1 | /* Statement Analysis and Transformation for Vectorization |
2 | Copyright (C) 2003-2024 Free Software Foundation, Inc. |
3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | |
6 | This file is part of GCC. |
7 | |
8 | GCC is free software; you can redistribute it and/or modify it under |
9 | the terms of the GNU General Public License as published by the Free |
10 | Software Foundation; either version 3, or (at your option) any later |
11 | version. |
12 | |
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
16 | for more details. |
17 | |
18 | You should have received a copy of the GNU General Public License |
19 | along with GCC; see the file COPYING3. If not see |
20 | <http://www.gnu.org/licenses/>. */ |
21 | |
22 | #include "config.h" |
23 | #include "system.h" |
24 | #include "coretypes.h" |
25 | #include "backend.h" |
26 | #include "target.h" |
27 | #include "rtl.h" |
28 | #include "tree.h" |
29 | #include "gimple.h" |
30 | #include "ssa.h" |
31 | #include "optabs-tree.h" |
32 | #include "insn-config.h" |
33 | #include "recog.h" /* FIXME: for insn_data */ |
34 | #include "cgraph.h" |
35 | #include "dumpfile.h" |
36 | #include "alias.h" |
37 | #include "fold-const.h" |
38 | #include "stor-layout.h" |
39 | #include "tree-eh.h" |
40 | #include "gimplify.h" |
41 | #include "gimple-iterator.h" |
42 | #include "gimplify-me.h" |
43 | #include "tree-cfg.h" |
44 | #include "tree-ssa-loop-manip.h" |
45 | #include "cfgloop.h" |
46 | #include "explow.h" |
47 | #include "tree-ssa-loop.h" |
48 | #include "tree-scalar-evolution.h" |
49 | #include "tree-vectorizer.h" |
50 | #include "builtins.h" |
51 | #include "internal-fn.h" |
52 | #include "tree-vector-builder.h" |
53 | #include "vec-perm-indices.h" |
54 | #include "gimple-range.h" |
55 | #include "tree-ssa-loop-niter.h" |
56 | #include "gimple-fold.h" |
57 | #include "regs.h" |
58 | #include "attribs.h" |
59 | #include "optabs-libfuncs.h" |
60 | |
61 | /* For lang_hooks.types.type_for_mode. */ |
62 | #include "langhooks.h" |
63 | |
64 | /* Return the vectorized type for the given statement. */ |
65 | |
66 | tree |
67 | stmt_vectype (class _stmt_vec_info *stmt_info) |
68 | { |
69 | return STMT_VINFO_VECTYPE (stmt_info); |
70 | } |
71 | |
72 | /* Return TRUE iff the given statement is in an inner loop relative to |
73 | the loop being vectorized. */ |
74 | bool |
75 | stmt_in_inner_loop_p (vec_info *vinfo, class _stmt_vec_info *stmt_info) |
76 | { |
77 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
78 | basic_block bb = gimple_bb (g: stmt); |
79 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
80 | class loop* loop; |
81 | |
82 | if (!loop_vinfo) |
83 | return false; |
84 | |
85 | loop = LOOP_VINFO_LOOP (loop_vinfo); |
86 | |
87 | return (bb->loop_father == loop->inner); |
88 | } |
89 | |
90 | /* Record the cost of a statement, either by directly informing the |
91 | target model or by saving it in a vector for later processing. |
92 | Return a preliminary estimate of the statement's cost. */ |
93 | |
94 | static unsigned |
95 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
96 | enum vect_cost_for_stmt kind, |
97 | stmt_vec_info stmt_info, slp_tree node, |
98 | tree vectype, int misalign, |
99 | enum vect_cost_model_location where) |
100 | { |
101 | if ((kind == vector_load || kind == unaligned_load) |
102 | && (stmt_info && STMT_VINFO_GATHER_SCATTER_P (stmt_info))) |
103 | kind = vector_gather_load; |
104 | if ((kind == vector_store || kind == unaligned_store) |
105 | && (stmt_info && STMT_VINFO_GATHER_SCATTER_P (stmt_info))) |
106 | kind = vector_scatter_store; |
107 | |
108 | stmt_info_for_cost si |
109 | = { .count: count, .kind: kind, .where: where, .stmt_info: stmt_info, .node: node, .vectype: vectype, .misalign: misalign }; |
110 | body_cost_vec->safe_push (obj: si); |
111 | |
112 | return (unsigned) |
113 | (builtin_vectorization_cost (type_of_cost: kind, vectype, misalign) * count); |
114 | } |
115 | |
116 | unsigned |
117 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
118 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
119 | tree vectype, int misalign, |
120 | enum vect_cost_model_location where) |
121 | { |
122 | return record_stmt_cost (body_cost_vec, count, kind, stmt_info, NULL, |
123 | vectype, misalign, where); |
124 | } |
125 | |
126 | unsigned |
127 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
128 | enum vect_cost_for_stmt kind, slp_tree node, |
129 | tree vectype, int misalign, |
130 | enum vect_cost_model_location where) |
131 | { |
132 | return record_stmt_cost (body_cost_vec, count, kind, NULL, node, |
133 | vectype, misalign, where); |
134 | } |
135 | |
136 | unsigned |
137 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
138 | enum vect_cost_for_stmt kind, |
139 | enum vect_cost_model_location where) |
140 | { |
141 | gcc_assert (kind == cond_branch_taken || kind == cond_branch_not_taken |
142 | || kind == scalar_stmt); |
143 | return record_stmt_cost (body_cost_vec, count, kind, NULL, NULL, |
144 | NULL_TREE, misalign: 0, where); |
145 | } |
146 | |
147 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
148 | |
149 | static tree |
150 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) |
151 | { |
152 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), |
153 | "vect_array" ); |
154 | } |
155 | |
156 | /* ARRAY is an array of vectors created by create_vector_array. |
157 | Return an SSA_NAME for the vector in index N. The reference |
158 | is part of the vectorization of STMT_INFO and the vector is associated |
159 | with scalar destination SCALAR_DEST. */ |
160 | |
161 | static tree |
162 | read_vector_array (vec_info *vinfo, |
163 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
164 | tree scalar_dest, tree array, unsigned HOST_WIDE_INT n) |
165 | { |
166 | tree vect_type, vect, vect_name, array_ref; |
167 | gimple *new_stmt; |
168 | |
169 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); |
170 | vect_type = TREE_TYPE (TREE_TYPE (array)); |
171 | vect = vect_create_destination_var (scalar_dest, vect_type); |
172 | array_ref = build4 (ARRAY_REF, vect_type, array, |
173 | build_int_cst (size_type_node, n), |
174 | NULL_TREE, NULL_TREE); |
175 | |
176 | new_stmt = gimple_build_assign (vect, array_ref); |
177 | vect_name = make_ssa_name (var: vect, stmt: new_stmt); |
178 | gimple_assign_set_lhs (gs: new_stmt, lhs: vect_name); |
179 | vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi); |
180 | |
181 | return vect_name; |
182 | } |
183 | |
184 | /* ARRAY is an array of vectors created by create_vector_array. |
185 | Emit code to store SSA_NAME VECT in index N of the array. |
186 | The store is part of the vectorization of STMT_INFO. */ |
187 | |
188 | static void |
189 | write_vector_array (vec_info *vinfo, |
190 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
191 | tree vect, tree array, unsigned HOST_WIDE_INT n) |
192 | { |
193 | tree array_ref; |
194 | gimple *new_stmt; |
195 | |
196 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, |
197 | build_int_cst (size_type_node, n), |
198 | NULL_TREE, NULL_TREE); |
199 | |
200 | new_stmt = gimple_build_assign (array_ref, vect); |
201 | vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi); |
202 | } |
203 | |
204 | /* PTR is a pointer to an array of type TYPE. Return a representation |
205 | of *PTR. The memory reference replaces those in FIRST_DR |
206 | (and its group). */ |
207 | |
208 | static tree |
209 | create_array_ref (tree type, tree ptr, tree alias_ptr_type) |
210 | { |
211 | tree mem_ref; |
212 | |
213 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); |
214 | /* Arrays have the same alignment as their type. */ |
215 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
216 | return mem_ref; |
217 | } |
218 | |
219 | /* Add a clobber of variable VAR to the vectorization of STMT_INFO. |
220 | Emit the clobber before *GSI. */ |
221 | |
222 | static void |
223 | vect_clobber_variable (vec_info *vinfo, stmt_vec_info stmt_info, |
224 | gimple_stmt_iterator *gsi, tree var) |
225 | { |
226 | tree clobber = build_clobber (TREE_TYPE (var)); |
227 | gimple *new_stmt = gimple_build_assign (var, clobber); |
228 | vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi); |
229 | } |
230 | |
231 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
232 | |
233 | /* Function vect_mark_relevant. |
234 | |
235 | Mark STMT_INFO as "relevant for vectorization" and add it to WORKLIST. */ |
236 | |
237 | static void |
238 | vect_mark_relevant (vec<stmt_vec_info> *worklist, stmt_vec_info stmt_info, |
239 | enum vect_relevant relevant, bool live_p) |
240 | { |
241 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); |
242 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); |
243 | |
244 | if (dump_enabled_p ()) |
245 | dump_printf_loc (MSG_NOTE, vect_location, |
246 | "mark relevant %d, live %d: %G" , relevant, live_p, |
247 | stmt_info->stmt); |
248 | |
249 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
250 | related pattern stmt instead of the original stmt. However, such stmts |
251 | may have their own uses that are not in any pattern, in such cases the |
252 | stmt itself should be marked. */ |
253 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
254 | { |
255 | /* This is the last stmt in a sequence that was detected as a |
256 | pattern that can potentially be vectorized. Don't mark the stmt |
257 | as relevant/live because it's not going to be vectorized. |
258 | Instead mark the pattern-stmt that replaces it. */ |
259 | |
260 | if (dump_enabled_p ()) |
261 | dump_printf_loc (MSG_NOTE, vect_location, |
262 | "last stmt in pattern. don't mark" |
263 | " relevant/live.\n" ); |
264 | |
265 | stmt_vec_info old_stmt_info = stmt_info; |
266 | stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); |
267 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == old_stmt_info); |
268 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); |
269 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); |
270 | |
271 | if (live_p && relevant == vect_unused_in_scope) |
272 | { |
273 | if (dump_enabled_p ()) |
274 | dump_printf_loc (MSG_NOTE, vect_location, |
275 | "vec_stmt_relevant_p: forcing live pattern stmt " |
276 | "relevant.\n" ); |
277 | relevant = vect_used_only_live; |
278 | } |
279 | |
280 | if (dump_enabled_p ()) |
281 | dump_printf_loc (MSG_NOTE, vect_location, |
282 | "mark relevant %d, live %d: %G" , relevant, live_p, |
283 | stmt_info->stmt); |
284 | } |
285 | |
286 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; |
287 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) |
288 | STMT_VINFO_RELEVANT (stmt_info) = relevant; |
289 | |
290 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant |
291 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) |
292 | { |
293 | if (dump_enabled_p ()) |
294 | dump_printf_loc (MSG_NOTE, vect_location, |
295 | "already marked relevant/live.\n" ); |
296 | return; |
297 | } |
298 | |
299 | worklist->safe_push (obj: stmt_info); |
300 | } |
301 | |
302 | |
303 | /* Function is_simple_and_all_uses_invariant |
304 | |
305 | Return true if STMT_INFO is simple and all uses of it are invariant. */ |
306 | |
307 | bool |
308 | is_simple_and_all_uses_invariant (stmt_vec_info stmt_info, |
309 | loop_vec_info loop_vinfo) |
310 | { |
311 | tree op; |
312 | ssa_op_iter iter; |
313 | |
314 | gassign *stmt = dyn_cast <gassign *> (p: stmt_info->stmt); |
315 | if (!stmt) |
316 | return false; |
317 | |
318 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) |
319 | { |
320 | enum vect_def_type dt = vect_uninitialized_def; |
321 | |
322 | if (!vect_is_simple_use (op, loop_vinfo, &dt)) |
323 | { |
324 | if (dump_enabled_p ()) |
325 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
326 | "use not simple.\n" ); |
327 | return false; |
328 | } |
329 | |
330 | if (dt != vect_external_def && dt != vect_constant_def) |
331 | return false; |
332 | } |
333 | return true; |
334 | } |
335 | |
336 | /* Function vect_stmt_relevant_p. |
337 | |
338 | Return true if STMT_INFO, in the loop that is represented by LOOP_VINFO, |
339 | is "relevant for vectorization". |
340 | |
341 | A stmt is considered "relevant for vectorization" if: |
342 | - it has uses outside the loop. |
343 | - it has vdefs (it alters memory). |
344 | - control stmts in the loop (except for the exit condition). |
345 | - it is an induction and we have multiple exits. |
346 | |
347 | CHECKME: what other side effects would the vectorizer allow? */ |
348 | |
349 | static bool |
350 | vect_stmt_relevant_p (stmt_vec_info stmt_info, loop_vec_info loop_vinfo, |
351 | enum vect_relevant *relevant, bool *live_p) |
352 | { |
353 | class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
354 | ssa_op_iter op_iter; |
355 | imm_use_iterator imm_iter; |
356 | use_operand_p use_p; |
357 | def_operand_p def_p; |
358 | |
359 | *relevant = vect_unused_in_scope; |
360 | *live_p = false; |
361 | |
362 | /* cond stmt other than loop exit cond. */ |
363 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
364 | if (is_ctrl_stmt (stmt) |
365 | && LOOP_VINFO_LOOP_IV_COND (loop_vinfo) != stmt |
366 | && (!loop->inner || gimple_bb (g: stmt)->loop_father == loop)) |
367 | *relevant = vect_used_in_scope; |
368 | |
369 | /* changing memory. */ |
370 | if (gimple_code (g: stmt_info->stmt) != GIMPLE_PHI) |
371 | if (gimple_vdef (g: stmt_info->stmt) |
372 | && !gimple_clobber_p (s: stmt_info->stmt)) |
373 | { |
374 | if (dump_enabled_p ()) |
375 | dump_printf_loc (MSG_NOTE, vect_location, |
376 | "vec_stmt_relevant_p: stmt has vdefs.\n" ); |
377 | *relevant = vect_used_in_scope; |
378 | } |
379 | |
380 | /* uses outside the loop. */ |
381 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt_info->stmt, op_iter, SSA_OP_DEF) |
382 | { |
383 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) |
384 | { |
385 | basic_block bb = gimple_bb (USE_STMT (use_p)); |
386 | if (!flow_bb_inside_loop_p (loop, bb)) |
387 | { |
388 | if (is_gimple_debug (USE_STMT (use_p))) |
389 | continue; |
390 | |
391 | if (dump_enabled_p ()) |
392 | dump_printf_loc (MSG_NOTE, vect_location, |
393 | "vec_stmt_relevant_p: used out of loop.\n" ); |
394 | |
395 | /* We expect all such uses to be in the loop exit phis |
396 | (because of loop closed form) */ |
397 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); |
398 | |
399 | *live_p = true; |
400 | } |
401 | } |
402 | } |
403 | |
404 | /* Check if it's an induction and multiple exits. In this case there will be |
405 | a usage later on after peeling which is needed for the alternate exit. */ |
406 | if (LOOP_VINFO_EARLY_BREAKS (loop_vinfo) |
407 | && STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def) |
408 | { |
409 | if (dump_enabled_p ()) |
410 | dump_printf_loc (MSG_NOTE, vect_location, |
411 | "vec_stmt_relevant_p: induction forced for " |
412 | "early break.\n" ); |
413 | *live_p = true; |
414 | |
415 | } |
416 | |
417 | if (*live_p && *relevant == vect_unused_in_scope |
418 | && !is_simple_and_all_uses_invariant (stmt_info, loop_vinfo)) |
419 | { |
420 | if (dump_enabled_p ()) |
421 | dump_printf_loc (MSG_NOTE, vect_location, |
422 | "vec_stmt_relevant_p: stmt live but not relevant.\n" ); |
423 | *relevant = vect_used_only_live; |
424 | } |
425 | |
426 | return (*live_p || *relevant); |
427 | } |
428 | |
429 | |
430 | /* Function exist_non_indexing_operands_for_use_p |
431 | |
432 | USE is one of the uses attached to STMT_INFO. Check if USE is |
433 | used in STMT_INFO for anything other than indexing an array. */ |
434 | |
435 | static bool |
436 | exist_non_indexing_operands_for_use_p (tree use, stmt_vec_info stmt_info) |
437 | { |
438 | tree operand; |
439 | |
440 | /* USE corresponds to some operand in STMT. If there is no data |
441 | reference in STMT, then any operand that corresponds to USE |
442 | is not indexing an array. */ |
443 | if (!STMT_VINFO_DATA_REF (stmt_info)) |
444 | return true; |
445 | |
446 | /* STMT has a data_ref. FORNOW this means that its of one of |
447 | the following forms: |
448 | -1- ARRAY_REF = var |
449 | -2- var = ARRAY_REF |
450 | (This should have been verified in analyze_data_refs). |
451 | |
452 | 'var' in the second case corresponds to a def, not a use, |
453 | so USE cannot correspond to any operands that are not used |
454 | for array indexing. |
455 | |
456 | Therefore, all we need to check is if STMT falls into the |
457 | first case, and whether var corresponds to USE. */ |
458 | |
459 | gassign *assign = dyn_cast <gassign *> (p: stmt_info->stmt); |
460 | if (!assign || !gimple_assign_copy_p (assign)) |
461 | { |
462 | gcall *call = dyn_cast <gcall *> (p: stmt_info->stmt); |
463 | if (call && gimple_call_internal_p (gs: call)) |
464 | { |
465 | internal_fn ifn = gimple_call_internal_fn (gs: call); |
466 | int mask_index = internal_fn_mask_index (ifn); |
467 | if (mask_index >= 0 |
468 | && use == gimple_call_arg (gs: call, index: mask_index)) |
469 | return true; |
470 | int stored_value_index = internal_fn_stored_value_index (ifn); |
471 | if (stored_value_index >= 0 |
472 | && use == gimple_call_arg (gs: call, index: stored_value_index)) |
473 | return true; |
474 | if (internal_gather_scatter_fn_p (ifn) |
475 | && use == gimple_call_arg (gs: call, index: 1)) |
476 | return true; |
477 | } |
478 | return false; |
479 | } |
480 | |
481 | if (TREE_CODE (gimple_assign_lhs (assign)) == SSA_NAME) |
482 | return false; |
483 | operand = gimple_assign_rhs1 (gs: assign); |
484 | if (TREE_CODE (operand) != SSA_NAME) |
485 | return false; |
486 | |
487 | if (operand == use) |
488 | return true; |
489 | |
490 | return false; |
491 | } |
492 | |
493 | |
494 | /* |
495 | Function process_use. |
496 | |
497 | Inputs: |
498 | - a USE in STMT_VINFO in a loop represented by LOOP_VINFO |
499 | - RELEVANT - enum value to be set in the STMT_VINFO of the stmt |
500 | that defined USE. This is done by calling mark_relevant and passing it |
501 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
502 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
503 | be performed. |
504 | |
505 | Outputs: |
506 | Generally, LIVE_P and RELEVANT are used to define the liveness and |
507 | relevance info of the DEF_STMT of this USE: |
508 | STMT_VINFO_LIVE_P (DEF_stmt_vinfo) <-- live_p |
509 | STMT_VINFO_RELEVANT (DEF_stmt_vinfo) <-- relevant |
510 | Exceptions: |
511 | - case 1: If USE is used only for address computations (e.g. array indexing), |
512 | which does not need to be directly vectorized, then the liveness/relevance |
513 | of the respective DEF_STMT is left unchanged. |
514 | - case 2: If STMT_VINFO is a reduction phi and DEF_STMT is a reduction stmt, |
515 | we skip DEF_STMT cause it had already been processed. |
516 | - case 3: If DEF_STMT and STMT_VINFO are in different nests, then |
517 | "relevant" will be modified accordingly. |
518 | |
519 | Return true if everything is as expected. Return false otherwise. */ |
520 | |
521 | static opt_result |
522 | process_use (stmt_vec_info stmt_vinfo, tree use, loop_vec_info loop_vinfo, |
523 | enum vect_relevant relevant, vec<stmt_vec_info> *worklist, |
524 | bool force) |
525 | { |
526 | stmt_vec_info dstmt_vinfo; |
527 | enum vect_def_type dt; |
528 | |
529 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
530 | that are used for address computation are not considered relevant. */ |
531 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt_info: stmt_vinfo)) |
532 | return opt_result::success (); |
533 | |
534 | if (!vect_is_simple_use (use, loop_vinfo, &dt, &dstmt_vinfo)) |
535 | return opt_result::failure_at (loc: stmt_vinfo->stmt, |
536 | fmt: "not vectorized:" |
537 | " unsupported use in stmt.\n" ); |
538 | |
539 | if (!dstmt_vinfo) |
540 | return opt_result::success (); |
541 | |
542 | basic_block def_bb = gimple_bb (g: dstmt_vinfo->stmt); |
543 | basic_block bb = gimple_bb (g: stmt_vinfo->stmt); |
544 | |
545 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DSTMT_VINFO). |
546 | We have to force the stmt live since the epilogue loop needs it to |
547 | continue computing the reduction. */ |
548 | if (gimple_code (g: stmt_vinfo->stmt) == GIMPLE_PHI |
549 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
550 | && gimple_code (g: dstmt_vinfo->stmt) != GIMPLE_PHI |
551 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def |
552 | && bb->loop_father == def_bb->loop_father) |
553 | { |
554 | if (dump_enabled_p ()) |
555 | dump_printf_loc (MSG_NOTE, vect_location, |
556 | "reduc-stmt defining reduc-phi in the same nest.\n" ); |
557 | vect_mark_relevant (worklist, stmt_info: dstmt_vinfo, relevant, live_p: true); |
558 | return opt_result::success (); |
559 | } |
560 | |
561 | /* case 3a: outer-loop stmt defining an inner-loop stmt: |
562 | outer-loop-header-bb: |
563 | d = dstmt_vinfo |
564 | inner-loop: |
565 | stmt # use (d) |
566 | outer-loop-tail-bb: |
567 | ... */ |
568 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) |
569 | { |
570 | if (dump_enabled_p ()) |
571 | dump_printf_loc (MSG_NOTE, vect_location, |
572 | "outer-loop def-stmt defining inner-loop stmt.\n" ); |
573 | |
574 | switch (relevant) |
575 | { |
576 | case vect_unused_in_scope: |
577 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
578 | vect_used_in_scope : vect_unused_in_scope; |
579 | break; |
580 | |
581 | case vect_used_in_outer_by_reduction: |
582 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
583 | relevant = vect_used_by_reduction; |
584 | break; |
585 | |
586 | case vect_used_in_outer: |
587 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
588 | relevant = vect_used_in_scope; |
589 | break; |
590 | |
591 | case vect_used_in_scope: |
592 | break; |
593 | |
594 | default: |
595 | gcc_unreachable (); |
596 | } |
597 | } |
598 | |
599 | /* case 3b: inner-loop stmt defining an outer-loop stmt: |
600 | outer-loop-header-bb: |
601 | ... |
602 | inner-loop: |
603 | d = dstmt_vinfo |
604 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
605 | stmt # use (d) */ |
606 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) |
607 | { |
608 | if (dump_enabled_p ()) |
609 | dump_printf_loc (MSG_NOTE, vect_location, |
610 | "inner-loop def-stmt defining outer-loop stmt.\n" ); |
611 | |
612 | switch (relevant) |
613 | { |
614 | case vect_unused_in_scope: |
615 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
616 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
617 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
618 | break; |
619 | |
620 | case vect_used_by_reduction: |
621 | case vect_used_only_live: |
622 | relevant = vect_used_in_outer_by_reduction; |
623 | break; |
624 | |
625 | case vect_used_in_scope: |
626 | relevant = vect_used_in_outer; |
627 | break; |
628 | |
629 | default: |
630 | gcc_unreachable (); |
631 | } |
632 | } |
633 | /* We are also not interested in uses on loop PHI backedges that are |
634 | inductions. Otherwise we'll needlessly vectorize the IV increment |
635 | and cause hybrid SLP for SLP inductions. Unless the PHI is live |
636 | of course. */ |
637 | else if (gimple_code (g: stmt_vinfo->stmt) == GIMPLE_PHI |
638 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_induction_def |
639 | && ! STMT_VINFO_LIVE_P (stmt_vinfo) |
640 | && (PHI_ARG_DEF_FROM_EDGE (stmt_vinfo->stmt, |
641 | loop_latch_edge (bb->loop_father)) |
642 | == use)) |
643 | { |
644 | if (dump_enabled_p ()) |
645 | dump_printf_loc (MSG_NOTE, vect_location, |
646 | "induction value on backedge.\n" ); |
647 | return opt_result::success (); |
648 | } |
649 | |
650 | |
651 | vect_mark_relevant (worklist, stmt_info: dstmt_vinfo, relevant, live_p: false); |
652 | return opt_result::success (); |
653 | } |
654 | |
655 | |
656 | /* Function vect_mark_stmts_to_be_vectorized. |
657 | |
658 | Not all stmts in the loop need to be vectorized. For example: |
659 | |
660 | for i... |
661 | for j... |
662 | 1. T0 = i + j |
663 | 2. T1 = a[T0] |
664 | |
665 | 3. j = j + 1 |
666 | |
667 | Stmt 1 and 3 do not need to be vectorized, because loop control and |
668 | addressing of vectorized data-refs are handled differently. |
669 | |
670 | This pass detects such stmts. */ |
671 | |
672 | opt_result |
673 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo, bool *fatal) |
674 | { |
675 | class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
676 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); |
677 | unsigned int nbbs = loop->num_nodes; |
678 | gimple_stmt_iterator si; |
679 | unsigned int i; |
680 | basic_block bb; |
681 | bool live_p; |
682 | enum vect_relevant relevant; |
683 | |
684 | DUMP_VECT_SCOPE ("vect_mark_stmts_to_be_vectorized" ); |
685 | |
686 | auto_vec<stmt_vec_info, 64> worklist; |
687 | |
688 | /* 1. Init worklist. */ |
689 | for (i = 0; i < nbbs; i++) |
690 | { |
691 | bb = bbs[i]; |
692 | for (si = gsi_start_phis (bb); !gsi_end_p (i: si); gsi_next (i: &si)) |
693 | { |
694 | stmt_vec_info phi_info = loop_vinfo->lookup_stmt (gsi_stmt (i: si)); |
695 | if (dump_enabled_p ()) |
696 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? %G" , |
697 | phi_info->stmt); |
698 | |
699 | if (vect_stmt_relevant_p (stmt_info: phi_info, loop_vinfo, relevant: &relevant, live_p: &live_p)) |
700 | vect_mark_relevant (worklist: &worklist, stmt_info: phi_info, relevant, live_p); |
701 | } |
702 | for (si = gsi_start_bb (bb); !gsi_end_p (i: si); gsi_next (i: &si)) |
703 | { |
704 | if (is_gimple_debug (gs: gsi_stmt (i: si))) |
705 | continue; |
706 | stmt_vec_info stmt_info = loop_vinfo->lookup_stmt (gsi_stmt (i: si)); |
707 | if (dump_enabled_p ()) |
708 | dump_printf_loc (MSG_NOTE, vect_location, |
709 | "init: stmt relevant? %G" , stmt_info->stmt); |
710 | |
711 | if (vect_stmt_relevant_p (stmt_info, loop_vinfo, relevant: &relevant, live_p: &live_p)) |
712 | vect_mark_relevant (worklist: &worklist, stmt_info, relevant, live_p); |
713 | } |
714 | } |
715 | |
716 | /* 2. Process_worklist */ |
717 | while (worklist.length () > 0) |
718 | { |
719 | use_operand_p use_p; |
720 | ssa_op_iter iter; |
721 | |
722 | stmt_vec_info stmt_vinfo = worklist.pop (); |
723 | if (dump_enabled_p ()) |
724 | dump_printf_loc (MSG_NOTE, vect_location, |
725 | "worklist: examine stmt: %G" , stmt_vinfo->stmt); |
726 | |
727 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
728 | (DEF_STMT) as relevant/irrelevant according to the relevance property |
729 | of STMT. */ |
730 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); |
731 | |
732 | /* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is |
733 | propagated as is to the DEF_STMTs of its USEs. |
734 | |
735 | One exception is when STMT has been identified as defining a reduction |
736 | variable; in this case we set the relevance to vect_used_by_reduction. |
737 | This is because we distinguish between two kinds of relevant stmts - |
738 | those that are used by a reduction computation, and those that are |
739 | (also) used by a regular computation. This allows us later on to |
740 | identify stmts that are used solely by a reduction, and therefore the |
741 | order of the results that they produce does not have to be kept. */ |
742 | |
743 | switch (STMT_VINFO_DEF_TYPE (stmt_vinfo)) |
744 | { |
745 | case vect_reduction_def: |
746 | gcc_assert (relevant != vect_unused_in_scope); |
747 | if (relevant != vect_unused_in_scope |
748 | && relevant != vect_used_in_scope |
749 | && relevant != vect_used_by_reduction |
750 | && relevant != vect_used_only_live) |
751 | return opt_result::failure_at |
752 | (loc: stmt_vinfo->stmt, fmt: "unsupported use of reduction.\n" ); |
753 | break; |
754 | |
755 | case vect_nested_cycle: |
756 | if (relevant != vect_unused_in_scope |
757 | && relevant != vect_used_in_outer_by_reduction |
758 | && relevant != vect_used_in_outer) |
759 | return opt_result::failure_at |
760 | (loc: stmt_vinfo->stmt, fmt: "unsupported use of nested cycle.\n" ); |
761 | break; |
762 | |
763 | case vect_double_reduction_def: |
764 | if (relevant != vect_unused_in_scope |
765 | && relevant != vect_used_by_reduction |
766 | && relevant != vect_used_only_live) |
767 | return opt_result::failure_at |
768 | (loc: stmt_vinfo->stmt, fmt: "unsupported use of double reduction.\n" ); |
769 | break; |
770 | |
771 | default: |
772 | break; |
773 | } |
774 | |
775 | if (is_pattern_stmt_p (stmt_info: stmt_vinfo)) |
776 | { |
777 | /* Pattern statements are not inserted into the code, so |
778 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we |
779 | have to scan the RHS or function arguments instead. */ |
780 | if (gassign *assign = dyn_cast <gassign *> (p: stmt_vinfo->stmt)) |
781 | { |
782 | enum tree_code rhs_code = gimple_assign_rhs_code (gs: assign); |
783 | tree op = gimple_assign_rhs1 (gs: assign); |
784 | |
785 | i = 1; |
786 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) |
787 | { |
788 | opt_result res |
789 | = process_use (stmt_vinfo, TREE_OPERAND (op, 0), |
790 | loop_vinfo, relevant, worklist: &worklist, force: false); |
791 | if (!res) |
792 | return res; |
793 | res = process_use (stmt_vinfo, TREE_OPERAND (op, 1), |
794 | loop_vinfo, relevant, worklist: &worklist, force: false); |
795 | if (!res) |
796 | return res; |
797 | i = 2; |
798 | } |
799 | for (; i < gimple_num_ops (gs: assign); i++) |
800 | { |
801 | op = gimple_op (gs: assign, i); |
802 | if (TREE_CODE (op) == SSA_NAME) |
803 | { |
804 | opt_result res |
805 | = process_use (stmt_vinfo, use: op, loop_vinfo, relevant, |
806 | worklist: &worklist, force: false); |
807 | if (!res) |
808 | return res; |
809 | } |
810 | } |
811 | } |
812 | else if (gcond *cond = dyn_cast <gcond *> (p: stmt_vinfo->stmt)) |
813 | { |
814 | tree_code rhs_code = gimple_cond_code (gs: cond); |
815 | gcc_assert (TREE_CODE_CLASS (rhs_code) == tcc_comparison); |
816 | opt_result res |
817 | = process_use (stmt_vinfo, use: gimple_cond_lhs (gs: cond), |
818 | loop_vinfo, relevant, worklist: &worklist, force: false); |
819 | if (!res) |
820 | return res; |
821 | res = process_use (stmt_vinfo, use: gimple_cond_rhs (gs: cond), |
822 | loop_vinfo, relevant, worklist: &worklist, force: false); |
823 | if (!res) |
824 | return res; |
825 | } |
826 | else if (gcall *call = dyn_cast <gcall *> (p: stmt_vinfo->stmt)) |
827 | { |
828 | for (i = 0; i < gimple_call_num_args (gs: call); i++) |
829 | { |
830 | tree arg = gimple_call_arg (gs: call, index: i); |
831 | opt_result res |
832 | = process_use (stmt_vinfo, use: arg, loop_vinfo, relevant, |
833 | worklist: &worklist, force: false); |
834 | if (!res) |
835 | return res; |
836 | } |
837 | } |
838 | else |
839 | gcc_unreachable (); |
840 | } |
841 | else |
842 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt_vinfo->stmt, iter, SSA_OP_USE) |
843 | { |
844 | tree op = USE_FROM_PTR (use_p); |
845 | opt_result res |
846 | = process_use (stmt_vinfo, use: op, loop_vinfo, relevant, |
847 | worklist: &worklist, force: false); |
848 | if (!res) |
849 | return res; |
850 | } |
851 | |
852 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
853 | { |
854 | gather_scatter_info gs_info; |
855 | if (!vect_check_gather_scatter (stmt_vinfo, loop_vinfo, &gs_info)) |
856 | gcc_unreachable (); |
857 | opt_result res |
858 | = process_use (stmt_vinfo, use: gs_info.offset, loop_vinfo, relevant, |
859 | worklist: &worklist, force: true); |
860 | if (!res) |
861 | { |
862 | if (fatal) |
863 | *fatal = false; |
864 | return res; |
865 | } |
866 | } |
867 | } /* while worklist */ |
868 | |
869 | return opt_result::success (); |
870 | } |
871 | |
872 | /* Function vect_model_simple_cost. |
873 | |
874 | Models cost for simple operations, i.e. those that only emit ncopies of a |
875 | single op. Right now, this does not account for multiple insns that could |
876 | be generated for the single vector op. We will handle that shortly. */ |
877 | |
878 | static void |
879 | vect_model_simple_cost (vec_info *, |
880 | stmt_vec_info stmt_info, int ncopies, |
881 | enum vect_def_type *dt, |
882 | int ndts, |
883 | slp_tree node, |
884 | stmt_vector_for_cost *cost_vec, |
885 | vect_cost_for_stmt kind = vector_stmt) |
886 | { |
887 | int inside_cost = 0, prologue_cost = 0; |
888 | |
889 | gcc_assert (cost_vec != NULL); |
890 | |
891 | /* ??? Somehow we need to fix this at the callers. */ |
892 | if (node) |
893 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (node); |
894 | |
895 | if (!node) |
896 | /* Cost the "broadcast" of a scalar operand in to a vector operand. |
897 | Use scalar_to_vec to cost the broadcast, as elsewhere in the vector |
898 | cost model. */ |
899 | for (int i = 0; i < ndts; i++) |
900 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
901 | prologue_cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: scalar_to_vec, |
902 | stmt_info, misalign: 0, where: vect_prologue); |
903 | |
904 | /* Pass the inside-of-loop statements to the target-specific cost model. */ |
905 | inside_cost += record_stmt_cost (body_cost_vec: cost_vec, count: ncopies, kind, |
906 | stmt_info, misalign: 0, where: vect_body); |
907 | |
908 | if (dump_enabled_p ()) |
909 | dump_printf_loc (MSG_NOTE, vect_location, |
910 | "vect_model_simple_cost: inside_cost = %d, " |
911 | "prologue_cost = %d .\n" , inside_cost, prologue_cost); |
912 | } |
913 | |
914 | |
915 | /* Model cost for type demotion and promotion operations. PWR is |
916 | normally zero for single-step promotions and demotions. It will be |
917 | one if two-step promotion/demotion is required, and so on. NCOPIES |
918 | is the number of vector results (and thus number of instructions) |
919 | for the narrowest end of the operation chain. Each additional |
920 | step doubles the number of instructions required. If WIDEN_ARITH |
921 | is true the stmt is doing widening arithmetic. */ |
922 | |
923 | static void |
924 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, |
925 | enum vect_def_type *dt, |
926 | unsigned int ncopies, int pwr, |
927 | stmt_vector_for_cost *cost_vec, |
928 | bool widen_arith) |
929 | { |
930 | int i; |
931 | int inside_cost = 0, prologue_cost = 0; |
932 | |
933 | for (i = 0; i < pwr + 1; i++) |
934 | { |
935 | inside_cost += record_stmt_cost (body_cost_vec: cost_vec, count: ncopies, |
936 | kind: widen_arith |
937 | ? vector_stmt : vec_promote_demote, |
938 | stmt_info, misalign: 0, where: vect_body); |
939 | ncopies *= 2; |
940 | } |
941 | |
942 | /* FORNOW: Assuming maximum 2 args per stmts. */ |
943 | for (i = 0; i < 2; i++) |
944 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
945 | prologue_cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: vector_stmt, |
946 | stmt_info, misalign: 0, where: vect_prologue); |
947 | |
948 | if (dump_enabled_p ()) |
949 | dump_printf_loc (MSG_NOTE, vect_location, |
950 | "vect_model_promotion_demotion_cost: inside_cost = %d, " |
951 | "prologue_cost = %d .\n" , inside_cost, prologue_cost); |
952 | } |
953 | |
954 | /* Returns true if the current function returns DECL. */ |
955 | |
956 | static bool |
957 | cfun_returns (tree decl) |
958 | { |
959 | edge_iterator ei; |
960 | edge e; |
961 | FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds) |
962 | { |
963 | greturn *ret = safe_dyn_cast <greturn *> (p: *gsi_last_bb (bb: e->src)); |
964 | if (!ret) |
965 | continue; |
966 | if (gimple_return_retval (gs: ret) == decl) |
967 | return true; |
968 | /* We often end up with an aggregate copy to the result decl, |
969 | handle that case as well. First skip intermediate clobbers |
970 | though. */ |
971 | gimple *def = ret; |
972 | do |
973 | { |
974 | def = SSA_NAME_DEF_STMT (gimple_vuse (def)); |
975 | } |
976 | while (gimple_clobber_p (s: def)); |
977 | if (is_a <gassign *> (p: def) |
978 | && gimple_assign_lhs (gs: def) == gimple_return_retval (gs: ret) |
979 | && gimple_assign_rhs1 (gs: def) == decl) |
980 | return true; |
981 | } |
982 | return false; |
983 | } |
984 | |
985 | /* Calculate cost of DR's memory access. */ |
986 | void |
987 | vect_get_store_cost (vec_info *, stmt_vec_info stmt_info, int ncopies, |
988 | dr_alignment_support alignment_support_scheme, |
989 | int misalignment, |
990 | unsigned int *inside_cost, |
991 | stmt_vector_for_cost *body_cost_vec) |
992 | { |
993 | switch (alignment_support_scheme) |
994 | { |
995 | case dr_aligned: |
996 | { |
997 | *inside_cost += record_stmt_cost (body_cost_vec, count: ncopies, |
998 | kind: vector_store, stmt_info, misalign: 0, |
999 | where: vect_body); |
1000 | |
1001 | if (dump_enabled_p ()) |
1002 | dump_printf_loc (MSG_NOTE, vect_location, |
1003 | "vect_model_store_cost: aligned.\n" ); |
1004 | break; |
1005 | } |
1006 | |
1007 | case dr_unaligned_supported: |
1008 | { |
1009 | /* Here, we assign an additional cost for the unaligned store. */ |
1010 | *inside_cost += record_stmt_cost (body_cost_vec, count: ncopies, |
1011 | kind: unaligned_store, stmt_info, |
1012 | misalign: misalignment, where: vect_body); |
1013 | if (dump_enabled_p ()) |
1014 | dump_printf_loc (MSG_NOTE, vect_location, |
1015 | "vect_model_store_cost: unaligned supported by " |
1016 | "hardware.\n" ); |
1017 | break; |
1018 | } |
1019 | |
1020 | case dr_unaligned_unsupported: |
1021 | { |
1022 | *inside_cost = VECT_MAX_COST; |
1023 | |
1024 | if (dump_enabled_p ()) |
1025 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1026 | "vect_model_store_cost: unsupported access.\n" ); |
1027 | break; |
1028 | } |
1029 | |
1030 | default: |
1031 | gcc_unreachable (); |
1032 | } |
1033 | } |
1034 | |
1035 | /* Calculate cost of DR's memory access. */ |
1036 | void |
1037 | vect_get_load_cost (vec_info *, stmt_vec_info stmt_info, int ncopies, |
1038 | dr_alignment_support alignment_support_scheme, |
1039 | int misalignment, |
1040 | bool add_realign_cost, unsigned int *inside_cost, |
1041 | unsigned int *prologue_cost, |
1042 | stmt_vector_for_cost *prologue_cost_vec, |
1043 | stmt_vector_for_cost *body_cost_vec, |
1044 | bool record_prologue_costs) |
1045 | { |
1046 | switch (alignment_support_scheme) |
1047 | { |
1048 | case dr_aligned: |
1049 | { |
1050 | *inside_cost += record_stmt_cost (body_cost_vec, count: ncopies, kind: vector_load, |
1051 | stmt_info, misalign: 0, where: vect_body); |
1052 | |
1053 | if (dump_enabled_p ()) |
1054 | dump_printf_loc (MSG_NOTE, vect_location, |
1055 | "vect_model_load_cost: aligned.\n" ); |
1056 | |
1057 | break; |
1058 | } |
1059 | case dr_unaligned_supported: |
1060 | { |
1061 | /* Here, we assign an additional cost for the unaligned load. */ |
1062 | *inside_cost += record_stmt_cost (body_cost_vec, count: ncopies, |
1063 | kind: unaligned_load, stmt_info, |
1064 | misalign: misalignment, where: vect_body); |
1065 | |
1066 | if (dump_enabled_p ()) |
1067 | dump_printf_loc (MSG_NOTE, vect_location, |
1068 | "vect_model_load_cost: unaligned supported by " |
1069 | "hardware.\n" ); |
1070 | |
1071 | break; |
1072 | } |
1073 | case dr_explicit_realign: |
1074 | { |
1075 | *inside_cost += record_stmt_cost (body_cost_vec, count: ncopies * 2, |
1076 | kind: vector_load, stmt_info, misalign: 0, where: vect_body); |
1077 | *inside_cost += record_stmt_cost (body_cost_vec, count: ncopies, |
1078 | kind: vec_perm, stmt_info, misalign: 0, where: vect_body); |
1079 | |
1080 | /* FIXME: If the misalignment remains fixed across the iterations of |
1081 | the containing loop, the following cost should be added to the |
1082 | prologue costs. */ |
1083 | if (targetm.vectorize.builtin_mask_for_load) |
1084 | *inside_cost += record_stmt_cost (body_cost_vec, count: 1, kind: vector_stmt, |
1085 | stmt_info, misalign: 0, where: vect_body); |
1086 | |
1087 | if (dump_enabled_p ()) |
1088 | dump_printf_loc (MSG_NOTE, vect_location, |
1089 | "vect_model_load_cost: explicit realign\n" ); |
1090 | |
1091 | break; |
1092 | } |
1093 | case dr_explicit_realign_optimized: |
1094 | { |
1095 | if (dump_enabled_p ()) |
1096 | dump_printf_loc (MSG_NOTE, vect_location, |
1097 | "vect_model_load_cost: unaligned software " |
1098 | "pipelined.\n" ); |
1099 | |
1100 | /* Unaligned software pipeline has a load of an address, an initial |
1101 | load, and possibly a mask operation to "prime" the loop. However, |
1102 | if this is an access in a group of loads, which provide grouped |
1103 | access, then the above cost should only be considered for one |
1104 | access in the group. Inside the loop, there is a load op |
1105 | and a realignment op. */ |
1106 | |
1107 | if (add_realign_cost && record_prologue_costs) |
1108 | { |
1109 | *prologue_cost += record_stmt_cost (body_cost_vec: prologue_cost_vec, count: 2, |
1110 | kind: vector_stmt, stmt_info, |
1111 | misalign: 0, where: vect_prologue); |
1112 | if (targetm.vectorize.builtin_mask_for_load) |
1113 | *prologue_cost += record_stmt_cost (body_cost_vec: prologue_cost_vec, count: 1, |
1114 | kind: vector_stmt, stmt_info, |
1115 | misalign: 0, where: vect_prologue); |
1116 | } |
1117 | |
1118 | *inside_cost += record_stmt_cost (body_cost_vec, count: ncopies, kind: vector_load, |
1119 | stmt_info, misalign: 0, where: vect_body); |
1120 | *inside_cost += record_stmt_cost (body_cost_vec, count: ncopies, kind: vec_perm, |
1121 | stmt_info, misalign: 0, where: vect_body); |
1122 | |
1123 | if (dump_enabled_p ()) |
1124 | dump_printf_loc (MSG_NOTE, vect_location, |
1125 | "vect_model_load_cost: explicit realign optimized" |
1126 | "\n" ); |
1127 | |
1128 | break; |
1129 | } |
1130 | |
1131 | case dr_unaligned_unsupported: |
1132 | { |
1133 | *inside_cost = VECT_MAX_COST; |
1134 | |
1135 | if (dump_enabled_p ()) |
1136 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1137 | "vect_model_load_cost: unsupported access.\n" ); |
1138 | break; |
1139 | } |
1140 | |
1141 | default: |
1142 | gcc_unreachable (); |
1143 | } |
1144 | } |
1145 | |
1146 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1147 | the loop preheader for the vectorized stmt STMT_VINFO. */ |
1148 | |
1149 | static void |
1150 | vect_init_vector_1 (vec_info *vinfo, stmt_vec_info stmt_vinfo, gimple *new_stmt, |
1151 | gimple_stmt_iterator *gsi) |
1152 | { |
1153 | if (gsi) |
1154 | vect_finish_stmt_generation (vinfo, stmt_vinfo, new_stmt, gsi); |
1155 | else |
1156 | vinfo->insert_on_entry (stmt_vinfo, new_stmt); |
1157 | |
1158 | if (dump_enabled_p ()) |
1159 | dump_printf_loc (MSG_NOTE, vect_location, |
1160 | "created new init_stmt: %G" , new_stmt); |
1161 | } |
1162 | |
1163 | /* Function vect_init_vector. |
1164 | |
1165 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1166 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have |
1167 | vector type a vector with all elements equal to VAL is created first. |
1168 | Place the initialization at GSI if it is not NULL. Otherwise, place the |
1169 | initialization at the loop preheader. |
1170 | Return the DEF of INIT_STMT. |
1171 | It will be used in the vectorization of STMT_INFO. */ |
1172 | |
1173 | tree |
1174 | vect_init_vector (vec_info *vinfo, stmt_vec_info stmt_info, tree val, tree type, |
1175 | gimple_stmt_iterator *gsi) |
1176 | { |
1177 | gimple *init_stmt; |
1178 | tree new_temp; |
1179 | |
1180 | /* We abuse this function to push sth to a SSA name with initial 'val'. */ |
1181 | if (! useless_type_conversion_p (type, TREE_TYPE (val))) |
1182 | { |
1183 | gcc_assert (VECTOR_TYPE_P (type)); |
1184 | if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) |
1185 | { |
1186 | /* Scalar boolean value should be transformed into |
1187 | all zeros or all ones value before building a vector. */ |
1188 | if (VECTOR_BOOLEAN_TYPE_P (type)) |
1189 | { |
1190 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1191 | tree false_val = build_zero_cst (TREE_TYPE (type)); |
1192 | |
1193 | if (CONSTANT_CLASS_P (val)) |
1194 | val = integer_zerop (val) ? false_val : true_val; |
1195 | else |
1196 | { |
1197 | new_temp = make_ssa_name (TREE_TYPE (type)); |
1198 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, |
1199 | val, true_val, false_val); |
1200 | vect_init_vector_1 (vinfo, stmt_vinfo: stmt_info, new_stmt: init_stmt, gsi); |
1201 | val = new_temp; |
1202 | } |
1203 | } |
1204 | else |
1205 | { |
1206 | gimple_seq stmts = NULL; |
1207 | if (! INTEGRAL_TYPE_P (TREE_TYPE (val))) |
1208 | val = gimple_build (seq: &stmts, code: VIEW_CONVERT_EXPR, |
1209 | TREE_TYPE (type), ops: val); |
1210 | else |
1211 | /* ??? Condition vectorization expects us to do |
1212 | promotion of invariant/external defs. */ |
1213 | val = gimple_convert (seq: &stmts, TREE_TYPE (type), op: val); |
1214 | for (gimple_stmt_iterator gsi2 = gsi_start (seq&: stmts); |
1215 | !gsi_end_p (i: gsi2); ) |
1216 | { |
1217 | init_stmt = gsi_stmt (i: gsi2); |
1218 | gsi_remove (&gsi2, false); |
1219 | vect_init_vector_1 (vinfo, stmt_vinfo: stmt_info, new_stmt: init_stmt, gsi); |
1220 | } |
1221 | } |
1222 | } |
1223 | val = build_vector_from_val (type, val); |
1224 | } |
1225 | |
1226 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_" ); |
1227 | init_stmt = gimple_build_assign (new_temp, val); |
1228 | vect_init_vector_1 (vinfo, stmt_vinfo: stmt_info, new_stmt: init_stmt, gsi); |
1229 | return new_temp; |
1230 | } |
1231 | |
1232 | |
1233 | /* Function vect_get_vec_defs_for_operand. |
1234 | |
1235 | OP is an operand in STMT_VINFO. This function returns a vector of |
1236 | NCOPIES defs that will be used in the vectorized stmts for STMT_VINFO. |
1237 | |
1238 | In the case that OP is an SSA_NAME which is defined in the loop, then |
1239 | STMT_VINFO_VEC_STMTS of the defining stmt holds the relevant defs. |
1240 | |
1241 | In case OP is an invariant or constant, a new stmt that creates a vector def |
1242 | needs to be introduced. VECTYPE may be used to specify a required type for |
1243 | vector invariant. */ |
1244 | |
1245 | void |
1246 | vect_get_vec_defs_for_operand (vec_info *vinfo, stmt_vec_info stmt_vinfo, |
1247 | unsigned ncopies, |
1248 | tree op, vec<tree> *vec_oprnds, tree vectype) |
1249 | { |
1250 | gimple *def_stmt; |
1251 | enum vect_def_type dt; |
1252 | bool is_simple_use; |
1253 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
1254 | |
1255 | if (dump_enabled_p ()) |
1256 | dump_printf_loc (MSG_NOTE, vect_location, |
1257 | "vect_get_vec_defs_for_operand: %T\n" , op); |
1258 | |
1259 | stmt_vec_info def_stmt_info; |
1260 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &dt, |
1261 | &def_stmt_info, &def_stmt); |
1262 | gcc_assert (is_simple_use); |
1263 | if (def_stmt && dump_enabled_p ()) |
1264 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = %G" , def_stmt); |
1265 | |
1266 | vec_oprnds->create (nelems: ncopies); |
1267 | if (dt == vect_constant_def || dt == vect_external_def) |
1268 | { |
1269 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); |
1270 | tree vector_type; |
1271 | |
1272 | if (vectype) |
1273 | vector_type = vectype; |
1274 | else if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op)) |
1275 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) |
1276 | vector_type = truth_type_for (stmt_vectype); |
1277 | else |
1278 | vector_type = get_vectype_for_scalar_type (loop_vinfo, TREE_TYPE (op)); |
1279 | |
1280 | gcc_assert (vector_type); |
1281 | tree vop = vect_init_vector (vinfo, stmt_info: stmt_vinfo, val: op, type: vector_type, NULL); |
1282 | while (ncopies--) |
1283 | vec_oprnds->quick_push (obj: vop); |
1284 | } |
1285 | else |
1286 | { |
1287 | def_stmt_info = vect_stmt_to_vectorize (stmt_info: def_stmt_info); |
1288 | gcc_assert (STMT_VINFO_VEC_STMTS (def_stmt_info).length () == ncopies); |
1289 | for (unsigned i = 0; i < ncopies; ++i) |
1290 | vec_oprnds->quick_push (obj: gimple_get_lhs |
1291 | (STMT_VINFO_VEC_STMTS (def_stmt_info)[i])); |
1292 | } |
1293 | } |
1294 | |
1295 | |
1296 | /* Get vectorized definitions for OP0 and OP1. */ |
1297 | |
1298 | void |
1299 | vect_get_vec_defs (vec_info *vinfo, stmt_vec_info stmt_info, slp_tree slp_node, |
1300 | unsigned ncopies, |
1301 | tree op0, tree vectype0, vec<tree> *vec_oprnds0, |
1302 | tree op1, tree vectype1, vec<tree> *vec_oprnds1, |
1303 | tree op2, tree vectype2, vec<tree> *vec_oprnds2, |
1304 | tree op3, tree vectype3, vec<tree> *vec_oprnds3) |
1305 | { |
1306 | if (slp_node) |
1307 | { |
1308 | if (op0) |
1309 | vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[0], vec_oprnds0); |
1310 | if (op1) |
1311 | vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[1], vec_oprnds1); |
1312 | if (op2) |
1313 | vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[2], vec_oprnds2); |
1314 | if (op3) |
1315 | vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[3], vec_oprnds3); |
1316 | } |
1317 | else |
1318 | { |
1319 | if (op0) |
1320 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies, |
1321 | op: op0, vec_oprnds: vec_oprnds0, vectype: vectype0); |
1322 | if (op1) |
1323 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies, |
1324 | op: op1, vec_oprnds: vec_oprnds1, vectype: vectype1); |
1325 | if (op2) |
1326 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies, |
1327 | op: op2, vec_oprnds: vec_oprnds2, vectype: vectype2); |
1328 | if (op3) |
1329 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies, |
1330 | op: op3, vec_oprnds: vec_oprnds3, vectype: vectype3); |
1331 | } |
1332 | } |
1333 | |
1334 | void |
1335 | vect_get_vec_defs (vec_info *vinfo, stmt_vec_info stmt_info, slp_tree slp_node, |
1336 | unsigned ncopies, |
1337 | tree op0, vec<tree> *vec_oprnds0, |
1338 | tree op1, vec<tree> *vec_oprnds1, |
1339 | tree op2, vec<tree> *vec_oprnds2, |
1340 | tree op3, vec<tree> *vec_oprnds3) |
1341 | { |
1342 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, |
1343 | op0, NULL_TREE, vec_oprnds0, |
1344 | op1, NULL_TREE, vec_oprnds1, |
1345 | op2, NULL_TREE, vec_oprnds2, |
1346 | op3, NULL_TREE, vec_oprnds3); |
1347 | } |
1348 | |
1349 | /* Helper function called by vect_finish_replace_stmt and |
1350 | vect_finish_stmt_generation. Set the location of the new |
1351 | statement and create and return a stmt_vec_info for it. */ |
1352 | |
1353 | static void |
1354 | vect_finish_stmt_generation_1 (vec_info *, |
1355 | stmt_vec_info stmt_info, gimple *vec_stmt) |
1356 | { |
1357 | if (dump_enabled_p ()) |
1358 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: %G" , vec_stmt); |
1359 | |
1360 | if (stmt_info) |
1361 | { |
1362 | gimple_set_location (g: vec_stmt, location: gimple_location (g: stmt_info->stmt)); |
1363 | |
1364 | /* While EH edges will generally prevent vectorization, stmt might |
1365 | e.g. be in a must-not-throw region. Ensure newly created stmts |
1366 | that could throw are part of the same region. */ |
1367 | int lp_nr = lookup_stmt_eh_lp (stmt_info->stmt); |
1368 | if (lp_nr != 0 && stmt_could_throw_p (cfun, vec_stmt)) |
1369 | add_stmt_to_eh_lp (vec_stmt, lp_nr); |
1370 | } |
1371 | else |
1372 | gcc_assert (!stmt_could_throw_p (cfun, vec_stmt)); |
1373 | } |
1374 | |
1375 | /* Replace the scalar statement STMT_INFO with a new vector statement VEC_STMT, |
1376 | which sets the same scalar result as STMT_INFO did. Create and return a |
1377 | stmt_vec_info for VEC_STMT. */ |
1378 | |
1379 | void |
1380 | vect_finish_replace_stmt (vec_info *vinfo, |
1381 | stmt_vec_info stmt_info, gimple *vec_stmt) |
1382 | { |
1383 | gimple *scalar_stmt = vect_orig_stmt (stmt_info)->stmt; |
1384 | gcc_assert (gimple_get_lhs (scalar_stmt) == gimple_get_lhs (vec_stmt)); |
1385 | |
1386 | gimple_stmt_iterator gsi = gsi_for_stmt (scalar_stmt); |
1387 | gsi_replace (&gsi, vec_stmt, true); |
1388 | |
1389 | vect_finish_stmt_generation_1 (vinfo, stmt_info, vec_stmt); |
1390 | } |
1391 | |
1392 | /* Add VEC_STMT to the vectorized implementation of STMT_INFO and insert it |
1393 | before *GSI. Create and return a stmt_vec_info for VEC_STMT. */ |
1394 | |
1395 | void |
1396 | vect_finish_stmt_generation (vec_info *vinfo, |
1397 | stmt_vec_info stmt_info, gimple *vec_stmt, |
1398 | gimple_stmt_iterator *gsi) |
1399 | { |
1400 | gcc_assert (!stmt_info || gimple_code (stmt_info->stmt) != GIMPLE_LABEL); |
1401 | |
1402 | if (!gsi_end_p (i: *gsi) |
1403 | && gimple_has_mem_ops (g: vec_stmt)) |
1404 | { |
1405 | gimple *at_stmt = gsi_stmt (i: *gsi); |
1406 | tree vuse = gimple_vuse (g: at_stmt); |
1407 | if (vuse && TREE_CODE (vuse) == SSA_NAME) |
1408 | { |
1409 | tree vdef = gimple_vdef (g: at_stmt); |
1410 | gimple_set_vuse (g: vec_stmt, vuse: gimple_vuse (g: at_stmt)); |
1411 | gimple_set_modified (s: vec_stmt, modifiedp: true); |
1412 | /* If we have an SSA vuse and insert a store, update virtual |
1413 | SSA form to avoid triggering the renamer. Do so only |
1414 | if we can easily see all uses - which is what almost always |
1415 | happens with the way vectorized stmts are inserted. */ |
1416 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) |
1417 | && ((is_gimple_assign (gs: vec_stmt) |
1418 | && !is_gimple_reg (gimple_assign_lhs (gs: vec_stmt))) |
1419 | || (is_gimple_call (gs: vec_stmt) |
1420 | && (!(gimple_call_flags (vec_stmt) |
1421 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS)) |
1422 | || (gimple_call_lhs (gs: vec_stmt) |
1423 | && !is_gimple_reg (gimple_call_lhs (gs: vec_stmt))))))) |
1424 | { |
1425 | tree new_vdef = copy_ssa_name (var: vuse, stmt: vec_stmt); |
1426 | gimple_set_vdef (g: vec_stmt, vdef: new_vdef); |
1427 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); |
1428 | } |
1429 | } |
1430 | } |
1431 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1432 | vect_finish_stmt_generation_1 (vinfo, stmt_info, vec_stmt); |
1433 | } |
1434 | |
1435 | /* We want to vectorize a call to combined function CFN with function |
1436 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN |
1437 | as the types of all inputs. Check whether this is possible using |
1438 | an internal function, returning its code if so or IFN_LAST if not. */ |
1439 | |
1440 | static internal_fn |
1441 | vectorizable_internal_function (combined_fn cfn, tree fndecl, |
1442 | tree vectype_out, tree vectype_in) |
1443 | { |
1444 | internal_fn ifn; |
1445 | if (internal_fn_p (code: cfn)) |
1446 | ifn = as_internal_fn (code: cfn); |
1447 | else |
1448 | ifn = associated_internal_fn (fndecl); |
1449 | if (ifn != IFN_LAST && direct_internal_fn_p (fn: ifn)) |
1450 | { |
1451 | const direct_internal_fn_info &info = direct_internal_fn (fn: ifn); |
1452 | if (info.vectorizable) |
1453 | { |
1454 | bool same_size_p = TYPE_SIZE (vectype_in) == TYPE_SIZE (vectype_out); |
1455 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); |
1456 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); |
1457 | |
1458 | /* The type size of both the vectype_in and vectype_out should be |
1459 | exactly the same when vectype_out isn't participating the optab. |
1460 | While there is no restriction for type size when vectype_out |
1461 | is part of the optab query. */ |
1462 | if (type0 != vectype_out && type1 != vectype_out && !same_size_p) |
1463 | return IFN_LAST; |
1464 | |
1465 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1466 | OPTIMIZE_FOR_SPEED)) |
1467 | return ifn; |
1468 | } |
1469 | } |
1470 | return IFN_LAST; |
1471 | } |
1472 | |
1473 | |
1474 | static tree permute_vec_elements (vec_info *, tree, tree, tree, stmt_vec_info, |
1475 | gimple_stmt_iterator *); |
1476 | |
1477 | /* Check whether a load or store statement in the loop described by |
1478 | LOOP_VINFO is possible in a loop using partial vectors. This is |
1479 | testing whether the vectorizer pass has the appropriate support, |
1480 | as well as whether the target does. |
1481 | |
1482 | VLS_TYPE says whether the statement is a load or store and VECTYPE |
1483 | is the type of the vector being loaded or stored. SLP_NODE is the SLP |
1484 | node that contains the statement, or null if none. MEMORY_ACCESS_TYPE |
1485 | says how the load or store is going to be implemented and GROUP_SIZE |
1486 | is the number of load or store statements in the containing group. |
1487 | If the access is a gather load or scatter store, GS_INFO describes |
1488 | its arguments. If the load or store is conditional, SCALAR_MASK is the |
1489 | condition under which it occurs. |
1490 | |
1491 | Clear LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P if a loop using partial |
1492 | vectors is not supported, otherwise record the required rgroup control |
1493 | types. */ |
1494 | |
1495 | static void |
1496 | check_load_store_for_partial_vectors (loop_vec_info loop_vinfo, tree vectype, |
1497 | slp_tree slp_node, |
1498 | vec_load_store_type vls_type, |
1499 | int group_size, |
1500 | vect_memory_access_type |
1501 | memory_access_type, |
1502 | gather_scatter_info *gs_info, |
1503 | tree scalar_mask) |
1504 | { |
1505 | /* Invariant loads need no special support. */ |
1506 | if (memory_access_type == VMAT_INVARIANT) |
1507 | return; |
1508 | |
1509 | unsigned int nvectors; |
1510 | if (slp_node) |
1511 | nvectors = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
1512 | else |
1513 | nvectors = vect_get_num_copies (loop_vinfo, vectype); |
1514 | |
1515 | vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); |
1516 | vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo); |
1517 | machine_mode vecmode = TYPE_MODE (vectype); |
1518 | bool is_load = (vls_type == VLS_LOAD); |
1519 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
1520 | { |
1521 | internal_fn ifn |
1522 | = (is_load ? vect_load_lanes_supported (vectype, group_size, true) |
1523 | : vect_store_lanes_supported (vectype, group_size, true)); |
1524 | if (ifn == IFN_MASK_LEN_LOAD_LANES || ifn == IFN_MASK_LEN_STORE_LANES) |
1525 | vect_record_loop_len (loop_vinfo, lens, nvectors, vectype, 1); |
1526 | else if (ifn == IFN_MASK_LOAD_LANES || ifn == IFN_MASK_STORE_LANES) |
1527 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype, |
1528 | scalar_mask); |
1529 | else |
1530 | { |
1531 | if (dump_enabled_p ()) |
1532 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1533 | "can't operate on partial vectors because" |
1534 | " the target doesn't have an appropriate" |
1535 | " load/store-lanes instruction.\n" ); |
1536 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
1537 | } |
1538 | return; |
1539 | } |
1540 | |
1541 | if (memory_access_type == VMAT_GATHER_SCATTER) |
1542 | { |
1543 | internal_fn ifn = (is_load |
1544 | ? IFN_MASK_GATHER_LOAD |
1545 | : IFN_MASK_SCATTER_STORE); |
1546 | internal_fn len_ifn = (is_load |
1547 | ? IFN_MASK_LEN_GATHER_LOAD |
1548 | : IFN_MASK_LEN_SCATTER_STORE); |
1549 | if (internal_gather_scatter_fn_supported_p (len_ifn, vectype, |
1550 | gs_info->memory_type, |
1551 | gs_info->offset_vectype, |
1552 | gs_info->scale)) |
1553 | vect_record_loop_len (loop_vinfo, lens, nvectors, vectype, 1); |
1554 | else if (internal_gather_scatter_fn_supported_p (ifn, vectype, |
1555 | gs_info->memory_type, |
1556 | gs_info->offset_vectype, |
1557 | gs_info->scale)) |
1558 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype, |
1559 | scalar_mask); |
1560 | else |
1561 | { |
1562 | if (dump_enabled_p ()) |
1563 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1564 | "can't operate on partial vectors because" |
1565 | " the target doesn't have an appropriate" |
1566 | " gather load or scatter store instruction.\n" ); |
1567 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
1568 | } |
1569 | return; |
1570 | } |
1571 | |
1572 | if (memory_access_type != VMAT_CONTIGUOUS |
1573 | && memory_access_type != VMAT_CONTIGUOUS_PERMUTE) |
1574 | { |
1575 | /* Element X of the data must come from iteration i * VF + X of the |
1576 | scalar loop. We need more work to support other mappings. */ |
1577 | if (dump_enabled_p ()) |
1578 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1579 | "can't operate on partial vectors because an" |
1580 | " access isn't contiguous.\n" ); |
1581 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
1582 | return; |
1583 | } |
1584 | |
1585 | if (!VECTOR_MODE_P (vecmode)) |
1586 | { |
1587 | if (dump_enabled_p ()) |
1588 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1589 | "can't operate on partial vectors when emulating" |
1590 | " vector operations.\n" ); |
1591 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
1592 | return; |
1593 | } |
1594 | |
1595 | /* We might load more scalars than we need for permuting SLP loads. |
1596 | We checked in get_group_load_store_type that the extra elements |
1597 | don't leak into a new vector. */ |
1598 | auto group_memory_nvectors = [](poly_uint64 size, poly_uint64 nunits) |
1599 | { |
1600 | unsigned int nvectors; |
1601 | if (can_div_away_from_zero_p (a: size, b: nunits, quotient: &nvectors)) |
1602 | return nvectors; |
1603 | gcc_unreachable (); |
1604 | }; |
1605 | |
1606 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
1607 | poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); |
1608 | machine_mode mask_mode; |
1609 | machine_mode vmode; |
1610 | bool using_partial_vectors_p = false; |
1611 | if (get_len_load_store_mode (vecmode, is_load).exists (mode: &vmode)) |
1612 | { |
1613 | nvectors = group_memory_nvectors (group_size * vf, nunits); |
1614 | unsigned factor = (vecmode == vmode) ? 1 : GET_MODE_UNIT_SIZE (vecmode); |
1615 | vect_record_loop_len (loop_vinfo, lens, nvectors, vectype, factor); |
1616 | using_partial_vectors_p = true; |
1617 | } |
1618 | else if (targetm.vectorize.get_mask_mode (vecmode).exists (mode: &mask_mode) |
1619 | && can_vec_mask_load_store_p (vecmode, mask_mode, is_load)) |
1620 | { |
1621 | nvectors = group_memory_nvectors (group_size * vf, nunits); |
1622 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype, scalar_mask); |
1623 | using_partial_vectors_p = true; |
1624 | } |
1625 | |
1626 | if (!using_partial_vectors_p) |
1627 | { |
1628 | if (dump_enabled_p ()) |
1629 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1630 | "can't operate on partial vectors because the" |
1631 | " target doesn't have the appropriate partial" |
1632 | " vectorization load or store.\n" ); |
1633 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
1634 | } |
1635 | } |
1636 | |
1637 | /* Return the mask input to a masked load or store. VEC_MASK is the vectorized |
1638 | form of the scalar mask condition and LOOP_MASK, if nonnull, is the mask |
1639 | that needs to be applied to all loads and stores in a vectorized loop. |
1640 | Return VEC_MASK if LOOP_MASK is null or if VEC_MASK is already masked, |
1641 | otherwise return VEC_MASK & LOOP_MASK. |
1642 | |
1643 | MASK_TYPE is the type of both masks. If new statements are needed, |
1644 | insert them before GSI. */ |
1645 | |
1646 | static tree |
1647 | prepare_vec_mask (loop_vec_info loop_vinfo, tree mask_type, tree loop_mask, |
1648 | tree vec_mask, gimple_stmt_iterator *gsi) |
1649 | { |
1650 | gcc_assert (useless_type_conversion_p (mask_type, TREE_TYPE (vec_mask))); |
1651 | if (!loop_mask) |
1652 | return vec_mask; |
1653 | |
1654 | gcc_assert (TREE_TYPE (loop_mask) == mask_type); |
1655 | |
1656 | if (loop_vinfo->vec_cond_masked_set.contains (k: { vec_mask, loop_mask })) |
1657 | return vec_mask; |
1658 | |
1659 | tree and_res = make_temp_ssa_name (type: mask_type, NULL, name: "vec_mask_and" ); |
1660 | gimple *and_stmt = gimple_build_assign (and_res, BIT_AND_EXPR, |
1661 | vec_mask, loop_mask); |
1662 | |
1663 | gsi_insert_before (gsi, and_stmt, GSI_SAME_STMT); |
1664 | return and_res; |
1665 | } |
1666 | |
1667 | /* Determine whether we can use a gather load or scatter store to vectorize |
1668 | strided load or store STMT_INFO by truncating the current offset to a |
1669 | smaller width. We need to be able to construct an offset vector: |
1670 | |
1671 | { 0, X, X*2, X*3, ... } |
1672 | |
1673 | without loss of precision, where X is STMT_INFO's DR_STEP. |
1674 | |
1675 | Return true if this is possible, describing the gather load or scatter |
1676 | store in GS_INFO. MASKED_P is true if the load or store is conditional. */ |
1677 | |
1678 | static bool |
1679 | vect_truncate_gather_scatter_offset (stmt_vec_info stmt_info, |
1680 | loop_vec_info loop_vinfo, bool masked_p, |
1681 | gather_scatter_info *gs_info) |
1682 | { |
1683 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info); |
1684 | data_reference *dr = dr_info->dr; |
1685 | tree step = DR_STEP (dr); |
1686 | if (TREE_CODE (step) != INTEGER_CST) |
1687 | { |
1688 | /* ??? Perhaps we could use range information here? */ |
1689 | if (dump_enabled_p ()) |
1690 | dump_printf_loc (MSG_NOTE, vect_location, |
1691 | "cannot truncate variable step.\n" ); |
1692 | return false; |
1693 | } |
1694 | |
1695 | /* Get the number of bits in an element. */ |
1696 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
1697 | scalar_mode element_mode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); |
1698 | unsigned int element_bits = GET_MODE_BITSIZE (mode: element_mode); |
1699 | |
1700 | /* Set COUNT to the upper limit on the number of elements - 1. |
1701 | Start with the maximum vectorization factor. */ |
1702 | unsigned HOST_WIDE_INT count = vect_max_vf (loop_vinfo) - 1; |
1703 | |
1704 | /* Try lowering COUNT to the number of scalar latch iterations. */ |
1705 | class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
1706 | widest_int max_iters; |
1707 | if (max_loop_iterations (loop, &max_iters) |
1708 | && max_iters < count) |
1709 | count = max_iters.to_shwi (); |
1710 | |
1711 | /* Try scales of 1 and the element size. */ |
1712 | int scales[] = { 1, vect_get_scalar_dr_size (dr_info) }; |
1713 | wi::overflow_type overflow = wi::OVF_NONE; |
1714 | for (int i = 0; i < 2; ++i) |
1715 | { |
1716 | int scale = scales[i]; |
1717 | widest_int factor; |
1718 | if (!wi::multiple_of_p (x: wi::to_widest (t: step), y: scale, sgn: SIGNED, res: &factor)) |
1719 | continue; |
1720 | |
1721 | /* Determine the minimum precision of (COUNT - 1) * STEP / SCALE. */ |
1722 | widest_int range = wi::mul (x: count, y: factor, sgn: SIGNED, overflow: &overflow); |
1723 | if (overflow) |
1724 | continue; |
1725 | signop sign = range >= 0 ? UNSIGNED : SIGNED; |
1726 | unsigned int min_offset_bits = wi::min_precision (x: range, sgn: sign); |
1727 | |
1728 | /* Find the narrowest viable offset type. */ |
1729 | unsigned int offset_bits = 1U << ceil_log2 (x: min_offset_bits); |
1730 | tree offset_type = build_nonstandard_integer_type (offset_bits, |
1731 | sign == UNSIGNED); |
1732 | |
1733 | /* See whether the target supports the operation with an offset |
1734 | no narrower than OFFSET_TYPE. */ |
1735 | tree memory_type = TREE_TYPE (DR_REF (dr)); |
1736 | if (!vect_gather_scatter_fn_p (loop_vinfo, DR_IS_READ (dr), masked_p, |
1737 | vectype, memory_type, offset_type, scale, |
1738 | &gs_info->ifn, &gs_info->offset_vectype) |
1739 | || gs_info->ifn == IFN_LAST) |
1740 | continue; |
1741 | |
1742 | gs_info->decl = NULL_TREE; |
1743 | /* Logically the sum of DR_BASE_ADDRESS, DR_INIT and DR_OFFSET, |
1744 | but we don't need to store that here. */ |
1745 | gs_info->base = NULL_TREE; |
1746 | gs_info->element_type = TREE_TYPE (vectype); |
1747 | gs_info->offset = fold_convert (offset_type, step); |
1748 | gs_info->offset_dt = vect_constant_def; |
1749 | gs_info->scale = scale; |
1750 | gs_info->memory_type = memory_type; |
1751 | return true; |
1752 | } |
1753 | |
1754 | if (overflow && dump_enabled_p ()) |
1755 | dump_printf_loc (MSG_NOTE, vect_location, |
1756 | "truncating gather/scatter offset to %d bits" |
1757 | " might change its value.\n" , element_bits); |
1758 | |
1759 | return false; |
1760 | } |
1761 | |
1762 | /* Return true if we can use gather/scatter internal functions to |
1763 | vectorize STMT_INFO, which is a grouped or strided load or store. |
1764 | MASKED_P is true if load or store is conditional. When returning |
1765 | true, fill in GS_INFO with the information required to perform the |
1766 | operation. */ |
1767 | |
1768 | static bool |
1769 | vect_use_strided_gather_scatters_p (stmt_vec_info stmt_info, |
1770 | loop_vec_info loop_vinfo, bool masked_p, |
1771 | gather_scatter_info *gs_info) |
1772 | { |
1773 | if (!vect_check_gather_scatter (stmt_info, loop_vinfo, gs_info) |
1774 | || gs_info->ifn == IFN_LAST) |
1775 | return vect_truncate_gather_scatter_offset (stmt_info, loop_vinfo, |
1776 | masked_p, gs_info); |
1777 | |
1778 | tree old_offset_type = TREE_TYPE (gs_info->offset); |
1779 | tree new_offset_type = TREE_TYPE (gs_info->offset_vectype); |
1780 | |
1781 | gcc_assert (TYPE_PRECISION (new_offset_type) |
1782 | >= TYPE_PRECISION (old_offset_type)); |
1783 | gs_info->offset = fold_convert (new_offset_type, gs_info->offset); |
1784 | |
1785 | if (dump_enabled_p ()) |
1786 | dump_printf_loc (MSG_NOTE, vect_location, |
1787 | "using gather/scatter for strided/grouped access," |
1788 | " scale = %d\n" , gs_info->scale); |
1789 | |
1790 | return true; |
1791 | } |
1792 | |
1793 | /* STMT_INFO is a non-strided load or store, meaning that it accesses |
1794 | elements with a known constant step. Return -1 if that step |
1795 | is negative, 0 if it is zero, and 1 if it is greater than zero. */ |
1796 | |
1797 | static int |
1798 | compare_step_with_zero (vec_info *vinfo, stmt_vec_info stmt_info) |
1799 | { |
1800 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info); |
1801 | return tree_int_cst_compare (t1: vect_dr_behavior (vinfo, dr_info)->step, |
1802 | size_zero_node); |
1803 | } |
1804 | |
1805 | /* If the target supports a permute mask that reverses the elements in |
1806 | a vector of type VECTYPE, return that mask, otherwise return null. */ |
1807 | |
1808 | tree |
1809 | perm_mask_for_reverse (tree vectype) |
1810 | { |
1811 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
1812 | |
1813 | /* The encoding has a single stepped pattern. */ |
1814 | vec_perm_builder sel (nunits, 1, 3); |
1815 | for (int i = 0; i < 3; ++i) |
1816 | sel.quick_push (obj: nunits - 1 - i); |
1817 | |
1818 | vec_perm_indices indices (sel, 1, nunits); |
1819 | if (!can_vec_perm_const_p (TYPE_MODE (vectype), TYPE_MODE (vectype), |
1820 | indices)) |
1821 | return NULL_TREE; |
1822 | return vect_gen_perm_mask_checked (vectype, indices); |
1823 | } |
1824 | |
1825 | /* A subroutine of get_load_store_type, with a subset of the same |
1826 | arguments. Handle the case where STMT_INFO is a load or store that |
1827 | accesses consecutive elements with a negative step. Sets *POFFSET |
1828 | to the offset to be applied to the DR for the first access. */ |
1829 | |
1830 | static vect_memory_access_type |
1831 | get_negative_load_store_type (vec_info *vinfo, |
1832 | stmt_vec_info stmt_info, tree vectype, |
1833 | vec_load_store_type vls_type, |
1834 | unsigned int ncopies, poly_int64 *poffset) |
1835 | { |
1836 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info); |
1837 | dr_alignment_support alignment_support_scheme; |
1838 | |
1839 | if (ncopies > 1) |
1840 | { |
1841 | if (dump_enabled_p ()) |
1842 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1843 | "multiple types with negative step.\n" ); |
1844 | return VMAT_ELEMENTWISE; |
1845 | } |
1846 | |
1847 | /* For backward running DRs the first access in vectype actually is |
1848 | N-1 elements before the address of the DR. */ |
1849 | *poffset = ((-TYPE_VECTOR_SUBPARTS (node: vectype) + 1) |
1850 | * TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (vectype)))); |
1851 | |
1852 | int misalignment = dr_misalignment (dr_info, vectype, offset: *poffset); |
1853 | alignment_support_scheme |
1854 | = vect_supportable_dr_alignment (vinfo, dr_info, vectype, misalignment); |
1855 | if (alignment_support_scheme != dr_aligned |
1856 | && alignment_support_scheme != dr_unaligned_supported) |
1857 | { |
1858 | if (dump_enabled_p ()) |
1859 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1860 | "negative step but alignment required.\n" ); |
1861 | *poffset = 0; |
1862 | return VMAT_ELEMENTWISE; |
1863 | } |
1864 | |
1865 | if (vls_type == VLS_STORE_INVARIANT) |
1866 | { |
1867 | if (dump_enabled_p ()) |
1868 | dump_printf_loc (MSG_NOTE, vect_location, |
1869 | "negative step with invariant source;" |
1870 | " no permute needed.\n" ); |
1871 | return VMAT_CONTIGUOUS_DOWN; |
1872 | } |
1873 | |
1874 | if (!perm_mask_for_reverse (vectype)) |
1875 | { |
1876 | if (dump_enabled_p ()) |
1877 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1878 | "negative step and reversing not supported.\n" ); |
1879 | *poffset = 0; |
1880 | return VMAT_ELEMENTWISE; |
1881 | } |
1882 | |
1883 | return VMAT_CONTIGUOUS_REVERSE; |
1884 | } |
1885 | |
1886 | /* STMT_INFO is either a masked or unconditional store. Return the value |
1887 | being stored. */ |
1888 | |
1889 | tree |
1890 | vect_get_store_rhs (stmt_vec_info stmt_info) |
1891 | { |
1892 | if (gassign *assign = dyn_cast <gassign *> (p: stmt_info->stmt)) |
1893 | { |
1894 | gcc_assert (gimple_assign_single_p (assign)); |
1895 | return gimple_assign_rhs1 (gs: assign); |
1896 | } |
1897 | if (gcall *call = dyn_cast <gcall *> (p: stmt_info->stmt)) |
1898 | { |
1899 | internal_fn ifn = gimple_call_internal_fn (gs: call); |
1900 | int index = internal_fn_stored_value_index (ifn); |
1901 | gcc_assert (index >= 0); |
1902 | return gimple_call_arg (gs: call, index); |
1903 | } |
1904 | gcc_unreachable (); |
1905 | } |
1906 | |
1907 | /* Function VECTOR_VECTOR_COMPOSITION_TYPE |
1908 | |
1909 | This function returns a vector type which can be composed with NETLS pieces, |
1910 | whose type is recorded in PTYPE. VTYPE should be a vector type, and has the |
1911 | same vector size as the return vector. It checks target whether supports |
1912 | pieces-size vector mode for construction firstly, if target fails to, check |
1913 | pieces-size scalar mode for construction further. It returns NULL_TREE if |
1914 | fails to find the available composition. |
1915 | |
1916 | For example, for (vtype=V16QI, nelts=4), we can probably get: |
1917 | - V16QI with PTYPE V4QI. |
1918 | - V4SI with PTYPE SI. |
1919 | - NULL_TREE. */ |
1920 | |
1921 | static tree |
1922 | vector_vector_composition_type (tree vtype, poly_uint64 nelts, tree *ptype) |
1923 | { |
1924 | gcc_assert (VECTOR_TYPE_P (vtype)); |
1925 | gcc_assert (known_gt (nelts, 0U)); |
1926 | |
1927 | machine_mode vmode = TYPE_MODE (vtype); |
1928 | if (!VECTOR_MODE_P (vmode)) |
1929 | return NULL_TREE; |
1930 | |
1931 | /* When we are asked to compose the vector from its components let |
1932 | that happen directly. */ |
1933 | if (known_eq (TYPE_VECTOR_SUBPARTS (vtype), nelts)) |
1934 | { |
1935 | *ptype = TREE_TYPE (vtype); |
1936 | return vtype; |
1937 | } |
1938 | |
1939 | poly_uint64 vbsize = GET_MODE_BITSIZE (mode: vmode); |
1940 | unsigned int pbsize; |
1941 | if (constant_multiple_p (a: vbsize, b: nelts, multiple: &pbsize)) |
1942 | { |
1943 | /* First check if vec_init optab supports construction from |
1944 | vector pieces directly. */ |
1945 | scalar_mode elmode = SCALAR_TYPE_MODE (TREE_TYPE (vtype)); |
1946 | poly_uint64 inelts = pbsize / GET_MODE_BITSIZE (mode: elmode); |
1947 | machine_mode rmode; |
1948 | if (related_vector_mode (vmode, elmode, inelts).exists (mode: &rmode) |
1949 | && (convert_optab_handler (op: vec_init_optab, to_mode: vmode, from_mode: rmode) |
1950 | != CODE_FOR_nothing)) |
1951 | { |
1952 | *ptype = build_vector_type (TREE_TYPE (vtype), inelts); |
1953 | return vtype; |
1954 | } |
1955 | |
1956 | /* Otherwise check if exists an integer type of the same piece size and |
1957 | if vec_init optab supports construction from it directly. */ |
1958 | if (int_mode_for_size (size: pbsize, limit: 0).exists (mode: &elmode) |
1959 | && related_vector_mode (vmode, elmode, nelts).exists (mode: &rmode) |
1960 | && (convert_optab_handler (op: vec_init_optab, to_mode: rmode, from_mode: elmode) |
1961 | != CODE_FOR_nothing)) |
1962 | { |
1963 | *ptype = build_nonstandard_integer_type (pbsize, 1); |
1964 | return build_vector_type (*ptype, nelts); |
1965 | } |
1966 | } |
1967 | |
1968 | return NULL_TREE; |
1969 | } |
1970 | |
1971 | /* A subroutine of get_load_store_type, with a subset of the same |
1972 | arguments. Handle the case where STMT_INFO is part of a grouped load |
1973 | or store. |
1974 | |
1975 | For stores, the statements in the group are all consecutive |
1976 | and there is no gap at the end. For loads, the statements in the |
1977 | group might not be consecutive; there can be gaps between statements |
1978 | as well as at the end. */ |
1979 | |
1980 | static bool |
1981 | get_group_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info, |
1982 | tree vectype, slp_tree slp_node, |
1983 | bool masked_p, vec_load_store_type vls_type, |
1984 | vect_memory_access_type *memory_access_type, |
1985 | poly_int64 *poffset, |
1986 | dr_alignment_support *alignment_support_scheme, |
1987 | int *misalignment, |
1988 | gather_scatter_info *gs_info, |
1989 | internal_fn *lanes_ifn) |
1990 | { |
1991 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
1992 | class loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
1993 | stmt_vec_info first_stmt_info; |
1994 | unsigned int group_size; |
1995 | unsigned HOST_WIDE_INT gap; |
1996 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
1997 | { |
1998 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
1999 | group_size = DR_GROUP_SIZE (first_stmt_info); |
2000 | gap = DR_GROUP_GAP (first_stmt_info); |
2001 | } |
2002 | else |
2003 | { |
2004 | first_stmt_info = stmt_info; |
2005 | group_size = 1; |
2006 | gap = 0; |
2007 | } |
2008 | dr_vec_info *first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info); |
2009 | bool single_element_p = (stmt_info == first_stmt_info |
2010 | && !DR_GROUP_NEXT_ELEMENT (stmt_info)); |
2011 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
2012 | |
2013 | /* True if the vectorized statements would access beyond the last |
2014 | statement in the group. */ |
2015 | bool overrun_p = false; |
2016 | |
2017 | /* True if we can cope with such overrun by peeling for gaps, so that |
2018 | there is at least one final scalar iteration after the vector loop. */ |
2019 | bool can_overrun_p = (!masked_p |
2020 | && vls_type == VLS_LOAD |
2021 | && loop_vinfo |
2022 | && !loop->inner); |
2023 | |
2024 | /* There can only be a gap at the end of the group if the stride is |
2025 | known at compile time. */ |
2026 | gcc_assert (!STMT_VINFO_STRIDED_P (first_stmt_info) || gap == 0); |
2027 | |
2028 | /* Stores can't yet have gaps. */ |
2029 | gcc_assert (slp_node || vls_type == VLS_LOAD || gap == 0); |
2030 | |
2031 | if (slp_node) |
2032 | { |
2033 | /* For SLP vectorization we directly vectorize a subchain |
2034 | without permutation. */ |
2035 | if (! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
2036 | first_dr_info |
2037 | = STMT_VINFO_DR_INFO (SLP_TREE_SCALAR_STMTS (slp_node)[0]); |
2038 | if (STMT_VINFO_STRIDED_P (first_stmt_info)) |
2039 | { |
2040 | /* Try to use consecutive accesses of DR_GROUP_SIZE elements, |
2041 | separated by the stride, until we have a complete vector. |
2042 | Fall back to scalar accesses if that isn't possible. */ |
2043 | if (multiple_p (a: nunits, b: group_size)) |
2044 | *memory_access_type = VMAT_STRIDED_SLP; |
2045 | else |
2046 | *memory_access_type = VMAT_ELEMENTWISE; |
2047 | } |
2048 | else |
2049 | { |
2050 | overrun_p = loop_vinfo && gap != 0; |
2051 | if (overrun_p && vls_type != VLS_LOAD) |
2052 | { |
2053 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2054 | "Grouped store with gaps requires" |
2055 | " non-consecutive accesses\n" ); |
2056 | return false; |
2057 | } |
2058 | /* An overrun is fine if the trailing elements are smaller |
2059 | than the alignment boundary B. Every vector access will |
2060 | be a multiple of B and so we are guaranteed to access a |
2061 | non-gap element in the same B-sized block. */ |
2062 | if (overrun_p |
2063 | && gap < (vect_known_alignment_in_bytes (dr_info: first_dr_info, |
2064 | vectype) |
2065 | / vect_get_scalar_dr_size (dr_info: first_dr_info))) |
2066 | overrun_p = false; |
2067 | |
2068 | /* If the gap splits the vector in half and the target |
2069 | can do half-vector operations avoid the epilogue peeling |
2070 | by simply loading half of the vector only. Usually |
2071 | the construction with an upper zero half will be elided. */ |
2072 | dr_alignment_support alss; |
2073 | int misalign = dr_misalignment (dr_info: first_dr_info, vectype); |
2074 | tree half_vtype; |
2075 | if (overrun_p |
2076 | && !masked_p |
2077 | && (((alss = vect_supportable_dr_alignment (vinfo, first_dr_info, |
2078 | vectype, misalign))) |
2079 | == dr_aligned |
2080 | || alss == dr_unaligned_supported) |
2081 | && known_eq (nunits, (group_size - gap) * 2) |
2082 | && known_eq (nunits, group_size) |
2083 | && (vector_vector_composition_type (vtype: vectype, nelts: 2, ptype: &half_vtype) |
2084 | != NULL_TREE)) |
2085 | overrun_p = false; |
2086 | |
2087 | if (overrun_p && !can_overrun_p) |
2088 | { |
2089 | if (dump_enabled_p ()) |
2090 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2091 | "Peeling for outer loop is not supported\n" ); |
2092 | return false; |
2093 | } |
2094 | int cmp = compare_step_with_zero (vinfo, stmt_info); |
2095 | if (cmp < 0) |
2096 | { |
2097 | if (single_element_p) |
2098 | /* ??? The VMAT_CONTIGUOUS_REVERSE code generation is |
2099 | only correct for single element "interleaving" SLP. */ |
2100 | *memory_access_type = get_negative_load_store_type |
2101 | (vinfo, stmt_info, vectype, vls_type, ncopies: 1, poffset); |
2102 | else |
2103 | { |
2104 | /* Try to use consecutive accesses of DR_GROUP_SIZE elements, |
2105 | separated by the stride, until we have a complete vector. |
2106 | Fall back to scalar accesses if that isn't possible. */ |
2107 | if (multiple_p (a: nunits, b: group_size)) |
2108 | *memory_access_type = VMAT_STRIDED_SLP; |
2109 | else |
2110 | *memory_access_type = VMAT_ELEMENTWISE; |
2111 | } |
2112 | } |
2113 | else if (cmp == 0 && loop_vinfo) |
2114 | { |
2115 | gcc_assert (vls_type == VLS_LOAD); |
2116 | *memory_access_type = VMAT_INVARIANT; |
2117 | /* Invariant accesses perform only component accesses, alignment |
2118 | is irrelevant for them. */ |
2119 | *alignment_support_scheme = dr_unaligned_supported; |
2120 | } |
2121 | else |
2122 | *memory_access_type = VMAT_CONTIGUOUS; |
2123 | |
2124 | /* When we have a contiguous access across loop iterations |
2125 | but the access in the loop doesn't cover the full vector |
2126 | we can end up with no gap recorded but still excess |
2127 | elements accessed, see PR103116. Make sure we peel for |
2128 | gaps if necessary and sufficient and give up if not. |
2129 | |
2130 | If there is a combination of the access not covering the full |
2131 | vector and a gap recorded then we may need to peel twice. */ |
2132 | if (loop_vinfo |
2133 | && *memory_access_type == VMAT_CONTIGUOUS |
2134 | && SLP_TREE_LOAD_PERMUTATION (slp_node).exists () |
2135 | && !multiple_p (a: group_size * LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
2136 | b: nunits)) |
2137 | { |
2138 | unsigned HOST_WIDE_INT cnunits, cvf; |
2139 | if (!can_overrun_p |
2140 | || !nunits.is_constant (const_value: &cnunits) |
2141 | || !LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (const_value: &cvf) |
2142 | /* Peeling for gaps assumes that a single scalar iteration |
2143 | is enough to make sure the last vector iteration doesn't |
2144 | access excess elements. |
2145 | ??? Enhancements include peeling multiple iterations |
2146 | or using masked loads with a static mask. */ |
2147 | || (group_size * cvf) % cnunits + group_size - gap < cnunits) |
2148 | { |
2149 | if (dump_enabled_p ()) |
2150 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2151 | "peeling for gaps insufficient for " |
2152 | "access\n" ); |
2153 | return false; |
2154 | } |
2155 | overrun_p = true; |
2156 | } |
2157 | } |
2158 | } |
2159 | else |
2160 | { |
2161 | /* We can always handle this case using elementwise accesses, |
2162 | but see if something more efficient is available. */ |
2163 | *memory_access_type = VMAT_ELEMENTWISE; |
2164 | |
2165 | /* If there is a gap at the end of the group then these optimizations |
2166 | would access excess elements in the last iteration. */ |
2167 | bool would_overrun_p = (gap != 0); |
2168 | /* An overrun is fine if the trailing elements are smaller than the |
2169 | alignment boundary B. Every vector access will be a multiple of B |
2170 | and so we are guaranteed to access a non-gap element in the |
2171 | same B-sized block. */ |
2172 | if (would_overrun_p |
2173 | && !masked_p |
2174 | && gap < (vect_known_alignment_in_bytes (dr_info: first_dr_info, vectype) |
2175 | / vect_get_scalar_dr_size (dr_info: first_dr_info))) |
2176 | would_overrun_p = false; |
2177 | |
2178 | if (!STMT_VINFO_STRIDED_P (first_stmt_info) |
2179 | && (can_overrun_p || !would_overrun_p) |
2180 | && compare_step_with_zero (vinfo, stmt_info) > 0) |
2181 | { |
2182 | /* First cope with the degenerate case of a single-element |
2183 | vector. */ |
2184 | if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U)) |
2185 | ; |
2186 | |
2187 | else |
2188 | { |
2189 | /* Otherwise try using LOAD/STORE_LANES. */ |
2190 | *lanes_ifn |
2191 | = vls_type == VLS_LOAD |
2192 | ? vect_load_lanes_supported (vectype, group_size, masked_p) |
2193 | : vect_store_lanes_supported (vectype, group_size, |
2194 | masked_p); |
2195 | if (*lanes_ifn != IFN_LAST) |
2196 | { |
2197 | *memory_access_type = VMAT_LOAD_STORE_LANES; |
2198 | overrun_p = would_overrun_p; |
2199 | } |
2200 | |
2201 | /* If that fails, try using permuting loads. */ |
2202 | else if (vls_type == VLS_LOAD |
2203 | ? vect_grouped_load_supported (vectype, |
2204 | single_element_p, |
2205 | group_size) |
2206 | : vect_grouped_store_supported (vectype, group_size)) |
2207 | { |
2208 | *memory_access_type = VMAT_CONTIGUOUS_PERMUTE; |
2209 | overrun_p = would_overrun_p; |
2210 | } |
2211 | } |
2212 | } |
2213 | |
2214 | /* As a last resort, trying using a gather load or scatter store. |
2215 | |
2216 | ??? Although the code can handle all group sizes correctly, |
2217 | it probably isn't a win to use separate strided accesses based |
2218 | on nearby locations. Or, even if it's a win over scalar code, |
2219 | it might not be a win over vectorizing at a lower VF, if that |
2220 | allows us to use contiguous accesses. */ |
2221 | if (*memory_access_type == VMAT_ELEMENTWISE |
2222 | && single_element_p |
2223 | && loop_vinfo |
2224 | && vect_use_strided_gather_scatters_p (stmt_info, loop_vinfo, |
2225 | masked_p, gs_info)) |
2226 | *memory_access_type = VMAT_GATHER_SCATTER; |
2227 | } |
2228 | |
2229 | if (*memory_access_type == VMAT_GATHER_SCATTER |
2230 | || *memory_access_type == VMAT_ELEMENTWISE) |
2231 | { |
2232 | *alignment_support_scheme = dr_unaligned_supported; |
2233 | *misalignment = DR_MISALIGNMENT_UNKNOWN; |
2234 | } |
2235 | else |
2236 | { |
2237 | *misalignment = dr_misalignment (dr_info: first_dr_info, vectype, offset: *poffset); |
2238 | *alignment_support_scheme |
2239 | = vect_supportable_dr_alignment (vinfo, first_dr_info, vectype, |
2240 | *misalignment); |
2241 | } |
2242 | |
2243 | if (vls_type != VLS_LOAD && first_stmt_info == stmt_info) |
2244 | { |
2245 | /* STMT is the leader of the group. Check the operands of all the |
2246 | stmts of the group. */ |
2247 | stmt_vec_info next_stmt_info = DR_GROUP_NEXT_ELEMENT (stmt_info); |
2248 | while (next_stmt_info) |
2249 | { |
2250 | tree op = vect_get_store_rhs (stmt_info: next_stmt_info); |
2251 | enum vect_def_type dt; |
2252 | if (!vect_is_simple_use (op, vinfo, &dt)) |
2253 | { |
2254 | if (dump_enabled_p ()) |
2255 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2256 | "use not simple.\n" ); |
2257 | return false; |
2258 | } |
2259 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
2260 | } |
2261 | } |
2262 | |
2263 | if (overrun_p) |
2264 | { |
2265 | gcc_assert (can_overrun_p); |
2266 | if (dump_enabled_p ()) |
2267 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2268 | "Data access with gaps requires scalar " |
2269 | "epilogue loop\n" ); |
2270 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; |
2271 | } |
2272 | |
2273 | return true; |
2274 | } |
2275 | |
2276 | /* Analyze load or store statement STMT_INFO of type VLS_TYPE. Return true |
2277 | if there is a memory access type that the vectorized form can use, |
2278 | storing it in *MEMORY_ACCESS_TYPE if so. If we decide to use gathers |
2279 | or scatters, fill in GS_INFO accordingly. In addition |
2280 | *ALIGNMENT_SUPPORT_SCHEME is filled out and false is returned if |
2281 | the target does not support the alignment scheme. *MISALIGNMENT |
2282 | is set according to the alignment of the access (including |
2283 | DR_MISALIGNMENT_UNKNOWN when it is unknown). |
2284 | |
2285 | SLP says whether we're performing SLP rather than loop vectorization. |
2286 | MASKED_P is true if the statement is conditional on a vectorized mask. |
2287 | VECTYPE is the vector type that the vectorized statements will use. |
2288 | NCOPIES is the number of vector statements that will be needed. */ |
2289 | |
2290 | static bool |
2291 | get_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info, |
2292 | tree vectype, slp_tree slp_node, |
2293 | bool masked_p, vec_load_store_type vls_type, |
2294 | unsigned int ncopies, |
2295 | vect_memory_access_type *memory_access_type, |
2296 | poly_int64 *poffset, |
2297 | dr_alignment_support *alignment_support_scheme, |
2298 | int *misalignment, |
2299 | gather_scatter_info *gs_info, |
2300 | internal_fn *lanes_ifn) |
2301 | { |
2302 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
2303 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
2304 | *misalignment = DR_MISALIGNMENT_UNKNOWN; |
2305 | *poffset = 0; |
2306 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
2307 | { |
2308 | *memory_access_type = VMAT_GATHER_SCATTER; |
2309 | if (!vect_check_gather_scatter (stmt_info, loop_vinfo, gs_info)) |
2310 | gcc_unreachable (); |
2311 | /* When using internal functions, we rely on pattern recognition |
2312 | to convert the type of the offset to the type that the target |
2313 | requires, with the result being a call to an internal function. |
2314 | If that failed for some reason (e.g. because another pattern |
2315 | took priority), just handle cases in which the offset already |
2316 | has the right type. */ |
2317 | else if (gs_info->ifn != IFN_LAST |
2318 | && !is_gimple_call (gs: stmt_info->stmt) |
2319 | && !tree_nop_conversion_p (TREE_TYPE (gs_info->offset), |
2320 | TREE_TYPE (gs_info->offset_vectype))) |
2321 | { |
2322 | if (dump_enabled_p ()) |
2323 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2324 | "%s offset requires a conversion\n" , |
2325 | vls_type == VLS_LOAD ? "gather" : "scatter" ); |
2326 | return false; |
2327 | } |
2328 | else if (!vect_is_simple_use (gs_info->offset, vinfo, |
2329 | &gs_info->offset_dt, |
2330 | &gs_info->offset_vectype)) |
2331 | { |
2332 | if (dump_enabled_p ()) |
2333 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2334 | "%s index use not simple.\n" , |
2335 | vls_type == VLS_LOAD ? "gather" : "scatter" ); |
2336 | return false; |
2337 | } |
2338 | else if (gs_info->ifn == IFN_LAST && !gs_info->decl) |
2339 | { |
2340 | if (!TYPE_VECTOR_SUBPARTS (node: vectype).is_constant () |
2341 | || !TYPE_VECTOR_SUBPARTS (node: gs_info->offset_vectype).is_constant () |
2342 | || !constant_multiple_p (a: TYPE_VECTOR_SUBPARTS |
2343 | (node: gs_info->offset_vectype), |
2344 | b: TYPE_VECTOR_SUBPARTS (node: vectype))) |
2345 | { |
2346 | if (dump_enabled_p ()) |
2347 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2348 | "unsupported vector types for emulated " |
2349 | "gather.\n" ); |
2350 | return false; |
2351 | } |
2352 | } |
2353 | /* Gather-scatter accesses perform only component accesses, alignment |
2354 | is irrelevant for them. */ |
2355 | *alignment_support_scheme = dr_unaligned_supported; |
2356 | } |
2357 | else if (STMT_VINFO_GROUPED_ACCESS (stmt_info) || slp_node) |
2358 | { |
2359 | if (!get_group_load_store_type (vinfo, stmt_info, vectype, slp_node, |
2360 | masked_p, |
2361 | vls_type, memory_access_type, poffset, |
2362 | alignment_support_scheme, |
2363 | misalignment, gs_info, lanes_ifn)) |
2364 | return false; |
2365 | } |
2366 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
2367 | { |
2368 | gcc_assert (!slp_node); |
2369 | if (loop_vinfo |
2370 | && vect_use_strided_gather_scatters_p (stmt_info, loop_vinfo, |
2371 | masked_p, gs_info)) |
2372 | *memory_access_type = VMAT_GATHER_SCATTER; |
2373 | else |
2374 | *memory_access_type = VMAT_ELEMENTWISE; |
2375 | /* Alignment is irrelevant here. */ |
2376 | *alignment_support_scheme = dr_unaligned_supported; |
2377 | } |
2378 | else |
2379 | { |
2380 | int cmp = compare_step_with_zero (vinfo, stmt_info); |
2381 | if (cmp == 0) |
2382 | { |
2383 | gcc_assert (vls_type == VLS_LOAD); |
2384 | *memory_access_type = VMAT_INVARIANT; |
2385 | /* Invariant accesses perform only component accesses, alignment |
2386 | is irrelevant for them. */ |
2387 | *alignment_support_scheme = dr_unaligned_supported; |
2388 | } |
2389 | else |
2390 | { |
2391 | if (cmp < 0) |
2392 | *memory_access_type = get_negative_load_store_type |
2393 | (vinfo, stmt_info, vectype, vls_type, ncopies, poffset); |
2394 | else |
2395 | *memory_access_type = VMAT_CONTIGUOUS; |
2396 | *misalignment = dr_misalignment (STMT_VINFO_DR_INFO (stmt_info), |
2397 | vectype, offset: *poffset); |
2398 | *alignment_support_scheme |
2399 | = vect_supportable_dr_alignment (vinfo, |
2400 | STMT_VINFO_DR_INFO (stmt_info), |
2401 | vectype, *misalignment); |
2402 | } |
2403 | } |
2404 | |
2405 | if ((*memory_access_type == VMAT_ELEMENTWISE |
2406 | || *memory_access_type == VMAT_STRIDED_SLP) |
2407 | && !nunits.is_constant ()) |
2408 | { |
2409 | if (dump_enabled_p ()) |
2410 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2411 | "Not using elementwise accesses due to variable " |
2412 | "vectorization factor.\n" ); |
2413 | return false; |
2414 | } |
2415 | |
2416 | if (*alignment_support_scheme == dr_unaligned_unsupported) |
2417 | { |
2418 | if (dump_enabled_p ()) |
2419 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2420 | "unsupported unaligned access\n" ); |
2421 | return false; |
2422 | } |
2423 | |
2424 | /* FIXME: At the moment the cost model seems to underestimate the |
2425 | cost of using elementwise accesses. This check preserves the |
2426 | traditional behavior until that can be fixed. */ |
2427 | stmt_vec_info first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
2428 | if (!first_stmt_info) |
2429 | first_stmt_info = stmt_info; |
2430 | if (*memory_access_type == VMAT_ELEMENTWISE |
2431 | && !STMT_VINFO_STRIDED_P (first_stmt_info) |
2432 | && !(stmt_info == DR_GROUP_FIRST_ELEMENT (stmt_info) |
2433 | && !DR_GROUP_NEXT_ELEMENT (stmt_info) |
2434 | && !pow2p_hwi (DR_GROUP_SIZE (stmt_info)))) |
2435 | { |
2436 | if (dump_enabled_p ()) |
2437 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2438 | "not falling back to elementwise accesses\n" ); |
2439 | return false; |
2440 | } |
2441 | return true; |
2442 | } |
2443 | |
2444 | /* Return true if boolean argument at MASK_INDEX is suitable for vectorizing |
2445 | conditional operation STMT_INFO. When returning true, store the mask |
2446 | in *MASK, the type of its definition in *MASK_DT_OUT, the type of the |
2447 | vectorized mask in *MASK_VECTYPE_OUT and the SLP node corresponding |
2448 | to the mask in *MASK_NODE if MASK_NODE is not NULL. */ |
2449 | |
2450 | static bool |
2451 | vect_check_scalar_mask (vec_info *vinfo, stmt_vec_info stmt_info, |
2452 | slp_tree slp_node, unsigned mask_index, |
2453 | tree *mask, slp_tree *mask_node, |
2454 | vect_def_type *mask_dt_out, tree *mask_vectype_out) |
2455 | { |
2456 | enum vect_def_type mask_dt; |
2457 | tree mask_vectype; |
2458 | slp_tree mask_node_1; |
2459 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, mask_index, |
2460 | mask, &mask_node_1, &mask_dt, &mask_vectype)) |
2461 | { |
2462 | if (dump_enabled_p ()) |
2463 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2464 | "mask use not simple.\n" ); |
2465 | return false; |
2466 | } |
2467 | |
2468 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (*mask))) |
2469 | { |
2470 | if (dump_enabled_p ()) |
2471 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2472 | "mask argument is not a boolean.\n" ); |
2473 | return false; |
2474 | } |
2475 | |
2476 | /* If the caller is not prepared for adjusting an external/constant |
2477 | SLP mask vector type fail. */ |
2478 | if (slp_node |
2479 | && !mask_node |
2480 | && SLP_TREE_DEF_TYPE (mask_node_1) != vect_internal_def) |
2481 | { |
2482 | if (dump_enabled_p ()) |
2483 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2484 | "SLP mask argument is not vectorized.\n" ); |
2485 | return false; |
2486 | } |
2487 | |
2488 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2489 | if (!mask_vectype) |
2490 | mask_vectype = get_mask_type_for_scalar_type (vinfo, TREE_TYPE (vectype), |
2491 | mask_node_1); |
2492 | |
2493 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype)) |
2494 | { |
2495 | if (dump_enabled_p ()) |
2496 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2497 | "could not find an appropriate vector mask type.\n" ); |
2498 | return false; |
2499 | } |
2500 | |
2501 | if (maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: mask_vectype), |
2502 | b: TYPE_VECTOR_SUBPARTS (node: vectype))) |
2503 | { |
2504 | if (dump_enabled_p ()) |
2505 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2506 | "vector mask type %T" |
2507 | " does not match vector data type %T.\n" , |
2508 | mask_vectype, vectype); |
2509 | |
2510 | return false; |
2511 | } |
2512 | |
2513 | *mask_dt_out = mask_dt; |
2514 | *mask_vectype_out = mask_vectype; |
2515 | if (mask_node) |
2516 | *mask_node = mask_node_1; |
2517 | return true; |
2518 | } |
2519 | |
2520 | /* Return true if stored value is suitable for vectorizing store |
2521 | statement STMT_INFO. When returning true, store the scalar stored |
2522 | in *RHS and *RHS_NODE, the type of the definition in *RHS_DT_OUT, |
2523 | the type of the vectorized store value in |
2524 | *RHS_VECTYPE_OUT and the type of the store in *VLS_TYPE_OUT. */ |
2525 | |
2526 | static bool |
2527 | vect_check_store_rhs (vec_info *vinfo, stmt_vec_info stmt_info, |
2528 | slp_tree slp_node, tree *rhs, slp_tree *rhs_node, |
2529 | vect_def_type *rhs_dt_out, tree *rhs_vectype_out, |
2530 | vec_load_store_type *vls_type_out) |
2531 | { |
2532 | int op_no = 0; |
2533 | if (gcall *call = dyn_cast <gcall *> (p: stmt_info->stmt)) |
2534 | { |
2535 | if (gimple_call_internal_p (gs: call) |
2536 | && internal_store_fn_p (gimple_call_internal_fn (gs: call))) |
2537 | op_no = internal_fn_stored_value_index (gimple_call_internal_fn (gs: call)); |
2538 | } |
2539 | if (slp_node) |
2540 | op_no = vect_slp_child_index_for_operand |
2541 | (stmt_info->stmt, op: op_no, STMT_VINFO_GATHER_SCATTER_P (stmt_info)); |
2542 | |
2543 | enum vect_def_type rhs_dt; |
2544 | tree rhs_vectype; |
2545 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, op_no, |
2546 | rhs, rhs_node, &rhs_dt, &rhs_vectype)) |
2547 | { |
2548 | if (dump_enabled_p ()) |
2549 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2550 | "use not simple.\n" ); |
2551 | return false; |
2552 | } |
2553 | |
2554 | /* In the case this is a store from a constant make sure |
2555 | native_encode_expr can handle it. */ |
2556 | if (CONSTANT_CLASS_P (*rhs) && native_encode_expr (*rhs, NULL, 64) == 0) |
2557 | { |
2558 | if (dump_enabled_p ()) |
2559 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2560 | "cannot encode constant as a byte sequence.\n" ); |
2561 | return false; |
2562 | } |
2563 | |
2564 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2565 | if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype)) |
2566 | { |
2567 | if (dump_enabled_p ()) |
2568 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2569 | "incompatible vector types.\n" ); |
2570 | return false; |
2571 | } |
2572 | |
2573 | *rhs_dt_out = rhs_dt; |
2574 | *rhs_vectype_out = rhs_vectype; |
2575 | if (rhs_dt == vect_constant_def || rhs_dt == vect_external_def) |
2576 | *vls_type_out = VLS_STORE_INVARIANT; |
2577 | else |
2578 | *vls_type_out = VLS_STORE; |
2579 | return true; |
2580 | } |
2581 | |
2582 | /* Build an all-ones vector mask of type MASKTYPE while vectorizing STMT_INFO. |
2583 | Note that we support masks with floating-point type, in which case the |
2584 | floats are interpreted as a bitmask. */ |
2585 | |
2586 | static tree |
2587 | vect_build_all_ones_mask (vec_info *vinfo, |
2588 | stmt_vec_info stmt_info, tree masktype) |
2589 | { |
2590 | if (TREE_CODE (masktype) == INTEGER_TYPE) |
2591 | return build_int_cst (masktype, -1); |
2592 | else if (VECTOR_BOOLEAN_TYPE_P (masktype) |
2593 | || TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) |
2594 | { |
2595 | tree mask = build_int_cst (TREE_TYPE (masktype), -1); |
2596 | mask = build_vector_from_val (masktype, mask); |
2597 | return vect_init_vector (vinfo, stmt_info, val: mask, type: masktype, NULL); |
2598 | } |
2599 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) |
2600 | { |
2601 | REAL_VALUE_TYPE r; |
2602 | long tmp[6]; |
2603 | for (int j = 0; j < 6; ++j) |
2604 | tmp[j] = -1; |
2605 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); |
2606 | tree mask = build_real (TREE_TYPE (masktype), r); |
2607 | mask = build_vector_from_val (masktype, mask); |
2608 | return vect_init_vector (vinfo, stmt_info, val: mask, type: masktype, NULL); |
2609 | } |
2610 | gcc_unreachable (); |
2611 | } |
2612 | |
2613 | /* Build an all-zero merge value of type VECTYPE while vectorizing |
2614 | STMT_INFO as a gather load. */ |
2615 | |
2616 | static tree |
2617 | vect_build_zero_merge_argument (vec_info *vinfo, |
2618 | stmt_vec_info stmt_info, tree vectype) |
2619 | { |
2620 | tree merge; |
2621 | if (TREE_CODE (TREE_TYPE (vectype)) == INTEGER_TYPE) |
2622 | merge = build_int_cst (TREE_TYPE (vectype), 0); |
2623 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (vectype))) |
2624 | { |
2625 | REAL_VALUE_TYPE r; |
2626 | long tmp[6]; |
2627 | for (int j = 0; j < 6; ++j) |
2628 | tmp[j] = 0; |
2629 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (vectype))); |
2630 | merge = build_real (TREE_TYPE (vectype), r); |
2631 | } |
2632 | else |
2633 | gcc_unreachable (); |
2634 | merge = build_vector_from_val (vectype, merge); |
2635 | return vect_init_vector (vinfo, stmt_info, val: merge, type: vectype, NULL); |
2636 | } |
2637 | |
2638 | /* Build a gather load call while vectorizing STMT_INFO. Insert new |
2639 | instructions before GSI and add them to VEC_STMT. GS_INFO describes |
2640 | the gather load operation. If the load is conditional, MASK is the |
2641 | vectorized condition, otherwise MASK is null. PTR is the base |
2642 | pointer and OFFSET is the vectorized offset. */ |
2643 | |
2644 | static gimple * |
2645 | vect_build_one_gather_load_call (vec_info *vinfo, stmt_vec_info stmt_info, |
2646 | gimple_stmt_iterator *gsi, |
2647 | gather_scatter_info *gs_info, |
2648 | tree ptr, tree offset, tree mask) |
2649 | { |
2650 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2651 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info->decl)); |
2652 | tree rettype = TREE_TYPE (TREE_TYPE (gs_info->decl)); |
2653 | tree srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
2654 | /* ptrtype */ arglist = TREE_CHAIN (arglist); |
2655 | tree idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
2656 | tree masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
2657 | tree scaletype = TREE_VALUE (arglist); |
2658 | tree var; |
2659 | gcc_checking_assert (types_compatible_p (srctype, rettype) |
2660 | && (!mask |
2661 | || TREE_CODE (masktype) == INTEGER_TYPE |
2662 | || types_compatible_p (srctype, masktype))); |
2663 | |
2664 | tree op = offset; |
2665 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) |
2666 | { |
2667 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), |
2668 | TYPE_VECTOR_SUBPARTS (idxtype))); |
2669 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
2670 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
2671 | gassign *new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
2672 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
2673 | op = var; |
2674 | } |
2675 | |
2676 | tree src_op = NULL_TREE; |
2677 | tree mask_op = NULL_TREE; |
2678 | if (mask) |
2679 | { |
2680 | if (!useless_type_conversion_p (masktype, TREE_TYPE (mask))) |
2681 | { |
2682 | tree utype, optype = TREE_TYPE (mask); |
2683 | if (VECTOR_TYPE_P (masktype) |
2684 | || TYPE_MODE (masktype) == TYPE_MODE (optype)) |
2685 | utype = masktype; |
2686 | else |
2687 | utype = lang_hooks.types.type_for_mode (TYPE_MODE (optype), 1); |
2688 | var = vect_get_new_ssa_name (utype, vect_scalar_var); |
2689 | tree mask_arg = build1 (VIEW_CONVERT_EXPR, utype, mask); |
2690 | gassign *new_stmt |
2691 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_arg); |
2692 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
2693 | mask_arg = var; |
2694 | if (!useless_type_conversion_p (masktype, utype)) |
2695 | { |
2696 | gcc_assert (TYPE_PRECISION (utype) |
2697 | <= TYPE_PRECISION (masktype)); |
2698 | var = vect_get_new_ssa_name (masktype, vect_scalar_var); |
2699 | new_stmt = gimple_build_assign (var, NOP_EXPR, mask_arg); |
2700 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
2701 | mask_arg = var; |
2702 | } |
2703 | src_op = build_zero_cst (srctype); |
2704 | mask_op = mask_arg; |
2705 | } |
2706 | else |
2707 | { |
2708 | src_op = mask; |
2709 | mask_op = mask; |
2710 | } |
2711 | } |
2712 | else |
2713 | { |
2714 | src_op = vect_build_zero_merge_argument (vinfo, stmt_info, vectype: rettype); |
2715 | mask_op = vect_build_all_ones_mask (vinfo, stmt_info, masktype); |
2716 | } |
2717 | |
2718 | tree scale = build_int_cst (scaletype, gs_info->scale); |
2719 | gimple *new_stmt = gimple_build_call (gs_info->decl, 5, src_op, ptr, op, |
2720 | mask_op, scale); |
2721 | |
2722 | if (!useless_type_conversion_p (vectype, rettype)) |
2723 | { |
2724 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (vectype), |
2725 | TYPE_VECTOR_SUBPARTS (rettype))); |
2726 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
2727 | gimple_call_set_lhs (gs: new_stmt, lhs: op); |
2728 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
2729 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
2730 | new_stmt = gimple_build_assign (NULL_TREE, VIEW_CONVERT_EXPR, op); |
2731 | } |
2732 | |
2733 | return new_stmt; |
2734 | } |
2735 | |
2736 | /* Build a scatter store call while vectorizing STMT_INFO. Insert new |
2737 | instructions before GSI. GS_INFO describes the scatter store operation. |
2738 | PTR is the base pointer, OFFSET the vectorized offsets and OPRND the |
2739 | vectorized data to store. |
2740 | If the store is conditional, MASK is the vectorized condition, otherwise |
2741 | MASK is null. */ |
2742 | |
2743 | static gimple * |
2744 | vect_build_one_scatter_store_call (vec_info *vinfo, stmt_vec_info stmt_info, |
2745 | gimple_stmt_iterator *gsi, |
2746 | gather_scatter_info *gs_info, |
2747 | tree ptr, tree offset, tree oprnd, tree mask) |
2748 | { |
2749 | tree rettype = TREE_TYPE (TREE_TYPE (gs_info->decl)); |
2750 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info->decl)); |
2751 | /* tree ptrtype = TREE_VALUE (arglist); */ arglist = TREE_CHAIN (arglist); |
2752 | tree masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
2753 | tree idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
2754 | tree srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
2755 | tree scaletype = TREE_VALUE (arglist); |
2756 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE |
2757 | && TREE_CODE (rettype) == VOID_TYPE); |
2758 | |
2759 | tree mask_arg = NULL_TREE; |
2760 | if (mask) |
2761 | { |
2762 | mask_arg = mask; |
2763 | tree optype = TREE_TYPE (mask_arg); |
2764 | tree utype; |
2765 | if (TYPE_MODE (masktype) == TYPE_MODE (optype)) |
2766 | utype = masktype; |
2767 | else |
2768 | utype = lang_hooks.types.type_for_mode (TYPE_MODE (optype), 1); |
2769 | tree var = vect_get_new_ssa_name (utype, vect_scalar_var); |
2770 | mask_arg = build1 (VIEW_CONVERT_EXPR, utype, mask_arg); |
2771 | gassign *new_stmt |
2772 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_arg); |
2773 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
2774 | mask_arg = var; |
2775 | if (!useless_type_conversion_p (masktype, utype)) |
2776 | { |
2777 | gcc_assert (TYPE_PRECISION (utype) <= TYPE_PRECISION (masktype)); |
2778 | tree var = vect_get_new_ssa_name (masktype, vect_scalar_var); |
2779 | new_stmt = gimple_build_assign (var, NOP_EXPR, mask_arg); |
2780 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
2781 | mask_arg = var; |
2782 | } |
2783 | } |
2784 | else |
2785 | { |
2786 | mask_arg = build_int_cst (masktype, -1); |
2787 | mask_arg = vect_init_vector (vinfo, stmt_info, val: mask_arg, type: masktype, NULL); |
2788 | } |
2789 | |
2790 | tree src = oprnd; |
2791 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) |
2792 | { |
2793 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)), |
2794 | TYPE_VECTOR_SUBPARTS (srctype))); |
2795 | tree var = vect_get_new_ssa_name (srctype, vect_simple_var); |
2796 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
2797 | gassign *new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); |
2798 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
2799 | src = var; |
2800 | } |
2801 | |
2802 | tree op = offset; |
2803 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) |
2804 | { |
2805 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), |
2806 | TYPE_VECTOR_SUBPARTS (idxtype))); |
2807 | tree var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
2808 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
2809 | gassign *new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
2810 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
2811 | op = var; |
2812 | } |
2813 | |
2814 | tree scale = build_int_cst (scaletype, gs_info->scale); |
2815 | gcall *new_stmt |
2816 | = gimple_build_call (gs_info->decl, 5, ptr, mask_arg, op, src, scale); |
2817 | return new_stmt; |
2818 | } |
2819 | |
2820 | /* Prepare the base and offset in GS_INFO for vectorization. |
2821 | Set *DATAREF_PTR to the loop-invariant base address and *VEC_OFFSET |
2822 | to the vectorized offset argument for the first copy of STMT_INFO. |
2823 | STMT_INFO is the statement described by GS_INFO and LOOP is the |
2824 | containing loop. */ |
2825 | |
2826 | static void |
2827 | vect_get_gather_scatter_ops (loop_vec_info loop_vinfo, |
2828 | class loop *loop, stmt_vec_info stmt_info, |
2829 | slp_tree slp_node, gather_scatter_info *gs_info, |
2830 | tree *dataref_ptr, vec<tree> *vec_offset) |
2831 | { |
2832 | gimple_seq stmts = NULL; |
2833 | *dataref_ptr = force_gimple_operand (gs_info->base, &stmts, true, NULL_TREE); |
2834 | if (stmts != NULL) |
2835 | { |
2836 | basic_block new_bb; |
2837 | edge pe = loop_preheader_edge (loop); |
2838 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); |
2839 | gcc_assert (!new_bb); |
2840 | } |
2841 | if (slp_node) |
2842 | vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[0], vec_offset); |
2843 | else |
2844 | { |
2845 | unsigned ncopies |
2846 | = vect_get_num_copies (loop_vinfo, vectype: gs_info->offset_vectype); |
2847 | vect_get_vec_defs_for_operand (vinfo: loop_vinfo, stmt_vinfo: stmt_info, ncopies, |
2848 | op: gs_info->offset, vec_oprnds: vec_offset, |
2849 | vectype: gs_info->offset_vectype); |
2850 | } |
2851 | } |
2852 | |
2853 | /* Prepare to implement a grouped or strided load or store using |
2854 | the gather load or scatter store operation described by GS_INFO. |
2855 | STMT_INFO is the load or store statement. |
2856 | |
2857 | Set *DATAREF_BUMP to the amount that should be added to the base |
2858 | address after each copy of the vectorized statement. Set *VEC_OFFSET |
2859 | to an invariant offset vector in which element I has the value |
2860 | I * DR_STEP / SCALE. */ |
2861 | |
2862 | static void |
2863 | vect_get_strided_load_store_ops (stmt_vec_info stmt_info, |
2864 | loop_vec_info loop_vinfo, |
2865 | gimple_stmt_iterator *gsi, |
2866 | gather_scatter_info *gs_info, |
2867 | tree *dataref_bump, tree *vec_offset, |
2868 | vec_loop_lens *loop_lens) |
2869 | { |
2870 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
2871 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2872 | |
2873 | if (LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)) |
2874 | { |
2875 | /* _31 = .SELECT_VL (ivtmp_29, POLY_INT_CST [4, 4]); |
2876 | ivtmp_8 = _31 * 16 (step in bytes); |
2877 | .MASK_LEN_SCATTER_STORE (vectp_a.9_7, ... ); |
2878 | vectp_a.9_26 = vectp_a.9_7 + ivtmp_8; */ |
2879 | tree loop_len |
2880 | = vect_get_loop_len (loop_vinfo, gsi, loop_lens, 1, vectype, 0, 0); |
2881 | tree tmp |
2882 | = fold_build2 (MULT_EXPR, sizetype, |
2883 | fold_convert (sizetype, unshare_expr (DR_STEP (dr))), |
2884 | loop_len); |
2885 | *dataref_bump = force_gimple_operand_gsi (gsi, tmp, true, NULL_TREE, true, |
2886 | GSI_SAME_STMT); |
2887 | } |
2888 | else |
2889 | { |
2890 | tree bump |
2891 | = size_binop (MULT_EXPR, |
2892 | fold_convert (sizetype, unshare_expr (DR_STEP (dr))), |
2893 | size_int (TYPE_VECTOR_SUBPARTS (vectype))); |
2894 | *dataref_bump = cse_and_gimplify_to_preheader (loop_vinfo, bump); |
2895 | } |
2896 | |
2897 | /* The offset given in GS_INFO can have pointer type, so use the element |
2898 | type of the vector instead. */ |
2899 | tree offset_type = TREE_TYPE (gs_info->offset_vectype); |
2900 | |
2901 | /* Calculate X = DR_STEP / SCALE and convert it to the appropriate type. */ |
2902 | tree step = size_binop (EXACT_DIV_EXPR, unshare_expr (DR_STEP (dr)), |
2903 | ssize_int (gs_info->scale)); |
2904 | step = fold_convert (offset_type, step); |
2905 | |
2906 | /* Create {0, X, X*2, X*3, ...}. */ |
2907 | tree offset = fold_build2 (VEC_SERIES_EXPR, gs_info->offset_vectype, |
2908 | build_zero_cst (offset_type), step); |
2909 | *vec_offset = cse_and_gimplify_to_preheader (loop_vinfo, offset); |
2910 | } |
2911 | |
2912 | /* Prepare the pointer IVs which needs to be updated by a variable amount. |
2913 | Such variable amount is the outcome of .SELECT_VL. In this case, we can |
2914 | allow each iteration process the flexible number of elements as long as |
2915 | the number <= vf elments. |
2916 | |
2917 | Return data reference according to SELECT_VL. |
2918 | If new statements are needed, insert them before GSI. */ |
2919 | |
2920 | static tree |
2921 | vect_get_loop_variant_data_ptr_increment ( |
2922 | vec_info *vinfo, tree aggr_type, gimple_stmt_iterator *gsi, |
2923 | vec_loop_lens *loop_lens, dr_vec_info *dr_info, |
2924 | vect_memory_access_type memory_access_type) |
2925 | { |
2926 | loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (p: vinfo); |
2927 | tree step = vect_dr_behavior (vinfo, dr_info)->step; |
2928 | |
2929 | /* gather/scatter never reach here. */ |
2930 | gcc_assert (memory_access_type != VMAT_GATHER_SCATTER); |
2931 | |
2932 | /* When we support SELECT_VL pattern, we dynamic adjust |
2933 | the memory address by .SELECT_VL result. |
2934 | |
2935 | The result of .SELECT_VL is the number of elements to |
2936 | be processed of each iteration. So the memory address |
2937 | adjustment operation should be: |
2938 | |
2939 | addr = addr + .SELECT_VL (ARG..) * step; |
2940 | */ |
2941 | tree loop_len |
2942 | = vect_get_loop_len (loop_vinfo, gsi, loop_lens, 1, aggr_type, 0, 0); |
2943 | tree len_type = TREE_TYPE (loop_len); |
2944 | /* Since the outcome of .SELECT_VL is element size, we should adjust |
2945 | it into bytesize so that it can be used in address pointer variable |
2946 | amount IVs adjustment. */ |
2947 | tree tmp = fold_build2 (MULT_EXPR, len_type, loop_len, |
2948 | wide_int_to_tree (len_type, wi::to_widest (step))); |
2949 | tree bump = make_temp_ssa_name (type: len_type, NULL, name: "ivtmp" ); |
2950 | gassign *assign = gimple_build_assign (bump, tmp); |
2951 | gsi_insert_before (gsi, assign, GSI_SAME_STMT); |
2952 | return bump; |
2953 | } |
2954 | |
2955 | /* Return the amount that should be added to a vector pointer to move |
2956 | to the next or previous copy of AGGR_TYPE. DR_INFO is the data reference |
2957 | being vectorized and MEMORY_ACCESS_TYPE describes the type of |
2958 | vectorization. */ |
2959 | |
2960 | static tree |
2961 | vect_get_data_ptr_increment (vec_info *vinfo, gimple_stmt_iterator *gsi, |
2962 | dr_vec_info *dr_info, tree aggr_type, |
2963 | vect_memory_access_type memory_access_type, |
2964 | vec_loop_lens *loop_lens = nullptr) |
2965 | { |
2966 | if (memory_access_type == VMAT_INVARIANT) |
2967 | return size_zero_node; |
2968 | |
2969 | loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (p: vinfo); |
2970 | if (loop_vinfo && LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)) |
2971 | return vect_get_loop_variant_data_ptr_increment (vinfo, aggr_type, gsi, |
2972 | loop_lens, dr_info, |
2973 | memory_access_type); |
2974 | |
2975 | tree iv_step = TYPE_SIZE_UNIT (aggr_type); |
2976 | tree step = vect_dr_behavior (vinfo, dr_info)->step; |
2977 | if (tree_int_cst_sgn (step) == -1) |
2978 | iv_step = fold_build1 (NEGATE_EXPR, TREE_TYPE (iv_step), iv_step); |
2979 | return iv_step; |
2980 | } |
2981 | |
2982 | /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64,128}. */ |
2983 | |
2984 | static bool |
2985 | vectorizable_bswap (vec_info *vinfo, |
2986 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
2987 | gimple **vec_stmt, slp_tree slp_node, |
2988 | slp_tree *slp_op, |
2989 | tree vectype_in, stmt_vector_for_cost *cost_vec) |
2990 | { |
2991 | tree op, vectype; |
2992 | gcall *stmt = as_a <gcall *> (p: stmt_info->stmt); |
2993 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
2994 | unsigned ncopies; |
2995 | |
2996 | op = gimple_call_arg (gs: stmt, index: 0); |
2997 | vectype = STMT_VINFO_VECTYPE (stmt_info); |
2998 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
2999 | |
3000 | /* Multiple types in SLP are handled by creating the appropriate number of |
3001 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
3002 | case of SLP. */ |
3003 | if (slp_node) |
3004 | ncopies = 1; |
3005 | else |
3006 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
3007 | |
3008 | gcc_assert (ncopies >= 1); |
3009 | |
3010 | if (TYPE_SIZE (vectype_in) != TYPE_SIZE (vectype)) |
3011 | { |
3012 | if (dump_enabled_p ()) |
3013 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3014 | "mismatched vector sizes %T and %T\n" , |
3015 | vectype_in, vectype); |
3016 | return false; |
3017 | } |
3018 | |
3019 | tree char_vectype = get_same_sized_vectype (char_type_node, vectype_in); |
3020 | if (! char_vectype) |
3021 | return false; |
3022 | |
3023 | poly_uint64 num_bytes = TYPE_VECTOR_SUBPARTS (node: char_vectype); |
3024 | unsigned word_bytes; |
3025 | if (!constant_multiple_p (a: num_bytes, b: nunits, multiple: &word_bytes)) |
3026 | return false; |
3027 | |
3028 | /* The encoding uses one stepped pattern for each byte in the word. */ |
3029 | vec_perm_builder elts (num_bytes, word_bytes, 3); |
3030 | for (unsigned i = 0; i < 3; ++i) |
3031 | for (unsigned j = 0; j < word_bytes; ++j) |
3032 | elts.quick_push (obj: (i + 1) * word_bytes - j - 1); |
3033 | |
3034 | vec_perm_indices indices (elts, 1, num_bytes); |
3035 | machine_mode vmode = TYPE_MODE (char_vectype); |
3036 | if (!can_vec_perm_const_p (vmode, vmode, indices)) |
3037 | return false; |
3038 | |
3039 | if (! vec_stmt) |
3040 | { |
3041 | if (slp_node |
3042 | && !vect_maybe_update_slp_op_vectype (slp_op[0], vectype_in)) |
3043 | { |
3044 | if (dump_enabled_p ()) |
3045 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3046 | "incompatible vector types for invariants\n" ); |
3047 | return false; |
3048 | } |
3049 | |
3050 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; |
3051 | DUMP_VECT_SCOPE ("vectorizable_bswap" ); |
3052 | record_stmt_cost (body_cost_vec: cost_vec, |
3053 | count: 1, kind: vector_stmt, stmt_info, misalign: 0, where: vect_prologue); |
3054 | record_stmt_cost (body_cost_vec: cost_vec, |
3055 | count: slp_node |
3056 | ? SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node) : ncopies, |
3057 | kind: vec_perm, stmt_info, misalign: 0, where: vect_body); |
3058 | return true; |
3059 | } |
3060 | |
3061 | tree bswap_vconst = vec_perm_indices_to_tree (char_vectype, indices); |
3062 | |
3063 | /* Transform. */ |
3064 | vec<tree> vec_oprnds = vNULL; |
3065 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, |
3066 | op0: op, vec_oprnds0: &vec_oprnds); |
3067 | /* Arguments are ready. create the new vector stmt. */ |
3068 | unsigned i; |
3069 | tree vop; |
3070 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
3071 | { |
3072 | gimple *new_stmt; |
3073 | tree tem = make_ssa_name (var: char_vectype); |
3074 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, |
3075 | char_vectype, vop)); |
3076 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
3077 | tree tem2 = make_ssa_name (var: char_vectype); |
3078 | new_stmt = gimple_build_assign (tem2, VEC_PERM_EXPR, |
3079 | tem, tem, bswap_vconst); |
3080 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
3081 | tem = make_ssa_name (var: vectype); |
3082 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, |
3083 | vectype, tem2)); |
3084 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
3085 | if (slp_node) |
3086 | slp_node->push_vec_def (def: new_stmt); |
3087 | else |
3088 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
3089 | } |
3090 | |
3091 | if (!slp_node) |
3092 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
3093 | |
3094 | vec_oprnds.release (); |
3095 | return true; |
3096 | } |
3097 | |
3098 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
3099 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT |
3100 | in a single step. On success, store the binary pack code in |
3101 | *CONVERT_CODE. */ |
3102 | |
3103 | static bool |
3104 | simple_integer_narrowing (tree vectype_out, tree vectype_in, |
3105 | code_helper *convert_code) |
3106 | { |
3107 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) |
3108 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) |
3109 | return false; |
3110 | |
3111 | code_helper code; |
3112 | int multi_step_cvt = 0; |
3113 | auto_vec <tree, 8> interm_types; |
3114 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, |
3115 | &code, &multi_step_cvt, &interm_types) |
3116 | || multi_step_cvt) |
3117 | return false; |
3118 | |
3119 | *convert_code = code; |
3120 | return true; |
3121 | } |
3122 | |
3123 | /* Function vectorizable_call. |
3124 | |
3125 | Check if STMT_INFO performs a function call that can be vectorized. |
3126 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
3127 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
3128 | Return true if STMT_INFO is vectorizable in this way. */ |
3129 | |
3130 | static bool |
3131 | vectorizable_call (vec_info *vinfo, |
3132 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
3133 | gimple **vec_stmt, slp_tree slp_node, |
3134 | stmt_vector_for_cost *cost_vec) |
3135 | { |
3136 | gcall *stmt; |
3137 | tree vec_dest; |
3138 | tree scalar_dest; |
3139 | tree op; |
3140 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
3141 | tree vectype_out, vectype_in; |
3142 | poly_uint64 nunits_in; |
3143 | poly_uint64 nunits_out; |
3144 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
3145 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
3146 | tree fndecl, new_temp, rhs_type; |
3147 | enum vect_def_type dt[4] |
3148 | = { vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type, |
3149 | vect_unknown_def_type }; |
3150 | tree vectypes[ARRAY_SIZE (dt)] = {}; |
3151 | slp_tree slp_op[ARRAY_SIZE (dt)] = {}; |
3152 | int ndts = ARRAY_SIZE (dt); |
3153 | int ncopies, j; |
3154 | auto_vec<tree, 8> vargs; |
3155 | enum { NARROW, NONE, WIDEN } modifier; |
3156 | size_t i, nargs; |
3157 | tree lhs; |
3158 | tree clz_ctz_arg1 = NULL_TREE; |
3159 | |
3160 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
3161 | return false; |
3162 | |
3163 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3164 | && ! vec_stmt) |
3165 | return false; |
3166 | |
3167 | /* Is STMT_INFO a vectorizable call? */ |
3168 | stmt = dyn_cast <gcall *> (p: stmt_info->stmt); |
3169 | if (!stmt) |
3170 | return false; |
3171 | |
3172 | if (gimple_call_internal_p (gs: stmt) |
3173 | && (internal_load_fn_p (gimple_call_internal_fn (gs: stmt)) |
3174 | || internal_store_fn_p (gimple_call_internal_fn (gs: stmt)))) |
3175 | /* Handled by vectorizable_load and vectorizable_store. */ |
3176 | return false; |
3177 | |
3178 | if (gimple_call_lhs (gs: stmt) == NULL_TREE |
3179 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) |
3180 | return false; |
3181 | |
3182 | gcc_checking_assert (!stmt_can_throw_internal (cfun, stmt)); |
3183 | |
3184 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3185 | |
3186 | /* Process function arguments. */ |
3187 | rhs_type = NULL_TREE; |
3188 | vectype_in = NULL_TREE; |
3189 | nargs = gimple_call_num_args (gs: stmt); |
3190 | |
3191 | /* Bail out if the function has more than four arguments, we do not have |
3192 | interesting builtin functions to vectorize with more than two arguments |
3193 | except for fma. No arguments is also not good. */ |
3194 | if (nargs == 0 || nargs > 4) |
3195 | return false; |
3196 | |
3197 | /* Ignore the arguments of IFN_GOMP_SIMD_LANE, they are magic. */ |
3198 | combined_fn cfn = gimple_call_combined_fn (stmt); |
3199 | if (cfn == CFN_GOMP_SIMD_LANE) |
3200 | { |
3201 | nargs = 0; |
3202 | rhs_type = unsigned_type_node; |
3203 | } |
3204 | /* Similarly pretend IFN_CLZ and IFN_CTZ only has one argument, the second |
3205 | argument just says whether it is well-defined at zero or not and what |
3206 | value should be returned for it. */ |
3207 | if ((cfn == CFN_CLZ || cfn == CFN_CTZ) && nargs == 2) |
3208 | { |
3209 | nargs = 1; |
3210 | clz_ctz_arg1 = gimple_call_arg (gs: stmt, index: 1); |
3211 | } |
3212 | |
3213 | int mask_opno = -1; |
3214 | if (internal_fn_p (code: cfn)) |
3215 | mask_opno = internal_fn_mask_index (as_internal_fn (code: cfn)); |
3216 | |
3217 | for (i = 0; i < nargs; i++) |
3218 | { |
3219 | if ((int) i == mask_opno) |
3220 | { |
3221 | if (!vect_check_scalar_mask (vinfo, stmt_info, slp_node, mask_index: mask_opno, |
3222 | mask: &op, mask_node: &slp_op[i], mask_dt_out: &dt[i], mask_vectype_out: &vectypes[i])) |
3223 | return false; |
3224 | continue; |
3225 | } |
3226 | |
3227 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
3228 | i, &op, &slp_op[i], &dt[i], &vectypes[i])) |
3229 | { |
3230 | if (dump_enabled_p ()) |
3231 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3232 | "use not simple.\n" ); |
3233 | return false; |
3234 | } |
3235 | |
3236 | /* We can only handle calls with arguments of the same type. */ |
3237 | if (rhs_type |
3238 | && !types_compatible_p (type1: rhs_type, TREE_TYPE (op))) |
3239 | { |
3240 | if (dump_enabled_p ()) |
3241 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3242 | "argument types differ.\n" ); |
3243 | return false; |
3244 | } |
3245 | if (!rhs_type) |
3246 | rhs_type = TREE_TYPE (op); |
3247 | |
3248 | if (!vectype_in) |
3249 | vectype_in = vectypes[i]; |
3250 | else if (vectypes[i] |
3251 | && !types_compatible_p (type1: vectypes[i], type2: vectype_in)) |
3252 | { |
3253 | if (dump_enabled_p ()) |
3254 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3255 | "argument vector types differ.\n" ); |
3256 | return false; |
3257 | } |
3258 | } |
3259 | /* If all arguments are external or constant defs, infer the vector type |
3260 | from the scalar type. */ |
3261 | if (!vectype_in) |
3262 | vectype_in = get_vectype_for_scalar_type (vinfo, rhs_type, slp_node); |
3263 | if (vec_stmt) |
3264 | gcc_assert (vectype_in); |
3265 | if (!vectype_in) |
3266 | { |
3267 | if (dump_enabled_p ()) |
3268 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3269 | "no vectype for scalar type %T\n" , rhs_type); |
3270 | |
3271 | return false; |
3272 | } |
3273 | |
3274 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
3275 | != VECTOR_BOOLEAN_TYPE_P (vectype_in)) |
3276 | { |
3277 | if (dump_enabled_p ()) |
3278 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3279 | "mixed mask and nonmask vector types\n" ); |
3280 | return false; |
3281 | } |
3282 | |
3283 | if (vect_emulated_vector_p (vectype_in) || vect_emulated_vector_p (vectype_out)) |
3284 | { |
3285 | if (dump_enabled_p ()) |
3286 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3287 | "use emulated vector type for call\n" ); |
3288 | return false; |
3289 | } |
3290 | |
3291 | /* FORNOW */ |
3292 | nunits_in = TYPE_VECTOR_SUBPARTS (node: vectype_in); |
3293 | nunits_out = TYPE_VECTOR_SUBPARTS (node: vectype_out); |
3294 | if (known_eq (nunits_in * 2, nunits_out)) |
3295 | modifier = NARROW; |
3296 | else if (known_eq (nunits_out, nunits_in)) |
3297 | modifier = NONE; |
3298 | else if (known_eq (nunits_out * 2, nunits_in)) |
3299 | modifier = WIDEN; |
3300 | else |
3301 | return false; |
3302 | |
3303 | /* We only handle functions that do not read or clobber memory. */ |
3304 | if (gimple_vuse (g: stmt)) |
3305 | { |
3306 | if (dump_enabled_p ()) |
3307 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3308 | "function reads from or writes to memory.\n" ); |
3309 | return false; |
3310 | } |
3311 | |
3312 | /* For now, we only vectorize functions if a target specific builtin |
3313 | is available. TODO -- in some cases, it might be profitable to |
3314 | insert the calls for pieces of the vector, in order to be able |
3315 | to vectorize other operations in the loop. */ |
3316 | fndecl = NULL_TREE; |
3317 | internal_fn ifn = IFN_LAST; |
3318 | tree callee = gimple_call_fndecl (gs: stmt); |
3319 | |
3320 | /* First try using an internal function. */ |
3321 | code_helper convert_code = MAX_TREE_CODES; |
3322 | if (cfn != CFN_LAST |
3323 | && (modifier == NONE |
3324 | || (modifier == NARROW |
3325 | && simple_integer_narrowing (vectype_out, vectype_in, |
3326 | convert_code: &convert_code)))) |
3327 | ifn = vectorizable_internal_function (cfn, fndecl: callee, vectype_out, |
3328 | vectype_in); |
3329 | |
3330 | /* If that fails, try asking for a target-specific built-in function. */ |
3331 | if (ifn == IFN_LAST) |
3332 | { |
3333 | if (cfn != CFN_LAST) |
3334 | fndecl = targetm.vectorize.builtin_vectorized_function |
3335 | (cfn, vectype_out, vectype_in); |
3336 | else if (callee && fndecl_built_in_p (node: callee, klass: BUILT_IN_MD)) |
3337 | fndecl = targetm.vectorize.builtin_md_vectorized_function |
3338 | (callee, vectype_out, vectype_in); |
3339 | } |
3340 | |
3341 | if (ifn == IFN_LAST && !fndecl) |
3342 | { |
3343 | if (cfn == CFN_GOMP_SIMD_LANE |
3344 | && !slp_node |
3345 | && loop_vinfo |
3346 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid |
3347 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME |
3348 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid |
3349 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) |
3350 | { |
3351 | /* We can handle IFN_GOMP_SIMD_LANE by returning a |
3352 | { 0, 1, 2, ... vf - 1 } vector. */ |
3353 | gcc_assert (nargs == 0); |
3354 | } |
3355 | else if (modifier == NONE |
3356 | && (gimple_call_builtin_p (stmt, BUILT_IN_BSWAP16) |
3357 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP32) |
3358 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP64) |
3359 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP128))) |
3360 | return vectorizable_bswap (vinfo, stmt_info, gsi, vec_stmt, slp_node, |
3361 | slp_op, vectype_in, cost_vec); |
3362 | else |
3363 | { |
3364 | if (dump_enabled_p ()) |
3365 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3366 | "function is not vectorizable.\n" ); |
3367 | return false; |
3368 | } |
3369 | } |
3370 | |
3371 | if (slp_node) |
3372 | ncopies = 1; |
3373 | else if (modifier == NARROW && ifn == IFN_LAST) |
3374 | ncopies = vect_get_num_copies (loop_vinfo, vectype: vectype_out); |
3375 | else |
3376 | ncopies = vect_get_num_copies (loop_vinfo, vectype: vectype_in); |
3377 | |
3378 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
3379 | needs to be generated. */ |
3380 | gcc_assert (ncopies >= 1); |
3381 | |
3382 | int reduc_idx = STMT_VINFO_REDUC_IDX (stmt_info); |
3383 | internal_fn cond_fn = get_conditional_internal_fn (ifn); |
3384 | internal_fn cond_len_fn = get_len_internal_fn (ifn); |
3385 | int len_opno = internal_fn_len_index (cond_len_fn); |
3386 | vec_loop_masks *masks = (loop_vinfo ? &LOOP_VINFO_MASKS (loop_vinfo) : NULL); |
3387 | vec_loop_lens *lens = (loop_vinfo ? &LOOP_VINFO_LENS (loop_vinfo) : NULL); |
3388 | if (!vec_stmt) /* transformation not required. */ |
3389 | { |
3390 | if (slp_node) |
3391 | for (i = 0; i < nargs; ++i) |
3392 | if (!vect_maybe_update_slp_op_vectype (slp_op[i], |
3393 | vectypes[i] |
3394 | ? vectypes[i] : vectype_in)) |
3395 | { |
3396 | if (dump_enabled_p ()) |
3397 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3398 | "incompatible vector types for invariants\n" ); |
3399 | return false; |
3400 | } |
3401 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; |
3402 | DUMP_VECT_SCOPE ("vectorizable_call" ); |
3403 | vect_model_simple_cost (vinfo, stmt_info, |
3404 | ncopies, dt, ndts, node: slp_node, cost_vec); |
3405 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) |
3406 | record_stmt_cost (body_cost_vec: cost_vec, count: ncopies / 2, |
3407 | kind: vec_promote_demote, stmt_info, misalign: 0, where: vect_body); |
3408 | |
3409 | if (loop_vinfo |
3410 | && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) |
3411 | && (reduc_idx >= 0 || mask_opno >= 0)) |
3412 | { |
3413 | if (reduc_idx >= 0 |
3414 | && (cond_fn == IFN_LAST |
3415 | || !direct_internal_fn_supported_p (cond_fn, vectype_out, |
3416 | OPTIMIZE_FOR_SPEED)) |
3417 | && (cond_len_fn == IFN_LAST |
3418 | || !direct_internal_fn_supported_p (cond_len_fn, vectype_out, |
3419 | OPTIMIZE_FOR_SPEED))) |
3420 | { |
3421 | if (dump_enabled_p ()) |
3422 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3423 | "can't use a fully-masked loop because no" |
3424 | " conditional operation is available.\n" ); |
3425 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
3426 | } |
3427 | else |
3428 | { |
3429 | unsigned int nvectors |
3430 | = (slp_node |
3431 | ? SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node) |
3432 | : ncopies); |
3433 | tree scalar_mask = NULL_TREE; |
3434 | if (mask_opno >= 0) |
3435 | scalar_mask = gimple_call_arg (gs: stmt_info->stmt, index: mask_opno); |
3436 | if (cond_len_fn != IFN_LAST |
3437 | && direct_internal_fn_supported_p (cond_len_fn, vectype_out, |
3438 | OPTIMIZE_FOR_SPEED)) |
3439 | vect_record_loop_len (loop_vinfo, lens, nvectors, vectype_out, |
3440 | 1); |
3441 | else |
3442 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype_out, |
3443 | scalar_mask); |
3444 | } |
3445 | } |
3446 | return true; |
3447 | } |
3448 | |
3449 | /* Transform. */ |
3450 | |
3451 | if (dump_enabled_p ()) |
3452 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n" ); |
3453 | |
3454 | /* Handle def. */ |
3455 | scalar_dest = gimple_call_lhs (gs: stmt); |
3456 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); |
3457 | |
3458 | bool masked_loop_p = loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo); |
3459 | bool len_loop_p = loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo); |
3460 | unsigned int vect_nargs = nargs; |
3461 | if (len_loop_p) |
3462 | { |
3463 | if (len_opno >= 0) |
3464 | { |
3465 | ifn = cond_len_fn; |
3466 | /* COND_* -> COND_LEN_* takes 2 extra arguments:LEN,BIAS. */ |
3467 | vect_nargs += 2; |
3468 | } |
3469 | else if (reduc_idx >= 0) |
3470 | gcc_unreachable (); |
3471 | } |
3472 | else if (masked_loop_p && reduc_idx >= 0) |
3473 | { |
3474 | ifn = cond_fn; |
3475 | vect_nargs += 2; |
3476 | } |
3477 | if (clz_ctz_arg1) |
3478 | ++vect_nargs; |
3479 | |
3480 | if (modifier == NONE || ifn != IFN_LAST) |
3481 | { |
3482 | tree prev_res = NULL_TREE; |
3483 | vargs.safe_grow (len: vect_nargs, exact: true); |
3484 | auto_vec<vec<tree> > vec_defs (nargs); |
3485 | for (j = 0; j < ncopies; ++j) |
3486 | { |
3487 | /* Build argument list for the vectorized call. */ |
3488 | if (slp_node) |
3489 | { |
3490 | vec<tree> vec_oprnds0; |
3491 | |
3492 | vect_get_slp_defs (vinfo, slp_node, &vec_defs); |
3493 | vec_oprnds0 = vec_defs[0]; |
3494 | |
3495 | /* Arguments are ready. Create the new vector stmt. */ |
3496 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
3497 | { |
3498 | int varg = 0; |
3499 | if (masked_loop_p && reduc_idx >= 0) |
3500 | { |
3501 | unsigned int vec_num = vec_oprnds0.length (); |
3502 | /* Always true for SLP. */ |
3503 | gcc_assert (ncopies == 1); |
3504 | vargs[varg++] = vect_get_loop_mask (loop_vinfo, |
3505 | gsi, masks, vec_num, |
3506 | vectype_out, i); |
3507 | } |
3508 | size_t k; |
3509 | for (k = 0; k < nargs; k++) |
3510 | { |
3511 | vec<tree> vec_oprndsk = vec_defs[k]; |
3512 | vargs[varg++] = vec_oprndsk[i]; |
3513 | } |
3514 | if (masked_loop_p && reduc_idx >= 0) |
3515 | vargs[varg++] = vargs[reduc_idx + 1]; |
3516 | if (clz_ctz_arg1) |
3517 | vargs[varg++] = clz_ctz_arg1; |
3518 | |
3519 | gimple *new_stmt; |
3520 | if (modifier == NARROW) |
3521 | { |
3522 | /* We don't define any narrowing conditional functions |
3523 | at present. */ |
3524 | gcc_assert (mask_opno < 0); |
3525 | tree half_res = make_ssa_name (var: vectype_in); |
3526 | gcall *call |
3527 | = gimple_build_call_internal_vec (ifn, vargs); |
3528 | gimple_call_set_lhs (gs: call, lhs: half_res); |
3529 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
3530 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
3531 | if ((i & 1) == 0) |
3532 | { |
3533 | prev_res = half_res; |
3534 | continue; |
3535 | } |
3536 | new_temp = make_ssa_name (var: vec_dest); |
3537 | new_stmt = vect_gimple_build (new_temp, convert_code, |
3538 | prev_res, half_res); |
3539 | vect_finish_stmt_generation (vinfo, stmt_info, |
3540 | vec_stmt: new_stmt, gsi); |
3541 | } |
3542 | else |
3543 | { |
3544 | if (len_opno >= 0 && len_loop_p) |
3545 | { |
3546 | unsigned int vec_num = vec_oprnds0.length (); |
3547 | /* Always true for SLP. */ |
3548 | gcc_assert (ncopies == 1); |
3549 | tree len |
3550 | = vect_get_loop_len (loop_vinfo, gsi, lens, vec_num, |
3551 | vectype_out, i, 1); |
3552 | signed char biasval |
3553 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
3554 | tree bias = build_int_cst (intQI_type_node, biasval); |
3555 | vargs[len_opno] = len; |
3556 | vargs[len_opno + 1] = bias; |
3557 | } |
3558 | else if (mask_opno >= 0 && masked_loop_p) |
3559 | { |
3560 | unsigned int vec_num = vec_oprnds0.length (); |
3561 | /* Always true for SLP. */ |
3562 | gcc_assert (ncopies == 1); |
3563 | tree mask = vect_get_loop_mask (loop_vinfo, |
3564 | gsi, masks, vec_num, |
3565 | vectype_out, i); |
3566 | vargs[mask_opno] = prepare_vec_mask |
3567 | (loop_vinfo, TREE_TYPE (mask), loop_mask: mask, |
3568 | vec_mask: vargs[mask_opno], gsi); |
3569 | } |
3570 | |
3571 | gcall *call; |
3572 | if (ifn != IFN_LAST) |
3573 | call = gimple_build_call_internal_vec (ifn, vargs); |
3574 | else |
3575 | call = gimple_build_call_vec (fndecl, vargs); |
3576 | new_temp = make_ssa_name (var: vec_dest, stmt: call); |
3577 | gimple_call_set_lhs (gs: call, lhs: new_temp); |
3578 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
3579 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
3580 | new_stmt = call; |
3581 | } |
3582 | slp_node->push_vec_def (def: new_stmt); |
3583 | } |
3584 | continue; |
3585 | } |
3586 | |
3587 | int varg = 0; |
3588 | if (masked_loop_p && reduc_idx >= 0) |
3589 | vargs[varg++] = vect_get_loop_mask (loop_vinfo, gsi, masks, ncopies, |
3590 | vectype_out, j); |
3591 | for (i = 0; i < nargs; i++) |
3592 | { |
3593 | op = gimple_call_arg (gs: stmt, index: i); |
3594 | if (j == 0) |
3595 | { |
3596 | vec_defs.quick_push (obj: vNULL); |
3597 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies, |
3598 | op, vec_oprnds: &vec_defs[i], |
3599 | vectype: vectypes[i]); |
3600 | } |
3601 | vargs[varg++] = vec_defs[i][j]; |
3602 | } |
3603 | if (masked_loop_p && reduc_idx >= 0) |
3604 | vargs[varg++] = vargs[reduc_idx + 1]; |
3605 | if (clz_ctz_arg1) |
3606 | vargs[varg++] = clz_ctz_arg1; |
3607 | |
3608 | if (len_opno >= 0 && len_loop_p) |
3609 | { |
3610 | tree len = vect_get_loop_len (loop_vinfo, gsi, lens, ncopies, |
3611 | vectype_out, j, 1); |
3612 | signed char biasval |
3613 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
3614 | tree bias = build_int_cst (intQI_type_node, biasval); |
3615 | vargs[len_opno] = len; |
3616 | vargs[len_opno + 1] = bias; |
3617 | } |
3618 | else if (mask_opno >= 0 && masked_loop_p) |
3619 | { |
3620 | tree mask = vect_get_loop_mask (loop_vinfo, gsi, masks, ncopies, |
3621 | vectype_out, j); |
3622 | vargs[mask_opno] |
3623 | = prepare_vec_mask (loop_vinfo, TREE_TYPE (mask), loop_mask: mask, |
3624 | vec_mask: vargs[mask_opno], gsi); |
3625 | } |
3626 | |
3627 | gimple *new_stmt; |
3628 | if (cfn == CFN_GOMP_SIMD_LANE) |
3629 | { |
3630 | tree cst = build_index_vector (vectype_out, j * nunits_out, 1); |
3631 | tree new_var |
3632 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_" ); |
3633 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
3634 | vect_init_vector_1 (vinfo, stmt_vinfo: stmt_info, new_stmt: init_stmt, NULL); |
3635 | new_temp = make_ssa_name (var: vec_dest); |
3636 | new_stmt = gimple_build_assign (new_temp, new_var); |
3637 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
3638 | } |
3639 | else if (modifier == NARROW) |
3640 | { |
3641 | /* We don't define any narrowing conditional functions at |
3642 | present. */ |
3643 | gcc_assert (mask_opno < 0); |
3644 | tree half_res = make_ssa_name (var: vectype_in); |
3645 | gcall *call = gimple_build_call_internal_vec (ifn, vargs); |
3646 | gimple_call_set_lhs (gs: call, lhs: half_res); |
3647 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
3648 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
3649 | if ((j & 1) == 0) |
3650 | { |
3651 | prev_res = half_res; |
3652 | continue; |
3653 | } |
3654 | new_temp = make_ssa_name (var: vec_dest); |
3655 | new_stmt = vect_gimple_build (new_temp, convert_code, prev_res, |
3656 | half_res); |
3657 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
3658 | } |
3659 | else |
3660 | { |
3661 | gcall *call; |
3662 | if (ifn != IFN_LAST) |
3663 | call = gimple_build_call_internal_vec (ifn, vargs); |
3664 | else |
3665 | call = gimple_build_call_vec (fndecl, vargs); |
3666 | new_temp = make_ssa_name (var: vec_dest, stmt: call); |
3667 | gimple_call_set_lhs (gs: call, lhs: new_temp); |
3668 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
3669 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
3670 | new_stmt = call; |
3671 | } |
3672 | |
3673 | if (j == (modifier == NARROW ? 1 : 0)) |
3674 | *vec_stmt = new_stmt; |
3675 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
3676 | } |
3677 | for (i = 0; i < nargs; i++) |
3678 | { |
3679 | vec<tree> vec_oprndsi = vec_defs[i]; |
3680 | vec_oprndsi.release (); |
3681 | } |
3682 | } |
3683 | else if (modifier == NARROW) |
3684 | { |
3685 | auto_vec<vec<tree> > vec_defs (nargs); |
3686 | /* We don't define any narrowing conditional functions at present. */ |
3687 | gcc_assert (mask_opno < 0); |
3688 | for (j = 0; j < ncopies; ++j) |
3689 | { |
3690 | /* Build argument list for the vectorized call. */ |
3691 | if (j == 0) |
3692 | vargs.create (nelems: nargs * 2); |
3693 | else |
3694 | vargs.truncate (size: 0); |
3695 | |
3696 | if (slp_node) |
3697 | { |
3698 | vec<tree> vec_oprnds0; |
3699 | |
3700 | vect_get_slp_defs (vinfo, slp_node, &vec_defs); |
3701 | vec_oprnds0 = vec_defs[0]; |
3702 | |
3703 | /* Arguments are ready. Create the new vector stmt. */ |
3704 | for (i = 0; vec_oprnds0.iterate (ix: i, ptr: &vec_oprnd0); i += 2) |
3705 | { |
3706 | size_t k; |
3707 | vargs.truncate (size: 0); |
3708 | for (k = 0; k < nargs; k++) |
3709 | { |
3710 | vec<tree> vec_oprndsk = vec_defs[k]; |
3711 | vargs.quick_push (obj: vec_oprndsk[i]); |
3712 | vargs.quick_push (obj: vec_oprndsk[i + 1]); |
3713 | } |
3714 | gcall *call; |
3715 | if (ifn != IFN_LAST) |
3716 | call = gimple_build_call_internal_vec (ifn, vargs); |
3717 | else |
3718 | call = gimple_build_call_vec (fndecl, vargs); |
3719 | new_temp = make_ssa_name (var: vec_dest, stmt: call); |
3720 | gimple_call_set_lhs (gs: call, lhs: new_temp); |
3721 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
3722 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
3723 | slp_node->push_vec_def (def: call); |
3724 | } |
3725 | continue; |
3726 | } |
3727 | |
3728 | for (i = 0; i < nargs; i++) |
3729 | { |
3730 | op = gimple_call_arg (gs: stmt, index: i); |
3731 | if (j == 0) |
3732 | { |
3733 | vec_defs.quick_push (obj: vNULL); |
3734 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies: 2 * ncopies, |
3735 | op, vec_oprnds: &vec_defs[i], vectype: vectypes[i]); |
3736 | } |
3737 | vec_oprnd0 = vec_defs[i][2*j]; |
3738 | vec_oprnd1 = vec_defs[i][2*j+1]; |
3739 | |
3740 | vargs.quick_push (obj: vec_oprnd0); |
3741 | vargs.quick_push (obj: vec_oprnd1); |
3742 | } |
3743 | |
3744 | gcall *new_stmt = gimple_build_call_vec (fndecl, vargs); |
3745 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
3746 | gimple_call_set_lhs (gs: new_stmt, lhs: new_temp); |
3747 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
3748 | |
3749 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
3750 | } |
3751 | |
3752 | if (!slp_node) |
3753 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
3754 | |
3755 | for (i = 0; i < nargs; i++) |
3756 | { |
3757 | vec<tree> vec_oprndsi = vec_defs[i]; |
3758 | vec_oprndsi.release (); |
3759 | } |
3760 | } |
3761 | else |
3762 | /* No current target implements this case. */ |
3763 | return false; |
3764 | |
3765 | vargs.release (); |
3766 | |
3767 | /* The call in STMT might prevent it from being removed in dce. |
3768 | We however cannot remove it here, due to the way the ssa name |
3769 | it defines is mapped to the new definition. So just replace |
3770 | rhs of the statement with something harmless. */ |
3771 | |
3772 | if (slp_node) |
3773 | return true; |
3774 | |
3775 | stmt_info = vect_orig_stmt (stmt_info); |
3776 | lhs = gimple_get_lhs (stmt_info->stmt); |
3777 | |
3778 | gassign *new_stmt |
3779 | = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); |
3780 | vinfo->replace_stmt (gsi, stmt_info, new_stmt); |
3781 | |
3782 | return true; |
3783 | } |
3784 | |
3785 | |
3786 | struct simd_call_arg_info |
3787 | { |
3788 | tree vectype; |
3789 | tree op; |
3790 | HOST_WIDE_INT linear_step; |
3791 | enum vect_def_type dt; |
3792 | unsigned int align; |
3793 | bool simd_lane_linear; |
3794 | }; |
3795 | |
3796 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
3797 | is linear within simd lane (but not within whole loop), note it in |
3798 | *ARGINFO. */ |
3799 | |
3800 | static void |
3801 | vect_simd_lane_linear (tree op, class loop *loop, |
3802 | struct simd_call_arg_info *arginfo) |
3803 | { |
3804 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
3805 | |
3806 | if (!is_gimple_assign (gs: def_stmt) |
3807 | || gimple_assign_rhs_code (gs: def_stmt) != POINTER_PLUS_EXPR |
3808 | || !is_gimple_min_invariant (gimple_assign_rhs1 (gs: def_stmt))) |
3809 | return; |
3810 | |
3811 | tree base = gimple_assign_rhs1 (gs: def_stmt); |
3812 | HOST_WIDE_INT linear_step = 0; |
3813 | tree v = gimple_assign_rhs2 (gs: def_stmt); |
3814 | while (TREE_CODE (v) == SSA_NAME) |
3815 | { |
3816 | tree t; |
3817 | def_stmt = SSA_NAME_DEF_STMT (v); |
3818 | if (is_gimple_assign (gs: def_stmt)) |
3819 | switch (gimple_assign_rhs_code (gs: def_stmt)) |
3820 | { |
3821 | case PLUS_EXPR: |
3822 | t = gimple_assign_rhs2 (gs: def_stmt); |
3823 | if (linear_step || TREE_CODE (t) != INTEGER_CST) |
3824 | return; |
3825 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); |
3826 | v = gimple_assign_rhs1 (gs: def_stmt); |
3827 | continue; |
3828 | case MULT_EXPR: |
3829 | t = gimple_assign_rhs2 (gs: def_stmt); |
3830 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) |
3831 | return; |
3832 | linear_step = tree_to_shwi (t); |
3833 | v = gimple_assign_rhs1 (gs: def_stmt); |
3834 | continue; |
3835 | CASE_CONVERT: |
3836 | t = gimple_assign_rhs1 (gs: def_stmt); |
3837 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE |
3838 | || (TYPE_PRECISION (TREE_TYPE (v)) |
3839 | < TYPE_PRECISION (TREE_TYPE (t)))) |
3840 | return; |
3841 | if (!linear_step) |
3842 | linear_step = 1; |
3843 | v = t; |
3844 | continue; |
3845 | default: |
3846 | return; |
3847 | } |
3848 | else if (gimple_call_internal_p (gs: def_stmt, fn: IFN_GOMP_SIMD_LANE) |
3849 | && loop->simduid |
3850 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME |
3851 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) |
3852 | == loop->simduid)) |
3853 | { |
3854 | if (!linear_step) |
3855 | linear_step = 1; |
3856 | arginfo->linear_step = linear_step; |
3857 | arginfo->op = base; |
3858 | arginfo->simd_lane_linear = true; |
3859 | return; |
3860 | } |
3861 | } |
3862 | } |
3863 | |
3864 | /* Function vectorizable_simd_clone_call. |
3865 | |
3866 | Check if STMT_INFO performs a function call that can be vectorized |
3867 | by calling a simd clone of the function. |
3868 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
3869 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
3870 | Return true if STMT_INFO is vectorizable in this way. */ |
3871 | |
3872 | static bool |
3873 | vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info, |
3874 | gimple_stmt_iterator *gsi, |
3875 | gimple **vec_stmt, slp_tree slp_node, |
3876 | stmt_vector_for_cost *) |
3877 | { |
3878 | tree vec_dest; |
3879 | tree scalar_dest; |
3880 | tree op, type; |
3881 | tree vec_oprnd0 = NULL_TREE; |
3882 | tree vectype; |
3883 | poly_uint64 nunits; |
3884 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
3885 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
3886 | class loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
3887 | tree fndecl, new_temp; |
3888 | int ncopies, j; |
3889 | auto_vec<simd_call_arg_info> arginfo; |
3890 | vec<tree> vargs = vNULL; |
3891 | size_t i, nargs; |
3892 | tree lhs, rtype, ratype; |
3893 | vec<constructor_elt, va_gc> *ret_ctor_elts = NULL; |
3894 | int masked_call_offset = 0; |
3895 | |
3896 | /* Is STMT a vectorizable call? */ |
3897 | gcall *stmt = dyn_cast <gcall *> (p: stmt_info->stmt); |
3898 | if (!stmt) |
3899 | return false; |
3900 | |
3901 | fndecl = gimple_call_fndecl (gs: stmt); |
3902 | if (fndecl == NULL_TREE |
3903 | && gimple_call_internal_p (gs: stmt, fn: IFN_MASK_CALL)) |
3904 | { |
3905 | fndecl = gimple_call_arg (gs: stmt, index: 0); |
3906 | gcc_checking_assert (TREE_CODE (fndecl) == ADDR_EXPR); |
3907 | fndecl = TREE_OPERAND (fndecl, 0); |
3908 | gcc_checking_assert (TREE_CODE (fndecl) == FUNCTION_DECL); |
3909 | masked_call_offset = 1; |
3910 | } |
3911 | if (fndecl == NULL_TREE) |
3912 | return false; |
3913 | |
3914 | struct cgraph_node *node = cgraph_node::get (decl: fndecl); |
3915 | if (node == NULL || node->simd_clones == NULL) |
3916 | return false; |
3917 | |
3918 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
3919 | return false; |
3920 | |
3921 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3922 | && ! vec_stmt) |
3923 | return false; |
3924 | |
3925 | if (gimple_call_lhs (gs: stmt) |
3926 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) |
3927 | return false; |
3928 | |
3929 | gcc_checking_assert (!stmt_can_throw_internal (cfun, stmt)); |
3930 | |
3931 | vectype = STMT_VINFO_VECTYPE (stmt_info); |
3932 | |
3933 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt_info)) |
3934 | return false; |
3935 | |
3936 | /* Process function arguments. */ |
3937 | nargs = gimple_call_num_args (gs: stmt) - masked_call_offset; |
3938 | |
3939 | /* Bail out if the function has zero arguments. */ |
3940 | if (nargs == 0) |
3941 | return false; |
3942 | |
3943 | vec<tree>& simd_clone_info = (slp_node ? SLP_TREE_SIMD_CLONE_INFO (slp_node) |
3944 | : STMT_VINFO_SIMD_CLONE_INFO (stmt_info)); |
3945 | arginfo.reserve (nelems: nargs, exact: true); |
3946 | auto_vec<slp_tree> slp_op; |
3947 | slp_op.safe_grow_cleared (len: nargs); |
3948 | |
3949 | for (i = 0; i < nargs; i++) |
3950 | { |
3951 | simd_call_arg_info thisarginfo; |
3952 | affine_iv iv; |
3953 | |
3954 | thisarginfo.linear_step = 0; |
3955 | thisarginfo.align = 0; |
3956 | thisarginfo.op = NULL_TREE; |
3957 | thisarginfo.simd_lane_linear = false; |
3958 | |
3959 | int op_no = i + masked_call_offset; |
3960 | if (slp_node) |
3961 | op_no = vect_slp_child_index_for_operand (stmt, op: op_no, false); |
3962 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
3963 | op_no, &op, &slp_op[i], |
3964 | &thisarginfo.dt, &thisarginfo.vectype) |
3965 | || thisarginfo.dt == vect_uninitialized_def) |
3966 | { |
3967 | if (dump_enabled_p ()) |
3968 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3969 | "use not simple.\n" ); |
3970 | return false; |
3971 | } |
3972 | |
3973 | if (thisarginfo.dt == vect_constant_def |
3974 | || thisarginfo.dt == vect_external_def) |
3975 | { |
3976 | /* With SLP we determine the vector type of constants/externals |
3977 | at analysis time, handling conflicts via |
3978 | vect_maybe_update_slp_op_vectype. At transform time |
3979 | we have a vector type recorded for SLP. */ |
3980 | gcc_assert (!vec_stmt |
3981 | || !slp_node |
3982 | || thisarginfo.vectype != NULL_TREE); |
3983 | if (!vec_stmt) |
3984 | thisarginfo.vectype = get_vectype_for_scalar_type (vinfo, |
3985 | TREE_TYPE (op), |
3986 | slp_node); |
3987 | } |
3988 | else |
3989 | gcc_assert (thisarginfo.vectype != NULL_TREE); |
3990 | |
3991 | /* For linear arguments, the analyze phase should have saved |
3992 | the base and step in {STMT_VINFO,SLP_TREE}_SIMD_CLONE_INFO. */ |
3993 | if (i * 3 + 4 <= simd_clone_info.length () |
3994 | && simd_clone_info[i * 3 + 2]) |
3995 | { |
3996 | gcc_assert (vec_stmt); |
3997 | thisarginfo.linear_step = tree_to_shwi (simd_clone_info[i * 3 + 2]); |
3998 | thisarginfo.op = simd_clone_info[i * 3 + 1]; |
3999 | thisarginfo.simd_lane_linear |
4000 | = (simd_clone_info[i * 3 + 3] == boolean_true_node); |
4001 | /* If loop has been peeled for alignment, we need to adjust it. */ |
4002 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); |
4003 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); |
4004 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
4005 | { |
4006 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); |
4007 | tree step = simd_clone_info[i * 3 + 2]; |
4008 | tree opt = TREE_TYPE (thisarginfo.op); |
4009 | bias = fold_convert (TREE_TYPE (step), bias); |
4010 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); |
4011 | thisarginfo.op |
4012 | = fold_build2 (POINTER_TYPE_P (opt) |
4013 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, |
4014 | thisarginfo.op, bias); |
4015 | } |
4016 | } |
4017 | else if (!vec_stmt |
4018 | && thisarginfo.dt != vect_constant_def |
4019 | && thisarginfo.dt != vect_external_def |
4020 | && loop_vinfo |
4021 | && TREE_CODE (op) == SSA_NAME |
4022 | && simple_iv (loop, loop_containing_stmt (stmt), op, |
4023 | &iv, false) |
4024 | && tree_fits_shwi_p (iv.step)) |
4025 | { |
4026 | thisarginfo.linear_step = tree_to_shwi (iv.step); |
4027 | thisarginfo.op = iv.base; |
4028 | } |
4029 | else if ((thisarginfo.dt == vect_constant_def |
4030 | || thisarginfo.dt == vect_external_def) |
4031 | && POINTER_TYPE_P (TREE_TYPE (op))) |
4032 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; |
4033 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
4034 | linear too. */ |
4035 | if (POINTER_TYPE_P (TREE_TYPE (op)) |
4036 | && !thisarginfo.linear_step |
4037 | && !vec_stmt |
4038 | && thisarginfo.dt != vect_constant_def |
4039 | && thisarginfo.dt != vect_external_def |
4040 | && loop_vinfo |
4041 | && TREE_CODE (op) == SSA_NAME) |
4042 | vect_simd_lane_linear (op, loop, arginfo: &thisarginfo); |
4043 | |
4044 | arginfo.quick_push (obj: thisarginfo); |
4045 | } |
4046 | |
4047 | poly_uint64 vf = loop_vinfo ? LOOP_VINFO_VECT_FACTOR (loop_vinfo) : 1; |
4048 | unsigned group_size = slp_node ? SLP_TREE_LANES (slp_node) : 1; |
4049 | unsigned int badness = 0; |
4050 | struct cgraph_node *bestn = NULL; |
4051 | if (simd_clone_info.exists ()) |
4052 | bestn = cgraph_node::get (decl: simd_clone_info[0]); |
4053 | else |
4054 | for (struct cgraph_node *n = node->simd_clones; n != NULL; |
4055 | n = n->simdclone->next_clone) |
4056 | { |
4057 | unsigned int this_badness = 0; |
4058 | unsigned int num_calls; |
4059 | /* The number of arguments in the call and the number of parameters in |
4060 | the simdclone should match. However, when the simdclone is |
4061 | 'inbranch', it could have one more paramater than nargs when using |
4062 | an inbranch simdclone to call a non-inbranch call, either in a |
4063 | non-masked loop using a all true constant mask, or inside a masked |
4064 | loop using it's mask. */ |
4065 | size_t simd_nargs = n->simdclone->nargs; |
4066 | if (!masked_call_offset && n->simdclone->inbranch) |
4067 | simd_nargs--; |
4068 | if (!constant_multiple_p (a: vf * group_size, b: n->simdclone->simdlen, |
4069 | multiple: &num_calls) |
4070 | || (!n->simdclone->inbranch && (masked_call_offset > 0)) |
4071 | || (nargs != simd_nargs)) |
4072 | continue; |
4073 | if (num_calls != 1) |
4074 | this_badness += floor_log2 (x: num_calls) * 4096; |
4075 | if (n->simdclone->inbranch) |
4076 | this_badness += 8192; |
4077 | int target_badness = targetm.simd_clone.usable (n); |
4078 | if (target_badness < 0) |
4079 | continue; |
4080 | this_badness += target_badness * 512; |
4081 | for (i = 0; i < nargs; i++) |
4082 | { |
4083 | switch (n->simdclone->args[i].arg_type) |
4084 | { |
4085 | case SIMD_CLONE_ARG_TYPE_VECTOR: |
4086 | if (!useless_type_conversion_p |
4087 | (n->simdclone->args[i].orig_type, |
4088 | TREE_TYPE (gimple_call_arg (stmt, |
4089 | i + masked_call_offset)))) |
4090 | i = -1; |
4091 | else if (arginfo[i].dt == vect_constant_def |
4092 | || arginfo[i].dt == vect_external_def |
4093 | || arginfo[i].linear_step) |
4094 | this_badness += 64; |
4095 | break; |
4096 | case SIMD_CLONE_ARG_TYPE_UNIFORM: |
4097 | if (arginfo[i].dt != vect_constant_def |
4098 | && arginfo[i].dt != vect_external_def) |
4099 | i = -1; |
4100 | break; |
4101 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: |
4102 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
4103 | if (arginfo[i].dt == vect_constant_def |
4104 | || arginfo[i].dt == vect_external_def |
4105 | || (arginfo[i].linear_step |
4106 | != n->simdclone->args[i].linear_step)) |
4107 | i = -1; |
4108 | break; |
4109 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
4110 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
4111 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: |
4112 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
4113 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: |
4114 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: |
4115 | /* FORNOW */ |
4116 | i = -1; |
4117 | break; |
4118 | case SIMD_CLONE_ARG_TYPE_MASK: |
4119 | /* While we can create a traditional data vector from |
4120 | an incoming integer mode mask we have no good way to |
4121 | force generate an integer mode mask from a traditional |
4122 | boolean vector input. */ |
4123 | if (SCALAR_INT_MODE_P (n->simdclone->mask_mode) |
4124 | && !SCALAR_INT_MODE_P (TYPE_MODE (arginfo[i].vectype))) |
4125 | i = -1; |
4126 | else if (!SCALAR_INT_MODE_P (n->simdclone->mask_mode) |
4127 | && SCALAR_INT_MODE_P (TYPE_MODE (arginfo[i].vectype))) |
4128 | this_badness += 2048; |
4129 | break; |
4130 | } |
4131 | if (i == (size_t) -1) |
4132 | break; |
4133 | if (n->simdclone->args[i].alignment > arginfo[i].align) |
4134 | { |
4135 | i = -1; |
4136 | break; |
4137 | } |
4138 | if (arginfo[i].align) |
4139 | this_badness += (exact_log2 (x: arginfo[i].align) |
4140 | - exact_log2 (x: n->simdclone->args[i].alignment)); |
4141 | } |
4142 | if (i == (size_t) -1) |
4143 | continue; |
4144 | if (masked_call_offset == 0 |
4145 | && n->simdclone->inbranch |
4146 | && n->simdclone->nargs > nargs) |
4147 | { |
4148 | gcc_assert (n->simdclone->args[n->simdclone->nargs - 1].arg_type == |
4149 | SIMD_CLONE_ARG_TYPE_MASK); |
4150 | /* Penalize using a masked SIMD clone in a non-masked loop, that is |
4151 | not in a branch, as we'd have to construct an all-true mask. */ |
4152 | if (!loop_vinfo || !LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) |
4153 | this_badness += 64; |
4154 | } |
4155 | if (bestn == NULL || this_badness < badness) |
4156 | { |
4157 | bestn = n; |
4158 | badness = this_badness; |
4159 | } |
4160 | } |
4161 | |
4162 | if (bestn == NULL) |
4163 | return false; |
4164 | |
4165 | unsigned int num_mask_args = 0; |
4166 | if (SCALAR_INT_MODE_P (bestn->simdclone->mask_mode)) |
4167 | for (i = 0; i < nargs; i++) |
4168 | if (bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_MASK) |
4169 | num_mask_args++; |
4170 | |
4171 | for (i = 0; i < nargs; i++) |
4172 | { |
4173 | if ((arginfo[i].dt == vect_constant_def |
4174 | || arginfo[i].dt == vect_external_def) |
4175 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) |
4176 | { |
4177 | tree arg_type = TREE_TYPE (gimple_call_arg (stmt, |
4178 | i + masked_call_offset)); |
4179 | arginfo[i].vectype = get_vectype_for_scalar_type (vinfo, arg_type, |
4180 | slp_node); |
4181 | if (arginfo[i].vectype == NULL |
4182 | || !constant_multiple_p (a: bestn->simdclone->simdlen, |
4183 | b: TYPE_VECTOR_SUBPARTS (node: arginfo[i].vectype))) |
4184 | return false; |
4185 | } |
4186 | |
4187 | if (bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR |
4188 | && VECTOR_BOOLEAN_TYPE_P (bestn->simdclone->args[i].vector_type)) |
4189 | { |
4190 | if (dump_enabled_p ()) |
4191 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4192 | "vector mask arguments are not supported.\n" ); |
4193 | return false; |
4194 | } |
4195 | |
4196 | if (bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_MASK) |
4197 | { |
4198 | tree clone_arg_vectype = bestn->simdclone->args[i].vector_type; |
4199 | if (bestn->simdclone->mask_mode == VOIDmode) |
4200 | { |
4201 | if (maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: clone_arg_vectype), |
4202 | b: TYPE_VECTOR_SUBPARTS (node: arginfo[i].vectype))) |
4203 | { |
4204 | /* FORNOW we only have partial support for vector-type masks |
4205 | that can't hold all of simdlen. */ |
4206 | if (dump_enabled_p ()) |
4207 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, |
4208 | vect_location, |
4209 | "in-branch vector clones are not yet" |
4210 | " supported for mismatched vector sizes.\n" ); |
4211 | return false; |
4212 | } |
4213 | if (!expand_vec_cond_expr_p (clone_arg_vectype, |
4214 | arginfo[i].vectype, ERROR_MARK)) |
4215 | { |
4216 | if (dump_enabled_p ()) |
4217 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, |
4218 | vect_location, |
4219 | "cannot compute mask argument for" |
4220 | " in-branch vector clones.\n" ); |
4221 | return false; |
4222 | } |
4223 | } |
4224 | else if (SCALAR_INT_MODE_P (bestn->simdclone->mask_mode)) |
4225 | { |
4226 | if (!SCALAR_INT_MODE_P (TYPE_MODE (arginfo[i].vectype)) |
4227 | || maybe_ne (a: exact_div (a: bestn->simdclone->simdlen, |
4228 | b: num_mask_args), |
4229 | b: TYPE_VECTOR_SUBPARTS (node: arginfo[i].vectype))) |
4230 | { |
4231 | /* FORNOW we only have partial support for integer-type masks |
4232 | that represent the same number of lanes as the |
4233 | vectorized mask inputs. */ |
4234 | if (dump_enabled_p ()) |
4235 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, |
4236 | vect_location, |
4237 | "in-branch vector clones are not yet " |
4238 | "supported for mismatched vector sizes.\n" ); |
4239 | return false; |
4240 | } |
4241 | } |
4242 | else |
4243 | { |
4244 | if (dump_enabled_p ()) |
4245 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, |
4246 | vect_location, |
4247 | "in-branch vector clones not supported" |
4248 | " on this target.\n" ); |
4249 | return false; |
4250 | } |
4251 | } |
4252 | } |
4253 | |
4254 | fndecl = bestn->decl; |
4255 | nunits = bestn->simdclone->simdlen; |
4256 | if (slp_node) |
4257 | ncopies = vector_unroll_factor (vf * group_size, nunits); |
4258 | else |
4259 | ncopies = vector_unroll_factor (vf, nunits); |
4260 | |
4261 | /* If the function isn't const, only allow it in simd loops where user |
4262 | has asserted that at least nunits consecutive iterations can be |
4263 | performed using SIMD instructions. */ |
4264 | if ((loop == NULL || maybe_lt (a: (unsigned) loop->safelen, b: nunits)) |
4265 | && gimple_vuse (g: stmt)) |
4266 | return false; |
4267 | |
4268 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
4269 | needs to be generated. */ |
4270 | gcc_assert (ncopies >= 1); |
4271 | |
4272 | if (!vec_stmt) /* transformation not required. */ |
4273 | { |
4274 | if (slp_node) |
4275 | for (unsigned i = 0; i < nargs; ++i) |
4276 | if (!vect_maybe_update_slp_op_vectype (slp_op[i], arginfo[i].vectype)) |
4277 | { |
4278 | if (dump_enabled_p ()) |
4279 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4280 | "incompatible vector types for invariants\n" ); |
4281 | return false; |
4282 | } |
4283 | /* When the original call is pure or const but the SIMD ABI dictates |
4284 | an aggregate return we will have to use a virtual definition and |
4285 | in a loop eventually even need to add a virtual PHI. That's |
4286 | not straight-forward so allow to fix this up via renaming. */ |
4287 | if (gimple_call_lhs (gs: stmt) |
4288 | && !gimple_vdef (g: stmt) |
4289 | && TREE_CODE (TREE_TYPE (TREE_TYPE (bestn->decl))) == ARRAY_TYPE) |
4290 | vinfo->any_known_not_updated_vssa = true; |
4291 | /* ??? For SLP code-gen we end up inserting after the last |
4292 | vector argument def rather than at the original call position |
4293 | so automagic virtual operand updating doesn't work. */ |
4294 | if (gimple_vuse (g: stmt) && slp_node) |
4295 | vinfo->any_known_not_updated_vssa = true; |
4296 | simd_clone_info.safe_push (obj: bestn->decl); |
4297 | for (i = 0; i < bestn->simdclone->nargs; i++) |
4298 | { |
4299 | switch (bestn->simdclone->args[i].arg_type) |
4300 | { |
4301 | default: |
4302 | continue; |
4303 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: |
4304 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
4305 | { |
4306 | simd_clone_info.safe_grow_cleared (len: i * 3 + 1, exact: true); |
4307 | simd_clone_info.safe_push (obj: arginfo[i].op); |
4308 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) |
4309 | ? size_type_node : TREE_TYPE (arginfo[i].op); |
4310 | tree ls = build_int_cst (lst, arginfo[i].linear_step); |
4311 | simd_clone_info.safe_push (obj: ls); |
4312 | tree sll = arginfo[i].simd_lane_linear |
4313 | ? boolean_true_node : boolean_false_node; |
4314 | simd_clone_info.safe_push (obj: sll); |
4315 | } |
4316 | break; |
4317 | case SIMD_CLONE_ARG_TYPE_MASK: |
4318 | if (loop_vinfo |
4319 | && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) |
4320 | vect_record_loop_mask (loop_vinfo, |
4321 | &LOOP_VINFO_MASKS (loop_vinfo), |
4322 | ncopies, vectype, op); |
4323 | |
4324 | break; |
4325 | } |
4326 | } |
4327 | |
4328 | if (!bestn->simdclone->inbranch && loop_vinfo) |
4329 | { |
4330 | if (dump_enabled_p () |
4331 | && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) |
4332 | dump_printf_loc (MSG_NOTE, vect_location, |
4333 | "can't use a fully-masked loop because a" |
4334 | " non-masked simd clone was selected.\n" ); |
4335 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
4336 | } |
4337 | |
4338 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
4339 | DUMP_VECT_SCOPE ("vectorizable_simd_clone_call" ); |
4340 | /* vect_model_simple_cost (vinfo, stmt_info, ncopies, |
4341 | dt, slp_node, cost_vec); */ |
4342 | return true; |
4343 | } |
4344 | |
4345 | /* Transform. */ |
4346 | |
4347 | if (dump_enabled_p ()) |
4348 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n" ); |
4349 | |
4350 | /* Handle def. */ |
4351 | scalar_dest = gimple_call_lhs (gs: stmt); |
4352 | vec_dest = NULL_TREE; |
4353 | rtype = NULL_TREE; |
4354 | ratype = NULL_TREE; |
4355 | if (scalar_dest) |
4356 | { |
4357 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
4358 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); |
4359 | if (TREE_CODE (rtype) == ARRAY_TYPE) |
4360 | { |
4361 | ratype = rtype; |
4362 | rtype = TREE_TYPE (ratype); |
4363 | } |
4364 | } |
4365 | |
4366 | auto_vec<vec<tree> > vec_oprnds; |
4367 | auto_vec<unsigned> vec_oprnds_i; |
4368 | vec_oprnds_i.safe_grow_cleared (len: nargs, exact: true); |
4369 | if (slp_node) |
4370 | { |
4371 | vec_oprnds.reserve_exact (nelems: nargs); |
4372 | vect_get_slp_defs (vinfo, slp_node, &vec_oprnds); |
4373 | } |
4374 | else |
4375 | vec_oprnds.safe_grow_cleared (len: nargs, exact: true); |
4376 | for (j = 0; j < ncopies; ++j) |
4377 | { |
4378 | poly_uint64 callee_nelements; |
4379 | poly_uint64 caller_nelements; |
4380 | /* Build argument list for the vectorized call. */ |
4381 | if (j == 0) |
4382 | vargs.create (nelems: nargs); |
4383 | else |
4384 | vargs.truncate (size: 0); |
4385 | |
4386 | for (i = 0; i < nargs; i++) |
4387 | { |
4388 | unsigned int k, l, m, o; |
4389 | tree atype; |
4390 | op = gimple_call_arg (gs: stmt, index: i + masked_call_offset); |
4391 | switch (bestn->simdclone->args[i].arg_type) |
4392 | { |
4393 | case SIMD_CLONE_ARG_TYPE_VECTOR: |
4394 | atype = bestn->simdclone->args[i].vector_type; |
4395 | caller_nelements = TYPE_VECTOR_SUBPARTS (node: arginfo[i].vectype); |
4396 | callee_nelements = TYPE_VECTOR_SUBPARTS (node: atype); |
4397 | o = vector_unroll_factor (nunits, callee_nelements); |
4398 | for (m = j * o; m < (j + 1) * o; m++) |
4399 | { |
4400 | if (known_lt (callee_nelements, caller_nelements)) |
4401 | { |
4402 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); |
4403 | if (!constant_multiple_p (a: caller_nelements, |
4404 | b: callee_nelements, multiple: &k)) |
4405 | gcc_unreachable (); |
4406 | |
4407 | gcc_assert ((k & (k - 1)) == 0); |
4408 | if (m == 0) |
4409 | { |
4410 | if (!slp_node) |
4411 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, |
4412 | ncopies: ncopies * o / k, op, |
4413 | vec_oprnds: &vec_oprnds[i]); |
4414 | vec_oprnds_i[i] = 0; |
4415 | vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++]; |
4416 | } |
4417 | else |
4418 | { |
4419 | vec_oprnd0 = arginfo[i].op; |
4420 | if ((m & (k - 1)) == 0) |
4421 | vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++]; |
4422 | } |
4423 | arginfo[i].op = vec_oprnd0; |
4424 | vec_oprnd0 |
4425 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, |
4426 | bitsize_int (prec), |
4427 | bitsize_int ((m & (k - 1)) * prec)); |
4428 | gassign *new_stmt |
4429 | = gimple_build_assign (make_ssa_name (var: atype), |
4430 | vec_oprnd0); |
4431 | vect_finish_stmt_generation (vinfo, stmt_info, |
4432 | vec_stmt: new_stmt, gsi); |
4433 | vargs.safe_push (obj: gimple_assign_lhs (gs: new_stmt)); |
4434 | } |
4435 | else |
4436 | { |
4437 | if (!constant_multiple_p (a: callee_nelements, |
4438 | b: caller_nelements, multiple: &k)) |
4439 | gcc_unreachable (); |
4440 | gcc_assert ((k & (k - 1)) == 0); |
4441 | vec<constructor_elt, va_gc> *ctor_elts; |
4442 | if (k != 1) |
4443 | vec_alloc (v&: ctor_elts, nelems: k); |
4444 | else |
4445 | ctor_elts = NULL; |
4446 | for (l = 0; l < k; l++) |
4447 | { |
4448 | if (m == 0 && l == 0) |
4449 | { |
4450 | if (!slp_node) |
4451 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, |
4452 | ncopies: k * o * ncopies, |
4453 | op, |
4454 | vec_oprnds: &vec_oprnds[i]); |
4455 | vec_oprnds_i[i] = 0; |
4456 | vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++]; |
4457 | } |
4458 | else |
4459 | vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++]; |
4460 | arginfo[i].op = vec_oprnd0; |
4461 | if (k == 1) |
4462 | break; |
4463 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, |
4464 | vec_oprnd0); |
4465 | } |
4466 | if (k == 1) |
4467 | if (!useless_type_conversion_p (TREE_TYPE (vec_oprnd0), |
4468 | atype)) |
4469 | { |
4470 | vec_oprnd0 = build1 (VIEW_CONVERT_EXPR, atype, |
4471 | vec_oprnd0); |
4472 | gassign *new_stmt |
4473 | = gimple_build_assign (make_ssa_name (var: atype), |
4474 | vec_oprnd0); |
4475 | vect_finish_stmt_generation (vinfo, stmt_info, |
4476 | vec_stmt: new_stmt, gsi); |
4477 | vargs.safe_push (obj: gimple_get_lhs (new_stmt)); |
4478 | } |
4479 | else |
4480 | vargs.safe_push (obj: vec_oprnd0); |
4481 | else |
4482 | { |
4483 | vec_oprnd0 = build_constructor (atype, ctor_elts); |
4484 | gassign *new_stmt |
4485 | = gimple_build_assign (make_ssa_name (var: atype), |
4486 | vec_oprnd0); |
4487 | vect_finish_stmt_generation (vinfo, stmt_info, |
4488 | vec_stmt: new_stmt, gsi); |
4489 | vargs.safe_push (obj: gimple_assign_lhs (gs: new_stmt)); |
4490 | } |
4491 | } |
4492 | } |
4493 | break; |
4494 | case SIMD_CLONE_ARG_TYPE_MASK: |
4495 | if (bestn->simdclone->mask_mode == VOIDmode) |
4496 | { |
4497 | atype = bestn->simdclone->args[i].vector_type; |
4498 | tree elt_type = TREE_TYPE (atype); |
4499 | tree one = fold_convert (elt_type, integer_one_node); |
4500 | tree zero = fold_convert (elt_type, integer_zero_node); |
4501 | callee_nelements = TYPE_VECTOR_SUBPARTS (node: atype); |
4502 | caller_nelements = TYPE_VECTOR_SUBPARTS (node: arginfo[i].vectype); |
4503 | o = vector_unroll_factor (nunits, callee_nelements); |
4504 | for (m = j * o; m < (j + 1) * o; m++) |
4505 | { |
4506 | if (maybe_lt (a: callee_nelements, b: caller_nelements)) |
4507 | { |
4508 | /* The mask type has fewer elements than simdlen. */ |
4509 | |
4510 | /* FORNOW */ |
4511 | gcc_unreachable (); |
4512 | } |
4513 | else if (known_eq (callee_nelements, caller_nelements)) |
4514 | { |
4515 | /* The SIMD clone function has the same number of |
4516 | elements as the current function. */ |
4517 | if (m == 0) |
4518 | { |
4519 | if (!slp_node) |
4520 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, |
4521 | ncopies: o * ncopies, |
4522 | op, |
4523 | vec_oprnds: &vec_oprnds[i]); |
4524 | vec_oprnds_i[i] = 0; |
4525 | } |
4526 | vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++]; |
4527 | if (loop_vinfo |
4528 | && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) |
4529 | { |
4530 | vec_loop_masks *loop_masks |
4531 | = &LOOP_VINFO_MASKS (loop_vinfo); |
4532 | tree loop_mask |
4533 | = vect_get_loop_mask (loop_vinfo, gsi, |
4534 | loop_masks, ncopies, |
4535 | vectype, j); |
4536 | vec_oprnd0 |
4537 | = prepare_vec_mask (loop_vinfo, |
4538 | TREE_TYPE (loop_mask), |
4539 | loop_mask, vec_mask: vec_oprnd0, |
4540 | gsi); |
4541 | loop_vinfo->vec_cond_masked_set.add (k: { vec_oprnd0, |
4542 | loop_mask }); |
4543 | |
4544 | } |
4545 | vec_oprnd0 |
4546 | = build3 (VEC_COND_EXPR, atype, vec_oprnd0, |
4547 | build_vector_from_val (atype, one), |
4548 | build_vector_from_val (atype, zero)); |
4549 | gassign *new_stmt |
4550 | = gimple_build_assign (make_ssa_name (var: atype), |
4551 | vec_oprnd0); |
4552 | vect_finish_stmt_generation (vinfo, stmt_info, |
4553 | vec_stmt: new_stmt, gsi); |
4554 | vargs.safe_push (obj: gimple_assign_lhs (gs: new_stmt)); |
4555 | } |
4556 | else |
4557 | { |
4558 | /* The mask type has more elements than simdlen. */ |
4559 | |
4560 | /* FORNOW */ |
4561 | gcc_unreachable (); |
4562 | } |
4563 | } |
4564 | } |
4565 | else if (SCALAR_INT_MODE_P (bestn->simdclone->mask_mode)) |
4566 | { |
4567 | atype = bestn->simdclone->args[i].vector_type; |
4568 | /* Guess the number of lanes represented by atype. */ |
4569 | poly_uint64 atype_subparts |
4570 | = exact_div (a: bestn->simdclone->simdlen, |
4571 | b: num_mask_args); |
4572 | o = vector_unroll_factor (nunits, atype_subparts); |
4573 | for (m = j * o; m < (j + 1) * o; m++) |
4574 | { |
4575 | if (m == 0) |
4576 | { |
4577 | if (!slp_node) |
4578 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, |
4579 | ncopies: o * ncopies, |
4580 | op, |
4581 | vec_oprnds: &vec_oprnds[i]); |
4582 | vec_oprnds_i[i] = 0; |
4583 | } |
4584 | if (maybe_lt (a: atype_subparts, |
4585 | b: TYPE_VECTOR_SUBPARTS (node: arginfo[i].vectype))) |
4586 | { |
4587 | /* The mask argument has fewer elements than the |
4588 | input vector. */ |
4589 | /* FORNOW */ |
4590 | gcc_unreachable (); |
4591 | } |
4592 | else if (known_eq (atype_subparts, |
4593 | TYPE_VECTOR_SUBPARTS (arginfo[i].vectype))) |
4594 | { |
4595 | /* The vector mask argument matches the input |
4596 | in the number of lanes, but not necessarily |
4597 | in the mode. */ |
4598 | vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++]; |
4599 | tree st = lang_hooks.types.type_for_mode |
4600 | (TYPE_MODE (TREE_TYPE (vec_oprnd0)), 1); |
4601 | vec_oprnd0 = build1 (VIEW_CONVERT_EXPR, st, |
4602 | vec_oprnd0); |
4603 | gassign *new_stmt |
4604 | = gimple_build_assign (make_ssa_name (var: st), |
4605 | vec_oprnd0); |
4606 | vect_finish_stmt_generation (vinfo, stmt_info, |
4607 | vec_stmt: new_stmt, gsi); |
4608 | if (!types_compatible_p (type1: atype, type2: st)) |
4609 | { |
4610 | new_stmt |
4611 | = gimple_build_assign (make_ssa_name (var: atype), |
4612 | NOP_EXPR, |
4613 | gimple_assign_lhs |
4614 | (gs: new_stmt)); |
4615 | vect_finish_stmt_generation (vinfo, stmt_info, |
4616 | vec_stmt: new_stmt, gsi); |
4617 | } |
4618 | vargs.safe_push (obj: gimple_assign_lhs (gs: new_stmt)); |
4619 | } |
4620 | else |
4621 | { |
4622 | /* The mask argument has more elements than the |
4623 | input vector. */ |
4624 | /* FORNOW */ |
4625 | gcc_unreachable (); |
4626 | } |
4627 | } |
4628 | } |
4629 | else |
4630 | gcc_unreachable (); |
4631 | break; |
4632 | case SIMD_CLONE_ARG_TYPE_UNIFORM: |
4633 | vargs.safe_push (obj: op); |
4634 | break; |
4635 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: |
4636 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
4637 | if (j == 0) |
4638 | { |
4639 | gimple_seq stmts; |
4640 | arginfo[i].op |
4641 | = force_gimple_operand (unshare_expr (arginfo[i].op), |
4642 | &stmts, true, NULL_TREE); |
4643 | if (stmts != NULL) |
4644 | { |
4645 | basic_block new_bb; |
4646 | edge pe = loop_preheader_edge (loop); |
4647 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); |
4648 | gcc_assert (!new_bb); |
4649 | } |
4650 | if (arginfo[i].simd_lane_linear) |
4651 | { |
4652 | vargs.safe_push (obj: arginfo[i].op); |
4653 | break; |
4654 | } |
4655 | tree phi_res = copy_ssa_name (var: op); |
4656 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
4657 | add_phi_arg (new_phi, arginfo[i].op, |
4658 | loop_preheader_edge (loop), UNKNOWN_LOCATION); |
4659 | enum tree_code code |
4660 | = POINTER_TYPE_P (TREE_TYPE (op)) |
4661 | ? POINTER_PLUS_EXPR : PLUS_EXPR; |
4662 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) |
4663 | ? sizetype : TREE_TYPE (op); |
4664 | poly_widest_int cst |
4665 | = wi::mul (a: bestn->simdclone->args[i].linear_step, |
4666 | b: ncopies * nunits); |
4667 | tree tcst = wide_int_to_tree (type, cst); |
4668 | tree phi_arg = copy_ssa_name (var: op); |
4669 | gassign *new_stmt |
4670 | = gimple_build_assign (phi_arg, code, phi_res, tcst); |
4671 | gimple_stmt_iterator si = gsi_after_labels (bb: loop->header); |
4672 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); |
4673 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
4674 | UNKNOWN_LOCATION); |
4675 | arginfo[i].op = phi_res; |
4676 | vargs.safe_push (obj: phi_res); |
4677 | } |
4678 | else |
4679 | { |
4680 | enum tree_code code |
4681 | = POINTER_TYPE_P (TREE_TYPE (op)) |
4682 | ? POINTER_PLUS_EXPR : PLUS_EXPR; |
4683 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) |
4684 | ? sizetype : TREE_TYPE (op); |
4685 | poly_widest_int cst |
4686 | = wi::mul (a: bestn->simdclone->args[i].linear_step, |
4687 | b: j * nunits); |
4688 | tree tcst = wide_int_to_tree (type, cst); |
4689 | new_temp = make_ssa_name (TREE_TYPE (op)); |
4690 | gassign *new_stmt |
4691 | = gimple_build_assign (new_temp, code, |
4692 | arginfo[i].op, tcst); |
4693 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
4694 | vargs.safe_push (obj: new_temp); |
4695 | } |
4696 | break; |
4697 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
4698 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: |
4699 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
4700 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
4701 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: |
4702 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: |
4703 | default: |
4704 | gcc_unreachable (); |
4705 | } |
4706 | } |
4707 | |
4708 | if (masked_call_offset == 0 |
4709 | && bestn->simdclone->inbranch |
4710 | && bestn->simdclone->nargs > nargs) |
4711 | { |
4712 | unsigned long m, o; |
4713 | size_t mask_i = bestn->simdclone->nargs - 1; |
4714 | tree mask; |
4715 | gcc_assert (bestn->simdclone->args[mask_i].arg_type == |
4716 | SIMD_CLONE_ARG_TYPE_MASK); |
4717 | |
4718 | tree masktype = bestn->simdclone->args[mask_i].vector_type; |
4719 | callee_nelements = TYPE_VECTOR_SUBPARTS (node: masktype); |
4720 | o = vector_unroll_factor (nunits, callee_nelements); |
4721 | for (m = j * o; m < (j + 1) * o; m++) |
4722 | { |
4723 | if (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) |
4724 | { |
4725 | vec_loop_masks *loop_masks = &LOOP_VINFO_MASKS (loop_vinfo); |
4726 | mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks, |
4727 | ncopies, vectype, j); |
4728 | } |
4729 | else |
4730 | mask = vect_build_all_ones_mask (vinfo, stmt_info, masktype); |
4731 | |
4732 | gassign *new_stmt; |
4733 | if (SCALAR_INT_MODE_P (bestn->simdclone->mask_mode)) |
4734 | { |
4735 | /* This means we are dealing with integer mask modes. |
4736 | First convert to an integer type with the same size as |
4737 | the current vector type. */ |
4738 | unsigned HOST_WIDE_INT intermediate_size |
4739 | = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (mask))); |
4740 | tree mid_int_type = |
4741 | build_nonstandard_integer_type (intermediate_size, 1); |
4742 | mask = build1 (VIEW_CONVERT_EXPR, mid_int_type, mask); |
4743 | new_stmt |
4744 | = gimple_build_assign (make_ssa_name (var: mid_int_type), |
4745 | mask); |
4746 | gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT); |
4747 | /* Then zero-extend to the mask mode. */ |
4748 | mask = fold_build1 (NOP_EXPR, masktype, |
4749 | gimple_get_lhs (new_stmt)); |
4750 | } |
4751 | else if (bestn->simdclone->mask_mode == VOIDmode) |
4752 | { |
4753 | tree one = fold_convert (TREE_TYPE (masktype), |
4754 | integer_one_node); |
4755 | tree zero = fold_convert (TREE_TYPE (masktype), |
4756 | integer_zero_node); |
4757 | mask = build3 (VEC_COND_EXPR, masktype, mask, |
4758 | build_vector_from_val (masktype, one), |
4759 | build_vector_from_val (masktype, zero)); |
4760 | } |
4761 | else |
4762 | gcc_unreachable (); |
4763 | |
4764 | new_stmt = gimple_build_assign (make_ssa_name (var: masktype), mask); |
4765 | vect_finish_stmt_generation (vinfo, stmt_info, |
4766 | vec_stmt: new_stmt, gsi); |
4767 | mask = gimple_assign_lhs (gs: new_stmt); |
4768 | vargs.safe_push (obj: mask); |
4769 | } |
4770 | } |
4771 | |
4772 | gcall *new_call = gimple_build_call_vec (fndecl, vargs); |
4773 | if (vec_dest) |
4774 | { |
4775 | gcc_assert (ratype |
4776 | || known_eq (TYPE_VECTOR_SUBPARTS (rtype), nunits)); |
4777 | if (ratype) |
4778 | new_temp = create_tmp_var (ratype); |
4779 | else if (useless_type_conversion_p (vectype, rtype)) |
4780 | new_temp = make_ssa_name (var: vec_dest, stmt: new_call); |
4781 | else |
4782 | new_temp = make_ssa_name (var: rtype, stmt: new_call); |
4783 | gimple_call_set_lhs (gs: new_call, lhs: new_temp); |
4784 | } |
4785 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_call, gsi); |
4786 | gimple *new_stmt = new_call; |
4787 | |
4788 | if (vec_dest) |
4789 | { |
4790 | if (!multiple_p (a: TYPE_VECTOR_SUBPARTS (node: vectype), b: nunits)) |
4791 | { |
4792 | unsigned int k, l; |
4793 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); |
4794 | poly_uint64 bytes = GET_MODE_SIZE (TYPE_MODE (vectype)); |
4795 | k = vector_unroll_factor (nunits, |
4796 | TYPE_VECTOR_SUBPARTS (vectype)); |
4797 | gcc_assert ((k & (k - 1)) == 0); |
4798 | for (l = 0; l < k; l++) |
4799 | { |
4800 | tree t; |
4801 | if (ratype) |
4802 | { |
4803 | t = build_fold_addr_expr (new_temp); |
4804 | t = build2 (MEM_REF, vectype, t, |
4805 | build_int_cst (TREE_TYPE (t), l * bytes)); |
4806 | } |
4807 | else |
4808 | t = build3 (BIT_FIELD_REF, vectype, new_temp, |
4809 | bitsize_int (prec), bitsize_int (l * prec)); |
4810 | new_stmt = gimple_build_assign (make_ssa_name (var: vectype), t); |
4811 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
4812 | |
4813 | if (j == 0 && l == 0) |
4814 | *vec_stmt = new_stmt; |
4815 | if (slp_node) |
4816 | SLP_TREE_VEC_DEFS (slp_node) |
4817 | .quick_push (obj: gimple_assign_lhs (gs: new_stmt)); |
4818 | else |
4819 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
4820 | } |
4821 | |
4822 | if (ratype) |
4823 | vect_clobber_variable (vinfo, stmt_info, gsi, var: new_temp); |
4824 | continue; |
4825 | } |
4826 | else if (!multiple_p (a: nunits, b: TYPE_VECTOR_SUBPARTS (node: vectype))) |
4827 | { |
4828 | unsigned int k; |
4829 | if (!constant_multiple_p (a: TYPE_VECTOR_SUBPARTS (node: vectype), |
4830 | b: TYPE_VECTOR_SUBPARTS (node: rtype), multiple: &k)) |
4831 | gcc_unreachable (); |
4832 | gcc_assert ((k & (k - 1)) == 0); |
4833 | if ((j & (k - 1)) == 0) |
4834 | vec_alloc (v&: ret_ctor_elts, nelems: k); |
4835 | if (ratype) |
4836 | { |
4837 | unsigned int m, o; |
4838 | o = vector_unroll_factor (nunits, |
4839 | TYPE_VECTOR_SUBPARTS (rtype)); |
4840 | for (m = 0; m < o; m++) |
4841 | { |
4842 | tree tem = build4 (ARRAY_REF, rtype, new_temp, |
4843 | size_int (m), NULL_TREE, NULL_TREE); |
4844 | new_stmt = gimple_build_assign (make_ssa_name (var: rtype), |
4845 | tem); |
4846 | vect_finish_stmt_generation (vinfo, stmt_info, |
4847 | vec_stmt: new_stmt, gsi); |
4848 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, |
4849 | gimple_assign_lhs (new_stmt)); |
4850 | } |
4851 | vect_clobber_variable (vinfo, stmt_info, gsi, var: new_temp); |
4852 | } |
4853 | else |
4854 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); |
4855 | if ((j & (k - 1)) != k - 1) |
4856 | continue; |
4857 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); |
4858 | new_stmt |
4859 | = gimple_build_assign (make_ssa_name (var: vec_dest), vec_oprnd0); |
4860 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
4861 | |
4862 | if ((unsigned) j == k - 1) |
4863 | *vec_stmt = new_stmt; |
4864 | if (slp_node) |
4865 | SLP_TREE_VEC_DEFS (slp_node) |
4866 | .quick_push (obj: gimple_assign_lhs (gs: new_stmt)); |
4867 | else |
4868 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
4869 | continue; |
4870 | } |
4871 | else if (ratype) |
4872 | { |
4873 | tree t = build_fold_addr_expr (new_temp); |
4874 | t = build2 (MEM_REF, vectype, t, |
4875 | build_int_cst (TREE_TYPE (t), 0)); |
4876 | new_stmt = gimple_build_assign (make_ssa_name (var: vec_dest), t); |
4877 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
4878 | vect_clobber_variable (vinfo, stmt_info, gsi, var: new_temp); |
4879 | } |
4880 | else if (!useless_type_conversion_p (vectype, rtype)) |
4881 | { |
4882 | vec_oprnd0 = build1 (VIEW_CONVERT_EXPR, vectype, new_temp); |
4883 | new_stmt |
4884 | = gimple_build_assign (make_ssa_name (var: vec_dest), vec_oprnd0); |
4885 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
4886 | } |
4887 | } |
4888 | |
4889 | if (j == 0) |
4890 | *vec_stmt = new_stmt; |
4891 | if (slp_node) |
4892 | SLP_TREE_VEC_DEFS (slp_node).quick_push (obj: gimple_get_lhs (new_stmt)); |
4893 | else |
4894 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
4895 | } |
4896 | |
4897 | for (i = 0; i < nargs; ++i) |
4898 | { |
4899 | vec<tree> oprndsi = vec_oprnds[i]; |
4900 | oprndsi.release (); |
4901 | } |
4902 | vargs.release (); |
4903 | |
4904 | /* Mark the clone as no longer being a candidate for GC. */ |
4905 | bestn->gc_candidate = false; |
4906 | |
4907 | /* The call in STMT might prevent it from being removed in dce. |
4908 | We however cannot remove it here, due to the way the ssa name |
4909 | it defines is mapped to the new definition. So just replace |
4910 | rhs of the statement with something harmless. */ |
4911 | |
4912 | if (slp_node) |
4913 | return true; |
4914 | |
4915 | gimple *new_stmt; |
4916 | if (scalar_dest) |
4917 | { |
4918 | type = TREE_TYPE (scalar_dest); |
4919 | lhs = gimple_call_lhs (gs: vect_orig_stmt (stmt_info)->stmt); |
4920 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
4921 | } |
4922 | else |
4923 | new_stmt = gimple_build_nop (); |
4924 | vinfo->replace_stmt (gsi, vect_orig_stmt (stmt_info), new_stmt); |
4925 | unlink_stmt_vdef (stmt); |
4926 | |
4927 | return true; |
4928 | } |
4929 | |
4930 | |
4931 | /* Function vect_gen_widened_results_half |
4932 | |
4933 | Create a vector stmt whose code, type, number of arguments, and result |
4934 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
4935 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at GSI. |
4936 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
4937 | needs to be created (DECL is a function-decl of a target-builtin). |
4938 | STMT_INFO is the original scalar stmt that we are vectorizing. */ |
4939 | |
4940 | static gimple * |
4941 | vect_gen_widened_results_half (vec_info *vinfo, code_helper ch, |
4942 | tree vec_oprnd0, tree vec_oprnd1, int op_type, |
4943 | tree vec_dest, gimple_stmt_iterator *gsi, |
4944 | stmt_vec_info stmt_info) |
4945 | { |
4946 | gimple *new_stmt; |
4947 | tree new_temp; |
4948 | |
4949 | /* Generate half of the widened result: */ |
4950 | if (op_type != binary_op) |
4951 | vec_oprnd1 = NULL; |
4952 | new_stmt = vect_gimple_build (vec_dest, ch, vec_oprnd0, vec_oprnd1); |
4953 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
4954 | gimple_set_lhs (new_stmt, new_temp); |
4955 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
4956 | |
4957 | return new_stmt; |
4958 | } |
4959 | |
4960 | |
4961 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. |
4962 | For multi-step conversions store the resulting vectors and call the function |
4963 | recursively. When NARROW_SRC_P is true, there's still a conversion after |
4964 | narrowing, don't store the vectors in the SLP_NODE or in vector info of |
4965 | the scalar statement(or in STMT_VINFO_RELATED_STMT chain). */ |
4966 | |
4967 | static void |
4968 | vect_create_vectorized_demotion_stmts (vec_info *vinfo, vec<tree> *vec_oprnds, |
4969 | int multi_step_cvt, |
4970 | stmt_vec_info stmt_info, |
4971 | vec<tree> &vec_dsts, |
4972 | gimple_stmt_iterator *gsi, |
4973 | slp_tree slp_node, code_helper code, |
4974 | bool narrow_src_p) |
4975 | { |
4976 | unsigned int i; |
4977 | tree vop0, vop1, new_tmp, vec_dest; |
4978 | |
4979 | vec_dest = vec_dsts.pop (); |
4980 | |
4981 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4982 | { |
4983 | /* Create demotion operation. */ |
4984 | vop0 = (*vec_oprnds)[i]; |
4985 | vop1 = (*vec_oprnds)[i + 1]; |
4986 | gimple *new_stmt = vect_gimple_build (vec_dest, code, vop0, vop1); |
4987 | new_tmp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
4988 | gimple_set_lhs (new_stmt, new_tmp); |
4989 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
4990 | if (multi_step_cvt || narrow_src_p) |
4991 | /* Store the resulting vector for next recursive call, |
4992 | or return the resulting vector_tmp for NARROW FLOAT_EXPR. */ |
4993 | (*vec_oprnds)[i/2] = new_tmp; |
4994 | else |
4995 | { |
4996 | /* This is the last step of the conversion sequence. Store the |
4997 | vectors in SLP_NODE or in vector info of the scalar statement |
4998 | (or in STMT_VINFO_RELATED_STMT chain). */ |
4999 | if (slp_node) |
5000 | slp_node->push_vec_def (def: new_stmt); |
5001 | else |
5002 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
5003 | } |
5004 | } |
5005 | |
5006 | /* For multi-step demotion operations we first generate demotion operations |
5007 | from the source type to the intermediate types, and then combine the |
5008 | results (stored in VEC_OPRNDS) in demotion operation to the destination |
5009 | type. */ |
5010 | if (multi_step_cvt) |
5011 | { |
5012 | /* At each level of recursion we have half of the operands we had at the |
5013 | previous level. */ |
5014 | vec_oprnds->truncate (size: (i+1)/2); |
5015 | vect_create_vectorized_demotion_stmts (vinfo, vec_oprnds, |
5016 | multi_step_cvt: multi_step_cvt - 1, |
5017 | stmt_info, vec_dsts, gsi, |
5018 | slp_node, code: VEC_PACK_TRUNC_EXPR, |
5019 | narrow_src_p); |
5020 | } |
5021 | |
5022 | vec_dsts.quick_push (obj: vec_dest); |
5023 | } |
5024 | |
5025 | |
5026 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 |
5027 | and VEC_OPRNDS1, for a binary operation associated with scalar statement |
5028 | STMT_INFO. For multi-step conversions store the resulting vectors and |
5029 | call the function recursively. */ |
5030 | |
5031 | static void |
5032 | vect_create_vectorized_promotion_stmts (vec_info *vinfo, |
5033 | vec<tree> *vec_oprnds0, |
5034 | vec<tree> *vec_oprnds1, |
5035 | stmt_vec_info stmt_info, tree vec_dest, |
5036 | gimple_stmt_iterator *gsi, |
5037 | code_helper ch1, |
5038 | code_helper ch2, int op_type) |
5039 | { |
5040 | int i; |
5041 | tree vop0, vop1, new_tmp1, new_tmp2; |
5042 | gimple *new_stmt1, *new_stmt2; |
5043 | vec<tree> vec_tmp = vNULL; |
5044 | |
5045 | vec_tmp.create (nelems: vec_oprnds0->length () * 2); |
5046 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) |
5047 | { |
5048 | if (op_type == binary_op) |
5049 | vop1 = (*vec_oprnds1)[i]; |
5050 | else |
5051 | vop1 = NULL_TREE; |
5052 | |
5053 | /* Generate the two halves of promotion operation. */ |
5054 | new_stmt1 = vect_gen_widened_results_half (vinfo, ch: ch1, vec_oprnd0: vop0, vec_oprnd1: vop1, |
5055 | op_type, vec_dest, gsi, |
5056 | stmt_info); |
5057 | new_stmt2 = vect_gen_widened_results_half (vinfo, ch: ch2, vec_oprnd0: vop0, vec_oprnd1: vop1, |
5058 | op_type, vec_dest, gsi, |
5059 | stmt_info); |
5060 | if (is_gimple_call (gs: new_stmt1)) |
5061 | { |
5062 | new_tmp1 = gimple_call_lhs (gs: new_stmt1); |
5063 | new_tmp2 = gimple_call_lhs (gs: new_stmt2); |
5064 | } |
5065 | else |
5066 | { |
5067 | new_tmp1 = gimple_assign_lhs (gs: new_stmt1); |
5068 | new_tmp2 = gimple_assign_lhs (gs: new_stmt2); |
5069 | } |
5070 | |
5071 | /* Store the results for the next step. */ |
5072 | vec_tmp.quick_push (obj: new_tmp1); |
5073 | vec_tmp.quick_push (obj: new_tmp2); |
5074 | } |
5075 | |
5076 | vec_oprnds0->release (); |
5077 | *vec_oprnds0 = vec_tmp; |
5078 | } |
5079 | |
5080 | /* Create vectorized promotion stmts for widening stmts using only half the |
5081 | potential vector size for input. */ |
5082 | static void |
5083 | vect_create_half_widening_stmts (vec_info *vinfo, |
5084 | vec<tree> *vec_oprnds0, |
5085 | vec<tree> *vec_oprnds1, |
5086 | stmt_vec_info stmt_info, tree vec_dest, |
5087 | gimple_stmt_iterator *gsi, |
5088 | code_helper code1, |
5089 | int op_type) |
5090 | { |
5091 | int i; |
5092 | tree vop0, vop1; |
5093 | gimple *new_stmt1; |
5094 | gimple *new_stmt2; |
5095 | gimple *new_stmt3; |
5096 | vec<tree> vec_tmp = vNULL; |
5097 | |
5098 | vec_tmp.create (nelems: vec_oprnds0->length ()); |
5099 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) |
5100 | { |
5101 | tree new_tmp1, new_tmp2, new_tmp3, out_type; |
5102 | |
5103 | gcc_assert (op_type == binary_op); |
5104 | vop1 = (*vec_oprnds1)[i]; |
5105 | |
5106 | /* Widen the first vector input. */ |
5107 | out_type = TREE_TYPE (vec_dest); |
5108 | new_tmp1 = make_ssa_name (var: out_type); |
5109 | new_stmt1 = gimple_build_assign (new_tmp1, NOP_EXPR, vop0); |
5110 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt1, gsi); |
5111 | if (VECTOR_TYPE_P (TREE_TYPE (vop1))) |
5112 | { |
5113 | /* Widen the second vector input. */ |
5114 | new_tmp2 = make_ssa_name (var: out_type); |
5115 | new_stmt2 = gimple_build_assign (new_tmp2, NOP_EXPR, vop1); |
5116 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt2, gsi); |
5117 | /* Perform the operation. With both vector inputs widened. */ |
5118 | new_stmt3 = vect_gimple_build (vec_dest, code1, new_tmp1, new_tmp2); |
5119 | } |
5120 | else |
5121 | { |
5122 | /* Perform the operation. With the single vector input widened. */ |
5123 | new_stmt3 = vect_gimple_build (vec_dest, code1, new_tmp1, vop1); |
5124 | } |
5125 | |
5126 | new_tmp3 = make_ssa_name (var: vec_dest, stmt: new_stmt3); |
5127 | gimple_assign_set_lhs (gs: new_stmt3, lhs: new_tmp3); |
5128 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt3, gsi); |
5129 | |
5130 | /* Store the results for the next step. */ |
5131 | vec_tmp.quick_push (obj: new_tmp3); |
5132 | } |
5133 | |
5134 | vec_oprnds0->release (); |
5135 | *vec_oprnds0 = vec_tmp; |
5136 | } |
5137 | |
5138 | |
5139 | /* Check if STMT_INFO performs a conversion operation that can be vectorized. |
5140 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
5141 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
5142 | Return true if STMT_INFO is vectorizable in this way. */ |
5143 | |
5144 | static bool |
5145 | vectorizable_conversion (vec_info *vinfo, |
5146 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
5147 | gimple **vec_stmt, slp_tree slp_node, |
5148 | stmt_vector_for_cost *cost_vec) |
5149 | { |
5150 | tree vec_dest, cvt_op = NULL_TREE; |
5151 | tree scalar_dest; |
5152 | tree op0, op1 = NULL_TREE; |
5153 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
5154 | tree_code tc1, tc2; |
5155 | code_helper code, code1, code2; |
5156 | code_helper codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
5157 | tree new_temp; |
5158 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
5159 | int ndts = 2; |
5160 | poly_uint64 nunits_in; |
5161 | poly_uint64 nunits_out; |
5162 | tree vectype_out, vectype_in; |
5163 | int ncopies, i; |
5164 | tree lhs_type, rhs_type; |
5165 | /* For conversions between floating point and integer, there're 2 NARROW |
5166 | cases. NARROW_SRC is for FLOAT_EXPR, means |
5167 | integer --DEMOTION--> integer --FLOAT_EXPR--> floating point. |
5168 | This is safe when the range of the source integer can fit into the lower |
5169 | precision. NARROW_DST is for FIX_TRUNC_EXPR, means |
5170 | floating point --FIX_TRUNC_EXPR--> integer --DEMOTION--> INTEGER. |
5171 | For other conversions, when there's narrowing, NARROW_DST is used as |
5172 | default. */ |
5173 | enum { NARROW_SRC, NARROW_DST, NONE, WIDEN } modifier; |
5174 | vec<tree> vec_oprnds0 = vNULL; |
5175 | vec<tree> vec_oprnds1 = vNULL; |
5176 | tree vop0; |
5177 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
5178 | int multi_step_cvt = 0; |
5179 | vec<tree> interm_types = vNULL; |
5180 | tree intermediate_type, cvt_type = NULL_TREE; |
5181 | int op_type; |
5182 | unsigned short fltsz; |
5183 | |
5184 | /* Is STMT a vectorizable conversion? */ |
5185 | |
5186 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
5187 | return false; |
5188 | |
5189 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5190 | && ! vec_stmt) |
5191 | return false; |
5192 | |
5193 | gimple* stmt = stmt_info->stmt; |
5194 | if (!(is_gimple_assign (gs: stmt) || is_gimple_call (gs: stmt))) |
5195 | return false; |
5196 | |
5197 | if (gimple_get_lhs (stmt) == NULL_TREE |
5198 | || TREE_CODE (gimple_get_lhs (stmt)) != SSA_NAME) |
5199 | return false; |
5200 | |
5201 | if (TREE_CODE (gimple_get_lhs (stmt)) != SSA_NAME) |
5202 | return false; |
5203 | |
5204 | if (is_gimple_assign (gs: stmt)) |
5205 | { |
5206 | code = gimple_assign_rhs_code (gs: stmt); |
5207 | op_type = TREE_CODE_LENGTH ((tree_code) code); |
5208 | } |
5209 | else if (gimple_call_internal_p (gs: stmt)) |
5210 | { |
5211 | code = gimple_call_internal_fn (gs: stmt); |
5212 | op_type = gimple_call_num_args (gs: stmt); |
5213 | } |
5214 | else |
5215 | return false; |
5216 | |
5217 | bool widen_arith = (code == WIDEN_MULT_EXPR |
5218 | || code == WIDEN_LSHIFT_EXPR |
5219 | || widening_fn_p (code)); |
5220 | |
5221 | if (!widen_arith |
5222 | && !CONVERT_EXPR_CODE_P (code) |
5223 | && code != FIX_TRUNC_EXPR |
5224 | && code != FLOAT_EXPR) |
5225 | return false; |
5226 | |
5227 | /* Check types of lhs and rhs. */ |
5228 | scalar_dest = gimple_get_lhs (stmt); |
5229 | lhs_type = TREE_TYPE (scalar_dest); |
5230 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
5231 | |
5232 | /* Check the operands of the operation. */ |
5233 | slp_tree slp_op0, slp_op1 = NULL; |
5234 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
5235 | 0, &op0, &slp_op0, &dt[0], &vectype_in)) |
5236 | { |
5237 | if (dump_enabled_p ()) |
5238 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5239 | "use not simple.\n" ); |
5240 | return false; |
5241 | } |
5242 | |
5243 | rhs_type = TREE_TYPE (op0); |
5244 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) |
5245 | && !((INTEGRAL_TYPE_P (lhs_type) |
5246 | && INTEGRAL_TYPE_P (rhs_type)) |
5247 | || (SCALAR_FLOAT_TYPE_P (lhs_type) |
5248 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) |
5249 | return false; |
5250 | |
5251 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
5252 | && ((INTEGRAL_TYPE_P (lhs_type) |
5253 | && !type_has_mode_precision_p (t: lhs_type)) |
5254 | || (INTEGRAL_TYPE_P (rhs_type) |
5255 | && !type_has_mode_precision_p (t: rhs_type)))) |
5256 | { |
5257 | if (dump_enabled_p ()) |
5258 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5259 | "type conversion to/from bit-precision unsupported." |
5260 | "\n" ); |
5261 | return false; |
5262 | } |
5263 | |
5264 | if (op_type == binary_op) |
5265 | { |
5266 | gcc_assert (code == WIDEN_MULT_EXPR |
5267 | || code == WIDEN_LSHIFT_EXPR |
5268 | || widening_fn_p (code)); |
5269 | |
5270 | op1 = is_gimple_assign (gs: stmt) ? gimple_assign_rhs2 (gs: stmt) : |
5271 | gimple_call_arg (gs: stmt, index: 0); |
5272 | tree vectype1_in; |
5273 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 1, |
5274 | &op1, &slp_op1, &dt[1], &vectype1_in)) |
5275 | { |
5276 | if (dump_enabled_p ()) |
5277 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5278 | "use not simple.\n" ); |
5279 | return false; |
5280 | } |
5281 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of |
5282 | OP1. */ |
5283 | if (!vectype_in) |
5284 | vectype_in = vectype1_in; |
5285 | } |
5286 | |
5287 | /* If op0 is an external or constant def, infer the vector type |
5288 | from the scalar type. */ |
5289 | if (!vectype_in) |
5290 | vectype_in = get_vectype_for_scalar_type (vinfo, rhs_type, slp_node); |
5291 | if (vec_stmt) |
5292 | gcc_assert (vectype_in); |
5293 | if (!vectype_in) |
5294 | { |
5295 | if (dump_enabled_p ()) |
5296 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5297 | "no vectype for scalar type %T\n" , rhs_type); |
5298 | |
5299 | return false; |
5300 | } |
5301 | |
5302 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
5303 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) |
5304 | { |
5305 | if (dump_enabled_p ()) |
5306 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5307 | "can't convert between boolean and non " |
5308 | "boolean vectors %T\n" , rhs_type); |
5309 | |
5310 | return false; |
5311 | } |
5312 | |
5313 | nunits_in = TYPE_VECTOR_SUBPARTS (node: vectype_in); |
5314 | nunits_out = TYPE_VECTOR_SUBPARTS (node: vectype_out); |
5315 | if (known_eq (nunits_out, nunits_in)) |
5316 | if (widen_arith) |
5317 | modifier = WIDEN; |
5318 | else |
5319 | modifier = NONE; |
5320 | else if (multiple_p (a: nunits_out, b: nunits_in)) |
5321 | modifier = NARROW_DST; |
5322 | else |
5323 | { |
5324 | gcc_checking_assert (multiple_p (nunits_in, nunits_out)); |
5325 | modifier = WIDEN; |
5326 | } |
5327 | |
5328 | /* Multiple types in SLP are handled by creating the appropriate number of |
5329 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
5330 | case of SLP. */ |
5331 | if (slp_node) |
5332 | ncopies = 1; |
5333 | else if (modifier == NARROW_DST) |
5334 | ncopies = vect_get_num_copies (loop_vinfo, vectype: vectype_out); |
5335 | else |
5336 | ncopies = vect_get_num_copies (loop_vinfo, vectype: vectype_in); |
5337 | |
5338 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
5339 | needs to be generated. */ |
5340 | gcc_assert (ncopies >= 1); |
5341 | |
5342 | bool found_mode = false; |
5343 | scalar_mode lhs_mode = SCALAR_TYPE_MODE (lhs_type); |
5344 | scalar_mode rhs_mode = SCALAR_TYPE_MODE (rhs_type); |
5345 | opt_scalar_mode rhs_mode_iter; |
5346 | |
5347 | /* Supportable by target? */ |
5348 | switch (modifier) |
5349 | { |
5350 | case NONE: |
5351 | if (code != FIX_TRUNC_EXPR |
5352 | && code != FLOAT_EXPR |
5353 | && !CONVERT_EXPR_CODE_P (code)) |
5354 | return false; |
5355 | gcc_assert (code.is_tree_code ()); |
5356 | if (supportable_convert_operation ((tree_code) code, vectype_out, |
5357 | vectype_in, &tc1)) |
5358 | { |
5359 | code1 = tc1; |
5360 | break; |
5361 | } |
5362 | |
5363 | /* For conversions between float and integer types try whether |
5364 | we can use intermediate signed integer types to support the |
5365 | conversion. */ |
5366 | if (GET_MODE_SIZE (mode: lhs_mode) != GET_MODE_SIZE (mode: rhs_mode) |
5367 | && (code == FLOAT_EXPR || |
5368 | (code == FIX_TRUNC_EXPR && !flag_trapping_math))) |
5369 | { |
5370 | bool demotion = GET_MODE_SIZE (mode: rhs_mode) > GET_MODE_SIZE (mode: lhs_mode); |
5371 | bool float_expr_p = code == FLOAT_EXPR; |
5372 | unsigned short target_size; |
5373 | scalar_mode intermediate_mode; |
5374 | if (demotion) |
5375 | { |
5376 | intermediate_mode = lhs_mode; |
5377 | target_size = GET_MODE_SIZE (mode: rhs_mode); |
5378 | } |
5379 | else |
5380 | { |
5381 | target_size = GET_MODE_SIZE (mode: lhs_mode); |
5382 | if (!int_mode_for_size |
5383 | (size: GET_MODE_BITSIZE (mode: rhs_mode), limit: 0).exists (mode: &intermediate_mode)) |
5384 | goto unsupported; |
5385 | } |
5386 | code1 = float_expr_p ? code : NOP_EXPR; |
5387 | codecvt1 = float_expr_p ? NOP_EXPR : code; |
5388 | opt_scalar_mode mode_iter; |
5389 | FOR_EACH_2XWIDER_MODE (mode_iter, intermediate_mode) |
5390 | { |
5391 | intermediate_mode = mode_iter.require (); |
5392 | |
5393 | if (GET_MODE_SIZE (mode: intermediate_mode) > target_size) |
5394 | break; |
5395 | |
5396 | scalar_mode cvt_mode; |
5397 | if (!int_mode_for_size |
5398 | (size: GET_MODE_BITSIZE (mode: intermediate_mode), limit: 0).exists (mode: &cvt_mode)) |
5399 | break; |
5400 | |
5401 | cvt_type = build_nonstandard_integer_type |
5402 | (GET_MODE_BITSIZE (mode: cvt_mode), 0); |
5403 | |
5404 | /* Check if the intermediate type can hold OP0's range. |
5405 | When converting from float to integer this is not necessary |
5406 | because values that do not fit the (smaller) target type are |
5407 | unspecified anyway. */ |
5408 | if (demotion && float_expr_p) |
5409 | { |
5410 | wide_int op_min_value, op_max_value; |
5411 | if (!vect_get_range_info (op0, &op_min_value, &op_max_value)) |
5412 | break; |
5413 | |
5414 | if (cvt_type == NULL_TREE |
5415 | || (wi::min_precision (x: op_max_value, sgn: SIGNED) |
5416 | > TYPE_PRECISION (cvt_type)) |
5417 | || (wi::min_precision (x: op_min_value, sgn: SIGNED) |
5418 | > TYPE_PRECISION (cvt_type))) |
5419 | continue; |
5420 | } |
5421 | |
5422 | cvt_type = get_vectype_for_scalar_type (vinfo, cvt_type, slp_node); |
5423 | /* This should only happened for SLP as long as loop vectorizer |
5424 | only supports same-sized vector. */ |
5425 | if (cvt_type == NULL_TREE |
5426 | || maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: cvt_type), b: nunits_in) |
5427 | || !supportable_convert_operation ((tree_code) code1, |
5428 | vectype_out, |
5429 | cvt_type, &tc1) |
5430 | || !supportable_convert_operation ((tree_code) codecvt1, |
5431 | cvt_type, |
5432 | vectype_in, &tc2)) |
5433 | continue; |
5434 | |
5435 | found_mode = true; |
5436 | break; |
5437 | } |
5438 | |
5439 | if (found_mode) |
5440 | { |
5441 | multi_step_cvt++; |
5442 | interm_types.safe_push (obj: cvt_type); |
5443 | cvt_type = NULL_TREE; |
5444 | code1 = tc1; |
5445 | codecvt1 = tc2; |
5446 | break; |
5447 | } |
5448 | } |
5449 | /* FALLTHRU */ |
5450 | unsupported: |
5451 | if (dump_enabled_p ()) |
5452 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5453 | "conversion not supported by target.\n" ); |
5454 | return false; |
5455 | |
5456 | case WIDEN: |
5457 | if (known_eq (nunits_in, nunits_out)) |
5458 | { |
5459 | if (!(code.is_tree_code () |
5460 | && supportable_half_widening_operation ((tree_code) code, |
5461 | vectype_out, vectype_in, |
5462 | &tc1))) |
5463 | goto unsupported; |
5464 | code1 = tc1; |
5465 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); |
5466 | break; |
5467 | } |
5468 | if (supportable_widening_operation (vinfo, code, stmt_info, |
5469 | vectype_out, vectype_in, &code1, |
5470 | &code2, &multi_step_cvt, |
5471 | &interm_types)) |
5472 | { |
5473 | /* Binary widening operation can only be supported directly by the |
5474 | architecture. */ |
5475 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); |
5476 | break; |
5477 | } |
5478 | |
5479 | if (code != FLOAT_EXPR |
5480 | || GET_MODE_SIZE (mode: lhs_mode) <= GET_MODE_SIZE (mode: rhs_mode)) |
5481 | goto unsupported; |
5482 | |
5483 | fltsz = GET_MODE_SIZE (mode: lhs_mode); |
5484 | FOR_EACH_2XWIDER_MODE (rhs_mode_iter, rhs_mode) |
5485 | { |
5486 | rhs_mode = rhs_mode_iter.require (); |
5487 | if (GET_MODE_SIZE (mode: rhs_mode) > fltsz) |
5488 | break; |
5489 | |
5490 | cvt_type |
5491 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (mode: rhs_mode), 0); |
5492 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); |
5493 | if (cvt_type == NULL_TREE) |
5494 | goto unsupported; |
5495 | |
5496 | if (GET_MODE_SIZE (mode: rhs_mode) == fltsz) |
5497 | { |
5498 | tc1 = ERROR_MARK; |
5499 | gcc_assert (code.is_tree_code ()); |
5500 | if (!supportable_convert_operation ((tree_code) code, vectype_out, |
5501 | cvt_type, &tc1)) |
5502 | goto unsupported; |
5503 | codecvt1 = tc1; |
5504 | } |
5505 | else if (!supportable_widening_operation (vinfo, code, |
5506 | stmt_info, vectype_out, |
5507 | cvt_type, &codecvt1, |
5508 | &codecvt2, &multi_step_cvt, |
5509 | &interm_types)) |
5510 | continue; |
5511 | else |
5512 | gcc_assert (multi_step_cvt == 0); |
5513 | |
5514 | if (supportable_widening_operation (vinfo, NOP_EXPR, stmt_info, |
5515 | cvt_type, |
5516 | vectype_in, &code1, |
5517 | &code2, &multi_step_cvt, |
5518 | &interm_types)) |
5519 | { |
5520 | found_mode = true; |
5521 | break; |
5522 | } |
5523 | } |
5524 | |
5525 | if (!found_mode) |
5526 | goto unsupported; |
5527 | |
5528 | if (GET_MODE_SIZE (mode: rhs_mode) == fltsz) |
5529 | codecvt2 = ERROR_MARK; |
5530 | else |
5531 | { |
5532 | multi_step_cvt++; |
5533 | interm_types.safe_push (obj: cvt_type); |
5534 | cvt_type = NULL_TREE; |
5535 | } |
5536 | break; |
5537 | |
5538 | case NARROW_DST: |
5539 | gcc_assert (op_type == unary_op); |
5540 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, |
5541 | &code1, &multi_step_cvt, |
5542 | &interm_types)) |
5543 | break; |
5544 | |
5545 | if (GET_MODE_SIZE (mode: lhs_mode) >= GET_MODE_SIZE (mode: rhs_mode)) |
5546 | goto unsupported; |
5547 | |
5548 | if (code == FIX_TRUNC_EXPR) |
5549 | { |
5550 | cvt_type |
5551 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (mode: rhs_mode), 0); |
5552 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); |
5553 | if (cvt_type == NULL_TREE) |
5554 | goto unsupported; |
5555 | if (supportable_convert_operation ((tree_code) code, cvt_type, vectype_in, |
5556 | &tc1)) |
5557 | codecvt1 = tc1; |
5558 | else |
5559 | goto unsupported; |
5560 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, |
5561 | &code1, &multi_step_cvt, |
5562 | &interm_types)) |
5563 | break; |
5564 | } |
5565 | /* If op0 can be represented with low precision integer, |
5566 | truncate it to cvt_type and the do FLOAT_EXPR. */ |
5567 | else if (code == FLOAT_EXPR) |
5568 | { |
5569 | wide_int op_min_value, op_max_value; |
5570 | if (!vect_get_range_info (op0, &op_min_value, &op_max_value)) |
5571 | goto unsupported; |
5572 | |
5573 | cvt_type |
5574 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (mode: lhs_mode), 0); |
5575 | if (cvt_type == NULL_TREE |
5576 | || (wi::min_precision (x: op_max_value, sgn: SIGNED) |
5577 | > TYPE_PRECISION (cvt_type)) |
5578 | || (wi::min_precision (x: op_min_value, sgn: SIGNED) |
5579 | > TYPE_PRECISION (cvt_type))) |
5580 | goto unsupported; |
5581 | |
5582 | cvt_type = get_same_sized_vectype (cvt_type, vectype_out); |
5583 | if (cvt_type == NULL_TREE) |
5584 | goto unsupported; |
5585 | if (!supportable_narrowing_operation (NOP_EXPR, cvt_type, vectype_in, |
5586 | &code1, &multi_step_cvt, |
5587 | &interm_types)) |
5588 | goto unsupported; |
5589 | if (supportable_convert_operation ((tree_code) code, vectype_out, |
5590 | cvt_type, &tc1)) |
5591 | { |
5592 | codecvt1 = tc1; |
5593 | modifier = NARROW_SRC; |
5594 | break; |
5595 | } |
5596 | } |
5597 | |
5598 | goto unsupported; |
5599 | |
5600 | default: |
5601 | gcc_unreachable (); |
5602 | } |
5603 | |
5604 | if (!vec_stmt) /* transformation not required. */ |
5605 | { |
5606 | if (slp_node |
5607 | && (!vect_maybe_update_slp_op_vectype (slp_op0, vectype_in) |
5608 | || !vect_maybe_update_slp_op_vectype (slp_op1, vectype_in))) |
5609 | { |
5610 | if (dump_enabled_p ()) |
5611 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5612 | "incompatible vector types for invariants\n" ); |
5613 | return false; |
5614 | } |
5615 | DUMP_VECT_SCOPE ("vectorizable_conversion" ); |
5616 | if (modifier == NONE) |
5617 | { |
5618 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; |
5619 | vect_model_simple_cost (vinfo, stmt_info, |
5620 | ncopies: ncopies * (1 + multi_step_cvt), |
5621 | dt, ndts, node: slp_node, cost_vec); |
5622 | } |
5623 | else if (modifier == NARROW_SRC || modifier == NARROW_DST) |
5624 | { |
5625 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; |
5626 | /* The final packing step produces one vector result per copy. */ |
5627 | unsigned int nvectors |
5628 | = (slp_node ? SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node) : ncopies); |
5629 | vect_model_promotion_demotion_cost (stmt_info, dt, ncopies: nvectors, |
5630 | pwr: multi_step_cvt, cost_vec, |
5631 | widen_arith); |
5632 | } |
5633 | else |
5634 | { |
5635 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; |
5636 | /* The initial unpacking step produces two vector results |
5637 | per copy. MULTI_STEP_CVT is 0 for a single conversion, |
5638 | so >> MULTI_STEP_CVT divides by 2^(number of steps - 1). */ |
5639 | unsigned int nvectors |
5640 | = (slp_node |
5641 | ? SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node) >> multi_step_cvt |
5642 | : ncopies * 2); |
5643 | vect_model_promotion_demotion_cost (stmt_info, dt, ncopies: nvectors, |
5644 | pwr: multi_step_cvt, cost_vec, |
5645 | widen_arith); |
5646 | } |
5647 | interm_types.release (); |
5648 | return true; |
5649 | } |
5650 | |
5651 | /* Transform. */ |
5652 | if (dump_enabled_p ()) |
5653 | dump_printf_loc (MSG_NOTE, vect_location, |
5654 | "transform conversion. ncopies = %d.\n" , ncopies); |
5655 | |
5656 | if (op_type == binary_op) |
5657 | { |
5658 | if (CONSTANT_CLASS_P (op0)) |
5659 | op0 = fold_convert (TREE_TYPE (op1), op0); |
5660 | else if (CONSTANT_CLASS_P (op1)) |
5661 | op1 = fold_convert (TREE_TYPE (op0), op1); |
5662 | } |
5663 | |
5664 | /* In case of multi-step conversion, we first generate conversion operations |
5665 | to the intermediate types, and then from that types to the final one. |
5666 | We create vector destinations for the intermediate type (TYPES) received |
5667 | from supportable_*_operation, and store them in the correct order |
5668 | for future use in vect_create_vectorized_*_stmts (). */ |
5669 | auto_vec<tree> vec_dsts (multi_step_cvt + 1); |
5670 | bool widen_or_narrow_float_p |
5671 | = cvt_type && (modifier == WIDEN || modifier == NARROW_SRC); |
5672 | vec_dest = vect_create_destination_var (scalar_dest, |
5673 | widen_or_narrow_float_p |
5674 | ? cvt_type : vectype_out); |
5675 | vec_dsts.quick_push (obj: vec_dest); |
5676 | |
5677 | if (multi_step_cvt) |
5678 | { |
5679 | for (i = interm_types.length () - 1; |
5680 | interm_types.iterate (ix: i, ptr: &intermediate_type); i--) |
5681 | { |
5682 | vec_dest = vect_create_destination_var (scalar_dest, |
5683 | intermediate_type); |
5684 | vec_dsts.quick_push (obj: vec_dest); |
5685 | } |
5686 | } |
5687 | |
5688 | if (cvt_type) |
5689 | vec_dest = vect_create_destination_var (scalar_dest, |
5690 | widen_or_narrow_float_p |
5691 | ? vectype_out : cvt_type); |
5692 | |
5693 | int ninputs = 1; |
5694 | if (!slp_node) |
5695 | { |
5696 | if (modifier == WIDEN) |
5697 | ; |
5698 | else if (modifier == NARROW_SRC || modifier == NARROW_DST) |
5699 | { |
5700 | if (multi_step_cvt) |
5701 | ninputs = vect_pow2 (x: multi_step_cvt); |
5702 | ninputs *= 2; |
5703 | } |
5704 | } |
5705 | |
5706 | switch (modifier) |
5707 | { |
5708 | case NONE: |
5709 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, |
5710 | op0, vectype0: vectype_in, vec_oprnds0: &vec_oprnds0); |
5711 | /* vec_dest is intermediate type operand when multi_step_cvt. */ |
5712 | if (multi_step_cvt) |
5713 | { |
5714 | cvt_op = vec_dest; |
5715 | vec_dest = vec_dsts[0]; |
5716 | } |
5717 | |
5718 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
5719 | { |
5720 | /* Arguments are ready, create the new vector stmt. */ |
5721 | gimple* new_stmt; |
5722 | if (multi_step_cvt) |
5723 | { |
5724 | gcc_assert (multi_step_cvt == 1); |
5725 | new_stmt = vect_gimple_build (cvt_op, codecvt1, vop0); |
5726 | new_temp = make_ssa_name (var: cvt_op, stmt: new_stmt); |
5727 | gimple_assign_set_lhs (gs: new_stmt, lhs: new_temp); |
5728 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
5729 | vop0 = new_temp; |
5730 | } |
5731 | new_stmt = vect_gimple_build (vec_dest, code1, vop0); |
5732 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
5733 | gimple_set_lhs (new_stmt, new_temp); |
5734 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
5735 | |
5736 | if (slp_node) |
5737 | slp_node->push_vec_def (def: new_stmt); |
5738 | else |
5739 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
5740 | } |
5741 | break; |
5742 | |
5743 | case WIDEN: |
5744 | /* In case the vectorization factor (VF) is bigger than the number |
5745 | of elements that we can fit in a vectype (nunits), we have to |
5746 | generate more than one vector stmt - i.e - we need to "unroll" |
5747 | the vector stmt by a factor VF/nunits. */ |
5748 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies: ncopies * ninputs, |
5749 | op0, vectype0: vectype_in, vec_oprnds0: &vec_oprnds0, |
5750 | op1: code == WIDEN_LSHIFT_EXPR ? NULL_TREE : op1, |
5751 | vectype1: vectype_in, vec_oprnds1: &vec_oprnds1); |
5752 | if (code == WIDEN_LSHIFT_EXPR) |
5753 | { |
5754 | int oprnds_size = vec_oprnds0.length (); |
5755 | vec_oprnds1.create (nelems: oprnds_size); |
5756 | for (i = 0; i < oprnds_size; ++i) |
5757 | vec_oprnds1.quick_push (obj: op1); |
5758 | } |
5759 | /* Arguments are ready. Create the new vector stmts. */ |
5760 | for (i = multi_step_cvt; i >= 0; i--) |
5761 | { |
5762 | tree this_dest = vec_dsts[i]; |
5763 | code_helper c1 = code1, c2 = code2; |
5764 | if (i == 0 && codecvt2 != ERROR_MARK) |
5765 | { |
5766 | c1 = codecvt1; |
5767 | c2 = codecvt2; |
5768 | } |
5769 | if (known_eq (nunits_out, nunits_in)) |
5770 | vect_create_half_widening_stmts (vinfo, vec_oprnds0: &vec_oprnds0, vec_oprnds1: &vec_oprnds1, |
5771 | stmt_info, vec_dest: this_dest, gsi, code1: c1, |
5772 | op_type); |
5773 | else |
5774 | vect_create_vectorized_promotion_stmts (vinfo, vec_oprnds0: &vec_oprnds0, |
5775 | vec_oprnds1: &vec_oprnds1, stmt_info, |
5776 | vec_dest: this_dest, gsi, |
5777 | ch1: c1, ch2: c2, op_type); |
5778 | } |
5779 | |
5780 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
5781 | { |
5782 | gimple *new_stmt; |
5783 | if (cvt_type) |
5784 | { |
5785 | new_temp = make_ssa_name (var: vec_dest); |
5786 | new_stmt = vect_gimple_build (new_temp, codecvt1, vop0); |
5787 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
5788 | } |
5789 | else |
5790 | new_stmt = SSA_NAME_DEF_STMT (vop0); |
5791 | |
5792 | if (slp_node) |
5793 | slp_node->push_vec_def (def: new_stmt); |
5794 | else |
5795 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
5796 | } |
5797 | break; |
5798 | |
5799 | case NARROW_SRC: |
5800 | case NARROW_DST: |
5801 | /* In case the vectorization factor (VF) is bigger than the number |
5802 | of elements that we can fit in a vectype (nunits), we have to |
5803 | generate more than one vector stmt - i.e - we need to "unroll" |
5804 | the vector stmt by a factor VF/nunits. */ |
5805 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies: ncopies * ninputs, |
5806 | op0, vectype0: vectype_in, vec_oprnds0: &vec_oprnds0); |
5807 | /* Arguments are ready. Create the new vector stmts. */ |
5808 | if (cvt_type && modifier == NARROW_DST) |
5809 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
5810 | { |
5811 | new_temp = make_ssa_name (var: vec_dest); |
5812 | gimple *new_stmt = vect_gimple_build (new_temp, codecvt1, vop0); |
5813 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
5814 | vec_oprnds0[i] = new_temp; |
5815 | } |
5816 | |
5817 | vect_create_vectorized_demotion_stmts (vinfo, vec_oprnds: &vec_oprnds0, |
5818 | multi_step_cvt, |
5819 | stmt_info, vec_dsts, gsi, |
5820 | slp_node, code: code1, |
5821 | narrow_src_p: modifier == NARROW_SRC); |
5822 | /* After demoting op0 to cvt_type, convert it to dest. */ |
5823 | if (cvt_type && code == FLOAT_EXPR) |
5824 | { |
5825 | for (unsigned int i = 0; i != vec_oprnds0.length() / 2; i++) |
5826 | { |
5827 | /* Arguments are ready, create the new vector stmt. */ |
5828 | gcc_assert (TREE_CODE_LENGTH ((tree_code) codecvt1) == unary_op); |
5829 | gimple *new_stmt |
5830 | = vect_gimple_build (vec_dest, codecvt1, vec_oprnds0[i]); |
5831 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
5832 | gimple_set_lhs (new_stmt, new_temp); |
5833 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
5834 | |
5835 | /* This is the last step of the conversion sequence. Store the |
5836 | vectors in SLP_NODE or in vector info of the scalar statement |
5837 | (or in STMT_VINFO_RELATED_STMT chain). */ |
5838 | if (slp_node) |
5839 | slp_node->push_vec_def (def: new_stmt); |
5840 | else |
5841 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
5842 | } |
5843 | } |
5844 | break; |
5845 | } |
5846 | if (!slp_node) |
5847 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
5848 | |
5849 | vec_oprnds0.release (); |
5850 | vec_oprnds1.release (); |
5851 | interm_types.release (); |
5852 | |
5853 | return true; |
5854 | } |
5855 | |
5856 | /* Return true if we can assume from the scalar form of STMT_INFO that |
5857 | neither the scalar nor the vector forms will generate code. STMT_INFO |
5858 | is known not to involve a data reference. */ |
5859 | |
5860 | bool |
5861 | vect_nop_conversion_p (stmt_vec_info stmt_info) |
5862 | { |
5863 | gassign *stmt = dyn_cast <gassign *> (p: stmt_info->stmt); |
5864 | if (!stmt) |
5865 | return false; |
5866 | |
5867 | tree lhs = gimple_assign_lhs (gs: stmt); |
5868 | tree_code code = gimple_assign_rhs_code (gs: stmt); |
5869 | tree rhs = gimple_assign_rhs1 (gs: stmt); |
5870 | |
5871 | if (code == SSA_NAME || code == VIEW_CONVERT_EXPR) |
5872 | return true; |
5873 | |
5874 | if (CONVERT_EXPR_CODE_P (code)) |
5875 | return tree_nop_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)); |
5876 | |
5877 | return false; |
5878 | } |
5879 | |
5880 | /* Function vectorizable_assignment. |
5881 | |
5882 | Check if STMT_INFO performs an assignment (copy) that can be vectorized. |
5883 | If VEC_STMT is also passed, vectorize the STMT_INFO: create a vectorized |
5884 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
5885 | Return true if STMT_INFO is vectorizable in this way. */ |
5886 | |
5887 | static bool |
5888 | vectorizable_assignment (vec_info *vinfo, |
5889 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
5890 | gimple **vec_stmt, slp_tree slp_node, |
5891 | stmt_vector_for_cost *cost_vec) |
5892 | { |
5893 | tree vec_dest; |
5894 | tree scalar_dest; |
5895 | tree op; |
5896 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
5897 | tree new_temp; |
5898 | enum vect_def_type dt[1] = {vect_unknown_def_type}; |
5899 | int ndts = 1; |
5900 | int ncopies; |
5901 | int i; |
5902 | vec<tree> vec_oprnds = vNULL; |
5903 | tree vop; |
5904 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
5905 | enum tree_code code; |
5906 | tree vectype_in; |
5907 | |
5908 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
5909 | return false; |
5910 | |
5911 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5912 | && ! vec_stmt) |
5913 | return false; |
5914 | |
5915 | /* Is vectorizable assignment? */ |
5916 | gassign *stmt = dyn_cast <gassign *> (p: stmt_info->stmt); |
5917 | if (!stmt) |
5918 | return false; |
5919 | |
5920 | scalar_dest = gimple_assign_lhs (gs: stmt); |
5921 | if (TREE_CODE (scalar_dest) != SSA_NAME) |
5922 | return false; |
5923 | |
5924 | if (STMT_VINFO_DATA_REF (stmt_info)) |
5925 | return false; |
5926 | |
5927 | code = gimple_assign_rhs_code (gs: stmt); |
5928 | if (!(gimple_assign_single_p (gs: stmt) |
5929 | || code == PAREN_EXPR |
5930 | || CONVERT_EXPR_CODE_P (code))) |
5931 | return false; |
5932 | |
5933 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
5934 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
5935 | |
5936 | /* Multiple types in SLP are handled by creating the appropriate number of |
5937 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
5938 | case of SLP. */ |
5939 | if (slp_node) |
5940 | ncopies = 1; |
5941 | else |
5942 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
5943 | |
5944 | gcc_assert (ncopies >= 1); |
5945 | |
5946 | slp_tree slp_op; |
5947 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 0, &op, &slp_op, |
5948 | &dt[0], &vectype_in)) |
5949 | { |
5950 | if (dump_enabled_p ()) |
5951 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5952 | "use not simple.\n" ); |
5953 | return false; |
5954 | } |
5955 | if (!vectype_in) |
5956 | vectype_in = get_vectype_for_scalar_type (vinfo, TREE_TYPE (op), slp_node); |
5957 | |
5958 | /* We can handle NOP_EXPR conversions that do not change the number |
5959 | of elements or the vector size. */ |
5960 | if ((CONVERT_EXPR_CODE_P (code) |
5961 | || code == VIEW_CONVERT_EXPR) |
5962 | && (!vectype_in |
5963 | || maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: vectype_in), b: nunits) |
5964 | || maybe_ne (a: GET_MODE_SIZE (TYPE_MODE (vectype)), |
5965 | b: GET_MODE_SIZE (TYPE_MODE (vectype_in))))) |
5966 | return false; |
5967 | |
5968 | if (VECTOR_BOOLEAN_TYPE_P (vectype) != VECTOR_BOOLEAN_TYPE_P (vectype_in)) |
5969 | { |
5970 | if (dump_enabled_p ()) |
5971 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5972 | "can't convert between boolean and non " |
5973 | "boolean vectors %T\n" , TREE_TYPE (op)); |
5974 | |
5975 | return false; |
5976 | } |
5977 | |
5978 | /* We do not handle bit-precision changes. */ |
5979 | if ((CONVERT_EXPR_CODE_P (code) |
5980 | || code == VIEW_CONVERT_EXPR) |
5981 | && ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) |
5982 | && !type_has_mode_precision_p (TREE_TYPE (scalar_dest))) |
5983 | || (INTEGRAL_TYPE_P (TREE_TYPE (op)) |
5984 | && !type_has_mode_precision_p (TREE_TYPE (op)))) |
5985 | /* But a conversion that does not change the bit-pattern is ok. */ |
5986 | && !(INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) |
5987 | && INTEGRAL_TYPE_P (TREE_TYPE (op)) |
5988 | && (((TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
5989 | > TYPE_PRECISION (TREE_TYPE (op))) |
5990 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
5991 | || (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
5992 | == TYPE_PRECISION (TREE_TYPE (op)))))) |
5993 | { |
5994 | if (dump_enabled_p ()) |
5995 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5996 | "type conversion to/from bit-precision " |
5997 | "unsupported.\n" ); |
5998 | return false; |
5999 | } |
6000 | |
6001 | if (!vec_stmt) /* transformation not required. */ |
6002 | { |
6003 | if (slp_node |
6004 | && !vect_maybe_update_slp_op_vectype (slp_op, vectype_in)) |
6005 | { |
6006 | if (dump_enabled_p ()) |
6007 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6008 | "incompatible vector types for invariants\n" ); |
6009 | return false; |
6010 | } |
6011 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; |
6012 | DUMP_VECT_SCOPE ("vectorizable_assignment" ); |
6013 | if (!vect_nop_conversion_p (stmt_info)) |
6014 | vect_model_simple_cost (vinfo, stmt_info, ncopies, dt, ndts, node: slp_node, |
6015 | cost_vec); |
6016 | return true; |
6017 | } |
6018 | |
6019 | /* Transform. */ |
6020 | if (dump_enabled_p ()) |
6021 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n" ); |
6022 | |
6023 | /* Handle def. */ |
6024 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
6025 | |
6026 | /* Handle use. */ |
6027 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, op0: op, vec_oprnds0: &vec_oprnds); |
6028 | |
6029 | /* Arguments are ready. create the new vector stmt. */ |
6030 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
6031 | { |
6032 | if (CONVERT_EXPR_CODE_P (code) |
6033 | || code == VIEW_CONVERT_EXPR) |
6034 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
6035 | gassign *new_stmt = gimple_build_assign (vec_dest, vop); |
6036 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
6037 | gimple_assign_set_lhs (gs: new_stmt, lhs: new_temp); |
6038 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6039 | if (slp_node) |
6040 | slp_node->push_vec_def (def: new_stmt); |
6041 | else |
6042 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
6043 | } |
6044 | if (!slp_node) |
6045 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
6046 | |
6047 | vec_oprnds.release (); |
6048 | return true; |
6049 | } |
6050 | |
6051 | |
6052 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
6053 | either as shift by a scalar or by a vector. */ |
6054 | |
6055 | bool |
6056 | vect_supportable_shift (vec_info *vinfo, enum tree_code code, tree scalar_type) |
6057 | { |
6058 | |
6059 | machine_mode vec_mode; |
6060 | optab optab; |
6061 | int icode; |
6062 | tree vectype; |
6063 | |
6064 | vectype = get_vectype_for_scalar_type (vinfo, scalar_type); |
6065 | if (!vectype) |
6066 | return false; |
6067 | |
6068 | optab = optab_for_tree_code (code, vectype, optab_scalar); |
6069 | if (!optab |
6070 | || optab_handler (op: optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) |
6071 | { |
6072 | optab = optab_for_tree_code (code, vectype, optab_vector); |
6073 | if (!optab |
6074 | || (optab_handler (op: optab, TYPE_MODE (vectype)) |
6075 | == CODE_FOR_nothing)) |
6076 | return false; |
6077 | } |
6078 | |
6079 | vec_mode = TYPE_MODE (vectype); |
6080 | icode = (int) optab_handler (op: optab, mode: vec_mode); |
6081 | if (icode == CODE_FOR_nothing) |
6082 | return false; |
6083 | |
6084 | return true; |
6085 | } |
6086 | |
6087 | |
6088 | /* Function vectorizable_shift. |
6089 | |
6090 | Check if STMT_INFO performs a shift operation that can be vectorized. |
6091 | If VEC_STMT is also passed, vectorize the STMT_INFO: create a vectorized |
6092 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
6093 | Return true if STMT_INFO is vectorizable in this way. */ |
6094 | |
6095 | static bool |
6096 | vectorizable_shift (vec_info *vinfo, |
6097 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
6098 | gimple **vec_stmt, slp_tree slp_node, |
6099 | stmt_vector_for_cost *cost_vec) |
6100 | { |
6101 | tree vec_dest; |
6102 | tree scalar_dest; |
6103 | tree op0, op1 = NULL; |
6104 | tree vec_oprnd1 = NULL_TREE; |
6105 | tree vectype; |
6106 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
6107 | enum tree_code code; |
6108 | machine_mode vec_mode; |
6109 | tree new_temp; |
6110 | optab optab; |
6111 | int icode; |
6112 | machine_mode optab_op2_mode; |
6113 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
6114 | int ndts = 2; |
6115 | poly_uint64 nunits_in; |
6116 | poly_uint64 nunits_out; |
6117 | tree vectype_out; |
6118 | tree op1_vectype; |
6119 | int ncopies; |
6120 | int i; |
6121 | vec<tree> vec_oprnds0 = vNULL; |
6122 | vec<tree> vec_oprnds1 = vNULL; |
6123 | tree vop0, vop1; |
6124 | unsigned int k; |
6125 | bool scalar_shift_arg = true; |
6126 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
6127 | bool incompatible_op1_vectype_p = false; |
6128 | |
6129 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
6130 | return false; |
6131 | |
6132 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6133 | && STMT_VINFO_DEF_TYPE (stmt_info) != vect_nested_cycle |
6134 | && ! vec_stmt) |
6135 | return false; |
6136 | |
6137 | /* Is STMT a vectorizable binary/unary operation? */ |
6138 | gassign *stmt = dyn_cast <gassign *> (p: stmt_info->stmt); |
6139 | if (!stmt) |
6140 | return false; |
6141 | |
6142 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) |
6143 | return false; |
6144 | |
6145 | code = gimple_assign_rhs_code (gs: stmt); |
6146 | |
6147 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
6148 | || code == RROTATE_EXPR)) |
6149 | return false; |
6150 | |
6151 | scalar_dest = gimple_assign_lhs (gs: stmt); |
6152 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
6153 | if (!type_has_mode_precision_p (TREE_TYPE (scalar_dest))) |
6154 | { |
6155 | if (dump_enabled_p ()) |
6156 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6157 | "bit-precision shifts not supported.\n" ); |
6158 | return false; |
6159 | } |
6160 | |
6161 | slp_tree slp_op0; |
6162 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
6163 | 0, &op0, &slp_op0, &dt[0], &vectype)) |
6164 | { |
6165 | if (dump_enabled_p ()) |
6166 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6167 | "use not simple.\n" ); |
6168 | return false; |
6169 | } |
6170 | /* If op0 is an external or constant def, infer the vector type |
6171 | from the scalar type. */ |
6172 | if (!vectype) |
6173 | vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (op0), slp_node); |
6174 | if (vec_stmt) |
6175 | gcc_assert (vectype); |
6176 | if (!vectype) |
6177 | { |
6178 | if (dump_enabled_p ()) |
6179 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6180 | "no vectype for scalar type\n" ); |
6181 | return false; |
6182 | } |
6183 | |
6184 | nunits_out = TYPE_VECTOR_SUBPARTS (node: vectype_out); |
6185 | nunits_in = TYPE_VECTOR_SUBPARTS (node: vectype); |
6186 | if (maybe_ne (a: nunits_out, b: nunits_in)) |
6187 | return false; |
6188 | |
6189 | stmt_vec_info op1_def_stmt_info; |
6190 | slp_tree slp_op1; |
6191 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 1, &op1, &slp_op1, |
6192 | &dt[1], &op1_vectype, &op1_def_stmt_info)) |
6193 | { |
6194 | if (dump_enabled_p ()) |
6195 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6196 | "use not simple.\n" ); |
6197 | return false; |
6198 | } |
6199 | |
6200 | /* Multiple types in SLP are handled by creating the appropriate number of |
6201 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
6202 | case of SLP. */ |
6203 | if (slp_node) |
6204 | ncopies = 1; |
6205 | else |
6206 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
6207 | |
6208 | gcc_assert (ncopies >= 1); |
6209 | |
6210 | /* Determine whether the shift amount is a vector, or scalar. If the |
6211 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ |
6212 | |
6213 | if ((dt[1] == vect_internal_def |
6214 | || dt[1] == vect_induction_def |
6215 | || dt[1] == vect_nested_cycle) |
6216 | && !slp_node) |
6217 | scalar_shift_arg = false; |
6218 | else if (dt[1] == vect_constant_def |
6219 | || dt[1] == vect_external_def |
6220 | || dt[1] == vect_internal_def) |
6221 | { |
6222 | /* In SLP, need to check whether the shift count is the same, |
6223 | in loops if it is a constant or invariant, it is always |
6224 | a scalar shift. */ |
6225 | if (slp_node) |
6226 | { |
6227 | vec<stmt_vec_info> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
6228 | stmt_vec_info slpstmt_info; |
6229 | |
6230 | FOR_EACH_VEC_ELT (stmts, k, slpstmt_info) |
6231 | { |
6232 | gassign *slpstmt = as_a <gassign *> (p: slpstmt_info->stmt); |
6233 | if (!operand_equal_p (gimple_assign_rhs2 (gs: slpstmt), op1, flags: 0)) |
6234 | scalar_shift_arg = false; |
6235 | } |
6236 | |
6237 | /* For internal SLP defs we have to make sure we see scalar stmts |
6238 | for all vector elements. |
6239 | ??? For different vectors we could resort to a different |
6240 | scalar shift operand but code-generation below simply always |
6241 | takes the first. */ |
6242 | if (dt[1] == vect_internal_def |
6243 | && maybe_ne (a: nunits_out * SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node), |
6244 | b: stmts.length ())) |
6245 | scalar_shift_arg = false; |
6246 | } |
6247 | |
6248 | /* If the shift amount is computed by a pattern stmt we cannot |
6249 | use the scalar amount directly thus give up and use a vector |
6250 | shift. */ |
6251 | if (op1_def_stmt_info && is_pattern_stmt_p (stmt_info: op1_def_stmt_info)) |
6252 | scalar_shift_arg = false; |
6253 | } |
6254 | else |
6255 | { |
6256 | if (dump_enabled_p ()) |
6257 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6258 | "operand mode requires invariant argument.\n" ); |
6259 | return false; |
6260 | } |
6261 | |
6262 | /* Vector shifted by vector. */ |
6263 | bool was_scalar_shift_arg = scalar_shift_arg; |
6264 | if (!scalar_shift_arg) |
6265 | { |
6266 | optab = optab_for_tree_code (code, vectype, optab_vector); |
6267 | if (dump_enabled_p ()) |
6268 | dump_printf_loc (MSG_NOTE, vect_location, |
6269 | "vector/vector shift/rotate found.\n" ); |
6270 | |
6271 | if (!op1_vectype) |
6272 | op1_vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (op1), |
6273 | slp_op1); |
6274 | incompatible_op1_vectype_p |
6275 | = (op1_vectype == NULL_TREE |
6276 | || maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: op1_vectype), |
6277 | b: TYPE_VECTOR_SUBPARTS (node: vectype)) |
6278 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)); |
6279 | if (incompatible_op1_vectype_p |
6280 | && (!slp_node |
6281 | || SLP_TREE_DEF_TYPE (slp_op1) != vect_constant_def |
6282 | || slp_op1->refcnt != 1)) |
6283 | { |
6284 | if (dump_enabled_p ()) |
6285 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6286 | "unusable type for last operand in" |
6287 | " vector/vector shift/rotate.\n" ); |
6288 | return false; |
6289 | } |
6290 | } |
6291 | /* See if the machine has a vector shifted by scalar insn and if not |
6292 | then see if it has a vector shifted by vector insn. */ |
6293 | else |
6294 | { |
6295 | optab = optab_for_tree_code (code, vectype, optab_scalar); |
6296 | if (optab |
6297 | && optab_handler (op: optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) |
6298 | { |
6299 | if (dump_enabled_p ()) |
6300 | dump_printf_loc (MSG_NOTE, vect_location, |
6301 | "vector/scalar shift/rotate found.\n" ); |
6302 | } |
6303 | else |
6304 | { |
6305 | optab = optab_for_tree_code (code, vectype, optab_vector); |
6306 | if (optab |
6307 | && (optab_handler (op: optab, TYPE_MODE (vectype)) |
6308 | != CODE_FOR_nothing)) |
6309 | { |
6310 | scalar_shift_arg = false; |
6311 | |
6312 | if (dump_enabled_p ()) |
6313 | dump_printf_loc (MSG_NOTE, vect_location, |
6314 | "vector/vector shift/rotate found.\n" ); |
6315 | |
6316 | if (!op1_vectype) |
6317 | op1_vectype = get_vectype_for_scalar_type (vinfo, |
6318 | TREE_TYPE (op1), |
6319 | slp_op1); |
6320 | |
6321 | /* Unlike the other binary operators, shifts/rotates have |
6322 | the rhs being int, instead of the same type as the lhs, |
6323 | so make sure the scalar is the right type if we are |
6324 | dealing with vectors of long long/long/short/char. */ |
6325 | incompatible_op1_vectype_p |
6326 | = (!op1_vectype |
6327 | || !tree_nop_conversion_p (TREE_TYPE (vectype), |
6328 | TREE_TYPE (op1))); |
6329 | if (incompatible_op1_vectype_p |
6330 | && dt[1] == vect_internal_def) |
6331 | { |
6332 | if (dump_enabled_p ()) |
6333 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6334 | "unusable type for last operand in" |
6335 | " vector/vector shift/rotate.\n" ); |
6336 | return false; |
6337 | } |
6338 | } |
6339 | } |
6340 | } |
6341 | |
6342 | /* Supportable by target? */ |
6343 | if (!optab) |
6344 | { |
6345 | if (dump_enabled_p ()) |
6346 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6347 | "no optab.\n" ); |
6348 | return false; |
6349 | } |
6350 | vec_mode = TYPE_MODE (vectype); |
6351 | icode = (int) optab_handler (op: optab, mode: vec_mode); |
6352 | if (icode == CODE_FOR_nothing) |
6353 | { |
6354 | if (dump_enabled_p ()) |
6355 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6356 | "op not supported by target.\n" ); |
6357 | return false; |
6358 | } |
6359 | /* vector lowering cannot optimize vector shifts using word arithmetic. */ |
6360 | if (vect_emulated_vector_p (vectype)) |
6361 | return false; |
6362 | |
6363 | if (!vec_stmt) /* transformation not required. */ |
6364 | { |
6365 | if (slp_node |
6366 | && (!vect_maybe_update_slp_op_vectype (slp_op0, vectype) |
6367 | || ((!scalar_shift_arg || dt[1] == vect_internal_def) |
6368 | && (!incompatible_op1_vectype_p |
6369 | || dt[1] == vect_constant_def) |
6370 | && !vect_maybe_update_slp_op_vectype |
6371 | (slp_op1, |
6372 | incompatible_op1_vectype_p ? vectype : op1_vectype)))) |
6373 | { |
6374 | if (dump_enabled_p ()) |
6375 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6376 | "incompatible vector types for invariants\n" ); |
6377 | return false; |
6378 | } |
6379 | /* Now adjust the constant shift amount in place. */ |
6380 | if (slp_node |
6381 | && incompatible_op1_vectype_p |
6382 | && dt[1] == vect_constant_def) |
6383 | { |
6384 | for (unsigned i = 0; |
6385 | i < SLP_TREE_SCALAR_OPS (slp_op1).length (); ++i) |
6386 | { |
6387 | SLP_TREE_SCALAR_OPS (slp_op1)[i] |
6388 | = fold_convert (TREE_TYPE (vectype), |
6389 | SLP_TREE_SCALAR_OPS (slp_op1)[i]); |
6390 | gcc_assert ((TREE_CODE (SLP_TREE_SCALAR_OPS (slp_op1)[i]) |
6391 | == INTEGER_CST)); |
6392 | } |
6393 | } |
6394 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; |
6395 | DUMP_VECT_SCOPE ("vectorizable_shift" ); |
6396 | vect_model_simple_cost (vinfo, stmt_info, ncopies, dt, |
6397 | ndts: scalar_shift_arg ? 1 : ndts, node: slp_node, cost_vec); |
6398 | return true; |
6399 | } |
6400 | |
6401 | /* Transform. */ |
6402 | |
6403 | if (dump_enabled_p ()) |
6404 | dump_printf_loc (MSG_NOTE, vect_location, |
6405 | "transform binary/unary operation.\n" ); |
6406 | |
6407 | if (incompatible_op1_vectype_p && !slp_node) |
6408 | { |
6409 | gcc_assert (!scalar_shift_arg && was_scalar_shift_arg); |
6410 | op1 = fold_convert (TREE_TYPE (vectype), op1); |
6411 | if (dt[1] != vect_constant_def) |
6412 | op1 = vect_init_vector (vinfo, stmt_info, val: op1, |
6413 | TREE_TYPE (vectype), NULL); |
6414 | } |
6415 | |
6416 | /* Handle def. */ |
6417 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
6418 | |
6419 | if (scalar_shift_arg && dt[1] != vect_internal_def) |
6420 | { |
6421 | /* Vector shl and shr insn patterns can be defined with scalar |
6422 | operand 2 (shift operand). In this case, use constant or loop |
6423 | invariant op1 directly, without extending it to vector mode |
6424 | first. */ |
6425 | optab_op2_mode = insn_data[icode].operand[2].mode; |
6426 | if (!VECTOR_MODE_P (optab_op2_mode)) |
6427 | { |
6428 | if (dump_enabled_p ()) |
6429 | dump_printf_loc (MSG_NOTE, vect_location, |
6430 | "operand 1 using scalar mode.\n" ); |
6431 | vec_oprnd1 = op1; |
6432 | vec_oprnds1.create (nelems: slp_node ? slp_node->vec_stmts_size : ncopies); |
6433 | vec_oprnds1.quick_push (obj: vec_oprnd1); |
6434 | /* Store vec_oprnd1 for every vector stmt to be created. |
6435 | We check during the analysis that all the shift arguments |
6436 | are the same. |
6437 | TODO: Allow different constants for different vector |
6438 | stmts generated for an SLP instance. */ |
6439 | for (k = 0; |
6440 | k < (slp_node ? slp_node->vec_stmts_size - 1 : ncopies - 1); k++) |
6441 | vec_oprnds1.quick_push (obj: vec_oprnd1); |
6442 | } |
6443 | } |
6444 | else if (!scalar_shift_arg && slp_node && incompatible_op1_vectype_p) |
6445 | { |
6446 | if (was_scalar_shift_arg) |
6447 | { |
6448 | /* If the argument was the same in all lanes create |
6449 | the correctly typed vector shift amount directly. */ |
6450 | op1 = fold_convert (TREE_TYPE (vectype), op1); |
6451 | op1 = vect_init_vector (vinfo, stmt_info, val: op1, TREE_TYPE (vectype), |
6452 | gsi: !loop_vinfo ? gsi : NULL); |
6453 | vec_oprnd1 = vect_init_vector (vinfo, stmt_info, val: op1, type: vectype, |
6454 | gsi: !loop_vinfo ? gsi : NULL); |
6455 | vec_oprnds1.create (nelems: slp_node->vec_stmts_size); |
6456 | for (k = 0; k < slp_node->vec_stmts_size; k++) |
6457 | vec_oprnds1.quick_push (obj: vec_oprnd1); |
6458 | } |
6459 | else if (dt[1] == vect_constant_def) |
6460 | /* The constant shift amount has been adjusted in place. */ |
6461 | ; |
6462 | else |
6463 | gcc_assert (TYPE_MODE (op1_vectype) == TYPE_MODE (vectype)); |
6464 | } |
6465 | |
6466 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type |
6467 | (a special case for certain kind of vector shifts); otherwise, |
6468 | operand 1 should be of a vector type (the usual case). */ |
6469 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, |
6470 | op0, vec_oprnds0: &vec_oprnds0, |
6471 | op1: vec_oprnd1 ? NULL_TREE : op1, vec_oprnds1: &vec_oprnds1); |
6472 | |
6473 | /* Arguments are ready. Create the new vector stmt. */ |
6474 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
6475 | { |
6476 | /* For internal defs where we need to use a scalar shift arg |
6477 | extract the first lane. */ |
6478 | if (scalar_shift_arg && dt[1] == vect_internal_def) |
6479 | { |
6480 | vop1 = vec_oprnds1[0]; |
6481 | new_temp = make_ssa_name (TREE_TYPE (TREE_TYPE (vop1))); |
6482 | gassign *new_stmt |
6483 | = gimple_build_assign (new_temp, |
6484 | build3 (BIT_FIELD_REF, TREE_TYPE (new_temp), |
6485 | vop1, |
6486 | TYPE_SIZE (TREE_TYPE (new_temp)), |
6487 | bitsize_zero_node)); |
6488 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6489 | vop1 = new_temp; |
6490 | } |
6491 | else |
6492 | vop1 = vec_oprnds1[i]; |
6493 | gassign *new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
6494 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
6495 | gimple_assign_set_lhs (gs: new_stmt, lhs: new_temp); |
6496 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6497 | if (slp_node) |
6498 | slp_node->push_vec_def (def: new_stmt); |
6499 | else |
6500 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
6501 | } |
6502 | |
6503 | if (!slp_node) |
6504 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
6505 | |
6506 | vec_oprnds0.release (); |
6507 | vec_oprnds1.release (); |
6508 | |
6509 | return true; |
6510 | } |
6511 | |
6512 | /* Function vectorizable_operation. |
6513 | |
6514 | Check if STMT_INFO performs a binary, unary or ternary operation that can |
6515 | be vectorized. |
6516 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
6517 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
6518 | Return true if STMT_INFO is vectorizable in this way. */ |
6519 | |
6520 | static bool |
6521 | vectorizable_operation (vec_info *vinfo, |
6522 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
6523 | gimple **vec_stmt, slp_tree slp_node, |
6524 | stmt_vector_for_cost *cost_vec) |
6525 | { |
6526 | tree vec_dest; |
6527 | tree scalar_dest; |
6528 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
6529 | tree vectype; |
6530 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
6531 | enum tree_code code, orig_code; |
6532 | machine_mode vec_mode; |
6533 | tree new_temp; |
6534 | int op_type; |
6535 | optab optab; |
6536 | bool target_support_p; |
6537 | enum vect_def_type dt[3] |
6538 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; |
6539 | int ndts = 3; |
6540 | poly_uint64 nunits_in; |
6541 | poly_uint64 nunits_out; |
6542 | tree vectype_out; |
6543 | int ncopies, vec_num; |
6544 | int i; |
6545 | vec<tree> vec_oprnds0 = vNULL; |
6546 | vec<tree> vec_oprnds1 = vNULL; |
6547 | vec<tree> vec_oprnds2 = vNULL; |
6548 | tree vop0, vop1, vop2; |
6549 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
6550 | |
6551 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
6552 | return false; |
6553 | |
6554 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6555 | && ! vec_stmt) |
6556 | return false; |
6557 | |
6558 | /* Is STMT a vectorizable binary/unary operation? */ |
6559 | gassign *stmt = dyn_cast <gassign *> (p: stmt_info->stmt); |
6560 | if (!stmt) |
6561 | return false; |
6562 | |
6563 | /* Loads and stores are handled in vectorizable_{load,store}. */ |
6564 | if (STMT_VINFO_DATA_REF (stmt_info)) |
6565 | return false; |
6566 | |
6567 | orig_code = code = gimple_assign_rhs_code (gs: stmt); |
6568 | |
6569 | /* Shifts are handled in vectorizable_shift. */ |
6570 | if (code == LSHIFT_EXPR |
6571 | || code == RSHIFT_EXPR |
6572 | || code == LROTATE_EXPR |
6573 | || code == RROTATE_EXPR) |
6574 | return false; |
6575 | |
6576 | /* Comparisons are handled in vectorizable_comparison. */ |
6577 | if (TREE_CODE_CLASS (code) == tcc_comparison) |
6578 | return false; |
6579 | |
6580 | /* Conditions are handled in vectorizable_condition. */ |
6581 | if (code == COND_EXPR) |
6582 | return false; |
6583 | |
6584 | /* For pointer addition and subtraction, we should use the normal |
6585 | plus and minus for the vector operation. */ |
6586 | if (code == POINTER_PLUS_EXPR) |
6587 | code = PLUS_EXPR; |
6588 | if (code == POINTER_DIFF_EXPR) |
6589 | code = MINUS_EXPR; |
6590 | |
6591 | /* Support only unary or binary operations. */ |
6592 | op_type = TREE_CODE_LENGTH (code); |
6593 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
6594 | { |
6595 | if (dump_enabled_p ()) |
6596 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6597 | "num. args = %d (not unary/binary/ternary op).\n" , |
6598 | op_type); |
6599 | return false; |
6600 | } |
6601 | |
6602 | scalar_dest = gimple_assign_lhs (gs: stmt); |
6603 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
6604 | |
6605 | /* Most operations cannot handle bit-precision types without extra |
6606 | truncations. */ |
6607 | bool mask_op_p = VECTOR_BOOLEAN_TYPE_P (vectype_out); |
6608 | if (!mask_op_p |
6609 | && !type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
6610 | /* Exception are bitwise binary operations. */ |
6611 | && code != BIT_IOR_EXPR |
6612 | && code != BIT_XOR_EXPR |
6613 | && code != BIT_AND_EXPR) |
6614 | { |
6615 | if (dump_enabled_p ()) |
6616 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6617 | "bit-precision arithmetic not supported.\n" ); |
6618 | return false; |
6619 | } |
6620 | |
6621 | slp_tree slp_op0; |
6622 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
6623 | 0, &op0, &slp_op0, &dt[0], &vectype)) |
6624 | { |
6625 | if (dump_enabled_p ()) |
6626 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6627 | "use not simple.\n" ); |
6628 | return false; |
6629 | } |
6630 | bool is_invariant = (dt[0] == vect_external_def |
6631 | || dt[0] == vect_constant_def); |
6632 | /* If op0 is an external or constant def, infer the vector type |
6633 | from the scalar type. */ |
6634 | if (!vectype) |
6635 | { |
6636 | /* For boolean type we cannot determine vectype by |
6637 | invariant value (don't know whether it is a vector |
6638 | of booleans or vector of integers). We use output |
6639 | vectype because operations on boolean don't change |
6640 | type. */ |
6641 | if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op0))) |
6642 | { |
6643 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (scalar_dest))) |
6644 | { |
6645 | if (dump_enabled_p ()) |
6646 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6647 | "not supported operation on bool value.\n" ); |
6648 | return false; |
6649 | } |
6650 | vectype = vectype_out; |
6651 | } |
6652 | else |
6653 | vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (op0), |
6654 | slp_node); |
6655 | } |
6656 | if (vec_stmt) |
6657 | gcc_assert (vectype); |
6658 | if (!vectype) |
6659 | { |
6660 | if (dump_enabled_p ()) |
6661 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6662 | "no vectype for scalar type %T\n" , |
6663 | TREE_TYPE (op0)); |
6664 | |
6665 | return false; |
6666 | } |
6667 | |
6668 | nunits_out = TYPE_VECTOR_SUBPARTS (node: vectype_out); |
6669 | nunits_in = TYPE_VECTOR_SUBPARTS (node: vectype); |
6670 | if (maybe_ne (a: nunits_out, b: nunits_in) |
6671 | || !tree_nop_conversion_p (TREE_TYPE (vectype_out), TREE_TYPE (vectype))) |
6672 | return false; |
6673 | |
6674 | tree vectype2 = NULL_TREE, vectype3 = NULL_TREE; |
6675 | slp_tree slp_op1 = NULL, slp_op2 = NULL; |
6676 | if (op_type == binary_op || op_type == ternary_op) |
6677 | { |
6678 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
6679 | 1, &op1, &slp_op1, &dt[1], &vectype2)) |
6680 | { |
6681 | if (dump_enabled_p ()) |
6682 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6683 | "use not simple.\n" ); |
6684 | return false; |
6685 | } |
6686 | is_invariant &= (dt[1] == vect_external_def |
6687 | || dt[1] == vect_constant_def); |
6688 | if (vectype2 |
6689 | && (maybe_ne (a: nunits_out, b: TYPE_VECTOR_SUBPARTS (node: vectype2)) |
6690 | || !tree_nop_conversion_p (TREE_TYPE (vectype_out), |
6691 | TREE_TYPE (vectype2)))) |
6692 | return false; |
6693 | } |
6694 | if (op_type == ternary_op) |
6695 | { |
6696 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
6697 | 2, &op2, &slp_op2, &dt[2], &vectype3)) |
6698 | { |
6699 | if (dump_enabled_p ()) |
6700 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6701 | "use not simple.\n" ); |
6702 | return false; |
6703 | } |
6704 | is_invariant &= (dt[2] == vect_external_def |
6705 | || dt[2] == vect_constant_def); |
6706 | if (vectype3 |
6707 | && (maybe_ne (a: nunits_out, b: TYPE_VECTOR_SUBPARTS (node: vectype3)) |
6708 | || !tree_nop_conversion_p (TREE_TYPE (vectype_out), |
6709 | TREE_TYPE (vectype3)))) |
6710 | return false; |
6711 | } |
6712 | |
6713 | /* Multiple types in SLP are handled by creating the appropriate number of |
6714 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
6715 | case of SLP. */ |
6716 | if (slp_node) |
6717 | { |
6718 | ncopies = 1; |
6719 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
6720 | } |
6721 | else |
6722 | { |
6723 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
6724 | vec_num = 1; |
6725 | } |
6726 | |
6727 | gcc_assert (ncopies >= 1); |
6728 | |
6729 | /* Reject attempts to combine mask types with nonmask types, e.g. if |
6730 | we have an AND between a (nonmask) boolean loaded from memory and |
6731 | a (mask) boolean result of a comparison. |
6732 | |
6733 | TODO: We could easily fix these cases up using pattern statements. */ |
6734 | if (VECTOR_BOOLEAN_TYPE_P (vectype) != mask_op_p |
6735 | || (vectype2 && VECTOR_BOOLEAN_TYPE_P (vectype2) != mask_op_p) |
6736 | || (vectype3 && VECTOR_BOOLEAN_TYPE_P (vectype3) != mask_op_p)) |
6737 | { |
6738 | if (dump_enabled_p ()) |
6739 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6740 | "mixed mask and nonmask vector types\n" ); |
6741 | return false; |
6742 | } |
6743 | |
6744 | /* Supportable by target? */ |
6745 | |
6746 | vec_mode = TYPE_MODE (vectype); |
6747 | if (code == MULT_HIGHPART_EXPR) |
6748 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
6749 | else |
6750 | { |
6751 | optab = optab_for_tree_code (code, vectype, optab_default); |
6752 | if (!optab) |
6753 | { |
6754 | if (dump_enabled_p ()) |
6755 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6756 | "no optab.\n" ); |
6757 | return false; |
6758 | } |
6759 | target_support_p = (optab_handler (op: optab, mode: vec_mode) != CODE_FOR_nothing |
6760 | || optab_libfunc (optab, vec_mode)); |
6761 | } |
6762 | |
6763 | bool using_emulated_vectors_p = vect_emulated_vector_p (vectype); |
6764 | if (!target_support_p || using_emulated_vectors_p) |
6765 | { |
6766 | if (dump_enabled_p ()) |
6767 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6768 | "op not supported by target.\n" ); |
6769 | /* When vec_mode is not a vector mode and we verified ops we |
6770 | do not have to lower like AND are natively supported let |
6771 | those through even when the mode isn't word_mode. For |
6772 | ops we have to lower the lowering code assumes we are |
6773 | dealing with word_mode. */ |
6774 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype)) |
6775 | || (((code == PLUS_EXPR || code == MINUS_EXPR || code == NEGATE_EXPR) |
6776 | || !target_support_p) |
6777 | && maybe_ne (a: GET_MODE_SIZE (mode: vec_mode), UNITS_PER_WORD)) |
6778 | /* Check only during analysis. */ |
6779 | || (!vec_stmt && !vect_can_vectorize_without_simd_p (code))) |
6780 | { |
6781 | if (dump_enabled_p ()) |
6782 | dump_printf (MSG_NOTE, "using word mode not possible.\n" ); |
6783 | return false; |
6784 | } |
6785 | if (dump_enabled_p ()) |
6786 | dump_printf_loc (MSG_NOTE, vect_location, |
6787 | "proceeding using word mode.\n" ); |
6788 | using_emulated_vectors_p = true; |
6789 | } |
6790 | |
6791 | int reduc_idx = STMT_VINFO_REDUC_IDX (stmt_info); |
6792 | vec_loop_masks *masks = (loop_vinfo ? &LOOP_VINFO_MASKS (loop_vinfo) : NULL); |
6793 | vec_loop_lens *lens = (loop_vinfo ? &LOOP_VINFO_LENS (loop_vinfo) : NULL); |
6794 | internal_fn cond_fn = get_conditional_internal_fn (code); |
6795 | internal_fn cond_len_fn = get_conditional_len_internal_fn (code); |
6796 | |
6797 | /* If operating on inactive elements could generate spurious traps, |
6798 | we need to restrict the operation to active lanes. Note that this |
6799 | specifically doesn't apply to unhoisted invariants, since they |
6800 | operate on the same value for every lane. |
6801 | |
6802 | Similarly, if this operation is part of a reduction, a fully-masked |
6803 | loop should only change the active lanes of the reduction chain, |
6804 | keeping the inactive lanes as-is. */ |
6805 | bool mask_out_inactive = ((!is_invariant && gimple_could_trap_p (stmt)) |
6806 | || reduc_idx >= 0); |
6807 | |
6808 | if (!vec_stmt) /* transformation not required. */ |
6809 | { |
6810 | if (loop_vinfo |
6811 | && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) |
6812 | && mask_out_inactive) |
6813 | { |
6814 | if (cond_len_fn != IFN_LAST |
6815 | && direct_internal_fn_supported_p (cond_len_fn, vectype, |
6816 | OPTIMIZE_FOR_SPEED)) |
6817 | vect_record_loop_len (loop_vinfo, lens, ncopies * vec_num, vectype, |
6818 | 1); |
6819 | else if (cond_fn != IFN_LAST |
6820 | && direct_internal_fn_supported_p (cond_fn, vectype, |
6821 | OPTIMIZE_FOR_SPEED)) |
6822 | vect_record_loop_mask (loop_vinfo, masks, ncopies * vec_num, |
6823 | vectype, NULL); |
6824 | else |
6825 | { |
6826 | if (dump_enabled_p ()) |
6827 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6828 | "can't use a fully-masked loop because no" |
6829 | " conditional operation is available.\n" ); |
6830 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
6831 | } |
6832 | } |
6833 | |
6834 | /* Put types on constant and invariant SLP children. */ |
6835 | if (slp_node |
6836 | && (!vect_maybe_update_slp_op_vectype (slp_op0, vectype) |
6837 | || !vect_maybe_update_slp_op_vectype (slp_op1, vectype) |
6838 | || !vect_maybe_update_slp_op_vectype (slp_op2, vectype))) |
6839 | { |
6840 | if (dump_enabled_p ()) |
6841 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6842 | "incompatible vector types for invariants\n" ); |
6843 | return false; |
6844 | } |
6845 | |
6846 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
6847 | DUMP_VECT_SCOPE ("vectorizable_operation" ); |
6848 | vect_model_simple_cost (vinfo, stmt_info, |
6849 | ncopies, dt, ndts, node: slp_node, cost_vec); |
6850 | if (using_emulated_vectors_p) |
6851 | { |
6852 | /* The above vect_model_simple_cost call handles constants |
6853 | in the prologue and (mis-)costs one of the stmts as |
6854 | vector stmt. See below for the actual lowering that will |
6855 | be applied. */ |
6856 | unsigned n |
6857 | = slp_node ? SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node) : ncopies; |
6858 | switch (code) |
6859 | { |
6860 | case PLUS_EXPR: |
6861 | n *= 5; |
6862 | break; |
6863 | case MINUS_EXPR: |
6864 | n *= 6; |
6865 | break; |
6866 | case NEGATE_EXPR: |
6867 | n *= 4; |
6868 | break; |
6869 | default: |
6870 | /* Bit operations do not have extra cost and are accounted |
6871 | as vector stmt by vect_model_simple_cost. */ |
6872 | n = 0; |
6873 | break; |
6874 | } |
6875 | if (n != 0) |
6876 | { |
6877 | /* We also need to materialize two large constants. */ |
6878 | record_stmt_cost (body_cost_vec: cost_vec, count: 2, kind: scalar_stmt, stmt_info, |
6879 | misalign: 0, where: vect_prologue); |
6880 | record_stmt_cost (body_cost_vec: cost_vec, count: n, kind: scalar_stmt, stmt_info, |
6881 | misalign: 0, where: vect_body); |
6882 | } |
6883 | } |
6884 | return true; |
6885 | } |
6886 | |
6887 | /* Transform. */ |
6888 | |
6889 | if (dump_enabled_p ()) |
6890 | dump_printf_loc (MSG_NOTE, vect_location, |
6891 | "transform binary/unary operation.\n" ); |
6892 | |
6893 | bool masked_loop_p = loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo); |
6894 | bool len_loop_p = loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo); |
6895 | |
6896 | /* POINTER_DIFF_EXPR has pointer arguments which are vectorized as |
6897 | vectors with unsigned elements, but the result is signed. So, we |
6898 | need to compute the MINUS_EXPR into vectype temporary and |
6899 | VIEW_CONVERT_EXPR it into the final vectype_out result. */ |
6900 | tree vec_cvt_dest = NULL_TREE; |
6901 | if (orig_code == POINTER_DIFF_EXPR) |
6902 | { |
6903 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
6904 | vec_cvt_dest = vect_create_destination_var (scalar_dest, vectype_out); |
6905 | } |
6906 | /* Handle def. */ |
6907 | else |
6908 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); |
6909 | |
6910 | /* In case the vectorization factor (VF) is bigger than the number |
6911 | of elements that we can fit in a vectype (nunits), we have to generate |
6912 | more than one vector stmt - i.e - we need to "unroll" the |
6913 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
6914 | from one copy of the vector stmt to the next, in the field |
6915 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
6916 | stages to find the correct vector defs to be used when vectorizing |
6917 | stmts that use the defs of the current stmt. The example below |
6918 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., |
6919 | we need to create 4 vectorized stmts): |
6920 | |
6921 | before vectorization: |
6922 | RELATED_STMT VEC_STMT |
6923 | S1: x = memref - - |
6924 | S2: z = x + 1 - - |
6925 | |
6926 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details |
6927 | there): |
6928 | RELATED_STMT VEC_STMT |
6929 | VS1_0: vx0 = memref0 VS1_1 - |
6930 | VS1_1: vx1 = memref1 VS1_2 - |
6931 | VS1_2: vx2 = memref2 VS1_3 - |
6932 | VS1_3: vx3 = memref3 - - |
6933 | S1: x = load - VS1_0 |
6934 | S2: z = x + 1 - - |
6935 | |
6936 | step2: vectorize stmt S2 (done here): |
6937 | To vectorize stmt S2 we first need to find the relevant vector |
6938 | def for the first operand 'x'. This is, as usual, obtained from |
6939 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt |
6940 | that defines 'x' (S1). This way we find the stmt VS1_0, and the |
6941 | relevant vector def 'vx0'. Having found 'vx0' we can generate |
6942 | the vector stmt VS2_0, and as usual, record it in the |
6943 | STMT_VINFO_VEC_STMT of stmt S2. |
6944 | When creating the second copy (VS2_1), we obtain the relevant vector |
6945 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of |
6946 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant |
6947 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a |
6948 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. |
6949 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting |
6950 | chain of stmts and pointers: |
6951 | RELATED_STMT VEC_STMT |
6952 | VS1_0: vx0 = memref0 VS1_1 - |
6953 | VS1_1: vx1 = memref1 VS1_2 - |
6954 | VS1_2: vx2 = memref2 VS1_3 - |
6955 | VS1_3: vx3 = memref3 - - |
6956 | S1: x = load - VS1_0 |
6957 | VS2_0: vz0 = vx0 + v1 VS2_1 - |
6958 | VS2_1: vz1 = vx1 + v1 VS2_2 - |
6959 | VS2_2: vz2 = vx2 + v1 VS2_3 - |
6960 | VS2_3: vz3 = vx3 + v1 - - |
6961 | S2: z = x + 1 - VS2_0 */ |
6962 | |
6963 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, |
6964 | op0, vec_oprnds0: &vec_oprnds0, op1, vec_oprnds1: &vec_oprnds1, op2, vec_oprnds2: &vec_oprnds2); |
6965 | /* Arguments are ready. Create the new vector stmt. */ |
6966 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
6967 | { |
6968 | gimple *new_stmt = NULL; |
6969 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
6970 | ? vec_oprnds1[i] : NULL_TREE); |
6971 | vop2 = ((op_type == ternary_op) ? vec_oprnds2[i] : NULL_TREE); |
6972 | if (using_emulated_vectors_p |
6973 | && (code == PLUS_EXPR || code == MINUS_EXPR || code == NEGATE_EXPR)) |
6974 | { |
6975 | /* Lower the operation. This follows vector lowering. */ |
6976 | unsigned int width = vector_element_bits (vectype); |
6977 | tree inner_type = TREE_TYPE (vectype); |
6978 | tree word_type |
6979 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (mode: word_mode), 1); |
6980 | HOST_WIDE_INT max = GET_MODE_MASK (TYPE_MODE (inner_type)); |
6981 | tree low_bits = build_replicated_int_cst (word_type, width, max >> 1); |
6982 | tree high_bits |
6983 | = build_replicated_int_cst (word_type, width, max & ~(max >> 1)); |
6984 | tree wvop0 = make_ssa_name (var: word_type); |
6985 | new_stmt = gimple_build_assign (wvop0, VIEW_CONVERT_EXPR, |
6986 | build1 (VIEW_CONVERT_EXPR, |
6987 | word_type, vop0)); |
6988 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6989 | tree result_low, signs; |
6990 | if (code == PLUS_EXPR || code == MINUS_EXPR) |
6991 | { |
6992 | tree wvop1 = make_ssa_name (var: word_type); |
6993 | new_stmt = gimple_build_assign (wvop1, VIEW_CONVERT_EXPR, |
6994 | build1 (VIEW_CONVERT_EXPR, |
6995 | word_type, vop1)); |
6996 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6997 | signs = make_ssa_name (var: word_type); |
6998 | new_stmt = gimple_build_assign (signs, |
6999 | BIT_XOR_EXPR, wvop0, wvop1); |
7000 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
7001 | tree b_low = make_ssa_name (var: word_type); |
7002 | new_stmt = gimple_build_assign (b_low, |
7003 | BIT_AND_EXPR, wvop1, low_bits); |
7004 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
7005 | tree a_low = make_ssa_name (var: word_type); |
7006 | if (code == PLUS_EXPR) |
7007 | new_stmt = gimple_build_assign (a_low, |
7008 | BIT_AND_EXPR, wvop0, low_bits); |
7009 | else |
7010 | new_stmt = gimple_build_assign (a_low, |
7011 | BIT_IOR_EXPR, wvop0, high_bits); |
7012 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
7013 | if (code == MINUS_EXPR) |
7014 | { |
7015 | new_stmt = gimple_build_assign (NULL_TREE, |
7016 | BIT_NOT_EXPR, signs); |
7017 | signs = make_ssa_name (var: word_type); |
7018 | gimple_assign_set_lhs (gs: new_stmt, lhs: signs); |
7019 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
7020 | } |
7021 | new_stmt = gimple_build_assign (NULL_TREE, |
7022 | BIT_AND_EXPR, signs, high_bits); |
7023 | signs = make_ssa_name (var: word_type); |
7024 | gimple_assign_set_lhs (gs: new_stmt, lhs: signs); |
7025 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
7026 | result_low = make_ssa_name (var: word_type); |
7027 | new_stmt = gimple_build_assign (result_low, code, a_low, b_low); |
7028 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
7029 | } |
7030 | else |
7031 | { |
7032 | tree a_low = make_ssa_name (var: word_type); |
7033 | new_stmt = gimple_build_assign (a_low, |
7034 | BIT_AND_EXPR, wvop0, low_bits); |
7035 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
7036 | signs = make_ssa_name (var: word_type); |
7037 | new_stmt = gimple_build_assign (signs, BIT_NOT_EXPR, wvop0); |
7038 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
7039 | new_stmt = gimple_build_assign (NULL_TREE, |
7040 | BIT_AND_EXPR, signs, high_bits); |
7041 | signs = make_ssa_name (var: word_type); |
7042 | gimple_assign_set_lhs (gs: new_stmt, lhs: signs); |
7043 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
7044 | result_low = make_ssa_name (var: word_type); |
7045 | new_stmt = gimple_build_assign (result_low, |
7046 | MINUS_EXPR, high_bits, a_low); |
7047 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
7048 | } |
7049 | new_stmt = gimple_build_assign (NULL_TREE, BIT_XOR_EXPR, result_low, |
7050 | signs); |
7051 | result_low = make_ssa_name (var: word_type); |
7052 | gimple_assign_set_lhs (gs: new_stmt, lhs: result_low); |
7053 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
7054 | new_stmt = gimple_build_assign (NULL_TREE, VIEW_CONVERT_EXPR, |
7055 | build1 (VIEW_CONVERT_EXPR, |
7056 | vectype, result_low)); |
7057 | new_temp = make_ssa_name (var: vectype); |
7058 | gimple_assign_set_lhs (gs: new_stmt, lhs: new_temp); |
7059 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
7060 | } |
7061 | else if ((masked_loop_p || len_loop_p) && mask_out_inactive) |
7062 | { |
7063 | tree mask; |
7064 | if (masked_loop_p) |
7065 | mask = vect_get_loop_mask (loop_vinfo, gsi, masks, |
7066 | vec_num * ncopies, vectype, i); |
7067 | else |
7068 | /* Dummy mask. */ |
7069 | mask = build_minus_one_cst (truth_type_for (vectype)); |
7070 | auto_vec<tree> vops (6); |
7071 | vops.quick_push (obj: mask); |
7072 | vops.quick_push (obj: vop0); |
7073 | if (vop1) |
7074 | vops.quick_push (obj: vop1); |
7075 | if (vop2) |
7076 | vops.quick_push (obj: vop2); |
7077 | if (reduc_idx >= 0) |
7078 | { |
7079 | /* Perform the operation on active elements only and take |
7080 | inactive elements from the reduction chain input. */ |
7081 | gcc_assert (!vop2); |
7082 | vops.quick_push (obj: reduc_idx == 1 ? vop1 : vop0); |
7083 | } |
7084 | else |
7085 | { |
7086 | auto else_value = targetm.preferred_else_value |
7087 | (cond_fn, vectype, vops.length () - 1, &vops[1]); |
7088 | vops.quick_push (obj: else_value); |
7089 | } |
7090 | if (len_loop_p) |
7091 | { |
7092 | tree len = vect_get_loop_len (loop_vinfo, gsi, lens, |
7093 | vec_num * ncopies, vectype, i, 1); |
7094 | signed char biasval |
7095 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
7096 | tree bias = build_int_cst (intQI_type_node, biasval); |
7097 | vops.quick_push (obj: len); |
7098 | vops.quick_push (obj: bias); |
7099 | } |
7100 | gcall *call |
7101 | = gimple_build_call_internal_vec (masked_loop_p ? cond_fn |
7102 | : cond_len_fn, |
7103 | vops); |
7104 | new_temp = make_ssa_name (var: vec_dest, stmt: call); |
7105 | gimple_call_set_lhs (gs: call, lhs: new_temp); |
7106 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
7107 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
7108 | new_stmt = call; |
7109 | } |
7110 | else |
7111 | { |
7112 | tree mask = NULL_TREE; |
7113 | /* When combining two masks check if either of them is elsewhere |
7114 | combined with a loop mask, if that's the case we can mark that the |
7115 | new combined mask doesn't need to be combined with a loop mask. */ |
7116 | if (masked_loop_p |
7117 | && code == BIT_AND_EXPR |
7118 | && VECTOR_BOOLEAN_TYPE_P (vectype)) |
7119 | { |
7120 | if (loop_vinfo->scalar_cond_masked_set.contains (k: { op0, |
7121 | ncopies})) |
7122 | { |
7123 | mask = vect_get_loop_mask (loop_vinfo, gsi, masks, |
7124 | vec_num * ncopies, vectype, i); |
7125 | |
7126 | vop0 = prepare_vec_mask (loop_vinfo, TREE_TYPE (mask), loop_mask: mask, |
7127 | vec_mask: vop0, gsi); |
7128 | } |
7129 | |
7130 | if (loop_vinfo->scalar_cond_masked_set.contains (k: { op1, |
7131 | ncopies })) |
7132 | { |
7133 | mask = vect_get_loop_mask (loop_vinfo, gsi, masks, |
7134 | vec_num * ncopies, vectype, i); |
7135 | |
7136 | vop1 = prepare_vec_mask (loop_vinfo, TREE_TYPE (mask), loop_mask: mask, |
7137 | vec_mask: vop1, gsi); |
7138 | } |
7139 | } |
7140 | |
7141 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
7142 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
7143 | gimple_assign_set_lhs (gs: new_stmt, lhs: new_temp); |
7144 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
7145 | if (using_emulated_vectors_p) |
7146 | suppress_warning (new_stmt, OPT_Wvector_operation_performance); |
7147 | |
7148 | /* Enter the combined value into the vector cond hash so we don't |
7149 | AND it with a loop mask again. */ |
7150 | if (mask) |
7151 | loop_vinfo->vec_cond_masked_set.add (k: { new_temp, mask }); |
7152 | } |
7153 | |
7154 | if (vec_cvt_dest) |
7155 | { |
7156 | new_temp = build1 (VIEW_CONVERT_EXPR, vectype_out, new_temp); |
7157 | new_stmt = gimple_build_assign (vec_cvt_dest, VIEW_CONVERT_EXPR, |
7158 | new_temp); |
7159 | new_temp = make_ssa_name (var: vec_cvt_dest, stmt: new_stmt); |
7160 | gimple_assign_set_lhs (gs: new_stmt, lhs: new_temp); |
7161 | vect_finish_stmt_generation (vinfo, stmt_info, |
7162 | vec_stmt: new_stmt, gsi); |
7163 | } |
7164 | |
7165 | if (slp_node) |
7166 | slp_node->push_vec_def (def: new_stmt); |
7167 | else |
7168 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
7169 | } |
7170 | |
7171 | if (!slp_node) |
7172 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
7173 | |
7174 | vec_oprnds0.release (); |
7175 | vec_oprnds1.release (); |
7176 | vec_oprnds2.release (); |
7177 | |
7178 | return true; |
7179 | } |
7180 | |
7181 | /* A helper function to ensure data reference DR_INFO's base alignment. */ |
7182 | |
7183 | static void |
7184 | ensure_base_align (dr_vec_info *dr_info) |
7185 | { |
7186 | /* Alignment is only analyzed for the first element of a DR group, |
7187 | use that to look at base alignment we need to enforce. */ |
7188 | if (STMT_VINFO_GROUPED_ACCESS (dr_info->stmt)) |
7189 | dr_info = STMT_VINFO_DR_INFO (DR_GROUP_FIRST_ELEMENT (dr_info->stmt)); |
7190 | |
7191 | gcc_assert (dr_info->misalignment != DR_MISALIGNMENT_UNINITIALIZED); |
7192 | |
7193 | if (dr_info->base_misaligned) |
7194 | { |
7195 | tree base_decl = dr_info->base_decl; |
7196 | |
7197 | // We should only be able to increase the alignment of a base object if |
7198 | // we know what its new alignment should be at compile time. |
7199 | unsigned HOST_WIDE_INT align_base_to = |
7200 | DR_TARGET_ALIGNMENT (dr_info).to_constant () * BITS_PER_UNIT; |
7201 | |
7202 | if (decl_in_symtab_p (decl: base_decl)) |
7203 | symtab_node::get (decl: base_decl)->increase_alignment (align: align_base_to); |
7204 | else if (DECL_ALIGN (base_decl) < align_base_to) |
7205 | { |
7206 | SET_DECL_ALIGN (base_decl, align_base_to); |
7207 | DECL_USER_ALIGN (base_decl) = 1; |
7208 | } |
7209 | dr_info->base_misaligned = false; |
7210 | } |
7211 | } |
7212 | |
7213 | |
7214 | /* Function get_group_alias_ptr_type. |
7215 | |
7216 | Return the alias type for the group starting at FIRST_STMT_INFO. */ |
7217 | |
7218 | static tree |
7219 | get_group_alias_ptr_type (stmt_vec_info first_stmt_info) |
7220 | { |
7221 | struct data_reference *first_dr, *next_dr; |
7222 | |
7223 | first_dr = STMT_VINFO_DATA_REF (first_stmt_info); |
7224 | stmt_vec_info next_stmt_info = DR_GROUP_NEXT_ELEMENT (first_stmt_info); |
7225 | while (next_stmt_info) |
7226 | { |
7227 | next_dr = STMT_VINFO_DATA_REF (next_stmt_info); |
7228 | if (get_alias_set (DR_REF (first_dr)) |
7229 | != get_alias_set (DR_REF (next_dr))) |
7230 | { |
7231 | if (dump_enabled_p ()) |
7232 | dump_printf_loc (MSG_NOTE, vect_location, |
7233 | "conflicting alias set types.\n" ); |
7234 | return ptr_type_node; |
7235 | } |
7236 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
7237 | } |
7238 | return reference_alias_ptr_type (DR_REF (first_dr)); |
7239 | } |
7240 | |
7241 | |
7242 | /* Function scan_operand_equal_p. |
7243 | |
7244 | Helper function for check_scan_store. Compare two references |
7245 | with .GOMP_SIMD_LANE bases. */ |
7246 | |
7247 | static bool |
7248 | scan_operand_equal_p (tree ref1, tree ref2) |
7249 | { |
7250 | tree ref[2] = { ref1, ref2 }; |
7251 | poly_int64 bitsize[2], bitpos[2]; |
7252 | tree offset[2], base[2]; |
7253 | for (int i = 0; i < 2; ++i) |
7254 | { |
7255 | machine_mode mode; |
7256 | int unsignedp, reversep, volatilep = 0; |
7257 | base[i] = get_inner_reference (ref[i], &bitsize[i], &bitpos[i], |
7258 | &offset[i], &mode, &unsignedp, |
7259 | &reversep, &volatilep); |
7260 | if (reversep || volatilep || maybe_ne (a: bitpos[i], b: 0)) |
7261 | return false; |
7262 | if (TREE_CODE (base[i]) == MEM_REF |
7263 | && offset[i] == NULL_TREE |
7264 | && TREE_CODE (TREE_OPERAND (base[i], 0)) == SSA_NAME) |
7265 | { |
7266 | gimple *def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (base[i], 0)); |
7267 | if (is_gimple_assign (gs: def_stmt) |
7268 | && gimple_assign_rhs_code (gs: def_stmt) == POINTER_PLUS_EXPR |
7269 | && TREE_CODE (gimple_assign_rhs1 (def_stmt)) == ADDR_EXPR |
7270 | && TREE_CODE (gimple_assign_rhs2 (def_stmt)) == SSA_NAME) |
7271 | { |
7272 | if (maybe_ne (a: mem_ref_offset (base[i]), b: 0)) |
7273 | return false; |
7274 | base[i] = TREE_OPERAND (gimple_assign_rhs1 (def_stmt), 0); |
7275 | offset[i] = gimple_assign_rhs2 (gs: def_stmt); |
7276 | } |
7277 | } |
7278 | } |
7279 | |
7280 | if (!operand_equal_p (base[0], base[1], flags: 0)) |
7281 | return false; |
7282 | if (maybe_ne (a: bitsize[0], b: bitsize[1])) |
7283 | return false; |
7284 | if (offset[0] != offset[1]) |
7285 | { |
7286 | if (!offset[0] || !offset[1]) |
7287 | return false; |
7288 | if (!operand_equal_p (offset[0], offset[1], flags: 0)) |
7289 | { |
7290 | tree step[2]; |
7291 | for (int i = 0; i < 2; ++i) |
7292 | { |
7293 | step[i] = integer_one_node; |
7294 | if (TREE_CODE (offset[i]) == SSA_NAME) |
7295 | { |
7296 | gimple *def_stmt = SSA_NAME_DEF_STMT (offset[i]); |
7297 | if (is_gimple_assign (gs: def_stmt) |
7298 | && gimple_assign_rhs_code (gs: def_stmt) == MULT_EXPR |
7299 | && (TREE_CODE (gimple_assign_rhs2 (def_stmt)) |
7300 | == INTEGER_CST)) |
7301 | { |
7302 | step[i] = gimple_assign_rhs2 (gs: def_stmt); |
7303 | offset[i] = gimple_assign_rhs1 (gs: def_stmt); |
7304 | } |
7305 | } |
7306 | else if (TREE_CODE (offset[i]) == MULT_EXPR) |
7307 | { |
7308 | step[i] = TREE_OPERAND (offset[i], 1); |
7309 | offset[i] = TREE_OPERAND (offset[i], 0); |
7310 | } |
7311 | tree rhs1 = NULL_TREE; |
7312 | if (TREE_CODE (offset[i]) == SSA_NAME) |
7313 | { |
7314 | gimple *def_stmt = SSA_NAME_DEF_STMT (offset[i]); |
7315 | if (gimple_assign_cast_p (s: def_stmt)) |
7316 | rhs1 = gimple_assign_rhs1 (gs: def_stmt); |
7317 | } |
7318 | else if (CONVERT_EXPR_P (offset[i])) |
7319 | rhs1 = TREE_OPERAND (offset[i], 0); |
7320 | if (rhs1 |
7321 | && INTEGRAL_TYPE_P (TREE_TYPE (rhs1)) |
7322 | && INTEGRAL_TYPE_P (TREE_TYPE (offset[i])) |
7323 | && (TYPE_PRECISION (TREE_TYPE (offset[i])) |
7324 | >= TYPE_PRECISION (TREE_TYPE (rhs1)))) |
7325 | offset[i] = rhs1; |
7326 | } |
7327 | if (!operand_equal_p (offset[0], offset[1], flags: 0) |
7328 | || !operand_equal_p (step[0], step[1], flags: 0)) |
7329 | return false; |
7330 | } |
7331 | } |
7332 | return true; |
7333 | } |
7334 | |
7335 | |
7336 | enum scan_store_kind { |
7337 | /* Normal permutation. */ |
7338 | scan_store_kind_perm, |
7339 | |
7340 | /* Whole vector left shift permutation with zero init. */ |
7341 | scan_store_kind_lshift_zero, |
7342 | |
7343 | /* Whole vector left shift permutation and VEC_COND_EXPR. */ |
7344 | scan_store_kind_lshift_cond |
7345 | }; |
7346 | |
7347 | /* Function check_scan_store. |
7348 | |
7349 | Verify if we can perform the needed permutations or whole vector shifts. |
7350 | Return -1 on failure, otherwise exact log2 of vectype's nunits. |
7351 | USE_WHOLE_VECTOR is a vector of enum scan_store_kind which operation |
7352 | to do at each step. */ |
7353 | |
7354 | static int |
7355 | scan_store_can_perm_p (tree vectype, tree init, |
7356 | vec<enum scan_store_kind> *use_whole_vector = NULL) |
7357 | { |
7358 | enum machine_mode vec_mode = TYPE_MODE (vectype); |
7359 | unsigned HOST_WIDE_INT nunits; |
7360 | if (!TYPE_VECTOR_SUBPARTS (node: vectype).is_constant (const_value: &nunits)) |
7361 | return -1; |
7362 | int units_log2 = exact_log2 (x: nunits); |
7363 | if (units_log2 <= 0) |
7364 | return -1; |
7365 | |
7366 | int i; |
7367 | enum scan_store_kind whole_vector_shift_kind = scan_store_kind_perm; |
7368 | for (i = 0; i <= units_log2; ++i) |
7369 | { |
7370 | unsigned HOST_WIDE_INT j, k; |
7371 | enum scan_store_kind kind = scan_store_kind_perm; |
7372 | vec_perm_builder sel (nunits, nunits, 1); |
7373 | sel.quick_grow (len: nunits); |
7374 | if (i == units_log2) |
7375 | { |
7376 | for (j = 0; j < nunits; ++j) |
7377 | sel[j] = nunits - 1; |
7378 | } |
7379 | else |
7380 | { |
7381 | for (j = 0; j < (HOST_WIDE_INT_1U << i); ++j) |
7382 | sel[j] = j; |
7383 | for (k = 0; j < nunits; ++j, ++k) |
7384 | sel[j] = nunits + k; |
7385 | } |
7386 | vec_perm_indices indices (sel, i == units_log2 ? 1 : 2, nunits); |
7387 | if (!can_vec_perm_const_p (vec_mode, vec_mode, indices)) |
7388 | { |
7389 | if (i == units_log2) |
7390 | return -1; |
7391 | |
7392 | if (whole_vector_shift_kind == scan_store_kind_perm) |
7393 | { |
7394 | if (optab_handler (op: vec_shl_optab, mode: vec_mode) == CODE_FOR_nothing) |
7395 | return -1; |
7396 | whole_vector_shift_kind = scan_store_kind_lshift_zero; |
7397 | /* Whole vector shifts shift in zeros, so if init is all zero |
7398 | constant, there is no need to do anything further. */ |
7399 | if ((TREE_CODE (init) != INTEGER_CST |
7400 | && TREE_CODE (init) != REAL_CST) |
7401 | || !initializer_zerop (init)) |
7402 | { |
7403 | tree masktype = truth_type_for (vectype); |
7404 | if (!expand_vec_cond_expr_p (vectype, masktype, VECTOR_CST)) |
7405 | return -1; |
7406 | whole_vector_shift_kind = scan_store_kind_lshift_cond; |
7407 | } |
7408 | } |
7409 | kind = whole_vector_shift_kind; |
7410 | } |
7411 | if (use_whole_vector) |
7412 | { |
7413 | if (kind != scan_store_kind_perm && use_whole_vector->is_empty ()) |
7414 | use_whole_vector->safe_grow_cleared (len: i, exact: true); |
7415 | if (kind != scan_store_kind_perm || !use_whole_vector->is_empty ()) |
7416 | use_whole_vector->safe_push (obj: kind); |
7417 | } |
7418 | } |
7419 | |
7420 | return units_log2; |
7421 | } |
7422 | |
7423 | |
7424 | /* Function check_scan_store. |
7425 | |
7426 | Check magic stores for #pragma omp scan {in,ex}clusive reductions. */ |
7427 | |
7428 | static bool |
7429 | check_scan_store (vec_info *vinfo, stmt_vec_info stmt_info, tree vectype, |
7430 | enum vect_def_type rhs_dt, bool slp, tree mask, |
7431 | vect_memory_access_type memory_access_type) |
7432 | { |
7433 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
7434 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info); |
7435 | tree ref_type; |
7436 | |
7437 | gcc_assert (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) > 1); |
7438 | if (slp |
7439 | || mask |
7440 | || memory_access_type != VMAT_CONTIGUOUS |
7441 | || TREE_CODE (DR_BASE_ADDRESS (dr_info->dr)) != ADDR_EXPR |
7442 | || !VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (dr_info->dr), 0)) |
7443 | || loop_vinfo == NULL |
7444 | || LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) |
7445 | || STMT_VINFO_GROUPED_ACCESS (stmt_info) |
7446 | || !integer_zerop (get_dr_vinfo_offset (vinfo, dr_info)) |
7447 | || !integer_zerop (DR_INIT (dr_info->dr)) |
7448 | || !(ref_type = reference_alias_ptr_type (DR_REF (dr_info->dr))) |
7449 | || !alias_sets_conflict_p (get_alias_set (vectype), |
7450 | get_alias_set (TREE_TYPE (ref_type)))) |
7451 | { |
7452 | if (dump_enabled_p ()) |
7453 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7454 | "unsupported OpenMP scan store.\n" ); |
7455 | return false; |
7456 | } |
7457 | |
7458 | /* We need to pattern match code built by OpenMP lowering and simplified |
7459 | by following optimizations into something we can handle. |
7460 | #pragma omp simd reduction(inscan,+:r) |
7461 | for (...) |
7462 | { |
7463 | r += something (); |
7464 | #pragma omp scan inclusive (r) |
7465 | use (r); |
7466 | } |
7467 | shall have body with: |
7468 | // Initialization for input phase, store the reduction initializer: |
7469 | _20 = .GOMP_SIMD_LANE (simduid.3_14(D), 0); |
7470 | _21 = .GOMP_SIMD_LANE (simduid.3_14(D), 1); |
7471 | D.2042[_21] = 0; |
7472 | // Actual input phase: |
7473 | ... |
7474 | r.0_5 = D.2042[_20]; |
7475 | _6 = _4 + r.0_5; |
7476 | D.2042[_20] = _6; |
7477 | // Initialization for scan phase: |
7478 | _25 = .GOMP_SIMD_LANE (simduid.3_14(D), 2); |
7479 | _26 = D.2043[_25]; |
7480 | _27 = D.2042[_25]; |
7481 | _28 = _26 + _27; |
7482 | D.2043[_25] = _28; |
7483 | D.2042[_25] = _28; |
7484 | // Actual scan phase: |
7485 | ... |
7486 | r.1_8 = D.2042[_20]; |
7487 | ... |
7488 | The "omp simd array" variable D.2042 holds the privatized copy used |
7489 | inside of the loop and D.2043 is another one that holds copies of |
7490 | the current original list item. The separate GOMP_SIMD_LANE ifn |
7491 | kinds are there in order to allow optimizing the initializer store |
7492 | and combiner sequence, e.g. if it is originally some C++ish user |
7493 | defined reduction, but allow the vectorizer to pattern recognize it |
7494 | and turn into the appropriate vectorized scan. |
7495 | |
7496 | For exclusive scan, this is slightly different: |
7497 | #pragma omp simd reduction(inscan,+:r) |
7498 | for (...) |
7499 | { |
7500 | use (r); |
7501 | #pragma omp scan exclusive (r) |
7502 | r += something (); |
7503 | } |
7504 | shall have body with: |
7505 | // Initialization for input phase, store the reduction initializer: |
7506 | _20 = .GOMP_SIMD_LANE (simduid.3_14(D), 0); |
7507 | _21 = .GOMP_SIMD_LANE (simduid.3_14(D), 1); |
7508 | D.2042[_21] = 0; |
7509 | // Actual input phase: |
7510 | ... |
7511 | r.0_5 = D.2042[_20]; |
7512 | _6 = _4 + r.0_5; |
7513 | D.2042[_20] = _6; |
7514 | // Initialization for scan phase: |
7515 | _25 = .GOMP_SIMD_LANE (simduid.3_14(D), 3); |
7516 | _26 = D.2043[_25]; |
7517 | D.2044[_25] = _26; |
7518 | _27 = D.2042[_25]; |
7519 | _28 = _26 + _27; |
7520 | D.2043[_25] = _28; |
7521 | // Actual scan phase: |
7522 | ... |
7523 | r.1_8 = D.2044[_20]; |
7524 | ... */ |
7525 | |
7526 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 2) |
7527 | { |
7528 | /* Match the D.2042[_21] = 0; store above. Just require that |
7529 | it is a constant or external definition store. */ |
7530 | if (rhs_dt != vect_constant_def && rhs_dt != vect_external_def) |
7531 | { |
7532 | fail_init: |
7533 | if (dump_enabled_p ()) |
7534 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7535 | "unsupported OpenMP scan initializer store.\n" ); |
7536 | return false; |
7537 | } |
7538 | |
7539 | if (! loop_vinfo->scan_map) |
7540 | loop_vinfo->scan_map = new hash_map<tree, tree>; |
7541 | tree var = TREE_OPERAND (DR_BASE_ADDRESS (dr_info->dr), 0); |
7542 | tree &cached = loop_vinfo->scan_map->get_or_insert (k: var); |
7543 | if (cached) |
7544 | goto fail_init; |
7545 | cached = gimple_assign_rhs1 (STMT_VINFO_STMT (stmt_info)); |
7546 | |
7547 | /* These stores can be vectorized normally. */ |
7548 | return true; |
7549 | } |
7550 | |
7551 | if (rhs_dt != vect_internal_def) |
7552 | { |
7553 | fail: |
7554 | if (dump_enabled_p ()) |
7555 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7556 | "unsupported OpenMP scan combiner pattern.\n" ); |
7557 | return false; |
7558 | } |
7559 | |
7560 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
7561 | tree rhs = gimple_assign_rhs1 (gs: stmt); |
7562 | if (TREE_CODE (rhs) != SSA_NAME) |
7563 | goto fail; |
7564 | |
7565 | gimple *other_store_stmt = NULL; |
7566 | tree var = TREE_OPERAND (DR_BASE_ADDRESS (dr_info->dr), 0); |
7567 | bool inscan_var_store |
7568 | = lookup_attribute (attr_name: "omp simd inscan" , DECL_ATTRIBUTES (var)) != NULL; |
7569 | |
7570 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4) |
7571 | { |
7572 | if (!inscan_var_store) |
7573 | { |
7574 | use_operand_p use_p; |
7575 | imm_use_iterator iter; |
7576 | FOR_EACH_IMM_USE_FAST (use_p, iter, rhs) |
7577 | { |
7578 | gimple *use_stmt = USE_STMT (use_p); |
7579 | if (use_stmt == stmt || is_gimple_debug (gs: use_stmt)) |
7580 | continue; |
7581 | if (gimple_bb (g: use_stmt) != gimple_bb (g: stmt) |
7582 | || !is_gimple_assign (gs: use_stmt) |
7583 | || gimple_assign_rhs_class (gs: use_stmt) != GIMPLE_BINARY_RHS |
7584 | || other_store_stmt |
7585 | || TREE_CODE (gimple_assign_lhs (use_stmt)) != SSA_NAME) |
7586 | goto fail; |
7587 | other_store_stmt = use_stmt; |
7588 | } |
7589 | if (other_store_stmt == NULL) |
7590 | goto fail; |
7591 | rhs = gimple_assign_lhs (gs: other_store_stmt); |
7592 | if (!single_imm_use (var: rhs, use_p: &use_p, stmt: &other_store_stmt)) |
7593 | goto fail; |
7594 | } |
7595 | } |
7596 | else if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 3) |
7597 | { |
7598 | use_operand_p use_p; |
7599 | imm_use_iterator iter; |
7600 | FOR_EACH_IMM_USE_FAST (use_p, iter, rhs) |
7601 | { |
7602 | gimple *use_stmt = USE_STMT (use_p); |
7603 | if (use_stmt == stmt || is_gimple_debug (gs: use_stmt)) |
7604 | continue; |
7605 | if (other_store_stmt) |
7606 | goto fail; |
7607 | other_store_stmt = use_stmt; |
7608 | } |
7609 | } |
7610 | else |
7611 | goto fail; |
7612 | |
7613 | gimple *def_stmt = SSA_NAME_DEF_STMT (rhs); |
7614 | if (gimple_bb (g: def_stmt) != gimple_bb (g: stmt) |
7615 | || !is_gimple_assign (gs: def_stmt) |
7616 | || gimple_assign_rhs_class (gs: def_stmt) != GIMPLE_BINARY_RHS) |
7617 | goto fail; |
7618 | |
7619 | enum tree_code code = gimple_assign_rhs_code (gs: def_stmt); |
7620 | /* For pointer addition, we should use the normal plus for the vector |
7621 | operation. */ |
7622 | switch (code) |
7623 | { |
7624 | case POINTER_PLUS_EXPR: |
7625 | code = PLUS_EXPR; |
7626 | break; |
7627 | case MULT_HIGHPART_EXPR: |
7628 | goto fail; |
7629 | default: |
7630 | break; |
7631 | } |
7632 | if (TREE_CODE_LENGTH (code) != binary_op || !commutative_tree_code (code)) |
7633 | goto fail; |
7634 | |
7635 | tree rhs1 = gimple_assign_rhs1 (gs: def_stmt); |
7636 | tree rhs2 = gimple_assign_rhs2 (gs: def_stmt); |
7637 | if (TREE_CODE (rhs1) != SSA_NAME || TREE_CODE (rhs2) != SSA_NAME) |
7638 | goto fail; |
7639 | |
7640 | gimple *load1_stmt = SSA_NAME_DEF_STMT (rhs1); |
7641 | gimple *load2_stmt = SSA_NAME_DEF_STMT (rhs2); |
7642 | if (gimple_bb (g: load1_stmt) != gimple_bb (g: stmt) |
7643 | || !gimple_assign_load_p (load1_stmt) |
7644 | || gimple_bb (g: load2_stmt) != gimple_bb (g: stmt) |
7645 | || !gimple_assign_load_p (load2_stmt)) |
7646 | goto fail; |
7647 | |
7648 | stmt_vec_info load1_stmt_info = loop_vinfo->lookup_stmt (load1_stmt); |
7649 | stmt_vec_info load2_stmt_info = loop_vinfo->lookup_stmt (load2_stmt); |
7650 | if (load1_stmt_info == NULL |
7651 | || load2_stmt_info == NULL |
7652 | || (STMT_VINFO_SIMD_LANE_ACCESS_P (load1_stmt_info) |
7653 | != STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info)) |
7654 | || (STMT_VINFO_SIMD_LANE_ACCESS_P (load2_stmt_info) |
7655 | != STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info))) |
7656 | goto fail; |
7657 | |
7658 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4 && inscan_var_store) |
7659 | { |
7660 | dr_vec_info *load1_dr_info = STMT_VINFO_DR_INFO (load1_stmt_info); |
7661 | if (TREE_CODE (DR_BASE_ADDRESS (load1_dr_info->dr)) != ADDR_EXPR |
7662 | || !VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (load1_dr_info->dr), 0))) |
7663 | goto fail; |
7664 | tree var1 = TREE_OPERAND (DR_BASE_ADDRESS (load1_dr_info->dr), 0); |
7665 | tree lrhs; |
7666 | if (lookup_attribute (attr_name: "omp simd inscan" , DECL_ATTRIBUTES (var1))) |
7667 | lrhs = rhs1; |
7668 | else |
7669 | lrhs = rhs2; |
7670 | use_operand_p use_p; |
7671 | imm_use_iterator iter; |
7672 | FOR_EACH_IMM_USE_FAST (use_p, iter, lrhs) |
7673 | { |
7674 | gimple *use_stmt = USE_STMT (use_p); |
7675 | if (use_stmt == def_stmt || is_gimple_debug (gs: use_stmt)) |
7676 | continue; |
7677 | if (other_store_stmt) |
7678 | goto fail; |
7679 | other_store_stmt = use_stmt; |
7680 | } |
7681 | } |
7682 | |
7683 | if (other_store_stmt == NULL) |
7684 | goto fail; |
7685 | if (gimple_bb (g: other_store_stmt) != gimple_bb (g: stmt) |
7686 | || !gimple_store_p (gs: other_store_stmt)) |
7687 | goto fail; |
7688 | |
7689 | stmt_vec_info other_store_stmt_info |
7690 | = loop_vinfo->lookup_stmt (other_store_stmt); |
7691 | if (other_store_stmt_info == NULL |
7692 | || (STMT_VINFO_SIMD_LANE_ACCESS_P (other_store_stmt_info) |
7693 | != STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info))) |
7694 | goto fail; |
7695 | |
7696 | gimple *stmt1 = stmt; |
7697 | gimple *stmt2 = other_store_stmt; |
7698 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4 && !inscan_var_store) |
7699 | std::swap (a&: stmt1, b&: stmt2); |
7700 | if (scan_operand_equal_p (ref1: gimple_assign_lhs (gs: stmt1), |
7701 | ref2: gimple_assign_rhs1 (gs: load2_stmt))) |
7702 | { |
7703 | std::swap (a&: rhs1, b&: rhs2); |
7704 | std::swap (a&: load1_stmt, b&: load2_stmt); |
7705 | std::swap (a&: load1_stmt_info, b&: load2_stmt_info); |
7706 | } |
7707 | if (!scan_operand_equal_p (ref1: gimple_assign_lhs (gs: stmt1), |
7708 | ref2: gimple_assign_rhs1 (gs: load1_stmt))) |
7709 | goto fail; |
7710 | |
7711 | tree var3 = NULL_TREE; |
7712 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 3 |
7713 | && !scan_operand_equal_p (ref1: gimple_assign_lhs (gs: stmt2), |
7714 | ref2: gimple_assign_rhs1 (gs: load2_stmt))) |
7715 | goto fail; |
7716 | else if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4) |
7717 | { |
7718 | dr_vec_info *load2_dr_info = STMT_VINFO_DR_INFO (load2_stmt_info); |
7719 | if (TREE_CODE (DR_BASE_ADDRESS (load2_dr_info->dr)) != ADDR_EXPR |
7720 | || !VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (load2_dr_info->dr), 0))) |
7721 | goto fail; |
7722 | var3 = TREE_OPERAND (DR_BASE_ADDRESS (load2_dr_info->dr), 0); |
7723 | if (!lookup_attribute (attr_name: "omp simd array" , DECL_ATTRIBUTES (var3)) |
7724 | || lookup_attribute (attr_name: "omp simd inscan" , DECL_ATTRIBUTES (var3)) |
7725 | || lookup_attribute (attr_name: "omp simd inscan exclusive" , |
7726 | DECL_ATTRIBUTES (var3))) |
7727 | goto fail; |
7728 | } |
7729 | |
7730 | dr_vec_info *other_dr_info = STMT_VINFO_DR_INFO (other_store_stmt_info); |
7731 | if (TREE_CODE (DR_BASE_ADDRESS (other_dr_info->dr)) != ADDR_EXPR |
7732 | || !VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (other_dr_info->dr), 0))) |
7733 | goto fail; |
7734 | |
7735 | tree var1 = TREE_OPERAND (DR_BASE_ADDRESS (dr_info->dr), 0); |
7736 | tree var2 = TREE_OPERAND (DR_BASE_ADDRESS (other_dr_info->dr), 0); |
7737 | if (!lookup_attribute (attr_name: "omp simd array" , DECL_ATTRIBUTES (var1)) |
7738 | || !lookup_attribute (attr_name: "omp simd array" , DECL_ATTRIBUTES (var2)) |
7739 | || (!lookup_attribute (attr_name: "omp simd inscan" , DECL_ATTRIBUTES (var1))) |
7740 | == (!lookup_attribute (attr_name: "omp simd inscan" , DECL_ATTRIBUTES (var2)))) |
7741 | goto fail; |
7742 | |
7743 | if (lookup_attribute (attr_name: "omp simd inscan" , DECL_ATTRIBUTES (var1))) |
7744 | std::swap (a&: var1, b&: var2); |
7745 | |
7746 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4) |
7747 | { |
7748 | if (!lookup_attribute (attr_name: "omp simd inscan exclusive" , |
7749 | DECL_ATTRIBUTES (var1))) |
7750 | goto fail; |
7751 | var1 = var3; |
7752 | } |
7753 | |
7754 | if (loop_vinfo->scan_map == NULL) |
7755 | goto fail; |
7756 | tree *init = loop_vinfo->scan_map->get (k: var1); |
7757 | if (init == NULL) |
7758 | goto fail; |
7759 | |
7760 | /* The IL is as expected, now check if we can actually vectorize it. |
7761 | Inclusive scan: |
7762 | _26 = D.2043[_25]; |
7763 | _27 = D.2042[_25]; |
7764 | _28 = _26 + _27; |
7765 | D.2043[_25] = _28; |
7766 | D.2042[_25] = _28; |
7767 | should be vectorized as (where _40 is the vectorized rhs |
7768 | from the D.2042[_21] = 0; store): |
7769 | _30 = MEM <vector(8) int> [(int *)&D.2043]; |
7770 | _31 = MEM <vector(8) int> [(int *)&D.2042]; |
7771 | _32 = VEC_PERM_EXPR <_40, _31, { 0, 8, 9, 10, 11, 12, 13, 14 }>; |
7772 | _33 = _31 + _32; |
7773 | // _33 = { _31[0], _31[0]+_31[1], _31[1]+_31[2], ..., _31[6]+_31[7] }; |
7774 | _34 = VEC_PERM_EXPR <_40, _33, { 0, 1, 8, 9, 10, 11, 12, 13 }>; |
7775 | _35 = _33 + _34; |
7776 | // _35 = { _31[0], _31[0]+_31[1], _31[0]+.._31[2], _31[0]+.._31[3], |
7777 | // _31[1]+.._31[4], ... _31[4]+.._31[7] }; |
7778 | _36 = VEC_PERM_EXPR <_40, _35, { 0, 1, 2, 3, 8, 9, 10, 11 }>; |
7779 | _37 = _35 + _36; |
7780 | // _37 = { _31[0], _31[0]+_31[1], _31[0]+.._31[2], _31[0]+.._31[3], |
7781 | // _31[0]+.._31[4], ... _31[0]+.._31[7] }; |
7782 | _38 = _30 + _37; |
7783 | _39 = VEC_PERM_EXPR <_38, _38, { 7, 7, 7, 7, 7, 7, 7, 7 }>; |
7784 | MEM <vector(8) int> [(int *)&D.2043] = _39; |
7785 | MEM <vector(8) int> [(int *)&D.2042] = _38; |
7786 | Exclusive scan: |
7787 | _26 = D.2043[_25]; |
7788 | D.2044[_25] = _26; |
7789 | _27 = D.2042[_25]; |
7790 | _28 = _26 + _27; |
7791 | D.2043[_25] = _28; |
7792 | should be vectorized as (where _40 is the vectorized rhs |
7793 | from the D.2042[_21] = 0; store): |
7794 | _30 = MEM <vector(8) int> [(int *)&D.2043]; |
7795 | _31 = MEM <vector(8) int> [(int *)&D.2042]; |
7796 | _32 = VEC_PERM_EXPR <_40, _31, { 0, 8, 9, 10, 11, 12, 13, 14 }>; |
7797 | _33 = VEC_PERM_EXPR <_40, _32, { 0, 8, 9, 10, 11, 12, 13, 14 }>; |
7798 | _34 = _32 + _33; |
7799 | // _34 = { 0, _31[0], _31[0]+_31[1], _31[1]+_31[2], _31[2]+_31[3], |
7800 | // _31[3]+_31[4], ... _31[5]+.._31[6] }; |
7801 | _35 = VEC_PERM_EXPR <_40, _34, { 0, 1, 8, 9, 10, 11, 12, 13 }>; |
7802 | _36 = _34 + _35; |
7803 | // _36 = { 0, _31[0], _31[0]+_31[1], _31[0]+.._31[2], _31[0]+.._31[3], |
7804 | // _31[1]+.._31[4], ... _31[3]+.._31[6] }; |
7805 | _37 = VEC_PERM_EXPR <_40, _36, { 0, 1, 2, 3, 8, 9, 10, 11 }>; |
7806 | _38 = _36 + _37; |
7807 | // _38 = { 0, _31[0], _31[0]+_31[1], _31[0]+.._31[2], _31[0]+.._31[3], |
7808 | // _31[0]+.._31[4], ... _31[0]+.._31[6] }; |
7809 | _39 = _30 + _38; |
7810 | _50 = _31 + _39; |
7811 | _51 = VEC_PERM_EXPR <_50, _50, { 7, 7, 7, 7, 7, 7, 7, 7 }>; |
7812 | MEM <vector(8) int> [(int *)&D.2044] = _39; |
7813 | MEM <vector(8) int> [(int *)&D.2042] = _51; */ |
7814 | enum machine_mode vec_mode = TYPE_MODE (vectype); |
7815 | optab optab = optab_for_tree_code (code, vectype, optab_default); |
7816 | if (!optab || optab_handler (op: optab, mode: vec_mode) == CODE_FOR_nothing) |
7817 | goto fail; |
7818 | |
7819 | int units_log2 = scan_store_can_perm_p (vectype, init: *init); |
7820 | if (units_log2 == -1) |
7821 | goto fail; |
7822 | |
7823 | return true; |
7824 | } |
7825 | |
7826 | |
7827 | /* Function vectorizable_scan_store. |
7828 | |
7829 | Helper of vectorizable_score, arguments like on vectorizable_store. |
7830 | Handle only the transformation, checking is done in check_scan_store. */ |
7831 | |
7832 | static bool |
7833 | vectorizable_scan_store (vec_info *vinfo, |
7834 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
7835 | gimple **vec_stmt, int ncopies) |
7836 | { |
7837 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
7838 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info); |
7839 | tree ref_type = reference_alias_ptr_type (DR_REF (dr_info->dr)); |
7840 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
7841 | |
7842 | if (dump_enabled_p ()) |
7843 | dump_printf_loc (MSG_NOTE, vect_location, |
7844 | "transform scan store. ncopies = %d\n" , ncopies); |
7845 | |
7846 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
7847 | tree rhs = gimple_assign_rhs1 (gs: stmt); |
7848 | gcc_assert (TREE_CODE (rhs) == SSA_NAME); |
7849 | |
7850 | tree var = TREE_OPERAND (DR_BASE_ADDRESS (dr_info->dr), 0); |
7851 | bool inscan_var_store |
7852 | = lookup_attribute (attr_name: "omp simd inscan" , DECL_ATTRIBUTES (var)) != NULL; |
7853 | |
7854 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4 && !inscan_var_store) |
7855 | { |
7856 | use_operand_p use_p; |
7857 | imm_use_iterator iter; |
7858 | FOR_EACH_IMM_USE_FAST (use_p, iter, rhs) |
7859 | { |
7860 | gimple *use_stmt = USE_STMT (use_p); |
7861 | if (use_stmt == stmt || is_gimple_debug (gs: use_stmt)) |
7862 | continue; |
7863 | rhs = gimple_assign_lhs (gs: use_stmt); |
7864 | break; |
7865 | } |
7866 | } |
7867 | |
7868 | gimple *def_stmt = SSA_NAME_DEF_STMT (rhs); |
7869 | enum tree_code code = gimple_assign_rhs_code (gs: def_stmt); |
7870 | if (code == POINTER_PLUS_EXPR) |
7871 | code = PLUS_EXPR; |
7872 | gcc_assert (TREE_CODE_LENGTH (code) == binary_op |
7873 | && commutative_tree_code (code)); |
7874 | tree rhs1 = gimple_assign_rhs1 (gs: def_stmt); |
7875 | tree rhs2 = gimple_assign_rhs2 (gs: def_stmt); |
7876 | gcc_assert (TREE_CODE (rhs1) == SSA_NAME && TREE_CODE (rhs2) == SSA_NAME); |
7877 | gimple *load1_stmt = SSA_NAME_DEF_STMT (rhs1); |
7878 | gimple *load2_stmt = SSA_NAME_DEF_STMT (rhs2); |
7879 | stmt_vec_info load1_stmt_info = loop_vinfo->lookup_stmt (load1_stmt); |
7880 | stmt_vec_info load2_stmt_info = loop_vinfo->lookup_stmt (load2_stmt); |
7881 | dr_vec_info *load1_dr_info = STMT_VINFO_DR_INFO (load1_stmt_info); |
7882 | dr_vec_info *load2_dr_info = STMT_VINFO_DR_INFO (load2_stmt_info); |
7883 | tree var1 = TREE_OPERAND (DR_BASE_ADDRESS (load1_dr_info->dr), 0); |
7884 | tree var2 = TREE_OPERAND (DR_BASE_ADDRESS (load2_dr_info->dr), 0); |
7885 | |
7886 | if (lookup_attribute (attr_name: "omp simd inscan" , DECL_ATTRIBUTES (var1))) |
7887 | { |
7888 | std::swap (a&: rhs1, b&: rhs2); |
7889 | std::swap (a&: var1, b&: var2); |
7890 | std::swap (a&: load1_dr_info, b&: load2_dr_info); |
7891 | } |
7892 | |
7893 | tree *init = loop_vinfo->scan_map->get (k: var1); |
7894 | gcc_assert (init); |
7895 | |
7896 | unsigned HOST_WIDE_INT nunits; |
7897 | if (!TYPE_VECTOR_SUBPARTS (node: vectype).is_constant (const_value: &nunits)) |
7898 | gcc_unreachable (); |
7899 | auto_vec<enum scan_store_kind, 16> use_whole_vector; |
7900 | int units_log2 = scan_store_can_perm_p (vectype, init: *init, use_whole_vector: &use_whole_vector); |
7901 | gcc_assert (units_log2 > 0); |
7902 | auto_vec<tree, 16> perms; |
7903 | perms.quick_grow (len: units_log2 + 1); |
7904 | tree zero_vec = NULL_TREE, masktype = NULL_TREE; |
7905 | for (int i = 0; i <= units_log2; ++i) |
7906 | { |
7907 | unsigned HOST_WIDE_INT j, k; |
7908 | vec_perm_builder sel (nunits, nunits, 1); |
7909 | sel.quick_grow (len: nunits); |
7910 | if (i == units_log2) |
7911 | for (j = 0; j < nunits; ++j) |
7912 | sel[j] = nunits - 1; |
7913 | else |
7914 | { |
7915 | for (j = 0; j < (HOST_WIDE_INT_1U << i); ++j) |
7916 | sel[j] = j; |
7917 | for (k = 0; j < nunits; ++j, ++k) |
7918 | sel[j] = nunits + k; |
7919 | } |
7920 | vec_perm_indices indices (sel, i == units_log2 ? 1 : 2, nunits); |
7921 | if (!use_whole_vector.is_empty () |
7922 | && use_whole_vector[i] != scan_store_kind_perm) |
7923 | { |
7924 | if (zero_vec == NULL_TREE) |
7925 | zero_vec = build_zero_cst (vectype); |
7926 | if (masktype == NULL_TREE |
7927 | && use_whole_vector[i] == scan_store_kind_lshift_cond) |
7928 | masktype = truth_type_for (vectype); |
7929 | perms[i] = vect_gen_perm_mask_any (vectype, indices); |
7930 | } |
7931 | else |
7932 | perms[i] = vect_gen_perm_mask_checked (vectype, indices); |
7933 | } |
7934 | |
7935 | tree vec_oprnd1 = NULL_TREE; |
7936 | tree vec_oprnd2 = NULL_TREE; |
7937 | tree vec_oprnd3 = NULL_TREE; |
7938 | tree dataref_ptr = DR_BASE_ADDRESS (dr_info->dr); |
7939 | tree dataref_offset = build_int_cst (ref_type, 0); |
7940 | tree bump = vect_get_data_ptr_increment (vinfo, gsi, dr_info, |
7941 | aggr_type: vectype, memory_access_type: VMAT_CONTIGUOUS); |
7942 | tree ldataref_ptr = NULL_TREE; |
7943 | tree orig = NULL_TREE; |
7944 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4 && !inscan_var_store) |
7945 | ldataref_ptr = DR_BASE_ADDRESS (load1_dr_info->dr); |
7946 | auto_vec<tree> vec_oprnds1; |
7947 | auto_vec<tree> vec_oprnds2; |
7948 | auto_vec<tree> vec_oprnds3; |
7949 | vect_get_vec_defs (vinfo, stmt_info, NULL, ncopies, |
7950 | op0: *init, vec_oprnds0: &vec_oprnds1, |
7951 | op1: ldataref_ptr == NULL ? rhs1 : NULL, vec_oprnds1: &vec_oprnds2, |
7952 | op2: rhs2, vec_oprnds2: &vec_oprnds3); |
7953 | for (int j = 0; j < ncopies; j++) |
7954 | { |
7955 | vec_oprnd1 = vec_oprnds1[j]; |
7956 | if (ldataref_ptr == NULL) |
7957 | vec_oprnd2 = vec_oprnds2[j]; |
7958 | vec_oprnd3 = vec_oprnds3[j]; |
7959 | if (j == 0) |
7960 | orig = vec_oprnd3; |
7961 | else if (!inscan_var_store) |
7962 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, bump); |
7963 | |
7964 | if (ldataref_ptr) |
7965 | { |
7966 | vec_oprnd2 = make_ssa_name (var: vectype); |
7967 | tree data_ref = fold_build2 (MEM_REF, vectype, |
7968 | unshare_expr (ldataref_ptr), |
7969 | dataref_offset); |
7970 | vect_copy_ref_info (data_ref, DR_REF (load1_dr_info->dr)); |
7971 | gimple *g = gimple_build_assign (vec_oprnd2, data_ref); |
7972 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
7973 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
7974 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
7975 | } |
7976 | |
7977 | tree v = vec_oprnd2; |
7978 | for (int i = 0; i < units_log2; ++i) |
7979 | { |
7980 | tree new_temp = make_ssa_name (var: vectype); |
7981 | gimple *g = gimple_build_assign (new_temp, VEC_PERM_EXPR, |
7982 | (zero_vec |
7983 | && (use_whole_vector[i] |
7984 | != scan_store_kind_perm)) |
7985 | ? zero_vec : vec_oprnd1, v, |
7986 | perms[i]); |
7987 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
7988 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
7989 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
7990 | |
7991 | if (zero_vec && use_whole_vector[i] == scan_store_kind_lshift_cond) |
7992 | { |
7993 | /* Whole vector shift shifted in zero bits, but if *init |
7994 | is not initializer_zerop, we need to replace those elements |
7995 | with elements from vec_oprnd1. */ |
7996 | tree_vector_builder vb (masktype, nunits, 1); |
7997 | for (unsigned HOST_WIDE_INT k = 0; k < nunits; ++k) |
7998 | vb.quick_push (obj: k < (HOST_WIDE_INT_1U << i) |
7999 | ? boolean_false_node : boolean_true_node); |
8000 | |
8001 | tree new_temp2 = make_ssa_name (var: vectype); |
8002 | g = gimple_build_assign (new_temp2, VEC_COND_EXPR, vb.build (), |
8003 | new_temp, vec_oprnd1); |
8004 | vect_finish_stmt_generation (vinfo, stmt_info, |
8005 | vec_stmt: g, gsi); |
8006 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
8007 | new_temp = new_temp2; |
8008 | } |
8009 | |
8010 | /* For exclusive scan, perform the perms[i] permutation once |
8011 | more. */ |
8012 | if (i == 0 |
8013 | && STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4 |
8014 | && v == vec_oprnd2) |
8015 | { |
8016 | v = new_temp; |
8017 | --i; |
8018 | continue; |
8019 | } |
8020 | |
8021 | tree new_temp2 = make_ssa_name (var: vectype); |
8022 | g = gimple_build_assign (new_temp2, code, v, new_temp); |
8023 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
8024 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
8025 | |
8026 | v = new_temp2; |
8027 | } |
8028 | |
8029 | tree new_temp = make_ssa_name (var: vectype); |
8030 | gimple *g = gimple_build_assign (new_temp, code, orig, v); |
8031 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
8032 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
8033 | |
8034 | tree last_perm_arg = new_temp; |
8035 | /* For exclusive scan, new_temp computed above is the exclusive scan |
8036 | prefix sum. Turn it into inclusive prefix sum for the broadcast |
8037 | of the last element into orig. */ |
8038 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4) |
8039 | { |
8040 | last_perm_arg = make_ssa_name (var: vectype); |
8041 | g = gimple_build_assign (last_perm_arg, code, new_temp, vec_oprnd2); |
8042 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
8043 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
8044 | } |
8045 | |
8046 | orig = make_ssa_name (var: vectype); |
8047 | g = gimple_build_assign (orig, VEC_PERM_EXPR, last_perm_arg, |
8048 | last_perm_arg, perms[units_log2]); |
8049 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
8050 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
8051 | |
8052 | if (!inscan_var_store) |
8053 | { |
8054 | tree data_ref = fold_build2 (MEM_REF, vectype, |
8055 | unshare_expr (dataref_ptr), |
8056 | dataref_offset); |
8057 | vect_copy_ref_info (data_ref, DR_REF (dr_info->dr)); |
8058 | g = gimple_build_assign (data_ref, new_temp); |
8059 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
8060 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
8061 | } |
8062 | } |
8063 | |
8064 | if (inscan_var_store) |
8065 | for (int j = 0; j < ncopies; j++) |
8066 | { |
8067 | if (j != 0) |
8068 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, bump); |
8069 | |
8070 | tree data_ref = fold_build2 (MEM_REF, vectype, |
8071 | unshare_expr (dataref_ptr), |
8072 | dataref_offset); |
8073 | vect_copy_ref_info (data_ref, DR_REF (dr_info->dr)); |
8074 | gimple *g = gimple_build_assign (data_ref, orig); |
8075 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
8076 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
8077 | } |
8078 | return true; |
8079 | } |
8080 | |
8081 | |
8082 | /* Function vectorizable_store. |
8083 | |
8084 | Check if STMT_INFO defines a non scalar data-ref (array/pointer/structure) |
8085 | that can be vectorized. |
8086 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
8087 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
8088 | Return true if STMT_INFO is vectorizable in this way. */ |
8089 | |
8090 | static bool |
8091 | vectorizable_store (vec_info *vinfo, |
8092 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
8093 | gimple **vec_stmt, slp_tree slp_node, |
8094 | stmt_vector_for_cost *cost_vec) |
8095 | { |
8096 | tree data_ref; |
8097 | tree vec_oprnd = NULL_TREE; |
8098 | tree elem_type; |
8099 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
8100 | class loop *loop = NULL; |
8101 | machine_mode vec_mode; |
8102 | tree dummy; |
8103 | enum vect_def_type rhs_dt = vect_unknown_def_type; |
8104 | enum vect_def_type mask_dt = vect_unknown_def_type; |
8105 | tree dataref_ptr = NULL_TREE; |
8106 | tree dataref_offset = NULL_TREE; |
8107 | gimple *ptr_incr = NULL; |
8108 | int ncopies; |
8109 | int j; |
8110 | stmt_vec_info first_stmt_info; |
8111 | bool grouped_store; |
8112 | unsigned int group_size, i; |
8113 | bool slp = (slp_node != NULL); |
8114 | unsigned int vec_num; |
8115 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
8116 | tree aggr_type; |
8117 | gather_scatter_info gs_info; |
8118 | poly_uint64 vf; |
8119 | vec_load_store_type vls_type; |
8120 | tree ref_type; |
8121 | |
8122 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
8123 | return false; |
8124 | |
8125 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
8126 | && ! vec_stmt) |
8127 | return false; |
8128 | |
8129 | /* Is vectorizable store? */ |
8130 | |
8131 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
8132 | slp_tree mask_node = NULL; |
8133 | if (gassign *assign = dyn_cast <gassign *> (p: stmt_info->stmt)) |
8134 | { |
8135 | tree scalar_dest = gimple_assign_lhs (gs: assign); |
8136 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
8137 | && is_pattern_stmt_p (stmt_info)) |
8138 | scalar_dest = TREE_OPERAND (scalar_dest, 0); |
8139 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
8140 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
8141 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
8142 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
8143 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR |
8144 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
8145 | && TREE_CODE (scalar_dest) != MEM_REF) |
8146 | return false; |
8147 | } |
8148 | else |
8149 | { |
8150 | gcall *call = dyn_cast <gcall *> (p: stmt_info->stmt); |
8151 | if (!call || !gimple_call_internal_p (gs: call)) |
8152 | return false; |
8153 | |
8154 | internal_fn ifn = gimple_call_internal_fn (gs: call); |
8155 | if (!internal_store_fn_p (ifn)) |
8156 | return false; |
8157 | |
8158 | int mask_index = internal_fn_mask_index (ifn); |
8159 | if (mask_index >= 0 && slp_node) |
8160 | mask_index = vect_slp_child_index_for_operand |
8161 | (call, op: mask_index, STMT_VINFO_GATHER_SCATTER_P (stmt_info)); |
8162 | if (mask_index >= 0 |
8163 | && !vect_check_scalar_mask (vinfo, stmt_info, slp_node, mask_index, |
8164 | mask: &mask, mask_node: &mask_node, mask_dt_out: &mask_dt, |
8165 | mask_vectype_out: &mask_vectype)) |
8166 | return false; |
8167 | } |
8168 | |
8169 | /* Cannot have hybrid store SLP -- that would mean storing to the |
8170 | same location twice. */ |
8171 | gcc_assert (slp == PURE_SLP_STMT (stmt_info)); |
8172 | |
8173 | tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE; |
8174 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
8175 | |
8176 | if (loop_vinfo) |
8177 | { |
8178 | loop = LOOP_VINFO_LOOP (loop_vinfo); |
8179 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); |
8180 | } |
8181 | else |
8182 | vf = 1; |
8183 | |
8184 | /* Multiple types in SLP are handled by creating the appropriate number of |
8185 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
8186 | case of SLP. */ |
8187 | if (slp) |
8188 | ncopies = 1; |
8189 | else |
8190 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
8191 | |
8192 | gcc_assert (ncopies >= 1); |
8193 | |
8194 | /* FORNOW. This restriction should be relaxed. */ |
8195 | if (loop && nested_in_vect_loop_p (loop, stmt_info) && ncopies > 1) |
8196 | { |
8197 | if (dump_enabled_p ()) |
8198 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8199 | "multiple types in nested loop.\n" ); |
8200 | return false; |
8201 | } |
8202 | |
8203 | tree op; |
8204 | slp_tree op_node; |
8205 | if (!vect_check_store_rhs (vinfo, stmt_info, slp_node, |
8206 | rhs: &op, rhs_node: &op_node, rhs_dt_out: &rhs_dt, rhs_vectype_out: &rhs_vectype, vls_type_out: &vls_type)) |
8207 | return false; |
8208 | |
8209 | elem_type = TREE_TYPE (vectype); |
8210 | vec_mode = TYPE_MODE (vectype); |
8211 | |
8212 | if (!STMT_VINFO_DATA_REF (stmt_info)) |
8213 | return false; |
8214 | |
8215 | vect_memory_access_type memory_access_type; |
8216 | enum dr_alignment_support alignment_support_scheme; |
8217 | int misalignment; |
8218 | poly_int64 poffset; |
8219 | internal_fn lanes_ifn; |
8220 | if (!get_load_store_type (vinfo, stmt_info, vectype, slp_node, masked_p: mask, vls_type, |
8221 | ncopies, memory_access_type: &memory_access_type, poffset: &poffset, |
8222 | alignment_support_scheme: &alignment_support_scheme, misalignment: &misalignment, gs_info: &gs_info, |
8223 | lanes_ifn: &lanes_ifn)) |
8224 | return false; |
8225 | |
8226 | if (mask) |
8227 | { |
8228 | if (memory_access_type == VMAT_CONTIGUOUS) |
8229 | { |
8230 | if (!VECTOR_MODE_P (vec_mode) |
8231 | || !can_vec_mask_load_store_p (vec_mode, |
8232 | TYPE_MODE (mask_vectype), false)) |
8233 | return false; |
8234 | } |
8235 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
8236 | && (memory_access_type != VMAT_GATHER_SCATTER |
8237 | || (gs_info.decl && !VECTOR_BOOLEAN_TYPE_P (mask_vectype)))) |
8238 | { |
8239 | if (dump_enabled_p ()) |
8240 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8241 | "unsupported access type for masked store.\n" ); |
8242 | return false; |
8243 | } |
8244 | else if (memory_access_type == VMAT_GATHER_SCATTER |
8245 | && gs_info.ifn == IFN_LAST |
8246 | && !gs_info.decl) |
8247 | { |
8248 | if (dump_enabled_p ()) |
8249 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8250 | "unsupported masked emulated scatter.\n" ); |
8251 | return false; |
8252 | } |
8253 | } |
8254 | else |
8255 | { |
8256 | /* FORNOW. In some cases can vectorize even if data-type not supported |
8257 | (e.g. - array initialization with 0). */ |
8258 | if (optab_handler (op: mov_optab, mode: vec_mode) == CODE_FOR_nothing) |
8259 | return false; |
8260 | } |
8261 | |
8262 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info), *first_dr_info = NULL; |
8263 | grouped_store = (STMT_VINFO_GROUPED_ACCESS (stmt_info) |
8264 | && memory_access_type != VMAT_GATHER_SCATTER |
8265 | && (slp || memory_access_type != VMAT_CONTIGUOUS)); |
8266 | if (grouped_store) |
8267 | { |
8268 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
8269 | first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info); |
8270 | group_size = DR_GROUP_SIZE (first_stmt_info); |
8271 | } |
8272 | else |
8273 | { |
8274 | first_stmt_info = stmt_info; |
8275 | first_dr_info = dr_info; |
8276 | group_size = vec_num = 1; |
8277 | } |
8278 | |
8279 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) > 1 && !vec_stmt) |
8280 | { |
8281 | if (!check_scan_store (vinfo, stmt_info, vectype, rhs_dt, slp, mask, |
8282 | memory_access_type)) |
8283 | return false; |
8284 | } |
8285 | |
8286 | bool costing_p = !vec_stmt; |
8287 | if (costing_p) /* transformation not required. */ |
8288 | { |
8289 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
8290 | |
8291 | if (loop_vinfo |
8292 | && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) |
8293 | check_load_store_for_partial_vectors (loop_vinfo, vectype, slp_node, |
8294 | vls_type, group_size, |
8295 | memory_access_type, gs_info: &gs_info, |
8296 | scalar_mask: mask); |
8297 | |
8298 | if (slp_node |
8299 | && (!vect_maybe_update_slp_op_vectype (op_node, vectype) |
8300 | || (mask |
8301 | && !vect_maybe_update_slp_op_vectype (mask_node, |
8302 | mask_vectype)))) |
8303 | { |
8304 | if (dump_enabled_p ()) |
8305 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8306 | "incompatible vector types for invariants\n" ); |
8307 | return false; |
8308 | } |
8309 | |
8310 | if (dump_enabled_p () |
8311 | && memory_access_type != VMAT_ELEMENTWISE |
8312 | && memory_access_type != VMAT_GATHER_SCATTER |
8313 | && alignment_support_scheme != dr_aligned) |
8314 | dump_printf_loc (MSG_NOTE, vect_location, |
8315 | "Vectorizing an unaligned access.\n" ); |
8316 | |
8317 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; |
8318 | |
8319 | /* As function vect_transform_stmt shows, for interleaving stores |
8320 | the whole chain is vectorized when the last store in the chain |
8321 | is reached, the other stores in the group are skipped. So we |
8322 | want to only cost the last one here, but it's not trivial to |
8323 | get the last, as it's equivalent to use the first one for |
8324 | costing, use the first one instead. */ |
8325 | if (grouped_store |
8326 | && !slp |
8327 | && first_stmt_info != stmt_info) |
8328 | return true; |
8329 | } |
8330 | gcc_assert (memory_access_type == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
8331 | |
8332 | /* Transform. */ |
8333 | |
8334 | ensure_base_align (dr_info); |
8335 | |
8336 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) >= 3) |
8337 | { |
8338 | gcc_assert (memory_access_type == VMAT_CONTIGUOUS); |
8339 | gcc_assert (!slp); |
8340 | if (costing_p) |
8341 | { |
8342 | unsigned int inside_cost = 0, prologue_cost = 0; |
8343 | if (vls_type == VLS_STORE_INVARIANT) |
8344 | prologue_cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: scalar_to_vec, |
8345 | stmt_info, misalign: 0, where: vect_prologue); |
8346 | vect_get_store_cost (vinfo, stmt_info, ncopies, |
8347 | alignment_support_scheme, misalignment, |
8348 | inside_cost: &inside_cost, body_cost_vec: cost_vec); |
8349 | |
8350 | if (dump_enabled_p ()) |
8351 | dump_printf_loc (MSG_NOTE, vect_location, |
8352 | "vect_model_store_cost: inside_cost = %d, " |
8353 | "prologue_cost = %d .\n" , |
8354 | inside_cost, prologue_cost); |
8355 | |
8356 | return true; |
8357 | } |
8358 | return vectorizable_scan_store (vinfo, stmt_info, gsi, vec_stmt, ncopies); |
8359 | } |
8360 | |
8361 | if (grouped_store) |
8362 | { |
8363 | /* FORNOW */ |
8364 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt_info)); |
8365 | |
8366 | if (slp) |
8367 | { |
8368 | grouped_store = false; |
8369 | /* VEC_NUM is the number of vect stmts to be created for this |
8370 | group. */ |
8371 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
8372 | first_stmt_info = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
8373 | gcc_assert (DR_GROUP_FIRST_ELEMENT (first_stmt_info) |
8374 | == first_stmt_info); |
8375 | first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info); |
8376 | op = vect_get_store_rhs (stmt_info: first_stmt_info); |
8377 | } |
8378 | else |
8379 | /* VEC_NUM is the number of vect stmts to be created for this |
8380 | group. */ |
8381 | vec_num = group_size; |
8382 | |
8383 | ref_type = get_group_alias_ptr_type (first_stmt_info); |
8384 | } |
8385 | else |
8386 | ref_type = reference_alias_ptr_type (DR_REF (first_dr_info->dr)); |
8387 | |
8388 | if (!costing_p && dump_enabled_p ()) |
8389 | dump_printf_loc (MSG_NOTE, vect_location, "transform store. ncopies = %d\n" , |
8390 | ncopies); |
8391 | |
8392 | /* Check if we need to update prologue cost for invariant, |
8393 | and update it accordingly if so. If it's not for |
8394 | interleaving store, we can just check vls_type; but if |
8395 | it's for interleaving store, need to check the def_type |
8396 | of the stored value since the current vls_type is just |
8397 | for first_stmt_info. */ |
8398 | auto update_prologue_cost = [&](unsigned *prologue_cost, tree store_rhs) |
8399 | { |
8400 | gcc_assert (costing_p); |
8401 | if (slp) |
8402 | return; |
8403 | if (grouped_store) |
8404 | { |
8405 | gcc_assert (store_rhs); |
8406 | enum vect_def_type cdt; |
8407 | gcc_assert (vect_is_simple_use (store_rhs, vinfo, &cdt)); |
8408 | if (cdt != vect_constant_def && cdt != vect_external_def) |
8409 | return; |
8410 | } |
8411 | else if (vls_type != VLS_STORE_INVARIANT) |
8412 | return; |
8413 | *prologue_cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: scalar_to_vec, stmt_info, |
8414 | misalign: 0, where: vect_prologue); |
8415 | }; |
8416 | |
8417 | if (memory_access_type == VMAT_ELEMENTWISE |
8418 | || memory_access_type == VMAT_STRIDED_SLP) |
8419 | { |
8420 | unsigned inside_cost = 0, prologue_cost = 0; |
8421 | gimple_stmt_iterator incr_gsi; |
8422 | bool insert_after; |
8423 | gimple *incr; |
8424 | tree offvar; |
8425 | tree ivstep; |
8426 | tree running_off; |
8427 | tree stride_base, stride_step, alias_off; |
8428 | tree vec_oprnd = NULL_TREE; |
8429 | tree dr_offset; |
8430 | unsigned int g; |
8431 | /* Checked by get_load_store_type. */ |
8432 | unsigned int const_nunits = nunits.to_constant (); |
8433 | |
8434 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
8435 | gcc_assert (!nested_in_vect_loop_p (loop, stmt_info)); |
8436 | |
8437 | dr_offset = get_dr_vinfo_offset (vinfo, dr_info: first_dr_info); |
8438 | stride_base |
8439 | = fold_build_pointer_plus |
8440 | (DR_BASE_ADDRESS (first_dr_info->dr), |
8441 | size_binop (PLUS_EXPR, |
8442 | convert_to_ptrofftype (dr_offset), |
8443 | convert_to_ptrofftype (DR_INIT (first_dr_info->dr)))); |
8444 | stride_step = fold_convert (sizetype, DR_STEP (first_dr_info->dr)); |
8445 | |
8446 | /* For a store with loop-invariant (but other than power-of-2) |
8447 | stride (i.e. not a grouped access) like so: |
8448 | |
8449 | for (i = 0; i < n; i += stride) |
8450 | array[i] = ...; |
8451 | |
8452 | we generate a new induction variable and new stores from |
8453 | the components of the (vectorized) rhs: |
8454 | |
8455 | for (j = 0; ; j += VF*stride) |
8456 | vectemp = ...; |
8457 | tmp1 = vectemp[0]; |
8458 | array[j] = tmp1; |
8459 | tmp2 = vectemp[1]; |
8460 | array[j + stride] = tmp2; |
8461 | ... |
8462 | */ |
8463 | |
8464 | unsigned nstores = const_nunits; |
8465 | unsigned lnel = 1; |
8466 | tree ltype = elem_type; |
8467 | tree lvectype = vectype; |
8468 | if (slp) |
8469 | { |
8470 | if (group_size < const_nunits |
8471 | && const_nunits % group_size == 0) |
8472 | { |
8473 | nstores = const_nunits / group_size; |
8474 | lnel = group_size; |
8475 | ltype = build_vector_type (elem_type, group_size); |
8476 | lvectype = vectype; |
8477 | |
8478 | /* First check if vec_extract optab doesn't support extraction |
8479 | of vector elts directly. */ |
8480 | scalar_mode elmode = SCALAR_TYPE_MODE (elem_type); |
8481 | machine_mode vmode; |
8482 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) |
8483 | || !related_vector_mode (TYPE_MODE (vectype), elmode, |
8484 | group_size).exists (mode: &vmode) |
8485 | || (convert_optab_handler (op: vec_extract_optab, |
8486 | TYPE_MODE (vectype), from_mode: vmode) |
8487 | == CODE_FOR_nothing)) |
8488 | { |
8489 | /* Try to avoid emitting an extract of vector elements |
8490 | by performing the extracts using an integer type of the |
8491 | same size, extracting from a vector of those and then |
8492 | re-interpreting it as the original vector type if |
8493 | supported. */ |
8494 | unsigned lsize |
8495 | = group_size * GET_MODE_BITSIZE (mode: elmode); |
8496 | unsigned int lnunits = const_nunits / group_size; |
8497 | /* If we can't construct such a vector fall back to |
8498 | element extracts from the original vector type and |
8499 | element size stores. */ |
8500 | if (int_mode_for_size (size: lsize, limit: 0).exists (mode: &elmode) |
8501 | && VECTOR_MODE_P (TYPE_MODE (vectype)) |
8502 | && related_vector_mode (TYPE_MODE (vectype), elmode, |
8503 | lnunits).exists (mode: &vmode) |
8504 | && (convert_optab_handler (op: vec_extract_optab, |
8505 | to_mode: vmode, from_mode: elmode) |
8506 | != CODE_FOR_nothing)) |
8507 | { |
8508 | nstores = lnunits; |
8509 | lnel = group_size; |
8510 | ltype = build_nonstandard_integer_type (lsize, 1); |
8511 | lvectype = build_vector_type (ltype, nstores); |
8512 | } |
8513 | /* Else fall back to vector extraction anyway. |
8514 | Fewer stores are more important than avoiding spilling |
8515 | of the vector we extract from. Compared to the |
8516 | construction case in vectorizable_load no store-forwarding |
8517 | issue exists here for reasonable archs. */ |
8518 | } |
8519 | } |
8520 | else if (group_size >= const_nunits |
8521 | && group_size % const_nunits == 0) |
8522 | { |
8523 | int mis_align = dr_misalignment (dr_info: first_dr_info, vectype); |
8524 | dr_alignment_support dr_align |
8525 | = vect_supportable_dr_alignment (vinfo, dr_info, vectype, |
8526 | mis_align); |
8527 | if (dr_align == dr_aligned |
8528 | || dr_align == dr_unaligned_supported) |
8529 | { |
8530 | nstores = 1; |
8531 | lnel = const_nunits; |
8532 | ltype = vectype; |
8533 | lvectype = vectype; |
8534 | alignment_support_scheme = dr_align; |
8535 | misalignment = mis_align; |
8536 | } |
8537 | } |
8538 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); |
8539 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
8540 | } |
8541 | |
8542 | if (!costing_p) |
8543 | { |
8544 | ivstep = stride_step; |
8545 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, |
8546 | build_int_cst (TREE_TYPE (ivstep), vf)); |
8547 | |
8548 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); |
8549 | |
8550 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
8551 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); |
8552 | create_iv (stride_base, PLUS_EXPR, ivstep, NULL, loop, &incr_gsi, |
8553 | insert_after, &offvar, NULL); |
8554 | incr = gsi_stmt (i: incr_gsi); |
8555 | |
8556 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
8557 | } |
8558 | |
8559 | alias_off = build_int_cst (ref_type, 0); |
8560 | stmt_vec_info next_stmt_info = first_stmt_info; |
8561 | auto_vec<tree> vec_oprnds; |
8562 | /* For costing some adjacent vector stores, we'd like to cost with |
8563 | the total number of them once instead of cost each one by one. */ |
8564 | unsigned int n_adjacent_stores = 0; |
8565 | for (g = 0; g < group_size; g++) |
8566 | { |
8567 | running_off = offvar; |
8568 | if (!costing_p) |
8569 | { |
8570 | if (g) |
8571 | { |
8572 | tree size = TYPE_SIZE_UNIT (ltype); |
8573 | tree pos |
8574 | = fold_build2 (MULT_EXPR, sizetype, size_int (g), size); |
8575 | tree newoff = copy_ssa_name (var: running_off, NULL); |
8576 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
8577 | running_off, pos); |
8578 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: incr, gsi); |
8579 | running_off = newoff; |
8580 | } |
8581 | } |
8582 | if (!slp) |
8583 | op = vect_get_store_rhs (stmt_info: next_stmt_info); |
8584 | if (!costing_p) |
8585 | vect_get_vec_defs (vinfo, stmt_info: next_stmt_info, slp_node, ncopies, op0: op, |
8586 | vec_oprnds0: &vec_oprnds); |
8587 | else |
8588 | update_prologue_cost (&prologue_cost, op); |
8589 | unsigned int group_el = 0; |
8590 | unsigned HOST_WIDE_INT |
8591 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); |
8592 | for (j = 0; j < ncopies; j++) |
8593 | { |
8594 | if (!costing_p) |
8595 | { |
8596 | vec_oprnd = vec_oprnds[j]; |
8597 | /* Pun the vector to extract from if necessary. */ |
8598 | if (lvectype != vectype) |
8599 | { |
8600 | tree tem = make_ssa_name (var: lvectype); |
8601 | tree cvt |
8602 | = build1 (VIEW_CONVERT_EXPR, lvectype, vec_oprnd); |
8603 | gimple *pun = gimple_build_assign (tem, cvt); |
8604 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: pun, gsi); |
8605 | vec_oprnd = tem; |
8606 | } |
8607 | } |
8608 | for (i = 0; i < nstores; i++) |
8609 | { |
8610 | if (costing_p) |
8611 | { |
8612 | /* Only need vector extracting when there are more |
8613 | than one stores. */ |
8614 | if (nstores > 1) |
8615 | inside_cost |
8616 | += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: vec_to_scalar, |
8617 | stmt_info, misalign: 0, where: vect_body); |
8618 | /* Take a single lane vector type store as scalar |
8619 | store to avoid ICE like 110776. */ |
8620 | if (VECTOR_TYPE_P (ltype) |
8621 | && known_ne (TYPE_VECTOR_SUBPARTS (ltype), 1U)) |
8622 | n_adjacent_stores++; |
8623 | else |
8624 | inside_cost |
8625 | += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: scalar_store, |
8626 | stmt_info, misalign: 0, where: vect_body); |
8627 | continue; |
8628 | } |
8629 | tree newref, newoff; |
8630 | gimple *incr, *assign; |
8631 | tree size = TYPE_SIZE (ltype); |
8632 | /* Extract the i'th component. */ |
8633 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, |
8634 | bitsize_int (i), size); |
8635 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, |
8636 | size, pos); |
8637 | |
8638 | elem = force_gimple_operand_gsi (gsi, elem, true, |
8639 | NULL_TREE, true, |
8640 | GSI_SAME_STMT); |
8641 | |
8642 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
8643 | group_el * elsz); |
8644 | newref = build2 (MEM_REF, ltype, |
8645 | running_off, this_off); |
8646 | vect_copy_ref_info (newref, DR_REF (first_dr_info->dr)); |
8647 | |
8648 | /* And store it to *running_off. */ |
8649 | assign = gimple_build_assign (newref, elem); |
8650 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: assign, gsi); |
8651 | |
8652 | group_el += lnel; |
8653 | if (! slp |
8654 | || group_el == group_size) |
8655 | { |
8656 | newoff = copy_ssa_name (var: running_off, NULL); |
8657 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
8658 | running_off, stride_step); |
8659 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: incr, gsi); |
8660 | |
8661 | running_off = newoff; |
8662 | group_el = 0; |
8663 | } |
8664 | if (g == group_size - 1 |
8665 | && !slp) |
8666 | { |
8667 | if (j == 0 && i == 0) |
8668 | *vec_stmt = assign; |
8669 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: assign); |
8670 | } |
8671 | } |
8672 | } |
8673 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
8674 | vec_oprnds.truncate(size: 0); |
8675 | if (slp) |
8676 | break; |
8677 | } |
8678 | |
8679 | if (costing_p) |
8680 | { |
8681 | if (n_adjacent_stores > 0) |
8682 | vect_get_store_cost (vinfo, stmt_info, ncopies: n_adjacent_stores, |
8683 | alignment_support_scheme, misalignment, |
8684 | inside_cost: &inside_cost, body_cost_vec: cost_vec); |
8685 | if (dump_enabled_p ()) |
8686 | dump_printf_loc (MSG_NOTE, vect_location, |
8687 | "vect_model_store_cost: inside_cost = %d, " |
8688 | "prologue_cost = %d .\n" , |
8689 | inside_cost, prologue_cost); |
8690 | } |
8691 | |
8692 | return true; |
8693 | } |
8694 | |
8695 | gcc_assert (alignment_support_scheme); |
8696 | vec_loop_masks *loop_masks |
8697 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) |
8698 | ? &LOOP_VINFO_MASKS (loop_vinfo) |
8699 | : NULL); |
8700 | vec_loop_lens *loop_lens |
8701 | = (loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo) |
8702 | ? &LOOP_VINFO_LENS (loop_vinfo) |
8703 | : NULL); |
8704 | |
8705 | /* The vect_transform_stmt and vect_analyze_stmt will go here but there |
8706 | are some difference here. We cannot enable both the lens and masks |
8707 | during transform but it is allowed during analysis. |
8708 | Shouldn't go with length-based approach if fully masked. */ |
8709 | if (cost_vec == NULL) |
8710 | /* The cost_vec is NULL during transfrom. */ |
8711 | gcc_assert ((!loop_lens || !loop_masks)); |
8712 | |
8713 | /* Targets with store-lane instructions must not require explicit |
8714 | realignment. vect_supportable_dr_alignment always returns either |
8715 | dr_aligned or dr_unaligned_supported for masked operations. */ |
8716 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES |
8717 | && !mask |
8718 | && !loop_masks) |
8719 | || alignment_support_scheme == dr_aligned |
8720 | || alignment_support_scheme == dr_unaligned_supported); |
8721 | |
8722 | tree offset = NULL_TREE; |
8723 | if (!known_eq (poffset, 0)) |
8724 | offset = size_int (poffset); |
8725 | |
8726 | tree bump; |
8727 | tree vec_offset = NULL_TREE; |
8728 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
8729 | { |
8730 | aggr_type = NULL_TREE; |
8731 | bump = NULL_TREE; |
8732 | } |
8733 | else if (memory_access_type == VMAT_GATHER_SCATTER) |
8734 | { |
8735 | aggr_type = elem_type; |
8736 | if (!costing_p) |
8737 | vect_get_strided_load_store_ops (stmt_info, loop_vinfo, gsi, gs_info: &gs_info, |
8738 | dataref_bump: &bump, vec_offset: &vec_offset, loop_lens); |
8739 | } |
8740 | else |
8741 | { |
8742 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
8743 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); |
8744 | else |
8745 | aggr_type = vectype; |
8746 | bump = vect_get_data_ptr_increment (vinfo, gsi, dr_info, aggr_type, |
8747 | memory_access_type, loop_lens); |
8748 | } |
8749 | |
8750 | if (mask && !costing_p) |
8751 | LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true; |
8752 | |
8753 | /* In case the vectorization factor (VF) is bigger than the number |
8754 | of elements that we can fit in a vectype (nunits), we have to generate |
8755 | more than one vector stmt - i.e - we need to "unroll" the |
8756 | vector stmt by a factor VF/nunits. */ |
8757 | |
8758 | /* In case of interleaving (non-unit grouped access): |
8759 | |
8760 | S1: &base + 2 = x2 |
8761 | S2: &base = x0 |
8762 | S3: &base + 1 = x1 |
8763 | S4: &base + 3 = x3 |
8764 | |
8765 | We create vectorized stores starting from base address (the access of the |
8766 | first stmt in the chain (S2 in the above example), when the last store stmt |
8767 | of the chain (S4) is reached: |
8768 | |
8769 | VS1: &base = vx2 |
8770 | VS2: &base + vec_size*1 = vx0 |
8771 | VS3: &base + vec_size*2 = vx1 |
8772 | VS4: &base + vec_size*3 = vx3 |
8773 | |
8774 | Then permutation statements are generated: |
8775 | |
8776 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
8777 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > |
8778 | ... |
8779 | |
8780 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
8781 | (the order of the data-refs in the output of vect_permute_store_chain |
8782 | corresponds to the order of scalar stmts in the interleaving chain - see |
8783 | the documentation of vect_permute_store_chain()). |
8784 | |
8785 | In case of both multiple types and interleaving, above vector stores and |
8786 | permutation stmts are created for every copy. The result vector stmts are |
8787 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
8788 | STMT_VINFO_RELATED_STMT for the next copies. |
8789 | */ |
8790 | |
8791 | auto_vec<tree> dr_chain (group_size); |
8792 | auto_vec<tree> vec_masks; |
8793 | tree vec_mask = NULL; |
8794 | auto_delete_vec<auto_vec<tree>> gvec_oprnds (group_size); |
8795 | for (i = 0; i < group_size; i++) |
8796 | gvec_oprnds.quick_push (obj: new auto_vec<tree> ()); |
8797 | |
8798 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
8799 | { |
8800 | gcc_assert (!slp && grouped_store); |
8801 | unsigned inside_cost = 0, prologue_cost = 0; |
8802 | /* For costing some adjacent vector stores, we'd like to cost with |
8803 | the total number of them once instead of cost each one by one. */ |
8804 | unsigned int n_adjacent_stores = 0; |
8805 | for (j = 0; j < ncopies; j++) |
8806 | { |
8807 | gimple *new_stmt; |
8808 | if (j == 0) |
8809 | { |
8810 | /* For interleaved stores we collect vectorized defs for all |
8811 | the stores in the group in DR_CHAIN. DR_CHAIN is then used |
8812 | as an input to vect_permute_store_chain(). */ |
8813 | stmt_vec_info next_stmt_info = first_stmt_info; |
8814 | for (i = 0; i < group_size; i++) |
8815 | { |
8816 | /* Since gaps are not supported for interleaved stores, |
8817 | DR_GROUP_SIZE is the exact number of stmts in the |
8818 | chain. Therefore, NEXT_STMT_INFO can't be NULL_TREE. */ |
8819 | op = vect_get_store_rhs (stmt_info: next_stmt_info); |
8820 | if (costing_p) |
8821 | update_prologue_cost (&prologue_cost, op); |
8822 | else |
8823 | { |
8824 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: next_stmt_info, |
8825 | ncopies, op, |
8826 | vec_oprnds: gvec_oprnds[i]); |
8827 | vec_oprnd = (*gvec_oprnds[i])[0]; |
8828 | dr_chain.quick_push (obj: vec_oprnd); |
8829 | } |
8830 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
8831 | } |
8832 | |
8833 | if (!costing_p) |
8834 | { |
8835 | if (mask) |
8836 | { |
8837 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies, |
8838 | op: mask, vec_oprnds: &vec_masks, |
8839 | vectype: mask_vectype); |
8840 | vec_mask = vec_masks[0]; |
8841 | } |
8842 | |
8843 | /* We should have catched mismatched types earlier. */ |
8844 | gcc_assert ( |
8845 | useless_type_conversion_p (vectype, TREE_TYPE (vec_oprnd))); |
8846 | dataref_ptr |
8847 | = vect_create_data_ref_ptr (vinfo, first_stmt_info, |
8848 | aggr_type, NULL, offset, &dummy, |
8849 | gsi, &ptr_incr, false, bump); |
8850 | } |
8851 | } |
8852 | else if (!costing_p) |
8853 | { |
8854 | gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)); |
8855 | /* DR_CHAIN is then used as an input to |
8856 | vect_permute_store_chain(). */ |
8857 | for (i = 0; i < group_size; i++) |
8858 | { |
8859 | vec_oprnd = (*gvec_oprnds[i])[j]; |
8860 | dr_chain[i] = vec_oprnd; |
8861 | } |
8862 | if (mask) |
8863 | vec_mask = vec_masks[j]; |
8864 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi, |
8865 | stmt_info, bump); |
8866 | } |
8867 | |
8868 | if (costing_p) |
8869 | { |
8870 | n_adjacent_stores += vec_num; |
8871 | continue; |
8872 | } |
8873 | |
8874 | /* Get an array into which we can store the individual vectors. */ |
8875 | tree vec_array = create_vector_array (elem_type: vectype, nelems: vec_num); |
8876 | |
8877 | /* Invalidate the current contents of VEC_ARRAY. This should |
8878 | become an RTL clobber too, which prevents the vector registers |
8879 | from being upward-exposed. */ |
8880 | vect_clobber_variable (vinfo, stmt_info, gsi, var: vec_array); |
8881 | |
8882 | /* Store the individual vectors into the array. */ |
8883 | for (i = 0; i < vec_num; i++) |
8884 | { |
8885 | vec_oprnd = dr_chain[i]; |
8886 | write_vector_array (vinfo, stmt_info, gsi, vect: vec_oprnd, array: vec_array, |
8887 | n: i); |
8888 | } |
8889 | |
8890 | tree final_mask = NULL; |
8891 | tree final_len = NULL; |
8892 | tree bias = NULL; |
8893 | if (loop_masks) |
8894 | final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks, |
8895 | ncopies, vectype, j); |
8896 | if (vec_mask) |
8897 | final_mask = prepare_vec_mask (loop_vinfo, mask_type: mask_vectype, loop_mask: final_mask, |
8898 | vec_mask, gsi); |
8899 | |
8900 | if (lanes_ifn == IFN_MASK_LEN_STORE_LANES) |
8901 | { |
8902 | if (loop_lens) |
8903 | final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens, |
8904 | ncopies, vectype, j, 1); |
8905 | else |
8906 | final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
8907 | signed char biasval |
8908 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
8909 | bias = build_int_cst (intQI_type_node, biasval); |
8910 | if (!final_mask) |
8911 | { |
8912 | mask_vectype = truth_type_for (vectype); |
8913 | final_mask = build_minus_one_cst (mask_vectype); |
8914 | } |
8915 | } |
8916 | |
8917 | gcall *call; |
8918 | if (final_len && final_mask) |
8919 | { |
8920 | /* Emit: |
8921 | MASK_LEN_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK, |
8922 | LEN, BIAS, VEC_ARRAY). */ |
8923 | unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype)); |
8924 | tree alias_ptr = build_int_cst (ref_type, align); |
8925 | call = gimple_build_call_internal (IFN_MASK_LEN_STORE_LANES, 6, |
8926 | dataref_ptr, alias_ptr, |
8927 | final_mask, final_len, bias, |
8928 | vec_array); |
8929 | } |
8930 | else if (final_mask) |
8931 | { |
8932 | /* Emit: |
8933 | MASK_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK, |
8934 | VEC_ARRAY). */ |
8935 | unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype)); |
8936 | tree alias_ptr = build_int_cst (ref_type, align); |
8937 | call = gimple_build_call_internal (IFN_MASK_STORE_LANES, 4, |
8938 | dataref_ptr, alias_ptr, |
8939 | final_mask, vec_array); |
8940 | } |
8941 | else |
8942 | { |
8943 | /* Emit: |
8944 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ |
8945 | data_ref = create_array_ref (type: aggr_type, ptr: dataref_ptr, alias_ptr_type: ref_type); |
8946 | call = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); |
8947 | gimple_call_set_lhs (gs: call, lhs: data_ref); |
8948 | } |
8949 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
8950 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
8951 | new_stmt = call; |
8952 | |
8953 | /* Record that VEC_ARRAY is now dead. */ |
8954 | vect_clobber_variable (vinfo, stmt_info, gsi, var: vec_array); |
8955 | if (j == 0) |
8956 | *vec_stmt = new_stmt; |
8957 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
8958 | } |
8959 | |
8960 | if (costing_p) |
8961 | { |
8962 | if (n_adjacent_stores > 0) |
8963 | vect_get_store_cost (vinfo, stmt_info, ncopies: n_adjacent_stores, |
8964 | alignment_support_scheme, misalignment, |
8965 | inside_cost: &inside_cost, body_cost_vec: cost_vec); |
8966 | if (dump_enabled_p ()) |
8967 | dump_printf_loc (MSG_NOTE, vect_location, |
8968 | "vect_model_store_cost: inside_cost = %d, " |
8969 | "prologue_cost = %d .\n" , |
8970 | inside_cost, prologue_cost); |
8971 | } |
8972 | |
8973 | return true; |
8974 | } |
8975 | |
8976 | if (memory_access_type == VMAT_GATHER_SCATTER) |
8977 | { |
8978 | gcc_assert (!grouped_store); |
8979 | auto_vec<tree> vec_offsets; |
8980 | unsigned int inside_cost = 0, prologue_cost = 0; |
8981 | for (j = 0; j < ncopies; j++) |
8982 | { |
8983 | gimple *new_stmt; |
8984 | if (j == 0) |
8985 | { |
8986 | if (costing_p && vls_type == VLS_STORE_INVARIANT) |
8987 | prologue_cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: scalar_to_vec, |
8988 | stmt_info, misalign: 0, where: vect_prologue); |
8989 | else if (!costing_p) |
8990 | { |
8991 | /* Since the store is not grouped, DR_GROUP_SIZE is 1, and |
8992 | DR_CHAIN is of size 1. */ |
8993 | gcc_assert (group_size == 1); |
8994 | if (slp_node) |
8995 | vect_get_slp_defs (op_node, gvec_oprnds[0]); |
8996 | else |
8997 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: first_stmt_info, |
8998 | ncopies, op, vec_oprnds: gvec_oprnds[0]); |
8999 | if (mask) |
9000 | { |
9001 | if (slp_node) |
9002 | vect_get_slp_defs (mask_node, &vec_masks); |
9003 | else |
9004 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, |
9005 | ncopies, |
9006 | op: mask, vec_oprnds: &vec_masks, |
9007 | vectype: mask_vectype); |
9008 | } |
9009 | |
9010 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
9011 | vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info, |
9012 | slp_node, gs_info: &gs_info, |
9013 | dataref_ptr: &dataref_ptr, vec_offset: &vec_offsets); |
9014 | else |
9015 | dataref_ptr |
9016 | = vect_create_data_ref_ptr (vinfo, first_stmt_info, |
9017 | aggr_type, NULL, offset, |
9018 | &dummy, gsi, &ptr_incr, false, |
9019 | bump); |
9020 | } |
9021 | } |
9022 | else if (!costing_p) |
9023 | { |
9024 | gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)); |
9025 | if (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
9026 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, |
9027 | gsi, stmt_info, bump); |
9028 | } |
9029 | |
9030 | new_stmt = NULL; |
9031 | for (i = 0; i < vec_num; ++i) |
9032 | { |
9033 | if (!costing_p) |
9034 | { |
9035 | vec_oprnd = (*gvec_oprnds[0])[vec_num * j + i]; |
9036 | if (mask) |
9037 | vec_mask = vec_masks[vec_num * j + i]; |
9038 | /* We should have catched mismatched types earlier. */ |
9039 | gcc_assert (useless_type_conversion_p (vectype, |
9040 | TREE_TYPE (vec_oprnd))); |
9041 | } |
9042 | unsigned HOST_WIDE_INT align; |
9043 | tree final_mask = NULL_TREE; |
9044 | tree final_len = NULL_TREE; |
9045 | tree bias = NULL_TREE; |
9046 | if (!costing_p) |
9047 | { |
9048 | if (loop_masks) |
9049 | final_mask = vect_get_loop_mask (loop_vinfo, gsi, |
9050 | loop_masks, ncopies, |
9051 | vectype, j); |
9052 | if (vec_mask) |
9053 | final_mask = prepare_vec_mask (loop_vinfo, mask_type: mask_vectype, |
9054 | loop_mask: final_mask, vec_mask, gsi); |
9055 | } |
9056 | |
9057 | if (gs_info.ifn != IFN_LAST) |
9058 | { |
9059 | if (costing_p) |
9060 | { |
9061 | unsigned int cnunits = vect_nunits_for_cost (vec_type: vectype); |
9062 | inside_cost |
9063 | += record_stmt_cost (body_cost_vec: cost_vec, count: cnunits, kind: scalar_store, |
9064 | stmt_info, misalign: 0, where: vect_body); |
9065 | continue; |
9066 | } |
9067 | |
9068 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
9069 | vec_offset = vec_offsets[vec_num * j + i]; |
9070 | tree scale = size_int (gs_info.scale); |
9071 | |
9072 | if (gs_info.ifn == IFN_MASK_LEN_SCATTER_STORE) |
9073 | { |
9074 | if (loop_lens) |
9075 | final_len = vect_get_loop_len (loop_vinfo, gsi, |
9076 | loop_lens, ncopies, |
9077 | vectype, j, 1); |
9078 | else |
9079 | final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
9080 | signed char biasval |
9081 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
9082 | bias = build_int_cst (intQI_type_node, biasval); |
9083 | if (!final_mask) |
9084 | { |
9085 | mask_vectype = truth_type_for (vectype); |
9086 | final_mask = build_minus_one_cst (mask_vectype); |
9087 | } |
9088 | } |
9089 | |
9090 | gcall *call; |
9091 | if (final_len && final_mask) |
9092 | call = gimple_build_call_internal |
9093 | (IFN_MASK_LEN_SCATTER_STORE, 7, dataref_ptr, |
9094 | vec_offset, scale, vec_oprnd, final_mask, |
9095 | final_len, bias); |
9096 | else if (final_mask) |
9097 | call = gimple_build_call_internal |
9098 | (IFN_MASK_SCATTER_STORE, 5, dataref_ptr, |
9099 | vec_offset, scale, vec_oprnd, final_mask); |
9100 | else |
9101 | call = gimple_build_call_internal (IFN_SCATTER_STORE, 4, |
9102 | dataref_ptr, vec_offset, |
9103 | scale, vec_oprnd); |
9104 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
9105 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
9106 | new_stmt = call; |
9107 | } |
9108 | else if (gs_info.decl) |
9109 | { |
9110 | /* The builtin decls path for scatter is legacy, x86 only. */ |
9111 | gcc_assert (nunits.is_constant () |
9112 | && (!final_mask |
9113 | || SCALAR_INT_MODE_P |
9114 | (TYPE_MODE (TREE_TYPE (final_mask))))); |
9115 | if (costing_p) |
9116 | { |
9117 | unsigned int cnunits = vect_nunits_for_cost (vec_type: vectype); |
9118 | inside_cost |
9119 | += record_stmt_cost (body_cost_vec: cost_vec, count: cnunits, kind: scalar_store, |
9120 | stmt_info, misalign: 0, where: vect_body); |
9121 | continue; |
9122 | } |
9123 | poly_uint64 offset_nunits |
9124 | = TYPE_VECTOR_SUBPARTS (node: gs_info.offset_vectype); |
9125 | if (known_eq (nunits, offset_nunits)) |
9126 | { |
9127 | new_stmt = vect_build_one_scatter_store_call |
9128 | (vinfo, stmt_info, gsi, gs_info: &gs_info, |
9129 | ptr: dataref_ptr, offset: vec_offsets[vec_num * j + i], |
9130 | oprnd: vec_oprnd, mask: final_mask); |
9131 | vect_finish_stmt_generation (vinfo, stmt_info, |
9132 | vec_stmt: new_stmt, gsi); |
9133 | } |
9134 | else if (known_eq (nunits, offset_nunits * 2)) |
9135 | { |
9136 | /* We have a offset vector with half the number of |
9137 | lanes but the builtins will store full vectype |
9138 | data from the lower lanes. */ |
9139 | new_stmt = vect_build_one_scatter_store_call |
9140 | (vinfo, stmt_info, gsi, gs_info: &gs_info, |
9141 | ptr: dataref_ptr, |
9142 | offset: vec_offsets[2 * vec_num * j + 2 * i], |
9143 | oprnd: vec_oprnd, mask: final_mask); |
9144 | vect_finish_stmt_generation (vinfo, stmt_info, |
9145 | vec_stmt: new_stmt, gsi); |
9146 | int count = nunits.to_constant (); |
9147 | vec_perm_builder sel (count, count, 1); |
9148 | sel.quick_grow (len: count); |
9149 | for (int i = 0; i < count; ++i) |
9150 | sel[i] = i | (count / 2); |
9151 | vec_perm_indices indices (sel, 2, count); |
9152 | tree perm_mask |
9153 | = vect_gen_perm_mask_checked (vectype, indices); |
9154 | new_stmt = gimple_build_assign (NULL_TREE, VEC_PERM_EXPR, |
9155 | vec_oprnd, vec_oprnd, |
9156 | perm_mask); |
9157 | vec_oprnd = make_ssa_name (var: vectype); |
9158 | gimple_set_lhs (new_stmt, vec_oprnd); |
9159 | vect_finish_stmt_generation (vinfo, stmt_info, |
9160 | vec_stmt: new_stmt, gsi); |
9161 | if (final_mask) |
9162 | { |
9163 | new_stmt = gimple_build_assign (NULL_TREE, |
9164 | VEC_UNPACK_HI_EXPR, |
9165 | final_mask); |
9166 | final_mask = make_ssa_name |
9167 | (var: truth_type_for (gs_info.offset_vectype)); |
9168 | gimple_set_lhs (new_stmt, final_mask); |
9169 | vect_finish_stmt_generation (vinfo, stmt_info, |
9170 | vec_stmt: new_stmt, gsi); |
9171 | } |
9172 | new_stmt = vect_build_one_scatter_store_call |
9173 | (vinfo, stmt_info, gsi, gs_info: &gs_info, |
9174 | ptr: dataref_ptr, |
9175 | offset: vec_offsets[2 * vec_num * j + 2 * i + 1], |
9176 | oprnd: vec_oprnd, mask: final_mask); |
9177 | vect_finish_stmt_generation (vinfo, stmt_info, |
9178 | vec_stmt: new_stmt, gsi); |
9179 | } |
9180 | else if (known_eq (nunits * 2, offset_nunits)) |
9181 | { |
9182 | /* We have a offset vector with double the number of |
9183 | lanes. Select the low/high part accordingly. */ |
9184 | vec_offset = vec_offsets[(vec_num * j + i) / 2]; |
9185 | if ((vec_num * j + i) & 1) |
9186 | { |
9187 | int count = offset_nunits.to_constant (); |
9188 | vec_perm_builder sel (count, count, 1); |
9189 | sel.quick_grow (len: count); |
9190 | for (int i = 0; i < count; ++i) |
9191 | sel[i] = i | (count / 2); |
9192 | vec_perm_indices indices (sel, 2, count); |
9193 | tree perm_mask = vect_gen_perm_mask_checked |
9194 | (TREE_TYPE (vec_offset), indices); |
9195 | new_stmt = gimple_build_assign (NULL_TREE, |
9196 | VEC_PERM_EXPR, |
9197 | vec_offset, |
9198 | vec_offset, |
9199 | perm_mask); |
9200 | vec_offset = make_ssa_name (TREE_TYPE (vec_offset)); |
9201 | gimple_set_lhs (new_stmt, vec_offset); |
9202 | vect_finish_stmt_generation (vinfo, stmt_info, |
9203 | vec_stmt: new_stmt, gsi); |
9204 | } |
9205 | new_stmt = vect_build_one_scatter_store_call |
9206 | (vinfo, stmt_info, gsi, gs_info: &gs_info, |
9207 | ptr: dataref_ptr, offset: vec_offset, |
9208 | oprnd: vec_oprnd, mask: final_mask); |
9209 | vect_finish_stmt_generation (vinfo, stmt_info, |
9210 | vec_stmt: new_stmt, gsi); |
9211 | } |
9212 | else |
9213 | gcc_unreachable (); |
9214 | } |
9215 | else |
9216 | { |
9217 | /* Emulated scatter. */ |
9218 | gcc_assert (!final_mask); |
9219 | if (costing_p) |
9220 | { |
9221 | unsigned int cnunits = vect_nunits_for_cost (vec_type: vectype); |
9222 | /* For emulated scatter N offset vector element extracts |
9223 | (we assume the scalar scaling and ptr + offset add is |
9224 | consumed by the load). */ |
9225 | inside_cost |
9226 | += record_stmt_cost (body_cost_vec: cost_vec, count: cnunits, kind: vec_to_scalar, |
9227 | stmt_info, misalign: 0, where: vect_body); |
9228 | /* N scalar stores plus extracting the elements. */ |
9229 | inside_cost |
9230 | += record_stmt_cost (body_cost_vec: cost_vec, count: cnunits, kind: vec_to_scalar, |
9231 | stmt_info, misalign: 0, where: vect_body); |
9232 | inside_cost |
9233 | += record_stmt_cost (body_cost_vec: cost_vec, count: cnunits, kind: scalar_store, |
9234 | stmt_info, misalign: 0, where: vect_body); |
9235 | continue; |
9236 | } |
9237 | |
9238 | unsigned HOST_WIDE_INT const_nunits = nunits.to_constant (); |
9239 | unsigned HOST_WIDE_INT const_offset_nunits |
9240 | = TYPE_VECTOR_SUBPARTS (node: gs_info.offset_vectype).to_constant (); |
9241 | vec<constructor_elt, va_gc> *ctor_elts; |
9242 | vec_alloc (v&: ctor_elts, nelems: const_nunits); |
9243 | gimple_seq stmts = NULL; |
9244 | tree elt_type = TREE_TYPE (vectype); |
9245 | unsigned HOST_WIDE_INT elt_size |
9246 | = tree_to_uhwi (TYPE_SIZE (elt_type)); |
9247 | /* We support offset vectors with more elements |
9248 | than the data vector for now. */ |
9249 | unsigned HOST_WIDE_INT factor |
9250 | = const_offset_nunits / const_nunits; |
9251 | vec_offset = vec_offsets[(vec_num * j + i) / factor]; |
9252 | unsigned elt_offset |
9253 | = ((vec_num * j + i) % factor) * const_nunits; |
9254 | tree idx_type = TREE_TYPE (TREE_TYPE (vec_offset)); |
9255 | tree scale = size_int (gs_info.scale); |
9256 | align = get_object_alignment (DR_REF (first_dr_info->dr)); |
9257 | tree ltype = build_aligned_type (TREE_TYPE (vectype), align); |
9258 | for (unsigned k = 0; k < const_nunits; ++k) |
9259 | { |
9260 | /* Compute the offsetted pointer. */ |
9261 | tree boff = size_binop (MULT_EXPR, TYPE_SIZE (idx_type), |
9262 | bitsize_int (k + elt_offset)); |
9263 | tree idx |
9264 | = gimple_build (seq: &stmts, code: BIT_FIELD_REF, type: idx_type, |
9265 | ops: vec_offset, TYPE_SIZE (idx_type), ops: boff); |
9266 | idx = gimple_convert (seq: &stmts, sizetype, op: idx); |
9267 | idx = gimple_build (seq: &stmts, code: MULT_EXPR, sizetype, |
9268 | ops: idx, ops: scale); |
9269 | tree ptr |
9270 | = gimple_build (seq: &stmts, code: PLUS_EXPR, |
9271 | TREE_TYPE (dataref_ptr), |
9272 | ops: dataref_ptr, ops: idx); |
9273 | ptr = gimple_convert (seq: &stmts, ptr_type_node, op: ptr); |
9274 | /* Extract the element to be stored. */ |
9275 | tree elt |
9276 | = gimple_build (seq: &stmts, code: BIT_FIELD_REF, |
9277 | TREE_TYPE (vectype), |
9278 | ops: vec_oprnd, TYPE_SIZE (elt_type), |
9279 | bitsize_int (k * elt_size)); |
9280 | gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT); |
9281 | stmts = NULL; |
9282 | tree ref |
9283 | = build2 (MEM_REF, ltype, ptr, |
9284 | build_int_cst (ref_type, 0)); |
9285 | new_stmt = gimple_build_assign (ref, elt); |
9286 | vect_finish_stmt_generation (vinfo, stmt_info, |
9287 | vec_stmt: new_stmt, gsi); |
9288 | } |
9289 | if (slp) |
9290 | slp_node->push_vec_def (def: new_stmt); |
9291 | } |
9292 | } |
9293 | if (!slp && !costing_p) |
9294 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
9295 | } |
9296 | |
9297 | if (!slp && !costing_p) |
9298 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
9299 | |
9300 | if (costing_p && dump_enabled_p ()) |
9301 | dump_printf_loc (MSG_NOTE, vect_location, |
9302 | "vect_model_store_cost: inside_cost = %d, " |
9303 | "prologue_cost = %d .\n" , |
9304 | inside_cost, prologue_cost); |
9305 | |
9306 | return true; |
9307 | } |
9308 | |
9309 | gcc_assert (memory_access_type == VMAT_CONTIGUOUS |
9310 | || memory_access_type == VMAT_CONTIGUOUS_DOWN |
9311 | || memory_access_type == VMAT_CONTIGUOUS_PERMUTE |
9312 | || memory_access_type == VMAT_CONTIGUOUS_REVERSE); |
9313 | |
9314 | unsigned inside_cost = 0, prologue_cost = 0; |
9315 | /* For costing some adjacent vector stores, we'd like to cost with |
9316 | the total number of them once instead of cost each one by one. */ |
9317 | unsigned int n_adjacent_stores = 0; |
9318 | auto_vec<tree> result_chain (group_size); |
9319 | auto_vec<tree, 1> vec_oprnds; |
9320 | for (j = 0; j < ncopies; j++) |
9321 | { |
9322 | gimple *new_stmt; |
9323 | if (j == 0) |
9324 | { |
9325 | if (slp && !costing_p) |
9326 | { |
9327 | /* Get vectorized arguments for SLP_NODE. */ |
9328 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies: 1, op0: op, |
9329 | vec_oprnds0: &vec_oprnds, op1: mask, vec_oprnds1: &vec_masks); |
9330 | vec_oprnd = vec_oprnds[0]; |
9331 | if (mask) |
9332 | vec_mask = vec_masks[0]; |
9333 | } |
9334 | else |
9335 | { |
9336 | /* For interleaved stores we collect vectorized defs for all the |
9337 | stores in the group in DR_CHAIN. DR_CHAIN is then used as an |
9338 | input to vect_permute_store_chain(). |
9339 | |
9340 | If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN |
9341 | is of size 1. */ |
9342 | stmt_vec_info next_stmt_info = first_stmt_info; |
9343 | for (i = 0; i < group_size; i++) |
9344 | { |
9345 | /* Since gaps are not supported for interleaved stores, |
9346 | DR_GROUP_SIZE is the exact number of stmts in the chain. |
9347 | Therefore, NEXT_STMT_INFO can't be NULL_TREE. In case |
9348 | that there is no interleaving, DR_GROUP_SIZE is 1, |
9349 | and only one iteration of the loop will be executed. */ |
9350 | op = vect_get_store_rhs (stmt_info: next_stmt_info); |
9351 | if (costing_p) |
9352 | update_prologue_cost (&prologue_cost, op); |
9353 | else |
9354 | { |
9355 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: next_stmt_info, |
9356 | ncopies, op, |
9357 | vec_oprnds: gvec_oprnds[i]); |
9358 | vec_oprnd = (*gvec_oprnds[i])[0]; |
9359 | dr_chain.quick_push (obj: vec_oprnd); |
9360 | } |
9361 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
9362 | } |
9363 | if (mask && !costing_p) |
9364 | { |
9365 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies, |
9366 | op: mask, vec_oprnds: &vec_masks, |
9367 | vectype: mask_vectype); |
9368 | vec_mask = vec_masks[0]; |
9369 | } |
9370 | } |
9371 | |
9372 | /* We should have catched mismatched types earlier. */ |
9373 | gcc_assert (costing_p |
9374 | || useless_type_conversion_p (vectype, |
9375 | TREE_TYPE (vec_oprnd))); |
9376 | bool simd_lane_access_p |
9377 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) != 0; |
9378 | if (!costing_p |
9379 | && simd_lane_access_p |
9380 | && !loop_masks |
9381 | && TREE_CODE (DR_BASE_ADDRESS (first_dr_info->dr)) == ADDR_EXPR |
9382 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr_info->dr), 0)) |
9383 | && integer_zerop (get_dr_vinfo_offset (vinfo, dr_info: first_dr_info)) |
9384 | && integer_zerop (DR_INIT (first_dr_info->dr)) |
9385 | && alias_sets_conflict_p (get_alias_set (aggr_type), |
9386 | get_alias_set (TREE_TYPE (ref_type)))) |
9387 | { |
9388 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr_info->dr)); |
9389 | dataref_offset = build_int_cst (ref_type, 0); |
9390 | } |
9391 | else if (!costing_p) |
9392 | dataref_ptr |
9393 | = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type, |
9394 | simd_lane_access_p ? loop : NULL, |
9395 | offset, &dummy, gsi, &ptr_incr, |
9396 | simd_lane_access_p, bump); |
9397 | } |
9398 | else if (!costing_p) |
9399 | { |
9400 | gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)); |
9401 | /* DR_CHAIN is then used as an input to vect_permute_store_chain(). |
9402 | If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN is |
9403 | of size 1. */ |
9404 | for (i = 0; i < group_size; i++) |
9405 | { |
9406 | vec_oprnd = (*gvec_oprnds[i])[j]; |
9407 | dr_chain[i] = vec_oprnd; |
9408 | } |
9409 | if (mask) |
9410 | vec_mask = vec_masks[j]; |
9411 | if (dataref_offset) |
9412 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, bump); |
9413 | else |
9414 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi, |
9415 | stmt_info, bump); |
9416 | } |
9417 | |
9418 | new_stmt = NULL; |
9419 | if (grouped_store) |
9420 | { |
9421 | /* Permute. */ |
9422 | gcc_assert (memory_access_type == VMAT_CONTIGUOUS_PERMUTE); |
9423 | if (costing_p) |
9424 | { |
9425 | int group_size = DR_GROUP_SIZE (first_stmt_info); |
9426 | int nstmts = ceil_log2 (x: group_size) * group_size; |
9427 | inside_cost += record_stmt_cost (body_cost_vec: cost_vec, count: nstmts, kind: vec_perm, |
9428 | stmt_info, misalign: 0, where: vect_body); |
9429 | if (dump_enabled_p ()) |
9430 | dump_printf_loc (MSG_NOTE, vect_location, |
9431 | "vect_model_store_cost: " |
9432 | "strided group_size = %d .\n" , |
9433 | group_size); |
9434 | } |
9435 | else |
9436 | vect_permute_store_chain (vinfo, dr_chain, group_size, stmt_info, |
9437 | gsi, &result_chain); |
9438 | } |
9439 | |
9440 | stmt_vec_info next_stmt_info = first_stmt_info; |
9441 | for (i = 0; i < vec_num; i++) |
9442 | { |
9443 | if (!costing_p) |
9444 | { |
9445 | if (slp) |
9446 | vec_oprnd = vec_oprnds[i]; |
9447 | else if (grouped_store) |
9448 | /* For grouped stores vectorized defs are interleaved in |
9449 | vect_permute_store_chain(). */ |
9450 | vec_oprnd = result_chain[i]; |
9451 | } |
9452 | |
9453 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
9454 | { |
9455 | if (costing_p) |
9456 | inside_cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: vec_perm, |
9457 | stmt_info, misalign: 0, where: vect_body); |
9458 | else |
9459 | { |
9460 | tree perm_mask = perm_mask_for_reverse (vectype); |
9461 | tree perm_dest = vect_create_destination_var ( |
9462 | vect_get_store_rhs (stmt_info), vectype); |
9463 | tree new_temp = make_ssa_name (var: perm_dest); |
9464 | |
9465 | /* Generate the permute statement. */ |
9466 | gimple *perm_stmt |
9467 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
9468 | vec_oprnd, perm_mask); |
9469 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: perm_stmt, |
9470 | gsi); |
9471 | |
9472 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); |
9473 | vec_oprnd = new_temp; |
9474 | } |
9475 | } |
9476 | |
9477 | if (costing_p) |
9478 | { |
9479 | n_adjacent_stores++; |
9480 | |
9481 | if (!slp) |
9482 | { |
9483 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
9484 | if (!next_stmt_info) |
9485 | break; |
9486 | } |
9487 | |
9488 | continue; |
9489 | } |
9490 | |
9491 | tree final_mask = NULL_TREE; |
9492 | tree final_len = NULL_TREE; |
9493 | tree bias = NULL_TREE; |
9494 | if (loop_masks) |
9495 | final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks, |
9496 | vec_num * ncopies, vectype, |
9497 | vec_num * j + i); |
9498 | if (slp && vec_mask) |
9499 | vec_mask = vec_masks[i]; |
9500 | if (vec_mask) |
9501 | final_mask = prepare_vec_mask (loop_vinfo, mask_type: mask_vectype, loop_mask: final_mask, |
9502 | vec_mask, gsi); |
9503 | |
9504 | if (i > 0) |
9505 | /* Bump the vector pointer. */ |
9506 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi, |
9507 | stmt_info, bump); |
9508 | |
9509 | unsigned misalign; |
9510 | unsigned HOST_WIDE_INT align; |
9511 | align = known_alignment (DR_TARGET_ALIGNMENT (first_dr_info)); |
9512 | if (alignment_support_scheme == dr_aligned) |
9513 | misalign = 0; |
9514 | else if (misalignment == DR_MISALIGNMENT_UNKNOWN) |
9515 | { |
9516 | align = dr_alignment (vect_dr_behavior (vinfo, dr_info: first_dr_info)); |
9517 | misalign = 0; |
9518 | } |
9519 | else |
9520 | misalign = misalignment; |
9521 | if (dataref_offset == NULL_TREE |
9522 | && TREE_CODE (dataref_ptr) == SSA_NAME) |
9523 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
9524 | misalign); |
9525 | align = least_bit_hwi (x: misalign | align); |
9526 | |
9527 | /* Compute IFN when LOOP_LENS or final_mask valid. */ |
9528 | machine_mode vmode = TYPE_MODE (vectype); |
9529 | machine_mode new_vmode = vmode; |
9530 | internal_fn partial_ifn = IFN_LAST; |
9531 | if (loop_lens) |
9532 | { |
9533 | opt_machine_mode new_ovmode |
9534 | = get_len_load_store_mode (vmode, false, &partial_ifn); |
9535 | new_vmode = new_ovmode.require (); |
9536 | unsigned factor |
9537 | = (new_ovmode == vmode) ? 1 : GET_MODE_UNIT_SIZE (vmode); |
9538 | final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens, |
9539 | vec_num * ncopies, vectype, |
9540 | vec_num * j + i, factor); |
9541 | } |
9542 | else if (final_mask) |
9543 | { |
9544 | if (!can_vec_mask_load_store_p ( |
9545 | vmode, TYPE_MODE (TREE_TYPE (final_mask)), false, |
9546 | &partial_ifn)) |
9547 | gcc_unreachable (); |
9548 | } |
9549 | |
9550 | if (partial_ifn == IFN_MASK_LEN_STORE) |
9551 | { |
9552 | if (!final_len) |
9553 | { |
9554 | /* Pass VF value to 'len' argument of |
9555 | MASK_LEN_STORE if LOOP_LENS is invalid. */ |
9556 | final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
9557 | } |
9558 | if (!final_mask) |
9559 | { |
9560 | /* Pass all ones value to 'mask' argument of |
9561 | MASK_LEN_STORE if final_mask is invalid. */ |
9562 | mask_vectype = truth_type_for (vectype); |
9563 | final_mask = build_minus_one_cst (mask_vectype); |
9564 | } |
9565 | } |
9566 | if (final_len) |
9567 | { |
9568 | signed char biasval |
9569 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
9570 | |
9571 | bias = build_int_cst (intQI_type_node, biasval); |
9572 | } |
9573 | |
9574 | /* Arguments are ready. Create the new vector stmt. */ |
9575 | if (final_len) |
9576 | { |
9577 | gcall *call; |
9578 | tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT); |
9579 | /* Need conversion if it's wrapped with VnQI. */ |
9580 | if (vmode != new_vmode) |
9581 | { |
9582 | tree new_vtype |
9583 | = build_vector_type_for_mode (unsigned_intQI_type_node, |
9584 | new_vmode); |
9585 | tree var = vect_get_new_ssa_name (new_vtype, vect_simple_var); |
9586 | vec_oprnd = build1 (VIEW_CONVERT_EXPR, new_vtype, vec_oprnd); |
9587 | gassign *new_stmt |
9588 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, vec_oprnd); |
9589 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
9590 | vec_oprnd = var; |
9591 | } |
9592 | |
9593 | if (partial_ifn == IFN_MASK_LEN_STORE) |
9594 | call = gimple_build_call_internal (IFN_MASK_LEN_STORE, 6, |
9595 | dataref_ptr, ptr, final_mask, |
9596 | final_len, bias, vec_oprnd); |
9597 | else |
9598 | call = gimple_build_call_internal (IFN_LEN_STORE, 5, |
9599 | dataref_ptr, ptr, final_len, |
9600 | bias, vec_oprnd); |
9601 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
9602 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
9603 | new_stmt = call; |
9604 | } |
9605 | else if (final_mask) |
9606 | { |
9607 | tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT); |
9608 | gcall *call |
9609 | = gimple_build_call_internal (IFN_MASK_STORE, 4, dataref_ptr, |
9610 | ptr, final_mask, vec_oprnd); |
9611 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
9612 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
9613 | new_stmt = call; |
9614 | } |
9615 | else |
9616 | { |
9617 | data_ref |
9618 | = fold_build2 (MEM_REF, vectype, dataref_ptr, |
9619 | dataref_offset ? dataref_offset |
9620 | : build_int_cst (ref_type, 0)); |
9621 | if (alignment_support_scheme == dr_aligned) |
9622 | ; |
9623 | else |
9624 | TREE_TYPE (data_ref) |
9625 | = build_aligned_type (TREE_TYPE (data_ref), |
9626 | align * BITS_PER_UNIT); |
9627 | vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr)); |
9628 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); |
9629 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
9630 | } |
9631 | |
9632 | if (slp) |
9633 | continue; |
9634 | |
9635 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
9636 | if (!next_stmt_info) |
9637 | break; |
9638 | } |
9639 | if (!slp && !costing_p) |
9640 | { |
9641 | if (j == 0) |
9642 | *vec_stmt = new_stmt; |
9643 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
9644 | } |
9645 | } |
9646 | |
9647 | if (costing_p) |
9648 | { |
9649 | if (n_adjacent_stores > 0) |
9650 | vect_get_store_cost (vinfo, stmt_info, ncopies: n_adjacent_stores, |
9651 | alignment_support_scheme, misalignment, |
9652 | inside_cost: &inside_cost, body_cost_vec: cost_vec); |
9653 | |
9654 | /* When vectorizing a store into the function result assign |
9655 | a penalty if the function returns in a multi-register location. |
9656 | In this case we assume we'll end up with having to spill the |
9657 | vector result and do piecewise loads as a conservative estimate. */ |
9658 | tree base = get_base_address (STMT_VINFO_DATA_REF (stmt_info)->ref); |
9659 | if (base |
9660 | && (TREE_CODE (base) == RESULT_DECL |
9661 | || (DECL_P (base) && cfun_returns (decl: base))) |
9662 | && !aggregate_value_p (base, cfun->decl)) |
9663 | { |
9664 | rtx reg = hard_function_value (TREE_TYPE (base), cfun->decl, 0, 1); |
9665 | /* ??? Handle PARALLEL in some way. */ |
9666 | if (REG_P (reg)) |
9667 | { |
9668 | int nregs = hard_regno_nregs (REGNO (reg), GET_MODE (reg)); |
9669 | /* Assume that a single reg-reg move is possible and cheap, |
9670 | do not account for vector to gp register move cost. */ |
9671 | if (nregs > 1) |
9672 | { |
9673 | /* Spill. */ |
9674 | prologue_cost |
9675 | += record_stmt_cost (body_cost_vec: cost_vec, count: ncopies, kind: vector_store, |
9676 | stmt_info, misalign: 0, where: vect_epilogue); |
9677 | /* Loads. */ |
9678 | prologue_cost |
9679 | += record_stmt_cost (body_cost_vec: cost_vec, count: ncopies * nregs, kind: scalar_load, |
9680 | stmt_info, misalign: 0, where: vect_epilogue); |
9681 | } |
9682 | } |
9683 | } |
9684 | if (dump_enabled_p ()) |
9685 | dump_printf_loc (MSG_NOTE, vect_location, |
9686 | "vect_model_store_cost: inside_cost = %d, " |
9687 | "prologue_cost = %d .\n" , |
9688 | inside_cost, prologue_cost); |
9689 | } |
9690 | |
9691 | return true; |
9692 | } |
9693 | |
9694 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
9695 | VECTOR_CST mask. No checks are made that the target platform supports the |
9696 | mask, so callers may wish to test can_vec_perm_const_p separately, or use |
9697 | vect_gen_perm_mask_checked. */ |
9698 | |
9699 | tree |
9700 | vect_gen_perm_mask_any (tree vectype, const vec_perm_indices &sel) |
9701 | { |
9702 | tree mask_type; |
9703 | |
9704 | poly_uint64 nunits = sel.length (); |
9705 | gcc_assert (known_eq (nunits, TYPE_VECTOR_SUBPARTS (vectype))); |
9706 | |
9707 | mask_type = build_vector_type (ssizetype, nunits); |
9708 | return vec_perm_indices_to_tree (mask_type, sel); |
9709 | } |
9710 | |
9711 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_const_p, |
9712 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ |
9713 | |
9714 | tree |
9715 | vect_gen_perm_mask_checked (tree vectype, const vec_perm_indices &sel) |
9716 | { |
9717 | machine_mode vmode = TYPE_MODE (vectype); |
9718 | gcc_assert (can_vec_perm_const_p (vmode, vmode, sel)); |
9719 | return vect_gen_perm_mask_any (vectype, sel); |
9720 | } |
9721 | |
9722 | /* Given a vector variable X and Y, that was generated for the scalar |
9723 | STMT_INFO, generate instructions to permute the vector elements of X and Y |
9724 | using permutation mask MASK_VEC, insert them at *GSI and return the |
9725 | permuted vector variable. */ |
9726 | |
9727 | static tree |
9728 | permute_vec_elements (vec_info *vinfo, |
9729 | tree x, tree y, tree mask_vec, stmt_vec_info stmt_info, |
9730 | gimple_stmt_iterator *gsi) |
9731 | { |
9732 | tree vectype = TREE_TYPE (x); |
9733 | tree perm_dest, data_ref; |
9734 | gimple *perm_stmt; |
9735 | |
9736 | tree scalar_dest = gimple_get_lhs (stmt_info->stmt); |
9737 | if (scalar_dest && TREE_CODE (scalar_dest) == SSA_NAME) |
9738 | perm_dest = vect_create_destination_var (scalar_dest, vectype); |
9739 | else |
9740 | perm_dest = vect_get_new_vect_var (vectype, vect_simple_var, NULL); |
9741 | data_ref = make_ssa_name (var: perm_dest); |
9742 | |
9743 | /* Generate the permute statement. */ |
9744 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
9745 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: perm_stmt, gsi); |
9746 | |
9747 | return data_ref; |
9748 | } |
9749 | |
9750 | /* Hoist the definitions of all SSA uses on STMT_INFO out of the loop LOOP, |
9751 | inserting them on the loops preheader edge. Returns true if we |
9752 | were successful in doing so (and thus STMT_INFO can be moved then), |
9753 | otherwise returns false. HOIST_P indicates if we want to hoist the |
9754 | definitions of all SSA uses, it would be false when we are costing. */ |
9755 | |
9756 | static bool |
9757 | hoist_defs_of_uses (stmt_vec_info stmt_info, class loop *loop, bool hoist_p) |
9758 | { |
9759 | ssa_op_iter i; |
9760 | tree op; |
9761 | bool any = false; |
9762 | |
9763 | FOR_EACH_SSA_TREE_OPERAND (op, stmt_info->stmt, i, SSA_OP_USE) |
9764 | { |
9765 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
9766 | if (!gimple_nop_p (g: def_stmt) |
9767 | && flow_bb_inside_loop_p (loop, gimple_bb (g: def_stmt))) |
9768 | { |
9769 | /* Make sure we don't need to recurse. While we could do |
9770 | so in simple cases when there are more complex use webs |
9771 | we don't have an easy way to preserve stmt order to fulfil |
9772 | dependencies within them. */ |
9773 | tree op2; |
9774 | ssa_op_iter i2; |
9775 | if (gimple_code (g: def_stmt) == GIMPLE_PHI) |
9776 | return false; |
9777 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
9778 | { |
9779 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
9780 | if (!gimple_nop_p (g: def_stmt2) |
9781 | && flow_bb_inside_loop_p (loop, gimple_bb (g: def_stmt2))) |
9782 | return false; |
9783 | } |
9784 | any = true; |
9785 | } |
9786 | } |
9787 | |
9788 | if (!any) |
9789 | return true; |
9790 | |
9791 | if (!hoist_p) |
9792 | return true; |
9793 | |
9794 | FOR_EACH_SSA_TREE_OPERAND (op, stmt_info->stmt, i, SSA_OP_USE) |
9795 | { |
9796 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
9797 | if (!gimple_nop_p (g: def_stmt) |
9798 | && flow_bb_inside_loop_p (loop, gimple_bb (g: def_stmt))) |
9799 | { |
9800 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); |
9801 | gsi_remove (&gsi, false); |
9802 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); |
9803 | } |
9804 | } |
9805 | |
9806 | return true; |
9807 | } |
9808 | |
9809 | /* vectorizable_load. |
9810 | |
9811 | Check if STMT_INFO reads a non scalar data-ref (array/pointer/structure) |
9812 | that can be vectorized. |
9813 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
9814 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
9815 | Return true if STMT_INFO is vectorizable in this way. */ |
9816 | |
9817 | static bool |
9818 | vectorizable_load (vec_info *vinfo, |
9819 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
9820 | gimple **vec_stmt, slp_tree slp_node, |
9821 | stmt_vector_for_cost *cost_vec) |
9822 | { |
9823 | tree scalar_dest; |
9824 | tree vec_dest = NULL; |
9825 | tree data_ref = NULL; |
9826 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
9827 | class loop *loop = NULL; |
9828 | class loop *containing_loop = gimple_bb (g: stmt_info->stmt)->loop_father; |
9829 | bool nested_in_vect_loop = false; |
9830 | tree elem_type; |
9831 | /* Avoid false positive uninitialized warning, see PR110652. */ |
9832 | tree new_temp = NULL_TREE; |
9833 | machine_mode mode; |
9834 | tree dummy; |
9835 | tree dataref_ptr = NULL_TREE; |
9836 | tree dataref_offset = NULL_TREE; |
9837 | gimple *ptr_incr = NULL; |
9838 | int ncopies; |
9839 | int i, j; |
9840 | unsigned int group_size; |
9841 | poly_uint64 group_gap_adj; |
9842 | tree msq = NULL_TREE, lsq; |
9843 | tree realignment_token = NULL_TREE; |
9844 | gphi *phi = NULL; |
9845 | vec<tree> dr_chain = vNULL; |
9846 | bool grouped_load = false; |
9847 | stmt_vec_info first_stmt_info; |
9848 | stmt_vec_info first_stmt_info_for_drptr = NULL; |
9849 | bool compute_in_loop = false; |
9850 | class loop *at_loop; |
9851 | int vec_num; |
9852 | bool slp = (slp_node != NULL); |
9853 | bool slp_perm = false; |
9854 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
9855 | poly_uint64 vf; |
9856 | tree aggr_type; |
9857 | gather_scatter_info gs_info; |
9858 | tree ref_type; |
9859 | enum vect_def_type mask_dt = vect_unknown_def_type; |
9860 | |
9861 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
9862 | return false; |
9863 | |
9864 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
9865 | && ! vec_stmt) |
9866 | return false; |
9867 | |
9868 | if (!STMT_VINFO_DATA_REF (stmt_info)) |
9869 | return false; |
9870 | |
9871 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
9872 | int mask_index = -1; |
9873 | slp_tree slp_op = NULL; |
9874 | if (gassign *assign = dyn_cast <gassign *> (p: stmt_info->stmt)) |
9875 | { |
9876 | scalar_dest = gimple_assign_lhs (gs: assign); |
9877 | if (TREE_CODE (scalar_dest) != SSA_NAME) |
9878 | return false; |
9879 | |
9880 | tree_code code = gimple_assign_rhs_code (gs: assign); |
9881 | if (code != ARRAY_REF |
9882 | && code != BIT_FIELD_REF |
9883 | && code != INDIRECT_REF |
9884 | && code != COMPONENT_REF |
9885 | && code != IMAGPART_EXPR |
9886 | && code != REALPART_EXPR |
9887 | && code != MEM_REF |
9888 | && TREE_CODE_CLASS (code) != tcc_declaration) |
9889 | return false; |
9890 | } |
9891 | else |
9892 | { |
9893 | gcall *call = dyn_cast <gcall *> (p: stmt_info->stmt); |
9894 | if (!call || !gimple_call_internal_p (gs: call)) |
9895 | return false; |
9896 | |
9897 | internal_fn ifn = gimple_call_internal_fn (gs: call); |
9898 | if (!internal_load_fn_p (ifn)) |
9899 | return false; |
9900 | |
9901 | scalar_dest = gimple_call_lhs (gs: call); |
9902 | if (!scalar_dest) |
9903 | return false; |
9904 | |
9905 | mask_index = internal_fn_mask_index (ifn); |
9906 | if (mask_index >= 0 && slp_node) |
9907 | mask_index = vect_slp_child_index_for_operand |
9908 | (call, op: mask_index, STMT_VINFO_GATHER_SCATTER_P (stmt_info)); |
9909 | if (mask_index >= 0 |
9910 | && !vect_check_scalar_mask (vinfo, stmt_info, slp_node, mask_index, |
9911 | mask: &mask, mask_node: &slp_op, mask_dt_out: &mask_dt, mask_vectype_out: &mask_vectype)) |
9912 | return false; |
9913 | } |
9914 | |
9915 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
9916 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
9917 | |
9918 | if (loop_vinfo) |
9919 | { |
9920 | loop = LOOP_VINFO_LOOP (loop_vinfo); |
9921 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt_info); |
9922 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); |
9923 | } |
9924 | else |
9925 | vf = 1; |
9926 | |
9927 | /* Multiple types in SLP are handled by creating the appropriate number of |
9928 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
9929 | case of SLP. */ |
9930 | if (slp) |
9931 | ncopies = 1; |
9932 | else |
9933 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
9934 | |
9935 | gcc_assert (ncopies >= 1); |
9936 | |
9937 | /* FORNOW. This restriction should be relaxed. */ |
9938 | if (nested_in_vect_loop && ncopies > 1) |
9939 | { |
9940 | if (dump_enabled_p ()) |
9941 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9942 | "multiple types in nested loop.\n" ); |
9943 | return false; |
9944 | } |
9945 | |
9946 | /* Invalidate assumptions made by dependence analysis when vectorization |
9947 | on the unrolled body effectively re-orders stmts. */ |
9948 | if (ncopies > 1 |
9949 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 |
9950 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
9951 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) |
9952 | { |
9953 | if (dump_enabled_p ()) |
9954 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9955 | "cannot perform implicit CSE when unrolling " |
9956 | "with negative dependence distance\n" ); |
9957 | return false; |
9958 | } |
9959 | |
9960 | elem_type = TREE_TYPE (vectype); |
9961 | mode = TYPE_MODE (vectype); |
9962 | |
9963 | /* FORNOW. In some cases can vectorize even if data-type not supported |
9964 | (e.g. - data copies). */ |
9965 | if (optab_handler (op: mov_optab, mode) == CODE_FOR_nothing) |
9966 | { |
9967 | if (dump_enabled_p ()) |
9968 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9969 | "Aligned load, but unsupported type.\n" ); |
9970 | return false; |
9971 | } |
9972 | |
9973 | /* Check if the load is a part of an interleaving chain. */ |
9974 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
9975 | { |
9976 | grouped_load = true; |
9977 | /* FORNOW */ |
9978 | gcc_assert (!nested_in_vect_loop); |
9979 | gcc_assert (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)); |
9980 | |
9981 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
9982 | group_size = DR_GROUP_SIZE (first_stmt_info); |
9983 | |
9984 | /* Refuse non-SLP vectorization of SLP-only groups. */ |
9985 | if (!slp && STMT_VINFO_SLP_VECT_ONLY (first_stmt_info)) |
9986 | { |
9987 | if (dump_enabled_p ()) |
9988 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9989 | "cannot vectorize load in non-SLP mode.\n" ); |
9990 | return false; |
9991 | } |
9992 | |
9993 | /* Invalidate assumptions made by dependence analysis when vectorization |
9994 | on the unrolled body effectively re-orders stmts. */ |
9995 | if (!PURE_SLP_STMT (stmt_info) |
9996 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 |
9997 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
9998 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) |
9999 | { |
10000 | if (dump_enabled_p ()) |
10001 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
10002 | "cannot perform implicit CSE when performing " |
10003 | "group loads with negative dependence distance\n" ); |
10004 | return false; |
10005 | } |
10006 | } |
10007 | else |
10008 | group_size = 1; |
10009 | |
10010 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
10011 | { |
10012 | slp_perm = true; |
10013 | |
10014 | if (!loop_vinfo) |
10015 | { |
10016 | /* In BB vectorization we may not actually use a loaded vector |
10017 | accessing elements in excess of DR_GROUP_SIZE. */ |
10018 | stmt_vec_info group_info = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
10019 | group_info = DR_GROUP_FIRST_ELEMENT (group_info); |
10020 | unsigned HOST_WIDE_INT nunits; |
10021 | unsigned j, k, maxk = 0; |
10022 | FOR_EACH_VEC_ELT (SLP_TREE_LOAD_PERMUTATION (slp_node), j, k) |
10023 | if (k > maxk) |
10024 | maxk = k; |
10025 | tree vectype = SLP_TREE_VECTYPE (slp_node); |
10026 | if (!TYPE_VECTOR_SUBPARTS (node: vectype).is_constant (const_value: &nunits) |
10027 | || maxk >= (DR_GROUP_SIZE (group_info) & ~(nunits - 1))) |
10028 | { |
10029 | if (dump_enabled_p ()) |
10030 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
10031 | "BB vectorization with gaps at the end of " |
10032 | "a load is not supported\n" ); |
10033 | return false; |
10034 | } |
10035 | } |
10036 | |
10037 | auto_vec<tree> tem; |
10038 | unsigned n_perms; |
10039 | if (!vect_transform_slp_perm_load (vinfo, slp_node, tem, NULL, vf, |
10040 | true, &n_perms)) |
10041 | { |
10042 | if (dump_enabled_p ()) |
10043 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, |
10044 | vect_location, |
10045 | "unsupported load permutation\n" ); |
10046 | return false; |
10047 | } |
10048 | } |
10049 | |
10050 | vect_memory_access_type memory_access_type; |
10051 | enum dr_alignment_support alignment_support_scheme; |
10052 | int misalignment; |
10053 | poly_int64 poffset; |
10054 | internal_fn lanes_ifn; |
10055 | if (!get_load_store_type (vinfo, stmt_info, vectype, slp_node, masked_p: mask, vls_type: VLS_LOAD, |
10056 | ncopies, memory_access_type: &memory_access_type, poffset: &poffset, |
10057 | alignment_support_scheme: &alignment_support_scheme, misalignment: &misalignment, gs_info: &gs_info, |
10058 | lanes_ifn: &lanes_ifn)) |
10059 | return false; |
10060 | |
10061 | if (mask) |
10062 | { |
10063 | if (memory_access_type == VMAT_CONTIGUOUS) |
10064 | { |
10065 | machine_mode vec_mode = TYPE_MODE (vectype); |
10066 | if (!VECTOR_MODE_P (vec_mode) |
10067 | || !can_vec_mask_load_store_p (vec_mode, |
10068 | TYPE_MODE (mask_vectype), true)) |
10069 | return false; |
10070 | } |
10071 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
10072 | && memory_access_type != VMAT_GATHER_SCATTER) |
10073 | { |
10074 | if (dump_enabled_p ()) |
10075 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
10076 | "unsupported access type for masked load.\n" ); |
10077 | return false; |
10078 | } |
10079 | else if (memory_access_type == VMAT_GATHER_SCATTER |
10080 | && gs_info.ifn == IFN_LAST |
10081 | && !gs_info.decl) |
10082 | { |
10083 | if (dump_enabled_p ()) |
10084 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
10085 | "unsupported masked emulated gather.\n" ); |
10086 | return false; |
10087 | } |
10088 | else if (memory_access_type == VMAT_ELEMENTWISE |
10089 | || memory_access_type == VMAT_STRIDED_SLP) |
10090 | { |
10091 | if (dump_enabled_p ()) |
10092 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
10093 | "unsupported masked strided access.\n" ); |
10094 | return false; |
10095 | } |
10096 | } |
10097 | |
10098 | bool costing_p = !vec_stmt; |
10099 | |
10100 | if (costing_p) /* transformation not required. */ |
10101 | { |
10102 | if (slp_node |
10103 | && mask |
10104 | && !vect_maybe_update_slp_op_vectype (slp_op, |
10105 | mask_vectype)) |
10106 | { |
10107 | if (dump_enabled_p ()) |
10108 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
10109 | "incompatible vector types for invariants\n" ); |
10110 | return false; |
10111 | } |
10112 | |
10113 | if (!slp) |
10114 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
10115 | |
10116 | if (loop_vinfo |
10117 | && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) |
10118 | check_load_store_for_partial_vectors (loop_vinfo, vectype, slp_node, |
10119 | vls_type: VLS_LOAD, group_size, |
10120 | memory_access_type, gs_info: &gs_info, |
10121 | scalar_mask: mask); |
10122 | |
10123 | if (dump_enabled_p () |
10124 | && memory_access_type != VMAT_ELEMENTWISE |
10125 | && memory_access_type != VMAT_GATHER_SCATTER |
10126 | && alignment_support_scheme != dr_aligned) |
10127 | dump_printf_loc (MSG_NOTE, vect_location, |
10128 | "Vectorizing an unaligned access.\n" ); |
10129 | |
10130 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
10131 | vinfo->any_known_not_updated_vssa = true; |
10132 | |
10133 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; |
10134 | } |
10135 | |
10136 | if (!slp) |
10137 | gcc_assert (memory_access_type |
10138 | == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
10139 | |
10140 | if (dump_enabled_p () && !costing_p) |
10141 | dump_printf_loc (MSG_NOTE, vect_location, |
10142 | "transform load. ncopies = %d\n" , ncopies); |
10143 | |
10144 | /* Transform. */ |
10145 | |
10146 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info), *first_dr_info = NULL; |
10147 | ensure_base_align (dr_info); |
10148 | |
10149 | if (memory_access_type == VMAT_INVARIANT) |
10150 | { |
10151 | gcc_assert (!grouped_load && !mask && !bb_vinfo); |
10152 | /* If we have versioned for aliasing or the loop doesn't |
10153 | have any data dependencies that would preclude this, |
10154 | then we are sure this is a loop invariant load and |
10155 | thus we can insert it on the preheader edge. |
10156 | TODO: hoist_defs_of_uses should ideally be computed |
10157 | once at analysis time, remembered and used in the |
10158 | transform time. */ |
10159 | bool hoist_p = (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) |
10160 | && !nested_in_vect_loop |
10161 | && hoist_defs_of_uses (stmt_info, loop, hoist_p: !costing_p)); |
10162 | if (costing_p) |
10163 | { |
10164 | enum vect_cost_model_location cost_loc |
10165 | = hoist_p ? vect_prologue : vect_body; |
10166 | unsigned int cost = record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: scalar_load, |
10167 | stmt_info, misalign: 0, where: cost_loc); |
10168 | cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: scalar_to_vec, stmt_info, misalign: 0, |
10169 | where: cost_loc); |
10170 | unsigned int prologue_cost = hoist_p ? cost : 0; |
10171 | unsigned int inside_cost = hoist_p ? 0 : cost; |
10172 | if (dump_enabled_p ()) |
10173 | dump_printf_loc (MSG_NOTE, vect_location, |
10174 | "vect_model_load_cost: inside_cost = %d, " |
10175 | "prologue_cost = %d .\n" , |
10176 | inside_cost, prologue_cost); |
10177 | return true; |
10178 | } |
10179 | if (hoist_p) |
10180 | { |
10181 | gassign *stmt = as_a <gassign *> (p: stmt_info->stmt); |
10182 | if (dump_enabled_p ()) |
10183 | dump_printf_loc (MSG_NOTE, vect_location, |
10184 | "hoisting out of the vectorized loop: %G" , |
10185 | (gimple *) stmt); |
10186 | scalar_dest = copy_ssa_name (var: scalar_dest); |
10187 | tree rhs = unshare_expr (gimple_assign_rhs1 (gs: stmt)); |
10188 | edge pe = loop_preheader_edge (loop); |
10189 | gphi *vphi = get_virtual_phi (loop->header); |
10190 | tree vuse; |
10191 | if (vphi) |
10192 | vuse = PHI_ARG_DEF_FROM_EDGE (vphi, pe); |
10193 | else |
10194 | vuse = gimple_vuse (g: gsi_stmt (i: *gsi)); |
10195 | gimple *new_stmt = gimple_build_assign (scalar_dest, rhs); |
10196 | gimple_set_vuse (g: new_stmt, vuse); |
10197 | gsi_insert_on_edge_immediate (pe, new_stmt); |
10198 | } |
10199 | /* These copies are all equivalent. */ |
10200 | if (hoist_p) |
10201 | new_temp = vect_init_vector (vinfo, stmt_info, val: scalar_dest, |
10202 | type: vectype, NULL); |
10203 | else |
10204 | { |
10205 | gimple_stmt_iterator gsi2 = *gsi; |
10206 | gsi_next (i: &gsi2); |
10207 | new_temp = vect_init_vector (vinfo, stmt_info, val: scalar_dest, |
10208 | type: vectype, gsi: &gsi2); |
10209 | } |
10210 | gimple *new_stmt = SSA_NAME_DEF_STMT (new_temp); |
10211 | if (slp) |
10212 | for (j = 0; j < (int) SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); ++j) |
10213 | slp_node->push_vec_def (def: new_stmt); |
10214 | else |
10215 | { |
10216 | for (j = 0; j < ncopies; ++j) |
10217 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
10218 | *vec_stmt = new_stmt; |
10219 | } |
10220 | return true; |
10221 | } |
10222 | |
10223 | if (memory_access_type == VMAT_ELEMENTWISE |
10224 | || memory_access_type == VMAT_STRIDED_SLP) |
10225 | { |
10226 | gimple_stmt_iterator incr_gsi; |
10227 | bool insert_after; |
10228 | tree offvar; |
10229 | tree ivstep; |
10230 | tree running_off; |
10231 | vec<constructor_elt, va_gc> *v = NULL; |
10232 | tree stride_base, stride_step, alias_off; |
10233 | /* Checked by get_load_store_type. */ |
10234 | unsigned int const_nunits = nunits.to_constant (); |
10235 | unsigned HOST_WIDE_INT cst_offset = 0; |
10236 | tree dr_offset; |
10237 | unsigned int inside_cost = 0; |
10238 | |
10239 | gcc_assert (!LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo)); |
10240 | gcc_assert (!nested_in_vect_loop); |
10241 | |
10242 | if (grouped_load) |
10243 | { |
10244 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
10245 | first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info); |
10246 | } |
10247 | else |
10248 | { |
10249 | first_stmt_info = stmt_info; |
10250 | first_dr_info = dr_info; |
10251 | } |
10252 | |
10253 | if (slp && grouped_load) |
10254 | { |
10255 | group_size = DR_GROUP_SIZE (first_stmt_info); |
10256 | ref_type = get_group_alias_ptr_type (first_stmt_info); |
10257 | } |
10258 | else |
10259 | { |
10260 | if (grouped_load) |
10261 | cst_offset |
10262 | = (tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))) |
10263 | * vect_get_place_in_interleaving_chain (stmt_info, |
10264 | first_stmt_info)); |
10265 | group_size = 1; |
10266 | ref_type = reference_alias_ptr_type (DR_REF (dr_info->dr)); |
10267 | } |
10268 | |
10269 | if (!costing_p) |
10270 | { |
10271 | dr_offset = get_dr_vinfo_offset (vinfo, dr_info: first_dr_info); |
10272 | stride_base = fold_build_pointer_plus ( |
10273 | DR_BASE_ADDRESS (first_dr_info->dr), |
10274 | size_binop (PLUS_EXPR, convert_to_ptrofftype (dr_offset), |
10275 | convert_to_ptrofftype (DR_INIT (first_dr_info->dr)))); |
10276 | stride_step = fold_convert (sizetype, DR_STEP (first_dr_info->dr)); |
10277 | |
10278 | /* For a load with loop-invariant (but other than power-of-2) |
10279 | stride (i.e. not a grouped access) like so: |
10280 | |
10281 | for (i = 0; i < n; i += stride) |
10282 | ... = array[i]; |
10283 | |
10284 | we generate a new induction variable and new accesses to |
10285 | form a new vector (or vectors, depending on ncopies): |
10286 | |
10287 | for (j = 0; ; j += VF*stride) |
10288 | tmp1 = array[j]; |
10289 | tmp2 = array[j + stride]; |
10290 | ... |
10291 | vectemp = {tmp1, tmp2, ...} |
10292 | */ |
10293 | |
10294 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
10295 | build_int_cst (TREE_TYPE (stride_step), vf)); |
10296 | |
10297 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); |
10298 | |
10299 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
10300 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); |
10301 | create_iv (stride_base, PLUS_EXPR, ivstep, NULL, |
10302 | loop, &incr_gsi, insert_after, |
10303 | &offvar, NULL); |
10304 | |
10305 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
10306 | } |
10307 | |
10308 | running_off = offvar; |
10309 | alias_off = build_int_cst (ref_type, 0); |
10310 | int nloads = const_nunits; |
10311 | int lnel = 1; |
10312 | tree ltype = TREE_TYPE (vectype); |
10313 | tree lvectype = vectype; |
10314 | auto_vec<tree> dr_chain; |
10315 | if (memory_access_type == VMAT_STRIDED_SLP) |
10316 | { |
10317 | if (group_size < const_nunits) |
10318 | { |
10319 | /* First check if vec_init optab supports construction from vector |
10320 | elts directly. Otherwise avoid emitting a constructor of |
10321 | vector elements by performing the loads using an integer type |
10322 | of the same size, constructing a vector of those and then |
10323 | re-interpreting it as the original vector type. This avoids a |
10324 | huge runtime penalty due to the general inability to perform |
10325 | store forwarding from smaller stores to a larger load. */ |
10326 | tree ptype; |
10327 | tree vtype |
10328 | = vector_vector_composition_type (vtype: vectype, |
10329 | nelts: const_nunits / group_size, |
10330 | ptype: &ptype); |
10331 | if (vtype != NULL_TREE) |
10332 | { |
10333 | nloads = const_nunits / group_size; |
10334 | lnel = group_size; |
10335 | lvectype = vtype; |
10336 | ltype = ptype; |
10337 | } |
10338 | } |
10339 | else |
10340 | { |
10341 | nloads = 1; |
10342 | lnel = const_nunits; |
10343 | ltype = vectype; |
10344 | } |
10345 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); |
10346 | } |
10347 | /* Load vector(1) scalar_type if it's 1 element-wise vectype. */ |
10348 | else if (nloads == 1) |
10349 | ltype = vectype; |
10350 | |
10351 | if (slp) |
10352 | { |
10353 | /* For SLP permutation support we need to load the whole group, |
10354 | not only the number of vector stmts the permutation result |
10355 | fits in. */ |
10356 | if (slp_perm) |
10357 | { |
10358 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
10359 | variable VF. */ |
10360 | unsigned int const_vf = vf.to_constant (); |
10361 | ncopies = CEIL (group_size * const_vf, const_nunits); |
10362 | dr_chain.create (nelems: ncopies); |
10363 | } |
10364 | else |
10365 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
10366 | } |
10367 | unsigned int group_el = 0; |
10368 | unsigned HOST_WIDE_INT |
10369 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); |
10370 | unsigned int n_groups = 0; |
10371 | /* For costing some adjacent vector loads, we'd like to cost with |
10372 | the total number of them once instead of cost each one by one. */ |
10373 | unsigned int n_adjacent_loads = 0; |
10374 | for (j = 0; j < ncopies; j++) |
10375 | { |
10376 | if (nloads > 1 && !costing_p) |
10377 | vec_alloc (v, nelems: nloads); |
10378 | gimple *new_stmt = NULL; |
10379 | for (i = 0; i < nloads; i++) |
10380 | { |
10381 | if (costing_p) |
10382 | { |
10383 | /* For VMAT_ELEMENTWISE, just cost it as scalar_load to |
10384 | avoid ICE, see PR110776. */ |
10385 | if (VECTOR_TYPE_P (ltype) |
10386 | && memory_access_type != VMAT_ELEMENTWISE) |
10387 | n_adjacent_loads++; |
10388 | else |
10389 | inside_cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: scalar_load, |
10390 | stmt_info, misalign: 0, where: vect_body); |
10391 | continue; |
10392 | } |
10393 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
10394 | group_el * elsz + cst_offset); |
10395 | tree data_ref = build2 (MEM_REF, ltype, running_off, this_off); |
10396 | vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr)); |
10397 | new_stmt = gimple_build_assign (make_ssa_name (var: ltype), data_ref); |
10398 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
10399 | if (nloads > 1) |
10400 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, |
10401 | gimple_assign_lhs (new_stmt)); |
10402 | |
10403 | group_el += lnel; |
10404 | if (! slp |
10405 | || group_el == group_size) |
10406 | { |
10407 | n_groups++; |
10408 | /* When doing SLP make sure to not load elements from |
10409 | the next vector iteration, those will not be accessed |
10410 | so just use the last element again. See PR107451. */ |
10411 | if (!slp || known_lt (n_groups, vf)) |
10412 | { |
10413 | tree newoff = copy_ssa_name (var: running_off); |
10414 | gimple *incr |
10415 | = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
10416 | running_off, stride_step); |
10417 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: incr, gsi); |
10418 | running_off = newoff; |
10419 | } |
10420 | group_el = 0; |
10421 | } |
10422 | } |
10423 | |
10424 | if (nloads > 1) |
10425 | { |
10426 | if (costing_p) |
10427 | inside_cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: vec_construct, |
10428 | stmt_info, misalign: 0, where: vect_body); |
10429 | else |
10430 | { |
10431 | tree vec_inv = build_constructor (lvectype, v); |
10432 | new_temp = vect_init_vector (vinfo, stmt_info, val: vec_inv, |
10433 | type: lvectype, gsi); |
10434 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
10435 | if (lvectype != vectype) |
10436 | { |
10437 | new_stmt |
10438 | = gimple_build_assign (make_ssa_name (var: vectype), |
10439 | VIEW_CONVERT_EXPR, |
10440 | build1 (VIEW_CONVERT_EXPR, |
10441 | vectype, new_temp)); |
10442 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, |
10443 | gsi); |
10444 | } |
10445 | } |
10446 | } |
10447 | |
10448 | if (!costing_p) |
10449 | { |
10450 | if (slp) |
10451 | { |
10452 | if (slp_perm) |
10453 | dr_chain.quick_push (obj: gimple_assign_lhs (gs: new_stmt)); |
10454 | else |
10455 | slp_node->push_vec_def (def: new_stmt); |
10456 | } |
10457 | else |
10458 | { |
10459 | if (j == 0) |
10460 | *vec_stmt = new_stmt; |
10461 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
10462 | } |
10463 | } |
10464 | } |
10465 | if (slp_perm) |
10466 | { |
10467 | unsigned n_perms; |
10468 | if (costing_p) |
10469 | { |
10470 | unsigned n_loads; |
10471 | vect_transform_slp_perm_load (vinfo, slp_node, vNULL, NULL, vf, |
10472 | true, &n_perms, &n_loads); |
10473 | inside_cost += record_stmt_cost (body_cost_vec: cost_vec, count: n_perms, kind: vec_perm, |
10474 | stmt_info: first_stmt_info, misalign: 0, where: vect_body); |
10475 | } |
10476 | else |
10477 | vect_transform_slp_perm_load (vinfo, slp_node, dr_chain, gsi, vf, |
10478 | false, &n_perms); |
10479 | } |
10480 | |
10481 | if (costing_p) |
10482 | { |
10483 | if (n_adjacent_loads > 0) |
10484 | vect_get_load_cost (vinfo, stmt_info, ncopies: n_adjacent_loads, |
10485 | alignment_support_scheme, misalignment, add_realign_cost: false, |
10486 | inside_cost: &inside_cost, prologue_cost: nullptr, prologue_cost_vec: cost_vec, body_cost_vec: cost_vec, |
10487 | record_prologue_costs: true); |
10488 | if (dump_enabled_p ()) |
10489 | dump_printf_loc (MSG_NOTE, vect_location, |
10490 | "vect_model_load_cost: inside_cost = %u, " |
10491 | "prologue_cost = 0 .\n" , |
10492 | inside_cost); |
10493 | } |
10494 | |
10495 | return true; |
10496 | } |
10497 | |
10498 | if (memory_access_type == VMAT_GATHER_SCATTER |
10499 | || (!slp && memory_access_type == VMAT_CONTIGUOUS)) |
10500 | grouped_load = false; |
10501 | |
10502 | if (grouped_load |
10503 | || (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ())) |
10504 | { |
10505 | if (grouped_load) |
10506 | { |
10507 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
10508 | group_size = DR_GROUP_SIZE (first_stmt_info); |
10509 | } |
10510 | else |
10511 | { |
10512 | first_stmt_info = stmt_info; |
10513 | group_size = 1; |
10514 | } |
10515 | /* For SLP vectorization we directly vectorize a subchain |
10516 | without permutation. */ |
10517 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
10518 | first_stmt_info = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
10519 | /* For BB vectorization always use the first stmt to base |
10520 | the data ref pointer on. */ |
10521 | if (bb_vinfo) |
10522 | first_stmt_info_for_drptr |
10523 | = vect_find_first_scalar_stmt_in_slp (slp_node); |
10524 | |
10525 | /* Check if the chain of loads is already vectorized. */ |
10526 | if (STMT_VINFO_VEC_STMTS (first_stmt_info).exists () |
10527 | /* For SLP we would need to copy over SLP_TREE_VEC_DEFS. |
10528 | ??? But we can only do so if there is exactly one |
10529 | as we have no way to get at the rest. Leave the CSE |
10530 | opportunity alone. |
10531 | ??? With the group load eventually participating |
10532 | in multiple different permutations (having multiple |
10533 | slp nodes which refer to the same group) the CSE |
10534 | is even wrong code. See PR56270. */ |
10535 | && !slp) |
10536 | { |
10537 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
10538 | return true; |
10539 | } |
10540 | first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info); |
10541 | group_gap_adj = 0; |
10542 | |
10543 | /* VEC_NUM is the number of vect stmts to be created for this group. */ |
10544 | if (slp) |
10545 | { |
10546 | grouped_load = false; |
10547 | /* If an SLP permutation is from N elements to N elements, |
10548 | and if one vector holds a whole number of N, we can load |
10549 | the inputs to the permutation in the same way as an |
10550 | unpermuted sequence. In other cases we need to load the |
10551 | whole group, not only the number of vector stmts the |
10552 | permutation result fits in. */ |
10553 | unsigned scalar_lanes = SLP_TREE_LANES (slp_node); |
10554 | if (slp_perm |
10555 | && (group_size != scalar_lanes |
10556 | || !multiple_p (a: nunits, b: group_size))) |
10557 | { |
10558 | /* We don't yet generate such SLP_TREE_LOAD_PERMUTATIONs for |
10559 | variable VF; see vect_transform_slp_perm_load. */ |
10560 | unsigned int const_vf = vf.to_constant (); |
10561 | unsigned int const_nunits = nunits.to_constant (); |
10562 | vec_num = CEIL (group_size * const_vf, const_nunits); |
10563 | group_gap_adj = vf * group_size - nunits * vec_num; |
10564 | } |
10565 | else |
10566 | { |
10567 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
10568 | group_gap_adj |
10569 | = group_size - scalar_lanes; |
10570 | } |
10571 | } |
10572 | else |
10573 | vec_num = group_size; |
10574 | |
10575 | ref_type = get_group_alias_ptr_type (first_stmt_info); |
10576 | } |
10577 | else |
10578 | { |
10579 | first_stmt_info = stmt_info; |
10580 | first_dr_info = dr_info; |
10581 | group_size = vec_num = 1; |
10582 | group_gap_adj = 0; |
10583 | ref_type = reference_alias_ptr_type (DR_REF (first_dr_info->dr)); |
10584 | if (slp) |
10585 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
10586 | } |
10587 | |
10588 | gcc_assert (alignment_support_scheme); |
10589 | vec_loop_masks *loop_masks |
10590 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) |
10591 | ? &LOOP_VINFO_MASKS (loop_vinfo) |
10592 | : NULL); |
10593 | vec_loop_lens *loop_lens |
10594 | = (loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo) |
10595 | ? &LOOP_VINFO_LENS (loop_vinfo) |
10596 | : NULL); |
10597 | |
10598 | /* The vect_transform_stmt and vect_analyze_stmt will go here but there |
10599 | are some difference here. We cannot enable both the lens and masks |
10600 | during transform but it is allowed during analysis. |
10601 | Shouldn't go with length-based approach if fully masked. */ |
10602 | if (cost_vec == NULL) |
10603 | /* The cost_vec is NULL during transfrom. */ |
10604 | gcc_assert ((!loop_lens || !loop_masks)); |
10605 | |
10606 | /* Targets with store-lane instructions must not require explicit |
10607 | realignment. vect_supportable_dr_alignment always returns either |
10608 | dr_aligned or dr_unaligned_supported for masked operations. */ |
10609 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES |
10610 | && !mask |
10611 | && !loop_masks) |
10612 | || alignment_support_scheme == dr_aligned |
10613 | || alignment_support_scheme == dr_unaligned_supported); |
10614 | |
10615 | /* In case the vectorization factor (VF) is bigger than the number |
10616 | of elements that we can fit in a vectype (nunits), we have to generate |
10617 | more than one vector stmt - i.e - we need to "unroll" the |
10618 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
10619 | from one copy of the vector stmt to the next, in the field |
10620 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
10621 | stages to find the correct vector defs to be used when vectorizing |
10622 | stmts that use the defs of the current stmt. The example below |
10623 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we |
10624 | need to create 4 vectorized stmts): |
10625 | |
10626 | before vectorization: |
10627 | RELATED_STMT VEC_STMT |
10628 | S1: x = memref - - |
10629 | S2: z = x + 1 - - |
10630 | |
10631 | step 1: vectorize stmt S1: |
10632 | We first create the vector stmt VS1_0, and, as usual, record a |
10633 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. |
10634 | Next, we create the vector stmt VS1_1, and record a pointer to |
10635 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. |
10636 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
10637 | stmts and pointers: |
10638 | RELATED_STMT VEC_STMT |
10639 | VS1_0: vx0 = memref0 VS1_1 - |
10640 | VS1_1: vx1 = memref1 VS1_2 - |
10641 | VS1_2: vx2 = memref2 VS1_3 - |
10642 | VS1_3: vx3 = memref3 - - |
10643 | S1: x = load - VS1_0 |
10644 | S2: z = x + 1 - - |
10645 | */ |
10646 | |
10647 | /* In case of interleaving (non-unit grouped access): |
10648 | |
10649 | S1: x2 = &base + 2 |
10650 | S2: x0 = &base |
10651 | S3: x1 = &base + 1 |
10652 | S4: x3 = &base + 3 |
10653 | |
10654 | Vectorized loads are created in the order of memory accesses |
10655 | starting from the access of the first stmt of the chain: |
10656 | |
10657 | VS1: vx0 = &base |
10658 | VS2: vx1 = &base + vec_size*1 |
10659 | VS3: vx3 = &base + vec_size*2 |
10660 | VS4: vx4 = &base + vec_size*3 |
10661 | |
10662 | Then permutation statements are generated: |
10663 | |
10664 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
10665 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > |
10666 | ... |
10667 | |
10668 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
10669 | (the order of the data-refs in the output of vect_permute_load_chain |
10670 | corresponds to the order of scalar stmts in the interleaving chain - see |
10671 | the documentation of vect_permute_load_chain()). |
10672 | The generation of permutation stmts and recording them in |
10673 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
10674 | |
10675 | In case of both multiple types and interleaving, the vector loads and |
10676 | permutation stmts above are created for every copy. The result vector |
10677 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the |
10678 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ |
10679 | |
10680 | /* If the data reference is aligned (dr_aligned) or potentially unaligned |
10681 | on a target that supports unaligned accesses (dr_unaligned_supported) |
10682 | we generate the following code: |
10683 | p = initial_addr; |
10684 | indx = 0; |
10685 | loop { |
10686 | p = p + indx * vectype_size; |
10687 | vec_dest = *(p); |
10688 | indx = indx + 1; |
10689 | } |
10690 | |
10691 | Otherwise, the data reference is potentially unaligned on a target that |
10692 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
10693 | then generate the following code, in which the data in each iteration is |
10694 | obtained by two vector loads, one from the previous iteration, and one |
10695 | from the current iteration: |
10696 | p1 = initial_addr; |
10697 | msq_init = *(floor(p1)) |
10698 | p2 = initial_addr + VS - 1; |
10699 | realignment_token = call target_builtin; |
10700 | indx = 0; |
10701 | loop { |
10702 | p2 = p2 + indx * vectype_size |
10703 | lsq = *(floor(p2)) |
10704 | vec_dest = realign_load (msq, lsq, realignment_token) |
10705 | indx = indx + 1; |
10706 | msq = lsq; |
10707 | } */ |
10708 | |
10709 | /* If the misalignment remains the same throughout the execution of the |
10710 | loop, we can create the init_addr and permutation mask at the loop |
10711 | preheader. Otherwise, it needs to be created inside the loop. |
10712 | This can only occur when vectorizing memory accesses in the inner-loop |
10713 | nested within an outer-loop that is being vectorized. */ |
10714 | |
10715 | if (nested_in_vect_loop |
10716 | && !multiple_p (DR_STEP_ALIGNMENT (dr_info->dr), |
10717 | b: GET_MODE_SIZE (TYPE_MODE (vectype)))) |
10718 | { |
10719 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); |
10720 | compute_in_loop = true; |
10721 | } |
10722 | |
10723 | bool diff_first_stmt_info |
10724 | = first_stmt_info_for_drptr && first_stmt_info != first_stmt_info_for_drptr; |
10725 | |
10726 | tree offset = NULL_TREE; |
10727 | if ((alignment_support_scheme == dr_explicit_realign_optimized |
10728 | || alignment_support_scheme == dr_explicit_realign) |
10729 | && !compute_in_loop) |
10730 | { |
10731 | /* If we have different first_stmt_info, we can't set up realignment |
10732 | here, since we can't guarantee first_stmt_info DR has been |
10733 | initialized yet, use first_stmt_info_for_drptr DR by bumping the |
10734 | distance from first_stmt_info DR instead as below. */ |
10735 | if (!costing_p) |
10736 | { |
10737 | if (!diff_first_stmt_info) |
10738 | msq = vect_setup_realignment (vinfo, first_stmt_info, gsi, |
10739 | &realignment_token, |
10740 | alignment_support_scheme, NULL_TREE, |
10741 | &at_loop); |
10742 | if (alignment_support_scheme == dr_explicit_realign_optimized) |
10743 | { |
10744 | phi = as_a<gphi *> (SSA_NAME_DEF_STMT (msq)); |
10745 | offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
10746 | size_one_node); |
10747 | gcc_assert (!first_stmt_info_for_drptr); |
10748 | } |
10749 | } |
10750 | } |
10751 | else |
10752 | at_loop = loop; |
10753 | |
10754 | if (!known_eq (poffset, 0)) |
10755 | offset = (offset |
10756 | ? size_binop (PLUS_EXPR, offset, size_int (poffset)) |
10757 | : size_int (poffset)); |
10758 | |
10759 | tree bump; |
10760 | tree vec_offset = NULL_TREE; |
10761 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
10762 | { |
10763 | aggr_type = NULL_TREE; |
10764 | bump = NULL_TREE; |
10765 | } |
10766 | else if (memory_access_type == VMAT_GATHER_SCATTER) |
10767 | { |
10768 | aggr_type = elem_type; |
10769 | if (!costing_p) |
10770 | vect_get_strided_load_store_ops (stmt_info, loop_vinfo, gsi, gs_info: &gs_info, |
10771 | dataref_bump: &bump, vec_offset: &vec_offset, loop_lens); |
10772 | } |
10773 | else |
10774 | { |
10775 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
10776 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); |
10777 | else |
10778 | aggr_type = vectype; |
10779 | bump = vect_get_data_ptr_increment (vinfo, gsi, dr_info, aggr_type, |
10780 | memory_access_type, loop_lens); |
10781 | } |
10782 | |
10783 | auto_vec<tree> vec_offsets; |
10784 | auto_vec<tree> vec_masks; |
10785 | if (mask && !costing_p) |
10786 | { |
10787 | if (slp_node) |
10788 | vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[mask_index], |
10789 | &vec_masks); |
10790 | else |
10791 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies, op: mask, |
10792 | vec_oprnds: &vec_masks, vectype: mask_vectype); |
10793 | } |
10794 | |
10795 | tree vec_mask = NULL_TREE; |
10796 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
10797 | { |
10798 | gcc_assert (alignment_support_scheme == dr_aligned |
10799 | || alignment_support_scheme == dr_unaligned_supported); |
10800 | gcc_assert (grouped_load && !slp); |
10801 | |
10802 | unsigned int inside_cost = 0, prologue_cost = 0; |
10803 | /* For costing some adjacent vector loads, we'd like to cost with |
10804 | the total number of them once instead of cost each one by one. */ |
10805 | unsigned int n_adjacent_loads = 0; |
10806 | for (j = 0; j < ncopies; j++) |
10807 | { |
10808 | if (costing_p) |
10809 | { |
10810 | /* An IFN_LOAD_LANES will load all its vector results, |
10811 | regardless of which ones we actually need. Account |
10812 | for the cost of unused results. */ |
10813 | if (first_stmt_info == stmt_info) |
10814 | { |
10815 | unsigned int gaps = DR_GROUP_SIZE (first_stmt_info); |
10816 | stmt_vec_info next_stmt_info = first_stmt_info; |
10817 | do |
10818 | { |
10819 | gaps -= 1; |
10820 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
10821 | } |
10822 | while (next_stmt_info); |
10823 | if (gaps) |
10824 | { |
10825 | if (dump_enabled_p ()) |
10826 | dump_printf_loc (MSG_NOTE, vect_location, |
10827 | "vect_model_load_cost: %d " |
10828 | "unused vectors.\n" , |
10829 | gaps); |
10830 | vect_get_load_cost (vinfo, stmt_info, ncopies: gaps, |
10831 | alignment_support_scheme, |
10832 | misalignment, add_realign_cost: false, inside_cost: &inside_cost, |
10833 | prologue_cost: &prologue_cost, prologue_cost_vec: cost_vec, body_cost_vec: cost_vec, |
10834 | record_prologue_costs: true); |
10835 | } |
10836 | } |
10837 | n_adjacent_loads++; |
10838 | continue; |
10839 | } |
10840 | |
10841 | /* 1. Create the vector or array pointer update chain. */ |
10842 | if (j == 0) |
10843 | dataref_ptr |
10844 | = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type, |
10845 | at_loop, offset, &dummy, gsi, |
10846 | &ptr_incr, false, bump); |
10847 | else |
10848 | { |
10849 | gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)); |
10850 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi, |
10851 | stmt_info, bump); |
10852 | } |
10853 | if (mask) |
10854 | vec_mask = vec_masks[j]; |
10855 | |
10856 | tree vec_array = create_vector_array (elem_type: vectype, nelems: vec_num); |
10857 | |
10858 | tree final_mask = NULL_TREE; |
10859 | tree final_len = NULL_TREE; |
10860 | tree bias = NULL_TREE; |
10861 | if (loop_masks) |
10862 | final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks, |
10863 | ncopies, vectype, j); |
10864 | if (vec_mask) |
10865 | final_mask = prepare_vec_mask (loop_vinfo, mask_type: mask_vectype, loop_mask: final_mask, |
10866 | vec_mask, gsi); |
10867 | |
10868 | if (lanes_ifn == IFN_MASK_LEN_LOAD_LANES) |
10869 | { |
10870 | if (loop_lens) |
10871 | final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens, |
10872 | ncopies, vectype, j, 1); |
10873 | else |
10874 | final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
10875 | signed char biasval |
10876 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
10877 | bias = build_int_cst (intQI_type_node, biasval); |
10878 | if (!final_mask) |
10879 | { |
10880 | mask_vectype = truth_type_for (vectype); |
10881 | final_mask = build_minus_one_cst (mask_vectype); |
10882 | } |
10883 | } |
10884 | |
10885 | gcall *call; |
10886 | if (final_len && final_mask) |
10887 | { |
10888 | /* Emit: |
10889 | VEC_ARRAY = MASK_LEN_LOAD_LANES (DATAREF_PTR, ALIAS_PTR, |
10890 | VEC_MASK, LEN, BIAS). */ |
10891 | unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype)); |
10892 | tree alias_ptr = build_int_cst (ref_type, align); |
10893 | call = gimple_build_call_internal (IFN_MASK_LEN_LOAD_LANES, 5, |
10894 | dataref_ptr, alias_ptr, |
10895 | final_mask, final_len, bias); |
10896 | } |
10897 | else if (final_mask) |
10898 | { |
10899 | /* Emit: |
10900 | VEC_ARRAY = MASK_LOAD_LANES (DATAREF_PTR, ALIAS_PTR, |
10901 | VEC_MASK). */ |
10902 | unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype)); |
10903 | tree alias_ptr = build_int_cst (ref_type, align); |
10904 | call = gimple_build_call_internal (IFN_MASK_LOAD_LANES, 3, |
10905 | dataref_ptr, alias_ptr, |
10906 | final_mask); |
10907 | } |
10908 | else |
10909 | { |
10910 | /* Emit: |
10911 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ |
10912 | data_ref = create_array_ref (type: aggr_type, ptr: dataref_ptr, alias_ptr_type: ref_type); |
10913 | call = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); |
10914 | } |
10915 | gimple_call_set_lhs (gs: call, lhs: vec_array); |
10916 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
10917 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
10918 | |
10919 | dr_chain.create (nelems: vec_num); |
10920 | /* Extract each vector into an SSA_NAME. */ |
10921 | for (i = 0; i < vec_num; i++) |
10922 | { |
10923 | new_temp = read_vector_array (vinfo, stmt_info, gsi, scalar_dest, |
10924 | array: vec_array, n: i); |
10925 | dr_chain.quick_push (obj: new_temp); |
10926 | } |
10927 | |
10928 | /* Record the mapping between SSA_NAMEs and statements. */ |
10929 | vect_record_grouped_load_vectors (vinfo, stmt_info, dr_chain); |
10930 | |
10931 | /* Record that VEC_ARRAY is now dead. */ |
10932 | vect_clobber_variable (vinfo, stmt_info, gsi, var: vec_array); |
10933 | |
10934 | dr_chain.release (); |
10935 | |
10936 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
10937 | } |
10938 | |
10939 | if (costing_p) |
10940 | { |
10941 | if (n_adjacent_loads > 0) |
10942 | vect_get_load_cost (vinfo, stmt_info, ncopies: n_adjacent_loads, |
10943 | alignment_support_scheme, misalignment, add_realign_cost: false, |
10944 | inside_cost: &inside_cost, prologue_cost: &prologue_cost, prologue_cost_vec: cost_vec, |
10945 | body_cost_vec: cost_vec, record_prologue_costs: true); |
10946 | if (dump_enabled_p ()) |
10947 | dump_printf_loc (MSG_NOTE, vect_location, |
10948 | "vect_model_load_cost: inside_cost = %u, " |
10949 | "prologue_cost = %u .\n" , |
10950 | inside_cost, prologue_cost); |
10951 | } |
10952 | |
10953 | return true; |
10954 | } |
10955 | |
10956 | if (memory_access_type == VMAT_GATHER_SCATTER) |
10957 | { |
10958 | gcc_assert (alignment_support_scheme == dr_aligned |
10959 | || alignment_support_scheme == dr_unaligned_supported); |
10960 | gcc_assert (!grouped_load && !slp_perm); |
10961 | |
10962 | unsigned int inside_cost = 0, prologue_cost = 0; |
10963 | for (j = 0; j < ncopies; j++) |
10964 | { |
10965 | /* 1. Create the vector or array pointer update chain. */ |
10966 | if (j == 0 && !costing_p) |
10967 | { |
10968 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
10969 | vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info, |
10970 | slp_node, gs_info: &gs_info, dataref_ptr: &dataref_ptr, |
10971 | vec_offset: &vec_offsets); |
10972 | else |
10973 | dataref_ptr |
10974 | = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type, |
10975 | at_loop, offset, &dummy, gsi, |
10976 | &ptr_incr, false, bump); |
10977 | } |
10978 | else if (!costing_p) |
10979 | { |
10980 | gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)); |
10981 | if (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
10982 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, |
10983 | gsi, stmt_info, bump); |
10984 | } |
10985 | |
10986 | gimple *new_stmt = NULL; |
10987 | for (i = 0; i < vec_num; i++) |
10988 | { |
10989 | tree final_mask = NULL_TREE; |
10990 | tree final_len = NULL_TREE; |
10991 | tree bias = NULL_TREE; |
10992 | if (!costing_p) |
10993 | { |
10994 | if (mask) |
10995 | vec_mask = vec_masks[vec_num * j + i]; |
10996 | if (loop_masks) |
10997 | final_mask |
10998 | = vect_get_loop_mask (loop_vinfo, gsi, loop_masks, |
10999 | vec_num * ncopies, vectype, |
11000 | vec_num * j + i); |
11001 | if (vec_mask) |
11002 | final_mask = prepare_vec_mask (loop_vinfo, mask_type: mask_vectype, |
11003 | loop_mask: final_mask, vec_mask, gsi); |
11004 | |
11005 | if (i > 0 && !STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
11006 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, |
11007 | gsi, stmt_info, bump); |
11008 | } |
11009 | |
11010 | /* 2. Create the vector-load in the loop. */ |
11011 | unsigned HOST_WIDE_INT align; |
11012 | if (gs_info.ifn != IFN_LAST) |
11013 | { |
11014 | if (costing_p) |
11015 | { |
11016 | unsigned int cnunits = vect_nunits_for_cost (vec_type: vectype); |
11017 | inside_cost |
11018 | = record_stmt_cost (body_cost_vec: cost_vec, count: cnunits, kind: scalar_load, |
11019 | stmt_info, misalign: 0, where: vect_body); |
11020 | continue; |
11021 | } |
11022 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
11023 | vec_offset = vec_offsets[vec_num * j + i]; |
11024 | tree zero = build_zero_cst (vectype); |
11025 | tree scale = size_int (gs_info.scale); |
11026 | |
11027 | if (gs_info.ifn == IFN_MASK_LEN_GATHER_LOAD) |
11028 | { |
11029 | if (loop_lens) |
11030 | final_len |
11031 | = vect_get_loop_len (loop_vinfo, gsi, loop_lens, |
11032 | vec_num * ncopies, vectype, |
11033 | vec_num * j + i, 1); |
11034 | else |
11035 | final_len |
11036 | = build_int_cst (sizetype, |
11037 | TYPE_VECTOR_SUBPARTS (node: vectype)); |
11038 | signed char biasval |
11039 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
11040 | bias = build_int_cst (intQI_type_node, biasval); |
11041 | if (!final_mask) |
11042 | { |
11043 | mask_vectype = truth_type_for (vectype); |
11044 | final_mask = build_minus_one_cst (mask_vectype); |
11045 | } |
11046 | } |
11047 | |
11048 | gcall *call; |
11049 | if (final_len && final_mask) |
11050 | call |
11051 | = gimple_build_call_internal (IFN_MASK_LEN_GATHER_LOAD, 7, |
11052 | dataref_ptr, vec_offset, |
11053 | scale, zero, final_mask, |
11054 | final_len, bias); |
11055 | else if (final_mask) |
11056 | call = gimple_build_call_internal (IFN_MASK_GATHER_LOAD, 5, |
11057 | dataref_ptr, vec_offset, |
11058 | scale, zero, final_mask); |
11059 | else |
11060 | call = gimple_build_call_internal (IFN_GATHER_LOAD, 4, |
11061 | dataref_ptr, vec_offset, |
11062 | scale, zero); |
11063 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
11064 | new_stmt = call; |
11065 | data_ref = NULL_TREE; |
11066 | } |
11067 | else if (gs_info.decl) |
11068 | { |
11069 | /* The builtin decls path for gather is legacy, x86 only. */ |
11070 | gcc_assert (!final_len && nunits.is_constant ()); |
11071 | if (costing_p) |
11072 | { |
11073 | unsigned int cnunits = vect_nunits_for_cost (vec_type: vectype); |
11074 | inside_cost |
11075 | = record_stmt_cost (body_cost_vec: cost_vec, count: cnunits, kind: scalar_load, |
11076 | stmt_info, misalign: 0, where: vect_body); |
11077 | continue; |
11078 | } |
11079 | poly_uint64 offset_nunits |
11080 | = TYPE_VECTOR_SUBPARTS (node: gs_info.offset_vectype); |
11081 | if (known_eq (nunits, offset_nunits)) |
11082 | { |
11083 | new_stmt = vect_build_one_gather_load_call |
11084 | (vinfo, stmt_info, gsi, gs_info: &gs_info, |
11085 | ptr: dataref_ptr, offset: vec_offsets[vec_num * j + i], |
11086 | mask: final_mask); |
11087 | data_ref = NULL_TREE; |
11088 | } |
11089 | else if (known_eq (nunits, offset_nunits * 2)) |
11090 | { |
11091 | /* We have a offset vector with half the number of |
11092 | lanes but the builtins will produce full vectype |
11093 | data with just the lower lanes filled. */ |
11094 | new_stmt = vect_build_one_gather_load_call |
11095 | (vinfo, stmt_info, gsi, gs_info: &gs_info, |
11096 | ptr: dataref_ptr, offset: vec_offsets[2 * vec_num * j + 2 * i], |
11097 | mask: final_mask); |
11098 | tree low = make_ssa_name (var: vectype); |
11099 | gimple_set_lhs (new_stmt, low); |
11100 | vect_finish_stmt_generation (vinfo, stmt_info, |
11101 | vec_stmt: new_stmt, gsi); |
11102 | |
11103 | /* now put upper half of final_mask in final_mask low. */ |
11104 | if (final_mask |
11105 | && !SCALAR_INT_MODE_P |
11106 | (TYPE_MODE (TREE_TYPE (final_mask)))) |
11107 | { |
11108 | int count = nunits.to_constant (); |
11109 | vec_perm_builder sel (count, count, 1); |
11110 | sel.quick_grow (len: count); |
11111 | for (int i = 0; i < count; ++i) |
11112 | sel[i] = i | (count / 2); |
11113 | vec_perm_indices indices (sel, 2, count); |
11114 | tree perm_mask = vect_gen_perm_mask_checked |
11115 | (TREE_TYPE (final_mask), sel: indices); |
11116 | new_stmt = gimple_build_assign (NULL_TREE, |
11117 | VEC_PERM_EXPR, |
11118 | final_mask, |
11119 | final_mask, |
11120 | perm_mask); |
11121 | final_mask = make_ssa_name (TREE_TYPE (final_mask)); |
11122 | gimple_set_lhs (new_stmt, final_mask); |
11123 | vect_finish_stmt_generation (vinfo, stmt_info, |
11124 | vec_stmt: new_stmt, gsi); |
11125 | } |
11126 | else if (final_mask) |
11127 | { |
11128 | new_stmt = gimple_build_assign (NULL_TREE, |
11129 | VEC_UNPACK_HI_EXPR, |
11130 | final_mask); |
11131 | final_mask = make_ssa_name |
11132 | (var: truth_type_for (gs_info.offset_vectype)); |
11133 | gimple_set_lhs (new_stmt, final_mask); |
11134 | vect_finish_stmt_generation (vinfo, stmt_info, |
11135 | vec_stmt: new_stmt, gsi); |
11136 | } |
11137 | |
11138 | new_stmt = vect_build_one_gather_load_call |
11139 | (vinfo, stmt_info, gsi, gs_info: &gs_info, |
11140 | ptr: dataref_ptr, |
11141 | offset: vec_offsets[2 * vec_num * j + 2 * i + 1], |
11142 | mask: final_mask); |
11143 | tree high = make_ssa_name (var: vectype); |
11144 | gimple_set_lhs (new_stmt, high); |
11145 | vect_finish_stmt_generation (vinfo, stmt_info, |
11146 | vec_stmt: new_stmt, gsi); |
11147 | |
11148 | /* compose low + high. */ |
11149 | int count = nunits.to_constant (); |
11150 | vec_perm_builder sel (count, count, 1); |
11151 | sel.quick_grow (len: count); |
11152 | for (int i = 0; i < count; ++i) |
11153 | sel[i] = i < count / 2 ? i : i + count / 2; |
11154 | vec_perm_indices indices (sel, 2, count); |
11155 | tree perm_mask |
11156 | = vect_gen_perm_mask_checked (vectype, sel: indices); |
11157 | new_stmt = gimple_build_assign (NULL_TREE, |
11158 | VEC_PERM_EXPR, |
11159 | low, high, perm_mask); |
11160 | data_ref = NULL_TREE; |
11161 | } |
11162 | else if (known_eq (nunits * 2, offset_nunits)) |
11163 | { |
11164 | /* We have a offset vector with double the number of |
11165 | lanes. Select the low/high part accordingly. */ |
11166 | vec_offset = vec_offsets[(vec_num * j + i) / 2]; |
11167 | if ((vec_num * j + i) & 1) |
11168 | { |
11169 | int count = offset_nunits.to_constant (); |
11170 | vec_perm_builder sel (count, count, 1); |
11171 | sel.quick_grow (len: count); |
11172 | for (int i = 0; i < count; ++i) |
11173 | sel[i] = i | (count / 2); |
11174 | vec_perm_indices indices (sel, 2, count); |
11175 | tree perm_mask = vect_gen_perm_mask_checked |
11176 | (TREE_TYPE (vec_offset), sel: indices); |
11177 | new_stmt = gimple_build_assign (NULL_TREE, |
11178 | VEC_PERM_EXPR, |
11179 | vec_offset, |
11180 | vec_offset, |
11181 | perm_mask); |
11182 | vec_offset = make_ssa_name (TREE_TYPE (vec_offset)); |
11183 | gimple_set_lhs (new_stmt, vec_offset); |
11184 | vect_finish_stmt_generation (vinfo, stmt_info, |
11185 | vec_stmt: new_stmt, gsi); |
11186 | } |
11187 | new_stmt = vect_build_one_gather_load_call |
11188 | (vinfo, stmt_info, gsi, gs_info: &gs_info, |
11189 | ptr: dataref_ptr, offset: vec_offset, mask: final_mask); |
11190 | data_ref = NULL_TREE; |
11191 | } |
11192 | else |
11193 | gcc_unreachable (); |
11194 | } |
11195 | else |
11196 | { |
11197 | /* Emulated gather-scatter. */ |
11198 | gcc_assert (!final_mask); |
11199 | unsigned HOST_WIDE_INT const_nunits = nunits.to_constant (); |
11200 | if (costing_p) |
11201 | { |
11202 | /* For emulated gathers N offset vector element |
11203 | offset add is consumed by the load). */ |
11204 | inside_cost = record_stmt_cost (body_cost_vec: cost_vec, count: const_nunits, |
11205 | kind: vec_to_scalar, stmt_info, |
11206 | misalign: 0, where: vect_body); |
11207 | /* N scalar loads plus gathering them into a |
11208 | vector. */ |
11209 | inside_cost |
11210 | = record_stmt_cost (body_cost_vec: cost_vec, count: const_nunits, kind: scalar_load, |
11211 | stmt_info, misalign: 0, where: vect_body); |
11212 | inside_cost |
11213 | = record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: vec_construct, |
11214 | stmt_info, misalign: 0, where: vect_body); |
11215 | continue; |
11216 | } |
11217 | unsigned HOST_WIDE_INT const_offset_nunits |
11218 | = TYPE_VECTOR_SUBPARTS (node: gs_info.offset_vectype) |
11219 | .to_constant (); |
11220 | vec<constructor_elt, va_gc> *ctor_elts; |
11221 | vec_alloc (v&: ctor_elts, nelems: const_nunits); |
11222 | gimple_seq stmts = NULL; |
11223 | /* We support offset vectors with more elements |
11224 | than the data vector for now. */ |
11225 | unsigned HOST_WIDE_INT factor |
11226 | = const_offset_nunits / const_nunits; |
11227 | vec_offset = vec_offsets[(vec_num * j + i) / factor]; |
11228 | unsigned elt_offset |
11229 | = ((vec_num * j + i) % factor) * const_nunits; |
11230 | tree idx_type = TREE_TYPE (TREE_TYPE (vec_offset)); |
11231 | tree scale = size_int (gs_info.scale); |
11232 | align = get_object_alignment (DR_REF (first_dr_info->dr)); |
11233 | tree ltype = build_aligned_type (TREE_TYPE (vectype), align); |
11234 | for (unsigned k = 0; k < const_nunits; ++k) |
11235 | { |
11236 | tree boff = size_binop (MULT_EXPR, TYPE_SIZE (idx_type), |
11237 | bitsize_int (k + elt_offset)); |
11238 | tree idx |
11239 | = gimple_build (seq: &stmts, code: BIT_FIELD_REF, type: idx_type, |
11240 | ops: vec_offset, TYPE_SIZE (idx_type), ops: boff); |
11241 | idx = gimple_convert (seq: &stmts, sizetype, op: idx); |
11242 | idx = gimple_build (seq: &stmts, code: MULT_EXPR, sizetype, ops: idx, |
11243 | ops: scale); |
11244 | tree ptr = gimple_build (seq: &stmts, code: PLUS_EXPR, |
11245 | TREE_TYPE (dataref_ptr), |
11246 | ops: dataref_ptr, ops: idx); |
11247 | ptr = gimple_convert (seq: &stmts, ptr_type_node, op: ptr); |
11248 | tree elt = make_ssa_name (TREE_TYPE (vectype)); |
11249 | tree ref = build2 (MEM_REF, ltype, ptr, |
11250 | build_int_cst (ref_type, 0)); |
11251 | new_stmt = gimple_build_assign (elt, ref); |
11252 | gimple_set_vuse (g: new_stmt, vuse: gimple_vuse (g: gsi_stmt (i: *gsi))); |
11253 | gimple_seq_add_stmt (&stmts, new_stmt); |
11254 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, elt); |
11255 | } |
11256 | gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT); |
11257 | new_stmt = gimple_build_assign ( |
11258 | NULL_TREE, build_constructor (vectype, ctor_elts)); |
11259 | data_ref = NULL_TREE; |
11260 | } |
11261 | |
11262 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
11263 | /* DATA_REF is null if we've already built the statement. */ |
11264 | if (data_ref) |
11265 | { |
11266 | vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr)); |
11267 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
11268 | } |
11269 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
11270 | gimple_set_lhs (new_stmt, new_temp); |
11271 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
11272 | |
11273 | /* Store vector loads in the corresponding SLP_NODE. */ |
11274 | if (slp) |
11275 | slp_node->push_vec_def (def: new_stmt); |
11276 | } |
11277 | |
11278 | if (!slp && !costing_p) |
11279 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
11280 | } |
11281 | |
11282 | if (!slp && !costing_p) |
11283 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
11284 | |
11285 | if (costing_p && dump_enabled_p ()) |
11286 | dump_printf_loc (MSG_NOTE, vect_location, |
11287 | "vect_model_load_cost: inside_cost = %u, " |
11288 | "prologue_cost = %u .\n" , |
11289 | inside_cost, prologue_cost); |
11290 | return true; |
11291 | } |
11292 | |
11293 | poly_uint64 group_elt = 0; |
11294 | unsigned int inside_cost = 0, prologue_cost = 0; |
11295 | /* For costing some adjacent vector loads, we'd like to cost with |
11296 | the total number of them once instead of cost each one by one. */ |
11297 | unsigned int n_adjacent_loads = 0; |
11298 | for (j = 0; j < ncopies; j++) |
11299 | { |
11300 | /* 1. Create the vector or array pointer update chain. */ |
11301 | if (j == 0 && !costing_p) |
11302 | { |
11303 | bool simd_lane_access_p |
11304 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) != 0; |
11305 | if (simd_lane_access_p |
11306 | && TREE_CODE (DR_BASE_ADDRESS (first_dr_info->dr)) == ADDR_EXPR |
11307 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr_info->dr), 0)) |
11308 | && integer_zerop (get_dr_vinfo_offset (vinfo, dr_info: first_dr_info)) |
11309 | && integer_zerop (DR_INIT (first_dr_info->dr)) |
11310 | && alias_sets_conflict_p (get_alias_set (aggr_type), |
11311 | get_alias_set (TREE_TYPE (ref_type))) |
11312 | && (alignment_support_scheme == dr_aligned |
11313 | || alignment_support_scheme == dr_unaligned_supported)) |
11314 | { |
11315 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr_info->dr)); |
11316 | dataref_offset = build_int_cst (ref_type, 0); |
11317 | } |
11318 | else if (diff_first_stmt_info) |
11319 | { |
11320 | dataref_ptr |
11321 | = vect_create_data_ref_ptr (vinfo, first_stmt_info_for_drptr, |
11322 | aggr_type, at_loop, offset, &dummy, |
11323 | gsi, &ptr_incr, simd_lane_access_p, |
11324 | bump); |
11325 | /* Adjust the pointer by the difference to first_stmt. */ |
11326 | data_reference_p ptrdr |
11327 | = STMT_VINFO_DATA_REF (first_stmt_info_for_drptr); |
11328 | tree diff |
11329 | = fold_convert (sizetype, |
11330 | size_binop (MINUS_EXPR, |
11331 | DR_INIT (first_dr_info->dr), |
11332 | DR_INIT (ptrdr))); |
11333 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi, |
11334 | stmt_info, diff); |
11335 | if (alignment_support_scheme == dr_explicit_realign) |
11336 | { |
11337 | msq = vect_setup_realignment (vinfo, |
11338 | first_stmt_info_for_drptr, gsi, |
11339 | &realignment_token, |
11340 | alignment_support_scheme, |
11341 | dataref_ptr, &at_loop); |
11342 | gcc_assert (!compute_in_loop); |
11343 | } |
11344 | } |
11345 | else |
11346 | dataref_ptr |
11347 | = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type, |
11348 | at_loop, |
11349 | offset, &dummy, gsi, &ptr_incr, |
11350 | simd_lane_access_p, bump); |
11351 | } |
11352 | else if (!costing_p) |
11353 | { |
11354 | gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)); |
11355 | if (dataref_offset) |
11356 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, |
11357 | bump); |
11358 | else |
11359 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi, |
11360 | stmt_info, bump); |
11361 | } |
11362 | |
11363 | if (grouped_load || slp_perm) |
11364 | dr_chain.create (nelems: vec_num); |
11365 | |
11366 | gimple *new_stmt = NULL; |
11367 | for (i = 0; i < vec_num; i++) |
11368 | { |
11369 | tree final_mask = NULL_TREE; |
11370 | tree final_len = NULL_TREE; |
11371 | tree bias = NULL_TREE; |
11372 | if (!costing_p) |
11373 | { |
11374 | if (mask) |
11375 | vec_mask = vec_masks[vec_num * j + i]; |
11376 | if (loop_masks) |
11377 | final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks, |
11378 | vec_num * ncopies, vectype, |
11379 | vec_num * j + i); |
11380 | if (vec_mask) |
11381 | final_mask = prepare_vec_mask (loop_vinfo, mask_type: mask_vectype, |
11382 | loop_mask: final_mask, vec_mask, gsi); |
11383 | |
11384 | if (i > 0) |
11385 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, |
11386 | gsi, stmt_info, bump); |
11387 | } |
11388 | |
11389 | /* 2. Create the vector-load in the loop. */ |
11390 | switch (alignment_support_scheme) |
11391 | { |
11392 | case dr_aligned: |
11393 | case dr_unaligned_supported: |
11394 | { |
11395 | if (costing_p) |
11396 | break; |
11397 | |
11398 | unsigned int misalign; |
11399 | unsigned HOST_WIDE_INT align; |
11400 | align = known_alignment (DR_TARGET_ALIGNMENT (first_dr_info)); |
11401 | if (alignment_support_scheme == dr_aligned) |
11402 | misalign = 0; |
11403 | else if (misalignment == DR_MISALIGNMENT_UNKNOWN) |
11404 | { |
11405 | align |
11406 | = dr_alignment (vect_dr_behavior (vinfo, dr_info: first_dr_info)); |
11407 | misalign = 0; |
11408 | } |
11409 | else |
11410 | misalign = misalignment; |
11411 | if (dataref_offset == NULL_TREE |
11412 | && TREE_CODE (dataref_ptr) == SSA_NAME) |
11413 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
11414 | misalign); |
11415 | align = least_bit_hwi (x: misalign | align); |
11416 | |
11417 | /* Compute IFN when LOOP_LENS or final_mask valid. */ |
11418 | machine_mode vmode = TYPE_MODE (vectype); |
11419 | machine_mode new_vmode = vmode; |
11420 | internal_fn partial_ifn = IFN_LAST; |
11421 | if (loop_lens) |
11422 | { |
11423 | opt_machine_mode new_ovmode |
11424 | = get_len_load_store_mode (vmode, true, &partial_ifn); |
11425 | new_vmode = new_ovmode.require (); |
11426 | unsigned factor |
11427 | = (new_ovmode == vmode) ? 1 : GET_MODE_UNIT_SIZE (vmode); |
11428 | final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens, |
11429 | vec_num * ncopies, vectype, |
11430 | vec_num * j + i, factor); |
11431 | } |
11432 | else if (final_mask) |
11433 | { |
11434 | if (!can_vec_mask_load_store_p ( |
11435 | vmode, TYPE_MODE (TREE_TYPE (final_mask)), true, |
11436 | &partial_ifn)) |
11437 | gcc_unreachable (); |
11438 | } |
11439 | |
11440 | if (partial_ifn == IFN_MASK_LEN_LOAD) |
11441 | { |
11442 | if (!final_len) |
11443 | { |
11444 | /* Pass VF value to 'len' argument of |
11445 | MASK_LEN_LOAD if LOOP_LENS is invalid. */ |
11446 | final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
11447 | } |
11448 | if (!final_mask) |
11449 | { |
11450 | /* Pass all ones value to 'mask' argument of |
11451 | MASK_LEN_LOAD if final_mask is invalid. */ |
11452 | mask_vectype = truth_type_for (vectype); |
11453 | final_mask = build_minus_one_cst (mask_vectype); |
11454 | } |
11455 | } |
11456 | if (final_len) |
11457 | { |
11458 | signed char biasval |
11459 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
11460 | |
11461 | bias = build_int_cst (intQI_type_node, biasval); |
11462 | } |
11463 | |
11464 | if (final_len) |
11465 | { |
11466 | tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT); |
11467 | gcall *call; |
11468 | if (partial_ifn == IFN_MASK_LEN_LOAD) |
11469 | call = gimple_build_call_internal (IFN_MASK_LEN_LOAD, 5, |
11470 | dataref_ptr, ptr, |
11471 | final_mask, final_len, |
11472 | bias); |
11473 | else |
11474 | call = gimple_build_call_internal (IFN_LEN_LOAD, 4, |
11475 | dataref_ptr, ptr, |
11476 | final_len, bias); |
11477 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
11478 | new_stmt = call; |
11479 | data_ref = NULL_TREE; |
11480 | |
11481 | /* Need conversion if it's wrapped with VnQI. */ |
11482 | if (vmode != new_vmode) |
11483 | { |
11484 | tree new_vtype = build_vector_type_for_mode ( |
11485 | unsigned_intQI_type_node, new_vmode); |
11486 | tree var |
11487 | = vect_get_new_ssa_name (new_vtype, vect_simple_var); |
11488 | gimple_set_lhs (call, var); |
11489 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, |
11490 | gsi); |
11491 | tree op = build1 (VIEW_CONVERT_EXPR, vectype, var); |
11492 | new_stmt = gimple_build_assign (vec_dest, |
11493 | VIEW_CONVERT_EXPR, op); |
11494 | } |
11495 | } |
11496 | else if (final_mask) |
11497 | { |
11498 | tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT); |
11499 | gcall *call = gimple_build_call_internal (IFN_MASK_LOAD, 3, |
11500 | dataref_ptr, ptr, |
11501 | final_mask); |
11502 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
11503 | new_stmt = call; |
11504 | data_ref = NULL_TREE; |
11505 | } |
11506 | else |
11507 | { |
11508 | tree ltype = vectype; |
11509 | tree new_vtype = NULL_TREE; |
11510 | unsigned HOST_WIDE_INT gap = DR_GROUP_GAP (first_stmt_info); |
11511 | unsigned int vect_align |
11512 | = vect_known_alignment_in_bytes (dr_info: first_dr_info, vectype); |
11513 | unsigned int scalar_dr_size |
11514 | = vect_get_scalar_dr_size (dr_info: first_dr_info); |
11515 | /* If there's no peeling for gaps but we have a gap |
11516 | with slp loads then load the lower half of the |
11517 | vector only. See get_group_load_store_type for |
11518 | when we apply this optimization. */ |
11519 | if (slp |
11520 | && loop_vinfo |
11521 | && !LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) && gap != 0 |
11522 | && known_eq (nunits, (group_size - gap) * 2) |
11523 | && known_eq (nunits, group_size) |
11524 | && gap >= (vect_align / scalar_dr_size)) |
11525 | { |
11526 | tree half_vtype; |
11527 | new_vtype |
11528 | = vector_vector_composition_type (vtype: vectype, nelts: 2, |
11529 | ptype: &half_vtype); |
11530 | if (new_vtype != NULL_TREE) |
11531 | ltype = half_vtype; |
11532 | } |
11533 | /* Try to use a single smaller load when we are about |
11534 | to load excess elements compared to the unrolled |
11535 | scalar loop. |
11536 | ??? This should cover the above case as well. */ |
11537 | else if (known_gt ((vec_num * j + i + 1) * nunits, |
11538 | (group_size * vf - gap))) |
11539 | { |
11540 | if (known_ge ((vec_num * j + i + 1) * nunits |
11541 | - (group_size * vf - gap), nunits)) |
11542 | /* DR will be unused. */ |
11543 | ltype = NULL_TREE; |
11544 | else if (known_ge (vect_align, |
11545 | tree_to_poly_uint64 |
11546 | (TYPE_SIZE_UNIT (vectype)))) |
11547 | /* Aligned access to excess elements is OK if |
11548 | at least one element is accessed in the |
11549 | scalar loop. */ |
11550 | ; |
11551 | else |
11552 | { |
11553 | auto remain |
11554 | = ((group_size * vf - gap) |
11555 | - (vec_num * j + i) * nunits); |
11556 | /* remain should now be > 0 and < nunits. */ |
11557 | unsigned num; |
11558 | if (constant_multiple_p (a: nunits, b: remain, multiple: &num)) |
11559 | { |
11560 | tree ptype; |
11561 | new_vtype |
11562 | = vector_vector_composition_type (vtype: vectype, |
11563 | nelts: num, |
11564 | ptype: &ptype); |
11565 | if (new_vtype) |
11566 | ltype = ptype; |
11567 | } |
11568 | /* Else use multiple loads or a masked load? */ |
11569 | } |
11570 | } |
11571 | tree offset |
11572 | = (dataref_offset ? dataref_offset |
11573 | : build_int_cst (ref_type, 0)); |
11574 | if (!ltype) |
11575 | ; |
11576 | else if (ltype != vectype |
11577 | && memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
11578 | { |
11579 | poly_uint64 gap_offset |
11580 | = (tree_to_poly_uint64 (TYPE_SIZE_UNIT (vectype)) |
11581 | - tree_to_poly_uint64 (TYPE_SIZE_UNIT (ltype))); |
11582 | tree gapcst = build_int_cstu (type: ref_type, gap_offset); |
11583 | offset = size_binop (PLUS_EXPR, offset, gapcst); |
11584 | } |
11585 | if (ltype) |
11586 | { |
11587 | data_ref |
11588 | = fold_build2 (MEM_REF, ltype, dataref_ptr, offset); |
11589 | if (alignment_support_scheme == dr_aligned) |
11590 | ; |
11591 | else |
11592 | TREE_TYPE (data_ref) |
11593 | = build_aligned_type (TREE_TYPE (data_ref), |
11594 | align * BITS_PER_UNIT); |
11595 | } |
11596 | if (!ltype) |
11597 | data_ref = build_constructor (vectype, NULL); |
11598 | else if (ltype != vectype) |
11599 | { |
11600 | vect_copy_ref_info (data_ref, |
11601 | DR_REF (first_dr_info->dr)); |
11602 | tree tem = make_ssa_name (var: ltype); |
11603 | new_stmt = gimple_build_assign (tem, data_ref); |
11604 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, |
11605 | gsi); |
11606 | data_ref = NULL; |
11607 | vec<constructor_elt, va_gc> *v; |
11608 | /* We've computed 'num' above to statically two |
11609 | or via constant_multiple_p. */ |
11610 | unsigned num |
11611 | = (exact_div (a: tree_to_poly_uint64 |
11612 | (TYPE_SIZE_UNIT (vectype)), |
11613 | b: tree_to_poly_uint64 |
11614 | (TYPE_SIZE_UNIT (ltype))) |
11615 | .to_constant ()); |
11616 | vec_alloc (v, nelems: num); |
11617 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
11618 | { |
11619 | while (--num) |
11620 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, |
11621 | build_zero_cst (ltype)); |
11622 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, tem); |
11623 | } |
11624 | else |
11625 | { |
11626 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, tem); |
11627 | while (--num) |
11628 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, |
11629 | build_zero_cst (ltype)); |
11630 | } |
11631 | gcc_assert (new_vtype != NULL_TREE); |
11632 | if (new_vtype == vectype) |
11633 | new_stmt = gimple_build_assign ( |
11634 | vec_dest, build_constructor (vectype, v)); |
11635 | else |
11636 | { |
11637 | tree new_vname = make_ssa_name (var: new_vtype); |
11638 | new_stmt = gimple_build_assign ( |
11639 | new_vname, build_constructor (new_vtype, v)); |
11640 | vect_finish_stmt_generation (vinfo, stmt_info, |
11641 | vec_stmt: new_stmt, gsi); |
11642 | new_stmt = gimple_build_assign ( |
11643 | vec_dest, |
11644 | build1 (VIEW_CONVERT_EXPR, vectype, new_vname)); |
11645 | } |
11646 | } |
11647 | } |
11648 | break; |
11649 | } |
11650 | case dr_explicit_realign: |
11651 | { |
11652 | if (costing_p) |
11653 | break; |
11654 | tree ptr, bump; |
11655 | |
11656 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
11657 | |
11658 | if (compute_in_loop) |
11659 | msq = vect_setup_realignment (vinfo, first_stmt_info, gsi, |
11660 | &realignment_token, |
11661 | dr_explicit_realign, |
11662 | dataref_ptr, NULL); |
11663 | |
11664 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
11665 | ptr = copy_ssa_name (var: dataref_ptr); |
11666 | else |
11667 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); |
11668 | // For explicit realign the target alignment should be |
11669 | // known at compile time. |
11670 | unsigned HOST_WIDE_INT align |
11671 | = DR_TARGET_ALIGNMENT (first_dr_info).to_constant (); |
11672 | new_stmt = gimple_build_assign ( |
11673 | ptr, BIT_AND_EXPR, dataref_ptr, |
11674 | build_int_cst (TREE_TYPE (dataref_ptr), |
11675 | -(HOST_WIDE_INT) align)); |
11676 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
11677 | data_ref |
11678 | = build2 (MEM_REF, vectype, ptr, build_int_cst (ref_type, 0)); |
11679 | vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr)); |
11680 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
11681 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
11682 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
11683 | gimple_assign_set_lhs (gs: new_stmt, lhs: new_temp); |
11684 | gimple_move_vops (new_stmt, stmt_info->stmt); |
11685 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
11686 | msq = new_temp; |
11687 | |
11688 | bump = size_binop (MULT_EXPR, vs, TYPE_SIZE_UNIT (elem_type)); |
11689 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
11690 | ptr = bump_vector_ptr (vinfo, dataref_ptr, NULL, gsi, stmt_info, |
11691 | bump); |
11692 | new_stmt = gimple_build_assign ( |
11693 | NULL_TREE, BIT_AND_EXPR, ptr, |
11694 | build_int_cst (TREE_TYPE (ptr), -(HOST_WIDE_INT) align)); |
11695 | if (TREE_CODE (ptr) == SSA_NAME) |
11696 | ptr = copy_ssa_name (var: ptr, stmt: new_stmt); |
11697 | else |
11698 | ptr = make_ssa_name (TREE_TYPE (ptr), stmt: new_stmt); |
11699 | gimple_assign_set_lhs (gs: new_stmt, lhs: ptr); |
11700 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
11701 | data_ref |
11702 | = build2 (MEM_REF, vectype, ptr, build_int_cst (ref_type, 0)); |
11703 | break; |
11704 | } |
11705 | case dr_explicit_realign_optimized: |
11706 | { |
11707 | if (costing_p) |
11708 | break; |
11709 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
11710 | new_temp = copy_ssa_name (var: dataref_ptr); |
11711 | else |
11712 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); |
11713 | // We should only be doing this if we know the target |
11714 | // alignment at compile time. |
11715 | unsigned HOST_WIDE_INT align |
11716 | = DR_TARGET_ALIGNMENT (first_dr_info).to_constant (); |
11717 | new_stmt = gimple_build_assign ( |
11718 | new_temp, BIT_AND_EXPR, dataref_ptr, |
11719 | build_int_cst (TREE_TYPE (dataref_ptr), |
11720 | -(HOST_WIDE_INT) align)); |
11721 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
11722 | data_ref = build2 (MEM_REF, vectype, new_temp, |
11723 | build_int_cst (ref_type, 0)); |
11724 | break; |
11725 | } |
11726 | default: |
11727 | gcc_unreachable (); |
11728 | } |
11729 | |
11730 | /* One common place to cost the above vect load for different |
11731 | alignment support schemes. */ |
11732 | if (costing_p) |
11733 | { |
11734 | /* For VMAT_CONTIGUOUS_PERMUTE if it's grouped load, we |
11735 | only need to take care of the first stmt, whose |
11736 | stmt_info is first_stmt_info, vec_num iterating on it |
11737 | will cover the cost for the remaining, it's consistent |
11738 | with transforming. For the prologue cost for realign, |
11739 | we only need to count it once for the whole group. */ |
11740 | bool first_stmt_info_p = first_stmt_info == stmt_info; |
11741 | bool add_realign_cost = first_stmt_info_p && i == 0; |
11742 | if (memory_access_type == VMAT_CONTIGUOUS |
11743 | || memory_access_type == VMAT_CONTIGUOUS_REVERSE |
11744 | || (memory_access_type == VMAT_CONTIGUOUS_PERMUTE |
11745 | && (!grouped_load || first_stmt_info_p))) |
11746 | { |
11747 | /* Leave realign cases alone to keep them simple. */ |
11748 | if (alignment_support_scheme == dr_explicit_realign_optimized |
11749 | || alignment_support_scheme == dr_explicit_realign) |
11750 | vect_get_load_cost (vinfo, stmt_info, ncopies: 1, |
11751 | alignment_support_scheme, misalignment, |
11752 | add_realign_cost, inside_cost: &inside_cost, |
11753 | prologue_cost: &prologue_cost, prologue_cost_vec: cost_vec, body_cost_vec: cost_vec, |
11754 | record_prologue_costs: true); |
11755 | else |
11756 | n_adjacent_loads++; |
11757 | } |
11758 | } |
11759 | else |
11760 | { |
11761 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
11762 | /* DATA_REF is null if we've already built the statement. */ |
11763 | if (data_ref) |
11764 | { |
11765 | vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr)); |
11766 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
11767 | } |
11768 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
11769 | gimple_set_lhs (new_stmt, new_temp); |
11770 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
11771 | } |
11772 | |
11773 | /* 3. Handle explicit realignment if necessary/supported. |
11774 | Create in loop: |
11775 | vec_dest = realign_load (msq, lsq, realignment_token) */ |
11776 | if (!costing_p |
11777 | && (alignment_support_scheme == dr_explicit_realign_optimized |
11778 | || alignment_support_scheme == dr_explicit_realign)) |
11779 | { |
11780 | lsq = gimple_assign_lhs (gs: new_stmt); |
11781 | if (!realignment_token) |
11782 | realignment_token = dataref_ptr; |
11783 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
11784 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, msq, |
11785 | lsq, realignment_token); |
11786 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
11787 | gimple_assign_set_lhs (gs: new_stmt, lhs: new_temp); |
11788 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
11789 | |
11790 | if (alignment_support_scheme == dr_explicit_realign_optimized) |
11791 | { |
11792 | gcc_assert (phi); |
11793 | if (i == vec_num - 1 && j == ncopies - 1) |
11794 | add_phi_arg (phi, lsq, loop_latch_edge (containing_loop), |
11795 | UNKNOWN_LOCATION); |
11796 | msq = lsq; |
11797 | } |
11798 | } |
11799 | |
11800 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
11801 | { |
11802 | if (costing_p) |
11803 | inside_cost = record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: vec_perm, |
11804 | stmt_info, misalign: 0, where: vect_body); |
11805 | else |
11806 | { |
11807 | tree perm_mask = perm_mask_for_reverse (vectype); |
11808 | new_temp = permute_vec_elements (vinfo, x: new_temp, y: new_temp, |
11809 | mask_vec: perm_mask, stmt_info, gsi); |
11810 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
11811 | } |
11812 | } |
11813 | |
11814 | /* Collect vector loads and later create their permutation in |
11815 | vect_transform_grouped_load (). */ |
11816 | if (!costing_p && (grouped_load || slp_perm)) |
11817 | dr_chain.quick_push (obj: new_temp); |
11818 | |
11819 | /* Store vector loads in the corresponding SLP_NODE. */ |
11820 | if (!costing_p && slp && !slp_perm) |
11821 | slp_node->push_vec_def (def: new_stmt); |
11822 | |
11823 | /* With SLP permutation we load the gaps as well, without |
11824 | we need to skip the gaps after we manage to fully load |
11825 | all elements. group_gap_adj is DR_GROUP_SIZE here. */ |
11826 | group_elt += nunits; |
11827 | if (!costing_p |
11828 | && maybe_ne (a: group_gap_adj, b: 0U) |
11829 | && !slp_perm |
11830 | && known_eq (group_elt, group_size - group_gap_adj)) |
11831 | { |
11832 | poly_wide_int bump_val |
11833 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) * group_gap_adj); |
11834 | if (tree_int_cst_sgn (vect_dr_behavior (vinfo, dr_info)->step) |
11835 | == -1) |
11836 | bump_val = -bump_val; |
11837 | tree bump = wide_int_to_tree (sizetype, cst: bump_val); |
11838 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi, |
11839 | stmt_info, bump); |
11840 | group_elt = 0; |
11841 | } |
11842 | } |
11843 | /* Bump the vector pointer to account for a gap or for excess |
11844 | elements loaded for a permuted SLP load. */ |
11845 | if (!costing_p |
11846 | && maybe_ne (a: group_gap_adj, b: 0U) |
11847 | && slp_perm) |
11848 | { |
11849 | poly_wide_int bump_val |
11850 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) * group_gap_adj); |
11851 | if (tree_int_cst_sgn (vect_dr_behavior (vinfo, dr_info)->step) == -1) |
11852 | bump_val = -bump_val; |
11853 | tree bump = wide_int_to_tree (sizetype, cst: bump_val); |
11854 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi, |
11855 | stmt_info, bump); |
11856 | } |
11857 | |
11858 | if (slp && !slp_perm) |
11859 | continue; |
11860 | |
11861 | if (slp_perm) |
11862 | { |
11863 | unsigned n_perms; |
11864 | /* For SLP we know we've seen all possible uses of dr_chain so |
11865 | direct vect_transform_slp_perm_load to DCE the unused parts. |
11866 | ??? This is a hack to prevent compile-time issues as seen |
11867 | in PR101120 and friends. */ |
11868 | if (costing_p) |
11869 | { |
11870 | vect_transform_slp_perm_load (vinfo, slp_node, vNULL, nullptr, vf, |
11871 | true, &n_perms, nullptr); |
11872 | inside_cost = record_stmt_cost (body_cost_vec: cost_vec, count: n_perms, kind: vec_perm, |
11873 | stmt_info, misalign: 0, where: vect_body); |
11874 | } |
11875 | else |
11876 | { |
11877 | bool ok = vect_transform_slp_perm_load (vinfo, slp_node, dr_chain, |
11878 | gsi, vf, false, &n_perms, |
11879 | nullptr, true); |
11880 | gcc_assert (ok); |
11881 | } |
11882 | } |
11883 | else |
11884 | { |
11885 | if (grouped_load) |
11886 | { |
11887 | gcc_assert (memory_access_type == VMAT_CONTIGUOUS_PERMUTE); |
11888 | /* We assume that the cost of a single load-lanes instruction |
11889 | is equivalent to the cost of DR_GROUP_SIZE separate loads. |
11890 | If a grouped access is instead being provided by a |
11891 | load-and-permute operation, include the cost of the |
11892 | permutes. */ |
11893 | if (costing_p && first_stmt_info == stmt_info) |
11894 | { |
11895 | /* Uses an even and odd extract operations or shuffle |
11896 | operations for each needed permute. */ |
11897 | int group_size = DR_GROUP_SIZE (first_stmt_info); |
11898 | int nstmts = ceil_log2 (x: group_size) * group_size; |
11899 | inside_cost += record_stmt_cost (body_cost_vec: cost_vec, count: nstmts, kind: vec_perm, |
11900 | stmt_info, misalign: 0, where: vect_body); |
11901 | |
11902 | if (dump_enabled_p ()) |
11903 | dump_printf_loc (MSG_NOTE, vect_location, |
11904 | "vect_model_load_cost:" |
11905 | "strided group_size = %d .\n" , |
11906 | group_size); |
11907 | } |
11908 | else if (!costing_p) |
11909 | { |
11910 | vect_transform_grouped_load (vinfo, stmt_info, dr_chain, |
11911 | group_size, gsi); |
11912 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
11913 | } |
11914 | } |
11915 | else if (!costing_p) |
11916 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
11917 | } |
11918 | dr_chain.release (); |
11919 | } |
11920 | if (!slp && !costing_p) |
11921 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
11922 | |
11923 | if (costing_p) |
11924 | { |
11925 | gcc_assert (memory_access_type == VMAT_CONTIGUOUS |
11926 | || memory_access_type == VMAT_CONTIGUOUS_REVERSE |
11927 | || memory_access_type == VMAT_CONTIGUOUS_PERMUTE); |
11928 | if (n_adjacent_loads > 0) |
11929 | vect_get_load_cost (vinfo, stmt_info, ncopies: n_adjacent_loads, |
11930 | alignment_support_scheme, misalignment, add_realign_cost: false, |
11931 | inside_cost: &inside_cost, prologue_cost: &prologue_cost, prologue_cost_vec: cost_vec, body_cost_vec: cost_vec, |
11932 | record_prologue_costs: true); |
11933 | if (dump_enabled_p ()) |
11934 | dump_printf_loc (MSG_NOTE, vect_location, |
11935 | "vect_model_load_cost: inside_cost = %u, " |
11936 | "prologue_cost = %u .\n" , |
11937 | inside_cost, prologue_cost); |
11938 | } |
11939 | |
11940 | return true; |
11941 | } |
11942 | |
11943 | /* Function vect_is_simple_cond. |
11944 | |
11945 | Input: |
11946 | LOOP - the loop that is being vectorized. |
11947 | COND - Condition that is checked for simple use. |
11948 | |
11949 | Output: |
11950 | *COMP_VECTYPE - the vector type for the comparison. |
11951 | *DTS - The def types for the arguments of the comparison |
11952 | |
11953 | Returns whether a COND can be vectorized. Checks whether |
11954 | condition operands are supportable using vec_is_simple_use. */ |
11955 | |
11956 | static bool |
11957 | vect_is_simple_cond (tree cond, vec_info *vinfo, stmt_vec_info stmt_info, |
11958 | slp_tree slp_node, tree *comp_vectype, |
11959 | enum vect_def_type *dts, tree vectype) |
11960 | { |
11961 | tree lhs, rhs; |
11962 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
11963 | slp_tree slp_op; |
11964 | |
11965 | /* Mask case. */ |
11966 | if (TREE_CODE (cond) == SSA_NAME |
11967 | && VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (cond))) |
11968 | { |
11969 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 0, &cond, |
11970 | &slp_op, &dts[0], comp_vectype) |
11971 | || !*comp_vectype |
11972 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) |
11973 | return false; |
11974 | return true; |
11975 | } |
11976 | |
11977 | if (!COMPARISON_CLASS_P (cond)) |
11978 | return false; |
11979 | |
11980 | lhs = TREE_OPERAND (cond, 0); |
11981 | rhs = TREE_OPERAND (cond, 1); |
11982 | |
11983 | if (TREE_CODE (lhs) == SSA_NAME) |
11984 | { |
11985 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 0, |
11986 | &lhs, &slp_op, &dts[0], &vectype1)) |
11987 | return false; |
11988 | } |
11989 | else if (TREE_CODE (lhs) == INTEGER_CST || TREE_CODE (lhs) == REAL_CST |
11990 | || TREE_CODE (lhs) == FIXED_CST) |
11991 | dts[0] = vect_constant_def; |
11992 | else |
11993 | return false; |
11994 | |
11995 | if (TREE_CODE (rhs) == SSA_NAME) |
11996 | { |
11997 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 1, |
11998 | &rhs, &slp_op, &dts[1], &vectype2)) |
11999 | return false; |
12000 | } |
12001 | else if (TREE_CODE (rhs) == INTEGER_CST || TREE_CODE (rhs) == REAL_CST |
12002 | || TREE_CODE (rhs) == FIXED_CST) |
12003 | dts[1] = vect_constant_def; |
12004 | else |
12005 | return false; |
12006 | |
12007 | if (vectype1 && vectype2 |
12008 | && maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: vectype1), |
12009 | b: TYPE_VECTOR_SUBPARTS (node: vectype2))) |
12010 | return false; |
12011 | |
12012 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
12013 | /* Invariant comparison. */ |
12014 | if (! *comp_vectype) |
12015 | { |
12016 | tree scalar_type = TREE_TYPE (lhs); |
12017 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type)) |
12018 | *comp_vectype = truth_type_for (vectype); |
12019 | else |
12020 | { |
12021 | /* If we can widen the comparison to match vectype do so. */ |
12022 | if (INTEGRAL_TYPE_P (scalar_type) |
12023 | && !slp_node |
12024 | && tree_int_cst_lt (TYPE_SIZE (scalar_type), |
12025 | TYPE_SIZE (TREE_TYPE (vectype)))) |
12026 | scalar_type = build_nonstandard_integer_type |
12027 | (vector_element_bits (vectype), TYPE_UNSIGNED (scalar_type)); |
12028 | *comp_vectype = get_vectype_for_scalar_type (vinfo, scalar_type, |
12029 | slp_node); |
12030 | } |
12031 | } |
12032 | |
12033 | return true; |
12034 | } |
12035 | |
12036 | /* vectorizable_condition. |
12037 | |
12038 | Check if STMT_INFO is conditional modify expression that can be vectorized. |
12039 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
12040 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it |
12041 | at GSI. |
12042 | |
12043 | When STMT_INFO is vectorized as a nested cycle, for_reduction is true. |
12044 | |
12045 | Return true if STMT_INFO is vectorizable in this way. */ |
12046 | |
12047 | static bool |
12048 | vectorizable_condition (vec_info *vinfo, |
12049 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
12050 | gimple **vec_stmt, |
12051 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
12052 | { |
12053 | tree scalar_dest = NULL_TREE; |
12054 | tree vec_dest = NULL_TREE; |
12055 | tree cond_expr, cond_expr0 = NULL_TREE, cond_expr1 = NULL_TREE; |
12056 | tree then_clause, else_clause; |
12057 | tree comp_vectype = NULL_TREE; |
12058 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
12059 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; |
12060 | tree vec_compare; |
12061 | tree new_temp; |
12062 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
12063 | enum vect_def_type dts[4] |
12064 | = {vect_unknown_def_type, vect_unknown_def_type, |
12065 | vect_unknown_def_type, vect_unknown_def_type}; |
12066 | int ndts = 4; |
12067 | int ncopies; |
12068 | int vec_num; |
12069 | enum tree_code code, cond_code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
12070 | int i; |
12071 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
12072 | vec<tree> vec_oprnds0 = vNULL; |
12073 | vec<tree> vec_oprnds1 = vNULL; |
12074 | vec<tree> vec_oprnds2 = vNULL; |
12075 | vec<tree> vec_oprnds3 = vNULL; |
12076 | tree vec_cmp_type; |
12077 | bool masked = false; |
12078 | |
12079 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
12080 | return false; |
12081 | |
12082 | /* Is vectorizable conditional operation? */ |
12083 | gassign *stmt = dyn_cast <gassign *> (p: stmt_info->stmt); |
12084 | if (!stmt) |
12085 | return false; |
12086 | |
12087 | code = gimple_assign_rhs_code (gs: stmt); |
12088 | if (code != COND_EXPR) |
12089 | return false; |
12090 | |
12091 | stmt_vec_info reduc_info = NULL; |
12092 | int reduc_index = -1; |
12093 | vect_reduction_type reduction_type = TREE_CODE_REDUCTION; |
12094 | bool for_reduction |
12095 | = STMT_VINFO_REDUC_DEF (vect_orig_stmt (stmt_info)) != NULL; |
12096 | if (for_reduction) |
12097 | { |
12098 | if (slp_node) |
12099 | return false; |
12100 | reduc_info = info_for_reduction (vinfo, stmt_info); |
12101 | reduction_type = STMT_VINFO_REDUC_TYPE (reduc_info); |
12102 | reduc_index = STMT_VINFO_REDUC_IDX (stmt_info); |
12103 | gcc_assert (reduction_type != EXTRACT_LAST_REDUCTION |
12104 | || reduc_index != -1); |
12105 | } |
12106 | else |
12107 | { |
12108 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
12109 | return false; |
12110 | } |
12111 | |
12112 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
12113 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
12114 | |
12115 | if (slp_node) |
12116 | { |
12117 | ncopies = 1; |
12118 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
12119 | } |
12120 | else |
12121 | { |
12122 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
12123 | vec_num = 1; |
12124 | } |
12125 | |
12126 | gcc_assert (ncopies >= 1); |
12127 | if (for_reduction && ncopies > 1) |
12128 | return false; /* FORNOW */ |
12129 | |
12130 | cond_expr = gimple_assign_rhs1 (gs: stmt); |
12131 | |
12132 | if (!vect_is_simple_cond (cond: cond_expr, vinfo, stmt_info, slp_node, |
12133 | comp_vectype: &comp_vectype, dts: &dts[0], vectype) |
12134 | || !comp_vectype) |
12135 | return false; |
12136 | |
12137 | unsigned op_adjust = COMPARISON_CLASS_P (cond_expr) ? 1 : 0; |
12138 | slp_tree then_slp_node, else_slp_node; |
12139 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 1 + op_adjust, |
12140 | &then_clause, &then_slp_node, &dts[2], &vectype1)) |
12141 | return false; |
12142 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 2 + op_adjust, |
12143 | &else_clause, &else_slp_node, &dts[3], &vectype2)) |
12144 | return false; |
12145 | |
12146 | if (vectype1 && !useless_type_conversion_p (vectype, vectype1)) |
12147 | return false; |
12148 | |
12149 | if (vectype2 && !useless_type_conversion_p (vectype, vectype2)) |
12150 | return false; |
12151 | |
12152 | masked = !COMPARISON_CLASS_P (cond_expr); |
12153 | vec_cmp_type = truth_type_for (comp_vectype); |
12154 | |
12155 | if (vec_cmp_type == NULL_TREE) |
12156 | return false; |
12157 | |
12158 | cond_code = TREE_CODE (cond_expr); |
12159 | if (!masked) |
12160 | { |
12161 | cond_expr0 = TREE_OPERAND (cond_expr, 0); |
12162 | cond_expr1 = TREE_OPERAND (cond_expr, 1); |
12163 | } |
12164 | |
12165 | /* For conditional reductions, the "then" value needs to be the candidate |
12166 | value calculated by this iteration while the "else" value needs to be |
12167 | the result carried over from previous iterations. If the COND_EXPR |
12168 | is the other way around, we need to swap it. */ |
12169 | bool must_invert_cmp_result = false; |
12170 | if (reduction_type == EXTRACT_LAST_REDUCTION && reduc_index == 1) |
12171 | { |
12172 | if (masked) |
12173 | must_invert_cmp_result = true; |
12174 | else |
12175 | { |
12176 | bool honor_nans = HONOR_NANS (TREE_TYPE (cond_expr0)); |
12177 | tree_code new_code = invert_tree_comparison (cond_code, honor_nans); |
12178 | if (new_code == ERROR_MARK) |
12179 | must_invert_cmp_result = true; |
12180 | else |
12181 | { |
12182 | cond_code = new_code; |
12183 | /* Make sure we don't accidentally use the old condition. */ |
12184 | cond_expr = NULL_TREE; |
12185 | } |
12186 | } |
12187 | std::swap (a&: then_clause, b&: else_clause); |
12188 | } |
12189 | |
12190 | if (!masked && VECTOR_BOOLEAN_TYPE_P (comp_vectype)) |
12191 | { |
12192 | /* Boolean values may have another representation in vectors |
12193 | and therefore we prefer bit operations over comparison for |
12194 | them (which also works for scalar masks). We store opcodes |
12195 | to use in bitop1 and bitop2. Statement is vectorized as |
12196 | BITOP2 (rhs1 BITOP1 rhs2) or rhs1 BITOP2 (BITOP1 rhs2) |
12197 | depending on bitop1 and bitop2 arity. */ |
12198 | switch (cond_code) |
12199 | { |
12200 | case GT_EXPR: |
12201 | bitop1 = BIT_NOT_EXPR; |
12202 | bitop2 = BIT_AND_EXPR; |
12203 | break; |
12204 | case GE_EXPR: |
12205 | bitop1 = BIT_NOT_EXPR; |
12206 | bitop2 = BIT_IOR_EXPR; |
12207 | break; |
12208 | case LT_EXPR: |
12209 | bitop1 = BIT_NOT_EXPR; |
12210 | bitop2 = BIT_AND_EXPR; |
12211 | std::swap (a&: cond_expr0, b&: cond_expr1); |
12212 | break; |
12213 | case LE_EXPR: |
12214 | bitop1 = BIT_NOT_EXPR; |
12215 | bitop2 = BIT_IOR_EXPR; |
12216 | std::swap (a&: cond_expr0, b&: cond_expr1); |
12217 | break; |
12218 | case NE_EXPR: |
12219 | bitop1 = BIT_XOR_EXPR; |
12220 | break; |
12221 | case EQ_EXPR: |
12222 | bitop1 = BIT_XOR_EXPR; |
12223 | bitop2 = BIT_NOT_EXPR; |
12224 | break; |
12225 | default: |
12226 | return false; |
12227 | } |
12228 | cond_code = SSA_NAME; |
12229 | } |
12230 | |
12231 | if (TREE_CODE_CLASS (cond_code) == tcc_comparison |
12232 | && reduction_type == EXTRACT_LAST_REDUCTION |
12233 | && !expand_vec_cmp_expr_p (comp_vectype, vec_cmp_type, cond_code)) |
12234 | { |
12235 | if (dump_enabled_p ()) |
12236 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
12237 | "reduction comparison operation not supported.\n" ); |
12238 | return false; |
12239 | } |
12240 | |
12241 | if (!vec_stmt) |
12242 | { |
12243 | if (bitop1 != NOP_EXPR) |
12244 | { |
12245 | machine_mode mode = TYPE_MODE (comp_vectype); |
12246 | optab optab; |
12247 | |
12248 | optab = optab_for_tree_code (bitop1, comp_vectype, optab_default); |
12249 | if (!optab || optab_handler (op: optab, mode) == CODE_FOR_nothing) |
12250 | return false; |
12251 | |
12252 | if (bitop2 != NOP_EXPR) |
12253 | { |
12254 | optab = optab_for_tree_code (bitop2, comp_vectype, |
12255 | optab_default); |
12256 | if (!optab || optab_handler (op: optab, mode) == CODE_FOR_nothing) |
12257 | return false; |
12258 | } |
12259 | } |
12260 | |
12261 | vect_cost_for_stmt kind = vector_stmt; |
12262 | if (reduction_type == EXTRACT_LAST_REDUCTION) |
12263 | /* Count one reduction-like operation per vector. */ |
12264 | kind = vec_to_scalar; |
12265 | else if (!expand_vec_cond_expr_p (vectype, comp_vectype, cond_code) |
12266 | && (masked |
12267 | || (!expand_vec_cmp_expr_p (comp_vectype, vec_cmp_type, |
12268 | cond_code) |
12269 | || !expand_vec_cond_expr_p (vectype, vec_cmp_type, |
12270 | ERROR_MARK)))) |
12271 | return false; |
12272 | |
12273 | if (slp_node |
12274 | && (!vect_maybe_update_slp_op_vectype |
12275 | (SLP_TREE_CHILDREN (slp_node)[0], comp_vectype) |
12276 | || (op_adjust == 1 |
12277 | && !vect_maybe_update_slp_op_vectype |
12278 | (SLP_TREE_CHILDREN (slp_node)[1], comp_vectype)) |
12279 | || !vect_maybe_update_slp_op_vectype (then_slp_node, vectype) |
12280 | || !vect_maybe_update_slp_op_vectype (else_slp_node, vectype))) |
12281 | { |
12282 | if (dump_enabled_p ()) |
12283 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
12284 | "incompatible vector types for invariants\n" ); |
12285 | return false; |
12286 | } |
12287 | |
12288 | if (loop_vinfo && for_reduction |
12289 | && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) |
12290 | { |
12291 | if (reduction_type == EXTRACT_LAST_REDUCTION) |
12292 | { |
12293 | if (direct_internal_fn_supported_p (IFN_LEN_FOLD_EXTRACT_LAST, |
12294 | vectype, OPTIMIZE_FOR_SPEED)) |
12295 | vect_record_loop_len (loop_vinfo, |
12296 | &LOOP_VINFO_LENS (loop_vinfo), |
12297 | ncopies * vec_num, vectype, 1); |
12298 | else |
12299 | vect_record_loop_mask (loop_vinfo, |
12300 | &LOOP_VINFO_MASKS (loop_vinfo), |
12301 | ncopies * vec_num, vectype, NULL); |
12302 | } |
12303 | /* Extra inactive lanes should be safe for vect_nested_cycle. */ |
12304 | else if (STMT_VINFO_DEF_TYPE (reduc_info) != vect_nested_cycle) |
12305 | { |
12306 | if (dump_enabled_p ()) |
12307 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
12308 | "conditional reduction prevents the use" |
12309 | " of partial vectors.\n" ); |
12310 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
12311 | } |
12312 | } |
12313 | |
12314 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; |
12315 | vect_model_simple_cost (vinfo, stmt_info, ncopies, dt: dts, ndts, node: slp_node, |
12316 | cost_vec, kind); |
12317 | return true; |
12318 | } |
12319 | |
12320 | /* Transform. */ |
12321 | |
12322 | /* Handle def. */ |
12323 | scalar_dest = gimple_assign_lhs (gs: stmt); |
12324 | if (reduction_type != EXTRACT_LAST_REDUCTION) |
12325 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
12326 | |
12327 | bool swap_cond_operands = false; |
12328 | |
12329 | /* See whether another part of the vectorized code applies a loop |
12330 | mask to the condition, or to its inverse. */ |
12331 | |
12332 | vec_loop_masks *masks = NULL; |
12333 | vec_loop_lens *lens = NULL; |
12334 | if (loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo)) |
12335 | { |
12336 | if (reduction_type == EXTRACT_LAST_REDUCTION) |
12337 | lens = &LOOP_VINFO_LENS (loop_vinfo); |
12338 | } |
12339 | else if (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) |
12340 | { |
12341 | if (reduction_type == EXTRACT_LAST_REDUCTION) |
12342 | masks = &LOOP_VINFO_MASKS (loop_vinfo); |
12343 | else |
12344 | { |
12345 | scalar_cond_masked_key cond (cond_expr, ncopies); |
12346 | if (loop_vinfo->scalar_cond_masked_set.contains (k: cond)) |
12347 | masks = &LOOP_VINFO_MASKS (loop_vinfo); |
12348 | else |
12349 | { |
12350 | bool honor_nans = HONOR_NANS (TREE_TYPE (cond.op0)); |
12351 | tree_code orig_code = cond.code; |
12352 | cond.code = invert_tree_comparison (cond.code, honor_nans); |
12353 | if (!masked && loop_vinfo->scalar_cond_masked_set.contains (k: cond)) |
12354 | { |
12355 | masks = &LOOP_VINFO_MASKS (loop_vinfo); |
12356 | cond_code = cond.code; |
12357 | swap_cond_operands = true; |
12358 | } |
12359 | else |
12360 | { |
12361 | /* Try the inverse of the current mask. We check if the |
12362 | inverse mask is live and if so we generate a negate of |
12363 | the current mask such that we still honor NaNs. */ |
12364 | cond.inverted_p = true; |
12365 | cond.code = orig_code; |
12366 | if (loop_vinfo->scalar_cond_masked_set.contains (k: cond)) |
12367 | { |
12368 | masks = &LOOP_VINFO_MASKS (loop_vinfo); |
12369 | cond_code = cond.code; |
12370 | swap_cond_operands = true; |
12371 | must_invert_cmp_result = true; |
12372 | } |
12373 | } |
12374 | } |
12375 | } |
12376 | } |
12377 | |
12378 | /* Handle cond expr. */ |
12379 | if (masked) |
12380 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, |
12381 | op0: cond_expr, vectype0: comp_vectype, vec_oprnds0: &vec_oprnds0, |
12382 | op1: then_clause, vectype1: vectype, vec_oprnds1: &vec_oprnds2, |
12383 | op2: reduction_type != EXTRACT_LAST_REDUCTION |
12384 | ? else_clause : NULL, vectype2: vectype, vec_oprnds2: &vec_oprnds3); |
12385 | else |
12386 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, |
12387 | op0: cond_expr0, vectype0: comp_vectype, vec_oprnds0: &vec_oprnds0, |
12388 | op1: cond_expr1, vectype1: comp_vectype, vec_oprnds1: &vec_oprnds1, |
12389 | op2: then_clause, vectype2: vectype, vec_oprnds2: &vec_oprnds2, |
12390 | op3: reduction_type != EXTRACT_LAST_REDUCTION |
12391 | ? else_clause : NULL, vectype3: vectype, vec_oprnds3: &vec_oprnds3); |
12392 | |
12393 | /* Arguments are ready. Create the new vector stmt. */ |
12394 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
12395 | { |
12396 | vec_then_clause = vec_oprnds2[i]; |
12397 | if (reduction_type != EXTRACT_LAST_REDUCTION) |
12398 | vec_else_clause = vec_oprnds3[i]; |
12399 | |
12400 | if (swap_cond_operands) |
12401 | std::swap (a&: vec_then_clause, b&: vec_else_clause); |
12402 | |
12403 | if (masked) |
12404 | vec_compare = vec_cond_lhs; |
12405 | else |
12406 | { |
12407 | vec_cond_rhs = vec_oprnds1[i]; |
12408 | if (bitop1 == NOP_EXPR) |
12409 | { |
12410 | gimple_seq stmts = NULL; |
12411 | vec_compare = gimple_build (seq: &stmts, code: cond_code, type: vec_cmp_type, |
12412 | ops: vec_cond_lhs, ops: vec_cond_rhs); |
12413 | gsi_insert_before (gsi, stmts, GSI_SAME_STMT); |
12414 | } |
12415 | else |
12416 | { |
12417 | new_temp = make_ssa_name (var: vec_cmp_type); |
12418 | gassign *new_stmt; |
12419 | if (bitop1 == BIT_NOT_EXPR) |
12420 | new_stmt = gimple_build_assign (new_temp, bitop1, |
12421 | vec_cond_rhs); |
12422 | else |
12423 | new_stmt |
12424 | = gimple_build_assign (new_temp, bitop1, vec_cond_lhs, |
12425 | vec_cond_rhs); |
12426 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
12427 | if (bitop2 == NOP_EXPR) |
12428 | vec_compare = new_temp; |
12429 | else if (bitop2 == BIT_NOT_EXPR |
12430 | && reduction_type != EXTRACT_LAST_REDUCTION) |
12431 | { |
12432 | /* Instead of doing ~x ? y : z do x ? z : y. */ |
12433 | vec_compare = new_temp; |
12434 | std::swap (a&: vec_then_clause, b&: vec_else_clause); |
12435 | } |
12436 | else |
12437 | { |
12438 | vec_compare = make_ssa_name (var: vec_cmp_type); |
12439 | if (bitop2 == BIT_NOT_EXPR) |
12440 | new_stmt |
12441 | = gimple_build_assign (vec_compare, bitop2, new_temp); |
12442 | else |
12443 | new_stmt |
12444 | = gimple_build_assign (vec_compare, bitop2, |
12445 | vec_cond_lhs, new_temp); |
12446 | vect_finish_stmt_generation (vinfo, stmt_info, |
12447 | vec_stmt: new_stmt, gsi); |
12448 | } |
12449 | } |
12450 | } |
12451 | |
12452 | /* If we decided to apply a loop mask to the result of the vector |
12453 | comparison, AND the comparison with the mask now. Later passes |
12454 | should then be able to reuse the AND results between mulitple |
12455 | vector statements. |
12456 | |
12457 | For example: |
12458 | for (int i = 0; i < 100; ++i) |
12459 | x[i] = y[i] ? z[i] : 10; |
12460 | |
12461 | results in following optimized GIMPLE: |
12462 | |
12463 | mask__35.8_43 = vect__4.7_41 != { 0, ... }; |
12464 | vec_mask_and_46 = loop_mask_40 & mask__35.8_43; |
12465 | _19 = &MEM[base: z_12(D), index: ivtmp_56, step: 4, offset: 0B]; |
12466 | vect_iftmp.11_47 = .MASK_LOAD (_19, 4B, vec_mask_and_46); |
12467 | vect_iftmp.12_52 = VEC_COND_EXPR <vec_mask_and_46, |
12468 | vect_iftmp.11_47, { 10, ... }>; |
12469 | |
12470 | instead of using a masked and unmasked forms of |
12471 | vec != { 0, ... } (masked in the MASK_LOAD, |
12472 | unmasked in the VEC_COND_EXPR). */ |
12473 | |
12474 | /* Force vec_compare to be an SSA_NAME rather than a comparison, |
12475 | in cases where that's necessary. */ |
12476 | |
12477 | tree len = NULL_TREE, bias = NULL_TREE; |
12478 | if (masks || lens || reduction_type == EXTRACT_LAST_REDUCTION) |
12479 | { |
12480 | if (!is_gimple_val (vec_compare)) |
12481 | { |
12482 | tree vec_compare_name = make_ssa_name (var: vec_cmp_type); |
12483 | gassign *new_stmt = gimple_build_assign (vec_compare_name, |
12484 | vec_compare); |
12485 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
12486 | vec_compare = vec_compare_name; |
12487 | } |
12488 | |
12489 | if (must_invert_cmp_result) |
12490 | { |
12491 | tree vec_compare_name = make_ssa_name (var: vec_cmp_type); |
12492 | gassign *new_stmt = gimple_build_assign (vec_compare_name, |
12493 | BIT_NOT_EXPR, |
12494 | vec_compare); |
12495 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
12496 | vec_compare = vec_compare_name; |
12497 | } |
12498 | |
12499 | if (direct_internal_fn_supported_p (IFN_LEN_FOLD_EXTRACT_LAST, |
12500 | vectype, OPTIMIZE_FOR_SPEED)) |
12501 | { |
12502 | if (lens) |
12503 | { |
12504 | len = vect_get_loop_len (loop_vinfo, gsi, lens, |
12505 | vec_num * ncopies, vectype, i, 1); |
12506 | signed char biasval |
12507 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
12508 | bias = build_int_cst (intQI_type_node, biasval); |
12509 | } |
12510 | else |
12511 | { |
12512 | len = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
12513 | bias = build_int_cst (intQI_type_node, 0); |
12514 | } |
12515 | } |
12516 | if (masks) |
12517 | { |
12518 | tree loop_mask |
12519 | = vect_get_loop_mask (loop_vinfo, gsi, masks, vec_num * ncopies, |
12520 | vectype, i); |
12521 | tree tmp2 = make_ssa_name (var: vec_cmp_type); |
12522 | gassign *g |
12523 | = gimple_build_assign (tmp2, BIT_AND_EXPR, vec_compare, |
12524 | loop_mask); |
12525 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
12526 | vec_compare = tmp2; |
12527 | } |
12528 | } |
12529 | |
12530 | gimple *new_stmt; |
12531 | if (reduction_type == EXTRACT_LAST_REDUCTION) |
12532 | { |
12533 | gimple *old_stmt = vect_orig_stmt (stmt_info)->stmt; |
12534 | tree lhs = gimple_get_lhs (old_stmt); |
12535 | if (len) |
12536 | new_stmt = gimple_build_call_internal |
12537 | (IFN_LEN_FOLD_EXTRACT_LAST, 5, else_clause, vec_compare, |
12538 | vec_then_clause, len, bias); |
12539 | else |
12540 | new_stmt = gimple_build_call_internal |
12541 | (IFN_FOLD_EXTRACT_LAST, 3, else_clause, vec_compare, |
12542 | vec_then_clause); |
12543 | gimple_call_set_lhs (gs: new_stmt, lhs); |
12544 | SSA_NAME_DEF_STMT (lhs) = new_stmt; |
12545 | if (old_stmt == gsi_stmt (i: *gsi)) |
12546 | vect_finish_replace_stmt (vinfo, stmt_info, vec_stmt: new_stmt); |
12547 | else |
12548 | { |
12549 | /* In this case we're moving the definition to later in the |
12550 | block. That doesn't matter because the only uses of the |
12551 | lhs are in phi statements. */ |
12552 | gimple_stmt_iterator old_gsi = gsi_for_stmt (old_stmt); |
12553 | gsi_remove (&old_gsi, true); |
12554 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
12555 | } |
12556 | } |
12557 | else |
12558 | { |
12559 | new_temp = make_ssa_name (var: vec_dest); |
12560 | new_stmt = gimple_build_assign (new_temp, VEC_COND_EXPR, vec_compare, |
12561 | vec_then_clause, vec_else_clause); |
12562 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
12563 | } |
12564 | if (slp_node) |
12565 | slp_node->push_vec_def (def: new_stmt); |
12566 | else |
12567 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
12568 | } |
12569 | |
12570 | if (!slp_node) |
12571 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
12572 | |
12573 | vec_oprnds0.release (); |
12574 | vec_oprnds1.release (); |
12575 | vec_oprnds2.release (); |
12576 | vec_oprnds3.release (); |
12577 | |
12578 | return true; |
12579 | } |
12580 | |
12581 | /* Helper of vectorizable_comparison. |
12582 | |
12583 | Check if STMT_INFO is comparison expression CODE that can be vectorized. |
12584 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
12585 | comparison, put it in VEC_STMT, and insert it at GSI. |
12586 | |
12587 | Return true if STMT_INFO is vectorizable in this way. */ |
12588 | |
12589 | static bool |
12590 | vectorizable_comparison_1 (vec_info *vinfo, tree vectype, |
12591 | stmt_vec_info stmt_info, tree_code code, |
12592 | gimple_stmt_iterator *gsi, gimple **vec_stmt, |
12593 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
12594 | { |
12595 | tree lhs, rhs1, rhs2; |
12596 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
12597 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; |
12598 | tree new_temp; |
12599 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
12600 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
12601 | int ndts = 2; |
12602 | poly_uint64 nunits; |
12603 | int ncopies; |
12604 | enum tree_code bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
12605 | int i; |
12606 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
12607 | vec<tree> vec_oprnds0 = vNULL; |
12608 | vec<tree> vec_oprnds1 = vNULL; |
12609 | tree mask_type; |
12610 | tree mask = NULL_TREE; |
12611 | |
12612 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
12613 | return false; |
12614 | |
12615 | if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)) |
12616 | return false; |
12617 | |
12618 | mask_type = vectype; |
12619 | nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
12620 | |
12621 | if (slp_node) |
12622 | ncopies = 1; |
12623 | else |
12624 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
12625 | |
12626 | gcc_assert (ncopies >= 1); |
12627 | |
12628 | if (TREE_CODE_CLASS (code) != tcc_comparison) |
12629 | return false; |
12630 | |
12631 | slp_tree slp_rhs1, slp_rhs2; |
12632 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
12633 | 0, &rhs1, &slp_rhs1, &dts[0], &vectype1)) |
12634 | return false; |
12635 | |
12636 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
12637 | 1, &rhs2, &slp_rhs2, &dts[1], &vectype2)) |
12638 | return false; |
12639 | |
12640 | if (vectype1 && vectype2 |
12641 | && maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: vectype1), |
12642 | b: TYPE_VECTOR_SUBPARTS (node: vectype2))) |
12643 | return false; |
12644 | |
12645 | vectype = vectype1 ? vectype1 : vectype2; |
12646 | |
12647 | /* Invariant comparison. */ |
12648 | if (!vectype) |
12649 | { |
12650 | if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (rhs1))) |
12651 | vectype = mask_type; |
12652 | else |
12653 | vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (rhs1), |
12654 | slp_node); |
12655 | if (!vectype || maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: vectype), b: nunits)) |
12656 | return false; |
12657 | } |
12658 | else if (maybe_ne (a: nunits, b: TYPE_VECTOR_SUBPARTS (node: vectype))) |
12659 | return false; |
12660 | |
12661 | /* Can't compare mask and non-mask types. */ |
12662 | if (vectype1 && vectype2 |
12663 | && (VECTOR_BOOLEAN_TYPE_P (vectype1) ^ VECTOR_BOOLEAN_TYPE_P (vectype2))) |
12664 | return false; |
12665 | |
12666 | /* Boolean values may have another representation in vectors |
12667 | and therefore we prefer bit operations over comparison for |
12668 | them (which also works for scalar masks). We store opcodes |
12669 | to use in bitop1 and bitop2. Statement is vectorized as |
12670 | BITOP2 (rhs1 BITOP1 rhs2) or |
12671 | rhs1 BITOP2 (BITOP1 rhs2) |
12672 | depending on bitop1 and bitop2 arity. */ |
12673 | bool swap_p = false; |
12674 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) |
12675 | { |
12676 | if (code == GT_EXPR) |
12677 | { |
12678 | bitop1 = BIT_NOT_EXPR; |
12679 | bitop2 = BIT_AND_EXPR; |
12680 | } |
12681 | else if (code == GE_EXPR) |
12682 | { |
12683 | bitop1 = BIT_NOT_EXPR; |
12684 | bitop2 = BIT_IOR_EXPR; |
12685 | } |
12686 | else if (code == LT_EXPR) |
12687 | { |
12688 | bitop1 = BIT_NOT_EXPR; |
12689 | bitop2 = BIT_AND_EXPR; |
12690 | swap_p = true; |
12691 | } |
12692 | else if (code == LE_EXPR) |
12693 | { |
12694 | bitop1 = BIT_NOT_EXPR; |
12695 | bitop2 = BIT_IOR_EXPR; |
12696 | swap_p = true; |
12697 | } |
12698 | else |
12699 | { |
12700 | bitop1 = BIT_XOR_EXPR; |
12701 | if (code == EQ_EXPR) |
12702 | bitop2 = BIT_NOT_EXPR; |
12703 | } |
12704 | } |
12705 | |
12706 | if (!vec_stmt) |
12707 | { |
12708 | if (bitop1 == NOP_EXPR) |
12709 | { |
12710 | if (!expand_vec_cmp_expr_p (vectype, mask_type, code)) |
12711 | return false; |
12712 | } |
12713 | else |
12714 | { |
12715 | machine_mode mode = TYPE_MODE (vectype); |
12716 | optab optab; |
12717 | |
12718 | optab = optab_for_tree_code (bitop1, vectype, optab_default); |
12719 | if (!optab || optab_handler (op: optab, mode) == CODE_FOR_nothing) |
12720 | return false; |
12721 | |
12722 | if (bitop2 != NOP_EXPR) |
12723 | { |
12724 | optab = optab_for_tree_code (bitop2, vectype, optab_default); |
12725 | if (!optab || optab_handler (op: optab, mode) == CODE_FOR_nothing) |
12726 | return false; |
12727 | } |
12728 | } |
12729 | |
12730 | /* Put types on constant and invariant SLP children. */ |
12731 | if (slp_node |
12732 | && (!vect_maybe_update_slp_op_vectype (slp_rhs1, vectype) |
12733 | || !vect_maybe_update_slp_op_vectype (slp_rhs2, vectype))) |
12734 | { |
12735 | if (dump_enabled_p ()) |
12736 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
12737 | "incompatible vector types for invariants\n" ); |
12738 | return false; |
12739 | } |
12740 | |
12741 | vect_model_simple_cost (vinfo, stmt_info, |
12742 | ncopies: ncopies * (1 + (bitop2 != NOP_EXPR)), |
12743 | dt: dts, ndts, node: slp_node, cost_vec); |
12744 | return true; |
12745 | } |
12746 | |
12747 | /* Transform. */ |
12748 | |
12749 | /* Handle def. */ |
12750 | lhs = gimple_get_lhs (STMT_VINFO_STMT (stmt_info)); |
12751 | if (lhs) |
12752 | mask = vect_create_destination_var (lhs, mask_type); |
12753 | |
12754 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, |
12755 | op0: rhs1, vectype0: vectype, vec_oprnds0: &vec_oprnds0, |
12756 | op1: rhs2, vectype1: vectype, vec_oprnds1: &vec_oprnds1); |
12757 | if (swap_p) |
12758 | std::swap (a&: vec_oprnds0, b&: vec_oprnds1); |
12759 | |
12760 | /* Arguments are ready. Create the new vector stmt. */ |
12761 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) |
12762 | { |
12763 | gimple *new_stmt; |
12764 | vec_rhs2 = vec_oprnds1[i]; |
12765 | |
12766 | if (lhs) |
12767 | new_temp = make_ssa_name (var: mask); |
12768 | else |
12769 | new_temp = make_temp_ssa_name (type: mask_type, NULL, name: "cmp" ); |
12770 | if (bitop1 == NOP_EXPR) |
12771 | { |
12772 | new_stmt = gimple_build_assign (new_temp, code, |
12773 | vec_rhs1, vec_rhs2); |
12774 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
12775 | } |
12776 | else |
12777 | { |
12778 | if (bitop1 == BIT_NOT_EXPR) |
12779 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs2); |
12780 | else |
12781 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs1, |
12782 | vec_rhs2); |
12783 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
12784 | if (bitop2 != NOP_EXPR) |
12785 | { |
12786 | tree res = make_ssa_name (var: mask); |
12787 | if (bitop2 == BIT_NOT_EXPR) |
12788 | new_stmt = gimple_build_assign (res, bitop2, new_temp); |
12789 | else |
12790 | new_stmt = gimple_build_assign (res, bitop2, vec_rhs1, |
12791 | new_temp); |
12792 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
12793 | } |
12794 | } |
12795 | if (slp_node) |
12796 | slp_node->push_vec_def (def: new_stmt); |
12797 | else |
12798 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
12799 | } |
12800 | |
12801 | if (!slp_node) |
12802 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
12803 | |
12804 | vec_oprnds0.release (); |
12805 | vec_oprnds1.release (); |
12806 | |
12807 | return true; |
12808 | } |
12809 | |
12810 | /* vectorizable_comparison. |
12811 | |
12812 | Check if STMT_INFO is comparison expression that can be vectorized. |
12813 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
12814 | comparison, put it in VEC_STMT, and insert it at GSI. |
12815 | |
12816 | Return true if STMT_INFO is vectorizable in this way. */ |
12817 | |
12818 | static bool |
12819 | vectorizable_comparison (vec_info *vinfo, |
12820 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
12821 | gimple **vec_stmt, |
12822 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
12823 | { |
12824 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
12825 | |
12826 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
12827 | return false; |
12828 | |
12829 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
12830 | return false; |
12831 | |
12832 | gassign *stmt = dyn_cast <gassign *> (p: stmt_info->stmt); |
12833 | if (!stmt) |
12834 | return false; |
12835 | |
12836 | enum tree_code code = gimple_assign_rhs_code (gs: stmt); |
12837 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
12838 | if (!vectorizable_comparison_1 (vinfo, vectype, stmt_info, code, gsi, |
12839 | vec_stmt, slp_node, cost_vec)) |
12840 | return false; |
12841 | |
12842 | if (!vec_stmt) |
12843 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; |
12844 | |
12845 | return true; |
12846 | } |
12847 | |
12848 | /* Check to see if the current early break given in STMT_INFO is valid for |
12849 | vectorization. */ |
12850 | |
12851 | static bool |
12852 | vectorizable_early_exit (vec_info *vinfo, stmt_vec_info stmt_info, |
12853 | gimple_stmt_iterator *gsi, gimple **vec_stmt, |
12854 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
12855 | { |
12856 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
12857 | if (!loop_vinfo |
12858 | || !is_a <gcond *> (STMT_VINFO_STMT (stmt_info))) |
12859 | return false; |
12860 | |
12861 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_condition_def) |
12862 | return false; |
12863 | |
12864 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
12865 | return false; |
12866 | |
12867 | DUMP_VECT_SCOPE ("vectorizable_early_exit" ); |
12868 | |
12869 | auto code = gimple_cond_code (STMT_VINFO_STMT (stmt_info)); |
12870 | |
12871 | tree vectype = NULL_TREE; |
12872 | slp_tree slp_op0; |
12873 | tree op0; |
12874 | enum vect_def_type dt0; |
12875 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 0, &op0, &slp_op0, &dt0, |
12876 | &vectype)) |
12877 | { |
12878 | if (dump_enabled_p ()) |
12879 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
12880 | "use not simple.\n" ); |
12881 | return false; |
12882 | } |
12883 | |
12884 | if (!vectype) |
12885 | return false; |
12886 | |
12887 | machine_mode mode = TYPE_MODE (vectype); |
12888 | int ncopies; |
12889 | |
12890 | if (slp_node) |
12891 | ncopies = 1; |
12892 | else |
12893 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
12894 | |
12895 | vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); |
12896 | bool masked_loop_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo); |
12897 | |
12898 | /* Now build the new conditional. Pattern gimple_conds get dropped during |
12899 | codegen so we must replace the original insn. */ |
12900 | gimple *orig_stmt = STMT_VINFO_STMT (vect_orig_stmt (stmt_info)); |
12901 | gcond *cond_stmt = as_a <gcond *>(p: orig_stmt); |
12902 | /* When vectorizing we assume that if the branch edge is taken that we're |
12903 | exiting the loop. This is not however always the case as the compiler will |
12904 | rewrite conditions to always be a comparison against 0. To do this it |
12905 | sometimes flips the edges. This is fine for scalar, but for vector we |
12906 | then have to flip the test, as we're still assuming that if you take the |
12907 | branch edge that we found the exit condition. i.e. we need to know whether |
12908 | we are generating a `forall` or an `exist` condition. */ |
12909 | auto new_code = NE_EXPR; |
12910 | auto reduc_optab = ior_optab; |
12911 | auto reduc_op = BIT_IOR_EXPR; |
12912 | tree cst = build_zero_cst (vectype); |
12913 | edge exit_true_edge = EDGE_SUCC (gimple_bb (cond_stmt), 0); |
12914 | if (exit_true_edge->flags & EDGE_FALSE_VALUE) |
12915 | exit_true_edge = EDGE_SUCC (gimple_bb (cond_stmt), 1); |
12916 | gcc_assert (exit_true_edge->flags & EDGE_TRUE_VALUE); |
12917 | if (flow_bb_inside_loop_p (LOOP_VINFO_LOOP (loop_vinfo), |
12918 | exit_true_edge->dest)) |
12919 | { |
12920 | new_code = EQ_EXPR; |
12921 | reduc_optab = and_optab; |
12922 | reduc_op = BIT_AND_EXPR; |
12923 | cst = build_minus_one_cst (vectype); |
12924 | } |
12925 | |
12926 | /* Analyze only. */ |
12927 | if (!vec_stmt) |
12928 | { |
12929 | if (direct_optab_handler (op: cbranch_optab, mode) == CODE_FOR_nothing) |
12930 | { |
12931 | if (dump_enabled_p ()) |
12932 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
12933 | "can't vectorize early exit because the " |
12934 | "target doesn't support flag setting vector " |
12935 | "comparisons.\n" ); |
12936 | return false; |
12937 | } |
12938 | |
12939 | if (ncopies > 1 |
12940 | && direct_optab_handler (op: reduc_optab, mode) == CODE_FOR_nothing) |
12941 | { |
12942 | if (dump_enabled_p ()) |
12943 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
12944 | "can't vectorize early exit because the " |
12945 | "target does not support boolean vector %s " |
12946 | "for type %T.\n" , |
12947 | reduc_optab == ior_optab ? "OR" : "AND" , |
12948 | vectype); |
12949 | return false; |
12950 | } |
12951 | |
12952 | if (!vectorizable_comparison_1 (vinfo, vectype, stmt_info, code, gsi, |
12953 | vec_stmt, slp_node, cost_vec)) |
12954 | return false; |
12955 | |
12956 | if (LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) |
12957 | { |
12958 | if (direct_internal_fn_supported_p (IFN_VCOND_MASK_LEN, vectype, |
12959 | OPTIMIZE_FOR_SPEED)) |
12960 | return false; |
12961 | else |
12962 | vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype, NULL); |
12963 | } |
12964 | |
12965 | |
12966 | return true; |
12967 | } |
12968 | |
12969 | /* Tranform. */ |
12970 | |
12971 | tree new_temp = NULL_TREE; |
12972 | gimple *new_stmt = NULL; |
12973 | |
12974 | if (dump_enabled_p ()) |
12975 | dump_printf_loc (MSG_NOTE, vect_location, "transform early-exit.\n" ); |
12976 | |
12977 | if (!vectorizable_comparison_1 (vinfo, vectype, stmt_info, code, gsi, |
12978 | vec_stmt, slp_node, cost_vec)) |
12979 | gcc_unreachable (); |
12980 | |
12981 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
12982 | basic_block cond_bb = gimple_bb (g: stmt); |
12983 | gimple_stmt_iterator cond_gsi = gsi_last_bb (bb: cond_bb); |
12984 | |
12985 | auto_vec<tree> stmts; |
12986 | |
12987 | if (slp_node) |
12988 | stmts.safe_splice (SLP_TREE_VEC_DEFS (slp_node)); |
12989 | else |
12990 | { |
12991 | auto vec_stmts = STMT_VINFO_VEC_STMTS (stmt_info); |
12992 | stmts.reserve_exact (nelems: vec_stmts.length ()); |
12993 | for (auto stmt : vec_stmts) |
12994 | stmts.quick_push (obj: gimple_assign_lhs (gs: stmt)); |
12995 | } |
12996 | |
12997 | /* Determine if we need to reduce the final value. */ |
12998 | if (stmts.length () > 1) |
12999 | { |
13000 | /* We build the reductions in a way to maintain as much parallelism as |
13001 | possible. */ |
13002 | auto_vec<tree> workset (stmts.length ()); |
13003 | |
13004 | /* Mask the statements as we queue them up. Normally we loop over |
13005 | vec_num, but since we inspect the exact results of vectorization |
13006 | we don't need to and instead can just use the stmts themselves. */ |
13007 | if (masked_loop_p) |
13008 | for (unsigned i = 0; i < stmts.length (); i++) |
13009 | { |
13010 | tree stmt_mask |
13011 | = vect_get_loop_mask (loop_vinfo, gsi, masks, ncopies, vectype, |
13012 | i); |
13013 | stmt_mask |
13014 | = prepare_vec_mask (loop_vinfo, TREE_TYPE (stmt_mask), loop_mask: stmt_mask, |
13015 | vec_mask: stmts[i], gsi: &cond_gsi); |
13016 | workset.quick_push (obj: stmt_mask); |
13017 | } |
13018 | else |
13019 | workset.splice (src: stmts); |
13020 | |
13021 | while (workset.length () > 1) |
13022 | { |
13023 | new_temp = make_temp_ssa_name (type: vectype, NULL, name: "vexit_reduc" ); |
13024 | tree arg0 = workset.pop (); |
13025 | tree arg1 = workset.pop (); |
13026 | new_stmt = gimple_build_assign (new_temp, reduc_op, arg0, arg1); |
13027 | vect_finish_stmt_generation (vinfo: loop_vinfo, stmt_info, vec_stmt: new_stmt, |
13028 | gsi: &cond_gsi); |
13029 | workset.quick_insert (ix: 0, obj: new_temp); |
13030 | } |
13031 | } |
13032 | else |
13033 | { |
13034 | new_temp = stmts[0]; |
13035 | if (masked_loop_p) |
13036 | { |
13037 | tree mask |
13038 | = vect_get_loop_mask (loop_vinfo, gsi, masks, ncopies, vectype, 0); |
13039 | new_temp = prepare_vec_mask (loop_vinfo, TREE_TYPE (mask), loop_mask: mask, |
13040 | vec_mask: new_temp, gsi: &cond_gsi); |
13041 | } |
13042 | } |
13043 | |
13044 | gcc_assert (new_temp); |
13045 | |
13046 | gimple_cond_set_condition (stmt: cond_stmt, code: new_code, lhs: new_temp, rhs: cst); |
13047 | update_stmt (s: orig_stmt); |
13048 | |
13049 | if (slp_node) |
13050 | SLP_TREE_VEC_DEFS (slp_node).truncate (size: 0); |
13051 | else |
13052 | STMT_VINFO_VEC_STMTS (stmt_info).truncate (size: 0); |
13053 | |
13054 | if (!slp_node) |
13055 | *vec_stmt = orig_stmt; |
13056 | |
13057 | return true; |
13058 | } |
13059 | |
13060 | /* If SLP_NODE is nonnull, return true if vectorizable_live_operation |
13061 | can handle all live statements in the node. Otherwise return true |
13062 | if STMT_INFO is not live or if vectorizable_live_operation can handle it. |
13063 | VEC_STMT_P is as for vectorizable_live_operation. */ |
13064 | |
13065 | static bool |
13066 | can_vectorize_live_stmts (vec_info *vinfo, stmt_vec_info stmt_info, |
13067 | slp_tree slp_node, slp_instance slp_node_instance, |
13068 | bool vec_stmt_p, |
13069 | stmt_vector_for_cost *cost_vec) |
13070 | { |
13071 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
13072 | if (slp_node) |
13073 | { |
13074 | stmt_vec_info slp_stmt_info; |
13075 | unsigned int i; |
13076 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, slp_stmt_info) |
13077 | { |
13078 | if ((STMT_VINFO_LIVE_P (slp_stmt_info) |
13079 | || (loop_vinfo |
13080 | && LOOP_VINFO_EARLY_BREAKS (loop_vinfo) |
13081 | && STMT_VINFO_DEF_TYPE (slp_stmt_info) |
13082 | == vect_induction_def)) |
13083 | && !vectorizable_live_operation (vinfo, slp_stmt_info, slp_node, |
13084 | slp_node_instance, i, |
13085 | vec_stmt_p, cost_vec)) |
13086 | return false; |
13087 | } |
13088 | } |
13089 | else if ((STMT_VINFO_LIVE_P (stmt_info) |
13090 | || (LOOP_VINFO_EARLY_BREAKS (loop_vinfo) |
13091 | && STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def)) |
13092 | && !vectorizable_live_operation (vinfo, stmt_info, |
13093 | slp_node, slp_node_instance, -1, |
13094 | vec_stmt_p, cost_vec)) |
13095 | return false; |
13096 | |
13097 | return true; |
13098 | } |
13099 | |
13100 | /* Make sure the statement is vectorizable. */ |
13101 | |
13102 | opt_result |
13103 | vect_analyze_stmt (vec_info *vinfo, |
13104 | stmt_vec_info stmt_info, bool *need_to_vectorize, |
13105 | slp_tree node, slp_instance node_instance, |
13106 | stmt_vector_for_cost *cost_vec) |
13107 | { |
13108 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
13109 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
13110 | bool ok; |
13111 | gimple_seq pattern_def_seq; |
13112 | |
13113 | if (dump_enabled_p ()) |
13114 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: %G" , |
13115 | stmt_info->stmt); |
13116 | |
13117 | if (gimple_has_volatile_ops (stmt: stmt_info->stmt)) |
13118 | return opt_result::failure_at (loc: stmt_info->stmt, |
13119 | fmt: "not vectorized:" |
13120 | " stmt has volatile operands: %G\n" , |
13121 | stmt_info->stmt); |
13122 | |
13123 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
13124 | && node == NULL |
13125 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
13126 | { |
13127 | gimple_stmt_iterator si; |
13128 | |
13129 | for (si = gsi_start (seq&: pattern_def_seq); !gsi_end_p (i: si); gsi_next (i: &si)) |
13130 | { |
13131 | stmt_vec_info pattern_def_stmt_info |
13132 | = vinfo->lookup_stmt (gsi_stmt (i: si)); |
13133 | if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info) |
13134 | || STMT_VINFO_LIVE_P (pattern_def_stmt_info)) |
13135 | { |
13136 | /* Analyze def stmt of STMT if it's a pattern stmt. */ |
13137 | if (dump_enabled_p ()) |
13138 | dump_printf_loc (MSG_NOTE, vect_location, |
13139 | "==> examining pattern def statement: %G" , |
13140 | pattern_def_stmt_info->stmt); |
13141 | |
13142 | opt_result res |
13143 | = vect_analyze_stmt (vinfo, stmt_info: pattern_def_stmt_info, |
13144 | need_to_vectorize, node, node_instance, |
13145 | cost_vec); |
13146 | if (!res) |
13147 | return res; |
13148 | } |
13149 | } |
13150 | } |
13151 | |
13152 | /* Skip stmts that do not need to be vectorized. In loops this is expected |
13153 | to include: |
13154 | - the COND_EXPR which is the loop exit condition |
13155 | - any LABEL_EXPRs in the loop |
13156 | - computations that are used only for array indexing or loop control. |
13157 | In basic blocks we only analyze statements that are a part of some SLP |
13158 | instance, therefore, all the statements are relevant. |
13159 | |
13160 | Pattern statement needs to be analyzed instead of the original statement |
13161 | if the original statement is not relevant. Otherwise, we analyze both |
13162 | statements. In basic blocks we are called from some SLP instance |
13163 | traversal, don't analyze pattern stmts instead, the pattern stmts |
13164 | already will be part of SLP instance. */ |
13165 | |
13166 | stmt_vec_info pattern_stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); |
13167 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
13168 | && !STMT_VINFO_LIVE_P (stmt_info)) |
13169 | { |
13170 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
13171 | && pattern_stmt_info |
13172 | && (STMT_VINFO_RELEVANT_P (pattern_stmt_info) |
13173 | || STMT_VINFO_LIVE_P (pattern_stmt_info))) |
13174 | { |
13175 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
13176 | stmt_info = pattern_stmt_info; |
13177 | if (dump_enabled_p ()) |
13178 | dump_printf_loc (MSG_NOTE, vect_location, |
13179 | "==> examining pattern statement: %G" , |
13180 | stmt_info->stmt); |
13181 | } |
13182 | else |
13183 | { |
13184 | if (dump_enabled_p ()) |
13185 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n" ); |
13186 | |
13187 | return opt_result::success (); |
13188 | } |
13189 | } |
13190 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
13191 | && node == NULL |
13192 | && pattern_stmt_info |
13193 | && (STMT_VINFO_RELEVANT_P (pattern_stmt_info) |
13194 | || STMT_VINFO_LIVE_P (pattern_stmt_info))) |
13195 | { |
13196 | /* Analyze PATTERN_STMT too. */ |
13197 | if (dump_enabled_p ()) |
13198 | dump_printf_loc (MSG_NOTE, vect_location, |
13199 | "==> examining pattern statement: %G" , |
13200 | pattern_stmt_info->stmt); |
13201 | |
13202 | opt_result res |
13203 | = vect_analyze_stmt (vinfo, stmt_info: pattern_stmt_info, need_to_vectorize, node, |
13204 | node_instance, cost_vec); |
13205 | if (!res) |
13206 | return res; |
13207 | } |
13208 | |
13209 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
13210 | { |
13211 | case vect_internal_def: |
13212 | case vect_condition_def: |
13213 | break; |
13214 | |
13215 | case vect_reduction_def: |
13216 | case vect_nested_cycle: |
13217 | gcc_assert (!bb_vinfo |
13218 | && (relevance == vect_used_in_outer |
13219 | || relevance == vect_used_in_outer_by_reduction |
13220 | || relevance == vect_used_by_reduction |
13221 | || relevance == vect_unused_in_scope |
13222 | || relevance == vect_used_only_live)); |
13223 | break; |
13224 | |
13225 | case vect_induction_def: |
13226 | case vect_first_order_recurrence: |
13227 | gcc_assert (!bb_vinfo); |
13228 | break; |
13229 | |
13230 | case vect_constant_def: |
13231 | case vect_external_def: |
13232 | case vect_unknown_def_type: |
13233 | default: |
13234 | gcc_unreachable (); |
13235 | } |
13236 | |
13237 | tree saved_vectype = STMT_VINFO_VECTYPE (stmt_info); |
13238 | if (node) |
13239 | STMT_VINFO_VECTYPE (stmt_info) = SLP_TREE_VECTYPE (node); |
13240 | |
13241 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
13242 | { |
13243 | gcall *call = dyn_cast <gcall *> (p: stmt_info->stmt); |
13244 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
13245 | || gimple_code (stmt_info->stmt) == GIMPLE_COND |
13246 | || (call && gimple_call_lhs (call) == NULL_TREE)); |
13247 | *need_to_vectorize = true; |
13248 | } |
13249 | |
13250 | if (PURE_SLP_STMT (stmt_info) && !node) |
13251 | { |
13252 | if (dump_enabled_p ()) |
13253 | dump_printf_loc (MSG_NOTE, vect_location, |
13254 | "handled only by SLP analysis\n" ); |
13255 | return opt_result::success (); |
13256 | } |
13257 | |
13258 | ok = true; |
13259 | if (!bb_vinfo |
13260 | && (STMT_VINFO_RELEVANT_P (stmt_info) |
13261 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) |
13262 | /* Prefer vectorizable_call over vectorizable_simd_clone_call so |
13263 | -mveclibabi= takes preference over library functions with |
13264 | the simd attribute. */ |
13265 | ok = (vectorizable_call (vinfo, stmt_info, NULL, NULL, slp_node: node, cost_vec) |
13266 | || vectorizable_simd_clone_call (vinfo, stmt_info, NULL, NULL, slp_node: node, |
13267 | cost_vec) |
13268 | || vectorizable_conversion (vinfo, stmt_info, |
13269 | NULL, NULL, slp_node: node, cost_vec) |
13270 | || vectorizable_operation (vinfo, stmt_info, |
13271 | NULL, NULL, slp_node: node, cost_vec) |
13272 | || vectorizable_assignment (vinfo, stmt_info, |
13273 | NULL, NULL, slp_node: node, cost_vec) |
13274 | || vectorizable_load (vinfo, stmt_info, NULL, NULL, slp_node: node, cost_vec) |
13275 | || vectorizable_store (vinfo, stmt_info, NULL, NULL, slp_node: node, cost_vec) |
13276 | || vectorizable_reduction (as_a <loop_vec_info> (p: vinfo), stmt_info, |
13277 | node, node_instance, cost_vec) |
13278 | || vectorizable_induction (as_a <loop_vec_info> (p: vinfo), stmt_info, |
13279 | NULL, node, cost_vec) |
13280 | || vectorizable_shift (vinfo, stmt_info, NULL, NULL, slp_node: node, cost_vec) |
13281 | || vectorizable_condition (vinfo, stmt_info, |
13282 | NULL, NULL, slp_node: node, cost_vec) |
13283 | || vectorizable_comparison (vinfo, stmt_info, NULL, NULL, slp_node: node, |
13284 | cost_vec) |
13285 | || vectorizable_lc_phi (as_a <loop_vec_info> (p: vinfo), |
13286 | stmt_info, NULL, node) |
13287 | || vectorizable_recurr (as_a <loop_vec_info> (p: vinfo), |
13288 | stmt_info, NULL, node, cost_vec) |
13289 | || vectorizable_early_exit (vinfo, stmt_info, NULL, NULL, slp_node: node, |
13290 | cost_vec)); |
13291 | else |
13292 | { |
13293 | if (bb_vinfo) |
13294 | ok = (vectorizable_call (vinfo, stmt_info, NULL, NULL, slp_node: node, cost_vec) |
13295 | || vectorizable_simd_clone_call (vinfo, stmt_info, |
13296 | NULL, NULL, slp_node: node, cost_vec) |
13297 | || vectorizable_conversion (vinfo, stmt_info, NULL, NULL, slp_node: node, |
13298 | cost_vec) |
13299 | || vectorizable_shift (vinfo, stmt_info, |
13300 | NULL, NULL, slp_node: node, cost_vec) |
13301 | || vectorizable_operation (vinfo, stmt_info, |
13302 | NULL, NULL, slp_node: node, cost_vec) |
13303 | || vectorizable_assignment (vinfo, stmt_info, NULL, NULL, slp_node: node, |
13304 | cost_vec) |
13305 | || vectorizable_load (vinfo, stmt_info, |
13306 | NULL, NULL, slp_node: node, cost_vec) |
13307 | || vectorizable_store (vinfo, stmt_info, |
13308 | NULL, NULL, slp_node: node, cost_vec) |
13309 | || vectorizable_condition (vinfo, stmt_info, |
13310 | NULL, NULL, slp_node: node, cost_vec) |
13311 | || vectorizable_comparison (vinfo, stmt_info, NULL, NULL, slp_node: node, |
13312 | cost_vec) |
13313 | || vectorizable_phi (vinfo, stmt_info, NULL, node, cost_vec) |
13314 | || vectorizable_early_exit (vinfo, stmt_info, NULL, NULL, slp_node: node, |
13315 | cost_vec)); |
13316 | |
13317 | } |
13318 | |
13319 | if (node) |
13320 | STMT_VINFO_VECTYPE (stmt_info) = saved_vectype; |
13321 | |
13322 | if (!ok) |
13323 | return opt_result::failure_at (loc: stmt_info->stmt, |
13324 | fmt: "not vectorized:" |
13325 | " relevant stmt not supported: %G" , |
13326 | stmt_info->stmt); |
13327 | |
13328 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
13329 | need extra handling, except for vectorizable reductions. */ |
13330 | if (!bb_vinfo |
13331 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
13332 | && STMT_VINFO_TYPE (stmt_info) != lc_phi_info_type |
13333 | && !can_vectorize_live_stmts (vinfo: as_a <loop_vec_info> (p: vinfo), |
13334 | stmt_info, slp_node: node, slp_node_instance: node_instance, |
13335 | vec_stmt_p: false, cost_vec)) |
13336 | return opt_result::failure_at (loc: stmt_info->stmt, |
13337 | fmt: "not vectorized:" |
13338 | " live stmt not supported: %G" , |
13339 | stmt_info->stmt); |
13340 | |
13341 | return opt_result::success (); |
13342 | } |
13343 | |
13344 | |
13345 | /* Function vect_transform_stmt. |
13346 | |
13347 | Create a vectorized stmt to replace STMT_INFO, and insert it at GSI. */ |
13348 | |
13349 | bool |
13350 | vect_transform_stmt (vec_info *vinfo, |
13351 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
13352 | slp_tree slp_node, slp_instance slp_node_instance) |
13353 | { |
13354 | bool is_store = false; |
13355 | gimple *vec_stmt = NULL; |
13356 | bool done; |
13357 | |
13358 | gcc_assert (slp_node || !PURE_SLP_STMT (stmt_info)); |
13359 | |
13360 | tree saved_vectype = STMT_VINFO_VECTYPE (stmt_info); |
13361 | if (slp_node) |
13362 | STMT_VINFO_VECTYPE (stmt_info) = SLP_TREE_VECTYPE (slp_node); |
13363 | |
13364 | switch (STMT_VINFO_TYPE (stmt_info)) |
13365 | { |
13366 | case type_demotion_vec_info_type: |
13367 | case type_promotion_vec_info_type: |
13368 | case type_conversion_vec_info_type: |
13369 | done = vectorizable_conversion (vinfo, stmt_info, |
13370 | gsi, vec_stmt: &vec_stmt, slp_node, NULL); |
13371 | gcc_assert (done); |
13372 | break; |
13373 | |
13374 | case induc_vec_info_type: |
13375 | done = vectorizable_induction (as_a <loop_vec_info> (p: vinfo), |
13376 | stmt_info, &vec_stmt, slp_node, |
13377 | NULL); |
13378 | gcc_assert (done); |
13379 | break; |
13380 | |
13381 | case shift_vec_info_type: |
13382 | done = vectorizable_shift (vinfo, stmt_info, |
13383 | gsi, vec_stmt: &vec_stmt, slp_node, NULL); |
13384 | gcc_assert (done); |
13385 | break; |
13386 | |
13387 | case op_vec_info_type: |
13388 | done = vectorizable_operation (vinfo, stmt_info, gsi, vec_stmt: &vec_stmt, slp_node, |
13389 | NULL); |
13390 | gcc_assert (done); |
13391 | break; |
13392 | |
13393 | case assignment_vec_info_type: |
13394 | done = vectorizable_assignment (vinfo, stmt_info, |
13395 | gsi, vec_stmt: &vec_stmt, slp_node, NULL); |
13396 | gcc_assert (done); |
13397 | break; |
13398 | |
13399 | case load_vec_info_type: |
13400 | done = vectorizable_load (vinfo, stmt_info, gsi, vec_stmt: &vec_stmt, slp_node, |
13401 | NULL); |
13402 | gcc_assert (done); |
13403 | break; |
13404 | |
13405 | case store_vec_info_type: |
13406 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) |
13407 | && !slp_node |
13408 | && (++DR_GROUP_STORE_COUNT (DR_GROUP_FIRST_ELEMENT (stmt_info)) |
13409 | < DR_GROUP_SIZE (DR_GROUP_FIRST_ELEMENT (stmt_info)))) |
13410 | /* In case of interleaving, the whole chain is vectorized when the |
13411 | last store in the chain is reached. Store stmts before the last |
13412 | one are skipped, and there vec_stmt_info shouldn't be freed |
13413 | meanwhile. */ |
13414 | ; |
13415 | else |
13416 | { |
13417 | done = vectorizable_store (vinfo, stmt_info, |
13418 | gsi, vec_stmt: &vec_stmt, slp_node, NULL); |
13419 | gcc_assert (done); |
13420 | is_store = true; |
13421 | } |
13422 | break; |
13423 | |
13424 | case condition_vec_info_type: |
13425 | done = vectorizable_condition (vinfo, stmt_info, |
13426 | gsi, vec_stmt: &vec_stmt, slp_node, NULL); |
13427 | gcc_assert (done); |
13428 | break; |
13429 | |
13430 | case comparison_vec_info_type: |
13431 | done = vectorizable_comparison (vinfo, stmt_info, gsi, vec_stmt: &vec_stmt, |
13432 | slp_node, NULL); |
13433 | gcc_assert (done); |
13434 | break; |
13435 | |
13436 | case call_vec_info_type: |
13437 | done = vectorizable_call (vinfo, stmt_info, |
13438 | gsi, vec_stmt: &vec_stmt, slp_node, NULL); |
13439 | break; |
13440 | |
13441 | case call_simd_clone_vec_info_type: |
13442 | done = vectorizable_simd_clone_call (vinfo, stmt_info, gsi, vec_stmt: &vec_stmt, |
13443 | slp_node, NULL); |
13444 | break; |
13445 | |
13446 | case reduc_vec_info_type: |
13447 | done = vect_transform_reduction (as_a <loop_vec_info> (p: vinfo), stmt_info, |
13448 | gsi, &vec_stmt, slp_node); |
13449 | gcc_assert (done); |
13450 | break; |
13451 | |
13452 | case cycle_phi_info_type: |
13453 | done = vect_transform_cycle_phi (as_a <loop_vec_info> (p: vinfo), stmt_info, |
13454 | &vec_stmt, slp_node, slp_node_instance); |
13455 | gcc_assert (done); |
13456 | break; |
13457 | |
13458 | case lc_phi_info_type: |
13459 | done = vectorizable_lc_phi (as_a <loop_vec_info> (p: vinfo), |
13460 | stmt_info, &vec_stmt, slp_node); |
13461 | gcc_assert (done); |
13462 | break; |
13463 | |
13464 | case recurr_info_type: |
13465 | done = vectorizable_recurr (as_a <loop_vec_info> (p: vinfo), |
13466 | stmt_info, &vec_stmt, slp_node, NULL); |
13467 | gcc_assert (done); |
13468 | break; |
13469 | |
13470 | case phi_info_type: |
13471 | done = vectorizable_phi (vinfo, stmt_info, &vec_stmt, slp_node, NULL); |
13472 | gcc_assert (done); |
13473 | break; |
13474 | |
13475 | case loop_exit_ctrl_vec_info_type: |
13476 | done = vectorizable_early_exit (vinfo, stmt_info, gsi, vec_stmt: &vec_stmt, |
13477 | slp_node, NULL); |
13478 | gcc_assert (done); |
13479 | break; |
13480 | |
13481 | default: |
13482 | if (!STMT_VINFO_LIVE_P (stmt_info)) |
13483 | { |
13484 | if (dump_enabled_p ()) |
13485 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
13486 | "stmt not supported.\n" ); |
13487 | gcc_unreachable (); |
13488 | } |
13489 | done = true; |
13490 | } |
13491 | |
13492 | if (!slp_node && vec_stmt) |
13493 | gcc_assert (STMT_VINFO_VEC_STMTS (stmt_info).exists ()); |
13494 | |
13495 | if (STMT_VINFO_TYPE (stmt_info) != store_vec_info_type) |
13496 | { |
13497 | /* Handle stmts whose DEF is used outside the loop-nest that is |
13498 | being vectorized. */ |
13499 | done = can_vectorize_live_stmts (vinfo, stmt_info, slp_node, |
13500 | slp_node_instance, vec_stmt_p: true, NULL); |
13501 | gcc_assert (done); |
13502 | } |
13503 | |
13504 | if (slp_node) |
13505 | STMT_VINFO_VECTYPE (stmt_info) = saved_vectype; |
13506 | |
13507 | return is_store; |
13508 | } |
13509 | |
13510 | |
13511 | /* Remove a group of stores (for SLP or interleaving), free their |
13512 | stmt_vec_info. */ |
13513 | |
13514 | void |
13515 | vect_remove_stores (vec_info *vinfo, stmt_vec_info first_stmt_info) |
13516 | { |
13517 | stmt_vec_info next_stmt_info = first_stmt_info; |
13518 | |
13519 | while (next_stmt_info) |
13520 | { |
13521 | stmt_vec_info tmp = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
13522 | next_stmt_info = vect_orig_stmt (stmt_info: next_stmt_info); |
13523 | /* Free the attached stmt_vec_info and remove the stmt. */ |
13524 | vinfo->remove_stmt (next_stmt_info); |
13525 | next_stmt_info = tmp; |
13526 | } |
13527 | } |
13528 | |
13529 | /* If NUNITS is nonzero, return a vector type that contains NUNITS |
13530 | elements of type SCALAR_TYPE, or null if the target doesn't support |
13531 | such a type. |
13532 | |
13533 | If NUNITS is zero, return a vector type that contains elements of |
13534 | type SCALAR_TYPE, choosing whichever vector size the target prefers. |
13535 | |
13536 | If PREVAILING_MODE is VOIDmode, we have not yet chosen a vector mode |
13537 | for this vectorization region and want to "autodetect" the best choice. |
13538 | Otherwise, PREVAILING_MODE is a previously-chosen vector TYPE_MODE |
13539 | and we want the new type to be interoperable with it. PREVAILING_MODE |
13540 | in this case can be a scalar integer mode or a vector mode; when it |
13541 | is a vector mode, the function acts like a tree-level version of |
13542 | related_vector_mode. */ |
13543 | |
13544 | tree |
13545 | get_related_vectype_for_scalar_type (machine_mode prevailing_mode, |
13546 | tree scalar_type, poly_uint64 nunits) |
13547 | { |
13548 | tree orig_scalar_type = scalar_type; |
13549 | scalar_mode inner_mode; |
13550 | machine_mode simd_mode; |
13551 | tree vectype; |
13552 | |
13553 | if ((!INTEGRAL_TYPE_P (scalar_type) |
13554 | && !POINTER_TYPE_P (scalar_type) |
13555 | && !SCALAR_FLOAT_TYPE_P (scalar_type)) |
13556 | || (!is_int_mode (TYPE_MODE (scalar_type), int_mode: &inner_mode) |
13557 | && !is_float_mode (TYPE_MODE (scalar_type), float_mode: &inner_mode))) |
13558 | return NULL_TREE; |
13559 | |
13560 | unsigned int nbytes = GET_MODE_SIZE (mode: inner_mode); |
13561 | |
13562 | /* Interoperability between modes requires one to be a constant multiple |
13563 | of the other, so that the number of vectors required for each operation |
13564 | is a compile-time constant. */ |
13565 | if (prevailing_mode != VOIDmode |
13566 | && !constant_multiple_p (a: nunits * nbytes, |
13567 | b: GET_MODE_SIZE (mode: prevailing_mode)) |
13568 | && !constant_multiple_p (a: GET_MODE_SIZE (mode: prevailing_mode), |
13569 | b: nunits * nbytes)) |
13570 | return NULL_TREE; |
13571 | |
13572 | /* For vector types of elements whose mode precision doesn't |
13573 | match their types precision we use a element type of mode |
13574 | precision. The vectorization routines will have to make sure |
13575 | they support the proper result truncation/extension. |
13576 | We also make sure to build vector types with INTEGER_TYPE |
13577 | component type only. */ |
13578 | if (INTEGRAL_TYPE_P (scalar_type) |
13579 | && (GET_MODE_BITSIZE (mode: inner_mode) != TYPE_PRECISION (scalar_type) |
13580 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) |
13581 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (mode: inner_mode), |
13582 | TYPE_UNSIGNED (scalar_type)); |
13583 | |
13584 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
13585 | When the component mode passes the above test simply use a type |
13586 | corresponding to that mode. The theory is that any use that |
13587 | would cause problems with this will disable vectorization anyway. */ |
13588 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
13589 | && !INTEGRAL_TYPE_P (scalar_type)) |
13590 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
13591 | |
13592 | /* We can't build a vector type of elements with alignment bigger than |
13593 | their size. */ |
13594 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
13595 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
13596 | TYPE_UNSIGNED (scalar_type)); |
13597 | |
13598 | /* If we felt back to using the mode fail if there was |
13599 | no scalar type for it. */ |
13600 | if (scalar_type == NULL_TREE) |
13601 | return NULL_TREE; |
13602 | |
13603 | /* If no prevailing mode was supplied, use the mode the target prefers. |
13604 | Otherwise lookup a vector mode based on the prevailing mode. */ |
13605 | if (prevailing_mode == VOIDmode) |
13606 | { |
13607 | gcc_assert (known_eq (nunits, 0U)); |
13608 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); |
13609 | if (SCALAR_INT_MODE_P (simd_mode)) |
13610 | { |
13611 | /* Traditional behavior is not to take the integer mode |
13612 | literally, but simply to use it as a way of determining |
13613 | the vector size. It is up to mode_for_vector to decide |
13614 | what the TYPE_MODE should be. |
13615 | |
13616 | Note that nunits == 1 is allowed in order to support single |
13617 | element vector types. */ |
13618 | if (!multiple_p (a: GET_MODE_SIZE (mode: simd_mode), b: nbytes, multiple: &nunits) |
13619 | || !mode_for_vector (inner_mode, nunits).exists (mode: &simd_mode)) |
13620 | return NULL_TREE; |
13621 | } |
13622 | } |
13623 | else if (SCALAR_INT_MODE_P (prevailing_mode) |
13624 | || !related_vector_mode (prevailing_mode, |
13625 | inner_mode, nunits).exists (mode: &simd_mode)) |
13626 | { |
13627 | /* Fall back to using mode_for_vector, mostly in the hope of being |
13628 | able to use an integer mode. */ |
13629 | if (known_eq (nunits, 0U) |
13630 | && !multiple_p (a: GET_MODE_SIZE (mode: prevailing_mode), b: nbytes, multiple: &nunits)) |
13631 | return NULL_TREE; |
13632 | |
13633 | if (!mode_for_vector (inner_mode, nunits).exists (mode: &simd_mode)) |
13634 | return NULL_TREE; |
13635 | } |
13636 | |
13637 | vectype = build_vector_type_for_mode (scalar_type, simd_mode); |
13638 | |
13639 | /* In cases where the mode was chosen by mode_for_vector, check that |
13640 | the target actually supports the chosen mode, or that it at least |
13641 | allows the vector mode to be replaced by a like-sized integer. */ |
13642 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) |
13643 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) |
13644 | return NULL_TREE; |
13645 | |
13646 | /* Re-attach the address-space qualifier if we canonicalized the scalar |
13647 | type. */ |
13648 | if (TYPE_ADDR_SPACE (orig_scalar_type) != TYPE_ADDR_SPACE (vectype)) |
13649 | return build_qualified_type |
13650 | (vectype, KEEP_QUAL_ADDR_SPACE (TYPE_QUALS (orig_scalar_type))); |
13651 | |
13652 | return vectype; |
13653 | } |
13654 | |
13655 | /* Function get_vectype_for_scalar_type. |
13656 | |
13657 | Returns the vector type corresponding to SCALAR_TYPE as supported |
13658 | by the target. If GROUP_SIZE is nonzero and we're performing BB |
13659 | vectorization, make sure that the number of elements in the vector |
13660 | is no bigger than GROUP_SIZE. */ |
13661 | |
13662 | tree |
13663 | get_vectype_for_scalar_type (vec_info *vinfo, tree scalar_type, |
13664 | unsigned int group_size) |
13665 | { |
13666 | /* For BB vectorization, we should always have a group size once we've |
13667 | constructed the SLP tree; the only valid uses of zero GROUP_SIZEs |
13668 | are tentative requests during things like early data reference |
13669 | analysis and pattern recognition. */ |
13670 | if (is_a <bb_vec_info> (p: vinfo)) |
13671 | gcc_assert (vinfo->slp_instances.is_empty () || group_size != 0); |
13672 | else |
13673 | group_size = 0; |
13674 | |
13675 | tree vectype = get_related_vectype_for_scalar_type (prevailing_mode: vinfo->vector_mode, |
13676 | scalar_type); |
13677 | if (vectype && vinfo->vector_mode == VOIDmode) |
13678 | vinfo->vector_mode = TYPE_MODE (vectype); |
13679 | |
13680 | /* Register the natural choice of vector type, before the group size |
13681 | has been applied. */ |
13682 | if (vectype) |
13683 | vinfo->used_vector_modes.add (TYPE_MODE (vectype)); |
13684 | |
13685 | /* If the natural choice of vector type doesn't satisfy GROUP_SIZE, |
13686 | try again with an explicit number of elements. */ |
13687 | if (vectype |
13688 | && group_size |
13689 | && maybe_ge (TYPE_VECTOR_SUBPARTS (vectype), group_size)) |
13690 | { |
13691 | /* Start with the biggest number of units that fits within |
13692 | GROUP_SIZE and halve it until we find a valid vector type. |
13693 | Usually either the first attempt will succeed or all will |
13694 | fail (in the latter case because GROUP_SIZE is too small |
13695 | for the target), but it's possible that a target could have |
13696 | a hole between supported vector types. |
13697 | |
13698 | If GROUP_SIZE is not a power of 2, this has the effect of |
13699 | trying the largest power of 2 that fits within the group, |
13700 | even though the group is not a multiple of that vector size. |
13701 | The BB vectorizer will then try to carve up the group into |
13702 | smaller pieces. */ |
13703 | unsigned int nunits = 1 << floor_log2 (x: group_size); |
13704 | do |
13705 | { |
13706 | vectype = get_related_vectype_for_scalar_type (prevailing_mode: vinfo->vector_mode, |
13707 | scalar_type, nunits); |
13708 | nunits /= 2; |
13709 | } |
13710 | while (nunits > 1 && !vectype); |
13711 | } |
13712 | |
13713 | return vectype; |
13714 | } |
13715 | |
13716 | /* Return the vector type corresponding to SCALAR_TYPE as supported |
13717 | by the target. NODE, if nonnull, is the SLP tree node that will |
13718 | use the returned vector type. */ |
13719 | |
13720 | tree |
13721 | get_vectype_for_scalar_type (vec_info *vinfo, tree scalar_type, slp_tree node) |
13722 | { |
13723 | unsigned int group_size = 0; |
13724 | if (node) |
13725 | group_size = SLP_TREE_LANES (node); |
13726 | return get_vectype_for_scalar_type (vinfo, scalar_type, group_size); |
13727 | } |
13728 | |
13729 | /* Function get_mask_type_for_scalar_type. |
13730 | |
13731 | Returns the mask type corresponding to a result of comparison |
13732 | of vectors of specified SCALAR_TYPE as supported by target. |
13733 | If GROUP_SIZE is nonzero and we're performing BB vectorization, |
13734 | make sure that the number of elements in the vector is no bigger |
13735 | than GROUP_SIZE. */ |
13736 | |
13737 | tree |
13738 | get_mask_type_for_scalar_type (vec_info *vinfo, tree scalar_type, |
13739 | unsigned int group_size) |
13740 | { |
13741 | tree vectype = get_vectype_for_scalar_type (vinfo, scalar_type, group_size); |
13742 | |
13743 | if (!vectype) |
13744 | return NULL; |
13745 | |
13746 | return truth_type_for (vectype); |
13747 | } |
13748 | |
13749 | /* Function get_mask_type_for_scalar_type. |
13750 | |
13751 | Returns the mask type corresponding to a result of comparison |
13752 | of vectors of specified SCALAR_TYPE as supported by target. |
13753 | NODE, if nonnull, is the SLP tree node that will use the returned |
13754 | vector type. */ |
13755 | |
13756 | tree |
13757 | get_mask_type_for_scalar_type (vec_info *vinfo, tree scalar_type, |
13758 | slp_tree node) |
13759 | { |
13760 | tree vectype = get_vectype_for_scalar_type (vinfo, scalar_type, node); |
13761 | |
13762 | if (!vectype) |
13763 | return NULL; |
13764 | |
13765 | return truth_type_for (vectype); |
13766 | } |
13767 | |
13768 | /* Function get_same_sized_vectype |
13769 | |
13770 | Returns a vector type corresponding to SCALAR_TYPE of size |
13771 | VECTOR_TYPE if supported by the target. */ |
13772 | |
13773 | tree |
13774 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
13775 | { |
13776 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type)) |
13777 | return truth_type_for (vector_type); |
13778 | |
13779 | poly_uint64 nunits; |
13780 | if (!multiple_p (a: GET_MODE_SIZE (TYPE_MODE (vector_type)), |
13781 | b: GET_MODE_SIZE (TYPE_MODE (scalar_type)), multiple: &nunits)) |
13782 | return NULL_TREE; |
13783 | |
13784 | return get_related_vectype_for_scalar_type (TYPE_MODE (vector_type), |
13785 | scalar_type, nunits); |
13786 | } |
13787 | |
13788 | /* Return true if replacing LOOP_VINFO->vector_mode with VECTOR_MODE |
13789 | would not change the chosen vector modes. */ |
13790 | |
13791 | bool |
13792 | vect_chooses_same_modes_p (vec_info *vinfo, machine_mode vector_mode) |
13793 | { |
13794 | for (vec_info::mode_set::iterator i = vinfo->used_vector_modes.begin (); |
13795 | i != vinfo->used_vector_modes.end (); ++i) |
13796 | if (!VECTOR_MODE_P (*i) |
13797 | || related_vector_mode (vector_mode, GET_MODE_INNER (*i), 0) != *i) |
13798 | return false; |
13799 | return true; |
13800 | } |
13801 | |
13802 | /* Function vect_is_simple_use. |
13803 | |
13804 | Input: |
13805 | VINFO - the vect info of the loop or basic block that is being vectorized. |
13806 | OPERAND - operand in the loop or bb. |
13807 | Output: |
13808 | DEF_STMT_INFO_OUT (optional) - information about the defining stmt in |
13809 | case OPERAND is an SSA_NAME that is defined in the vectorizable region |
13810 | DEF_STMT_OUT (optional) - the defining stmt in case OPERAND is an SSA_NAME; |
13811 | the definition could be anywhere in the function |
13812 | DT - the type of definition |
13813 | |
13814 | Returns whether a stmt with OPERAND can be vectorized. |
13815 | For loops, supportable operands are constants, loop invariants, and operands |
13816 | that are defined by the current iteration of the loop. Unsupportable |
13817 | operands are those that are defined by a previous iteration of the loop (as |
13818 | is the case in reduction/induction computations). |
13819 | For basic blocks, supportable operands are constants and bb invariants. |
13820 | For now, operands defined outside the basic block are not supported. */ |
13821 | |
13822 | bool |
13823 | vect_is_simple_use (tree operand, vec_info *vinfo, enum vect_def_type *dt, |
13824 | stmt_vec_info *def_stmt_info_out, gimple **def_stmt_out) |
13825 | { |
13826 | if (def_stmt_info_out) |
13827 | *def_stmt_info_out = NULL; |
13828 | if (def_stmt_out) |
13829 | *def_stmt_out = NULL; |
13830 | *dt = vect_unknown_def_type; |
13831 | |
13832 | if (dump_enabled_p ()) |
13833 | { |
13834 | dump_printf_loc (MSG_NOTE, vect_location, |
13835 | "vect_is_simple_use: operand " ); |
13836 | if (TREE_CODE (operand) == SSA_NAME |
13837 | && !SSA_NAME_IS_DEFAULT_DEF (operand)) |
13838 | dump_gimple_expr (MSG_NOTE, TDF_SLIM, SSA_NAME_DEF_STMT (operand), 0); |
13839 | else |
13840 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); |
13841 | } |
13842 | |
13843 | if (CONSTANT_CLASS_P (operand)) |
13844 | *dt = vect_constant_def; |
13845 | else if (is_gimple_min_invariant (operand)) |
13846 | *dt = vect_external_def; |
13847 | else if (TREE_CODE (operand) != SSA_NAME) |
13848 | *dt = vect_unknown_def_type; |
13849 | else if (SSA_NAME_IS_DEFAULT_DEF (operand)) |
13850 | *dt = vect_external_def; |
13851 | else |
13852 | { |
13853 | gimple *def_stmt = SSA_NAME_DEF_STMT (operand); |
13854 | stmt_vec_info stmt_vinfo = vinfo->lookup_def (operand); |
13855 | if (!stmt_vinfo) |
13856 | *dt = vect_external_def; |
13857 | else |
13858 | { |
13859 | stmt_vinfo = vect_stmt_to_vectorize (stmt_info: stmt_vinfo); |
13860 | def_stmt = stmt_vinfo->stmt; |
13861 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
13862 | if (def_stmt_info_out) |
13863 | *def_stmt_info_out = stmt_vinfo; |
13864 | } |
13865 | if (def_stmt_out) |
13866 | *def_stmt_out = def_stmt; |
13867 | } |
13868 | |
13869 | if (dump_enabled_p ()) |
13870 | { |
13871 | dump_printf (MSG_NOTE, ", type of def: " ); |
13872 | switch (*dt) |
13873 | { |
13874 | case vect_uninitialized_def: |
13875 | dump_printf (MSG_NOTE, "uninitialized\n" ); |
13876 | break; |
13877 | case vect_constant_def: |
13878 | dump_printf (MSG_NOTE, "constant\n" ); |
13879 | break; |
13880 | case vect_external_def: |
13881 | dump_printf (MSG_NOTE, "external\n" ); |
13882 | break; |
13883 | case vect_internal_def: |
13884 | dump_printf (MSG_NOTE, "internal\n" ); |
13885 | break; |
13886 | case vect_induction_def: |
13887 | dump_printf (MSG_NOTE, "induction\n" ); |
13888 | break; |
13889 | case vect_reduction_def: |
13890 | dump_printf (MSG_NOTE, "reduction\n" ); |
13891 | break; |
13892 | case vect_double_reduction_def: |
13893 | dump_printf (MSG_NOTE, "double reduction\n" ); |
13894 | break; |
13895 | case vect_nested_cycle: |
13896 | dump_printf (MSG_NOTE, "nested cycle\n" ); |
13897 | break; |
13898 | case vect_first_order_recurrence: |
13899 | dump_printf (MSG_NOTE, "first order recurrence\n" ); |
13900 | break; |
13901 | case vect_condition_def: |
13902 | dump_printf (MSG_NOTE, "control flow\n" ); |
13903 | break; |
13904 | case vect_unknown_def_type: |
13905 | dump_printf (MSG_NOTE, "unknown\n" ); |
13906 | break; |
13907 | } |
13908 | } |
13909 | |
13910 | if (*dt == vect_unknown_def_type) |
13911 | { |
13912 | if (dump_enabled_p ()) |
13913 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
13914 | "Unsupported pattern.\n" ); |
13915 | return false; |
13916 | } |
13917 | |
13918 | return true; |
13919 | } |
13920 | |
13921 | /* Function vect_is_simple_use. |
13922 | |
13923 | Same as vect_is_simple_use but also determines the vector operand |
13924 | type of OPERAND and stores it to *VECTYPE. If the definition of |
13925 | OPERAND is vect_uninitialized_def, vect_constant_def or |
13926 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller |
13927 | is responsible to compute the best suited vector type for the |
13928 | scalar operand. */ |
13929 | |
13930 | bool |
13931 | vect_is_simple_use (tree operand, vec_info *vinfo, enum vect_def_type *dt, |
13932 | tree *vectype, stmt_vec_info *def_stmt_info_out, |
13933 | gimple **def_stmt_out) |
13934 | { |
13935 | stmt_vec_info def_stmt_info; |
13936 | gimple *def_stmt; |
13937 | if (!vect_is_simple_use (operand, vinfo, dt, def_stmt_info_out: &def_stmt_info, def_stmt_out: &def_stmt)) |
13938 | return false; |
13939 | |
13940 | if (def_stmt_out) |
13941 | *def_stmt_out = def_stmt; |
13942 | if (def_stmt_info_out) |
13943 | *def_stmt_info_out = def_stmt_info; |
13944 | |
13945 | /* Now get a vector type if the def is internal, otherwise supply |
13946 | NULL_TREE and leave it up to the caller to figure out a proper |
13947 | type for the use stmt. */ |
13948 | if (*dt == vect_internal_def |
13949 | || *dt == vect_induction_def |
13950 | || *dt == vect_reduction_def |
13951 | || *dt == vect_double_reduction_def |
13952 | || *dt == vect_nested_cycle |
13953 | || *dt == vect_first_order_recurrence) |
13954 | { |
13955 | *vectype = STMT_VINFO_VECTYPE (def_stmt_info); |
13956 | gcc_assert (*vectype != NULL_TREE); |
13957 | if (dump_enabled_p ()) |
13958 | dump_printf_loc (MSG_NOTE, vect_location, |
13959 | "vect_is_simple_use: vectype %T\n" , *vectype); |
13960 | } |
13961 | else if (*dt == vect_uninitialized_def |
13962 | || *dt == vect_constant_def |
13963 | || *dt == vect_external_def) |
13964 | *vectype = NULL_TREE; |
13965 | else |
13966 | gcc_unreachable (); |
13967 | |
13968 | return true; |
13969 | } |
13970 | |
13971 | /* Function vect_is_simple_use. |
13972 | |
13973 | Same as vect_is_simple_use but determines the operand by operand |
13974 | position OPERAND from either STMT or SLP_NODE, filling in *OP |
13975 | and *SLP_DEF (when SLP_NODE is not NULL). */ |
13976 | |
13977 | bool |
13978 | vect_is_simple_use (vec_info *vinfo, stmt_vec_info stmt, slp_tree slp_node, |
13979 | unsigned operand, tree *op, slp_tree *slp_def, |
13980 | enum vect_def_type *dt, |
13981 | tree *vectype, stmt_vec_info *def_stmt_info_out) |
13982 | { |
13983 | if (slp_node) |
13984 | { |
13985 | slp_tree child = SLP_TREE_CHILDREN (slp_node)[operand]; |
13986 | *slp_def = child; |
13987 | *vectype = SLP_TREE_VECTYPE (child); |
13988 | if (SLP_TREE_DEF_TYPE (child) == vect_internal_def) |
13989 | { |
13990 | *op = gimple_get_lhs (SLP_TREE_REPRESENTATIVE (child)->stmt); |
13991 | return vect_is_simple_use (operand: *op, vinfo, dt, def_stmt_info_out); |
13992 | } |
13993 | else |
13994 | { |
13995 | if (def_stmt_info_out) |
13996 | *def_stmt_info_out = NULL; |
13997 | *op = SLP_TREE_SCALAR_OPS (child)[0]; |
13998 | *dt = SLP_TREE_DEF_TYPE (child); |
13999 | return true; |
14000 | } |
14001 | } |
14002 | else |
14003 | { |
14004 | *slp_def = NULL; |
14005 | if (gassign *ass = dyn_cast <gassign *> (p: stmt->stmt)) |
14006 | { |
14007 | if (gimple_assign_rhs_code (gs: ass) == COND_EXPR |
14008 | && COMPARISON_CLASS_P (gimple_assign_rhs1 (ass))) |
14009 | { |
14010 | if (operand < 2) |
14011 | *op = TREE_OPERAND (gimple_assign_rhs1 (ass), operand); |
14012 | else |
14013 | *op = gimple_op (gs: ass, i: operand); |
14014 | } |
14015 | else if (gimple_assign_rhs_code (gs: ass) == VIEW_CONVERT_EXPR) |
14016 | *op = TREE_OPERAND (gimple_assign_rhs1 (ass), 0); |
14017 | else |
14018 | *op = gimple_op (gs: ass, i: operand + 1); |
14019 | } |
14020 | else if (gcond *cond = dyn_cast <gcond *> (p: stmt->stmt)) |
14021 | *op = gimple_op (gs: cond, i: operand); |
14022 | else if (gcall *call = dyn_cast <gcall *> (p: stmt->stmt)) |
14023 | *op = gimple_call_arg (gs: call, index: operand); |
14024 | else |
14025 | gcc_unreachable (); |
14026 | return vect_is_simple_use (operand: *op, vinfo, dt, vectype, def_stmt_info_out); |
14027 | } |
14028 | } |
14029 | |
14030 | /* If OP is not NULL and is external or constant update its vector |
14031 | type with VECTYPE. Returns true if successful or false if not, |
14032 | for example when conflicting vector types are present. */ |
14033 | |
14034 | bool |
14035 | vect_maybe_update_slp_op_vectype (slp_tree op, tree vectype) |
14036 | { |
14037 | if (!op || SLP_TREE_DEF_TYPE (op) == vect_internal_def) |
14038 | return true; |
14039 | if (SLP_TREE_VECTYPE (op)) |
14040 | return types_compatible_p (SLP_TREE_VECTYPE (op), type2: vectype); |
14041 | /* For external defs refuse to produce VECTOR_BOOLEAN_TYPE_P, those |
14042 | should be handled by patters. Allow vect_constant_def for now. */ |
14043 | if (VECTOR_BOOLEAN_TYPE_P (vectype) |
14044 | && SLP_TREE_DEF_TYPE (op) == vect_external_def) |
14045 | return false; |
14046 | SLP_TREE_VECTYPE (op) = vectype; |
14047 | return true; |
14048 | } |
14049 | |
14050 | /* Function supportable_widening_operation |
14051 | |
14052 | Check whether an operation represented by the code CODE is a |
14053 | widening operation that is supported by the target platform in |
14054 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
14055 | producing a result of type VECTYPE_OUT). |
14056 | |
14057 | Widening operations we currently support are NOP (CONVERT), FLOAT, |
14058 | FIX_TRUNC and WIDEN_MULT. This function checks if these operations |
14059 | are supported by the target platform either directly (via vector |
14060 | tree-codes), or via target builtins. |
14061 | |
14062 | Output: |
14063 | - CODE1 and CODE2 are codes of vector operations to be used when |
14064 | vectorizing the operation, if available. |
14065 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
14066 | case of multi-step conversion (like char->short->int - in that case |
14067 | MULTI_STEP_CVT will be 1). |
14068 | - INTERM_TYPES contains the intermediate type required to perform the |
14069 | widening operation (short in the above example). */ |
14070 | |
14071 | bool |
14072 | supportable_widening_operation (vec_info *vinfo, |
14073 | code_helper code, |
14074 | stmt_vec_info stmt_info, |
14075 | tree vectype_out, tree vectype_in, |
14076 | code_helper *code1, |
14077 | code_helper *code2, |
14078 | int *multi_step_cvt, |
14079 | vec<tree> *interm_types) |
14080 | { |
14081 | loop_vec_info loop_info = dyn_cast <loop_vec_info> (p: vinfo); |
14082 | class loop *vect_loop = NULL; |
14083 | machine_mode vec_mode; |
14084 | enum insn_code icode1, icode2; |
14085 | optab optab1 = unknown_optab, optab2 = unknown_optab; |
14086 | tree vectype = vectype_in; |
14087 | tree wide_vectype = vectype_out; |
14088 | tree_code c1 = MAX_TREE_CODES, c2 = MAX_TREE_CODES; |
14089 | int i; |
14090 | tree prev_type, intermediate_type; |
14091 | machine_mode intermediate_mode, prev_mode; |
14092 | optab optab3, optab4; |
14093 | |
14094 | *multi_step_cvt = 0; |
14095 | if (loop_info) |
14096 | vect_loop = LOOP_VINFO_LOOP (loop_info); |
14097 | |
14098 | switch (code.safe_as_tree_code ()) |
14099 | { |
14100 | case MAX_TREE_CODES: |
14101 | /* Don't set c1 and c2 if code is not a tree_code. */ |
14102 | break; |
14103 | |
14104 | case WIDEN_MULT_EXPR: |
14105 | /* The result of a vectorized widening operation usually requires |
14106 | two vectors (because the widened results do not fit into one vector). |
14107 | The generated vector results would normally be expected to be |
14108 | generated in the same order as in the original scalar computation, |
14109 | i.e. if 8 results are generated in each vector iteration, they are |
14110 | to be organized as follows: |
14111 | vect1: [res1,res2,res3,res4], |
14112 | vect2: [res5,res6,res7,res8]. |
14113 | |
14114 | However, in the special case that the result of the widening |
14115 | operation is used in a reduction computation only, the order doesn't |
14116 | matter (because when vectorizing a reduction we change the order of |
14117 | the computation). Some targets can take advantage of this and |
14118 | generate more efficient code. For example, targets like Altivec, |
14119 | that support widen_mult using a sequence of {mult_even,mult_odd} |
14120 | generate the following vectors: |
14121 | vect1: [res1,res3,res5,res7], |
14122 | vect2: [res2,res4,res6,res8]. |
14123 | |
14124 | When vectorizing outer-loops, we execute the inner-loop sequentially |
14125 | (each vectorized inner-loop iteration contributes to VF outer-loop |
14126 | iterations in parallel). We therefore don't allow to change the |
14127 | order of the computation in the inner-loop during outer-loop |
14128 | vectorization. */ |
14129 | /* TODO: Another case in which order doesn't *really* matter is when we |
14130 | widen and then contract again, e.g. (short)((int)x * y >> 8). |
14131 | Normally, pack_trunc performs an even/odd permute, whereas the |
14132 | repack from an even/odd expansion would be an interleave, which |
14133 | would be significantly simpler for e.g. AVX2. */ |
14134 | /* In any case, in order to avoid duplicating the code below, recurse |
14135 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values |
14136 | are properly set up for the caller. If we fail, we'll continue with |
14137 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ |
14138 | if (vect_loop |
14139 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction |
14140 | && !nested_in_vect_loop_p (loop: vect_loop, stmt_info) |
14141 | && supportable_widening_operation (vinfo, code: VEC_WIDEN_MULT_EVEN_EXPR, |
14142 | stmt_info, vectype_out, |
14143 | vectype_in, code1, |
14144 | code2, multi_step_cvt, |
14145 | interm_types)) |
14146 | { |
14147 | /* Elements in a vector with vect_used_by_reduction property cannot |
14148 | be reordered if the use chain with this property does not have the |
14149 | same operation. One such an example is s += a * b, where elements |
14150 | in a and b cannot be reordered. Here we check if the vector defined |
14151 | by STMT is only directly used in the reduction statement. */ |
14152 | tree lhs = gimple_assign_lhs (gs: stmt_info->stmt); |
14153 | stmt_vec_info use_stmt_info = loop_info->lookup_single_use (lhs); |
14154 | if (use_stmt_info |
14155 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) |
14156 | return true; |
14157 | } |
14158 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
14159 | c2 = VEC_WIDEN_MULT_HI_EXPR; |
14160 | break; |
14161 | |
14162 | case DOT_PROD_EXPR: |
14163 | c1 = DOT_PROD_EXPR; |
14164 | c2 = DOT_PROD_EXPR; |
14165 | break; |
14166 | |
14167 | case SAD_EXPR: |
14168 | c1 = SAD_EXPR; |
14169 | c2 = SAD_EXPR; |
14170 | break; |
14171 | |
14172 | case VEC_WIDEN_MULT_EVEN_EXPR: |
14173 | /* Support the recursion induced just above. */ |
14174 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; |
14175 | c2 = VEC_WIDEN_MULT_ODD_EXPR; |
14176 | break; |
14177 | |
14178 | case WIDEN_LSHIFT_EXPR: |
14179 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
14180 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; |
14181 | break; |
14182 | |
14183 | CASE_CONVERT: |
14184 | c1 = VEC_UNPACK_LO_EXPR; |
14185 | c2 = VEC_UNPACK_HI_EXPR; |
14186 | break; |
14187 | |
14188 | case FLOAT_EXPR: |
14189 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
14190 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; |
14191 | break; |
14192 | |
14193 | case FIX_TRUNC_EXPR: |
14194 | c1 = VEC_UNPACK_FIX_TRUNC_LO_EXPR; |
14195 | c2 = VEC_UNPACK_FIX_TRUNC_HI_EXPR; |
14196 | break; |
14197 | |
14198 | default: |
14199 | gcc_unreachable (); |
14200 | } |
14201 | |
14202 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
14203 | std::swap (a&: c1, b&: c2); |
14204 | |
14205 | if (code == FIX_TRUNC_EXPR) |
14206 | { |
14207 | /* The signedness is determined from output operand. */ |
14208 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
14209 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); |
14210 | } |
14211 | else if (CONVERT_EXPR_CODE_P (code.safe_as_tree_code ()) |
14212 | && VECTOR_BOOLEAN_TYPE_P (wide_vectype) |
14213 | && VECTOR_BOOLEAN_TYPE_P (vectype) |
14214 | && TYPE_MODE (wide_vectype) == TYPE_MODE (vectype) |
14215 | && SCALAR_INT_MODE_P (TYPE_MODE (vectype))) |
14216 | { |
14217 | /* If the input and result modes are the same, a different optab |
14218 | is needed where we pass in the number of units in vectype. */ |
14219 | optab1 = vec_unpacks_sbool_lo_optab; |
14220 | optab2 = vec_unpacks_sbool_hi_optab; |
14221 | } |
14222 | |
14223 | vec_mode = TYPE_MODE (vectype); |
14224 | if (widening_fn_p (code)) |
14225 | { |
14226 | /* If this is an internal fn then we must check whether the target |
14227 | supports either a low-high split or an even-odd split. */ |
14228 | internal_fn ifn = as_internal_fn (code: (combined_fn) code); |
14229 | |
14230 | internal_fn lo, hi, even, odd; |
14231 | lookup_hilo_internal_fn (ifn, &lo, &hi); |
14232 | *code1 = as_combined_fn (fn: lo); |
14233 | *code2 = as_combined_fn (fn: hi); |
14234 | optab1 = direct_internal_fn_optab (lo, {vectype, vectype}); |
14235 | optab2 = direct_internal_fn_optab (hi, {vectype, vectype}); |
14236 | |
14237 | /* If we don't support low-high, then check for even-odd. */ |
14238 | if (!optab1 |
14239 | || (icode1 = optab_handler (op: optab1, mode: vec_mode)) == CODE_FOR_nothing |
14240 | || !optab2 |
14241 | || (icode2 = optab_handler (op: optab2, mode: vec_mode)) == CODE_FOR_nothing) |
14242 | { |
14243 | lookup_evenodd_internal_fn (ifn, &even, &odd); |
14244 | *code1 = as_combined_fn (fn: even); |
14245 | *code2 = as_combined_fn (fn: odd); |
14246 | optab1 = direct_internal_fn_optab (even, {vectype, vectype}); |
14247 | optab2 = direct_internal_fn_optab (odd, {vectype, vectype}); |
14248 | } |
14249 | } |
14250 | else if (code.is_tree_code ()) |
14251 | { |
14252 | if (code == FIX_TRUNC_EXPR) |
14253 | { |
14254 | /* The signedness is determined from output operand. */ |
14255 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
14256 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); |
14257 | } |
14258 | else if (CONVERT_EXPR_CODE_P ((tree_code) code.safe_as_tree_code ()) |
14259 | && VECTOR_BOOLEAN_TYPE_P (wide_vectype) |
14260 | && VECTOR_BOOLEAN_TYPE_P (vectype) |
14261 | && TYPE_MODE (wide_vectype) == TYPE_MODE (vectype) |
14262 | && SCALAR_INT_MODE_P (TYPE_MODE (vectype))) |
14263 | { |
14264 | /* If the input and result modes are the same, a different optab |
14265 | is needed where we pass in the number of units in vectype. */ |
14266 | optab1 = vec_unpacks_sbool_lo_optab; |
14267 | optab2 = vec_unpacks_sbool_hi_optab; |
14268 | } |
14269 | else |
14270 | { |
14271 | optab1 = optab_for_tree_code (c1, vectype, optab_default); |
14272 | optab2 = optab_for_tree_code (c2, vectype, optab_default); |
14273 | } |
14274 | *code1 = c1; |
14275 | *code2 = c2; |
14276 | } |
14277 | |
14278 | if (!optab1 || !optab2) |
14279 | return false; |
14280 | |
14281 | if ((icode1 = optab_handler (op: optab1, mode: vec_mode)) == CODE_FOR_nothing |
14282 | || (icode2 = optab_handler (op: optab2, mode: vec_mode)) == CODE_FOR_nothing) |
14283 | return false; |
14284 | |
14285 | |
14286 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) |
14287 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) |
14288 | { |
14289 | if (!VECTOR_BOOLEAN_TYPE_P (vectype)) |
14290 | return true; |
14291 | /* For scalar masks we may have different boolean |
14292 | vector types having the same QImode. Thus we |
14293 | add additional check for elements number. */ |
14294 | if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), |
14295 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)) |
14296 | return true; |
14297 | } |
14298 | |
14299 | /* Check if it's a multi-step conversion that can be done using intermediate |
14300 | types. */ |
14301 | |
14302 | prev_type = vectype; |
14303 | prev_mode = vec_mode; |
14304 | |
14305 | if (!CONVERT_EXPR_CODE_P (code.safe_as_tree_code ())) |
14306 | return false; |
14307 | |
14308 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
14309 | intermediate steps in promotion sequence. We try |
14310 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do |
14311 | not. */ |
14312 | interm_types->create (MAX_INTERM_CVT_STEPS); |
14313 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
14314 | { |
14315 | intermediate_mode = insn_data[icode1].operand[0].mode; |
14316 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
14317 | intermediate_type |
14318 | = vect_halve_mask_nunits (prev_type, intermediate_mode); |
14319 | else if (VECTOR_MODE_P (intermediate_mode)) |
14320 | { |
14321 | tree intermediate_element_type |
14322 | = lang_hooks.types.type_for_mode (GET_MODE_INNER (intermediate_mode), |
14323 | TYPE_UNSIGNED (prev_type)); |
14324 | intermediate_type |
14325 | = build_vector_type_for_mode (intermediate_element_type, |
14326 | intermediate_mode); |
14327 | } |
14328 | else |
14329 | intermediate_type |
14330 | = lang_hooks.types.type_for_mode (intermediate_mode, |
14331 | TYPE_UNSIGNED (prev_type)); |
14332 | |
14333 | if (VECTOR_BOOLEAN_TYPE_P (intermediate_type) |
14334 | && VECTOR_BOOLEAN_TYPE_P (prev_type) |
14335 | && intermediate_mode == prev_mode |
14336 | && SCALAR_INT_MODE_P (prev_mode)) |
14337 | { |
14338 | /* If the input and result modes are the same, a different optab |
14339 | is needed where we pass in the number of units in vectype. */ |
14340 | optab3 = vec_unpacks_sbool_lo_optab; |
14341 | optab4 = vec_unpacks_sbool_hi_optab; |
14342 | } |
14343 | else |
14344 | { |
14345 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); |
14346 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); |
14347 | } |
14348 | |
14349 | if (!optab3 || !optab4 |
14350 | || (icode1 = optab_handler (op: optab1, mode: prev_mode)) == CODE_FOR_nothing |
14351 | || insn_data[icode1].operand[0].mode != intermediate_mode |
14352 | || (icode2 = optab_handler (op: optab2, mode: prev_mode)) == CODE_FOR_nothing |
14353 | || insn_data[icode2].operand[0].mode != intermediate_mode |
14354 | || ((icode1 = optab_handler (op: optab3, mode: intermediate_mode)) |
14355 | == CODE_FOR_nothing) |
14356 | || ((icode2 = optab_handler (op: optab4, mode: intermediate_mode)) |
14357 | == CODE_FOR_nothing)) |
14358 | break; |
14359 | |
14360 | interm_types->quick_push (obj: intermediate_type); |
14361 | (*multi_step_cvt)++; |
14362 | |
14363 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) |
14364 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) |
14365 | { |
14366 | if (!VECTOR_BOOLEAN_TYPE_P (vectype)) |
14367 | return true; |
14368 | if (known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type), |
14369 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)) |
14370 | return true; |
14371 | } |
14372 | |
14373 | prev_type = intermediate_type; |
14374 | prev_mode = intermediate_mode; |
14375 | } |
14376 | |
14377 | interm_types->release (); |
14378 | return false; |
14379 | } |
14380 | |
14381 | |
14382 | /* Function supportable_narrowing_operation |
14383 | |
14384 | Check whether an operation represented by the code CODE is a |
14385 | narrowing operation that is supported by the target platform in |
14386 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
14387 | and producing a result of type VECTYPE_OUT). |
14388 | |
14389 | Narrowing operations we currently support are NOP (CONVERT), FIX_TRUNC |
14390 | and FLOAT. This function checks if these operations are supported by |
14391 | the target platform directly via vector tree-codes. |
14392 | |
14393 | Output: |
14394 | - CODE1 is the code of a vector operation to be used when |
14395 | vectorizing the operation, if available. |
14396 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
14397 | case of multi-step conversion (like int->short->char - in that case |
14398 | MULTI_STEP_CVT will be 1). |
14399 | - INTERM_TYPES contains the intermediate type required to perform the |
14400 | narrowing operation (short in the above example). */ |
14401 | |
14402 | bool |
14403 | supportable_narrowing_operation (code_helper code, |
14404 | tree vectype_out, tree vectype_in, |
14405 | code_helper *code1, int *multi_step_cvt, |
14406 | vec<tree> *interm_types) |
14407 | { |
14408 | machine_mode vec_mode; |
14409 | enum insn_code icode1; |
14410 | optab optab1, interm_optab; |
14411 | tree vectype = vectype_in; |
14412 | tree narrow_vectype = vectype_out; |
14413 | enum tree_code c1; |
14414 | tree intermediate_type, prev_type; |
14415 | machine_mode intermediate_mode, prev_mode; |
14416 | int i; |
14417 | unsigned HOST_WIDE_INT n_elts; |
14418 | bool uns; |
14419 | |
14420 | if (!code.is_tree_code ()) |
14421 | return false; |
14422 | |
14423 | *multi_step_cvt = 0; |
14424 | switch ((tree_code) code) |
14425 | { |
14426 | CASE_CONVERT: |
14427 | c1 = VEC_PACK_TRUNC_EXPR; |
14428 | if (VECTOR_BOOLEAN_TYPE_P (narrow_vectype) |
14429 | && VECTOR_BOOLEAN_TYPE_P (vectype) |
14430 | && SCALAR_INT_MODE_P (TYPE_MODE (vectype)) |
14431 | && TYPE_VECTOR_SUBPARTS (node: vectype).is_constant (const_value: &n_elts) |
14432 | && n_elts < BITS_PER_UNIT) |
14433 | optab1 = vec_pack_sbool_trunc_optab; |
14434 | else |
14435 | optab1 = optab_for_tree_code (c1, vectype, optab_default); |
14436 | break; |
14437 | |
14438 | case FIX_TRUNC_EXPR: |
14439 | c1 = VEC_PACK_FIX_TRUNC_EXPR; |
14440 | /* The signedness is determined from output operand. */ |
14441 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
14442 | break; |
14443 | |
14444 | case FLOAT_EXPR: |
14445 | c1 = VEC_PACK_FLOAT_EXPR; |
14446 | optab1 = optab_for_tree_code (c1, vectype, optab_default); |
14447 | break; |
14448 | |
14449 | default: |
14450 | gcc_unreachable (); |
14451 | } |
14452 | |
14453 | if (!optab1) |
14454 | return false; |
14455 | |
14456 | vec_mode = TYPE_MODE (vectype); |
14457 | if ((icode1 = optab_handler (op: optab1, mode: vec_mode)) == CODE_FOR_nothing) |
14458 | return false; |
14459 | |
14460 | *code1 = c1; |
14461 | |
14462 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) |
14463 | { |
14464 | if (!VECTOR_BOOLEAN_TYPE_P (vectype)) |
14465 | return true; |
14466 | /* For scalar masks we may have different boolean |
14467 | vector types having the same QImode. Thus we |
14468 | add additional check for elements number. */ |
14469 | if (known_eq (TYPE_VECTOR_SUBPARTS (vectype) * 2, |
14470 | TYPE_VECTOR_SUBPARTS (narrow_vectype))) |
14471 | return true; |
14472 | } |
14473 | |
14474 | if (code == FLOAT_EXPR) |
14475 | return false; |
14476 | |
14477 | /* Check if it's a multi-step conversion that can be done using intermediate |
14478 | types. */ |
14479 | prev_mode = vec_mode; |
14480 | prev_type = vectype; |
14481 | if (code == FIX_TRUNC_EXPR) |
14482 | uns = TYPE_UNSIGNED (vectype_out); |
14483 | else |
14484 | uns = TYPE_UNSIGNED (vectype); |
14485 | |
14486 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer |
14487 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more |
14488 | costly than signed. */ |
14489 | if (code == FIX_TRUNC_EXPR && uns) |
14490 | { |
14491 | enum insn_code icode2; |
14492 | |
14493 | intermediate_type |
14494 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); |
14495 | interm_optab |
14496 | = optab_for_tree_code (c1, intermediate_type, optab_default); |
14497 | if (interm_optab != unknown_optab |
14498 | && (icode2 = optab_handler (op: optab1, mode: vec_mode)) != CODE_FOR_nothing |
14499 | && insn_data[icode1].operand[0].mode |
14500 | == insn_data[icode2].operand[0].mode) |
14501 | { |
14502 | uns = false; |
14503 | optab1 = interm_optab; |
14504 | icode1 = icode2; |
14505 | } |
14506 | } |
14507 | |
14508 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
14509 | intermediate steps in promotion sequence. We try |
14510 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ |
14511 | interm_types->create (MAX_INTERM_CVT_STEPS); |
14512 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
14513 | { |
14514 | intermediate_mode = insn_data[icode1].operand[0].mode; |
14515 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
14516 | intermediate_type |
14517 | = vect_double_mask_nunits (prev_type, intermediate_mode); |
14518 | else |
14519 | intermediate_type |
14520 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); |
14521 | if (VECTOR_BOOLEAN_TYPE_P (intermediate_type) |
14522 | && VECTOR_BOOLEAN_TYPE_P (prev_type) |
14523 | && SCALAR_INT_MODE_P (prev_mode) |
14524 | && TYPE_VECTOR_SUBPARTS (node: intermediate_type).is_constant (const_value: &n_elts) |
14525 | && n_elts < BITS_PER_UNIT) |
14526 | interm_optab = vec_pack_sbool_trunc_optab; |
14527 | else |
14528 | interm_optab |
14529 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, |
14530 | optab_default); |
14531 | if (!interm_optab |
14532 | || ((icode1 = optab_handler (op: optab1, mode: prev_mode)) == CODE_FOR_nothing) |
14533 | || insn_data[icode1].operand[0].mode != intermediate_mode |
14534 | || ((icode1 = optab_handler (op: interm_optab, mode: intermediate_mode)) |
14535 | == CODE_FOR_nothing)) |
14536 | break; |
14537 | |
14538 | interm_types->quick_push (obj: intermediate_type); |
14539 | (*multi_step_cvt)++; |
14540 | |
14541 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) |
14542 | { |
14543 | if (!VECTOR_BOOLEAN_TYPE_P (vectype)) |
14544 | return true; |
14545 | if (known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type) * 2, |
14546 | TYPE_VECTOR_SUBPARTS (narrow_vectype))) |
14547 | return true; |
14548 | } |
14549 | |
14550 | prev_mode = intermediate_mode; |
14551 | prev_type = intermediate_type; |
14552 | optab1 = interm_optab; |
14553 | } |
14554 | |
14555 | interm_types->release (); |
14556 | return false; |
14557 | } |
14558 | |
14559 | /* Generate and return a vector mask of MASK_TYPE such that |
14560 | mask[I] is true iff J + START_INDEX < END_INDEX for all J <= I. |
14561 | Add the statements to SEQ. */ |
14562 | |
14563 | tree |
14564 | vect_gen_while (gimple_seq *seq, tree mask_type, tree start_index, |
14565 | tree end_index, const char *name) |
14566 | { |
14567 | tree cmp_type = TREE_TYPE (start_index); |
14568 | gcc_checking_assert (direct_internal_fn_supported_p (IFN_WHILE_ULT, |
14569 | cmp_type, mask_type, |
14570 | OPTIMIZE_FOR_SPEED)); |
14571 | gcall *call = gimple_build_call_internal (IFN_WHILE_ULT, 3, |
14572 | start_index, end_index, |
14573 | build_zero_cst (mask_type)); |
14574 | tree tmp; |
14575 | if (name) |
14576 | tmp = make_temp_ssa_name (type: mask_type, NULL, name); |
14577 | else |
14578 | tmp = make_ssa_name (var: mask_type); |
14579 | gimple_call_set_lhs (gs: call, lhs: tmp); |
14580 | gimple_seq_add_stmt (seq, call); |
14581 | return tmp; |
14582 | } |
14583 | |
14584 | /* Generate a vector mask of type MASK_TYPE for which index I is false iff |
14585 | J + START_INDEX < END_INDEX for all J <= I. Add the statements to SEQ. */ |
14586 | |
14587 | tree |
14588 | vect_gen_while_not (gimple_seq *seq, tree mask_type, tree start_index, |
14589 | tree end_index) |
14590 | { |
14591 | tree tmp = vect_gen_while (seq, mask_type, start_index, end_index); |
14592 | return gimple_build (seq, code: BIT_NOT_EXPR, type: mask_type, ops: tmp); |
14593 | } |
14594 | |
14595 | /* Try to compute the vector types required to vectorize STMT_INFO, |
14596 | returning true on success and false if vectorization isn't possible. |
14597 | If GROUP_SIZE is nonzero and we're performing BB vectorization, |
14598 | take sure that the number of elements in the vectors is no bigger |
14599 | than GROUP_SIZE. |
14600 | |
14601 | On success: |
14602 | |
14603 | - Set *STMT_VECTYPE_OUT to: |
14604 | - NULL_TREE if the statement doesn't need to be vectorized; |
14605 | - the equivalent of STMT_VINFO_VECTYPE otherwise. |
14606 | |
14607 | - Set *NUNITS_VECTYPE_OUT to the vector type that contains the maximum |
14608 | number of units needed to vectorize STMT_INFO, or NULL_TREE if the |
14609 | statement does not help to determine the overall number of units. */ |
14610 | |
14611 | opt_result |
14612 | vect_get_vector_types_for_stmt (vec_info *vinfo, stmt_vec_info stmt_info, |
14613 | tree *stmt_vectype_out, |
14614 | tree *nunits_vectype_out, |
14615 | unsigned int group_size) |
14616 | { |
14617 | gimple *stmt = stmt_info->stmt; |
14618 | |
14619 | /* For BB vectorization, we should always have a group size once we've |
14620 | constructed the SLP tree; the only valid uses of zero GROUP_SIZEs |
14621 | are tentative requests during things like early data reference |
14622 | analysis and pattern recognition. */ |
14623 | if (is_a <bb_vec_info> (p: vinfo)) |
14624 | gcc_assert (vinfo->slp_instances.is_empty () || group_size != 0); |
14625 | else |
14626 | group_size = 0; |
14627 | |
14628 | *stmt_vectype_out = NULL_TREE; |
14629 | *nunits_vectype_out = NULL_TREE; |
14630 | |
14631 | if (gimple_get_lhs (stmt) == NULL_TREE |
14632 | /* Allow vector conditionals through here. */ |
14633 | && !is_a <gcond *> (p: stmt) |
14634 | /* MASK_STORE has no lhs, but is ok. */ |
14635 | && !gimple_call_internal_p (gs: stmt, fn: IFN_MASK_STORE)) |
14636 | { |
14637 | if (is_a <gcall *> (p: stmt)) |
14638 | { |
14639 | /* Ignore calls with no lhs. These must be calls to |
14640 | #pragma omp simd functions, and what vectorization factor |
14641 | it really needs can't be determined until |
14642 | vectorizable_simd_clone_call. */ |
14643 | if (dump_enabled_p ()) |
14644 | dump_printf_loc (MSG_NOTE, vect_location, |
14645 | "defer to SIMD clone analysis.\n" ); |
14646 | return opt_result::success (); |
14647 | } |
14648 | |
14649 | return opt_result::failure_at (loc: stmt, |
14650 | fmt: "not vectorized: irregular stmt: %G" , stmt); |
14651 | } |
14652 | |
14653 | tree vectype; |
14654 | tree scalar_type = NULL_TREE; |
14655 | if (group_size == 0 && STMT_VINFO_VECTYPE (stmt_info)) |
14656 | { |
14657 | vectype = STMT_VINFO_VECTYPE (stmt_info); |
14658 | if (dump_enabled_p ()) |
14659 | dump_printf_loc (MSG_NOTE, vect_location, |
14660 | "precomputed vectype: %T\n" , vectype); |
14661 | } |
14662 | else if (vect_use_mask_type_p (stmt_info)) |
14663 | { |
14664 | unsigned int precision = stmt_info->mask_precision; |
14665 | scalar_type = build_nonstandard_integer_type (precision, 1); |
14666 | vectype = get_mask_type_for_scalar_type (vinfo, scalar_type, group_size); |
14667 | if (!vectype) |
14668 | return opt_result::failure_at (loc: stmt, fmt: "not vectorized: unsupported" |
14669 | " data-type %T\n" , scalar_type); |
14670 | if (dump_enabled_p ()) |
14671 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: %T\n" , vectype); |
14672 | } |
14673 | else |
14674 | { |
14675 | /* If we got here with a gcond it means that the target had no available vector |
14676 | mode for the scalar type. We can't vectorize so abort. */ |
14677 | if (is_a <gcond *> (p: stmt)) |
14678 | return opt_result::failure_at (loc: stmt, |
14679 | fmt: "not vectorized:" |
14680 | " unsupported data-type for gcond %T\n" , |
14681 | scalar_type); |
14682 | |
14683 | if (data_reference *dr = STMT_VINFO_DATA_REF (stmt_info)) |
14684 | scalar_type = TREE_TYPE (DR_REF (dr)); |
14685 | else if (gimple_call_internal_p (gs: stmt, fn: IFN_MASK_STORE)) |
14686 | scalar_type = TREE_TYPE (gimple_call_arg (stmt, 3)); |
14687 | else |
14688 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
14689 | |
14690 | if (dump_enabled_p ()) |
14691 | { |
14692 | if (group_size) |
14693 | dump_printf_loc (MSG_NOTE, vect_location, |
14694 | "get vectype for scalar type (group size %d):" |
14695 | " %T\n" , group_size, scalar_type); |
14696 | else |
14697 | dump_printf_loc (MSG_NOTE, vect_location, |
14698 | "get vectype for scalar type: %T\n" , scalar_type); |
14699 | } |
14700 | vectype = get_vectype_for_scalar_type (vinfo, scalar_type, group_size); |
14701 | if (!vectype) |
14702 | return opt_result::failure_at (loc: stmt, |
14703 | fmt: "not vectorized:" |
14704 | " unsupported data-type %T\n" , |
14705 | scalar_type); |
14706 | |
14707 | if (dump_enabled_p ()) |
14708 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: %T\n" , vectype); |
14709 | } |
14710 | |
14711 | if (scalar_type && VECTOR_MODE_P (TYPE_MODE (scalar_type))) |
14712 | return opt_result::failure_at (loc: stmt, |
14713 | fmt: "not vectorized: vector stmt in loop:%G" , |
14714 | stmt); |
14715 | |
14716 | *stmt_vectype_out = vectype; |
14717 | |
14718 | /* Don't try to compute scalar types if the stmt produces a boolean |
14719 | vector; use the existing vector type instead. */ |
14720 | tree nunits_vectype = vectype; |
14721 | if (!VECTOR_BOOLEAN_TYPE_P (vectype)) |
14722 | { |
14723 | /* The number of units is set according to the smallest scalar |
14724 | type (or the largest vector size, but we only support one |
14725 | vector size per vectorization). */ |
14726 | scalar_type = vect_get_smallest_scalar_type (stmt_info, |
14727 | TREE_TYPE (vectype)); |
14728 | if (scalar_type != TREE_TYPE (vectype)) |
14729 | { |
14730 | if (dump_enabled_p ()) |
14731 | dump_printf_loc (MSG_NOTE, vect_location, |
14732 | "get vectype for smallest scalar type: %T\n" , |
14733 | scalar_type); |
14734 | nunits_vectype = get_vectype_for_scalar_type (vinfo, scalar_type, |
14735 | group_size); |
14736 | if (!nunits_vectype) |
14737 | return opt_result::failure_at |
14738 | (loc: stmt, fmt: "not vectorized: unsupported data-type %T\n" , |
14739 | scalar_type); |
14740 | if (dump_enabled_p ()) |
14741 | dump_printf_loc (MSG_NOTE, vect_location, "nunits vectype: %T\n" , |
14742 | nunits_vectype); |
14743 | } |
14744 | } |
14745 | |
14746 | if (!multiple_p (a: TYPE_VECTOR_SUBPARTS (node: nunits_vectype), |
14747 | b: TYPE_VECTOR_SUBPARTS (node: *stmt_vectype_out))) |
14748 | return opt_result::failure_at (loc: stmt, |
14749 | fmt: "Not vectorized: Incompatible number " |
14750 | "of vector subparts between %T and %T\n" , |
14751 | nunits_vectype, *stmt_vectype_out); |
14752 | |
14753 | if (dump_enabled_p ()) |
14754 | { |
14755 | dump_printf_loc (MSG_NOTE, vect_location, "nunits = " ); |
14756 | dump_dec (MSG_NOTE, TYPE_VECTOR_SUBPARTS (node: nunits_vectype)); |
14757 | dump_printf (MSG_NOTE, "\n" ); |
14758 | } |
14759 | |
14760 | *nunits_vectype_out = nunits_vectype; |
14761 | return opt_result::success (); |
14762 | } |
14763 | |
14764 | /* Generate and return statement sequence that sets vector length LEN that is: |
14765 | |
14766 | min_of_start_and_end = min (START_INDEX, END_INDEX); |
14767 | left_len = END_INDEX - min_of_start_and_end; |
14768 | rhs = min (left_len, LEN_LIMIT); |
14769 | LEN = rhs; |
14770 | |
14771 | Note: the cost of the code generated by this function is modeled |
14772 | by vect_estimate_min_profitable_iters, so changes here may need |
14773 | corresponding changes there. */ |
14774 | |
14775 | gimple_seq |
14776 | vect_gen_len (tree len, tree start_index, tree end_index, tree len_limit) |
14777 | { |
14778 | gimple_seq stmts = NULL; |
14779 | tree len_type = TREE_TYPE (len); |
14780 | gcc_assert (TREE_TYPE (start_index) == len_type); |
14781 | |
14782 | tree min = gimple_build (seq: &stmts, code: MIN_EXPR, type: len_type, ops: start_index, ops: end_index); |
14783 | tree left_len = gimple_build (seq: &stmts, code: MINUS_EXPR, type: len_type, ops: end_index, ops: min); |
14784 | tree rhs = gimple_build (seq: &stmts, code: MIN_EXPR, type: len_type, ops: left_len, ops: len_limit); |
14785 | gimple* stmt = gimple_build_assign (len, rhs); |
14786 | gimple_seq_add_stmt (&stmts, stmt); |
14787 | |
14788 | return stmts; |
14789 | } |
14790 | |
14791 | |