]>
Commit | Line | Data |
---|---|---|
4502f5d0 | 1 | /* Straight-line strength reduction. |
3aea1f79 | 2 | Copyright (C) 2012-2014 Free Software Foundation, Inc. |
4502f5d0 | 3 | Contributed by Bill Schmidt, IBM <wschmidt@linux.ibm.com> |
4 | ||
5 | This file is part of GCC. | |
6 | ||
7 | GCC is free software; you can redistribute it and/or modify it under | |
8 | the terms of the GNU General Public License as published by the Free | |
9 | Software Foundation; either version 3, or (at your option) any later | |
10 | version. | |
11 | ||
12 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
13 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 | for more details. | |
16 | ||
17 | You should have received a copy of the GNU General Public License | |
18 | along with GCC; see the file COPYING3. If not see | |
19 | <http://www.gnu.org/licenses/>. */ | |
20 | ||
21 | /* There are many algorithms for performing strength reduction on | |
22 | loops. This is not one of them. IVOPTS handles strength reduction | |
23 | of induction variables just fine. This pass is intended to pick | |
24 | up the crumbs it leaves behind, by considering opportunities for | |
25 | strength reduction along dominator paths. | |
26 | ||
1168caca | 27 | Strength reduction addresses explicit multiplies, and certain |
28 | multiplies implicit in addressing expressions. It would also be | |
29 | possible to apply strength reduction to divisions and modulos, | |
30 | but such opportunities are relatively uncommon. | |
4502f5d0 | 31 | |
32 | Strength reduction is also currently restricted to integer operations. | |
33 | If desired, it could be extended to floating-point operations under | |
34 | control of something like -funsafe-math-optimizations. */ | |
35 | ||
36 | #include "config.h" | |
37 | #include "system.h" | |
38 | #include "coretypes.h" | |
39 | #include "tree.h" | |
06ecf488 | 40 | #include "hash-map.h" |
bc61cadb | 41 | #include "hash-table.h" |
94ea8568 | 42 | #include "predict.h" |
43 | #include "vec.h" | |
44 | #include "hashtab.h" | |
45 | #include "hash-set.h" | |
46 | #include "machmode.h" | |
47 | #include "tm.h" | |
48 | #include "hard-reg-set.h" | |
49 | #include "input.h" | |
50 | #include "function.h" | |
51 | #include "dominance.h" | |
52 | #include "cfg.h" | |
bc61cadb | 53 | #include "basic-block.h" |
54 | #include "tree-ssa-alias.h" | |
55 | #include "internal-fn.h" | |
56 | #include "gimple-expr.h" | |
57 | #include "is-a.h" | |
4502f5d0 | 58 | #include "gimple.h" |
dcf1a1ec | 59 | #include "gimple-iterator.h" |
e795d6e1 | 60 | #include "gimplify-me.h" |
9ed99284 | 61 | #include "stor-layout.h" |
62 | #include "expr.h" | |
4502f5d0 | 63 | #include "tree-pass.h" |
4502f5d0 | 64 | #include "cfgloop.h" |
4502f5d0 | 65 | #include "gimple-pretty-print.h" |
073c1fd5 | 66 | #include "gimple-ssa.h" |
67 | #include "tree-cfg.h" | |
68 | #include "tree-phinodes.h" | |
69 | #include "ssa-iterators.h" | |
9ed99284 | 70 | #include "stringpool.h" |
073c1fd5 | 71 | #include "tree-ssanames.h" |
4502f5d0 | 72 | #include "domwalk.h" |
72655676 | 73 | #include "expmed.h" |
df4c32b3 | 74 | #include "params.h" |
424a4a92 | 75 | #include "tree-ssa-address.h" |
0ca43779 | 76 | #include "tree-affine.h" |
e913b5cd | 77 | #include "wide-int-print.h" |
f7715905 | 78 | #include "builtins.h" |
4502f5d0 | 79 | \f |
80 | /* Information about a strength reduction candidate. Each statement | |
81 | in the candidate table represents an expression of one of the | |
82 | following forms (the special case of CAND_REF will be described | |
83 | later): | |
84 | ||
85 | (CAND_MULT) S1: X = (B + i) * S | |
86 | (CAND_ADD) S1: X = B + (i * S) | |
87 | ||
88 | Here X and B are SSA names, i is an integer constant, and S is | |
89 | either an SSA name or a constant. We call B the "base," i the | |
90 | "index", and S the "stride." | |
91 | ||
92 | Any statement S0 that dominates S1 and is of the form: | |
93 | ||
94 | (CAND_MULT) S0: Y = (B + i') * S | |
95 | (CAND_ADD) S0: Y = B + (i' * S) | |
96 | ||
97 | is called a "basis" for S1. In both cases, S1 may be replaced by | |
98 | ||
99 | S1': X = Y + (i - i') * S, | |
100 | ||
101 | where (i - i') * S is folded to the extent possible. | |
102 | ||
103 | All gimple statements are visited in dominator order, and each | |
104 | statement that may contribute to one of the forms of S1 above is | |
105 | given at least one entry in the candidate table. Such statements | |
106 | include addition, pointer addition, subtraction, multiplication, | |
107 | negation, copies, and nontrivial type casts. If a statement may | |
108 | represent more than one expression of the forms of S1 above, | |
109 | multiple "interpretations" are stored in the table and chained | |
110 | together. Examples: | |
111 | ||
112 | * An add of two SSA names may treat either operand as the base. | |
113 | * A multiply of two SSA names, likewise. | |
114 | * A copy or cast may be thought of as either a CAND_MULT with | |
115 | i = 0 and S = 1, or as a CAND_ADD with i = 0 or S = 0. | |
116 | ||
117 | Candidate records are allocated from an obstack. They are addressed | |
118 | both from a hash table keyed on S1, and from a vector of candidate | |
119 | pointers arranged in predominator order. | |
120 | ||
121 | Opportunity note | |
122 | ---------------- | |
123 | Currently we don't recognize: | |
124 | ||
125 | S0: Y = (S * i') - B | |
126 | S1: X = (S * i) - B | |
127 | ||
128 | as a strength reduction opportunity, even though this S1 would | |
129 | also be replaceable by the S1' above. This can be added if it | |
070bf980 | 130 | comes up in practice. |
131 | ||
132 | Strength reduction in addressing | |
133 | -------------------------------- | |
134 | There is another kind of candidate known as CAND_REF. A CAND_REF | |
135 | describes a statement containing a memory reference having | |
136 | complex addressing that might benefit from strength reduction. | |
137 | Specifically, we are interested in references for which | |
138 | get_inner_reference returns a base address, offset, and bitpos as | |
139 | follows: | |
140 | ||
141 | base: MEM_REF (T1, C1) | |
142 | offset: MULT_EXPR (PLUS_EXPR (T2, C2), C3) | |
143 | bitpos: C4 * BITS_PER_UNIT | |
144 | ||
145 | Here T1 and T2 are arbitrary trees, and C1, C2, C3, C4 are | |
146 | arbitrary integer constants. Note that C2 may be zero, in which | |
147 | case the offset will be MULT_EXPR (T2, C3). | |
148 | ||
149 | When this pattern is recognized, the original memory reference | |
150 | can be replaced with: | |
151 | ||
152 | MEM_REF (POINTER_PLUS_EXPR (T1, MULT_EXPR (T2, C3)), | |
153 | C1 + (C2 * C3) + C4) | |
154 | ||
155 | which distributes the multiply to allow constant folding. When | |
156 | two or more addressing expressions can be represented by MEM_REFs | |
157 | of this form, differing only in the constants C1, C2, and C4, | |
158 | making this substitution produces more efficient addressing during | |
159 | the RTL phases. When there are not at least two expressions with | |
160 | the same values of T1, T2, and C3, there is nothing to be gained | |
161 | by the replacement. | |
162 | ||
163 | Strength reduction of CAND_REFs uses the same infrastructure as | |
164 | that used by CAND_MULTs and CAND_ADDs. We record T1 in the base (B) | |
165 | field, MULT_EXPR (T2, C3) in the stride (S) field, and | |
166 | C1 + (C2 * C3) + C4 in the index (i) field. A basis for a CAND_REF | |
167 | is thus another CAND_REF with the same B and S values. When at | |
168 | least two CAND_REFs are chained together using the basis relation, | |
169 | each of them is replaced as above, resulting in improved code | |
1168caca | 170 | generation for addressing. |
171 | ||
172 | Conditional candidates | |
173 | ====================== | |
174 | ||
175 | Conditional candidates are best illustrated with an example. | |
176 | Consider the code sequence: | |
177 | ||
178 | (1) x_0 = ...; | |
179 | (2) a_0 = x_0 * 5; MULT (B: x_0; i: 0; S: 5) | |
180 | if (...) | |
181 | (3) x_1 = x_0 + 1; ADD (B: x_0, i: 1; S: 1) | |
182 | (4) x_2 = PHI <x_0, x_1>; PHI (B: x_0, i: 0, S: 1) | |
183 | (5) x_3 = x_2 + 1; ADD (B: x_2, i: 1, S: 1) | |
184 | (6) a_1 = x_3 * 5; MULT (B: x_2, i: 1; S: 5) | |
185 | ||
186 | Here strength reduction is complicated by the uncertain value of x_2. | |
187 | A legitimate transformation is: | |
188 | ||
189 | (1) x_0 = ...; | |
190 | (2) a_0 = x_0 * 5; | |
191 | if (...) | |
192 | { | |
193 | (3) [x_1 = x_0 + 1;] | |
194 | (3a) t_1 = a_0 + 5; | |
195 | } | |
196 | (4) [x_2 = PHI <x_0, x_1>;] | |
197 | (4a) t_2 = PHI <a_0, t_1>; | |
198 | (5) [x_3 = x_2 + 1;] | |
199 | (6r) a_1 = t_2 + 5; | |
200 | ||
201 | where the bracketed instructions may go dead. | |
202 | ||
203 | To recognize this opportunity, we have to observe that statement (6) | |
204 | has a "hidden basis" (2). The hidden basis is unlike a normal basis | |
205 | in that the statement and the hidden basis have different base SSA | |
206 | names (x_2 and x_0, respectively). The relationship is established | |
207 | when a statement's base name (x_2) is defined by a phi statement (4), | |
208 | each argument of which (x_0, x_1) has an identical "derived base name." | |
209 | If the argument is defined by a candidate (as x_1 is by (3)) that is a | |
210 | CAND_ADD having a stride of 1, the derived base name of the argument is | |
211 | the base name of the candidate (x_0). Otherwise, the argument itself | |
212 | is its derived base name (as is the case with argument x_0). | |
213 | ||
214 | The hidden basis for statement (6) is the nearest dominating candidate | |
215 | whose base name is the derived base name (x_0) of the feeding phi (4), | |
216 | and whose stride is identical to that of the statement. We can then | |
217 | create the new "phi basis" (4a) and feeding adds along incoming arcs (3a), | |
218 | allowing the final replacement of (6) by the strength-reduced (6r). | |
219 | ||
220 | To facilitate this, a new kind of candidate (CAND_PHI) is introduced. | |
221 | A CAND_PHI is not a candidate for replacement, but is maintained in the | |
222 | candidate table to ease discovery of hidden bases. Any phi statement | |
223 | whose arguments share a common derived base name is entered into the | |
224 | table with the derived base name, an (arbitrary) index of zero, and a | |
225 | stride of 1. A statement with a hidden basis can then be detected by | |
226 | simply looking up its feeding phi definition in the candidate table, | |
227 | extracting the derived base name, and searching for a basis in the | |
228 | usual manner after substituting the derived base name. | |
229 | ||
230 | Note that the transformation is only valid when the original phi and | |
231 | the statements that define the phi's arguments are all at the same | |
232 | position in the loop hierarchy. */ | |
4502f5d0 | 233 | |
234 | ||
235 | /* Index into the candidate vector, offset by 1. VECs are zero-based, | |
236 | while cand_idx's are one-based, with zero indicating null. */ | |
237 | typedef unsigned cand_idx; | |
238 | ||
239 | /* The kind of candidate. */ | |
240 | enum cand_kind | |
241 | { | |
242 | CAND_MULT, | |
070bf980 | 243 | CAND_ADD, |
1168caca | 244 | CAND_REF, |
245 | CAND_PHI | |
4502f5d0 | 246 | }; |
247 | ||
248 | struct slsr_cand_d | |
249 | { | |
250 | /* The candidate statement S1. */ | |
251 | gimple cand_stmt; | |
252 | ||
c99e471e | 253 | /* The base expression B: often an SSA name, but not always. */ |
254 | tree base_expr; | |
4502f5d0 | 255 | |
256 | /* The stride S. */ | |
257 | tree stride; | |
258 | ||
259 | /* The index constant i. */ | |
5de9d3ed | 260 | widest_int index; |
4502f5d0 | 261 | |
c99e471e | 262 | /* The type of the candidate. This is normally the type of base_expr, |
4502f5d0 | 263 | but casts may have occurred when combining feeding instructions. |
070bf980 | 264 | A candidate can only be a basis for candidates of the same final type. |
265 | (For CAND_REFs, this is the type to be used for operand 1 of the | |
266 | replacement MEM_REF.) */ | |
4502f5d0 | 267 | tree cand_type; |
268 | ||
269 | /* The kind of candidate (CAND_MULT, etc.). */ | |
270 | enum cand_kind kind; | |
271 | ||
272 | /* Index of this candidate in the candidate vector. */ | |
273 | cand_idx cand_num; | |
274 | ||
275 | /* Index of the next candidate record for the same statement. | |
276 | A statement may be useful in more than one way (e.g., due to | |
277 | commutativity). So we can have multiple "interpretations" | |
278 | of a statement. */ | |
279 | cand_idx next_interp; | |
280 | ||
281 | /* Index of the basis statement S0, if any, in the candidate vector. */ | |
282 | cand_idx basis; | |
283 | ||
284 | /* First candidate for which this candidate is a basis, if one exists. */ | |
285 | cand_idx dependent; | |
286 | ||
287 | /* Next candidate having the same basis as this one. */ | |
288 | cand_idx sibling; | |
289 | ||
1168caca | 290 | /* If this is a conditional candidate, the CAND_PHI candidate |
291 | that defines the base SSA name B. */ | |
292 | cand_idx def_phi; | |
4502f5d0 | 293 | |
294 | /* Savings that can be expected from eliminating dead code if this | |
295 | candidate is replaced. */ | |
296 | int dead_savings; | |
297 | }; | |
298 | ||
299 | typedef struct slsr_cand_d slsr_cand, *slsr_cand_t; | |
300 | typedef const struct slsr_cand_d *const_slsr_cand_t; | |
301 | ||
302 | /* Pointers to candidates are chained together as part of a mapping | |
c99e471e | 303 | from base expressions to the candidates that use them. */ |
4502f5d0 | 304 | |
305 | struct cand_chain_d | |
306 | { | |
c99e471e | 307 | /* Base expression for the chain of candidates: often, but not |
308 | always, an SSA name. */ | |
309 | tree base_expr; | |
4502f5d0 | 310 | |
311 | /* Pointer to a candidate. */ | |
312 | slsr_cand_t cand; | |
313 | ||
314 | /* Chain pointer. */ | |
315 | struct cand_chain_d *next; | |
316 | ||
317 | }; | |
318 | ||
319 | typedef struct cand_chain_d cand_chain, *cand_chain_t; | |
320 | typedef const struct cand_chain_d *const_cand_chain_t; | |
321 | ||
9ce0359e | 322 | /* Information about a unique "increment" associated with candidates |
323 | having an SSA name for a stride. An increment is the difference | |
324 | between the index of the candidate and the index of its basis, | |
325 | i.e., (i - i') as discussed in the module commentary. | |
326 | ||
327 | When we are not going to generate address arithmetic we treat | |
328 | increments that differ only in sign as the same, allowing sharing | |
329 | of the cost of initializers. The absolute value of the increment | |
330 | is stored in the incr_info. */ | |
331 | ||
332 | struct incr_info_d | |
333 | { | |
334 | /* The increment that relates a candidate to its basis. */ | |
5de9d3ed | 335 | widest_int incr; |
9ce0359e | 336 | |
337 | /* How many times the increment occurs in the candidate tree. */ | |
338 | unsigned count; | |
339 | ||
340 | /* Cost of replacing candidates using this increment. Negative and | |
341 | zero costs indicate replacement should be performed. */ | |
342 | int cost; | |
343 | ||
344 | /* If this increment is profitable but is not -1, 0, or 1, it requires | |
345 | an initializer T_0 = stride * incr to be found or introduced in the | |
346 | nearest common dominator of all candidates. This field holds T_0 | |
347 | for subsequent use. */ | |
348 | tree initializer; | |
349 | ||
350 | /* If the initializer was found to already exist, this is the block | |
351 | where it was found. */ | |
352 | basic_block init_bb; | |
353 | }; | |
354 | ||
355 | typedef struct incr_info_d incr_info, *incr_info_t; | |
356 | ||
4502f5d0 | 357 | /* Candidates are maintained in a vector. If candidate X dominates |
358 | candidate Y, then X appears before Y in the vector; but the | |
359 | converse does not necessarily hold. */ | |
f1f41a6c | 360 | static vec<slsr_cand_t> cand_vec; |
4502f5d0 | 361 | |
362 | enum cost_consts | |
363 | { | |
364 | COST_NEUTRAL = 0, | |
365 | COST_INFINITE = 1000 | |
366 | }; | |
367 | ||
1168caca | 368 | enum stride_status |
369 | { | |
370 | UNKNOWN_STRIDE = 0, | |
371 | KNOWN_STRIDE = 1 | |
372 | }; | |
373 | ||
374 | enum phi_adjust_status | |
375 | { | |
376 | NOT_PHI_ADJUST = 0, | |
377 | PHI_ADJUST = 1 | |
378 | }; | |
379 | ||
380 | enum count_phis_status | |
381 | { | |
382 | DONT_COUNT_PHIS = 0, | |
383 | COUNT_PHIS = 1 | |
384 | }; | |
385 | ||
4502f5d0 | 386 | /* Pointer map embodying a mapping from statements to candidates. */ |
06ecf488 | 387 | static hash_map<gimple, slsr_cand_t> *stmt_cand_map; |
4502f5d0 | 388 | |
389 | /* Obstack for candidates. */ | |
390 | static struct obstack cand_obstack; | |
391 | ||
4502f5d0 | 392 | /* Obstack for candidate chains. */ |
393 | static struct obstack chain_obstack; | |
9ce0359e | 394 | |
395 | /* An array INCR_VEC of incr_infos is used during analysis of related | |
396 | candidates having an SSA name for a stride. INCR_VEC_LEN describes | |
d62a8480 | 397 | its current length. MAX_INCR_VEC_LEN is used to avoid costly |
398 | pathological cases. */ | |
9ce0359e | 399 | static incr_info_t incr_vec; |
400 | static unsigned incr_vec_len; | |
d62a8480 | 401 | const int MAX_INCR_VEC_LEN = 16; |
9ce0359e | 402 | |
403 | /* For a chain of candidates with unknown stride, indicates whether or not | |
404 | we must generate pointer arithmetic when replacing statements. */ | |
405 | static bool address_arithmetic_p; | |
1168caca | 406 | |
407 | /* Forward function declarations. */ | |
408 | static slsr_cand_t base_cand_from_table (tree); | |
d54ab372 | 409 | static tree introduce_cast_before_cand (slsr_cand_t, tree, tree); |
ee4e4f98 | 410 | static bool legal_cast_p_1 (tree, tree); |
4502f5d0 | 411 | \f |
412 | /* Produce a pointer to the IDX'th candidate in the candidate vector. */ | |
413 | ||
414 | static slsr_cand_t | |
415 | lookup_cand (cand_idx idx) | |
416 | { | |
f1f41a6c | 417 | return cand_vec[idx - 1]; |
4502f5d0 | 418 | } |
419 | ||
d9dd21a8 | 420 | /* Helper for hashing a candidate chain header. */ |
070bf980 | 421 | |
d9dd21a8 | 422 | struct cand_chain_hasher : typed_noop_remove <cand_chain> |
070bf980 | 423 | { |
d9dd21a8 | 424 | typedef cand_chain value_type; |
425 | typedef cand_chain compare_type; | |
426 | static inline hashval_t hash (const value_type *); | |
427 | static inline bool equal (const value_type *, const compare_type *); | |
428 | }; | |
070bf980 | 429 | |
d9dd21a8 | 430 | inline hashval_t |
431 | cand_chain_hasher::hash (const value_type *p) | |
070bf980 | 432 | { |
d9dd21a8 | 433 | tree base_expr = p->base_expr; |
434 | return iterative_hash_expr (base_expr, 0); | |
070bf980 | 435 | } |
436 | ||
d9dd21a8 | 437 | inline bool |
438 | cand_chain_hasher::equal (const value_type *chain1, const compare_type *chain2) | |
070bf980 | 439 | { |
c99e471e | 440 | return operand_equal_p (chain1->base_expr, chain2->base_expr, 0); |
070bf980 | 441 | } |
d9dd21a8 | 442 | |
443 | /* Hash table embodying a mapping from base exprs to chains of candidates. */ | |
c1f445d2 | 444 | static hash_table<cand_chain_hasher> *base_cand_map; |
070bf980 | 445 | \f |
0ca43779 | 446 | /* Pointer map used by tree_to_aff_combination_expand. */ |
5f8841a5 | 447 | static hash_map<tree, name_expansion *> *name_expansions; |
0ca43779 | 448 | /* Pointer map embodying a mapping from bases to alternative bases. */ |
06ecf488 | 449 | static hash_map<tree, tree> *alt_base_map; |
0ca43779 | 450 | |
451 | /* Given BASE, use the tree affine combiniation facilities to | |
452 | find the underlying tree expression for BASE, with any | |
cf6458c3 | 453 | immediate offset excluded. |
454 | ||
455 | N.B. we should eliminate this backtracking with better forward | |
456 | analysis in a future release. */ | |
0ca43779 | 457 | |
458 | static tree | |
459 | get_alternative_base (tree base) | |
460 | { | |
06ecf488 | 461 | tree *result = alt_base_map->get (base); |
0ca43779 | 462 | |
463 | if (result == NULL) | |
464 | { | |
465 | tree expr; | |
466 | aff_tree aff; | |
467 | ||
468 | tree_to_aff_combination_expand (base, TREE_TYPE (base), | |
469 | &aff, &name_expansions); | |
bdff91a1 | 470 | aff.offset = 0; |
0ca43779 | 471 | expr = aff_combination_to_tree (&aff); |
472 | ||
06ecf488 | 473 | gcc_assert (!alt_base_map->put (base, base == expr ? NULL : expr)); |
0ca43779 | 474 | |
06ecf488 | 475 | return expr == base ? NULL : expr; |
0ca43779 | 476 | } |
477 | ||
478 | return *result; | |
479 | } | |
480 | ||
1168caca | 481 | /* Look in the candidate table for a CAND_PHI that defines BASE and |
8858a76d | 482 | return it if found; otherwise return NULL. */ |
4502f5d0 | 483 | |
1168caca | 484 | static cand_idx |
8858a76d | 485 | find_phi_def (tree base) |
1168caca | 486 | { |
487 | slsr_cand_t c; | |
488 | ||
489 | if (TREE_CODE (base) != SSA_NAME) | |
490 | return 0; | |
cebae8f7 | 491 | |
1168caca | 492 | c = base_cand_from_table (base); |
493 | ||
494 | if (!c || c->kind != CAND_PHI) | |
495 | return 0; | |
496 | ||
497 | return c->cand_num; | |
498 | } | |
499 | ||
500 | /* Helper routine for find_basis_for_candidate. May be called twice: | |
0ca43779 | 501 | once for the candidate's base expr, and optionally again either for |
502 | the candidate's phi definition or for a CAND_REF's alternative base | |
503 | expression. */ | |
1168caca | 504 | |
505 | static slsr_cand_t | |
506 | find_basis_for_base_expr (slsr_cand_t c, tree base_expr) | |
4502f5d0 | 507 | { |
070bf980 | 508 | cand_chain mapping_key; |
4502f5d0 | 509 | cand_chain_t chain; |
510 | slsr_cand_t basis = NULL; | |
511 | ||
df4c32b3 | 512 | // Limit potential of N^2 behavior for long candidate chains. |
513 | int iters = 0; | |
514 | int max_iters = PARAM_VALUE (PARAM_MAX_SLSR_CANDIDATE_SCAN); | |
515 | ||
1168caca | 516 | mapping_key.base_expr = base_expr; |
c1f445d2 | 517 | chain = base_cand_map->find (&mapping_key); |
4502f5d0 | 518 | |
df4c32b3 | 519 | for (; chain && iters < max_iters; chain = chain->next, ++iters) |
4502f5d0 | 520 | { |
521 | slsr_cand_t one_basis = chain->cand; | |
522 | ||
523 | if (one_basis->kind != c->kind | |
9cc0ec41 | 524 | || one_basis->cand_stmt == c->cand_stmt |
4502f5d0 | 525 | || !operand_equal_p (one_basis->stride, c->stride, 0) |
526 | || !types_compatible_p (one_basis->cand_type, c->cand_type) | |
527 | || !dominated_by_p (CDI_DOMINATORS, | |
528 | gimple_bb (c->cand_stmt), | |
529 | gimple_bb (one_basis->cand_stmt))) | |
530 | continue; | |
531 | ||
532 | if (!basis || basis->cand_num < one_basis->cand_num) | |
533 | basis = one_basis; | |
534 | } | |
535 | ||
1168caca | 536 | return basis; |
537 | } | |
538 | ||
539 | /* Use the base expr from candidate C to look for possible candidates | |
540 | that can serve as a basis for C. Each potential basis must also | |
541 | appear in a block that dominates the candidate statement and have | |
542 | the same stride and type. If more than one possible basis exists, | |
543 | the one with highest index in the vector is chosen; this will be | |
544 | the most immediately dominating basis. */ | |
545 | ||
546 | static int | |
547 | find_basis_for_candidate (slsr_cand_t c) | |
548 | { | |
549 | slsr_cand_t basis = find_basis_for_base_expr (c, c->base_expr); | |
550 | ||
551 | /* If a candidate doesn't have a basis using its base expression, | |
552 | it may have a basis hidden by one or more intervening phis. */ | |
553 | if (!basis && c->def_phi) | |
554 | { | |
555 | basic_block basis_bb, phi_bb; | |
556 | slsr_cand_t phi_cand = lookup_cand (c->def_phi); | |
557 | basis = find_basis_for_base_expr (c, phi_cand->base_expr); | |
558 | ||
559 | if (basis) | |
560 | { | |
561 | /* A hidden basis must dominate the phi-definition of the | |
562 | candidate's base name. */ | |
563 | phi_bb = gimple_bb (phi_cand->cand_stmt); | |
564 | basis_bb = gimple_bb (basis->cand_stmt); | |
565 | ||
566 | if (phi_bb == basis_bb | |
567 | || !dominated_by_p (CDI_DOMINATORS, phi_bb, basis_bb)) | |
568 | { | |
569 | basis = NULL; | |
570 | c->basis = 0; | |
571 | } | |
572 | ||
573 | /* If we found a hidden basis, estimate additional dead-code | |
574 | savings if the phi and its feeding statements can be removed. */ | |
575 | if (basis && has_single_use (gimple_phi_result (phi_cand->cand_stmt))) | |
576 | c->dead_savings += phi_cand->dead_savings; | |
577 | } | |
578 | } | |
579 | ||
cf6458c3 | 580 | if (flag_expensive_optimizations && !basis && c->kind == CAND_REF) |
0ca43779 | 581 | { |
582 | tree alt_base_expr = get_alternative_base (c->base_expr); | |
583 | if (alt_base_expr) | |
584 | basis = find_basis_for_base_expr (c, alt_base_expr); | |
585 | } | |
586 | ||
4502f5d0 | 587 | if (basis) |
588 | { | |
589 | c->sibling = basis->dependent; | |
590 | basis->dependent = c->cand_num; | |
591 | return basis->cand_num; | |
592 | } | |
593 | ||
594 | return 0; | |
595 | } | |
596 | ||
0ca43779 | 597 | /* Record a mapping from BASE to C, indicating that C may potentially serve |
598 | as a basis using that base expression. BASE may be the same as | |
599 | C->BASE_EXPR; alternatively BASE can be a different tree that share the | |
600 | underlining expression of C->BASE_EXPR. */ | |
4502f5d0 | 601 | |
602 | static void | |
0ca43779 | 603 | record_potential_basis (slsr_cand_t c, tree base) |
4502f5d0 | 604 | { |
070bf980 | 605 | cand_chain_t node; |
d9dd21a8 | 606 | cand_chain **slot; |
4502f5d0 | 607 | |
0ca43779 | 608 | gcc_assert (base); |
609 | ||
4502f5d0 | 610 | node = (cand_chain_t) obstack_alloc (&chain_obstack, sizeof (cand_chain)); |
0ca43779 | 611 | node->base_expr = base; |
4502f5d0 | 612 | node->cand = c; |
613 | node->next = NULL; | |
c1f445d2 | 614 | slot = base_cand_map->find_slot (node, INSERT); |
4502f5d0 | 615 | |
070bf980 | 616 | if (*slot) |
4502f5d0 | 617 | { |
070bf980 | 618 | cand_chain_t head = (cand_chain_t) (*slot); |
4502f5d0 | 619 | node->next = head->next; |
620 | head->next = node; | |
621 | } | |
622 | else | |
070bf980 | 623 | *slot = node; |
4502f5d0 | 624 | } |
625 | ||
626 | /* Allocate storage for a new candidate and initialize its fields. | |
0ca43779 | 627 | Attempt to find a basis for the candidate. |
628 | ||
629 | For CAND_REF, an alternative base may also be recorded and used | |
630 | to find a basis. This helps cases where the expression hidden | |
631 | behind BASE (which is usually an SSA_NAME) has immediate offset, | |
632 | e.g. | |
633 | ||
634 | a2[i][j] = 1; | |
635 | a2[i + 20][j] = 2; */ | |
4502f5d0 | 636 | |
637 | static slsr_cand_t | |
0ca43779 | 638 | alloc_cand_and_find_basis (enum cand_kind kind, gimple gs, tree base, |
5de9d3ed | 639 | const widest_int &index, tree stride, tree ctype, |
4502f5d0 | 640 | unsigned savings) |
641 | { | |
642 | slsr_cand_t c = (slsr_cand_t) obstack_alloc (&cand_obstack, | |
643 | sizeof (slsr_cand)); | |
644 | c->cand_stmt = gs; | |
c99e471e | 645 | c->base_expr = base; |
4502f5d0 | 646 | c->stride = stride; |
647 | c->index = index; | |
648 | c->cand_type = ctype; | |
649 | c->kind = kind; | |
f1f41a6c | 650 | c->cand_num = cand_vec.length () + 1; |
4502f5d0 | 651 | c->next_interp = 0; |
652 | c->dependent = 0; | |
653 | c->sibling = 0; | |
8858a76d | 654 | c->def_phi = kind == CAND_MULT ? find_phi_def (base) : 0; |
4502f5d0 | 655 | c->dead_savings = savings; |
656 | ||
f1f41a6c | 657 | cand_vec.safe_push (c); |
1168caca | 658 | |
659 | if (kind == CAND_PHI) | |
660 | c->basis = 0; | |
661 | else | |
662 | c->basis = find_basis_for_candidate (c); | |
663 | ||
0ca43779 | 664 | record_potential_basis (c, base); |
cf6458c3 | 665 | if (flag_expensive_optimizations && kind == CAND_REF) |
0ca43779 | 666 | { |
667 | tree alt_base = get_alternative_base (base); | |
668 | if (alt_base) | |
669 | record_potential_basis (c, alt_base); | |
670 | } | |
4502f5d0 | 671 | |
672 | return c; | |
673 | } | |
674 | ||
675 | /* Determine the target cost of statement GS when compiling according | |
676 | to SPEED. */ | |
677 | ||
678 | static int | |
679 | stmt_cost (gimple gs, bool speed) | |
680 | { | |
681 | tree lhs, rhs1, rhs2; | |
3754d046 | 682 | machine_mode lhs_mode; |
4502f5d0 | 683 | |
684 | gcc_assert (is_gimple_assign (gs)); | |
685 | lhs = gimple_assign_lhs (gs); | |
686 | rhs1 = gimple_assign_rhs1 (gs); | |
687 | lhs_mode = TYPE_MODE (TREE_TYPE (lhs)); | |
688 | ||
689 | switch (gimple_assign_rhs_code (gs)) | |
690 | { | |
691 | case MULT_EXPR: | |
692 | rhs2 = gimple_assign_rhs2 (gs); | |
693 | ||
e913b5cd | 694 | if (tree_fits_shwi_p (rhs2)) |
695 | return mult_by_coeff_cost (tree_to_shwi (rhs2), lhs_mode, speed); | |
4502f5d0 | 696 | |
697 | gcc_assert (TREE_CODE (rhs1) != INTEGER_CST); | |
49a71e58 | 698 | return mul_cost (speed, lhs_mode); |
4502f5d0 | 699 | |
700 | case PLUS_EXPR: | |
701 | case POINTER_PLUS_EXPR: | |
702 | case MINUS_EXPR: | |
49a71e58 | 703 | return add_cost (speed, lhs_mode); |
4502f5d0 | 704 | |
705 | case NEGATE_EXPR: | |
49a71e58 | 706 | return neg_cost (speed, lhs_mode); |
4502f5d0 | 707 | |
708 | case NOP_EXPR: | |
72655676 | 709 | return convert_cost (lhs_mode, TYPE_MODE (TREE_TYPE (rhs1)), speed); |
4502f5d0 | 710 | |
711 | /* Note that we don't assign costs to copies that in most cases | |
712 | will go away. */ | |
713 | default: | |
714 | ; | |
715 | } | |
716 | ||
717 | gcc_unreachable (); | |
718 | return 0; | |
719 | } | |
720 | ||
721 | /* Look up the defining statement for BASE_IN and return a pointer | |
722 | to its candidate in the candidate table, if any; otherwise NULL. | |
723 | Only CAND_ADD and CAND_MULT candidates are returned. */ | |
724 | ||
725 | static slsr_cand_t | |
726 | base_cand_from_table (tree base_in) | |
727 | { | |
728 | slsr_cand_t *result; | |
729 | ||
730 | gimple def = SSA_NAME_DEF_STMT (base_in); | |
731 | if (!def) | |
732 | return (slsr_cand_t) NULL; | |
733 | ||
06ecf488 | 734 | result = stmt_cand_map->get (def); |
070bf980 | 735 | |
736 | if (result && (*result)->kind != CAND_REF) | |
737 | return *result; | |
4502f5d0 | 738 | |
070bf980 | 739 | return (slsr_cand_t) NULL; |
4502f5d0 | 740 | } |
741 | ||
742 | /* Add an entry to the statement-to-candidate mapping. */ | |
743 | ||
744 | static void | |
745 | add_cand_for_stmt (gimple gs, slsr_cand_t c) | |
746 | { | |
06ecf488 | 747 | gcc_assert (!stmt_cand_map->put (gs, c)); |
4502f5d0 | 748 | } |
749 | \f | |
1168caca | 750 | /* Given PHI which contains a phi statement, determine whether it |
751 | satisfies all the requirements of a phi candidate. If so, create | |
752 | a candidate. Note that a CAND_PHI never has a basis itself, but | |
753 | is used to help find a basis for subsequent candidates. */ | |
754 | ||
755 | static void | |
756 | slsr_process_phi (gimple phi, bool speed) | |
757 | { | |
758 | unsigned i; | |
759 | tree arg0_base = NULL_TREE, base_type; | |
760 | slsr_cand_t c; | |
761 | struct loop *cand_loop = gimple_bb (phi)->loop_father; | |
762 | unsigned savings = 0; | |
763 | ||
764 | /* A CAND_PHI requires each of its arguments to have the same | |
765 | derived base name. (See the module header commentary for a | |
766 | definition of derived base names.) Furthermore, all feeding | |
767 | definitions must be in the same position in the loop hierarchy | |
768 | as PHI. */ | |
769 | ||
770 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
771 | { | |
772 | slsr_cand_t arg_cand; | |
773 | tree arg = gimple_phi_arg_def (phi, i); | |
774 | tree derived_base_name = NULL_TREE; | |
775 | gimple arg_stmt = NULL; | |
776 | basic_block arg_bb = NULL; | |
777 | ||
778 | if (TREE_CODE (arg) != SSA_NAME) | |
779 | return; | |
780 | ||
781 | arg_cand = base_cand_from_table (arg); | |
782 | ||
783 | if (arg_cand) | |
784 | { | |
785 | while (arg_cand->kind != CAND_ADD && arg_cand->kind != CAND_PHI) | |
786 | { | |
787 | if (!arg_cand->next_interp) | |
788 | return; | |
789 | ||
790 | arg_cand = lookup_cand (arg_cand->next_interp); | |
791 | } | |
792 | ||
793 | if (!integer_onep (arg_cand->stride)) | |
794 | return; | |
795 | ||
796 | derived_base_name = arg_cand->base_expr; | |
797 | arg_stmt = arg_cand->cand_stmt; | |
798 | arg_bb = gimple_bb (arg_stmt); | |
799 | ||
800 | /* Gather potential dead code savings if the phi statement | |
801 | can be removed later on. */ | |
802 | if (has_single_use (arg)) | |
803 | { | |
804 | if (gimple_code (arg_stmt) == GIMPLE_PHI) | |
805 | savings += arg_cand->dead_savings; | |
806 | else | |
807 | savings += stmt_cost (arg_stmt, speed); | |
808 | } | |
809 | } | |
810 | else | |
811 | { | |
812 | derived_base_name = arg; | |
813 | ||
814 | if (SSA_NAME_IS_DEFAULT_DEF (arg)) | |
34154e27 | 815 | arg_bb = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)); |
1168caca | 816 | else |
817 | gimple_bb (SSA_NAME_DEF_STMT (arg)); | |
818 | } | |
819 | ||
820 | if (!arg_bb || arg_bb->loop_father != cand_loop) | |
821 | return; | |
822 | ||
823 | if (i == 0) | |
824 | arg0_base = derived_base_name; | |
825 | else if (!operand_equal_p (derived_base_name, arg0_base, 0)) | |
826 | return; | |
827 | } | |
828 | ||
829 | /* Create the candidate. "alloc_cand_and_find_basis" is named | |
830 | misleadingly for this case, as no basis will be sought for a | |
831 | CAND_PHI. */ | |
832 | base_type = TREE_TYPE (arg0_base); | |
833 | ||
ddb1be65 | 834 | c = alloc_cand_and_find_basis (CAND_PHI, phi, arg0_base, |
e913b5cd | 835 | 0, integer_one_node, base_type, savings); |
1168caca | 836 | |
837 | /* Add the candidate to the statement-candidate mapping. */ | |
838 | add_cand_for_stmt (phi, c); | |
839 | } | |
840 | ||
3f8e7554 | 841 | /* Given PBASE which is a pointer to tree, look up the defining |
842 | statement for it and check whether the candidate is in the | |
843 | form of: | |
844 | ||
845 | X = B + (1 * S), S is integer constant | |
846 | X = B + (i * S), S is integer one | |
847 | ||
848 | If so, set PBASE to the candidate's base_expr and return double | |
849 | int (i * S). | |
850 | Otherwise, just return double int zero. */ | |
851 | ||
5de9d3ed | 852 | static widest_int |
3f8e7554 | 853 | backtrace_base_for_ref (tree *pbase) |
854 | { | |
855 | tree base_in = *pbase; | |
856 | slsr_cand_t base_cand; | |
857 | ||
858 | STRIP_NOPS (base_in); | |
ee4e4f98 | 859 | |
860 | /* Strip off widening conversion(s) to handle cases where | |
861 | e.g. 'B' is widened from an 'int' in order to calculate | |
862 | a 64-bit address. */ | |
863 | if (CONVERT_EXPR_P (base_in) | |
864 | && legal_cast_p_1 (base_in, TREE_OPERAND (base_in, 0))) | |
865 | base_in = get_unwidened (base_in, NULL_TREE); | |
866 | ||
3f8e7554 | 867 | if (TREE_CODE (base_in) != SSA_NAME) |
087b03ca | 868 | return 0; |
3f8e7554 | 869 | |
870 | base_cand = base_cand_from_table (base_in); | |
871 | ||
872 | while (base_cand && base_cand->kind != CAND_PHI) | |
873 | { | |
874 | if (base_cand->kind == CAND_ADD | |
087b03ca | 875 | && base_cand->index == 1 |
3f8e7554 | 876 | && TREE_CODE (base_cand->stride) == INTEGER_CST) |
877 | { | |
878 | /* X = B + (1 * S), S is integer constant. */ | |
879 | *pbase = base_cand->base_expr; | |
5de9d3ed | 880 | return wi::to_widest (base_cand->stride); |
3f8e7554 | 881 | } |
882 | else if (base_cand->kind == CAND_ADD | |
883 | && TREE_CODE (base_cand->stride) == INTEGER_CST | |
884 | && integer_onep (base_cand->stride)) | |
6da7e110 | 885 | { |
3f8e7554 | 886 | /* X = B + (i * S), S is integer one. */ |
887 | *pbase = base_cand->base_expr; | |
888 | return base_cand->index; | |
889 | } | |
890 | ||
891 | if (base_cand->next_interp) | |
892 | base_cand = lookup_cand (base_cand->next_interp); | |
893 | else | |
894 | base_cand = NULL; | |
895 | } | |
896 | ||
087b03ca | 897 | return 0; |
3f8e7554 | 898 | } |
899 | ||
070bf980 | 900 | /* Look for the following pattern: |
901 | ||
902 | *PBASE: MEM_REF (T1, C1) | |
903 | ||
904 | *POFFSET: MULT_EXPR (T2, C3) [C2 is zero] | |
905 | or | |
906 | MULT_EXPR (PLUS_EXPR (T2, C2), C3) | |
907 | or | |
908 | MULT_EXPR (MINUS_EXPR (T2, -C2), C3) | |
909 | ||
910 | *PINDEX: C4 * BITS_PER_UNIT | |
911 | ||
912 | If not present, leave the input values unchanged and return FALSE. | |
913 | Otherwise, modify the input values as follows and return TRUE: | |
914 | ||
915 | *PBASE: T1 | |
916 | *POFFSET: MULT_EXPR (T2, C3) | |
3f8e7554 | 917 | *PINDEX: C1 + (C2 * C3) + C4 |
918 | ||
919 | When T2 is recorded by a CAND_ADD in the form of (T2' + C5), it | |
920 | will be further restructured to: | |
921 | ||
922 | *PBASE: T1 | |
923 | *POFFSET: MULT_EXPR (T2', C3) | |
924 | *PINDEX: C1 + (C2 * C3) + C4 + (C5 * C3) */ | |
070bf980 | 925 | |
926 | static bool | |
5de9d3ed | 927 | restructure_reference (tree *pbase, tree *poffset, widest_int *pindex, |
070bf980 | 928 | tree *ptype) |
929 | { | |
930 | tree base = *pbase, offset = *poffset; | |
5de9d3ed | 931 | widest_int index = *pindex; |
e913b5cd | 932 | tree mult_op0, t1, t2, type; |
5de9d3ed | 933 | widest_int c1, c2, c3, c4, c5; |
070bf980 | 934 | |
935 | if (!base | |
936 | || !offset | |
937 | || TREE_CODE (base) != MEM_REF | |
938 | || TREE_CODE (offset) != MULT_EXPR | |
939 | || TREE_CODE (TREE_OPERAND (offset, 1)) != INTEGER_CST | |
796b6678 | 940 | || wi::umod_floor (index, BITS_PER_UNIT) != 0) |
070bf980 | 941 | return false; |
942 | ||
943 | t1 = TREE_OPERAND (base, 0); | |
5de9d3ed | 944 | c1 = widest_int::from (mem_ref_offset (base), SIGNED); |
070bf980 | 945 | type = TREE_TYPE (TREE_OPERAND (base, 1)); |
946 | ||
947 | mult_op0 = TREE_OPERAND (offset, 0); | |
5de9d3ed | 948 | c3 = wi::to_widest (TREE_OPERAND (offset, 1)); |
070bf980 | 949 | |
950 | if (TREE_CODE (mult_op0) == PLUS_EXPR) | |
951 | ||
952 | if (TREE_CODE (TREE_OPERAND (mult_op0, 1)) == INTEGER_CST) | |
953 | { | |
954 | t2 = TREE_OPERAND (mult_op0, 0); | |
5de9d3ed | 955 | c2 = wi::to_widest (TREE_OPERAND (mult_op0, 1)); |
070bf980 | 956 | } |
957 | else | |
958 | return false; | |
959 | ||
960 | else if (TREE_CODE (mult_op0) == MINUS_EXPR) | |
961 | ||
962 | if (TREE_CODE (TREE_OPERAND (mult_op0, 1)) == INTEGER_CST) | |
963 | { | |
964 | t2 = TREE_OPERAND (mult_op0, 0); | |
5de9d3ed | 965 | c2 = -wi::to_widest (TREE_OPERAND (mult_op0, 1)); |
070bf980 | 966 | } |
967 | else | |
968 | return false; | |
969 | ||
970 | else | |
971 | { | |
972 | t2 = mult_op0; | |
e913b5cd | 973 | c2 = 0; |
070bf980 | 974 | } |
975 | ||
2315e038 | 976 | c4 = wi::lrshift (index, LOG2_BITS_PER_UNIT); |
3f8e7554 | 977 | c5 = backtrace_base_for_ref (&t2); |
070bf980 | 978 | |
979 | *pbase = t1; | |
3f8e7554 | 980 | *poffset = fold_build2 (MULT_EXPR, sizetype, fold_convert (sizetype, t2), |
e913b5cd | 981 | wide_int_to_tree (sizetype, c3)); |
3f8e7554 | 982 | *pindex = c1 + c2 * c3 + c4 + c5 * c3; |
070bf980 | 983 | *ptype = type; |
984 | ||
985 | return true; | |
986 | } | |
987 | ||
988 | /* Given GS which contains a data reference, create a CAND_REF entry in | |
989 | the candidate table and attempt to find a basis. */ | |
990 | ||
991 | static void | |
992 | slsr_process_ref (gimple gs) | |
993 | { | |
994 | tree ref_expr, base, offset, type; | |
995 | HOST_WIDE_INT bitsize, bitpos; | |
3754d046 | 996 | machine_mode mode; |
070bf980 | 997 | int unsignedp, volatilep; |
070bf980 | 998 | slsr_cand_t c; |
999 | ||
1000 | if (gimple_vdef (gs)) | |
1001 | ref_expr = gimple_assign_lhs (gs); | |
1002 | else | |
1003 | ref_expr = gimple_assign_rhs1 (gs); | |
1004 | ||
1005 | if (!handled_component_p (ref_expr) | |
1006 | || TREE_CODE (ref_expr) == BIT_FIELD_REF | |
1007 | || (TREE_CODE (ref_expr) == COMPONENT_REF | |
1008 | && DECL_BIT_FIELD (TREE_OPERAND (ref_expr, 1)))) | |
1009 | return; | |
1010 | ||
1011 | base = get_inner_reference (ref_expr, &bitsize, &bitpos, &offset, &mode, | |
1012 | &unsignedp, &volatilep, false); | |
ab2c1de8 | 1013 | widest_int index = bitpos; |
070bf980 | 1014 | |
1015 | if (!restructure_reference (&base, &offset, &index, &type)) | |
1016 | return; | |
1017 | ||
1018 | c = alloc_cand_and_find_basis (CAND_REF, gs, base, index, offset, | |
1019 | type, 0); | |
1020 | ||
1021 | /* Add the candidate to the statement-candidate mapping. */ | |
1022 | add_cand_for_stmt (gs, c); | |
1023 | } | |
1024 | ||
4502f5d0 | 1025 | /* Create a candidate entry for a statement GS, where GS multiplies |
1026 | two SSA names BASE_IN and STRIDE_IN. Propagate any known information | |
1027 | about the two SSA names into the new candidate. Return the new | |
1028 | candidate. */ | |
1029 | ||
1030 | static slsr_cand_t | |
1031 | create_mul_ssa_cand (gimple gs, tree base_in, tree stride_in, bool speed) | |
1032 | { | |
1033 | tree base = NULL_TREE, stride = NULL_TREE, ctype = NULL_TREE; | |
5de9d3ed | 1034 | widest_int index; |
4502f5d0 | 1035 | unsigned savings = 0; |
1036 | slsr_cand_t c; | |
1037 | slsr_cand_t base_cand = base_cand_from_table (base_in); | |
1038 | ||
1039 | /* Look at all interpretations of the base candidate, if necessary, | |
1040 | to find information to propagate into this candidate. */ | |
1168caca | 1041 | while (base_cand && !base && base_cand->kind != CAND_PHI) |
4502f5d0 | 1042 | { |
1043 | ||
1168caca | 1044 | if (base_cand->kind == CAND_MULT && integer_onep (base_cand->stride)) |
4502f5d0 | 1045 | { |
1046 | /* Y = (B + i') * 1 | |
1047 | X = Y * Z | |
1048 | ================ | |
1049 | X = (B + i') * Z */ | |
c99e471e | 1050 | base = base_cand->base_expr; |
4502f5d0 | 1051 | index = base_cand->index; |
1052 | stride = stride_in; | |
1053 | ctype = base_cand->cand_type; | |
1054 | if (has_single_use (base_in)) | |
1055 | savings = (base_cand->dead_savings | |
1056 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1057 | } | |
1058 | else if (base_cand->kind == CAND_ADD | |
1059 | && TREE_CODE (base_cand->stride) == INTEGER_CST) | |
1060 | { | |
1061 | /* Y = B + (i' * S), S constant | |
1062 | X = Y * Z | |
1063 | ============================ | |
1064 | X = B + ((i' * S) * Z) */ | |
c99e471e | 1065 | base = base_cand->base_expr; |
5de9d3ed | 1066 | index = base_cand->index * wi::to_widest (base_cand->stride); |
4502f5d0 | 1067 | stride = stride_in; |
1068 | ctype = base_cand->cand_type; | |
1069 | if (has_single_use (base_in)) | |
1070 | savings = (base_cand->dead_savings | |
1071 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1072 | } | |
1073 | ||
1074 | if (base_cand->next_interp) | |
1075 | base_cand = lookup_cand (base_cand->next_interp); | |
1076 | else | |
1077 | base_cand = NULL; | |
1078 | } | |
1079 | ||
1080 | if (!base) | |
1081 | { | |
1082 | /* No interpretations had anything useful to propagate, so | |
1083 | produce X = (Y + 0) * Z. */ | |
1084 | base = base_in; | |
e913b5cd | 1085 | index = 0; |
4502f5d0 | 1086 | stride = stride_in; |
3b652cc1 | 1087 | ctype = TREE_TYPE (base_in); |
4502f5d0 | 1088 | } |
1089 | ||
1090 | c = alloc_cand_and_find_basis (CAND_MULT, gs, base, index, stride, | |
1091 | ctype, savings); | |
1092 | return c; | |
1093 | } | |
1094 | ||
1095 | /* Create a candidate entry for a statement GS, where GS multiplies | |
1096 | SSA name BASE_IN by constant STRIDE_IN. Propagate any known | |
1097 | information about BASE_IN into the new candidate. Return the new | |
1098 | candidate. */ | |
1099 | ||
1100 | static slsr_cand_t | |
1101 | create_mul_imm_cand (gimple gs, tree base_in, tree stride_in, bool speed) | |
1102 | { | |
1103 | tree base = NULL_TREE, stride = NULL_TREE, ctype = NULL_TREE; | |
5de9d3ed | 1104 | widest_int index, temp; |
4502f5d0 | 1105 | unsigned savings = 0; |
1106 | slsr_cand_t c; | |
1107 | slsr_cand_t base_cand = base_cand_from_table (base_in); | |
1108 | ||
1109 | /* Look at all interpretations of the base candidate, if necessary, | |
1110 | to find information to propagate into this candidate. */ | |
1168caca | 1111 | while (base_cand && !base && base_cand->kind != CAND_PHI) |
4502f5d0 | 1112 | { |
1113 | if (base_cand->kind == CAND_MULT | |
1114 | && TREE_CODE (base_cand->stride) == INTEGER_CST) | |
1115 | { | |
1116 | /* Y = (B + i') * S, S constant | |
1117 | X = Y * c | |
1118 | ============================ | |
1119 | X = (B + i') * (S * c) */ | |
5de9d3ed | 1120 | temp = wi::to_widest (base_cand->stride) * wi::to_widest (stride_in); |
149a91a1 | 1121 | if (wi::fits_to_tree_p (temp, TREE_TYPE (stride_in))) |
05a4798a | 1122 | { |
1123 | base = base_cand->base_expr; | |
1124 | index = base_cand->index; | |
149a91a1 | 1125 | stride = wide_int_to_tree (TREE_TYPE (stride_in), temp); |
05a4798a | 1126 | ctype = base_cand->cand_type; |
1127 | if (has_single_use (base_in)) | |
1128 | savings = (base_cand->dead_savings | |
1129 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1130 | } | |
4502f5d0 | 1131 | } |
1168caca | 1132 | else if (base_cand->kind == CAND_ADD && integer_onep (base_cand->stride)) |
4502f5d0 | 1133 | { |
1134 | /* Y = B + (i' * 1) | |
1135 | X = Y * c | |
1136 | =========================== | |
1137 | X = (B + i') * c */ | |
c99e471e | 1138 | base = base_cand->base_expr; |
4502f5d0 | 1139 | index = base_cand->index; |
1140 | stride = stride_in; | |
1141 | ctype = base_cand->cand_type; | |
1142 | if (has_single_use (base_in)) | |
1143 | savings = (base_cand->dead_savings | |
1144 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1145 | } | |
1146 | else if (base_cand->kind == CAND_ADD | |
796b6678 | 1147 | && base_cand->index == 1 |
4502f5d0 | 1148 | && TREE_CODE (base_cand->stride) == INTEGER_CST) |
1149 | { | |
1150 | /* Y = B + (1 * S), S constant | |
1151 | X = Y * c | |
1152 | =========================== | |
1153 | X = (B + S) * c */ | |
c99e471e | 1154 | base = base_cand->base_expr; |
5de9d3ed | 1155 | index = wi::to_widest (base_cand->stride); |
4502f5d0 | 1156 | stride = stride_in; |
1157 | ctype = base_cand->cand_type; | |
1158 | if (has_single_use (base_in)) | |
1159 | savings = (base_cand->dead_savings | |
1160 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1161 | } | |
1162 | ||
1163 | if (base_cand->next_interp) | |
1164 | base_cand = lookup_cand (base_cand->next_interp); | |
1165 | else | |
1166 | base_cand = NULL; | |
1167 | } | |
1168 | ||
1169 | if (!base) | |
1170 | { | |
1171 | /* No interpretations had anything useful to propagate, so | |
1172 | produce X = (Y + 0) * c. */ | |
1173 | base = base_in; | |
e913b5cd | 1174 | index = 0; |
4502f5d0 | 1175 | stride = stride_in; |
3b652cc1 | 1176 | ctype = TREE_TYPE (base_in); |
4502f5d0 | 1177 | } |
1178 | ||
1179 | c = alloc_cand_and_find_basis (CAND_MULT, gs, base, index, stride, | |
1180 | ctype, savings); | |
1181 | return c; | |
1182 | } | |
1183 | ||
1184 | /* Given GS which is a multiply of scalar integers, make an appropriate | |
1185 | entry in the candidate table. If this is a multiply of two SSA names, | |
1186 | create two CAND_MULT interpretations and attempt to find a basis for | |
1187 | each of them. Otherwise, create a single CAND_MULT and attempt to | |
1188 | find a basis. */ | |
1189 | ||
1190 | static void | |
1191 | slsr_process_mul (gimple gs, tree rhs1, tree rhs2, bool speed) | |
1192 | { | |
1193 | slsr_cand_t c, c2; | |
1194 | ||
1195 | /* If this is a multiply of an SSA name with itself, it is highly | |
1196 | unlikely that we will get a strength reduction opportunity, so | |
1197 | don't record it as a candidate. This simplifies the logic for | |
1198 | finding a basis, so if this is removed that must be considered. */ | |
1199 | if (rhs1 == rhs2) | |
1200 | return; | |
1201 | ||
1202 | if (TREE_CODE (rhs2) == SSA_NAME) | |
1203 | { | |
1204 | /* Record an interpretation of this statement in the candidate table | |
c99e471e | 1205 | assuming RHS1 is the base expression and RHS2 is the stride. */ |
4502f5d0 | 1206 | c = create_mul_ssa_cand (gs, rhs1, rhs2, speed); |
1207 | ||
1208 | /* Add the first interpretation to the statement-candidate mapping. */ | |
1209 | add_cand_for_stmt (gs, c); | |
1210 | ||
1211 | /* Record another interpretation of this statement assuming RHS1 | |
c99e471e | 1212 | is the stride and RHS2 is the base expression. */ |
4502f5d0 | 1213 | c2 = create_mul_ssa_cand (gs, rhs2, rhs1, speed); |
1214 | c->next_interp = c2->cand_num; | |
1215 | } | |
1216 | else | |
1217 | { | |
1218 | /* Record an interpretation for the multiply-immediate. */ | |
1219 | c = create_mul_imm_cand (gs, rhs1, rhs2, speed); | |
1220 | ||
1221 | /* Add the interpretation to the statement-candidate mapping. */ | |
1222 | add_cand_for_stmt (gs, c); | |
1223 | } | |
1224 | } | |
1225 | ||
1226 | /* Create a candidate entry for a statement GS, where GS adds two | |
1227 | SSA names BASE_IN and ADDEND_IN if SUBTRACT_P is false, and | |
1228 | subtracts ADDEND_IN from BASE_IN otherwise. Propagate any known | |
1229 | information about the two SSA names into the new candidate. | |
1230 | Return the new candidate. */ | |
1231 | ||
1232 | static slsr_cand_t | |
1233 | create_add_ssa_cand (gimple gs, tree base_in, tree addend_in, | |
1234 | bool subtract_p, bool speed) | |
1235 | { | |
1236 | tree base = NULL_TREE, stride = NULL_TREE, ctype = NULL; | |
5de9d3ed | 1237 | widest_int index; |
4502f5d0 | 1238 | unsigned savings = 0; |
1239 | slsr_cand_t c; | |
1240 | slsr_cand_t base_cand = base_cand_from_table (base_in); | |
1241 | slsr_cand_t addend_cand = base_cand_from_table (addend_in); | |
1242 | ||
1243 | /* The most useful transformation is a multiply-immediate feeding | |
1244 | an add or subtract. Look for that first. */ | |
1168caca | 1245 | while (addend_cand && !base && addend_cand->kind != CAND_PHI) |
4502f5d0 | 1246 | { |
1247 | if (addend_cand->kind == CAND_MULT | |
796b6678 | 1248 | && addend_cand->index == 0 |
4502f5d0 | 1249 | && TREE_CODE (addend_cand->stride) == INTEGER_CST) |
1250 | { | |
1251 | /* Z = (B + 0) * S, S constant | |
1252 | X = Y +/- Z | |
1253 | =========================== | |
1254 | X = Y + ((+/-1 * S) * B) */ | |
1255 | base = base_in; | |
5de9d3ed | 1256 | index = wi::to_widest (addend_cand->stride); |
4502f5d0 | 1257 | if (subtract_p) |
cf8f0e63 | 1258 | index = -index; |
c99e471e | 1259 | stride = addend_cand->base_expr; |
3b652cc1 | 1260 | ctype = TREE_TYPE (base_in); |
4502f5d0 | 1261 | if (has_single_use (addend_in)) |
1262 | savings = (addend_cand->dead_savings | |
1263 | + stmt_cost (addend_cand->cand_stmt, speed)); | |
1264 | } | |
1265 | ||
1266 | if (addend_cand->next_interp) | |
1267 | addend_cand = lookup_cand (addend_cand->next_interp); | |
1268 | else | |
1269 | addend_cand = NULL; | |
1270 | } | |
1271 | ||
1168caca | 1272 | while (base_cand && !base && base_cand->kind != CAND_PHI) |
4502f5d0 | 1273 | { |
1274 | if (base_cand->kind == CAND_ADD | |
796b6678 | 1275 | && (base_cand->index == 0 |
4502f5d0 | 1276 | || operand_equal_p (base_cand->stride, |
1277 | integer_zero_node, 0))) | |
1278 | { | |
1279 | /* Y = B + (i' * S), i' * S = 0 | |
1280 | X = Y +/- Z | |
1281 | ============================ | |
1282 | X = B + (+/-1 * Z) */ | |
c99e471e | 1283 | base = base_cand->base_expr; |
e913b5cd | 1284 | index = subtract_p ? -1 : 1; |
4502f5d0 | 1285 | stride = addend_in; |
1286 | ctype = base_cand->cand_type; | |
1287 | if (has_single_use (base_in)) | |
1288 | savings = (base_cand->dead_savings | |
1289 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1290 | } | |
1291 | else if (subtract_p) | |
1292 | { | |
1293 | slsr_cand_t subtrahend_cand = base_cand_from_table (addend_in); | |
1294 | ||
1168caca | 1295 | while (subtrahend_cand && !base && subtrahend_cand->kind != CAND_PHI) |
4502f5d0 | 1296 | { |
1297 | if (subtrahend_cand->kind == CAND_MULT | |
796b6678 | 1298 | && subtrahend_cand->index == 0 |
4502f5d0 | 1299 | && TREE_CODE (subtrahend_cand->stride) == INTEGER_CST) |
1300 | { | |
1301 | /* Z = (B + 0) * S, S constant | |
1302 | X = Y - Z | |
1303 | =========================== | |
1304 | Value: X = Y + ((-1 * S) * B) */ | |
1305 | base = base_in; | |
5de9d3ed | 1306 | index = wi::to_widest (subtrahend_cand->stride); |
cf8f0e63 | 1307 | index = -index; |
c99e471e | 1308 | stride = subtrahend_cand->base_expr; |
3b652cc1 | 1309 | ctype = TREE_TYPE (base_in); |
4502f5d0 | 1310 | if (has_single_use (addend_in)) |
1311 | savings = (subtrahend_cand->dead_savings | |
1312 | + stmt_cost (subtrahend_cand->cand_stmt, speed)); | |
1313 | } | |
1314 | ||
1315 | if (subtrahend_cand->next_interp) | |
1316 | subtrahend_cand = lookup_cand (subtrahend_cand->next_interp); | |
1317 | else | |
1318 | subtrahend_cand = NULL; | |
1319 | } | |
1320 | } | |
1321 | ||
1322 | if (base_cand->next_interp) | |
1323 | base_cand = lookup_cand (base_cand->next_interp); | |
1324 | else | |
1325 | base_cand = NULL; | |
1326 | } | |
1327 | ||
1328 | if (!base) | |
1329 | { | |
1330 | /* No interpretations had anything useful to propagate, so | |
1331 | produce X = Y + (1 * Z). */ | |
1332 | base = base_in; | |
e913b5cd | 1333 | index = subtract_p ? -1 : 1; |
4502f5d0 | 1334 | stride = addend_in; |
3b652cc1 | 1335 | ctype = TREE_TYPE (base_in); |
4502f5d0 | 1336 | } |
1337 | ||
1338 | c = alloc_cand_and_find_basis (CAND_ADD, gs, base, index, stride, | |
1339 | ctype, savings); | |
1340 | return c; | |
1341 | } | |
1342 | ||
1343 | /* Create a candidate entry for a statement GS, where GS adds SSA | |
1344 | name BASE_IN to constant INDEX_IN. Propagate any known information | |
1345 | about BASE_IN into the new candidate. Return the new candidate. */ | |
1346 | ||
1347 | static slsr_cand_t | |
ab2c1de8 | 1348 | create_add_imm_cand (gimple gs, tree base_in, const widest_int &index_in, |
1349 | bool speed) | |
4502f5d0 | 1350 | { |
1351 | enum cand_kind kind = CAND_ADD; | |
1352 | tree base = NULL_TREE, stride = NULL_TREE, ctype = NULL_TREE; | |
5de9d3ed | 1353 | widest_int index, multiple; |
4502f5d0 | 1354 | unsigned savings = 0; |
1355 | slsr_cand_t c; | |
1356 | slsr_cand_t base_cand = base_cand_from_table (base_in); | |
1357 | ||
1168caca | 1358 | while (base_cand && !base && base_cand->kind != CAND_PHI) |
4502f5d0 | 1359 | { |
e913b5cd | 1360 | signop sign = TYPE_SIGN (TREE_TYPE (base_cand->stride)); |
4502f5d0 | 1361 | |
1362 | if (TREE_CODE (base_cand->stride) == INTEGER_CST | |
5de9d3ed | 1363 | && wi::multiple_of_p (index_in, wi::to_widest (base_cand->stride), |
cc5bf449 | 1364 | sign, &multiple)) |
4502f5d0 | 1365 | { |
1366 | /* Y = (B + i') * S, S constant, c = kS for some integer k | |
1367 | X = Y + c | |
1368 | ============================ | |
1369 | X = (B + (i'+ k)) * S | |
1370 | OR | |
1371 | Y = B + (i' * S), S constant, c = kS for some integer k | |
1372 | X = Y + c | |
1373 | ============================ | |
1374 | X = (B + (i'+ k)) * S */ | |
1375 | kind = base_cand->kind; | |
c99e471e | 1376 | base = base_cand->base_expr; |
cf8f0e63 | 1377 | index = base_cand->index + multiple; |
4502f5d0 | 1378 | stride = base_cand->stride; |
1379 | ctype = base_cand->cand_type; | |
1380 | if (has_single_use (base_in)) | |
1381 | savings = (base_cand->dead_savings | |
1382 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1383 | } | |
1384 | ||
1385 | if (base_cand->next_interp) | |
1386 | base_cand = lookup_cand (base_cand->next_interp); | |
1387 | else | |
1388 | base_cand = NULL; | |
1389 | } | |
1390 | ||
1391 | if (!base) | |
1392 | { | |
1393 | /* No interpretations had anything useful to propagate, so | |
1394 | produce X = Y + (c * 1). */ | |
1395 | kind = CAND_ADD; | |
1396 | base = base_in; | |
1397 | index = index_in; | |
1398 | stride = integer_one_node; | |
3b652cc1 | 1399 | ctype = TREE_TYPE (base_in); |
4502f5d0 | 1400 | } |
1401 | ||
1402 | c = alloc_cand_and_find_basis (kind, gs, base, index, stride, | |
1403 | ctype, savings); | |
1404 | return c; | |
1405 | } | |
1406 | ||
1407 | /* Given GS which is an add or subtract of scalar integers or pointers, | |
1408 | make at least one appropriate entry in the candidate table. */ | |
1409 | ||
1410 | static void | |
1411 | slsr_process_add (gimple gs, tree rhs1, tree rhs2, bool speed) | |
1412 | { | |
1413 | bool subtract_p = gimple_assign_rhs_code (gs) == MINUS_EXPR; | |
1414 | slsr_cand_t c = NULL, c2; | |
1415 | ||
1416 | if (TREE_CODE (rhs2) == SSA_NAME) | |
1417 | { | |
c99e471e | 1418 | /* First record an interpretation assuming RHS1 is the base expression |
4502f5d0 | 1419 | and RHS2 is the stride. But it doesn't make sense for the |
1420 | stride to be a pointer, so don't record a candidate in that case. */ | |
3b652cc1 | 1421 | if (!POINTER_TYPE_P (TREE_TYPE (rhs2))) |
4502f5d0 | 1422 | { |
1423 | c = create_add_ssa_cand (gs, rhs1, rhs2, subtract_p, speed); | |
1424 | ||
1425 | /* Add the first interpretation to the statement-candidate | |
1426 | mapping. */ | |
1427 | add_cand_for_stmt (gs, c); | |
1428 | } | |
1429 | ||
1430 | /* If the two RHS operands are identical, or this is a subtract, | |
1431 | we're done. */ | |
1432 | if (operand_equal_p (rhs1, rhs2, 0) || subtract_p) | |
1433 | return; | |
1434 | ||
1435 | /* Otherwise, record another interpretation assuming RHS2 is the | |
c99e471e | 1436 | base expression and RHS1 is the stride, again provided that the |
4502f5d0 | 1437 | stride is not a pointer. */ |
3b652cc1 | 1438 | if (!POINTER_TYPE_P (TREE_TYPE (rhs1))) |
4502f5d0 | 1439 | { |
1440 | c2 = create_add_ssa_cand (gs, rhs2, rhs1, false, speed); | |
1441 | if (c) | |
1442 | c->next_interp = c2->cand_num; | |
1443 | else | |
1444 | add_cand_for_stmt (gs, c2); | |
1445 | } | |
1446 | } | |
1447 | else | |
1448 | { | |
4502f5d0 | 1449 | /* Record an interpretation for the add-immediate. */ |
ab2c1de8 | 1450 | widest_int index = wi::to_widest (rhs2); |
4502f5d0 | 1451 | if (subtract_p) |
cf8f0e63 | 1452 | index = -index; |
4502f5d0 | 1453 | |
1454 | c = create_add_imm_cand (gs, rhs1, index, speed); | |
1455 | ||
1456 | /* Add the interpretation to the statement-candidate mapping. */ | |
1457 | add_cand_for_stmt (gs, c); | |
1458 | } | |
1459 | } | |
1460 | ||
1461 | /* Given GS which is a negate of a scalar integer, make an appropriate | |
1462 | entry in the candidate table. A negate is equivalent to a multiply | |
1463 | by -1. */ | |
1464 | ||
1465 | static void | |
1466 | slsr_process_neg (gimple gs, tree rhs1, bool speed) | |
1467 | { | |
1468 | /* Record a CAND_MULT interpretation for the multiply by -1. */ | |
1469 | slsr_cand_t c = create_mul_imm_cand (gs, rhs1, integer_minus_one_node, speed); | |
1470 | ||
1471 | /* Add the interpretation to the statement-candidate mapping. */ | |
1472 | add_cand_for_stmt (gs, c); | |
1473 | } | |
1474 | ||
38304570 | 1475 | /* Help function for legal_cast_p, operating on two trees. Checks |
1476 | whether it's allowable to cast from RHS to LHS. See legal_cast_p | |
1477 | for more details. */ | |
1478 | ||
1479 | static bool | |
1480 | legal_cast_p_1 (tree lhs, tree rhs) | |
1481 | { | |
1482 | tree lhs_type, rhs_type; | |
1483 | unsigned lhs_size, rhs_size; | |
1484 | bool lhs_wraps, rhs_wraps; | |
1485 | ||
1486 | lhs_type = TREE_TYPE (lhs); | |
1487 | rhs_type = TREE_TYPE (rhs); | |
1488 | lhs_size = TYPE_PRECISION (lhs_type); | |
1489 | rhs_size = TYPE_PRECISION (rhs_type); | |
1490 | lhs_wraps = TYPE_OVERFLOW_WRAPS (lhs_type); | |
1491 | rhs_wraps = TYPE_OVERFLOW_WRAPS (rhs_type); | |
1492 | ||
1493 | if (lhs_size < rhs_size | |
1494 | || (rhs_wraps && !lhs_wraps) | |
1495 | || (rhs_wraps && lhs_wraps && rhs_size != lhs_size)) | |
1496 | return false; | |
1497 | ||
1498 | return true; | |
1499 | } | |
1500 | ||
4502f5d0 | 1501 | /* Return TRUE if GS is a statement that defines an SSA name from |
1502 | a conversion and is legal for us to combine with an add and multiply | |
1503 | in the candidate table. For example, suppose we have: | |
1504 | ||
1505 | A = B + i; | |
1506 | C = (type) A; | |
1507 | D = C * S; | |
1508 | ||
1509 | Without the type-cast, we would create a CAND_MULT for D with base B, | |
1510 | index i, and stride S. We want to record this candidate only if it | |
1511 | is equivalent to apply the type cast following the multiply: | |
1512 | ||
1513 | A = B + i; | |
1514 | E = A * S; | |
1515 | D = (type) E; | |
1516 | ||
1517 | We will record the type with the candidate for D. This allows us | |
1518 | to use a similar previous candidate as a basis. If we have earlier seen | |
1519 | ||
1520 | A' = B + i'; | |
1521 | C' = (type) A'; | |
1522 | D' = C' * S; | |
1523 | ||
1524 | we can replace D with | |
1525 | ||
1526 | D = D' + (i - i') * S; | |
1527 | ||
1528 | But if moving the type-cast would change semantics, we mustn't do this. | |
1529 | ||
1530 | This is legitimate for casts from a non-wrapping integral type to | |
1531 | any integral type of the same or larger size. It is not legitimate | |
1532 | to convert a wrapping type to a non-wrapping type, or to a wrapping | |
1533 | type of a different size. I.e., with a wrapping type, we must | |
1534 | assume that the addition B + i could wrap, in which case performing | |
1535 | the multiply before or after one of the "illegal" type casts will | |
1536 | have different semantics. */ | |
1537 | ||
1538 | static bool | |
1539 | legal_cast_p (gimple gs, tree rhs) | |
1540 | { | |
4502f5d0 | 1541 | if (!is_gimple_assign (gs) |
1542 | || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (gs))) | |
1543 | return false; | |
1544 | ||
38304570 | 1545 | return legal_cast_p_1 (gimple_assign_lhs (gs), rhs); |
4502f5d0 | 1546 | } |
1547 | ||
1548 | /* Given GS which is a cast to a scalar integer type, determine whether | |
1549 | the cast is legal for strength reduction. If so, make at least one | |
1550 | appropriate entry in the candidate table. */ | |
1551 | ||
1552 | static void | |
1553 | slsr_process_cast (gimple gs, tree rhs1, bool speed) | |
1554 | { | |
1555 | tree lhs, ctype; | |
1556 | slsr_cand_t base_cand, c, c2; | |
1557 | unsigned savings = 0; | |
1558 | ||
1559 | if (!legal_cast_p (gs, rhs1)) | |
1560 | return; | |
1561 | ||
1562 | lhs = gimple_assign_lhs (gs); | |
1563 | base_cand = base_cand_from_table (rhs1); | |
1564 | ctype = TREE_TYPE (lhs); | |
1565 | ||
1168caca | 1566 | if (base_cand && base_cand->kind != CAND_PHI) |
4502f5d0 | 1567 | { |
1568 | while (base_cand) | |
1569 | { | |
1570 | /* Propagate all data from the base candidate except the type, | |
1571 | which comes from the cast, and the base candidate's cast, | |
1572 | which is no longer applicable. */ | |
1573 | if (has_single_use (rhs1)) | |
1574 | savings = (base_cand->dead_savings | |
1575 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1576 | ||
1577 | c = alloc_cand_and_find_basis (base_cand->kind, gs, | |
c99e471e | 1578 | base_cand->base_expr, |
4502f5d0 | 1579 | base_cand->index, base_cand->stride, |
1580 | ctype, savings); | |
1581 | if (base_cand->next_interp) | |
1582 | base_cand = lookup_cand (base_cand->next_interp); | |
1583 | else | |
1584 | base_cand = NULL; | |
1585 | } | |
1586 | } | |
1587 | else | |
1588 | { | |
1589 | /* If nothing is known about the RHS, create fresh CAND_ADD and | |
1590 | CAND_MULT interpretations: | |
1591 | ||
1592 | X = Y + (0 * 1) | |
1593 | X = (Y + 0) * 1 | |
1594 | ||
1595 | The first of these is somewhat arbitrary, but the choice of | |
1596 | 1 for the stride simplifies the logic for propagating casts | |
1597 | into their uses. */ | |
ddb1be65 | 1598 | c = alloc_cand_and_find_basis (CAND_ADD, gs, rhs1, |
e913b5cd | 1599 | 0, integer_one_node, ctype, 0); |
ddb1be65 | 1600 | c2 = alloc_cand_and_find_basis (CAND_MULT, gs, rhs1, |
e913b5cd | 1601 | 0, integer_one_node, ctype, 0); |
4502f5d0 | 1602 | c->next_interp = c2->cand_num; |
1603 | } | |
1604 | ||
1605 | /* Add the first (or only) interpretation to the statement-candidate | |
1606 | mapping. */ | |
1607 | add_cand_for_stmt (gs, c); | |
1608 | } | |
1609 | ||
1610 | /* Given GS which is a copy of a scalar integer type, make at least one | |
1611 | appropriate entry in the candidate table. | |
1612 | ||
1613 | This interface is included for completeness, but is unnecessary | |
1614 | if this pass immediately follows a pass that performs copy | |
1615 | propagation, such as DOM. */ | |
1616 | ||
1617 | static void | |
1618 | slsr_process_copy (gimple gs, tree rhs1, bool speed) | |
1619 | { | |
1620 | slsr_cand_t base_cand, c, c2; | |
1621 | unsigned savings = 0; | |
1622 | ||
1623 | base_cand = base_cand_from_table (rhs1); | |
1624 | ||
1168caca | 1625 | if (base_cand && base_cand->kind != CAND_PHI) |
4502f5d0 | 1626 | { |
1627 | while (base_cand) | |
1628 | { | |
1629 | /* Propagate all data from the base candidate. */ | |
1630 | if (has_single_use (rhs1)) | |
1631 | savings = (base_cand->dead_savings | |
1632 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1633 | ||
1634 | c = alloc_cand_and_find_basis (base_cand->kind, gs, | |
c99e471e | 1635 | base_cand->base_expr, |
4502f5d0 | 1636 | base_cand->index, base_cand->stride, |
1637 | base_cand->cand_type, savings); | |
1638 | if (base_cand->next_interp) | |
1639 | base_cand = lookup_cand (base_cand->next_interp); | |
1640 | else | |
1641 | base_cand = NULL; | |
1642 | } | |
1643 | } | |
1644 | else | |
1645 | { | |
1646 | /* If nothing is known about the RHS, create fresh CAND_ADD and | |
1647 | CAND_MULT interpretations: | |
1648 | ||
1649 | X = Y + (0 * 1) | |
1650 | X = (Y + 0) * 1 | |
1651 | ||
1652 | The first of these is somewhat arbitrary, but the choice of | |
1653 | 1 for the stride simplifies the logic for propagating casts | |
1654 | into their uses. */ | |
ddb1be65 | 1655 | c = alloc_cand_and_find_basis (CAND_ADD, gs, rhs1, |
e913b5cd | 1656 | 0, integer_one_node, TREE_TYPE (rhs1), 0); |
ddb1be65 | 1657 | c2 = alloc_cand_and_find_basis (CAND_MULT, gs, rhs1, |
e913b5cd | 1658 | 0, integer_one_node, TREE_TYPE (rhs1), 0); |
4502f5d0 | 1659 | c->next_interp = c2->cand_num; |
1660 | } | |
1661 | ||
1662 | /* Add the first (or only) interpretation to the statement-candidate | |
1663 | mapping. */ | |
1664 | add_cand_for_stmt (gs, c); | |
1665 | } | |
1666 | \f | |
54c91640 | 1667 | class find_candidates_dom_walker : public dom_walker |
1668 | { | |
1669 | public: | |
1670 | find_candidates_dom_walker (cdi_direction direction) | |
1671 | : dom_walker (direction) {} | |
1672 | virtual void before_dom_children (basic_block); | |
1673 | }; | |
1674 | ||
4502f5d0 | 1675 | /* Find strength-reduction candidates in block BB. */ |
1676 | ||
54c91640 | 1677 | void |
1678 | find_candidates_dom_walker::before_dom_children (basic_block bb) | |
4502f5d0 | 1679 | { |
1680 | bool speed = optimize_bb_for_speed_p (bb); | |
1681 | gimple_stmt_iterator gsi; | |
1682 | ||
1168caca | 1683 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
1684 | slsr_process_phi (gsi_stmt (gsi), speed); | |
1685 | ||
4502f5d0 | 1686 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
1687 | { | |
1688 | gimple gs = gsi_stmt (gsi); | |
1689 | ||
070bf980 | 1690 | if (gimple_vuse (gs) && gimple_assign_single_p (gs)) |
1691 | slsr_process_ref (gs); | |
1692 | ||
1693 | else if (is_gimple_assign (gs) | |
1694 | && SCALAR_INT_MODE_P | |
1695 | (TYPE_MODE (TREE_TYPE (gimple_assign_lhs (gs))))) | |
4502f5d0 | 1696 | { |
1697 | tree rhs1 = NULL_TREE, rhs2 = NULL_TREE; | |
1698 | ||
1699 | switch (gimple_assign_rhs_code (gs)) | |
1700 | { | |
1701 | case MULT_EXPR: | |
1702 | case PLUS_EXPR: | |
1703 | rhs1 = gimple_assign_rhs1 (gs); | |
1704 | rhs2 = gimple_assign_rhs2 (gs); | |
1705 | /* Should never happen, but currently some buggy situations | |
1706 | in earlier phases put constants in rhs1. */ | |
1707 | if (TREE_CODE (rhs1) != SSA_NAME) | |
1708 | continue; | |
1709 | break; | |
1710 | ||
1711 | /* Possible future opportunity: rhs1 of a ptr+ can be | |
1712 | an ADDR_EXPR. */ | |
1713 | case POINTER_PLUS_EXPR: | |
1714 | case MINUS_EXPR: | |
1715 | rhs2 = gimple_assign_rhs2 (gs); | |
1716 | /* Fall-through. */ | |
1717 | ||
1718 | case NOP_EXPR: | |
1719 | case MODIFY_EXPR: | |
1720 | case NEGATE_EXPR: | |
1721 | rhs1 = gimple_assign_rhs1 (gs); | |
1722 | if (TREE_CODE (rhs1) != SSA_NAME) | |
1723 | continue; | |
1724 | break; | |
1725 | ||
1726 | default: | |
1727 | ; | |
1728 | } | |
1729 | ||
1730 | switch (gimple_assign_rhs_code (gs)) | |
1731 | { | |
1732 | case MULT_EXPR: | |
1733 | slsr_process_mul (gs, rhs1, rhs2, speed); | |
1734 | break; | |
1735 | ||
1736 | case PLUS_EXPR: | |
1737 | case POINTER_PLUS_EXPR: | |
1738 | case MINUS_EXPR: | |
1739 | slsr_process_add (gs, rhs1, rhs2, speed); | |
1740 | break; | |
1741 | ||
1742 | case NEGATE_EXPR: | |
1743 | slsr_process_neg (gs, rhs1, speed); | |
1744 | break; | |
1745 | ||
1746 | case NOP_EXPR: | |
1747 | slsr_process_cast (gs, rhs1, speed); | |
1748 | break; | |
1749 | ||
1750 | case MODIFY_EXPR: | |
1751 | slsr_process_copy (gs, rhs1, speed); | |
1752 | break; | |
1753 | ||
1754 | default: | |
1755 | ; | |
1756 | } | |
1757 | } | |
1758 | } | |
1759 | } | |
1760 | \f | |
1761 | /* Dump a candidate for debug. */ | |
1762 | ||
1763 | static void | |
1764 | dump_candidate (slsr_cand_t c) | |
1765 | { | |
1766 | fprintf (dump_file, "%3d [%d] ", c->cand_num, | |
1767 | gimple_bb (c->cand_stmt)->index); | |
1768 | print_gimple_stmt (dump_file, c->cand_stmt, 0, 0); | |
1769 | switch (c->kind) | |
1770 | { | |
1771 | case CAND_MULT: | |
1772 | fputs (" MULT : (", dump_file); | |
c99e471e | 1773 | print_generic_expr (dump_file, c->base_expr, 0); |
4502f5d0 | 1774 | fputs (" + ", dump_file); |
e913b5cd | 1775 | print_decs (c->index, dump_file); |
4502f5d0 | 1776 | fputs (") * ", dump_file); |
1777 | print_generic_expr (dump_file, c->stride, 0); | |
1778 | fputs (" : ", dump_file); | |
1779 | break; | |
1780 | case CAND_ADD: | |
1781 | fputs (" ADD : ", dump_file); | |
c99e471e | 1782 | print_generic_expr (dump_file, c->base_expr, 0); |
4502f5d0 | 1783 | fputs (" + (", dump_file); |
e913b5cd | 1784 | print_decs (c->index, dump_file); |
4502f5d0 | 1785 | fputs (" * ", dump_file); |
1786 | print_generic_expr (dump_file, c->stride, 0); | |
1787 | fputs (") : ", dump_file); | |
1788 | break; | |
070bf980 | 1789 | case CAND_REF: |
1790 | fputs (" REF : ", dump_file); | |
c99e471e | 1791 | print_generic_expr (dump_file, c->base_expr, 0); |
070bf980 | 1792 | fputs (" + (", dump_file); |
1793 | print_generic_expr (dump_file, c->stride, 0); | |
1794 | fputs (") + ", dump_file); | |
e913b5cd | 1795 | print_decs (c->index, dump_file); |
070bf980 | 1796 | fputs (" : ", dump_file); |
1797 | break; | |
1168caca | 1798 | case CAND_PHI: |
1799 | fputs (" PHI : ", dump_file); | |
1800 | print_generic_expr (dump_file, c->base_expr, 0); | |
1801 | fputs (" + (unknown * ", dump_file); | |
1802 | print_generic_expr (dump_file, c->stride, 0); | |
1803 | fputs (") : ", dump_file); | |
1804 | break; | |
4502f5d0 | 1805 | default: |
1806 | gcc_unreachable (); | |
1807 | } | |
1808 | print_generic_expr (dump_file, c->cand_type, 0); | |
1809 | fprintf (dump_file, "\n basis: %d dependent: %d sibling: %d\n", | |
1810 | c->basis, c->dependent, c->sibling); | |
1811 | fprintf (dump_file, " next-interp: %d dead-savings: %d\n", | |
1812 | c->next_interp, c->dead_savings); | |
1813 | if (c->def_phi) | |
1168caca | 1814 | fprintf (dump_file, " phi: %d\n", c->def_phi); |
4502f5d0 | 1815 | fputs ("\n", dump_file); |
1816 | } | |
1817 | ||
1818 | /* Dump the candidate vector for debug. */ | |
1819 | ||
1820 | static void | |
1821 | dump_cand_vec (void) | |
1822 | { | |
1823 | unsigned i; | |
1824 | slsr_cand_t c; | |
1825 | ||
1826 | fprintf (dump_file, "\nStrength reduction candidate vector:\n\n"); | |
1827 | ||
f1f41a6c | 1828 | FOR_EACH_VEC_ELT (cand_vec, i, c) |
4502f5d0 | 1829 | dump_candidate (c); |
1830 | } | |
1831 | ||
070bf980 | 1832 | /* Callback used to dump the candidate chains hash table. */ |
4502f5d0 | 1833 | |
d9dd21a8 | 1834 | int |
1835 | ssa_base_cand_dump_callback (cand_chain **slot, void *ignored ATTRIBUTE_UNUSED) | |
4502f5d0 | 1836 | { |
d9dd21a8 | 1837 | const_cand_chain_t chain = *slot; |
070bf980 | 1838 | cand_chain_t p; |
4502f5d0 | 1839 | |
c99e471e | 1840 | print_generic_expr (dump_file, chain->base_expr, 0); |
070bf980 | 1841 | fprintf (dump_file, " -> %d", chain->cand->cand_num); |
4502f5d0 | 1842 | |
070bf980 | 1843 | for (p = chain->next; p; p = p->next) |
1844 | fprintf (dump_file, " -> %d", p->cand->cand_num); | |
4502f5d0 | 1845 | |
070bf980 | 1846 | fputs ("\n", dump_file); |
1847 | return 1; | |
1848 | } | |
4502f5d0 | 1849 | |
070bf980 | 1850 | /* Dump the candidate chains. */ |
4502f5d0 | 1851 | |
070bf980 | 1852 | static void |
1853 | dump_cand_chains (void) | |
1854 | { | |
1855 | fprintf (dump_file, "\nStrength reduction candidate chains:\n\n"); | |
c1f445d2 | 1856 | base_cand_map->traverse_noresize <void *, ssa_base_cand_dump_callback> |
1857 | (NULL); | |
4502f5d0 | 1858 | fputs ("\n", dump_file); |
1859 | } | |
9ce0359e | 1860 | |
1861 | /* Dump the increment vector for debug. */ | |
1862 | ||
1863 | static void | |
1864 | dump_incr_vec (void) | |
1865 | { | |
1866 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1867 | { | |
1868 | unsigned i; | |
1869 | ||
1870 | fprintf (dump_file, "\nIncrement vector:\n\n"); | |
1871 | ||
1872 | for (i = 0; i < incr_vec_len; i++) | |
1873 | { | |
1874 | fprintf (dump_file, "%3d increment: ", i); | |
e913b5cd | 1875 | print_decs (incr_vec[i].incr, dump_file); |
9ce0359e | 1876 | fprintf (dump_file, "\n count: %d", incr_vec[i].count); |
1877 | fprintf (dump_file, "\n cost: %d", incr_vec[i].cost); | |
1878 | fputs ("\n initializer: ", dump_file); | |
1879 | print_generic_expr (dump_file, incr_vec[i].initializer, 0); | |
1880 | fputs ("\n\n", dump_file); | |
1881 | } | |
1882 | } | |
1883 | } | |
4502f5d0 | 1884 | \f |
070bf980 | 1885 | /* Replace *EXPR in candidate C with an equivalent strength-reduced |
1886 | data reference. */ | |
1887 | ||
1888 | static void | |
1889 | replace_ref (tree *expr, slsr_cand_t c) | |
1890 | { | |
dcf9d898 | 1891 | tree add_expr, mem_ref, acc_type = TREE_TYPE (*expr); |
1892 | unsigned HOST_WIDE_INT misalign; | |
1893 | unsigned align; | |
1894 | ||
1895 | /* Ensure the memory reference carries the minimum alignment | |
1896 | requirement for the data type. See PR58041. */ | |
1897 | get_object_alignment_1 (*expr, &align, &misalign); | |
1898 | if (misalign != 0) | |
1899 | align = (misalign & -misalign); | |
1900 | if (align < TYPE_ALIGN (acc_type)) | |
1901 | acc_type = build_aligned_type (acc_type, align); | |
1902 | ||
1903 | add_expr = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (c->base_expr), | |
1904 | c->base_expr, c->stride); | |
1905 | mem_ref = fold_build2 (MEM_REF, acc_type, add_expr, | |
1906 | wide_int_to_tree (c->cand_type, c->index)); | |
072a243c | 1907 | |
070bf980 | 1908 | /* Gimplify the base addressing expression for the new MEM_REF tree. */ |
1909 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); | |
1910 | TREE_OPERAND (mem_ref, 0) | |
1911 | = force_gimple_operand_gsi (&gsi, TREE_OPERAND (mem_ref, 0), | |
1912 | /*simple_p=*/true, NULL, | |
1913 | /*before=*/true, GSI_SAME_STMT); | |
1914 | copy_ref_info (mem_ref, *expr); | |
1915 | *expr = mem_ref; | |
1916 | update_stmt (c->cand_stmt); | |
1917 | } | |
1918 | ||
1919 | /* Replace CAND_REF candidate C, each sibling of candidate C, and each | |
1920 | dependent of candidate C with an equivalent strength-reduced data | |
1921 | reference. */ | |
1922 | ||
1923 | static void | |
1924 | replace_refs (slsr_cand_t c) | |
1925 | { | |
0ca43779 | 1926 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1927 | { | |
1928 | fputs ("Replacing reference: ", dump_file); | |
1929 | print_gimple_stmt (dump_file, c->cand_stmt, 0, 0); | |
1930 | } | |
1931 | ||
070bf980 | 1932 | if (gimple_vdef (c->cand_stmt)) |
1933 | { | |
1934 | tree *lhs = gimple_assign_lhs_ptr (c->cand_stmt); | |
1935 | replace_ref (lhs, c); | |
1936 | } | |
1937 | else | |
1938 | { | |
1939 | tree *rhs = gimple_assign_rhs1_ptr (c->cand_stmt); | |
1940 | replace_ref (rhs, c); | |
1941 | } | |
1942 | ||
0ca43779 | 1943 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1944 | { | |
1945 | fputs ("With: ", dump_file); | |
1946 | print_gimple_stmt (dump_file, c->cand_stmt, 0, 0); | |
1947 | fputs ("\n", dump_file); | |
1948 | } | |
1949 | ||
070bf980 | 1950 | if (c->sibling) |
1951 | replace_refs (lookup_cand (c->sibling)); | |
1952 | ||
1953 | if (c->dependent) | |
1954 | replace_refs (lookup_cand (c->dependent)); | |
1955 | } | |
1956 | ||
1168caca | 1957 | /* Return TRUE if candidate C is dependent upon a PHI. */ |
1958 | ||
1959 | static bool | |
1960 | phi_dependent_cand_p (slsr_cand_t c) | |
1961 | { | |
1962 | /* A candidate is not necessarily dependent upon a PHI just because | |
1963 | it has a phi definition for its base name. It may have a basis | |
1964 | that relies upon the same phi definition, in which case the PHI | |
1965 | is irrelevant to this candidate. */ | |
1966 | return (c->def_phi | |
1967 | && c->basis | |
1968 | && lookup_cand (c->basis)->def_phi != c->def_phi); | |
1969 | } | |
1970 | ||
4502f5d0 | 1971 | /* Calculate the increment required for candidate C relative to |
1972 | its basis. */ | |
1973 | ||
5de9d3ed | 1974 | static widest_int |
4502f5d0 | 1975 | cand_increment (slsr_cand_t c) |
1976 | { | |
1977 | slsr_cand_t basis; | |
1978 | ||
1979 | /* If the candidate doesn't have a basis, just return its own | |
1980 | index. This is useful in record_increments to help us find | |
1168caca | 1981 | an existing initializer. Also, if the candidate's basis is |
1982 | hidden by a phi, then its own index will be the increment | |
1983 | from the newly introduced phi basis. */ | |
1984 | if (!c->basis || phi_dependent_cand_p (c)) | |
4502f5d0 | 1985 | return c->index; |
1986 | ||
1987 | basis = lookup_cand (c->basis); | |
c99e471e | 1988 | gcc_assert (operand_equal_p (c->base_expr, basis->base_expr, 0)); |
cf8f0e63 | 1989 | return c->index - basis->index; |
4502f5d0 | 1990 | } |
1991 | ||
9ce0359e | 1992 | /* Calculate the increment required for candidate C relative to |
1993 | its basis. If we aren't going to generate pointer arithmetic | |
1994 | for this candidate, return the absolute value of that increment | |
1995 | instead. */ | |
1996 | ||
5de9d3ed | 1997 | static inline widest_int |
9ce0359e | 1998 | cand_abs_increment (slsr_cand_t c) |
1999 | { | |
5de9d3ed | 2000 | widest_int increment = cand_increment (c); |
9ce0359e | 2001 | |
796b6678 | 2002 | if (!address_arithmetic_p && wi::neg_p (increment)) |
cf8f0e63 | 2003 | increment = -increment; |
9ce0359e | 2004 | |
2005 | return increment; | |
2006 | } | |
2007 | ||
4502f5d0 | 2008 | /* Return TRUE iff candidate C has already been replaced under |
2009 | another interpretation. */ | |
2010 | ||
2011 | static inline bool | |
2012 | cand_already_replaced (slsr_cand_t c) | |
2013 | { | |
2014 | return (gimple_bb (c->cand_stmt) == 0); | |
2015 | } | |
2016 | ||
1168caca | 2017 | /* Common logic used by replace_unconditional_candidate and |
2018 | replace_conditional_candidate. */ | |
4502f5d0 | 2019 | |
2020 | static void | |
ab2c1de8 | 2021 | replace_mult_candidate (slsr_cand_t c, tree basis_name, widest_int bump) |
4502f5d0 | 2022 | { |
1168caca | 2023 | tree target_type = TREE_TYPE (gimple_assign_lhs (c->cand_stmt)); |
2024 | enum tree_code cand_code = gimple_assign_rhs_code (c->cand_stmt); | |
2025 | ||
4502f5d0 | 2026 | /* It is highly unlikely, but possible, that the resulting |
2027 | bump doesn't fit in a HWI. Abandon the replacement | |
1168caca | 2028 | in this case. This does not affect siblings or dependents |
2029 | of C. Restriction to signed HWI is conservative for unsigned | |
2030 | types but allows for safe negation without twisted logic. */ | |
796b6678 | 2031 | if (wi::fits_shwi_p (bump) |
1168caca | 2032 | && bump.to_shwi () != HOST_WIDE_INT_MIN |
2033 | /* It is not useful to replace casts, copies, or adds of | |
2034 | an SSA name and a constant. */ | |
2035 | && cand_code != MODIFY_EXPR | |
2036 | && cand_code != NOP_EXPR | |
2037 | && cand_code != PLUS_EXPR | |
2038 | && cand_code != POINTER_PLUS_EXPR | |
2039 | && cand_code != MINUS_EXPR) | |
2040 | { | |
2041 | enum tree_code code = PLUS_EXPR; | |
2042 | tree bump_tree; | |
2043 | gimple stmt_to_print = NULL; | |
2044 | ||
2045 | /* If the basis name and the candidate's LHS have incompatible | |
2046 | types, introduce a cast. */ | |
2047 | if (!useless_type_conversion_p (target_type, TREE_TYPE (basis_name))) | |
d54ab372 | 2048 | basis_name = introduce_cast_before_cand (c, target_type, basis_name); |
796b6678 | 2049 | if (wi::neg_p (bump)) |
1168caca | 2050 | { |
2051 | code = MINUS_EXPR; | |
2052 | bump = -bump; | |
2053 | } | |
2054 | ||
e913b5cd | 2055 | bump_tree = wide_int_to_tree (target_type, bump); |
1168caca | 2056 | |
2057 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2058 | { | |
2059 | fputs ("Replacing: ", dump_file); | |
2060 | print_gimple_stmt (dump_file, c->cand_stmt, 0, 0); | |
2061 | } | |
2062 | ||
796b6678 | 2063 | if (bump == 0) |
1168caca | 2064 | { |
2065 | tree lhs = gimple_assign_lhs (c->cand_stmt); | |
2066 | gimple copy_stmt = gimple_build_assign (lhs, basis_name); | |
2067 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); | |
2068 | gimple_set_location (copy_stmt, gimple_location (c->cand_stmt)); | |
2069 | gsi_replace (&gsi, copy_stmt, false); | |
8f0676e4 | 2070 | c->cand_stmt = copy_stmt; |
1168caca | 2071 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2072 | stmt_to_print = copy_stmt; | |
2073 | } | |
2074 | else | |
2075 | { | |
2076 | tree rhs1, rhs2; | |
2077 | if (cand_code != NEGATE_EXPR) { | |
2078 | rhs1 = gimple_assign_rhs1 (c->cand_stmt); | |
2079 | rhs2 = gimple_assign_rhs2 (c->cand_stmt); | |
2080 | } | |
2081 | if (cand_code != NEGATE_EXPR | |
2082 | && ((operand_equal_p (rhs1, basis_name, 0) | |
2083 | && operand_equal_p (rhs2, bump_tree, 0)) | |
2084 | || (operand_equal_p (rhs1, bump_tree, 0) | |
2085 | && operand_equal_p (rhs2, basis_name, 0)))) | |
2086 | { | |
2087 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2088 | { | |
2089 | fputs ("(duplicate, not actually replacing)", dump_file); | |
2090 | stmt_to_print = c->cand_stmt; | |
2091 | } | |
2092 | } | |
2093 | else | |
2094 | { | |
2095 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); | |
2096 | gimple_assign_set_rhs_with_ops (&gsi, code, | |
2097 | basis_name, bump_tree); | |
2098 | update_stmt (gsi_stmt (gsi)); | |
dc77dea1 | 2099 | c->cand_stmt = gsi_stmt (gsi); |
1168caca | 2100 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2101 | stmt_to_print = gsi_stmt (gsi); | |
2102 | } | |
2103 | } | |
2104 | ||
2105 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2106 | { | |
2107 | fputs ("With: ", dump_file); | |
2108 | print_gimple_stmt (dump_file, stmt_to_print, 0, 0); | |
2109 | fputs ("\n", dump_file); | |
2110 | } | |
2111 | } | |
2112 | } | |
2113 | ||
2114 | /* Replace candidate C with an add or subtract. Note that we only | |
2115 | operate on CAND_MULTs with known strides, so we will never generate | |
2116 | a POINTER_PLUS_EXPR. Each candidate X = (B + i) * S is replaced by | |
2117 | X = Y + ((i - i') * S), as described in the module commentary. The | |
2118 | folded value ((i - i') * S) is referred to here as the "bump." */ | |
2119 | ||
2120 | static void | |
2121 | replace_unconditional_candidate (slsr_cand_t c) | |
2122 | { | |
2123 | slsr_cand_t basis; | |
1168caca | 2124 | |
2125 | if (cand_already_replaced (c)) | |
4502f5d0 | 2126 | return; |
2127 | ||
2128 | basis = lookup_cand (c->basis); | |
ab2c1de8 | 2129 | widest_int bump = cand_increment (c) * wi::to_widest (c->stride); |
1168caca | 2130 | |
d54ab372 | 2131 | replace_mult_candidate (c, gimple_assign_lhs (basis->cand_stmt), bump); |
1168caca | 2132 | } |
2133 | \f | |
d62a8480 | 2134 | /* Return the index in the increment vector of the given INCREMENT, |
2135 | or -1 if not found. The latter can occur if more than | |
2136 | MAX_INCR_VEC_LEN increments have been found. */ | |
1168caca | 2137 | |
d62a8480 | 2138 | static inline int |
10c3fe8d | 2139 | incr_vec_index (const widest_int &increment) |
1168caca | 2140 | { |
2141 | unsigned i; | |
2142 | ||
2143 | for (i = 0; i < incr_vec_len && increment != incr_vec[i].incr; i++) | |
2144 | ; | |
2145 | ||
d62a8480 | 2146 | if (i < incr_vec_len) |
2147 | return i; | |
2148 | else | |
2149 | return -1; | |
1168caca | 2150 | } |
2151 | ||
2152 | /* Create a new statement along edge E to add BASIS_NAME to the product | |
2153 | of INCREMENT and the stride of candidate C. Create and return a new | |
2154 | SSA name from *VAR to be used as the LHS of the new statement. | |
2155 | KNOWN_STRIDE is true iff C's stride is a constant. */ | |
2156 | ||
2157 | static tree | |
2158 | create_add_on_incoming_edge (slsr_cand_t c, tree basis_name, | |
5de9d3ed | 2159 | widest_int increment, edge e, location_t loc, |
1168caca | 2160 | bool known_stride) |
2161 | { | |
2162 | basic_block insert_bb; | |
2163 | gimple_stmt_iterator gsi; | |
2164 | tree lhs, basis_type; | |
2165 | gimple new_stmt; | |
2166 | ||
2167 | /* If the add candidate along this incoming edge has the same | |
2168 | index as C's hidden basis, the hidden basis represents this | |
2169 | edge correctly. */ | |
796b6678 | 2170 | if (increment == 0) |
1168caca | 2171 | return basis_name; |
2172 | ||
2173 | basis_type = TREE_TYPE (basis_name); | |
2174 | lhs = make_temp_ssa_name (basis_type, NULL, "slsr"); | |
2175 | ||
2176 | if (known_stride) | |
8e6331c6 | 2177 | { |
1168caca | 2178 | tree bump_tree; |
2179 | enum tree_code code = PLUS_EXPR; | |
5de9d3ed | 2180 | widest_int bump = increment * wi::to_widest (c->stride); |
796b6678 | 2181 | if (wi::neg_p (bump)) |
1168caca | 2182 | { |
2183 | code = MINUS_EXPR; | |
2184 | bump = -bump; | |
2185 | } | |
2186 | ||
e913b5cd | 2187 | bump_tree = wide_int_to_tree (basis_type, bump); |
1168caca | 2188 | new_stmt = gimple_build_assign_with_ops (code, lhs, basis_name, |
2189 | bump_tree); | |
8e6331c6 | 2190 | } |
2191 | else | |
2192 | { | |
d62a8480 | 2193 | int i; |
796b6678 | 2194 | bool negate_incr = (!address_arithmetic_p && wi::neg_p (increment)); |
1168caca | 2195 | i = incr_vec_index (negate_incr ? -increment : increment); |
d62a8480 | 2196 | gcc_assert (i >= 0); |
4502f5d0 | 2197 | |
1168caca | 2198 | if (incr_vec[i].initializer) |
2199 | { | |
2200 | enum tree_code code = negate_incr ? MINUS_EXPR : PLUS_EXPR; | |
2201 | new_stmt = gimple_build_assign_with_ops (code, lhs, basis_name, | |
2202 | incr_vec[i].initializer); | |
2203 | } | |
796b6678 | 2204 | else if (increment == 1) |
1168caca | 2205 | new_stmt = gimple_build_assign_with_ops (PLUS_EXPR, lhs, basis_name, |
2206 | c->stride); | |
796b6678 | 2207 | else if (increment == -1) |
1168caca | 2208 | new_stmt = gimple_build_assign_with_ops (MINUS_EXPR, lhs, basis_name, |
2209 | c->stride); | |
2210 | else | |
2211 | gcc_unreachable (); | |
4502f5d0 | 2212 | } |
2213 | ||
1168caca | 2214 | insert_bb = single_succ_p (e->src) ? e->src : split_edge (e); |
2215 | gsi = gsi_last_bb (insert_bb); | |
2216 | ||
2217 | if (!gsi_end_p (gsi) && is_ctrl_stmt (gsi_stmt (gsi))) | |
2218 | gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); | |
2219 | else | |
2220 | gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT); | |
2221 | ||
2222 | gimple_set_location (new_stmt, loc); | |
4502f5d0 | 2223 | |
2224 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2225 | { | |
1168caca | 2226 | fprintf (dump_file, "Inserting in block %d: ", insert_bb->index); |
2227 | print_gimple_stmt (dump_file, new_stmt, 0, 0); | |
4502f5d0 | 2228 | } |
2229 | ||
1168caca | 2230 | return lhs; |
2231 | } | |
2232 | ||
2233 | /* Given a candidate C with BASIS_NAME being the LHS of C's basis which | |
2234 | is hidden by the phi node FROM_PHI, create a new phi node in the same | |
2235 | block as FROM_PHI. The new phi is suitable for use as a basis by C, | |
2236 | with its phi arguments representing conditional adjustments to the | |
2237 | hidden basis along conditional incoming paths. Those adjustments are | |
2238 | made by creating add statements (and sometimes recursively creating | |
2239 | phis) along those incoming paths. LOC is the location to attach to | |
2240 | the introduced statements. KNOWN_STRIDE is true iff C's stride is a | |
2241 | constant. */ | |
2242 | ||
2243 | static tree | |
2244 | create_phi_basis (slsr_cand_t c, gimple from_phi, tree basis_name, | |
2245 | location_t loc, bool known_stride) | |
2246 | { | |
2247 | int i; | |
2248 | tree name, phi_arg; | |
2249 | gimple phi; | |
2250 | vec<tree> phi_args; | |
2251 | slsr_cand_t basis = lookup_cand (c->basis); | |
2252 | int nargs = gimple_phi_num_args (from_phi); | |
2253 | basic_block phi_bb = gimple_bb (from_phi); | |
2254 | slsr_cand_t phi_cand = base_cand_from_table (gimple_phi_result (from_phi)); | |
2255 | phi_args.create (nargs); | |
2256 | ||
2257 | /* Process each argument of the existing phi that represents | |
2258 | conditionally-executed add candidates. */ | |
2259 | for (i = 0; i < nargs; i++) | |
4502f5d0 | 2260 | { |
1168caca | 2261 | edge e = (*phi_bb->preds)[i]; |
2262 | tree arg = gimple_phi_arg_def (from_phi, i); | |
2263 | tree feeding_def; | |
2264 | ||
2265 | /* If the phi argument is the base name of the CAND_PHI, then | |
2266 | this incoming arc should use the hidden basis. */ | |
2267 | if (operand_equal_p (arg, phi_cand->base_expr, 0)) | |
796b6678 | 2268 | if (basis->index == 0) |
1168caca | 2269 | feeding_def = gimple_assign_lhs (basis->cand_stmt); |
2270 | else | |
2271 | { | |
5de9d3ed | 2272 | widest_int incr = -basis->index; |
1168caca | 2273 | feeding_def = create_add_on_incoming_edge (c, basis_name, incr, |
2274 | e, loc, known_stride); | |
2275 | } | |
2276 | else | |
4502f5d0 | 2277 | { |
1168caca | 2278 | gimple arg_def = SSA_NAME_DEF_STMT (arg); |
2279 | ||
2280 | /* If there is another phi along this incoming edge, we must | |
2281 | process it in the same fashion to ensure that all basis | |
2282 | adjustments are made along its incoming edges. */ | |
2283 | if (gimple_code (arg_def) == GIMPLE_PHI) | |
2284 | feeding_def = create_phi_basis (c, arg_def, basis_name, | |
2285 | loc, known_stride); | |
2286 | else | |
4502f5d0 | 2287 | { |
1168caca | 2288 | slsr_cand_t arg_cand = base_cand_from_table (arg); |
5de9d3ed | 2289 | widest_int diff = arg_cand->index - basis->index; |
1168caca | 2290 | feeding_def = create_add_on_incoming_edge (c, basis_name, diff, |
2291 | e, loc, known_stride); | |
4502f5d0 | 2292 | } |
2293 | } | |
1168caca | 2294 | |
2295 | /* Because of recursion, we need to save the arguments in a vector | |
2296 | so we can create the PHI statement all at once. Otherwise the | |
2297 | storage for the half-created PHI can be reclaimed. */ | |
2298 | phi_args.safe_push (feeding_def); | |
4502f5d0 | 2299 | } |
1168caca | 2300 | |
2301 | /* Create the new phi basis. */ | |
2302 | name = make_temp_ssa_name (TREE_TYPE (basis_name), NULL, "slsr"); | |
2303 | phi = create_phi_node (name, phi_bb); | |
2304 | SSA_NAME_DEF_STMT (name) = phi; | |
2305 | ||
2306 | FOR_EACH_VEC_ELT (phi_args, i, phi_arg) | |
2307 | { | |
2308 | edge e = (*phi_bb->preds)[i]; | |
2309 | add_phi_arg (phi, phi_arg, e, loc); | |
2310 | } | |
2311 | ||
2312 | update_stmt (phi); | |
2313 | ||
4502f5d0 | 2314 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2315 | { | |
1168caca | 2316 | fputs ("Introducing new phi basis: ", dump_file); |
2317 | print_gimple_stmt (dump_file, phi, 0, 0); | |
4502f5d0 | 2318 | } |
1168caca | 2319 | |
2320 | return name; | |
4502f5d0 | 2321 | } |
2322 | ||
1168caca | 2323 | /* Given a candidate C whose basis is hidden by at least one intervening |
2324 | phi, introduce a matching number of new phis to represent its basis | |
2325 | adjusted by conditional increments along possible incoming paths. Then | |
2326 | replace C as though it were an unconditional candidate, using the new | |
2327 | basis. */ | |
4502f5d0 | 2328 | |
2329 | static void | |
1168caca | 2330 | replace_conditional_candidate (slsr_cand_t c) |
4502f5d0 | 2331 | { |
d54ab372 | 2332 | tree basis_name, name; |
1168caca | 2333 | slsr_cand_t basis; |
2334 | location_t loc; | |
4502f5d0 | 2335 | |
1168caca | 2336 | /* Look up the LHS SSA name from C's basis. This will be the |
2337 | RHS1 of the adds we will introduce to create new phi arguments. */ | |
2338 | basis = lookup_cand (c->basis); | |
2339 | basis_name = gimple_assign_lhs (basis->cand_stmt); | |
4502f5d0 | 2340 | |
1168caca | 2341 | /* Create a new phi statement which will represent C's true basis |
2342 | after the transformation is complete. */ | |
2343 | loc = gimple_location (c->cand_stmt); | |
2344 | name = create_phi_basis (c, lookup_cand (c->def_phi)->cand_stmt, | |
2345 | basis_name, loc, KNOWN_STRIDE); | |
2346 | /* Replace C with an add of the new basis phi and a constant. */ | |
ab2c1de8 | 2347 | widest_int bump = c->index * wi::to_widest (c->stride); |
4502f5d0 | 2348 | |
d54ab372 | 2349 | replace_mult_candidate (c, name, bump); |
4502f5d0 | 2350 | } |
9ce0359e | 2351 | |
1168caca | 2352 | /* Compute the expected costs of inserting basis adjustments for |
2353 | candidate C with phi-definition PHI. The cost of inserting | |
2354 | one adjustment is given by ONE_ADD_COST. If PHI has arguments | |
2355 | which are themselves phi results, recursively calculate costs | |
2356 | for those phis as well. */ | |
2357 | ||
2358 | static int | |
2359 | phi_add_costs (gimple phi, slsr_cand_t c, int one_add_cost) | |
9ce0359e | 2360 | { |
2361 | unsigned i; | |
1168caca | 2362 | int cost = 0; |
2363 | slsr_cand_t phi_cand = base_cand_from_table (gimple_phi_result (phi)); | |
9ce0359e | 2364 | |
8f0676e4 | 2365 | /* If we work our way back to a phi that isn't dominated by the hidden |
2366 | basis, this isn't a candidate for replacement. Indicate this by | |
2367 | returning an unreasonably high cost. It's not easy to detect | |
2368 | these situations when determining the basis, so we defer the | |
2369 | decision until now. */ | |
2370 | basic_block phi_bb = gimple_bb (phi); | |
2371 | slsr_cand_t basis = lookup_cand (c->basis); | |
2372 | basic_block basis_bb = gimple_bb (basis->cand_stmt); | |
2373 | ||
2374 | if (phi_bb == basis_bb || !dominated_by_p (CDI_DOMINATORS, phi_bb, basis_bb)) | |
2375 | return COST_INFINITE; | |
2376 | ||
1168caca | 2377 | for (i = 0; i < gimple_phi_num_args (phi); i++) |
2378 | { | |
2379 | tree arg = gimple_phi_arg_def (phi, i); | |
2380 | ||
2381 | if (arg != phi_cand->base_expr) | |
2382 | { | |
2383 | gimple arg_def = SSA_NAME_DEF_STMT (arg); | |
2384 | ||
2385 | if (gimple_code (arg_def) == GIMPLE_PHI) | |
2386 | cost += phi_add_costs (arg_def, c, one_add_cost); | |
2387 | else | |
2388 | { | |
2389 | slsr_cand_t arg_cand = base_cand_from_table (arg); | |
2390 | ||
2391 | if (arg_cand->index != c->index) | |
2392 | cost += one_add_cost; | |
2393 | } | |
2394 | } | |
2395 | } | |
2396 | ||
2397 | return cost; | |
9ce0359e | 2398 | } |
2399 | ||
1168caca | 2400 | /* For candidate C, each sibling of candidate C, and each dependent of |
2401 | candidate C, determine whether the candidate is dependent upon a | |
2402 | phi that hides its basis. If not, replace the candidate unconditionally. | |
2403 | Otherwise, determine whether the cost of introducing compensation code | |
2404 | for the candidate is offset by the gains from strength reduction. If | |
2405 | so, replace the candidate and introduce the compensation code. */ | |
2406 | ||
2407 | static void | |
2408 | replace_uncond_cands_and_profitable_phis (slsr_cand_t c) | |
2409 | { | |
2410 | if (phi_dependent_cand_p (c)) | |
2411 | { | |
2412 | if (c->kind == CAND_MULT) | |
2413 | { | |
2414 | /* A candidate dependent upon a phi will replace a multiply by | |
2415 | a constant with an add, and will insert at most one add for | |
2416 | each phi argument. Add these costs with the potential dead-code | |
2417 | savings to determine profitability. */ | |
2418 | bool speed = optimize_bb_for_speed_p (gimple_bb (c->cand_stmt)); | |
2419 | int mult_savings = stmt_cost (c->cand_stmt, speed); | |
2420 | gimple phi = lookup_cand (c->def_phi)->cand_stmt; | |
2421 | tree phi_result = gimple_phi_result (phi); | |
2422 | int one_add_cost = add_cost (speed, | |
2423 | TYPE_MODE (TREE_TYPE (phi_result))); | |
2424 | int add_costs = one_add_cost + phi_add_costs (phi, c, one_add_cost); | |
2425 | int cost = add_costs - mult_savings - c->dead_savings; | |
2426 | ||
2427 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2428 | { | |
2429 | fprintf (dump_file, " Conditional candidate %d:\n", c->cand_num); | |
2430 | fprintf (dump_file, " add_costs = %d\n", add_costs); | |
2431 | fprintf (dump_file, " mult_savings = %d\n", mult_savings); | |
2432 | fprintf (dump_file, " dead_savings = %d\n", c->dead_savings); | |
2433 | fprintf (dump_file, " cost = %d\n", cost); | |
2434 | if (cost <= COST_NEUTRAL) | |
2435 | fputs (" Replacing...\n", dump_file); | |
2436 | else | |
2437 | fputs (" Not replaced.\n", dump_file); | |
2438 | } | |
2439 | ||
2440 | if (cost <= COST_NEUTRAL) | |
2441 | replace_conditional_candidate (c); | |
2442 | } | |
2443 | } | |
2444 | else | |
2445 | replace_unconditional_candidate (c); | |
2446 | ||
2447 | if (c->sibling) | |
2448 | replace_uncond_cands_and_profitable_phis (lookup_cand (c->sibling)); | |
2449 | ||
2450 | if (c->dependent) | |
2451 | replace_uncond_cands_and_profitable_phis (lookup_cand (c->dependent)); | |
2452 | } | |
2453 | \f | |
9ce0359e | 2454 | /* Count the number of candidates in the tree rooted at C that have |
2455 | not already been replaced under other interpretations. */ | |
2456 | ||
d873e825 | 2457 | static int |
9ce0359e | 2458 | count_candidates (slsr_cand_t c) |
2459 | { | |
2460 | unsigned count = cand_already_replaced (c) ? 0 : 1; | |
2461 | ||
2462 | if (c->sibling) | |
2463 | count += count_candidates (lookup_cand (c->sibling)); | |
2464 | ||
2465 | if (c->dependent) | |
2466 | count += count_candidates (lookup_cand (c->dependent)); | |
2467 | ||
2468 | return count; | |
2469 | } | |
2470 | ||
2471 | /* Increase the count of INCREMENT by one in the increment vector. | |
1168caca | 2472 | INCREMENT is associated with candidate C. If INCREMENT is to be |
2473 | conditionally executed as part of a conditional candidate replacement, | |
2474 | IS_PHI_ADJUST is true, otherwise false. If an initializer | |
9ce0359e | 2475 | T_0 = stride * I is provided by a candidate that dominates all |
2476 | candidates with the same increment, also record T_0 for subsequent use. */ | |
2477 | ||
2478 | static void | |
ab2c1de8 | 2479 | record_increment (slsr_cand_t c, widest_int increment, bool is_phi_adjust) |
9ce0359e | 2480 | { |
2481 | bool found = false; | |
2482 | unsigned i; | |
2483 | ||
2484 | /* Treat increments that differ only in sign as identical so as to | |
2485 | share initializers, unless we are generating pointer arithmetic. */ | |
796b6678 | 2486 | if (!address_arithmetic_p && wi::neg_p (increment)) |
cf8f0e63 | 2487 | increment = -increment; |
9ce0359e | 2488 | |
2489 | for (i = 0; i < incr_vec_len; i++) | |
2490 | { | |
cf8f0e63 | 2491 | if (incr_vec[i].incr == increment) |
9ce0359e | 2492 | { |
2493 | incr_vec[i].count++; | |
2494 | found = true; | |
2495 | ||
2496 | /* If we previously recorded an initializer that doesn't | |
2497 | dominate this candidate, it's not going to be useful to | |
2498 | us after all. */ | |
2499 | if (incr_vec[i].initializer | |
2500 | && !dominated_by_p (CDI_DOMINATORS, | |
2501 | gimple_bb (c->cand_stmt), | |
2502 | incr_vec[i].init_bb)) | |
2503 | { | |
2504 | incr_vec[i].initializer = NULL_TREE; | |
2505 | incr_vec[i].init_bb = NULL; | |
2506 | } | |
2507 | ||
2508 | break; | |
2509 | } | |
2510 | } | |
2511 | ||
d62a8480 | 2512 | if (!found && incr_vec_len < MAX_INCR_VEC_LEN - 1) |
9ce0359e | 2513 | { |
2514 | /* The first time we see an increment, create the entry for it. | |
2515 | If this is the root candidate which doesn't have a basis, set | |
2516 | the count to zero. We're only processing it so it can possibly | |
2517 | provide an initializer for other candidates. */ | |
2518 | incr_vec[incr_vec_len].incr = increment; | |
1168caca | 2519 | incr_vec[incr_vec_len].count = c->basis || is_phi_adjust ? 1 : 0; |
9ce0359e | 2520 | incr_vec[incr_vec_len].cost = COST_INFINITE; |
2521 | ||
2522 | /* Optimistically record the first occurrence of this increment | |
2523 | as providing an initializer (if it does); we will revise this | |
2524 | opinion later if it doesn't dominate all other occurrences. | |
1168caca | 2525 | Exception: increments of -1, 0, 1 never need initializers; |
2526 | and phi adjustments don't ever provide initializers. */ | |
9ce0359e | 2527 | if (c->kind == CAND_ADD |
1168caca | 2528 | && !is_phi_adjust |
cf8f0e63 | 2529 | && c->index == increment |
796b6678 | 2530 | && (wi::gts_p (increment, 1) |
2531 | || wi::lts_p (increment, -1)) | |
9de38bc3 | 2532 | && (gimple_assign_rhs_code (c->cand_stmt) == PLUS_EXPR |
2533 | || gimple_assign_rhs_code (c->cand_stmt) == POINTER_PLUS_EXPR)) | |
9ce0359e | 2534 | { |
9de38bc3 | 2535 | tree t0 = NULL_TREE; |
9ce0359e | 2536 | tree rhs1 = gimple_assign_rhs1 (c->cand_stmt); |
2537 | tree rhs2 = gimple_assign_rhs2 (c->cand_stmt); | |
2538 | if (operand_equal_p (rhs1, c->base_expr, 0)) | |
2539 | t0 = rhs2; | |
9de38bc3 | 2540 | else if (operand_equal_p (rhs2, c->base_expr, 0)) |
9ce0359e | 2541 | t0 = rhs1; |
9de38bc3 | 2542 | if (t0 |
2543 | && SSA_NAME_DEF_STMT (t0) | |
2544 | && gimple_bb (SSA_NAME_DEF_STMT (t0))) | |
9ce0359e | 2545 | { |
2546 | incr_vec[incr_vec_len].initializer = t0; | |
2547 | incr_vec[incr_vec_len++].init_bb | |
2548 | = gimple_bb (SSA_NAME_DEF_STMT (t0)); | |
2549 | } | |
2550 | else | |
2551 | { | |
2552 | incr_vec[incr_vec_len].initializer = NULL_TREE; | |
2553 | incr_vec[incr_vec_len++].init_bb = NULL; | |
2554 | } | |
2555 | } | |
2556 | else | |
2557 | { | |
2558 | incr_vec[incr_vec_len].initializer = NULL_TREE; | |
2559 | incr_vec[incr_vec_len++].init_bb = NULL; | |
2560 | } | |
2561 | } | |
2562 | } | |
2563 | ||
1168caca | 2564 | /* Given phi statement PHI that hides a candidate from its BASIS, find |
2565 | the increments along each incoming arc (recursively handling additional | |
2566 | phis that may be present) and record them. These increments are the | |
2567 | difference in index between the index-adjusting statements and the | |
2568 | index of the basis. */ | |
2569 | ||
2570 | static void | |
2571 | record_phi_increments (slsr_cand_t basis, gimple phi) | |
2572 | { | |
2573 | unsigned i; | |
2574 | slsr_cand_t phi_cand = base_cand_from_table (gimple_phi_result (phi)); | |
2575 | ||
2576 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2577 | { | |
2578 | tree arg = gimple_phi_arg_def (phi, i); | |
2579 | ||
2580 | if (!operand_equal_p (arg, phi_cand->base_expr, 0)) | |
2581 | { | |
2582 | gimple arg_def = SSA_NAME_DEF_STMT (arg); | |
2583 | ||
2584 | if (gimple_code (arg_def) == GIMPLE_PHI) | |
2585 | record_phi_increments (basis, arg_def); | |
2586 | else | |
2587 | { | |
2588 | slsr_cand_t arg_cand = base_cand_from_table (arg); | |
5de9d3ed | 2589 | widest_int diff = arg_cand->index - basis->index; |
1168caca | 2590 | record_increment (arg_cand, diff, PHI_ADJUST); |
2591 | } | |
2592 | } | |
2593 | } | |
2594 | } | |
2595 | ||
9ce0359e | 2596 | /* Determine how many times each unique increment occurs in the set |
2597 | of candidates rooted at C's parent, recording the data in the | |
2598 | increment vector. For each unique increment I, if an initializer | |
2599 | T_0 = stride * I is provided by a candidate that dominates all | |
2600 | candidates with the same increment, also record T_0 for subsequent | |
2601 | use. */ | |
2602 | ||
2603 | static void | |
2604 | record_increments (slsr_cand_t c) | |
2605 | { | |
2606 | if (!cand_already_replaced (c)) | |
1168caca | 2607 | { |
2608 | if (!phi_dependent_cand_p (c)) | |
2609 | record_increment (c, cand_increment (c), NOT_PHI_ADJUST); | |
2610 | else | |
2611 | { | |
2612 | /* A candidate with a basis hidden by a phi will have one | |
2613 | increment for its relationship to the index represented by | |
2614 | the phi, and potentially additional increments along each | |
2615 | incoming edge. For the root of the dependency tree (which | |
2616 | has no basis), process just the initial index in case it has | |
2617 | an initializer that can be used by subsequent candidates. */ | |
2618 | record_increment (c, c->index, NOT_PHI_ADJUST); | |
2619 | ||
2620 | if (c->basis) | |
2621 | record_phi_increments (lookup_cand (c->basis), | |
2622 | lookup_cand (c->def_phi)->cand_stmt); | |
2623 | } | |
2624 | } | |
9ce0359e | 2625 | |
2626 | if (c->sibling) | |
2627 | record_increments (lookup_cand (c->sibling)); | |
2628 | ||
2629 | if (c->dependent) | |
2630 | record_increments (lookup_cand (c->dependent)); | |
2631 | } | |
2632 | ||
1168caca | 2633 | /* Add up and return the costs of introducing add statements that |
2634 | require the increment INCR on behalf of candidate C and phi | |
2635 | statement PHI. Accumulate into *SAVINGS the potential savings | |
2636 | from removing existing statements that feed PHI and have no other | |
2637 | uses. */ | |
2638 | ||
2639 | static int | |
5de9d3ed | 2640 | phi_incr_cost (slsr_cand_t c, const widest_int &incr, gimple phi, int *savings) |
1168caca | 2641 | { |
2642 | unsigned i; | |
2643 | int cost = 0; | |
2644 | slsr_cand_t basis = lookup_cand (c->basis); | |
2645 | slsr_cand_t phi_cand = base_cand_from_table (gimple_phi_result (phi)); | |
2646 | ||
2647 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2648 | { | |
2649 | tree arg = gimple_phi_arg_def (phi, i); | |
2650 | ||
2651 | if (!operand_equal_p (arg, phi_cand->base_expr, 0)) | |
2652 | { | |
2653 | gimple arg_def = SSA_NAME_DEF_STMT (arg); | |
2654 | ||
2655 | if (gimple_code (arg_def) == GIMPLE_PHI) | |
2656 | { | |
2657 | int feeding_savings = 0; | |
2658 | cost += phi_incr_cost (c, incr, arg_def, &feeding_savings); | |
2659 | if (has_single_use (gimple_phi_result (arg_def))) | |
2660 | *savings += feeding_savings; | |
2661 | } | |
2662 | else | |
2663 | { | |
2664 | slsr_cand_t arg_cand = base_cand_from_table (arg); | |
5de9d3ed | 2665 | widest_int diff = arg_cand->index - basis->index; |
1168caca | 2666 | |
2667 | if (incr == diff) | |
2668 | { | |
2669 | tree basis_lhs = gimple_assign_lhs (basis->cand_stmt); | |
2670 | tree lhs = gimple_assign_lhs (arg_cand->cand_stmt); | |
2671 | cost += add_cost (true, TYPE_MODE (TREE_TYPE (basis_lhs))); | |
2672 | if (has_single_use (lhs)) | |
2673 | *savings += stmt_cost (arg_cand->cand_stmt, true); | |
2674 | } | |
2675 | } | |
2676 | } | |
2677 | } | |
2678 | ||
2679 | return cost; | |
2680 | } | |
2681 | ||
9ce0359e | 2682 | /* Return the first candidate in the tree rooted at C that has not |
2683 | already been replaced, favoring siblings over dependents. */ | |
2684 | ||
2685 | static slsr_cand_t | |
2686 | unreplaced_cand_in_tree (slsr_cand_t c) | |
2687 | { | |
2688 | if (!cand_already_replaced (c)) | |
2689 | return c; | |
2690 | ||
2691 | if (c->sibling) | |
2692 | { | |
2693 | slsr_cand_t sib = unreplaced_cand_in_tree (lookup_cand (c->sibling)); | |
2694 | if (sib) | |
2695 | return sib; | |
2696 | } | |
2697 | ||
2698 | if (c->dependent) | |
2699 | { | |
2700 | slsr_cand_t dep = unreplaced_cand_in_tree (lookup_cand (c->dependent)); | |
2701 | if (dep) | |
2702 | return dep; | |
2703 | } | |
2704 | ||
2705 | return NULL; | |
2706 | } | |
2707 | ||
2708 | /* Return TRUE if the candidates in the tree rooted at C should be | |
2709 | optimized for speed, else FALSE. We estimate this based on the block | |
2710 | containing the most dominant candidate in the tree that has not yet | |
2711 | been replaced. */ | |
2712 | ||
2713 | static bool | |
2714 | optimize_cands_for_speed_p (slsr_cand_t c) | |
2715 | { | |
2716 | slsr_cand_t c2 = unreplaced_cand_in_tree (c); | |
2717 | gcc_assert (c2); | |
2718 | return optimize_bb_for_speed_p (gimple_bb (c2->cand_stmt)); | |
2719 | } | |
2720 | ||
2721 | /* Add COST_IN to the lowest cost of any dependent path starting at | |
2722 | candidate C or any of its siblings, counting only candidates along | |
2723 | such paths with increment INCR. Assume that replacing a candidate | |
2724 | reduces cost by REPL_SAVINGS. Also account for savings from any | |
1168caca | 2725 | statements that would go dead. If COUNT_PHIS is true, include |
2726 | costs of introducing feeding statements for conditional candidates. */ | |
9ce0359e | 2727 | |
2728 | static int | |
1168caca | 2729 | lowest_cost_path (int cost_in, int repl_savings, slsr_cand_t c, |
5de9d3ed | 2730 | const widest_int &incr, bool count_phis) |
9ce0359e | 2731 | { |
1168caca | 2732 | int local_cost, sib_cost, savings = 0; |
5de9d3ed | 2733 | widest_int cand_incr = cand_abs_increment (c); |
9ce0359e | 2734 | |
2735 | if (cand_already_replaced (c)) | |
2736 | local_cost = cost_in; | |
cf8f0e63 | 2737 | else if (incr == cand_incr) |
9ce0359e | 2738 | local_cost = cost_in - repl_savings - c->dead_savings; |
2739 | else | |
2740 | local_cost = cost_in - c->dead_savings; | |
2741 | ||
1168caca | 2742 | if (count_phis |
2743 | && phi_dependent_cand_p (c) | |
2744 | && !cand_already_replaced (c)) | |
2745 | { | |
2746 | gimple phi = lookup_cand (c->def_phi)->cand_stmt; | |
2747 | local_cost += phi_incr_cost (c, incr, phi, &savings); | |
2748 | ||
2749 | if (has_single_use (gimple_phi_result (phi))) | |
2750 | local_cost -= savings; | |
2751 | } | |
2752 | ||
9ce0359e | 2753 | if (c->dependent) |
2754 | local_cost = lowest_cost_path (local_cost, repl_savings, | |
1168caca | 2755 | lookup_cand (c->dependent), incr, |
2756 | count_phis); | |
9ce0359e | 2757 | |
2758 | if (c->sibling) | |
2759 | { | |
2760 | sib_cost = lowest_cost_path (cost_in, repl_savings, | |
1168caca | 2761 | lookup_cand (c->sibling), incr, |
2762 | count_phis); | |
9ce0359e | 2763 | local_cost = MIN (local_cost, sib_cost); |
2764 | } | |
2765 | ||
2766 | return local_cost; | |
2767 | } | |
2768 | ||
2769 | /* Compute the total savings that would accrue from all replacements | |
2770 | in the candidate tree rooted at C, counting only candidates with | |
2771 | increment INCR. Assume that replacing a candidate reduces cost | |
2772 | by REPL_SAVINGS. Also account for savings from statements that | |
2773 | would go dead. */ | |
2774 | ||
2775 | static int | |
5de9d3ed | 2776 | total_savings (int repl_savings, slsr_cand_t c, const widest_int &incr, |
1168caca | 2777 | bool count_phis) |
9ce0359e | 2778 | { |
2779 | int savings = 0; | |
5de9d3ed | 2780 | widest_int cand_incr = cand_abs_increment (c); |
9ce0359e | 2781 | |
cf8f0e63 | 2782 | if (incr == cand_incr && !cand_already_replaced (c)) |
9ce0359e | 2783 | savings += repl_savings + c->dead_savings; |
2784 | ||
1168caca | 2785 | if (count_phis |
2786 | && phi_dependent_cand_p (c) | |
2787 | && !cand_already_replaced (c)) | |
2788 | { | |
2789 | int phi_savings = 0; | |
2790 | gimple phi = lookup_cand (c->def_phi)->cand_stmt; | |
2791 | savings -= phi_incr_cost (c, incr, phi, &phi_savings); | |
2792 | ||
2793 | if (has_single_use (gimple_phi_result (phi))) | |
2794 | savings += phi_savings; | |
2795 | } | |
2796 | ||
9ce0359e | 2797 | if (c->dependent) |
1168caca | 2798 | savings += total_savings (repl_savings, lookup_cand (c->dependent), incr, |
2799 | count_phis); | |
9ce0359e | 2800 | |
2801 | if (c->sibling) | |
1168caca | 2802 | savings += total_savings (repl_savings, lookup_cand (c->sibling), incr, |
2803 | count_phis); | |
9ce0359e | 2804 | |
2805 | return savings; | |
2806 | } | |
2807 | ||
2808 | /* Use target-specific costs to determine and record which increments | |
2809 | in the current candidate tree are profitable to replace, assuming | |
2810 | MODE and SPEED. FIRST_DEP is the first dependent of the root of | |
2811 | the candidate tree. | |
2812 | ||
2813 | One slight limitation here is that we don't account for the possible | |
2814 | introduction of casts in some cases. See replace_one_candidate for | |
2815 | the cases where these are introduced. This should probably be cleaned | |
2816 | up sometime. */ | |
2817 | ||
2818 | static void | |
3754d046 | 2819 | analyze_increments (slsr_cand_t first_dep, machine_mode mode, bool speed) |
9ce0359e | 2820 | { |
2821 | unsigned i; | |
2822 | ||
2823 | for (i = 0; i < incr_vec_len; i++) | |
2824 | { | |
cf8f0e63 | 2825 | HOST_WIDE_INT incr = incr_vec[i].incr.to_shwi (); |
9ce0359e | 2826 | |
2827 | /* If somehow this increment is bigger than a HWI, we won't | |
2828 | be optimizing candidates that use it. And if the increment | |
2829 | has a count of zero, nothing will be done with it. */ | |
796b6678 | 2830 | if (!wi::fits_shwi_p (incr_vec[i].incr) || !incr_vec[i].count) |
9ce0359e | 2831 | incr_vec[i].cost = COST_INFINITE; |
2832 | ||
2833 | /* Increments of 0, 1, and -1 are always profitable to replace, | |
2834 | because they always replace a multiply or add with an add or | |
2835 | copy, and may cause one or more existing instructions to go | |
2836 | dead. Exception: -1 can't be assumed to be profitable for | |
2837 | pointer addition. */ | |
2838 | else if (incr == 0 | |
2839 | || incr == 1 | |
2840 | || (incr == -1 | |
2841 | && (gimple_assign_rhs_code (first_dep->cand_stmt) | |
2842 | != POINTER_PLUS_EXPR))) | |
2843 | incr_vec[i].cost = COST_NEUTRAL; | |
2844 | ||
38304570 | 2845 | /* FORNOW: If we need to add an initializer, give up if a cast from |
2846 | the candidate's type to its stride's type can lose precision. | |
2847 | This could eventually be handled better by expressly retaining the | |
2848 | result of a cast to a wider type in the stride. Example: | |
2849 | ||
2850 | short int _1; | |
2851 | _2 = (int) _1; | |
2852 | _3 = _2 * 10; | |
2853 | _4 = x + _3; ADD: x + (10 * _1) : int | |
2854 | _5 = _2 * 15; | |
2855 | _6 = x + _3; ADD: x + (15 * _1) : int | |
2856 | ||
2857 | Right now replacing _6 would cause insertion of an initializer | |
2858 | of the form "short int T = _1 * 5;" followed by a cast to | |
2859 | int, which could overflow incorrectly. Had we recorded _2 or | |
2860 | (int)_1 as the stride, this wouldn't happen. However, doing | |
2861 | this breaks other opportunities, so this will require some | |
2862 | care. */ | |
2863 | else if (!incr_vec[i].initializer | |
2864 | && TREE_CODE (first_dep->stride) != INTEGER_CST | |
2865 | && !legal_cast_p_1 (first_dep->stride, | |
2866 | gimple_assign_lhs (first_dep->cand_stmt))) | |
2867 | ||
2868 | incr_vec[i].cost = COST_INFINITE; | |
2869 | ||
781e2ff3 | 2870 | /* If we need to add an initializer, make sure we don't introduce |
2871 | a multiply by a pointer type, which can happen in certain cast | |
2872 | scenarios. FIXME: When cleaning up these cast issues, we can | |
2873 | afford to introduce the multiply provided we cast out to an | |
2874 | unsigned int of appropriate size. */ | |
2875 | else if (!incr_vec[i].initializer | |
2876 | && TREE_CODE (first_dep->stride) != INTEGER_CST | |
2877 | && POINTER_TYPE_P (TREE_TYPE (first_dep->stride))) | |
2878 | ||
2879 | incr_vec[i].cost = COST_INFINITE; | |
2880 | ||
9ce0359e | 2881 | /* For any other increment, if this is a multiply candidate, we |
2882 | must introduce a temporary T and initialize it with | |
2883 | T_0 = stride * increment. When optimizing for speed, walk the | |
2884 | candidate tree to calculate the best cost reduction along any | |
2885 | path; if it offsets the fixed cost of inserting the initializer, | |
2886 | replacing the increment is profitable. When optimizing for | |
2887 | size, instead calculate the total cost reduction from replacing | |
2888 | all candidates with this increment. */ | |
2889 | else if (first_dep->kind == CAND_MULT) | |
2890 | { | |
2891 | int cost = mult_by_coeff_cost (incr, mode, speed); | |
2892 | int repl_savings = mul_cost (speed, mode) - add_cost (speed, mode); | |
2893 | if (speed) | |
2894 | cost = lowest_cost_path (cost, repl_savings, first_dep, | |
1168caca | 2895 | incr_vec[i].incr, COUNT_PHIS); |
9ce0359e | 2896 | else |
1168caca | 2897 | cost -= total_savings (repl_savings, first_dep, incr_vec[i].incr, |
2898 | COUNT_PHIS); | |
9ce0359e | 2899 | |
2900 | incr_vec[i].cost = cost; | |
2901 | } | |
2902 | ||
2903 | /* If this is an add candidate, the initializer may already | |
2904 | exist, so only calculate the cost of the initializer if it | |
2905 | doesn't. We are replacing one add with another here, so the | |
2906 | known replacement savings is zero. We will account for removal | |
2907 | of dead instructions in lowest_cost_path or total_savings. */ | |
2908 | else | |
2909 | { | |
2910 | int cost = 0; | |
2911 | if (!incr_vec[i].initializer) | |
2912 | cost = mult_by_coeff_cost (incr, mode, speed); | |
2913 | ||
2914 | if (speed) | |
1168caca | 2915 | cost = lowest_cost_path (cost, 0, first_dep, incr_vec[i].incr, |
2916 | DONT_COUNT_PHIS); | |
9ce0359e | 2917 | else |
1168caca | 2918 | cost -= total_savings (0, first_dep, incr_vec[i].incr, |
2919 | DONT_COUNT_PHIS); | |
9ce0359e | 2920 | |
2921 | incr_vec[i].cost = cost; | |
2922 | } | |
2923 | } | |
2924 | } | |
2925 | ||
2926 | /* Return the nearest common dominator of BB1 and BB2. If the blocks | |
2927 | are identical, return the earlier of C1 and C2 in *WHERE. Otherwise, | |
2928 | if the NCD matches BB1, return C1 in *WHERE; if the NCD matches BB2, | |
2929 | return C2 in *WHERE; and if the NCD matches neither, return NULL in | |
2930 | *WHERE. Note: It is possible for one of C1 and C2 to be NULL. */ | |
2931 | ||
2932 | static basic_block | |
2933 | ncd_for_two_cands (basic_block bb1, basic_block bb2, | |
2934 | slsr_cand_t c1, slsr_cand_t c2, slsr_cand_t *where) | |
2935 | { | |
2936 | basic_block ncd; | |
2937 | ||
2938 | if (!bb1) | |
2939 | { | |
2940 | *where = c2; | |
2941 | return bb2; | |
2942 | } | |
2943 | ||
2944 | if (!bb2) | |
2945 | { | |
2946 | *where = c1; | |
2947 | return bb1; | |
2948 | } | |
2949 | ||
2950 | ncd = nearest_common_dominator (CDI_DOMINATORS, bb1, bb2); | |
2951 | ||
2952 | /* If both candidates are in the same block, the earlier | |
2953 | candidate wins. */ | |
2954 | if (bb1 == ncd && bb2 == ncd) | |
2955 | { | |
2956 | if (!c1 || (c2 && c2->cand_num < c1->cand_num)) | |
2957 | *where = c2; | |
2958 | else | |
2959 | *where = c1; | |
2960 | } | |
2961 | ||
2962 | /* Otherwise, if one of them produced a candidate in the | |
2963 | dominator, that one wins. */ | |
2964 | else if (bb1 == ncd) | |
2965 | *where = c1; | |
2966 | ||
2967 | else if (bb2 == ncd) | |
2968 | *where = c2; | |
2969 | ||
2970 | /* If neither matches the dominator, neither wins. */ | |
2971 | else | |
2972 | *where = NULL; | |
2973 | ||
2974 | return ncd; | |
2975 | } | |
2976 | ||
1168caca | 2977 | /* Consider all candidates that feed PHI. Find the nearest common |
2978 | dominator of those candidates requiring the given increment INCR. | |
2979 | Further find and return the nearest common dominator of this result | |
2980 | with block NCD. If the returned block contains one or more of the | |
2981 | candidates, return the earliest candidate in the block in *WHERE. */ | |
2982 | ||
2983 | static basic_block | |
5de9d3ed | 2984 | ncd_with_phi (slsr_cand_t c, const widest_int &incr, gimple phi, |
1168caca | 2985 | basic_block ncd, slsr_cand_t *where) |
2986 | { | |
2987 | unsigned i; | |
2988 | slsr_cand_t basis = lookup_cand (c->basis); | |
2989 | slsr_cand_t phi_cand = base_cand_from_table (gimple_phi_result (phi)); | |
2990 | ||
2991 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2992 | { | |
2993 | tree arg = gimple_phi_arg_def (phi, i); | |
2994 | ||
2995 | if (!operand_equal_p (arg, phi_cand->base_expr, 0)) | |
2996 | { | |
2997 | gimple arg_def = SSA_NAME_DEF_STMT (arg); | |
2998 | ||
2999 | if (gimple_code (arg_def) == GIMPLE_PHI) | |
3000 | ncd = ncd_with_phi (c, incr, arg_def, ncd, where); | |
3001 | else | |
3002 | { | |
3003 | slsr_cand_t arg_cand = base_cand_from_table (arg); | |
5de9d3ed | 3004 | widest_int diff = arg_cand->index - basis->index; |
f626787a | 3005 | basic_block pred = gimple_phi_arg_edge (phi, i)->src; |
1168caca | 3006 | |
3007 | if ((incr == diff) || (!address_arithmetic_p && incr == -diff)) | |
f626787a | 3008 | ncd = ncd_for_two_cands (ncd, pred, *where, NULL, where); |
1168caca | 3009 | } |
3010 | } | |
3011 | } | |
3012 | ||
3013 | return ncd; | |
3014 | } | |
3015 | ||
3016 | /* Consider the candidate C together with any candidates that feed | |
3017 | C's phi dependence (if any). Find and return the nearest common | |
3018 | dominator of those candidates requiring the given increment INCR. | |
3019 | If the returned block contains one or more of the candidates, | |
3020 | return the earliest candidate in the block in *WHERE. */ | |
3021 | ||
3022 | static basic_block | |
5de9d3ed | 3023 | ncd_of_cand_and_phis (slsr_cand_t c, const widest_int &incr, slsr_cand_t *where) |
1168caca | 3024 | { |
3025 | basic_block ncd = NULL; | |
3026 | ||
3027 | if (cand_abs_increment (c) == incr) | |
3028 | { | |
3029 | ncd = gimple_bb (c->cand_stmt); | |
3030 | *where = c; | |
3031 | } | |
3032 | ||
3033 | if (phi_dependent_cand_p (c)) | |
3034 | ncd = ncd_with_phi (c, incr, lookup_cand (c->def_phi)->cand_stmt, | |
3035 | ncd, where); | |
3036 | ||
3037 | return ncd; | |
3038 | } | |
3039 | ||
9ce0359e | 3040 | /* Consider all candidates in the tree rooted at C for which INCR |
3041 | represents the required increment of C relative to its basis. | |
3042 | Find and return the basic block that most nearly dominates all | |
3043 | such candidates. If the returned block contains one or more of | |
3044 | the candidates, return the earliest candidate in the block in | |
3045 | *WHERE. */ | |
3046 | ||
3047 | static basic_block | |
5de9d3ed | 3048 | nearest_common_dominator_for_cands (slsr_cand_t c, const widest_int &incr, |
9ce0359e | 3049 | slsr_cand_t *where) |
3050 | { | |
3051 | basic_block sib_ncd = NULL, dep_ncd = NULL, this_ncd = NULL, ncd; | |
3052 | slsr_cand_t sib_where = NULL, dep_where = NULL, this_where = NULL, new_where; | |
9ce0359e | 3053 | |
3054 | /* First find the NCD of all siblings and dependents. */ | |
3055 | if (c->sibling) | |
3056 | sib_ncd = nearest_common_dominator_for_cands (lookup_cand (c->sibling), | |
3057 | incr, &sib_where); | |
3058 | if (c->dependent) | |
3059 | dep_ncd = nearest_common_dominator_for_cands (lookup_cand (c->dependent), | |
3060 | incr, &dep_where); | |
3061 | if (!sib_ncd && !dep_ncd) | |
3062 | { | |
3063 | new_where = NULL; | |
3064 | ncd = NULL; | |
3065 | } | |
3066 | else if (sib_ncd && !dep_ncd) | |
3067 | { | |
3068 | new_where = sib_where; | |
3069 | ncd = sib_ncd; | |
3070 | } | |
3071 | else if (dep_ncd && !sib_ncd) | |
3072 | { | |
3073 | new_where = dep_where; | |
3074 | ncd = dep_ncd; | |
3075 | } | |
3076 | else | |
3077 | ncd = ncd_for_two_cands (sib_ncd, dep_ncd, sib_where, | |
3078 | dep_where, &new_where); | |
3079 | ||
3080 | /* If the candidate's increment doesn't match the one we're interested | |
1168caca | 3081 | in (and nor do any increments for feeding defs of a phi-dependence), |
3082 | then the result depends only on siblings and dependents. */ | |
3083 | this_ncd = ncd_of_cand_and_phis (c, incr, &this_where); | |
9ce0359e | 3084 | |
1168caca | 3085 | if (!this_ncd || cand_already_replaced (c)) |
9ce0359e | 3086 | { |
3087 | *where = new_where; | |
3088 | return ncd; | |
3089 | } | |
3090 | ||
3091 | /* Otherwise, compare this candidate with the result from all siblings | |
3092 | and dependents. */ | |
9ce0359e | 3093 | ncd = ncd_for_two_cands (ncd, this_ncd, new_where, this_where, where); |
3094 | ||
3095 | return ncd; | |
3096 | } | |
3097 | ||
3098 | /* Return TRUE if the increment indexed by INDEX is profitable to replace. */ | |
3099 | ||
3100 | static inline bool | |
3101 | profitable_increment_p (unsigned index) | |
3102 | { | |
3103 | return (incr_vec[index].cost <= COST_NEUTRAL); | |
3104 | } | |
3105 | ||
3106 | /* For each profitable increment in the increment vector not equal to | |
3107 | 0 or 1 (or -1, for non-pointer arithmetic), find the nearest common | |
3108 | dominator of all statements in the candidate chain rooted at C | |
3109 | that require that increment, and insert an initializer | |
3110 | T_0 = stride * increment at that location. Record T_0 with the | |
3111 | increment record. */ | |
3112 | ||
3113 | static void | |
3114 | insert_initializers (slsr_cand_t c) | |
3115 | { | |
3116 | unsigned i; | |
9ce0359e | 3117 | |
3118 | for (i = 0; i < incr_vec_len; i++) | |
3119 | { | |
3120 | basic_block bb; | |
3121 | slsr_cand_t where = NULL; | |
3122 | gimple init_stmt; | |
3123 | tree stride_type, new_name, incr_tree; | |
5de9d3ed | 3124 | widest_int incr = incr_vec[i].incr; |
9ce0359e | 3125 | |
3126 | if (!profitable_increment_p (i) | |
796b6678 | 3127 | || incr == 1 |
3128 | || (incr == -1 | |
9ce0359e | 3129 | && gimple_assign_rhs_code (c->cand_stmt) != POINTER_PLUS_EXPR) |
796b6678 | 3130 | || incr == 0) |
9ce0359e | 3131 | continue; |
3132 | ||
3133 | /* We may have already identified an existing initializer that | |
3134 | will suffice. */ | |
3135 | if (incr_vec[i].initializer) | |
3136 | { | |
3137 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3138 | { | |
3139 | fputs ("Using existing initializer: ", dump_file); | |
3140 | print_gimple_stmt (dump_file, | |
3141 | SSA_NAME_DEF_STMT (incr_vec[i].initializer), | |
3142 | 0, 0); | |
3143 | } | |
3144 | continue; | |
3145 | } | |
3146 | ||
3147 | /* Find the block that most closely dominates all candidates | |
3148 | with this increment. If there is at least one candidate in | |
3149 | that block, the earliest one will be returned in WHERE. */ | |
3150 | bb = nearest_common_dominator_for_cands (c, incr, &where); | |
3151 | ||
3152 | /* Create a new SSA name to hold the initializer's value. */ | |
3153 | stride_type = TREE_TYPE (c->stride); | |
d54ab372 | 3154 | new_name = make_temp_ssa_name (stride_type, NULL, "slsr"); |
9ce0359e | 3155 | incr_vec[i].initializer = new_name; |
3156 | ||
3157 | /* Create the initializer and insert it in the latest possible | |
3158 | dominating position. */ | |
e913b5cd | 3159 | incr_tree = wide_int_to_tree (stride_type, incr); |
9ce0359e | 3160 | init_stmt = gimple_build_assign_with_ops (MULT_EXPR, new_name, |
3161 | c->stride, incr_tree); | |
3162 | if (where) | |
3163 | { | |
3164 | gimple_stmt_iterator gsi = gsi_for_stmt (where->cand_stmt); | |
3165 | gsi_insert_before (&gsi, init_stmt, GSI_SAME_STMT); | |
3166 | gimple_set_location (init_stmt, gimple_location (where->cand_stmt)); | |
3167 | } | |
3168 | else | |
3169 | { | |
3170 | gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
3171 | gimple basis_stmt = lookup_cand (c->basis)->cand_stmt; | |
3172 | ||
3173 | if (!gsi_end_p (gsi) && is_ctrl_stmt (gsi_stmt (gsi))) | |
3174 | gsi_insert_before (&gsi, init_stmt, GSI_SAME_STMT); | |
3175 | else | |
3176 | gsi_insert_after (&gsi, init_stmt, GSI_SAME_STMT); | |
3177 | ||
3178 | gimple_set_location (init_stmt, gimple_location (basis_stmt)); | |
3179 | } | |
3180 | ||
3181 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3182 | { | |
3183 | fputs ("Inserting initializer: ", dump_file); | |
3184 | print_gimple_stmt (dump_file, init_stmt, 0, 0); | |
3185 | } | |
3186 | } | |
3187 | } | |
3188 | ||
1168caca | 3189 | /* Return TRUE iff all required increments for candidates feeding PHI |
3190 | are profitable to replace on behalf of candidate C. */ | |
3191 | ||
3192 | static bool | |
3193 | all_phi_incrs_profitable (slsr_cand_t c, gimple phi) | |
3194 | { | |
3195 | unsigned i; | |
3196 | slsr_cand_t basis = lookup_cand (c->basis); | |
3197 | slsr_cand_t phi_cand = base_cand_from_table (gimple_phi_result (phi)); | |
3198 | ||
3199 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
3200 | { | |
3201 | tree arg = gimple_phi_arg_def (phi, i); | |
3202 | ||
3203 | if (!operand_equal_p (arg, phi_cand->base_expr, 0)) | |
3204 | { | |
3205 | gimple arg_def = SSA_NAME_DEF_STMT (arg); | |
3206 | ||
3207 | if (gimple_code (arg_def) == GIMPLE_PHI) | |
3208 | { | |
3209 | if (!all_phi_incrs_profitable (c, arg_def)) | |
3210 | return false; | |
3211 | } | |
3212 | else | |
3213 | { | |
d62a8480 | 3214 | int j; |
1168caca | 3215 | slsr_cand_t arg_cand = base_cand_from_table (arg); |
5de9d3ed | 3216 | widest_int increment = arg_cand->index - basis->index; |
1168caca | 3217 | |
796b6678 | 3218 | if (!address_arithmetic_p && wi::neg_p (increment)) |
1168caca | 3219 | increment = -increment; |
3220 | ||
3221 | j = incr_vec_index (increment); | |
3222 | ||
3223 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3224 | { | |
3225 | fprintf (dump_file, " Conditional candidate %d, phi: ", | |
3226 | c->cand_num); | |
3227 | print_gimple_stmt (dump_file, phi, 0, 0); | |
3228 | fputs (" increment: ", dump_file); | |
e913b5cd | 3229 | print_decs (increment, dump_file); |
d62a8480 | 3230 | if (j < 0) |
3231 | fprintf (dump_file, | |
3232 | "\n Not replaced; incr_vec overflow.\n"); | |
3233 | else { | |
3234 | fprintf (dump_file, "\n cost: %d\n", incr_vec[j].cost); | |
3235 | if (profitable_increment_p (j)) | |
3236 | fputs (" Replacing...\n", dump_file); | |
3237 | else | |
3238 | fputs (" Not replaced.\n", dump_file); | |
3239 | } | |
1168caca | 3240 | } |
3241 | ||
d62a8480 | 3242 | if (j < 0 || !profitable_increment_p (j)) |
1168caca | 3243 | return false; |
3244 | } | |
3245 | } | |
3246 | } | |
3247 | ||
3248 | return true; | |
3249 | } | |
3250 | ||
9ce0359e | 3251 | /* Create a NOP_EXPR that copies FROM_EXPR into a new SSA name of |
3252 | type TO_TYPE, and insert it in front of the statement represented | |
3253 | by candidate C. Use *NEW_VAR to create the new SSA name. Return | |
3254 | the new SSA name. */ | |
3255 | ||
3256 | static tree | |
d54ab372 | 3257 | introduce_cast_before_cand (slsr_cand_t c, tree to_type, tree from_expr) |
9ce0359e | 3258 | { |
3259 | tree cast_lhs; | |
3260 | gimple cast_stmt; | |
3261 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); | |
3262 | ||
d54ab372 | 3263 | cast_lhs = make_temp_ssa_name (to_type, NULL, "slsr"); |
9ce0359e | 3264 | cast_stmt = gimple_build_assign_with_ops (NOP_EXPR, cast_lhs, |
3265 | from_expr, NULL_TREE); | |
3266 | gimple_set_location (cast_stmt, gimple_location (c->cand_stmt)); | |
3267 | gsi_insert_before (&gsi, cast_stmt, GSI_SAME_STMT); | |
3268 | ||
3269 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3270 | { | |
3271 | fputs (" Inserting: ", dump_file); | |
3272 | print_gimple_stmt (dump_file, cast_stmt, 0, 0); | |
3273 | } | |
3274 | ||
3275 | return cast_lhs; | |
3276 | } | |
3277 | ||
3278 | /* Replace the RHS of the statement represented by candidate C with | |
3279 | NEW_CODE, NEW_RHS1, and NEW_RHS2, provided that to do so doesn't | |
3280 | leave C unchanged or just interchange its operands. The original | |
3281 | operation and operands are in OLD_CODE, OLD_RHS1, and OLD_RHS2. | |
3282 | If the replacement was made and we are doing a details dump, | |
3283 | return the revised statement, else NULL. */ | |
3284 | ||
3285 | static gimple | |
3286 | replace_rhs_if_not_dup (enum tree_code new_code, tree new_rhs1, tree new_rhs2, | |
3287 | enum tree_code old_code, tree old_rhs1, tree old_rhs2, | |
3288 | slsr_cand_t c) | |
3289 | { | |
3290 | if (new_code != old_code | |
3291 | || ((!operand_equal_p (new_rhs1, old_rhs1, 0) | |
3292 | || !operand_equal_p (new_rhs2, old_rhs2, 0)) | |
3293 | && (!operand_equal_p (new_rhs1, old_rhs2, 0) | |
3294 | || !operand_equal_p (new_rhs2, old_rhs1, 0)))) | |
3295 | { | |
3296 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); | |
3297 | gimple_assign_set_rhs_with_ops (&gsi, new_code, new_rhs1, new_rhs2); | |
3298 | update_stmt (gsi_stmt (gsi)); | |
dc77dea1 | 3299 | c->cand_stmt = gsi_stmt (gsi); |
9ce0359e | 3300 | |
3301 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3302 | return gsi_stmt (gsi); | |
3303 | } | |
3304 | ||
3305 | else if (dump_file && (dump_flags & TDF_DETAILS)) | |
3306 | fputs (" (duplicate, not actually replacing)\n", dump_file); | |
3307 | ||
3308 | return NULL; | |
3309 | } | |
3310 | ||
3311 | /* Strength-reduce the statement represented by candidate C by replacing | |
3312 | it with an equivalent addition or subtraction. I is the index into | |
3313 | the increment vector identifying C's increment. NEW_VAR is used to | |
3314 | create a new SSA name if a cast needs to be introduced. BASIS_NAME | |
3315 | is the rhs1 to use in creating the add/subtract. */ | |
3316 | ||
3317 | static void | |
d54ab372 | 3318 | replace_one_candidate (slsr_cand_t c, unsigned i, tree basis_name) |
9ce0359e | 3319 | { |
3320 | gimple stmt_to_print = NULL; | |
3321 | tree orig_rhs1, orig_rhs2; | |
3322 | tree rhs2; | |
3323 | enum tree_code orig_code, repl_code; | |
5de9d3ed | 3324 | widest_int cand_incr; |
9ce0359e | 3325 | |
3326 | orig_code = gimple_assign_rhs_code (c->cand_stmt); | |
3327 | orig_rhs1 = gimple_assign_rhs1 (c->cand_stmt); | |
3328 | orig_rhs2 = gimple_assign_rhs2 (c->cand_stmt); | |
3329 | cand_incr = cand_increment (c); | |
3330 | ||
3331 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3332 | { | |
3333 | fputs ("Replacing: ", dump_file); | |
3334 | print_gimple_stmt (dump_file, c->cand_stmt, 0, 0); | |
3335 | stmt_to_print = c->cand_stmt; | |
3336 | } | |
3337 | ||
3338 | if (address_arithmetic_p) | |
3339 | repl_code = POINTER_PLUS_EXPR; | |
3340 | else | |
3341 | repl_code = PLUS_EXPR; | |
3342 | ||
3343 | /* If the increment has an initializer T_0, replace the candidate | |
3344 | statement with an add of the basis name and the initializer. */ | |
3345 | if (incr_vec[i].initializer) | |
3346 | { | |
3347 | tree init_type = TREE_TYPE (incr_vec[i].initializer); | |
3348 | tree orig_type = TREE_TYPE (orig_rhs2); | |
3349 | ||
3350 | if (types_compatible_p (orig_type, init_type)) | |
3351 | rhs2 = incr_vec[i].initializer; | |
3352 | else | |
3353 | rhs2 = introduce_cast_before_cand (c, orig_type, | |
d54ab372 | 3354 | incr_vec[i].initializer); |
9ce0359e | 3355 | |
cf8f0e63 | 3356 | if (incr_vec[i].incr != cand_incr) |
9ce0359e | 3357 | { |
3358 | gcc_assert (repl_code == PLUS_EXPR); | |
3359 | repl_code = MINUS_EXPR; | |
3360 | } | |
3361 | ||
3362 | stmt_to_print = replace_rhs_if_not_dup (repl_code, basis_name, rhs2, | |
3363 | orig_code, orig_rhs1, orig_rhs2, | |
3364 | c); | |
3365 | } | |
3366 | ||
3367 | /* Otherwise, the increment is one of -1, 0, and 1. Replace | |
3368 | with a subtract of the stride from the basis name, a copy | |
3369 | from the basis name, or an add of the stride to the basis | |
3370 | name, respectively. It may be necessary to introduce a | |
3371 | cast (or reuse an existing cast). */ | |
796b6678 | 3372 | else if (cand_incr == 1) |
9ce0359e | 3373 | { |
3374 | tree stride_type = TREE_TYPE (c->stride); | |
3375 | tree orig_type = TREE_TYPE (orig_rhs2); | |
3376 | ||
3377 | if (types_compatible_p (orig_type, stride_type)) | |
3378 | rhs2 = c->stride; | |
3379 | else | |
d54ab372 | 3380 | rhs2 = introduce_cast_before_cand (c, orig_type, c->stride); |
9ce0359e | 3381 | |
3382 | stmt_to_print = replace_rhs_if_not_dup (repl_code, basis_name, rhs2, | |
3383 | orig_code, orig_rhs1, orig_rhs2, | |
3384 | c); | |
3385 | } | |
3386 | ||
796b6678 | 3387 | else if (cand_incr == -1) |
9ce0359e | 3388 | { |
3389 | tree stride_type = TREE_TYPE (c->stride); | |
3390 | tree orig_type = TREE_TYPE (orig_rhs2); | |
3391 | gcc_assert (repl_code != POINTER_PLUS_EXPR); | |
3392 | ||
3393 | if (types_compatible_p (orig_type, stride_type)) | |
3394 | rhs2 = c->stride; | |
3395 | else | |
d54ab372 | 3396 | rhs2 = introduce_cast_before_cand (c, orig_type, c->stride); |
9ce0359e | 3397 | |
3398 | if (orig_code != MINUS_EXPR | |
3399 | || !operand_equal_p (basis_name, orig_rhs1, 0) | |
3400 | || !operand_equal_p (rhs2, orig_rhs2, 0)) | |
3401 | { | |
3402 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); | |
3403 | gimple_assign_set_rhs_with_ops (&gsi, MINUS_EXPR, basis_name, rhs2); | |
3404 | update_stmt (gsi_stmt (gsi)); | |
dc77dea1 | 3405 | c->cand_stmt = gsi_stmt (gsi); |
9ce0359e | 3406 | |
3407 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3408 | stmt_to_print = gsi_stmt (gsi); | |
3409 | } | |
3410 | else if (dump_file && (dump_flags & TDF_DETAILS)) | |
3411 | fputs (" (duplicate, not actually replacing)\n", dump_file); | |
3412 | } | |
3413 | ||
796b6678 | 3414 | else if (cand_incr == 0) |
9ce0359e | 3415 | { |
3416 | tree lhs = gimple_assign_lhs (c->cand_stmt); | |
3417 | tree lhs_type = TREE_TYPE (lhs); | |
3418 | tree basis_type = TREE_TYPE (basis_name); | |
3419 | ||
3420 | if (types_compatible_p (lhs_type, basis_type)) | |
3421 | { | |
3422 | gimple copy_stmt = gimple_build_assign (lhs, basis_name); | |
3423 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); | |
3424 | gimple_set_location (copy_stmt, gimple_location (c->cand_stmt)); | |
3425 | gsi_replace (&gsi, copy_stmt, false); | |
8f0676e4 | 3426 | c->cand_stmt = copy_stmt; |
9ce0359e | 3427 | |
3428 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3429 | stmt_to_print = copy_stmt; | |
3430 | } | |
3431 | else | |
3432 | { | |
3433 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); | |
3434 | gimple cast_stmt = gimple_build_assign_with_ops (NOP_EXPR, lhs, | |
3435 | basis_name, | |
3436 | NULL_TREE); | |
3437 | gimple_set_location (cast_stmt, gimple_location (c->cand_stmt)); | |
3438 | gsi_replace (&gsi, cast_stmt, false); | |
8f0676e4 | 3439 | c->cand_stmt = cast_stmt; |
9ce0359e | 3440 | |
3441 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3442 | stmt_to_print = cast_stmt; | |
3443 | } | |
3444 | } | |
3445 | else | |
3446 | gcc_unreachable (); | |
3447 | ||
3448 | if (dump_file && (dump_flags & TDF_DETAILS) && stmt_to_print) | |
3449 | { | |
3450 | fputs ("With: ", dump_file); | |
3451 | print_gimple_stmt (dump_file, stmt_to_print, 0, 0); | |
3452 | fputs ("\n", dump_file); | |
3453 | } | |
3454 | } | |
3455 | ||
3456 | /* For each candidate in the tree rooted at C, replace it with | |
3457 | an increment if such has been shown to be profitable. */ | |
3458 | ||
3459 | static void | |
3460 | replace_profitable_candidates (slsr_cand_t c) | |
3461 | { | |
3462 | if (!cand_already_replaced (c)) | |
3463 | { | |
5de9d3ed | 3464 | widest_int increment = cand_abs_increment (c); |
9ce0359e | 3465 | enum tree_code orig_code = gimple_assign_rhs_code (c->cand_stmt); |
d62a8480 | 3466 | int i; |
9ce0359e | 3467 | |
3468 | i = incr_vec_index (increment); | |
3469 | ||
3470 | /* Only process profitable increments. Nothing useful can be done | |
3471 | to a cast or copy. */ | |
d62a8480 | 3472 | if (i >= 0 |
3473 | && profitable_increment_p (i) | |
9ce0359e | 3474 | && orig_code != MODIFY_EXPR |
3475 | && orig_code != NOP_EXPR) | |
3476 | { | |
1168caca | 3477 | if (phi_dependent_cand_p (c)) |
3478 | { | |
3479 | gimple phi = lookup_cand (c->def_phi)->cand_stmt; | |
3480 | ||
3481 | if (all_phi_incrs_profitable (c, phi)) | |
3482 | { | |
3483 | /* Look up the LHS SSA name from C's basis. This will be | |
3484 | the RHS1 of the adds we will introduce to create new | |
3485 | phi arguments. */ | |
3486 | slsr_cand_t basis = lookup_cand (c->basis); | |
3487 | tree basis_name = gimple_assign_lhs (basis->cand_stmt); | |
3488 | ||
3489 | /* Create a new phi statement that will represent C's true | |
3490 | basis after the transformation is complete. */ | |
3491 | location_t loc = gimple_location (c->cand_stmt); | |
3492 | tree name = create_phi_basis (c, phi, basis_name, | |
3493 | loc, UNKNOWN_STRIDE); | |
3494 | ||
3495 | /* Replace C with an add of the new basis phi and the | |
3496 | increment. */ | |
d54ab372 | 3497 | replace_one_candidate (c, i, name); |
1168caca | 3498 | } |
3499 | } | |
3500 | else | |
3501 | { | |
3502 | slsr_cand_t basis = lookup_cand (c->basis); | |
3503 | tree basis_name = gimple_assign_lhs (basis->cand_stmt); | |
d54ab372 | 3504 | replace_one_candidate (c, i, basis_name); |
1168caca | 3505 | } |
9ce0359e | 3506 | } |
3507 | } | |
3508 | ||
3509 | if (c->sibling) | |
3510 | replace_profitable_candidates (lookup_cand (c->sibling)); | |
3511 | ||
3512 | if (c->dependent) | |
3513 | replace_profitable_candidates (lookup_cand (c->dependent)); | |
3514 | } | |
3515 | \f | |
4502f5d0 | 3516 | /* Analyze costs of related candidates in the candidate vector, |
3517 | and make beneficial replacements. */ | |
3518 | ||
3519 | static void | |
3520 | analyze_candidates_and_replace (void) | |
3521 | { | |
3522 | unsigned i; | |
3523 | slsr_cand_t c; | |
3524 | ||
3525 | /* Each candidate that has a null basis and a non-null | |
3526 | dependent is the root of a tree of related statements. | |
3527 | Analyze each tree to determine a subset of those | |
3528 | statements that can be replaced with maximum benefit. */ | |
f1f41a6c | 3529 | FOR_EACH_VEC_ELT (cand_vec, i, c) |
4502f5d0 | 3530 | { |
3531 | slsr_cand_t first_dep; | |
3532 | ||
3533 | if (c->basis != 0 || c->dependent == 0) | |
3534 | continue; | |
3535 | ||
3536 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3537 | fprintf (dump_file, "\nProcessing dependency tree rooted at %d.\n", | |
3538 | c->cand_num); | |
3539 | ||
3540 | first_dep = lookup_cand (c->dependent); | |
3541 | ||
070bf980 | 3542 | /* If this is a chain of CAND_REFs, unconditionally replace |
3543 | each of them with a strength-reduced data reference. */ | |
3544 | if (c->kind == CAND_REF) | |
3545 | replace_refs (c); | |
3546 | ||
1168caca | 3547 | /* If the common stride of all related candidates is a known |
3548 | constant, each candidate without a phi-dependence can be | |
3549 | profitably replaced. Each replaces a multiply by a single | |
3550 | add, with the possibility that a feeding add also goes dead. | |
3551 | A candidate with a phi-dependence is replaced only if the | |
3552 | compensation code it requires is offset by the strength | |
3553 | reduction savings. */ | |
3554 | else if (TREE_CODE (c->stride) == INTEGER_CST) | |
3555 | replace_uncond_cands_and_profitable_phis (first_dep); | |
4502f5d0 | 3556 | |
9ce0359e | 3557 | /* When the stride is an SSA name, it may still be profitable |
3558 | to replace some or all of the dependent candidates, depending | |
3559 | on whether the introduced increments can be reused, or are | |
3560 | less expensive to calculate than the replaced statements. */ | |
3561 | else | |
3562 | { | |
3754d046 | 3563 | machine_mode mode; |
9ce0359e | 3564 | bool speed; |
3565 | ||
3566 | /* Determine whether we'll be generating pointer arithmetic | |
3567 | when replacing candidates. */ | |
3568 | address_arithmetic_p = (c->kind == CAND_ADD | |
fd291195 | 3569 | && POINTER_TYPE_P (c->cand_type)); |
9ce0359e | 3570 | |
3571 | /* If all candidates have already been replaced under other | |
3572 | interpretations, nothing remains to be done. */ | |
259c15e6 | 3573 | if (!count_candidates (c)) |
9ce0359e | 3574 | continue; |
3575 | ||
3576 | /* Construct an array of increments for this candidate chain. */ | |
259c15e6 | 3577 | incr_vec = XNEWVEC (incr_info, MAX_INCR_VEC_LEN); |
9ce0359e | 3578 | incr_vec_len = 0; |
3579 | record_increments (c); | |
3580 | ||
3581 | /* Determine which increments are profitable to replace. */ | |
3582 | mode = TYPE_MODE (TREE_TYPE (gimple_assign_lhs (c->cand_stmt))); | |
3583 | speed = optimize_cands_for_speed_p (c); | |
3584 | analyze_increments (first_dep, mode, speed); | |
3585 | ||
3586 | /* Insert initializers of the form T_0 = stride * increment | |
3587 | for use in profitable replacements. */ | |
3588 | insert_initializers (first_dep); | |
3589 | dump_incr_vec (); | |
3590 | ||
3591 | /* Perform the replacements. */ | |
3592 | replace_profitable_candidates (first_dep); | |
3593 | free (incr_vec); | |
3594 | } | |
4502f5d0 | 3595 | } |
3596 | } | |
3597 | ||
65b0537f | 3598 | namespace { |
3599 | ||
3600 | const pass_data pass_data_strength_reduction = | |
3601 | { | |
3602 | GIMPLE_PASS, /* type */ | |
3603 | "slsr", /* name */ | |
3604 | OPTGROUP_NONE, /* optinfo_flags */ | |
65b0537f | 3605 | TV_GIMPLE_SLSR, /* tv_id */ |
3606 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
3607 | 0, /* properties_provided */ | |
3608 | 0, /* properties_destroyed */ | |
3609 | 0, /* todo_flags_start */ | |
8b88439e | 3610 | 0, /* todo_flags_finish */ |
65b0537f | 3611 | }; |
3612 | ||
3613 | class pass_strength_reduction : public gimple_opt_pass | |
3614 | { | |
3615 | public: | |
3616 | pass_strength_reduction (gcc::context *ctxt) | |
3617 | : gimple_opt_pass (pass_data_strength_reduction, ctxt) | |
3618 | {} | |
3619 | ||
3620 | /* opt_pass methods: */ | |
3621 | virtual bool gate (function *) { return flag_tree_slsr; } | |
3622 | virtual unsigned int execute (function *); | |
3623 | ||
3624 | }; // class pass_strength_reduction | |
3625 | ||
3626 | unsigned | |
3627 | pass_strength_reduction::execute (function *fun) | |
4502f5d0 | 3628 | { |
4502f5d0 | 3629 | /* Create the obstack where candidates will reside. */ |
3630 | gcc_obstack_init (&cand_obstack); | |
3631 | ||
3632 | /* Allocate the candidate vector. */ | |
f1f41a6c | 3633 | cand_vec.create (128); |
4502f5d0 | 3634 | |
3635 | /* Allocate the mapping from statements to candidate indices. */ | |
06ecf488 | 3636 | stmt_cand_map = new hash_map<gimple, slsr_cand_t>; |
4502f5d0 | 3637 | |
3638 | /* Create the obstack where candidate chains will reside. */ | |
3639 | gcc_obstack_init (&chain_obstack); | |
3640 | ||
c99e471e | 3641 | /* Allocate the mapping from base expressions to candidate chains. */ |
c1f445d2 | 3642 | base_cand_map = new hash_table<cand_chain_hasher> (500); |
4502f5d0 | 3643 | |
0ca43779 | 3644 | /* Allocate the mapping from bases to alternative bases. */ |
06ecf488 | 3645 | alt_base_map = new hash_map<tree, tree>; |
0ca43779 | 3646 | |
4502f5d0 | 3647 | /* Initialize the loop optimizer. We need to detect flow across |
3648 | back edges, and this gives us dominator information as well. */ | |
3649 | loop_optimizer_init (AVOID_CFG_MODIFICATIONS); | |
3650 | ||
4502f5d0 | 3651 | /* Walk the CFG in predominator order looking for strength reduction |
3652 | candidates. */ | |
54c91640 | 3653 | find_candidates_dom_walker (CDI_DOMINATORS) |
65b0537f | 3654 | .walk (fun->cfg->x_entry_block_ptr); |
4502f5d0 | 3655 | |
3656 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3657 | { | |
3658 | dump_cand_vec (); | |
3659 | dump_cand_chains (); | |
3660 | } | |
3661 | ||
06ecf488 | 3662 | delete alt_base_map; |
0ca43779 | 3663 | free_affine_expand_cache (&name_expansions); |
3664 | ||
4502f5d0 | 3665 | /* Analyze costs and make appropriate replacements. */ |
3666 | analyze_candidates_and_replace (); | |
3667 | ||
4502f5d0 | 3668 | loop_optimizer_finalize (); |
c1f445d2 | 3669 | delete base_cand_map; |
3670 | base_cand_map = NULL; | |
4502f5d0 | 3671 | obstack_free (&chain_obstack, NULL); |
06ecf488 | 3672 | delete stmt_cand_map; |
f1f41a6c | 3673 | cand_vec.release (); |
4502f5d0 | 3674 | obstack_free (&cand_obstack, NULL); |
4502f5d0 | 3675 | |
3676 | return 0; | |
3677 | } | |
3678 | ||
cbe8bda8 | 3679 | } // anon namespace |
3680 | ||
3681 | gimple_opt_pass * | |
3682 | make_pass_strength_reduction (gcc::context *ctxt) | |
3683 | { | |
3684 | return new pass_strength_reduction (ctxt); | |
3685 | } |