]>
Commit | Line | Data |
---|---|---|
abacb398 | 1 | /* Global, SSA-based optimizations using mathematical identities. |
e78306af | 2 | Copyright (C) 2005, 2006, 2007, 2008, 2009, 2010, 2011 |
7cf0dbf3 | 3 | Free Software Foundation, Inc. |
48e1416a | 4 | |
abacb398 | 5 | This file is part of GCC. |
48e1416a | 6 | |
abacb398 | 7 | GCC is free software; you can redistribute it and/or modify it |
8 | under the terms of the GNU General Public License as published by the | |
8c4c00c1 | 9 | Free Software Foundation; either version 3, or (at your option) any |
abacb398 | 10 | later version. |
48e1416a | 11 | |
abacb398 | 12 | GCC is distributed in the hope that it will be useful, but WITHOUT |
13 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 | for more details. | |
48e1416a | 16 | |
abacb398 | 17 | You should have received a copy of the GNU General Public License |
8c4c00c1 | 18 | along with GCC; see the file COPYING3. If not see |
19 | <http://www.gnu.org/licenses/>. */ | |
abacb398 | 20 | |
21 | /* Currently, the only mini-pass in this file tries to CSE reciprocal | |
22 | operations. These are common in sequences such as this one: | |
23 | ||
24 | modulus = sqrt(x*x + y*y + z*z); | |
25 | x = x / modulus; | |
26 | y = y / modulus; | |
27 | z = z / modulus; | |
28 | ||
29 | that can be optimized to | |
30 | ||
31 | modulus = sqrt(x*x + y*y + z*z); | |
32 | rmodulus = 1.0 / modulus; | |
33 | x = x * rmodulus; | |
34 | y = y * rmodulus; | |
35 | z = z * rmodulus; | |
36 | ||
37 | We do this for loop invariant divisors, and with this pass whenever | |
ac70caad | 38 | we notice that a division has the same divisor multiple times. |
39 | ||
40 | Of course, like in PRE, we don't insert a division if a dominator | |
41 | already has one. However, this cannot be done as an extension of | |
42 | PRE for several reasons. | |
43 | ||
44 | First of all, with some experiments it was found out that the | |
45 | transformation is not always useful if there are only two divisions | |
46 | hy the same divisor. This is probably because modern processors | |
47 | can pipeline the divisions; on older, in-order processors it should | |
48 | still be effective to optimize two divisions by the same number. | |
49 | We make this a param, and it shall be called N in the remainder of | |
50 | this comment. | |
51 | ||
52 | Second, if trapping math is active, we have less freedom on where | |
53 | to insert divisions: we can only do so in basic blocks that already | |
54 | contain one. (If divisions don't trap, instead, we can insert | |
55 | divisions elsewhere, which will be in blocks that are common dominators | |
56 | of those that have the division). | |
57 | ||
58 | We really don't want to compute the reciprocal unless a division will | |
59 | be found. To do this, we won't insert the division in a basic block | |
60 | that has less than N divisions *post-dominating* it. | |
61 | ||
62 | The algorithm constructs a subset of the dominator tree, holding the | |
63 | blocks containing the divisions and the common dominators to them, | |
64 | and walk it twice. The first walk is in post-order, and it annotates | |
65 | each block with the number of divisions that post-dominate it: this | |
66 | gives information on where divisions can be inserted profitably. | |
67 | The second walk is in pre-order, and it inserts divisions as explained | |
68 | above, and replaces divisions by multiplications. | |
69 | ||
70 | In the best case, the cost of the pass is O(n_statements). In the | |
71 | worst-case, the cost is due to creating the dominator tree subset, | |
72 | with a cost of O(n_basic_blocks ^ 2); however this can only happen | |
73 | for n_statements / n_basic_blocks statements. So, the amortized cost | |
74 | of creating the dominator tree subset is O(n_basic_blocks) and the | |
75 | worst-case cost of the pass is O(n_statements * n_basic_blocks). | |
76 | ||
77 | More practically, the cost will be small because there are few | |
78 | divisions, and they tend to be in the same basic block, so insert_bb | |
79 | is called very few times. | |
80 | ||
81 | If we did this using domwalk.c, an efficient implementation would have | |
82 | to work on all the variables in a single pass, because we could not | |
83 | work on just a subset of the dominator tree, as we do now, and the | |
84 | cost would also be something like O(n_statements * n_basic_blocks). | |
85 | The data structures would be more complex in order to work on all the | |
86 | variables in a single pass. */ | |
abacb398 | 87 | |
88 | #include "config.h" | |
89 | #include "system.h" | |
90 | #include "coretypes.h" | |
91 | #include "tm.h" | |
92 | #include "flags.h" | |
93 | #include "tree.h" | |
94 | #include "tree-flow.h" | |
abacb398 | 95 | #include "tree-pass.h" |
ac70caad | 96 | #include "alloc-pool.h" |
97 | #include "basic-block.h" | |
98 | #include "target.h" | |
ce084dfc | 99 | #include "gimple-pretty-print.h" |
a7a46268 | 100 | |
101 | /* FIXME: RTL headers have to be included here for optabs. */ | |
102 | #include "rtl.h" /* Because optabs.h wants enum rtx_code. */ | |
103 | #include "expr.h" /* Because optabs.h wants sepops. */ | |
84cc784c | 104 | #include "optabs.h" |
ac70caad | 105 | |
106 | /* This structure represents one basic block that either computes a | |
107 | division, or is a common dominator for basic block that compute a | |
108 | division. */ | |
109 | struct occurrence { | |
110 | /* The basic block represented by this structure. */ | |
111 | basic_block bb; | |
112 | ||
113 | /* If non-NULL, the SSA_NAME holding the definition for a reciprocal | |
114 | inserted in BB. */ | |
115 | tree recip_def; | |
116 | ||
75a70cf9 | 117 | /* If non-NULL, the GIMPLE_ASSIGN for a reciprocal computation that |
ac70caad | 118 | was inserted in BB. */ |
75a70cf9 | 119 | gimple recip_def_stmt; |
ac70caad | 120 | |
121 | /* Pointer to a list of "struct occurrence"s for blocks dominated | |
122 | by BB. */ | |
123 | struct occurrence *children; | |
124 | ||
125 | /* Pointer to the next "struct occurrence"s in the list of blocks | |
126 | sharing a common dominator. */ | |
127 | struct occurrence *next; | |
128 | ||
129 | /* The number of divisions that are in BB before compute_merit. The | |
130 | number of divisions that are in BB or post-dominate it after | |
131 | compute_merit. */ | |
132 | int num_divisions; | |
133 | ||
134 | /* True if the basic block has a division, false if it is a common | |
135 | dominator for basic blocks that do. If it is false and trapping | |
136 | math is active, BB is not a candidate for inserting a reciprocal. */ | |
137 | bool bb_has_division; | |
138 | }; | |
139 | ||
30c4e60d | 140 | static struct |
141 | { | |
142 | /* Number of 1.0/X ops inserted. */ | |
143 | int rdivs_inserted; | |
144 | ||
145 | /* Number of 1.0/FUNC ops inserted. */ | |
146 | int rfuncs_inserted; | |
147 | } reciprocal_stats; | |
148 | ||
149 | static struct | |
150 | { | |
151 | /* Number of cexpi calls inserted. */ | |
152 | int inserted; | |
153 | } sincos_stats; | |
154 | ||
155 | static struct | |
156 | { | |
157 | /* Number of hand-written 32-bit bswaps found. */ | |
158 | int found_32bit; | |
159 | ||
160 | /* Number of hand-written 64-bit bswaps found. */ | |
161 | int found_64bit; | |
162 | } bswap_stats; | |
163 | ||
164 | static struct | |
165 | { | |
166 | /* Number of widening multiplication ops inserted. */ | |
167 | int widen_mults_inserted; | |
168 | ||
169 | /* Number of integer multiply-and-accumulate ops inserted. */ | |
170 | int maccs_inserted; | |
171 | ||
172 | /* Number of fp fused multiply-add ops inserted. */ | |
173 | int fmas_inserted; | |
174 | } widen_mul_stats; | |
ac70caad | 175 | |
176 | /* The instance of "struct occurrence" representing the highest | |
177 | interesting block in the dominator tree. */ | |
178 | static struct occurrence *occ_head; | |
179 | ||
180 | /* Allocation pool for getting instances of "struct occurrence". */ | |
181 | static alloc_pool occ_pool; | |
182 | ||
183 | ||
184 | ||
185 | /* Allocate and return a new struct occurrence for basic block BB, and | |
186 | whose children list is headed by CHILDREN. */ | |
187 | static struct occurrence * | |
188 | occ_new (basic_block bb, struct occurrence *children) | |
abacb398 | 189 | { |
ac70caad | 190 | struct occurrence *occ; |
191 | ||
f0d6e81c | 192 | bb->aux = occ = (struct occurrence *) pool_alloc (occ_pool); |
ac70caad | 193 | memset (occ, 0, sizeof (struct occurrence)); |
194 | ||
195 | occ->bb = bb; | |
196 | occ->children = children; | |
197 | return occ; | |
abacb398 | 198 | } |
199 | ||
ac70caad | 200 | |
201 | /* Insert NEW_OCC into our subset of the dominator tree. P_HEAD points to a | |
202 | list of "struct occurrence"s, one per basic block, having IDOM as | |
203 | their common dominator. | |
204 | ||
205 | We try to insert NEW_OCC as deep as possible in the tree, and we also | |
206 | insert any other block that is a common dominator for BB and one | |
207 | block already in the tree. */ | |
208 | ||
209 | static void | |
210 | insert_bb (struct occurrence *new_occ, basic_block idom, | |
211 | struct occurrence **p_head) | |
9e583fac | 212 | { |
ac70caad | 213 | struct occurrence *occ, **p_occ; |
9e583fac | 214 | |
ac70caad | 215 | for (p_occ = p_head; (occ = *p_occ) != NULL; ) |
216 | { | |
217 | basic_block bb = new_occ->bb, occ_bb = occ->bb; | |
218 | basic_block dom = nearest_common_dominator (CDI_DOMINATORS, occ_bb, bb); | |
219 | if (dom == bb) | |
220 | { | |
221 | /* BB dominates OCC_BB. OCC becomes NEW_OCC's child: remove OCC | |
222 | from its list. */ | |
223 | *p_occ = occ->next; | |
224 | occ->next = new_occ->children; | |
225 | new_occ->children = occ; | |
226 | ||
227 | /* Try the next block (it may as well be dominated by BB). */ | |
228 | } | |
229 | ||
230 | else if (dom == occ_bb) | |
231 | { | |
232 | /* OCC_BB dominates BB. Tail recurse to look deeper. */ | |
233 | insert_bb (new_occ, dom, &occ->children); | |
234 | return; | |
235 | } | |
236 | ||
237 | else if (dom != idom) | |
238 | { | |
239 | gcc_assert (!dom->aux); | |
240 | ||
241 | /* There is a dominator between IDOM and BB, add it and make | |
242 | two children out of NEW_OCC and OCC. First, remove OCC from | |
243 | its list. */ | |
244 | *p_occ = occ->next; | |
245 | new_occ->next = occ; | |
246 | occ->next = NULL; | |
247 | ||
248 | /* None of the previous blocks has DOM as a dominator: if we tail | |
249 | recursed, we would reexamine them uselessly. Just switch BB with | |
250 | DOM, and go on looking for blocks dominated by DOM. */ | |
251 | new_occ = occ_new (dom, new_occ); | |
252 | } | |
253 | ||
254 | else | |
255 | { | |
256 | /* Nothing special, go on with the next element. */ | |
257 | p_occ = &occ->next; | |
258 | } | |
259 | } | |
260 | ||
261 | /* No place was found as a child of IDOM. Make BB a sibling of IDOM. */ | |
262 | new_occ->next = *p_head; | |
263 | *p_head = new_occ; | |
264 | } | |
265 | ||
266 | /* Register that we found a division in BB. */ | |
267 | ||
268 | static inline void | |
269 | register_division_in (basic_block bb) | |
270 | { | |
271 | struct occurrence *occ; | |
272 | ||
273 | occ = (struct occurrence *) bb->aux; | |
274 | if (!occ) | |
275 | { | |
276 | occ = occ_new (bb, NULL); | |
277 | insert_bb (occ, ENTRY_BLOCK_PTR, &occ_head); | |
278 | } | |
279 | ||
280 | occ->bb_has_division = true; | |
281 | occ->num_divisions++; | |
282 | } | |
283 | ||
284 | ||
285 | /* Compute the number of divisions that postdominate each block in OCC and | |
286 | its children. */ | |
abacb398 | 287 | |
abacb398 | 288 | static void |
ac70caad | 289 | compute_merit (struct occurrence *occ) |
abacb398 | 290 | { |
ac70caad | 291 | struct occurrence *occ_child; |
292 | basic_block dom = occ->bb; | |
abacb398 | 293 | |
ac70caad | 294 | for (occ_child = occ->children; occ_child; occ_child = occ_child->next) |
abacb398 | 295 | { |
ac70caad | 296 | basic_block bb; |
297 | if (occ_child->children) | |
298 | compute_merit (occ_child); | |
299 | ||
300 | if (flag_exceptions) | |
301 | bb = single_noncomplex_succ (dom); | |
302 | else | |
303 | bb = dom; | |
304 | ||
305 | if (dominated_by_p (CDI_POST_DOMINATORS, bb, occ_child->bb)) | |
306 | occ->num_divisions += occ_child->num_divisions; | |
307 | } | |
308 | } | |
309 | ||
310 | ||
311 | /* Return whether USE_STMT is a floating-point division by DEF. */ | |
312 | static inline bool | |
75a70cf9 | 313 | is_division_by (gimple use_stmt, tree def) |
ac70caad | 314 | { |
75a70cf9 | 315 | return is_gimple_assign (use_stmt) |
316 | && gimple_assign_rhs_code (use_stmt) == RDIV_EXPR | |
317 | && gimple_assign_rhs2 (use_stmt) == def | |
119368d7 | 318 | /* Do not recognize x / x as valid division, as we are getting |
319 | confused later by replacing all immediate uses x in such | |
320 | a stmt. */ | |
75a70cf9 | 321 | && gimple_assign_rhs1 (use_stmt) != def; |
ac70caad | 322 | } |
323 | ||
324 | /* Walk the subset of the dominator tree rooted at OCC, setting the | |
325 | RECIP_DEF field to a definition of 1.0 / DEF that can be used in | |
326 | the given basic block. The field may be left NULL, of course, | |
327 | if it is not possible or profitable to do the optimization. | |
328 | ||
329 | DEF_BSI is an iterator pointing at the statement defining DEF. | |
330 | If RECIP_DEF is set, a dominator already has a computation that can | |
331 | be used. */ | |
332 | ||
333 | static void | |
75a70cf9 | 334 | insert_reciprocals (gimple_stmt_iterator *def_gsi, struct occurrence *occ, |
ac70caad | 335 | tree def, tree recip_def, int threshold) |
336 | { | |
75a70cf9 | 337 | tree type; |
338 | gimple new_stmt; | |
339 | gimple_stmt_iterator gsi; | |
ac70caad | 340 | struct occurrence *occ_child; |
341 | ||
342 | if (!recip_def | |
343 | && (occ->bb_has_division || !flag_trapping_math) | |
344 | && occ->num_divisions >= threshold) | |
345 | { | |
346 | /* Make a variable with the replacement and substitute it. */ | |
347 | type = TREE_TYPE (def); | |
072f7ab1 | 348 | recip_def = create_tmp_reg (type, "reciptmp"); |
75a70cf9 | 349 | new_stmt = gimple_build_assign_with_ops (RDIV_EXPR, recip_def, |
350 | build_one_cst (type), def); | |
48e1416a | 351 | |
ac70caad | 352 | if (occ->bb_has_division) |
353 | { | |
354 | /* Case 1: insert before an existing division. */ | |
75a70cf9 | 355 | gsi = gsi_after_labels (occ->bb); |
356 | while (!gsi_end_p (gsi) && !is_division_by (gsi_stmt (gsi), def)) | |
357 | gsi_next (&gsi); | |
ac70caad | 358 | |
75a70cf9 | 359 | gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT); |
ac70caad | 360 | } |
75a70cf9 | 361 | else if (def_gsi && occ->bb == def_gsi->bb) |
685b24f5 | 362 | { |
ac70caad | 363 | /* Case 2: insert right after the definition. Note that this will |
364 | never happen if the definition statement can throw, because in | |
365 | that case the sole successor of the statement's basic block will | |
366 | dominate all the uses as well. */ | |
75a70cf9 | 367 | gsi_insert_after (def_gsi, new_stmt, GSI_NEW_STMT); |
685b24f5 | 368 | } |
ac70caad | 369 | else |
370 | { | |
371 | /* Case 3: insert in a basic block not containing defs/uses. */ | |
75a70cf9 | 372 | gsi = gsi_after_labels (occ->bb); |
373 | gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT); | |
ac70caad | 374 | } |
375 | ||
30c4e60d | 376 | reciprocal_stats.rdivs_inserted++; |
377 | ||
ac70caad | 378 | occ->recip_def_stmt = new_stmt; |
abacb398 | 379 | } |
380 | ||
ac70caad | 381 | occ->recip_def = recip_def; |
382 | for (occ_child = occ->children; occ_child; occ_child = occ_child->next) | |
75a70cf9 | 383 | insert_reciprocals (def_gsi, occ_child, def, recip_def, threshold); |
ac70caad | 384 | } |
385 | ||
386 | ||
387 | /* Replace the division at USE_P with a multiplication by the reciprocal, if | |
388 | possible. */ | |
389 | ||
390 | static inline void | |
391 | replace_reciprocal (use_operand_p use_p) | |
392 | { | |
75a70cf9 | 393 | gimple use_stmt = USE_STMT (use_p); |
394 | basic_block bb = gimple_bb (use_stmt); | |
ac70caad | 395 | struct occurrence *occ = (struct occurrence *) bb->aux; |
396 | ||
0bfd8d5c | 397 | if (optimize_bb_for_speed_p (bb) |
398 | && occ->recip_def && use_stmt != occ->recip_def_stmt) | |
ac70caad | 399 | { |
50aacf4c | 400 | gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt); |
75a70cf9 | 401 | gimple_assign_set_rhs_code (use_stmt, MULT_EXPR); |
ac70caad | 402 | SET_USE (use_p, occ->recip_def); |
50aacf4c | 403 | fold_stmt_inplace (&gsi); |
ac70caad | 404 | update_stmt (use_stmt); |
405 | } | |
406 | } | |
407 | ||
408 | ||
409 | /* Free OCC and return one more "struct occurrence" to be freed. */ | |
410 | ||
411 | static struct occurrence * | |
412 | free_bb (struct occurrence *occ) | |
413 | { | |
414 | struct occurrence *child, *next; | |
415 | ||
416 | /* First get the two pointers hanging off OCC. */ | |
417 | next = occ->next; | |
418 | child = occ->children; | |
419 | occ->bb->aux = NULL; | |
420 | pool_free (occ_pool, occ); | |
421 | ||
422 | /* Now ensure that we don't recurse unless it is necessary. */ | |
423 | if (!child) | |
424 | return next; | |
9e583fac | 425 | else |
ac70caad | 426 | { |
427 | while (next) | |
428 | next = free_bb (next); | |
429 | ||
430 | return child; | |
431 | } | |
432 | } | |
433 | ||
434 | ||
435 | /* Look for floating-point divisions among DEF's uses, and try to | |
436 | replace them by multiplications with the reciprocal. Add | |
437 | as many statements computing the reciprocal as needed. | |
438 | ||
439 | DEF must be a GIMPLE register of a floating-point type. */ | |
440 | ||
441 | static void | |
75a70cf9 | 442 | execute_cse_reciprocals_1 (gimple_stmt_iterator *def_gsi, tree def) |
ac70caad | 443 | { |
444 | use_operand_p use_p; | |
445 | imm_use_iterator use_iter; | |
446 | struct occurrence *occ; | |
447 | int count = 0, threshold; | |
abacb398 | 448 | |
ac70caad | 449 | gcc_assert (FLOAT_TYPE_P (TREE_TYPE (def)) && is_gimple_reg (def)); |
450 | ||
451 | FOR_EACH_IMM_USE_FAST (use_p, use_iter, def) | |
abacb398 | 452 | { |
75a70cf9 | 453 | gimple use_stmt = USE_STMT (use_p); |
ac70caad | 454 | if (is_division_by (use_stmt, def)) |
abacb398 | 455 | { |
75a70cf9 | 456 | register_division_in (gimple_bb (use_stmt)); |
ac70caad | 457 | count++; |
abacb398 | 458 | } |
459 | } | |
48e1416a | 460 | |
ac70caad | 461 | /* Do the expensive part only if we can hope to optimize something. */ |
462 | threshold = targetm.min_divisions_for_recip_mul (TYPE_MODE (TREE_TYPE (def))); | |
463 | if (count >= threshold) | |
464 | { | |
75a70cf9 | 465 | gimple use_stmt; |
ac70caad | 466 | for (occ = occ_head; occ; occ = occ->next) |
467 | { | |
468 | compute_merit (occ); | |
75a70cf9 | 469 | insert_reciprocals (def_gsi, occ, def, NULL, threshold); |
ac70caad | 470 | } |
471 | ||
09aca5bc | 472 | FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, def) |
ac70caad | 473 | { |
ac70caad | 474 | if (is_division_by (use_stmt, def)) |
09aca5bc | 475 | { |
476 | FOR_EACH_IMM_USE_ON_STMT (use_p, use_iter) | |
477 | replace_reciprocal (use_p); | |
478 | } | |
ac70caad | 479 | } |
480 | } | |
481 | ||
482 | for (occ = occ_head; occ; ) | |
483 | occ = free_bb (occ); | |
484 | ||
485 | occ_head = NULL; | |
abacb398 | 486 | } |
487 | ||
ac70caad | 488 | static bool |
489 | gate_cse_reciprocals (void) | |
490 | { | |
0bfd8d5c | 491 | return optimize && flag_reciprocal_math; |
ac70caad | 492 | } |
493 | ||
ac70caad | 494 | /* Go through all the floating-point SSA_NAMEs, and call |
495 | execute_cse_reciprocals_1 on each of them. */ | |
2a1990e9 | 496 | static unsigned int |
abacb398 | 497 | execute_cse_reciprocals (void) |
498 | { | |
499 | basic_block bb; | |
51b60a11 | 500 | tree arg; |
685b24f5 | 501 | |
ac70caad | 502 | occ_pool = create_alloc_pool ("dominators for recip", |
503 | sizeof (struct occurrence), | |
504 | n_basic_blocks / 3 + 1); | |
685b24f5 | 505 | |
30c4e60d | 506 | memset (&reciprocal_stats, 0, sizeof (reciprocal_stats)); |
c136ae61 | 507 | calculate_dominance_info (CDI_DOMINATORS); |
508 | calculate_dominance_info (CDI_POST_DOMINATORS); | |
ac70caad | 509 | |
510 | #ifdef ENABLE_CHECKING | |
511 | FOR_EACH_BB (bb) | |
512 | gcc_assert (!bb->aux); | |
513 | #endif | |
514 | ||
1767a056 | 515 | for (arg = DECL_ARGUMENTS (cfun->decl); arg; arg = DECL_CHAIN (arg)) |
c6dfe037 | 516 | if (FLOAT_TYPE_P (TREE_TYPE (arg)) |
ac70caad | 517 | && is_gimple_reg (arg)) |
c6dfe037 | 518 | { |
519 | tree name = ssa_default_def (cfun, arg); | |
520 | if (name) | |
521 | execute_cse_reciprocals_1 (NULL, name); | |
522 | } | |
51b60a11 | 523 | |
abacb398 | 524 | FOR_EACH_BB (bb) |
525 | { | |
75a70cf9 | 526 | gimple_stmt_iterator gsi; |
527 | gimple phi; | |
528 | tree def; | |
abacb398 | 529 | |
75a70cf9 | 530 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
abacb398 | 531 | { |
75a70cf9 | 532 | phi = gsi_stmt (gsi); |
abacb398 | 533 | def = PHI_RESULT (phi); |
534 | if (FLOAT_TYPE_P (TREE_TYPE (def)) | |
535 | && is_gimple_reg (def)) | |
ac70caad | 536 | execute_cse_reciprocals_1 (NULL, def); |
abacb398 | 537 | } |
538 | ||
75a70cf9 | 539 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
abacb398 | 540 | { |
75a70cf9 | 541 | gimple stmt = gsi_stmt (gsi); |
a0315874 | 542 | |
75a70cf9 | 543 | if (gimple_has_lhs (stmt) |
abacb398 | 544 | && (def = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_DEF)) != NULL |
545 | && FLOAT_TYPE_P (TREE_TYPE (def)) | |
51b60a11 | 546 | && TREE_CODE (def) == SSA_NAME) |
75a70cf9 | 547 | execute_cse_reciprocals_1 (&gsi, def); |
abacb398 | 548 | } |
e174638f | 549 | |
0bfd8d5c | 550 | if (optimize_bb_for_size_p (bb)) |
551 | continue; | |
552 | ||
e174638f | 553 | /* Scan for a/func(b) and convert it to reciprocal a*rfunc(b). */ |
75a70cf9 | 554 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
e174638f | 555 | { |
75a70cf9 | 556 | gimple stmt = gsi_stmt (gsi); |
e174638f | 557 | tree fndecl; |
558 | ||
75a70cf9 | 559 | if (is_gimple_assign (stmt) |
560 | && gimple_assign_rhs_code (stmt) == RDIV_EXPR) | |
e174638f | 561 | { |
75a70cf9 | 562 | tree arg1 = gimple_assign_rhs2 (stmt); |
563 | gimple stmt1; | |
2cd360b6 | 564 | |
565 | if (TREE_CODE (arg1) != SSA_NAME) | |
566 | continue; | |
567 | ||
568 | stmt1 = SSA_NAME_DEF_STMT (arg1); | |
e174638f | 569 | |
75a70cf9 | 570 | if (is_gimple_call (stmt1) |
571 | && gimple_call_lhs (stmt1) | |
572 | && (fndecl = gimple_call_fndecl (stmt1)) | |
e174638f | 573 | && (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL |
574 | || DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_MD)) | |
575 | { | |
576 | enum built_in_function code; | |
774b1cdd | 577 | bool md_code, fail; |
578 | imm_use_iterator ui; | |
579 | use_operand_p use_p; | |
e174638f | 580 | |
581 | code = DECL_FUNCTION_CODE (fndecl); | |
2cd360b6 | 582 | md_code = DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_MD; |
583 | ||
584 | fndecl = targetm.builtin_reciprocal (code, md_code, false); | |
e174638f | 585 | if (!fndecl) |
586 | continue; | |
587 | ||
774b1cdd | 588 | /* Check that all uses of the SSA name are divisions, |
589 | otherwise replacing the defining statement will do | |
590 | the wrong thing. */ | |
591 | fail = false; | |
592 | FOR_EACH_IMM_USE_FAST (use_p, ui, arg1) | |
593 | { | |
594 | gimple stmt2 = USE_STMT (use_p); | |
595 | if (is_gimple_debug (stmt2)) | |
596 | continue; | |
597 | if (!is_gimple_assign (stmt2) | |
598 | || gimple_assign_rhs_code (stmt2) != RDIV_EXPR | |
599 | || gimple_assign_rhs1 (stmt2) == arg1 | |
600 | || gimple_assign_rhs2 (stmt2) != arg1) | |
601 | { | |
602 | fail = true; | |
603 | break; | |
604 | } | |
605 | } | |
606 | if (fail) | |
607 | continue; | |
608 | ||
5fb3d93f | 609 | gimple_replace_lhs (stmt1, arg1); |
0acacf9e | 610 | gimple_call_set_fndecl (stmt1, fndecl); |
e174638f | 611 | update_stmt (stmt1); |
30c4e60d | 612 | reciprocal_stats.rfuncs_inserted++; |
e174638f | 613 | |
774b1cdd | 614 | FOR_EACH_IMM_USE_STMT (stmt, ui, arg1) |
615 | { | |
50aacf4c | 616 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); |
774b1cdd | 617 | gimple_assign_set_rhs_code (stmt, MULT_EXPR); |
50aacf4c | 618 | fold_stmt_inplace (&gsi); |
774b1cdd | 619 | update_stmt (stmt); |
620 | } | |
e174638f | 621 | } |
622 | } | |
623 | } | |
abacb398 | 624 | } |
685b24f5 | 625 | |
30c4e60d | 626 | statistics_counter_event (cfun, "reciprocal divs inserted", |
627 | reciprocal_stats.rdivs_inserted); | |
628 | statistics_counter_event (cfun, "reciprocal functions inserted", | |
629 | reciprocal_stats.rfuncs_inserted); | |
630 | ||
c136ae61 | 631 | free_dominance_info (CDI_DOMINATORS); |
632 | free_dominance_info (CDI_POST_DOMINATORS); | |
ac70caad | 633 | free_alloc_pool (occ_pool); |
2a1990e9 | 634 | return 0; |
abacb398 | 635 | } |
636 | ||
20099e35 | 637 | struct gimple_opt_pass pass_cse_reciprocals = |
abacb398 | 638 | { |
20099e35 | 639 | { |
640 | GIMPLE_PASS, | |
abacb398 | 641 | "recip", /* name */ |
642 | gate_cse_reciprocals, /* gate */ | |
643 | execute_cse_reciprocals, /* execute */ | |
644 | NULL, /* sub */ | |
645 | NULL, /* next */ | |
646 | 0, /* static_pass_number */ | |
0b1615c1 | 647 | TV_NONE, /* tv_id */ |
abacb398 | 648 | PROP_ssa, /* properties_required */ |
649 | 0, /* properties_provided */ | |
650 | 0, /* properties_destroyed */ | |
651 | 0, /* todo_flags_start */ | |
771e2890 | 652 | TODO_update_ssa | TODO_verify_ssa |
20099e35 | 653 | | TODO_verify_stmts /* todo_flags_finish */ |
654 | } | |
abacb398 | 655 | }; |
a0315874 | 656 | |
0d424440 | 657 | /* Records an occurrence at statement USE_STMT in the vector of trees |
a0315874 | 658 | STMTS if it is dominated by *TOP_BB or dominates it or this basic block |
0d424440 | 659 | is not yet initialized. Returns true if the occurrence was pushed on |
a0315874 | 660 | the vector. Adjusts *TOP_BB to be the basic block dominating all |
661 | statements in the vector. */ | |
662 | ||
663 | static bool | |
75a70cf9 | 664 | maybe_record_sincos (VEC(gimple, heap) **stmts, |
665 | basic_block *top_bb, gimple use_stmt) | |
a0315874 | 666 | { |
75a70cf9 | 667 | basic_block use_bb = gimple_bb (use_stmt); |
a0315874 | 668 | if (*top_bb |
669 | && (*top_bb == use_bb | |
670 | || dominated_by_p (CDI_DOMINATORS, use_bb, *top_bb))) | |
75a70cf9 | 671 | VEC_safe_push (gimple, heap, *stmts, use_stmt); |
a0315874 | 672 | else if (!*top_bb |
673 | || dominated_by_p (CDI_DOMINATORS, *top_bb, use_bb)) | |
674 | { | |
75a70cf9 | 675 | VEC_safe_push (gimple, heap, *stmts, use_stmt); |
a0315874 | 676 | *top_bb = use_bb; |
677 | } | |
678 | else | |
679 | return false; | |
680 | ||
681 | return true; | |
682 | } | |
683 | ||
684 | /* Look for sin, cos and cexpi calls with the same argument NAME and | |
685 | create a single call to cexpi CSEing the result in this case. | |
686 | We first walk over all immediate uses of the argument collecting | |
687 | statements that we can CSE in a vector and in a second pass replace | |
688 | the statement rhs with a REALPART or IMAGPART expression on the | |
689 | result of the cexpi call we insert before the use statement that | |
690 | dominates all other candidates. */ | |
691 | ||
4c80086d | 692 | static bool |
a0315874 | 693 | execute_cse_sincos_1 (tree name) |
694 | { | |
75a70cf9 | 695 | gimple_stmt_iterator gsi; |
a0315874 | 696 | imm_use_iterator use_iter; |
75a70cf9 | 697 | tree fndecl, res, type; |
698 | gimple def_stmt, use_stmt, stmt; | |
a0315874 | 699 | int seen_cos = 0, seen_sin = 0, seen_cexpi = 0; |
75a70cf9 | 700 | VEC(gimple, heap) *stmts = NULL; |
a0315874 | 701 | basic_block top_bb = NULL; |
702 | int i; | |
4c80086d | 703 | bool cfg_changed = false; |
a0315874 | 704 | |
705 | type = TREE_TYPE (name); | |
706 | FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, name) | |
707 | { | |
75a70cf9 | 708 | if (gimple_code (use_stmt) != GIMPLE_CALL |
709 | || !gimple_call_lhs (use_stmt) | |
710 | || !(fndecl = gimple_call_fndecl (use_stmt)) | |
a0315874 | 711 | || DECL_BUILT_IN_CLASS (fndecl) != BUILT_IN_NORMAL) |
712 | continue; | |
713 | ||
714 | switch (DECL_FUNCTION_CODE (fndecl)) | |
715 | { | |
716 | CASE_FLT_FN (BUILT_IN_COS): | |
717 | seen_cos |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0; | |
718 | break; | |
719 | ||
720 | CASE_FLT_FN (BUILT_IN_SIN): | |
721 | seen_sin |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0; | |
722 | break; | |
723 | ||
724 | CASE_FLT_FN (BUILT_IN_CEXPI): | |
725 | seen_cexpi |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0; | |
726 | break; | |
727 | ||
728 | default:; | |
729 | } | |
730 | } | |
731 | ||
732 | if (seen_cos + seen_sin + seen_cexpi <= 1) | |
733 | { | |
75a70cf9 | 734 | VEC_free(gimple, heap, stmts); |
4c80086d | 735 | return false; |
a0315874 | 736 | } |
737 | ||
738 | /* Simply insert cexpi at the beginning of top_bb but not earlier than | |
739 | the name def statement. */ | |
740 | fndecl = mathfn_built_in (type, BUILT_IN_CEXPI); | |
741 | if (!fndecl) | |
4c80086d | 742 | return false; |
75a70cf9 | 743 | stmt = gimple_build_call (fndecl, 1, name); |
03d37e4e | 744 | res = make_temp_ssa_name (TREE_TYPE (TREE_TYPE (fndecl)), stmt, "sincostmp"); |
75a70cf9 | 745 | gimple_call_set_lhs (stmt, res); |
746 | ||
a0315874 | 747 | def_stmt = SSA_NAME_DEF_STMT (name); |
8090c12d | 748 | if (!SSA_NAME_IS_DEFAULT_DEF (name) |
75a70cf9 | 749 | && gimple_code (def_stmt) != GIMPLE_PHI |
750 | && gimple_bb (def_stmt) == top_bb) | |
a0315874 | 751 | { |
75a70cf9 | 752 | gsi = gsi_for_stmt (def_stmt); |
753 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); | |
a0315874 | 754 | } |
755 | else | |
756 | { | |
75a70cf9 | 757 | gsi = gsi_after_labels (top_bb); |
758 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
a0315874 | 759 | } |
30c4e60d | 760 | sincos_stats.inserted++; |
a0315874 | 761 | |
762 | /* And adjust the recorded old call sites. */ | |
75a70cf9 | 763 | for (i = 0; VEC_iterate(gimple, stmts, i, use_stmt); ++i) |
a0315874 | 764 | { |
75a70cf9 | 765 | tree rhs = NULL; |
766 | fndecl = gimple_call_fndecl (use_stmt); | |
767 | ||
a0315874 | 768 | switch (DECL_FUNCTION_CODE (fndecl)) |
769 | { | |
770 | CASE_FLT_FN (BUILT_IN_COS): | |
75a70cf9 | 771 | rhs = fold_build1 (REALPART_EXPR, type, res); |
a0315874 | 772 | break; |
773 | ||
774 | CASE_FLT_FN (BUILT_IN_SIN): | |
75a70cf9 | 775 | rhs = fold_build1 (IMAGPART_EXPR, type, res); |
a0315874 | 776 | break; |
777 | ||
778 | CASE_FLT_FN (BUILT_IN_CEXPI): | |
75a70cf9 | 779 | rhs = res; |
a0315874 | 780 | break; |
781 | ||
782 | default:; | |
783 | gcc_unreachable (); | |
784 | } | |
785 | ||
75a70cf9 | 786 | /* Replace call with a copy. */ |
787 | stmt = gimple_build_assign (gimple_call_lhs (use_stmt), rhs); | |
788 | ||
789 | gsi = gsi_for_stmt (use_stmt); | |
4c80086d | 790 | gsi_replace (&gsi, stmt, true); |
791 | if (gimple_purge_dead_eh_edges (gimple_bb (stmt))) | |
792 | cfg_changed = true; | |
a0315874 | 793 | } |
794 | ||
75a70cf9 | 795 | VEC_free(gimple, heap, stmts); |
4c80086d | 796 | |
797 | return cfg_changed; | |
a0315874 | 798 | } |
799 | ||
e9a6c4bc | 800 | /* To evaluate powi(x,n), the floating point value x raised to the |
801 | constant integer exponent n, we use a hybrid algorithm that | |
802 | combines the "window method" with look-up tables. For an | |
803 | introduction to exponentiation algorithms and "addition chains", | |
804 | see section 4.6.3, "Evaluation of Powers" of Donald E. Knuth, | |
805 | "Seminumerical Algorithms", Vol. 2, "The Art of Computer Programming", | |
806 | 3rd Edition, 1998, and Daniel M. Gordon, "A Survey of Fast Exponentiation | |
807 | Methods", Journal of Algorithms, Vol. 27, pp. 129-146, 1998. */ | |
808 | ||
809 | /* Provide a default value for POWI_MAX_MULTS, the maximum number of | |
810 | multiplications to inline before calling the system library's pow | |
811 | function. powi(x,n) requires at worst 2*bits(n)-2 multiplications, | |
812 | so this default never requires calling pow, powf or powl. */ | |
813 | ||
814 | #ifndef POWI_MAX_MULTS | |
815 | #define POWI_MAX_MULTS (2*HOST_BITS_PER_WIDE_INT-2) | |
816 | #endif | |
817 | ||
818 | /* The size of the "optimal power tree" lookup table. All | |
819 | exponents less than this value are simply looked up in the | |
820 | powi_table below. This threshold is also used to size the | |
821 | cache of pseudo registers that hold intermediate results. */ | |
822 | #define POWI_TABLE_SIZE 256 | |
823 | ||
824 | /* The size, in bits of the window, used in the "window method" | |
825 | exponentiation algorithm. This is equivalent to a radix of | |
826 | (1<<POWI_WINDOW_SIZE) in the corresponding "m-ary method". */ | |
827 | #define POWI_WINDOW_SIZE 3 | |
828 | ||
829 | /* The following table is an efficient representation of an | |
830 | "optimal power tree". For each value, i, the corresponding | |
831 | value, j, in the table states than an optimal evaluation | |
832 | sequence for calculating pow(x,i) can be found by evaluating | |
833 | pow(x,j)*pow(x,i-j). An optimal power tree for the first | |
834 | 100 integers is given in Knuth's "Seminumerical algorithms". */ | |
835 | ||
836 | static const unsigned char powi_table[POWI_TABLE_SIZE] = | |
837 | { | |
838 | 0, 1, 1, 2, 2, 3, 3, 4, /* 0 - 7 */ | |
839 | 4, 6, 5, 6, 6, 10, 7, 9, /* 8 - 15 */ | |
840 | 8, 16, 9, 16, 10, 12, 11, 13, /* 16 - 23 */ | |
841 | 12, 17, 13, 18, 14, 24, 15, 26, /* 24 - 31 */ | |
842 | 16, 17, 17, 19, 18, 33, 19, 26, /* 32 - 39 */ | |
843 | 20, 25, 21, 40, 22, 27, 23, 44, /* 40 - 47 */ | |
844 | 24, 32, 25, 34, 26, 29, 27, 44, /* 48 - 55 */ | |
845 | 28, 31, 29, 34, 30, 60, 31, 36, /* 56 - 63 */ | |
846 | 32, 64, 33, 34, 34, 46, 35, 37, /* 64 - 71 */ | |
847 | 36, 65, 37, 50, 38, 48, 39, 69, /* 72 - 79 */ | |
848 | 40, 49, 41, 43, 42, 51, 43, 58, /* 80 - 87 */ | |
849 | 44, 64, 45, 47, 46, 59, 47, 76, /* 88 - 95 */ | |
850 | 48, 65, 49, 66, 50, 67, 51, 66, /* 96 - 103 */ | |
851 | 52, 70, 53, 74, 54, 104, 55, 74, /* 104 - 111 */ | |
852 | 56, 64, 57, 69, 58, 78, 59, 68, /* 112 - 119 */ | |
853 | 60, 61, 61, 80, 62, 75, 63, 68, /* 120 - 127 */ | |
854 | 64, 65, 65, 128, 66, 129, 67, 90, /* 128 - 135 */ | |
855 | 68, 73, 69, 131, 70, 94, 71, 88, /* 136 - 143 */ | |
856 | 72, 128, 73, 98, 74, 132, 75, 121, /* 144 - 151 */ | |
857 | 76, 102, 77, 124, 78, 132, 79, 106, /* 152 - 159 */ | |
858 | 80, 97, 81, 160, 82, 99, 83, 134, /* 160 - 167 */ | |
859 | 84, 86, 85, 95, 86, 160, 87, 100, /* 168 - 175 */ | |
860 | 88, 113, 89, 98, 90, 107, 91, 122, /* 176 - 183 */ | |
861 | 92, 111, 93, 102, 94, 126, 95, 150, /* 184 - 191 */ | |
862 | 96, 128, 97, 130, 98, 133, 99, 195, /* 192 - 199 */ | |
863 | 100, 128, 101, 123, 102, 164, 103, 138, /* 200 - 207 */ | |
864 | 104, 145, 105, 146, 106, 109, 107, 149, /* 208 - 215 */ | |
865 | 108, 200, 109, 146, 110, 170, 111, 157, /* 216 - 223 */ | |
866 | 112, 128, 113, 130, 114, 182, 115, 132, /* 224 - 231 */ | |
867 | 116, 200, 117, 132, 118, 158, 119, 206, /* 232 - 239 */ | |
868 | 120, 240, 121, 162, 122, 147, 123, 152, /* 240 - 247 */ | |
869 | 124, 166, 125, 214, 126, 138, 127, 153, /* 248 - 255 */ | |
870 | }; | |
871 | ||
872 | ||
873 | /* Return the number of multiplications required to calculate | |
874 | powi(x,n) where n is less than POWI_TABLE_SIZE. This is a | |
875 | subroutine of powi_cost. CACHE is an array indicating | |
876 | which exponents have already been calculated. */ | |
877 | ||
878 | static int | |
879 | powi_lookup_cost (unsigned HOST_WIDE_INT n, bool *cache) | |
880 | { | |
881 | /* If we've already calculated this exponent, then this evaluation | |
882 | doesn't require any additional multiplications. */ | |
883 | if (cache[n]) | |
884 | return 0; | |
885 | ||
886 | cache[n] = true; | |
887 | return powi_lookup_cost (n - powi_table[n], cache) | |
888 | + powi_lookup_cost (powi_table[n], cache) + 1; | |
889 | } | |
890 | ||
891 | /* Return the number of multiplications required to calculate | |
892 | powi(x,n) for an arbitrary x, given the exponent N. This | |
893 | function needs to be kept in sync with powi_as_mults below. */ | |
894 | ||
895 | static int | |
896 | powi_cost (HOST_WIDE_INT n) | |
897 | { | |
898 | bool cache[POWI_TABLE_SIZE]; | |
899 | unsigned HOST_WIDE_INT digit; | |
900 | unsigned HOST_WIDE_INT val; | |
901 | int result; | |
902 | ||
903 | if (n == 0) | |
904 | return 0; | |
905 | ||
906 | /* Ignore the reciprocal when calculating the cost. */ | |
907 | val = (n < 0) ? -n : n; | |
908 | ||
909 | /* Initialize the exponent cache. */ | |
910 | memset (cache, 0, POWI_TABLE_SIZE * sizeof (bool)); | |
911 | cache[1] = true; | |
912 | ||
913 | result = 0; | |
914 | ||
915 | while (val >= POWI_TABLE_SIZE) | |
916 | { | |
917 | if (val & 1) | |
918 | { | |
919 | digit = val & ((1 << POWI_WINDOW_SIZE) - 1); | |
920 | result += powi_lookup_cost (digit, cache) | |
921 | + POWI_WINDOW_SIZE + 1; | |
922 | val >>= POWI_WINDOW_SIZE; | |
923 | } | |
924 | else | |
925 | { | |
926 | val >>= 1; | |
927 | result++; | |
928 | } | |
929 | } | |
930 | ||
931 | return result + powi_lookup_cost (val, cache); | |
932 | } | |
933 | ||
934 | /* Recursive subroutine of powi_as_mults. This function takes the | |
935 | array, CACHE, of already calculated exponents and an exponent N and | |
936 | returns a tree that corresponds to CACHE[1]**N, with type TYPE. */ | |
937 | ||
938 | static tree | |
939 | powi_as_mults_1 (gimple_stmt_iterator *gsi, location_t loc, tree type, | |
03d37e4e | 940 | HOST_WIDE_INT n, tree *cache) |
e9a6c4bc | 941 | { |
942 | tree op0, op1, ssa_target; | |
943 | unsigned HOST_WIDE_INT digit; | |
944 | gimple mult_stmt; | |
945 | ||
946 | if (n < POWI_TABLE_SIZE && cache[n]) | |
947 | return cache[n]; | |
948 | ||
03d37e4e | 949 | ssa_target = make_temp_ssa_name (type, NULL, "powmult"); |
e9a6c4bc | 950 | |
951 | if (n < POWI_TABLE_SIZE) | |
952 | { | |
953 | cache[n] = ssa_target; | |
03d37e4e | 954 | op0 = powi_as_mults_1 (gsi, loc, type, n - powi_table[n], cache); |
955 | op1 = powi_as_mults_1 (gsi, loc, type, powi_table[n], cache); | |
e9a6c4bc | 956 | } |
957 | else if (n & 1) | |
958 | { | |
959 | digit = n & ((1 << POWI_WINDOW_SIZE) - 1); | |
03d37e4e | 960 | op0 = powi_as_mults_1 (gsi, loc, type, n - digit, cache); |
961 | op1 = powi_as_mults_1 (gsi, loc, type, digit, cache); | |
e9a6c4bc | 962 | } |
963 | else | |
964 | { | |
03d37e4e | 965 | op0 = powi_as_mults_1 (gsi, loc, type, n >> 1, cache); |
e9a6c4bc | 966 | op1 = op0; |
967 | } | |
968 | ||
969 | mult_stmt = gimple_build_assign_with_ops (MULT_EXPR, ssa_target, op0, op1); | |
ae43b05e | 970 | gimple_set_location (mult_stmt, loc); |
e9a6c4bc | 971 | gsi_insert_before (gsi, mult_stmt, GSI_SAME_STMT); |
972 | ||
973 | return ssa_target; | |
974 | } | |
975 | ||
976 | /* Convert ARG0**N to a tree of multiplications of ARG0 with itself. | |
977 | This function needs to be kept in sync with powi_cost above. */ | |
978 | ||
979 | static tree | |
980 | powi_as_mults (gimple_stmt_iterator *gsi, location_t loc, | |
981 | tree arg0, HOST_WIDE_INT n) | |
982 | { | |
03d37e4e | 983 | tree cache[POWI_TABLE_SIZE], result, type = TREE_TYPE (arg0); |
e9a6c4bc | 984 | gimple div_stmt; |
03d37e4e | 985 | tree target; |
e9a6c4bc | 986 | |
987 | if (n == 0) | |
988 | return build_real (type, dconst1); | |
989 | ||
990 | memset (cache, 0, sizeof (cache)); | |
991 | cache[1] = arg0; | |
992 | ||
03d37e4e | 993 | result = powi_as_mults_1 (gsi, loc, type, (n < 0) ? -n : n, cache); |
e9a6c4bc | 994 | if (n >= 0) |
995 | return result; | |
996 | ||
997 | /* If the original exponent was negative, reciprocate the result. */ | |
03d37e4e | 998 | target = make_temp_ssa_name (type, NULL, "powmult"); |
e9a6c4bc | 999 | div_stmt = gimple_build_assign_with_ops (RDIV_EXPR, target, |
1000 | build_real (type, dconst1), | |
1001 | result); | |
ae43b05e | 1002 | gimple_set_location (div_stmt, loc); |
e9a6c4bc | 1003 | gsi_insert_before (gsi, div_stmt, GSI_SAME_STMT); |
1004 | ||
1005 | return target; | |
1006 | } | |
1007 | ||
1008 | /* ARG0 and N are the two arguments to a powi builtin in GSI with | |
1009 | location info LOC. If the arguments are appropriate, create an | |
1010 | equivalent sequence of statements prior to GSI using an optimal | |
1011 | number of multiplications, and return an expession holding the | |
1012 | result. */ | |
1013 | ||
1014 | static tree | |
1015 | gimple_expand_builtin_powi (gimple_stmt_iterator *gsi, location_t loc, | |
1016 | tree arg0, HOST_WIDE_INT n) | |
1017 | { | |
1018 | /* Avoid largest negative number. */ | |
1019 | if (n != -n | |
1020 | && ((n >= -1 && n <= 2) | |
1021 | || (optimize_function_for_speed_p (cfun) | |
1022 | && powi_cost (n) <= POWI_MAX_MULTS))) | |
1023 | return powi_as_mults (gsi, loc, arg0, n); | |
1024 | ||
1025 | return NULL_TREE; | |
1026 | } | |
1027 | ||
ae43b05e | 1028 | /* Build a gimple call statement that calls FN with argument ARG. |
03d37e4e | 1029 | Set the lhs of the call statement to a fresh SSA name. Insert the |
ae43b05e | 1030 | statement prior to GSI's current position, and return the fresh |
1031 | SSA name. */ | |
1032 | ||
1033 | static tree | |
ca12eb68 | 1034 | build_and_insert_call (gimple_stmt_iterator *gsi, location_t loc, |
03d37e4e | 1035 | tree fn, tree arg) |
ae43b05e | 1036 | { |
1037 | gimple call_stmt; | |
1038 | tree ssa_target; | |
1039 | ||
ae43b05e | 1040 | call_stmt = gimple_build_call (fn, 1, arg); |
03d37e4e | 1041 | ssa_target = make_temp_ssa_name (TREE_TYPE (arg), NULL, "powroot"); |
ae43b05e | 1042 | gimple_set_lhs (call_stmt, ssa_target); |
1043 | gimple_set_location (call_stmt, loc); | |
1044 | gsi_insert_before (gsi, call_stmt, GSI_SAME_STMT); | |
1045 | ||
1046 | return ssa_target; | |
1047 | } | |
1048 | ||
ca12eb68 | 1049 | /* Build a gimple binary operation with the given CODE and arguments |
1050 | ARG0, ARG1, assigning the result to a new SSA name for variable | |
1051 | TARGET. Insert the statement prior to GSI's current position, and | |
1052 | return the fresh SSA name.*/ | |
1053 | ||
1054 | static tree | |
1055 | build_and_insert_binop (gimple_stmt_iterator *gsi, location_t loc, | |
03d37e4e | 1056 | const char *name, enum tree_code code, |
1057 | tree arg0, tree arg1) | |
ca12eb68 | 1058 | { |
03d37e4e | 1059 | tree result = make_temp_ssa_name (TREE_TYPE (arg0), NULL, name); |
ca12eb68 | 1060 | gimple stmt = gimple_build_assign_with_ops (code, result, arg0, arg1); |
1061 | gimple_set_location (stmt, loc); | |
1062 | gsi_insert_before (gsi, stmt, GSI_SAME_STMT); | |
1063 | return result; | |
1064 | } | |
1065 | ||
a5c384c1 | 1066 | /* Build a gimple reference operation with the given CODE and argument |
03d37e4e | 1067 | ARG, assigning the result to a new SSA name of TYPE with NAME. |
a5c384c1 | 1068 | Insert the statement prior to GSI's current position, and return |
1069 | the fresh SSA name. */ | |
1070 | ||
1071 | static inline tree | |
1072 | build_and_insert_ref (gimple_stmt_iterator *gsi, location_t loc, tree type, | |
03d37e4e | 1073 | const char *name, enum tree_code code, tree arg0) |
a5c384c1 | 1074 | { |
03d37e4e | 1075 | tree result = make_temp_ssa_name (type, NULL, name); |
a5c384c1 | 1076 | gimple stmt = gimple_build_assign (result, build1 (code, type, arg0)); |
1077 | gimple_set_location (stmt, loc); | |
1078 | gsi_insert_before (gsi, stmt, GSI_SAME_STMT); | |
1079 | return result; | |
1080 | } | |
1081 | ||
03d37e4e | 1082 | /* Build a gimple assignment to cast VAL to TYPE. Insert the statement |
aff5fb4d | 1083 | prior to GSI's current position, and return the fresh SSA name. */ |
1084 | ||
1085 | static tree | |
1086 | build_and_insert_cast (gimple_stmt_iterator *gsi, location_t loc, | |
03d37e4e | 1087 | tree type, tree val) |
aff5fb4d | 1088 | { |
03d37e4e | 1089 | tree result = make_ssa_name (type, NULL); |
1090 | gimple stmt = gimple_build_assign_with_ops (NOP_EXPR, result, val, NULL_TREE); | |
1091 | gimple_set_location (stmt, loc); | |
1092 | gsi_insert_before (gsi, stmt, GSI_SAME_STMT); | |
1093 | return result; | |
aff5fb4d | 1094 | } |
1095 | ||
e78306af | 1096 | /* ARG0 and ARG1 are the two arguments to a pow builtin call in GSI |
1097 | with location info LOC. If possible, create an equivalent and | |
1098 | less expensive sequence of statements prior to GSI, and return an | |
1099 | expession holding the result. */ | |
1100 | ||
1101 | static tree | |
1102 | gimple_expand_builtin_pow (gimple_stmt_iterator *gsi, location_t loc, | |
1103 | tree arg0, tree arg1) | |
1104 | { | |
ae43b05e | 1105 | REAL_VALUE_TYPE c, cint, dconst1_4, dconst3_4, dconst1_3, dconst1_6; |
ca12eb68 | 1106 | REAL_VALUE_TYPE c2, dconst3; |
e78306af | 1107 | HOST_WIDE_INT n; |
ca12eb68 | 1108 | tree type, sqrtfn, cbrtfn, sqrt_arg0, sqrt_sqrt, result, cbrt_x, powi_cbrt_x; |
ae43b05e | 1109 | enum machine_mode mode; |
1110 | bool hw_sqrt_exists; | |
e78306af | 1111 | |
1112 | /* If the exponent isn't a constant, there's nothing of interest | |
1113 | to be done. */ | |
1114 | if (TREE_CODE (arg1) != REAL_CST) | |
1115 | return NULL_TREE; | |
1116 | ||
ae43b05e | 1117 | /* If the exponent is equivalent to an integer, expand to an optimal |
1118 | multiplication sequence when profitable. */ | |
e78306af | 1119 | c = TREE_REAL_CST (arg1); |
1120 | n = real_to_integer (&c); | |
1121 | real_from_integer (&cint, VOIDmode, n, n < 0 ? -1 : 0, 0); | |
1122 | ||
1123 | if (real_identical (&c, &cint) | |
1124 | && ((n >= -1 && n <= 2) | |
1125 | || (flag_unsafe_math_optimizations | |
1126 | && optimize_insn_for_speed_p () | |
1127 | && powi_cost (n) <= POWI_MAX_MULTS))) | |
1128 | return gimple_expand_builtin_powi (gsi, loc, arg0, n); | |
1129 | ||
ae43b05e | 1130 | /* Attempt various optimizations using sqrt and cbrt. */ |
1131 | type = TREE_TYPE (arg0); | |
1132 | mode = TYPE_MODE (type); | |
1133 | sqrtfn = mathfn_built_in (type, BUILT_IN_SQRT); | |
1134 | ||
1135 | /* Optimize pow(x,0.5) = sqrt(x). This replacement is always safe | |
1136 | unless signed zeros must be maintained. pow(-0,0.5) = +0, while | |
1137 | sqrt(-0) = -0. */ | |
1138 | if (sqrtfn | |
1139 | && REAL_VALUES_EQUAL (c, dconsthalf) | |
1140 | && !HONOR_SIGNED_ZEROS (mode)) | |
03d37e4e | 1141 | return build_and_insert_call (gsi, loc, sqrtfn, arg0); |
ae43b05e | 1142 | |
1143 | /* Optimize pow(x,0.25) = sqrt(sqrt(x)). Assume on most machines that | |
1144 | a builtin sqrt instruction is smaller than a call to pow with 0.25, | |
1145 | so do this optimization even if -Os. Don't do this optimization | |
1146 | if we don't have a hardware sqrt insn. */ | |
1147 | dconst1_4 = dconst1; | |
1148 | SET_REAL_EXP (&dconst1_4, REAL_EXP (&dconst1_4) - 2); | |
a5c384c1 | 1149 | hw_sqrt_exists = optab_handler (sqrt_optab, mode) != CODE_FOR_nothing; |
ae43b05e | 1150 | |
1151 | if (flag_unsafe_math_optimizations | |
1152 | && sqrtfn | |
1153 | && REAL_VALUES_EQUAL (c, dconst1_4) | |
1154 | && hw_sqrt_exists) | |
1155 | { | |
1156 | /* sqrt(x) */ | |
03d37e4e | 1157 | sqrt_arg0 = build_and_insert_call (gsi, loc, sqrtfn, arg0); |
ae43b05e | 1158 | |
1159 | /* sqrt(sqrt(x)) */ | |
03d37e4e | 1160 | return build_and_insert_call (gsi, loc, sqrtfn, sqrt_arg0); |
ae43b05e | 1161 | } |
1162 | ||
1163 | /* Optimize pow(x,0.75) = sqrt(x) * sqrt(sqrt(x)) unless we are | |
1164 | optimizing for space. Don't do this optimization if we don't have | |
1165 | a hardware sqrt insn. */ | |
1166 | real_from_integer (&dconst3_4, VOIDmode, 3, 0, 0); | |
1167 | SET_REAL_EXP (&dconst3_4, REAL_EXP (&dconst3_4) - 2); | |
1168 | ||
1169 | if (flag_unsafe_math_optimizations | |
1170 | && sqrtfn | |
1171 | && optimize_function_for_speed_p (cfun) | |
1172 | && REAL_VALUES_EQUAL (c, dconst3_4) | |
1173 | && hw_sqrt_exists) | |
1174 | { | |
1175 | /* sqrt(x) */ | |
03d37e4e | 1176 | sqrt_arg0 = build_and_insert_call (gsi, loc, sqrtfn, arg0); |
ae43b05e | 1177 | |
1178 | /* sqrt(sqrt(x)) */ | |
03d37e4e | 1179 | sqrt_sqrt = build_and_insert_call (gsi, loc, sqrtfn, sqrt_arg0); |
ae43b05e | 1180 | |
1181 | /* sqrt(x) * sqrt(sqrt(x)) */ | |
03d37e4e | 1182 | return build_and_insert_binop (gsi, loc, "powroot", MULT_EXPR, |
ca12eb68 | 1183 | sqrt_arg0, sqrt_sqrt); |
ae43b05e | 1184 | } |
1185 | ||
1186 | /* Optimize pow(x,1./3.) = cbrt(x). This requires unsafe math | |
1187 | optimizations since 1./3. is not exactly representable. If x | |
1188 | is negative and finite, the correct value of pow(x,1./3.) is | |
1189 | a NaN with the "invalid" exception raised, because the value | |
1190 | of 1./3. actually has an even denominator. The correct value | |
1191 | of cbrt(x) is a negative real value. */ | |
1192 | cbrtfn = mathfn_built_in (type, BUILT_IN_CBRT); | |
1193 | dconst1_3 = real_value_truncate (mode, dconst_third ()); | |
1194 | ||
1195 | if (flag_unsafe_math_optimizations | |
1196 | && cbrtfn | |
0b7ad900 | 1197 | && (gimple_val_nonnegative_real_p (arg0) || !HONOR_NANS (mode)) |
ae43b05e | 1198 | && REAL_VALUES_EQUAL (c, dconst1_3)) |
03d37e4e | 1199 | return build_and_insert_call (gsi, loc, cbrtfn, arg0); |
ae43b05e | 1200 | |
1201 | /* Optimize pow(x,1./6.) = cbrt(sqrt(x)). Don't do this optimization | |
1202 | if we don't have a hardware sqrt insn. */ | |
1203 | dconst1_6 = dconst1_3; | |
1204 | SET_REAL_EXP (&dconst1_6, REAL_EXP (&dconst1_6) - 1); | |
1205 | ||
1206 | if (flag_unsafe_math_optimizations | |
1207 | && sqrtfn | |
1208 | && cbrtfn | |
0b7ad900 | 1209 | && (gimple_val_nonnegative_real_p (arg0) || !HONOR_NANS (mode)) |
ae43b05e | 1210 | && optimize_function_for_speed_p (cfun) |
1211 | && hw_sqrt_exists | |
1212 | && REAL_VALUES_EQUAL (c, dconst1_6)) | |
1213 | { | |
1214 | /* sqrt(x) */ | |
03d37e4e | 1215 | sqrt_arg0 = build_and_insert_call (gsi, loc, sqrtfn, arg0); |
ae43b05e | 1216 | |
1217 | /* cbrt(sqrt(x)) */ | |
03d37e4e | 1218 | return build_and_insert_call (gsi, loc, cbrtfn, sqrt_arg0); |
ca12eb68 | 1219 | } |
1220 | ||
1221 | /* Optimize pow(x,c), where n = 2c for some nonzero integer n, into | |
1222 | ||
1223 | sqrt(x) * powi(x, n/2), n > 0; | |
1224 | 1.0 / (sqrt(x) * powi(x, abs(n/2))), n < 0. | |
1225 | ||
1226 | Do not calculate the powi factor when n/2 = 0. */ | |
1227 | real_arithmetic (&c2, MULT_EXPR, &c, &dconst2); | |
1228 | n = real_to_integer (&c2); | |
1229 | real_from_integer (&cint, VOIDmode, n, n < 0 ? -1 : 0, 0); | |
1230 | ||
1231 | if (flag_unsafe_math_optimizations | |
1232 | && sqrtfn | |
1233 | && real_identical (&c2, &cint)) | |
1234 | { | |
1235 | tree powi_x_ndiv2 = NULL_TREE; | |
1236 | ||
1237 | /* Attempt to fold powi(arg0, abs(n/2)) into multiplies. If not | |
1238 | possible or profitable, give up. Skip the degenerate case when | |
1239 | n is 1 or -1, where the result is always 1. */ | |
b1757d46 | 1240 | if (absu_hwi (n) != 1) |
ca12eb68 | 1241 | { |
5ebd604f | 1242 | powi_x_ndiv2 = gimple_expand_builtin_powi (gsi, loc, arg0, |
1243 | abs_hwi (n / 2)); | |
ca12eb68 | 1244 | if (!powi_x_ndiv2) |
1245 | return NULL_TREE; | |
1246 | } | |
1247 | ||
1248 | /* Calculate sqrt(x). When n is not 1 or -1, multiply it by the | |
1249 | result of the optimal multiply sequence just calculated. */ | |
03d37e4e | 1250 | sqrt_arg0 = build_and_insert_call (gsi, loc, sqrtfn, arg0); |
ca12eb68 | 1251 | |
b1757d46 | 1252 | if (absu_hwi (n) == 1) |
ca12eb68 | 1253 | result = sqrt_arg0; |
1254 | else | |
03d37e4e | 1255 | result = build_and_insert_binop (gsi, loc, "powroot", MULT_EXPR, |
ca12eb68 | 1256 | sqrt_arg0, powi_x_ndiv2); |
1257 | ||
1258 | /* If n is negative, reciprocate the result. */ | |
1259 | if (n < 0) | |
03d37e4e | 1260 | result = build_and_insert_binop (gsi, loc, "powroot", RDIV_EXPR, |
ca12eb68 | 1261 | build_real (type, dconst1), result); |
1262 | return result; | |
1263 | } | |
1264 | ||
1265 | /* Optimize pow(x,c), where 3c = n for some nonzero integer n, into | |
1266 | ||
1267 | powi(x, n/3) * powi(cbrt(x), n%3), n > 0; | |
1268 | 1.0 / (powi(x, abs(n)/3) * powi(cbrt(x), abs(n)%3)), n < 0. | |
1269 | ||
1270 | Do not calculate the first factor when n/3 = 0. As cbrt(x) is | |
1271 | different from pow(x, 1./3.) due to rounding and behavior with | |
1272 | negative x, we need to constrain this transformation to unsafe | |
1273 | math and positive x or finite math. */ | |
1274 | real_from_integer (&dconst3, VOIDmode, 3, 0, 0); | |
1275 | real_arithmetic (&c2, MULT_EXPR, &c, &dconst3); | |
1276 | real_round (&c2, mode, &c2); | |
1277 | n = real_to_integer (&c2); | |
1278 | real_from_integer (&cint, VOIDmode, n, n < 0 ? -1 : 0, 0); | |
1279 | real_arithmetic (&c2, RDIV_EXPR, &cint, &dconst3); | |
1280 | real_convert (&c2, mode, &c2); | |
1281 | ||
1282 | if (flag_unsafe_math_optimizations | |
1283 | && cbrtfn | |
0b7ad900 | 1284 | && (gimple_val_nonnegative_real_p (arg0) || !HONOR_NANS (mode)) |
ca12eb68 | 1285 | && real_identical (&c2, &c) |
1286 | && optimize_function_for_speed_p (cfun) | |
1287 | && powi_cost (n / 3) <= POWI_MAX_MULTS) | |
1288 | { | |
1289 | tree powi_x_ndiv3 = NULL_TREE; | |
1290 | ||
1291 | /* Attempt to fold powi(arg0, abs(n/3)) into multiplies. If not | |
1292 | possible or profitable, give up. Skip the degenerate case when | |
1293 | abs(n) < 3, where the result is always 1. */ | |
b1757d46 | 1294 | if (absu_hwi (n) >= 3) |
ca12eb68 | 1295 | { |
1296 | powi_x_ndiv3 = gimple_expand_builtin_powi (gsi, loc, arg0, | |
5ebd604f | 1297 | abs_hwi (n / 3)); |
ca12eb68 | 1298 | if (!powi_x_ndiv3) |
1299 | return NULL_TREE; | |
1300 | } | |
1301 | ||
1302 | /* Calculate powi(cbrt(x), n%3). Don't use gimple_expand_builtin_powi | |
1303 | as that creates an unnecessary variable. Instead, just produce | |
1304 | either cbrt(x) or cbrt(x) * cbrt(x). */ | |
03d37e4e | 1305 | cbrt_x = build_and_insert_call (gsi, loc, cbrtfn, arg0); |
ca12eb68 | 1306 | |
b1757d46 | 1307 | if (absu_hwi (n) % 3 == 1) |
ca12eb68 | 1308 | powi_cbrt_x = cbrt_x; |
1309 | else | |
03d37e4e | 1310 | powi_cbrt_x = build_and_insert_binop (gsi, loc, "powroot", MULT_EXPR, |
ca12eb68 | 1311 | cbrt_x, cbrt_x); |
1312 | ||
1313 | /* Multiply the two subexpressions, unless powi(x,abs(n)/3) = 1. */ | |
b1757d46 | 1314 | if (absu_hwi (n) < 3) |
ca12eb68 | 1315 | result = powi_cbrt_x; |
1316 | else | |
03d37e4e | 1317 | result = build_and_insert_binop (gsi, loc, "powroot", MULT_EXPR, |
ca12eb68 | 1318 | powi_x_ndiv3, powi_cbrt_x); |
1319 | ||
1320 | /* If n is negative, reciprocate the result. */ | |
1321 | if (n < 0) | |
03d37e4e | 1322 | result = build_and_insert_binop (gsi, loc, "powroot", RDIV_EXPR, |
ca12eb68 | 1323 | build_real (type, dconst1), result); |
1324 | ||
1325 | return result; | |
ae43b05e | 1326 | } |
1327 | ||
ca12eb68 | 1328 | /* No optimizations succeeded. */ |
e78306af | 1329 | return NULL_TREE; |
1330 | } | |
1331 | ||
a5c384c1 | 1332 | /* ARG is the argument to a cabs builtin call in GSI with location info |
1333 | LOC. Create a sequence of statements prior to GSI that calculates | |
1334 | sqrt(R*R + I*I), where R and I are the real and imaginary components | |
1335 | of ARG, respectively. Return an expression holding the result. */ | |
1336 | ||
1337 | static tree | |
1338 | gimple_expand_builtin_cabs (gimple_stmt_iterator *gsi, location_t loc, tree arg) | |
1339 | { | |
03d37e4e | 1340 | tree real_part, imag_part, addend1, addend2, sum, result; |
a5c384c1 | 1341 | tree type = TREE_TYPE (TREE_TYPE (arg)); |
1342 | tree sqrtfn = mathfn_built_in (type, BUILT_IN_SQRT); | |
1343 | enum machine_mode mode = TYPE_MODE (type); | |
1344 | ||
1345 | if (!flag_unsafe_math_optimizations | |
1346 | || !optimize_bb_for_speed_p (gimple_bb (gsi_stmt (*gsi))) | |
1347 | || !sqrtfn | |
1348 | || optab_handler (sqrt_optab, mode) == CODE_FOR_nothing) | |
1349 | return NULL_TREE; | |
1350 | ||
03d37e4e | 1351 | real_part = build_and_insert_ref (gsi, loc, type, "cabs", |
a5c384c1 | 1352 | REALPART_EXPR, arg); |
03d37e4e | 1353 | addend1 = build_and_insert_binop (gsi, loc, "cabs", MULT_EXPR, |
a5c384c1 | 1354 | real_part, real_part); |
03d37e4e | 1355 | imag_part = build_and_insert_ref (gsi, loc, type, "cabs", |
a5c384c1 | 1356 | IMAGPART_EXPR, arg); |
03d37e4e | 1357 | addend2 = build_and_insert_binop (gsi, loc, "cabs", MULT_EXPR, |
a5c384c1 | 1358 | imag_part, imag_part); |
03d37e4e | 1359 | sum = build_and_insert_binop (gsi, loc, "cabs", PLUS_EXPR, addend1, addend2); |
1360 | result = build_and_insert_call (gsi, loc, sqrtfn, sum); | |
a5c384c1 | 1361 | |
1362 | return result; | |
1363 | } | |
1364 | ||
a0315874 | 1365 | /* Go through all calls to sin, cos and cexpi and call execute_cse_sincos_1 |
e9a6c4bc | 1366 | on the SSA_NAME argument of each of them. Also expand powi(x,n) into |
1367 | an optimal number of multiplies, when n is a constant. */ | |
a0315874 | 1368 | |
1369 | static unsigned int | |
1370 | execute_cse_sincos (void) | |
1371 | { | |
1372 | basic_block bb; | |
4c80086d | 1373 | bool cfg_changed = false; |
a0315874 | 1374 | |
1375 | calculate_dominance_info (CDI_DOMINATORS); | |
30c4e60d | 1376 | memset (&sincos_stats, 0, sizeof (sincos_stats)); |
a0315874 | 1377 | |
1378 | FOR_EACH_BB (bb) | |
1379 | { | |
75a70cf9 | 1380 | gimple_stmt_iterator gsi; |
a0315874 | 1381 | |
75a70cf9 | 1382 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
a0315874 | 1383 | { |
75a70cf9 | 1384 | gimple stmt = gsi_stmt (gsi); |
a0315874 | 1385 | tree fndecl; |
1386 | ||
75a70cf9 | 1387 | if (is_gimple_call (stmt) |
1388 | && gimple_call_lhs (stmt) | |
1389 | && (fndecl = gimple_call_fndecl (stmt)) | |
a0315874 | 1390 | && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) |
1391 | { | |
e9a6c4bc | 1392 | tree arg, arg0, arg1, result; |
1393 | HOST_WIDE_INT n; | |
1394 | location_t loc; | |
a0315874 | 1395 | |
1396 | switch (DECL_FUNCTION_CODE (fndecl)) | |
1397 | { | |
1398 | CASE_FLT_FN (BUILT_IN_COS): | |
1399 | CASE_FLT_FN (BUILT_IN_SIN): | |
1400 | CASE_FLT_FN (BUILT_IN_CEXPI): | |
d312d7df | 1401 | /* Make sure we have either sincos or cexp. */ |
1402 | if (!TARGET_HAS_SINCOS && !TARGET_C99_FUNCTIONS) | |
1403 | break; | |
1404 | ||
75a70cf9 | 1405 | arg = gimple_call_arg (stmt, 0); |
a0315874 | 1406 | if (TREE_CODE (arg) == SSA_NAME) |
4c80086d | 1407 | cfg_changed |= execute_cse_sincos_1 (arg); |
a0315874 | 1408 | break; |
1409 | ||
e78306af | 1410 | CASE_FLT_FN (BUILT_IN_POW): |
1411 | arg0 = gimple_call_arg (stmt, 0); | |
1412 | arg1 = gimple_call_arg (stmt, 1); | |
1413 | ||
1414 | loc = gimple_location (stmt); | |
1415 | result = gimple_expand_builtin_pow (&gsi, loc, arg0, arg1); | |
1416 | ||
1417 | if (result) | |
1418 | { | |
1419 | tree lhs = gimple_get_lhs (stmt); | |
1420 | gimple new_stmt = gimple_build_assign (lhs, result); | |
1421 | gimple_set_location (new_stmt, loc); | |
1422 | unlink_stmt_vdef (stmt); | |
1423 | gsi_replace (&gsi, new_stmt, true); | |
bc8a8451 | 1424 | if (gimple_vdef (stmt)) |
1425 | release_ssa_name (gimple_vdef (stmt)); | |
e78306af | 1426 | } |
1427 | break; | |
1428 | ||
e9a6c4bc | 1429 | CASE_FLT_FN (BUILT_IN_POWI): |
1430 | arg0 = gimple_call_arg (stmt, 0); | |
1431 | arg1 = gimple_call_arg (stmt, 1); | |
1432 | if (!host_integerp (arg1, 0)) | |
1433 | break; | |
1434 | ||
1435 | n = TREE_INT_CST_LOW (arg1); | |
1436 | loc = gimple_location (stmt); | |
1437 | result = gimple_expand_builtin_powi (&gsi, loc, arg0, n); | |
1438 | ||
1439 | if (result) | |
1440 | { | |
1441 | tree lhs = gimple_get_lhs (stmt); | |
1442 | gimple new_stmt = gimple_build_assign (lhs, result); | |
1443 | gimple_set_location (new_stmt, loc); | |
a5c384c1 | 1444 | unlink_stmt_vdef (stmt); |
1445 | gsi_replace (&gsi, new_stmt, true); | |
bc8a8451 | 1446 | if (gimple_vdef (stmt)) |
1447 | release_ssa_name (gimple_vdef (stmt)); | |
a5c384c1 | 1448 | } |
1449 | break; | |
1450 | ||
1451 | CASE_FLT_FN (BUILT_IN_CABS): | |
1452 | arg0 = gimple_call_arg (stmt, 0); | |
1453 | loc = gimple_location (stmt); | |
1454 | result = gimple_expand_builtin_cabs (&gsi, loc, arg0); | |
1455 | ||
1456 | if (result) | |
1457 | { | |
1458 | tree lhs = gimple_get_lhs (stmt); | |
1459 | gimple new_stmt = gimple_build_assign (lhs, result); | |
1460 | gimple_set_location (new_stmt, loc); | |
e9a6c4bc | 1461 | unlink_stmt_vdef (stmt); |
1462 | gsi_replace (&gsi, new_stmt, true); | |
bc8a8451 | 1463 | if (gimple_vdef (stmt)) |
1464 | release_ssa_name (gimple_vdef (stmt)); | |
e9a6c4bc | 1465 | } |
1466 | break; | |
1467 | ||
a0315874 | 1468 | default:; |
1469 | } | |
1470 | } | |
1471 | } | |
1472 | } | |
1473 | ||
30c4e60d | 1474 | statistics_counter_event (cfun, "sincos statements inserted", |
1475 | sincos_stats.inserted); | |
1476 | ||
a0315874 | 1477 | free_dominance_info (CDI_DOMINATORS); |
4c80086d | 1478 | return cfg_changed ? TODO_cleanup_cfg : 0; |
a0315874 | 1479 | } |
1480 | ||
1481 | static bool | |
1482 | gate_cse_sincos (void) | |
1483 | { | |
e9a6c4bc | 1484 | /* We no longer require either sincos or cexp, since powi expansion |
1485 | piggybacks on this pass. */ | |
1486 | return optimize; | |
a0315874 | 1487 | } |
1488 | ||
20099e35 | 1489 | struct gimple_opt_pass pass_cse_sincos = |
a0315874 | 1490 | { |
20099e35 | 1491 | { |
1492 | GIMPLE_PASS, | |
a0315874 | 1493 | "sincos", /* name */ |
1494 | gate_cse_sincos, /* gate */ | |
1495 | execute_cse_sincos, /* execute */ | |
1496 | NULL, /* sub */ | |
1497 | NULL, /* next */ | |
1498 | 0, /* static_pass_number */ | |
0b1615c1 | 1499 | TV_NONE, /* tv_id */ |
a0315874 | 1500 | PROP_ssa, /* properties_required */ |
1501 | 0, /* properties_provided */ | |
1502 | 0, /* properties_destroyed */ | |
1503 | 0, /* todo_flags_start */ | |
771e2890 | 1504 | TODO_update_ssa | TODO_verify_ssa |
20099e35 | 1505 | | TODO_verify_stmts /* todo_flags_finish */ |
1506 | } | |
a0315874 | 1507 | }; |
e174638f | 1508 | |
84cc784c | 1509 | /* A symbolic number is used to detect byte permutation and selection |
1510 | patterns. Therefore the field N contains an artificial number | |
1511 | consisting of byte size markers: | |
1512 | ||
1513 | 0 - byte has the value 0 | |
1514 | 1..size - byte contains the content of the byte | |
1515 | number indexed with that value minus one */ | |
1516 | ||
1517 | struct symbolic_number { | |
1518 | unsigned HOST_WIDEST_INT n; | |
1519 | int size; | |
1520 | }; | |
1521 | ||
1522 | /* Perform a SHIFT or ROTATE operation by COUNT bits on symbolic | |
1523 | number N. Return false if the requested operation is not permitted | |
1524 | on a symbolic number. */ | |
1525 | ||
1526 | static inline bool | |
1527 | do_shift_rotate (enum tree_code code, | |
1528 | struct symbolic_number *n, | |
1529 | int count) | |
1530 | { | |
1531 | if (count % 8 != 0) | |
1532 | return false; | |
1533 | ||
1534 | /* Zero out the extra bits of N in order to avoid them being shifted | |
1535 | into the significant bits. */ | |
1536 | if (n->size < (int)sizeof (HOST_WIDEST_INT)) | |
1537 | n->n &= ((unsigned HOST_WIDEST_INT)1 << (n->size * BITS_PER_UNIT)) - 1; | |
1538 | ||
1539 | switch (code) | |
1540 | { | |
1541 | case LSHIFT_EXPR: | |
1542 | n->n <<= count; | |
1543 | break; | |
1544 | case RSHIFT_EXPR: | |
1545 | n->n >>= count; | |
1546 | break; | |
1547 | case LROTATE_EXPR: | |
1548 | n->n = (n->n << count) | (n->n >> ((n->size * BITS_PER_UNIT) - count)); | |
1549 | break; | |
1550 | case RROTATE_EXPR: | |
1551 | n->n = (n->n >> count) | (n->n << ((n->size * BITS_PER_UNIT) - count)); | |
1552 | break; | |
1553 | default: | |
1554 | return false; | |
1555 | } | |
0f09ed00 | 1556 | /* Zero unused bits for size. */ |
1557 | if (n->size < (int)sizeof (HOST_WIDEST_INT)) | |
1558 | n->n &= ((unsigned HOST_WIDEST_INT)1 << (n->size * BITS_PER_UNIT)) - 1; | |
84cc784c | 1559 | return true; |
1560 | } | |
1561 | ||
1562 | /* Perform sanity checking for the symbolic number N and the gimple | |
1563 | statement STMT. */ | |
1564 | ||
1565 | static inline bool | |
1566 | verify_symbolic_number_p (struct symbolic_number *n, gimple stmt) | |
1567 | { | |
1568 | tree lhs_type; | |
1569 | ||
1570 | lhs_type = gimple_expr_type (stmt); | |
1571 | ||
1572 | if (TREE_CODE (lhs_type) != INTEGER_TYPE) | |
1573 | return false; | |
1574 | ||
1575 | if (TYPE_PRECISION (lhs_type) != n->size * BITS_PER_UNIT) | |
1576 | return false; | |
1577 | ||
1578 | return true; | |
1579 | } | |
1580 | ||
1581 | /* find_bswap_1 invokes itself recursively with N and tries to perform | |
1582 | the operation given by the rhs of STMT on the result. If the | |
1583 | operation could successfully be executed the function returns the | |
1584 | tree expression of the source operand and NULL otherwise. */ | |
1585 | ||
1586 | static tree | |
1587 | find_bswap_1 (gimple stmt, struct symbolic_number *n, int limit) | |
1588 | { | |
1589 | enum tree_code code; | |
1590 | tree rhs1, rhs2 = NULL; | |
1591 | gimple rhs1_stmt, rhs2_stmt; | |
1592 | tree source_expr1; | |
1593 | enum gimple_rhs_class rhs_class; | |
1594 | ||
1595 | if (!limit || !is_gimple_assign (stmt)) | |
1596 | return NULL_TREE; | |
1597 | ||
1598 | rhs1 = gimple_assign_rhs1 (stmt); | |
1599 | ||
1600 | if (TREE_CODE (rhs1) != SSA_NAME) | |
1601 | return NULL_TREE; | |
1602 | ||
1603 | code = gimple_assign_rhs_code (stmt); | |
1604 | rhs_class = gimple_assign_rhs_class (stmt); | |
1605 | rhs1_stmt = SSA_NAME_DEF_STMT (rhs1); | |
1606 | ||
1607 | if (rhs_class == GIMPLE_BINARY_RHS) | |
1608 | rhs2 = gimple_assign_rhs2 (stmt); | |
1609 | ||
1610 | /* Handle unary rhs and binary rhs with integer constants as second | |
1611 | operand. */ | |
1612 | ||
1613 | if (rhs_class == GIMPLE_UNARY_RHS | |
1614 | || (rhs_class == GIMPLE_BINARY_RHS | |
1615 | && TREE_CODE (rhs2) == INTEGER_CST)) | |
1616 | { | |
1617 | if (code != BIT_AND_EXPR | |
1618 | && code != LSHIFT_EXPR | |
1619 | && code != RSHIFT_EXPR | |
1620 | && code != LROTATE_EXPR | |
1621 | && code != RROTATE_EXPR | |
1622 | && code != NOP_EXPR | |
1623 | && code != CONVERT_EXPR) | |
1624 | return NULL_TREE; | |
1625 | ||
1626 | source_expr1 = find_bswap_1 (rhs1_stmt, n, limit - 1); | |
1627 | ||
1628 | /* If find_bswap_1 returned NULL STMT is a leaf node and we have | |
1629 | to initialize the symbolic number. */ | |
1630 | if (!source_expr1) | |
1631 | { | |
1632 | /* Set up the symbolic number N by setting each byte to a | |
1633 | value between 1 and the byte size of rhs1. The highest | |
f9a210c9 | 1634 | order byte is set to n->size and the lowest order |
1635 | byte to 1. */ | |
84cc784c | 1636 | n->size = TYPE_PRECISION (TREE_TYPE (rhs1)); |
1637 | if (n->size % BITS_PER_UNIT != 0) | |
1638 | return NULL_TREE; | |
1639 | n->size /= BITS_PER_UNIT; | |
1640 | n->n = (sizeof (HOST_WIDEST_INT) < 8 ? 0 : | |
f9a210c9 | 1641 | (unsigned HOST_WIDEST_INT)0x08070605 << 32 | 0x04030201); |
1642 | ||
1643 | if (n->size < (int)sizeof (HOST_WIDEST_INT)) | |
1644 | n->n &= ((unsigned HOST_WIDEST_INT)1 << | |
1645 | (n->size * BITS_PER_UNIT)) - 1; | |
84cc784c | 1646 | |
1647 | source_expr1 = rhs1; | |
1648 | } | |
1649 | ||
1650 | switch (code) | |
1651 | { | |
1652 | case BIT_AND_EXPR: | |
1653 | { | |
1654 | int i; | |
1655 | unsigned HOST_WIDEST_INT val = widest_int_cst_value (rhs2); | |
1656 | unsigned HOST_WIDEST_INT tmp = val; | |
1657 | ||
1658 | /* Only constants masking full bytes are allowed. */ | |
1659 | for (i = 0; i < n->size; i++, tmp >>= BITS_PER_UNIT) | |
1660 | if ((tmp & 0xff) != 0 && (tmp & 0xff) != 0xff) | |
1661 | return NULL_TREE; | |
1662 | ||
1663 | n->n &= val; | |
1664 | } | |
1665 | break; | |
1666 | case LSHIFT_EXPR: | |
1667 | case RSHIFT_EXPR: | |
1668 | case LROTATE_EXPR: | |
1669 | case RROTATE_EXPR: | |
1670 | if (!do_shift_rotate (code, n, (int)TREE_INT_CST_LOW (rhs2))) | |
1671 | return NULL_TREE; | |
1672 | break; | |
1673 | CASE_CONVERT: | |
1674 | { | |
1675 | int type_size; | |
1676 | ||
1677 | type_size = TYPE_PRECISION (gimple_expr_type (stmt)); | |
1678 | if (type_size % BITS_PER_UNIT != 0) | |
1679 | return NULL_TREE; | |
1680 | ||
84cc784c | 1681 | if (type_size / BITS_PER_UNIT < (int)(sizeof (HOST_WIDEST_INT))) |
1682 | { | |
1683 | /* If STMT casts to a smaller type mask out the bits not | |
1684 | belonging to the target type. */ | |
84cc784c | 1685 | n->n &= ((unsigned HOST_WIDEST_INT)1 << type_size) - 1; |
1686 | } | |
f9a210c9 | 1687 | n->size = type_size / BITS_PER_UNIT; |
84cc784c | 1688 | } |
1689 | break; | |
1690 | default: | |
1691 | return NULL_TREE; | |
1692 | }; | |
1693 | return verify_symbolic_number_p (n, stmt) ? source_expr1 : NULL; | |
1694 | } | |
1695 | ||
1696 | /* Handle binary rhs. */ | |
1697 | ||
1698 | if (rhs_class == GIMPLE_BINARY_RHS) | |
1699 | { | |
1700 | struct symbolic_number n1, n2; | |
1701 | tree source_expr2; | |
1702 | ||
1703 | if (code != BIT_IOR_EXPR) | |
1704 | return NULL_TREE; | |
1705 | ||
1706 | if (TREE_CODE (rhs2) != SSA_NAME) | |
1707 | return NULL_TREE; | |
1708 | ||
1709 | rhs2_stmt = SSA_NAME_DEF_STMT (rhs2); | |
1710 | ||
1711 | switch (code) | |
1712 | { | |
1713 | case BIT_IOR_EXPR: | |
1714 | source_expr1 = find_bswap_1 (rhs1_stmt, &n1, limit - 1); | |
1715 | ||
1716 | if (!source_expr1) | |
1717 | return NULL_TREE; | |
1718 | ||
1719 | source_expr2 = find_bswap_1 (rhs2_stmt, &n2, limit - 1); | |
1720 | ||
1721 | if (source_expr1 != source_expr2 | |
1722 | || n1.size != n2.size) | |
1723 | return NULL_TREE; | |
1724 | ||
1725 | n->size = n1.size; | |
1726 | n->n = n1.n | n2.n; | |
1727 | ||
1728 | if (!verify_symbolic_number_p (n, stmt)) | |
1729 | return NULL_TREE; | |
1730 | ||
1731 | break; | |
1732 | default: | |
1733 | return NULL_TREE; | |
1734 | } | |
1735 | return source_expr1; | |
1736 | } | |
1737 | return NULL_TREE; | |
1738 | } | |
1739 | ||
1740 | /* Check if STMT completes a bswap implementation consisting of ORs, | |
1741 | SHIFTs and ANDs. Return the source tree expression on which the | |
1742 | byte swap is performed and NULL if no bswap was found. */ | |
1743 | ||
1744 | static tree | |
1745 | find_bswap (gimple stmt) | |
1746 | { | |
1747 | /* The number which the find_bswap result should match in order to | |
f9a210c9 | 1748 | have a full byte swap. The number is shifted to the left according |
1749 | to the size of the symbolic number before using it. */ | |
84cc784c | 1750 | unsigned HOST_WIDEST_INT cmp = |
1751 | sizeof (HOST_WIDEST_INT) < 8 ? 0 : | |
f9a210c9 | 1752 | (unsigned HOST_WIDEST_INT)0x01020304 << 32 | 0x05060708; |
84cc784c | 1753 | |
1754 | struct symbolic_number n; | |
1755 | tree source_expr; | |
0f09ed00 | 1756 | int limit; |
84cc784c | 1757 | |
9bc1852a | 1758 | /* The last parameter determines the depth search limit. It usually |
1759 | correlates directly to the number of bytes to be touched. We | |
0f09ed00 | 1760 | increase that number by three here in order to also |
1761 | cover signed -> unsigned converions of the src operand as can be seen | |
1762 | in libgcc, and for initial shift/and operation of the src operand. */ | |
1763 | limit = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (gimple_expr_type (stmt))); | |
1764 | limit += 1 + (int) ceil_log2 ((unsigned HOST_WIDE_INT) limit); | |
1765 | source_expr = find_bswap_1 (stmt, &n, limit); | |
84cc784c | 1766 | |
1767 | if (!source_expr) | |
1768 | return NULL_TREE; | |
1769 | ||
1770 | /* Zero out the extra bits of N and CMP. */ | |
1771 | if (n.size < (int)sizeof (HOST_WIDEST_INT)) | |
1772 | { | |
1773 | unsigned HOST_WIDEST_INT mask = | |
1774 | ((unsigned HOST_WIDEST_INT)1 << (n.size * BITS_PER_UNIT)) - 1; | |
1775 | ||
1776 | n.n &= mask; | |
f9a210c9 | 1777 | cmp >>= (sizeof (HOST_WIDEST_INT) - n.size) * BITS_PER_UNIT; |
84cc784c | 1778 | } |
1779 | ||
1780 | /* A complete byte swap should make the symbolic number to start | |
1781 | with the largest digit in the highest order byte. */ | |
1782 | if (cmp != n.n) | |
1783 | return NULL_TREE; | |
1784 | ||
1785 | return source_expr; | |
1786 | } | |
1787 | ||
1788 | /* Find manual byte swap implementations and turn them into a bswap | |
1789 | builtin invokation. */ | |
1790 | ||
1791 | static unsigned int | |
1792 | execute_optimize_bswap (void) | |
1793 | { | |
1794 | basic_block bb; | |
1795 | bool bswap32_p, bswap64_p; | |
1796 | bool changed = false; | |
0af25806 | 1797 | tree bswap32_type = NULL_TREE, bswap64_type = NULL_TREE; |
84cc784c | 1798 | |
1799 | if (BITS_PER_UNIT != 8) | |
1800 | return 0; | |
1801 | ||
1802 | if (sizeof (HOST_WIDEST_INT) < 8) | |
1803 | return 0; | |
1804 | ||
b9a16870 | 1805 | bswap32_p = (builtin_decl_explicit_p (BUILT_IN_BSWAP32) |
d6bf3b14 | 1806 | && optab_handler (bswap_optab, SImode) != CODE_FOR_nothing); |
b9a16870 | 1807 | bswap64_p = (builtin_decl_explicit_p (BUILT_IN_BSWAP64) |
d6bf3b14 | 1808 | && (optab_handler (bswap_optab, DImode) != CODE_FOR_nothing |
3328b1fb | 1809 | || (bswap32_p && word_mode == SImode))); |
84cc784c | 1810 | |
1811 | if (!bswap32_p && !bswap64_p) | |
1812 | return 0; | |
1813 | ||
0af25806 | 1814 | /* Determine the argument type of the builtins. The code later on |
1815 | assumes that the return and argument type are the same. */ | |
1816 | if (bswap32_p) | |
1817 | { | |
b9a16870 | 1818 | tree fndecl = builtin_decl_explicit (BUILT_IN_BSWAP32); |
0af25806 | 1819 | bswap32_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl))); |
1820 | } | |
1821 | ||
1822 | if (bswap64_p) | |
1823 | { | |
b9a16870 | 1824 | tree fndecl = builtin_decl_explicit (BUILT_IN_BSWAP64); |
0af25806 | 1825 | bswap64_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl))); |
1826 | } | |
1827 | ||
30c4e60d | 1828 | memset (&bswap_stats, 0, sizeof (bswap_stats)); |
1829 | ||
84cc784c | 1830 | FOR_EACH_BB (bb) |
1831 | { | |
1832 | gimple_stmt_iterator gsi; | |
1833 | ||
0ec31268 | 1834 | /* We do a reverse scan for bswap patterns to make sure we get the |
1835 | widest match. As bswap pattern matching doesn't handle | |
1836 | previously inserted smaller bswap replacements as sub- | |
1837 | patterns, the wider variant wouldn't be detected. */ | |
1838 | for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi)) | |
84cc784c | 1839 | { |
1840 | gimple stmt = gsi_stmt (gsi); | |
0af25806 | 1841 | tree bswap_src, bswap_type; |
1842 | tree bswap_tmp; | |
84cc784c | 1843 | tree fndecl = NULL_TREE; |
1844 | int type_size; | |
1845 | gimple call; | |
1846 | ||
1847 | if (!is_gimple_assign (stmt) | |
1848 | || gimple_assign_rhs_code (stmt) != BIT_IOR_EXPR) | |
1849 | continue; | |
1850 | ||
1851 | type_size = TYPE_PRECISION (gimple_expr_type (stmt)); | |
1852 | ||
1853 | switch (type_size) | |
1854 | { | |
1855 | case 32: | |
1856 | if (bswap32_p) | |
0af25806 | 1857 | { |
b9a16870 | 1858 | fndecl = builtin_decl_explicit (BUILT_IN_BSWAP32); |
0af25806 | 1859 | bswap_type = bswap32_type; |
1860 | } | |
84cc784c | 1861 | break; |
1862 | case 64: | |
1863 | if (bswap64_p) | |
0af25806 | 1864 | { |
b9a16870 | 1865 | fndecl = builtin_decl_explicit (BUILT_IN_BSWAP64); |
0af25806 | 1866 | bswap_type = bswap64_type; |
1867 | } | |
84cc784c | 1868 | break; |
1869 | default: | |
1870 | continue; | |
1871 | } | |
1872 | ||
1873 | if (!fndecl) | |
1874 | continue; | |
1875 | ||
1876 | bswap_src = find_bswap (stmt); | |
1877 | ||
1878 | if (!bswap_src) | |
1879 | continue; | |
1880 | ||
1881 | changed = true; | |
30c4e60d | 1882 | if (type_size == 32) |
1883 | bswap_stats.found_32bit++; | |
1884 | else | |
1885 | bswap_stats.found_64bit++; | |
0af25806 | 1886 | |
1887 | bswap_tmp = bswap_src; | |
1888 | ||
1889 | /* Convert the src expression if necessary. */ | |
1890 | if (!useless_type_conversion_p (TREE_TYPE (bswap_tmp), bswap_type)) | |
1891 | { | |
1892 | gimple convert_stmt; | |
03d37e4e | 1893 | bswap_tmp = make_temp_ssa_name (bswap_type, NULL, "bswapsrc"); |
1894 | convert_stmt = gimple_build_assign_with_ops | |
1895 | (NOP_EXPR, bswap_tmp, bswap_src, NULL); | |
0af25806 | 1896 | gsi_insert_before (&gsi, convert_stmt, GSI_SAME_STMT); |
1897 | } | |
1898 | ||
1899 | call = gimple_build_call (fndecl, 1, bswap_tmp); | |
1900 | ||
1901 | bswap_tmp = gimple_assign_lhs (stmt); | |
1902 | ||
1903 | /* Convert the result if necessary. */ | |
1904 | if (!useless_type_conversion_p (TREE_TYPE (bswap_tmp), bswap_type)) | |
1905 | { | |
1906 | gimple convert_stmt; | |
03d37e4e | 1907 | bswap_tmp = make_temp_ssa_name (bswap_type, NULL, "bswapdst"); |
1908 | convert_stmt = gimple_build_assign_with_ops | |
1909 | (NOP_EXPR, gimple_assign_lhs (stmt), bswap_tmp, NULL); | |
0af25806 | 1910 | gsi_insert_after (&gsi, convert_stmt, GSI_SAME_STMT); |
1911 | } | |
1912 | ||
1913 | gimple_call_set_lhs (call, bswap_tmp); | |
84cc784c | 1914 | |
1915 | if (dump_file) | |
1916 | { | |
1917 | fprintf (dump_file, "%d bit bswap implementation found at: ", | |
1918 | (int)type_size); | |
1919 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
1920 | } | |
1921 | ||
1922 | gsi_insert_after (&gsi, call, GSI_SAME_STMT); | |
1923 | gsi_remove (&gsi, true); | |
1924 | } | |
1925 | } | |
1926 | ||
30c4e60d | 1927 | statistics_counter_event (cfun, "32-bit bswap implementations found", |
1928 | bswap_stats.found_32bit); | |
1929 | statistics_counter_event (cfun, "64-bit bswap implementations found", | |
1930 | bswap_stats.found_64bit); | |
1931 | ||
771e2890 | 1932 | return (changed ? TODO_update_ssa | TODO_verify_ssa |
84cc784c | 1933 | | TODO_verify_stmts : 0); |
1934 | } | |
1935 | ||
1936 | static bool | |
1937 | gate_optimize_bswap (void) | |
1938 | { | |
1939 | return flag_expensive_optimizations && optimize; | |
1940 | } | |
1941 | ||
1942 | struct gimple_opt_pass pass_optimize_bswap = | |
1943 | { | |
1944 | { | |
1945 | GIMPLE_PASS, | |
1946 | "bswap", /* name */ | |
1947 | gate_optimize_bswap, /* gate */ | |
1948 | execute_optimize_bswap, /* execute */ | |
1949 | NULL, /* sub */ | |
1950 | NULL, /* next */ | |
1951 | 0, /* static_pass_number */ | |
1952 | TV_NONE, /* tv_id */ | |
1953 | PROP_ssa, /* properties_required */ | |
1954 | 0, /* properties_provided */ | |
1955 | 0, /* properties_destroyed */ | |
1956 | 0, /* todo_flags_start */ | |
1957 | 0 /* todo_flags_finish */ | |
1958 | } | |
1959 | }; | |
62be004c | 1960 | |
0989f516 | 1961 | /* Return true if RHS is a suitable operand for a widening multiplication, |
1962 | assuming a target type of TYPE. | |
7e4c867e | 1963 | There are two cases: |
1964 | ||
aff5fb4d | 1965 | - RHS makes some value at least twice as wide. Store that value |
1966 | in *NEW_RHS_OUT if so, and store its type in *TYPE_OUT. | |
7e4c867e | 1967 | |
1968 | - RHS is an integer constant. Store that value in *NEW_RHS_OUT if so, | |
1969 | but leave *TYPE_OUT untouched. */ | |
00f4f705 | 1970 | |
1971 | static bool | |
0989f516 | 1972 | is_widening_mult_rhs_p (tree type, tree rhs, tree *type_out, |
1973 | tree *new_rhs_out) | |
7e4c867e | 1974 | { |
1975 | gimple stmt; | |
0989f516 | 1976 | tree type1, rhs1; |
7e4c867e | 1977 | enum tree_code rhs_code; |
1978 | ||
1979 | if (TREE_CODE (rhs) == SSA_NAME) | |
1980 | { | |
7e4c867e | 1981 | stmt = SSA_NAME_DEF_STMT (rhs); |
0989f516 | 1982 | if (is_gimple_assign (stmt)) |
1983 | { | |
1984 | rhs_code = gimple_assign_rhs_code (stmt); | |
1985 | if (TREE_CODE (type) == INTEGER_TYPE | |
1986 | ? !CONVERT_EXPR_CODE_P (rhs_code) | |
1987 | : rhs_code != FIXED_CONVERT_EXPR) | |
1988 | rhs1 = rhs; | |
1989 | else | |
ffebd9c5 | 1990 | { |
1991 | rhs1 = gimple_assign_rhs1 (stmt); | |
1992 | ||
1993 | if (TREE_CODE (rhs1) == INTEGER_CST) | |
1994 | { | |
1995 | *new_rhs_out = rhs1; | |
1996 | *type_out = NULL; | |
1997 | return true; | |
1998 | } | |
1999 | } | |
0989f516 | 2000 | } |
2001 | else | |
2002 | rhs1 = rhs; | |
7e4c867e | 2003 | |
7e4c867e | 2004 | type1 = TREE_TYPE (rhs1); |
0989f516 | 2005 | |
7e4c867e | 2006 | if (TREE_CODE (type1) != TREE_CODE (type) |
aff5fb4d | 2007 | || TYPE_PRECISION (type1) * 2 > TYPE_PRECISION (type)) |
7e4c867e | 2008 | return false; |
2009 | ||
2010 | *new_rhs_out = rhs1; | |
2011 | *type_out = type1; | |
2012 | return true; | |
2013 | } | |
2014 | ||
2015 | if (TREE_CODE (rhs) == INTEGER_CST) | |
2016 | { | |
2017 | *new_rhs_out = rhs; | |
2018 | *type_out = NULL; | |
2019 | return true; | |
2020 | } | |
2021 | ||
2022 | return false; | |
2023 | } | |
2024 | ||
0989f516 | 2025 | /* Return true if STMT performs a widening multiplication, assuming the |
2026 | output type is TYPE. If so, store the unwidened types of the operands | |
2027 | in *TYPE1_OUT and *TYPE2_OUT respectively. Also fill *RHS1_OUT and | |
2028 | *RHS2_OUT such that converting those operands to types *TYPE1_OUT | |
2029 | and *TYPE2_OUT would give the operands of the multiplication. */ | |
7e4c867e | 2030 | |
2031 | static bool | |
4333b41f | 2032 | is_widening_mult_p (gimple stmt, |
7e4c867e | 2033 | tree *type1_out, tree *rhs1_out, |
2034 | tree *type2_out, tree *rhs2_out) | |
00f4f705 | 2035 | { |
4333b41f | 2036 | tree type = TREE_TYPE (gimple_assign_lhs (stmt)); |
2037 | ||
7e4c867e | 2038 | if (TREE_CODE (type) != INTEGER_TYPE |
2039 | && TREE_CODE (type) != FIXED_POINT_TYPE) | |
2040 | return false; | |
00f4f705 | 2041 | |
0989f516 | 2042 | if (!is_widening_mult_rhs_p (type, gimple_assign_rhs1 (stmt), type1_out, |
2043 | rhs1_out)) | |
00f4f705 | 2044 | return false; |
2045 | ||
0989f516 | 2046 | if (!is_widening_mult_rhs_p (type, gimple_assign_rhs2 (stmt), type2_out, |
2047 | rhs2_out)) | |
7e4c867e | 2048 | return false; |
00f4f705 | 2049 | |
7e4c867e | 2050 | if (*type1_out == NULL) |
00f4f705 | 2051 | { |
7e4c867e | 2052 | if (*type2_out == NULL || !int_fits_type_p (*rhs1_out, *type2_out)) |
00f4f705 | 2053 | return false; |
7e4c867e | 2054 | *type1_out = *type2_out; |
00f4f705 | 2055 | } |
00f4f705 | 2056 | |
7e4c867e | 2057 | if (*type2_out == NULL) |
00f4f705 | 2058 | { |
7e4c867e | 2059 | if (!int_fits_type_p (*rhs2_out, *type1_out)) |
00f4f705 | 2060 | return false; |
7e4c867e | 2061 | *type2_out = *type1_out; |
00f4f705 | 2062 | } |
00f4f705 | 2063 | |
287c271c | 2064 | /* Ensure that the larger of the two operands comes first. */ |
2065 | if (TYPE_PRECISION (*type1_out) < TYPE_PRECISION (*type2_out)) | |
2066 | { | |
2067 | tree tmp; | |
2068 | tmp = *type1_out; | |
2069 | *type1_out = *type2_out; | |
2070 | *type2_out = tmp; | |
2071 | tmp = *rhs1_out; | |
2072 | *rhs1_out = *rhs2_out; | |
2073 | *rhs2_out = tmp; | |
2074 | } | |
aff5fb4d | 2075 | |
7e4c867e | 2076 | return true; |
2077 | } | |
00f4f705 | 2078 | |
7e4c867e | 2079 | /* Process a single gimple statement STMT, which has a MULT_EXPR as |
2080 | its rhs, and try to convert it into a WIDEN_MULT_EXPR. The return | |
2081 | value is true iff we converted the statement. */ | |
2082 | ||
2083 | static bool | |
aff5fb4d | 2084 | convert_mult_to_widen (gimple stmt, gimple_stmt_iterator *gsi) |
7e4c867e | 2085 | { |
03d37e4e | 2086 | tree lhs, rhs1, rhs2, type, type1, type2; |
7e4c867e | 2087 | enum insn_code handler; |
aff5fb4d | 2088 | enum machine_mode to_mode, from_mode, actual_mode; |
5a574e8b | 2089 | optab op; |
aff5fb4d | 2090 | int actual_precision; |
2091 | location_t loc = gimple_location (stmt); | |
3f2ab719 | 2092 | bool from_unsigned1, from_unsigned2; |
7e4c867e | 2093 | |
2094 | lhs = gimple_assign_lhs (stmt); | |
2095 | type = TREE_TYPE (lhs); | |
2096 | if (TREE_CODE (type) != INTEGER_TYPE) | |
00f4f705 | 2097 | return false; |
2098 | ||
4333b41f | 2099 | if (!is_widening_mult_p (stmt, &type1, &rhs1, &type2, &rhs2)) |
00f4f705 | 2100 | return false; |
2101 | ||
5a574e8b | 2102 | to_mode = TYPE_MODE (type); |
2103 | from_mode = TYPE_MODE (type1); | |
3f2ab719 | 2104 | from_unsigned1 = TYPE_UNSIGNED (type1); |
2105 | from_unsigned2 = TYPE_UNSIGNED (type2); | |
5a574e8b | 2106 | |
3f2ab719 | 2107 | if (from_unsigned1 && from_unsigned2) |
5a574e8b | 2108 | op = umul_widen_optab; |
3f2ab719 | 2109 | else if (!from_unsigned1 && !from_unsigned2) |
5a574e8b | 2110 | op = smul_widen_optab; |
00f4f705 | 2111 | else |
5a574e8b | 2112 | op = usmul_widen_optab; |
2113 | ||
aff5fb4d | 2114 | handler = find_widening_optab_handler_and_mode (op, to_mode, from_mode, |
2115 | 0, &actual_mode); | |
7e4c867e | 2116 | |
2117 | if (handler == CODE_FOR_nothing) | |
3f2ab719 | 2118 | { |
2119 | if (op != smul_widen_optab) | |
2120 | { | |
22ffd684 | 2121 | /* We can use a signed multiply with unsigned types as long as |
2122 | there is a wider mode to use, or it is the smaller of the two | |
2123 | types that is unsigned. Note that type1 >= type2, always. */ | |
2124 | if ((TYPE_UNSIGNED (type1) | |
2125 | && TYPE_PRECISION (type1) == GET_MODE_PRECISION (from_mode)) | |
2126 | || (TYPE_UNSIGNED (type2) | |
2127 | && TYPE_PRECISION (type2) == GET_MODE_PRECISION (from_mode))) | |
2128 | { | |
2129 | from_mode = GET_MODE_WIDER_MODE (from_mode); | |
2130 | if (GET_MODE_SIZE (to_mode) <= GET_MODE_SIZE (from_mode)) | |
2131 | return false; | |
2132 | } | |
3f2ab719 | 2133 | |
2134 | op = smul_widen_optab; | |
2135 | handler = find_widening_optab_handler_and_mode (op, to_mode, | |
2136 | from_mode, 0, | |
2137 | &actual_mode); | |
2138 | ||
2139 | if (handler == CODE_FOR_nothing) | |
2140 | return false; | |
2141 | ||
2142 | from_unsigned1 = from_unsigned2 = false; | |
2143 | } | |
2144 | else | |
2145 | return false; | |
2146 | } | |
7e4c867e | 2147 | |
aff5fb4d | 2148 | /* Ensure that the inputs to the handler are in the correct precison |
2149 | for the opcode. This will be the full mode size. */ | |
2150 | actual_precision = GET_MODE_PRECISION (actual_mode); | |
b36be69d | 2151 | if (2 * actual_precision > TYPE_PRECISION (type)) |
2152 | return false; | |
3f2ab719 | 2153 | if (actual_precision != TYPE_PRECISION (type1) |
2154 | || from_unsigned1 != TYPE_UNSIGNED (type1)) | |
03d37e4e | 2155 | rhs1 = build_and_insert_cast (gsi, loc, |
2156 | build_nonstandard_integer_type | |
2157 | (actual_precision, from_unsigned1), rhs1); | |
3f2ab719 | 2158 | if (actual_precision != TYPE_PRECISION (type2) |
2159 | || from_unsigned2 != TYPE_UNSIGNED (type2)) | |
03d37e4e | 2160 | rhs2 = build_and_insert_cast (gsi, loc, |
2161 | build_nonstandard_integer_type | |
2162 | (actual_precision, from_unsigned2), rhs2); | |
aff5fb4d | 2163 | |
ffebd9c5 | 2164 | /* Handle constants. */ |
2165 | if (TREE_CODE (rhs1) == INTEGER_CST) | |
2166 | rhs1 = fold_convert (type1, rhs1); | |
2167 | if (TREE_CODE (rhs2) == INTEGER_CST) | |
2168 | rhs2 = fold_convert (type2, rhs2); | |
2169 | ||
aff5fb4d | 2170 | gimple_assign_set_rhs1 (stmt, rhs1); |
2171 | gimple_assign_set_rhs2 (stmt, rhs2); | |
00f4f705 | 2172 | gimple_assign_set_rhs_code (stmt, WIDEN_MULT_EXPR); |
2173 | update_stmt (stmt); | |
30c4e60d | 2174 | widen_mul_stats.widen_mults_inserted++; |
00f4f705 | 2175 | return true; |
2176 | } | |
2177 | ||
2178 | /* Process a single gimple statement STMT, which is found at the | |
2179 | iterator GSI and has a either a PLUS_EXPR or a MINUS_EXPR as its | |
2180 | rhs (given by CODE), and try to convert it into a | |
2181 | WIDEN_MULT_PLUS_EXPR or a WIDEN_MULT_MINUS_EXPR. The return value | |
2182 | is true iff we converted the statement. */ | |
2183 | ||
2184 | static bool | |
2185 | convert_plusminus_to_widen (gimple_stmt_iterator *gsi, gimple stmt, | |
2186 | enum tree_code code) | |
2187 | { | |
2188 | gimple rhs1_stmt = NULL, rhs2_stmt = NULL; | |
07ea3e5c | 2189 | gimple conv1_stmt = NULL, conv2_stmt = NULL, conv_stmt; |
03d37e4e | 2190 | tree type, type1, type2, optype; |
00f4f705 | 2191 | tree lhs, rhs1, rhs2, mult_rhs1, mult_rhs2, add_rhs; |
2192 | enum tree_code rhs1_code = ERROR_MARK, rhs2_code = ERROR_MARK; | |
2193 | optab this_optab; | |
2194 | enum tree_code wmult_code; | |
aff5fb4d | 2195 | enum insn_code handler; |
2196 | enum machine_mode to_mode, from_mode, actual_mode; | |
2197 | location_t loc = gimple_location (stmt); | |
2198 | int actual_precision; | |
3f2ab719 | 2199 | bool from_unsigned1, from_unsigned2; |
00f4f705 | 2200 | |
2201 | lhs = gimple_assign_lhs (stmt); | |
2202 | type = TREE_TYPE (lhs); | |
7e4c867e | 2203 | if (TREE_CODE (type) != INTEGER_TYPE |
2204 | && TREE_CODE (type) != FIXED_POINT_TYPE) | |
00f4f705 | 2205 | return false; |
2206 | ||
2207 | if (code == MINUS_EXPR) | |
2208 | wmult_code = WIDEN_MULT_MINUS_EXPR; | |
2209 | else | |
2210 | wmult_code = WIDEN_MULT_PLUS_EXPR; | |
2211 | ||
00f4f705 | 2212 | rhs1 = gimple_assign_rhs1 (stmt); |
2213 | rhs2 = gimple_assign_rhs2 (stmt); | |
2214 | ||
2215 | if (TREE_CODE (rhs1) == SSA_NAME) | |
2216 | { | |
2217 | rhs1_stmt = SSA_NAME_DEF_STMT (rhs1); | |
2218 | if (is_gimple_assign (rhs1_stmt)) | |
2219 | rhs1_code = gimple_assign_rhs_code (rhs1_stmt); | |
2220 | } | |
00f4f705 | 2221 | |
2222 | if (TREE_CODE (rhs2) == SSA_NAME) | |
2223 | { | |
2224 | rhs2_stmt = SSA_NAME_DEF_STMT (rhs2); | |
2225 | if (is_gimple_assign (rhs2_stmt)) | |
2226 | rhs2_code = gimple_assign_rhs_code (rhs2_stmt); | |
2227 | } | |
00f4f705 | 2228 | |
07ea3e5c | 2229 | /* Allow for one conversion statement between the multiply |
2230 | and addition/subtraction statement. If there are more than | |
2231 | one conversions then we assume they would invalidate this | |
2232 | transformation. If that's not the case then they should have | |
2233 | been folded before now. */ | |
2234 | if (CONVERT_EXPR_CODE_P (rhs1_code)) | |
2235 | { | |
2236 | conv1_stmt = rhs1_stmt; | |
2237 | rhs1 = gimple_assign_rhs1 (rhs1_stmt); | |
2238 | if (TREE_CODE (rhs1) == SSA_NAME) | |
2239 | { | |
2240 | rhs1_stmt = SSA_NAME_DEF_STMT (rhs1); | |
2241 | if (is_gimple_assign (rhs1_stmt)) | |
2242 | rhs1_code = gimple_assign_rhs_code (rhs1_stmt); | |
2243 | } | |
2244 | else | |
2245 | return false; | |
2246 | } | |
2247 | if (CONVERT_EXPR_CODE_P (rhs2_code)) | |
2248 | { | |
2249 | conv2_stmt = rhs2_stmt; | |
2250 | rhs2 = gimple_assign_rhs1 (rhs2_stmt); | |
2251 | if (TREE_CODE (rhs2) == SSA_NAME) | |
2252 | { | |
2253 | rhs2_stmt = SSA_NAME_DEF_STMT (rhs2); | |
2254 | if (is_gimple_assign (rhs2_stmt)) | |
2255 | rhs2_code = gimple_assign_rhs_code (rhs2_stmt); | |
2256 | } | |
2257 | else | |
2258 | return false; | |
2259 | } | |
2260 | ||
aff5fb4d | 2261 | /* If code is WIDEN_MULT_EXPR then it would seem unnecessary to call |
2262 | is_widening_mult_p, but we still need the rhs returns. | |
2263 | ||
2264 | It might also appear that it would be sufficient to use the existing | |
2265 | operands of the widening multiply, but that would limit the choice of | |
2266 | multiply-and-accumulate instructions. */ | |
2267 | if (code == PLUS_EXPR | |
2268 | && (rhs1_code == MULT_EXPR || rhs1_code == WIDEN_MULT_EXPR)) | |
00f4f705 | 2269 | { |
4333b41f | 2270 | if (!is_widening_mult_p (rhs1_stmt, &type1, &mult_rhs1, |
7e4c867e | 2271 | &type2, &mult_rhs2)) |
00f4f705 | 2272 | return false; |
7e4c867e | 2273 | add_rhs = rhs2; |
07ea3e5c | 2274 | conv_stmt = conv1_stmt; |
00f4f705 | 2275 | } |
aff5fb4d | 2276 | else if (rhs2_code == MULT_EXPR || rhs2_code == WIDEN_MULT_EXPR) |
00f4f705 | 2277 | { |
4333b41f | 2278 | if (!is_widening_mult_p (rhs2_stmt, &type1, &mult_rhs1, |
7e4c867e | 2279 | &type2, &mult_rhs2)) |
00f4f705 | 2280 | return false; |
7e4c867e | 2281 | add_rhs = rhs1; |
07ea3e5c | 2282 | conv_stmt = conv2_stmt; |
00f4f705 | 2283 | } |
00f4f705 | 2284 | else |
2285 | return false; | |
2286 | ||
aff5fb4d | 2287 | to_mode = TYPE_MODE (type); |
2288 | from_mode = TYPE_MODE (type1); | |
3f2ab719 | 2289 | from_unsigned1 = TYPE_UNSIGNED (type1); |
2290 | from_unsigned2 = TYPE_UNSIGNED (type2); | |
4ccf368d | 2291 | optype = type1; |
aff5fb4d | 2292 | |
3f2ab719 | 2293 | /* There's no such thing as a mixed sign madd yet, so use a wider mode. */ |
2294 | if (from_unsigned1 != from_unsigned2) | |
2295 | { | |
4ccf368d | 2296 | if (!INTEGRAL_TYPE_P (type)) |
2297 | return false; | |
22ffd684 | 2298 | /* We can use a signed multiply with unsigned types as long as |
2299 | there is a wider mode to use, or it is the smaller of the two | |
2300 | types that is unsigned. Note that type1 >= type2, always. */ | |
2301 | if ((from_unsigned1 | |
2302 | && TYPE_PRECISION (type1) == GET_MODE_PRECISION (from_mode)) | |
2303 | || (from_unsigned2 | |
2304 | && TYPE_PRECISION (type2) == GET_MODE_PRECISION (from_mode))) | |
3f2ab719 | 2305 | { |
22ffd684 | 2306 | from_mode = GET_MODE_WIDER_MODE (from_mode); |
2307 | if (GET_MODE_SIZE (from_mode) >= GET_MODE_SIZE (to_mode)) | |
2308 | return false; | |
3f2ab719 | 2309 | } |
22ffd684 | 2310 | |
2311 | from_unsigned1 = from_unsigned2 = false; | |
4ccf368d | 2312 | optype = build_nonstandard_integer_type (GET_MODE_PRECISION (from_mode), |
2313 | false); | |
3f2ab719 | 2314 | } |
815a0224 | 2315 | |
07ea3e5c | 2316 | /* If there was a conversion between the multiply and addition |
2317 | then we need to make sure it fits a multiply-and-accumulate. | |
2318 | The should be a single mode change which does not change the | |
2319 | value. */ | |
2320 | if (conv_stmt) | |
2321 | { | |
3f2ab719 | 2322 | /* We use the original, unmodified data types for this. */ |
07ea3e5c | 2323 | tree from_type = TREE_TYPE (gimple_assign_rhs1 (conv_stmt)); |
2324 | tree to_type = TREE_TYPE (gimple_assign_lhs (conv_stmt)); | |
2325 | int data_size = TYPE_PRECISION (type1) + TYPE_PRECISION (type2); | |
2326 | bool is_unsigned = TYPE_UNSIGNED (type1) && TYPE_UNSIGNED (type2); | |
2327 | ||
2328 | if (TYPE_PRECISION (from_type) > TYPE_PRECISION (to_type)) | |
2329 | { | |
2330 | /* Conversion is a truncate. */ | |
2331 | if (TYPE_PRECISION (to_type) < data_size) | |
2332 | return false; | |
2333 | } | |
2334 | else if (TYPE_PRECISION (from_type) < TYPE_PRECISION (to_type)) | |
2335 | { | |
2336 | /* Conversion is an extend. Check it's the right sort. */ | |
2337 | if (TYPE_UNSIGNED (from_type) != is_unsigned | |
2338 | && !(is_unsigned && TYPE_PRECISION (from_type) > data_size)) | |
2339 | return false; | |
2340 | } | |
2341 | /* else convert is a no-op for our purposes. */ | |
2342 | } | |
2343 | ||
815a0224 | 2344 | /* Verify that the machine can perform a widening multiply |
2345 | accumulate in this mode/signedness combination, otherwise | |
2346 | this transformation is likely to pessimize code. */ | |
3f2ab719 | 2347 | this_optab = optab_for_tree_code (wmult_code, optype, optab_default); |
aff5fb4d | 2348 | handler = find_widening_optab_handler_and_mode (this_optab, to_mode, |
2349 | from_mode, 0, &actual_mode); | |
2350 | ||
2351 | if (handler == CODE_FOR_nothing) | |
815a0224 | 2352 | return false; |
2353 | ||
aff5fb4d | 2354 | /* Ensure that the inputs to the handler are in the correct precison |
2355 | for the opcode. This will be the full mode size. */ | |
2356 | actual_precision = GET_MODE_PRECISION (actual_mode); | |
3f2ab719 | 2357 | if (actual_precision != TYPE_PRECISION (type1) |
2358 | || from_unsigned1 != TYPE_UNSIGNED (type1)) | |
03d37e4e | 2359 | mult_rhs1 = build_and_insert_cast (gsi, loc, |
2360 | build_nonstandard_integer_type | |
2361 | (actual_precision, from_unsigned1), | |
2362 | mult_rhs1); | |
3f2ab719 | 2363 | if (actual_precision != TYPE_PRECISION (type2) |
2364 | || from_unsigned2 != TYPE_UNSIGNED (type2)) | |
03d37e4e | 2365 | mult_rhs2 = build_and_insert_cast (gsi, loc, |
2366 | build_nonstandard_integer_type | |
2367 | (actual_precision, from_unsigned2), | |
2368 | mult_rhs2); | |
00f4f705 | 2369 | |
12421545 | 2370 | if (!useless_type_conversion_p (type, TREE_TYPE (add_rhs))) |
03d37e4e | 2371 | add_rhs = build_and_insert_cast (gsi, loc, type, add_rhs); |
12421545 | 2372 | |
ffebd9c5 | 2373 | /* Handle constants. */ |
2374 | if (TREE_CODE (mult_rhs1) == INTEGER_CST) | |
d5a3bb10 | 2375 | mult_rhs1 = fold_convert (type1, mult_rhs1); |
ffebd9c5 | 2376 | if (TREE_CODE (mult_rhs2) == INTEGER_CST) |
d5a3bb10 | 2377 | mult_rhs2 = fold_convert (type2, mult_rhs2); |
ffebd9c5 | 2378 | |
aff5fb4d | 2379 | gimple_assign_set_rhs_with_ops_1 (gsi, wmult_code, mult_rhs1, mult_rhs2, |
00f4f705 | 2380 | add_rhs); |
2381 | update_stmt (gsi_stmt (*gsi)); | |
30c4e60d | 2382 | widen_mul_stats.maccs_inserted++; |
00f4f705 | 2383 | return true; |
2384 | } | |
2385 | ||
15dbdc8f | 2386 | /* Combine the multiplication at MUL_STMT with operands MULOP1 and MULOP2 |
2387 | with uses in additions and subtractions to form fused multiply-add | |
2388 | operations. Returns true if successful and MUL_STMT should be removed. */ | |
b9be572e | 2389 | |
2390 | static bool | |
15dbdc8f | 2391 | convert_mult_to_fma (gimple mul_stmt, tree op1, tree op2) |
b9be572e | 2392 | { |
15dbdc8f | 2393 | tree mul_result = gimple_get_lhs (mul_stmt); |
b9be572e | 2394 | tree type = TREE_TYPE (mul_result); |
44579526 | 2395 | gimple use_stmt, neguse_stmt, fma_stmt; |
b9be572e | 2396 | use_operand_p use_p; |
2397 | imm_use_iterator imm_iter; | |
2398 | ||
2399 | if (FLOAT_TYPE_P (type) | |
2400 | && flag_fp_contract_mode == FP_CONTRACT_OFF) | |
2401 | return false; | |
2402 | ||
2403 | /* We don't want to do bitfield reduction ops. */ | |
2404 | if (INTEGRAL_TYPE_P (type) | |
2405 | && (TYPE_PRECISION (type) | |
2406 | != GET_MODE_PRECISION (TYPE_MODE (type)))) | |
2407 | return false; | |
2408 | ||
2409 | /* If the target doesn't support it, don't generate it. We assume that | |
2410 | if fma isn't available then fms, fnma or fnms are not either. */ | |
2411 | if (optab_handler (fma_optab, TYPE_MODE (type)) == CODE_FOR_nothing) | |
2412 | return false; | |
2413 | ||
5ed3d3b8 | 2414 | /* If the multiplication has zero uses, it is kept around probably because |
2415 | of -fnon-call-exceptions. Don't optimize it away in that case, | |
2416 | it is DCE job. */ | |
2417 | if (has_zero_uses (mul_result)) | |
2418 | return false; | |
2419 | ||
b9be572e | 2420 | /* Make sure that the multiplication statement becomes dead after |
2421 | the transformation, thus that all uses are transformed to FMAs. | |
2422 | This means we assume that an FMA operation has the same cost | |
2423 | as an addition. */ | |
2424 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, mul_result) | |
2425 | { | |
2426 | enum tree_code use_code; | |
44579526 | 2427 | tree result = mul_result; |
2428 | bool negate_p = false; | |
b9be572e | 2429 | |
2430 | use_stmt = USE_STMT (use_p); | |
2431 | ||
17a2c727 | 2432 | if (is_gimple_debug (use_stmt)) |
2433 | continue; | |
2434 | ||
b9be572e | 2435 | /* For now restrict this operations to single basic blocks. In theory |
2436 | we would want to support sinking the multiplication in | |
2437 | m = a*b; | |
2438 | if () | |
2439 | ma = m + c; | |
2440 | else | |
2441 | d = m; | |
2442 | to form a fma in the then block and sink the multiplication to the | |
2443 | else block. */ | |
2444 | if (gimple_bb (use_stmt) != gimple_bb (mul_stmt)) | |
2445 | return false; | |
2446 | ||
44579526 | 2447 | if (!is_gimple_assign (use_stmt)) |
b9be572e | 2448 | return false; |
2449 | ||
44579526 | 2450 | use_code = gimple_assign_rhs_code (use_stmt); |
2451 | ||
2452 | /* A negate on the multiplication leads to FNMA. */ | |
2453 | if (use_code == NEGATE_EXPR) | |
2454 | { | |
805ad414 | 2455 | ssa_op_iter iter; |
5715c09b | 2456 | use_operand_p usep; |
805ad414 | 2457 | |
44579526 | 2458 | result = gimple_assign_lhs (use_stmt); |
2459 | ||
2460 | /* Make sure the negate statement becomes dead with this | |
2461 | single transformation. */ | |
2462 | if (!single_imm_use (gimple_assign_lhs (use_stmt), | |
2463 | &use_p, &neguse_stmt)) | |
2464 | return false; | |
2465 | ||
805ad414 | 2466 | /* Make sure the multiplication isn't also used on that stmt. */ |
5715c09b | 2467 | FOR_EACH_PHI_OR_STMT_USE (usep, neguse_stmt, iter, SSA_OP_USE) |
2468 | if (USE_FROM_PTR (usep) == mul_result) | |
805ad414 | 2469 | return false; |
2470 | ||
44579526 | 2471 | /* Re-validate. */ |
2472 | use_stmt = neguse_stmt; | |
2473 | if (gimple_bb (use_stmt) != gimple_bb (mul_stmt)) | |
2474 | return false; | |
2475 | if (!is_gimple_assign (use_stmt)) | |
2476 | return false; | |
2477 | ||
2478 | use_code = gimple_assign_rhs_code (use_stmt); | |
2479 | negate_p = true; | |
2480 | } | |
b9be572e | 2481 | |
44579526 | 2482 | switch (use_code) |
2483 | { | |
2484 | case MINUS_EXPR: | |
8a9d0572 | 2485 | if (gimple_assign_rhs2 (use_stmt) == result) |
2486 | negate_p = !negate_p; | |
2487 | break; | |
44579526 | 2488 | case PLUS_EXPR: |
44579526 | 2489 | break; |
44579526 | 2490 | default: |
2491 | /* FMA can only be formed from PLUS and MINUS. */ | |
2492 | return false; | |
2493 | } | |
b9be572e | 2494 | |
44579526 | 2495 | /* We can't handle a * b + a * b. */ |
2496 | if (gimple_assign_rhs1 (use_stmt) == gimple_assign_rhs2 (use_stmt)) | |
2497 | return false; | |
8a9d0572 | 2498 | |
2499 | /* While it is possible to validate whether or not the exact form | |
2500 | that we've recognized is available in the backend, the assumption | |
2501 | is that the transformation is never a loss. For instance, suppose | |
2502 | the target only has the plain FMA pattern available. Consider | |
2503 | a*b-c -> fma(a,b,-c): we've exchanged MUL+SUB for FMA+NEG, which | |
2504 | is still two operations. Consider -(a*b)-c -> fma(-a,b,-c): we | |
2505 | still have 3 operations, but in the FMA form the two NEGs are | |
9d75589a | 2506 | independent and could be run in parallel. */ |
b9be572e | 2507 | } |
2508 | ||
2509 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, mul_result) | |
2510 | { | |
b9be572e | 2511 | gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt); |
17a2c727 | 2512 | enum tree_code use_code; |
15dbdc8f | 2513 | tree addop, mulop1 = op1, result = mul_result; |
44579526 | 2514 | bool negate_p = false; |
b9be572e | 2515 | |
17a2c727 | 2516 | if (is_gimple_debug (use_stmt)) |
2517 | continue; | |
2518 | ||
2519 | use_code = gimple_assign_rhs_code (use_stmt); | |
44579526 | 2520 | if (use_code == NEGATE_EXPR) |
2521 | { | |
2522 | result = gimple_assign_lhs (use_stmt); | |
2523 | single_imm_use (gimple_assign_lhs (use_stmt), &use_p, &neguse_stmt); | |
2524 | gsi_remove (&gsi, true); | |
2525 | release_defs (use_stmt); | |
2526 | ||
2527 | use_stmt = neguse_stmt; | |
2528 | gsi = gsi_for_stmt (use_stmt); | |
2529 | use_code = gimple_assign_rhs_code (use_stmt); | |
2530 | negate_p = true; | |
2531 | } | |
2532 | ||
2533 | if (gimple_assign_rhs1 (use_stmt) == result) | |
b9be572e | 2534 | { |
2535 | addop = gimple_assign_rhs2 (use_stmt); | |
2536 | /* a * b - c -> a * b + (-c) */ | |
2537 | if (gimple_assign_rhs_code (use_stmt) == MINUS_EXPR) | |
2538 | addop = force_gimple_operand_gsi (&gsi, | |
2539 | build1 (NEGATE_EXPR, | |
2540 | type, addop), | |
2541 | true, NULL_TREE, true, | |
2542 | GSI_SAME_STMT); | |
2543 | } | |
2544 | else | |
2545 | { | |
2546 | addop = gimple_assign_rhs1 (use_stmt); | |
2547 | /* a - b * c -> (-b) * c + a */ | |
2548 | if (gimple_assign_rhs_code (use_stmt) == MINUS_EXPR) | |
44579526 | 2549 | negate_p = !negate_p; |
b9be572e | 2550 | } |
2551 | ||
44579526 | 2552 | if (negate_p) |
2553 | mulop1 = force_gimple_operand_gsi (&gsi, | |
2554 | build1 (NEGATE_EXPR, | |
2555 | type, mulop1), | |
2556 | true, NULL_TREE, true, | |
2557 | GSI_SAME_STMT); | |
2558 | ||
b9be572e | 2559 | fma_stmt = gimple_build_assign_with_ops3 (FMA_EXPR, |
2560 | gimple_assign_lhs (use_stmt), | |
15dbdc8f | 2561 | mulop1, op2, |
b9be572e | 2562 | addop); |
2563 | gsi_replace (&gsi, fma_stmt, true); | |
30c4e60d | 2564 | widen_mul_stats.fmas_inserted++; |
b9be572e | 2565 | } |
2566 | ||
2567 | return true; | |
2568 | } | |
2569 | ||
62be004c | 2570 | /* Find integer multiplications where the operands are extended from |
2571 | smaller types, and replace the MULT_EXPR with a WIDEN_MULT_EXPR | |
2572 | where appropriate. */ | |
2573 | ||
2574 | static unsigned int | |
2575 | execute_optimize_widening_mul (void) | |
2576 | { | |
62be004c | 2577 | basic_block bb; |
15dbdc8f | 2578 | bool cfg_changed = false; |
62be004c | 2579 | |
30c4e60d | 2580 | memset (&widen_mul_stats, 0, sizeof (widen_mul_stats)); |
2581 | ||
62be004c | 2582 | FOR_EACH_BB (bb) |
2583 | { | |
2584 | gimple_stmt_iterator gsi; | |
2585 | ||
b9be572e | 2586 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi);) |
62be004c | 2587 | { |
2588 | gimple stmt = gsi_stmt (gsi); | |
00f4f705 | 2589 | enum tree_code code; |
62be004c | 2590 | |
b9be572e | 2591 | if (is_gimple_assign (stmt)) |
2592 | { | |
2593 | code = gimple_assign_rhs_code (stmt); | |
2594 | switch (code) | |
2595 | { | |
2596 | case MULT_EXPR: | |
aff5fb4d | 2597 | if (!convert_mult_to_widen (stmt, &gsi) |
15dbdc8f | 2598 | && convert_mult_to_fma (stmt, |
2599 | gimple_assign_rhs1 (stmt), | |
2600 | gimple_assign_rhs2 (stmt))) | |
b9be572e | 2601 | { |
2602 | gsi_remove (&gsi, true); | |
2603 | release_defs (stmt); | |
2604 | continue; | |
2605 | } | |
2606 | break; | |
2607 | ||
2608 | case PLUS_EXPR: | |
2609 | case MINUS_EXPR: | |
2610 | convert_plusminus_to_widen (&gsi, stmt, code); | |
2611 | break; | |
62be004c | 2612 | |
b9be572e | 2613 | default:; |
2614 | } | |
2615 | } | |
d4af184a | 2616 | else if (is_gimple_call (stmt) |
2617 | && gimple_call_lhs (stmt)) | |
15dbdc8f | 2618 | { |
2619 | tree fndecl = gimple_call_fndecl (stmt); | |
2620 | if (fndecl | |
2621 | && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) | |
2622 | { | |
2623 | switch (DECL_FUNCTION_CODE (fndecl)) | |
2624 | { | |
2625 | case BUILT_IN_POWF: | |
2626 | case BUILT_IN_POW: | |
2627 | case BUILT_IN_POWL: | |
2628 | if (TREE_CODE (gimple_call_arg (stmt, 1)) == REAL_CST | |
2629 | && REAL_VALUES_EQUAL | |
2630 | (TREE_REAL_CST (gimple_call_arg (stmt, 1)), | |
2631 | dconst2) | |
2632 | && convert_mult_to_fma (stmt, | |
2633 | gimple_call_arg (stmt, 0), | |
2634 | gimple_call_arg (stmt, 0))) | |
2635 | { | |
6716f635 | 2636 | unlink_stmt_vdef (stmt); |
13ff78a4 | 2637 | if (gsi_remove (&gsi, true) |
2638 | && gimple_purge_dead_eh_edges (bb)) | |
15dbdc8f | 2639 | cfg_changed = true; |
13ff78a4 | 2640 | release_defs (stmt); |
15dbdc8f | 2641 | continue; |
2642 | } | |
2643 | break; | |
2644 | ||
2645 | default:; | |
2646 | } | |
2647 | } | |
2648 | } | |
b9be572e | 2649 | gsi_next (&gsi); |
62be004c | 2650 | } |
2651 | } | |
00f4f705 | 2652 | |
30c4e60d | 2653 | statistics_counter_event (cfun, "widening multiplications inserted", |
2654 | widen_mul_stats.widen_mults_inserted); | |
2655 | statistics_counter_event (cfun, "widening maccs inserted", | |
2656 | widen_mul_stats.maccs_inserted); | |
2657 | statistics_counter_event (cfun, "fused multiply-adds inserted", | |
2658 | widen_mul_stats.fmas_inserted); | |
2659 | ||
15dbdc8f | 2660 | return cfg_changed ? TODO_cleanup_cfg : 0; |
62be004c | 2661 | } |
2662 | ||
2663 | static bool | |
2664 | gate_optimize_widening_mul (void) | |
2665 | { | |
2666 | return flag_expensive_optimizations && optimize; | |
2667 | } | |
2668 | ||
2669 | struct gimple_opt_pass pass_optimize_widening_mul = | |
2670 | { | |
2671 | { | |
2672 | GIMPLE_PASS, | |
2673 | "widening_mul", /* name */ | |
2674 | gate_optimize_widening_mul, /* gate */ | |
2675 | execute_optimize_widening_mul, /* execute */ | |
2676 | NULL, /* sub */ | |
2677 | NULL, /* next */ | |
2678 | 0, /* static_pass_number */ | |
2679 | TV_NONE, /* tv_id */ | |
2680 | PROP_ssa, /* properties_required */ | |
2681 | 0, /* properties_provided */ | |
2682 | 0, /* properties_destroyed */ | |
2683 | 0, /* todo_flags_start */ | |
b9be572e | 2684 | TODO_verify_ssa |
2685 | | TODO_verify_stmts | |
b9be572e | 2686 | | TODO_update_ssa /* todo_flags_finish */ |
62be004c | 2687 | } |
2688 | }; |