]>
Commit | Line | Data |
---|---|---|
6c2a63a3 | 1 | /* Global, SSA-based optimizations using mathematical identities. |
85ec4feb | 2 | Copyright (C) 2005-2018 Free Software Foundation, Inc. |
b8698a0f | 3 | |
6c2a63a3 | 4 | This file is part of GCC. |
b8698a0f | 5 | |
6c2a63a3 PB |
6 | GCC is free software; you can redistribute it and/or modify it |
7 | under the terms of the GNU General Public License as published by the | |
9dcd6f09 | 8 | Free Software Foundation; either version 3, or (at your option) any |
6c2a63a3 | 9 | later version. |
b8698a0f | 10 | |
6c2a63a3 PB |
11 | GCC is distributed in the hope that it will be useful, but WITHOUT |
12 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
14 | for more details. | |
b8698a0f | 15 | |
6c2a63a3 | 16 | You should have received a copy of the GNU General Public License |
9dcd6f09 NC |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
6c2a63a3 PB |
19 | |
20 | /* Currently, the only mini-pass in this file tries to CSE reciprocal | |
21 | operations. These are common in sequences such as this one: | |
22 | ||
23 | modulus = sqrt(x*x + y*y + z*z); | |
24 | x = x / modulus; | |
25 | y = y / modulus; | |
26 | z = z / modulus; | |
27 | ||
28 | that can be optimized to | |
29 | ||
30 | modulus = sqrt(x*x + y*y + z*z); | |
31 | rmodulus = 1.0 / modulus; | |
32 | x = x * rmodulus; | |
33 | y = y * rmodulus; | |
34 | z = z * rmodulus; | |
35 | ||
36 | We do this for loop invariant divisors, and with this pass whenever | |
bc23502b PB |
37 | we notice that a division has the same divisor multiple times. |
38 | ||
39 | Of course, like in PRE, we don't insert a division if a dominator | |
40 | already has one. However, this cannot be done as an extension of | |
41 | PRE for several reasons. | |
42 | ||
43 | First of all, with some experiments it was found out that the | |
44 | transformation is not always useful if there are only two divisions | |
6edbcfc3 | 45 | by the same divisor. This is probably because modern processors |
bc23502b PB |
46 | can pipeline the divisions; on older, in-order processors it should |
47 | still be effective to optimize two divisions by the same number. | |
48 | We make this a param, and it shall be called N in the remainder of | |
49 | this comment. | |
50 | ||
51 | Second, if trapping math is active, we have less freedom on where | |
52 | to insert divisions: we can only do so in basic blocks that already | |
53 | contain one. (If divisions don't trap, instead, we can insert | |
54 | divisions elsewhere, which will be in blocks that are common dominators | |
55 | of those that have the division). | |
56 | ||
57 | We really don't want to compute the reciprocal unless a division will | |
58 | be found. To do this, we won't insert the division in a basic block | |
59 | that has less than N divisions *post-dominating* it. | |
60 | ||
61 | The algorithm constructs a subset of the dominator tree, holding the | |
62 | blocks containing the divisions and the common dominators to them, | |
63 | and walk it twice. The first walk is in post-order, and it annotates | |
64 | each block with the number of divisions that post-dominate it: this | |
65 | gives information on where divisions can be inserted profitably. | |
66 | The second walk is in pre-order, and it inserts divisions as explained | |
67 | above, and replaces divisions by multiplications. | |
68 | ||
69 | In the best case, the cost of the pass is O(n_statements). In the | |
70 | worst-case, the cost is due to creating the dominator tree subset, | |
71 | with a cost of O(n_basic_blocks ^ 2); however this can only happen | |
72 | for n_statements / n_basic_blocks statements. So, the amortized cost | |
73 | of creating the dominator tree subset is O(n_basic_blocks) and the | |
74 | worst-case cost of the pass is O(n_statements * n_basic_blocks). | |
75 | ||
76 | More practically, the cost will be small because there are few | |
77 | divisions, and they tend to be in the same basic block, so insert_bb | |
78 | is called very few times. | |
79 | ||
80 | If we did this using domwalk.c, an efficient implementation would have | |
81 | to work on all the variables in a single pass, because we could not | |
82 | work on just a subset of the dominator tree, as we do now, and the | |
83 | cost would also be something like O(n_statements * n_basic_blocks). | |
84 | The data structures would be more complex in order to work on all the | |
85 | variables in a single pass. */ | |
6c2a63a3 PB |
86 | |
87 | #include "config.h" | |
88 | #include "system.h" | |
89 | #include "coretypes.h" | |
c7131fb2 | 90 | #include "backend.h" |
957060b5 AM |
91 | #include "target.h" |
92 | #include "rtl.h" | |
c7131fb2 AM |
93 | #include "tree.h" |
94 | #include "gimple.h" | |
957060b5 AM |
95 | #include "predict.h" |
96 | #include "alloc-pool.h" | |
97 | #include "tree-pass.h" | |
c7131fb2 | 98 | #include "ssa.h" |
957060b5 AM |
99 | #include "optabs-tree.h" |
100 | #include "gimple-pretty-print.h" | |
40e23961 | 101 | #include "alias.h" |
40e23961 | 102 | #include "fold-const.h" |
2fb9a547 | 103 | #include "gimple-fold.h" |
5be5c238 | 104 | #include "gimple-iterator.h" |
73984f84 | 105 | #include "gimplify.h" |
18f429e2 | 106 | #include "gimplify-me.h" |
d8a2d370 | 107 | #include "stor-layout.h" |
442b4905 | 108 | #include "tree-cfg.h" |
442b4905 | 109 | #include "tree-dfa.h" |
7a300452 | 110 | #include "tree-ssa.h" |
9b2b7279 | 111 | #include "builtins.h" |
b7dce216 | 112 | #include "params.h" |
ee62a5a6 | 113 | #include "internal-fn.h" |
c97d1c9d | 114 | #include "case-cfn-macros.h" |
e72531b9 PK |
115 | #include "optabs-libfuncs.h" |
116 | #include "tree-eh.h" | |
117 | #include "targhooks.h" | |
4a0d0ed2 | 118 | #include "domwalk.h" |
bc23502b PB |
119 | |
120 | /* This structure represents one basic block that either computes a | |
121 | division, or is a common dominator for basic block that compute a | |
122 | division. */ | |
123 | struct occurrence { | |
124 | /* The basic block represented by this structure. */ | |
125 | basic_block bb; | |
126 | ||
127 | /* If non-NULL, the SSA_NAME holding the definition for a reciprocal | |
128 | inserted in BB. */ | |
129 | tree recip_def; | |
130 | ||
19cf3a36 JW |
131 | /* If non-NULL, the SSA_NAME holding the definition for a squared |
132 | reciprocal inserted in BB. */ | |
133 | tree square_recip_def; | |
134 | ||
726a989a | 135 | /* If non-NULL, the GIMPLE_ASSIGN for a reciprocal computation that |
bc23502b | 136 | was inserted in BB. */ |
355fe088 | 137 | gimple *recip_def_stmt; |
bc23502b PB |
138 | |
139 | /* Pointer to a list of "struct occurrence"s for blocks dominated | |
140 | by BB. */ | |
141 | struct occurrence *children; | |
142 | ||
143 | /* Pointer to the next "struct occurrence"s in the list of blocks | |
144 | sharing a common dominator. */ | |
145 | struct occurrence *next; | |
146 | ||
147 | /* The number of divisions that are in BB before compute_merit. The | |
148 | number of divisions that are in BB or post-dominate it after | |
149 | compute_merit. */ | |
150 | int num_divisions; | |
151 | ||
152 | /* True if the basic block has a division, false if it is a common | |
153 | dominator for basic blocks that do. If it is false and trapping | |
154 | math is active, BB is not a candidate for inserting a reciprocal. */ | |
155 | bool bb_has_division; | |
156 | }; | |
157 | ||
4da3b811 NF |
158 | static struct |
159 | { | |
160 | /* Number of 1.0/X ops inserted. */ | |
161 | int rdivs_inserted; | |
162 | ||
163 | /* Number of 1.0/FUNC ops inserted. */ | |
164 | int rfuncs_inserted; | |
165 | } reciprocal_stats; | |
166 | ||
167 | static struct | |
168 | { | |
169 | /* Number of cexpi calls inserted. */ | |
170 | int inserted; | |
171 | } sincos_stats; | |
172 | ||
4da3b811 NF |
173 | static struct |
174 | { | |
175 | /* Number of widening multiplication ops inserted. */ | |
176 | int widen_mults_inserted; | |
177 | ||
178 | /* Number of integer multiply-and-accumulate ops inserted. */ | |
179 | int maccs_inserted; | |
180 | ||
181 | /* Number of fp fused multiply-add ops inserted. */ | |
182 | int fmas_inserted; | |
e72531b9 PK |
183 | |
184 | /* Number of divmod calls inserted. */ | |
185 | int divmod_calls_inserted; | |
4da3b811 | 186 | } widen_mul_stats; |
bc23502b PB |
187 | |
188 | /* The instance of "struct occurrence" representing the highest | |
189 | interesting block in the dominator tree. */ | |
190 | static struct occurrence *occ_head; | |
191 | ||
192 | /* Allocation pool for getting instances of "struct occurrence". */ | |
fb0b2914 | 193 | static object_allocator<occurrence> *occ_pool; |
bc23502b PB |
194 | |
195 | ||
196 | ||
197 | /* Allocate and return a new struct occurrence for basic block BB, and | |
198 | whose children list is headed by CHILDREN. */ | |
199 | static struct occurrence * | |
200 | occ_new (basic_block bb, struct occurrence *children) | |
6c2a63a3 | 201 | { |
bc23502b PB |
202 | struct occurrence *occ; |
203 | ||
2cc777fb | 204 | bb->aux = occ = occ_pool->allocate (); |
bc23502b PB |
205 | memset (occ, 0, sizeof (struct occurrence)); |
206 | ||
207 | occ->bb = bb; | |
208 | occ->children = children; | |
209 | return occ; | |
6c2a63a3 PB |
210 | } |
211 | ||
bc23502b PB |
212 | |
213 | /* Insert NEW_OCC into our subset of the dominator tree. P_HEAD points to a | |
214 | list of "struct occurrence"s, one per basic block, having IDOM as | |
215 | their common dominator. | |
216 | ||
217 | We try to insert NEW_OCC as deep as possible in the tree, and we also | |
218 | insert any other block that is a common dominator for BB and one | |
219 | block already in the tree. */ | |
220 | ||
221 | static void | |
222 | insert_bb (struct occurrence *new_occ, basic_block idom, | |
223 | struct occurrence **p_head) | |
2ef571e2 | 224 | { |
bc23502b | 225 | struct occurrence *occ, **p_occ; |
2ef571e2 | 226 | |
bc23502b PB |
227 | for (p_occ = p_head; (occ = *p_occ) != NULL; ) |
228 | { | |
229 | basic_block bb = new_occ->bb, occ_bb = occ->bb; | |
230 | basic_block dom = nearest_common_dominator (CDI_DOMINATORS, occ_bb, bb); | |
231 | if (dom == bb) | |
232 | { | |
233 | /* BB dominates OCC_BB. OCC becomes NEW_OCC's child: remove OCC | |
234 | from its list. */ | |
235 | *p_occ = occ->next; | |
236 | occ->next = new_occ->children; | |
237 | new_occ->children = occ; | |
238 | ||
239 | /* Try the next block (it may as well be dominated by BB). */ | |
240 | } | |
241 | ||
242 | else if (dom == occ_bb) | |
243 | { | |
244 | /* OCC_BB dominates BB. Tail recurse to look deeper. */ | |
245 | insert_bb (new_occ, dom, &occ->children); | |
246 | return; | |
247 | } | |
248 | ||
249 | else if (dom != idom) | |
250 | { | |
251 | gcc_assert (!dom->aux); | |
252 | ||
253 | /* There is a dominator between IDOM and BB, add it and make | |
254 | two children out of NEW_OCC and OCC. First, remove OCC from | |
255 | its list. */ | |
256 | *p_occ = occ->next; | |
257 | new_occ->next = occ; | |
258 | occ->next = NULL; | |
259 | ||
260 | /* None of the previous blocks has DOM as a dominator: if we tail | |
261 | recursed, we would reexamine them uselessly. Just switch BB with | |
262 | DOM, and go on looking for blocks dominated by DOM. */ | |
263 | new_occ = occ_new (dom, new_occ); | |
264 | } | |
265 | ||
266 | else | |
267 | { | |
268 | /* Nothing special, go on with the next element. */ | |
269 | p_occ = &occ->next; | |
270 | } | |
271 | } | |
272 | ||
273 | /* No place was found as a child of IDOM. Make BB a sibling of IDOM. */ | |
274 | new_occ->next = *p_head; | |
275 | *p_head = new_occ; | |
276 | } | |
277 | ||
19cf3a36 JW |
278 | /* Register that we found a division in BB. |
279 | IMPORTANCE is a measure of how much weighting to give | |
280 | that division. Use IMPORTANCE = 2 to register a single | |
281 | division. If the division is going to be found multiple | |
282 | times use 1 (as it is with squares). */ | |
bc23502b PB |
283 | |
284 | static inline void | |
19cf3a36 | 285 | register_division_in (basic_block bb, int importance) |
bc23502b PB |
286 | { |
287 | struct occurrence *occ; | |
288 | ||
289 | occ = (struct occurrence *) bb->aux; | |
290 | if (!occ) | |
291 | { | |
292 | occ = occ_new (bb, NULL); | |
fefa31b5 | 293 | insert_bb (occ, ENTRY_BLOCK_PTR_FOR_FN (cfun), &occ_head); |
bc23502b PB |
294 | } |
295 | ||
296 | occ->bb_has_division = true; | |
19cf3a36 | 297 | occ->num_divisions += importance; |
bc23502b PB |
298 | } |
299 | ||
300 | ||
301 | /* Compute the number of divisions that postdominate each block in OCC and | |
302 | its children. */ | |
6c2a63a3 | 303 | |
6c2a63a3 | 304 | static void |
bc23502b | 305 | compute_merit (struct occurrence *occ) |
6c2a63a3 | 306 | { |
bc23502b PB |
307 | struct occurrence *occ_child; |
308 | basic_block dom = occ->bb; | |
6c2a63a3 | 309 | |
bc23502b | 310 | for (occ_child = occ->children; occ_child; occ_child = occ_child->next) |
6c2a63a3 | 311 | { |
bc23502b PB |
312 | basic_block bb; |
313 | if (occ_child->children) | |
314 | compute_merit (occ_child); | |
315 | ||
316 | if (flag_exceptions) | |
317 | bb = single_noncomplex_succ (dom); | |
318 | else | |
319 | bb = dom; | |
320 | ||
321 | if (dominated_by_p (CDI_POST_DOMINATORS, bb, occ_child->bb)) | |
322 | occ->num_divisions += occ_child->num_divisions; | |
323 | } | |
324 | } | |
325 | ||
326 | ||
327 | /* Return whether USE_STMT is a floating-point division by DEF. */ | |
328 | static inline bool | |
355fe088 | 329 | is_division_by (gimple *use_stmt, tree def) |
bc23502b | 330 | { |
726a989a RB |
331 | return is_gimple_assign (use_stmt) |
332 | && gimple_assign_rhs_code (use_stmt) == RDIV_EXPR | |
333 | && gimple_assign_rhs2 (use_stmt) == def | |
8a5b57cd RG |
334 | /* Do not recognize x / x as valid division, as we are getting |
335 | confused later by replacing all immediate uses x in such | |
336 | a stmt. */ | |
726a989a | 337 | && gimple_assign_rhs1 (use_stmt) != def; |
bc23502b PB |
338 | } |
339 | ||
19cf3a36 JW |
340 | /* Return whether USE_STMT is DEF * DEF. */ |
341 | static inline bool | |
342 | is_square_of (gimple *use_stmt, tree def) | |
343 | { | |
344 | if (gimple_code (use_stmt) == GIMPLE_ASSIGN | |
345 | && gimple_assign_rhs_code (use_stmt) == MULT_EXPR) | |
346 | { | |
347 | tree op0 = gimple_assign_rhs1 (use_stmt); | |
348 | tree op1 = gimple_assign_rhs2 (use_stmt); | |
349 | ||
350 | return op0 == op1 && op0 == def; | |
351 | } | |
352 | return 0; | |
353 | } | |
354 | ||
355 | /* Return whether USE_STMT is a floating-point division by | |
356 | DEF * DEF. */ | |
357 | static inline bool | |
358 | is_division_by_square (gimple *use_stmt, tree def) | |
359 | { | |
360 | if (gimple_code (use_stmt) == GIMPLE_ASSIGN | |
361 | && gimple_assign_rhs_code (use_stmt) == RDIV_EXPR | |
362 | && gimple_assign_rhs1 (use_stmt) != gimple_assign_rhs2 (use_stmt)) | |
363 | { | |
364 | tree denominator = gimple_assign_rhs2 (use_stmt); | |
365 | if (TREE_CODE (denominator) == SSA_NAME) | |
366 | { | |
367 | return is_square_of (SSA_NAME_DEF_STMT (denominator), def); | |
368 | } | |
369 | } | |
370 | return 0; | |
371 | } | |
372 | ||
bc23502b PB |
373 | /* Walk the subset of the dominator tree rooted at OCC, setting the |
374 | RECIP_DEF field to a definition of 1.0 / DEF that can be used in | |
375 | the given basic block. The field may be left NULL, of course, | |
376 | if it is not possible or profitable to do the optimization. | |
377 | ||
378 | DEF_BSI is an iterator pointing at the statement defining DEF. | |
379 | If RECIP_DEF is set, a dominator already has a computation that can | |
19cf3a36 JW |
380 | be used. |
381 | ||
382 | If should_insert_square_recip is set, then this also inserts | |
383 | the square of the reciprocal immediately after the definition | |
384 | of the reciprocal. */ | |
bc23502b PB |
385 | |
386 | static void | |
726a989a | 387 | insert_reciprocals (gimple_stmt_iterator *def_gsi, struct occurrence *occ, |
19cf3a36 JW |
388 | tree def, tree recip_def, tree square_recip_def, |
389 | int should_insert_square_recip, int threshold) | |
bc23502b | 390 | { |
726a989a | 391 | tree type; |
19cf3a36 | 392 | gassign *new_stmt, *new_square_stmt; |
726a989a | 393 | gimple_stmt_iterator gsi; |
bc23502b PB |
394 | struct occurrence *occ_child; |
395 | ||
396 | if (!recip_def | |
397 | && (occ->bb_has_division || !flag_trapping_math) | |
19cf3a36 JW |
398 | /* Divide by two as all divisions are counted twice in |
399 | the costing loop. */ | |
400 | && occ->num_divisions / 2 >= threshold) | |
bc23502b PB |
401 | { |
402 | /* Make a variable with the replacement and substitute it. */ | |
403 | type = TREE_TYPE (def); | |
7cc434a3 | 404 | recip_def = create_tmp_reg (type, "reciptmp"); |
0d0e4a03 JJ |
405 | new_stmt = gimple_build_assign (recip_def, RDIV_EXPR, |
406 | build_one_cst (type), def); | |
b8698a0f | 407 | |
19cf3a36 JW |
408 | if (should_insert_square_recip) |
409 | { | |
410 | square_recip_def = create_tmp_reg (type, "powmult_reciptmp"); | |
411 | new_square_stmt = gimple_build_assign (square_recip_def, MULT_EXPR, | |
412 | recip_def, recip_def); | |
413 | } | |
414 | ||
bc23502b | 415 | if (occ->bb_has_division) |
19cf3a36 JW |
416 | { |
417 | /* Case 1: insert before an existing division. */ | |
418 | gsi = gsi_after_labels (occ->bb); | |
419 | while (!gsi_end_p (gsi) | |
420 | && (!is_division_by (gsi_stmt (gsi), def)) | |
421 | && (!is_division_by_square (gsi_stmt (gsi), def))) | |
726a989a | 422 | gsi_next (&gsi); |
bc23502b | 423 | |
19cf3a36 | 424 | gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT); |
8be591a9 JJ |
425 | if (should_insert_square_recip) |
426 | gsi_insert_before (&gsi, new_square_stmt, GSI_SAME_STMT); | |
19cf3a36 | 427 | } |
726a989a | 428 | else if (def_gsi && occ->bb == def_gsi->bb) |
19cf3a36 JW |
429 | { |
430 | /* Case 2: insert right after the definition. Note that this will | |
bc23502b PB |
431 | never happen if the definition statement can throw, because in |
432 | that case the sole successor of the statement's basic block will | |
433 | dominate all the uses as well. */ | |
19cf3a36 | 434 | gsi_insert_after (def_gsi, new_stmt, GSI_NEW_STMT); |
8be591a9 JJ |
435 | if (should_insert_square_recip) |
436 | gsi_insert_after (def_gsi, new_square_stmt, GSI_NEW_STMT); | |
19cf3a36 | 437 | } |
bc23502b | 438 | else |
19cf3a36 JW |
439 | { |
440 | /* Case 3: insert in a basic block not containing defs/uses. */ | |
441 | gsi = gsi_after_labels (occ->bb); | |
442 | gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT); | |
8be591a9 JJ |
443 | if (should_insert_square_recip) |
444 | gsi_insert_before (&gsi, new_square_stmt, GSI_SAME_STMT); | |
19cf3a36 JW |
445 | } |
446 | ||
4da3b811 NF |
447 | reciprocal_stats.rdivs_inserted++; |
448 | ||
bc23502b | 449 | occ->recip_def_stmt = new_stmt; |
6c2a63a3 PB |
450 | } |
451 | ||
bc23502b | 452 | occ->recip_def = recip_def; |
19cf3a36 | 453 | occ->square_recip_def = square_recip_def; |
bc23502b | 454 | for (occ_child = occ->children; occ_child; occ_child = occ_child->next) |
19cf3a36 JW |
455 | insert_reciprocals (def_gsi, occ_child, def, recip_def, |
456 | square_recip_def, should_insert_square_recip, | |
457 | threshold); | |
458 | } | |
459 | ||
460 | /* Replace occurrences of expr / (x * x) with expr * ((1 / x) * (1 / x)). | |
461 | Take as argument the use for (x * x). */ | |
462 | static inline void | |
463 | replace_reciprocal_squares (use_operand_p use_p) | |
464 | { | |
465 | gimple *use_stmt = USE_STMT (use_p); | |
466 | basic_block bb = gimple_bb (use_stmt); | |
467 | struct occurrence *occ = (struct occurrence *) bb->aux; | |
468 | ||
469 | if (optimize_bb_for_speed_p (bb) && occ->square_recip_def | |
470 | && occ->recip_def) | |
471 | { | |
472 | gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt); | |
473 | gimple_assign_set_rhs_code (use_stmt, MULT_EXPR); | |
474 | gimple_assign_set_rhs2 (use_stmt, occ->square_recip_def); | |
475 | SET_USE (use_p, occ->square_recip_def); | |
476 | fold_stmt_inplace (&gsi); | |
477 | update_stmt (use_stmt); | |
478 | } | |
bc23502b PB |
479 | } |
480 | ||
481 | ||
482 | /* Replace the division at USE_P with a multiplication by the reciprocal, if | |
483 | possible. */ | |
484 | ||
485 | static inline void | |
486 | replace_reciprocal (use_operand_p use_p) | |
487 | { | |
355fe088 | 488 | gimple *use_stmt = USE_STMT (use_p); |
726a989a | 489 | basic_block bb = gimple_bb (use_stmt); |
bc23502b PB |
490 | struct occurrence *occ = (struct occurrence *) bb->aux; |
491 | ||
efd8f750 JH |
492 | if (optimize_bb_for_speed_p (bb) |
493 | && occ->recip_def && use_stmt != occ->recip_def_stmt) | |
bc23502b | 494 | { |
59401b92 | 495 | gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt); |
726a989a | 496 | gimple_assign_set_rhs_code (use_stmt, MULT_EXPR); |
bc23502b | 497 | SET_USE (use_p, occ->recip_def); |
59401b92 | 498 | fold_stmt_inplace (&gsi); |
bc23502b PB |
499 | update_stmt (use_stmt); |
500 | } | |
501 | } | |
502 | ||
503 | ||
504 | /* Free OCC and return one more "struct occurrence" to be freed. */ | |
505 | ||
506 | static struct occurrence * | |
507 | free_bb (struct occurrence *occ) | |
508 | { | |
509 | struct occurrence *child, *next; | |
510 | ||
511 | /* First get the two pointers hanging off OCC. */ | |
512 | next = occ->next; | |
513 | child = occ->children; | |
514 | occ->bb->aux = NULL; | |
2cc777fb | 515 | occ_pool->remove (occ); |
bc23502b PB |
516 | |
517 | /* Now ensure that we don't recurse unless it is necessary. */ | |
518 | if (!child) | |
519 | return next; | |
2ef571e2 | 520 | else |
bc23502b PB |
521 | { |
522 | while (next) | |
523 | next = free_bb (next); | |
524 | ||
525 | return child; | |
526 | } | |
527 | } | |
528 | ||
529 | ||
530 | /* Look for floating-point divisions among DEF's uses, and try to | |
531 | replace them by multiplications with the reciprocal. Add | |
532 | as many statements computing the reciprocal as needed. | |
533 | ||
534 | DEF must be a GIMPLE register of a floating-point type. */ | |
535 | ||
536 | static void | |
726a989a | 537 | execute_cse_reciprocals_1 (gimple_stmt_iterator *def_gsi, tree def) |
bc23502b | 538 | { |
19cf3a36 JW |
539 | use_operand_p use_p, square_use_p; |
540 | imm_use_iterator use_iter, square_use_iter; | |
541 | tree square_def; | |
bc23502b | 542 | struct occurrence *occ; |
19cf3a36 JW |
543 | int count = 0; |
544 | int threshold; | |
545 | int square_recip_count = 0; | |
546 | int sqrt_recip_count = 0; | |
6c2a63a3 | 547 | |
682820cc | 548 | gcc_assert (FLOAT_TYPE_P (TREE_TYPE (def)) && TREE_CODE (def) == SSA_NAME); |
19cf3a36 JW |
549 | threshold = targetm.min_divisions_for_recip_mul (TYPE_MODE (TREE_TYPE (def))); |
550 | ||
5f84eb27 WD |
551 | /* If DEF is a square (x * x), count the number of divisions by x. |
552 | If there are more divisions by x than by (DEF * DEF), prefer to optimize | |
553 | the reciprocal of x instead of DEF. This improves cases like: | |
554 | def = x * x | |
555 | t0 = a / def | |
556 | t1 = b / def | |
557 | t2 = c / x | |
558 | Reciprocal optimization of x results in 1 division rather than 2 or 3. */ | |
559 | gimple *def_stmt = SSA_NAME_DEF_STMT (def); | |
560 | ||
561 | if (is_gimple_assign (def_stmt) | |
562 | && gimple_assign_rhs_code (def_stmt) == MULT_EXPR | |
563 | && TREE_CODE (gimple_assign_rhs1 (def_stmt)) == SSA_NAME | |
564 | && gimple_assign_rhs1 (def_stmt) == gimple_assign_rhs2 (def_stmt)) | |
19cf3a36 | 565 | { |
5f84eb27 | 566 | tree op0 = gimple_assign_rhs1 (def_stmt); |
19cf3a36 | 567 | |
5f84eb27 | 568 | FOR_EACH_IMM_USE_FAST (use_p, use_iter, op0) |
19cf3a36 | 569 | { |
5f84eb27 WD |
570 | gimple *use_stmt = USE_STMT (use_p); |
571 | if (is_division_by (use_stmt, op0)) | |
572 | sqrt_recip_count++; | |
19cf3a36 JW |
573 | } |
574 | } | |
bc23502b PB |
575 | |
576 | FOR_EACH_IMM_USE_FAST (use_p, use_iter, def) | |
6c2a63a3 | 577 | { |
355fe088 | 578 | gimple *use_stmt = USE_STMT (use_p); |
bc23502b | 579 | if (is_division_by (use_stmt, def)) |
6c2a63a3 | 580 | { |
19cf3a36 | 581 | register_division_in (gimple_bb (use_stmt), 2); |
bc23502b | 582 | count++; |
6c2a63a3 | 583 | } |
19cf3a36 JW |
584 | |
585 | if (is_square_of (use_stmt, def)) | |
586 | { | |
587 | square_def = gimple_assign_lhs (use_stmt); | |
588 | FOR_EACH_IMM_USE_FAST (square_use_p, square_use_iter, square_def) | |
589 | { | |
590 | gimple *square_use_stmt = USE_STMT (square_use_p); | |
591 | if (is_division_by (square_use_stmt, square_def)) | |
592 | { | |
5f84eb27 | 593 | /* This is executed twice for each division by a square. */ |
19cf3a36 | 594 | register_division_in (gimple_bb (square_use_stmt), 1); |
5f84eb27 | 595 | square_recip_count++; |
19cf3a36 JW |
596 | } |
597 | } | |
598 | } | |
6c2a63a3 | 599 | } |
b8698a0f | 600 | |
5f84eb27 | 601 | /* Square reciprocals were counted twice above. */ |
19cf3a36 JW |
602 | square_recip_count /= 2; |
603 | ||
5f84eb27 | 604 | /* If it is more profitable to optimize 1 / x, don't optimize 1 / (x * x). */ |
19cf3a36 | 605 | if (sqrt_recip_count > square_recip_count) |
19cf3a36 JW |
606 | return; |
607 | ||
bc23502b | 608 | /* Do the expensive part only if we can hope to optimize something. */ |
5f84eb27 | 609 | if (count + square_recip_count >= threshold && count >= 1) |
bc23502b | 610 | { |
355fe088 | 611 | gimple *use_stmt; |
bc23502b PB |
612 | for (occ = occ_head; occ; occ = occ->next) |
613 | { | |
614 | compute_merit (occ); | |
19cf3a36 JW |
615 | insert_reciprocals (def_gsi, occ, def, NULL, NULL, |
616 | square_recip_count, threshold); | |
bc23502b PB |
617 | } |
618 | ||
6c00f606 | 619 | FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, def) |
bc23502b | 620 | { |
bc23502b | 621 | if (is_division_by (use_stmt, def)) |
6c00f606 AM |
622 | { |
623 | FOR_EACH_IMM_USE_ON_STMT (use_p, use_iter) | |
624 | replace_reciprocal (use_p); | |
625 | } | |
5f84eb27 | 626 | else if (square_recip_count > 0 && is_square_of (use_stmt, def)) |
19cf3a36 JW |
627 | { |
628 | FOR_EACH_IMM_USE_ON_STMT (use_p, use_iter) | |
629 | { | |
630 | /* Find all uses of the square that are divisions and | |
631 | * replace them by multiplications with the inverse. */ | |
632 | imm_use_iterator square_iterator; | |
633 | gimple *powmult_use_stmt = USE_STMT (use_p); | |
634 | tree powmult_def_name = gimple_assign_lhs (powmult_use_stmt); | |
635 | ||
636 | FOR_EACH_IMM_USE_STMT (powmult_use_stmt, | |
637 | square_iterator, powmult_def_name) | |
638 | FOR_EACH_IMM_USE_ON_STMT (square_use_p, square_iterator) | |
639 | { | |
640 | gimple *powmult_use_stmt = USE_STMT (square_use_p); | |
641 | if (is_division_by (powmult_use_stmt, powmult_def_name)) | |
642 | replace_reciprocal_squares (square_use_p); | |
643 | } | |
644 | } | |
645 | } | |
bc23502b PB |
646 | } |
647 | } | |
648 | ||
649 | for (occ = occ_head; occ; ) | |
650 | occ = free_bb (occ); | |
651 | ||
652 | occ_head = NULL; | |
6c2a63a3 PB |
653 | } |
654 | ||
ee62a5a6 RS |
655 | /* Return an internal function that implements the reciprocal of CALL, |
656 | or IFN_LAST if there is no such function that the target supports. */ | |
657 | ||
658 | internal_fn | |
659 | internal_fn_reciprocal (gcall *call) | |
660 | { | |
661 | internal_fn ifn; | |
662 | ||
663 | switch (gimple_call_combined_fn (call)) | |
664 | { | |
665 | CASE_CFN_SQRT: | |
ee5fd23a | 666 | CASE_CFN_SQRT_FN: |
ee62a5a6 RS |
667 | ifn = IFN_RSQRT; |
668 | break; | |
669 | ||
670 | default: | |
671 | return IFN_LAST; | |
672 | } | |
673 | ||
674 | tree_pair types = direct_internal_fn_types (ifn, call); | |
675 | if (!direct_internal_fn_supported_p (ifn, types, OPTIMIZE_FOR_SPEED)) | |
676 | return IFN_LAST; | |
677 | ||
678 | return ifn; | |
679 | } | |
680 | ||
bc23502b PB |
681 | /* Go through all the floating-point SSA_NAMEs, and call |
682 | execute_cse_reciprocals_1 on each of them. */ | |
be55bfe6 TS |
683 | namespace { |
684 | ||
685 | const pass_data pass_data_cse_reciprocals = | |
686 | { | |
687 | GIMPLE_PASS, /* type */ | |
688 | "recip", /* name */ | |
689 | OPTGROUP_NONE, /* optinfo_flags */ | |
1b6546cc | 690 | TV_TREE_RECIP, /* tv_id */ |
be55bfe6 TS |
691 | PROP_ssa, /* properties_required */ |
692 | 0, /* properties_provided */ | |
693 | 0, /* properties_destroyed */ | |
694 | 0, /* todo_flags_start */ | |
3bea341f | 695 | TODO_update_ssa, /* todo_flags_finish */ |
be55bfe6 TS |
696 | }; |
697 | ||
698 | class pass_cse_reciprocals : public gimple_opt_pass | |
699 | { | |
700 | public: | |
701 | pass_cse_reciprocals (gcc::context *ctxt) | |
702 | : gimple_opt_pass (pass_data_cse_reciprocals, ctxt) | |
703 | {} | |
704 | ||
705 | /* opt_pass methods: */ | |
706 | virtual bool gate (function *) { return optimize && flag_reciprocal_math; } | |
707 | virtual unsigned int execute (function *); | |
708 | ||
709 | }; // class pass_cse_reciprocals | |
710 | ||
711 | unsigned int | |
712 | pass_cse_reciprocals::execute (function *fun) | |
6c2a63a3 PB |
713 | { |
714 | basic_block bb; | |
a8f82ec4 | 715 | tree arg; |
ac264fef | 716 | |
fcb87c50 | 717 | occ_pool = new object_allocator<occurrence> ("dominators for recip"); |
ac264fef | 718 | |
4da3b811 | 719 | memset (&reciprocal_stats, 0, sizeof (reciprocal_stats)); |
d898f3ce PB |
720 | calculate_dominance_info (CDI_DOMINATORS); |
721 | calculate_dominance_info (CDI_POST_DOMINATORS); | |
bc23502b | 722 | |
b2b29377 MM |
723 | if (flag_checking) |
724 | FOR_EACH_BB_FN (bb, fun) | |
725 | gcc_assert (!bb->aux); | |
bc23502b | 726 | |
be55bfe6 | 727 | for (arg = DECL_ARGUMENTS (fun->decl); arg; arg = DECL_CHAIN (arg)) |
32244553 | 728 | if (FLOAT_TYPE_P (TREE_TYPE (arg)) |
bc23502b | 729 | && is_gimple_reg (arg)) |
32244553 | 730 | { |
be55bfe6 | 731 | tree name = ssa_default_def (fun, arg); |
32244553 RG |
732 | if (name) |
733 | execute_cse_reciprocals_1 (NULL, name); | |
734 | } | |
a8f82ec4 | 735 | |
be55bfe6 | 736 | FOR_EACH_BB_FN (bb, fun) |
6c2a63a3 | 737 | { |
726a989a | 738 | tree def; |
6c2a63a3 | 739 | |
538dd0b7 DM |
740 | for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi); |
741 | gsi_next (&gsi)) | |
6c2a63a3 | 742 | { |
538dd0b7 | 743 | gphi *phi = gsi.phi (); |
6c2a63a3 | 744 | def = PHI_RESULT (phi); |
ea057359 RG |
745 | if (! virtual_operand_p (def) |
746 | && FLOAT_TYPE_P (TREE_TYPE (def))) | |
bc23502b | 747 | execute_cse_reciprocals_1 (NULL, def); |
6c2a63a3 PB |
748 | } |
749 | ||
538dd0b7 DM |
750 | for (gimple_stmt_iterator gsi = gsi_after_labels (bb); !gsi_end_p (gsi); |
751 | gsi_next (&gsi)) | |
6c2a63a3 | 752 | { |
355fe088 | 753 | gimple *stmt = gsi_stmt (gsi); |
2f397a93 | 754 | |
726a989a | 755 | if (gimple_has_lhs (stmt) |
6c2a63a3 PB |
756 | && (def = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_DEF)) != NULL |
757 | && FLOAT_TYPE_P (TREE_TYPE (def)) | |
a8f82ec4 | 758 | && TREE_CODE (def) == SSA_NAME) |
726a989a | 759 | execute_cse_reciprocals_1 (&gsi, def); |
6c2a63a3 | 760 | } |
6b889d89 | 761 | |
efd8f750 JH |
762 | if (optimize_bb_for_size_p (bb)) |
763 | continue; | |
764 | ||
6b889d89 | 765 | /* Scan for a/func(b) and convert it to reciprocal a*rfunc(b). */ |
538dd0b7 DM |
766 | for (gimple_stmt_iterator gsi = gsi_after_labels (bb); !gsi_end_p (gsi); |
767 | gsi_next (&gsi)) | |
6b889d89 | 768 | { |
355fe088 | 769 | gimple *stmt = gsi_stmt (gsi); |
6b889d89 | 770 | |
726a989a RB |
771 | if (is_gimple_assign (stmt) |
772 | && gimple_assign_rhs_code (stmt) == RDIV_EXPR) | |
6b889d89 | 773 | { |
726a989a | 774 | tree arg1 = gimple_assign_rhs2 (stmt); |
355fe088 | 775 | gimple *stmt1; |
ac10986f UB |
776 | |
777 | if (TREE_CODE (arg1) != SSA_NAME) | |
778 | continue; | |
779 | ||
780 | stmt1 = SSA_NAME_DEF_STMT (arg1); | |
6b889d89 | 781 | |
726a989a | 782 | if (is_gimple_call (stmt1) |
ee62a5a6 | 783 | && gimple_call_lhs (stmt1)) |
6b889d89 | 784 | { |
7b90c63a | 785 | bool fail; |
79af7c1f MM |
786 | imm_use_iterator ui; |
787 | use_operand_p use_p; | |
ee62a5a6 | 788 | tree fndecl = NULL_TREE; |
6b889d89 | 789 | |
ee62a5a6 RS |
790 | gcall *call = as_a <gcall *> (stmt1); |
791 | internal_fn ifn = internal_fn_reciprocal (call); | |
792 | if (ifn == IFN_LAST) | |
793 | { | |
794 | fndecl = gimple_call_fndecl (call); | |
795 | if (!fndecl | |
3d78e008 | 796 | || !fndecl_built_in_p (fndecl, BUILT_IN_MD)) |
ee62a5a6 RS |
797 | continue; |
798 | fndecl = targetm.builtin_reciprocal (fndecl); | |
799 | if (!fndecl) | |
800 | continue; | |
801 | } | |
6b889d89 | 802 | |
79af7c1f MM |
803 | /* Check that all uses of the SSA name are divisions, |
804 | otherwise replacing the defining statement will do | |
805 | the wrong thing. */ | |
806 | fail = false; | |
807 | FOR_EACH_IMM_USE_FAST (use_p, ui, arg1) | |
808 | { | |
355fe088 | 809 | gimple *stmt2 = USE_STMT (use_p); |
79af7c1f MM |
810 | if (is_gimple_debug (stmt2)) |
811 | continue; | |
812 | if (!is_gimple_assign (stmt2) | |
813 | || gimple_assign_rhs_code (stmt2) != RDIV_EXPR | |
814 | || gimple_assign_rhs1 (stmt2) == arg1 | |
815 | || gimple_assign_rhs2 (stmt2) != arg1) | |
816 | { | |
817 | fail = true; | |
818 | break; | |
819 | } | |
820 | } | |
821 | if (fail) | |
822 | continue; | |
823 | ||
ee62a5a6 RS |
824 | gimple_replace_ssa_lhs (call, arg1); |
825 | if (gimple_call_internal_p (call) != (ifn != IFN_LAST)) | |
7b90c63a JJ |
826 | { |
827 | auto_vec<tree, 4> args; | |
828 | for (unsigned int i = 0; | |
ee62a5a6 RS |
829 | i < gimple_call_num_args (call); i++) |
830 | args.safe_push (gimple_call_arg (call, i)); | |
831 | gcall *stmt2; | |
832 | if (ifn == IFN_LAST) | |
833 | stmt2 = gimple_build_call_vec (fndecl, args); | |
834 | else | |
835 | stmt2 = gimple_build_call_internal_vec (ifn, args); | |
7b90c63a | 836 | gimple_call_set_lhs (stmt2, arg1); |
ee62a5a6 | 837 | if (gimple_vdef (call)) |
7b90c63a | 838 | { |
ee62a5a6 | 839 | gimple_set_vdef (stmt2, gimple_vdef (call)); |
7b90c63a JJ |
840 | SSA_NAME_DEF_STMT (gimple_vdef (stmt2)) = stmt2; |
841 | } | |
a844293d RS |
842 | gimple_call_set_nothrow (stmt2, |
843 | gimple_call_nothrow_p (call)); | |
ee62a5a6 RS |
844 | gimple_set_vuse (stmt2, gimple_vuse (call)); |
845 | gimple_stmt_iterator gsi2 = gsi_for_stmt (call); | |
7b90c63a JJ |
846 | gsi_replace (&gsi2, stmt2, true); |
847 | } | |
848 | else | |
849 | { | |
ee62a5a6 RS |
850 | if (ifn == IFN_LAST) |
851 | gimple_call_set_fndecl (call, fndecl); | |
852 | else | |
853 | gimple_call_set_internal_fn (call, ifn); | |
854 | update_stmt (call); | |
7b90c63a | 855 | } |
4da3b811 | 856 | reciprocal_stats.rfuncs_inserted++; |
6b889d89 | 857 | |
79af7c1f MM |
858 | FOR_EACH_IMM_USE_STMT (stmt, ui, arg1) |
859 | { | |
59401b92 | 860 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); |
79af7c1f | 861 | gimple_assign_set_rhs_code (stmt, MULT_EXPR); |
59401b92 | 862 | fold_stmt_inplace (&gsi); |
79af7c1f MM |
863 | update_stmt (stmt); |
864 | } | |
6b889d89 UB |
865 | } |
866 | } | |
867 | } | |
6c2a63a3 | 868 | } |
ac264fef | 869 | |
be55bfe6 | 870 | statistics_counter_event (fun, "reciprocal divs inserted", |
4da3b811 | 871 | reciprocal_stats.rdivs_inserted); |
be55bfe6 | 872 | statistics_counter_event (fun, "reciprocal functions inserted", |
4da3b811 NF |
873 | reciprocal_stats.rfuncs_inserted); |
874 | ||
d898f3ce PB |
875 | free_dominance_info (CDI_DOMINATORS); |
876 | free_dominance_info (CDI_POST_DOMINATORS); | |
2cc777fb | 877 | delete occ_pool; |
c2924966 | 878 | return 0; |
6c2a63a3 PB |
879 | } |
880 | ||
27a4cd48 DM |
881 | } // anon namespace |
882 | ||
883 | gimple_opt_pass * | |
884 | make_pass_cse_reciprocals (gcc::context *ctxt) | |
885 | { | |
886 | return new pass_cse_reciprocals (ctxt); | |
887 | } | |
888 | ||
88512ba0 | 889 | /* Records an occurrence at statement USE_STMT in the vector of trees |
2f397a93 | 890 | STMTS if it is dominated by *TOP_BB or dominates it or this basic block |
88512ba0 | 891 | is not yet initialized. Returns true if the occurrence was pushed on |
2f397a93 RG |
892 | the vector. Adjusts *TOP_BB to be the basic block dominating all |
893 | statements in the vector. */ | |
894 | ||
895 | static bool | |
355fe088 TS |
896 | maybe_record_sincos (vec<gimple *> *stmts, |
897 | basic_block *top_bb, gimple *use_stmt) | |
2f397a93 | 898 | { |
726a989a | 899 | basic_block use_bb = gimple_bb (use_stmt); |
2f397a93 RG |
900 | if (*top_bb |
901 | && (*top_bb == use_bb | |
902 | || dominated_by_p (CDI_DOMINATORS, use_bb, *top_bb))) | |
9771b263 | 903 | stmts->safe_push (use_stmt); |
2f397a93 RG |
904 | else if (!*top_bb |
905 | || dominated_by_p (CDI_DOMINATORS, *top_bb, use_bb)) | |
906 | { | |
9771b263 | 907 | stmts->safe_push (use_stmt); |
2f397a93 RG |
908 | *top_bb = use_bb; |
909 | } | |
910 | else | |
911 | return false; | |
912 | ||
913 | return true; | |
914 | } | |
915 | ||
916 | /* Look for sin, cos and cexpi calls with the same argument NAME and | |
917 | create a single call to cexpi CSEing the result in this case. | |
918 | We first walk over all immediate uses of the argument collecting | |
919 | statements that we can CSE in a vector and in a second pass replace | |
920 | the statement rhs with a REALPART or IMAGPART expression on the | |
921 | result of the cexpi call we insert before the use statement that | |
922 | dominates all other candidates. */ | |
923 | ||
90bc1cb8 | 924 | static bool |
2f397a93 RG |
925 | execute_cse_sincos_1 (tree name) |
926 | { | |
726a989a | 927 | gimple_stmt_iterator gsi; |
2f397a93 | 928 | imm_use_iterator use_iter; |
726a989a | 929 | tree fndecl, res, type; |
355fe088 | 930 | gimple *def_stmt, *use_stmt, *stmt; |
2f397a93 | 931 | int seen_cos = 0, seen_sin = 0, seen_cexpi = 0; |
355fe088 | 932 | auto_vec<gimple *> stmts; |
2f397a93 RG |
933 | basic_block top_bb = NULL; |
934 | int i; | |
90bc1cb8 | 935 | bool cfg_changed = false; |
2f397a93 RG |
936 | |
937 | type = TREE_TYPE (name); | |
938 | FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, name) | |
939 | { | |
726a989a | 940 | if (gimple_code (use_stmt) != GIMPLE_CALL |
c97d1c9d | 941 | || !gimple_call_lhs (use_stmt)) |
2f397a93 RG |
942 | continue; |
943 | ||
c97d1c9d | 944 | switch (gimple_call_combined_fn (use_stmt)) |
2f397a93 | 945 | { |
c97d1c9d | 946 | CASE_CFN_COS: |
2f397a93 RG |
947 | seen_cos |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0; |
948 | break; | |
949 | ||
c97d1c9d | 950 | CASE_CFN_SIN: |
2f397a93 RG |
951 | seen_sin |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0; |
952 | break; | |
953 | ||
c97d1c9d | 954 | CASE_CFN_CEXPI: |
2f397a93 RG |
955 | seen_cexpi |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0; |
956 | break; | |
957 | ||
958 | default:; | |
959 | } | |
960 | } | |
961 | ||
962 | if (seen_cos + seen_sin + seen_cexpi <= 1) | |
9370adeb | 963 | return false; |
2f397a93 RG |
964 | |
965 | /* Simply insert cexpi at the beginning of top_bb but not earlier than | |
966 | the name def statement. */ | |
967 | fndecl = mathfn_built_in (type, BUILT_IN_CEXPI); | |
968 | if (!fndecl) | |
90bc1cb8 | 969 | return false; |
726a989a | 970 | stmt = gimple_build_call (fndecl, 1, name); |
83d5977e | 971 | res = make_temp_ssa_name (TREE_TYPE (TREE_TYPE (fndecl)), stmt, "sincostmp"); |
726a989a RB |
972 | gimple_call_set_lhs (stmt, res); |
973 | ||
2f397a93 | 974 | def_stmt = SSA_NAME_DEF_STMT (name); |
59805c3b | 975 | if (!SSA_NAME_IS_DEFAULT_DEF (name) |
726a989a RB |
976 | && gimple_code (def_stmt) != GIMPLE_PHI |
977 | && gimple_bb (def_stmt) == top_bb) | |
2f397a93 | 978 | { |
726a989a RB |
979 | gsi = gsi_for_stmt (def_stmt); |
980 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); | |
2f397a93 RG |
981 | } |
982 | else | |
983 | { | |
726a989a RB |
984 | gsi = gsi_after_labels (top_bb); |
985 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
2f397a93 | 986 | } |
4da3b811 | 987 | sincos_stats.inserted++; |
2f397a93 RG |
988 | |
989 | /* And adjust the recorded old call sites. */ | |
9771b263 | 990 | for (i = 0; stmts.iterate (i, &use_stmt); ++i) |
2f397a93 | 991 | { |
726a989a | 992 | tree rhs = NULL; |
726a989a | 993 | |
c97d1c9d | 994 | switch (gimple_call_combined_fn (use_stmt)) |
2f397a93 | 995 | { |
c97d1c9d | 996 | CASE_CFN_COS: |
726a989a | 997 | rhs = fold_build1 (REALPART_EXPR, type, res); |
2f397a93 RG |
998 | break; |
999 | ||
c97d1c9d | 1000 | CASE_CFN_SIN: |
726a989a | 1001 | rhs = fold_build1 (IMAGPART_EXPR, type, res); |
2f397a93 RG |
1002 | break; |
1003 | ||
c97d1c9d | 1004 | CASE_CFN_CEXPI: |
726a989a | 1005 | rhs = res; |
2f397a93 RG |
1006 | break; |
1007 | ||
1008 | default:; | |
1009 | gcc_unreachable (); | |
1010 | } | |
1011 | ||
726a989a RB |
1012 | /* Replace call with a copy. */ |
1013 | stmt = gimple_build_assign (gimple_call_lhs (use_stmt), rhs); | |
1014 | ||
1015 | gsi = gsi_for_stmt (use_stmt); | |
90bc1cb8 RG |
1016 | gsi_replace (&gsi, stmt, true); |
1017 | if (gimple_purge_dead_eh_edges (gimple_bb (stmt))) | |
1018 | cfg_changed = true; | |
2f397a93 RG |
1019 | } |
1020 | ||
90bc1cb8 | 1021 | return cfg_changed; |
2f397a93 RG |
1022 | } |
1023 | ||
78be79d5 BS |
1024 | /* To evaluate powi(x,n), the floating point value x raised to the |
1025 | constant integer exponent n, we use a hybrid algorithm that | |
1026 | combines the "window method" with look-up tables. For an | |
1027 | introduction to exponentiation algorithms and "addition chains", | |
1028 | see section 4.6.3, "Evaluation of Powers" of Donald E. Knuth, | |
1029 | "Seminumerical Algorithms", Vol. 2, "The Art of Computer Programming", | |
1030 | 3rd Edition, 1998, and Daniel M. Gordon, "A Survey of Fast Exponentiation | |
1031 | Methods", Journal of Algorithms, Vol. 27, pp. 129-146, 1998. */ | |
1032 | ||
1033 | /* Provide a default value for POWI_MAX_MULTS, the maximum number of | |
1034 | multiplications to inline before calling the system library's pow | |
1035 | function. powi(x,n) requires at worst 2*bits(n)-2 multiplications, | |
1036 | so this default never requires calling pow, powf or powl. */ | |
1037 | ||
1038 | #ifndef POWI_MAX_MULTS | |
1039 | #define POWI_MAX_MULTS (2*HOST_BITS_PER_WIDE_INT-2) | |
1040 | #endif | |
1041 | ||
1042 | /* The size of the "optimal power tree" lookup table. All | |
1043 | exponents less than this value are simply looked up in the | |
1044 | powi_table below. This threshold is also used to size the | |
1045 | cache of pseudo registers that hold intermediate results. */ | |
1046 | #define POWI_TABLE_SIZE 256 | |
1047 | ||
1048 | /* The size, in bits of the window, used in the "window method" | |
1049 | exponentiation algorithm. This is equivalent to a radix of | |
1050 | (1<<POWI_WINDOW_SIZE) in the corresponding "m-ary method". */ | |
1051 | #define POWI_WINDOW_SIZE 3 | |
1052 | ||
1053 | /* The following table is an efficient representation of an | |
1054 | "optimal power tree". For each value, i, the corresponding | |
1055 | value, j, in the table states than an optimal evaluation | |
1056 | sequence for calculating pow(x,i) can be found by evaluating | |
1057 | pow(x,j)*pow(x,i-j). An optimal power tree for the first | |
1058 | 100 integers is given in Knuth's "Seminumerical algorithms". */ | |
1059 | ||
1060 | static const unsigned char powi_table[POWI_TABLE_SIZE] = | |
1061 | { | |
1062 | 0, 1, 1, 2, 2, 3, 3, 4, /* 0 - 7 */ | |
1063 | 4, 6, 5, 6, 6, 10, 7, 9, /* 8 - 15 */ | |
1064 | 8, 16, 9, 16, 10, 12, 11, 13, /* 16 - 23 */ | |
1065 | 12, 17, 13, 18, 14, 24, 15, 26, /* 24 - 31 */ | |
1066 | 16, 17, 17, 19, 18, 33, 19, 26, /* 32 - 39 */ | |
1067 | 20, 25, 21, 40, 22, 27, 23, 44, /* 40 - 47 */ | |
1068 | 24, 32, 25, 34, 26, 29, 27, 44, /* 48 - 55 */ | |
1069 | 28, 31, 29, 34, 30, 60, 31, 36, /* 56 - 63 */ | |
1070 | 32, 64, 33, 34, 34, 46, 35, 37, /* 64 - 71 */ | |
1071 | 36, 65, 37, 50, 38, 48, 39, 69, /* 72 - 79 */ | |
1072 | 40, 49, 41, 43, 42, 51, 43, 58, /* 80 - 87 */ | |
1073 | 44, 64, 45, 47, 46, 59, 47, 76, /* 88 - 95 */ | |
1074 | 48, 65, 49, 66, 50, 67, 51, 66, /* 96 - 103 */ | |
1075 | 52, 70, 53, 74, 54, 104, 55, 74, /* 104 - 111 */ | |
1076 | 56, 64, 57, 69, 58, 78, 59, 68, /* 112 - 119 */ | |
1077 | 60, 61, 61, 80, 62, 75, 63, 68, /* 120 - 127 */ | |
1078 | 64, 65, 65, 128, 66, 129, 67, 90, /* 128 - 135 */ | |
1079 | 68, 73, 69, 131, 70, 94, 71, 88, /* 136 - 143 */ | |
1080 | 72, 128, 73, 98, 74, 132, 75, 121, /* 144 - 151 */ | |
1081 | 76, 102, 77, 124, 78, 132, 79, 106, /* 152 - 159 */ | |
1082 | 80, 97, 81, 160, 82, 99, 83, 134, /* 160 - 167 */ | |
1083 | 84, 86, 85, 95, 86, 160, 87, 100, /* 168 - 175 */ | |
1084 | 88, 113, 89, 98, 90, 107, 91, 122, /* 176 - 183 */ | |
1085 | 92, 111, 93, 102, 94, 126, 95, 150, /* 184 - 191 */ | |
1086 | 96, 128, 97, 130, 98, 133, 99, 195, /* 192 - 199 */ | |
1087 | 100, 128, 101, 123, 102, 164, 103, 138, /* 200 - 207 */ | |
1088 | 104, 145, 105, 146, 106, 109, 107, 149, /* 208 - 215 */ | |
1089 | 108, 200, 109, 146, 110, 170, 111, 157, /* 216 - 223 */ | |
1090 | 112, 128, 113, 130, 114, 182, 115, 132, /* 224 - 231 */ | |
1091 | 116, 200, 117, 132, 118, 158, 119, 206, /* 232 - 239 */ | |
1092 | 120, 240, 121, 162, 122, 147, 123, 152, /* 240 - 247 */ | |
1093 | 124, 166, 125, 214, 126, 138, 127, 153, /* 248 - 255 */ | |
1094 | }; | |
1095 | ||
1096 | ||
1097 | /* Return the number of multiplications required to calculate | |
1098 | powi(x,n) where n is less than POWI_TABLE_SIZE. This is a | |
1099 | subroutine of powi_cost. CACHE is an array indicating | |
1100 | which exponents have already been calculated. */ | |
1101 | ||
1102 | static int | |
1103 | powi_lookup_cost (unsigned HOST_WIDE_INT n, bool *cache) | |
1104 | { | |
1105 | /* If we've already calculated this exponent, then this evaluation | |
1106 | doesn't require any additional multiplications. */ | |
1107 | if (cache[n]) | |
1108 | return 0; | |
1109 | ||
1110 | cache[n] = true; | |
1111 | return powi_lookup_cost (n - powi_table[n], cache) | |
1112 | + powi_lookup_cost (powi_table[n], cache) + 1; | |
1113 | } | |
1114 | ||
1115 | /* Return the number of multiplications required to calculate | |
1116 | powi(x,n) for an arbitrary x, given the exponent N. This | |
1117 | function needs to be kept in sync with powi_as_mults below. */ | |
1118 | ||
1119 | static int | |
1120 | powi_cost (HOST_WIDE_INT n) | |
1121 | { | |
1122 | bool cache[POWI_TABLE_SIZE]; | |
1123 | unsigned HOST_WIDE_INT digit; | |
1124 | unsigned HOST_WIDE_INT val; | |
1125 | int result; | |
1126 | ||
1127 | if (n == 0) | |
1128 | return 0; | |
1129 | ||
1130 | /* Ignore the reciprocal when calculating the cost. */ | |
1131 | val = (n < 0) ? -n : n; | |
1132 | ||
1133 | /* Initialize the exponent cache. */ | |
1134 | memset (cache, 0, POWI_TABLE_SIZE * sizeof (bool)); | |
1135 | cache[1] = true; | |
1136 | ||
1137 | result = 0; | |
1138 | ||
1139 | while (val >= POWI_TABLE_SIZE) | |
1140 | { | |
1141 | if (val & 1) | |
1142 | { | |
1143 | digit = val & ((1 << POWI_WINDOW_SIZE) - 1); | |
1144 | result += powi_lookup_cost (digit, cache) | |
1145 | + POWI_WINDOW_SIZE + 1; | |
1146 | val >>= POWI_WINDOW_SIZE; | |
1147 | } | |
1148 | else | |
1149 | { | |
1150 | val >>= 1; | |
1151 | result++; | |
1152 | } | |
1153 | } | |
1154 | ||
1155 | return result + powi_lookup_cost (val, cache); | |
1156 | } | |
1157 | ||
1158 | /* Recursive subroutine of powi_as_mults. This function takes the | |
1159 | array, CACHE, of already calculated exponents and an exponent N and | |
1160 | returns a tree that corresponds to CACHE[1]**N, with type TYPE. */ | |
1161 | ||
1162 | static tree | |
1163 | powi_as_mults_1 (gimple_stmt_iterator *gsi, location_t loc, tree type, | |
83d5977e | 1164 | HOST_WIDE_INT n, tree *cache) |
78be79d5 BS |
1165 | { |
1166 | tree op0, op1, ssa_target; | |
1167 | unsigned HOST_WIDE_INT digit; | |
538dd0b7 | 1168 | gassign *mult_stmt; |
78be79d5 BS |
1169 | |
1170 | if (n < POWI_TABLE_SIZE && cache[n]) | |
1171 | return cache[n]; | |
1172 | ||
83d5977e | 1173 | ssa_target = make_temp_ssa_name (type, NULL, "powmult"); |
78be79d5 BS |
1174 | |
1175 | if (n < POWI_TABLE_SIZE) | |
1176 | { | |
1177 | cache[n] = ssa_target; | |
83d5977e RG |
1178 | op0 = powi_as_mults_1 (gsi, loc, type, n - powi_table[n], cache); |
1179 | op1 = powi_as_mults_1 (gsi, loc, type, powi_table[n], cache); | |
78be79d5 BS |
1180 | } |
1181 | else if (n & 1) | |
1182 | { | |
1183 | digit = n & ((1 << POWI_WINDOW_SIZE) - 1); | |
83d5977e RG |
1184 | op0 = powi_as_mults_1 (gsi, loc, type, n - digit, cache); |
1185 | op1 = powi_as_mults_1 (gsi, loc, type, digit, cache); | |
78be79d5 BS |
1186 | } |
1187 | else | |
1188 | { | |
83d5977e | 1189 | op0 = powi_as_mults_1 (gsi, loc, type, n >> 1, cache); |
78be79d5 BS |
1190 | op1 = op0; |
1191 | } | |
1192 | ||
0d0e4a03 | 1193 | mult_stmt = gimple_build_assign (ssa_target, MULT_EXPR, op0, op1); |
ba869341 | 1194 | gimple_set_location (mult_stmt, loc); |
78be79d5 BS |
1195 | gsi_insert_before (gsi, mult_stmt, GSI_SAME_STMT); |
1196 | ||
1197 | return ssa_target; | |
1198 | } | |
1199 | ||
1200 | /* Convert ARG0**N to a tree of multiplications of ARG0 with itself. | |
1201 | This function needs to be kept in sync with powi_cost above. */ | |
1202 | ||
1203 | static tree | |
1204 | powi_as_mults (gimple_stmt_iterator *gsi, location_t loc, | |
1205 | tree arg0, HOST_WIDE_INT n) | |
1206 | { | |
83d5977e | 1207 | tree cache[POWI_TABLE_SIZE], result, type = TREE_TYPE (arg0); |
538dd0b7 | 1208 | gassign *div_stmt; |
83d5977e | 1209 | tree target; |
78be79d5 BS |
1210 | |
1211 | if (n == 0) | |
1212 | return build_real (type, dconst1); | |
1213 | ||
1214 | memset (cache, 0, sizeof (cache)); | |
1215 | cache[1] = arg0; | |
1216 | ||
83d5977e | 1217 | result = powi_as_mults_1 (gsi, loc, type, (n < 0) ? -n : n, cache); |
78be79d5 BS |
1218 | if (n >= 0) |
1219 | return result; | |
1220 | ||
1221 | /* If the original exponent was negative, reciprocate the result. */ | |
83d5977e | 1222 | target = make_temp_ssa_name (type, NULL, "powmult"); |
0d0e4a03 JJ |
1223 | div_stmt = gimple_build_assign (target, RDIV_EXPR, |
1224 | build_real (type, dconst1), result); | |
ba869341 | 1225 | gimple_set_location (div_stmt, loc); |
78be79d5 BS |
1226 | gsi_insert_before (gsi, div_stmt, GSI_SAME_STMT); |
1227 | ||
1228 | return target; | |
1229 | } | |
1230 | ||
1231 | /* ARG0 and N are the two arguments to a powi builtin in GSI with | |
1232 | location info LOC. If the arguments are appropriate, create an | |
1233 | equivalent sequence of statements prior to GSI using an optimal | |
1234 | number of multiplications, and return an expession holding the | |
1235 | result. */ | |
1236 | ||
1237 | static tree | |
1238 | gimple_expand_builtin_powi (gimple_stmt_iterator *gsi, location_t loc, | |
1239 | tree arg0, HOST_WIDE_INT n) | |
1240 | { | |
1241 | /* Avoid largest negative number. */ | |
1242 | if (n != -n | |
1243 | && ((n >= -1 && n <= 2) | |
1244 | || (optimize_function_for_speed_p (cfun) | |
1245 | && powi_cost (n) <= POWI_MAX_MULTS))) | |
1246 | return powi_as_mults (gsi, loc, arg0, n); | |
1247 | ||
1248 | return NULL_TREE; | |
1249 | } | |
1250 | ||
ba869341 | 1251 | /* Build a gimple call statement that calls FN with argument ARG. |
83d5977e | 1252 | Set the lhs of the call statement to a fresh SSA name. Insert the |
ba869341 BS |
1253 | statement prior to GSI's current position, and return the fresh |
1254 | SSA name. */ | |
1255 | ||
1256 | static tree | |
6e96f98a | 1257 | build_and_insert_call (gimple_stmt_iterator *gsi, location_t loc, |
83d5977e | 1258 | tree fn, tree arg) |
ba869341 | 1259 | { |
538dd0b7 | 1260 | gcall *call_stmt; |
ba869341 BS |
1261 | tree ssa_target; |
1262 | ||
ba869341 | 1263 | call_stmt = gimple_build_call (fn, 1, arg); |
83d5977e | 1264 | ssa_target = make_temp_ssa_name (TREE_TYPE (arg), NULL, "powroot"); |
ba869341 BS |
1265 | gimple_set_lhs (call_stmt, ssa_target); |
1266 | gimple_set_location (call_stmt, loc); | |
1267 | gsi_insert_before (gsi, call_stmt, GSI_SAME_STMT); | |
1268 | ||
1269 | return ssa_target; | |
1270 | } | |
1271 | ||
6e96f98a BS |
1272 | /* Build a gimple binary operation with the given CODE and arguments |
1273 | ARG0, ARG1, assigning the result to a new SSA name for variable | |
1274 | TARGET. Insert the statement prior to GSI's current position, and | |
1275 | return the fresh SSA name.*/ | |
1276 | ||
1277 | static tree | |
1278 | build_and_insert_binop (gimple_stmt_iterator *gsi, location_t loc, | |
83d5977e RG |
1279 | const char *name, enum tree_code code, |
1280 | tree arg0, tree arg1) | |
6e96f98a | 1281 | { |
83d5977e | 1282 | tree result = make_temp_ssa_name (TREE_TYPE (arg0), NULL, name); |
0d0e4a03 | 1283 | gassign *stmt = gimple_build_assign (result, code, arg0, arg1); |
6e96f98a BS |
1284 | gimple_set_location (stmt, loc); |
1285 | gsi_insert_before (gsi, stmt, GSI_SAME_STMT); | |
1286 | return result; | |
1287 | } | |
1288 | ||
d7e2a1c1 | 1289 | /* Build a gimple reference operation with the given CODE and argument |
83d5977e | 1290 | ARG, assigning the result to a new SSA name of TYPE with NAME. |
d7e2a1c1 BS |
1291 | Insert the statement prior to GSI's current position, and return |
1292 | the fresh SSA name. */ | |
1293 | ||
1294 | static inline tree | |
1295 | build_and_insert_ref (gimple_stmt_iterator *gsi, location_t loc, tree type, | |
83d5977e | 1296 | const char *name, enum tree_code code, tree arg0) |
d7e2a1c1 | 1297 | { |
83d5977e | 1298 | tree result = make_temp_ssa_name (type, NULL, name); |
355fe088 | 1299 | gimple *stmt = gimple_build_assign (result, build1 (code, type, arg0)); |
d7e2a1c1 BS |
1300 | gimple_set_location (stmt, loc); |
1301 | gsi_insert_before (gsi, stmt, GSI_SAME_STMT); | |
1302 | return result; | |
1303 | } | |
1304 | ||
83d5977e | 1305 | /* Build a gimple assignment to cast VAL to TYPE. Insert the statement |
5dfe80ba AS |
1306 | prior to GSI's current position, and return the fresh SSA name. */ |
1307 | ||
1308 | static tree | |
1309 | build_and_insert_cast (gimple_stmt_iterator *gsi, location_t loc, | |
83d5977e | 1310 | tree type, tree val) |
5dfe80ba | 1311 | { |
b731b390 | 1312 | tree result = make_ssa_name (type); |
0d0e4a03 | 1313 | gassign *stmt = gimple_build_assign (result, NOP_EXPR, val); |
83d5977e RG |
1314 | gimple_set_location (stmt, loc); |
1315 | gsi_insert_before (gsi, stmt, GSI_SAME_STMT); | |
1316 | return result; | |
5dfe80ba AS |
1317 | } |
1318 | ||
b7dce216 KT |
1319 | struct pow_synth_sqrt_info |
1320 | { | |
1321 | bool *factors; | |
1322 | unsigned int deepest; | |
1323 | unsigned int num_mults; | |
1324 | }; | |
1325 | ||
1326 | /* Return true iff the real value C can be represented as a | |
1327 | sum of powers of 0.5 up to N. That is: | |
1328 | C == SUM<i from 1..N> (a[i]*(0.5**i)) where a[i] is either 0 or 1. | |
1329 | Record in INFO the various parameters of the synthesis algorithm such | |
1330 | as the factors a[i], the maximum 0.5 power and the number of | |
1331 | multiplications that will be required. */ | |
1332 | ||
1333 | bool | |
1334 | representable_as_half_series_p (REAL_VALUE_TYPE c, unsigned n, | |
1335 | struct pow_synth_sqrt_info *info) | |
1336 | { | |
1337 | REAL_VALUE_TYPE factor = dconsthalf; | |
1338 | REAL_VALUE_TYPE remainder = c; | |
1339 | ||
1340 | info->deepest = 0; | |
1341 | info->num_mults = 0; | |
1342 | memset (info->factors, 0, n * sizeof (bool)); | |
1343 | ||
1344 | for (unsigned i = 0; i < n; i++) | |
1345 | { | |
1346 | REAL_VALUE_TYPE res; | |
1347 | ||
1348 | /* If something inexact happened bail out now. */ | |
5c88ea94 | 1349 | if (real_arithmetic (&res, MINUS_EXPR, &remainder, &factor)) |
b7dce216 KT |
1350 | return false; |
1351 | ||
1352 | /* We have hit zero. The number is representable as a sum | |
1353 | of powers of 0.5. */ | |
624d31fe | 1354 | if (real_equal (&res, &dconst0)) |
b7dce216 KT |
1355 | { |
1356 | info->factors[i] = true; | |
1357 | info->deepest = i + 1; | |
1358 | return true; | |
1359 | } | |
1360 | else if (!REAL_VALUE_NEGATIVE (res)) | |
1361 | { | |
1362 | remainder = res; | |
1363 | info->factors[i] = true; | |
1364 | info->num_mults++; | |
1365 | } | |
1366 | else | |
1367 | info->factors[i] = false; | |
1368 | ||
5c88ea94 | 1369 | real_arithmetic (&factor, MULT_EXPR, &factor, &dconsthalf); |
b7dce216 KT |
1370 | } |
1371 | return false; | |
1372 | } | |
1373 | ||
1374 | /* Return the tree corresponding to FN being applied | |
1375 | to ARG N times at GSI and LOC. | |
1376 | Look up previous results from CACHE if need be. | |
1377 | cache[0] should contain just plain ARG i.e. FN applied to ARG 0 times. */ | |
1378 | ||
1379 | static tree | |
1380 | get_fn_chain (tree arg, unsigned int n, gimple_stmt_iterator *gsi, | |
1381 | tree fn, location_t loc, tree *cache) | |
1382 | { | |
1383 | tree res = cache[n]; | |
1384 | if (!res) | |
1385 | { | |
1386 | tree prev = get_fn_chain (arg, n - 1, gsi, fn, loc, cache); | |
1387 | res = build_and_insert_call (gsi, loc, fn, prev); | |
1388 | cache[n] = res; | |
1389 | } | |
1390 | ||
1391 | return res; | |
1392 | } | |
1393 | ||
1394 | /* Print to STREAM the repeated application of function FNAME to ARG | |
1395 | N times. So, for FNAME = "foo", ARG = "x", N = 2 it would print: | |
1396 | "foo (foo (x))". */ | |
1397 | ||
1398 | static void | |
1399 | print_nested_fn (FILE* stream, const char *fname, const char* arg, | |
1400 | unsigned int n) | |
1401 | { | |
1402 | if (n == 0) | |
1403 | fprintf (stream, "%s", arg); | |
1404 | else | |
1405 | { | |
1406 | fprintf (stream, "%s (", fname); | |
1407 | print_nested_fn (stream, fname, arg, n - 1); | |
1408 | fprintf (stream, ")"); | |
1409 | } | |
1410 | } | |
1411 | ||
1412 | /* Print to STREAM the fractional sequence of sqrt chains | |
1413 | applied to ARG, described by INFO. Used for the dump file. */ | |
1414 | ||
1415 | static void | |
1416 | dump_fractional_sqrt_sequence (FILE *stream, const char *arg, | |
1417 | struct pow_synth_sqrt_info *info) | |
1418 | { | |
1419 | for (unsigned int i = 0; i < info->deepest; i++) | |
1420 | { | |
1421 | bool is_set = info->factors[i]; | |
1422 | if (is_set) | |
1423 | { | |
1424 | print_nested_fn (stream, "sqrt", arg, i + 1); | |
1425 | if (i != info->deepest - 1) | |
1426 | fprintf (stream, " * "); | |
1427 | } | |
1428 | } | |
1429 | } | |
1430 | ||
1431 | /* Print to STREAM a representation of raising ARG to an integer | |
1432 | power N. Used for the dump file. */ | |
1433 | ||
1434 | static void | |
1435 | dump_integer_part (FILE *stream, const char* arg, HOST_WIDE_INT n) | |
1436 | { | |
1437 | if (n > 1) | |
1438 | fprintf (stream, "powi (%s, " HOST_WIDE_INT_PRINT_DEC ")", arg, n); | |
1439 | else if (n == 1) | |
1440 | fprintf (stream, "%s", arg); | |
1441 | } | |
1442 | ||
1443 | /* Attempt to synthesize a POW[F] (ARG0, ARG1) call using chains of | |
1444 | square roots. Place at GSI and LOC. Limit the maximum depth | |
1445 | of the sqrt chains to MAX_DEPTH. Return the tree holding the | |
1446 | result of the expanded sequence or NULL_TREE if the expansion failed. | |
1447 | ||
1448 | This routine assumes that ARG1 is a real number with a fractional part | |
1449 | (the integer exponent case will have been handled earlier in | |
1450 | gimple_expand_builtin_pow). | |
1451 | ||
1452 | For ARG1 > 0.0: | |
1453 | * For ARG1 composed of a whole part WHOLE_PART and a fractional part | |
1454 | FRAC_PART i.e. WHOLE_PART == floor (ARG1) and | |
1455 | FRAC_PART == ARG1 - WHOLE_PART: | |
1456 | Produce POWI (ARG0, WHOLE_PART) * POW (ARG0, FRAC_PART) where | |
1457 | POW (ARG0, FRAC_PART) is expanded as a product of square root chains | |
1458 | if it can be expressed as such, that is if FRAC_PART satisfies: | |
1459 | FRAC_PART == <SUM from i = 1 until MAX_DEPTH> (a[i] * (0.5**i)) | |
1460 | where integer a[i] is either 0 or 1. | |
1461 | ||
1462 | Example: | |
1463 | POW (x, 3.625) == POWI (x, 3) * POW (x, 0.625) | |
1464 | --> POWI (x, 3) * SQRT (x) * SQRT (SQRT (SQRT (x))) | |
1465 | ||
1466 | For ARG1 < 0.0 there are two approaches: | |
1467 | * (A) Expand to 1.0 / POW (ARG0, -ARG1) where POW (ARG0, -ARG1) | |
1468 | is calculated as above. | |
1469 | ||
1470 | Example: | |
1471 | POW (x, -5.625) == 1.0 / POW (x, 5.625) | |
1472 | --> 1.0 / (POWI (x, 5) * SQRT (x) * SQRT (SQRT (SQRT (x)))) | |
1473 | ||
1474 | * (B) : WHOLE_PART := - ceil (abs (ARG1)) | |
1475 | FRAC_PART := ARG1 - WHOLE_PART | |
1476 | and expand to POW (x, FRAC_PART) / POWI (x, WHOLE_PART). | |
1477 | Example: | |
1478 | POW (x, -5.875) == POW (x, 0.125) / POWI (X, 6) | |
1479 | --> SQRT (SQRT (SQRT (x))) / (POWI (x, 6)) | |
1480 | ||
1481 | For ARG1 < 0.0 we choose between (A) and (B) depending on | |
1482 | how many multiplications we'd have to do. | |
1483 | So, for the example in (B): POW (x, -5.875), if we were to | |
1484 | follow algorithm (A) we would produce: | |
1485 | 1.0 / POWI (X, 5) * SQRT (X) * SQRT (SQRT (X)) * SQRT (SQRT (SQRT (X))) | |
1486 | which contains more multiplications than approach (B). | |
1487 | ||
1488 | Hopefully, this approach will eliminate potentially expensive POW library | |
1489 | calls when unsafe floating point math is enabled and allow the compiler to | |
1490 | further optimise the multiplies, square roots and divides produced by this | |
1491 | function. */ | |
1492 | ||
1493 | static tree | |
1494 | expand_pow_as_sqrts (gimple_stmt_iterator *gsi, location_t loc, | |
1495 | tree arg0, tree arg1, HOST_WIDE_INT max_depth) | |
1496 | { | |
1497 | tree type = TREE_TYPE (arg0); | |
1498 | machine_mode mode = TYPE_MODE (type); | |
1499 | tree sqrtfn = mathfn_built_in (type, BUILT_IN_SQRT); | |
1500 | bool one_over = true; | |
1501 | ||
1502 | if (!sqrtfn) | |
1503 | return NULL_TREE; | |
1504 | ||
1505 | if (TREE_CODE (arg1) != REAL_CST) | |
1506 | return NULL_TREE; | |
1507 | ||
1508 | REAL_VALUE_TYPE exp_init = TREE_REAL_CST (arg1); | |
1509 | ||
1510 | gcc_assert (max_depth > 0); | |
1511 | tree *cache = XALLOCAVEC (tree, max_depth + 1); | |
1512 | ||
1513 | struct pow_synth_sqrt_info synth_info; | |
1514 | synth_info.factors = XALLOCAVEC (bool, max_depth + 1); | |
1515 | synth_info.deepest = 0; | |
1516 | synth_info.num_mults = 0; | |
1517 | ||
1518 | bool neg_exp = REAL_VALUE_NEGATIVE (exp_init); | |
1519 | REAL_VALUE_TYPE exp = real_value_abs (&exp_init); | |
1520 | ||
1521 | /* The whole and fractional parts of exp. */ | |
1522 | REAL_VALUE_TYPE whole_part; | |
1523 | REAL_VALUE_TYPE frac_part; | |
1524 | ||
1525 | real_floor (&whole_part, mode, &exp); | |
5c88ea94 | 1526 | real_arithmetic (&frac_part, MINUS_EXPR, &exp, &whole_part); |
b7dce216 KT |
1527 | |
1528 | ||
1529 | REAL_VALUE_TYPE ceil_whole = dconst0; | |
1530 | REAL_VALUE_TYPE ceil_fract = dconst0; | |
1531 | ||
1532 | if (neg_exp) | |
1533 | { | |
1534 | real_ceil (&ceil_whole, mode, &exp); | |
5c88ea94 | 1535 | real_arithmetic (&ceil_fract, MINUS_EXPR, &ceil_whole, &exp); |
b7dce216 KT |
1536 | } |
1537 | ||
1538 | if (!representable_as_half_series_p (frac_part, max_depth, &synth_info)) | |
1539 | return NULL_TREE; | |
1540 | ||
1541 | /* Check whether it's more profitable to not use 1.0 / ... */ | |
1542 | if (neg_exp) | |
1543 | { | |
1544 | struct pow_synth_sqrt_info alt_synth_info; | |
1545 | alt_synth_info.factors = XALLOCAVEC (bool, max_depth + 1); | |
1546 | alt_synth_info.deepest = 0; | |
1547 | alt_synth_info.num_mults = 0; | |
1548 | ||
1549 | if (representable_as_half_series_p (ceil_fract, max_depth, | |
1550 | &alt_synth_info) | |
1551 | && alt_synth_info.deepest <= synth_info.deepest | |
1552 | && alt_synth_info.num_mults < synth_info.num_mults) | |
1553 | { | |
1554 | whole_part = ceil_whole; | |
1555 | frac_part = ceil_fract; | |
1556 | synth_info.deepest = alt_synth_info.deepest; | |
1557 | synth_info.num_mults = alt_synth_info.num_mults; | |
1558 | memcpy (synth_info.factors, alt_synth_info.factors, | |
1559 | (max_depth + 1) * sizeof (bool)); | |
1560 | one_over = false; | |
1561 | } | |
1562 | } | |
1563 | ||
1564 | HOST_WIDE_INT n = real_to_integer (&whole_part); | |
1565 | REAL_VALUE_TYPE cint; | |
1566 | real_from_integer (&cint, VOIDmode, n, SIGNED); | |
1567 | ||
1568 | if (!real_identical (&whole_part, &cint)) | |
1569 | return NULL_TREE; | |
1570 | ||
1571 | if (powi_cost (n) + synth_info.num_mults > POWI_MAX_MULTS) | |
1572 | return NULL_TREE; | |
1573 | ||
1574 | memset (cache, 0, (max_depth + 1) * sizeof (tree)); | |
1575 | ||
1576 | tree integer_res = n == 0 ? build_real (type, dconst1) : arg0; | |
1577 | ||
1578 | /* Calculate the integer part of the exponent. */ | |
1579 | if (n > 1) | |
1580 | { | |
1581 | integer_res = gimple_expand_builtin_powi (gsi, loc, arg0, n); | |
1582 | if (!integer_res) | |
1583 | return NULL_TREE; | |
1584 | } | |
1585 | ||
1586 | if (dump_file) | |
1587 | { | |
1588 | char string[64]; | |
1589 | ||
1590 | real_to_decimal (string, &exp_init, sizeof (string), 0, 1); | |
1591 | fprintf (dump_file, "synthesizing pow (x, %s) as:\n", string); | |
1592 | ||
1593 | if (neg_exp) | |
1594 | { | |
1595 | if (one_over) | |
1596 | { | |
1597 | fprintf (dump_file, "1.0 / ("); | |
1598 | dump_integer_part (dump_file, "x", n); | |
1599 | if (n > 0) | |
1600 | fprintf (dump_file, " * "); | |
1601 | dump_fractional_sqrt_sequence (dump_file, "x", &synth_info); | |
1602 | fprintf (dump_file, ")"); | |
1603 | } | |
1604 | else | |
1605 | { | |
1606 | dump_fractional_sqrt_sequence (dump_file, "x", &synth_info); | |
1607 | fprintf (dump_file, " / ("); | |
1608 | dump_integer_part (dump_file, "x", n); | |
1609 | fprintf (dump_file, ")"); | |
1610 | } | |
1611 | } | |
1612 | else | |
1613 | { | |
1614 | dump_fractional_sqrt_sequence (dump_file, "x", &synth_info); | |
1615 | if (n > 0) | |
1616 | fprintf (dump_file, " * "); | |
1617 | dump_integer_part (dump_file, "x", n); | |
1618 | } | |
1619 | ||
1620 | fprintf (dump_file, "\ndeepest sqrt chain: %d\n", synth_info.deepest); | |
1621 | } | |
1622 | ||
1623 | ||
1624 | tree fract_res = NULL_TREE; | |
1625 | cache[0] = arg0; | |
1626 | ||
1627 | /* Calculate the fractional part of the exponent. */ | |
1628 | for (unsigned i = 0; i < synth_info.deepest; i++) | |
1629 | { | |
1630 | if (synth_info.factors[i]) | |
1631 | { | |
1632 | tree sqrt_chain = get_fn_chain (arg0, i + 1, gsi, sqrtfn, loc, cache); | |
1633 | ||
1634 | if (!fract_res) | |
1635 | fract_res = sqrt_chain; | |
1636 | ||
1637 | else | |
1638 | fract_res = build_and_insert_binop (gsi, loc, "powroot", MULT_EXPR, | |
1639 | fract_res, sqrt_chain); | |
1640 | } | |
1641 | } | |
1642 | ||
1643 | tree res = NULL_TREE; | |
1644 | ||
1645 | if (neg_exp) | |
1646 | { | |
1647 | if (one_over) | |
1648 | { | |
1649 | if (n > 0) | |
1650 | res = build_and_insert_binop (gsi, loc, "powroot", MULT_EXPR, | |
1651 | fract_res, integer_res); | |
1652 | else | |
1653 | res = fract_res; | |
1654 | ||
1655 | res = build_and_insert_binop (gsi, loc, "powrootrecip", RDIV_EXPR, | |
1656 | build_real (type, dconst1), res); | |
1657 | } | |
1658 | else | |
1659 | { | |
1660 | res = build_and_insert_binop (gsi, loc, "powroot", RDIV_EXPR, | |
1661 | fract_res, integer_res); | |
1662 | } | |
1663 | } | |
1664 | else | |
1665 | res = build_and_insert_binop (gsi, loc, "powroot", MULT_EXPR, | |
1666 | fract_res, integer_res); | |
1667 | return res; | |
1668 | } | |
1669 | ||
d24ad7d6 BS |
1670 | /* ARG0 and ARG1 are the two arguments to a pow builtin call in GSI |
1671 | with location info LOC. If possible, create an equivalent and | |
1672 | less expensive sequence of statements prior to GSI, and return an | |
1673 | expession holding the result. */ | |
1674 | ||
1675 | static tree | |
1676 | gimple_expand_builtin_pow (gimple_stmt_iterator *gsi, location_t loc, | |
1677 | tree arg0, tree arg1) | |
1678 | { | |
b7dce216 | 1679 | REAL_VALUE_TYPE c, cint, dconst1_3, dconst1_4, dconst1_6; |
6e96f98a | 1680 | REAL_VALUE_TYPE c2, dconst3; |
d24ad7d6 | 1681 | HOST_WIDE_INT n; |
b7dce216 | 1682 | tree type, sqrtfn, cbrtfn, sqrt_arg0, result, cbrt_x, powi_cbrt_x; |
ef4bddc2 | 1683 | machine_mode mode; |
b7dce216 | 1684 | bool speed_p = optimize_bb_for_speed_p (gsi_bb (*gsi)); |
0bfbca58 | 1685 | bool hw_sqrt_exists, c_is_int, c2_is_int; |
d24ad7d6 | 1686 | |
b7dce216 KT |
1687 | dconst1_4 = dconst1; |
1688 | SET_REAL_EXP (&dconst1_4, REAL_EXP (&dconst1_4) - 2); | |
1689 | ||
d24ad7d6 BS |
1690 | /* If the exponent isn't a constant, there's nothing of interest |
1691 | to be done. */ | |
1692 | if (TREE_CODE (arg1) != REAL_CST) | |
1693 | return NULL_TREE; | |
1694 | ||
5a00b0aa SS |
1695 | /* Don't perform the operation if flag_signaling_nans is on |
1696 | and the operand is a signaling NaN. */ | |
1697 | if (HONOR_SNANS (TYPE_MODE (TREE_TYPE (arg1))) | |
942a1319 JJ |
1698 | && ((TREE_CODE (arg0) == REAL_CST |
1699 | && REAL_VALUE_ISSIGNALING_NAN (TREE_REAL_CST (arg0))) | |
5a00b0aa SS |
1700 | || REAL_VALUE_ISSIGNALING_NAN (TREE_REAL_CST (arg1)))) |
1701 | return NULL_TREE; | |
1702 | ||
ba869341 BS |
1703 | /* If the exponent is equivalent to an integer, expand to an optimal |
1704 | multiplication sequence when profitable. */ | |
d24ad7d6 BS |
1705 | c = TREE_REAL_CST (arg1); |
1706 | n = real_to_integer (&c); | |
807e902e | 1707 | real_from_integer (&cint, VOIDmode, n, SIGNED); |
0bfbca58 | 1708 | c_is_int = real_identical (&c, &cint); |
d24ad7d6 | 1709 | |
0bfbca58 | 1710 | if (c_is_int |
d24ad7d6 BS |
1711 | && ((n >= -1 && n <= 2) |
1712 | || (flag_unsafe_math_optimizations | |
b7dce216 | 1713 | && speed_p |
d24ad7d6 BS |
1714 | && powi_cost (n) <= POWI_MAX_MULTS))) |
1715 | return gimple_expand_builtin_powi (gsi, loc, arg0, n); | |
1716 | ||
ba869341 BS |
1717 | /* Attempt various optimizations using sqrt and cbrt. */ |
1718 | type = TREE_TYPE (arg0); | |
1719 | mode = TYPE_MODE (type); | |
1720 | sqrtfn = mathfn_built_in (type, BUILT_IN_SQRT); | |
1721 | ||
1722 | /* Optimize pow(x,0.5) = sqrt(x). This replacement is always safe | |
1723 | unless signed zeros must be maintained. pow(-0,0.5) = +0, while | |
1724 | sqrt(-0) = -0. */ | |
1725 | if (sqrtfn | |
624d31fe | 1726 | && real_equal (&c, &dconsthalf) |
ba869341 | 1727 | && !HONOR_SIGNED_ZEROS (mode)) |
83d5977e | 1728 | return build_and_insert_call (gsi, loc, sqrtfn, arg0); |
ba869341 | 1729 | |
d7e2a1c1 | 1730 | hw_sqrt_exists = optab_handler (sqrt_optab, mode) != CODE_FOR_nothing; |
ba869341 | 1731 | |
ba869341 BS |
1732 | /* Optimize pow(x,1./3.) = cbrt(x). This requires unsafe math |
1733 | optimizations since 1./3. is not exactly representable. If x | |
1734 | is negative and finite, the correct value of pow(x,1./3.) is | |
1735 | a NaN with the "invalid" exception raised, because the value | |
1736 | of 1./3. actually has an even denominator. The correct value | |
1737 | of cbrt(x) is a negative real value. */ | |
1738 | cbrtfn = mathfn_built_in (type, BUILT_IN_CBRT); | |
1739 | dconst1_3 = real_value_truncate (mode, dconst_third ()); | |
1740 | ||
1741 | if (flag_unsafe_math_optimizations | |
1742 | && cbrtfn | |
68e57f04 | 1743 | && (!HONOR_NANS (mode) || tree_expr_nonnegative_p (arg0)) |
624d31fe | 1744 | && real_equal (&c, &dconst1_3)) |
83d5977e | 1745 | return build_and_insert_call (gsi, loc, cbrtfn, arg0); |
ba869341 BS |
1746 | |
1747 | /* Optimize pow(x,1./6.) = cbrt(sqrt(x)). Don't do this optimization | |
1748 | if we don't have a hardware sqrt insn. */ | |
1749 | dconst1_6 = dconst1_3; | |
1750 | SET_REAL_EXP (&dconst1_6, REAL_EXP (&dconst1_6) - 1); | |
1751 | ||
1752 | if (flag_unsafe_math_optimizations | |
1753 | && sqrtfn | |
1754 | && cbrtfn | |
68e57f04 | 1755 | && (!HONOR_NANS (mode) || tree_expr_nonnegative_p (arg0)) |
b7dce216 | 1756 | && speed_p |
ba869341 | 1757 | && hw_sqrt_exists |
624d31fe | 1758 | && real_equal (&c, &dconst1_6)) |
ba869341 BS |
1759 | { |
1760 | /* sqrt(x) */ | |
83d5977e | 1761 | sqrt_arg0 = build_and_insert_call (gsi, loc, sqrtfn, arg0); |
ba869341 BS |
1762 | |
1763 | /* cbrt(sqrt(x)) */ | |
83d5977e | 1764 | return build_and_insert_call (gsi, loc, cbrtfn, sqrt_arg0); |
6e96f98a BS |
1765 | } |
1766 | ||
6e96f98a | 1767 | |
b7dce216 KT |
1768 | /* Attempt to expand the POW as a product of square root chains. |
1769 | Expand the 0.25 case even when otpimising for size. */ | |
6e96f98a BS |
1770 | if (flag_unsafe_math_optimizations |
1771 | && sqrtfn | |
b7dce216 | 1772 | && hw_sqrt_exists |
624d31fe | 1773 | && (speed_p || real_equal (&c, &dconst1_4)) |
b7dce216 | 1774 | && !HONOR_SIGNED_ZEROS (mode)) |
6e96f98a | 1775 | { |
b7dce216 KT |
1776 | unsigned int max_depth = speed_p |
1777 | ? PARAM_VALUE (PARAM_MAX_POW_SQRT_DEPTH) | |
1778 | : 2; | |
6e96f98a | 1779 | |
b7dce216 KT |
1780 | tree expand_with_sqrts |
1781 | = expand_pow_as_sqrts (gsi, loc, arg0, arg1, max_depth); | |
6e96f98a | 1782 | |
b7dce216 KT |
1783 | if (expand_with_sqrts) |
1784 | return expand_with_sqrts; | |
6e96f98a BS |
1785 | } |
1786 | ||
b7dce216 KT |
1787 | real_arithmetic (&c2, MULT_EXPR, &c, &dconst2); |
1788 | n = real_to_integer (&c2); | |
1789 | real_from_integer (&cint, VOIDmode, n, SIGNED); | |
1790 | c2_is_int = real_identical (&c2, &cint); | |
1791 | ||
6e96f98a BS |
1792 | /* Optimize pow(x,c), where 3c = n for some nonzero integer n, into |
1793 | ||
1794 | powi(x, n/3) * powi(cbrt(x), n%3), n > 0; | |
1795 | 1.0 / (powi(x, abs(n)/3) * powi(cbrt(x), abs(n)%3)), n < 0. | |
1796 | ||
1797 | Do not calculate the first factor when n/3 = 0. As cbrt(x) is | |
1798 | different from pow(x, 1./3.) due to rounding and behavior with | |
1799 | negative x, we need to constrain this transformation to unsafe | |
1800 | math and positive x or finite math. */ | |
807e902e | 1801 | real_from_integer (&dconst3, VOIDmode, 3, SIGNED); |
6e96f98a BS |
1802 | real_arithmetic (&c2, MULT_EXPR, &c, &dconst3); |
1803 | real_round (&c2, mode, &c2); | |
1804 | n = real_to_integer (&c2); | |
807e902e | 1805 | real_from_integer (&cint, VOIDmode, n, SIGNED); |
6e96f98a BS |
1806 | real_arithmetic (&c2, RDIV_EXPR, &cint, &dconst3); |
1807 | real_convert (&c2, mode, &c2); | |
1808 | ||
1809 | if (flag_unsafe_math_optimizations | |
1810 | && cbrtfn | |
68e57f04 | 1811 | && (!HONOR_NANS (mode) || tree_expr_nonnegative_p (arg0)) |
6e96f98a | 1812 | && real_identical (&c2, &c) |
0bfbca58 | 1813 | && !c2_is_int |
6e96f98a BS |
1814 | && optimize_function_for_speed_p (cfun) |
1815 | && powi_cost (n / 3) <= POWI_MAX_MULTS) | |
1816 | { | |
1817 | tree powi_x_ndiv3 = NULL_TREE; | |
1818 | ||
1819 | /* Attempt to fold powi(arg0, abs(n/3)) into multiplies. If not | |
1820 | possible or profitable, give up. Skip the degenerate case when | |
1821 | abs(n) < 3, where the result is always 1. */ | |
4c9cf7af | 1822 | if (absu_hwi (n) >= 3) |
6e96f98a BS |
1823 | { |
1824 | powi_x_ndiv3 = gimple_expand_builtin_powi (gsi, loc, arg0, | |
9f813990 | 1825 | abs_hwi (n / 3)); |
6e96f98a BS |
1826 | if (!powi_x_ndiv3) |
1827 | return NULL_TREE; | |
1828 | } | |
1829 | ||
1830 | /* Calculate powi(cbrt(x), n%3). Don't use gimple_expand_builtin_powi | |
1831 | as that creates an unnecessary variable. Instead, just produce | |
1832 | either cbrt(x) or cbrt(x) * cbrt(x). */ | |
83d5977e | 1833 | cbrt_x = build_and_insert_call (gsi, loc, cbrtfn, arg0); |
6e96f98a | 1834 | |
4c9cf7af | 1835 | if (absu_hwi (n) % 3 == 1) |
6e96f98a BS |
1836 | powi_cbrt_x = cbrt_x; |
1837 | else | |
83d5977e | 1838 | powi_cbrt_x = build_and_insert_binop (gsi, loc, "powroot", MULT_EXPR, |
6e96f98a BS |
1839 | cbrt_x, cbrt_x); |
1840 | ||
1841 | /* Multiply the two subexpressions, unless powi(x,abs(n)/3) = 1. */ | |
4c9cf7af | 1842 | if (absu_hwi (n) < 3) |
6e96f98a BS |
1843 | result = powi_cbrt_x; |
1844 | else | |
83d5977e | 1845 | result = build_and_insert_binop (gsi, loc, "powroot", MULT_EXPR, |
6e96f98a BS |
1846 | powi_x_ndiv3, powi_cbrt_x); |
1847 | ||
1848 | /* If n is negative, reciprocate the result. */ | |
1849 | if (n < 0) | |
83d5977e | 1850 | result = build_and_insert_binop (gsi, loc, "powroot", RDIV_EXPR, |
6e96f98a BS |
1851 | build_real (type, dconst1), result); |
1852 | ||
1853 | return result; | |
ba869341 BS |
1854 | } |
1855 | ||
6e96f98a | 1856 | /* No optimizations succeeded. */ |
d24ad7d6 BS |
1857 | return NULL_TREE; |
1858 | } | |
1859 | ||
d7e2a1c1 BS |
1860 | /* ARG is the argument to a cabs builtin call in GSI with location info |
1861 | LOC. Create a sequence of statements prior to GSI that calculates | |
1862 | sqrt(R*R + I*I), where R and I are the real and imaginary components | |
1863 | of ARG, respectively. Return an expression holding the result. */ | |
1864 | ||
1865 | static tree | |
1866 | gimple_expand_builtin_cabs (gimple_stmt_iterator *gsi, location_t loc, tree arg) | |
1867 | { | |
83d5977e | 1868 | tree real_part, imag_part, addend1, addend2, sum, result; |
d7e2a1c1 BS |
1869 | tree type = TREE_TYPE (TREE_TYPE (arg)); |
1870 | tree sqrtfn = mathfn_built_in (type, BUILT_IN_SQRT); | |
ef4bddc2 | 1871 | machine_mode mode = TYPE_MODE (type); |
d7e2a1c1 BS |
1872 | |
1873 | if (!flag_unsafe_math_optimizations | |
1874 | || !optimize_bb_for_speed_p (gimple_bb (gsi_stmt (*gsi))) | |
1875 | || !sqrtfn | |
1876 | || optab_handler (sqrt_optab, mode) == CODE_FOR_nothing) | |
1877 | return NULL_TREE; | |
1878 | ||
83d5977e | 1879 | real_part = build_and_insert_ref (gsi, loc, type, "cabs", |
d7e2a1c1 | 1880 | REALPART_EXPR, arg); |
83d5977e | 1881 | addend1 = build_and_insert_binop (gsi, loc, "cabs", MULT_EXPR, |
d7e2a1c1 | 1882 | real_part, real_part); |
83d5977e | 1883 | imag_part = build_and_insert_ref (gsi, loc, type, "cabs", |
d7e2a1c1 | 1884 | IMAGPART_EXPR, arg); |
83d5977e | 1885 | addend2 = build_and_insert_binop (gsi, loc, "cabs", MULT_EXPR, |
d7e2a1c1 | 1886 | imag_part, imag_part); |
83d5977e RG |
1887 | sum = build_and_insert_binop (gsi, loc, "cabs", PLUS_EXPR, addend1, addend2); |
1888 | result = build_and_insert_call (gsi, loc, sqrtfn, sum); | |
d7e2a1c1 BS |
1889 | |
1890 | return result; | |
1891 | } | |
1892 | ||
2f397a93 | 1893 | /* Go through all calls to sin, cos and cexpi and call execute_cse_sincos_1 |
78be79d5 BS |
1894 | on the SSA_NAME argument of each of them. Also expand powi(x,n) into |
1895 | an optimal number of multiplies, when n is a constant. */ | |
2f397a93 | 1896 | |
be55bfe6 TS |
1897 | namespace { |
1898 | ||
1899 | const pass_data pass_data_cse_sincos = | |
1900 | { | |
1901 | GIMPLE_PASS, /* type */ | |
1902 | "sincos", /* name */ | |
1903 | OPTGROUP_NONE, /* optinfo_flags */ | |
1b6546cc | 1904 | TV_TREE_SINCOS, /* tv_id */ |
be55bfe6 | 1905 | PROP_ssa, /* properties_required */ |
53f3cd25 | 1906 | PROP_gimple_opt_math, /* properties_provided */ |
be55bfe6 TS |
1907 | 0, /* properties_destroyed */ |
1908 | 0, /* todo_flags_start */ | |
3bea341f | 1909 | TODO_update_ssa, /* todo_flags_finish */ |
be55bfe6 TS |
1910 | }; |
1911 | ||
1912 | class pass_cse_sincos : public gimple_opt_pass | |
1913 | { | |
1914 | public: | |
1915 | pass_cse_sincos (gcc::context *ctxt) | |
1916 | : gimple_opt_pass (pass_data_cse_sincos, ctxt) | |
1917 | {} | |
1918 | ||
1919 | /* opt_pass methods: */ | |
1920 | virtual bool gate (function *) | |
1921 | { | |
1922 | /* We no longer require either sincos or cexp, since powi expansion | |
1923 | piggybacks on this pass. */ | |
1924 | return optimize; | |
1925 | } | |
1926 | ||
1927 | virtual unsigned int execute (function *); | |
1928 | ||
1929 | }; // class pass_cse_sincos | |
1930 | ||
1931 | unsigned int | |
1932 | pass_cse_sincos::execute (function *fun) | |
2f397a93 RG |
1933 | { |
1934 | basic_block bb; | |
90bc1cb8 | 1935 | bool cfg_changed = false; |
2f397a93 RG |
1936 | |
1937 | calculate_dominance_info (CDI_DOMINATORS); | |
4da3b811 | 1938 | memset (&sincos_stats, 0, sizeof (sincos_stats)); |
2f397a93 | 1939 | |
be55bfe6 | 1940 | FOR_EACH_BB_FN (bb, fun) |
2f397a93 | 1941 | { |
726a989a | 1942 | gimple_stmt_iterator gsi; |
3b9ee1cc | 1943 | bool cleanup_eh = false; |
2f397a93 | 1944 | |
726a989a | 1945 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
2f397a93 | 1946 | { |
355fe088 | 1947 | gimple *stmt = gsi_stmt (gsi); |
2f397a93 | 1948 | |
3b9ee1cc JJ |
1949 | /* Only the last stmt in a bb could throw, no need to call |
1950 | gimple_purge_dead_eh_edges if we change something in the middle | |
1951 | of a basic block. */ | |
1952 | cleanup_eh = false; | |
1953 | ||
c97d1c9d | 1954 | if (is_gimple_call (stmt) |
eada851c | 1955 | && gimple_call_lhs (stmt)) |
2f397a93 | 1956 | { |
78be79d5 BS |
1957 | tree arg, arg0, arg1, result; |
1958 | HOST_WIDE_INT n; | |
1959 | location_t loc; | |
2f397a93 | 1960 | |
c97d1c9d | 1961 | switch (gimple_call_combined_fn (stmt)) |
2f397a93 | 1962 | { |
c97d1c9d RS |
1963 | CASE_CFN_COS: |
1964 | CASE_CFN_SIN: | |
1965 | CASE_CFN_CEXPI: | |
fa65a9cf | 1966 | /* Make sure we have either sincos or cexp. */ |
d33d9e47 AI |
1967 | if (!targetm.libc_has_function (function_c99_math_complex) |
1968 | && !targetm.libc_has_function (function_sincos)) | |
fa65a9cf BS |
1969 | break; |
1970 | ||
726a989a | 1971 | arg = gimple_call_arg (stmt, 0); |
2f397a93 | 1972 | if (TREE_CODE (arg) == SSA_NAME) |
90bc1cb8 | 1973 | cfg_changed |= execute_cse_sincos_1 (arg); |
2f397a93 RG |
1974 | break; |
1975 | ||
c97d1c9d | 1976 | CASE_CFN_POW: |
d24ad7d6 BS |
1977 | arg0 = gimple_call_arg (stmt, 0); |
1978 | arg1 = gimple_call_arg (stmt, 1); | |
1979 | ||
1980 | loc = gimple_location (stmt); | |
1981 | result = gimple_expand_builtin_pow (&gsi, loc, arg0, arg1); | |
1982 | ||
1983 | if (result) | |
1984 | { | |
1985 | tree lhs = gimple_get_lhs (stmt); | |
538dd0b7 | 1986 | gassign *new_stmt = gimple_build_assign (lhs, result); |
d24ad7d6 BS |
1987 | gimple_set_location (new_stmt, loc); |
1988 | unlink_stmt_vdef (stmt); | |
1989 | gsi_replace (&gsi, new_stmt, true); | |
3b9ee1cc | 1990 | cleanup_eh = true; |
3d3f2249 RG |
1991 | if (gimple_vdef (stmt)) |
1992 | release_ssa_name (gimple_vdef (stmt)); | |
d24ad7d6 BS |
1993 | } |
1994 | break; | |
1995 | ||
c97d1c9d | 1996 | CASE_CFN_POWI: |
78be79d5 BS |
1997 | arg0 = gimple_call_arg (stmt, 0); |
1998 | arg1 = gimple_call_arg (stmt, 1); | |
78be79d5 | 1999 | loc = gimple_location (stmt); |
0fa6e0ef | 2000 | |
e3530904 | 2001 | if (real_minus_onep (arg0)) |
0fa6e0ef TB |
2002 | { |
2003 | tree t0, t1, cond, one, minus_one; | |
538dd0b7 | 2004 | gassign *stmt; |
0fa6e0ef TB |
2005 | |
2006 | t0 = TREE_TYPE (arg0); | |
2007 | t1 = TREE_TYPE (arg1); | |
2008 | one = build_real (t0, dconst1); | |
2009 | minus_one = build_real (t0, dconstm1); | |
2010 | ||
2011 | cond = make_temp_ssa_name (t1, NULL, "powi_cond"); | |
0d0e4a03 JJ |
2012 | stmt = gimple_build_assign (cond, BIT_AND_EXPR, |
2013 | arg1, build_int_cst (t1, 1)); | |
0fa6e0ef TB |
2014 | gimple_set_location (stmt, loc); |
2015 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
2016 | ||
2017 | result = make_temp_ssa_name (t0, NULL, "powi"); | |
0d0e4a03 JJ |
2018 | stmt = gimple_build_assign (result, COND_EXPR, cond, |
2019 | minus_one, one); | |
0fa6e0ef TB |
2020 | gimple_set_location (stmt, loc); |
2021 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
2022 | } | |
2023 | else | |
2024 | { | |
9541ffee | 2025 | if (!tree_fits_shwi_p (arg1)) |
daf4e940 TB |
2026 | break; |
2027 | ||
eb1ce453 | 2028 | n = tree_to_shwi (arg1); |
0fa6e0ef TB |
2029 | result = gimple_expand_builtin_powi (&gsi, loc, arg0, n); |
2030 | } | |
78be79d5 BS |
2031 | |
2032 | if (result) | |
2033 | { | |
2034 | tree lhs = gimple_get_lhs (stmt); | |
538dd0b7 | 2035 | gassign *new_stmt = gimple_build_assign (lhs, result); |
78be79d5 | 2036 | gimple_set_location (new_stmt, loc); |
d7e2a1c1 BS |
2037 | unlink_stmt_vdef (stmt); |
2038 | gsi_replace (&gsi, new_stmt, true); | |
3b9ee1cc | 2039 | cleanup_eh = true; |
3d3f2249 RG |
2040 | if (gimple_vdef (stmt)) |
2041 | release_ssa_name (gimple_vdef (stmt)); | |
d7e2a1c1 BS |
2042 | } |
2043 | break; | |
2044 | ||
c97d1c9d | 2045 | CASE_CFN_CABS: |
d7e2a1c1 BS |
2046 | arg0 = gimple_call_arg (stmt, 0); |
2047 | loc = gimple_location (stmt); | |
2048 | result = gimple_expand_builtin_cabs (&gsi, loc, arg0); | |
2049 | ||
2050 | if (result) | |
2051 | { | |
2052 | tree lhs = gimple_get_lhs (stmt); | |
538dd0b7 | 2053 | gassign *new_stmt = gimple_build_assign (lhs, result); |
d7e2a1c1 | 2054 | gimple_set_location (new_stmt, loc); |
78be79d5 BS |
2055 | unlink_stmt_vdef (stmt); |
2056 | gsi_replace (&gsi, new_stmt, true); | |
3b9ee1cc | 2057 | cleanup_eh = true; |
3d3f2249 RG |
2058 | if (gimple_vdef (stmt)) |
2059 | release_ssa_name (gimple_vdef (stmt)); | |
78be79d5 BS |
2060 | } |
2061 | break; | |
2062 | ||
2f397a93 RG |
2063 | default:; |
2064 | } | |
2065 | } | |
2066 | } | |
3b9ee1cc JJ |
2067 | if (cleanup_eh) |
2068 | cfg_changed |= gimple_purge_dead_eh_edges (bb); | |
2f397a93 RG |
2069 | } |
2070 | ||
be55bfe6 | 2071 | statistics_counter_event (fun, "sincos statements inserted", |
4da3b811 NF |
2072 | sincos_stats.inserted); |
2073 | ||
90bc1cb8 | 2074 | return cfg_changed ? TODO_cleanup_cfg : 0; |
2f397a93 RG |
2075 | } |
2076 | ||
27a4cd48 DM |
2077 | } // anon namespace |
2078 | ||
2079 | gimple_opt_pass * | |
2080 | make_pass_cse_sincos (gcc::context *ctxt) | |
2081 | { | |
2082 | return new pass_cse_sincos (ctxt); | |
2083 | } | |
2084 | ||
7ab6a828 RE |
2085 | /* Return true if stmt is a type conversion operation that can be stripped |
2086 | when used in a widening multiply operation. */ | |
2087 | static bool | |
355fe088 | 2088 | widening_mult_conversion_strippable_p (tree result_type, gimple *stmt) |
7ab6a828 RE |
2089 | { |
2090 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); | |
2091 | ||
2092 | if (TREE_CODE (result_type) == INTEGER_TYPE) | |
2093 | { | |
2094 | tree op_type; | |
2095 | tree inner_op_type; | |
2096 | ||
2097 | if (!CONVERT_EXPR_CODE_P (rhs_code)) | |
2098 | return false; | |
2099 | ||
2100 | op_type = TREE_TYPE (gimple_assign_lhs (stmt)); | |
2101 | ||
2102 | /* If the type of OP has the same precision as the result, then | |
2103 | we can strip this conversion. The multiply operation will be | |
2104 | selected to create the correct extension as a by-product. */ | |
2105 | if (TYPE_PRECISION (result_type) == TYPE_PRECISION (op_type)) | |
2106 | return true; | |
2107 | ||
2108 | /* We can also strip a conversion if it preserves the signed-ness of | |
2109 | the operation and doesn't narrow the range. */ | |
2110 | inner_op_type = TREE_TYPE (gimple_assign_rhs1 (stmt)); | |
2111 | ||
e919e5bf RE |
2112 | /* If the inner-most type is unsigned, then we can strip any |
2113 | intermediate widening operation. If it's signed, then the | |
2114 | intermediate widening operation must also be signed. */ | |
2115 | if ((TYPE_UNSIGNED (inner_op_type) | |
2116 | || TYPE_UNSIGNED (op_type) == TYPE_UNSIGNED (inner_op_type)) | |
7ab6a828 RE |
2117 | && TYPE_PRECISION (op_type) > TYPE_PRECISION (inner_op_type)) |
2118 | return true; | |
2119 | ||
2120 | return false; | |
2121 | } | |
2122 | ||
2123 | return rhs_code == FIXED_CONVERT_EXPR; | |
2124 | } | |
2125 | ||
26a855d7 AS |
2126 | /* Return true if RHS is a suitable operand for a widening multiplication, |
2127 | assuming a target type of TYPE. | |
1a39adae RS |
2128 | There are two cases: |
2129 | ||
5dfe80ba AS |
2130 | - RHS makes some value at least twice as wide. Store that value |
2131 | in *NEW_RHS_OUT if so, and store its type in *TYPE_OUT. | |
1a39adae RS |
2132 | |
2133 | - RHS is an integer constant. Store that value in *NEW_RHS_OUT if so, | |
2134 | but leave *TYPE_OUT untouched. */ | |
0354c0c7 BS |
2135 | |
2136 | static bool | |
26a855d7 AS |
2137 | is_widening_mult_rhs_p (tree type, tree rhs, tree *type_out, |
2138 | tree *new_rhs_out) | |
1a39adae | 2139 | { |
355fe088 | 2140 | gimple *stmt; |
26a855d7 | 2141 | tree type1, rhs1; |
1a39adae RS |
2142 | |
2143 | if (TREE_CODE (rhs) == SSA_NAME) | |
2144 | { | |
1a39adae | 2145 | stmt = SSA_NAME_DEF_STMT (rhs); |
26a855d7 AS |
2146 | if (is_gimple_assign (stmt)) |
2147 | { | |
7ab6a828 | 2148 | if (! widening_mult_conversion_strippable_p (type, stmt)) |
26a855d7 AS |
2149 | rhs1 = rhs; |
2150 | else | |
a6f969f4 AS |
2151 | { |
2152 | rhs1 = gimple_assign_rhs1 (stmt); | |
2153 | ||
2154 | if (TREE_CODE (rhs1) == INTEGER_CST) | |
2155 | { | |
2156 | *new_rhs_out = rhs1; | |
2157 | *type_out = NULL; | |
2158 | return true; | |
2159 | } | |
2160 | } | |
26a855d7 AS |
2161 | } |
2162 | else | |
2163 | rhs1 = rhs; | |
1a39adae | 2164 | |
1a39adae | 2165 | type1 = TREE_TYPE (rhs1); |
26a855d7 | 2166 | |
1a39adae | 2167 | if (TREE_CODE (type1) != TREE_CODE (type) |
5dfe80ba | 2168 | || TYPE_PRECISION (type1) * 2 > TYPE_PRECISION (type)) |
1a39adae RS |
2169 | return false; |
2170 | ||
2171 | *new_rhs_out = rhs1; | |
2172 | *type_out = type1; | |
2173 | return true; | |
2174 | } | |
2175 | ||
2176 | if (TREE_CODE (rhs) == INTEGER_CST) | |
2177 | { | |
2178 | *new_rhs_out = rhs; | |
2179 | *type_out = NULL; | |
2180 | return true; | |
2181 | } | |
2182 | ||
2183 | return false; | |
2184 | } | |
2185 | ||
26a855d7 AS |
2186 | /* Return true if STMT performs a widening multiplication, assuming the |
2187 | output type is TYPE. If so, store the unwidened types of the operands | |
2188 | in *TYPE1_OUT and *TYPE2_OUT respectively. Also fill *RHS1_OUT and | |
2189 | *RHS2_OUT such that converting those operands to types *TYPE1_OUT | |
2190 | and *TYPE2_OUT would give the operands of the multiplication. */ | |
1a39adae RS |
2191 | |
2192 | static bool | |
355fe088 | 2193 | is_widening_mult_p (gimple *stmt, |
1a39adae RS |
2194 | tree *type1_out, tree *rhs1_out, |
2195 | tree *type2_out, tree *rhs2_out) | |
0354c0c7 | 2196 | { |
3d71881d AS |
2197 | tree type = TREE_TYPE (gimple_assign_lhs (stmt)); |
2198 | ||
9b8e85a5 JJ |
2199 | if (TREE_CODE (type) == INTEGER_TYPE) |
2200 | { | |
2201 | if (TYPE_OVERFLOW_TRAPS (type)) | |
2202 | return false; | |
2203 | } | |
2204 | else if (TREE_CODE (type) != FIXED_POINT_TYPE) | |
1a39adae | 2205 | return false; |
0354c0c7 | 2206 | |
26a855d7 AS |
2207 | if (!is_widening_mult_rhs_p (type, gimple_assign_rhs1 (stmt), type1_out, |
2208 | rhs1_out)) | |
0354c0c7 BS |
2209 | return false; |
2210 | ||
26a855d7 AS |
2211 | if (!is_widening_mult_rhs_p (type, gimple_assign_rhs2 (stmt), type2_out, |
2212 | rhs2_out)) | |
1a39adae | 2213 | return false; |
0354c0c7 | 2214 | |
1a39adae | 2215 | if (*type1_out == NULL) |
0354c0c7 | 2216 | { |
1a39adae | 2217 | if (*type2_out == NULL || !int_fits_type_p (*rhs1_out, *type2_out)) |
0354c0c7 | 2218 | return false; |
1a39adae | 2219 | *type1_out = *type2_out; |
0354c0c7 | 2220 | } |
0354c0c7 | 2221 | |
1a39adae | 2222 | if (*type2_out == NULL) |
0354c0c7 | 2223 | { |
1a39adae | 2224 | if (!int_fits_type_p (*rhs2_out, *type1_out)) |
0354c0c7 | 2225 | return false; |
1a39adae | 2226 | *type2_out = *type1_out; |
0354c0c7 | 2227 | } |
0354c0c7 | 2228 | |
ff63d754 AS |
2229 | /* Ensure that the larger of the two operands comes first. */ |
2230 | if (TYPE_PRECISION (*type1_out) < TYPE_PRECISION (*type2_out)) | |
2231 | { | |
fab27f52 MM |
2232 | std::swap (*type1_out, *type2_out); |
2233 | std::swap (*rhs1_out, *rhs2_out); | |
ff63d754 | 2234 | } |
5dfe80ba | 2235 | |
1a39adae RS |
2236 | return true; |
2237 | } | |
0354c0c7 | 2238 | |
336a06a1 TC |
2239 | /* Check to see if the CALL statement is an invocation of copysign |
2240 | with 1. being the first argument. */ | |
2241 | static bool | |
2242 | is_copysign_call_with_1 (gimple *call) | |
2243 | { | |
2244 | gcall *c = dyn_cast <gcall *> (call); | |
2245 | if (! c) | |
2246 | return false; | |
2247 | ||
2248 | enum combined_fn code = gimple_call_combined_fn (c); | |
2249 | ||
2250 | if (code == CFN_LAST) | |
2251 | return false; | |
2252 | ||
2253 | if (builtin_fn_p (code)) | |
2254 | { | |
2255 | switch (as_builtin_fn (code)) | |
2256 | { | |
2257 | CASE_FLT_FN (BUILT_IN_COPYSIGN): | |
2258 | CASE_FLT_FN_FLOATN_NX (BUILT_IN_COPYSIGN): | |
2259 | return real_onep (gimple_call_arg (c, 0)); | |
2260 | default: | |
2261 | return false; | |
2262 | } | |
2263 | } | |
2264 | ||
2265 | if (internal_fn_p (code)) | |
2266 | { | |
2267 | switch (as_internal_fn (code)) | |
2268 | { | |
2269 | case IFN_COPYSIGN: | |
2270 | return real_onep (gimple_call_arg (c, 0)); | |
2271 | default: | |
2272 | return false; | |
2273 | } | |
2274 | } | |
2275 | ||
2276 | return false; | |
2277 | } | |
2278 | ||
2279 | /* Try to expand the pattern x * copysign (1, y) into xorsign (x, y). | |
2280 | This only happens when the the xorsign optab is defined, if the | |
2281 | pattern is not a xorsign pattern or if expansion fails FALSE is | |
2282 | returned, otherwise TRUE is returned. */ | |
2283 | static bool | |
2284 | convert_expand_mult_copysign (gimple *stmt, gimple_stmt_iterator *gsi) | |
2285 | { | |
2286 | tree treeop0, treeop1, lhs, type; | |
2287 | location_t loc = gimple_location (stmt); | |
2288 | lhs = gimple_assign_lhs (stmt); | |
2289 | treeop0 = gimple_assign_rhs1 (stmt); | |
2290 | treeop1 = gimple_assign_rhs2 (stmt); | |
2291 | type = TREE_TYPE (lhs); | |
2292 | machine_mode mode = TYPE_MODE (type); | |
2293 | ||
9880acc1 | 2294 | if (HONOR_SNANS (type)) |
336a06a1 TC |
2295 | return false; |
2296 | ||
2297 | if (TREE_CODE (treeop0) == SSA_NAME && TREE_CODE (treeop1) == SSA_NAME) | |
2298 | { | |
2299 | gimple *call0 = SSA_NAME_DEF_STMT (treeop0); | |
9880acc1 | 2300 | if (!has_single_use (treeop0) || !is_copysign_call_with_1 (call0)) |
336a06a1 TC |
2301 | { |
2302 | call0 = SSA_NAME_DEF_STMT (treeop1); | |
9880acc1 | 2303 | if (!has_single_use (treeop1) || !is_copysign_call_with_1 (call0)) |
336a06a1 TC |
2304 | return false; |
2305 | ||
2306 | treeop1 = treeop0; | |
2307 | } | |
336a06a1 TC |
2308 | if (optab_handler (xorsign_optab, mode) == CODE_FOR_nothing) |
2309 | return false; | |
2310 | ||
2311 | gcall *c = as_a<gcall*> (call0); | |
2312 | treeop0 = gimple_call_arg (c, 1); | |
2313 | ||
2314 | gcall *call_stmt | |
2315 | = gimple_build_call_internal (IFN_XORSIGN, 2, treeop1, treeop0); | |
2316 | gimple_set_lhs (call_stmt, lhs); | |
2317 | gimple_set_location (call_stmt, loc); | |
2318 | gsi_replace (gsi, call_stmt, true); | |
2319 | return true; | |
2320 | } | |
2321 | ||
2322 | return false; | |
2323 | } | |
2324 | ||
1a39adae RS |
2325 | /* Process a single gimple statement STMT, which has a MULT_EXPR as |
2326 | its rhs, and try to convert it into a WIDEN_MULT_EXPR. The return | |
2327 | value is true iff we converted the statement. */ | |
2328 | ||
2329 | static bool | |
355fe088 | 2330 | convert_mult_to_widen (gimple *stmt, gimple_stmt_iterator *gsi) |
1a39adae | 2331 | { |
83d5977e | 2332 | tree lhs, rhs1, rhs2, type, type1, type2; |
1a39adae | 2333 | enum insn_code handler; |
4b926fea | 2334 | scalar_int_mode to_mode, from_mode, actual_mode; |
a484f6ba | 2335 | optab op; |
5dfe80ba AS |
2336 | int actual_precision; |
2337 | location_t loc = gimple_location (stmt); | |
db719f50 | 2338 | bool from_unsigned1, from_unsigned2; |
1a39adae RS |
2339 | |
2340 | lhs = gimple_assign_lhs (stmt); | |
2341 | type = TREE_TYPE (lhs); | |
2342 | if (TREE_CODE (type) != INTEGER_TYPE) | |
0354c0c7 BS |
2343 | return false; |
2344 | ||
3d71881d | 2345 | if (!is_widening_mult_p (stmt, &type1, &rhs1, &type2, &rhs2)) |
0354c0c7 BS |
2346 | return false; |
2347 | ||
7a504f33 RS |
2348 | to_mode = SCALAR_INT_TYPE_MODE (type); |
2349 | from_mode = SCALAR_INT_TYPE_MODE (type1); | |
9134df2c RS |
2350 | if (to_mode == from_mode) |
2351 | return false; | |
2352 | ||
db719f50 AS |
2353 | from_unsigned1 = TYPE_UNSIGNED (type1); |
2354 | from_unsigned2 = TYPE_UNSIGNED (type2); | |
a484f6ba | 2355 | |
db719f50 | 2356 | if (from_unsigned1 && from_unsigned2) |
a484f6ba | 2357 | op = umul_widen_optab; |
db719f50 | 2358 | else if (!from_unsigned1 && !from_unsigned2) |
a484f6ba | 2359 | op = smul_widen_optab; |
0354c0c7 | 2360 | else |
a484f6ba AS |
2361 | op = usmul_widen_optab; |
2362 | ||
5dfe80ba | 2363 | handler = find_widening_optab_handler_and_mode (op, to_mode, from_mode, |
4b926fea | 2364 | &actual_mode); |
1a39adae RS |
2365 | |
2366 | if (handler == CODE_FOR_nothing) | |
db719f50 AS |
2367 | { |
2368 | if (op != smul_widen_optab) | |
2369 | { | |
6a228c2c AS |
2370 | /* We can use a signed multiply with unsigned types as long as |
2371 | there is a wider mode to use, or it is the smaller of the two | |
2372 | types that is unsigned. Note that type1 >= type2, always. */ | |
2373 | if ((TYPE_UNSIGNED (type1) | |
2374 | && TYPE_PRECISION (type1) == GET_MODE_PRECISION (from_mode)) | |
2375 | || (TYPE_UNSIGNED (type2) | |
2376 | && TYPE_PRECISION (type2) == GET_MODE_PRECISION (from_mode))) | |
2377 | { | |
490d0f6c RS |
2378 | if (!GET_MODE_WIDER_MODE (from_mode).exists (&from_mode) |
2379 | || GET_MODE_SIZE (to_mode) <= GET_MODE_SIZE (from_mode)) | |
6a228c2c AS |
2380 | return false; |
2381 | } | |
db719f50 AS |
2382 | |
2383 | op = smul_widen_optab; | |
2384 | handler = find_widening_optab_handler_and_mode (op, to_mode, | |
4b926fea | 2385 | from_mode, |
db719f50 AS |
2386 | &actual_mode); |
2387 | ||
2388 | if (handler == CODE_FOR_nothing) | |
2389 | return false; | |
2390 | ||
2391 | from_unsigned1 = from_unsigned2 = false; | |
2392 | } | |
2393 | else | |
2394 | return false; | |
2395 | } | |
1a39adae | 2396 | |
5dfe80ba AS |
2397 | /* Ensure that the inputs to the handler are in the correct precison |
2398 | for the opcode. This will be the full mode size. */ | |
2399 | actual_precision = GET_MODE_PRECISION (actual_mode); | |
f409d239 RG |
2400 | if (2 * actual_precision > TYPE_PRECISION (type)) |
2401 | return false; | |
db719f50 AS |
2402 | if (actual_precision != TYPE_PRECISION (type1) |
2403 | || from_unsigned1 != TYPE_UNSIGNED (type1)) | |
83d5977e RG |
2404 | rhs1 = build_and_insert_cast (gsi, loc, |
2405 | build_nonstandard_integer_type | |
2406 | (actual_precision, from_unsigned1), rhs1); | |
db719f50 AS |
2407 | if (actual_precision != TYPE_PRECISION (type2) |
2408 | || from_unsigned2 != TYPE_UNSIGNED (type2)) | |
83d5977e RG |
2409 | rhs2 = build_and_insert_cast (gsi, loc, |
2410 | build_nonstandard_integer_type | |
2411 | (actual_precision, from_unsigned2), rhs2); | |
5dfe80ba | 2412 | |
a6f969f4 AS |
2413 | /* Handle constants. */ |
2414 | if (TREE_CODE (rhs1) == INTEGER_CST) | |
2415 | rhs1 = fold_convert (type1, rhs1); | |
2416 | if (TREE_CODE (rhs2) == INTEGER_CST) | |
2417 | rhs2 = fold_convert (type2, rhs2); | |
2418 | ||
5dfe80ba AS |
2419 | gimple_assign_set_rhs1 (stmt, rhs1); |
2420 | gimple_assign_set_rhs2 (stmt, rhs2); | |
0354c0c7 BS |
2421 | gimple_assign_set_rhs_code (stmt, WIDEN_MULT_EXPR); |
2422 | update_stmt (stmt); | |
4da3b811 | 2423 | widen_mul_stats.widen_mults_inserted++; |
0354c0c7 BS |
2424 | return true; |
2425 | } | |
2426 | ||
2427 | /* Process a single gimple statement STMT, which is found at the | |
2428 | iterator GSI and has a either a PLUS_EXPR or a MINUS_EXPR as its | |
2429 | rhs (given by CODE), and try to convert it into a | |
2430 | WIDEN_MULT_PLUS_EXPR or a WIDEN_MULT_MINUS_EXPR. The return value | |
2431 | is true iff we converted the statement. */ | |
2432 | ||
2433 | static bool | |
355fe088 | 2434 | convert_plusminus_to_widen (gimple_stmt_iterator *gsi, gimple *stmt, |
0354c0c7 BS |
2435 | enum tree_code code) |
2436 | { | |
355fe088 TS |
2437 | gimple *rhs1_stmt = NULL, *rhs2_stmt = NULL; |
2438 | gimple *conv1_stmt = NULL, *conv2_stmt = NULL, *conv_stmt; | |
83d5977e | 2439 | tree type, type1, type2, optype; |
0354c0c7 BS |
2440 | tree lhs, rhs1, rhs2, mult_rhs1, mult_rhs2, add_rhs; |
2441 | enum tree_code rhs1_code = ERROR_MARK, rhs2_code = ERROR_MARK; | |
2442 | optab this_optab; | |
2443 | enum tree_code wmult_code; | |
5dfe80ba | 2444 | enum insn_code handler; |
4b926fea | 2445 | scalar_mode to_mode, from_mode, actual_mode; |
5dfe80ba AS |
2446 | location_t loc = gimple_location (stmt); |
2447 | int actual_precision; | |
db719f50 | 2448 | bool from_unsigned1, from_unsigned2; |
0354c0c7 BS |
2449 | |
2450 | lhs = gimple_assign_lhs (stmt); | |
2451 | type = TREE_TYPE (lhs); | |
1a39adae RS |
2452 | if (TREE_CODE (type) != INTEGER_TYPE |
2453 | && TREE_CODE (type) != FIXED_POINT_TYPE) | |
0354c0c7 BS |
2454 | return false; |
2455 | ||
2456 | if (code == MINUS_EXPR) | |
2457 | wmult_code = WIDEN_MULT_MINUS_EXPR; | |
2458 | else | |
2459 | wmult_code = WIDEN_MULT_PLUS_EXPR; | |
2460 | ||
0354c0c7 BS |
2461 | rhs1 = gimple_assign_rhs1 (stmt); |
2462 | rhs2 = gimple_assign_rhs2 (stmt); | |
2463 | ||
2464 | if (TREE_CODE (rhs1) == SSA_NAME) | |
2465 | { | |
2466 | rhs1_stmt = SSA_NAME_DEF_STMT (rhs1); | |
2467 | if (is_gimple_assign (rhs1_stmt)) | |
2468 | rhs1_code = gimple_assign_rhs_code (rhs1_stmt); | |
2469 | } | |
0354c0c7 BS |
2470 | |
2471 | if (TREE_CODE (rhs2) == SSA_NAME) | |
2472 | { | |
2473 | rhs2_stmt = SSA_NAME_DEF_STMT (rhs2); | |
2474 | if (is_gimple_assign (rhs2_stmt)) | |
2475 | rhs2_code = gimple_assign_rhs_code (rhs2_stmt); | |
2476 | } | |
0354c0c7 | 2477 | |
cefb4d4f AS |
2478 | /* Allow for one conversion statement between the multiply |
2479 | and addition/subtraction statement. If there are more than | |
2480 | one conversions then we assume they would invalidate this | |
2481 | transformation. If that's not the case then they should have | |
2482 | been folded before now. */ | |
2483 | if (CONVERT_EXPR_CODE_P (rhs1_code)) | |
2484 | { | |
2485 | conv1_stmt = rhs1_stmt; | |
2486 | rhs1 = gimple_assign_rhs1 (rhs1_stmt); | |
2487 | if (TREE_CODE (rhs1) == SSA_NAME) | |
2488 | { | |
2489 | rhs1_stmt = SSA_NAME_DEF_STMT (rhs1); | |
2490 | if (is_gimple_assign (rhs1_stmt)) | |
2491 | rhs1_code = gimple_assign_rhs_code (rhs1_stmt); | |
2492 | } | |
2493 | else | |
2494 | return false; | |
2495 | } | |
2496 | if (CONVERT_EXPR_CODE_P (rhs2_code)) | |
2497 | { | |
2498 | conv2_stmt = rhs2_stmt; | |
2499 | rhs2 = gimple_assign_rhs1 (rhs2_stmt); | |
2500 | if (TREE_CODE (rhs2) == SSA_NAME) | |
2501 | { | |
2502 | rhs2_stmt = SSA_NAME_DEF_STMT (rhs2); | |
2503 | if (is_gimple_assign (rhs2_stmt)) | |
2504 | rhs2_code = gimple_assign_rhs_code (rhs2_stmt); | |
2505 | } | |
2506 | else | |
2507 | return false; | |
2508 | } | |
2509 | ||
5dfe80ba AS |
2510 | /* If code is WIDEN_MULT_EXPR then it would seem unnecessary to call |
2511 | is_widening_mult_p, but we still need the rhs returns. | |
2512 | ||
2513 | It might also appear that it would be sufficient to use the existing | |
2514 | operands of the widening multiply, but that would limit the choice of | |
42917d01 YZ |
2515 | multiply-and-accumulate instructions. |
2516 | ||
2517 | If the widened-multiplication result has more than one uses, it is | |
2518 | probably wiser not to do the conversion. */ | |
5dfe80ba AS |
2519 | if (code == PLUS_EXPR |
2520 | && (rhs1_code == MULT_EXPR || rhs1_code == WIDEN_MULT_EXPR)) | |
0354c0c7 | 2521 | { |
42917d01 YZ |
2522 | if (!has_single_use (rhs1) |
2523 | || !is_widening_mult_p (rhs1_stmt, &type1, &mult_rhs1, | |
2524 | &type2, &mult_rhs2)) | |
0354c0c7 | 2525 | return false; |
1a39adae | 2526 | add_rhs = rhs2; |
cefb4d4f | 2527 | conv_stmt = conv1_stmt; |
0354c0c7 | 2528 | } |
5dfe80ba | 2529 | else if (rhs2_code == MULT_EXPR || rhs2_code == WIDEN_MULT_EXPR) |
0354c0c7 | 2530 | { |
42917d01 YZ |
2531 | if (!has_single_use (rhs2) |
2532 | || !is_widening_mult_p (rhs2_stmt, &type1, &mult_rhs1, | |
2533 | &type2, &mult_rhs2)) | |
0354c0c7 | 2534 | return false; |
1a39adae | 2535 | add_rhs = rhs1; |
cefb4d4f | 2536 | conv_stmt = conv2_stmt; |
0354c0c7 | 2537 | } |
0354c0c7 BS |
2538 | else |
2539 | return false; | |
2540 | ||
b397965c RS |
2541 | to_mode = SCALAR_TYPE_MODE (type); |
2542 | from_mode = SCALAR_TYPE_MODE (type1); | |
9134df2c RS |
2543 | if (to_mode == from_mode) |
2544 | return false; | |
2545 | ||
db719f50 AS |
2546 | from_unsigned1 = TYPE_UNSIGNED (type1); |
2547 | from_unsigned2 = TYPE_UNSIGNED (type2); | |
3752b2ab | 2548 | optype = type1; |
5dfe80ba | 2549 | |
db719f50 AS |
2550 | /* There's no such thing as a mixed sign madd yet, so use a wider mode. */ |
2551 | if (from_unsigned1 != from_unsigned2) | |
2552 | { | |
3752b2ab RS |
2553 | if (!INTEGRAL_TYPE_P (type)) |
2554 | return false; | |
6a228c2c AS |
2555 | /* We can use a signed multiply with unsigned types as long as |
2556 | there is a wider mode to use, or it is the smaller of the two | |
2557 | types that is unsigned. Note that type1 >= type2, always. */ | |
2558 | if ((from_unsigned1 | |
2559 | && TYPE_PRECISION (type1) == GET_MODE_PRECISION (from_mode)) | |
2560 | || (from_unsigned2 | |
2561 | && TYPE_PRECISION (type2) == GET_MODE_PRECISION (from_mode))) | |
db719f50 | 2562 | { |
490d0f6c RS |
2563 | if (!GET_MODE_WIDER_MODE (from_mode).exists (&from_mode) |
2564 | || GET_MODE_SIZE (from_mode) >= GET_MODE_SIZE (to_mode)) | |
6a228c2c | 2565 | return false; |
db719f50 | 2566 | } |
6a228c2c AS |
2567 | |
2568 | from_unsigned1 = from_unsigned2 = false; | |
3752b2ab RS |
2569 | optype = build_nonstandard_integer_type (GET_MODE_PRECISION (from_mode), |
2570 | false); | |
db719f50 | 2571 | } |
9eab7f91 | 2572 | |
cefb4d4f AS |
2573 | /* If there was a conversion between the multiply and addition |
2574 | then we need to make sure it fits a multiply-and-accumulate. | |
2575 | The should be a single mode change which does not change the | |
2576 | value. */ | |
2577 | if (conv_stmt) | |
2578 | { | |
db719f50 | 2579 | /* We use the original, unmodified data types for this. */ |
cefb4d4f AS |
2580 | tree from_type = TREE_TYPE (gimple_assign_rhs1 (conv_stmt)); |
2581 | tree to_type = TREE_TYPE (gimple_assign_lhs (conv_stmt)); | |
2582 | int data_size = TYPE_PRECISION (type1) + TYPE_PRECISION (type2); | |
2583 | bool is_unsigned = TYPE_UNSIGNED (type1) && TYPE_UNSIGNED (type2); | |
2584 | ||
2585 | if (TYPE_PRECISION (from_type) > TYPE_PRECISION (to_type)) | |
2586 | { | |
2587 | /* Conversion is a truncate. */ | |
2588 | if (TYPE_PRECISION (to_type) < data_size) | |
2589 | return false; | |
2590 | } | |
2591 | else if (TYPE_PRECISION (from_type) < TYPE_PRECISION (to_type)) | |
2592 | { | |
2593 | /* Conversion is an extend. Check it's the right sort. */ | |
2594 | if (TYPE_UNSIGNED (from_type) != is_unsigned | |
2595 | && !(is_unsigned && TYPE_PRECISION (from_type) > data_size)) | |
2596 | return false; | |
2597 | } | |
2598 | /* else convert is a no-op for our purposes. */ | |
2599 | } | |
2600 | ||
9eab7f91 RS |
2601 | /* Verify that the machine can perform a widening multiply |
2602 | accumulate in this mode/signedness combination, otherwise | |
2603 | this transformation is likely to pessimize code. */ | |
db719f50 | 2604 | this_optab = optab_for_tree_code (wmult_code, optype, optab_default); |
5dfe80ba | 2605 | handler = find_widening_optab_handler_and_mode (this_optab, to_mode, |
4b926fea | 2606 | from_mode, &actual_mode); |
5dfe80ba AS |
2607 | |
2608 | if (handler == CODE_FOR_nothing) | |
9eab7f91 RS |
2609 | return false; |
2610 | ||
5dfe80ba AS |
2611 | /* Ensure that the inputs to the handler are in the correct precison |
2612 | for the opcode. This will be the full mode size. */ | |
2613 | actual_precision = GET_MODE_PRECISION (actual_mode); | |
db719f50 AS |
2614 | if (actual_precision != TYPE_PRECISION (type1) |
2615 | || from_unsigned1 != TYPE_UNSIGNED (type1)) | |
83d5977e RG |
2616 | mult_rhs1 = build_and_insert_cast (gsi, loc, |
2617 | build_nonstandard_integer_type | |
2618 | (actual_precision, from_unsigned1), | |
2619 | mult_rhs1); | |
db719f50 AS |
2620 | if (actual_precision != TYPE_PRECISION (type2) |
2621 | || from_unsigned2 != TYPE_UNSIGNED (type2)) | |
83d5977e RG |
2622 | mult_rhs2 = build_and_insert_cast (gsi, loc, |
2623 | build_nonstandard_integer_type | |
2624 | (actual_precision, from_unsigned2), | |
2625 | mult_rhs2); | |
0354c0c7 | 2626 | |
75161d2c | 2627 | if (!useless_type_conversion_p (type, TREE_TYPE (add_rhs))) |
83d5977e | 2628 | add_rhs = build_and_insert_cast (gsi, loc, type, add_rhs); |
75161d2c | 2629 | |
a6f969f4 AS |
2630 | /* Handle constants. */ |
2631 | if (TREE_CODE (mult_rhs1) == INTEGER_CST) | |
c3c5a1cc | 2632 | mult_rhs1 = fold_convert (type1, mult_rhs1); |
a6f969f4 | 2633 | if (TREE_CODE (mult_rhs2) == INTEGER_CST) |
c3c5a1cc | 2634 | mult_rhs2 = fold_convert (type2, mult_rhs2); |
a6f969f4 | 2635 | |
00d66391 JJ |
2636 | gimple_assign_set_rhs_with_ops (gsi, wmult_code, mult_rhs1, mult_rhs2, |
2637 | add_rhs); | |
0354c0c7 | 2638 | update_stmt (gsi_stmt (*gsi)); |
4da3b811 | 2639 | widen_mul_stats.maccs_inserted++; |
0354c0c7 BS |
2640 | return true; |
2641 | } | |
2642 | ||
4a0d0ed2 MJ |
2643 | /* Given a result MUL_RESULT which is a result of a multiplication of OP1 and |
2644 | OP2 and which we know is used in statements that can be, together with the | |
2645 | multiplication, converted to FMAs, perform the transformation. */ | |
2646 | ||
2647 | static void | |
2648 | convert_mult_to_fma_1 (tree mul_result, tree op1, tree op2) | |
2649 | { | |
2650 | tree type = TREE_TYPE (mul_result); | |
2651 | gimple *use_stmt; | |
2652 | imm_use_iterator imm_iter; | |
c566cc9f | 2653 | gcall *fma_stmt; |
4a0d0ed2 MJ |
2654 | |
2655 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, mul_result) | |
2656 | { | |
2657 | gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt); | |
4a0d0ed2 MJ |
2658 | tree addop, mulop1 = op1, result = mul_result; |
2659 | bool negate_p = false; | |
c566cc9f | 2660 | gimple_seq seq = NULL; |
4a0d0ed2 MJ |
2661 | |
2662 | if (is_gimple_debug (use_stmt)) | |
2663 | continue; | |
2664 | ||
0936858f RS |
2665 | if (is_gimple_assign (use_stmt) |
2666 | && gimple_assign_rhs_code (use_stmt) == NEGATE_EXPR) | |
4a0d0ed2 MJ |
2667 | { |
2668 | result = gimple_assign_lhs (use_stmt); | |
2669 | use_operand_p use_p; | |
2670 | gimple *neguse_stmt; | |
2671 | single_imm_use (gimple_assign_lhs (use_stmt), &use_p, &neguse_stmt); | |
2672 | gsi_remove (&gsi, true); | |
2673 | release_defs (use_stmt); | |
2674 | ||
2675 | use_stmt = neguse_stmt; | |
2676 | gsi = gsi_for_stmt (use_stmt); | |
4a0d0ed2 MJ |
2677 | negate_p = true; |
2678 | } | |
2679 | ||
0936858f RS |
2680 | tree cond, else_value, ops[3]; |
2681 | tree_code code; | |
2682 | if (!can_interpret_as_conditional_op_p (use_stmt, &cond, &code, | |
2683 | ops, &else_value)) | |
2684 | gcc_unreachable (); | |
2685 | addop = ops[0] == result ? ops[1] : ops[0]; | |
2686 | ||
2687 | if (code == MINUS_EXPR) | |
4a0d0ed2 | 2688 | { |
0936858f RS |
2689 | if (ops[0] == result) |
2690 | /* a * b - c -> a * b + (-c) */ | |
c566cc9f | 2691 | addop = gimple_build (&seq, NEGATE_EXPR, type, addop); |
0936858f RS |
2692 | else |
2693 | /* a - b * c -> (-b) * c + a */ | |
4a0d0ed2 MJ |
2694 | negate_p = !negate_p; |
2695 | } | |
2696 | ||
2697 | if (negate_p) | |
c566cc9f | 2698 | mulop1 = gimple_build (&seq, NEGATE_EXPR, type, mulop1); |
4a0d0ed2 | 2699 | |
c566cc9f RS |
2700 | if (seq) |
2701 | gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT); | |
0936858f RS |
2702 | |
2703 | if (cond) | |
2704 | fma_stmt = gimple_build_call_internal (IFN_COND_FMA, 5, cond, mulop1, | |
2705 | op2, addop, else_value); | |
2706 | else | |
2707 | fma_stmt = gimple_build_call_internal (IFN_FMA, 3, mulop1, op2, addop); | |
2708 | gimple_set_lhs (fma_stmt, gimple_get_lhs (use_stmt)); | |
c566cc9f RS |
2709 | gimple_call_set_nothrow (fma_stmt, !stmt_can_throw_internal (use_stmt)); |
2710 | gsi_replace (&gsi, fma_stmt, true); | |
2711 | /* Follow all SSA edges so that we generate FMS, FNMA and FNMS | |
2712 | regardless of where the negation occurs. */ | |
2713 | if (fold_stmt (&gsi, follow_all_ssa_edges)) | |
2714 | update_stmt (gsi_stmt (gsi)); | |
4a0d0ed2 MJ |
2715 | |
2716 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2717 | { | |
2718 | fprintf (dump_file, "Generated FMA "); | |
4af78ef8 | 2719 | print_gimple_stmt (dump_file, gsi_stmt (gsi), 0, TDF_NONE); |
4a0d0ed2 MJ |
2720 | fprintf (dump_file, "\n"); |
2721 | } | |
2722 | ||
4a0d0ed2 MJ |
2723 | widen_mul_stats.fmas_inserted++; |
2724 | } | |
2725 | } | |
2726 | ||
2727 | /* Data necessary to perform the actual transformation from a multiplication | |
2728 | and an addition to an FMA after decision is taken it should be done and to | |
2729 | then delete the multiplication statement from the function IL. */ | |
2730 | ||
2731 | struct fma_transformation_info | |
2732 | { | |
2733 | gimple *mul_stmt; | |
2734 | tree mul_result; | |
2735 | tree op1; | |
2736 | tree op2; | |
2737 | }; | |
2738 | ||
2739 | /* Structure containing the current state of FMA deferring, i.e. whether we are | |
2740 | deferring, whether to continue deferring, and all data necessary to come | |
2741 | back and perform all deferred transformations. */ | |
2742 | ||
2743 | class fma_deferring_state | |
2744 | { | |
2745 | public: | |
2746 | /* Class constructor. Pass true as PERFORM_DEFERRING in order to actually | |
2747 | do any deferring. */ | |
2748 | ||
2749 | fma_deferring_state (bool perform_deferring) | |
2750 | : m_candidates (), m_mul_result_set (), m_initial_phi (NULL), | |
2751 | m_last_result (NULL_TREE), m_deferring_p (perform_deferring) {} | |
2752 | ||
2753 | /* List of FMA candidates for which we the transformation has been determined | |
2754 | possible but we at this point in BB analysis we do not consider them | |
2755 | beneficial. */ | |
2756 | auto_vec<fma_transformation_info, 8> m_candidates; | |
2757 | ||
2758 | /* Set of results of multiplication that are part of an already deferred FMA | |
2759 | candidates. */ | |
2760 | hash_set<tree> m_mul_result_set; | |
2761 | ||
2762 | /* The PHI that supposedly feeds back result of a FMA to another over loop | |
2763 | boundary. */ | |
2764 | gphi *m_initial_phi; | |
2765 | ||
2766 | /* Result of the last produced FMA candidate or NULL if there has not been | |
2767 | one. */ | |
2768 | tree m_last_result; | |
2769 | ||
2770 | /* If true, deferring might still be profitable. If false, transform all | |
2771 | candidates and no longer defer. */ | |
2772 | bool m_deferring_p; | |
2773 | }; | |
2774 | ||
2775 | /* Transform all deferred FMA candidates and mark STATE as no longer | |
2776 | deferring. */ | |
2777 | ||
2778 | static void | |
2779 | cancel_fma_deferring (fma_deferring_state *state) | |
2780 | { | |
2781 | if (!state->m_deferring_p) | |
2782 | return; | |
2783 | ||
2784 | for (unsigned i = 0; i < state->m_candidates.length (); i++) | |
2785 | { | |
2786 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2787 | fprintf (dump_file, "Generating deferred FMA\n"); | |
2788 | ||
2789 | const fma_transformation_info &fti = state->m_candidates[i]; | |
2790 | convert_mult_to_fma_1 (fti.mul_result, fti.op1, fti.op2); | |
2791 | ||
2792 | gimple_stmt_iterator gsi = gsi_for_stmt (fti.mul_stmt); | |
2793 | gsi_remove (&gsi, true); | |
2794 | release_defs (fti.mul_stmt); | |
2795 | } | |
2796 | state->m_deferring_p = false; | |
2797 | } | |
2798 | ||
2799 | /* If OP is an SSA name defined by a PHI node, return the PHI statement. | |
2800 | Otherwise return NULL. */ | |
2801 | ||
2802 | static gphi * | |
2803 | result_of_phi (tree op) | |
2804 | { | |
2805 | if (TREE_CODE (op) != SSA_NAME) | |
2806 | return NULL; | |
2807 | ||
2808 | return dyn_cast <gphi *> (SSA_NAME_DEF_STMT (op)); | |
2809 | } | |
2810 | ||
2811 | /* After processing statements of a BB and recording STATE, return true if the | |
2812 | initial phi is fed by the last FMA candidate result ore one such result from | |
2813 | previously processed BBs marked in LAST_RESULT_SET. */ | |
2814 | ||
2815 | static bool | |
2816 | last_fma_candidate_feeds_initial_phi (fma_deferring_state *state, | |
2817 | hash_set<tree> *last_result_set) | |
2818 | { | |
2819 | ssa_op_iter iter; | |
2820 | use_operand_p use; | |
2821 | FOR_EACH_PHI_ARG (use, state->m_initial_phi, iter, SSA_OP_USE) | |
2822 | { | |
2823 | tree t = USE_FROM_PTR (use); | |
2824 | if (t == state->m_last_result | |
2825 | || last_result_set->contains (t)) | |
2826 | return true; | |
2827 | } | |
2828 | ||
2829 | return false; | |
2830 | } | |
2831 | ||
4dbed5f6 RG |
2832 | /* Combine the multiplication at MUL_STMT with operands MULOP1 and MULOP2 |
2833 | with uses in additions and subtractions to form fused multiply-add | |
4a0d0ed2 MJ |
2834 | operations. Returns true if successful and MUL_STMT should be removed. |
2835 | ||
2836 | If STATE indicates that we are deferring FMA transformation, that means | |
2837 | that we do not produce FMAs for basic blocks which look like: | |
2838 | ||
2839 | <bb 6> | |
2840 | # accumulator_111 = PHI <0.0(5), accumulator_66(6)> | |
2841 | _65 = _14 * _16; | |
2842 | accumulator_66 = _65 + accumulator_111; | |
2843 | ||
2844 | or its unrolled version, i.e. with several FMA candidates that feed result | |
2845 | of one into the addend of another. Instead, we add them to a list in STATE | |
2846 | and if we later discover an FMA candidate that is not part of such a chain, | |
2847 | we go back and perform all deferred past candidates. */ | |
16949072 RG |
2848 | |
2849 | static bool | |
4a0d0ed2 MJ |
2850 | convert_mult_to_fma (gimple *mul_stmt, tree op1, tree op2, |
2851 | fma_deferring_state *state) | |
16949072 | 2852 | { |
4dbed5f6 | 2853 | tree mul_result = gimple_get_lhs (mul_stmt); |
16949072 | 2854 | tree type = TREE_TYPE (mul_result); |
355fe088 | 2855 | gimple *use_stmt, *neguse_stmt; |
16949072 RG |
2856 | use_operand_p use_p; |
2857 | imm_use_iterator imm_iter; | |
2858 | ||
2859 | if (FLOAT_TYPE_P (type) | |
2860 | && flag_fp_contract_mode == FP_CONTRACT_OFF) | |
2861 | return false; | |
2862 | ||
2863 | /* We don't want to do bitfield reduction ops. */ | |
2864 | if (INTEGRAL_TYPE_P (type) | |
9b8e85a5 | 2865 | && (!type_has_mode_precision_p (type) || TYPE_OVERFLOW_TRAPS (type))) |
16949072 RG |
2866 | return false; |
2867 | ||
2868 | /* If the target doesn't support it, don't generate it. We assume that | |
2869 | if fma isn't available then fms, fnma or fnms are not either. */ | |
c566cc9f RS |
2870 | optimization_type opt_type = bb_optimization_type (gimple_bb (mul_stmt)); |
2871 | if (!direct_internal_fn_supported_p (IFN_FMA, type, opt_type)) | |
16949072 RG |
2872 | return false; |
2873 | ||
0fb808ea JJ |
2874 | /* If the multiplication has zero uses, it is kept around probably because |
2875 | of -fnon-call-exceptions. Don't optimize it away in that case, | |
2876 | it is DCE job. */ | |
2877 | if (has_zero_uses (mul_result)) | |
2878 | return false; | |
2879 | ||
4a0d0ed2 MJ |
2880 | bool check_defer |
2881 | = (state->m_deferring_p | |
2882 | && (tree_to_shwi (TYPE_SIZE (type)) | |
2883 | <= PARAM_VALUE (PARAM_AVOID_FMA_MAX_BITS))); | |
2884 | bool defer = check_defer; | |
16949072 RG |
2885 | /* Make sure that the multiplication statement becomes dead after |
2886 | the transformation, thus that all uses are transformed to FMAs. | |
2887 | This means we assume that an FMA operation has the same cost | |
2888 | as an addition. */ | |
2889 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, mul_result) | |
2890 | { | |
a5f09e73 RH |
2891 | tree result = mul_result; |
2892 | bool negate_p = false; | |
16949072 RG |
2893 | |
2894 | use_stmt = USE_STMT (use_p); | |
2895 | ||
76b14c29 RG |
2896 | if (is_gimple_debug (use_stmt)) |
2897 | continue; | |
2898 | ||
16949072 RG |
2899 | /* For now restrict this operations to single basic blocks. In theory |
2900 | we would want to support sinking the multiplication in | |
2901 | m = a*b; | |
2902 | if () | |
2903 | ma = m + c; | |
2904 | else | |
2905 | d = m; | |
2906 | to form a fma in the then block and sink the multiplication to the | |
2907 | else block. */ | |
2908 | if (gimple_bb (use_stmt) != gimple_bb (mul_stmt)) | |
2909 | return false; | |
2910 | ||
a5f09e73 | 2911 | /* A negate on the multiplication leads to FNMA. */ |
0936858f RS |
2912 | if (is_gimple_assign (use_stmt) |
2913 | && gimple_assign_rhs_code (use_stmt) == NEGATE_EXPR) | |
a5f09e73 | 2914 | { |
a758fd67 | 2915 | ssa_op_iter iter; |
dae957ae | 2916 | use_operand_p usep; |
a758fd67 | 2917 | |
a5f09e73 RH |
2918 | result = gimple_assign_lhs (use_stmt); |
2919 | ||
2920 | /* Make sure the negate statement becomes dead with this | |
2921 | single transformation. */ | |
2922 | if (!single_imm_use (gimple_assign_lhs (use_stmt), | |
2923 | &use_p, &neguse_stmt)) | |
2924 | return false; | |
2925 | ||
a758fd67 | 2926 | /* Make sure the multiplication isn't also used on that stmt. */ |
dae957ae RG |
2927 | FOR_EACH_PHI_OR_STMT_USE (usep, neguse_stmt, iter, SSA_OP_USE) |
2928 | if (USE_FROM_PTR (usep) == mul_result) | |
a758fd67 RG |
2929 | return false; |
2930 | ||
a5f09e73 RH |
2931 | /* Re-validate. */ |
2932 | use_stmt = neguse_stmt; | |
2933 | if (gimple_bb (use_stmt) != gimple_bb (mul_stmt)) | |
2934 | return false; | |
a5f09e73 | 2935 | |
a5f09e73 RH |
2936 | negate_p = true; |
2937 | } | |
16949072 | 2938 | |
0936858f RS |
2939 | tree cond, else_value, ops[3]; |
2940 | tree_code code; | |
2941 | if (!can_interpret_as_conditional_op_p (use_stmt, &cond, &code, ops, | |
2942 | &else_value)) | |
2943 | return false; | |
2944 | ||
2945 | switch (code) | |
a5f09e73 RH |
2946 | { |
2947 | case MINUS_EXPR: | |
0936858f | 2948 | if (ops[1] == result) |
a1d8aa4b RH |
2949 | negate_p = !negate_p; |
2950 | break; | |
a5f09e73 | 2951 | case PLUS_EXPR: |
a5f09e73 | 2952 | break; |
a5f09e73 RH |
2953 | default: |
2954 | /* FMA can only be formed from PLUS and MINUS. */ | |
2955 | return false; | |
2956 | } | |
16949072 | 2957 | |
0936858f RS |
2958 | if (cond) |
2959 | { | |
2960 | if (cond == result || else_value == result) | |
2961 | return false; | |
2962 | if (!direct_internal_fn_supported_p (IFN_COND_FMA, type, opt_type)) | |
2963 | return false; | |
2964 | } | |
2965 | ||
2966 | /* If the subtrahend (OPS[1]) is computed by a MULT_EXPR that | |
2967 | we'll visit later, we might be able to get a more profitable | |
2968 | match with fnma. | |
ee8a9b7b JR |
2969 | OTOH, if we don't, a negate / fma pair has likely lower latency |
2970 | that a mult / subtract pair. */ | |
0936858f RS |
2971 | if (code == MINUS_EXPR |
2972 | && !negate_p | |
2973 | && ops[0] == result | |
c566cc9f | 2974 | && !direct_internal_fn_supported_p (IFN_FMS, type, opt_type) |
0936858f RS |
2975 | && direct_internal_fn_supported_p (IFN_FNMA, type, opt_type) |
2976 | && TREE_CODE (ops[1]) == SSA_NAME | |
2977 | && has_single_use (ops[1])) | |
ee8a9b7b | 2978 | { |
0936858f RS |
2979 | gimple *stmt2 = SSA_NAME_DEF_STMT (ops[1]); |
2980 | if (is_gimple_assign (stmt2) | |
2981 | && gimple_assign_rhs_code (stmt2) == MULT_EXPR) | |
2982 | return false; | |
ee8a9b7b JR |
2983 | } |
2984 | ||
a5f09e73 | 2985 | /* We can't handle a * b + a * b. */ |
0936858f | 2986 | if (ops[0] == ops[1]) |
4a0d0ed2 MJ |
2987 | return false; |
2988 | /* If deferring, make sure we are not looking at an instruction that | |
2989 | wouldn't have existed if we were not. */ | |
2990 | if (state->m_deferring_p | |
0936858f RS |
2991 | && (state->m_mul_result_set.contains (ops[0]) |
2992 | || state->m_mul_result_set.contains (ops[1]))) | |
a5f09e73 | 2993 | return false; |
a1d8aa4b | 2994 | |
4a0d0ed2 | 2995 | if (check_defer) |
a5f09e73 | 2996 | { |
0936858f | 2997 | tree use_lhs = gimple_get_lhs (use_stmt); |
4a0d0ed2 MJ |
2998 | if (state->m_last_result) |
2999 | { | |
0936858f RS |
3000 | if (ops[1] == state->m_last_result |
3001 | || ops[0] == state->m_last_result) | |
4a0d0ed2 MJ |
3002 | defer = true; |
3003 | else | |
3004 | defer = false; | |
3005 | } | |
3006 | else | |
3007 | { | |
3008 | gcc_checking_assert (!state->m_initial_phi); | |
3009 | gphi *phi; | |
0936858f RS |
3010 | if (ops[0] == result) |
3011 | phi = result_of_phi (ops[1]); | |
4a0d0ed2 MJ |
3012 | else |
3013 | { | |
0936858f RS |
3014 | gcc_assert (ops[1] == result); |
3015 | phi = result_of_phi (ops[0]); | |
4a0d0ed2 | 3016 | } |
a5f09e73 | 3017 | |
4a0d0ed2 MJ |
3018 | if (phi) |
3019 | { | |
3020 | state->m_initial_phi = phi; | |
3021 | defer = true; | |
3022 | } | |
3023 | else | |
3024 | defer = false; | |
3025 | } | |
a5f09e73 | 3026 | |
4a0d0ed2 MJ |
3027 | state->m_last_result = use_lhs; |
3028 | check_defer = false; | |
16949072 RG |
3029 | } |
3030 | else | |
4a0d0ed2 MJ |
3031 | defer = false; |
3032 | ||
3033 | /* While it is possible to validate whether or not the exact form that | |
3034 | we've recognized is available in the backend, the assumption is that | |
3035 | if the deferring logic above did not trigger, the transformation is | |
3036 | never a loss. For instance, suppose the target only has the plain FMA | |
3037 | pattern available. Consider a*b-c -> fma(a,b,-c): we've exchanged | |
3038 | MUL+SUB for FMA+NEG, which is still two operations. Consider | |
3039 | -(a*b)-c -> fma(-a,b,-c): we still have 3 operations, but in the FMA | |
3040 | form the two NEGs are independent and could be run in parallel. */ | |
3041 | } | |
3042 | ||
3043 | if (defer) | |
3044 | { | |
3045 | fma_transformation_info fti; | |
3046 | fti.mul_stmt = mul_stmt; | |
3047 | fti.mul_result = mul_result; | |
3048 | fti.op1 = op1; | |
3049 | fti.op2 = op2; | |
3050 | state->m_candidates.safe_push (fti); | |
3051 | state->m_mul_result_set.add (mul_result); | |
3052 | ||
3053 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
16949072 | 3054 | { |
4a0d0ed2 | 3055 | fprintf (dump_file, "Deferred generating FMA for multiplication "); |
4af78ef8 | 3056 | print_gimple_stmt (dump_file, mul_stmt, 0, TDF_NONE); |
4a0d0ed2 | 3057 | fprintf (dump_file, "\n"); |
16949072 RG |
3058 | } |
3059 | ||
4a0d0ed2 MJ |
3060 | return false; |
3061 | } | |
3062 | else | |
3063 | { | |
3064 | if (state->m_deferring_p) | |
3065 | cancel_fma_deferring (state); | |
3066 | convert_mult_to_fma_1 (mul_result, op1, op2); | |
3067 | return true; | |
16949072 | 3068 | } |
16949072 RG |
3069 | } |
3070 | ||
6837d899 JJ |
3071 | |
3072 | /* Helper function of match_uaddsub_overflow. Return 1 | |
3073 | if USE_STMT is unsigned overflow check ovf != 0 for | |
3074 | STMT, -1 if USE_STMT is unsigned overflow check ovf == 0 | |
3075 | and 0 otherwise. */ | |
3076 | ||
3077 | static int | |
3078 | uaddsub_overflow_check_p (gimple *stmt, gimple *use_stmt) | |
3079 | { | |
3080 | enum tree_code ccode = ERROR_MARK; | |
3081 | tree crhs1 = NULL_TREE, crhs2 = NULL_TREE; | |
3082 | if (gimple_code (use_stmt) == GIMPLE_COND) | |
3083 | { | |
3084 | ccode = gimple_cond_code (use_stmt); | |
3085 | crhs1 = gimple_cond_lhs (use_stmt); | |
3086 | crhs2 = gimple_cond_rhs (use_stmt); | |
3087 | } | |
3088 | else if (is_gimple_assign (use_stmt)) | |
3089 | { | |
3090 | if (gimple_assign_rhs_class (use_stmt) == GIMPLE_BINARY_RHS) | |
3091 | { | |
3092 | ccode = gimple_assign_rhs_code (use_stmt); | |
3093 | crhs1 = gimple_assign_rhs1 (use_stmt); | |
3094 | crhs2 = gimple_assign_rhs2 (use_stmt); | |
3095 | } | |
3096 | else if (gimple_assign_rhs_code (use_stmt) == COND_EXPR) | |
3097 | { | |
3098 | tree cond = gimple_assign_rhs1 (use_stmt); | |
3099 | if (COMPARISON_CLASS_P (cond)) | |
3100 | { | |
3101 | ccode = TREE_CODE (cond); | |
3102 | crhs1 = TREE_OPERAND (cond, 0); | |
3103 | crhs2 = TREE_OPERAND (cond, 1); | |
3104 | } | |
3105 | else | |
3106 | return 0; | |
3107 | } | |
3108 | else | |
3109 | return 0; | |
3110 | } | |
3111 | else | |
3112 | return 0; | |
3113 | ||
3114 | if (TREE_CODE_CLASS (ccode) != tcc_comparison) | |
3115 | return 0; | |
3116 | ||
3117 | enum tree_code code = gimple_assign_rhs_code (stmt); | |
3118 | tree lhs = gimple_assign_lhs (stmt); | |
3119 | tree rhs1 = gimple_assign_rhs1 (stmt); | |
3120 | tree rhs2 = gimple_assign_rhs2 (stmt); | |
3121 | ||
3122 | switch (ccode) | |
3123 | { | |
3124 | case GT_EXPR: | |
3125 | case LE_EXPR: | |
3126 | /* r = a - b; r > a or r <= a | |
3127 | r = a + b; a > r or a <= r or b > r or b <= r. */ | |
3128 | if ((code == MINUS_EXPR && crhs1 == lhs && crhs2 == rhs1) | |
3129 | || (code == PLUS_EXPR && (crhs1 == rhs1 || crhs1 == rhs2) | |
3130 | && crhs2 == lhs)) | |
3131 | return ccode == GT_EXPR ? 1 : -1; | |
3132 | break; | |
3133 | case LT_EXPR: | |
3134 | case GE_EXPR: | |
3135 | /* r = a - b; a < r or a >= r | |
3136 | r = a + b; r < a or r >= a or r < b or r >= b. */ | |
3137 | if ((code == MINUS_EXPR && crhs1 == rhs1 && crhs2 == lhs) | |
3138 | || (code == PLUS_EXPR && crhs1 == lhs | |
3139 | && (crhs2 == rhs1 || crhs2 == rhs2))) | |
3140 | return ccode == LT_EXPR ? 1 : -1; | |
3141 | break; | |
3142 | default: | |
3143 | break; | |
3144 | } | |
3145 | return 0; | |
3146 | } | |
3147 | ||
3148 | /* Recognize for unsigned x | |
3149 | x = y - z; | |
3150 | if (x > y) | |
3151 | where there are other uses of x and replace it with | |
3152 | _7 = SUB_OVERFLOW (y, z); | |
3153 | x = REALPART_EXPR <_7>; | |
3154 | _8 = IMAGPART_EXPR <_7>; | |
3155 | if (_8) | |
3156 | and similarly for addition. */ | |
3157 | ||
3158 | static bool | |
3159 | match_uaddsub_overflow (gimple_stmt_iterator *gsi, gimple *stmt, | |
3160 | enum tree_code code) | |
3161 | { | |
3162 | tree lhs = gimple_assign_lhs (stmt); | |
3163 | tree type = TREE_TYPE (lhs); | |
3164 | use_operand_p use_p; | |
3165 | imm_use_iterator iter; | |
3166 | bool use_seen = false; | |
3167 | bool ovf_use_seen = false; | |
3168 | gimple *use_stmt; | |
3169 | ||
3170 | gcc_checking_assert (code == PLUS_EXPR || code == MINUS_EXPR); | |
3171 | if (!INTEGRAL_TYPE_P (type) | |
3172 | || !TYPE_UNSIGNED (type) | |
3173 | || has_zero_uses (lhs) | |
3174 | || has_single_use (lhs) | |
3175 | || optab_handler (code == PLUS_EXPR ? uaddv4_optab : usubv4_optab, | |
3176 | TYPE_MODE (type)) == CODE_FOR_nothing) | |
3177 | return false; | |
3178 | ||
3179 | FOR_EACH_IMM_USE_FAST (use_p, iter, lhs) | |
3180 | { | |
3181 | use_stmt = USE_STMT (use_p); | |
3182 | if (is_gimple_debug (use_stmt)) | |
3183 | continue; | |
3184 | ||
3185 | if (uaddsub_overflow_check_p (stmt, use_stmt)) | |
3186 | ovf_use_seen = true; | |
3187 | else | |
3188 | use_seen = true; | |
3189 | if (ovf_use_seen && use_seen) | |
3190 | break; | |
3191 | } | |
3192 | ||
3193 | if (!ovf_use_seen || !use_seen) | |
3194 | return false; | |
3195 | ||
3196 | tree ctype = build_complex_type (type); | |
3197 | tree rhs1 = gimple_assign_rhs1 (stmt); | |
3198 | tree rhs2 = gimple_assign_rhs2 (stmt); | |
3199 | gcall *g = gimple_build_call_internal (code == PLUS_EXPR | |
3200 | ? IFN_ADD_OVERFLOW : IFN_SUB_OVERFLOW, | |
3201 | 2, rhs1, rhs2); | |
3202 | tree ctmp = make_ssa_name (ctype); | |
3203 | gimple_call_set_lhs (g, ctmp); | |
3204 | gsi_insert_before (gsi, g, GSI_SAME_STMT); | |
3205 | gassign *g2 = gimple_build_assign (lhs, REALPART_EXPR, | |
3206 | build1 (REALPART_EXPR, type, ctmp)); | |
3207 | gsi_replace (gsi, g2, true); | |
3208 | tree ovf = make_ssa_name (type); | |
3209 | g2 = gimple_build_assign (ovf, IMAGPART_EXPR, | |
3210 | build1 (IMAGPART_EXPR, type, ctmp)); | |
3211 | gsi_insert_after (gsi, g2, GSI_NEW_STMT); | |
3212 | ||
3213 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs) | |
3214 | { | |
3215 | if (is_gimple_debug (use_stmt)) | |
3216 | continue; | |
3217 | ||
3218 | int ovf_use = uaddsub_overflow_check_p (stmt, use_stmt); | |
3219 | if (ovf_use == 0) | |
3220 | continue; | |
3221 | if (gimple_code (use_stmt) == GIMPLE_COND) | |
3222 | { | |
3223 | gcond *cond_stmt = as_a <gcond *> (use_stmt); | |
3224 | gimple_cond_set_lhs (cond_stmt, ovf); | |
3225 | gimple_cond_set_rhs (cond_stmt, build_int_cst (type, 0)); | |
3226 | gimple_cond_set_code (cond_stmt, ovf_use == 1 ? NE_EXPR : EQ_EXPR); | |
3227 | } | |
3228 | else | |
3229 | { | |
3230 | gcc_checking_assert (is_gimple_assign (use_stmt)); | |
3231 | if (gimple_assign_rhs_class (use_stmt) == GIMPLE_BINARY_RHS) | |
3232 | { | |
3233 | gimple_assign_set_rhs1 (use_stmt, ovf); | |
3234 | gimple_assign_set_rhs2 (use_stmt, build_int_cst (type, 0)); | |
3235 | gimple_assign_set_rhs_code (use_stmt, | |
3236 | ovf_use == 1 ? NE_EXPR : EQ_EXPR); | |
3237 | } | |
3238 | else | |
3239 | { | |
3240 | gcc_checking_assert (gimple_assign_rhs_code (use_stmt) | |
3241 | == COND_EXPR); | |
3242 | tree cond = build2 (ovf_use == 1 ? NE_EXPR : EQ_EXPR, | |
3243 | boolean_type_node, ovf, | |
3244 | build_int_cst (type, 0)); | |
3245 | gimple_assign_set_rhs1 (use_stmt, cond); | |
3246 | } | |
3247 | } | |
3248 | update_stmt (use_stmt); | |
3249 | } | |
3250 | return true; | |
3251 | } | |
3252 | ||
e72531b9 PK |
3253 | /* Return true if target has support for divmod. */ |
3254 | ||
3255 | static bool | |
3256 | target_supports_divmod_p (optab divmod_optab, optab div_optab, machine_mode mode) | |
3257 | { | |
3258 | /* If target supports hardware divmod insn, use it for divmod. */ | |
3259 | if (optab_handler (divmod_optab, mode) != CODE_FOR_nothing) | |
3260 | return true; | |
3261 | ||
3262 | /* Check if libfunc for divmod is available. */ | |
3263 | rtx libfunc = optab_libfunc (divmod_optab, mode); | |
3264 | if (libfunc != NULL_RTX) | |
3265 | { | |
3266 | /* If optab_handler exists for div_optab, perhaps in a wider mode, | |
3267 | we don't want to use the libfunc even if it exists for given mode. */ | |
c94843d2 RS |
3268 | machine_mode div_mode; |
3269 | FOR_EACH_MODE_FROM (div_mode, mode) | |
e72531b9 PK |
3270 | if (optab_handler (div_optab, div_mode) != CODE_FOR_nothing) |
3271 | return false; | |
3272 | ||
3273 | return targetm.expand_divmod_libfunc != NULL; | |
3274 | } | |
3275 | ||
3276 | return false; | |
3277 | } | |
3278 | ||
3279 | /* Check if stmt is candidate for divmod transform. */ | |
3280 | ||
3281 | static bool | |
3282 | divmod_candidate_p (gassign *stmt) | |
3283 | { | |
3284 | tree type = TREE_TYPE (gimple_assign_lhs (stmt)); | |
b8506a8a | 3285 | machine_mode mode = TYPE_MODE (type); |
e72531b9 PK |
3286 | optab divmod_optab, div_optab; |
3287 | ||
3288 | if (TYPE_UNSIGNED (type)) | |
3289 | { | |
3290 | divmod_optab = udivmod_optab; | |
3291 | div_optab = udiv_optab; | |
3292 | } | |
3293 | else | |
3294 | { | |
3295 | divmod_optab = sdivmod_optab; | |
3296 | div_optab = sdiv_optab; | |
3297 | } | |
3298 | ||
3299 | tree op1 = gimple_assign_rhs1 (stmt); | |
3300 | tree op2 = gimple_assign_rhs2 (stmt); | |
3301 | ||
3302 | /* Disable the transform if either is a constant, since division-by-constant | |
3303 | may have specialized expansion. */ | |
3304 | if (CONSTANT_CLASS_P (op1) || CONSTANT_CLASS_P (op2)) | |
3305 | return false; | |
3306 | ||
3307 | /* Exclude the case where TYPE_OVERFLOW_TRAPS (type) as that should | |
3308 | expand using the [su]divv optabs. */ | |
3309 | if (TYPE_OVERFLOW_TRAPS (type)) | |
3310 | return false; | |
3311 | ||
3312 | if (!target_supports_divmod_p (divmod_optab, div_optab, mode)) | |
3313 | return false; | |
3314 | ||
3315 | return true; | |
3316 | } | |
3317 | ||
3318 | /* This function looks for: | |
3319 | t1 = a TRUNC_DIV_EXPR b; | |
3320 | t2 = a TRUNC_MOD_EXPR b; | |
3321 | and transforms it to the following sequence: | |
3322 | complex_tmp = DIVMOD (a, b); | |
3323 | t1 = REALPART_EXPR(a); | |
3324 | t2 = IMAGPART_EXPR(b); | |
3325 | For conditions enabling the transform see divmod_candidate_p(). | |
3326 | ||
3327 | The pass has three parts: | |
3328 | 1) Find top_stmt which is trunc_div or trunc_mod stmt and dominates all | |
3329 | other trunc_div_expr and trunc_mod_expr stmts. | |
3330 | 2) Add top_stmt and all trunc_div and trunc_mod stmts dominated by top_stmt | |
3331 | to stmts vector. | |
3332 | 3) Insert DIVMOD call just before top_stmt and update entries in | |
3333 | stmts vector to use return value of DIMOVD (REALEXPR_PART for div, | |
3334 | IMAGPART_EXPR for mod). */ | |
3335 | ||
3336 | static bool | |
3337 | convert_to_divmod (gassign *stmt) | |
3338 | { | |
3339 | if (stmt_can_throw_internal (stmt) | |
3340 | || !divmod_candidate_p (stmt)) | |
3341 | return false; | |
3342 | ||
3343 | tree op1 = gimple_assign_rhs1 (stmt); | |
3344 | tree op2 = gimple_assign_rhs2 (stmt); | |
3345 | ||
3346 | imm_use_iterator use_iter; | |
3347 | gimple *use_stmt; | |
3348 | auto_vec<gimple *> stmts; | |
3349 | ||
3350 | gimple *top_stmt = stmt; | |
3351 | basic_block top_bb = gimple_bb (stmt); | |
3352 | ||
3353 | /* Part 1: Try to set top_stmt to "topmost" stmt that dominates | |
3354 | at-least stmt and possibly other trunc_div/trunc_mod stmts | |
3355 | having same operands as stmt. */ | |
3356 | ||
3357 | FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, op1) | |
3358 | { | |
3359 | if (is_gimple_assign (use_stmt) | |
3360 | && (gimple_assign_rhs_code (use_stmt) == TRUNC_DIV_EXPR | |
3361 | || gimple_assign_rhs_code (use_stmt) == TRUNC_MOD_EXPR) | |
3362 | && operand_equal_p (op1, gimple_assign_rhs1 (use_stmt), 0) | |
3363 | && operand_equal_p (op2, gimple_assign_rhs2 (use_stmt), 0)) | |
3364 | { | |
3365 | if (stmt_can_throw_internal (use_stmt)) | |
3366 | continue; | |
3367 | ||
3368 | basic_block bb = gimple_bb (use_stmt); | |
3369 | ||
3370 | if (bb == top_bb) | |
3371 | { | |
3372 | if (gimple_uid (use_stmt) < gimple_uid (top_stmt)) | |
3373 | top_stmt = use_stmt; | |
3374 | } | |
3375 | else if (dominated_by_p (CDI_DOMINATORS, top_bb, bb)) | |
3376 | { | |
3377 | top_bb = bb; | |
3378 | top_stmt = use_stmt; | |
3379 | } | |
3380 | } | |
3381 | } | |
3382 | ||
3383 | tree top_op1 = gimple_assign_rhs1 (top_stmt); | |
3384 | tree top_op2 = gimple_assign_rhs2 (top_stmt); | |
3385 | ||
3386 | stmts.safe_push (top_stmt); | |
3387 | bool div_seen = (gimple_assign_rhs_code (top_stmt) == TRUNC_DIV_EXPR); | |
3388 | ||
3389 | /* Part 2: Add all trunc_div/trunc_mod statements domianted by top_bb | |
3390 | to stmts vector. The 2nd loop will always add stmt to stmts vector, since | |
3391 | gimple_bb (top_stmt) dominates gimple_bb (stmt), so the | |
3392 | 2nd loop ends up adding at-least single trunc_mod_expr stmt. */ | |
3393 | ||
3394 | FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, top_op1) | |
3395 | { | |
3396 | if (is_gimple_assign (use_stmt) | |
3397 | && (gimple_assign_rhs_code (use_stmt) == TRUNC_DIV_EXPR | |
3398 | || gimple_assign_rhs_code (use_stmt) == TRUNC_MOD_EXPR) | |
3399 | && operand_equal_p (top_op1, gimple_assign_rhs1 (use_stmt), 0) | |
3400 | && operand_equal_p (top_op2, gimple_assign_rhs2 (use_stmt), 0)) | |
3401 | { | |
3402 | if (use_stmt == top_stmt | |
3403 | || stmt_can_throw_internal (use_stmt) | |
3404 | || !dominated_by_p (CDI_DOMINATORS, gimple_bb (use_stmt), top_bb)) | |
3405 | continue; | |
3406 | ||
3407 | stmts.safe_push (use_stmt); | |
3408 | if (gimple_assign_rhs_code (use_stmt) == TRUNC_DIV_EXPR) | |
3409 | div_seen = true; | |
3410 | } | |
3411 | } | |
3412 | ||
3413 | if (!div_seen) | |
3414 | return false; | |
3415 | ||
3416 | /* Part 3: Create libcall to internal fn DIVMOD: | |
3417 | divmod_tmp = DIVMOD (op1, op2). */ | |
3418 | ||
3419 | gcall *call_stmt = gimple_build_call_internal (IFN_DIVMOD, 2, op1, op2); | |
3420 | tree res = make_temp_ssa_name (build_complex_type (TREE_TYPE (op1)), | |
3421 | call_stmt, "divmod_tmp"); | |
3422 | gimple_call_set_lhs (call_stmt, res); | |
a844293d RS |
3423 | /* We rejected throwing statements above. */ |
3424 | gimple_call_set_nothrow (call_stmt, true); | |
e72531b9 PK |
3425 | |
3426 | /* Insert the call before top_stmt. */ | |
3427 | gimple_stmt_iterator top_stmt_gsi = gsi_for_stmt (top_stmt); | |
3428 | gsi_insert_before (&top_stmt_gsi, call_stmt, GSI_SAME_STMT); | |
3429 | ||
3430 | widen_mul_stats.divmod_calls_inserted++; | |
3431 | ||
3432 | /* Update all statements in stmts vector: | |
3433 | lhs = op1 TRUNC_DIV_EXPR op2 -> lhs = REALPART_EXPR<divmod_tmp> | |
3434 | lhs = op1 TRUNC_MOD_EXPR op2 -> lhs = IMAGPART_EXPR<divmod_tmp>. */ | |
3435 | ||
3436 | for (unsigned i = 0; stmts.iterate (i, &use_stmt); ++i) | |
3437 | { | |
3438 | tree new_rhs; | |
3439 | ||
3440 | switch (gimple_assign_rhs_code (use_stmt)) | |
3441 | { | |
3442 | case TRUNC_DIV_EXPR: | |
3443 | new_rhs = fold_build1 (REALPART_EXPR, TREE_TYPE (op1), res); | |
3444 | break; | |
3445 | ||
3446 | case TRUNC_MOD_EXPR: | |
3447 | new_rhs = fold_build1 (IMAGPART_EXPR, TREE_TYPE (op1), res); | |
3448 | break; | |
3449 | ||
3450 | default: | |
3451 | gcc_unreachable (); | |
3452 | } | |
3453 | ||
3454 | gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt); | |
3455 | gimple_assign_set_rhs_from_tree (&gsi, new_rhs); | |
3456 | update_stmt (use_stmt); | |
3457 | } | |
3458 | ||
3459 | return true; | |
3460 | } | |
6837d899 | 3461 | |
5b58b39b BS |
3462 | /* Find integer multiplications where the operands are extended from |
3463 | smaller types, and replace the MULT_EXPR with a WIDEN_MULT_EXPR | |
3464 | where appropriate. */ | |
3465 | ||
be55bfe6 TS |
3466 | namespace { |
3467 | ||
3468 | const pass_data pass_data_optimize_widening_mul = | |
3469 | { | |
3470 | GIMPLE_PASS, /* type */ | |
3471 | "widening_mul", /* name */ | |
3472 | OPTGROUP_NONE, /* optinfo_flags */ | |
1b6546cc | 3473 | TV_TREE_WIDEN_MUL, /* tv_id */ |
be55bfe6 TS |
3474 | PROP_ssa, /* properties_required */ |
3475 | 0, /* properties_provided */ | |
3476 | 0, /* properties_destroyed */ | |
3477 | 0, /* todo_flags_start */ | |
3bea341f | 3478 | TODO_update_ssa, /* todo_flags_finish */ |
be55bfe6 TS |
3479 | }; |
3480 | ||
3481 | class pass_optimize_widening_mul : public gimple_opt_pass | |
3482 | { | |
3483 | public: | |
3484 | pass_optimize_widening_mul (gcc::context *ctxt) | |
3485 | : gimple_opt_pass (pass_data_optimize_widening_mul, ctxt) | |
3486 | {} | |
3487 | ||
3488 | /* opt_pass methods: */ | |
3489 | virtual bool gate (function *) | |
3490 | { | |
3491 | return flag_expensive_optimizations && optimize; | |
3492 | } | |
3493 | ||
3494 | virtual unsigned int execute (function *); | |
3495 | ||
3496 | }; // class pass_optimize_widening_mul | |
3497 | ||
4a0d0ed2 MJ |
3498 | /* Walker class to perform the transformation in reverse dominance order. */ |
3499 | ||
3500 | class math_opts_dom_walker : public dom_walker | |
5b58b39b | 3501 | { |
4a0d0ed2 MJ |
3502 | public: |
3503 | /* Constructor, CFG_CHANGED is a pointer to a boolean flag that will be set | |
3504 | if walking modidifes the CFG. */ | |
5b58b39b | 3505 | |
4a0d0ed2 MJ |
3506 | math_opts_dom_walker (bool *cfg_changed_p) |
3507 | : dom_walker (CDI_DOMINATORS), m_last_result_set (), | |
3508 | m_cfg_changed_p (cfg_changed_p) {} | |
4da3b811 | 3509 | |
4a0d0ed2 MJ |
3510 | /* The actual actions performed in the walk. */ |
3511 | ||
3512 | virtual void after_dom_children (basic_block); | |
3513 | ||
3514 | /* Set of results of chains of multiply and add statement combinations that | |
3515 | were not transformed into FMAs because of active deferring. */ | |
3516 | hash_set<tree> m_last_result_set; | |
3517 | ||
3518 | /* Pointer to a flag of the user that needs to be set if CFG has been | |
3519 | modified. */ | |
3520 | bool *m_cfg_changed_p; | |
3521 | }; | |
3522 | ||
3523 | void | |
3524 | math_opts_dom_walker::after_dom_children (basic_block bb) | |
3525 | { | |
3526 | gimple_stmt_iterator gsi; | |
3527 | ||
3528 | fma_deferring_state fma_state (PARAM_VALUE (PARAM_AVOID_FMA_MAX_BITS) > 0); | |
3529 | ||
3530 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi);) | |
5b58b39b | 3531 | { |
4a0d0ed2 MJ |
3532 | gimple *stmt = gsi_stmt (gsi); |
3533 | enum tree_code code; | |
5b58b39b | 3534 | |
4a0d0ed2 MJ |
3535 | if (is_gimple_assign (stmt)) |
3536 | { | |
3537 | code = gimple_assign_rhs_code (stmt); | |
3538 | switch (code) | |
3539 | { | |
3540 | case MULT_EXPR: | |
3541 | if (!convert_mult_to_widen (stmt, &gsi) | |
3542 | && !convert_expand_mult_copysign (stmt, &gsi) | |
3543 | && convert_mult_to_fma (stmt, | |
3544 | gimple_assign_rhs1 (stmt), | |
3545 | gimple_assign_rhs2 (stmt), | |
3546 | &fma_state)) | |
3547 | { | |
3548 | gsi_remove (&gsi, true); | |
3549 | release_defs (stmt); | |
3550 | continue; | |
3551 | } | |
3552 | break; | |
3553 | ||
3554 | case PLUS_EXPR: | |
3555 | case MINUS_EXPR: | |
3556 | if (!convert_plusminus_to_widen (&gsi, stmt, code)) | |
3557 | match_uaddsub_overflow (&gsi, stmt, code); | |
3558 | break; | |
5b58b39b | 3559 | |
4a0d0ed2 MJ |
3560 | case TRUNC_MOD_EXPR: |
3561 | convert_to_divmod (as_a<gassign *> (stmt)); | |
3562 | break; | |
3563 | ||
3564 | default:; | |
3565 | } | |
3566 | } | |
3567 | else if (is_gimple_call (stmt)) | |
3568 | { | |
3569 | tree fndecl = gimple_call_fndecl (stmt); | |
3570 | if (fndecl && gimple_call_builtin_p (stmt, BUILT_IN_NORMAL)) | |
16949072 | 3571 | { |
4a0d0ed2 | 3572 | switch (DECL_FUNCTION_CODE (fndecl)) |
16949072 | 3573 | { |
4a0d0ed2 MJ |
3574 | case BUILT_IN_POWF: |
3575 | case BUILT_IN_POW: | |
3576 | case BUILT_IN_POWL: | |
3577 | if (gimple_call_lhs (stmt) | |
3578 | && TREE_CODE (gimple_call_arg (stmt, 1)) == REAL_CST | |
3579 | && real_equal | |
3580 | (&TREE_REAL_CST (gimple_call_arg (stmt, 1)), | |
3581 | &dconst2) | |
4dbed5f6 | 3582 | && convert_mult_to_fma (stmt, |
4a0d0ed2 MJ |
3583 | gimple_call_arg (stmt, 0), |
3584 | gimple_call_arg (stmt, 0), | |
3585 | &fma_state)) | |
16949072 | 3586 | { |
4a0d0ed2 MJ |
3587 | unlink_stmt_vdef (stmt); |
3588 | if (gsi_remove (&gsi, true) | |
3589 | && gimple_purge_dead_eh_edges (bb)) | |
3590 | *m_cfg_changed_p = true; | |
16949072 RG |
3591 | release_defs (stmt); |
3592 | continue; | |
3593 | } | |
3594 | break; | |
3595 | ||
16949072 RG |
3596 | default:; |
3597 | } | |
3598 | } | |
4a0d0ed2 MJ |
3599 | else |
3600 | cancel_fma_deferring (&fma_state); | |
5b58b39b | 3601 | } |
4a0d0ed2 | 3602 | gsi_next (&gsi); |
5b58b39b | 3603 | } |
4a0d0ed2 MJ |
3604 | if (fma_state.m_deferring_p |
3605 | && fma_state.m_initial_phi) | |
3606 | { | |
3607 | gcc_checking_assert (fma_state.m_last_result); | |
3608 | if (!last_fma_candidate_feeds_initial_phi (&fma_state, | |
3609 | &m_last_result_set)) | |
3610 | cancel_fma_deferring (&fma_state); | |
3611 | else | |
3612 | m_last_result_set.add (fma_state.m_last_result); | |
3613 | } | |
3614 | } | |
3615 | ||
3616 | ||
3617 | unsigned int | |
3618 | pass_optimize_widening_mul::execute (function *fun) | |
3619 | { | |
3620 | bool cfg_changed = false; | |
3621 | ||
3622 | memset (&widen_mul_stats, 0, sizeof (widen_mul_stats)); | |
3623 | calculate_dominance_info (CDI_DOMINATORS); | |
3624 | renumber_gimple_stmt_uids (); | |
3625 | ||
3626 | math_opts_dom_walker (&cfg_changed).walk (ENTRY_BLOCK_PTR_FOR_FN (cfun)); | |
0354c0c7 | 3627 | |
be55bfe6 | 3628 | statistics_counter_event (fun, "widening multiplications inserted", |
4da3b811 | 3629 | widen_mul_stats.widen_mults_inserted); |
be55bfe6 | 3630 | statistics_counter_event (fun, "widening maccs inserted", |
4da3b811 | 3631 | widen_mul_stats.maccs_inserted); |
be55bfe6 | 3632 | statistics_counter_event (fun, "fused multiply-adds inserted", |
4da3b811 | 3633 | widen_mul_stats.fmas_inserted); |
e72531b9 PK |
3634 | statistics_counter_event (fun, "divmod calls inserted", |
3635 | widen_mul_stats.divmod_calls_inserted); | |
4da3b811 | 3636 | |
4dbed5f6 | 3637 | return cfg_changed ? TODO_cleanup_cfg : 0; |
5b58b39b BS |
3638 | } |
3639 | ||
27a4cd48 DM |
3640 | } // anon namespace |
3641 | ||
3642 | gimple_opt_pass * | |
3643 | make_pass_optimize_widening_mul (gcc::context *ctxt) | |
3644 | { | |
3645 | return new pass_optimize_widening_mul (ctxt); | |
3646 | } |