]>
Commit | Line | Data |
---|---|---|
8b11a64c | 1 | /* Induction variable optimizations. |
5624e564 | 2 | Copyright (C) 2003-2015 Free Software Foundation, Inc. |
b8698a0f | 3 | |
8b11a64c | 4 | This file is part of GCC. |
b8698a0f | 5 | |
8b11a64c ZD |
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 | |
9dcd6f09 | 8 | Free Software Foundation; either version 3, or (at your option) any |
8b11a64c | 9 | later version. |
b8698a0f | 10 | |
8b11a64c ZD |
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. | |
b8698a0f | 15 | |
8b11a64c | 16 | You should have received a copy of the GNU General Public License |
9dcd6f09 NC |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
8b11a64c ZD |
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 | 2) Candidates for the induction variables are found. This includes | |
34 | ||
35 | -- old induction variables | |
36 | -- the variables defined by expressions derived from the "interesting | |
37 | uses" above | |
38 | ||
39 | 3) The optimal (w.r. to a cost function) set of variables is chosen. The | |
40 | cost function assigns a cost to sets of induction variables and consists | |
41 | of three parts: | |
42 | ||
43 | -- The use costs. Each of the interesting uses chooses the best induction | |
44 | variable in the set and adds its cost to the sum. The cost reflects | |
45 | the time spent on modifying the induction variables value to be usable | |
46 | for the given purpose (adding base and offset for arrays, etc.). | |
47 | -- The variable costs. Each of the variables has a cost assigned that | |
48 | reflects the costs associated with incrementing the value of the | |
49 | variable. The original variables are somewhat preferred. | |
50 | -- The set cost. Depending on the size of the set, extra cost may be | |
51 | added to reflect register pressure. | |
52 | ||
53 | All the costs are defined in a machine-specific way, using the target | |
54 | hooks and machine descriptions to determine them. | |
55 | ||
56 | 4) The trees are transformed to use the new variables, the dead code is | |
57 | removed. | |
b8698a0f | 58 | |
8b11a64c ZD |
59 | All of this is done loop by loop. Doing it globally is theoretically |
60 | possible, it might give a better performance and it might enable us | |
61 | to decide costs more precisely, but getting all the interactions right | |
62 | would be complicated. */ | |
63 | ||
64 | #include "config.h" | |
65 | #include "system.h" | |
66 | #include "coretypes.h" | |
c7131fb2 | 67 | #include "backend.h" |
8b11a64c | 68 | #include "tree.h" |
c7131fb2 AM |
69 | #include "gimple.h" |
70 | #include "rtl.h" | |
71 | #include "ssa.h" | |
72 | #include "alias.h" | |
40e23961 | 73 | #include "fold-const.h" |
d8a2d370 | 74 | #include "stor-layout.h" |
8b11a64c | 75 | #include "tm_p.h" |
cf835838 | 76 | #include "gimple-pretty-print.h" |
2fb9a547 AM |
77 | #include "internal-fn.h" |
78 | #include "tree-eh.h" | |
45b0be94 | 79 | #include "gimplify.h" |
5be5c238 | 80 | #include "gimple-iterator.h" |
18f429e2 | 81 | #include "gimplify-me.h" |
442b4905 AM |
82 | #include "cgraph.h" |
83 | #include "tree-cfg.h" | |
e28030cf AM |
84 | #include "tree-ssa-loop-ivopts.h" |
85 | #include "tree-ssa-loop-manip.h" | |
86 | #include "tree-ssa-loop-niter.h" | |
442b4905 | 87 | #include "tree-ssa-loop.h" |
36566b39 | 88 | #include "flags.h" |
36566b39 PK |
89 | #include "insn-config.h" |
90 | #include "expmed.h" | |
91 | #include "dojump.h" | |
92 | #include "explow.h" | |
93 | #include "calls.h" | |
94 | #include "emit-rtl.h" | |
95 | #include "varasm.h" | |
96 | #include "stmt.h" | |
d8a2d370 | 97 | #include "expr.h" |
442b4905 | 98 | #include "tree-dfa.h" |
7a300452 | 99 | #include "tree-ssa.h" |
8b11a64c | 100 | #include "cfgloop.h" |
8b11a64c | 101 | #include "tree-pass.h" |
8b11a64c ZD |
102 | #include "tree-chrec.h" |
103 | #include "tree-scalar-evolution.h" | |
8b11a64c | 104 | #include "params.h" |
39b4020c | 105 | #include "langhooks.h" |
73f30c63 | 106 | #include "tree-affine.h" |
8318b0d9 | 107 | #include "target.h" |
18081149 | 108 | #include "tree-inline.h" |
17fc049f | 109 | #include "tree-ssa-propagate.h" |
4484a35a | 110 | #include "tree-ssa-address.h" |
9b2b7279 | 111 | #include "builtins.h" |
28002f1a | 112 | #include "tree-vectorizer.h" |
7735d6c7 | 113 | |
2eb79bbb SB |
114 | /* FIXME: Expressions are expanded to RTL in this pass to determine the |
115 | cost of different addressing modes. This should be moved to a TBD | |
116 | interface between the GIMPLE and RTL worlds. */ | |
1c1ad7bb | 117 | #include "recog.h" |
2eb79bbb | 118 | |
8b11a64c ZD |
119 | /* The infinite cost. */ |
120 | #define INFTY 10000000 | |
121 | ||
8b11a64c ZD |
122 | #define AVG_LOOP_NITER(LOOP) 5 |
123 | ||
18081149 XDL |
124 | /* Returns the expected number of loop iterations for LOOP. |
125 | The average trip count is computed from profile data if it | |
126 | exists. */ | |
127 | ||
128 | static inline HOST_WIDE_INT | |
129 | avg_loop_niter (struct loop *loop) | |
130 | { | |
652c4c71 | 131 | HOST_WIDE_INT niter = estimated_stmt_executions_int (loop); |
18081149 XDL |
132 | if (niter == -1) |
133 | return AVG_LOOP_NITER (loop); | |
134 | ||
135 | return niter; | |
136 | } | |
8b11a64c ZD |
137 | |
138 | /* Representation of the induction variable. */ | |
139 | struct iv | |
140 | { | |
141 | tree base; /* Initial value of the iv. */ | |
e6845c23 | 142 | tree base_object; /* A memory object to that the induction variable points. */ |
8b11a64c ZD |
143 | tree step; /* Step of the iv (constant only). */ |
144 | tree ssa_name; /* The ssa name with the value. */ | |
c70ed622 | 145 | unsigned use_id; /* The identifier in the use if it is the case. */ |
8b11a64c ZD |
146 | bool biv_p; /* Is it a biv? */ |
147 | bool have_use_for; /* Do we already have a use for it? */ | |
c70ed622 | 148 | bool no_overflow; /* True if the iv doesn't overflow. */ |
8b11a64c ZD |
149 | }; |
150 | ||
151 | /* Per-ssa version information (induction variable descriptions, etc.). */ | |
152 | struct version_info | |
153 | { | |
154 | tree name; /* The ssa name. */ | |
155 | struct iv *iv; /* Induction variable description. */ | |
156 | bool has_nonlin_use; /* For a loop-level invariant, whether it is used in | |
157 | an expression that is not an induction variable. */ | |
8b11a64c | 158 | bool preserve_biv; /* For the original biv, whether to preserve it. */ |
448f65db | 159 | unsigned inv_id; /* Id of an invariant. */ |
8b11a64c ZD |
160 | }; |
161 | ||
8b11a64c ZD |
162 | /* Types of uses. */ |
163 | enum use_type | |
164 | { | |
165 | USE_NONLINEAR_EXPR, /* Use in a nonlinear expression. */ | |
8b11a64c ZD |
166 | USE_ADDRESS, /* Use in an address. */ |
167 | USE_COMPARE /* Use is a compare. */ | |
168 | }; | |
169 | ||
6e8c65f6 ZD |
170 | /* Cost of a computation. */ |
171 | typedef struct | |
172 | { | |
2c08497a | 173 | int cost; /* The runtime cost. */ |
6e8c65f6 ZD |
174 | unsigned complexity; /* The estimate of the complexity of the code for |
175 | the computation (in no concrete units -- | |
176 | complexity field should be larger for more | |
177 | complex expressions and addressing modes). */ | |
178 | } comp_cost; | |
179 | ||
7735d6c7 | 180 | static const comp_cost no_cost = {0, 0}; |
6e8c65f6 ZD |
181 | static const comp_cost infinite_cost = {INFTY, INFTY}; |
182 | ||
8b11a64c ZD |
183 | /* The candidate - cost pair. */ |
184 | struct cost_pair | |
185 | { | |
186 | struct iv_cand *cand; /* The candidate. */ | |
6e8c65f6 | 187 | comp_cost cost; /* The cost. */ |
8b11a64c ZD |
188 | bitmap depends_on; /* The list of invariants that have to be |
189 | preserved. */ | |
f5f12961 ZD |
190 | tree value; /* For final value elimination, the expression for |
191 | the final value of the iv. For iv elimination, | |
192 | the new bound to compare with. */ | |
d8af4ba3 | 193 | enum tree_code comp; /* For iv elimination, the comparison. */ |
18081149 | 194 | int inv_expr_id; /* Loop invariant expression id. */ |
8b11a64c ZD |
195 | }; |
196 | ||
197 | /* Use. */ | |
198 | struct iv_use | |
199 | { | |
200 | unsigned id; /* The id of the use. */ | |
a7e43c57 | 201 | unsigned sub_id; /* The id of the sub use. */ |
8b11a64c ZD |
202 | enum use_type type; /* Type of the use. */ |
203 | struct iv *iv; /* The induction variable it is based on. */ | |
726a989a | 204 | gimple stmt; /* Statement in that it occurs. */ |
8b11a64c | 205 | tree *op_p; /* The place where it occurs. */ |
b1b02be2 ZD |
206 | bitmap related_cands; /* The set of "related" iv candidates, plus the common |
207 | important ones. */ | |
8b11a64c ZD |
208 | |
209 | unsigned n_map_members; /* Number of candidates in the cost_map list. */ | |
210 | struct cost_pair *cost_map; | |
211 | /* The costs wrto the iv candidates. */ | |
212 | ||
213 | struct iv_cand *selected; | |
214 | /* The selected candidate. */ | |
a7e43c57 BC |
215 | |
216 | struct iv_use *next; /* The next sub use. */ | |
217 | tree addr_base; /* Base address with const offset stripped. */ | |
218 | unsigned HOST_WIDE_INT addr_offset; | |
219 | /* Const offset stripped from base address. */ | |
8b11a64c ZD |
220 | }; |
221 | ||
222 | /* The position where the iv is computed. */ | |
223 | enum iv_position | |
224 | { | |
225 | IP_NORMAL, /* At the end, just before the exit condition. */ | |
226 | IP_END, /* At the end of the latch block. */ | |
2c08497a BS |
227 | IP_BEFORE_USE, /* Immediately before a specific use. */ |
228 | IP_AFTER_USE, /* Immediately after a specific use. */ | |
8b11a64c ZD |
229 | IP_ORIGINAL /* The original biv. */ |
230 | }; | |
231 | ||
232 | /* The induction variable candidate. */ | |
233 | struct iv_cand | |
234 | { | |
235 | unsigned id; /* The number of the candidate. */ | |
236 | bool important; /* Whether this is an "important" candidate, i.e. such | |
237 | that it should be considered by all uses. */ | |
448f65db | 238 | ENUM_BITFIELD(iv_position) pos : 8; /* Where it is computed. */ |
726a989a | 239 | gimple incremented_at;/* For original biv, the statement where it is |
8b11a64c ZD |
240 | incremented. */ |
241 | tree var_before; /* The variable used for it before increment. */ | |
242 | tree var_after; /* The variable used for it after increment. */ | |
243 | struct iv *iv; /* The value of the candidate. NULL for | |
244 | "pseudocandidate" used to indicate the possibility | |
245 | to replace the final value of an iv by direct | |
246 | computation of the value. */ | |
247 | unsigned cost; /* Cost of the candidate. */ | |
2c08497a BS |
248 | unsigned cost_step; /* Cost of the candidate's increment operation. */ |
249 | struct iv_use *ainc_use; /* For IP_{BEFORE,AFTER}_USE candidates, the place | |
250 | where it is incremented. */ | |
9be872b7 ZD |
251 | bitmap depends_on; /* The list of invariants that are used in step of the |
252 | biv. */ | |
8b11a64c ZD |
253 | }; |
254 | ||
18081149 XDL |
255 | /* Loop invariant expression hashtable entry. */ |
256 | struct iv_inv_expr_ent | |
257 | { | |
258 | tree expr; | |
259 | int id; | |
260 | hashval_t hash; | |
261 | }; | |
262 | ||
8b11a64c ZD |
263 | /* The data used by the induction variable optimizations. */ |
264 | ||
69ebd99d | 265 | typedef struct iv_use *iv_use_p; |
69ebd99d KH |
266 | |
267 | typedef struct iv_cand *iv_cand_p; | |
69ebd99d | 268 | |
4a8fb1a1 LC |
269 | /* Hashtable helpers. */ |
270 | ||
95fbe13e | 271 | struct iv_inv_expr_hasher : free_ptr_hash <iv_inv_expr_ent> |
4a8fb1a1 | 272 | { |
67f58944 TS |
273 | static inline hashval_t hash (const iv_inv_expr_ent *); |
274 | static inline bool equal (const iv_inv_expr_ent *, const iv_inv_expr_ent *); | |
4a8fb1a1 LC |
275 | }; |
276 | ||
277 | /* Hash function for loop invariant expressions. */ | |
278 | ||
279 | inline hashval_t | |
67f58944 | 280 | iv_inv_expr_hasher::hash (const iv_inv_expr_ent *expr) |
4a8fb1a1 LC |
281 | { |
282 | return expr->hash; | |
283 | } | |
284 | ||
285 | /* Hash table equality function for expressions. */ | |
286 | ||
287 | inline bool | |
67f58944 TS |
288 | iv_inv_expr_hasher::equal (const iv_inv_expr_ent *expr1, |
289 | const iv_inv_expr_ent *expr2) | |
4a8fb1a1 LC |
290 | { |
291 | return expr1->hash == expr2->hash | |
292 | && operand_equal_p (expr1->expr, expr2->expr, 0); | |
293 | } | |
294 | ||
8b11a64c ZD |
295 | struct ivopts_data |
296 | { | |
297 | /* The currently optimized loop. */ | |
298 | struct loop *current_loop; | |
28002f1a | 299 | source_location loop_loc; |
8b11a64c | 300 | |
8f5929e1 | 301 | /* Numbers of iterations for all exits of the current loop. */ |
b787e7a2 | 302 | hash_map<edge, tree_niter_desc *> *niters; |
f40751dd | 303 | |
9a2ef6b8 ZD |
304 | /* Number of registers used in it. */ |
305 | unsigned regs_used; | |
306 | ||
8b11a64c ZD |
307 | /* The size of version_info array allocated. */ |
308 | unsigned version_info_size; | |
309 | ||
310 | /* The array of information for the ssa names. */ | |
311 | struct version_info *version_info; | |
312 | ||
18081149 XDL |
313 | /* The hashtable of loop invariant expressions created |
314 | by ivopt. */ | |
c203e8a7 | 315 | hash_table<iv_inv_expr_hasher> *inv_expr_tab; |
18081149 XDL |
316 | |
317 | /* Loop invariant expression id. */ | |
318 | int inv_expr_id; | |
319 | ||
8b11a64c ZD |
320 | /* The bitmap of indices in version_info whose value was changed. */ |
321 | bitmap relevant; | |
322 | ||
8b11a64c | 323 | /* The uses of induction variables. */ |
9771b263 | 324 | vec<iv_use_p> iv_uses; |
8b11a64c ZD |
325 | |
326 | /* The candidates. */ | |
9771b263 | 327 | vec<iv_cand_p> iv_candidates; |
8b11a64c | 328 | |
80cad5fa ZD |
329 | /* A bitmap of important candidates. */ |
330 | bitmap important_candidates; | |
331 | ||
3230c614 BC |
332 | /* Cache used by tree_to_aff_combination_expand. */ |
333 | hash_map<tree, name_expansion *> *name_expansion_cache; | |
334 | ||
8f5929e1 JJ |
335 | /* The maximum invariant id. */ |
336 | unsigned max_inv_id; | |
337 | ||
6f929985 BC |
338 | /* Obstack for iv structure. */ |
339 | struct obstack iv_obstack; | |
340 | ||
8b11a64c ZD |
341 | /* Whether to consider just related and important candidates when replacing a |
342 | use. */ | |
343 | bool consider_all_candidates; | |
8f5929e1 JJ |
344 | |
345 | /* Are we optimizing for speed? */ | |
346 | bool speed; | |
bec922f0 SL |
347 | |
348 | /* Whether the loop body includes any function calls. */ | |
349 | bool body_includes_call; | |
d8af4ba3 ZD |
350 | |
351 | /* Whether the loop body can only be exited via single exit. */ | |
352 | bool loop_single_exit_p; | |
8b11a64c ZD |
353 | }; |
354 | ||
b1b02be2 ZD |
355 | /* An assignment of iv candidates to uses. */ |
356 | ||
357 | struct iv_ca | |
358 | { | |
359 | /* The number of uses covered by the assignment. */ | |
360 | unsigned upto; | |
361 | ||
362 | /* Number of uses that cannot be expressed by the candidates in the set. */ | |
363 | unsigned bad_uses; | |
364 | ||
365 | /* Candidate assigned to a use, together with the related costs. */ | |
366 | struct cost_pair **cand_for_use; | |
367 | ||
368 | /* Number of times each candidate is used. */ | |
369 | unsigned *n_cand_uses; | |
370 | ||
371 | /* The candidates used. */ | |
372 | bitmap cands; | |
373 | ||
36f5ada1 ZD |
374 | /* The number of candidates in the set. */ |
375 | unsigned n_cands; | |
376 | ||
b1b02be2 ZD |
377 | /* Total number of registers needed. */ |
378 | unsigned n_regs; | |
379 | ||
380 | /* Total cost of expressing uses. */ | |
6e8c65f6 | 381 | comp_cost cand_use_cost; |
b1b02be2 ZD |
382 | |
383 | /* Total cost of candidates. */ | |
384 | unsigned cand_cost; | |
385 | ||
386 | /* Number of times each invariant is used. */ | |
387 | unsigned *n_invariant_uses; | |
388 | ||
f06e400f XDL |
389 | /* The array holding the number of uses of each loop |
390 | invariant expressions created by ivopt. */ | |
391 | unsigned *used_inv_expr; | |
392 | ||
393 | /* The number of created loop invariants. */ | |
394 | unsigned num_used_inv_expr; | |
395 | ||
b1b02be2 | 396 | /* Total cost of the assignment. */ |
6e8c65f6 | 397 | comp_cost cost; |
b1b02be2 ZD |
398 | }; |
399 | ||
400 | /* Difference of two iv candidate assignments. */ | |
401 | ||
402 | struct iv_ca_delta | |
403 | { | |
404 | /* Changed use. */ | |
405 | struct iv_use *use; | |
406 | ||
407 | /* An old assignment (for rollback purposes). */ | |
408 | struct cost_pair *old_cp; | |
409 | ||
410 | /* A new assignment. */ | |
411 | struct cost_pair *new_cp; | |
412 | ||
413 | /* Next change in the list. */ | |
414 | struct iv_ca_delta *next_change; | |
415 | }; | |
416 | ||
8b11a64c ZD |
417 | /* Bound on number of candidates below that all candidates are considered. */ |
418 | ||
419 | #define CONSIDER_ALL_CANDIDATES_BOUND \ | |
420 | ((unsigned) PARAM_VALUE (PARAM_IV_CONSIDER_ALL_CANDIDATES_BOUND)) | |
421 | ||
2a7e31df | 422 | /* If there are more iv occurrences, we just give up (it is quite unlikely that |
8b11a64c ZD |
423 | optimizing such a loop would help, and it would take ages). */ |
424 | ||
425 | #define MAX_CONSIDERED_USES \ | |
426 | ((unsigned) PARAM_VALUE (PARAM_IV_MAX_CONSIDERED_USES)) | |
427 | ||
36f5ada1 ZD |
428 | /* If there are at most this number of ivs in the set, try removing unnecessary |
429 | ivs from the set always. */ | |
430 | ||
431 | #define ALWAYS_PRUNE_CAND_SET_BOUND \ | |
432 | ((unsigned) PARAM_VALUE (PARAM_IV_ALWAYS_PRUNE_CAND_SET_BOUND)) | |
433 | ||
8b11a64c ZD |
434 | /* The list of trees for that the decl_rtl field must be reset is stored |
435 | here. */ | |
436 | ||
9771b263 | 437 | static vec<tree> decl_rtl_to_reset; |
8b11a64c | 438 | |
e6450c11 TV |
439 | static comp_cost force_expr_to_var_cost (tree, bool); |
440 | ||
8b11a64c ZD |
441 | /* Number of uses recorded in DATA. */ |
442 | ||
443 | static inline unsigned | |
444 | n_iv_uses (struct ivopts_data *data) | |
445 | { | |
9771b263 | 446 | return data->iv_uses.length (); |
8b11a64c ZD |
447 | } |
448 | ||
449 | /* Ith use recorded in DATA. */ | |
450 | ||
451 | static inline struct iv_use * | |
452 | iv_use (struct ivopts_data *data, unsigned i) | |
453 | { | |
9771b263 | 454 | return data->iv_uses[i]; |
8b11a64c ZD |
455 | } |
456 | ||
457 | /* Number of candidates recorded in DATA. */ | |
458 | ||
459 | static inline unsigned | |
460 | n_iv_cands (struct ivopts_data *data) | |
461 | { | |
9771b263 | 462 | return data->iv_candidates.length (); |
8b11a64c ZD |
463 | } |
464 | ||
465 | /* Ith candidate recorded in DATA. */ | |
466 | ||
467 | static inline struct iv_cand * | |
468 | iv_cand (struct ivopts_data *data, unsigned i) | |
469 | { | |
9771b263 | 470 | return data->iv_candidates[i]; |
8b11a64c ZD |
471 | } |
472 | ||
8b11a64c ZD |
473 | /* The single loop exit if it dominates the latch, NULL otherwise. */ |
474 | ||
b7eae7b8 | 475 | edge |
8b11a64c ZD |
476 | single_dom_exit (struct loop *loop) |
477 | { | |
ac8f6c69 | 478 | edge exit = single_exit (loop); |
8b11a64c ZD |
479 | |
480 | if (!exit) | |
481 | return NULL; | |
482 | ||
483 | if (!just_once_each_iteration_p (loop, exit->src)) | |
484 | return NULL; | |
485 | ||
486 | return exit; | |
487 | } | |
488 | ||
489 | /* Dumps information about the induction variable IV to FILE. */ | |
490 | ||
8b11a64c | 491 | void |
e185f450 | 492 | dump_iv (FILE *file, struct iv *iv, bool dump_name) |
8b11a64c | 493 | { |
e185f450 | 494 | if (iv->ssa_name && dump_name) |
e6845c23 ZD |
495 | { |
496 | fprintf (file, "ssa name "); | |
497 | print_generic_expr (file, iv->ssa_name, TDF_SLIM); | |
498 | fprintf (file, "\n"); | |
499 | } | |
8b11a64c | 500 | |
2f4675b4 ZD |
501 | fprintf (file, " type "); |
502 | print_generic_expr (file, TREE_TYPE (iv->base), TDF_SLIM); | |
503 | fprintf (file, "\n"); | |
504 | ||
8b11a64c ZD |
505 | if (iv->step) |
506 | { | |
507 | fprintf (file, " base "); | |
508 | print_generic_expr (file, iv->base, TDF_SLIM); | |
509 | fprintf (file, "\n"); | |
510 | ||
511 | fprintf (file, " step "); | |
512 | print_generic_expr (file, iv->step, TDF_SLIM); | |
513 | fprintf (file, "\n"); | |
514 | } | |
515 | else | |
516 | { | |
517 | fprintf (file, " invariant "); | |
518 | print_generic_expr (file, iv->base, TDF_SLIM); | |
519 | fprintf (file, "\n"); | |
520 | } | |
521 | ||
e6845c23 ZD |
522 | if (iv->base_object) |
523 | { | |
524 | fprintf (file, " base object "); | |
525 | print_generic_expr (file, iv->base_object, TDF_SLIM); | |
526 | fprintf (file, "\n"); | |
527 | } | |
528 | ||
8b11a64c ZD |
529 | if (iv->biv_p) |
530 | fprintf (file, " is a biv\n"); | |
531 | } | |
532 | ||
533 | /* Dumps information about the USE to FILE. */ | |
534 | ||
8b11a64c ZD |
535 | void |
536 | dump_use (FILE *file, struct iv_use *use) | |
537 | { | |
a7e43c57 BC |
538 | fprintf (file, "use %d", use->id); |
539 | if (use->sub_id) | |
540 | fprintf (file, ".%d", use->sub_id); | |
541 | ||
542 | fprintf (file, "\n"); | |
8b11a64c ZD |
543 | |
544 | switch (use->type) | |
545 | { | |
546 | case USE_NONLINEAR_EXPR: | |
547 | fprintf (file, " generic\n"); | |
548 | break; | |
549 | ||
8b11a64c ZD |
550 | case USE_ADDRESS: |
551 | fprintf (file, " address\n"); | |
552 | break; | |
553 | ||
554 | case USE_COMPARE: | |
555 | fprintf (file, " compare\n"); | |
556 | break; | |
557 | ||
558 | default: | |
1e128c5f | 559 | gcc_unreachable (); |
8b11a64c ZD |
560 | } |
561 | ||
2f4675b4 | 562 | fprintf (file, " in statement "); |
726a989a | 563 | print_gimple_stmt (file, use->stmt, 0, 0); |
2f4675b4 ZD |
564 | fprintf (file, "\n"); |
565 | ||
566 | fprintf (file, " at position "); | |
567 | if (use->op_p) | |
568 | print_generic_expr (file, *use->op_p, TDF_SLIM); | |
569 | fprintf (file, "\n"); | |
570 | ||
e185f450 | 571 | dump_iv (file, use->iv, false); |
2f4675b4 | 572 | |
eec5fec9 ZD |
573 | if (use->related_cands) |
574 | { | |
575 | fprintf (file, " related candidates "); | |
576 | dump_bitmap (file, use->related_cands); | |
577 | } | |
8b11a64c ZD |
578 | } |
579 | ||
580 | /* Dumps information about the uses to FILE. */ | |
581 | ||
8b11a64c ZD |
582 | void |
583 | dump_uses (FILE *file, struct ivopts_data *data) | |
584 | { | |
585 | unsigned i; | |
586 | struct iv_use *use; | |
587 | ||
588 | for (i = 0; i < n_iv_uses (data); i++) | |
589 | { | |
590 | use = iv_use (data, i); | |
a7e43c57 BC |
591 | do |
592 | { | |
593 | dump_use (file, use); | |
594 | use = use->next; | |
595 | } | |
596 | while (use); | |
8b11a64c ZD |
597 | fprintf (file, "\n"); |
598 | } | |
599 | } | |
600 | ||
601 | /* Dumps information about induction variable candidate CAND to FILE. */ | |
602 | ||
8b11a64c ZD |
603 | void |
604 | dump_cand (FILE *file, struct iv_cand *cand) | |
605 | { | |
606 | struct iv *iv = cand->iv; | |
607 | ||
608 | fprintf (file, "candidate %d%s\n", | |
609 | cand->id, cand->important ? " (important)" : ""); | |
610 | ||
9be872b7 ZD |
611 | if (cand->depends_on) |
612 | { | |
613 | fprintf (file, " depends on "); | |
614 | dump_bitmap (file, cand->depends_on); | |
615 | } | |
616 | ||
8b11a64c ZD |
617 | if (!iv) |
618 | { | |
619 | fprintf (file, " final value replacement\n"); | |
620 | return; | |
621 | } | |
622 | ||
18081149 XDL |
623 | if (cand->var_before) |
624 | { | |
625 | fprintf (file, " var_before "); | |
626 | print_generic_expr (file, cand->var_before, TDF_SLIM); | |
627 | fprintf (file, "\n"); | |
628 | } | |
629 | if (cand->var_after) | |
630 | { | |
631 | fprintf (file, " var_after "); | |
632 | print_generic_expr (file, cand->var_after, TDF_SLIM); | |
633 | fprintf (file, "\n"); | |
634 | } | |
635 | ||
8b11a64c ZD |
636 | switch (cand->pos) |
637 | { | |
638 | case IP_NORMAL: | |
639 | fprintf (file, " incremented before exit test\n"); | |
640 | break; | |
641 | ||
2c08497a BS |
642 | case IP_BEFORE_USE: |
643 | fprintf (file, " incremented before use %d\n", cand->ainc_use->id); | |
644 | break; | |
645 | ||
646 | case IP_AFTER_USE: | |
647 | fprintf (file, " incremented after use %d\n", cand->ainc_use->id); | |
648 | break; | |
649 | ||
8b11a64c ZD |
650 | case IP_END: |
651 | fprintf (file, " incremented at end\n"); | |
652 | break; | |
653 | ||
654 | case IP_ORIGINAL: | |
655 | fprintf (file, " original biv\n"); | |
656 | break; | |
657 | } | |
658 | ||
e185f450 | 659 | dump_iv (file, iv, false); |
8b11a64c ZD |
660 | } |
661 | ||
662 | /* Returns the info for ssa version VER. */ | |
663 | ||
664 | static inline struct version_info * | |
665 | ver_info (struct ivopts_data *data, unsigned ver) | |
666 | { | |
667 | return data->version_info + ver; | |
668 | } | |
669 | ||
670 | /* Returns the info for ssa name NAME. */ | |
671 | ||
672 | static inline struct version_info * | |
673 | name_info (struct ivopts_data *data, tree name) | |
674 | { | |
675 | return ver_info (data, SSA_NAME_VERSION (name)); | |
676 | } | |
677 | ||
8b11a64c ZD |
678 | /* Returns true if STMT is after the place where the IP_NORMAL ivs will be |
679 | emitted in LOOP. */ | |
680 | ||
681 | static bool | |
726a989a | 682 | stmt_after_ip_normal_pos (struct loop *loop, gimple stmt) |
8b11a64c | 683 | { |
726a989a | 684 | basic_block bb = ip_normal_pos (loop), sbb = gimple_bb (stmt); |
8b11a64c | 685 | |
1e128c5f | 686 | gcc_assert (bb); |
8b11a64c ZD |
687 | |
688 | if (sbb == loop->latch) | |
689 | return true; | |
690 | ||
691 | if (sbb != bb) | |
692 | return false; | |
693 | ||
694 | return stmt == last_stmt (bb); | |
695 | } | |
696 | ||
697 | /* Returns true if STMT if after the place where the original induction | |
2c08497a BS |
698 | variable CAND is incremented. If TRUE_IF_EQUAL is set, we return true |
699 | if the positions are identical. */ | |
8b11a64c ZD |
700 | |
701 | static bool | |
2c08497a | 702 | stmt_after_inc_pos (struct iv_cand *cand, gimple stmt, bool true_if_equal) |
8b11a64c | 703 | { |
726a989a RB |
704 | basic_block cand_bb = gimple_bb (cand->incremented_at); |
705 | basic_block stmt_bb = gimple_bb (stmt); | |
8b11a64c ZD |
706 | |
707 | if (!dominated_by_p (CDI_DOMINATORS, stmt_bb, cand_bb)) | |
708 | return false; | |
709 | ||
710 | if (stmt_bb != cand_bb) | |
711 | return true; | |
712 | ||
2c08497a BS |
713 | if (true_if_equal |
714 | && gimple_uid (stmt) == gimple_uid (cand->incremented_at)) | |
715 | return true; | |
716 | return gimple_uid (stmt) > gimple_uid (cand->incremented_at); | |
8b11a64c ZD |
717 | } |
718 | ||
719 | /* Returns true if STMT if after the place where the induction variable | |
720 | CAND is incremented in LOOP. */ | |
721 | ||
722 | static bool | |
726a989a | 723 | stmt_after_increment (struct loop *loop, struct iv_cand *cand, gimple stmt) |
8b11a64c ZD |
724 | { |
725 | switch (cand->pos) | |
726 | { | |
727 | case IP_END: | |
728 | return false; | |
729 | ||
730 | case IP_NORMAL: | |
731 | return stmt_after_ip_normal_pos (loop, stmt); | |
732 | ||
733 | case IP_ORIGINAL: | |
2c08497a BS |
734 | case IP_AFTER_USE: |
735 | return stmt_after_inc_pos (cand, stmt, false); | |
736 | ||
737 | case IP_BEFORE_USE: | |
738 | return stmt_after_inc_pos (cand, stmt, true); | |
8b11a64c ZD |
739 | |
740 | default: | |
1e128c5f | 741 | gcc_unreachable (); |
8b11a64c ZD |
742 | } |
743 | } | |
744 | ||
dcccd88d ZD |
745 | /* Returns true if EXP is a ssa name that occurs in an abnormal phi node. */ |
746 | ||
747 | static bool | |
748 | abnormal_ssa_name_p (tree exp) | |
749 | { | |
750 | if (!exp) | |
751 | return false; | |
752 | ||
753 | if (TREE_CODE (exp) != SSA_NAME) | |
754 | return false; | |
755 | ||
756 | return SSA_NAME_OCCURS_IN_ABNORMAL_PHI (exp) != 0; | |
757 | } | |
758 | ||
759 | /* Returns false if BASE or INDEX contains a ssa name that occurs in an | |
760 | abnormal phi node. Callback for for_each_index. */ | |
761 | ||
762 | static bool | |
763 | idx_contains_abnormal_ssa_name_p (tree base, tree *index, | |
764 | void *data ATTRIBUTE_UNUSED) | |
765 | { | |
9f7ccf69 | 766 | if (TREE_CODE (base) == ARRAY_REF || TREE_CODE (base) == ARRAY_RANGE_REF) |
dcccd88d ZD |
767 | { |
768 | if (abnormal_ssa_name_p (TREE_OPERAND (base, 2))) | |
769 | return false; | |
770 | if (abnormal_ssa_name_p (TREE_OPERAND (base, 3))) | |
771 | return false; | |
772 | } | |
773 | ||
774 | return !abnormal_ssa_name_p (*index); | |
775 | } | |
776 | ||
777 | /* Returns true if EXPR contains a ssa name that occurs in an | |
778 | abnormal phi node. */ | |
779 | ||
e5db3515 | 780 | bool |
dcccd88d ZD |
781 | contains_abnormal_ssa_name_p (tree expr) |
782 | { | |
783 | enum tree_code code; | |
c22940cd | 784 | enum tree_code_class codeclass; |
dcccd88d ZD |
785 | |
786 | if (!expr) | |
787 | return false; | |
788 | ||
789 | code = TREE_CODE (expr); | |
c22940cd | 790 | codeclass = TREE_CODE_CLASS (code); |
dcccd88d ZD |
791 | |
792 | if (code == SSA_NAME) | |
793 | return SSA_NAME_OCCURS_IN_ABNORMAL_PHI (expr) != 0; | |
794 | ||
795 | if (code == INTEGER_CST | |
796 | || is_gimple_min_invariant (expr)) | |
797 | return false; | |
798 | ||
799 | if (code == ADDR_EXPR) | |
800 | return !for_each_index (&TREE_OPERAND (expr, 0), | |
801 | idx_contains_abnormal_ssa_name_p, | |
802 | NULL); | |
803 | ||
0a74c758 SP |
804 | if (code == COND_EXPR) |
805 | return contains_abnormal_ssa_name_p (TREE_OPERAND (expr, 0)) | |
806 | || contains_abnormal_ssa_name_p (TREE_OPERAND (expr, 1)) | |
807 | || contains_abnormal_ssa_name_p (TREE_OPERAND (expr, 2)); | |
808 | ||
c22940cd | 809 | switch (codeclass) |
dcccd88d ZD |
810 | { |
811 | case tcc_binary: | |
812 | case tcc_comparison: | |
813 | if (contains_abnormal_ssa_name_p (TREE_OPERAND (expr, 1))) | |
814 | return true; | |
815 | ||
816 | /* Fallthru. */ | |
817 | case tcc_unary: | |
818 | if (contains_abnormal_ssa_name_p (TREE_OPERAND (expr, 0))) | |
819 | return true; | |
820 | ||
821 | break; | |
822 | ||
823 | default: | |
824 | gcc_unreachable (); | |
825 | } | |
826 | ||
827 | return false; | |
828 | } | |
829 | ||
d8af4ba3 | 830 | /* Returns the structure describing number of iterations determined from |
ca4c3169 ZD |
831 | EXIT of DATA->current_loop, or NULL if something goes wrong. */ |
832 | ||
d8af4ba3 ZD |
833 | static struct tree_niter_desc * |
834 | niter_for_exit (struct ivopts_data *data, edge exit) | |
ca4c3169 | 835 | { |
d8af4ba3 | 836 | struct tree_niter_desc *desc; |
b787e7a2 | 837 | tree_niter_desc **slot; |
ca4c3169 | 838 | |
15814ba0 | 839 | if (!data->niters) |
ca4c3169 | 840 | { |
b787e7a2 | 841 | data->niters = new hash_map<edge, tree_niter_desc *>; |
15814ba0 PB |
842 | slot = NULL; |
843 | } | |
844 | else | |
b787e7a2 | 845 | slot = data->niters->get (exit); |
dcccd88d | 846 | |
15814ba0 PB |
847 | if (!slot) |
848 | { | |
d8af4ba3 ZD |
849 | /* Try to determine number of iterations. We cannot safely work with ssa |
850 | names that appear in phi nodes on abnormal edges, so that we do not | |
851 | create overlapping life ranges for them (PR 27283). */ | |
e2102efc | 852 | desc = XNEW (struct tree_niter_desc); |
d8af4ba3 ZD |
853 | if (!number_of_iterations_exit (data->current_loop, |
854 | exit, desc, true) | |
855 | || contains_abnormal_ssa_name_p (desc->niter)) | |
856 | { | |
857 | XDELETE (desc); | |
858 | desc = NULL; | |
859 | } | |
b787e7a2 | 860 | data->niters->put (exit, desc); |
ca4c3169 ZD |
861 | } |
862 | else | |
b787e7a2 | 863 | desc = *slot; |
ca4c3169 | 864 | |
d8af4ba3 | 865 | return desc; |
ca4c3169 ZD |
866 | } |
867 | ||
d8af4ba3 | 868 | /* Returns the structure describing number of iterations determined from |
ca4c3169 ZD |
869 | single dominating exit of DATA->current_loop, or NULL if something |
870 | goes wrong. */ | |
871 | ||
d8af4ba3 | 872 | static struct tree_niter_desc * |
ca4c3169 ZD |
873 | niter_for_single_dom_exit (struct ivopts_data *data) |
874 | { | |
875 | edge exit = single_dom_exit (data->current_loop); | |
876 | ||
877 | if (!exit) | |
878 | return NULL; | |
879 | ||
d8af4ba3 | 880 | return niter_for_exit (data, exit); |
ca4c3169 ZD |
881 | } |
882 | ||
8b11a64c | 883 | /* Initializes data structures used by the iv optimization pass, stored |
9a2ef6b8 | 884 | in DATA. */ |
8b11a64c ZD |
885 | |
886 | static void | |
9a2ef6b8 | 887 | tree_ssa_iv_optimize_init (struct ivopts_data *data) |
8b11a64c | 888 | { |
8b11a64c | 889 | data->version_info_size = 2 * num_ssa_names; |
5ed6ace5 | 890 | data->version_info = XCNEWVEC (struct version_info, data->version_info_size); |
8bdbfff5 NS |
891 | data->relevant = BITMAP_ALLOC (NULL); |
892 | data->important_candidates = BITMAP_ALLOC (NULL); | |
8b11a64c | 893 | data->max_inv_id = 0; |
15814ba0 | 894 | data->niters = NULL; |
9771b263 DN |
895 | data->iv_uses.create (20); |
896 | data->iv_candidates.create (20); | |
c203e8a7 | 897 | data->inv_expr_tab = new hash_table<iv_inv_expr_hasher> (10); |
18081149 | 898 | data->inv_expr_id = 0; |
3230c614 | 899 | data->name_expansion_cache = NULL; |
9771b263 | 900 | decl_rtl_to_reset.create (20); |
6f929985 | 901 | gcc_obstack_init (&data->iv_obstack); |
8b11a64c ZD |
902 | } |
903 | ||
e6845c23 ZD |
904 | /* Returns a memory object to that EXPR points. In case we are able to |
905 | determine that it does not point to any such object, NULL is returned. */ | |
906 | ||
907 | static tree | |
908 | determine_base_object (tree expr) | |
909 | { | |
910 | enum tree_code code = TREE_CODE (expr); | |
5be014d5 | 911 | tree base, obj; |
e6845c23 | 912 | |
975626a7 ZD |
913 | /* If this is a pointer casted to any type, we need to determine |
914 | the base object for the pointer; so handle conversions before | |
915 | throwing away non-pointer expressions. */ | |
1043771b | 916 | if (CONVERT_EXPR_P (expr)) |
975626a7 ZD |
917 | return determine_base_object (TREE_OPERAND (expr, 0)); |
918 | ||
e6845c23 ZD |
919 | if (!POINTER_TYPE_P (TREE_TYPE (expr))) |
920 | return NULL_TREE; | |
921 | ||
922 | switch (code) | |
923 | { | |
924 | case INTEGER_CST: | |
925 | return NULL_TREE; | |
926 | ||
927 | case ADDR_EXPR: | |
928 | obj = TREE_OPERAND (expr, 0); | |
929 | base = get_base_address (obj); | |
930 | ||
931 | if (!base) | |
f5e2738c | 932 | return expr; |
e6845c23 | 933 | |
70f34814 | 934 | if (TREE_CODE (base) == MEM_REF) |
f5e2738c | 935 | return determine_base_object (TREE_OPERAND (base, 0)); |
7299dbfb | 936 | |
62b37d91 RG |
937 | return fold_convert (ptr_type_node, |
938 | build_fold_addr_expr (base)); | |
e6845c23 | 939 | |
5be014d5 AP |
940 | case POINTER_PLUS_EXPR: |
941 | return determine_base_object (TREE_OPERAND (expr, 0)); | |
942 | ||
e6845c23 ZD |
943 | case PLUS_EXPR: |
944 | case MINUS_EXPR: | |
5be014d5 AP |
945 | /* Pointer addition is done solely using POINTER_PLUS_EXPR. */ |
946 | gcc_unreachable (); | |
e6845c23 ZD |
947 | |
948 | default: | |
949 | return fold_convert (ptr_type_node, expr); | |
950 | } | |
951 | } | |
952 | ||
be9a0da5 BC |
953 | /* Return true if address expression with non-DECL_P operand appears |
954 | in EXPR. */ | |
955 | ||
956 | static bool | |
957 | contain_complex_addr_expr (tree expr) | |
958 | { | |
959 | bool res = false; | |
960 | ||
961 | STRIP_NOPS (expr); | |
962 | switch (TREE_CODE (expr)) | |
963 | { | |
964 | case POINTER_PLUS_EXPR: | |
965 | case PLUS_EXPR: | |
966 | case MINUS_EXPR: | |
967 | res |= contain_complex_addr_expr (TREE_OPERAND (expr, 0)); | |
968 | res |= contain_complex_addr_expr (TREE_OPERAND (expr, 1)); | |
969 | break; | |
970 | ||
971 | case ADDR_EXPR: | |
972 | return (!DECL_P (TREE_OPERAND (expr, 0))); | |
973 | ||
974 | default: | |
975 | return false; | |
976 | } | |
977 | ||
978 | return res; | |
979 | } | |
980 | ||
8b11a64c | 981 | /* Allocates an induction variable with given initial value BASE and step STEP |
c70ed622 | 982 | for loop LOOP. NO_OVERFLOW implies the iv doesn't overflow. */ |
8b11a64c ZD |
983 | |
984 | static struct iv * | |
6f929985 BC |
985 | alloc_iv (struct ivopts_data *data, tree base, tree step, |
986 | bool no_overflow = false) | |
8b11a64c | 987 | { |
be9a0da5 | 988 | tree expr = base; |
6f929985 BC |
989 | struct iv *iv = (struct iv*) obstack_alloc (&data->iv_obstack, |
990 | sizeof (struct iv)); | |
6e42ce54 | 991 | gcc_assert (step != NULL_TREE); |
8b11a64c | 992 | |
be9a0da5 | 993 | /* Lower address expression in base except ones with DECL_P as operand. |
be8c1c8c BC |
994 | By doing this: |
995 | 1) More accurate cost can be computed for address expressions; | |
996 | 2) Duplicate candidates won't be created for bases in different | |
997 | forms, like &a[0] and &a. */ | |
be9a0da5 BC |
998 | STRIP_NOPS (expr); |
999 | if ((TREE_CODE (expr) == ADDR_EXPR && !DECL_P (TREE_OPERAND (expr, 0))) | |
1000 | || contain_complex_addr_expr (expr)) | |
be8c1c8c BC |
1001 | { |
1002 | aff_tree comb; | |
be9a0da5 | 1003 | tree_to_aff_combination (expr, TREE_TYPE (base), &comb); |
be8c1c8c BC |
1004 | base = fold_convert (TREE_TYPE (base), aff_combination_to_tree (&comb)); |
1005 | } | |
1006 | ||
8b11a64c | 1007 | iv->base = base; |
be9a0da5 | 1008 | iv->base_object = determine_base_object (base); |
8b11a64c ZD |
1009 | iv->step = step; |
1010 | iv->biv_p = false; | |
1011 | iv->have_use_for = false; | |
1012 | iv->use_id = 0; | |
1013 | iv->ssa_name = NULL_TREE; | |
c70ed622 | 1014 | iv->no_overflow = no_overflow; |
8b11a64c ZD |
1015 | |
1016 | return iv; | |
1017 | } | |
1018 | ||
c70ed622 BC |
1019 | /* Sets STEP and BASE for induction variable IV. NO_OVERFLOW implies the IV |
1020 | doesn't overflow. */ | |
8b11a64c ZD |
1021 | |
1022 | static void | |
c70ed622 BC |
1023 | set_iv (struct ivopts_data *data, tree iv, tree base, tree step, |
1024 | bool no_overflow) | |
8b11a64c ZD |
1025 | { |
1026 | struct version_info *info = name_info (data, iv); | |
1027 | ||
1e128c5f | 1028 | gcc_assert (!info->iv); |
8b11a64c ZD |
1029 | |
1030 | bitmap_set_bit (data->relevant, SSA_NAME_VERSION (iv)); | |
6f929985 | 1031 | info->iv = alloc_iv (data, base, step, no_overflow); |
8b11a64c ZD |
1032 | info->iv->ssa_name = iv; |
1033 | } | |
1034 | ||
1035 | /* Finds induction variable declaration for VAR. */ | |
1036 | ||
1037 | static struct iv * | |
1038 | get_iv (struct ivopts_data *data, tree var) | |
1039 | { | |
1040 | basic_block bb; | |
6e42ce54 ZD |
1041 | tree type = TREE_TYPE (var); |
1042 | ||
1043 | if (!POINTER_TYPE_P (type) | |
1044 | && !INTEGRAL_TYPE_P (type)) | |
1045 | return NULL; | |
1046 | ||
8b11a64c ZD |
1047 | if (!name_info (data, var)->iv) |
1048 | { | |
726a989a | 1049 | bb = gimple_bb (SSA_NAME_DEF_STMT (var)); |
8b11a64c ZD |
1050 | |
1051 | if (!bb | |
1052 | || !flow_bb_inside_loop_p (data->current_loop, bb)) | |
c70ed622 | 1053 | set_iv (data, var, var, build_int_cst (type, 0), true); |
8b11a64c ZD |
1054 | } |
1055 | ||
1056 | return name_info (data, var)->iv; | |
1057 | } | |
1058 | ||
fc06280e BC |
1059 | /* Return the first non-invariant ssa var found in EXPR. */ |
1060 | ||
1061 | static tree | |
1062 | extract_single_var_from_expr (tree expr) | |
1063 | { | |
1064 | int i, n; | |
1065 | tree tmp; | |
1066 | enum tree_code code; | |
1067 | ||
1068 | if (!expr || is_gimple_min_invariant (expr)) | |
1069 | return NULL; | |
1070 | ||
1071 | code = TREE_CODE (expr); | |
1072 | if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code))) | |
1073 | { | |
1074 | n = TREE_OPERAND_LENGTH (expr); | |
1075 | for (i = 0; i < n; i++) | |
1076 | { | |
1077 | tmp = extract_single_var_from_expr (TREE_OPERAND (expr, i)); | |
1078 | ||
1079 | if (tmp) | |
1080 | return tmp; | |
1081 | } | |
1082 | } | |
1083 | return (TREE_CODE (expr) == SSA_NAME) ? expr : NULL; | |
1084 | } | |
1085 | ||
8b11a64c ZD |
1086 | /* Finds basic ivs. */ |
1087 | ||
1088 | static bool | |
1089 | find_bivs (struct ivopts_data *data) | |
1090 | { | |
538dd0b7 | 1091 | gphi *phi; |
c70ed622 | 1092 | affine_iv iv; |
fc06280e | 1093 | tree step, type, base, stop; |
8b11a64c ZD |
1094 | bool found = false; |
1095 | struct loop *loop = data->current_loop; | |
538dd0b7 | 1096 | gphi_iterator psi; |
8b11a64c | 1097 | |
726a989a | 1098 | for (psi = gsi_start_phis (loop->header); !gsi_end_p (psi); gsi_next (&psi)) |
8b11a64c | 1099 | { |
538dd0b7 | 1100 | phi = psi.phi (); |
726a989a | 1101 | |
8b11a64c ZD |
1102 | if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (PHI_RESULT (phi))) |
1103 | continue; | |
1104 | ||
c70ed622 | 1105 | if (virtual_operand_p (PHI_RESULT (phi))) |
8b11a64c | 1106 | continue; |
8b11a64c | 1107 | |
c70ed622 BC |
1108 | if (!simple_iv (loop, loop, PHI_RESULT (phi), &iv, true)) |
1109 | continue; | |
1110 | ||
1111 | if (integer_zerop (iv.step)) | |
1112 | continue; | |
1113 | ||
1114 | step = iv.step; | |
8b11a64c | 1115 | base = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); |
fc06280e BC |
1116 | /* Stop expanding iv base at the first ssa var referred by iv step. |
1117 | Ideally we should stop at any ssa var, because that's expensive | |
1118 | and unusual to happen, we just do it on the first one. | |
1119 | ||
1120 | See PR64705 for the rationale. */ | |
1121 | stop = extract_single_var_from_expr (step); | |
1122 | base = expand_simple_operations (base, stop); | |
9be872b7 ZD |
1123 | if (contains_abnormal_ssa_name_p (base) |
1124 | || contains_abnormal_ssa_name_p (step)) | |
8b11a64c ZD |
1125 | continue; |
1126 | ||
1127 | type = TREE_TYPE (PHI_RESULT (phi)); | |
1128 | base = fold_convert (type, base); | |
9be872b7 | 1129 | if (step) |
1ffe34d9 AP |
1130 | { |
1131 | if (POINTER_TYPE_P (type)) | |
0d82a1c8 | 1132 | step = convert_to_ptrofftype (step); |
1ffe34d9 AP |
1133 | else |
1134 | step = fold_convert (type, step); | |
1135 | } | |
8b11a64c | 1136 | |
c70ed622 | 1137 | set_iv (data, PHI_RESULT (phi), base, step, iv.no_overflow); |
8b11a64c ZD |
1138 | found = true; |
1139 | } | |
1140 | ||
1141 | return found; | |
1142 | } | |
1143 | ||
1144 | /* Marks basic ivs. */ | |
1145 | ||
1146 | static void | |
1147 | mark_bivs (struct ivopts_data *data) | |
1148 | { | |
538dd0b7 DM |
1149 | gphi *phi; |
1150 | gimple def; | |
726a989a | 1151 | tree var; |
8b11a64c ZD |
1152 | struct iv *iv, *incr_iv; |
1153 | struct loop *loop = data->current_loop; | |
1154 | basic_block incr_bb; | |
538dd0b7 | 1155 | gphi_iterator psi; |
8b11a64c | 1156 | |
726a989a | 1157 | for (psi = gsi_start_phis (loop->header); !gsi_end_p (psi); gsi_next (&psi)) |
8b11a64c | 1158 | { |
538dd0b7 | 1159 | phi = psi.phi (); |
726a989a | 1160 | |
8b11a64c ZD |
1161 | iv = get_iv (data, PHI_RESULT (phi)); |
1162 | if (!iv) | |
1163 | continue; | |
1164 | ||
1165 | var = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); | |
b83b5507 BC |
1166 | def = SSA_NAME_DEF_STMT (var); |
1167 | /* Don't mark iv peeled from other one as biv. */ | |
1168 | if (def | |
1169 | && gimple_code (def) == GIMPLE_PHI | |
1170 | && gimple_bb (def) == loop->header) | |
1171 | continue; | |
1172 | ||
8b11a64c ZD |
1173 | incr_iv = get_iv (data, var); |
1174 | if (!incr_iv) | |
1175 | continue; | |
1176 | ||
1177 | /* If the increment is in the subloop, ignore it. */ | |
726a989a | 1178 | incr_bb = gimple_bb (SSA_NAME_DEF_STMT (var)); |
8b11a64c ZD |
1179 | if (incr_bb->loop_father != data->current_loop |
1180 | || (incr_bb->flags & BB_IRREDUCIBLE_LOOP)) | |
1181 | continue; | |
1182 | ||
1183 | iv->biv_p = true; | |
1184 | incr_iv->biv_p = true; | |
1185 | } | |
1186 | } | |
1187 | ||
1188 | /* Checks whether STMT defines a linear induction variable and stores its | |
a6f778b2 | 1189 | parameters to IV. */ |
8b11a64c ZD |
1190 | |
1191 | static bool | |
726a989a | 1192 | find_givs_in_stmt_scev (struct ivopts_data *data, gimple stmt, affine_iv *iv) |
8b11a64c | 1193 | { |
fc06280e | 1194 | tree lhs, stop; |
8b11a64c ZD |
1195 | struct loop *loop = data->current_loop; |
1196 | ||
a6f778b2 ZD |
1197 | iv->base = NULL_TREE; |
1198 | iv->step = NULL_TREE; | |
8b11a64c | 1199 | |
726a989a | 1200 | if (gimple_code (stmt) != GIMPLE_ASSIGN) |
8b11a64c ZD |
1201 | return false; |
1202 | ||
726a989a | 1203 | lhs = gimple_assign_lhs (stmt); |
8b11a64c ZD |
1204 | if (TREE_CODE (lhs) != SSA_NAME) |
1205 | return false; | |
1206 | ||
f017bf5e | 1207 | if (!simple_iv (loop, loop_containing_stmt (stmt), lhs, iv, true)) |
8b11a64c ZD |
1208 | return false; |
1209 | ||
fc06280e BC |
1210 | /* Stop expanding iv base at the first ssa var referred by iv step. |
1211 | Ideally we should stop at any ssa var, because that's expensive | |
1212 | and unusual to happen, we just do it on the first one. | |
1213 | ||
1214 | See PR64705 for the rationale. */ | |
1215 | stop = extract_single_var_from_expr (iv->step); | |
1216 | iv->base = expand_simple_operations (iv->base, stop); | |
a6f778b2 ZD |
1217 | if (contains_abnormal_ssa_name_p (iv->base) |
1218 | || contains_abnormal_ssa_name_p (iv->step)) | |
8b11a64c ZD |
1219 | return false; |
1220 | ||
fc06280e | 1221 | /* If STMT could throw, then do not consider STMT as defining a GIV. |
9f9ca914 JL |
1222 | While this will suppress optimizations, we can not safely delete this |
1223 | GIV and associated statements, even if it appears it is not used. */ | |
1224 | if (stmt_could_throw_p (stmt)) | |
1225 | return false; | |
1226 | ||
8b11a64c ZD |
1227 | return true; |
1228 | } | |
1229 | ||
1230 | /* Finds general ivs in statement STMT. */ | |
1231 | ||
1232 | static void | |
726a989a | 1233 | find_givs_in_stmt (struct ivopts_data *data, gimple stmt) |
8b11a64c | 1234 | { |
a6f778b2 | 1235 | affine_iv iv; |
8b11a64c | 1236 | |
a6f778b2 | 1237 | if (!find_givs_in_stmt_scev (data, stmt, &iv)) |
8b11a64c ZD |
1238 | return; |
1239 | ||
c70ed622 | 1240 | set_iv (data, gimple_assign_lhs (stmt), iv.base, iv.step, iv.no_overflow); |
8b11a64c ZD |
1241 | } |
1242 | ||
1243 | /* Finds general ivs in basic block BB. */ | |
1244 | ||
1245 | static void | |
1246 | find_givs_in_bb (struct ivopts_data *data, basic_block bb) | |
1247 | { | |
726a989a | 1248 | gimple_stmt_iterator bsi; |
8b11a64c | 1249 | |
726a989a RB |
1250 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
1251 | find_givs_in_stmt (data, gsi_stmt (bsi)); | |
8b11a64c ZD |
1252 | } |
1253 | ||
1254 | /* Finds general ivs. */ | |
1255 | ||
1256 | static void | |
1257 | find_givs (struct ivopts_data *data) | |
1258 | { | |
1259 | struct loop *loop = data->current_loop; | |
1260 | basic_block *body = get_loop_body_in_dom_order (loop); | |
1261 | unsigned i; | |
1262 | ||
1263 | for (i = 0; i < loop->num_nodes; i++) | |
1264 | find_givs_in_bb (data, body[i]); | |
1265 | free (body); | |
1266 | } | |
1267 | ||
8b11a64c ZD |
1268 | /* For each ssa name defined in LOOP determines whether it is an induction |
1269 | variable and if so, its initial value and step. */ | |
1270 | ||
1271 | static bool | |
1272 | find_induction_variables (struct ivopts_data *data) | |
1273 | { | |
1274 | unsigned i; | |
87c476a2 | 1275 | bitmap_iterator bi; |
8b11a64c ZD |
1276 | |
1277 | if (!find_bivs (data)) | |
1278 | return false; | |
1279 | ||
1280 | find_givs (data); | |
1281 | mark_bivs (data); | |
8b11a64c ZD |
1282 | |
1283 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1284 | { | |
d8af4ba3 | 1285 | struct tree_niter_desc *niter = niter_for_single_dom_exit (data); |
ca4c3169 ZD |
1286 | |
1287 | if (niter) | |
8b11a64c ZD |
1288 | { |
1289 | fprintf (dump_file, " number of iterations "); | |
d8af4ba3 ZD |
1290 | print_generic_expr (dump_file, niter->niter, TDF_SLIM); |
1291 | if (!integer_zerop (niter->may_be_zero)) | |
1292 | { | |
1293 | fprintf (dump_file, "; zero if "); | |
1294 | print_generic_expr (dump_file, niter->may_be_zero, TDF_SLIM); | |
1295 | } | |
dcccd88d | 1296 | fprintf (dump_file, "\n\n"); |
8b11a64c | 1297 | }; |
b8698a0f | 1298 | |
8b11a64c ZD |
1299 | fprintf (dump_file, "Induction variables:\n\n"); |
1300 | ||
87c476a2 | 1301 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, i, bi) |
8b11a64c ZD |
1302 | { |
1303 | if (ver_info (data, i)->iv) | |
e185f450 | 1304 | dump_iv (dump_file, ver_info (data, i)->iv, true); |
87c476a2 | 1305 | } |
8b11a64c ZD |
1306 | } |
1307 | ||
1308 | return true; | |
1309 | } | |
1310 | ||
a7e43c57 BC |
1311 | /* Records a use of type USE_TYPE at *USE_P in STMT whose value is IV. |
1312 | For address type use, ADDR_BASE is the stripped IV base, ADDR_OFFSET | |
1313 | is the const offset stripped from IV base. For uses of other types, | |
1314 | ADDR_BASE and ADDR_OFFSET are zero by default. */ | |
8b11a64c ZD |
1315 | |
1316 | static struct iv_use * | |
1317 | record_use (struct ivopts_data *data, tree *use_p, struct iv *iv, | |
a7e43c57 BC |
1318 | gimple stmt, enum use_type use_type, tree addr_base = NULL, |
1319 | unsigned HOST_WIDE_INT addr_offset = 0) | |
8b11a64c | 1320 | { |
5ed6ace5 | 1321 | struct iv_use *use = XCNEW (struct iv_use); |
8b11a64c ZD |
1322 | |
1323 | use->id = n_iv_uses (data); | |
a7e43c57 | 1324 | use->sub_id = 0; |
8b11a64c ZD |
1325 | use->type = use_type; |
1326 | use->iv = iv; | |
1327 | use->stmt = stmt; | |
1328 | use->op_p = use_p; | |
8bdbfff5 | 1329 | use->related_cands = BITMAP_ALLOC (NULL); |
a7e43c57 BC |
1330 | use->next = NULL; |
1331 | use->addr_base = addr_base; | |
1332 | use->addr_offset = addr_offset; | |
8b11a64c | 1333 | |
9771b263 | 1334 | data->iv_uses.safe_push (use); |
8b11a64c ZD |
1335 | |
1336 | return use; | |
1337 | } | |
1338 | ||
a7e43c57 BC |
1339 | /* Records a sub use of type USE_TYPE at *USE_P in STMT whose value is IV. |
1340 | The sub use is recorded under the one whose use id is ID_GROUP. */ | |
1341 | ||
1342 | static struct iv_use * | |
1343 | record_sub_use (struct ivopts_data *data, tree *use_p, | |
1344 | struct iv *iv, gimple stmt, enum use_type use_type, | |
1345 | tree addr_base, unsigned HOST_WIDE_INT addr_offset, | |
1346 | unsigned int id_group) | |
1347 | { | |
1348 | struct iv_use *use = XCNEW (struct iv_use); | |
1349 | struct iv_use *group = iv_use (data, id_group); | |
1350 | ||
1351 | use->id = group->id; | |
1352 | use->sub_id = 0; | |
1353 | use->type = use_type; | |
1354 | use->iv = iv; | |
1355 | use->stmt = stmt; | |
1356 | use->op_p = use_p; | |
1357 | use->related_cands = NULL; | |
1358 | use->addr_base = addr_base; | |
1359 | use->addr_offset = addr_offset; | |
1360 | ||
1361 | /* Sub use list is maintained in offset ascending order. */ | |
1362 | if (addr_offset <= group->addr_offset) | |
1363 | { | |
1364 | use->related_cands = group->related_cands; | |
1365 | group->related_cands = NULL; | |
1366 | use->next = group; | |
1367 | data->iv_uses[id_group] = use; | |
1368 | } | |
1369 | else | |
1370 | { | |
1371 | struct iv_use *pre; | |
1372 | do | |
1373 | { | |
1374 | pre = group; | |
1375 | group = group->next; | |
1376 | } | |
1377 | while (group && addr_offset > group->addr_offset); | |
1378 | use->next = pre->next; | |
1379 | pre->next = use; | |
1380 | } | |
1381 | ||
a7e43c57 BC |
1382 | return use; |
1383 | } | |
1384 | ||
8b11a64c ZD |
1385 | /* Checks whether OP is a loop-level invariant and if so, records it. |
1386 | NONLINEAR_USE is true if the invariant is used in a way we do not | |
1387 | handle specially. */ | |
1388 | ||
1389 | static void | |
1390 | record_invariant (struct ivopts_data *data, tree op, bool nonlinear_use) | |
1391 | { | |
1392 | basic_block bb; | |
1393 | struct version_info *info; | |
1394 | ||
1395 | if (TREE_CODE (op) != SSA_NAME | |
ea057359 | 1396 | || virtual_operand_p (op)) |
8b11a64c ZD |
1397 | return; |
1398 | ||
726a989a | 1399 | bb = gimple_bb (SSA_NAME_DEF_STMT (op)); |
8b11a64c ZD |
1400 | if (bb |
1401 | && flow_bb_inside_loop_p (data->current_loop, bb)) | |
1402 | return; | |
1403 | ||
1404 | info = name_info (data, op); | |
1405 | info->name = op; | |
1406 | info->has_nonlin_use |= nonlinear_use; | |
1407 | if (!info->inv_id) | |
1408 | info->inv_id = ++data->max_inv_id; | |
1409 | bitmap_set_bit (data->relevant, SSA_NAME_VERSION (op)); | |
1410 | } | |
1411 | ||
50cc9802 | 1412 | /* Checks whether the use OP is interesting and if so, records it. */ |
8b11a64c ZD |
1413 | |
1414 | static struct iv_use * | |
50cc9802 | 1415 | find_interesting_uses_op (struct ivopts_data *data, tree op) |
8b11a64c ZD |
1416 | { |
1417 | struct iv *iv; | |
726a989a | 1418 | gimple stmt; |
8b11a64c ZD |
1419 | struct iv_use *use; |
1420 | ||
1421 | if (TREE_CODE (op) != SSA_NAME) | |
1422 | return NULL; | |
1423 | ||
1424 | iv = get_iv (data, op); | |
1425 | if (!iv) | |
1426 | return NULL; | |
b8698a0f | 1427 | |
8b11a64c ZD |
1428 | if (iv->have_use_for) |
1429 | { | |
1430 | use = iv_use (data, iv->use_id); | |
1431 | ||
50cc9802 | 1432 | gcc_assert (use->type == USE_NONLINEAR_EXPR); |
8b11a64c ZD |
1433 | return use; |
1434 | } | |
1435 | ||
6e42ce54 | 1436 | if (integer_zerop (iv->step)) |
8b11a64c ZD |
1437 | { |
1438 | record_invariant (data, op, true); | |
1439 | return NULL; | |
1440 | } | |
1441 | iv->have_use_for = true; | |
1442 | ||
8b11a64c | 1443 | stmt = SSA_NAME_DEF_STMT (op); |
726a989a RB |
1444 | gcc_assert (gimple_code (stmt) == GIMPLE_PHI |
1445 | || is_gimple_assign (stmt)); | |
8b11a64c | 1446 | |
6f929985 | 1447 | use = record_use (data, NULL, iv, stmt, USE_NONLINEAR_EXPR); |
8b11a64c ZD |
1448 | iv->use_id = use->id; |
1449 | ||
1450 | return use; | |
1451 | } | |
1452 | ||
726a989a RB |
1453 | /* Given a condition in statement STMT, checks whether it is a compare |
1454 | of an induction variable and an invariant. If this is the case, | |
1455 | CONTROL_VAR is set to location of the iv, BOUND to the location of | |
1456 | the invariant, IV_VAR and IV_BOUND are set to the corresponding | |
1457 | induction variable descriptions, and true is returned. If this is not | |
1458 | the case, CONTROL_VAR and BOUND are set to the arguments of the | |
1459 | condition and false is returned. */ | |
8b11a64c | 1460 | |
b697aed4 | 1461 | static bool |
726a989a | 1462 | extract_cond_operands (struct ivopts_data *data, gimple stmt, |
b697aed4 ZD |
1463 | tree **control_var, tree **bound, |
1464 | struct iv **iv_var, struct iv **iv_bound) | |
1465 | { | |
726a989a | 1466 | /* The objects returned when COND has constant operands. */ |
b697aed4 ZD |
1467 | static struct iv const_iv; |
1468 | static tree zero; | |
6b4db501 MM |
1469 | tree *op0 = &zero, *op1 = &zero; |
1470 | struct iv *iv0 = &const_iv, *iv1 = &const_iv; | |
b697aed4 ZD |
1471 | bool ret = false; |
1472 | ||
726a989a | 1473 | if (gimple_code (stmt) == GIMPLE_COND) |
8b11a64c | 1474 | { |
538dd0b7 DM |
1475 | gcond *cond_stmt = as_a <gcond *> (stmt); |
1476 | op0 = gimple_cond_lhs_ptr (cond_stmt); | |
1477 | op1 = gimple_cond_rhs_ptr (cond_stmt); | |
8b11a64c | 1478 | } |
726a989a | 1479 | else |
8b11a64c | 1480 | { |
726a989a RB |
1481 | op0 = gimple_assign_rhs1_ptr (stmt); |
1482 | op1 = gimple_assign_rhs2_ptr (stmt); | |
8b11a64c ZD |
1483 | } |
1484 | ||
726a989a RB |
1485 | zero = integer_zero_node; |
1486 | const_iv.step = integer_zero_node; | |
1487 | ||
b697aed4 ZD |
1488 | if (TREE_CODE (*op0) == SSA_NAME) |
1489 | iv0 = get_iv (data, *op0); | |
1490 | if (TREE_CODE (*op1) == SSA_NAME) | |
1491 | iv1 = get_iv (data, *op1); | |
8b11a64c | 1492 | |
b697aed4 ZD |
1493 | /* Exactly one of the compared values must be an iv, and the other one must |
1494 | be an invariant. */ | |
1495 | if (!iv0 || !iv1) | |
1496 | goto end; | |
1497 | ||
1498 | if (integer_zerop (iv0->step)) | |
1499 | { | |
1500 | /* Control variable may be on the other side. */ | |
6b4db501 MM |
1501 | std::swap (op0, op1); |
1502 | std::swap (iv0, iv1); | |
8b11a64c | 1503 | } |
b697aed4 ZD |
1504 | ret = !integer_zerop (iv0->step) && integer_zerop (iv1->step); |
1505 | ||
1506 | end: | |
1507 | if (control_var) | |
6f3d1a5e | 1508 | *control_var = op0; |
b697aed4 | 1509 | if (iv_var) |
6f3d1a5e | 1510 | *iv_var = iv0; |
b697aed4 ZD |
1511 | if (bound) |
1512 | *bound = op1; | |
1513 | if (iv_bound) | |
1514 | *iv_bound = iv1; | |
1515 | ||
1516 | return ret; | |
1517 | } | |
1518 | ||
726a989a RB |
1519 | /* Checks whether the condition in STMT is interesting and if so, |
1520 | records it. */ | |
b697aed4 ZD |
1521 | |
1522 | static void | |
726a989a | 1523 | find_interesting_uses_cond (struct ivopts_data *data, gimple stmt) |
b697aed4 ZD |
1524 | { |
1525 | tree *var_p, *bound_p; | |
6f929985 | 1526 | struct iv *var_iv; |
8b11a64c | 1527 | |
726a989a | 1528 | if (!extract_cond_operands (data, stmt, &var_p, &bound_p, &var_iv, NULL)) |
8b11a64c | 1529 | { |
b697aed4 ZD |
1530 | find_interesting_uses_op (data, *var_p); |
1531 | find_interesting_uses_op (data, *bound_p); | |
8b11a64c ZD |
1532 | return; |
1533 | } | |
1534 | ||
6f929985 | 1535 | record_use (data, NULL, var_iv, stmt, USE_COMPARE); |
8b11a64c ZD |
1536 | } |
1537 | ||
4ba5ea11 RB |
1538 | /* Returns the outermost loop EXPR is obviously invariant in |
1539 | relative to the loop LOOP, i.e. if all its operands are defined | |
1540 | outside of the returned loop. Returns NULL if EXPR is not | |
1541 | even obviously invariant in LOOP. */ | |
1542 | ||
1543 | struct loop * | |
1544 | outermost_invariant_loop_for_expr (struct loop *loop, tree expr) | |
1545 | { | |
1546 | basic_block def_bb; | |
1547 | unsigned i, len; | |
1548 | ||
1549 | if (is_gimple_min_invariant (expr)) | |
1550 | return current_loops->tree_root; | |
1551 | ||
1552 | if (TREE_CODE (expr) == SSA_NAME) | |
1553 | { | |
1554 | def_bb = gimple_bb (SSA_NAME_DEF_STMT (expr)); | |
1555 | if (def_bb) | |
1556 | { | |
1557 | if (flow_bb_inside_loop_p (loop, def_bb)) | |
1558 | return NULL; | |
1559 | return superloop_at_depth (loop, | |
1560 | loop_depth (def_bb->loop_father) + 1); | |
1561 | } | |
1562 | ||
1563 | return current_loops->tree_root; | |
1564 | } | |
1565 | ||
1566 | if (!EXPR_P (expr)) | |
1567 | return NULL; | |
1568 | ||
1569 | unsigned maxdepth = 0; | |
1570 | len = TREE_OPERAND_LENGTH (expr); | |
1571 | for (i = 0; i < len; i++) | |
1572 | { | |
1573 | struct loop *ivloop; | |
1574 | if (!TREE_OPERAND (expr, i)) | |
1575 | continue; | |
1576 | ||
1577 | ivloop = outermost_invariant_loop_for_expr (loop, TREE_OPERAND (expr, i)); | |
1578 | if (!ivloop) | |
1579 | return NULL; | |
1580 | maxdepth = MAX (maxdepth, loop_depth (ivloop)); | |
1581 | } | |
1582 | ||
1583 | return superloop_at_depth (loop, maxdepth); | |
1584 | } | |
1585 | ||
be35cf60 | 1586 | /* Returns true if expression EXPR is obviously invariant in LOOP, |
6a732743 SP |
1587 | i.e. if all its operands are defined outside of the LOOP. LOOP |
1588 | should not be the function body. */ | |
be35cf60 | 1589 | |
feb075f4 | 1590 | bool |
be35cf60 ZD |
1591 | expr_invariant_in_loop_p (struct loop *loop, tree expr) |
1592 | { | |
1593 | basic_block def_bb; | |
1594 | unsigned i, len; | |
1595 | ||
6a732743 SP |
1596 | gcc_assert (loop_depth (loop) > 0); |
1597 | ||
be35cf60 ZD |
1598 | if (is_gimple_min_invariant (expr)) |
1599 | return true; | |
1600 | ||
1601 | if (TREE_CODE (expr) == SSA_NAME) | |
1602 | { | |
726a989a | 1603 | def_bb = gimple_bb (SSA_NAME_DEF_STMT (expr)); |
be35cf60 ZD |
1604 | if (def_bb |
1605 | && flow_bb_inside_loop_p (loop, def_bb)) | |
1606 | return false; | |
1607 | ||
1608 | return true; | |
1609 | } | |
1610 | ||
726a989a | 1611 | if (!EXPR_P (expr)) |
be35cf60 ZD |
1612 | return false; |
1613 | ||
5039610b | 1614 | len = TREE_OPERAND_LENGTH (expr); |
be35cf60 | 1615 | for (i = 0; i < len; i++) |
837a549b JH |
1616 | if (TREE_OPERAND (expr, i) |
1617 | && !expr_invariant_in_loop_p (loop, TREE_OPERAND (expr, i))) | |
be35cf60 ZD |
1618 | return false; |
1619 | ||
1620 | return true; | |
1621 | } | |
1622 | ||
8b11a64c ZD |
1623 | /* Cumulates the steps of indices into DATA and replaces their values with the |
1624 | initial ones. Returns false when the value of the index cannot be determined. | |
1625 | Callback for for_each_index. */ | |
1626 | ||
1627 | struct ifs_ivopts_data | |
1628 | { | |
1629 | struct ivopts_data *ivopts_data; | |
726a989a | 1630 | gimple stmt; |
6e42ce54 | 1631 | tree step; |
8b11a64c ZD |
1632 | }; |
1633 | ||
1634 | static bool | |
1635 | idx_find_step (tree base, tree *idx, void *data) | |
1636 | { | |
c22940cd | 1637 | struct ifs_ivopts_data *dta = (struct ifs_ivopts_data *) data; |
8b11a64c | 1638 | struct iv *iv; |
c70ed622 | 1639 | bool use_overflow_semantics = false; |
d7f5de76 | 1640 | tree step, iv_base, iv_step, lbound, off; |
2f4675b4 | 1641 | struct loop *loop = dta->ivopts_data->current_loop; |
be35cf60 | 1642 | |
be35cf60 | 1643 | /* If base is a component ref, require that the offset of the reference |
3a7c155d | 1644 | be invariant. */ |
be35cf60 ZD |
1645 | if (TREE_CODE (base) == COMPONENT_REF) |
1646 | { | |
1647 | off = component_ref_field_offset (base); | |
1648 | return expr_invariant_in_loop_p (loop, off); | |
1649 | } | |
1650 | ||
1651 | /* If base is array, first check whether we will be able to move the | |
1652 | reference out of the loop (in order to take its address in strength | |
1653 | reduction). In order for this to work we need both lower bound | |
1654 | and step to be loop invariants. */ | |
9f7ccf69 | 1655 | if (TREE_CODE (base) == ARRAY_REF || TREE_CODE (base) == ARRAY_RANGE_REF) |
be35cf60 | 1656 | { |
9f7ccf69 EB |
1657 | /* Moreover, for a range, the size needs to be invariant as well. */ |
1658 | if (TREE_CODE (base) == ARRAY_RANGE_REF | |
1659 | && !expr_invariant_in_loop_p (loop, TYPE_SIZE (TREE_TYPE (base)))) | |
1660 | return false; | |
1661 | ||
be35cf60 ZD |
1662 | step = array_ref_element_size (base); |
1663 | lbound = array_ref_low_bound (base); | |
1664 | ||
1665 | if (!expr_invariant_in_loop_p (loop, step) | |
1666 | || !expr_invariant_in_loop_p (loop, lbound)) | |
1667 | return false; | |
1668 | } | |
1669 | ||
8b11a64c ZD |
1670 | if (TREE_CODE (*idx) != SSA_NAME) |
1671 | return true; | |
1672 | ||
1673 | iv = get_iv (dta->ivopts_data, *idx); | |
1674 | if (!iv) | |
1675 | return false; | |
1676 | ||
ea643120 RG |
1677 | /* XXX We produce for a base of *D42 with iv->base being &x[0] |
1678 | *&x[0], which is not folded and does not trigger the | |
1679 | ARRAY_REF path below. */ | |
8b11a64c ZD |
1680 | *idx = iv->base; |
1681 | ||
6e42ce54 | 1682 | if (integer_zerop (iv->step)) |
8b11a64c ZD |
1683 | return true; |
1684 | ||
9f7ccf69 | 1685 | if (TREE_CODE (base) == ARRAY_REF || TREE_CODE (base) == ARRAY_RANGE_REF) |
2f4675b4 ZD |
1686 | { |
1687 | step = array_ref_element_size (base); | |
2f4675b4 ZD |
1688 | |
1689 | /* We only handle addresses whose step is an integer constant. */ | |
1690 | if (TREE_CODE (step) != INTEGER_CST) | |
1691 | return false; | |
2f4675b4 | 1692 | } |
8b11a64c | 1693 | else |
5212068f | 1694 | /* The step for pointer arithmetics already is 1 byte. */ |
9a9d280e | 1695 | step = size_one_node; |
8b11a64c | 1696 | |
d7f5de76 ZD |
1697 | iv_base = iv->base; |
1698 | iv_step = iv->step; | |
c70ed622 BC |
1699 | if (iv->no_overflow && nowrap_type_p (TREE_TYPE (iv_step))) |
1700 | use_overflow_semantics = true; | |
1701 | ||
d7f5de76 ZD |
1702 | if (!convert_affine_scev (dta->ivopts_data->current_loop, |
1703 | sizetype, &iv_base, &iv_step, dta->stmt, | |
c70ed622 | 1704 | use_overflow_semantics)) |
8b11a64c ZD |
1705 | { |
1706 | /* The index might wrap. */ | |
1707 | return false; | |
1708 | } | |
1709 | ||
1c1205fb | 1710 | step = fold_build2 (MULT_EXPR, sizetype, step, iv_step); |
6e42ce54 | 1711 | dta->step = fold_build2 (PLUS_EXPR, sizetype, dta->step, step); |
8b11a64c ZD |
1712 | |
1713 | return true; | |
1714 | } | |
1715 | ||
1716 | /* Records use in index IDX. Callback for for_each_index. Ivopts data | |
1717 | object is passed to it in DATA. */ | |
1718 | ||
1719 | static bool | |
2f4675b4 | 1720 | idx_record_use (tree base, tree *idx, |
c22940cd | 1721 | void *vdata) |
8b11a64c | 1722 | { |
c22940cd | 1723 | struct ivopts_data *data = (struct ivopts_data *) vdata; |
8b11a64c | 1724 | find_interesting_uses_op (data, *idx); |
9f7ccf69 | 1725 | if (TREE_CODE (base) == ARRAY_REF || TREE_CODE (base) == ARRAY_RANGE_REF) |
2f4675b4 ZD |
1726 | { |
1727 | find_interesting_uses_op (data, array_ref_element_size (base)); | |
1728 | find_interesting_uses_op (data, array_ref_low_bound (base)); | |
1729 | } | |
8b11a64c ZD |
1730 | return true; |
1731 | } | |
1732 | ||
32159434 CB |
1733 | /* If we can prove that TOP = cst * BOT for some constant cst, |
1734 | store cst to MUL and return true. Otherwise return false. | |
1735 | The returned value is always sign-extended, regardless of the | |
1736 | signedness of TOP and BOT. */ | |
1737 | ||
1738 | static bool | |
807e902e | 1739 | constant_multiple_of (tree top, tree bot, widest_int *mul) |
32159434 CB |
1740 | { |
1741 | tree mby; | |
1742 | enum tree_code code; | |
32159434 | 1743 | unsigned precision = TYPE_PRECISION (TREE_TYPE (top)); |
807e902e | 1744 | widest_int res, p0, p1; |
32159434 CB |
1745 | |
1746 | STRIP_NOPS (top); | |
1747 | STRIP_NOPS (bot); | |
1748 | ||
1749 | if (operand_equal_p (top, bot, 0)) | |
1750 | { | |
807e902e | 1751 | *mul = 1; |
32159434 CB |
1752 | return true; |
1753 | } | |
1754 | ||
1755 | code = TREE_CODE (top); | |
1756 | switch (code) | |
1757 | { | |
1758 | case MULT_EXPR: | |
1759 | mby = TREE_OPERAND (top, 1); | |
1760 | if (TREE_CODE (mby) != INTEGER_CST) | |
1761 | return false; | |
1762 | ||
1763 | if (!constant_multiple_of (TREE_OPERAND (top, 0), bot, &res)) | |
1764 | return false; | |
1765 | ||
807e902e | 1766 | *mul = wi::sext (res * wi::to_widest (mby), precision); |
32159434 CB |
1767 | return true; |
1768 | ||
1769 | case PLUS_EXPR: | |
1770 | case MINUS_EXPR: | |
1771 | if (!constant_multiple_of (TREE_OPERAND (top, 0), bot, &p0) | |
1772 | || !constant_multiple_of (TREE_OPERAND (top, 1), bot, &p1)) | |
1773 | return false; | |
1774 | ||
1775 | if (code == MINUS_EXPR) | |
27bcd47c | 1776 | p1 = -p1; |
807e902e | 1777 | *mul = wi::sext (p0 + p1, precision); |
32159434 CB |
1778 | return true; |
1779 | ||
1780 | case INTEGER_CST: | |
1781 | if (TREE_CODE (bot) != INTEGER_CST) | |
1782 | return false; | |
1783 | ||
807e902e KZ |
1784 | p0 = widest_int::from (top, SIGNED); |
1785 | p1 = widest_int::from (bot, SIGNED); | |
1786 | if (p1 == 0) | |
32159434 | 1787 | return false; |
807e902e KZ |
1788 | *mul = wi::sext (wi::divmod_trunc (p0, p1, SIGNED, &res), precision); |
1789 | return res == 0; | |
32159434 CB |
1790 | |
1791 | default: | |
1792 | return false; | |
1793 | } | |
1794 | } | |
1795 | ||
e75fde1a | 1796 | /* Return true if memory reference REF with step STEP may be unaligned. */ |
0a915e3d ZD |
1797 | |
1798 | static bool | |
32159434 | 1799 | may_be_unaligned_p (tree ref, tree step) |
0a915e3d | 1800 | { |
ac182688 | 1801 | /* TARGET_MEM_REFs are translated directly to valid MEMs on the target, |
607fb860 | 1802 | thus they are not misaligned. */ |
ac182688 ZD |
1803 | if (TREE_CODE (ref) == TARGET_MEM_REF) |
1804 | return false; | |
1805 | ||
e75fde1a | 1806 | unsigned int align = TYPE_ALIGN (TREE_TYPE (ref)); |
abf30454 RB |
1807 | if (GET_MODE_ALIGNMENT (TYPE_MODE (TREE_TYPE (ref))) > align) |
1808 | align = GET_MODE_ALIGNMENT (TYPE_MODE (TREE_TYPE (ref))); | |
b8698a0f | 1809 | |
e75fde1a EB |
1810 | unsigned HOST_WIDE_INT bitpos; |
1811 | unsigned int ref_align; | |
1812 | get_object_alignment_1 (ref, &ref_align, &bitpos); | |
1813 | if (ref_align < align | |
1814 | || (bitpos % align) != 0 | |
1815 | || (bitpos % BITS_PER_UNIT) != 0) | |
1816 | return true; | |
ce276b61 | 1817 | |
e75fde1a EB |
1818 | unsigned int trailing_zeros = tree_ctz (step); |
1819 | if (trailing_zeros < HOST_BITS_PER_INT | |
1820 | && (1U << trailing_zeros) * BITS_PER_UNIT < align) | |
1821 | return true; | |
0a915e3d ZD |
1822 | |
1823 | return false; | |
1824 | } | |
1825 | ||
75715cf6 EB |
1826 | /* Return true if EXPR may be non-addressable. */ |
1827 | ||
bc068a23 | 1828 | bool |
75715cf6 EB |
1829 | may_be_nonaddressable_p (tree expr) |
1830 | { | |
1831 | switch (TREE_CODE (expr)) | |
1832 | { | |
928bc34f EB |
1833 | case TARGET_MEM_REF: |
1834 | /* TARGET_MEM_REFs are translated directly to valid MEMs on the | |
1835 | target, thus they are always addressable. */ | |
1836 | return false; | |
1837 | ||
75715cf6 EB |
1838 | case COMPONENT_REF: |
1839 | return DECL_NONADDRESSABLE_P (TREE_OPERAND (expr, 1)) | |
1840 | || may_be_nonaddressable_p (TREE_OPERAND (expr, 0)); | |
1841 | ||
75715cf6 EB |
1842 | case VIEW_CONVERT_EXPR: |
1843 | /* This kind of view-conversions may wrap non-addressable objects | |
1844 | and make them look addressable. After some processing the | |
1845 | non-addressability may be uncovered again, causing ADDR_EXPRs | |
1846 | of inappropriate objects to be built. */ | |
7a4fbffc EB |
1847 | if (is_gimple_reg (TREE_OPERAND (expr, 0)) |
1848 | || !is_gimple_addressable (TREE_OPERAND (expr, 0))) | |
1849 | return true; | |
1850 | ||
1851 | /* ... fall through ... */ | |
928bc34f EB |
1852 | |
1853 | case ARRAY_REF: | |
1854 | case ARRAY_RANGE_REF: | |
7a4fbffc | 1855 | return may_be_nonaddressable_p (TREE_OPERAND (expr, 0)); |
928bc34f | 1856 | |
1043771b | 1857 | CASE_CONVERT: |
928bc34f | 1858 | return true; |
75715cf6 EB |
1859 | |
1860 | default: | |
1861 | break; | |
1862 | } | |
1863 | ||
1864 | return false; | |
1865 | } | |
1866 | ||
a7e43c57 BC |
1867 | static tree |
1868 | strip_offset (tree expr, unsigned HOST_WIDE_INT *offset); | |
1869 | ||
1870 | /* Record a use of type USE_TYPE at *USE_P in STMT whose value is IV. | |
1871 | If there is an existing use which has same stripped iv base and step, | |
1872 | this function records this one as a sub use to that; otherwise records | |
1873 | it as a normal one. */ | |
1874 | ||
1875 | static struct iv_use * | |
1876 | record_group_use (struct ivopts_data *data, tree *use_p, | |
1877 | struct iv *iv, gimple stmt, enum use_type use_type) | |
1878 | { | |
1879 | unsigned int i; | |
1880 | struct iv_use *use; | |
1881 | tree addr_base; | |
1882 | unsigned HOST_WIDE_INT addr_offset; | |
1883 | ||
1884 | /* Only support sub use for address type uses, that is, with base | |
1885 | object. */ | |
1886 | if (!iv->base_object) | |
1887 | return record_use (data, use_p, iv, stmt, use_type); | |
1888 | ||
1889 | addr_base = strip_offset (iv->base, &addr_offset); | |
1890 | for (i = 0; i < n_iv_uses (data); i++) | |
1891 | { | |
1892 | use = iv_use (data, i); | |
1893 | if (use->type != USE_ADDRESS || !use->iv->base_object) | |
1894 | continue; | |
1895 | ||
1896 | /* Check if it has the same stripped base and step. */ | |
1897 | if (operand_equal_p (iv->base_object, use->iv->base_object, 0) | |
1898 | && operand_equal_p (iv->step, use->iv->step, 0) | |
1899 | && operand_equal_p (addr_base, use->addr_base, 0)) | |
1900 | break; | |
1901 | } | |
1902 | ||
1903 | if (i == n_iv_uses (data)) | |
1904 | return record_use (data, use_p, iv, stmt, | |
1905 | use_type, addr_base, addr_offset); | |
1906 | else | |
1907 | return record_sub_use (data, use_p, iv, stmt, | |
1908 | use_type, addr_base, addr_offset, i); | |
1909 | } | |
1910 | ||
8b11a64c ZD |
1911 | /* Finds addresses in *OP_P inside STMT. */ |
1912 | ||
1913 | static void | |
726a989a | 1914 | find_interesting_uses_address (struct ivopts_data *data, gimple stmt, tree *op_p) |
8b11a64c | 1915 | { |
9a9d280e | 1916 | tree base = *op_p, step = size_zero_node; |
8b11a64c ZD |
1917 | struct iv *civ; |
1918 | struct ifs_ivopts_data ifs_ivopts_data; | |
1919 | ||
e3cc7254 ZD |
1920 | /* Do not play with volatile memory references. A bit too conservative, |
1921 | perhaps, but safe. */ | |
726a989a | 1922 | if (gimple_has_volatile_ops (stmt)) |
e3cc7254 ZD |
1923 | goto fail; |
1924 | ||
8b11a64c ZD |
1925 | /* Ignore bitfields for now. Not really something terribly complicated |
1926 | to handle. TODO. */ | |
75715cf6 EB |
1927 | if (TREE_CODE (base) == BIT_FIELD_REF) |
1928 | goto fail; | |
1929 | ||
ac182688 ZD |
1930 | base = unshare_expr (base); |
1931 | ||
1932 | if (TREE_CODE (base) == TARGET_MEM_REF) | |
1933 | { | |
1934 | tree type = build_pointer_type (TREE_TYPE (base)); | |
1935 | tree astep; | |
1936 | ||
1937 | if (TMR_BASE (base) | |
1938 | && TREE_CODE (TMR_BASE (base)) == SSA_NAME) | |
1939 | { | |
1940 | civ = get_iv (data, TMR_BASE (base)); | |
1941 | if (!civ) | |
1942 | goto fail; | |
1943 | ||
1944 | TMR_BASE (base) = civ->base; | |
1945 | step = civ->step; | |
1946 | } | |
4d948885 RG |
1947 | if (TMR_INDEX2 (base) |
1948 | && TREE_CODE (TMR_INDEX2 (base)) == SSA_NAME) | |
1949 | { | |
1950 | civ = get_iv (data, TMR_INDEX2 (base)); | |
1951 | if (!civ) | |
1952 | goto fail; | |
1953 | ||
1954 | TMR_INDEX2 (base) = civ->base; | |
1955 | step = civ->step; | |
1956 | } | |
ac182688 ZD |
1957 | if (TMR_INDEX (base) |
1958 | && TREE_CODE (TMR_INDEX (base)) == SSA_NAME) | |
1959 | { | |
1960 | civ = get_iv (data, TMR_INDEX (base)); | |
1961 | if (!civ) | |
1962 | goto fail; | |
8b11a64c | 1963 | |
ac182688 ZD |
1964 | TMR_INDEX (base) = civ->base; |
1965 | astep = civ->step; | |
be35cf60 | 1966 | |
ac182688 ZD |
1967 | if (astep) |
1968 | { | |
1969 | if (TMR_STEP (base)) | |
1970 | astep = fold_build2 (MULT_EXPR, type, TMR_STEP (base), astep); | |
1971 | ||
6e42ce54 | 1972 | step = fold_build2 (PLUS_EXPR, type, step, astep); |
ac182688 ZD |
1973 | } |
1974 | } | |
1975 | ||
6e42ce54 | 1976 | if (integer_zerop (step)) |
ac182688 ZD |
1977 | goto fail; |
1978 | base = tree_mem_ref_addr (type, base); | |
1979 | } | |
1980 | else | |
1981 | { | |
1982 | ifs_ivopts_data.ivopts_data = data; | |
1983 | ifs_ivopts_data.stmt = stmt; | |
9a9d280e | 1984 | ifs_ivopts_data.step = size_zero_node; |
ac182688 | 1985 | if (!for_each_index (&base, idx_find_step, &ifs_ivopts_data) |
6e42ce54 | 1986 | || integer_zerop (ifs_ivopts_data.step)) |
ac182688 | 1987 | goto fail; |
6e42ce54 | 1988 | step = ifs_ivopts_data.step; |
ac182688 | 1989 | |
928bc34f EB |
1990 | /* Check that the base expression is addressable. This needs |
1991 | to be done after substituting bases of IVs into it. */ | |
1992 | if (may_be_nonaddressable_p (base)) | |
1993 | goto fail; | |
1994 | ||
1995 | /* Moreover, on strict alignment platforms, check that it is | |
1996 | sufficiently aligned. */ | |
32159434 | 1997 | if (STRICT_ALIGNMENT && may_be_unaligned_p (base, step)) |
928bc34f EB |
1998 | goto fail; |
1999 | ||
ac182688 | 2000 | base = build_fold_addr_expr (base); |
ea643120 RG |
2001 | |
2002 | /* Substituting bases of IVs into the base expression might | |
2003 | have caused folding opportunities. */ | |
2004 | if (TREE_CODE (base) == ADDR_EXPR) | |
2005 | { | |
2006 | tree *ref = &TREE_OPERAND (base, 0); | |
2007 | while (handled_component_p (*ref)) | |
2008 | ref = &TREE_OPERAND (*ref, 0); | |
70f34814 | 2009 | if (TREE_CODE (*ref) == MEM_REF) |
cb6b911a | 2010 | { |
bcf71673 RG |
2011 | tree tem = fold_binary (MEM_REF, TREE_TYPE (*ref), |
2012 | TREE_OPERAND (*ref, 0), | |
2013 | TREE_OPERAND (*ref, 1)); | |
cb6b911a RG |
2014 | if (tem) |
2015 | *ref = tem; | |
2016 | } | |
ea643120 | 2017 | } |
ac182688 | 2018 | } |
8b11a64c | 2019 | |
6f929985 | 2020 | civ = alloc_iv (data, base, step); |
a7e43c57 | 2021 | record_group_use (data, op_p, civ, stmt, USE_ADDRESS); |
8b11a64c ZD |
2022 | return; |
2023 | ||
2024 | fail: | |
2025 | for_each_index (op_p, idx_record_use, data); | |
2026 | } | |
2027 | ||
2028 | /* Finds and records invariants used in STMT. */ | |
2029 | ||
2030 | static void | |
726a989a | 2031 | find_invariants_stmt (struct ivopts_data *data, gimple stmt) |
8b11a64c | 2032 | { |
f47c96aa AM |
2033 | ssa_op_iter iter; |
2034 | use_operand_p use_p; | |
8b11a64c ZD |
2035 | tree op; |
2036 | ||
f47c96aa | 2037 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) |
8b11a64c | 2038 | { |
f47c96aa | 2039 | op = USE_FROM_PTR (use_p); |
8b11a64c ZD |
2040 | record_invariant (data, op, false); |
2041 | } | |
2042 | } | |
2043 | ||
2044 | /* Finds interesting uses of induction variables in the statement STMT. */ | |
2045 | ||
2046 | static void | |
726a989a | 2047 | find_interesting_uses_stmt (struct ivopts_data *data, gimple stmt) |
8b11a64c ZD |
2048 | { |
2049 | struct iv *iv; | |
726a989a | 2050 | tree op, *lhs, *rhs; |
f47c96aa AM |
2051 | ssa_op_iter iter; |
2052 | use_operand_p use_p; | |
726a989a | 2053 | enum tree_code code; |
8b11a64c ZD |
2054 | |
2055 | find_invariants_stmt (data, stmt); | |
2056 | ||
726a989a | 2057 | if (gimple_code (stmt) == GIMPLE_COND) |
8b11a64c | 2058 | { |
726a989a | 2059 | find_interesting_uses_cond (data, stmt); |
8b11a64c ZD |
2060 | return; |
2061 | } | |
2062 | ||
726a989a | 2063 | if (is_gimple_assign (stmt)) |
8b11a64c | 2064 | { |
726a989a RB |
2065 | lhs = gimple_assign_lhs_ptr (stmt); |
2066 | rhs = gimple_assign_rhs1_ptr (stmt); | |
8b11a64c | 2067 | |
726a989a | 2068 | if (TREE_CODE (*lhs) == SSA_NAME) |
8b11a64c ZD |
2069 | { |
2070 | /* If the statement defines an induction variable, the uses are not | |
2071 | interesting by themselves. */ | |
2072 | ||
726a989a | 2073 | iv = get_iv (data, *lhs); |
8b11a64c | 2074 | |
6e42ce54 | 2075 | if (iv && !integer_zerop (iv->step)) |
8b11a64c ZD |
2076 | return; |
2077 | } | |
2078 | ||
726a989a RB |
2079 | code = gimple_assign_rhs_code (stmt); |
2080 | if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS | |
2081 | && (REFERENCE_CLASS_P (*rhs) | |
2082 | || is_gimple_val (*rhs))) | |
8b11a64c | 2083 | { |
726a989a RB |
2084 | if (REFERENCE_CLASS_P (*rhs)) |
2085 | find_interesting_uses_address (data, stmt, rhs); | |
2086 | else | |
2087 | find_interesting_uses_op (data, *rhs); | |
8b11a64c | 2088 | |
726a989a RB |
2089 | if (REFERENCE_CLASS_P (*lhs)) |
2090 | find_interesting_uses_address (data, stmt, lhs); | |
8b11a64c | 2091 | return; |
8b11a64c | 2092 | } |
726a989a | 2093 | else if (TREE_CODE_CLASS (code) == tcc_comparison) |
8b11a64c | 2094 | { |
726a989a | 2095 | find_interesting_uses_cond (data, stmt); |
8b11a64c ZD |
2096 | return; |
2097 | } | |
2f4675b4 ZD |
2098 | |
2099 | /* TODO -- we should also handle address uses of type | |
2100 | ||
2101 | memory = call (whatever); | |
2102 | ||
2103 | and | |
2104 | ||
2105 | call (memory). */ | |
8b11a64c ZD |
2106 | } |
2107 | ||
726a989a RB |
2108 | if (gimple_code (stmt) == GIMPLE_PHI |
2109 | && gimple_bb (stmt) == data->current_loop->header) | |
8b11a64c | 2110 | { |
726a989a | 2111 | iv = get_iv (data, PHI_RESULT (stmt)); |
8b11a64c | 2112 | |
6e42ce54 | 2113 | if (iv && !integer_zerop (iv->step)) |
8b11a64c ZD |
2114 | return; |
2115 | } | |
2116 | ||
f47c96aa | 2117 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) |
8b11a64c | 2118 | { |
f47c96aa | 2119 | op = USE_FROM_PTR (use_p); |
8b11a64c ZD |
2120 | |
2121 | if (TREE_CODE (op) != SSA_NAME) | |
2122 | continue; | |
2123 | ||
2124 | iv = get_iv (data, op); | |
2125 | if (!iv) | |
2126 | continue; | |
2127 | ||
2128 | find_interesting_uses_op (data, op); | |
2129 | } | |
2130 | } | |
2131 | ||
2132 | /* Finds interesting uses of induction variables outside of loops | |
2133 | on loop exit edge EXIT. */ | |
2134 | ||
2135 | static void | |
2136 | find_interesting_uses_outside (struct ivopts_data *data, edge exit) | |
2137 | { | |
538dd0b7 DM |
2138 | gphi *phi; |
2139 | gphi_iterator psi; | |
726a989a | 2140 | tree def; |
8b11a64c | 2141 | |
726a989a | 2142 | for (psi = gsi_start_phis (exit->dest); !gsi_end_p (psi); gsi_next (&psi)) |
8b11a64c | 2143 | { |
538dd0b7 | 2144 | phi = psi.phi (); |
8b11a64c | 2145 | def = PHI_ARG_DEF_FROM_EDGE (phi, exit); |
ea057359 | 2146 | if (!virtual_operand_p (def)) |
18081149 | 2147 | find_interesting_uses_op (data, def); |
8b11a64c ZD |
2148 | } |
2149 | } | |
2150 | ||
2151 | /* Finds uses of the induction variables that are interesting. */ | |
2152 | ||
2153 | static void | |
2154 | find_interesting_uses (struct ivopts_data *data) | |
2155 | { | |
2156 | basic_block bb; | |
726a989a | 2157 | gimple_stmt_iterator bsi; |
8b11a64c ZD |
2158 | basic_block *body = get_loop_body (data->current_loop); |
2159 | unsigned i; | |
2160 | struct version_info *info; | |
2161 | edge e; | |
2162 | ||
2163 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2164 | fprintf (dump_file, "Uses:\n\n"); | |
2165 | ||
2166 | for (i = 0; i < data->current_loop->num_nodes; i++) | |
2167 | { | |
628f6a4e | 2168 | edge_iterator ei; |
8b11a64c ZD |
2169 | bb = body[i]; |
2170 | ||
628f6a4e | 2171 | FOR_EACH_EDGE (e, ei, bb->succs) |
fefa31b5 | 2172 | if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
8b11a64c ZD |
2173 | && !flow_bb_inside_loop_p (data->current_loop, e->dest)) |
2174 | find_interesting_uses_outside (data, e); | |
2175 | ||
726a989a RB |
2176 | for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
2177 | find_interesting_uses_stmt (data, gsi_stmt (bsi)); | |
2178 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
b5b8b0ac AO |
2179 | if (!is_gimple_debug (gsi_stmt (bsi))) |
2180 | find_interesting_uses_stmt (data, gsi_stmt (bsi)); | |
8b11a64c ZD |
2181 | } |
2182 | ||
2183 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2184 | { | |
87c476a2 ZD |
2185 | bitmap_iterator bi; |
2186 | ||
8b11a64c ZD |
2187 | fprintf (dump_file, "\n"); |
2188 | ||
87c476a2 | 2189 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, i, bi) |
8b11a64c ZD |
2190 | { |
2191 | info = ver_info (data, i); | |
2192 | if (info->inv_id) | |
2193 | { | |
2194 | fprintf (dump_file, " "); | |
2195 | print_generic_expr (dump_file, info->name, TDF_SLIM); | |
2196 | fprintf (dump_file, " is invariant (%d)%s\n", | |
2197 | info->inv_id, info->has_nonlin_use ? "" : ", eliminable"); | |
2198 | } | |
87c476a2 | 2199 | } |
8b11a64c ZD |
2200 | |
2201 | fprintf (dump_file, "\n"); | |
2202 | } | |
2203 | ||
2204 | free (body); | |
2205 | } | |
2206 | ||
a7e43c57 BC |
2207 | /* Compute maximum offset of [base + offset] addressing mode |
2208 | for memory reference represented by USE. */ | |
2209 | ||
2210 | static HOST_WIDE_INT | |
2211 | compute_max_addr_offset (struct iv_use *use) | |
2212 | { | |
2213 | int width; | |
2214 | rtx reg, addr; | |
2215 | HOST_WIDE_INT i, off; | |
2216 | unsigned list_index, num; | |
2217 | addr_space_t as; | |
2218 | machine_mode mem_mode, addr_mode; | |
2219 | static vec<HOST_WIDE_INT> max_offset_list; | |
2220 | ||
2221 | as = TYPE_ADDR_SPACE (TREE_TYPE (use->iv->base)); | |
2222 | mem_mode = TYPE_MODE (TREE_TYPE (*use->op_p)); | |
2223 | ||
2224 | num = max_offset_list.length (); | |
2225 | list_index = (unsigned) as * MAX_MACHINE_MODE + (unsigned) mem_mode; | |
2226 | if (list_index >= num) | |
2227 | { | |
2228 | max_offset_list.safe_grow (list_index + MAX_MACHINE_MODE); | |
2229 | for (; num < max_offset_list.length (); num++) | |
2230 | max_offset_list[num] = -1; | |
2231 | } | |
2232 | ||
2233 | off = max_offset_list[list_index]; | |
2234 | if (off != -1) | |
2235 | return off; | |
2236 | ||
2237 | addr_mode = targetm.addr_space.address_mode (as); | |
2238 | reg = gen_raw_REG (addr_mode, LAST_VIRTUAL_REGISTER + 1); | |
2239 | addr = gen_rtx_fmt_ee (PLUS, addr_mode, reg, NULL_RTX); | |
2240 | ||
2241 | width = GET_MODE_BITSIZE (addr_mode) - 1; | |
2242 | if (width > (HOST_BITS_PER_WIDE_INT - 1)) | |
2243 | width = HOST_BITS_PER_WIDE_INT - 1; | |
2244 | ||
2245 | for (i = width; i > 0; i--) | |
2246 | { | |
2247 | off = ((unsigned HOST_WIDE_INT) 1 << i) - 1; | |
2248 | XEXP (addr, 1) = gen_int_mode (off, addr_mode); | |
2249 | if (memory_address_addr_space_p (mem_mode, addr, as)) | |
2250 | break; | |
2251 | ||
2252 | /* For some strict-alignment targets, the offset must be naturally | |
2253 | aligned. Try an aligned offset if mem_mode is not QImode. */ | |
2254 | off = ((unsigned HOST_WIDE_INT) 1 << i); | |
2255 | if (off > GET_MODE_SIZE (mem_mode) && mem_mode != QImode) | |
2256 | { | |
2257 | off -= GET_MODE_SIZE (mem_mode); | |
2258 | XEXP (addr, 1) = gen_int_mode (off, addr_mode); | |
2259 | if (memory_address_addr_space_p (mem_mode, addr, as)) | |
2260 | break; | |
2261 | } | |
2262 | } | |
2263 | if (i == 0) | |
2264 | off = 0; | |
2265 | ||
2266 | max_offset_list[list_index] = off; | |
2267 | return off; | |
2268 | } | |
2269 | ||
2270 | /* Check if all small groups should be split. Return true if and | |
2271 | only if: | |
2272 | ||
2273 | 1) At least one groups contain two uses with different offsets. | |
2274 | 2) No group contains more than two uses with different offsets. | |
2275 | ||
2276 | Return false otherwise. We want to split such groups because: | |
2277 | ||
2278 | 1) Small groups don't have much benefit and may interfer with | |
2279 | general candidate selection. | |
2280 | 2) Size for problem with only small groups is usually small and | |
2281 | general algorithm can handle it well. | |
2282 | ||
2283 | TODO -- Above claim may not hold when auto increment is supported. */ | |
2284 | ||
2285 | static bool | |
2286 | split_all_small_groups (struct ivopts_data *data) | |
2287 | { | |
2288 | bool split_p = false; | |
2289 | unsigned int i, n, distinct; | |
2290 | struct iv_use *pre, *use; | |
2291 | ||
2292 | n = n_iv_uses (data); | |
2293 | for (i = 0; i < n; i++) | |
2294 | { | |
2295 | use = iv_use (data, i); | |
2296 | if (!use->next) | |
2297 | continue; | |
2298 | ||
2299 | distinct = 1; | |
2300 | gcc_assert (use->type == USE_ADDRESS); | |
2301 | for (pre = use, use = use->next; use; pre = use, use = use->next) | |
2302 | { | |
2303 | if (pre->addr_offset != use->addr_offset) | |
2304 | distinct++; | |
2305 | ||
2306 | if (distinct > 2) | |
2307 | return false; | |
2308 | } | |
2309 | if (distinct == 2) | |
2310 | split_p = true; | |
2311 | } | |
2312 | ||
2313 | return split_p; | |
2314 | } | |
2315 | ||
2316 | /* For each group of address type uses, this function further groups | |
2317 | these uses according to the maximum offset supported by target's | |
2318 | [base + offset] addressing mode. */ | |
2319 | ||
2320 | static void | |
2321 | group_address_uses (struct ivopts_data *data) | |
2322 | { | |
2323 | HOST_WIDE_INT max_offset = -1; | |
2324 | unsigned int i, n, sub_id; | |
2325 | struct iv_use *pre, *use; | |
2326 | unsigned HOST_WIDE_INT addr_offset_first; | |
2327 | ||
2328 | /* Reset max offset to split all small groups. */ | |
2329 | if (split_all_small_groups (data)) | |
2330 | max_offset = 0; | |
2331 | ||
2332 | n = n_iv_uses (data); | |
2333 | for (i = 0; i < n; i++) | |
2334 | { | |
2335 | use = iv_use (data, i); | |
2336 | if (!use->next) | |
2337 | continue; | |
2338 | ||
2339 | gcc_assert (use->type == USE_ADDRESS); | |
2340 | if (max_offset != 0) | |
2341 | max_offset = compute_max_addr_offset (use); | |
2342 | ||
2343 | while (use) | |
2344 | { | |
2345 | sub_id = 0; | |
2346 | addr_offset_first = use->addr_offset; | |
2347 | /* Only uses with offset that can fit in offset part against | |
2348 | the first use can be grouped together. */ | |
2349 | for (pre = use, use = use->next; | |
2350 | use && (use->addr_offset - addr_offset_first | |
2351 | <= (unsigned HOST_WIDE_INT) max_offset); | |
2352 | pre = use, use = use->next) | |
2353 | { | |
2354 | use->id = pre->id; | |
2355 | use->sub_id = ++sub_id; | |
2356 | } | |
2357 | ||
2358 | /* Break the list and create new group. */ | |
2359 | if (use) | |
2360 | { | |
2361 | pre->next = NULL; | |
2362 | use->id = n_iv_uses (data); | |
2363 | use->related_cands = BITMAP_ALLOC (NULL); | |
2364 | data->iv_uses.safe_push (use); | |
2365 | } | |
2366 | } | |
2367 | } | |
2368 | ||
2369 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2370 | dump_uses (dump_file, data); | |
2371 | } | |
2372 | ||
f5e2738c | 2373 | /* Strips constant offsets from EXPR and stores them to OFFSET. If INSIDE_ADDR |
9be872b7 ZD |
2374 | is true, assume we are inside an address. If TOP_COMPREF is true, assume |
2375 | we are at the top-level of the processed address. */ | |
f5e2738c ZD |
2376 | |
2377 | static tree | |
9be872b7 | 2378 | strip_offset_1 (tree expr, bool inside_addr, bool top_compref, |
58fe50d5 | 2379 | HOST_WIDE_INT *offset) |
f5e2738c | 2380 | { |
9be872b7 | 2381 | tree op0 = NULL_TREE, op1 = NULL_TREE, tmp, step; |
f5e2738c ZD |
2382 | enum tree_code code; |
2383 | tree type, orig_type = TREE_TYPE (expr); | |
58fe50d5 | 2384 | HOST_WIDE_INT off0, off1, st; |
f5e2738c ZD |
2385 | tree orig_expr = expr; |
2386 | ||
2387 | STRIP_NOPS (expr); | |
9be872b7 | 2388 | |
f5e2738c ZD |
2389 | type = TREE_TYPE (expr); |
2390 | code = TREE_CODE (expr); | |
2391 | *offset = 0; | |
2392 | ||
2393 | switch (code) | |
2394 | { | |
2395 | case INTEGER_CST: | |
2396 | if (!cst_and_fits_in_hwi (expr) | |
6e682d7e | 2397 | || integer_zerop (expr)) |
f5e2738c ZD |
2398 | return orig_expr; |
2399 | ||
2400 | *offset = int_cst_value (expr); | |
ff5e9a94 | 2401 | return build_int_cst (orig_type, 0); |
f5e2738c | 2402 | |
2d1a1007 | 2403 | case POINTER_PLUS_EXPR: |
f5e2738c ZD |
2404 | case PLUS_EXPR: |
2405 | case MINUS_EXPR: | |
2406 | op0 = TREE_OPERAND (expr, 0); | |
2407 | op1 = TREE_OPERAND (expr, 1); | |
2408 | ||
9be872b7 ZD |
2409 | op0 = strip_offset_1 (op0, false, false, &off0); |
2410 | op1 = strip_offset_1 (op1, false, false, &off1); | |
f5e2738c | 2411 | |
2d1a1007 | 2412 | *offset = (code == MINUS_EXPR ? off0 - off1 : off0 + off1); |
f5e2738c ZD |
2413 | if (op0 == TREE_OPERAND (expr, 0) |
2414 | && op1 == TREE_OPERAND (expr, 1)) | |
2415 | return orig_expr; | |
2416 | ||
6e682d7e | 2417 | if (integer_zerop (op1)) |
f5e2738c | 2418 | expr = op0; |
6e682d7e | 2419 | else if (integer_zerop (op0)) |
f5e2738c | 2420 | { |
2d1a1007 | 2421 | if (code == MINUS_EXPR) |
9be872b7 | 2422 | expr = fold_build1 (NEGATE_EXPR, type, op1); |
2d1a1007 AP |
2423 | else |
2424 | expr = op1; | |
f5e2738c ZD |
2425 | } |
2426 | else | |
9be872b7 | 2427 | expr = fold_build2 (code, type, op0, op1); |
f5e2738c ZD |
2428 | |
2429 | return fold_convert (orig_type, expr); | |
2430 | ||
7a2faca1 EB |
2431 | case MULT_EXPR: |
2432 | op1 = TREE_OPERAND (expr, 1); | |
2433 | if (!cst_and_fits_in_hwi (op1)) | |
2434 | return orig_expr; | |
2435 | ||
2436 | op0 = TREE_OPERAND (expr, 0); | |
2437 | op0 = strip_offset_1 (op0, false, false, &off0); | |
2438 | if (op0 == TREE_OPERAND (expr, 0)) | |
2439 | return orig_expr; | |
2440 | ||
2441 | *offset = off0 * int_cst_value (op1); | |
2442 | if (integer_zerop (op0)) | |
2443 | expr = op0; | |
2444 | else | |
2445 | expr = fold_build2 (MULT_EXPR, type, op0, op1); | |
2446 | ||
2447 | return fold_convert (orig_type, expr); | |
2448 | ||
f5e2738c | 2449 | case ARRAY_REF: |
9f7ccf69 | 2450 | case ARRAY_RANGE_REF: |
f5e2738c ZD |
2451 | if (!inside_addr) |
2452 | return orig_expr; | |
2453 | ||
2454 | step = array_ref_element_size (expr); | |
2455 | if (!cst_and_fits_in_hwi (step)) | |
2456 | break; | |
2457 | ||
2458 | st = int_cst_value (step); | |
2459 | op1 = TREE_OPERAND (expr, 1); | |
9be872b7 | 2460 | op1 = strip_offset_1 (op1, false, false, &off1); |
f5e2738c | 2461 | *offset = off1 * st; |
9be872b7 ZD |
2462 | |
2463 | if (top_compref | |
6e682d7e | 2464 | && integer_zerop (op1)) |
9be872b7 ZD |
2465 | { |
2466 | /* Strip the component reference completely. */ | |
2467 | op0 = TREE_OPERAND (expr, 0); | |
2468 | op0 = strip_offset_1 (op0, inside_addr, top_compref, &off0); | |
2469 | *offset += off0; | |
2470 | return op0; | |
2471 | } | |
f5e2738c ZD |
2472 | break; |
2473 | ||
2474 | case COMPONENT_REF: | |
58fe50d5 BC |
2475 | { |
2476 | tree field; | |
2477 | ||
2478 | if (!inside_addr) | |
2479 | return orig_expr; | |
2480 | ||
2481 | tmp = component_ref_field_offset (expr); | |
2482 | field = TREE_OPERAND (expr, 1); | |
2483 | if (top_compref | |
2484 | && cst_and_fits_in_hwi (tmp) | |
2485 | && cst_and_fits_in_hwi (DECL_FIELD_BIT_OFFSET (field))) | |
2486 | { | |
2487 | HOST_WIDE_INT boffset, abs_off; | |
2488 | ||
2489 | /* Strip the component reference completely. */ | |
2490 | op0 = TREE_OPERAND (expr, 0); | |
2491 | op0 = strip_offset_1 (op0, inside_addr, top_compref, &off0); | |
2492 | boffset = int_cst_value (DECL_FIELD_BIT_OFFSET (field)); | |
2493 | abs_off = abs_hwi (boffset) / BITS_PER_UNIT; | |
2494 | if (boffset < 0) | |
2495 | abs_off = -abs_off; | |
2496 | ||
2497 | *offset = off0 + int_cst_value (tmp) + abs_off; | |
2498 | return op0; | |
2499 | } | |
2500 | } | |
f5e2738c ZD |
2501 | break; |
2502 | ||
2503 | case ADDR_EXPR: | |
9be872b7 ZD |
2504 | op0 = TREE_OPERAND (expr, 0); |
2505 | op0 = strip_offset_1 (op0, true, true, &off0); | |
2506 | *offset += off0; | |
2507 | ||
2508 | if (op0 == TREE_OPERAND (expr, 0)) | |
2509 | return orig_expr; | |
2510 | ||
d5dc1717 | 2511 | expr = build_fold_addr_expr (op0); |
9be872b7 ZD |
2512 | return fold_convert (orig_type, expr); |
2513 | ||
70f34814 RG |
2514 | case MEM_REF: |
2515 | /* ??? Offset operand? */ | |
9be872b7 | 2516 | inside_addr = false; |
f5e2738c ZD |
2517 | break; |
2518 | ||
2519 | default: | |
2520 | return orig_expr; | |
2521 | } | |
2522 | ||
2523 | /* Default handling of expressions for that we want to recurse into | |
2524 | the first operand. */ | |
2525 | op0 = TREE_OPERAND (expr, 0); | |
9be872b7 | 2526 | op0 = strip_offset_1 (op0, inside_addr, false, &off0); |
f5e2738c ZD |
2527 | *offset += off0; |
2528 | ||
2529 | if (op0 == TREE_OPERAND (expr, 0) | |
2530 | && (!op1 || op1 == TREE_OPERAND (expr, 1))) | |
2531 | return orig_expr; | |
2532 | ||
2533 | expr = copy_node (expr); | |
2534 | TREE_OPERAND (expr, 0) = op0; | |
2535 | if (op1) | |
2536 | TREE_OPERAND (expr, 1) = op1; | |
2537 | ||
9be872b7 | 2538 | /* Inside address, we might strip the top level component references, |
0fa2e4df | 2539 | thus changing type of the expression. Handling of ADDR_EXPR |
9be872b7 ZD |
2540 | will fix that. */ |
2541 | expr = fold_convert (orig_type, expr); | |
2542 | ||
2543 | return expr; | |
2544 | } | |
2545 | ||
2546 | /* Strips constant offsets from EXPR and stores them to OFFSET. */ | |
2547 | ||
2548 | static tree | |
2549 | strip_offset (tree expr, unsigned HOST_WIDE_INT *offset) | |
2550 | { | |
58fe50d5 BC |
2551 | HOST_WIDE_INT off; |
2552 | tree core = strip_offset_1 (expr, false, false, &off); | |
2553 | *offset = off; | |
2554 | return core; | |
f5e2738c ZD |
2555 | } |
2556 | ||
d482f417 | 2557 | /* Returns variant of TYPE that can be used as base for different uses. |
20527215 ZD |
2558 | We return unsigned type with the same precision, which avoids problems |
2559 | with overflows. */ | |
d482f417 ZD |
2560 | |
2561 | static tree | |
2562 | generic_type_for (tree type) | |
2563 | { | |
2564 | if (POINTER_TYPE_P (type)) | |
20527215 | 2565 | return unsigned_type_for (type); |
d482f417 ZD |
2566 | |
2567 | if (TYPE_UNSIGNED (type)) | |
2568 | return type; | |
2569 | ||
2570 | return unsigned_type_for (type); | |
2571 | } | |
2572 | ||
9be872b7 ZD |
2573 | /* Records invariants in *EXPR_P. Callback for walk_tree. DATA contains |
2574 | the bitmap to that we should store it. */ | |
2575 | ||
2576 | static struct ivopts_data *fd_ivopts_data; | |
2577 | static tree | |
2578 | find_depends (tree *expr_p, int *ws ATTRIBUTE_UNUSED, void *data) | |
2579 | { | |
c22940cd | 2580 | bitmap *depends_on = (bitmap *) data; |
9be872b7 ZD |
2581 | struct version_info *info; |
2582 | ||
2583 | if (TREE_CODE (*expr_p) != SSA_NAME) | |
2584 | return NULL_TREE; | |
2585 | info = name_info (fd_ivopts_data, *expr_p); | |
2586 | ||
2587 | if (!info->inv_id || info->has_nonlin_use) | |
2588 | return NULL_TREE; | |
2589 | ||
2590 | if (!*depends_on) | |
2591 | *depends_on = BITMAP_ALLOC (NULL); | |
2592 | bitmap_set_bit (*depends_on, info->inv_id); | |
2593 | ||
2594 | return NULL_TREE; | |
2595 | } | |
2596 | ||
8b11a64c ZD |
2597 | /* Adds a candidate BASE + STEP * i. Important field is set to IMPORTANT and |
2598 | position to POS. If USE is not NULL, the candidate is set as related to | |
2599 | it. If both BASE and STEP are NULL, we add a pseudocandidate for the | |
2600 | replacement of the final value of the iv by a direct computation. */ | |
2601 | ||
2602 | static struct iv_cand * | |
2603 | add_candidate_1 (struct ivopts_data *data, | |
2604 | tree base, tree step, bool important, enum iv_position pos, | |
726a989a | 2605 | struct iv_use *use, gimple incremented_at) |
8b11a64c ZD |
2606 | { |
2607 | unsigned i; | |
2608 | struct iv_cand *cand = NULL; | |
d482f417 | 2609 | tree type, orig_type; |
b8698a0f | 2610 | |
d8af4ba3 ZD |
2611 | /* For non-original variables, make sure their values are computed in a type |
2612 | that does not invoke undefined behavior on overflows (since in general, | |
2613 | we cannot prove that these induction variables are non-wrapping). */ | |
2614 | if (pos != IP_ORIGINAL) | |
8b11a64c | 2615 | { |
d482f417 ZD |
2616 | orig_type = TREE_TYPE (base); |
2617 | type = generic_type_for (orig_type); | |
71adbef3 | 2618 | if (type != orig_type) |
8b11a64c | 2619 | { |
8b11a64c | 2620 | base = fold_convert (type, base); |
6e42ce54 | 2621 | step = fold_convert (type, step); |
8b11a64c ZD |
2622 | } |
2623 | } | |
2624 | ||
2625 | for (i = 0; i < n_iv_cands (data); i++) | |
2626 | { | |
2627 | cand = iv_cand (data, i); | |
2628 | ||
2629 | if (cand->pos != pos) | |
2630 | continue; | |
2631 | ||
2c08497a BS |
2632 | if (cand->incremented_at != incremented_at |
2633 | || ((pos == IP_AFTER_USE || pos == IP_BEFORE_USE) | |
2634 | && cand->ainc_use != use)) | |
8b11a64c ZD |
2635 | continue; |
2636 | ||
2637 | if (!cand->iv) | |
2638 | { | |
2639 | if (!base && !step) | |
2640 | break; | |
2641 | ||
2642 | continue; | |
2643 | } | |
2644 | ||
2645 | if (!base && !step) | |
2646 | continue; | |
2647 | ||
6e42ce54 | 2648 | if (operand_equal_p (base, cand->iv->base, 0) |
18081149 XDL |
2649 | && operand_equal_p (step, cand->iv->step, 0) |
2650 | && (TYPE_PRECISION (TREE_TYPE (base)) | |
2651 | == TYPE_PRECISION (TREE_TYPE (cand->iv->base)))) | |
6e42ce54 | 2652 | break; |
8b11a64c ZD |
2653 | } |
2654 | ||
2655 | if (i == n_iv_cands (data)) | |
2656 | { | |
5ed6ace5 | 2657 | cand = XCNEW (struct iv_cand); |
8b11a64c ZD |
2658 | cand->id = i; |
2659 | ||
2660 | if (!base && !step) | |
2661 | cand->iv = NULL; | |
2662 | else | |
6f929985 | 2663 | cand->iv = alloc_iv (data, base, step); |
8b11a64c ZD |
2664 | |
2665 | cand->pos = pos; | |
2666 | if (pos != IP_ORIGINAL && cand->iv) | |
2667 | { | |
2668 | cand->var_before = create_tmp_var_raw (TREE_TYPE (base), "ivtmp"); | |
2669 | cand->var_after = cand->var_before; | |
2670 | } | |
2671 | cand->important = important; | |
2672 | cand->incremented_at = incremented_at; | |
9771b263 | 2673 | data->iv_candidates.safe_push (cand); |
8b11a64c | 2674 | |
9be872b7 ZD |
2675 | if (step |
2676 | && TREE_CODE (step) != INTEGER_CST) | |
2677 | { | |
2678 | fd_ivopts_data = data; | |
2679 | walk_tree (&step, find_depends, &cand->depends_on, NULL); | |
2680 | } | |
2681 | ||
2c08497a BS |
2682 | if (pos == IP_AFTER_USE || pos == IP_BEFORE_USE) |
2683 | cand->ainc_use = use; | |
2684 | else | |
2685 | cand->ainc_use = NULL; | |
2686 | ||
8b11a64c ZD |
2687 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2688 | dump_cand (dump_file, cand); | |
2689 | } | |
2690 | ||
2691 | if (important && !cand->important) | |
2692 | { | |
2693 | cand->important = true; | |
2694 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2695 | fprintf (dump_file, "Candidate %d is important\n", cand->id); | |
2696 | } | |
2697 | ||
2698 | if (use) | |
2699 | { | |
2700 | bitmap_set_bit (use->related_cands, i); | |
2701 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2702 | fprintf (dump_file, "Candidate %d is related to use %d\n", | |
2703 | cand->id, use->id); | |
2704 | } | |
2705 | ||
2706 | return cand; | |
2707 | } | |
2708 | ||
4366cf6d ZD |
2709 | /* Returns true if incrementing the induction variable at the end of the LOOP |
2710 | is allowed. | |
2711 | ||
2712 | The purpose is to avoid splitting latch edge with a biv increment, thus | |
2713 | creating a jump, possibly confusing other optimization passes and leaving | |
2714 | less freedom to scheduler. So we allow IP_END_POS only if IP_NORMAL_POS | |
2715 | is not available (so we do not have a better alternative), or if the latch | |
2716 | edge is already nonempty. */ | |
2717 | ||
2718 | static bool | |
2719 | allow_ip_end_pos_p (struct loop *loop) | |
2720 | { | |
2721 | if (!ip_normal_pos (loop)) | |
2722 | return true; | |
2723 | ||
2724 | if (!empty_block_p (ip_end_pos (loop))) | |
2725 | return true; | |
2726 | ||
2727 | return false; | |
2728 | } | |
2729 | ||
2c08497a BS |
2730 | /* If possible, adds autoincrement candidates BASE + STEP * i based on use USE. |
2731 | Important field is set to IMPORTANT. */ | |
2732 | ||
2733 | static void | |
2734 | add_autoinc_candidates (struct ivopts_data *data, tree base, tree step, | |
2735 | bool important, struct iv_use *use) | |
2736 | { | |
2737 | basic_block use_bb = gimple_bb (use->stmt); | |
ef4bddc2 | 2738 | machine_mode mem_mode; |
2c08497a BS |
2739 | unsigned HOST_WIDE_INT cstepi; |
2740 | ||
2741 | /* If we insert the increment in any position other than the standard | |
2742 | ones, we must ensure that it is incremented once per iteration. | |
2743 | It must not be in an inner nested loop, or one side of an if | |
2744 | statement. */ | |
2745 | if (use_bb->loop_father != data->current_loop | |
2746 | || !dominated_by_p (CDI_DOMINATORS, data->current_loop->latch, use_bb) | |
2747 | || stmt_could_throw_p (use->stmt) | |
2748 | || !cst_and_fits_in_hwi (step)) | |
2749 | return; | |
2750 | ||
2751 | cstepi = int_cst_value (step); | |
2752 | ||
2753 | mem_mode = TYPE_MODE (TREE_TYPE (*use->op_p)); | |
8875e939 RR |
2754 | if (((USE_LOAD_PRE_INCREMENT (mem_mode) |
2755 | || USE_STORE_PRE_INCREMENT (mem_mode)) | |
2756 | && GET_MODE_SIZE (mem_mode) == cstepi) | |
2757 | || ((USE_LOAD_PRE_DECREMENT (mem_mode) | |
2758 | || USE_STORE_PRE_DECREMENT (mem_mode)) | |
2759 | && GET_MODE_SIZE (mem_mode) == -cstepi)) | |
2c08497a BS |
2760 | { |
2761 | enum tree_code code = MINUS_EXPR; | |
2762 | tree new_base; | |
2763 | tree new_step = step; | |
2764 | ||
2765 | if (POINTER_TYPE_P (TREE_TYPE (base))) | |
2766 | { | |
2767 | new_step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step); | |
2768 | code = POINTER_PLUS_EXPR; | |
2769 | } | |
2770 | else | |
2771 | new_step = fold_convert (TREE_TYPE (base), new_step); | |
2772 | new_base = fold_build2 (code, TREE_TYPE (base), base, new_step); | |
2773 | add_candidate_1 (data, new_base, step, important, IP_BEFORE_USE, use, | |
2774 | use->stmt); | |
2775 | } | |
8875e939 RR |
2776 | if (((USE_LOAD_POST_INCREMENT (mem_mode) |
2777 | || USE_STORE_POST_INCREMENT (mem_mode)) | |
2778 | && GET_MODE_SIZE (mem_mode) == cstepi) | |
2779 | || ((USE_LOAD_POST_DECREMENT (mem_mode) | |
2780 | || USE_STORE_POST_DECREMENT (mem_mode)) | |
2781 | && GET_MODE_SIZE (mem_mode) == -cstepi)) | |
2c08497a BS |
2782 | { |
2783 | add_candidate_1 (data, base, step, important, IP_AFTER_USE, use, | |
2784 | use->stmt); | |
2785 | } | |
2786 | } | |
2787 | ||
8b11a64c ZD |
2788 | /* Adds a candidate BASE + STEP * i. Important field is set to IMPORTANT and |
2789 | position to POS. If USE is not NULL, the candidate is set as related to | |
2790 | it. The candidate computation is scheduled on all available positions. */ | |
2791 | ||
2792 | static void | |
b8698a0f | 2793 | add_candidate (struct ivopts_data *data, |
8b11a64c ZD |
2794 | tree base, tree step, bool important, struct iv_use *use) |
2795 | { | |
a7e43c57 BC |
2796 | gcc_assert (use == NULL || use->sub_id == 0); |
2797 | ||
8b11a64c | 2798 | if (ip_normal_pos (data->current_loop)) |
726a989a | 2799 | add_candidate_1 (data, base, step, important, IP_NORMAL, use, NULL); |
4366cf6d ZD |
2800 | if (ip_end_pos (data->current_loop) |
2801 | && allow_ip_end_pos_p (data->current_loop)) | |
726a989a | 2802 | add_candidate_1 (data, base, step, important, IP_END, use, NULL); |
2c08497a BS |
2803 | |
2804 | if (use != NULL && use->type == USE_ADDRESS) | |
2805 | add_autoinc_candidates (data, base, step, important, use); | |
8b11a64c ZD |
2806 | } |
2807 | ||
2808 | /* Adds standard iv candidates. */ | |
2809 | ||
2810 | static void | |
2811 | add_standard_iv_candidates (struct ivopts_data *data) | |
2812 | { | |
0f250839 RG |
2813 | add_candidate (data, integer_zero_node, integer_one_node, true, NULL); |
2814 | ||
2815 | /* The same for a double-integer type if it is still fast enough. */ | |
2816 | if (TYPE_PRECISION | |
2817 | (long_integer_type_node) > TYPE_PRECISION (integer_type_node) | |
2818 | && TYPE_PRECISION (long_integer_type_node) <= BITS_PER_WORD) | |
2819 | add_candidate (data, build_int_cst (long_integer_type_node, 0), | |
2820 | build_int_cst (long_integer_type_node, 1), true, NULL); | |
8b11a64c | 2821 | |
39b4020c | 2822 | /* The same for a double-integer type if it is still fast enough. */ |
0f250839 RG |
2823 | if (TYPE_PRECISION |
2824 | (long_long_integer_type_node) > TYPE_PRECISION (long_integer_type_node) | |
2825 | && TYPE_PRECISION (long_long_integer_type_node) <= BITS_PER_WORD) | |
2826 | add_candidate (data, build_int_cst (long_long_integer_type_node, 0), | |
2827 | build_int_cst (long_long_integer_type_node, 1), true, NULL); | |
8b11a64c ZD |
2828 | } |
2829 | ||
2830 | ||
2831 | /* Adds candidates bases on the old induction variable IV. */ | |
2832 | ||
2833 | static void | |
2834 | add_old_iv_candidates (struct ivopts_data *data, struct iv *iv) | |
2835 | { | |
726a989a RB |
2836 | gimple phi; |
2837 | tree def; | |
8b11a64c ZD |
2838 | struct iv_cand *cand; |
2839 | ||
2840 | add_candidate (data, iv->base, iv->step, true, NULL); | |
2841 | ||
2842 | /* The same, but with initial value zero. */ | |
1a00e5f7 JJ |
2843 | if (POINTER_TYPE_P (TREE_TYPE (iv->base))) |
2844 | add_candidate (data, size_int (0), iv->step, true, NULL); | |
2845 | else | |
2846 | add_candidate (data, build_int_cst (TREE_TYPE (iv->base), 0), | |
2847 | iv->step, true, NULL); | |
8b11a64c ZD |
2848 | |
2849 | phi = SSA_NAME_DEF_STMT (iv->ssa_name); | |
726a989a | 2850 | if (gimple_code (phi) == GIMPLE_PHI) |
8b11a64c ZD |
2851 | { |
2852 | /* Additionally record the possibility of leaving the original iv | |
2853 | untouched. */ | |
2854 | def = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (data->current_loop)); | |
b83b5507 BC |
2855 | /* Don't add candidate if it's from another PHI node because |
2856 | it's an affine iv appearing in the form of PEELED_CHREC. */ | |
2857 | phi = SSA_NAME_DEF_STMT (def); | |
2858 | if (gimple_code (phi) != GIMPLE_PHI) | |
2859 | { | |
2860 | cand = add_candidate_1 (data, | |
2861 | iv->base, iv->step, true, IP_ORIGINAL, NULL, | |
2862 | SSA_NAME_DEF_STMT (def)); | |
2863 | cand->var_before = iv->ssa_name; | |
2864 | cand->var_after = def; | |
2865 | } | |
2866 | else | |
2867 | gcc_assert (gimple_bb (phi) == data->current_loop->header); | |
8b11a64c ZD |
2868 | } |
2869 | } | |
2870 | ||
2871 | /* Adds candidates based on the old induction variables. */ | |
2872 | ||
2873 | static void | |
2874 | add_old_ivs_candidates (struct ivopts_data *data) | |
2875 | { | |
2876 | unsigned i; | |
2877 | struct iv *iv; | |
87c476a2 | 2878 | bitmap_iterator bi; |
8b11a64c | 2879 | |
87c476a2 | 2880 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, i, bi) |
8b11a64c ZD |
2881 | { |
2882 | iv = ver_info (data, i)->iv; | |
6e42ce54 | 2883 | if (iv && iv->biv_p && !integer_zerop (iv->step)) |
8b11a64c | 2884 | add_old_iv_candidates (data, iv); |
87c476a2 | 2885 | } |
8b11a64c ZD |
2886 | } |
2887 | ||
2888 | /* Adds candidates based on the value of the induction variable IV and USE. */ | |
2889 | ||
2890 | static void | |
2891 | add_iv_value_candidates (struct ivopts_data *data, | |
2892 | struct iv *iv, struct iv_use *use) | |
2893 | { | |
f5e2738c | 2894 | unsigned HOST_WIDE_INT offset; |
9be872b7 | 2895 | tree base; |
1ffe34d9 | 2896 | tree basetype; |
8b11a64c | 2897 | |
9be872b7 | 2898 | add_candidate (data, iv->base, iv->step, false, use); |
be35cf60 | 2899 | |
9be872b7 ZD |
2900 | /* The same, but with initial value zero. Make such variable important, |
2901 | since it is generic enough so that possibly many uses may be based | |
2902 | on it. */ | |
1ffe34d9 AP |
2903 | basetype = TREE_TYPE (iv->base); |
2904 | if (POINTER_TYPE_P (basetype)) | |
2905 | basetype = sizetype; | |
2906 | add_candidate (data, build_int_cst (basetype, 0), | |
9be872b7 | 2907 | iv->step, true, use); |
8b11a64c | 2908 | |
7eeef08e RG |
2909 | /* Third, try removing the constant offset. Make sure to even |
2910 | add a candidate for &a[0] vs. (T *)&a. */ | |
9be872b7 | 2911 | base = strip_offset (iv->base, &offset); |
7eeef08e RG |
2912 | if (offset |
2913 | || base != iv->base) | |
f5e2738c | 2914 | add_candidate (data, base, iv->step, false, use); |
8b11a64c ZD |
2915 | } |
2916 | ||
8b11a64c ZD |
2917 | /* Adds candidates based on the uses. */ |
2918 | ||
2919 | static void | |
2920 | add_derived_ivs_candidates (struct ivopts_data *data) | |
2921 | { | |
2922 | unsigned i; | |
2923 | ||
2924 | for (i = 0; i < n_iv_uses (data); i++) | |
2925 | { | |
2926 | struct iv_use *use = iv_use (data, i); | |
2927 | ||
2928 | if (!use) | |
2929 | continue; | |
2930 | ||
2931 | switch (use->type) | |
2932 | { | |
2933 | case USE_NONLINEAR_EXPR: | |
2934 | case USE_COMPARE: | |
9be872b7 | 2935 | case USE_ADDRESS: |
8b11a64c ZD |
2936 | /* Just add the ivs based on the value of the iv used here. */ |
2937 | add_iv_value_candidates (data, use->iv, use); | |
2938 | break; | |
2939 | ||
8b11a64c | 2940 | default: |
1e128c5f | 2941 | gcc_unreachable (); |
8b11a64c ZD |
2942 | } |
2943 | } | |
2944 | } | |
2945 | ||
b1b02be2 ZD |
2946 | /* Record important candidates and add them to related_cands bitmaps |
2947 | if needed. */ | |
2948 | ||
2949 | static void | |
2950 | record_important_candidates (struct ivopts_data *data) | |
2951 | { | |
2952 | unsigned i; | |
2953 | struct iv_use *use; | |
2954 | ||
2955 | for (i = 0; i < n_iv_cands (data); i++) | |
2956 | { | |
2957 | struct iv_cand *cand = iv_cand (data, i); | |
2958 | ||
2959 | if (cand->important) | |
2960 | bitmap_set_bit (data->important_candidates, i); | |
2961 | } | |
2962 | ||
2963 | data->consider_all_candidates = (n_iv_cands (data) | |
2964 | <= CONSIDER_ALL_CANDIDATES_BOUND); | |
2965 | ||
2966 | if (data->consider_all_candidates) | |
2967 | { | |
2968 | /* We will not need "related_cands" bitmaps in this case, | |
2969 | so release them to decrease peak memory consumption. */ | |
2970 | for (i = 0; i < n_iv_uses (data); i++) | |
2971 | { | |
2972 | use = iv_use (data, i); | |
8bdbfff5 | 2973 | BITMAP_FREE (use->related_cands); |
b1b02be2 ZD |
2974 | } |
2975 | } | |
2976 | else | |
2977 | { | |
2978 | /* Add important candidates to the related_cands bitmaps. */ | |
2979 | for (i = 0; i < n_iv_uses (data); i++) | |
2980 | bitmap_ior_into (iv_use (data, i)->related_cands, | |
2981 | data->important_candidates); | |
2982 | } | |
2983 | } | |
2984 | ||
8b11a64c ZD |
2985 | /* Allocates the data structure mapping the (use, candidate) pairs to costs. |
2986 | If consider_all_candidates is true, we use a two-dimensional array, otherwise | |
2987 | we allocate a simple list to every use. */ | |
2988 | ||
2989 | static void | |
2990 | alloc_use_cost_map (struct ivopts_data *data) | |
2991 | { | |
79836a12 | 2992 | unsigned i, size, s; |
8b11a64c ZD |
2993 | |
2994 | for (i = 0; i < n_iv_uses (data); i++) | |
2995 | { | |
2996 | struct iv_use *use = iv_use (data, i); | |
2997 | ||
2998 | if (data->consider_all_candidates) | |
b1b02be2 | 2999 | size = n_iv_cands (data); |
8b11a64c ZD |
3000 | else |
3001 | { | |
79836a12 | 3002 | s = bitmap_count_bits (use->related_cands); |
b1b02be2 ZD |
3003 | |
3004 | /* Round up to the power of two, so that moduling by it is fast. */ | |
79836a12 | 3005 | size = s ? (1 << ceil_log2 (s)) : 1; |
8b11a64c ZD |
3006 | } |
3007 | ||
b1b02be2 | 3008 | use->n_map_members = size; |
5ed6ace5 | 3009 | use->cost_map = XCNEWVEC (struct cost_pair, size); |
8b11a64c ZD |
3010 | } |
3011 | } | |
3012 | ||
6e8c65f6 ZD |
3013 | /* Returns description of computation cost of expression whose runtime |
3014 | cost is RUNTIME and complexity corresponds to COMPLEXITY. */ | |
3015 | ||
3016 | static comp_cost | |
3017 | new_cost (unsigned runtime, unsigned complexity) | |
3018 | { | |
3019 | comp_cost cost; | |
3020 | ||
3021 | cost.cost = runtime; | |
3022 | cost.complexity = complexity; | |
3023 | ||
3024 | return cost; | |
3025 | } | |
3026 | ||
a7e43c57 BC |
3027 | /* Returns true if COST is infinite. */ |
3028 | ||
3029 | static bool | |
3030 | infinite_cost_p (comp_cost cost) | |
3031 | { | |
3032 | return cost.cost == INFTY; | |
3033 | } | |
3034 | ||
6e8c65f6 ZD |
3035 | /* Adds costs COST1 and COST2. */ |
3036 | ||
3037 | static comp_cost | |
3038 | add_costs (comp_cost cost1, comp_cost cost2) | |
3039 | { | |
a7e43c57 BC |
3040 | if (infinite_cost_p (cost1) || infinite_cost_p (cost2)) |
3041 | return infinite_cost; | |
3042 | ||
6e8c65f6 ZD |
3043 | cost1.cost += cost2.cost; |
3044 | cost1.complexity += cost2.complexity; | |
3045 | ||
3046 | return cost1; | |
3047 | } | |
3048 | /* Subtracts costs COST1 and COST2. */ | |
3049 | ||
3050 | static comp_cost | |
3051 | sub_costs (comp_cost cost1, comp_cost cost2) | |
3052 | { | |
3053 | cost1.cost -= cost2.cost; | |
3054 | cost1.complexity -= cost2.complexity; | |
3055 | ||
3056 | return cost1; | |
3057 | } | |
3058 | ||
3059 | /* Returns a negative number if COST1 < COST2, a positive number if | |
3060 | COST1 > COST2, and 0 if COST1 = COST2. */ | |
3061 | ||
3062 | static int | |
3063 | compare_costs (comp_cost cost1, comp_cost cost2) | |
3064 | { | |
3065 | if (cost1.cost == cost2.cost) | |
3066 | return cost1.complexity - cost2.complexity; | |
3067 | ||
3068 | return cost1.cost - cost2.cost; | |
3069 | } | |
3070 | ||
8b11a64c | 3071 | /* Sets cost of (USE, CANDIDATE) pair to COST and record that it depends |
f5f12961 | 3072 | on invariants DEPENDS_ON and that the value used in expressing it |
d8af4ba3 | 3073 | is VALUE, and in case of iv elimination the comparison operator is COMP. */ |
8b11a64c ZD |
3074 | |
3075 | static void | |
3076 | set_use_iv_cost (struct ivopts_data *data, | |
6e8c65f6 | 3077 | struct iv_use *use, struct iv_cand *cand, |
18081149 | 3078 | comp_cost cost, bitmap depends_on, tree value, |
d8af4ba3 | 3079 | enum tree_code comp, int inv_expr_id) |
8b11a64c | 3080 | { |
b1b02be2 ZD |
3081 | unsigned i, s; |
3082 | ||
6e8c65f6 | 3083 | if (infinite_cost_p (cost)) |
8b11a64c | 3084 | { |
8bdbfff5 | 3085 | BITMAP_FREE (depends_on); |
b1b02be2 | 3086 | return; |
8b11a64c ZD |
3087 | } |
3088 | ||
3089 | if (data->consider_all_candidates) | |
3090 | { | |
3091 | use->cost_map[cand->id].cand = cand; | |
3092 | use->cost_map[cand->id].cost = cost; | |
3093 | use->cost_map[cand->id].depends_on = depends_on; | |
f5f12961 | 3094 | use->cost_map[cand->id].value = value; |
d8af4ba3 | 3095 | use->cost_map[cand->id].comp = comp; |
18081149 | 3096 | use->cost_map[cand->id].inv_expr_id = inv_expr_id; |
8b11a64c ZD |
3097 | return; |
3098 | } | |
3099 | ||
b1b02be2 ZD |
3100 | /* n_map_members is a power of two, so this computes modulo. */ |
3101 | s = cand->id & (use->n_map_members - 1); | |
3102 | for (i = s; i < use->n_map_members; i++) | |
3103 | if (!use->cost_map[i].cand) | |
3104 | goto found; | |
3105 | for (i = 0; i < s; i++) | |
3106 | if (!use->cost_map[i].cand) | |
3107 | goto found; | |
3108 | ||
3109 | gcc_unreachable (); | |
8b11a64c | 3110 | |
b1b02be2 ZD |
3111 | found: |
3112 | use->cost_map[i].cand = cand; | |
3113 | use->cost_map[i].cost = cost; | |
3114 | use->cost_map[i].depends_on = depends_on; | |
f5f12961 | 3115 | use->cost_map[i].value = value; |
d8af4ba3 | 3116 | use->cost_map[i].comp = comp; |
18081149 | 3117 | use->cost_map[i].inv_expr_id = inv_expr_id; |
8b11a64c ZD |
3118 | } |
3119 | ||
b1b02be2 | 3120 | /* Gets cost of (USE, CANDIDATE) pair. */ |
8b11a64c | 3121 | |
b1b02be2 ZD |
3122 | static struct cost_pair * |
3123 | get_use_iv_cost (struct ivopts_data *data, struct iv_use *use, | |
3124 | struct iv_cand *cand) | |
8b11a64c | 3125 | { |
b1b02be2 ZD |
3126 | unsigned i, s; |
3127 | struct cost_pair *ret; | |
8b11a64c ZD |
3128 | |
3129 | if (!cand) | |
b1b02be2 | 3130 | return NULL; |
8b11a64c ZD |
3131 | |
3132 | if (data->consider_all_candidates) | |
8b11a64c | 3133 | { |
b1b02be2 ZD |
3134 | ret = use->cost_map + cand->id; |
3135 | if (!ret->cand) | |
3136 | return NULL; | |
8b11a64c | 3137 | |
b1b02be2 | 3138 | return ret; |
8b11a64c | 3139 | } |
b8698a0f | 3140 | |
b1b02be2 ZD |
3141 | /* n_map_members is a power of two, so this computes modulo. */ |
3142 | s = cand->id & (use->n_map_members - 1); | |
3143 | for (i = s; i < use->n_map_members; i++) | |
3144 | if (use->cost_map[i].cand == cand) | |
3145 | return use->cost_map + i; | |
79836a12 RB |
3146 | else if (use->cost_map[i].cand == NULL) |
3147 | return NULL; | |
b1b02be2 ZD |
3148 | for (i = 0; i < s; i++) |
3149 | if (use->cost_map[i].cand == cand) | |
3150 | return use->cost_map + i; | |
79836a12 RB |
3151 | else if (use->cost_map[i].cand == NULL) |
3152 | return NULL; | |
b1b02be2 ZD |
3153 | |
3154 | return NULL; | |
8b11a64c ZD |
3155 | } |
3156 | ||
8679c649 JH |
3157 | /* Produce DECL_RTL for object obj so it looks like it is stored in memory. */ |
3158 | static rtx | |
3159 | produce_memory_decl_rtl (tree obj, int *regno) | |
3160 | { | |
09e881c9 | 3161 | addr_space_t as = TYPE_ADDR_SPACE (TREE_TYPE (obj)); |
ef4bddc2 | 3162 | machine_mode address_mode = targetm.addr_space.address_mode (as); |
8679c649 | 3163 | rtx x; |
b8698a0f | 3164 | |
e16acfcd | 3165 | gcc_assert (obj); |
8679c649 JH |
3166 | if (TREE_STATIC (obj) || DECL_EXTERNAL (obj)) |
3167 | { | |
3168 | const char *name = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (obj)); | |
d4ebfa65 | 3169 | x = gen_rtx_SYMBOL_REF (address_mode, name); |
8318b0d9 RH |
3170 | SET_SYMBOL_REF_DECL (x, obj); |
3171 | x = gen_rtx_MEM (DECL_MODE (obj), x); | |
09e881c9 | 3172 | set_mem_addr_space (x, as); |
8318b0d9 | 3173 | targetm.encode_section_info (obj, x, true); |
8679c649 JH |
3174 | } |
3175 | else | |
8318b0d9 | 3176 | { |
d4ebfa65 | 3177 | x = gen_raw_REG (address_mode, (*regno)++); |
8318b0d9 | 3178 | x = gen_rtx_MEM (DECL_MODE (obj), x); |
09e881c9 | 3179 | set_mem_addr_space (x, as); |
8318b0d9 | 3180 | } |
8679c649 | 3181 | |
8318b0d9 | 3182 | return x; |
8679c649 JH |
3183 | } |
3184 | ||
8b11a64c ZD |
3185 | /* Prepares decl_rtl for variables referred in *EXPR_P. Callback for |
3186 | walk_tree. DATA contains the actual fake register number. */ | |
3187 | ||
3188 | static tree | |
3189 | prepare_decl_rtl (tree *expr_p, int *ws, void *data) | |
3190 | { | |
3191 | tree obj = NULL_TREE; | |
3192 | rtx x = NULL_RTX; | |
c22940cd | 3193 | int *regno = (int *) data; |
8b11a64c ZD |
3194 | |
3195 | switch (TREE_CODE (*expr_p)) | |
3196 | { | |
8679c649 JH |
3197 | case ADDR_EXPR: |
3198 | for (expr_p = &TREE_OPERAND (*expr_p, 0); | |
afe84921 RH |
3199 | handled_component_p (*expr_p); |
3200 | expr_p = &TREE_OPERAND (*expr_p, 0)) | |
3201 | continue; | |
8679c649 | 3202 | obj = *expr_p; |
c401fb6f | 3203 | if (DECL_P (obj) && HAS_RTL_P (obj) && !DECL_RTL_SET_P (obj)) |
8679c649 JH |
3204 | x = produce_memory_decl_rtl (obj, regno); |
3205 | break; | |
3206 | ||
8b11a64c ZD |
3207 | case SSA_NAME: |
3208 | *ws = 0; | |
3209 | obj = SSA_NAME_VAR (*expr_p); | |
70b5e7dc RG |
3210 | /* Defer handling of anonymous SSA_NAMEs to the expander. */ |
3211 | if (!obj) | |
3212 | return NULL_TREE; | |
8b11a64c ZD |
3213 | if (!DECL_RTL_SET_P (obj)) |
3214 | x = gen_raw_REG (DECL_MODE (obj), (*regno)++); | |
3215 | break; | |
3216 | ||
3217 | case VAR_DECL: | |
3218 | case PARM_DECL: | |
3219 | case RESULT_DECL: | |
3220 | *ws = 0; | |
3221 | obj = *expr_p; | |
3222 | ||
3223 | if (DECL_RTL_SET_P (obj)) | |
3224 | break; | |
3225 | ||
3226 | if (DECL_MODE (obj) == BLKmode) | |
8679c649 | 3227 | x = produce_memory_decl_rtl (obj, regno); |
8b11a64c ZD |
3228 | else |
3229 | x = gen_raw_REG (DECL_MODE (obj), (*regno)++); | |
3230 | ||
3231 | break; | |
3232 | ||
3233 | default: | |
3234 | break; | |
3235 | } | |
3236 | ||
3237 | if (x) | |
3238 | { | |
9771b263 | 3239 | decl_rtl_to_reset.safe_push (obj); |
8b11a64c ZD |
3240 | SET_DECL_RTL (obj, x); |
3241 | } | |
3242 | ||
3243 | return NULL_TREE; | |
3244 | } | |
3245 | ||
3246 | /* Determines cost of the computation of EXPR. */ | |
3247 | ||
3248 | static unsigned | |
f40751dd | 3249 | computation_cost (tree expr, bool speed) |
8b11a64c | 3250 | { |
b32d5189 DM |
3251 | rtx_insn *seq; |
3252 | rtx rslt; | |
8b11a64c ZD |
3253 | tree type = TREE_TYPE (expr); |
3254 | unsigned cost; | |
1d27fed4 DD |
3255 | /* Avoid using hard regs in ways which may be unsupported. */ |
3256 | int regno = LAST_VIRTUAL_REGISTER + 1; | |
d52f5295 | 3257 | struct cgraph_node *node = cgraph_node::get (current_function_decl); |
5fefcf92 | 3258 | enum node_frequency real_frequency = node->frequency; |
8b11a64c | 3259 | |
5fefcf92 | 3260 | node->frequency = NODE_FREQUENCY_NORMAL; |
f40751dd | 3261 | crtl->maybe_hot_insn_p = speed; |
8b11a64c ZD |
3262 | walk_tree (&expr, prepare_decl_rtl, ®no, NULL); |
3263 | start_sequence (); | |
3264 | rslt = expand_expr (expr, NULL_RTX, TYPE_MODE (type), EXPAND_NORMAL); | |
3265 | seq = get_insns (); | |
3266 | end_sequence (); | |
f40751dd | 3267 | default_rtl_profile (); |
5fefcf92 | 3268 | node->frequency = real_frequency; |
8b11a64c | 3269 | |
f40751dd | 3270 | cost = seq_cost (seq, speed); |
2ca202e7 | 3271 | if (MEM_P (rslt)) |
09e881c9 BE |
3272 | cost += address_cost (XEXP (rslt, 0), TYPE_MODE (type), |
3273 | TYPE_ADDR_SPACE (type), speed); | |
b5ee6752 | 3274 | else if (!REG_P (rslt)) |
e548c9df | 3275 | cost += set_src_cost (rslt, TYPE_MODE (type), speed); |
8b11a64c ZD |
3276 | |
3277 | return cost; | |
3278 | } | |
3279 | ||
3280 | /* Returns variable containing the value of candidate CAND at statement AT. */ | |
3281 | ||
3282 | static tree | |
726a989a | 3283 | var_at_stmt (struct loop *loop, struct iv_cand *cand, gimple stmt) |
8b11a64c ZD |
3284 | { |
3285 | if (stmt_after_increment (loop, cand, stmt)) | |
3286 | return cand->var_after; | |
3287 | else | |
3288 | return cand->var_before; | |
3289 | } | |
3290 | ||
b67102ae ZD |
3291 | /* If A is (TYPE) BA and B is (TYPE) BB, and the types of BA and BB have the |
3292 | same precision that is at least as wide as the precision of TYPE, stores | |
3293 | BA to A and BB to B, and returns the type of BA. Otherwise, returns the | |
3294 | type of A and B. */ | |
3295 | ||
3296 | static tree | |
3297 | determine_common_wider_type (tree *a, tree *b) | |
3298 | { | |
3299 | tree wider_type = NULL; | |
3300 | tree suba, subb; | |
3301 | tree atype = TREE_TYPE (*a); | |
3302 | ||
1043771b | 3303 | if (CONVERT_EXPR_P (*a)) |
b67102ae ZD |
3304 | { |
3305 | suba = TREE_OPERAND (*a, 0); | |
3306 | wider_type = TREE_TYPE (suba); | |
3307 | if (TYPE_PRECISION (wider_type) < TYPE_PRECISION (atype)) | |
3308 | return atype; | |
3309 | } | |
3310 | else | |
3311 | return atype; | |
3312 | ||
1043771b | 3313 | if (CONVERT_EXPR_P (*b)) |
b67102ae ZD |
3314 | { |
3315 | subb = TREE_OPERAND (*b, 0); | |
3316 | if (TYPE_PRECISION (wider_type) != TYPE_PRECISION (TREE_TYPE (subb))) | |
3317 | return atype; | |
3318 | } | |
3319 | else | |
3320 | return atype; | |
3321 | ||
3322 | *a = suba; | |
3323 | *b = subb; | |
3324 | return wider_type; | |
3325 | } | |
3326 | ||
8b11a64c | 3327 | /* Determines the expression by that USE is expressed from induction variable |
ac182688 ZD |
3328 | CAND at statement AT in LOOP. The expression is stored in a decomposed |
3329 | form into AFF. Returns false if USE cannot be expressed using CAND. */ | |
8b11a64c | 3330 | |
ac182688 ZD |
3331 | static bool |
3332 | get_computation_aff (struct loop *loop, | |
726a989a | 3333 | struct iv_use *use, struct iv_cand *cand, gimple at, |
84562394 | 3334 | struct aff_tree *aff) |
8b11a64c | 3335 | { |
2f4675b4 ZD |
3336 | tree ubase = use->iv->base; |
3337 | tree ustep = use->iv->step; | |
3338 | tree cbase = cand->iv->base; | |
73f30c63 | 3339 | tree cstep = cand->iv->step, cstep_common; |
8b11a64c | 3340 | tree utype = TREE_TYPE (ubase), ctype = TREE_TYPE (cbase); |
73f30c63 | 3341 | tree common_type, var; |
8b11a64c | 3342 | tree uutype; |
73f30c63 | 3343 | aff_tree cbase_aff, var_aff; |
807e902e | 3344 | widest_int rat; |
8b11a64c ZD |
3345 | |
3346 | if (TYPE_PRECISION (utype) > TYPE_PRECISION (ctype)) | |
3347 | { | |
3348 | /* We do not have a precision to express the values of use. */ | |
ac182688 | 3349 | return false; |
8b11a64c ZD |
3350 | } |
3351 | ||
73f30c63 ZD |
3352 | var = var_at_stmt (loop, cand, at); |
3353 | uutype = unsigned_type_for (utype); | |
8b11a64c | 3354 | |
73f30c63 ZD |
3355 | /* If the conversion is not noop, perform it. */ |
3356 | if (TYPE_PRECISION (utype) < TYPE_PRECISION (ctype)) | |
8b11a64c | 3357 | { |
8b11a64c | 3358 | cstep = fold_convert (uutype, cstep); |
73f30c63 ZD |
3359 | cbase = fold_convert (uutype, cbase); |
3360 | var = fold_convert (uutype, var); | |
9be872b7 | 3361 | } |
9be872b7 | 3362 | |
73f30c63 ZD |
3363 | if (!constant_multiple_of (ustep, cstep, &rat)) |
3364 | return false; | |
8b11a64c | 3365 | |
b67102ae ZD |
3366 | /* In case both UBASE and CBASE are shortened to UUTYPE from some common |
3367 | type, we achieve better folding by computing their difference in this | |
3368 | wider type, and cast the result to UUTYPE. We do not need to worry about | |
3369 | overflows, as all the arithmetics will in the end be performed in UUTYPE | |
3370 | anyway. */ | |
3371 | common_type = determine_common_wider_type (&ubase, &cbase); | |
3372 | ||
73f30c63 ZD |
3373 | /* use = ubase - ratio * cbase + ratio * var. */ |
3374 | tree_to_aff_combination (ubase, common_type, aff); | |
3375 | tree_to_aff_combination (cbase, common_type, &cbase_aff); | |
3376 | tree_to_aff_combination (var, uutype, &var_aff); | |
8b11a64c | 3377 | |
73f30c63 ZD |
3378 | /* We need to shift the value if we are after the increment. */ |
3379 | if (stmt_after_increment (loop, cand, at)) | |
8b11a64c | 3380 | { |
73f30c63 | 3381 | aff_tree cstep_aff; |
b8698a0f | 3382 | |
73f30c63 ZD |
3383 | if (common_type != uutype) |
3384 | cstep_common = fold_convert (common_type, cstep); | |
9be872b7 | 3385 | else |
73f30c63 | 3386 | cstep_common = cstep; |
ac182688 | 3387 | |
73f30c63 ZD |
3388 | tree_to_aff_combination (cstep_common, common_type, &cstep_aff); |
3389 | aff_combination_add (&cbase_aff, &cstep_aff); | |
8b11a64c | 3390 | } |
8b11a64c | 3391 | |
27bcd47c | 3392 | aff_combination_scale (&cbase_aff, -rat); |
ac182688 | 3393 | aff_combination_add (aff, &cbase_aff); |
b67102ae | 3394 | if (common_type != uutype) |
73f30c63 ZD |
3395 | aff_combination_convert (aff, uutype); |
3396 | ||
3397 | aff_combination_scale (&var_aff, rat); | |
3398 | aff_combination_add (aff, &var_aff); | |
ac182688 ZD |
3399 | |
3400 | return true; | |
3401 | } | |
3402 | ||
0c604a61 TV |
3403 | /* Return the type of USE. */ |
3404 | ||
3405 | static tree | |
3406 | get_use_type (struct iv_use *use) | |
3407 | { | |
3408 | tree base_type = TREE_TYPE (use->iv->base); | |
3409 | tree type; | |
3410 | ||
3411 | if (use->type == USE_ADDRESS) | |
3412 | { | |
3413 | /* The base_type may be a void pointer. Create a pointer type based on | |
3414 | the mem_ref instead. */ | |
3415 | type = build_pointer_type (TREE_TYPE (*use->op_p)); | |
3416 | gcc_assert (TYPE_ADDR_SPACE (TREE_TYPE (type)) | |
3417 | == TYPE_ADDR_SPACE (TREE_TYPE (base_type))); | |
3418 | } | |
3419 | else | |
3420 | type = base_type; | |
3421 | ||
3422 | return type; | |
3423 | } | |
3424 | ||
ac182688 ZD |
3425 | /* Determines the expression by that USE is expressed from induction variable |
3426 | CAND at statement AT in LOOP. The computation is unshared. */ | |
3427 | ||
3428 | static tree | |
3429 | get_computation_at (struct loop *loop, | |
726a989a | 3430 | struct iv_use *use, struct iv_cand *cand, gimple at) |
ac182688 | 3431 | { |
73f30c63 | 3432 | aff_tree aff; |
0c604a61 | 3433 | tree type = get_use_type (use); |
ac182688 ZD |
3434 | |
3435 | if (!get_computation_aff (loop, use, cand, at, &aff)) | |
3436 | return NULL_TREE; | |
3437 | unshare_aff_combination (&aff); | |
3438 | return fold_convert (type, aff_combination_to_tree (&aff)); | |
8b11a64c ZD |
3439 | } |
3440 | ||
3441 | /* Determines the expression by that USE is expressed from induction variable | |
ac182688 | 3442 | CAND in LOOP. The computation is unshared. */ |
8b11a64c ZD |
3443 | |
3444 | static tree | |
3445 | get_computation (struct loop *loop, struct iv_use *use, struct iv_cand *cand) | |
3446 | { | |
3447 | return get_computation_at (loop, use, cand, use->stmt); | |
3448 | } | |
3449 | ||
6521ac85 SL |
3450 | /* Adjust the cost COST for being in loop setup rather than loop body. |
3451 | If we're optimizing for space, the loop setup overhead is constant; | |
3452 | if we're optimizing for speed, amortize it over the per-iteration cost. */ | |
3453 | static unsigned | |
3454 | adjust_setup_cost (struct ivopts_data *data, unsigned cost) | |
3455 | { | |
3456 | if (cost == INFTY) | |
3457 | return cost; | |
3458 | else if (optimize_loop_for_speed_p (data->current_loop)) | |
18081149 | 3459 | return cost / avg_loop_niter (data->current_loop); |
6521ac85 SL |
3460 | else |
3461 | return cost; | |
3462 | } | |
3463 | ||
b1dd92fd | 3464 | /* Returns true if multiplying by RATIO is allowed in an address. Test the |
09e881c9 BE |
3465 | validity for a memory reference accessing memory of mode MODE in |
3466 | address space AS. */ | |
3467 | ||
ac182688 ZD |
3468 | |
3469 | bool | |
ef4bddc2 | 3470 | multiplier_allowed_in_address_p (HOST_WIDE_INT ratio, machine_mode mode, |
09e881c9 | 3471 | addr_space_t as) |
ac182688 ZD |
3472 | { |
3473 | #define MAX_RATIO 128 | |
09e881c9 | 3474 | unsigned int data_index = (int) as * MAX_MACHINE_MODE + (int) mode; |
9771b263 | 3475 | static vec<sbitmap> valid_mult_list; |
09e881c9 BE |
3476 | sbitmap valid_mult; |
3477 | ||
9771b263 DN |
3478 | if (data_index >= valid_mult_list.length ()) |
3479 | valid_mult_list.safe_grow_cleared (data_index + 1); | |
09e881c9 | 3480 | |
9771b263 | 3481 | valid_mult = valid_mult_list[data_index]; |
09e881c9 | 3482 | if (!valid_mult) |
ac182688 | 3483 | { |
ef4bddc2 | 3484 | machine_mode address_mode = targetm.addr_space.address_mode (as); |
d4ebfa65 | 3485 | rtx reg1 = gen_raw_REG (address_mode, LAST_VIRTUAL_REGISTER + 1); |
f258df2d IS |
3486 | rtx reg2 = gen_raw_REG (address_mode, LAST_VIRTUAL_REGISTER + 2); |
3487 | rtx addr, scaled; | |
ac182688 ZD |
3488 | HOST_WIDE_INT i; |
3489 | ||
09e881c9 | 3490 | valid_mult = sbitmap_alloc (2 * MAX_RATIO + 1); |
f61e445a | 3491 | bitmap_clear (valid_mult); |
f258df2d IS |
3492 | scaled = gen_rtx_fmt_ee (MULT, address_mode, reg1, NULL_RTX); |
3493 | addr = gen_rtx_fmt_ee (PLUS, address_mode, scaled, reg2); | |
ac182688 ZD |
3494 | for (i = -MAX_RATIO; i <= MAX_RATIO; i++) |
3495 | { | |
f258df2d IS |
3496 | XEXP (scaled, 1) = gen_int_mode (i, address_mode); |
3497 | if (memory_address_addr_space_p (mode, addr, as) | |
3498 | || memory_address_addr_space_p (mode, scaled, as)) | |
d7c028c0 | 3499 | bitmap_set_bit (valid_mult, i + MAX_RATIO); |
ac182688 ZD |
3500 | } |
3501 | ||
3502 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3503 | { | |
3504 | fprintf (dump_file, " allowed multipliers:"); | |
3505 | for (i = -MAX_RATIO; i <= MAX_RATIO; i++) | |
d7c028c0 | 3506 | if (bitmap_bit_p (valid_mult, i + MAX_RATIO)) |
ac182688 ZD |
3507 | fprintf (dump_file, " %d", (int) i); |
3508 | fprintf (dump_file, "\n"); | |
3509 | fprintf (dump_file, "\n"); | |
3510 | } | |
09e881c9 | 3511 | |
9771b263 | 3512 | valid_mult_list[data_index] = valid_mult; |
ac182688 ZD |
3513 | } |
3514 | ||
3515 | if (ratio > MAX_RATIO || ratio < -MAX_RATIO) | |
3516 | return false; | |
3517 | ||
d7c028c0 | 3518 | return bitmap_bit_p (valid_mult, ratio + MAX_RATIO); |
ac182688 ZD |
3519 | } |
3520 | ||
8b11a64c ZD |
3521 | /* Returns cost of address in shape symbol + var + OFFSET + RATIO * index. |
3522 | If SYMBOL_PRESENT is false, symbol is omitted. If VAR_PRESENT is false, | |
b1dd92fd | 3523 | variable is omitted. Compute the cost for a memory reference that accesses |
09e881c9 | 3524 | a memory location of mode MEM_MODE in address space AS. |
b1dd92fd | 3525 | |
2c08497a BS |
3526 | MAY_AUTOINC is set to true if the autoincrement (increasing index by |
3527 | size of MEM_MODE / RATIO) is available. To make this determination, we | |
3528 | look at the size of the increment to be made, which is given in CSTEP. | |
3529 | CSTEP may be zero if the step is unknown. | |
3530 | STMT_AFTER_INC is true iff the statement we're looking at is after the | |
3531 | increment of the original biv. | |
3532 | ||
8b11a64c ZD |
3533 | TODO -- there must be some better way. This all is quite crude. */ |
3534 | ||
cd047059 BC |
3535 | enum ainc_type |
3536 | { | |
3537 | AINC_PRE_INC, /* Pre increment. */ | |
3538 | AINC_PRE_DEC, /* Pre decrement. */ | |
3539 | AINC_POST_INC, /* Post increment. */ | |
3540 | AINC_POST_DEC, /* Post decrement. */ | |
3541 | AINC_NONE /* Also the number of auto increment types. */ | |
3542 | }; | |
3543 | ||
0823efed | 3544 | typedef struct address_cost_data_s |
09e881c9 BE |
3545 | { |
3546 | HOST_WIDE_INT min_offset, max_offset; | |
3547 | unsigned costs[2][2][2][2]; | |
cd047059 | 3548 | unsigned ainc_costs[AINC_NONE]; |
09e881c9 BE |
3549 | } *address_cost_data; |
3550 | ||
09e881c9 | 3551 | |
6e8c65f6 | 3552 | static comp_cost |
8b11a64c | 3553 | get_address_cost (bool symbol_present, bool var_present, |
dd5f63f8 | 3554 | unsigned HOST_WIDE_INT offset, HOST_WIDE_INT ratio, |
ef4bddc2 | 3555 | HOST_WIDE_INT cstep, machine_mode mem_mode, |
09e881c9 | 3556 | addr_space_t as, bool speed, |
2c08497a | 3557 | bool stmt_after_inc, bool *may_autoinc) |
8b11a64c | 3558 | { |
ef4bddc2 | 3559 | machine_mode address_mode = targetm.addr_space.address_mode (as); |
9771b263 | 3560 | static vec<address_cost_data> address_cost_data_list; |
09e881c9 BE |
3561 | unsigned int data_index = (int) as * MAX_MACHINE_MODE + (int) mem_mode; |
3562 | address_cost_data data; | |
2c08497a BS |
3563 | static bool has_preinc[MAX_MACHINE_MODE], has_postinc[MAX_MACHINE_MODE]; |
3564 | static bool has_predec[MAX_MACHINE_MODE], has_postdec[MAX_MACHINE_MODE]; | |
6e8c65f6 | 3565 | unsigned cost, acost, complexity; |
cd047059 | 3566 | enum ainc_type autoinc_type; |
2c08497a BS |
3567 | bool offset_p, ratio_p, autoinc; |
3568 | HOST_WIDE_INT s_offset, autoinc_offset, msize; | |
8b11a64c ZD |
3569 | unsigned HOST_WIDE_INT mask; |
3570 | unsigned bits; | |
3571 | ||
9771b263 DN |
3572 | if (data_index >= address_cost_data_list.length ()) |
3573 | address_cost_data_list.safe_grow_cleared (data_index + 1); | |
09e881c9 | 3574 | |
9771b263 | 3575 | data = address_cost_data_list[data_index]; |
09e881c9 | 3576 | if (!data) |
8b11a64c ZD |
3577 | { |
3578 | HOST_WIDE_INT i; | |
d6fa2d17 XDL |
3579 | HOST_WIDE_INT rat, off = 0; |
3580 | int old_cse_not_expected, width; | |
975626a7 | 3581 | unsigned sym_p, var_p, off_p, rat_p, add_c; |
b32d5189 DM |
3582 | rtx_insn *seq; |
3583 | rtx addr, base; | |
975626a7 ZD |
3584 | rtx reg0, reg1; |
3585 | ||
09e881c9 | 3586 | data = (address_cost_data) xcalloc (1, sizeof (*data)); |
8b11a64c | 3587 | |
d4ebfa65 | 3588 | reg1 = gen_raw_REG (address_mode, LAST_VIRTUAL_REGISTER + 1); |
8b11a64c | 3589 | |
d6fa2d17 XDL |
3590 | width = GET_MODE_BITSIZE (address_mode) - 1; |
3591 | if (width > (HOST_BITS_PER_WIDE_INT - 1)) | |
3592 | width = HOST_BITS_PER_WIDE_INT - 1; | |
d4ebfa65 | 3593 | addr = gen_rtx_fmt_ee (PLUS, address_mode, reg1, NULL_RTX); |
d6fa2d17 XDL |
3594 | |
3595 | for (i = width; i >= 0; i--) | |
8b11a64c | 3596 | { |
eb87c7c4 | 3597 | off = -((unsigned HOST_WIDE_INT) 1 << i); |
d6fa2d17 XDL |
3598 | XEXP (addr, 1) = gen_int_mode (off, address_mode); |
3599 | if (memory_address_addr_space_p (mem_mode, addr, as)) | |
8b11a64c ZD |
3600 | break; |
3601 | } | |
d6fa2d17 | 3602 | data->min_offset = (i == -1? 0 : off); |
8b11a64c | 3603 | |
d6fa2d17 | 3604 | for (i = width; i >= 0; i--) |
8b11a64c | 3605 | { |
eb87c7c4 | 3606 | off = ((unsigned HOST_WIDE_INT) 1 << i) - 1; |
d6fa2d17 XDL |
3607 | XEXP (addr, 1) = gen_int_mode (off, address_mode); |
3608 | if (memory_address_addr_space_p (mem_mode, addr, as)) | |
8b11a64c | 3609 | break; |
5a33f47d EB |
3610 | /* For some strict-alignment targets, the offset must be naturally |
3611 | aligned. Try an aligned offset if mem_mode is not QImode. */ | |
3612 | off = mem_mode != QImode | |
3613 | ? ((unsigned HOST_WIDE_INT) 1 << i) | |
3614 | - GET_MODE_SIZE (mem_mode) | |
3615 | : 0; | |
b88fe5e9 ZC |
3616 | if (off > 0) |
3617 | { | |
3618 | XEXP (addr, 1) = gen_int_mode (off, address_mode); | |
3619 | if (memory_address_addr_space_p (mem_mode, addr, as)) | |
3620 | break; | |
3621 | } | |
8b11a64c | 3622 | } |
d6fa2d17 XDL |
3623 | if (i == -1) |
3624 | off = 0; | |
3625 | data->max_offset = off; | |
8b11a64c ZD |
3626 | |
3627 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3628 | { | |
3629 | fprintf (dump_file, "get_address_cost:\n"); | |
d6fa2d17 | 3630 | fprintf (dump_file, " min offset %s " HOST_WIDE_INT_PRINT_DEC "\n", |
dd5f63f8 | 3631 | GET_MODE_NAME (mem_mode), |
d6fa2d17 XDL |
3632 | data->min_offset); |
3633 | fprintf (dump_file, " max offset %s " HOST_WIDE_INT_PRINT_DEC "\n", | |
dd5f63f8 | 3634 | GET_MODE_NAME (mem_mode), |
d6fa2d17 | 3635 | data->max_offset); |
8b11a64c ZD |
3636 | } |
3637 | ||
09e881c9 | 3638 | rat = 1; |
ac182688 | 3639 | for (i = 2; i <= MAX_RATIO; i++) |
09e881c9 | 3640 | if (multiplier_allowed_in_address_p (i, mem_mode, as)) |
ac182688 | 3641 | { |
09e881c9 | 3642 | rat = i; |
ac182688 ZD |
3643 | break; |
3644 | } | |
975626a7 ZD |
3645 | |
3646 | /* Compute the cost of various addressing modes. */ | |
3647 | acost = 0; | |
d4ebfa65 BE |
3648 | reg0 = gen_raw_REG (address_mode, LAST_VIRTUAL_REGISTER + 1); |
3649 | reg1 = gen_raw_REG (address_mode, LAST_VIRTUAL_REGISTER + 2); | |
975626a7 | 3650 | |
cd047059 | 3651 | if (USE_LOAD_PRE_DECREMENT (mem_mode) |
8875e939 | 3652 | || USE_STORE_PRE_DECREMENT (mem_mode)) |
2c08497a | 3653 | { |
d4ebfa65 | 3654 | addr = gen_rtx_PRE_DEC (address_mode, reg0); |
09e881c9 BE |
3655 | has_predec[mem_mode] |
3656 | = memory_address_addr_space_p (mem_mode, addr, as); | |
cd047059 BC |
3657 | |
3658 | if (has_predec[mem_mode]) | |
3659 | data->ainc_costs[AINC_PRE_DEC] | |
3660 | = address_cost (addr, mem_mode, as, speed); | |
2c08497a | 3661 | } |
cd047059 | 3662 | if (USE_LOAD_POST_DECREMENT (mem_mode) |
8875e939 | 3663 | || USE_STORE_POST_DECREMENT (mem_mode)) |
2c08497a | 3664 | { |
d4ebfa65 | 3665 | addr = gen_rtx_POST_DEC (address_mode, reg0); |
09e881c9 BE |
3666 | has_postdec[mem_mode] |
3667 | = memory_address_addr_space_p (mem_mode, addr, as); | |
cd047059 BC |
3668 | |
3669 | if (has_postdec[mem_mode]) | |
3670 | data->ainc_costs[AINC_POST_DEC] | |
3671 | = address_cost (addr, mem_mode, as, speed); | |
2c08497a | 3672 | } |
cd047059 | 3673 | if (USE_LOAD_PRE_INCREMENT (mem_mode) |
8875e939 | 3674 | || USE_STORE_PRE_DECREMENT (mem_mode)) |
2c08497a | 3675 | { |
d4ebfa65 | 3676 | addr = gen_rtx_PRE_INC (address_mode, reg0); |
09e881c9 BE |
3677 | has_preinc[mem_mode] |
3678 | = memory_address_addr_space_p (mem_mode, addr, as); | |
cd047059 BC |
3679 | |
3680 | if (has_preinc[mem_mode]) | |
3681 | data->ainc_costs[AINC_PRE_INC] | |
3682 | = address_cost (addr, mem_mode, as, speed); | |
2c08497a | 3683 | } |
cd047059 | 3684 | if (USE_LOAD_POST_INCREMENT (mem_mode) |
8875e939 | 3685 | || USE_STORE_POST_INCREMENT (mem_mode)) |
2c08497a | 3686 | { |
d4ebfa65 | 3687 | addr = gen_rtx_POST_INC (address_mode, reg0); |
09e881c9 BE |
3688 | has_postinc[mem_mode] |
3689 | = memory_address_addr_space_p (mem_mode, addr, as); | |
cd047059 BC |
3690 | |
3691 | if (has_postinc[mem_mode]) | |
3692 | data->ainc_costs[AINC_POST_INC] | |
3693 | = address_cost (addr, mem_mode, as, speed); | |
2c08497a | 3694 | } |
975626a7 ZD |
3695 | for (i = 0; i < 16; i++) |
3696 | { | |
3697 | sym_p = i & 1; | |
3698 | var_p = (i >> 1) & 1; | |
3699 | off_p = (i >> 2) & 1; | |
3700 | rat_p = (i >> 3) & 1; | |
3701 | ||
3702 | addr = reg0; | |
3703 | if (rat_p) | |
d4ebfa65 BE |
3704 | addr = gen_rtx_fmt_ee (MULT, address_mode, addr, |
3705 | gen_int_mode (rat, address_mode)); | |
975626a7 ZD |
3706 | |
3707 | if (var_p) | |
d4ebfa65 | 3708 | addr = gen_rtx_fmt_ee (PLUS, address_mode, addr, reg1); |
975626a7 ZD |
3709 | |
3710 | if (sym_p) | |
3711 | { | |
d4ebfa65 | 3712 | base = gen_rtx_SYMBOL_REF (address_mode, ggc_strdup ("")); |
8318b0d9 | 3713 | /* ??? We can run into trouble with some backends by presenting |
fa10beec | 3714 | it with symbols which haven't been properly passed through |
8318b0d9 RH |
3715 | targetm.encode_section_info. By setting the local bit, we |
3716 | enhance the probability of things working. */ | |
3717 | SYMBOL_REF_FLAGS (base) = SYMBOL_FLAG_LOCAL; | |
3718 | ||
975626a7 | 3719 | if (off_p) |
d4ebfa65 | 3720 | base = gen_rtx_fmt_e (CONST, address_mode, |
09e881c9 | 3721 | gen_rtx_fmt_ee |
d4ebfa65 BE |
3722 | (PLUS, address_mode, base, |
3723 | gen_int_mode (off, address_mode))); | |
975626a7 ZD |
3724 | } |
3725 | else if (off_p) | |
d4ebfa65 | 3726 | base = gen_int_mode (off, address_mode); |
975626a7 ZD |
3727 | else |
3728 | base = NULL_RTX; | |
b8698a0f | 3729 | |
975626a7 | 3730 | if (base) |
d4ebfa65 | 3731 | addr = gen_rtx_fmt_ee (PLUS, address_mode, addr, base); |
2c08497a | 3732 | |
975626a7 ZD |
3733 | start_sequence (); |
3734 | /* To avoid splitting addressing modes, pretend that no cse will | |
3735 | follow. */ | |
3736 | old_cse_not_expected = cse_not_expected; | |
3737 | cse_not_expected = true; | |
09e881c9 | 3738 | addr = memory_address_addr_space (mem_mode, addr, as); |
975626a7 ZD |
3739 | cse_not_expected = old_cse_not_expected; |
3740 | seq = get_insns (); | |
3741 | end_sequence (); | |
3742 | ||
f40751dd | 3743 | acost = seq_cost (seq, speed); |
09e881c9 | 3744 | acost += address_cost (addr, mem_mode, as, speed); |
975626a7 ZD |
3745 | |
3746 | if (!acost) | |
3747 | acost = 1; | |
09e881c9 | 3748 | data->costs[sym_p][var_p][off_p][rat_p] = acost; |
975626a7 ZD |
3749 | } |
3750 | ||
3751 | /* On some targets, it is quite expensive to load symbol to a register, | |
3752 | which makes addresses that contain symbols look much more expensive. | |
3753 | However, the symbol will have to be loaded in any case before the | |
3754 | loop (and quite likely we have it in register already), so it does not | |
3755 | make much sense to penalize them too heavily. So make some final | |
3756 | tweaks for the SYMBOL_PRESENT modes: | |
3757 | ||
3758 | If VAR_PRESENT is false, and the mode obtained by changing symbol to | |
3759 | var is cheaper, use this mode with small penalty. | |
3760 | If VAR_PRESENT is true, try whether the mode with | |
3761 | SYMBOL_PRESENT = false is cheaper even with cost of addition, and | |
3762 | if this is the case, use it. */ | |
5322d07e | 3763 | add_c = add_cost (speed, address_mode); |
975626a7 ZD |
3764 | for (i = 0; i < 8; i++) |
3765 | { | |
3766 | var_p = i & 1; | |
3767 | off_p = (i >> 1) & 1; | |
3768 | rat_p = (i >> 2) & 1; | |
3769 | ||
09e881c9 | 3770 | acost = data->costs[0][1][off_p][rat_p] + 1; |
975626a7 ZD |
3771 | if (var_p) |
3772 | acost += add_c; | |
3773 | ||
09e881c9 BE |
3774 | if (acost < data->costs[1][var_p][off_p][rat_p]) |
3775 | data->costs[1][var_p][off_p][rat_p] = acost; | |
975626a7 | 3776 | } |
2c08497a | 3777 | |
975626a7 ZD |
3778 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3779 | { | |
3780 | fprintf (dump_file, "Address costs:\n"); | |
b8698a0f | 3781 | |
975626a7 ZD |
3782 | for (i = 0; i < 16; i++) |
3783 | { | |
3784 | sym_p = i & 1; | |
3785 | var_p = (i >> 1) & 1; | |
3786 | off_p = (i >> 2) & 1; | |
3787 | rat_p = (i >> 3) & 1; | |
3788 | ||
3789 | fprintf (dump_file, " "); | |
3790 | if (sym_p) | |
3791 | fprintf (dump_file, "sym + "); | |
3792 | if (var_p) | |
3793 | fprintf (dump_file, "var + "); | |
3794 | if (off_p) | |
3795 | fprintf (dump_file, "cst + "); | |
3796 | if (rat_p) | |
3797 | fprintf (dump_file, "rat * "); | |
3798 | ||
09e881c9 | 3799 | acost = data->costs[sym_p][var_p][off_p][rat_p]; |
975626a7 ZD |
3800 | fprintf (dump_file, "index costs %d\n", acost); |
3801 | } | |
2c08497a BS |
3802 | if (has_predec[mem_mode] || has_postdec[mem_mode] |
3803 | || has_preinc[mem_mode] || has_postinc[mem_mode]) | |
3804 | fprintf (dump_file, " May include autoinc/dec\n"); | |
975626a7 ZD |
3805 | fprintf (dump_file, "\n"); |
3806 | } | |
09e881c9 | 3807 | |
9771b263 | 3808 | address_cost_data_list[data_index] = data; |
8b11a64c ZD |
3809 | } |
3810 | ||
d4ebfa65 | 3811 | bits = GET_MODE_BITSIZE (address_mode); |
8b11a64c ZD |
3812 | mask = ~(~(unsigned HOST_WIDE_INT) 0 << (bits - 1) << 1); |
3813 | offset &= mask; | |
3814 | if ((offset >> (bits - 1) & 1)) | |
3815 | offset |= ~mask; | |
3816 | s_offset = offset; | |
3817 | ||
2c08497a | 3818 | autoinc = false; |
cd047059 | 3819 | autoinc_type = AINC_NONE; |
2c08497a BS |
3820 | msize = GET_MODE_SIZE (mem_mode); |
3821 | autoinc_offset = offset; | |
3822 | if (stmt_after_inc) | |
3823 | autoinc_offset += ratio * cstep; | |
3824 | if (symbol_present || var_present || ratio != 1) | |
3825 | autoinc = false; | |
cd047059 BC |
3826 | else |
3827 | { | |
3828 | if (has_postinc[mem_mode] && autoinc_offset == 0 | |
3829 | && msize == cstep) | |
3830 | autoinc_type = AINC_POST_INC; | |
3831 | else if (has_postdec[mem_mode] && autoinc_offset == 0 | |
2c08497a | 3832 | && msize == -cstep) |
cd047059 BC |
3833 | autoinc_type = AINC_POST_DEC; |
3834 | else if (has_preinc[mem_mode] && autoinc_offset == msize | |
2c08497a | 3835 | && msize == cstep) |
cd047059 BC |
3836 | autoinc_type = AINC_PRE_INC; |
3837 | else if (has_predec[mem_mode] && autoinc_offset == -msize | |
3838 | && msize == -cstep) | |
3839 | autoinc_type = AINC_PRE_DEC; | |
3840 | ||
3841 | if (autoinc_type != AINC_NONE) | |
3842 | autoinc = true; | |
3843 | } | |
2c08497a | 3844 | |
8b11a64c | 3845 | cost = 0; |
a85a9024 | 3846 | offset_p = (s_offset != 0 |
09e881c9 BE |
3847 | && data->min_offset <= s_offset |
3848 | && s_offset <= data->max_offset); | |
8b11a64c | 3849 | ratio_p = (ratio != 1 |
09e881c9 | 3850 | && multiplier_allowed_in_address_p (ratio, mem_mode, as)); |
8b11a64c ZD |
3851 | |
3852 | if (ratio != 1 && !ratio_p) | |
6dd8f4bb | 3853 | cost += mult_by_coeff_cost (ratio, address_mode, speed); |
8b11a64c ZD |
3854 | |
3855 | if (s_offset && !offset_p && !symbol_present) | |
5322d07e | 3856 | cost += add_cost (speed, address_mode); |
8b11a64c | 3857 | |
2c08497a BS |
3858 | if (may_autoinc) |
3859 | *may_autoinc = autoinc; | |
cd047059 BC |
3860 | if (autoinc) |
3861 | acost = data->ainc_costs[autoinc_type]; | |
3862 | else | |
3863 | acost = data->costs[symbol_present][var_present][offset_p][ratio_p]; | |
6e8c65f6 ZD |
3864 | complexity = (symbol_present != 0) + (var_present != 0) + offset_p + ratio_p; |
3865 | return new_cost (cost + acost, complexity); | |
8b11a64c | 3866 | } |
8b11a64c | 3867 | |
e6450c11 TV |
3868 | /* Calculate the SPEED or size cost of shiftadd EXPR in MODE. MULT is the |
3869 | the EXPR operand holding the shift. COST0 and COST1 are the costs for | |
3870 | calculating the operands of EXPR. Returns true if successful, and returns | |
3871 | the cost in COST. */ | |
3872 | ||
3873 | static bool | |
ef4bddc2 | 3874 | get_shiftadd_cost (tree expr, machine_mode mode, comp_cost cost0, |
e6450c11 TV |
3875 | comp_cost cost1, tree mult, bool speed, comp_cost *cost) |
3876 | { | |
3877 | comp_cost res; | |
3878 | tree op1 = TREE_OPERAND (expr, 1); | |
3879 | tree cst = TREE_OPERAND (mult, 1); | |
3880 | tree multop = TREE_OPERAND (mult, 0); | |
3881 | int m = exact_log2 (int_cst_value (cst)); | |
3882 | int maxm = MIN (BITS_PER_WORD, GET_MODE_BITSIZE (mode)); | |
9fb87eb0 EB |
3883 | int as_cost, sa_cost; |
3884 | bool mult_in_op1; | |
e6450c11 TV |
3885 | |
3886 | if (!(m >= 0 && m < maxm)) | |
3887 | return false; | |
3888 | ||
9fb87eb0 | 3889 | mult_in_op1 = operand_equal_p (op1, mult, 0); |
1a1a5534 | 3890 | |
9fb87eb0 EB |
3891 | as_cost = add_cost (speed, mode) + shift_cost (speed, mode, m); |
3892 | ||
3893 | /* If the target has a cheap shift-and-add or shift-and-sub instruction, | |
3894 | use that in preference to a shift insn followed by an add insn. */ | |
e6450c11 | 3895 | sa_cost = (TREE_CODE (expr) != MINUS_EXPR |
5322d07e | 3896 | ? shiftadd_cost (speed, mode, m) |
9fb87eb0 | 3897 | : (mult_in_op1 |
5322d07e NF |
3898 | ? shiftsub1_cost (speed, mode, m) |
3899 | : shiftsub0_cost (speed, mode, m))); | |
9fb87eb0 EB |
3900 | |
3901 | res = new_cost (MIN (as_cost, sa_cost), 0); | |
3902 | res = add_costs (res, mult_in_op1 ? cost0 : cost1); | |
e6450c11 TV |
3903 | |
3904 | STRIP_NOPS (multop); | |
3905 | if (!is_gimple_val (multop)) | |
3906 | res = add_costs (res, force_expr_to_var_cost (multop, speed)); | |
3907 | ||
3908 | *cost = res; | |
3909 | return true; | |
3910 | } | |
3911 | ||
3ac01fde ZD |
3912 | /* Estimates cost of forcing expression EXPR into a variable. */ |
3913 | ||
6e8c65f6 | 3914 | static comp_cost |
f40751dd | 3915 | force_expr_to_var_cost (tree expr, bool speed) |
8b11a64c ZD |
3916 | { |
3917 | static bool costs_initialized = false; | |
f40751dd JH |
3918 | static unsigned integer_cost [2]; |
3919 | static unsigned symbol_cost [2]; | |
3920 | static unsigned address_cost [2]; | |
7299dbfb | 3921 | tree op0, op1; |
6e8c65f6 | 3922 | comp_cost cost0, cost1, cost; |
ef4bddc2 | 3923 | machine_mode mode; |
8b11a64c ZD |
3924 | |
3925 | if (!costs_initialized) | |
3926 | { | |
8b11a64c | 3927 | tree type = build_pointer_type (integer_type_node); |
8318b0d9 RH |
3928 | tree var, addr; |
3929 | rtx x; | |
f40751dd | 3930 | int i; |
8318b0d9 RH |
3931 | |
3932 | var = create_tmp_var_raw (integer_type_node, "test_var"); | |
3933 | TREE_STATIC (var) = 1; | |
3934 | x = produce_memory_decl_rtl (var, NULL); | |
3935 | SET_DECL_RTL (var, x); | |
8b11a64c | 3936 | |
8b11a64c | 3937 | addr = build1 (ADDR_EXPR, type, var); |
8b11a64c | 3938 | |
f40751dd JH |
3939 | |
3940 | for (i = 0; i < 2; i++) | |
8b11a64c | 3941 | { |
f40751dd JH |
3942 | integer_cost[i] = computation_cost (build_int_cst (integer_type_node, |
3943 | 2000), i); | |
3944 | ||
3945 | symbol_cost[i] = computation_cost (addr, i) + 1; | |
3946 | ||
3947 | address_cost[i] | |
5d49b6a7 | 3948 | = computation_cost (fold_build_pointer_plus_hwi (addr, 2000), i) + 1; |
f40751dd JH |
3949 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3950 | { | |
3951 | fprintf (dump_file, "force_expr_to_var_cost %s costs:\n", i ? "speed" : "size"); | |
3952 | fprintf (dump_file, " integer %d\n", (int) integer_cost[i]); | |
3953 | fprintf (dump_file, " symbol %d\n", (int) symbol_cost[i]); | |
3954 | fprintf (dump_file, " address %d\n", (int) address_cost[i]); | |
3955 | fprintf (dump_file, " other %d\n", (int) target_spill_cost[i]); | |
3956 | fprintf (dump_file, "\n"); | |
3957 | } | |
8b11a64c ZD |
3958 | } |
3959 | ||
3960 | costs_initialized = true; | |
3961 | } | |
3962 | ||
f5e2738c ZD |
3963 | STRIP_NOPS (expr); |
3964 | ||
8b11a64c | 3965 | if (SSA_VAR_P (expr)) |
7735d6c7 | 3966 | return no_cost; |
8b11a64c | 3967 | |
ad6003f2 | 3968 | if (is_gimple_min_invariant (expr)) |
8b11a64c ZD |
3969 | { |
3970 | if (TREE_CODE (expr) == INTEGER_CST) | |
f40751dd | 3971 | return new_cost (integer_cost [speed], 0); |
8b11a64c ZD |
3972 | |
3973 | if (TREE_CODE (expr) == ADDR_EXPR) | |
3974 | { | |
3975 | tree obj = TREE_OPERAND (expr, 0); | |
3976 | ||
3977 | if (TREE_CODE (obj) == VAR_DECL | |
3978 | || TREE_CODE (obj) == PARM_DECL | |
3979 | || TREE_CODE (obj) == RESULT_DECL) | |
f40751dd | 3980 | return new_cost (symbol_cost [speed], 0); |
8b11a64c ZD |
3981 | } |
3982 | ||
f40751dd | 3983 | return new_cost (address_cost [speed], 0); |
8b11a64c ZD |
3984 | } |
3985 | ||
7299dbfb ZD |
3986 | switch (TREE_CODE (expr)) |
3987 | { | |
5be014d5 | 3988 | case POINTER_PLUS_EXPR: |
7299dbfb ZD |
3989 | case PLUS_EXPR: |
3990 | case MINUS_EXPR: | |
3991 | case MULT_EXPR: | |
3992 | op0 = TREE_OPERAND (expr, 0); | |
3993 | op1 = TREE_OPERAND (expr, 1); | |
f5e2738c ZD |
3994 | STRIP_NOPS (op0); |
3995 | STRIP_NOPS (op1); | |
7299dbfb ZD |
3996 | break; |
3997 | ||
801a661c | 3998 | CASE_CONVERT: |
7a2faca1 EB |
3999 | case NEGATE_EXPR: |
4000 | op0 = TREE_OPERAND (expr, 0); | |
4001 | STRIP_NOPS (op0); | |
4002 | op1 = NULL_TREE; | |
7a2faca1 EB |
4003 | break; |
4004 | ||
7299dbfb ZD |
4005 | default: |
4006 | /* Just an arbitrary value, FIXME. */ | |
f40751dd | 4007 | return new_cost (target_spill_cost[speed], 0); |
7299dbfb ZD |
4008 | } |
4009 | ||
801a661c BC |
4010 | if (op0 == NULL_TREE |
4011 | || TREE_CODE (op0) == SSA_NAME || CONSTANT_CLASS_P (op0)) | |
4012 | cost0 = no_cost; | |
4013 | else | |
4014 | cost0 = force_expr_to_var_cost (op0, speed); | |
4015 | ||
4016 | if (op1 == NULL_TREE | |
4017 | || TREE_CODE (op1) == SSA_NAME || CONSTANT_CLASS_P (op1)) | |
4018 | cost1 = no_cost; | |
4019 | else | |
4020 | cost1 = force_expr_to_var_cost (op1, speed); | |
4021 | ||
7299dbfb ZD |
4022 | mode = TYPE_MODE (TREE_TYPE (expr)); |
4023 | switch (TREE_CODE (expr)) | |
4024 | { | |
5be014d5 | 4025 | case POINTER_PLUS_EXPR: |
7299dbfb ZD |
4026 | case PLUS_EXPR: |
4027 | case MINUS_EXPR: | |
7a2faca1 | 4028 | case NEGATE_EXPR: |
5322d07e | 4029 | cost = new_cost (add_cost (speed, mode), 0); |
e6450c11 TV |
4030 | if (TREE_CODE (expr) != NEGATE_EXPR) |
4031 | { | |
4032 | tree mult = NULL_TREE; | |
4033 | comp_cost sa_cost; | |
4034 | if (TREE_CODE (op1) == MULT_EXPR) | |
4035 | mult = op1; | |
4036 | else if (TREE_CODE (op0) == MULT_EXPR) | |
4037 | mult = op0; | |
4038 | ||
4039 | if (mult != NULL_TREE | |
50ad7db2 | 4040 | && cst_and_fits_in_hwi (TREE_OPERAND (mult, 1)) |
5322d07e NF |
4041 | && get_shiftadd_cost (expr, mode, cost0, cost1, mult, |
4042 | speed, &sa_cost)) | |
e6450c11 TV |
4043 | return sa_cost; |
4044 | } | |
7299dbfb ZD |
4045 | break; |
4046 | ||
801a661c BC |
4047 | CASE_CONVERT: |
4048 | { | |
4049 | tree inner_mode, outer_mode; | |
4050 | outer_mode = TREE_TYPE (expr); | |
4051 | inner_mode = TREE_TYPE (op0); | |
4052 | cost = new_cost (convert_cost (TYPE_MODE (outer_mode), | |
4053 | TYPE_MODE (inner_mode), speed), 0); | |
4054 | } | |
4055 | break; | |
4056 | ||
7299dbfb ZD |
4057 | case MULT_EXPR: |
4058 | if (cst_and_fits_in_hwi (op0)) | |
6dd8f4bb BS |
4059 | cost = new_cost (mult_by_coeff_cost (int_cst_value (op0), |
4060 | mode, speed), 0); | |
b8698a0f | 4061 | else if (cst_and_fits_in_hwi (op1)) |
6dd8f4bb BS |
4062 | cost = new_cost (mult_by_coeff_cost (int_cst_value (op1), |
4063 | mode, speed), 0); | |
7299dbfb | 4064 | else |
f40751dd | 4065 | return new_cost (target_spill_cost [speed], 0); |
7299dbfb ZD |
4066 | break; |
4067 | ||
4068 | default: | |
4069 | gcc_unreachable (); | |
4070 | } | |
4071 | ||
6e8c65f6 ZD |
4072 | cost = add_costs (cost, cost0); |
4073 | cost = add_costs (cost, cost1); | |
7299dbfb ZD |
4074 | |
4075 | /* Bound the cost by target_spill_cost. The parts of complicated | |
4076 | computations often are either loop invariant or at least can | |
4077 | be shared between several iv uses, so letting this grow without | |
4078 | limits would not give reasonable results. */ | |
2c08497a | 4079 | if (cost.cost > (int) target_spill_cost [speed]) |
f40751dd | 4080 | cost.cost = target_spill_cost [speed]; |
6e8c65f6 ZD |
4081 | |
4082 | return cost; | |
8b11a64c ZD |
4083 | } |
4084 | ||
3ac01fde ZD |
4085 | /* Estimates cost of forcing EXPR into a variable. DEPENDS_ON is a set of the |
4086 | invariants the computation depends on. */ | |
4087 | ||
6e8c65f6 | 4088 | static comp_cost |
3ac01fde ZD |
4089 | force_var_cost (struct ivopts_data *data, |
4090 | tree expr, bitmap *depends_on) | |
4091 | { | |
4092 | if (depends_on) | |
4093 | { | |
4094 | fd_ivopts_data = data; | |
4095 | walk_tree (&expr, find_depends, depends_on, NULL); | |
4096 | } | |
4097 | ||
f40751dd | 4098 | return force_expr_to_var_cost (expr, data->speed); |
3ac01fde ZD |
4099 | } |
4100 | ||
8b11a64c ZD |
4101 | /* Estimates cost of expressing address ADDR as var + symbol + offset. The |
4102 | value of offset is added to OFFSET, SYMBOL_PRESENT and VAR_PRESENT are set | |
4103 | to false if the corresponding part is missing. DEPENDS_ON is a set of the | |
4104 | invariants the computation depends on. */ | |
4105 | ||
6e8c65f6 | 4106 | static comp_cost |
8b11a64c ZD |
4107 | split_address_cost (struct ivopts_data *data, |
4108 | tree addr, bool *symbol_present, bool *var_present, | |
4109 | unsigned HOST_WIDE_INT *offset, bitmap *depends_on) | |
4110 | { | |
2f4675b4 ZD |
4111 | tree core; |
4112 | HOST_WIDE_INT bitsize; | |
4113 | HOST_WIDE_INT bitpos; | |
4114 | tree toffset; | |
ef4bddc2 | 4115 | machine_mode mode; |
2f4675b4 | 4116 | int unsignedp, volatilep; |
b8698a0f | 4117 | |
2f4675b4 | 4118 | core = get_inner_reference (addr, &bitsize, &bitpos, &toffset, &mode, |
b3ecff82 | 4119 | &unsignedp, &volatilep, false); |
8b11a64c | 4120 | |
2f4675b4 ZD |
4121 | if (toffset != 0 |
4122 | || bitpos % BITS_PER_UNIT != 0 | |
4123 | || TREE_CODE (core) != VAR_DECL) | |
8b11a64c ZD |
4124 | { |
4125 | *symbol_present = false; | |
4126 | *var_present = true; | |
4127 | fd_ivopts_data = data; | |
4128 | walk_tree (&addr, find_depends, depends_on, NULL); | |
f40751dd | 4129 | return new_cost (target_spill_cost[data->speed], 0); |
2f4675b4 ZD |
4130 | } |
4131 | ||
4132 | *offset += bitpos / BITS_PER_UNIT; | |
8b11a64c ZD |
4133 | if (TREE_STATIC (core) |
4134 | || DECL_EXTERNAL (core)) | |
4135 | { | |
4136 | *symbol_present = true; | |
4137 | *var_present = false; | |
7735d6c7 | 4138 | return no_cost; |
8b11a64c | 4139 | } |
b8698a0f | 4140 | |
8b11a64c ZD |
4141 | *symbol_present = false; |
4142 | *var_present = true; | |
7735d6c7 | 4143 | return no_cost; |
8b11a64c ZD |
4144 | } |
4145 | ||
4146 | /* Estimates cost of expressing difference of addresses E1 - E2 as | |
4147 | var + symbol + offset. The value of offset is added to OFFSET, | |
4148 | SYMBOL_PRESENT and VAR_PRESENT are set to false if the corresponding | |
4149 | part is missing. DEPENDS_ON is a set of the invariants the computation | |
4150 | depends on. */ | |
4151 | ||
6e8c65f6 | 4152 | static comp_cost |
8b11a64c ZD |
4153 | ptr_difference_cost (struct ivopts_data *data, |
4154 | tree e1, tree e2, bool *symbol_present, bool *var_present, | |
4155 | unsigned HOST_WIDE_INT *offset, bitmap *depends_on) | |
4156 | { | |
2f4675b4 | 4157 | HOST_WIDE_INT diff = 0; |
7a2faca1 EB |
4158 | aff_tree aff_e1, aff_e2; |
4159 | tree type; | |
8b11a64c | 4160 | |
1e128c5f | 4161 | gcc_assert (TREE_CODE (e1) == ADDR_EXPR); |
8b11a64c | 4162 | |
7299dbfb | 4163 | if (ptr_difference_const (e1, e2, &diff)) |
8b11a64c ZD |
4164 | { |
4165 | *offset += diff; | |
4166 | *symbol_present = false; | |
4167 | *var_present = false; | |
7735d6c7 | 4168 | return no_cost; |
8b11a64c ZD |
4169 | } |
4170 | ||
6e8c65f6 | 4171 | if (integer_zerop (e2)) |
8b11a64c ZD |
4172 | return split_address_cost (data, TREE_OPERAND (e1, 0), |
4173 | symbol_present, var_present, offset, depends_on); | |
4174 | ||
4175 | *symbol_present = false; | |
4176 | *var_present = true; | |
8b11a64c | 4177 | |
7a2faca1 EB |
4178 | type = signed_type_for (TREE_TYPE (e1)); |
4179 | tree_to_aff_combination (e1, type, &aff_e1); | |
4180 | tree_to_aff_combination (e2, type, &aff_e2); | |
807e902e | 4181 | aff_combination_scale (&aff_e2, -1); |
7a2faca1 EB |
4182 | aff_combination_add (&aff_e1, &aff_e2); |
4183 | ||
4184 | return force_var_cost (data, aff_combination_to_tree (&aff_e1), depends_on); | |
8b11a64c ZD |
4185 | } |
4186 | ||
4187 | /* Estimates cost of expressing difference E1 - E2 as | |
4188 | var + symbol + offset. The value of offset is added to OFFSET, | |
4189 | SYMBOL_PRESENT and VAR_PRESENT are set to false if the corresponding | |
4190 | part is missing. DEPENDS_ON is a set of the invariants the computation | |
4191 | depends on. */ | |
4192 | ||
6e8c65f6 | 4193 | static comp_cost |
8b11a64c ZD |
4194 | difference_cost (struct ivopts_data *data, |
4195 | tree e1, tree e2, bool *symbol_present, bool *var_present, | |
4196 | unsigned HOST_WIDE_INT *offset, bitmap *depends_on) | |
4197 | { | |
ef4bddc2 | 4198 | machine_mode mode = TYPE_MODE (TREE_TYPE (e1)); |
f5e2738c | 4199 | unsigned HOST_WIDE_INT off1, off2; |
7a2faca1 EB |
4200 | aff_tree aff_e1, aff_e2; |
4201 | tree type; | |
f5e2738c | 4202 | |
9be872b7 ZD |
4203 | e1 = strip_offset (e1, &off1); |
4204 | e2 = strip_offset (e2, &off2); | |
f5e2738c | 4205 | *offset += off1 - off2; |
8b11a64c | 4206 | |
f5e2738c ZD |
4207 | STRIP_NOPS (e1); |
4208 | STRIP_NOPS (e2); | |
8b11a64c ZD |
4209 | |
4210 | if (TREE_CODE (e1) == ADDR_EXPR) | |
7a2faca1 EB |
4211 | return ptr_difference_cost (data, e1, e2, symbol_present, var_present, |
4212 | offset, depends_on); | |
8b11a64c ZD |
4213 | *symbol_present = false; |
4214 | ||
4215 | if (operand_equal_p (e1, e2, 0)) | |
4216 | { | |
4217 | *var_present = false; | |
7735d6c7 | 4218 | return no_cost; |
8b11a64c | 4219 | } |
7a2faca1 | 4220 | |
8b11a64c | 4221 | *var_present = true; |
7a2faca1 | 4222 | |
6e682d7e | 4223 | if (integer_zerop (e2)) |
8b11a64c ZD |
4224 | return force_var_cost (data, e1, depends_on); |
4225 | ||
6e682d7e | 4226 | if (integer_zerop (e1)) |
8b11a64c | 4227 | { |
7a2faca1 | 4228 | comp_cost cost = force_var_cost (data, e2, depends_on); |
6dd8f4bb | 4229 | cost.cost += mult_by_coeff_cost (-1, mode, data->speed); |
8b11a64c ZD |
4230 | return cost; |
4231 | } | |
4232 | ||
7a2faca1 EB |
4233 | type = signed_type_for (TREE_TYPE (e1)); |
4234 | tree_to_aff_combination (e1, type, &aff_e1); | |
4235 | tree_to_aff_combination (e2, type, &aff_e2); | |
807e902e | 4236 | aff_combination_scale (&aff_e2, -1); |
7a2faca1 | 4237 | aff_combination_add (&aff_e1, &aff_e2); |
8b11a64c | 4238 | |
7a2faca1 | 4239 | return force_var_cost (data, aff_combination_to_tree (&aff_e1), depends_on); |
8b11a64c ZD |
4240 | } |
4241 | ||
18081149 XDL |
4242 | /* Returns true if AFF1 and AFF2 are identical. */ |
4243 | ||
4244 | static bool | |
4245 | compare_aff_trees (aff_tree *aff1, aff_tree *aff2) | |
4246 | { | |
4247 | unsigned i; | |
4248 | ||
4249 | if (aff1->n != aff2->n) | |
4250 | return false; | |
4251 | ||
4252 | for (i = 0; i < aff1->n; i++) | |
4253 | { | |
27bcd47c | 4254 | if (aff1->elts[i].coef != aff2->elts[i].coef) |
18081149 XDL |
4255 | return false; |
4256 | ||
4257 | if (!operand_equal_p (aff1->elts[i].val, aff2->elts[i].val, 0)) | |
4258 | return false; | |
4259 | } | |
4260 | return true; | |
4261 | } | |
4262 | ||
bb8d292d TV |
4263 | /* Stores EXPR in DATA->inv_expr_tab, and assigns it an inv_expr_id. */ |
4264 | ||
4265 | static int | |
4266 | get_expr_id (struct ivopts_data *data, tree expr) | |
4267 | { | |
4268 | struct iv_inv_expr_ent ent; | |
4269 | struct iv_inv_expr_ent **slot; | |
4270 | ||
4271 | ent.expr = expr; | |
4272 | ent.hash = iterative_hash_expr (expr, 0); | |
c203e8a7 | 4273 | slot = data->inv_expr_tab->find_slot (&ent, INSERT); |
bb8d292d TV |
4274 | if (*slot) |
4275 | return (*slot)->id; | |
4276 | ||
4277 | *slot = XNEW (struct iv_inv_expr_ent); | |
4278 | (*slot)->expr = expr; | |
4279 | (*slot)->hash = ent.hash; | |
4280 | (*slot)->id = data->inv_expr_id++; | |
4281 | return (*slot)->id; | |
4282 | } | |
4283 | ||
18081149 XDL |
4284 | /* Returns the pseudo expr id if expression UBASE - RATIO * CBASE |
4285 | requires a new compiler generated temporary. Returns -1 otherwise. | |
4286 | ADDRESS_P is a flag indicating if the expression is for address | |
4287 | computation. */ | |
4288 | ||
4289 | static int | |
4290 | get_loop_invariant_expr_id (struct ivopts_data *data, tree ubase, | |
4291 | tree cbase, HOST_WIDE_INT ratio, | |
4292 | bool address_p) | |
4293 | { | |
4294 | aff_tree ubase_aff, cbase_aff; | |
4295 | tree expr, ub, cb; | |
18081149 XDL |
4296 | |
4297 | STRIP_NOPS (ubase); | |
4298 | STRIP_NOPS (cbase); | |
4299 | ub = ubase; | |
4300 | cb = cbase; | |
4301 | ||
4302 | if ((TREE_CODE (ubase) == INTEGER_CST) | |
4303 | && (TREE_CODE (cbase) == INTEGER_CST)) | |
4304 | return -1; | |
4305 | ||
4306 | /* Strips the constant part. */ | |
4307 | if (TREE_CODE (ubase) == PLUS_EXPR | |
4308 | || TREE_CODE (ubase) == MINUS_EXPR | |
4309 | || TREE_CODE (ubase) == POINTER_PLUS_EXPR) | |
4310 | { | |
4311 | if (TREE_CODE (TREE_OPERAND (ubase, 1)) == INTEGER_CST) | |
4312 | ubase = TREE_OPERAND (ubase, 0); | |
4313 | } | |
4314 | ||
4315 | /* Strips the constant part. */ | |
4316 | if (TREE_CODE (cbase) == PLUS_EXPR | |
4317 | || TREE_CODE (cbase) == MINUS_EXPR | |
4318 | || TREE_CODE (cbase) == POINTER_PLUS_EXPR) | |
4319 | { | |
4320 | if (TREE_CODE (TREE_OPERAND (cbase, 1)) == INTEGER_CST) | |
4321 | cbase = TREE_OPERAND (cbase, 0); | |
4322 | } | |
4323 | ||
4324 | if (address_p) | |
4325 | { | |
4326 | if (((TREE_CODE (ubase) == SSA_NAME) | |
4327 | || (TREE_CODE (ubase) == ADDR_EXPR | |
4328 | && is_gimple_min_invariant (ubase))) | |
4329 | && (TREE_CODE (cbase) == INTEGER_CST)) | |
4330 | return -1; | |
4331 | ||
4332 | if (((TREE_CODE (cbase) == SSA_NAME) | |
4333 | || (TREE_CODE (cbase) == ADDR_EXPR | |
4334 | && is_gimple_min_invariant (cbase))) | |
4335 | && (TREE_CODE (ubase) == INTEGER_CST)) | |
4336 | return -1; | |
4337 | } | |
4338 | ||
4339 | if (ratio == 1) | |
4340 | { | |
c3284718 | 4341 | if (operand_equal_p (ubase, cbase, 0)) |
18081149 XDL |
4342 | return -1; |
4343 | ||
4344 | if (TREE_CODE (ubase) == ADDR_EXPR | |
4345 | && TREE_CODE (cbase) == ADDR_EXPR) | |
4346 | { | |
4347 | tree usym, csym; | |
4348 | ||
4349 | usym = TREE_OPERAND (ubase, 0); | |
4350 | csym = TREE_OPERAND (cbase, 0); | |
4351 | if (TREE_CODE (usym) == ARRAY_REF) | |
4352 | { | |
4353 | tree ind = TREE_OPERAND (usym, 1); | |
4354 | if (TREE_CODE (ind) == INTEGER_CST | |
9541ffee | 4355 | && tree_fits_shwi_p (ind) |
eb1ce453 | 4356 | && tree_to_shwi (ind) == 0) |
18081149 XDL |
4357 | usym = TREE_OPERAND (usym, 0); |
4358 | } | |
4359 | if (TREE_CODE (csym) == ARRAY_REF) | |
4360 | { | |
4361 | tree ind = TREE_OPERAND (csym, 1); | |
4362 | if (TREE_CODE (ind) == INTEGER_CST | |
9541ffee | 4363 | && tree_fits_shwi_p (ind) |
eb1ce453 | 4364 | && tree_to_shwi (ind) == 0) |
18081149 XDL |
4365 | csym = TREE_OPERAND (csym, 0); |
4366 | } | |
4367 | if (operand_equal_p (usym, csym, 0)) | |
4368 | return -1; | |
4369 | } | |
4370 | /* Now do more complex comparison */ | |
4371 | tree_to_aff_combination (ubase, TREE_TYPE (ubase), &ubase_aff); | |
4372 | tree_to_aff_combination (cbase, TREE_TYPE (cbase), &cbase_aff); | |
4373 | if (compare_aff_trees (&ubase_aff, &cbase_aff)) | |
4374 | return -1; | |
4375 | } | |
4376 | ||
4377 | tree_to_aff_combination (ub, TREE_TYPE (ub), &ubase_aff); | |
4378 | tree_to_aff_combination (cb, TREE_TYPE (cb), &cbase_aff); | |
4379 | ||
807e902e | 4380 | aff_combination_scale (&cbase_aff, -1 * ratio); |
18081149 XDL |
4381 | aff_combination_add (&ubase_aff, &cbase_aff); |
4382 | expr = aff_combination_to_tree (&ubase_aff); | |
bb8d292d | 4383 | return get_expr_id (data, expr); |
18081149 XDL |
4384 | } |
4385 | ||
4386 | ||
4387 | ||
8b11a64c ZD |
4388 | /* Determines the cost of the computation by that USE is expressed |
4389 | from induction variable CAND. If ADDRESS_P is true, we just need | |
4390 | to create an address from it, otherwise we want to get it into | |
4391 | register. A set of invariants we depend on is stored in | |
2c08497a BS |
4392 | DEPENDS_ON. AT is the statement at that the value is computed. |
4393 | If CAN_AUTOINC is nonnull, use it to record whether autoinc | |
4394 | addressing is likely. */ | |
8b11a64c | 4395 | |
6e8c65f6 | 4396 | static comp_cost |
8b11a64c ZD |
4397 | get_computation_cost_at (struct ivopts_data *data, |
4398 | struct iv_use *use, struct iv_cand *cand, | |
2c08497a | 4399 | bool address_p, bitmap *depends_on, gimple at, |
18081149 XDL |
4400 | bool *can_autoinc, |
4401 | int *inv_expr_id) | |
8b11a64c ZD |
4402 | { |
4403 | tree ubase = use->iv->base, ustep = use->iv->step; | |
4404 | tree cbase, cstep; | |
4405 | tree utype = TREE_TYPE (ubase), ctype; | |
73f30c63 | 4406 | unsigned HOST_WIDE_INT cstepi, offset = 0; |
8b11a64c | 4407 | HOST_WIDE_INT ratio, aratio; |
2c08497a | 4408 | bool var_present, symbol_present, stmt_is_after_inc; |
6e8c65f6 | 4409 | comp_cost cost; |
807e902e | 4410 | widest_int rat; |
f40751dd | 4411 | bool speed = optimize_bb_for_speed_p (gimple_bb (at)); |
ef4bddc2 | 4412 | machine_mode mem_mode = (address_p |
0c604a61 TV |
4413 | ? TYPE_MODE (TREE_TYPE (*use->op_p)) |
4414 | : VOIDmode); | |
8b11a64c ZD |
4415 | |
4416 | *depends_on = NULL; | |
4417 | ||
4418 | /* Only consider real candidates. */ | |
4419 | if (!cand->iv) | |
6e8c65f6 | 4420 | return infinite_cost; |
8b11a64c ZD |
4421 | |
4422 | cbase = cand->iv->base; | |
4423 | cstep = cand->iv->step; | |
4424 | ctype = TREE_TYPE (cbase); | |
4425 | ||
4426 | if (TYPE_PRECISION (utype) > TYPE_PRECISION (ctype)) | |
4427 | { | |
4428 | /* We do not have a precision to express the values of use. */ | |
6e8c65f6 | 4429 | return infinite_cost; |
8b11a64c ZD |
4430 | } |
4431 | ||
1d30a09a RG |
4432 | if (address_p |
4433 | || (use->iv->base_object | |
4434 | && cand->iv->base_object | |
4435 | && POINTER_TYPE_P (TREE_TYPE (use->iv->base_object)) | |
4436 | && POINTER_TYPE_P (TREE_TYPE (cand->iv->base_object)))) | |
e6845c23 ZD |
4437 | { |
4438 | /* Do not try to express address of an object with computation based | |
4439 | on address of a different object. This may cause problems in rtl | |
4440 | level alias analysis (that does not expect this to be happening, | |
4441 | as this is illegal in C), and would be unlikely to be useful | |
4442 | anyway. */ | |
4443 | if (use->iv->base_object | |
4444 | && cand->iv->base_object | |
4445 | && !operand_equal_p (use->iv->base_object, cand->iv->base_object, 0)) | |
6e8c65f6 | 4446 | return infinite_cost; |
e6845c23 ZD |
4447 | } |
4448 | ||
7a2faca1 | 4449 | if (TYPE_PRECISION (utype) < TYPE_PRECISION (ctype)) |
8b11a64c ZD |
4450 | { |
4451 | /* TODO -- add direct handling of this case. */ | |
4452 | goto fallback; | |
4453 | } | |
4454 | ||
9be872b7 ZD |
4455 | /* CSTEPI is removed from the offset in case statement is after the |
4456 | increment. If the step is not constant, we use zero instead. | |
1e1def1c | 4457 | This is a bit imprecise (there is the extra addition), but |
9be872b7 ZD |
4458 | redundancy elimination is likely to transform the code so that |
4459 | it uses value of the variable before increment anyway, | |
4460 | so it is not that much unrealistic. */ | |
4461 | if (cst_and_fits_in_hwi (cstep)) | |
4462 | cstepi = int_cst_value (cstep); | |
4463 | else | |
4464 | cstepi = 0; | |
4465 | ||
73f30c63 | 4466 | if (!constant_multiple_of (ustep, cstep, &rat)) |
6e8c65f6 | 4467 | return infinite_cost; |
b8698a0f | 4468 | |
807e902e | 4469 | if (wi::fits_shwi_p (rat)) |
27bcd47c | 4470 | ratio = rat.to_shwi (); |
73f30c63 | 4471 | else |
6e8c65f6 | 4472 | return infinite_cost; |
8b11a64c | 4473 | |
7a2faca1 EB |
4474 | STRIP_NOPS (cbase); |
4475 | ctype = TREE_TYPE (cbase); | |
4476 | ||
2fa692c0 XDL |
4477 | stmt_is_after_inc = stmt_after_increment (data->current_loop, cand, at); |
4478 | ||
8b11a64c ZD |
4479 | /* use = ubase + ratio * (var - cbase). If either cbase is a constant |
4480 | or ratio == 1, it is better to handle this like | |
b8698a0f | 4481 | |
8b11a64c | 4482 | ubase - ratio * cbase + ratio * var |
b8698a0f | 4483 | |
8b11a64c ZD |
4484 | (also holds in the case ratio == -1, TODO. */ |
4485 | ||
9be872b7 | 4486 | if (cst_and_fits_in_hwi (cbase)) |
8b11a64c | 4487 | { |
d7ca26e4 | 4488 | offset = - ratio * (unsigned HOST_WIDE_INT) int_cst_value (cbase); |
6e8c65f6 ZD |
4489 | cost = difference_cost (data, |
4490 | ubase, build_int_cst (utype, 0), | |
4491 | &symbol_present, &var_present, &offset, | |
4492 | depends_on); | |
18081149 | 4493 | cost.cost /= avg_loop_niter (data->current_loop); |
8b11a64c ZD |
4494 | } |
4495 | else if (ratio == 1) | |
4496 | { | |
2fa692c0 XDL |
4497 | tree real_cbase = cbase; |
4498 | ||
4499 | /* Check to see if any adjustment is needed. */ | |
4500 | if (cstepi == 0 && stmt_is_after_inc) | |
4501 | { | |
4502 | aff_tree real_cbase_aff; | |
4503 | aff_tree cstep_aff; | |
4504 | ||
4505 | tree_to_aff_combination (cbase, TREE_TYPE (real_cbase), | |
4506 | &real_cbase_aff); | |
4507 | tree_to_aff_combination (cstep, TREE_TYPE (cstep), &cstep_aff); | |
4508 | ||
4509 | aff_combination_add (&real_cbase_aff, &cstep_aff); | |
4510 | real_cbase = aff_combination_to_tree (&real_cbase_aff); | |
4511 | } | |
4512 | ||
6e8c65f6 | 4513 | cost = difference_cost (data, |
2fa692c0 | 4514 | ubase, real_cbase, |
6e8c65f6 ZD |
4515 | &symbol_present, &var_present, &offset, |
4516 | depends_on); | |
18081149 | 4517 | cost.cost /= avg_loop_niter (data->current_loop); |
8b11a64c | 4518 | } |
7a2faca1 EB |
4519 | else if (address_p |
4520 | && !POINTER_TYPE_P (ctype) | |
09e881c9 | 4521 | && multiplier_allowed_in_address_p |
0c604a61 | 4522 | (ratio, mem_mode, |
09e881c9 | 4523 | TYPE_ADDR_SPACE (TREE_TYPE (utype)))) |
7a2faca1 EB |
4524 | { |
4525 | cbase | |
4526 | = fold_build2 (MULT_EXPR, ctype, cbase, build_int_cst (ctype, ratio)); | |
4527 | cost = difference_cost (data, | |
4528 | ubase, cbase, | |
4529 | &symbol_present, &var_present, &offset, | |
4530 | depends_on); | |
18081149 | 4531 | cost.cost /= avg_loop_niter (data->current_loop); |
7a2faca1 | 4532 | } |
8b11a64c ZD |
4533 | else |
4534 | { | |
6e8c65f6 | 4535 | cost = force_var_cost (data, cbase, depends_on); |
6e8c65f6 ZD |
4536 | cost = add_costs (cost, |
4537 | difference_cost (data, | |
4538 | ubase, build_int_cst (utype, 0), | |
4539 | &symbol_present, &var_present, | |
4540 | &offset, depends_on)); | |
18081149 | 4541 | cost.cost /= avg_loop_niter (data->current_loop); |
5322d07e | 4542 | cost.cost += add_cost (data->speed, TYPE_MODE (ctype)); |
18081149 XDL |
4543 | } |
4544 | ||
a7e43c57 BC |
4545 | /* Set of invariants depended on by sub use has already been computed |
4546 | for the first use in the group. */ | |
4547 | if (use->sub_id) | |
4548 | { | |
4549 | cost.cost = 0; | |
4550 | if (depends_on && *depends_on) | |
4551 | bitmap_clear (*depends_on); | |
4552 | } | |
4553 | else if (inv_expr_id) | |
18081149 XDL |
4554 | { |
4555 | *inv_expr_id = | |
4556 | get_loop_invariant_expr_id (data, ubase, cbase, ratio, address_p); | |
4557 | /* Clear depends on. */ | |
4558 | if (*inv_expr_id != -1 && depends_on && *depends_on) | |
4559 | bitmap_clear (*depends_on); | |
8b11a64c ZD |
4560 | } |
4561 | ||
4562 | /* If we are after the increment, the value of the candidate is higher by | |
4563 | one iteration. */ | |
2c08497a | 4564 | if (stmt_is_after_inc) |
8b11a64c ZD |
4565 | offset -= ratio * cstepi; |
4566 | ||
4567 | /* Now the computation is in shape symbol + var1 + const + ratio * var2. | |
7a2faca1 EB |
4568 | (symbol/var1/const parts may be omitted). If we are looking for an |
4569 | address, find the cost of addressing this. */ | |
8b11a64c | 4570 | if (address_p) |
7a2faca1 EB |
4571 | return add_costs (cost, |
4572 | get_address_cost (symbol_present, var_present, | |
2c08497a | 4573 | offset, ratio, cstepi, |
0c604a61 | 4574 | mem_mode, |
09e881c9 | 4575 | TYPE_ADDR_SPACE (TREE_TYPE (utype)), |
2c08497a BS |
4576 | speed, stmt_is_after_inc, |
4577 | can_autoinc)); | |
8b11a64c ZD |
4578 | |
4579 | /* Otherwise estimate the costs for computing the expression. */ | |
8b11a64c ZD |
4580 | if (!symbol_present && !var_present && !offset) |
4581 | { | |
4582 | if (ratio != 1) | |
6dd8f4bb | 4583 | cost.cost += mult_by_coeff_cost (ratio, TYPE_MODE (ctype), speed); |
8b11a64c ZD |
4584 | return cost; |
4585 | } | |
4586 | ||
7a2faca1 EB |
4587 | /* Symbol + offset should be compile-time computable so consider that they |
4588 | are added once to the variable, if present. */ | |
4589 | if (var_present && (symbol_present || offset)) | |
6521ac85 | 4590 | cost.cost += adjust_setup_cost (data, |
5322d07e | 4591 | add_cost (speed, TYPE_MODE (ctype))); |
8b11a64c | 4592 | |
6e8c65f6 ZD |
4593 | /* Having offset does not affect runtime cost in case it is added to |
4594 | symbol, but it increases complexity. */ | |
4595 | if (offset) | |
4596 | cost.complexity++; | |
4597 | ||
5322d07e | 4598 | cost.cost += add_cost (speed, TYPE_MODE (ctype)); |
7a2faca1 EB |
4599 | |
4600 | aratio = ratio > 0 ? ratio : -ratio; | |
4601 | if (aratio != 1) | |
6dd8f4bb | 4602 | cost.cost += mult_by_coeff_cost (aratio, TYPE_MODE (ctype), speed); |
5914a70f | 4603 | return cost; |
8b11a64c ZD |
4604 | |
4605 | fallback: | |
2c08497a BS |
4606 | if (can_autoinc) |
4607 | *can_autoinc = false; | |
4608 | ||
8b11a64c ZD |
4609 | { |
4610 | /* Just get the expression, expand it and measure the cost. */ | |
4611 | tree comp = get_computation_at (data->current_loop, use, cand, at); | |
4612 | ||
4613 | if (!comp) | |
6e8c65f6 | 4614 | return infinite_cost; |
8b11a64c ZD |
4615 | |
4616 | if (address_p) | |
70f34814 | 4617 | comp = build_simple_mem_ref (comp); |
8b11a64c | 4618 | |
f40751dd | 4619 | return new_cost (computation_cost (comp, speed), 0); |
8b11a64c ZD |
4620 | } |
4621 | } | |
4622 | ||
4623 | /* Determines the cost of the computation by that USE is expressed | |
4624 | from induction variable CAND. If ADDRESS_P is true, we just need | |
4625 | to create an address from it, otherwise we want to get it into | |
4626 | register. A set of invariants we depend on is stored in | |
2c08497a BS |
4627 | DEPENDS_ON. If CAN_AUTOINC is nonnull, use it to record whether |
4628 | autoinc addressing is likely. */ | |
8b11a64c | 4629 | |
6e8c65f6 | 4630 | static comp_cost |
8b11a64c ZD |
4631 | get_computation_cost (struct ivopts_data *data, |
4632 | struct iv_use *use, struct iv_cand *cand, | |
18081149 XDL |
4633 | bool address_p, bitmap *depends_on, |
4634 | bool *can_autoinc, int *inv_expr_id) | |
8b11a64c ZD |
4635 | { |
4636 | return get_computation_cost_at (data, | |
2c08497a | 4637 | use, cand, address_p, depends_on, use->stmt, |
18081149 | 4638 | can_autoinc, inv_expr_id); |
8b11a64c ZD |
4639 | } |
4640 | ||
4641 | /* Determines cost of basing replacement of USE on CAND in a generic | |
4642 | expression. */ | |
4643 | ||
b1b02be2 | 4644 | static bool |
8b11a64c ZD |
4645 | determine_use_iv_cost_generic (struct ivopts_data *data, |
4646 | struct iv_use *use, struct iv_cand *cand) | |
4647 | { | |
4648 | bitmap depends_on; | |
6e8c65f6 | 4649 | comp_cost cost; |
18081149 | 4650 | int inv_expr_id = -1; |
eec5fec9 ZD |
4651 | |
4652 | /* The simple case first -- if we need to express value of the preserved | |
4653 | original biv, the cost is 0. This also prevents us from counting the | |
4654 | cost of increment twice -- once at this use and once in the cost of | |
4655 | the candidate. */ | |
4656 | if (cand->pos == IP_ORIGINAL | |
4657 | && cand->incremented_at == use->stmt) | |
4658 | { | |
7735d6c7 | 4659 | set_use_iv_cost (data, use, cand, no_cost, NULL, NULL_TREE, |
d8af4ba3 | 4660 | ERROR_MARK, -1); |
eec5fec9 ZD |
4661 | return true; |
4662 | } | |
8b11a64c | 4663 | |
18081149 XDL |
4664 | cost = get_computation_cost (data, use, cand, false, &depends_on, |
4665 | NULL, &inv_expr_id); | |
4666 | ||
d8af4ba3 | 4667 | set_use_iv_cost (data, use, cand, cost, depends_on, NULL_TREE, ERROR_MARK, |
18081149 | 4668 | inv_expr_id); |
b1b02be2 | 4669 | |
6e8c65f6 | 4670 | return !infinite_cost_p (cost); |
8b11a64c ZD |
4671 | } |
4672 | ||
4673 | /* Determines cost of basing replacement of USE on CAND in an address. */ | |
4674 | ||
b1b02be2 | 4675 | static bool |
8b11a64c ZD |
4676 | determine_use_iv_cost_address (struct ivopts_data *data, |
4677 | struct iv_use *use, struct iv_cand *cand) | |
4678 | { | |
4679 | bitmap depends_on; | |
2c08497a | 4680 | bool can_autoinc; |
18081149 | 4681 | int inv_expr_id = -1; |
a7e43c57 BC |
4682 | struct iv_use *sub_use; |
4683 | comp_cost sub_cost; | |
2c08497a | 4684 | comp_cost cost = get_computation_cost (data, use, cand, true, &depends_on, |
18081149 | 4685 | &can_autoinc, &inv_expr_id); |
8b11a64c | 4686 | |
2c08497a BS |
4687 | if (cand->ainc_use == use) |
4688 | { | |
4689 | if (can_autoinc) | |
4690 | cost.cost -= cand->cost_step; | |
4691 | /* If we generated the candidate solely for exploiting autoincrement | |
4692 | opportunities, and it turns out it can't be used, set the cost to | |
4693 | infinity to make sure we ignore it. */ | |
4694 | else if (cand->pos == IP_AFTER_USE || cand->pos == IP_BEFORE_USE) | |
4695 | cost = infinite_cost; | |
4696 | } | |
a7e43c57 BC |
4697 | for (sub_use = use->next; |
4698 | sub_use && !infinite_cost_p (cost); | |
4699 | sub_use = sub_use->next) | |
4700 | { | |
4701 | sub_cost = get_computation_cost (data, sub_use, cand, true, &depends_on, | |
4702 | &can_autoinc, &inv_expr_id); | |
4703 | cost = add_costs (cost, sub_cost); | |
4704 | } | |
4705 | ||
d8af4ba3 | 4706 | set_use_iv_cost (data, use, cand, cost, depends_on, NULL_TREE, ERROR_MARK, |
18081149 | 4707 | inv_expr_id); |
b1b02be2 | 4708 | |
6e8c65f6 | 4709 | return !infinite_cost_p (cost); |
8b11a64c ZD |
4710 | } |
4711 | ||
7e2ac86c ZD |
4712 | /* Computes value of candidate CAND at position AT in iteration NITER, and |
4713 | stores it to VAL. */ | |
8b11a64c | 4714 | |
7e2ac86c | 4715 | static void |
726a989a | 4716 | cand_value_at (struct loop *loop, struct iv_cand *cand, gimple at, tree niter, |
7e2ac86c | 4717 | aff_tree *val) |
8b11a64c | 4718 | { |
7e2ac86c ZD |
4719 | aff_tree step, delta, nit; |
4720 | struct iv *iv = cand->iv; | |
8b11a64c | 4721 | tree type = TREE_TYPE (iv->base); |
1ffe34d9 AP |
4722 | tree steptype = type; |
4723 | if (POINTER_TYPE_P (type)) | |
4724 | steptype = sizetype; | |
d6adff07 | 4725 | steptype = unsigned_type_for (type); |
8b11a64c | 4726 | |
d6adff07 RB |
4727 | tree_to_aff_combination (iv->step, TREE_TYPE (iv->step), &step); |
4728 | aff_combination_convert (&step, steptype); | |
7e2ac86c | 4729 | tree_to_aff_combination (niter, TREE_TYPE (niter), &nit); |
1ffe34d9 | 4730 | aff_combination_convert (&nit, steptype); |
7e2ac86c | 4731 | aff_combination_mult (&nit, &step, &delta); |
8b11a64c | 4732 | if (stmt_after_increment (loop, cand, at)) |
7e2ac86c | 4733 | aff_combination_add (&delta, &step); |
8b11a64c | 4734 | |
7e2ac86c | 4735 | tree_to_aff_combination (iv->base, type, val); |
d6adff07 RB |
4736 | if (!POINTER_TYPE_P (type)) |
4737 | aff_combination_convert (val, steptype); | |
7e2ac86c | 4738 | aff_combination_add (val, &delta); |
8b11a64c ZD |
4739 | } |
4740 | ||
ca4c3169 ZD |
4741 | /* Returns period of induction variable iv. */ |
4742 | ||
4743 | static tree | |
4744 | iv_period (struct iv *iv) | |
4745 | { | |
4746 | tree step = iv->step, period, type; | |
4747 | tree pow2div; | |
4748 | ||
4749 | gcc_assert (step && TREE_CODE (step) == INTEGER_CST); | |
4750 | ||
ca4c3169 | 4751 | type = unsigned_type_for (TREE_TYPE (step)); |
e2102efc XDL |
4752 | /* Period of the iv is lcm (step, type_range)/step -1, |
4753 | i.e., N*type_range/step - 1. Since type range is power | |
4754 | of two, N == (step >> num_of_ending_zeros_binary (step), | |
4755 | so the final result is | |
4756 | ||
4757 | (type_range >> num_of_ending_zeros_binary (step)) - 1 | |
4758 | ||
4759 | */ | |
4760 | pow2div = num_ending_zeros (step); | |
ca4c3169 ZD |
4761 | |
4762 | period = build_low_bits_mask (type, | |
e2102efc | 4763 | (TYPE_PRECISION (type) |
ae7e9ddd | 4764 | - tree_to_uhwi (pow2div))); |
ca4c3169 ZD |
4765 | |
4766 | return period; | |
4767 | } | |
4768 | ||
f5f12961 ZD |
4769 | /* Returns the comparison operator used when eliminating the iv USE. */ |
4770 | ||
4771 | static enum tree_code | |
4772 | iv_elimination_compare (struct ivopts_data *data, struct iv_use *use) | |
4773 | { | |
4774 | struct loop *loop = data->current_loop; | |
4775 | basic_block ex_bb; | |
4776 | edge exit; | |
4777 | ||
726a989a | 4778 | ex_bb = gimple_bb (use->stmt); |
f5f12961 ZD |
4779 | exit = EDGE_SUCC (ex_bb, 0); |
4780 | if (flow_bb_inside_loop_p (loop, exit->dest)) | |
4781 | exit = EDGE_SUCC (ex_bb, 1); | |
4782 | ||
4783 | return (exit->flags & EDGE_TRUE_VALUE ? EQ_EXPR : NE_EXPR); | |
4784 | } | |
4785 | ||
d8af4ba3 ZD |
4786 | /* Returns true if we can prove that BASE - OFFSET does not overflow. For now, |
4787 | we only detect the situation that BASE = SOMETHING + OFFSET, where the | |
4788 | calculation is performed in non-wrapping type. | |
4789 | ||
4790 | TODO: More generally, we could test for the situation that | |
4791 | BASE = SOMETHING + OFFSET' and OFFSET is between OFFSET' and zero. | |
3230c614 | 4792 | This would require knowing the sign of OFFSET. */ |
d8af4ba3 ZD |
4793 | |
4794 | static bool | |
3230c614 | 4795 | difference_cannot_overflow_p (struct ivopts_data *data, tree base, tree offset) |
d8af4ba3 ZD |
4796 | { |
4797 | enum tree_code code; | |
4798 | tree e1, e2; | |
3230c614 | 4799 | aff_tree aff_e1, aff_e2, aff_offset; |
d8af4ba3 ZD |
4800 | |
4801 | if (!nowrap_type_p (TREE_TYPE (base))) | |
4802 | return false; | |
4803 | ||
4804 | base = expand_simple_operations (base); | |
4805 | ||
4806 | if (TREE_CODE (base) == SSA_NAME) | |
4807 | { | |
4808 | gimple stmt = SSA_NAME_DEF_STMT (base); | |
4809 | ||
4810 | if (gimple_code (stmt) != GIMPLE_ASSIGN) | |
4811 | return false; | |
4812 | ||
4813 | code = gimple_assign_rhs_code (stmt); | |
4814 | if (get_gimple_rhs_class (code) != GIMPLE_BINARY_RHS) | |
4815 | return false; | |
4816 | ||
4817 | e1 = gimple_assign_rhs1 (stmt); | |
4818 | e2 = gimple_assign_rhs2 (stmt); | |
4819 | } | |
4820 | else | |
4821 | { | |
4822 | code = TREE_CODE (base); | |
4823 | if (get_gimple_rhs_class (code) != GIMPLE_BINARY_RHS) | |
4824 | return false; | |
4825 | e1 = TREE_OPERAND (base, 0); | |
4826 | e2 = TREE_OPERAND (base, 1); | |
4827 | } | |
4828 | ||
3230c614 BC |
4829 | /* Use affine expansion as deeper inspection to prove the equality. */ |
4830 | tree_to_aff_combination_expand (e2, TREE_TYPE (e2), | |
4831 | &aff_e2, &data->name_expansion_cache); | |
4832 | tree_to_aff_combination_expand (offset, TREE_TYPE (offset), | |
4833 | &aff_offset, &data->name_expansion_cache); | |
4834 | aff_combination_scale (&aff_offset, -1); | |
d8af4ba3 ZD |
4835 | switch (code) |
4836 | { | |
4837 | case PLUS_EXPR: | |
3230c614 BC |
4838 | aff_combination_add (&aff_e2, &aff_offset); |
4839 | if (aff_combination_zero_p (&aff_e2)) | |
4840 | return true; | |
4841 | ||
4842 | tree_to_aff_combination_expand (e1, TREE_TYPE (e1), | |
4843 | &aff_e1, &data->name_expansion_cache); | |
4844 | aff_combination_add (&aff_e1, &aff_offset); | |
4845 | return aff_combination_zero_p (&aff_e1); | |
4846 | ||
d8af4ba3 | 4847 | case POINTER_PLUS_EXPR: |
3230c614 BC |
4848 | aff_combination_add (&aff_e2, &aff_offset); |
4849 | return aff_combination_zero_p (&aff_e2); | |
d8af4ba3 ZD |
4850 | |
4851 | default: | |
4852 | return false; | |
4853 | } | |
4854 | } | |
4855 | ||
4856 | /* Tries to replace loop exit by one formulated in terms of a LT_EXPR | |
4857 | comparison with CAND. NITER describes the number of iterations of | |
4858 | the loops. If successful, the comparison in COMP_P is altered accordingly. | |
4859 | ||
4860 | We aim to handle the following situation: | |
4861 | ||
4862 | sometype *base, *p; | |
4863 | int a, b, i; | |
4864 | ||
4865 | i = a; | |
4866 | p = p_0 = base + a; | |
4867 | ||
4868 | do | |
4869 | { | |
4870 | bla (*p); | |
4871 | p++; | |
4872 | i++; | |
4873 | } | |
4874 | while (i < b); | |
4875 | ||
4876 | Here, the number of iterations of the loop is (a + 1 > b) ? 0 : b - a - 1. | |
4877 | We aim to optimize this to | |
4878 | ||
4879 | p = p_0 = base + a; | |
4880 | do | |
4881 | { | |
4882 | bla (*p); | |
4883 | p++; | |
4884 | } | |
4885 | while (p < p_0 - a + b); | |
4886 | ||
4887 | This preserves the correctness, since the pointer arithmetics does not | |
4888 | overflow. More precisely: | |
4889 | ||
4890 | 1) if a + 1 <= b, then p_0 - a + b is the final value of p, hence there is no | |
4891 | overflow in computing it or the values of p. | |
4892 | 2) if a + 1 > b, then we need to verify that the expression p_0 - a does not | |
4893 | overflow. To prove this, we use the fact that p_0 = base + a. */ | |
4894 | ||
4895 | static bool | |
4896 | iv_elimination_compare_lt (struct ivopts_data *data, | |
4897 | struct iv_cand *cand, enum tree_code *comp_p, | |
4898 | struct tree_niter_desc *niter) | |
4899 | { | |
4900 | tree cand_type, a, b, mbz, nit_type = TREE_TYPE (niter->niter), offset; | |
84562394 | 4901 | struct aff_tree nit, tmpa, tmpb; |
d8af4ba3 ZD |
4902 | enum tree_code comp; |
4903 | HOST_WIDE_INT step; | |
4904 | ||
4905 | /* We need to know that the candidate induction variable does not overflow. | |
4906 | While more complex analysis may be used to prove this, for now just | |
4907 | check that the variable appears in the original program and that it | |
4908 | is computed in a type that guarantees no overflows. */ | |
4909 | cand_type = TREE_TYPE (cand->iv->base); | |
4910 | if (cand->pos != IP_ORIGINAL || !nowrap_type_p (cand_type)) | |
4911 | return false; | |
4912 | ||
4913 | /* Make sure that the loop iterates till the loop bound is hit, as otherwise | |
4914 | the calculation of the BOUND could overflow, making the comparison | |
4915 | invalid. */ | |
4916 | if (!data->loop_single_exit_p) | |
4917 | return false; | |
4918 | ||
4919 | /* We need to be able to decide whether candidate is increasing or decreasing | |
4920 | in order to choose the right comparison operator. */ | |
4921 | if (!cst_and_fits_in_hwi (cand->iv->step)) | |
4922 | return false; | |
4923 | step = int_cst_value (cand->iv->step); | |
4924 | ||
4925 | /* Check that the number of iterations matches the expected pattern: | |
4926 | a + 1 > b ? 0 : b - a - 1. */ | |
4927 | mbz = niter->may_be_zero; | |
4928 | if (TREE_CODE (mbz) == GT_EXPR) | |
4929 | { | |
4930 | /* Handle a + 1 > b. */ | |
4931 | tree op0 = TREE_OPERAND (mbz, 0); | |
4932 | if (TREE_CODE (op0) == PLUS_EXPR && integer_onep (TREE_OPERAND (op0, 1))) | |
4933 | { | |
4934 | a = TREE_OPERAND (op0, 0); | |
4935 | b = TREE_OPERAND (mbz, 1); | |
4936 | } | |
4937 | else | |
4938 | return false; | |
4939 | } | |
4940 | else if (TREE_CODE (mbz) == LT_EXPR) | |
4941 | { | |
4942 | tree op1 = TREE_OPERAND (mbz, 1); | |
4943 | ||
4944 | /* Handle b < a + 1. */ | |
4945 | if (TREE_CODE (op1) == PLUS_EXPR && integer_onep (TREE_OPERAND (op1, 1))) | |
4946 | { | |
4947 | a = TREE_OPERAND (op1, 0); | |
4948 | b = TREE_OPERAND (mbz, 0); | |
4949 | } | |
4950 | else | |
4951 | return false; | |
4952 | } | |
4953 | else | |
4954 | return false; | |
4955 | ||
4956 | /* Expected number of iterations is B - A - 1. Check that it matches | |
4957 | the actual number, i.e., that B - A - NITER = 1. */ | |
4958 | tree_to_aff_combination (niter->niter, nit_type, &nit); | |
4959 | tree_to_aff_combination (fold_convert (nit_type, a), nit_type, &tmpa); | |
4960 | tree_to_aff_combination (fold_convert (nit_type, b), nit_type, &tmpb); | |
807e902e KZ |
4961 | aff_combination_scale (&nit, -1); |
4962 | aff_combination_scale (&tmpa, -1); | |
d8af4ba3 ZD |
4963 | aff_combination_add (&tmpb, &tmpa); |
4964 | aff_combination_add (&tmpb, &nit); | |
807e902e | 4965 | if (tmpb.n != 0 || tmpb.offset != 1) |
d8af4ba3 ZD |
4966 | return false; |
4967 | ||
4968 | /* Finally, check that CAND->IV->BASE - CAND->IV->STEP * A does not | |
4969 | overflow. */ | |
4970 | offset = fold_build2 (MULT_EXPR, TREE_TYPE (cand->iv->step), | |
4971 | cand->iv->step, | |
4972 | fold_convert (TREE_TYPE (cand->iv->step), a)); | |
3230c614 | 4973 | if (!difference_cannot_overflow_p (data, cand->iv->base, offset)) |
d8af4ba3 ZD |
4974 | return false; |
4975 | ||
4976 | /* Determine the new comparison operator. */ | |
4977 | comp = step < 0 ? GT_EXPR : LT_EXPR; | |
4978 | if (*comp_p == NE_EXPR) | |
4979 | *comp_p = comp; | |
4980 | else if (*comp_p == EQ_EXPR) | |
4981 | *comp_p = invert_tree_comparison (comp, false); | |
4982 | else | |
4983 | gcc_unreachable (); | |
4984 | ||
4985 | return true; | |
4986 | } | |
4987 | ||
8b11a64c | 4988 | /* Check whether it is possible to express the condition in USE by comparison |
d8af4ba3 ZD |
4989 | of candidate CAND. If so, store the value compared with to BOUND, and the |
4990 | comparison operator to COMP. */ | |
8b11a64c ZD |
4991 | |
4992 | static bool | |
ca4c3169 | 4993 | may_eliminate_iv (struct ivopts_data *data, |
d8af4ba3 ZD |
4994 | struct iv_use *use, struct iv_cand *cand, tree *bound, |
4995 | enum tree_code *comp) | |
8b11a64c | 4996 | { |
e6845c23 | 4997 | basic_block ex_bb; |
8b11a64c | 4998 | edge exit; |
d8af4ba3 | 4999 | tree period; |
ca4c3169 | 5000 | struct loop *loop = data->current_loop; |
7e2ac86c | 5001 | aff_tree bnd; |
e2102efc | 5002 | struct tree_niter_desc *desc = NULL; |
cbc012d5 | 5003 | |
9be872b7 ZD |
5004 | if (TREE_CODE (cand->iv->step) != INTEGER_CST) |
5005 | return false; | |
5006 | ||
52778e2a EB |
5007 | /* For now works only for exits that dominate the loop latch. |
5008 | TODO: extend to other conditions inside loop body. */ | |
726a989a | 5009 | ex_bb = gimple_bb (use->stmt); |
e6845c23 | 5010 | if (use->stmt != last_stmt (ex_bb) |
726a989a RB |
5011 | || gimple_code (use->stmt) != GIMPLE_COND |
5012 | || !dominated_by_p (CDI_DOMINATORS, loop->latch, ex_bb)) | |
8b11a64c ZD |
5013 | return false; |
5014 | ||
e6845c23 ZD |
5015 | exit = EDGE_SUCC (ex_bb, 0); |
5016 | if (flow_bb_inside_loop_p (loop, exit->dest)) | |
5017 | exit = EDGE_SUCC (ex_bb, 1); | |
5018 | if (flow_bb_inside_loop_p (loop, exit->dest)) | |
5019 | return false; | |
5020 | ||
d8af4ba3 ZD |
5021 | desc = niter_for_exit (data, exit); |
5022 | if (!desc) | |
8b11a64c ZD |
5023 | return false; |
5024 | ||
52778e2a EB |
5025 | /* Determine whether we can use the variable to test the exit condition. |
5026 | This is the case iff the period of the induction variable is greater | |
5027 | than the number of iterations for which the exit condition is true. */ | |
ca4c3169 | 5028 | period = iv_period (cand->iv); |
8b11a64c | 5029 | |
52778e2a | 5030 | /* If the number of iterations is constant, compare against it directly. */ |
d8af4ba3 | 5031 | if (TREE_CODE (desc->niter) == INTEGER_CST) |
52778e2a | 5032 | { |
e2102efc XDL |
5033 | /* See cand_value_at. */ |
5034 | if (stmt_after_increment (loop, cand, use->stmt)) | |
5035 | { | |
d8af4ba3 | 5036 | if (!tree_int_cst_lt (desc->niter, period)) |
e2102efc XDL |
5037 | return false; |
5038 | } | |
5039 | else | |
5040 | { | |
d8af4ba3 | 5041 | if (tree_int_cst_lt (period, desc->niter)) |
e2102efc XDL |
5042 | return false; |
5043 | } | |
52778e2a EB |
5044 | } |
5045 | ||
5046 | /* If not, and if this is the only possible exit of the loop, see whether | |
5047 | we can get a conservative estimate on the number of iterations of the | |
5048 | entire loop and compare against that instead. */ | |
e2102efc | 5049 | else |
52778e2a | 5050 | { |
807e902e | 5051 | widest_int period_value, max_niter; |
e2102efc XDL |
5052 | |
5053 | max_niter = desc->max; | |
5054 | if (stmt_after_increment (loop, cand, use->stmt)) | |
807e902e KZ |
5055 | max_niter += 1; |
5056 | period_value = wi::to_widest (period); | |
5057 | if (wi::gtu_p (max_niter, period_value)) | |
e2102efc | 5058 | { |
073a8998 | 5059 | /* See if we can take advantage of inferred loop bound information. */ |
d8af4ba3 | 5060 | if (data->loop_single_exit_p) |
e2102efc | 5061 | { |
652c4c71 | 5062 | if (!max_loop_iterations (loop, &max_niter)) |
e2102efc XDL |
5063 | return false; |
5064 | /* The loop bound is already adjusted by adding 1. */ | |
807e902e | 5065 | if (wi::gtu_p (max_niter, period_value)) |
e2102efc XDL |
5066 | return false; |
5067 | } | |
5068 | else | |
5069 | return false; | |
5070 | } | |
52778e2a EB |
5071 | } |
5072 | ||
d8af4ba3 | 5073 | cand_value_at (loop, cand, use->stmt, desc->niter, &bnd); |
771f882e | 5074 | |
d6adff07 RB |
5075 | *bound = fold_convert (TREE_TYPE (cand->iv->base), |
5076 | aff_combination_to_tree (&bnd)); | |
d8af4ba3 ZD |
5077 | *comp = iv_elimination_compare (data, use); |
5078 | ||
771f882e ZD |
5079 | /* It is unlikely that computing the number of iterations using division |
5080 | would be more profitable than keeping the original induction variable. */ | |
5081 | if (expression_expensive_p (*bound)) | |
5082 | return false; | |
d8af4ba3 ZD |
5083 | |
5084 | /* Sometimes, it is possible to handle the situation that the number of | |
5085 | iterations may be zero unless additional assumtions by using < | |
5086 | instead of != in the exit condition. | |
5087 | ||
5088 | TODO: we could also calculate the value MAY_BE_ZERO ? 0 : NITER and | |
5089 | base the exit condition on it. However, that is often too | |
5090 | expensive. */ | |
5091 | if (!integer_zerop (desc->may_be_zero)) | |
5092 | return iv_elimination_compare_lt (data, cand, comp, desc); | |
5093 | ||
8b11a64c ZD |
5094 | return true; |
5095 | } | |
5096 | ||
bb8d292d TV |
5097 | /* Calculates the cost of BOUND, if it is a PARM_DECL. A PARM_DECL must |
5098 | be copied, if is is used in the loop body and DATA->body_includes_call. */ | |
5099 | ||
5100 | static int | |
5101 | parm_decl_cost (struct ivopts_data *data, tree bound) | |
5102 | { | |
5103 | tree sbound = bound; | |
5104 | STRIP_NOPS (sbound); | |
5105 | ||
5106 | if (TREE_CODE (sbound) == SSA_NAME | |
67386041 | 5107 | && SSA_NAME_IS_DEFAULT_DEF (sbound) |
bb8d292d | 5108 | && TREE_CODE (SSA_NAME_VAR (sbound)) == PARM_DECL |
bb8d292d TV |
5109 | && data->body_includes_call) |
5110 | return COSTS_N_INSNS (1); | |
5111 | ||
5112 | return 0; | |
5113 | } | |
e2102efc | 5114 | |
8b11a64c ZD |
5115 | /* Determines cost of basing replacement of USE on CAND in a condition. */ |
5116 | ||
b1b02be2 | 5117 | static bool |
8b11a64c ZD |
5118 | determine_use_iv_cost_condition (struct ivopts_data *data, |
5119 | struct iv_use *use, struct iv_cand *cand) | |
5120 | { | |
b697aed4 ZD |
5121 | tree bound = NULL_TREE; |
5122 | struct iv *cmp_iv; | |
5123 | bitmap depends_on_elim = NULL, depends_on_express = NULL, depends_on; | |
bb8d292d | 5124 | comp_cost elim_cost, express_cost, cost, bound_cost; |
b697aed4 | 5125 | bool ok; |
bb8d292d | 5126 | int elim_inv_expr_id = -1, express_inv_expr_id = -1, inv_expr_id; |
04eadb13 | 5127 | tree *control_var, *bound_cst; |
76725a03 | 5128 | enum tree_code comp = ERROR_MARK; |
8b11a64c ZD |
5129 | |
5130 | /* Only consider real candidates. */ | |
5131 | if (!cand->iv) | |
5132 | { | |
d8af4ba3 ZD |
5133 | set_use_iv_cost (data, use, cand, infinite_cost, NULL, NULL_TREE, |
5134 | ERROR_MARK, -1); | |
b1b02be2 | 5135 | return false; |
8b11a64c ZD |
5136 | } |
5137 | ||
b697aed4 | 5138 | /* Try iv elimination. */ |
d8af4ba3 | 5139 | if (may_eliminate_iv (data, use, cand, &bound, &comp)) |
cbc012d5 ZD |
5140 | { |
5141 | elim_cost = force_var_cost (data, bound, &depends_on_elim); | |
bb8d292d TV |
5142 | if (elim_cost.cost == 0) |
5143 | elim_cost.cost = parm_decl_cost (data, bound); | |
5144 | else if (TREE_CODE (bound) == INTEGER_CST) | |
5145 | elim_cost.cost = 0; | |
5146 | /* If we replace a loop condition 'i < n' with 'p < base + n', | |
5147 | depends_on_elim will have 'base' and 'n' set, which implies | |
5148 | that both 'base' and 'n' will be live during the loop. More likely, | |
5149 | 'base + n' will be loop invariant, resulting in only one live value | |
5150 | during the loop. So in that case we clear depends_on_elim and set | |
5151 | elim_inv_expr_id instead. */ | |
5152 | if (depends_on_elim && bitmap_count_bits (depends_on_elim) > 1) | |
5153 | { | |
5154 | elim_inv_expr_id = get_expr_id (data, bound); | |
5155 | bitmap_clear (depends_on_elim); | |
5156 | } | |
cbc012d5 ZD |
5157 | /* The bound is a loop invariant, so it will be only computed |
5158 | once. */ | |
6521ac85 | 5159 | elim_cost.cost = adjust_setup_cost (data, elim_cost.cost); |
cbc012d5 | 5160 | } |
b697aed4 | 5161 | else |
6e8c65f6 | 5162 | elim_cost = infinite_cost; |
8b11a64c | 5163 | |
b697aed4 ZD |
5164 | /* Try expressing the original giv. If it is compared with an invariant, |
5165 | note that we cannot get rid of it. */ | |
04eadb13 SP |
5166 | ok = extract_cond_operands (data, use->stmt, &control_var, &bound_cst, |
5167 | NULL, &cmp_iv); | |
b697aed4 | 5168 | gcc_assert (ok); |
8b11a64c | 5169 | |
04eadb13 SP |
5170 | /* When the condition is a comparison of the candidate IV against |
5171 | zero, prefer this IV. | |
5172 | ||
073a8998 | 5173 | TODO: The constant that we're subtracting from the cost should |
04eadb13 SP |
5174 | be target-dependent. This information should be added to the |
5175 | target costs for each backend. */ | |
572ae476 CF |
5176 | if (!infinite_cost_p (elim_cost) /* Do not try to decrease infinite! */ |
5177 | && integer_zerop (*bound_cst) | |
04eadb13 SP |
5178 | && (operand_equal_p (*control_var, cand->var_after, 0) |
5179 | || operand_equal_p (*control_var, cand->var_before, 0))) | |
5180 | elim_cost.cost -= 1; | |
5181 | ||
b697aed4 | 5182 | express_cost = get_computation_cost (data, use, cand, false, |
18081149 | 5183 | &depends_on_express, NULL, |
bb8d292d | 5184 | &express_inv_expr_id); |
b697aed4 ZD |
5185 | fd_ivopts_data = data; |
5186 | walk_tree (&cmp_iv->base, find_depends, &depends_on_express, NULL); | |
f5f12961 | 5187 | |
bb8d292d TV |
5188 | /* Count the cost of the original bound as well. */ |
5189 | bound_cost = force_var_cost (data, *bound_cst, NULL); | |
5190 | if (bound_cost.cost == 0) | |
5191 | bound_cost.cost = parm_decl_cost (data, *bound_cst); | |
5192 | else if (TREE_CODE (*bound_cst) == INTEGER_CST) | |
5193 | bound_cost.cost = 0; | |
5194 | express_cost.cost += bound_cost.cost; | |
5195 | ||
b9ff6079 AP |
5196 | /* Choose the better approach, preferring the eliminated IV. */ |
5197 | if (compare_costs (elim_cost, express_cost) <= 0) | |
8b11a64c | 5198 | { |
b697aed4 ZD |
5199 | cost = elim_cost; |
5200 | depends_on = depends_on_elim; | |
5201 | depends_on_elim = NULL; | |
bb8d292d | 5202 | inv_expr_id = elim_inv_expr_id; |
b697aed4 ZD |
5203 | } |
5204 | else | |
5205 | { | |
5206 | cost = express_cost; | |
5207 | depends_on = depends_on_express; | |
5208 | depends_on_express = NULL; | |
5209 | bound = NULL_TREE; | |
d8af4ba3 | 5210 | comp = ERROR_MARK; |
bb8d292d | 5211 | inv_expr_id = express_inv_expr_id; |
8b11a64c ZD |
5212 | } |
5213 | ||
d8af4ba3 | 5214 | set_use_iv_cost (data, use, cand, cost, depends_on, bound, comp, inv_expr_id); |
b697aed4 ZD |
5215 | |
5216 | if (depends_on_elim) | |
5217 | BITMAP_FREE (depends_on_elim); | |
5218 | if (depends_on_express) | |
5219 | BITMAP_FREE (depends_on_express); | |
5220 | ||
6e8c65f6 | 5221 | return !infinite_cost_p (cost); |
8b11a64c ZD |
5222 | } |
5223 | ||
b1b02be2 ZD |
5224 | /* Determines cost of basing replacement of USE on CAND. Returns false |
5225 | if USE cannot be based on CAND. */ | |
8b11a64c | 5226 | |
b1b02be2 | 5227 | static bool |
8b11a64c ZD |
5228 | determine_use_iv_cost (struct ivopts_data *data, |
5229 | struct iv_use *use, struct iv_cand *cand) | |
5230 | { | |
5231 | switch (use->type) | |
5232 | { | |
5233 | case USE_NONLINEAR_EXPR: | |
b1b02be2 | 5234 | return determine_use_iv_cost_generic (data, use, cand); |
8b11a64c | 5235 | |
8b11a64c | 5236 | case USE_ADDRESS: |
b1b02be2 | 5237 | return determine_use_iv_cost_address (data, use, cand); |
8b11a64c ZD |
5238 | |
5239 | case USE_COMPARE: | |
b1b02be2 | 5240 | return determine_use_iv_cost_condition (data, use, cand); |
8b11a64c ZD |
5241 | |
5242 | default: | |
1e128c5f | 5243 | gcc_unreachable (); |
8b11a64c ZD |
5244 | } |
5245 | } | |
5246 | ||
2c08497a BS |
5247 | /* Return true if get_computation_cost indicates that autoincrement is |
5248 | a possibility for the pair of USE and CAND, false otherwise. */ | |
5249 | ||
5250 | static bool | |
5251 | autoinc_possible_for_pair (struct ivopts_data *data, struct iv_use *use, | |
5252 | struct iv_cand *cand) | |
5253 | { | |
5254 | bitmap depends_on; | |
5255 | bool can_autoinc; | |
5256 | comp_cost cost; | |
5257 | ||
5258 | if (use->type != USE_ADDRESS) | |
5259 | return false; | |
5260 | ||
5261 | cost = get_computation_cost (data, use, cand, true, &depends_on, | |
18081149 | 5262 | &can_autoinc, NULL); |
2c08497a BS |
5263 | |
5264 | BITMAP_FREE (depends_on); | |
5265 | ||
5266 | return !infinite_cost_p (cost) && can_autoinc; | |
5267 | } | |
5268 | ||
5269 | /* Examine IP_ORIGINAL candidates to see if they are incremented next to a | |
5270 | use that allows autoincrement, and set their AINC_USE if possible. */ | |
5271 | ||
5272 | static void | |
5273 | set_autoinc_for_original_candidates (struct ivopts_data *data) | |
5274 | { | |
5275 | unsigned i, j; | |
5276 | ||
5277 | for (i = 0; i < n_iv_cands (data); i++) | |
5278 | { | |
5279 | struct iv_cand *cand = iv_cand (data, i); | |
85ff4ec6 BC |
5280 | struct iv_use *closest_before = NULL; |
5281 | struct iv_use *closest_after = NULL; | |
2c08497a BS |
5282 | if (cand->pos != IP_ORIGINAL) |
5283 | continue; | |
85ff4ec6 | 5284 | |
2c08497a BS |
5285 | for (j = 0; j < n_iv_uses (data); j++) |
5286 | { | |
5287 | struct iv_use *use = iv_use (data, j); | |
5288 | unsigned uid = gimple_uid (use->stmt); | |
85ff4ec6 BC |
5289 | |
5290 | if (gimple_bb (use->stmt) != gimple_bb (cand->incremented_at)) | |
2c08497a | 5291 | continue; |
85ff4ec6 BC |
5292 | |
5293 | if (uid < gimple_uid (cand->incremented_at) | |
5294 | && (closest_before == NULL | |
5295 | || uid > gimple_uid (closest_before->stmt))) | |
5296 | closest_before = use; | |
5297 | ||
5298 | if (uid > gimple_uid (cand->incremented_at) | |
5299 | && (closest_after == NULL | |
5300 | || uid < gimple_uid (closest_after->stmt))) | |
5301 | closest_after = use; | |
2c08497a | 5302 | } |
85ff4ec6 BC |
5303 | |
5304 | if (closest_before != NULL | |
5305 | && autoinc_possible_for_pair (data, closest_before, cand)) | |
5306 | cand->ainc_use = closest_before; | |
5307 | else if (closest_after != NULL | |
5308 | && autoinc_possible_for_pair (data, closest_after, cand)) | |
5309 | cand->ainc_use = closest_after; | |
2c08497a BS |
5310 | } |
5311 | } | |
5312 | ||
5313 | /* Finds the candidates for the induction variables. */ | |
5314 | ||
5315 | static void | |
5316 | find_iv_candidates (struct ivopts_data *data) | |
5317 | { | |
5318 | /* Add commonly used ivs. */ | |
5319 | add_standard_iv_candidates (data); | |
5320 | ||
5321 | /* Add old induction variables. */ | |
5322 | add_old_ivs_candidates (data); | |
5323 | ||
5324 | /* Add induction variables derived from uses. */ | |
5325 | add_derived_ivs_candidates (data); | |
5326 | ||
5327 | set_autoinc_for_original_candidates (data); | |
5328 | ||
5329 | /* Record the important candidates. */ | |
5330 | record_important_candidates (data); | |
5331 | } | |
5332 | ||
8b11a64c ZD |
5333 | /* Determines costs of basing the use of the iv on an iv candidate. */ |
5334 | ||
5335 | static void | |
5336 | determine_use_iv_costs (struct ivopts_data *data) | |
5337 | { | |
5338 | unsigned i, j; | |
5339 | struct iv_use *use; | |
5340 | struct iv_cand *cand; | |
8bdbfff5 | 5341 | bitmap to_clear = BITMAP_ALLOC (NULL); |
8b11a64c ZD |
5342 | |
5343 | alloc_use_cost_map (data); | |
5344 | ||
8b11a64c ZD |
5345 | for (i = 0; i < n_iv_uses (data); i++) |
5346 | { | |
5347 | use = iv_use (data, i); | |
5348 | ||
5349 | if (data->consider_all_candidates) | |
5350 | { | |
5351 | for (j = 0; j < n_iv_cands (data); j++) | |
5352 | { | |
5353 | cand = iv_cand (data, j); | |
5354 | determine_use_iv_cost (data, use, cand); | |
5355 | } | |
5356 | } | |
5357 | else | |
5358 | { | |
87c476a2 ZD |
5359 | bitmap_iterator bi; |
5360 | ||
5361 | EXECUTE_IF_SET_IN_BITMAP (use->related_cands, 0, j, bi) | |
8b11a64c ZD |
5362 | { |
5363 | cand = iv_cand (data, j); | |
b1b02be2 ZD |
5364 | if (!determine_use_iv_cost (data, use, cand)) |
5365 | bitmap_set_bit (to_clear, j); | |
87c476a2 | 5366 | } |
b1b02be2 ZD |
5367 | |
5368 | /* Remove the candidates for that the cost is infinite from | |
5369 | the list of related candidates. */ | |
5370 | bitmap_and_compl_into (use->related_cands, to_clear); | |
5371 | bitmap_clear (to_clear); | |
8b11a64c ZD |
5372 | } |
5373 | } | |
5374 | ||
8bdbfff5 | 5375 | BITMAP_FREE (to_clear); |
b1b02be2 | 5376 | |
8b11a64c ZD |
5377 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5378 | { | |
5379 | fprintf (dump_file, "Use-candidate costs:\n"); | |
5380 | ||
5381 | for (i = 0; i < n_iv_uses (data); i++) | |
5382 | { | |
5383 | use = iv_use (data, i); | |
5384 | ||
5385 | fprintf (dump_file, "Use %d:\n", i); | |
6e8c65f6 | 5386 | fprintf (dump_file, " cand\tcost\tcompl.\tdepends on\n"); |
8b11a64c ZD |
5387 | for (j = 0; j < use->n_map_members; j++) |
5388 | { | |
5389 | if (!use->cost_map[j].cand | |
6e8c65f6 | 5390 | || infinite_cost_p (use->cost_map[j].cost)) |
8b11a64c ZD |
5391 | continue; |
5392 | ||
6e8c65f6 | 5393 | fprintf (dump_file, " %d\t%d\t%d\t", |
8b11a64c | 5394 | use->cost_map[j].cand->id, |
6e8c65f6 ZD |
5395 | use->cost_map[j].cost.cost, |
5396 | use->cost_map[j].cost.complexity); | |
8b11a64c ZD |
5397 | if (use->cost_map[j].depends_on) |
5398 | bitmap_print (dump_file, | |
5399 | use->cost_map[j].depends_on, "",""); | |
18081149 XDL |
5400 | if (use->cost_map[j].inv_expr_id != -1) |
5401 | fprintf (dump_file, " inv_expr:%d", use->cost_map[j].inv_expr_id); | |
8b11a64c ZD |
5402 | fprintf (dump_file, "\n"); |
5403 | } | |
5404 | ||
5405 | fprintf (dump_file, "\n"); | |
5406 | } | |
5407 | fprintf (dump_file, "\n"); | |
5408 | } | |
5409 | } | |
5410 | ||
5411 | /* Determines cost of the candidate CAND. */ | |
5412 | ||
5413 | static void | |
5414 | determine_iv_cost (struct ivopts_data *data, struct iv_cand *cand) | |
5415 | { | |
6e8c65f6 ZD |
5416 | comp_cost cost_base; |
5417 | unsigned cost, cost_step; | |
4366cf6d | 5418 | tree base; |
8b11a64c ZD |
5419 | |
5420 | if (!cand->iv) | |
5421 | { | |
5422 | cand->cost = 0; | |
5423 | return; | |
5424 | } | |
5425 | ||
5426 | /* There are two costs associated with the candidate -- its increment | |
5427 | and its initialization. The second is almost negligible for any loop | |
5428 | that rolls enough, so we take it just very little into account. */ | |
5429 | ||
5430 | base = cand->iv->base; | |
5431 | cost_base = force_var_cost (data, base, NULL); | |
a53c5024 TV |
5432 | /* It will be exceptional that the iv register happens to be initialized with |
5433 | the proper value at no cost. In general, there will at least be a regcopy | |
5434 | or a const set. */ | |
5435 | if (cost_base.cost == 0) | |
5436 | cost_base.cost = COSTS_N_INSNS (1); | |
5322d07e | 5437 | cost_step = add_cost (data->speed, TYPE_MODE (TREE_TYPE (base))); |
8b11a64c | 5438 | |
6521ac85 | 5439 | cost = cost_step + adjust_setup_cost (data, cost_base.cost); |
8b11a64c | 5440 | |
6e8c65f6 | 5441 | /* Prefer the original ivs unless we may gain something by replacing it. |
fa10beec | 5442 | The reason is to make debugging simpler; so this is not relevant for |
6e8c65f6 ZD |
5443 | artificial ivs created by other optimization passes. */ |
5444 | if (cand->pos != IP_ORIGINAL | |
70b5e7dc | 5445 | || !SSA_NAME_VAR (cand->var_before) |
6e8c65f6 ZD |
5446 | || DECL_ARTIFICIAL (SSA_NAME_VAR (cand->var_before))) |
5447 | cost++; | |
b8698a0f | 5448 | |
8b11a64c ZD |
5449 | /* Prefer not to insert statements into latch unless there are some |
5450 | already (so that we do not create unnecessary jumps). */ | |
4366cf6d ZD |
5451 | if (cand->pos == IP_END |
5452 | && empty_block_p (ip_end_pos (data->current_loop))) | |
6e8c65f6 ZD |
5453 | cost++; |
5454 | ||
5455 | cand->cost = cost; | |
2c08497a | 5456 | cand->cost_step = cost_step; |
8b11a64c ZD |
5457 | } |
5458 | ||
5459 | /* Determines costs of computation of the candidates. */ | |
5460 | ||
5461 | static void | |
5462 | determine_iv_costs (struct ivopts_data *data) | |
5463 | { | |
5464 | unsigned i; | |
5465 | ||
5466 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
5467 | { | |
5468 | fprintf (dump_file, "Candidate costs:\n"); | |
5469 | fprintf (dump_file, " cand\tcost\n"); | |
5470 | } | |
5471 | ||
5472 | for (i = 0; i < n_iv_cands (data); i++) | |
5473 | { | |
5474 | struct iv_cand *cand = iv_cand (data, i); | |
5475 | ||
5476 | determine_iv_cost (data, cand); | |
5477 | ||
5478 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
5479 | fprintf (dump_file, " %d\t%d\n", i, cand->cost); | |
5480 | } | |
2c08497a | 5481 | |
0f14009a BS |
5482 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5483 | fprintf (dump_file, "\n"); | |
8b11a64c ZD |
5484 | } |
5485 | ||
5486 | /* Calculates cost for having SIZE induction variables. */ | |
5487 | ||
5488 | static unsigned | |
5489 | ivopts_global_cost_for_size (struct ivopts_data *data, unsigned size) | |
5490 | { | |
a154b43a ZD |
5491 | /* We add size to the cost, so that we prefer eliminating ivs |
5492 | if possible. */ | |
bec922f0 SL |
5493 | return size + estimate_reg_pressure_cost (size, data->regs_used, data->speed, |
5494 | data->body_includes_call); | |
8b11a64c ZD |
5495 | } |
5496 | ||
5497 | /* For each size of the induction variable set determine the penalty. */ | |
5498 | ||
5499 | static void | |
5500 | determine_set_costs (struct ivopts_data *data) | |
5501 | { | |
5502 | unsigned j, n; | |
538dd0b7 DM |
5503 | gphi *phi; |
5504 | gphi_iterator psi; | |
726a989a | 5505 | tree op; |
8b11a64c | 5506 | struct loop *loop = data->current_loop; |
87c476a2 | 5507 | bitmap_iterator bi; |
8b11a64c | 5508 | |
8b11a64c ZD |
5509 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5510 | { | |
5511 | fprintf (dump_file, "Global costs:\n"); | |
5512 | fprintf (dump_file, " target_avail_regs %d\n", target_avail_regs); | |
bec922f0 | 5513 | fprintf (dump_file, " target_clobbered_regs %d\n", target_clobbered_regs); |
f40751dd JH |
5514 | fprintf (dump_file, " target_reg_cost %d\n", target_reg_cost[data->speed]); |
5515 | fprintf (dump_file, " target_spill_cost %d\n", target_spill_cost[data->speed]); | |
8b11a64c ZD |
5516 | } |
5517 | ||
5518 | n = 0; | |
726a989a | 5519 | for (psi = gsi_start_phis (loop->header); !gsi_end_p (psi); gsi_next (&psi)) |
8b11a64c | 5520 | { |
538dd0b7 | 5521 | phi = psi.phi (); |
8b11a64c ZD |
5522 | op = PHI_RESULT (phi); |
5523 | ||
ea057359 | 5524 | if (virtual_operand_p (op)) |
8b11a64c ZD |
5525 | continue; |
5526 | ||
5527 | if (get_iv (data, op)) | |
5528 | continue; | |
5529 | ||
5530 | n++; | |
5531 | } | |
5532 | ||
87c476a2 | 5533 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, j, bi) |
8b11a64c ZD |
5534 | { |
5535 | struct version_info *info = ver_info (data, j); | |
5536 | ||
5537 | if (info->inv_id && info->has_nonlin_use) | |
5538 | n++; | |
87c476a2 | 5539 | } |
8b11a64c | 5540 | |
9a2ef6b8 | 5541 | data->regs_used = n; |
8b11a64c ZD |
5542 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5543 | fprintf (dump_file, " regs_used %d\n", n); | |
5544 | ||
5545 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
5546 | { | |
5547 | fprintf (dump_file, " cost for size:\n"); | |
5548 | fprintf (dump_file, " ivs\tcost\n"); | |
5549 | for (j = 0; j <= 2 * target_avail_regs; j++) | |
5550 | fprintf (dump_file, " %d\t%d\n", j, | |
5551 | ivopts_global_cost_for_size (data, j)); | |
5552 | fprintf (dump_file, "\n"); | |
5553 | } | |
5554 | } | |
5555 | ||
b1b02be2 | 5556 | /* Returns true if A is a cheaper cost pair than B. */ |
8b11a64c | 5557 | |
b1b02be2 ZD |
5558 | static bool |
5559 | cheaper_cost_pair (struct cost_pair *a, struct cost_pair *b) | |
8b11a64c | 5560 | { |
6e8c65f6 ZD |
5561 | int cmp; |
5562 | ||
b1b02be2 ZD |
5563 | if (!a) |
5564 | return false; | |
8b11a64c | 5565 | |
b1b02be2 ZD |
5566 | if (!b) |
5567 | return true; | |
5568 | ||
6e8c65f6 ZD |
5569 | cmp = compare_costs (a->cost, b->cost); |
5570 | if (cmp < 0) | |
b1b02be2 ZD |
5571 | return true; |
5572 | ||
6e8c65f6 | 5573 | if (cmp > 0) |
b1b02be2 ZD |
5574 | return false; |
5575 | ||
5576 | /* In case the costs are the same, prefer the cheaper candidate. */ | |
5577 | if (a->cand->cost < b->cand->cost) | |
5578 | return true; | |
5579 | ||
5580 | return false; | |
5581 | } | |
5582 | ||
18081149 XDL |
5583 | |
5584 | /* Returns candidate by that USE is expressed in IVS. */ | |
5585 | ||
5586 | static struct cost_pair * | |
5587 | iv_ca_cand_for_use (struct iv_ca *ivs, struct iv_use *use) | |
5588 | { | |
5589 | return ivs->cand_for_use[use->id]; | |
5590 | } | |
5591 | ||
b1b02be2 ZD |
5592 | /* Computes the cost field of IVS structure. */ |
5593 | ||
5594 | static void | |
5595 | iv_ca_recount_cost (struct ivopts_data *data, struct iv_ca *ivs) | |
5596 | { | |
6e8c65f6 | 5597 | comp_cost cost = ivs->cand_use_cost; |
18081149 | 5598 | |
6e8c65f6 | 5599 | cost.cost += ivs->cand_cost; |
18081149 | 5600 | |
18081149 | 5601 | cost.cost += ivopts_global_cost_for_size (data, |
f06e400f | 5602 | ivs->n_regs + ivs->num_used_inv_expr); |
b1b02be2 ZD |
5603 | |
5604 | ivs->cost = cost; | |
5605 | } | |
5606 | ||
9be872b7 ZD |
5607 | /* Remove invariants in set INVS to set IVS. */ |
5608 | ||
5609 | static void | |
5610 | iv_ca_set_remove_invariants (struct iv_ca *ivs, bitmap invs) | |
5611 | { | |
5612 | bitmap_iterator bi; | |
5613 | unsigned iid; | |
5614 | ||
5615 | if (!invs) | |
5616 | return; | |
5617 | ||
5618 | EXECUTE_IF_SET_IN_BITMAP (invs, 0, iid, bi) | |
5619 | { | |
5620 | ivs->n_invariant_uses[iid]--; | |
5621 | if (ivs->n_invariant_uses[iid] == 0) | |
18081149 | 5622 | ivs->n_regs--; |
9be872b7 ZD |
5623 | } |
5624 | } | |
5625 | ||
b1b02be2 ZD |
5626 | /* Set USE not to be expressed by any candidate in IVS. */ |
5627 | ||
5628 | static void | |
5629 | iv_ca_set_no_cp (struct ivopts_data *data, struct iv_ca *ivs, | |
5630 | struct iv_use *use) | |
5631 | { | |
9be872b7 | 5632 | unsigned uid = use->id, cid; |
b1b02be2 | 5633 | struct cost_pair *cp; |
b1b02be2 ZD |
5634 | |
5635 | cp = ivs->cand_for_use[uid]; | |
5636 | if (!cp) | |
5637 | return; | |
5638 | cid = cp->cand->id; | |
5639 | ||
5640 | ivs->bad_uses++; | |
5641 | ivs->cand_for_use[uid] = NULL; | |
5642 | ivs->n_cand_uses[cid]--; | |
5643 | ||
5644 | if (ivs->n_cand_uses[cid] == 0) | |
8b11a64c | 5645 | { |
b1b02be2 ZD |
5646 | bitmap_clear_bit (ivs->cands, cid); |
5647 | /* Do not count the pseudocandidates. */ | |
5648 | if (cp->cand->iv) | |
5649 | ivs->n_regs--; | |
36f5ada1 | 5650 | ivs->n_cands--; |
b1b02be2 | 5651 | ivs->cand_cost -= cp->cand->cost; |
9be872b7 ZD |
5652 | |
5653 | iv_ca_set_remove_invariants (ivs, cp->cand->depends_on); | |
b1b02be2 ZD |
5654 | } |
5655 | ||
6e8c65f6 | 5656 | ivs->cand_use_cost = sub_costs (ivs->cand_use_cost, cp->cost); |
b1b02be2 | 5657 | |
9be872b7 | 5658 | iv_ca_set_remove_invariants (ivs, cp->depends_on); |
f06e400f XDL |
5659 | |
5660 | if (cp->inv_expr_id != -1) | |
5661 | { | |
5662 | ivs->used_inv_expr[cp->inv_expr_id]--; | |
5663 | if (ivs->used_inv_expr[cp->inv_expr_id] == 0) | |
5664 | ivs->num_used_inv_expr--; | |
5665 | } | |
9be872b7 ZD |
5666 | iv_ca_recount_cost (data, ivs); |
5667 | } | |
5668 | ||
5669 | /* Add invariants in set INVS to set IVS. */ | |
80cad5fa | 5670 | |
9be872b7 ZD |
5671 | static void |
5672 | iv_ca_set_add_invariants (struct iv_ca *ivs, bitmap invs) | |
5673 | { | |
5674 | bitmap_iterator bi; | |
5675 | unsigned iid; | |
5676 | ||
5677 | if (!invs) | |
5678 | return; | |
5679 | ||
5680 | EXECUTE_IF_SET_IN_BITMAP (invs, 0, iid, bi) | |
b1b02be2 | 5681 | { |
9be872b7 ZD |
5682 | ivs->n_invariant_uses[iid]++; |
5683 | if (ivs->n_invariant_uses[iid] == 1) | |
18081149 | 5684 | ivs->n_regs++; |
8b11a64c | 5685 | } |
b1b02be2 ZD |
5686 | } |
5687 | ||
5688 | /* Set cost pair for USE in set IVS to CP. */ | |
5689 | ||
5690 | static void | |
5691 | iv_ca_set_cp (struct ivopts_data *data, struct iv_ca *ivs, | |
5692 | struct iv_use *use, struct cost_pair *cp) | |
5693 | { | |
9be872b7 | 5694 | unsigned uid = use->id, cid; |
b1b02be2 ZD |
5695 | |
5696 | if (ivs->cand_for_use[uid] == cp) | |
5697 | return; | |
5698 | ||
5699 | if (ivs->cand_for_use[uid]) | |
5700 | iv_ca_set_no_cp (data, ivs, use); | |
5701 | ||
5702 | if (cp) | |
8b11a64c | 5703 | { |
b1b02be2 | 5704 | cid = cp->cand->id; |
8b11a64c | 5705 | |
b1b02be2 ZD |
5706 | ivs->bad_uses--; |
5707 | ivs->cand_for_use[uid] = cp; | |
5708 | ivs->n_cand_uses[cid]++; | |
5709 | if (ivs->n_cand_uses[cid] == 1) | |
8b11a64c | 5710 | { |
b1b02be2 ZD |
5711 | bitmap_set_bit (ivs->cands, cid); |
5712 | /* Do not count the pseudocandidates. */ | |
5713 | if (cp->cand->iv) | |
5714 | ivs->n_regs++; | |
36f5ada1 | 5715 | ivs->n_cands++; |
b1b02be2 | 5716 | ivs->cand_cost += cp->cand->cost; |
b1b02be2 | 5717 | |
9be872b7 | 5718 | iv_ca_set_add_invariants (ivs, cp->cand->depends_on); |
8b11a64c ZD |
5719 | } |
5720 | ||
6e8c65f6 | 5721 | ivs->cand_use_cost = add_costs (ivs->cand_use_cost, cp->cost); |
9be872b7 | 5722 | iv_ca_set_add_invariants (ivs, cp->depends_on); |
f06e400f XDL |
5723 | |
5724 | if (cp->inv_expr_id != -1) | |
5725 | { | |
5726 | ivs->used_inv_expr[cp->inv_expr_id]++; | |
5727 | if (ivs->used_inv_expr[cp->inv_expr_id] == 1) | |
5728 | ivs->num_used_inv_expr++; | |
5729 | } | |
b1b02be2 | 5730 | iv_ca_recount_cost (data, ivs); |
87c476a2 | 5731 | } |
b1b02be2 ZD |
5732 | } |
5733 | ||
5734 | /* Extend set IVS by expressing USE by some of the candidates in it | |
f22ae1ec BC |
5735 | if possible. Consider all important candidates if candidates in |
5736 | set IVS don't give any result. */ | |
b1b02be2 ZD |
5737 | |
5738 | static void | |
5739 | iv_ca_add_use (struct ivopts_data *data, struct iv_ca *ivs, | |
f22ae1ec | 5740 | struct iv_use *use) |
b1b02be2 ZD |
5741 | { |
5742 | struct cost_pair *best_cp = NULL, *cp; | |
5743 | bitmap_iterator bi; | |
5744 | unsigned i; | |
f22ae1ec | 5745 | struct iv_cand *cand; |
8b11a64c | 5746 | |
b1b02be2 | 5747 | gcc_assert (ivs->upto >= use->id); |
f22ae1ec BC |
5748 | ivs->upto++; |
5749 | ivs->bad_uses++; | |
b1b02be2 | 5750 | |
f22ae1ec | 5751 | EXECUTE_IF_SET_IN_BITMAP (ivs->cands, 0, i, bi) |
b1b02be2 | 5752 | { |
f22ae1ec | 5753 | cand = iv_cand (data, i); |
18081149 | 5754 | cp = get_use_iv_cost (data, use, cand); |
b1b02be2 ZD |
5755 | if (cheaper_cost_pair (cp, best_cp)) |
5756 | best_cp = cp; | |
5757 | } | |
f22ae1ec BC |
5758 | |
5759 | if (best_cp == NULL) | |
5760 | { | |
5761 | EXECUTE_IF_SET_IN_BITMAP (data->important_candidates, 0, i, bi) | |
5762 | { | |
5763 | cand = iv_cand (data, i); | |
5764 | cp = get_use_iv_cost (data, use, cand); | |
5765 | if (cheaper_cost_pair (cp, best_cp)) | |
5766 | best_cp = cp; | |
5767 | } | |
5768 | } | |
8b11a64c | 5769 | |
b1b02be2 | 5770 | iv_ca_set_cp (data, ivs, use, best_cp); |
8b11a64c ZD |
5771 | } |
5772 | ||
b1b02be2 | 5773 | /* Get cost for assignment IVS. */ |
8b11a64c | 5774 | |
6e8c65f6 | 5775 | static comp_cost |
b1b02be2 ZD |
5776 | iv_ca_cost (struct iv_ca *ivs) |
5777 | { | |
c4e93e28 AH |
5778 | /* This was a conditional expression but it triggered a bug in |
5779 | Sun C 5.5. */ | |
cb4ad180 AH |
5780 | if (ivs->bad_uses) |
5781 | return infinite_cost; | |
5782 | else | |
5783 | return ivs->cost; | |
b1b02be2 ZD |
5784 | } |
5785 | ||
5786 | /* Returns true if all dependences of CP are among invariants in IVS. */ | |
5787 | ||
5788 | static bool | |
5789 | iv_ca_has_deps (struct iv_ca *ivs, struct cost_pair *cp) | |
8b11a64c ZD |
5790 | { |
5791 | unsigned i; | |
87c476a2 | 5792 | bitmap_iterator bi; |
8b11a64c | 5793 | |
b1b02be2 ZD |
5794 | if (!cp->depends_on) |
5795 | return true; | |
5796 | ||
5797 | EXECUTE_IF_SET_IN_BITMAP (cp->depends_on, 0, i, bi) | |
8b11a64c | 5798 | { |
b1b02be2 ZD |
5799 | if (ivs->n_invariant_uses[i] == 0) |
5800 | return false; | |
5801 | } | |
5802 | ||
5803 | return true; | |
5804 | } | |
5805 | ||
5806 | /* Creates change of expressing USE by NEW_CP instead of OLD_CP and chains | |
5807 | it before NEXT_CHANGE. */ | |
5808 | ||
5809 | static struct iv_ca_delta * | |
5810 | iv_ca_delta_add (struct iv_use *use, struct cost_pair *old_cp, | |
5811 | struct cost_pair *new_cp, struct iv_ca_delta *next_change) | |
5812 | { | |
5ed6ace5 | 5813 | struct iv_ca_delta *change = XNEW (struct iv_ca_delta); |
b1b02be2 ZD |
5814 | |
5815 | change->use = use; | |
5816 | change->old_cp = old_cp; | |
5817 | change->new_cp = new_cp; | |
5818 | change->next_change = next_change; | |
5819 | ||
5820 | return change; | |
5821 | } | |
5822 | ||
36f5ada1 | 5823 | /* Joins two lists of changes L1 and L2. Destructive -- old lists |
6c6cfbfd | 5824 | are rewritten. */ |
36f5ada1 ZD |
5825 | |
5826 | static struct iv_ca_delta * | |
5827 | iv_ca_delta_join (struct iv_ca_delta *l1, struct iv_ca_delta *l2) | |
5828 | { | |
5829 | struct iv_ca_delta *last; | |
5830 | ||
5831 | if (!l2) | |
5832 | return l1; | |
5833 | ||
5834 | if (!l1) | |
5835 | return l2; | |
5836 | ||
5837 | for (last = l1; last->next_change; last = last->next_change) | |
5838 | continue; | |
5839 | last->next_change = l2; | |
5840 | ||
5841 | return l1; | |
5842 | } | |
5843 | ||
36f5ada1 ZD |
5844 | /* Reverse the list of changes DELTA, forming the inverse to it. */ |
5845 | ||
5846 | static struct iv_ca_delta * | |
5847 | iv_ca_delta_reverse (struct iv_ca_delta *delta) | |
5848 | { | |
5849 | struct iv_ca_delta *act, *next, *prev = NULL; | |
36f5ada1 ZD |
5850 | |
5851 | for (act = delta; act; act = next) | |
5852 | { | |
5853 | next = act->next_change; | |
5854 | act->next_change = prev; | |
5855 | prev = act; | |
5856 | ||
fab27f52 | 5857 | std::swap (act->old_cp, act->new_cp); |
36f5ada1 ZD |
5858 | } |
5859 | ||
5860 | return prev; | |
5861 | } | |
5862 | ||
b1b02be2 ZD |
5863 | /* Commit changes in DELTA to IVS. If FORWARD is false, the changes are |
5864 | reverted instead. */ | |
5865 | ||
5866 | static void | |
5867 | iv_ca_delta_commit (struct ivopts_data *data, struct iv_ca *ivs, | |
5868 | struct iv_ca_delta *delta, bool forward) | |
5869 | { | |
5870 | struct cost_pair *from, *to; | |
36f5ada1 | 5871 | struct iv_ca_delta *act; |
b1b02be2 | 5872 | |
36f5ada1 ZD |
5873 | if (!forward) |
5874 | delta = iv_ca_delta_reverse (delta); | |
b1b02be2 | 5875 | |
36f5ada1 ZD |
5876 | for (act = delta; act; act = act->next_change) |
5877 | { | |
5878 | from = act->old_cp; | |
5879 | to = act->new_cp; | |
5880 | gcc_assert (iv_ca_cand_for_use (ivs, act->use) == from); | |
5881 | iv_ca_set_cp (data, ivs, act->use, to); | |
8b11a64c | 5882 | } |
36f5ada1 ZD |
5883 | |
5884 | if (!forward) | |
5885 | iv_ca_delta_reverse (delta); | |
b1b02be2 | 5886 | } |
8b11a64c | 5887 | |
b1b02be2 | 5888 | /* Returns true if CAND is used in IVS. */ |
8b11a64c | 5889 | |
b1b02be2 ZD |
5890 | static bool |
5891 | iv_ca_cand_used_p (struct iv_ca *ivs, struct iv_cand *cand) | |
5892 | { | |
5893 | return ivs->n_cand_uses[cand->id] > 0; | |
5894 | } | |
8b11a64c | 5895 | |
36f5ada1 ZD |
5896 | /* Returns number of induction variable candidates in the set IVS. */ |
5897 | ||
5898 | static unsigned | |
5899 | iv_ca_n_cands (struct iv_ca *ivs) | |
5900 | { | |
5901 | return ivs->n_cands; | |
5902 | } | |
5903 | ||
b1b02be2 ZD |
5904 | /* Free the list of changes DELTA. */ |
5905 | ||
5906 | static void | |
5907 | iv_ca_delta_free (struct iv_ca_delta **delta) | |
5908 | { | |
5909 | struct iv_ca_delta *act, *next; | |
5910 | ||
5911 | for (act = *delta; act; act = next) | |
87c476a2 | 5912 | { |
b1b02be2 ZD |
5913 | next = act->next_change; |
5914 | free (act); | |
87c476a2 | 5915 | } |
8b11a64c | 5916 | |
b1b02be2 ZD |
5917 | *delta = NULL; |
5918 | } | |
5919 | ||
5920 | /* Allocates new iv candidates assignment. */ | |
5921 | ||
5922 | static struct iv_ca * | |
5923 | iv_ca_new (struct ivopts_data *data) | |
5924 | { | |
5ed6ace5 | 5925 | struct iv_ca *nw = XNEW (struct iv_ca); |
8b11a64c | 5926 | |
b1b02be2 ZD |
5927 | nw->upto = 0; |
5928 | nw->bad_uses = 0; | |
5ed6ace5 MD |
5929 | nw->cand_for_use = XCNEWVEC (struct cost_pair *, n_iv_uses (data)); |
5930 | nw->n_cand_uses = XCNEWVEC (unsigned, n_iv_cands (data)); | |
8bdbfff5 | 5931 | nw->cands = BITMAP_ALLOC (NULL); |
36f5ada1 | 5932 | nw->n_cands = 0; |
b1b02be2 | 5933 | nw->n_regs = 0; |
7735d6c7 | 5934 | nw->cand_use_cost = no_cost; |
b1b02be2 | 5935 | nw->cand_cost = 0; |
5ed6ace5 | 5936 | nw->n_invariant_uses = XCNEWVEC (unsigned, data->max_inv_id + 1); |
7735d6c7 | 5937 | nw->cost = no_cost; |
f06e400f XDL |
5938 | nw->used_inv_expr = XCNEWVEC (unsigned, data->inv_expr_id + 1); |
5939 | nw->num_used_inv_expr = 0; | |
b1b02be2 ZD |
5940 | |
5941 | return nw; | |
5942 | } | |
5943 | ||
5944 | /* Free memory occupied by the set IVS. */ | |
5945 | ||
5946 | static void | |
5947 | iv_ca_free (struct iv_ca **ivs) | |
5948 | { | |
5949 | free ((*ivs)->cand_for_use); | |
5950 | free ((*ivs)->n_cand_uses); | |
8bdbfff5 | 5951 | BITMAP_FREE ((*ivs)->cands); |
b1b02be2 | 5952 | free ((*ivs)->n_invariant_uses); |
f06e400f | 5953 | free ((*ivs)->used_inv_expr); |
b1b02be2 ZD |
5954 | free (*ivs); |
5955 | *ivs = NULL; | |
5956 | } | |
5957 | ||
5958 | /* Dumps IVS to FILE. */ | |
5959 | ||
5960 | static void | |
5961 | iv_ca_dump (struct ivopts_data *data, FILE *file, struct iv_ca *ivs) | |
5962 | { | |
5963 | const char *pref = " invariants "; | |
5964 | unsigned i; | |
6e8c65f6 | 5965 | comp_cost cost = iv_ca_cost (ivs); |
b1b02be2 | 5966 | |
18081149 XDL |
5967 | fprintf (file, " cost: %d (complexity %d)\n", cost.cost, cost.complexity); |
5968 | fprintf (file, " cand_cost: %d\n cand_use_cost: %d (complexity %d)\n", | |
5969 | ivs->cand_cost, ivs->cand_use_cost.cost, ivs->cand_use_cost.complexity); | |
5970 | bitmap_print (file, ivs->cands, " candidates: ","\n"); | |
5971 | ||
5972 | for (i = 0; i < ivs->upto; i++) | |
5973 | { | |
5974 | struct iv_use *use = iv_use (data, i); | |
5975 | struct cost_pair *cp = iv_ca_cand_for_use (ivs, use); | |
5976 | if (cp) | |
5977 | fprintf (file, " use:%d --> iv_cand:%d, cost=(%d,%d)\n", | |
5978 | use->id, cp->cand->id, cp->cost.cost, cp->cost.complexity); | |
5979 | else | |
5980 | fprintf (file, " use:%d --> ??\n", use->id); | |
5981 | } | |
b1b02be2 ZD |
5982 | |
5983 | for (i = 1; i <= data->max_inv_id; i++) | |
5984 | if (ivs->n_invariant_uses[i]) | |
5985 | { | |
5986 | fprintf (file, "%s%d", pref, i); | |
5987 | pref = ", "; | |
5988 | } | |
18081149 | 5989 | fprintf (file, "\n\n"); |
b1b02be2 ZD |
5990 | } |
5991 | ||
5992 | /* Try changing candidate in IVS to CAND for each use. Return cost of the | |
36f5ada1 | 5993 | new set, and store differences in DELTA. Number of induction variables |
18081149 XDL |
5994 | in the new set is stored to N_IVS. MIN_NCAND is a flag. When it is true |
5995 | the function will try to find a solution with mimimal iv candidates. */ | |
b1b02be2 | 5996 | |
6e8c65f6 | 5997 | static comp_cost |
b1b02be2 | 5998 | iv_ca_extend (struct ivopts_data *data, struct iv_ca *ivs, |
36f5ada1 | 5999 | struct iv_cand *cand, struct iv_ca_delta **delta, |
18081149 | 6000 | unsigned *n_ivs, bool min_ncand) |
b1b02be2 | 6001 | { |
6e8c65f6 ZD |
6002 | unsigned i; |
6003 | comp_cost cost; | |
b1b02be2 ZD |
6004 | struct iv_use *use; |
6005 | struct cost_pair *old_cp, *new_cp; | |
6006 | ||
6007 | *delta = NULL; | |
6008 | for (i = 0; i < ivs->upto; i++) | |
6009 | { | |
6010 | use = iv_use (data, i); | |
6011 | old_cp = iv_ca_cand_for_use (ivs, use); | |
6012 | ||
6013 | if (old_cp | |
6014 | && old_cp->cand == cand) | |
6015 | continue; | |
6016 | ||
6017 | new_cp = get_use_iv_cost (data, use, cand); | |
6018 | if (!new_cp) | |
6019 | continue; | |
6020 | ||
18081149 | 6021 | if (!min_ncand && !iv_ca_has_deps (ivs, new_cp)) |
b1b02be2 | 6022 | continue; |
b8698a0f | 6023 | |
18081149 XDL |
6024 | if (!min_ncand && !cheaper_cost_pair (new_cp, old_cp)) |
6025 | continue; | |
b1b02be2 ZD |
6026 | |
6027 | *delta = iv_ca_delta_add (use, old_cp, new_cp, *delta); | |
6028 | } | |
6029 | ||
6030 | iv_ca_delta_commit (data, ivs, *delta, true); | |
6031 | cost = iv_ca_cost (ivs); | |
36f5ada1 ZD |
6032 | if (n_ivs) |
6033 | *n_ivs = iv_ca_n_cands (ivs); | |
b1b02be2 | 6034 | iv_ca_delta_commit (data, ivs, *delta, false); |
8b11a64c ZD |
6035 | |
6036 | return cost; | |
6037 | } | |
6038 | ||
a0eca485 | 6039 | /* Try narrowing set IVS by removing CAND. Return the cost of |
2c407426 BC |
6040 | the new set and store the differences in DELTA. START is |
6041 | the candidate with which we start narrowing. */ | |
8b11a64c | 6042 | |
6e8c65f6 | 6043 | static comp_cost |
b1b02be2 | 6044 | iv_ca_narrow (struct ivopts_data *data, struct iv_ca *ivs, |
2c407426 BC |
6045 | struct iv_cand *cand, struct iv_cand *start, |
6046 | struct iv_ca_delta **delta) | |
8b11a64c | 6047 | { |
b1b02be2 ZD |
6048 | unsigned i, ci; |
6049 | struct iv_use *use; | |
6050 | struct cost_pair *old_cp, *new_cp, *cp; | |
6051 | bitmap_iterator bi; | |
6052 | struct iv_cand *cnd; | |
2c407426 | 6053 | comp_cost cost, best_cost, acost; |
b1b02be2 ZD |
6054 | |
6055 | *delta = NULL; | |
6056 | for (i = 0; i < n_iv_uses (data); i++) | |
6057 | { | |
6058 | use = iv_use (data, i); | |
6059 | ||
6060 | old_cp = iv_ca_cand_for_use (ivs, use); | |
6061 | if (old_cp->cand != cand) | |
6062 | continue; | |
6063 | ||
2c407426 BC |
6064 | best_cost = iv_ca_cost (ivs); |
6065 | /* Start narrowing with START. */ | |
6066 | new_cp = get_use_iv_cost (data, use, start); | |
b1b02be2 ZD |
6067 | |
6068 | if (data->consider_all_candidates) | |
6069 | { | |
6070 | EXECUTE_IF_SET_IN_BITMAP (ivs->cands, 0, ci, bi) | |
6071 | { | |
2c407426 | 6072 | if (ci == cand->id || (start && ci == start->id)) |
b1b02be2 ZD |
6073 | continue; |
6074 | ||
6075 | cnd = iv_cand (data, ci); | |
6076 | ||
6077 | cp = get_use_iv_cost (data, use, cnd); | |
6078 | if (!cp) | |
6079 | continue; | |
18081149 | 6080 | |
2c407426 BC |
6081 | iv_ca_set_cp (data, ivs, use, cp); |
6082 | acost = iv_ca_cost (ivs); | |
b1b02be2 | 6083 | |
2c407426 BC |
6084 | if (compare_costs (acost, best_cost) < 0) |
6085 | { | |
6086 | best_cost = acost; | |
6087 | new_cp = cp; | |
6088 | } | |
b1b02be2 ZD |
6089 | } |
6090 | } | |
6091 | else | |
6092 | { | |
6093 | EXECUTE_IF_AND_IN_BITMAP (use->related_cands, ivs->cands, 0, ci, bi) | |
6094 | { | |
2c407426 | 6095 | if (ci == cand->id || (start && ci == start->id)) |
b1b02be2 ZD |
6096 | continue; |
6097 | ||
6098 | cnd = iv_cand (data, ci); | |
6099 | ||
6100 | cp = get_use_iv_cost (data, use, cnd); | |
6101 | if (!cp) | |
6102 | continue; | |
b8698a0f | 6103 | |
2c407426 BC |
6104 | iv_ca_set_cp (data, ivs, use, cp); |
6105 | acost = iv_ca_cost (ivs); | |
b1b02be2 | 6106 | |
2c407426 BC |
6107 | if (compare_costs (acost, best_cost) < 0) |
6108 | { | |
6109 | best_cost = acost; | |
6110 | new_cp = cp; | |
6111 | } | |
b1b02be2 ZD |
6112 | } |
6113 | } | |
2c407426 BC |
6114 | /* Restore to old cp for use. */ |
6115 | iv_ca_set_cp (data, ivs, use, old_cp); | |
b1b02be2 ZD |
6116 | |
6117 | if (!new_cp) | |
6118 | { | |
6119 | iv_ca_delta_free (delta); | |
6e8c65f6 | 6120 | return infinite_cost; |
b1b02be2 ZD |
6121 | } |
6122 | ||
6123 | *delta = iv_ca_delta_add (use, old_cp, new_cp, *delta); | |
6124 | } | |
6125 | ||
6126 | iv_ca_delta_commit (data, ivs, *delta, true); | |
6127 | cost = iv_ca_cost (ivs); | |
6128 | iv_ca_delta_commit (data, ivs, *delta, false); | |
6129 | ||
6130 | return cost; | |
8b11a64c ZD |
6131 | } |
6132 | ||
36f5ada1 ZD |
6133 | /* Try optimizing the set of candidates IVS by removing candidates different |
6134 | from to EXCEPT_CAND from it. Return cost of the new set, and store | |
6135 | differences in DELTA. */ | |
6136 | ||
6e8c65f6 | 6137 | static comp_cost |
36f5ada1 ZD |
6138 | iv_ca_prune (struct ivopts_data *data, struct iv_ca *ivs, |
6139 | struct iv_cand *except_cand, struct iv_ca_delta **delta) | |
6140 | { | |
6141 | bitmap_iterator bi; | |
6142 | struct iv_ca_delta *act_delta, *best_delta; | |
6e8c65f6 ZD |
6143 | unsigned i; |
6144 | comp_cost best_cost, acost; | |
36f5ada1 ZD |
6145 | struct iv_cand *cand; |
6146 | ||
6147 | best_delta = NULL; | |
6148 | best_cost = iv_ca_cost (ivs); | |
6149 | ||
6150 | EXECUTE_IF_SET_IN_BITMAP (ivs->cands, 0, i, bi) | |
6151 | { | |
6152 | cand = iv_cand (data, i); | |
6153 | ||
6154 | if (cand == except_cand) | |
6155 | continue; | |
6156 | ||
2c407426 | 6157 | acost = iv_ca_narrow (data, ivs, cand, except_cand, &act_delta); |
36f5ada1 | 6158 | |
6e8c65f6 | 6159 | if (compare_costs (acost, best_cost) < 0) |
36f5ada1 ZD |
6160 | { |
6161 | best_cost = acost; | |
6162 | iv_ca_delta_free (&best_delta); | |
6163 | best_delta = act_delta; | |
6164 | } | |
6165 | else | |
6166 | iv_ca_delta_free (&act_delta); | |
6167 | } | |
6168 | ||
6169 | if (!best_delta) | |
6170 | { | |
6171 | *delta = NULL; | |
6172 | return best_cost; | |
6173 | } | |
6174 | ||
6175 | /* Recurse to possibly remove other unnecessary ivs. */ | |
6176 | iv_ca_delta_commit (data, ivs, best_delta, true); | |
6177 | best_cost = iv_ca_prune (data, ivs, except_cand, delta); | |
6178 | iv_ca_delta_commit (data, ivs, best_delta, false); | |
6179 | *delta = iv_ca_delta_join (best_delta, *delta); | |
6180 | return best_cost; | |
6181 | } | |
6182 | ||
6326a5f5 BC |
6183 | /* Check if CAND_IDX is a candidate other than OLD_CAND and has |
6184 | cheaper local cost for USE than BEST_CP. Return pointer to | |
6185 | the corresponding cost_pair, otherwise just return BEST_CP. */ | |
6186 | ||
6187 | static struct cost_pair* | |
6188 | cheaper_cost_with_cand (struct ivopts_data *data, struct iv_use *use, | |
6189 | unsigned int cand_idx, struct iv_cand *old_cand, | |
6190 | struct cost_pair *best_cp) | |
6191 | { | |
6192 | struct iv_cand *cand; | |
6193 | struct cost_pair *cp; | |
6194 | ||
6195 | gcc_assert (old_cand != NULL && best_cp != NULL); | |
6196 | if (cand_idx == old_cand->id) | |
6197 | return best_cp; | |
6198 | ||
6199 | cand = iv_cand (data, cand_idx); | |
6200 | cp = get_use_iv_cost (data, use, cand); | |
6201 | if (cp != NULL && cheaper_cost_pair (cp, best_cp)) | |
6202 | return cp; | |
6203 | ||
6204 | return best_cp; | |
6205 | } | |
6206 | ||
6207 | /* Try breaking local optimal fixed-point for IVS by replacing candidates | |
6208 | which are used by more than one iv uses. For each of those candidates, | |
6209 | this function tries to represent iv uses under that candidate using | |
6210 | other ones with lower local cost, then tries to prune the new set. | |
6211 | If the new set has lower cost, It returns the new cost after recording | |
6212 | candidate replacement in list DELTA. */ | |
6213 | ||
6214 | static comp_cost | |
6215 | iv_ca_replace (struct ivopts_data *data, struct iv_ca *ivs, | |
6216 | struct iv_ca_delta **delta) | |
6217 | { | |
6218 | bitmap_iterator bi, bj; | |
6219 | unsigned int i, j, k; | |
6220 | struct iv_use *use; | |
6221 | struct iv_cand *cand; | |
6222 | comp_cost orig_cost, acost; | |
6223 | struct iv_ca_delta *act_delta, *tmp_delta; | |
6224 | struct cost_pair *old_cp, *best_cp = NULL; | |
6225 | ||
6226 | *delta = NULL; | |
6227 | orig_cost = iv_ca_cost (ivs); | |
6228 | ||
6229 | EXECUTE_IF_SET_IN_BITMAP (ivs->cands, 0, i, bi) | |
6230 | { | |
6231 | if (ivs->n_cand_uses[i] == 1 | |
6232 | || ivs->n_cand_uses[i] > ALWAYS_PRUNE_CAND_SET_BOUND) | |
6233 | continue; | |
6234 | ||
6235 | cand = iv_cand (data, i); | |
6236 | ||
6237 | act_delta = NULL; | |
6238 | /* Represent uses under current candidate using other ones with | |
6239 | lower local cost. */ | |
6240 | for (j = 0; j < ivs->upto; j++) | |
6241 | { | |
6242 | use = iv_use (data, j); | |
6243 | old_cp = iv_ca_cand_for_use (ivs, use); | |
6244 | ||
6245 | if (old_cp->cand != cand) | |
6246 | continue; | |
6247 | ||
6248 | best_cp = old_cp; | |
6249 | if (data->consider_all_candidates) | |
6250 | for (k = 0; k < n_iv_cands (data); k++) | |
6251 | best_cp = cheaper_cost_with_cand (data, use, k, | |
6252 | old_cp->cand, best_cp); | |
6253 | else | |
6254 | EXECUTE_IF_SET_IN_BITMAP (use->related_cands, 0, k, bj) | |
6255 | best_cp = cheaper_cost_with_cand (data, use, k, | |
6256 | old_cp->cand, best_cp); | |
6257 | ||
6258 | if (best_cp == old_cp) | |
6259 | continue; | |
6260 | ||
6261 | act_delta = iv_ca_delta_add (use, old_cp, best_cp, act_delta); | |
6262 | } | |
6263 | /* No need for further prune. */ | |
6264 | if (!act_delta) | |
6265 | continue; | |
6266 | ||
6267 | /* Prune the new candidate set. */ | |
6268 | iv_ca_delta_commit (data, ivs, act_delta, true); | |
6269 | acost = iv_ca_prune (data, ivs, NULL, &tmp_delta); | |
6270 | iv_ca_delta_commit (data, ivs, act_delta, false); | |
6271 | act_delta = iv_ca_delta_join (act_delta, tmp_delta); | |
6272 | ||
6273 | if (compare_costs (acost, orig_cost) < 0) | |
6274 | { | |
6275 | *delta = act_delta; | |
6276 | return acost; | |
6277 | } | |
6278 | else | |
6279 | iv_ca_delta_free (&act_delta); | |
6280 | } | |
6281 | ||
6282 | return orig_cost; | |
6283 | } | |
6284 | ||
b1b02be2 | 6285 | /* Tries to extend the sets IVS in the best possible way in order |
16ad8025 SL |
6286 | to express the USE. If ORIGINALP is true, prefer candidates from |
6287 | the original set of IVs, otherwise favor important candidates not | |
6288 | based on any memory object. */ | |
8b11a64c ZD |
6289 | |
6290 | static bool | |
b1b02be2 | 6291 | try_add_cand_for (struct ivopts_data *data, struct iv_ca *ivs, |
16ad8025 | 6292 | struct iv_use *use, bool originalp) |
8b11a64c | 6293 | { |
6e8c65f6 | 6294 | comp_cost best_cost, act_cost; |
8b11a64c | 6295 | unsigned i; |
38b0dcb8 ZD |
6296 | bitmap_iterator bi; |
6297 | struct iv_cand *cand; | |
b1b02be2 ZD |
6298 | struct iv_ca_delta *best_delta = NULL, *act_delta; |
6299 | struct cost_pair *cp; | |
6300 | ||
f22ae1ec | 6301 | iv_ca_add_use (data, ivs, use); |
b1b02be2 | 6302 | best_cost = iv_ca_cost (ivs); |
b1b02be2 ZD |
6303 | cp = iv_ca_cand_for_use (ivs, use); |
6304 | if (cp) | |
6305 | { | |
6306 | best_delta = iv_ca_delta_add (use, NULL, cp, NULL); | |
6307 | iv_ca_set_no_cp (data, ivs, use); | |
6308 | } | |
8b11a64c | 6309 | |
16ad8025 SL |
6310 | /* If ORIGINALP is true, try to find the original IV for the use. Otherwise |
6311 | first try important candidates not based on any memory object. Only if | |
6e8c65f6 ZD |
6312 | this fails, try the specific ones. Rationale -- in loops with many |
6313 | variables the best choice often is to use just one generic biv. If we | |
6314 | added here many ivs specific to the uses, the optimization algorithm later | |
6315 | would be likely to get stuck in a local minimum, thus causing us to create | |
6316 | too many ivs. The approach from few ivs to more seems more likely to be | |
6317 | successful -- starting from few ivs, replacing an expensive use by a | |
6318 | specific iv should always be a win. */ | |
38b0dcb8 | 6319 | EXECUTE_IF_SET_IN_BITMAP (data->important_candidates, 0, i, bi) |
8b11a64c | 6320 | { |
38b0dcb8 ZD |
6321 | cand = iv_cand (data, i); |
6322 | ||
16ad8025 SL |
6323 | if (originalp && cand->pos !=IP_ORIGINAL) |
6324 | continue; | |
6325 | ||
6326 | if (!originalp && cand->iv->base_object != NULL_TREE) | |
6e8c65f6 ZD |
6327 | continue; |
6328 | ||
b1b02be2 | 6329 | if (iv_ca_cand_used_p (ivs, cand)) |
18081149 | 6330 | continue; |
8b11a64c | 6331 | |
b1b02be2 ZD |
6332 | cp = get_use_iv_cost (data, use, cand); |
6333 | if (!cp) | |
6334 | continue; | |
6335 | ||
6336 | iv_ca_set_cp (data, ivs, use, cp); | |
18081149 XDL |
6337 | act_cost = iv_ca_extend (data, ivs, cand, &act_delta, NULL, |
6338 | true); | |
b1b02be2 ZD |
6339 | iv_ca_set_no_cp (data, ivs, use); |
6340 | act_delta = iv_ca_delta_add (use, NULL, cp, act_delta); | |
8b11a64c | 6341 | |
6e8c65f6 | 6342 | if (compare_costs (act_cost, best_cost) < 0) |
8b11a64c ZD |
6343 | { |
6344 | best_cost = act_cost; | |
b1b02be2 ZD |
6345 | |
6346 | iv_ca_delta_free (&best_delta); | |
6347 | best_delta = act_delta; | |
8b11a64c | 6348 | } |
b1b02be2 ZD |
6349 | else |
6350 | iv_ca_delta_free (&act_delta); | |
8b11a64c ZD |
6351 | } |
6352 | ||
6e8c65f6 | 6353 | if (infinite_cost_p (best_cost)) |
38b0dcb8 ZD |
6354 | { |
6355 | for (i = 0; i < use->n_map_members; i++) | |
6356 | { | |
6357 | cp = use->cost_map + i; | |
b1b02be2 ZD |
6358 | cand = cp->cand; |
6359 | if (!cand) | |
38b0dcb8 ZD |
6360 | continue; |
6361 | ||
6362 | /* Already tried this. */ | |
16ad8025 SL |
6363 | if (cand->important) |
6364 | { | |
6365 | if (originalp && cand->pos == IP_ORIGINAL) | |
6366 | continue; | |
6367 | if (!originalp && cand->iv->base_object == NULL_TREE) | |
6368 | continue; | |
6369 | } | |
b8698a0f | 6370 | |
b1b02be2 | 6371 | if (iv_ca_cand_used_p (ivs, cand)) |
38b0dcb8 ZD |
6372 | continue; |
6373 | ||
b1b02be2 ZD |
6374 | act_delta = NULL; |
6375 | iv_ca_set_cp (data, ivs, use, cp); | |
18081149 | 6376 | act_cost = iv_ca_extend (data, ivs, cand, &act_delta, NULL, true); |
b1b02be2 ZD |
6377 | iv_ca_set_no_cp (data, ivs, use); |
6378 | act_delta = iv_ca_delta_add (use, iv_ca_cand_for_use (ivs, use), | |
6379 | cp, act_delta); | |
38b0dcb8 | 6380 | |
6e8c65f6 | 6381 | if (compare_costs (act_cost, best_cost) < 0) |
38b0dcb8 ZD |
6382 | { |
6383 | best_cost = act_cost; | |
b1b02be2 ZD |
6384 | |
6385 | if (best_delta) | |
6386 | iv_ca_delta_free (&best_delta); | |
6387 | best_delta = act_delta; | |
38b0dcb8 | 6388 | } |
b1b02be2 ZD |
6389 | else |
6390 | iv_ca_delta_free (&act_delta); | |
38b0dcb8 ZD |
6391 | } |
6392 | } | |
6393 | ||
b1b02be2 ZD |
6394 | iv_ca_delta_commit (data, ivs, best_delta, true); |
6395 | iv_ca_delta_free (&best_delta); | |
8b11a64c | 6396 | |
6e8c65f6 | 6397 | return !infinite_cost_p (best_cost); |
8b11a64c ZD |
6398 | } |
6399 | ||
b1b02be2 | 6400 | /* Finds an initial assignment of candidates to uses. */ |
8b11a64c | 6401 | |
b1b02be2 | 6402 | static struct iv_ca * |
16ad8025 | 6403 | get_initial_solution (struct ivopts_data *data, bool originalp) |
8b11a64c | 6404 | { |
b1b02be2 | 6405 | struct iv_ca *ivs = iv_ca_new (data); |
8b11a64c ZD |
6406 | unsigned i; |
6407 | ||
6408 | for (i = 0; i < n_iv_uses (data); i++) | |
16ad8025 | 6409 | if (!try_add_cand_for (data, ivs, iv_use (data, i), originalp)) |
b1b02be2 ZD |
6410 | { |
6411 | iv_ca_free (&ivs); | |
6412 | return NULL; | |
6413 | } | |
8b11a64c | 6414 | |
b1b02be2 | 6415 | return ivs; |
8b11a64c ZD |
6416 | } |
6417 | ||
6326a5f5 BC |
6418 | /* Tries to improve set of induction variables IVS. TRY_REPLACE_P |
6419 | points to a bool variable, this function tries to break local | |
6420 | optimal fixed-point by replacing candidates in IVS if it's true. */ | |
8b11a64c ZD |
6421 | |
6422 | static bool | |
6326a5f5 BC |
6423 | try_improve_iv_set (struct ivopts_data *data, |
6424 | struct iv_ca *ivs, bool *try_replace_p) | |
8b11a64c | 6425 | { |
6e8c65f6 ZD |
6426 | unsigned i, n_ivs; |
6427 | comp_cost acost, best_cost = iv_ca_cost (ivs); | |
36f5ada1 | 6428 | struct iv_ca_delta *best_delta = NULL, *act_delta, *tmp_delta; |
b1b02be2 | 6429 | struct iv_cand *cand; |
8b11a64c | 6430 | |
36f5ada1 | 6431 | /* Try extending the set of induction variables by one. */ |
8b11a64c ZD |
6432 | for (i = 0; i < n_iv_cands (data); i++) |
6433 | { | |
b1b02be2 | 6434 | cand = iv_cand (data, i); |
b8698a0f | 6435 | |
b1b02be2 | 6436 | if (iv_ca_cand_used_p (ivs, cand)) |
36f5ada1 ZD |
6437 | continue; |
6438 | ||
18081149 | 6439 | acost = iv_ca_extend (data, ivs, cand, &act_delta, &n_ivs, false); |
36f5ada1 ZD |
6440 | if (!act_delta) |
6441 | continue; | |
6442 | ||
6443 | /* If we successfully added the candidate and the set is small enough, | |
6444 | try optimizing it by removing other candidates. */ | |
6445 | if (n_ivs <= ALWAYS_PRUNE_CAND_SET_BOUND) | |
6446 | { | |
6447 | iv_ca_delta_commit (data, ivs, act_delta, true); | |
6448 | acost = iv_ca_prune (data, ivs, cand, &tmp_delta); | |
6449 | iv_ca_delta_commit (data, ivs, act_delta, false); | |
6450 | act_delta = iv_ca_delta_join (act_delta, tmp_delta); | |
6451 | } | |
8b11a64c | 6452 | |
6e8c65f6 | 6453 | if (compare_costs (acost, best_cost) < 0) |
8b11a64c | 6454 | { |
b1b02be2 | 6455 | best_cost = acost; |
36f5ada1 | 6456 | iv_ca_delta_free (&best_delta); |
b1b02be2 | 6457 | best_delta = act_delta; |
8b11a64c | 6458 | } |
8b11a64c | 6459 | else |
b1b02be2 | 6460 | iv_ca_delta_free (&act_delta); |
8b11a64c ZD |
6461 | } |
6462 | ||
b1b02be2 | 6463 | if (!best_delta) |
36f5ada1 ZD |
6464 | { |
6465 | /* Try removing the candidates from the set instead. */ | |
6466 | best_cost = iv_ca_prune (data, ivs, NULL, &best_delta); | |
6467 | ||
6326a5f5 BC |
6468 | if (!best_delta && *try_replace_p) |
6469 | { | |
6470 | *try_replace_p = false; | |
6471 | /* So far candidate selecting algorithm tends to choose fewer IVs | |
6472 | so that it can handle cases in which loops have many variables | |
6473 | but the best choice is often to use only one general biv. One | |
6474 | weakness is it can't handle opposite cases, in which different | |
6475 | candidates should be chosen with respect to each use. To solve | |
6476 | the problem, we replace candidates in a manner described by the | |
6477 | comments of iv_ca_replace, thus give general algorithm a chance | |
6478 | to break local optimal fixed-point in these cases. */ | |
6479 | best_cost = iv_ca_replace (data, ivs, &best_delta); | |
6480 | } | |
6481 | ||
36f5ada1 ZD |
6482 | if (!best_delta) |
6483 | return false; | |
6484 | } | |
8b11a64c | 6485 | |
b1b02be2 | 6486 | iv_ca_delta_commit (data, ivs, best_delta, true); |
6e8c65f6 | 6487 | gcc_assert (compare_costs (best_cost, iv_ca_cost (ivs)) == 0); |
b1b02be2 | 6488 | iv_ca_delta_free (&best_delta); |
8b11a64c ZD |
6489 | return true; |
6490 | } | |
6491 | ||
6492 | /* Attempts to find the optimal set of induction variables. We do simple | |
6493 | greedy heuristic -- we try to replace at most one candidate in the selected | |
6494 | solution and remove the unused ivs while this improves the cost. */ | |
6495 | ||
b1b02be2 | 6496 | static struct iv_ca * |
16ad8025 | 6497 | find_optimal_iv_set_1 (struct ivopts_data *data, bool originalp) |
8b11a64c | 6498 | { |
b1b02be2 | 6499 | struct iv_ca *set; |
6326a5f5 | 6500 | bool try_replace_p = true; |
8b11a64c | 6501 | |
b1b02be2 | 6502 | /* Get the initial solution. */ |
16ad8025 | 6503 | set = get_initial_solution (data, originalp); |
b1b02be2 | 6504 | if (!set) |
8b11a64c ZD |
6505 | { |
6506 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
6507 | fprintf (dump_file, "Unable to substitute for ivs, failed.\n"); | |
8b11a64c ZD |
6508 | return NULL; |
6509 | } | |
6510 | ||
6511 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
6512 | { | |
b1b02be2 ZD |
6513 | fprintf (dump_file, "Initial set of candidates:\n"); |
6514 | iv_ca_dump (data, dump_file, set); | |
8b11a64c ZD |
6515 | } |
6516 | ||
6326a5f5 | 6517 | while (try_improve_iv_set (data, set, &try_replace_p)) |
8b11a64c ZD |
6518 | { |
6519 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
6520 | { | |
b1b02be2 ZD |
6521 | fprintf (dump_file, "Improved to:\n"); |
6522 | iv_ca_dump (data, dump_file, set); | |
8b11a64c ZD |
6523 | } |
6524 | } | |
6525 | ||
16ad8025 SL |
6526 | return set; |
6527 | } | |
6528 | ||
6529 | static struct iv_ca * | |
6530 | find_optimal_iv_set (struct ivopts_data *data) | |
6531 | { | |
6532 | unsigned i; | |
6533 | struct iv_ca *set, *origset; | |
6534 | struct iv_use *use; | |
6535 | comp_cost cost, origcost; | |
6536 | ||
6537 | /* Determine the cost based on a strategy that starts with original IVs, | |
6538 | and try again using a strategy that prefers candidates not based | |
6539 | on any IVs. */ | |
6540 | origset = find_optimal_iv_set_1 (data, true); | |
6541 | set = find_optimal_iv_set_1 (data, false); | |
6542 | ||
6543 | if (!origset && !set) | |
6544 | return NULL; | |
6545 | ||
6546 | origcost = origset ? iv_ca_cost (origset) : infinite_cost; | |
6547 | cost = set ? iv_ca_cost (set) : infinite_cost; | |
6548 | ||
8b11a64c | 6549 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6e8c65f6 | 6550 | { |
16ad8025 SL |
6551 | fprintf (dump_file, "Original cost %d (complexity %d)\n\n", |
6552 | origcost.cost, origcost.complexity); | |
6553 | fprintf (dump_file, "Final cost %d (complexity %d)\n\n", | |
6554 | cost.cost, cost.complexity); | |
6555 | } | |
6556 | ||
6557 | /* Choose the one with the best cost. */ | |
6558 | if (compare_costs (origcost, cost) <= 0) | |
6559 | { | |
6560 | if (set) | |
6561 | iv_ca_free (&set); | |
6562 | set = origset; | |
6e8c65f6 | 6563 | } |
16ad8025 SL |
6564 | else if (origset) |
6565 | iv_ca_free (&origset); | |
8b11a64c ZD |
6566 | |
6567 | for (i = 0; i < n_iv_uses (data); i++) | |
6568 | { | |
6569 | use = iv_use (data, i); | |
b1b02be2 | 6570 | use->selected = iv_ca_cand_for_use (set, use)->cand; |
8b11a64c ZD |
6571 | } |
6572 | ||
8b11a64c ZD |
6573 | return set; |
6574 | } | |
6575 | ||
6576 | /* Creates a new induction variable corresponding to CAND. */ | |
6577 | ||
6578 | static void | |
6579 | create_new_iv (struct ivopts_data *data, struct iv_cand *cand) | |
6580 | { | |
726a989a | 6581 | gimple_stmt_iterator incr_pos; |
8b11a64c ZD |
6582 | tree base; |
6583 | bool after = false; | |
6584 | ||
6585 | if (!cand->iv) | |
6586 | return; | |
6587 | ||
6588 | switch (cand->pos) | |
6589 | { | |
6590 | case IP_NORMAL: | |
726a989a | 6591 | incr_pos = gsi_last_bb (ip_normal_pos (data->current_loop)); |
8b11a64c ZD |
6592 | break; |
6593 | ||
6594 | case IP_END: | |
726a989a | 6595 | incr_pos = gsi_last_bb (ip_end_pos (data->current_loop)); |
8b11a64c ZD |
6596 | after = true; |
6597 | break; | |
6598 | ||
2c08497a BS |
6599 | case IP_AFTER_USE: |
6600 | after = true; | |
6601 | /* fall through */ | |
6602 | case IP_BEFORE_USE: | |
6603 | incr_pos = gsi_for_stmt (cand->incremented_at); | |
6604 | break; | |
6605 | ||
8b11a64c ZD |
6606 | case IP_ORIGINAL: |
6607 | /* Mark that the iv is preserved. */ | |
6608 | name_info (data, cand->var_before)->preserve_biv = true; | |
6609 | name_info (data, cand->var_after)->preserve_biv = true; | |
6610 | ||
6611 | /* Rewrite the increment so that it uses var_before directly. */ | |
6612 | find_interesting_uses_op (data, cand->var_after)->selected = cand; | |
8b11a64c ZD |
6613 | return; |
6614 | } | |
b8698a0f | 6615 | |
8b11a64c | 6616 | gimple_add_tmp_var (cand->var_before); |
8b11a64c ZD |
6617 | |
6618 | base = unshare_expr (cand->iv->base); | |
6619 | ||
9be872b7 ZD |
6620 | create_iv (base, unshare_expr (cand->iv->step), |
6621 | cand->var_before, data->current_loop, | |
8b11a64c ZD |
6622 | &incr_pos, after, &cand->var_before, &cand->var_after); |
6623 | } | |
6624 | ||
6625 | /* Creates new induction variables described in SET. */ | |
6626 | ||
6627 | static void | |
b1b02be2 | 6628 | create_new_ivs (struct ivopts_data *data, struct iv_ca *set) |
8b11a64c ZD |
6629 | { |
6630 | unsigned i; | |
6631 | struct iv_cand *cand; | |
87c476a2 | 6632 | bitmap_iterator bi; |
8b11a64c | 6633 | |
b1b02be2 | 6634 | EXECUTE_IF_SET_IN_BITMAP (set->cands, 0, i, bi) |
8b11a64c ZD |
6635 | { |
6636 | cand = iv_cand (data, i); | |
6637 | create_new_iv (data, cand); | |
87c476a2 | 6638 | } |
8b11a64c | 6639 | |
18081149 XDL |
6640 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6641 | { | |
28002f1a RB |
6642 | fprintf (dump_file, "Selected IV set for loop %d", |
6643 | data->current_loop->num); | |
6644 | if (data->loop_loc != UNKNOWN_LOCATION) | |
6645 | fprintf (dump_file, " at %s:%d", LOCATION_FILE (data->loop_loc), | |
6646 | LOCATION_LINE (data->loop_loc)); | |
6647 | fprintf (dump_file, ", %lu IVs:\n", bitmap_count_bits (set->cands)); | |
18081149 XDL |
6648 | EXECUTE_IF_SET_IN_BITMAP (set->cands, 0, i, bi) |
6649 | { | |
6650 | cand = iv_cand (data, i); | |
6651 | dump_cand (dump_file, cand); | |
6652 | } | |
6653 | fprintf (dump_file, "\n"); | |
6654 | } | |
6655 | } | |
8b11a64c ZD |
6656 | |
6657 | /* Rewrites USE (definition of iv used in a nonlinear expression) | |
6658 | using candidate CAND. */ | |
6659 | ||
6660 | static void | |
6661 | rewrite_use_nonlinear_expr (struct ivopts_data *data, | |
6662 | struct iv_use *use, struct iv_cand *cand) | |
6663 | { | |
3520b745 | 6664 | tree comp; |
726a989a | 6665 | tree op, tgt; |
538dd0b7 | 6666 | gassign *ass; |
726a989a | 6667 | gimple_stmt_iterator bsi; |
3520b745 ZD |
6668 | |
6669 | /* An important special case -- if we are asked to express value of | |
6670 | the original iv by itself, just exit; there is no need to | |
6671 | introduce a new computation (that might also need casting the | |
6672 | variable to unsigned and back). */ | |
6673 | if (cand->pos == IP_ORIGINAL | |
7b9d4f70 | 6674 | && cand->incremented_at == use->stmt) |
3520b745 | 6675 | { |
d06a01bf | 6676 | enum tree_code stmt_code; |
7b9d4f70 | 6677 | |
726a989a RB |
6678 | gcc_assert (is_gimple_assign (use->stmt)); |
6679 | gcc_assert (gimple_assign_lhs (use->stmt) == cand->var_after); | |
7b9d4f70 | 6680 | |
7b9d4f70 ZD |
6681 | /* Check whether we may leave the computation unchanged. |
6682 | This is the case only if it does not rely on other | |
6683 | computations in the loop -- otherwise, the computation | |
6684 | we rely upon may be removed in remove_unused_ivs, | |
6685 | thus leading to ICE. */ | |
d06a01bf ZD |
6686 | stmt_code = gimple_assign_rhs_code (use->stmt); |
6687 | if (stmt_code == PLUS_EXPR | |
6688 | || stmt_code == MINUS_EXPR | |
6689 | || stmt_code == POINTER_PLUS_EXPR) | |
7b9d4f70 | 6690 | { |
726a989a RB |
6691 | if (gimple_assign_rhs1 (use->stmt) == cand->var_before) |
6692 | op = gimple_assign_rhs2 (use->stmt); | |
d06a01bf | 6693 | else if (gimple_assign_rhs2 (use->stmt) == cand->var_before) |
726a989a | 6694 | op = gimple_assign_rhs1 (use->stmt); |
7b9d4f70 ZD |
6695 | else |
6696 | op = NULL_TREE; | |
6697 | } | |
6698 | else | |
6699 | op = NULL_TREE; | |
3520b745 | 6700 | |
d06a01bf | 6701 | if (op && expr_invariant_in_loop_p (data->current_loop, op)) |
3520b745 | 6702 | return; |
73f30c63 | 6703 | } |
3520b745 | 6704 | |
d06a01bf ZD |
6705 | comp = get_computation (data->current_loop, use, cand); |
6706 | gcc_assert (comp != NULL_TREE); | |
6707 | ||
726a989a | 6708 | switch (gimple_code (use->stmt)) |
8b11a64c | 6709 | { |
726a989a | 6710 | case GIMPLE_PHI: |
8b11a64c ZD |
6711 | tgt = PHI_RESULT (use->stmt); |
6712 | ||
6713 | /* If we should keep the biv, do not replace it. */ | |
6714 | if (name_info (data, tgt)->preserve_biv) | |
6715 | return; | |
6716 | ||
726a989a | 6717 | bsi = gsi_after_labels (gimple_bb (use->stmt)); |
1e128c5f GB |
6718 | break; |
6719 | ||
726a989a RB |
6720 | case GIMPLE_ASSIGN: |
6721 | tgt = gimple_assign_lhs (use->stmt); | |
6722 | bsi = gsi_for_stmt (use->stmt); | |
1e128c5f GB |
6723 | break; |
6724 | ||
6725 | default: | |
6726 | gcc_unreachable (); | |
8b11a64c | 6727 | } |
8b11a64c | 6728 | |
17fc049f RG |
6729 | if (!valid_gimple_rhs_p (comp) |
6730 | || (gimple_code (use->stmt) != GIMPLE_PHI | |
6731 | /* We can't allow re-allocating the stmt as it might be pointed | |
6732 | to still. */ | |
6733 | && (get_gimple_rhs_num_ops (TREE_CODE (comp)) | |
6734 | >= gimple_num_ops (gsi_stmt (bsi))))) | |
bdf0f819 RG |
6735 | { |
6736 | comp = force_gimple_operand_gsi (&bsi, comp, true, NULL_TREE, | |
6737 | true, GSI_SAME_STMT); | |
6738 | if (POINTER_TYPE_P (TREE_TYPE (tgt))) | |
b5c878a5 RG |
6739 | { |
6740 | duplicate_ssa_name_ptr_info (comp, SSA_NAME_PTR_INFO (tgt)); | |
6741 | /* As this isn't a plain copy we have to reset alignment | |
6742 | information. */ | |
6743 | if (SSA_NAME_PTR_INFO (comp)) | |
644ffefd | 6744 | mark_ptr_info_alignment_unknown (SSA_NAME_PTR_INFO (comp)); |
b5c878a5 | 6745 | } |
bdf0f819 | 6746 | } |
8b11a64c | 6747 | |
726a989a | 6748 | if (gimple_code (use->stmt) == GIMPLE_PHI) |
8b11a64c | 6749 | { |
17fc049f | 6750 | ass = gimple_build_assign (tgt, comp); |
726a989a | 6751 | gsi_insert_before (&bsi, ass, GSI_SAME_STMT); |
ae0a4449 AO |
6752 | |
6753 | bsi = gsi_for_stmt (use->stmt); | |
6754 | remove_phi_node (&bsi, false); | |
8b11a64c ZD |
6755 | } |
6756 | else | |
726a989a | 6757 | { |
17fc049f | 6758 | gimple_assign_set_rhs_from_tree (&bsi, comp); |
726a989a RB |
6759 | use->stmt = gsi_stmt (bsi); |
6760 | } | |
8b11a64c ZD |
6761 | } |
6762 | ||
55791fcd XDL |
6763 | /* Performs a peephole optimization to reorder the iv update statement with |
6764 | a mem ref to enable instruction combining in later phases. The mem ref uses | |
6765 | the iv value before the update, so the reordering transformation requires | |
6766 | adjustment of the offset. CAND is the selected IV_CAND. | |
6767 | ||
6768 | Example: | |
6769 | ||
6770 | t = MEM_REF (base, iv1, 8, 16); // base, index, stride, offset | |
6771 | iv2 = iv1 + 1; | |
6772 | ||
6773 | if (t < val) (1) | |
6774 | goto L; | |
6775 | goto Head; | |
6776 | ||
6777 | ||
6778 | directly propagating t over to (1) will introduce overlapping live range | |
6779 | thus increase register pressure. This peephole transform it into: | |
6780 | ||
6781 | ||
6782 | iv2 = iv1 + 1; | |
6783 | t = MEM_REF (base, iv2, 8, 8); | |
6784 | if (t < val) | |
6785 | goto L; | |
6786 | goto Head; | |
6787 | */ | |
6788 | ||
6789 | static void | |
6790 | adjust_iv_update_pos (struct iv_cand *cand, struct iv_use *use) | |
6791 | { | |
6792 | tree var_after; | |
6793 | gimple iv_update, stmt; | |
6794 | basic_block bb; | |
6795 | gimple_stmt_iterator gsi, gsi_iv; | |
6796 | ||
6797 | if (cand->pos != IP_NORMAL) | |
6798 | return; | |
6799 | ||
6800 | var_after = cand->var_after; | |
6801 | iv_update = SSA_NAME_DEF_STMT (var_after); | |
6802 | ||
6803 | bb = gimple_bb (iv_update); | |
6804 | gsi = gsi_last_nondebug_bb (bb); | |
6805 | stmt = gsi_stmt (gsi); | |
6806 | ||
6807 | /* Only handle conditional statement for now. */ | |
6808 | if (gimple_code (stmt) != GIMPLE_COND) | |
6809 | return; | |
6810 | ||
6811 | gsi_prev_nondebug (&gsi); | |
6812 | stmt = gsi_stmt (gsi); | |
6813 | if (stmt != iv_update) | |
6814 | return; | |
6815 | ||
6816 | gsi_prev_nondebug (&gsi); | |
6817 | if (gsi_end_p (gsi)) | |
6818 | return; | |
6819 | ||
6820 | stmt = gsi_stmt (gsi); | |
6821 | if (gimple_code (stmt) != GIMPLE_ASSIGN) | |
6822 | return; | |
6823 | ||
6824 | if (stmt != use->stmt) | |
6825 | return; | |
6826 | ||
6827 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
6828 | return; | |
6829 | ||
6830 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
6831 | { | |
6832 | fprintf (dump_file, "Reordering \n"); | |
6833 | print_gimple_stmt (dump_file, iv_update, 0, 0); | |
6834 | print_gimple_stmt (dump_file, use->stmt, 0, 0); | |
6835 | fprintf (dump_file, "\n"); | |
6836 | } | |
6837 | ||
6838 | gsi = gsi_for_stmt (use->stmt); | |
6839 | gsi_iv = gsi_for_stmt (iv_update); | |
6840 | gsi_move_before (&gsi_iv, &gsi); | |
6841 | ||
6842 | cand->pos = IP_BEFORE_USE; | |
6843 | cand->incremented_at = use->stmt; | |
6844 | } | |
6845 | ||
8b11a64c ZD |
6846 | /* Rewrites USE (address that is an iv) using candidate CAND. */ |
6847 | ||
6848 | static void | |
a7e43c57 BC |
6849 | rewrite_use_address_1 (struct ivopts_data *data, |
6850 | struct iv_use *use, struct iv_cand *cand) | |
8b11a64c | 6851 | { |
73f30c63 | 6852 | aff_tree aff; |
726a989a | 6853 | gimple_stmt_iterator bsi = gsi_for_stmt (use->stmt); |
d7c0c068 | 6854 | tree base_hint = NULL_TREE; |
880a1451 | 6855 | tree ref, iv; |
73f30c63 | 6856 | bool ok; |
8b11a64c | 6857 | |
55791fcd | 6858 | adjust_iv_update_pos (cand, use); |
73f30c63 ZD |
6859 | ok = get_computation_aff (data->current_loop, use, cand, use->stmt, &aff); |
6860 | gcc_assert (ok); | |
ac182688 | 6861 | unshare_aff_combination (&aff); |
8b11a64c | 6862 | |
d7c0c068 UW |
6863 | /* To avoid undefined overflow problems, all IV candidates use unsigned |
6864 | integer types. The drawback is that this makes it impossible for | |
6865 | create_mem_ref to distinguish an IV that is based on a memory object | |
6866 | from one that represents simply an offset. | |
6867 | ||
6868 | To work around this problem, we pass a hint to create_mem_ref that | |
6869 | indicates which variable (if any) in aff is an IV based on a memory | |
6870 | object. Note that we only consider the candidate. If this is not | |
6871 | based on an object, the base of the reference is in some subexpression | |
6872 | of the use -- but these will use pointer types, so they are recognized | |
6873 | by the create_mem_ref heuristics anyway. */ | |
6874 | if (cand->iv->base_object) | |
6875 | base_hint = var_at_stmt (data->current_loop, cand, use->stmt); | |
6876 | ||
880a1451 XDL |
6877 | iv = var_at_stmt (data->current_loop, cand, use->stmt); |
6878 | ref = create_mem_ref (&bsi, TREE_TYPE (*use->op_p), &aff, | |
a41e5e86 | 6879 | reference_alias_ptr_type (*use->op_p), |
880a1451 | 6880 | iv, base_hint, data->speed); |
ac182688 ZD |
6881 | copy_ref_info (ref, *use->op_p); |
6882 | *use->op_p = ref; | |
8b11a64c ZD |
6883 | } |
6884 | ||
a7e43c57 BC |
6885 | /* Rewrites USE (address that is an iv) using candidate CAND. If it's the |
6886 | first use of a group, rewrites sub uses in the group too. */ | |
6887 | ||
6888 | static void | |
6889 | rewrite_use_address (struct ivopts_data *data, | |
6890 | struct iv_use *use, struct iv_cand *cand) | |
6891 | { | |
6892 | struct iv_use *next; | |
6893 | ||
6894 | gcc_assert (use->sub_id == 0); | |
6895 | rewrite_use_address_1 (data, use, cand); | |
6896 | update_stmt (use->stmt); | |
6897 | ||
6898 | for (next = use->next; next != NULL; next = next->next) | |
6899 | { | |
6900 | rewrite_use_address_1 (data, next, cand); | |
6901 | update_stmt (next->stmt); | |
6902 | } | |
6903 | ||
6904 | return; | |
6905 | } | |
6906 | ||
8b11a64c ZD |
6907 | /* Rewrites USE (the condition such that one of the arguments is an iv) using |
6908 | candidate CAND. */ | |
6909 | ||
6910 | static void | |
6911 | rewrite_use_compare (struct ivopts_data *data, | |
6912 | struct iv_use *use, struct iv_cand *cand) | |
6913 | { | |
b697aed4 | 6914 | tree comp, *var_p, op, bound; |
726a989a | 6915 | gimple_stmt_iterator bsi = gsi_for_stmt (use->stmt); |
8b11a64c | 6916 | enum tree_code compare; |
f5f12961 | 6917 | struct cost_pair *cp = get_use_iv_cost (data, use, cand); |
b697aed4 ZD |
6918 | bool ok; |
6919 | ||
f5f12961 ZD |
6920 | bound = cp->value; |
6921 | if (bound) | |
8b11a64c | 6922 | { |
9e7376e5 ZD |
6923 | tree var = var_at_stmt (data->current_loop, cand, use->stmt); |
6924 | tree var_type = TREE_TYPE (var); | |
dc5b3407 | 6925 | gimple_seq stmts; |
9e7376e5 | 6926 | |
18081149 XDL |
6927 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6928 | { | |
6929 | fprintf (dump_file, "Replacing exit test: "); | |
6930 | print_gimple_stmt (dump_file, use->stmt, 0, TDF_SLIM); | |
6931 | } | |
d8af4ba3 | 6932 | compare = cp->comp; |
b697aed4 | 6933 | bound = unshare_expr (fold_convert (var_type, bound)); |
dc5b3407 ZD |
6934 | op = force_gimple_operand (bound, &stmts, true, NULL_TREE); |
6935 | if (stmts) | |
6936 | gsi_insert_seq_on_edge_immediate ( | |
6937 | loop_preheader_edge (data->current_loop), | |
6938 | stmts); | |
8b11a64c | 6939 | |
538dd0b7 DM |
6940 | gcond *cond_stmt = as_a <gcond *> (use->stmt); |
6941 | gimple_cond_set_lhs (cond_stmt, var); | |
6942 | gimple_cond_set_code (cond_stmt, compare); | |
6943 | gimple_cond_set_rhs (cond_stmt, op); | |
8b11a64c ZD |
6944 | return; |
6945 | } | |
6946 | ||
6947 | /* The induction variable elimination failed; just express the original | |
6948 | giv. */ | |
ac182688 | 6949 | comp = get_computation (data->current_loop, use, cand); |
73f30c63 | 6950 | gcc_assert (comp != NULL_TREE); |
8b11a64c | 6951 | |
726a989a | 6952 | ok = extract_cond_operands (data, use->stmt, &var_p, NULL, NULL, NULL); |
b697aed4 | 6953 | gcc_assert (ok); |
8b11a64c | 6954 | |
726a989a RB |
6955 | *var_p = force_gimple_operand_gsi (&bsi, comp, true, SSA_NAME_VAR (*var_p), |
6956 | true, GSI_SAME_STMT); | |
8b11a64c ZD |
6957 | } |
6958 | ||
8b11a64c ZD |
6959 | /* Rewrites USE using candidate CAND. */ |
6960 | ||
6961 | static void | |
cfaab3a9 | 6962 | rewrite_use (struct ivopts_data *data, struct iv_use *use, struct iv_cand *cand) |
8b11a64c ZD |
6963 | { |
6964 | switch (use->type) | |
6965 | { | |
6966 | case USE_NONLINEAR_EXPR: | |
6967 | rewrite_use_nonlinear_expr (data, use, cand); | |
6968 | break; | |
6969 | ||
8b11a64c ZD |
6970 | case USE_ADDRESS: |
6971 | rewrite_use_address (data, use, cand); | |
6972 | break; | |
6973 | ||
6974 | case USE_COMPARE: | |
6975 | rewrite_use_compare (data, use, cand); | |
6976 | break; | |
6977 | ||
6978 | default: | |
1e128c5f | 6979 | gcc_unreachable (); |
8b11a64c | 6980 | } |
b8698a0f | 6981 | |
cff4e50d | 6982 | update_stmt (use->stmt); |
8b11a64c ZD |
6983 | } |
6984 | ||
6985 | /* Rewrite the uses using the selected induction variables. */ | |
6986 | ||
6987 | static void | |
6988 | rewrite_uses (struct ivopts_data *data) | |
6989 | { | |
6990 | unsigned i; | |
6991 | struct iv_cand *cand; | |
6992 | struct iv_use *use; | |
6993 | ||
6994 | for (i = 0; i < n_iv_uses (data); i++) | |
6995 | { | |
6996 | use = iv_use (data, i); | |
6997 | cand = use->selected; | |
1e128c5f | 6998 | gcc_assert (cand); |
8b11a64c ZD |
6999 | |
7000 | rewrite_use (data, use, cand); | |
7001 | } | |
7002 | } | |
7003 | ||
7004 | /* Removes the ivs that are not used after rewriting. */ | |
7005 | ||
7006 | static void | |
7007 | remove_unused_ivs (struct ivopts_data *data) | |
7008 | { | |
7009 | unsigned j; | |
87c476a2 | 7010 | bitmap_iterator bi; |
ae0a4449 | 7011 | bitmap toremove = BITMAP_ALLOC (NULL); |
8b11a64c | 7012 | |
ae0a4449 AO |
7013 | /* Figure out an order in which to release SSA DEFs so that we don't |
7014 | release something that we'd have to propagate into a debug stmt | |
7015 | afterwards. */ | |
87c476a2 | 7016 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, j, bi) |
8b11a64c ZD |
7017 | { |
7018 | struct version_info *info; | |
7019 | ||
7020 | info = ver_info (data, j); | |
7021 | if (info->iv | |
6e42ce54 | 7022 | && !integer_zerop (info->iv->step) |
8b11a64c ZD |
7023 | && !info->inv_id |
7024 | && !info->iv->have_use_for | |
7025 | && !info->preserve_biv) | |
e1066560 AO |
7026 | { |
7027 | bitmap_set_bit (toremove, SSA_NAME_VERSION (info->iv->ssa_name)); | |
7028 | ||
7029 | tree def = info->iv->ssa_name; | |
7030 | ||
7031 | if (MAY_HAVE_DEBUG_STMTS && SSA_NAME_DEF_STMT (def)) | |
7032 | { | |
7033 | imm_use_iterator imm_iter; | |
7034 | use_operand_p use_p; | |
7035 | gimple stmt; | |
7036 | int count = 0; | |
7037 | ||
7038 | FOR_EACH_IMM_USE_STMT (stmt, imm_iter, def) | |
7039 | { | |
7040 | if (!gimple_debug_bind_p (stmt)) | |
7041 | continue; | |
7042 | ||
7043 | /* We just want to determine whether to do nothing | |
7044 | (count == 0), to substitute the computed | |
7045 | expression into a single use of the SSA DEF by | |
7046 | itself (count == 1), or to use a debug temp | |
7047 | because the SSA DEF is used multiple times or as | |
7048 | part of a larger expression (count > 1). */ | |
7049 | count++; | |
7050 | if (gimple_debug_bind_get_value (stmt) != def) | |
7051 | count++; | |
7052 | ||
7053 | if (count > 1) | |
7054 | BREAK_FROM_IMM_USE_STMT (imm_iter); | |
7055 | } | |
7056 | ||
7057 | if (!count) | |
7058 | continue; | |
7059 | ||
7060 | struct iv_use dummy_use; | |
7061 | struct iv_cand *best_cand = NULL, *cand; | |
7062 | unsigned i, best_pref = 0, cand_pref; | |
7063 | ||
7064 | memset (&dummy_use, 0, sizeof (dummy_use)); | |
7065 | dummy_use.iv = info->iv; | |
7066 | for (i = 0; i < n_iv_uses (data) && i < 64; i++) | |
7067 | { | |
7068 | cand = iv_use (data, i)->selected; | |
7069 | if (cand == best_cand) | |
7070 | continue; | |
7071 | cand_pref = operand_equal_p (cand->iv->step, | |
7072 | info->iv->step, 0) | |
7073 | ? 4 : 0; | |
7074 | cand_pref | |
7075 | += TYPE_MODE (TREE_TYPE (cand->iv->base)) | |
7076 | == TYPE_MODE (TREE_TYPE (info->iv->base)) | |
7077 | ? 2 : 0; | |
7078 | cand_pref | |
7079 | += TREE_CODE (cand->iv->base) == INTEGER_CST | |
7080 | ? 1 : 0; | |
7081 | if (best_cand == NULL || best_pref < cand_pref) | |
7082 | { | |
7083 | best_cand = cand; | |
7084 | best_pref = cand_pref; | |
7085 | } | |
7086 | } | |
7087 | ||
7088 | if (!best_cand) | |
7089 | continue; | |
7090 | ||
7091 | tree comp = get_computation_at (data->current_loop, | |
7092 | &dummy_use, best_cand, | |
7093 | SSA_NAME_DEF_STMT (def)); | |
7094 | if (!comp) | |
7095 | continue; | |
7096 | ||
7097 | if (count > 1) | |
7098 | { | |
7099 | tree vexpr = make_node (DEBUG_EXPR_DECL); | |
7100 | DECL_ARTIFICIAL (vexpr) = 1; | |
7101 | TREE_TYPE (vexpr) = TREE_TYPE (comp); | |
7102 | if (SSA_NAME_VAR (def)) | |
7103 | DECL_MODE (vexpr) = DECL_MODE (SSA_NAME_VAR (def)); | |
7104 | else | |
7105 | DECL_MODE (vexpr) = TYPE_MODE (TREE_TYPE (vexpr)); | |
538dd0b7 DM |
7106 | gdebug *def_temp |
7107 | = gimple_build_debug_bind (vexpr, comp, NULL); | |
e1066560 AO |
7108 | gimple_stmt_iterator gsi; |
7109 | ||
7110 | if (gimple_code (SSA_NAME_DEF_STMT (def)) == GIMPLE_PHI) | |
7111 | gsi = gsi_after_labels (gimple_bb | |
7112 | (SSA_NAME_DEF_STMT (def))); | |
7113 | else | |
7114 | gsi = gsi_for_stmt (SSA_NAME_DEF_STMT (def)); | |
7115 | ||
7116 | gsi_insert_before (&gsi, def_temp, GSI_SAME_STMT); | |
7117 | comp = vexpr; | |
7118 | } | |
7119 | ||
7120 | FOR_EACH_IMM_USE_STMT (stmt, imm_iter, def) | |
7121 | { | |
7122 | if (!gimple_debug_bind_p (stmt)) | |
7123 | continue; | |
7124 | ||
7125 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) | |
7126 | SET_USE (use_p, comp); | |
7127 | ||
7128 | update_stmt (stmt); | |
7129 | } | |
7130 | } | |
7131 | } | |
87c476a2 | 7132 | } |
ae0a4449 AO |
7133 | |
7134 | release_defs_bitset (toremove); | |
7135 | ||
7136 | BITMAP_FREE (toremove); | |
8b11a64c ZD |
7137 | } |
7138 | ||
e2102efc | 7139 | /* Frees memory occupied by struct tree_niter_desc in *VALUE. Callback |
b787e7a2 | 7140 | for hash_map::traverse. */ |
e2102efc | 7141 | |
b787e7a2 TS |
7142 | bool |
7143 | free_tree_niter_desc (edge const &, tree_niter_desc *const &value, void *) | |
e2102efc | 7144 | { |
b787e7a2 | 7145 | free (value); |
e2102efc XDL |
7146 | return true; |
7147 | } | |
7148 | ||
8b11a64c ZD |
7149 | /* Frees data allocated by the optimization of a single loop. */ |
7150 | ||
7151 | static void | |
7152 | free_loop_data (struct ivopts_data *data) | |
7153 | { | |
7154 | unsigned i, j; | |
87c476a2 | 7155 | bitmap_iterator bi; |
69ebd99d | 7156 | tree obj; |
8b11a64c | 7157 | |
15814ba0 PB |
7158 | if (data->niters) |
7159 | { | |
b787e7a2 TS |
7160 | data->niters->traverse<void *, free_tree_niter_desc> (NULL); |
7161 | delete data->niters; | |
15814ba0 PB |
7162 | data->niters = NULL; |
7163 | } | |
ca4c3169 | 7164 | |
87c476a2 | 7165 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, i, bi) |
8b11a64c ZD |
7166 | { |
7167 | struct version_info *info; | |
7168 | ||
7169 | info = ver_info (data, i); | |
8b11a64c ZD |
7170 | info->iv = NULL; |
7171 | info->has_nonlin_use = false; | |
7172 | info->preserve_biv = false; | |
7173 | info->inv_id = 0; | |
87c476a2 | 7174 | } |
8b11a64c | 7175 | bitmap_clear (data->relevant); |
b1b02be2 | 7176 | bitmap_clear (data->important_candidates); |
8b11a64c ZD |
7177 | |
7178 | for (i = 0; i < n_iv_uses (data); i++) | |
7179 | { | |
7180 | struct iv_use *use = iv_use (data, i); | |
a7e43c57 BC |
7181 | struct iv_use *pre = use, *sub = use->next; |
7182 | ||
7183 | while (sub) | |
7184 | { | |
7185 | gcc_assert (sub->related_cands == NULL); | |
7186 | gcc_assert (sub->n_map_members == 0 && sub->cost_map == NULL); | |
7187 | ||
a7e43c57 BC |
7188 | pre = sub; |
7189 | sub = sub->next; | |
7190 | free (pre); | |
7191 | } | |
8b11a64c | 7192 | |
8bdbfff5 | 7193 | BITMAP_FREE (use->related_cands); |
8b11a64c ZD |
7194 | for (j = 0; j < use->n_map_members; j++) |
7195 | if (use->cost_map[j].depends_on) | |
8bdbfff5 | 7196 | BITMAP_FREE (use->cost_map[j].depends_on); |
8b11a64c ZD |
7197 | free (use->cost_map); |
7198 | free (use); | |
7199 | } | |
9771b263 | 7200 | data->iv_uses.truncate (0); |
8b11a64c ZD |
7201 | |
7202 | for (i = 0; i < n_iv_cands (data); i++) | |
7203 | { | |
7204 | struct iv_cand *cand = iv_cand (data, i); | |
7205 | ||
9be872b7 ZD |
7206 | if (cand->depends_on) |
7207 | BITMAP_FREE (cand->depends_on); | |
8b11a64c ZD |
7208 | free (cand); |
7209 | } | |
9771b263 | 7210 | data->iv_candidates.truncate (0); |
8b11a64c ZD |
7211 | |
7212 | if (data->version_info_size < num_ssa_names) | |
7213 | { | |
7214 | data->version_info_size = 2 * num_ssa_names; | |
7215 | free (data->version_info); | |
5ed6ace5 | 7216 | data->version_info = XCNEWVEC (struct version_info, data->version_info_size); |
8b11a64c ZD |
7217 | } |
7218 | ||
7219 | data->max_inv_id = 0; | |
7220 | ||
9771b263 | 7221 | FOR_EACH_VEC_ELT (decl_rtl_to_reset, i, obj) |
69ebd99d | 7222 | SET_DECL_RTL (obj, NULL_RTX); |
8b11a64c | 7223 | |
9771b263 | 7224 | decl_rtl_to_reset.truncate (0); |
18081149 | 7225 | |
c203e8a7 | 7226 | data->inv_expr_tab->empty (); |
18081149 | 7227 | data->inv_expr_id = 0; |
8b11a64c ZD |
7228 | } |
7229 | ||
7230 | /* Finalizes data structures used by the iv optimization pass. LOOPS is the | |
7231 | loop tree. */ | |
7232 | ||
7233 | static void | |
9a2ef6b8 | 7234 | tree_ssa_iv_optimize_finalize (struct ivopts_data *data) |
8b11a64c | 7235 | { |
8b11a64c ZD |
7236 | free_loop_data (data); |
7237 | free (data->version_info); | |
8bdbfff5 NS |
7238 | BITMAP_FREE (data->relevant); |
7239 | BITMAP_FREE (data->important_candidates); | |
8b11a64c | 7240 | |
9771b263 DN |
7241 | decl_rtl_to_reset.release (); |
7242 | data->iv_uses.release (); | |
7243 | data->iv_candidates.release (); | |
c203e8a7 TS |
7244 | delete data->inv_expr_tab; |
7245 | data->inv_expr_tab = NULL; | |
3230c614 | 7246 | free_affine_expand_cache (&data->name_expansion_cache); |
6f929985 | 7247 | obstack_free (&data->iv_obstack, NULL); |
8b11a64c ZD |
7248 | } |
7249 | ||
bec922f0 SL |
7250 | /* Returns true if the loop body BODY includes any function calls. */ |
7251 | ||
7252 | static bool | |
7253 | loop_body_includes_call (basic_block *body, unsigned num_nodes) | |
7254 | { | |
7255 | gimple_stmt_iterator gsi; | |
7256 | unsigned i; | |
7257 | ||
7258 | for (i = 0; i < num_nodes; i++) | |
7259 | for (gsi = gsi_start_bb (body[i]); !gsi_end_p (gsi); gsi_next (&gsi)) | |
7260 | { | |
7261 | gimple stmt = gsi_stmt (gsi); | |
7262 | if (is_gimple_call (stmt) | |
7263 | && !is_inexpensive_builtin (gimple_call_fndecl (stmt))) | |
7264 | return true; | |
7265 | } | |
7266 | return false; | |
7267 | } | |
7268 | ||
8b11a64c ZD |
7269 | /* Optimizes the LOOP. Returns true if anything changed. */ |
7270 | ||
7271 | static bool | |
7272 | tree_ssa_iv_optimize_loop (struct ivopts_data *data, struct loop *loop) | |
7273 | { | |
7274 | bool changed = false; | |
b1b02be2 | 7275 | struct iv_ca *iv_ca; |
d8af4ba3 | 7276 | edge exit = single_dom_exit (loop); |
2c08497a | 7277 | basic_block *body; |
8b11a64c | 7278 | |
15814ba0 | 7279 | gcc_assert (!data->niters); |
8b11a64c | 7280 | data->current_loop = loop; |
28002f1a | 7281 | data->loop_loc = find_loop_location (loop); |
f40751dd | 7282 | data->speed = optimize_loop_for_speed_p (loop); |
8b11a64c ZD |
7283 | |
7284 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
7285 | { | |
28002f1a RB |
7286 | fprintf (dump_file, "Processing loop %d", loop->num); |
7287 | if (data->loop_loc != UNKNOWN_LOCATION) | |
7288 | fprintf (dump_file, " at %s:%d", LOCATION_FILE (data->loop_loc), | |
7289 | LOCATION_LINE (data->loop_loc)); | |
7290 | fprintf (dump_file, "\n"); | |
b8698a0f | 7291 | |
8b11a64c ZD |
7292 | if (exit) |
7293 | { | |
7294 | fprintf (dump_file, " single exit %d -> %d, exit condition ", | |
7295 | exit->src->index, exit->dest->index); | |
726a989a | 7296 | print_gimple_stmt (dump_file, last_stmt (exit->src), 0, TDF_SLIM); |
8b11a64c ZD |
7297 | fprintf (dump_file, "\n"); |
7298 | } | |
7299 | ||
7300 | fprintf (dump_file, "\n"); | |
7301 | } | |
7302 | ||
2c08497a | 7303 | body = get_loop_body (loop); |
bec922f0 | 7304 | data->body_includes_call = loop_body_includes_call (body, loop->num_nodes); |
2c08497a BS |
7305 | renumber_gimple_stmt_uids_in_blocks (body, loop->num_nodes); |
7306 | free (body); | |
7307 | ||
d8af4ba3 ZD |
7308 | data->loop_single_exit_p = exit != NULL && loop_only_exit_p (loop, exit); |
7309 | ||
8b11a64c ZD |
7310 | /* For each ssa name determines whether it behaves as an induction variable |
7311 | in some loop. */ | |
7312 | if (!find_induction_variables (data)) | |
7313 | goto finish; | |
7314 | ||
7315 | /* Finds interesting uses (item 1). */ | |
7316 | find_interesting_uses (data); | |
a7e43c57 | 7317 | group_address_uses (data); |
8b11a64c ZD |
7318 | if (n_iv_uses (data) > MAX_CONSIDERED_USES) |
7319 | goto finish; | |
7320 | ||
7321 | /* Finds candidates for the induction variables (item 2). */ | |
7322 | find_iv_candidates (data); | |
7323 | ||
7324 | /* Calculates the costs (item 3, part 1). */ | |
8b11a64c | 7325 | determine_iv_costs (data); |
2c08497a | 7326 | determine_use_iv_costs (data); |
8b11a64c ZD |
7327 | determine_set_costs (data); |
7328 | ||
7329 | /* Find the optimal set of induction variables (item 3, part 2). */ | |
b1b02be2 ZD |
7330 | iv_ca = find_optimal_iv_set (data); |
7331 | if (!iv_ca) | |
8b11a64c ZD |
7332 | goto finish; |
7333 | changed = true; | |
7334 | ||
7335 | /* Create the new induction variables (item 4, part 1). */ | |
b1b02be2 ZD |
7336 | create_new_ivs (data, iv_ca); |
7337 | iv_ca_free (&iv_ca); | |
b8698a0f | 7338 | |
8b11a64c ZD |
7339 | /* Rewrite the uses (item 4, part 2). */ |
7340 | rewrite_uses (data); | |
7341 | ||
7342 | /* Remove the ivs that are unused after rewriting. */ | |
7343 | remove_unused_ivs (data); | |
7344 | ||
8b11a64c ZD |
7345 | /* We have changed the structure of induction variables; it might happen |
7346 | that definitions in the scev database refer to some of them that were | |
7347 | eliminated. */ | |
7348 | scev_reset (); | |
7349 | ||
7350 | finish: | |
7351 | free_loop_data (data); | |
7352 | ||
7353 | return changed; | |
7354 | } | |
7355 | ||
d73be268 | 7356 | /* Main entry point. Optimizes induction variables in loops. */ |
8b11a64c ZD |
7357 | |
7358 | void | |
d73be268 | 7359 | tree_ssa_iv_optimize (void) |
8b11a64c ZD |
7360 | { |
7361 | struct loop *loop; | |
7362 | struct ivopts_data data; | |
7363 | ||
9a2ef6b8 | 7364 | tree_ssa_iv_optimize_init (&data); |
8b11a64c ZD |
7365 | |
7366 | /* Optimize the loops starting with the innermost ones. */ | |
f0bd40b1 | 7367 | FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) |
8b11a64c | 7368 | { |
8679c649 JH |
7369 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7370 | flow_loop_dump (loop, dump_file, NULL, 1); | |
e9472263 ZD |
7371 | |
7372 | tree_ssa_iv_optimize_loop (&data, loop); | |
8b11a64c ZD |
7373 | } |
7374 | ||
9a2ef6b8 | 7375 | tree_ssa_iv_optimize_finalize (&data); |
8b11a64c | 7376 | } |