]>
Commit | Line | Data |
---|---|---|
ebfd146a IR |
1 | /* Statement Analysis and Transformation for Vectorization |
2 | Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009 Free Software | |
3 | Foundation, Inc. | |
4 | Contributed by Dorit Naishlos <dorit@il.ibm.com> | |
5 | and Ira Rosen <irar@il.ibm.com> | |
6 | ||
7 | This file is part of GCC. | |
8 | ||
9 | GCC is free software; you can redistribute it and/or modify it under | |
10 | the terms of the GNU General Public License as published by the Free | |
11 | Software Foundation; either version 3, or (at your option) any later | |
12 | version. | |
13 | ||
14 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
15 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
16 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
17 | for more details. | |
18 | ||
19 | You should have received a copy of the GNU General Public License | |
20 | along with GCC; see the file COPYING3. If not see | |
21 | <http://www.gnu.org/licenses/>. */ | |
22 | ||
23 | #include "config.h" | |
24 | #include "system.h" | |
25 | #include "coretypes.h" | |
26 | #include "tm.h" | |
27 | #include "ggc.h" | |
28 | #include "tree.h" | |
29 | #include "target.h" | |
30 | #include "basic-block.h" | |
31 | #include "diagnostic.h" | |
32 | #include "tree-flow.h" | |
33 | #include "tree-dump.h" | |
34 | #include "cfgloop.h" | |
35 | #include "cfglayout.h" | |
36 | #include "expr.h" | |
37 | #include "recog.h" | |
38 | #include "optabs.h" | |
39 | #include "toplev.h" | |
40 | #include "tree-vectorizer.h" | |
41 | #include "langhooks.h" | |
42 | ||
43 | ||
44 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ | |
45 | ||
46 | /* Function vect_mark_relevant. | |
47 | ||
48 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
49 | ||
50 | static void | |
51 | vect_mark_relevant (VEC(gimple,heap) **worklist, gimple stmt, | |
52 | enum vect_relevant relevant, bool live_p) | |
53 | { | |
54 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
55 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
56 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
57 | ||
58 | if (vect_print_dump_info (REPORT_DETAILS)) | |
59 | fprintf (vect_dump, "mark relevant %d, live %d.", relevant, live_p); | |
60 | ||
61 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
62 | { | |
63 | gimple pattern_stmt; | |
64 | ||
65 | /* This is the last stmt in a sequence that was detected as a | |
66 | pattern that can potentially be vectorized. Don't mark the stmt | |
67 | as relevant/live because it's not going to be vectorized. | |
68 | Instead mark the pattern-stmt that replaces it. */ | |
69 | ||
70 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
71 | ||
72 | if (vect_print_dump_info (REPORT_DETAILS)) | |
73 | fprintf (vect_dump, "last stmt in pattern. don't mark relevant/live."); | |
74 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
75 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
76 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
77 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
78 | stmt = pattern_stmt; | |
79 | } | |
80 | ||
81 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
82 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
83 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
84 | ||
85 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
86 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
87 | { | |
88 | if (vect_print_dump_info (REPORT_DETAILS)) | |
89 | fprintf (vect_dump, "already marked relevant/live."); | |
90 | return; | |
91 | } | |
92 | ||
93 | VEC_safe_push (gimple, heap, *worklist, stmt); | |
94 | } | |
95 | ||
96 | ||
97 | /* Function vect_stmt_relevant_p. | |
98 | ||
99 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
100 | "relevant for vectorization". | |
101 | ||
102 | A stmt is considered "relevant for vectorization" if: | |
103 | - it has uses outside the loop. | |
104 | - it has vdefs (it alters memory). | |
105 | - control stmts in the loop (except for the exit condition). | |
106 | ||
107 | CHECKME: what other side effects would the vectorizer allow? */ | |
108 | ||
109 | static bool | |
110 | vect_stmt_relevant_p (gimple stmt, loop_vec_info loop_vinfo, | |
111 | enum vect_relevant *relevant, bool *live_p) | |
112 | { | |
113 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
114 | ssa_op_iter op_iter; | |
115 | imm_use_iterator imm_iter; | |
116 | use_operand_p use_p; | |
117 | def_operand_p def_p; | |
118 | ||
8644a673 | 119 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
120 | *live_p = false; |
121 | ||
122 | /* cond stmt other than loop exit cond. */ | |
123 | if (is_ctrl_stmt (stmt) | |
8644a673 IR |
124 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) |
125 | != loop_exit_ctrl_vec_info_type) | |
126 | *relevant = vect_used_in_scope; | |
ebfd146a IR |
127 | |
128 | /* changing memory. */ | |
129 | if (gimple_code (stmt) != GIMPLE_PHI) | |
5006671f | 130 | if (gimple_vdef (stmt)) |
ebfd146a IR |
131 | { |
132 | if (vect_print_dump_info (REPORT_DETAILS)) | |
133 | fprintf (vect_dump, "vec_stmt_relevant_p: stmt has vdefs."); | |
8644a673 | 134 | *relevant = vect_used_in_scope; |
ebfd146a IR |
135 | } |
136 | ||
137 | /* uses outside the loop. */ | |
138 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
139 | { | |
140 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
141 | { | |
142 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
143 | if (!flow_bb_inside_loop_p (loop, bb)) | |
144 | { | |
145 | if (vect_print_dump_info (REPORT_DETAILS)) | |
146 | fprintf (vect_dump, "vec_stmt_relevant_p: used out of loop."); | |
147 | ||
148 | /* We expect all such uses to be in the loop exit phis | |
149 | (because of loop closed form) */ | |
150 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
151 | gcc_assert (bb == single_exit (loop)->dest); | |
152 | ||
153 | *live_p = true; | |
154 | } | |
155 | } | |
156 | } | |
157 | ||
158 | return (*live_p || *relevant); | |
159 | } | |
160 | ||
161 | ||
162 | /* Function exist_non_indexing_operands_for_use_p | |
163 | ||
164 | USE is one of the uses attached to STMT. Check if USE is | |
165 | used in STMT for anything other than indexing an array. */ | |
166 | ||
167 | static bool | |
168 | exist_non_indexing_operands_for_use_p (tree use, gimple stmt) | |
169 | { | |
170 | tree operand; | |
171 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
172 | ||
173 | /* USE corresponds to some operand in STMT. If there is no data | |
174 | reference in STMT, then any operand that corresponds to USE | |
175 | is not indexing an array. */ | |
176 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
177 | return true; | |
178 | ||
179 | /* STMT has a data_ref. FORNOW this means that its of one of | |
180 | the following forms: | |
181 | -1- ARRAY_REF = var | |
182 | -2- var = ARRAY_REF | |
183 | (This should have been verified in analyze_data_refs). | |
184 | ||
185 | 'var' in the second case corresponds to a def, not a use, | |
186 | so USE cannot correspond to any operands that are not used | |
187 | for array indexing. | |
188 | ||
189 | Therefore, all we need to check is if STMT falls into the | |
190 | first case, and whether var corresponds to USE. */ | |
191 | ||
192 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) | |
193 | return false; | |
194 | ||
195 | if (!gimple_assign_copy_p (stmt)) | |
196 | return false; | |
197 | operand = gimple_assign_rhs1 (stmt); | |
198 | ||
199 | if (TREE_CODE (operand) != SSA_NAME) | |
200 | return false; | |
201 | ||
202 | if (operand == use) | |
203 | return true; | |
204 | ||
205 | return false; | |
206 | } | |
207 | ||
208 | ||
209 | /* | |
210 | Function process_use. | |
211 | ||
212 | Inputs: | |
213 | - a USE in STMT in a loop represented by LOOP_VINFO | |
214 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt | |
215 | that defined USE. This is done by calling mark_relevant and passing it | |
216 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). | |
217 | ||
218 | Outputs: | |
219 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
220 | relevance info of the DEF_STMT of this USE: | |
221 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
222 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
223 | Exceptions: | |
224 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
225 | which does not need to be directly vectorized, then the liveness/relevance | |
226 | of the respective DEF_STMT is left unchanged. | |
227 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we | |
228 | skip DEF_STMT cause it had already been processed. | |
229 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will | |
230 | be modified accordingly. | |
231 | ||
232 | Return true if everything is as expected. Return false otherwise. */ | |
233 | ||
234 | static bool | |
235 | process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p, | |
236 | enum vect_relevant relevant, VEC(gimple,heap) **worklist) | |
237 | { | |
238 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
239 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
240 | stmt_vec_info dstmt_vinfo; | |
241 | basic_block bb, def_bb; | |
242 | tree def; | |
243 | gimple def_stmt; | |
244 | enum vect_def_type dt; | |
245 | ||
246 | /* case 1: we are only interested in uses that need to be vectorized. Uses | |
247 | that are used for address computation are not considered relevant. */ | |
248 | if (!exist_non_indexing_operands_for_use_p (use, stmt)) | |
249 | return true; | |
250 | ||
a70d6342 | 251 | if (!vect_is_simple_use (use, loop_vinfo, NULL, &def_stmt, &def, &dt)) |
ebfd146a | 252 | { |
8644a673 | 253 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
ebfd146a IR |
254 | fprintf (vect_dump, "not vectorized: unsupported use in stmt."); |
255 | return false; | |
256 | } | |
257 | ||
258 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
259 | return true; | |
260 | ||
261 | def_bb = gimple_bb (def_stmt); | |
262 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
263 | { | |
264 | if (vect_print_dump_info (REPORT_DETAILS)) | |
265 | fprintf (vect_dump, "def_stmt is out of loop."); | |
266 | return true; | |
267 | } | |
268 | ||
269 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). | |
270 | DEF_STMT must have already been processed, because this should be the | |
271 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
272 | as there should be no other uses for DEF_STMT in the loop. So we just | |
273 | check that everything is as expected, and we are done. */ | |
274 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
275 | bb = gimple_bb (stmt); | |
276 | if (gimple_code (stmt) == GIMPLE_PHI | |
277 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
278 | && gimple_code (def_stmt) != GIMPLE_PHI | |
279 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
280 | && bb->loop_father == def_bb->loop_father) | |
281 | { | |
282 | if (vect_print_dump_info (REPORT_DETAILS)) | |
283 | fprintf (vect_dump, "reduc-stmt defining reduc-phi in the same nest."); | |
284 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) | |
285 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
286 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
287 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) | |
8644a673 | 288 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
289 | return true; |
290 | } | |
291 | ||
292 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
293 | outer-loop-header-bb: | |
294 | d = def_stmt | |
295 | inner-loop: | |
296 | stmt # use (d) | |
297 | outer-loop-tail-bb: | |
298 | ... */ | |
299 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
300 | { | |
301 | if (vect_print_dump_info (REPORT_DETAILS)) | |
302 | fprintf (vect_dump, "outer-loop def-stmt defining inner-loop stmt."); | |
7c5222ff | 303 | |
ebfd146a IR |
304 | switch (relevant) |
305 | { | |
8644a673 | 306 | case vect_unused_in_scope: |
7c5222ff IR |
307 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
308 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 309 | break; |
7c5222ff | 310 | |
ebfd146a | 311 | case vect_used_in_outer_by_reduction: |
7c5222ff | 312 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
313 | relevant = vect_used_by_reduction; |
314 | break; | |
7c5222ff | 315 | |
ebfd146a | 316 | case vect_used_in_outer: |
7c5222ff | 317 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 318 | relevant = vect_used_in_scope; |
ebfd146a | 319 | break; |
7c5222ff | 320 | |
8644a673 | 321 | case vect_used_in_scope: |
ebfd146a IR |
322 | break; |
323 | ||
324 | default: | |
325 | gcc_unreachable (); | |
326 | } | |
327 | } | |
328 | ||
329 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
330 | outer-loop-header-bb: | |
331 | ... | |
332 | inner-loop: | |
333 | d = def_stmt | |
06066f92 | 334 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
335 | stmt # use (d) */ |
336 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
337 | { | |
338 | if (vect_print_dump_info (REPORT_DETAILS)) | |
339 | fprintf (vect_dump, "inner-loop def-stmt defining outer-loop stmt."); | |
7c5222ff | 340 | |
ebfd146a IR |
341 | switch (relevant) |
342 | { | |
8644a673 | 343 | case vect_unused_in_scope: |
06066f92 IR |
344 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
345 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? | |
a70d6342 | 346 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
347 | break; |
348 | ||
ebfd146a IR |
349 | case vect_used_by_reduction: |
350 | relevant = vect_used_in_outer_by_reduction; | |
351 | break; | |
352 | ||
8644a673 | 353 | case vect_used_in_scope: |
ebfd146a IR |
354 | relevant = vect_used_in_outer; |
355 | break; | |
356 | ||
357 | default: | |
358 | gcc_unreachable (); | |
359 | } | |
360 | } | |
361 | ||
362 | vect_mark_relevant (worklist, def_stmt, relevant, live_p); | |
363 | return true; | |
364 | } | |
365 | ||
366 | ||
367 | /* Function vect_mark_stmts_to_be_vectorized. | |
368 | ||
369 | Not all stmts in the loop need to be vectorized. For example: | |
370 | ||
371 | for i... | |
372 | for j... | |
373 | 1. T0 = i + j | |
374 | 2. T1 = a[T0] | |
375 | ||
376 | 3. j = j + 1 | |
377 | ||
378 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
379 | addressing of vectorized data-refs are handled differently. | |
380 | ||
381 | This pass detects such stmts. */ | |
382 | ||
383 | bool | |
384 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
385 | { | |
386 | VEC(gimple,heap) *worklist; | |
387 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
388 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
389 | unsigned int nbbs = loop->num_nodes; | |
390 | gimple_stmt_iterator si; | |
391 | gimple stmt; | |
392 | unsigned int i; | |
393 | stmt_vec_info stmt_vinfo; | |
394 | basic_block bb; | |
395 | gimple phi; | |
396 | bool live_p; | |
06066f92 IR |
397 | enum vect_relevant relevant, tmp_relevant; |
398 | enum vect_def_type def_type; | |
ebfd146a IR |
399 | |
400 | if (vect_print_dump_info (REPORT_DETAILS)) | |
401 | fprintf (vect_dump, "=== vect_mark_stmts_to_be_vectorized ==="); | |
402 | ||
403 | worklist = VEC_alloc (gimple, heap, 64); | |
404 | ||
405 | /* 1. Init worklist. */ | |
406 | for (i = 0; i < nbbs; i++) | |
407 | { | |
408 | bb = bbs[i]; | |
409 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
410 | { | |
411 | phi = gsi_stmt (si); | |
412 | if (vect_print_dump_info (REPORT_DETAILS)) | |
413 | { | |
414 | fprintf (vect_dump, "init: phi relevant? "); | |
415 | print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM); | |
416 | } | |
417 | ||
418 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
419 | vect_mark_relevant (&worklist, phi, relevant, live_p); | |
420 | } | |
421 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
422 | { | |
423 | stmt = gsi_stmt (si); | |
424 | if (vect_print_dump_info (REPORT_DETAILS)) | |
425 | { | |
426 | fprintf (vect_dump, "init: stmt relevant? "); | |
427 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
428 | } | |
429 | ||
430 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
431 | vect_mark_relevant (&worklist, stmt, relevant, live_p); | |
432 | } | |
433 | } | |
434 | ||
435 | /* 2. Process_worklist */ | |
436 | while (VEC_length (gimple, worklist) > 0) | |
437 | { | |
438 | use_operand_p use_p; | |
439 | ssa_op_iter iter; | |
440 | ||
441 | stmt = VEC_pop (gimple, worklist); | |
442 | if (vect_print_dump_info (REPORT_DETAILS)) | |
443 | { | |
444 | fprintf (vect_dump, "worklist: examine stmt: "); | |
445 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
446 | } | |
447 | ||
448 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it | |
449 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
450 | liveness and relevance properties of STMT. */ | |
451 | stmt_vinfo = vinfo_for_stmt (stmt); | |
452 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
453 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
454 | ||
455 | /* Generally, the liveness and relevance properties of STMT are | |
456 | propagated as is to the DEF_STMTs of its USEs: | |
457 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
458 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
459 | ||
460 | One exception is when STMT has been identified as defining a reduction | |
461 | variable; in this case we set the liveness/relevance as follows: | |
462 | live_p = false | |
463 | relevant = vect_used_by_reduction | |
464 | This is because we distinguish between two kinds of relevant stmts - | |
465 | those that are used by a reduction computation, and those that are | |
466 | (also) used by a regular computation. This allows us later on to | |
467 | identify stmts that are used solely by a reduction, and therefore the | |
7c5222ff | 468 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 469 | |
06066f92 IR |
470 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
471 | tmp_relevant = relevant; | |
472 | switch (def_type) | |
ebfd146a | 473 | { |
06066f92 IR |
474 | case vect_reduction_def: |
475 | switch (tmp_relevant) | |
476 | { | |
477 | case vect_unused_in_scope: | |
478 | relevant = vect_used_by_reduction; | |
479 | break; | |
480 | ||
481 | case vect_used_by_reduction: | |
482 | if (gimple_code (stmt) == GIMPLE_PHI) | |
483 | break; | |
484 | /* fall through */ | |
485 | ||
486 | default: | |
487 | if (vect_print_dump_info (REPORT_DETAILS)) | |
488 | fprintf (vect_dump, "unsupported use of reduction."); | |
489 | ||
490 | VEC_free (gimple, heap, worklist); | |
491 | return false; | |
492 | } | |
493 | ||
494 | live_p = false; | |
495 | break; | |
496 | ||
497 | case vect_nested_cycle: | |
498 | if (tmp_relevant != vect_unused_in_scope | |
499 | && tmp_relevant != vect_used_in_outer_by_reduction | |
500 | && tmp_relevant != vect_used_in_outer) | |
501 | { | |
502 | if (vect_print_dump_info (REPORT_DETAILS)) | |
503 | fprintf (vect_dump, "unsupported use of nested cycle."); | |
7c5222ff | 504 | |
06066f92 IR |
505 | VEC_free (gimple, heap, worklist); |
506 | return false; | |
507 | } | |
7c5222ff | 508 | |
06066f92 IR |
509 | live_p = false; |
510 | break; | |
511 | ||
512 | case vect_double_reduction_def: | |
513 | if (tmp_relevant != vect_unused_in_scope | |
514 | && tmp_relevant != vect_used_by_reduction) | |
515 | { | |
7c5222ff | 516 | if (vect_print_dump_info (REPORT_DETAILS)) |
06066f92 | 517 | fprintf (vect_dump, "unsupported use of double reduction."); |
7c5222ff IR |
518 | |
519 | VEC_free (gimple, heap, worklist); | |
520 | return false; | |
06066f92 IR |
521 | } |
522 | ||
523 | live_p = false; | |
524 | break; | |
7c5222ff | 525 | |
06066f92 IR |
526 | default: |
527 | break; | |
7c5222ff IR |
528 | } |
529 | ||
ebfd146a IR |
530 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) |
531 | { | |
532 | tree op = USE_FROM_PTR (use_p); | |
533 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, &worklist)) | |
534 | { | |
535 | VEC_free (gimple, heap, worklist); | |
536 | return false; | |
537 | } | |
538 | } | |
539 | } /* while worklist */ | |
540 | ||
541 | VEC_free (gimple, heap, worklist); | |
542 | return true; | |
543 | } | |
544 | ||
545 | ||
546 | int | |
547 | cost_for_stmt (gimple stmt) | |
548 | { | |
549 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
550 | ||
551 | switch (STMT_VINFO_TYPE (stmt_info)) | |
552 | { | |
553 | case load_vec_info_type: | |
554 | return TARG_SCALAR_LOAD_COST; | |
555 | case store_vec_info_type: | |
556 | return TARG_SCALAR_STORE_COST; | |
557 | case op_vec_info_type: | |
558 | case condition_vec_info_type: | |
559 | case assignment_vec_info_type: | |
560 | case reduc_vec_info_type: | |
561 | case induc_vec_info_type: | |
562 | case type_promotion_vec_info_type: | |
563 | case type_demotion_vec_info_type: | |
564 | case type_conversion_vec_info_type: | |
565 | case call_vec_info_type: | |
566 | return TARG_SCALAR_STMT_COST; | |
567 | case undef_vec_info_type: | |
568 | default: | |
569 | gcc_unreachable (); | |
570 | } | |
571 | } | |
572 | ||
573 | /* Function vect_model_simple_cost. | |
574 | ||
575 | Models cost for simple operations, i.e. those that only emit ncopies of a | |
576 | single op. Right now, this does not account for multiple insns that could | |
577 | be generated for the single vector op. We will handle that shortly. */ | |
578 | ||
579 | void | |
580 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, | |
581 | enum vect_def_type *dt, slp_tree slp_node) | |
582 | { | |
583 | int i; | |
584 | int inside_cost = 0, outside_cost = 0; | |
585 | ||
586 | /* The SLP costs were already calculated during SLP tree build. */ | |
587 | if (PURE_SLP_STMT (stmt_info)) | |
588 | return; | |
589 | ||
590 | inside_cost = ncopies * TARG_VEC_STMT_COST; | |
591 | ||
592 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
593 | for (i = 0; i < 2; i++) | |
594 | { | |
8644a673 | 595 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
ebfd146a IR |
596 | outside_cost += TARG_SCALAR_TO_VEC_COST; |
597 | } | |
598 | ||
599 | if (vect_print_dump_info (REPORT_COST)) | |
600 | fprintf (vect_dump, "vect_model_simple_cost: inside_cost = %d, " | |
601 | "outside_cost = %d .", inside_cost, outside_cost); | |
602 | ||
603 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
604 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
605 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
606 | } | |
607 | ||
608 | ||
609 | /* Function vect_cost_strided_group_size | |
610 | ||
611 | For strided load or store, return the group_size only if it is the first | |
612 | load or store of a group, else return 1. This ensures that group size is | |
613 | only returned once per group. */ | |
614 | ||
615 | static int | |
616 | vect_cost_strided_group_size (stmt_vec_info stmt_info) | |
617 | { | |
618 | gimple first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
619 | ||
620 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
621 | return DR_GROUP_SIZE (stmt_info); | |
622 | ||
623 | return 1; | |
624 | } | |
625 | ||
626 | ||
627 | /* Function vect_model_store_cost | |
628 | ||
629 | Models cost for stores. In the case of strided accesses, one access | |
630 | has the overhead of the strided access attributed to it. */ | |
631 | ||
632 | void | |
633 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, | |
634 | enum vect_def_type dt, slp_tree slp_node) | |
635 | { | |
636 | int group_size; | |
637 | int inside_cost = 0, outside_cost = 0; | |
638 | ||
639 | /* The SLP costs were already calculated during SLP tree build. */ | |
640 | if (PURE_SLP_STMT (stmt_info)) | |
641 | return; | |
642 | ||
8644a673 | 643 | if (dt == vect_constant_def || dt == vect_external_def) |
ebfd146a IR |
644 | outside_cost = TARG_SCALAR_TO_VEC_COST; |
645 | ||
646 | /* Strided access? */ | |
647 | if (DR_GROUP_FIRST_DR (stmt_info) && !slp_node) | |
648 | group_size = vect_cost_strided_group_size (stmt_info); | |
649 | /* Not a strided access. */ | |
650 | else | |
651 | group_size = 1; | |
652 | ||
653 | /* Is this an access in a group of stores, which provide strided access? | |
654 | If so, add in the cost of the permutes. */ | |
655 | if (group_size > 1) | |
656 | { | |
657 | /* Uses a high and low interleave operation for each needed permute. */ | |
658 | inside_cost = ncopies * exact_log2(group_size) * group_size | |
659 | * TARG_VEC_STMT_COST; | |
660 | ||
661 | if (vect_print_dump_info (REPORT_COST)) | |
662 | fprintf (vect_dump, "vect_model_store_cost: strided group_size = %d .", | |
663 | group_size); | |
664 | ||
665 | } | |
666 | ||
667 | /* Costs of the stores. */ | |
668 | inside_cost += ncopies * TARG_VEC_STORE_COST; | |
669 | ||
670 | if (vect_print_dump_info (REPORT_COST)) | |
671 | fprintf (vect_dump, "vect_model_store_cost: inside_cost = %d, " | |
672 | "outside_cost = %d .", inside_cost, outside_cost); | |
673 | ||
674 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
675 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
676 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
677 | } | |
678 | ||
679 | ||
680 | /* Function vect_model_load_cost | |
681 | ||
682 | Models cost for loads. In the case of strided accesses, the last access | |
683 | has the overhead of the strided access attributed to it. Since unaligned | |
684 | accesses are supported for loads, we also account for the costs of the | |
685 | access scheme chosen. */ | |
686 | ||
687 | void | |
688 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, slp_tree slp_node) | |
689 | ||
690 | { | |
691 | int group_size; | |
692 | int alignment_support_cheme; | |
693 | gimple first_stmt; | |
694 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; | |
695 | int inside_cost = 0, outside_cost = 0; | |
696 | ||
697 | /* The SLP costs were already calculated during SLP tree build. */ | |
698 | if (PURE_SLP_STMT (stmt_info)) | |
699 | return; | |
700 | ||
701 | /* Strided accesses? */ | |
702 | first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
703 | if (first_stmt && !slp_node) | |
704 | { | |
705 | group_size = vect_cost_strided_group_size (stmt_info); | |
706 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
707 | } | |
708 | /* Not a strided access. */ | |
709 | else | |
710 | { | |
711 | group_size = 1; | |
712 | first_dr = dr; | |
713 | } | |
714 | ||
715 | alignment_support_cheme = vect_supportable_dr_alignment (first_dr); | |
716 | ||
717 | /* Is this an access in a group of loads providing strided access? | |
718 | If so, add in the cost of the permutes. */ | |
719 | if (group_size > 1) | |
720 | { | |
721 | /* Uses an even and odd extract operations for each needed permute. */ | |
722 | inside_cost = ncopies * exact_log2(group_size) * group_size | |
723 | * TARG_VEC_STMT_COST; | |
724 | ||
725 | if (vect_print_dump_info (REPORT_COST)) | |
726 | fprintf (vect_dump, "vect_model_load_cost: strided group_size = %d .", | |
727 | group_size); | |
728 | ||
729 | } | |
730 | ||
731 | /* The loads themselves. */ | |
732 | switch (alignment_support_cheme) | |
733 | { | |
734 | case dr_aligned: | |
735 | { | |
736 | inside_cost += ncopies * TARG_VEC_LOAD_COST; | |
737 | ||
738 | if (vect_print_dump_info (REPORT_COST)) | |
739 | fprintf (vect_dump, "vect_model_load_cost: aligned."); | |
740 | ||
741 | break; | |
742 | } | |
743 | case dr_unaligned_supported: | |
744 | { | |
745 | /* Here, we assign an additional cost for the unaligned load. */ | |
746 | inside_cost += ncopies * TARG_VEC_UNALIGNED_LOAD_COST; | |
747 | ||
748 | if (vect_print_dump_info (REPORT_COST)) | |
749 | fprintf (vect_dump, "vect_model_load_cost: unaligned supported by " | |
750 | "hardware."); | |
751 | ||
752 | break; | |
753 | } | |
754 | case dr_explicit_realign: | |
755 | { | |
756 | inside_cost += ncopies * (2*TARG_VEC_LOAD_COST + TARG_VEC_STMT_COST); | |
757 | ||
758 | /* FIXME: If the misalignment remains fixed across the iterations of | |
759 | the containing loop, the following cost should be added to the | |
760 | outside costs. */ | |
761 | if (targetm.vectorize.builtin_mask_for_load) | |
762 | inside_cost += TARG_VEC_STMT_COST; | |
763 | ||
764 | break; | |
765 | } | |
766 | case dr_explicit_realign_optimized: | |
767 | { | |
768 | if (vect_print_dump_info (REPORT_COST)) | |
769 | fprintf (vect_dump, "vect_model_load_cost: unaligned software " | |
770 | "pipelined."); | |
771 | ||
772 | /* Unaligned software pipeline has a load of an address, an initial | |
773 | load, and possibly a mask operation to "prime" the loop. However, | |
774 | if this is an access in a group of loads, which provide strided | |
775 | access, then the above cost should only be considered for one | |
776 | access in the group. Inside the loop, there is a load op | |
777 | and a realignment op. */ | |
778 | ||
779 | if ((!DR_GROUP_FIRST_DR (stmt_info)) || group_size > 1 || slp_node) | |
780 | { | |
781 | outside_cost = 2*TARG_VEC_STMT_COST; | |
782 | if (targetm.vectorize.builtin_mask_for_load) | |
783 | outside_cost += TARG_VEC_STMT_COST; | |
784 | } | |
785 | ||
786 | inside_cost += ncopies * (TARG_VEC_LOAD_COST + TARG_VEC_STMT_COST); | |
787 | ||
788 | break; | |
789 | } | |
790 | ||
791 | default: | |
792 | gcc_unreachable (); | |
793 | } | |
794 | ||
795 | if (vect_print_dump_info (REPORT_COST)) | |
796 | fprintf (vect_dump, "vect_model_load_cost: inside_cost = %d, " | |
797 | "outside_cost = %d .", inside_cost, outside_cost); | |
798 | ||
799 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
800 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
801 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
802 | } | |
803 | ||
804 | ||
805 | /* Function vect_init_vector. | |
806 | ||
807 | Insert a new stmt (INIT_STMT) that initializes a new vector variable with | |
808 | the vector elements of VECTOR_VAR. Place the initialization at BSI if it | |
809 | is not NULL. Otherwise, place the initialization at the loop preheader. | |
810 | Return the DEF of INIT_STMT. | |
811 | It will be used in the vectorization of STMT. */ | |
812 | ||
813 | tree | |
814 | vect_init_vector (gimple stmt, tree vector_var, tree vector_type, | |
815 | gimple_stmt_iterator *gsi) | |
816 | { | |
817 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
818 | tree new_var; | |
819 | gimple init_stmt; | |
820 | tree vec_oprnd; | |
821 | edge pe; | |
822 | tree new_temp; | |
823 | basic_block new_bb; | |
824 | ||
825 | new_var = vect_get_new_vect_var (vector_type, vect_simple_var, "cst_"); | |
826 | add_referenced_var (new_var); | |
827 | init_stmt = gimple_build_assign (new_var, vector_var); | |
828 | new_temp = make_ssa_name (new_var, init_stmt); | |
829 | gimple_assign_set_lhs (init_stmt, new_temp); | |
830 | ||
831 | if (gsi) | |
832 | vect_finish_stmt_generation (stmt, init_stmt, gsi); | |
833 | else | |
834 | { | |
835 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
a70d6342 IR |
836 | |
837 | if (loop_vinfo) | |
838 | { | |
839 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
840 | ||
841 | if (nested_in_vect_loop_p (loop, stmt)) | |
842 | loop = loop->inner; | |
843 | ||
844 | pe = loop_preheader_edge (loop); | |
845 | new_bb = gsi_insert_on_edge_immediate (pe, init_stmt); | |
846 | gcc_assert (!new_bb); | |
847 | } | |
848 | else | |
849 | { | |
850 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
851 | basic_block bb; | |
852 | gimple_stmt_iterator gsi_bb_start; | |
853 | ||
854 | gcc_assert (bb_vinfo); | |
855 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 856 | gsi_bb_start = gsi_after_labels (bb); |
a70d6342 IR |
857 | gsi_insert_before (&gsi_bb_start, init_stmt, GSI_SAME_STMT); |
858 | } | |
ebfd146a IR |
859 | } |
860 | ||
861 | if (vect_print_dump_info (REPORT_DETAILS)) | |
862 | { | |
863 | fprintf (vect_dump, "created new init_stmt: "); | |
864 | print_gimple_stmt (vect_dump, init_stmt, 0, TDF_SLIM); | |
865 | } | |
866 | ||
867 | vec_oprnd = gimple_assign_lhs (init_stmt); | |
868 | return vec_oprnd; | |
869 | } | |
870 | ||
a70d6342 | 871 | |
ebfd146a IR |
872 | /* Function vect_get_vec_def_for_operand. |
873 | ||
874 | OP is an operand in STMT. This function returns a (vector) def that will be | |
875 | used in the vectorized stmt for STMT. | |
876 | ||
877 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
878 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
879 | ||
880 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
881 | needs to be introduced. */ | |
882 | ||
883 | tree | |
884 | vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def) | |
885 | { | |
886 | tree vec_oprnd; | |
887 | gimple vec_stmt; | |
888 | gimple def_stmt; | |
889 | stmt_vec_info def_stmt_info = NULL; | |
890 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
891 | tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
892 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
893 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
894 | tree vec_inv; | |
895 | tree vec_cst; | |
896 | tree t = NULL_TREE; | |
897 | tree def; | |
898 | int i; | |
899 | enum vect_def_type dt; | |
900 | bool is_simple_use; | |
901 | tree vector_type; | |
902 | ||
903 | if (vect_print_dump_info (REPORT_DETAILS)) | |
904 | { | |
905 | fprintf (vect_dump, "vect_get_vec_def_for_operand: "); | |
906 | print_generic_expr (vect_dump, op, TDF_SLIM); | |
907 | } | |
908 | ||
a70d6342 IR |
909 | is_simple_use = vect_is_simple_use (op, loop_vinfo, NULL, &def_stmt, &def, |
910 | &dt); | |
ebfd146a IR |
911 | gcc_assert (is_simple_use); |
912 | if (vect_print_dump_info (REPORT_DETAILS)) | |
913 | { | |
914 | if (def) | |
915 | { | |
916 | fprintf (vect_dump, "def = "); | |
917 | print_generic_expr (vect_dump, def, TDF_SLIM); | |
918 | } | |
919 | if (def_stmt) | |
920 | { | |
921 | fprintf (vect_dump, " def_stmt = "); | |
922 | print_gimple_stmt (vect_dump, def_stmt, 0, TDF_SLIM); | |
923 | } | |
924 | } | |
925 | ||
926 | switch (dt) | |
927 | { | |
928 | /* Case 1: operand is a constant. */ | |
929 | case vect_constant_def: | |
930 | { | |
7569a6cc RG |
931 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
932 | gcc_assert (vector_type); | |
933 | ||
ebfd146a IR |
934 | if (scalar_def) |
935 | *scalar_def = op; | |
936 | ||
937 | /* Create 'vect_cst_ = {cst,cst,...,cst}' */ | |
938 | if (vect_print_dump_info (REPORT_DETAILS)) | |
939 | fprintf (vect_dump, "Create vector_cst. nunits = %d", nunits); | |
940 | ||
941 | for (i = nunits - 1; i >= 0; --i) | |
942 | { | |
943 | t = tree_cons (NULL_TREE, op, t); | |
944 | } | |
7569a6cc RG |
945 | vec_cst = build_vector (vector_type, t); |
946 | return vect_init_vector (stmt, vec_cst, vector_type, NULL); | |
ebfd146a IR |
947 | } |
948 | ||
949 | /* Case 2: operand is defined outside the loop - loop invariant. */ | |
8644a673 | 950 | case vect_external_def: |
ebfd146a IR |
951 | { |
952 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); | |
953 | gcc_assert (vector_type); | |
954 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); | |
955 | ||
956 | if (scalar_def) | |
957 | *scalar_def = def; | |
958 | ||
959 | /* Create 'vec_inv = {inv,inv,..,inv}' */ | |
960 | if (vect_print_dump_info (REPORT_DETAILS)) | |
961 | fprintf (vect_dump, "Create vector_inv."); | |
962 | ||
963 | for (i = nunits - 1; i >= 0; --i) | |
964 | { | |
965 | t = tree_cons (NULL_TREE, def, t); | |
966 | } | |
967 | ||
968 | /* FIXME: use build_constructor directly. */ | |
969 | vec_inv = build_constructor_from_list (vector_type, t); | |
970 | return vect_init_vector (stmt, vec_inv, vector_type, NULL); | |
971 | } | |
972 | ||
973 | /* Case 3: operand is defined inside the loop. */ | |
8644a673 | 974 | case vect_internal_def: |
ebfd146a IR |
975 | { |
976 | if (scalar_def) | |
977 | *scalar_def = NULL/* FIXME tuples: def_stmt*/; | |
978 | ||
979 | /* Get the def from the vectorized stmt. */ | |
980 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
981 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
982 | gcc_assert (vec_stmt); | |
983 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
984 | vec_oprnd = PHI_RESULT (vec_stmt); | |
985 | else if (is_gimple_call (vec_stmt)) | |
986 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
987 | else | |
988 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
989 | return vec_oprnd; | |
990 | } | |
991 | ||
992 | /* Case 4: operand is defined by a loop header phi - reduction */ | |
993 | case vect_reduction_def: | |
06066f92 | 994 | case vect_double_reduction_def: |
7c5222ff | 995 | case vect_nested_cycle: |
ebfd146a IR |
996 | { |
997 | struct loop *loop; | |
998 | ||
999 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1000 | loop = (gimple_bb (def_stmt))->loop_father; | |
1001 | ||
1002 | /* Get the def before the loop */ | |
1003 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); | |
1004 | return get_initial_def_for_reduction (stmt, op, scalar_def); | |
1005 | } | |
1006 | ||
1007 | /* Case 5: operand is defined by loop-header phi - induction. */ | |
1008 | case vect_induction_def: | |
1009 | { | |
1010 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1011 | ||
1012 | /* Get the def from the vectorized stmt. */ | |
1013 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1014 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
1015 | gcc_assert (vec_stmt && gimple_code (vec_stmt) == GIMPLE_PHI); | |
1016 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1017 | return vec_oprnd; | |
1018 | } | |
1019 | ||
1020 | default: | |
1021 | gcc_unreachable (); | |
1022 | } | |
1023 | } | |
1024 | ||
1025 | ||
1026 | /* Function vect_get_vec_def_for_stmt_copy | |
1027 | ||
1028 | Return a vector-def for an operand. This function is used when the | |
1029 | vectorized stmt to be created (by the caller to this function) is a "copy" | |
1030 | created in case the vectorized result cannot fit in one vector, and several | |
1031 | copies of the vector-stmt are required. In this case the vector-def is | |
1032 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field | |
1033 | of the stmt that defines VEC_OPRND. | |
1034 | DT is the type of the vector def VEC_OPRND. | |
1035 | ||
1036 | Context: | |
1037 | In case the vectorization factor (VF) is bigger than the number | |
1038 | of elements that can fit in a vectype (nunits), we have to generate | |
1039 | more than one vector stmt to vectorize the scalar stmt. This situation | |
1040 | arises when there are multiple data-types operated upon in the loop; the | |
1041 | smallest data-type determines the VF, and as a result, when vectorizing | |
1042 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1043 | vector stmt (each computing a vector of 'nunits' results, and together | |
1044 | computing 'VF' results in each iteration). This function is called when | |
1045 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in | |
1046 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1047 | ||
1048 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
1049 | ||
1050 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 | |
1051 | VS1.1: vx.1 = memref1 VS1.2 | |
1052 | VS1.2: vx.2 = memref2 VS1.3 | |
1053 | VS1.3: vx.3 = memref3 | |
1054 | ||
1055 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1056 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1057 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1058 | VSnew.3: vz3 = vx.3 + ... | |
1059 | ||
1060 | The vectorization of S1 is explained in vectorizable_load. | |
1061 | The vectorization of S2: | |
1062 | To create the first vector-stmt out of the 4 copies - VSnew.0 - | |
1063 | the function 'vect_get_vec_def_for_operand' is called to | |
1064 | get the relevant vector-def for each operand of S2. For operand x it | |
1065 | returns the vector-def 'vx.0'. | |
1066 | ||
1067 | To create the remaining copies of the vector-stmt (VSnew.j), this | |
1068 | function is called to get the relevant vector-def for each operand. It is | |
1069 | obtained from the respective VS1.j stmt, which is recorded in the | |
1070 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. | |
1071 | ||
1072 | For example, to obtain the vector-def 'vx.1' in order to create the | |
1073 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1074 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
1075 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', | |
1076 | and return its def ('vx.1'). | |
1077 | Overall, to create the above sequence this function will be called 3 times: | |
1078 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1079 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1080 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1081 | ||
1082 | tree | |
1083 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1084 | { | |
1085 | gimple vec_stmt_for_operand; | |
1086 | stmt_vec_info def_stmt_info; | |
1087 | ||
1088 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1089 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1090 | return vec_oprnd; |
1091 | ||
1092 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1093 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1094 | gcc_assert (def_stmt_info); | |
1095 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1096 | gcc_assert (vec_stmt_for_operand); | |
1097 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1098 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) | |
1099 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1100 | else | |
1101 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1102 | return vec_oprnd; | |
1103 | } | |
1104 | ||
1105 | ||
1106 | /* Get vectorized definitions for the operands to create a copy of an original | |
1107 | stmt. See vect_get_vec_def_for_stmt_copy() for details. */ | |
1108 | ||
1109 | static void | |
1110 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, | |
1111 | VEC(tree,heap) **vec_oprnds0, | |
1112 | VEC(tree,heap) **vec_oprnds1) | |
1113 | { | |
1114 | tree vec_oprnd = VEC_pop (tree, *vec_oprnds0); | |
1115 | ||
1116 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
1117 | VEC_quick_push (tree, *vec_oprnds0, vec_oprnd); | |
1118 | ||
1119 | if (vec_oprnds1 && *vec_oprnds1) | |
1120 | { | |
1121 | vec_oprnd = VEC_pop (tree, *vec_oprnds1); | |
1122 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); | |
1123 | VEC_quick_push (tree, *vec_oprnds1, vec_oprnd); | |
1124 | } | |
1125 | } | |
1126 | ||
1127 | ||
1128 | /* Get vectorized definitions for OP0 and OP1, or SLP_NODE if it is not NULL. */ | |
1129 | ||
1130 | static void | |
1131 | vect_get_vec_defs (tree op0, tree op1, gimple stmt, | |
1132 | VEC(tree,heap) **vec_oprnds0, VEC(tree,heap) **vec_oprnds1, | |
1133 | slp_tree slp_node) | |
1134 | { | |
1135 | if (slp_node) | |
1136 | vect_get_slp_defs (slp_node, vec_oprnds0, vec_oprnds1); | |
1137 | else | |
1138 | { | |
1139 | tree vec_oprnd; | |
1140 | ||
1141 | *vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
1142 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
1143 | VEC_quick_push (tree, *vec_oprnds0, vec_oprnd); | |
1144 | ||
1145 | if (op1) | |
1146 | { | |
1147 | *vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
1148 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL); | |
1149 | VEC_quick_push (tree, *vec_oprnds1, vec_oprnd); | |
1150 | } | |
1151 | } | |
1152 | } | |
1153 | ||
1154 | ||
1155 | /* Function vect_finish_stmt_generation. | |
1156 | ||
1157 | Insert a new stmt. */ | |
1158 | ||
1159 | void | |
1160 | vect_finish_stmt_generation (gimple stmt, gimple vec_stmt, | |
1161 | gimple_stmt_iterator *gsi) | |
1162 | { | |
1163 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1164 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1165 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1166 | |
1167 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1168 | ||
1169 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); | |
1170 | ||
a70d6342 IR |
1171 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo, |
1172 | bb_vinfo)); | |
ebfd146a IR |
1173 | |
1174 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1175 | { | |
1176 | fprintf (vect_dump, "add new stmt: "); | |
1177 | print_gimple_stmt (vect_dump, vec_stmt, 0, TDF_SLIM); | |
1178 | } | |
1179 | ||
1180 | gimple_set_location (vec_stmt, gimple_location (gsi_stmt (*gsi))); | |
1181 | } | |
1182 | ||
1183 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1184 | a function declaration if the target has a vectorized version | |
1185 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1186 | ||
1187 | tree | |
1188 | vectorizable_function (gimple call, tree vectype_out, tree vectype_in) | |
1189 | { | |
1190 | tree fndecl = gimple_call_fndecl (call); | |
1191 | enum built_in_function code; | |
1192 | ||
1193 | /* We only handle functions that do not read or clobber memory -- i.e. | |
1194 | const or novops ones. */ | |
1195 | if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS))) | |
1196 | return NULL_TREE; | |
1197 | ||
1198 | if (!fndecl | |
1199 | || TREE_CODE (fndecl) != FUNCTION_DECL | |
1200 | || !DECL_BUILT_IN (fndecl)) | |
1201 | return NULL_TREE; | |
1202 | ||
1203 | code = DECL_FUNCTION_CODE (fndecl); | |
1204 | return targetm.vectorize.builtin_vectorized_function (code, vectype_out, | |
1205 | vectype_in); | |
1206 | } | |
1207 | ||
1208 | /* Function vectorizable_call. | |
1209 | ||
1210 | Check if STMT performs a function call that can be vectorized. | |
1211 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1212 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
1213 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1214 | ||
1215 | static bool | |
1216 | vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt) | |
1217 | { | |
1218 | tree vec_dest; | |
1219 | tree scalar_dest; | |
1220 | tree op, type; | |
1221 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1222 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
1223 | tree vectype_out, vectype_in; | |
1224 | int nunits_in; | |
1225 | int nunits_out; | |
1226 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1227 | tree fndecl, new_temp, def, rhs_type, lhs_type; | |
1228 | gimple def_stmt; | |
1229 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
63827fb8 | 1230 | gimple new_stmt = NULL; |
ebfd146a IR |
1231 | int ncopies, j; |
1232 | VEC(tree, heap) *vargs = NULL; | |
1233 | enum { NARROW, NONE, WIDEN } modifier; | |
1234 | size_t i, nargs; | |
1235 | ||
a70d6342 IR |
1236 | /* FORNOW: unsupported in basic block SLP. */ |
1237 | gcc_assert (loop_vinfo); | |
1238 | ||
ebfd146a IR |
1239 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
1240 | return false; | |
1241 | ||
8644a673 | 1242 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1243 | return false; |
1244 | ||
1245 | /* FORNOW: SLP not supported. */ | |
1246 | if (STMT_SLP_TYPE (stmt_info)) | |
1247 | return false; | |
1248 | ||
1249 | /* Is STMT a vectorizable call? */ | |
1250 | if (!is_gimple_call (stmt)) | |
1251 | return false; | |
1252 | ||
1253 | if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
1254 | return false; | |
1255 | ||
1256 | /* Process function arguments. */ | |
1257 | rhs_type = NULL_TREE; | |
1258 | nargs = gimple_call_num_args (stmt); | |
1259 | ||
1260 | /* Bail out if the function has more than two arguments, we | |
1261 | do not have interesting builtin functions to vectorize with | |
1262 | more than two arguments. No arguments is also not good. */ | |
1263 | if (nargs == 0 || nargs > 2) | |
1264 | return false; | |
1265 | ||
1266 | for (i = 0; i < nargs; i++) | |
1267 | { | |
1268 | op = gimple_call_arg (stmt, i); | |
1269 | ||
1270 | /* We can only handle calls with arguments of the same type. */ | |
1271 | if (rhs_type | |
1272 | && rhs_type != TREE_TYPE (op)) | |
1273 | { | |
1274 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1275 | fprintf (vect_dump, "argument types differ."); | |
1276 | return false; | |
1277 | } | |
1278 | rhs_type = TREE_TYPE (op); | |
1279 | ||
a70d6342 | 1280 | if (!vect_is_simple_use (op, loop_vinfo, NULL, &def_stmt, &def, &dt[i])) |
ebfd146a IR |
1281 | { |
1282 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1283 | fprintf (vect_dump, "use not simple."); | |
1284 | return false; | |
1285 | } | |
1286 | } | |
1287 | ||
1288 | vectype_in = get_vectype_for_scalar_type (rhs_type); | |
1289 | if (!vectype_in) | |
1290 | return false; | |
1291 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); | |
1292 | ||
1293 | lhs_type = TREE_TYPE (gimple_call_lhs (stmt)); | |
1294 | vectype_out = get_vectype_for_scalar_type (lhs_type); | |
1295 | if (!vectype_out) | |
1296 | return false; | |
1297 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
1298 | ||
1299 | /* FORNOW */ | |
1300 | if (nunits_in == nunits_out / 2) | |
1301 | modifier = NARROW; | |
1302 | else if (nunits_out == nunits_in) | |
1303 | modifier = NONE; | |
1304 | else if (nunits_out == nunits_in / 2) | |
1305 | modifier = WIDEN; | |
1306 | else | |
1307 | return false; | |
1308 | ||
1309 | /* For now, we only vectorize functions if a target specific builtin | |
1310 | is available. TODO -- in some cases, it might be profitable to | |
1311 | insert the calls for pieces of the vector, in order to be able | |
1312 | to vectorize other operations in the loop. */ | |
1313 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
1314 | if (fndecl == NULL_TREE) | |
1315 | { | |
1316 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1317 | fprintf (vect_dump, "function is not vectorizable."); | |
1318 | ||
1319 | return false; | |
1320 | } | |
1321 | ||
5006671f | 1322 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a IR |
1323 | |
1324 | if (modifier == NARROW) | |
1325 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
1326 | else | |
1327 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1328 | ||
1329 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
1330 | needs to be generated. */ | |
1331 | gcc_assert (ncopies >= 1); | |
1332 | ||
1333 | if (!vec_stmt) /* transformation not required. */ | |
1334 | { | |
1335 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1336 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1337 | fprintf (vect_dump, "=== vectorizable_call ==="); | |
1338 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
1339 | return true; | |
1340 | } | |
1341 | ||
1342 | /** Transform. **/ | |
1343 | ||
1344 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1345 | fprintf (vect_dump, "transform operation."); | |
1346 | ||
1347 | /* Handle def. */ | |
1348 | scalar_dest = gimple_call_lhs (stmt); | |
1349 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1350 | ||
1351 | prev_stmt_info = NULL; | |
1352 | switch (modifier) | |
1353 | { | |
1354 | case NONE: | |
1355 | for (j = 0; j < ncopies; ++j) | |
1356 | { | |
1357 | /* Build argument list for the vectorized call. */ | |
1358 | if (j == 0) | |
1359 | vargs = VEC_alloc (tree, heap, nargs); | |
1360 | else | |
1361 | VEC_truncate (tree, vargs, 0); | |
1362 | ||
1363 | for (i = 0; i < nargs; i++) | |
1364 | { | |
1365 | op = gimple_call_arg (stmt, i); | |
1366 | if (j == 0) | |
1367 | vec_oprnd0 | |
1368 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1369 | else | |
63827fb8 IR |
1370 | { |
1371 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
1372 | vec_oprnd0 | |
1373 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
1374 | } | |
ebfd146a IR |
1375 | |
1376 | VEC_quick_push (tree, vargs, vec_oprnd0); | |
1377 | } | |
1378 | ||
1379 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1380 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1381 | gimple_call_set_lhs (new_stmt, new_temp); | |
1382 | ||
1383 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1384 | ||
1385 | if (j == 0) | |
1386 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1387 | else | |
1388 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1389 | ||
1390 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1391 | } | |
1392 | ||
1393 | break; | |
1394 | ||
1395 | case NARROW: | |
1396 | for (j = 0; j < ncopies; ++j) | |
1397 | { | |
1398 | /* Build argument list for the vectorized call. */ | |
1399 | if (j == 0) | |
1400 | vargs = VEC_alloc (tree, heap, nargs * 2); | |
1401 | else | |
1402 | VEC_truncate (tree, vargs, 0); | |
1403 | ||
1404 | for (i = 0; i < nargs; i++) | |
1405 | { | |
1406 | op = gimple_call_arg (stmt, i); | |
1407 | if (j == 0) | |
1408 | { | |
1409 | vec_oprnd0 | |
1410 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1411 | vec_oprnd1 | |
63827fb8 | 1412 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
1413 | } |
1414 | else | |
1415 | { | |
63827fb8 | 1416 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i); |
ebfd146a | 1417 | vec_oprnd0 |
63827fb8 | 1418 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 1419 | vec_oprnd1 |
63827fb8 | 1420 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
1421 | } |
1422 | ||
1423 | VEC_quick_push (tree, vargs, vec_oprnd0); | |
1424 | VEC_quick_push (tree, vargs, vec_oprnd1); | |
1425 | } | |
1426 | ||
1427 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1428 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1429 | gimple_call_set_lhs (new_stmt, new_temp); | |
1430 | ||
1431 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1432 | ||
1433 | if (j == 0) | |
1434 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1435 | else | |
1436 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1437 | ||
1438 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1439 | } | |
1440 | ||
1441 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
1442 | ||
1443 | break; | |
1444 | ||
1445 | case WIDEN: | |
1446 | /* No current target implements this case. */ | |
1447 | return false; | |
1448 | } | |
1449 | ||
1450 | VEC_free (tree, heap, vargs); | |
1451 | ||
1452 | /* Update the exception handling table with the vector stmt if necessary. */ | |
1453 | if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt)) | |
1454 | gimple_purge_dead_eh_edges (gimple_bb (stmt)); | |
1455 | ||
1456 | /* The call in STMT might prevent it from being removed in dce. | |
1457 | We however cannot remove it here, due to the way the ssa name | |
1458 | it defines is mapped to the new definition. So just replace | |
1459 | rhs of the statement with something harmless. */ | |
1460 | ||
1461 | type = TREE_TYPE (scalar_dest); | |
1462 | new_stmt = gimple_build_assign (gimple_call_lhs (stmt), | |
1463 | fold_convert (type, integer_zero_node)); | |
1464 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
1465 | set_vinfo_for_stmt (stmt, NULL); | |
1466 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
1467 | gsi_replace (gsi, new_stmt, false); | |
1468 | SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt; | |
1469 | ||
1470 | return true; | |
1471 | } | |
1472 | ||
1473 | ||
1474 | /* Function vect_gen_widened_results_half | |
1475 | ||
1476 | Create a vector stmt whose code, type, number of arguments, and result | |
1477 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are | |
1478 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. | |
1479 | In the case that CODE is a CALL_EXPR, this means that a call to DECL | |
1480 | needs to be created (DECL is a function-decl of a target-builtin). | |
1481 | STMT is the original scalar stmt that we are vectorizing. */ | |
1482 | ||
1483 | static gimple | |
1484 | vect_gen_widened_results_half (enum tree_code code, | |
1485 | tree decl, | |
1486 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
1487 | tree vec_dest, gimple_stmt_iterator *gsi, | |
1488 | gimple stmt) | |
1489 | { | |
1490 | gimple new_stmt; | |
1491 | tree new_temp; | |
ebfd146a IR |
1492 | |
1493 | /* Generate half of the widened result: */ | |
1494 | if (code == CALL_EXPR) | |
1495 | { | |
1496 | /* Target specific support */ | |
1497 | if (op_type == binary_op) | |
1498 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
1499 | else | |
1500 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
1501 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1502 | gimple_call_set_lhs (new_stmt, new_temp); | |
1503 | } | |
1504 | else | |
1505 | { | |
1506 | /* Generic support */ | |
1507 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
1508 | if (op_type != binary_op) | |
1509 | vec_oprnd1 = NULL; | |
1510 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0, | |
1511 | vec_oprnd1); | |
1512 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1513 | gimple_assign_set_lhs (new_stmt, new_temp); | |
1514 | } | |
1515 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1516 | ||
ebfd146a IR |
1517 | return new_stmt; |
1518 | } | |
1519 | ||
1520 | ||
1521 | /* Check if STMT performs a conversion operation, that can be vectorized. | |
1522 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1523 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
1524 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1525 | ||
1526 | static bool | |
1527 | vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi, | |
1528 | gimple *vec_stmt, slp_tree slp_node) | |
1529 | { | |
1530 | tree vec_dest; | |
1531 | tree scalar_dest; | |
1532 | tree op0; | |
1533 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1534 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1535 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1536 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
1537 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; | |
1538 | tree new_temp; | |
1539 | tree def; | |
1540 | gimple def_stmt; | |
1541 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
1542 | gimple new_stmt = NULL; | |
1543 | stmt_vec_info prev_stmt_info; | |
1544 | int nunits_in; | |
1545 | int nunits_out; | |
1546 | tree vectype_out, vectype_in; | |
1547 | int ncopies, j; | |
1548 | tree expr; | |
1549 | tree rhs_type, lhs_type; | |
1550 | tree builtin_decl; | |
1551 | enum { NARROW, NONE, WIDEN } modifier; | |
1552 | int i; | |
1553 | VEC(tree,heap) *vec_oprnds0 = NULL; | |
1554 | tree vop0; | |
1555 | tree integral_type; | |
1556 | VEC(tree,heap) *dummy = NULL; | |
1557 | int dummy_int; | |
1558 | ||
1559 | /* Is STMT a vectorizable conversion? */ | |
1560 | ||
a70d6342 IR |
1561 | /* FORNOW: unsupported in basic block SLP. */ |
1562 | gcc_assert (loop_vinfo); | |
1563 | ||
ebfd146a IR |
1564 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
1565 | return false; | |
1566 | ||
8644a673 | 1567 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1568 | return false; |
1569 | ||
1570 | if (!is_gimple_assign (stmt)) | |
1571 | return false; | |
1572 | ||
1573 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
1574 | return false; | |
1575 | ||
1576 | code = gimple_assign_rhs_code (stmt); | |
1577 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
1578 | return false; | |
1579 | ||
1580 | /* Check types of lhs and rhs. */ | |
1581 | op0 = gimple_assign_rhs1 (stmt); | |
1582 | rhs_type = TREE_TYPE (op0); | |
1583 | vectype_in = get_vectype_for_scalar_type (rhs_type); | |
1584 | if (!vectype_in) | |
1585 | return false; | |
1586 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); | |
1587 | ||
1588 | scalar_dest = gimple_assign_lhs (stmt); | |
1589 | lhs_type = TREE_TYPE (scalar_dest); | |
1590 | vectype_out = get_vectype_for_scalar_type (lhs_type); | |
1591 | if (!vectype_out) | |
1592 | return false; | |
1593 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
1594 | ||
1595 | /* FORNOW */ | |
1596 | if (nunits_in == nunits_out / 2) | |
1597 | modifier = NARROW; | |
1598 | else if (nunits_out == nunits_in) | |
1599 | modifier = NONE; | |
1600 | else if (nunits_out == nunits_in / 2) | |
1601 | modifier = WIDEN; | |
1602 | else | |
1603 | return false; | |
1604 | ||
1605 | if (modifier == NONE) | |
1606 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) == vectype_out); | |
1607 | ||
1608 | /* Bail out if the types are both integral or non-integral. */ | |
1609 | if ((INTEGRAL_TYPE_P (rhs_type) && INTEGRAL_TYPE_P (lhs_type)) | |
1610 | || (!INTEGRAL_TYPE_P (rhs_type) && !INTEGRAL_TYPE_P (lhs_type))) | |
1611 | return false; | |
1612 | ||
1613 | integral_type = INTEGRAL_TYPE_P (rhs_type) ? vectype_in : vectype_out; | |
1614 | ||
1615 | if (modifier == NARROW) | |
1616 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
1617 | else | |
1618 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1619 | ||
1620 | /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies | |
1621 | this, so we can safely override NCOPIES with 1 here. */ | |
1622 | if (slp_node) | |
1623 | ncopies = 1; | |
1624 | ||
1625 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
1626 | needs to be generated. */ | |
1627 | gcc_assert (ncopies >= 1); | |
1628 | ||
1629 | /* Check the operands of the operation. */ | |
a70d6342 | 1630 | if (!vect_is_simple_use (op0, loop_vinfo, NULL, &def_stmt, &def, &dt[0])) |
ebfd146a IR |
1631 | { |
1632 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1633 | fprintf (vect_dump, "use not simple."); | |
1634 | return false; | |
1635 | } | |
1636 | ||
1637 | /* Supportable by target? */ | |
1638 | if ((modifier == NONE | |
1639 | && !targetm.vectorize.builtin_conversion (code, integral_type)) | |
1640 | || (modifier == WIDEN | |
1641 | && !supportable_widening_operation (code, stmt, vectype_in, | |
1642 | &decl1, &decl2, | |
1643 | &code1, &code2, | |
1644 | &dummy_int, &dummy)) | |
1645 | || (modifier == NARROW | |
1646 | && !supportable_narrowing_operation (code, stmt, vectype_in, | |
1647 | &code1, &dummy_int, &dummy))) | |
1648 | { | |
1649 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1650 | fprintf (vect_dump, "conversion not supported by target."); | |
1651 | return false; | |
1652 | } | |
1653 | ||
1654 | if (modifier != NONE) | |
1655 | { | |
1656 | STMT_VINFO_VECTYPE (stmt_info) = vectype_in; | |
1657 | /* FORNOW: SLP not supported. */ | |
1658 | if (STMT_SLP_TYPE (stmt_info)) | |
1659 | return false; | |
1660 | } | |
1661 | ||
1662 | if (!vec_stmt) /* transformation not required. */ | |
1663 | { | |
1664 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
1665 | return true; | |
1666 | } | |
1667 | ||
1668 | /** Transform. **/ | |
1669 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1670 | fprintf (vect_dump, "transform conversion."); | |
1671 | ||
1672 | /* Handle def. */ | |
1673 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1674 | ||
1675 | if (modifier == NONE && !slp_node) | |
1676 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
1677 | ||
1678 | prev_stmt_info = NULL; | |
1679 | switch (modifier) | |
1680 | { | |
1681 | case NONE: | |
1682 | for (j = 0; j < ncopies; j++) | |
1683 | { | |
ebfd146a IR |
1684 | if (j == 0) |
1685 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node); | |
1686 | else | |
1687 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
1688 | ||
1689 | builtin_decl = | |
1690 | targetm.vectorize.builtin_conversion (code, integral_type); | |
1691 | for (i = 0; VEC_iterate (tree, vec_oprnds0, i, vop0); i++) | |
1692 | { | |
1693 | /* Arguments are ready. create the new vector stmt. */ | |
1694 | new_stmt = gimple_build_call (builtin_decl, 1, vop0); | |
1695 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1696 | gimple_call_set_lhs (new_stmt, new_temp); | |
1697 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a IR |
1698 | if (slp_node) |
1699 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
1700 | } | |
1701 | ||
1702 | if (j == 0) | |
1703 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1704 | else | |
1705 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1706 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1707 | } | |
1708 | break; | |
1709 | ||
1710 | case WIDEN: | |
1711 | /* In case the vectorization factor (VF) is bigger than the number | |
1712 | of elements that we can fit in a vectype (nunits), we have to | |
1713 | generate more than one vector stmt - i.e - we need to "unroll" | |
1714 | the vector stmt by a factor VF/nunits. */ | |
1715 | for (j = 0; j < ncopies; j++) | |
1716 | { | |
1717 | if (j == 0) | |
1718 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
1719 | else | |
1720 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1721 | ||
1722 | STMT_VINFO_VECTYPE (stmt_info) = vectype_in; | |
1723 | ||
1724 | /* Generate first half of the widened result: */ | |
1725 | new_stmt | |
1726 | = vect_gen_widened_results_half (code1, decl1, | |
1727 | vec_oprnd0, vec_oprnd1, | |
1728 | unary_op, vec_dest, gsi, stmt); | |
1729 | if (j == 0) | |
1730 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1731 | else | |
1732 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1733 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1734 | ||
1735 | /* Generate second half of the widened result: */ | |
1736 | new_stmt | |
1737 | = vect_gen_widened_results_half (code2, decl2, | |
1738 | vec_oprnd0, vec_oprnd1, | |
1739 | unary_op, vec_dest, gsi, stmt); | |
1740 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1741 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1742 | } | |
1743 | break; | |
1744 | ||
1745 | case NARROW: | |
1746 | /* In case the vectorization factor (VF) is bigger than the number | |
1747 | of elements that we can fit in a vectype (nunits), we have to | |
1748 | generate more than one vector stmt - i.e - we need to "unroll" | |
1749 | the vector stmt by a factor VF/nunits. */ | |
1750 | for (j = 0; j < ncopies; j++) | |
1751 | { | |
1752 | /* Handle uses. */ | |
1753 | if (j == 0) | |
1754 | { | |
1755 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
1756 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1757 | } | |
1758 | else | |
1759 | { | |
1760 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd1); | |
1761 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1762 | } | |
1763 | ||
1764 | /* Arguments are ready. Create the new vector stmt. */ | |
1765 | expr = build2 (code1, vectype_out, vec_oprnd0, vec_oprnd1); | |
1766 | new_stmt = gimple_build_assign_with_ops (code1, vec_dest, vec_oprnd0, | |
1767 | vec_oprnd1); | |
1768 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1769 | gimple_assign_set_lhs (new_stmt, new_temp); | |
1770 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1771 | ||
1772 | if (j == 0) | |
1773 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1774 | else | |
1775 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1776 | ||
1777 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1778 | } | |
1779 | ||
1780 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
1781 | } | |
1782 | ||
1783 | if (vec_oprnds0) | |
1784 | VEC_free (tree, heap, vec_oprnds0); | |
1785 | ||
1786 | return true; | |
1787 | } | |
1788 | /* Function vectorizable_assignment. | |
1789 | ||
1790 | Check if STMT performs an assignment (copy) that can be vectorized. | |
1791 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1792 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
1793 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1794 | ||
1795 | static bool | |
1796 | vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi, | |
1797 | gimple *vec_stmt, slp_tree slp_node) | |
1798 | { | |
1799 | tree vec_dest; | |
1800 | tree scalar_dest; | |
1801 | tree op; | |
1802 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1803 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
1804 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1805 | tree new_temp; | |
1806 | tree def; | |
1807 | gimple def_stmt; | |
1808 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
1809 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
1810 | int ncopies; | |
1811 | int i; | |
1812 | VEC(tree,heap) *vec_oprnds = NULL; | |
1813 | tree vop; | |
a70d6342 | 1814 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1815 | |
1816 | /* Multiple types in SLP are handled by creating the appropriate number of | |
1817 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
1818 | case of SLP. */ | |
1819 | if (slp_node) | |
1820 | ncopies = 1; | |
1821 | else | |
1822 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
1823 | ||
1824 | gcc_assert (ncopies >= 1); | |
1825 | if (ncopies > 1) | |
1826 | return false; /* FORNOW */ | |
1827 | ||
a70d6342 | 1828 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
1829 | return false; |
1830 | ||
8644a673 | 1831 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1832 | return false; |
1833 | ||
1834 | /* Is vectorizable assignment? */ | |
1835 | if (!is_gimple_assign (stmt)) | |
1836 | return false; | |
1837 | ||
1838 | scalar_dest = gimple_assign_lhs (stmt); | |
1839 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
1840 | return false; | |
1841 | ||
1842 | if (gimple_assign_single_p (stmt) | |
1843 | || gimple_assign_rhs_code (stmt) == PAREN_EXPR) | |
1844 | op = gimple_assign_rhs1 (stmt); | |
1845 | else | |
1846 | return false; | |
1847 | ||
a70d6342 | 1848 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt[0])) |
ebfd146a IR |
1849 | { |
1850 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1851 | fprintf (vect_dump, "use not simple."); | |
1852 | return false; | |
1853 | } | |
1854 | ||
1855 | if (!vec_stmt) /* transformation not required. */ | |
1856 | { | |
1857 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
1858 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1859 | fprintf (vect_dump, "=== vectorizable_assignment ==="); | |
1860 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
1861 | return true; | |
1862 | } | |
1863 | ||
1864 | /** Transform. **/ | |
1865 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1866 | fprintf (vect_dump, "transform assignment."); | |
1867 | ||
1868 | /* Handle def. */ | |
1869 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
1870 | ||
1871 | /* Handle use. */ | |
1872 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); | |
1873 | ||
1874 | /* Arguments are ready. create the new vector stmt. */ | |
1875 | for (i = 0; VEC_iterate (tree, vec_oprnds, i, vop); i++) | |
1876 | { | |
1877 | *vec_stmt = gimple_build_assign (vec_dest, vop); | |
1878 | new_temp = make_ssa_name (vec_dest, *vec_stmt); | |
1879 | gimple_assign_set_lhs (*vec_stmt, new_temp); | |
1880 | vect_finish_stmt_generation (stmt, *vec_stmt, gsi); | |
1881 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt; | |
1882 | ||
1883 | if (slp_node) | |
1884 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), *vec_stmt); | |
1885 | } | |
1886 | ||
1887 | VEC_free (tree, heap, vec_oprnds); | |
1888 | return true; | |
1889 | } | |
1890 | ||
1891 | /* Function vectorizable_operation. | |
1892 | ||
1893 | Check if STMT performs a binary or unary operation that can be vectorized. | |
1894 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1895 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
1896 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1897 | ||
1898 | static bool | |
1899 | vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi, | |
1900 | gimple *vec_stmt, slp_tree slp_node) | |
1901 | { | |
1902 | tree vec_dest; | |
1903 | tree scalar_dest; | |
1904 | tree op0, op1 = NULL; | |
1905 | tree vec_oprnd1 = NULL_TREE; | |
1906 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1907 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
1908 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1909 | enum tree_code code; | |
1910 | enum machine_mode vec_mode; | |
1911 | tree new_temp; | |
1912 | int op_type; | |
1913 | optab optab; | |
1914 | int icode; | |
1915 | enum machine_mode optab_op2_mode; | |
1916 | tree def; | |
1917 | gimple def_stmt; | |
1918 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
1919 | gimple new_stmt = NULL; | |
1920 | stmt_vec_info prev_stmt_info; | |
1921 | int nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
1922 | int nunits_out; | |
1923 | tree vectype_out; | |
1924 | int ncopies; | |
1925 | int j, i; | |
1926 | VEC(tree,heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; | |
1927 | tree vop0, vop1; | |
1928 | unsigned int k; | |
1929 | bool shift_p = false; | |
1930 | bool scalar_shift_arg = false; | |
a70d6342 IR |
1931 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
1932 | int vf; | |
1933 | ||
1934 | if (loop_vinfo) | |
1935 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
1936 | else | |
1937 | /* FORNOW: multiple types are not supported in basic block SLP. */ | |
1938 | vf = nunits_in; | |
ebfd146a IR |
1939 | |
1940 | /* Multiple types in SLP are handled by creating the appropriate number of | |
1941 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
1942 | case of SLP. */ | |
1943 | if (slp_node) | |
1944 | ncopies = 1; | |
1945 | else | |
1946 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1947 | ||
1948 | gcc_assert (ncopies >= 1); | |
1949 | ||
a70d6342 | 1950 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
1951 | return false; |
1952 | ||
8644a673 | 1953 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1954 | return false; |
1955 | ||
1956 | /* Is STMT a vectorizable binary/unary operation? */ | |
1957 | if (!is_gimple_assign (stmt)) | |
1958 | return false; | |
1959 | ||
1960 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
1961 | return false; | |
1962 | ||
1963 | scalar_dest = gimple_assign_lhs (stmt); | |
1964 | vectype_out = get_vectype_for_scalar_type (TREE_TYPE (scalar_dest)); | |
1965 | if (!vectype_out) | |
1966 | return false; | |
1967 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
1968 | if (nunits_out != nunits_in) | |
1969 | return false; | |
1970 | ||
1971 | code = gimple_assign_rhs_code (stmt); | |
1972 | ||
1973 | /* For pointer addition, we should use the normal plus for | |
1974 | the vector addition. */ | |
1975 | if (code == POINTER_PLUS_EXPR) | |
1976 | code = PLUS_EXPR; | |
1977 | ||
1978 | /* Support only unary or binary operations. */ | |
1979 | op_type = TREE_CODE_LENGTH (code); | |
1980 | if (op_type != unary_op && op_type != binary_op) | |
1981 | { | |
1982 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1983 | fprintf (vect_dump, "num. args = %d (not unary/binary op).", op_type); | |
1984 | return false; | |
1985 | } | |
1986 | ||
1987 | op0 = gimple_assign_rhs1 (stmt); | |
a70d6342 | 1988 | if (!vect_is_simple_use (op0, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt[0])) |
ebfd146a IR |
1989 | { |
1990 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1991 | fprintf (vect_dump, "use not simple."); | |
1992 | return false; | |
1993 | } | |
1994 | ||
1995 | if (op_type == binary_op) | |
1996 | { | |
1997 | op1 = gimple_assign_rhs2 (stmt); | |
a70d6342 IR |
1998 | if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def, |
1999 | &dt[1])) | |
ebfd146a IR |
2000 | { |
2001 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2002 | fprintf (vect_dump, "use not simple."); | |
2003 | return false; | |
2004 | } | |
2005 | } | |
2006 | ||
2007 | /* If this is a shift/rotate, determine whether the shift amount is a vector, | |
2008 | or scalar. If the shift/rotate amount is a vector, use the vector/vector | |
2009 | shift optabs. */ | |
2010 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
2011 | || code == RROTATE_EXPR) | |
2012 | { | |
2013 | shift_p = true; | |
2014 | ||
2015 | /* vector shifted by vector */ | |
8644a673 | 2016 | if (dt[1] == vect_internal_def) |
ebfd146a IR |
2017 | { |
2018 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
2019 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2020 | fprintf (vect_dump, "vector/vector shift/rotate found."); | |
2021 | } | |
2022 | ||
2023 | /* See if the machine has a vector shifted by scalar insn and if not | |
2024 | then see if it has a vector shifted by vector insn */ | |
8644a673 | 2025 | else if (dt[1] == vect_constant_def || dt[1] == vect_external_def) |
ebfd146a IR |
2026 | { |
2027 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
2028 | if (optab | |
2029 | && (optab_handler (optab, TYPE_MODE (vectype))->insn_code | |
2030 | != CODE_FOR_nothing)) | |
2031 | { | |
2032 | scalar_shift_arg = true; | |
2033 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2034 | fprintf (vect_dump, "vector/scalar shift/rotate found."); | |
2035 | } | |
2036 | else | |
2037 | { | |
2038 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
ad6c0864 | 2039 | if (optab |
ebfd146a IR |
2040 | && (optab_handler (optab, TYPE_MODE (vectype))->insn_code |
2041 | != CODE_FOR_nothing)) | |
ad6c0864 MM |
2042 | { |
2043 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2044 | fprintf (vect_dump, "vector/vector shift/rotate found."); | |
2045 | ||
2046 | /* Unlike the other binary operators, shifts/rotates have | |
2047 | the rhs being int, instead of the same type as the lhs, | |
2048 | so make sure the scalar is the right type if we are | |
2049 | dealing with vectors of short/char. */ | |
2050 | if (dt[1] == vect_constant_def) | |
2051 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
2052 | } | |
ebfd146a IR |
2053 | } |
2054 | } | |
2055 | ||
2056 | else | |
2057 | { | |
2058 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2059 | fprintf (vect_dump, "operand mode requires invariant argument."); | |
2060 | return false; | |
2061 | } | |
2062 | } | |
2063 | else | |
2064 | optab = optab_for_tree_code (code, vectype, optab_default); | |
2065 | ||
2066 | /* Supportable by target? */ | |
2067 | if (!optab) | |
2068 | { | |
2069 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2070 | fprintf (vect_dump, "no optab."); | |
2071 | return false; | |
2072 | } | |
2073 | vec_mode = TYPE_MODE (vectype); | |
2074 | icode = (int) optab_handler (optab, vec_mode)->insn_code; | |
2075 | if (icode == CODE_FOR_nothing) | |
2076 | { | |
2077 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2078 | fprintf (vect_dump, "op not supported by target."); | |
2079 | /* Check only during analysis. */ | |
2080 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
a70d6342 | 2081 | || (vf < vect_min_worthwhile_factor (code) |
ebfd146a IR |
2082 | && !vec_stmt)) |
2083 | return false; | |
2084 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2085 | fprintf (vect_dump, "proceeding using word mode."); | |
2086 | } | |
2087 | ||
2088 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
2089 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
a70d6342 | 2090 | && vf < vect_min_worthwhile_factor (code) |
ebfd146a IR |
2091 | && !vec_stmt) |
2092 | { | |
2093 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2094 | fprintf (vect_dump, "not worthwhile without SIMD support."); | |
2095 | return false; | |
2096 | } | |
2097 | ||
2098 | if (!vec_stmt) /* transformation not required. */ | |
2099 | { | |
2100 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; | |
2101 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2102 | fprintf (vect_dump, "=== vectorizable_operation ==="); | |
2103 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2104 | return true; | |
2105 | } | |
2106 | ||
2107 | /** Transform. **/ | |
2108 | ||
2109 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2110 | fprintf (vect_dump, "transform binary/unary operation."); | |
2111 | ||
2112 | /* Handle def. */ | |
2113 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2114 | ||
2115 | /* Allocate VECs for vector operands. In case of SLP, vector operands are | |
2116 | created in the previous stages of the recursion, so no allocation is | |
2117 | needed, except for the case of shift with scalar shift argument. In that | |
2118 | case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to | |
2119 | be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE. | |
2120 | In case of loop-based vectorization we allocate VECs of size 1. We | |
2121 | allocate VEC_OPRNDS1 only in case of binary operation. */ | |
2122 | if (!slp_node) | |
2123 | { | |
2124 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
2125 | if (op_type == binary_op) | |
2126 | vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
2127 | } | |
2128 | else if (scalar_shift_arg) | |
2129 | vec_oprnds1 = VEC_alloc (tree, heap, slp_node->vec_stmts_size); | |
2130 | ||
2131 | /* In case the vectorization factor (VF) is bigger than the number | |
2132 | of elements that we can fit in a vectype (nunits), we have to generate | |
2133 | more than one vector stmt - i.e - we need to "unroll" the | |
2134 | vector stmt by a factor VF/nunits. In doing so, we record a pointer | |
2135 | from one copy of the vector stmt to the next, in the field | |
2136 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
2137 | stages to find the correct vector defs to be used when vectorizing | |
2138 | stmts that use the defs of the current stmt. The example below illustrates | |
2139 | the vectorization process when VF=16 and nunits=4 (i.e - we need to create | |
2140 | 4 vectorized stmts): | |
2141 | ||
2142 | before vectorization: | |
2143 | RELATED_STMT VEC_STMT | |
2144 | S1: x = memref - - | |
2145 | S2: z = x + 1 - - | |
2146 | ||
2147 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
2148 | there): | |
2149 | RELATED_STMT VEC_STMT | |
2150 | VS1_0: vx0 = memref0 VS1_1 - | |
2151 | VS1_1: vx1 = memref1 VS1_2 - | |
2152 | VS1_2: vx2 = memref2 VS1_3 - | |
2153 | VS1_3: vx3 = memref3 - - | |
2154 | S1: x = load - VS1_0 | |
2155 | S2: z = x + 1 - - | |
2156 | ||
2157 | step2: vectorize stmt S2 (done here): | |
2158 | To vectorize stmt S2 we first need to find the relevant vector | |
2159 | def for the first operand 'x'. This is, as usual, obtained from | |
2160 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
2161 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
2162 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
2163 | the vector stmt VS2_0, and as usual, record it in the | |
2164 | STMT_VINFO_VEC_STMT of stmt S2. | |
2165 | When creating the second copy (VS2_1), we obtain the relevant vector | |
2166 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
2167 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
2168 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
2169 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
2170 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
2171 | chain of stmts and pointers: | |
2172 | RELATED_STMT VEC_STMT | |
2173 | VS1_0: vx0 = memref0 VS1_1 - | |
2174 | VS1_1: vx1 = memref1 VS1_2 - | |
2175 | VS1_2: vx2 = memref2 VS1_3 - | |
2176 | VS1_3: vx3 = memref3 - - | |
2177 | S1: x = load - VS1_0 | |
2178 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
2179 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
2180 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
2181 | VS2_3: vz3 = vx3 + v1 - - | |
2182 | S2: z = x + 1 - VS2_0 */ | |
2183 | ||
2184 | prev_stmt_info = NULL; | |
2185 | for (j = 0; j < ncopies; j++) | |
2186 | { | |
2187 | /* Handle uses. */ | |
2188 | if (j == 0) | |
2189 | { | |
2190 | if (op_type == binary_op && scalar_shift_arg) | |
2191 | { | |
2192 | /* Vector shl and shr insn patterns can be defined with scalar | |
2193 | operand 2 (shift operand). In this case, use constant or loop | |
2194 | invariant op1 directly, without extending it to vector mode | |
2195 | first. */ | |
2196 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
2197 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
2198 | { | |
2199 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2200 | fprintf (vect_dump, "operand 1 using scalar mode."); | |
2201 | vec_oprnd1 = op1; | |
2202 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2203 | if (slp_node) | |
2204 | { | |
2205 | /* Store vec_oprnd1 for every vector stmt to be created | |
2206 | for SLP_NODE. We check during the analysis that all the | |
2207 | shift arguments are the same. | |
2208 | TODO: Allow different constants for different vector | |
2209 | stmts generated for an SLP instance. */ | |
2210 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
2211 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2212 | } | |
2213 | } | |
2214 | } | |
2215 | ||
2216 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
2217 | (a special case for certain kind of vector shifts); otherwise, | |
2218 | operand 1 should be of a vector type (the usual case). */ | |
2219 | if (op_type == binary_op && !vec_oprnd1) | |
2220 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
2221 | slp_node); | |
2222 | else | |
2223 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2224 | slp_node); | |
2225 | } | |
2226 | else | |
2227 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
2228 | ||
2229 | /* Arguments are ready. Create the new vector stmt. */ | |
2230 | for (i = 0; VEC_iterate (tree, vec_oprnds0, i, vop0); i++) | |
2231 | { | |
2232 | vop1 = ((op_type == binary_op) | |
2233 | ? VEC_index (tree, vec_oprnds1, i) : NULL); | |
2234 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); | |
2235 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2236 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2237 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2238 | if (slp_node) | |
2239 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2240 | } | |
2241 | ||
2242 | if (slp_node) | |
2243 | continue; | |
2244 | ||
2245 | if (j == 0) | |
2246 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2247 | else | |
2248 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2249 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2250 | } | |
2251 | ||
2252 | VEC_free (tree, heap, vec_oprnds0); | |
2253 | if (vec_oprnds1) | |
2254 | VEC_free (tree, heap, vec_oprnds1); | |
2255 | ||
2256 | return true; | |
2257 | } | |
2258 | ||
2259 | ||
2260 | /* Get vectorized definitions for loop-based vectorization. For the first | |
2261 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
2262 | scalar operand), and for the rest we get a copy with | |
2263 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
2264 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
2265 | The vectors are collected into VEC_OPRNDS. */ | |
2266 | ||
2267 | static void | |
2268 | vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt, | |
2269 | VEC (tree, heap) **vec_oprnds, int multi_step_cvt) | |
2270 | { | |
2271 | tree vec_oprnd; | |
2272 | ||
2273 | /* Get first vector operand. */ | |
2274 | /* All the vector operands except the very first one (that is scalar oprnd) | |
2275 | are stmt copies. */ | |
2276 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
2277 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL); | |
2278 | else | |
2279 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
2280 | ||
2281 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
2282 | ||
2283 | /* Get second vector operand. */ | |
2284 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
2285 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
2286 | ||
2287 | *oprnd = vec_oprnd; | |
2288 | ||
2289 | /* For conversion in multiple steps, continue to get operands | |
2290 | recursively. */ | |
2291 | if (multi_step_cvt) | |
2292 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
2293 | } | |
2294 | ||
2295 | ||
2296 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
2297 | For multi-step conversions store the resulting vectors and call the function | |
2298 | recursively. */ | |
2299 | ||
2300 | static void | |
2301 | vect_create_vectorized_demotion_stmts (VEC (tree, heap) **vec_oprnds, | |
2302 | int multi_step_cvt, gimple stmt, | |
2303 | VEC (tree, heap) *vec_dsts, | |
2304 | gimple_stmt_iterator *gsi, | |
2305 | slp_tree slp_node, enum tree_code code, | |
2306 | stmt_vec_info *prev_stmt_info) | |
2307 | { | |
2308 | unsigned int i; | |
2309 | tree vop0, vop1, new_tmp, vec_dest; | |
2310 | gimple new_stmt; | |
2311 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2312 | ||
2313 | vec_dest = VEC_pop (tree, vec_dsts); | |
2314 | ||
2315 | for (i = 0; i < VEC_length (tree, *vec_oprnds); i += 2) | |
2316 | { | |
2317 | /* Create demotion operation. */ | |
2318 | vop0 = VEC_index (tree, *vec_oprnds, i); | |
2319 | vop1 = VEC_index (tree, *vec_oprnds, i + 1); | |
2320 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); | |
2321 | new_tmp = make_ssa_name (vec_dest, new_stmt); | |
2322 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
2323 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2324 | ||
2325 | if (multi_step_cvt) | |
2326 | /* Store the resulting vector for next recursive call. */ | |
2327 | VEC_replace (tree, *vec_oprnds, i/2, new_tmp); | |
2328 | else | |
2329 | { | |
2330 | /* This is the last step of the conversion sequence. Store the | |
2331 | vectors in SLP_NODE or in vector info of the scalar statement | |
2332 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
2333 | if (slp_node) | |
2334 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2335 | else | |
2336 | { | |
2337 | if (!*prev_stmt_info) | |
2338 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2339 | else | |
2340 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
2341 | ||
2342 | *prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2343 | } | |
2344 | } | |
2345 | } | |
2346 | ||
2347 | /* For multi-step demotion operations we first generate demotion operations | |
2348 | from the source type to the intermediate types, and then combine the | |
2349 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
2350 | type. */ | |
2351 | if (multi_step_cvt) | |
2352 | { | |
2353 | /* At each level of recursion we have have of the operands we had at the | |
2354 | previous level. */ | |
2355 | VEC_truncate (tree, *vec_oprnds, (i+1)/2); | |
2356 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, | |
2357 | stmt, vec_dsts, gsi, slp_node, | |
2358 | code, prev_stmt_info); | |
2359 | } | |
2360 | } | |
2361 | ||
2362 | ||
2363 | /* Function vectorizable_type_demotion | |
2364 | ||
2365 | Check if STMT performs a binary or unary operation that involves | |
2366 | type demotion, and if it can be vectorized. | |
2367 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2368 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2369 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2370 | ||
2371 | static bool | |
2372 | vectorizable_type_demotion (gimple stmt, gimple_stmt_iterator *gsi, | |
2373 | gimple *vec_stmt, slp_tree slp_node) | |
2374 | { | |
2375 | tree vec_dest; | |
2376 | tree scalar_dest; | |
2377 | tree op0; | |
2378 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2379 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2380 | enum tree_code code, code1 = ERROR_MARK; | |
2381 | tree def; | |
2382 | gimple def_stmt; | |
2383 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2384 | stmt_vec_info prev_stmt_info; | |
2385 | int nunits_in; | |
2386 | int nunits_out; | |
2387 | tree vectype_out; | |
2388 | int ncopies; | |
2389 | int j, i; | |
2390 | tree vectype_in; | |
2391 | int multi_step_cvt = 0; | |
2392 | VEC (tree, heap) *vec_oprnds0 = NULL; | |
2393 | VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL; | |
2394 | tree last_oprnd, intermediate_type; | |
2395 | ||
a70d6342 IR |
2396 | /* FORNOW: not supported by basic block SLP vectorization. */ |
2397 | gcc_assert (loop_vinfo); | |
2398 | ||
ebfd146a IR |
2399 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
2400 | return false; | |
2401 | ||
8644a673 | 2402 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2403 | return false; |
2404 | ||
2405 | /* Is STMT a vectorizable type-demotion operation? */ | |
2406 | if (!is_gimple_assign (stmt)) | |
2407 | return false; | |
2408 | ||
2409 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2410 | return false; | |
2411 | ||
2412 | code = gimple_assign_rhs_code (stmt); | |
2413 | if (!CONVERT_EXPR_CODE_P (code)) | |
2414 | return false; | |
2415 | ||
2416 | op0 = gimple_assign_rhs1 (stmt); | |
2417 | vectype_in = get_vectype_for_scalar_type (TREE_TYPE (op0)); | |
2418 | if (!vectype_in) | |
2419 | return false; | |
2420 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); | |
2421 | ||
2422 | scalar_dest = gimple_assign_lhs (stmt); | |
2423 | vectype_out = get_vectype_for_scalar_type (TREE_TYPE (scalar_dest)); | |
2424 | if (!vectype_out) | |
2425 | return false; | |
2426 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
2427 | if (nunits_in >= nunits_out) | |
2428 | return false; | |
2429 | ||
2430 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2431 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2432 | case of SLP. */ | |
2433 | if (slp_node) | |
2434 | ncopies = 1; | |
2435 | else | |
2436 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
ebfd146a IR |
2437 | gcc_assert (ncopies >= 1); |
2438 | ||
2439 | if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2440 | && INTEGRAL_TYPE_P (TREE_TYPE (op0))) | |
2441 | || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest)) | |
2442 | && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0)) | |
2443 | && CONVERT_EXPR_CODE_P (code)))) | |
2444 | return false; | |
2445 | ||
2446 | /* Check the operands of the operation. */ | |
a70d6342 | 2447 | if (!vect_is_simple_use (op0, loop_vinfo, NULL, &def_stmt, &def, &dt[0])) |
ebfd146a IR |
2448 | { |
2449 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2450 | fprintf (vect_dump, "use not simple."); | |
2451 | return false; | |
2452 | } | |
2453 | ||
2454 | /* Supportable by target? */ | |
2455 | if (!supportable_narrowing_operation (code, stmt, vectype_in, &code1, | |
2456 | &multi_step_cvt, &interm_types)) | |
2457 | return false; | |
2458 | ||
2459 | STMT_VINFO_VECTYPE (stmt_info) = vectype_in; | |
2460 | ||
2461 | if (!vec_stmt) /* transformation not required. */ | |
2462 | { | |
2463 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
2464 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2465 | fprintf (vect_dump, "=== vectorizable_demotion ==="); | |
2466 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2467 | return true; | |
2468 | } | |
2469 | ||
2470 | /** Transform. **/ | |
2471 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2472 | fprintf (vect_dump, "transform type demotion operation. ncopies = %d.", | |
2473 | ncopies); | |
2474 | ||
2475 | /* In case of multi-step demotion, we first generate demotion operations to | |
2476 | the intermediate types, and then from that types to the final one. | |
2477 | We create vector destinations for the intermediate type (TYPES) received | |
2478 | from supportable_narrowing_operation, and store them in the correct order | |
2479 | for future use in vect_create_vectorized_demotion_stmts(). */ | |
2480 | if (multi_step_cvt) | |
2481 | vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1); | |
2482 | else | |
2483 | vec_dsts = VEC_alloc (tree, heap, 1); | |
2484 | ||
2485 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2486 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2487 | ||
2488 | if (multi_step_cvt) | |
2489 | { | |
2490 | for (i = VEC_length (tree, interm_types) - 1; | |
2491 | VEC_iterate (tree, interm_types, i, intermediate_type); i--) | |
2492 | { | |
2493 | vec_dest = vect_create_destination_var (scalar_dest, | |
2494 | intermediate_type); | |
2495 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2496 | } | |
2497 | } | |
2498 | ||
2499 | /* In case the vectorization factor (VF) is bigger than the number | |
2500 | of elements that we can fit in a vectype (nunits), we have to generate | |
2501 | more than one vector stmt - i.e - we need to "unroll" the | |
2502 | vector stmt by a factor VF/nunits. */ | |
2503 | last_oprnd = op0; | |
2504 | prev_stmt_info = NULL; | |
2505 | for (j = 0; j < ncopies; j++) | |
2506 | { | |
2507 | /* Handle uses. */ | |
2508 | if (slp_node) | |
2509 | vect_get_slp_defs (slp_node, &vec_oprnds0, NULL); | |
2510 | else | |
2511 | { | |
2512 | VEC_free (tree, heap, vec_oprnds0); | |
2513 | vec_oprnds0 = VEC_alloc (tree, heap, | |
2514 | (multi_step_cvt ? vect_pow2 (multi_step_cvt) * 2 : 2)); | |
2515 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, | |
2516 | vect_pow2 (multi_step_cvt) - 1); | |
2517 | } | |
2518 | ||
2519 | /* Arguments are ready. Create the new vector stmts. */ | |
2520 | tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts); | |
2521 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, | |
2522 | multi_step_cvt, stmt, tmp_vec_dsts, | |
2523 | gsi, slp_node, code1, | |
2524 | &prev_stmt_info); | |
2525 | } | |
2526 | ||
2527 | VEC_free (tree, heap, vec_oprnds0); | |
2528 | VEC_free (tree, heap, vec_dsts); | |
2529 | VEC_free (tree, heap, tmp_vec_dsts); | |
2530 | VEC_free (tree, heap, interm_types); | |
2531 | ||
2532 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2533 | return true; | |
2534 | } | |
2535 | ||
2536 | ||
2537 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
2538 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
2539 | the resulting vectors and call the function recursively. */ | |
2540 | ||
2541 | static void | |
2542 | vect_create_vectorized_promotion_stmts (VEC (tree, heap) **vec_oprnds0, | |
2543 | VEC (tree, heap) **vec_oprnds1, | |
2544 | int multi_step_cvt, gimple stmt, | |
2545 | VEC (tree, heap) *vec_dsts, | |
2546 | gimple_stmt_iterator *gsi, | |
2547 | slp_tree slp_node, enum tree_code code1, | |
2548 | enum tree_code code2, tree decl1, | |
2549 | tree decl2, int op_type, | |
2550 | stmt_vec_info *prev_stmt_info) | |
2551 | { | |
2552 | int i; | |
2553 | tree vop0, vop1, new_tmp1, new_tmp2, vec_dest; | |
2554 | gimple new_stmt1, new_stmt2; | |
2555 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2556 | VEC (tree, heap) *vec_tmp; | |
2557 | ||
2558 | vec_dest = VEC_pop (tree, vec_dsts); | |
2559 | vec_tmp = VEC_alloc (tree, heap, VEC_length (tree, *vec_oprnds0) * 2); | |
2560 | ||
2561 | for (i = 0; VEC_iterate (tree, *vec_oprnds0, i, vop0); i++) | |
2562 | { | |
2563 | if (op_type == binary_op) | |
2564 | vop1 = VEC_index (tree, *vec_oprnds1, i); | |
2565 | else | |
2566 | vop1 = NULL_TREE; | |
2567 | ||
2568 | /* Generate the two halves of promotion operation. */ | |
2569 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
2570 | op_type, vec_dest, gsi, stmt); | |
2571 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
2572 | op_type, vec_dest, gsi, stmt); | |
2573 | if (is_gimple_call (new_stmt1)) | |
2574 | { | |
2575 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
2576 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
2577 | } | |
2578 | else | |
2579 | { | |
2580 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
2581 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
2582 | } | |
2583 | ||
2584 | if (multi_step_cvt) | |
2585 | { | |
2586 | /* Store the results for the recursive call. */ | |
2587 | VEC_quick_push (tree, vec_tmp, new_tmp1); | |
2588 | VEC_quick_push (tree, vec_tmp, new_tmp2); | |
2589 | } | |
2590 | else | |
2591 | { | |
2592 | /* Last step of promotion sequience - store the results. */ | |
2593 | if (slp_node) | |
2594 | { | |
2595 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt1); | |
2596 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt2); | |
2597 | } | |
2598 | else | |
2599 | { | |
2600 | if (!*prev_stmt_info) | |
2601 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt1; | |
2602 | else | |
2603 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt1; | |
2604 | ||
2605 | *prev_stmt_info = vinfo_for_stmt (new_stmt1); | |
2606 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt2; | |
2607 | *prev_stmt_info = vinfo_for_stmt (new_stmt2); | |
2608 | } | |
2609 | } | |
2610 | } | |
2611 | ||
2612 | if (multi_step_cvt) | |
2613 | { | |
2614 | /* For multi-step promotion operation we first generate we call the | |
2615 | function recurcively for every stage. We start from the input type, | |
2616 | create promotion operations to the intermediate types, and then | |
2617 | create promotions to the output type. */ | |
2618 | *vec_oprnds0 = VEC_copy (tree, heap, vec_tmp); | |
2619 | VEC_free (tree, heap, vec_tmp); | |
2620 | vect_create_vectorized_promotion_stmts (vec_oprnds0, vec_oprnds1, | |
2621 | multi_step_cvt - 1, stmt, | |
2622 | vec_dsts, gsi, slp_node, code1, | |
2623 | code2, decl2, decl2, op_type, | |
2624 | prev_stmt_info); | |
2625 | } | |
2626 | } | |
a70d6342 | 2627 | |
ebfd146a IR |
2628 | |
2629 | /* Function vectorizable_type_promotion | |
2630 | ||
2631 | Check if STMT performs a binary or unary operation that involves | |
2632 | type promotion, and if it can be vectorized. | |
2633 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2634 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2635 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2636 | ||
2637 | static bool | |
2638 | vectorizable_type_promotion (gimple stmt, gimple_stmt_iterator *gsi, | |
2639 | gimple *vec_stmt, slp_tree slp_node) | |
2640 | { | |
2641 | tree vec_dest; | |
2642 | tree scalar_dest; | |
2643 | tree op0, op1 = NULL; | |
2644 | tree vec_oprnd0=NULL, vec_oprnd1=NULL; | |
2645 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2646 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2647 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
2648 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; | |
2649 | int op_type; | |
2650 | tree def; | |
2651 | gimple def_stmt; | |
2652 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2653 | stmt_vec_info prev_stmt_info; | |
2654 | int nunits_in; | |
2655 | int nunits_out; | |
2656 | tree vectype_out; | |
2657 | int ncopies; | |
2658 | int j, i; | |
2659 | tree vectype_in; | |
2660 | tree intermediate_type = NULL_TREE; | |
2661 | int multi_step_cvt = 0; | |
2662 | VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; | |
2663 | VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL; | |
2664 | ||
a70d6342 IR |
2665 | /* FORNOW: not supported by basic block SLP vectorization. */ |
2666 | gcc_assert (loop_vinfo); | |
2667 | ||
ebfd146a IR |
2668 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
2669 | return false; | |
2670 | ||
8644a673 | 2671 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2672 | return false; |
2673 | ||
2674 | /* Is STMT a vectorizable type-promotion operation? */ | |
2675 | if (!is_gimple_assign (stmt)) | |
2676 | return false; | |
2677 | ||
2678 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2679 | return false; | |
2680 | ||
2681 | code = gimple_assign_rhs_code (stmt); | |
2682 | if (!CONVERT_EXPR_CODE_P (code) | |
2683 | && code != WIDEN_MULT_EXPR) | |
2684 | return false; | |
2685 | ||
2686 | op0 = gimple_assign_rhs1 (stmt); | |
2687 | vectype_in = get_vectype_for_scalar_type (TREE_TYPE (op0)); | |
2688 | if (!vectype_in) | |
2689 | return false; | |
2690 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); | |
2691 | ||
2692 | scalar_dest = gimple_assign_lhs (stmt); | |
2693 | vectype_out = get_vectype_for_scalar_type (TREE_TYPE (scalar_dest)); | |
2694 | if (!vectype_out) | |
2695 | return false; | |
2696 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
2697 | if (nunits_in <= nunits_out) | |
2698 | return false; | |
2699 | ||
2700 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2701 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2702 | case of SLP. */ | |
2703 | if (slp_node) | |
2704 | ncopies = 1; | |
2705 | else | |
2706 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2707 | ||
2708 | gcc_assert (ncopies >= 1); | |
2709 | ||
2710 | if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2711 | && INTEGRAL_TYPE_P (TREE_TYPE (op0))) | |
2712 | || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest)) | |
2713 | && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0)) | |
2714 | && CONVERT_EXPR_CODE_P (code)))) | |
2715 | return false; | |
2716 | ||
2717 | /* Check the operands of the operation. */ | |
a70d6342 | 2718 | if (!vect_is_simple_use (op0, loop_vinfo, NULL, &def_stmt, &def, &dt[0])) |
ebfd146a IR |
2719 | { |
2720 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2721 | fprintf (vect_dump, "use not simple."); | |
2722 | return false; | |
2723 | } | |
2724 | ||
2725 | op_type = TREE_CODE_LENGTH (code); | |
2726 | if (op_type == binary_op) | |
2727 | { | |
2728 | op1 = gimple_assign_rhs2 (stmt); | |
a70d6342 | 2729 | if (!vect_is_simple_use (op1, loop_vinfo, NULL, &def_stmt, &def, &dt[1])) |
ebfd146a IR |
2730 | { |
2731 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2732 | fprintf (vect_dump, "use not simple."); | |
2733 | return false; | |
2734 | } | |
2735 | } | |
2736 | ||
2737 | /* Supportable by target? */ | |
2738 | if (!supportable_widening_operation (code, stmt, vectype_in, | |
2739 | &decl1, &decl2, &code1, &code2, | |
2740 | &multi_step_cvt, &interm_types)) | |
2741 | return false; | |
2742 | ||
2743 | /* Binary widening operation can only be supported directly by the | |
2744 | architecture. */ | |
2745 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
2746 | ||
2747 | STMT_VINFO_VECTYPE (stmt_info) = vectype_in; | |
2748 | ||
2749 | if (!vec_stmt) /* transformation not required. */ | |
2750 | { | |
2751 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
2752 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2753 | fprintf (vect_dump, "=== vectorizable_promotion ==="); | |
2754 | vect_model_simple_cost (stmt_info, 2*ncopies, dt, NULL); | |
2755 | return true; | |
2756 | } | |
2757 | ||
2758 | /** Transform. **/ | |
2759 | ||
2760 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2761 | fprintf (vect_dump, "transform type promotion operation. ncopies = %d.", | |
2762 | ncopies); | |
2763 | ||
2764 | /* Handle def. */ | |
2765 | /* In case of multi-step promotion, we first generate promotion operations | |
2766 | to the intermediate types, and then from that types to the final one. | |
2767 | We store vector destination in VEC_DSTS in the correct order for | |
2768 | recursive creation of promotion operations in | |
2769 | vect_create_vectorized_promotion_stmts(). Vector destinations are created | |
2770 | according to TYPES recieved from supportable_widening_operation(). */ | |
2771 | if (multi_step_cvt) | |
2772 | vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1); | |
2773 | else | |
2774 | vec_dsts = VEC_alloc (tree, heap, 1); | |
2775 | ||
2776 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2777 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2778 | ||
2779 | if (multi_step_cvt) | |
2780 | { | |
2781 | for (i = VEC_length (tree, interm_types) - 1; | |
2782 | VEC_iterate (tree, interm_types, i, intermediate_type); i--) | |
2783 | { | |
2784 | vec_dest = vect_create_destination_var (scalar_dest, | |
2785 | intermediate_type); | |
2786 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2787 | } | |
2788 | } | |
2789 | ||
2790 | if (!slp_node) | |
2791 | { | |
2792 | vec_oprnds0 = VEC_alloc (tree, heap, | |
2793 | (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
2794 | if (op_type == binary_op) | |
2795 | vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
2796 | } | |
2797 | ||
2798 | /* In case the vectorization factor (VF) is bigger than the number | |
2799 | of elements that we can fit in a vectype (nunits), we have to generate | |
2800 | more than one vector stmt - i.e - we need to "unroll" the | |
2801 | vector stmt by a factor VF/nunits. */ | |
2802 | ||
2803 | prev_stmt_info = NULL; | |
2804 | for (j = 0; j < ncopies; j++) | |
2805 | { | |
2806 | /* Handle uses. */ | |
2807 | if (j == 0) | |
2808 | { | |
2809 | if (slp_node) | |
2810 | vect_get_slp_defs (slp_node, &vec_oprnds0, &vec_oprnds1); | |
2811 | else | |
2812 | { | |
2813 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
2814 | VEC_quick_push (tree, vec_oprnds0, vec_oprnd0); | |
2815 | if (op_type == binary_op) | |
2816 | { | |
2817 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, NULL); | |
2818 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2819 | } | |
2820 | } | |
2821 | } | |
2822 | else | |
2823 | { | |
2824 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
2825 | VEC_replace (tree, vec_oprnds0, 0, vec_oprnd0); | |
2826 | if (op_type == binary_op) | |
2827 | { | |
2828 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd1); | |
2829 | VEC_replace (tree, vec_oprnds1, 0, vec_oprnd1); | |
2830 | } | |
2831 | } | |
2832 | ||
2833 | /* Arguments are ready. Create the new vector stmts. */ | |
2834 | tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts); | |
2835 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, &vec_oprnds1, | |
2836 | multi_step_cvt, stmt, | |
2837 | tmp_vec_dsts, | |
2838 | gsi, slp_node, code1, code2, | |
2839 | decl1, decl2, op_type, | |
2840 | &prev_stmt_info); | |
2841 | } | |
2842 | ||
2843 | VEC_free (tree, heap, vec_dsts); | |
2844 | VEC_free (tree, heap, tmp_vec_dsts); | |
2845 | VEC_free (tree, heap, interm_types); | |
2846 | VEC_free (tree, heap, vec_oprnds0); | |
2847 | VEC_free (tree, heap, vec_oprnds1); | |
2848 | ||
2849 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2850 | return true; | |
2851 | } | |
2852 | ||
2853 | ||
2854 | /* Function vectorizable_store. | |
2855 | ||
2856 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that | |
2857 | can be vectorized. | |
2858 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2859 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2860 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2861 | ||
2862 | static bool | |
2863 | vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
2864 | slp_tree slp_node) | |
2865 | { | |
2866 | tree scalar_dest; | |
2867 | tree data_ref; | |
2868 | tree op; | |
2869 | tree vec_oprnd = NULL_TREE; | |
2870 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2871 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
2872 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2873 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 2874 | struct loop *loop = NULL; |
ebfd146a IR |
2875 | enum machine_mode vec_mode; |
2876 | tree dummy; | |
2877 | enum dr_alignment_support alignment_support_scheme; | |
2878 | tree def; | |
2879 | gimple def_stmt; | |
2880 | enum vect_def_type dt; | |
2881 | stmt_vec_info prev_stmt_info = NULL; | |
2882 | tree dataref_ptr = NULL_TREE; | |
2883 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2884 | int ncopies; | |
2885 | int j; | |
2886 | gimple next_stmt, first_stmt = NULL; | |
2887 | bool strided_store = false; | |
2888 | unsigned int group_size, i; | |
2889 | VEC(tree,heap) *dr_chain = NULL, *oprnds = NULL, *result_chain = NULL; | |
2890 | bool inv_p; | |
2891 | VEC(tree,heap) *vec_oprnds = NULL; | |
2892 | bool slp = (slp_node != NULL); | |
2893 | stmt_vec_info first_stmt_vinfo; | |
2894 | unsigned int vec_num; | |
a70d6342 IR |
2895 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2896 | ||
2897 | if (loop_vinfo) | |
2898 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
2899 | |
2900 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2901 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2902 | case of SLP. */ | |
2903 | if (slp) | |
2904 | ncopies = 1; | |
2905 | else | |
2906 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2907 | ||
2908 | gcc_assert (ncopies >= 1); | |
2909 | ||
2910 | /* FORNOW. This restriction should be relaxed. */ | |
a70d6342 | 2911 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) |
ebfd146a IR |
2912 | { |
2913 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2914 | fprintf (vect_dump, "multiple types in nested loop."); | |
2915 | return false; | |
2916 | } | |
2917 | ||
a70d6342 | 2918 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2919 | return false; |
2920 | ||
8644a673 | 2921 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2922 | return false; |
2923 | ||
2924 | /* Is vectorizable store? */ | |
2925 | ||
2926 | if (!is_gimple_assign (stmt)) | |
2927 | return false; | |
2928 | ||
2929 | scalar_dest = gimple_assign_lhs (stmt); | |
2930 | if (TREE_CODE (scalar_dest) != ARRAY_REF | |
2931 | && TREE_CODE (scalar_dest) != INDIRECT_REF | |
e9dbe7bb IR |
2932 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
2933 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
2934 | && TREE_CODE (scalar_dest) != REALPART_EXPR) | |
ebfd146a IR |
2935 | return false; |
2936 | ||
2937 | gcc_assert (gimple_assign_single_p (stmt)); | |
2938 | op = gimple_assign_rhs1 (stmt); | |
a70d6342 | 2939 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt)) |
ebfd146a IR |
2940 | { |
2941 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2942 | fprintf (vect_dump, "use not simple."); | |
2943 | return false; | |
2944 | } | |
2945 | ||
2946 | /* The scalar rhs type needs to be trivially convertible to the vector | |
2947 | component type. This should always be the case. */ | |
2948 | if (!useless_type_conversion_p (TREE_TYPE (vectype), TREE_TYPE (op))) | |
2949 | { | |
2950 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2951 | fprintf (vect_dump, "??? operands of different types"); | |
2952 | return false; | |
2953 | } | |
2954 | ||
2955 | vec_mode = TYPE_MODE (vectype); | |
2956 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
2957 | (e.g. - array initialization with 0). */ | |
2958 | if (optab_handler (mov_optab, (int)vec_mode)->insn_code == CODE_FOR_nothing) | |
2959 | return false; | |
2960 | ||
2961 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
2962 | return false; | |
2963 | ||
2964 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) | |
2965 | { | |
2966 | strided_store = true; | |
2967 | first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
2968 | if (!vect_strided_store_supported (vectype) | |
2969 | && !PURE_SLP_STMT (stmt_info) && !slp) | |
2970 | return false; | |
2971 | ||
2972 | if (first_stmt == stmt) | |
2973 | { | |
2974 | /* STMT is the leader of the group. Check the operands of all the | |
2975 | stmts of the group. */ | |
2976 | next_stmt = DR_GROUP_NEXT_DR (stmt_info); | |
2977 | while (next_stmt) | |
2978 | { | |
2979 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
2980 | op = gimple_assign_rhs1 (next_stmt); | |
a70d6342 IR |
2981 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, |
2982 | &def, &dt)) | |
ebfd146a IR |
2983 | { |
2984 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2985 | fprintf (vect_dump, "use not simple."); | |
2986 | return false; | |
2987 | } | |
2988 | next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); | |
2989 | } | |
2990 | } | |
2991 | } | |
2992 | ||
2993 | if (!vec_stmt) /* transformation not required. */ | |
2994 | { | |
2995 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
2996 | vect_model_store_cost (stmt_info, ncopies, dt, NULL); | |
2997 | return true; | |
2998 | } | |
2999 | ||
3000 | /** Transform. **/ | |
3001 | ||
3002 | if (strided_store) | |
3003 | { | |
3004 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
3005 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
3006 | ||
3007 | DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; | |
3008 | ||
3009 | /* FORNOW */ | |
a70d6342 | 3010 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
3011 | |
3012 | /* We vectorize all the stmts of the interleaving group when we | |
3013 | reach the last stmt in the group. */ | |
3014 | if (DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) | |
3015 | < DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
3016 | && !slp) | |
3017 | { | |
3018 | *vec_stmt = NULL; | |
3019 | return true; | |
3020 | } | |
3021 | ||
3022 | if (slp) | |
3023 | strided_store = false; | |
3024 | ||
3025 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
3026 | if (slp) | |
3027 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
3028 | else | |
3029 | vec_num = group_size; | |
3030 | } | |
3031 | else | |
3032 | { | |
3033 | first_stmt = stmt; | |
3034 | first_dr = dr; | |
3035 | group_size = vec_num = 1; | |
3036 | first_stmt_vinfo = stmt_info; | |
3037 | } | |
3038 | ||
3039 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3040 | fprintf (vect_dump, "transform store. ncopies = %d",ncopies); | |
3041 | ||
3042 | dr_chain = VEC_alloc (tree, heap, group_size); | |
3043 | oprnds = VEC_alloc (tree, heap, group_size); | |
3044 | ||
3045 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr); | |
3046 | gcc_assert (alignment_support_scheme); | |
ebfd146a IR |
3047 | |
3048 | /* In case the vectorization factor (VF) is bigger than the number | |
3049 | of elements that we can fit in a vectype (nunits), we have to generate | |
3050 | more than one vector stmt - i.e - we need to "unroll" the | |
3051 | vector stmt by a factor VF/nunits. For more details see documentation in | |
3052 | vect_get_vec_def_for_copy_stmt. */ | |
3053 | ||
3054 | /* In case of interleaving (non-unit strided access): | |
3055 | ||
3056 | S1: &base + 2 = x2 | |
3057 | S2: &base = x0 | |
3058 | S3: &base + 1 = x1 | |
3059 | S4: &base + 3 = x3 | |
3060 | ||
3061 | We create vectorized stores starting from base address (the access of the | |
3062 | first stmt in the chain (S2 in the above example), when the last store stmt | |
3063 | of the chain (S4) is reached: | |
3064 | ||
3065 | VS1: &base = vx2 | |
3066 | VS2: &base + vec_size*1 = vx0 | |
3067 | VS3: &base + vec_size*2 = vx1 | |
3068 | VS4: &base + vec_size*3 = vx3 | |
3069 | ||
3070 | Then permutation statements are generated: | |
3071 | ||
3072 | VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 > | |
3073 | VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 > | |
3074 | ... | |
3075 | ||
3076 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
3077 | (the order of the data-refs in the output of vect_permute_store_chain | |
3078 | corresponds to the order of scalar stmts in the interleaving chain - see | |
3079 | the documentation of vect_permute_store_chain()). | |
3080 | ||
3081 | In case of both multiple types and interleaving, above vector stores and | |
3082 | permutation stmts are created for every copy. The result vector stmts are | |
3083 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding | |
3084 | STMT_VINFO_RELATED_STMT for the next copies. | |
3085 | */ | |
3086 | ||
3087 | prev_stmt_info = NULL; | |
3088 | for (j = 0; j < ncopies; j++) | |
3089 | { | |
3090 | gimple new_stmt; | |
3091 | gimple ptr_incr; | |
3092 | ||
3093 | if (j == 0) | |
3094 | { | |
3095 | if (slp) | |
3096 | { | |
3097 | /* Get vectorized arguments for SLP_NODE. */ | |
3098 | vect_get_slp_defs (slp_node, &vec_oprnds, NULL); | |
3099 | ||
3100 | vec_oprnd = VEC_index (tree, vec_oprnds, 0); | |
3101 | } | |
3102 | else | |
3103 | { | |
3104 | /* For interleaved stores we collect vectorized defs for all the | |
3105 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
3106 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
3107 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. | |
3108 | ||
3109 | If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and | |
3110 | OPRNDS are of size 1. */ | |
3111 | next_stmt = first_stmt; | |
3112 | for (i = 0; i < group_size; i++) | |
3113 | { | |
3114 | /* Since gaps are not supported for interleaved stores, | |
3115 | GROUP_SIZE is the exact number of stmts in the chain. | |
3116 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
3117 | there is no interleaving, GROUP_SIZE is 1, and only one | |
3118 | iteration of the loop will be executed. */ | |
3119 | gcc_assert (next_stmt | |
3120 | && gimple_assign_single_p (next_stmt)); | |
3121 | op = gimple_assign_rhs1 (next_stmt); | |
3122 | ||
3123 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt, | |
3124 | NULL); | |
3125 | VEC_quick_push(tree, dr_chain, vec_oprnd); | |
3126 | VEC_quick_push(tree, oprnds, vec_oprnd); | |
3127 | next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); | |
3128 | } | |
3129 | } | |
3130 | ||
3131 | /* We should have catched mismatched types earlier. */ | |
3132 | gcc_assert (useless_type_conversion_p (vectype, | |
3133 | TREE_TYPE (vec_oprnd))); | |
3134 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, NULL, NULL_TREE, | |
3135 | &dummy, &ptr_incr, false, | |
5006671f | 3136 | &inv_p); |
a70d6342 | 3137 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a IR |
3138 | } |
3139 | else | |
3140 | { | |
3141 | /* For interleaved stores we created vectorized defs for all the | |
3142 | defs stored in OPRNDS in the previous iteration (previous copy). | |
3143 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
3144 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the | |
3145 | next copy. | |
3146 | If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and | |
3147 | OPRNDS are of size 1. */ | |
3148 | for (i = 0; i < group_size; i++) | |
3149 | { | |
3150 | op = VEC_index (tree, oprnds, i); | |
a70d6342 IR |
3151 | vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, |
3152 | &dt); | |
ebfd146a IR |
3153 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
3154 | VEC_replace(tree, dr_chain, i, vec_oprnd); | |
3155 | VEC_replace(tree, oprnds, i, vec_oprnd); | |
3156 | } | |
3157 | dataref_ptr = | |
3158 | bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE); | |
3159 | } | |
3160 | ||
3161 | if (strided_store) | |
3162 | { | |
3163 | result_chain = VEC_alloc (tree, heap, group_size); | |
3164 | /* Permute. */ | |
3165 | if (!vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
3166 | &result_chain)) | |
3167 | return false; | |
3168 | } | |
3169 | ||
3170 | next_stmt = first_stmt; | |
3171 | for (i = 0; i < vec_num; i++) | |
3172 | { | |
3173 | if (i > 0) | |
3174 | /* Bump the vector pointer. */ | |
3175 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
3176 | NULL_TREE); | |
3177 | ||
3178 | if (slp) | |
3179 | vec_oprnd = VEC_index (tree, vec_oprnds, i); | |
3180 | else if (strided_store) | |
3181 | /* For strided stores vectorized defs are interleaved in | |
3182 | vect_permute_store_chain(). */ | |
3183 | vec_oprnd = VEC_index (tree, result_chain, i); | |
3184 | ||
8f439681 RE |
3185 | if (aligned_access_p (first_dr)) |
3186 | data_ref = build_fold_indirect_ref (dataref_ptr); | |
3187 | else | |
3188 | { | |
3189 | int mis = DR_MISALIGNMENT (first_dr); | |
3190 | tree tmis = (mis == -1 ? size_zero_node : size_int (mis)); | |
3191 | tmis = size_binop (MULT_EXPR, tmis, size_int (BITS_PER_UNIT)); | |
3192 | data_ref = build2 (MISALIGNED_INDIRECT_REF, vectype, dataref_ptr, tmis); | |
3193 | } | |
3194 | ||
5006671f RG |
3195 | /* If accesses through a pointer to vectype do not alias the original |
3196 | memory reference we have a problem. This should never happen. */ | |
3197 | gcc_assert (alias_sets_conflict_p (get_alias_set (data_ref), | |
3198 | get_alias_set (gimple_assign_lhs (stmt)))); | |
ebfd146a IR |
3199 | |
3200 | /* Arguments are ready. Create the new vector stmt. */ | |
3201 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
3202 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3203 | mark_symbols_for_renaming (new_stmt); | |
3204 | ||
3205 | if (slp) | |
3206 | continue; | |
3207 | ||
3208 | if (j == 0) | |
3209 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3210 | else | |
3211 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3212 | ||
3213 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3214 | next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); | |
3215 | if (!next_stmt) | |
3216 | break; | |
3217 | } | |
3218 | } | |
3219 | ||
3220 | VEC_free (tree, heap, dr_chain); | |
3221 | VEC_free (tree, heap, oprnds); | |
3222 | if (result_chain) | |
3223 | VEC_free (tree, heap, result_chain); | |
3224 | ||
3225 | return true; | |
3226 | } | |
3227 | ||
3228 | /* vectorizable_load. | |
3229 | ||
3230 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that | |
3231 | can be vectorized. | |
3232 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3233 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3234 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3235 | ||
3236 | static bool | |
3237 | vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
3238 | slp_tree slp_node, slp_instance slp_node_instance) | |
3239 | { | |
3240 | tree scalar_dest; | |
3241 | tree vec_dest = NULL; | |
3242 | tree data_ref = NULL; | |
3243 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3244 | stmt_vec_info prev_stmt_info; | |
3245 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 3246 | struct loop *loop = NULL; |
ebfd146a | 3247 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 3248 | bool nested_in_vect_loop = false; |
ebfd146a IR |
3249 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; |
3250 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3251 | tree new_temp; | |
3252 | int mode; | |
3253 | gimple new_stmt = NULL; | |
3254 | tree dummy; | |
3255 | enum dr_alignment_support alignment_support_scheme; | |
3256 | tree dataref_ptr = NULL_TREE; | |
3257 | gimple ptr_incr; | |
3258 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
3259 | int ncopies; | |
3260 | int i, j, group_size; | |
3261 | tree msq = NULL_TREE, lsq; | |
3262 | tree offset = NULL_TREE; | |
3263 | tree realignment_token = NULL_TREE; | |
3264 | gimple phi = NULL; | |
3265 | VEC(tree,heap) *dr_chain = NULL; | |
3266 | bool strided_load = false; | |
3267 | gimple first_stmt; | |
3268 | tree scalar_type; | |
3269 | bool inv_p; | |
3270 | bool compute_in_loop = false; | |
3271 | struct loop *at_loop; | |
3272 | int vec_num; | |
3273 | bool slp = (slp_node != NULL); | |
3274 | bool slp_perm = false; | |
3275 | enum tree_code code; | |
a70d6342 IR |
3276 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3277 | int vf; | |
3278 | ||
3279 | if (loop_vinfo) | |
3280 | { | |
3281 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
3282 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
3283 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3284 | } | |
3285 | else | |
3286 | /* FORNOW: multiple types are not supported in basic block SLP. */ | |
3287 | vf = nunits; | |
ebfd146a IR |
3288 | |
3289 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3290 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3291 | case of SLP. */ | |
3292 | if (slp) | |
3293 | ncopies = 1; | |
3294 | else | |
3295 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3296 | ||
3297 | gcc_assert (ncopies >= 1); | |
3298 | ||
3299 | /* FORNOW. This restriction should be relaxed. */ | |
3300 | if (nested_in_vect_loop && ncopies > 1) | |
3301 | { | |
3302 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3303 | fprintf (vect_dump, "multiple types in nested loop."); | |
3304 | return false; | |
3305 | } | |
3306 | ||
a70d6342 | 3307 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3308 | return false; |
3309 | ||
8644a673 | 3310 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3311 | return false; |
3312 | ||
3313 | /* Is vectorizable load? */ | |
3314 | if (!is_gimple_assign (stmt)) | |
3315 | return false; | |
3316 | ||
3317 | scalar_dest = gimple_assign_lhs (stmt); | |
3318 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
3319 | return false; | |
3320 | ||
3321 | code = gimple_assign_rhs_code (stmt); | |
3322 | if (code != ARRAY_REF | |
3323 | && code != INDIRECT_REF | |
e9dbe7bb IR |
3324 | && code != COMPONENT_REF |
3325 | && code != IMAGPART_EXPR | |
3326 | && code != REALPART_EXPR) | |
ebfd146a IR |
3327 | return false; |
3328 | ||
3329 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
3330 | return false; | |
3331 | ||
3332 | scalar_type = TREE_TYPE (DR_REF (dr)); | |
3333 | mode = (int) TYPE_MODE (vectype); | |
3334 | ||
3335 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
3336 | (e.g. - data copies). */ | |
3337 | if (optab_handler (mov_optab, mode)->insn_code == CODE_FOR_nothing) | |
3338 | { | |
3339 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3340 | fprintf (vect_dump, "Aligned load, but unsupported type."); | |
3341 | return false; | |
3342 | } | |
3343 | ||
3344 | /* The vector component type needs to be trivially convertible to the | |
3345 | scalar lhs. This should always be the case. */ | |
3346 | if (!useless_type_conversion_p (TREE_TYPE (scalar_dest), TREE_TYPE (vectype))) | |
3347 | { | |
3348 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3349 | fprintf (vect_dump, "??? operands of different types"); | |
3350 | return false; | |
3351 | } | |
3352 | ||
3353 | /* Check if the load is a part of an interleaving chain. */ | |
3354 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) | |
3355 | { | |
3356 | strided_load = true; | |
3357 | /* FORNOW */ | |
3358 | gcc_assert (! nested_in_vect_loop); | |
3359 | ||
3360 | /* Check if interleaving is supported. */ | |
3361 | if (!vect_strided_load_supported (vectype) | |
3362 | && !PURE_SLP_STMT (stmt_info) && !slp) | |
3363 | return false; | |
3364 | } | |
3365 | ||
3366 | if (!vec_stmt) /* transformation not required. */ | |
3367 | { | |
3368 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
3369 | vect_model_load_cost (stmt_info, ncopies, NULL); | |
3370 | return true; | |
3371 | } | |
3372 | ||
3373 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3374 | fprintf (vect_dump, "transform load."); | |
3375 | ||
3376 | /** Transform. **/ | |
3377 | ||
3378 | if (strided_load) | |
3379 | { | |
3380 | first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
3381 | /* Check if the chain of loads is already vectorized. */ | |
3382 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt))) | |
3383 | { | |
3384 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
3385 | return true; | |
3386 | } | |
3387 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
3388 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
3389 | ||
3390 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
3391 | if (slp) | |
3392 | { | |
3393 | strided_load = false; | |
3394 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
a70d6342 IR |
3395 | if (SLP_INSTANCE_LOAD_PERMUTATION (slp_node_instance)) |
3396 | slp_perm = true; | |
3397 | } | |
ebfd146a IR |
3398 | else |
3399 | vec_num = group_size; | |
3400 | ||
3401 | dr_chain = VEC_alloc (tree, heap, vec_num); | |
3402 | } | |
3403 | else | |
3404 | { | |
3405 | first_stmt = stmt; | |
3406 | first_dr = dr; | |
3407 | group_size = vec_num = 1; | |
3408 | } | |
3409 | ||
3410 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr); | |
3411 | gcc_assert (alignment_support_scheme); | |
3412 | ||
3413 | /* In case the vectorization factor (VF) is bigger than the number | |
3414 | of elements that we can fit in a vectype (nunits), we have to generate | |
3415 | more than one vector stmt - i.e - we need to "unroll" the | |
3416 | vector stmt by a factor VF/nunits. In doing so, we record a pointer | |
3417 | from one copy of the vector stmt to the next, in the field | |
3418 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
3419 | stages to find the correct vector defs to be used when vectorizing | |
3420 | stmts that use the defs of the current stmt. The example below illustrates | |
3421 | the vectorization process when VF=16 and nunits=4 (i.e - we need to create | |
3422 | 4 vectorized stmts): | |
3423 | ||
3424 | before vectorization: | |
3425 | RELATED_STMT VEC_STMT | |
3426 | S1: x = memref - - | |
3427 | S2: z = x + 1 - - | |
3428 | ||
3429 | step 1: vectorize stmt S1: | |
3430 | We first create the vector stmt VS1_0, and, as usual, record a | |
3431 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
3432 | Next, we create the vector stmt VS1_1, and record a pointer to | |
3433 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
3434 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of | |
3435 | stmts and pointers: | |
3436 | RELATED_STMT VEC_STMT | |
3437 | VS1_0: vx0 = memref0 VS1_1 - | |
3438 | VS1_1: vx1 = memref1 VS1_2 - | |
3439 | VS1_2: vx2 = memref2 VS1_3 - | |
3440 | VS1_3: vx3 = memref3 - - | |
3441 | S1: x = load - VS1_0 | |
3442 | S2: z = x + 1 - - | |
3443 | ||
3444 | See in documentation in vect_get_vec_def_for_stmt_copy for how the | |
3445 | information we recorded in RELATED_STMT field is used to vectorize | |
3446 | stmt S2. */ | |
3447 | ||
3448 | /* In case of interleaving (non-unit strided access): | |
3449 | ||
3450 | S1: x2 = &base + 2 | |
3451 | S2: x0 = &base | |
3452 | S3: x1 = &base + 1 | |
3453 | S4: x3 = &base + 3 | |
3454 | ||
3455 | Vectorized loads are created in the order of memory accesses | |
3456 | starting from the access of the first stmt of the chain: | |
3457 | ||
3458 | VS1: vx0 = &base | |
3459 | VS2: vx1 = &base + vec_size*1 | |
3460 | VS3: vx3 = &base + vec_size*2 | |
3461 | VS4: vx4 = &base + vec_size*3 | |
3462 | ||
3463 | Then permutation statements are generated: | |
3464 | ||
3465 | VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 > | |
3466 | VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 > | |
3467 | ... | |
3468 | ||
3469 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
3470 | (the order of the data-refs in the output of vect_permute_load_chain | |
3471 | corresponds to the order of scalar stmts in the interleaving chain - see | |
3472 | the documentation of vect_permute_load_chain()). | |
3473 | The generation of permutation stmts and recording them in | |
3474 | STMT_VINFO_VEC_STMT is done in vect_transform_strided_load(). | |
3475 | ||
3476 | In case of both multiple types and interleaving, the vector loads and | |
3477 | permutation stmts above are created for every copy. The result vector stmts | |
3478 | are put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding | |
3479 | STMT_VINFO_RELATED_STMT for the next copies. */ | |
3480 | ||
3481 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
3482 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
3483 | we generate the following code: | |
3484 | p = initial_addr; | |
3485 | indx = 0; | |
3486 | loop { | |
3487 | p = p + indx * vectype_size; | |
3488 | vec_dest = *(p); | |
3489 | indx = indx + 1; | |
3490 | } | |
3491 | ||
3492 | Otherwise, the data reference is potentially unaligned on a target that | |
3493 | does not support unaligned accesses (dr_explicit_realign_optimized) - | |
3494 | then generate the following code, in which the data in each iteration is | |
3495 | obtained by two vector loads, one from the previous iteration, and one | |
3496 | from the current iteration: | |
3497 | p1 = initial_addr; | |
3498 | msq_init = *(floor(p1)) | |
3499 | p2 = initial_addr + VS - 1; | |
3500 | realignment_token = call target_builtin; | |
3501 | indx = 0; | |
3502 | loop { | |
3503 | p2 = p2 + indx * vectype_size | |
3504 | lsq = *(floor(p2)) | |
3505 | vec_dest = realign_load (msq, lsq, realignment_token) | |
3506 | indx = indx + 1; | |
3507 | msq = lsq; | |
3508 | } */ | |
3509 | ||
3510 | /* If the misalignment remains the same throughout the execution of the | |
3511 | loop, we can create the init_addr and permutation mask at the loop | |
3512 | preheader. Otherwise, it needs to be created inside the loop. | |
3513 | This can only occur when vectorizing memory accesses in the inner-loop | |
3514 | nested within an outer-loop that is being vectorized. */ | |
3515 | ||
a70d6342 | 3516 | if (loop && nested_in_vect_loop_p (loop, stmt) |
ebfd146a IR |
3517 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
3518 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) | |
3519 | { | |
3520 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
3521 | compute_in_loop = true; | |
3522 | } | |
3523 | ||
3524 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
3525 | || alignment_support_scheme == dr_explicit_realign) | |
3526 | && !compute_in_loop) | |
3527 | { | |
3528 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
3529 | alignment_support_scheme, NULL_TREE, | |
3530 | &at_loop); | |
3531 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
3532 | { | |
3533 | phi = SSA_NAME_DEF_STMT (msq); | |
3534 | offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
3535 | } | |
3536 | } | |
3537 | else | |
3538 | at_loop = loop; | |
3539 | ||
3540 | prev_stmt_info = NULL; | |
3541 | for (j = 0; j < ncopies; j++) | |
3542 | { | |
3543 | /* 1. Create the vector pointer update chain. */ | |
3544 | if (j == 0) | |
3545 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, | |
3546 | at_loop, offset, | |
3547 | &dummy, &ptr_incr, false, | |
5006671f | 3548 | &inv_p); |
ebfd146a IR |
3549 | else |
3550 | dataref_ptr = | |
3551 | bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE); | |
3552 | ||
3553 | for (i = 0; i < vec_num; i++) | |
3554 | { | |
3555 | if (i > 0) | |
3556 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
3557 | NULL_TREE); | |
3558 | ||
3559 | /* 2. Create the vector-load in the loop. */ | |
3560 | switch (alignment_support_scheme) | |
3561 | { | |
3562 | case dr_aligned: | |
3563 | gcc_assert (aligned_access_p (first_dr)); | |
3564 | data_ref = build_fold_indirect_ref (dataref_ptr); | |
3565 | break; | |
3566 | case dr_unaligned_supported: | |
3567 | { | |
3568 | int mis = DR_MISALIGNMENT (first_dr); | |
3569 | tree tmis = (mis == -1 ? size_zero_node : size_int (mis)); | |
3570 | ||
3571 | tmis = size_binop (MULT_EXPR, tmis, size_int(BITS_PER_UNIT)); | |
3572 | data_ref = | |
3573 | build2 (MISALIGNED_INDIRECT_REF, vectype, dataref_ptr, tmis); | |
3574 | break; | |
3575 | } | |
3576 | case dr_explicit_realign: | |
3577 | { | |
3578 | tree ptr, bump; | |
3579 | tree vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
3580 | ||
3581 | if (compute_in_loop) | |
3582 | msq = vect_setup_realignment (first_stmt, gsi, | |
3583 | &realignment_token, | |
3584 | dr_explicit_realign, | |
3585 | dataref_ptr, NULL); | |
3586 | ||
3587 | data_ref = build1 (ALIGN_INDIRECT_REF, vectype, dataref_ptr); | |
3588 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3589 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
3590 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3591 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5006671f RG |
3592 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); |
3593 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
ebfd146a | 3594 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
3595 | msq = new_temp; |
3596 | ||
3597 | bump = size_binop (MULT_EXPR, vs_minus_1, | |
3598 | TYPE_SIZE_UNIT (scalar_type)); | |
3599 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); | |
3600 | data_ref = build1 (ALIGN_INDIRECT_REF, vectype, ptr); | |
3601 | break; | |
3602 | } | |
3603 | case dr_explicit_realign_optimized: | |
3604 | data_ref = build1 (ALIGN_INDIRECT_REF, vectype, dataref_ptr); | |
3605 | break; | |
3606 | default: | |
3607 | gcc_unreachable (); | |
3608 | } | |
5006671f RG |
3609 | /* If accesses through a pointer to vectype do not alias the original |
3610 | memory reference we have a problem. This should never happen. */ | |
3611 | gcc_assert (alias_sets_conflict_p (get_alias_set (data_ref), | |
3612 | get_alias_set (gimple_assign_rhs1 (stmt)))); | |
ebfd146a IR |
3613 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
3614 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
3615 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3616 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3617 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3618 | mark_symbols_for_renaming (new_stmt); | |
3619 | ||
3620 | /* 3. Handle explicit realignment if necessary/supported. Create in | |
3621 | loop: vec_dest = realign_load (msq, lsq, realignment_token) */ | |
3622 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
3623 | || alignment_support_scheme == dr_explicit_realign) | |
3624 | { | |
3625 | tree tmp; | |
3626 | ||
3627 | lsq = gimple_assign_lhs (new_stmt); | |
3628 | if (!realignment_token) | |
3629 | realignment_token = dataref_ptr; | |
3630 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3631 | tmp = build3 (REALIGN_LOAD_EXPR, vectype, msq, lsq, | |
3632 | realignment_token); | |
3633 | new_stmt = gimple_build_assign (vec_dest, tmp); | |
3634 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3635 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3636 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3637 | ||
3638 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
3639 | { | |
3640 | gcc_assert (phi); | |
3641 | if (i == vec_num - 1 && j == ncopies - 1) | |
f5045c96 AM |
3642 | add_phi_arg (phi, lsq, loop_latch_edge (containing_loop), |
3643 | UNKNOWN_LOCATION); | |
ebfd146a IR |
3644 | msq = lsq; |
3645 | } | |
3646 | } | |
3647 | ||
3648 | /* 4. Handle invariant-load. */ | |
a70d6342 | 3649 | if (inv_p && !bb_vinfo) |
ebfd146a IR |
3650 | { |
3651 | gcc_assert (!strided_load); | |
3652 | gcc_assert (nested_in_vect_loop_p (loop, stmt)); | |
3653 | if (j == 0) | |
3654 | { | |
3655 | int k; | |
3656 | tree t = NULL_TREE; | |
3657 | tree vec_inv, bitpos, bitsize = TYPE_SIZE (scalar_type); | |
3658 | ||
3659 | /* CHECKME: bitpos depends on endianess? */ | |
3660 | bitpos = bitsize_zero_node; | |
3661 | vec_inv = build3 (BIT_FIELD_REF, scalar_type, new_temp, | |
3662 | bitsize, bitpos); | |
3663 | vec_dest = | |
3664 | vect_create_destination_var (scalar_dest, NULL_TREE); | |
3665 | new_stmt = gimple_build_assign (vec_dest, vec_inv); | |
3666 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3667 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3668 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3669 | ||
3670 | for (k = nunits - 1; k >= 0; --k) | |
3671 | t = tree_cons (NULL_TREE, new_temp, t); | |
3672 | /* FIXME: use build_constructor directly. */ | |
3673 | vec_inv = build_constructor_from_list (vectype, t); | |
3674 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
3675 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
3676 | } | |
3677 | else | |
3678 | gcc_unreachable (); /* FORNOW. */ | |
3679 | } | |
3680 | ||
3681 | /* Collect vector loads and later create their permutation in | |
3682 | vect_transform_strided_load (). */ | |
3683 | if (strided_load || slp_perm) | |
3684 | VEC_quick_push (tree, dr_chain, new_temp); | |
3685 | ||
3686 | /* Store vector loads in the corresponding SLP_NODE. */ | |
3687 | if (slp && !slp_perm) | |
3688 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
3689 | } | |
3690 | ||
3691 | if (slp && !slp_perm) | |
3692 | continue; | |
3693 | ||
3694 | if (slp_perm) | |
3695 | { | |
a70d6342 | 3696 | if (!vect_transform_slp_perm_load (stmt, dr_chain, gsi, vf, |
ebfd146a IR |
3697 | slp_node_instance, false)) |
3698 | { | |
3699 | VEC_free (tree, heap, dr_chain); | |
3700 | return false; | |
3701 | } | |
3702 | } | |
3703 | else | |
3704 | { | |
3705 | if (strided_load) | |
3706 | { | |
3707 | if (!vect_transform_strided_load (stmt, dr_chain, group_size, gsi)) | |
3708 | return false; | |
3709 | ||
3710 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
3711 | VEC_free (tree, heap, dr_chain); | |
3712 | dr_chain = VEC_alloc (tree, heap, group_size); | |
3713 | } | |
3714 | else | |
3715 | { | |
3716 | if (j == 0) | |
3717 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3718 | else | |
3719 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3720 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3721 | } | |
3722 | } | |
3723 | } | |
3724 | ||
3725 | if (dr_chain) | |
3726 | VEC_free (tree, heap, dr_chain); | |
3727 | ||
3728 | return true; | |
3729 | } | |
3730 | ||
3731 | /* Function vect_is_simple_cond. | |
3732 | ||
3733 | Input: | |
3734 | LOOP - the loop that is being vectorized. | |
3735 | COND - Condition that is checked for simple use. | |
3736 | ||
3737 | Returns whether a COND can be vectorized. Checks whether | |
3738 | condition operands are supportable using vec_is_simple_use. */ | |
3739 | ||
3740 | static bool | |
3741 | vect_is_simple_cond (tree cond, loop_vec_info loop_vinfo) | |
3742 | { | |
3743 | tree lhs, rhs; | |
3744 | tree def; | |
3745 | enum vect_def_type dt; | |
3746 | ||
3747 | if (!COMPARISON_CLASS_P (cond)) | |
3748 | return false; | |
3749 | ||
3750 | lhs = TREE_OPERAND (cond, 0); | |
3751 | rhs = TREE_OPERAND (cond, 1); | |
3752 | ||
3753 | if (TREE_CODE (lhs) == SSA_NAME) | |
3754 | { | |
3755 | gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); | |
a70d6342 IR |
3756 | if (!vect_is_simple_use (lhs, loop_vinfo, NULL, &lhs_def_stmt, &def, |
3757 | &dt)) | |
ebfd146a IR |
3758 | return false; |
3759 | } | |
3760 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
3761 | && TREE_CODE (lhs) != FIXED_CST) | |
3762 | return false; | |
3763 | ||
3764 | if (TREE_CODE (rhs) == SSA_NAME) | |
3765 | { | |
3766 | gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); | |
a70d6342 IR |
3767 | if (!vect_is_simple_use (rhs, loop_vinfo, NULL, &rhs_def_stmt, &def, |
3768 | &dt)) | |
ebfd146a IR |
3769 | return false; |
3770 | } | |
3771 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST | |
3772 | && TREE_CODE (rhs) != FIXED_CST) | |
3773 | return false; | |
3774 | ||
3775 | return true; | |
3776 | } | |
3777 | ||
3778 | /* vectorizable_condition. | |
3779 | ||
3780 | Check if STMT is conditional modify expression that can be vectorized. | |
3781 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3782 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
3783 | at GSI. |
3784 | ||
3785 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
3786 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
3787 | else caluse if it is 2). | |
ebfd146a IR |
3788 | |
3789 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3790 | ||
4bbe8262 | 3791 | bool |
ebfd146a | 3792 | vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi, |
4bbe8262 | 3793 | gimple *vec_stmt, tree reduc_def, int reduc_index) |
ebfd146a IR |
3794 | { |
3795 | tree scalar_dest = NULL_TREE; | |
3796 | tree vec_dest = NULL_TREE; | |
3797 | tree op = NULL_TREE; | |
3798 | tree cond_expr, then_clause, else_clause; | |
3799 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3800 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3801 | tree vec_cond_lhs, vec_cond_rhs, vec_then_clause, vec_else_clause; | |
3802 | tree vec_compare, vec_cond_expr; | |
3803 | tree new_temp; | |
3804 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3805 | enum machine_mode vec_mode; | |
3806 | tree def; | |
3807 | enum vect_def_type dt; | |
3808 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
3809 | int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3810 | enum tree_code code; | |
3811 | ||
a70d6342 IR |
3812 | /* FORNOW: unsupported in basic block SLP. */ |
3813 | gcc_assert (loop_vinfo); | |
3814 | ||
ebfd146a IR |
3815 | gcc_assert (ncopies >= 1); |
3816 | if (ncopies > 1) | |
3817 | return false; /* FORNOW */ | |
3818 | ||
3819 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
3820 | return false; | |
3821 | ||
4bbe8262 IR |
3822 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3823 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
3824 | && reduc_def)) | |
ebfd146a IR |
3825 | return false; |
3826 | ||
3827 | /* FORNOW: SLP not supported. */ | |
3828 | if (STMT_SLP_TYPE (stmt_info)) | |
3829 | return false; | |
3830 | ||
3831 | /* FORNOW: not yet supported. */ | |
4bbe8262 | 3832 | if (STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a IR |
3833 | { |
3834 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3835 | fprintf (vect_dump, "value used after loop."); | |
3836 | return false; | |
3837 | } | |
3838 | ||
3839 | /* Is vectorizable conditional operation? */ | |
3840 | if (!is_gimple_assign (stmt)) | |
3841 | return false; | |
3842 | ||
3843 | code = gimple_assign_rhs_code (stmt); | |
3844 | ||
3845 | if (code != COND_EXPR) | |
3846 | return false; | |
3847 | ||
3848 | gcc_assert (gimple_assign_single_p (stmt)); | |
3849 | op = gimple_assign_rhs1 (stmt); | |
3850 | cond_expr = TREE_OPERAND (op, 0); | |
3851 | then_clause = TREE_OPERAND (op, 1); | |
3852 | else_clause = TREE_OPERAND (op, 2); | |
3853 | ||
3854 | if (!vect_is_simple_cond (cond_expr, loop_vinfo)) | |
3855 | return false; | |
3856 | ||
3857 | /* We do not handle two different vector types for the condition | |
3858 | and the values. */ | |
3859 | if (TREE_TYPE (TREE_OPERAND (cond_expr, 0)) != TREE_TYPE (vectype)) | |
3860 | return false; | |
3861 | ||
3862 | if (TREE_CODE (then_clause) == SSA_NAME) | |
3863 | { | |
3864 | gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause); | |
a70d6342 | 3865 | if (!vect_is_simple_use (then_clause, loop_vinfo, NULL, |
ebfd146a IR |
3866 | &then_def_stmt, &def, &dt)) |
3867 | return false; | |
3868 | } | |
3869 | else if (TREE_CODE (then_clause) != INTEGER_CST | |
3870 | && TREE_CODE (then_clause) != REAL_CST | |
3871 | && TREE_CODE (then_clause) != FIXED_CST) | |
3872 | return false; | |
3873 | ||
3874 | if (TREE_CODE (else_clause) == SSA_NAME) | |
3875 | { | |
3876 | gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause); | |
a70d6342 | 3877 | if (!vect_is_simple_use (else_clause, loop_vinfo, NULL, |
ebfd146a IR |
3878 | &else_def_stmt, &def, &dt)) |
3879 | return false; | |
3880 | } | |
3881 | else if (TREE_CODE (else_clause) != INTEGER_CST | |
3882 | && TREE_CODE (else_clause) != REAL_CST | |
3883 | && TREE_CODE (else_clause) != FIXED_CST) | |
3884 | return false; | |
3885 | ||
3886 | ||
3887 | vec_mode = TYPE_MODE (vectype); | |
3888 | ||
3889 | if (!vec_stmt) | |
3890 | { | |
3891 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
8e7aa1f9 | 3892 | return expand_vec_cond_expr_p (TREE_TYPE (op), vec_mode); |
ebfd146a IR |
3893 | } |
3894 | ||
3895 | /* Transform */ | |
3896 | ||
3897 | /* Handle def. */ | |
3898 | scalar_dest = gimple_assign_lhs (stmt); | |
3899 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3900 | ||
3901 | /* Handle cond expr. */ | |
3902 | vec_cond_lhs = | |
3903 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), stmt, NULL); | |
3904 | vec_cond_rhs = | |
3905 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), stmt, NULL); | |
4bbe8262 IR |
3906 | if (reduc_index == 1) |
3907 | vec_then_clause = reduc_def; | |
3908 | else | |
3909 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, stmt, NULL); | |
3910 | if (reduc_index == 2) | |
3911 | vec_else_clause = reduc_def; | |
3912 | else | |
3913 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, stmt, NULL); | |
ebfd146a IR |
3914 | |
3915 | /* Arguments are ready. Create the new vector stmt. */ | |
3916 | vec_compare = build2 (TREE_CODE (cond_expr), vectype, | |
3917 | vec_cond_lhs, vec_cond_rhs); | |
3918 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, | |
3919 | vec_compare, vec_then_clause, vec_else_clause); | |
3920 | ||
3921 | *vec_stmt = gimple_build_assign (vec_dest, vec_cond_expr); | |
3922 | new_temp = make_ssa_name (vec_dest, *vec_stmt); | |
3923 | gimple_assign_set_lhs (*vec_stmt, new_temp); | |
3924 | vect_finish_stmt_generation (stmt, *vec_stmt, gsi); | |
3925 | ||
3926 | return true; | |
3927 | } | |
3928 | ||
3929 | ||
8644a673 | 3930 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
3931 | |
3932 | bool | |
a70d6342 | 3933 | vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 3934 | { |
8644a673 | 3935 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 IR |
3936 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3937 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); | |
ebfd146a | 3938 | bool ok; |
a70d6342 IR |
3939 | HOST_WIDE_INT dummy; |
3940 | tree scalar_type, vectype; | |
ebfd146a IR |
3941 | |
3942 | if (vect_print_dump_info (REPORT_DETAILS)) | |
ebfd146a | 3943 | { |
8644a673 IR |
3944 | fprintf (vect_dump, "==> examining statement: "); |
3945 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
3946 | } | |
ebfd146a | 3947 | |
1825a1f3 IR |
3948 | if (gimple_has_volatile_ops (stmt)) |
3949 | { | |
3950 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
3951 | fprintf (vect_dump, "not vectorized: stmt has volatile operands"); | |
3952 | ||
3953 | return false; | |
3954 | } | |
3955 | ||
a70d6342 | 3956 | /* Skip stmts that do not need to be vectorized. In loops this is expected |
8644a673 IR |
3957 | to include: |
3958 | - the COND_EXPR which is the loop exit condition | |
3959 | - any LABEL_EXPRs in the loop | |
a70d6342 | 3960 | - computations that are used only for array indexing or loop control. |
8644a673 IR |
3961 | In basic blocks we only analyze statements that are a part of some SLP |
3962 | instance, therefore, all the statements are relevant. */ | |
ebfd146a | 3963 | |
a70d6342 | 3964 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 3965 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a IR |
3966 | { |
3967 | if (vect_print_dump_info (REPORT_DETAILS)) | |
8644a673 | 3968 | fprintf (vect_dump, "irrelevant."); |
ebfd146a | 3969 | |
8644a673 IR |
3970 | return true; |
3971 | } | |
ebfd146a | 3972 | |
8644a673 IR |
3973 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
3974 | { | |
3975 | case vect_internal_def: | |
3976 | break; | |
ebfd146a | 3977 | |
8644a673 | 3978 | case vect_reduction_def: |
7c5222ff | 3979 | case vect_nested_cycle: |
a70d6342 | 3980 | gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer |
8644a673 | 3981 | || relevance == vect_used_in_outer_by_reduction |
a70d6342 | 3982 | || relevance == vect_unused_in_scope)); |
8644a673 IR |
3983 | break; |
3984 | ||
3985 | case vect_induction_def: | |
3986 | case vect_constant_def: | |
3987 | case vect_external_def: | |
3988 | case vect_unknown_def_type: | |
3989 | default: | |
3990 | gcc_unreachable (); | |
3991 | } | |
ebfd146a | 3992 | |
a70d6342 IR |
3993 | if (bb_vinfo) |
3994 | { | |
3995 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
3996 | ||
3997 | scalar_type = vect_get_smallest_scalar_type (stmt, &dummy, &dummy); | |
3998 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3999 | { | |
4000 | fprintf (vect_dump, "get vectype for scalar type: "); | |
4001 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
4002 | } | |
4003 | ||
4004 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4005 | if (!vectype) | |
4006 | { | |
4007 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4008 | { | |
4009 | fprintf (vect_dump, "not SLPed: unsupported data-type "); | |
4010 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
4011 | } | |
4012 | return false; | |
4013 | } | |
4014 | ||
4015 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4016 | { | |
4017 | fprintf (vect_dump, "vectype: "); | |
4018 | print_generic_expr (vect_dump, vectype, TDF_SLIM); | |
4019 | } | |
4020 | ||
4021 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
4022 | } | |
4023 | ||
8644a673 | 4024 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 4025 | { |
8644a673 IR |
4026 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
4027 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info)); | |
4028 | *need_to_vectorize = true; | |
ebfd146a IR |
4029 | } |
4030 | ||
8644a673 | 4031 | ok = true; |
a70d6342 IR |
4032 | if (!bb_vinfo |
4033 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
4034 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
8644a673 IR |
4035 | ok = (vectorizable_type_promotion (stmt, NULL, NULL, NULL) |
4036 | || vectorizable_type_demotion (stmt, NULL, NULL, NULL) | |
4037 | || vectorizable_conversion (stmt, NULL, NULL, NULL) | |
4038 | || vectorizable_operation (stmt, NULL, NULL, NULL) | |
4039 | || vectorizable_assignment (stmt, NULL, NULL, NULL) | |
4040 | || vectorizable_load (stmt, NULL, NULL, NULL, NULL) | |
4041 | || vectorizable_call (stmt, NULL, NULL) | |
4042 | || vectorizable_store (stmt, NULL, NULL, NULL) | |
4bbe8262 IR |
4043 | || vectorizable_reduction (stmt, NULL, NULL) |
4044 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0)); | |
a70d6342 IR |
4045 | else |
4046 | { | |
4047 | if (bb_vinfo) | |
4048 | ok = (vectorizable_operation (stmt, NULL, NULL, node) | |
4049 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
4050 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
4051 | || vectorizable_store (stmt, NULL, NULL, node)); | |
4052 | } | |
8644a673 IR |
4053 | |
4054 | if (!ok) | |
ebfd146a | 4055 | { |
8644a673 IR |
4056 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
4057 | { | |
4058 | fprintf (vect_dump, "not vectorized: relevant stmt not "); | |
4059 | fprintf (vect_dump, "supported: "); | |
4060 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4061 | } | |
a70d6342 | 4062 | |
ebfd146a IR |
4063 | return false; |
4064 | } | |
4065 | ||
a70d6342 IR |
4066 | if (bb_vinfo) |
4067 | return true; | |
4068 | ||
8644a673 IR |
4069 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
4070 | need extra handling, except for vectorizable reductions. */ | |
4071 | if (STMT_VINFO_LIVE_P (stmt_info) | |
4072 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
4073 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 4074 | |
8644a673 | 4075 | if (!ok) |
ebfd146a | 4076 | { |
8644a673 IR |
4077 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
4078 | { | |
4079 | fprintf (vect_dump, "not vectorized: live stmt not "); | |
4080 | fprintf (vect_dump, "supported: "); | |
4081 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4082 | } | |
a70d6342 | 4083 | |
8644a673 | 4084 | return false; |
ebfd146a IR |
4085 | } |
4086 | ||
8644a673 | 4087 | if (!PURE_SLP_STMT (stmt_info)) |
ebfd146a | 4088 | { |
a70d6342 IR |
4089 | /* Groups of strided accesses whose size is not a power of 2 are not |
4090 | vectorizable yet using loop-vectorization. Therefore, if this stmt | |
4091 | feeds non-SLP-able stmts (i.e., this stmt has to be both SLPed and | |
4092 | loop-based vectorized), the loop cannot be vectorized. */ | |
8644a673 IR |
4093 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) |
4094 | && exact_log2 (DR_GROUP_SIZE (vinfo_for_stmt ( | |
4095 | DR_GROUP_FIRST_DR (stmt_info)))) == -1) | |
ebfd146a | 4096 | { |
8644a673 IR |
4097 | if (vect_print_dump_info (REPORT_DETAILS)) |
4098 | { | |
4099 | fprintf (vect_dump, "not vectorized: the size of group " | |
4100 | "of strided accesses is not a power of 2"); | |
4101 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4102 | } | |
4103 | ||
ebfd146a IR |
4104 | return false; |
4105 | } | |
4106 | } | |
a70d6342 | 4107 | |
ebfd146a IR |
4108 | return true; |
4109 | } | |
4110 | ||
4111 | ||
4112 | /* Function vect_transform_stmt. | |
4113 | ||
4114 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
4115 | ||
4116 | bool | |
4117 | vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi, | |
4118 | bool *strided_store, slp_tree slp_node, | |
4119 | slp_instance slp_node_instance) | |
4120 | { | |
4121 | bool is_store = false; | |
4122 | gimple vec_stmt = NULL; | |
4123 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4124 | gimple orig_stmt_in_pattern; | |
4125 | bool done; | |
ebfd146a IR |
4126 | |
4127 | switch (STMT_VINFO_TYPE (stmt_info)) | |
4128 | { | |
4129 | case type_demotion_vec_info_type: | |
4130 | done = vectorizable_type_demotion (stmt, gsi, &vec_stmt, slp_node); | |
4131 | gcc_assert (done); | |
4132 | break; | |
4133 | ||
4134 | case type_promotion_vec_info_type: | |
4135 | done = vectorizable_type_promotion (stmt, gsi, &vec_stmt, slp_node); | |
4136 | gcc_assert (done); | |
4137 | break; | |
4138 | ||
4139 | case type_conversion_vec_info_type: | |
4140 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
4141 | gcc_assert (done); | |
4142 | break; | |
4143 | ||
4144 | case induc_vec_info_type: | |
4145 | gcc_assert (!slp_node); | |
4146 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
4147 | gcc_assert (done); | |
4148 | break; | |
4149 | ||
4150 | case op_vec_info_type: | |
4151 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
4152 | gcc_assert (done); | |
4153 | break; | |
4154 | ||
4155 | case assignment_vec_info_type: | |
4156 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
4157 | gcc_assert (done); | |
4158 | break; | |
4159 | ||
4160 | case load_vec_info_type: | |
4161 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, | |
4162 | slp_node_instance); | |
4163 | gcc_assert (done); | |
4164 | break; | |
4165 | ||
4166 | case store_vec_info_type: | |
4167 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
4168 | gcc_assert (done); | |
4169 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) && !slp_node) | |
4170 | { | |
4171 | /* In case of interleaving, the whole chain is vectorized when the | |
4172 | last store in the chain is reached. Store stmts before the last | |
4173 | one are skipped, and there vec_stmt_info shouldn't be freed | |
4174 | meanwhile. */ | |
4175 | *strided_store = true; | |
4176 | if (STMT_VINFO_VEC_STMT (stmt_info)) | |
4177 | is_store = true; | |
4178 | } | |
4179 | else | |
4180 | is_store = true; | |
4181 | break; | |
4182 | ||
4183 | case condition_vec_info_type: | |
4184 | gcc_assert (!slp_node); | |
4bbe8262 | 4185 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0); |
ebfd146a IR |
4186 | gcc_assert (done); |
4187 | break; | |
4188 | ||
4189 | case call_vec_info_type: | |
4190 | gcc_assert (!slp_node); | |
4191 | done = vectorizable_call (stmt, gsi, &vec_stmt); | |
4192 | break; | |
4193 | ||
4194 | case reduc_vec_info_type: | |
4195 | gcc_assert (!slp_node); | |
4196 | done = vectorizable_reduction (stmt, gsi, &vec_stmt); | |
4197 | gcc_assert (done); | |
4198 | break; | |
4199 | ||
4200 | default: | |
4201 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
4202 | { | |
4203 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4204 | fprintf (vect_dump, "stmt not supported."); | |
4205 | gcc_unreachable (); | |
4206 | } | |
4207 | } | |
4208 | ||
4209 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that | |
4210 | is being vectorized, but outside the immediately enclosing loop. */ | |
4211 | if (vec_stmt | |
a70d6342 IR |
4212 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
4213 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
4214 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
4215 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
4216 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
a70d6342 IR |
4217 | || STMT_VINFO_RELEVANT (stmt_info) == |
4218 | vect_used_in_outer_by_reduction)) | |
ebfd146a | 4219 | { |
a70d6342 IR |
4220 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
4221 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
4222 | imm_use_iterator imm_iter; |
4223 | use_operand_p use_p; | |
4224 | tree scalar_dest; | |
4225 | gimple exit_phi; | |
4226 | ||
4227 | if (vect_print_dump_info (REPORT_DETAILS)) | |
a70d6342 | 4228 | fprintf (vect_dump, "Record the vdef for outer-loop vectorization."); |
ebfd146a IR |
4229 | |
4230 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
4231 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
4232 | STMT). */ | |
4233 | if (gimple_code (stmt) == GIMPLE_PHI) | |
4234 | scalar_dest = PHI_RESULT (stmt); | |
4235 | else | |
4236 | scalar_dest = gimple_assign_lhs (stmt); | |
4237 | ||
4238 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
4239 | { | |
4240 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
4241 | { | |
4242 | exit_phi = USE_STMT (use_p); | |
4243 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
4244 | } | |
4245 | } | |
4246 | } | |
4247 | ||
4248 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
4249 | being vectorized. */ | |
4250 | if (STMT_VINFO_LIVE_P (stmt_info) | |
4251 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
4252 | { | |
4253 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
4254 | gcc_assert (done); | |
4255 | } | |
4256 | ||
4257 | if (vec_stmt) | |
4258 | { | |
4259 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; | |
4260 | orig_stmt_in_pattern = STMT_VINFO_RELATED_STMT (stmt_info); | |
4261 | if (orig_stmt_in_pattern) | |
4262 | { | |
4263 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt_in_pattern); | |
4264 | /* STMT was inserted by the vectorizer to replace a computation idiom. | |
4265 | ORIG_STMT_IN_PATTERN is a stmt in the original sequence that | |
4266 | computed this idiom. We need to record a pointer to VEC_STMT in | |
4267 | the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the | |
4268 | documentation of vect_pattern_recog. */ | |
4269 | if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo)) | |
4270 | { | |
4271 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) == stmt); | |
4272 | STMT_VINFO_VEC_STMT (stmt_vinfo) = vec_stmt; | |
4273 | } | |
4274 | } | |
4275 | } | |
4276 | ||
4277 | return is_store; | |
4278 | } | |
4279 | ||
4280 | ||
4281 | /* Remove a group of stores (for SLP or interleaving), free their | |
4282 | stmt_vec_info. */ | |
4283 | ||
4284 | void | |
4285 | vect_remove_stores (gimple first_stmt) | |
4286 | { | |
4287 | gimple next = first_stmt; | |
4288 | gimple tmp; | |
4289 | gimple_stmt_iterator next_si; | |
4290 | ||
4291 | while (next) | |
4292 | { | |
4293 | /* Free the attached stmt_vec_info and remove the stmt. */ | |
4294 | next_si = gsi_for_stmt (next); | |
4295 | gsi_remove (&next_si, true); | |
4296 | tmp = DR_GROUP_NEXT_DR (vinfo_for_stmt (next)); | |
4297 | free_stmt_vec_info (next); | |
4298 | next = tmp; | |
4299 | } | |
4300 | } | |
4301 | ||
4302 | ||
4303 | /* Function new_stmt_vec_info. | |
4304 | ||
4305 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
4306 | ||
4307 | stmt_vec_info | |
a70d6342 IR |
4308 | new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo, |
4309 | bb_vec_info bb_vinfo) | |
ebfd146a IR |
4310 | { |
4311 | stmt_vec_info res; | |
4312 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
4313 | ||
4314 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
4315 | STMT_VINFO_STMT (res) = stmt; | |
4316 | STMT_VINFO_LOOP_VINFO (res) = loop_vinfo; | |
a70d6342 | 4317 | STMT_VINFO_BB_VINFO (res) = bb_vinfo; |
8644a673 | 4318 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
4319 | STMT_VINFO_LIVE_P (res) = false; |
4320 | STMT_VINFO_VECTYPE (res) = NULL; | |
4321 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4322 | STMT_VINFO_IN_PATTERN_P (res) = false; | |
4323 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
4324 | STMT_VINFO_DATA_REF (res) = NULL; | |
4325 | ||
4326 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
4327 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
4328 | STMT_VINFO_DR_INIT (res) = NULL; | |
4329 | STMT_VINFO_DR_STEP (res) = NULL; | |
4330 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
4331 | ||
4332 | if (gimple_code (stmt) == GIMPLE_PHI | |
4333 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
4334 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
4335 | else | |
8644a673 IR |
4336 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
4337 | ||
ebfd146a IR |
4338 | STMT_VINFO_SAME_ALIGN_REFS (res) = VEC_alloc (dr_p, heap, 5); |
4339 | STMT_VINFO_INSIDE_OF_LOOP_COST (res) = 0; | |
4340 | STMT_VINFO_OUTSIDE_OF_LOOP_COST (res) = 0; | |
32e8bb8e | 4341 | STMT_SLP_TYPE (res) = loop_vect; |
ebfd146a IR |
4342 | DR_GROUP_FIRST_DR (res) = NULL; |
4343 | DR_GROUP_NEXT_DR (res) = NULL; | |
4344 | DR_GROUP_SIZE (res) = 0; | |
4345 | DR_GROUP_STORE_COUNT (res) = 0; | |
4346 | DR_GROUP_GAP (res) = 0; | |
4347 | DR_GROUP_SAME_DR_STMT (res) = NULL; | |
4348 | DR_GROUP_READ_WRITE_DEPENDENCE (res) = false; | |
4349 | ||
4350 | return res; | |
4351 | } | |
4352 | ||
4353 | ||
4354 | /* Create a hash table for stmt_vec_info. */ | |
4355 | ||
4356 | void | |
4357 | init_stmt_vec_info_vec (void) | |
4358 | { | |
4359 | gcc_assert (!stmt_vec_info_vec); | |
4360 | stmt_vec_info_vec = VEC_alloc (vec_void_p, heap, 50); | |
4361 | } | |
4362 | ||
4363 | ||
4364 | /* Free hash table for stmt_vec_info. */ | |
4365 | ||
4366 | void | |
4367 | free_stmt_vec_info_vec (void) | |
4368 | { | |
4369 | gcc_assert (stmt_vec_info_vec); | |
4370 | VEC_free (vec_void_p, heap, stmt_vec_info_vec); | |
4371 | } | |
4372 | ||
4373 | ||
4374 | /* Free stmt vectorization related info. */ | |
4375 | ||
4376 | void | |
4377 | free_stmt_vec_info (gimple stmt) | |
4378 | { | |
4379 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4380 | ||
4381 | if (!stmt_info) | |
4382 | return; | |
4383 | ||
4384 | VEC_free (dr_p, heap, STMT_VINFO_SAME_ALIGN_REFS (stmt_info)); | |
4385 | set_vinfo_for_stmt (stmt, NULL); | |
4386 | free (stmt_info); | |
4387 | } | |
4388 | ||
4389 | ||
4390 | /* Function get_vectype_for_scalar_type. | |
4391 | ||
4392 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
4393 | by the target. */ | |
4394 | ||
4395 | tree | |
4396 | get_vectype_for_scalar_type (tree scalar_type) | |
4397 | { | |
4398 | enum machine_mode inner_mode = TYPE_MODE (scalar_type); | |
4399 | int nbytes = GET_MODE_SIZE (inner_mode); | |
4400 | int nunits; | |
4401 | tree vectype; | |
4402 | ||
4403 | if (nbytes == 0 || nbytes >= UNITS_PER_SIMD_WORD (inner_mode)) | |
4404 | return NULL_TREE; | |
4405 | ||
4406 | /* FORNOW: Only a single vector size per mode (UNITS_PER_SIMD_WORD) | |
4407 | is expected. */ | |
4408 | nunits = UNITS_PER_SIMD_WORD (inner_mode) / nbytes; | |
4409 | ||
4410 | vectype = build_vector_type (scalar_type, nunits); | |
4411 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4412 | { | |
4413 | fprintf (vect_dump, "get vectype with %d units of type ", nunits); | |
4414 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
4415 | } | |
4416 | ||
4417 | if (!vectype) | |
4418 | return NULL_TREE; | |
4419 | ||
4420 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4421 | { | |
4422 | fprintf (vect_dump, "vectype: "); | |
4423 | print_generic_expr (vect_dump, vectype, TDF_SLIM); | |
4424 | } | |
4425 | ||
4426 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
4427 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
4428 | { | |
4429 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4430 | fprintf (vect_dump, "mode not supported by target."); | |
4431 | return NULL_TREE; | |
4432 | } | |
4433 | ||
4434 | return vectype; | |
4435 | } | |
4436 | ||
4437 | /* Function vect_is_simple_use. | |
4438 | ||
4439 | Input: | |
a70d6342 IR |
4440 | LOOP_VINFO - the vect info of the loop that is being vectorized. |
4441 | BB_VINFO - the vect info of the basic block that is being vectorized. | |
4442 | OPERAND - operand of a stmt in the loop or bb. | |
ebfd146a IR |
4443 | DEF - the defining stmt in case OPERAND is an SSA_NAME. |
4444 | ||
4445 | Returns whether a stmt with OPERAND can be vectorized. | |
a70d6342 IR |
4446 | For loops, supportable operands are constants, loop invariants, and operands |
4447 | that are defined by the current iteration of the loop. Unsupportable | |
4448 | operands are those that are defined by a previous iteration of the loop (as | |
4449 | is the case in reduction/induction computations). | |
4450 | For basic blocks, supportable operands are constants and bb invariants. | |
4451 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
4452 | |
4453 | bool | |
a70d6342 IR |
4454 | vect_is_simple_use (tree operand, loop_vec_info loop_vinfo, |
4455 | bb_vec_info bb_vinfo, gimple *def_stmt, | |
ebfd146a IR |
4456 | tree *def, enum vect_def_type *dt) |
4457 | { | |
4458 | basic_block bb; | |
4459 | stmt_vec_info stmt_vinfo; | |
a70d6342 IR |
4460 | struct loop *loop = NULL; |
4461 | ||
4462 | if (loop_vinfo) | |
4463 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
4464 | |
4465 | *def_stmt = NULL; | |
4466 | *def = NULL_TREE; | |
4467 | ||
4468 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4469 | { | |
4470 | fprintf (vect_dump, "vect_is_simple_use: operand "); | |
4471 | print_generic_expr (vect_dump, operand, TDF_SLIM); | |
4472 | } | |
4473 | ||
4474 | if (TREE_CODE (operand) == INTEGER_CST || TREE_CODE (operand) == REAL_CST) | |
4475 | { | |
4476 | *dt = vect_constant_def; | |
4477 | return true; | |
4478 | } | |
a70d6342 | 4479 | |
ebfd146a IR |
4480 | if (is_gimple_min_invariant (operand)) |
4481 | { | |
4482 | *def = operand; | |
8644a673 | 4483 | *dt = vect_external_def; |
ebfd146a IR |
4484 | return true; |
4485 | } | |
4486 | ||
4487 | if (TREE_CODE (operand) == PAREN_EXPR) | |
4488 | { | |
4489 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4490 | fprintf (vect_dump, "non-associatable copy."); | |
4491 | operand = TREE_OPERAND (operand, 0); | |
4492 | } | |
a70d6342 | 4493 | |
ebfd146a IR |
4494 | if (TREE_CODE (operand) != SSA_NAME) |
4495 | { | |
4496 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4497 | fprintf (vect_dump, "not ssa-name."); | |
4498 | return false; | |
4499 | } | |
4500 | ||
4501 | *def_stmt = SSA_NAME_DEF_STMT (operand); | |
4502 | if (*def_stmt == NULL) | |
4503 | { | |
4504 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4505 | fprintf (vect_dump, "no def_stmt."); | |
4506 | return false; | |
4507 | } | |
4508 | ||
4509 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4510 | { | |
4511 | fprintf (vect_dump, "def_stmt: "); | |
4512 | print_gimple_stmt (vect_dump, *def_stmt, 0, TDF_SLIM); | |
4513 | } | |
4514 | ||
8644a673 | 4515 | /* Empty stmt is expected only in case of a function argument. |
ebfd146a IR |
4516 | (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */ |
4517 | if (gimple_nop_p (*def_stmt)) | |
4518 | { | |
4519 | *def = operand; | |
8644a673 | 4520 | *dt = vect_external_def; |
ebfd146a IR |
4521 | return true; |
4522 | } | |
4523 | ||
4524 | bb = gimple_bb (*def_stmt); | |
a70d6342 IR |
4525 | |
4526 | if ((loop && !flow_bb_inside_loop_p (loop, bb)) | |
4527 | || (!loop && bb != BB_VINFO_BB (bb_vinfo)) | |
4528 | || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI)) | |
8644a673 | 4529 | *dt = vect_external_def; |
ebfd146a IR |
4530 | else |
4531 | { | |
4532 | stmt_vinfo = vinfo_for_stmt (*def_stmt); | |
4533 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
4534 | } | |
4535 | ||
4536 | if (*dt == vect_unknown_def_type) | |
4537 | { | |
4538 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4539 | fprintf (vect_dump, "Unsupported pattern."); | |
4540 | return false; | |
4541 | } | |
4542 | ||
4543 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4544 | fprintf (vect_dump, "type of def: %d.",*dt); | |
4545 | ||
4546 | switch (gimple_code (*def_stmt)) | |
4547 | { | |
4548 | case GIMPLE_PHI: | |
4549 | *def = gimple_phi_result (*def_stmt); | |
4550 | break; | |
4551 | ||
4552 | case GIMPLE_ASSIGN: | |
4553 | *def = gimple_assign_lhs (*def_stmt); | |
4554 | break; | |
4555 | ||
4556 | case GIMPLE_CALL: | |
4557 | *def = gimple_call_lhs (*def_stmt); | |
4558 | if (*def != NULL) | |
4559 | break; | |
4560 | /* FALLTHRU */ | |
4561 | default: | |
4562 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4563 | fprintf (vect_dump, "unsupported defining stmt: "); | |
4564 | return false; | |
4565 | } | |
4566 | ||
4567 | return true; | |
4568 | } | |
4569 | ||
4570 | ||
4571 | /* Function supportable_widening_operation | |
4572 | ||
4573 | Check whether an operation represented by the code CODE is a | |
4574 | widening operation that is supported by the target platform in | |
4575 | vector form (i.e., when operating on arguments of type VECTYPE). | |
4576 | ||
4577 | Widening operations we currently support are NOP (CONVERT), FLOAT | |
4578 | and WIDEN_MULT. This function checks if these operations are supported | |
4579 | by the target platform either directly (via vector tree-codes), or via | |
4580 | target builtins. | |
4581 | ||
4582 | Output: | |
4583 | - CODE1 and CODE2 are codes of vector operations to be used when | |
4584 | vectorizing the operation, if available. | |
4585 | - DECL1 and DECL2 are decls of target builtin functions to be used | |
4586 | when vectorizing the operation, if available. In this case, | |
4587 | CODE1 and CODE2 are CALL_EXPR. | |
4588 | - MULTI_STEP_CVT determines the number of required intermediate steps in | |
4589 | case of multi-step conversion (like char->short->int - in that case | |
4590 | MULTI_STEP_CVT will be 1). | |
4591 | - INTERM_TYPES contains the intermediate type required to perform the | |
4592 | widening operation (short in the above example). */ | |
4593 | ||
4594 | bool | |
4595 | supportable_widening_operation (enum tree_code code, gimple stmt, tree vectype, | |
4596 | tree *decl1, tree *decl2, | |
4597 | enum tree_code *code1, enum tree_code *code2, | |
4598 | int *multi_step_cvt, | |
4599 | VEC (tree, heap) **interm_types) | |
4600 | { | |
4601 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4602 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4603 | struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info); | |
4604 | bool ordered_p; | |
4605 | enum machine_mode vec_mode; | |
81f40b79 | 4606 | enum insn_code icode1, icode2; |
ebfd146a IR |
4607 | optab optab1, optab2; |
4608 | tree type = gimple_expr_type (stmt); | |
4609 | tree wide_vectype = get_vectype_for_scalar_type (type); | |
4610 | enum tree_code c1, c2; | |
4611 | ||
4612 | /* The result of a vectorized widening operation usually requires two vectors | |
4613 | (because the widened results do not fit int one vector). The generated | |
4614 | vector results would normally be expected to be generated in the same | |
4615 | order as in the original scalar computation, i.e. if 8 results are | |
4616 | generated in each vector iteration, they are to be organized as follows: | |
4617 | vect1: [res1,res2,res3,res4], vect2: [res5,res6,res7,res8]. | |
4618 | ||
4619 | However, in the special case that the result of the widening operation is | |
4620 | used in a reduction computation only, the order doesn't matter (because | |
4621 | when vectorizing a reduction we change the order of the computation). | |
4622 | Some targets can take advantage of this and generate more efficient code. | |
4623 | For example, targets like Altivec, that support widen_mult using a sequence | |
4624 | of {mult_even,mult_odd} generate the following vectors: | |
4625 | vect1: [res1,res3,res5,res7], vect2: [res2,res4,res6,res8]. | |
4626 | ||
4627 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
4628 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
4629 | iterations in parallel). We therefore don't allow to change the order | |
4630 | of the computation in the inner-loop during outer-loop vectorization. */ | |
4631 | ||
4632 | if (STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
4633 | && !nested_in_vect_loop_p (vect_loop, stmt)) | |
4634 | ordered_p = false; | |
4635 | else | |
4636 | ordered_p = true; | |
4637 | ||
4638 | if (!ordered_p | |
4639 | && code == WIDEN_MULT_EXPR | |
4640 | && targetm.vectorize.builtin_mul_widen_even | |
4641 | && targetm.vectorize.builtin_mul_widen_even (vectype) | |
4642 | && targetm.vectorize.builtin_mul_widen_odd | |
4643 | && targetm.vectorize.builtin_mul_widen_odd (vectype)) | |
4644 | { | |
4645 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4646 | fprintf (vect_dump, "Unordered widening operation detected."); | |
4647 | ||
4648 | *code1 = *code2 = CALL_EXPR; | |
4649 | *decl1 = targetm.vectorize.builtin_mul_widen_even (vectype); | |
4650 | *decl2 = targetm.vectorize.builtin_mul_widen_odd (vectype); | |
4651 | return true; | |
4652 | } | |
4653 | ||
4654 | switch (code) | |
4655 | { | |
4656 | case WIDEN_MULT_EXPR: | |
4657 | if (BYTES_BIG_ENDIAN) | |
4658 | { | |
4659 | c1 = VEC_WIDEN_MULT_HI_EXPR; | |
4660 | c2 = VEC_WIDEN_MULT_LO_EXPR; | |
4661 | } | |
4662 | else | |
4663 | { | |
4664 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
4665 | c1 = VEC_WIDEN_MULT_LO_EXPR; | |
4666 | } | |
4667 | break; | |
4668 | ||
4669 | CASE_CONVERT: | |
4670 | if (BYTES_BIG_ENDIAN) | |
4671 | { | |
4672 | c1 = VEC_UNPACK_HI_EXPR; | |
4673 | c2 = VEC_UNPACK_LO_EXPR; | |
4674 | } | |
4675 | else | |
4676 | { | |
4677 | c2 = VEC_UNPACK_HI_EXPR; | |
4678 | c1 = VEC_UNPACK_LO_EXPR; | |
4679 | } | |
4680 | break; | |
4681 | ||
4682 | case FLOAT_EXPR: | |
4683 | if (BYTES_BIG_ENDIAN) | |
4684 | { | |
4685 | c1 = VEC_UNPACK_FLOAT_HI_EXPR; | |
4686 | c2 = VEC_UNPACK_FLOAT_LO_EXPR; | |
4687 | } | |
4688 | else | |
4689 | { | |
4690 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
4691 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; | |
4692 | } | |
4693 | break; | |
4694 | ||
4695 | case FIX_TRUNC_EXPR: | |
4696 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
4697 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
4698 | computing the operation. */ | |
4699 | return false; | |
4700 | ||
4701 | default: | |
4702 | gcc_unreachable (); | |
4703 | } | |
4704 | ||
4705 | if (code == FIX_TRUNC_EXPR) | |
4706 | { | |
4707 | /* The signedness is determined from output operand. */ | |
4708 | optab1 = optab_for_tree_code (c1, type, optab_default); | |
4709 | optab2 = optab_for_tree_code (c2, type, optab_default); | |
4710 | } | |
4711 | else | |
4712 | { | |
4713 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
4714 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
4715 | } | |
4716 | ||
4717 | if (!optab1 || !optab2) | |
4718 | return false; | |
4719 | ||
4720 | vec_mode = TYPE_MODE (vectype); | |
4721 | if ((icode1 = optab_handler (optab1, vec_mode)->insn_code) == CODE_FOR_nothing | |
4722 | || (icode2 = optab_handler (optab2, vec_mode)->insn_code) | |
4723 | == CODE_FOR_nothing) | |
4724 | return false; | |
4725 | ||
4726 | /* Check if it's a multi-step conversion that can be done using intermediate | |
4727 | types. */ | |
4728 | if (insn_data[icode1].operand[0].mode != TYPE_MODE (wide_vectype) | |
4729 | || insn_data[icode2].operand[0].mode != TYPE_MODE (wide_vectype)) | |
4730 | { | |
4731 | int i; | |
4732 | tree prev_type = vectype, intermediate_type; | |
4733 | enum machine_mode intermediate_mode, prev_mode = vec_mode; | |
4734 | optab optab3, optab4; | |
4735 | ||
4736 | if (!CONVERT_EXPR_CODE_P (code)) | |
4737 | return false; | |
4738 | ||
4739 | *code1 = c1; | |
4740 | *code2 = c2; | |
4741 | ||
4742 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS | |
4743 | intermediate steps in promotion sequence. We try MAX_INTERM_CVT_STEPS | |
4744 | to get to NARROW_VECTYPE, and fail if we do not. */ | |
4745 | *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS); | |
4746 | for (i = 0; i < 3; i++) | |
4747 | { | |
4748 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
4749 | intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode, | |
4750 | TYPE_UNSIGNED (prev_type)); | |
4751 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
4752 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
4753 | ||
4754 | if (!optab3 || !optab4 | |
4755 | || (icode1 = optab1->handlers[(int) prev_mode].insn_code) | |
4756 | == CODE_FOR_nothing | |
4757 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
4758 | || (icode2 = optab2->handlers[(int) prev_mode].insn_code) | |
4759 | == CODE_FOR_nothing | |
4760 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
4761 | || (icode1 = optab3->handlers[(int) intermediate_mode].insn_code) | |
4762 | == CODE_FOR_nothing | |
4763 | || (icode2 = optab4->handlers[(int) intermediate_mode].insn_code) | |
4764 | == CODE_FOR_nothing) | |
4765 | return false; | |
4766 | ||
4767 | VEC_quick_push (tree, *interm_types, intermediate_type); | |
4768 | (*multi_step_cvt)++; | |
4769 | ||
4770 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
4771 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
4772 | return true; | |
4773 | ||
4774 | prev_type = intermediate_type; | |
4775 | prev_mode = intermediate_mode; | |
4776 | } | |
4777 | ||
4778 | return false; | |
4779 | } | |
4780 | ||
4781 | *code1 = c1; | |
4782 | *code2 = c2; | |
4783 | return true; | |
4784 | } | |
4785 | ||
4786 | ||
4787 | /* Function supportable_narrowing_operation | |
4788 | ||
4789 | Check whether an operation represented by the code CODE is a | |
4790 | narrowing operation that is supported by the target platform in | |
4791 | vector form (i.e., when operating on arguments of type VECTYPE). | |
4792 | ||
4793 | Narrowing operations we currently support are NOP (CONVERT) and | |
4794 | FIX_TRUNC. This function checks if these operations are supported by | |
4795 | the target platform directly via vector tree-codes. | |
4796 | ||
4797 | Output: | |
4798 | - CODE1 is the code of a vector operation to be used when | |
4799 | vectorizing the operation, if available. | |
4800 | - MULTI_STEP_CVT determines the number of required intermediate steps in | |
4801 | case of multi-step conversion (like int->short->char - in that case | |
4802 | MULTI_STEP_CVT will be 1). | |
4803 | - INTERM_TYPES contains the intermediate type required to perform the | |
4804 | narrowing operation (short in the above example). */ | |
4805 | ||
4806 | bool | |
4807 | supportable_narrowing_operation (enum tree_code code, | |
4808 | const_gimple stmt, tree vectype, | |
4809 | enum tree_code *code1, int *multi_step_cvt, | |
4810 | VEC (tree, heap) **interm_types) | |
4811 | { | |
4812 | enum machine_mode vec_mode; | |
4813 | enum insn_code icode1; | |
4814 | optab optab1, interm_optab; | |
4815 | tree type = gimple_expr_type (stmt); | |
4816 | tree narrow_vectype = get_vectype_for_scalar_type (type); | |
4817 | enum tree_code c1; | |
4818 | tree intermediate_type, prev_type; | |
4819 | int i; | |
4820 | ||
4821 | switch (code) | |
4822 | { | |
4823 | CASE_CONVERT: | |
4824 | c1 = VEC_PACK_TRUNC_EXPR; | |
4825 | break; | |
4826 | ||
4827 | case FIX_TRUNC_EXPR: | |
4828 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
4829 | break; | |
4830 | ||
4831 | case FLOAT_EXPR: | |
4832 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
4833 | tree code and optabs used for computing the operation. */ | |
4834 | return false; | |
4835 | ||
4836 | default: | |
4837 | gcc_unreachable (); | |
4838 | } | |
4839 | ||
4840 | if (code == FIX_TRUNC_EXPR) | |
4841 | /* The signedness is determined from output operand. */ | |
4842 | optab1 = optab_for_tree_code (c1, type, optab_default); | |
4843 | else | |
4844 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
4845 | ||
4846 | if (!optab1) | |
4847 | return false; | |
4848 | ||
4849 | vec_mode = TYPE_MODE (vectype); | |
4850 | if ((icode1 = optab_handler (optab1, vec_mode)->insn_code) | |
4851 | == CODE_FOR_nothing) | |
4852 | return false; | |
4853 | ||
4854 | /* Check if it's a multi-step conversion that can be done using intermediate | |
4855 | types. */ | |
4856 | if (insn_data[icode1].operand[0].mode != TYPE_MODE (narrow_vectype)) | |
4857 | { | |
4858 | enum machine_mode intermediate_mode, prev_mode = vec_mode; | |
4859 | ||
4860 | *code1 = c1; | |
4861 | prev_type = vectype; | |
4862 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS | |
4863 | intermediate steps in promotion sequence. We try MAX_INTERM_CVT_STEPS | |
4864 | to get to NARROW_VECTYPE, and fail if we do not. */ | |
4865 | *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS); | |
4866 | for (i = 0; i < 3; i++) | |
4867 | { | |
4868 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
4869 | intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode, | |
4870 | TYPE_UNSIGNED (prev_type)); | |
4871 | interm_optab = optab_for_tree_code (c1, intermediate_type, | |
4872 | optab_default); | |
4873 | if (!interm_optab | |
4874 | || (icode1 = optab1->handlers[(int) prev_mode].insn_code) | |
4875 | == CODE_FOR_nothing | |
4876 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
4877 | || (icode1 | |
4878 | = interm_optab->handlers[(int) intermediate_mode].insn_code) | |
4879 | == CODE_FOR_nothing) | |
4880 | return false; | |
4881 | ||
4882 | VEC_quick_push (tree, *interm_types, intermediate_type); | |
4883 | (*multi_step_cvt)++; | |
4884 | ||
4885 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
4886 | return true; | |
4887 | ||
4888 | prev_type = intermediate_type; | |
4889 | prev_mode = intermediate_mode; | |
4890 | } | |
4891 | ||
4892 | return false; | |
4893 | } | |
4894 | ||
4895 | *code1 = c1; | |
4896 | return true; | |
4897 | } |