1 | /* Induction variable optimizations. |
2 | Copyright (C) 2003-2023 Free Software Foundation, Inc. |
3 | |
4 | This file is part of GCC. |
5 | |
6 | GCC is free software; you can redistribute it and/or modify it |
7 | under the terms of the GNU General Public License as published by the |
8 | Free Software Foundation; either version 3, or (at your option) any |
9 | later version. |
10 | |
11 | GCC is distributed in the hope that it will be useful, but WITHOUT |
12 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
14 | for more details. |
15 | |
16 | You should have received a copy of the GNU General Public License |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ |
19 | |
20 | /* This pass tries to find the optimal set of induction variables for the loop. |
21 | It optimizes just the basic linear induction variables (although adding |
22 | support for other types should not be too hard). It includes the |
23 | optimizations commonly known as strength reduction, induction variable |
24 | coalescing and induction variable elimination. It does it in the |
25 | following steps: |
26 | |
27 | 1) The interesting uses of induction variables are found. This includes |
28 | |
29 | -- uses of induction variables in non-linear expressions |
30 | -- addresses of arrays |
31 | -- comparisons of induction variables |
32 | |
33 | Note the interesting uses are categorized and handled in group. |
34 | Generally, address type uses are grouped together if their iv bases |
35 | are different in constant offset. |
36 | |
37 | 2) Candidates for the induction variables are found. This includes |
38 | |
39 | -- old induction variables |
40 | -- the variables defined by expressions derived from the "interesting |
41 | groups/uses" above |
42 | |
43 | 3) The optimal (w.r. to a cost function) set of variables is chosen. The |
44 | cost function assigns a cost to sets of induction variables and consists |
45 | of three parts: |
46 | |
47 | -- The group/use costs. Each of the interesting groups/uses chooses |
48 | the best induction variable in the set and adds its cost to the sum. |
49 | The cost reflects the time spent on modifying the induction variables |
50 | value to be usable for the given purpose (adding base and offset for |
51 | arrays, etc.). |
52 | -- The variable costs. Each of the variables has a cost assigned that |
53 | reflects the costs associated with incrementing the value of the |
54 | variable. The original variables are somewhat preferred. |
55 | -- The set cost. Depending on the size of the set, extra cost may be |
56 | added to reflect register pressure. |
57 | |
58 | All the costs are defined in a machine-specific way, using the target |
59 | hooks and machine descriptions to determine them. |
60 | |
61 | 4) The trees are transformed to use the new variables, the dead code is |
62 | removed. |
63 | |
64 | All of this is done loop by loop. Doing it globally is theoretically |
65 | possible, it might give a better performance and it might enable us |
66 | to decide costs more precisely, but getting all the interactions right |
67 | would be complicated. |
68 | |
69 | For the targets supporting low-overhead loops, IVOPTs has to take care of |
70 | the loops which will probably be transformed in RTL doloop optimization, |
71 | to try to make selected IV candidate set optimal. The process of doloop |
72 | support includes: |
73 | |
74 | 1) Analyze the current loop will be transformed to doloop or not, find and |
75 | mark its compare type IV use as doloop use (iv_group field doloop_p), and |
76 | set flag doloop_use_p of ivopts_data to notify subsequent processings on |
77 | doloop. See analyze_and_mark_doloop_use and its callees for the details. |
78 | The target hook predict_doloop_p can be used for target specific checks. |
79 | |
80 | 2) Add one doloop dedicated IV cand {(may_be_zero ? 1 : (niter + 1)), +, -1}, |
81 | set flag doloop_p of iv_cand, step cost is set as zero and no extra cost |
82 | like biv. For cost determination between doloop IV cand and IV use, the |
83 | target hooks doloop_cost_for_generic and doloop_cost_for_address are |
84 | provided to add on extra costs for generic type and address type IV use. |
85 | Zero cost is assigned to the pair between doloop IV cand and doloop IV |
86 | use, and bound zero is set for IV elimination. |
87 | |
88 | 3) With the cost setting in step 2), the current cost model based IV |
89 | selection algorithm will process as usual, pick up doloop dedicated IV if |
90 | profitable. */ |
91 | |
92 | #include "config.h" |
93 | #include "system.h" |
94 | #include "coretypes.h" |
95 | #include "backend.h" |
96 | #include "rtl.h" |
97 | #include "tree.h" |
98 | #include "gimple.h" |
99 | #include "cfghooks.h" |
100 | #include "tree-pass.h" |
101 | #include "memmodel.h" |
102 | #include "tm_p.h" |
103 | #include "ssa.h" |
104 | #include "expmed.h" |
105 | #include "insn-config.h" |
106 | #include "emit-rtl.h" |
107 | #include "recog.h" |
108 | #include "cgraph.h" |
109 | #include "gimple-pretty-print.h" |
110 | #include "alias.h" |
111 | #include "fold-const.h" |
112 | #include "stor-layout.h" |
113 | #include "tree-eh.h" |
114 | #include "gimplify.h" |
115 | #include "gimple-iterator.h" |
116 | #include "gimplify-me.h" |
117 | #include "tree-cfg.h" |
118 | #include "tree-ssa-loop-ivopts.h" |
119 | #include "tree-ssa-loop-manip.h" |
120 | #include "tree-ssa-loop-niter.h" |
121 | #include "tree-ssa-loop.h" |
122 | #include "explow.h" |
123 | #include "expr.h" |
124 | #include "tree-dfa.h" |
125 | #include "tree-ssa.h" |
126 | #include "cfgloop.h" |
127 | #include "tree-scalar-evolution.h" |
128 | #include "tree-affine.h" |
129 | #include "tree-ssa-propagate.h" |
130 | #include "tree-ssa-address.h" |
131 | #include "builtins.h" |
132 | #include "tree-vectorizer.h" |
133 | #include "dbgcnt.h" |
134 | #include "cfganal.h" |
135 | |
136 | /* For lang_hooks.types.type_for_mode. */ |
137 | #include "langhooks.h" |
138 | |
139 | /* FIXME: Expressions are expanded to RTL in this pass to determine the |
140 | cost of different addressing modes. This should be moved to a TBD |
141 | interface between the GIMPLE and RTL worlds. */ |
142 | |
143 | /* The infinite cost. */ |
144 | #define INFTY 1000000000 |
145 | |
146 | /* Returns the expected number of loop iterations for LOOP. |
147 | The average trip count is computed from profile data if it |
148 | exists. */ |
149 | |
150 | static inline HOST_WIDE_INT |
151 | avg_loop_niter (class loop *loop) |
152 | { |
153 | HOST_WIDE_INT niter = estimated_stmt_executions_int (loop); |
154 | if (niter == -1) |
155 | { |
156 | niter = likely_max_stmt_executions_int (loop); |
157 | |
158 | if (niter == -1 || niter > param_avg_loop_niter) |
159 | return param_avg_loop_niter; |
160 | } |
161 | |
162 | return niter; |
163 | } |
164 | |
165 | struct iv_use; |
166 | |
167 | /* Representation of the induction variable. */ |
168 | struct iv |
169 | { |
170 | tree base; /* Initial value of the iv. */ |
171 | tree base_object; /* A memory object to that the induction variable points. */ |
172 | tree step; /* Step of the iv (constant only). */ |
173 | tree ssa_name; /* The ssa name with the value. */ |
174 | struct iv_use *nonlin_use; /* The identifier in the use if it is the case. */ |
175 | bool biv_p; /* Is it a biv? */ |
176 | bool no_overflow; /* True if the iv doesn't overflow. */ |
177 | bool have_address_use;/* For biv, indicate if it's used in any address |
178 | type use. */ |
179 | }; |
180 | |
181 | /* Per-ssa version information (induction variable descriptions, etc.). */ |
182 | struct version_info |
183 | { |
184 | tree name; /* The ssa name. */ |
185 | struct iv *iv; /* Induction variable description. */ |
186 | bool has_nonlin_use; /* For a loop-level invariant, whether it is used in |
187 | an expression that is not an induction variable. */ |
188 | bool preserve_biv; /* For the original biv, whether to preserve it. */ |
189 | unsigned inv_id; /* Id of an invariant. */ |
190 | }; |
191 | |
192 | /* Types of uses. */ |
193 | enum use_type |
194 | { |
195 | USE_NONLINEAR_EXPR, /* Use in a nonlinear expression. */ |
196 | USE_REF_ADDRESS, /* Use is an address for an explicit memory |
197 | reference. */ |
198 | USE_PTR_ADDRESS, /* Use is a pointer argument to a function in |
199 | cases where the expansion of the function |
200 | will turn the argument into a normal address. */ |
201 | USE_COMPARE /* Use is a compare. */ |
202 | }; |
203 | |
204 | /* Cost of a computation. */ |
205 | class comp_cost |
206 | { |
207 | public: |
208 | comp_cost (): cost (0), complexity (0), scratch (0) |
209 | {} |
210 | |
211 | comp_cost (int64_t cost, unsigned complexity, int64_t scratch = 0) |
212 | : cost (cost), complexity (complexity), scratch (scratch) |
213 | {} |
214 | |
215 | /* Returns true if COST is infinite. */ |
216 | bool infinite_cost_p (); |
217 | |
218 | /* Adds costs COST1 and COST2. */ |
219 | friend comp_cost operator+ (comp_cost cost1, comp_cost cost2); |
220 | |
221 | /* Adds COST to the comp_cost. */ |
222 | comp_cost operator+= (comp_cost cost); |
223 | |
224 | /* Adds constant C to this comp_cost. */ |
225 | comp_cost operator+= (HOST_WIDE_INT c); |
226 | |
227 | /* Subtracts constant C to this comp_cost. */ |
228 | comp_cost operator-= (HOST_WIDE_INT c); |
229 | |
230 | /* Divide the comp_cost by constant C. */ |
231 | comp_cost operator/= (HOST_WIDE_INT c); |
232 | |
233 | /* Multiply the comp_cost by constant C. */ |
234 | comp_cost operator*= (HOST_WIDE_INT c); |
235 | |
236 | /* Subtracts costs COST1 and COST2. */ |
237 | friend comp_cost operator- (comp_cost cost1, comp_cost cost2); |
238 | |
239 | /* Subtracts COST from this comp_cost. */ |
240 | comp_cost operator-= (comp_cost cost); |
241 | |
242 | /* Returns true if COST1 is smaller than COST2. */ |
243 | friend bool operator< (comp_cost cost1, comp_cost cost2); |
244 | |
245 | /* Returns true if COST1 and COST2 are equal. */ |
246 | friend bool operator== (comp_cost cost1, comp_cost cost2); |
247 | |
248 | /* Returns true if COST1 is smaller or equal than COST2. */ |
249 | friend bool operator<= (comp_cost cost1, comp_cost cost2); |
250 | |
251 | int64_t cost; /* The runtime cost. */ |
252 | unsigned complexity; /* The estimate of the complexity of the code for |
253 | the computation (in no concrete units -- |
254 | complexity field should be larger for more |
255 | complex expressions and addressing modes). */ |
256 | int64_t scratch; /* Scratch used during cost computation. */ |
257 | }; |
258 | |
259 | static const comp_cost no_cost; |
260 | static const comp_cost infinite_cost (INFTY, 0, INFTY); |
261 | |
262 | bool |
263 | comp_cost::infinite_cost_p () |
264 | { |
265 | return cost == INFTY; |
266 | } |
267 | |
268 | comp_cost |
269 | operator+ (comp_cost cost1, comp_cost cost2) |
270 | { |
271 | if (cost1.infinite_cost_p () || cost2.infinite_cost_p ()) |
272 | return infinite_cost; |
273 | |
274 | gcc_assert (cost1.cost + cost2.cost < infinite_cost.cost); |
275 | cost1.cost += cost2.cost; |
276 | cost1.complexity += cost2.complexity; |
277 | |
278 | return cost1; |
279 | } |
280 | |
281 | comp_cost |
282 | operator- (comp_cost cost1, comp_cost cost2) |
283 | { |
284 | if (cost1.infinite_cost_p ()) |
285 | return infinite_cost; |
286 | |
287 | gcc_assert (!cost2.infinite_cost_p ()); |
288 | gcc_assert (cost1.cost - cost2.cost < infinite_cost.cost); |
289 | |
290 | cost1.cost -= cost2.cost; |
291 | cost1.complexity -= cost2.complexity; |
292 | |
293 | return cost1; |
294 | } |
295 | |
296 | comp_cost |
297 | comp_cost::operator+= (comp_cost cost) |
298 | { |
299 | *this = *this + cost; |
300 | return *this; |
301 | } |
302 | |
303 | comp_cost |
304 | comp_cost::operator+= (HOST_WIDE_INT c) |
305 | { |
306 | if (c >= INFTY) |
307 | this->cost = INFTY; |
308 | |
309 | if (infinite_cost_p ()) |
310 | return *this; |
311 | |
312 | gcc_assert (this->cost + c < infinite_cost.cost); |
313 | this->cost += c; |
314 | |
315 | return *this; |
316 | } |
317 | |
318 | comp_cost |
319 | comp_cost::operator-= (HOST_WIDE_INT c) |
320 | { |
321 | if (infinite_cost_p ()) |
322 | return *this; |
323 | |
324 | gcc_assert (this->cost - c < infinite_cost.cost); |
325 | this->cost -= c; |
326 | |
327 | return *this; |
328 | } |
329 | |
330 | comp_cost |
331 | comp_cost::operator/= (HOST_WIDE_INT c) |
332 | { |
333 | gcc_assert (c != 0); |
334 | if (infinite_cost_p ()) |
335 | return *this; |
336 | |
337 | this->cost /= c; |
338 | |
339 | return *this; |
340 | } |
341 | |
342 | comp_cost |
343 | comp_cost::operator*= (HOST_WIDE_INT c) |
344 | { |
345 | if (infinite_cost_p ()) |
346 | return *this; |
347 | |
348 | gcc_assert (this->cost * c < infinite_cost.cost); |
349 | this->cost *= c; |
350 | |
351 | return *this; |
352 | } |
353 | |
354 | comp_cost |
355 | comp_cost::operator-= (comp_cost cost) |
356 | { |
357 | *this = *this - cost; |
358 | return *this; |
359 | } |
360 | |
361 | bool |
362 | operator< (comp_cost cost1, comp_cost cost2) |
363 | { |
364 | if (cost1.cost == cost2.cost) |
365 | return cost1.complexity < cost2.complexity; |
366 | |
367 | return cost1.cost < cost2.cost; |
368 | } |
369 | |
370 | bool |
371 | operator== (comp_cost cost1, comp_cost cost2) |
372 | { |
373 | return cost1.cost == cost2.cost |
374 | && cost1.complexity == cost2.complexity; |
375 | } |
376 | |
377 | bool |
378 | operator<= (comp_cost cost1, comp_cost cost2) |
379 | { |
380 | return cost1 < cost2 || cost1 == cost2; |
381 | } |
382 | |
383 | struct iv_inv_expr_ent; |
384 | |
385 | /* The candidate - cost pair. */ |
386 | class cost_pair |
387 | { |
388 | public: |
389 | struct iv_cand *cand; /* The candidate. */ |
390 | comp_cost cost; /* The cost. */ |
391 | enum tree_code comp; /* For iv elimination, the comparison. */ |
392 | bitmap inv_vars; /* The list of invariant ssa_vars that have to be |
393 | preserved when representing iv_use with iv_cand. */ |
394 | bitmap inv_exprs; /* The list of newly created invariant expressions |
395 | when representing iv_use with iv_cand. */ |
396 | tree value; /* For final value elimination, the expression for |
397 | the final value of the iv. For iv elimination, |
398 | the new bound to compare with. */ |
399 | }; |
400 | |
401 | /* Use. */ |
402 | struct iv_use |
403 | { |
404 | unsigned id; /* The id of the use. */ |
405 | unsigned group_id; /* The group id the use belongs to. */ |
406 | enum use_type type; /* Type of the use. */ |
407 | tree mem_type; /* The memory type to use when testing whether an |
408 | address is legitimate, and what the address's |
409 | cost is. */ |
410 | struct iv *iv; /* The induction variable it is based on. */ |
411 | gimple *stmt; /* Statement in that it occurs. */ |
412 | tree *op_p; /* The place where it occurs. */ |
413 | |
414 | tree addr_base; /* Base address with const offset stripped. */ |
415 | poly_uint64 addr_offset; |
416 | /* Const offset stripped from base address. */ |
417 | }; |
418 | |
419 | /* Group of uses. */ |
420 | struct iv_group |
421 | { |
422 | /* The id of the group. */ |
423 | unsigned id; |
424 | /* Uses of the group are of the same type. */ |
425 | enum use_type type; |
426 | /* The set of "related" IV candidates, plus the important ones. */ |
427 | bitmap related_cands; |
428 | /* Number of IV candidates in the cost_map. */ |
429 | unsigned n_map_members; |
430 | /* The costs wrto the iv candidates. */ |
431 | class cost_pair *cost_map; |
432 | /* The selected candidate for the group. */ |
433 | struct iv_cand *selected; |
434 | /* To indicate this is a doloop use group. */ |
435 | bool doloop_p; |
436 | /* Uses in the group. */ |
437 | vec<struct iv_use *> vuses; |
438 | }; |
439 | |
440 | /* The position where the iv is computed. */ |
441 | enum iv_position |
442 | { |
443 | IP_NORMAL, /* At the end, just before the exit condition. */ |
444 | IP_END, /* At the end of the latch block. */ |
445 | IP_BEFORE_USE, /* Immediately before a specific use. */ |
446 | IP_AFTER_USE, /* Immediately after a specific use. */ |
447 | IP_ORIGINAL /* The original biv. */ |
448 | }; |
449 | |
450 | /* The induction variable candidate. */ |
451 | struct iv_cand |
452 | { |
453 | unsigned id; /* The number of the candidate. */ |
454 | bool important; /* Whether this is an "important" candidate, i.e. such |
455 | that it should be considered by all uses. */ |
456 | bool involves_undefs; /* Whether the IV involves undefined values. */ |
457 | ENUM_BITFIELD(iv_position) pos : 8; /* Where it is computed. */ |
458 | gimple *incremented_at;/* For original biv, the statement where it is |
459 | incremented. */ |
460 | tree var_before; /* The variable used for it before increment. */ |
461 | tree var_after; /* The variable used for it after increment. */ |
462 | struct iv *iv; /* The value of the candidate. NULL for |
463 | "pseudocandidate" used to indicate the possibility |
464 | to replace the final value of an iv by direct |
465 | computation of the value. */ |
466 | unsigned cost; /* Cost of the candidate. */ |
467 | unsigned cost_step; /* Cost of the candidate's increment operation. */ |
468 | struct iv_use *ainc_use; /* For IP_{BEFORE,AFTER}_USE candidates, the place |
469 | where it is incremented. */ |
470 | bitmap inv_vars; /* The list of invariant ssa_vars used in step of the |
471 | iv_cand. */ |
472 | bitmap inv_exprs; /* If step is more complicated than a single ssa_var, |
473 | handle it as a new invariant expression which will |
474 | be hoisted out of loop. */ |
475 | struct iv *orig_iv; /* The original iv if this cand is added from biv with |
476 | smaller type. */ |
477 | bool doloop_p; /* Whether this is a doloop candidate. */ |
478 | }; |
479 | |
480 | /* Hashtable entry for common candidate derived from iv uses. */ |
481 | class iv_common_cand |
482 | { |
483 | public: |
484 | tree base; |
485 | tree step; |
486 | /* IV uses from which this common candidate is derived. */ |
487 | auto_vec<struct iv_use *> uses; |
488 | hashval_t hash; |
489 | }; |
490 | |
491 | /* Hashtable helpers. */ |
492 | |
493 | struct iv_common_cand_hasher : delete_ptr_hash <iv_common_cand> |
494 | { |
495 | static inline hashval_t hash (const iv_common_cand *); |
496 | static inline bool equal (const iv_common_cand *, const iv_common_cand *); |
497 | }; |
498 | |
499 | /* Hash function for possible common candidates. */ |
500 | |
501 | inline hashval_t |
502 | iv_common_cand_hasher::hash (const iv_common_cand *ccand) |
503 | { |
504 | return ccand->hash; |
505 | } |
506 | |
507 | /* Hash table equality function for common candidates. */ |
508 | |
509 | inline bool |
510 | iv_common_cand_hasher::equal (const iv_common_cand *ccand1, |
511 | const iv_common_cand *ccand2) |
512 | { |
513 | return (ccand1->hash == ccand2->hash |
514 | && operand_equal_p (ccand1->base, ccand2->base, flags: 0) |
515 | && operand_equal_p (ccand1->step, ccand2->step, flags: 0) |
516 | && (TYPE_PRECISION (TREE_TYPE (ccand1->base)) |
517 | == TYPE_PRECISION (TREE_TYPE (ccand2->base)))); |
518 | } |
519 | |
520 | /* Loop invariant expression hashtable entry. */ |
521 | |
522 | struct iv_inv_expr_ent |
523 | { |
524 | /* Tree expression of the entry. */ |
525 | tree expr; |
526 | /* Unique indentifier. */ |
527 | int id; |
528 | /* Hash value. */ |
529 | hashval_t hash; |
530 | }; |
531 | |
532 | /* Sort iv_inv_expr_ent pair A and B by id field. */ |
533 | |
534 | static int |
535 | sort_iv_inv_expr_ent (const void *a, const void *b) |
536 | { |
537 | const iv_inv_expr_ent * const *e1 = (const iv_inv_expr_ent * const *) (a); |
538 | const iv_inv_expr_ent * const *e2 = (const iv_inv_expr_ent * const *) (b); |
539 | |
540 | unsigned id1 = (*e1)->id; |
541 | unsigned id2 = (*e2)->id; |
542 | |
543 | if (id1 < id2) |
544 | return -1; |
545 | else if (id1 > id2) |
546 | return 1; |
547 | else |
548 | return 0; |
549 | } |
550 | |
551 | /* Hashtable helpers. */ |
552 | |
553 | struct iv_inv_expr_hasher : free_ptr_hash <iv_inv_expr_ent> |
554 | { |
555 | static inline hashval_t hash (const iv_inv_expr_ent *); |
556 | static inline bool equal (const iv_inv_expr_ent *, const iv_inv_expr_ent *); |
557 | }; |
558 | |
559 | /* Return true if uses of type TYPE represent some form of address. */ |
560 | |
561 | inline bool |
562 | address_p (use_type type) |
563 | { |
564 | return type == USE_REF_ADDRESS || type == USE_PTR_ADDRESS; |
565 | } |
566 | |
567 | /* Hash function for loop invariant expressions. */ |
568 | |
569 | inline hashval_t |
570 | iv_inv_expr_hasher::hash (const iv_inv_expr_ent *expr) |
571 | { |
572 | return expr->hash; |
573 | } |
574 | |
575 | /* Hash table equality function for expressions. */ |
576 | |
577 | inline bool |
578 | iv_inv_expr_hasher::equal (const iv_inv_expr_ent *expr1, |
579 | const iv_inv_expr_ent *expr2) |
580 | { |
581 | return expr1->hash == expr2->hash |
582 | && operand_equal_p (expr1->expr, expr2->expr, flags: 0); |
583 | } |
584 | |
585 | struct ivopts_data |
586 | { |
587 | /* The currently optimized loop. */ |
588 | class loop *current_loop; |
589 | location_t loop_loc; |
590 | |
591 | /* Numbers of iterations for all exits of the current loop. */ |
592 | hash_map<edge, tree_niter_desc *> *niters; |
593 | |
594 | /* Number of registers used in it. */ |
595 | unsigned regs_used; |
596 | |
597 | /* The size of version_info array allocated. */ |
598 | unsigned version_info_size; |
599 | |
600 | /* The array of information for the ssa names. */ |
601 | struct version_info *version_info; |
602 | |
603 | /* The hashtable of loop invariant expressions created |
604 | by ivopt. */ |
605 | hash_table<iv_inv_expr_hasher> *inv_expr_tab; |
606 | |
607 | /* The bitmap of indices in version_info whose value was changed. */ |
608 | bitmap relevant; |
609 | |
610 | /* The uses of induction variables. */ |
611 | vec<iv_group *> vgroups; |
612 | |
613 | /* The candidates. */ |
614 | vec<iv_cand *> vcands; |
615 | |
616 | /* A bitmap of important candidates. */ |
617 | bitmap important_candidates; |
618 | |
619 | /* Cache used by tree_to_aff_combination_expand. */ |
620 | hash_map<tree, name_expansion *> *name_expansion_cache; |
621 | |
622 | /* The hashtable of common candidates derived from iv uses. */ |
623 | hash_table<iv_common_cand_hasher> *iv_common_cand_tab; |
624 | |
625 | /* The common candidates. */ |
626 | vec<iv_common_cand *> iv_common_cands; |
627 | |
628 | /* Hash map recording base object information of tree exp. */ |
629 | hash_map<tree, tree> *base_object_map; |
630 | |
631 | /* The maximum invariant variable id. */ |
632 | unsigned max_inv_var_id; |
633 | |
634 | /* The maximum invariant expression id. */ |
635 | unsigned max_inv_expr_id; |
636 | |
637 | /* Number of no_overflow BIVs which are not used in memory address. */ |
638 | unsigned bivs_not_used_in_addr; |
639 | |
640 | /* Obstack for iv structure. */ |
641 | struct obstack iv_obstack; |
642 | |
643 | /* Whether to consider just related and important candidates when replacing a |
644 | use. */ |
645 | bool consider_all_candidates; |
646 | |
647 | /* Are we optimizing for speed? */ |
648 | bool speed; |
649 | |
650 | /* Whether the loop body includes any function calls. */ |
651 | bool body_includes_call; |
652 | |
653 | /* Whether the loop body can only be exited via single exit. */ |
654 | bool loop_single_exit_p; |
655 | |
656 | /* Whether the loop has doloop comparison use. */ |
657 | bool doloop_use_p; |
658 | }; |
659 | |
660 | /* An assignment of iv candidates to uses. */ |
661 | |
662 | class iv_ca |
663 | { |
664 | public: |
665 | /* The number of uses covered by the assignment. */ |
666 | unsigned upto; |
667 | |
668 | /* Number of uses that cannot be expressed by the candidates in the set. */ |
669 | unsigned bad_groups; |
670 | |
671 | /* Candidate assigned to a use, together with the related costs. */ |
672 | class cost_pair **cand_for_group; |
673 | |
674 | /* Number of times each candidate is used. */ |
675 | unsigned *n_cand_uses; |
676 | |
677 | /* The candidates used. */ |
678 | bitmap cands; |
679 | |
680 | /* The number of candidates in the set. */ |
681 | unsigned n_cands; |
682 | |
683 | /* The number of invariants needed, including both invariant variants and |
684 | invariant expressions. */ |
685 | unsigned n_invs; |
686 | |
687 | /* Total cost of expressing uses. */ |
688 | comp_cost cand_use_cost; |
689 | |
690 | /* Total cost of candidates. */ |
691 | int64_t cand_cost; |
692 | |
693 | /* Number of times each invariant variable is used. */ |
694 | unsigned *n_inv_var_uses; |
695 | |
696 | /* Number of times each invariant expression is used. */ |
697 | unsigned *n_inv_expr_uses; |
698 | |
699 | /* Total cost of the assignment. */ |
700 | comp_cost cost; |
701 | }; |
702 | |
703 | /* Difference of two iv candidate assignments. */ |
704 | |
705 | struct iv_ca_delta |
706 | { |
707 | /* Changed group. */ |
708 | struct iv_group *group; |
709 | |
710 | /* An old assignment (for rollback purposes). */ |
711 | class cost_pair *old_cp; |
712 | |
713 | /* A new assignment. */ |
714 | class cost_pair *new_cp; |
715 | |
716 | /* Next change in the list. */ |
717 | struct iv_ca_delta *next; |
718 | }; |
719 | |
720 | /* Bound on number of candidates below that all candidates are considered. */ |
721 | |
722 | #define CONSIDER_ALL_CANDIDATES_BOUND \ |
723 | ((unsigned) param_iv_consider_all_candidates_bound) |
724 | |
725 | /* If there are more iv occurrences, we just give up (it is quite unlikely that |
726 | optimizing such a loop would help, and it would take ages). */ |
727 | |
728 | #define MAX_CONSIDERED_GROUPS \ |
729 | ((unsigned) param_iv_max_considered_uses) |
730 | |
731 | /* If there are at most this number of ivs in the set, try removing unnecessary |
732 | ivs from the set always. */ |
733 | |
734 | #define ALWAYS_PRUNE_CAND_SET_BOUND \ |
735 | ((unsigned) param_iv_always_prune_cand_set_bound) |
736 | |
737 | /* The list of trees for that the decl_rtl field must be reset is stored |
738 | here. */ |
739 | |
740 | static vec<tree> decl_rtl_to_reset; |
741 | |
742 | static comp_cost force_expr_to_var_cost (tree, bool); |
743 | |
744 | /* The single loop exit if it dominates the latch, NULL otherwise. */ |
745 | |
746 | edge |
747 | single_dom_exit (class loop *loop) |
748 | { |
749 | edge exit = single_exit (loop); |
750 | |
751 | if (!exit) |
752 | return NULL; |
753 | |
754 | if (!just_once_each_iteration_p (loop, exit->src)) |
755 | return NULL; |
756 | |
757 | return exit; |
758 | } |
759 | |
760 | /* Dumps information about the induction variable IV to FILE. Don't dump |
761 | variable's name if DUMP_NAME is FALSE. The information is dumped with |
762 | preceding spaces indicated by INDENT_LEVEL. */ |
763 | |
764 | void |
765 | dump_iv (FILE *file, struct iv *iv, bool dump_name, unsigned indent_level) |
766 | { |
767 | const char *p; |
768 | const char spaces[9] = {' ', ' ', ' ', ' ', ' ', ' ', ' ', ' ', '\0'}; |
769 | |
770 | if (indent_level > 4) |
771 | indent_level = 4; |
772 | p = spaces + 8 - (indent_level << 1); |
773 | |
774 | fprintf (stream: file, format: "%sIV struct:\n" , p); |
775 | if (iv->ssa_name && dump_name) |
776 | { |
777 | fprintf (stream: file, format: "%s SSA_NAME:\t" , p); |
778 | print_generic_expr (file, iv->ssa_name, TDF_SLIM); |
779 | fprintf (stream: file, format: "\n" ); |
780 | } |
781 | |
782 | fprintf (stream: file, format: "%s Type:\t" , p); |
783 | print_generic_expr (file, TREE_TYPE (iv->base), TDF_SLIM); |
784 | fprintf (stream: file, format: "\n" ); |
785 | |
786 | fprintf (stream: file, format: "%s Base:\t" , p); |
787 | print_generic_expr (file, iv->base, TDF_SLIM); |
788 | fprintf (stream: file, format: "\n" ); |
789 | |
790 | fprintf (stream: file, format: "%s Step:\t" , p); |
791 | print_generic_expr (file, iv->step, TDF_SLIM); |
792 | fprintf (stream: file, format: "\n" ); |
793 | |
794 | if (iv->base_object) |
795 | { |
796 | fprintf (stream: file, format: "%s Object:\t" , p); |
797 | print_generic_expr (file, iv->base_object, TDF_SLIM); |
798 | fprintf (stream: file, format: "\n" ); |
799 | } |
800 | |
801 | fprintf (stream: file, format: "%s Biv:\t%c\n" , p, iv->biv_p ? 'Y' : 'N'); |
802 | |
803 | fprintf (stream: file, format: "%s Overflowness wrto loop niter:\t%s\n" , |
804 | p, iv->no_overflow ? "No-overflow" : "Overflow" ); |
805 | } |
806 | |
807 | /* Dumps information about the USE to FILE. */ |
808 | |
809 | void |
810 | dump_use (FILE *file, struct iv_use *use) |
811 | { |
812 | fprintf (stream: file, format: " Use %d.%d:\n" , use->group_id, use->id); |
813 | fprintf (stream: file, format: " At stmt:\t" ); |
814 | print_gimple_stmt (file, use->stmt, 0); |
815 | fprintf (stream: file, format: " At pos:\t" ); |
816 | if (use->op_p) |
817 | print_generic_expr (file, *use->op_p, TDF_SLIM); |
818 | fprintf (stream: file, format: "\n" ); |
819 | dump_iv (file, iv: use->iv, dump_name: false, indent_level: 2); |
820 | } |
821 | |
822 | /* Dumps information about the uses to FILE. */ |
823 | |
824 | void |
825 | dump_groups (FILE *file, struct ivopts_data *data) |
826 | { |
827 | unsigned i, j; |
828 | struct iv_group *group; |
829 | |
830 | for (i = 0; i < data->vgroups.length (); i++) |
831 | { |
832 | group = data->vgroups[i]; |
833 | fprintf (stream: file, format: "Group %d:\n" , group->id); |
834 | if (group->type == USE_NONLINEAR_EXPR) |
835 | fprintf (stream: file, format: " Type:\tGENERIC\n" ); |
836 | else if (group->type == USE_REF_ADDRESS) |
837 | fprintf (stream: file, format: " Type:\tREFERENCE ADDRESS\n" ); |
838 | else if (group->type == USE_PTR_ADDRESS) |
839 | fprintf (stream: file, format: " Type:\tPOINTER ARGUMENT ADDRESS\n" ); |
840 | else |
841 | { |
842 | gcc_assert (group->type == USE_COMPARE); |
843 | fprintf (stream: file, format: " Type:\tCOMPARE\n" ); |
844 | } |
845 | for (j = 0; j < group->vuses.length (); j++) |
846 | dump_use (file, use: group->vuses[j]); |
847 | } |
848 | } |
849 | |
850 | /* Dumps information about induction variable candidate CAND to FILE. */ |
851 | |
852 | void |
853 | dump_cand (FILE *file, struct iv_cand *cand) |
854 | { |
855 | struct iv *iv = cand->iv; |
856 | |
857 | fprintf (stream: file, format: "Candidate %d:\n" , cand->id); |
858 | if (cand->inv_vars) |
859 | { |
860 | fprintf (stream: file, format: " Depend on inv.vars: " ); |
861 | dump_bitmap (file, map: cand->inv_vars); |
862 | } |
863 | if (cand->inv_exprs) |
864 | { |
865 | fprintf (stream: file, format: " Depend on inv.exprs: " ); |
866 | dump_bitmap (file, map: cand->inv_exprs); |
867 | } |
868 | |
869 | if (cand->var_before) |
870 | { |
871 | fprintf (stream: file, format: " Var befor: " ); |
872 | print_generic_expr (file, cand->var_before, TDF_SLIM); |
873 | fprintf (stream: file, format: "\n" ); |
874 | } |
875 | if (cand->var_after) |
876 | { |
877 | fprintf (stream: file, format: " Var after: " ); |
878 | print_generic_expr (file, cand->var_after, TDF_SLIM); |
879 | fprintf (stream: file, format: "\n" ); |
880 | } |
881 | |
882 | switch (cand->pos) |
883 | { |
884 | case IP_NORMAL: |
885 | fprintf (stream: file, format: " Incr POS: before exit test\n" ); |
886 | break; |
887 | |
888 | case IP_BEFORE_USE: |
889 | fprintf (stream: file, format: " Incr POS: before use %d\n" , cand->ainc_use->id); |
890 | break; |
891 | |
892 | case IP_AFTER_USE: |
893 | fprintf (stream: file, format: " Incr POS: after use %d\n" , cand->ainc_use->id); |
894 | break; |
895 | |
896 | case IP_END: |
897 | fprintf (stream: file, format: " Incr POS: at end\n" ); |
898 | break; |
899 | |
900 | case IP_ORIGINAL: |
901 | fprintf (stream: file, format: " Incr POS: orig biv\n" ); |
902 | break; |
903 | } |
904 | |
905 | dump_iv (file, iv, dump_name: false, indent_level: 1); |
906 | } |
907 | |
908 | /* Returns the info for ssa version VER. */ |
909 | |
910 | static inline struct version_info * |
911 | ver_info (struct ivopts_data *data, unsigned ver) |
912 | { |
913 | return data->version_info + ver; |
914 | } |
915 | |
916 | /* Returns the info for ssa name NAME. */ |
917 | |
918 | static inline struct version_info * |
919 | name_info (struct ivopts_data *data, tree name) |
920 | { |
921 | return ver_info (data, SSA_NAME_VERSION (name)); |
922 | } |
923 | |
924 | /* Returns true if STMT is after the place where the IP_NORMAL ivs will be |
925 | emitted in LOOP. */ |
926 | |
927 | static bool |
928 | stmt_after_ip_normal_pos (class loop *loop, gimple *stmt) |
929 | { |
930 | basic_block bb = ip_normal_pos (loop), sbb = gimple_bb (g: stmt); |
931 | |
932 | gcc_assert (bb); |
933 | |
934 | if (sbb == loop->latch) |
935 | return true; |
936 | |
937 | if (sbb != bb) |
938 | return false; |
939 | |
940 | return stmt == last_nondebug_stmt (bb); |
941 | } |
942 | |
943 | /* Returns true if STMT if after the place where the original induction |
944 | variable CAND is incremented. If TRUE_IF_EQUAL is set, we return true |
945 | if the positions are identical. */ |
946 | |
947 | static bool |
948 | stmt_after_inc_pos (struct iv_cand *cand, gimple *stmt, bool true_if_equal) |
949 | { |
950 | basic_block cand_bb = gimple_bb (g: cand->incremented_at); |
951 | basic_block stmt_bb = gimple_bb (g: stmt); |
952 | |
953 | if (!dominated_by_p (CDI_DOMINATORS, stmt_bb, cand_bb)) |
954 | return false; |
955 | |
956 | if (stmt_bb != cand_bb) |
957 | return true; |
958 | |
959 | if (true_if_equal |
960 | && gimple_uid (g: stmt) == gimple_uid (g: cand->incremented_at)) |
961 | return true; |
962 | return gimple_uid (g: stmt) > gimple_uid (g: cand->incremented_at); |
963 | } |
964 | |
965 | /* Returns true if STMT if after the place where the induction variable |
966 | CAND is incremented in LOOP. */ |
967 | |
968 | static bool |
969 | stmt_after_increment (class loop *loop, struct iv_cand *cand, gimple *stmt) |
970 | { |
971 | switch (cand->pos) |
972 | { |
973 | case IP_END: |
974 | return false; |
975 | |
976 | case IP_NORMAL: |
977 | return stmt_after_ip_normal_pos (loop, stmt); |
978 | |
979 | case IP_ORIGINAL: |
980 | case IP_AFTER_USE: |
981 | return stmt_after_inc_pos (cand, stmt, true_if_equal: false); |
982 | |
983 | case IP_BEFORE_USE: |
984 | return stmt_after_inc_pos (cand, stmt, true_if_equal: true); |
985 | |
986 | default: |
987 | gcc_unreachable (); |
988 | } |
989 | } |
990 | |
991 | /* walk_tree callback for contains_abnormal_ssa_name_p. */ |
992 | |
993 | static tree |
994 | contains_abnormal_ssa_name_p_1 (tree *tp, int *walk_subtrees, void *) |
995 | { |
996 | if (TREE_CODE (*tp) == SSA_NAME |
997 | && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (*tp)) |
998 | return *tp; |
999 | |
1000 | if (!EXPR_P (*tp)) |
1001 | *walk_subtrees = 0; |
1002 | |
1003 | return NULL_TREE; |
1004 | } |
1005 | |
1006 | /* Returns true if EXPR contains a ssa name that occurs in an |
1007 | abnormal phi node. */ |
1008 | |
1009 | bool |
1010 | contains_abnormal_ssa_name_p (tree expr) |
1011 | { |
1012 | return walk_tree_without_duplicates |
1013 | (&expr, contains_abnormal_ssa_name_p_1, NULL) != NULL_TREE; |
1014 | } |
1015 | |
1016 | /* Returns the structure describing number of iterations determined from |
1017 | EXIT of DATA->current_loop, or NULL if something goes wrong. */ |
1018 | |
1019 | static class tree_niter_desc * |
1020 | niter_for_exit (struct ivopts_data *data, edge exit) |
1021 | { |
1022 | class tree_niter_desc *desc; |
1023 | tree_niter_desc **slot; |
1024 | |
1025 | if (!data->niters) |
1026 | { |
1027 | data->niters = new hash_map<edge, tree_niter_desc *>; |
1028 | slot = NULL; |
1029 | } |
1030 | else |
1031 | slot = data->niters->get (k: exit); |
1032 | |
1033 | if (!slot) |
1034 | { |
1035 | /* Try to determine number of iterations. We cannot safely work with ssa |
1036 | names that appear in phi nodes on abnormal edges, so that we do not |
1037 | create overlapping life ranges for them (PR 27283). */ |
1038 | desc = XNEW (class tree_niter_desc); |
1039 | ::new (static_cast<void*> (desc)) tree_niter_desc (); |
1040 | if (!number_of_iterations_exit (data->current_loop, |
1041 | exit, niter: desc, true) |
1042 | || contains_abnormal_ssa_name_p (expr: desc->niter)) |
1043 | { |
1044 | desc->~tree_niter_desc (); |
1045 | XDELETE (desc); |
1046 | desc = NULL; |
1047 | } |
1048 | data->niters->put (k: exit, v: desc); |
1049 | } |
1050 | else |
1051 | desc = *slot; |
1052 | |
1053 | return desc; |
1054 | } |
1055 | |
1056 | /* Returns the structure describing number of iterations determined from |
1057 | single dominating exit of DATA->current_loop, or NULL if something |
1058 | goes wrong. */ |
1059 | |
1060 | static class tree_niter_desc * |
1061 | niter_for_single_dom_exit (struct ivopts_data *data) |
1062 | { |
1063 | edge exit = single_dom_exit (loop: data->current_loop); |
1064 | |
1065 | if (!exit) |
1066 | return NULL; |
1067 | |
1068 | return niter_for_exit (data, exit); |
1069 | } |
1070 | |
1071 | /* Initializes data structures used by the iv optimization pass, stored |
1072 | in DATA. */ |
1073 | |
1074 | static void |
1075 | tree_ssa_iv_optimize_init (struct ivopts_data *data) |
1076 | { |
1077 | data->version_info_size = 2 * num_ssa_names; |
1078 | data->version_info = XCNEWVEC (struct version_info, data->version_info_size); |
1079 | data->relevant = BITMAP_ALLOC (NULL); |
1080 | data->important_candidates = BITMAP_ALLOC (NULL); |
1081 | data->max_inv_var_id = 0; |
1082 | data->max_inv_expr_id = 0; |
1083 | data->niters = NULL; |
1084 | data->vgroups.create (nelems: 20); |
1085 | data->vcands.create (nelems: 20); |
1086 | data->inv_expr_tab = new hash_table<iv_inv_expr_hasher> (10); |
1087 | data->name_expansion_cache = NULL; |
1088 | data->base_object_map = NULL; |
1089 | data->iv_common_cand_tab = new hash_table<iv_common_cand_hasher> (10); |
1090 | data->iv_common_cands.create (nelems: 20); |
1091 | decl_rtl_to_reset.create (nelems: 20); |
1092 | gcc_obstack_init (&data->iv_obstack); |
1093 | } |
1094 | |
1095 | /* walk_tree callback for determine_base_object. */ |
1096 | |
1097 | static tree |
1098 | determine_base_object_1 (tree *tp, int *walk_subtrees, void *wdata) |
1099 | { |
1100 | tree_code code = TREE_CODE (*tp); |
1101 | tree obj = NULL_TREE; |
1102 | if (code == ADDR_EXPR) |
1103 | { |
1104 | tree base = get_base_address (TREE_OPERAND (*tp, 0)); |
1105 | if (!base) |
1106 | obj = *tp; |
1107 | else if (TREE_CODE (base) != MEM_REF) |
1108 | obj = fold_convert (ptr_type_node, build_fold_addr_expr (base)); |
1109 | } |
1110 | else if (code == SSA_NAME && POINTER_TYPE_P (TREE_TYPE (*tp))) |
1111 | obj = fold_convert (ptr_type_node, *tp); |
1112 | |
1113 | if (!obj) |
1114 | { |
1115 | if (!EXPR_P (*tp)) |
1116 | *walk_subtrees = 0; |
1117 | |
1118 | return NULL_TREE; |
1119 | } |
1120 | /* Record special node for multiple base objects and stop. */ |
1121 | if (*static_cast<tree *> (wdata)) |
1122 | { |
1123 | *static_cast<tree *> (wdata) = integer_zero_node; |
1124 | return integer_zero_node; |
1125 | } |
1126 | /* Record the base object and continue looking. */ |
1127 | *static_cast<tree *> (wdata) = obj; |
1128 | return NULL_TREE; |
1129 | } |
1130 | |
1131 | /* Returns a memory object to that EXPR points with caching. Return NULL if we |
1132 | are able to determine that it does not point to any such object; specially |
1133 | return integer_zero_node if EXPR contains multiple base objects. */ |
1134 | |
1135 | static tree |
1136 | determine_base_object (struct ivopts_data *data, tree expr) |
1137 | { |
1138 | tree *slot, obj = NULL_TREE; |
1139 | if (data->base_object_map) |
1140 | { |
1141 | if ((slot = data->base_object_map->get(k: expr)) != NULL) |
1142 | return *slot; |
1143 | } |
1144 | else |
1145 | data->base_object_map = new hash_map<tree, tree>; |
1146 | |
1147 | (void) walk_tree_without_duplicates (&expr, determine_base_object_1, &obj); |
1148 | data->base_object_map->put (k: expr, v: obj); |
1149 | return obj; |
1150 | } |
1151 | |
1152 | /* Return true if address expression with non-DECL_P operand appears |
1153 | in EXPR. */ |
1154 | |
1155 | static bool |
1156 | contain_complex_addr_expr (tree expr) |
1157 | { |
1158 | bool res = false; |
1159 | |
1160 | STRIP_NOPS (expr); |
1161 | switch (TREE_CODE (expr)) |
1162 | { |
1163 | case POINTER_PLUS_EXPR: |
1164 | case PLUS_EXPR: |
1165 | case MINUS_EXPR: |
1166 | res |= contain_complex_addr_expr (TREE_OPERAND (expr, 0)); |
1167 | res |= contain_complex_addr_expr (TREE_OPERAND (expr, 1)); |
1168 | break; |
1169 | |
1170 | case ADDR_EXPR: |
1171 | return (!DECL_P (TREE_OPERAND (expr, 0))); |
1172 | |
1173 | default: |
1174 | return false; |
1175 | } |
1176 | |
1177 | return res; |
1178 | } |
1179 | |
1180 | /* Allocates an induction variable with given initial value BASE and step STEP |
1181 | for loop LOOP. NO_OVERFLOW implies the iv doesn't overflow. */ |
1182 | |
1183 | static struct iv * |
1184 | alloc_iv (struct ivopts_data *data, tree base, tree step, |
1185 | bool no_overflow = false) |
1186 | { |
1187 | tree expr = base; |
1188 | struct iv *iv = (struct iv*) obstack_alloc (&data->iv_obstack, |
1189 | sizeof (struct iv)); |
1190 | gcc_assert (step != NULL_TREE); |
1191 | |
1192 | /* Lower address expression in base except ones with DECL_P as operand. |
1193 | By doing this: |
1194 | 1) More accurate cost can be computed for address expressions; |
1195 | 2) Duplicate candidates won't be created for bases in different |
1196 | forms, like &a[0] and &a. */ |
1197 | STRIP_NOPS (expr); |
1198 | if ((TREE_CODE (expr) == ADDR_EXPR && !DECL_P (TREE_OPERAND (expr, 0))) |
1199 | || contain_complex_addr_expr (expr)) |
1200 | { |
1201 | aff_tree comb; |
1202 | tree_to_aff_combination (expr, TREE_TYPE (expr), &comb); |
1203 | base = fold_convert (TREE_TYPE (base), aff_combination_to_tree (&comb)); |
1204 | } |
1205 | |
1206 | iv->base = base; |
1207 | iv->base_object = determine_base_object (data, expr: base); |
1208 | iv->step = step; |
1209 | iv->biv_p = false; |
1210 | iv->nonlin_use = NULL; |
1211 | iv->ssa_name = NULL_TREE; |
1212 | if (!no_overflow |
1213 | && !iv_can_overflow_p (data->current_loop, TREE_TYPE (base), |
1214 | base, step)) |
1215 | no_overflow = true; |
1216 | iv->no_overflow = no_overflow; |
1217 | iv->have_address_use = false; |
1218 | |
1219 | return iv; |
1220 | } |
1221 | |
1222 | /* Sets STEP and BASE for induction variable IV. NO_OVERFLOW implies the IV |
1223 | doesn't overflow. */ |
1224 | |
1225 | static void |
1226 | set_iv (struct ivopts_data *data, tree iv, tree base, tree step, |
1227 | bool no_overflow) |
1228 | { |
1229 | struct version_info *info = name_info (data, name: iv); |
1230 | |
1231 | gcc_assert (!info->iv); |
1232 | |
1233 | bitmap_set_bit (data->relevant, SSA_NAME_VERSION (iv)); |
1234 | info->iv = alloc_iv (data, base, step, no_overflow); |
1235 | info->iv->ssa_name = iv; |
1236 | } |
1237 | |
1238 | /* Finds induction variable declaration for VAR. */ |
1239 | |
1240 | static struct iv * |
1241 | get_iv (struct ivopts_data *data, tree var) |
1242 | { |
1243 | basic_block bb; |
1244 | tree type = TREE_TYPE (var); |
1245 | |
1246 | if (!POINTER_TYPE_P (type) |
1247 | && !INTEGRAL_TYPE_P (type)) |
1248 | return NULL; |
1249 | |
1250 | if (!name_info (data, name: var)->iv) |
1251 | { |
1252 | bb = gimple_bb (SSA_NAME_DEF_STMT (var)); |
1253 | |
1254 | if (!bb |
1255 | || !flow_bb_inside_loop_p (data->current_loop, bb)) |
1256 | { |
1257 | if (POINTER_TYPE_P (type)) |
1258 | type = sizetype; |
1259 | set_iv (data, iv: var, base: var, step: build_int_cst (type, 0), no_overflow: true); |
1260 | } |
1261 | } |
1262 | |
1263 | return name_info (data, name: var)->iv; |
1264 | } |
1265 | |
1266 | /* Return the first non-invariant ssa var found in EXPR. */ |
1267 | |
1268 | static tree |
1269 | (tree expr) |
1270 | { |
1271 | int i, n; |
1272 | tree tmp; |
1273 | enum tree_code code; |
1274 | |
1275 | if (!expr || is_gimple_min_invariant (expr)) |
1276 | return NULL; |
1277 | |
1278 | code = TREE_CODE (expr); |
1279 | if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code))) |
1280 | { |
1281 | n = TREE_OPERAND_LENGTH (expr); |
1282 | for (i = 0; i < n; i++) |
1283 | { |
1284 | tmp = extract_single_var_from_expr (TREE_OPERAND (expr, i)); |
1285 | |
1286 | if (tmp) |
1287 | return tmp; |
1288 | } |
1289 | } |
1290 | return (TREE_CODE (expr) == SSA_NAME) ? expr : NULL; |
1291 | } |
1292 | |
1293 | /* Finds basic ivs. */ |
1294 | |
1295 | static bool |
1296 | find_bivs (struct ivopts_data *data) |
1297 | { |
1298 | gphi *phi; |
1299 | affine_iv iv; |
1300 | tree step, type, base, stop; |
1301 | bool found = false; |
1302 | class loop *loop = data->current_loop; |
1303 | gphi_iterator psi; |
1304 | |
1305 | for (psi = gsi_start_phis (loop->header); !gsi_end_p (i: psi); gsi_next (i: &psi)) |
1306 | { |
1307 | phi = psi.phi (); |
1308 | |
1309 | if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (PHI_RESULT (phi))) |
1310 | continue; |
1311 | |
1312 | if (virtual_operand_p (PHI_RESULT (phi))) |
1313 | continue; |
1314 | |
1315 | if (!simple_iv (loop, loop, PHI_RESULT (phi), &iv, true)) |
1316 | continue; |
1317 | |
1318 | if (integer_zerop (iv.step)) |
1319 | continue; |
1320 | |
1321 | step = iv.step; |
1322 | base = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); |
1323 | /* Stop expanding iv base at the first ssa var referred by iv step. |
1324 | Ideally we should stop at any ssa var, because that's expensive |
1325 | and unusual to happen, we just do it on the first one. |
1326 | |
1327 | See PR64705 for the rationale. */ |
1328 | stop = extract_single_var_from_expr (expr: step); |
1329 | base = expand_simple_operations (base, stop); |
1330 | if (contains_abnormal_ssa_name_p (expr: base) |
1331 | || contains_abnormal_ssa_name_p (expr: step)) |
1332 | continue; |
1333 | |
1334 | type = TREE_TYPE (PHI_RESULT (phi)); |
1335 | base = fold_convert (type, base); |
1336 | if (step) |
1337 | { |
1338 | if (POINTER_TYPE_P (type)) |
1339 | step = convert_to_ptrofftype (step); |
1340 | else |
1341 | step = fold_convert (type, step); |
1342 | } |
1343 | |
1344 | set_iv (data, PHI_RESULT (phi), base, step, no_overflow: iv.no_overflow); |
1345 | found = true; |
1346 | } |
1347 | |
1348 | return found; |
1349 | } |
1350 | |
1351 | /* Marks basic ivs. */ |
1352 | |
1353 | static void |
1354 | mark_bivs (struct ivopts_data *data) |
1355 | { |
1356 | gphi *phi; |
1357 | gimple *def; |
1358 | tree var; |
1359 | struct iv *iv, *incr_iv; |
1360 | class loop *loop = data->current_loop; |
1361 | basic_block incr_bb; |
1362 | gphi_iterator psi; |
1363 | |
1364 | data->bivs_not_used_in_addr = 0; |
1365 | for (psi = gsi_start_phis (loop->header); !gsi_end_p (i: psi); gsi_next (i: &psi)) |
1366 | { |
1367 | phi = psi.phi (); |
1368 | |
1369 | iv = get_iv (data, PHI_RESULT (phi)); |
1370 | if (!iv) |
1371 | continue; |
1372 | |
1373 | var = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); |
1374 | def = SSA_NAME_DEF_STMT (var); |
1375 | /* Don't mark iv peeled from other one as biv. */ |
1376 | if (def |
1377 | && gimple_code (g: def) == GIMPLE_PHI |
1378 | && gimple_bb (g: def) == loop->header) |
1379 | continue; |
1380 | |
1381 | incr_iv = get_iv (data, var); |
1382 | if (!incr_iv) |
1383 | continue; |
1384 | |
1385 | /* If the increment is in the subloop, ignore it. */ |
1386 | incr_bb = gimple_bb (SSA_NAME_DEF_STMT (var)); |
1387 | if (incr_bb->loop_father != data->current_loop |
1388 | || (incr_bb->flags & BB_IRREDUCIBLE_LOOP)) |
1389 | continue; |
1390 | |
1391 | iv->biv_p = true; |
1392 | incr_iv->biv_p = true; |
1393 | if (iv->no_overflow) |
1394 | data->bivs_not_used_in_addr++; |
1395 | if (incr_iv->no_overflow) |
1396 | data->bivs_not_used_in_addr++; |
1397 | } |
1398 | } |
1399 | |
1400 | /* Checks whether STMT defines a linear induction variable and stores its |
1401 | parameters to IV. */ |
1402 | |
1403 | static bool |
1404 | find_givs_in_stmt_scev (struct ivopts_data *data, gimple *stmt, affine_iv *iv) |
1405 | { |
1406 | tree lhs, stop; |
1407 | class loop *loop = data->current_loop; |
1408 | |
1409 | iv->base = NULL_TREE; |
1410 | iv->step = NULL_TREE; |
1411 | |
1412 | if (gimple_code (g: stmt) != GIMPLE_ASSIGN) |
1413 | return false; |
1414 | |
1415 | lhs = gimple_assign_lhs (gs: stmt); |
1416 | if (TREE_CODE (lhs) != SSA_NAME) |
1417 | return false; |
1418 | |
1419 | if (!simple_iv (loop, loop_containing_stmt (stmt), lhs, iv, true)) |
1420 | return false; |
1421 | |
1422 | /* Stop expanding iv base at the first ssa var referred by iv step. |
1423 | Ideally we should stop at any ssa var, because that's expensive |
1424 | and unusual to happen, we just do it on the first one. |
1425 | |
1426 | See PR64705 for the rationale. */ |
1427 | stop = extract_single_var_from_expr (expr: iv->step); |
1428 | iv->base = expand_simple_operations (iv->base, stop); |
1429 | if (contains_abnormal_ssa_name_p (expr: iv->base) |
1430 | || contains_abnormal_ssa_name_p (expr: iv->step)) |
1431 | return false; |
1432 | |
1433 | /* If STMT could throw, then do not consider STMT as defining a GIV. |
1434 | While this will suppress optimizations, we cannot safely delete this |
1435 | GIV and associated statements, even if it appears it is not used. */ |
1436 | if (stmt_could_throw_p (cfun, stmt)) |
1437 | return false; |
1438 | |
1439 | return true; |
1440 | } |
1441 | |
1442 | /* Finds general ivs in statement STMT. */ |
1443 | |
1444 | static void |
1445 | find_givs_in_stmt (struct ivopts_data *data, gimple *stmt) |
1446 | { |
1447 | affine_iv iv; |
1448 | |
1449 | if (!find_givs_in_stmt_scev (data, stmt, iv: &iv)) |
1450 | return; |
1451 | |
1452 | set_iv (data, iv: gimple_assign_lhs (gs: stmt), base: iv.base, step: iv.step, no_overflow: iv.no_overflow); |
1453 | } |
1454 | |
1455 | /* Finds general ivs in basic block BB. */ |
1456 | |
1457 | static void |
1458 | find_givs_in_bb (struct ivopts_data *data, basic_block bb) |
1459 | { |
1460 | gimple_stmt_iterator bsi; |
1461 | |
1462 | for (bsi = gsi_start_bb (bb); !gsi_end_p (i: bsi); gsi_next (i: &bsi)) |
1463 | find_givs_in_stmt (data, stmt: gsi_stmt (i: bsi)); |
1464 | } |
1465 | |
1466 | /* Finds general ivs. */ |
1467 | |
1468 | static void |
1469 | find_givs (struct ivopts_data *data, basic_block *body) |
1470 | { |
1471 | class loop *loop = data->current_loop; |
1472 | unsigned i; |
1473 | |
1474 | for (i = 0; i < loop->num_nodes; i++) |
1475 | find_givs_in_bb (data, bb: body[i]); |
1476 | } |
1477 | |
1478 | /* For each ssa name defined in LOOP determines whether it is an induction |
1479 | variable and if so, its initial value and step. */ |
1480 | |
1481 | static bool |
1482 | find_induction_variables (struct ivopts_data *data, basic_block *body) |
1483 | { |
1484 | unsigned i; |
1485 | bitmap_iterator bi; |
1486 | |
1487 | if (!find_bivs (data)) |
1488 | return false; |
1489 | |
1490 | find_givs (data, body); |
1491 | mark_bivs (data); |
1492 | |
1493 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1494 | { |
1495 | class tree_niter_desc *niter = niter_for_single_dom_exit (data); |
1496 | |
1497 | if (niter) |
1498 | { |
1499 | fprintf (stream: dump_file, format: " number of iterations " ); |
1500 | print_generic_expr (dump_file, niter->niter, TDF_SLIM); |
1501 | if (!integer_zerop (niter->may_be_zero)) |
1502 | { |
1503 | fprintf (stream: dump_file, format: "; zero if " ); |
1504 | print_generic_expr (dump_file, niter->may_be_zero, TDF_SLIM); |
1505 | } |
1506 | fprintf (stream: dump_file, format: "\n" ); |
1507 | }; |
1508 | |
1509 | fprintf (stream: dump_file, format: "\n<Induction Vars>:\n" ); |
1510 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, i, bi) |
1511 | { |
1512 | struct version_info *info = ver_info (data, ver: i); |
1513 | if (info->iv && info->iv->step && !integer_zerop (info->iv->step)) |
1514 | dump_iv (file: dump_file, iv: ver_info (data, ver: i)->iv, dump_name: true, indent_level: 0); |
1515 | } |
1516 | } |
1517 | |
1518 | return true; |
1519 | } |
1520 | |
1521 | /* Records a use of TYPE at *USE_P in STMT whose value is IV in GROUP. |
1522 | For address type use, ADDR_BASE is the stripped IV base, ADDR_OFFSET |
1523 | is the const offset stripped from IV base and MEM_TYPE is the type |
1524 | of the memory being addressed. For uses of other types, ADDR_BASE |
1525 | and ADDR_OFFSET are zero by default and MEM_TYPE is NULL_TREE. */ |
1526 | |
1527 | static struct iv_use * |
1528 | record_use (struct iv_group *group, tree *use_p, struct iv *iv, |
1529 | gimple *stmt, enum use_type type, tree mem_type, |
1530 | tree addr_base, poly_uint64 addr_offset) |
1531 | { |
1532 | struct iv_use *use = XCNEW (struct iv_use); |
1533 | |
1534 | use->id = group->vuses.length (); |
1535 | use->group_id = group->id; |
1536 | use->type = type; |
1537 | use->mem_type = mem_type; |
1538 | use->iv = iv; |
1539 | use->stmt = stmt; |
1540 | use->op_p = use_p; |
1541 | use->addr_base = addr_base; |
1542 | use->addr_offset = addr_offset; |
1543 | |
1544 | group->vuses.safe_push (obj: use); |
1545 | return use; |
1546 | } |
1547 | |
1548 | /* Checks whether OP is a loop-level invariant and if so, records it. |
1549 | NONLINEAR_USE is true if the invariant is used in a way we do not |
1550 | handle specially. */ |
1551 | |
1552 | static void |
1553 | record_invariant (struct ivopts_data *data, tree op, bool nonlinear_use) |
1554 | { |
1555 | basic_block bb; |
1556 | struct version_info *info; |
1557 | |
1558 | if (TREE_CODE (op) != SSA_NAME |
1559 | || virtual_operand_p (op)) |
1560 | return; |
1561 | |
1562 | bb = gimple_bb (SSA_NAME_DEF_STMT (op)); |
1563 | if (bb |
1564 | && flow_bb_inside_loop_p (data->current_loop, bb)) |
1565 | return; |
1566 | |
1567 | info = name_info (data, name: op); |
1568 | info->name = op; |
1569 | info->has_nonlin_use |= nonlinear_use; |
1570 | if (!info->inv_id) |
1571 | info->inv_id = ++data->max_inv_var_id; |
1572 | bitmap_set_bit (data->relevant, SSA_NAME_VERSION (op)); |
1573 | } |
1574 | |
1575 | /* Record a group of TYPE. */ |
1576 | |
1577 | static struct iv_group * |
1578 | record_group (struct ivopts_data *data, enum use_type type) |
1579 | { |
1580 | struct iv_group *group = XCNEW (struct iv_group); |
1581 | |
1582 | group->id = data->vgroups.length (); |
1583 | group->type = type; |
1584 | group->related_cands = BITMAP_ALLOC (NULL); |
1585 | group->vuses.create (nelems: 1); |
1586 | group->doloop_p = false; |
1587 | |
1588 | data->vgroups.safe_push (obj: group); |
1589 | return group; |
1590 | } |
1591 | |
1592 | /* Record a use of TYPE at *USE_P in STMT whose value is IV in a group. |
1593 | New group will be created if there is no existing group for the use. |
1594 | MEM_TYPE is the type of memory being addressed, or NULL if this |
1595 | isn't an address reference. */ |
1596 | |
1597 | static struct iv_use * |
1598 | record_group_use (struct ivopts_data *data, tree *use_p, |
1599 | struct iv *iv, gimple *stmt, enum use_type type, |
1600 | tree mem_type) |
1601 | { |
1602 | tree addr_base = NULL; |
1603 | struct iv_group *group = NULL; |
1604 | poly_uint64 addr_offset = 0; |
1605 | |
1606 | /* Record non address type use in a new group. */ |
1607 | if (address_p (type)) |
1608 | { |
1609 | unsigned int i; |
1610 | |
1611 | gcc_assert (POINTER_TYPE_P (TREE_TYPE (iv->base))); |
1612 | tree addr_toffset; |
1613 | split_constant_offset (iv->base, &addr_base, &addr_toffset); |
1614 | addr_offset = int_cst_value (addr_toffset); |
1615 | for (i = 0; i < data->vgroups.length (); i++) |
1616 | { |
1617 | struct iv_use *use; |
1618 | |
1619 | group = data->vgroups[i]; |
1620 | use = group->vuses[0]; |
1621 | if (!address_p (type: use->type)) |
1622 | continue; |
1623 | |
1624 | /* Check if it has the same stripped base and step. */ |
1625 | if (operand_equal_p (iv->base_object, use->iv->base_object, flags: 0) |
1626 | && operand_equal_p (iv->step, use->iv->step, flags: 0) |
1627 | && operand_equal_p (addr_base, use->addr_base, flags: 0)) |
1628 | break; |
1629 | } |
1630 | if (i == data->vgroups.length ()) |
1631 | group = NULL; |
1632 | } |
1633 | |
1634 | if (!group) |
1635 | group = record_group (data, type); |
1636 | |
1637 | return record_use (group, use_p, iv, stmt, type, mem_type, |
1638 | addr_base, addr_offset); |
1639 | } |
1640 | |
1641 | /* Checks whether the use OP is interesting and if so, records it. */ |
1642 | |
1643 | static struct iv_use * |
1644 | find_interesting_uses_op (struct ivopts_data *data, tree op) |
1645 | { |
1646 | struct iv *iv; |
1647 | gimple *stmt; |
1648 | struct iv_use *use; |
1649 | |
1650 | if (TREE_CODE (op) != SSA_NAME) |
1651 | return NULL; |
1652 | |
1653 | iv = get_iv (data, var: op); |
1654 | if (!iv) |
1655 | return NULL; |
1656 | |
1657 | if (iv->nonlin_use) |
1658 | { |
1659 | gcc_assert (iv->nonlin_use->type == USE_NONLINEAR_EXPR); |
1660 | return iv->nonlin_use; |
1661 | } |
1662 | |
1663 | if (integer_zerop (iv->step)) |
1664 | { |
1665 | record_invariant (data, op, nonlinear_use: true); |
1666 | return NULL; |
1667 | } |
1668 | |
1669 | stmt = SSA_NAME_DEF_STMT (op); |
1670 | gcc_assert (gimple_code (stmt) == GIMPLE_PHI || is_gimple_assign (stmt)); |
1671 | |
1672 | use = record_group_use (data, NULL, iv, stmt, type: USE_NONLINEAR_EXPR, NULL_TREE); |
1673 | iv->nonlin_use = use; |
1674 | return use; |
1675 | } |
1676 | |
1677 | /* Indicate how compare type iv_use can be handled. */ |
1678 | enum comp_iv_rewrite |
1679 | { |
1680 | COMP_IV_NA, |
1681 | /* We may rewrite compare type iv_use by expressing value of the iv_use. */ |
1682 | COMP_IV_EXPR, |
1683 | /* We may rewrite compare type iv_uses on both sides of comparison by |
1684 | expressing value of each iv_use. */ |
1685 | COMP_IV_EXPR_2, |
1686 | /* We may rewrite compare type iv_use by expressing value of the iv_use |
1687 | or by eliminating it with other iv_cand. */ |
1688 | COMP_IV_ELIM |
1689 | }; |
1690 | |
1691 | /* Given a condition in statement STMT, checks whether it is a compare |
1692 | of an induction variable and an invariant. If this is the case, |
1693 | CONTROL_VAR is set to location of the iv, BOUND to the location of |
1694 | the invariant, IV_VAR and IV_BOUND are set to the corresponding |
1695 | induction variable descriptions, and true is returned. If this is not |
1696 | the case, CONTROL_VAR and BOUND are set to the arguments of the |
1697 | condition and false is returned. */ |
1698 | |
1699 | static enum comp_iv_rewrite |
1700 | extract_cond_operands (struct ivopts_data *data, gimple *stmt, |
1701 | tree **control_var, tree **bound, |
1702 | struct iv **iv_var, struct iv **iv_bound) |
1703 | { |
1704 | /* The objects returned when COND has constant operands. */ |
1705 | static struct iv const_iv; |
1706 | static tree zero; |
1707 | tree *op0 = &zero, *op1 = &zero; |
1708 | struct iv *iv0 = &const_iv, *iv1 = &const_iv; |
1709 | enum comp_iv_rewrite rewrite_type = COMP_IV_NA; |
1710 | |
1711 | if (gimple_code (g: stmt) == GIMPLE_COND) |
1712 | { |
1713 | gcond *cond_stmt = as_a <gcond *> (p: stmt); |
1714 | op0 = gimple_cond_lhs_ptr (gs: cond_stmt); |
1715 | op1 = gimple_cond_rhs_ptr (gs: cond_stmt); |
1716 | } |
1717 | else |
1718 | { |
1719 | op0 = gimple_assign_rhs1_ptr (gs: stmt); |
1720 | op1 = gimple_assign_rhs2_ptr (gs: stmt); |
1721 | } |
1722 | |
1723 | zero = integer_zero_node; |
1724 | const_iv.step = integer_zero_node; |
1725 | |
1726 | if (TREE_CODE (*op0) == SSA_NAME) |
1727 | iv0 = get_iv (data, var: *op0); |
1728 | if (TREE_CODE (*op1) == SSA_NAME) |
1729 | iv1 = get_iv (data, var: *op1); |
1730 | |
1731 | /* If both sides of comparison are IVs. We can express ivs on both end. */ |
1732 | if (iv0 && iv1 && !integer_zerop (iv0->step) && !integer_zerop (iv1->step)) |
1733 | { |
1734 | rewrite_type = COMP_IV_EXPR_2; |
1735 | goto end; |
1736 | } |
1737 | |
1738 | /* If none side of comparison is IV. */ |
1739 | if ((!iv0 || integer_zerop (iv0->step)) |
1740 | && (!iv1 || integer_zerop (iv1->step))) |
1741 | goto end; |
1742 | |
1743 | /* Control variable may be on the other side. */ |
1744 | if (!iv0 || integer_zerop (iv0->step)) |
1745 | { |
1746 | std::swap (a&: op0, b&: op1); |
1747 | std::swap (a&: iv0, b&: iv1); |
1748 | } |
1749 | /* If one side is IV and the other side isn't loop invariant. */ |
1750 | if (!iv1) |
1751 | rewrite_type = COMP_IV_EXPR; |
1752 | /* If one side is IV and the other side is loop invariant. */ |
1753 | else if (!integer_zerop (iv0->step) && integer_zerop (iv1->step)) |
1754 | rewrite_type = COMP_IV_ELIM; |
1755 | |
1756 | end: |
1757 | if (control_var) |
1758 | *control_var = op0; |
1759 | if (iv_var) |
1760 | *iv_var = iv0; |
1761 | if (bound) |
1762 | *bound = op1; |
1763 | if (iv_bound) |
1764 | *iv_bound = iv1; |
1765 | |
1766 | return rewrite_type; |
1767 | } |
1768 | |
1769 | /* Checks whether the condition in STMT is interesting and if so, |
1770 | records it. */ |
1771 | |
1772 | static void |
1773 | find_interesting_uses_cond (struct ivopts_data *data, gimple *stmt) |
1774 | { |
1775 | tree *var_p, *bound_p; |
1776 | struct iv *var_iv, *bound_iv; |
1777 | enum comp_iv_rewrite ret; |
1778 | |
1779 | ret = extract_cond_operands (data, stmt, |
1780 | control_var: &var_p, bound: &bound_p, iv_var: &var_iv, iv_bound: &bound_iv); |
1781 | if (ret == COMP_IV_NA) |
1782 | { |
1783 | find_interesting_uses_op (data, op: *var_p); |
1784 | find_interesting_uses_op (data, op: *bound_p); |
1785 | return; |
1786 | } |
1787 | |
1788 | record_group_use (data, use_p: var_p, iv: var_iv, stmt, type: USE_COMPARE, NULL_TREE); |
1789 | /* Record compare type iv_use for iv on the other side of comparison. */ |
1790 | if (ret == COMP_IV_EXPR_2) |
1791 | record_group_use (data, use_p: bound_p, iv: bound_iv, stmt, type: USE_COMPARE, NULL_TREE); |
1792 | } |
1793 | |
1794 | /* Returns the outermost loop EXPR is obviously invariant in |
1795 | relative to the loop LOOP, i.e. if all its operands are defined |
1796 | outside of the returned loop. Returns NULL if EXPR is not |
1797 | even obviously invariant in LOOP. */ |
1798 | |
1799 | class loop * |
1800 | outermost_invariant_loop_for_expr (class loop *loop, tree expr) |
1801 | { |
1802 | basic_block def_bb; |
1803 | unsigned i, len; |
1804 | |
1805 | if (is_gimple_min_invariant (expr)) |
1806 | return current_loops->tree_root; |
1807 | |
1808 | if (TREE_CODE (expr) == SSA_NAME) |
1809 | { |
1810 | def_bb = gimple_bb (SSA_NAME_DEF_STMT (expr)); |
1811 | if (def_bb) |
1812 | { |
1813 | if (flow_bb_inside_loop_p (loop, def_bb)) |
1814 | return NULL; |
1815 | return superloop_at_depth (loop, |
1816 | loop_depth (loop: def_bb->loop_father) + 1); |
1817 | } |
1818 | |
1819 | return current_loops->tree_root; |
1820 | } |
1821 | |
1822 | if (!EXPR_P (expr)) |
1823 | return NULL; |
1824 | |
1825 | unsigned maxdepth = 0; |
1826 | len = TREE_OPERAND_LENGTH (expr); |
1827 | for (i = 0; i < len; i++) |
1828 | { |
1829 | class loop *ivloop; |
1830 | if (!TREE_OPERAND (expr, i)) |
1831 | continue; |
1832 | |
1833 | ivloop = outermost_invariant_loop_for_expr (loop, TREE_OPERAND (expr, i)); |
1834 | if (!ivloop) |
1835 | return NULL; |
1836 | maxdepth = MAX (maxdepth, loop_depth (ivloop)); |
1837 | } |
1838 | |
1839 | return superloop_at_depth (loop, maxdepth); |
1840 | } |
1841 | |
1842 | /* Returns true if expression EXPR is obviously invariant in LOOP, |
1843 | i.e. if all its operands are defined outside of the LOOP. LOOP |
1844 | should not be the function body. */ |
1845 | |
1846 | bool |
1847 | expr_invariant_in_loop_p (class loop *loop, tree expr) |
1848 | { |
1849 | basic_block def_bb; |
1850 | unsigned i, len; |
1851 | |
1852 | gcc_assert (loop_depth (loop) > 0); |
1853 | |
1854 | if (is_gimple_min_invariant (expr)) |
1855 | return true; |
1856 | |
1857 | if (TREE_CODE (expr) == SSA_NAME) |
1858 | { |
1859 | def_bb = gimple_bb (SSA_NAME_DEF_STMT (expr)); |
1860 | if (def_bb |
1861 | && flow_bb_inside_loop_p (loop, def_bb)) |
1862 | return false; |
1863 | |
1864 | return true; |
1865 | } |
1866 | |
1867 | if (!EXPR_P (expr)) |
1868 | return false; |
1869 | |
1870 | len = TREE_OPERAND_LENGTH (expr); |
1871 | for (i = 0; i < len; i++) |
1872 | if (TREE_OPERAND (expr, i) |
1873 | && !expr_invariant_in_loop_p (loop, TREE_OPERAND (expr, i))) |
1874 | return false; |
1875 | |
1876 | return true; |
1877 | } |
1878 | |
1879 | /* Given expression EXPR which computes inductive values with respect |
1880 | to loop recorded in DATA, this function returns biv from which EXPR |
1881 | is derived by tracing definition chains of ssa variables in EXPR. */ |
1882 | |
1883 | static struct iv* |
1884 | find_deriving_biv_for_expr (struct ivopts_data *data, tree expr) |
1885 | { |
1886 | struct iv *iv; |
1887 | unsigned i, n; |
1888 | tree e2, e1; |
1889 | enum tree_code code; |
1890 | gimple *stmt; |
1891 | |
1892 | if (expr == NULL_TREE) |
1893 | return NULL; |
1894 | |
1895 | if (is_gimple_min_invariant (expr)) |
1896 | return NULL; |
1897 | |
1898 | code = TREE_CODE (expr); |
1899 | if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code))) |
1900 | { |
1901 | n = TREE_OPERAND_LENGTH (expr); |
1902 | for (i = 0; i < n; i++) |
1903 | { |
1904 | iv = find_deriving_biv_for_expr (data, TREE_OPERAND (expr, i)); |
1905 | if (iv) |
1906 | return iv; |
1907 | } |
1908 | } |
1909 | |
1910 | /* Stop if it's not ssa name. */ |
1911 | if (code != SSA_NAME) |
1912 | return NULL; |
1913 | |
1914 | iv = get_iv (data, var: expr); |
1915 | if (!iv || integer_zerop (iv->step)) |
1916 | return NULL; |
1917 | else if (iv->biv_p) |
1918 | return iv; |
1919 | |
1920 | stmt = SSA_NAME_DEF_STMT (expr); |
1921 | if (gphi *phi = dyn_cast <gphi *> (p: stmt)) |
1922 | { |
1923 | ssa_op_iter iter; |
1924 | use_operand_p use_p; |
1925 | basic_block phi_bb = gimple_bb (g: phi); |
1926 | |
1927 | /* Skip loop header PHI that doesn't define biv. */ |
1928 | if (phi_bb->loop_father == data->current_loop) |
1929 | return NULL; |
1930 | |
1931 | if (virtual_operand_p (op: gimple_phi_result (gs: phi))) |
1932 | return NULL; |
1933 | |
1934 | FOR_EACH_PHI_ARG (use_p, phi, iter, SSA_OP_USE) |
1935 | { |
1936 | tree use = USE_FROM_PTR (use_p); |
1937 | iv = find_deriving_biv_for_expr (data, expr: use); |
1938 | if (iv) |
1939 | return iv; |
1940 | } |
1941 | return NULL; |
1942 | } |
1943 | if (gimple_code (g: stmt) != GIMPLE_ASSIGN) |
1944 | return NULL; |
1945 | |
1946 | e1 = gimple_assign_rhs1 (gs: stmt); |
1947 | code = gimple_assign_rhs_code (gs: stmt); |
1948 | if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS) |
1949 | return find_deriving_biv_for_expr (data, expr: e1); |
1950 | |
1951 | switch (code) |
1952 | { |
1953 | case MULT_EXPR: |
1954 | case PLUS_EXPR: |
1955 | case MINUS_EXPR: |
1956 | case POINTER_PLUS_EXPR: |
1957 | /* Increments, decrements and multiplications by a constant |
1958 | are simple. */ |
1959 | e2 = gimple_assign_rhs2 (gs: stmt); |
1960 | iv = find_deriving_biv_for_expr (data, expr: e2); |
1961 | if (iv) |
1962 | return iv; |
1963 | gcc_fallthrough (); |
1964 | |
1965 | CASE_CONVERT: |
1966 | /* Casts are simple. */ |
1967 | return find_deriving_biv_for_expr (data, expr: e1); |
1968 | |
1969 | default: |
1970 | break; |
1971 | } |
1972 | |
1973 | return NULL; |
1974 | } |
1975 | |
1976 | /* Record BIV, its predecessor and successor that they are used in |
1977 | address type uses. */ |
1978 | |
1979 | static void |
1980 | record_biv_for_address_use (struct ivopts_data *data, struct iv *biv) |
1981 | { |
1982 | unsigned i; |
1983 | tree type, base_1, base_2; |
1984 | bitmap_iterator bi; |
1985 | |
1986 | if (!biv || !biv->biv_p || integer_zerop (biv->step) |
1987 | || biv->have_address_use || !biv->no_overflow) |
1988 | return; |
1989 | |
1990 | type = TREE_TYPE (biv->base); |
1991 | if (!INTEGRAL_TYPE_P (type)) |
1992 | return; |
1993 | |
1994 | biv->have_address_use = true; |
1995 | data->bivs_not_used_in_addr--; |
1996 | base_1 = fold_build2 (PLUS_EXPR, type, biv->base, biv->step); |
1997 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, i, bi) |
1998 | { |
1999 | struct iv *iv = ver_info (data, ver: i)->iv; |
2000 | |
2001 | if (!iv || !iv->biv_p || integer_zerop (iv->step) |
2002 | || iv->have_address_use || !iv->no_overflow) |
2003 | continue; |
2004 | |
2005 | if (type != TREE_TYPE (iv->base) |
2006 | || !INTEGRAL_TYPE_P (TREE_TYPE (iv->base))) |
2007 | continue; |
2008 | |
2009 | if (!operand_equal_p (biv->step, iv->step, flags: 0)) |
2010 | continue; |
2011 | |
2012 | base_2 = fold_build2 (PLUS_EXPR, type, iv->base, iv->step); |
2013 | if (operand_equal_p (base_1, iv->base, flags: 0) |
2014 | || operand_equal_p (base_2, biv->base, flags: 0)) |
2015 | { |
2016 | iv->have_address_use = true; |
2017 | data->bivs_not_used_in_addr--; |
2018 | } |
2019 | } |
2020 | } |
2021 | |
2022 | /* Cumulates the steps of indices into DATA and replaces their values with the |
2023 | initial ones. Returns false when the value of the index cannot be determined. |
2024 | Callback for for_each_index. */ |
2025 | |
2026 | struct ifs_ivopts_data |
2027 | { |
2028 | struct ivopts_data *ivopts_data; |
2029 | gimple *stmt; |
2030 | tree step; |
2031 | }; |
2032 | |
2033 | static bool |
2034 | idx_find_step (tree base, tree *idx, void *data) |
2035 | { |
2036 | struct ifs_ivopts_data *dta = (struct ifs_ivopts_data *) data; |
2037 | struct iv *iv; |
2038 | bool use_overflow_semantics = false; |
2039 | tree step, iv_base, iv_step, lbound, off; |
2040 | class loop *loop = dta->ivopts_data->current_loop; |
2041 | |
2042 | /* If base is a component ref, require that the offset of the reference |
2043 | be invariant. */ |
2044 | if (TREE_CODE (base) == COMPONENT_REF) |
2045 | { |
2046 | off = component_ref_field_offset (base); |
2047 | return expr_invariant_in_loop_p (loop, expr: off); |
2048 | } |
2049 | |
2050 | /* If base is array, first check whether we will be able to move the |
2051 | reference out of the loop (in order to take its address in strength |
2052 | reduction). In order for this to work we need both lower bound |
2053 | and step to be loop invariants. */ |
2054 | if (TREE_CODE (base) == ARRAY_REF || TREE_CODE (base) == ARRAY_RANGE_REF) |
2055 | { |
2056 | /* Moreover, for a range, the size needs to be invariant as well. */ |
2057 | if (TREE_CODE (base) == ARRAY_RANGE_REF |
2058 | && !expr_invariant_in_loop_p (loop, TYPE_SIZE (TREE_TYPE (base)))) |
2059 | return false; |
2060 | |
2061 | step = array_ref_element_size (base); |
2062 | lbound = array_ref_low_bound (base); |
2063 | |
2064 | if (!expr_invariant_in_loop_p (loop, expr: step) |
2065 | || !expr_invariant_in_loop_p (loop, expr: lbound)) |
2066 | return false; |
2067 | } |
2068 | |
2069 | if (TREE_CODE (*idx) != SSA_NAME) |
2070 | return true; |
2071 | |
2072 | iv = get_iv (data: dta->ivopts_data, var: *idx); |
2073 | if (!iv) |
2074 | return false; |
2075 | |
2076 | /* XXX We produce for a base of *D42 with iv->base being &x[0] |
2077 | *&x[0], which is not folded and does not trigger the |
2078 | ARRAY_REF path below. */ |
2079 | *idx = iv->base; |
2080 | |
2081 | if (integer_zerop (iv->step)) |
2082 | return true; |
2083 | |
2084 | if (TREE_CODE (base) == ARRAY_REF || TREE_CODE (base) == ARRAY_RANGE_REF) |
2085 | { |
2086 | step = array_ref_element_size (base); |
2087 | |
2088 | /* We only handle addresses whose step is an integer constant. */ |
2089 | if (TREE_CODE (step) != INTEGER_CST) |
2090 | return false; |
2091 | } |
2092 | else |
2093 | /* The step for pointer arithmetics already is 1 byte. */ |
2094 | step = size_one_node; |
2095 | |
2096 | iv_base = iv->base; |
2097 | iv_step = iv->step; |
2098 | if (iv->no_overflow && nowrap_type_p (TREE_TYPE (iv_step))) |
2099 | use_overflow_semantics = true; |
2100 | |
2101 | if (!convert_affine_scev (dta->ivopts_data->current_loop, |
2102 | sizetype, &iv_base, &iv_step, dta->stmt, |
2103 | use_overflow_semantics)) |
2104 | { |
2105 | /* The index might wrap. */ |
2106 | return false; |
2107 | } |
2108 | |
2109 | step = fold_build2 (MULT_EXPR, sizetype, step, iv_step); |
2110 | dta->step = fold_build2 (PLUS_EXPR, sizetype, dta->step, step); |
2111 | |
2112 | if (dta->ivopts_data->bivs_not_used_in_addr) |
2113 | { |
2114 | if (!iv->biv_p) |
2115 | iv = find_deriving_biv_for_expr (data: dta->ivopts_data, expr: iv->ssa_name); |
2116 | |
2117 | record_biv_for_address_use (data: dta->ivopts_data, biv: iv); |
2118 | } |
2119 | return true; |
2120 | } |
2121 | |
2122 | /* Records use in index IDX. Callback for for_each_index. Ivopts data |
2123 | object is passed to it in DATA. */ |
2124 | |
2125 | static bool |
2126 | idx_record_use (tree base, tree *idx, |
2127 | void *vdata) |
2128 | { |
2129 | struct ivopts_data *data = (struct ivopts_data *) vdata; |
2130 | find_interesting_uses_op (data, op: *idx); |
2131 | if (TREE_CODE (base) == ARRAY_REF || TREE_CODE (base) == ARRAY_RANGE_REF) |
2132 | { |
2133 | if (TREE_OPERAND (base, 2)) |
2134 | find_interesting_uses_op (data, TREE_OPERAND (base, 2)); |
2135 | if (TREE_OPERAND (base, 3)) |
2136 | find_interesting_uses_op (data, TREE_OPERAND (base, 3)); |
2137 | } |
2138 | return true; |
2139 | } |
2140 | |
2141 | /* If we can prove that TOP = cst * BOT for some constant cst, |
2142 | store cst to MUL and return true. Otherwise return false. |
2143 | The returned value is always sign-extended, regardless of the |
2144 | signedness of TOP and BOT. */ |
2145 | |
2146 | static bool |
2147 | constant_multiple_of (tree top, tree bot, widest_int *mul) |
2148 | { |
2149 | tree mby; |
2150 | enum tree_code code; |
2151 | unsigned precision = TYPE_PRECISION (TREE_TYPE (top)); |
2152 | widest_int res, p0, p1; |
2153 | |
2154 | STRIP_NOPS (top); |
2155 | STRIP_NOPS (bot); |
2156 | |
2157 | if (operand_equal_p (top, bot, flags: 0)) |
2158 | { |
2159 | *mul = 1; |
2160 | return true; |
2161 | } |
2162 | |
2163 | code = TREE_CODE (top); |
2164 | switch (code) |
2165 | { |
2166 | case MULT_EXPR: |
2167 | mby = TREE_OPERAND (top, 1); |
2168 | if (TREE_CODE (mby) != INTEGER_CST) |
2169 | return false; |
2170 | |
2171 | if (!constant_multiple_of (TREE_OPERAND (top, 0), bot, mul: &res)) |
2172 | return false; |
2173 | |
2174 | *mul = wi::sext (x: res * wi::to_widest (t: mby), offset: precision); |
2175 | return true; |
2176 | |
2177 | case PLUS_EXPR: |
2178 | case MINUS_EXPR: |
2179 | if (!constant_multiple_of (TREE_OPERAND (top, 0), bot, mul: &p0) |
2180 | || !constant_multiple_of (TREE_OPERAND (top, 1), bot, mul: &p1)) |
2181 | return false; |
2182 | |
2183 | if (code == MINUS_EXPR) |
2184 | p1 = -p1; |
2185 | *mul = wi::sext (x: p0 + p1, offset: precision); |
2186 | return true; |
2187 | |
2188 | case INTEGER_CST: |
2189 | if (TREE_CODE (bot) != INTEGER_CST) |
2190 | return false; |
2191 | |
2192 | p0 = widest_int::from (x: wi::to_wide (t: top), sgn: SIGNED); |
2193 | p1 = widest_int::from (x: wi::to_wide (t: bot), sgn: SIGNED); |
2194 | if (p1 == 0) |
2195 | return false; |
2196 | *mul = wi::sext (x: wi::divmod_trunc (x: p0, y: p1, sgn: SIGNED, remainder_ptr: &res), offset: precision); |
2197 | return res == 0; |
2198 | |
2199 | default: |
2200 | if (POLY_INT_CST_P (top) |
2201 | && POLY_INT_CST_P (bot) |
2202 | && constant_multiple_p (a: wi::to_poly_widest (t: top), |
2203 | b: wi::to_poly_widest (t: bot), multiple: mul)) |
2204 | return true; |
2205 | |
2206 | return false; |
2207 | } |
2208 | } |
2209 | |
2210 | /* Return true if memory reference REF with step STEP may be unaligned. */ |
2211 | |
2212 | static bool |
2213 | may_be_unaligned_p (tree ref, tree step) |
2214 | { |
2215 | /* TARGET_MEM_REFs are translated directly to valid MEMs on the target, |
2216 | thus they are not misaligned. */ |
2217 | if (TREE_CODE (ref) == TARGET_MEM_REF) |
2218 | return false; |
2219 | |
2220 | unsigned int align = TYPE_ALIGN (TREE_TYPE (ref)); |
2221 | if (GET_MODE_ALIGNMENT (TYPE_MODE (TREE_TYPE (ref))) > align) |
2222 | align = GET_MODE_ALIGNMENT (TYPE_MODE (TREE_TYPE (ref))); |
2223 | |
2224 | unsigned HOST_WIDE_INT bitpos; |
2225 | unsigned int ref_align; |
2226 | get_object_alignment_1 (ref, &ref_align, &bitpos); |
2227 | if (ref_align < align |
2228 | || (bitpos % align) != 0 |
2229 | || (bitpos % BITS_PER_UNIT) != 0) |
2230 | return true; |
2231 | |
2232 | unsigned int trailing_zeros = tree_ctz (step); |
2233 | if (trailing_zeros < HOST_BITS_PER_INT |
2234 | && (1U << trailing_zeros) * BITS_PER_UNIT < align) |
2235 | return true; |
2236 | |
2237 | return false; |
2238 | } |
2239 | |
2240 | /* Return true if EXPR may be non-addressable. */ |
2241 | |
2242 | bool |
2243 | may_be_nonaddressable_p (tree expr) |
2244 | { |
2245 | switch (TREE_CODE (expr)) |
2246 | { |
2247 | case VAR_DECL: |
2248 | /* Check if it's a register variable. */ |
2249 | return DECL_HARD_REGISTER (expr); |
2250 | |
2251 | case TARGET_MEM_REF: |
2252 | /* TARGET_MEM_REFs are translated directly to valid MEMs on the |
2253 | target, thus they are always addressable. */ |
2254 | return false; |
2255 | |
2256 | case MEM_REF: |
2257 | /* Likewise for MEM_REFs, modulo the storage order. */ |
2258 | return REF_REVERSE_STORAGE_ORDER (expr); |
2259 | |
2260 | case BIT_FIELD_REF: |
2261 | if (REF_REVERSE_STORAGE_ORDER (expr)) |
2262 | return true; |
2263 | return may_be_nonaddressable_p (TREE_OPERAND (expr, 0)); |
2264 | |
2265 | case COMPONENT_REF: |
2266 | if (TYPE_REVERSE_STORAGE_ORDER (TREE_TYPE (TREE_OPERAND (expr, 0)))) |
2267 | return true; |
2268 | return DECL_NONADDRESSABLE_P (TREE_OPERAND (expr, 1)) |
2269 | || may_be_nonaddressable_p (TREE_OPERAND (expr, 0)); |
2270 | |
2271 | case ARRAY_REF: |
2272 | case ARRAY_RANGE_REF: |
2273 | if (TYPE_REVERSE_STORAGE_ORDER (TREE_TYPE (TREE_OPERAND (expr, 0)))) |
2274 | return true; |
2275 | return may_be_nonaddressable_p (TREE_OPERAND (expr, 0)); |
2276 | |
2277 | case VIEW_CONVERT_EXPR: |
2278 | /* This kind of view-conversions may wrap non-addressable objects |
2279 | and make them look addressable. After some processing the |
2280 | non-addressability may be uncovered again, causing ADDR_EXPRs |
2281 | of inappropriate objects to be built. */ |
2282 | if (is_gimple_reg (TREE_OPERAND (expr, 0)) |
2283 | || !is_gimple_addressable (TREE_OPERAND (expr, 0))) |
2284 | return true; |
2285 | return may_be_nonaddressable_p (TREE_OPERAND (expr, 0)); |
2286 | |
2287 | CASE_CONVERT: |
2288 | return true; |
2289 | |
2290 | default: |
2291 | break; |
2292 | } |
2293 | |
2294 | return false; |
2295 | } |
2296 | |
2297 | /* Finds addresses in *OP_P inside STMT. */ |
2298 | |
2299 | static void |
2300 | find_interesting_uses_address (struct ivopts_data *data, gimple *stmt, |
2301 | tree *op_p) |
2302 | { |
2303 | tree base = *op_p, step = size_zero_node; |
2304 | struct iv *civ; |
2305 | struct ifs_ivopts_data ifs_ivopts_data; |
2306 | |
2307 | /* Do not play with volatile memory references. A bit too conservative, |
2308 | perhaps, but safe. */ |
2309 | if (gimple_has_volatile_ops (stmt)) |
2310 | goto fail; |
2311 | |
2312 | /* Ignore bitfields for now. Not really something terribly complicated |
2313 | to handle. TODO. */ |
2314 | if (TREE_CODE (base) == BIT_FIELD_REF) |
2315 | goto fail; |
2316 | |
2317 | base = unshare_expr (base); |
2318 | |
2319 | if (TREE_CODE (base) == TARGET_MEM_REF) |
2320 | { |
2321 | tree type = build_pointer_type (TREE_TYPE (base)); |
2322 | tree astep; |
2323 | |
2324 | if (TMR_BASE (base) |
2325 | && TREE_CODE (TMR_BASE (base)) == SSA_NAME) |
2326 | { |
2327 | civ = get_iv (data, TMR_BASE (base)); |
2328 | if (!civ) |
2329 | goto fail; |
2330 | |
2331 | TMR_BASE (base) = civ->base; |
2332 | step = civ->step; |
2333 | } |
2334 | if (TMR_INDEX2 (base) |
2335 | && TREE_CODE (TMR_INDEX2 (base)) == SSA_NAME) |
2336 | { |
2337 | civ = get_iv (data, TMR_INDEX2 (base)); |
2338 | if (!civ) |
2339 | goto fail; |
2340 | |
2341 | TMR_INDEX2 (base) = civ->base; |
2342 | step = civ->step; |
2343 | } |
2344 | if (TMR_INDEX (base) |
2345 | && TREE_CODE (TMR_INDEX (base)) == SSA_NAME) |
2346 | { |
2347 | civ = get_iv (data, TMR_INDEX (base)); |
2348 | if (!civ) |
2349 | goto fail; |
2350 | |
2351 | TMR_INDEX (base) = civ->base; |
2352 | astep = civ->step; |
2353 | |
2354 | if (astep) |
2355 | { |
2356 | if (TMR_STEP (base)) |
2357 | astep = fold_build2 (MULT_EXPR, type, TMR_STEP (base), astep); |
2358 | |
2359 | step = fold_build2 (PLUS_EXPR, type, step, astep); |
2360 | } |
2361 | } |
2362 | |
2363 | if (integer_zerop (step)) |
2364 | goto fail; |
2365 | base = tree_mem_ref_addr (type, base); |
2366 | } |
2367 | else |
2368 | { |
2369 | ifs_ivopts_data.ivopts_data = data; |
2370 | ifs_ivopts_data.stmt = stmt; |
2371 | ifs_ivopts_data.step = size_zero_node; |
2372 | if (!for_each_index (&base, idx_find_step, &ifs_ivopts_data) |
2373 | || integer_zerop (ifs_ivopts_data.step)) |
2374 | goto fail; |
2375 | step = ifs_ivopts_data.step; |
2376 | |
2377 | /* Check that the base expression is addressable. This needs |
2378 | to be done after substituting bases of IVs into it. */ |
2379 | if (may_be_nonaddressable_p (expr: base)) |
2380 | goto fail; |
2381 | |
2382 | /* Moreover, on strict alignment platforms, check that it is |
2383 | sufficiently aligned. */ |
2384 | if (STRICT_ALIGNMENT && may_be_unaligned_p (ref: base, step)) |
2385 | goto fail; |
2386 | |
2387 | base = build_fold_addr_expr (base); |
2388 | |
2389 | /* Substituting bases of IVs into the base expression might |
2390 | have caused folding opportunities. */ |
2391 | if (TREE_CODE (base) == ADDR_EXPR) |
2392 | { |
2393 | tree *ref = &TREE_OPERAND (base, 0); |
2394 | while (handled_component_p (t: *ref)) |
2395 | ref = &TREE_OPERAND (*ref, 0); |
2396 | if (TREE_CODE (*ref) == MEM_REF) |
2397 | { |
2398 | tree tem = fold_binary (MEM_REF, TREE_TYPE (*ref), |
2399 | TREE_OPERAND (*ref, 0), |
2400 | TREE_OPERAND (*ref, 1)); |
2401 | if (tem) |
2402 | *ref = tem; |
2403 | } |
2404 | } |
2405 | } |
2406 | |
2407 | civ = alloc_iv (data, base, step); |
2408 | /* Fail if base object of this memory reference is unknown. */ |
2409 | if (civ->base_object == NULL_TREE) |
2410 | goto fail; |
2411 | |
2412 | record_group_use (data, use_p: op_p, iv: civ, stmt, type: USE_REF_ADDRESS, TREE_TYPE (*op_p)); |
2413 | return; |
2414 | |
2415 | fail: |
2416 | for_each_index (op_p, idx_record_use, data); |
2417 | } |
2418 | |
2419 | /* Finds and records invariants used in STMT. */ |
2420 | |
2421 | static void |
2422 | find_invariants_stmt (struct ivopts_data *data, gimple *stmt) |
2423 | { |
2424 | ssa_op_iter iter; |
2425 | use_operand_p use_p; |
2426 | tree op; |
2427 | |
2428 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) |
2429 | { |
2430 | op = USE_FROM_PTR (use_p); |
2431 | record_invariant (data, op, nonlinear_use: false); |
2432 | } |
2433 | } |
2434 | |
2435 | /* CALL calls an internal function. If operand *OP_P will become an |
2436 | address when the call is expanded, return the type of the memory |
2437 | being addressed, otherwise return null. */ |
2438 | |
2439 | static tree |
2440 | get_mem_type_for_internal_fn (gcall *call, tree *op_p) |
2441 | { |
2442 | switch (gimple_call_internal_fn (gs: call)) |
2443 | { |
2444 | case IFN_MASK_LOAD: |
2445 | case IFN_MASK_LOAD_LANES: |
2446 | case IFN_MASK_LEN_LOAD_LANES: |
2447 | case IFN_LEN_LOAD: |
2448 | case IFN_MASK_LEN_LOAD: |
2449 | if (op_p == gimple_call_arg_ptr (gs: call, index: 0)) |
2450 | return TREE_TYPE (gimple_call_lhs (call)); |
2451 | return NULL_TREE; |
2452 | |
2453 | case IFN_MASK_STORE: |
2454 | case IFN_MASK_STORE_LANES: |
2455 | case IFN_MASK_LEN_STORE_LANES: |
2456 | case IFN_LEN_STORE: |
2457 | case IFN_MASK_LEN_STORE: |
2458 | { |
2459 | if (op_p == gimple_call_arg_ptr (gs: call, index: 0)) |
2460 | { |
2461 | internal_fn ifn = gimple_call_internal_fn (gs: call); |
2462 | int index = internal_fn_stored_value_index (ifn); |
2463 | return TREE_TYPE (gimple_call_arg (call, index)); |
2464 | } |
2465 | return NULL_TREE; |
2466 | } |
2467 | |
2468 | default: |
2469 | return NULL_TREE; |
2470 | } |
2471 | } |
2472 | |
2473 | /* IV is a (non-address) iv that describes operand *OP_P of STMT. |
2474 | Return true if the operand will become an address when STMT |
2475 | is expanded and record the associated address use if so. */ |
2476 | |
2477 | static bool |
2478 | find_address_like_use (struct ivopts_data *data, gimple *stmt, tree *op_p, |
2479 | struct iv *iv) |
2480 | { |
2481 | /* Fail if base object of this memory reference is unknown. */ |
2482 | if (iv->base_object == NULL_TREE) |
2483 | return false; |
2484 | |
2485 | tree mem_type = NULL_TREE; |
2486 | if (gcall *call = dyn_cast <gcall *> (p: stmt)) |
2487 | if (gimple_call_internal_p (gs: call)) |
2488 | mem_type = get_mem_type_for_internal_fn (call, op_p); |
2489 | if (mem_type) |
2490 | { |
2491 | iv = alloc_iv (data, base: iv->base, step: iv->step); |
2492 | record_group_use (data, use_p: op_p, iv, stmt, type: USE_PTR_ADDRESS, mem_type); |
2493 | return true; |
2494 | } |
2495 | return false; |
2496 | } |
2497 | |
2498 | /* Finds interesting uses of induction variables in the statement STMT. */ |
2499 | |
2500 | static void |
2501 | find_interesting_uses_stmt (struct ivopts_data *data, gimple *stmt) |
2502 | { |
2503 | struct iv *iv; |
2504 | tree op, *lhs, *rhs; |
2505 | ssa_op_iter iter; |
2506 | use_operand_p use_p; |
2507 | enum tree_code code; |
2508 | |
2509 | find_invariants_stmt (data, stmt); |
2510 | |
2511 | if (gimple_code (g: stmt) == GIMPLE_COND) |
2512 | { |
2513 | find_interesting_uses_cond (data, stmt); |
2514 | return; |
2515 | } |
2516 | |
2517 | if (is_gimple_assign (gs: stmt)) |
2518 | { |
2519 | lhs = gimple_assign_lhs_ptr (gs: stmt); |
2520 | rhs = gimple_assign_rhs1_ptr (gs: stmt); |
2521 | |
2522 | if (TREE_CODE (*lhs) == SSA_NAME) |
2523 | { |
2524 | /* If the statement defines an induction variable, the uses are not |
2525 | interesting by themselves. */ |
2526 | |
2527 | iv = get_iv (data, var: *lhs); |
2528 | |
2529 | if (iv && !integer_zerop (iv->step)) |
2530 | return; |
2531 | } |
2532 | |
2533 | code = gimple_assign_rhs_code (gs: stmt); |
2534 | if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS |
2535 | && (REFERENCE_CLASS_P (*rhs) |
2536 | || is_gimple_val (*rhs))) |
2537 | { |
2538 | if (REFERENCE_CLASS_P (*rhs)) |
2539 | find_interesting_uses_address (data, stmt, op_p: rhs); |
2540 | else |
2541 | find_interesting_uses_op (data, op: *rhs); |
2542 | |
2543 | if (REFERENCE_CLASS_P (*lhs)) |
2544 | find_interesting_uses_address (data, stmt, op_p: lhs); |
2545 | return; |
2546 | } |
2547 | else if (TREE_CODE_CLASS (code) == tcc_comparison) |
2548 | { |
2549 | find_interesting_uses_cond (data, stmt); |
2550 | return; |
2551 | } |
2552 | |
2553 | /* TODO -- we should also handle address uses of type |
2554 | |
2555 | memory = call (whatever); |
2556 | |
2557 | and |
2558 | |
2559 | call (memory). */ |
2560 | } |
2561 | |
2562 | if (gimple_code (g: stmt) == GIMPLE_PHI |
2563 | && gimple_bb (g: stmt) == data->current_loop->header) |
2564 | { |
2565 | iv = get_iv (data, PHI_RESULT (stmt)); |
2566 | |
2567 | if (iv && !integer_zerop (iv->step)) |
2568 | return; |
2569 | } |
2570 | |
2571 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) |
2572 | { |
2573 | op = USE_FROM_PTR (use_p); |
2574 | |
2575 | if (TREE_CODE (op) != SSA_NAME) |
2576 | continue; |
2577 | |
2578 | iv = get_iv (data, var: op); |
2579 | if (!iv) |
2580 | continue; |
2581 | |
2582 | if (!find_address_like_use (data, stmt, op_p: use_p->use, iv)) |
2583 | find_interesting_uses_op (data, op); |
2584 | } |
2585 | } |
2586 | |
2587 | /* Finds interesting uses of induction variables outside of loops |
2588 | on loop exit edge EXIT. */ |
2589 | |
2590 | static void |
2591 | find_interesting_uses_outside (struct ivopts_data *data, edge exit) |
2592 | { |
2593 | gphi *phi; |
2594 | gphi_iterator psi; |
2595 | tree def; |
2596 | |
2597 | for (psi = gsi_start_phis (exit->dest); !gsi_end_p (i: psi); gsi_next (i: &psi)) |
2598 | { |
2599 | phi = psi.phi (); |
2600 | def = PHI_ARG_DEF_FROM_EDGE (phi, exit); |
2601 | if (!virtual_operand_p (op: def)) |
2602 | find_interesting_uses_op (data, op: def); |
2603 | } |
2604 | } |
2605 | |
2606 | /* Return TRUE if OFFSET is within the range of [base + offset] addressing |
2607 | mode for memory reference represented by USE. */ |
2608 | |
2609 | static GTY (()) vec<rtx, va_gc> *addr_list; |
2610 | |
2611 | static bool |
2612 | addr_offset_valid_p (struct iv_use *use, poly_int64 offset) |
2613 | { |
2614 | rtx reg, addr; |
2615 | unsigned list_index; |
2616 | addr_space_t as = TYPE_ADDR_SPACE (TREE_TYPE (use->iv->base)); |
2617 | machine_mode addr_mode, mem_mode = TYPE_MODE (use->mem_type); |
2618 | |
2619 | list_index = (unsigned) as * MAX_MACHINE_MODE + (unsigned) mem_mode; |
2620 | if (list_index >= vec_safe_length (v: addr_list)) |
2621 | vec_safe_grow_cleared (v&: addr_list, len: list_index + MAX_MACHINE_MODE, exact: true); |
2622 | |
2623 | addr = (*addr_list)[list_index]; |
2624 | if (!addr) |
2625 | { |
2626 | addr_mode = targetm.addr_space.address_mode (as); |
2627 | reg = gen_raw_REG (addr_mode, LAST_VIRTUAL_REGISTER + 1); |
2628 | addr = gen_rtx_fmt_ee (PLUS, addr_mode, reg, NULL_RTX); |
2629 | (*addr_list)[list_index] = addr; |
2630 | } |
2631 | else |
2632 | addr_mode = GET_MODE (addr); |
2633 | |
2634 | XEXP (addr, 1) = gen_int_mode (offset, addr_mode); |
2635 | return (memory_address_addr_space_p (mem_mode, addr, as)); |
2636 | } |
2637 | |
2638 | /* Comparison function to sort group in ascending order of addr_offset. */ |
2639 | |
2640 | static int |
2641 | group_compare_offset (const void *a, const void *b) |
2642 | { |
2643 | const struct iv_use *const *u1 = (const struct iv_use *const *) a; |
2644 | const struct iv_use *const *u2 = (const struct iv_use *const *) b; |
2645 | |
2646 | return compare_sizes_for_sort (a: (*u1)->addr_offset, b: (*u2)->addr_offset); |
2647 | } |
2648 | |
2649 | /* Check if small groups should be split. Return true if no group |
2650 | contains more than two uses with distinct addr_offsets. Return |
2651 | false otherwise. We want to split such groups because: |
2652 | |
2653 | 1) Small groups don't have much benefit and may interfer with |
2654 | general candidate selection. |
2655 | 2) Size for problem with only small groups is usually small and |
2656 | general algorithm can handle it well. |
2657 | |
2658 | TODO -- Above claim may not hold when we want to merge memory |
2659 | accesses with conseuctive addresses. */ |
2660 | |
2661 | static bool |
2662 | split_small_address_groups_p (struct ivopts_data *data) |
2663 | { |
2664 | unsigned int i, j, distinct = 1; |
2665 | struct iv_use *pre; |
2666 | struct iv_group *group; |
2667 | |
2668 | for (i = 0; i < data->vgroups.length (); i++) |
2669 | { |
2670 | group = data->vgroups[i]; |
2671 | if (group->vuses.length () == 1) |
2672 | continue; |
2673 | |
2674 | gcc_assert (address_p (group->type)); |
2675 | if (group->vuses.length () == 2) |
2676 | { |
2677 | if (compare_sizes_for_sort (a: group->vuses[0]->addr_offset, |
2678 | b: group->vuses[1]->addr_offset) > 0) |
2679 | std::swap (a&: group->vuses[0], b&: group->vuses[1]); |
2680 | } |
2681 | else |
2682 | group->vuses.qsort (group_compare_offset); |
2683 | |
2684 | if (distinct > 2) |
2685 | continue; |
2686 | |
2687 | distinct = 1; |
2688 | for (pre = group->vuses[0], j = 1; j < group->vuses.length (); j++) |
2689 | { |
2690 | if (maybe_ne (a: group->vuses[j]->addr_offset, b: pre->addr_offset)) |
2691 | { |
2692 | pre = group->vuses[j]; |
2693 | distinct++; |
2694 | } |
2695 | |
2696 | if (distinct > 2) |
2697 | break; |
2698 | } |
2699 | } |
2700 | |
2701 | return (distinct <= 2); |
2702 | } |
2703 | |
2704 | /* For each group of address type uses, this function further groups |
2705 | these uses according to the maximum offset supported by target's |
2706 | [base + offset] addressing mode. */ |
2707 | |
2708 | static void |
2709 | split_address_groups (struct ivopts_data *data) |
2710 | { |
2711 | unsigned int i, j; |
2712 | /* Always split group. */ |
2713 | bool split_p = split_small_address_groups_p (data); |
2714 | |
2715 | for (i = 0; i < data->vgroups.length (); i++) |
2716 | { |
2717 | struct iv_group *new_group = NULL; |
2718 | struct iv_group *group = data->vgroups[i]; |
2719 | struct iv_use *use = group->vuses[0]; |
2720 | |
2721 | use->id = 0; |
2722 | use->group_id = group->id; |
2723 | if (group->vuses.length () == 1) |
2724 | continue; |
2725 | |
2726 | gcc_assert (address_p (use->type)); |
2727 | |
2728 | for (j = 1; j < group->vuses.length ();) |
2729 | { |
2730 | struct iv_use *next = group->vuses[j]; |
2731 | poly_int64 offset = next->addr_offset - use->addr_offset; |
2732 | |
2733 | /* Split group if aksed to, or the offset against the first |
2734 | use can't fit in offset part of addressing mode. IV uses |
2735 | having the same offset are still kept in one group. */ |
2736 | if (maybe_ne (a: offset, b: 0) |
2737 | && (split_p || !addr_offset_valid_p (use, offset))) |
2738 | { |
2739 | if (!new_group) |
2740 | new_group = record_group (data, type: group->type); |
2741 | group->vuses.ordered_remove (ix: j); |
2742 | new_group->vuses.safe_push (obj: next); |
2743 | continue; |
2744 | } |
2745 | |
2746 | next->id = j; |
2747 | next->group_id = group->id; |
2748 | j++; |
2749 | } |
2750 | } |
2751 | } |
2752 | |
2753 | /* Finds uses of the induction variables that are interesting. */ |
2754 | |
2755 | static void |
2756 | find_interesting_uses (struct ivopts_data *data, basic_block *body) |
2757 | { |
2758 | basic_block bb; |
2759 | gimple_stmt_iterator bsi; |
2760 | unsigned i; |
2761 | edge e; |
2762 | |
2763 | for (i = 0; i < data->current_loop->num_nodes; i++) |
2764 | { |
2765 | edge_iterator ei; |
2766 | bb = body[i]; |
2767 | |
2768 | FOR_EACH_EDGE (e, ei, bb->succs) |
2769 | if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
2770 | && !flow_bb_inside_loop_p (data->current_loop, e->dest)) |
2771 | find_interesting_uses_outside (data, exit: e); |
2772 | |
2773 | for (bsi = gsi_start_phis (bb); !gsi_end_p (i: bsi); gsi_next (i: &bsi)) |
2774 | find_interesting_uses_stmt (data, stmt: gsi_stmt (i: bsi)); |
2775 | for (bsi = gsi_start_bb (bb); !gsi_end_p (i: bsi); gsi_next (i: &bsi)) |
2776 | if (!is_gimple_debug (gs: gsi_stmt (i: bsi))) |
2777 | find_interesting_uses_stmt (data, stmt: gsi_stmt (i: bsi)); |
2778 | } |
2779 | |
2780 | split_address_groups (data); |
2781 | |
2782 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2783 | { |
2784 | fprintf (stream: dump_file, format: "\n<IV Groups>:\n" ); |
2785 | dump_groups (file: dump_file, data); |
2786 | fprintf (stream: dump_file, format: "\n" ); |
2787 | } |
2788 | } |
2789 | |
2790 | /* Strips constant offsets from EXPR and stores them to OFFSET. If INSIDE_ADDR |
2791 | is true, assume we are inside an address. If TOP_COMPREF is true, assume |
2792 | we are at the top-level of the processed address. */ |
2793 | |
2794 | static tree |
2795 | strip_offset_1 (tree expr, bool inside_addr, bool top_compref, |
2796 | poly_int64 *offset) |
2797 | { |
2798 | tree op0 = NULL_TREE, op1 = NULL_TREE, tmp, step; |
2799 | enum tree_code code; |
2800 | tree type, orig_type = TREE_TYPE (expr); |
2801 | poly_int64 off0, off1; |
2802 | HOST_WIDE_INT st; |
2803 | tree orig_expr = expr; |
2804 | |
2805 | STRIP_NOPS (expr); |
2806 | |
2807 | type = TREE_TYPE (expr); |
2808 | code = TREE_CODE (expr); |
2809 | *offset = 0; |
2810 | |
2811 | switch (code) |
2812 | { |
2813 | case POINTER_PLUS_EXPR: |
2814 | case PLUS_EXPR: |
2815 | case MINUS_EXPR: |
2816 | op0 = TREE_OPERAND (expr, 0); |
2817 | op1 = TREE_OPERAND (expr, 1); |
2818 | |
2819 | op0 = strip_offset_1 (expr: op0, inside_addr: false, top_compref: false, offset: &off0); |
2820 | op1 = strip_offset_1 (expr: op1, inside_addr: false, top_compref: false, offset: &off1); |
2821 | |
2822 | *offset = (code == MINUS_EXPR ? off0 - off1 : off0 + off1); |
2823 | if (op0 == TREE_OPERAND (expr, 0) |
2824 | && op1 == TREE_OPERAND (expr, 1)) |
2825 | return orig_expr; |
2826 | |
2827 | if (integer_zerop (op1)) |
2828 | expr = op0; |
2829 | else if (integer_zerop (op0)) |
2830 | { |
2831 | if (code == MINUS_EXPR) |
2832 | { |
2833 | if (TYPE_OVERFLOW_UNDEFINED (type)) |
2834 | { |
2835 | type = unsigned_type_for (type); |
2836 | op1 = fold_convert (type, op1); |
2837 | } |
2838 | expr = fold_build1 (NEGATE_EXPR, type, op1); |
2839 | } |
2840 | else |
2841 | expr = op1; |
2842 | } |
2843 | else |
2844 | { |
2845 | if (TYPE_OVERFLOW_UNDEFINED (type)) |
2846 | { |
2847 | type = unsigned_type_for (type); |
2848 | if (code == POINTER_PLUS_EXPR) |
2849 | code = PLUS_EXPR; |
2850 | op0 = fold_convert (type, op0); |
2851 | op1 = fold_convert (type, op1); |
2852 | } |
2853 | expr = fold_build2 (code, type, op0, op1); |
2854 | } |
2855 | |
2856 | return fold_convert (orig_type, expr); |
2857 | |
2858 | case MULT_EXPR: |
2859 | op1 = TREE_OPERAND (expr, 1); |
2860 | if (!cst_and_fits_in_hwi (op1)) |
2861 | return orig_expr; |
2862 | |
2863 | op0 = TREE_OPERAND (expr, 0); |
2864 | op0 = strip_offset_1 (expr: op0, inside_addr: false, top_compref: false, offset: &off0); |
2865 | if (op0 == TREE_OPERAND (expr, 0)) |
2866 | return orig_expr; |
2867 | |
2868 | *offset = off0 * int_cst_value (op1); |
2869 | if (integer_zerop (op0)) |
2870 | expr = op0; |
2871 | else |
2872 | { |
2873 | if (TYPE_OVERFLOW_UNDEFINED (type)) |
2874 | { |
2875 | type = unsigned_type_for (type); |
2876 | op0 = fold_convert (type, op0); |
2877 | op1 = fold_convert (type, op1); |
2878 | } |
2879 | expr = fold_build2 (MULT_EXPR, type, op0, op1); |
2880 | } |
2881 | |
2882 | return fold_convert (orig_type, expr); |
2883 | |
2884 | case ARRAY_REF: |
2885 | case ARRAY_RANGE_REF: |
2886 | if (!inside_addr) |
2887 | return orig_expr; |
2888 | |
2889 | step = array_ref_element_size (expr); |
2890 | if (!cst_and_fits_in_hwi (step)) |
2891 | break; |
2892 | |
2893 | st = int_cst_value (step); |
2894 | op1 = TREE_OPERAND (expr, 1); |
2895 | op1 = strip_offset_1 (expr: op1, inside_addr: false, top_compref: false, offset: &off1); |
2896 | *offset = off1 * st; |
2897 | |
2898 | if (top_compref |
2899 | && integer_zerop (op1)) |
2900 | { |
2901 | /* Strip the component reference completely. */ |
2902 | op0 = TREE_OPERAND (expr, 0); |
2903 | op0 = strip_offset_1 (expr: op0, inside_addr, top_compref, offset: &off0); |
2904 | *offset += off0; |
2905 | return op0; |
2906 | } |
2907 | break; |
2908 | |
2909 | case COMPONENT_REF: |
2910 | { |
2911 | tree field; |
2912 | |
2913 | if (!inside_addr) |
2914 | return orig_expr; |
2915 | |
2916 | tmp = component_ref_field_offset (expr); |
2917 | field = TREE_OPERAND (expr, 1); |
2918 | if (top_compref |
2919 | && cst_and_fits_in_hwi (tmp) |
2920 | && cst_and_fits_in_hwi (DECL_FIELD_BIT_OFFSET (field))) |
2921 | { |
2922 | HOST_WIDE_INT boffset, abs_off; |
2923 | |
2924 | /* Strip the component reference completely. */ |
2925 | op0 = TREE_OPERAND (expr, 0); |
2926 | op0 = strip_offset_1 (expr: op0, inside_addr, top_compref, offset: &off0); |
2927 | boffset = int_cst_value (DECL_FIELD_BIT_OFFSET (field)); |
2928 | abs_off = abs_hwi (x: boffset) / BITS_PER_UNIT; |
2929 | if (boffset < 0) |
2930 | abs_off = -abs_off; |
2931 | |
2932 | *offset = off0 + int_cst_value (tmp) + abs_off; |
2933 | return op0; |
2934 | } |
2935 | } |
2936 | break; |
2937 | |
2938 | case ADDR_EXPR: |
2939 | op0 = TREE_OPERAND (expr, 0); |
2940 | op0 = strip_offset_1 (expr: op0, inside_addr: true, top_compref: true, offset: &off0); |
2941 | *offset += off0; |
2942 | |
2943 | if (op0 == TREE_OPERAND (expr, 0)) |
2944 | return orig_expr; |
2945 | |
2946 | expr = build_fold_addr_expr (op0); |
2947 | return fold_convert (orig_type, expr); |
2948 | |
2949 | case MEM_REF: |
2950 | /* ??? Offset operand? */ |
2951 | inside_addr = false; |
2952 | break; |
2953 | |
2954 | default: |
2955 | if (ptrdiff_tree_p (expr, offset) && maybe_ne (a: *offset, b: 0)) |
2956 | return build_int_cst (orig_type, 0); |
2957 | return orig_expr; |
2958 | } |
2959 | |
2960 | /* Default handling of expressions for that we want to recurse into |
2961 | the first operand. */ |
2962 | op0 = TREE_OPERAND (expr, 0); |
2963 | op0 = strip_offset_1 (expr: op0, inside_addr, top_compref: false, offset: &off0); |
2964 | *offset += off0; |
2965 | |
2966 | if (op0 == TREE_OPERAND (expr, 0) |
2967 | && (!op1 || op1 == TREE_OPERAND (expr, 1))) |
2968 | return orig_expr; |
2969 | |
2970 | expr = copy_node (expr); |
2971 | TREE_OPERAND (expr, 0) = op0; |
2972 | if (op1) |
2973 | TREE_OPERAND (expr, 1) = op1; |
2974 | |
2975 | /* Inside address, we might strip the top level component references, |
2976 | thus changing type of the expression. Handling of ADDR_EXPR |
2977 | will fix that. */ |
2978 | expr = fold_convert (orig_type, expr); |
2979 | |
2980 | return expr; |
2981 | } |
2982 | |
2983 | /* Strips constant offsets from EXPR and stores them to OFFSET. */ |
2984 | |
2985 | static tree |
2986 | strip_offset (tree expr, poly_uint64 *offset) |
2987 | { |
2988 | poly_int64 off; |
2989 | tree core = strip_offset_1 (expr, inside_addr: false, top_compref: false, offset: &off); |
2990 | *offset = off; |
2991 | return core; |
2992 | } |
2993 | |
2994 | /* Returns variant of TYPE that can be used as base for different uses. |
2995 | We return unsigned type with the same precision, which avoids problems |
2996 | with overflows. */ |
2997 | |
2998 | static tree |
2999 | generic_type_for (tree type) |
3000 | { |
3001 | if (POINTER_TYPE_P (type)) |
3002 | return unsigned_type_for (type); |
3003 | |
3004 | if (TYPE_UNSIGNED (type)) |
3005 | return type; |
3006 | |
3007 | return unsigned_type_for (type); |
3008 | } |
3009 | |
3010 | /* Private data for walk_tree. */ |
3011 | |
3012 | struct walk_tree_data |
3013 | { |
3014 | bitmap *inv_vars; |
3015 | struct ivopts_data *idata; |
3016 | }; |
3017 | |
3018 | /* Callback function for walk_tree, it records invariants and symbol |
3019 | reference in *EXPR_P. DATA is the structure storing result info. */ |
3020 | |
3021 | static tree |
3022 | find_inv_vars_cb (tree *expr_p, int *ws ATTRIBUTE_UNUSED, void *data) |
3023 | { |
3024 | tree op = *expr_p; |
3025 | struct version_info *info; |
3026 | struct walk_tree_data *wdata = (struct walk_tree_data*) data; |
3027 | |
3028 | if (TREE_CODE (op) != SSA_NAME) |
3029 | return NULL_TREE; |
3030 | |
3031 | info = name_info (data: wdata->idata, name: op); |
3032 | /* Because we expand simple operations when finding IVs, loop invariant |
3033 | variable that isn't referred by the original loop could be used now. |
3034 | Record such invariant variables here. */ |
3035 | if (!info->iv) |
3036 | { |
3037 | struct ivopts_data *idata = wdata->idata; |
3038 | basic_block bb = gimple_bb (SSA_NAME_DEF_STMT (op)); |
3039 | |
3040 | if (!bb || !flow_bb_inside_loop_p (idata->current_loop, bb)) |
3041 | { |
3042 | tree steptype = TREE_TYPE (op); |
3043 | if (POINTER_TYPE_P (steptype)) |
3044 | steptype = sizetype; |
3045 | set_iv (data: idata, iv: op, base: op, step: build_int_cst (steptype, 0), no_overflow: true); |
3046 | record_invariant (data: idata, op, nonlinear_use: false); |
3047 | } |
3048 | } |
3049 | if (!info->inv_id || info->has_nonlin_use) |
3050 | return NULL_TREE; |
3051 | |
3052 | if (!*wdata->inv_vars) |
3053 | *wdata->inv_vars = BITMAP_ALLOC (NULL); |
3054 | bitmap_set_bit (*wdata->inv_vars, info->inv_id); |
3055 | |
3056 | return NULL_TREE; |
3057 | } |
3058 | |
3059 | /* Records invariants in *EXPR_P. INV_VARS is the bitmap to that we should |
3060 | store it. */ |
3061 | |
3062 | static inline void |
3063 | find_inv_vars (struct ivopts_data *data, tree *expr_p, bitmap *inv_vars) |
3064 | { |
3065 | struct walk_tree_data wdata; |
3066 | |
3067 | if (!inv_vars) |
3068 | return; |
3069 | |
3070 | wdata.idata = data; |
3071 | wdata.inv_vars = inv_vars; |
3072 | walk_tree (expr_p, find_inv_vars_cb, &wdata, NULL); |
3073 | } |
3074 | |
3075 | /* Get entry from invariant expr hash table for INV_EXPR. New entry |
3076 | will be recorded if it doesn't exist yet. Given below two exprs: |
3077 | inv_expr + cst1, inv_expr + cst2 |
3078 | It's hard to make decision whether constant part should be stripped |
3079 | or not. We choose to not strip based on below facts: |
3080 | 1) We need to count ADD cost for constant part if it's stripped, |
3081 | which isn't always trivial where this functions is called. |
3082 | 2) Stripping constant away may be conflict with following loop |
3083 | invariant hoisting pass. |
3084 | 3) Not stripping constant away results in more invariant exprs, |
3085 | which usually leads to decision preferring lower reg pressure. */ |
3086 | |
3087 | static iv_inv_expr_ent * |
3088 | get_loop_invariant_expr (struct ivopts_data *data, tree inv_expr) |
3089 | { |
3090 | STRIP_NOPS (inv_expr); |
3091 | |
3092 | if (poly_int_tree_p (t: inv_expr) |
3093 | || TREE_CODE (inv_expr) == SSA_NAME) |
3094 | return NULL; |
3095 | |
3096 | /* Don't strip constant part away as we used to. */ |
3097 | |
3098 | /* Stores EXPR in DATA->inv_expr_tab, return pointer to iv_inv_expr_ent. */ |
3099 | struct iv_inv_expr_ent ent; |
3100 | ent.expr = inv_expr; |
3101 | ent.hash = iterative_hash_expr (tree: inv_expr, seed: 0); |
3102 | struct iv_inv_expr_ent **slot = data->inv_expr_tab->find_slot (value: &ent, insert: INSERT); |
3103 | |
3104 | if (!*slot) |
3105 | { |
3106 | *slot = XNEW (struct iv_inv_expr_ent); |
3107 | (*slot)->expr = inv_expr; |
3108 | (*slot)->hash = ent.hash; |
3109 | (*slot)->id = ++data->max_inv_expr_id; |
3110 | } |
3111 | |
3112 | return *slot; |
3113 | } |
3114 | |
3115 | |
3116 | /* Return *TP if it is an SSA_NAME marked with TREE_VISITED, i.e., as |
3117 | unsuitable as ivopts candidates for potentially involving undefined |
3118 | behavior. */ |
3119 | |
3120 | static tree |
3121 | find_ssa_undef (tree *tp, int *walk_subtrees, void *bb_) |
3122 | { |
3123 | basic_block bb = (basic_block) bb_; |
3124 | if (TREE_CODE (*tp) == SSA_NAME |
3125 | && ssa_name_maybe_undef_p (var: *tp) |
3126 | && !ssa_name_any_use_dominates_bb_p (var: *tp, bb)) |
3127 | return *tp; |
3128 | if (!EXPR_P (*tp)) |
3129 | *walk_subtrees = 0; |
3130 | return NULL; |
3131 | } |
3132 | |
3133 | /* Adds a candidate BASE + STEP * i. Important field is set to IMPORTANT and |
3134 | position to POS. If USE is not NULL, the candidate is set as related to |
3135 | it. If both BASE and STEP are NULL, we add a pseudocandidate for the |
3136 | replacement of the final value of the iv by a direct computation. */ |
3137 | |
3138 | static struct iv_cand * |
3139 | add_candidate_1 (struct ivopts_data *data, tree base, tree step, bool important, |
3140 | enum iv_position pos, struct iv_use *use, |
3141 | gimple *incremented_at, struct iv *orig_iv = NULL, |
3142 | bool doloop = false) |
3143 | { |
3144 | unsigned i; |
3145 | struct iv_cand *cand = NULL; |
3146 | tree type, orig_type; |
3147 | |
3148 | gcc_assert (base && step); |
3149 | |
3150 | /* -fkeep-gc-roots-live means that we have to keep a real pointer |
3151 | live, but the ivopts code may replace a real pointer with one |
3152 | pointing before or after the memory block that is then adjusted |
3153 | into the memory block during the loop. FIXME: It would likely be |
3154 | better to actually force the pointer live and still use ivopts; |
3155 | for example, it would be enough to write the pointer into memory |
3156 | and keep it there until after the loop. */ |
3157 | if (flag_keep_gc_roots_live && POINTER_TYPE_P (TREE_TYPE (base))) |
3158 | return NULL; |
3159 | |
3160 | /* If BASE contains undefined SSA names make sure we only record |
3161 | the original IV. */ |
3162 | bool involves_undefs = false; |
3163 | if (walk_tree (&base, find_ssa_undef, data->current_loop->header, NULL)) |
3164 | { |
3165 | if (pos != IP_ORIGINAL) |
3166 | return NULL; |
3167 | important = false; |
3168 | involves_undefs = true; |
3169 | } |
3170 | |
3171 | /* For non-original variables, make sure their values are computed in a type |
3172 | that does not invoke undefined behavior on overflows (since in general, |
3173 | we cannot prove that these induction variables are non-wrapping). */ |
3174 | if (pos != IP_ORIGINAL) |
3175 | { |
3176 | orig_type = TREE_TYPE (base); |
3177 | type = generic_type_for (type: orig_type); |
3178 | if (type != orig_type) |
3179 | { |
3180 | base = fold_convert (type, base); |
3181 | step = fold_convert (type, step); |
3182 | } |
3183 | } |
3184 | |
3185 | for (i = 0; i < data->vcands.length (); i++) |
3186 | { |
3187 | cand = data->vcands[i]; |
3188 | |
3189 | if (cand->pos != pos) |
3190 | continue; |
3191 | |
3192 | if (cand->incremented_at != incremented_at |
3193 | || ((pos == IP_AFTER_USE || pos == IP_BEFORE_USE) |
3194 | && cand->ainc_use != use)) |
3195 | continue; |
3196 | |
3197 | if (operand_equal_p (base, cand->iv->base, flags: 0) |
3198 | && operand_equal_p (step, cand->iv->step, flags: 0) |
3199 | && (TYPE_PRECISION (TREE_TYPE (base)) |
3200 | == TYPE_PRECISION (TREE_TYPE (cand->iv->base)))) |
3201 | break; |
3202 | } |
3203 | |
3204 | if (i == data->vcands.length ()) |
3205 | { |
3206 | cand = XCNEW (struct iv_cand); |
3207 | cand->id = i; |
3208 | cand->iv = alloc_iv (data, base, step); |
3209 | cand->pos = pos; |
3210 | if (pos != IP_ORIGINAL) |
3211 | { |
3212 | if (doloop) |
3213 | cand->var_before = create_tmp_var_raw (TREE_TYPE (base), "doloop" ); |
3214 | else |
3215 | cand->var_before = create_tmp_var_raw (TREE_TYPE (base), "ivtmp" ); |
3216 | cand->var_after = cand->var_before; |
3217 | } |
3218 | cand->important = important; |
3219 | cand->involves_undefs = involves_undefs; |
3220 | cand->incremented_at = incremented_at; |
3221 | cand->doloop_p = doloop; |
3222 | data->vcands.safe_push (obj: cand); |
3223 | |
3224 | if (!poly_int_tree_p (t: step)) |
3225 | { |
3226 | find_inv_vars (data, expr_p: &step, inv_vars: &cand->inv_vars); |
3227 | |
3228 | iv_inv_expr_ent *inv_expr = get_loop_invariant_expr (data, inv_expr: step); |
3229 | /* Share bitmap between inv_vars and inv_exprs for cand. */ |
3230 | if (inv_expr != NULL) |
3231 | { |
3232 | cand->inv_exprs = cand->inv_vars; |
3233 | cand->inv_vars = NULL; |
3234 | if (cand->inv_exprs) |
3235 | bitmap_clear (cand->inv_exprs); |
3236 | else |
3237 | cand->inv_exprs = BITMAP_ALLOC (NULL); |
3238 | |
3239 | bitmap_set_bit (cand->inv_exprs, inv_expr->id); |
3240 | } |
3241 | } |
3242 | |
3243 | if (pos == IP_AFTER_USE || pos == IP_BEFORE_USE) |
3244 | cand->ainc_use = use; |
3245 | else |
3246 | cand->ainc_use = NULL; |
3247 | |
3248 | cand->orig_iv = orig_iv; |
3249 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3250 | dump_cand (file: dump_file, cand); |
3251 | } |
3252 | |
3253 | cand->important |= important; |
3254 | cand->doloop_p |= doloop; |
3255 | |
3256 | /* Relate candidate to the group for which it is added. */ |
3257 | if (use) |
3258 | bitmap_set_bit (data->vgroups[use->group_id]->related_cands, i); |
3259 | |
3260 | return cand; |
3261 | } |
3262 | |
3263 | /* Returns true if incrementing the induction variable at the end of the LOOP |
3264 | is allowed. |
3265 | |
3266 | The purpose is to avoid splitting latch edge with a biv increment, thus |
3267 | creating a jump, possibly confusing other optimization passes and leaving |
3268 | less freedom to scheduler. So we allow IP_END only if IP_NORMAL is not |
3269 | available (so we do not have a better alternative), or if the latch edge |
3270 | is already nonempty. */ |
3271 | |
3272 | static bool |
3273 | allow_ip_end_pos_p (class loop *loop) |
3274 | { |
3275 | if (!ip_normal_pos (loop)) |
3276 | return true; |
3277 | |
3278 | if (!empty_block_p (ip_end_pos (loop))) |
3279 | return true; |
3280 | |
3281 | return false; |
3282 | } |
3283 | |
3284 | /* If possible, adds autoincrement candidates BASE + STEP * i based on use USE. |
3285 | Important field is set to IMPORTANT. */ |
3286 | |
3287 | static void |
3288 | add_autoinc_candidates (struct ivopts_data *data, tree base, tree step, |
3289 | bool important, struct iv_use *use) |
3290 | { |
3291 | basic_block use_bb = gimple_bb (g: use->stmt); |
3292 | machine_mode mem_mode; |
3293 | unsigned HOST_WIDE_INT cstepi; |
3294 | |
3295 | /* If we insert the increment in any position other than the standard |
3296 | ones, we must ensure that it is incremented once per iteration. |
3297 | It must not be in an inner nested loop, or one side of an if |
3298 | statement. */ |
3299 | if (use_bb->loop_father != data->current_loop |
3300 | || !dominated_by_p (CDI_DOMINATORS, data->current_loop->latch, use_bb) |
3301 | || stmt_can_throw_internal (cfun, use->stmt) |
3302 | || !cst_and_fits_in_hwi (step)) |
3303 | return; |
3304 | |
3305 | cstepi = int_cst_value (step); |
3306 | |
3307 | mem_mode = TYPE_MODE (use->mem_type); |
3308 | if (((USE_LOAD_PRE_INCREMENT (mem_mode) |
3309 | || USE_STORE_PRE_INCREMENT (mem_mode)) |
3310 | && known_eq (GET_MODE_SIZE (mem_mode), cstepi)) |
3311 | || ((USE_LOAD_PRE_DECREMENT (mem_mode) |
3312 | || USE_STORE_PRE_DECREMENT (mem_mode)) |
3313 | && known_eq (GET_MODE_SIZE (mem_mode), -cstepi))) |
3314 | { |
3315 | enum tree_code code = MINUS_EXPR; |
3316 | tree new_base; |
3317 | tree new_step = step; |
3318 | |
3319 | if (POINTER_TYPE_P (TREE_TYPE (base))) |
3320 | { |
3321 | new_step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step); |
3322 | code = POINTER_PLUS_EXPR; |
3323 | } |
3324 | else |
3325 | new_step = fold_convert (TREE_TYPE (base), new_step); |
3326 | new_base = fold_build2 (code, TREE_TYPE (base), base, new_step); |
3327 | add_candidate_1 (data, base: new_base, step, important, pos: IP_BEFORE_USE, use, |
3328 | incremented_at: use->stmt); |
3329 | } |
3330 | if (((USE_LOAD_POST_INCREMENT (mem_mode) |
3331 | || USE_STORE_POST_INCREMENT (mem_mode)) |
3332 | && known_eq (GET_MODE_SIZE (mem_mode), cstepi)) |
3333 | || ((USE_LOAD_POST_DECREMENT (mem_mode) |
3334 | || USE_STORE_POST_DECREMENT (mem_mode)) |
3335 | && known_eq (GET_MODE_SIZE (mem_mode), -cstepi))) |
3336 | { |
3337 | add_candidate_1 (data, base, step, important, pos: IP_AFTER_USE, use, |
3338 | incremented_at: use->stmt); |
3339 | } |
3340 | } |
3341 | |
3342 | /* Adds a candidate BASE + STEP * i. Important field is set to IMPORTANT and |
3343 | position to POS. If USE is not NULL, the candidate is set as related to |
3344 | it. The candidate computation is scheduled before exit condition and at |
3345 | the end of loop. */ |
3346 | |
3347 | static void |
3348 | add_candidate (struct ivopts_data *data, tree base, tree step, bool important, |
3349 | struct iv_use *use, struct iv *orig_iv = NULL, |
3350 | bool doloop = false) |
3351 | { |
3352 | if (ip_normal_pos (data->current_loop)) |
3353 | add_candidate_1 (data, base, step, important, pos: IP_NORMAL, use, NULL, orig_iv, |
3354 | doloop); |
3355 | /* Exclude doloop candidate here since it requires decrement then comparison |
3356 | and jump, the IP_END position doesn't match. */ |
3357 | if (!doloop && ip_end_pos (data->current_loop) |
3358 | && allow_ip_end_pos_p (loop: data->current_loop)) |
3359 | add_candidate_1 (data, base, step, important, pos: IP_END, use, NULL, orig_iv); |
3360 | } |
3361 | |
3362 | /* Adds standard iv candidates. */ |
3363 | |
3364 | static void |
3365 | add_standard_iv_candidates (struct ivopts_data *data) |
3366 | { |
3367 | add_candidate (data, integer_zero_node, integer_one_node, important: true, NULL); |
3368 | |
3369 | /* The same for a double-integer type if it is still fast enough. */ |
3370 | if (TYPE_PRECISION |
3371 | (long_integer_type_node) > TYPE_PRECISION (integer_type_node) |
3372 | && TYPE_PRECISION (long_integer_type_node) <= BITS_PER_WORD) |
3373 | add_candidate (data, base: build_int_cst (long_integer_type_node, 0), |
3374 | step: build_int_cst (long_integer_type_node, 1), important: true, NULL); |
3375 | |
3376 | /* The same for a double-integer type if it is still fast enough. */ |
3377 | if (TYPE_PRECISION |
3378 | (long_long_integer_type_node) > TYPE_PRECISION (long_integer_type_node) |
3379 | && TYPE_PRECISION (long_long_integer_type_node) <= BITS_PER_WORD) |
3380 | add_candidate (data, base: build_int_cst (long_long_integer_type_node, 0), |
3381 | step: build_int_cst (long_long_integer_type_node, 1), important: true, NULL); |
3382 | } |
3383 | |
3384 | |
3385 | /* Adds candidates bases on the old induction variable IV. */ |
3386 | |
3387 | static void |
3388 | add_iv_candidate_for_biv (struct ivopts_data *data, struct iv *iv) |
3389 | { |
3390 | gimple *phi; |
3391 | tree def; |
3392 | struct iv_cand *cand; |
3393 | |
3394 | /* Check if this biv is used in address type use. */ |
3395 | if (iv->no_overflow && iv->have_address_use |
3396 | && INTEGRAL_TYPE_P (TREE_TYPE (iv->base)) |
3397 | && TYPE_PRECISION (TREE_TYPE (iv->base)) < TYPE_PRECISION (sizetype)) |
3398 | { |
3399 | tree base = fold_convert (sizetype, iv->base); |
3400 | tree step = fold_convert (sizetype, iv->step); |
3401 | |
3402 | /* Add iv cand of same precision as index part in TARGET_MEM_REF. */ |
3403 | add_candidate (data, base, step, important: true, NULL, orig_iv: iv); |
3404 | /* Add iv cand of the original type only if it has nonlinear use. */ |
3405 | if (iv->nonlin_use) |
3406 | add_candidate (data, base: iv->base, step: iv->step, important: true, NULL); |
3407 | } |
3408 | else |
3409 | add_candidate (data, base: iv->base, step: iv->step, important: true, NULL); |
3410 | |
3411 | /* The same, but with initial value zero. */ |
3412 | if (POINTER_TYPE_P (TREE_TYPE (iv->base))) |
3413 | add_candidate (data, size_int (0), step: iv->step, important: true, NULL); |
3414 | else |
3415 | add_candidate (data, base: build_int_cst (TREE_TYPE (iv->base), 0), |
3416 | step: iv->step, important: true, NULL); |
3417 | |
3418 | phi = SSA_NAME_DEF_STMT (iv->ssa_name); |
3419 | if (gimple_code (g: phi) == GIMPLE_PHI) |
3420 | { |
3421 | /* Additionally record the possibility of leaving the original iv |
3422 | untouched. */ |
3423 | def = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (data->current_loop)); |
3424 | /* Don't add candidate if it's from another PHI node because |
3425 | it's an affine iv appearing in the form of PEELED_CHREC. */ |
3426 | phi = SSA_NAME_DEF_STMT (def); |
3427 | if (gimple_code (g: phi) != GIMPLE_PHI) |
3428 | { |
3429 | cand = add_candidate_1 (data, |
3430 | base: iv->base, step: iv->step, important: true, pos: IP_ORIGINAL, NULL, |
3431 | SSA_NAME_DEF_STMT (def)); |
3432 | if (cand) |
3433 | { |
3434 | cand->var_before = iv->ssa_name; |
3435 | cand->var_after = def; |
3436 | } |
3437 | } |
3438 | else |
3439 | gcc_assert (gimple_bb (phi) == data->current_loop->header); |
3440 | } |
3441 | } |
3442 | |
3443 | /* Adds candidates based on the old induction variables. */ |
3444 | |
3445 | static void |
3446 | add_iv_candidate_for_bivs (struct ivopts_data *data) |
3447 | { |
3448 | unsigned i; |
3449 | struct iv *iv; |
3450 | bitmap_iterator bi; |
3451 | |
3452 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, i, bi) |
3453 | { |
3454 | iv = ver_info (data, ver: i)->iv; |
3455 | if (iv && iv->biv_p && !integer_zerop (iv->step)) |
3456 | add_iv_candidate_for_biv (data, iv); |
3457 | } |
3458 | } |
3459 | |
3460 | /* Record common candidate {BASE, STEP} derived from USE in hashtable. */ |
3461 | |
3462 | static void |
3463 | record_common_cand (struct ivopts_data *data, tree base, |
3464 | tree step, struct iv_use *use) |
3465 | { |
3466 | class iv_common_cand ent; |
3467 | class iv_common_cand **slot; |
3468 | |
3469 | ent.base = base; |
3470 | ent.step = step; |
3471 | ent.hash = iterative_hash_expr (tree: base, seed: 0); |
3472 | ent.hash = iterative_hash_expr (tree: step, seed: ent.hash); |
3473 | |
3474 | slot = data->iv_common_cand_tab->find_slot (value: &ent, insert: INSERT); |
3475 | if (*slot == NULL) |
3476 | { |
3477 | *slot = new iv_common_cand (); |
3478 | (*slot)->base = base; |
3479 | (*slot)->step = step; |
3480 | (*slot)->uses.create (nelems: 8); |
3481 | (*slot)->hash = ent.hash; |
3482 | data->iv_common_cands.safe_push (obj: (*slot)); |
3483 | } |
3484 | |
3485 | gcc_assert (use != NULL); |
3486 | (*slot)->uses.safe_push (obj: use); |
3487 | return; |
3488 | } |
3489 | |
3490 | /* Comparison function used to sort common candidates. */ |
3491 | |
3492 | static int |
3493 | common_cand_cmp (const void *p1, const void *p2) |
3494 | { |
3495 | unsigned n1, n2; |
3496 | const class iv_common_cand *const *const ccand1 |
3497 | = (const class iv_common_cand *const *)p1; |
3498 | const class iv_common_cand *const *const ccand2 |
3499 | = (const class iv_common_cand *const *)p2; |
3500 | |
3501 | n1 = (*ccand1)->uses.length (); |
3502 | n2 = (*ccand2)->uses.length (); |
3503 | return n2 - n1; |
3504 | } |
3505 | |
3506 | /* Adds IV candidates based on common candidated recorded. */ |
3507 | |
3508 | static void |
3509 | add_iv_candidate_derived_from_uses (struct ivopts_data *data) |
3510 | { |
3511 | unsigned i, j; |
3512 | struct iv_cand *cand_1, *cand_2; |
3513 | |
3514 | data->iv_common_cands.qsort (common_cand_cmp); |
3515 | for (i = 0; i < data->iv_common_cands.length (); i++) |
3516 | { |
3517 | class iv_common_cand *ptr = data->iv_common_cands[i]; |
3518 | |
3519 | /* Only add IV candidate if it's derived from multiple uses. */ |
3520 | if (ptr->uses.length () <= 1) |
3521 | break; |
3522 | |
3523 | cand_1 = NULL; |
3524 | cand_2 = NULL; |
3525 | if (ip_normal_pos (data->current_loop)) |
3526 | cand_1 = add_candidate_1 (data, base: ptr->base, step: ptr->step, |
3527 | important: false, pos: IP_NORMAL, NULL, NULL); |
3528 | |
3529 | if (ip_end_pos (data->current_loop) |
3530 | && allow_ip_end_pos_p (loop: data->current_loop)) |
3531 | cand_2 = add_candidate_1 (data, base: ptr->base, step: ptr->step, |
3532 | important: false, pos: IP_END, NULL, NULL); |
3533 | |
3534 | /* Bind deriving uses and the new candidates. */ |
3535 | for (j = 0; j < ptr->uses.length (); j++) |
3536 | { |
3537 | struct iv_group *group = data->vgroups[ptr->uses[j]->group_id]; |
3538 | if (cand_1) |
3539 | bitmap_set_bit (group->related_cands, cand_1->id); |
3540 | if (cand_2) |
3541 | bitmap_set_bit (group->related_cands, cand_2->id); |
3542 | } |
3543 | } |
3544 | |
3545 | /* Release data since it is useless from this point. */ |
3546 | data->iv_common_cand_tab->empty (); |
3547 | data->iv_common_cands.truncate (size: 0); |
3548 | } |
3549 | |
3550 | /* Adds candidates based on the value of USE's iv. */ |
3551 | |
3552 | static void |
3553 | add_iv_candidate_for_use (struct ivopts_data *data, struct iv_use *use) |
3554 | { |
3555 | poly_uint64 offset; |
3556 | tree base; |
3557 | struct iv *iv = use->iv; |
3558 | tree basetype = TREE_TYPE (iv->base); |
3559 | |
3560 | /* Don't add candidate for iv_use with non integer, pointer or non-mode |
3561 | precision types, instead, add candidate for the corresponding scev in |
3562 | unsigned type with the same precision. See PR93674 for more info. */ |
3563 | if ((TREE_CODE (basetype) != INTEGER_TYPE && !POINTER_TYPE_P (basetype)) |
3564 | || !type_has_mode_precision_p (t: basetype)) |
3565 | { |
3566 | basetype = lang_hooks.types.type_for_mode (TYPE_MODE (basetype), |
3567 | TYPE_UNSIGNED (basetype)); |
3568 | add_candidate (data, fold_convert (basetype, iv->base), |
3569 | fold_convert (basetype, iv->step), important: false, NULL); |
3570 | return; |
3571 | } |
3572 | |
3573 | add_candidate (data, base: iv->base, step: iv->step, important: false, use); |
3574 | |
3575 | /* Record common candidate for use in case it can be shared by others. */ |
3576 | record_common_cand (data, base: iv->base, step: iv->step, use); |
3577 | |
3578 | /* Record common candidate with initial value zero. */ |
3579 | basetype = TREE_TYPE (iv->base); |
3580 | if (POINTER_TYPE_P (basetype)) |
3581 | basetype = sizetype; |
3582 | record_common_cand (data, base: build_int_cst (basetype, 0), step: iv->step, use); |
3583 | |
3584 | /* Compare the cost of an address with an unscaled index with the cost of |
3585 | an address with a scaled index and add candidate if useful. */ |
3586 | poly_int64 step; |
3587 | if (use != NULL |
3588 | && poly_int_tree_p (t: iv->step, value: &step) |
3589 | && address_p (type: use->type)) |
3590 | { |
3591 | poly_int64 new_step; |
3592 | unsigned int fact = preferred_mem_scale_factor |
3593 | (base: use->iv->base, |
3594 | TYPE_MODE (use->mem_type), |
3595 | speed: optimize_loop_for_speed_p (data->current_loop)); |
3596 | |
3597 | if (fact != 1 |
3598 | && multiple_p (a: step, b: fact, multiple: &new_step)) |
3599 | add_candidate (data, size_int (0), |
3600 | step: wide_int_to_tree (sizetype, cst: new_step), |
3601 | important: true, NULL); |
3602 | } |
3603 | |
3604 | /* Record common candidate with constant offset stripped in base. |
3605 | Like the use itself, we also add candidate directly for it. */ |
3606 | base = strip_offset (expr: iv->base, offset: &offset); |
3607 | if (maybe_ne (a: offset, b: 0U) || base != iv->base) |
3608 | { |
3609 | record_common_cand (data, base, step: iv->step, use); |
3610 | add_candidate (data, base, step: iv->step, important: false, use); |
3611 | } |
3612 | |
3613 | /* Record common candidate with base_object removed in base. */ |
3614 | base = iv->base; |
3615 | STRIP_NOPS (base); |
3616 | if (iv->base_object != NULL && TREE_CODE (base) == POINTER_PLUS_EXPR) |
3617 | { |
3618 | tree step = iv->step; |
3619 | |
3620 | STRIP_NOPS (step); |
3621 | base = TREE_OPERAND (base, 1); |
3622 | step = fold_convert (sizetype, step); |
3623 | record_common_cand (data, base, step, use); |
3624 | /* Also record common candidate with offset stripped. */ |
3625 | tree alt_base, alt_offset; |
3626 | split_constant_offset (base, &alt_base, &alt_offset); |
3627 | if (!integer_zerop (alt_offset)) |
3628 | record_common_cand (data, base: alt_base, step, use); |
3629 | } |
3630 | |
3631 | /* At last, add auto-incremental candidates. Make such variables |
3632 | important since other iv uses with same base object may be based |
3633 | on it. */ |
3634 | if (use != NULL && address_p (type: use->type)) |
3635 | add_autoinc_candidates (data, base: iv->base, step: iv->step, important: true, use); |
3636 | } |
3637 | |
3638 | /* Adds candidates based on the uses. */ |
3639 | |
3640 | static void |
3641 | add_iv_candidate_for_groups (struct ivopts_data *data) |
3642 | { |
3643 | unsigned i; |
3644 | |
3645 | /* Only add candidate for the first use in group. */ |
3646 | for (i = 0; i < data->vgroups.length (); i++) |
3647 | { |
3648 | struct iv_group *group = data->vgroups[i]; |
3649 | |
3650 | gcc_assert (group->vuses[0] != NULL); |
3651 | add_iv_candidate_for_use (data, use: group->vuses[0]); |
3652 | } |
3653 | add_iv_candidate_derived_from_uses (data); |
3654 | } |
3655 | |
3656 | /* Record important candidates and add them to related_cands bitmaps. */ |
3657 | |
3658 | static void |
3659 | record_important_candidates (struct ivopts_data *data) |
3660 | { |
3661 | unsigned i; |
3662 | struct iv_group *group; |
3663 | |
3664 | for (i = 0; i < data->vcands.length (); i++) |
3665 | { |
3666 | struct iv_cand *cand = data->vcands[i]; |
3667 | |
3668 | if (cand->important) |
3669 | bitmap_set_bit (data->important_candidates, i); |
3670 | } |
3671 | |
3672 | data->consider_all_candidates = (data->vcands.length () |
3673 | <= CONSIDER_ALL_CANDIDATES_BOUND); |
3674 | |
3675 | /* Add important candidates to groups' related_cands bitmaps. */ |
3676 | for (i = 0; i < data->vgroups.length (); i++) |
3677 | { |
3678 | group = data->vgroups[i]; |
3679 | bitmap_ior_into (group->related_cands, data->important_candidates); |
3680 | } |
3681 | } |
3682 | |
3683 | /* Allocates the data structure mapping the (use, candidate) pairs to costs. |
3684 | If consider_all_candidates is true, we use a two-dimensional array, otherwise |
3685 | we allocate a simple list to every use. */ |
3686 | |
3687 | static void |
3688 | alloc_use_cost_map (struct ivopts_data *data) |
3689 | { |
3690 | unsigned i, size, s; |
3691 | |
3692 | for (i = 0; i < data->vgroups.length (); i++) |
3693 | { |
3694 | struct iv_group *group = data->vgroups[i]; |
3695 | |
3696 | if (data->consider_all_candidates) |
3697 | size = data->vcands.length (); |
3698 | else |
3699 | { |
3700 | s = bitmap_count_bits (group->related_cands); |
3701 | |
3702 | /* Round up to the power of two, so that moduling by it is fast. */ |
3703 | size = s ? (1 << ceil_log2 (x: s)) : 1; |
3704 | } |
3705 | |
3706 | group->n_map_members = size; |
3707 | group->cost_map = XCNEWVEC (class cost_pair, size); |
3708 | } |
3709 | } |
3710 | |
3711 | /* Sets cost of (GROUP, CAND) pair to COST and record that it depends |
3712 | on invariants INV_VARS and that the value used in expressing it is |
3713 | VALUE, and in case of iv elimination the comparison operator is COMP. */ |
3714 | |
3715 | static void |
3716 | set_group_iv_cost (struct ivopts_data *data, |
3717 | struct iv_group *group, struct iv_cand *cand, |
3718 | comp_cost cost, bitmap inv_vars, tree value, |
3719 | enum tree_code comp, bitmap inv_exprs) |
3720 | { |
3721 | unsigned i, s; |
3722 | |
3723 | if (cost.infinite_cost_p ()) |
3724 | { |
3725 | BITMAP_FREE (inv_vars); |
3726 | BITMAP_FREE (inv_exprs); |
3727 | return; |
3728 | } |
3729 | |
3730 | if (data->consider_all_candidates) |
3731 | { |
3732 | group->cost_map[cand->id].cand = cand; |
3733 | group->cost_map[cand->id].cost = cost; |
3734 | group->cost_map[cand->id].inv_vars = inv_vars; |
3735 | group->cost_map[cand->id].inv_exprs = inv_exprs; |
3736 | group->cost_map[cand->id].value = value; |
3737 | group->cost_map[cand->id].comp = comp; |
3738 | return; |
3739 | } |
3740 | |
3741 | /* n_map_members is a power of two, so this computes modulo. */ |
3742 | s = cand->id & (group->n_map_members - 1); |
3743 | for (i = s; i < group->n_map_members; i++) |
3744 | if (!group->cost_map[i].cand) |
3745 | goto found; |
3746 | for (i = 0; i < s; i++) |
3747 | if (!group->cost_map[i].cand) |
3748 | goto found; |
3749 | |
3750 | gcc_unreachable (); |
3751 | |
3752 | found: |
3753 | group->cost_map[i].cand = cand; |
3754 | group->cost_map[i].cost = cost; |
3755 | group->cost_map[i].inv_vars = inv_vars; |
3756 | group->cost_map[i].inv_exprs = inv_exprs; |
3757 | group->cost_map[i].value = value; |
3758 | group->cost_map[i].comp = comp; |
3759 | } |
3760 | |
3761 | /* Gets cost of (GROUP, CAND) pair. */ |
3762 | |
3763 | static class cost_pair * |
3764 | get_group_iv_cost (struct ivopts_data *data, struct iv_group *group, |
3765 | struct iv_cand *cand) |
3766 | { |
3767 | unsigned i, s; |
3768 | class cost_pair *ret; |
3769 | |
3770 | if (!cand) |
3771 | return NULL; |
3772 | |
3773 | if (data->consider_all_candidates) |
3774 | { |
3775 | ret = group->cost_map + cand->id; |
3776 | if (!ret->cand) |
3777 | return NULL; |
3778 | |
3779 | return ret; |
3780 | } |
3781 | |
3782 | /* n_map_members is a power of two, so this computes modulo. */ |
3783 | s = cand->id & (group->n_map_members - 1); |
3784 | for (i = s; i < group->n_map_members; i++) |
3785 | if (group->cost_map[i].cand == cand) |
3786 | return group->cost_map + i; |
3787 | else if (group->cost_map[i].cand == NULL) |
3788 | return NULL; |
3789 | for (i = 0; i < s; i++) |
3790 | if (group->cost_map[i].cand == cand) |
3791 | return group->cost_map + i; |
3792 | else if (group->cost_map[i].cand == NULL) |
3793 | return NULL; |
3794 | |
3795 | return NULL; |
3796 | } |
3797 | |
3798 | /* Produce DECL_RTL for object obj so it looks like it is stored in memory. */ |
3799 | static rtx |
3800 | produce_memory_decl_rtl (tree obj, int *regno) |
3801 | { |
3802 | addr_space_t as = TYPE_ADDR_SPACE (TREE_TYPE (obj)); |
3803 | machine_mode address_mode = targetm.addr_space.address_mode (as); |
3804 | rtx x; |
3805 | |
3806 | gcc_assert (obj); |
3807 | if (TREE_STATIC (obj) || DECL_EXTERNAL (obj)) |
3808 | { |
3809 | const char *name = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (obj)); |
3810 | x = gen_rtx_SYMBOL_REF (address_mode, name); |
3811 | SET_SYMBOL_REF_DECL (x, obj); |
3812 | x = gen_rtx_MEM (DECL_MODE (obj), x); |
3813 | set_mem_addr_space (x, as); |
3814 | targetm.encode_section_info (obj, x, true); |
3815 | } |
3816 | else |
3817 | { |
3818 | x = gen_raw_REG (address_mode, (*regno)++); |
3819 | x = gen_rtx_MEM (DECL_MODE (obj), x); |
3820 | set_mem_addr_space (x, as); |
3821 | } |
3822 | |
3823 | return x; |
3824 | } |
3825 | |
3826 | /* Prepares decl_rtl for variables referred in *EXPR_P. Callback for |
3827 | walk_tree. DATA contains the actual fake register number. */ |
3828 | |
3829 | static tree |
3830 | prepare_decl_rtl (tree *expr_p, int *ws, void *data) |
3831 | { |
3832 | tree obj = NULL_TREE; |
3833 | rtx x = NULL_RTX; |
3834 | int *regno = (int *) data; |
3835 | |
3836 | switch (TREE_CODE (*expr_p)) |
3837 | { |
3838 | case ADDR_EXPR: |
3839 | for (expr_p = &TREE_OPERAND (*expr_p, 0); |
3840 | handled_component_p (t: *expr_p); |
3841 | expr_p = &TREE_OPERAND (*expr_p, 0)) |
3842 | continue; |
3843 | obj = *expr_p; |
3844 | if (DECL_P (obj) && HAS_RTL_P (obj) && !DECL_RTL_SET_P (obj)) |
3845 | x = produce_memory_decl_rtl (obj, regno); |
3846 | break; |
3847 | |
3848 | case SSA_NAME: |
3849 | *ws = 0; |
3850 | obj = SSA_NAME_VAR (*expr_p); |
3851 | /* Defer handling of anonymous SSA_NAMEs to the expander. */ |
3852 | if (!obj) |
3853 | return NULL_TREE; |
3854 | if (!DECL_RTL_SET_P (obj)) |
3855 | x = gen_raw_REG (DECL_MODE (obj), (*regno)++); |
3856 | break; |
3857 | |
3858 | case VAR_DECL: |
3859 | case PARM_DECL: |
3860 | case RESULT_DECL: |
3861 | *ws = 0; |
3862 | obj = *expr_p; |
3863 | |
3864 | if (DECL_RTL_SET_P (obj)) |
3865 | break; |
3866 | |
3867 | if (DECL_MODE (obj) == BLKmode) |
3868 | x = produce_memory_decl_rtl (obj, regno); |
3869 | else |
3870 | x = gen_raw_REG (DECL_MODE (obj), (*regno)++); |
3871 | |
3872 | break; |
3873 | |
3874 | default: |
3875 | break; |
3876 | } |
3877 | |
3878 | if (x) |
3879 | { |
3880 | decl_rtl_to_reset.safe_push (obj); |
3881 | SET_DECL_RTL (obj, x); |
3882 | } |
3883 | |
3884 | return NULL_TREE; |
3885 | } |
3886 | |
3887 | /* Predict whether the given loop will be transformed in the RTL |
3888 | doloop_optimize pass. Attempt to duplicate some doloop_optimize checks. |
3889 | This is only for target independent checks, see targetm.predict_doloop_p |
3890 | for the target dependent ones. |
3891 | |
3892 | Note that according to some initial investigation, some checks like costly |
3893 | niter check and invalid stmt scanning don't have much gains among general |
3894 | cases, so keep this as simple as possible first. |
3895 | |
3896 | Some RTL specific checks seems unable to be checked in gimple, if any new |
3897 | checks or easy checks _are_ missing here, please add them. */ |
3898 | |
3899 | static bool |
3900 | generic_predict_doloop_p (struct ivopts_data *data) |
3901 | { |
3902 | class loop *loop = data->current_loop; |
3903 | |
3904 | /* Call target hook for target dependent checks. */ |
3905 | if (!targetm.predict_doloop_p (loop)) |
3906 | { |
3907 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3908 | fprintf (stream: dump_file, format: "Predict doloop failure due to" |
3909 | " target specific checks.\n" ); |
3910 | return false; |
3911 | } |
3912 | |
3913 | /* Similar to doloop_optimize, check iteration description to know it's |
3914 | suitable or not. Keep it as simple as possible, feel free to extend it |
3915 | if you find any multiple exits cases matter. */ |
3916 | edge exit = single_dom_exit (loop); |
3917 | class tree_niter_desc *niter_desc; |
3918 | if (!exit || !(niter_desc = niter_for_exit (data, exit))) |
3919 | { |
3920 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3921 | fprintf (stream: dump_file, format: "Predict doloop failure due to" |
3922 | " unexpected niters.\n" ); |
3923 | return false; |
3924 | } |
3925 | |
3926 | /* Similar to doloop_optimize, check whether iteration count too small |
3927 | and not profitable. */ |
3928 | HOST_WIDE_INT est_niter = get_estimated_loop_iterations_int (loop); |
3929 | if (est_niter == -1) |
3930 | est_niter = get_likely_max_loop_iterations_int (loop); |
3931 | if (est_niter >= 0 && est_niter < 3) |
3932 | { |
3933 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3934 | fprintf (stream: dump_file, |
3935 | format: "Predict doloop failure due to" |
3936 | " too few iterations (%u).\n" , |
3937 | (unsigned int) est_niter); |
3938 | return false; |
3939 | } |
3940 | |
3941 | return true; |
3942 | } |
3943 | |
3944 | /* Determines cost of the computation of EXPR. */ |
3945 | |
3946 | static unsigned |
3947 | computation_cost (tree expr, bool speed) |
3948 | { |
3949 | rtx_insn *seq; |
3950 | rtx rslt; |
3951 | tree type = TREE_TYPE (expr); |
3952 | unsigned cost; |
3953 | /* Avoid using hard regs in ways which may be unsupported. */ |
3954 | int regno = LAST_VIRTUAL_REGISTER + 1; |
3955 | struct cgraph_node *node = cgraph_node::get (decl: current_function_decl); |
3956 | enum node_frequency real_frequency = node->frequency; |
3957 | |
3958 | node->frequency = NODE_FREQUENCY_NORMAL; |
3959 | crtl->maybe_hot_insn_p = speed; |
3960 | walk_tree (&expr, prepare_decl_rtl, ®no, NULL); |
3961 | start_sequence (); |
3962 | rslt = expand_expr (exp: expr, NULL_RTX, TYPE_MODE (type), modifier: EXPAND_NORMAL); |
3963 | seq = get_insns (); |
3964 | end_sequence (); |
3965 | default_rtl_profile (); |
3966 | node->frequency = real_frequency; |
3967 | |
3968 | cost = seq_cost (seq, speed); |
3969 | if (MEM_P (rslt)) |
3970 | cost += address_cost (XEXP (rslt, 0), TYPE_MODE (type), |
3971 | TYPE_ADDR_SPACE (type), speed); |
3972 | else if (!REG_P (rslt)) |
3973 | cost += set_src_cost (x: rslt, TYPE_MODE (type), speed_p: speed); |
3974 | |
3975 | return cost; |
3976 | } |
3977 | |
3978 | /* Returns variable containing the value of candidate CAND at statement AT. */ |
3979 | |
3980 | static tree |
3981 | var_at_stmt (class loop *loop, struct iv_cand *cand, gimple *stmt) |
3982 | { |
3983 | if (stmt_after_increment (loop, cand, stmt)) |
3984 | return cand->var_after; |
3985 | else |
3986 | return cand->var_before; |
3987 | } |
3988 | |
3989 | /* If A is (TYPE) BA and B is (TYPE) BB, and the types of BA and BB have the |
3990 | same precision that is at least as wide as the precision of TYPE, stores |
3991 | BA to A and BB to B, and returns the type of BA. Otherwise, returns the |
3992 | type of A and B. */ |
3993 | |
3994 | static tree |
3995 | determine_common_wider_type (tree *a, tree *b) |
3996 | { |
3997 | tree wider_type = NULL; |
3998 | tree suba, subb; |
3999 | tree atype = TREE_TYPE (*a); |
4000 | |
4001 | if (CONVERT_EXPR_P (*a)) |
4002 | { |
4003 | suba = TREE_OPERAND (*a, 0); |
4004 | wider_type = TREE_TYPE (suba); |
4005 | if (TYPE_PRECISION (wider_type) < TYPE_PRECISION (atype)) |
4006 | return atype; |
4007 | } |
4008 | else |
4009 | return atype; |
4010 | |
4011 | if (CONVERT_EXPR_P (*b)) |
4012 | { |
4013 | subb = TREE_OPERAND (*b, 0); |
4014 | if (TYPE_PRECISION (wider_type) != TYPE_PRECISION (TREE_TYPE (subb))) |
4015 | return atype; |
4016 | } |
4017 | else |
4018 | return atype; |
4019 | |
4020 | *a = suba; |
4021 | *b = subb; |
4022 | return wider_type; |
4023 | } |
4024 | |
4025 | /* Determines the expression by that USE is expressed from induction variable |
4026 | CAND at statement AT in LOOP. The expression is stored in two parts in a |
4027 | decomposed form. The invariant part is stored in AFF_INV; while variant |
4028 | part in AFF_VAR. Store ratio of CAND.step over USE.step in PRAT if it's |
4029 | non-null. Returns false if USE cannot be expressed using CAND. */ |
4030 | |
4031 | static bool |
4032 | get_computation_aff_1 (class loop *loop, gimple *at, struct iv_use *use, |
4033 | struct iv_cand *cand, class aff_tree *aff_inv, |
4034 | class aff_tree *aff_var, widest_int *prat = NULL) |
4035 | { |
4036 | tree ubase = use->iv->base, ustep = use->iv->step; |
4037 | tree cbase = cand->iv->base, cstep = cand->iv->step; |
4038 | tree common_type, uutype, var, cstep_common; |
4039 | tree utype = TREE_TYPE (ubase), ctype = TREE_TYPE (cbase); |
4040 | aff_tree aff_cbase; |
4041 | widest_int rat; |
4042 | |
4043 | /* We must have a precision to express the values of use. */ |
4044 | if (TYPE_PRECISION (utype) > TYPE_PRECISION (ctype)) |
4045 | return false; |
4046 | |
4047 | var = var_at_stmt (loop, cand, stmt: at); |
4048 | uutype = unsigned_type_for (utype); |
4049 | |
4050 | /* If the conversion is not noop, perform it. */ |
4051 | if (TYPE_PRECISION (utype) < TYPE_PRECISION (ctype)) |
4052 | { |
4053 | if (cand->orig_iv != NULL && CONVERT_EXPR_P (cbase) |
4054 | && (CONVERT_EXPR_P (cstep) || poly_int_tree_p (t: cstep))) |
4055 | { |
4056 | tree inner_base, inner_step, inner_type; |
4057 | inner_base = TREE_OPERAND (cbase, 0); |
4058 | if (CONVERT_EXPR_P (cstep)) |
4059 | inner_step = TREE_OPERAND (cstep, 0); |
4060 | else |
4061 | inner_step = cstep; |
4062 | |
4063 | inner_type = TREE_TYPE (inner_base); |
4064 | /* If candidate is added from a biv whose type is smaller than |
4065 | ctype, we know both candidate and the biv won't overflow. |
4066 | In this case, it's safe to skip the convertion in candidate. |
4067 | As an example, (unsigned short)((unsigned long)A) equals to |
4068 | (unsigned short)A, if A has a type no larger than short. */ |
4069 | if (TYPE_PRECISION (inner_type) <= TYPE_PRECISION (uutype)) |
4070 | { |
4071 | cbase = inner_base; |
4072 | cstep = inner_step; |
4073 | } |
4074 | } |
4075 | cbase = fold_convert (uutype, cbase); |
4076 | cstep = fold_convert (uutype, cstep); |
4077 | var = fold_convert (uutype, var); |
4078 | } |
4079 | |
4080 | /* Ratio is 1 when computing the value of biv cand by itself. |
4081 | We can't rely on constant_multiple_of in this case because the |
4082 | use is created after the original biv is selected. The call |
4083 | could fail because of inconsistent fold behavior. See PR68021 |
4084 | for more information. */ |
4085 | if (cand->pos == IP_ORIGINAL && cand->incremented_at == use->stmt) |
4086 | { |
4087 | gcc_assert (is_gimple_assign (use->stmt)); |
4088 | gcc_assert (use->iv->ssa_name == cand->var_after); |
4089 | gcc_assert (gimple_assign_lhs (use->stmt) == cand->var_after); |
4090 | rat = 1; |
4091 | } |
4092 | else if (!constant_multiple_of (top: ustep, bot: cstep, mul: &rat)) |
4093 | return false; |
4094 | |
4095 | if (prat) |
4096 | *prat = rat; |
4097 | |
4098 | /* In case both UBASE and CBASE are shortened to UUTYPE from some common |
4099 | type, we achieve better folding by computing their difference in this |
4100 | wider type, and cast the result to UUTYPE. We do not need to worry about |
4101 | overflows, as all the arithmetics will in the end be performed in UUTYPE |
4102 | anyway. */ |
4103 | common_type = determine_common_wider_type (a: &ubase, b: &cbase); |
4104 | |
4105 | /* use = ubase - ratio * cbase + ratio * var. */ |
4106 | tree_to_aff_combination (ubase, common_type, aff_inv); |
4107 | tree_to_aff_combination (cbase, common_type, &aff_cbase); |
4108 | tree_to_aff_combination (var, uutype, aff_var); |
4109 | |
4110 | /* We need to shift the value if we are after the increment. */ |
4111 | if (stmt_after_increment (loop, cand, stmt: at)) |
4112 | { |
4113 | aff_tree cstep_aff; |
4114 | |
4115 | if (common_type != uutype) |
4116 | cstep_common = fold_convert (common_type, cstep); |
4117 | else |
4118 | cstep_common = cstep; |
4119 | |
4120 | tree_to_aff_combination (cstep_common, common_type, &cstep_aff); |
4121 | aff_combination_add (&aff_cbase, &cstep_aff); |
4122 | } |
4123 | |
4124 | aff_combination_scale (&aff_cbase, -rat); |
4125 | aff_combination_add (aff_inv, &aff_cbase); |
4126 | if (common_type != uutype) |
4127 | aff_combination_convert (aff_inv, uutype); |
4128 | |
4129 | aff_combination_scale (aff_var, rat); |
4130 | return true; |
4131 | } |
4132 | |
4133 | /* Determines the expression by that USE is expressed from induction variable |
4134 | CAND at statement AT in LOOP. The expression is stored in a decomposed |
4135 | form into AFF. Returns false if USE cannot be expressed using CAND. */ |
4136 | |
4137 | static bool |
4138 | get_computation_aff (class loop *loop, gimple *at, struct iv_use *use, |
4139 | struct iv_cand *cand, class aff_tree *aff) |
4140 | { |
4141 | aff_tree aff_var; |
4142 | |
4143 | if (!get_computation_aff_1 (loop, at, use, cand, aff_inv: aff, aff_var: &aff_var)) |
4144 | return false; |
4145 | |
4146 | aff_combination_add (aff, &aff_var); |
4147 | return true; |
4148 | } |
4149 | |
4150 | /* Return the type of USE. */ |
4151 | |
4152 | static tree |
4153 | get_use_type (struct iv_use *use) |
4154 | { |
4155 | tree base_type = TREE_TYPE (use->iv->base); |
4156 | tree type; |
4157 | |
4158 | if (use->type == USE_REF_ADDRESS) |
4159 | { |
4160 | /* The base_type may be a void pointer. Create a pointer type based on |
4161 | the mem_ref instead. */ |
4162 | type = build_pointer_type (TREE_TYPE (*use->op_p)); |
4163 | gcc_assert (TYPE_ADDR_SPACE (TREE_TYPE (type)) |
4164 | == TYPE_ADDR_SPACE (TREE_TYPE (base_type))); |
4165 | } |
4166 | else |
4167 | type = base_type; |
4168 | |
4169 | return type; |
4170 | } |
4171 | |
4172 | /* Determines the expression by that USE is expressed from induction variable |
4173 | CAND at statement AT in LOOP. The computation is unshared. */ |
4174 | |
4175 | static tree |
4176 | get_computation_at (class loop *loop, gimple *at, |
4177 | struct iv_use *use, struct iv_cand *cand) |
4178 | { |
4179 | aff_tree aff; |
4180 | tree type = get_use_type (use); |
4181 | |
4182 | if (!get_computation_aff (loop, at, use, cand, aff: &aff)) |
4183 | return NULL_TREE; |
4184 | unshare_aff_combination (&aff); |
4185 | return fold_convert (type, aff_combination_to_tree (&aff)); |
4186 | } |
4187 | |
4188 | /* Like get_computation_at, but try harder, even if the computation |
4189 | is more expensive. Intended for debug stmts. */ |
4190 | |
4191 | static tree |
4192 | get_debug_computation_at (class loop *loop, gimple *at, |
4193 | struct iv_use *use, struct iv_cand *cand) |
4194 | { |
4195 | if (tree ret = get_computation_at (loop, at, use, cand)) |
4196 | return ret; |
4197 | |
4198 | tree ubase = use->iv->base, ustep = use->iv->step; |
4199 | tree cbase = cand->iv->base, cstep = cand->iv->step; |
4200 | tree var; |
4201 | tree utype = TREE_TYPE (ubase), ctype = TREE_TYPE (cbase); |
4202 | widest_int rat; |
4203 | |
4204 | /* We must have a precision to express the values of use. */ |
4205 | if (TYPE_PRECISION (utype) >= TYPE_PRECISION (ctype)) |
4206 | return NULL_TREE; |
4207 | |
4208 | /* Try to handle the case that get_computation_at doesn't, |
4209 | try to express |
4210 | use = ubase + (var - cbase) / ratio. */ |
4211 | if (!constant_multiple_of (top: cstep, fold_convert (TREE_TYPE (cstep), ustep), |
4212 | mul: &rat)) |
4213 | return NULL_TREE; |
4214 | |
4215 | bool neg_p = false; |
4216 | if (wi::neg_p (x: rat)) |
4217 | { |
4218 | if (TYPE_UNSIGNED (ctype)) |
4219 | return NULL_TREE; |
4220 | neg_p = true; |
4221 | rat = wi::neg (x: rat); |
4222 | } |
4223 | |
4224 | /* If both IVs can wrap around and CAND doesn't have a power of two step, |
4225 | it is unsafe. Consider uint16_t CAND with step 9, when wrapping around, |
4226 | the values will be ... 0xfff0, 0xfff9, 2, 11 ... and when use is say |
4227 | uint8_t with step 3, those values divided by 3 cast to uint8_t will be |
4228 | ... 0x50, 0x53, 0, 3 ... rather than expected 0x50, 0x53, 0x56, 0x59. */ |
4229 | if (!use->iv->no_overflow |
4230 | && !cand->iv->no_overflow |
4231 | && !integer_pow2p (cstep)) |
4232 | return NULL_TREE; |
4233 | |
4234 | int bits = wi::exact_log2 (rat); |
4235 | if (bits == -1) |
4236 | bits = wi::floor_log2 (rat) + 1; |
4237 | if (!cand->iv->no_overflow |
4238 | && TYPE_PRECISION (utype) + bits > TYPE_PRECISION (ctype)) |
4239 | return NULL_TREE; |
4240 | |
4241 | var = var_at_stmt (loop, cand, stmt: at); |
4242 | |
4243 | if (POINTER_TYPE_P (ctype)) |
4244 | { |
4245 | ctype = unsigned_type_for (ctype); |
4246 | cbase = fold_convert (ctype, cbase); |
4247 | cstep = fold_convert (ctype, cstep); |
4248 | var = fold_convert (ctype, var); |
4249 | } |
4250 | |
4251 | if (stmt_after_increment (loop, cand, stmt: at)) |
4252 | var = fold_build2 (MINUS_EXPR, TREE_TYPE (var), var, |
4253 | unshare_expr (cstep)); |
4254 | |
4255 | var = fold_build2 (MINUS_EXPR, TREE_TYPE (var), var, cbase); |
4256 | var = fold_build2 (EXACT_DIV_EXPR, TREE_TYPE (var), var, |
4257 | wide_int_to_tree (TREE_TYPE (var), rat)); |
4258 | if (POINTER_TYPE_P (utype)) |
4259 | { |
4260 | var = fold_convert (sizetype, var); |
4261 | if (neg_p) |
4262 | var = fold_build1 (NEGATE_EXPR, sizetype, var); |
4263 | var = fold_build2 (POINTER_PLUS_EXPR, utype, ubase, var); |
4264 | } |
4265 | else |
4266 | { |
4267 | var = fold_convert (utype, var); |
4268 | var = fold_build2 (neg_p ? MINUS_EXPR : PLUS_EXPR, utype, |
4269 | ubase, var); |
4270 | } |
4271 | return var; |
4272 | } |
4273 | |
4274 | /* Adjust the cost COST for being in loop setup rather than loop body. |
4275 | If we're optimizing for space, the loop setup overhead is constant; |
4276 | if we're optimizing for speed, amortize it over the per-iteration cost. |
4277 | If ROUND_UP_P is true, the result is round up rather than to zero when |
4278 | optimizing for speed. */ |
4279 | static int64_t |
4280 | adjust_setup_cost (struct ivopts_data *data, int64_t cost, |
4281 | bool round_up_p = false) |
4282 | { |
4283 | if (cost == INFTY) |
4284 | return cost; |
4285 | else if (optimize_loop_for_speed_p (data->current_loop)) |
4286 | { |
4287 | int64_t niters = (int64_t) avg_loop_niter (loop: data->current_loop); |
4288 | return (cost + (round_up_p ? niters - 1 : 0)) / niters; |
4289 | } |
4290 | else |
4291 | return cost; |
4292 | } |
4293 | |
4294 | /* Calculate the SPEED or size cost of shiftadd EXPR in MODE. MULT is the |
4295 | EXPR operand holding the shift. COST0 and COST1 are the costs for |
4296 | calculating the operands of EXPR. Returns true if successful, and returns |
4297 | the cost in COST. */ |
4298 | |
4299 | static bool |
4300 | get_shiftadd_cost (tree expr, scalar_int_mode mode, comp_cost cost0, |
4301 | comp_cost cost1, tree mult, bool speed, comp_cost *cost) |
4302 | { |
4303 | comp_cost res; |
4304 | tree op1 = TREE_OPERAND (expr, 1); |
4305 | tree cst = TREE_OPERAND (mult, 1); |
4306 | tree multop = TREE_OPERAND (mult, 0); |
4307 | int m = exact_log2 (x: int_cst_value (cst)); |
4308 | int maxm = MIN (BITS_PER_WORD, GET_MODE_BITSIZE (mode)); |
4309 | int as_cost, sa_cost; |
4310 | bool mult_in_op1; |
4311 | |
4312 | if (!(m >= 0 && m < maxm)) |
4313 | return false; |
4314 | |
4315 | STRIP_NOPS (op1); |
4316 | mult_in_op1 = operand_equal_p (op1, mult, flags: 0); |
4317 | |
4318 | as_cost = add_cost (speed, mode) + shift_cost (speed, mode, bits: m); |
4319 | |
4320 | /* If the target has a cheap shift-and-add or shift-and-sub instruction, |
4321 | use that in preference to a shift insn followed by an add insn. */ |
4322 | sa_cost = (TREE_CODE (expr) != MINUS_EXPR |
4323 | ? shiftadd_cost (speed, mode, bits: m) |
4324 | : (mult_in_op1 |
4325 | ? shiftsub1_cost (speed, mode, bits: m) |
4326 | : shiftsub0_cost (speed, mode, bits: m))); |
4327 | |
4328 | res = comp_cost (MIN (as_cost, sa_cost), 0); |
4329 | res += (mult_in_op1 ? cost0 : cost1); |
4330 | |
4331 | STRIP_NOPS (multop); |
4332 | if (!is_gimple_val (multop)) |
4333 | res += force_expr_to_var_cost (multop, speed); |
4334 | |
4335 | *cost = res; |
4336 | return true; |
4337 | } |
4338 | |
4339 | /* Estimates cost of forcing expression EXPR into a variable. */ |
4340 | |
4341 | static comp_cost |
4342 | force_expr_to_var_cost (tree expr, bool speed) |
4343 | { |
4344 | static bool costs_initialized = false; |
4345 | static unsigned integer_cost [2]; |
4346 | static unsigned symbol_cost [2]; |
4347 | static unsigned address_cost [2]; |
4348 | tree op0, op1; |
4349 | comp_cost cost0, cost1, cost; |
4350 | machine_mode mode; |
4351 | scalar_int_mode int_mode; |
4352 | |
4353 | if (!costs_initialized) |
4354 | { |
4355 | tree type = build_pointer_type (integer_type_node); |
4356 | tree var, addr; |
4357 | rtx x; |
4358 | int i; |
4359 | |
4360 | var = create_tmp_var_raw (integer_type_node, "test_var" ); |
4361 | TREE_STATIC (var) = 1; |
4362 | x = produce_memory_decl_rtl (obj: var, NULL); |
4363 | SET_DECL_RTL (var, x); |
4364 | |
4365 | addr = build1 (ADDR_EXPR, type, var); |
4366 | |
4367 | |
4368 | for (i = 0; i < 2; i++) |
4369 | { |
4370 | integer_cost[i] = computation_cost (expr: build_int_cst (integer_type_node, |
4371 | 2000), speed: i); |
4372 | |
4373 | symbol_cost[i] = computation_cost (expr: addr, speed: i) + 1; |
4374 | |
4375 | address_cost[i] |
4376 | = computation_cost (fold_build_pointer_plus_hwi (addr, 2000), speed: i) + 1; |
4377 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4378 | { |
4379 | fprintf (stream: dump_file, format: "force_expr_to_var_cost %s costs:\n" , i ? "speed" : "size" ); |
4380 | fprintf (stream: dump_file, format: " integer %d\n" , (int) integer_cost[i]); |
4381 | fprintf (stream: dump_file, format: " symbol %d\n" , (int) symbol_cost[i]); |
4382 | fprintf (stream: dump_file, format: " address %d\n" , (int) address_cost[i]); |
4383 | fprintf (stream: dump_file, format: " other %d\n" , (int) target_spill_cost[i]); |
4384 | fprintf (stream: dump_file, format: "\n" ); |
4385 | } |
4386 | } |
4387 | |
4388 | costs_initialized = true; |
4389 | } |
4390 | |
4391 | STRIP_NOPS (expr); |
4392 | |
4393 | if (SSA_VAR_P (expr)) |
4394 | return no_cost; |
4395 | |
4396 | if (is_gimple_min_invariant (expr)) |
4397 | { |
4398 | if (poly_int_tree_p (t: expr)) |
4399 | return comp_cost (integer_cost [speed], 0); |
4400 | |
4401 | if (TREE_CODE (expr) == ADDR_EXPR) |
4402 | { |
4403 | tree obj = TREE_OPERAND (expr, 0); |
4404 | |
4405 | if (VAR_P (obj) |
4406 | || TREE_CODE (obj) == PARM_DECL |
4407 | || TREE_CODE (obj) == RESULT_DECL) |
4408 | return comp_cost (symbol_cost [speed], 0); |
4409 | } |
4410 | |
4411 | return comp_cost (address_cost [speed], 0); |
4412 | } |
4413 | |
4414 | switch (TREE_CODE (expr)) |
4415 | { |
4416 | case POINTER_PLUS_EXPR: |
4417 | case PLUS_EXPR: |
4418 | case MINUS_EXPR: |
4419 | case MULT_EXPR: |
4420 | case TRUNC_DIV_EXPR: |
4421 | case BIT_AND_EXPR: |
4422 | case BIT_IOR_EXPR: |
4423 | case LSHIFT_EXPR: |
4424 | case RSHIFT_EXPR: |
4425 | op0 = TREE_OPERAND (expr, 0); |
4426 | op1 = TREE_OPERAND (expr, 1); |
4427 | STRIP_NOPS (op0); |
4428 | STRIP_NOPS (op1); |
4429 | break; |
4430 | |
4431 | CASE_CONVERT: |
4432 | case NEGATE_EXPR: |
4433 | case BIT_NOT_EXPR: |
4434 | op0 = TREE_OPERAND (expr, 0); |
4435 | STRIP_NOPS (op0); |
4436 | op1 = NULL_TREE; |
4437 | break; |
4438 | /* See add_iv_candidate_for_doloop, for doloop may_be_zero case, we |
4439 | introduce COND_EXPR for IV base, need to support better cost estimation |
4440 | for this COND_EXPR and tcc_comparison. */ |
4441 | case COND_EXPR: |
4442 | op0 = TREE_OPERAND (expr, 1); |
4443 | STRIP_NOPS (op0); |
4444 | op1 = TREE_OPERAND (expr, 2); |
4445 | STRIP_NOPS (op1); |
4446 | break; |
4447 | case LT_EXPR: |
4448 | case LE_EXPR: |
4449 | case GT_EXPR: |
4450 | case GE_EXPR: |
4451 | case EQ_EXPR: |
4452 | case NE_EXPR: |
4453 | case UNORDERED_EXPR: |
4454 | case ORDERED_EXPR: |
4455 | case UNLT_EXPR: |
4456 | case UNLE_EXPR: |
4457 | case UNGT_EXPR: |
4458 | case UNGE_EXPR: |
4459 | case UNEQ_EXPR: |
4460 | case LTGT_EXPR: |
4461 | case MAX_EXPR: |
4462 | case MIN_EXPR: |
4463 | op0 = TREE_OPERAND (expr, 0); |
4464 | STRIP_NOPS (op0); |
4465 | op1 = TREE_OPERAND (expr, 1); |
4466 | STRIP_NOPS (op1); |
4467 | break; |
4468 | |
4469 | default: |
4470 | /* Just an arbitrary value, FIXME. */ |
4471 | return comp_cost (target_spill_cost[speed], 0); |
4472 | } |
4473 | |
4474 | if (op0 == NULL_TREE |
4475 | || TREE_CODE (op0) == SSA_NAME || CONSTANT_CLASS_P (op0)) |
4476 | cost0 = no_cost; |
4477 | else |
4478 | cost0 = force_expr_to_var_cost (expr: op0, speed); |
4479 | |
4480 | if (op1 == NULL_TREE |
4481 | || TREE_CODE (op1) == SSA_NAME || CONSTANT_CLASS_P (op1)) |
4482 | cost1 = no_cost; |
4483 | else |
4484 | cost1 = force_expr_to_var_cost (expr: op1, speed); |
4485 | |
4486 | mode = TYPE_MODE (TREE_TYPE (expr)); |
4487 | switch (TREE_CODE (expr)) |
4488 | { |
4489 | case POINTER_PLUS_EXPR: |
4490 | case PLUS_EXPR: |
4491 | case MINUS_EXPR: |
4492 | case NEGATE_EXPR: |
4493 | cost = comp_cost (add_cost (speed, mode), 0); |
4494 | if (TREE_CODE (expr) != NEGATE_EXPR) |
4495 | { |
4496 | tree mult = NULL_TREE; |
4497 | comp_cost sa_cost; |
4498 | if (TREE_CODE (op1) == MULT_EXPR) |
4499 | mult = op1; |
4500 | else if (TREE_CODE (op0) == MULT_EXPR) |
4501 | mult = op0; |
4502 | |
4503 | if (mult != NULL_TREE |
4504 | && is_a <scalar_int_mode> (m: mode, result: &int_mode) |
4505 | && cst_and_fits_in_hwi (TREE_OPERAND (mult, 1)) |
4506 | && get_shiftadd_cost (expr, mode: int_mode, cost0, cost1, mult, |
4507 | speed, cost: &sa_cost)) |
4508 | return sa_cost; |
4509 | } |
4510 | break; |
4511 | |
4512 | CASE_CONVERT: |
4513 | { |
4514 | tree inner_mode, outer_mode; |
4515 | outer_mode = TREE_TYPE (expr); |
4516 | inner_mode = TREE_TYPE (op0); |
4517 | cost = comp_cost (convert_cost (TYPE_MODE (outer_mode), |
4518 | TYPE_MODE (inner_mode), speed), 0); |
4519 | } |
4520 | break; |
4521 | |
4522 | case MULT_EXPR: |
4523 | if (cst_and_fits_in_hwi (op0)) |
4524 | cost = comp_cost (mult_by_coeff_cost (int_cst_value (op0), |
4525 | mode, speed), 0); |
4526 | else if (cst_and_fits_in_hwi (op1)) |
4527 | cost = comp_cost (mult_by_coeff_cost (int_cst_value (op1), |
4528 | mode, speed), 0); |
4529 | else |
4530 | return comp_cost (target_spill_cost [speed], 0); |
4531 | break; |
4532 | |
4533 | case TRUNC_DIV_EXPR: |
4534 | /* Division by power of two is usually cheap, so we allow it. Forbid |
4535 | anything else. */ |
4536 | if (integer_pow2p (TREE_OPERAND (expr, 1))) |
4537 | cost = comp_cost (add_cost (speed, mode), 0); |
4538 | else |
4539 | cost = comp_cost (target_spill_cost[speed], 0); |
4540 | break; |
4541 | |
4542 | case BIT_AND_EXPR: |
4543 | case BIT_IOR_EXPR: |
4544 | case BIT_NOT_EXPR: |
4545 | case LSHIFT_EXPR: |
4546 | case RSHIFT_EXPR: |
4547 | cost = comp_cost (add_cost (speed, mode), 0); |
4548 | break; |
4549 | case COND_EXPR: |
4550 | op0 = TREE_OPERAND (expr, 0); |
4551 | STRIP_NOPS (op0); |
4552 | if (op0 == NULL_TREE || TREE_CODE (op0) == SSA_NAME |
4553 | || CONSTANT_CLASS_P (op0)) |
4554 | cost = no_cost; |
4555 | else |
4556 | cost = force_expr_to_var_cost (expr: op0, speed); |
4557 | break; |
4558 | case LT_EXPR: |
4559 | case LE_EXPR: |
4560 | case GT_EXPR: |
4561 | case GE_EXPR: |
4562 | case EQ_EXPR: |
4563 | case NE_EXPR: |
4564 | case UNORDERED_EXPR: |
4565 | case ORDERED_EXPR: |
4566 | case UNLT_EXPR: |
4567 | case UNLE_EXPR: |
4568 | case UNGT_EXPR: |
4569 | case UNGE_EXPR: |
4570 | case UNEQ_EXPR: |
4571 | case LTGT_EXPR: |
4572 | case MAX_EXPR: |
4573 | case MIN_EXPR: |
4574 | /* Simply use add cost for now, FIXME if there is some more accurate cost |
4575 | evaluation way. */ |
4576 | cost = comp_cost (add_cost (speed, mode), 0); |
4577 | break; |
4578 | |
4579 | default: |
4580 | gcc_unreachable (); |
4581 | } |
4582 | |
4583 | cost += cost0; |
4584 | cost += cost1; |
4585 | return cost; |
4586 | } |
4587 | |
4588 | /* Estimates cost of forcing EXPR into a variable. INV_VARS is a set of the |
4589 | invariants the computation depends on. */ |
4590 | |
4591 | static comp_cost |
4592 | force_var_cost (struct ivopts_data *data, tree expr, bitmap *inv_vars) |
4593 | { |
4594 | if (!expr) |
4595 | return no_cost; |
4596 | |
4597 | find_inv_vars (data, expr_p: &expr, inv_vars); |
4598 | return force_expr_to_var_cost (expr, speed: data->speed); |
4599 | } |
4600 | |
4601 | /* Returns cost of auto-modifying address expression in shape base + offset. |
4602 | AINC_STEP is step size of the address IV. AINC_OFFSET is offset of the |
4603 | address expression. The address expression has ADDR_MODE in addr space |
4604 | AS. The memory access has MEM_MODE. SPEED means we are optimizing for |
4605 | speed or size. */ |
4606 | |
4607 | enum ainc_type |
4608 | { |
4609 | AINC_PRE_INC, /* Pre increment. */ |
4610 | AINC_PRE_DEC, /* Pre decrement. */ |
4611 | AINC_POST_INC, /* Post increment. */ |
4612 | AINC_POST_DEC, /* Post decrement. */ |
4613 | AINC_NONE /* Also the number of auto increment types. */ |
4614 | }; |
4615 | |
4616 | struct ainc_cost_data |
4617 | { |
4618 | int64_t costs[AINC_NONE]; |
4619 | }; |
4620 | |
4621 | static comp_cost |
4622 | get_address_cost_ainc (poly_int64 ainc_step, poly_int64 ainc_offset, |
4623 | machine_mode addr_mode, machine_mode mem_mode, |
4624 | addr_space_t as, bool speed) |
4625 | { |
4626 | if (!USE_LOAD_PRE_DECREMENT (mem_mode) |
4627 | && !USE_STORE_PRE_DECREMENT (mem_mode) |
4628 | && !USE_LOAD_POST_DECREMENT (mem_mode) |
4629 | && !USE_STORE_POST_DECREMENT (mem_mode) |
4630 | && !USE_LOAD_PRE_INCREMENT (mem_mode) |
4631 | && !USE_STORE_PRE_INCREMENT (mem_mode) |
4632 | && !USE_LOAD_POST_INCREMENT (mem_mode) |
4633 | && !USE_STORE_POST_INCREMENT (mem_mode)) |
4634 | return infinite_cost; |
4635 | |
4636 | static vec<ainc_cost_data *> ainc_cost_data_list; |
4637 | unsigned idx = (unsigned) as * MAX_MACHINE_MODE + (unsigned) mem_mode; |
4638 | if (idx >= ainc_cost_data_list.length ()) |
4639 | { |
4640 | unsigned nsize = ((unsigned) as + 1) *MAX_MACHINE_MODE; |
4641 | |
4642 | gcc_assert (nsize > idx); |
4643 | ainc_cost_data_list.safe_grow_cleared (len: nsize, exact: true); |
4644 | } |
4645 | |
4646 | ainc_cost_data *data = ainc_cost_data_list[idx]; |
4647 | if (data == NULL) |
4648 | { |
4649 | rtx reg = gen_raw_REG (addr_mode, LAST_VIRTUAL_REGISTER + 1); |
4650 | |
4651 | data = (ainc_cost_data *) xcalloc (1, sizeof (*data)); |
4652 | data->costs[AINC_PRE_DEC] = INFTY; |
4653 | data->costs[AINC_POST_DEC] = INFTY; |
4654 | data->costs[AINC_PRE_INC] = INFTY; |
4655 | data->costs[AINC_POST_INC] = INFTY; |
4656 | if (USE_LOAD_PRE_DECREMENT (mem_mode) |
4657 | || USE_STORE_PRE_DECREMENT (mem_mode)) |
4658 | { |
4659 | rtx addr = gen_rtx_PRE_DEC (addr_mode, reg); |
4660 | |
4661 | if (memory_address_addr_space_p (mem_mode, addr, as)) |
4662 | data->costs[AINC_PRE_DEC] |
4663 | = address_cost (addr, mem_mode, as, speed); |
4664 | } |
4665 | if (USE_LOAD_POST_DECREMENT (mem_mode) |
4666 | || USE_STORE_POST_DECREMENT (mem_mode)) |
4667 | { |
4668 | rtx addr = gen_rtx_POST_DEC (addr_mode, reg); |
4669 | |
4670 | if (memory_address_addr_space_p (mem_mode, addr, as)) |
4671 | data->costs[AINC_POST_DEC] |
4672 | = address_cost (addr, mem_mode, as, speed); |
4673 | } |
4674 | if (USE_LOAD_PRE_INCREMENT (mem_mode) |
4675 | || USE_STORE_PRE_INCREMENT (mem_mode)) |
4676 | { |
4677 | rtx addr = gen_rtx_PRE_INC (addr_mode, reg); |
4678 | |
4679 | if (memory_address_addr_space_p (mem_mode, addr, as)) |
4680 | data->costs[AINC_PRE_INC] |
4681 | = address_cost (addr, mem_mode, as, speed); |
4682 | } |
4683 | if (USE_LOAD_POST_INCREMENT (mem_mode) |
4684 | || USE_STORE_POST_INCREMENT (mem_mode)) |
4685 | { |
4686 | rtx addr = gen_rtx_POST_INC (addr_mode, reg); |
4687 | |
4688 | if (memory_address_addr_space_p (mem_mode, addr, as)) |
4689 | data->costs[AINC_POST_INC] |
4690 | = address_cost (addr, mem_mode, as, speed); |
4691 | } |
4692 | ainc_cost_data_list[idx] = data; |
4693 | } |
4694 | |
4695 | poly_int64 msize = GET_MODE_SIZE (mode: mem_mode); |
4696 | if (known_eq (ainc_offset, 0) && known_eq (msize, ainc_step)) |
4697 | return comp_cost (data->costs[AINC_POST_INC], 0); |
4698 | if (known_eq (ainc_offset, 0) && known_eq (msize, -ainc_step)) |
4699 | return comp_cost (data->costs[AINC_POST_DEC], 0); |
4700 | if (known_eq (ainc_offset, msize) && known_eq (msize, ainc_step)) |
4701 | return comp_cost (data->costs[AINC_PRE_INC], 0); |
4702 | if (known_eq (ainc_offset, -msize) && known_eq (msize, -ainc_step)) |
4703 | return comp_cost (data->costs[AINC_PRE_DEC], 0); |
4704 | |
4705 | return infinite_cost; |
4706 | } |
4707 | |
4708 | /* Return cost of computing USE's address expression by using CAND. |
4709 | AFF_INV and AFF_VAR represent invariant and variant parts of the |
4710 | address expression, respectively. If AFF_INV is simple, store |
4711 | the loop invariant variables which are depended by it in INV_VARS; |
4712 | if AFF_INV is complicated, handle it as a new invariant expression |
4713 | and record it in INV_EXPR. RATIO indicates multiple times between |
4714 | steps of USE and CAND. If CAN_AUTOINC is nonNULL, store boolean |
4715 | value to it indicating if this is an auto-increment address. */ |
4716 | |
4717 | static comp_cost |
4718 | get_address_cost (struct ivopts_data *data, struct iv_use *use, |
4719 | struct iv_cand *cand, aff_tree *aff_inv, |
4720 | aff_tree *aff_var, HOST_WIDE_INT ratio, |
4721 | bitmap *inv_vars, iv_inv_expr_ent **inv_expr, |
4722 | bool *can_autoinc, bool speed) |
4723 | { |
4724 | rtx addr; |
4725 | bool simple_inv = true; |
4726 | tree comp_inv = NULL_TREE, type = aff_var->type; |
4727 | comp_cost var_cost = no_cost, cost = no_cost; |
4728 | struct mem_address parts = {NULL_TREE, integer_one_node, |
4729 | NULL_TREE, NULL_TREE, NULL_TREE}; |
4730 | machine_mode addr_mode = TYPE_MODE (type); |
4731 | machine_mode mem_mode = TYPE_MODE (use->mem_type); |
4732 | addr_space_t as = TYPE_ADDR_SPACE (TREE_TYPE (use->iv->base)); |
4733 | /* Only true if ratio != 1. */ |
4734 | bool ok_with_ratio_p = false; |
4735 | bool ok_without_ratio_p = false; |
4736 | code_helper code = ERROR_MARK; |
4737 | |
4738 | if (use->type == USE_PTR_ADDRESS) |
4739 | { |
4740 | gcall *call = as_a<gcall *> (p: use->stmt); |
4741 | gcc_assert (gimple_call_internal_p (call)); |
4742 | code = gimple_call_internal_fn (gs: call); |
4743 | } |
4744 | |
4745 | if (!aff_combination_const_p (aff: aff_inv)) |
4746 | { |
4747 | parts.index = integer_one_node; |
4748 | /* Addressing mode "base + index". */ |
4749 | ok_without_ratio_p = valid_mem_ref_p (mem_mode, as, &parts, code); |
4750 | if (ratio != 1) |
4751 | { |
4752 | parts.step = wide_int_to_tree (type, cst: ratio); |
4753 | /* Addressing mode "base + index << scale". */ |
4754 | ok_with_ratio_p = valid_mem_ref_p (mem_mode, as, &parts, code); |
4755 | if (!ok_with_ratio_p) |
4756 | parts.step = NULL_TREE; |
4757 | } |
4758 | if (ok_with_ratio_p || ok_without_ratio_p) |
4759 | { |
4760 | if (maybe_ne (a: aff_inv->offset, b: 0)) |
4761 | { |
4762 | parts.offset = wide_int_to_tree (sizetype, cst: aff_inv->offset); |
4763 | /* Addressing mode "base + index [<< scale] + offset". */ |
4764 | if (!valid_mem_ref_p (mem_mode, as, &parts, code)) |
4765 | parts.offset = NULL_TREE; |
4766 | else |
4767 | aff_inv->offset = 0; |
4768 | } |
4769 | |
4770 | move_fixed_address_to_symbol (&parts, aff_inv); |
4771 | /* Base is fixed address and is moved to symbol part. */ |
4772 | if (parts.symbol != NULL_TREE && aff_combination_zero_p (aff: aff_inv)) |
4773 | parts.base = NULL_TREE; |
4774 | |
4775 | /* Addressing mode "symbol + base + index [<< scale] [+ offset]". */ |
4776 | if (parts.symbol != NULL_TREE |
4777 | && !valid_mem_ref_p (mem_mode, as, &parts, code)) |
4778 | { |
4779 | aff_combination_add_elt (aff_inv, parts.symbol, 1); |
4780 | parts.symbol = NULL_TREE; |
4781 | /* Reset SIMPLE_INV since symbol address needs to be computed |
4782 | outside of address expression in this case. */ |
4783 | simple_inv = false; |
4784 | /* Symbol part is moved back to base part, it can't be NULL. */ |
4785 | parts.base = integer_one_node; |
4786 | } |
4787 | } |
4788 | else |
4789 | parts.index = NULL_TREE; |
4790 | } |
4791 | else |
4792 | { |
4793 | poly_int64 ainc_step; |
4794 | if (can_autoinc |
4795 | && ratio == 1 |
4796 | && ptrdiff_tree_p (cand->iv->step, &ainc_step)) |
4797 | { |
4798 | poly_int64 ainc_offset = (aff_inv->offset).force_shwi (); |
4799 | |
4800 | if (stmt_after_increment (loop: data->current_loop, cand, stmt: use->stmt)) |
4801 | ainc_offset += ainc_step; |
4802 | cost = get_address_cost_ainc (ainc_step, ainc_offset, |
4803 | addr_mode, mem_mode, as, speed); |
4804 | if (!cost.infinite_cost_p ()) |
4805 | { |
4806 | *can_autoinc = true; |
4807 | return cost; |
4808 | } |
4809 | cost = no_cost; |
4810 | } |
4811 | if (!aff_combination_zero_p (aff: aff_inv)) |
4812 | { |
4813 | parts.offset = wide_int_to_tree (sizetype, cst: aff_inv->offset); |
4814 | /* Addressing mode "base + offset". */ |
4815 | if (!valid_mem_ref_p (mem_mode, as, &parts, code)) |
4816 | parts.offset = NULL_TREE; |
4817 | else |
4818 | aff_inv->offset = 0; |
4819 | } |
4820 | } |
4821 | |
4822 | if (simple_inv) |
4823 | simple_inv = (aff_inv == NULL |
4824 | || aff_combination_const_p (aff: aff_inv) |
4825 | || aff_combination_singleton_var_p (aff: aff_inv)); |
4826 | if (!aff_combination_zero_p (aff: aff_inv)) |
4827 | comp_inv = aff_combination_to_tree (aff_inv); |
4828 | if (comp_inv != NULL_TREE) |
4829 | cost = force_var_cost (data, expr: comp_inv, inv_vars); |
4830 | if (ratio != 1 && parts.step == NULL_TREE) |
4831 | var_cost += mult_by_coeff_cost (ratio, addr_mode, speed); |
4832 | if (comp_inv != NULL_TREE && parts.index == NULL_TREE) |
4833 | var_cost += add_cost (speed, mode: addr_mode); |
4834 | |
4835 | if (comp_inv && inv_expr && !simple_inv) |
4836 | { |
4837 | *inv_expr = get_loop_invariant_expr (data, inv_expr: comp_inv); |
4838 | /* Clear depends on. */ |
4839 | if (*inv_expr != NULL && inv_vars && *inv_vars) |
4840 | bitmap_clear (*inv_vars); |
4841 | |
4842 | /* Cost of small invariant expression adjusted against loop niters |
4843 | is usually zero, which makes it difficult to be differentiated |
4844 | from candidate based on loop invariant variables. Secondly, the |
4845 | generated invariant expression may not be hoisted out of loop by |
4846 | following pass. We penalize the cost by rounding up in order to |
4847 | neutralize such effects. */ |
4848 | cost.cost = adjust_setup_cost (data, cost: cost.cost, round_up_p: true); |
4849 | cost.scratch = cost.cost; |
4850 | } |
4851 | |
4852 | cost += var_cost; |
4853 | addr = addr_for_mem_ref (&parts, as, false); |
4854 | gcc_assert (memory_address_addr_space_p (mem_mode, addr, as)); |
4855 | cost += address_cost (addr, mem_mode, as, speed); |
4856 | |
4857 | if (parts.symbol != NULL_TREE) |
4858 | cost.complexity += 1; |
4859 | /* Don't increase the complexity of adding a scaled index if it's |
4860 | the only kind of index that the target allows. */ |
4861 | if (parts.step != NULL_TREE && ok_without_ratio_p) |
4862 | cost.complexity += 1; |
4863 | if (parts.base != NULL_TREE && parts.index != NULL_TREE) |
4864 | cost.complexity += 1; |
4865 | if (parts.offset != NULL_TREE && !integer_zerop (parts.offset)) |
4866 | cost.complexity += 1; |
4867 | |
4868 | return cost; |
4869 | } |
4870 | |
4871 | /* Scale (multiply) the computed COST (except scratch part that should be |
4872 | hoisted out a loop) by header->frequency / AT->frequency, which makes |
4873 | expected cost more accurate. */ |
4874 | |
4875 | static comp_cost |
4876 | get_scaled_computation_cost_at (ivopts_data *data, gimple *at, comp_cost cost) |
4877 | { |
4878 | if (data->speed |
4879 | && data->current_loop->header->count.to_frequency (cfun) > 0) |
4880 | { |
4881 | basic_block bb = gimple_bb (g: at); |
4882 | gcc_assert (cost.scratch <= cost.cost); |
4883 | int scale_factor = (int)(intptr_t) bb->aux; |
4884 | if (scale_factor == 1) |
4885 | return cost; |
4886 | |
4887 | int64_t scaled_cost |
4888 | = cost.scratch + (cost.cost - cost.scratch) * scale_factor; |
4889 | |
4890 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4891 | fprintf (stream: dump_file, format: "Scaling cost based on bb prob by %2.2f: " |
4892 | "%" PRId64 " (scratch: %" PRId64 ") -> %" PRId64 "\n" , |
4893 | 1.0f * scale_factor, cost.cost, cost.scratch, scaled_cost); |
4894 | |
4895 | cost.cost = scaled_cost; |
4896 | } |
4897 | |
4898 | return cost; |
4899 | } |
4900 | |
4901 | /* Determines the cost of the computation by that USE is expressed |
4902 | from induction variable CAND. If ADDRESS_P is true, we just need |
4903 | to create an address from it, otherwise we want to get it into |
4904 | register. A set of invariants we depend on is stored in INV_VARS. |
4905 | If CAN_AUTOINC is nonnull, use it to record whether autoinc |
4906 | addressing is likely. If INV_EXPR is nonnull, record invariant |
4907 | expr entry in it. */ |
4908 | |
4909 | static comp_cost |
4910 | get_computation_cost (struct ivopts_data *data, struct iv_use *use, |
4911 | struct iv_cand *cand, bool address_p, bitmap *inv_vars, |
4912 | bool *can_autoinc, iv_inv_expr_ent **inv_expr) |
4913 | { |
4914 | gimple *at = use->stmt; |
4915 | tree ubase = use->iv->base, cbase = cand->iv->base; |
4916 | tree utype = TREE_TYPE (ubase), ctype = TREE_TYPE (cbase); |
4917 | tree comp_inv = NULL_TREE; |
4918 | HOST_WIDE_INT ratio, aratio; |
4919 | comp_cost cost; |
4920 | widest_int rat; |
4921 | aff_tree aff_inv, aff_var; |
4922 | bool speed = optimize_bb_for_speed_p (gimple_bb (g: at)); |
4923 | |
4924 | if (inv_vars) |
4925 | *inv_vars = NULL; |
4926 | if (can_autoinc) |
4927 | *can_autoinc = false; |
4928 | if (inv_expr) |
4929 | *inv_expr = NULL; |
4930 | |
4931 | /* Check if we have enough precision to express the values of use. */ |
4932 | if (TYPE_PRECISION (utype) > TYPE_PRECISION (ctype)) |
4933 | return infinite_cost; |
4934 | |
4935 | if (address_p |
4936 | || (use->iv->base_object |
4937 | && cand->iv->base_object |
4938 | && POINTER_TYPE_P (TREE_TYPE (use->iv->base_object)) |
4939 | && POINTER_TYPE_P (TREE_TYPE (cand->iv->base_object)))) |
4940 | { |
4941 | /* Do not try to express address of an object with computation based |
4942 | on address of a different object. This may cause problems in rtl |
4943 | level alias analysis (that does not expect this to be happening, |
4944 | as this is illegal in C), and would be unlikely to be useful |
4945 | anyway. */ |
4946 | if (use->iv->base_object |
4947 | && cand->iv->base_object |
4948 | && !operand_equal_p (use->iv->base_object, cand->iv->base_object, flags: 0)) |
4949 | return infinite_cost; |
4950 | } |
4951 | |
4952 | if (!get_computation_aff_1 (loop: data->current_loop, at, use, |
4953 | cand, aff_inv: &aff_inv, aff_var: &aff_var, prat: &rat) |
4954 | || !wi::fits_shwi_p (x: rat)) |
4955 | return infinite_cost; |
4956 | |
4957 | ratio = rat.to_shwi (); |
4958 | if (address_p) |
4959 | { |
4960 | cost = get_address_cost (data, use, cand, aff_inv: &aff_inv, aff_var: &aff_var, ratio, |
4961 | inv_vars, inv_expr, can_autoinc, speed); |
4962 | cost = get_scaled_computation_cost_at (data, at, cost); |
4963 | /* For doloop IV cand, add on the extra cost. */ |
4964 | cost += cand->doloop_p ? targetm.doloop_cost_for_address : 0; |
4965 | return cost; |
4966 | } |
4967 | |
4968 | bool simple_inv = (aff_combination_const_p (aff: &aff_inv) |
4969 | || aff_combination_singleton_var_p (aff: &aff_inv)); |
4970 | tree signed_type = signed_type_for (aff_combination_type (aff: &aff_inv)); |
4971 | aff_combination_convert (&aff_inv, signed_type); |
4972 | if (!aff_combination_zero_p (aff: &aff_inv)) |
4973 | comp_inv = aff_combination_to_tree (&aff_inv); |
4974 | |
4975 | cost = force_var_cost (data, expr: comp_inv, inv_vars); |
4976 | if (comp_inv && inv_expr && !simple_inv) |
4977 | { |
4978 | *inv_expr = get_loop_invariant_expr (data, inv_expr: comp_inv); |
4979 | /* Clear depends on. */ |
4980 | if (*inv_expr != NULL && inv_vars && *inv_vars) |
4981 | bitmap_clear (*inv_vars); |
4982 | |
4983 | cost.cost = adjust_setup_cost (data, cost: cost.cost); |
4984 | /* Record setup cost in scratch field. */ |
4985 | cost.scratch = cost.cost; |
4986 | } |
4987 | /* Cost of constant integer can be covered when adding invariant part to |
4988 | variant part. */ |
4989 | else if (comp_inv && CONSTANT_CLASS_P (comp_inv)) |
4990 | cost = no_cost; |
4991 | |
4992 | /* Need type narrowing to represent use with cand. */ |
4993 | if (TYPE_PRECISION (utype) < TYPE_PRECISION (ctype)) |
4994 | { |
4995 | machine_mode outer_mode = TYPE_MODE (utype); |
4996 | machine_mode inner_mode = TYPE_MODE (ctype); |
4997 | cost += comp_cost (convert_cost (to_mode: outer_mode, from_mode: inner_mode, speed), 0); |
4998 | } |
4999 | |
5000 | /* Turn a + i * (-c) into a - i * c. */ |
5001 | if (ratio < 0 && comp_inv && !integer_zerop (comp_inv)) |
5002 | aratio = -ratio; |
5003 | else |
5004 | aratio = ratio; |
5005 | |
5006 | if (ratio != 1) |
5007 | cost += mult_by_coeff_cost (aratio, TYPE_MODE (utype), speed); |
5008 | |
5009 | /* TODO: We may also need to check if we can compute a + i * 4 in one |
5010 | instruction. */ |
5011 | /* Need to add up the invariant and variant parts. */ |
5012 | if (comp_inv && !integer_zerop (comp_inv)) |
5013 | cost += add_cost (speed, TYPE_MODE (utype)); |
5014 | |
5015 | cost = get_scaled_computation_cost_at (data, at, cost); |
5016 | |
5017 | /* For doloop IV cand, add on the extra cost. */ |
5018 | if (cand->doloop_p && use->type == USE_NONLINEAR_EXPR) |
5019 | cost += targetm.doloop_cost_for_generic; |
5020 | |
5021 | return cost; |
5022 | } |
5023 | |
5024 | /* Determines cost of computing the use in GROUP with CAND in a generic |
5025 | expression. */ |
5026 | |
5027 | static bool |
5028 | determine_group_iv_cost_generic (struct ivopts_data *data, |
5029 | struct iv_group *group, struct iv_cand *cand) |
5030 | { |
5031 | comp_cost cost; |
5032 | iv_inv_expr_ent *inv_expr = NULL; |
5033 | bitmap inv_vars = NULL, inv_exprs = NULL; |
5034 | struct iv_use *use = group->vuses[0]; |
5035 | |
5036 | /* The simple case first -- if we need to express value of the preserved |
5037 | original biv, the cost is 0. This also prevents us from counting the |
5038 | cost of increment twice -- once at this use and once in the cost of |
5039 | the candidate. */ |
5040 | if (cand->pos == IP_ORIGINAL && cand->incremented_at == use->stmt) |
5041 | cost = no_cost; |
5042 | /* If the IV candidate involves undefined SSA values and is not the |
5043 | same IV as on the USE avoid using that candidate here. */ |
5044 | else if (cand->involves_undefs |
5045 | && (!use->iv || !operand_equal_p (cand->iv->base, use->iv->base, flags: 0))) |
5046 | return false; |
5047 | else |
5048 | cost = get_computation_cost (data, use, cand, address_p: false, |
5049 | inv_vars: &inv_vars, NULL, inv_expr: &inv_expr); |
5050 | |
5051 | if (inv_expr) |
5052 | { |
5053 | inv_exprs = BITMAP_ALLOC (NULL); |
5054 | bitmap_set_bit (inv_exprs, inv_expr->id); |
5055 | } |
5056 | set_group_iv_cost (data, group, cand, cost, inv_vars, |
5057 | NULL_TREE, comp: ERROR_MARK, inv_exprs); |
5058 | return !cost.infinite_cost_p (); |
5059 | } |
5060 | |
5061 | /* Determines cost of computing uses in GROUP with CAND in addresses. */ |
5062 | |
5063 | static bool |
5064 | determine_group_iv_cost_address (struct ivopts_data *data, |
5065 | struct iv_group *group, struct iv_cand *cand) |
5066 | { |
5067 | unsigned i; |
5068 | bitmap inv_vars = NULL, inv_exprs = NULL; |
5069 | bool can_autoinc; |
5070 | iv_inv_expr_ent *inv_expr = NULL; |
5071 | struct iv_use *use = group->vuses[0]; |
5072 | comp_cost sum_cost = no_cost, cost; |
5073 | |
5074 | cost = get_computation_cost (data, use, cand, address_p: true, |
5075 | inv_vars: &inv_vars, can_autoinc: &can_autoinc, inv_expr: &inv_expr); |
5076 | |
5077 | if (inv_expr) |
5078 | { |
5079 | inv_exprs = BITMAP_ALLOC (NULL); |
5080 | bitmap_set_bit (inv_exprs, inv_expr->id); |
5081 | } |
5082 | sum_cost = cost; |
5083 | if (!sum_cost.infinite_cost_p () && cand->ainc_use == use) |
5084 | { |
5085 | if (can_autoinc) |
5086 | sum_cost -= cand->cost_step; |
5087 | /* If we generated the candidate solely for exploiting autoincrement |
5088 | opportunities, and it turns out it can't be used, set the cost to |
5089 | infinity to make sure we ignore it. */ |
5090 | else if (cand->pos == IP_AFTER_USE || cand->pos == IP_BEFORE_USE) |
5091 | sum_cost = infinite_cost; |
5092 | } |
5093 | |
5094 | /* Uses in a group can share setup code, so only add setup cost once. */ |
5095 | cost -= cost.scratch; |
5096 | /* Compute and add costs for rest uses of this group. */ |
5097 | for (i = 1; i < group->vuses.length () && !sum_cost.infinite_cost_p (); i++) |
5098 | { |
5099 | struct iv_use *next = group->vuses[i]; |
5100 | |
5101 | /* TODO: We could skip computing cost for sub iv_use when it has the |
5102 | same cost as the first iv_use, but the cost really depends on the |
5103 | offset and where the iv_use is. */ |
5104 | cost = get_computation_cost (data, use: next, cand, address_p: true, |
5105 | NULL, can_autoinc: &can_autoinc, inv_expr: &inv_expr); |
5106 | if (inv_expr) |
5107 | { |
5108 | if (!inv_exprs) |
5109 | inv_exprs = BITMAP_ALLOC (NULL); |
5110 | |
5111 | bitmap_set_bit (inv_exprs, inv_expr->id); |
5112 | } |
5113 | sum_cost += cost; |
5114 | } |
5115 | set_group_iv_cost (data, group, cand, cost: sum_cost, inv_vars, |
5116 | NULL_TREE, comp: ERROR_MARK, inv_exprs); |
5117 | |
5118 | return !sum_cost.infinite_cost_p (); |
5119 | } |
5120 | |
5121 | /* Computes value of candidate CAND at position AT in iteration DESC->NITER, |
5122 | and stores it to VAL. */ |
5123 | |
5124 | static void |
5125 | cand_value_at (class loop *loop, struct iv_cand *cand, gimple *at, |
5126 | class tree_niter_desc *desc, aff_tree *val) |
5127 | { |
5128 | aff_tree step, delta, nit; |
5129 | struct iv *iv = cand->iv; |
5130 | tree type = TREE_TYPE (iv->base); |
5131 | tree niter = desc->niter; |
5132 | bool after_adjust = stmt_after_increment (loop, cand, stmt: at); |
5133 | tree steptype; |
5134 | |
5135 | if (POINTER_TYPE_P (type)) |
5136 | steptype = sizetype; |
5137 | else |
5138 | steptype = unsigned_type_for (type); |
5139 | |
5140 | /* If AFTER_ADJUST is required, the code below generates the equivalent |
5141 | of BASE + NITER * STEP + STEP, when ideally we'd prefer the expression |
5142 | BASE + (NITER + 1) * STEP, especially when NITER is often of the form |
5143 | SSA_NAME - 1. Unfortunately, guaranteeing that adding 1 to NITER |
5144 | doesn't overflow is tricky, so we peek inside the TREE_NITER_DESC |
5145 | class for common idioms that we know are safe. */ |
5146 | if (after_adjust |
5147 | && desc->control.no_overflow |
5148 | && integer_onep (desc->control.step) |
5149 | && (desc->cmp == LT_EXPR |
5150 | || desc->cmp == NE_EXPR) |
5151 | && TREE_CODE (desc->bound) == SSA_NAME) |
5152 | { |
5153 | if (integer_onep (desc->control.base)) |
5154 | { |
5155 | niter = desc->bound; |
5156 | after_adjust = false; |
5157 | } |
5158 | else if (TREE_CODE (niter) == MINUS_EXPR |
5159 | && integer_onep (TREE_OPERAND (niter, 1))) |
5160 | { |
5161 | niter = TREE_OPERAND (niter, 0); |
5162 | after_adjust = false; |
5163 | } |
5164 | } |
5165 | |
5166 | tree_to_aff_combination (iv->step, TREE_TYPE (iv->step), &step); |
5167 | aff_combination_convert (&step, steptype); |
5168 | tree_to_aff_combination (niter, TREE_TYPE (niter), &nit); |
5169 | aff_combination_convert (&nit, steptype); |
5170 | aff_combination_mult (&nit, &step, &delta); |
5171 | if (after_adjust) |
5172 | aff_combination_add (&delta, &step); |
5173 | |
5174 | tree_to_aff_combination (iv->base, type, val); |
5175 | if (!POINTER_TYPE_P (type)) |
5176 | aff_combination_convert (val, steptype); |
5177 | aff_combination_add (val, &delta); |
5178 | } |
5179 | |
5180 | /* Returns period of induction variable iv. */ |
5181 | |
5182 | static tree |
5183 | iv_period (struct iv *iv) |
5184 | { |
5185 | tree step = iv->step, period, type; |
5186 | tree pow2div; |
5187 | |
5188 | gcc_assert (step && TREE_CODE (step) == INTEGER_CST); |
5189 | |
5190 | type = unsigned_type_for (TREE_TYPE (step)); |
5191 | /* Period of the iv is lcm (step, type_range)/step -1, |
5192 | i.e., N*type_range/step - 1. Since type range is power |
5193 | of two, N == (step >> num_of_ending_zeros_binary (step), |
5194 | so the final result is |
5195 | |
5196 | (type_range >> num_of_ending_zeros_binary (step)) - 1 |
5197 | |
5198 | */ |
5199 | pow2div = num_ending_zeros (step); |
5200 | |
5201 | period = build_low_bits_mask (type, |
5202 | (TYPE_PRECISION (type) |
5203 | - tree_to_uhwi (pow2div))); |
5204 | |
5205 | return period; |
5206 | } |
5207 | |
5208 | /* Returns the comparison operator used when eliminating the iv USE. */ |
5209 | |
5210 | static enum tree_code |
5211 | iv_elimination_compare (struct ivopts_data *data, struct iv_use *use) |
5212 | { |
5213 | class loop *loop = data->current_loop; |
5214 | basic_block ex_bb; |
5215 | edge exit; |
5216 | |
5217 | ex_bb = gimple_bb (g: use->stmt); |
5218 | exit = EDGE_SUCC (ex_bb, 0); |
5219 | if (flow_bb_inside_loop_p (loop, exit->dest)) |
5220 | exit = EDGE_SUCC (ex_bb, 1); |
5221 | |
5222 | return (exit->flags & EDGE_TRUE_VALUE ? EQ_EXPR : NE_EXPR); |
5223 | } |
5224 | |
5225 | /* Returns true if we can prove that BASE - OFFSET does not overflow. For now, |
5226 | we only detect the situation that BASE = SOMETHING + OFFSET, where the |
5227 | calculation is performed in non-wrapping type. |
5228 | |
5229 | TODO: More generally, we could test for the situation that |
5230 | BASE = SOMETHING + OFFSET' and OFFSET is between OFFSET' and zero. |
5231 | This would require knowing the sign of OFFSET. */ |
5232 | |
5233 | static bool |
5234 | difference_cannot_overflow_p (struct ivopts_data *data, tree base, tree offset) |
5235 | { |
5236 | enum tree_code code; |
5237 | tree e1, e2; |
5238 | aff_tree aff_e1, aff_e2, aff_offset; |
5239 | |
5240 | if (!nowrap_type_p (TREE_TYPE (base))) |
5241 | return false; |
5242 | |
5243 | base = expand_simple_operations (base); |
5244 | |
5245 | if (TREE_CODE (base) == SSA_NAME) |
5246 | { |
5247 | gimple *stmt = SSA_NAME_DEF_STMT (base); |
5248 | |
5249 | if (gimple_code (g: stmt) != GIMPLE_ASSIGN) |
5250 | return false; |
5251 | |
5252 | code = gimple_assign_rhs_code (gs: stmt); |
5253 | if (get_gimple_rhs_class (code) != GIMPLE_BINARY_RHS) |
5254 | return false; |
5255 | |
5256 | e1 = gimple_assign_rhs1 (gs: stmt); |
5257 | e2 = gimple_assign_rhs2 (gs: stmt); |
5258 | } |
5259 | else |
5260 | { |
5261 | code = TREE_CODE (base); |
5262 | if (get_gimple_rhs_class (code) != GIMPLE_BINARY_RHS) |
5263 | return false; |
5264 | e1 = TREE_OPERAND (base, 0); |
5265 | e2 = TREE_OPERAND (base, 1); |
5266 | } |
5267 | |
5268 | /* Use affine expansion as deeper inspection to prove the equality. */ |
5269 | tree_to_aff_combination_expand (e2, TREE_TYPE (e2), |
5270 | &aff_e2, &data->name_expansion_cache); |
5271 | tree_to_aff_combination_expand (offset, TREE_TYPE (offset), |
5272 | &aff_offset, &data->name_expansion_cache); |
5273 | aff_combination_scale (&aff_offset, -1); |
5274 | switch (code) |
5275 | { |
5276 | case PLUS_EXPR: |
5277 | aff_combination_add (&aff_e2, &aff_offset); |
5278 | if (aff_combination_zero_p (aff: &aff_e2)) |
5279 | return true; |
5280 | |
5281 | tree_to_aff_combination_expand (e1, TREE_TYPE (e1), |
5282 | &aff_e1, &data->name_expansion_cache); |
5283 | aff_combination_add (&aff_e1, &aff_offset); |
5284 | return aff_combination_zero_p (aff: &aff_e1); |
5285 | |
5286 | case POINTER_PLUS_EXPR: |
5287 | aff_combination_add (&aff_e2, &aff_offset); |
5288 | return aff_combination_zero_p (aff: &aff_e2); |
5289 | |
5290 | default: |
5291 | return false; |
5292 | } |
5293 | } |
5294 | |
5295 | /* Tries to replace loop exit by one formulated in terms of a LT_EXPR |
5296 | comparison with CAND. NITER describes the number of iterations of |
5297 | the loops. If successful, the comparison in COMP_P is altered accordingly. |
5298 | |
5299 | We aim to handle the following situation: |
5300 | |
5301 | sometype *base, *p; |
5302 | int a, b, i; |
5303 | |
5304 | i = a; |
5305 | p = p_0 = base + a; |
5306 | |
5307 | do |
5308 | { |
5309 | bla (*p); |
5310 | p++; |
5311 | i++; |
5312 | } |
5313 | while (i < b); |
5314 | |
5315 | Here, the number of iterations of the loop is (a + 1 > b) ? 0 : b - a - 1. |
5316 | We aim to optimize this to |
5317 | |
5318 | p = p_0 = base + a; |
5319 | do |
5320 | { |
5321 | bla (*p); |
5322 | p++; |
5323 | } |
5324 | while (p < p_0 - a + b); |
5325 | |
5326 | This preserves the correctness, since the pointer arithmetics does not |
5327 | overflow. More precisely: |
5328 | |
5329 | 1) if a + 1 <= b, then p_0 - a + b is the final value of p, hence there is no |
5330 | overflow in computing it or the values of p. |
5331 | 2) if a + 1 > b, then we need to verify that the expression p_0 - a does not |
5332 | overflow. To prove this, we use the fact that p_0 = base + a. */ |
5333 | |
5334 | static bool |
5335 | iv_elimination_compare_lt (struct ivopts_data *data, |
5336 | struct iv_cand *cand, enum tree_code *comp_p, |
5337 | class tree_niter_desc *niter) |
5338 | { |
5339 | tree cand_type, a, b, mbz, nit_type = TREE_TYPE (niter->niter), offset; |
5340 | class aff_tree nit, tmpa, tmpb; |
5341 | enum tree_code comp; |
5342 | HOST_WIDE_INT step; |
5343 | |
5344 | /* We need to know that the candidate induction variable does not overflow. |
5345 | While more complex analysis may be used to prove this, for now just |
5346 | check that the variable appears in the original program and that it |
5347 | is computed in a type that guarantees no overflows. */ |
5348 | cand_type = TREE_TYPE (cand->iv->base); |
5349 | if (cand->pos != IP_ORIGINAL || !nowrap_type_p (cand_type)) |
5350 | return false; |
5351 | |
5352 | /* Make sure that the loop iterates till the loop bound is hit, as otherwise |
5353 | the calculation of the BOUND could overflow, making the comparison |
5354 | invalid. */ |
5355 | if (!data->loop_single_exit_p) |
5356 | return false; |
5357 | |
5358 | /* We need to be able to decide whether candidate is increasing or decreasing |
5359 | in order to choose the right comparison operator. */ |
5360 | if (!cst_and_fits_in_hwi (cand->iv->step)) |
5361 | return false; |
5362 | step = int_cst_value (cand->iv->step); |
5363 | |
5364 | /* Check that the number of iterations matches the expected pattern: |
5365 | a + 1 > b ? 0 : b - a - 1. */ |
5366 | mbz = niter->may_be_zero; |
5367 | if (TREE_CODE (mbz) == GT_EXPR) |
5368 | { |
5369 | /* Handle a + 1 > b. */ |
5370 | tree op0 = TREE_OPERAND (mbz, 0); |
5371 | if (TREE_CODE (op0) == PLUS_EXPR && integer_onep (TREE_OPERAND (op0, 1))) |
5372 | { |
5373 | a = TREE_OPERAND (op0, 0); |
5374 | b = TREE_OPERAND (mbz, 1); |
5375 | } |
5376 | else |
5377 | return false; |
5378 | } |
5379 | else if (TREE_CODE (mbz) == LT_EXPR) |
5380 | { |
5381 | tree op1 = TREE_OPERAND (mbz, 1); |
5382 | |
5383 | /* Handle b < a + 1. */ |
5384 | if (TREE_CODE (op1) == PLUS_EXPR && integer_onep (TREE_OPERAND (op1, 1))) |
5385 | { |
5386 | a = TREE_OPERAND (op1, 0); |
5387 | b = TREE_OPERAND (mbz, 0); |
5388 | } |
5389 | else |
5390 | return false; |
5391 | } |
5392 | else |
5393 | return false; |
5394 | |
5395 | /* Expected number of iterations is B - A - 1. Check that it matches |
5396 | the actual number, i.e., that B - A - NITER = 1. */ |
5397 | tree_to_aff_combination (niter->niter, nit_type, &nit); |
5398 | tree_to_aff_combination (fold_convert (nit_type, a), nit_type, &tmpa); |
5399 | tree_to_aff_combination (fold_convert (nit_type, b), nit_type, &tmpb); |
5400 | aff_combination_scale (&nit, -1); |
5401 | aff_combination_scale (&tmpa, -1); |
5402 | aff_combination_add (&tmpb, &tmpa); |
5403 | aff_combination_add (&tmpb, &nit); |
5404 | if (tmpb.n != 0 || maybe_ne (a: tmpb.offset, b: 1)) |
5405 | return false; |
5406 | |
5407 | /* Finally, check that CAND->IV->BASE - CAND->IV->STEP * A does not |
5408 | overflow. */ |
5409 | offset = fold_build2 (MULT_EXPR, TREE_TYPE (cand->iv->step), |
5410 | cand->iv->step, |
5411 | fold_convert (TREE_TYPE (cand->iv->step), a)); |
5412 | if (!difference_cannot_overflow_p (data, base: cand->iv->base, offset)) |
5413 | return false; |
5414 | |
5415 | /* Determine the new comparison operator. */ |
5416 | comp = step < 0 ? GT_EXPR : LT_EXPR; |
5417 | if (*comp_p == NE_EXPR) |
5418 | *comp_p = comp; |
5419 | else if (*comp_p == EQ_EXPR) |
5420 | *comp_p = invert_tree_comparison (comp, false); |
5421 | else |
5422 | gcc_unreachable (); |
5423 | |
5424 | return true; |
5425 | } |
5426 | |
5427 | /* Check whether it is possible to express the condition in USE by comparison |
5428 | of candidate CAND. If so, store the value compared with to BOUND, and the |
5429 | comparison operator to COMP. */ |
5430 | |
5431 | static bool |
5432 | may_eliminate_iv (struct ivopts_data *data, |
5433 | struct iv_use *use, struct iv_cand *cand, tree *bound, |
5434 | enum tree_code *comp) |
5435 | { |
5436 | basic_block ex_bb; |
5437 | edge exit; |
5438 | tree period; |
5439 | class loop *loop = data->current_loop; |
5440 | aff_tree bnd; |
5441 | class tree_niter_desc *desc = NULL; |
5442 | |
5443 | if (TREE_CODE (cand->iv->step) != INTEGER_CST) |
5444 | return false; |
5445 | |
5446 | /* For now works only for exits that dominate the loop latch. |
5447 | TODO: extend to other conditions inside loop body. */ |
5448 | ex_bb = gimple_bb (g: use->stmt); |
5449 | if (use->stmt != last_nondebug_stmt (ex_bb) |
5450 | || gimple_code (g: use->stmt) != GIMPLE_COND |
5451 | || !dominated_by_p (CDI_DOMINATORS, loop->latch, ex_bb)) |
5452 | return false; |
5453 | |
5454 | exit = EDGE_SUCC (ex_bb, 0); |
5455 | if (flow_bb_inside_loop_p (loop, exit->dest)) |
5456 | exit = EDGE_SUCC (ex_bb, 1); |
5457 | if (flow_bb_inside_loop_p (loop, exit->dest)) |
5458 | return false; |
5459 | |
5460 | desc = niter_for_exit (data, exit); |
5461 | if (!desc) |
5462 | return false; |
5463 | |
5464 | /* Determine whether we can use the variable to test the exit condition. |
5465 | This is the case iff the period of the induction variable is greater |
5466 | than the number of iterations for which the exit condition is true. */ |
5467 | period = iv_period (iv: cand->iv); |
5468 | |
5469 | /* If the number of iterations is constant, compare against it directly. */ |
5470 | if (TREE_CODE (desc->niter) == INTEGER_CST) |
5471 | { |
5472 | /* See cand_value_at. */ |
5473 | if (stmt_after_increment (loop, cand, stmt: use->stmt)) |
5474 | { |
5475 | if (!tree_int_cst_lt (t1: desc->niter, t2: period)) |
5476 | return false; |
5477 | } |
5478 | else |
5479 | { |
5480 | if (tree_int_cst_lt (t1: period, t2: desc->niter)) |
5481 | return false; |
5482 | } |
5483 | } |
5484 | |
5485 | /* If not, and if this is the only possible exit of the loop, see whether |
5486 | we can get a conservative estimate on the number of iterations of the |
5487 | entire loop and compare against that instead. */ |
5488 | else |
5489 | { |
5490 | widest_int period_value, max_niter; |
5491 | |
5492 | max_niter = desc->max; |
5493 | if (stmt_after_increment (loop, cand, stmt: use->stmt)) |
5494 | max_niter += 1; |
5495 | period_value = wi::to_widest (t: period); |
5496 | if (wi::gtu_p (x: max_niter, y: period_value)) |
5497 | { |
5498 | /* See if we can take advantage of inferred loop bound |
5499 | information. */ |
5500 | if (data->loop_single_exit_p) |
5501 | { |
5502 | if (!max_loop_iterations (loop, &max_niter)) |
5503 | return false; |
5504 | /* The loop bound is already adjusted by adding 1. */ |
5505 | if (wi::gtu_p (x: max_niter, y: period_value)) |
5506 | return false; |
5507 | } |
5508 | else |
5509 | return false; |
5510 | } |
5511 | } |
5512 | |
5513 | /* For doloop IV cand, the bound would be zero. It's safe whether |
5514 | may_be_zero set or not. */ |
5515 | if (cand->doloop_p) |
5516 | { |
5517 | *bound = build_int_cst (TREE_TYPE (cand->iv->base), 0); |
5518 | *comp = iv_elimination_compare (data, use); |
5519 | return true; |
5520 | } |
5521 | |
5522 | cand_value_at (loop, cand, at: use->stmt, desc, val: &bnd); |
5523 | |
5524 | *bound = fold_convert (TREE_TYPE (cand->iv->base), |
5525 | aff_combination_to_tree (&bnd)); |
5526 | *comp = iv_elimination_compare (data, use); |
5527 | |
5528 | /* It is unlikely that computing the number of iterations using division |
5529 | would be more profitable than keeping the original induction variable. */ |
5530 | bool cond_overflow_p; |
5531 | if (expression_expensive_p (*bound, &cond_overflow_p)) |
5532 | return false; |
5533 | |
5534 | /* Sometimes, it is possible to handle the situation that the number of |
5535 | iterations may be zero unless additional assumptions by using < |
5536 | instead of != in the exit condition. |
5537 | |
5538 | TODO: we could also calculate the value MAY_BE_ZERO ? 0 : NITER and |
5539 | base the exit condition on it. However, that is often too |
5540 | expensive. */ |
5541 | if (!integer_zerop (desc->may_be_zero)) |
5542 | return iv_elimination_compare_lt (data, cand, comp_p: comp, niter: desc); |
5543 | |
5544 | return true; |
5545 | } |
5546 | |
5547 | /* Calculates the cost of BOUND, if it is a PARM_DECL. A PARM_DECL must |
5548 | be copied, if it is used in the loop body and DATA->body_includes_call. */ |
5549 | |
5550 | static int |
5551 | parm_decl_cost (struct ivopts_data *data, tree bound) |
5552 | { |
5553 | tree sbound = bound; |
5554 | STRIP_NOPS (sbound); |
5555 | |
5556 | if (TREE_CODE (sbound) == SSA_NAME |
5557 | && SSA_NAME_IS_DEFAULT_DEF (sbound) |
5558 | && TREE_CODE (SSA_NAME_VAR (sbound)) == PARM_DECL |
5559 | && data->body_includes_call) |
5560 | return COSTS_N_INSNS (1); |
5561 | |
5562 | return 0; |
5563 | } |
5564 | |
5565 | /* Determines cost of computing the use in GROUP with CAND in a condition. */ |
5566 | |
5567 | static bool |
5568 | determine_group_iv_cost_cond (struct ivopts_data *data, |
5569 | struct iv_group *group, struct iv_cand *cand) |
5570 | { |
5571 | tree bound = NULL_TREE; |
5572 | struct iv *cmp_iv; |
5573 | bitmap inv_exprs = NULL; |
5574 | bitmap inv_vars_elim = NULL, inv_vars_express = NULL, inv_vars; |
5575 | comp_cost elim_cost = infinite_cost, express_cost, cost, bound_cost; |
5576 | enum comp_iv_rewrite rewrite_type; |
5577 | iv_inv_expr_ent *inv_expr_elim = NULL, *inv_expr_express = NULL, *inv_expr; |
5578 | tree *control_var, *bound_cst; |
5579 | enum tree_code comp = ERROR_MARK; |
5580 | struct iv_use *use = group->vuses[0]; |
5581 | |
5582 | /* Extract condition operands. */ |
5583 | rewrite_type = extract_cond_operands (data, stmt: use->stmt, control_var: &control_var, |
5584 | bound: &bound_cst, NULL, iv_bound: &cmp_iv); |
5585 | gcc_assert (rewrite_type != COMP_IV_NA); |
5586 | |
5587 | /* Try iv elimination. */ |
5588 | if (rewrite_type == COMP_IV_ELIM |
5589 | && may_eliminate_iv (data, use, cand, bound: &bound, comp: &comp)) |
5590 | { |
5591 | elim_cost = force_var_cost (data, expr: bound, inv_vars: &inv_vars_elim); |
5592 | if (elim_cost.cost == 0) |
5593 | elim_cost.cost = parm_decl_cost (data, bound); |
5594 | else if (TREE_CODE (bound) == INTEGER_CST) |
5595 | elim_cost.cost = 0; |
5596 | /* If we replace a loop condition 'i < n' with 'p < base + n', |
5597 | inv_vars_elim will have 'base' and 'n' set, which implies that both |
5598 | 'base' and 'n' will be live during the loop. More likely, |
5599 | 'base + n' will be loop invariant, resulting in only one live value |
5600 | during the loop. So in that case we clear inv_vars_elim and set |
5601 | inv_expr_elim instead. */ |
5602 | if (inv_vars_elim && bitmap_count_bits (inv_vars_elim) > 1) |
5603 | { |
5604 | inv_expr_elim = get_loop_invariant_expr (data, inv_expr: bound); |
5605 | bitmap_clear (inv_vars_elim); |
5606 | } |
5607 | /* The bound is a loop invariant, so it will be only computed |
5608 | once. */ |
5609 | elim_cost.cost = adjust_setup_cost (data, cost: elim_cost.cost); |
5610 | } |
5611 | |
5612 | /* When the condition is a comparison of the candidate IV against |
5613 | zero, prefer this IV. |
5614 | |
5615 | TODO: The constant that we're subtracting from the cost should |
5616 | be target-dependent. This information should be added to the |
5617 | target costs for each backend. */ |
5618 | if (!elim_cost.infinite_cost_p () /* Do not try to decrease infinite! */ |
5619 | && integer_zerop (*bound_cst) |
5620 | && (operand_equal_p (*control_var, cand->var_after, flags: 0) |
5621 | || operand_equal_p (*control_var, cand->var_before, flags: 0))) |
5622 | elim_cost -= 1; |
5623 | |
5624 | express_cost = get_computation_cost (data, use, cand, address_p: false, |
5625 | inv_vars: &inv_vars_express, NULL, |
5626 | inv_expr: &inv_expr_express); |
5627 | if (cmp_iv != NULL) |
5628 | find_inv_vars (data, expr_p: &cmp_iv->base, inv_vars: &inv_vars_express); |
5629 | |
5630 | /* Count the cost of the original bound as well. */ |
5631 | bound_cost = force_var_cost (data, expr: *bound_cst, NULL); |
5632 | if (bound_cost.cost == 0) |
5633 | bound_cost.cost = parm_decl_cost (data, bound: *bound_cst); |
5634 | else if (TREE_CODE (*bound_cst) == INTEGER_CST) |
5635 | bound_cost.cost = 0; |
5636 | express_cost += bound_cost; |
5637 | |
5638 | /* Choose the better approach, preferring the eliminated IV. */ |
5639 | if (elim_cost <= express_cost) |
5640 | { |
5641 | cost = elim_cost; |
5642 | inv_vars = inv_vars_elim; |
5643 | inv_vars_elim = NULL; |
5644 | inv_expr = inv_expr_elim; |
5645 | /* For doloop candidate/use pair, adjust to zero cost. */ |
5646 | if (group->doloop_p && cand->doloop_p && elim_cost.cost > no_cost.cost) |
5647 | cost = no_cost; |
5648 | } |
5649 | else |
5650 | { |
5651 | cost = express_cost; |
5652 | inv_vars = inv_vars_express; |
5653 | inv_vars_express = NULL; |
5654 | bound = NULL_TREE; |
5655 | comp = ERROR_MARK; |
5656 | inv_expr = inv_expr_express; |
5657 | } |
5658 | |
5659 | if (inv_expr) |
5660 | { |
5661 | inv_exprs = BITMAP_ALLOC (NULL); |
5662 | bitmap_set_bit (inv_exprs, inv_expr->id); |
5663 | } |
5664 | set_group_iv_cost (data, group, cand, cost, |
5665 | inv_vars, value: bound, comp, inv_exprs); |
5666 | |
5667 | if (inv_vars_elim) |
5668 | BITMAP_FREE (inv_vars_elim); |
5669 | if (inv_vars_express) |
5670 | BITMAP_FREE (inv_vars_express); |
5671 | |
5672 | return !cost.infinite_cost_p (); |
5673 | } |
5674 | |
5675 | /* Determines cost of computing uses in GROUP with CAND. Returns false |
5676 | if USE cannot be represented with CAND. */ |
5677 | |
5678 | static bool |
5679 | determine_group_iv_cost (struct ivopts_data *data, |
5680 | struct iv_group *group, struct iv_cand *cand) |
5681 | { |
5682 | switch (group->type) |
5683 | { |
5684 | case USE_NONLINEAR_EXPR: |
5685 | return determine_group_iv_cost_generic (data, group, cand); |
5686 | |
5687 | case USE_REF_ADDRESS: |
5688 | case USE_PTR_ADDRESS: |
5689 | return determine_group_iv_cost_address (data, group, cand); |
5690 | |
5691 | case USE_COMPARE: |
5692 | return determine_group_iv_cost_cond (data, group, cand); |
5693 | |
5694 | default: |
5695 | gcc_unreachable (); |
5696 | } |
5697 | } |
5698 | |
5699 | /* Return true if get_computation_cost indicates that autoincrement is |
5700 | a possibility for the pair of USE and CAND, false otherwise. */ |
5701 | |
5702 | static bool |
5703 | autoinc_possible_for_pair (struct ivopts_data *data, struct iv_use *use, |
5704 | struct iv_cand *cand) |
5705 | { |
5706 | if (!address_p (type: use->type)) |
5707 | return false; |
5708 | |
5709 | bool can_autoinc = false; |
5710 | get_computation_cost (data, use, cand, address_p: true, NULL, can_autoinc: &can_autoinc, NULL); |
5711 | return can_autoinc; |
5712 | } |
5713 | |
5714 | /* Examine IP_ORIGINAL candidates to see if they are incremented next to a |
5715 | use that allows autoincrement, and set their AINC_USE if possible. */ |
5716 | |
5717 | static void |
5718 | set_autoinc_for_original_candidates (struct ivopts_data *data) |
5719 | { |
5720 | unsigned i, j; |
5721 | |
5722 | for (i = 0; i < data->vcands.length (); i++) |
5723 | { |
5724 | struct iv_cand *cand = data->vcands[i]; |
5725 | struct iv_use *closest_before = NULL; |
5726 | struct iv_use *closest_after = NULL; |
5727 | if (cand->pos != IP_ORIGINAL) |
5728 | continue; |
5729 | |
5730 | for (j = 0; j < data->vgroups.length (); j++) |
5731 | { |
5732 | struct iv_group *group = data->vgroups[j]; |
5733 | struct iv_use *use = group->vuses[0]; |
5734 | unsigned uid = gimple_uid (g: use->stmt); |
5735 | |
5736 | if (gimple_bb (g: use->stmt) != gimple_bb (g: cand->incremented_at)) |
5737 | continue; |
5738 | |
5739 | if (uid < gimple_uid (g: cand->incremented_at) |
5740 | && (closest_before == NULL |
5741 | || uid > gimple_uid (g: closest_before->stmt))) |
5742 | closest_before = use; |
5743 | |
5744 | if (uid > gimple_uid (g: cand->incremented_at) |
5745 | && (closest_after == NULL |
5746 | || uid < gimple_uid (g: closest_after->stmt))) |
5747 | closest_after = use; |
5748 | } |
5749 | |
5750 | if (closest_before != NULL |
5751 | && autoinc_possible_for_pair (data, use: closest_before, cand)) |
5752 | cand->ainc_use = closest_before; |
5753 | else if (closest_after != NULL |
5754 | && autoinc_possible_for_pair (data, use: closest_after, cand)) |
5755 | cand->ainc_use = closest_after; |
5756 | } |
5757 | } |
5758 | |
5759 | /* Relate compare use with all candidates. */ |
5760 | |
5761 | static void |
5762 | relate_compare_use_with_all_cands (struct ivopts_data *data) |
5763 | { |
5764 | unsigned i, count = data->vcands.length (); |
5765 | for (i = 0; i < data->vgroups.length (); i++) |
5766 | { |
5767 | struct iv_group *group = data->vgroups[i]; |
5768 | |
5769 | if (group->type == USE_COMPARE) |
5770 | bitmap_set_range (group->related_cands, 0, count); |
5771 | } |
5772 | } |
5773 | |
5774 | /* If PREFERRED_MODE is suitable and profitable, use the preferred |
5775 | PREFERRED_MODE to compute doloop iv base from niter: base = niter + 1. */ |
5776 | |
5777 | static tree |
5778 | compute_doloop_base_on_mode (machine_mode preferred_mode, tree niter, |
5779 | const widest_int &iterations_max) |
5780 | { |
5781 | tree ntype = TREE_TYPE (niter); |
5782 | tree pref_type = lang_hooks.types.type_for_mode (preferred_mode, 1); |
5783 | if (!pref_type) |
5784 | return fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), |
5785 | build_int_cst (ntype, 1)); |
5786 | |
5787 | gcc_assert (TREE_CODE (pref_type) == INTEGER_TYPE); |
5788 | |
5789 | int prec = TYPE_PRECISION (ntype); |
5790 | int pref_prec = TYPE_PRECISION (pref_type); |
5791 | |
5792 | tree base; |
5793 | |
5794 | /* Check if the PREFERRED_MODED is able to present niter. */ |
5795 | if (pref_prec > prec |
5796 | || wi::ltu_p (x: iterations_max, |
5797 | y: widest_int::from (x: wi::max_value (pref_prec, UNSIGNED), |
5798 | sgn: UNSIGNED))) |
5799 | { |
5800 | /* No wrap, it is safe to use preferred type after niter + 1. */ |
5801 | if (wi::ltu_p (x: iterations_max, |
5802 | y: widest_int::from (x: wi::max_value (prec, UNSIGNED), |
5803 | sgn: UNSIGNED))) |
5804 | { |
5805 | /* This could help to optimize "-1 +1" pair when niter looks |
5806 | like "n-1": n is in original mode. "base = (n - 1) + 1" |
5807 | in PREFERRED_MODED: it could be base = (PREFERRED_TYPE)n. */ |
5808 | base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), |
5809 | build_int_cst (ntype, 1)); |
5810 | base = fold_convert (pref_type, base); |
5811 | } |
5812 | |
5813 | /* To avoid wrap, convert niter to preferred type before plus 1. */ |
5814 | else |
5815 | { |
5816 | niter = fold_convert (pref_type, niter); |
5817 | base = fold_build2 (PLUS_EXPR, pref_type, unshare_expr (niter), |
5818 | build_int_cst (pref_type, 1)); |
5819 | } |
5820 | } |
5821 | else |
5822 | base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), |
5823 | build_int_cst (ntype, 1)); |
5824 | return base; |
5825 | } |
5826 | |
5827 | /* Add one doloop dedicated IV candidate: |
5828 | - Base is (may_be_zero ? 1 : (niter + 1)). |
5829 | - Step is -1. */ |
5830 | |
5831 | static void |
5832 | add_iv_candidate_for_doloop (struct ivopts_data *data) |
5833 | { |
5834 | tree_niter_desc *niter_desc = niter_for_single_dom_exit (data); |
5835 | gcc_assert (niter_desc && niter_desc->assumptions); |
5836 | |
5837 | tree niter = niter_desc->niter; |
5838 | tree ntype = TREE_TYPE (niter); |
5839 | gcc_assert (TREE_CODE (ntype) == INTEGER_TYPE); |
5840 | |
5841 | tree may_be_zero = niter_desc->may_be_zero; |
5842 | if (may_be_zero && integer_zerop (may_be_zero)) |
5843 | may_be_zero = NULL_TREE; |
5844 | if (may_be_zero) |
5845 | { |
5846 | if (COMPARISON_CLASS_P (may_be_zero)) |
5847 | { |
5848 | niter = fold_build3 (COND_EXPR, ntype, may_be_zero, |
5849 | build_int_cst (ntype, 0), |
5850 | rewrite_to_non_trapping_overflow (niter)); |
5851 | } |
5852 | /* Don't try to obtain the iteration count expression when may_be_zero is |
5853 | integer_nonzerop (actually iteration count is one) or else. */ |
5854 | else |
5855 | return; |
5856 | } |
5857 | |
5858 | machine_mode mode = TYPE_MODE (ntype); |
5859 | machine_mode pref_mode = targetm.preferred_doloop_mode (mode); |
5860 | |
5861 | tree base; |
5862 | if (mode != pref_mode) |
5863 | { |
5864 | base = compute_doloop_base_on_mode (preferred_mode: pref_mode, niter, iterations_max: niter_desc->max); |
5865 | ntype = TREE_TYPE (base); |
5866 | } |
5867 | else |
5868 | base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), |
5869 | build_int_cst (ntype, 1)); |
5870 | |
5871 | |
5872 | add_candidate (data, base, step: build_int_cst (ntype, -1), important: true, NULL, NULL, doloop: true); |
5873 | } |
5874 | |
5875 | /* Finds the candidates for the induction variables. */ |
5876 | |
5877 | static void |
5878 | find_iv_candidates (struct ivopts_data *data) |
5879 | { |
5880 | /* Add commonly used ivs. */ |
5881 | add_standard_iv_candidates (data); |
5882 | |
5883 | /* Add doloop dedicated ivs. */ |
5884 | if (data->doloop_use_p) |
5885 | add_iv_candidate_for_doloop (data); |
5886 | |
5887 | /* Add old induction variables. */ |
5888 | add_iv_candidate_for_bivs (data); |
5889 | |
5890 | /* Add induction variables derived from uses. */ |
5891 | add_iv_candidate_for_groups (data); |
5892 | |
5893 | set_autoinc_for_original_candidates (data); |
5894 | |
5895 | /* Record the important candidates. */ |
5896 | record_important_candidates (data); |
5897 | |
5898 | /* Relate compare iv_use with all candidates. */ |
5899 | if (!data->consider_all_candidates) |
5900 | relate_compare_use_with_all_cands (data); |
5901 | |
5902 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5903 | { |
5904 | unsigned i; |
5905 | |
5906 | fprintf (stream: dump_file, format: "\n<Important Candidates>:\t" ); |
5907 | for (i = 0; i < data->vcands.length (); i++) |
5908 | if (data->vcands[i]->important) |
5909 | fprintf (stream: dump_file, format: " %d," , data->vcands[i]->id); |
5910 | fprintf (stream: dump_file, format: "\n" ); |
5911 | |
5912 | fprintf (stream: dump_file, format: "\n<Group, Cand> Related:\n" ); |
5913 | for (i = 0; i < data->vgroups.length (); i++) |
5914 | { |
5915 | struct iv_group *group = data->vgroups[i]; |
5916 | |
5917 | if (group->related_cands) |
5918 | { |
5919 | fprintf (stream: dump_file, format: " Group %d:\t" , group->id); |
5920 | dump_bitmap (file: dump_file, map: group->related_cands); |
5921 | } |
5922 | } |
5923 | fprintf (stream: dump_file, format: "\n" ); |
5924 | } |
5925 | } |
5926 | |
5927 | /* Determines costs of computing use of iv with an iv candidate. */ |
5928 | |
5929 | static void |
5930 | determine_group_iv_costs (struct ivopts_data *data) |
5931 | { |
5932 | unsigned i, j; |
5933 | struct iv_cand *cand; |
5934 | struct iv_group *group; |
5935 | bitmap to_clear = BITMAP_ALLOC (NULL); |
5936 | |
5937 | alloc_use_cost_map (data); |
5938 | |
5939 | for (i = 0; i < data->vgroups.length (); i++) |
5940 | { |
5941 | group = data->vgroups[i]; |
5942 | |
5943 | if (data->consider_all_candidates) |
5944 | { |
5945 | for (j = 0; j < data->vcands.length (); j++) |
5946 | { |
5947 | cand = data->vcands[j]; |
5948 | determine_group_iv_cost (data, group, cand); |
5949 | } |
5950 | } |
5951 | else |
5952 | { |
5953 | bitmap_iterator bi; |
5954 | |
5955 | EXECUTE_IF_SET_IN_BITMAP (group->related_cands, 0, j, bi) |
5956 | { |
5957 | cand = data->vcands[j]; |
5958 | if (!determine_group_iv_cost (data, group, cand)) |
5959 | bitmap_set_bit (to_clear, j); |
5960 | } |
5961 | |
5962 | /* Remove the candidates for that the cost is infinite from |
5963 | the list of related candidates. */ |
5964 | bitmap_and_compl_into (group->related_cands, to_clear); |
5965 | bitmap_clear (to_clear); |
5966 | } |
5967 | } |
5968 | |
5969 | BITMAP_FREE (to_clear); |
5970 | |
5971 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5972 | { |
5973 | bitmap_iterator bi; |
5974 | |
5975 | /* Dump invariant variables. */ |
5976 | fprintf (stream: dump_file, format: "\n<Invariant Vars>:\n" ); |
5977 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, i, bi) |
5978 | { |
5979 | struct version_info *info = ver_info (data, ver: i); |
5980 | if (info->inv_id) |
5981 | { |
5982 | fprintf (stream: dump_file, format: "Inv %d:\t" , info->inv_id); |
5983 | print_generic_expr (dump_file, info->name, TDF_SLIM); |
5984 | fprintf (stream: dump_file, format: "%s\n" , |
5985 | info->has_nonlin_use ? "" : "\t(eliminable)" ); |
5986 | } |
5987 | } |
5988 | |
5989 | /* Dump invariant expressions. */ |
5990 | fprintf (stream: dump_file, format: "\n<Invariant Expressions>:\n" ); |
5991 | auto_vec <iv_inv_expr_ent *> list (data->inv_expr_tab->elements ()); |
5992 | |
5993 | for (hash_table<iv_inv_expr_hasher>::iterator it |
5994 | = data->inv_expr_tab->begin (); it != data->inv_expr_tab->end (); |
5995 | ++it) |
5996 | list.safe_push (obj: *it); |
5997 | |
5998 | list.qsort (sort_iv_inv_expr_ent); |
5999 | |
6000 | for (i = 0; i < list.length (); ++i) |
6001 | { |
6002 | fprintf (stream: dump_file, format: "inv_expr %d: \t" , list[i]->id); |
6003 | print_generic_expr (dump_file, list[i]->expr, TDF_SLIM); |
6004 | fprintf (stream: dump_file, format: "\n" ); |
6005 | } |
6006 | |
6007 | fprintf (stream: dump_file, format: "\n<Group-candidate Costs>:\n" ); |
6008 | |
6009 | for (i = 0; i < data->vgroups.length (); i++) |
6010 | { |
6011 | group = data->vgroups[i]; |
6012 | |
6013 | fprintf (stream: dump_file, format: "Group %d:\n" , i); |
6014 | fprintf (stream: dump_file, format: " cand\tcost\tcompl.\tinv.expr.\tinv.vars\n" ); |
6015 | for (j = 0; j < group->n_map_members; j++) |
6016 | { |
6017 | if (!group->cost_map[j].cand |
6018 | || group->cost_map[j].cost.infinite_cost_p ()) |
6019 | continue; |
6020 | |
6021 | fprintf (stream: dump_file, format: " %d\t%" PRId64 "\t%d\t" , |
6022 | group->cost_map[j].cand->id, |
6023 | group->cost_map[j].cost.cost, |
6024 | group->cost_map[j].cost.complexity); |
6025 | if (!group->cost_map[j].inv_exprs |
6026 | || bitmap_empty_p (map: group->cost_map[j].inv_exprs)) |
6027 | fprintf (stream: dump_file, format: "NIL;\t" ); |
6028 | else |
6029 | bitmap_print (dump_file, |
6030 | group->cost_map[j].inv_exprs, "" , ";\t" ); |
6031 | if (!group->cost_map[j].inv_vars |
6032 | || bitmap_empty_p (map: group->cost_map[j].inv_vars)) |
6033 | fprintf (stream: dump_file, format: "NIL;\n" ); |
6034 | else |
6035 | bitmap_print (dump_file, |
6036 | group->cost_map[j].inv_vars, "" , "\n" ); |
6037 | } |
6038 | |
6039 | fprintf (stream: dump_file, format: "\n" ); |
6040 | } |
6041 | fprintf (stream: dump_file, format: "\n" ); |
6042 | } |
6043 | } |
6044 | |
6045 | /* Determines cost of the candidate CAND. */ |
6046 | |
6047 | static void |
6048 | determine_iv_cost (struct ivopts_data *data, struct iv_cand *cand) |
6049 | { |
6050 | comp_cost cost_base; |
6051 | int64_t cost, cost_step; |
6052 | tree base; |
6053 | |
6054 | gcc_assert (cand->iv != NULL); |
6055 | |
6056 | /* There are two costs associated with the candidate -- its increment |
6057 | and its initialization. The second is almost negligible for any loop |
6058 | that rolls enough, so we take it just very little into account. */ |
6059 | |
6060 | base = cand->iv->base; |
6061 | cost_base = force_var_cost (data, expr: base, NULL); |
6062 | /* It will be exceptional that the iv register happens to be initialized with |
6063 | the proper value at no cost. In general, there will at least be a regcopy |
6064 | or a const set. */ |
6065 | if (cost_base.cost == 0) |
6066 | cost_base.cost = COSTS_N_INSNS (1); |
6067 | /* Doloop decrement should be considered as zero cost. */ |
6068 | if (cand->doloop_p) |
6069 | cost_step = 0; |
6070 | else |
6071 | cost_step = add_cost (speed: data->speed, TYPE_MODE (TREE_TYPE (base))); |
6072 | cost = cost_step + adjust_setup_cost (data, cost: cost_base.cost); |
6073 | |
6074 | /* Prefer the original ivs unless we may gain something by replacing it. |
6075 | The reason is to make debugging simpler; so this is not relevant for |
6076 | artificial ivs created by other optimization passes. */ |
6077 | if ((cand->pos != IP_ORIGINAL |
6078 | || !SSA_NAME_VAR (cand->var_before) |
6079 | || DECL_ARTIFICIAL (SSA_NAME_VAR (cand->var_before))) |
6080 | /* Prefer doloop as well. */ |
6081 | && !cand->doloop_p) |
6082 | cost++; |
6083 | |
6084 | /* Prefer not to insert statements into latch unless there are some |
6085 | already (so that we do not create unnecessary jumps). */ |
6086 | if (cand->pos == IP_END |
6087 | && empty_block_p (ip_end_pos (data->current_loop))) |
6088 | cost++; |
6089 | |
6090 | cand->cost = cost; |
6091 | cand->cost_step = cost_step; |
6092 | } |
6093 | |
6094 | /* Determines costs of computation of the candidates. */ |
6095 | |
6096 | static void |
6097 | determine_iv_costs (struct ivopts_data *data) |
6098 | { |
6099 | unsigned i; |
6100 | |
6101 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6102 | { |
6103 | fprintf (stream: dump_file, format: "<Candidate Costs>:\n" ); |
6104 | fprintf (stream: dump_file, format: " cand\tcost\n" ); |
6105 | } |
6106 | |
6107 | for (i = 0; i < data->vcands.length (); i++) |
6108 | { |
6109 | struct iv_cand *cand = data->vcands[i]; |
6110 | |
6111 | determine_iv_cost (data, cand); |
6112 | |
6113 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6114 | fprintf (stream: dump_file, format: " %d\t%d\n" , i, cand->cost); |
6115 | } |
6116 | |
6117 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6118 | fprintf (stream: dump_file, format: "\n" ); |
6119 | } |
6120 | |
6121 | /* Estimate register pressure for loop having N_INVS invariants and N_CANDS |
6122 | induction variables. Note N_INVS includes both invariant variables and |
6123 | invariant expressions. */ |
6124 | |
6125 | static unsigned |
6126 | ivopts_estimate_reg_pressure (struct ivopts_data *data, unsigned n_invs, |
6127 | unsigned n_cands) |
6128 | { |
6129 | unsigned cost; |
6130 | unsigned n_old = data->regs_used, n_new = n_invs + n_cands; |
6131 | unsigned regs_needed = n_new + n_old, available_regs = target_avail_regs; |
6132 | bool speed = data->speed; |
6133 | |
6134 | /* If there is a call in the loop body, the call-clobbered registers |
6135 | are not available for loop invariants. */ |
6136 | if (data->body_includes_call) |
6137 | available_regs = available_regs - target_clobbered_regs; |
6138 | |
6139 | /* If we have enough registers. */ |
6140 | if (regs_needed + target_res_regs < available_regs) |
6141 | cost = n_new; |
6142 | /* If close to running out of registers, try to preserve them. */ |
6143 | else if (regs_needed <= available_regs) |
6144 | cost = target_reg_cost [speed] * regs_needed; |
6145 | /* If we run out of available registers but the number of candidates |
6146 | does not, we penalize extra registers using target_spill_cost. */ |
6147 | else if (n_cands <= available_regs) |
6148 | cost = target_reg_cost [speed] * available_regs |
6149 | + target_spill_cost [speed] * (regs_needed - available_regs); |
6150 | /* If the number of candidates runs out available registers, we penalize |
6151 | extra candidate registers using target_spill_cost * 2. Because it is |
6152 | more expensive to spill induction variable than invariant. */ |
6153 | else |
6154 | cost = target_reg_cost [speed] * available_regs |
6155 | + target_spill_cost [speed] * (n_cands - available_regs) * 2 |
6156 | + target_spill_cost [speed] * (regs_needed - n_cands); |
6157 | |
6158 | /* Finally, add the number of candidates, so that we prefer eliminating |
6159 | induction variables if possible. */ |
6160 | return cost + n_cands; |
6161 | } |
6162 | |
6163 | /* For each size of the induction variable set determine the penalty. */ |
6164 | |
6165 | static void |
6166 | determine_set_costs (struct ivopts_data *data) |
6167 | { |
6168 | unsigned j, n; |
6169 | gphi *phi; |
6170 | gphi_iterator psi; |
6171 | tree op; |
6172 | class loop *loop = data->current_loop; |
6173 | bitmap_iterator bi; |
6174 | |
6175 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6176 | { |
6177 | fprintf (stream: dump_file, format: "<Global Costs>:\n" ); |
6178 | fprintf (stream: dump_file, format: " target_avail_regs %d\n" , target_avail_regs); |
6179 | fprintf (stream: dump_file, format: " target_clobbered_regs %d\n" , target_clobbered_regs); |
6180 | fprintf (stream: dump_file, format: " target_reg_cost %d\n" , target_reg_cost[data->speed]); |
6181 | fprintf (stream: dump_file, format: " target_spill_cost %d\n" , target_spill_cost[data->speed]); |
6182 | } |
6183 | |
6184 | n = 0; |
6185 | for (psi = gsi_start_phis (loop->header); !gsi_end_p (i: psi); gsi_next (i: &psi)) |
6186 | { |
6187 | phi = psi.phi (); |
6188 | op = PHI_RESULT (phi); |
6189 | |
6190 | if (virtual_operand_p (op)) |
6191 | continue; |
6192 | |
6193 | if (get_iv (data, var: op)) |
6194 | continue; |
6195 | |
6196 | if (!POINTER_TYPE_P (TREE_TYPE (op)) |
6197 | && !INTEGRAL_TYPE_P (TREE_TYPE (op))) |
6198 | continue; |
6199 | |
6200 | n++; |
6201 | } |
6202 | |
6203 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, j, bi) |
6204 | { |
6205 | struct version_info *info = ver_info (data, ver: j); |
6206 | |
6207 | if (info->inv_id && info->has_nonlin_use) |
6208 | n++; |
6209 | } |
6210 | |
6211 | data->regs_used = n; |
6212 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6213 | fprintf (stream: dump_file, format: " regs_used %d\n" , n); |
6214 | |
6215 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6216 | { |
6217 | fprintf (stream: dump_file, format: " cost for size:\n" ); |
6218 | fprintf (stream: dump_file, format: " ivs\tcost\n" ); |
6219 | for (j = 0; j <= 2 * target_avail_regs; j++) |
6220 | fprintf (stream: dump_file, format: " %d\t%d\n" , j, |
6221 | ivopts_estimate_reg_pressure (data, n_invs: 0, n_cands: j)); |
6222 | fprintf (stream: dump_file, format: "\n" ); |
6223 | } |
6224 | } |
6225 | |
6226 | /* Returns true if A is a cheaper cost pair than B. */ |
6227 | |
6228 | static bool |
6229 | cheaper_cost_pair (class cost_pair *a, class cost_pair *b) |
6230 | { |
6231 | if (!a) |
6232 | return false; |
6233 | |
6234 | if (!b) |
6235 | return true; |
6236 | |
6237 | if (a->cost < b->cost) |
6238 | return true; |
6239 | |
6240 | if (b->cost < a->cost) |
6241 | return false; |
6242 | |
6243 | /* In case the costs are the same, prefer the cheaper candidate. */ |
6244 | if (a->cand->cost < b->cand->cost) |
6245 | return true; |
6246 | |
6247 | return false; |
6248 | } |
6249 | |
6250 | /* Compare if A is a more expensive cost pair than B. Return 1, 0 and -1 |
6251 | for more expensive, equal and cheaper respectively. */ |
6252 | |
6253 | static int |
6254 | compare_cost_pair (class cost_pair *a, class cost_pair *b) |
6255 | { |
6256 | if (cheaper_cost_pair (a, b)) |
6257 | return -1; |
6258 | if (cheaper_cost_pair (a: b, b: a)) |
6259 | return 1; |
6260 | |
6261 | return 0; |
6262 | } |
6263 | |
6264 | /* Returns candidate by that USE is expressed in IVS. */ |
6265 | |
6266 | static class cost_pair * |
6267 | iv_ca_cand_for_group (class iv_ca *ivs, struct iv_group *group) |
6268 | { |
6269 | return ivs->cand_for_group[group->id]; |
6270 | } |
6271 | |
6272 | /* Computes the cost field of IVS structure. */ |
6273 | |
6274 | static void |
6275 | iv_ca_recount_cost (struct ivopts_data *data, class iv_ca *ivs) |
6276 | { |
6277 | comp_cost cost = ivs->cand_use_cost; |
6278 | |
6279 | cost += ivs->cand_cost; |
6280 | cost += ivopts_estimate_reg_pressure (data, n_invs: ivs->n_invs, n_cands: ivs->n_cands); |
6281 | ivs->cost = cost; |
6282 | } |
6283 | |
6284 | /* Remove use of invariants in set INVS by decreasing counter in N_INV_USES |
6285 | and IVS. */ |
6286 | |
6287 | static void |
6288 | iv_ca_set_remove_invs (class iv_ca *ivs, bitmap invs, unsigned *n_inv_uses) |
6289 | { |
6290 | bitmap_iterator bi; |
6291 | unsigned iid; |
6292 | |
6293 | if (!invs) |
6294 | return; |
6295 | |
6296 | gcc_assert (n_inv_uses != NULL); |
6297 | EXECUTE_IF_SET_IN_BITMAP (invs, 0, iid, bi) |
6298 | { |
6299 | n_inv_uses[iid]--; |
6300 | if (n_inv_uses[iid] == 0) |
6301 | ivs->n_invs--; |
6302 | } |
6303 | } |
6304 | |
6305 | /* Set USE not to be expressed by any candidate in IVS. */ |
6306 | |
6307 | static void |
6308 | iv_ca_set_no_cp (struct ivopts_data *data, class iv_ca *ivs, |
6309 | struct iv_group *group) |
6310 | { |
6311 | unsigned gid = group->id, cid; |
6312 | class cost_pair *cp; |
6313 | |
6314 | cp = ivs->cand_for_group[gid]; |
6315 | if (!cp) |
6316 | return; |
6317 | cid = cp->cand->id; |
6318 | |
6319 | ivs->bad_groups++; |
6320 | ivs->cand_for_group[gid] = NULL; |
6321 | ivs->n_cand_uses[cid]--; |
6322 | |
6323 | if (ivs->n_cand_uses[cid] == 0) |
6324 | { |
6325 | bitmap_clear_bit (ivs->cands, cid); |
6326 | if (!cp->cand->doloop_p || !targetm.have_count_reg_decr_p) |
6327 | ivs->n_cands--; |
6328 | ivs->cand_cost -= cp->cand->cost; |
6329 | iv_ca_set_remove_invs (ivs, invs: cp->cand->inv_vars, n_inv_uses: ivs->n_inv_var_uses); |
6330 | iv_ca_set_remove_invs (ivs, invs: cp->cand->inv_exprs, n_inv_uses: ivs->n_inv_expr_uses); |
6331 | } |
6332 | |
6333 | ivs->cand_use_cost -= cp->cost; |
6334 | iv_ca_set_remove_invs (ivs, invs: cp->inv_vars, n_inv_uses: ivs->n_inv_var_uses); |
6335 | iv_ca_set_remove_invs (ivs, invs: cp->inv_exprs, n_inv_uses: ivs->n_inv_expr_uses); |
6336 | iv_ca_recount_cost (data, ivs); |
6337 | } |
6338 | |
6339 | /* Add use of invariants in set INVS by increasing counter in N_INV_USES and |
6340 | IVS. */ |
6341 | |
6342 | static void |
6343 | iv_ca_set_add_invs (class iv_ca *ivs, bitmap invs, unsigned *n_inv_uses) |
6344 | { |
6345 | bitmap_iterator bi; |
6346 | unsigned iid; |
6347 | |
6348 | if (!invs) |
6349 | return; |
6350 | |
6351 | gcc_assert (n_inv_uses != NULL); |
6352 | EXECUTE_IF_SET_IN_BITMAP (invs, 0, iid, bi) |
6353 | { |
6354 | n_inv_uses[iid]++; |
6355 | if (n_inv_uses[iid] == 1) |
6356 | ivs->n_invs++; |
6357 | } |
6358 | } |
6359 | |
6360 | /* Set cost pair for GROUP in set IVS to CP. */ |
6361 | |
6362 | static void |
6363 | iv_ca_set_cp (struct ivopts_data *data, class iv_ca *ivs, |
6364 | struct iv_group *group, class cost_pair *cp) |
6365 | { |
6366 | unsigned gid = group->id, cid; |
6367 | |
6368 | if (ivs->cand_for_group[gid] == cp) |
6369 | return; |
6370 | |
6371 | if (ivs->cand_for_group[gid]) |
6372 | iv_ca_set_no_cp (data, ivs, group); |
6373 | |
6374 | if (cp) |
6375 | { |
6376 | cid = cp->cand->id; |
6377 | |
6378 | ivs->bad_groups--; |
6379 | ivs->cand_for_group[gid] = cp; |
6380 | ivs->n_cand_uses[cid]++; |
6381 | if (ivs->n_cand_uses[cid] == 1) |
6382 | { |
6383 | bitmap_set_bit (ivs->cands, cid); |
6384 | if (!cp->cand->doloop_p || !targetm.have_count_reg_decr_p) |
6385 | ivs->n_cands++; |
6386 | ivs->cand_cost += cp->cand->cost; |
6387 | iv_ca_set_add_invs (ivs, invs: cp->cand->inv_vars, n_inv_uses: ivs->n_inv_var_uses); |
6388 | iv_ca_set_add_invs (ivs, invs: cp->cand->inv_exprs, n_inv_uses: ivs->n_inv_expr_uses); |
6389 | } |
6390 | |
6391 | ivs->cand_use_cost += cp->cost; |
6392 | iv_ca_set_add_invs (ivs, invs: cp->inv_vars, n_inv_uses: ivs->n_inv_var_uses); |
6393 | iv_ca_set_add_invs (ivs, invs: cp->inv_exprs, n_inv_uses: ivs->n_inv_expr_uses); |
6394 | iv_ca_recount_cost (data, ivs); |
6395 | } |
6396 | } |
6397 | |
6398 | /* Extend set IVS by expressing USE by some of the candidates in it |
6399 | if possible. Consider all important candidates if candidates in |
6400 | set IVS don't give any result. */ |
6401 | |
6402 | static void |
6403 | iv_ca_add_group (struct ivopts_data *data, class iv_ca *ivs, |
6404 | struct iv_group *group) |
6405 | { |
6406 | class cost_pair *best_cp = NULL, *cp; |
6407 | bitmap_iterator bi; |
6408 | unsigned i; |
6409 | struct iv_cand *cand; |
6410 | |
6411 | gcc_assert (ivs->upto >= group->id); |
6412 | ivs->upto++; |
6413 | ivs->bad_groups++; |
6414 | |
6415 | EXECUTE_IF_SET_IN_BITMAP (ivs->cands, 0, i, bi) |
6416 | { |
6417 | cand = data->vcands[i]; |
6418 | cp = get_group_iv_cost (data, group, cand); |
6419 | if (cheaper_cost_pair (a: cp, b: best_cp)) |
6420 | best_cp = cp; |
6421 | } |
6422 | |
6423 | if (best_cp == NULL) |
6424 | { |
6425 | EXECUTE_IF_SET_IN_BITMAP (data->important_candidates, 0, i, bi) |
6426 | { |
6427 | cand = data->vcands[i]; |
6428 | cp = get_group_iv_cost (data, group, cand); |
6429 | if (cheaper_cost_pair (a: cp, b: best_cp)) |
6430 | best_cp = cp; |
6431 | } |
6432 | } |
6433 | |
6434 | iv_ca_set_cp (data, ivs, group, cp: best_cp); |
6435 | } |
6436 | |
6437 | /* Get cost for assignment IVS. */ |
6438 | |
6439 | static comp_cost |
6440 | iv_ca_cost (class iv_ca *ivs) |
6441 | { |
6442 | /* This was a conditional expression but it triggered a bug in |
6443 | Sun C 5.5. */ |
6444 | if (ivs->bad_groups) |
6445 | return infinite_cost; |
6446 | else |
6447 | return ivs->cost; |
6448 | } |
6449 | |
6450 | /* Compare if applying NEW_CP to GROUP for IVS introduces more invariants |
6451 | than OLD_CP. Return 1, 0 and -1 for more, equal and fewer invariants |
6452 | respectively. */ |
6453 | |
6454 | static int |
6455 | iv_ca_compare_deps (struct ivopts_data *data, class iv_ca *ivs, |
6456 | struct iv_group *group, class cost_pair *old_cp, |
6457 | class cost_pair *new_cp) |
6458 | { |
6459 | gcc_assert (old_cp && new_cp && old_cp != new_cp); |
6460 | unsigned old_n_invs = ivs->n_invs; |
6461 | iv_ca_set_cp (data, ivs, group, cp: new_cp); |
6462 | unsigned new_n_invs = ivs->n_invs; |
6463 | iv_ca_set_cp (data, ivs, group, cp: old_cp); |
6464 | |
6465 | return new_n_invs > old_n_invs ? 1 : (new_n_invs < old_n_invs ? -1 : 0); |
6466 | } |
6467 | |
6468 | /* Creates change of expressing GROUP by NEW_CP instead of OLD_CP and chains |
6469 | it before NEXT. */ |
6470 | |
6471 | static struct iv_ca_delta * |
6472 | iv_ca_delta_add (struct iv_group *group, class cost_pair *old_cp, |
6473 | class cost_pair *new_cp, struct iv_ca_delta *next) |
6474 | { |
6475 | struct iv_ca_delta *change = XNEW (struct iv_ca_delta); |
6476 | |
6477 | change->group = group; |
6478 | change->old_cp = old_cp; |
6479 | change->new_cp = new_cp; |
6480 | change->next = next; |
6481 | |
6482 | return change; |
6483 | } |
6484 | |
6485 | /* Joins two lists of changes L1 and L2. Destructive -- old lists |
6486 | are rewritten. */ |
6487 | |
6488 | static struct iv_ca_delta * |
6489 | iv_ca_delta_join (struct iv_ca_delta *l1, struct iv_ca_delta *l2) |
6490 | { |
6491 | struct iv_ca_delta *last; |
6492 | |
6493 | if (!l2) |
6494 | return l1; |
6495 | |
6496 | if (!l1) |
6497 | return l2; |
6498 | |
6499 | for (last = l1; last->next; last = last->next) |
6500 | continue; |
6501 | last->next = l2; |
6502 | |
6503 | return l1; |
6504 | } |
6505 | |
6506 | /* Reverse the list of changes DELTA, forming the inverse to it. */ |
6507 | |
6508 | static struct iv_ca_delta * |
6509 | iv_ca_delta_reverse (struct iv_ca_delta *delta) |
6510 | { |
6511 | struct iv_ca_delta *act, *next, *prev = NULL; |
6512 | |
6513 | for (act = delta; act; act = next) |
6514 | { |
6515 | next = act->next; |
6516 | act->next = prev; |
6517 | prev = act; |
6518 | |
6519 | std::swap (a&: act->old_cp, b&: act->new_cp); |
6520 | } |
6521 | |
6522 | return prev; |
6523 | } |
6524 | |
6525 | /* Commit changes in DELTA to IVS. If FORWARD is false, the changes are |
6526 | reverted instead. */ |
6527 | |
6528 | static void |
6529 | iv_ca_delta_commit (struct ivopts_data *data, class iv_ca *ivs, |
6530 | struct iv_ca_delta *delta, bool forward) |
6531 | { |
6532 | class cost_pair *from, *to; |
6533 | struct iv_ca_delta *act; |
6534 | |
6535 | if (!forward) |
6536 | delta = iv_ca_delta_reverse (delta); |
6537 | |
6538 | for (act = delta; act; act = act->next) |
6539 | { |
6540 | from = act->old_cp; |
6541 | to = act->new_cp; |
6542 | gcc_assert (iv_ca_cand_for_group (ivs, act->group) == from); |
6543 | iv_ca_set_cp (data, ivs, group: act->group, cp: to); |
6544 | } |
6545 | |
6546 | if (!forward) |
6547 | iv_ca_delta_reverse (delta); |
6548 | } |
6549 | |
6550 | /* Returns true if CAND is used in IVS. */ |
6551 | |
6552 | static bool |
6553 | iv_ca_cand_used_p (class iv_ca *ivs, struct iv_cand *cand) |
6554 | { |
6555 | return ivs->n_cand_uses[cand->id] > 0; |
6556 | } |
6557 | |
6558 | /* Returns number of induction variable candidates in the set IVS. */ |
6559 | |
6560 | static unsigned |
6561 | iv_ca_n_cands (class iv_ca *ivs) |
6562 | { |
6563 | return ivs->n_cands; |
6564 | } |
6565 | |
6566 | /* Free the list of changes DELTA. */ |
6567 | |
6568 | static void |
6569 | iv_ca_delta_free (struct iv_ca_delta **delta) |
6570 | { |
6571 | struct iv_ca_delta *act, *next; |
6572 | |
6573 | for (act = *delta; act; act = next) |
6574 | { |
6575 | next = act->next; |
6576 | free (ptr: act); |
6577 | } |
6578 | |
6579 | *delta = NULL; |
6580 | } |
6581 | |
6582 | /* Allocates new iv candidates assignment. */ |
6583 | |
6584 | static class iv_ca * |
6585 | iv_ca_new (struct ivopts_data *data) |
6586 | { |
6587 | class iv_ca *nw = XNEW (class iv_ca); |
6588 | |
6589 | nw->upto = 0; |
6590 | nw->bad_groups = 0; |
6591 | nw->cand_for_group = XCNEWVEC (class cost_pair *, |
6592 | data->vgroups.length ()); |
6593 | nw->n_cand_uses = XCNEWVEC (unsigned, data->vcands.length ()); |
6594 | nw->cands = BITMAP_ALLOC (NULL); |
6595 | nw->n_cands = 0; |
6596 | nw->n_invs = 0; |
6597 | nw->cand_use_cost = no_cost; |
6598 | nw->cand_cost = 0; |
6599 | nw->n_inv_var_uses = XCNEWVEC (unsigned, data->max_inv_var_id + 1); |
6600 | nw->n_inv_expr_uses = XCNEWVEC (unsigned, data->max_inv_expr_id + 1); |
6601 | nw->cost = no_cost; |
6602 | |
6603 | return nw; |
6604 | } |
6605 | |
6606 | /* Free memory occupied by the set IVS. */ |
6607 | |
6608 | static void |
6609 | iv_ca_free (class iv_ca **ivs) |
6610 | { |
6611 | free (ptr: (*ivs)->cand_for_group); |
6612 | free (ptr: (*ivs)->n_cand_uses); |
6613 | BITMAP_FREE ((*ivs)->cands); |
6614 | free (ptr: (*ivs)->n_inv_var_uses); |
6615 | free (ptr: (*ivs)->n_inv_expr_uses); |
6616 | free (ptr: *ivs); |
6617 | *ivs = NULL; |
6618 | } |
6619 | |
6620 | /* Dumps IVS to FILE. */ |
6621 | |
6622 | static void |
6623 | iv_ca_dump (struct ivopts_data *data, FILE *file, class iv_ca *ivs) |
6624 | { |
6625 | unsigned i; |
6626 | comp_cost cost = iv_ca_cost (ivs); |
6627 | |
6628 | fprintf (stream: file, format: " cost: %" PRId64 " (complexity %d)\n" , cost.cost, |
6629 | cost.complexity); |
6630 | fprintf (stream: file, format: " reg_cost: %d\n" , |
6631 | ivopts_estimate_reg_pressure (data, n_invs: ivs->n_invs, n_cands: ivs->n_cands)); |
6632 | fprintf (stream: file, format: " cand_cost: %" PRId64 "\n cand_group_cost: " |
6633 | "%" PRId64 " (complexity %d)\n" , ivs->cand_cost, |
6634 | ivs->cand_use_cost.cost, ivs->cand_use_cost.complexity); |
6635 | bitmap_print (file, ivs->cands, " candidates: " ,"\n" ); |
6636 | |
6637 | for (i = 0; i < ivs->upto; i++) |
6638 | { |
6639 | struct iv_group *group = data->vgroups[i]; |
6640 | class cost_pair *cp = iv_ca_cand_for_group (ivs, group); |
6641 | if (cp) |
6642 | fprintf (stream: file, format: " group:%d --> iv_cand:%d, cost=(" |
6643 | "%" PRId64 ",%d)\n" , group->id, cp->cand->id, |
6644 | cp->cost.cost, cp->cost.complexity); |
6645 | else |
6646 | fprintf (stream: file, format: " group:%d --> ??\n" , group->id); |
6647 | } |
6648 | |
6649 | const char *pref = "" ; |
6650 | fprintf (stream: file, format: " invariant variables: " ); |
6651 | for (i = 1; i <= data->max_inv_var_id; i++) |
6652 | if (ivs->n_inv_var_uses[i]) |
6653 | { |
6654 | fprintf (stream: file, format: "%s%d" , pref, i); |
6655 | pref = ", " ; |
6656 | } |
6657 | |
6658 | pref = "" ; |
6659 | fprintf (stream: file, format: "\n invariant expressions: " ); |
6660 | for (i = 1; i <= data->max_inv_expr_id; i++) |
6661 | if (ivs->n_inv_expr_uses[i]) |
6662 | { |
6663 | fprintf (stream: file, format: "%s%d" , pref, i); |
6664 | pref = ", " ; |
6665 | } |
6666 | |
6667 | fprintf (stream: file, format: "\n\n" ); |
6668 | } |
6669 | |
6670 | /* Try changing candidate in IVS to CAND for each use. Return cost of the |
6671 | new set, and store differences in DELTA. Number of induction variables |
6672 | in the new set is stored to N_IVS. MIN_NCAND is a flag. When it is true |
6673 | the function will try to find a solution with mimimal iv candidates. */ |
6674 | |
6675 | static comp_cost |
6676 | iv_ca_extend (struct ivopts_data *data, class iv_ca *ivs, |
6677 | struct iv_cand *cand, struct iv_ca_delta **delta, |
6678 | unsigned *n_ivs, bool min_ncand) |
6679 | { |
6680 | unsigned i; |
6681 | comp_cost cost; |
6682 | struct iv_group *group; |
6683 | class cost_pair *old_cp, *new_cp; |
6684 | |
6685 | *delta = NULL; |
6686 | for (i = 0; i < ivs->upto; i++) |
6687 | { |
6688 | group = data->vgroups[i]; |
6689 | old_cp = iv_ca_cand_for_group (ivs, group); |
6690 | |
6691 | if (old_cp |
6692 | && old_cp->cand == cand) |
6693 | continue; |
6694 | |
6695 | new_cp = get_group_iv_cost (data, group, cand); |
6696 | if (!new_cp) |
6697 | continue; |
6698 | |
6699 | if (!min_ncand) |
6700 | { |
6701 | int cmp_invs = iv_ca_compare_deps (data, ivs, group, old_cp, new_cp); |
6702 | /* Skip if new_cp depends on more invariants. */ |
6703 | if (cmp_invs > 0) |
6704 | continue; |
6705 | |
6706 | int cmp_cost = compare_cost_pair (a: new_cp, b: old_cp); |
6707 | /* Skip if new_cp is not cheaper. */ |
6708 | if (cmp_cost > 0 || (cmp_cost == 0 && cmp_invs == 0)) |
6709 | continue; |
6710 | } |
6711 | |
6712 | *delta = iv_ca_delta_add (group, old_cp, new_cp, next: *delta); |
6713 | } |
6714 | |
6715 | iv_ca_delta_commit (data, ivs, delta: *delta, forward: true); |
6716 | cost = iv_ca_cost (ivs); |
6717 | if (n_ivs) |
6718 | *n_ivs = iv_ca_n_cands (ivs); |
6719 | iv_ca_delta_commit (data, ivs, delta: *delta, forward: false); |
6720 | |
6721 | return cost; |
6722 | } |
6723 | |
6724 | /* Try narrowing set IVS by removing CAND. Return the cost of |
6725 | the new set and store the differences in DELTA. START is |
6726 | the candidate with which we start narrowing. */ |
6727 | |
6728 | static comp_cost |
6729 | iv_ca_narrow (struct ivopts_data *data, class iv_ca *ivs, |
6730 | struct iv_cand *cand, struct iv_cand *start, |
6731 | struct iv_ca_delta **delta) |
6732 | { |
6733 | unsigned i, ci; |
6734 | struct iv_group *group; |
6735 | class cost_pair *old_cp, *new_cp, *cp; |
6736 | bitmap_iterator bi; |
6737 | struct iv_cand *cnd; |
6738 | comp_cost cost, best_cost, acost; |
6739 | |
6740 | *delta = NULL; |
6741 | for (i = 0; i < data->vgroups.length (); i++) |
6742 | { |
6743 | group = data->vgroups[i]; |
6744 | |
6745 | old_cp = iv_ca_cand_for_group (ivs, group); |
6746 | if (old_cp->cand != cand) |
6747 | continue; |
6748 | |
6749 | best_cost = iv_ca_cost (ivs); |
6750 | /* Start narrowing with START. */ |
6751 | new_cp = get_group_iv_cost (data, group, cand: start); |
6752 | |
6753 | if (data->consider_all_candidates) |
6754 | { |
6755 | EXECUTE_IF_SET_IN_BITMAP (ivs->cands, 0, ci, bi) |
6756 | { |
6757 | if (ci == cand->id || (start && ci == start->id)) |
6758 | continue; |
6759 | |
6760 | cnd = data->vcands[ci]; |
6761 | |
6762 | cp = get_group_iv_cost (data, group, cand: cnd); |
6763 | if (!cp) |
6764 | continue; |
6765 | |
6766 | iv_ca_set_cp (data, ivs, group, cp); |
6767 | acost = iv_ca_cost (ivs); |
6768 | |
6769 | if (acost < best_cost) |
6770 | { |
6771 | best_cost = acost; |
6772 | new_cp = cp; |
6773 | } |
6774 | } |
6775 | } |
6776 | else |
6777 | { |
6778 | EXECUTE_IF_AND_IN_BITMAP (group->related_cands, ivs->cands, 0, ci, bi) |
6779 | { |
6780 | if (ci == cand->id || (start && ci == start->id)) |
6781 | continue; |
6782 | |
6783 | cnd = data->vcands[ci]; |
6784 | |
6785 | cp = get_group_iv_cost (data, group, cand: cnd); |
6786 | if (!cp) |
6787 | continue; |
6788 | |
6789 | iv_ca_set_cp (data, ivs, group, cp); |
6790 | acost = iv_ca_cost (ivs); |
6791 | |
6792 | if (acost < best_cost) |
6793 | { |
6794 | best_cost = acost; |
6795 | new_cp = cp; |
6796 | } |
6797 | } |
6798 | } |
6799 | /* Restore to old cp for use. */ |
6800 | iv_ca_set_cp (data, ivs, group, cp: old_cp); |
6801 | |
6802 | if (!new_cp) |
6803 | { |
6804 | iv_ca_delta_free (delta); |
6805 | return infinite_cost; |
6806 | } |
6807 | |
6808 | *delta = iv_ca_delta_add (group, old_cp, new_cp, next: *delta); |
6809 | } |
6810 | |
6811 | iv_ca_delta_commit (data, ivs, delta: *delta, forward: true); |
6812 | cost = iv_ca_cost (ivs); |
6813 | iv_ca_delta_commit (data, ivs, delta: *delta, forward: false); |
6814 | |
6815 | return cost; |
6816 | } |
6817 | |
6818 | /* Try optimizing the set of candidates IVS by removing candidates different |
6819 | from to EXCEPT_CAND from it. Return cost of the new set, and store |
6820 | differences in DELTA. */ |
6821 | |
6822 | static comp_cost |
6823 | iv_ca_prune (struct ivopts_data *data, class iv_ca *ivs, |
6824 | struct iv_cand *except_cand, struct iv_ca_delta **delta) |
6825 | { |
6826 | bitmap_iterator bi; |
6827 | struct iv_ca_delta *act_delta, *best_delta; |
6828 | unsigned i; |
6829 | comp_cost best_cost, acost; |
6830 | struct iv_cand *cand; |
6831 | |
6832 | best_delta = NULL; |
6833 | best_cost = iv_ca_cost (ivs); |
6834 | |
6835 | EXECUTE_IF_SET_IN_BITMAP (ivs->cands, 0, i, bi) |
6836 | { |
6837 | cand = data->vcands[i]; |
6838 | |
6839 | if (cand == except_cand) |
6840 | continue; |
6841 | |
6842 | acost = iv_ca_narrow (data, ivs, cand, start: except_cand, delta: &act_delta); |
6843 | |
6844 | if (acost < best_cost) |
6845 | { |
6846 | best_cost = acost; |
6847 | iv_ca_delta_free (delta: &best_delta); |
6848 | best_delta = act_delta; |
6849 | } |
6850 | else |
6851 | iv_ca_delta_free (delta: &act_delta); |
6852 | } |
6853 | |
6854 | if (!best_delta) |
6855 | { |
6856 | *delta = NULL; |
6857 | return best_cost; |
6858 | } |
6859 | |
6860 | /* Recurse to possibly remove other unnecessary ivs. */ |
6861 | iv_ca_delta_commit (data, ivs, delta: best_delta, forward: true); |
6862 | best_cost = iv_ca_prune (data, ivs, except_cand, delta); |
6863 | iv_ca_delta_commit (data, ivs, delta: best_delta, forward: false); |
6864 | *delta = iv_ca_delta_join (l1: best_delta, l2: *delta); |
6865 | return best_cost; |
6866 | } |
6867 | |
6868 | /* Check if CAND_IDX is a candidate other than OLD_CAND and has |
6869 | cheaper local cost for GROUP than BEST_CP. Return pointer to |
6870 | the corresponding cost_pair, otherwise just return BEST_CP. */ |
6871 | |
6872 | static class cost_pair* |
6873 | cheaper_cost_with_cand (struct ivopts_data *data, struct iv_group *group, |
6874 | unsigned int cand_idx, struct iv_cand *old_cand, |
6875 | class cost_pair *best_cp) |
6876 | { |
6877 | struct iv_cand *cand; |
6878 | class cost_pair *cp; |
6879 | |
6880 | gcc_assert (old_cand != NULL && best_cp != NULL); |
6881 | if (cand_idx == old_cand->id) |
6882 | return best_cp; |
6883 | |
6884 | cand = data->vcands[cand_idx]; |
6885 | cp = get_group_iv_cost (data, group, cand); |
6886 | if (cp != NULL && cheaper_cost_pair (a: cp, b: best_cp)) |
6887 | return cp; |
6888 | |
6889 | return best_cp; |
6890 | } |
6891 | |
6892 | /* Try breaking local optimal fixed-point for IVS by replacing candidates |
6893 | which are used by more than one iv uses. For each of those candidates, |
6894 | this function tries to represent iv uses under that candidate using |
6895 | other ones with lower local cost, then tries to prune the new set. |
6896 | If the new set has lower cost, It returns the new cost after recording |
6897 | candidate replacement in list DELTA. */ |
6898 | |
6899 | static comp_cost |
6900 | iv_ca_replace (struct ivopts_data *data, class iv_ca *ivs, |
6901 | struct iv_ca_delta **delta) |
6902 | { |
6903 | bitmap_iterator bi, bj; |
6904 | unsigned int i, j, k; |
6905 | struct iv_cand *cand; |
6906 | comp_cost orig_cost, acost; |
6907 | struct iv_ca_delta *act_delta, *tmp_delta; |
6908 | class cost_pair *old_cp, *best_cp = NULL; |
6909 | |
6910 | *delta = NULL; |
6911 | orig_cost = iv_ca_cost (ivs); |
6912 | |
6913 | EXECUTE_IF_SET_IN_BITMAP (ivs->cands, 0, i, bi) |
6914 | { |
6915 | if (ivs->n_cand_uses[i] == 1 |
6916 | || ivs->n_cand_uses[i] > ALWAYS_PRUNE_CAND_SET_BOUND) |
6917 | continue; |
6918 | |
6919 | cand = data->vcands[i]; |
6920 | |
6921 | act_delta = NULL; |
6922 | /* Represent uses under current candidate using other ones with |
6923 | lower local cost. */ |
6924 | for (j = 0; j < ivs->upto; j++) |
6925 | { |
6926 | struct iv_group *group = data->vgroups[j]; |
6927 | old_cp = iv_ca_cand_for_group (ivs, group); |
6928 | |
6929 | if (old_cp->cand != cand) |
6930 | continue; |
6931 | |
6932 | best_cp = old_cp; |
6933 | if (data->consider_all_candidates) |
6934 | for (k = 0; k < data->vcands.length (); k++) |
6935 | best_cp = cheaper_cost_with_cand (data, group, cand_idx: k, |
6936 | old_cand: old_cp->cand, best_cp); |
6937 | else |
6938 | EXECUTE_IF_SET_IN_BITMAP (group->related_cands, 0, k, bj) |
6939 | best_cp = cheaper_cost_with_cand (data, group, cand_idx: k, |
6940 | old_cand: old_cp->cand, best_cp); |
6941 | |
6942 | if (best_cp == old_cp) |
6943 | continue; |
6944 | |
6945 | act_delta = iv_ca_delta_add (group, old_cp, new_cp: best_cp, next: act_delta); |
6946 | } |
6947 | /* No need for further prune. */ |
6948 | if (!act_delta) |
6949 | continue; |
6950 | |
6951 | /* Prune the new candidate set. */ |
6952 | iv_ca_delta_commit (data, ivs, delta: act_delta, forward: true); |
6953 | acost = iv_ca_prune (data, ivs, NULL, delta: &tmp_delta); |
6954 | iv_ca_delta_commit (data, ivs, delta: act_delta, forward: false); |
6955 | act_delta = iv_ca_delta_join (l1: act_delta, l2: tmp_delta); |
6956 | |
6957 | if (acost < orig_cost) |
6958 | { |
6959 | *delta = act_delta; |
6960 | return acost; |
6961 | } |
6962 | else |
6963 | iv_ca_delta_free (delta: &act_delta); |
6964 | } |
6965 | |
6966 | return orig_cost; |
6967 | } |
6968 | |
6969 | /* Tries to extend the sets IVS in the best possible way in order to |
6970 | express the GROUP. If ORIGINALP is true, prefer candidates from |
6971 | the original set of IVs, otherwise favor important candidates not |
6972 | based on any memory object. */ |
6973 | |
6974 | static bool |
6975 | try_add_cand_for (struct ivopts_data *data, class iv_ca *ivs, |
6976 | struct iv_group *group, bool originalp) |
6977 | { |
6978 | comp_cost best_cost, act_cost; |
6979 | unsigned i; |
6980 | bitmap_iterator bi; |
6981 | struct iv_cand *cand; |
6982 | struct iv_ca_delta *best_delta = NULL, *act_delta; |
6983 | class cost_pair *cp; |
6984 | |
6985 | iv_ca_add_group (data, ivs, group); |
6986 | best_cost = iv_ca_cost (ivs); |
6987 | cp = iv_ca_cand_for_group (ivs, group); |
6988 | if (cp) |
6989 | { |
6990 | best_delta = iv_ca_delta_add (group, NULL, new_cp: cp, NULL); |
6991 | iv_ca_set_no_cp (data, ivs, group); |
6992 | } |
6993 | |
6994 | /* If ORIGINALP is true, try to find the original IV for the use. Otherwise |
6995 | first try important candidates not based on any memory object. Only if |
6996 | this fails, try the specific ones. Rationale -- in loops with many |
6997 | variables the best choice often is to use just one generic biv. If we |
6998 | added here many ivs specific to the uses, the optimization algorithm later |
6999 | would be likely to get stuck in a local minimum, thus causing us to create |
7000 | too many ivs. The approach from few ivs to more seems more likely to be |
7001 | successful -- starting from few ivs, replacing an expensive use by a |
7002 | specific iv should always be a win. */ |
7003 | EXECUTE_IF_SET_IN_BITMAP (group->related_cands, 0, i, bi) |
7004 | { |
7005 | cand = data->vcands[i]; |
7006 | |
7007 | if (originalp && cand->pos !=IP_ORIGINAL) |
7008 | continue; |
7009 | |
7010 | if (!originalp && cand->iv->base_object != NULL_TREE) |
7011 | continue; |
7012 | |
7013 | if (iv_ca_cand_used_p (ivs, cand)) |
7014 | continue; |
7015 | |
7016 | cp = get_group_iv_cost (data, group, cand); |
7017 | if (!cp) |
7018 | continue; |
7019 | |
7020 | iv_ca_set_cp (data, ivs, group, cp); |
7021 | act_cost = iv_ca_extend (data, ivs, cand, delta: &act_delta, NULL, |
7022 | min_ncand: true); |
7023 | iv_ca_set_no_cp (data, ivs, group); |
7024 | act_delta = iv_ca_delta_add (group, NULL, new_cp: cp, next: act_delta); |
7025 | |
7026 | if (act_cost < best_cost) |
7027 | { |
7028 | best_cost = act_cost; |
7029 | |
7030 | iv_ca_delta_free (delta: &best_delta); |
7031 | best_delta = act_delta; |
7032 | } |
7033 | else |
7034 | iv_ca_delta_free (delta: &act_delta); |
7035 | } |
7036 | |
7037 | if (best_cost.infinite_cost_p ()) |
7038 | { |
7039 | for (i = 0; i < group->n_map_members; i++) |
7040 | { |
7041 | cp = group->cost_map + i; |
7042 | cand = cp->cand; |
7043 | if (!cand) |
7044 | continue; |
7045 | |
7046 | /* Already tried this. */ |
7047 | if (cand->important) |
7048 | { |
7049 | if (originalp && cand->pos == IP_ORIGINAL) |
7050 | continue; |
7051 | if (!originalp && cand->iv->base_object == NULL_TREE) |
7052 | continue; |
7053 | } |
7054 | |
7055 | if (iv_ca_cand_used_p (ivs, cand)) |
7056 | continue; |
7057 | |
7058 | act_delta = NULL; |
7059 | iv_ca_set_cp (data, ivs, group, cp); |
7060 | act_cost = iv_ca_extend (data, ivs, cand, delta: &act_delta, NULL, min_ncand: true); |
7061 | iv_ca_set_no_cp (data, ivs, group); |
7062 | act_delta = iv_ca_delta_add (group, |
7063 | old_cp: iv_ca_cand_for_group (ivs, group), |
7064 | new_cp: cp, next: act_delta); |
7065 | |
7066 | if (act_cost < best_cost) |
7067 | { |
7068 | best_cost = act_cost; |
7069 | |
7070 | if (best_delta) |
7071 | iv_ca_delta_free (delta: &best_delta); |
7072 | best_delta = act_delta; |
7073 | } |
7074 | else |
7075 | iv_ca_delta_free (delta: &act_delta); |
7076 | } |
7077 | } |
7078 | |
7079 | iv_ca_delta_commit (data, ivs, delta: best_delta, forward: true); |
7080 | iv_ca_delta_free (delta: &best_delta); |
7081 | |
7082 | return !best_cost.infinite_cost_p (); |
7083 | } |
7084 | |
7085 | /* Finds an initial assignment of candidates to uses. */ |
7086 | |
7087 | static class iv_ca * |
7088 | get_initial_solution (struct ivopts_data *data, bool originalp) |
7089 | { |
7090 | unsigned i; |
7091 | class iv_ca *ivs = iv_ca_new (data); |
7092 | |
7093 | for (i = 0; i < data->vgroups.length (); i++) |
7094 | if (!try_add_cand_for (data, ivs, group: data->vgroups[i], originalp)) |
7095 | { |
7096 | iv_ca_free (ivs: &ivs); |
7097 | return NULL; |
7098 | } |
7099 | |
7100 | return ivs; |
7101 | } |
7102 | |
7103 | /* Tries to improve set of induction variables IVS. TRY_REPLACE_P |
7104 | points to a bool variable, this function tries to break local |
7105 | optimal fixed-point by replacing candidates in IVS if it's true. */ |
7106 | |
7107 | static bool |
7108 | try_improve_iv_set (struct ivopts_data *data, |
7109 | class iv_ca *ivs, bool *try_replace_p) |
7110 | { |
7111 | unsigned i, n_ivs; |
7112 | comp_cost acost, best_cost = iv_ca_cost (ivs); |
7113 | struct iv_ca_delta *best_delta = NULL, *act_delta, *tmp_delta; |
7114 | struct iv_cand *cand; |
7115 | |
7116 | /* Try extending the set of induction variables by one. */ |
7117 | for (i = 0; i < data->vcands.length (); i++) |
7118 | { |
7119 | cand = data->vcands[i]; |
7120 | |
7121 | if (iv_ca_cand_used_p (ivs, cand)) |
7122 | continue; |
7123 | |
7124 | acost = iv_ca_extend (data, ivs, cand, delta: &act_delta, n_ivs: &n_ivs, min_ncand: false); |
7125 | if (!act_delta) |
7126 | continue; |
7127 | |
7128 | /* If we successfully added the candidate and the set is small enough, |
7129 | try optimizing it by removing other candidates. */ |
7130 | if (n_ivs <= ALWAYS_PRUNE_CAND_SET_BOUND) |
7131 | { |
7132 | iv_ca_delta_commit (data, ivs, delta: act_delta, forward: true); |
7133 | acost = iv_ca_prune (data, ivs, except_cand: cand, delta: &tmp_delta); |
7134 | iv_ca_delta_commit (data, ivs, delta: act_delta, forward: false); |
7135 | act_delta = iv_ca_delta_join (l1: act_delta, l2: tmp_delta); |
7136 | } |
7137 | |
7138 | if (acost < best_cost) |
7139 | { |
7140 | best_cost = acost; |
7141 | iv_ca_delta_free (delta: &best_delta); |
7142 | best_delta = act_delta; |
7143 | } |
7144 | else |
7145 | iv_ca_delta_free (delta: &act_delta); |
7146 | } |
7147 | |
7148 | if (!best_delta) |
7149 | { |
7150 | /* Try removing the candidates from the set instead. */ |
7151 | best_cost = iv_ca_prune (data, ivs, NULL, delta: &best_delta); |
7152 | |
7153 | if (!best_delta && *try_replace_p) |
7154 | { |
7155 | *try_replace_p = false; |
7156 | /* So far candidate selecting algorithm tends to choose fewer IVs |
7157 | so that it can handle cases in which loops have many variables |
7158 | but the best choice is often to use only one general biv. One |
7159 | weakness is it can't handle opposite cases, in which different |
7160 | candidates should be chosen with respect to each use. To solve |
7161 | the problem, we replace candidates in a manner described by the |
7162 | comments of iv_ca_replace, thus give general algorithm a chance |
7163 | to break local optimal fixed-point in these cases. */ |
7164 | best_cost = iv_ca_replace (data, ivs, delta: &best_delta); |
7165 | } |
7166 | |
7167 | if (!best_delta) |
7168 | return false; |
7169 | } |
7170 | |
7171 | iv_ca_delta_commit (data, ivs, delta: best_delta, forward: true); |
7172 | iv_ca_delta_free (delta: &best_delta); |
7173 | return best_cost == iv_ca_cost (ivs); |
7174 | } |
7175 | |
7176 | /* Attempts to find the optimal set of induction variables. We do simple |
7177 | greedy heuristic -- we try to replace at most one candidate in the selected |
7178 | solution and remove the unused ivs while this improves the cost. */ |
7179 | |
7180 | static class iv_ca * |
7181 | find_optimal_iv_set_1 (struct ivopts_data *data, bool originalp) |
7182 | { |
7183 | class iv_ca *set; |
7184 | bool try_replace_p = true; |
7185 | |
7186 | /* Get the initial solution. */ |
7187 | set = get_initial_solution (data, originalp); |
7188 | if (!set) |
7189 | { |
7190 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7191 | fprintf (stream: dump_file, format: "Unable to substitute for ivs, failed.\n" ); |
7192 | return NULL; |
7193 | } |
7194 | |
7195 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7196 | { |
7197 | fprintf (stream: dump_file, format: "Initial set of candidates:\n" ); |
7198 | iv_ca_dump (data, file: dump_file, ivs: set); |
7199 | } |
7200 | |
7201 | while (try_improve_iv_set (data, ivs: set, try_replace_p: &try_replace_p)) |
7202 | { |
7203 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7204 | { |
7205 | fprintf (stream: dump_file, format: "Improved to:\n" ); |
7206 | iv_ca_dump (data, file: dump_file, ivs: set); |
7207 | } |
7208 | } |
7209 | |
7210 | /* If the set has infinite_cost, it can't be optimal. */ |
7211 | if (iv_ca_cost (ivs: set).infinite_cost_p ()) |
7212 | { |
7213 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7214 | fprintf (stream: dump_file, |
7215 | format: "Overflow to infinite cost in try_improve_iv_set.\n" ); |
7216 | iv_ca_free (ivs: &set); |
7217 | } |
7218 | return set; |
7219 | } |
7220 | |
7221 | static class iv_ca * |
7222 | find_optimal_iv_set (struct ivopts_data *data) |
7223 | { |
7224 | unsigned i; |
7225 | comp_cost cost, origcost; |
7226 | class iv_ca *set, *origset; |
7227 | |
7228 | /* Determine the cost based on a strategy that starts with original IVs, |
7229 | and try again using a strategy that prefers candidates not based |
7230 | on any IVs. */ |
7231 | origset = find_optimal_iv_set_1 (data, originalp: true); |
7232 | set = find_optimal_iv_set_1 (data, originalp: false); |
7233 | |
7234 | if (!origset && !set) |
7235 | return NULL; |
7236 | |
7237 | origcost = origset ? iv_ca_cost (ivs: origset) : infinite_cost; |
7238 | cost = set ? iv_ca_cost (ivs: set) : infinite_cost; |
7239 | |
7240 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7241 | { |
7242 | fprintf (stream: dump_file, format: "Original cost %" PRId64 " (complexity %d)\n\n" , |
7243 | origcost.cost, origcost.complexity); |
7244 | fprintf (stream: dump_file, format: "Final cost %" PRId64 " (complexity %d)\n\n" , |
7245 | cost.cost, cost.complexity); |
7246 | } |
7247 | |
7248 | /* Choose the one with the best cost. */ |
7249 | if (origcost <= cost) |
7250 | { |
7251 | if (set) |
7252 | iv_ca_free (ivs: &set); |
7253 | set = origset; |
7254 | } |
7255 | else if (origset) |
7256 | iv_ca_free (ivs: &origset); |
7257 | |
7258 | for (i = 0; i < data->vgroups.length (); i++) |
7259 | { |
7260 | struct iv_group *group = data->vgroups[i]; |
7261 | group->selected = iv_ca_cand_for_group (ivs: set, group)->cand; |
7262 | } |
7263 | |
7264 | return set; |
7265 | } |
7266 | |
7267 | /* Creates a new induction variable corresponding to CAND. */ |
7268 | |
7269 | static void |
7270 | create_new_iv (struct ivopts_data *data, struct iv_cand *cand) |
7271 | { |
7272 | gimple_stmt_iterator incr_pos; |
7273 | tree base; |
7274 | struct iv_use *use; |
7275 | struct iv_group *group; |
7276 | bool after = false; |
7277 | |
7278 | gcc_assert (cand->iv != NULL); |
7279 | |
7280 | switch (cand->pos) |
7281 | { |
7282 | case IP_NORMAL: |
7283 | incr_pos = gsi_last_bb (bb: ip_normal_pos (data->current_loop)); |
7284 | break; |
7285 | |
7286 | case IP_END: |
7287 | incr_pos = gsi_last_bb (bb: ip_end_pos (data->current_loop)); |
7288 | after = true; |
7289 | if (!gsi_end_p (i: incr_pos) && stmt_ends_bb_p (gsi_stmt (i: incr_pos))) |
7290 | { |
7291 | edge e = find_edge (gsi_bb (i: incr_pos), data->current_loop->header); |
7292 | incr_pos = gsi_after_labels (bb: split_edge (e)); |
7293 | after = false; |
7294 | } |
7295 | break; |
7296 | |
7297 | case IP_AFTER_USE: |
7298 | after = true; |
7299 | /* fall through */ |
7300 | case IP_BEFORE_USE: |
7301 | incr_pos = gsi_for_stmt (cand->incremented_at); |
7302 | break; |
7303 | |
7304 | case IP_ORIGINAL: |
7305 | /* Mark that the iv is preserved. */ |
7306 | name_info (data, name: cand->var_before)->preserve_biv = true; |
7307 | name_info (data, name: cand->var_after)->preserve_biv = true; |
7308 | |
7309 | /* Rewrite the increment so that it uses var_before directly. */ |
7310 | use = find_interesting_uses_op (data, op: cand->var_after); |
7311 | group = data->vgroups[use->group_id]; |
7312 | group->selected = cand; |
7313 | return; |
7314 | } |
7315 | |
7316 | gimple_add_tmp_var (cand->var_before); |
7317 | |
7318 | base = unshare_expr (cand->iv->base); |
7319 | |
7320 | create_iv (base, PLUS_EXPR, unshare_expr (cand->iv->step), |
7321 | cand->var_before, data->current_loop, |
7322 | &incr_pos, after, &cand->var_before, &cand->var_after); |
7323 | } |
7324 | |
7325 | /* Creates new induction variables described in SET. */ |
7326 | |
7327 | static void |
7328 | create_new_ivs (struct ivopts_data *data, class iv_ca *set) |
7329 | { |
7330 | unsigned i; |
7331 | struct iv_cand *cand; |
7332 | bitmap_iterator bi; |
7333 | |
7334 | EXECUTE_IF_SET_IN_BITMAP (set->cands, 0, i, bi) |
7335 | { |
7336 | cand = data->vcands[i]; |
7337 | create_new_iv (data, cand); |
7338 | } |
7339 | |
7340 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7341 | { |
7342 | fprintf (stream: dump_file, format: "Selected IV set for loop %d" , |
7343 | data->current_loop->num); |
7344 | if (data->loop_loc != UNKNOWN_LOCATION) |
7345 | fprintf (stream: dump_file, format: " at %s:%d" , LOCATION_FILE (data->loop_loc), |
7346 | LOCATION_LINE (data->loop_loc)); |
7347 | fprintf (stream: dump_file, format: ", " HOST_WIDE_INT_PRINT_DEC " avg niters" , |
7348 | avg_loop_niter (loop: data->current_loop)); |
7349 | fprintf (stream: dump_file, format: ", %lu IVs:\n" , bitmap_count_bits (set->cands)); |
7350 | EXECUTE_IF_SET_IN_BITMAP (set->cands, 0, i, bi) |
7351 | { |
7352 | cand = data->vcands[i]; |
7353 | dump_cand (file: dump_file, cand); |
7354 | } |
7355 | fprintf (stream: dump_file, format: "\n" ); |
7356 | } |
7357 | } |
7358 | |
7359 | /* Rewrites USE (definition of iv used in a nonlinear expression) |
7360 | using candidate CAND. */ |
7361 | |
7362 | static void |
7363 | rewrite_use_nonlinear_expr (struct ivopts_data *data, |
7364 | struct iv_use *use, struct iv_cand *cand) |
7365 | { |
7366 | gassign *ass; |
7367 | gimple_stmt_iterator bsi; |
7368 | tree comp, type = get_use_type (use), tgt; |
7369 | |
7370 | /* An important special case -- if we are asked to express value of |
7371 | the original iv by itself, just exit; there is no need to |
7372 | introduce a new computation (that might also need casting the |
7373 | variable to unsigned and back). */ |
7374 | if (cand->pos == IP_ORIGINAL |
7375 | && cand->incremented_at == use->stmt) |
7376 | { |
7377 | tree op = NULL_TREE; |
7378 | enum tree_code stmt_code; |
7379 | |
7380 | gcc_assert (is_gimple_assign (use->stmt)); |
7381 | gcc_assert (gimple_assign_lhs (use->stmt) == cand->var_after); |
7382 | |
7383 | /* Check whether we may leave the computation unchanged. |
7384 | This is the case only if it does not rely on other |
7385 | computations in the loop -- otherwise, the computation |
7386 | we rely upon may be removed in remove_unused_ivs, |
7387 | thus leading to ICE. */ |
7388 | stmt_code = gimple_assign_rhs_code (gs: use->stmt); |
7389 | if (stmt_code == PLUS_EXPR |
7390 | || stmt_code == MINUS_EXPR |
7391 | || stmt_code == POINTER_PLUS_EXPR) |
7392 | { |
7393 | if (gimple_assign_rhs1 (gs: use->stmt) == cand->var_before) |
7394 | op = gimple_assign_rhs2 (gs: use->stmt); |
7395 | else if (gimple_assign_rhs2 (gs: use->stmt) == cand->var_before) |
7396 | op = gimple_assign_rhs1 (gs: use->stmt); |
7397 | } |
7398 | |
7399 | if (op != NULL_TREE) |
7400 | { |
7401 | if (expr_invariant_in_loop_p (loop: data->current_loop, expr: op)) |
7402 | return; |
7403 | if (TREE_CODE (op) == SSA_NAME) |
7404 | { |
7405 | struct iv *iv = get_iv (data, var: op); |
7406 | if (iv != NULL && integer_zerop (iv->step)) |
7407 | return; |
7408 | } |
7409 | } |
7410 | } |
7411 | |
7412 | switch (gimple_code (g: use->stmt)) |
7413 | { |
7414 | case GIMPLE_PHI: |
7415 | tgt = PHI_RESULT (use->stmt); |
7416 | |
7417 | /* If we should keep the biv, do not replace it. */ |
7418 | if (name_info (data, name: tgt)->preserve_biv) |
7419 | return; |
7420 | |
7421 | bsi = gsi_after_labels (bb: gimple_bb (g: use->stmt)); |
7422 | break; |
7423 | |
7424 | case GIMPLE_ASSIGN: |
7425 | tgt = gimple_assign_lhs (gs: use->stmt); |
7426 | bsi = gsi_for_stmt (use->stmt); |
7427 | break; |
7428 | |
7429 | default: |
7430 | gcc_unreachable (); |
7431 | } |
7432 | |
7433 | aff_tree aff_inv, aff_var; |
7434 | if (!get_computation_aff_1 (loop: data->current_loop, at: use->stmt, |
7435 | use, cand, aff_inv: &aff_inv, aff_var: &aff_var)) |
7436 | gcc_unreachable (); |
7437 | |
7438 | unshare_aff_combination (&aff_inv); |
7439 | unshare_aff_combination (&aff_var); |
7440 | /* Prefer CSE opportunity than loop invariant by adding offset at last |
7441 | so that iv_uses have different offsets can be CSEed. */ |
7442 | poly_widest_int offset = aff_inv.offset; |
7443 | aff_inv.offset = 0; |
7444 | |
7445 | gimple_seq stmt_list = NULL, seq = NULL; |
7446 | tree comp_op1 = aff_combination_to_tree (&aff_inv); |
7447 | tree comp_op2 = aff_combination_to_tree (&aff_var); |
7448 | gcc_assert (comp_op1 && comp_op2); |
7449 | |
7450 | comp_op1 = force_gimple_operand (comp_op1, &seq, true, NULL); |
7451 | gimple_seq_add_seq (&stmt_list, seq); |
7452 | comp_op2 = force_gimple_operand (comp_op2, &seq, true, NULL); |
7453 | gimple_seq_add_seq (&stmt_list, seq); |
7454 | |
7455 | if (POINTER_TYPE_P (TREE_TYPE (comp_op2))) |
7456 | std::swap (a&: comp_op1, b&: comp_op2); |
7457 | |
7458 | if (POINTER_TYPE_P (TREE_TYPE (comp_op1))) |
7459 | { |
7460 | comp = fold_build_pointer_plus (comp_op1, |
7461 | fold_convert (sizetype, comp_op2)); |
7462 | comp = fold_build_pointer_plus (comp, |
7463 | wide_int_to_tree (sizetype, offset)); |
7464 | } |
7465 | else |
7466 | { |
7467 | comp = fold_build2 (PLUS_EXPR, TREE_TYPE (comp_op1), comp_op1, |
7468 | fold_convert (TREE_TYPE (comp_op1), comp_op2)); |
7469 | comp = fold_build2 (PLUS_EXPR, TREE_TYPE (comp_op1), comp, |
7470 | wide_int_to_tree (TREE_TYPE (comp_op1), offset)); |
7471 | } |
7472 | |
7473 | comp = fold_convert (type, comp); |
7474 | comp = force_gimple_operand (comp, &seq, false, NULL); |
7475 | gimple_seq_add_seq (&stmt_list, seq); |
7476 | if (gimple_code (g: use->stmt) != GIMPLE_PHI |
7477 | /* We can't allow re-allocating the stmt as it might be pointed |
7478 | to still. */ |
7479 | && (get_gimple_rhs_num_ops (TREE_CODE (comp)) |
7480 | >= gimple_num_ops (gs: gsi_stmt (i: bsi)))) |
7481 | { |
7482 | comp = force_gimple_operand (comp, &seq, true, NULL); |
7483 | gimple_seq_add_seq (&stmt_list, seq); |
7484 | if (POINTER_TYPE_P (TREE_TYPE (tgt))) |
7485 | { |
7486 | duplicate_ssa_name_ptr_info (comp, SSA_NAME_PTR_INFO (tgt)); |
7487 | /* As this isn't a plain copy we have to reset alignment |
7488 | information. */ |
7489 | if (SSA_NAME_PTR_INFO (comp)) |
7490 | mark_ptr_info_alignment_unknown (SSA_NAME_PTR_INFO (comp)); |
7491 | } |
7492 | } |
7493 | |
7494 | gsi_insert_seq_before (&bsi, stmt_list, GSI_SAME_STMT); |
7495 | if (gimple_code (g: use->stmt) == GIMPLE_PHI) |
7496 | { |
7497 | ass = gimple_build_assign (tgt, comp); |
7498 | gsi_insert_before (&bsi, ass, GSI_SAME_STMT); |
7499 | |
7500 | bsi = gsi_for_stmt (use->stmt); |
7501 | remove_phi_node (&bsi, false); |
7502 | } |
7503 | else |
7504 | { |
7505 | gimple_assign_set_rhs_from_tree (&bsi, comp); |
7506 | use->stmt = gsi_stmt (i: bsi); |
7507 | } |
7508 | } |
7509 | |
7510 | /* Performs a peephole optimization to reorder the iv update statement with |
7511 | a mem ref to enable instruction combining in later phases. The mem ref uses |
7512 | the iv value before the update, so the reordering transformation requires |
7513 | adjustment of the offset. CAND is the selected IV_CAND. |
7514 | |
7515 | Example: |
7516 | |
7517 | t = MEM_REF (base, iv1, 8, 16); // base, index, stride, offset |
7518 | iv2 = iv1 + 1; |
7519 | |
7520 | if (t < val) (1) |
7521 | goto L; |
7522 | goto Head; |
7523 | |
7524 | |
7525 | directly propagating t over to (1) will introduce overlapping live range |
7526 | thus increase register pressure. This peephole transform it into: |
7527 | |
7528 | |
7529 | iv2 = iv1 + 1; |
7530 | t = MEM_REF (base, iv2, 8, 8); |
7531 | if (t < val) |
7532 | goto L; |
7533 | goto Head; |
7534 | */ |
7535 | |
7536 | static void |
7537 | adjust_iv_update_pos (struct iv_cand *cand, struct iv_use *use) |
7538 | { |
7539 | tree var_after; |
7540 | gimple *iv_update, *stmt; |
7541 | basic_block bb; |
7542 | gimple_stmt_iterator gsi, gsi_iv; |
7543 | |
7544 | if (cand->pos != IP_NORMAL) |
7545 | return; |
7546 | |
7547 | var_after = cand->var_after; |
7548 | iv_update = SSA_NAME_DEF_STMT (var_after); |
7549 | |
7550 | bb = gimple_bb (g: iv_update); |
7551 | gsi = gsi_last_nondebug_bb (bb); |
7552 | stmt = gsi_stmt (i: gsi); |
7553 | |
7554 | /* Only handle conditional statement for now. */ |
7555 | if (gimple_code (g: stmt) != GIMPLE_COND) |
7556 | return; |
7557 | |
7558 | gsi_prev_nondebug (i: &gsi); |
7559 | stmt = gsi_stmt (i: gsi); |
7560 | if (stmt != iv_update) |
7561 | return; |
7562 | |
7563 | gsi_prev_nondebug (i: &gsi); |
7564 | if (gsi_end_p (i: gsi)) |
7565 | return; |
7566 | |
7567 | stmt = gsi_stmt (i: gsi); |
7568 | if (gimple_code (g: stmt) != GIMPLE_ASSIGN) |
7569 | return; |
7570 | |
7571 | if (stmt != use->stmt) |
7572 | return; |
7573 | |
7574 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) |
7575 | return; |
7576 | |
7577 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7578 | { |
7579 | fprintf (stream: dump_file, format: "Reordering \n" ); |
7580 | print_gimple_stmt (dump_file, iv_update, 0); |
7581 | print_gimple_stmt (dump_file, use->stmt, 0); |
7582 | fprintf (stream: dump_file, format: "\n" ); |
7583 | } |
7584 | |
7585 | gsi = gsi_for_stmt (use->stmt); |
7586 | gsi_iv = gsi_for_stmt (iv_update); |
7587 | gsi_move_before (&gsi_iv, &gsi); |
7588 | |
7589 | cand->pos = IP_BEFORE_USE; |
7590 | cand->incremented_at = use->stmt; |
7591 | } |
7592 | |
7593 | /* Return the alias pointer type that should be used for a MEM_REF |
7594 | associated with USE, which has type USE_PTR_ADDRESS. */ |
7595 | |
7596 | static tree |
7597 | get_alias_ptr_type_for_ptr_address (iv_use *use) |
7598 | { |
7599 | gcall *call = as_a <gcall *> (p: use->stmt); |
7600 | switch (gimple_call_internal_fn (gs: call)) |
7601 | { |
7602 | case IFN_MASK_LOAD: |
7603 | case IFN_MASK_STORE: |
7604 | case IFN_MASK_LOAD_LANES: |
7605 | case IFN_MASK_STORE_LANES: |
7606 | case IFN_MASK_LEN_LOAD_LANES: |
7607 | case IFN_MASK_LEN_STORE_LANES: |
7608 | case IFN_LEN_LOAD: |
7609 | case IFN_LEN_STORE: |
7610 | case IFN_MASK_LEN_LOAD: |
7611 | case IFN_MASK_LEN_STORE: |
7612 | /* The second argument contains the correct alias type. */ |
7613 | gcc_assert (use->op_p = gimple_call_arg_ptr (call, 0)); |
7614 | return TREE_TYPE (gimple_call_arg (call, 1)); |
7615 | |
7616 | default: |
7617 | gcc_unreachable (); |
7618 | } |
7619 | } |
7620 | |
7621 | |
7622 | /* Rewrites USE (address that is an iv) using candidate CAND. */ |
7623 | |
7624 | static void |
7625 | rewrite_use_address (struct ivopts_data *data, |
7626 | struct iv_use *use, struct iv_cand *cand) |
7627 | { |
7628 | aff_tree aff; |
7629 | bool ok; |
7630 | |
7631 | adjust_iv_update_pos (cand, use); |
7632 | ok = get_computation_aff (loop: data->current_loop, at: use->stmt, use, cand, aff: &aff); |
7633 | gcc_assert (ok); |
7634 | unshare_aff_combination (&aff); |
7635 | |
7636 | /* To avoid undefined overflow problems, all IV candidates use unsigned |
7637 | integer types. The drawback is that this makes it impossible for |
7638 | create_mem_ref to distinguish an IV that is based on a memory object |
7639 | from one that represents simply an offset. |
7640 | |
7641 | To work around this problem, we pass a hint to create_mem_ref that |
7642 | indicates which variable (if any) in aff is an IV based on a memory |
7643 | object. Note that we only consider the candidate. If this is not |
7644 | based on an object, the base of the reference is in some subexpression |
7645 | of the use -- but these will use pointer types, so they are recognized |
7646 | by the create_mem_ref heuristics anyway. */ |
7647 | tree iv = var_at_stmt (loop: data->current_loop, cand, stmt: use->stmt); |
7648 | tree base_hint = (cand->iv->base_object) ? iv : NULL_TREE; |
7649 | gimple_stmt_iterator bsi = gsi_for_stmt (use->stmt); |
7650 | tree type = use->mem_type; |
7651 | tree alias_ptr_type; |
7652 | if (use->type == USE_PTR_ADDRESS) |
7653 | alias_ptr_type = get_alias_ptr_type_for_ptr_address (use); |
7654 | else |
7655 | { |
7656 | gcc_assert (type == TREE_TYPE (*use->op_p)); |
7657 | unsigned int align = get_object_alignment (*use->op_p); |
7658 | if (align != TYPE_ALIGN (type)) |
7659 | type = build_aligned_type (type, align); |
7660 | alias_ptr_type = reference_alias_ptr_type (*use->op_p); |
7661 | } |
7662 | tree ref = create_mem_ref (&bsi, type, &aff, alias_ptr_type, |
7663 | iv, base_hint, data->speed); |
7664 | |
7665 | if (use->type == USE_PTR_ADDRESS) |
7666 | { |
7667 | ref = fold_build1 (ADDR_EXPR, build_pointer_type (use->mem_type), ref); |
7668 | ref = fold_convert (get_use_type (use), ref); |
7669 | ref = force_gimple_operand_gsi (&bsi, ref, true, NULL_TREE, |
7670 | true, GSI_SAME_STMT); |
7671 | } |
7672 | else |
7673 | { |
7674 | /* When we end up confused enough and have no suitable base but |
7675 | stuffed everything to index2 use a LEA for the address and |
7676 | create a plain MEM_REF to avoid basing a memory reference |
7677 | on address zero which create_mem_ref_raw does as fallback. */ |
7678 | if (TREE_CODE (ref) == TARGET_MEM_REF |
7679 | && TMR_INDEX2 (ref) != NULL_TREE |
7680 | && integer_zerop (TREE_OPERAND (ref, 0))) |
7681 | { |
7682 | ref = fold_build1 (ADDR_EXPR, TREE_TYPE (TREE_OPERAND (ref, 0)), ref); |
7683 | ref = force_gimple_operand_gsi (&bsi, ref, true, NULL_TREE, |
7684 | true, GSI_SAME_STMT); |
7685 | ref = build2 (MEM_REF, type, ref, build_zero_cst (alias_ptr_type)); |
7686 | } |
7687 | copy_ref_info (ref, *use->op_p); |
7688 | } |
7689 | |
7690 | *use->op_p = ref; |
7691 | } |
7692 | |
7693 | /* Rewrites USE (the condition such that one of the arguments is an iv) using |
7694 | candidate CAND. */ |
7695 | |
7696 | static void |
7697 | rewrite_use_compare (struct ivopts_data *data, |
7698 | struct iv_use *use, struct iv_cand *cand) |
7699 | { |
7700 | tree comp, op, bound; |
7701 | gimple_stmt_iterator bsi = gsi_for_stmt (use->stmt); |
7702 | enum tree_code compare; |
7703 | struct iv_group *group = data->vgroups[use->group_id]; |
7704 | class cost_pair *cp = get_group_iv_cost (data, group, cand); |
7705 | |
7706 | bound = cp->value; |
7707 | if (bound) |
7708 | { |
7709 | tree var = var_at_stmt (loop: data->current_loop, cand, stmt: use->stmt); |
7710 | tree var_type = TREE_TYPE (var); |
7711 | gimple_seq stmts; |
7712 | |
7713 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7714 | { |
7715 | fprintf (stream: dump_file, format: "Replacing exit test: " ); |
7716 | print_gimple_stmt (dump_file, use->stmt, 0, TDF_SLIM); |
7717 | } |
7718 | compare = cp->comp; |
7719 | bound = unshare_expr (fold_convert (var_type, bound)); |
7720 | op = force_gimple_operand (bound, &stmts, true, NULL_TREE); |
7721 | if (stmts) |
7722 | gsi_insert_seq_on_edge_immediate ( |
7723 | loop_preheader_edge (data->current_loop), |
7724 | stmts); |
7725 | |
7726 | gcond *cond_stmt = as_a <gcond *> (p: use->stmt); |
7727 | gimple_cond_set_lhs (gs: cond_stmt, lhs: var); |
7728 | gimple_cond_set_code (gs: cond_stmt, code: compare); |
7729 | gimple_cond_set_rhs (gs: cond_stmt, rhs: op); |
7730 | return; |
7731 | } |
7732 | |
7733 | /* The induction variable elimination failed; just express the original |
7734 | giv. */ |
7735 | comp = get_computation_at (loop: data->current_loop, at: use->stmt, use, cand); |
7736 | gcc_assert (comp != NULL_TREE); |
7737 | gcc_assert (use->op_p != NULL); |
7738 | *use->op_p = force_gimple_operand_gsi (&bsi, comp, true, |
7739 | SSA_NAME_VAR (*use->op_p), |
7740 | true, GSI_SAME_STMT); |
7741 | } |
7742 | |
7743 | /* Rewrite the groups using the selected induction variables. */ |
7744 | |
7745 | static void |
7746 | rewrite_groups (struct ivopts_data *data) |
7747 | { |
7748 | unsigned i, j; |
7749 | |
7750 | for (i = 0; i < data->vgroups.length (); i++) |
7751 | { |
7752 | struct iv_group *group = data->vgroups[i]; |
7753 | struct iv_cand *cand = group->selected; |
7754 | |
7755 | gcc_assert (cand); |
7756 | |
7757 | if (group->type == USE_NONLINEAR_EXPR) |
7758 | { |
7759 | for (j = 0; j < group->vuses.length (); j++) |
7760 | { |
7761 | rewrite_use_nonlinear_expr (data, use: group->vuses[j], cand); |
7762 | update_stmt (s: group->vuses[j]->stmt); |
7763 | } |
7764 | } |
7765 | else if (address_p (type: group->type)) |
7766 | { |
7767 | for (j = 0; j < group->vuses.length (); j++) |
7768 | { |
7769 | rewrite_use_address (data, use: group->vuses[j], cand); |
7770 | update_stmt (s: group->vuses[j]->stmt); |
7771 | } |
7772 | } |
7773 | else |
7774 | { |
7775 | gcc_assert (group->type == USE_COMPARE); |
7776 | |
7777 | for (j = 0; j < group->vuses.length (); j++) |
7778 | { |
7779 | rewrite_use_compare (data, use: group->vuses[j], cand); |
7780 | update_stmt (s: group->vuses[j]->stmt); |
7781 | } |
7782 | } |
7783 | } |
7784 | } |
7785 | |
7786 | /* Removes the ivs that are not used after rewriting. */ |
7787 | |
7788 | static void |
7789 | remove_unused_ivs (struct ivopts_data *data, bitmap toremove) |
7790 | { |
7791 | unsigned j; |
7792 | bitmap_iterator bi; |
7793 | |
7794 | /* Figure out an order in which to release SSA DEFs so that we don't |
7795 | release something that we'd have to propagate into a debug stmt |
7796 | afterwards. */ |
7797 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, j, bi) |
7798 | { |
7799 | struct version_info *info; |
7800 | |
7801 | info = ver_info (data, ver: j); |
7802 | if (info->iv |
7803 | && !integer_zerop (info->iv->step) |
7804 | && !info->inv_id |
7805 | && !info->iv->nonlin_use |
7806 | && !info->preserve_biv) |
7807 | { |
7808 | bitmap_set_bit (toremove, SSA_NAME_VERSION (info->iv->ssa_name)); |
7809 | |
7810 | tree def = info->iv->ssa_name; |
7811 | |
7812 | if (MAY_HAVE_DEBUG_BIND_STMTS && SSA_NAME_DEF_STMT (def)) |
7813 | { |
7814 | imm_use_iterator imm_iter; |
7815 | use_operand_p use_p; |
7816 | gimple *stmt; |
7817 | int count = 0; |
7818 | |
7819 | FOR_EACH_IMM_USE_STMT (stmt, imm_iter, def) |
7820 | { |
7821 | if (!gimple_debug_bind_p (s: stmt)) |
7822 | continue; |
7823 | |
7824 | /* We just want to determine whether to do nothing |
7825 | (count == 0), to substitute the computed |
7826 | expression into a single use of the SSA DEF by |
7827 | itself (count == 1), or to use a debug temp |
7828 | because the SSA DEF is used multiple times or as |
7829 | part of a larger expression (count > 1). */ |
7830 | count++; |
7831 | if (gimple_debug_bind_get_value (dbg: stmt) != def) |
7832 | count++; |
7833 | |
7834 | if (count > 1) |
7835 | break; |
7836 | } |
7837 | |
7838 | if (!count) |
7839 | continue; |
7840 | |
7841 | struct iv_use dummy_use; |
7842 | struct iv_cand *best_cand = NULL, *cand; |
7843 | unsigned i, best_pref = 0, cand_pref; |
7844 | tree comp = NULL_TREE; |
7845 | |
7846 | memset (s: &dummy_use, c: 0, n: sizeof (dummy_use)); |
7847 | dummy_use.iv = info->iv; |
7848 | for (i = 0; i < data->vgroups.length () && i < 64; i++) |
7849 | { |
7850 | cand = data->vgroups[i]->selected; |
7851 | if (cand == best_cand) |
7852 | continue; |
7853 | cand_pref = operand_equal_p (cand->iv->step, |
7854 | info->iv->step, flags: 0) |
7855 | ? 4 : 0; |
7856 | cand_pref |
7857 | += TYPE_MODE (TREE_TYPE (cand->iv->base)) |
7858 | == TYPE_MODE (TREE_TYPE (info->iv->base)) |
7859 | ? 2 : 0; |
7860 | cand_pref |
7861 | += TREE_CODE (cand->iv->base) == INTEGER_CST |
7862 | ? 1 : 0; |
7863 | if (best_cand == NULL || best_pref < cand_pref) |
7864 | { |
7865 | tree this_comp |
7866 | = get_debug_computation_at (loop: data->current_loop, |
7867 | SSA_NAME_DEF_STMT (def), |
7868 | use: &dummy_use, cand); |
7869 | if (this_comp) |
7870 | { |
7871 | best_cand = cand; |
7872 | best_pref = cand_pref; |
7873 | comp = this_comp; |
7874 | } |
7875 | } |
7876 | } |
7877 | |
7878 | if (!best_cand) |
7879 | continue; |
7880 | |
7881 | comp = unshare_expr (comp); |
7882 | if (count > 1) |
7883 | { |
7884 | tree vexpr = build_debug_expr_decl (TREE_TYPE (comp)); |
7885 | /* FIXME: Is setting the mode really necessary? */ |
7886 | if (SSA_NAME_VAR (def)) |
7887 | SET_DECL_MODE (vexpr, DECL_MODE (SSA_NAME_VAR (def))); |
7888 | else |
7889 | SET_DECL_MODE (vexpr, TYPE_MODE (TREE_TYPE (vexpr))); |
7890 | gdebug *def_temp |
7891 | = gimple_build_debug_bind (vexpr, comp, NULL); |
7892 | gimple_stmt_iterator gsi; |
7893 | |
7894 | if (gimple_code (SSA_NAME_DEF_STMT (def)) == GIMPLE_PHI) |
7895 | gsi = gsi_after_labels (bb: gimple_bb |
7896 | (SSA_NAME_DEF_STMT (def))); |
7897 | else |
7898 | gsi = gsi_for_stmt (SSA_NAME_DEF_STMT (def)); |
7899 | |
7900 | gsi_insert_before (&gsi, def_temp, GSI_SAME_STMT); |
7901 | comp = vexpr; |
7902 | } |
7903 | |
7904 | FOR_EACH_IMM_USE_STMT (stmt, imm_iter, def) |
7905 | { |
7906 | if (!gimple_debug_bind_p (s: stmt)) |
7907 | continue; |
7908 | |
7909 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) |
7910 | SET_USE (use_p, comp); |
7911 | |
7912 | update_stmt (s: stmt); |
7913 | } |
7914 | } |
7915 | } |
7916 | } |
7917 | } |
7918 | |
7919 | /* Frees memory occupied by class tree_niter_desc in *VALUE. Callback |
7920 | for hash_map::traverse. */ |
7921 | |
7922 | bool |
7923 | free_tree_niter_desc (edge const &, tree_niter_desc *const &value, void *) |
7924 | { |
7925 | if (value) |
7926 | { |
7927 | value->~tree_niter_desc (); |
7928 | free (ptr: value); |
7929 | } |
7930 | return true; |
7931 | } |
7932 | |
7933 | /* Frees data allocated by the optimization of a single loop. */ |
7934 | |
7935 | static void |
7936 | free_loop_data (struct ivopts_data *data) |
7937 | { |
7938 | unsigned i, j; |
7939 | bitmap_iterator bi; |
7940 | tree obj; |
7941 | |
7942 | if (data->niters) |
7943 | { |
7944 | data->niters->traverse<void *, free_tree_niter_desc> (NULL); |
7945 | delete data->niters; |
7946 | data->niters = NULL; |
7947 | } |
7948 | |
7949 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, i, bi) |
7950 | { |
7951 | struct version_info *info; |
7952 | |
7953 | info = ver_info (data, ver: i); |
7954 | info->iv = NULL; |
7955 | info->has_nonlin_use = false; |
7956 | info->preserve_biv = false; |
7957 | info->inv_id = 0; |
7958 | } |
7959 | bitmap_clear (data->relevant); |
7960 | bitmap_clear (data->important_candidates); |
7961 | |
7962 | for (i = 0; i < data->vgroups.length (); i++) |
7963 | { |
7964 | struct iv_group *group = data->vgroups[i]; |
7965 | |
7966 | for (j = 0; j < group->vuses.length (); j++) |
7967 | free (ptr: group->vuses[j]); |
7968 | group->vuses.release (); |
7969 | |
7970 | BITMAP_FREE (group->related_cands); |
7971 | for (j = 0; j < group->n_map_members; j++) |
7972 | { |
7973 | if (group->cost_map[j].inv_vars) |
7974 | BITMAP_FREE (group->cost_map[j].inv_vars); |
7975 | if (group->cost_map[j].inv_exprs) |
7976 | BITMAP_FREE (group->cost_map[j].inv_exprs); |
7977 | } |
7978 | |
7979 | free (ptr: group->cost_map); |
7980 | free (ptr: group); |
7981 | } |
7982 | data->vgroups.truncate (size: 0); |
7983 | |
7984 | for (i = 0; i < data->vcands.length (); i++) |
7985 | { |
7986 | struct iv_cand *cand = data->vcands[i]; |
7987 | |
7988 | if (cand->inv_vars) |
7989 | BITMAP_FREE (cand->inv_vars); |
7990 | if (cand->inv_exprs) |
7991 | BITMAP_FREE (cand->inv_exprs); |
7992 | free (ptr: cand); |
7993 | } |
7994 | data->vcands.truncate (size: 0); |
7995 | |
7996 | if (data->version_info_size < num_ssa_names) |
7997 | { |
7998 | data->version_info_size = 2 * num_ssa_names; |
7999 | free (ptr: data->version_info); |
8000 | data->version_info = XCNEWVEC (struct version_info, data->version_info_size); |
8001 | } |
8002 | |
8003 | data->max_inv_var_id = 0; |
8004 | data->max_inv_expr_id = 0; |
8005 | |
8006 | FOR_EACH_VEC_ELT (decl_rtl_to_reset, i, obj) |
8007 | SET_DECL_RTL (obj, NULL_RTX); |
8008 | |
8009 | decl_rtl_to_reset.truncate (size: 0); |
8010 | |
8011 | data->inv_expr_tab->empty (); |
8012 | |
8013 | data->iv_common_cand_tab->empty (); |
8014 | data->iv_common_cands.truncate (size: 0); |
8015 | } |
8016 | |
8017 | /* Finalizes data structures used by the iv optimization pass. LOOPS is the |
8018 | loop tree. */ |
8019 | |
8020 | static void |
8021 | tree_ssa_iv_optimize_finalize (struct ivopts_data *data) |
8022 | { |
8023 | free_loop_data (data); |
8024 | free (ptr: data->version_info); |
8025 | BITMAP_FREE (data->relevant); |
8026 | BITMAP_FREE (data->important_candidates); |
8027 | |
8028 | decl_rtl_to_reset.release (); |
8029 | data->vgroups.release (); |
8030 | data->vcands.release (); |
8031 | delete data->inv_expr_tab; |
8032 | data->inv_expr_tab = NULL; |
8033 | free_affine_expand_cache (&data->name_expansion_cache); |
8034 | if (data->base_object_map) |
8035 | delete data->base_object_map; |
8036 | delete data->iv_common_cand_tab; |
8037 | data->iv_common_cand_tab = NULL; |
8038 | data->iv_common_cands.release (); |
8039 | obstack_free (&data->iv_obstack, NULL); |
8040 | } |
8041 | |
8042 | /* Returns true if the loop body BODY includes any function calls. */ |
8043 | |
8044 | static bool |
8045 | loop_body_includes_call (basic_block *body, unsigned num_nodes) |
8046 | { |
8047 | gimple_stmt_iterator gsi; |
8048 | unsigned i; |
8049 | |
8050 | for (i = 0; i < num_nodes; i++) |
8051 | for (gsi = gsi_start_bb (bb: body[i]); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
8052 | { |
8053 | gimple *stmt = gsi_stmt (i: gsi); |
8054 | if (is_gimple_call (gs: stmt) |
8055 | && !gimple_call_internal_p (gs: stmt) |
8056 | && !is_inexpensive_builtin (gimple_call_fndecl (gs: stmt))) |
8057 | return true; |
8058 | } |
8059 | return false; |
8060 | } |
8061 | |
8062 | /* Determine cost scaling factor for basic blocks in loop. */ |
8063 | #define COST_SCALING_FACTOR_BOUND (20) |
8064 | |
8065 | static void |
8066 | determine_scaling_factor (struct ivopts_data *data, basic_block *body) |
8067 | { |
8068 | int lfreq = data->current_loop->header->count.to_frequency (cfun); |
8069 | if (!data->speed || lfreq <= 0) |
8070 | return; |
8071 | |
8072 | int max_freq = lfreq; |
8073 | for (unsigned i = 0; i < data->current_loop->num_nodes; i++) |
8074 | { |
8075 | body[i]->aux = (void *)(intptr_t) 1; |
8076 | if (max_freq < body[i]->count.to_frequency (cfun)) |
8077 | max_freq = body[i]->count.to_frequency (cfun); |
8078 | } |
8079 | if (max_freq > lfreq) |
8080 | { |
8081 | int divisor, factor; |
8082 | /* Check if scaling factor itself needs to be scaled by the bound. This |
8083 | is to avoid overflow when scaling cost according to profile info. */ |
8084 | if (max_freq / lfreq > COST_SCALING_FACTOR_BOUND) |
8085 | { |
8086 | divisor = max_freq; |
8087 | factor = COST_SCALING_FACTOR_BOUND; |
8088 | } |
8089 | else |
8090 | { |
8091 | divisor = lfreq; |
8092 | factor = 1; |
8093 | } |
8094 | for (unsigned i = 0; i < data->current_loop->num_nodes; i++) |
8095 | { |
8096 | int bfreq = body[i]->count.to_frequency (cfun); |
8097 | if (bfreq <= lfreq) |
8098 | continue; |
8099 | |
8100 | body[i]->aux = (void*)(intptr_t) (factor * bfreq / divisor); |
8101 | } |
8102 | } |
8103 | } |
8104 | |
8105 | /* Find doloop comparison use and set its doloop_p on if found. */ |
8106 | |
8107 | static bool |
8108 | find_doloop_use (struct ivopts_data *data) |
8109 | { |
8110 | struct loop *loop = data->current_loop; |
8111 | |
8112 | for (unsigned i = 0; i < data->vgroups.length (); i++) |
8113 | { |
8114 | struct iv_group *group = data->vgroups[i]; |
8115 | if (group->type == USE_COMPARE) |
8116 | { |
8117 | gcc_assert (group->vuses.length () == 1); |
8118 | struct iv_use *use = group->vuses[0]; |
8119 | gimple *stmt = use->stmt; |
8120 | if (gimple_code (g: stmt) == GIMPLE_COND) |
8121 | { |
8122 | basic_block bb = gimple_bb (g: stmt); |
8123 | edge true_edge, false_edge; |
8124 | extract_true_false_edges_from_block (bb, &true_edge, &false_edge); |
8125 | /* This comparison is used for loop latch. Require latch is empty |
8126 | for now. */ |
8127 | if ((loop->latch == true_edge->dest |
8128 | || loop->latch == false_edge->dest) |
8129 | && empty_block_p (loop->latch)) |
8130 | { |
8131 | group->doloop_p = true; |
8132 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8133 | { |
8134 | fprintf (stream: dump_file, format: "Doloop cmp iv use: " ); |
8135 | print_gimple_stmt (dump_file, stmt, TDF_DETAILS); |
8136 | } |
8137 | return true; |
8138 | } |
8139 | } |
8140 | } |
8141 | } |
8142 | |
8143 | return false; |
8144 | } |
8145 | |
8146 | /* For the targets which support doloop, to predict whether later RTL doloop |
8147 | transformation will perform on this loop, further detect the doloop use and |
8148 | mark the flag doloop_use_p if predicted. */ |
8149 | |
8150 | void |
8151 | analyze_and_mark_doloop_use (struct ivopts_data *data) |
8152 | { |
8153 | data->doloop_use_p = false; |
8154 | |
8155 | if (!flag_branch_on_count_reg) |
8156 | return; |
8157 | |
8158 | if (data->current_loop->unroll == USHRT_MAX) |
8159 | return; |
8160 | |
8161 | if (!generic_predict_doloop_p (data)) |
8162 | return; |
8163 | |
8164 | if (find_doloop_use (data)) |
8165 | { |
8166 | data->doloop_use_p = true; |
8167 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8168 | { |
8169 | struct loop *loop = data->current_loop; |
8170 | fprintf (stream: dump_file, |
8171 | format: "Predict loop %d can perform" |
8172 | " doloop optimization later.\n" , |
8173 | loop->num); |
8174 | flow_loop_dump (loop, dump_file, NULL, 1); |
8175 | } |
8176 | } |
8177 | } |
8178 | |
8179 | /* Optimizes the LOOP. Returns true if anything changed. */ |
8180 | |
8181 | static bool |
8182 | tree_ssa_iv_optimize_loop (struct ivopts_data *data, class loop *loop, |
8183 | bitmap toremove) |
8184 | { |
8185 | bool changed = false; |
8186 | class iv_ca *iv_ca; |
8187 | edge exit = single_dom_exit (loop); |
8188 | basic_block *body; |
8189 | |
8190 | gcc_assert (!data->niters); |
8191 | data->current_loop = loop; |
8192 | data->loop_loc = find_loop_location (loop).get_location_t (); |
8193 | data->speed = optimize_loop_for_speed_p (loop); |
8194 | |
8195 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8196 | { |
8197 | fprintf (stream: dump_file, format: "Processing loop %d" , loop->num); |
8198 | if (data->loop_loc != UNKNOWN_LOCATION) |
8199 | fprintf (stream: dump_file, format: " at %s:%d" , LOCATION_FILE (data->loop_loc), |
8200 | LOCATION_LINE (data->loop_loc)); |
8201 | fprintf (stream: dump_file, format: "\n" ); |
8202 | |
8203 | if (exit) |
8204 | { |
8205 | fprintf (stream: dump_file, format: " single exit %d -> %d, exit condition " , |
8206 | exit->src->index, exit->dest->index); |
8207 | print_gimple_stmt (dump_file, *gsi_last_bb (bb: exit->src), |
8208 | 0, TDF_SLIM); |
8209 | fprintf (stream: dump_file, format: "\n" ); |
8210 | } |
8211 | |
8212 | fprintf (stream: dump_file, format: "\n" ); |
8213 | } |
8214 | |
8215 | body = get_loop_body (loop); |
8216 | data->body_includes_call = loop_body_includes_call (body, num_nodes: loop->num_nodes); |
8217 | renumber_gimple_stmt_uids_in_blocks (body, loop->num_nodes); |
8218 | |
8219 | data->loop_single_exit_p |
8220 | = exit != NULL && loop_only_exit_p (loop, body, exit); |
8221 | |
8222 | /* For each ssa name determines whether it behaves as an induction variable |
8223 | in some loop. */ |
8224 | if (!find_induction_variables (data, body)) |
8225 | goto finish; |
8226 | |
8227 | /* Finds interesting uses (item 1). */ |
8228 | find_interesting_uses (data, body); |
8229 | if (data->vgroups.length () > MAX_CONSIDERED_GROUPS) |
8230 | goto finish; |
8231 | |
8232 | /* Determine cost scaling factor for basic blocks in loop. */ |
8233 | determine_scaling_factor (data, body); |
8234 | |
8235 | /* Analyze doloop possibility and mark the doloop use if predicted. */ |
8236 | analyze_and_mark_doloop_use (data); |
8237 | |
8238 | /* Finds candidates for the induction variables (item 2). */ |
8239 | find_iv_candidates (data); |
8240 | |
8241 | /* Calculates the costs (item 3, part 1). */ |
8242 | determine_iv_costs (data); |
8243 | determine_group_iv_costs (data); |
8244 | determine_set_costs (data); |
8245 | |
8246 | /* Find the optimal set of induction variables (item 3, part 2). */ |
8247 | iv_ca = find_optimal_iv_set (data); |
8248 | /* Cleanup basic block aux field. */ |
8249 | for (unsigned i = 0; i < data->current_loop->num_nodes; i++) |
8250 | body[i]->aux = NULL; |
8251 | if (!iv_ca) |
8252 | goto finish; |
8253 | changed = true; |
8254 | |
8255 | /* Create the new induction variables (item 4, part 1). */ |
8256 | create_new_ivs (data, set: iv_ca); |
8257 | iv_ca_free (ivs: &iv_ca); |
8258 | |
8259 | /* Rewrite the uses (item 4, part 2). */ |
8260 | rewrite_groups (data); |
8261 | |
8262 | /* Remove the ivs that are unused after rewriting. */ |
8263 | remove_unused_ivs (data, toremove); |
8264 | |
8265 | finish: |
8266 | free (ptr: body); |
8267 | free_loop_data (data); |
8268 | |
8269 | return changed; |
8270 | } |
8271 | |
8272 | /* Main entry point. Optimizes induction variables in loops. */ |
8273 | |
8274 | void |
8275 | tree_ssa_iv_optimize (void) |
8276 | { |
8277 | struct ivopts_data data; |
8278 | auto_bitmap toremove; |
8279 | |
8280 | tree_ssa_iv_optimize_init (data: &data); |
8281 | mark_ssa_maybe_undefs (); |
8282 | |
8283 | /* Optimize the loops starting with the innermost ones. */ |
8284 | for (auto loop : loops_list (cfun, LI_FROM_INNERMOST)) |
8285 | { |
8286 | if (!dbg_cnt (index: ivopts_loop)) |
8287 | continue; |
8288 | |
8289 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8290 | flow_loop_dump (loop, dump_file, NULL, 1); |
8291 | |
8292 | tree_ssa_iv_optimize_loop (data: &data, loop, toremove); |
8293 | } |
8294 | |
8295 | /* Remove eliminated IV defs. */ |
8296 | release_defs_bitset (toremove); |
8297 | |
8298 | /* We have changed the structure of induction variables; it might happen |
8299 | that definitions in the scev database refer to some of them that were |
8300 | eliminated. */ |
8301 | scev_reset_htab (); |
8302 | /* Likewise niter and control-IV information. */ |
8303 | free_numbers_of_iterations_estimates (cfun); |
8304 | |
8305 | tree_ssa_iv_optimize_finalize (data: &data); |
8306 | } |
8307 | |
8308 | #include "gt-tree-ssa-loop-ivopts.h" |
8309 | |