]>
Commit | Line | Data |
---|---|---|
f9453c07 | 1 | /* Straight-line strength reduction. |
a5544970 | 2 | Copyright (C) 2012-2019 Free Software Foundation, Inc. |
f9453c07 BS |
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 | ||
9b92d12b BS |
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. | |
f9453c07 BS |
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" | |
c7131fb2 | 39 | #include "backend.h" |
957060b5 | 40 | #include "rtl.h" |
40e23961 | 41 | #include "tree.h" |
c7131fb2 | 42 | #include "gimple.h" |
957060b5 AM |
43 | #include "cfghooks.h" |
44 | #include "tree-pass.h" | |
c7131fb2 | 45 | #include "ssa.h" |
957060b5 | 46 | #include "expmed.h" |
957060b5 | 47 | #include "gimple-pretty-print.h" |
40e23961 | 48 | #include "fold-const.h" |
5be5c238 | 49 | #include "gimple-iterator.h" |
18f429e2 | 50 | #include "gimplify-me.h" |
d8a2d370 | 51 | #include "stor-layout.h" |
f9453c07 | 52 | #include "cfgloop.h" |
442b4905 | 53 | #include "tree-cfg.h" |
f9453c07 | 54 | #include "domwalk.h" |
ccdbfe93 | 55 | #include "params.h" |
4484a35a | 56 | #include "tree-ssa-address.h" |
96d75a2c | 57 | #include "tree-affine.h" |
bef52a68 | 58 | #include "tree-eh.h" |
9b2b7279 | 59 | #include "builtins.h" |
f9453c07 BS |
60 | \f |
61 | /* Information about a strength reduction candidate. Each statement | |
62 | in the candidate table represents an expression of one of the | |
63 | following forms (the special case of CAND_REF will be described | |
64 | later): | |
65 | ||
66 | (CAND_MULT) S1: X = (B + i) * S | |
67 | (CAND_ADD) S1: X = B + (i * S) | |
68 | ||
69 | Here X and B are SSA names, i is an integer constant, and S is | |
70 | either an SSA name or a constant. We call B the "base," i the | |
71 | "index", and S the "stride." | |
72 | ||
73 | Any statement S0 that dominates S1 and is of the form: | |
74 | ||
75 | (CAND_MULT) S0: Y = (B + i') * S | |
76 | (CAND_ADD) S0: Y = B + (i' * S) | |
77 | ||
78 | is called a "basis" for S1. In both cases, S1 may be replaced by | |
79 | ||
80 | S1': X = Y + (i - i') * S, | |
81 | ||
82 | where (i - i') * S is folded to the extent possible. | |
83 | ||
84 | All gimple statements are visited in dominator order, and each | |
85 | statement that may contribute to one of the forms of S1 above is | |
86 | given at least one entry in the candidate table. Such statements | |
87 | include addition, pointer addition, subtraction, multiplication, | |
88 | negation, copies, and nontrivial type casts. If a statement may | |
89 | represent more than one expression of the forms of S1 above, | |
90 | multiple "interpretations" are stored in the table and chained | |
91 | together. Examples: | |
92 | ||
93 | * An add of two SSA names may treat either operand as the base. | |
94 | * A multiply of two SSA names, likewise. | |
95 | * A copy or cast may be thought of as either a CAND_MULT with | |
96 | i = 0 and S = 1, or as a CAND_ADD with i = 0 or S = 0. | |
97 | ||
98 | Candidate records are allocated from an obstack. They are addressed | |
99 | both from a hash table keyed on S1, and from a vector of candidate | |
100 | pointers arranged in predominator order. | |
101 | ||
102 | Opportunity note | |
103 | ---------------- | |
104 | Currently we don't recognize: | |
105 | ||
106 | S0: Y = (S * i') - B | |
107 | S1: X = (S * i) - B | |
108 | ||
109 | as a strength reduction opportunity, even though this S1 would | |
110 | also be replaceable by the S1' above. This can be added if it | |
2749c8f6 BS |
111 | comes up in practice. |
112 | ||
113 | Strength reduction in addressing | |
114 | -------------------------------- | |
115 | There is another kind of candidate known as CAND_REF. A CAND_REF | |
116 | describes a statement containing a memory reference having | |
117 | complex addressing that might benefit from strength reduction. | |
118 | Specifically, we are interested in references for which | |
119 | get_inner_reference returns a base address, offset, and bitpos as | |
120 | follows: | |
121 | ||
122 | base: MEM_REF (T1, C1) | |
123 | offset: MULT_EXPR (PLUS_EXPR (T2, C2), C3) | |
124 | bitpos: C4 * BITS_PER_UNIT | |
125 | ||
126 | Here T1 and T2 are arbitrary trees, and C1, C2, C3, C4 are | |
127 | arbitrary integer constants. Note that C2 may be zero, in which | |
128 | case the offset will be MULT_EXPR (T2, C3). | |
129 | ||
130 | When this pattern is recognized, the original memory reference | |
131 | can be replaced with: | |
132 | ||
133 | MEM_REF (POINTER_PLUS_EXPR (T1, MULT_EXPR (T2, C3)), | |
134 | C1 + (C2 * C3) + C4) | |
135 | ||
136 | which distributes the multiply to allow constant folding. When | |
137 | two or more addressing expressions can be represented by MEM_REFs | |
138 | of this form, differing only in the constants C1, C2, and C4, | |
139 | making this substitution produces more efficient addressing during | |
140 | the RTL phases. When there are not at least two expressions with | |
141 | the same values of T1, T2, and C3, there is nothing to be gained | |
142 | by the replacement. | |
143 | ||
144 | Strength reduction of CAND_REFs uses the same infrastructure as | |
145 | that used by CAND_MULTs and CAND_ADDs. We record T1 in the base (B) | |
146 | field, MULT_EXPR (T2, C3) in the stride (S) field, and | |
147 | C1 + (C2 * C3) + C4 in the index (i) field. A basis for a CAND_REF | |
148 | is thus another CAND_REF with the same B and S values. When at | |
149 | least two CAND_REFs are chained together using the basis relation, | |
150 | each of them is replaced as above, resulting in improved code | |
9b92d12b BS |
151 | generation for addressing. |
152 | ||
153 | Conditional candidates | |
154 | ====================== | |
155 | ||
156 | Conditional candidates are best illustrated with an example. | |
157 | Consider the code sequence: | |
158 | ||
159 | (1) x_0 = ...; | |
160 | (2) a_0 = x_0 * 5; MULT (B: x_0; i: 0; S: 5) | |
161 | if (...) | |
162 | (3) x_1 = x_0 + 1; ADD (B: x_0, i: 1; S: 1) | |
163 | (4) x_2 = PHI <x_0, x_1>; PHI (B: x_0, i: 0, S: 1) | |
164 | (5) x_3 = x_2 + 1; ADD (B: x_2, i: 1, S: 1) | |
165 | (6) a_1 = x_3 * 5; MULT (B: x_2, i: 1; S: 5) | |
166 | ||
167 | Here strength reduction is complicated by the uncertain value of x_2. | |
168 | A legitimate transformation is: | |
169 | ||
170 | (1) x_0 = ...; | |
171 | (2) a_0 = x_0 * 5; | |
172 | if (...) | |
173 | { | |
174 | (3) [x_1 = x_0 + 1;] | |
175 | (3a) t_1 = a_0 + 5; | |
176 | } | |
177 | (4) [x_2 = PHI <x_0, x_1>;] | |
178 | (4a) t_2 = PHI <a_0, t_1>; | |
179 | (5) [x_3 = x_2 + 1;] | |
180 | (6r) a_1 = t_2 + 5; | |
181 | ||
182 | where the bracketed instructions may go dead. | |
183 | ||
184 | To recognize this opportunity, we have to observe that statement (6) | |
185 | has a "hidden basis" (2). The hidden basis is unlike a normal basis | |
186 | in that the statement and the hidden basis have different base SSA | |
187 | names (x_2 and x_0, respectively). The relationship is established | |
188 | when a statement's base name (x_2) is defined by a phi statement (4), | |
189 | each argument of which (x_0, x_1) has an identical "derived base name." | |
190 | If the argument is defined by a candidate (as x_1 is by (3)) that is a | |
191 | CAND_ADD having a stride of 1, the derived base name of the argument is | |
192 | the base name of the candidate (x_0). Otherwise, the argument itself | |
193 | is its derived base name (as is the case with argument x_0). | |
194 | ||
195 | The hidden basis for statement (6) is the nearest dominating candidate | |
196 | whose base name is the derived base name (x_0) of the feeding phi (4), | |
197 | and whose stride is identical to that of the statement. We can then | |
198 | create the new "phi basis" (4a) and feeding adds along incoming arcs (3a), | |
199 | allowing the final replacement of (6) by the strength-reduced (6r). | |
200 | ||
201 | To facilitate this, a new kind of candidate (CAND_PHI) is introduced. | |
202 | A CAND_PHI is not a candidate for replacement, but is maintained in the | |
203 | candidate table to ease discovery of hidden bases. Any phi statement | |
204 | whose arguments share a common derived base name is entered into the | |
205 | table with the derived base name, an (arbitrary) index of zero, and a | |
206 | stride of 1. A statement with a hidden basis can then be detected by | |
207 | simply looking up its feeding phi definition in the candidate table, | |
208 | extracting the derived base name, and searching for a basis in the | |
209 | usual manner after substituting the derived base name. | |
210 | ||
211 | Note that the transformation is only valid when the original phi and | |
212 | the statements that define the phi's arguments are all at the same | |
213 | position in the loop hierarchy. */ | |
f9453c07 BS |
214 | |
215 | ||
216 | /* Index into the candidate vector, offset by 1. VECs are zero-based, | |
217 | while cand_idx's are one-based, with zero indicating null. */ | |
218 | typedef unsigned cand_idx; | |
219 | ||
220 | /* The kind of candidate. */ | |
221 | enum cand_kind | |
222 | { | |
223 | CAND_MULT, | |
2749c8f6 | 224 | CAND_ADD, |
9b92d12b BS |
225 | CAND_REF, |
226 | CAND_PHI | |
f9453c07 BS |
227 | }; |
228 | ||
229 | struct slsr_cand_d | |
230 | { | |
231 | /* The candidate statement S1. */ | |
355fe088 | 232 | gimple *cand_stmt; |
f9453c07 | 233 | |
3cfd4469 BS |
234 | /* The base expression B: often an SSA name, but not always. */ |
235 | tree base_expr; | |
f9453c07 BS |
236 | |
237 | /* The stride S. */ | |
238 | tree stride; | |
239 | ||
240 | /* The index constant i. */ | |
807e902e | 241 | widest_int index; |
f9453c07 | 242 | |
3cfd4469 | 243 | /* The type of the candidate. This is normally the type of base_expr, |
f9453c07 | 244 | but casts may have occurred when combining feeding instructions. |
2749c8f6 BS |
245 | A candidate can only be a basis for candidates of the same final type. |
246 | (For CAND_REFs, this is the type to be used for operand 1 of the | |
247 | replacement MEM_REF.) */ | |
f9453c07 BS |
248 | tree cand_type; |
249 | ||
0b56e9ad BS |
250 | /* The type to be used to interpret the stride field when the stride |
251 | is not a constant. Normally the same as the type of the recorded | |
252 | stride, but when the stride has been cast we need to maintain that | |
253 | knowledge in order to make legal substitutions without losing | |
254 | precision. When the stride is a constant, this will be sizetype. */ | |
255 | tree stride_type; | |
256 | ||
f9453c07 BS |
257 | /* The kind of candidate (CAND_MULT, etc.). */ |
258 | enum cand_kind kind; | |
259 | ||
260 | /* Index of this candidate in the candidate vector. */ | |
261 | cand_idx cand_num; | |
262 | ||
263 | /* Index of the next candidate record for the same statement. | |
264 | A statement may be useful in more than one way (e.g., due to | |
265 | commutativity). So we can have multiple "interpretations" | |
266 | of a statement. */ | |
267 | cand_idx next_interp; | |
268 | ||
d71dc990 BS |
269 | /* Index of the first candidate record in a chain for the same |
270 | statement. */ | |
271 | cand_idx first_interp; | |
272 | ||
f9453c07 BS |
273 | /* Index of the basis statement S0, if any, in the candidate vector. */ |
274 | cand_idx basis; | |
275 | ||
276 | /* First candidate for which this candidate is a basis, if one exists. */ | |
277 | cand_idx dependent; | |
278 | ||
279 | /* Next candidate having the same basis as this one. */ | |
280 | cand_idx sibling; | |
281 | ||
9b92d12b BS |
282 | /* If this is a conditional candidate, the CAND_PHI candidate |
283 | that defines the base SSA name B. */ | |
284 | cand_idx def_phi; | |
f9453c07 BS |
285 | |
286 | /* Savings that can be expected from eliminating dead code if this | |
287 | candidate is replaced. */ | |
288 | int dead_savings; | |
65d3dce8 BS |
289 | |
290 | /* For PHI candidates, use a visited flag to keep from processing the | |
291 | same PHI twice from multiple paths. */ | |
292 | int visited; | |
293 | ||
294 | /* We sometimes have to cache a phi basis with a phi candidate to | |
295 | avoid processing it twice. Valid only if visited==1. */ | |
296 | tree cached_basis; | |
f9453c07 BS |
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 | |
3cfd4469 | 303 | from base expressions to the candidates that use them. */ |
f9453c07 BS |
304 | |
305 | struct cand_chain_d | |
306 | { | |
3cfd4469 BS |
307 | /* Base expression for the chain of candidates: often, but not |
308 | always, an SSA name. */ | |
309 | tree base_expr; | |
f9453c07 BS |
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 | ||
88ca9ea1 BS |
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. */ | |
807e902e | 335 | widest_int incr; |
88ca9ea1 BS |
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 | ||
f9453c07 BS |
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. */ | |
9771b263 | 360 | static vec<slsr_cand_t> cand_vec; |
f9453c07 BS |
361 | |
362 | enum cost_consts | |
363 | { | |
364 | COST_NEUTRAL = 0, | |
365 | COST_INFINITE = 1000 | |
366 | }; | |
367 | ||
9b92d12b BS |
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 | }; | |
65d3dce8 BS |
385 | |
386 | /* Constrain how many PHI nodes we will visit for a conditional | |
387 | candidate (depth and breadth). */ | |
388 | const int MAX_SPREAD = 16; | |
389 | ||
f9453c07 | 390 | /* Pointer map embodying a mapping from statements to candidates. */ |
355fe088 | 391 | static hash_map<gimple *, slsr_cand_t> *stmt_cand_map; |
f9453c07 BS |
392 | |
393 | /* Obstack for candidates. */ | |
394 | static struct obstack cand_obstack; | |
395 | ||
f9453c07 BS |
396 | /* Obstack for candidate chains. */ |
397 | static struct obstack chain_obstack; | |
88ca9ea1 BS |
398 | |
399 | /* An array INCR_VEC of incr_infos is used during analysis of related | |
400 | candidates having an SSA name for a stride. INCR_VEC_LEN describes | |
7bf55a70 BS |
401 | its current length. MAX_INCR_VEC_LEN is used to avoid costly |
402 | pathological cases. */ | |
88ca9ea1 BS |
403 | static incr_info_t incr_vec; |
404 | static unsigned incr_vec_len; | |
7bf55a70 | 405 | const int MAX_INCR_VEC_LEN = 16; |
88ca9ea1 BS |
406 | |
407 | /* For a chain of candidates with unknown stride, indicates whether or not | |
408 | we must generate pointer arithmetic when replacing statements. */ | |
409 | static bool address_arithmetic_p; | |
9b92d12b BS |
410 | |
411 | /* Forward function declarations. */ | |
412 | static slsr_cand_t base_cand_from_table (tree); | |
a7a7d10e | 413 | static tree introduce_cast_before_cand (slsr_cand_t, tree, tree); |
0916f876 | 414 | static bool legal_cast_p_1 (tree, tree); |
f9453c07 BS |
415 | \f |
416 | /* Produce a pointer to the IDX'th candidate in the candidate vector. */ | |
417 | ||
418 | static slsr_cand_t | |
419 | lookup_cand (cand_idx idx) | |
420 | { | |
98aaa1a6 | 421 | return cand_vec[idx]; |
f9453c07 BS |
422 | } |
423 | ||
4a8fb1a1 | 424 | /* Helper for hashing a candidate chain header. */ |
2749c8f6 | 425 | |
8d67ee55 | 426 | struct cand_chain_hasher : nofree_ptr_hash <cand_chain> |
2749c8f6 | 427 | { |
67f58944 TS |
428 | static inline hashval_t hash (const cand_chain *); |
429 | static inline bool equal (const cand_chain *, const cand_chain *); | |
4a8fb1a1 | 430 | }; |
2749c8f6 | 431 | |
4a8fb1a1 | 432 | inline hashval_t |
67f58944 | 433 | cand_chain_hasher::hash (const cand_chain *p) |
2749c8f6 | 434 | { |
4a8fb1a1 LC |
435 | tree base_expr = p->base_expr; |
436 | return iterative_hash_expr (base_expr, 0); | |
2749c8f6 BS |
437 | } |
438 | ||
4a8fb1a1 | 439 | inline bool |
67f58944 | 440 | cand_chain_hasher::equal (const cand_chain *chain1, const cand_chain *chain2) |
2749c8f6 | 441 | { |
3cfd4469 | 442 | return operand_equal_p (chain1->base_expr, chain2->base_expr, 0); |
2749c8f6 | 443 | } |
4a8fb1a1 LC |
444 | |
445 | /* Hash table embodying a mapping from base exprs to chains of candidates. */ | |
c203e8a7 | 446 | static hash_table<cand_chain_hasher> *base_cand_map; |
2749c8f6 | 447 | \f |
96d75a2c | 448 | /* Pointer map used by tree_to_aff_combination_expand. */ |
39c8aaa4 | 449 | static hash_map<tree, name_expansion *> *name_expansions; |
96d75a2c | 450 | /* Pointer map embodying a mapping from bases to alternative bases. */ |
b787e7a2 | 451 | static hash_map<tree, tree> *alt_base_map; |
96d75a2c YZ |
452 | |
453 | /* Given BASE, use the tree affine combiniation facilities to | |
454 | find the underlying tree expression for BASE, with any | |
8fde427f YZ |
455 | immediate offset excluded. |
456 | ||
457 | N.B. we should eliminate this backtracking with better forward | |
458 | analysis in a future release. */ | |
96d75a2c YZ |
459 | |
460 | static tree | |
461 | get_alternative_base (tree base) | |
462 | { | |
b787e7a2 | 463 | tree *result = alt_base_map->get (base); |
96d75a2c YZ |
464 | |
465 | if (result == NULL) | |
466 | { | |
467 | tree expr; | |
468 | aff_tree aff; | |
469 | ||
470 | tree_to_aff_combination_expand (base, TREE_TYPE (base), | |
471 | &aff, &name_expansions); | |
807e902e | 472 | aff.offset = 0; |
96d75a2c YZ |
473 | expr = aff_combination_to_tree (&aff); |
474 | ||
b787e7a2 | 475 | gcc_assert (!alt_base_map->put (base, base == expr ? NULL : expr)); |
96d75a2c | 476 | |
b787e7a2 | 477 | return expr == base ? NULL : expr; |
96d75a2c YZ |
478 | } |
479 | ||
480 | return *result; | |
481 | } | |
482 | ||
9b92d12b | 483 | /* Look in the candidate table for a CAND_PHI that defines BASE and |
8b28cf47 | 484 | return it if found; otherwise return NULL. */ |
f9453c07 | 485 | |
9b92d12b | 486 | static cand_idx |
8b28cf47 | 487 | find_phi_def (tree base) |
9b92d12b BS |
488 | { |
489 | slsr_cand_t c; | |
490 | ||
491 | if (TREE_CODE (base) != SSA_NAME) | |
492 | return 0; | |
29105868 | 493 | |
9b92d12b BS |
494 | c = base_cand_from_table (base); |
495 | ||
e75210d6 BS |
496 | if (!c || c->kind != CAND_PHI |
497 | || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (gimple_phi_result (c->cand_stmt))) | |
9b92d12b BS |
498 | return 0; |
499 | ||
500 | return c->cand_num; | |
501 | } | |
502 | ||
df11b2ea BS |
503 | /* Determine whether all uses of NAME are directly or indirectly |
504 | used by STMT. That is, we want to know whether if STMT goes | |
505 | dead, the definition of NAME also goes dead. */ | |
506 | static bool | |
507 | uses_consumed_by_stmt (tree name, gimple *stmt, unsigned recurse = 0) | |
508 | { | |
509 | gimple *use_stmt; | |
510 | imm_use_iterator iter; | |
511 | bool retval = true; | |
512 | ||
513 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, name) | |
514 | { | |
515 | if (use_stmt == stmt || is_gimple_debug (use_stmt)) | |
516 | continue; | |
517 | ||
518 | if (!is_gimple_assign (use_stmt) | |
519 | || !gimple_get_lhs (use_stmt) | |
520 | || !is_gimple_reg (gimple_get_lhs (use_stmt)) | |
521 | || recurse >= 10 | |
522 | || !uses_consumed_by_stmt (gimple_get_lhs (use_stmt), stmt, | |
523 | recurse + 1)) | |
524 | { | |
525 | retval = false; | |
526 | BREAK_FROM_IMM_USE_STMT (iter); | |
527 | } | |
528 | } | |
529 | ||
530 | return retval; | |
531 | } | |
532 | ||
9b92d12b | 533 | /* Helper routine for find_basis_for_candidate. May be called twice: |
96d75a2c YZ |
534 | once for the candidate's base expr, and optionally again either for |
535 | the candidate's phi definition or for a CAND_REF's alternative base | |
536 | expression. */ | |
9b92d12b BS |
537 | |
538 | static slsr_cand_t | |
539 | find_basis_for_base_expr (slsr_cand_t c, tree base_expr) | |
f9453c07 | 540 | { |
2749c8f6 | 541 | cand_chain mapping_key; |
f9453c07 BS |
542 | cand_chain_t chain; |
543 | slsr_cand_t basis = NULL; | |
544 | ||
ccdbfe93 BS |
545 | // Limit potential of N^2 behavior for long candidate chains. |
546 | int iters = 0; | |
547 | int max_iters = PARAM_VALUE (PARAM_MAX_SLSR_CANDIDATE_SCAN); | |
548 | ||
9b92d12b | 549 | mapping_key.base_expr = base_expr; |
c203e8a7 | 550 | chain = base_cand_map->find (&mapping_key); |
f9453c07 | 551 | |
ccdbfe93 | 552 | for (; chain && iters < max_iters; chain = chain->next, ++iters) |
f9453c07 BS |
553 | { |
554 | slsr_cand_t one_basis = chain->cand; | |
555 | ||
556 | if (one_basis->kind != c->kind | |
69e1a1a3 | 557 | || one_basis->cand_stmt == c->cand_stmt |
f9453c07 BS |
558 | || !operand_equal_p (one_basis->stride, c->stride, 0) |
559 | || !types_compatible_p (one_basis->cand_type, c->cand_type) | |
0b56e9ad | 560 | || !types_compatible_p (one_basis->stride_type, c->stride_type) |
f9453c07 BS |
561 | || !dominated_by_p (CDI_DOMINATORS, |
562 | gimple_bb (c->cand_stmt), | |
563 | gimple_bb (one_basis->cand_stmt))) | |
564 | continue; | |
565 | ||
e75210d6 BS |
566 | tree lhs = gimple_assign_lhs (one_basis->cand_stmt); |
567 | if (lhs && TREE_CODE (lhs) == SSA_NAME | |
568 | && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs)) | |
569 | continue; | |
570 | ||
f9453c07 BS |
571 | if (!basis || basis->cand_num < one_basis->cand_num) |
572 | basis = one_basis; | |
573 | } | |
574 | ||
9b92d12b BS |
575 | return basis; |
576 | } | |
577 | ||
578 | /* Use the base expr from candidate C to look for possible candidates | |
579 | that can serve as a basis for C. Each potential basis must also | |
580 | appear in a block that dominates the candidate statement and have | |
581 | the same stride and type. If more than one possible basis exists, | |
582 | the one with highest index in the vector is chosen; this will be | |
583 | the most immediately dominating basis. */ | |
584 | ||
585 | static int | |
586 | find_basis_for_candidate (slsr_cand_t c) | |
587 | { | |
588 | slsr_cand_t basis = find_basis_for_base_expr (c, c->base_expr); | |
589 | ||
590 | /* If a candidate doesn't have a basis using its base expression, | |
591 | it may have a basis hidden by one or more intervening phis. */ | |
592 | if (!basis && c->def_phi) | |
593 | { | |
594 | basic_block basis_bb, phi_bb; | |
595 | slsr_cand_t phi_cand = lookup_cand (c->def_phi); | |
596 | basis = find_basis_for_base_expr (c, phi_cand->base_expr); | |
597 | ||
598 | if (basis) | |
599 | { | |
600 | /* A hidden basis must dominate the phi-definition of the | |
601 | candidate's base name. */ | |
602 | phi_bb = gimple_bb (phi_cand->cand_stmt); | |
603 | basis_bb = gimple_bb (basis->cand_stmt); | |
604 | ||
605 | if (phi_bb == basis_bb | |
606 | || !dominated_by_p (CDI_DOMINATORS, phi_bb, basis_bb)) | |
607 | { | |
608 | basis = NULL; | |
609 | c->basis = 0; | |
610 | } | |
611 | ||
612 | /* If we found a hidden basis, estimate additional dead-code | |
613 | savings if the phi and its feeding statements can be removed. */ | |
df11b2ea BS |
614 | tree feeding_var = gimple_phi_result (phi_cand->cand_stmt); |
615 | if (basis && uses_consumed_by_stmt (feeding_var, c->cand_stmt)) | |
9b92d12b BS |
616 | c->dead_savings += phi_cand->dead_savings; |
617 | } | |
618 | } | |
619 | ||
8fde427f | 620 | if (flag_expensive_optimizations && !basis && c->kind == CAND_REF) |
96d75a2c YZ |
621 | { |
622 | tree alt_base_expr = get_alternative_base (c->base_expr); | |
623 | if (alt_base_expr) | |
624 | basis = find_basis_for_base_expr (c, alt_base_expr); | |
625 | } | |
626 | ||
f9453c07 BS |
627 | if (basis) |
628 | { | |
629 | c->sibling = basis->dependent; | |
630 | basis->dependent = c->cand_num; | |
631 | return basis->cand_num; | |
632 | } | |
633 | ||
634 | return 0; | |
635 | } | |
636 | ||
96d75a2c YZ |
637 | /* Record a mapping from BASE to C, indicating that C may potentially serve |
638 | as a basis using that base expression. BASE may be the same as | |
639 | C->BASE_EXPR; alternatively BASE can be a different tree that share the | |
640 | underlining expression of C->BASE_EXPR. */ | |
f9453c07 BS |
641 | |
642 | static void | |
96d75a2c | 643 | record_potential_basis (slsr_cand_t c, tree base) |
f9453c07 | 644 | { |
2749c8f6 | 645 | cand_chain_t node; |
4a8fb1a1 | 646 | cand_chain **slot; |
f9453c07 | 647 | |
96d75a2c YZ |
648 | gcc_assert (base); |
649 | ||
f9453c07 | 650 | node = (cand_chain_t) obstack_alloc (&chain_obstack, sizeof (cand_chain)); |
96d75a2c | 651 | node->base_expr = base; |
f9453c07 BS |
652 | node->cand = c; |
653 | node->next = NULL; | |
c203e8a7 | 654 | slot = base_cand_map->find_slot (node, INSERT); |
f9453c07 | 655 | |
2749c8f6 | 656 | if (*slot) |
f9453c07 | 657 | { |
2749c8f6 | 658 | cand_chain_t head = (cand_chain_t) (*slot); |
f9453c07 BS |
659 | node->next = head->next; |
660 | head->next = node; | |
661 | } | |
662 | else | |
2749c8f6 | 663 | *slot = node; |
f9453c07 BS |
664 | } |
665 | ||
666 | /* Allocate storage for a new candidate and initialize its fields. | |
96d75a2c YZ |
667 | Attempt to find a basis for the candidate. |
668 | ||
669 | For CAND_REF, an alternative base may also be recorded and used | |
670 | to find a basis. This helps cases where the expression hidden | |
671 | behind BASE (which is usually an SSA_NAME) has immediate offset, | |
672 | e.g. | |
673 | ||
674 | a2[i][j] = 1; | |
675 | a2[i + 20][j] = 2; */ | |
f9453c07 BS |
676 | |
677 | static slsr_cand_t | |
355fe088 | 678 | alloc_cand_and_find_basis (enum cand_kind kind, gimple *gs, tree base, |
807e902e | 679 | const widest_int &index, tree stride, tree ctype, |
0b56e9ad | 680 | tree stype, unsigned savings) |
f9453c07 BS |
681 | { |
682 | slsr_cand_t c = (slsr_cand_t) obstack_alloc (&cand_obstack, | |
683 | sizeof (slsr_cand)); | |
684 | c->cand_stmt = gs; | |
3cfd4469 | 685 | c->base_expr = base; |
f9453c07 BS |
686 | c->stride = stride; |
687 | c->index = index; | |
688 | c->cand_type = ctype; | |
0b56e9ad | 689 | c->stride_type = stype; |
f9453c07 | 690 | c->kind = kind; |
98aaa1a6 | 691 | c->cand_num = cand_vec.length (); |
f9453c07 | 692 | c->next_interp = 0; |
d71dc990 | 693 | c->first_interp = c->cand_num; |
f9453c07 BS |
694 | c->dependent = 0; |
695 | c->sibling = 0; | |
8b28cf47 | 696 | c->def_phi = kind == CAND_MULT ? find_phi_def (base) : 0; |
f9453c07 | 697 | c->dead_savings = savings; |
65d3dce8 BS |
698 | c->visited = 0; |
699 | c->cached_basis = NULL_TREE; | |
f9453c07 | 700 | |
9771b263 | 701 | cand_vec.safe_push (c); |
9b92d12b BS |
702 | |
703 | if (kind == CAND_PHI) | |
704 | c->basis = 0; | |
705 | else | |
706 | c->basis = find_basis_for_candidate (c); | |
707 | ||
96d75a2c | 708 | record_potential_basis (c, base); |
8fde427f | 709 | if (flag_expensive_optimizations && kind == CAND_REF) |
96d75a2c YZ |
710 | { |
711 | tree alt_base = get_alternative_base (base); | |
712 | if (alt_base) | |
713 | record_potential_basis (c, alt_base); | |
714 | } | |
f9453c07 BS |
715 | |
716 | return c; | |
717 | } | |
718 | ||
719 | /* Determine the target cost of statement GS when compiling according | |
720 | to SPEED. */ | |
721 | ||
722 | static int | |
355fe088 | 723 | stmt_cost (gimple *gs, bool speed) |
f9453c07 BS |
724 | { |
725 | tree lhs, rhs1, rhs2; | |
ef4bddc2 | 726 | machine_mode lhs_mode; |
f9453c07 BS |
727 | |
728 | gcc_assert (is_gimple_assign (gs)); | |
729 | lhs = gimple_assign_lhs (gs); | |
730 | rhs1 = gimple_assign_rhs1 (gs); | |
731 | lhs_mode = TYPE_MODE (TREE_TYPE (lhs)); | |
732 | ||
733 | switch (gimple_assign_rhs_code (gs)) | |
734 | { | |
735 | case MULT_EXPR: | |
736 | rhs2 = gimple_assign_rhs2 (gs); | |
737 | ||
9541ffee | 738 | if (tree_fits_shwi_p (rhs2)) |
eb1ce453 | 739 | return mult_by_coeff_cost (tree_to_shwi (rhs2), lhs_mode, speed); |
f9453c07 BS |
740 | |
741 | gcc_assert (TREE_CODE (rhs1) != INTEGER_CST); | |
5322d07e | 742 | return mul_cost (speed, lhs_mode); |
f9453c07 BS |
743 | |
744 | case PLUS_EXPR: | |
745 | case POINTER_PLUS_EXPR: | |
746 | case MINUS_EXPR: | |
5322d07e | 747 | return add_cost (speed, lhs_mode); |
f9453c07 BS |
748 | |
749 | case NEGATE_EXPR: | |
5322d07e | 750 | return neg_cost (speed, lhs_mode); |
f9453c07 | 751 | |
d822570f | 752 | CASE_CONVERT: |
6dd8f4bb | 753 | return convert_cost (lhs_mode, TYPE_MODE (TREE_TYPE (rhs1)), speed); |
f9453c07 BS |
754 | |
755 | /* Note that we don't assign costs to copies that in most cases | |
756 | will go away. */ | |
fbcdc43e BS |
757 | case SSA_NAME: |
758 | return 0; | |
759 | ||
f9453c07 BS |
760 | default: |
761 | ; | |
762 | } | |
763 | ||
764 | gcc_unreachable (); | |
765 | return 0; | |
766 | } | |
767 | ||
768 | /* Look up the defining statement for BASE_IN and return a pointer | |
769 | to its candidate in the candidate table, if any; otherwise NULL. | |
770 | Only CAND_ADD and CAND_MULT candidates are returned. */ | |
771 | ||
772 | static slsr_cand_t | |
773 | base_cand_from_table (tree base_in) | |
774 | { | |
775 | slsr_cand_t *result; | |
776 | ||
355fe088 | 777 | gimple *def = SSA_NAME_DEF_STMT (base_in); |
f9453c07 BS |
778 | if (!def) |
779 | return (slsr_cand_t) NULL; | |
780 | ||
b787e7a2 | 781 | result = stmt_cand_map->get (def); |
2749c8f6 BS |
782 | |
783 | if (result && (*result)->kind != CAND_REF) | |
784 | return *result; | |
f9453c07 | 785 | |
2749c8f6 | 786 | return (slsr_cand_t) NULL; |
f9453c07 BS |
787 | } |
788 | ||
789 | /* Add an entry to the statement-to-candidate mapping. */ | |
790 | ||
791 | static void | |
355fe088 | 792 | add_cand_for_stmt (gimple *gs, slsr_cand_t c) |
f9453c07 | 793 | { |
b787e7a2 | 794 | gcc_assert (!stmt_cand_map->put (gs, c)); |
f9453c07 BS |
795 | } |
796 | \f | |
9b92d12b BS |
797 | /* Given PHI which contains a phi statement, determine whether it |
798 | satisfies all the requirements of a phi candidate. If so, create | |
799 | a candidate. Note that a CAND_PHI never has a basis itself, but | |
800 | is used to help find a basis for subsequent candidates. */ | |
801 | ||
802 | static void | |
538dd0b7 | 803 | slsr_process_phi (gphi *phi, bool speed) |
9b92d12b BS |
804 | { |
805 | unsigned i; | |
806 | tree arg0_base = NULL_TREE, base_type; | |
807 | slsr_cand_t c; | |
808 | struct loop *cand_loop = gimple_bb (phi)->loop_father; | |
809 | unsigned savings = 0; | |
810 | ||
811 | /* A CAND_PHI requires each of its arguments to have the same | |
812 | derived base name. (See the module header commentary for a | |
813 | definition of derived base names.) Furthermore, all feeding | |
814 | definitions must be in the same position in the loop hierarchy | |
815 | as PHI. */ | |
816 | ||
817 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
818 | { | |
819 | slsr_cand_t arg_cand; | |
820 | tree arg = gimple_phi_arg_def (phi, i); | |
821 | tree derived_base_name = NULL_TREE; | |
355fe088 | 822 | gimple *arg_stmt = NULL; |
9b92d12b BS |
823 | basic_block arg_bb = NULL; |
824 | ||
825 | if (TREE_CODE (arg) != SSA_NAME) | |
826 | return; | |
827 | ||
828 | arg_cand = base_cand_from_table (arg); | |
829 | ||
830 | if (arg_cand) | |
831 | { | |
832 | while (arg_cand->kind != CAND_ADD && arg_cand->kind != CAND_PHI) | |
833 | { | |
834 | if (!arg_cand->next_interp) | |
835 | return; | |
836 | ||
837 | arg_cand = lookup_cand (arg_cand->next_interp); | |
838 | } | |
839 | ||
840 | if (!integer_onep (arg_cand->stride)) | |
841 | return; | |
842 | ||
843 | derived_base_name = arg_cand->base_expr; | |
844 | arg_stmt = arg_cand->cand_stmt; | |
845 | arg_bb = gimple_bb (arg_stmt); | |
846 | ||
847 | /* Gather potential dead code savings if the phi statement | |
848 | can be removed later on. */ | |
df11b2ea | 849 | if (uses_consumed_by_stmt (arg, phi)) |
9b92d12b BS |
850 | { |
851 | if (gimple_code (arg_stmt) == GIMPLE_PHI) | |
852 | savings += arg_cand->dead_savings; | |
853 | else | |
854 | savings += stmt_cost (arg_stmt, speed); | |
855 | } | |
856 | } | |
6e281ce3 | 857 | else if (SSA_NAME_IS_DEFAULT_DEF (arg)) |
9b92d12b BS |
858 | { |
859 | derived_base_name = arg; | |
6e281ce3 | 860 | arg_bb = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)); |
9b92d12b BS |
861 | } |
862 | ||
863 | if (!arg_bb || arg_bb->loop_father != cand_loop) | |
864 | return; | |
865 | ||
866 | if (i == 0) | |
867 | arg0_base = derived_base_name; | |
868 | else if (!operand_equal_p (derived_base_name, arg0_base, 0)) | |
869 | return; | |
870 | } | |
871 | ||
872 | /* Create the candidate. "alloc_cand_and_find_basis" is named | |
873 | misleadingly for this case, as no basis will be sought for a | |
874 | CAND_PHI. */ | |
875 | base_type = TREE_TYPE (arg0_base); | |
876 | ||
807e902e | 877 | c = alloc_cand_and_find_basis (CAND_PHI, phi, arg0_base, |
0b56e9ad BS |
878 | 0, integer_one_node, base_type, |
879 | sizetype, savings); | |
9b92d12b BS |
880 | |
881 | /* Add the candidate to the statement-candidate mapping. */ | |
882 | add_cand_for_stmt (phi, c); | |
883 | } | |
884 | ||
c068654b BC |
885 | /* Given PBASE which is a pointer to tree, look up the defining |
886 | statement for it and check whether the candidate is in the | |
887 | form of: | |
888 | ||
889 | X = B + (1 * S), S is integer constant | |
890 | X = B + (i * S), S is integer one | |
891 | ||
892 | If so, set PBASE to the candidate's base_expr and return double | |
893 | int (i * S). | |
894 | Otherwise, just return double int zero. */ | |
895 | ||
807e902e | 896 | static widest_int |
c068654b BC |
897 | backtrace_base_for_ref (tree *pbase) |
898 | { | |
899 | tree base_in = *pbase; | |
900 | slsr_cand_t base_cand; | |
901 | ||
902 | STRIP_NOPS (base_in); | |
0916f876 YZ |
903 | |
904 | /* Strip off widening conversion(s) to handle cases where | |
905 | e.g. 'B' is widened from an 'int' in order to calculate | |
906 | a 64-bit address. */ | |
907 | if (CONVERT_EXPR_P (base_in) | |
0b56e9ad BS |
908 | && legal_cast_p_1 (TREE_TYPE (base_in), |
909 | TREE_TYPE (TREE_OPERAND (base_in, 0)))) | |
0916f876 YZ |
910 | base_in = get_unwidened (base_in, NULL_TREE); |
911 | ||
c068654b | 912 | if (TREE_CODE (base_in) != SSA_NAME) |
807e902e | 913 | return 0; |
c068654b BC |
914 | |
915 | base_cand = base_cand_from_table (base_in); | |
916 | ||
917 | while (base_cand && base_cand->kind != CAND_PHI) | |
918 | { | |
919 | if (base_cand->kind == CAND_ADD | |
807e902e | 920 | && base_cand->index == 1 |
c068654b BC |
921 | && TREE_CODE (base_cand->stride) == INTEGER_CST) |
922 | { | |
923 | /* X = B + (1 * S), S is integer constant. */ | |
924 | *pbase = base_cand->base_expr; | |
807e902e | 925 | return wi::to_widest (base_cand->stride); |
c068654b BC |
926 | } |
927 | else if (base_cand->kind == CAND_ADD | |
928 | && TREE_CODE (base_cand->stride) == INTEGER_CST | |
929 | && integer_onep (base_cand->stride)) | |
1d2151c6 | 930 | { |
c068654b BC |
931 | /* X = B + (i * S), S is integer one. */ |
932 | *pbase = base_cand->base_expr; | |
933 | return base_cand->index; | |
934 | } | |
935 | ||
98aaa1a6 | 936 | base_cand = lookup_cand (base_cand->next_interp); |
c068654b BC |
937 | } |
938 | ||
807e902e | 939 | return 0; |
c068654b BC |
940 | } |
941 | ||
2749c8f6 BS |
942 | /* Look for the following pattern: |
943 | ||
944 | *PBASE: MEM_REF (T1, C1) | |
945 | ||
946 | *POFFSET: MULT_EXPR (T2, C3) [C2 is zero] | |
947 | or | |
948 | MULT_EXPR (PLUS_EXPR (T2, C2), C3) | |
949 | or | |
950 | MULT_EXPR (MINUS_EXPR (T2, -C2), C3) | |
951 | ||
952 | *PINDEX: C4 * BITS_PER_UNIT | |
953 | ||
954 | If not present, leave the input values unchanged and return FALSE. | |
955 | Otherwise, modify the input values as follows and return TRUE: | |
956 | ||
957 | *PBASE: T1 | |
958 | *POFFSET: MULT_EXPR (T2, C3) | |
c068654b BC |
959 | *PINDEX: C1 + (C2 * C3) + C4 |
960 | ||
961 | When T2 is recorded by a CAND_ADD in the form of (T2' + C5), it | |
962 | will be further restructured to: | |
963 | ||
964 | *PBASE: T1 | |
965 | *POFFSET: MULT_EXPR (T2', C3) | |
966 | *PINDEX: C1 + (C2 * C3) + C4 + (C5 * C3) */ | |
2749c8f6 BS |
967 | |
968 | static bool | |
807e902e | 969 | restructure_reference (tree *pbase, tree *poffset, widest_int *pindex, |
2749c8f6 BS |
970 | tree *ptype) |
971 | { | |
972 | tree base = *pbase, offset = *poffset; | |
807e902e KZ |
973 | widest_int index = *pindex; |
974 | tree mult_op0, t1, t2, type; | |
975 | widest_int c1, c2, c3, c4, c5; | |
aca52e6f | 976 | offset_int mem_offset; |
2749c8f6 BS |
977 | |
978 | if (!base | |
979 | || !offset | |
980 | || TREE_CODE (base) != MEM_REF | |
aca52e6f | 981 | || !mem_ref_offset (base).is_constant (&mem_offset) |
2749c8f6 BS |
982 | || TREE_CODE (offset) != MULT_EXPR |
983 | || TREE_CODE (TREE_OPERAND (offset, 1)) != INTEGER_CST | |
807e902e | 984 | || wi::umod_floor (index, BITS_PER_UNIT) != 0) |
2749c8f6 BS |
985 | return false; |
986 | ||
987 | t1 = TREE_OPERAND (base, 0); | |
aca52e6f | 988 | c1 = widest_int::from (mem_offset, SIGNED); |
2749c8f6 BS |
989 | type = TREE_TYPE (TREE_OPERAND (base, 1)); |
990 | ||
991 | mult_op0 = TREE_OPERAND (offset, 0); | |
807e902e | 992 | c3 = wi::to_widest (TREE_OPERAND (offset, 1)); |
2749c8f6 BS |
993 | |
994 | if (TREE_CODE (mult_op0) == PLUS_EXPR) | |
995 | ||
996 | if (TREE_CODE (TREE_OPERAND (mult_op0, 1)) == INTEGER_CST) | |
997 | { | |
998 | t2 = TREE_OPERAND (mult_op0, 0); | |
807e902e | 999 | c2 = wi::to_widest (TREE_OPERAND (mult_op0, 1)); |
2749c8f6 BS |
1000 | } |
1001 | else | |
1002 | return false; | |
1003 | ||
1004 | else if (TREE_CODE (mult_op0) == MINUS_EXPR) | |
1005 | ||
1006 | if (TREE_CODE (TREE_OPERAND (mult_op0, 1)) == INTEGER_CST) | |
1007 | { | |
1008 | t2 = TREE_OPERAND (mult_op0, 0); | |
807e902e | 1009 | c2 = -wi::to_widest (TREE_OPERAND (mult_op0, 1)); |
2749c8f6 BS |
1010 | } |
1011 | else | |
1012 | return false; | |
1013 | ||
1014 | else | |
1015 | { | |
1016 | t2 = mult_op0; | |
807e902e | 1017 | c2 = 0; |
2749c8f6 BS |
1018 | } |
1019 | ||
8de73453 | 1020 | c4 = index >> LOG2_BITS_PER_UNIT; |
c068654b | 1021 | c5 = backtrace_base_for_ref (&t2); |
2749c8f6 BS |
1022 | |
1023 | *pbase = t1; | |
c068654b | 1024 | *poffset = fold_build2 (MULT_EXPR, sizetype, fold_convert (sizetype, t2), |
807e902e | 1025 | wide_int_to_tree (sizetype, c3)); |
c068654b | 1026 | *pindex = c1 + c2 * c3 + c4 + c5 * c3; |
2749c8f6 BS |
1027 | *ptype = type; |
1028 | ||
1029 | return true; | |
1030 | } | |
1031 | ||
1032 | /* Given GS which contains a data reference, create a CAND_REF entry in | |
1033 | the candidate table and attempt to find a basis. */ | |
1034 | ||
1035 | static void | |
355fe088 | 1036 | slsr_process_ref (gimple *gs) |
2749c8f6 BS |
1037 | { |
1038 | tree ref_expr, base, offset, type; | |
f37fac2b | 1039 | poly_int64 bitsize, bitpos; |
ef4bddc2 | 1040 | machine_mode mode; |
ee45a32d | 1041 | int unsignedp, reversep, volatilep; |
2749c8f6 BS |
1042 | slsr_cand_t c; |
1043 | ||
1044 | if (gimple_vdef (gs)) | |
1045 | ref_expr = gimple_assign_lhs (gs); | |
1046 | else | |
1047 | ref_expr = gimple_assign_rhs1 (gs); | |
1048 | ||
1049 | if (!handled_component_p (ref_expr) | |
1050 | || TREE_CODE (ref_expr) == BIT_FIELD_REF | |
1051 | || (TREE_CODE (ref_expr) == COMPONENT_REF | |
1052 | && DECL_BIT_FIELD (TREE_OPERAND (ref_expr, 1)))) | |
1053 | return; | |
1054 | ||
1055 | base = get_inner_reference (ref_expr, &bitsize, &bitpos, &offset, &mode, | |
25b75a48 | 1056 | &unsignedp, &reversep, &volatilep); |
f37fac2b RS |
1057 | HOST_WIDE_INT cbitpos; |
1058 | if (reversep || !bitpos.is_constant (&cbitpos)) | |
ee45a32d | 1059 | return; |
f37fac2b | 1060 | widest_int index = cbitpos; |
2749c8f6 BS |
1061 | |
1062 | if (!restructure_reference (&base, &offset, &index, &type)) | |
1063 | return; | |
1064 | ||
1065 | c = alloc_cand_and_find_basis (CAND_REF, gs, base, index, offset, | |
0b56e9ad | 1066 | type, sizetype, 0); |
2749c8f6 BS |
1067 | |
1068 | /* Add the candidate to the statement-candidate mapping. */ | |
1069 | add_cand_for_stmt (gs, c); | |
1070 | } | |
1071 | ||
f9453c07 BS |
1072 | /* Create a candidate entry for a statement GS, where GS multiplies |
1073 | two SSA names BASE_IN and STRIDE_IN. Propagate any known information | |
1074 | about the two SSA names into the new candidate. Return the new | |
1075 | candidate. */ | |
1076 | ||
1077 | static slsr_cand_t | |
355fe088 | 1078 | create_mul_ssa_cand (gimple *gs, tree base_in, tree stride_in, bool speed) |
f9453c07 BS |
1079 | { |
1080 | tree base = NULL_TREE, stride = NULL_TREE, ctype = NULL_TREE; | |
0b56e9ad | 1081 | tree stype = NULL_TREE; |
807e902e | 1082 | widest_int index; |
f9453c07 BS |
1083 | unsigned savings = 0; |
1084 | slsr_cand_t c; | |
1085 | slsr_cand_t base_cand = base_cand_from_table (base_in); | |
1086 | ||
1087 | /* Look at all interpretations of the base candidate, if necessary, | |
1088 | to find information to propagate into this candidate. */ | |
9b92d12b | 1089 | while (base_cand && !base && base_cand->kind != CAND_PHI) |
f9453c07 BS |
1090 | { |
1091 | ||
9b92d12b | 1092 | if (base_cand->kind == CAND_MULT && integer_onep (base_cand->stride)) |
f9453c07 BS |
1093 | { |
1094 | /* Y = (B + i') * 1 | |
1095 | X = Y * Z | |
1096 | ================ | |
1097 | X = (B + i') * Z */ | |
3cfd4469 | 1098 | base = base_cand->base_expr; |
f9453c07 BS |
1099 | index = base_cand->index; |
1100 | stride = stride_in; | |
1101 | ctype = base_cand->cand_type; | |
0b56e9ad | 1102 | stype = TREE_TYPE (stride_in); |
f9453c07 BS |
1103 | if (has_single_use (base_in)) |
1104 | savings = (base_cand->dead_savings | |
1105 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1106 | } | |
1107 | else if (base_cand->kind == CAND_ADD | |
1108 | && TREE_CODE (base_cand->stride) == INTEGER_CST) | |
1109 | { | |
1110 | /* Y = B + (i' * S), S constant | |
1111 | X = Y * Z | |
1112 | ============================ | |
1113 | X = B + ((i' * S) * Z) */ | |
3cfd4469 | 1114 | base = base_cand->base_expr; |
807e902e | 1115 | index = base_cand->index * wi::to_widest (base_cand->stride); |
f9453c07 BS |
1116 | stride = stride_in; |
1117 | ctype = base_cand->cand_type; | |
0b56e9ad | 1118 | stype = TREE_TYPE (stride_in); |
f9453c07 BS |
1119 | if (has_single_use (base_in)) |
1120 | savings = (base_cand->dead_savings | |
1121 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1122 | } | |
1123 | ||
98aaa1a6 | 1124 | base_cand = lookup_cand (base_cand->next_interp); |
f9453c07 BS |
1125 | } |
1126 | ||
1127 | if (!base) | |
1128 | { | |
1129 | /* No interpretations had anything useful to propagate, so | |
1130 | produce X = (Y + 0) * Z. */ | |
1131 | base = base_in; | |
807e902e | 1132 | index = 0; |
f9453c07 | 1133 | stride = stride_in; |
b2ec94d4 | 1134 | ctype = TREE_TYPE (base_in); |
0b56e9ad | 1135 | stype = TREE_TYPE (stride_in); |
f9453c07 BS |
1136 | } |
1137 | ||
1138 | c = alloc_cand_and_find_basis (CAND_MULT, gs, base, index, stride, | |
0b56e9ad | 1139 | ctype, stype, savings); |
f9453c07 BS |
1140 | return c; |
1141 | } | |
1142 | ||
1143 | /* Create a candidate entry for a statement GS, where GS multiplies | |
1144 | SSA name BASE_IN by constant STRIDE_IN. Propagate any known | |
1145 | information about BASE_IN into the new candidate. Return the new | |
1146 | candidate. */ | |
1147 | ||
1148 | static slsr_cand_t | |
355fe088 | 1149 | create_mul_imm_cand (gimple *gs, tree base_in, tree stride_in, bool speed) |
f9453c07 BS |
1150 | { |
1151 | tree base = NULL_TREE, stride = NULL_TREE, ctype = NULL_TREE; | |
807e902e | 1152 | widest_int index, temp; |
f9453c07 BS |
1153 | unsigned savings = 0; |
1154 | slsr_cand_t c; | |
1155 | slsr_cand_t base_cand = base_cand_from_table (base_in); | |
1156 | ||
1157 | /* Look at all interpretations of the base candidate, if necessary, | |
1158 | to find information to propagate into this candidate. */ | |
9b92d12b | 1159 | while (base_cand && !base && base_cand->kind != CAND_PHI) |
f9453c07 BS |
1160 | { |
1161 | if (base_cand->kind == CAND_MULT | |
1162 | && TREE_CODE (base_cand->stride) == INTEGER_CST) | |
1163 | { | |
1164 | /* Y = (B + i') * S, S constant | |
1165 | X = Y * c | |
1166 | ============================ | |
1167 | X = (B + i') * (S * c) */ | |
807e902e KZ |
1168 | temp = wi::to_widest (base_cand->stride) * wi::to_widest (stride_in); |
1169 | if (wi::fits_to_tree_p (temp, TREE_TYPE (stride_in))) | |
61ba7329 BS |
1170 | { |
1171 | base = base_cand->base_expr; | |
1172 | index = base_cand->index; | |
807e902e | 1173 | stride = wide_int_to_tree (TREE_TYPE (stride_in), temp); |
61ba7329 BS |
1174 | ctype = base_cand->cand_type; |
1175 | if (has_single_use (base_in)) | |
1176 | savings = (base_cand->dead_savings | |
1177 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1178 | } | |
f9453c07 | 1179 | } |
9b92d12b | 1180 | else if (base_cand->kind == CAND_ADD && integer_onep (base_cand->stride)) |
f9453c07 BS |
1181 | { |
1182 | /* Y = B + (i' * 1) | |
1183 | X = Y * c | |
1184 | =========================== | |
1185 | X = (B + i') * c */ | |
3cfd4469 | 1186 | base = base_cand->base_expr; |
f9453c07 BS |
1187 | index = base_cand->index; |
1188 | stride = stride_in; | |
1189 | ctype = base_cand->cand_type; | |
1190 | if (has_single_use (base_in)) | |
1191 | savings = (base_cand->dead_savings | |
1192 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1193 | } | |
1194 | else if (base_cand->kind == CAND_ADD | |
807e902e | 1195 | && base_cand->index == 1 |
f9453c07 BS |
1196 | && TREE_CODE (base_cand->stride) == INTEGER_CST) |
1197 | { | |
1198 | /* Y = B + (1 * S), S constant | |
1199 | X = Y * c | |
1200 | =========================== | |
1201 | X = (B + S) * c */ | |
3cfd4469 | 1202 | base = base_cand->base_expr; |
807e902e | 1203 | index = wi::to_widest (base_cand->stride); |
f9453c07 BS |
1204 | stride = stride_in; |
1205 | ctype = base_cand->cand_type; | |
1206 | if (has_single_use (base_in)) | |
1207 | savings = (base_cand->dead_savings | |
1208 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1209 | } | |
1210 | ||
98aaa1a6 | 1211 | base_cand = lookup_cand (base_cand->next_interp); |
f9453c07 BS |
1212 | } |
1213 | ||
1214 | if (!base) | |
1215 | { | |
1216 | /* No interpretations had anything useful to propagate, so | |
1217 | produce X = (Y + 0) * c. */ | |
1218 | base = base_in; | |
807e902e | 1219 | index = 0; |
f9453c07 | 1220 | stride = stride_in; |
b2ec94d4 | 1221 | ctype = TREE_TYPE (base_in); |
f9453c07 BS |
1222 | } |
1223 | ||
1224 | c = alloc_cand_and_find_basis (CAND_MULT, gs, base, index, stride, | |
0b56e9ad | 1225 | ctype, sizetype, savings); |
f9453c07 BS |
1226 | return c; |
1227 | } | |
1228 | ||
1229 | /* Given GS which is a multiply of scalar integers, make an appropriate | |
1230 | entry in the candidate table. If this is a multiply of two SSA names, | |
1231 | create two CAND_MULT interpretations and attempt to find a basis for | |
1232 | each of them. Otherwise, create a single CAND_MULT and attempt to | |
1233 | find a basis. */ | |
1234 | ||
1235 | static void | |
355fe088 | 1236 | slsr_process_mul (gimple *gs, tree rhs1, tree rhs2, bool speed) |
f9453c07 BS |
1237 | { |
1238 | slsr_cand_t c, c2; | |
1239 | ||
1240 | /* If this is a multiply of an SSA name with itself, it is highly | |
1241 | unlikely that we will get a strength reduction opportunity, so | |
1242 | don't record it as a candidate. This simplifies the logic for | |
1243 | finding a basis, so if this is removed that must be considered. */ | |
1244 | if (rhs1 == rhs2) | |
1245 | return; | |
1246 | ||
1247 | if (TREE_CODE (rhs2) == SSA_NAME) | |
1248 | { | |
1249 | /* Record an interpretation of this statement in the candidate table | |
3cfd4469 | 1250 | assuming RHS1 is the base expression and RHS2 is the stride. */ |
f9453c07 BS |
1251 | c = create_mul_ssa_cand (gs, rhs1, rhs2, speed); |
1252 | ||
1253 | /* Add the first interpretation to the statement-candidate mapping. */ | |
1254 | add_cand_for_stmt (gs, c); | |
1255 | ||
1256 | /* Record another interpretation of this statement assuming RHS1 | |
3cfd4469 | 1257 | is the stride and RHS2 is the base expression. */ |
f9453c07 BS |
1258 | c2 = create_mul_ssa_cand (gs, rhs2, rhs1, speed); |
1259 | c->next_interp = c2->cand_num; | |
d71dc990 | 1260 | c2->first_interp = c->cand_num; |
f9453c07 | 1261 | } |
f942ef18 | 1262 | else if (TREE_CODE (rhs2) == INTEGER_CST && !integer_zerop (rhs2)) |
f9453c07 BS |
1263 | { |
1264 | /* Record an interpretation for the multiply-immediate. */ | |
1265 | c = create_mul_imm_cand (gs, rhs1, rhs2, speed); | |
1266 | ||
1267 | /* Add the interpretation to the statement-candidate mapping. */ | |
1268 | add_cand_for_stmt (gs, c); | |
1269 | } | |
1270 | } | |
1271 | ||
1272 | /* Create a candidate entry for a statement GS, where GS adds two | |
1273 | SSA names BASE_IN and ADDEND_IN if SUBTRACT_P is false, and | |
1274 | subtracts ADDEND_IN from BASE_IN otherwise. Propagate any known | |
1275 | information about the two SSA names into the new candidate. | |
1276 | Return the new candidate. */ | |
1277 | ||
1278 | static slsr_cand_t | |
355fe088 | 1279 | create_add_ssa_cand (gimple *gs, tree base_in, tree addend_in, |
f9453c07 BS |
1280 | bool subtract_p, bool speed) |
1281 | { | |
0b56e9ad BS |
1282 | tree base = NULL_TREE, stride = NULL_TREE, ctype = NULL_TREE; |
1283 | tree stype = NULL_TREE; | |
807e902e | 1284 | widest_int index; |
f9453c07 BS |
1285 | unsigned savings = 0; |
1286 | slsr_cand_t c; | |
1287 | slsr_cand_t base_cand = base_cand_from_table (base_in); | |
1288 | slsr_cand_t addend_cand = base_cand_from_table (addend_in); | |
1289 | ||
1290 | /* The most useful transformation is a multiply-immediate feeding | |
1291 | an add or subtract. Look for that first. */ | |
9b92d12b | 1292 | while (addend_cand && !base && addend_cand->kind != CAND_PHI) |
f9453c07 BS |
1293 | { |
1294 | if (addend_cand->kind == CAND_MULT | |
807e902e | 1295 | && addend_cand->index == 0 |
f9453c07 BS |
1296 | && TREE_CODE (addend_cand->stride) == INTEGER_CST) |
1297 | { | |
1298 | /* Z = (B + 0) * S, S constant | |
1299 | X = Y +/- Z | |
1300 | =========================== | |
1301 | X = Y + ((+/-1 * S) * B) */ | |
1302 | base = base_in; | |
807e902e | 1303 | index = wi::to_widest (addend_cand->stride); |
f9453c07 | 1304 | if (subtract_p) |
27bcd47c | 1305 | index = -index; |
3cfd4469 | 1306 | stride = addend_cand->base_expr; |
b2ec94d4 | 1307 | ctype = TREE_TYPE (base_in); |
0b56e9ad | 1308 | stype = addend_cand->cand_type; |
f9453c07 BS |
1309 | if (has_single_use (addend_in)) |
1310 | savings = (addend_cand->dead_savings | |
1311 | + stmt_cost (addend_cand->cand_stmt, speed)); | |
1312 | } | |
1313 | ||
98aaa1a6 | 1314 | addend_cand = lookup_cand (addend_cand->next_interp); |
f9453c07 BS |
1315 | } |
1316 | ||
9b92d12b | 1317 | while (base_cand && !base && base_cand->kind != CAND_PHI) |
f9453c07 BS |
1318 | { |
1319 | if (base_cand->kind == CAND_ADD | |
807e902e | 1320 | && (base_cand->index == 0 |
f9453c07 BS |
1321 | || operand_equal_p (base_cand->stride, |
1322 | integer_zero_node, 0))) | |
1323 | { | |
1324 | /* Y = B + (i' * S), i' * S = 0 | |
1325 | X = Y +/- Z | |
1326 | ============================ | |
1327 | X = B + (+/-1 * Z) */ | |
3cfd4469 | 1328 | base = base_cand->base_expr; |
807e902e | 1329 | index = subtract_p ? -1 : 1; |
f9453c07 BS |
1330 | stride = addend_in; |
1331 | ctype = base_cand->cand_type; | |
0b56e9ad BS |
1332 | stype = (TREE_CODE (addend_in) == INTEGER_CST ? sizetype |
1333 | : TREE_TYPE (addend_in)); | |
f9453c07 BS |
1334 | if (has_single_use (base_in)) |
1335 | savings = (base_cand->dead_savings | |
1336 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1337 | } | |
1338 | else if (subtract_p) | |
1339 | { | |
1340 | slsr_cand_t subtrahend_cand = base_cand_from_table (addend_in); | |
1341 | ||
9b92d12b | 1342 | while (subtrahend_cand && !base && subtrahend_cand->kind != CAND_PHI) |
f9453c07 BS |
1343 | { |
1344 | if (subtrahend_cand->kind == CAND_MULT | |
807e902e | 1345 | && subtrahend_cand->index == 0 |
f9453c07 BS |
1346 | && TREE_CODE (subtrahend_cand->stride) == INTEGER_CST) |
1347 | { | |
1348 | /* Z = (B + 0) * S, S constant | |
1349 | X = Y - Z | |
1350 | =========================== | |
1351 | Value: X = Y + ((-1 * S) * B) */ | |
1352 | base = base_in; | |
807e902e | 1353 | index = wi::to_widest (subtrahend_cand->stride); |
27bcd47c | 1354 | index = -index; |
3cfd4469 | 1355 | stride = subtrahend_cand->base_expr; |
b2ec94d4 | 1356 | ctype = TREE_TYPE (base_in); |
0b56e9ad | 1357 | stype = subtrahend_cand->cand_type; |
f9453c07 BS |
1358 | if (has_single_use (addend_in)) |
1359 | savings = (subtrahend_cand->dead_savings | |
1360 | + stmt_cost (subtrahend_cand->cand_stmt, speed)); | |
1361 | } | |
98aaa1a6 BC |
1362 | |
1363 | subtrahend_cand = lookup_cand (subtrahend_cand->next_interp); | |
f9453c07 BS |
1364 | } |
1365 | } | |
1366 | ||
98aaa1a6 | 1367 | base_cand = lookup_cand (base_cand->next_interp); |
f9453c07 BS |
1368 | } |
1369 | ||
1370 | if (!base) | |
1371 | { | |
1372 | /* No interpretations had anything useful to propagate, so | |
1373 | produce X = Y + (1 * Z). */ | |
1374 | base = base_in; | |
807e902e | 1375 | index = subtract_p ? -1 : 1; |
f9453c07 | 1376 | stride = addend_in; |
b2ec94d4 | 1377 | ctype = TREE_TYPE (base_in); |
0b56e9ad BS |
1378 | stype = (TREE_CODE (addend_in) == INTEGER_CST ? sizetype |
1379 | : TREE_TYPE (addend_in)); | |
f9453c07 BS |
1380 | } |
1381 | ||
1382 | c = alloc_cand_and_find_basis (CAND_ADD, gs, base, index, stride, | |
0b56e9ad | 1383 | ctype, stype, savings); |
f9453c07 BS |
1384 | return c; |
1385 | } | |
1386 | ||
1387 | /* Create a candidate entry for a statement GS, where GS adds SSA | |
1388 | name BASE_IN to constant INDEX_IN. Propagate any known information | |
1389 | about BASE_IN into the new candidate. Return the new candidate. */ | |
1390 | ||
1391 | static slsr_cand_t | |
355fe088 | 1392 | create_add_imm_cand (gimple *gs, tree base_in, const widest_int &index_in, |
807e902e | 1393 | bool speed) |
f9453c07 BS |
1394 | { |
1395 | enum cand_kind kind = CAND_ADD; | |
1396 | tree base = NULL_TREE, stride = NULL_TREE, ctype = NULL_TREE; | |
0b56e9ad | 1397 | tree stype = NULL_TREE; |
807e902e | 1398 | widest_int index, multiple; |
f9453c07 BS |
1399 | unsigned savings = 0; |
1400 | slsr_cand_t c; | |
1401 | slsr_cand_t base_cand = base_cand_from_table (base_in); | |
1402 | ||
9b92d12b | 1403 | while (base_cand && !base && base_cand->kind != CAND_PHI) |
f9453c07 | 1404 | { |
807e902e | 1405 | signop sign = TYPE_SIGN (TREE_TYPE (base_cand->stride)); |
f9453c07 BS |
1406 | |
1407 | if (TREE_CODE (base_cand->stride) == INTEGER_CST | |
807e902e KZ |
1408 | && wi::multiple_of_p (index_in, wi::to_widest (base_cand->stride), |
1409 | sign, &multiple)) | |
f9453c07 BS |
1410 | { |
1411 | /* Y = (B + i') * S, S constant, c = kS for some integer k | |
1412 | X = Y + c | |
1413 | ============================ | |
1414 | X = (B + (i'+ k)) * S | |
1415 | OR | |
1416 | Y = B + (i' * S), S constant, c = kS for some integer k | |
1417 | X = Y + c | |
1418 | ============================ | |
1419 | X = (B + (i'+ k)) * S */ | |
1420 | kind = base_cand->kind; | |
3cfd4469 | 1421 | base = base_cand->base_expr; |
27bcd47c | 1422 | index = base_cand->index + multiple; |
f9453c07 BS |
1423 | stride = base_cand->stride; |
1424 | ctype = base_cand->cand_type; | |
0b56e9ad | 1425 | stype = base_cand->stride_type; |
f9453c07 BS |
1426 | if (has_single_use (base_in)) |
1427 | savings = (base_cand->dead_savings | |
1428 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1429 | } | |
1430 | ||
98aaa1a6 | 1431 | base_cand = lookup_cand (base_cand->next_interp); |
f9453c07 BS |
1432 | } |
1433 | ||
1434 | if (!base) | |
1435 | { | |
1436 | /* No interpretations had anything useful to propagate, so | |
1437 | produce X = Y + (c * 1). */ | |
1438 | kind = CAND_ADD; | |
1439 | base = base_in; | |
1440 | index = index_in; | |
1441 | stride = integer_one_node; | |
b2ec94d4 | 1442 | ctype = TREE_TYPE (base_in); |
0b56e9ad | 1443 | stype = sizetype; |
f9453c07 BS |
1444 | } |
1445 | ||
1446 | c = alloc_cand_and_find_basis (kind, gs, base, index, stride, | |
0b56e9ad | 1447 | ctype, stype, savings); |
f9453c07 BS |
1448 | return c; |
1449 | } | |
1450 | ||
1451 | /* Given GS which is an add or subtract of scalar integers or pointers, | |
1452 | make at least one appropriate entry in the candidate table. */ | |
1453 | ||
1454 | static void | |
355fe088 | 1455 | slsr_process_add (gimple *gs, tree rhs1, tree rhs2, bool speed) |
f9453c07 BS |
1456 | { |
1457 | bool subtract_p = gimple_assign_rhs_code (gs) == MINUS_EXPR; | |
1458 | slsr_cand_t c = NULL, c2; | |
1459 | ||
1460 | if (TREE_CODE (rhs2) == SSA_NAME) | |
1461 | { | |
3cfd4469 | 1462 | /* First record an interpretation assuming RHS1 is the base expression |
f9453c07 BS |
1463 | and RHS2 is the stride. But it doesn't make sense for the |
1464 | stride to be a pointer, so don't record a candidate in that case. */ | |
b2ec94d4 | 1465 | if (!POINTER_TYPE_P (TREE_TYPE (rhs2))) |
f9453c07 BS |
1466 | { |
1467 | c = create_add_ssa_cand (gs, rhs1, rhs2, subtract_p, speed); | |
1468 | ||
1469 | /* Add the first interpretation to the statement-candidate | |
1470 | mapping. */ | |
1471 | add_cand_for_stmt (gs, c); | |
1472 | } | |
1473 | ||
1474 | /* If the two RHS operands are identical, or this is a subtract, | |
1475 | we're done. */ | |
1476 | if (operand_equal_p (rhs1, rhs2, 0) || subtract_p) | |
1477 | return; | |
1478 | ||
1479 | /* Otherwise, record another interpretation assuming RHS2 is the | |
3cfd4469 | 1480 | base expression and RHS1 is the stride, again provided that the |
f9453c07 | 1481 | stride is not a pointer. */ |
b2ec94d4 | 1482 | if (!POINTER_TYPE_P (TREE_TYPE (rhs1))) |
f9453c07 BS |
1483 | { |
1484 | c2 = create_add_ssa_cand (gs, rhs2, rhs1, false, speed); | |
1485 | if (c) | |
d71dc990 BS |
1486 | { |
1487 | c->next_interp = c2->cand_num; | |
1488 | c2->first_interp = c->cand_num; | |
1489 | } | |
f9453c07 BS |
1490 | else |
1491 | add_cand_for_stmt (gs, c2); | |
1492 | } | |
1493 | } | |
36fd6408 | 1494 | else if (TREE_CODE (rhs2) == INTEGER_CST) |
f9453c07 | 1495 | { |
f9453c07 | 1496 | /* Record an interpretation for the add-immediate. */ |
807e902e | 1497 | widest_int index = wi::to_widest (rhs2); |
f9453c07 | 1498 | if (subtract_p) |
27bcd47c | 1499 | index = -index; |
f9453c07 BS |
1500 | |
1501 | c = create_add_imm_cand (gs, rhs1, index, speed); | |
1502 | ||
1503 | /* Add the interpretation to the statement-candidate mapping. */ | |
1504 | add_cand_for_stmt (gs, c); | |
1505 | } | |
1506 | } | |
1507 | ||
1508 | /* Given GS which is a negate of a scalar integer, make an appropriate | |
1509 | entry in the candidate table. A negate is equivalent to a multiply | |
1510 | by -1. */ | |
1511 | ||
1512 | static void | |
355fe088 | 1513 | slsr_process_neg (gimple *gs, tree rhs1, bool speed) |
f9453c07 BS |
1514 | { |
1515 | /* Record a CAND_MULT interpretation for the multiply by -1. */ | |
1516 | slsr_cand_t c = create_mul_imm_cand (gs, rhs1, integer_minus_one_node, speed); | |
1517 | ||
1518 | /* Add the interpretation to the statement-candidate mapping. */ | |
1519 | add_cand_for_stmt (gs, c); | |
1520 | } | |
1521 | ||
6b5eea61 BS |
1522 | /* Help function for legal_cast_p, operating on two trees. Checks |
1523 | whether it's allowable to cast from RHS to LHS. See legal_cast_p | |
1524 | for more details. */ | |
1525 | ||
1526 | static bool | |
0b56e9ad | 1527 | legal_cast_p_1 (tree lhs_type, tree rhs_type) |
6b5eea61 | 1528 | { |
6b5eea61 BS |
1529 | unsigned lhs_size, rhs_size; |
1530 | bool lhs_wraps, rhs_wraps; | |
1531 | ||
6b5eea61 BS |
1532 | lhs_size = TYPE_PRECISION (lhs_type); |
1533 | rhs_size = TYPE_PRECISION (rhs_type); | |
20bd649a MP |
1534 | lhs_wraps = ANY_INTEGRAL_TYPE_P (lhs_type) && TYPE_OVERFLOW_WRAPS (lhs_type); |
1535 | rhs_wraps = ANY_INTEGRAL_TYPE_P (rhs_type) && TYPE_OVERFLOW_WRAPS (rhs_type); | |
6b5eea61 BS |
1536 | |
1537 | if (lhs_size < rhs_size | |
1538 | || (rhs_wraps && !lhs_wraps) | |
1539 | || (rhs_wraps && lhs_wraps && rhs_size != lhs_size)) | |
1540 | return false; | |
1541 | ||
1542 | return true; | |
1543 | } | |
1544 | ||
f9453c07 BS |
1545 | /* Return TRUE if GS is a statement that defines an SSA name from |
1546 | a conversion and is legal for us to combine with an add and multiply | |
1547 | in the candidate table. For example, suppose we have: | |
1548 | ||
1549 | A = B + i; | |
1550 | C = (type) A; | |
1551 | D = C * S; | |
1552 | ||
1553 | Without the type-cast, we would create a CAND_MULT for D with base B, | |
1554 | index i, and stride S. We want to record this candidate only if it | |
1555 | is equivalent to apply the type cast following the multiply: | |
1556 | ||
1557 | A = B + i; | |
1558 | E = A * S; | |
1559 | D = (type) E; | |
1560 | ||
1561 | We will record the type with the candidate for D. This allows us | |
1562 | to use a similar previous candidate as a basis. If we have earlier seen | |
1563 | ||
1564 | A' = B + i'; | |
1565 | C' = (type) A'; | |
1566 | D' = C' * S; | |
1567 | ||
1568 | we can replace D with | |
1569 | ||
1570 | D = D' + (i - i') * S; | |
1571 | ||
1572 | But if moving the type-cast would change semantics, we mustn't do this. | |
1573 | ||
1574 | This is legitimate for casts from a non-wrapping integral type to | |
1575 | any integral type of the same or larger size. It is not legitimate | |
1576 | to convert a wrapping type to a non-wrapping type, or to a wrapping | |
1577 | type of a different size. I.e., with a wrapping type, we must | |
1578 | assume that the addition B + i could wrap, in which case performing | |
1579 | the multiply before or after one of the "illegal" type casts will | |
1580 | have different semantics. */ | |
1581 | ||
1582 | static bool | |
355fe088 | 1583 | legal_cast_p (gimple *gs, tree rhs) |
f9453c07 | 1584 | { |
f9453c07 BS |
1585 | if (!is_gimple_assign (gs) |
1586 | || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (gs))) | |
1587 | return false; | |
1588 | ||
0b56e9ad | 1589 | return legal_cast_p_1 (TREE_TYPE (gimple_assign_lhs (gs)), TREE_TYPE (rhs)); |
f9453c07 BS |
1590 | } |
1591 | ||
1592 | /* Given GS which is a cast to a scalar integer type, determine whether | |
1593 | the cast is legal for strength reduction. If so, make at least one | |
1594 | appropriate entry in the candidate table. */ | |
1595 | ||
1596 | static void | |
355fe088 | 1597 | slsr_process_cast (gimple *gs, tree rhs1, bool speed) |
f9453c07 BS |
1598 | { |
1599 | tree lhs, ctype; | |
ac2a97db | 1600 | slsr_cand_t base_cand, c = NULL, c2; |
f9453c07 BS |
1601 | unsigned savings = 0; |
1602 | ||
1603 | if (!legal_cast_p (gs, rhs1)) | |
1604 | return; | |
1605 | ||
1606 | lhs = gimple_assign_lhs (gs); | |
1607 | base_cand = base_cand_from_table (rhs1); | |
1608 | ctype = TREE_TYPE (lhs); | |
1609 | ||
9b92d12b | 1610 | if (base_cand && base_cand->kind != CAND_PHI) |
f9453c07 | 1611 | { |
d71dc990 BS |
1612 | slsr_cand_t first_cand = NULL; |
1613 | ||
f9453c07 BS |
1614 | while (base_cand) |
1615 | { | |
1616 | /* Propagate all data from the base candidate except the type, | |
1617 | which comes from the cast, and the base candidate's cast, | |
1618 | which is no longer applicable. */ | |
1619 | if (has_single_use (rhs1)) | |
1620 | savings = (base_cand->dead_savings | |
1621 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1622 | ||
1623 | c = alloc_cand_and_find_basis (base_cand->kind, gs, | |
3cfd4469 | 1624 | base_cand->base_expr, |
f9453c07 | 1625 | base_cand->index, base_cand->stride, |
0b56e9ad BS |
1626 | ctype, base_cand->stride_type, |
1627 | savings); | |
d71dc990 BS |
1628 | if (!first_cand) |
1629 | first_cand = c; | |
1630 | ||
1631 | if (first_cand != c) | |
1632 | c->first_interp = first_cand->cand_num; | |
1633 | ||
98aaa1a6 | 1634 | base_cand = lookup_cand (base_cand->next_interp); |
f9453c07 BS |
1635 | } |
1636 | } | |
1637 | else | |
1638 | { | |
1639 | /* If nothing is known about the RHS, create fresh CAND_ADD and | |
1640 | CAND_MULT interpretations: | |
1641 | ||
1642 | X = Y + (0 * 1) | |
1643 | X = (Y + 0) * 1 | |
1644 | ||
1645 | The first of these is somewhat arbitrary, but the choice of | |
1646 | 1 for the stride simplifies the logic for propagating casts | |
1647 | into their uses. */ | |
0b56e9ad BS |
1648 | c = alloc_cand_and_find_basis (CAND_ADD, gs, rhs1, 0, |
1649 | integer_one_node, ctype, sizetype, 0); | |
1650 | c2 = alloc_cand_and_find_basis (CAND_MULT, gs, rhs1, 0, | |
1651 | integer_one_node, ctype, sizetype, 0); | |
f9453c07 | 1652 | c->next_interp = c2->cand_num; |
d71dc990 | 1653 | c2->first_interp = c->cand_num; |
f9453c07 BS |
1654 | } |
1655 | ||
1656 | /* Add the first (or only) interpretation to the statement-candidate | |
1657 | mapping. */ | |
1658 | add_cand_for_stmt (gs, c); | |
1659 | } | |
1660 | ||
1661 | /* Given GS which is a copy of a scalar integer type, make at least one | |
1662 | appropriate entry in the candidate table. | |
1663 | ||
1664 | This interface is included for completeness, but is unnecessary | |
1665 | if this pass immediately follows a pass that performs copy | |
1666 | propagation, such as DOM. */ | |
1667 | ||
1668 | static void | |
355fe088 | 1669 | slsr_process_copy (gimple *gs, tree rhs1, bool speed) |
f9453c07 | 1670 | { |
ac2a97db | 1671 | slsr_cand_t base_cand, c = NULL, c2; |
f9453c07 BS |
1672 | unsigned savings = 0; |
1673 | ||
1674 | base_cand = base_cand_from_table (rhs1); | |
1675 | ||
9b92d12b | 1676 | if (base_cand && base_cand->kind != CAND_PHI) |
f9453c07 | 1677 | { |
d71dc990 BS |
1678 | slsr_cand_t first_cand = NULL; |
1679 | ||
f9453c07 BS |
1680 | while (base_cand) |
1681 | { | |
1682 | /* Propagate all data from the base candidate. */ | |
1683 | if (has_single_use (rhs1)) | |
1684 | savings = (base_cand->dead_savings | |
1685 | + stmt_cost (base_cand->cand_stmt, speed)); | |
1686 | ||
1687 | c = alloc_cand_and_find_basis (base_cand->kind, gs, | |
3cfd4469 | 1688 | base_cand->base_expr, |
f9453c07 | 1689 | base_cand->index, base_cand->stride, |
0b56e9ad BS |
1690 | base_cand->cand_type, |
1691 | base_cand->stride_type, savings); | |
d71dc990 BS |
1692 | if (!first_cand) |
1693 | first_cand = c; | |
1694 | ||
1695 | if (first_cand != c) | |
1696 | c->first_interp = first_cand->cand_num; | |
1697 | ||
98aaa1a6 | 1698 | base_cand = lookup_cand (base_cand->next_interp); |
f9453c07 BS |
1699 | } |
1700 | } | |
1701 | else | |
1702 | { | |
1703 | /* If nothing is known about the RHS, create fresh CAND_ADD and | |
1704 | CAND_MULT interpretations: | |
1705 | ||
1706 | X = Y + (0 * 1) | |
1707 | X = (Y + 0) * 1 | |
1708 | ||
1709 | The first of these is somewhat arbitrary, but the choice of | |
1710 | 1 for the stride simplifies the logic for propagating casts | |
1711 | into their uses. */ | |
0b56e9ad BS |
1712 | c = alloc_cand_and_find_basis (CAND_ADD, gs, rhs1, 0, |
1713 | integer_one_node, TREE_TYPE (rhs1), | |
1714 | sizetype, 0); | |
1715 | c2 = alloc_cand_and_find_basis (CAND_MULT, gs, rhs1, 0, | |
1716 | integer_one_node, TREE_TYPE (rhs1), | |
1717 | sizetype, 0); | |
f9453c07 | 1718 | c->next_interp = c2->cand_num; |
d71dc990 | 1719 | c2->first_interp = c->cand_num; |
f9453c07 BS |
1720 | } |
1721 | ||
1722 | /* Add the first (or only) interpretation to the statement-candidate | |
1723 | mapping. */ | |
1724 | add_cand_for_stmt (gs, c); | |
1725 | } | |
1726 | \f | |
4d9192b5 TS |
1727 | class find_candidates_dom_walker : public dom_walker |
1728 | { | |
1729 | public: | |
1730 | find_candidates_dom_walker (cdi_direction direction) | |
1731 | : dom_walker (direction) {} | |
3daacdcd | 1732 | virtual edge before_dom_children (basic_block); |
4d9192b5 TS |
1733 | }; |
1734 | ||
f9453c07 BS |
1735 | /* Find strength-reduction candidates in block BB. */ |
1736 | ||
3daacdcd | 1737 | edge |
4d9192b5 | 1738 | find_candidates_dom_walker::before_dom_children (basic_block bb) |
f9453c07 BS |
1739 | { |
1740 | bool speed = optimize_bb_for_speed_p (bb); | |
f9453c07 | 1741 | |
538dd0b7 DM |
1742 | for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi); |
1743 | gsi_next (&gsi)) | |
1744 | slsr_process_phi (gsi.phi (), speed); | |
9b92d12b | 1745 | |
538dd0b7 DM |
1746 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi); |
1747 | gsi_next (&gsi)) | |
f9453c07 | 1748 | { |
355fe088 | 1749 | gimple *gs = gsi_stmt (gsi); |
f9453c07 | 1750 | |
36bbc05d | 1751 | if (stmt_could_throw_p (cfun, gs)) |
bef52a68 JJ |
1752 | continue; |
1753 | ||
2749c8f6 BS |
1754 | if (gimple_vuse (gs) && gimple_assign_single_p (gs)) |
1755 | slsr_process_ref (gs); | |
1756 | ||
1757 | else if (is_gimple_assign (gs) | |
1ffbb016 RB |
1758 | && (INTEGRAL_TYPE_P (TREE_TYPE (gimple_assign_lhs (gs))) |
1759 | || POINTER_TYPE_P (TREE_TYPE (gimple_assign_lhs (gs))))) | |
f9453c07 BS |
1760 | { |
1761 | tree rhs1 = NULL_TREE, rhs2 = NULL_TREE; | |
1762 | ||
1763 | switch (gimple_assign_rhs_code (gs)) | |
1764 | { | |
1765 | case MULT_EXPR: | |
1766 | case PLUS_EXPR: | |
1767 | rhs1 = gimple_assign_rhs1 (gs); | |
1768 | rhs2 = gimple_assign_rhs2 (gs); | |
1769 | /* Should never happen, but currently some buggy situations | |
1770 | in earlier phases put constants in rhs1. */ | |
1771 | if (TREE_CODE (rhs1) != SSA_NAME) | |
1772 | continue; | |
1773 | break; | |
1774 | ||
1775 | /* Possible future opportunity: rhs1 of a ptr+ can be | |
1776 | an ADDR_EXPR. */ | |
1777 | case POINTER_PLUS_EXPR: | |
1778 | case MINUS_EXPR: | |
1779 | rhs2 = gimple_assign_rhs2 (gs); | |
81fea426 | 1780 | gcc_fallthrough (); |
f9453c07 | 1781 | |
d822570f | 1782 | CASE_CONVERT: |
fbcdc43e | 1783 | case SSA_NAME: |
f9453c07 BS |
1784 | case NEGATE_EXPR: |
1785 | rhs1 = gimple_assign_rhs1 (gs); | |
1786 | if (TREE_CODE (rhs1) != SSA_NAME) | |
1787 | continue; | |
1788 | break; | |
1789 | ||
1790 | default: | |
1791 | ; | |
1792 | } | |
1793 | ||
1794 | switch (gimple_assign_rhs_code (gs)) | |
1795 | { | |
1796 | case MULT_EXPR: | |
1797 | slsr_process_mul (gs, rhs1, rhs2, speed); | |
1798 | break; | |
1799 | ||
1800 | case PLUS_EXPR: | |
1801 | case POINTER_PLUS_EXPR: | |
1802 | case MINUS_EXPR: | |
1803 | slsr_process_add (gs, rhs1, rhs2, speed); | |
1804 | break; | |
1805 | ||
1806 | case NEGATE_EXPR: | |
1807 | slsr_process_neg (gs, rhs1, speed); | |
1808 | break; | |
1809 | ||
d822570f | 1810 | CASE_CONVERT: |
f9453c07 BS |
1811 | slsr_process_cast (gs, rhs1, speed); |
1812 | break; | |
1813 | ||
fbcdc43e | 1814 | case SSA_NAME: |
f9453c07 BS |
1815 | slsr_process_copy (gs, rhs1, speed); |
1816 | break; | |
1817 | ||
1818 | default: | |
1819 | ; | |
1820 | } | |
1821 | } | |
1822 | } | |
3daacdcd | 1823 | return NULL; |
f9453c07 BS |
1824 | } |
1825 | \f | |
1826 | /* Dump a candidate for debug. */ | |
1827 | ||
1828 | static void | |
1829 | dump_candidate (slsr_cand_t c) | |
1830 | { | |
1831 | fprintf (dump_file, "%3d [%d] ", c->cand_num, | |
1832 | gimple_bb (c->cand_stmt)->index); | |
ef6cb4c7 | 1833 | print_gimple_stmt (dump_file, c->cand_stmt, 0); |
f9453c07 BS |
1834 | switch (c->kind) |
1835 | { | |
1836 | case CAND_MULT: | |
1837 | fputs (" MULT : (", dump_file); | |
ef6cb4c7 | 1838 | print_generic_expr (dump_file, c->base_expr); |
f9453c07 | 1839 | fputs (" + ", dump_file); |
807e902e | 1840 | print_decs (c->index, dump_file); |
f9453c07 | 1841 | fputs (") * ", dump_file); |
0b56e9ad BS |
1842 | if (TREE_CODE (c->stride) != INTEGER_CST |
1843 | && c->stride_type != TREE_TYPE (c->stride)) | |
1844 | { | |
1845 | fputs ("(", dump_file); | |
ef6cb4c7 | 1846 | print_generic_expr (dump_file, c->stride_type); |
0b56e9ad BS |
1847 | fputs (")", dump_file); |
1848 | } | |
ef6cb4c7 | 1849 | print_generic_expr (dump_file, c->stride); |
f9453c07 BS |
1850 | fputs (" : ", dump_file); |
1851 | break; | |
1852 | case CAND_ADD: | |
1853 | fputs (" ADD : ", dump_file); | |
ef6cb4c7 | 1854 | print_generic_expr (dump_file, c->base_expr); |
f9453c07 | 1855 | fputs (" + (", dump_file); |
807e902e | 1856 | print_decs (c->index, dump_file); |
f9453c07 | 1857 | fputs (" * ", dump_file); |
0b56e9ad BS |
1858 | if (TREE_CODE (c->stride) != INTEGER_CST |
1859 | && c->stride_type != TREE_TYPE (c->stride)) | |
1860 | { | |
1861 | fputs ("(", dump_file); | |
ef6cb4c7 | 1862 | print_generic_expr (dump_file, c->stride_type); |
0b56e9ad BS |
1863 | fputs (")", dump_file); |
1864 | } | |
ef6cb4c7 | 1865 | print_generic_expr (dump_file, c->stride); |
f9453c07 BS |
1866 | fputs (") : ", dump_file); |
1867 | break; | |
2749c8f6 BS |
1868 | case CAND_REF: |
1869 | fputs (" REF : ", dump_file); | |
ef6cb4c7 | 1870 | print_generic_expr (dump_file, c->base_expr); |
2749c8f6 | 1871 | fputs (" + (", dump_file); |
ef6cb4c7 | 1872 | print_generic_expr (dump_file, c->stride); |
2749c8f6 | 1873 | fputs (") + ", dump_file); |
807e902e | 1874 | print_decs (c->index, dump_file); |
2749c8f6 BS |
1875 | fputs (" : ", dump_file); |
1876 | break; | |
9b92d12b BS |
1877 | case CAND_PHI: |
1878 | fputs (" PHI : ", dump_file); | |
ef6cb4c7 | 1879 | print_generic_expr (dump_file, c->base_expr); |
9b92d12b | 1880 | fputs (" + (unknown * ", dump_file); |
ef6cb4c7 | 1881 | print_generic_expr (dump_file, c->stride); |
9b92d12b BS |
1882 | fputs (") : ", dump_file); |
1883 | break; | |
f9453c07 BS |
1884 | default: |
1885 | gcc_unreachable (); | |
1886 | } | |
ef6cb4c7 | 1887 | print_generic_expr (dump_file, c->cand_type); |
f9453c07 BS |
1888 | fprintf (dump_file, "\n basis: %d dependent: %d sibling: %d\n", |
1889 | c->basis, c->dependent, c->sibling); | |
d71dc990 BS |
1890 | fprintf (dump_file, |
1891 | " next-interp: %d first-interp: %d dead-savings: %d\n", | |
1892 | c->next_interp, c->first_interp, c->dead_savings); | |
f9453c07 | 1893 | if (c->def_phi) |
9b92d12b | 1894 | fprintf (dump_file, " phi: %d\n", c->def_phi); |
f9453c07 BS |
1895 | fputs ("\n", dump_file); |
1896 | } | |
1897 | ||
1898 | /* Dump the candidate vector for debug. */ | |
1899 | ||
1900 | static void | |
1901 | dump_cand_vec (void) | |
1902 | { | |
1903 | unsigned i; | |
1904 | slsr_cand_t c; | |
1905 | ||
1906 | fprintf (dump_file, "\nStrength reduction candidate vector:\n\n"); | |
1907 | ||
9771b263 | 1908 | FOR_EACH_VEC_ELT (cand_vec, i, c) |
98aaa1a6 BC |
1909 | if (c != NULL) |
1910 | dump_candidate (c); | |
f9453c07 BS |
1911 | } |
1912 | ||
2749c8f6 | 1913 | /* Callback used to dump the candidate chains hash table. */ |
f9453c07 | 1914 | |
4a8fb1a1 LC |
1915 | int |
1916 | ssa_base_cand_dump_callback (cand_chain **slot, void *ignored ATTRIBUTE_UNUSED) | |
f9453c07 | 1917 | { |
4a8fb1a1 | 1918 | const_cand_chain_t chain = *slot; |
2749c8f6 | 1919 | cand_chain_t p; |
f9453c07 | 1920 | |
ef6cb4c7 | 1921 | print_generic_expr (dump_file, chain->base_expr); |
2749c8f6 | 1922 | fprintf (dump_file, " -> %d", chain->cand->cand_num); |
f9453c07 | 1923 | |
2749c8f6 BS |
1924 | for (p = chain->next; p; p = p->next) |
1925 | fprintf (dump_file, " -> %d", p->cand->cand_num); | |
f9453c07 | 1926 | |
2749c8f6 BS |
1927 | fputs ("\n", dump_file); |
1928 | return 1; | |
1929 | } | |
f9453c07 | 1930 | |
2749c8f6 | 1931 | /* Dump the candidate chains. */ |
f9453c07 | 1932 | |
2749c8f6 BS |
1933 | static void |
1934 | dump_cand_chains (void) | |
1935 | { | |
1936 | fprintf (dump_file, "\nStrength reduction candidate chains:\n\n"); | |
c203e8a7 TS |
1937 | base_cand_map->traverse_noresize <void *, ssa_base_cand_dump_callback> |
1938 | (NULL); | |
f9453c07 BS |
1939 | fputs ("\n", dump_file); |
1940 | } | |
88ca9ea1 BS |
1941 | |
1942 | /* Dump the increment vector for debug. */ | |
1943 | ||
1944 | static void | |
1945 | dump_incr_vec (void) | |
1946 | { | |
1947 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1948 | { | |
1949 | unsigned i; | |
1950 | ||
1951 | fprintf (dump_file, "\nIncrement vector:\n\n"); | |
1952 | ||
1953 | for (i = 0; i < incr_vec_len; i++) | |
1954 | { | |
1955 | fprintf (dump_file, "%3d increment: ", i); | |
807e902e | 1956 | print_decs (incr_vec[i].incr, dump_file); |
88ca9ea1 BS |
1957 | fprintf (dump_file, "\n count: %d", incr_vec[i].count); |
1958 | fprintf (dump_file, "\n cost: %d", incr_vec[i].cost); | |
1959 | fputs ("\n initializer: ", dump_file); | |
ef6cb4c7 | 1960 | print_generic_expr (dump_file, incr_vec[i].initializer); |
88ca9ea1 BS |
1961 | fputs ("\n\n", dump_file); |
1962 | } | |
1963 | } | |
1964 | } | |
f9453c07 | 1965 | \f |
2749c8f6 BS |
1966 | /* Replace *EXPR in candidate C with an equivalent strength-reduced |
1967 | data reference. */ | |
1968 | ||
1969 | static void | |
1970 | replace_ref (tree *expr, slsr_cand_t c) | |
1971 | { | |
78f6dd68 MJ |
1972 | tree add_expr, mem_ref, acc_type = TREE_TYPE (*expr); |
1973 | unsigned HOST_WIDE_INT misalign; | |
1974 | unsigned align; | |
1975 | ||
1976 | /* Ensure the memory reference carries the minimum alignment | |
1977 | requirement for the data type. See PR58041. */ | |
1978 | get_object_alignment_1 (*expr, &align, &misalign); | |
1979 | if (misalign != 0) | |
146ec50f | 1980 | align = least_bit_hwi (misalign); |
78f6dd68 MJ |
1981 | if (align < TYPE_ALIGN (acc_type)) |
1982 | acc_type = build_aligned_type (acc_type, align); | |
1983 | ||
8907a722 | 1984 | add_expr = fold_build2 (POINTER_PLUS_EXPR, c->cand_type, |
78f6dd68 MJ |
1985 | c->base_expr, c->stride); |
1986 | mem_ref = fold_build2 (MEM_REF, acc_type, add_expr, | |
807e902e | 1987 | wide_int_to_tree (c->cand_type, c->index)); |
78f6dd68 | 1988 | |
2749c8f6 BS |
1989 | /* Gimplify the base addressing expression for the new MEM_REF tree. */ |
1990 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); | |
1991 | TREE_OPERAND (mem_ref, 0) | |
1992 | = force_gimple_operand_gsi (&gsi, TREE_OPERAND (mem_ref, 0), | |
1993 | /*simple_p=*/true, NULL, | |
1994 | /*before=*/true, GSI_SAME_STMT); | |
1995 | copy_ref_info (mem_ref, *expr); | |
1996 | *expr = mem_ref; | |
1997 | update_stmt (c->cand_stmt); | |
1998 | } | |
1999 | ||
2000 | /* Replace CAND_REF candidate C, each sibling of candidate C, and each | |
2001 | dependent of candidate C with an equivalent strength-reduced data | |
2002 | reference. */ | |
2003 | ||
2004 | static void | |
2005 | replace_refs (slsr_cand_t c) | |
2006 | { | |
96d75a2c YZ |
2007 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2008 | { | |
2009 | fputs ("Replacing reference: ", dump_file); | |
ef6cb4c7 | 2010 | print_gimple_stmt (dump_file, c->cand_stmt, 0); |
96d75a2c YZ |
2011 | } |
2012 | ||
2749c8f6 BS |
2013 | if (gimple_vdef (c->cand_stmt)) |
2014 | { | |
2015 | tree *lhs = gimple_assign_lhs_ptr (c->cand_stmt); | |
2016 | replace_ref (lhs, c); | |
2017 | } | |
2018 | else | |
2019 | { | |
2020 | tree *rhs = gimple_assign_rhs1_ptr (c->cand_stmt); | |
2021 | replace_ref (rhs, c); | |
2022 | } | |
2023 | ||
96d75a2c YZ |
2024 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2025 | { | |
2026 | fputs ("With: ", dump_file); | |
ef6cb4c7 | 2027 | print_gimple_stmt (dump_file, c->cand_stmt, 0); |
96d75a2c YZ |
2028 | fputs ("\n", dump_file); |
2029 | } | |
2030 | ||
2749c8f6 BS |
2031 | if (c->sibling) |
2032 | replace_refs (lookup_cand (c->sibling)); | |
2033 | ||
2034 | if (c->dependent) | |
2035 | replace_refs (lookup_cand (c->dependent)); | |
2036 | } | |
2037 | ||
9b92d12b BS |
2038 | /* Return TRUE if candidate C is dependent upon a PHI. */ |
2039 | ||
2040 | static bool | |
2041 | phi_dependent_cand_p (slsr_cand_t c) | |
2042 | { | |
2043 | /* A candidate is not necessarily dependent upon a PHI just because | |
2044 | it has a phi definition for its base name. It may have a basis | |
2045 | that relies upon the same phi definition, in which case the PHI | |
2046 | is irrelevant to this candidate. */ | |
2047 | return (c->def_phi | |
2048 | && c->basis | |
2049 | && lookup_cand (c->basis)->def_phi != c->def_phi); | |
2050 | } | |
2051 | ||
f9453c07 BS |
2052 | /* Calculate the increment required for candidate C relative to |
2053 | its basis. */ | |
2054 | ||
807e902e | 2055 | static widest_int |
f9453c07 BS |
2056 | cand_increment (slsr_cand_t c) |
2057 | { | |
2058 | slsr_cand_t basis; | |
2059 | ||
2060 | /* If the candidate doesn't have a basis, just return its own | |
2061 | index. This is useful in record_increments to help us find | |
9b92d12b BS |
2062 | an existing initializer. Also, if the candidate's basis is |
2063 | hidden by a phi, then its own index will be the increment | |
2064 | from the newly introduced phi basis. */ | |
2065 | if (!c->basis || phi_dependent_cand_p (c)) | |
f9453c07 BS |
2066 | return c->index; |
2067 | ||
2068 | basis = lookup_cand (c->basis); | |
3cfd4469 | 2069 | gcc_assert (operand_equal_p (c->base_expr, basis->base_expr, 0)); |
27bcd47c | 2070 | return c->index - basis->index; |
f9453c07 BS |
2071 | } |
2072 | ||
88ca9ea1 BS |
2073 | /* Calculate the increment required for candidate C relative to |
2074 | its basis. If we aren't going to generate pointer arithmetic | |
2075 | for this candidate, return the absolute value of that increment | |
2076 | instead. */ | |
2077 | ||
807e902e | 2078 | static inline widest_int |
88ca9ea1 BS |
2079 | cand_abs_increment (slsr_cand_t c) |
2080 | { | |
807e902e | 2081 | widest_int increment = cand_increment (c); |
88ca9ea1 | 2082 | |
807e902e | 2083 | if (!address_arithmetic_p && wi::neg_p (increment)) |
27bcd47c | 2084 | increment = -increment; |
88ca9ea1 BS |
2085 | |
2086 | return increment; | |
2087 | } | |
2088 | ||
f9453c07 BS |
2089 | /* Return TRUE iff candidate C has already been replaced under |
2090 | another interpretation. */ | |
2091 | ||
2092 | static inline bool | |
2093 | cand_already_replaced (slsr_cand_t c) | |
2094 | { | |
2095 | return (gimple_bb (c->cand_stmt) == 0); | |
2096 | } | |
2097 | ||
9b92d12b BS |
2098 | /* Common logic used by replace_unconditional_candidate and |
2099 | replace_conditional_candidate. */ | |
f9453c07 BS |
2100 | |
2101 | static void | |
807e902e | 2102 | replace_mult_candidate (slsr_cand_t c, tree basis_name, widest_int bump) |
f9453c07 | 2103 | { |
9b92d12b BS |
2104 | tree target_type = TREE_TYPE (gimple_assign_lhs (c->cand_stmt)); |
2105 | enum tree_code cand_code = gimple_assign_rhs_code (c->cand_stmt); | |
2106 | ||
8eb91869 BS |
2107 | /* It is not useful to replace casts, copies, negates, or adds of |
2108 | an SSA name and a constant. */ | |
2109 | if (cand_code == SSA_NAME | |
2110 | || CONVERT_EXPR_CODE_P (cand_code) | |
2111 | || cand_code == PLUS_EXPR | |
2112 | || cand_code == POINTER_PLUS_EXPR | |
2113 | || cand_code == MINUS_EXPR | |
2114 | || cand_code == NEGATE_EXPR) | |
2115 | return; | |
2116 | ||
2117 | enum tree_code code = PLUS_EXPR; | |
2118 | tree bump_tree; | |
2119 | gimple *stmt_to_print = NULL; | |
2120 | ||
2121 | if (wi::neg_p (bump)) | |
9b92d12b | 2122 | { |
8eb91869 BS |
2123 | code = MINUS_EXPR; |
2124 | bump = -bump; | |
2125 | } | |
9b92d12b | 2126 | |
8eb91869 BS |
2127 | /* It is possible that the resulting bump doesn't fit in target_type. |
2128 | Abandon the replacement in this case. This does not affect | |
2129 | siblings or dependents of C. */ | |
2130 | if (bump != wi::ext (bump, TYPE_PRECISION (target_type), | |
2131 | TYPE_SIGN (target_type))) | |
2132 | return; | |
2133 | ||
2134 | bump_tree = wide_int_to_tree (target_type, bump); | |
2135 | ||
2136 | /* If the basis name and the candidate's LHS have incompatible types, | |
2137 | introduce a cast. */ | |
2138 | if (!useless_type_conversion_p (target_type, TREE_TYPE (basis_name))) | |
2139 | basis_name = introduce_cast_before_cand (c, target_type, basis_name); | |
9b92d12b | 2140 | |
8eb91869 BS |
2141 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2142 | { | |
2143 | fputs ("Replacing: ", dump_file); | |
2144 | print_gimple_stmt (dump_file, c->cand_stmt, 0); | |
2145 | } | |
2146 | ||
2147 | if (bump == 0) | |
2148 | { | |
2149 | tree lhs = gimple_assign_lhs (c->cand_stmt); | |
2150 | gassign *copy_stmt = gimple_build_assign (lhs, basis_name); | |
2151 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); | |
d71dc990 | 2152 | slsr_cand_t cc = lookup_cand (c->first_interp); |
8eb91869 BS |
2153 | gimple_set_location (copy_stmt, gimple_location (c->cand_stmt)); |
2154 | gsi_replace (&gsi, copy_stmt, false); | |
d71dc990 | 2155 | while (cc) |
8eb91869 | 2156 | { |
8eb91869 | 2157 | cc->cand_stmt = copy_stmt; |
98aaa1a6 | 2158 | cc = lookup_cand (cc->next_interp); |
8eb91869 | 2159 | } |
9b92d12b | 2160 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8eb91869 BS |
2161 | stmt_to_print = copy_stmt; |
2162 | } | |
2163 | else | |
2164 | { | |
2165 | tree rhs1, rhs2; | |
2166 | if (cand_code != NEGATE_EXPR) { | |
2167 | rhs1 = gimple_assign_rhs1 (c->cand_stmt); | |
2168 | rhs2 = gimple_assign_rhs2 (c->cand_stmt); | |
2169 | } | |
2170 | if (cand_code != NEGATE_EXPR | |
2171 | && ((operand_equal_p (rhs1, basis_name, 0) | |
2172 | && operand_equal_p (rhs2, bump_tree, 0)) | |
2173 | || (operand_equal_p (rhs1, bump_tree, 0) | |
2174 | && operand_equal_p (rhs2, basis_name, 0)))) | |
9b92d12b | 2175 | { |
8eb91869 BS |
2176 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2177 | { | |
2178 | fputs ("(duplicate, not actually replacing)", dump_file); | |
2179 | stmt_to_print = c->cand_stmt; | |
2180 | } | |
9b92d12b | 2181 | } |
8eb91869 | 2182 | else |
9b92d12b | 2183 | { |
9b92d12b | 2184 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); |
d71dc990 | 2185 | slsr_cand_t cc = lookup_cand (c->first_interp); |
8eb91869 BS |
2186 | gimple_assign_set_rhs_with_ops (&gsi, code, basis_name, bump_tree); |
2187 | update_stmt (gsi_stmt (gsi)); | |
d71dc990 | 2188 | while (cc) |
9fbbba71 | 2189 | { |
8eb91869 | 2190 | cc->cand_stmt = gsi_stmt (gsi); |
98aaa1a6 | 2191 | cc = lookup_cand (cc->next_interp); |
9fbbba71 | 2192 | } |
9b92d12b | 2193 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8eb91869 | 2194 | stmt_to_print = gsi_stmt (gsi); |
9b92d12b | 2195 | } |
8eb91869 | 2196 | } |
9b92d12b | 2197 | |
8eb91869 BS |
2198 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2199 | { | |
2200 | fputs ("With: ", dump_file); | |
2201 | print_gimple_stmt (dump_file, stmt_to_print, 0); | |
2202 | fputs ("\n", dump_file); | |
9b92d12b BS |
2203 | } |
2204 | } | |
2205 | ||
2206 | /* Replace candidate C with an add or subtract. Note that we only | |
2207 | operate on CAND_MULTs with known strides, so we will never generate | |
2208 | a POINTER_PLUS_EXPR. Each candidate X = (B + i) * S is replaced by | |
2209 | X = Y + ((i - i') * S), as described in the module commentary. The | |
2210 | folded value ((i - i') * S) is referred to here as the "bump." */ | |
2211 | ||
2212 | static void | |
2213 | replace_unconditional_candidate (slsr_cand_t c) | |
2214 | { | |
2215 | slsr_cand_t basis; | |
9b92d12b BS |
2216 | |
2217 | if (cand_already_replaced (c)) | |
f9453c07 BS |
2218 | return; |
2219 | ||
2220 | basis = lookup_cand (c->basis); | |
807e902e | 2221 | widest_int bump = cand_increment (c) * wi::to_widest (c->stride); |
9b92d12b | 2222 | |
a7a7d10e | 2223 | replace_mult_candidate (c, gimple_assign_lhs (basis->cand_stmt), bump); |
9b92d12b BS |
2224 | } |
2225 | \f | |
7bf55a70 BS |
2226 | /* Return the index in the increment vector of the given INCREMENT, |
2227 | or -1 if not found. The latter can occur if more than | |
2228 | MAX_INCR_VEC_LEN increments have been found. */ | |
9b92d12b | 2229 | |
7bf55a70 | 2230 | static inline int |
807e902e | 2231 | incr_vec_index (const widest_int &increment) |
9b92d12b BS |
2232 | { |
2233 | unsigned i; | |
2234 | ||
2235 | for (i = 0; i < incr_vec_len && increment != incr_vec[i].incr; i++) | |
2236 | ; | |
2237 | ||
7bf55a70 BS |
2238 | if (i < incr_vec_len) |
2239 | return i; | |
2240 | else | |
2241 | return -1; | |
9b92d12b BS |
2242 | } |
2243 | ||
2244 | /* Create a new statement along edge E to add BASIS_NAME to the product | |
2245 | of INCREMENT and the stride of candidate C. Create and return a new | |
2246 | SSA name from *VAR to be used as the LHS of the new statement. | |
2247 | KNOWN_STRIDE is true iff C's stride is a constant. */ | |
2248 | ||
2249 | static tree | |
2250 | create_add_on_incoming_edge (slsr_cand_t c, tree basis_name, | |
807e902e | 2251 | widest_int increment, edge e, location_t loc, |
9b92d12b BS |
2252 | bool known_stride) |
2253 | { | |
9b92d12b | 2254 | tree lhs, basis_type; |
0b56e9ad | 2255 | gassign *new_stmt, *cast_stmt = NULL; |
9b92d12b BS |
2256 | |
2257 | /* If the add candidate along this incoming edge has the same | |
2258 | index as C's hidden basis, the hidden basis represents this | |
2259 | edge correctly. */ | |
807e902e | 2260 | if (increment == 0) |
9b92d12b BS |
2261 | return basis_name; |
2262 | ||
2263 | basis_type = TREE_TYPE (basis_name); | |
2264 | lhs = make_temp_ssa_name (basis_type, NULL, "slsr"); | |
2265 | ||
a929f266 BS |
2266 | /* Occasionally people convert integers to pointers without a |
2267 | cast, leading us into trouble if we aren't careful. */ | |
2268 | enum tree_code plus_code | |
2269 | = POINTER_TYPE_P (basis_type) ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
2270 | ||
9b92d12b | 2271 | if (known_stride) |
7139194b | 2272 | { |
9b92d12b | 2273 | tree bump_tree; |
a929f266 | 2274 | enum tree_code code = plus_code; |
807e902e | 2275 | widest_int bump = increment * wi::to_widest (c->stride); |
a929f266 | 2276 | if (wi::neg_p (bump) && !POINTER_TYPE_P (basis_type)) |
9b92d12b BS |
2277 | { |
2278 | code = MINUS_EXPR; | |
2279 | bump = -bump; | |
2280 | } | |
2281 | ||
a929f266 BS |
2282 | tree stride_type = POINTER_TYPE_P (basis_type) ? sizetype : basis_type; |
2283 | bump_tree = wide_int_to_tree (stride_type, bump); | |
0d0e4a03 | 2284 | new_stmt = gimple_build_assign (lhs, code, basis_name, bump_tree); |
7139194b AH |
2285 | } |
2286 | else | |
2287 | { | |
7bf55a70 | 2288 | int i; |
a929f266 | 2289 | bool negate_incr = !POINTER_TYPE_P (basis_type) && wi::neg_p (increment); |
9b92d12b | 2290 | i = incr_vec_index (negate_incr ? -increment : increment); |
7bf55a70 | 2291 | gcc_assert (i >= 0); |
f9453c07 | 2292 | |
9b92d12b BS |
2293 | if (incr_vec[i].initializer) |
2294 | { | |
a929f266 | 2295 | enum tree_code code = negate_incr ? MINUS_EXPR : plus_code; |
0d0e4a03 JJ |
2296 | new_stmt = gimple_build_assign (lhs, code, basis_name, |
2297 | incr_vec[i].initializer); | |
9b92d12b | 2298 | } |
0b56e9ad BS |
2299 | else { |
2300 | tree stride; | |
2301 | ||
2302 | if (!types_compatible_p (TREE_TYPE (c->stride), c->stride_type)) | |
2303 | { | |
2304 | tree cast_stride = make_temp_ssa_name (c->stride_type, NULL, | |
2305 | "slsr"); | |
2306 | cast_stmt = gimple_build_assign (cast_stride, NOP_EXPR, | |
2307 | c->stride); | |
2308 | stride = cast_stride; | |
2309 | } | |
2310 | else | |
2311 | stride = c->stride; | |
2312 | ||
2313 | if (increment == 1) | |
2314 | new_stmt = gimple_build_assign (lhs, plus_code, basis_name, stride); | |
2315 | else if (increment == -1) | |
2316 | new_stmt = gimple_build_assign (lhs, MINUS_EXPR, basis_name, stride); | |
2317 | else | |
2318 | gcc_unreachable (); | |
2319 | } | |
f9453c07 BS |
2320 | } |
2321 | ||
b115e803 | 2322 | if (cast_stmt) |
0b56e9ad | 2323 | { |
b115e803 BS |
2324 | gimple_set_location (cast_stmt, loc); |
2325 | gsi_insert_on_edge (e, cast_stmt); | |
0b56e9ad | 2326 | } |
9b92d12b BS |
2327 | |
2328 | gimple_set_location (new_stmt, loc); | |
b115e803 | 2329 | gsi_insert_on_edge (e, new_stmt); |
f9453c07 BS |
2330 | |
2331 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2332 | { | |
0b56e9ad BS |
2333 | if (cast_stmt) |
2334 | { | |
b115e803 BS |
2335 | fprintf (dump_file, "Inserting cast on edge %d->%d: ", |
2336 | e->src->index, e->dest->index); | |
ef6cb4c7 | 2337 | print_gimple_stmt (dump_file, cast_stmt, 0); |
0b56e9ad | 2338 | } |
b115e803 BS |
2339 | fprintf (dump_file, "Inserting on edge %d->%d: ", e->src->index, |
2340 | e->dest->index); | |
ef6cb4c7 | 2341 | print_gimple_stmt (dump_file, new_stmt, 0); |
f9453c07 BS |
2342 | } |
2343 | ||
9b92d12b BS |
2344 | return lhs; |
2345 | } | |
2346 | ||
65d3dce8 BS |
2347 | /* Clear the visited field for a tree of PHI candidates. */ |
2348 | ||
2349 | static void | |
2350 | clear_visited (gphi *phi) | |
2351 | { | |
2352 | unsigned i; | |
2353 | slsr_cand_t phi_cand = *stmt_cand_map->get (phi); | |
2354 | ||
2355 | if (phi_cand->visited) | |
2356 | { | |
2357 | phi_cand->visited = 0; | |
2358 | ||
2359 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2360 | { | |
2361 | tree arg = gimple_phi_arg_def (phi, i); | |
2362 | gimple *arg_def = SSA_NAME_DEF_STMT (arg); | |
2363 | if (gimple_code (arg_def) == GIMPLE_PHI) | |
2364 | clear_visited (as_a <gphi *> (arg_def)); | |
2365 | } | |
2366 | } | |
2367 | } | |
2368 | ||
2369 | /* Recursive helper function for create_phi_basis. */ | |
9b92d12b BS |
2370 | |
2371 | static tree | |
65d3dce8 BS |
2372 | create_phi_basis_1 (slsr_cand_t c, gimple *from_phi, tree basis_name, |
2373 | location_t loc, bool known_stride) | |
9b92d12b BS |
2374 | { |
2375 | int i; | |
2376 | tree name, phi_arg; | |
538dd0b7 | 2377 | gphi *phi; |
9b92d12b BS |
2378 | slsr_cand_t basis = lookup_cand (c->basis); |
2379 | int nargs = gimple_phi_num_args (from_phi); | |
2380 | basic_block phi_bb = gimple_bb (from_phi); | |
5b3d5f76 | 2381 | slsr_cand_t phi_cand = *stmt_cand_map->get (from_phi); |
c8189787 | 2382 | auto_vec<tree> phi_args (nargs); |
9b92d12b | 2383 | |
65d3dce8 BS |
2384 | if (phi_cand->visited) |
2385 | return phi_cand->cached_basis; | |
2386 | phi_cand->visited = 1; | |
2387 | ||
9b92d12b BS |
2388 | /* Process each argument of the existing phi that represents |
2389 | conditionally-executed add candidates. */ | |
2390 | for (i = 0; i < nargs; i++) | |
f9453c07 | 2391 | { |
9b92d12b BS |
2392 | edge e = (*phi_bb->preds)[i]; |
2393 | tree arg = gimple_phi_arg_def (from_phi, i); | |
2394 | tree feeding_def; | |
2395 | ||
2396 | /* If the phi argument is the base name of the CAND_PHI, then | |
2397 | this incoming arc should use the hidden basis. */ | |
2398 | if (operand_equal_p (arg, phi_cand->base_expr, 0)) | |
807e902e | 2399 | if (basis->index == 0) |
9b92d12b BS |
2400 | feeding_def = gimple_assign_lhs (basis->cand_stmt); |
2401 | else | |
2402 | { | |
807e902e | 2403 | widest_int incr = -basis->index; |
9b92d12b BS |
2404 | feeding_def = create_add_on_incoming_edge (c, basis_name, incr, |
2405 | e, loc, known_stride); | |
2406 | } | |
2407 | else | |
f9453c07 | 2408 | { |
355fe088 | 2409 | gimple *arg_def = SSA_NAME_DEF_STMT (arg); |
9b92d12b BS |
2410 | |
2411 | /* If there is another phi along this incoming edge, we must | |
2412 | process it in the same fashion to ensure that all basis | |
2413 | adjustments are made along its incoming edges. */ | |
2414 | if (gimple_code (arg_def) == GIMPLE_PHI) | |
65d3dce8 BS |
2415 | feeding_def = create_phi_basis_1 (c, arg_def, basis_name, |
2416 | loc, known_stride); | |
9b92d12b | 2417 | else |
f9453c07 | 2418 | { |
9b92d12b | 2419 | slsr_cand_t arg_cand = base_cand_from_table (arg); |
807e902e | 2420 | widest_int diff = arg_cand->index - basis->index; |
9b92d12b BS |
2421 | feeding_def = create_add_on_incoming_edge (c, basis_name, diff, |
2422 | e, loc, known_stride); | |
f9453c07 BS |
2423 | } |
2424 | } | |
9b92d12b BS |
2425 | |
2426 | /* Because of recursion, we need to save the arguments in a vector | |
2427 | so we can create the PHI statement all at once. Otherwise the | |
2428 | storage for the half-created PHI can be reclaimed. */ | |
2429 | phi_args.safe_push (feeding_def); | |
f9453c07 | 2430 | } |
9b92d12b BS |
2431 | |
2432 | /* Create the new phi basis. */ | |
2433 | name = make_temp_ssa_name (TREE_TYPE (basis_name), NULL, "slsr"); | |
2434 | phi = create_phi_node (name, phi_bb); | |
2435 | SSA_NAME_DEF_STMT (name) = phi; | |
2436 | ||
2437 | FOR_EACH_VEC_ELT (phi_args, i, phi_arg) | |
2438 | { | |
2439 | edge e = (*phi_bb->preds)[i]; | |
2440 | add_phi_arg (phi, phi_arg, e, loc); | |
2441 | } | |
2442 | ||
2443 | update_stmt (phi); | |
2444 | ||
f9453c07 BS |
2445 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2446 | { | |
9b92d12b | 2447 | fputs ("Introducing new phi basis: ", dump_file); |
ef6cb4c7 | 2448 | print_gimple_stmt (dump_file, phi, 0); |
f9453c07 | 2449 | } |
9b92d12b | 2450 | |
65d3dce8 | 2451 | phi_cand->cached_basis = name; |
9b92d12b | 2452 | return name; |
f9453c07 BS |
2453 | } |
2454 | ||
65d3dce8 BS |
2455 | /* Given a candidate C with BASIS_NAME being the LHS of C's basis which |
2456 | is hidden by the phi node FROM_PHI, create a new phi node in the same | |
2457 | block as FROM_PHI. The new phi is suitable for use as a basis by C, | |
2458 | with its phi arguments representing conditional adjustments to the | |
2459 | hidden basis along conditional incoming paths. Those adjustments are | |
2460 | made by creating add statements (and sometimes recursively creating | |
2461 | phis) along those incoming paths. LOC is the location to attach to | |
2462 | the introduced statements. KNOWN_STRIDE is true iff C's stride is a | |
2463 | constant. */ | |
2464 | ||
2465 | static tree | |
2466 | create_phi_basis (slsr_cand_t c, gimple *from_phi, tree basis_name, | |
2467 | location_t loc, bool known_stride) | |
2468 | { | |
2469 | tree retval = create_phi_basis_1 (c, from_phi, basis_name, loc, | |
2470 | known_stride); | |
2471 | gcc_assert (retval); | |
2472 | clear_visited (as_a <gphi *> (from_phi)); | |
2473 | return retval; | |
2474 | } | |
2475 | ||
9b92d12b BS |
2476 | /* Given a candidate C whose basis is hidden by at least one intervening |
2477 | phi, introduce a matching number of new phis to represent its basis | |
2478 | adjusted by conditional increments along possible incoming paths. Then | |
2479 | replace C as though it were an unconditional candidate, using the new | |
2480 | basis. */ | |
f9453c07 BS |
2481 | |
2482 | static void | |
9b92d12b | 2483 | replace_conditional_candidate (slsr_cand_t c) |
f9453c07 | 2484 | { |
a7a7d10e | 2485 | tree basis_name, name; |
9b92d12b BS |
2486 | slsr_cand_t basis; |
2487 | location_t loc; | |
f9453c07 | 2488 | |
9b92d12b BS |
2489 | /* Look up the LHS SSA name from C's basis. This will be the |
2490 | RHS1 of the adds we will introduce to create new phi arguments. */ | |
2491 | basis = lookup_cand (c->basis); | |
2492 | basis_name = gimple_assign_lhs (basis->cand_stmt); | |
f9453c07 | 2493 | |
9b92d12b BS |
2494 | /* Create a new phi statement which will represent C's true basis |
2495 | after the transformation is complete. */ | |
2496 | loc = gimple_location (c->cand_stmt); | |
2497 | name = create_phi_basis (c, lookup_cand (c->def_phi)->cand_stmt, | |
2498 | basis_name, loc, KNOWN_STRIDE); | |
65d3dce8 | 2499 | |
9b92d12b | 2500 | /* Replace C with an add of the new basis phi and a constant. */ |
807e902e | 2501 | widest_int bump = c->index * wi::to_widest (c->stride); |
f9453c07 | 2502 | |
a7a7d10e | 2503 | replace_mult_candidate (c, name, bump); |
f9453c07 | 2504 | } |
88ca9ea1 | 2505 | |
65d3dce8 BS |
2506 | /* Recursive helper function for phi_add_costs. SPREAD is a measure of |
2507 | how many PHI nodes we have visited at this point in the tree walk. */ | |
9b92d12b BS |
2508 | |
2509 | static int | |
65d3dce8 | 2510 | phi_add_costs_1 (gimple *phi, slsr_cand_t c, int one_add_cost, int *spread) |
88ca9ea1 BS |
2511 | { |
2512 | unsigned i; | |
9b92d12b | 2513 | int cost = 0; |
5b3d5f76 | 2514 | slsr_cand_t phi_cand = *stmt_cand_map->get (phi); |
88ca9ea1 | 2515 | |
65d3dce8 BS |
2516 | if (phi_cand->visited) |
2517 | return 0; | |
2518 | ||
2519 | phi_cand->visited = 1; | |
2520 | (*spread)++; | |
2521 | ||
0100cd3f BS |
2522 | /* If we work our way back to a phi that isn't dominated by the hidden |
2523 | basis, this isn't a candidate for replacement. Indicate this by | |
2524 | returning an unreasonably high cost. It's not easy to detect | |
2525 | these situations when determining the basis, so we defer the | |
2526 | decision until now. */ | |
2527 | basic_block phi_bb = gimple_bb (phi); | |
2528 | slsr_cand_t basis = lookup_cand (c->basis); | |
2529 | basic_block basis_bb = gimple_bb (basis->cand_stmt); | |
2530 | ||
2531 | if (phi_bb == basis_bb || !dominated_by_p (CDI_DOMINATORS, phi_bb, basis_bb)) | |
2532 | return COST_INFINITE; | |
2533 | ||
9b92d12b BS |
2534 | for (i = 0; i < gimple_phi_num_args (phi); i++) |
2535 | { | |
2536 | tree arg = gimple_phi_arg_def (phi, i); | |
2537 | ||
2538 | if (arg != phi_cand->base_expr) | |
2539 | { | |
355fe088 | 2540 | gimple *arg_def = SSA_NAME_DEF_STMT (arg); |
9b92d12b BS |
2541 | |
2542 | if (gimple_code (arg_def) == GIMPLE_PHI) | |
65d3dce8 BS |
2543 | { |
2544 | cost += phi_add_costs_1 (arg_def, c, one_add_cost, spread); | |
2545 | ||
2546 | if (cost >= COST_INFINITE || *spread > MAX_SPREAD) | |
2547 | return COST_INFINITE; | |
2548 | } | |
9b92d12b BS |
2549 | else |
2550 | { | |
2551 | slsr_cand_t arg_cand = base_cand_from_table (arg); | |
2552 | ||
2553 | if (arg_cand->index != c->index) | |
2554 | cost += one_add_cost; | |
2555 | } | |
2556 | } | |
2557 | } | |
2558 | ||
2559 | return cost; | |
88ca9ea1 BS |
2560 | } |
2561 | ||
65d3dce8 BS |
2562 | /* Compute the expected costs of inserting basis adjustments for |
2563 | candidate C with phi-definition PHI. The cost of inserting | |
2564 | one adjustment is given by ONE_ADD_COST. If PHI has arguments | |
2565 | which are themselves phi results, recursively calculate costs | |
2566 | for those phis as well. */ | |
2567 | ||
2568 | static int | |
2569 | phi_add_costs (gimple *phi, slsr_cand_t c, int one_add_cost) | |
2570 | { | |
2571 | int spread = 0; | |
2572 | int retval = phi_add_costs_1 (phi, c, one_add_cost, &spread); | |
2573 | clear_visited (as_a <gphi *> (phi)); | |
2574 | return retval; | |
2575 | } | |
9b92d12b BS |
2576 | /* For candidate C, each sibling of candidate C, and each dependent of |
2577 | candidate C, determine whether the candidate is dependent upon a | |
2578 | phi that hides its basis. If not, replace the candidate unconditionally. | |
2579 | Otherwise, determine whether the cost of introducing compensation code | |
2580 | for the candidate is offset by the gains from strength reduction. If | |
2581 | so, replace the candidate and introduce the compensation code. */ | |
2582 | ||
2583 | static void | |
2584 | replace_uncond_cands_and_profitable_phis (slsr_cand_t c) | |
2585 | { | |
2586 | if (phi_dependent_cand_p (c)) | |
2587 | { | |
df11b2ea BS |
2588 | /* A multiply candidate with a stride of 1 is just an artifice |
2589 | of a copy or cast; there is no value in replacing it. */ | |
2590 | if (c->kind == CAND_MULT && wi::to_widest (c->stride) != 1) | |
9b92d12b BS |
2591 | { |
2592 | /* A candidate dependent upon a phi will replace a multiply by | |
2593 | a constant with an add, and will insert at most one add for | |
2594 | each phi argument. Add these costs with the potential dead-code | |
2595 | savings to determine profitability. */ | |
2596 | bool speed = optimize_bb_for_speed_p (gimple_bb (c->cand_stmt)); | |
2597 | int mult_savings = stmt_cost (c->cand_stmt, speed); | |
355fe088 | 2598 | gimple *phi = lookup_cand (c->def_phi)->cand_stmt; |
9b92d12b BS |
2599 | tree phi_result = gimple_phi_result (phi); |
2600 | int one_add_cost = add_cost (speed, | |
2601 | TYPE_MODE (TREE_TYPE (phi_result))); | |
2602 | int add_costs = one_add_cost + phi_add_costs (phi, c, one_add_cost); | |
2603 | int cost = add_costs - mult_savings - c->dead_savings; | |
2604 | ||
2605 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2606 | { | |
2607 | fprintf (dump_file, " Conditional candidate %d:\n", c->cand_num); | |
2608 | fprintf (dump_file, " add_costs = %d\n", add_costs); | |
2609 | fprintf (dump_file, " mult_savings = %d\n", mult_savings); | |
2610 | fprintf (dump_file, " dead_savings = %d\n", c->dead_savings); | |
2611 | fprintf (dump_file, " cost = %d\n", cost); | |
2612 | if (cost <= COST_NEUTRAL) | |
2613 | fputs (" Replacing...\n", dump_file); | |
2614 | else | |
2615 | fputs (" Not replaced.\n", dump_file); | |
2616 | } | |
2617 | ||
2618 | if (cost <= COST_NEUTRAL) | |
2619 | replace_conditional_candidate (c); | |
2620 | } | |
2621 | } | |
2622 | else | |
2623 | replace_unconditional_candidate (c); | |
2624 | ||
2625 | if (c->sibling) | |
2626 | replace_uncond_cands_and_profitable_phis (lookup_cand (c->sibling)); | |
2627 | ||
2628 | if (c->dependent) | |
2629 | replace_uncond_cands_and_profitable_phis (lookup_cand (c->dependent)); | |
2630 | } | |
2631 | \f | |
88ca9ea1 BS |
2632 | /* Count the number of candidates in the tree rooted at C that have |
2633 | not already been replaced under other interpretations. */ | |
2634 | ||
1dc3d6e9 | 2635 | static int |
88ca9ea1 BS |
2636 | count_candidates (slsr_cand_t c) |
2637 | { | |
2638 | unsigned count = cand_already_replaced (c) ? 0 : 1; | |
2639 | ||
2640 | if (c->sibling) | |
2641 | count += count_candidates (lookup_cand (c->sibling)); | |
2642 | ||
2643 | if (c->dependent) | |
2644 | count += count_candidates (lookup_cand (c->dependent)); | |
2645 | ||
2646 | return count; | |
2647 | } | |
2648 | ||
2649 | /* Increase the count of INCREMENT by one in the increment vector. | |
9b92d12b BS |
2650 | INCREMENT is associated with candidate C. If INCREMENT is to be |
2651 | conditionally executed as part of a conditional candidate replacement, | |
2652 | IS_PHI_ADJUST is true, otherwise false. If an initializer | |
88ca9ea1 BS |
2653 | T_0 = stride * I is provided by a candidate that dominates all |
2654 | candidates with the same increment, also record T_0 for subsequent use. */ | |
2655 | ||
2656 | static void | |
807e902e | 2657 | record_increment (slsr_cand_t c, widest_int increment, bool is_phi_adjust) |
88ca9ea1 BS |
2658 | { |
2659 | bool found = false; | |
2660 | unsigned i; | |
2661 | ||
2662 | /* Treat increments that differ only in sign as identical so as to | |
2663 | share initializers, unless we are generating pointer arithmetic. */ | |
807e902e | 2664 | if (!address_arithmetic_p && wi::neg_p (increment)) |
27bcd47c | 2665 | increment = -increment; |
88ca9ea1 BS |
2666 | |
2667 | for (i = 0; i < incr_vec_len; i++) | |
2668 | { | |
27bcd47c | 2669 | if (incr_vec[i].incr == increment) |
88ca9ea1 BS |
2670 | { |
2671 | incr_vec[i].count++; | |
2672 | found = true; | |
2673 | ||
2674 | /* If we previously recorded an initializer that doesn't | |
2675 | dominate this candidate, it's not going to be useful to | |
2676 | us after all. */ | |
2677 | if (incr_vec[i].initializer | |
2678 | && !dominated_by_p (CDI_DOMINATORS, | |
2679 | gimple_bb (c->cand_stmt), | |
2680 | incr_vec[i].init_bb)) | |
2681 | { | |
2682 | incr_vec[i].initializer = NULL_TREE; | |
2683 | incr_vec[i].init_bb = NULL; | |
2684 | } | |
2685 | ||
2686 | break; | |
2687 | } | |
2688 | } | |
2689 | ||
7bf55a70 | 2690 | if (!found && incr_vec_len < MAX_INCR_VEC_LEN - 1) |
88ca9ea1 BS |
2691 | { |
2692 | /* The first time we see an increment, create the entry for it. | |
2693 | If this is the root candidate which doesn't have a basis, set | |
2694 | the count to zero. We're only processing it so it can possibly | |
2695 | provide an initializer for other candidates. */ | |
2696 | incr_vec[incr_vec_len].incr = increment; | |
9b92d12b | 2697 | incr_vec[incr_vec_len].count = c->basis || is_phi_adjust ? 1 : 0; |
88ca9ea1 BS |
2698 | incr_vec[incr_vec_len].cost = COST_INFINITE; |
2699 | ||
2700 | /* Optimistically record the first occurrence of this increment | |
2701 | as providing an initializer (if it does); we will revise this | |
2702 | opinion later if it doesn't dominate all other occurrences. | |
a929f266 | 2703 | Exception: increments of 0, 1 never need initializers; |
806696eb | 2704 | and phi adjustments don't ever provide initializers. */ |
88ca9ea1 | 2705 | if (c->kind == CAND_ADD |
9b92d12b | 2706 | && !is_phi_adjust |
27bcd47c | 2707 | && c->index == increment |
a929f266 | 2708 | && (increment > 1 || increment < 0) |
7b8265ba JJ |
2709 | && (gimple_assign_rhs_code (c->cand_stmt) == PLUS_EXPR |
2710 | || gimple_assign_rhs_code (c->cand_stmt) == POINTER_PLUS_EXPR)) | |
88ca9ea1 | 2711 | { |
7b8265ba | 2712 | tree t0 = NULL_TREE; |
88ca9ea1 BS |
2713 | tree rhs1 = gimple_assign_rhs1 (c->cand_stmt); |
2714 | tree rhs2 = gimple_assign_rhs2 (c->cand_stmt); | |
2715 | if (operand_equal_p (rhs1, c->base_expr, 0)) | |
2716 | t0 = rhs2; | |
7b8265ba | 2717 | else if (operand_equal_p (rhs2, c->base_expr, 0)) |
88ca9ea1 | 2718 | t0 = rhs1; |
7b8265ba JJ |
2719 | if (t0 |
2720 | && SSA_NAME_DEF_STMT (t0) | |
2721 | && gimple_bb (SSA_NAME_DEF_STMT (t0))) | |
88ca9ea1 BS |
2722 | { |
2723 | incr_vec[incr_vec_len].initializer = t0; | |
2724 | incr_vec[incr_vec_len++].init_bb | |
2725 | = gimple_bb (SSA_NAME_DEF_STMT (t0)); | |
2726 | } | |
2727 | else | |
2728 | { | |
2729 | incr_vec[incr_vec_len].initializer = NULL_TREE; | |
2730 | incr_vec[incr_vec_len++].init_bb = NULL; | |
2731 | } | |
2732 | } | |
2733 | else | |
2734 | { | |
2735 | incr_vec[incr_vec_len].initializer = NULL_TREE; | |
2736 | incr_vec[incr_vec_len++].init_bb = NULL; | |
2737 | } | |
2738 | } | |
2739 | } | |
2740 | ||
65d3dce8 | 2741 | /* Recursive helper function for record_phi_increments. */ |
9b92d12b BS |
2742 | |
2743 | static void | |
65d3dce8 | 2744 | record_phi_increments_1 (slsr_cand_t basis, gimple *phi) |
9b92d12b BS |
2745 | { |
2746 | unsigned i; | |
5b3d5f76 | 2747 | slsr_cand_t phi_cand = *stmt_cand_map->get (phi); |
9b92d12b | 2748 | |
65d3dce8 BS |
2749 | if (phi_cand->visited) |
2750 | return; | |
2751 | phi_cand->visited = 1; | |
2752 | ||
9b92d12b BS |
2753 | for (i = 0; i < gimple_phi_num_args (phi); i++) |
2754 | { | |
2755 | tree arg = gimple_phi_arg_def (phi, i); | |
3146c60f | 2756 | gimple *arg_def = SSA_NAME_DEF_STMT (arg); |
9b92d12b | 2757 | |
3146c60f WS |
2758 | if (gimple_code (arg_def) == GIMPLE_PHI) |
2759 | record_phi_increments_1 (basis, arg_def); | |
2760 | else | |
9b92d12b | 2761 | { |
3146c60f | 2762 | widest_int diff; |
9b92d12b | 2763 | |
3146c60f WS |
2764 | if (operand_equal_p (arg, phi_cand->base_expr, 0)) |
2765 | { | |
2766 | diff = -basis->index; | |
2767 | record_increment (phi_cand, diff, PHI_ADJUST); | |
2768 | } | |
9b92d12b BS |
2769 | else |
2770 | { | |
2771 | slsr_cand_t arg_cand = base_cand_from_table (arg); | |
3146c60f | 2772 | diff = arg_cand->index - basis->index; |
9b92d12b BS |
2773 | record_increment (arg_cand, diff, PHI_ADJUST); |
2774 | } | |
2775 | } | |
2776 | } | |
2777 | } | |
2778 | ||
65d3dce8 BS |
2779 | /* Given phi statement PHI that hides a candidate from its BASIS, find |
2780 | the increments along each incoming arc (recursively handling additional | |
2781 | phis that may be present) and record them. These increments are the | |
2782 | difference in index between the index-adjusting statements and the | |
2783 | index of the basis. */ | |
2784 | ||
2785 | static void | |
2786 | record_phi_increments (slsr_cand_t basis, gimple *phi) | |
2787 | { | |
2788 | record_phi_increments_1 (basis, phi); | |
2789 | clear_visited (as_a <gphi *> (phi)); | |
2790 | } | |
2791 | ||
88ca9ea1 BS |
2792 | /* Determine how many times each unique increment occurs in the set |
2793 | of candidates rooted at C's parent, recording the data in the | |
2794 | increment vector. For each unique increment I, if an initializer | |
2795 | T_0 = stride * I is provided by a candidate that dominates all | |
2796 | candidates with the same increment, also record T_0 for subsequent | |
2797 | use. */ | |
2798 | ||
2799 | static void | |
2800 | record_increments (slsr_cand_t c) | |
2801 | { | |
2802 | if (!cand_already_replaced (c)) | |
9b92d12b BS |
2803 | { |
2804 | if (!phi_dependent_cand_p (c)) | |
2805 | record_increment (c, cand_increment (c), NOT_PHI_ADJUST); | |
2806 | else | |
2807 | { | |
2808 | /* A candidate with a basis hidden by a phi will have one | |
2809 | increment for its relationship to the index represented by | |
2810 | the phi, and potentially additional increments along each | |
2811 | incoming edge. For the root of the dependency tree (which | |
2812 | has no basis), process just the initial index in case it has | |
2813 | an initializer that can be used by subsequent candidates. */ | |
2814 | record_increment (c, c->index, NOT_PHI_ADJUST); | |
2815 | ||
2816 | if (c->basis) | |
2817 | record_phi_increments (lookup_cand (c->basis), | |
2818 | lookup_cand (c->def_phi)->cand_stmt); | |
2819 | } | |
2820 | } | |
88ca9ea1 BS |
2821 | |
2822 | if (c->sibling) | |
2823 | record_increments (lookup_cand (c->sibling)); | |
2824 | ||
2825 | if (c->dependent) | |
2826 | record_increments (lookup_cand (c->dependent)); | |
2827 | } | |
2828 | ||
65d3dce8 | 2829 | /* Recursive helper function for phi_incr_cost. */ |
9b92d12b BS |
2830 | |
2831 | static int | |
65d3dce8 BS |
2832 | phi_incr_cost_1 (slsr_cand_t c, const widest_int &incr, gimple *phi, |
2833 | int *savings) | |
9b92d12b BS |
2834 | { |
2835 | unsigned i; | |
2836 | int cost = 0; | |
2837 | slsr_cand_t basis = lookup_cand (c->basis); | |
5b3d5f76 | 2838 | slsr_cand_t phi_cand = *stmt_cand_map->get (phi); |
9b92d12b | 2839 | |
65d3dce8 BS |
2840 | if (phi_cand->visited) |
2841 | return 0; | |
2842 | phi_cand->visited = 1; | |
2843 | ||
9b92d12b BS |
2844 | for (i = 0; i < gimple_phi_num_args (phi); i++) |
2845 | { | |
2846 | tree arg = gimple_phi_arg_def (phi, i); | |
3146c60f | 2847 | gimple *arg_def = SSA_NAME_DEF_STMT (arg); |
9b92d12b | 2848 | |
3146c60f | 2849 | if (gimple_code (arg_def) == GIMPLE_PHI) |
9b92d12b | 2850 | { |
3146c60f WS |
2851 | int feeding_savings = 0; |
2852 | tree feeding_var = gimple_phi_result (arg_def); | |
2853 | cost += phi_incr_cost_1 (c, incr, arg_def, &feeding_savings); | |
2854 | if (uses_consumed_by_stmt (feeding_var, phi)) | |
2855 | *savings += feeding_savings; | |
2856 | } | |
2857 | else | |
2858 | { | |
2859 | widest_int diff; | |
2860 | slsr_cand_t arg_cand; | |
2861 | ||
2862 | /* When the PHI argument is just a pass-through to the base | |
2863 | expression of the hidden basis, the difference is zero minus | |
2864 | the index of the basis. There is no potential savings by | |
2865 | eliminating a statement in this case. */ | |
2866 | if (operand_equal_p (arg, phi_cand->base_expr, 0)) | |
9b92d12b | 2867 | { |
3146c60f WS |
2868 | arg_cand = (slsr_cand_t)NULL; |
2869 | diff = -basis->index; | |
9b92d12b BS |
2870 | } |
2871 | else | |
2872 | { | |
3146c60f WS |
2873 | arg_cand = base_cand_from_table (arg); |
2874 | diff = arg_cand->index - basis->index; | |
2875 | } | |
2876 | ||
2877 | if (incr == diff) | |
2878 | { | |
2879 | tree basis_lhs = gimple_assign_lhs (basis->cand_stmt); | |
2880 | cost += add_cost (true, TYPE_MODE (TREE_TYPE (basis_lhs))); | |
2881 | if (arg_cand) | |
9b92d12b | 2882 | { |
9b92d12b | 2883 | tree lhs = gimple_assign_lhs (arg_cand->cand_stmt); |
df11b2ea | 2884 | if (uses_consumed_by_stmt (lhs, phi)) |
9b92d12b BS |
2885 | *savings += stmt_cost (arg_cand->cand_stmt, true); |
2886 | } | |
2887 | } | |
2888 | } | |
2889 | } | |
2890 | ||
2891 | return cost; | |
2892 | } | |
2893 | ||
65d3dce8 BS |
2894 | /* Add up and return the costs of introducing add statements that |
2895 | require the increment INCR on behalf of candidate C and phi | |
2896 | statement PHI. Accumulate into *SAVINGS the potential savings | |
2897 | from removing existing statements that feed PHI and have no other | |
2898 | uses. */ | |
2899 | ||
2900 | static int | |
2901 | phi_incr_cost (slsr_cand_t c, const widest_int &incr, gimple *phi, | |
2902 | int *savings) | |
2903 | { | |
2904 | int retval = phi_incr_cost_1 (c, incr, phi, savings); | |
2905 | clear_visited (as_a <gphi *> (phi)); | |
2906 | return retval; | |
2907 | } | |
2908 | ||
88ca9ea1 BS |
2909 | /* Return the first candidate in the tree rooted at C that has not |
2910 | already been replaced, favoring siblings over dependents. */ | |
2911 | ||
2912 | static slsr_cand_t | |
2913 | unreplaced_cand_in_tree (slsr_cand_t c) | |
2914 | { | |
2915 | if (!cand_already_replaced (c)) | |
2916 | return c; | |
2917 | ||
2918 | if (c->sibling) | |
2919 | { | |
2920 | slsr_cand_t sib = unreplaced_cand_in_tree (lookup_cand (c->sibling)); | |
2921 | if (sib) | |
2922 | return sib; | |
2923 | } | |
2924 | ||
2925 | if (c->dependent) | |
2926 | { | |
2927 | slsr_cand_t dep = unreplaced_cand_in_tree (lookup_cand (c->dependent)); | |
2928 | if (dep) | |
2929 | return dep; | |
2930 | } | |
2931 | ||
2932 | return NULL; | |
2933 | } | |
2934 | ||
2935 | /* Return TRUE if the candidates in the tree rooted at C should be | |
2936 | optimized for speed, else FALSE. We estimate this based on the block | |
2937 | containing the most dominant candidate in the tree that has not yet | |
2938 | been replaced. */ | |
2939 | ||
2940 | static bool | |
2941 | optimize_cands_for_speed_p (slsr_cand_t c) | |
2942 | { | |
2943 | slsr_cand_t c2 = unreplaced_cand_in_tree (c); | |
2944 | gcc_assert (c2); | |
2945 | return optimize_bb_for_speed_p (gimple_bb (c2->cand_stmt)); | |
2946 | } | |
2947 | ||
2948 | /* Add COST_IN to the lowest cost of any dependent path starting at | |
2949 | candidate C or any of its siblings, counting only candidates along | |
2950 | such paths with increment INCR. Assume that replacing a candidate | |
2951 | reduces cost by REPL_SAVINGS. Also account for savings from any | |
9b92d12b BS |
2952 | statements that would go dead. If COUNT_PHIS is true, include |
2953 | costs of introducing feeding statements for conditional candidates. */ | |
88ca9ea1 BS |
2954 | |
2955 | static int | |
9b92d12b | 2956 | lowest_cost_path (int cost_in, int repl_savings, slsr_cand_t c, |
807e902e | 2957 | const widest_int &incr, bool count_phis) |
88ca9ea1 | 2958 | { |
9b92d12b | 2959 | int local_cost, sib_cost, savings = 0; |
807e902e | 2960 | widest_int cand_incr = cand_abs_increment (c); |
88ca9ea1 BS |
2961 | |
2962 | if (cand_already_replaced (c)) | |
2963 | local_cost = cost_in; | |
27bcd47c | 2964 | else if (incr == cand_incr) |
88ca9ea1 BS |
2965 | local_cost = cost_in - repl_savings - c->dead_savings; |
2966 | else | |
2967 | local_cost = cost_in - c->dead_savings; | |
2968 | ||
9b92d12b BS |
2969 | if (count_phis |
2970 | && phi_dependent_cand_p (c) | |
2971 | && !cand_already_replaced (c)) | |
2972 | { | |
355fe088 | 2973 | gimple *phi = lookup_cand (c->def_phi)->cand_stmt; |
9b92d12b BS |
2974 | local_cost += phi_incr_cost (c, incr, phi, &savings); |
2975 | ||
df11b2ea | 2976 | if (uses_consumed_by_stmt (gimple_phi_result (phi), c->cand_stmt)) |
9b92d12b BS |
2977 | local_cost -= savings; |
2978 | } | |
2979 | ||
88ca9ea1 BS |
2980 | if (c->dependent) |
2981 | local_cost = lowest_cost_path (local_cost, repl_savings, | |
9b92d12b BS |
2982 | lookup_cand (c->dependent), incr, |
2983 | count_phis); | |
88ca9ea1 BS |
2984 | |
2985 | if (c->sibling) | |
2986 | { | |
2987 | sib_cost = lowest_cost_path (cost_in, repl_savings, | |
9b92d12b BS |
2988 | lookup_cand (c->sibling), incr, |
2989 | count_phis); | |
88ca9ea1 BS |
2990 | local_cost = MIN (local_cost, sib_cost); |
2991 | } | |
2992 | ||
2993 | return local_cost; | |
2994 | } | |
2995 | ||
2996 | /* Compute the total savings that would accrue from all replacements | |
2997 | in the candidate tree rooted at C, counting only candidates with | |
2998 | increment INCR. Assume that replacing a candidate reduces cost | |
2999 | by REPL_SAVINGS. Also account for savings from statements that | |
3000 | would go dead. */ | |
3001 | ||
3002 | static int | |
807e902e | 3003 | total_savings (int repl_savings, slsr_cand_t c, const widest_int &incr, |
9b92d12b | 3004 | bool count_phis) |
88ca9ea1 BS |
3005 | { |
3006 | int savings = 0; | |
807e902e | 3007 | widest_int cand_incr = cand_abs_increment (c); |
88ca9ea1 | 3008 | |
27bcd47c | 3009 | if (incr == cand_incr && !cand_already_replaced (c)) |
88ca9ea1 BS |
3010 | savings += repl_savings + c->dead_savings; |
3011 | ||
9b92d12b BS |
3012 | if (count_phis |
3013 | && phi_dependent_cand_p (c) | |
3014 | && !cand_already_replaced (c)) | |
3015 | { | |
3016 | int phi_savings = 0; | |
355fe088 | 3017 | gimple *phi = lookup_cand (c->def_phi)->cand_stmt; |
9b92d12b BS |
3018 | savings -= phi_incr_cost (c, incr, phi, &phi_savings); |
3019 | ||
df11b2ea | 3020 | if (uses_consumed_by_stmt (gimple_phi_result (phi), c->cand_stmt)) |
9b92d12b BS |
3021 | savings += phi_savings; |
3022 | } | |
3023 | ||
88ca9ea1 | 3024 | if (c->dependent) |
9b92d12b BS |
3025 | savings += total_savings (repl_savings, lookup_cand (c->dependent), incr, |
3026 | count_phis); | |
88ca9ea1 BS |
3027 | |
3028 | if (c->sibling) | |
9b92d12b BS |
3029 | savings += total_savings (repl_savings, lookup_cand (c->sibling), incr, |
3030 | count_phis); | |
88ca9ea1 BS |
3031 | |
3032 | return savings; | |
3033 | } | |
3034 | ||
3035 | /* Use target-specific costs to determine and record which increments | |
3036 | in the current candidate tree are profitable to replace, assuming | |
3037 | MODE and SPEED. FIRST_DEP is the first dependent of the root of | |
3038 | the candidate tree. | |
3039 | ||
3040 | One slight limitation here is that we don't account for the possible | |
3041 | introduction of casts in some cases. See replace_one_candidate for | |
3042 | the cases where these are introduced. This should probably be cleaned | |
3043 | up sometime. */ | |
3044 | ||
3045 | static void | |
ef4bddc2 | 3046 | analyze_increments (slsr_cand_t first_dep, machine_mode mode, bool speed) |
88ca9ea1 BS |
3047 | { |
3048 | unsigned i; | |
3049 | ||
3050 | for (i = 0; i < incr_vec_len; i++) | |
3051 | { | |
27bcd47c | 3052 | HOST_WIDE_INT incr = incr_vec[i].incr.to_shwi (); |
88ca9ea1 BS |
3053 | |
3054 | /* If somehow this increment is bigger than a HWI, we won't | |
3055 | be optimizing candidates that use it. And if the increment | |
3056 | has a count of zero, nothing will be done with it. */ | |
807e902e | 3057 | if (!wi::fits_shwi_p (incr_vec[i].incr) || !incr_vec[i].count) |
88ca9ea1 BS |
3058 | incr_vec[i].cost = COST_INFINITE; |
3059 | ||
3060 | /* Increments of 0, 1, and -1 are always profitable to replace, | |
3061 | because they always replace a multiply or add with an add or | |
3062 | copy, and may cause one or more existing instructions to go | |
3063 | dead. Exception: -1 can't be assumed to be profitable for | |
3064 | pointer addition. */ | |
3065 | else if (incr == 0 | |
3066 | || incr == 1 | |
3067 | || (incr == -1 | |
457e189d | 3068 | && !POINTER_TYPE_P (first_dep->cand_type))) |
88ca9ea1 | 3069 | incr_vec[i].cost = COST_NEUTRAL; |
2d4035dc | 3070 | |
0b56e9ad BS |
3071 | /* If we need to add an initializer, give up if a cast from the |
3072 | candidate's type to its stride's type can lose precision. | |
3073 | Note that this already takes into account that the stride may | |
3074 | have been cast to a wider type, in which case this test won't | |
3075 | fire. Example: | |
6b5eea61 BS |
3076 | |
3077 | short int _1; | |
3078 | _2 = (int) _1; | |
3079 | _3 = _2 * 10; | |
0b56e9ad | 3080 | _4 = x + _3; ADD: x + (10 * (int)_1) : int |
6b5eea61 | 3081 | _5 = _2 * 15; |
0b56e9ad BS |
3082 | _6 = x + _5; ADD: x + (15 * (int)_1) : int |
3083 | ||
3084 | Although the stride was a short int initially, the stride | |
3085 | used in the analysis has been widened to an int, and such | |
3086 | widening will be done in the initializer as well. */ | |
6b5eea61 BS |
3087 | else if (!incr_vec[i].initializer |
3088 | && TREE_CODE (first_dep->stride) != INTEGER_CST | |
0b56e9ad BS |
3089 | && !legal_cast_p_1 (first_dep->stride_type, |
3090 | TREE_TYPE (gimple_assign_lhs | |
3091 | (first_dep->cand_stmt)))) | |
6b5eea61 BS |
3092 | incr_vec[i].cost = COST_INFINITE; |
3093 | ||
50251425 BS |
3094 | /* If we need to add an initializer, make sure we don't introduce |
3095 | a multiply by a pointer type, which can happen in certain cast | |
0b56e9ad | 3096 | scenarios. */ |
50251425 BS |
3097 | else if (!incr_vec[i].initializer |
3098 | && TREE_CODE (first_dep->stride) != INTEGER_CST | |
0b56e9ad | 3099 | && POINTER_TYPE_P (first_dep->stride_type)) |
50251425 BS |
3100 | incr_vec[i].cost = COST_INFINITE; |
3101 | ||
88ca9ea1 BS |
3102 | /* For any other increment, if this is a multiply candidate, we |
3103 | must introduce a temporary T and initialize it with | |
3104 | T_0 = stride * increment. When optimizing for speed, walk the | |
3105 | candidate tree to calculate the best cost reduction along any | |
3106 | path; if it offsets the fixed cost of inserting the initializer, | |
3107 | replacing the increment is profitable. When optimizing for | |
3108 | size, instead calculate the total cost reduction from replacing | |
3109 | all candidates with this increment. */ | |
3110 | else if (first_dep->kind == CAND_MULT) | |
3111 | { | |
3112 | int cost = mult_by_coeff_cost (incr, mode, speed); | |
d5f6f04c BS |
3113 | int repl_savings; |
3114 | ||
3115 | if (tree_fits_shwi_p (first_dep->stride)) | |
3116 | { | |
3117 | HOST_WIDE_INT hwi_stride = tree_to_shwi (first_dep->stride); | |
3118 | repl_savings = mult_by_coeff_cost (hwi_stride, mode, speed); | |
3119 | } | |
3120 | else | |
3121 | repl_savings = mul_cost (speed, mode); | |
3122 | repl_savings -= add_cost (speed, mode); | |
3123 | ||
88ca9ea1 BS |
3124 | if (speed) |
3125 | cost = lowest_cost_path (cost, repl_savings, first_dep, | |
9b92d12b | 3126 | incr_vec[i].incr, COUNT_PHIS); |
88ca9ea1 | 3127 | else |
9b92d12b BS |
3128 | cost -= total_savings (repl_savings, first_dep, incr_vec[i].incr, |
3129 | COUNT_PHIS); | |
88ca9ea1 BS |
3130 | |
3131 | incr_vec[i].cost = cost; | |
3132 | } | |
3133 | ||
3134 | /* If this is an add candidate, the initializer may already | |
3135 | exist, so only calculate the cost of the initializer if it | |
3136 | doesn't. We are replacing one add with another here, so the | |
3137 | known replacement savings is zero. We will account for removal | |
3138 | of dead instructions in lowest_cost_path or total_savings. */ | |
3139 | else | |
3140 | { | |
3141 | int cost = 0; | |
3142 | if (!incr_vec[i].initializer) | |
3143 | cost = mult_by_coeff_cost (incr, mode, speed); | |
3144 | ||
3145 | if (speed) | |
9b92d12b BS |
3146 | cost = lowest_cost_path (cost, 0, first_dep, incr_vec[i].incr, |
3147 | DONT_COUNT_PHIS); | |
88ca9ea1 | 3148 | else |
9b92d12b BS |
3149 | cost -= total_savings (0, first_dep, incr_vec[i].incr, |
3150 | DONT_COUNT_PHIS); | |
88ca9ea1 BS |
3151 | |
3152 | incr_vec[i].cost = cost; | |
3153 | } | |
3154 | } | |
3155 | } | |
3156 | ||
3157 | /* Return the nearest common dominator of BB1 and BB2. If the blocks | |
3158 | are identical, return the earlier of C1 and C2 in *WHERE. Otherwise, | |
3159 | if the NCD matches BB1, return C1 in *WHERE; if the NCD matches BB2, | |
3160 | return C2 in *WHERE; and if the NCD matches neither, return NULL in | |
3161 | *WHERE. Note: It is possible for one of C1 and C2 to be NULL. */ | |
3162 | ||
3163 | static basic_block | |
3164 | ncd_for_two_cands (basic_block bb1, basic_block bb2, | |
3165 | slsr_cand_t c1, slsr_cand_t c2, slsr_cand_t *where) | |
3166 | { | |
3167 | basic_block ncd; | |
3168 | ||
3169 | if (!bb1) | |
3170 | { | |
3171 | *where = c2; | |
3172 | return bb2; | |
3173 | } | |
3174 | ||
3175 | if (!bb2) | |
3176 | { | |
3177 | *where = c1; | |
3178 | return bb1; | |
3179 | } | |
3180 | ||
3181 | ncd = nearest_common_dominator (CDI_DOMINATORS, bb1, bb2); | |
3182 | ||
3183 | /* If both candidates are in the same block, the earlier | |
3184 | candidate wins. */ | |
3185 | if (bb1 == ncd && bb2 == ncd) | |
3186 | { | |
3187 | if (!c1 || (c2 && c2->cand_num < c1->cand_num)) | |
3188 | *where = c2; | |
3189 | else | |
3190 | *where = c1; | |
3191 | } | |
3192 | ||
3193 | /* Otherwise, if one of them produced a candidate in the | |
3194 | dominator, that one wins. */ | |
3195 | else if (bb1 == ncd) | |
3196 | *where = c1; | |
3197 | ||
3198 | else if (bb2 == ncd) | |
3199 | *where = c2; | |
3200 | ||
3201 | /* If neither matches the dominator, neither wins. */ | |
3202 | else | |
3203 | *where = NULL; | |
3204 | ||
3205 | return ncd; | |
3206 | } | |
3207 | ||
9b92d12b BS |
3208 | /* Consider all candidates that feed PHI. Find the nearest common |
3209 | dominator of those candidates requiring the given increment INCR. | |
3210 | Further find and return the nearest common dominator of this result | |
3211 | with block NCD. If the returned block contains one or more of the | |
3212 | candidates, return the earliest candidate in the block in *WHERE. */ | |
3213 | ||
3214 | static basic_block | |
538dd0b7 | 3215 | ncd_with_phi (slsr_cand_t c, const widest_int &incr, gphi *phi, |
9b92d12b BS |
3216 | basic_block ncd, slsr_cand_t *where) |
3217 | { | |
3218 | unsigned i; | |
3219 | slsr_cand_t basis = lookup_cand (c->basis); | |
5b3d5f76 | 3220 | slsr_cand_t phi_cand = *stmt_cand_map->get (phi); |
9b92d12b BS |
3221 | |
3222 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
3223 | { | |
3224 | tree arg = gimple_phi_arg_def (phi, i); | |
3146c60f | 3225 | gimple *arg_def = SSA_NAME_DEF_STMT (arg); |
9b92d12b | 3226 | |
3146c60f WS |
3227 | if (gimple_code (arg_def) == GIMPLE_PHI) |
3228 | ncd = ncd_with_phi (c, incr, as_a <gphi *> (arg_def), ncd, where); | |
3229 | else | |
9b92d12b | 3230 | { |
3146c60f | 3231 | widest_int diff; |
9b92d12b | 3232 | |
3146c60f WS |
3233 | if (operand_equal_p (arg, phi_cand->base_expr, 0)) |
3234 | diff = -basis->index; | |
3235 | else | |
9b92d12b BS |
3236 | { |
3237 | slsr_cand_t arg_cand = base_cand_from_table (arg); | |
3146c60f | 3238 | diff = arg_cand->index - basis->index; |
9b92d12b | 3239 | } |
3146c60f WS |
3240 | |
3241 | basic_block pred = gimple_phi_arg_edge (phi, i)->src; | |
3242 | ||
3243 | if ((incr == diff) || (!address_arithmetic_p && incr == -diff)) | |
3244 | ncd = ncd_for_two_cands (ncd, pred, *where, NULL, where); | |
9b92d12b BS |
3245 | } |
3246 | } | |
3247 | ||
3248 | return ncd; | |
3249 | } | |
3250 | ||
3251 | /* Consider the candidate C together with any candidates that feed | |
3252 | C's phi dependence (if any). Find and return the nearest common | |
3253 | dominator of those candidates requiring the given increment INCR. | |
3254 | If the returned block contains one or more of the candidates, | |
3255 | return the earliest candidate in the block in *WHERE. */ | |
3256 | ||
3257 | static basic_block | |
807e902e | 3258 | ncd_of_cand_and_phis (slsr_cand_t c, const widest_int &incr, slsr_cand_t *where) |
9b92d12b BS |
3259 | { |
3260 | basic_block ncd = NULL; | |
3261 | ||
3262 | if (cand_abs_increment (c) == incr) | |
3263 | { | |
3264 | ncd = gimple_bb (c->cand_stmt); | |
3265 | *where = c; | |
3266 | } | |
3267 | ||
3268 | if (phi_dependent_cand_p (c)) | |
538dd0b7 DM |
3269 | ncd = ncd_with_phi (c, incr, |
3270 | as_a <gphi *> (lookup_cand (c->def_phi)->cand_stmt), | |
9b92d12b BS |
3271 | ncd, where); |
3272 | ||
3273 | return ncd; | |
3274 | } | |
3275 | ||
88ca9ea1 BS |
3276 | /* Consider all candidates in the tree rooted at C for which INCR |
3277 | represents the required increment of C relative to its basis. | |
3278 | Find and return the basic block that most nearly dominates all | |
3279 | such candidates. If the returned block contains one or more of | |
3280 | the candidates, return the earliest candidate in the block in | |
3281 | *WHERE. */ | |
3282 | ||
3283 | static basic_block | |
807e902e | 3284 | nearest_common_dominator_for_cands (slsr_cand_t c, const widest_int &incr, |
88ca9ea1 BS |
3285 | slsr_cand_t *where) |
3286 | { | |
3287 | basic_block sib_ncd = NULL, dep_ncd = NULL, this_ncd = NULL, ncd; | |
3288 | slsr_cand_t sib_where = NULL, dep_where = NULL, this_where = NULL, new_where; | |
88ca9ea1 BS |
3289 | |
3290 | /* First find the NCD of all siblings and dependents. */ | |
3291 | if (c->sibling) | |
3292 | sib_ncd = nearest_common_dominator_for_cands (lookup_cand (c->sibling), | |
3293 | incr, &sib_where); | |
3294 | if (c->dependent) | |
3295 | dep_ncd = nearest_common_dominator_for_cands (lookup_cand (c->dependent), | |
3296 | incr, &dep_where); | |
3297 | if (!sib_ncd && !dep_ncd) | |
3298 | { | |
3299 | new_where = NULL; | |
3300 | ncd = NULL; | |
3301 | } | |
3302 | else if (sib_ncd && !dep_ncd) | |
3303 | { | |
3304 | new_where = sib_where; | |
3305 | ncd = sib_ncd; | |
3306 | } | |
3307 | else if (dep_ncd && !sib_ncd) | |
3308 | { | |
3309 | new_where = dep_where; | |
3310 | ncd = dep_ncd; | |
3311 | } | |
3312 | else | |
3313 | ncd = ncd_for_two_cands (sib_ncd, dep_ncd, sib_where, | |
3314 | dep_where, &new_where); | |
3315 | ||
3316 | /* If the candidate's increment doesn't match the one we're interested | |
9b92d12b BS |
3317 | in (and nor do any increments for feeding defs of a phi-dependence), |
3318 | then the result depends only on siblings and dependents. */ | |
3319 | this_ncd = ncd_of_cand_and_phis (c, incr, &this_where); | |
88ca9ea1 | 3320 | |
9b92d12b | 3321 | if (!this_ncd || cand_already_replaced (c)) |
88ca9ea1 BS |
3322 | { |
3323 | *where = new_where; | |
3324 | return ncd; | |
3325 | } | |
3326 | ||
3327 | /* Otherwise, compare this candidate with the result from all siblings | |
3328 | and dependents. */ | |
88ca9ea1 BS |
3329 | ncd = ncd_for_two_cands (ncd, this_ncd, new_where, this_where, where); |
3330 | ||
3331 | return ncd; | |
3332 | } | |
3333 | ||
3334 | /* Return TRUE if the increment indexed by INDEX is profitable to replace. */ | |
3335 | ||
3336 | static inline bool | |
3337 | profitable_increment_p (unsigned index) | |
3338 | { | |
3339 | return (incr_vec[index].cost <= COST_NEUTRAL); | |
3340 | } | |
3341 | ||
3342 | /* For each profitable increment in the increment vector not equal to | |
3343 | 0 or 1 (or -1, for non-pointer arithmetic), find the nearest common | |
3344 | dominator of all statements in the candidate chain rooted at C | |
3345 | that require that increment, and insert an initializer | |
3346 | T_0 = stride * increment at that location. Record T_0 with the | |
3347 | increment record. */ | |
3348 | ||
3349 | static void | |
3350 | insert_initializers (slsr_cand_t c) | |
3351 | { | |
3352 | unsigned i; | |
88ca9ea1 BS |
3353 | |
3354 | for (i = 0; i < incr_vec_len; i++) | |
3355 | { | |
3356 | basic_block bb; | |
3357 | slsr_cand_t where = NULL; | |
538dd0b7 | 3358 | gassign *init_stmt; |
0b56e9ad BS |
3359 | gassign *cast_stmt = NULL; |
3360 | tree new_name, incr_tree, init_stride; | |
807e902e | 3361 | widest_int incr = incr_vec[i].incr; |
88ca9ea1 BS |
3362 | |
3363 | if (!profitable_increment_p (i) | |
807e902e KZ |
3364 | || incr == 1 |
3365 | || (incr == -1 | |
6e3d8cb4 | 3366 | && (!POINTER_TYPE_P (lookup_cand (c->basis)->cand_type))) |
807e902e | 3367 | || incr == 0) |
88ca9ea1 BS |
3368 | continue; |
3369 | ||
3370 | /* We may have already identified an existing initializer that | |
3371 | will suffice. */ | |
3372 | if (incr_vec[i].initializer) | |
3373 | { | |
3374 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3375 | { | |
3376 | fputs ("Using existing initializer: ", dump_file); | |
3377 | print_gimple_stmt (dump_file, | |
3378 | SSA_NAME_DEF_STMT (incr_vec[i].initializer), | |
4af78ef8 | 3379 | 0, TDF_NONE); |
88ca9ea1 BS |
3380 | } |
3381 | continue; | |
3382 | } | |
3383 | ||
3384 | /* Find the block that most closely dominates all candidates | |
3385 | with this increment. If there is at least one candidate in | |
3386 | that block, the earliest one will be returned in WHERE. */ | |
3387 | bb = nearest_common_dominator_for_cands (c, incr, &where); | |
3388 | ||
3e75ec3f BS |
3389 | /* If the NCD is not dominated by the block containing the |
3390 | definition of the stride, we can't legally insert a | |
3391 | single initializer. Mark the increment as unprofitable | |
3392 | so we don't make any replacements. FIXME: Multiple | |
3393 | initializers could be placed with more analysis. */ | |
3394 | gimple *stride_def = SSA_NAME_DEF_STMT (c->stride); | |
3395 | basic_block stride_bb = gimple_bb (stride_def); | |
3396 | ||
3397 | if (stride_bb && !dominated_by_p (CDI_DOMINATORS, bb, stride_bb)) | |
3398 | { | |
3399 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3400 | fprintf (dump_file, | |
3401 | "Initializer #%d cannot be legally placed\n", i); | |
3402 | incr_vec[i].cost = COST_INFINITE; | |
3403 | continue; | |
3404 | } | |
3405 | ||
0b56e9ad BS |
3406 | /* If the nominal stride has a different type than the recorded |
3407 | stride type, build a cast from the nominal stride to that type. */ | |
3408 | if (!types_compatible_p (TREE_TYPE (c->stride), c->stride_type)) | |
3409 | { | |
3410 | init_stride = make_temp_ssa_name (c->stride_type, NULL, "slsr"); | |
3411 | cast_stmt = gimple_build_assign (init_stride, NOP_EXPR, c->stride); | |
3412 | } | |
3413 | else | |
3414 | init_stride = c->stride; | |
3415 | ||
88ca9ea1 | 3416 | /* Create a new SSA name to hold the initializer's value. */ |
0b56e9ad | 3417 | new_name = make_temp_ssa_name (c->stride_type, NULL, "slsr"); |
88ca9ea1 BS |
3418 | incr_vec[i].initializer = new_name; |
3419 | ||
3420 | /* Create the initializer and insert it in the latest possible | |
3421 | dominating position. */ | |
0b56e9ad | 3422 | incr_tree = wide_int_to_tree (c->stride_type, incr); |
0d0e4a03 | 3423 | init_stmt = gimple_build_assign (new_name, MULT_EXPR, |
0b56e9ad | 3424 | init_stride, incr_tree); |
88ca9ea1 BS |
3425 | if (where) |
3426 | { | |
3427 | gimple_stmt_iterator gsi = gsi_for_stmt (where->cand_stmt); | |
0b56e9ad BS |
3428 | location_t loc = gimple_location (where->cand_stmt); |
3429 | ||
3430 | if (cast_stmt) | |
3431 | { | |
3432 | gsi_insert_before (&gsi, cast_stmt, GSI_SAME_STMT); | |
3433 | gimple_set_location (cast_stmt, loc); | |
3434 | } | |
3435 | ||
88ca9ea1 | 3436 | gsi_insert_before (&gsi, init_stmt, GSI_SAME_STMT); |
0b56e9ad | 3437 | gimple_set_location (init_stmt, loc); |
88ca9ea1 BS |
3438 | } |
3439 | else | |
3440 | { | |
3441 | gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
355fe088 | 3442 | gimple *basis_stmt = lookup_cand (c->basis)->cand_stmt; |
0b56e9ad | 3443 | location_t loc = gimple_location (basis_stmt); |
88ca9ea1 | 3444 | |
1697df8c | 3445 | if (!gsi_end_p (gsi) && stmt_ends_bb_p (gsi_stmt (gsi))) |
0b56e9ad BS |
3446 | { |
3447 | if (cast_stmt) | |
3448 | { | |
3449 | gsi_insert_before (&gsi, cast_stmt, GSI_SAME_STMT); | |
3450 | gimple_set_location (cast_stmt, loc); | |
3451 | } | |
3452 | gsi_insert_before (&gsi, init_stmt, GSI_SAME_STMT); | |
3453 | } | |
88ca9ea1 | 3454 | else |
0b56e9ad BS |
3455 | { |
3456 | if (cast_stmt) | |
3457 | { | |
3458 | gsi_insert_after (&gsi, cast_stmt, GSI_NEW_STMT); | |
3459 | gimple_set_location (cast_stmt, loc); | |
3460 | } | |
06f97084 | 3461 | gsi_insert_after (&gsi, init_stmt, GSI_NEW_STMT); |
0b56e9ad | 3462 | } |
88ca9ea1 BS |
3463 | |
3464 | gimple_set_location (init_stmt, gimple_location (basis_stmt)); | |
3465 | } | |
3466 | ||
3467 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3468 | { | |
0b56e9ad BS |
3469 | if (cast_stmt) |
3470 | { | |
3471 | fputs ("Inserting stride cast: ", dump_file); | |
ef6cb4c7 | 3472 | print_gimple_stmt (dump_file, cast_stmt, 0); |
0b56e9ad | 3473 | } |
88ca9ea1 | 3474 | fputs ("Inserting initializer: ", dump_file); |
ef6cb4c7 | 3475 | print_gimple_stmt (dump_file, init_stmt, 0); |
88ca9ea1 BS |
3476 | } |
3477 | } | |
3478 | } | |
3479 | ||
65d3dce8 | 3480 | /* Recursive helper function for all_phi_incrs_profitable. */ |
9b92d12b BS |
3481 | |
3482 | static bool | |
65d3dce8 | 3483 | all_phi_incrs_profitable_1 (slsr_cand_t c, gphi *phi, int *spread) |
9b92d12b BS |
3484 | { |
3485 | unsigned i; | |
3486 | slsr_cand_t basis = lookup_cand (c->basis); | |
5b3d5f76 | 3487 | slsr_cand_t phi_cand = *stmt_cand_map->get (phi); |
9b92d12b | 3488 | |
65d3dce8 BS |
3489 | if (phi_cand->visited) |
3490 | return true; | |
3491 | ||
3492 | phi_cand->visited = 1; | |
3493 | (*spread)++; | |
3494 | ||
c34923c4 BS |
3495 | /* If the basis doesn't dominate the PHI (including when the PHI is |
3496 | in the same block as the basis), we won't be able to create a PHI | |
3497 | using the basis here. */ | |
3498 | basic_block basis_bb = gimple_bb (basis->cand_stmt); | |
3499 | basic_block phi_bb = gimple_bb (phi); | |
3500 | ||
3501 | if (phi_bb == basis_bb | |
3502 | || !dominated_by_p (CDI_DOMINATORS, phi_bb, basis_bb)) | |
3503 | return false; | |
3504 | ||
9b92d12b BS |
3505 | for (i = 0; i < gimple_phi_num_args (phi); i++) |
3506 | { | |
c34923c4 BS |
3507 | /* If the PHI arg resides in a block not dominated by the basis, |
3508 | we won't be able to create a PHI using the basis here. */ | |
3509 | basic_block pred_bb = gimple_phi_arg_edge (phi, i)->src; | |
3510 | ||
3511 | if (!dominated_by_p (CDI_DOMINATORS, pred_bb, basis_bb)) | |
3512 | return false; | |
3513 | ||
9b92d12b | 3514 | tree arg = gimple_phi_arg_def (phi, i); |
3146c60f | 3515 | gimple *arg_def = SSA_NAME_DEF_STMT (arg); |
9b92d12b | 3516 | |
3146c60f | 3517 | if (gimple_code (arg_def) == GIMPLE_PHI) |
9b92d12b | 3518 | { |
3146c60f WS |
3519 | if (!all_phi_incrs_profitable_1 (c, as_a <gphi *> (arg_def), spread) |
3520 | || *spread > MAX_SPREAD) | |
3521 | return false; | |
3522 | } | |
3523 | else | |
3524 | { | |
3525 | int j; | |
3526 | widest_int increment; | |
9b92d12b | 3527 | |
3146c60f WS |
3528 | if (operand_equal_p (arg, phi_cand->base_expr, 0)) |
3529 | increment = -basis->index; | |
9b92d12b BS |
3530 | else |
3531 | { | |
9b92d12b | 3532 | slsr_cand_t arg_cand = base_cand_from_table (arg); |
3146c60f WS |
3533 | increment = arg_cand->index - basis->index; |
3534 | } | |
9b92d12b | 3535 | |
3146c60f WS |
3536 | if (!address_arithmetic_p && wi::neg_p (increment)) |
3537 | increment = -increment; | |
9b92d12b | 3538 | |
3146c60f | 3539 | j = incr_vec_index (increment); |
9b92d12b | 3540 | |
3146c60f WS |
3541 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3542 | { | |
3543 | fprintf (dump_file, " Conditional candidate %d, phi: ", | |
3544 | c->cand_num); | |
3545 | print_gimple_stmt (dump_file, phi, 0); | |
3546 | fputs (" increment: ", dump_file); | |
3547 | print_decs (increment, dump_file); | |
3548 | if (j < 0) | |
3549 | fprintf (dump_file, | |
3550 | "\n Not replaced; incr_vec overflow.\n"); | |
3551 | else { | |
3552 | fprintf (dump_file, "\n cost: %d\n", incr_vec[j].cost); | |
3553 | if (profitable_increment_p (j)) | |
3554 | fputs (" Replacing...\n", dump_file); | |
3555 | else | |
3556 | fputs (" Not replaced.\n", dump_file); | |
3557 | } | |
9b92d12b | 3558 | } |
3146c60f WS |
3559 | |
3560 | if (j < 0 || !profitable_increment_p (j)) | |
3561 | return false; | |
9b92d12b BS |
3562 | } |
3563 | } | |
3564 | ||
3565 | return true; | |
3566 | } | |
3567 | ||
65d3dce8 BS |
3568 | /* Return TRUE iff all required increments for candidates feeding PHI |
3569 | are profitable (and legal!) to replace on behalf of candidate C. */ | |
3570 | ||
3571 | static bool | |
3572 | all_phi_incrs_profitable (slsr_cand_t c, gphi *phi) | |
3573 | { | |
3574 | int spread = 0; | |
3575 | bool retval = all_phi_incrs_profitable_1 (c, phi, &spread); | |
3576 | clear_visited (phi); | |
3577 | return retval; | |
3578 | } | |
3579 | ||
88ca9ea1 BS |
3580 | /* Create a NOP_EXPR that copies FROM_EXPR into a new SSA name of |
3581 | type TO_TYPE, and insert it in front of the statement represented | |
3582 | by candidate C. Use *NEW_VAR to create the new SSA name. Return | |
3583 | the new SSA name. */ | |
3584 | ||
3585 | static tree | |
a7a7d10e | 3586 | introduce_cast_before_cand (slsr_cand_t c, tree to_type, tree from_expr) |
88ca9ea1 BS |
3587 | { |
3588 | tree cast_lhs; | |
538dd0b7 | 3589 | gassign *cast_stmt; |
88ca9ea1 BS |
3590 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); |
3591 | ||
a7a7d10e | 3592 | cast_lhs = make_temp_ssa_name (to_type, NULL, "slsr"); |
0d0e4a03 | 3593 | cast_stmt = gimple_build_assign (cast_lhs, NOP_EXPR, from_expr); |
88ca9ea1 BS |
3594 | gimple_set_location (cast_stmt, gimple_location (c->cand_stmt)); |
3595 | gsi_insert_before (&gsi, cast_stmt, GSI_SAME_STMT); | |
3596 | ||
3597 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3598 | { | |
3599 | fputs (" Inserting: ", dump_file); | |
ef6cb4c7 | 3600 | print_gimple_stmt (dump_file, cast_stmt, 0); |
88ca9ea1 BS |
3601 | } |
3602 | ||
3603 | return cast_lhs; | |
3604 | } | |
3605 | ||
3606 | /* Replace the RHS of the statement represented by candidate C with | |
3607 | NEW_CODE, NEW_RHS1, and NEW_RHS2, provided that to do so doesn't | |
3608 | leave C unchanged or just interchange its operands. The original | |
3609 | operation and operands are in OLD_CODE, OLD_RHS1, and OLD_RHS2. | |
3610 | If the replacement was made and we are doing a details dump, | |
3611 | return the revised statement, else NULL. */ | |
3612 | ||
355fe088 | 3613 | static gimple * |
88ca9ea1 BS |
3614 | replace_rhs_if_not_dup (enum tree_code new_code, tree new_rhs1, tree new_rhs2, |
3615 | enum tree_code old_code, tree old_rhs1, tree old_rhs2, | |
3616 | slsr_cand_t c) | |
3617 | { | |
3618 | if (new_code != old_code | |
3619 | || ((!operand_equal_p (new_rhs1, old_rhs1, 0) | |
3620 | || !operand_equal_p (new_rhs2, old_rhs2, 0)) | |
3621 | && (!operand_equal_p (new_rhs1, old_rhs2, 0) | |
3622 | || !operand_equal_p (new_rhs2, old_rhs1, 0)))) | |
3623 | { | |
3624 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); | |
d71dc990 | 3625 | slsr_cand_t cc = lookup_cand (c->first_interp); |
88ca9ea1 BS |
3626 | gimple_assign_set_rhs_with_ops (&gsi, new_code, new_rhs1, new_rhs2); |
3627 | update_stmt (gsi_stmt (gsi)); | |
d71dc990 | 3628 | while (cc) |
9fbbba71 | 3629 | { |
9fbbba71 | 3630 | cc->cand_stmt = gsi_stmt (gsi); |
98aaa1a6 | 3631 | cc = lookup_cand (cc->next_interp); |
9fbbba71 | 3632 | } |
88ca9ea1 BS |
3633 | |
3634 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3635 | return gsi_stmt (gsi); | |
3636 | } | |
3637 | ||
3638 | else if (dump_file && (dump_flags & TDF_DETAILS)) | |
3639 | fputs (" (duplicate, not actually replacing)\n", dump_file); | |
3640 | ||
3641 | return NULL; | |
3642 | } | |
3643 | ||
3644 | /* Strength-reduce the statement represented by candidate C by replacing | |
3645 | it with an equivalent addition or subtraction. I is the index into | |
3646 | the increment vector identifying C's increment. NEW_VAR is used to | |
3647 | create a new SSA name if a cast needs to be introduced. BASIS_NAME | |
3648 | is the rhs1 to use in creating the add/subtract. */ | |
3649 | ||
3650 | static void | |
a7a7d10e | 3651 | replace_one_candidate (slsr_cand_t c, unsigned i, tree basis_name) |
88ca9ea1 | 3652 | { |
355fe088 | 3653 | gimple *stmt_to_print = NULL; |
88ca9ea1 BS |
3654 | tree orig_rhs1, orig_rhs2; |
3655 | tree rhs2; | |
3656 | enum tree_code orig_code, repl_code; | |
807e902e | 3657 | widest_int cand_incr; |
88ca9ea1 BS |
3658 | |
3659 | orig_code = gimple_assign_rhs_code (c->cand_stmt); | |
3660 | orig_rhs1 = gimple_assign_rhs1 (c->cand_stmt); | |
3661 | orig_rhs2 = gimple_assign_rhs2 (c->cand_stmt); | |
3662 | cand_incr = cand_increment (c); | |
3663 | ||
9d502741 BS |
3664 | /* If orig_rhs2 is NULL, we have already replaced this in situ with |
3665 | a copy statement under another interpretation. */ | |
3666 | if (!orig_rhs2) | |
3667 | return; | |
3668 | ||
88ca9ea1 BS |
3669 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3670 | { | |
3671 | fputs ("Replacing: ", dump_file); | |
ef6cb4c7 | 3672 | print_gimple_stmt (dump_file, c->cand_stmt, 0); |
88ca9ea1 BS |
3673 | stmt_to_print = c->cand_stmt; |
3674 | } | |
3675 | ||
3676 | if (address_arithmetic_p) | |
3677 | repl_code = POINTER_PLUS_EXPR; | |
3678 | else | |
3679 | repl_code = PLUS_EXPR; | |
3680 | ||
3681 | /* If the increment has an initializer T_0, replace the candidate | |
3682 | statement with an add of the basis name and the initializer. */ | |
3683 | if (incr_vec[i].initializer) | |
3684 | { | |
3685 | tree init_type = TREE_TYPE (incr_vec[i].initializer); | |
3686 | tree orig_type = TREE_TYPE (orig_rhs2); | |
3687 | ||
3688 | if (types_compatible_p (orig_type, init_type)) | |
3689 | rhs2 = incr_vec[i].initializer; | |
3690 | else | |
3691 | rhs2 = introduce_cast_before_cand (c, orig_type, | |
a7a7d10e | 3692 | incr_vec[i].initializer); |
88ca9ea1 | 3693 | |
27bcd47c | 3694 | if (incr_vec[i].incr != cand_incr) |
88ca9ea1 BS |
3695 | { |
3696 | gcc_assert (repl_code == PLUS_EXPR); | |
3697 | repl_code = MINUS_EXPR; | |
3698 | } | |
3699 | ||
3700 | stmt_to_print = replace_rhs_if_not_dup (repl_code, basis_name, rhs2, | |
3701 | orig_code, orig_rhs1, orig_rhs2, | |
3702 | c); | |
3703 | } | |
3704 | ||
3705 | /* Otherwise, the increment is one of -1, 0, and 1. Replace | |
3706 | with a subtract of the stride from the basis name, a copy | |
3707 | from the basis name, or an add of the stride to the basis | |
3708 | name, respectively. It may be necessary to introduce a | |
3709 | cast (or reuse an existing cast). */ | |
807e902e | 3710 | else if (cand_incr == 1) |
88ca9ea1 BS |
3711 | { |
3712 | tree stride_type = TREE_TYPE (c->stride); | |
3713 | tree orig_type = TREE_TYPE (orig_rhs2); | |
3714 | ||
3715 | if (types_compatible_p (orig_type, stride_type)) | |
3716 | rhs2 = c->stride; | |
3717 | else | |
a7a7d10e | 3718 | rhs2 = introduce_cast_before_cand (c, orig_type, c->stride); |
88ca9ea1 BS |
3719 | |
3720 | stmt_to_print = replace_rhs_if_not_dup (repl_code, basis_name, rhs2, | |
3721 | orig_code, orig_rhs1, orig_rhs2, | |
3722 | c); | |
3723 | } | |
3724 | ||
807e902e | 3725 | else if (cand_incr == -1) |
88ca9ea1 BS |
3726 | { |
3727 | tree stride_type = TREE_TYPE (c->stride); | |
3728 | tree orig_type = TREE_TYPE (orig_rhs2); | |
3729 | gcc_assert (repl_code != POINTER_PLUS_EXPR); | |
3730 | ||
3731 | if (types_compatible_p (orig_type, stride_type)) | |
3732 | rhs2 = c->stride; | |
3733 | else | |
a7a7d10e | 3734 | rhs2 = introduce_cast_before_cand (c, orig_type, c->stride); |
88ca9ea1 BS |
3735 | |
3736 | if (orig_code != MINUS_EXPR | |
3737 | || !operand_equal_p (basis_name, orig_rhs1, 0) | |
3738 | || !operand_equal_p (rhs2, orig_rhs2, 0)) | |
3739 | { | |
3740 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); | |
d71dc990 | 3741 | slsr_cand_t cc = lookup_cand (c->first_interp); |
88ca9ea1 BS |
3742 | gimple_assign_set_rhs_with_ops (&gsi, MINUS_EXPR, basis_name, rhs2); |
3743 | update_stmt (gsi_stmt (gsi)); | |
d71dc990 | 3744 | while (cc) |
9fbbba71 | 3745 | { |
9fbbba71 | 3746 | cc->cand_stmt = gsi_stmt (gsi); |
98aaa1a6 | 3747 | cc = lookup_cand (cc->next_interp); |
9fbbba71 | 3748 | } |
88ca9ea1 BS |
3749 | |
3750 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3751 | stmt_to_print = gsi_stmt (gsi); | |
3752 | } | |
3753 | else if (dump_file && (dump_flags & TDF_DETAILS)) | |
3754 | fputs (" (duplicate, not actually replacing)\n", dump_file); | |
3755 | } | |
3756 | ||
807e902e | 3757 | else if (cand_incr == 0) |
88ca9ea1 BS |
3758 | { |
3759 | tree lhs = gimple_assign_lhs (c->cand_stmt); | |
3760 | tree lhs_type = TREE_TYPE (lhs); | |
3761 | tree basis_type = TREE_TYPE (basis_name); | |
3762 | ||
3763 | if (types_compatible_p (lhs_type, basis_type)) | |
3764 | { | |
538dd0b7 | 3765 | gassign *copy_stmt = gimple_build_assign (lhs, basis_name); |
88ca9ea1 | 3766 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); |
d71dc990 | 3767 | slsr_cand_t cc = lookup_cand (c->first_interp); |
88ca9ea1 BS |
3768 | gimple_set_location (copy_stmt, gimple_location (c->cand_stmt)); |
3769 | gsi_replace (&gsi, copy_stmt, false); | |
d71dc990 | 3770 | while (cc) |
9fbbba71 | 3771 | { |
9fbbba71 | 3772 | cc->cand_stmt = copy_stmt; |
98aaa1a6 | 3773 | cc = lookup_cand (cc->next_interp); |
9fbbba71 | 3774 | } |
88ca9ea1 BS |
3775 | |
3776 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3777 | stmt_to_print = copy_stmt; | |
3778 | } | |
3779 | else | |
3780 | { | |
3781 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); | |
0d0e4a03 | 3782 | gassign *cast_stmt = gimple_build_assign (lhs, NOP_EXPR, basis_name); |
d71dc990 | 3783 | slsr_cand_t cc = lookup_cand (c->first_interp); |
88ca9ea1 BS |
3784 | gimple_set_location (cast_stmt, gimple_location (c->cand_stmt)); |
3785 | gsi_replace (&gsi, cast_stmt, false); | |
d71dc990 | 3786 | while (cc) |
9fbbba71 | 3787 | { |
9fbbba71 | 3788 | cc->cand_stmt = cast_stmt; |
98aaa1a6 | 3789 | cc = lookup_cand (cc->next_interp); |
9fbbba71 | 3790 | } |
88ca9ea1 BS |
3791 | |
3792 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3793 | stmt_to_print = cast_stmt; | |
3794 | } | |
3795 | } | |
3796 | else | |
3797 | gcc_unreachable (); | |
3798 | ||
3799 | if (dump_file && (dump_flags & TDF_DETAILS) && stmt_to_print) | |
3800 | { | |
3801 | fputs ("With: ", dump_file); | |
ef6cb4c7 | 3802 | print_gimple_stmt (dump_file, stmt_to_print, 0); |
88ca9ea1 BS |
3803 | fputs ("\n", dump_file); |
3804 | } | |
3805 | } | |
3806 | ||
3807 | /* For each candidate in the tree rooted at C, replace it with | |
3808 | an increment if such has been shown to be profitable. */ | |
3809 | ||
3810 | static void | |
3811 | replace_profitable_candidates (slsr_cand_t c) | |
3812 | { | |
3813 | if (!cand_already_replaced (c)) | |
3814 | { | |
807e902e | 3815 | widest_int increment = cand_abs_increment (c); |
88ca9ea1 | 3816 | enum tree_code orig_code = gimple_assign_rhs_code (c->cand_stmt); |
7bf55a70 | 3817 | int i; |
88ca9ea1 BS |
3818 | |
3819 | i = incr_vec_index (increment); | |
3820 | ||
3821 | /* Only process profitable increments. Nothing useful can be done | |
3822 | to a cast or copy. */ | |
7bf55a70 BS |
3823 | if (i >= 0 |
3824 | && profitable_increment_p (i) | |
fbcdc43e | 3825 | && orig_code != SSA_NAME |
d822570f | 3826 | && !CONVERT_EXPR_CODE_P (orig_code)) |
88ca9ea1 | 3827 | { |
9b92d12b BS |
3828 | if (phi_dependent_cand_p (c)) |
3829 | { | |
c34923c4 | 3830 | gphi *phi = as_a <gphi *> (lookup_cand (c->def_phi)->cand_stmt); |
9b92d12b BS |
3831 | |
3832 | if (all_phi_incrs_profitable (c, phi)) | |
3833 | { | |
3834 | /* Look up the LHS SSA name from C's basis. This will be | |
3835 | the RHS1 of the adds we will introduce to create new | |
3836 | phi arguments. */ | |
3837 | slsr_cand_t basis = lookup_cand (c->basis); | |
3838 | tree basis_name = gimple_assign_lhs (basis->cand_stmt); | |
3839 | ||
3840 | /* Create a new phi statement that will represent C's true | |
3841 | basis after the transformation is complete. */ | |
3842 | location_t loc = gimple_location (c->cand_stmt); | |
3843 | tree name = create_phi_basis (c, phi, basis_name, | |
3844 | loc, UNKNOWN_STRIDE); | |
3845 | ||
3846 | /* Replace C with an add of the new basis phi and the | |
3847 | increment. */ | |
a7a7d10e | 3848 | replace_one_candidate (c, i, name); |
9b92d12b BS |
3849 | } |
3850 | } | |
3851 | else | |
3852 | { | |
3853 | slsr_cand_t basis = lookup_cand (c->basis); | |
3854 | tree basis_name = gimple_assign_lhs (basis->cand_stmt); | |
a7a7d10e | 3855 | replace_one_candidate (c, i, basis_name); |
9b92d12b | 3856 | } |
88ca9ea1 BS |
3857 | } |
3858 | } | |
3859 | ||
3860 | if (c->sibling) | |
3861 | replace_profitable_candidates (lookup_cand (c->sibling)); | |
3862 | ||
3863 | if (c->dependent) | |
3864 | replace_profitable_candidates (lookup_cand (c->dependent)); | |
3865 | } | |
3866 | \f | |
f9453c07 BS |
3867 | /* Analyze costs of related candidates in the candidate vector, |
3868 | and make beneficial replacements. */ | |
3869 | ||
3870 | static void | |
3871 | analyze_candidates_and_replace (void) | |
3872 | { | |
3873 | unsigned i; | |
3874 | slsr_cand_t c; | |
3875 | ||
3876 | /* Each candidate that has a null basis and a non-null | |
3877 | dependent is the root of a tree of related statements. | |
3878 | Analyze each tree to determine a subset of those | |
98aaa1a6 BC |
3879 | statements that can be replaced with maximum benefit. |
3880 | ||
3881 | Note the first NULL element is skipped. */ | |
3882 | FOR_EACH_VEC_ELT_FROM (cand_vec, i, c, 1) | |
f9453c07 BS |
3883 | { |
3884 | slsr_cand_t first_dep; | |
3885 | ||
3886 | if (c->basis != 0 || c->dependent == 0) | |
3887 | continue; | |
3888 | ||
3889 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3890 | fprintf (dump_file, "\nProcessing dependency tree rooted at %d.\n", | |
3891 | c->cand_num); | |
3892 | ||
3893 | first_dep = lookup_cand (c->dependent); | |
3894 | ||
2749c8f6 BS |
3895 | /* If this is a chain of CAND_REFs, unconditionally replace |
3896 | each of them with a strength-reduced data reference. */ | |
3897 | if (c->kind == CAND_REF) | |
3898 | replace_refs (c); | |
3899 | ||
9b92d12b BS |
3900 | /* If the common stride of all related candidates is a known |
3901 | constant, each candidate without a phi-dependence can be | |
3902 | profitably replaced. Each replaces a multiply by a single | |
3903 | add, with the possibility that a feeding add also goes dead. | |
3904 | A candidate with a phi-dependence is replaced only if the | |
3905 | compensation code it requires is offset by the strength | |
3906 | reduction savings. */ | |
3907 | else if (TREE_CODE (c->stride) == INTEGER_CST) | |
3908 | replace_uncond_cands_and_profitable_phis (first_dep); | |
f9453c07 | 3909 | |
88ca9ea1 BS |
3910 | /* When the stride is an SSA name, it may still be profitable |
3911 | to replace some or all of the dependent candidates, depending | |
3912 | on whether the introduced increments can be reused, or are | |
3913 | less expensive to calculate than the replaced statements. */ | |
3914 | else | |
3915 | { | |
ef4bddc2 | 3916 | machine_mode mode; |
88ca9ea1 BS |
3917 | bool speed; |
3918 | ||
3919 | /* Determine whether we'll be generating pointer arithmetic | |
3920 | when replacing candidates. */ | |
3921 | address_arithmetic_p = (c->kind == CAND_ADD | |
99cababb | 3922 | && POINTER_TYPE_P (c->cand_type)); |
88ca9ea1 BS |
3923 | |
3924 | /* If all candidates have already been replaced under other | |
3925 | interpretations, nothing remains to be done. */ | |
4b847da9 | 3926 | if (!count_candidates (c)) |
88ca9ea1 BS |
3927 | continue; |
3928 | ||
3929 | /* Construct an array of increments for this candidate chain. */ | |
4b847da9 | 3930 | incr_vec = XNEWVEC (incr_info, MAX_INCR_VEC_LEN); |
88ca9ea1 BS |
3931 | incr_vec_len = 0; |
3932 | record_increments (c); | |
3933 | ||
3934 | /* Determine which increments are profitable to replace. */ | |
3935 | mode = TYPE_MODE (TREE_TYPE (gimple_assign_lhs (c->cand_stmt))); | |
3936 | speed = optimize_cands_for_speed_p (c); | |
3937 | analyze_increments (first_dep, mode, speed); | |
3938 | ||
3939 | /* Insert initializers of the form T_0 = stride * increment | |
3940 | for use in profitable replacements. */ | |
3941 | insert_initializers (first_dep); | |
3942 | dump_incr_vec (); | |
3943 | ||
3944 | /* Perform the replacements. */ | |
3945 | replace_profitable_candidates (first_dep); | |
3946 | free (incr_vec); | |
3947 | } | |
f9453c07 | 3948 | } |
b115e803 BS |
3949 | |
3950 | /* For conditional candidates, we may have uncommitted insertions | |
3951 | on edges to clean up. */ | |
3952 | gsi_commit_edge_inserts (); | |
f9453c07 BS |
3953 | } |
3954 | ||
17795822 TS |
3955 | namespace { |
3956 | ||
3957 | const pass_data pass_data_strength_reduction = | |
be55bfe6 TS |
3958 | { |
3959 | GIMPLE_PASS, /* type */ | |
3960 | "slsr", /* name */ | |
3961 | OPTGROUP_NONE, /* optinfo_flags */ | |
be55bfe6 TS |
3962 | TV_GIMPLE_SLSR, /* tv_id */ |
3963 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
3964 | 0, /* properties_provided */ | |
3965 | 0, /* properties_destroyed */ | |
3966 | 0, /* todo_flags_start */ | |
3bea341f | 3967 | 0, /* todo_flags_finish */ |
be55bfe6 TS |
3968 | }; |
3969 | ||
17795822 | 3970 | class pass_strength_reduction : public gimple_opt_pass |
be55bfe6 TS |
3971 | { |
3972 | public: | |
3973 | pass_strength_reduction (gcc::context *ctxt) | |
3974 | : gimple_opt_pass (pass_data_strength_reduction, ctxt) | |
3975 | {} | |
3976 | ||
3977 | /* opt_pass methods: */ | |
3978 | virtual bool gate (function *) { return flag_tree_slsr; } | |
3979 | virtual unsigned int execute (function *); | |
3980 | ||
3981 | }; // class pass_strength_reduction | |
3982 | ||
3983 | unsigned | |
3984 | pass_strength_reduction::execute (function *fun) | |
f9453c07 | 3985 | { |
f9453c07 BS |
3986 | /* Create the obstack where candidates will reside. */ |
3987 | gcc_obstack_init (&cand_obstack); | |
3988 | ||
98aaa1a6 | 3989 | /* Allocate the candidate vector and initialize the first NULL element. */ |
9771b263 | 3990 | cand_vec.create (128); |
98aaa1a6 | 3991 | cand_vec.safe_push (NULL); |
f9453c07 BS |
3992 | |
3993 | /* Allocate the mapping from statements to candidate indices. */ | |
355fe088 | 3994 | stmt_cand_map = new hash_map<gimple *, slsr_cand_t>; |
f9453c07 BS |
3995 | |
3996 | /* Create the obstack where candidate chains will reside. */ | |
3997 | gcc_obstack_init (&chain_obstack); | |
3998 | ||
3cfd4469 | 3999 | /* Allocate the mapping from base expressions to candidate chains. */ |
c203e8a7 | 4000 | base_cand_map = new hash_table<cand_chain_hasher> (500); |
f9453c07 | 4001 | |
96d75a2c | 4002 | /* Allocate the mapping from bases to alternative bases. */ |
b787e7a2 | 4003 | alt_base_map = new hash_map<tree, tree>; |
96d75a2c | 4004 | |
f9453c07 BS |
4005 | /* Initialize the loop optimizer. We need to detect flow across |
4006 | back edges, and this gives us dominator information as well. */ | |
4007 | loop_optimizer_init (AVOID_CFG_MODIFICATIONS); | |
4008 | ||
f9453c07 BS |
4009 | /* Walk the CFG in predominator order looking for strength reduction |
4010 | candidates. */ | |
4d9192b5 | 4011 | find_candidates_dom_walker (CDI_DOMINATORS) |
be55bfe6 | 4012 | .walk (fun->cfg->x_entry_block_ptr); |
f9453c07 BS |
4013 | |
4014 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4015 | { | |
4016 | dump_cand_vec (); | |
4017 | dump_cand_chains (); | |
4018 | } | |
4019 | ||
b787e7a2 | 4020 | delete alt_base_map; |
96d75a2c YZ |
4021 | free_affine_expand_cache (&name_expansions); |
4022 | ||
f9453c07 BS |
4023 | /* Analyze costs and make appropriate replacements. */ |
4024 | analyze_candidates_and_replace (); | |
4025 | ||
f9453c07 | 4026 | loop_optimizer_finalize (); |
c203e8a7 TS |
4027 | delete base_cand_map; |
4028 | base_cand_map = NULL; | |
f9453c07 | 4029 | obstack_free (&chain_obstack, NULL); |
b787e7a2 | 4030 | delete stmt_cand_map; |
9771b263 | 4031 | cand_vec.release (); |
f9453c07 | 4032 | obstack_free (&cand_obstack, NULL); |
f9453c07 BS |
4033 | |
4034 | return 0; | |
4035 | } | |
4036 | ||
17795822 TS |
4037 | } // anon namespace |
4038 | ||
27a4cd48 DM |
4039 | gimple_opt_pass * |
4040 | make_pass_strength_reduction (gcc::context *ctxt) | |
4041 | { | |
4042 | return new pass_strength_reduction (ctxt); | |
4043 | } |