]>
Commit | Line | Data |
---|---|---|
6c2a63a3 | 1 | /* Global, SSA-based optimizations using mathematical identities. |
23a5b65a | 2 | Copyright (C) 2005-2014 Free Software Foundation, Inc. |
b8698a0f | 3 | |
6c2a63a3 | 4 | This file is part of GCC. |
b8698a0f | 5 | |
6c2a63a3 PB |
6 | GCC is free software; you can redistribute it and/or modify it |
7 | under the terms of the GNU General Public License as published by the | |
9dcd6f09 | 8 | Free Software Foundation; either version 3, or (at your option) any |
6c2a63a3 | 9 | later version. |
b8698a0f | 10 | |
6c2a63a3 PB |
11 | GCC is distributed in the hope that it will be useful, but WITHOUT |
12 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
14 | for more details. | |
b8698a0f | 15 | |
6c2a63a3 | 16 | You should have received a copy of the GNU General Public License |
9dcd6f09 NC |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
6c2a63a3 PB |
19 | |
20 | /* Currently, the only mini-pass in this file tries to CSE reciprocal | |
21 | operations. These are common in sequences such as this one: | |
22 | ||
23 | modulus = sqrt(x*x + y*y + z*z); | |
24 | x = x / modulus; | |
25 | y = y / modulus; | |
26 | z = z / modulus; | |
27 | ||
28 | that can be optimized to | |
29 | ||
30 | modulus = sqrt(x*x + y*y + z*z); | |
31 | rmodulus = 1.0 / modulus; | |
32 | x = x * rmodulus; | |
33 | y = y * rmodulus; | |
34 | z = z * rmodulus; | |
35 | ||
36 | We do this for loop invariant divisors, and with this pass whenever | |
bc23502b PB |
37 | we notice that a division has the same divisor multiple times. |
38 | ||
39 | Of course, like in PRE, we don't insert a division if a dominator | |
40 | already has one. However, this cannot be done as an extension of | |
41 | PRE for several reasons. | |
42 | ||
43 | First of all, with some experiments it was found out that the | |
44 | transformation is not always useful if there are only two divisions | |
45 | hy the same divisor. This is probably because modern processors | |
46 | can pipeline the divisions; on older, in-order processors it should | |
47 | still be effective to optimize two divisions by the same number. | |
48 | We make this a param, and it shall be called N in the remainder of | |
49 | this comment. | |
50 | ||
51 | Second, if trapping math is active, we have less freedom on where | |
52 | to insert divisions: we can only do so in basic blocks that already | |
53 | contain one. (If divisions don't trap, instead, we can insert | |
54 | divisions elsewhere, which will be in blocks that are common dominators | |
55 | of those that have the division). | |
56 | ||
57 | We really don't want to compute the reciprocal unless a division will | |
58 | be found. To do this, we won't insert the division in a basic block | |
59 | that has less than N divisions *post-dominating* it. | |
60 | ||
61 | The algorithm constructs a subset of the dominator tree, holding the | |
62 | blocks containing the divisions and the common dominators to them, | |
63 | and walk it twice. The first walk is in post-order, and it annotates | |
64 | each block with the number of divisions that post-dominate it: this | |
65 | gives information on where divisions can be inserted profitably. | |
66 | The second walk is in pre-order, and it inserts divisions as explained | |
67 | above, and replaces divisions by multiplications. | |
68 | ||
69 | In the best case, the cost of the pass is O(n_statements). In the | |
70 | worst-case, the cost is due to creating the dominator tree subset, | |
71 | with a cost of O(n_basic_blocks ^ 2); however this can only happen | |
72 | for n_statements / n_basic_blocks statements. So, the amortized cost | |
73 | of creating the dominator tree subset is O(n_basic_blocks) and the | |
74 | worst-case cost of the pass is O(n_statements * n_basic_blocks). | |
75 | ||
76 | More practically, the cost will be small because there are few | |
77 | divisions, and they tend to be in the same basic block, so insert_bb | |
78 | is called very few times. | |
79 | ||
80 | If we did this using domwalk.c, an efficient implementation would have | |
81 | to work on all the variables in a single pass, because we could not | |
82 | work on just a subset of the dominator tree, as we do now, and the | |
83 | cost would also be something like O(n_statements * n_basic_blocks). | |
84 | The data structures would be more complex in order to work on all the | |
85 | variables in a single pass. */ | |
6c2a63a3 PB |
86 | |
87 | #include "config.h" | |
88 | #include "system.h" | |
89 | #include "coretypes.h" | |
90 | #include "tm.h" | |
91 | #include "flags.h" | |
92 | #include "tree.h" | |
60393bbc AM |
93 | #include "predict.h" |
94 | #include "vec.h" | |
95 | #include "hashtab.h" | |
96 | #include "hash-set.h" | |
97 | #include "machmode.h" | |
98 | #include "hard-reg-set.h" | |
99 | #include "input.h" | |
100 | #include "function.h" | |
101 | #include "dominance.h" | |
102 | #include "cfg.h" | |
2fb9a547 AM |
103 | #include "basic-block.h" |
104 | #include "tree-ssa-alias.h" | |
105 | #include "internal-fn.h" | |
106 | #include "gimple-fold.h" | |
107 | #include "gimple-expr.h" | |
108 | #include "is-a.h" | |
18f429e2 | 109 | #include "gimple.h" |
5be5c238 | 110 | #include "gimple-iterator.h" |
73984f84 | 111 | #include "gimplify.h" |
18f429e2 | 112 | #include "gimplify-me.h" |
d8a2d370 | 113 | #include "stor-layout.h" |
442b4905 AM |
114 | #include "gimple-ssa.h" |
115 | #include "tree-cfg.h" | |
116 | #include "tree-phinodes.h" | |
117 | #include "ssa-iterators.h" | |
d8a2d370 | 118 | #include "stringpool.h" |
442b4905 | 119 | #include "tree-ssanames.h" |
d8a2d370 | 120 | #include "expr.h" |
442b4905 | 121 | #include "tree-dfa.h" |
7a300452 | 122 | #include "tree-ssa.h" |
6c2a63a3 | 123 | #include "tree-pass.h" |
bc23502b | 124 | #include "alloc-pool.h" |
bc23502b | 125 | #include "target.h" |
cf835838 | 126 | #include "gimple-pretty-print.h" |
9b2b7279 | 127 | #include "builtins.h" |
40013784 SB |
128 | |
129 | /* FIXME: RTL headers have to be included here for optabs. */ | |
130 | #include "rtl.h" /* Because optabs.h wants enum rtx_code. */ | |
131 | #include "expr.h" /* Because optabs.h wants sepops. */ | |
b0710fe1 | 132 | #include "insn-codes.h" |
03bd2f1a | 133 | #include "optabs.h" |
bc23502b PB |
134 | |
135 | /* This structure represents one basic block that either computes a | |
136 | division, or is a common dominator for basic block that compute a | |
137 | division. */ | |
138 | struct occurrence { | |
139 | /* The basic block represented by this structure. */ | |
140 | basic_block bb; | |
141 | ||
142 | /* If non-NULL, the SSA_NAME holding the definition for a reciprocal | |
143 | inserted in BB. */ | |
144 | tree recip_def; | |
145 | ||
726a989a | 146 | /* If non-NULL, the GIMPLE_ASSIGN for a reciprocal computation that |
bc23502b | 147 | was inserted in BB. */ |
726a989a | 148 | gimple recip_def_stmt; |
bc23502b PB |
149 | |
150 | /* Pointer to a list of "struct occurrence"s for blocks dominated | |
151 | by BB. */ | |
152 | struct occurrence *children; | |
153 | ||
154 | /* Pointer to the next "struct occurrence"s in the list of blocks | |
155 | sharing a common dominator. */ | |
156 | struct occurrence *next; | |
157 | ||
158 | /* The number of divisions that are in BB before compute_merit. The | |
159 | number of divisions that are in BB or post-dominate it after | |
160 | compute_merit. */ | |
161 | int num_divisions; | |
162 | ||
163 | /* True if the basic block has a division, false if it is a common | |
164 | dominator for basic blocks that do. If it is false and trapping | |
165 | math is active, BB is not a candidate for inserting a reciprocal. */ | |
166 | bool bb_has_division; | |
167 | }; | |
168 | ||
4da3b811 NF |
169 | static struct |
170 | { | |
171 | /* Number of 1.0/X ops inserted. */ | |
172 | int rdivs_inserted; | |
173 | ||
174 | /* Number of 1.0/FUNC ops inserted. */ | |
175 | int rfuncs_inserted; | |
176 | } reciprocal_stats; | |
177 | ||
178 | static struct | |
179 | { | |
180 | /* Number of cexpi calls inserted. */ | |
181 | int inserted; | |
182 | } sincos_stats; | |
183 | ||
184 | static struct | |
185 | { | |
73984f84 | 186 | /* Number of hand-written 16-bit nop / bswaps found. */ |
1df855ce CL |
187 | int found_16bit; |
188 | ||
73984f84 | 189 | /* Number of hand-written 32-bit nop / bswaps found. */ |
4da3b811 NF |
190 | int found_32bit; |
191 | ||
73984f84 | 192 | /* Number of hand-written 64-bit nop / bswaps found. */ |
4da3b811 | 193 | int found_64bit; |
73984f84 | 194 | } nop_stats, bswap_stats; |
4da3b811 NF |
195 | |
196 | static struct | |
197 | { | |
198 | /* Number of widening multiplication ops inserted. */ | |
199 | int widen_mults_inserted; | |
200 | ||
201 | /* Number of integer multiply-and-accumulate ops inserted. */ | |
202 | int maccs_inserted; | |
203 | ||
204 | /* Number of fp fused multiply-add ops inserted. */ | |
205 | int fmas_inserted; | |
206 | } widen_mul_stats; | |
bc23502b PB |
207 | |
208 | /* The instance of "struct occurrence" representing the highest | |
209 | interesting block in the dominator tree. */ | |
210 | static struct occurrence *occ_head; | |
211 | ||
212 | /* Allocation pool for getting instances of "struct occurrence". */ | |
213 | static alloc_pool occ_pool; | |
214 | ||
215 | ||
216 | ||
217 | /* Allocate and return a new struct occurrence for basic block BB, and | |
218 | whose children list is headed by CHILDREN. */ | |
219 | static struct occurrence * | |
220 | occ_new (basic_block bb, struct occurrence *children) | |
6c2a63a3 | 221 | { |
bc23502b PB |
222 | struct occurrence *occ; |
223 | ||
c22940cd | 224 | bb->aux = occ = (struct occurrence *) pool_alloc (occ_pool); |
bc23502b PB |
225 | memset (occ, 0, sizeof (struct occurrence)); |
226 | ||
227 | occ->bb = bb; | |
228 | occ->children = children; | |
229 | return occ; | |
6c2a63a3 PB |
230 | } |
231 | ||
bc23502b PB |
232 | |
233 | /* Insert NEW_OCC into our subset of the dominator tree. P_HEAD points to a | |
234 | list of "struct occurrence"s, one per basic block, having IDOM as | |
235 | their common dominator. | |
236 | ||
237 | We try to insert NEW_OCC as deep as possible in the tree, and we also | |
238 | insert any other block that is a common dominator for BB and one | |
239 | block already in the tree. */ | |
240 | ||
241 | static void | |
242 | insert_bb (struct occurrence *new_occ, basic_block idom, | |
243 | struct occurrence **p_head) | |
2ef571e2 | 244 | { |
bc23502b | 245 | struct occurrence *occ, **p_occ; |
2ef571e2 | 246 | |
bc23502b PB |
247 | for (p_occ = p_head; (occ = *p_occ) != NULL; ) |
248 | { | |
249 | basic_block bb = new_occ->bb, occ_bb = occ->bb; | |
250 | basic_block dom = nearest_common_dominator (CDI_DOMINATORS, occ_bb, bb); | |
251 | if (dom == bb) | |
252 | { | |
253 | /* BB dominates OCC_BB. OCC becomes NEW_OCC's child: remove OCC | |
254 | from its list. */ | |
255 | *p_occ = occ->next; | |
256 | occ->next = new_occ->children; | |
257 | new_occ->children = occ; | |
258 | ||
259 | /* Try the next block (it may as well be dominated by BB). */ | |
260 | } | |
261 | ||
262 | else if (dom == occ_bb) | |
263 | { | |
264 | /* OCC_BB dominates BB. Tail recurse to look deeper. */ | |
265 | insert_bb (new_occ, dom, &occ->children); | |
266 | return; | |
267 | } | |
268 | ||
269 | else if (dom != idom) | |
270 | { | |
271 | gcc_assert (!dom->aux); | |
272 | ||
273 | /* There is a dominator between IDOM and BB, add it and make | |
274 | two children out of NEW_OCC and OCC. First, remove OCC from | |
275 | its list. */ | |
276 | *p_occ = occ->next; | |
277 | new_occ->next = occ; | |
278 | occ->next = NULL; | |
279 | ||
280 | /* None of the previous blocks has DOM as a dominator: if we tail | |
281 | recursed, we would reexamine them uselessly. Just switch BB with | |
282 | DOM, and go on looking for blocks dominated by DOM. */ | |
283 | new_occ = occ_new (dom, new_occ); | |
284 | } | |
285 | ||
286 | else | |
287 | { | |
288 | /* Nothing special, go on with the next element. */ | |
289 | p_occ = &occ->next; | |
290 | } | |
291 | } | |
292 | ||
293 | /* No place was found as a child of IDOM. Make BB a sibling of IDOM. */ | |
294 | new_occ->next = *p_head; | |
295 | *p_head = new_occ; | |
296 | } | |
297 | ||
298 | /* Register that we found a division in BB. */ | |
299 | ||
300 | static inline void | |
301 | register_division_in (basic_block bb) | |
302 | { | |
303 | struct occurrence *occ; | |
304 | ||
305 | occ = (struct occurrence *) bb->aux; | |
306 | if (!occ) | |
307 | { | |
308 | occ = occ_new (bb, NULL); | |
fefa31b5 | 309 | insert_bb (occ, ENTRY_BLOCK_PTR_FOR_FN (cfun), &occ_head); |
bc23502b PB |
310 | } |
311 | ||
312 | occ->bb_has_division = true; | |
313 | occ->num_divisions++; | |
314 | } | |
315 | ||
316 | ||
317 | /* Compute the number of divisions that postdominate each block in OCC and | |
318 | its children. */ | |
6c2a63a3 | 319 | |
6c2a63a3 | 320 | static void |
bc23502b | 321 | compute_merit (struct occurrence *occ) |
6c2a63a3 | 322 | { |
bc23502b PB |
323 | struct occurrence *occ_child; |
324 | basic_block dom = occ->bb; | |
6c2a63a3 | 325 | |
bc23502b | 326 | for (occ_child = occ->children; occ_child; occ_child = occ_child->next) |
6c2a63a3 | 327 | { |
bc23502b PB |
328 | basic_block bb; |
329 | if (occ_child->children) | |
330 | compute_merit (occ_child); | |
331 | ||
332 | if (flag_exceptions) | |
333 | bb = single_noncomplex_succ (dom); | |
334 | else | |
335 | bb = dom; | |
336 | ||
337 | if (dominated_by_p (CDI_POST_DOMINATORS, bb, occ_child->bb)) | |
338 | occ->num_divisions += occ_child->num_divisions; | |
339 | } | |
340 | } | |
341 | ||
342 | ||
343 | /* Return whether USE_STMT is a floating-point division by DEF. */ | |
344 | static inline bool | |
726a989a | 345 | is_division_by (gimple use_stmt, tree def) |
bc23502b | 346 | { |
726a989a RB |
347 | return is_gimple_assign (use_stmt) |
348 | && gimple_assign_rhs_code (use_stmt) == RDIV_EXPR | |
349 | && gimple_assign_rhs2 (use_stmt) == def | |
8a5b57cd RG |
350 | /* Do not recognize x / x as valid division, as we are getting |
351 | confused later by replacing all immediate uses x in such | |
352 | a stmt. */ | |
726a989a | 353 | && gimple_assign_rhs1 (use_stmt) != def; |
bc23502b PB |
354 | } |
355 | ||
356 | /* Walk the subset of the dominator tree rooted at OCC, setting the | |
357 | RECIP_DEF field to a definition of 1.0 / DEF that can be used in | |
358 | the given basic block. The field may be left NULL, of course, | |
359 | if it is not possible or profitable to do the optimization. | |
360 | ||
361 | DEF_BSI is an iterator pointing at the statement defining DEF. | |
362 | If RECIP_DEF is set, a dominator already has a computation that can | |
363 | be used. */ | |
364 | ||
365 | static void | |
726a989a | 366 | insert_reciprocals (gimple_stmt_iterator *def_gsi, struct occurrence *occ, |
bc23502b PB |
367 | tree def, tree recip_def, int threshold) |
368 | { | |
726a989a RB |
369 | tree type; |
370 | gimple new_stmt; | |
371 | gimple_stmt_iterator gsi; | |
bc23502b PB |
372 | struct occurrence *occ_child; |
373 | ||
374 | if (!recip_def | |
375 | && (occ->bb_has_division || !flag_trapping_math) | |
376 | && occ->num_divisions >= threshold) | |
377 | { | |
378 | /* Make a variable with the replacement and substitute it. */ | |
379 | type = TREE_TYPE (def); | |
7cc434a3 | 380 | recip_def = create_tmp_reg (type, "reciptmp"); |
726a989a RB |
381 | new_stmt = gimple_build_assign_with_ops (RDIV_EXPR, recip_def, |
382 | build_one_cst (type), def); | |
b8698a0f | 383 | |
bc23502b PB |
384 | if (occ->bb_has_division) |
385 | { | |
386 | /* Case 1: insert before an existing division. */ | |
726a989a RB |
387 | gsi = gsi_after_labels (occ->bb); |
388 | while (!gsi_end_p (gsi) && !is_division_by (gsi_stmt (gsi), def)) | |
389 | gsi_next (&gsi); | |
bc23502b | 390 | |
726a989a | 391 | gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT); |
bc23502b | 392 | } |
726a989a | 393 | else if (def_gsi && occ->bb == def_gsi->bb) |
ac264fef | 394 | { |
bc23502b PB |
395 | /* Case 2: insert right after the definition. Note that this will |
396 | never happen if the definition statement can throw, because in | |
397 | that case the sole successor of the statement's basic block will | |
398 | dominate all the uses as well. */ | |
726a989a | 399 | gsi_insert_after (def_gsi, new_stmt, GSI_NEW_STMT); |
ac264fef | 400 | } |
bc23502b PB |
401 | else |
402 | { | |
403 | /* Case 3: insert in a basic block not containing defs/uses. */ | |
726a989a RB |
404 | gsi = gsi_after_labels (occ->bb); |
405 | gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT); | |
bc23502b PB |
406 | } |
407 | ||
4da3b811 NF |
408 | reciprocal_stats.rdivs_inserted++; |
409 | ||
bc23502b | 410 | occ->recip_def_stmt = new_stmt; |
6c2a63a3 PB |
411 | } |
412 | ||
bc23502b PB |
413 | occ->recip_def = recip_def; |
414 | for (occ_child = occ->children; occ_child; occ_child = occ_child->next) | |
726a989a | 415 | insert_reciprocals (def_gsi, occ_child, def, recip_def, threshold); |
bc23502b PB |
416 | } |
417 | ||
418 | ||
419 | /* Replace the division at USE_P with a multiplication by the reciprocal, if | |
420 | possible. */ | |
421 | ||
422 | static inline void | |
423 | replace_reciprocal (use_operand_p use_p) | |
424 | { | |
726a989a RB |
425 | gimple use_stmt = USE_STMT (use_p); |
426 | basic_block bb = gimple_bb (use_stmt); | |
bc23502b PB |
427 | struct occurrence *occ = (struct occurrence *) bb->aux; |
428 | ||
efd8f750 JH |
429 | if (optimize_bb_for_speed_p (bb) |
430 | && occ->recip_def && use_stmt != occ->recip_def_stmt) | |
bc23502b | 431 | { |
59401b92 | 432 | gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt); |
726a989a | 433 | gimple_assign_set_rhs_code (use_stmt, MULT_EXPR); |
bc23502b | 434 | SET_USE (use_p, occ->recip_def); |
59401b92 | 435 | fold_stmt_inplace (&gsi); |
bc23502b PB |
436 | update_stmt (use_stmt); |
437 | } | |
438 | } | |
439 | ||
440 | ||
441 | /* Free OCC and return one more "struct occurrence" to be freed. */ | |
442 | ||
443 | static struct occurrence * | |
444 | free_bb (struct occurrence *occ) | |
445 | { | |
446 | struct occurrence *child, *next; | |
447 | ||
448 | /* First get the two pointers hanging off OCC. */ | |
449 | next = occ->next; | |
450 | child = occ->children; | |
451 | occ->bb->aux = NULL; | |
452 | pool_free (occ_pool, occ); | |
453 | ||
454 | /* Now ensure that we don't recurse unless it is necessary. */ | |
455 | if (!child) | |
456 | return next; | |
2ef571e2 | 457 | else |
bc23502b PB |
458 | { |
459 | while (next) | |
460 | next = free_bb (next); | |
461 | ||
462 | return child; | |
463 | } | |
464 | } | |
465 | ||
466 | ||
467 | /* Look for floating-point divisions among DEF's uses, and try to | |
468 | replace them by multiplications with the reciprocal. Add | |
469 | as many statements computing the reciprocal as needed. | |
470 | ||
471 | DEF must be a GIMPLE register of a floating-point type. */ | |
472 | ||
473 | static void | |
726a989a | 474 | execute_cse_reciprocals_1 (gimple_stmt_iterator *def_gsi, tree def) |
bc23502b PB |
475 | { |
476 | use_operand_p use_p; | |
477 | imm_use_iterator use_iter; | |
478 | struct occurrence *occ; | |
479 | int count = 0, threshold; | |
6c2a63a3 | 480 | |
bc23502b PB |
481 | gcc_assert (FLOAT_TYPE_P (TREE_TYPE (def)) && is_gimple_reg (def)); |
482 | ||
483 | FOR_EACH_IMM_USE_FAST (use_p, use_iter, def) | |
6c2a63a3 | 484 | { |
726a989a | 485 | gimple use_stmt = USE_STMT (use_p); |
bc23502b | 486 | if (is_division_by (use_stmt, def)) |
6c2a63a3 | 487 | { |
726a989a | 488 | register_division_in (gimple_bb (use_stmt)); |
bc23502b | 489 | count++; |
6c2a63a3 PB |
490 | } |
491 | } | |
b8698a0f | 492 | |
bc23502b PB |
493 | /* Do the expensive part only if we can hope to optimize something. */ |
494 | threshold = targetm.min_divisions_for_recip_mul (TYPE_MODE (TREE_TYPE (def))); | |
495 | if (count >= threshold) | |
496 | { | |
726a989a | 497 | gimple use_stmt; |
bc23502b PB |
498 | for (occ = occ_head; occ; occ = occ->next) |
499 | { | |
500 | compute_merit (occ); | |
726a989a | 501 | insert_reciprocals (def_gsi, occ, def, NULL, threshold); |
bc23502b PB |
502 | } |
503 | ||
6c00f606 | 504 | FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, def) |
bc23502b | 505 | { |
bc23502b | 506 | if (is_division_by (use_stmt, def)) |
6c00f606 AM |
507 | { |
508 | FOR_EACH_IMM_USE_ON_STMT (use_p, use_iter) | |
509 | replace_reciprocal (use_p); | |
510 | } | |
bc23502b PB |
511 | } |
512 | } | |
513 | ||
514 | for (occ = occ_head; occ; ) | |
515 | occ = free_bb (occ); | |
516 | ||
517 | occ_head = NULL; | |
6c2a63a3 PB |
518 | } |
519 | ||
bc23502b PB |
520 | /* Go through all the floating-point SSA_NAMEs, and call |
521 | execute_cse_reciprocals_1 on each of them. */ | |
be55bfe6 TS |
522 | namespace { |
523 | ||
524 | const pass_data pass_data_cse_reciprocals = | |
525 | { | |
526 | GIMPLE_PASS, /* type */ | |
527 | "recip", /* name */ | |
528 | OPTGROUP_NONE, /* optinfo_flags */ | |
be55bfe6 TS |
529 | TV_NONE, /* tv_id */ |
530 | PROP_ssa, /* properties_required */ | |
531 | 0, /* properties_provided */ | |
532 | 0, /* properties_destroyed */ | |
533 | 0, /* todo_flags_start */ | |
3bea341f | 534 | TODO_update_ssa, /* todo_flags_finish */ |
be55bfe6 TS |
535 | }; |
536 | ||
537 | class pass_cse_reciprocals : public gimple_opt_pass | |
538 | { | |
539 | public: | |
540 | pass_cse_reciprocals (gcc::context *ctxt) | |
541 | : gimple_opt_pass (pass_data_cse_reciprocals, ctxt) | |
542 | {} | |
543 | ||
544 | /* opt_pass methods: */ | |
545 | virtual bool gate (function *) { return optimize && flag_reciprocal_math; } | |
546 | virtual unsigned int execute (function *); | |
547 | ||
548 | }; // class pass_cse_reciprocals | |
549 | ||
550 | unsigned int | |
551 | pass_cse_reciprocals::execute (function *fun) | |
6c2a63a3 PB |
552 | { |
553 | basic_block bb; | |
a8f82ec4 | 554 | tree arg; |
ac264fef | 555 | |
bc23502b PB |
556 | occ_pool = create_alloc_pool ("dominators for recip", |
557 | sizeof (struct occurrence), | |
be55bfe6 | 558 | n_basic_blocks_for_fn (fun) / 3 + 1); |
ac264fef | 559 | |
4da3b811 | 560 | memset (&reciprocal_stats, 0, sizeof (reciprocal_stats)); |
d898f3ce PB |
561 | calculate_dominance_info (CDI_DOMINATORS); |
562 | calculate_dominance_info (CDI_POST_DOMINATORS); | |
bc23502b PB |
563 | |
564 | #ifdef ENABLE_CHECKING | |
be55bfe6 | 565 | FOR_EACH_BB_FN (bb, fun) |
bc23502b PB |
566 | gcc_assert (!bb->aux); |
567 | #endif | |
568 | ||
be55bfe6 | 569 | for (arg = DECL_ARGUMENTS (fun->decl); arg; arg = DECL_CHAIN (arg)) |
32244553 | 570 | if (FLOAT_TYPE_P (TREE_TYPE (arg)) |
bc23502b | 571 | && is_gimple_reg (arg)) |
32244553 | 572 | { |
be55bfe6 | 573 | tree name = ssa_default_def (fun, arg); |
32244553 RG |
574 | if (name) |
575 | execute_cse_reciprocals_1 (NULL, name); | |
576 | } | |
a8f82ec4 | 577 | |
be55bfe6 | 578 | FOR_EACH_BB_FN (bb, fun) |
6c2a63a3 | 579 | { |
726a989a RB |
580 | gimple_stmt_iterator gsi; |
581 | gimple phi; | |
582 | tree def; | |
6c2a63a3 | 583 | |
726a989a | 584 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
6c2a63a3 | 585 | { |
726a989a | 586 | phi = gsi_stmt (gsi); |
6c2a63a3 | 587 | def = PHI_RESULT (phi); |
ea057359 RG |
588 | if (! virtual_operand_p (def) |
589 | && FLOAT_TYPE_P (TREE_TYPE (def))) | |
bc23502b | 590 | execute_cse_reciprocals_1 (NULL, def); |
6c2a63a3 PB |
591 | } |
592 | ||
726a989a | 593 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
6c2a63a3 | 594 | { |
726a989a | 595 | gimple stmt = gsi_stmt (gsi); |
2f397a93 | 596 | |
726a989a | 597 | if (gimple_has_lhs (stmt) |
6c2a63a3 PB |
598 | && (def = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_DEF)) != NULL |
599 | && FLOAT_TYPE_P (TREE_TYPE (def)) | |
a8f82ec4 | 600 | && TREE_CODE (def) == SSA_NAME) |
726a989a | 601 | execute_cse_reciprocals_1 (&gsi, def); |
6c2a63a3 | 602 | } |
6b889d89 | 603 | |
efd8f750 JH |
604 | if (optimize_bb_for_size_p (bb)) |
605 | continue; | |
606 | ||
6b889d89 | 607 | /* Scan for a/func(b) and convert it to reciprocal a*rfunc(b). */ |
726a989a | 608 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
6b889d89 | 609 | { |
726a989a | 610 | gimple stmt = gsi_stmt (gsi); |
6b889d89 UB |
611 | tree fndecl; |
612 | ||
726a989a RB |
613 | if (is_gimple_assign (stmt) |
614 | && gimple_assign_rhs_code (stmt) == RDIV_EXPR) | |
6b889d89 | 615 | { |
726a989a RB |
616 | tree arg1 = gimple_assign_rhs2 (stmt); |
617 | gimple stmt1; | |
ac10986f UB |
618 | |
619 | if (TREE_CODE (arg1) != SSA_NAME) | |
620 | continue; | |
621 | ||
622 | stmt1 = SSA_NAME_DEF_STMT (arg1); | |
6b889d89 | 623 | |
726a989a RB |
624 | if (is_gimple_call (stmt1) |
625 | && gimple_call_lhs (stmt1) | |
626 | && (fndecl = gimple_call_fndecl (stmt1)) | |
6b889d89 UB |
627 | && (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL |
628 | || DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_MD)) | |
629 | { | |
630 | enum built_in_function code; | |
79af7c1f MM |
631 | bool md_code, fail; |
632 | imm_use_iterator ui; | |
633 | use_operand_p use_p; | |
6b889d89 UB |
634 | |
635 | code = DECL_FUNCTION_CODE (fndecl); | |
ac10986f UB |
636 | md_code = DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_MD; |
637 | ||
638 | fndecl = targetm.builtin_reciprocal (code, md_code, false); | |
6b889d89 UB |
639 | if (!fndecl) |
640 | continue; | |
641 | ||
79af7c1f MM |
642 | /* Check that all uses of the SSA name are divisions, |
643 | otherwise replacing the defining statement will do | |
644 | the wrong thing. */ | |
645 | fail = false; | |
646 | FOR_EACH_IMM_USE_FAST (use_p, ui, arg1) | |
647 | { | |
648 | gimple stmt2 = USE_STMT (use_p); | |
649 | if (is_gimple_debug (stmt2)) | |
650 | continue; | |
651 | if (!is_gimple_assign (stmt2) | |
652 | || gimple_assign_rhs_code (stmt2) != RDIV_EXPR | |
653 | || gimple_assign_rhs1 (stmt2) == arg1 | |
654 | || gimple_assign_rhs2 (stmt2) != arg1) | |
655 | { | |
656 | fail = true; | |
657 | break; | |
658 | } | |
659 | } | |
660 | if (fail) | |
661 | continue; | |
662 | ||
ff2a63a7 | 663 | gimple_replace_ssa_lhs (stmt1, arg1); |
7c9577be | 664 | gimple_call_set_fndecl (stmt1, fndecl); |
6b889d89 | 665 | update_stmt (stmt1); |
4da3b811 | 666 | reciprocal_stats.rfuncs_inserted++; |
6b889d89 | 667 | |
79af7c1f MM |
668 | FOR_EACH_IMM_USE_STMT (stmt, ui, arg1) |
669 | { | |
59401b92 | 670 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); |
79af7c1f | 671 | gimple_assign_set_rhs_code (stmt, MULT_EXPR); |
59401b92 | 672 | fold_stmt_inplace (&gsi); |
79af7c1f MM |
673 | update_stmt (stmt); |
674 | } | |
6b889d89 UB |
675 | } |
676 | } | |
677 | } | |
6c2a63a3 | 678 | } |
ac264fef | 679 | |
be55bfe6 | 680 | statistics_counter_event (fun, "reciprocal divs inserted", |
4da3b811 | 681 | reciprocal_stats.rdivs_inserted); |
be55bfe6 | 682 | statistics_counter_event (fun, "reciprocal functions inserted", |
4da3b811 NF |
683 | reciprocal_stats.rfuncs_inserted); |
684 | ||
d898f3ce PB |
685 | free_dominance_info (CDI_DOMINATORS); |
686 | free_dominance_info (CDI_POST_DOMINATORS); | |
bc23502b | 687 | free_alloc_pool (occ_pool); |
c2924966 | 688 | return 0; |
6c2a63a3 PB |
689 | } |
690 | ||
27a4cd48 DM |
691 | } // anon namespace |
692 | ||
693 | gimple_opt_pass * | |
694 | make_pass_cse_reciprocals (gcc::context *ctxt) | |
695 | { | |
696 | return new pass_cse_reciprocals (ctxt); | |
697 | } | |
698 | ||
88512ba0 | 699 | /* Records an occurrence at statement USE_STMT in the vector of trees |
2f397a93 | 700 | STMTS if it is dominated by *TOP_BB or dominates it or this basic block |
88512ba0 | 701 | is not yet initialized. Returns true if the occurrence was pushed on |
2f397a93 RG |
702 | the vector. Adjusts *TOP_BB to be the basic block dominating all |
703 | statements in the vector. */ | |
704 | ||
705 | static bool | |
9771b263 | 706 | maybe_record_sincos (vec<gimple> *stmts, |
726a989a | 707 | basic_block *top_bb, gimple use_stmt) |
2f397a93 | 708 | { |
726a989a | 709 | basic_block use_bb = gimple_bb (use_stmt); |
2f397a93 RG |
710 | if (*top_bb |
711 | && (*top_bb == use_bb | |
712 | || dominated_by_p (CDI_DOMINATORS, use_bb, *top_bb))) | |
9771b263 | 713 | stmts->safe_push (use_stmt); |
2f397a93 RG |
714 | else if (!*top_bb |
715 | || dominated_by_p (CDI_DOMINATORS, *top_bb, use_bb)) | |
716 | { | |
9771b263 | 717 | stmts->safe_push (use_stmt); |
2f397a93 RG |
718 | *top_bb = use_bb; |
719 | } | |
720 | else | |
721 | return false; | |
722 | ||
723 | return true; | |
724 | } | |
725 | ||
726 | /* Look for sin, cos and cexpi calls with the same argument NAME and | |
727 | create a single call to cexpi CSEing the result in this case. | |
728 | We first walk over all immediate uses of the argument collecting | |
729 | statements that we can CSE in a vector and in a second pass replace | |
730 | the statement rhs with a REALPART or IMAGPART expression on the | |
731 | result of the cexpi call we insert before the use statement that | |
732 | dominates all other candidates. */ | |
733 | ||
90bc1cb8 | 734 | static bool |
2f397a93 RG |
735 | execute_cse_sincos_1 (tree name) |
736 | { | |
726a989a | 737 | gimple_stmt_iterator gsi; |
2f397a93 | 738 | imm_use_iterator use_iter; |
726a989a RB |
739 | tree fndecl, res, type; |
740 | gimple def_stmt, use_stmt, stmt; | |
2f397a93 | 741 | int seen_cos = 0, seen_sin = 0, seen_cexpi = 0; |
6e1aa848 | 742 | vec<gimple> stmts = vNULL; |
2f397a93 RG |
743 | basic_block top_bb = NULL; |
744 | int i; | |
90bc1cb8 | 745 | bool cfg_changed = false; |
2f397a93 RG |
746 | |
747 | type = TREE_TYPE (name); | |
748 | FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, name) | |
749 | { | |
726a989a RB |
750 | if (gimple_code (use_stmt) != GIMPLE_CALL |
751 | || !gimple_call_lhs (use_stmt) | |
752 | || !(fndecl = gimple_call_fndecl (use_stmt)) | |
2f397a93 RG |
753 | || DECL_BUILT_IN_CLASS (fndecl) != BUILT_IN_NORMAL) |
754 | continue; | |
755 | ||
756 | switch (DECL_FUNCTION_CODE (fndecl)) | |
757 | { | |
758 | CASE_FLT_FN (BUILT_IN_COS): | |
759 | seen_cos |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0; | |
760 | break; | |
761 | ||
762 | CASE_FLT_FN (BUILT_IN_SIN): | |
763 | seen_sin |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0; | |
764 | break; | |
765 | ||
766 | CASE_FLT_FN (BUILT_IN_CEXPI): | |
767 | seen_cexpi |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0; | |
768 | break; | |
769 | ||
770 | default:; | |
771 | } | |
772 | } | |
773 | ||
774 | if (seen_cos + seen_sin + seen_cexpi <= 1) | |
775 | { | |
9771b263 | 776 | stmts.release (); |
90bc1cb8 | 777 | return false; |
2f397a93 RG |
778 | } |
779 | ||
780 | /* Simply insert cexpi at the beginning of top_bb but not earlier than | |
781 | the name def statement. */ | |
782 | fndecl = mathfn_built_in (type, BUILT_IN_CEXPI); | |
783 | if (!fndecl) | |
90bc1cb8 | 784 | return false; |
726a989a | 785 | stmt = gimple_build_call (fndecl, 1, name); |
83d5977e | 786 | res = make_temp_ssa_name (TREE_TYPE (TREE_TYPE (fndecl)), stmt, "sincostmp"); |
726a989a RB |
787 | gimple_call_set_lhs (stmt, res); |
788 | ||
2f397a93 | 789 | def_stmt = SSA_NAME_DEF_STMT (name); |
59805c3b | 790 | if (!SSA_NAME_IS_DEFAULT_DEF (name) |
726a989a RB |
791 | && gimple_code (def_stmt) != GIMPLE_PHI |
792 | && gimple_bb (def_stmt) == top_bb) | |
2f397a93 | 793 | { |
726a989a RB |
794 | gsi = gsi_for_stmt (def_stmt); |
795 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); | |
2f397a93 RG |
796 | } |
797 | else | |
798 | { | |
726a989a RB |
799 | gsi = gsi_after_labels (top_bb); |
800 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
2f397a93 | 801 | } |
4da3b811 | 802 | sincos_stats.inserted++; |
2f397a93 RG |
803 | |
804 | /* And adjust the recorded old call sites. */ | |
9771b263 | 805 | for (i = 0; stmts.iterate (i, &use_stmt); ++i) |
2f397a93 | 806 | { |
726a989a RB |
807 | tree rhs = NULL; |
808 | fndecl = gimple_call_fndecl (use_stmt); | |
809 | ||
2f397a93 RG |
810 | switch (DECL_FUNCTION_CODE (fndecl)) |
811 | { | |
812 | CASE_FLT_FN (BUILT_IN_COS): | |
726a989a | 813 | rhs = fold_build1 (REALPART_EXPR, type, res); |
2f397a93 RG |
814 | break; |
815 | ||
816 | CASE_FLT_FN (BUILT_IN_SIN): | |
726a989a | 817 | rhs = fold_build1 (IMAGPART_EXPR, type, res); |
2f397a93 RG |
818 | break; |
819 | ||
820 | CASE_FLT_FN (BUILT_IN_CEXPI): | |
726a989a | 821 | rhs = res; |
2f397a93 RG |
822 | break; |
823 | ||
824 | default:; | |
825 | gcc_unreachable (); | |
826 | } | |
827 | ||
726a989a RB |
828 | /* Replace call with a copy. */ |
829 | stmt = gimple_build_assign (gimple_call_lhs (use_stmt), rhs); | |
830 | ||
831 | gsi = gsi_for_stmt (use_stmt); | |
90bc1cb8 RG |
832 | gsi_replace (&gsi, stmt, true); |
833 | if (gimple_purge_dead_eh_edges (gimple_bb (stmt))) | |
834 | cfg_changed = true; | |
2f397a93 RG |
835 | } |
836 | ||
9771b263 | 837 | stmts.release (); |
90bc1cb8 RG |
838 | |
839 | return cfg_changed; | |
2f397a93 RG |
840 | } |
841 | ||
78be79d5 BS |
842 | /* To evaluate powi(x,n), the floating point value x raised to the |
843 | constant integer exponent n, we use a hybrid algorithm that | |
844 | combines the "window method" with look-up tables. For an | |
845 | introduction to exponentiation algorithms and "addition chains", | |
846 | see section 4.6.3, "Evaluation of Powers" of Donald E. Knuth, | |
847 | "Seminumerical Algorithms", Vol. 2, "The Art of Computer Programming", | |
848 | 3rd Edition, 1998, and Daniel M. Gordon, "A Survey of Fast Exponentiation | |
849 | Methods", Journal of Algorithms, Vol. 27, pp. 129-146, 1998. */ | |
850 | ||
851 | /* Provide a default value for POWI_MAX_MULTS, the maximum number of | |
852 | multiplications to inline before calling the system library's pow | |
853 | function. powi(x,n) requires at worst 2*bits(n)-2 multiplications, | |
854 | so this default never requires calling pow, powf or powl. */ | |
855 | ||
856 | #ifndef POWI_MAX_MULTS | |
857 | #define POWI_MAX_MULTS (2*HOST_BITS_PER_WIDE_INT-2) | |
858 | #endif | |
859 | ||
860 | /* The size of the "optimal power tree" lookup table. All | |
861 | exponents less than this value are simply looked up in the | |
862 | powi_table below. This threshold is also used to size the | |
863 | cache of pseudo registers that hold intermediate results. */ | |
864 | #define POWI_TABLE_SIZE 256 | |
865 | ||
866 | /* The size, in bits of the window, used in the "window method" | |
867 | exponentiation algorithm. This is equivalent to a radix of | |
868 | (1<<POWI_WINDOW_SIZE) in the corresponding "m-ary method". */ | |
869 | #define POWI_WINDOW_SIZE 3 | |
870 | ||
871 | /* The following table is an efficient representation of an | |
872 | "optimal power tree". For each value, i, the corresponding | |
873 | value, j, in the table states than an optimal evaluation | |
874 | sequence for calculating pow(x,i) can be found by evaluating | |
875 | pow(x,j)*pow(x,i-j). An optimal power tree for the first | |
876 | 100 integers is given in Knuth's "Seminumerical algorithms". */ | |
877 | ||
878 | static const unsigned char powi_table[POWI_TABLE_SIZE] = | |
879 | { | |
880 | 0, 1, 1, 2, 2, 3, 3, 4, /* 0 - 7 */ | |
881 | 4, 6, 5, 6, 6, 10, 7, 9, /* 8 - 15 */ | |
882 | 8, 16, 9, 16, 10, 12, 11, 13, /* 16 - 23 */ | |
883 | 12, 17, 13, 18, 14, 24, 15, 26, /* 24 - 31 */ | |
884 | 16, 17, 17, 19, 18, 33, 19, 26, /* 32 - 39 */ | |
885 | 20, 25, 21, 40, 22, 27, 23, 44, /* 40 - 47 */ | |
886 | 24, 32, 25, 34, 26, 29, 27, 44, /* 48 - 55 */ | |
887 | 28, 31, 29, 34, 30, 60, 31, 36, /* 56 - 63 */ | |
888 | 32, 64, 33, 34, 34, 46, 35, 37, /* 64 - 71 */ | |
889 | 36, 65, 37, 50, 38, 48, 39, 69, /* 72 - 79 */ | |
890 | 40, 49, 41, 43, 42, 51, 43, 58, /* 80 - 87 */ | |
891 | 44, 64, 45, 47, 46, 59, 47, 76, /* 88 - 95 */ | |
892 | 48, 65, 49, 66, 50, 67, 51, 66, /* 96 - 103 */ | |
893 | 52, 70, 53, 74, 54, 104, 55, 74, /* 104 - 111 */ | |
894 | 56, 64, 57, 69, 58, 78, 59, 68, /* 112 - 119 */ | |
895 | 60, 61, 61, 80, 62, 75, 63, 68, /* 120 - 127 */ | |
896 | 64, 65, 65, 128, 66, 129, 67, 90, /* 128 - 135 */ | |
897 | 68, 73, 69, 131, 70, 94, 71, 88, /* 136 - 143 */ | |
898 | 72, 128, 73, 98, 74, 132, 75, 121, /* 144 - 151 */ | |
899 | 76, 102, 77, 124, 78, 132, 79, 106, /* 152 - 159 */ | |
900 | 80, 97, 81, 160, 82, 99, 83, 134, /* 160 - 167 */ | |
901 | 84, 86, 85, 95, 86, 160, 87, 100, /* 168 - 175 */ | |
902 | 88, 113, 89, 98, 90, 107, 91, 122, /* 176 - 183 */ | |
903 | 92, 111, 93, 102, 94, 126, 95, 150, /* 184 - 191 */ | |
904 | 96, 128, 97, 130, 98, 133, 99, 195, /* 192 - 199 */ | |
905 | 100, 128, 101, 123, 102, 164, 103, 138, /* 200 - 207 */ | |
906 | 104, 145, 105, 146, 106, 109, 107, 149, /* 208 - 215 */ | |
907 | 108, 200, 109, 146, 110, 170, 111, 157, /* 216 - 223 */ | |
908 | 112, 128, 113, 130, 114, 182, 115, 132, /* 224 - 231 */ | |
909 | 116, 200, 117, 132, 118, 158, 119, 206, /* 232 - 239 */ | |
910 | 120, 240, 121, 162, 122, 147, 123, 152, /* 240 - 247 */ | |
911 | 124, 166, 125, 214, 126, 138, 127, 153, /* 248 - 255 */ | |
912 | }; | |
913 | ||
914 | ||
915 | /* Return the number of multiplications required to calculate | |
916 | powi(x,n) where n is less than POWI_TABLE_SIZE. This is a | |
917 | subroutine of powi_cost. CACHE is an array indicating | |
918 | which exponents have already been calculated. */ | |
919 | ||
920 | static int | |
921 | powi_lookup_cost (unsigned HOST_WIDE_INT n, bool *cache) | |
922 | { | |
923 | /* If we've already calculated this exponent, then this evaluation | |
924 | doesn't require any additional multiplications. */ | |
925 | if (cache[n]) | |
926 | return 0; | |
927 | ||
928 | cache[n] = true; | |
929 | return powi_lookup_cost (n - powi_table[n], cache) | |
930 | + powi_lookup_cost (powi_table[n], cache) + 1; | |
931 | } | |
932 | ||
933 | /* Return the number of multiplications required to calculate | |
934 | powi(x,n) for an arbitrary x, given the exponent N. This | |
935 | function needs to be kept in sync with powi_as_mults below. */ | |
936 | ||
937 | static int | |
938 | powi_cost (HOST_WIDE_INT n) | |
939 | { | |
940 | bool cache[POWI_TABLE_SIZE]; | |
941 | unsigned HOST_WIDE_INT digit; | |
942 | unsigned HOST_WIDE_INT val; | |
943 | int result; | |
944 | ||
945 | if (n == 0) | |
946 | return 0; | |
947 | ||
948 | /* Ignore the reciprocal when calculating the cost. */ | |
949 | val = (n < 0) ? -n : n; | |
950 | ||
951 | /* Initialize the exponent cache. */ | |
952 | memset (cache, 0, POWI_TABLE_SIZE * sizeof (bool)); | |
953 | cache[1] = true; | |
954 | ||
955 | result = 0; | |
956 | ||
957 | while (val >= POWI_TABLE_SIZE) | |
958 | { | |
959 | if (val & 1) | |
960 | { | |
961 | digit = val & ((1 << POWI_WINDOW_SIZE) - 1); | |
962 | result += powi_lookup_cost (digit, cache) | |
963 | + POWI_WINDOW_SIZE + 1; | |
964 | val >>= POWI_WINDOW_SIZE; | |
965 | } | |
966 | else | |
967 | { | |
968 | val >>= 1; | |
969 | result++; | |
970 | } | |
971 | } | |
972 | ||
973 | return result + powi_lookup_cost (val, cache); | |
974 | } | |
975 | ||
976 | /* Recursive subroutine of powi_as_mults. This function takes the | |
977 | array, CACHE, of already calculated exponents and an exponent N and | |
978 | returns a tree that corresponds to CACHE[1]**N, with type TYPE. */ | |
979 | ||
980 | static tree | |
981 | powi_as_mults_1 (gimple_stmt_iterator *gsi, location_t loc, tree type, | |
83d5977e | 982 | HOST_WIDE_INT n, tree *cache) |
78be79d5 BS |
983 | { |
984 | tree op0, op1, ssa_target; | |
985 | unsigned HOST_WIDE_INT digit; | |
986 | gimple mult_stmt; | |
987 | ||
988 | if (n < POWI_TABLE_SIZE && cache[n]) | |
989 | return cache[n]; | |
990 | ||
83d5977e | 991 | ssa_target = make_temp_ssa_name (type, NULL, "powmult"); |
78be79d5 BS |
992 | |
993 | if (n < POWI_TABLE_SIZE) | |
994 | { | |
995 | cache[n] = ssa_target; | |
83d5977e RG |
996 | op0 = powi_as_mults_1 (gsi, loc, type, n - powi_table[n], cache); |
997 | op1 = powi_as_mults_1 (gsi, loc, type, powi_table[n], cache); | |
78be79d5 BS |
998 | } |
999 | else if (n & 1) | |
1000 | { | |
1001 | digit = n & ((1 << POWI_WINDOW_SIZE) - 1); | |
83d5977e RG |
1002 | op0 = powi_as_mults_1 (gsi, loc, type, n - digit, cache); |
1003 | op1 = powi_as_mults_1 (gsi, loc, type, digit, cache); | |
78be79d5 BS |
1004 | } |
1005 | else | |
1006 | { | |
83d5977e | 1007 | op0 = powi_as_mults_1 (gsi, loc, type, n >> 1, cache); |
78be79d5 BS |
1008 | op1 = op0; |
1009 | } | |
1010 | ||
1011 | mult_stmt = gimple_build_assign_with_ops (MULT_EXPR, ssa_target, op0, op1); | |
ba869341 | 1012 | gimple_set_location (mult_stmt, loc); |
78be79d5 BS |
1013 | gsi_insert_before (gsi, mult_stmt, GSI_SAME_STMT); |
1014 | ||
1015 | return ssa_target; | |
1016 | } | |
1017 | ||
1018 | /* Convert ARG0**N to a tree of multiplications of ARG0 with itself. | |
1019 | This function needs to be kept in sync with powi_cost above. */ | |
1020 | ||
1021 | static tree | |
1022 | powi_as_mults (gimple_stmt_iterator *gsi, location_t loc, | |
1023 | tree arg0, HOST_WIDE_INT n) | |
1024 | { | |
83d5977e | 1025 | tree cache[POWI_TABLE_SIZE], result, type = TREE_TYPE (arg0); |
78be79d5 | 1026 | gimple div_stmt; |
83d5977e | 1027 | tree target; |
78be79d5 BS |
1028 | |
1029 | if (n == 0) | |
1030 | return build_real (type, dconst1); | |
1031 | ||
1032 | memset (cache, 0, sizeof (cache)); | |
1033 | cache[1] = arg0; | |
1034 | ||
83d5977e | 1035 | result = powi_as_mults_1 (gsi, loc, type, (n < 0) ? -n : n, cache); |
78be79d5 BS |
1036 | if (n >= 0) |
1037 | return result; | |
1038 | ||
1039 | /* If the original exponent was negative, reciprocate the result. */ | |
83d5977e | 1040 | target = make_temp_ssa_name (type, NULL, "powmult"); |
78be79d5 BS |
1041 | div_stmt = gimple_build_assign_with_ops (RDIV_EXPR, target, |
1042 | build_real (type, dconst1), | |
1043 | result); | |
ba869341 | 1044 | gimple_set_location (div_stmt, loc); |
78be79d5 BS |
1045 | gsi_insert_before (gsi, div_stmt, GSI_SAME_STMT); |
1046 | ||
1047 | return target; | |
1048 | } | |
1049 | ||
1050 | /* ARG0 and N are the two arguments to a powi builtin in GSI with | |
1051 | location info LOC. If the arguments are appropriate, create an | |
1052 | equivalent sequence of statements prior to GSI using an optimal | |
1053 | number of multiplications, and return an expession holding the | |
1054 | result. */ | |
1055 | ||
1056 | static tree | |
1057 | gimple_expand_builtin_powi (gimple_stmt_iterator *gsi, location_t loc, | |
1058 | tree arg0, HOST_WIDE_INT n) | |
1059 | { | |
1060 | /* Avoid largest negative number. */ | |
1061 | if (n != -n | |
1062 | && ((n >= -1 && n <= 2) | |
1063 | || (optimize_function_for_speed_p (cfun) | |
1064 | && powi_cost (n) <= POWI_MAX_MULTS))) | |
1065 | return powi_as_mults (gsi, loc, arg0, n); | |
1066 | ||
1067 | return NULL_TREE; | |
1068 | } | |
1069 | ||
ba869341 | 1070 | /* Build a gimple call statement that calls FN with argument ARG. |
83d5977e | 1071 | Set the lhs of the call statement to a fresh SSA name. Insert the |
ba869341 BS |
1072 | statement prior to GSI's current position, and return the fresh |
1073 | SSA name. */ | |
1074 | ||
1075 | static tree | |
6e96f98a | 1076 | build_and_insert_call (gimple_stmt_iterator *gsi, location_t loc, |
83d5977e | 1077 | tree fn, tree arg) |
ba869341 BS |
1078 | { |
1079 | gimple call_stmt; | |
1080 | tree ssa_target; | |
1081 | ||
ba869341 | 1082 | call_stmt = gimple_build_call (fn, 1, arg); |
83d5977e | 1083 | ssa_target = make_temp_ssa_name (TREE_TYPE (arg), NULL, "powroot"); |
ba869341 BS |
1084 | gimple_set_lhs (call_stmt, ssa_target); |
1085 | gimple_set_location (call_stmt, loc); | |
1086 | gsi_insert_before (gsi, call_stmt, GSI_SAME_STMT); | |
1087 | ||
1088 | return ssa_target; | |
1089 | } | |
1090 | ||
6e96f98a BS |
1091 | /* Build a gimple binary operation with the given CODE and arguments |
1092 | ARG0, ARG1, assigning the result to a new SSA name for variable | |
1093 | TARGET. Insert the statement prior to GSI's current position, and | |
1094 | return the fresh SSA name.*/ | |
1095 | ||
1096 | static tree | |
1097 | build_and_insert_binop (gimple_stmt_iterator *gsi, location_t loc, | |
83d5977e RG |
1098 | const char *name, enum tree_code code, |
1099 | tree arg0, tree arg1) | |
6e96f98a | 1100 | { |
83d5977e | 1101 | tree result = make_temp_ssa_name (TREE_TYPE (arg0), NULL, name); |
6e96f98a BS |
1102 | gimple stmt = gimple_build_assign_with_ops (code, result, arg0, arg1); |
1103 | gimple_set_location (stmt, loc); | |
1104 | gsi_insert_before (gsi, stmt, GSI_SAME_STMT); | |
1105 | return result; | |
1106 | } | |
1107 | ||
d7e2a1c1 | 1108 | /* Build a gimple reference operation with the given CODE and argument |
83d5977e | 1109 | ARG, assigning the result to a new SSA name of TYPE with NAME. |
d7e2a1c1 BS |
1110 | Insert the statement prior to GSI's current position, and return |
1111 | the fresh SSA name. */ | |
1112 | ||
1113 | static inline tree | |
1114 | build_and_insert_ref (gimple_stmt_iterator *gsi, location_t loc, tree type, | |
83d5977e | 1115 | const char *name, enum tree_code code, tree arg0) |
d7e2a1c1 | 1116 | { |
83d5977e | 1117 | tree result = make_temp_ssa_name (type, NULL, name); |
d7e2a1c1 BS |
1118 | gimple stmt = gimple_build_assign (result, build1 (code, type, arg0)); |
1119 | gimple_set_location (stmt, loc); | |
1120 | gsi_insert_before (gsi, stmt, GSI_SAME_STMT); | |
1121 | return result; | |
1122 | } | |
1123 | ||
83d5977e | 1124 | /* Build a gimple assignment to cast VAL to TYPE. Insert the statement |
5dfe80ba AS |
1125 | prior to GSI's current position, and return the fresh SSA name. */ |
1126 | ||
1127 | static tree | |
1128 | build_and_insert_cast (gimple_stmt_iterator *gsi, location_t loc, | |
83d5977e | 1129 | tree type, tree val) |
5dfe80ba | 1130 | { |
83d5977e RG |
1131 | tree result = make_ssa_name (type, NULL); |
1132 | gimple stmt = gimple_build_assign_with_ops (NOP_EXPR, result, val, NULL_TREE); | |
1133 | gimple_set_location (stmt, loc); | |
1134 | gsi_insert_before (gsi, stmt, GSI_SAME_STMT); | |
1135 | return result; | |
5dfe80ba AS |
1136 | } |
1137 | ||
d24ad7d6 BS |
1138 | /* ARG0 and ARG1 are the two arguments to a pow builtin call in GSI |
1139 | with location info LOC. If possible, create an equivalent and | |
1140 | less expensive sequence of statements prior to GSI, and return an | |
1141 | expession holding the result. */ | |
1142 | ||
1143 | static tree | |
1144 | gimple_expand_builtin_pow (gimple_stmt_iterator *gsi, location_t loc, | |
1145 | tree arg0, tree arg1) | |
1146 | { | |
ba869341 | 1147 | REAL_VALUE_TYPE c, cint, dconst1_4, dconst3_4, dconst1_3, dconst1_6; |
6e96f98a | 1148 | REAL_VALUE_TYPE c2, dconst3; |
d24ad7d6 | 1149 | HOST_WIDE_INT n; |
6e96f98a | 1150 | tree type, sqrtfn, cbrtfn, sqrt_arg0, sqrt_sqrt, result, cbrt_x, powi_cbrt_x; |
ef4bddc2 | 1151 | machine_mode mode; |
0bfbca58 | 1152 | bool hw_sqrt_exists, c_is_int, c2_is_int; |
d24ad7d6 BS |
1153 | |
1154 | /* If the exponent isn't a constant, there's nothing of interest | |
1155 | to be done. */ | |
1156 | if (TREE_CODE (arg1) != REAL_CST) | |
1157 | return NULL_TREE; | |
1158 | ||
ba869341 BS |
1159 | /* If the exponent is equivalent to an integer, expand to an optimal |
1160 | multiplication sequence when profitable. */ | |
d24ad7d6 BS |
1161 | c = TREE_REAL_CST (arg1); |
1162 | n = real_to_integer (&c); | |
807e902e | 1163 | real_from_integer (&cint, VOIDmode, n, SIGNED); |
0bfbca58 | 1164 | c_is_int = real_identical (&c, &cint); |
d24ad7d6 | 1165 | |
0bfbca58 | 1166 | if (c_is_int |
d24ad7d6 BS |
1167 | && ((n >= -1 && n <= 2) |
1168 | || (flag_unsafe_math_optimizations | |
72798784 | 1169 | && optimize_bb_for_speed_p (gsi_bb (*gsi)) |
d24ad7d6 BS |
1170 | && powi_cost (n) <= POWI_MAX_MULTS))) |
1171 | return gimple_expand_builtin_powi (gsi, loc, arg0, n); | |
1172 | ||
ba869341 BS |
1173 | /* Attempt various optimizations using sqrt and cbrt. */ |
1174 | type = TREE_TYPE (arg0); | |
1175 | mode = TYPE_MODE (type); | |
1176 | sqrtfn = mathfn_built_in (type, BUILT_IN_SQRT); | |
1177 | ||
1178 | /* Optimize pow(x,0.5) = sqrt(x). This replacement is always safe | |
1179 | unless signed zeros must be maintained. pow(-0,0.5) = +0, while | |
1180 | sqrt(-0) = -0. */ | |
1181 | if (sqrtfn | |
1182 | && REAL_VALUES_EQUAL (c, dconsthalf) | |
1183 | && !HONOR_SIGNED_ZEROS (mode)) | |
83d5977e | 1184 | return build_and_insert_call (gsi, loc, sqrtfn, arg0); |
ba869341 BS |
1185 | |
1186 | /* Optimize pow(x,0.25) = sqrt(sqrt(x)). Assume on most machines that | |
1187 | a builtin sqrt instruction is smaller than a call to pow with 0.25, | |
1188 | so do this optimization even if -Os. Don't do this optimization | |
1189 | if we don't have a hardware sqrt insn. */ | |
1190 | dconst1_4 = dconst1; | |
1191 | SET_REAL_EXP (&dconst1_4, REAL_EXP (&dconst1_4) - 2); | |
d7e2a1c1 | 1192 | hw_sqrt_exists = optab_handler (sqrt_optab, mode) != CODE_FOR_nothing; |
ba869341 BS |
1193 | |
1194 | if (flag_unsafe_math_optimizations | |
1195 | && sqrtfn | |
1196 | && REAL_VALUES_EQUAL (c, dconst1_4) | |
1197 | && hw_sqrt_exists) | |
1198 | { | |
1199 | /* sqrt(x) */ | |
83d5977e | 1200 | sqrt_arg0 = build_and_insert_call (gsi, loc, sqrtfn, arg0); |
ba869341 BS |
1201 | |
1202 | /* sqrt(sqrt(x)) */ | |
83d5977e | 1203 | return build_and_insert_call (gsi, loc, sqrtfn, sqrt_arg0); |
ba869341 BS |
1204 | } |
1205 | ||
1206 | /* Optimize pow(x,0.75) = sqrt(x) * sqrt(sqrt(x)) unless we are | |
1207 | optimizing for space. Don't do this optimization if we don't have | |
1208 | a hardware sqrt insn. */ | |
807e902e | 1209 | real_from_integer (&dconst3_4, VOIDmode, 3, SIGNED); |
ba869341 BS |
1210 | SET_REAL_EXP (&dconst3_4, REAL_EXP (&dconst3_4) - 2); |
1211 | ||
1212 | if (flag_unsafe_math_optimizations | |
1213 | && sqrtfn | |
1214 | && optimize_function_for_speed_p (cfun) | |
1215 | && REAL_VALUES_EQUAL (c, dconst3_4) | |
1216 | && hw_sqrt_exists) | |
1217 | { | |
1218 | /* sqrt(x) */ | |
83d5977e | 1219 | sqrt_arg0 = build_and_insert_call (gsi, loc, sqrtfn, arg0); |
ba869341 BS |
1220 | |
1221 | /* sqrt(sqrt(x)) */ | |
83d5977e | 1222 | sqrt_sqrt = build_and_insert_call (gsi, loc, sqrtfn, sqrt_arg0); |
ba869341 BS |
1223 | |
1224 | /* sqrt(x) * sqrt(sqrt(x)) */ | |
83d5977e | 1225 | return build_and_insert_binop (gsi, loc, "powroot", MULT_EXPR, |
6e96f98a | 1226 | sqrt_arg0, sqrt_sqrt); |
ba869341 BS |
1227 | } |
1228 | ||
1229 | /* Optimize pow(x,1./3.) = cbrt(x). This requires unsafe math | |
1230 | optimizations since 1./3. is not exactly representable. If x | |
1231 | is negative and finite, the correct value of pow(x,1./3.) is | |
1232 | a NaN with the "invalid" exception raised, because the value | |
1233 | of 1./3. actually has an even denominator. The correct value | |
1234 | of cbrt(x) is a negative real value. */ | |
1235 | cbrtfn = mathfn_built_in (type, BUILT_IN_CBRT); | |
1236 | dconst1_3 = real_value_truncate (mode, dconst_third ()); | |
1237 | ||
1238 | if (flag_unsafe_math_optimizations | |
1239 | && cbrtfn | |
06bc3ec7 | 1240 | && (gimple_val_nonnegative_real_p (arg0) || !HONOR_NANS (mode)) |
ba869341 | 1241 | && REAL_VALUES_EQUAL (c, dconst1_3)) |
83d5977e | 1242 | return build_and_insert_call (gsi, loc, cbrtfn, arg0); |
ba869341 BS |
1243 | |
1244 | /* Optimize pow(x,1./6.) = cbrt(sqrt(x)). Don't do this optimization | |
1245 | if we don't have a hardware sqrt insn. */ | |
1246 | dconst1_6 = dconst1_3; | |
1247 | SET_REAL_EXP (&dconst1_6, REAL_EXP (&dconst1_6) - 1); | |
1248 | ||
1249 | if (flag_unsafe_math_optimizations | |
1250 | && sqrtfn | |
1251 | && cbrtfn | |
06bc3ec7 | 1252 | && (gimple_val_nonnegative_real_p (arg0) || !HONOR_NANS (mode)) |
ba869341 BS |
1253 | && optimize_function_for_speed_p (cfun) |
1254 | && hw_sqrt_exists | |
1255 | && REAL_VALUES_EQUAL (c, dconst1_6)) | |
1256 | { | |
1257 | /* sqrt(x) */ | |
83d5977e | 1258 | sqrt_arg0 = build_and_insert_call (gsi, loc, sqrtfn, arg0); |
ba869341 BS |
1259 | |
1260 | /* cbrt(sqrt(x)) */ | |
83d5977e | 1261 | return build_and_insert_call (gsi, loc, cbrtfn, sqrt_arg0); |
6e96f98a BS |
1262 | } |
1263 | ||
0bfbca58 JJ |
1264 | /* Optimize pow(x,c), where n = 2c for some nonzero integer n |
1265 | and c not an integer, into | |
6e96f98a BS |
1266 | |
1267 | sqrt(x) * powi(x, n/2), n > 0; | |
1268 | 1.0 / (sqrt(x) * powi(x, abs(n/2))), n < 0. | |
1269 | ||
1270 | Do not calculate the powi factor when n/2 = 0. */ | |
1271 | real_arithmetic (&c2, MULT_EXPR, &c, &dconst2); | |
1272 | n = real_to_integer (&c2); | |
807e902e | 1273 | real_from_integer (&cint, VOIDmode, n, SIGNED); |
0bfbca58 | 1274 | c2_is_int = real_identical (&c2, &cint); |
6e96f98a BS |
1275 | |
1276 | if (flag_unsafe_math_optimizations | |
1277 | && sqrtfn | |
0bfbca58 JJ |
1278 | && c2_is_int |
1279 | && !c_is_int | |
1280 | && optimize_function_for_speed_p (cfun)) | |
6e96f98a BS |
1281 | { |
1282 | tree powi_x_ndiv2 = NULL_TREE; | |
1283 | ||
1284 | /* Attempt to fold powi(arg0, abs(n/2)) into multiplies. If not | |
1285 | possible or profitable, give up. Skip the degenerate case when | |
1286 | n is 1 or -1, where the result is always 1. */ | |
4c9cf7af | 1287 | if (absu_hwi (n) != 1) |
6e96f98a | 1288 | { |
9f813990 PC |
1289 | powi_x_ndiv2 = gimple_expand_builtin_powi (gsi, loc, arg0, |
1290 | abs_hwi (n / 2)); | |
6e96f98a BS |
1291 | if (!powi_x_ndiv2) |
1292 | return NULL_TREE; | |
1293 | } | |
1294 | ||
1295 | /* Calculate sqrt(x). When n is not 1 or -1, multiply it by the | |
1296 | result of the optimal multiply sequence just calculated. */ | |
83d5977e | 1297 | sqrt_arg0 = build_and_insert_call (gsi, loc, sqrtfn, arg0); |
6e96f98a | 1298 | |
4c9cf7af | 1299 | if (absu_hwi (n) == 1) |
6e96f98a BS |
1300 | result = sqrt_arg0; |
1301 | else | |
83d5977e | 1302 | result = build_and_insert_binop (gsi, loc, "powroot", MULT_EXPR, |
6e96f98a BS |
1303 | sqrt_arg0, powi_x_ndiv2); |
1304 | ||
1305 | /* If n is negative, reciprocate the result. */ | |
1306 | if (n < 0) | |
83d5977e | 1307 | result = build_and_insert_binop (gsi, loc, "powroot", RDIV_EXPR, |
6e96f98a BS |
1308 | build_real (type, dconst1), result); |
1309 | return result; | |
1310 | } | |
1311 | ||
1312 | /* Optimize pow(x,c), where 3c = n for some nonzero integer n, into | |
1313 | ||
1314 | powi(x, n/3) * powi(cbrt(x), n%3), n > 0; | |
1315 | 1.0 / (powi(x, abs(n)/3) * powi(cbrt(x), abs(n)%3)), n < 0. | |
1316 | ||
1317 | Do not calculate the first factor when n/3 = 0. As cbrt(x) is | |
1318 | different from pow(x, 1./3.) due to rounding and behavior with | |
1319 | negative x, we need to constrain this transformation to unsafe | |
1320 | math and positive x or finite math. */ | |
807e902e | 1321 | real_from_integer (&dconst3, VOIDmode, 3, SIGNED); |
6e96f98a BS |
1322 | real_arithmetic (&c2, MULT_EXPR, &c, &dconst3); |
1323 | real_round (&c2, mode, &c2); | |
1324 | n = real_to_integer (&c2); | |
807e902e | 1325 | real_from_integer (&cint, VOIDmode, n, SIGNED); |
6e96f98a BS |
1326 | real_arithmetic (&c2, RDIV_EXPR, &cint, &dconst3); |
1327 | real_convert (&c2, mode, &c2); | |
1328 | ||
1329 | if (flag_unsafe_math_optimizations | |
1330 | && cbrtfn | |
06bc3ec7 | 1331 | && (gimple_val_nonnegative_real_p (arg0) || !HONOR_NANS (mode)) |
6e96f98a | 1332 | && real_identical (&c2, &c) |
0bfbca58 | 1333 | && !c2_is_int |
6e96f98a BS |
1334 | && optimize_function_for_speed_p (cfun) |
1335 | && powi_cost (n / 3) <= POWI_MAX_MULTS) | |
1336 | { | |
1337 | tree powi_x_ndiv3 = NULL_TREE; | |
1338 | ||
1339 | /* Attempt to fold powi(arg0, abs(n/3)) into multiplies. If not | |
1340 | possible or profitable, give up. Skip the degenerate case when | |
1341 | abs(n) < 3, where the result is always 1. */ | |
4c9cf7af | 1342 | if (absu_hwi (n) >= 3) |
6e96f98a BS |
1343 | { |
1344 | powi_x_ndiv3 = gimple_expand_builtin_powi (gsi, loc, arg0, | |
9f813990 | 1345 | abs_hwi (n / 3)); |
6e96f98a BS |
1346 | if (!powi_x_ndiv3) |
1347 | return NULL_TREE; | |
1348 | } | |
1349 | ||
1350 | /* Calculate powi(cbrt(x), n%3). Don't use gimple_expand_builtin_powi | |
1351 | as that creates an unnecessary variable. Instead, just produce | |
1352 | either cbrt(x) or cbrt(x) * cbrt(x). */ | |
83d5977e | 1353 | cbrt_x = build_and_insert_call (gsi, loc, cbrtfn, arg0); |
6e96f98a | 1354 | |
4c9cf7af | 1355 | if (absu_hwi (n) % 3 == 1) |
6e96f98a BS |
1356 | powi_cbrt_x = cbrt_x; |
1357 | else | |
83d5977e | 1358 | powi_cbrt_x = build_and_insert_binop (gsi, loc, "powroot", MULT_EXPR, |
6e96f98a BS |
1359 | cbrt_x, cbrt_x); |
1360 | ||
1361 | /* Multiply the two subexpressions, unless powi(x,abs(n)/3) = 1. */ | |
4c9cf7af | 1362 | if (absu_hwi (n) < 3) |
6e96f98a BS |
1363 | result = powi_cbrt_x; |
1364 | else | |
83d5977e | 1365 | result = build_and_insert_binop (gsi, loc, "powroot", MULT_EXPR, |
6e96f98a BS |
1366 | powi_x_ndiv3, powi_cbrt_x); |
1367 | ||
1368 | /* If n is negative, reciprocate the result. */ | |
1369 | if (n < 0) | |
83d5977e | 1370 | result = build_and_insert_binop (gsi, loc, "powroot", RDIV_EXPR, |
6e96f98a BS |
1371 | build_real (type, dconst1), result); |
1372 | ||
1373 | return result; | |
ba869341 BS |
1374 | } |
1375 | ||
6e96f98a | 1376 | /* No optimizations succeeded. */ |
d24ad7d6 BS |
1377 | return NULL_TREE; |
1378 | } | |
1379 | ||
d7e2a1c1 BS |
1380 | /* ARG is the argument to a cabs builtin call in GSI with location info |
1381 | LOC. Create a sequence of statements prior to GSI that calculates | |
1382 | sqrt(R*R + I*I), where R and I are the real and imaginary components | |
1383 | of ARG, respectively. Return an expression holding the result. */ | |
1384 | ||
1385 | static tree | |
1386 | gimple_expand_builtin_cabs (gimple_stmt_iterator *gsi, location_t loc, tree arg) | |
1387 | { | |
83d5977e | 1388 | tree real_part, imag_part, addend1, addend2, sum, result; |
d7e2a1c1 BS |
1389 | tree type = TREE_TYPE (TREE_TYPE (arg)); |
1390 | tree sqrtfn = mathfn_built_in (type, BUILT_IN_SQRT); | |
ef4bddc2 | 1391 | machine_mode mode = TYPE_MODE (type); |
d7e2a1c1 BS |
1392 | |
1393 | if (!flag_unsafe_math_optimizations | |
1394 | || !optimize_bb_for_speed_p (gimple_bb (gsi_stmt (*gsi))) | |
1395 | || !sqrtfn | |
1396 | || optab_handler (sqrt_optab, mode) == CODE_FOR_nothing) | |
1397 | return NULL_TREE; | |
1398 | ||
83d5977e | 1399 | real_part = build_and_insert_ref (gsi, loc, type, "cabs", |
d7e2a1c1 | 1400 | REALPART_EXPR, arg); |
83d5977e | 1401 | addend1 = build_and_insert_binop (gsi, loc, "cabs", MULT_EXPR, |
d7e2a1c1 | 1402 | real_part, real_part); |
83d5977e | 1403 | imag_part = build_and_insert_ref (gsi, loc, type, "cabs", |
d7e2a1c1 | 1404 | IMAGPART_EXPR, arg); |
83d5977e | 1405 | addend2 = build_and_insert_binop (gsi, loc, "cabs", MULT_EXPR, |
d7e2a1c1 | 1406 | imag_part, imag_part); |
83d5977e RG |
1407 | sum = build_and_insert_binop (gsi, loc, "cabs", PLUS_EXPR, addend1, addend2); |
1408 | result = build_and_insert_call (gsi, loc, sqrtfn, sum); | |
d7e2a1c1 BS |
1409 | |
1410 | return result; | |
1411 | } | |
1412 | ||
2f397a93 | 1413 | /* Go through all calls to sin, cos and cexpi and call execute_cse_sincos_1 |
78be79d5 BS |
1414 | on the SSA_NAME argument of each of them. Also expand powi(x,n) into |
1415 | an optimal number of multiplies, when n is a constant. */ | |
2f397a93 | 1416 | |
be55bfe6 TS |
1417 | namespace { |
1418 | ||
1419 | const pass_data pass_data_cse_sincos = | |
1420 | { | |
1421 | GIMPLE_PASS, /* type */ | |
1422 | "sincos", /* name */ | |
1423 | OPTGROUP_NONE, /* optinfo_flags */ | |
be55bfe6 TS |
1424 | TV_NONE, /* tv_id */ |
1425 | PROP_ssa, /* properties_required */ | |
1426 | 0, /* properties_provided */ | |
1427 | 0, /* properties_destroyed */ | |
1428 | 0, /* todo_flags_start */ | |
3bea341f | 1429 | TODO_update_ssa, /* todo_flags_finish */ |
be55bfe6 TS |
1430 | }; |
1431 | ||
1432 | class pass_cse_sincos : public gimple_opt_pass | |
1433 | { | |
1434 | public: | |
1435 | pass_cse_sincos (gcc::context *ctxt) | |
1436 | : gimple_opt_pass (pass_data_cse_sincos, ctxt) | |
1437 | {} | |
1438 | ||
1439 | /* opt_pass methods: */ | |
1440 | virtual bool gate (function *) | |
1441 | { | |
1442 | /* We no longer require either sincos or cexp, since powi expansion | |
1443 | piggybacks on this pass. */ | |
1444 | return optimize; | |
1445 | } | |
1446 | ||
1447 | virtual unsigned int execute (function *); | |
1448 | ||
1449 | }; // class pass_cse_sincos | |
1450 | ||
1451 | unsigned int | |
1452 | pass_cse_sincos::execute (function *fun) | |
2f397a93 RG |
1453 | { |
1454 | basic_block bb; | |
90bc1cb8 | 1455 | bool cfg_changed = false; |
2f397a93 RG |
1456 | |
1457 | calculate_dominance_info (CDI_DOMINATORS); | |
4da3b811 | 1458 | memset (&sincos_stats, 0, sizeof (sincos_stats)); |
2f397a93 | 1459 | |
be55bfe6 | 1460 | FOR_EACH_BB_FN (bb, fun) |
2f397a93 | 1461 | { |
726a989a | 1462 | gimple_stmt_iterator gsi; |
3b9ee1cc | 1463 | bool cleanup_eh = false; |
2f397a93 | 1464 | |
726a989a | 1465 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
2f397a93 | 1466 | { |
726a989a | 1467 | gimple stmt = gsi_stmt (gsi); |
2f397a93 RG |
1468 | tree fndecl; |
1469 | ||
3b9ee1cc JJ |
1470 | /* Only the last stmt in a bb could throw, no need to call |
1471 | gimple_purge_dead_eh_edges if we change something in the middle | |
1472 | of a basic block. */ | |
1473 | cleanup_eh = false; | |
1474 | ||
726a989a RB |
1475 | if (is_gimple_call (stmt) |
1476 | && gimple_call_lhs (stmt) | |
1477 | && (fndecl = gimple_call_fndecl (stmt)) | |
2f397a93 RG |
1478 | && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) |
1479 | { | |
78be79d5 BS |
1480 | tree arg, arg0, arg1, result; |
1481 | HOST_WIDE_INT n; | |
1482 | location_t loc; | |
2f397a93 RG |
1483 | |
1484 | switch (DECL_FUNCTION_CODE (fndecl)) | |
1485 | { | |
1486 | CASE_FLT_FN (BUILT_IN_COS): | |
1487 | CASE_FLT_FN (BUILT_IN_SIN): | |
1488 | CASE_FLT_FN (BUILT_IN_CEXPI): | |
fa65a9cf | 1489 | /* Make sure we have either sincos or cexp. */ |
d33d9e47 AI |
1490 | if (!targetm.libc_has_function (function_c99_math_complex) |
1491 | && !targetm.libc_has_function (function_sincos)) | |
fa65a9cf BS |
1492 | break; |
1493 | ||
726a989a | 1494 | arg = gimple_call_arg (stmt, 0); |
2f397a93 | 1495 | if (TREE_CODE (arg) == SSA_NAME) |
90bc1cb8 | 1496 | cfg_changed |= execute_cse_sincos_1 (arg); |
2f397a93 RG |
1497 | break; |
1498 | ||
d24ad7d6 BS |
1499 | CASE_FLT_FN (BUILT_IN_POW): |
1500 | arg0 = gimple_call_arg (stmt, 0); | |
1501 | arg1 = gimple_call_arg (stmt, 1); | |
1502 | ||
1503 | loc = gimple_location (stmt); | |
1504 | result = gimple_expand_builtin_pow (&gsi, loc, arg0, arg1); | |
1505 | ||
1506 | if (result) | |
1507 | { | |
1508 | tree lhs = gimple_get_lhs (stmt); | |
1509 | gimple new_stmt = gimple_build_assign (lhs, result); | |
1510 | gimple_set_location (new_stmt, loc); | |
1511 | unlink_stmt_vdef (stmt); | |
1512 | gsi_replace (&gsi, new_stmt, true); | |
3b9ee1cc | 1513 | cleanup_eh = true; |
3d3f2249 RG |
1514 | if (gimple_vdef (stmt)) |
1515 | release_ssa_name (gimple_vdef (stmt)); | |
d24ad7d6 BS |
1516 | } |
1517 | break; | |
1518 | ||
78be79d5 BS |
1519 | CASE_FLT_FN (BUILT_IN_POWI): |
1520 | arg0 = gimple_call_arg (stmt, 0); | |
1521 | arg1 = gimple_call_arg (stmt, 1); | |
78be79d5 | 1522 | loc = gimple_location (stmt); |
0fa6e0ef | 1523 | |
e3530904 | 1524 | if (real_minus_onep (arg0)) |
0fa6e0ef TB |
1525 | { |
1526 | tree t0, t1, cond, one, minus_one; | |
1527 | gimple stmt; | |
1528 | ||
1529 | t0 = TREE_TYPE (arg0); | |
1530 | t1 = TREE_TYPE (arg1); | |
1531 | one = build_real (t0, dconst1); | |
1532 | minus_one = build_real (t0, dconstm1); | |
1533 | ||
1534 | cond = make_temp_ssa_name (t1, NULL, "powi_cond"); | |
1535 | stmt = gimple_build_assign_with_ops (BIT_AND_EXPR, cond, | |
1536 | arg1, | |
1537 | build_int_cst (t1, | |
1538 | 1)); | |
1539 | gimple_set_location (stmt, loc); | |
1540 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
1541 | ||
1542 | result = make_temp_ssa_name (t0, NULL, "powi"); | |
1543 | stmt = gimple_build_assign_with_ops (COND_EXPR, result, | |
1544 | cond, | |
1545 | minus_one, one); | |
1546 | gimple_set_location (stmt, loc); | |
1547 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
1548 | } | |
1549 | else | |
1550 | { | |
9541ffee | 1551 | if (!tree_fits_shwi_p (arg1)) |
daf4e940 TB |
1552 | break; |
1553 | ||
eb1ce453 | 1554 | n = tree_to_shwi (arg1); |
0fa6e0ef TB |
1555 | result = gimple_expand_builtin_powi (&gsi, loc, arg0, n); |
1556 | } | |
78be79d5 BS |
1557 | |
1558 | if (result) | |
1559 | { | |
1560 | tree lhs = gimple_get_lhs (stmt); | |
1561 | gimple new_stmt = gimple_build_assign (lhs, result); | |
1562 | gimple_set_location (new_stmt, loc); | |
d7e2a1c1 BS |
1563 | unlink_stmt_vdef (stmt); |
1564 | gsi_replace (&gsi, new_stmt, true); | |
3b9ee1cc | 1565 | cleanup_eh = true; |
3d3f2249 RG |
1566 | if (gimple_vdef (stmt)) |
1567 | release_ssa_name (gimple_vdef (stmt)); | |
d7e2a1c1 BS |
1568 | } |
1569 | break; | |
1570 | ||
1571 | CASE_FLT_FN (BUILT_IN_CABS): | |
1572 | arg0 = gimple_call_arg (stmt, 0); | |
1573 | loc = gimple_location (stmt); | |
1574 | result = gimple_expand_builtin_cabs (&gsi, loc, arg0); | |
1575 | ||
1576 | if (result) | |
1577 | { | |
1578 | tree lhs = gimple_get_lhs (stmt); | |
1579 | gimple new_stmt = gimple_build_assign (lhs, result); | |
1580 | gimple_set_location (new_stmt, loc); | |
78be79d5 BS |
1581 | unlink_stmt_vdef (stmt); |
1582 | gsi_replace (&gsi, new_stmt, true); | |
3b9ee1cc | 1583 | cleanup_eh = true; |
3d3f2249 RG |
1584 | if (gimple_vdef (stmt)) |
1585 | release_ssa_name (gimple_vdef (stmt)); | |
78be79d5 BS |
1586 | } |
1587 | break; | |
1588 | ||
2f397a93 RG |
1589 | default:; |
1590 | } | |
1591 | } | |
1592 | } | |
3b9ee1cc JJ |
1593 | if (cleanup_eh) |
1594 | cfg_changed |= gimple_purge_dead_eh_edges (bb); | |
2f397a93 RG |
1595 | } |
1596 | ||
be55bfe6 | 1597 | statistics_counter_event (fun, "sincos statements inserted", |
4da3b811 NF |
1598 | sincos_stats.inserted); |
1599 | ||
2f397a93 | 1600 | free_dominance_info (CDI_DOMINATORS); |
90bc1cb8 | 1601 | return cfg_changed ? TODO_cleanup_cfg : 0; |
2f397a93 RG |
1602 | } |
1603 | ||
27a4cd48 DM |
1604 | } // anon namespace |
1605 | ||
1606 | gimple_opt_pass * | |
1607 | make_pass_cse_sincos (gcc::context *ctxt) | |
1608 | { | |
1609 | return new pass_cse_sincos (ctxt); | |
1610 | } | |
1611 | ||
03bd2f1a AK |
1612 | /* A symbolic number is used to detect byte permutation and selection |
1613 | patterns. Therefore the field N contains an artificial number | |
e3ef4162 | 1614 | consisting of octet sized markers: |
03bd2f1a | 1615 | |
e3ef4162 | 1616 | 0 - target byte has the value 0 |
aa29ea0c | 1617 | FF - target byte has an unknown value (eg. due to sign extension) |
e3ef4162 | 1618 | 1..size - marker value is the target byte index minus one. |
73984f84 TP |
1619 | |
1620 | To detect permutations on memory sources (arrays and structures), a symbolic | |
1621 | number is also associated a base address (the array or structure the load is | |
1622 | made from), an offset from the base address and a range which gives the | |
1623 | difference between the highest and lowest accessed memory location to make | |
1624 | such a symbolic number. The range is thus different from size which reflects | |
1625 | the size of the type of current expression. Note that for non memory source, | |
1626 | range holds the same value as size. | |
1627 | ||
1628 | For instance, for an array char a[], (short) a[0] | (short) a[3] would have | |
1629 | a size of 2 but a range of 4 while (short) a[0] | ((short) a[0] << 1) would | |
1630 | still have a size of 2 but this time a range of 1. */ | |
03bd2f1a AK |
1631 | |
1632 | struct symbolic_number { | |
a9243bfc | 1633 | uint64_t n; |
698ff107 | 1634 | tree type; |
73984f84 TP |
1635 | tree base_addr; |
1636 | tree offset; | |
1637 | HOST_WIDE_INT bytepos; | |
1638 | tree alias_set; | |
1639 | tree vuse; | |
1640 | unsigned HOST_WIDE_INT range; | |
03bd2f1a AK |
1641 | }; |
1642 | ||
e3ef4162 | 1643 | #define BITS_PER_MARKER 8 |
aa29ea0c TP |
1644 | #define MARKER_MASK ((1 << BITS_PER_MARKER) - 1) |
1645 | #define MARKER_BYTE_UNKNOWN MARKER_MASK | |
1646 | #define HEAD_MARKER(n, size) \ | |
1647 | ((n) & ((uint64_t) MARKER_MASK << (((size) - 1) * BITS_PER_MARKER))) | |
e3ef4162 | 1648 | |
73984f84 TP |
1649 | /* The number which the find_bswap_or_nop_1 result should match in |
1650 | order to have a nop. The number is masked according to the size of | |
1651 | the symbolic number before using it. */ | |
a9243bfc RB |
1652 | #define CMPNOP (sizeof (int64_t) < 8 ? 0 : \ |
1653 | (uint64_t)0x08070605 << 32 | 0x04030201) | |
73984f84 TP |
1654 | |
1655 | /* The number which the find_bswap_or_nop_1 result should match in | |
1656 | order to have a byte swap. The number is masked according to the | |
1657 | size of the symbolic number before using it. */ | |
a9243bfc RB |
1658 | #define CMPXCHG (sizeof (int64_t) < 8 ? 0 : \ |
1659 | (uint64_t)0x01020304 << 32 | 0x05060708) | |
73984f84 | 1660 | |
03bd2f1a AK |
1661 | /* Perform a SHIFT or ROTATE operation by COUNT bits on symbolic |
1662 | number N. Return false if the requested operation is not permitted | |
1663 | on a symbolic number. */ | |
1664 | ||
1665 | static inline bool | |
1666 | do_shift_rotate (enum tree_code code, | |
1667 | struct symbolic_number *n, | |
1668 | int count) | |
1669 | { | |
aa29ea0c TP |
1670 | int i, size = TYPE_PRECISION (n->type) / BITS_PER_UNIT; |
1671 | unsigned head_marker; | |
698ff107 | 1672 | |
e3ef4162 | 1673 | if (count % BITS_PER_UNIT != 0) |
03bd2f1a | 1674 | return false; |
e3ef4162 | 1675 | count = (count / BITS_PER_UNIT) * BITS_PER_MARKER; |
03bd2f1a AK |
1676 | |
1677 | /* Zero out the extra bits of N in order to avoid them being shifted | |
1678 | into the significant bits. */ | |
e3ef4162 TP |
1679 | if (size < 64 / BITS_PER_MARKER) |
1680 | n->n &= ((uint64_t) 1 << (size * BITS_PER_MARKER)) - 1; | |
03bd2f1a AK |
1681 | |
1682 | switch (code) | |
1683 | { | |
1684 | case LSHIFT_EXPR: | |
1685 | n->n <<= count; | |
1686 | break; | |
1687 | case RSHIFT_EXPR: | |
aa29ea0c | 1688 | head_marker = HEAD_MARKER (n->n, size); |
03bd2f1a | 1689 | n->n >>= count; |
aa29ea0c TP |
1690 | /* Arithmetic shift of signed type: result is dependent on the value. */ |
1691 | if (!TYPE_UNSIGNED (n->type) && head_marker) | |
1692 | for (i = 0; i < count / BITS_PER_MARKER; i++) | |
1693 | n->n |= (uint64_t) MARKER_BYTE_UNKNOWN | |
1694 | << ((size - 1 - i) * BITS_PER_MARKER); | |
03bd2f1a AK |
1695 | break; |
1696 | case LROTATE_EXPR: | |
e3ef4162 | 1697 | n->n = (n->n << count) | (n->n >> ((size * BITS_PER_MARKER) - count)); |
03bd2f1a AK |
1698 | break; |
1699 | case RROTATE_EXPR: | |
e3ef4162 | 1700 | n->n = (n->n >> count) | (n->n << ((size * BITS_PER_MARKER) - count)); |
03bd2f1a AK |
1701 | break; |
1702 | default: | |
1703 | return false; | |
1704 | } | |
5da49a9d | 1705 | /* Zero unused bits for size. */ |
e3ef4162 TP |
1706 | if (size < 64 / BITS_PER_MARKER) |
1707 | n->n &= ((uint64_t) 1 << (size * BITS_PER_MARKER)) - 1; | |
03bd2f1a AK |
1708 | return true; |
1709 | } | |
1710 | ||
1711 | /* Perform sanity checking for the symbolic number N and the gimple | |
1712 | statement STMT. */ | |
1713 | ||
1714 | static inline bool | |
1715 | verify_symbolic_number_p (struct symbolic_number *n, gimple stmt) | |
1716 | { | |
1717 | tree lhs_type; | |
1718 | ||
1719 | lhs_type = gimple_expr_type (stmt); | |
1720 | ||
1721 | if (TREE_CODE (lhs_type) != INTEGER_TYPE) | |
1722 | return false; | |
1723 | ||
698ff107 | 1724 | if (TYPE_PRECISION (lhs_type) != TYPE_PRECISION (n->type)) |
03bd2f1a AK |
1725 | return false; |
1726 | ||
1727 | return true; | |
1728 | } | |
1729 | ||
3cc272c1 TP |
1730 | /* Initialize the symbolic number N for the bswap pass from the base element |
1731 | SRC manipulated by the bitwise OR expression. */ | |
1732 | ||
1733 | static bool | |
1734 | init_symbolic_number (struct symbolic_number *n, tree src) | |
1735 | { | |
698ff107 TP |
1736 | int size; |
1737 | ||
3cc272c1 TP |
1738 | n->base_addr = n->offset = n->alias_set = n->vuse = NULL_TREE; |
1739 | ||
1740 | /* Set up the symbolic number N by setting each byte to a value between 1 and | |
1741 | the byte size of rhs1. The highest order byte is set to n->size and the | |
1742 | lowest order byte to 1. */ | |
698ff107 TP |
1743 | n->type = TREE_TYPE (src); |
1744 | size = TYPE_PRECISION (n->type); | |
1745 | if (size % BITS_PER_UNIT != 0) | |
3cc272c1 | 1746 | return false; |
698ff107 | 1747 | size /= BITS_PER_UNIT; |
e3ef4162 | 1748 | if (size > 64 / BITS_PER_MARKER) |
ca6cbdca | 1749 | return false; |
698ff107 | 1750 | n->range = size; |
3cc272c1 TP |
1751 | n->n = CMPNOP; |
1752 | ||
e3ef4162 TP |
1753 | if (size < 64 / BITS_PER_MARKER) |
1754 | n->n &= ((uint64_t) 1 << (size * BITS_PER_MARKER)) - 1; | |
3cc272c1 TP |
1755 | |
1756 | return true; | |
1757 | } | |
1758 | ||
73984f84 TP |
1759 | /* Check if STMT might be a byte swap or a nop from a memory source and returns |
1760 | the answer. If so, REF is that memory source and the base of the memory area | |
1761 | accessed and the offset of the access from that base are recorded in N. */ | |
1762 | ||
1763 | bool | |
1764 | find_bswap_or_nop_load (gimple stmt, tree ref, struct symbolic_number *n) | |
1765 | { | |
1766 | /* Leaf node is an array or component ref. Memorize its base and | |
1767 | offset from base to compare to other such leaf node. */ | |
1768 | HOST_WIDE_INT bitsize, bitpos; | |
ef4bddc2 | 1769 | machine_mode mode; |
73984f84 | 1770 | int unsignedp, volatilep; |
3cc272c1 | 1771 | tree offset, base_addr; |
73984f84 TP |
1772 | |
1773 | if (!gimple_assign_load_p (stmt) || gimple_has_volatile_ops (stmt)) | |
1774 | return false; | |
1775 | ||
3cc272c1 TP |
1776 | base_addr = get_inner_reference (ref, &bitsize, &bitpos, &offset, &mode, |
1777 | &unsignedp, &volatilep, false); | |
73984f84 | 1778 | |
3cc272c1 | 1779 | if (TREE_CODE (base_addr) == MEM_REF) |
73984f84 TP |
1780 | { |
1781 | offset_int bit_offset = 0; | |
3cc272c1 | 1782 | tree off = TREE_OPERAND (base_addr, 1); |
73984f84 TP |
1783 | |
1784 | if (!integer_zerop (off)) | |
1785 | { | |
3cc272c1 | 1786 | offset_int boff, coff = mem_ref_offset (base_addr); |
73984f84 TP |
1787 | boff = wi::lshift (coff, LOG2_BITS_PER_UNIT); |
1788 | bit_offset += boff; | |
1789 | } | |
1790 | ||
3cc272c1 | 1791 | base_addr = TREE_OPERAND (base_addr, 0); |
73984f84 TP |
1792 | |
1793 | /* Avoid returning a negative bitpos as this may wreak havoc later. */ | |
1794 | if (wi::neg_p (bit_offset)) | |
1795 | { | |
1796 | offset_int mask = wi::mask <offset_int> (LOG2_BITS_PER_UNIT, false); | |
1797 | offset_int tem = bit_offset.and_not (mask); | |
1798 | /* TEM is the bitpos rounded to BITS_PER_UNIT towards -Inf. | |
1799 | Subtract it to BIT_OFFSET and add it (scaled) to OFFSET. */ | |
1800 | bit_offset -= tem; | |
1801 | tem = wi::arshift (tem, LOG2_BITS_PER_UNIT); | |
3cc272c1 TP |
1802 | if (offset) |
1803 | offset = size_binop (PLUS_EXPR, offset, | |
73984f84 TP |
1804 | wide_int_to_tree (sizetype, tem)); |
1805 | else | |
3cc272c1 | 1806 | offset = wide_int_to_tree (sizetype, tem); |
73984f84 TP |
1807 | } |
1808 | ||
1809 | bitpos += bit_offset.to_shwi (); | |
1810 | } | |
1811 | ||
1812 | if (bitpos % BITS_PER_UNIT) | |
1813 | return false; | |
1814 | if (bitsize % BITS_PER_UNIT) | |
1815 | return false; | |
1816 | ||
2cfa504a TP |
1817 | if (!init_symbolic_number (n, ref)) |
1818 | return false; | |
3cc272c1 TP |
1819 | n->base_addr = base_addr; |
1820 | n->offset = offset; | |
73984f84 TP |
1821 | n->bytepos = bitpos / BITS_PER_UNIT; |
1822 | n->alias_set = reference_alias_ptr_type (ref); | |
1823 | n->vuse = gimple_vuse (stmt); | |
1824 | return true; | |
1825 | } | |
1826 | ||
1827 | /* find_bswap_or_nop_1 invokes itself recursively with N and tries to perform | |
1828 | the operation given by the rhs of STMT on the result. If the operation | |
a31d2741 TP |
1829 | could successfully be executed the function returns a gimple stmt whose |
1830 | rhs's first tree is the expression of the source operand and NULL | |
1831 | otherwise. */ | |
03bd2f1a | 1832 | |
a31d2741 | 1833 | static gimple |
73984f84 | 1834 | find_bswap_or_nop_1 (gimple stmt, struct symbolic_number *n, int limit) |
03bd2f1a AK |
1835 | { |
1836 | enum tree_code code; | |
1837 | tree rhs1, rhs2 = NULL; | |
a31d2741 | 1838 | gimple rhs1_stmt, rhs2_stmt, source_stmt1; |
03bd2f1a AK |
1839 | enum gimple_rhs_class rhs_class; |
1840 | ||
1841 | if (!limit || !is_gimple_assign (stmt)) | |
a31d2741 | 1842 | return NULL; |
03bd2f1a AK |
1843 | |
1844 | rhs1 = gimple_assign_rhs1 (stmt); | |
1845 | ||
73984f84 | 1846 | if (find_bswap_or_nop_load (stmt, rhs1, n)) |
a31d2741 | 1847 | return stmt; |
73984f84 | 1848 | |
03bd2f1a | 1849 | if (TREE_CODE (rhs1) != SSA_NAME) |
a31d2741 | 1850 | return NULL; |
03bd2f1a AK |
1851 | |
1852 | code = gimple_assign_rhs_code (stmt); | |
1853 | rhs_class = gimple_assign_rhs_class (stmt); | |
1854 | rhs1_stmt = SSA_NAME_DEF_STMT (rhs1); | |
1855 | ||
1856 | if (rhs_class == GIMPLE_BINARY_RHS) | |
1857 | rhs2 = gimple_assign_rhs2 (stmt); | |
1858 | ||
1859 | /* Handle unary rhs and binary rhs with integer constants as second | |
1860 | operand. */ | |
1861 | ||
1862 | if (rhs_class == GIMPLE_UNARY_RHS | |
1863 | || (rhs_class == GIMPLE_BINARY_RHS | |
1864 | && TREE_CODE (rhs2) == INTEGER_CST)) | |
1865 | { | |
1866 | if (code != BIT_AND_EXPR | |
1867 | && code != LSHIFT_EXPR | |
1868 | && code != RSHIFT_EXPR | |
1869 | && code != LROTATE_EXPR | |
1870 | && code != RROTATE_EXPR | |
625a9766 | 1871 | && !CONVERT_EXPR_CODE_P (code)) |
a31d2741 | 1872 | return NULL; |
03bd2f1a | 1873 | |
a31d2741 | 1874 | source_stmt1 = find_bswap_or_nop_1 (rhs1_stmt, n, limit - 1); |
03bd2f1a | 1875 | |
73984f84 TP |
1876 | /* If find_bswap_or_nop_1 returned NULL, STMT is a leaf node and |
1877 | we have to initialize the symbolic number. */ | |
a31d2741 | 1878 | if (!source_stmt1) |
03bd2f1a | 1879 | { |
3cc272c1 TP |
1880 | if (gimple_assign_load_p (stmt) |
1881 | || !init_symbolic_number (n, rhs1)) | |
a31d2741 TP |
1882 | return NULL; |
1883 | source_stmt1 = stmt; | |
03bd2f1a AK |
1884 | } |
1885 | ||
1886 | switch (code) | |
1887 | { | |
1888 | case BIT_AND_EXPR: | |
1889 | { | |
698ff107 | 1890 | int i, size = TYPE_PRECISION (n->type) / BITS_PER_UNIT; |
e3ef4162 TP |
1891 | uint64_t val = int_cst_value (rhs2), mask = 0; |
1892 | uint64_t tmp = (1 << BITS_PER_UNIT) - 1; | |
03bd2f1a AK |
1893 | |
1894 | /* Only constants masking full bytes are allowed. */ | |
e3ef4162 TP |
1895 | for (i = 0; i < size; i++, tmp <<= BITS_PER_UNIT) |
1896 | if ((val & tmp) != 0 && (val & tmp) != tmp) | |
a31d2741 | 1897 | return NULL; |
e3ef4162 | 1898 | else if (val & tmp) |
aa29ea0c | 1899 | mask |= (uint64_t) MARKER_MASK << (i * BITS_PER_MARKER); |
03bd2f1a | 1900 | |
e3ef4162 | 1901 | n->n &= mask; |
03bd2f1a AK |
1902 | } |
1903 | break; | |
1904 | case LSHIFT_EXPR: | |
1905 | case RSHIFT_EXPR: | |
1906 | case LROTATE_EXPR: | |
1907 | case RROTATE_EXPR: | |
1908 | if (!do_shift_rotate (code, n, (int)TREE_INT_CST_LOW (rhs2))) | |
a31d2741 | 1909 | return NULL; |
03bd2f1a AK |
1910 | break; |
1911 | CASE_CONVERT: | |
1912 | { | |
aa29ea0c | 1913 | int i, type_size, old_type_size; |
698ff107 | 1914 | tree type; |
03bd2f1a | 1915 | |
698ff107 TP |
1916 | type = gimple_expr_type (stmt); |
1917 | type_size = TYPE_PRECISION (type); | |
03bd2f1a | 1918 | if (type_size % BITS_PER_UNIT != 0) |
a31d2741 | 1919 | return NULL; |
e3ef4162 TP |
1920 | type_size /= BITS_PER_UNIT; |
1921 | if (type_size > 64 / BITS_PER_MARKER) | |
a31d2741 | 1922 | return NULL; |
03bd2f1a | 1923 | |
698ff107 | 1924 | /* Sign extension: result is dependent on the value. */ |
e3ef4162 | 1925 | old_type_size = TYPE_PRECISION (n->type) / BITS_PER_UNIT; |
aa29ea0c TP |
1926 | if (!TYPE_UNSIGNED (n->type) && type_size > old_type_size |
1927 | && HEAD_MARKER (n->n, old_type_size)) | |
1928 | for (i = 0; i < type_size - old_type_size; i++) | |
8ae9ab2b | 1929 | n->n |= (uint64_t) MARKER_BYTE_UNKNOWN |
e4d2f1db | 1930 | << ((type_size - 1 - i) * BITS_PER_MARKER); |
698ff107 | 1931 | |
e3ef4162 | 1932 | if (type_size < 64 / BITS_PER_MARKER) |
03bd2f1a AK |
1933 | { |
1934 | /* If STMT casts to a smaller type mask out the bits not | |
1935 | belonging to the target type. */ | |
e3ef4162 | 1936 | n->n &= ((uint64_t) 1 << (type_size * BITS_PER_MARKER)) - 1; |
03bd2f1a | 1937 | } |
698ff107 | 1938 | n->type = type; |
73984f84 | 1939 | if (!n->base_addr) |
e3ef4162 | 1940 | n->range = type_size; |
03bd2f1a AK |
1941 | } |
1942 | break; | |
1943 | default: | |
a31d2741 | 1944 | return NULL; |
03bd2f1a | 1945 | }; |
a31d2741 | 1946 | return verify_symbolic_number_p (n, stmt) ? source_stmt1 : NULL; |
03bd2f1a AK |
1947 | } |
1948 | ||
1949 | /* Handle binary rhs. */ | |
1950 | ||
1951 | if (rhs_class == GIMPLE_BINARY_RHS) | |
1952 | { | |
698ff107 | 1953 | int i, size; |
03bd2f1a | 1954 | struct symbolic_number n1, n2; |
a9243bfc | 1955 | uint64_t mask; |
a31d2741 | 1956 | gimple source_stmt2; |
03bd2f1a AK |
1957 | |
1958 | if (code != BIT_IOR_EXPR) | |
a31d2741 | 1959 | return NULL; |
03bd2f1a AK |
1960 | |
1961 | if (TREE_CODE (rhs2) != SSA_NAME) | |
a31d2741 | 1962 | return NULL; |
03bd2f1a AK |
1963 | |
1964 | rhs2_stmt = SSA_NAME_DEF_STMT (rhs2); | |
1965 | ||
1966 | switch (code) | |
1967 | { | |
1968 | case BIT_IOR_EXPR: | |
a31d2741 | 1969 | source_stmt1 = find_bswap_or_nop_1 (rhs1_stmt, &n1, limit - 1); |
03bd2f1a | 1970 | |
a31d2741 TP |
1971 | if (!source_stmt1) |
1972 | return NULL; | |
03bd2f1a | 1973 | |
a31d2741 | 1974 | source_stmt2 = find_bswap_or_nop_1 (rhs2_stmt, &n2, limit - 1); |
73984f84 | 1975 | |
a31d2741 TP |
1976 | if (!source_stmt2) |
1977 | return NULL; | |
3cc272c1 | 1978 | |
698ff107 | 1979 | if (TYPE_PRECISION (n1.type) != TYPE_PRECISION (n2.type)) |
a31d2741 | 1980 | return NULL; |
03bd2f1a | 1981 | |
73984f84 TP |
1982 | if (!n1.vuse != !n2.vuse || |
1983 | (n1.vuse && !operand_equal_p (n1.vuse, n2.vuse, 0))) | |
a31d2741 | 1984 | return NULL; |
03bd2f1a | 1985 | |
a31d2741 TP |
1986 | if (gimple_assign_rhs1 (source_stmt1) |
1987 | != gimple_assign_rhs1 (source_stmt2)) | |
73984f84 | 1988 | { |
aa29ea0c | 1989 | int64_t inc; |
73984f84 TP |
1990 | HOST_WIDE_INT off_sub; |
1991 | struct symbolic_number *n_ptr; | |
1992 | ||
1993 | if (!n1.base_addr || !n2.base_addr | |
1994 | || !operand_equal_p (n1.base_addr, n2.base_addr, 0)) | |
a31d2741 | 1995 | return NULL; |
73984f84 TP |
1996 | if (!n1.offset != !n2.offset || |
1997 | (n1.offset && !operand_equal_p (n1.offset, n2.offset, 0))) | |
a31d2741 | 1998 | return NULL; |
73984f84 TP |
1999 | |
2000 | /* We swap n1 with n2 to have n1 < n2. */ | |
2001 | if (n2.bytepos < n1.bytepos) | |
2002 | { | |
2003 | struct symbolic_number tmpn; | |
2004 | ||
2005 | tmpn = n2; | |
2006 | n2 = n1; | |
2007 | n1 = tmpn; | |
a31d2741 | 2008 | source_stmt1 = source_stmt2; |
73984f84 TP |
2009 | } |
2010 | ||
2011 | off_sub = n2.bytepos - n1.bytepos; | |
2012 | ||
e3ef4162 TP |
2013 | /* Check that the range of memory covered can be represented by |
2014 | a symbolic number. */ | |
2015 | if (off_sub + n2.range > 64 / BITS_PER_MARKER) | |
a31d2741 | 2016 | return NULL; |
73984f84 TP |
2017 | n->range = n2.range + off_sub; |
2018 | ||
2019 | /* Reinterpret byte marks in symbolic number holding the value of | |
58126368 | 2020 | bigger weight according to target endianness. */ |
73984f84 | 2021 | inc = BYTES_BIG_ENDIAN ? off_sub + n2.range - n1.range : off_sub; |
e3ef4162 | 2022 | size = TYPE_PRECISION (n1.type) / BITS_PER_UNIT; |
73984f84 TP |
2023 | if (BYTES_BIG_ENDIAN) |
2024 | n_ptr = &n1; | |
2025 | else | |
2026 | n_ptr = &n2; | |
aa29ea0c | 2027 | for (i = 0; i < size; i++, inc <<= BITS_PER_MARKER) |
73984f84 | 2028 | { |
aa29ea0c TP |
2029 | unsigned marker = |
2030 | (n_ptr->n >> (i * BITS_PER_MARKER)) & MARKER_MASK; | |
2031 | if (marker && marker != MARKER_BYTE_UNKNOWN) | |
73984f84 TP |
2032 | n_ptr->n += inc; |
2033 | } | |
2034 | } | |
2035 | else | |
2036 | n->range = n1.range; | |
2037 | ||
2038 | if (!n1.alias_set | |
2039 | || alias_ptr_types_compatible_p (n1.alias_set, n2.alias_set)) | |
2040 | n->alias_set = n1.alias_set; | |
2041 | else | |
2042 | n->alias_set = ptr_type_node; | |
2043 | n->vuse = n1.vuse; | |
2044 | n->base_addr = n1.base_addr; | |
2045 | n->offset = n1.offset; | |
2046 | n->bytepos = n1.bytepos; | |
698ff107 TP |
2047 | n->type = n1.type; |
2048 | size = TYPE_PRECISION (n->type) / BITS_PER_UNIT; | |
aa29ea0c TP |
2049 | for (i = 0, mask = MARKER_MASK; i < size; |
2050 | i++, mask <<= BITS_PER_MARKER) | |
882a5fbe | 2051 | { |
a9243bfc | 2052 | uint64_t masked1, masked2; |
882a5fbe TP |
2053 | |
2054 | masked1 = n1.n & mask; | |
2055 | masked2 = n2.n & mask; | |
2056 | if (masked1 && masked2 && masked1 != masked2) | |
a31d2741 | 2057 | return NULL; |
882a5fbe | 2058 | } |
03bd2f1a AK |
2059 | n->n = n1.n | n2.n; |
2060 | ||
2061 | if (!verify_symbolic_number_p (n, stmt)) | |
a31d2741 | 2062 | return NULL; |
03bd2f1a AK |
2063 | |
2064 | break; | |
2065 | default: | |
a31d2741 | 2066 | return NULL; |
03bd2f1a | 2067 | } |
a31d2741 | 2068 | return source_stmt1; |
03bd2f1a | 2069 | } |
a31d2741 | 2070 | return NULL; |
03bd2f1a AK |
2071 | } |
2072 | ||
73984f84 TP |
2073 | /* Check if STMT completes a bswap implementation or a read in a given |
2074 | endianness consisting of ORs, SHIFTs and ANDs and sets *BSWAP | |
2075 | accordingly. It also sets N to represent the kind of operations | |
2076 | performed: size of the resulting expression and whether it works on | |
2077 | a memory source, and if so alias-set and vuse. At last, the | |
a31d2741 TP |
2078 | function returns a stmt whose rhs's first tree is the source |
2079 | expression. */ | |
03bd2f1a | 2080 | |
a31d2741 | 2081 | static gimple |
73984f84 | 2082 | find_bswap_or_nop (gimple stmt, struct symbolic_number *n, bool *bswap) |
03bd2f1a | 2083 | { |
73984f84 TP |
2084 | /* The number which the find_bswap_or_nop_1 result should match in order |
2085 | to have a full byte swap. The number is shifted to the right | |
2086 | according to the size of the symbolic number before using it. */ | |
a9243bfc RB |
2087 | uint64_t cmpxchg = CMPXCHG; |
2088 | uint64_t cmpnop = CMPNOP; | |
73984f84 | 2089 | |
a31d2741 | 2090 | gimple source_stmt; |
5da49a9d | 2091 | int limit; |
03bd2f1a | 2092 | |
fef015a8 | 2093 | /* The last parameter determines the depth search limit. It usually |
73984f84 TP |
2094 | correlates directly to the number n of bytes to be touched. We |
2095 | increase that number by log2(n) + 1 here in order to also | |
2096 | cover signed -> unsigned conversions of the src operand as can be seen | |
5da49a9d KT |
2097 | in libgcc, and for initial shift/and operation of the src operand. */ |
2098 | limit = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (gimple_expr_type (stmt))); | |
2099 | limit += 1 + (int) ceil_log2 ((unsigned HOST_WIDE_INT) limit); | |
a31d2741 | 2100 | source_stmt = find_bswap_or_nop_1 (stmt, n, limit); |
03bd2f1a | 2101 | |
a31d2741 TP |
2102 | if (!source_stmt) |
2103 | return NULL; | |
03bd2f1a | 2104 | |
73984f84 TP |
2105 | /* Find real size of result (highest non zero byte). */ |
2106 | if (n->base_addr) | |
03bd2f1a | 2107 | { |
73984f84 | 2108 | int rsize; |
a9243bfc | 2109 | uint64_t tmpn; |
03bd2f1a | 2110 | |
e3ef4162 | 2111 | for (tmpn = n->n, rsize = 0; tmpn; tmpn >>= BITS_PER_MARKER, rsize++); |
73984f84 | 2112 | n->range = rsize; |
03bd2f1a AK |
2113 | } |
2114 | ||
73984f84 | 2115 | /* Zero out the extra bits of N and CMP*. */ |
e3ef4162 | 2116 | if (n->range < (int) sizeof (int64_t)) |
73984f84 | 2117 | { |
a9243bfc | 2118 | uint64_t mask; |
73984f84 | 2119 | |
e3ef4162 TP |
2120 | mask = ((uint64_t) 1 << (n->range * BITS_PER_MARKER)) - 1; |
2121 | cmpxchg >>= (64 / BITS_PER_MARKER - n->range) * BITS_PER_MARKER; | |
73984f84 TP |
2122 | cmpnop &= mask; |
2123 | } | |
2124 | ||
2125 | /* A complete byte swap should make the symbolic number to start with | |
2126 | the largest digit in the highest order byte. Unchanged symbolic | |
58126368 | 2127 | number indicates a read with same endianness as target architecture. */ |
73984f84 TP |
2128 | if (n->n == cmpnop) |
2129 | *bswap = false; | |
2130 | else if (n->n == cmpxchg) | |
2131 | *bswap = true; | |
2132 | else | |
a31d2741 | 2133 | return NULL; |
73984f84 TP |
2134 | |
2135 | /* Useless bit manipulation performed by code. */ | |
2136 | if (!n->base_addr && n->n == cmpnop) | |
a31d2741 | 2137 | return NULL; |
03bd2f1a | 2138 | |
73984f84 | 2139 | n->range *= BITS_PER_UNIT; |
a31d2741 | 2140 | return source_stmt; |
03bd2f1a AK |
2141 | } |
2142 | ||
be55bfe6 TS |
2143 | namespace { |
2144 | ||
2145 | const pass_data pass_data_optimize_bswap = | |
2146 | { | |
2147 | GIMPLE_PASS, /* type */ | |
2148 | "bswap", /* name */ | |
2149 | OPTGROUP_NONE, /* optinfo_flags */ | |
be55bfe6 TS |
2150 | TV_NONE, /* tv_id */ |
2151 | PROP_ssa, /* properties_required */ | |
2152 | 0, /* properties_provided */ | |
2153 | 0, /* properties_destroyed */ | |
2154 | 0, /* todo_flags_start */ | |
2155 | 0, /* todo_flags_finish */ | |
2156 | }; | |
2157 | ||
2158 | class pass_optimize_bswap : public gimple_opt_pass | |
2159 | { | |
2160 | public: | |
2161 | pass_optimize_bswap (gcc::context *ctxt) | |
2162 | : gimple_opt_pass (pass_data_optimize_bswap, ctxt) | |
2163 | {} | |
2164 | ||
2165 | /* opt_pass methods: */ | |
2166 | virtual bool gate (function *) | |
2167 | { | |
2168 | return flag_expensive_optimizations && optimize; | |
2169 | } | |
2170 | ||
2171 | virtual unsigned int execute (function *); | |
2172 | ||
2173 | }; // class pass_optimize_bswap | |
2174 | ||
a31d2741 TP |
2175 | /* Perform the bswap optimization: replace the statement CUR_STMT at |
2176 | GSI with a load of type, VUSE and set-alias as described by N if a | |
2177 | memory source is involved (N->base_addr is non null), followed by | |
2178 | the builtin bswap invocation in FNDECL if BSWAP is true. SRC_STMT | |
2179 | gives where should the replacement be made. It also gives the | |
2180 | source on which CUR_STMT is operating via its rhs's first tree nad | |
2181 | N->range gives the size of the expression involved for maintaining | |
2182 | some statistics. */ | |
73984f84 TP |
2183 | |
2184 | static bool | |
a31d2741 TP |
2185 | bswap_replace (gimple cur_stmt, gimple_stmt_iterator gsi, gimple src_stmt, |
2186 | tree fndecl, tree bswap_type, tree load_type, | |
2187 | struct symbolic_number *n, bool bswap) | |
73984f84 | 2188 | { |
a31d2741 | 2189 | tree src, tmp, tgt; |
c6e3a931 | 2190 | gimple bswap_stmt; |
73984f84 | 2191 | |
a31d2741 TP |
2192 | src = gimple_assign_rhs1 (src_stmt); |
2193 | tgt = gimple_assign_lhs (cur_stmt); | |
73984f84 TP |
2194 | |
2195 | /* Need to load the value from memory first. */ | |
2196 | if (n->base_addr) | |
2197 | { | |
a31d2741 | 2198 | gimple_stmt_iterator gsi_ins = gsi_for_stmt (src_stmt); |
73984f84 TP |
2199 | tree addr_expr, addr_tmp, val_expr, val_tmp; |
2200 | tree load_offset_ptr, aligned_load_type; | |
2201 | gimple addr_stmt, load_stmt; | |
2202 | unsigned align; | |
2203 | ||
2204 | align = get_object_alignment (src); | |
3fd269db RB |
2205 | if (bswap |
2206 | && align < GET_MODE_ALIGNMENT (TYPE_MODE (load_type)) | |
2207 | && SLOW_UNALIGNED_ACCESS (TYPE_MODE (load_type), align)) | |
73984f84 TP |
2208 | return false; |
2209 | ||
a31d2741 TP |
2210 | gsi_move_before (&gsi, &gsi_ins); |
2211 | gsi = gsi_for_stmt (cur_stmt); | |
2212 | ||
73984f84 TP |
2213 | /* Compute address to load from and cast according to the size |
2214 | of the load. */ | |
2215 | addr_expr = build_fold_addr_expr (unshare_expr (src)); | |
2216 | if (is_gimple_min_invariant (addr_expr)) | |
2217 | addr_tmp = addr_expr; | |
2218 | else | |
2219 | { | |
2220 | addr_tmp = make_temp_ssa_name (TREE_TYPE (addr_expr), NULL, | |
2221 | "load_src"); | |
2222 | addr_stmt = gimple_build_assign (addr_tmp, addr_expr); | |
a31d2741 | 2223 | gsi_insert_before (&gsi, addr_stmt, GSI_SAME_STMT); |
73984f84 TP |
2224 | } |
2225 | ||
2226 | /* Perform the load. */ | |
2227 | aligned_load_type = load_type; | |
2228 | if (align < TYPE_ALIGN (load_type)) | |
2229 | aligned_load_type = build_aligned_type (load_type, align); | |
2230 | load_offset_ptr = build_int_cst (n->alias_set, 0); | |
2231 | val_expr = fold_build2 (MEM_REF, aligned_load_type, addr_tmp, | |
2232 | load_offset_ptr); | |
2233 | ||
2234 | if (!bswap) | |
2235 | { | |
2236 | if (n->range == 16) | |
2237 | nop_stats.found_16bit++; | |
2238 | else if (n->range == 32) | |
2239 | nop_stats.found_32bit++; | |
2240 | else | |
2241 | { | |
2242 | gcc_assert (n->range == 64); | |
2243 | nop_stats.found_64bit++; | |
2244 | } | |
2245 | ||
2246 | /* Convert the result of load if necessary. */ | |
2247 | if (!useless_type_conversion_p (TREE_TYPE (tgt), load_type)) | |
2248 | { | |
2249 | val_tmp = make_temp_ssa_name (aligned_load_type, NULL, | |
2250 | "load_dst"); | |
2251 | load_stmt = gimple_build_assign (val_tmp, val_expr); | |
2252 | gimple_set_vuse (load_stmt, n->vuse); | |
a31d2741 TP |
2253 | gsi_insert_before (&gsi, load_stmt, GSI_SAME_STMT); |
2254 | gimple_assign_set_rhs_with_ops_1 (&gsi, NOP_EXPR, val_tmp, | |
73984f84 TP |
2255 | NULL_TREE, NULL_TREE); |
2256 | } | |
2257 | else | |
a31d2741 TP |
2258 | { |
2259 | gimple_assign_set_rhs_with_ops_1 (&gsi, MEM_REF, val_expr, | |
2260 | NULL_TREE, NULL_TREE); | |
2261 | gimple_set_vuse (cur_stmt, n->vuse); | |
2262 | } | |
2263 | update_stmt (cur_stmt); | |
73984f84 TP |
2264 | |
2265 | if (dump_file) | |
2266 | { | |
2267 | fprintf (dump_file, | |
58126368 | 2268 | "%d bit load in target endianness found at: ", |
73984f84 | 2269 | (int)n->range); |
a31d2741 | 2270 | print_gimple_stmt (dump_file, cur_stmt, 0, 0); |
73984f84 TP |
2271 | } |
2272 | return true; | |
2273 | } | |
2274 | else | |
2275 | { | |
2276 | val_tmp = make_temp_ssa_name (aligned_load_type, NULL, "load_dst"); | |
2277 | load_stmt = gimple_build_assign (val_tmp, val_expr); | |
2278 | gimple_set_vuse (load_stmt, n->vuse); | |
a31d2741 | 2279 | gsi_insert_before (&gsi, load_stmt, GSI_SAME_STMT); |
73984f84 TP |
2280 | } |
2281 | src = val_tmp; | |
2282 | } | |
2283 | ||
2284 | if (n->range == 16) | |
2285 | bswap_stats.found_16bit++; | |
2286 | else if (n->range == 32) | |
2287 | bswap_stats.found_32bit++; | |
2288 | else | |
2289 | { | |
2290 | gcc_assert (n->range == 64); | |
2291 | bswap_stats.found_64bit++; | |
2292 | } | |
2293 | ||
2294 | tmp = src; | |
2295 | ||
c6e3a931 TP |
2296 | /* Canonical form for 16 bit bswap is a rotate expression. */ |
2297 | if (bswap && n->range == 16) | |
73984f84 | 2298 | { |
c6e3a931 TP |
2299 | tree count = build_int_cst (NULL, BITS_PER_UNIT); |
2300 | bswap_type = TREE_TYPE (src); | |
2301 | src = fold_build2 (LROTATE_EXPR, bswap_type, src, count); | |
2302 | bswap_stmt = gimple_build_assign (NULL, src); | |
73984f84 | 2303 | } |
c6e3a931 TP |
2304 | else |
2305 | { | |
2306 | /* Convert the src expression if necessary. */ | |
2307 | if (!useless_type_conversion_p (TREE_TYPE (tmp), bswap_type)) | |
2308 | { | |
2309 | gimple convert_stmt; | |
2310 | tmp = make_temp_ssa_name (bswap_type, NULL, "bswapsrc"); | |
2311 | convert_stmt = gimple_build_assign_with_ops (NOP_EXPR, tmp, src, | |
2312 | NULL); | |
2313 | gsi_insert_before (&gsi, convert_stmt, GSI_SAME_STMT); | |
2314 | } | |
73984f84 | 2315 | |
c6e3a931 TP |
2316 | bswap_stmt = gimple_build_call (fndecl, 1, tmp); |
2317 | } | |
73984f84 TP |
2318 | |
2319 | tmp = tgt; | |
2320 | ||
2321 | /* Convert the result if necessary. */ | |
2322 | if (!useless_type_conversion_p (TREE_TYPE (tgt), bswap_type)) | |
2323 | { | |
2324 | gimple convert_stmt; | |
2325 | tmp = make_temp_ssa_name (bswap_type, NULL, "bswapdst"); | |
2326 | convert_stmt = gimple_build_assign_with_ops (NOP_EXPR, tgt, tmp, NULL); | |
a31d2741 | 2327 | gsi_insert_after (&gsi, convert_stmt, GSI_SAME_STMT); |
73984f84 TP |
2328 | } |
2329 | ||
c6e3a931 | 2330 | gimple_set_lhs (bswap_stmt, tmp); |
73984f84 TP |
2331 | |
2332 | if (dump_file) | |
2333 | { | |
2334 | fprintf (dump_file, "%d bit bswap implementation found at: ", | |
2335 | (int)n->range); | |
a31d2741 | 2336 | print_gimple_stmt (dump_file, cur_stmt, 0, 0); |
73984f84 TP |
2337 | } |
2338 | ||
c6e3a931 | 2339 | gsi_insert_after (&gsi, bswap_stmt, GSI_SAME_STMT); |
a31d2741 | 2340 | gsi_remove (&gsi, true); |
73984f84 TP |
2341 | return true; |
2342 | } | |
2343 | ||
2344 | /* Find manual byte swap implementations as well as load in a given | |
2345 | endianness. Byte swaps are turned into a bswap builtin invokation | |
2346 | while endian loads are converted to bswap builtin invokation or | |
58126368 | 2347 | simple load according to the target endianness. */ |
73984f84 | 2348 | |
be55bfe6 TS |
2349 | unsigned int |
2350 | pass_optimize_bswap::execute (function *fun) | |
03bd2f1a AK |
2351 | { |
2352 | basic_block bb; | |
1df855ce | 2353 | bool bswap16_p, bswap32_p, bswap64_p; |
03bd2f1a | 2354 | bool changed = false; |
1df855ce | 2355 | tree bswap16_type = NULL_TREE, bswap32_type = NULL_TREE, bswap64_type = NULL_TREE; |
03bd2f1a AK |
2356 | |
2357 | if (BITS_PER_UNIT != 8) | |
2358 | return 0; | |
2359 | ||
1df855ce CL |
2360 | bswap16_p = (builtin_decl_explicit_p (BUILT_IN_BSWAP16) |
2361 | && optab_handler (bswap_optab, HImode) != CODE_FOR_nothing); | |
e79983f4 | 2362 | bswap32_p = (builtin_decl_explicit_p (BUILT_IN_BSWAP32) |
947131ba | 2363 | && optab_handler (bswap_optab, SImode) != CODE_FOR_nothing); |
e79983f4 | 2364 | bswap64_p = (builtin_decl_explicit_p (BUILT_IN_BSWAP64) |
947131ba | 2365 | && (optab_handler (bswap_optab, DImode) != CODE_FOR_nothing |
ba1ee228 | 2366 | || (bswap32_p && word_mode == SImode))); |
03bd2f1a | 2367 | |
fb6234e0 AK |
2368 | /* Determine the argument type of the builtins. The code later on |
2369 | assumes that the return and argument type are the same. */ | |
1df855ce CL |
2370 | if (bswap16_p) |
2371 | { | |
2372 | tree fndecl = builtin_decl_explicit (BUILT_IN_BSWAP16); | |
2373 | bswap16_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl))); | |
2374 | } | |
2375 | ||
fb6234e0 AK |
2376 | if (bswap32_p) |
2377 | { | |
e79983f4 | 2378 | tree fndecl = builtin_decl_explicit (BUILT_IN_BSWAP32); |
fb6234e0 AK |
2379 | bswap32_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl))); |
2380 | } | |
2381 | ||
2382 | if (bswap64_p) | |
2383 | { | |
e79983f4 | 2384 | tree fndecl = builtin_decl_explicit (BUILT_IN_BSWAP64); |
fb6234e0 AK |
2385 | bswap64_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl))); |
2386 | } | |
2387 | ||
73984f84 | 2388 | memset (&nop_stats, 0, sizeof (nop_stats)); |
4da3b811 NF |
2389 | memset (&bswap_stats, 0, sizeof (bswap_stats)); |
2390 | ||
be55bfe6 | 2391 | FOR_EACH_BB_FN (bb, fun) |
03bd2f1a AK |
2392 | { |
2393 | gimple_stmt_iterator gsi; | |
2394 | ||
72a32729 KT |
2395 | /* We do a reverse scan for bswap patterns to make sure we get the |
2396 | widest match. As bswap pattern matching doesn't handle | |
2397 | previously inserted smaller bswap replacements as sub- | |
2398 | patterns, the wider variant wouldn't be detected. */ | |
2399 | for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi)) | |
03bd2f1a | 2400 | { |
a31d2741 TP |
2401 | gimple src_stmt, cur_stmt = gsi_stmt (gsi); |
2402 | tree fndecl = NULL_TREE, bswap_type = NULL_TREE, load_type; | |
c6e3a931 | 2403 | enum tree_code code; |
73984f84 TP |
2404 | struct symbolic_number n; |
2405 | bool bswap; | |
03bd2f1a | 2406 | |
c6e3a931 | 2407 | if (!is_gimple_assign (cur_stmt)) |
03bd2f1a AK |
2408 | continue; |
2409 | ||
c6e3a931 TP |
2410 | code = gimple_assign_rhs_code (cur_stmt); |
2411 | switch (code) | |
2412 | { | |
2413 | case LROTATE_EXPR: | |
2414 | case RROTATE_EXPR: | |
2415 | if (!tree_fits_uhwi_p (gimple_assign_rhs2 (cur_stmt)) | |
2416 | || tree_to_uhwi (gimple_assign_rhs2 (cur_stmt)) | |
2417 | % BITS_PER_UNIT) | |
2418 | continue; | |
2419 | /* Fall through. */ | |
2420 | case BIT_IOR_EXPR: | |
2421 | break; | |
2422 | default: | |
2423 | continue; | |
2424 | } | |
2425 | ||
a31d2741 | 2426 | src_stmt = find_bswap_or_nop (cur_stmt, &n, &bswap); |
73984f84 | 2427 | |
a31d2741 | 2428 | if (!src_stmt) |
73984f84 | 2429 | continue; |
03bd2f1a | 2430 | |
73984f84 | 2431 | switch (n.range) |
03bd2f1a | 2432 | { |
1df855ce | 2433 | case 16: |
73984f84 | 2434 | load_type = uint16_type_node; |
1df855ce CL |
2435 | if (bswap16_p) |
2436 | { | |
2437 | fndecl = builtin_decl_explicit (BUILT_IN_BSWAP16); | |
2438 | bswap_type = bswap16_type; | |
2439 | } | |
2440 | break; | |
03bd2f1a | 2441 | case 32: |
73984f84 | 2442 | load_type = uint32_type_node; |
03bd2f1a | 2443 | if (bswap32_p) |
fb6234e0 | 2444 | { |
e79983f4 | 2445 | fndecl = builtin_decl_explicit (BUILT_IN_BSWAP32); |
fb6234e0 AK |
2446 | bswap_type = bswap32_type; |
2447 | } | |
03bd2f1a AK |
2448 | break; |
2449 | case 64: | |
73984f84 | 2450 | load_type = uint64_type_node; |
03bd2f1a | 2451 | if (bswap64_p) |
fb6234e0 | 2452 | { |
e79983f4 | 2453 | fndecl = builtin_decl_explicit (BUILT_IN_BSWAP64); |
fb6234e0 AK |
2454 | bswap_type = bswap64_type; |
2455 | } | |
03bd2f1a AK |
2456 | break; |
2457 | default: | |
2458 | continue; | |
2459 | } | |
2460 | ||
73984f84 | 2461 | if (bswap && !fndecl) |
03bd2f1a AK |
2462 | continue; |
2463 | ||
a31d2741 TP |
2464 | if (bswap_replace (cur_stmt, gsi, src_stmt, fndecl, bswap_type, |
2465 | load_type, &n, bswap)) | |
73984f84 | 2466 | changed = true; |
03bd2f1a AK |
2467 | } |
2468 | } | |
2469 | ||
73984f84 TP |
2470 | statistics_counter_event (fun, "16-bit nop implementations found", |
2471 | nop_stats.found_16bit); | |
2472 | statistics_counter_event (fun, "32-bit nop implementations found", | |
2473 | nop_stats.found_32bit); | |
2474 | statistics_counter_event (fun, "64-bit nop implementations found", | |
2475 | nop_stats.found_64bit); | |
be55bfe6 | 2476 | statistics_counter_event (fun, "16-bit bswap implementations found", |
1df855ce | 2477 | bswap_stats.found_16bit); |
be55bfe6 | 2478 | statistics_counter_event (fun, "32-bit bswap implementations found", |
4da3b811 | 2479 | bswap_stats.found_32bit); |
be55bfe6 | 2480 | statistics_counter_event (fun, "64-bit bswap implementations found", |
4da3b811 NF |
2481 | bswap_stats.found_64bit); |
2482 | ||
3bea341f | 2483 | return (changed ? TODO_update_ssa : 0); |
03bd2f1a AK |
2484 | } |
2485 | ||
27a4cd48 DM |
2486 | } // anon namespace |
2487 | ||
2488 | gimple_opt_pass * | |
2489 | make_pass_optimize_bswap (gcc::context *ctxt) | |
2490 | { | |
2491 | return new pass_optimize_bswap (ctxt); | |
2492 | } | |
2493 | ||
7ab6a828 RE |
2494 | /* Return true if stmt is a type conversion operation that can be stripped |
2495 | when used in a widening multiply operation. */ | |
2496 | static bool | |
2497 | widening_mult_conversion_strippable_p (tree result_type, gimple stmt) | |
2498 | { | |
2499 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); | |
2500 | ||
2501 | if (TREE_CODE (result_type) == INTEGER_TYPE) | |
2502 | { | |
2503 | tree op_type; | |
2504 | tree inner_op_type; | |
2505 | ||
2506 | if (!CONVERT_EXPR_CODE_P (rhs_code)) | |
2507 | return false; | |
2508 | ||
2509 | op_type = TREE_TYPE (gimple_assign_lhs (stmt)); | |
2510 | ||
2511 | /* If the type of OP has the same precision as the result, then | |
2512 | we can strip this conversion. The multiply operation will be | |
2513 | selected to create the correct extension as a by-product. */ | |
2514 | if (TYPE_PRECISION (result_type) == TYPE_PRECISION (op_type)) | |
2515 | return true; | |
2516 | ||
2517 | /* We can also strip a conversion if it preserves the signed-ness of | |
2518 | the operation and doesn't narrow the range. */ | |
2519 | inner_op_type = TREE_TYPE (gimple_assign_rhs1 (stmt)); | |
2520 | ||
e919e5bf RE |
2521 | /* If the inner-most type is unsigned, then we can strip any |
2522 | intermediate widening operation. If it's signed, then the | |
2523 | intermediate widening operation must also be signed. */ | |
2524 | if ((TYPE_UNSIGNED (inner_op_type) | |
2525 | || TYPE_UNSIGNED (op_type) == TYPE_UNSIGNED (inner_op_type)) | |
7ab6a828 RE |
2526 | && TYPE_PRECISION (op_type) > TYPE_PRECISION (inner_op_type)) |
2527 | return true; | |
2528 | ||
2529 | return false; | |
2530 | } | |
2531 | ||
2532 | return rhs_code == FIXED_CONVERT_EXPR; | |
2533 | } | |
2534 | ||
26a855d7 AS |
2535 | /* Return true if RHS is a suitable operand for a widening multiplication, |
2536 | assuming a target type of TYPE. | |
1a39adae RS |
2537 | There are two cases: |
2538 | ||
5dfe80ba AS |
2539 | - RHS makes some value at least twice as wide. Store that value |
2540 | in *NEW_RHS_OUT if so, and store its type in *TYPE_OUT. | |
1a39adae RS |
2541 | |
2542 | - RHS is an integer constant. Store that value in *NEW_RHS_OUT if so, | |
2543 | but leave *TYPE_OUT untouched. */ | |
0354c0c7 BS |
2544 | |
2545 | static bool | |
26a855d7 AS |
2546 | is_widening_mult_rhs_p (tree type, tree rhs, tree *type_out, |
2547 | tree *new_rhs_out) | |
1a39adae RS |
2548 | { |
2549 | gimple stmt; | |
26a855d7 | 2550 | tree type1, rhs1; |
1a39adae RS |
2551 | |
2552 | if (TREE_CODE (rhs) == SSA_NAME) | |
2553 | { | |
1a39adae | 2554 | stmt = SSA_NAME_DEF_STMT (rhs); |
26a855d7 AS |
2555 | if (is_gimple_assign (stmt)) |
2556 | { | |
7ab6a828 | 2557 | if (! widening_mult_conversion_strippable_p (type, stmt)) |
26a855d7 AS |
2558 | rhs1 = rhs; |
2559 | else | |
a6f969f4 AS |
2560 | { |
2561 | rhs1 = gimple_assign_rhs1 (stmt); | |
2562 | ||
2563 | if (TREE_CODE (rhs1) == INTEGER_CST) | |
2564 | { | |
2565 | *new_rhs_out = rhs1; | |
2566 | *type_out = NULL; | |
2567 | return true; | |
2568 | } | |
2569 | } | |
26a855d7 AS |
2570 | } |
2571 | else | |
2572 | rhs1 = rhs; | |
1a39adae | 2573 | |
1a39adae | 2574 | type1 = TREE_TYPE (rhs1); |
26a855d7 | 2575 | |
1a39adae | 2576 | if (TREE_CODE (type1) != TREE_CODE (type) |
5dfe80ba | 2577 | || TYPE_PRECISION (type1) * 2 > TYPE_PRECISION (type)) |
1a39adae RS |
2578 | return false; |
2579 | ||
2580 | *new_rhs_out = rhs1; | |
2581 | *type_out = type1; | |
2582 | return true; | |
2583 | } | |
2584 | ||
2585 | if (TREE_CODE (rhs) == INTEGER_CST) | |
2586 | { | |
2587 | *new_rhs_out = rhs; | |
2588 | *type_out = NULL; | |
2589 | return true; | |
2590 | } | |
2591 | ||
2592 | return false; | |
2593 | } | |
2594 | ||
26a855d7 AS |
2595 | /* Return true if STMT performs a widening multiplication, assuming the |
2596 | output type is TYPE. If so, store the unwidened types of the operands | |
2597 | in *TYPE1_OUT and *TYPE2_OUT respectively. Also fill *RHS1_OUT and | |
2598 | *RHS2_OUT such that converting those operands to types *TYPE1_OUT | |
2599 | and *TYPE2_OUT would give the operands of the multiplication. */ | |
1a39adae RS |
2600 | |
2601 | static bool | |
3d71881d | 2602 | is_widening_mult_p (gimple stmt, |
1a39adae RS |
2603 | tree *type1_out, tree *rhs1_out, |
2604 | tree *type2_out, tree *rhs2_out) | |
0354c0c7 | 2605 | { |
3d71881d AS |
2606 | tree type = TREE_TYPE (gimple_assign_lhs (stmt)); |
2607 | ||
1a39adae RS |
2608 | if (TREE_CODE (type) != INTEGER_TYPE |
2609 | && TREE_CODE (type) != FIXED_POINT_TYPE) | |
2610 | return false; | |
0354c0c7 | 2611 | |
26a855d7 AS |
2612 | if (!is_widening_mult_rhs_p (type, gimple_assign_rhs1 (stmt), type1_out, |
2613 | rhs1_out)) | |
0354c0c7 BS |
2614 | return false; |
2615 | ||
26a855d7 AS |
2616 | if (!is_widening_mult_rhs_p (type, gimple_assign_rhs2 (stmt), type2_out, |
2617 | rhs2_out)) | |
1a39adae | 2618 | return false; |
0354c0c7 | 2619 | |
1a39adae | 2620 | if (*type1_out == NULL) |
0354c0c7 | 2621 | { |
1a39adae | 2622 | if (*type2_out == NULL || !int_fits_type_p (*rhs1_out, *type2_out)) |
0354c0c7 | 2623 | return false; |
1a39adae | 2624 | *type1_out = *type2_out; |
0354c0c7 | 2625 | } |
0354c0c7 | 2626 | |
1a39adae | 2627 | if (*type2_out == NULL) |
0354c0c7 | 2628 | { |
1a39adae | 2629 | if (!int_fits_type_p (*rhs2_out, *type1_out)) |
0354c0c7 | 2630 | return false; |
1a39adae | 2631 | *type2_out = *type1_out; |
0354c0c7 | 2632 | } |
0354c0c7 | 2633 | |
ff63d754 AS |
2634 | /* Ensure that the larger of the two operands comes first. */ |
2635 | if (TYPE_PRECISION (*type1_out) < TYPE_PRECISION (*type2_out)) | |
2636 | { | |
2637 | tree tmp; | |
2638 | tmp = *type1_out; | |
2639 | *type1_out = *type2_out; | |
2640 | *type2_out = tmp; | |
2641 | tmp = *rhs1_out; | |
2642 | *rhs1_out = *rhs2_out; | |
2643 | *rhs2_out = tmp; | |
2644 | } | |
5dfe80ba | 2645 | |
1a39adae RS |
2646 | return true; |
2647 | } | |
0354c0c7 | 2648 | |
1a39adae RS |
2649 | /* Process a single gimple statement STMT, which has a MULT_EXPR as |
2650 | its rhs, and try to convert it into a WIDEN_MULT_EXPR. The return | |
2651 | value is true iff we converted the statement. */ | |
2652 | ||
2653 | static bool | |
5dfe80ba | 2654 | convert_mult_to_widen (gimple stmt, gimple_stmt_iterator *gsi) |
1a39adae | 2655 | { |
83d5977e | 2656 | tree lhs, rhs1, rhs2, type, type1, type2; |
1a39adae | 2657 | enum insn_code handler; |
ef4bddc2 | 2658 | machine_mode to_mode, from_mode, actual_mode; |
a484f6ba | 2659 | optab op; |
5dfe80ba AS |
2660 | int actual_precision; |
2661 | location_t loc = gimple_location (stmt); | |
db719f50 | 2662 | bool from_unsigned1, from_unsigned2; |
1a39adae RS |
2663 | |
2664 | lhs = gimple_assign_lhs (stmt); | |
2665 | type = TREE_TYPE (lhs); | |
2666 | if (TREE_CODE (type) != INTEGER_TYPE) | |
0354c0c7 BS |
2667 | return false; |
2668 | ||
3d71881d | 2669 | if (!is_widening_mult_p (stmt, &type1, &rhs1, &type2, &rhs2)) |
0354c0c7 BS |
2670 | return false; |
2671 | ||
a484f6ba AS |
2672 | to_mode = TYPE_MODE (type); |
2673 | from_mode = TYPE_MODE (type1); | |
db719f50 AS |
2674 | from_unsigned1 = TYPE_UNSIGNED (type1); |
2675 | from_unsigned2 = TYPE_UNSIGNED (type2); | |
a484f6ba | 2676 | |
db719f50 | 2677 | if (from_unsigned1 && from_unsigned2) |
a484f6ba | 2678 | op = umul_widen_optab; |
db719f50 | 2679 | else if (!from_unsigned1 && !from_unsigned2) |
a484f6ba | 2680 | op = smul_widen_optab; |
0354c0c7 | 2681 | else |
a484f6ba AS |
2682 | op = usmul_widen_optab; |
2683 | ||
5dfe80ba AS |
2684 | handler = find_widening_optab_handler_and_mode (op, to_mode, from_mode, |
2685 | 0, &actual_mode); | |
1a39adae RS |
2686 | |
2687 | if (handler == CODE_FOR_nothing) | |
db719f50 AS |
2688 | { |
2689 | if (op != smul_widen_optab) | |
2690 | { | |
6a228c2c AS |
2691 | /* We can use a signed multiply with unsigned types as long as |
2692 | there is a wider mode to use, or it is the smaller of the two | |
2693 | types that is unsigned. Note that type1 >= type2, always. */ | |
2694 | if ((TYPE_UNSIGNED (type1) | |
2695 | && TYPE_PRECISION (type1) == GET_MODE_PRECISION (from_mode)) | |
2696 | || (TYPE_UNSIGNED (type2) | |
2697 | && TYPE_PRECISION (type2) == GET_MODE_PRECISION (from_mode))) | |
2698 | { | |
2699 | from_mode = GET_MODE_WIDER_MODE (from_mode); | |
2700 | if (GET_MODE_SIZE (to_mode) <= GET_MODE_SIZE (from_mode)) | |
2701 | return false; | |
2702 | } | |
db719f50 AS |
2703 | |
2704 | op = smul_widen_optab; | |
2705 | handler = find_widening_optab_handler_and_mode (op, to_mode, | |
2706 | from_mode, 0, | |
2707 | &actual_mode); | |
2708 | ||
2709 | if (handler == CODE_FOR_nothing) | |
2710 | return false; | |
2711 | ||
2712 | from_unsigned1 = from_unsigned2 = false; | |
2713 | } | |
2714 | else | |
2715 | return false; | |
2716 | } | |
1a39adae | 2717 | |
5dfe80ba AS |
2718 | /* Ensure that the inputs to the handler are in the correct precison |
2719 | for the opcode. This will be the full mode size. */ | |
2720 | actual_precision = GET_MODE_PRECISION (actual_mode); | |
f409d239 RG |
2721 | if (2 * actual_precision > TYPE_PRECISION (type)) |
2722 | return false; | |
db719f50 AS |
2723 | if (actual_precision != TYPE_PRECISION (type1) |
2724 | || from_unsigned1 != TYPE_UNSIGNED (type1)) | |
83d5977e RG |
2725 | rhs1 = build_and_insert_cast (gsi, loc, |
2726 | build_nonstandard_integer_type | |
2727 | (actual_precision, from_unsigned1), rhs1); | |
db719f50 AS |
2728 | if (actual_precision != TYPE_PRECISION (type2) |
2729 | || from_unsigned2 != TYPE_UNSIGNED (type2)) | |
83d5977e RG |
2730 | rhs2 = build_and_insert_cast (gsi, loc, |
2731 | build_nonstandard_integer_type | |
2732 | (actual_precision, from_unsigned2), rhs2); | |
5dfe80ba | 2733 | |
a6f969f4 AS |
2734 | /* Handle constants. */ |
2735 | if (TREE_CODE (rhs1) == INTEGER_CST) | |
2736 | rhs1 = fold_convert (type1, rhs1); | |
2737 | if (TREE_CODE (rhs2) == INTEGER_CST) | |
2738 | rhs2 = fold_convert (type2, rhs2); | |
2739 | ||
5dfe80ba AS |
2740 | gimple_assign_set_rhs1 (stmt, rhs1); |
2741 | gimple_assign_set_rhs2 (stmt, rhs2); | |
0354c0c7 BS |
2742 | gimple_assign_set_rhs_code (stmt, WIDEN_MULT_EXPR); |
2743 | update_stmt (stmt); | |
4da3b811 | 2744 | widen_mul_stats.widen_mults_inserted++; |
0354c0c7 BS |
2745 | return true; |
2746 | } | |
2747 | ||
2748 | /* Process a single gimple statement STMT, which is found at the | |
2749 | iterator GSI and has a either a PLUS_EXPR or a MINUS_EXPR as its | |
2750 | rhs (given by CODE), and try to convert it into a | |
2751 | WIDEN_MULT_PLUS_EXPR or a WIDEN_MULT_MINUS_EXPR. The return value | |
2752 | is true iff we converted the statement. */ | |
2753 | ||
2754 | static bool | |
2755 | convert_plusminus_to_widen (gimple_stmt_iterator *gsi, gimple stmt, | |
2756 | enum tree_code code) | |
2757 | { | |
2758 | gimple rhs1_stmt = NULL, rhs2_stmt = NULL; | |
cefb4d4f | 2759 | gimple conv1_stmt = NULL, conv2_stmt = NULL, conv_stmt; |
83d5977e | 2760 | tree type, type1, type2, optype; |
0354c0c7 BS |
2761 | tree lhs, rhs1, rhs2, mult_rhs1, mult_rhs2, add_rhs; |
2762 | enum tree_code rhs1_code = ERROR_MARK, rhs2_code = ERROR_MARK; | |
2763 | optab this_optab; | |
2764 | enum tree_code wmult_code; | |
5dfe80ba | 2765 | enum insn_code handler; |
ef4bddc2 | 2766 | machine_mode to_mode, from_mode, actual_mode; |
5dfe80ba AS |
2767 | location_t loc = gimple_location (stmt); |
2768 | int actual_precision; | |
db719f50 | 2769 | bool from_unsigned1, from_unsigned2; |
0354c0c7 BS |
2770 | |
2771 | lhs = gimple_assign_lhs (stmt); | |
2772 | type = TREE_TYPE (lhs); | |
1a39adae RS |
2773 | if (TREE_CODE (type) != INTEGER_TYPE |
2774 | && TREE_CODE (type) != FIXED_POINT_TYPE) | |
0354c0c7 BS |
2775 | return false; |
2776 | ||
2777 | if (code == MINUS_EXPR) | |
2778 | wmult_code = WIDEN_MULT_MINUS_EXPR; | |
2779 | else | |
2780 | wmult_code = WIDEN_MULT_PLUS_EXPR; | |
2781 | ||
0354c0c7 BS |
2782 | rhs1 = gimple_assign_rhs1 (stmt); |
2783 | rhs2 = gimple_assign_rhs2 (stmt); | |
2784 | ||
2785 | if (TREE_CODE (rhs1) == SSA_NAME) | |
2786 | { | |
2787 | rhs1_stmt = SSA_NAME_DEF_STMT (rhs1); | |
2788 | if (is_gimple_assign (rhs1_stmt)) | |
2789 | rhs1_code = gimple_assign_rhs_code (rhs1_stmt); | |
2790 | } | |
0354c0c7 BS |
2791 | |
2792 | if (TREE_CODE (rhs2) == SSA_NAME) | |
2793 | { | |
2794 | rhs2_stmt = SSA_NAME_DEF_STMT (rhs2); | |
2795 | if (is_gimple_assign (rhs2_stmt)) | |
2796 | rhs2_code = gimple_assign_rhs_code (rhs2_stmt); | |
2797 | } | |
0354c0c7 | 2798 | |
cefb4d4f AS |
2799 | /* Allow for one conversion statement between the multiply |
2800 | and addition/subtraction statement. If there are more than | |
2801 | one conversions then we assume they would invalidate this | |
2802 | transformation. If that's not the case then they should have | |
2803 | been folded before now. */ | |
2804 | if (CONVERT_EXPR_CODE_P (rhs1_code)) | |
2805 | { | |
2806 | conv1_stmt = rhs1_stmt; | |
2807 | rhs1 = gimple_assign_rhs1 (rhs1_stmt); | |
2808 | if (TREE_CODE (rhs1) == SSA_NAME) | |
2809 | { | |
2810 | rhs1_stmt = SSA_NAME_DEF_STMT (rhs1); | |
2811 | if (is_gimple_assign (rhs1_stmt)) | |
2812 | rhs1_code = gimple_assign_rhs_code (rhs1_stmt); | |
2813 | } | |
2814 | else | |
2815 | return false; | |
2816 | } | |
2817 | if (CONVERT_EXPR_CODE_P (rhs2_code)) | |
2818 | { | |
2819 | conv2_stmt = rhs2_stmt; | |
2820 | rhs2 = gimple_assign_rhs1 (rhs2_stmt); | |
2821 | if (TREE_CODE (rhs2) == SSA_NAME) | |
2822 | { | |
2823 | rhs2_stmt = SSA_NAME_DEF_STMT (rhs2); | |
2824 | if (is_gimple_assign (rhs2_stmt)) | |
2825 | rhs2_code = gimple_assign_rhs_code (rhs2_stmt); | |
2826 | } | |
2827 | else | |
2828 | return false; | |
2829 | } | |
2830 | ||
5dfe80ba AS |
2831 | /* If code is WIDEN_MULT_EXPR then it would seem unnecessary to call |
2832 | is_widening_mult_p, but we still need the rhs returns. | |
2833 | ||
2834 | It might also appear that it would be sufficient to use the existing | |
2835 | operands of the widening multiply, but that would limit the choice of | |
42917d01 YZ |
2836 | multiply-and-accumulate instructions. |
2837 | ||
2838 | If the widened-multiplication result has more than one uses, it is | |
2839 | probably wiser not to do the conversion. */ | |
5dfe80ba AS |
2840 | if (code == PLUS_EXPR |
2841 | && (rhs1_code == MULT_EXPR || rhs1_code == WIDEN_MULT_EXPR)) | |
0354c0c7 | 2842 | { |
42917d01 YZ |
2843 | if (!has_single_use (rhs1) |
2844 | || !is_widening_mult_p (rhs1_stmt, &type1, &mult_rhs1, | |
2845 | &type2, &mult_rhs2)) | |
0354c0c7 | 2846 | return false; |
1a39adae | 2847 | add_rhs = rhs2; |
cefb4d4f | 2848 | conv_stmt = conv1_stmt; |
0354c0c7 | 2849 | } |
5dfe80ba | 2850 | else if (rhs2_code == MULT_EXPR || rhs2_code == WIDEN_MULT_EXPR) |
0354c0c7 | 2851 | { |
42917d01 YZ |
2852 | if (!has_single_use (rhs2) |
2853 | || !is_widening_mult_p (rhs2_stmt, &type1, &mult_rhs1, | |
2854 | &type2, &mult_rhs2)) | |
0354c0c7 | 2855 | return false; |
1a39adae | 2856 | add_rhs = rhs1; |
cefb4d4f | 2857 | conv_stmt = conv2_stmt; |
0354c0c7 | 2858 | } |
0354c0c7 BS |
2859 | else |
2860 | return false; | |
2861 | ||
5dfe80ba AS |
2862 | to_mode = TYPE_MODE (type); |
2863 | from_mode = TYPE_MODE (type1); | |
db719f50 AS |
2864 | from_unsigned1 = TYPE_UNSIGNED (type1); |
2865 | from_unsigned2 = TYPE_UNSIGNED (type2); | |
3752b2ab | 2866 | optype = type1; |
5dfe80ba | 2867 | |
db719f50 AS |
2868 | /* There's no such thing as a mixed sign madd yet, so use a wider mode. */ |
2869 | if (from_unsigned1 != from_unsigned2) | |
2870 | { | |
3752b2ab RS |
2871 | if (!INTEGRAL_TYPE_P (type)) |
2872 | return false; | |
6a228c2c AS |
2873 | /* We can use a signed multiply with unsigned types as long as |
2874 | there is a wider mode to use, or it is the smaller of the two | |
2875 | types that is unsigned. Note that type1 >= type2, always. */ | |
2876 | if ((from_unsigned1 | |
2877 | && TYPE_PRECISION (type1) == GET_MODE_PRECISION (from_mode)) | |
2878 | || (from_unsigned2 | |
2879 | && TYPE_PRECISION (type2) == GET_MODE_PRECISION (from_mode))) | |
db719f50 | 2880 | { |
6a228c2c AS |
2881 | from_mode = GET_MODE_WIDER_MODE (from_mode); |
2882 | if (GET_MODE_SIZE (from_mode) >= GET_MODE_SIZE (to_mode)) | |
2883 | return false; | |
db719f50 | 2884 | } |
6a228c2c AS |
2885 | |
2886 | from_unsigned1 = from_unsigned2 = false; | |
3752b2ab RS |
2887 | optype = build_nonstandard_integer_type (GET_MODE_PRECISION (from_mode), |
2888 | false); | |
db719f50 | 2889 | } |
9eab7f91 | 2890 | |
cefb4d4f AS |
2891 | /* If there was a conversion between the multiply and addition |
2892 | then we need to make sure it fits a multiply-and-accumulate. | |
2893 | The should be a single mode change which does not change the | |
2894 | value. */ | |
2895 | if (conv_stmt) | |
2896 | { | |
db719f50 | 2897 | /* We use the original, unmodified data types for this. */ |
cefb4d4f AS |
2898 | tree from_type = TREE_TYPE (gimple_assign_rhs1 (conv_stmt)); |
2899 | tree to_type = TREE_TYPE (gimple_assign_lhs (conv_stmt)); | |
2900 | int data_size = TYPE_PRECISION (type1) + TYPE_PRECISION (type2); | |
2901 | bool is_unsigned = TYPE_UNSIGNED (type1) && TYPE_UNSIGNED (type2); | |
2902 | ||
2903 | if (TYPE_PRECISION (from_type) > TYPE_PRECISION (to_type)) | |
2904 | { | |
2905 | /* Conversion is a truncate. */ | |
2906 | if (TYPE_PRECISION (to_type) < data_size) | |
2907 | return false; | |
2908 | } | |
2909 | else if (TYPE_PRECISION (from_type) < TYPE_PRECISION (to_type)) | |
2910 | { | |
2911 | /* Conversion is an extend. Check it's the right sort. */ | |
2912 | if (TYPE_UNSIGNED (from_type) != is_unsigned | |
2913 | && !(is_unsigned && TYPE_PRECISION (from_type) > data_size)) | |
2914 | return false; | |
2915 | } | |
2916 | /* else convert is a no-op for our purposes. */ | |
2917 | } | |
2918 | ||
9eab7f91 RS |
2919 | /* Verify that the machine can perform a widening multiply |
2920 | accumulate in this mode/signedness combination, otherwise | |
2921 | this transformation is likely to pessimize code. */ | |
db719f50 | 2922 | this_optab = optab_for_tree_code (wmult_code, optype, optab_default); |
5dfe80ba AS |
2923 | handler = find_widening_optab_handler_and_mode (this_optab, to_mode, |
2924 | from_mode, 0, &actual_mode); | |
2925 | ||
2926 | if (handler == CODE_FOR_nothing) | |
9eab7f91 RS |
2927 | return false; |
2928 | ||
5dfe80ba AS |
2929 | /* Ensure that the inputs to the handler are in the correct precison |
2930 | for the opcode. This will be the full mode size. */ | |
2931 | actual_precision = GET_MODE_PRECISION (actual_mode); | |
db719f50 AS |
2932 | if (actual_precision != TYPE_PRECISION (type1) |
2933 | || from_unsigned1 != TYPE_UNSIGNED (type1)) | |
83d5977e RG |
2934 | mult_rhs1 = build_and_insert_cast (gsi, loc, |
2935 | build_nonstandard_integer_type | |
2936 | (actual_precision, from_unsigned1), | |
2937 | mult_rhs1); | |
db719f50 AS |
2938 | if (actual_precision != TYPE_PRECISION (type2) |
2939 | || from_unsigned2 != TYPE_UNSIGNED (type2)) | |
83d5977e RG |
2940 | mult_rhs2 = build_and_insert_cast (gsi, loc, |
2941 | build_nonstandard_integer_type | |
2942 | (actual_precision, from_unsigned2), | |
2943 | mult_rhs2); | |
0354c0c7 | 2944 | |
75161d2c | 2945 | if (!useless_type_conversion_p (type, TREE_TYPE (add_rhs))) |
83d5977e | 2946 | add_rhs = build_and_insert_cast (gsi, loc, type, add_rhs); |
75161d2c | 2947 | |
a6f969f4 AS |
2948 | /* Handle constants. */ |
2949 | if (TREE_CODE (mult_rhs1) == INTEGER_CST) | |
c3c5a1cc | 2950 | mult_rhs1 = fold_convert (type1, mult_rhs1); |
a6f969f4 | 2951 | if (TREE_CODE (mult_rhs2) == INTEGER_CST) |
c3c5a1cc | 2952 | mult_rhs2 = fold_convert (type2, mult_rhs2); |
a6f969f4 | 2953 | |
5dfe80ba | 2954 | gimple_assign_set_rhs_with_ops_1 (gsi, wmult_code, mult_rhs1, mult_rhs2, |
0354c0c7 BS |
2955 | add_rhs); |
2956 | update_stmt (gsi_stmt (*gsi)); | |
4da3b811 | 2957 | widen_mul_stats.maccs_inserted++; |
0354c0c7 BS |
2958 | return true; |
2959 | } | |
2960 | ||
4dbed5f6 RG |
2961 | /* Combine the multiplication at MUL_STMT with operands MULOP1 and MULOP2 |
2962 | with uses in additions and subtractions to form fused multiply-add | |
2963 | operations. Returns true if successful and MUL_STMT should be removed. */ | |
16949072 RG |
2964 | |
2965 | static bool | |
4dbed5f6 | 2966 | convert_mult_to_fma (gimple mul_stmt, tree op1, tree op2) |
16949072 | 2967 | { |
4dbed5f6 | 2968 | tree mul_result = gimple_get_lhs (mul_stmt); |
16949072 | 2969 | tree type = TREE_TYPE (mul_result); |
a5f09e73 | 2970 | gimple use_stmt, neguse_stmt, fma_stmt; |
16949072 RG |
2971 | use_operand_p use_p; |
2972 | imm_use_iterator imm_iter; | |
2973 | ||
2974 | if (FLOAT_TYPE_P (type) | |
2975 | && flag_fp_contract_mode == FP_CONTRACT_OFF) | |
2976 | return false; | |
2977 | ||
2978 | /* We don't want to do bitfield reduction ops. */ | |
2979 | if (INTEGRAL_TYPE_P (type) | |
2980 | && (TYPE_PRECISION (type) | |
2981 | != GET_MODE_PRECISION (TYPE_MODE (type)))) | |
2982 | return false; | |
2983 | ||
2984 | /* If the target doesn't support it, don't generate it. We assume that | |
2985 | if fma isn't available then fms, fnma or fnms are not either. */ | |
2986 | if (optab_handler (fma_optab, TYPE_MODE (type)) == CODE_FOR_nothing) | |
2987 | return false; | |
2988 | ||
0fb808ea JJ |
2989 | /* If the multiplication has zero uses, it is kept around probably because |
2990 | of -fnon-call-exceptions. Don't optimize it away in that case, | |
2991 | it is DCE job. */ | |
2992 | if (has_zero_uses (mul_result)) | |
2993 | return false; | |
2994 | ||
16949072 RG |
2995 | /* Make sure that the multiplication statement becomes dead after |
2996 | the transformation, thus that all uses are transformed to FMAs. | |
2997 | This means we assume that an FMA operation has the same cost | |
2998 | as an addition. */ | |
2999 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, mul_result) | |
3000 | { | |
3001 | enum tree_code use_code; | |
a5f09e73 RH |
3002 | tree result = mul_result; |
3003 | bool negate_p = false; | |
16949072 RG |
3004 | |
3005 | use_stmt = USE_STMT (use_p); | |
3006 | ||
76b14c29 RG |
3007 | if (is_gimple_debug (use_stmt)) |
3008 | continue; | |
3009 | ||
16949072 RG |
3010 | /* For now restrict this operations to single basic blocks. In theory |
3011 | we would want to support sinking the multiplication in | |
3012 | m = a*b; | |
3013 | if () | |
3014 | ma = m + c; | |
3015 | else | |
3016 | d = m; | |
3017 | to form a fma in the then block and sink the multiplication to the | |
3018 | else block. */ | |
3019 | if (gimple_bb (use_stmt) != gimple_bb (mul_stmt)) | |
3020 | return false; | |
3021 | ||
a5f09e73 | 3022 | if (!is_gimple_assign (use_stmt)) |
16949072 RG |
3023 | return false; |
3024 | ||
a5f09e73 RH |
3025 | use_code = gimple_assign_rhs_code (use_stmt); |
3026 | ||
3027 | /* A negate on the multiplication leads to FNMA. */ | |
3028 | if (use_code == NEGATE_EXPR) | |
3029 | { | |
a758fd67 | 3030 | ssa_op_iter iter; |
dae957ae | 3031 | use_operand_p usep; |
a758fd67 | 3032 | |
a5f09e73 RH |
3033 | result = gimple_assign_lhs (use_stmt); |
3034 | ||
3035 | /* Make sure the negate statement becomes dead with this | |
3036 | single transformation. */ | |
3037 | if (!single_imm_use (gimple_assign_lhs (use_stmt), | |
3038 | &use_p, &neguse_stmt)) | |
3039 | return false; | |
3040 | ||
a758fd67 | 3041 | /* Make sure the multiplication isn't also used on that stmt. */ |
dae957ae RG |
3042 | FOR_EACH_PHI_OR_STMT_USE (usep, neguse_stmt, iter, SSA_OP_USE) |
3043 | if (USE_FROM_PTR (usep) == mul_result) | |
a758fd67 RG |
3044 | return false; |
3045 | ||
a5f09e73 RH |
3046 | /* Re-validate. */ |
3047 | use_stmt = neguse_stmt; | |
3048 | if (gimple_bb (use_stmt) != gimple_bb (mul_stmt)) | |
3049 | return false; | |
3050 | if (!is_gimple_assign (use_stmt)) | |
3051 | return false; | |
3052 | ||
3053 | use_code = gimple_assign_rhs_code (use_stmt); | |
3054 | negate_p = true; | |
3055 | } | |
16949072 | 3056 | |
a5f09e73 RH |
3057 | switch (use_code) |
3058 | { | |
3059 | case MINUS_EXPR: | |
a1d8aa4b RH |
3060 | if (gimple_assign_rhs2 (use_stmt) == result) |
3061 | negate_p = !negate_p; | |
3062 | break; | |
a5f09e73 | 3063 | case PLUS_EXPR: |
a5f09e73 | 3064 | break; |
a5f09e73 RH |
3065 | default: |
3066 | /* FMA can only be formed from PLUS and MINUS. */ | |
3067 | return false; | |
3068 | } | |
16949072 | 3069 | |
ee8a9b7b JR |
3070 | /* If the subtrahend (gimple_assign_rhs2 (use_stmt)) is computed |
3071 | by a MULT_EXPR that we'll visit later, we might be able to | |
3072 | get a more profitable match with fnma. | |
3073 | OTOH, if we don't, a negate / fma pair has likely lower latency | |
3074 | that a mult / subtract pair. */ | |
3075 | if (use_code == MINUS_EXPR && !negate_p | |
3076 | && gimple_assign_rhs1 (use_stmt) == result | |
3077 | && optab_handler (fms_optab, TYPE_MODE (type)) == CODE_FOR_nothing | |
3078 | && optab_handler (fnma_optab, TYPE_MODE (type)) != CODE_FOR_nothing) | |
3079 | { | |
3080 | tree rhs2 = gimple_assign_rhs2 (use_stmt); | |
ee8a9b7b | 3081 | |
95c03b36 JR |
3082 | if (TREE_CODE (rhs2) == SSA_NAME) |
3083 | { | |
3084 | gimple stmt2 = SSA_NAME_DEF_STMT (rhs2); | |
3085 | if (has_single_use (rhs2) | |
3086 | && is_gimple_assign (stmt2) | |
3087 | && gimple_assign_rhs_code (stmt2) == MULT_EXPR) | |
3088 | return false; | |
3089 | } | |
ee8a9b7b JR |
3090 | } |
3091 | ||
a5f09e73 RH |
3092 | /* We can't handle a * b + a * b. */ |
3093 | if (gimple_assign_rhs1 (use_stmt) == gimple_assign_rhs2 (use_stmt)) | |
3094 | return false; | |
a1d8aa4b RH |
3095 | |
3096 | /* While it is possible to validate whether or not the exact form | |
3097 | that we've recognized is available in the backend, the assumption | |
3098 | is that the transformation is never a loss. For instance, suppose | |
3099 | the target only has the plain FMA pattern available. Consider | |
3100 | a*b-c -> fma(a,b,-c): we've exchanged MUL+SUB for FMA+NEG, which | |
3101 | is still two operations. Consider -(a*b)-c -> fma(-a,b,-c): we | |
3102 | still have 3 operations, but in the FMA form the two NEGs are | |
073a8998 | 3103 | independent and could be run in parallel. */ |
16949072 RG |
3104 | } |
3105 | ||
3106 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, mul_result) | |
3107 | { | |
16949072 | 3108 | gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt); |
76b14c29 | 3109 | enum tree_code use_code; |
4dbed5f6 | 3110 | tree addop, mulop1 = op1, result = mul_result; |
a5f09e73 | 3111 | bool negate_p = false; |
16949072 | 3112 | |
76b14c29 RG |
3113 | if (is_gimple_debug (use_stmt)) |
3114 | continue; | |
3115 | ||
3116 | use_code = gimple_assign_rhs_code (use_stmt); | |
a5f09e73 RH |
3117 | if (use_code == NEGATE_EXPR) |
3118 | { | |
3119 | result = gimple_assign_lhs (use_stmt); | |
3120 | single_imm_use (gimple_assign_lhs (use_stmt), &use_p, &neguse_stmt); | |
3121 | gsi_remove (&gsi, true); | |
3122 | release_defs (use_stmt); | |
3123 | ||
3124 | use_stmt = neguse_stmt; | |
3125 | gsi = gsi_for_stmt (use_stmt); | |
3126 | use_code = gimple_assign_rhs_code (use_stmt); | |
3127 | negate_p = true; | |
3128 | } | |
3129 | ||
3130 | if (gimple_assign_rhs1 (use_stmt) == result) | |
16949072 RG |
3131 | { |
3132 | addop = gimple_assign_rhs2 (use_stmt); | |
3133 | /* a * b - c -> a * b + (-c) */ | |
3134 | if (gimple_assign_rhs_code (use_stmt) == MINUS_EXPR) | |
3135 | addop = force_gimple_operand_gsi (&gsi, | |
3136 | build1 (NEGATE_EXPR, | |
3137 | type, addop), | |
3138 | true, NULL_TREE, true, | |
3139 | GSI_SAME_STMT); | |
3140 | } | |
3141 | else | |
3142 | { | |
3143 | addop = gimple_assign_rhs1 (use_stmt); | |
3144 | /* a - b * c -> (-b) * c + a */ | |
3145 | if (gimple_assign_rhs_code (use_stmt) == MINUS_EXPR) | |
a5f09e73 | 3146 | negate_p = !negate_p; |
16949072 RG |
3147 | } |
3148 | ||
a5f09e73 RH |
3149 | if (negate_p) |
3150 | mulop1 = force_gimple_operand_gsi (&gsi, | |
3151 | build1 (NEGATE_EXPR, | |
3152 | type, mulop1), | |
3153 | true, NULL_TREE, true, | |
3154 | GSI_SAME_STMT); | |
3155 | ||
73804b12 RG |
3156 | fma_stmt = gimple_build_assign_with_ops (FMA_EXPR, |
3157 | gimple_assign_lhs (use_stmt), | |
3158 | mulop1, op2, | |
3159 | addop); | |
16949072 | 3160 | gsi_replace (&gsi, fma_stmt, true); |
4da3b811 | 3161 | widen_mul_stats.fmas_inserted++; |
16949072 RG |
3162 | } |
3163 | ||
3164 | return true; | |
3165 | } | |
3166 | ||
5b58b39b BS |
3167 | /* Find integer multiplications where the operands are extended from |
3168 | smaller types, and replace the MULT_EXPR with a WIDEN_MULT_EXPR | |
3169 | where appropriate. */ | |
3170 | ||
be55bfe6 TS |
3171 | namespace { |
3172 | ||
3173 | const pass_data pass_data_optimize_widening_mul = | |
3174 | { | |
3175 | GIMPLE_PASS, /* type */ | |
3176 | "widening_mul", /* name */ | |
3177 | OPTGROUP_NONE, /* optinfo_flags */ | |
be55bfe6 TS |
3178 | TV_NONE, /* tv_id */ |
3179 | PROP_ssa, /* properties_required */ | |
3180 | 0, /* properties_provided */ | |
3181 | 0, /* properties_destroyed */ | |
3182 | 0, /* todo_flags_start */ | |
3bea341f | 3183 | TODO_update_ssa, /* todo_flags_finish */ |
be55bfe6 TS |
3184 | }; |
3185 | ||
3186 | class pass_optimize_widening_mul : public gimple_opt_pass | |
3187 | { | |
3188 | public: | |
3189 | pass_optimize_widening_mul (gcc::context *ctxt) | |
3190 | : gimple_opt_pass (pass_data_optimize_widening_mul, ctxt) | |
3191 | {} | |
3192 | ||
3193 | /* opt_pass methods: */ | |
3194 | virtual bool gate (function *) | |
3195 | { | |
3196 | return flag_expensive_optimizations && optimize; | |
3197 | } | |
3198 | ||
3199 | virtual unsigned int execute (function *); | |
3200 | ||
3201 | }; // class pass_optimize_widening_mul | |
3202 | ||
3203 | unsigned int | |
3204 | pass_optimize_widening_mul::execute (function *fun) | |
5b58b39b | 3205 | { |
5b58b39b | 3206 | basic_block bb; |
4dbed5f6 | 3207 | bool cfg_changed = false; |
5b58b39b | 3208 | |
4da3b811 NF |
3209 | memset (&widen_mul_stats, 0, sizeof (widen_mul_stats)); |
3210 | ||
be55bfe6 | 3211 | FOR_EACH_BB_FN (bb, fun) |
5b58b39b BS |
3212 | { |
3213 | gimple_stmt_iterator gsi; | |
3214 | ||
16949072 | 3215 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi);) |
5b58b39b BS |
3216 | { |
3217 | gimple stmt = gsi_stmt (gsi); | |
0354c0c7 | 3218 | enum tree_code code; |
5b58b39b | 3219 | |
16949072 RG |
3220 | if (is_gimple_assign (stmt)) |
3221 | { | |
3222 | code = gimple_assign_rhs_code (stmt); | |
3223 | switch (code) | |
3224 | { | |
3225 | case MULT_EXPR: | |
5dfe80ba | 3226 | if (!convert_mult_to_widen (stmt, &gsi) |
4dbed5f6 RG |
3227 | && convert_mult_to_fma (stmt, |
3228 | gimple_assign_rhs1 (stmt), | |
3229 | gimple_assign_rhs2 (stmt))) | |
16949072 RG |
3230 | { |
3231 | gsi_remove (&gsi, true); | |
3232 | release_defs (stmt); | |
3233 | continue; | |
3234 | } | |
3235 | break; | |
3236 | ||
3237 | case PLUS_EXPR: | |
3238 | case MINUS_EXPR: | |
3239 | convert_plusminus_to_widen (&gsi, stmt, code); | |
3240 | break; | |
5b58b39b | 3241 | |
16949072 RG |
3242 | default:; |
3243 | } | |
3244 | } | |
85a47bed RG |
3245 | else if (is_gimple_call (stmt) |
3246 | && gimple_call_lhs (stmt)) | |
4dbed5f6 RG |
3247 | { |
3248 | tree fndecl = gimple_call_fndecl (stmt); | |
3249 | if (fndecl | |
3250 | && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) | |
3251 | { | |
3252 | switch (DECL_FUNCTION_CODE (fndecl)) | |
3253 | { | |
3254 | case BUILT_IN_POWF: | |
3255 | case BUILT_IN_POW: | |
3256 | case BUILT_IN_POWL: | |
3257 | if (TREE_CODE (gimple_call_arg (stmt, 1)) == REAL_CST | |
3258 | && REAL_VALUES_EQUAL | |
3259 | (TREE_REAL_CST (gimple_call_arg (stmt, 1)), | |
3260 | dconst2) | |
3261 | && convert_mult_to_fma (stmt, | |
3262 | gimple_call_arg (stmt, 0), | |
3263 | gimple_call_arg (stmt, 0))) | |
3264 | { | |
0b238a9b | 3265 | unlink_stmt_vdef (stmt); |
b5b3ec3e RG |
3266 | if (gsi_remove (&gsi, true) |
3267 | && gimple_purge_dead_eh_edges (bb)) | |
4dbed5f6 | 3268 | cfg_changed = true; |
b5b3ec3e | 3269 | release_defs (stmt); |
4dbed5f6 RG |
3270 | continue; |
3271 | } | |
3272 | break; | |
3273 | ||
3274 | default:; | |
3275 | } | |
3276 | } | |
3277 | } | |
16949072 | 3278 | gsi_next (&gsi); |
5b58b39b BS |
3279 | } |
3280 | } | |
0354c0c7 | 3281 | |
be55bfe6 | 3282 | statistics_counter_event (fun, "widening multiplications inserted", |
4da3b811 | 3283 | widen_mul_stats.widen_mults_inserted); |
be55bfe6 | 3284 | statistics_counter_event (fun, "widening maccs inserted", |
4da3b811 | 3285 | widen_mul_stats.maccs_inserted); |
be55bfe6 | 3286 | statistics_counter_event (fun, "fused multiply-adds inserted", |
4da3b811 NF |
3287 | widen_mul_stats.fmas_inserted); |
3288 | ||
4dbed5f6 | 3289 | return cfg_changed ? TODO_cleanup_cfg : 0; |
5b58b39b BS |
3290 | } |
3291 | ||
27a4cd48 DM |
3292 | } // anon namespace |
3293 | ||
3294 | gimple_opt_pass * | |
3295 | make_pass_optimize_widening_mul (gcc::context *ctxt) | |
3296 | { | |
3297 | return new pass_optimize_widening_mul (ctxt); | |
3298 | } |