]>
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 "timevar.h" |
96 | #include "tree-pass.h" | |
ac70caad | 97 | #include "alloc-pool.h" |
98 | #include "basic-block.h" | |
99 | #include "target.h" | |
ce084dfc | 100 | #include "gimple-pretty-print.h" |
a7a46268 | 101 | |
102 | /* FIXME: RTL headers have to be included here for optabs. */ | |
103 | #include "rtl.h" /* Because optabs.h wants enum rtx_code. */ | |
104 | #include "expr.h" /* Because optabs.h wants sepops. */ | |
84cc784c | 105 | #include "optabs.h" |
ac70caad | 106 | |
107 | /* This structure represents one basic block that either computes a | |
108 | division, or is a common dominator for basic block that compute a | |
109 | division. */ | |
110 | struct occurrence { | |
111 | /* The basic block represented by this structure. */ | |
112 | basic_block bb; | |
113 | ||
114 | /* If non-NULL, the SSA_NAME holding the definition for a reciprocal | |
115 | inserted in BB. */ | |
116 | tree recip_def; | |
117 | ||
75a70cf9 | 118 | /* If non-NULL, the GIMPLE_ASSIGN for a reciprocal computation that |
ac70caad | 119 | was inserted in BB. */ |
75a70cf9 | 120 | gimple recip_def_stmt; |
ac70caad | 121 | |
122 | /* Pointer to a list of "struct occurrence"s for blocks dominated | |
123 | by BB. */ | |
124 | struct occurrence *children; | |
125 | ||
126 | /* Pointer to the next "struct occurrence"s in the list of blocks | |
127 | sharing a common dominator. */ | |
128 | struct occurrence *next; | |
129 | ||
130 | /* The number of divisions that are in BB before compute_merit. The | |
131 | number of divisions that are in BB or post-dominate it after | |
132 | compute_merit. */ | |
133 | int num_divisions; | |
134 | ||
135 | /* True if the basic block has a division, false if it is a common | |
136 | dominator for basic blocks that do. If it is false and trapping | |
137 | math is active, BB is not a candidate for inserting a reciprocal. */ | |
138 | bool bb_has_division; | |
139 | }; | |
140 | ||
30c4e60d | 141 | static struct |
142 | { | |
143 | /* Number of 1.0/X ops inserted. */ | |
144 | int rdivs_inserted; | |
145 | ||
146 | /* Number of 1.0/FUNC ops inserted. */ | |
147 | int rfuncs_inserted; | |
148 | } reciprocal_stats; | |
149 | ||
150 | static struct | |
151 | { | |
152 | /* Number of cexpi calls inserted. */ | |
153 | int inserted; | |
154 | } sincos_stats; | |
155 | ||
156 | static struct | |
157 | { | |
158 | /* Number of hand-written 32-bit bswaps found. */ | |
159 | int found_32bit; | |
160 | ||
161 | /* Number of hand-written 64-bit bswaps found. */ | |
162 | int found_64bit; | |
163 | } bswap_stats; | |
164 | ||
165 | static struct | |
166 | { | |
167 | /* Number of widening multiplication ops inserted. */ | |
168 | int widen_mults_inserted; | |
169 | ||
170 | /* Number of integer multiply-and-accumulate ops inserted. */ | |
171 | int maccs_inserted; | |
172 | ||
173 | /* Number of fp fused multiply-add ops inserted. */ | |
174 | int fmas_inserted; | |
175 | } widen_mul_stats; | |
ac70caad | 176 | |
177 | /* The instance of "struct occurrence" representing the highest | |
178 | interesting block in the dominator tree. */ | |
179 | static struct occurrence *occ_head; | |
180 | ||
181 | /* Allocation pool for getting instances of "struct occurrence". */ | |
182 | static alloc_pool occ_pool; | |
183 | ||
184 | ||
185 | ||
186 | /* Allocate and return a new struct occurrence for basic block BB, and | |
187 | whose children list is headed by CHILDREN. */ | |
188 | static struct occurrence * | |
189 | occ_new (basic_block bb, struct occurrence *children) | |
abacb398 | 190 | { |
ac70caad | 191 | struct occurrence *occ; |
192 | ||
f0d6e81c | 193 | bb->aux = occ = (struct occurrence *) pool_alloc (occ_pool); |
ac70caad | 194 | memset (occ, 0, sizeof (struct occurrence)); |
195 | ||
196 | occ->bb = bb; | |
197 | occ->children = children; | |
198 | return occ; | |
abacb398 | 199 | } |
200 | ||
ac70caad | 201 | |
202 | /* Insert NEW_OCC into our subset of the dominator tree. P_HEAD points to a | |
203 | list of "struct occurrence"s, one per basic block, having IDOM as | |
204 | their common dominator. | |
205 | ||
206 | We try to insert NEW_OCC as deep as possible in the tree, and we also | |
207 | insert any other block that is a common dominator for BB and one | |
208 | block already in the tree. */ | |
209 | ||
210 | static void | |
211 | insert_bb (struct occurrence *new_occ, basic_block idom, | |
212 | struct occurrence **p_head) | |
9e583fac | 213 | { |
ac70caad | 214 | struct occurrence *occ, **p_occ; |
9e583fac | 215 | |
ac70caad | 216 | for (p_occ = p_head; (occ = *p_occ) != NULL; ) |
217 | { | |
218 | basic_block bb = new_occ->bb, occ_bb = occ->bb; | |
219 | basic_block dom = nearest_common_dominator (CDI_DOMINATORS, occ_bb, bb); | |
220 | if (dom == bb) | |
221 | { | |
222 | /* BB dominates OCC_BB. OCC becomes NEW_OCC's child: remove OCC | |
223 | from its list. */ | |
224 | *p_occ = occ->next; | |
225 | occ->next = new_occ->children; | |
226 | new_occ->children = occ; | |
227 | ||
228 | /* Try the next block (it may as well be dominated by BB). */ | |
229 | } | |
230 | ||
231 | else if (dom == occ_bb) | |
232 | { | |
233 | /* OCC_BB dominates BB. Tail recurse to look deeper. */ | |
234 | insert_bb (new_occ, dom, &occ->children); | |
235 | return; | |
236 | } | |
237 | ||
238 | else if (dom != idom) | |
239 | { | |
240 | gcc_assert (!dom->aux); | |
241 | ||
242 | /* There is a dominator between IDOM and BB, add it and make | |
243 | two children out of NEW_OCC and OCC. First, remove OCC from | |
244 | its list. */ | |
245 | *p_occ = occ->next; | |
246 | new_occ->next = occ; | |
247 | occ->next = NULL; | |
248 | ||
249 | /* None of the previous blocks has DOM as a dominator: if we tail | |
250 | recursed, we would reexamine them uselessly. Just switch BB with | |
251 | DOM, and go on looking for blocks dominated by DOM. */ | |
252 | new_occ = occ_new (dom, new_occ); | |
253 | } | |
254 | ||
255 | else | |
256 | { | |
257 | /* Nothing special, go on with the next element. */ | |
258 | p_occ = &occ->next; | |
259 | } | |
260 | } | |
261 | ||
262 | /* No place was found as a child of IDOM. Make BB a sibling of IDOM. */ | |
263 | new_occ->next = *p_head; | |
264 | *p_head = new_occ; | |
265 | } | |
266 | ||
267 | /* Register that we found a division in BB. */ | |
268 | ||
269 | static inline void | |
270 | register_division_in (basic_block bb) | |
271 | { | |
272 | struct occurrence *occ; | |
273 | ||
274 | occ = (struct occurrence *) bb->aux; | |
275 | if (!occ) | |
276 | { | |
277 | occ = occ_new (bb, NULL); | |
278 | insert_bb (occ, ENTRY_BLOCK_PTR, &occ_head); | |
279 | } | |
280 | ||
281 | occ->bb_has_division = true; | |
282 | occ->num_divisions++; | |
283 | } | |
284 | ||
285 | ||
286 | /* Compute the number of divisions that postdominate each block in OCC and | |
287 | its children. */ | |
abacb398 | 288 | |
abacb398 | 289 | static void |
ac70caad | 290 | compute_merit (struct occurrence *occ) |
abacb398 | 291 | { |
ac70caad | 292 | struct occurrence *occ_child; |
293 | basic_block dom = occ->bb; | |
abacb398 | 294 | |
ac70caad | 295 | for (occ_child = occ->children; occ_child; occ_child = occ_child->next) |
abacb398 | 296 | { |
ac70caad | 297 | basic_block bb; |
298 | if (occ_child->children) | |
299 | compute_merit (occ_child); | |
300 | ||
301 | if (flag_exceptions) | |
302 | bb = single_noncomplex_succ (dom); | |
303 | else | |
304 | bb = dom; | |
305 | ||
306 | if (dominated_by_p (CDI_POST_DOMINATORS, bb, occ_child->bb)) | |
307 | occ->num_divisions += occ_child->num_divisions; | |
308 | } | |
309 | } | |
310 | ||
311 | ||
312 | /* Return whether USE_STMT is a floating-point division by DEF. */ | |
313 | static inline bool | |
75a70cf9 | 314 | is_division_by (gimple use_stmt, tree def) |
ac70caad | 315 | { |
75a70cf9 | 316 | return is_gimple_assign (use_stmt) |
317 | && gimple_assign_rhs_code (use_stmt) == RDIV_EXPR | |
318 | && gimple_assign_rhs2 (use_stmt) == def | |
119368d7 | 319 | /* Do not recognize x / x as valid division, as we are getting |
320 | confused later by replacing all immediate uses x in such | |
321 | a stmt. */ | |
75a70cf9 | 322 | && gimple_assign_rhs1 (use_stmt) != def; |
ac70caad | 323 | } |
324 | ||
325 | /* Walk the subset of the dominator tree rooted at OCC, setting the | |
326 | RECIP_DEF field to a definition of 1.0 / DEF that can be used in | |
327 | the given basic block. The field may be left NULL, of course, | |
328 | if it is not possible or profitable to do the optimization. | |
329 | ||
330 | DEF_BSI is an iterator pointing at the statement defining DEF. | |
331 | If RECIP_DEF is set, a dominator already has a computation that can | |
332 | be used. */ | |
333 | ||
334 | static void | |
75a70cf9 | 335 | insert_reciprocals (gimple_stmt_iterator *def_gsi, struct occurrence *occ, |
ac70caad | 336 | tree def, tree recip_def, int threshold) |
337 | { | |
75a70cf9 | 338 | tree type; |
339 | gimple new_stmt; | |
340 | gimple_stmt_iterator gsi; | |
ac70caad | 341 | struct occurrence *occ_child; |
342 | ||
343 | if (!recip_def | |
344 | && (occ->bb_has_division || !flag_trapping_math) | |
345 | && occ->num_divisions >= threshold) | |
346 | { | |
347 | /* Make a variable with the replacement and substitute it. */ | |
348 | type = TREE_TYPE (def); | |
349 | recip_def = make_rename_temp (type, "reciptmp"); | |
75a70cf9 | 350 | new_stmt = gimple_build_assign_with_ops (RDIV_EXPR, recip_def, |
351 | build_one_cst (type), def); | |
48e1416a | 352 | |
ac70caad | 353 | if (occ->bb_has_division) |
354 | { | |
355 | /* Case 1: insert before an existing division. */ | |
75a70cf9 | 356 | gsi = gsi_after_labels (occ->bb); |
357 | while (!gsi_end_p (gsi) && !is_division_by (gsi_stmt (gsi), def)) | |
358 | gsi_next (&gsi); | |
ac70caad | 359 | |
75a70cf9 | 360 | gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT); |
ac70caad | 361 | } |
75a70cf9 | 362 | else if (def_gsi && occ->bb == def_gsi->bb) |
685b24f5 | 363 | { |
ac70caad | 364 | /* Case 2: insert right after the definition. Note that this will |
365 | never happen if the definition statement can throw, because in | |
366 | that case the sole successor of the statement's basic block will | |
367 | dominate all the uses as well. */ | |
75a70cf9 | 368 | gsi_insert_after (def_gsi, new_stmt, GSI_NEW_STMT); |
685b24f5 | 369 | } |
ac70caad | 370 | else |
371 | { | |
372 | /* Case 3: insert in a basic block not containing defs/uses. */ | |
75a70cf9 | 373 | gsi = gsi_after_labels (occ->bb); |
374 | gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT); | |
ac70caad | 375 | } |
376 | ||
30c4e60d | 377 | reciprocal_stats.rdivs_inserted++; |
378 | ||
ac70caad | 379 | occ->recip_def_stmt = new_stmt; |
abacb398 | 380 | } |
381 | ||
ac70caad | 382 | occ->recip_def = recip_def; |
383 | for (occ_child = occ->children; occ_child; occ_child = occ_child->next) | |
75a70cf9 | 384 | insert_reciprocals (def_gsi, occ_child, def, recip_def, threshold); |
ac70caad | 385 | } |
386 | ||
387 | ||
388 | /* Replace the division at USE_P with a multiplication by the reciprocal, if | |
389 | possible. */ | |
390 | ||
391 | static inline void | |
392 | replace_reciprocal (use_operand_p use_p) | |
393 | { | |
75a70cf9 | 394 | gimple use_stmt = USE_STMT (use_p); |
395 | basic_block bb = gimple_bb (use_stmt); | |
ac70caad | 396 | struct occurrence *occ = (struct occurrence *) bb->aux; |
397 | ||
0bfd8d5c | 398 | if (optimize_bb_for_speed_p (bb) |
399 | && occ->recip_def && use_stmt != occ->recip_def_stmt) | |
ac70caad | 400 | { |
75a70cf9 | 401 | gimple_assign_set_rhs_code (use_stmt, MULT_EXPR); |
ac70caad | 402 | SET_USE (use_p, occ->recip_def); |
403 | fold_stmt_inplace (use_stmt); | |
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)) |
2d04fd8d | 516 | if (gimple_default_def (cfun, arg) |
ac70caad | 517 | && FLOAT_TYPE_P (TREE_TYPE (arg)) |
518 | && is_gimple_reg (arg)) | |
2d04fd8d | 519 | execute_cse_reciprocals_1 (NULL, gimple_default_def (cfun, arg)); |
51b60a11 | 520 | |
abacb398 | 521 | FOR_EACH_BB (bb) |
522 | { | |
75a70cf9 | 523 | gimple_stmt_iterator gsi; |
524 | gimple phi; | |
525 | tree def; | |
abacb398 | 526 | |
75a70cf9 | 527 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
abacb398 | 528 | { |
75a70cf9 | 529 | phi = gsi_stmt (gsi); |
abacb398 | 530 | def = PHI_RESULT (phi); |
531 | if (FLOAT_TYPE_P (TREE_TYPE (def)) | |
532 | && is_gimple_reg (def)) | |
ac70caad | 533 | execute_cse_reciprocals_1 (NULL, def); |
abacb398 | 534 | } |
535 | ||
75a70cf9 | 536 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
abacb398 | 537 | { |
75a70cf9 | 538 | gimple stmt = gsi_stmt (gsi); |
a0315874 | 539 | |
75a70cf9 | 540 | if (gimple_has_lhs (stmt) |
abacb398 | 541 | && (def = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_DEF)) != NULL |
542 | && FLOAT_TYPE_P (TREE_TYPE (def)) | |
51b60a11 | 543 | && TREE_CODE (def) == SSA_NAME) |
75a70cf9 | 544 | execute_cse_reciprocals_1 (&gsi, def); |
abacb398 | 545 | } |
e174638f | 546 | |
0bfd8d5c | 547 | if (optimize_bb_for_size_p (bb)) |
548 | continue; | |
549 | ||
e174638f | 550 | /* Scan for a/func(b) and convert it to reciprocal a*rfunc(b). */ |
75a70cf9 | 551 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
e174638f | 552 | { |
75a70cf9 | 553 | gimple stmt = gsi_stmt (gsi); |
e174638f | 554 | tree fndecl; |
555 | ||
75a70cf9 | 556 | if (is_gimple_assign (stmt) |
557 | && gimple_assign_rhs_code (stmt) == RDIV_EXPR) | |
e174638f | 558 | { |
75a70cf9 | 559 | tree arg1 = gimple_assign_rhs2 (stmt); |
560 | gimple stmt1; | |
2cd360b6 | 561 | |
562 | if (TREE_CODE (arg1) != SSA_NAME) | |
563 | continue; | |
564 | ||
565 | stmt1 = SSA_NAME_DEF_STMT (arg1); | |
e174638f | 566 | |
75a70cf9 | 567 | if (is_gimple_call (stmt1) |
568 | && gimple_call_lhs (stmt1) | |
569 | && (fndecl = gimple_call_fndecl (stmt1)) | |
e174638f | 570 | && (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL |
571 | || DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_MD)) | |
572 | { | |
573 | enum built_in_function code; | |
774b1cdd | 574 | bool md_code, fail; |
575 | imm_use_iterator ui; | |
576 | use_operand_p use_p; | |
e174638f | 577 | |
578 | code = DECL_FUNCTION_CODE (fndecl); | |
2cd360b6 | 579 | md_code = DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_MD; |
580 | ||
581 | fndecl = targetm.builtin_reciprocal (code, md_code, false); | |
e174638f | 582 | if (!fndecl) |
583 | continue; | |
584 | ||
774b1cdd | 585 | /* Check that all uses of the SSA name are divisions, |
586 | otherwise replacing the defining statement will do | |
587 | the wrong thing. */ | |
588 | fail = false; | |
589 | FOR_EACH_IMM_USE_FAST (use_p, ui, arg1) | |
590 | { | |
591 | gimple stmt2 = USE_STMT (use_p); | |
592 | if (is_gimple_debug (stmt2)) | |
593 | continue; | |
594 | if (!is_gimple_assign (stmt2) | |
595 | || gimple_assign_rhs_code (stmt2) != RDIV_EXPR | |
596 | || gimple_assign_rhs1 (stmt2) == arg1 | |
597 | || gimple_assign_rhs2 (stmt2) != arg1) | |
598 | { | |
599 | fail = true; | |
600 | break; | |
601 | } | |
602 | } | |
603 | if (fail) | |
604 | continue; | |
605 | ||
5fb3d93f | 606 | gimple_replace_lhs (stmt1, arg1); |
0acacf9e | 607 | gimple_call_set_fndecl (stmt1, fndecl); |
e174638f | 608 | update_stmt (stmt1); |
30c4e60d | 609 | reciprocal_stats.rfuncs_inserted++; |
e174638f | 610 | |
774b1cdd | 611 | FOR_EACH_IMM_USE_STMT (stmt, ui, arg1) |
612 | { | |
613 | gimple_assign_set_rhs_code (stmt, MULT_EXPR); | |
614 | fold_stmt_inplace (stmt); | |
615 | update_stmt (stmt); | |
616 | } | |
e174638f | 617 | } |
618 | } | |
619 | } | |
abacb398 | 620 | } |
685b24f5 | 621 | |
30c4e60d | 622 | statistics_counter_event (cfun, "reciprocal divs inserted", |
623 | reciprocal_stats.rdivs_inserted); | |
624 | statistics_counter_event (cfun, "reciprocal functions inserted", | |
625 | reciprocal_stats.rfuncs_inserted); | |
626 | ||
c136ae61 | 627 | free_dominance_info (CDI_DOMINATORS); |
628 | free_dominance_info (CDI_POST_DOMINATORS); | |
ac70caad | 629 | free_alloc_pool (occ_pool); |
2a1990e9 | 630 | return 0; |
abacb398 | 631 | } |
632 | ||
20099e35 | 633 | struct gimple_opt_pass pass_cse_reciprocals = |
abacb398 | 634 | { |
20099e35 | 635 | { |
636 | GIMPLE_PASS, | |
abacb398 | 637 | "recip", /* name */ |
638 | gate_cse_reciprocals, /* gate */ | |
639 | execute_cse_reciprocals, /* execute */ | |
640 | NULL, /* sub */ | |
641 | NULL, /* next */ | |
642 | 0, /* static_pass_number */ | |
0b1615c1 | 643 | TV_NONE, /* tv_id */ |
abacb398 | 644 | PROP_ssa, /* properties_required */ |
645 | 0, /* properties_provided */ | |
646 | 0, /* properties_destroyed */ | |
647 | 0, /* todo_flags_start */ | |
648 | TODO_dump_func | TODO_update_ssa | TODO_verify_ssa | |
20099e35 | 649 | | TODO_verify_stmts /* todo_flags_finish */ |
650 | } | |
abacb398 | 651 | }; |
a0315874 | 652 | |
0d424440 | 653 | /* Records an occurrence at statement USE_STMT in the vector of trees |
a0315874 | 654 | STMTS if it is dominated by *TOP_BB or dominates it or this basic block |
0d424440 | 655 | is not yet initialized. Returns true if the occurrence was pushed on |
a0315874 | 656 | the vector. Adjusts *TOP_BB to be the basic block dominating all |
657 | statements in the vector. */ | |
658 | ||
659 | static bool | |
75a70cf9 | 660 | maybe_record_sincos (VEC(gimple, heap) **stmts, |
661 | basic_block *top_bb, gimple use_stmt) | |
a0315874 | 662 | { |
75a70cf9 | 663 | basic_block use_bb = gimple_bb (use_stmt); |
a0315874 | 664 | if (*top_bb |
665 | && (*top_bb == use_bb | |
666 | || dominated_by_p (CDI_DOMINATORS, use_bb, *top_bb))) | |
75a70cf9 | 667 | VEC_safe_push (gimple, heap, *stmts, use_stmt); |
a0315874 | 668 | else if (!*top_bb |
669 | || dominated_by_p (CDI_DOMINATORS, *top_bb, use_bb)) | |
670 | { | |
75a70cf9 | 671 | VEC_safe_push (gimple, heap, *stmts, use_stmt); |
a0315874 | 672 | *top_bb = use_bb; |
673 | } | |
674 | else | |
675 | return false; | |
676 | ||
677 | return true; | |
678 | } | |
679 | ||
680 | /* Look for sin, cos and cexpi calls with the same argument NAME and | |
681 | create a single call to cexpi CSEing the result in this case. | |
682 | We first walk over all immediate uses of the argument collecting | |
683 | statements that we can CSE in a vector and in a second pass replace | |
684 | the statement rhs with a REALPART or IMAGPART expression on the | |
685 | result of the cexpi call we insert before the use statement that | |
686 | dominates all other candidates. */ | |
687 | ||
4c80086d | 688 | static bool |
a0315874 | 689 | execute_cse_sincos_1 (tree name) |
690 | { | |
75a70cf9 | 691 | gimple_stmt_iterator gsi; |
a0315874 | 692 | imm_use_iterator use_iter; |
75a70cf9 | 693 | tree fndecl, res, type; |
694 | gimple def_stmt, use_stmt, stmt; | |
a0315874 | 695 | int seen_cos = 0, seen_sin = 0, seen_cexpi = 0; |
75a70cf9 | 696 | VEC(gimple, heap) *stmts = NULL; |
a0315874 | 697 | basic_block top_bb = NULL; |
698 | int i; | |
4c80086d | 699 | bool cfg_changed = false; |
a0315874 | 700 | |
701 | type = TREE_TYPE (name); | |
702 | FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, name) | |
703 | { | |
75a70cf9 | 704 | if (gimple_code (use_stmt) != GIMPLE_CALL |
705 | || !gimple_call_lhs (use_stmt) | |
706 | || !(fndecl = gimple_call_fndecl (use_stmt)) | |
a0315874 | 707 | || DECL_BUILT_IN_CLASS (fndecl) != BUILT_IN_NORMAL) |
708 | continue; | |
709 | ||
710 | switch (DECL_FUNCTION_CODE (fndecl)) | |
711 | { | |
712 | CASE_FLT_FN (BUILT_IN_COS): | |
713 | seen_cos |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0; | |
714 | break; | |
715 | ||
716 | CASE_FLT_FN (BUILT_IN_SIN): | |
717 | seen_sin |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0; | |
718 | break; | |
719 | ||
720 | CASE_FLT_FN (BUILT_IN_CEXPI): | |
721 | seen_cexpi |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0; | |
722 | break; | |
723 | ||
724 | default:; | |
725 | } | |
726 | } | |
727 | ||
728 | if (seen_cos + seen_sin + seen_cexpi <= 1) | |
729 | { | |
75a70cf9 | 730 | VEC_free(gimple, heap, stmts); |
4c80086d | 731 | return false; |
a0315874 | 732 | } |
733 | ||
734 | /* Simply insert cexpi at the beginning of top_bb but not earlier than | |
735 | the name def statement. */ | |
736 | fndecl = mathfn_built_in (type, BUILT_IN_CEXPI); | |
737 | if (!fndecl) | |
4c80086d | 738 | return false; |
739 | res = create_tmp_reg (TREE_TYPE (TREE_TYPE (fndecl)), "sincostmp"); | |
75a70cf9 | 740 | stmt = gimple_build_call (fndecl, 1, name); |
4c80086d | 741 | res = make_ssa_name (res, stmt); |
75a70cf9 | 742 | gimple_call_set_lhs (stmt, res); |
743 | ||
a0315874 | 744 | def_stmt = SSA_NAME_DEF_STMT (name); |
8090c12d | 745 | if (!SSA_NAME_IS_DEFAULT_DEF (name) |
75a70cf9 | 746 | && gimple_code (def_stmt) != GIMPLE_PHI |
747 | && gimple_bb (def_stmt) == top_bb) | |
a0315874 | 748 | { |
75a70cf9 | 749 | gsi = gsi_for_stmt (def_stmt); |
750 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); | |
a0315874 | 751 | } |
752 | else | |
753 | { | |
75a70cf9 | 754 | gsi = gsi_after_labels (top_bb); |
755 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
a0315874 | 756 | } |
757 | update_stmt (stmt); | |
30c4e60d | 758 | sincos_stats.inserted++; |
a0315874 | 759 | |
760 | /* And adjust the recorded old call sites. */ | |
75a70cf9 | 761 | for (i = 0; VEC_iterate(gimple, stmts, i, use_stmt); ++i) |
a0315874 | 762 | { |
75a70cf9 | 763 | tree rhs = NULL; |
764 | fndecl = gimple_call_fndecl (use_stmt); | |
765 | ||
a0315874 | 766 | switch (DECL_FUNCTION_CODE (fndecl)) |
767 | { | |
768 | CASE_FLT_FN (BUILT_IN_COS): | |
75a70cf9 | 769 | rhs = fold_build1 (REALPART_EXPR, type, res); |
a0315874 | 770 | break; |
771 | ||
772 | CASE_FLT_FN (BUILT_IN_SIN): | |
75a70cf9 | 773 | rhs = fold_build1 (IMAGPART_EXPR, type, res); |
a0315874 | 774 | break; |
775 | ||
776 | CASE_FLT_FN (BUILT_IN_CEXPI): | |
75a70cf9 | 777 | rhs = res; |
a0315874 | 778 | break; |
779 | ||
780 | default:; | |
781 | gcc_unreachable (); | |
782 | } | |
783 | ||
75a70cf9 | 784 | /* Replace call with a copy. */ |
785 | stmt = gimple_build_assign (gimple_call_lhs (use_stmt), rhs); | |
786 | ||
787 | gsi = gsi_for_stmt (use_stmt); | |
4c80086d | 788 | gsi_replace (&gsi, stmt, true); |
789 | if (gimple_purge_dead_eh_edges (gimple_bb (stmt))) | |
790 | cfg_changed = true; | |
a0315874 | 791 | } |
792 | ||
75a70cf9 | 793 | VEC_free(gimple, heap, stmts); |
4c80086d | 794 | |
795 | return cfg_changed; | |
a0315874 | 796 | } |
797 | ||
e9a6c4bc | 798 | /* To evaluate powi(x,n), the floating point value x raised to the |
799 | constant integer exponent n, we use a hybrid algorithm that | |
800 | combines the "window method" with look-up tables. For an | |
801 | introduction to exponentiation algorithms and "addition chains", | |
802 | see section 4.6.3, "Evaluation of Powers" of Donald E. Knuth, | |
803 | "Seminumerical Algorithms", Vol. 2, "The Art of Computer Programming", | |
804 | 3rd Edition, 1998, and Daniel M. Gordon, "A Survey of Fast Exponentiation | |
805 | Methods", Journal of Algorithms, Vol. 27, pp. 129-146, 1998. */ | |
806 | ||
807 | /* Provide a default value for POWI_MAX_MULTS, the maximum number of | |
808 | multiplications to inline before calling the system library's pow | |
809 | function. powi(x,n) requires at worst 2*bits(n)-2 multiplications, | |
810 | so this default never requires calling pow, powf or powl. */ | |
811 | ||
812 | #ifndef POWI_MAX_MULTS | |
813 | #define POWI_MAX_MULTS (2*HOST_BITS_PER_WIDE_INT-2) | |
814 | #endif | |
815 | ||
816 | /* The size of the "optimal power tree" lookup table. All | |
817 | exponents less than this value are simply looked up in the | |
818 | powi_table below. This threshold is also used to size the | |
819 | cache of pseudo registers that hold intermediate results. */ | |
820 | #define POWI_TABLE_SIZE 256 | |
821 | ||
822 | /* The size, in bits of the window, used in the "window method" | |
823 | exponentiation algorithm. This is equivalent to a radix of | |
824 | (1<<POWI_WINDOW_SIZE) in the corresponding "m-ary method". */ | |
825 | #define POWI_WINDOW_SIZE 3 | |
826 | ||
827 | /* The following table is an efficient representation of an | |
828 | "optimal power tree". For each value, i, the corresponding | |
829 | value, j, in the table states than an optimal evaluation | |
830 | sequence for calculating pow(x,i) can be found by evaluating | |
831 | pow(x,j)*pow(x,i-j). An optimal power tree for the first | |
832 | 100 integers is given in Knuth's "Seminumerical algorithms". */ | |
833 | ||
834 | static const unsigned char powi_table[POWI_TABLE_SIZE] = | |
835 | { | |
836 | 0, 1, 1, 2, 2, 3, 3, 4, /* 0 - 7 */ | |
837 | 4, 6, 5, 6, 6, 10, 7, 9, /* 8 - 15 */ | |
838 | 8, 16, 9, 16, 10, 12, 11, 13, /* 16 - 23 */ | |
839 | 12, 17, 13, 18, 14, 24, 15, 26, /* 24 - 31 */ | |
840 | 16, 17, 17, 19, 18, 33, 19, 26, /* 32 - 39 */ | |
841 | 20, 25, 21, 40, 22, 27, 23, 44, /* 40 - 47 */ | |
842 | 24, 32, 25, 34, 26, 29, 27, 44, /* 48 - 55 */ | |
843 | 28, 31, 29, 34, 30, 60, 31, 36, /* 56 - 63 */ | |
844 | 32, 64, 33, 34, 34, 46, 35, 37, /* 64 - 71 */ | |
845 | 36, 65, 37, 50, 38, 48, 39, 69, /* 72 - 79 */ | |
846 | 40, 49, 41, 43, 42, 51, 43, 58, /* 80 - 87 */ | |
847 | 44, 64, 45, 47, 46, 59, 47, 76, /* 88 - 95 */ | |
848 | 48, 65, 49, 66, 50, 67, 51, 66, /* 96 - 103 */ | |
849 | 52, 70, 53, 74, 54, 104, 55, 74, /* 104 - 111 */ | |
850 | 56, 64, 57, 69, 58, 78, 59, 68, /* 112 - 119 */ | |
851 | 60, 61, 61, 80, 62, 75, 63, 68, /* 120 - 127 */ | |
852 | 64, 65, 65, 128, 66, 129, 67, 90, /* 128 - 135 */ | |
853 | 68, 73, 69, 131, 70, 94, 71, 88, /* 136 - 143 */ | |
854 | 72, 128, 73, 98, 74, 132, 75, 121, /* 144 - 151 */ | |
855 | 76, 102, 77, 124, 78, 132, 79, 106, /* 152 - 159 */ | |
856 | 80, 97, 81, 160, 82, 99, 83, 134, /* 160 - 167 */ | |
857 | 84, 86, 85, 95, 86, 160, 87, 100, /* 168 - 175 */ | |
858 | 88, 113, 89, 98, 90, 107, 91, 122, /* 176 - 183 */ | |
859 | 92, 111, 93, 102, 94, 126, 95, 150, /* 184 - 191 */ | |
860 | 96, 128, 97, 130, 98, 133, 99, 195, /* 192 - 199 */ | |
861 | 100, 128, 101, 123, 102, 164, 103, 138, /* 200 - 207 */ | |
862 | 104, 145, 105, 146, 106, 109, 107, 149, /* 208 - 215 */ | |
863 | 108, 200, 109, 146, 110, 170, 111, 157, /* 216 - 223 */ | |
864 | 112, 128, 113, 130, 114, 182, 115, 132, /* 224 - 231 */ | |
865 | 116, 200, 117, 132, 118, 158, 119, 206, /* 232 - 239 */ | |
866 | 120, 240, 121, 162, 122, 147, 123, 152, /* 240 - 247 */ | |
867 | 124, 166, 125, 214, 126, 138, 127, 153, /* 248 - 255 */ | |
868 | }; | |
869 | ||
870 | ||
871 | /* Return the number of multiplications required to calculate | |
872 | powi(x,n) where n is less than POWI_TABLE_SIZE. This is a | |
873 | subroutine of powi_cost. CACHE is an array indicating | |
874 | which exponents have already been calculated. */ | |
875 | ||
876 | static int | |
877 | powi_lookup_cost (unsigned HOST_WIDE_INT n, bool *cache) | |
878 | { | |
879 | /* If we've already calculated this exponent, then this evaluation | |
880 | doesn't require any additional multiplications. */ | |
881 | if (cache[n]) | |
882 | return 0; | |
883 | ||
884 | cache[n] = true; | |
885 | return powi_lookup_cost (n - powi_table[n], cache) | |
886 | + powi_lookup_cost (powi_table[n], cache) + 1; | |
887 | } | |
888 | ||
889 | /* Return the number of multiplications required to calculate | |
890 | powi(x,n) for an arbitrary x, given the exponent N. This | |
891 | function needs to be kept in sync with powi_as_mults below. */ | |
892 | ||
893 | static int | |
894 | powi_cost (HOST_WIDE_INT n) | |
895 | { | |
896 | bool cache[POWI_TABLE_SIZE]; | |
897 | unsigned HOST_WIDE_INT digit; | |
898 | unsigned HOST_WIDE_INT val; | |
899 | int result; | |
900 | ||
901 | if (n == 0) | |
902 | return 0; | |
903 | ||
904 | /* Ignore the reciprocal when calculating the cost. */ | |
905 | val = (n < 0) ? -n : n; | |
906 | ||
907 | /* Initialize the exponent cache. */ | |
908 | memset (cache, 0, POWI_TABLE_SIZE * sizeof (bool)); | |
909 | cache[1] = true; | |
910 | ||
911 | result = 0; | |
912 | ||
913 | while (val >= POWI_TABLE_SIZE) | |
914 | { | |
915 | if (val & 1) | |
916 | { | |
917 | digit = val & ((1 << POWI_WINDOW_SIZE) - 1); | |
918 | result += powi_lookup_cost (digit, cache) | |
919 | + POWI_WINDOW_SIZE + 1; | |
920 | val >>= POWI_WINDOW_SIZE; | |
921 | } | |
922 | else | |
923 | { | |
924 | val >>= 1; | |
925 | result++; | |
926 | } | |
927 | } | |
928 | ||
929 | return result + powi_lookup_cost (val, cache); | |
930 | } | |
931 | ||
932 | /* Recursive subroutine of powi_as_mults. This function takes the | |
933 | array, CACHE, of already calculated exponents and an exponent N and | |
934 | returns a tree that corresponds to CACHE[1]**N, with type TYPE. */ | |
935 | ||
936 | static tree | |
937 | powi_as_mults_1 (gimple_stmt_iterator *gsi, location_t loc, tree type, | |
938 | HOST_WIDE_INT n, tree *cache, tree target) | |
939 | { | |
940 | tree op0, op1, ssa_target; | |
941 | unsigned HOST_WIDE_INT digit; | |
942 | gimple mult_stmt; | |
943 | ||
944 | if (n < POWI_TABLE_SIZE && cache[n]) | |
945 | return cache[n]; | |
946 | ||
947 | ssa_target = make_ssa_name (target, NULL); | |
948 | ||
949 | if (n < POWI_TABLE_SIZE) | |
950 | { | |
951 | cache[n] = ssa_target; | |
952 | op0 = powi_as_mults_1 (gsi, loc, type, n - powi_table[n], cache, target); | |
953 | op1 = powi_as_mults_1 (gsi, loc, type, powi_table[n], cache, target); | |
954 | } | |
955 | else if (n & 1) | |
956 | { | |
957 | digit = n & ((1 << POWI_WINDOW_SIZE) - 1); | |
958 | op0 = powi_as_mults_1 (gsi, loc, type, n - digit, cache, target); | |
959 | op1 = powi_as_mults_1 (gsi, loc, type, digit, cache, target); | |
960 | } | |
961 | else | |
962 | { | |
963 | op0 = powi_as_mults_1 (gsi, loc, type, n >> 1, cache, target); | |
964 | op1 = op0; | |
965 | } | |
966 | ||
967 | mult_stmt = gimple_build_assign_with_ops (MULT_EXPR, ssa_target, op0, op1); | |
ae43b05e | 968 | gimple_set_location (mult_stmt, loc); |
e9a6c4bc | 969 | gsi_insert_before (gsi, mult_stmt, GSI_SAME_STMT); |
970 | ||
971 | return ssa_target; | |
972 | } | |
973 | ||
974 | /* Convert ARG0**N to a tree of multiplications of ARG0 with itself. | |
975 | This function needs to be kept in sync with powi_cost above. */ | |
976 | ||
977 | static tree | |
978 | powi_as_mults (gimple_stmt_iterator *gsi, location_t loc, | |
979 | tree arg0, HOST_WIDE_INT n) | |
980 | { | |
981 | tree cache[POWI_TABLE_SIZE], result, type = TREE_TYPE (arg0), target; | |
982 | gimple div_stmt; | |
983 | ||
984 | if (n == 0) | |
985 | return build_real (type, dconst1); | |
986 | ||
987 | memset (cache, 0, sizeof (cache)); | |
988 | cache[1] = arg0; | |
989 | ||
990 | target = create_tmp_var (type, "powmult"); | |
991 | add_referenced_var (target); | |
992 | ||
993 | result = powi_as_mults_1 (gsi, loc, type, (n < 0) ? -n : n, cache, target); | |
994 | ||
995 | if (n >= 0) | |
996 | return result; | |
997 | ||
998 | /* If the original exponent was negative, reciprocate the result. */ | |
999 | target = make_ssa_name (target, NULL); | |
1000 | div_stmt = gimple_build_assign_with_ops (RDIV_EXPR, target, | |
1001 | build_real (type, dconst1), | |
1002 | result); | |
ae43b05e | 1003 | gimple_set_location (div_stmt, loc); |
e9a6c4bc | 1004 | gsi_insert_before (gsi, div_stmt, GSI_SAME_STMT); |
1005 | ||
1006 | return target; | |
1007 | } | |
1008 | ||
1009 | /* ARG0 and N are the two arguments to a powi builtin in GSI with | |
1010 | location info LOC. If the arguments are appropriate, create an | |
1011 | equivalent sequence of statements prior to GSI using an optimal | |
1012 | number of multiplications, and return an expession holding the | |
1013 | result. */ | |
1014 | ||
1015 | static tree | |
1016 | gimple_expand_builtin_powi (gimple_stmt_iterator *gsi, location_t loc, | |
1017 | tree arg0, HOST_WIDE_INT n) | |
1018 | { | |
1019 | /* Avoid largest negative number. */ | |
1020 | if (n != -n | |
1021 | && ((n >= -1 && n <= 2) | |
1022 | || (optimize_function_for_speed_p (cfun) | |
1023 | && powi_cost (n) <= POWI_MAX_MULTS))) | |
1024 | return powi_as_mults (gsi, loc, arg0, n); | |
1025 | ||
1026 | return NULL_TREE; | |
1027 | } | |
1028 | ||
ae43b05e | 1029 | /* Build a gimple call statement that calls FN with argument ARG. |
1030 | Set the lhs of the call statement to a fresh SSA name for | |
1031 | variable VAR. If VAR is NULL, first allocate it. Insert the | |
1032 | statement prior to GSI's current position, and return the fresh | |
1033 | SSA name. */ | |
1034 | ||
1035 | static tree | |
ca12eb68 | 1036 | build_and_insert_call (gimple_stmt_iterator *gsi, location_t loc, |
1037 | tree *var, tree fn, tree arg) | |
ae43b05e | 1038 | { |
1039 | gimple call_stmt; | |
1040 | tree ssa_target; | |
1041 | ||
1042 | if (!*var) | |
1043 | { | |
1044 | *var = create_tmp_var (TREE_TYPE (arg), "powroot"); | |
1045 | add_referenced_var (*var); | |
1046 | } | |
1047 | ||
1048 | call_stmt = gimple_build_call (fn, 1, arg); | |
1049 | ssa_target = make_ssa_name (*var, NULL); | |
1050 | gimple_set_lhs (call_stmt, ssa_target); | |
1051 | gimple_set_location (call_stmt, loc); | |
1052 | gsi_insert_before (gsi, call_stmt, GSI_SAME_STMT); | |
1053 | ||
1054 | return ssa_target; | |
1055 | } | |
1056 | ||
ca12eb68 | 1057 | /* Build a gimple binary operation with the given CODE and arguments |
1058 | ARG0, ARG1, assigning the result to a new SSA name for variable | |
1059 | TARGET. Insert the statement prior to GSI's current position, and | |
1060 | return the fresh SSA name.*/ | |
1061 | ||
1062 | static tree | |
1063 | build_and_insert_binop (gimple_stmt_iterator *gsi, location_t loc, | |
1064 | tree target, enum tree_code code, tree arg0, tree arg1) | |
1065 | { | |
1066 | tree result = make_ssa_name (target, NULL); | |
1067 | gimple stmt = gimple_build_assign_with_ops (code, result, arg0, arg1); | |
1068 | gimple_set_location (stmt, loc); | |
1069 | gsi_insert_before (gsi, stmt, GSI_SAME_STMT); | |
1070 | return result; | |
1071 | } | |
1072 | ||
e78306af | 1073 | /* ARG0 and ARG1 are the two arguments to a pow builtin call in GSI |
1074 | with location info LOC. If possible, create an equivalent and | |
1075 | less expensive sequence of statements prior to GSI, and return an | |
1076 | expession holding the result. */ | |
1077 | ||
1078 | static tree | |
1079 | gimple_expand_builtin_pow (gimple_stmt_iterator *gsi, location_t loc, | |
1080 | tree arg0, tree arg1) | |
1081 | { | |
ae43b05e | 1082 | REAL_VALUE_TYPE c, cint, dconst1_4, dconst3_4, dconst1_3, dconst1_6; |
ca12eb68 | 1083 | REAL_VALUE_TYPE c2, dconst3; |
e78306af | 1084 | HOST_WIDE_INT n; |
ca12eb68 | 1085 | tree type, sqrtfn, cbrtfn, sqrt_arg0, sqrt_sqrt, result, cbrt_x, powi_cbrt_x; |
ae43b05e | 1086 | tree target = NULL_TREE; |
1087 | enum machine_mode mode; | |
1088 | bool hw_sqrt_exists; | |
e78306af | 1089 | |
1090 | /* If the exponent isn't a constant, there's nothing of interest | |
1091 | to be done. */ | |
1092 | if (TREE_CODE (arg1) != REAL_CST) | |
1093 | return NULL_TREE; | |
1094 | ||
ae43b05e | 1095 | /* If the exponent is equivalent to an integer, expand to an optimal |
1096 | multiplication sequence when profitable. */ | |
e78306af | 1097 | c = TREE_REAL_CST (arg1); |
1098 | n = real_to_integer (&c); | |
1099 | real_from_integer (&cint, VOIDmode, n, n < 0 ? -1 : 0, 0); | |
1100 | ||
1101 | if (real_identical (&c, &cint) | |
1102 | && ((n >= -1 && n <= 2) | |
1103 | || (flag_unsafe_math_optimizations | |
1104 | && optimize_insn_for_speed_p () | |
1105 | && powi_cost (n) <= POWI_MAX_MULTS))) | |
1106 | return gimple_expand_builtin_powi (gsi, loc, arg0, n); | |
1107 | ||
ae43b05e | 1108 | /* Attempt various optimizations using sqrt and cbrt. */ |
1109 | type = TREE_TYPE (arg0); | |
1110 | mode = TYPE_MODE (type); | |
1111 | sqrtfn = mathfn_built_in (type, BUILT_IN_SQRT); | |
1112 | ||
1113 | /* Optimize pow(x,0.5) = sqrt(x). This replacement is always safe | |
1114 | unless signed zeros must be maintained. pow(-0,0.5) = +0, while | |
1115 | sqrt(-0) = -0. */ | |
1116 | if (sqrtfn | |
1117 | && REAL_VALUES_EQUAL (c, dconsthalf) | |
1118 | && !HONOR_SIGNED_ZEROS (mode)) | |
ca12eb68 | 1119 | return build_and_insert_call (gsi, loc, &target, sqrtfn, arg0); |
ae43b05e | 1120 | |
1121 | /* Optimize pow(x,0.25) = sqrt(sqrt(x)). Assume on most machines that | |
1122 | a builtin sqrt instruction is smaller than a call to pow with 0.25, | |
1123 | so do this optimization even if -Os. Don't do this optimization | |
1124 | if we don't have a hardware sqrt insn. */ | |
1125 | dconst1_4 = dconst1; | |
1126 | SET_REAL_EXP (&dconst1_4, REAL_EXP (&dconst1_4) - 2); | |
1127 | hw_sqrt_exists = optab_handler(sqrt_optab, mode) != CODE_FOR_nothing; | |
1128 | ||
1129 | if (flag_unsafe_math_optimizations | |
1130 | && sqrtfn | |
1131 | && REAL_VALUES_EQUAL (c, dconst1_4) | |
1132 | && hw_sqrt_exists) | |
1133 | { | |
1134 | /* sqrt(x) */ | |
ca12eb68 | 1135 | sqrt_arg0 = build_and_insert_call (gsi, loc, &target, sqrtfn, arg0); |
ae43b05e | 1136 | |
1137 | /* sqrt(sqrt(x)) */ | |
ca12eb68 | 1138 | return build_and_insert_call (gsi, loc, &target, sqrtfn, sqrt_arg0); |
ae43b05e | 1139 | } |
1140 | ||
1141 | /* Optimize pow(x,0.75) = sqrt(x) * sqrt(sqrt(x)) unless we are | |
1142 | optimizing for space. Don't do this optimization if we don't have | |
1143 | a hardware sqrt insn. */ | |
1144 | real_from_integer (&dconst3_4, VOIDmode, 3, 0, 0); | |
1145 | SET_REAL_EXP (&dconst3_4, REAL_EXP (&dconst3_4) - 2); | |
1146 | ||
1147 | if (flag_unsafe_math_optimizations | |
1148 | && sqrtfn | |
1149 | && optimize_function_for_speed_p (cfun) | |
1150 | && REAL_VALUES_EQUAL (c, dconst3_4) | |
1151 | && hw_sqrt_exists) | |
1152 | { | |
1153 | /* sqrt(x) */ | |
ca12eb68 | 1154 | sqrt_arg0 = build_and_insert_call (gsi, loc, &target, sqrtfn, arg0); |
ae43b05e | 1155 | |
1156 | /* sqrt(sqrt(x)) */ | |
ca12eb68 | 1157 | sqrt_sqrt = build_and_insert_call (gsi, loc, &target, sqrtfn, sqrt_arg0); |
ae43b05e | 1158 | |
1159 | /* sqrt(x) * sqrt(sqrt(x)) */ | |
ca12eb68 | 1160 | return build_and_insert_binop (gsi, loc, target, MULT_EXPR, |
1161 | sqrt_arg0, sqrt_sqrt); | |
ae43b05e | 1162 | } |
1163 | ||
1164 | /* Optimize pow(x,1./3.) = cbrt(x). This requires unsafe math | |
1165 | optimizations since 1./3. is not exactly representable. If x | |
1166 | is negative and finite, the correct value of pow(x,1./3.) is | |
1167 | a NaN with the "invalid" exception raised, because the value | |
1168 | of 1./3. actually has an even denominator. The correct value | |
1169 | of cbrt(x) is a negative real value. */ | |
1170 | cbrtfn = mathfn_built_in (type, BUILT_IN_CBRT); | |
1171 | dconst1_3 = real_value_truncate (mode, dconst_third ()); | |
1172 | ||
1173 | if (flag_unsafe_math_optimizations | |
1174 | && cbrtfn | |
0b7ad900 | 1175 | && (gimple_val_nonnegative_real_p (arg0) || !HONOR_NANS (mode)) |
ae43b05e | 1176 | && REAL_VALUES_EQUAL (c, dconst1_3)) |
ca12eb68 | 1177 | return build_and_insert_call (gsi, loc, &target, cbrtfn, arg0); |
ae43b05e | 1178 | |
1179 | /* Optimize pow(x,1./6.) = cbrt(sqrt(x)). Don't do this optimization | |
1180 | if we don't have a hardware sqrt insn. */ | |
1181 | dconst1_6 = dconst1_3; | |
1182 | SET_REAL_EXP (&dconst1_6, REAL_EXP (&dconst1_6) - 1); | |
1183 | ||
1184 | if (flag_unsafe_math_optimizations | |
1185 | && sqrtfn | |
1186 | && cbrtfn | |
0b7ad900 | 1187 | && (gimple_val_nonnegative_real_p (arg0) || !HONOR_NANS (mode)) |
ae43b05e | 1188 | && optimize_function_for_speed_p (cfun) |
1189 | && hw_sqrt_exists | |
1190 | && REAL_VALUES_EQUAL (c, dconst1_6)) | |
1191 | { | |
1192 | /* sqrt(x) */ | |
ca12eb68 | 1193 | sqrt_arg0 = build_and_insert_call (gsi, loc, &target, sqrtfn, arg0); |
ae43b05e | 1194 | |
1195 | /* cbrt(sqrt(x)) */ | |
ca12eb68 | 1196 | return build_and_insert_call (gsi, loc, &target, cbrtfn, sqrt_arg0); |
1197 | } | |
1198 | ||
1199 | /* Optimize pow(x,c), where n = 2c for some nonzero integer n, into | |
1200 | ||
1201 | sqrt(x) * powi(x, n/2), n > 0; | |
1202 | 1.0 / (sqrt(x) * powi(x, abs(n/2))), n < 0. | |
1203 | ||
1204 | Do not calculate the powi factor when n/2 = 0. */ | |
1205 | real_arithmetic (&c2, MULT_EXPR, &c, &dconst2); | |
1206 | n = real_to_integer (&c2); | |
1207 | real_from_integer (&cint, VOIDmode, n, n < 0 ? -1 : 0, 0); | |
1208 | ||
1209 | if (flag_unsafe_math_optimizations | |
1210 | && sqrtfn | |
1211 | && real_identical (&c2, &cint)) | |
1212 | { | |
1213 | tree powi_x_ndiv2 = NULL_TREE; | |
1214 | ||
1215 | /* Attempt to fold powi(arg0, abs(n/2)) into multiplies. If not | |
1216 | possible or profitable, give up. Skip the degenerate case when | |
1217 | n is 1 or -1, where the result is always 1. */ | |
1218 | if (abs (n) != 1) | |
1219 | { | |
1220 | powi_x_ndiv2 = gimple_expand_builtin_powi (gsi, loc, arg0, abs(n/2)); | |
1221 | if (!powi_x_ndiv2) | |
1222 | return NULL_TREE; | |
1223 | } | |
1224 | ||
1225 | /* Calculate sqrt(x). When n is not 1 or -1, multiply it by the | |
1226 | result of the optimal multiply sequence just calculated. */ | |
1227 | sqrt_arg0 = build_and_insert_call (gsi, loc, &target, sqrtfn, arg0); | |
1228 | ||
1229 | if (abs (n) == 1) | |
1230 | result = sqrt_arg0; | |
1231 | else | |
1232 | result = build_and_insert_binop (gsi, loc, target, MULT_EXPR, | |
1233 | sqrt_arg0, powi_x_ndiv2); | |
1234 | ||
1235 | /* If n is negative, reciprocate the result. */ | |
1236 | if (n < 0) | |
1237 | result = build_and_insert_binop (gsi, loc, target, RDIV_EXPR, | |
1238 | build_real (type, dconst1), result); | |
1239 | return result; | |
1240 | } | |
1241 | ||
1242 | /* Optimize pow(x,c), where 3c = n for some nonzero integer n, into | |
1243 | ||
1244 | powi(x, n/3) * powi(cbrt(x), n%3), n > 0; | |
1245 | 1.0 / (powi(x, abs(n)/3) * powi(cbrt(x), abs(n)%3)), n < 0. | |
1246 | ||
1247 | Do not calculate the first factor when n/3 = 0. As cbrt(x) is | |
1248 | different from pow(x, 1./3.) due to rounding and behavior with | |
1249 | negative x, we need to constrain this transformation to unsafe | |
1250 | math and positive x or finite math. */ | |
1251 | real_from_integer (&dconst3, VOIDmode, 3, 0, 0); | |
1252 | real_arithmetic (&c2, MULT_EXPR, &c, &dconst3); | |
1253 | real_round (&c2, mode, &c2); | |
1254 | n = real_to_integer (&c2); | |
1255 | real_from_integer (&cint, VOIDmode, n, n < 0 ? -1 : 0, 0); | |
1256 | real_arithmetic (&c2, RDIV_EXPR, &cint, &dconst3); | |
1257 | real_convert (&c2, mode, &c2); | |
1258 | ||
1259 | if (flag_unsafe_math_optimizations | |
1260 | && cbrtfn | |
0b7ad900 | 1261 | && (gimple_val_nonnegative_real_p (arg0) || !HONOR_NANS (mode)) |
ca12eb68 | 1262 | && real_identical (&c2, &c) |
1263 | && optimize_function_for_speed_p (cfun) | |
1264 | && powi_cost (n / 3) <= POWI_MAX_MULTS) | |
1265 | { | |
1266 | tree powi_x_ndiv3 = NULL_TREE; | |
1267 | ||
1268 | /* Attempt to fold powi(arg0, abs(n/3)) into multiplies. If not | |
1269 | possible or profitable, give up. Skip the degenerate case when | |
1270 | abs(n) < 3, where the result is always 1. */ | |
1271 | if (abs (n) >= 3) | |
1272 | { | |
1273 | powi_x_ndiv3 = gimple_expand_builtin_powi (gsi, loc, arg0, | |
1274 | abs (n / 3)); | |
1275 | if (!powi_x_ndiv3) | |
1276 | return NULL_TREE; | |
1277 | } | |
1278 | ||
1279 | /* Calculate powi(cbrt(x), n%3). Don't use gimple_expand_builtin_powi | |
1280 | as that creates an unnecessary variable. Instead, just produce | |
1281 | either cbrt(x) or cbrt(x) * cbrt(x). */ | |
1282 | cbrt_x = build_and_insert_call (gsi, loc, &target, cbrtfn, arg0); | |
1283 | ||
1284 | if (abs (n) % 3 == 1) | |
1285 | powi_cbrt_x = cbrt_x; | |
1286 | else | |
1287 | powi_cbrt_x = build_and_insert_binop (gsi, loc, target, MULT_EXPR, | |
1288 | cbrt_x, cbrt_x); | |
1289 | ||
1290 | /* Multiply the two subexpressions, unless powi(x,abs(n)/3) = 1. */ | |
1291 | if (abs (n) < 3) | |
1292 | result = powi_cbrt_x; | |
1293 | else | |
1294 | result = build_and_insert_binop (gsi, loc, target, MULT_EXPR, | |
1295 | powi_x_ndiv3, powi_cbrt_x); | |
1296 | ||
1297 | /* If n is negative, reciprocate the result. */ | |
1298 | if (n < 0) | |
1299 | result = build_and_insert_binop (gsi, loc, target, RDIV_EXPR, | |
1300 | build_real (type, dconst1), result); | |
1301 | ||
1302 | return result; | |
ae43b05e | 1303 | } |
1304 | ||
ca12eb68 | 1305 | /* No optimizations succeeded. */ |
e78306af | 1306 | return NULL_TREE; |
1307 | } | |
1308 | ||
a0315874 | 1309 | /* Go through all calls to sin, cos and cexpi and call execute_cse_sincos_1 |
e9a6c4bc | 1310 | on the SSA_NAME argument of each of them. Also expand powi(x,n) into |
1311 | an optimal number of multiplies, when n is a constant. */ | |
a0315874 | 1312 | |
1313 | static unsigned int | |
1314 | execute_cse_sincos (void) | |
1315 | { | |
1316 | basic_block bb; | |
4c80086d | 1317 | bool cfg_changed = false; |
a0315874 | 1318 | |
1319 | calculate_dominance_info (CDI_DOMINATORS); | |
30c4e60d | 1320 | memset (&sincos_stats, 0, sizeof (sincos_stats)); |
a0315874 | 1321 | |
1322 | FOR_EACH_BB (bb) | |
1323 | { | |
75a70cf9 | 1324 | gimple_stmt_iterator gsi; |
a0315874 | 1325 | |
75a70cf9 | 1326 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
a0315874 | 1327 | { |
75a70cf9 | 1328 | gimple stmt = gsi_stmt (gsi); |
a0315874 | 1329 | tree fndecl; |
1330 | ||
75a70cf9 | 1331 | if (is_gimple_call (stmt) |
1332 | && gimple_call_lhs (stmt) | |
1333 | && (fndecl = gimple_call_fndecl (stmt)) | |
a0315874 | 1334 | && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) |
1335 | { | |
e9a6c4bc | 1336 | tree arg, arg0, arg1, result; |
1337 | HOST_WIDE_INT n; | |
1338 | location_t loc; | |
a0315874 | 1339 | |
1340 | switch (DECL_FUNCTION_CODE (fndecl)) | |
1341 | { | |
1342 | CASE_FLT_FN (BUILT_IN_COS): | |
1343 | CASE_FLT_FN (BUILT_IN_SIN): | |
1344 | CASE_FLT_FN (BUILT_IN_CEXPI): | |
d312d7df | 1345 | /* Make sure we have either sincos or cexp. */ |
1346 | if (!TARGET_HAS_SINCOS && !TARGET_C99_FUNCTIONS) | |
1347 | break; | |
1348 | ||
75a70cf9 | 1349 | arg = gimple_call_arg (stmt, 0); |
a0315874 | 1350 | if (TREE_CODE (arg) == SSA_NAME) |
4c80086d | 1351 | cfg_changed |= execute_cse_sincos_1 (arg); |
a0315874 | 1352 | break; |
1353 | ||
e78306af | 1354 | CASE_FLT_FN (BUILT_IN_POW): |
1355 | arg0 = gimple_call_arg (stmt, 0); | |
1356 | arg1 = gimple_call_arg (stmt, 1); | |
1357 | ||
1358 | loc = gimple_location (stmt); | |
1359 | result = gimple_expand_builtin_pow (&gsi, loc, arg0, arg1); | |
1360 | ||
1361 | if (result) | |
1362 | { | |
1363 | tree lhs = gimple_get_lhs (stmt); | |
1364 | gimple new_stmt = gimple_build_assign (lhs, result); | |
1365 | gimple_set_location (new_stmt, loc); | |
1366 | unlink_stmt_vdef (stmt); | |
1367 | gsi_replace (&gsi, new_stmt, true); | |
1368 | } | |
1369 | break; | |
1370 | ||
e9a6c4bc | 1371 | CASE_FLT_FN (BUILT_IN_POWI): |
1372 | arg0 = gimple_call_arg (stmt, 0); | |
1373 | arg1 = gimple_call_arg (stmt, 1); | |
1374 | if (!host_integerp (arg1, 0)) | |
1375 | break; | |
1376 | ||
1377 | n = TREE_INT_CST_LOW (arg1); | |
1378 | loc = gimple_location (stmt); | |
1379 | result = gimple_expand_builtin_powi (&gsi, loc, arg0, n); | |
1380 | ||
1381 | if (result) | |
1382 | { | |
1383 | tree lhs = gimple_get_lhs (stmt); | |
1384 | gimple new_stmt = gimple_build_assign (lhs, result); | |
1385 | gimple_set_location (new_stmt, loc); | |
1386 | unlink_stmt_vdef (stmt); | |
1387 | gsi_replace (&gsi, new_stmt, true); | |
1388 | } | |
1389 | break; | |
1390 | ||
a0315874 | 1391 | default:; |
1392 | } | |
1393 | } | |
1394 | } | |
1395 | } | |
1396 | ||
30c4e60d | 1397 | statistics_counter_event (cfun, "sincos statements inserted", |
1398 | sincos_stats.inserted); | |
1399 | ||
a0315874 | 1400 | free_dominance_info (CDI_DOMINATORS); |
4c80086d | 1401 | return cfg_changed ? TODO_cleanup_cfg : 0; |
a0315874 | 1402 | } |
1403 | ||
1404 | static bool | |
1405 | gate_cse_sincos (void) | |
1406 | { | |
e9a6c4bc | 1407 | /* We no longer require either sincos or cexp, since powi expansion |
1408 | piggybacks on this pass. */ | |
1409 | return optimize; | |
a0315874 | 1410 | } |
1411 | ||
20099e35 | 1412 | struct gimple_opt_pass pass_cse_sincos = |
a0315874 | 1413 | { |
20099e35 | 1414 | { |
1415 | GIMPLE_PASS, | |
a0315874 | 1416 | "sincos", /* name */ |
1417 | gate_cse_sincos, /* gate */ | |
1418 | execute_cse_sincos, /* execute */ | |
1419 | NULL, /* sub */ | |
1420 | NULL, /* next */ | |
1421 | 0, /* static_pass_number */ | |
0b1615c1 | 1422 | TV_NONE, /* tv_id */ |
a0315874 | 1423 | PROP_ssa, /* properties_required */ |
1424 | 0, /* properties_provided */ | |
1425 | 0, /* properties_destroyed */ | |
1426 | 0, /* todo_flags_start */ | |
1427 | TODO_dump_func | TODO_update_ssa | TODO_verify_ssa | |
20099e35 | 1428 | | TODO_verify_stmts /* todo_flags_finish */ |
1429 | } | |
a0315874 | 1430 | }; |
e174638f | 1431 | |
84cc784c | 1432 | /* A symbolic number is used to detect byte permutation and selection |
1433 | patterns. Therefore the field N contains an artificial number | |
1434 | consisting of byte size markers: | |
1435 | ||
1436 | 0 - byte has the value 0 | |
1437 | 1..size - byte contains the content of the byte | |
1438 | number indexed with that value minus one */ | |
1439 | ||
1440 | struct symbolic_number { | |
1441 | unsigned HOST_WIDEST_INT n; | |
1442 | int size; | |
1443 | }; | |
1444 | ||
1445 | /* Perform a SHIFT or ROTATE operation by COUNT bits on symbolic | |
1446 | number N. Return false if the requested operation is not permitted | |
1447 | on a symbolic number. */ | |
1448 | ||
1449 | static inline bool | |
1450 | do_shift_rotate (enum tree_code code, | |
1451 | struct symbolic_number *n, | |
1452 | int count) | |
1453 | { | |
1454 | if (count % 8 != 0) | |
1455 | return false; | |
1456 | ||
1457 | /* Zero out the extra bits of N in order to avoid them being shifted | |
1458 | into the significant bits. */ | |
1459 | if (n->size < (int)sizeof (HOST_WIDEST_INT)) | |
1460 | n->n &= ((unsigned HOST_WIDEST_INT)1 << (n->size * BITS_PER_UNIT)) - 1; | |
1461 | ||
1462 | switch (code) | |
1463 | { | |
1464 | case LSHIFT_EXPR: | |
1465 | n->n <<= count; | |
1466 | break; | |
1467 | case RSHIFT_EXPR: | |
1468 | n->n >>= count; | |
1469 | break; | |
1470 | case LROTATE_EXPR: | |
1471 | n->n = (n->n << count) | (n->n >> ((n->size * BITS_PER_UNIT) - count)); | |
1472 | break; | |
1473 | case RROTATE_EXPR: | |
1474 | n->n = (n->n >> count) | (n->n << ((n->size * BITS_PER_UNIT) - count)); | |
1475 | break; | |
1476 | default: | |
1477 | return false; | |
1478 | } | |
1479 | return true; | |
1480 | } | |
1481 | ||
1482 | /* Perform sanity checking for the symbolic number N and the gimple | |
1483 | statement STMT. */ | |
1484 | ||
1485 | static inline bool | |
1486 | verify_symbolic_number_p (struct symbolic_number *n, gimple stmt) | |
1487 | { | |
1488 | tree lhs_type; | |
1489 | ||
1490 | lhs_type = gimple_expr_type (stmt); | |
1491 | ||
1492 | if (TREE_CODE (lhs_type) != INTEGER_TYPE) | |
1493 | return false; | |
1494 | ||
1495 | if (TYPE_PRECISION (lhs_type) != n->size * BITS_PER_UNIT) | |
1496 | return false; | |
1497 | ||
1498 | return true; | |
1499 | } | |
1500 | ||
1501 | /* find_bswap_1 invokes itself recursively with N and tries to perform | |
1502 | the operation given by the rhs of STMT on the result. If the | |
1503 | operation could successfully be executed the function returns the | |
1504 | tree expression of the source operand and NULL otherwise. */ | |
1505 | ||
1506 | static tree | |
1507 | find_bswap_1 (gimple stmt, struct symbolic_number *n, int limit) | |
1508 | { | |
1509 | enum tree_code code; | |
1510 | tree rhs1, rhs2 = NULL; | |
1511 | gimple rhs1_stmt, rhs2_stmt; | |
1512 | tree source_expr1; | |
1513 | enum gimple_rhs_class rhs_class; | |
1514 | ||
1515 | if (!limit || !is_gimple_assign (stmt)) | |
1516 | return NULL_TREE; | |
1517 | ||
1518 | rhs1 = gimple_assign_rhs1 (stmt); | |
1519 | ||
1520 | if (TREE_CODE (rhs1) != SSA_NAME) | |
1521 | return NULL_TREE; | |
1522 | ||
1523 | code = gimple_assign_rhs_code (stmt); | |
1524 | rhs_class = gimple_assign_rhs_class (stmt); | |
1525 | rhs1_stmt = SSA_NAME_DEF_STMT (rhs1); | |
1526 | ||
1527 | if (rhs_class == GIMPLE_BINARY_RHS) | |
1528 | rhs2 = gimple_assign_rhs2 (stmt); | |
1529 | ||
1530 | /* Handle unary rhs and binary rhs with integer constants as second | |
1531 | operand. */ | |
1532 | ||
1533 | if (rhs_class == GIMPLE_UNARY_RHS | |
1534 | || (rhs_class == GIMPLE_BINARY_RHS | |
1535 | && TREE_CODE (rhs2) == INTEGER_CST)) | |
1536 | { | |
1537 | if (code != BIT_AND_EXPR | |
1538 | && code != LSHIFT_EXPR | |
1539 | && code != RSHIFT_EXPR | |
1540 | && code != LROTATE_EXPR | |
1541 | && code != RROTATE_EXPR | |
1542 | && code != NOP_EXPR | |
1543 | && code != CONVERT_EXPR) | |
1544 | return NULL_TREE; | |
1545 | ||
1546 | source_expr1 = find_bswap_1 (rhs1_stmt, n, limit - 1); | |
1547 | ||
1548 | /* If find_bswap_1 returned NULL STMT is a leaf node and we have | |
1549 | to initialize the symbolic number. */ | |
1550 | if (!source_expr1) | |
1551 | { | |
1552 | /* Set up the symbolic number N by setting each byte to a | |
1553 | value between 1 and the byte size of rhs1. The highest | |
f9a210c9 | 1554 | order byte is set to n->size and the lowest order |
1555 | byte to 1. */ | |
84cc784c | 1556 | n->size = TYPE_PRECISION (TREE_TYPE (rhs1)); |
1557 | if (n->size % BITS_PER_UNIT != 0) | |
1558 | return NULL_TREE; | |
1559 | n->size /= BITS_PER_UNIT; | |
1560 | n->n = (sizeof (HOST_WIDEST_INT) < 8 ? 0 : | |
f9a210c9 | 1561 | (unsigned HOST_WIDEST_INT)0x08070605 << 32 | 0x04030201); |
1562 | ||
1563 | if (n->size < (int)sizeof (HOST_WIDEST_INT)) | |
1564 | n->n &= ((unsigned HOST_WIDEST_INT)1 << | |
1565 | (n->size * BITS_PER_UNIT)) - 1; | |
84cc784c | 1566 | |
1567 | source_expr1 = rhs1; | |
1568 | } | |
1569 | ||
1570 | switch (code) | |
1571 | { | |
1572 | case BIT_AND_EXPR: | |
1573 | { | |
1574 | int i; | |
1575 | unsigned HOST_WIDEST_INT val = widest_int_cst_value (rhs2); | |
1576 | unsigned HOST_WIDEST_INT tmp = val; | |
1577 | ||
1578 | /* Only constants masking full bytes are allowed. */ | |
1579 | for (i = 0; i < n->size; i++, tmp >>= BITS_PER_UNIT) | |
1580 | if ((tmp & 0xff) != 0 && (tmp & 0xff) != 0xff) | |
1581 | return NULL_TREE; | |
1582 | ||
1583 | n->n &= val; | |
1584 | } | |
1585 | break; | |
1586 | case LSHIFT_EXPR: | |
1587 | case RSHIFT_EXPR: | |
1588 | case LROTATE_EXPR: | |
1589 | case RROTATE_EXPR: | |
1590 | if (!do_shift_rotate (code, n, (int)TREE_INT_CST_LOW (rhs2))) | |
1591 | return NULL_TREE; | |
1592 | break; | |
1593 | CASE_CONVERT: | |
1594 | { | |
1595 | int type_size; | |
1596 | ||
1597 | type_size = TYPE_PRECISION (gimple_expr_type (stmt)); | |
1598 | if (type_size % BITS_PER_UNIT != 0) | |
1599 | return NULL_TREE; | |
1600 | ||
84cc784c | 1601 | if (type_size / BITS_PER_UNIT < (int)(sizeof (HOST_WIDEST_INT))) |
1602 | { | |
1603 | /* If STMT casts to a smaller type mask out the bits not | |
1604 | belonging to the target type. */ | |
84cc784c | 1605 | n->n &= ((unsigned HOST_WIDEST_INT)1 << type_size) - 1; |
1606 | } | |
f9a210c9 | 1607 | n->size = type_size / BITS_PER_UNIT; |
84cc784c | 1608 | } |
1609 | break; | |
1610 | default: | |
1611 | return NULL_TREE; | |
1612 | }; | |
1613 | return verify_symbolic_number_p (n, stmt) ? source_expr1 : NULL; | |
1614 | } | |
1615 | ||
1616 | /* Handle binary rhs. */ | |
1617 | ||
1618 | if (rhs_class == GIMPLE_BINARY_RHS) | |
1619 | { | |
1620 | struct symbolic_number n1, n2; | |
1621 | tree source_expr2; | |
1622 | ||
1623 | if (code != BIT_IOR_EXPR) | |
1624 | return NULL_TREE; | |
1625 | ||
1626 | if (TREE_CODE (rhs2) != SSA_NAME) | |
1627 | return NULL_TREE; | |
1628 | ||
1629 | rhs2_stmt = SSA_NAME_DEF_STMT (rhs2); | |
1630 | ||
1631 | switch (code) | |
1632 | { | |
1633 | case BIT_IOR_EXPR: | |
1634 | source_expr1 = find_bswap_1 (rhs1_stmt, &n1, limit - 1); | |
1635 | ||
1636 | if (!source_expr1) | |
1637 | return NULL_TREE; | |
1638 | ||
1639 | source_expr2 = find_bswap_1 (rhs2_stmt, &n2, limit - 1); | |
1640 | ||
1641 | if (source_expr1 != source_expr2 | |
1642 | || n1.size != n2.size) | |
1643 | return NULL_TREE; | |
1644 | ||
1645 | n->size = n1.size; | |
1646 | n->n = n1.n | n2.n; | |
1647 | ||
1648 | if (!verify_symbolic_number_p (n, stmt)) | |
1649 | return NULL_TREE; | |
1650 | ||
1651 | break; | |
1652 | default: | |
1653 | return NULL_TREE; | |
1654 | } | |
1655 | return source_expr1; | |
1656 | } | |
1657 | return NULL_TREE; | |
1658 | } | |
1659 | ||
1660 | /* Check if STMT completes a bswap implementation consisting of ORs, | |
1661 | SHIFTs and ANDs. Return the source tree expression on which the | |
1662 | byte swap is performed and NULL if no bswap was found. */ | |
1663 | ||
1664 | static tree | |
1665 | find_bswap (gimple stmt) | |
1666 | { | |
1667 | /* The number which the find_bswap result should match in order to | |
f9a210c9 | 1668 | have a full byte swap. The number is shifted to the left according |
1669 | to the size of the symbolic number before using it. */ | |
84cc784c | 1670 | unsigned HOST_WIDEST_INT cmp = |
1671 | sizeof (HOST_WIDEST_INT) < 8 ? 0 : | |
f9a210c9 | 1672 | (unsigned HOST_WIDEST_INT)0x01020304 << 32 | 0x05060708; |
84cc784c | 1673 | |
1674 | struct symbolic_number n; | |
1675 | tree source_expr; | |
1676 | ||
9bc1852a | 1677 | /* The last parameter determines the depth search limit. It usually |
1678 | correlates directly to the number of bytes to be touched. We | |
1679 | increase that number by one here in order to also cover signed -> | |
1680 | unsigned conversions of the src operand as can be seen in | |
1681 | libgcc. */ | |
84cc784c | 1682 | source_expr = find_bswap_1 (stmt, &n, |
1683 | TREE_INT_CST_LOW ( | |
9bc1852a | 1684 | TYPE_SIZE_UNIT (gimple_expr_type (stmt))) + 1); |
84cc784c | 1685 | |
1686 | if (!source_expr) | |
1687 | return NULL_TREE; | |
1688 | ||
1689 | /* Zero out the extra bits of N and CMP. */ | |
1690 | if (n.size < (int)sizeof (HOST_WIDEST_INT)) | |
1691 | { | |
1692 | unsigned HOST_WIDEST_INT mask = | |
1693 | ((unsigned HOST_WIDEST_INT)1 << (n.size * BITS_PER_UNIT)) - 1; | |
1694 | ||
1695 | n.n &= mask; | |
f9a210c9 | 1696 | cmp >>= (sizeof (HOST_WIDEST_INT) - n.size) * BITS_PER_UNIT; |
84cc784c | 1697 | } |
1698 | ||
1699 | /* A complete byte swap should make the symbolic number to start | |
1700 | with the largest digit in the highest order byte. */ | |
1701 | if (cmp != n.n) | |
1702 | return NULL_TREE; | |
1703 | ||
1704 | return source_expr; | |
1705 | } | |
1706 | ||
1707 | /* Find manual byte swap implementations and turn them into a bswap | |
1708 | builtin invokation. */ | |
1709 | ||
1710 | static unsigned int | |
1711 | execute_optimize_bswap (void) | |
1712 | { | |
1713 | basic_block bb; | |
1714 | bool bswap32_p, bswap64_p; | |
1715 | bool changed = false; | |
0af25806 | 1716 | tree bswap32_type = NULL_TREE, bswap64_type = NULL_TREE; |
84cc784c | 1717 | |
1718 | if (BITS_PER_UNIT != 8) | |
1719 | return 0; | |
1720 | ||
1721 | if (sizeof (HOST_WIDEST_INT) < 8) | |
1722 | return 0; | |
1723 | ||
1724 | bswap32_p = (built_in_decls[BUILT_IN_BSWAP32] | |
d6bf3b14 | 1725 | && optab_handler (bswap_optab, SImode) != CODE_FOR_nothing); |
84cc784c | 1726 | bswap64_p = (built_in_decls[BUILT_IN_BSWAP64] |
d6bf3b14 | 1727 | && (optab_handler (bswap_optab, DImode) != CODE_FOR_nothing |
3328b1fb | 1728 | || (bswap32_p && word_mode == SImode))); |
84cc784c | 1729 | |
1730 | if (!bswap32_p && !bswap64_p) | |
1731 | return 0; | |
1732 | ||
0af25806 | 1733 | /* Determine the argument type of the builtins. The code later on |
1734 | assumes that the return and argument type are the same. */ | |
1735 | if (bswap32_p) | |
1736 | { | |
1737 | tree fndecl = built_in_decls[BUILT_IN_BSWAP32]; | |
1738 | bswap32_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl))); | |
1739 | } | |
1740 | ||
1741 | if (bswap64_p) | |
1742 | { | |
1743 | tree fndecl = built_in_decls[BUILT_IN_BSWAP64]; | |
1744 | bswap64_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl))); | |
1745 | } | |
1746 | ||
30c4e60d | 1747 | memset (&bswap_stats, 0, sizeof (bswap_stats)); |
1748 | ||
84cc784c | 1749 | FOR_EACH_BB (bb) |
1750 | { | |
1751 | gimple_stmt_iterator gsi; | |
1752 | ||
1753 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1754 | { | |
1755 | gimple stmt = gsi_stmt (gsi); | |
0af25806 | 1756 | tree bswap_src, bswap_type; |
1757 | tree bswap_tmp; | |
84cc784c | 1758 | tree fndecl = NULL_TREE; |
1759 | int type_size; | |
1760 | gimple call; | |
1761 | ||
1762 | if (!is_gimple_assign (stmt) | |
1763 | || gimple_assign_rhs_code (stmt) != BIT_IOR_EXPR) | |
1764 | continue; | |
1765 | ||
1766 | type_size = TYPE_PRECISION (gimple_expr_type (stmt)); | |
1767 | ||
1768 | switch (type_size) | |
1769 | { | |
1770 | case 32: | |
1771 | if (bswap32_p) | |
0af25806 | 1772 | { |
1773 | fndecl = built_in_decls[BUILT_IN_BSWAP32]; | |
1774 | bswap_type = bswap32_type; | |
1775 | } | |
84cc784c | 1776 | break; |
1777 | case 64: | |
1778 | if (bswap64_p) | |
0af25806 | 1779 | { |
1780 | fndecl = built_in_decls[BUILT_IN_BSWAP64]; | |
1781 | bswap_type = bswap64_type; | |
1782 | } | |
84cc784c | 1783 | break; |
1784 | default: | |
1785 | continue; | |
1786 | } | |
1787 | ||
1788 | if (!fndecl) | |
1789 | continue; | |
1790 | ||
1791 | bswap_src = find_bswap (stmt); | |
1792 | ||
1793 | if (!bswap_src) | |
1794 | continue; | |
1795 | ||
1796 | changed = true; | |
30c4e60d | 1797 | if (type_size == 32) |
1798 | bswap_stats.found_32bit++; | |
1799 | else | |
1800 | bswap_stats.found_64bit++; | |
0af25806 | 1801 | |
1802 | bswap_tmp = bswap_src; | |
1803 | ||
1804 | /* Convert the src expression if necessary. */ | |
1805 | if (!useless_type_conversion_p (TREE_TYPE (bswap_tmp), bswap_type)) | |
1806 | { | |
1807 | gimple convert_stmt; | |
1808 | ||
1809 | bswap_tmp = create_tmp_var (bswap_type, "bswapsrc"); | |
1810 | add_referenced_var (bswap_tmp); | |
1811 | bswap_tmp = make_ssa_name (bswap_tmp, NULL); | |
1812 | ||
1813 | convert_stmt = gimple_build_assign_with_ops ( | |
1814 | CONVERT_EXPR, bswap_tmp, bswap_src, NULL); | |
1815 | gsi_insert_before (&gsi, convert_stmt, GSI_SAME_STMT); | |
1816 | } | |
1817 | ||
1818 | call = gimple_build_call (fndecl, 1, bswap_tmp); | |
1819 | ||
1820 | bswap_tmp = gimple_assign_lhs (stmt); | |
1821 | ||
1822 | /* Convert the result if necessary. */ | |
1823 | if (!useless_type_conversion_p (TREE_TYPE (bswap_tmp), bswap_type)) | |
1824 | { | |
1825 | gimple convert_stmt; | |
1826 | ||
1827 | bswap_tmp = create_tmp_var (bswap_type, "bswapdst"); | |
1828 | add_referenced_var (bswap_tmp); | |
1829 | bswap_tmp = make_ssa_name (bswap_tmp, NULL); | |
1830 | convert_stmt = gimple_build_assign_with_ops ( | |
1831 | CONVERT_EXPR, gimple_assign_lhs (stmt), bswap_tmp, NULL); | |
1832 | gsi_insert_after (&gsi, convert_stmt, GSI_SAME_STMT); | |
1833 | } | |
1834 | ||
1835 | gimple_call_set_lhs (call, bswap_tmp); | |
84cc784c | 1836 | |
1837 | if (dump_file) | |
1838 | { | |
1839 | fprintf (dump_file, "%d bit bswap implementation found at: ", | |
1840 | (int)type_size); | |
1841 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
1842 | } | |
1843 | ||
1844 | gsi_insert_after (&gsi, call, GSI_SAME_STMT); | |
1845 | gsi_remove (&gsi, true); | |
1846 | } | |
1847 | } | |
1848 | ||
30c4e60d | 1849 | statistics_counter_event (cfun, "32-bit bswap implementations found", |
1850 | bswap_stats.found_32bit); | |
1851 | statistics_counter_event (cfun, "64-bit bswap implementations found", | |
1852 | bswap_stats.found_64bit); | |
1853 | ||
84cc784c | 1854 | return (changed ? TODO_dump_func | TODO_update_ssa | TODO_verify_ssa |
1855 | | TODO_verify_stmts : 0); | |
1856 | } | |
1857 | ||
1858 | static bool | |
1859 | gate_optimize_bswap (void) | |
1860 | { | |
1861 | return flag_expensive_optimizations && optimize; | |
1862 | } | |
1863 | ||
1864 | struct gimple_opt_pass pass_optimize_bswap = | |
1865 | { | |
1866 | { | |
1867 | GIMPLE_PASS, | |
1868 | "bswap", /* name */ | |
1869 | gate_optimize_bswap, /* gate */ | |
1870 | execute_optimize_bswap, /* execute */ | |
1871 | NULL, /* sub */ | |
1872 | NULL, /* next */ | |
1873 | 0, /* static_pass_number */ | |
1874 | TV_NONE, /* tv_id */ | |
1875 | PROP_ssa, /* properties_required */ | |
1876 | 0, /* properties_provided */ | |
1877 | 0, /* properties_destroyed */ | |
1878 | 0, /* todo_flags_start */ | |
1879 | 0 /* todo_flags_finish */ | |
1880 | } | |
1881 | }; | |
62be004c | 1882 | |
7e4c867e | 1883 | /* Return true if RHS is a suitable operand for a widening multiplication. |
1884 | There are two cases: | |
1885 | ||
1886 | - RHS makes some value twice as wide. Store that value in *NEW_RHS_OUT | |
1887 | if so, and store its type in *TYPE_OUT. | |
1888 | ||
1889 | - RHS is an integer constant. Store that value in *NEW_RHS_OUT if so, | |
1890 | but leave *TYPE_OUT untouched. */ | |
00f4f705 | 1891 | |
1892 | static bool | |
7e4c867e | 1893 | is_widening_mult_rhs_p (tree rhs, tree *type_out, tree *new_rhs_out) |
1894 | { | |
1895 | gimple stmt; | |
1896 | tree type, type1, rhs1; | |
1897 | enum tree_code rhs_code; | |
1898 | ||
1899 | if (TREE_CODE (rhs) == SSA_NAME) | |
1900 | { | |
1901 | type = TREE_TYPE (rhs); | |
1902 | stmt = SSA_NAME_DEF_STMT (rhs); | |
1903 | if (!is_gimple_assign (stmt)) | |
1904 | return false; | |
1905 | ||
1906 | rhs_code = gimple_assign_rhs_code (stmt); | |
1907 | if (TREE_CODE (type) == INTEGER_TYPE | |
1908 | ? !CONVERT_EXPR_CODE_P (rhs_code) | |
1909 | : rhs_code != FIXED_CONVERT_EXPR) | |
1910 | return false; | |
1911 | ||
1912 | rhs1 = gimple_assign_rhs1 (stmt); | |
1913 | type1 = TREE_TYPE (rhs1); | |
1914 | if (TREE_CODE (type1) != TREE_CODE (type) | |
1915 | || TYPE_PRECISION (type1) * 2 != TYPE_PRECISION (type)) | |
1916 | return false; | |
1917 | ||
1918 | *new_rhs_out = rhs1; | |
1919 | *type_out = type1; | |
1920 | return true; | |
1921 | } | |
1922 | ||
1923 | if (TREE_CODE (rhs) == INTEGER_CST) | |
1924 | { | |
1925 | *new_rhs_out = rhs; | |
1926 | *type_out = NULL; | |
1927 | return true; | |
1928 | } | |
1929 | ||
1930 | return false; | |
1931 | } | |
1932 | ||
1933 | /* Return true if STMT performs a widening multiplication. If so, | |
1934 | store the unwidened types of the operands in *TYPE1_OUT and *TYPE2_OUT | |
1935 | respectively. Also fill *RHS1_OUT and *RHS2_OUT such that converting | |
1936 | those operands to types *TYPE1_OUT and *TYPE2_OUT would give the | |
1937 | operands of the multiplication. */ | |
1938 | ||
1939 | static bool | |
1940 | is_widening_mult_p (gimple stmt, | |
1941 | tree *type1_out, tree *rhs1_out, | |
1942 | tree *type2_out, tree *rhs2_out) | |
00f4f705 | 1943 | { |
00f4f705 | 1944 | tree type; |
1945 | ||
1946 | type = TREE_TYPE (gimple_assign_lhs (stmt)); | |
7e4c867e | 1947 | if (TREE_CODE (type) != INTEGER_TYPE |
1948 | && TREE_CODE (type) != FIXED_POINT_TYPE) | |
1949 | return false; | |
00f4f705 | 1950 | |
7e4c867e | 1951 | if (!is_widening_mult_rhs_p (gimple_assign_rhs1 (stmt), type1_out, rhs1_out)) |
00f4f705 | 1952 | return false; |
1953 | ||
7e4c867e | 1954 | if (!is_widening_mult_rhs_p (gimple_assign_rhs2 (stmt), type2_out, rhs2_out)) |
1955 | return false; | |
00f4f705 | 1956 | |
7e4c867e | 1957 | if (*type1_out == NULL) |
00f4f705 | 1958 | { |
7e4c867e | 1959 | if (*type2_out == NULL || !int_fits_type_p (*rhs1_out, *type2_out)) |
00f4f705 | 1960 | return false; |
7e4c867e | 1961 | *type1_out = *type2_out; |
00f4f705 | 1962 | } |
00f4f705 | 1963 | |
7e4c867e | 1964 | if (*type2_out == NULL) |
00f4f705 | 1965 | { |
7e4c867e | 1966 | if (!int_fits_type_p (*rhs2_out, *type1_out)) |
00f4f705 | 1967 | return false; |
7e4c867e | 1968 | *type2_out = *type1_out; |
00f4f705 | 1969 | } |
00f4f705 | 1970 | |
7e4c867e | 1971 | return true; |
1972 | } | |
00f4f705 | 1973 | |
7e4c867e | 1974 | /* Process a single gimple statement STMT, which has a MULT_EXPR as |
1975 | its rhs, and try to convert it into a WIDEN_MULT_EXPR. The return | |
1976 | value is true iff we converted the statement. */ | |
1977 | ||
1978 | static bool | |
1979 | convert_mult_to_widen (gimple stmt) | |
1980 | { | |
1981 | tree lhs, rhs1, rhs2, type, type1, type2; | |
1982 | enum insn_code handler; | |
1983 | ||
1984 | lhs = gimple_assign_lhs (stmt); | |
1985 | type = TREE_TYPE (lhs); | |
1986 | if (TREE_CODE (type) != INTEGER_TYPE) | |
00f4f705 | 1987 | return false; |
1988 | ||
7e4c867e | 1989 | if (!is_widening_mult_p (stmt, &type1, &rhs1, &type2, &rhs2)) |
00f4f705 | 1990 | return false; |
1991 | ||
7e4c867e | 1992 | if (TYPE_UNSIGNED (type1) && TYPE_UNSIGNED (type2)) |
1993 | handler = optab_handler (umul_widen_optab, TYPE_MODE (type)); | |
1994 | else if (!TYPE_UNSIGNED (type1) && !TYPE_UNSIGNED (type2)) | |
1995 | handler = optab_handler (smul_widen_optab, TYPE_MODE (type)); | |
00f4f705 | 1996 | else |
7e4c867e | 1997 | handler = optab_handler (usmul_widen_optab, TYPE_MODE (type)); |
1998 | ||
1999 | if (handler == CODE_FOR_nothing) | |
2000 | return false; | |
2001 | ||
2002 | gimple_assign_set_rhs1 (stmt, fold_convert (type1, rhs1)); | |
2003 | gimple_assign_set_rhs2 (stmt, fold_convert (type2, rhs2)); | |
00f4f705 | 2004 | gimple_assign_set_rhs_code (stmt, WIDEN_MULT_EXPR); |
2005 | update_stmt (stmt); | |
30c4e60d | 2006 | widen_mul_stats.widen_mults_inserted++; |
00f4f705 | 2007 | return true; |
2008 | } | |
2009 | ||
2010 | /* Process a single gimple statement STMT, which is found at the | |
2011 | iterator GSI and has a either a PLUS_EXPR or a MINUS_EXPR as its | |
2012 | rhs (given by CODE), and try to convert it into a | |
2013 | WIDEN_MULT_PLUS_EXPR or a WIDEN_MULT_MINUS_EXPR. The return value | |
2014 | is true iff we converted the statement. */ | |
2015 | ||
2016 | static bool | |
2017 | convert_plusminus_to_widen (gimple_stmt_iterator *gsi, gimple stmt, | |
2018 | enum tree_code code) | |
2019 | { | |
2020 | gimple rhs1_stmt = NULL, rhs2_stmt = NULL; | |
7e4c867e | 2021 | tree type, type1, type2; |
00f4f705 | 2022 | tree lhs, rhs1, rhs2, mult_rhs1, mult_rhs2, add_rhs; |
2023 | enum tree_code rhs1_code = ERROR_MARK, rhs2_code = ERROR_MARK; | |
2024 | optab this_optab; | |
2025 | enum tree_code wmult_code; | |
2026 | ||
2027 | lhs = gimple_assign_lhs (stmt); | |
2028 | type = TREE_TYPE (lhs); | |
7e4c867e | 2029 | if (TREE_CODE (type) != INTEGER_TYPE |
2030 | && TREE_CODE (type) != FIXED_POINT_TYPE) | |
00f4f705 | 2031 | return false; |
2032 | ||
2033 | if (code == MINUS_EXPR) | |
2034 | wmult_code = WIDEN_MULT_MINUS_EXPR; | |
2035 | else | |
2036 | wmult_code = WIDEN_MULT_PLUS_EXPR; | |
2037 | ||
00f4f705 | 2038 | rhs1 = gimple_assign_rhs1 (stmt); |
2039 | rhs2 = gimple_assign_rhs2 (stmt); | |
2040 | ||
2041 | if (TREE_CODE (rhs1) == SSA_NAME) | |
2042 | { | |
2043 | rhs1_stmt = SSA_NAME_DEF_STMT (rhs1); | |
2044 | if (is_gimple_assign (rhs1_stmt)) | |
2045 | rhs1_code = gimple_assign_rhs_code (rhs1_stmt); | |
2046 | } | |
2047 | else | |
2048 | return false; | |
2049 | ||
2050 | if (TREE_CODE (rhs2) == SSA_NAME) | |
2051 | { | |
2052 | rhs2_stmt = SSA_NAME_DEF_STMT (rhs2); | |
2053 | if (is_gimple_assign (rhs2_stmt)) | |
2054 | rhs2_code = gimple_assign_rhs_code (rhs2_stmt); | |
2055 | } | |
2056 | else | |
2057 | return false; | |
2058 | ||
7e4c867e | 2059 | if (code == PLUS_EXPR && rhs1_code == MULT_EXPR) |
00f4f705 | 2060 | { |
7e4c867e | 2061 | if (!is_widening_mult_p (rhs1_stmt, &type1, &mult_rhs1, |
2062 | &type2, &mult_rhs2)) | |
00f4f705 | 2063 | return false; |
7e4c867e | 2064 | add_rhs = rhs2; |
00f4f705 | 2065 | } |
7e4c867e | 2066 | else if (rhs2_code == MULT_EXPR) |
00f4f705 | 2067 | { |
815a0224 | 2068 | if (!is_widening_mult_p (rhs2_stmt, &type1, &mult_rhs1, |
7e4c867e | 2069 | &type2, &mult_rhs2)) |
00f4f705 | 2070 | return false; |
7e4c867e | 2071 | add_rhs = rhs1; |
00f4f705 | 2072 | } |
7e4c867e | 2073 | else if (code == PLUS_EXPR && rhs1_code == WIDEN_MULT_EXPR) |
00f4f705 | 2074 | { |
2075 | mult_rhs1 = gimple_assign_rhs1 (rhs1_stmt); | |
2076 | mult_rhs2 = gimple_assign_rhs2 (rhs1_stmt); | |
815a0224 | 2077 | type1 = TREE_TYPE (mult_rhs1); |
2078 | type2 = TREE_TYPE (mult_rhs2); | |
00f4f705 | 2079 | add_rhs = rhs2; |
2080 | } | |
2081 | else if (rhs2_code == WIDEN_MULT_EXPR) | |
2082 | { | |
2083 | mult_rhs1 = gimple_assign_rhs1 (rhs2_stmt); | |
2084 | mult_rhs2 = gimple_assign_rhs2 (rhs2_stmt); | |
815a0224 | 2085 | type1 = TREE_TYPE (mult_rhs1); |
2086 | type2 = TREE_TYPE (mult_rhs2); | |
00f4f705 | 2087 | add_rhs = rhs1; |
2088 | } | |
2089 | else | |
2090 | return false; | |
2091 | ||
815a0224 | 2092 | if (TYPE_UNSIGNED (type1) != TYPE_UNSIGNED (type2)) |
2093 | return false; | |
2094 | ||
2095 | /* Verify that the machine can perform a widening multiply | |
2096 | accumulate in this mode/signedness combination, otherwise | |
2097 | this transformation is likely to pessimize code. */ | |
2098 | this_optab = optab_for_tree_code (wmult_code, type1, optab_default); | |
2099 | if (optab_handler (this_optab, TYPE_MODE (type)) == CODE_FOR_nothing) | |
2100 | return false; | |
2101 | ||
00f4f705 | 2102 | /* ??? May need some type verification here? */ |
2103 | ||
815a0224 | 2104 | gimple_assign_set_rhs_with_ops_1 (gsi, wmult_code, |
2105 | fold_convert (type1, mult_rhs1), | |
2106 | fold_convert (type2, mult_rhs2), | |
00f4f705 | 2107 | add_rhs); |
2108 | update_stmt (gsi_stmt (*gsi)); | |
30c4e60d | 2109 | widen_mul_stats.maccs_inserted++; |
00f4f705 | 2110 | return true; |
2111 | } | |
2112 | ||
15dbdc8f | 2113 | /* Combine the multiplication at MUL_STMT with operands MULOP1 and MULOP2 |
2114 | with uses in additions and subtractions to form fused multiply-add | |
2115 | operations. Returns true if successful and MUL_STMT should be removed. */ | |
b9be572e | 2116 | |
2117 | static bool | |
15dbdc8f | 2118 | convert_mult_to_fma (gimple mul_stmt, tree op1, tree op2) |
b9be572e | 2119 | { |
15dbdc8f | 2120 | tree mul_result = gimple_get_lhs (mul_stmt); |
b9be572e | 2121 | tree type = TREE_TYPE (mul_result); |
44579526 | 2122 | gimple use_stmt, neguse_stmt, fma_stmt; |
b9be572e | 2123 | use_operand_p use_p; |
2124 | imm_use_iterator imm_iter; | |
2125 | ||
2126 | if (FLOAT_TYPE_P (type) | |
2127 | && flag_fp_contract_mode == FP_CONTRACT_OFF) | |
2128 | return false; | |
2129 | ||
2130 | /* We don't want to do bitfield reduction ops. */ | |
2131 | if (INTEGRAL_TYPE_P (type) | |
2132 | && (TYPE_PRECISION (type) | |
2133 | != GET_MODE_PRECISION (TYPE_MODE (type)))) | |
2134 | return false; | |
2135 | ||
2136 | /* If the target doesn't support it, don't generate it. We assume that | |
2137 | if fma isn't available then fms, fnma or fnms are not either. */ | |
2138 | if (optab_handler (fma_optab, TYPE_MODE (type)) == CODE_FOR_nothing) | |
2139 | return false; | |
2140 | ||
2141 | /* Make sure that the multiplication statement becomes dead after | |
2142 | the transformation, thus that all uses are transformed to FMAs. | |
2143 | This means we assume that an FMA operation has the same cost | |
2144 | as an addition. */ | |
2145 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, mul_result) | |
2146 | { | |
2147 | enum tree_code use_code; | |
44579526 | 2148 | tree result = mul_result; |
2149 | bool negate_p = false; | |
b9be572e | 2150 | |
2151 | use_stmt = USE_STMT (use_p); | |
2152 | ||
17a2c727 | 2153 | if (is_gimple_debug (use_stmt)) |
2154 | continue; | |
2155 | ||
b9be572e | 2156 | /* For now restrict this operations to single basic blocks. In theory |
2157 | we would want to support sinking the multiplication in | |
2158 | m = a*b; | |
2159 | if () | |
2160 | ma = m + c; | |
2161 | else | |
2162 | d = m; | |
2163 | to form a fma in the then block and sink the multiplication to the | |
2164 | else block. */ | |
2165 | if (gimple_bb (use_stmt) != gimple_bb (mul_stmt)) | |
2166 | return false; | |
2167 | ||
44579526 | 2168 | if (!is_gimple_assign (use_stmt)) |
b9be572e | 2169 | return false; |
2170 | ||
44579526 | 2171 | use_code = gimple_assign_rhs_code (use_stmt); |
2172 | ||
2173 | /* A negate on the multiplication leads to FNMA. */ | |
2174 | if (use_code == NEGATE_EXPR) | |
2175 | { | |
805ad414 | 2176 | ssa_op_iter iter; |
5715c09b | 2177 | use_operand_p usep; |
805ad414 | 2178 | |
44579526 | 2179 | result = gimple_assign_lhs (use_stmt); |
2180 | ||
2181 | /* Make sure the negate statement becomes dead with this | |
2182 | single transformation. */ | |
2183 | if (!single_imm_use (gimple_assign_lhs (use_stmt), | |
2184 | &use_p, &neguse_stmt)) | |
2185 | return false; | |
2186 | ||
805ad414 | 2187 | /* Make sure the multiplication isn't also used on that stmt. */ |
5715c09b | 2188 | FOR_EACH_PHI_OR_STMT_USE (usep, neguse_stmt, iter, SSA_OP_USE) |
2189 | if (USE_FROM_PTR (usep) == mul_result) | |
805ad414 | 2190 | return false; |
2191 | ||
44579526 | 2192 | /* Re-validate. */ |
2193 | use_stmt = neguse_stmt; | |
2194 | if (gimple_bb (use_stmt) != gimple_bb (mul_stmt)) | |
2195 | return false; | |
2196 | if (!is_gimple_assign (use_stmt)) | |
2197 | return false; | |
2198 | ||
2199 | use_code = gimple_assign_rhs_code (use_stmt); | |
2200 | negate_p = true; | |
2201 | } | |
b9be572e | 2202 | |
44579526 | 2203 | switch (use_code) |
2204 | { | |
2205 | case MINUS_EXPR: | |
8a9d0572 | 2206 | if (gimple_assign_rhs2 (use_stmt) == result) |
2207 | negate_p = !negate_p; | |
2208 | break; | |
44579526 | 2209 | case PLUS_EXPR: |
44579526 | 2210 | break; |
44579526 | 2211 | default: |
2212 | /* FMA can only be formed from PLUS and MINUS. */ | |
2213 | return false; | |
2214 | } | |
b9be572e | 2215 | |
44579526 | 2216 | /* We can't handle a * b + a * b. */ |
2217 | if (gimple_assign_rhs1 (use_stmt) == gimple_assign_rhs2 (use_stmt)) | |
2218 | return false; | |
8a9d0572 | 2219 | |
2220 | /* While it is possible to validate whether or not the exact form | |
2221 | that we've recognized is available in the backend, the assumption | |
2222 | is that the transformation is never a loss. For instance, suppose | |
2223 | the target only has the plain FMA pattern available. Consider | |
2224 | a*b-c -> fma(a,b,-c): we've exchanged MUL+SUB for FMA+NEG, which | |
2225 | is still two operations. Consider -(a*b)-c -> fma(-a,b,-c): we | |
2226 | still have 3 operations, but in the FMA form the two NEGs are | |
2227 | independant and could be run in parallel. */ | |
b9be572e | 2228 | } |
2229 | ||
2230 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, mul_result) | |
2231 | { | |
b9be572e | 2232 | gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt); |
17a2c727 | 2233 | enum tree_code use_code; |
15dbdc8f | 2234 | tree addop, mulop1 = op1, result = mul_result; |
44579526 | 2235 | bool negate_p = false; |
b9be572e | 2236 | |
17a2c727 | 2237 | if (is_gimple_debug (use_stmt)) |
2238 | continue; | |
2239 | ||
2240 | use_code = gimple_assign_rhs_code (use_stmt); | |
44579526 | 2241 | if (use_code == NEGATE_EXPR) |
2242 | { | |
2243 | result = gimple_assign_lhs (use_stmt); | |
2244 | single_imm_use (gimple_assign_lhs (use_stmt), &use_p, &neguse_stmt); | |
2245 | gsi_remove (&gsi, true); | |
2246 | release_defs (use_stmt); | |
2247 | ||
2248 | use_stmt = neguse_stmt; | |
2249 | gsi = gsi_for_stmt (use_stmt); | |
2250 | use_code = gimple_assign_rhs_code (use_stmt); | |
2251 | negate_p = true; | |
2252 | } | |
2253 | ||
2254 | if (gimple_assign_rhs1 (use_stmt) == result) | |
b9be572e | 2255 | { |
2256 | addop = gimple_assign_rhs2 (use_stmt); | |
2257 | /* a * b - c -> a * b + (-c) */ | |
2258 | if (gimple_assign_rhs_code (use_stmt) == MINUS_EXPR) | |
2259 | addop = force_gimple_operand_gsi (&gsi, | |
2260 | build1 (NEGATE_EXPR, | |
2261 | type, addop), | |
2262 | true, NULL_TREE, true, | |
2263 | GSI_SAME_STMT); | |
2264 | } | |
2265 | else | |
2266 | { | |
2267 | addop = gimple_assign_rhs1 (use_stmt); | |
2268 | /* a - b * c -> (-b) * c + a */ | |
2269 | if (gimple_assign_rhs_code (use_stmt) == MINUS_EXPR) | |
44579526 | 2270 | negate_p = !negate_p; |
b9be572e | 2271 | } |
2272 | ||
44579526 | 2273 | if (negate_p) |
2274 | mulop1 = force_gimple_operand_gsi (&gsi, | |
2275 | build1 (NEGATE_EXPR, | |
2276 | type, mulop1), | |
2277 | true, NULL_TREE, true, | |
2278 | GSI_SAME_STMT); | |
2279 | ||
b9be572e | 2280 | fma_stmt = gimple_build_assign_with_ops3 (FMA_EXPR, |
2281 | gimple_assign_lhs (use_stmt), | |
15dbdc8f | 2282 | mulop1, op2, |
b9be572e | 2283 | addop); |
2284 | gsi_replace (&gsi, fma_stmt, true); | |
30c4e60d | 2285 | widen_mul_stats.fmas_inserted++; |
b9be572e | 2286 | } |
2287 | ||
2288 | return true; | |
2289 | } | |
2290 | ||
62be004c | 2291 | /* Find integer multiplications where the operands are extended from |
2292 | smaller types, and replace the MULT_EXPR with a WIDEN_MULT_EXPR | |
2293 | where appropriate. */ | |
2294 | ||
2295 | static unsigned int | |
2296 | execute_optimize_widening_mul (void) | |
2297 | { | |
62be004c | 2298 | basic_block bb; |
15dbdc8f | 2299 | bool cfg_changed = false; |
62be004c | 2300 | |
30c4e60d | 2301 | memset (&widen_mul_stats, 0, sizeof (widen_mul_stats)); |
2302 | ||
62be004c | 2303 | FOR_EACH_BB (bb) |
2304 | { | |
2305 | gimple_stmt_iterator gsi; | |
2306 | ||
b9be572e | 2307 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi);) |
62be004c | 2308 | { |
2309 | gimple stmt = gsi_stmt (gsi); | |
00f4f705 | 2310 | enum tree_code code; |
62be004c | 2311 | |
b9be572e | 2312 | if (is_gimple_assign (stmt)) |
2313 | { | |
2314 | code = gimple_assign_rhs_code (stmt); | |
2315 | switch (code) | |
2316 | { | |
2317 | case MULT_EXPR: | |
2318 | if (!convert_mult_to_widen (stmt) | |
15dbdc8f | 2319 | && convert_mult_to_fma (stmt, |
2320 | gimple_assign_rhs1 (stmt), | |
2321 | gimple_assign_rhs2 (stmt))) | |
b9be572e | 2322 | { |
2323 | gsi_remove (&gsi, true); | |
2324 | release_defs (stmt); | |
2325 | continue; | |
2326 | } | |
2327 | break; | |
2328 | ||
2329 | case PLUS_EXPR: | |
2330 | case MINUS_EXPR: | |
2331 | convert_plusminus_to_widen (&gsi, stmt, code); | |
2332 | break; | |
62be004c | 2333 | |
b9be572e | 2334 | default:; |
2335 | } | |
2336 | } | |
d4af184a | 2337 | else if (is_gimple_call (stmt) |
2338 | && gimple_call_lhs (stmt)) | |
15dbdc8f | 2339 | { |
2340 | tree fndecl = gimple_call_fndecl (stmt); | |
2341 | if (fndecl | |
2342 | && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) | |
2343 | { | |
2344 | switch (DECL_FUNCTION_CODE (fndecl)) | |
2345 | { | |
2346 | case BUILT_IN_POWF: | |
2347 | case BUILT_IN_POW: | |
2348 | case BUILT_IN_POWL: | |
2349 | if (TREE_CODE (gimple_call_arg (stmt, 1)) == REAL_CST | |
2350 | && REAL_VALUES_EQUAL | |
2351 | (TREE_REAL_CST (gimple_call_arg (stmt, 1)), | |
2352 | dconst2) | |
2353 | && convert_mult_to_fma (stmt, | |
2354 | gimple_call_arg (stmt, 0), | |
2355 | gimple_call_arg (stmt, 0))) | |
2356 | { | |
6716f635 | 2357 | unlink_stmt_vdef (stmt); |
15dbdc8f | 2358 | gsi_remove (&gsi, true); |
2359 | release_defs (stmt); | |
2360 | if (gimple_purge_dead_eh_edges (bb)) | |
2361 | cfg_changed = true; | |
2362 | continue; | |
2363 | } | |
2364 | break; | |
2365 | ||
2366 | default:; | |
2367 | } | |
2368 | } | |
2369 | } | |
b9be572e | 2370 | gsi_next (&gsi); |
62be004c | 2371 | } |
2372 | } | |
00f4f705 | 2373 | |
30c4e60d | 2374 | statistics_counter_event (cfun, "widening multiplications inserted", |
2375 | widen_mul_stats.widen_mults_inserted); | |
2376 | statistics_counter_event (cfun, "widening maccs inserted", | |
2377 | widen_mul_stats.maccs_inserted); | |
2378 | statistics_counter_event (cfun, "fused multiply-adds inserted", | |
2379 | widen_mul_stats.fmas_inserted); | |
2380 | ||
15dbdc8f | 2381 | return cfg_changed ? TODO_cleanup_cfg : 0; |
62be004c | 2382 | } |
2383 | ||
2384 | static bool | |
2385 | gate_optimize_widening_mul (void) | |
2386 | { | |
2387 | return flag_expensive_optimizations && optimize; | |
2388 | } | |
2389 | ||
2390 | struct gimple_opt_pass pass_optimize_widening_mul = | |
2391 | { | |
2392 | { | |
2393 | GIMPLE_PASS, | |
2394 | "widening_mul", /* name */ | |
2395 | gate_optimize_widening_mul, /* gate */ | |
2396 | execute_optimize_widening_mul, /* execute */ | |
2397 | NULL, /* sub */ | |
2398 | NULL, /* next */ | |
2399 | 0, /* static_pass_number */ | |
2400 | TV_NONE, /* tv_id */ | |
2401 | PROP_ssa, /* properties_required */ | |
2402 | 0, /* properties_provided */ | |
2403 | 0, /* properties_destroyed */ | |
2404 | 0, /* todo_flags_start */ | |
b9be572e | 2405 | TODO_verify_ssa |
2406 | | TODO_verify_stmts | |
2407 | | TODO_dump_func | |
2408 | | TODO_update_ssa /* todo_flags_finish */ | |
62be004c | 2409 | } |
2410 | }; |