1 | /* Lower GIMPLE_SWITCH expressions to something more efficient than |
2 | a jump table. |
3 | Copyright (C) 2006-2024 Free Software Foundation, Inc. |
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, write to the Free |
19 | Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA |
20 | 02110-1301, USA. */ |
21 | |
22 | /* This file handles the lowering of GIMPLE_SWITCH to an indexed |
23 | load, or a series of bit-test-and-branch expressions. */ |
24 | |
25 | #include "config.h" |
26 | #include "system.h" |
27 | #include "coretypes.h" |
28 | #include "backend.h" |
29 | #include "insn-codes.h" |
30 | #include "rtl.h" |
31 | #include "tree.h" |
32 | #include "gimple.h" |
33 | #include "cfghooks.h" |
34 | #include "tree-pass.h" |
35 | #include "ssa.h" |
36 | #include "optabs-tree.h" |
37 | #include "cgraph.h" |
38 | #include "gimple-pretty-print.h" |
39 | #include "fold-const.h" |
40 | #include "varasm.h" |
41 | #include "stor-layout.h" |
42 | #include "cfganal.h" |
43 | #include "gimplify.h" |
44 | #include "gimple-iterator.h" |
45 | #include "gimplify-me.h" |
46 | #include "gimple-fold.h" |
47 | #include "tree-cfg.h" |
48 | #include "cfgloop.h" |
49 | #include "alloc-pool.h" |
50 | #include "target.h" |
51 | #include "tree-into-ssa.h" |
52 | #include "omp-general.h" |
53 | #include "gimple-range.h" |
54 | #include "tree-cfgcleanup.h" |
55 | |
56 | /* ??? For lang_hooks.types.type_for_mode, but is there a word_mode |
57 | type in the GIMPLE type system that is language-independent? */ |
58 | #include "langhooks.h" |
59 | |
60 | #include "tree-switch-conversion.h" |
61 | |
62 | using namespace tree_switch_conversion; |
63 | |
64 | /* Constructor. */ |
65 | |
66 | switch_conversion::switch_conversion (): m_final_bb (NULL), |
67 | m_constructors (NULL), m_default_values (NULL), |
68 | m_arr_ref_first (NULL), m_arr_ref_last (NULL), |
69 | m_reason (NULL), m_default_case_nonstandard (false), m_cfg_altered (false) |
70 | { |
71 | } |
72 | |
73 | /* Collection information about SWTCH statement. */ |
74 | |
75 | void |
76 | switch_conversion::collect (gswitch *swtch) |
77 | { |
78 | unsigned int branch_num = gimple_switch_num_labels (gs: swtch); |
79 | tree min_case, max_case; |
80 | unsigned int i; |
81 | edge e, e_default, e_first; |
82 | edge_iterator ei; |
83 | |
84 | m_switch = swtch; |
85 | |
86 | /* The gimplifier has already sorted the cases by CASE_LOW and ensured there |
87 | is a default label which is the first in the vector. |
88 | Collect the bits we can deduce from the CFG. */ |
89 | m_index_expr = gimple_switch_index (gs: swtch); |
90 | m_switch_bb = gimple_bb (g: swtch); |
91 | e_default = gimple_switch_default_edge (cfun, swtch); |
92 | m_default_bb = e_default->dest; |
93 | m_default_prob = e_default->probability; |
94 | |
95 | /* Get upper and lower bounds of case values, and the covered range. */ |
96 | min_case = gimple_switch_label (gs: swtch, index: 1); |
97 | max_case = gimple_switch_label (gs: swtch, index: branch_num - 1); |
98 | |
99 | m_range_min = CASE_LOW (min_case); |
100 | if (CASE_HIGH (max_case) != NULL_TREE) |
101 | m_range_max = CASE_HIGH (max_case); |
102 | else |
103 | m_range_max = CASE_LOW (max_case); |
104 | |
105 | m_contiguous_range = true; |
106 | tree last = CASE_HIGH (min_case) ? CASE_HIGH (min_case) : m_range_min; |
107 | for (i = 2; i < branch_num; i++) |
108 | { |
109 | tree elt = gimple_switch_label (gs: swtch, index: i); |
110 | if (wi::to_wide (t: last) + 1 != wi::to_wide (CASE_LOW (elt))) |
111 | { |
112 | m_contiguous_range = false; |
113 | break; |
114 | } |
115 | last = CASE_HIGH (elt) ? CASE_HIGH (elt) : CASE_LOW (elt); |
116 | } |
117 | |
118 | if (m_contiguous_range) |
119 | e_first = gimple_switch_edge (cfun, swtch, 1); |
120 | else |
121 | e_first = e_default; |
122 | |
123 | /* See if there is one common successor block for all branch |
124 | targets. If it exists, record it in FINAL_BB. |
125 | Start with the destination of the first non-default case |
126 | if the range is contiguous and default case otherwise as |
127 | guess or its destination in case it is a forwarder block. */ |
128 | if (! single_pred_p (bb: e_first->dest)) |
129 | m_final_bb = e_first->dest; |
130 | else if (single_succ_p (bb: e_first->dest) |
131 | && ! single_pred_p (bb: single_succ (bb: e_first->dest))) |
132 | m_final_bb = single_succ (bb: e_first->dest); |
133 | /* Require that all switch destinations are either that common |
134 | FINAL_BB or a forwarder to it, except for the default |
135 | case if contiguous range. */ |
136 | auto_vec<edge, 10> fw_edges; |
137 | m_uniq = 0; |
138 | if (m_final_bb) |
139 | FOR_EACH_EDGE (e, ei, m_switch_bb->succs) |
140 | { |
141 | edge phi_e = nullptr; |
142 | if (e->dest == m_final_bb) |
143 | phi_e = e; |
144 | else if (single_pred_p (bb: e->dest) |
145 | && single_succ_p (bb: e->dest) |
146 | && single_succ (bb: e->dest) == m_final_bb) |
147 | phi_e = single_succ_edge (bb: e->dest); |
148 | if (phi_e) |
149 | { |
150 | if (e == e_default) |
151 | ; |
152 | else if (phi_e == e || empty_block_p (e->dest)) |
153 | { |
154 | /* For empty blocks consider forwarders with equal |
155 | PHI arguments in m_final_bb as unique. */ |
156 | unsigned i; |
157 | for (i = 0; i < fw_edges.length (); ++i) |
158 | if (phi_alternatives_equal (m_final_bb, fw_edges[i], phi_e)) |
159 | break; |
160 | if (i == fw_edges.length ()) |
161 | { |
162 | /* But limit the above possibly quadratic search. */ |
163 | if (fw_edges.length () < 10) |
164 | fw_edges.quick_push (obj: phi_e); |
165 | m_uniq++; |
166 | } |
167 | } |
168 | else |
169 | m_uniq++; |
170 | continue; |
171 | } |
172 | |
173 | if (e == e_default && m_contiguous_range) |
174 | { |
175 | m_default_case_nonstandard = true; |
176 | continue; |
177 | } |
178 | |
179 | m_final_bb = NULL; |
180 | break; |
181 | } |
182 | |
183 | /* When there's not a single common successor block conservatively |
184 | approximate the number of unique non-default targets. */ |
185 | if (!m_final_bb) |
186 | m_uniq = EDGE_COUNT (gimple_bb (swtch)->succs) - 1; |
187 | |
188 | m_range_size |
189 | = int_const_binop (MINUS_EXPR, m_range_max, m_range_min); |
190 | |
191 | /* Get a count of the number of case labels. Single-valued case labels |
192 | simply count as one, but a case range counts double, since it may |
193 | require two compares if it gets lowered as a branching tree. */ |
194 | m_count = 0; |
195 | for (i = 1; i < branch_num; i++) |
196 | { |
197 | tree elt = gimple_switch_label (gs: swtch, index: i); |
198 | m_count++; |
199 | if (CASE_HIGH (elt) |
200 | && ! tree_int_cst_equal (CASE_LOW (elt), CASE_HIGH (elt))) |
201 | m_count++; |
202 | } |
203 | } |
204 | |
205 | /* Checks whether the range given by individual case statements of the switch |
206 | switch statement isn't too big and whether the number of branches actually |
207 | satisfies the size of the new array. */ |
208 | |
209 | bool |
210 | switch_conversion::check_range () |
211 | { |
212 | gcc_assert (m_range_size); |
213 | if (!tree_fits_uhwi_p (m_range_size)) |
214 | { |
215 | m_reason = "index range way too large or otherwise unusable" ; |
216 | return false; |
217 | } |
218 | |
219 | if (tree_to_uhwi (m_range_size) |
220 | > ((unsigned) m_count * param_switch_conversion_branch_ratio)) |
221 | { |
222 | m_reason = "the maximum range-branch ratio exceeded" ; |
223 | return false; |
224 | } |
225 | |
226 | return true; |
227 | } |
228 | |
229 | /* Checks whether all but the final BB basic blocks are empty. */ |
230 | |
231 | bool |
232 | switch_conversion::check_all_empty_except_final () |
233 | { |
234 | edge e, e_default = find_edge (m_switch_bb, m_default_bb); |
235 | edge_iterator ei; |
236 | |
237 | FOR_EACH_EDGE (e, ei, m_switch_bb->succs) |
238 | { |
239 | if (e->dest == m_final_bb) |
240 | continue; |
241 | |
242 | if (!empty_block_p (e->dest)) |
243 | { |
244 | if (m_contiguous_range && e == e_default) |
245 | { |
246 | m_default_case_nonstandard = true; |
247 | continue; |
248 | } |
249 | |
250 | m_reason = "bad case - a non-final BB not empty" ; |
251 | return false; |
252 | } |
253 | } |
254 | |
255 | return true; |
256 | } |
257 | |
258 | /* This function checks whether all required values in phi nodes in final_bb |
259 | are constants. Required values are those that correspond to a basic block |
260 | which is a part of the examined switch statement. It returns true if the |
261 | phi nodes are OK, otherwise false. */ |
262 | |
263 | bool |
264 | switch_conversion::check_final_bb () |
265 | { |
266 | gphi_iterator gsi; |
267 | |
268 | m_phi_count = 0; |
269 | for (gsi = gsi_start_phis (m_final_bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
270 | { |
271 | gphi *phi = gsi.phi (); |
272 | unsigned int i; |
273 | |
274 | if (virtual_operand_p (op: gimple_phi_result (gs: phi))) |
275 | continue; |
276 | |
277 | m_phi_count++; |
278 | |
279 | for (i = 0; i < gimple_phi_num_args (gs: phi); i++) |
280 | { |
281 | basic_block bb = gimple_phi_arg_edge (phi, i)->src; |
282 | |
283 | if (bb == m_switch_bb |
284 | || (single_pred_p (bb) |
285 | && single_pred (bb) == m_switch_bb |
286 | && (!m_default_case_nonstandard |
287 | || empty_block_p (bb)))) |
288 | { |
289 | tree reloc, val; |
290 | const char *reason = NULL; |
291 | |
292 | val = gimple_phi_arg_def (gs: phi, index: i); |
293 | if (!is_gimple_ip_invariant (val)) |
294 | reason = "non-invariant value from a case" ; |
295 | else |
296 | { |
297 | reloc = initializer_constant_valid_p (val, TREE_TYPE (val)); |
298 | if ((flag_pic && reloc != null_pointer_node) |
299 | || (!flag_pic && reloc == NULL_TREE)) |
300 | { |
301 | if (reloc) |
302 | reason |
303 | = "value from a case would need runtime relocations" ; |
304 | else |
305 | reason |
306 | = "value from a case is not a valid initializer" ; |
307 | } |
308 | } |
309 | if (reason) |
310 | { |
311 | /* For contiguous range, we can allow non-constant |
312 | or one that needs relocation, as long as it is |
313 | only reachable from the default case. */ |
314 | if (bb == m_switch_bb) |
315 | bb = m_final_bb; |
316 | if (!m_contiguous_range || bb != m_default_bb) |
317 | { |
318 | m_reason = reason; |
319 | return false; |
320 | } |
321 | |
322 | unsigned int branch_num = gimple_switch_num_labels (gs: m_switch); |
323 | for (unsigned int i = 1; i < branch_num; i++) |
324 | { |
325 | if (gimple_switch_label_bb (cfun, m_switch, i) == bb) |
326 | { |
327 | m_reason = reason; |
328 | return false; |
329 | } |
330 | } |
331 | m_default_case_nonstandard = true; |
332 | } |
333 | } |
334 | } |
335 | } |
336 | |
337 | return true; |
338 | } |
339 | |
340 | /* The following function allocates default_values, target_{in,out}_names and |
341 | constructors arrays. The last one is also populated with pointers to |
342 | vectors that will become constructors of new arrays. */ |
343 | |
344 | void |
345 | switch_conversion::create_temp_arrays () |
346 | { |
347 | int i; |
348 | |
349 | m_default_values = XCNEWVEC (tree, m_phi_count * 3); |
350 | /* ??? Macros do not support multi argument templates in their |
351 | argument list. We create a typedef to work around that problem. */ |
352 | typedef vec<constructor_elt, va_gc> *vec_constructor_elt_gc; |
353 | m_constructors = XCNEWVEC (vec_constructor_elt_gc, m_phi_count); |
354 | m_target_inbound_names = m_default_values + m_phi_count; |
355 | m_target_outbound_names = m_target_inbound_names + m_phi_count; |
356 | for (i = 0; i < m_phi_count; i++) |
357 | vec_alloc (v&: m_constructors[i], nelems: tree_to_uhwi (m_range_size) + 1); |
358 | } |
359 | |
360 | /* Populate the array of default values in the order of phi nodes. |
361 | DEFAULT_CASE is the CASE_LABEL_EXPR for the default switch branch |
362 | if the range is non-contiguous or the default case has standard |
363 | structure, otherwise it is the first non-default case instead. */ |
364 | |
365 | void |
366 | switch_conversion::gather_default_values (tree default_case) |
367 | { |
368 | gphi_iterator gsi; |
369 | basic_block bb = label_to_block (cfun, CASE_LABEL (default_case)); |
370 | edge e; |
371 | int i = 0; |
372 | |
373 | gcc_assert (CASE_LOW (default_case) == NULL_TREE |
374 | || m_default_case_nonstandard); |
375 | |
376 | if (bb == m_final_bb) |
377 | e = find_edge (m_switch_bb, bb); |
378 | else |
379 | e = single_succ_edge (bb); |
380 | |
381 | for (gsi = gsi_start_phis (m_final_bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
382 | { |
383 | gphi *phi = gsi.phi (); |
384 | if (virtual_operand_p (op: gimple_phi_result (gs: phi))) |
385 | continue; |
386 | tree val = PHI_ARG_DEF_FROM_EDGE (phi, e); |
387 | gcc_assert (val); |
388 | m_default_values[i++] = val; |
389 | } |
390 | } |
391 | |
392 | /* The following function populates the vectors in the constructors array with |
393 | future contents of the static arrays. The vectors are populated in the |
394 | order of phi nodes. */ |
395 | |
396 | void |
397 | switch_conversion::build_constructors () |
398 | { |
399 | unsigned i, branch_num = gimple_switch_num_labels (gs: m_switch); |
400 | tree pos = m_range_min; |
401 | tree pos_one = build_int_cst (TREE_TYPE (pos), 1); |
402 | |
403 | for (i = 1; i < branch_num; i++) |
404 | { |
405 | tree cs = gimple_switch_label (gs: m_switch, index: i); |
406 | basic_block bb = label_to_block (cfun, CASE_LABEL (cs)); |
407 | edge e; |
408 | tree high; |
409 | gphi_iterator gsi; |
410 | int j; |
411 | |
412 | if (bb == m_final_bb) |
413 | e = find_edge (m_switch_bb, bb); |
414 | else |
415 | e = single_succ_edge (bb); |
416 | gcc_assert (e); |
417 | |
418 | while (tree_int_cst_lt (t1: pos, CASE_LOW (cs))) |
419 | { |
420 | int k; |
421 | for (k = 0; k < m_phi_count; k++) |
422 | { |
423 | constructor_elt elt; |
424 | |
425 | elt.index = int_const_binop (MINUS_EXPR, pos, m_range_min); |
426 | if (TYPE_PRECISION (TREE_TYPE (elt.index)) |
427 | > TYPE_PRECISION (sizetype)) |
428 | elt.index = fold_convert (sizetype, elt.index); |
429 | elt.value |
430 | = unshare_expr_without_location (m_default_values[k]); |
431 | m_constructors[k]->quick_push (obj: elt); |
432 | } |
433 | |
434 | pos = int_const_binop (PLUS_EXPR, pos, pos_one); |
435 | } |
436 | gcc_assert (tree_int_cst_equal (pos, CASE_LOW (cs))); |
437 | |
438 | j = 0; |
439 | if (CASE_HIGH (cs)) |
440 | high = CASE_HIGH (cs); |
441 | else |
442 | high = CASE_LOW (cs); |
443 | for (gsi = gsi_start_phis (m_final_bb); |
444 | !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
445 | { |
446 | gphi *phi = gsi.phi (); |
447 | if (virtual_operand_p (op: gimple_phi_result (gs: phi))) |
448 | continue; |
449 | tree val = PHI_ARG_DEF_FROM_EDGE (phi, e); |
450 | tree low = CASE_LOW (cs); |
451 | pos = CASE_LOW (cs); |
452 | |
453 | do |
454 | { |
455 | constructor_elt elt; |
456 | |
457 | elt.index = int_const_binop (MINUS_EXPR, pos, m_range_min); |
458 | if (TYPE_PRECISION (TREE_TYPE (elt.index)) |
459 | > TYPE_PRECISION (sizetype)) |
460 | elt.index = fold_convert (sizetype, elt.index); |
461 | elt.value = unshare_expr_without_location (val); |
462 | m_constructors[j]->quick_push (obj: elt); |
463 | |
464 | pos = int_const_binop (PLUS_EXPR, pos, pos_one); |
465 | } while (!tree_int_cst_lt (t1: high, t2: pos) |
466 | && tree_int_cst_lt (t1: low, t2: pos)); |
467 | j++; |
468 | } |
469 | } |
470 | } |
471 | |
472 | /* If all values in the constructor vector are products of a linear function |
473 | a * x + b, then return true. When true, COEFF_A and COEFF_B and |
474 | coefficients of the linear function. Note that equal values are special |
475 | case of a linear function with a and b equal to zero. */ |
476 | |
477 | bool |
478 | switch_conversion::contains_linear_function_p (vec<constructor_elt, va_gc> *vec, |
479 | wide_int *coeff_a, |
480 | wide_int *coeff_b) |
481 | { |
482 | unsigned int i; |
483 | constructor_elt *elt; |
484 | |
485 | gcc_assert (vec->length () >= 2); |
486 | |
487 | /* Let's try to find any linear function a * x + y that can apply to |
488 | given values. 'a' can be calculated as follows: |
489 | |
490 | a = (y2 - y1) / (x2 - x1) where x2 - x1 = 1 (consecutive case indices) |
491 | a = y2 - y1 |
492 | |
493 | and |
494 | |
495 | b = y2 - a * x2 |
496 | |
497 | */ |
498 | |
499 | tree elt0 = (*vec)[0].value; |
500 | tree elt1 = (*vec)[1].value; |
501 | |
502 | if (TREE_CODE (elt0) != INTEGER_CST || TREE_CODE (elt1) != INTEGER_CST) |
503 | return false; |
504 | |
505 | wide_int range_min |
506 | = wide_int::from (x: wi::to_wide (t: m_range_min), |
507 | TYPE_PRECISION (TREE_TYPE (elt0)), |
508 | TYPE_SIGN (TREE_TYPE (m_range_min))); |
509 | wide_int y1 = wi::to_wide (t: elt0); |
510 | wide_int y2 = wi::to_wide (t: elt1); |
511 | wide_int a = y2 - y1; |
512 | wide_int b = y2 - a * (range_min + 1); |
513 | |
514 | /* Verify that all values fulfill the linear function. */ |
515 | FOR_EACH_VEC_SAFE_ELT (vec, i, elt) |
516 | { |
517 | if (TREE_CODE (elt->value) != INTEGER_CST) |
518 | return false; |
519 | |
520 | wide_int value = wi::to_wide (t: elt->value); |
521 | if (a * range_min + b != value) |
522 | return false; |
523 | |
524 | ++range_min; |
525 | } |
526 | |
527 | *coeff_a = a; |
528 | *coeff_b = b; |
529 | |
530 | return true; |
531 | } |
532 | |
533 | /* Return type which should be used for array elements, either TYPE's |
534 | main variant or, for integral types, some smaller integral type |
535 | that can still hold all the constants. */ |
536 | |
537 | tree |
538 | switch_conversion::array_value_type (tree type, int num) |
539 | { |
540 | unsigned int i, len = vec_safe_length (v: m_constructors[num]); |
541 | constructor_elt *elt; |
542 | int sign = 0; |
543 | tree smaller_type; |
544 | |
545 | /* Types with alignments greater than their size can reach here, e.g. out of |
546 | SRA. We couldn't use these as an array component type so get back to the |
547 | main variant first, which, for our purposes, is fine for other types as |
548 | well. */ |
549 | |
550 | type = TYPE_MAIN_VARIANT (type); |
551 | |
552 | if (!INTEGRAL_TYPE_P (type) |
553 | || (TREE_CODE (type) == BITINT_TYPE |
554 | && (TYPE_PRECISION (type) > MAX_FIXED_MODE_SIZE |
555 | || TYPE_MODE (type) == BLKmode))) |
556 | return type; |
557 | |
558 | scalar_int_mode type_mode = SCALAR_INT_TYPE_MODE (type); |
559 | scalar_int_mode mode = get_narrowest_mode (mode: type_mode); |
560 | if (GET_MODE_SIZE (mode: type_mode) <= GET_MODE_SIZE (mode)) |
561 | return type; |
562 | |
563 | if (len < (optimize_bb_for_size_p (gimple_bb (g: m_switch)) ? 2 : 32)) |
564 | return type; |
565 | |
566 | FOR_EACH_VEC_SAFE_ELT (m_constructors[num], i, elt) |
567 | { |
568 | wide_int cst; |
569 | |
570 | if (TREE_CODE (elt->value) != INTEGER_CST) |
571 | return type; |
572 | |
573 | cst = wi::to_wide (t: elt->value); |
574 | while (1) |
575 | { |
576 | unsigned int prec = GET_MODE_BITSIZE (mode); |
577 | if (prec > HOST_BITS_PER_WIDE_INT) |
578 | return type; |
579 | |
580 | if (sign >= 0 && cst == wi::zext (x: cst, offset: prec)) |
581 | { |
582 | if (sign == 0 && cst == wi::sext (x: cst, offset: prec)) |
583 | break; |
584 | sign = 1; |
585 | break; |
586 | } |
587 | if (sign <= 0 && cst == wi::sext (x: cst, offset: prec)) |
588 | { |
589 | sign = -1; |
590 | break; |
591 | } |
592 | |
593 | if (sign == 1) |
594 | sign = 0; |
595 | |
596 | if (!GET_MODE_WIDER_MODE (m: mode).exists (mode: &mode) |
597 | || GET_MODE_SIZE (mode) >= GET_MODE_SIZE (mode: type_mode)) |
598 | return type; |
599 | } |
600 | } |
601 | |
602 | if (sign == 0) |
603 | sign = TYPE_UNSIGNED (type) ? 1 : -1; |
604 | smaller_type = lang_hooks.types.type_for_mode (mode, sign >= 0); |
605 | if (GET_MODE_SIZE (mode: type_mode) |
606 | <= GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (smaller_type))) |
607 | return type; |
608 | |
609 | return smaller_type; |
610 | } |
611 | |
612 | /* Create an appropriate array type and declaration and assemble a static |
613 | array variable. Also create a load statement that initializes |
614 | the variable in question with a value from the static array. SWTCH is |
615 | the switch statement being converted, NUM is the index to |
616 | arrays of constructors, default values and target SSA names |
617 | for this particular array. ARR_INDEX_TYPE is the type of the index |
618 | of the new array, PHI is the phi node of the final BB that corresponds |
619 | to the value that will be loaded from the created array. TIDX |
620 | is an ssa name of a temporary variable holding the index for loads from the |
621 | new array. */ |
622 | |
623 | void |
624 | switch_conversion::build_one_array (int num, tree arr_index_type, |
625 | gphi *phi, tree tidx) |
626 | { |
627 | tree name; |
628 | gimple *load; |
629 | gimple_stmt_iterator gsi = gsi_for_stmt (m_switch); |
630 | location_t loc = gimple_location (g: m_switch); |
631 | |
632 | gcc_assert (m_default_values[num]); |
633 | |
634 | name = copy_ssa_name (PHI_RESULT (phi)); |
635 | m_target_inbound_names[num] = name; |
636 | |
637 | vec<constructor_elt, va_gc> *constructor = m_constructors[num]; |
638 | wide_int coeff_a, coeff_b; |
639 | bool linear_p = contains_linear_function_p (vec: constructor, coeff_a: &coeff_a, coeff_b: &coeff_b); |
640 | tree type; |
641 | if (linear_p |
642 | && (type = range_check_type (TREE_TYPE ((*constructor)[0].value)))) |
643 | { |
644 | if (dump_file && coeff_a.to_uhwi () > 0) |
645 | fprintf (stream: dump_file, format: "Linear transformation with A = %" PRId64 |
646 | " and B = %" PRId64 "\n" , coeff_a.to_shwi (), |
647 | coeff_b.to_shwi ()); |
648 | |
649 | /* We must use type of constructor values. */ |
650 | gimple_seq seq = NULL; |
651 | tree tmp = gimple_convert (seq: &seq, type, op: m_index_expr); |
652 | tree tmp2 = gimple_build (seq: &seq, code: MULT_EXPR, type, |
653 | ops: wide_int_to_tree (type, cst: coeff_a), ops: tmp); |
654 | tree tmp3 = gimple_build (seq: &seq, code: PLUS_EXPR, type, ops: tmp2, |
655 | ops: wide_int_to_tree (type, cst: coeff_b)); |
656 | tree tmp4 = gimple_convert (seq: &seq, TREE_TYPE (name), op: tmp3); |
657 | gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT); |
658 | load = gimple_build_assign (name, tmp4); |
659 | } |
660 | else |
661 | { |
662 | tree array_type, ctor, decl, value_type, fetch, default_type; |
663 | |
664 | default_type = TREE_TYPE (m_default_values[num]); |
665 | value_type = array_value_type (type: default_type, num); |
666 | array_type = build_array_type (value_type, arr_index_type); |
667 | if (default_type != value_type) |
668 | { |
669 | unsigned int i; |
670 | constructor_elt *elt; |
671 | |
672 | FOR_EACH_VEC_SAFE_ELT (constructor, i, elt) |
673 | elt->value = fold_convert (value_type, elt->value); |
674 | } |
675 | ctor = build_constructor (array_type, constructor); |
676 | TREE_CONSTANT (ctor) = true; |
677 | TREE_STATIC (ctor) = true; |
678 | |
679 | decl = build_decl (loc, VAR_DECL, NULL_TREE, array_type); |
680 | TREE_STATIC (decl) = 1; |
681 | DECL_INITIAL (decl) = ctor; |
682 | |
683 | DECL_NAME (decl) = create_tmp_var_name ("CSWTCH" ); |
684 | DECL_ARTIFICIAL (decl) = 1; |
685 | DECL_IGNORED_P (decl) = 1; |
686 | TREE_CONSTANT (decl) = 1; |
687 | TREE_READONLY (decl) = 1; |
688 | DECL_IGNORED_P (decl) = 1; |
689 | if (offloading_function_p (cfun->decl)) |
690 | DECL_ATTRIBUTES (decl) |
691 | = tree_cons (get_identifier ("omp declare target" ), NULL_TREE, |
692 | NULL_TREE); |
693 | varpool_node::finalize_decl (decl); |
694 | |
695 | fetch = build4 (ARRAY_REF, value_type, decl, tidx, NULL_TREE, |
696 | NULL_TREE); |
697 | if (default_type != value_type) |
698 | { |
699 | fetch = fold_convert (default_type, fetch); |
700 | fetch = force_gimple_operand_gsi (&gsi, fetch, true, NULL_TREE, |
701 | true, GSI_SAME_STMT); |
702 | } |
703 | load = gimple_build_assign (name, fetch); |
704 | } |
705 | |
706 | gsi_insert_before (&gsi, load, GSI_SAME_STMT); |
707 | update_stmt (s: load); |
708 | m_arr_ref_last = load; |
709 | } |
710 | |
711 | /* Builds and initializes static arrays initialized with values gathered from |
712 | the switch statement. Also creates statements that load values from |
713 | them. */ |
714 | |
715 | void |
716 | switch_conversion::build_arrays () |
717 | { |
718 | tree arr_index_type; |
719 | tree tidx, sub, utype, tidxtype; |
720 | gimple *stmt; |
721 | gimple_stmt_iterator gsi; |
722 | gphi_iterator gpi; |
723 | int i; |
724 | location_t loc = gimple_location (g: m_switch); |
725 | |
726 | gsi = gsi_for_stmt (m_switch); |
727 | |
728 | /* Make sure we do not generate arithmetics in a subrange. */ |
729 | utype = TREE_TYPE (m_index_expr); |
730 | if (TREE_TYPE (utype)) |
731 | utype = lang_hooks.types.type_for_mode (TYPE_MODE (TREE_TYPE (utype)), 1); |
732 | else if (TREE_CODE (utype) == BITINT_TYPE |
733 | && (TYPE_PRECISION (utype) > MAX_FIXED_MODE_SIZE |
734 | || TYPE_MODE (utype) == BLKmode)) |
735 | utype = unsigned_type_for (utype); |
736 | else |
737 | utype = lang_hooks.types.type_for_mode (TYPE_MODE (utype), 1); |
738 | if (TYPE_PRECISION (utype) > TYPE_PRECISION (sizetype)) |
739 | tidxtype = sizetype; |
740 | else |
741 | tidxtype = utype; |
742 | |
743 | arr_index_type = build_index_type (m_range_size); |
744 | tidx = make_ssa_name (var: tidxtype); |
745 | sub = fold_build2_loc (loc, MINUS_EXPR, utype, |
746 | fold_convert_loc (loc, utype, m_index_expr), |
747 | fold_convert_loc (loc, utype, m_range_min)); |
748 | sub = fold_convert (tidxtype, sub); |
749 | sub = force_gimple_operand_gsi (&gsi, sub, |
750 | false, NULL, true, GSI_SAME_STMT); |
751 | stmt = gimple_build_assign (tidx, sub); |
752 | |
753 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); |
754 | update_stmt (s: stmt); |
755 | m_arr_ref_first = stmt; |
756 | |
757 | for (gpi = gsi_start_phis (m_final_bb), i = 0; |
758 | !gsi_end_p (i: gpi); gsi_next (i: &gpi)) |
759 | { |
760 | gphi *phi = gpi.phi (); |
761 | if (!virtual_operand_p (op: gimple_phi_result (gs: phi))) |
762 | build_one_array (num: i++, arr_index_type, phi, tidx); |
763 | else |
764 | { |
765 | edge e; |
766 | edge_iterator ei; |
767 | FOR_EACH_EDGE (e, ei, m_switch_bb->succs) |
768 | { |
769 | if (e->dest == m_final_bb) |
770 | break; |
771 | if (!m_default_case_nonstandard |
772 | || e->dest != m_default_bb) |
773 | { |
774 | e = single_succ_edge (bb: e->dest); |
775 | break; |
776 | } |
777 | } |
778 | gcc_assert (e && e->dest == m_final_bb); |
779 | m_target_vop = PHI_ARG_DEF_FROM_EDGE (phi, e); |
780 | } |
781 | } |
782 | } |
783 | |
784 | /* Generates and appropriately inserts loads of default values at the position |
785 | given by GSI. Returns the last inserted statement. */ |
786 | |
787 | gassign * |
788 | switch_conversion::gen_def_assigns (gimple_stmt_iterator *gsi) |
789 | { |
790 | int i; |
791 | gassign *assign = NULL; |
792 | |
793 | for (i = 0; i < m_phi_count; i++) |
794 | { |
795 | tree name = copy_ssa_name (var: m_target_inbound_names[i]); |
796 | m_target_outbound_names[i] = name; |
797 | assign = gimple_build_assign (name, m_default_values[i]); |
798 | gsi_insert_before (gsi, assign, GSI_SAME_STMT); |
799 | update_stmt (s: assign); |
800 | } |
801 | return assign; |
802 | } |
803 | |
804 | /* Deletes the unused bbs and edges that now contain the switch statement and |
805 | its empty branch bbs. BBD is the now dead BB containing |
806 | the original switch statement, FINAL is the last BB of the converted |
807 | switch statement (in terms of succession). */ |
808 | |
809 | void |
810 | switch_conversion::prune_bbs (basic_block bbd, basic_block final, |
811 | basic_block default_bb) |
812 | { |
813 | edge_iterator ei; |
814 | edge e; |
815 | |
816 | for (ei = ei_start (bbd->succs); (e = ei_safe_edge (i: ei)); ) |
817 | { |
818 | basic_block bb; |
819 | bb = e->dest; |
820 | remove_edge (e); |
821 | if (bb != final && bb != default_bb) |
822 | delete_basic_block (bb); |
823 | } |
824 | delete_basic_block (bbd); |
825 | } |
826 | |
827 | /* Add values to phi nodes in final_bb for the two new edges. E1F is the edge |
828 | from the basic block loading values from an array and E2F from the basic |
829 | block loading default values. BBF is the last switch basic block (see the |
830 | bbf description in the comment below). */ |
831 | |
832 | void |
833 | switch_conversion::fix_phi_nodes (edge e1f, edge e2f, basic_block bbf) |
834 | { |
835 | gphi_iterator gsi; |
836 | int i; |
837 | |
838 | for (gsi = gsi_start_phis (bbf), i = 0; |
839 | !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
840 | { |
841 | gphi *phi = gsi.phi (); |
842 | tree inbound, outbound; |
843 | if (virtual_operand_p (op: gimple_phi_result (gs: phi))) |
844 | inbound = outbound = m_target_vop; |
845 | else |
846 | { |
847 | inbound = m_target_inbound_names[i]; |
848 | outbound = m_target_outbound_names[i++]; |
849 | } |
850 | add_phi_arg (phi, inbound, e1f, UNKNOWN_LOCATION); |
851 | if (!m_default_case_nonstandard) |
852 | add_phi_arg (phi, outbound, e2f, UNKNOWN_LOCATION); |
853 | } |
854 | } |
855 | |
856 | /* Creates a check whether the switch expression value actually falls into the |
857 | range given by all the cases. If it does not, the temporaries are loaded |
858 | with default values instead. */ |
859 | |
860 | void |
861 | switch_conversion::gen_inbound_check () |
862 | { |
863 | tree label_decl1 = create_artificial_label (UNKNOWN_LOCATION); |
864 | tree label_decl2 = create_artificial_label (UNKNOWN_LOCATION); |
865 | tree label_decl3 = create_artificial_label (UNKNOWN_LOCATION); |
866 | glabel *label1, *label2, *label3; |
867 | tree utype, tidx; |
868 | tree bound; |
869 | |
870 | gcond *cond_stmt; |
871 | |
872 | gassign *last_assign = NULL; |
873 | gimple_stmt_iterator gsi; |
874 | basic_block bb0, bb1, bb2, bbf, bbd; |
875 | edge e01 = NULL, e02, e21, e1d, e1f, e2f; |
876 | location_t loc = gimple_location (g: m_switch); |
877 | |
878 | gcc_assert (m_default_values); |
879 | |
880 | bb0 = gimple_bb (g: m_switch); |
881 | |
882 | tidx = gimple_assign_lhs (gs: m_arr_ref_first); |
883 | utype = TREE_TYPE (tidx); |
884 | |
885 | /* (end of) block 0 */ |
886 | gsi = gsi_for_stmt (m_arr_ref_first); |
887 | gsi_next (i: &gsi); |
888 | |
889 | bound = fold_convert_loc (loc, utype, m_range_size); |
890 | cond_stmt = gimple_build_cond (LE_EXPR, tidx, bound, NULL_TREE, NULL_TREE); |
891 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); |
892 | update_stmt (s: cond_stmt); |
893 | |
894 | /* block 2 */ |
895 | if (!m_default_case_nonstandard) |
896 | { |
897 | label2 = gimple_build_label (label: label_decl2); |
898 | gsi_insert_before (&gsi, label2, GSI_SAME_STMT); |
899 | last_assign = gen_def_assigns (gsi: &gsi); |
900 | } |
901 | |
902 | /* block 1 */ |
903 | label1 = gimple_build_label (label: label_decl1); |
904 | gsi_insert_before (&gsi, label1, GSI_SAME_STMT); |
905 | |
906 | /* block F */ |
907 | gsi = gsi_start_bb (bb: m_final_bb); |
908 | label3 = gimple_build_label (label: label_decl3); |
909 | gsi_insert_before (&gsi, label3, GSI_SAME_STMT); |
910 | |
911 | /* cfg fix */ |
912 | e02 = split_block (bb0, cond_stmt); |
913 | bb2 = e02->dest; |
914 | |
915 | if (m_default_case_nonstandard) |
916 | { |
917 | bb1 = bb2; |
918 | bb2 = m_default_bb; |
919 | e01 = e02; |
920 | e01->flags = EDGE_TRUE_VALUE; |
921 | e02 = make_edge (bb0, bb2, EDGE_FALSE_VALUE); |
922 | edge e_default = find_edge (bb1, bb2); |
923 | for (gphi_iterator gsi = gsi_start_phis (bb2); |
924 | !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
925 | { |
926 | gphi *phi = gsi.phi (); |
927 | tree arg = PHI_ARG_DEF_FROM_EDGE (phi, e_default); |
928 | add_phi_arg (phi, arg, e02, |
929 | gimple_phi_arg_location_from_edge (phi, e: e_default)); |
930 | } |
931 | /* Partially fix the dominator tree, if it is available. */ |
932 | if (dom_info_available_p (CDI_DOMINATORS)) |
933 | redirect_immediate_dominators (CDI_DOMINATORS, bb1, bb0); |
934 | } |
935 | else |
936 | { |
937 | e21 = split_block (bb2, last_assign); |
938 | bb1 = e21->dest; |
939 | remove_edge (e21); |
940 | } |
941 | |
942 | e1d = split_block (bb1, m_arr_ref_last); |
943 | bbd = e1d->dest; |
944 | remove_edge (e1d); |
945 | |
946 | /* Flags and profiles of the edge for in-range values. */ |
947 | if (!m_default_case_nonstandard) |
948 | e01 = make_edge (bb0, bb1, EDGE_TRUE_VALUE); |
949 | e01->probability = m_default_prob.invert (); |
950 | |
951 | /* Flags and profiles of the edge taking care of out-of-range values. */ |
952 | e02->flags &= ~EDGE_FALLTHRU; |
953 | e02->flags |= EDGE_FALSE_VALUE; |
954 | e02->probability = m_default_prob; |
955 | |
956 | bbf = m_final_bb; |
957 | |
958 | e1f = make_edge (bb1, bbf, EDGE_FALLTHRU); |
959 | e1f->probability = profile_probability::always (); |
960 | |
961 | if (m_default_case_nonstandard) |
962 | e2f = NULL; |
963 | else |
964 | { |
965 | e2f = make_edge (bb2, bbf, EDGE_FALLTHRU); |
966 | e2f->probability = profile_probability::always (); |
967 | } |
968 | |
969 | /* frequencies of the new BBs */ |
970 | bb1->count = e01->count (); |
971 | bb2->count = e02->count (); |
972 | if (!m_default_case_nonstandard) |
973 | bbf->count = e1f->count () + e2f->count (); |
974 | |
975 | /* Tidy blocks that have become unreachable. */ |
976 | prune_bbs (bbd, final: m_final_bb, |
977 | default_bb: m_default_case_nonstandard ? m_default_bb : NULL); |
978 | |
979 | /* Fixup the PHI nodes in bbF. */ |
980 | fix_phi_nodes (e1f, e2f, bbf); |
981 | |
982 | /* Fix the dominator tree, if it is available. */ |
983 | if (dom_info_available_p (CDI_DOMINATORS)) |
984 | { |
985 | vec<basic_block> bbs_to_fix_dom; |
986 | |
987 | set_immediate_dominator (CDI_DOMINATORS, bb1, bb0); |
988 | if (!m_default_case_nonstandard) |
989 | set_immediate_dominator (CDI_DOMINATORS, bb2, bb0); |
990 | if (! get_immediate_dominator (CDI_DOMINATORS, bbf)) |
991 | /* If bbD was the immediate dominator ... */ |
992 | set_immediate_dominator (CDI_DOMINATORS, bbf, bb0); |
993 | |
994 | bbs_to_fix_dom.create (nelems: 3 + (bb2 != bbf)); |
995 | bbs_to_fix_dom.quick_push (obj: bb0); |
996 | bbs_to_fix_dom.quick_push (obj: bb1); |
997 | if (bb2 != bbf) |
998 | bbs_to_fix_dom.quick_push (obj: bb2); |
999 | bbs_to_fix_dom.quick_push (obj: bbf); |
1000 | |
1001 | iterate_fix_dominators (CDI_DOMINATORS, bbs_to_fix_dom, true); |
1002 | bbs_to_fix_dom.release (); |
1003 | } |
1004 | } |
1005 | |
1006 | /* The following function is invoked on every switch statement (the current |
1007 | one is given in SWTCH) and runs the individual phases of switch |
1008 | conversion on it one after another until one fails or the conversion |
1009 | is completed. On success, NULL is in m_reason, otherwise points |
1010 | to a string with the reason why the conversion failed. */ |
1011 | |
1012 | void |
1013 | switch_conversion::expand (gswitch *swtch) |
1014 | { |
1015 | /* Group case labels so that we get the right results from the heuristics |
1016 | that decide on the code generation approach for this switch. */ |
1017 | m_cfg_altered |= group_case_labels_stmt (swtch); |
1018 | |
1019 | /* If this switch is now a degenerate case with only a default label, |
1020 | there is nothing left for us to do. */ |
1021 | if (gimple_switch_num_labels (gs: swtch) < 2) |
1022 | { |
1023 | m_reason = "switch is a degenerate case" ; |
1024 | return; |
1025 | } |
1026 | |
1027 | collect (swtch); |
1028 | |
1029 | /* No error markers should reach here (they should be filtered out |
1030 | during gimplification). */ |
1031 | gcc_checking_assert (TREE_TYPE (m_index_expr) != error_mark_node); |
1032 | |
1033 | /* Prefer bit test if possible. */ |
1034 | if (tree_fits_uhwi_p (m_range_size) |
1035 | && bit_test_cluster::can_be_handled (range: tree_to_uhwi (m_range_size), uniq: m_uniq) |
1036 | && bit_test_cluster::is_beneficial (count: m_count, uniq: m_uniq)) |
1037 | { |
1038 | m_reason = "expanding as bit test is preferable" ; |
1039 | return; |
1040 | } |
1041 | |
1042 | if (m_uniq <= 2) |
1043 | { |
1044 | /* This will be expanded as a decision tree . */ |
1045 | m_reason = "expanding as jumps is preferable" ; |
1046 | return; |
1047 | } |
1048 | |
1049 | /* If there is no common successor, we cannot do the transformation. */ |
1050 | if (!m_final_bb) |
1051 | { |
1052 | m_reason = "no common successor to all case label target blocks found" ; |
1053 | return; |
1054 | } |
1055 | |
1056 | /* Check the case label values are within reasonable range: */ |
1057 | if (!check_range ()) |
1058 | { |
1059 | gcc_assert (m_reason); |
1060 | return; |
1061 | } |
1062 | |
1063 | /* For all the cases, see whether they are empty, the assignments they |
1064 | represent constant and so on... */ |
1065 | if (!check_all_empty_except_final ()) |
1066 | { |
1067 | gcc_assert (m_reason); |
1068 | return; |
1069 | } |
1070 | if (!check_final_bb ()) |
1071 | { |
1072 | gcc_assert (m_reason); |
1073 | return; |
1074 | } |
1075 | |
1076 | /* At this point all checks have passed and we can proceed with the |
1077 | transformation. */ |
1078 | |
1079 | create_temp_arrays (); |
1080 | gather_default_values (default_case: m_default_case_nonstandard |
1081 | ? gimple_switch_label (gs: swtch, index: 1) |
1082 | : gimple_switch_default_label (gs: swtch)); |
1083 | build_constructors (); |
1084 | |
1085 | build_arrays (); /* Build the static arrays and assignments. */ |
1086 | gen_inbound_check (); /* Build the bounds check. */ |
1087 | |
1088 | m_cfg_altered = true; |
1089 | } |
1090 | |
1091 | /* Destructor. */ |
1092 | |
1093 | switch_conversion::~switch_conversion () |
1094 | { |
1095 | XDELETEVEC (m_constructors); |
1096 | XDELETEVEC (m_default_values); |
1097 | } |
1098 | |
1099 | /* Constructor. */ |
1100 | |
1101 | group_cluster::group_cluster (vec<cluster *> &clusters, |
1102 | unsigned start, unsigned end) |
1103 | { |
1104 | gcc_checking_assert (end - start + 1 >= 1); |
1105 | m_prob = profile_probability::never (); |
1106 | m_cases.create (nelems: end - start + 1); |
1107 | for (unsigned i = start; i <= end; i++) |
1108 | { |
1109 | m_cases.quick_push (obj: static_cast<simple_cluster *> (clusters[i])); |
1110 | m_prob += clusters[i]->m_prob; |
1111 | } |
1112 | m_subtree_prob = m_prob; |
1113 | } |
1114 | |
1115 | /* Destructor. */ |
1116 | |
1117 | group_cluster::~group_cluster () |
1118 | { |
1119 | for (unsigned i = 0; i < m_cases.length (); i++) |
1120 | delete m_cases[i]; |
1121 | |
1122 | m_cases.release (); |
1123 | } |
1124 | |
1125 | /* Dump content of a cluster. */ |
1126 | |
1127 | void |
1128 | group_cluster::dump (FILE *f, bool details) |
1129 | { |
1130 | unsigned total_values = 0; |
1131 | for (unsigned i = 0; i < m_cases.length (); i++) |
1132 | total_values += m_cases[i]->get_range (low: m_cases[i]->get_low (), |
1133 | high: m_cases[i]->get_high ()); |
1134 | |
1135 | unsigned comparison_count = 0; |
1136 | for (unsigned i = 0; i < m_cases.length (); i++) |
1137 | { |
1138 | simple_cluster *sc = static_cast<simple_cluster *> (m_cases[i]); |
1139 | comparison_count += sc->get_comparison_count (); |
1140 | } |
1141 | |
1142 | unsigned HOST_WIDE_INT range = get_range (low: get_low (), high: get_high ()); |
1143 | fprintf (stream: f, format: "%s" , get_type () == JUMP_TABLE ? "JT" : "BT" ); |
1144 | |
1145 | if (details) |
1146 | fprintf (stream: f, format: "(values:%d comparisons:%d range:" HOST_WIDE_INT_PRINT_DEC |
1147 | " density: %.2f%%)" , total_values, comparison_count, range, |
1148 | 100.0f * comparison_count / range); |
1149 | |
1150 | fprintf (stream: f, format: ":" ); |
1151 | PRINT_CASE (f, get_low ()); |
1152 | fprintf (stream: f, format: "-" ); |
1153 | PRINT_CASE (f, get_high ()); |
1154 | fprintf (stream: f, format: " " ); |
1155 | } |
1156 | |
1157 | /* Emit GIMPLE code to handle the cluster. */ |
1158 | |
1159 | void |
1160 | jump_table_cluster::emit (tree index_expr, tree, |
1161 | tree default_label_expr, basic_block default_bb, |
1162 | location_t loc) |
1163 | { |
1164 | tree low = get_low (); |
1165 | unsigned HOST_WIDE_INT range = get_range (low, high: get_high ()); |
1166 | unsigned HOST_WIDE_INT nondefault_range = 0; |
1167 | bool bitint = false; |
1168 | gimple_stmt_iterator gsi = gsi_start_bb (bb: m_case_bb); |
1169 | |
1170 | /* For large/huge _BitInt, subtract low from index_expr, cast to unsigned |
1171 | DImode type (get_range doesn't support ranges larger than 64-bits) |
1172 | and subtract low from all case values as well. */ |
1173 | if (TREE_CODE (TREE_TYPE (index_expr)) == BITINT_TYPE |
1174 | && TYPE_PRECISION (TREE_TYPE (index_expr)) > GET_MODE_PRECISION (DImode)) |
1175 | { |
1176 | bitint = true; |
1177 | tree this_low = low, type; |
1178 | gimple *g; |
1179 | gimple_seq seq = NULL; |
1180 | if (!TYPE_OVERFLOW_WRAPS (TREE_TYPE (index_expr))) |
1181 | { |
1182 | type = unsigned_type_for (TREE_TYPE (index_expr)); |
1183 | index_expr = gimple_convert (seq: &seq, type, op: index_expr); |
1184 | this_low = fold_convert (type, this_low); |
1185 | } |
1186 | this_low = const_unop (NEGATE_EXPR, TREE_TYPE (this_low), this_low); |
1187 | index_expr = gimple_build (seq: &seq, code: PLUS_EXPR, TREE_TYPE (index_expr), |
1188 | ops: index_expr, ops: this_low); |
1189 | type = build_nonstandard_integer_type (GET_MODE_PRECISION (DImode), 1); |
1190 | g = gimple_build_cond (GT_EXPR, index_expr, |
1191 | fold_convert (TREE_TYPE (index_expr), |
1192 | TYPE_MAX_VALUE (type)), |
1193 | NULL_TREE, NULL_TREE); |
1194 | gimple_seq_add_stmt (&seq, g); |
1195 | gimple_seq_set_location (seq, loc); |
1196 | gsi_insert_seq_after (&gsi, seq, GSI_NEW_STMT); |
1197 | edge e1 = split_block (m_case_bb, g); |
1198 | e1->flags = EDGE_FALSE_VALUE; |
1199 | e1->probability = profile_probability::likely (); |
1200 | edge e2 = make_edge (e1->src, default_bb, EDGE_TRUE_VALUE); |
1201 | e2->probability = e1->probability.invert (); |
1202 | gsi = gsi_start_bb (bb: e1->dest); |
1203 | seq = NULL; |
1204 | index_expr = gimple_convert (seq: &seq, type, op: index_expr); |
1205 | gimple_seq_set_location (seq, loc); |
1206 | gsi_insert_seq_after (&gsi, seq, GSI_NEW_STMT); |
1207 | } |
1208 | |
1209 | /* For jump table we just emit a new gswitch statement that will |
1210 | be latter lowered to jump table. */ |
1211 | auto_vec <tree> labels; |
1212 | labels.create (nelems: m_cases.length ()); |
1213 | |
1214 | basic_block case_bb = gsi_bb (i: gsi); |
1215 | make_edge (case_bb, default_bb, 0); |
1216 | for (unsigned i = 0; i < m_cases.length (); i++) |
1217 | { |
1218 | tree lab = unshare_expr (m_cases[i]->m_case_label_expr); |
1219 | if (bitint) |
1220 | { |
1221 | CASE_LOW (lab) |
1222 | = fold_convert (TREE_TYPE (index_expr), |
1223 | const_binop (MINUS_EXPR, |
1224 | TREE_TYPE (CASE_LOW (lab)), |
1225 | CASE_LOW (lab), low)); |
1226 | if (CASE_HIGH (lab)) |
1227 | CASE_HIGH (lab) |
1228 | = fold_convert (TREE_TYPE (index_expr), |
1229 | const_binop (MINUS_EXPR, |
1230 | TREE_TYPE (CASE_HIGH (lab)), |
1231 | CASE_HIGH (lab), low)); |
1232 | } |
1233 | labels.quick_push (obj: lab); |
1234 | make_edge (case_bb, m_cases[i]->m_case_bb, 0); |
1235 | } |
1236 | |
1237 | gswitch *s = gimple_build_switch (index_expr, |
1238 | unshare_expr (default_label_expr), labels); |
1239 | gimple_set_location (g: s, location: loc); |
1240 | gsi_insert_after (&gsi, s, GSI_NEW_STMT); |
1241 | |
1242 | /* Set up even probabilities for all cases. */ |
1243 | for (unsigned i = 0; i < m_cases.length (); i++) |
1244 | { |
1245 | simple_cluster *sc = static_cast<simple_cluster *> (m_cases[i]); |
1246 | edge case_edge = find_edge (case_bb, sc->m_case_bb); |
1247 | unsigned HOST_WIDE_INT case_range |
1248 | = sc->get_range (low: sc->get_low (), high: sc->get_high ()); |
1249 | nondefault_range += case_range; |
1250 | |
1251 | /* case_edge->aux is number of values in a jump-table that are covered |
1252 | by the case_edge. */ |
1253 | case_edge->aux = (void *) ((intptr_t) (case_edge->aux) + case_range); |
1254 | } |
1255 | |
1256 | edge default_edge = gimple_switch_default_edge (cfun, s); |
1257 | default_edge->probability = profile_probability::never (); |
1258 | |
1259 | for (unsigned i = 0; i < m_cases.length (); i++) |
1260 | { |
1261 | simple_cluster *sc = static_cast<simple_cluster *> (m_cases[i]); |
1262 | edge case_edge = find_edge (case_bb, sc->m_case_bb); |
1263 | case_edge->probability |
1264 | = profile_probability::always ().apply_scale (num: (intptr_t)case_edge->aux, |
1265 | den: range); |
1266 | } |
1267 | |
1268 | /* Number of non-default values is probability of default edge. */ |
1269 | default_edge->probability |
1270 | += profile_probability::always ().apply_scale (num: nondefault_range, |
1271 | den: range).invert (); |
1272 | |
1273 | switch_decision_tree::reset_out_edges_aux (swtch: s); |
1274 | } |
1275 | |
1276 | /* Find jump tables of given CLUSTERS, where all members of the vector |
1277 | are of type simple_cluster. New clusters are returned. */ |
1278 | |
1279 | vec<cluster *> |
1280 | jump_table_cluster::find_jump_tables (vec<cluster *> &clusters) |
1281 | { |
1282 | if (!is_enabled ()) |
1283 | return clusters.copy (); |
1284 | |
1285 | unsigned l = clusters.length (); |
1286 | auto_vec<min_cluster_item> min; |
1287 | min.reserve (nelems: l + 1); |
1288 | |
1289 | min.quick_push (obj: min_cluster_item (0, 0, 0)); |
1290 | |
1291 | unsigned HOST_WIDE_INT max_ratio |
1292 | = (optimize_insn_for_size_p () |
1293 | ? param_jump_table_max_growth_ratio_for_size |
1294 | : param_jump_table_max_growth_ratio_for_speed); |
1295 | |
1296 | for (unsigned i = 1; i <= l; i++) |
1297 | { |
1298 | /* Set minimal # of clusters with i-th item to infinite. */ |
1299 | min.quick_push (obj: min_cluster_item (INT_MAX, INT_MAX, INT_MAX)); |
1300 | |
1301 | /* Pre-calculate number of comparisons for the clusters. */ |
1302 | HOST_WIDE_INT comparison_count = 0; |
1303 | for (unsigned k = 0; k <= i - 1; k++) |
1304 | { |
1305 | simple_cluster *sc = static_cast<simple_cluster *> (clusters[k]); |
1306 | comparison_count += sc->get_comparison_count (); |
1307 | } |
1308 | |
1309 | for (unsigned j = 0; j < i; j++) |
1310 | { |
1311 | unsigned HOST_WIDE_INT s = min[j].m_non_jt_cases; |
1312 | if (i - j < case_values_threshold ()) |
1313 | s += i - j; |
1314 | |
1315 | /* Prefer clusters with smaller number of numbers covered. */ |
1316 | if ((min[j].m_count + 1 < min[i].m_count |
1317 | || (min[j].m_count + 1 == min[i].m_count |
1318 | && s < min[i].m_non_jt_cases)) |
1319 | && can_be_handled (clusters, start: j, end: i - 1, max_ratio, |
1320 | comparison_count)) |
1321 | min[i] = min_cluster_item (min[j].m_count + 1, j, s); |
1322 | |
1323 | simple_cluster *sc = static_cast<simple_cluster *> (clusters[j]); |
1324 | comparison_count -= sc->get_comparison_count (); |
1325 | } |
1326 | |
1327 | gcc_checking_assert (comparison_count == 0); |
1328 | gcc_checking_assert (min[i].m_count != INT_MAX); |
1329 | } |
1330 | |
1331 | /* No result. */ |
1332 | if (min[l].m_count == l) |
1333 | return clusters.copy (); |
1334 | |
1335 | vec<cluster *> output; |
1336 | output.create (nelems: 4); |
1337 | |
1338 | /* Find and build the clusters. */ |
1339 | for (unsigned int end = l;;) |
1340 | { |
1341 | int start = min[end].m_start; |
1342 | |
1343 | /* Do not allow clusters with small number of cases. */ |
1344 | if (is_beneficial (clusters, start, end: end - 1)) |
1345 | output.safe_push (obj: new jump_table_cluster (clusters, start, end - 1)); |
1346 | else |
1347 | for (int i = end - 1; i >= start; i--) |
1348 | output.safe_push (obj: clusters[i]); |
1349 | |
1350 | end = start; |
1351 | |
1352 | if (start <= 0) |
1353 | break; |
1354 | } |
1355 | |
1356 | output.reverse (); |
1357 | return output; |
1358 | } |
1359 | |
1360 | /* Return true when cluster starting at START and ending at END (inclusive) |
1361 | can build a jump-table. */ |
1362 | |
1363 | bool |
1364 | jump_table_cluster::can_be_handled (const vec<cluster *> &clusters, |
1365 | unsigned start, unsigned end, |
1366 | unsigned HOST_WIDE_INT max_ratio, |
1367 | unsigned HOST_WIDE_INT comparison_count) |
1368 | { |
1369 | /* If the switch is relatively small such that the cost of one |
1370 | indirect jump on the target are higher than the cost of a |
1371 | decision tree, go with the decision tree. |
1372 | |
1373 | If range of values is much bigger than number of values, |
1374 | or if it is too large to represent in a HOST_WIDE_INT, |
1375 | make a sequence of conditional branches instead of a dispatch. |
1376 | |
1377 | The definition of "much bigger" depends on whether we are |
1378 | optimizing for size or for speed. |
1379 | |
1380 | For algorithm correctness, jump table for a single case must return |
1381 | true. We bail out in is_beneficial if it's called just for |
1382 | a single case. */ |
1383 | if (start == end) |
1384 | return true; |
1385 | |
1386 | unsigned HOST_WIDE_INT range = get_range (low: clusters[start]->get_low (), |
1387 | high: clusters[end]->get_high ()); |
1388 | /* Check overflow. */ |
1389 | if (range == 0) |
1390 | return false; |
1391 | |
1392 | if (range > HOST_WIDE_INT_M1U / 100) |
1393 | return false; |
1394 | |
1395 | unsigned HOST_WIDE_INT lhs = 100 * range; |
1396 | if (lhs < range) |
1397 | return false; |
1398 | |
1399 | return lhs <= max_ratio * comparison_count; |
1400 | } |
1401 | |
1402 | /* Return true if cluster starting at START and ending at END (inclusive) |
1403 | is profitable transformation. */ |
1404 | |
1405 | bool |
1406 | jump_table_cluster::is_beneficial (const vec<cluster *> &, |
1407 | unsigned start, unsigned end) |
1408 | { |
1409 | /* Single case bail out. */ |
1410 | if (start == end) |
1411 | return false; |
1412 | |
1413 | return end - start + 1 >= case_values_threshold (); |
1414 | } |
1415 | |
1416 | /* Find bit tests of given CLUSTERS, where all members of the vector |
1417 | are of type simple_cluster. New clusters are returned. */ |
1418 | |
1419 | vec<cluster *> |
1420 | bit_test_cluster::find_bit_tests (vec<cluster *> &clusters) |
1421 | { |
1422 | if (!is_enabled ()) |
1423 | return clusters.copy (); |
1424 | |
1425 | unsigned l = clusters.length (); |
1426 | auto_vec<min_cluster_item> min; |
1427 | min.reserve (nelems: l + 1); |
1428 | |
1429 | min.quick_push (obj: min_cluster_item (0, 0, 0)); |
1430 | |
1431 | for (unsigned i = 1; i <= l; i++) |
1432 | { |
1433 | /* Set minimal # of clusters with i-th item to infinite. */ |
1434 | min.quick_push (obj: min_cluster_item (INT_MAX, INT_MAX, INT_MAX)); |
1435 | |
1436 | for (unsigned j = 0; j < i; j++) |
1437 | { |
1438 | if (min[j].m_count + 1 < min[i].m_count |
1439 | && can_be_handled (clusters, start: j, end: i - 1)) |
1440 | min[i] = min_cluster_item (min[j].m_count + 1, j, INT_MAX); |
1441 | } |
1442 | |
1443 | gcc_checking_assert (min[i].m_count != INT_MAX); |
1444 | } |
1445 | |
1446 | /* No result. */ |
1447 | if (min[l].m_count == l) |
1448 | return clusters.copy (); |
1449 | |
1450 | vec<cluster *> output; |
1451 | output.create (nelems: 4); |
1452 | |
1453 | /* Find and build the clusters. */ |
1454 | for (unsigned end = l;;) |
1455 | { |
1456 | int start = min[end].m_start; |
1457 | |
1458 | if (is_beneficial (clusters, start, end: end - 1)) |
1459 | { |
1460 | bool entire = start == 0 && end == clusters.length (); |
1461 | output.safe_push (obj: new bit_test_cluster (clusters, start, end - 1, |
1462 | entire)); |
1463 | } |
1464 | else |
1465 | for (int i = end - 1; i >= start; i--) |
1466 | output.safe_push (obj: clusters[i]); |
1467 | |
1468 | end = start; |
1469 | |
1470 | if (start <= 0) |
1471 | break; |
1472 | } |
1473 | |
1474 | output.reverse (); |
1475 | return output; |
1476 | } |
1477 | |
1478 | /* Return true when RANGE of case values with UNIQ labels |
1479 | can build a bit test. */ |
1480 | |
1481 | bool |
1482 | bit_test_cluster::can_be_handled (unsigned HOST_WIDE_INT range, |
1483 | unsigned int uniq) |
1484 | { |
1485 | /* Check overflow. */ |
1486 | if (range == 0) |
1487 | return false; |
1488 | |
1489 | if (range >= GET_MODE_BITSIZE (mode: word_mode)) |
1490 | return false; |
1491 | |
1492 | return uniq <= m_max_case_bit_tests; |
1493 | } |
1494 | |
1495 | /* Return true when cluster starting at START and ending at END (inclusive) |
1496 | can build a bit test. */ |
1497 | |
1498 | bool |
1499 | bit_test_cluster::can_be_handled (const vec<cluster *> &clusters, |
1500 | unsigned start, unsigned end) |
1501 | { |
1502 | auto_vec<int, m_max_case_bit_tests> dest_bbs; |
1503 | /* For algorithm correctness, bit test for a single case must return |
1504 | true. We bail out in is_beneficial if it's called just for |
1505 | a single case. */ |
1506 | if (start == end) |
1507 | return true; |
1508 | |
1509 | unsigned HOST_WIDE_INT range = get_range (low: clusters[start]->get_low (), |
1510 | high: clusters[end]->get_high ()); |
1511 | |
1512 | /* Make a guess first. */ |
1513 | if (!can_be_handled (range, uniq: m_max_case_bit_tests)) |
1514 | return false; |
1515 | |
1516 | for (unsigned i = start; i <= end; i++) |
1517 | { |
1518 | simple_cluster *sc = static_cast<simple_cluster *> (clusters[i]); |
1519 | /* m_max_case_bit_tests is very small integer, thus the operation |
1520 | is constant. */ |
1521 | if (!dest_bbs.contains (search: sc->m_case_bb->index)) |
1522 | { |
1523 | if (dest_bbs.length () >= m_max_case_bit_tests) |
1524 | return false; |
1525 | dest_bbs.quick_push (obj: sc->m_case_bb->index); |
1526 | } |
1527 | } |
1528 | |
1529 | return true; |
1530 | } |
1531 | |
1532 | /* Return true when COUNT of cases of UNIQ labels is beneficial for bit test |
1533 | transformation. */ |
1534 | |
1535 | bool |
1536 | bit_test_cluster::is_beneficial (unsigned count, unsigned uniq) |
1537 | { |
1538 | return (((uniq == 1 && count >= 3) |
1539 | || (uniq == 2 && count >= 5) |
1540 | || (uniq == 3 && count >= 6))); |
1541 | } |
1542 | |
1543 | /* Return true if cluster starting at START and ending at END (inclusive) |
1544 | is profitable transformation. */ |
1545 | |
1546 | bool |
1547 | bit_test_cluster::is_beneficial (const vec<cluster *> &clusters, |
1548 | unsigned start, unsigned end) |
1549 | { |
1550 | /* Single case bail out. */ |
1551 | if (start == end) |
1552 | return false; |
1553 | |
1554 | auto_bitmap dest_bbs; |
1555 | |
1556 | for (unsigned i = start; i <= end; i++) |
1557 | { |
1558 | simple_cluster *sc = static_cast<simple_cluster *> (clusters[i]); |
1559 | bitmap_set_bit (dest_bbs, sc->m_case_bb->index); |
1560 | } |
1561 | |
1562 | unsigned uniq = bitmap_count_bits (dest_bbs); |
1563 | unsigned count = end - start + 1; |
1564 | return is_beneficial (count, uniq); |
1565 | } |
1566 | |
1567 | /* Comparison function for qsort to order bit tests by decreasing |
1568 | probability of execution. */ |
1569 | |
1570 | int |
1571 | case_bit_test::cmp (const void *p1, const void *p2) |
1572 | { |
1573 | const case_bit_test *const d1 = (const case_bit_test *) p1; |
1574 | const case_bit_test *const d2 = (const case_bit_test *) p2; |
1575 | |
1576 | if (d2->bits != d1->bits) |
1577 | return d2->bits - d1->bits; |
1578 | |
1579 | /* Stabilize the sort. */ |
1580 | return (LABEL_DECL_UID (CASE_LABEL (d2->label)) |
1581 | - LABEL_DECL_UID (CASE_LABEL (d1->label))); |
1582 | } |
1583 | |
1584 | /* Expand a switch statement by a short sequence of bit-wise |
1585 | comparisons. "switch(x)" is effectively converted into |
1586 | "if ((1 << (x-MINVAL)) & CST)" where CST and MINVAL are |
1587 | integer constants. |
1588 | |
1589 | INDEX_EXPR is the value being switched on. |
1590 | |
1591 | MINVAL is the lowest case value of in the case nodes, |
1592 | and RANGE is highest value minus MINVAL. MINVAL and RANGE |
1593 | are not guaranteed to be of the same type as INDEX_EXPR |
1594 | (the gimplifier doesn't change the type of case label values, |
1595 | and MINVAL and RANGE are derived from those values). |
1596 | MAXVAL is MINVAL + RANGE. |
1597 | |
1598 | There *MUST* be max_case_bit_tests or less unique case |
1599 | node targets. */ |
1600 | |
1601 | void |
1602 | bit_test_cluster::emit (tree index_expr, tree index_type, |
1603 | tree, basic_block default_bb, location_t loc) |
1604 | { |
1605 | case_bit_test test[m_max_case_bit_tests] = { {} }; |
1606 | unsigned int i, j, k; |
1607 | unsigned int count; |
1608 | |
1609 | tree unsigned_index_type = range_check_type (index_type); |
1610 | |
1611 | gimple_stmt_iterator gsi; |
1612 | gassign *shift_stmt; |
1613 | |
1614 | tree idx, tmp, csui; |
1615 | tree word_type_node = lang_hooks.types.type_for_mode (word_mode, 1); |
1616 | tree word_mode_zero = fold_convert (word_type_node, integer_zero_node); |
1617 | tree word_mode_one = fold_convert (word_type_node, integer_one_node); |
1618 | int prec = TYPE_PRECISION (word_type_node); |
1619 | wide_int wone = wi::one (precision: prec); |
1620 | |
1621 | tree minval = get_low (); |
1622 | tree maxval = get_high (); |
1623 | |
1624 | /* Go through all case labels, and collect the case labels, profile |
1625 | counts, and other information we need to build the branch tests. */ |
1626 | count = 0; |
1627 | for (i = 0; i < m_cases.length (); i++) |
1628 | { |
1629 | unsigned int lo, hi; |
1630 | simple_cluster *n = static_cast<simple_cluster *> (m_cases[i]); |
1631 | for (k = 0; k < count; k++) |
1632 | if (n->m_case_bb == test[k].target_bb) |
1633 | break; |
1634 | |
1635 | if (k == count) |
1636 | { |
1637 | gcc_checking_assert (count < m_max_case_bit_tests); |
1638 | test[k].mask = wi::zero (precision: prec); |
1639 | test[k].target_bb = n->m_case_bb; |
1640 | test[k].label = n->m_case_label_expr; |
1641 | test[k].bits = 0; |
1642 | test[k].prob = profile_probability::never (); |
1643 | count++; |
1644 | } |
1645 | |
1646 | test[k].bits += n->get_range (low: n->get_low (), high: n->get_high ()); |
1647 | test[k].prob += n->m_prob; |
1648 | |
1649 | lo = tree_to_uhwi (int_const_binop (MINUS_EXPR, n->get_low (), minval)); |
1650 | if (n->get_high () == NULL_TREE) |
1651 | hi = lo; |
1652 | else |
1653 | hi = tree_to_uhwi (int_const_binop (MINUS_EXPR, n->get_high (), |
1654 | minval)); |
1655 | |
1656 | for (j = lo; j <= hi; j++) |
1657 | test[k].mask |= wi::lshift (x: wone, y: j); |
1658 | } |
1659 | |
1660 | qsort (test, count, sizeof (*test), case_bit_test::cmp); |
1661 | |
1662 | /* If every possible relative value of the index expression is a valid shift |
1663 | amount, then we can merge the entry test in the bit test. */ |
1664 | bool entry_test_needed; |
1665 | value_range r; |
1666 | if (TREE_CODE (index_expr) == SSA_NAME |
1667 | && get_range_query (cfun)->range_of_expr (r, expr: index_expr) |
1668 | && !r.undefined_p () |
1669 | && !r.varying_p () |
1670 | && wi::leu_p (x: r.upper_bound () - r.lower_bound (), y: prec - 1)) |
1671 | { |
1672 | wide_int min = r.lower_bound (); |
1673 | wide_int max = r.upper_bound (); |
1674 | tree index_type = TREE_TYPE (index_expr); |
1675 | minval = fold_convert (index_type, minval); |
1676 | wide_int iminval = wi::to_wide (t: minval); |
1677 | if (wi::lt_p (x: min, y: iminval, TYPE_SIGN (index_type))) |
1678 | { |
1679 | minval = wide_int_to_tree (type: index_type, cst: min); |
1680 | for (i = 0; i < count; i++) |
1681 | test[i].mask = wi::lshift (x: test[i].mask, y: iminval - min); |
1682 | } |
1683 | else if (wi::gt_p (x: min, y: iminval, TYPE_SIGN (index_type))) |
1684 | { |
1685 | minval = wide_int_to_tree (type: index_type, cst: min); |
1686 | for (i = 0; i < count; i++) |
1687 | test[i].mask = wi::lrshift (x: test[i].mask, y: min - iminval); |
1688 | } |
1689 | maxval = wide_int_to_tree (type: index_type, cst: max); |
1690 | entry_test_needed = false; |
1691 | } |
1692 | else |
1693 | entry_test_needed = true; |
1694 | |
1695 | /* If all values are in the 0 .. BITS_PER_WORD-1 range, we can get rid of |
1696 | the minval subtractions, but it might make the mask constants more |
1697 | expensive. So, compare the costs. */ |
1698 | if (compare_tree_int (minval, 0) > 0 && compare_tree_int (maxval, prec) < 0) |
1699 | { |
1700 | int cost_diff; |
1701 | HOST_WIDE_INT m = tree_to_uhwi (minval); |
1702 | rtx reg = gen_raw_REG (word_mode, 10000); |
1703 | bool speed_p = optimize_insn_for_speed_p (); |
1704 | cost_diff = set_src_cost (gen_rtx_PLUS (word_mode, reg, |
1705 | GEN_INT (-m)), |
1706 | mode: word_mode, speed_p); |
1707 | for (i = 0; i < count; i++) |
1708 | { |
1709 | rtx r = immed_wide_int_const (test[i].mask, word_mode); |
1710 | cost_diff += set_src_cost (gen_rtx_AND (word_mode, reg, r), |
1711 | mode: word_mode, speed_p); |
1712 | r = immed_wide_int_const (wi::lshift (x: test[i].mask, y: m), word_mode); |
1713 | cost_diff -= set_src_cost (gen_rtx_AND (word_mode, reg, r), |
1714 | mode: word_mode, speed_p); |
1715 | } |
1716 | if (cost_diff > 0) |
1717 | { |
1718 | for (i = 0; i < count; i++) |
1719 | test[i].mask = wi::lshift (x: test[i].mask, y: m); |
1720 | minval = build_zero_cst (TREE_TYPE (minval)); |
1721 | } |
1722 | } |
1723 | |
1724 | /* Now build the test-and-branch code. */ |
1725 | |
1726 | gsi = gsi_last_bb (bb: m_case_bb); |
1727 | |
1728 | /* idx = (unsigned)x - minval. */ |
1729 | idx = fold_convert_loc (loc, unsigned_index_type, index_expr); |
1730 | idx = fold_build2_loc (loc, MINUS_EXPR, unsigned_index_type, idx, |
1731 | fold_convert_loc (loc, unsigned_index_type, minval)); |
1732 | idx = force_gimple_operand_gsi (&gsi, idx, |
1733 | /*simple=*/true, NULL_TREE, |
1734 | /*before=*/true, GSI_SAME_STMT); |
1735 | |
1736 | profile_probability subtree_prob = m_subtree_prob; |
1737 | profile_probability default_prob = m_default_prob; |
1738 | if (!default_prob.initialized_p ()) |
1739 | default_prob = m_subtree_prob.invert (); |
1740 | |
1741 | if (m_handles_entire_switch && entry_test_needed) |
1742 | { |
1743 | tree range = int_const_binop (MINUS_EXPR, maxval, minval); |
1744 | /* if (idx > range) goto default */ |
1745 | range |
1746 | = force_gimple_operand_gsi (&gsi, |
1747 | fold_convert (unsigned_index_type, range), |
1748 | /*simple=*/true, NULL_TREE, |
1749 | /*before=*/true, GSI_SAME_STMT); |
1750 | tmp = fold_build2 (GT_EXPR, boolean_type_node, idx, range); |
1751 | default_prob = default_prob / 2; |
1752 | basic_block new_bb |
1753 | = hoist_edge_and_branch_if_true (gsip: &gsi, cond: tmp, case_bb: default_bb, |
1754 | prob: default_prob, loc); |
1755 | gsi = gsi_last_bb (bb: new_bb); |
1756 | } |
1757 | |
1758 | tmp = fold_build2_loc (loc, LSHIFT_EXPR, word_type_node, word_mode_one, |
1759 | fold_convert_loc (loc, word_type_node, idx)); |
1760 | |
1761 | /* csui = (1 << (word_mode) idx) */ |
1762 | if (count > 1) |
1763 | { |
1764 | csui = make_ssa_name (var: word_type_node); |
1765 | tmp = force_gimple_operand_gsi (&gsi, tmp, |
1766 | /*simple=*/false, NULL_TREE, |
1767 | /*before=*/true, GSI_SAME_STMT); |
1768 | shift_stmt = gimple_build_assign (csui, tmp); |
1769 | gsi_insert_before (&gsi, shift_stmt, GSI_SAME_STMT); |
1770 | update_stmt (s: shift_stmt); |
1771 | } |
1772 | else |
1773 | csui = tmp; |
1774 | |
1775 | /* for each unique set of cases: |
1776 | if (const & csui) goto target */ |
1777 | for (k = 0; k < count; k++) |
1778 | { |
1779 | profile_probability prob = test[k].prob / (subtree_prob + default_prob); |
1780 | subtree_prob -= test[k].prob; |
1781 | tmp = wide_int_to_tree (type: word_type_node, cst: test[k].mask); |
1782 | tmp = fold_build2_loc (loc, BIT_AND_EXPR, word_type_node, csui, tmp); |
1783 | tmp = fold_build2_loc (loc, NE_EXPR, boolean_type_node, |
1784 | tmp, word_mode_zero); |
1785 | tmp = force_gimple_operand_gsi (&gsi, tmp, |
1786 | /*simple=*/true, NULL_TREE, |
1787 | /*before=*/true, GSI_SAME_STMT); |
1788 | basic_block new_bb |
1789 | = hoist_edge_and_branch_if_true (gsip: &gsi, cond: tmp, case_bb: test[k].target_bb, |
1790 | prob, loc); |
1791 | gsi = gsi_last_bb (bb: new_bb); |
1792 | } |
1793 | |
1794 | /* We should have removed all edges now. */ |
1795 | gcc_assert (EDGE_COUNT (gsi_bb (gsi)->succs) == 0); |
1796 | |
1797 | /* If nothing matched, go to the default label. */ |
1798 | edge e = make_edge (gsi_bb (i: gsi), default_bb, EDGE_FALLTHRU); |
1799 | e->probability = profile_probability::always (); |
1800 | } |
1801 | |
1802 | /* Split the basic block at the statement pointed to by GSIP, and insert |
1803 | a branch to the target basic block of E_TRUE conditional on tree |
1804 | expression COND. |
1805 | |
1806 | It is assumed that there is already an edge from the to-be-split |
1807 | basic block to E_TRUE->dest block. This edge is removed, and the |
1808 | profile information on the edge is re-used for the new conditional |
1809 | jump. |
1810 | |
1811 | The CFG is updated. The dominator tree will not be valid after |
1812 | this transformation, but the immediate dominators are updated if |
1813 | UPDATE_DOMINATORS is true. |
1814 | |
1815 | Returns the newly created basic block. */ |
1816 | |
1817 | basic_block |
1818 | bit_test_cluster::hoist_edge_and_branch_if_true (gimple_stmt_iterator *gsip, |
1819 | tree cond, basic_block case_bb, |
1820 | profile_probability prob, |
1821 | location_t loc) |
1822 | { |
1823 | tree tmp; |
1824 | gcond *cond_stmt; |
1825 | edge e_false; |
1826 | basic_block new_bb, split_bb = gsi_bb (i: *gsip); |
1827 | |
1828 | edge e_true = make_edge (split_bb, case_bb, EDGE_TRUE_VALUE); |
1829 | e_true->probability = prob; |
1830 | gcc_assert (e_true->src == split_bb); |
1831 | |
1832 | tmp = force_gimple_operand_gsi (gsip, cond, /*simple=*/true, NULL, |
1833 | /*before=*/true, GSI_SAME_STMT); |
1834 | cond_stmt = gimple_build_cond_from_tree (tmp, NULL_TREE, NULL_TREE); |
1835 | gimple_set_location (g: cond_stmt, location: loc); |
1836 | gsi_insert_before (gsip, cond_stmt, GSI_SAME_STMT); |
1837 | |
1838 | e_false = split_block (split_bb, cond_stmt); |
1839 | new_bb = e_false->dest; |
1840 | redirect_edge_pred (e_true, split_bb); |
1841 | |
1842 | e_false->flags &= ~EDGE_FALLTHRU; |
1843 | e_false->flags |= EDGE_FALSE_VALUE; |
1844 | e_false->probability = e_true->probability.invert (); |
1845 | new_bb->count = e_false->count (); |
1846 | |
1847 | return new_bb; |
1848 | } |
1849 | |
1850 | /* Compute the number of case labels that correspond to each outgoing edge of |
1851 | switch statement. Record this information in the aux field of the edge. */ |
1852 | |
1853 | void |
1854 | switch_decision_tree::compute_cases_per_edge () |
1855 | { |
1856 | reset_out_edges_aux (swtch: m_switch); |
1857 | int ncases = gimple_switch_num_labels (gs: m_switch); |
1858 | for (int i = ncases - 1; i >= 1; --i) |
1859 | { |
1860 | edge case_edge = gimple_switch_edge (cfun, m_switch, i); |
1861 | case_edge->aux = (void *) ((intptr_t) (case_edge->aux) + 1); |
1862 | } |
1863 | } |
1864 | |
1865 | /* Analyze switch statement and return true when the statement is expanded |
1866 | as decision tree. */ |
1867 | |
1868 | bool |
1869 | switch_decision_tree::analyze_switch_statement () |
1870 | { |
1871 | unsigned l = gimple_switch_num_labels (gs: m_switch); |
1872 | basic_block bb = gimple_bb (g: m_switch); |
1873 | auto_vec<cluster *> clusters; |
1874 | clusters.create (nelems: l - 1); |
1875 | |
1876 | basic_block default_bb = gimple_switch_default_bb (cfun, m_switch); |
1877 | m_case_bbs.reserve (nelems: l); |
1878 | m_case_bbs.quick_push (obj: default_bb); |
1879 | |
1880 | compute_cases_per_edge (); |
1881 | |
1882 | for (unsigned i = 1; i < l; i++) |
1883 | { |
1884 | tree elt = gimple_switch_label (gs: m_switch, index: i); |
1885 | tree lab = CASE_LABEL (elt); |
1886 | basic_block case_bb = label_to_block (cfun, lab); |
1887 | edge case_edge = find_edge (bb, case_bb); |
1888 | tree low = CASE_LOW (elt); |
1889 | tree high = CASE_HIGH (elt); |
1890 | |
1891 | profile_probability p |
1892 | = case_edge->probability / ((intptr_t) (case_edge->aux)); |
1893 | clusters.quick_push (obj: new simple_cluster (low, high, elt, case_edge->dest, |
1894 | p)); |
1895 | m_case_bbs.quick_push (obj: case_edge->dest); |
1896 | } |
1897 | |
1898 | reset_out_edges_aux (swtch: m_switch); |
1899 | |
1900 | /* Find bit-test clusters. */ |
1901 | vec<cluster *> output = bit_test_cluster::find_bit_tests (clusters); |
1902 | |
1903 | /* Find jump table clusters. */ |
1904 | vec<cluster *> output2; |
1905 | auto_vec<cluster *> tmp; |
1906 | output2.create (nelems: 1); |
1907 | tmp.create (nelems: 1); |
1908 | |
1909 | for (unsigned i = 0; i < output.length (); i++) |
1910 | { |
1911 | cluster *c = output[i]; |
1912 | if (c->get_type () != SIMPLE_CASE) |
1913 | { |
1914 | if (!tmp.is_empty ()) |
1915 | { |
1916 | vec<cluster *> n = jump_table_cluster::find_jump_tables (clusters&: tmp); |
1917 | output2.safe_splice (src: n); |
1918 | n.release (); |
1919 | tmp.truncate (size: 0); |
1920 | } |
1921 | output2.safe_push (obj: c); |
1922 | } |
1923 | else |
1924 | tmp.safe_push (obj: c); |
1925 | } |
1926 | |
1927 | /* We still can have a temporary vector to test. */ |
1928 | if (!tmp.is_empty ()) |
1929 | { |
1930 | vec<cluster *> n = jump_table_cluster::find_jump_tables (clusters&: tmp); |
1931 | output2.safe_splice (src: n); |
1932 | n.release (); |
1933 | } |
1934 | |
1935 | if (dump_file) |
1936 | { |
1937 | fprintf (stream: dump_file, format: ";; GIMPLE switch case clusters: " ); |
1938 | for (unsigned i = 0; i < output2.length (); i++) |
1939 | output2[i]->dump (f: dump_file, details: dump_flags & TDF_DETAILS); |
1940 | fprintf (stream: dump_file, format: "\n" ); |
1941 | } |
1942 | |
1943 | output.release (); |
1944 | |
1945 | bool expanded = try_switch_expansion (clusters&: output2); |
1946 | release_clusters (clusters&: output2); |
1947 | return expanded; |
1948 | } |
1949 | |
1950 | /* Attempt to expand CLUSTERS as a decision tree. Return true when |
1951 | expanded. */ |
1952 | |
1953 | bool |
1954 | switch_decision_tree::try_switch_expansion (vec<cluster *> &clusters) |
1955 | { |
1956 | tree index_expr = gimple_switch_index (gs: m_switch); |
1957 | tree index_type = TREE_TYPE (index_expr); |
1958 | basic_block bb = gimple_bb (g: m_switch); |
1959 | |
1960 | if (gimple_switch_num_labels (gs: m_switch) == 1 |
1961 | || range_check_type (index_type) == NULL_TREE) |
1962 | return false; |
1963 | |
1964 | /* Find the default case target label. */ |
1965 | edge default_edge = gimple_switch_default_edge (cfun, m_switch); |
1966 | m_default_bb = default_edge->dest; |
1967 | |
1968 | /* Do the insertion of a case label into m_case_list. The labels are |
1969 | fed to us in descending order from the sorted vector of case labels used |
1970 | in the tree part of the middle end. So the list we construct is |
1971 | sorted in ascending order. */ |
1972 | |
1973 | for (int i = clusters.length () - 1; i >= 0; i--) |
1974 | { |
1975 | case_tree_node *r = m_case_list; |
1976 | m_case_list = m_case_node_pool.allocate (); |
1977 | m_case_list->m_right = r; |
1978 | m_case_list->m_c = clusters[i]; |
1979 | } |
1980 | |
1981 | record_phi_operand_mapping (); |
1982 | |
1983 | /* Split basic block that contains the gswitch statement. */ |
1984 | gimple_stmt_iterator gsi = gsi_last_bb (bb); |
1985 | edge e; |
1986 | if (gsi_end_p (i: gsi)) |
1987 | e = split_block_after_labels (bb); |
1988 | else |
1989 | { |
1990 | gsi_prev (i: &gsi); |
1991 | e = split_block (bb, gsi_stmt (i: gsi)); |
1992 | } |
1993 | bb = split_edge (e); |
1994 | |
1995 | /* Create new basic blocks for non-case clusters where specific expansion |
1996 | needs to happen. */ |
1997 | for (unsigned i = 0; i < clusters.length (); i++) |
1998 | if (clusters[i]->get_type () != SIMPLE_CASE) |
1999 | { |
2000 | clusters[i]->m_case_bb = create_empty_bb (bb); |
2001 | clusters[i]->m_case_bb->count = bb->count; |
2002 | clusters[i]->m_case_bb->loop_father = bb->loop_father; |
2003 | } |
2004 | |
2005 | /* Do not do an extra work for a single cluster. */ |
2006 | if (clusters.length () == 1 |
2007 | && clusters[0]->get_type () != SIMPLE_CASE) |
2008 | { |
2009 | cluster *c = clusters[0]; |
2010 | c->emit (index_expr, index_type, |
2011 | gimple_switch_default_label (gs: m_switch), m_default_bb, |
2012 | gimple_location (g: m_switch)); |
2013 | redirect_edge_succ (single_succ_edge (bb), c->m_case_bb); |
2014 | } |
2015 | else |
2016 | { |
2017 | emit (bb, index_expr, default_prob: default_edge->probability, index_type); |
2018 | |
2019 | /* Emit cluster-specific switch handling. */ |
2020 | for (unsigned i = 0; i < clusters.length (); i++) |
2021 | if (clusters[i]->get_type () != SIMPLE_CASE) |
2022 | { |
2023 | edge e = single_pred_edge (bb: clusters[i]->m_case_bb); |
2024 | e->dest->count = e->src->count.apply_probability (prob: e->probability); |
2025 | clusters[i]->emit (index_expr, index_type, |
2026 | gimple_switch_default_label (gs: m_switch), |
2027 | m_default_bb, gimple_location (g: m_switch)); |
2028 | } |
2029 | } |
2030 | |
2031 | fix_phi_operands_for_edges (); |
2032 | |
2033 | return true; |
2034 | } |
2035 | |
2036 | /* Before switch transformation, record all SSA_NAMEs defined in switch BB |
2037 | and used in a label basic block. */ |
2038 | |
2039 | void |
2040 | switch_decision_tree::record_phi_operand_mapping () |
2041 | { |
2042 | basic_block switch_bb = gimple_bb (g: m_switch); |
2043 | /* Record all PHI nodes that have to be fixed after conversion. */ |
2044 | for (unsigned i = 0; i < m_case_bbs.length (); i++) |
2045 | { |
2046 | gphi_iterator gsi; |
2047 | basic_block bb = m_case_bbs[i]; |
2048 | for (gsi = gsi_start_phis (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
2049 | { |
2050 | gphi *phi = gsi.phi (); |
2051 | |
2052 | for (unsigned i = 0; i < gimple_phi_num_args (gs: phi); i++) |
2053 | { |
2054 | basic_block phi_src_bb = gimple_phi_arg_edge (phi, i)->src; |
2055 | if (phi_src_bb == switch_bb) |
2056 | { |
2057 | tree def = gimple_phi_arg_def (gs: phi, index: i); |
2058 | tree result = gimple_phi_result (gs: phi); |
2059 | m_phi_mapping.put (k: result, v: def); |
2060 | break; |
2061 | } |
2062 | } |
2063 | } |
2064 | } |
2065 | } |
2066 | |
2067 | /* Append new operands to PHI statements that were introduced due to |
2068 | addition of new edges to case labels. */ |
2069 | |
2070 | void |
2071 | switch_decision_tree::fix_phi_operands_for_edges () |
2072 | { |
2073 | gphi_iterator gsi; |
2074 | |
2075 | for (unsigned i = 0; i < m_case_bbs.length (); i++) |
2076 | { |
2077 | basic_block bb = m_case_bbs[i]; |
2078 | for (gsi = gsi_start_phis (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
2079 | { |
2080 | gphi *phi = gsi.phi (); |
2081 | for (unsigned j = 0; j < gimple_phi_num_args (gs: phi); j++) |
2082 | { |
2083 | tree def = gimple_phi_arg_def (gs: phi, index: j); |
2084 | if (def == NULL_TREE) |
2085 | { |
2086 | edge e = gimple_phi_arg_edge (phi, i: j); |
2087 | tree *definition |
2088 | = m_phi_mapping.get (k: gimple_phi_result (gs: phi)); |
2089 | gcc_assert (definition); |
2090 | add_phi_arg (phi, *definition, e, UNKNOWN_LOCATION); |
2091 | } |
2092 | } |
2093 | } |
2094 | } |
2095 | } |
2096 | |
2097 | /* Generate a decision tree, switching on INDEX_EXPR and jumping to |
2098 | one of the labels in CASE_LIST or to the DEFAULT_LABEL. |
2099 | |
2100 | We generate a binary decision tree to select the appropriate target |
2101 | code. */ |
2102 | |
2103 | void |
2104 | switch_decision_tree::emit (basic_block bb, tree index_expr, |
2105 | profile_probability default_prob, tree index_type) |
2106 | { |
2107 | balance_case_nodes (head: &m_case_list, NULL); |
2108 | |
2109 | if (dump_file) |
2110 | dump_function_to_file (current_function_decl, dump_file, dump_flags); |
2111 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2112 | { |
2113 | int indent_step = ceil_log2 (TYPE_PRECISION (index_type)) + 2; |
2114 | fprintf (stream: dump_file, format: ";; Expanding GIMPLE switch as decision tree:\n" ); |
2115 | gcc_assert (m_case_list != NULL); |
2116 | dump_case_nodes (f: dump_file, root: m_case_list, indent_step, indent_level: 0); |
2117 | } |
2118 | |
2119 | bb = emit_case_nodes (bb, index: index_expr, node: m_case_list, default_prob, index_type, |
2120 | gimple_location (g: m_switch)); |
2121 | |
2122 | if (bb) |
2123 | emit_jump (bb, case_bb: m_default_bb); |
2124 | |
2125 | /* Remove all edges and do just an edge that will reach default_bb. */ |
2126 | bb = gimple_bb (g: m_switch); |
2127 | gimple_stmt_iterator gsi = gsi_last_bb (bb); |
2128 | gsi_remove (&gsi, true); |
2129 | |
2130 | delete_basic_block (bb); |
2131 | } |
2132 | |
2133 | /* Take an ordered list of case nodes |
2134 | and transform them into a near optimal binary tree, |
2135 | on the assumption that any target code selection value is as |
2136 | likely as any other. |
2137 | |
2138 | The transformation is performed by splitting the ordered |
2139 | list into two equal sections plus a pivot. The parts are |
2140 | then attached to the pivot as left and right branches. Each |
2141 | branch is then transformed recursively. */ |
2142 | |
2143 | void |
2144 | switch_decision_tree::balance_case_nodes (case_tree_node **head, |
2145 | case_tree_node *parent) |
2146 | { |
2147 | case_tree_node *np; |
2148 | |
2149 | np = *head; |
2150 | if (np) |
2151 | { |
2152 | int i = 0; |
2153 | case_tree_node **npp; |
2154 | case_tree_node *left; |
2155 | profile_probability prob = profile_probability::never (); |
2156 | |
2157 | /* Count the number of entries on branch. */ |
2158 | |
2159 | while (np) |
2160 | { |
2161 | i++; |
2162 | prob += np->m_c->m_prob; |
2163 | np = np->m_right; |
2164 | } |
2165 | |
2166 | if (i > 2) |
2167 | { |
2168 | /* Split this list if it is long enough for that to help. */ |
2169 | npp = head; |
2170 | left = *npp; |
2171 | profile_probability pivot_prob = prob / 2; |
2172 | |
2173 | /* Find the place in the list that bisects the list's total cost |
2174 | by probability. */ |
2175 | while (1) |
2176 | { |
2177 | /* Skip nodes while their probability does not reach |
2178 | that amount. */ |
2179 | prob -= (*npp)->m_c->m_prob; |
2180 | if ((prob.initialized_p () && prob < pivot_prob) |
2181 | || ! (*npp)->m_right) |
2182 | break; |
2183 | npp = &(*npp)->m_right; |
2184 | } |
2185 | |
2186 | np = *npp; |
2187 | *npp = 0; |
2188 | *head = np; |
2189 | np->m_parent = parent; |
2190 | np->m_left = left == np ? NULL : left; |
2191 | |
2192 | /* Optimize each of the two split parts. */ |
2193 | balance_case_nodes (head: &np->m_left, parent: np); |
2194 | balance_case_nodes (head: &np->m_right, parent: np); |
2195 | np->m_c->m_subtree_prob = np->m_c->m_prob; |
2196 | if (np->m_left) |
2197 | np->m_c->m_subtree_prob += np->m_left->m_c->m_subtree_prob; |
2198 | if (np->m_right) |
2199 | np->m_c->m_subtree_prob += np->m_right->m_c->m_subtree_prob; |
2200 | } |
2201 | else |
2202 | { |
2203 | /* Else leave this branch as one level, |
2204 | but fill in `parent' fields. */ |
2205 | np = *head; |
2206 | np->m_parent = parent; |
2207 | np->m_c->m_subtree_prob = np->m_c->m_prob; |
2208 | for (; np->m_right; np = np->m_right) |
2209 | { |
2210 | np->m_right->m_parent = np; |
2211 | (*head)->m_c->m_subtree_prob += np->m_right->m_c->m_subtree_prob; |
2212 | } |
2213 | } |
2214 | } |
2215 | } |
2216 | |
2217 | /* Dump ROOT, a list or tree of case nodes, to file. */ |
2218 | |
2219 | void |
2220 | switch_decision_tree::dump_case_nodes (FILE *f, case_tree_node *root, |
2221 | int indent_step, int indent_level) |
2222 | { |
2223 | if (root == 0) |
2224 | return; |
2225 | indent_level++; |
2226 | |
2227 | dump_case_nodes (f, root: root->m_left, indent_step, indent_level); |
2228 | |
2229 | fputs (s: ";; " , stream: f); |
2230 | fprintf (stream: f, format: "%*s" , indent_step * indent_level, "" ); |
2231 | root->m_c->dump (f); |
2232 | root->m_c->m_prob.dump (f); |
2233 | fputs (s: " subtree: " , stream: f); |
2234 | root->m_c->m_subtree_prob.dump (f); |
2235 | fputs (s: ")\n" , stream: f); |
2236 | |
2237 | dump_case_nodes (f, root: root->m_right, indent_step, indent_level); |
2238 | } |
2239 | |
2240 | |
2241 | /* Add an unconditional jump to CASE_BB that happens in basic block BB. */ |
2242 | |
2243 | void |
2244 | switch_decision_tree::emit_jump (basic_block bb, basic_block case_bb) |
2245 | { |
2246 | edge e = single_succ_edge (bb); |
2247 | redirect_edge_succ (e, case_bb); |
2248 | } |
2249 | |
2250 | /* Generate code to compare OP0 with OP1 so that the condition codes are |
2251 | set and to jump to LABEL_BB if the condition is true. |
2252 | COMPARISON is the GIMPLE comparison (EQ, NE, GT, etc.). |
2253 | PROB is the probability of jumping to LABEL_BB. */ |
2254 | |
2255 | basic_block |
2256 | switch_decision_tree::emit_cmp_and_jump_insns (basic_block bb, tree op0, |
2257 | tree op1, tree_code comparison, |
2258 | basic_block label_bb, |
2259 | profile_probability prob, |
2260 | location_t loc) |
2261 | { |
2262 | // TODO: it's once called with lhs != index. |
2263 | op1 = fold_convert (TREE_TYPE (op0), op1); |
2264 | |
2265 | gcond *cond = gimple_build_cond (comparison, op0, op1, NULL_TREE, NULL_TREE); |
2266 | gimple_set_location (g: cond, location: loc); |
2267 | gimple_stmt_iterator gsi = gsi_last_bb (bb); |
2268 | gsi_insert_after (&gsi, cond, GSI_NEW_STMT); |
2269 | |
2270 | gcc_assert (single_succ_p (bb)); |
2271 | |
2272 | /* Make a new basic block where false branch will take place. */ |
2273 | edge false_edge = split_block (bb, cond); |
2274 | false_edge->flags = EDGE_FALSE_VALUE; |
2275 | false_edge->probability = prob.invert (); |
2276 | false_edge->dest->count = bb->count.apply_probability (prob: prob.invert ()); |
2277 | |
2278 | edge true_edge = make_edge (bb, label_bb, EDGE_TRUE_VALUE); |
2279 | true_edge->probability = prob; |
2280 | |
2281 | return false_edge->dest; |
2282 | } |
2283 | |
2284 | /* Generate code to jump to LABEL if OP0 and OP1 are equal. |
2285 | PROB is the probability of jumping to LABEL_BB. |
2286 | BB is a basic block where the new condition will be placed. */ |
2287 | |
2288 | basic_block |
2289 | switch_decision_tree::do_jump_if_equal (basic_block bb, tree op0, tree op1, |
2290 | basic_block label_bb, |
2291 | profile_probability prob, |
2292 | location_t loc) |
2293 | { |
2294 | op1 = fold_convert (TREE_TYPE (op0), op1); |
2295 | |
2296 | gcond *cond = gimple_build_cond (EQ_EXPR, op0, op1, NULL_TREE, NULL_TREE); |
2297 | gimple_set_location (g: cond, location: loc); |
2298 | gimple_stmt_iterator gsi = gsi_last_bb (bb); |
2299 | gsi_insert_before (&gsi, cond, GSI_SAME_STMT); |
2300 | |
2301 | gcc_assert (single_succ_p (bb)); |
2302 | |
2303 | /* Make a new basic block where false branch will take place. */ |
2304 | edge false_edge = split_block (bb, cond); |
2305 | false_edge->flags = EDGE_FALSE_VALUE; |
2306 | false_edge->probability = prob.invert (); |
2307 | false_edge->dest->count = bb->count.apply_probability (prob: prob.invert ()); |
2308 | |
2309 | edge true_edge = make_edge (bb, label_bb, EDGE_TRUE_VALUE); |
2310 | true_edge->probability = prob; |
2311 | |
2312 | return false_edge->dest; |
2313 | } |
2314 | |
2315 | /* Emit step-by-step code to select a case for the value of INDEX. |
2316 | The thus generated decision tree follows the form of the |
2317 | case-node binary tree NODE, whose nodes represent test conditions. |
2318 | DEFAULT_PROB is probability of cases leading to default BB. |
2319 | INDEX_TYPE is the type of the index of the switch. */ |
2320 | |
2321 | basic_block |
2322 | switch_decision_tree::emit_case_nodes (basic_block bb, tree index, |
2323 | case_tree_node *node, |
2324 | profile_probability default_prob, |
2325 | tree index_type, location_t loc) |
2326 | { |
2327 | profile_probability p; |
2328 | |
2329 | /* If node is null, we are done. */ |
2330 | if (node == NULL) |
2331 | return bb; |
2332 | |
2333 | /* Single value case. */ |
2334 | if (node->m_c->is_single_value_p ()) |
2335 | { |
2336 | /* Node is single valued. First see if the index expression matches |
2337 | this node and then check our children, if any. */ |
2338 | p = node->m_c->m_prob / (node->m_c->m_subtree_prob + default_prob); |
2339 | bb = do_jump_if_equal (bb, op0: index, op1: node->m_c->get_low (), |
2340 | label_bb: node->m_c->m_case_bb, prob: p, loc); |
2341 | /* Since this case is taken at this point, reduce its weight from |
2342 | subtree_weight. */ |
2343 | node->m_c->m_subtree_prob -= node->m_c->m_prob; |
2344 | |
2345 | if (node->m_left != NULL && node->m_right != NULL) |
2346 | { |
2347 | /* 1) the node has both children |
2348 | |
2349 | If both children are single-valued cases with no |
2350 | children, finish up all the work. This way, we can save |
2351 | one ordered comparison. */ |
2352 | |
2353 | if (!node->m_left->has_child () |
2354 | && node->m_left->m_c->is_single_value_p () |
2355 | && !node->m_right->has_child () |
2356 | && node->m_right->m_c->is_single_value_p ()) |
2357 | { |
2358 | p = (node->m_right->m_c->m_prob |
2359 | / (node->m_c->m_subtree_prob + default_prob)); |
2360 | bb = do_jump_if_equal (bb, op0: index, op1: node->m_right->m_c->get_low (), |
2361 | label_bb: node->m_right->m_c->m_case_bb, prob: p, loc); |
2362 | node->m_c->m_subtree_prob -= node->m_right->m_c->m_prob; |
2363 | |
2364 | p = (node->m_left->m_c->m_prob |
2365 | / (node->m_c->m_subtree_prob + default_prob)); |
2366 | bb = do_jump_if_equal (bb, op0: index, op1: node->m_left->m_c->get_low (), |
2367 | label_bb: node->m_left->m_c->m_case_bb, prob: p, loc); |
2368 | } |
2369 | else |
2370 | { |
2371 | /* Branch to a label where we will handle it later. */ |
2372 | basic_block test_bb = split_edge (single_succ_edge (bb)); |
2373 | redirect_edge_succ (single_pred_edge (bb: test_bb), |
2374 | single_succ_edge (bb)->dest); |
2375 | |
2376 | p = ((node->m_right->m_c->m_subtree_prob + default_prob / 2) |
2377 | / (node->m_c->m_subtree_prob + default_prob)); |
2378 | test_bb->count = bb->count.apply_probability (prob: p); |
2379 | bb = emit_cmp_and_jump_insns (bb, op0: index, op1: node->m_c->get_high (), |
2380 | comparison: GT_EXPR, label_bb: test_bb, prob: p, loc); |
2381 | default_prob /= 2; |
2382 | |
2383 | /* Handle the left-hand subtree. */ |
2384 | bb = emit_case_nodes (bb, index, node: node->m_left, |
2385 | default_prob, index_type, loc); |
2386 | |
2387 | /* If the left-hand subtree fell through, |
2388 | don't let it fall into the right-hand subtree. */ |
2389 | if (bb && m_default_bb) |
2390 | emit_jump (bb, case_bb: m_default_bb); |
2391 | |
2392 | bb = emit_case_nodes (bb: test_bb, index, node: node->m_right, |
2393 | default_prob, index_type, loc); |
2394 | } |
2395 | } |
2396 | else if (node->m_left == NULL && node->m_right != NULL) |
2397 | { |
2398 | /* 2) the node has only right child. */ |
2399 | |
2400 | /* Here we have a right child but no left so we issue a conditional |
2401 | branch to default and process the right child. |
2402 | |
2403 | Omit the conditional branch to default if the right child |
2404 | does not have any children and is single valued; it would |
2405 | cost too much space to save so little time. */ |
2406 | |
2407 | if (node->m_right->has_child () |
2408 | || !node->m_right->m_c->is_single_value_p ()) |
2409 | { |
2410 | p = ((default_prob / 2) |
2411 | / (node->m_c->m_subtree_prob + default_prob)); |
2412 | bb = emit_cmp_and_jump_insns (bb, op0: index, op1: node->m_c->get_low (), |
2413 | comparison: LT_EXPR, label_bb: m_default_bb, prob: p, loc); |
2414 | default_prob /= 2; |
2415 | |
2416 | bb = emit_case_nodes (bb, index, node: node->m_right, default_prob, |
2417 | index_type, loc); |
2418 | } |
2419 | else |
2420 | { |
2421 | /* We cannot process node->right normally |
2422 | since we haven't ruled out the numbers less than |
2423 | this node's value. So handle node->right explicitly. */ |
2424 | p = (node->m_right->m_c->m_subtree_prob |
2425 | / (node->m_c->m_subtree_prob + default_prob)); |
2426 | bb = do_jump_if_equal (bb, op0: index, op1: node->m_right->m_c->get_low (), |
2427 | label_bb: node->m_right->m_c->m_case_bb, prob: p, loc); |
2428 | } |
2429 | } |
2430 | else if (node->m_left != NULL && node->m_right == NULL) |
2431 | { |
2432 | /* 3) just one subtree, on the left. Similar case as previous. */ |
2433 | |
2434 | if (node->m_left->has_child () |
2435 | || !node->m_left->m_c->is_single_value_p ()) |
2436 | { |
2437 | p = ((default_prob / 2) |
2438 | / (node->m_c->m_subtree_prob + default_prob)); |
2439 | bb = emit_cmp_and_jump_insns (bb, op0: index, op1: node->m_c->get_high (), |
2440 | comparison: GT_EXPR, label_bb: m_default_bb, prob: p, loc); |
2441 | default_prob /= 2; |
2442 | |
2443 | bb = emit_case_nodes (bb, index, node: node->m_left, default_prob, |
2444 | index_type, loc); |
2445 | } |
2446 | else |
2447 | { |
2448 | /* We cannot process node->left normally |
2449 | since we haven't ruled out the numbers less than |
2450 | this node's value. So handle node->left explicitly. */ |
2451 | p = (node->m_left->m_c->m_subtree_prob |
2452 | / (node->m_c->m_subtree_prob + default_prob)); |
2453 | bb = do_jump_if_equal (bb, op0: index, op1: node->m_left->m_c->get_low (), |
2454 | label_bb: node->m_left->m_c->m_case_bb, prob: p, loc); |
2455 | } |
2456 | } |
2457 | } |
2458 | else |
2459 | { |
2460 | /* Node is a range. These cases are very similar to those for a single |
2461 | value, except that we do not start by testing whether this node |
2462 | is the one to branch to. */ |
2463 | if (node->has_child () || node->m_c->get_type () != SIMPLE_CASE) |
2464 | { |
2465 | bool is_bt = node->m_c->get_type () == BIT_TEST; |
2466 | int parts = is_bt ? 3 : 2; |
2467 | |
2468 | /* Branch to a label where we will handle it later. */ |
2469 | basic_block test_bb = split_edge (single_succ_edge (bb)); |
2470 | redirect_edge_succ (single_pred_edge (bb: test_bb), |
2471 | single_succ_edge (bb)->dest); |
2472 | |
2473 | profile_probability right_prob = profile_probability::never (); |
2474 | if (node->m_right) |
2475 | right_prob = node->m_right->m_c->m_subtree_prob; |
2476 | p = ((right_prob + default_prob / parts) |
2477 | / (node->m_c->m_subtree_prob + default_prob)); |
2478 | test_bb->count = bb->count.apply_probability (prob: p); |
2479 | |
2480 | bb = emit_cmp_and_jump_insns (bb, op0: index, op1: node->m_c->get_high (), |
2481 | comparison: GT_EXPR, label_bb: test_bb, prob: p, loc); |
2482 | |
2483 | default_prob /= parts; |
2484 | node->m_c->m_subtree_prob -= right_prob; |
2485 | if (is_bt) |
2486 | node->m_c->m_default_prob = default_prob; |
2487 | |
2488 | /* Value belongs to this node or to the left-hand subtree. */ |
2489 | p = node->m_c->m_prob / (node->m_c->m_subtree_prob + default_prob); |
2490 | bb = emit_cmp_and_jump_insns (bb, op0: index, op1: node->m_c->get_low (), |
2491 | comparison: GE_EXPR, label_bb: node->m_c->m_case_bb, prob: p, loc); |
2492 | |
2493 | /* Handle the left-hand subtree. */ |
2494 | bb = emit_case_nodes (bb, index, node: node->m_left, default_prob, |
2495 | index_type, loc); |
2496 | |
2497 | /* If the left-hand subtree fell through, |
2498 | don't let it fall into the right-hand subtree. */ |
2499 | if (bb && m_default_bb) |
2500 | emit_jump (bb, case_bb: m_default_bb); |
2501 | |
2502 | bb = emit_case_nodes (bb: test_bb, index, node: node->m_right, default_prob, |
2503 | index_type, loc); |
2504 | } |
2505 | else |
2506 | { |
2507 | /* Node has no children so we check low and high bounds to remove |
2508 | redundant tests. Only one of the bounds can exist, |
2509 | since otherwise this node is bounded--a case tested already. */ |
2510 | tree lhs, rhs; |
2511 | generate_range_test (bb, index, low: node->m_c->get_low (), |
2512 | high: node->m_c->get_high (), lhs: &lhs, rhs: &rhs); |
2513 | p = default_prob / (node->m_c->m_subtree_prob + default_prob); |
2514 | |
2515 | bb = emit_cmp_and_jump_insns (bb, op0: lhs, op1: rhs, comparison: GT_EXPR, |
2516 | label_bb: m_default_bb, prob: p, loc); |
2517 | |
2518 | emit_jump (bb, case_bb: node->m_c->m_case_bb); |
2519 | return NULL; |
2520 | } |
2521 | } |
2522 | |
2523 | return bb; |
2524 | } |
2525 | |
2526 | /* The main function of the pass scans statements for switches and invokes |
2527 | process_switch on them. */ |
2528 | |
2529 | namespace { |
2530 | |
2531 | const pass_data pass_data_convert_switch = |
2532 | { |
2533 | .type: GIMPLE_PASS, /* type */ |
2534 | .name: "switchconv" , /* name */ |
2535 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
2536 | .tv_id: TV_TREE_SWITCH_CONVERSION, /* tv_id */ |
2537 | .properties_required: ( PROP_cfg | PROP_ssa ), /* properties_required */ |
2538 | .properties_provided: 0, /* properties_provided */ |
2539 | .properties_destroyed: 0, /* properties_destroyed */ |
2540 | .todo_flags_start: 0, /* todo_flags_start */ |
2541 | TODO_update_ssa, /* todo_flags_finish */ |
2542 | }; |
2543 | |
2544 | class pass_convert_switch : public gimple_opt_pass |
2545 | { |
2546 | public: |
2547 | pass_convert_switch (gcc::context *ctxt) |
2548 | : gimple_opt_pass (pass_data_convert_switch, ctxt) |
2549 | {} |
2550 | |
2551 | /* opt_pass methods: */ |
2552 | bool gate (function *) final override |
2553 | { |
2554 | return flag_tree_switch_conversion != 0; |
2555 | } |
2556 | unsigned int execute (function *) final override; |
2557 | |
2558 | }; // class pass_convert_switch |
2559 | |
2560 | unsigned int |
2561 | pass_convert_switch::execute (function *fun) |
2562 | { |
2563 | basic_block bb; |
2564 | bool cfg_altered = false; |
2565 | |
2566 | FOR_EACH_BB_FN (bb, fun) |
2567 | { |
2568 | if (gswitch *stmt = safe_dyn_cast <gswitch *> (p: *gsi_last_bb (bb))) |
2569 | { |
2570 | if (dump_file) |
2571 | { |
2572 | expanded_location loc = expand_location (gimple_location (g: stmt)); |
2573 | |
2574 | fprintf (stream: dump_file, format: "beginning to process the following " |
2575 | "SWITCH statement (%s:%d) : ------- \n" , |
2576 | loc.file, loc.line); |
2577 | print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); |
2578 | putc (c: '\n', stream: dump_file); |
2579 | } |
2580 | |
2581 | switch_conversion sconv; |
2582 | sconv.expand (swtch: stmt); |
2583 | cfg_altered |= sconv.m_cfg_altered; |
2584 | if (!sconv.m_reason) |
2585 | { |
2586 | if (dump_file) |
2587 | { |
2588 | fputs (s: "Switch converted\n" , stream: dump_file); |
2589 | fputs (s: "--------------------------------\n" , stream: dump_file); |
2590 | } |
2591 | |
2592 | /* Make no effort to update the post-dominator tree. |
2593 | It is actually not that hard for the transformations |
2594 | we have performed, but it is not supported |
2595 | by iterate_fix_dominators. */ |
2596 | free_dominance_info (CDI_POST_DOMINATORS); |
2597 | } |
2598 | else |
2599 | { |
2600 | if (dump_file) |
2601 | { |
2602 | fputs (s: "Bailing out - " , stream: dump_file); |
2603 | fputs (s: sconv.m_reason, stream: dump_file); |
2604 | fputs (s: "\n--------------------------------\n" , stream: dump_file); |
2605 | } |
2606 | } |
2607 | } |
2608 | } |
2609 | |
2610 | return cfg_altered ? TODO_cleanup_cfg : 0;; |
2611 | } |
2612 | |
2613 | } // anon namespace |
2614 | |
2615 | gimple_opt_pass * |
2616 | make_pass_convert_switch (gcc::context *ctxt) |
2617 | { |
2618 | return new pass_convert_switch (ctxt); |
2619 | } |
2620 | |
2621 | /* The main function of the pass scans statements for switches and invokes |
2622 | process_switch on them. */ |
2623 | |
2624 | namespace { |
2625 | |
2626 | template <bool O0> class pass_lower_switch: public gimple_opt_pass |
2627 | { |
2628 | public: |
2629 | pass_lower_switch (gcc::context *ctxt) : gimple_opt_pass (data, ctxt) {} |
2630 | |
2631 | static const pass_data data; |
2632 | opt_pass * |
2633 | clone () final override |
2634 | { |
2635 | return new pass_lower_switch<O0> (m_ctxt); |
2636 | } |
2637 | |
2638 | bool |
2639 | gate (function *) final override |
2640 | { |
2641 | return !O0 || !optimize; |
2642 | } |
2643 | |
2644 | unsigned int execute (function *fun) final override; |
2645 | }; // class pass_lower_switch |
2646 | |
2647 | template <bool O0> |
2648 | const pass_data pass_lower_switch<O0>::data = { |
2649 | .type: .type: .type: GIMPLE_PASS, /* type */ |
2650 | .name: .name: .name: O0 ? "switchlower_O0" : "switchlower" , /* name */ |
2651 | .optinfo_flags: .optinfo_flags: .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
2652 | .tv_id: .tv_id: .tv_id: TV_TREE_SWITCH_LOWERING, /* tv_id */ |
2653 | .properties_required: .properties_required: .properties_required: ( PROP_cfg | PROP_ssa ), /* properties_required */ |
2654 | .properties_provided: .properties_provided: .properties_provided: 0, /* properties_provided */ |
2655 | .properties_destroyed: .properties_destroyed: .properties_destroyed: 0, /* properties_destroyed */ |
2656 | .todo_flags_start: .todo_flags_start: .todo_flags_start: 0, /* todo_flags_start */ |
2657 | TODO_update_ssa | TODO_cleanup_cfg, /* todo_flags_finish */ |
2658 | }; |
2659 | |
2660 | template <bool O0> |
2661 | unsigned int |
2662 | pass_lower_switch<O0>::execute (function *fun) |
2663 | { |
2664 | basic_block bb; |
2665 | bool expanded = false; |
2666 | |
2667 | auto_vec<gimple *> switch_statements; |
2668 | switch_statements.create (nelems: 1); |
2669 | |
2670 | FOR_EACH_BB_FN (bb, fun) |
2671 | { |
2672 | if (gswitch *swtch = safe_dyn_cast <gswitch *> (p: *gsi_last_bb (bb))) |
2673 | { |
2674 | if (!O0) |
2675 | group_case_labels_stmt (swtch); |
2676 | switch_statements.safe_push (obj: swtch); |
2677 | } |
2678 | } |
2679 | |
2680 | for (unsigned i = 0; i < switch_statements.length (); i++) |
2681 | { |
2682 | gimple *stmt = switch_statements[i]; |
2683 | if (dump_file) |
2684 | { |
2685 | expanded_location loc = expand_location (gimple_location (g: stmt)); |
2686 | |
2687 | fprintf (stream: dump_file, format: "beginning to process the following " |
2688 | "SWITCH statement (%s:%d) : ------- \n" , |
2689 | loc.file, loc.line); |
2690 | print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); |
2691 | putc (c: '\n', stream: dump_file); |
2692 | } |
2693 | |
2694 | gswitch *swtch = dyn_cast<gswitch *> (p: stmt); |
2695 | if (swtch) |
2696 | { |
2697 | switch_decision_tree dt (swtch); |
2698 | expanded |= dt.analyze_switch_statement (); |
2699 | } |
2700 | } |
2701 | |
2702 | if (expanded) |
2703 | { |
2704 | free_dominance_info (CDI_DOMINATORS); |
2705 | free_dominance_info (CDI_POST_DOMINATORS); |
2706 | mark_virtual_operands_for_renaming (cfun); |
2707 | } |
2708 | |
2709 | return 0; |
2710 | } |
2711 | |
2712 | } // anon namespace |
2713 | |
2714 | gimple_opt_pass * |
2715 | make_pass_lower_switch_O0 (gcc::context *ctxt) |
2716 | { |
2717 | return new pass_lower_switch<true> (ctxt); |
2718 | } |
2719 | gimple_opt_pass * |
2720 | make_pass_lower_switch (gcc::context *ctxt) |
2721 | { |
2722 | return new pass_lower_switch<false> (ctxt); |
2723 | } |
2724 | |