1 | /* Loop interchange. |
2 | Copyright (C) 2017-2023 Free Software Foundation, Inc. |
3 | Contributed by ARM Ltd. |
4 | |
5 | This file is part of GCC. |
6 | |
7 | GCC is free software; you can redistribute it and/or modify it |
8 | under the terms of the GNU General Public License as published by the |
9 | Free Software Foundation; either version 3, or (at your option) any |
10 | later version. |
11 | |
12 | GCC is distributed in the hope that it will be useful, but WITHOUT |
13 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
15 | for more details. |
16 | |
17 | You should have received a copy of the GNU General Public License |
18 | along with GCC; see the file COPYING3. If not see |
19 | <http://www.gnu.org/licenses/>. */ |
20 | |
21 | #include "config.h" |
22 | #include "system.h" |
23 | #include "coretypes.h" |
24 | #include "backend.h" |
25 | #include "is-a.h" |
26 | #include "tree.h" |
27 | #include "gimple.h" |
28 | #include "tree-pass.h" |
29 | #include "ssa.h" |
30 | #include "gimple-pretty-print.h" |
31 | #include "fold-const.h" |
32 | #include "gimplify.h" |
33 | #include "gimple-iterator.h" |
34 | #include "gimplify-me.h" |
35 | #include "cfgloop.h" |
36 | #include "tree-ssa.h" |
37 | #include "tree-scalar-evolution.h" |
38 | #include "tree-ssa-loop-manip.h" |
39 | #include "tree-ssa-loop-niter.h" |
40 | #include "tree-ssa-loop-ivopts.h" |
41 | #include "tree-ssa-dce.h" |
42 | #include "tree-data-ref.h" |
43 | #include "tree-vectorizer.h" |
44 | |
45 | /* This pass performs loop interchange: for example, the loop nest |
46 | |
47 | for (int j = 0; j < N; j++) |
48 | for (int k = 0; k < N; k++) |
49 | for (int i = 0; i < N; i++) |
50 | c[i][j] = c[i][j] + a[i][k]*b[k][j]; |
51 | |
52 | is transformed to |
53 | |
54 | for (int i = 0; i < N; i++) |
55 | for (int j = 0; j < N; j++) |
56 | for (int k = 0; k < N; k++) |
57 | c[i][j] = c[i][j] + a[i][k]*b[k][j]; |
58 | |
59 | This pass implements loop interchange in the following steps: |
60 | |
61 | 1) Find perfect loop nest for each innermost loop and compute data |
62 | dependence relations for it. For above example, loop nest is |
63 | <loop_j, loop_k, loop_i>. |
64 | 2) From innermost to outermost loop, this pass tries to interchange |
65 | each loop pair. For above case, it firstly tries to interchange |
66 | <loop_k, loop_i> and loop nest becomes <loop_j, loop_i, loop_k>. |
67 | Then it tries to interchange <loop_j, loop_i> and loop nest becomes |
68 | <loop_i, loop_j, loop_k>. The overall effect is to move innermost |
69 | loop to the outermost position. For loop pair <loop_i, loop_j> |
70 | to be interchanged, we: |
71 | 3) Check if data dependence relations are valid for loop interchange. |
72 | 4) Check if both loops can be interchanged in terms of transformation. |
73 | 5) Check if interchanging the two loops is profitable. |
74 | 6) Interchange the two loops by mapping induction variables. |
75 | |
76 | This pass also handles reductions in loop nest. So far we only support |
77 | simple reduction of inner loop and double reduction of the loop nest. */ |
78 | |
79 | /* Maximum number of stmts in each loop that should be interchanged. */ |
80 | #define MAX_NUM_STMT (param_loop_interchange_max_num_stmts) |
81 | /* Maximum number of data references in loop nest. */ |
82 | #define MAX_DATAREFS (param_loop_max_datarefs_for_datadeps) |
83 | |
84 | /* Comparison ratio of access stride between inner/outer loops to be |
85 | interchanged. This is the minimum stride ratio for loop interchange |
86 | to be profitable. */ |
87 | #define OUTER_STRIDE_RATIO (param_loop_interchange_stride_ratio) |
88 | /* The same as above, but we require higher ratio for interchanging the |
89 | innermost two loops. */ |
90 | #define INNER_STRIDE_RATIO ((OUTER_STRIDE_RATIO) + 1) |
91 | |
92 | /* Comparison ratio of stmt cost between inner/outer loops. Loops won't |
93 | be interchanged if outer loop has too many stmts. */ |
94 | #define STMT_COST_RATIO (3) |
95 | |
96 | /* Vector of strides that DR accesses in each level loop of a loop nest. */ |
97 | #define DR_ACCESS_STRIDE(dr) ((vec<tree> *) dr->aux) |
98 | |
99 | /* Structure recording loop induction variable. */ |
100 | typedef struct induction |
101 | { |
102 | /* IV itself. */ |
103 | tree var; |
104 | /* IV's initializing value, which is the init arg of the IV PHI node. */ |
105 | tree init_val; |
106 | /* IV's initializing expr, which is (the expanded result of) init_val. */ |
107 | tree init_expr; |
108 | /* IV's step. */ |
109 | tree step; |
110 | } *induction_p; |
111 | |
112 | /* Enum type for loop reduction variable. */ |
113 | enum reduction_type |
114 | { |
115 | UNKNOWN_RTYPE = 0, |
116 | SIMPLE_RTYPE, |
117 | DOUBLE_RTYPE |
118 | }; |
119 | |
120 | /* Structure recording loop reduction variable. */ |
121 | typedef struct reduction |
122 | { |
123 | /* Reduction itself. */ |
124 | tree var; |
125 | /* PHI node defining reduction variable. */ |
126 | gphi *phi; |
127 | /* Init and next variables of the reduction. */ |
128 | tree init; |
129 | tree next; |
130 | /* Lcssa PHI node if reduction is used outside of its definition loop. */ |
131 | gphi *lcssa_phi; |
132 | /* Stmts defining init and next. */ |
133 | gimple *producer; |
134 | gimple *consumer; |
135 | /* If init is loaded from memory, this is the loading memory reference. */ |
136 | tree init_ref; |
137 | /* If reduction is finally stored to memory, this is the stored memory |
138 | reference. */ |
139 | tree fini_ref; |
140 | enum reduction_type type; |
141 | } *reduction_p; |
142 | |
143 | |
144 | /* Dump reduction RE. */ |
145 | |
146 | static void |
147 | dump_reduction (reduction_p re) |
148 | { |
149 | if (re->type == SIMPLE_RTYPE) |
150 | fprintf (stream: dump_file, format: " Simple reduction: " ); |
151 | else if (re->type == DOUBLE_RTYPE) |
152 | fprintf (stream: dump_file, format: " Double reduction: " ); |
153 | else |
154 | fprintf (stream: dump_file, format: " Unknown reduction: " ); |
155 | |
156 | print_gimple_stmt (dump_file, re->phi, 0); |
157 | } |
158 | |
159 | /* Dump LOOP's induction IV. */ |
160 | static void |
161 | dump_induction (class loop *loop, induction_p iv) |
162 | { |
163 | fprintf (stream: dump_file, format: " Induction: " ); |
164 | print_generic_expr (dump_file, iv->var, TDF_SLIM); |
165 | fprintf (stream: dump_file, format: " = {" ); |
166 | print_generic_expr (dump_file, iv->init_expr, TDF_SLIM); |
167 | fprintf (stream: dump_file, format: ", " ); |
168 | print_generic_expr (dump_file, iv->step, TDF_SLIM); |
169 | fprintf (stream: dump_file, format: "}_%d\n" , loop->num); |
170 | } |
171 | |
172 | /* Loop candidate for interchange. */ |
173 | |
174 | class loop_cand |
175 | { |
176 | public: |
177 | loop_cand (class loop *, class loop *); |
178 | ~loop_cand (); |
179 | |
180 | reduction_p find_reduction_by_stmt (gimple *); |
181 | void classify_simple_reduction (reduction_p); |
182 | bool analyze_iloop_reduction_var (tree); |
183 | bool analyze_oloop_reduction_var (loop_cand *, tree); |
184 | bool analyze_induction_var (tree, tree); |
185 | bool analyze_carried_vars (loop_cand *); |
186 | bool analyze_lcssa_phis (void); |
187 | bool can_interchange_p (loop_cand *); |
188 | void undo_simple_reduction (reduction_p, bitmap); |
189 | |
190 | /* The loop itself. */ |
191 | class loop *m_loop; |
192 | /* The outer loop for interchange. It equals to loop if this loop cand |
193 | itself represents the outer loop. */ |
194 | class loop *m_outer; |
195 | /* Vector of induction variables in loop. */ |
196 | vec<induction_p> m_inductions; |
197 | /* Vector of reduction variables in loop. */ |
198 | vec<reduction_p> m_reductions; |
199 | /* Lcssa PHI nodes of this loop. */ |
200 | vec<gphi *> m_lcssa_nodes; |
201 | /* Single exit edge of this loop. */ |
202 | edge m_exit; |
203 | /* Basic blocks of this loop. */ |
204 | basic_block *m_bbs; |
205 | /* Number of stmts of this loop. Inner loops' stmts are not included. */ |
206 | int m_num_stmts; |
207 | /* Number of constant initialized simple reduction. */ |
208 | int m_const_init_reduc; |
209 | }; |
210 | |
211 | /* Constructor. */ |
212 | |
213 | loop_cand::loop_cand (class loop *loop, class loop *outer) |
214 | : m_loop (loop), m_outer (outer), m_exit (single_exit (loop)), |
215 | m_bbs (get_loop_body (loop)), m_num_stmts (0), m_const_init_reduc (0) |
216 | { |
217 | m_inductions.create (nelems: 3); |
218 | m_reductions.create (nelems: 3); |
219 | m_lcssa_nodes.create (nelems: 3); |
220 | } |
221 | |
222 | /* Destructor. */ |
223 | |
224 | loop_cand::~loop_cand () |
225 | { |
226 | induction_p iv; |
227 | for (unsigned i = 0; m_inductions.iterate (ix: i, ptr: &iv); ++i) |
228 | free (ptr: iv); |
229 | |
230 | reduction_p re; |
231 | for (unsigned i = 0; m_reductions.iterate (ix: i, ptr: &re); ++i) |
232 | free (ptr: re); |
233 | |
234 | m_inductions.release (); |
235 | m_reductions.release (); |
236 | m_lcssa_nodes.release (); |
237 | free (ptr: m_bbs); |
238 | } |
239 | |
240 | /* Return single use stmt of VAR in LOOP, otherwise return NULL. */ |
241 | |
242 | static gimple * |
243 | single_use_in_loop (tree var, class loop *loop) |
244 | { |
245 | gimple *stmt, *res = NULL; |
246 | use_operand_p use_p; |
247 | imm_use_iterator iterator; |
248 | |
249 | FOR_EACH_IMM_USE_FAST (use_p, iterator, var) |
250 | { |
251 | stmt = USE_STMT (use_p); |
252 | if (is_gimple_debug (gs: stmt)) |
253 | continue; |
254 | |
255 | if (!flow_bb_inside_loop_p (loop, gimple_bb (g: stmt))) |
256 | continue; |
257 | |
258 | if (res) |
259 | return NULL; |
260 | |
261 | res = stmt; |
262 | } |
263 | return res; |
264 | } |
265 | |
266 | /* Return true if E is unsupported in loop interchange, i.e, E is a complex |
267 | edge or part of irreducible loop. */ |
268 | |
269 | static inline bool |
270 | unsupported_edge (edge e) |
271 | { |
272 | return (e->flags & (EDGE_COMPLEX | EDGE_IRREDUCIBLE_LOOP)); |
273 | } |
274 | |
275 | /* Return the reduction if STMT is one of its lcssa PHI, producer or consumer |
276 | stmt. */ |
277 | |
278 | reduction_p |
279 | loop_cand::find_reduction_by_stmt (gimple *stmt) |
280 | { |
281 | gphi *phi = dyn_cast <gphi *> (p: stmt); |
282 | reduction_p re; |
283 | |
284 | for (unsigned i = 0; m_reductions.iterate (ix: i, ptr: &re); ++i) |
285 | if ((phi != NULL && phi == re->lcssa_phi) |
286 | || (stmt == re->producer || stmt == re->consumer)) |
287 | return re; |
288 | |
289 | return NULL; |
290 | } |
291 | |
292 | /* Return true if current loop_cand be interchanged. ILOOP is not NULL if |
293 | current loop_cand is outer loop in loop nest. */ |
294 | |
295 | bool |
296 | loop_cand::can_interchange_p (loop_cand *iloop) |
297 | { |
298 | /* For now we only support at most one reduction. */ |
299 | unsigned allowed_reduction_num = 1; |
300 | |
301 | /* Only support reduction if the loop nest to be interchanged is the |
302 | innermostin two loops. */ |
303 | if ((iloop == NULL && m_loop->inner != NULL) |
304 | || (iloop != NULL && iloop->m_loop->inner != NULL)) |
305 | allowed_reduction_num = 0; |
306 | |
307 | if (m_reductions.length () > allowed_reduction_num |
308 | || (m_reductions.length () == 1 |
309 | && m_reductions[0]->type == UNKNOWN_RTYPE)) |
310 | return false; |
311 | |
312 | /* Only support lcssa PHI node which is for reduction. */ |
313 | if (m_lcssa_nodes.length () > allowed_reduction_num) |
314 | return false; |
315 | |
316 | /* Check if basic block has any unsupported operation. Note basic blocks |
317 | of inner loops are not checked here. */ |
318 | for (unsigned i = 0; i < m_loop->num_nodes; i++) |
319 | { |
320 | basic_block bb = m_bbs[i]; |
321 | gphi_iterator psi; |
322 | gimple_stmt_iterator gsi; |
323 | |
324 | /* Skip basic blocks of inner loops. */ |
325 | if (bb->loop_father != m_loop) |
326 | continue; |
327 | |
328 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
329 | { |
330 | gimple *stmt = gsi_stmt (i: gsi); |
331 | if (is_gimple_debug (gs: stmt)) |
332 | continue; |
333 | |
334 | if (gimple_has_side_effects (stmt)) |
335 | return false; |
336 | |
337 | m_num_stmts++; |
338 | if (gcall *call = dyn_cast <gcall *> (p: stmt)) |
339 | { |
340 | /* In basic block of outer loop, the call should be cheap since |
341 | it will be moved to inner loop. */ |
342 | if (iloop != NULL |
343 | && !gimple_inexpensive_call_p (call)) |
344 | return false; |
345 | continue; |
346 | } |
347 | |
348 | if (!iloop || !gimple_vuse (g: stmt)) |
349 | continue; |
350 | |
351 | /* Support stmt accessing memory in outer loop only if it is for |
352 | inner loop's reduction. */ |
353 | if (iloop->find_reduction_by_stmt (stmt)) |
354 | continue; |
355 | |
356 | tree lhs; |
357 | /* Support loop invariant memory reference if it's only used once by |
358 | inner loop. */ |
359 | /* ??? How's this checking for invariantness? */ |
360 | if (gimple_assign_single_p (gs: stmt) |
361 | && (lhs = gimple_assign_lhs (gs: stmt)) != NULL_TREE |
362 | && TREE_CODE (lhs) == SSA_NAME |
363 | && single_use_in_loop (var: lhs, loop: iloop->m_loop)) |
364 | continue; |
365 | |
366 | return false; |
367 | } |
368 | /* Check if loop has too many stmts. */ |
369 | if (m_num_stmts > MAX_NUM_STMT) |
370 | return false; |
371 | |
372 | /* Allow PHI nodes in any basic block of inner loop, PHI nodes in outer |
373 | loop's header, or PHI nodes in dest bb of inner loop's exit edge. */ |
374 | if (!iloop || bb == m_loop->header |
375 | || bb == iloop->m_exit->dest) |
376 | continue; |
377 | |
378 | /* Don't allow any other PHI nodes. */ |
379 | for (psi = gsi_start_phis (bb); !gsi_end_p (i: psi); gsi_next (i: &psi)) |
380 | if (!virtual_operand_p (PHI_RESULT (psi.phi ()))) |
381 | return false; |
382 | } |
383 | |
384 | return true; |
385 | } |
386 | |
387 | /* Programmers and optimizers (like loop store motion) may optimize code: |
388 | |
389 | for (int i = 0; i < N; i++) |
390 | for (int j = 0; j < N; j++) |
391 | a[i] += b[j][i] * c[j][i]; |
392 | |
393 | into reduction: |
394 | |
395 | for (int i = 0; i < N; i++) |
396 | { |
397 | // producer. Note sum can be intitialized to a constant. |
398 | int sum = a[i]; |
399 | for (int j = 0; j < N; j++) |
400 | { |
401 | sum += b[j][i] * c[j][i]; |
402 | } |
403 | // consumer. |
404 | a[i] = sum; |
405 | } |
406 | |
407 | The result code can't be interchanged without undoing the optimization. |
408 | This function classifies this kind reduction and records information so |
409 | that we can undo the store motion during interchange. */ |
410 | |
411 | void |
412 | loop_cand::classify_simple_reduction (reduction_p re) |
413 | { |
414 | gimple *producer, *consumer; |
415 | |
416 | /* Check init variable of reduction and how it is initialized. */ |
417 | if (TREE_CODE (re->init) == SSA_NAME) |
418 | { |
419 | producer = SSA_NAME_DEF_STMT (re->init); |
420 | re->producer = producer; |
421 | basic_block bb = gimple_bb (g: producer); |
422 | if (!bb || bb->loop_father != m_outer) |
423 | return; |
424 | |
425 | if (!gimple_assign_load_p (producer)) |
426 | return; |
427 | |
428 | re->init_ref = gimple_assign_rhs1 (gs: producer); |
429 | } |
430 | else if (CONSTANT_CLASS_P (re->init)) |
431 | m_const_init_reduc++; |
432 | else |
433 | return; |
434 | |
435 | /* Check how reduction variable is used. */ |
436 | consumer = single_use_in_loop (PHI_RESULT (re->lcssa_phi), loop: m_outer); |
437 | if (!consumer |
438 | || !gimple_store_p (gs: consumer)) |
439 | return; |
440 | |
441 | re->fini_ref = gimple_get_lhs (consumer); |
442 | re->consumer = consumer; |
443 | |
444 | /* Simple reduction with constant initializer. */ |
445 | if (!re->init_ref) |
446 | { |
447 | gcc_assert (CONSTANT_CLASS_P (re->init)); |
448 | re->init_ref = unshare_expr (re->fini_ref); |
449 | } |
450 | |
451 | /* Require memory references in producer and consumer are the same so |
452 | that we can undo reduction during interchange. */ |
453 | if (re->init_ref && !operand_equal_p (re->init_ref, re->fini_ref, flags: 0)) |
454 | return; |
455 | |
456 | re->type = SIMPLE_RTYPE; |
457 | } |
458 | |
459 | /* Analyze reduction variable VAR for inner loop of the loop nest to be |
460 | interchanged. Return true if analysis succeeds. */ |
461 | |
462 | bool |
463 | loop_cand::analyze_iloop_reduction_var (tree var) |
464 | { |
465 | gphi *phi = as_a <gphi *> (SSA_NAME_DEF_STMT (var)); |
466 | gphi *lcssa_phi = NULL, *use_phi; |
467 | tree init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (m_loop)); |
468 | tree next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (m_loop)); |
469 | reduction_p re; |
470 | gimple *stmt, *next_def, *single_use = NULL; |
471 | use_operand_p use_p; |
472 | imm_use_iterator iterator; |
473 | |
474 | if (TREE_CODE (next) != SSA_NAME) |
475 | return false; |
476 | |
477 | next_def = SSA_NAME_DEF_STMT (next); |
478 | basic_block bb = gimple_bb (g: next_def); |
479 | if (!bb || !flow_bb_inside_loop_p (m_loop, bb)) |
480 | return false; |
481 | |
482 | /* In restricted reduction, the var is (and must be) used in defining |
483 | the updated var. The process can be depicted as below: |
484 | |
485 | var ;; = PHI<init, next> |
486 | | |
487 | | |
488 | v |
489 | +---------------------+ |
490 | | reduction operators | <-- other operands |
491 | +---------------------+ |
492 | | |
493 | | |
494 | v |
495 | next |
496 | |
497 | In terms loop interchange, we don't change how NEXT is computed based |
498 | on VAR and OTHER OPERANDS. In case of double reduction in loop nest |
499 | to be interchanged, we don't changed it at all. In the case of simple |
500 | reduction in inner loop, we only make change how VAR/NEXT is loaded or |
501 | stored. With these conditions, we can relax restrictions on reduction |
502 | in a way that reduction operation is seen as black box. In general, |
503 | we can ignore reassociation of reduction operator; we can handle fake |
504 | reductions in which VAR is not even used to compute NEXT. */ |
505 | if (! single_imm_use (var, use_p: &use_p, stmt: &single_use) |
506 | || ! flow_bb_inside_loop_p (m_loop, gimple_bb (g: single_use))) |
507 | return false; |
508 | |
509 | /* Check the reduction operation. We require a left-associative operation. |
510 | For FP math we also need to be allowed to associate operations. */ |
511 | if (gassign *ass = dyn_cast <gassign *> (p: single_use)) |
512 | { |
513 | enum tree_code code = gimple_assign_rhs_code (gs: ass); |
514 | if (! (associative_tree_code (code) |
515 | || (code == MINUS_EXPR |
516 | && use_p->use == gimple_assign_rhs1_ptr (gs: ass))) |
517 | || (FLOAT_TYPE_P (TREE_TYPE (var)) |
518 | && ! flag_associative_math)) |
519 | return false; |
520 | } |
521 | else |
522 | return false; |
523 | |
524 | /* Handle and verify a series of stmts feeding the reduction op. */ |
525 | if (single_use != next_def |
526 | && !check_reduction_path (dump_user_location_t (), m_loop, phi, next, |
527 | gimple_assign_rhs_code (gs: single_use))) |
528 | return false; |
529 | |
530 | /* Only support cases in which INIT is used in inner loop. */ |
531 | if (TREE_CODE (init) == SSA_NAME) |
532 | FOR_EACH_IMM_USE_FAST (use_p, iterator, init) |
533 | { |
534 | stmt = USE_STMT (use_p); |
535 | if (is_gimple_debug (gs: stmt)) |
536 | continue; |
537 | |
538 | if (!flow_bb_inside_loop_p (m_loop, gimple_bb (g: stmt))) |
539 | return false; |
540 | } |
541 | |
542 | FOR_EACH_IMM_USE_FAST (use_p, iterator, next) |
543 | { |
544 | stmt = USE_STMT (use_p); |
545 | if (is_gimple_debug (gs: stmt)) |
546 | continue; |
547 | |
548 | /* Or else it's used in PHI itself. */ |
549 | use_phi = dyn_cast <gphi *> (p: stmt); |
550 | if (use_phi == phi) |
551 | continue; |
552 | |
553 | if (use_phi != NULL |
554 | && lcssa_phi == NULL |
555 | && gimple_bb (g: stmt) == m_exit->dest |
556 | && PHI_ARG_DEF_FROM_EDGE (use_phi, m_exit) == next) |
557 | lcssa_phi = use_phi; |
558 | else |
559 | return false; |
560 | } |
561 | if (!lcssa_phi) |
562 | return false; |
563 | |
564 | re = XCNEW (struct reduction); |
565 | re->var = var; |
566 | re->init = init; |
567 | re->next = next; |
568 | re->phi = phi; |
569 | re->lcssa_phi = lcssa_phi; |
570 | |
571 | classify_simple_reduction (re); |
572 | |
573 | if (dump_file && (dump_flags & TDF_DETAILS)) |
574 | dump_reduction (re); |
575 | |
576 | m_reductions.safe_push (obj: re); |
577 | return true; |
578 | } |
579 | |
580 | /* Analyze reduction variable VAR for outer loop of the loop nest to be |
581 | interchanged. ILOOP is not NULL and points to inner loop. For the |
582 | moment, we only support double reduction for outer loop, like: |
583 | |
584 | for (int i = 0; i < n; i++) |
585 | { |
586 | int sum = 0; |
587 | |
588 | for (int j = 0; j < n; j++) // outer loop |
589 | for (int k = 0; k < n; k++) // inner loop |
590 | sum += a[i][k]*b[k][j]; |
591 | |
592 | s[i] = sum; |
593 | } |
594 | |
595 | Note the innermost two loops are the loop nest to be interchanged. |
596 | Return true if analysis succeeds. */ |
597 | |
598 | bool |
599 | loop_cand::analyze_oloop_reduction_var (loop_cand *iloop, tree var) |
600 | { |
601 | gphi *phi = as_a <gphi *> (SSA_NAME_DEF_STMT (var)); |
602 | gphi *lcssa_phi = NULL, *use_phi; |
603 | tree init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (m_loop)); |
604 | tree next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (m_loop)); |
605 | reduction_p re; |
606 | gimple *stmt, *next_def; |
607 | use_operand_p use_p; |
608 | imm_use_iterator iterator; |
609 | |
610 | if (TREE_CODE (next) != SSA_NAME) |
611 | return false; |
612 | |
613 | next_def = SSA_NAME_DEF_STMT (next); |
614 | basic_block bb = gimple_bb (g: next_def); |
615 | if (!bb || !flow_bb_inside_loop_p (m_loop, bb)) |
616 | return false; |
617 | |
618 | /* Find inner loop's simple reduction that uses var as initializer. */ |
619 | reduction_p inner_re = NULL; |
620 | for (unsigned i = 0; iloop->m_reductions.iterate (ix: i, ptr: &inner_re); ++i) |
621 | if (inner_re->init == var || operand_equal_p (inner_re->init, var, flags: 0)) |
622 | break; |
623 | |
624 | if (inner_re == NULL |
625 | || inner_re->type != UNKNOWN_RTYPE |
626 | || inner_re->producer != phi) |
627 | return false; |
628 | |
629 | /* In case of double reduction, outer loop's reduction should be updated |
630 | by inner loop's simple reduction. */ |
631 | if (next_def != inner_re->lcssa_phi) |
632 | return false; |
633 | |
634 | /* Outer loop's reduction should only be used to initialize inner loop's |
635 | simple reduction. */ |
636 | if (! single_imm_use (var, use_p: &use_p, stmt: &stmt) |
637 | || stmt != inner_re->phi) |
638 | return false; |
639 | |
640 | /* Check this reduction is correctly used outside of loop via lcssa phi. */ |
641 | FOR_EACH_IMM_USE_FAST (use_p, iterator, next) |
642 | { |
643 | stmt = USE_STMT (use_p); |
644 | if (is_gimple_debug (gs: stmt)) |
645 | continue; |
646 | |
647 | /* Or else it's used in PHI itself. */ |
648 | use_phi = dyn_cast <gphi *> (p: stmt); |
649 | if (use_phi == phi) |
650 | continue; |
651 | |
652 | if (lcssa_phi == NULL |
653 | && use_phi != NULL |
654 | && gimple_bb (g: stmt) == m_exit->dest |
655 | && PHI_ARG_DEF_FROM_EDGE (use_phi, m_exit) == next) |
656 | lcssa_phi = use_phi; |
657 | else |
658 | return false; |
659 | } |
660 | if (!lcssa_phi) |
661 | return false; |
662 | |
663 | re = XCNEW (struct reduction); |
664 | re->var = var; |
665 | re->init = init; |
666 | re->next = next; |
667 | re->phi = phi; |
668 | re->lcssa_phi = lcssa_phi; |
669 | re->type = DOUBLE_RTYPE; |
670 | inner_re->type = DOUBLE_RTYPE; |
671 | |
672 | if (dump_file && (dump_flags & TDF_DETAILS)) |
673 | dump_reduction (re); |
674 | |
675 | m_reductions.safe_push (obj: re); |
676 | return true; |
677 | } |
678 | |
679 | /* Return true if VAR is induction variable of current loop whose scev is |
680 | specified by CHREC. */ |
681 | |
682 | bool |
683 | loop_cand::analyze_induction_var (tree var, tree chrec) |
684 | { |
685 | gphi *phi = as_a <gphi *> (SSA_NAME_DEF_STMT (var)); |
686 | tree init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (m_loop)); |
687 | |
688 | /* Var is loop invariant, though it's unlikely to happen. */ |
689 | if (tree_does_not_contain_chrecs (expr: chrec)) |
690 | { |
691 | /* Punt on floating point invariants if honoring signed zeros, |
692 | representing that as + 0.0 would change the result if init |
693 | is -0.0. Similarly for SNaNs it can raise exception. */ |
694 | if (HONOR_SIGNED_ZEROS (chrec) || HONOR_SNANS (chrec)) |
695 | return false; |
696 | struct induction *iv = XCNEW (struct induction); |
697 | iv->var = var; |
698 | iv->init_val = init; |
699 | iv->init_expr = chrec; |
700 | iv->step = build_zero_cst (TREE_TYPE (chrec)); |
701 | m_inductions.safe_push (obj: iv); |
702 | return true; |
703 | } |
704 | |
705 | if (TREE_CODE (chrec) != POLYNOMIAL_CHREC |
706 | || CHREC_VARIABLE (chrec) != (unsigned) m_loop->num |
707 | || tree_contains_chrecs (CHREC_LEFT (chrec), NULL) |
708 | || tree_contains_chrecs (CHREC_RIGHT (chrec), NULL)) |
709 | return false; |
710 | |
711 | struct induction *iv = XCNEW (struct induction); |
712 | iv->var = var; |
713 | iv->init_val = init; |
714 | iv->init_expr = CHREC_LEFT (chrec); |
715 | iv->step = CHREC_RIGHT (chrec); |
716 | |
717 | if (dump_file && (dump_flags & TDF_DETAILS)) |
718 | dump_induction (loop: m_loop, iv); |
719 | |
720 | m_inductions.safe_push (obj: iv); |
721 | return true; |
722 | } |
723 | |
724 | /* Return true if all loop carried variables defined in loop header can |
725 | be successfully analyzed. */ |
726 | |
727 | bool |
728 | loop_cand::analyze_carried_vars (loop_cand *iloop) |
729 | { |
730 | edge e = loop_preheader_edge (m_outer); |
731 | gphi_iterator gsi; |
732 | |
733 | if (dump_file && (dump_flags & TDF_DETAILS)) |
734 | fprintf (stream: dump_file, format: "\nLoop(%d) carried vars:\n" , m_loop->num); |
735 | |
736 | for (gsi = gsi_start_phis (m_loop->header); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
737 | { |
738 | gphi *phi = gsi.phi (); |
739 | |
740 | tree var = PHI_RESULT (phi); |
741 | if (virtual_operand_p (op: var)) |
742 | continue; |
743 | |
744 | tree chrec = analyze_scalar_evolution (m_loop, var); |
745 | chrec = instantiate_scev (e, m_loop, chrec); |
746 | |
747 | /* Analyze var as reduction variable. */ |
748 | if (chrec_contains_undetermined (chrec) |
749 | || chrec_contains_symbols_defined_in_loop (chrec, m_outer->num)) |
750 | { |
751 | if (iloop && !analyze_oloop_reduction_var (iloop, var)) |
752 | return false; |
753 | if (!iloop && !analyze_iloop_reduction_var (var)) |
754 | return false; |
755 | } |
756 | /* Analyze var as induction variable. */ |
757 | else if (!analyze_induction_var (var, chrec)) |
758 | return false; |
759 | } |
760 | |
761 | return true; |
762 | } |
763 | |
764 | /* Return TRUE if loop closed PHI nodes can be analyzed successfully. */ |
765 | |
766 | bool |
767 | loop_cand::analyze_lcssa_phis (void) |
768 | { |
769 | gphi_iterator gsi; |
770 | for (gsi = gsi_start_phis (m_exit->dest); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
771 | { |
772 | gphi *phi = gsi.phi (); |
773 | |
774 | if (virtual_operand_p (PHI_RESULT (phi))) |
775 | continue; |
776 | |
777 | /* TODO: We only support lcssa phi for reduction for now. */ |
778 | if (!find_reduction_by_stmt (stmt: phi)) |
779 | return false; |
780 | } |
781 | |
782 | return true; |
783 | } |
784 | |
785 | /* CONSUMER is a stmt in BB storing reduction result into memory object. |
786 | When the reduction is intialized from constant value, we need to add |
787 | a stmt loading from the memory object to target basic block in inner |
788 | loop during undoing the reduction. Problem is that memory reference |
789 | may use ssa variables not dominating the target basic block. This |
790 | function finds all stmts on which CONSUMER depends in basic block BB, |
791 | records and returns them via STMTS. */ |
792 | |
793 | static void |
794 | find_deps_in_bb_for_stmt (gimple_seq *stmts, basic_block bb, gimple *consumer) |
795 | { |
796 | auto_vec<gimple *, 4> worklist; |
797 | use_operand_p use_p; |
798 | ssa_op_iter iter; |
799 | gimple *stmt, *def_stmt; |
800 | gimple_stmt_iterator gsi; |
801 | |
802 | /* First clear flag for stmts in bb. */ |
803 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
804 | gimple_set_plf (stmt: gsi_stmt (i: gsi), plf: GF_PLF_1, val_p: false); |
805 | |
806 | /* DFS search all depended stmts in bb and mark flag for these stmts. */ |
807 | worklist.safe_push (obj: consumer); |
808 | while (!worklist.is_empty ()) |
809 | { |
810 | stmt = worklist.pop (); |
811 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE) |
812 | { |
813 | def_stmt = SSA_NAME_DEF_STMT (USE_FROM_PTR (use_p)); |
814 | |
815 | if (is_a <gphi *> (p: def_stmt) |
816 | || gimple_bb (g: def_stmt) != bb |
817 | || gimple_plf (stmt: def_stmt, plf: GF_PLF_1)) |
818 | continue; |
819 | |
820 | worklist.safe_push (obj: def_stmt); |
821 | } |
822 | gimple_set_plf (stmt, plf: GF_PLF_1, val_p: true); |
823 | } |
824 | for (gsi = gsi_start_nondebug_bb (bb); |
825 | !gsi_end_p (i: gsi) && (stmt = gsi_stmt (i: gsi)) != consumer;) |
826 | { |
827 | /* Move dep stmts to sequence STMTS. */ |
828 | if (gimple_plf (stmt, plf: GF_PLF_1)) |
829 | { |
830 | gsi_remove (&gsi, false); |
831 | gimple_seq_add_stmt_without_update (stmts, stmt); |
832 | } |
833 | else |
834 | gsi_next_nondebug (i: &gsi); |
835 | } |
836 | } |
837 | |
838 | /* User can write, optimizers can generate simple reduction RE for inner |
839 | loop. In order to make interchange valid, we have to undo reduction by |
840 | moving producer and consumer stmts into the inner loop. For example, |
841 | below code: |
842 | |
843 | init = MEM_REF[idx]; //producer |
844 | loop: |
845 | var = phi<init, next> |
846 | next = var op ... |
847 | reduc_sum = phi<next> |
848 | MEM_REF[idx] = reduc_sum //consumer |
849 | |
850 | is transformed into: |
851 | |
852 | loop: |
853 | new_var = MEM_REF[idx]; //producer after moving |
854 | next = new_var op ... |
855 | MEM_REF[idx] = next; //consumer after moving |
856 | |
857 | Note if the reduction variable is initialized to constant, like: |
858 | |
859 | var = phi<0.0, next> |
860 | |
861 | we compute new_var as below: |
862 | |
863 | loop: |
864 | tmp = MEM_REF[idx]; |
865 | new_var = !first_iteration ? tmp : 0.0; |
866 | |
867 | so that the initial const is used in the first iteration of loop. Also |
868 | record ssa variables for dead code elimination in DCE_SEEDS. */ |
869 | |
870 | void |
871 | loop_cand::undo_simple_reduction (reduction_p re, bitmap dce_seeds) |
872 | { |
873 | gimple *stmt; |
874 | gimple_stmt_iterator from, to = gsi_after_labels (bb: m_loop->header); |
875 | gimple_seq stmts = NULL; |
876 | tree new_var; |
877 | |
878 | /* Prepare the initialization stmts and insert it to inner loop. */ |
879 | if (re->producer != NULL) |
880 | { |
881 | gimple_set_vuse (g: re->producer, NULL_TREE); |
882 | update_stmt (s: re->producer); |
883 | from = gsi_for_stmt (re->producer); |
884 | gsi_remove (&from, false); |
885 | gimple_seq_add_stmt_without_update (&stmts, re->producer); |
886 | new_var = re->init; |
887 | } |
888 | else |
889 | { |
890 | /* Find all stmts on which expression "MEM_REF[idx]" depends. */ |
891 | find_deps_in_bb_for_stmt (stmts: &stmts, bb: gimple_bb (g: re->consumer), consumer: re->consumer); |
892 | /* Because we generate new stmt loading from the MEM_REF to TMP. */ |
893 | tree cond, tmp = copy_ssa_name (var: re->var); |
894 | stmt = gimple_build_assign (tmp, re->init_ref); |
895 | gimple_seq_add_stmt_without_update (&stmts, stmt); |
896 | |
897 | /* Init new_var to MEM_REF or CONST depending on if it is the first |
898 | iteration. */ |
899 | induction_p iv = m_inductions[0]; |
900 | cond = make_ssa_name (boolean_type_node); |
901 | stmt = gimple_build_assign (cond, NE_EXPR, iv->var, iv->init_val); |
902 | gimple_seq_add_stmt_without_update (&stmts, stmt); |
903 | new_var = copy_ssa_name (var: re->var); |
904 | stmt = gimple_build_assign (new_var, COND_EXPR, cond, tmp, re->init); |
905 | gimple_seq_add_stmt_without_update (&stmts, stmt); |
906 | } |
907 | gsi_insert_seq_before (&to, stmts, GSI_SAME_STMT); |
908 | |
909 | /* Replace all uses of reduction var with new variable. */ |
910 | use_operand_p use_p; |
911 | imm_use_iterator iterator; |
912 | FOR_EACH_IMM_USE_STMT (stmt, iterator, re->var) |
913 | { |
914 | FOR_EACH_IMM_USE_ON_STMT (use_p, iterator) |
915 | SET_USE (use_p, new_var); |
916 | |
917 | update_stmt (s: stmt); |
918 | } |
919 | |
920 | /* Move consumer stmt into inner loop, just after reduction next's def. */ |
921 | unlink_stmt_vdef (re->consumer); |
922 | release_ssa_name (name: gimple_vdef (g: re->consumer)); |
923 | gimple_set_vdef (g: re->consumer, NULL_TREE); |
924 | gimple_set_vuse (g: re->consumer, NULL_TREE); |
925 | gimple_assign_set_rhs1 (gs: re->consumer, rhs: re->next); |
926 | update_stmt (s: re->consumer); |
927 | from = gsi_for_stmt (re->consumer); |
928 | to = gsi_for_stmt (SSA_NAME_DEF_STMT (re->next)); |
929 | gsi_move_after (&from, &to); |
930 | |
931 | /* Mark the reduction variables for DCE. */ |
932 | bitmap_set_bit (dce_seeds, SSA_NAME_VERSION (re->var)); |
933 | bitmap_set_bit (dce_seeds, SSA_NAME_VERSION (PHI_RESULT (re->lcssa_phi))); |
934 | } |
935 | |
936 | /* Free DATAREFS and its auxiliary memory. */ |
937 | |
938 | static void |
939 | free_data_refs_with_aux (vec<data_reference_p> datarefs) |
940 | { |
941 | data_reference_p dr; |
942 | for (unsigned i = 0; datarefs.iterate (ix: i, ptr: &dr); ++i) |
943 | if (dr->aux != NULL) |
944 | { |
945 | DR_ACCESS_STRIDE (dr)->release (); |
946 | delete (vec<tree> *) dr->aux; |
947 | } |
948 | |
949 | free_data_refs (datarefs); |
950 | } |
951 | |
952 | /* Class for loop interchange transformation. */ |
953 | |
954 | class tree_loop_interchange |
955 | { |
956 | public: |
957 | tree_loop_interchange (vec<class loop *> loop_nest) |
958 | : m_loop_nest (loop_nest), m_niters_iv_var (NULL_TREE), |
959 | m_dce_seeds (BITMAP_ALLOC (NULL)) { } |
960 | ~tree_loop_interchange () { BITMAP_FREE (m_dce_seeds); } |
961 | bool interchange (vec<data_reference_p>, vec<ddr_p>); |
962 | |
963 | private: |
964 | void update_data_info (unsigned, unsigned, vec<data_reference_p>, vec<ddr_p>); |
965 | bool valid_data_dependences (unsigned, unsigned, vec<ddr_p>); |
966 | void interchange_loops (loop_cand &, loop_cand &); |
967 | void map_inductions_to_loop (loop_cand &, loop_cand &); |
968 | void move_code_to_inner_loop (class loop *, class loop *, basic_block *); |
969 | |
970 | /* The whole loop nest in which interchange is ongoing. */ |
971 | vec<class loop *> m_loop_nest; |
972 | /* We create new IV which is only used in loop's exit condition check. |
973 | In case of 3-level loop nest interchange, when we interchange the |
974 | innermost two loops, new IV created in the middle level loop does |
975 | not need to be preserved in interchanging the outermost two loops |
976 | later. We record the IV so that it can be skipped. */ |
977 | tree m_niters_iv_var; |
978 | /* Bitmap of seed variables for dead code elimination after interchange. */ |
979 | bitmap m_dce_seeds; |
980 | }; |
981 | |
982 | /* Update data refs' access stride and dependence information after loop |
983 | interchange. I_IDX/O_IDX gives indices of interchanged loops in loop |
984 | nest. DATAREFS are data references. DDRS are data dependences. */ |
985 | |
986 | void |
987 | tree_loop_interchange::update_data_info (unsigned i_idx, unsigned o_idx, |
988 | vec<data_reference_p> datarefs, |
989 | vec<ddr_p> ddrs) |
990 | { |
991 | struct data_reference *dr; |
992 | struct data_dependence_relation *ddr; |
993 | |
994 | /* Update strides of data references. */ |
995 | for (unsigned i = 0; datarefs.iterate (ix: i, ptr: &dr); ++i) |
996 | { |
997 | vec<tree> *stride = DR_ACCESS_STRIDE (dr); |
998 | gcc_assert (stride->length () > i_idx); |
999 | std::swap (a&: (*stride)[i_idx], b&: (*stride)[o_idx]); |
1000 | } |
1001 | /* Update data dependences. */ |
1002 | for (unsigned i = 0; ddrs.iterate (ix: i, ptr: &ddr); ++i) |
1003 | if (DDR_ARE_DEPENDENT (ddr) != chrec_known) |
1004 | { |
1005 | for (unsigned j = 0; j < DDR_NUM_DIST_VECTS (ddr); ++j) |
1006 | { |
1007 | lambda_vector dist_vect = DDR_DIST_VECT (ddr, j); |
1008 | std::swap (a&: dist_vect[i_idx], b&: dist_vect[o_idx]); |
1009 | } |
1010 | } |
1011 | } |
1012 | |
1013 | /* Check data dependence relations, return TRUE if it's valid to interchange |
1014 | two loops specified by I_IDX/O_IDX. Theoretically, interchanging the two |
1015 | loops is valid only if dist vector, after interchanging, doesn't have '>' |
1016 | as the leftmost non-'=' direction. Practically, this function have been |
1017 | conservative here by not checking some valid cases. */ |
1018 | |
1019 | bool |
1020 | tree_loop_interchange::valid_data_dependences (unsigned i_idx, unsigned o_idx, |
1021 | vec<ddr_p> ddrs) |
1022 | { |
1023 | struct data_dependence_relation *ddr; |
1024 | |
1025 | for (unsigned i = 0; ddrs.iterate (ix: i, ptr: &ddr); ++i) |
1026 | { |
1027 | /* Skip no-dependence case. */ |
1028 | if (DDR_ARE_DEPENDENT (ddr) == chrec_known) |
1029 | continue; |
1030 | |
1031 | for (unsigned j = 0; j < DDR_NUM_DIST_VECTS (ddr); ++j) |
1032 | { |
1033 | lambda_vector dist_vect = DDR_DIST_VECT (ddr, j); |
1034 | unsigned level = dependence_level (dist_vect, length: m_loop_nest.length ()); |
1035 | |
1036 | /* If there is no carried dependence. */ |
1037 | if (level == 0) |
1038 | continue; |
1039 | |
1040 | level --; |
1041 | |
1042 | /* If dependence is not carried by any loop in between the two |
1043 | loops [oloop, iloop] to interchange. */ |
1044 | if (level < o_idx || level > i_idx) |
1045 | continue; |
1046 | |
1047 | /* Be conservative, skip case if either direction at i_idx/o_idx |
1048 | levels is not '=' or '<'. */ |
1049 | if ((!DDR_REVERSED_P (ddr) && dist_vect[i_idx] < 0) |
1050 | || (DDR_REVERSED_P (ddr) && dist_vect[i_idx] > 0) |
1051 | || (!DDR_REVERSED_P (ddr) && dist_vect[o_idx] < 0) |
1052 | || (DDR_REVERSED_P (ddr) && dist_vect[o_idx] > 0)) |
1053 | return false; |
1054 | } |
1055 | } |
1056 | |
1057 | return true; |
1058 | } |
1059 | |
1060 | /* Interchange two loops specified by ILOOP and OLOOP. */ |
1061 | |
1062 | void |
1063 | tree_loop_interchange::interchange_loops (loop_cand &iloop, loop_cand &oloop) |
1064 | { |
1065 | reduction_p re; |
1066 | gimple_stmt_iterator gsi; |
1067 | tree i_niters, o_niters, var_after; |
1068 | |
1069 | /* Undo inner loop's simple reduction. */ |
1070 | for (unsigned i = 0; iloop.m_reductions.iterate (ix: i, ptr: &re); ++i) |
1071 | if (re->type != DOUBLE_RTYPE) |
1072 | { |
1073 | if (re->producer) |
1074 | reset_debug_uses (re->producer); |
1075 | |
1076 | iloop.undo_simple_reduction (re, dce_seeds: m_dce_seeds); |
1077 | } |
1078 | |
1079 | /* Only need to reset debug uses for double reduction. */ |
1080 | for (unsigned i = 0; oloop.m_reductions.iterate (ix: i, ptr: &re); ++i) |
1081 | { |
1082 | gcc_assert (re->type == DOUBLE_RTYPE); |
1083 | reset_debug_uses (SSA_NAME_DEF_STMT (re->var)); |
1084 | reset_debug_uses (SSA_NAME_DEF_STMT (re->next)); |
1085 | } |
1086 | |
1087 | /* Prepare niters for both loops. */ |
1088 | class loop *loop_nest = m_loop_nest[0]; |
1089 | edge instantiate_below = loop_preheader_edge (loop_nest); |
1090 | gsi = gsi_last_bb (bb: loop_preheader_edge (loop_nest)->src); |
1091 | i_niters = number_of_latch_executions (iloop.m_loop); |
1092 | i_niters = analyze_scalar_evolution (loop_outer (loop: iloop.m_loop), i_niters); |
1093 | i_niters = instantiate_scev (instantiate_below, loop_outer (loop: iloop.m_loop), |
1094 | i_niters); |
1095 | i_niters = force_gimple_operand_gsi (&gsi, unshare_expr (i_niters), true, |
1096 | NULL_TREE, false, GSI_CONTINUE_LINKING); |
1097 | o_niters = number_of_latch_executions (oloop.m_loop); |
1098 | if (oloop.m_loop != loop_nest) |
1099 | { |
1100 | o_niters = analyze_scalar_evolution (loop_outer (loop: oloop.m_loop), o_niters); |
1101 | o_niters = instantiate_scev (instantiate_below, loop_outer (loop: oloop.m_loop), |
1102 | o_niters); |
1103 | } |
1104 | o_niters = force_gimple_operand_gsi (&gsi, unshare_expr (o_niters), true, |
1105 | NULL_TREE, false, GSI_CONTINUE_LINKING); |
1106 | |
1107 | /* Move src's code to tgt loop. This is necessary when src is the outer |
1108 | loop and tgt is the inner loop. */ |
1109 | move_code_to_inner_loop (oloop.m_loop, iloop.m_loop, oloop.m_bbs); |
1110 | |
1111 | /* Map outer loop's IV to inner loop, and vice versa. */ |
1112 | map_inductions_to_loop (oloop, iloop); |
1113 | map_inductions_to_loop (iloop, oloop); |
1114 | |
1115 | /* Create canonical IV for both loops. Note canonical IV for outer/inner |
1116 | loop is actually from inner/outer loop. Also we record the new IV |
1117 | created for the outer loop so that it can be skipped in later loop |
1118 | interchange. */ |
1119 | create_canonical_iv (oloop.m_loop, oloop.m_exit, |
1120 | i_niters, &m_niters_iv_var, &var_after); |
1121 | bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_after)); |
1122 | create_canonical_iv (iloop.m_loop, iloop.m_exit, |
1123 | o_niters, NULL, &var_after); |
1124 | bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_after)); |
1125 | |
1126 | /* Scrap niters estimation of interchanged loops. */ |
1127 | iloop.m_loop->any_upper_bound = false; |
1128 | iloop.m_loop->any_likely_upper_bound = false; |
1129 | free_numbers_of_iterations_estimates (iloop.m_loop); |
1130 | oloop.m_loop->any_upper_bound = false; |
1131 | oloop.m_loop->any_likely_upper_bound = false; |
1132 | free_numbers_of_iterations_estimates (oloop.m_loop); |
1133 | |
1134 | /* Clear all cached scev information. This is expensive but shouldn't be |
1135 | a problem given we interchange in very limited times. */ |
1136 | scev_reset_htab (); |
1137 | |
1138 | /* ??? The association between the loop data structure and the |
1139 | CFG changed, so what was loop N at the source level is now |
1140 | loop M. We should think of retaining the association or breaking |
1141 | it fully by creating a new loop instead of re-using the "wrong" one. */ |
1142 | } |
1143 | |
1144 | /* Map induction variables of SRC loop to TGT loop. The function firstly |
1145 | creates the same IV of SRC loop in TGT loop, then deletes the original |
1146 | IV and re-initialize it using the newly created IV. For example, loop |
1147 | nest: |
1148 | |
1149 | for (i = 0; i < N; i++) |
1150 | for (j = 0; j < M; j++) |
1151 | { |
1152 | //use of i; |
1153 | //use of j; |
1154 | } |
1155 | |
1156 | will be transformed into: |
1157 | |
1158 | for (jj = 0; jj < M; jj++) |
1159 | for (ii = 0; ii < N; ii++) |
1160 | { |
1161 | //use of ii; |
1162 | //use of jj; |
1163 | } |
1164 | |
1165 | after loop interchange. */ |
1166 | |
1167 | void |
1168 | tree_loop_interchange::map_inductions_to_loop (loop_cand &src, loop_cand &tgt) |
1169 | { |
1170 | induction_p iv; |
1171 | edge e = tgt.m_exit; |
1172 | gimple_stmt_iterator incr_pos = gsi_last_bb (bb: e->src), gsi; |
1173 | |
1174 | /* Map source loop's IV to target loop. */ |
1175 | for (unsigned i = 0; src.m_inductions.iterate (ix: i, ptr: &iv); ++i) |
1176 | { |
1177 | gimple *use_stmt, *stmt = SSA_NAME_DEF_STMT (iv->var); |
1178 | gcc_assert (is_a <gphi *> (stmt)); |
1179 | |
1180 | use_operand_p use_p; |
1181 | /* Only map original IV to target loop. */ |
1182 | if (m_niters_iv_var != iv->var) |
1183 | { |
1184 | /* Map the IV by creating the same one in target loop. */ |
1185 | tree var_before, var_after; |
1186 | tree base = unshare_expr (iv->init_expr); |
1187 | tree step = unshare_expr (iv->step); |
1188 | create_iv (base, PLUS_EXPR, step, SSA_NAME_VAR (iv->var), |
1189 | tgt.m_loop, &incr_pos, false, &var_before, &var_after); |
1190 | bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_before)); |
1191 | bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_after)); |
1192 | |
1193 | /* Replace uses of the original IV var with newly created IV var. */ |
1194 | imm_use_iterator imm_iter; |
1195 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, iv->var) |
1196 | { |
1197 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) |
1198 | SET_USE (use_p, var_before); |
1199 | |
1200 | update_stmt (s: use_stmt); |
1201 | } |
1202 | } |
1203 | |
1204 | /* Mark all uses for DCE. */ |
1205 | ssa_op_iter op_iter; |
1206 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, op_iter, SSA_OP_USE) |
1207 | { |
1208 | tree use = USE_FROM_PTR (use_p); |
1209 | if (TREE_CODE (use) == SSA_NAME |
1210 | && ! SSA_NAME_IS_DEFAULT_DEF (use)) |
1211 | bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (use)); |
1212 | } |
1213 | |
1214 | /* Delete definition of the original IV in the source loop. */ |
1215 | gsi = gsi_for_stmt (stmt); |
1216 | remove_phi_node (&gsi, true); |
1217 | } |
1218 | } |
1219 | |
1220 | /* Move stmts of outer loop to inner loop. */ |
1221 | |
1222 | void |
1223 | tree_loop_interchange::move_code_to_inner_loop (class loop *outer, |
1224 | class loop *inner, |
1225 | basic_block *outer_bbs) |
1226 | { |
1227 | basic_block oloop_exit_bb = single_exit (outer)->src; |
1228 | gimple_stmt_iterator gsi, to; |
1229 | |
1230 | for (unsigned i = 0; i < outer->num_nodes; i++) |
1231 | { |
1232 | basic_block bb = outer_bbs[i]; |
1233 | |
1234 | /* Skip basic blocks of inner loop. */ |
1235 | if (flow_bb_inside_loop_p (inner, bb)) |
1236 | continue; |
1237 | |
1238 | /* Move code from header/latch to header/latch. */ |
1239 | if (bb == outer->header) |
1240 | to = gsi_after_labels (bb: inner->header); |
1241 | else if (bb == outer->latch) |
1242 | to = gsi_after_labels (bb: inner->latch); |
1243 | else |
1244 | /* Otherwise, simply move to exit->src. */ |
1245 | to = gsi_last_bb (bb: single_exit (inner)->src); |
1246 | |
1247 | for (gsi = gsi_after_labels (bb); !gsi_end_p (i: gsi);) |
1248 | { |
1249 | gimple *stmt = gsi_stmt (i: gsi); |
1250 | |
1251 | if (oloop_exit_bb == bb |
1252 | && stmt == gsi_stmt (i: gsi_last_bb (bb: oloop_exit_bb))) |
1253 | { |
1254 | gsi_next (i: &gsi); |
1255 | continue; |
1256 | } |
1257 | |
1258 | if (gimple_vdef (g: stmt)) |
1259 | { |
1260 | unlink_stmt_vdef (stmt); |
1261 | release_ssa_name (name: gimple_vdef (g: stmt)); |
1262 | gimple_set_vdef (g: stmt, NULL_TREE); |
1263 | } |
1264 | if (gimple_vuse (g: stmt)) |
1265 | { |
1266 | gimple_set_vuse (g: stmt, NULL_TREE); |
1267 | update_stmt (s: stmt); |
1268 | } |
1269 | |
1270 | reset_debug_uses (stmt); |
1271 | gsi_move_before (&gsi, &to); |
1272 | } |
1273 | } |
1274 | } |
1275 | |
1276 | /* Given data reference DR in LOOP_NEST, the function computes DR's access |
1277 | stride at each level of loop from innermost LOOP to outer. On success, |
1278 | it saves access stride at each level loop in a vector which is pointed |
1279 | by DR->aux. For example: |
1280 | |
1281 | int arr[100][100][100]; |
1282 | for (i = 0; i < 100; i++) ;(DR->aux)strides[0] = 40000 |
1283 | for (j = 100; j > 0; j--) ;(DR->aux)strides[1] = 400 |
1284 | for (k = 0; k < 100; k++) ;(DR->aux)strides[2] = 4 |
1285 | arr[i][j - 1][k] = 0; */ |
1286 | |
1287 | static void |
1288 | compute_access_stride (class loop *&loop_nest, class loop *loop, |
1289 | data_reference_p dr) |
1290 | { |
1291 | vec<tree> *strides = new vec<tree> (); |
1292 | dr->aux = strides; |
1293 | |
1294 | basic_block bb = gimple_bb (DR_STMT (dr)); |
1295 | if (!flow_bb_inside_loop_p (loop_nest, bb)) |
1296 | return; |
1297 | while (!flow_bb_inside_loop_p (loop, bb)) |
1298 | { |
1299 | strides->safe_push (obj: build_int_cst (sizetype, 0)); |
1300 | loop = loop_outer (loop); |
1301 | } |
1302 | gcc_assert (loop == bb->loop_father); |
1303 | |
1304 | tree ref = DR_REF (dr); |
1305 | if (TREE_CODE (ref) == COMPONENT_REF |
1306 | && DECL_BIT_FIELD (TREE_OPERAND (ref, 1))) |
1307 | { |
1308 | /* We can't take address of bitfields. If the bitfield is at constant |
1309 | offset from the start of the struct, just use address of the |
1310 | struct, for analysis of the strides that shouldn't matter. */ |
1311 | if (!TREE_OPERAND (ref, 2) |
1312 | || TREE_CODE (TREE_OPERAND (ref, 2)) == INTEGER_CST) |
1313 | ref = TREE_OPERAND (ref, 0); |
1314 | /* Otherwise, if we have a bit field representative, use that. */ |
1315 | else if (DECL_BIT_FIELD_REPRESENTATIVE (TREE_OPERAND (ref, 1)) |
1316 | != NULL_TREE) |
1317 | { |
1318 | tree repr = DECL_BIT_FIELD_REPRESENTATIVE (TREE_OPERAND (ref, 1)); |
1319 | ref = build3 (COMPONENT_REF, TREE_TYPE (repr), TREE_OPERAND (ref, 0), |
1320 | repr, TREE_OPERAND (ref, 2)); |
1321 | } |
1322 | /* Otherwise punt. */ |
1323 | else |
1324 | return; |
1325 | } |
1326 | tree scev_base = build_fold_addr_expr (ref); |
1327 | tree scev = analyze_scalar_evolution (loop, scev_base); |
1328 | if (chrec_contains_undetermined (scev)) |
1329 | return; |
1330 | |
1331 | tree orig_scev = scev; |
1332 | do |
1333 | { |
1334 | scev = instantiate_scev (loop_preheader_edge (loop_nest), |
1335 | loop, orig_scev); |
1336 | if (! chrec_contains_undetermined (scev)) |
1337 | break; |
1338 | |
1339 | /* If we couldn't instantiate for the desired nest, shrink it. */ |
1340 | if (loop_nest == loop) |
1341 | return; |
1342 | loop_nest = loop_nest->inner; |
1343 | } while (1); |
1344 | |
1345 | tree sl = scev; |
1346 | class loop *expected = loop; |
1347 | while (TREE_CODE (sl) == POLYNOMIAL_CHREC) |
1348 | { |
1349 | class loop *sl_loop = get_chrec_loop (chrec: sl); |
1350 | while (sl_loop != expected) |
1351 | { |
1352 | strides->safe_push (size_int (0)); |
1353 | expected = loop_outer (loop: expected); |
1354 | } |
1355 | strides->safe_push (CHREC_RIGHT (sl)); |
1356 | sl = CHREC_LEFT (sl); |
1357 | expected = loop_outer (loop: expected); |
1358 | } |
1359 | if (! tree_contains_chrecs (sl, NULL)) |
1360 | while (expected != loop_outer (loop: loop_nest)) |
1361 | { |
1362 | strides->safe_push (size_int (0)); |
1363 | expected = loop_outer (loop: expected); |
1364 | } |
1365 | } |
1366 | |
1367 | /* Given loop nest LOOP_NEST with innermost LOOP, the function computes |
1368 | access strides with respect to each level loop for all data refs in |
1369 | DATAREFS from inner loop to outer loop. On success, it returns the |
1370 | outermost loop that access strides can be computed successfully for |
1371 | all data references. If access strides cannot be computed at least |
1372 | for two levels of loop for any data reference, it returns NULL. */ |
1373 | |
1374 | static class loop * |
1375 | compute_access_strides (class loop *loop_nest, class loop *loop, |
1376 | vec<data_reference_p> datarefs) |
1377 | { |
1378 | unsigned i, j, num_loops = (unsigned) -1; |
1379 | data_reference_p dr; |
1380 | vec<tree> *stride; |
1381 | |
1382 | class loop *interesting_loop_nest = loop_nest; |
1383 | for (i = 0; datarefs.iterate (ix: i, ptr: &dr); ++i) |
1384 | { |
1385 | compute_access_stride (loop_nest&: interesting_loop_nest, loop, dr); |
1386 | stride = DR_ACCESS_STRIDE (dr); |
1387 | if (stride->length () < num_loops) |
1388 | { |
1389 | num_loops = stride->length (); |
1390 | if (num_loops < 2) |
1391 | return NULL; |
1392 | } |
1393 | } |
1394 | |
1395 | for (i = 0; datarefs.iterate (ix: i, ptr: &dr); ++i) |
1396 | { |
1397 | stride = DR_ACCESS_STRIDE (dr); |
1398 | if (stride->length () > num_loops) |
1399 | stride->truncate (size: num_loops); |
1400 | |
1401 | for (j = 0; j < (num_loops >> 1); ++j) |
1402 | std::swap (a&: (*stride)[j], b&: (*stride)[num_loops - j - 1]); |
1403 | } |
1404 | |
1405 | loop = superloop_at_depth (loop, loop_depth (loop) + 1 - num_loops); |
1406 | gcc_assert (loop_nest == loop || flow_loop_nested_p (loop_nest, loop)); |
1407 | return loop; |
1408 | } |
1409 | |
1410 | /* Prune access strides for data references in DATAREFS by removing strides |
1411 | of loops that isn't in current LOOP_NEST. */ |
1412 | |
1413 | static void |
1414 | prune_access_strides_not_in_loop (class loop *loop_nest, |
1415 | class loop *innermost, |
1416 | vec<data_reference_p> datarefs) |
1417 | { |
1418 | data_reference_p dr; |
1419 | unsigned num_loops = loop_depth (loop: innermost) - loop_depth (loop: loop_nest) + 1; |
1420 | gcc_assert (num_loops > 1); |
1421 | |
1422 | /* Block remove strides of loops that is not in current loop nest. */ |
1423 | for (unsigned i = 0; datarefs.iterate (ix: i, ptr: &dr); ++i) |
1424 | { |
1425 | vec<tree> *stride = DR_ACCESS_STRIDE (dr); |
1426 | if (stride->length () > num_loops) |
1427 | stride->block_remove (ix: 0, len: stride->length () - num_loops); |
1428 | } |
1429 | } |
1430 | |
1431 | /* Dump access strides for all DATAREFS. */ |
1432 | |
1433 | static void |
1434 | dump_access_strides (vec<data_reference_p> datarefs) |
1435 | { |
1436 | data_reference_p dr; |
1437 | fprintf (stream: dump_file, format: "Access Strides for DRs:\n" ); |
1438 | for (unsigned i = 0; datarefs.iterate (ix: i, ptr: &dr); ++i) |
1439 | { |
1440 | fprintf (stream: dump_file, format: " " ); |
1441 | print_generic_expr (dump_file, DR_REF (dr), TDF_SLIM); |
1442 | fprintf (stream: dump_file, format: ":\t\t<" ); |
1443 | |
1444 | vec<tree> *stride = DR_ACCESS_STRIDE (dr); |
1445 | unsigned num_loops = stride->length (); |
1446 | for (unsigned j = 0; j < num_loops; ++j) |
1447 | { |
1448 | print_generic_expr (dump_file, (*stride)[j], TDF_SLIM); |
1449 | fprintf (stream: dump_file, format: "%s" , (j < num_loops - 1) ? ",\t" : ">\n" ); |
1450 | } |
1451 | } |
1452 | } |
1453 | |
1454 | /* Return true if it's profitable to interchange two loops whose index |
1455 | in whole loop nest vector are I_IDX/O_IDX respectively. The function |
1456 | computes and compares three types information from all DATAREFS: |
1457 | 1) Access stride for loop I_IDX and O_IDX. |
1458 | 2) Number of invariant memory references with respect to I_IDX before |
1459 | and after loop interchange. |
1460 | 3) Flags indicating if all memory references access sequential memory |
1461 | in ILOOP, before and after loop interchange. |
1462 | If INNMOST_LOOP_P is true, the two loops for interchanging are the two |
1463 | innermost loops in loop nest. This function also dumps information if |
1464 | DUMP_INFO_P is true. */ |
1465 | |
1466 | static bool |
1467 | should_interchange_loops (unsigned i_idx, unsigned o_idx, |
1468 | vec<data_reference_p> datarefs, |
1469 | unsigned i_stmt_cost, unsigned o_stmt_cost, |
1470 | bool innermost_loops_p, bool dump_info_p = true) |
1471 | { |
1472 | unsigned HOST_WIDE_INT ratio; |
1473 | unsigned i, j, num_old_inv_drs = 0, num_new_inv_drs = 0; |
1474 | struct data_reference *dr; |
1475 | bool all_seq_dr_before_p = true, all_seq_dr_after_p = true; |
1476 | widest_int iloop_strides = 0, oloop_strides = 0; |
1477 | unsigned num_unresolved_drs = 0; |
1478 | unsigned num_resolved_ok_drs = 0; |
1479 | unsigned num_resolved_not_ok_drs = 0; |
1480 | |
1481 | if (dump_info_p && dump_file && (dump_flags & TDF_DETAILS)) |
1482 | fprintf (stream: dump_file, format: "\nData ref strides:\n\tmem_ref:\t\tiloop\toloop\n" ); |
1483 | |
1484 | for (i = 0; datarefs.iterate (ix: i, ptr: &dr); ++i) |
1485 | { |
1486 | vec<tree> *stride = DR_ACCESS_STRIDE (dr); |
1487 | tree iloop_stride = (*stride)[i_idx], oloop_stride = (*stride)[o_idx]; |
1488 | |
1489 | bool subloop_stride_p = false; |
1490 | /* Data ref can't be invariant or sequential access at current loop if |
1491 | its address changes with respect to any subloops. */ |
1492 | for (j = i_idx + 1; j < stride->length (); ++j) |
1493 | if (!integer_zerop ((*stride)[j])) |
1494 | { |
1495 | subloop_stride_p = true; |
1496 | break; |
1497 | } |
1498 | |
1499 | if (integer_zerop (iloop_stride)) |
1500 | { |
1501 | if (!subloop_stride_p) |
1502 | num_old_inv_drs++; |
1503 | } |
1504 | if (integer_zerop (oloop_stride)) |
1505 | { |
1506 | if (!subloop_stride_p) |
1507 | num_new_inv_drs++; |
1508 | } |
1509 | |
1510 | if (TREE_CODE (iloop_stride) == INTEGER_CST |
1511 | && TREE_CODE (oloop_stride) == INTEGER_CST) |
1512 | { |
1513 | iloop_strides = wi::add (x: iloop_strides, y: wi::to_widest (t: iloop_stride)); |
1514 | oloop_strides = wi::add (x: oloop_strides, y: wi::to_widest (t: oloop_stride)); |
1515 | } |
1516 | else if (multiple_of_p (TREE_TYPE (iloop_stride), |
1517 | iloop_stride, oloop_stride)) |
1518 | num_resolved_ok_drs++; |
1519 | else if (multiple_of_p (TREE_TYPE (iloop_stride), |
1520 | oloop_stride, iloop_stride)) |
1521 | num_resolved_not_ok_drs++; |
1522 | else |
1523 | num_unresolved_drs++; |
1524 | |
1525 | /* Data ref can't be sequential access if its address changes in sub |
1526 | loop. */ |
1527 | if (subloop_stride_p) |
1528 | { |
1529 | all_seq_dr_before_p = false; |
1530 | all_seq_dr_after_p = false; |
1531 | continue; |
1532 | } |
1533 | /* Track if all data references are sequential accesses before/after loop |
1534 | interchange. Note invariant is considered sequential here. */ |
1535 | tree access_size = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr))); |
1536 | if (all_seq_dr_before_p |
1537 | && ! (integer_zerop (iloop_stride) |
1538 | || operand_equal_p (access_size, iloop_stride, flags: 0))) |
1539 | all_seq_dr_before_p = false; |
1540 | if (all_seq_dr_after_p |
1541 | && ! (integer_zerop (oloop_stride) |
1542 | || operand_equal_p (access_size, oloop_stride, flags: 0))) |
1543 | all_seq_dr_after_p = false; |
1544 | } |
1545 | |
1546 | if (dump_info_p && dump_file && (dump_flags & TDF_DETAILS)) |
1547 | { |
1548 | fprintf (stream: dump_file, format: "\toverall:\t\t" ); |
1549 | print_decu (wi: iloop_strides, file: dump_file); |
1550 | fprintf (stream: dump_file, format: "\t" ); |
1551 | print_decu (wi: oloop_strides, file: dump_file); |
1552 | fprintf (stream: dump_file, format: "\n" ); |
1553 | |
1554 | fprintf (stream: dump_file, format: "Invariant data ref: before(%d), after(%d)\n" , |
1555 | num_old_inv_drs, num_new_inv_drs); |
1556 | fprintf (stream: dump_file, format: "All consecutive stride: before(%s), after(%s)\n" , |
1557 | all_seq_dr_before_p ? "true" : "false" , |
1558 | all_seq_dr_after_p ? "true" : "false" ); |
1559 | fprintf (stream: dump_file, format: "OK to interchage with variable strides: %d\n" , |
1560 | num_resolved_ok_drs); |
1561 | fprintf (stream: dump_file, format: "Not OK to interchage with variable strides: %d\n" , |
1562 | num_resolved_not_ok_drs); |
1563 | fprintf (stream: dump_file, format: "Variable strides we cannot decide: %d\n" , |
1564 | num_unresolved_drs); |
1565 | fprintf (stream: dump_file, format: "Stmt cost of inner loop: %d\n" , i_stmt_cost); |
1566 | fprintf (stream: dump_file, format: "Stmt cost of outer loop: %d\n" , o_stmt_cost); |
1567 | } |
1568 | |
1569 | if (num_unresolved_drs != 0 || num_resolved_not_ok_drs != 0) |
1570 | return false; |
1571 | |
1572 | /* Stmts of outer loop will be moved to inner loop. If there are two many |
1573 | such stmts, it could make inner loop costly. Here we compare stmt cost |
1574 | between outer and inner loops. */ |
1575 | if (i_stmt_cost && o_stmt_cost |
1576 | && num_old_inv_drs + o_stmt_cost > num_new_inv_drs |
1577 | && o_stmt_cost * STMT_COST_RATIO > i_stmt_cost) |
1578 | return false; |
1579 | |
1580 | /* We use different stride comparison ratio for interchanging innermost |
1581 | two loops or not. The idea is to be conservative in interchange for |
1582 | the innermost loops. */ |
1583 | ratio = innermost_loops_p ? INNER_STRIDE_RATIO : OUTER_STRIDE_RATIO; |
1584 | /* Do interchange if it gives better data locality behavior. */ |
1585 | if (wi::gtu_p (x: iloop_strides, y: wi::mul (x: oloop_strides, y: ratio))) |
1586 | return true; |
1587 | if (wi::gtu_p (x: iloop_strides, y: oloop_strides)) |
1588 | { |
1589 | /* Or it creates more invariant memory references. */ |
1590 | if ((!all_seq_dr_before_p || all_seq_dr_after_p) |
1591 | && num_new_inv_drs > num_old_inv_drs) |
1592 | return true; |
1593 | /* Or it makes all memory references sequential. */ |
1594 | if (num_new_inv_drs >= num_old_inv_drs |
1595 | && !all_seq_dr_before_p && all_seq_dr_after_p) |
1596 | return true; |
1597 | } |
1598 | |
1599 | return false; |
1600 | } |
1601 | |
1602 | /* Try to interchange inner loop of a loop nest to outer level. */ |
1603 | |
1604 | bool |
1605 | tree_loop_interchange::interchange (vec<data_reference_p> datarefs, |
1606 | vec<ddr_p> ddrs) |
1607 | { |
1608 | dump_user_location_t loc = find_loop_location (m_loop_nest[0]); |
1609 | bool changed_p = false; |
1610 | /* In each iteration we try to interchange I-th loop with (I+1)-th loop. |
1611 | The overall effect is to push inner loop to outermost level in whole |
1612 | loop nest. */ |
1613 | for (unsigned i = m_loop_nest.length (); i > 1; --i) |
1614 | { |
1615 | unsigned i_idx = i - 1, o_idx = i - 2; |
1616 | |
1617 | /* Check validity for loop interchange. */ |
1618 | if (!valid_data_dependences (i_idx, o_idx, ddrs)) |
1619 | break; |
1620 | |
1621 | loop_cand iloop (m_loop_nest[i_idx], m_loop_nest[o_idx]); |
1622 | loop_cand oloop (m_loop_nest[o_idx], m_loop_nest[o_idx]); |
1623 | |
1624 | /* Check if we can do transformation for loop interchange. */ |
1625 | if (!iloop.analyze_carried_vars (NULL) |
1626 | || !iloop.analyze_lcssa_phis () |
1627 | || !oloop.analyze_carried_vars (iloop: &iloop) |
1628 | || !oloop.analyze_lcssa_phis () |
1629 | || !iloop.can_interchange_p (NULL) |
1630 | || !oloop.can_interchange_p (iloop: &iloop)) |
1631 | break; |
1632 | |
1633 | /* Outer loop's stmts will be moved to inner loop during interchange. |
1634 | If there are many of them, it may make inner loop very costly. We |
1635 | need to check number of outer loop's stmts in profit cost model of |
1636 | interchange. */ |
1637 | int stmt_cost = oloop.m_num_stmts; |
1638 | /* Count out the exit checking stmt of outer loop. */ |
1639 | stmt_cost --; |
1640 | /* Count out IV's increasing stmt, IVOPTs takes care if it. */ |
1641 | stmt_cost -= oloop.m_inductions.length (); |
1642 | /* Count in the additional load and cond_expr stmts caused by inner |
1643 | loop's constant initialized reduction. */ |
1644 | stmt_cost += iloop.m_const_init_reduc * 2; |
1645 | if (stmt_cost < 0) |
1646 | stmt_cost = 0; |
1647 | |
1648 | /* Check profitability for loop interchange. */ |
1649 | if (should_interchange_loops (i_idx, o_idx, datarefs, |
1650 | i_stmt_cost: (unsigned) iloop.m_num_stmts, |
1651 | o_stmt_cost: (unsigned) stmt_cost, |
1652 | innermost_loops_p: iloop.m_loop->inner == NULL)) |
1653 | { |
1654 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1655 | fprintf (stream: dump_file, |
1656 | format: "Loop_pair<outer:%d, inner:%d> is interchanged\n\n" , |
1657 | oloop.m_loop->num, iloop.m_loop->num); |
1658 | |
1659 | changed_p = true; |
1660 | interchange_loops (iloop, oloop); |
1661 | /* No need to update if there is no further loop interchange. */ |
1662 | if (o_idx > 0) |
1663 | update_data_info (i_idx, o_idx, datarefs, ddrs); |
1664 | } |
1665 | else |
1666 | { |
1667 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1668 | fprintf (stream: dump_file, |
1669 | format: "Loop_pair<outer:%d, inner:%d> is not interchanged\n\n" , |
1670 | oloop.m_loop->num, iloop.m_loop->num); |
1671 | } |
1672 | } |
1673 | simple_dce_from_worklist (m_dce_seeds); |
1674 | |
1675 | if (changed_p && dump_enabled_p ()) |
1676 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loc, |
1677 | "loops interchanged in loop nest\n" ); |
1678 | |
1679 | return changed_p; |
1680 | } |
1681 | |
1682 | |
1683 | /* Loop interchange pass. */ |
1684 | |
1685 | namespace { |
1686 | |
1687 | const pass_data pass_data_linterchange = |
1688 | { |
1689 | .type: GIMPLE_PASS, /* type */ |
1690 | .name: "linterchange" , /* name */ |
1691 | .optinfo_flags: OPTGROUP_LOOP, /* optinfo_flags */ |
1692 | .tv_id: TV_LINTERCHANGE, /* tv_id */ |
1693 | PROP_cfg, /* properties_required */ |
1694 | .properties_provided: 0, /* properties_provided */ |
1695 | .properties_destroyed: 0, /* properties_destroyed */ |
1696 | .todo_flags_start: 0, /* todo_flags_start */ |
1697 | .todo_flags_finish: 0, /* todo_flags_finish */ |
1698 | }; |
1699 | |
1700 | class pass_linterchange : public gimple_opt_pass |
1701 | { |
1702 | public: |
1703 | pass_linterchange (gcc::context *ctxt) |
1704 | : gimple_opt_pass (pass_data_linterchange, ctxt) |
1705 | {} |
1706 | |
1707 | /* opt_pass methods: */ |
1708 | opt_pass * clone () final override { return new pass_linterchange (m_ctxt); } |
1709 | bool gate (function *) final override { return flag_loop_interchange; } |
1710 | unsigned int execute (function *) final override; |
1711 | |
1712 | }; // class pass_linterchange |
1713 | |
1714 | |
1715 | /* Return true if LOOP has proper form for interchange. We check three |
1716 | conditions in the function: |
1717 | 1) In general, a loop can be interchanged only if it doesn't have |
1718 | basic blocks other than header, exit and latch besides possible |
1719 | inner loop nest. This basically restricts loop interchange to |
1720 | below form loop nests: |
1721 | |
1722 | header<---+ |
1723 | | | |
1724 | v | |
1725 | INNER_LOOP | |
1726 | | | |
1727 | v | |
1728 | exit--->latch |
1729 | |
1730 | 2) Data reference in basic block that executes in different times |
1731 | than loop head/exit is not allowed. |
1732 | 3) Record the innermost outer loop that doesn't form rectangle loop |
1733 | nest with LOOP. */ |
1734 | |
1735 | static bool |
1736 | proper_loop_form_for_interchange (class loop *loop, class loop **min_outer) |
1737 | { |
1738 | edge e0, e1, exit; |
1739 | |
1740 | /* Don't interchange if loop has unsupported information for the moment. */ |
1741 | if (loop->safelen > 0 |
1742 | || loop->constraints != 0 |
1743 | || loop->can_be_parallel |
1744 | || loop->dont_vectorize |
1745 | || loop->force_vectorize |
1746 | || loop->in_oacc_kernels_region |
1747 | || loop->orig_loop_num != 0 |
1748 | || loop->simduid != NULL_TREE) |
1749 | return false; |
1750 | |
1751 | /* Don't interchange if outer loop has basic block other than header, exit |
1752 | and latch. */ |
1753 | if (loop->inner != NULL |
1754 | && loop->num_nodes != loop->inner->num_nodes + 3) |
1755 | return false; |
1756 | |
1757 | if ((exit = single_dom_exit (loop)) == NULL) |
1758 | return false; |
1759 | |
1760 | /* Check control flow on loop header/exit blocks. */ |
1761 | if (loop->header == exit->src |
1762 | && (EDGE_COUNT (loop->header->preds) != 2 |
1763 | || EDGE_COUNT (loop->header->succs) != 2)) |
1764 | return false; |
1765 | else if (loop->header != exit->src |
1766 | && (EDGE_COUNT (loop->header->preds) != 2 |
1767 | || !single_succ_p (bb: loop->header) |
1768 | || unsupported_edge (e: single_succ_edge (bb: loop->header)) |
1769 | || EDGE_COUNT (exit->src->succs) != 2 |
1770 | || !single_pred_p (bb: exit->src) |
1771 | || unsupported_edge (e: single_pred_edge (bb: exit->src)))) |
1772 | return false; |
1773 | |
1774 | e0 = EDGE_PRED (loop->header, 0); |
1775 | e1 = EDGE_PRED (loop->header, 1); |
1776 | if (unsupported_edge (e: e0) || unsupported_edge (e: e1) |
1777 | || (e0->src != loop->latch && e1->src != loop->latch) |
1778 | || (e0->src->loop_father == loop && e1->src->loop_father == loop)) |
1779 | return false; |
1780 | |
1781 | e0 = EDGE_SUCC (exit->src, 0); |
1782 | e1 = EDGE_SUCC (exit->src, 1); |
1783 | if (unsupported_edge (e: e0) || unsupported_edge (e: e1) |
1784 | || (e0->dest != loop->latch && e1->dest != loop->latch) |
1785 | || (e0->dest->loop_father == loop && e1->dest->loop_father == loop)) |
1786 | return false; |
1787 | |
1788 | /* Don't interchange if any reference is in basic block that doesn't |
1789 | dominate exit block. */ |
1790 | basic_block *bbs = get_loop_body (loop); |
1791 | for (unsigned i = 0; i < loop->num_nodes; i++) |
1792 | { |
1793 | basic_block bb = bbs[i]; |
1794 | |
1795 | if (bb->loop_father != loop |
1796 | || bb == loop->header || bb == exit->src |
1797 | || dominated_by_p (CDI_DOMINATORS, exit->src, bb)) |
1798 | continue; |
1799 | |
1800 | for (gimple_stmt_iterator gsi = gsi_start_nondebug_bb (bb); |
1801 | !gsi_end_p (i: gsi); gsi_next_nondebug (i: &gsi)) |
1802 | if (gimple_vuse (g: gsi_stmt (i: gsi))) |
1803 | { |
1804 | free (ptr: bbs); |
1805 | return false; |
1806 | } |
1807 | } |
1808 | free (ptr: bbs); |
1809 | |
1810 | tree niters = number_of_latch_executions (loop); |
1811 | niters = analyze_scalar_evolution (loop_outer (loop), niters); |
1812 | if (!niters || chrec_contains_undetermined (niters)) |
1813 | return false; |
1814 | |
1815 | /* Record the innermost outer loop that doesn't form rectangle loop nest. */ |
1816 | for (loop_p loop2 = loop_outer (loop); |
1817 | loop2 && flow_loop_nested_p (*min_outer, loop2); |
1818 | loop2 = loop_outer (loop: loop2)) |
1819 | { |
1820 | niters = instantiate_scev (loop_preheader_edge (loop2), |
1821 | loop_outer (loop), niters); |
1822 | if (!evolution_function_is_invariant_p (niters, loop2->num)) |
1823 | { |
1824 | *min_outer = loop2; |
1825 | break; |
1826 | } |
1827 | } |
1828 | return true; |
1829 | } |
1830 | |
1831 | /* Return true if any two adjacent loops in loop nest [INNERMOST, LOOP_NEST] |
1832 | should be interchanged by looking into all DATAREFS. */ |
1833 | |
1834 | static bool |
1835 | should_interchange_loop_nest (class loop *loop_nest, class loop *innermost, |
1836 | vec<data_reference_p> datarefs) |
1837 | { |
1838 | unsigned idx = loop_depth (loop: innermost) - loop_depth (loop: loop_nest); |
1839 | gcc_assert (idx > 0); |
1840 | |
1841 | /* Check if any two adjacent loops should be interchanged. */ |
1842 | for (class loop *loop = innermost; |
1843 | loop != loop_nest; loop = loop_outer (loop), idx--) |
1844 | if (should_interchange_loops (i_idx: idx, o_idx: idx - 1, datarefs, i_stmt_cost: 0, o_stmt_cost: 0, |
1845 | innermost_loops_p: loop == innermost, dump_info_p: false)) |
1846 | return true; |
1847 | |
1848 | return false; |
1849 | } |
1850 | |
1851 | /* Given loop nest LOOP_NEST and data references DATAREFS, compute data |
1852 | dependences for loop interchange and store it in DDRS. Note we compute |
1853 | dependences directly rather than call generic interface so that we can |
1854 | return on unknown dependence instantly. */ |
1855 | |
1856 | static bool |
1857 | tree_loop_interchange_compute_ddrs (vec<loop_p> loop_nest, |
1858 | vec<data_reference_p> datarefs, |
1859 | vec<ddr_p> *ddrs) |
1860 | { |
1861 | struct data_reference *a, *b; |
1862 | class loop *innermost = loop_nest.last (); |
1863 | |
1864 | for (unsigned i = 0; datarefs.iterate (ix: i, ptr: &a); ++i) |
1865 | { |
1866 | bool a_outer_p = gimple_bb (DR_STMT (a))->loop_father != innermost; |
1867 | for (unsigned j = i + 1; datarefs.iterate (ix: j, ptr: &b); ++j) |
1868 | if (DR_IS_WRITE (a) || DR_IS_WRITE (b)) |
1869 | { |
1870 | bool b_outer_p = gimple_bb (DR_STMT (b))->loop_father != innermost; |
1871 | /* Don't support multiple write references in outer loop. */ |
1872 | if (a_outer_p && b_outer_p && DR_IS_WRITE (a) && DR_IS_WRITE (b)) |
1873 | return false; |
1874 | |
1875 | ddr_p ddr = initialize_data_dependence_relation (a, b, loop_nest); |
1876 | ddrs->safe_push (obj: ddr); |
1877 | compute_affine_dependence (ddr, loop_nest[0]); |
1878 | |
1879 | /* Give up if ddr is unknown dependence or classic direct vector |
1880 | is not available. */ |
1881 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know |
1882 | || (DDR_ARE_DEPENDENT (ddr) == NULL_TREE |
1883 | && DDR_NUM_DIR_VECTS (ddr) == 0)) |
1884 | return false; |
1885 | |
1886 | /* If either data references is in outer loop of nest, we require |
1887 | no dependence here because the data reference need to be moved |
1888 | into inner loop during interchange. */ |
1889 | if (a_outer_p && b_outer_p |
1890 | && operand_equal_p (DR_REF (a), DR_REF (b), flags: 0)) |
1891 | continue; |
1892 | if (DDR_ARE_DEPENDENT (ddr) != chrec_known |
1893 | && (a_outer_p || b_outer_p)) |
1894 | return false; |
1895 | } |
1896 | } |
1897 | |
1898 | return true; |
1899 | } |
1900 | |
1901 | /* Prune DATAREFS by removing any data reference not inside of LOOP. */ |
1902 | |
1903 | static inline void |
1904 | prune_datarefs_not_in_loop (class loop *loop, vec<data_reference_p> datarefs) |
1905 | { |
1906 | unsigned i, j; |
1907 | struct data_reference *dr; |
1908 | |
1909 | for (i = 0, j = 0; datarefs.iterate (ix: i, ptr: &dr); ++i) |
1910 | { |
1911 | if (flow_bb_inside_loop_p (loop, gimple_bb (DR_STMT (dr)))) |
1912 | datarefs[j++] = dr; |
1913 | else |
1914 | { |
1915 | if (dr->aux) |
1916 | { |
1917 | DR_ACCESS_STRIDE (dr)->release (); |
1918 | delete (vec<tree> *) dr->aux; |
1919 | } |
1920 | free_data_ref (dr); |
1921 | } |
1922 | } |
1923 | datarefs.truncate (size: j); |
1924 | } |
1925 | |
1926 | /* Find and store data references in DATAREFS for LOOP nest. If there's |
1927 | difficult data reference in a basic block, we shrink the loop nest to |
1928 | inner loop of that basic block's father loop. On success, return the |
1929 | outer loop of the result loop nest. */ |
1930 | |
1931 | static class loop * |
1932 | prepare_data_references (class loop *loop, vec<data_reference_p> *datarefs) |
1933 | { |
1934 | class loop *loop_nest = loop; |
1935 | vec<data_reference_p> *bb_refs; |
1936 | basic_block bb, *bbs = get_loop_body_in_dom_order (loop); |
1937 | |
1938 | for (unsigned i = 0; i < loop->num_nodes; i++) |
1939 | bbs[i]->aux = NULL; |
1940 | |
1941 | /* Find data references for all basic blocks. Shrink loop nest on difficult |
1942 | data reference. */ |
1943 | for (unsigned i = 0; loop_nest && i < loop->num_nodes; ++i) |
1944 | { |
1945 | bb = bbs[i]; |
1946 | if (!flow_bb_inside_loop_p (loop_nest, bb)) |
1947 | continue; |
1948 | |
1949 | bb_refs = new vec<data_reference_p> (); |
1950 | if (find_data_references_in_bb (loop, bb, bb_refs) == chrec_dont_know) |
1951 | { |
1952 | loop_nest = bb->loop_father->inner; |
1953 | if (loop_nest && !loop_nest->inner) |
1954 | loop_nest = NULL; |
1955 | |
1956 | free_data_refs (*bb_refs); |
1957 | delete bb_refs; |
1958 | } |
1959 | else if (bb_refs->is_empty ()) |
1960 | { |
1961 | bb_refs->release (); |
1962 | delete bb_refs; |
1963 | } |
1964 | else |
1965 | bb->aux = bb_refs; |
1966 | } |
1967 | |
1968 | /* Collect all data references in loop nest. */ |
1969 | for (unsigned i = 0; i < loop->num_nodes; i++) |
1970 | { |
1971 | bb = bbs[i]; |
1972 | if (!bb->aux) |
1973 | continue; |
1974 | |
1975 | bb_refs = (vec<data_reference_p> *) bb->aux; |
1976 | if (loop_nest && flow_bb_inside_loop_p (loop_nest, bb)) |
1977 | { |
1978 | datarefs->safe_splice (src: *bb_refs); |
1979 | bb_refs->release (); |
1980 | } |
1981 | else |
1982 | free_data_refs (*bb_refs); |
1983 | |
1984 | delete bb_refs; |
1985 | bb->aux = NULL; |
1986 | } |
1987 | free (ptr: bbs); |
1988 | |
1989 | return loop_nest; |
1990 | } |
1991 | |
1992 | /* Given innermost LOOP, return true if perfect loop nest can be found and |
1993 | data dependences can be computed. If succeed, record the perfect loop |
1994 | nest in LOOP_NEST; record all data references in DATAREFS and record all |
1995 | data dependence relations in DDRS. |
1996 | |
1997 | We do support a restricted form of imperfect loop nest, i.e, loop nest |
1998 | with load/store in outer loop initializing/finalizing simple reduction |
1999 | of the innermost loop. For such outer loop reference, we require that |
2000 | it has no dependence with others sinve it will be moved to inner loop |
2001 | in interchange. */ |
2002 | |
2003 | static bool |
2004 | prepare_perfect_loop_nest (class loop *loop, vec<loop_p> *loop_nest, |
2005 | vec<data_reference_p> *datarefs, vec<ddr_p> *ddrs) |
2006 | { |
2007 | class loop *start_loop = NULL, *innermost = loop; |
2008 | class loop *outermost = loops_for_fn (cfun)->tree_root; |
2009 | |
2010 | /* Find loop nest from the innermost loop. The outermost is the innermost |
2011 | outer*/ |
2012 | while (loop->num != 0 && loop->inner == start_loop |
2013 | && flow_loop_nested_p (outermost, loop)) |
2014 | { |
2015 | if (!proper_loop_form_for_interchange (loop, min_outer: &outermost)) |
2016 | break; |
2017 | |
2018 | start_loop = loop; |
2019 | /* If this loop has sibling loop, the father loop won't be in perfect |
2020 | loop nest. */ |
2021 | if (loop->next != NULL) |
2022 | break; |
2023 | |
2024 | loop = loop_outer (loop); |
2025 | } |
2026 | if (!start_loop || !start_loop->inner) |
2027 | return false; |
2028 | |
2029 | /* Prepare the data reference vector for the loop nest, pruning outer |
2030 | loops we cannot handle. */ |
2031 | start_loop = prepare_data_references (loop: start_loop, datarefs); |
2032 | if (!start_loop |
2033 | /* Check if there is no data reference. */ |
2034 | || datarefs->is_empty () |
2035 | /* Check if there are too many of data references. */ |
2036 | || (int) datarefs->length () > MAX_DATAREFS) |
2037 | return false; |
2038 | |
2039 | /* Compute access strides for all data references, pruning outer |
2040 | loops we cannot analyze refs in. */ |
2041 | start_loop = compute_access_strides (loop_nest: start_loop, loop: innermost, datarefs: *datarefs); |
2042 | if (!start_loop) |
2043 | return false; |
2044 | |
2045 | /* Check if any interchange is profitable in the loop nest. */ |
2046 | if (!should_interchange_loop_nest (loop_nest: start_loop, innermost, datarefs: *datarefs)) |
2047 | return false; |
2048 | |
2049 | /* Check if data dependences can be computed for loop nest starting from |
2050 | start_loop. */ |
2051 | loop = start_loop; |
2052 | do { |
2053 | loop_nest->truncate (size: 0); |
2054 | |
2055 | if (loop != start_loop) |
2056 | prune_datarefs_not_in_loop (loop: start_loop, datarefs: *datarefs); |
2057 | |
2058 | if (find_loop_nest (start_loop, loop_nest) |
2059 | && tree_loop_interchange_compute_ddrs (loop_nest: *loop_nest, datarefs: *datarefs, ddrs)) |
2060 | { |
2061 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2062 | fprintf (stream: dump_file, |
2063 | format: "\nConsider loop interchange for loop_nest<%d - %d>\n" , |
2064 | start_loop->num, innermost->num); |
2065 | |
2066 | if (loop != start_loop) |
2067 | prune_access_strides_not_in_loop (loop_nest: start_loop, innermost, datarefs: *datarefs); |
2068 | |
2069 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2070 | dump_access_strides (datarefs: *datarefs); |
2071 | |
2072 | return true; |
2073 | } |
2074 | |
2075 | free_dependence_relations (*ddrs); |
2076 | *ddrs = vNULL; |
2077 | /* Try to compute data dependences with the outermost loop stripped. */ |
2078 | loop = start_loop; |
2079 | start_loop = start_loop->inner; |
2080 | } while (start_loop && start_loop->inner); |
2081 | |
2082 | return false; |
2083 | } |
2084 | |
2085 | /* Main entry for loop interchange pass. */ |
2086 | |
2087 | unsigned int |
2088 | pass_linterchange::execute (function *fun) |
2089 | { |
2090 | if (number_of_loops (fn: fun) <= 2) |
2091 | return 0; |
2092 | |
2093 | bool changed_p = false; |
2094 | for (auto loop : loops_list (cfun, LI_ONLY_INNERMOST)) |
2095 | { |
2096 | vec<loop_p> loop_nest = vNULL; |
2097 | vec<data_reference_p> datarefs = vNULL; |
2098 | vec<ddr_p> ddrs = vNULL; |
2099 | if (prepare_perfect_loop_nest (loop, loop_nest: &loop_nest, datarefs: &datarefs, ddrs: &ddrs)) |
2100 | { |
2101 | tree_loop_interchange loop_interchange (loop_nest); |
2102 | changed_p |= loop_interchange.interchange (datarefs, ddrs); |
2103 | } |
2104 | free_dependence_relations (ddrs); |
2105 | free_data_refs_with_aux (datarefs); |
2106 | loop_nest.release (); |
2107 | } |
2108 | |
2109 | if (changed_p) |
2110 | { |
2111 | unsigned todo = TODO_update_ssa_only_virtuals; |
2112 | todo |= loop_invariant_motion_in_fun (cfun, false); |
2113 | scev_reset (); |
2114 | return todo; |
2115 | } |
2116 | return 0; |
2117 | } |
2118 | |
2119 | } // anon namespace |
2120 | |
2121 | gimple_opt_pass * |
2122 | make_pass_linterchange (gcc::context *ctxt) |
2123 | { |
2124 | return new pass_linterchange (ctxt); |
2125 | } |
2126 | |