]>
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 | ||
3157b0c2 AO |
148 | if (is_gimple_debug (USE_STMT (use_p))) |
149 | continue; | |
150 | ||
ebfd146a IR |
151 | /* We expect all such uses to be in the loop exit phis |
152 | (because of loop closed form) */ | |
153 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
154 | gcc_assert (bb == single_exit (loop)->dest); | |
155 | ||
156 | *live_p = true; | |
157 | } | |
158 | } | |
159 | } | |
160 | ||
161 | return (*live_p || *relevant); | |
162 | } | |
163 | ||
164 | ||
165 | /* Function exist_non_indexing_operands_for_use_p | |
166 | ||
167 | USE is one of the uses attached to STMT. Check if USE is | |
168 | used in STMT for anything other than indexing an array. */ | |
169 | ||
170 | static bool | |
171 | exist_non_indexing_operands_for_use_p (tree use, gimple stmt) | |
172 | { | |
173 | tree operand; | |
174 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 175 | |
ebfd146a IR |
176 | /* USE corresponds to some operand in STMT. If there is no data |
177 | reference in STMT, then any operand that corresponds to USE | |
178 | is not indexing an array. */ | |
179 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
180 | return true; | |
59a05b0c | 181 | |
ebfd146a IR |
182 | /* STMT has a data_ref. FORNOW this means that its of one of |
183 | the following forms: | |
184 | -1- ARRAY_REF = var | |
185 | -2- var = ARRAY_REF | |
186 | (This should have been verified in analyze_data_refs). | |
187 | ||
188 | 'var' in the second case corresponds to a def, not a use, | |
189 | so USE cannot correspond to any operands that are not used | |
190 | for array indexing. | |
191 | ||
192 | Therefore, all we need to check is if STMT falls into the | |
193 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
194 | |
195 | if (!gimple_assign_copy_p (stmt)) | |
196 | return false; | |
59a05b0c EB |
197 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
198 | return false; | |
ebfd146a | 199 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
200 | if (TREE_CODE (operand) != SSA_NAME) |
201 | return false; | |
202 | ||
203 | if (operand == use) | |
204 | return true; | |
205 | ||
206 | return false; | |
207 | } | |
208 | ||
209 | ||
210 | /* | |
211 | Function process_use. | |
212 | ||
213 | Inputs: | |
214 | - a USE in STMT in a loop represented by LOOP_VINFO | |
215 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt | |
216 | that defined USE. This is done by calling mark_relevant and passing it | |
217 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). | |
218 | ||
219 | Outputs: | |
220 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
221 | relevance info of the DEF_STMT of this USE: | |
222 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
223 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
224 | Exceptions: | |
225 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
226 | which does not need to be directly vectorized, then the liveness/relevance | |
227 | of the respective DEF_STMT is left unchanged. | |
228 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we | |
229 | skip DEF_STMT cause it had already been processed. | |
230 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will | |
231 | be modified accordingly. | |
232 | ||
233 | Return true if everything is as expected. Return false otherwise. */ | |
234 | ||
235 | static bool | |
236 | process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p, | |
237 | enum vect_relevant relevant, VEC(gimple,heap) **worklist) | |
238 | { | |
239 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
240 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
241 | stmt_vec_info dstmt_vinfo; | |
242 | basic_block bb, def_bb; | |
243 | tree def; | |
244 | gimple def_stmt; | |
245 | enum vect_def_type dt; | |
246 | ||
247 | /* case 1: we are only interested in uses that need to be vectorized. Uses | |
248 | that are used for address computation are not considered relevant. */ | |
249 | if (!exist_non_indexing_operands_for_use_p (use, stmt)) | |
250 | return true; | |
251 | ||
a70d6342 | 252 | if (!vect_is_simple_use (use, loop_vinfo, NULL, &def_stmt, &def, &dt)) |
ebfd146a | 253 | { |
8644a673 | 254 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
ebfd146a IR |
255 | fprintf (vect_dump, "not vectorized: unsupported use in stmt."); |
256 | return false; | |
257 | } | |
258 | ||
259 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
260 | return true; | |
261 | ||
262 | def_bb = gimple_bb (def_stmt); | |
263 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
264 | { | |
265 | if (vect_print_dump_info (REPORT_DETAILS)) | |
266 | fprintf (vect_dump, "def_stmt is out of loop."); | |
267 | return true; | |
268 | } | |
269 | ||
270 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). | |
271 | DEF_STMT must have already been processed, because this should be the | |
272 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
273 | as there should be no other uses for DEF_STMT in the loop. So we just | |
274 | check that everything is as expected, and we are done. */ | |
275 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
276 | bb = gimple_bb (stmt); | |
277 | if (gimple_code (stmt) == GIMPLE_PHI | |
278 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
279 | && gimple_code (def_stmt) != GIMPLE_PHI | |
280 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
281 | && bb->loop_father == def_bb->loop_father) | |
282 | { | |
283 | if (vect_print_dump_info (REPORT_DETAILS)) | |
284 | fprintf (vect_dump, "reduc-stmt defining reduc-phi in the same nest."); | |
285 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) | |
286 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
287 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
288 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) | |
8644a673 | 289 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
290 | return true; |
291 | } | |
292 | ||
293 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
294 | outer-loop-header-bb: | |
295 | d = def_stmt | |
296 | inner-loop: | |
297 | stmt # use (d) | |
298 | outer-loop-tail-bb: | |
299 | ... */ | |
300 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
301 | { | |
302 | if (vect_print_dump_info (REPORT_DETAILS)) | |
303 | fprintf (vect_dump, "outer-loop def-stmt defining inner-loop stmt."); | |
7c5222ff | 304 | |
ebfd146a IR |
305 | switch (relevant) |
306 | { | |
8644a673 | 307 | case vect_unused_in_scope: |
7c5222ff IR |
308 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
309 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 310 | break; |
7c5222ff | 311 | |
ebfd146a | 312 | case vect_used_in_outer_by_reduction: |
7c5222ff | 313 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
314 | relevant = vect_used_by_reduction; |
315 | break; | |
7c5222ff | 316 | |
ebfd146a | 317 | case vect_used_in_outer: |
7c5222ff | 318 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 319 | relevant = vect_used_in_scope; |
ebfd146a | 320 | break; |
7c5222ff | 321 | |
8644a673 | 322 | case vect_used_in_scope: |
ebfd146a IR |
323 | break; |
324 | ||
325 | default: | |
326 | gcc_unreachable (); | |
327 | } | |
328 | } | |
329 | ||
330 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
331 | outer-loop-header-bb: | |
332 | ... | |
333 | inner-loop: | |
334 | d = def_stmt | |
06066f92 | 335 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
336 | stmt # use (d) */ |
337 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
338 | { | |
339 | if (vect_print_dump_info (REPORT_DETAILS)) | |
340 | fprintf (vect_dump, "inner-loop def-stmt defining outer-loop stmt."); | |
7c5222ff | 341 | |
ebfd146a IR |
342 | switch (relevant) |
343 | { | |
8644a673 | 344 | case vect_unused_in_scope: |
06066f92 IR |
345 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
346 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? | |
a70d6342 | 347 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
348 | break; |
349 | ||
ebfd146a IR |
350 | case vect_used_by_reduction: |
351 | relevant = vect_used_in_outer_by_reduction; | |
352 | break; | |
353 | ||
8644a673 | 354 | case vect_used_in_scope: |
ebfd146a IR |
355 | relevant = vect_used_in_outer; |
356 | break; | |
357 | ||
358 | default: | |
359 | gcc_unreachable (); | |
360 | } | |
361 | } | |
362 | ||
363 | vect_mark_relevant (worklist, def_stmt, relevant, live_p); | |
364 | return true; | |
365 | } | |
366 | ||
367 | ||
368 | /* Function vect_mark_stmts_to_be_vectorized. | |
369 | ||
370 | Not all stmts in the loop need to be vectorized. For example: | |
371 | ||
372 | for i... | |
373 | for j... | |
374 | 1. T0 = i + j | |
375 | 2. T1 = a[T0] | |
376 | ||
377 | 3. j = j + 1 | |
378 | ||
379 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
380 | addressing of vectorized data-refs are handled differently. | |
381 | ||
382 | This pass detects such stmts. */ | |
383 | ||
384 | bool | |
385 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
386 | { | |
387 | VEC(gimple,heap) *worklist; | |
388 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
389 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
390 | unsigned int nbbs = loop->num_nodes; | |
391 | gimple_stmt_iterator si; | |
392 | gimple stmt; | |
393 | unsigned int i; | |
394 | stmt_vec_info stmt_vinfo; | |
395 | basic_block bb; | |
396 | gimple phi; | |
397 | bool live_p; | |
06066f92 IR |
398 | enum vect_relevant relevant, tmp_relevant; |
399 | enum vect_def_type def_type; | |
ebfd146a IR |
400 | |
401 | if (vect_print_dump_info (REPORT_DETAILS)) | |
402 | fprintf (vect_dump, "=== vect_mark_stmts_to_be_vectorized ==="); | |
403 | ||
404 | worklist = VEC_alloc (gimple, heap, 64); | |
405 | ||
406 | /* 1. Init worklist. */ | |
407 | for (i = 0; i < nbbs; i++) | |
408 | { | |
409 | bb = bbs[i]; | |
410 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
411 | { | |
412 | phi = gsi_stmt (si); | |
413 | if (vect_print_dump_info (REPORT_DETAILS)) | |
414 | { | |
415 | fprintf (vect_dump, "init: phi relevant? "); | |
416 | print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM); | |
417 | } | |
418 | ||
419 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
420 | vect_mark_relevant (&worklist, phi, relevant, live_p); | |
421 | } | |
422 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
423 | { | |
424 | stmt = gsi_stmt (si); | |
425 | if (vect_print_dump_info (REPORT_DETAILS)) | |
426 | { | |
427 | fprintf (vect_dump, "init: stmt relevant? "); | |
428 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
429 | } | |
430 | ||
431 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
432 | vect_mark_relevant (&worklist, stmt, relevant, live_p); | |
433 | } | |
434 | } | |
435 | ||
436 | /* 2. Process_worklist */ | |
437 | while (VEC_length (gimple, worklist) > 0) | |
438 | { | |
439 | use_operand_p use_p; | |
440 | ssa_op_iter iter; | |
441 | ||
442 | stmt = VEC_pop (gimple, worklist); | |
443 | if (vect_print_dump_info (REPORT_DETAILS)) | |
444 | { | |
445 | fprintf (vect_dump, "worklist: examine stmt: "); | |
446 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
447 | } | |
448 | ||
449 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it | |
450 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
451 | liveness and relevance properties of STMT. */ | |
452 | stmt_vinfo = vinfo_for_stmt (stmt); | |
453 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
454 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
455 | ||
456 | /* Generally, the liveness and relevance properties of STMT are | |
457 | propagated as is to the DEF_STMTs of its USEs: | |
458 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
459 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
460 | ||
461 | One exception is when STMT has been identified as defining a reduction | |
462 | variable; in this case we set the liveness/relevance as follows: | |
463 | live_p = false | |
464 | relevant = vect_used_by_reduction | |
465 | This is because we distinguish between two kinds of relevant stmts - | |
466 | those that are used by a reduction computation, and those that are | |
467 | (also) used by a regular computation. This allows us later on to | |
468 | identify stmts that are used solely by a reduction, and therefore the | |
7c5222ff | 469 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 470 | |
06066f92 IR |
471 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
472 | tmp_relevant = relevant; | |
473 | switch (def_type) | |
ebfd146a | 474 | { |
06066f92 IR |
475 | case vect_reduction_def: |
476 | switch (tmp_relevant) | |
477 | { | |
478 | case vect_unused_in_scope: | |
479 | relevant = vect_used_by_reduction; | |
480 | break; | |
481 | ||
482 | case vect_used_by_reduction: | |
483 | if (gimple_code (stmt) == GIMPLE_PHI) | |
484 | break; | |
485 | /* fall through */ | |
486 | ||
487 | default: | |
488 | if (vect_print_dump_info (REPORT_DETAILS)) | |
489 | fprintf (vect_dump, "unsupported use of reduction."); | |
490 | ||
491 | VEC_free (gimple, heap, worklist); | |
492 | return false; | |
493 | } | |
494 | ||
495 | live_p = false; | |
496 | break; | |
497 | ||
498 | case vect_nested_cycle: | |
499 | if (tmp_relevant != vect_unused_in_scope | |
500 | && tmp_relevant != vect_used_in_outer_by_reduction | |
501 | && tmp_relevant != vect_used_in_outer) | |
502 | { | |
503 | if (vect_print_dump_info (REPORT_DETAILS)) | |
504 | fprintf (vect_dump, "unsupported use of nested cycle."); | |
7c5222ff | 505 | |
06066f92 IR |
506 | VEC_free (gimple, heap, worklist); |
507 | return false; | |
508 | } | |
7c5222ff | 509 | |
06066f92 IR |
510 | live_p = false; |
511 | break; | |
512 | ||
513 | case vect_double_reduction_def: | |
514 | if (tmp_relevant != vect_unused_in_scope | |
515 | && tmp_relevant != vect_used_by_reduction) | |
516 | { | |
7c5222ff | 517 | if (vect_print_dump_info (REPORT_DETAILS)) |
06066f92 | 518 | fprintf (vect_dump, "unsupported use of double reduction."); |
7c5222ff IR |
519 | |
520 | VEC_free (gimple, heap, worklist); | |
521 | return false; | |
06066f92 IR |
522 | } |
523 | ||
524 | live_p = false; | |
525 | break; | |
7c5222ff | 526 | |
06066f92 IR |
527 | default: |
528 | break; | |
7c5222ff IR |
529 | } |
530 | ||
ebfd146a IR |
531 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) |
532 | { | |
533 | tree op = USE_FROM_PTR (use_p); | |
534 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, &worklist)) | |
535 | { | |
536 | VEC_free (gimple, heap, worklist); | |
537 | return false; | |
538 | } | |
539 | } | |
540 | } /* while worklist */ | |
541 | ||
542 | VEC_free (gimple, heap, worklist); | |
543 | return true; | |
544 | } | |
545 | ||
546 | ||
547 | int | |
548 | cost_for_stmt (gimple stmt) | |
549 | { | |
550 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
551 | ||
552 | switch (STMT_VINFO_TYPE (stmt_info)) | |
553 | { | |
554 | case load_vec_info_type: | |
555 | return TARG_SCALAR_LOAD_COST; | |
556 | case store_vec_info_type: | |
557 | return TARG_SCALAR_STORE_COST; | |
558 | case op_vec_info_type: | |
559 | case condition_vec_info_type: | |
560 | case assignment_vec_info_type: | |
561 | case reduc_vec_info_type: | |
562 | case induc_vec_info_type: | |
563 | case type_promotion_vec_info_type: | |
564 | case type_demotion_vec_info_type: | |
565 | case type_conversion_vec_info_type: | |
566 | case call_vec_info_type: | |
567 | return TARG_SCALAR_STMT_COST; | |
568 | case undef_vec_info_type: | |
569 | default: | |
570 | gcc_unreachable (); | |
571 | } | |
572 | } | |
573 | ||
574 | /* Function vect_model_simple_cost. | |
575 | ||
576 | Models cost for simple operations, i.e. those that only emit ncopies of a | |
577 | single op. Right now, this does not account for multiple insns that could | |
578 | be generated for the single vector op. We will handle that shortly. */ | |
579 | ||
580 | void | |
581 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, | |
582 | enum vect_def_type *dt, slp_tree slp_node) | |
583 | { | |
584 | int i; | |
585 | int inside_cost = 0, outside_cost = 0; | |
586 | ||
587 | /* The SLP costs were already calculated during SLP tree build. */ | |
588 | if (PURE_SLP_STMT (stmt_info)) | |
589 | return; | |
590 | ||
591 | inside_cost = ncopies * TARG_VEC_STMT_COST; | |
592 | ||
593 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
594 | for (i = 0; i < 2; i++) | |
595 | { | |
8644a673 | 596 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
ebfd146a IR |
597 | outside_cost += TARG_SCALAR_TO_VEC_COST; |
598 | } | |
599 | ||
600 | if (vect_print_dump_info (REPORT_COST)) | |
601 | fprintf (vect_dump, "vect_model_simple_cost: inside_cost = %d, " | |
602 | "outside_cost = %d .", inside_cost, outside_cost); | |
603 | ||
604 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
605 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
606 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
607 | } | |
608 | ||
609 | ||
610 | /* Function vect_cost_strided_group_size | |
611 | ||
612 | For strided load or store, return the group_size only if it is the first | |
613 | load or store of a group, else return 1. This ensures that group size is | |
614 | only returned once per group. */ | |
615 | ||
616 | static int | |
617 | vect_cost_strided_group_size (stmt_vec_info stmt_info) | |
618 | { | |
619 | gimple first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
620 | ||
621 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
622 | return DR_GROUP_SIZE (stmt_info); | |
623 | ||
624 | return 1; | |
625 | } | |
626 | ||
627 | ||
628 | /* Function vect_model_store_cost | |
629 | ||
630 | Models cost for stores. In the case of strided accesses, one access | |
631 | has the overhead of the strided access attributed to it. */ | |
632 | ||
633 | void | |
634 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, | |
635 | enum vect_def_type dt, slp_tree slp_node) | |
636 | { | |
637 | int group_size; | |
638 | int inside_cost = 0, outside_cost = 0; | |
639 | ||
640 | /* The SLP costs were already calculated during SLP tree build. */ | |
641 | if (PURE_SLP_STMT (stmt_info)) | |
642 | return; | |
643 | ||
8644a673 | 644 | if (dt == vect_constant_def || dt == vect_external_def) |
ebfd146a IR |
645 | outside_cost = TARG_SCALAR_TO_VEC_COST; |
646 | ||
647 | /* Strided access? */ | |
648 | if (DR_GROUP_FIRST_DR (stmt_info) && !slp_node) | |
649 | group_size = vect_cost_strided_group_size (stmt_info); | |
650 | /* Not a strided access. */ | |
651 | else | |
652 | group_size = 1; | |
653 | ||
654 | /* Is this an access in a group of stores, which provide strided access? | |
655 | If so, add in the cost of the permutes. */ | |
656 | if (group_size > 1) | |
657 | { | |
658 | /* Uses a high and low interleave operation for each needed permute. */ | |
659 | inside_cost = ncopies * exact_log2(group_size) * group_size | |
660 | * TARG_VEC_STMT_COST; | |
661 | ||
662 | if (vect_print_dump_info (REPORT_COST)) | |
663 | fprintf (vect_dump, "vect_model_store_cost: strided group_size = %d .", | |
664 | group_size); | |
665 | ||
666 | } | |
667 | ||
668 | /* Costs of the stores. */ | |
669 | inside_cost += ncopies * TARG_VEC_STORE_COST; | |
670 | ||
671 | if (vect_print_dump_info (REPORT_COST)) | |
672 | fprintf (vect_dump, "vect_model_store_cost: inside_cost = %d, " | |
673 | "outside_cost = %d .", inside_cost, outside_cost); | |
674 | ||
675 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
676 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
677 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
678 | } | |
679 | ||
680 | ||
681 | /* Function vect_model_load_cost | |
682 | ||
683 | Models cost for loads. In the case of strided accesses, the last access | |
684 | has the overhead of the strided access attributed to it. Since unaligned | |
685 | accesses are supported for loads, we also account for the costs of the | |
686 | access scheme chosen. */ | |
687 | ||
688 | void | |
689 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, slp_tree slp_node) | |
690 | ||
691 | { | |
692 | int group_size; | |
693 | int alignment_support_cheme; | |
694 | gimple first_stmt; | |
695 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; | |
696 | int inside_cost = 0, outside_cost = 0; | |
697 | ||
698 | /* The SLP costs were already calculated during SLP tree build. */ | |
699 | if (PURE_SLP_STMT (stmt_info)) | |
700 | return; | |
701 | ||
702 | /* Strided accesses? */ | |
703 | first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
704 | if (first_stmt && !slp_node) | |
705 | { | |
706 | group_size = vect_cost_strided_group_size (stmt_info); | |
707 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
708 | } | |
709 | /* Not a strided access. */ | |
710 | else | |
711 | { | |
712 | group_size = 1; | |
713 | first_dr = dr; | |
714 | } | |
715 | ||
716 | alignment_support_cheme = vect_supportable_dr_alignment (first_dr); | |
717 | ||
718 | /* Is this an access in a group of loads providing strided access? | |
719 | If so, add in the cost of the permutes. */ | |
720 | if (group_size > 1) | |
721 | { | |
722 | /* Uses an even and odd extract operations for each needed permute. */ | |
723 | inside_cost = ncopies * exact_log2(group_size) * group_size | |
724 | * TARG_VEC_STMT_COST; | |
725 | ||
726 | if (vect_print_dump_info (REPORT_COST)) | |
727 | fprintf (vect_dump, "vect_model_load_cost: strided group_size = %d .", | |
728 | group_size); | |
729 | ||
730 | } | |
731 | ||
732 | /* The loads themselves. */ | |
733 | switch (alignment_support_cheme) | |
734 | { | |
735 | case dr_aligned: | |
736 | { | |
737 | inside_cost += ncopies * TARG_VEC_LOAD_COST; | |
738 | ||
739 | if (vect_print_dump_info (REPORT_COST)) | |
740 | fprintf (vect_dump, "vect_model_load_cost: aligned."); | |
741 | ||
742 | break; | |
743 | } | |
744 | case dr_unaligned_supported: | |
745 | { | |
746 | /* Here, we assign an additional cost for the unaligned load. */ | |
747 | inside_cost += ncopies * TARG_VEC_UNALIGNED_LOAD_COST; | |
748 | ||
749 | if (vect_print_dump_info (REPORT_COST)) | |
750 | fprintf (vect_dump, "vect_model_load_cost: unaligned supported by " | |
751 | "hardware."); | |
752 | ||
753 | break; | |
754 | } | |
755 | case dr_explicit_realign: | |
756 | { | |
757 | inside_cost += ncopies * (2*TARG_VEC_LOAD_COST + TARG_VEC_STMT_COST); | |
758 | ||
759 | /* FIXME: If the misalignment remains fixed across the iterations of | |
760 | the containing loop, the following cost should be added to the | |
761 | outside costs. */ | |
762 | if (targetm.vectorize.builtin_mask_for_load) | |
763 | inside_cost += TARG_VEC_STMT_COST; | |
764 | ||
765 | break; | |
766 | } | |
767 | case dr_explicit_realign_optimized: | |
768 | { | |
769 | if (vect_print_dump_info (REPORT_COST)) | |
770 | fprintf (vect_dump, "vect_model_load_cost: unaligned software " | |
771 | "pipelined."); | |
772 | ||
773 | /* Unaligned software pipeline has a load of an address, an initial | |
774 | load, and possibly a mask operation to "prime" the loop. However, | |
775 | if this is an access in a group of loads, which provide strided | |
776 | access, then the above cost should only be considered for one | |
777 | access in the group. Inside the loop, there is a load op | |
778 | and a realignment op. */ | |
779 | ||
780 | if ((!DR_GROUP_FIRST_DR (stmt_info)) || group_size > 1 || slp_node) | |
781 | { | |
782 | outside_cost = 2*TARG_VEC_STMT_COST; | |
783 | if (targetm.vectorize.builtin_mask_for_load) | |
784 | outside_cost += TARG_VEC_STMT_COST; | |
785 | } | |
786 | ||
787 | inside_cost += ncopies * (TARG_VEC_LOAD_COST + TARG_VEC_STMT_COST); | |
788 | ||
789 | break; | |
790 | } | |
791 | ||
792 | default: | |
793 | gcc_unreachable (); | |
794 | } | |
795 | ||
796 | if (vect_print_dump_info (REPORT_COST)) | |
797 | fprintf (vect_dump, "vect_model_load_cost: inside_cost = %d, " | |
798 | "outside_cost = %d .", inside_cost, outside_cost); | |
799 | ||
800 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
801 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
802 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
803 | } | |
804 | ||
805 | ||
806 | /* Function vect_init_vector. | |
807 | ||
808 | Insert a new stmt (INIT_STMT) that initializes a new vector variable with | |
809 | the vector elements of VECTOR_VAR. Place the initialization at BSI if it | |
810 | is not NULL. Otherwise, place the initialization at the loop preheader. | |
811 | Return the DEF of INIT_STMT. | |
812 | It will be used in the vectorization of STMT. */ | |
813 | ||
814 | tree | |
815 | vect_init_vector (gimple stmt, tree vector_var, tree vector_type, | |
816 | gimple_stmt_iterator *gsi) | |
817 | { | |
818 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
819 | tree new_var; | |
820 | gimple init_stmt; | |
821 | tree vec_oprnd; | |
822 | edge pe; | |
823 | tree new_temp; | |
824 | basic_block new_bb; | |
825 | ||
826 | new_var = vect_get_new_vect_var (vector_type, vect_simple_var, "cst_"); | |
827 | add_referenced_var (new_var); | |
828 | init_stmt = gimple_build_assign (new_var, vector_var); | |
829 | new_temp = make_ssa_name (new_var, init_stmt); | |
830 | gimple_assign_set_lhs (init_stmt, new_temp); | |
831 | ||
832 | if (gsi) | |
833 | vect_finish_stmt_generation (stmt, init_stmt, gsi); | |
834 | else | |
835 | { | |
836 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
a70d6342 IR |
837 | |
838 | if (loop_vinfo) | |
839 | { | |
840 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
841 | ||
842 | if (nested_in_vect_loop_p (loop, stmt)) | |
843 | loop = loop->inner; | |
844 | ||
845 | pe = loop_preheader_edge (loop); | |
846 | new_bb = gsi_insert_on_edge_immediate (pe, init_stmt); | |
847 | gcc_assert (!new_bb); | |
848 | } | |
849 | else | |
850 | { | |
851 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
852 | basic_block bb; | |
853 | gimple_stmt_iterator gsi_bb_start; | |
854 | ||
855 | gcc_assert (bb_vinfo); | |
856 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 857 | gsi_bb_start = gsi_after_labels (bb); |
a70d6342 IR |
858 | gsi_insert_before (&gsi_bb_start, init_stmt, GSI_SAME_STMT); |
859 | } | |
ebfd146a IR |
860 | } |
861 | ||
862 | if (vect_print_dump_info (REPORT_DETAILS)) | |
863 | { | |
864 | fprintf (vect_dump, "created new init_stmt: "); | |
865 | print_gimple_stmt (vect_dump, init_stmt, 0, TDF_SLIM); | |
866 | } | |
867 | ||
868 | vec_oprnd = gimple_assign_lhs (init_stmt); | |
869 | return vec_oprnd; | |
870 | } | |
871 | ||
a70d6342 | 872 | |
ebfd146a IR |
873 | /* Function vect_get_vec_def_for_operand. |
874 | ||
875 | OP is an operand in STMT. This function returns a (vector) def that will be | |
876 | used in the vectorized stmt for STMT. | |
877 | ||
878 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
879 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
880 | ||
881 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
882 | needs to be introduced. */ | |
883 | ||
884 | tree | |
885 | vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def) | |
886 | { | |
887 | tree vec_oprnd; | |
888 | gimple vec_stmt; | |
889 | gimple def_stmt; | |
890 | stmt_vec_info def_stmt_info = NULL; | |
891 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
892 | tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
893 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
894 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
895 | tree vec_inv; | |
896 | tree vec_cst; | |
897 | tree t = NULL_TREE; | |
898 | tree def; | |
899 | int i; | |
900 | enum vect_def_type dt; | |
901 | bool is_simple_use; | |
902 | tree vector_type; | |
903 | ||
904 | if (vect_print_dump_info (REPORT_DETAILS)) | |
905 | { | |
906 | fprintf (vect_dump, "vect_get_vec_def_for_operand: "); | |
907 | print_generic_expr (vect_dump, op, TDF_SLIM); | |
908 | } | |
909 | ||
a70d6342 IR |
910 | is_simple_use = vect_is_simple_use (op, loop_vinfo, NULL, &def_stmt, &def, |
911 | &dt); | |
ebfd146a IR |
912 | gcc_assert (is_simple_use); |
913 | if (vect_print_dump_info (REPORT_DETAILS)) | |
914 | { | |
915 | if (def) | |
916 | { | |
917 | fprintf (vect_dump, "def = "); | |
918 | print_generic_expr (vect_dump, def, TDF_SLIM); | |
919 | } | |
920 | if (def_stmt) | |
921 | { | |
922 | fprintf (vect_dump, " def_stmt = "); | |
923 | print_gimple_stmt (vect_dump, def_stmt, 0, TDF_SLIM); | |
924 | } | |
925 | } | |
926 | ||
927 | switch (dt) | |
928 | { | |
929 | /* Case 1: operand is a constant. */ | |
930 | case vect_constant_def: | |
931 | { | |
7569a6cc RG |
932 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
933 | gcc_assert (vector_type); | |
934 | ||
ebfd146a IR |
935 | if (scalar_def) |
936 | *scalar_def = op; | |
937 | ||
938 | /* Create 'vect_cst_ = {cst,cst,...,cst}' */ | |
939 | if (vect_print_dump_info (REPORT_DETAILS)) | |
940 | fprintf (vect_dump, "Create vector_cst. nunits = %d", nunits); | |
941 | ||
942 | for (i = nunits - 1; i >= 0; --i) | |
943 | { | |
944 | t = tree_cons (NULL_TREE, op, t); | |
945 | } | |
7569a6cc RG |
946 | vec_cst = build_vector (vector_type, t); |
947 | return vect_init_vector (stmt, vec_cst, vector_type, NULL); | |
ebfd146a IR |
948 | } |
949 | ||
950 | /* Case 2: operand is defined outside the loop - loop invariant. */ | |
8644a673 | 951 | case vect_external_def: |
ebfd146a IR |
952 | { |
953 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); | |
954 | gcc_assert (vector_type); | |
955 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); | |
956 | ||
957 | if (scalar_def) | |
958 | *scalar_def = def; | |
959 | ||
960 | /* Create 'vec_inv = {inv,inv,..,inv}' */ | |
961 | if (vect_print_dump_info (REPORT_DETAILS)) | |
962 | fprintf (vect_dump, "Create vector_inv."); | |
963 | ||
964 | for (i = nunits - 1; i >= 0; --i) | |
965 | { | |
966 | t = tree_cons (NULL_TREE, def, t); | |
967 | } | |
968 | ||
969 | /* FIXME: use build_constructor directly. */ | |
970 | vec_inv = build_constructor_from_list (vector_type, t); | |
971 | return vect_init_vector (stmt, vec_inv, vector_type, NULL); | |
972 | } | |
973 | ||
974 | /* Case 3: operand is defined inside the loop. */ | |
8644a673 | 975 | case vect_internal_def: |
ebfd146a IR |
976 | { |
977 | if (scalar_def) | |
978 | *scalar_def = NULL/* FIXME tuples: def_stmt*/; | |
979 | ||
980 | /* Get the def from the vectorized stmt. */ | |
981 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
982 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
983 | gcc_assert (vec_stmt); | |
984 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
985 | vec_oprnd = PHI_RESULT (vec_stmt); | |
986 | else if (is_gimple_call (vec_stmt)) | |
987 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
988 | else | |
989 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
990 | return vec_oprnd; | |
991 | } | |
992 | ||
993 | /* Case 4: operand is defined by a loop header phi - reduction */ | |
994 | case vect_reduction_def: | |
06066f92 | 995 | case vect_double_reduction_def: |
7c5222ff | 996 | case vect_nested_cycle: |
ebfd146a IR |
997 | { |
998 | struct loop *loop; | |
999 | ||
1000 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1001 | loop = (gimple_bb (def_stmt))->loop_father; | |
1002 | ||
1003 | /* Get the def before the loop */ | |
1004 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); | |
1005 | return get_initial_def_for_reduction (stmt, op, scalar_def); | |
1006 | } | |
1007 | ||
1008 | /* Case 5: operand is defined by loop-header phi - induction. */ | |
1009 | case vect_induction_def: | |
1010 | { | |
1011 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1012 | ||
1013 | /* Get the def from the vectorized stmt. */ | |
1014 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1015 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
1016 | gcc_assert (vec_stmt && gimple_code (vec_stmt) == GIMPLE_PHI); | |
1017 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1018 | return vec_oprnd; | |
1019 | } | |
1020 | ||
1021 | default: | |
1022 | gcc_unreachable (); | |
1023 | } | |
1024 | } | |
1025 | ||
1026 | ||
1027 | /* Function vect_get_vec_def_for_stmt_copy | |
1028 | ||
1029 | Return a vector-def for an operand. This function is used when the | |
1030 | vectorized stmt to be created (by the caller to this function) is a "copy" | |
1031 | created in case the vectorized result cannot fit in one vector, and several | |
1032 | copies of the vector-stmt are required. In this case the vector-def is | |
1033 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field | |
1034 | of the stmt that defines VEC_OPRND. | |
1035 | DT is the type of the vector def VEC_OPRND. | |
1036 | ||
1037 | Context: | |
1038 | In case the vectorization factor (VF) is bigger than the number | |
1039 | of elements that can fit in a vectype (nunits), we have to generate | |
1040 | more than one vector stmt to vectorize the scalar stmt. This situation | |
1041 | arises when there are multiple data-types operated upon in the loop; the | |
1042 | smallest data-type determines the VF, and as a result, when vectorizing | |
1043 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1044 | vector stmt (each computing a vector of 'nunits' results, and together | |
1045 | computing 'VF' results in each iteration). This function is called when | |
1046 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in | |
1047 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1048 | ||
1049 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
1050 | ||
1051 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 | |
1052 | VS1.1: vx.1 = memref1 VS1.2 | |
1053 | VS1.2: vx.2 = memref2 VS1.3 | |
1054 | VS1.3: vx.3 = memref3 | |
1055 | ||
1056 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1057 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1058 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1059 | VSnew.3: vz3 = vx.3 + ... | |
1060 | ||
1061 | The vectorization of S1 is explained in vectorizable_load. | |
1062 | The vectorization of S2: | |
1063 | To create the first vector-stmt out of the 4 copies - VSnew.0 - | |
1064 | the function 'vect_get_vec_def_for_operand' is called to | |
1065 | get the relevant vector-def for each operand of S2. For operand x it | |
1066 | returns the vector-def 'vx.0'. | |
1067 | ||
1068 | To create the remaining copies of the vector-stmt (VSnew.j), this | |
1069 | function is called to get the relevant vector-def for each operand. It is | |
1070 | obtained from the respective VS1.j stmt, which is recorded in the | |
1071 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. | |
1072 | ||
1073 | For example, to obtain the vector-def 'vx.1' in order to create the | |
1074 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1075 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
1076 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', | |
1077 | and return its def ('vx.1'). | |
1078 | Overall, to create the above sequence this function will be called 3 times: | |
1079 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1080 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1081 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1082 | ||
1083 | tree | |
1084 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1085 | { | |
1086 | gimple vec_stmt_for_operand; | |
1087 | stmt_vec_info def_stmt_info; | |
1088 | ||
1089 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1090 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1091 | return vec_oprnd; |
1092 | ||
1093 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1094 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1095 | gcc_assert (def_stmt_info); | |
1096 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1097 | gcc_assert (vec_stmt_for_operand); | |
1098 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1099 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) | |
1100 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1101 | else | |
1102 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1103 | return vec_oprnd; | |
1104 | } | |
1105 | ||
1106 | ||
1107 | /* Get vectorized definitions for the operands to create a copy of an original | |
1108 | stmt. See vect_get_vec_def_for_stmt_copy() for details. */ | |
1109 | ||
1110 | static void | |
1111 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, | |
1112 | VEC(tree,heap) **vec_oprnds0, | |
1113 | VEC(tree,heap) **vec_oprnds1) | |
1114 | { | |
1115 | tree vec_oprnd = VEC_pop (tree, *vec_oprnds0); | |
1116 | ||
1117 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
1118 | VEC_quick_push (tree, *vec_oprnds0, vec_oprnd); | |
1119 | ||
1120 | if (vec_oprnds1 && *vec_oprnds1) | |
1121 | { | |
1122 | vec_oprnd = VEC_pop (tree, *vec_oprnds1); | |
1123 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); | |
1124 | VEC_quick_push (tree, *vec_oprnds1, vec_oprnd); | |
1125 | } | |
1126 | } | |
1127 | ||
1128 | ||
1129 | /* Get vectorized definitions for OP0 and OP1, or SLP_NODE if it is not NULL. */ | |
1130 | ||
1131 | static void | |
1132 | vect_get_vec_defs (tree op0, tree op1, gimple stmt, | |
1133 | VEC(tree,heap) **vec_oprnds0, VEC(tree,heap) **vec_oprnds1, | |
1134 | slp_tree slp_node) | |
1135 | { | |
1136 | if (slp_node) | |
1137 | vect_get_slp_defs (slp_node, vec_oprnds0, vec_oprnds1); | |
1138 | else | |
1139 | { | |
1140 | tree vec_oprnd; | |
1141 | ||
1142 | *vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
1143 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
1144 | VEC_quick_push (tree, *vec_oprnds0, vec_oprnd); | |
1145 | ||
1146 | if (op1) | |
1147 | { | |
1148 | *vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
1149 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL); | |
1150 | VEC_quick_push (tree, *vec_oprnds1, vec_oprnd); | |
1151 | } | |
1152 | } | |
1153 | } | |
1154 | ||
1155 | ||
1156 | /* Function vect_finish_stmt_generation. | |
1157 | ||
1158 | Insert a new stmt. */ | |
1159 | ||
1160 | void | |
1161 | vect_finish_stmt_generation (gimple stmt, gimple vec_stmt, | |
1162 | gimple_stmt_iterator *gsi) | |
1163 | { | |
1164 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1165 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1166 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1167 | |
1168 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1169 | ||
1170 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); | |
1171 | ||
a70d6342 IR |
1172 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo, |
1173 | bb_vinfo)); | |
ebfd146a IR |
1174 | |
1175 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1176 | { | |
1177 | fprintf (vect_dump, "add new stmt: "); | |
1178 | print_gimple_stmt (vect_dump, vec_stmt, 0, TDF_SLIM); | |
1179 | } | |
1180 | ||
1181 | gimple_set_location (vec_stmt, gimple_location (gsi_stmt (*gsi))); | |
1182 | } | |
1183 | ||
1184 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1185 | a function declaration if the target has a vectorized version | |
1186 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1187 | ||
1188 | tree | |
1189 | vectorizable_function (gimple call, tree vectype_out, tree vectype_in) | |
1190 | { | |
1191 | tree fndecl = gimple_call_fndecl (call); | |
1192 | enum built_in_function code; | |
1193 | ||
1194 | /* We only handle functions that do not read or clobber memory -- i.e. | |
1195 | const or novops ones. */ | |
1196 | if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS))) | |
1197 | return NULL_TREE; | |
1198 | ||
1199 | if (!fndecl | |
1200 | || TREE_CODE (fndecl) != FUNCTION_DECL | |
1201 | || !DECL_BUILT_IN (fndecl)) | |
1202 | return NULL_TREE; | |
1203 | ||
1204 | code = DECL_FUNCTION_CODE (fndecl); | |
1205 | return targetm.vectorize.builtin_vectorized_function (code, vectype_out, | |
1206 | vectype_in); | |
1207 | } | |
1208 | ||
1209 | /* Function vectorizable_call. | |
1210 | ||
1211 | Check if STMT performs a function call that can be vectorized. | |
1212 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1213 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
1214 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1215 | ||
1216 | static bool | |
1217 | vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt) | |
1218 | { | |
1219 | tree vec_dest; | |
1220 | tree scalar_dest; | |
1221 | tree op, type; | |
1222 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1223 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
1224 | tree vectype_out, vectype_in; | |
1225 | int nunits_in; | |
1226 | int nunits_out; | |
1227 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1228 | tree fndecl, new_temp, def, rhs_type, lhs_type; | |
1229 | gimple def_stmt; | |
1230 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
63827fb8 | 1231 | gimple new_stmt = NULL; |
ebfd146a IR |
1232 | int ncopies, j; |
1233 | VEC(tree, heap) *vargs = NULL; | |
1234 | enum { NARROW, NONE, WIDEN } modifier; | |
1235 | size_t i, nargs; | |
1236 | ||
a70d6342 IR |
1237 | /* FORNOW: unsupported in basic block SLP. */ |
1238 | gcc_assert (loop_vinfo); | |
1239 | ||
ebfd146a IR |
1240 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
1241 | return false; | |
1242 | ||
8644a673 | 1243 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1244 | return false; |
1245 | ||
1246 | /* FORNOW: SLP not supported. */ | |
1247 | if (STMT_SLP_TYPE (stmt_info)) | |
1248 | return false; | |
1249 | ||
1250 | /* Is STMT a vectorizable call? */ | |
1251 | if (!is_gimple_call (stmt)) | |
1252 | return false; | |
1253 | ||
1254 | if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
1255 | return false; | |
1256 | ||
1257 | /* Process function arguments. */ | |
1258 | rhs_type = NULL_TREE; | |
1259 | nargs = gimple_call_num_args (stmt); | |
1260 | ||
1261 | /* Bail out if the function has more than two arguments, we | |
1262 | do not have interesting builtin functions to vectorize with | |
1263 | more than two arguments. No arguments is also not good. */ | |
1264 | if (nargs == 0 || nargs > 2) | |
1265 | return false; | |
1266 | ||
1267 | for (i = 0; i < nargs; i++) | |
1268 | { | |
1269 | op = gimple_call_arg (stmt, i); | |
1270 | ||
1271 | /* We can only handle calls with arguments of the same type. */ | |
1272 | if (rhs_type | |
1273 | && rhs_type != TREE_TYPE (op)) | |
1274 | { | |
1275 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1276 | fprintf (vect_dump, "argument types differ."); | |
1277 | return false; | |
1278 | } | |
1279 | rhs_type = TREE_TYPE (op); | |
1280 | ||
a70d6342 | 1281 | if (!vect_is_simple_use (op, loop_vinfo, NULL, &def_stmt, &def, &dt[i])) |
ebfd146a IR |
1282 | { |
1283 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1284 | fprintf (vect_dump, "use not simple."); | |
1285 | return false; | |
1286 | } | |
1287 | } | |
1288 | ||
1289 | vectype_in = get_vectype_for_scalar_type (rhs_type); | |
1290 | if (!vectype_in) | |
1291 | return false; | |
1292 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); | |
1293 | ||
1294 | lhs_type = TREE_TYPE (gimple_call_lhs (stmt)); | |
1295 | vectype_out = get_vectype_for_scalar_type (lhs_type); | |
1296 | if (!vectype_out) | |
1297 | return false; | |
1298 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
1299 | ||
1300 | /* FORNOW */ | |
1301 | if (nunits_in == nunits_out / 2) | |
1302 | modifier = NARROW; | |
1303 | else if (nunits_out == nunits_in) | |
1304 | modifier = NONE; | |
1305 | else if (nunits_out == nunits_in / 2) | |
1306 | modifier = WIDEN; | |
1307 | else | |
1308 | return false; | |
1309 | ||
1310 | /* For now, we only vectorize functions if a target specific builtin | |
1311 | is available. TODO -- in some cases, it might be profitable to | |
1312 | insert the calls for pieces of the vector, in order to be able | |
1313 | to vectorize other operations in the loop. */ | |
1314 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
1315 | if (fndecl == NULL_TREE) | |
1316 | { | |
1317 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1318 | fprintf (vect_dump, "function is not vectorizable."); | |
1319 | ||
1320 | return false; | |
1321 | } | |
1322 | ||
5006671f | 1323 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a IR |
1324 | |
1325 | if (modifier == NARROW) | |
1326 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
1327 | else | |
1328 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1329 | ||
1330 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
1331 | needs to be generated. */ | |
1332 | gcc_assert (ncopies >= 1); | |
1333 | ||
1334 | if (!vec_stmt) /* transformation not required. */ | |
1335 | { | |
1336 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1337 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1338 | fprintf (vect_dump, "=== vectorizable_call ==="); | |
1339 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
1340 | return true; | |
1341 | } | |
1342 | ||
1343 | /** Transform. **/ | |
1344 | ||
1345 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1346 | fprintf (vect_dump, "transform operation."); | |
1347 | ||
1348 | /* Handle def. */ | |
1349 | scalar_dest = gimple_call_lhs (stmt); | |
1350 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1351 | ||
1352 | prev_stmt_info = NULL; | |
1353 | switch (modifier) | |
1354 | { | |
1355 | case NONE: | |
1356 | for (j = 0; j < ncopies; ++j) | |
1357 | { | |
1358 | /* Build argument list for the vectorized call. */ | |
1359 | if (j == 0) | |
1360 | vargs = VEC_alloc (tree, heap, nargs); | |
1361 | else | |
1362 | VEC_truncate (tree, vargs, 0); | |
1363 | ||
1364 | for (i = 0; i < nargs; i++) | |
1365 | { | |
1366 | op = gimple_call_arg (stmt, i); | |
1367 | if (j == 0) | |
1368 | vec_oprnd0 | |
1369 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1370 | else | |
63827fb8 IR |
1371 | { |
1372 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
1373 | vec_oprnd0 | |
1374 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
1375 | } | |
ebfd146a IR |
1376 | |
1377 | VEC_quick_push (tree, vargs, vec_oprnd0); | |
1378 | } | |
1379 | ||
1380 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1381 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1382 | gimple_call_set_lhs (new_stmt, new_temp); | |
1383 | ||
1384 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7411b8f0 | 1385 | mark_symbols_for_renaming (new_stmt); |
ebfd146a IR |
1386 | |
1387 | if (j == 0) | |
1388 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1389 | else | |
1390 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1391 | ||
1392 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1393 | } | |
1394 | ||
1395 | break; | |
1396 | ||
1397 | case NARROW: | |
1398 | for (j = 0; j < ncopies; ++j) | |
1399 | { | |
1400 | /* Build argument list for the vectorized call. */ | |
1401 | if (j == 0) | |
1402 | vargs = VEC_alloc (tree, heap, nargs * 2); | |
1403 | else | |
1404 | VEC_truncate (tree, vargs, 0); | |
1405 | ||
1406 | for (i = 0; i < nargs; i++) | |
1407 | { | |
1408 | op = gimple_call_arg (stmt, i); | |
1409 | if (j == 0) | |
1410 | { | |
1411 | vec_oprnd0 | |
1412 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1413 | vec_oprnd1 | |
63827fb8 | 1414 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
1415 | } |
1416 | else | |
1417 | { | |
63827fb8 | 1418 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i); |
ebfd146a | 1419 | vec_oprnd0 |
63827fb8 | 1420 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 1421 | vec_oprnd1 |
63827fb8 | 1422 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
1423 | } |
1424 | ||
1425 | VEC_quick_push (tree, vargs, vec_oprnd0); | |
1426 | VEC_quick_push (tree, vargs, vec_oprnd1); | |
1427 | } | |
1428 | ||
1429 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1430 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1431 | gimple_call_set_lhs (new_stmt, new_temp); | |
1432 | ||
1433 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7411b8f0 | 1434 | mark_symbols_for_renaming (new_stmt); |
ebfd146a IR |
1435 | |
1436 | if (j == 0) | |
1437 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1438 | else | |
1439 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1440 | ||
1441 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1442 | } | |
1443 | ||
1444 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
1445 | ||
1446 | break; | |
1447 | ||
1448 | case WIDEN: | |
1449 | /* No current target implements this case. */ | |
1450 | return false; | |
1451 | } | |
1452 | ||
1453 | VEC_free (tree, heap, vargs); | |
1454 | ||
1455 | /* Update the exception handling table with the vector stmt if necessary. */ | |
1456 | if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt)) | |
1457 | gimple_purge_dead_eh_edges (gimple_bb (stmt)); | |
1458 | ||
1459 | /* The call in STMT might prevent it from being removed in dce. | |
1460 | We however cannot remove it here, due to the way the ssa name | |
1461 | it defines is mapped to the new definition. So just replace | |
1462 | rhs of the statement with something harmless. */ | |
1463 | ||
1464 | type = TREE_TYPE (scalar_dest); | |
1465 | new_stmt = gimple_build_assign (gimple_call_lhs (stmt), | |
1466 | fold_convert (type, integer_zero_node)); | |
1467 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
1468 | set_vinfo_for_stmt (stmt, NULL); | |
1469 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
1470 | gsi_replace (gsi, new_stmt, false); | |
1471 | SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt; | |
1472 | ||
1473 | return true; | |
1474 | } | |
1475 | ||
1476 | ||
1477 | /* Function vect_gen_widened_results_half | |
1478 | ||
1479 | Create a vector stmt whose code, type, number of arguments, and result | |
1480 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are | |
1481 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. | |
1482 | In the case that CODE is a CALL_EXPR, this means that a call to DECL | |
1483 | needs to be created (DECL is a function-decl of a target-builtin). | |
1484 | STMT is the original scalar stmt that we are vectorizing. */ | |
1485 | ||
1486 | static gimple | |
1487 | vect_gen_widened_results_half (enum tree_code code, | |
1488 | tree decl, | |
1489 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
1490 | tree vec_dest, gimple_stmt_iterator *gsi, | |
1491 | gimple stmt) | |
1492 | { | |
1493 | gimple new_stmt; | |
1494 | tree new_temp; | |
ebfd146a IR |
1495 | |
1496 | /* Generate half of the widened result: */ | |
1497 | if (code == CALL_EXPR) | |
1498 | { | |
1499 | /* Target specific support */ | |
1500 | if (op_type == binary_op) | |
1501 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
1502 | else | |
1503 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
1504 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1505 | gimple_call_set_lhs (new_stmt, new_temp); | |
1506 | } | |
1507 | else | |
1508 | { | |
1509 | /* Generic support */ | |
1510 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
1511 | if (op_type != binary_op) | |
1512 | vec_oprnd1 = NULL; | |
1513 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0, | |
1514 | vec_oprnd1); | |
1515 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1516 | gimple_assign_set_lhs (new_stmt, new_temp); | |
1517 | } | |
1518 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1519 | ||
ebfd146a IR |
1520 | return new_stmt; |
1521 | } | |
1522 | ||
1523 | ||
1524 | /* Check if STMT performs a conversion operation, that can be vectorized. | |
1525 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1526 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
1527 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1528 | ||
1529 | static bool | |
1530 | vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi, | |
1531 | gimple *vec_stmt, slp_tree slp_node) | |
1532 | { | |
1533 | tree vec_dest; | |
1534 | tree scalar_dest; | |
1535 | tree op0; | |
1536 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1537 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1538 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1539 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
1540 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; | |
1541 | tree new_temp; | |
1542 | tree def; | |
1543 | gimple def_stmt; | |
1544 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
1545 | gimple new_stmt = NULL; | |
1546 | stmt_vec_info prev_stmt_info; | |
1547 | int nunits_in; | |
1548 | int nunits_out; | |
1549 | tree vectype_out, vectype_in; | |
1550 | int ncopies, j; | |
1551 | tree expr; | |
1552 | tree rhs_type, lhs_type; | |
1553 | tree builtin_decl; | |
1554 | enum { NARROW, NONE, WIDEN } modifier; | |
1555 | int i; | |
1556 | VEC(tree,heap) *vec_oprnds0 = NULL; | |
1557 | tree vop0; | |
1558 | tree integral_type; | |
1559 | VEC(tree,heap) *dummy = NULL; | |
1560 | int dummy_int; | |
1561 | ||
1562 | /* Is STMT a vectorizable conversion? */ | |
1563 | ||
a70d6342 IR |
1564 | /* FORNOW: unsupported in basic block SLP. */ |
1565 | gcc_assert (loop_vinfo); | |
1566 | ||
ebfd146a IR |
1567 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
1568 | return false; | |
1569 | ||
8644a673 | 1570 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1571 | return false; |
1572 | ||
1573 | if (!is_gimple_assign (stmt)) | |
1574 | return false; | |
1575 | ||
1576 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
1577 | return false; | |
1578 | ||
1579 | code = gimple_assign_rhs_code (stmt); | |
1580 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
1581 | return false; | |
1582 | ||
1583 | /* Check types of lhs and rhs. */ | |
1584 | op0 = gimple_assign_rhs1 (stmt); | |
1585 | rhs_type = TREE_TYPE (op0); | |
1586 | vectype_in = get_vectype_for_scalar_type (rhs_type); | |
1587 | if (!vectype_in) | |
1588 | return false; | |
1589 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); | |
1590 | ||
1591 | scalar_dest = gimple_assign_lhs (stmt); | |
1592 | lhs_type = TREE_TYPE (scalar_dest); | |
1593 | vectype_out = get_vectype_for_scalar_type (lhs_type); | |
1594 | if (!vectype_out) | |
1595 | return false; | |
1596 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
1597 | ||
1598 | /* FORNOW */ | |
1599 | if (nunits_in == nunits_out / 2) | |
1600 | modifier = NARROW; | |
1601 | else if (nunits_out == nunits_in) | |
1602 | modifier = NONE; | |
1603 | else if (nunits_out == nunits_in / 2) | |
1604 | modifier = WIDEN; | |
1605 | else | |
1606 | return false; | |
1607 | ||
1608 | if (modifier == NONE) | |
1609 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) == vectype_out); | |
1610 | ||
1611 | /* Bail out if the types are both integral or non-integral. */ | |
1612 | if ((INTEGRAL_TYPE_P (rhs_type) && INTEGRAL_TYPE_P (lhs_type)) | |
1613 | || (!INTEGRAL_TYPE_P (rhs_type) && !INTEGRAL_TYPE_P (lhs_type))) | |
1614 | return false; | |
1615 | ||
1616 | integral_type = INTEGRAL_TYPE_P (rhs_type) ? vectype_in : vectype_out; | |
1617 | ||
1618 | if (modifier == NARROW) | |
1619 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
1620 | else | |
1621 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1622 | ||
1623 | /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies | |
1624 | this, so we can safely override NCOPIES with 1 here. */ | |
1625 | if (slp_node) | |
1626 | ncopies = 1; | |
1627 | ||
1628 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
1629 | needs to be generated. */ | |
1630 | gcc_assert (ncopies >= 1); | |
1631 | ||
1632 | /* Check the operands of the operation. */ | |
a70d6342 | 1633 | if (!vect_is_simple_use (op0, loop_vinfo, NULL, &def_stmt, &def, &dt[0])) |
ebfd146a IR |
1634 | { |
1635 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1636 | fprintf (vect_dump, "use not simple."); | |
1637 | return false; | |
1638 | } | |
1639 | ||
1640 | /* Supportable by target? */ | |
1641 | if ((modifier == NONE | |
1642 | && !targetm.vectorize.builtin_conversion (code, integral_type)) | |
1643 | || (modifier == WIDEN | |
1644 | && !supportable_widening_operation (code, stmt, vectype_in, | |
1645 | &decl1, &decl2, | |
1646 | &code1, &code2, | |
1647 | &dummy_int, &dummy)) | |
1648 | || (modifier == NARROW | |
1649 | && !supportable_narrowing_operation (code, stmt, vectype_in, | |
1650 | &code1, &dummy_int, &dummy))) | |
1651 | { | |
1652 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1653 | fprintf (vect_dump, "conversion not supported by target."); | |
1654 | return false; | |
1655 | } | |
1656 | ||
1657 | if (modifier != NONE) | |
1658 | { | |
1659 | STMT_VINFO_VECTYPE (stmt_info) = vectype_in; | |
1660 | /* FORNOW: SLP not supported. */ | |
1661 | if (STMT_SLP_TYPE (stmt_info)) | |
1662 | return false; | |
1663 | } | |
1664 | ||
1665 | if (!vec_stmt) /* transformation not required. */ | |
1666 | { | |
1667 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
1668 | return true; | |
1669 | } | |
1670 | ||
1671 | /** Transform. **/ | |
1672 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1673 | fprintf (vect_dump, "transform conversion."); | |
1674 | ||
1675 | /* Handle def. */ | |
1676 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1677 | ||
1678 | if (modifier == NONE && !slp_node) | |
1679 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
1680 | ||
1681 | prev_stmt_info = NULL; | |
1682 | switch (modifier) | |
1683 | { | |
1684 | case NONE: | |
1685 | for (j = 0; j < ncopies; j++) | |
1686 | { | |
ebfd146a IR |
1687 | if (j == 0) |
1688 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node); | |
1689 | else | |
1690 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
1691 | ||
1692 | builtin_decl = | |
1693 | targetm.vectorize.builtin_conversion (code, integral_type); | |
1694 | for (i = 0; VEC_iterate (tree, vec_oprnds0, i, vop0); i++) | |
1695 | { | |
1696 | /* Arguments are ready. create the new vector stmt. */ | |
1697 | new_stmt = gimple_build_call (builtin_decl, 1, vop0); | |
1698 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1699 | gimple_call_set_lhs (new_stmt, new_temp); | |
1700 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a IR |
1701 | if (slp_node) |
1702 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
1703 | } | |
1704 | ||
1705 | if (j == 0) | |
1706 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1707 | else | |
1708 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1709 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1710 | } | |
1711 | break; | |
1712 | ||
1713 | case WIDEN: | |
1714 | /* In case the vectorization factor (VF) is bigger than the number | |
1715 | of elements that we can fit in a vectype (nunits), we have to | |
1716 | generate more than one vector stmt - i.e - we need to "unroll" | |
1717 | the vector stmt by a factor VF/nunits. */ | |
1718 | for (j = 0; j < ncopies; j++) | |
1719 | { | |
1720 | if (j == 0) | |
1721 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
1722 | else | |
1723 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1724 | ||
1725 | STMT_VINFO_VECTYPE (stmt_info) = vectype_in; | |
1726 | ||
1727 | /* Generate first half of the widened result: */ | |
1728 | new_stmt | |
1729 | = vect_gen_widened_results_half (code1, decl1, | |
1730 | vec_oprnd0, vec_oprnd1, | |
1731 | unary_op, vec_dest, gsi, stmt); | |
1732 | if (j == 0) | |
1733 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1734 | else | |
1735 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1736 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1737 | ||
1738 | /* Generate second half of the widened result: */ | |
1739 | new_stmt | |
1740 | = vect_gen_widened_results_half (code2, decl2, | |
1741 | vec_oprnd0, vec_oprnd1, | |
1742 | unary_op, vec_dest, gsi, stmt); | |
1743 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1744 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1745 | } | |
1746 | break; | |
1747 | ||
1748 | case NARROW: | |
1749 | /* In case the vectorization factor (VF) is bigger than the number | |
1750 | of elements that we can fit in a vectype (nunits), we have to | |
1751 | generate more than one vector stmt - i.e - we need to "unroll" | |
1752 | the vector stmt by a factor VF/nunits. */ | |
1753 | for (j = 0; j < ncopies; j++) | |
1754 | { | |
1755 | /* Handle uses. */ | |
1756 | if (j == 0) | |
1757 | { | |
1758 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
1759 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1760 | } | |
1761 | else | |
1762 | { | |
1763 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd1); | |
1764 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1765 | } | |
1766 | ||
1767 | /* Arguments are ready. Create the new vector stmt. */ | |
1768 | expr = build2 (code1, vectype_out, vec_oprnd0, vec_oprnd1); | |
1769 | new_stmt = gimple_build_assign_with_ops (code1, vec_dest, vec_oprnd0, | |
1770 | vec_oprnd1); | |
1771 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1772 | gimple_assign_set_lhs (new_stmt, new_temp); | |
1773 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1774 | ||
1775 | if (j == 0) | |
1776 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1777 | else | |
1778 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1779 | ||
1780 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1781 | } | |
1782 | ||
1783 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
1784 | } | |
1785 | ||
1786 | if (vec_oprnds0) | |
1787 | VEC_free (tree, heap, vec_oprnds0); | |
1788 | ||
1789 | return true; | |
1790 | } | |
1791 | /* Function vectorizable_assignment. | |
1792 | ||
1793 | Check if STMT performs an assignment (copy) that can be vectorized. | |
1794 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1795 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
1796 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1797 | ||
1798 | static bool | |
1799 | vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi, | |
1800 | gimple *vec_stmt, slp_tree slp_node) | |
1801 | { | |
1802 | tree vec_dest; | |
1803 | tree scalar_dest; | |
1804 | tree op; | |
1805 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1806 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
1807 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1808 | tree new_temp; | |
1809 | tree def; | |
1810 | gimple def_stmt; | |
1811 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
1812 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
1813 | int ncopies; | |
1814 | int i; | |
1815 | VEC(tree,heap) *vec_oprnds = NULL; | |
1816 | tree vop; | |
a70d6342 | 1817 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1818 | |
1819 | /* Multiple types in SLP are handled by creating the appropriate number of | |
1820 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
1821 | case of SLP. */ | |
1822 | if (slp_node) | |
1823 | ncopies = 1; | |
1824 | else | |
1825 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
1826 | ||
1827 | gcc_assert (ncopies >= 1); | |
1828 | if (ncopies > 1) | |
1829 | return false; /* FORNOW */ | |
1830 | ||
a70d6342 | 1831 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
1832 | return false; |
1833 | ||
8644a673 | 1834 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1835 | return false; |
1836 | ||
1837 | /* Is vectorizable assignment? */ | |
1838 | if (!is_gimple_assign (stmt)) | |
1839 | return false; | |
1840 | ||
1841 | scalar_dest = gimple_assign_lhs (stmt); | |
1842 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
1843 | return false; | |
1844 | ||
1845 | if (gimple_assign_single_p (stmt) | |
1846 | || gimple_assign_rhs_code (stmt) == PAREN_EXPR) | |
1847 | op = gimple_assign_rhs1 (stmt); | |
1848 | else | |
1849 | return false; | |
1850 | ||
a70d6342 | 1851 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt[0])) |
ebfd146a IR |
1852 | { |
1853 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1854 | fprintf (vect_dump, "use not simple."); | |
1855 | return false; | |
1856 | } | |
1857 | ||
1858 | if (!vec_stmt) /* transformation not required. */ | |
1859 | { | |
1860 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
1861 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1862 | fprintf (vect_dump, "=== vectorizable_assignment ==="); | |
1863 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
1864 | return true; | |
1865 | } | |
1866 | ||
1867 | /** Transform. **/ | |
1868 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1869 | fprintf (vect_dump, "transform assignment."); | |
1870 | ||
1871 | /* Handle def. */ | |
1872 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
1873 | ||
1874 | /* Handle use. */ | |
1875 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); | |
1876 | ||
1877 | /* Arguments are ready. create the new vector stmt. */ | |
1878 | for (i = 0; VEC_iterate (tree, vec_oprnds, i, vop); i++) | |
1879 | { | |
1880 | *vec_stmt = gimple_build_assign (vec_dest, vop); | |
1881 | new_temp = make_ssa_name (vec_dest, *vec_stmt); | |
1882 | gimple_assign_set_lhs (*vec_stmt, new_temp); | |
1883 | vect_finish_stmt_generation (stmt, *vec_stmt, gsi); | |
1884 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt; | |
1885 | ||
1886 | if (slp_node) | |
1887 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), *vec_stmt); | |
1888 | } | |
1889 | ||
1890 | VEC_free (tree, heap, vec_oprnds); | |
1891 | return true; | |
1892 | } | |
1893 | ||
1894 | /* Function vectorizable_operation. | |
1895 | ||
1896 | Check if STMT performs a binary or unary operation that can be vectorized. | |
1897 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1898 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
1899 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1900 | ||
1901 | static bool | |
1902 | vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi, | |
1903 | gimple *vec_stmt, slp_tree slp_node) | |
1904 | { | |
1905 | tree vec_dest; | |
1906 | tree scalar_dest; | |
1907 | tree op0, op1 = NULL; | |
1908 | tree vec_oprnd1 = NULL_TREE; | |
1909 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1910 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
1911 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1912 | enum tree_code code; | |
1913 | enum machine_mode vec_mode; | |
1914 | tree new_temp; | |
1915 | int op_type; | |
1916 | optab optab; | |
1917 | int icode; | |
1918 | enum machine_mode optab_op2_mode; | |
1919 | tree def; | |
1920 | gimple def_stmt; | |
1921 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
1922 | gimple new_stmt = NULL; | |
1923 | stmt_vec_info prev_stmt_info; | |
1924 | int nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
1925 | int nunits_out; | |
1926 | tree vectype_out; | |
1927 | int ncopies; | |
1928 | int j, i; | |
1929 | VEC(tree,heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; | |
1930 | tree vop0, vop1; | |
1931 | unsigned int k; | |
1932 | bool shift_p = false; | |
1933 | bool scalar_shift_arg = false; | |
a70d6342 IR |
1934 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
1935 | int vf; | |
1936 | ||
1937 | if (loop_vinfo) | |
1938 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
1939 | else | |
1940 | /* FORNOW: multiple types are not supported in basic block SLP. */ | |
1941 | vf = nunits_in; | |
ebfd146a IR |
1942 | |
1943 | /* Multiple types in SLP are handled by creating the appropriate number of | |
1944 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
1945 | case of SLP. */ | |
1946 | if (slp_node) | |
1947 | ncopies = 1; | |
1948 | else | |
1949 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1950 | ||
1951 | gcc_assert (ncopies >= 1); | |
1952 | ||
a70d6342 | 1953 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
1954 | return false; |
1955 | ||
8644a673 | 1956 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1957 | return false; |
1958 | ||
1959 | /* Is STMT a vectorizable binary/unary operation? */ | |
1960 | if (!is_gimple_assign (stmt)) | |
1961 | return false; | |
1962 | ||
1963 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
1964 | return false; | |
1965 | ||
1966 | scalar_dest = gimple_assign_lhs (stmt); | |
1967 | vectype_out = get_vectype_for_scalar_type (TREE_TYPE (scalar_dest)); | |
1968 | if (!vectype_out) | |
1969 | return false; | |
1970 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
1971 | if (nunits_out != nunits_in) | |
1972 | return false; | |
1973 | ||
1974 | code = gimple_assign_rhs_code (stmt); | |
1975 | ||
1976 | /* For pointer addition, we should use the normal plus for | |
1977 | the vector addition. */ | |
1978 | if (code == POINTER_PLUS_EXPR) | |
1979 | code = PLUS_EXPR; | |
1980 | ||
1981 | /* Support only unary or binary operations. */ | |
1982 | op_type = TREE_CODE_LENGTH (code); | |
1983 | if (op_type != unary_op && op_type != binary_op) | |
1984 | { | |
1985 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1986 | fprintf (vect_dump, "num. args = %d (not unary/binary op).", op_type); | |
1987 | return false; | |
1988 | } | |
1989 | ||
1990 | op0 = gimple_assign_rhs1 (stmt); | |
a70d6342 | 1991 | if (!vect_is_simple_use (op0, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt[0])) |
ebfd146a IR |
1992 | { |
1993 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1994 | fprintf (vect_dump, "use not simple."); | |
1995 | return false; | |
1996 | } | |
1997 | ||
1998 | if (op_type == binary_op) | |
1999 | { | |
2000 | op1 = gimple_assign_rhs2 (stmt); | |
a70d6342 IR |
2001 | if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def, |
2002 | &dt[1])) | |
ebfd146a IR |
2003 | { |
2004 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2005 | fprintf (vect_dump, "use not simple."); | |
2006 | return false; | |
2007 | } | |
2008 | } | |
2009 | ||
2010 | /* If this is a shift/rotate, determine whether the shift amount is a vector, | |
2011 | or scalar. If the shift/rotate amount is a vector, use the vector/vector | |
2012 | shift optabs. */ | |
2013 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
2014 | || code == RROTATE_EXPR) | |
2015 | { | |
2016 | shift_p = true; | |
2017 | ||
2018 | /* vector shifted by vector */ | |
8644a673 | 2019 | if (dt[1] == vect_internal_def) |
ebfd146a IR |
2020 | { |
2021 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
2022 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2023 | fprintf (vect_dump, "vector/vector shift/rotate found."); | |
2024 | } | |
2025 | ||
2026 | /* See if the machine has a vector shifted by scalar insn and if not | |
2027 | then see if it has a vector shifted by vector insn */ | |
8644a673 | 2028 | else if (dt[1] == vect_constant_def || dt[1] == vect_external_def) |
ebfd146a IR |
2029 | { |
2030 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
2031 | if (optab | |
2032 | && (optab_handler (optab, TYPE_MODE (vectype))->insn_code | |
2033 | != CODE_FOR_nothing)) | |
2034 | { | |
2035 | scalar_shift_arg = true; | |
2036 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2037 | fprintf (vect_dump, "vector/scalar shift/rotate found."); | |
2038 | } | |
2039 | else | |
2040 | { | |
2041 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
ad6c0864 | 2042 | if (optab |
ebfd146a IR |
2043 | && (optab_handler (optab, TYPE_MODE (vectype))->insn_code |
2044 | != CODE_FOR_nothing)) | |
ad6c0864 MM |
2045 | { |
2046 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2047 | fprintf (vect_dump, "vector/vector shift/rotate found."); | |
2048 | ||
2049 | /* Unlike the other binary operators, shifts/rotates have | |
2050 | the rhs being int, instead of the same type as the lhs, | |
2051 | so make sure the scalar is the right type if we are | |
2052 | dealing with vectors of short/char. */ | |
2053 | if (dt[1] == vect_constant_def) | |
2054 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
2055 | } | |
ebfd146a IR |
2056 | } |
2057 | } | |
2058 | ||
2059 | else | |
2060 | { | |
2061 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2062 | fprintf (vect_dump, "operand mode requires invariant argument."); | |
2063 | return false; | |
2064 | } | |
2065 | } | |
2066 | else | |
2067 | optab = optab_for_tree_code (code, vectype, optab_default); | |
2068 | ||
2069 | /* Supportable by target? */ | |
2070 | if (!optab) | |
2071 | { | |
2072 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2073 | fprintf (vect_dump, "no optab."); | |
2074 | return false; | |
2075 | } | |
2076 | vec_mode = TYPE_MODE (vectype); | |
2077 | icode = (int) optab_handler (optab, vec_mode)->insn_code; | |
2078 | if (icode == CODE_FOR_nothing) | |
2079 | { | |
2080 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2081 | fprintf (vect_dump, "op not supported by target."); | |
2082 | /* Check only during analysis. */ | |
2083 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
a70d6342 | 2084 | || (vf < vect_min_worthwhile_factor (code) |
ebfd146a IR |
2085 | && !vec_stmt)) |
2086 | return false; | |
2087 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2088 | fprintf (vect_dump, "proceeding using word mode."); | |
2089 | } | |
2090 | ||
2091 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
2092 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
a70d6342 | 2093 | && vf < vect_min_worthwhile_factor (code) |
ebfd146a IR |
2094 | && !vec_stmt) |
2095 | { | |
2096 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2097 | fprintf (vect_dump, "not worthwhile without SIMD support."); | |
2098 | return false; | |
2099 | } | |
2100 | ||
2101 | if (!vec_stmt) /* transformation not required. */ | |
2102 | { | |
2103 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; | |
2104 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2105 | fprintf (vect_dump, "=== vectorizable_operation ==="); | |
2106 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2107 | return true; | |
2108 | } | |
2109 | ||
2110 | /** Transform. **/ | |
2111 | ||
2112 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2113 | fprintf (vect_dump, "transform binary/unary operation."); | |
2114 | ||
2115 | /* Handle def. */ | |
2116 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2117 | ||
2118 | /* Allocate VECs for vector operands. In case of SLP, vector operands are | |
2119 | created in the previous stages of the recursion, so no allocation is | |
2120 | needed, except for the case of shift with scalar shift argument. In that | |
2121 | case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to | |
2122 | be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE. | |
2123 | In case of loop-based vectorization we allocate VECs of size 1. We | |
2124 | allocate VEC_OPRNDS1 only in case of binary operation. */ | |
2125 | if (!slp_node) | |
2126 | { | |
2127 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
2128 | if (op_type == binary_op) | |
2129 | vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
2130 | } | |
2131 | else if (scalar_shift_arg) | |
2132 | vec_oprnds1 = VEC_alloc (tree, heap, slp_node->vec_stmts_size); | |
2133 | ||
2134 | /* In case the vectorization factor (VF) is bigger than the number | |
2135 | of elements that we can fit in a vectype (nunits), we have to generate | |
2136 | more than one vector stmt - i.e - we need to "unroll" the | |
2137 | vector stmt by a factor VF/nunits. In doing so, we record a pointer | |
2138 | from one copy of the vector stmt to the next, in the field | |
2139 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
2140 | stages to find the correct vector defs to be used when vectorizing | |
2141 | stmts that use the defs of the current stmt. The example below illustrates | |
2142 | the vectorization process when VF=16 and nunits=4 (i.e - we need to create | |
2143 | 4 vectorized stmts): | |
2144 | ||
2145 | before vectorization: | |
2146 | RELATED_STMT VEC_STMT | |
2147 | S1: x = memref - - | |
2148 | S2: z = x + 1 - - | |
2149 | ||
2150 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
2151 | there): | |
2152 | RELATED_STMT VEC_STMT | |
2153 | VS1_0: vx0 = memref0 VS1_1 - | |
2154 | VS1_1: vx1 = memref1 VS1_2 - | |
2155 | VS1_2: vx2 = memref2 VS1_3 - | |
2156 | VS1_3: vx3 = memref3 - - | |
2157 | S1: x = load - VS1_0 | |
2158 | S2: z = x + 1 - - | |
2159 | ||
2160 | step2: vectorize stmt S2 (done here): | |
2161 | To vectorize stmt S2 we first need to find the relevant vector | |
2162 | def for the first operand 'x'. This is, as usual, obtained from | |
2163 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
2164 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
2165 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
2166 | the vector stmt VS2_0, and as usual, record it in the | |
2167 | STMT_VINFO_VEC_STMT of stmt S2. | |
2168 | When creating the second copy (VS2_1), we obtain the relevant vector | |
2169 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
2170 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
2171 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
2172 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
2173 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
2174 | chain of stmts and pointers: | |
2175 | RELATED_STMT VEC_STMT | |
2176 | VS1_0: vx0 = memref0 VS1_1 - | |
2177 | VS1_1: vx1 = memref1 VS1_2 - | |
2178 | VS1_2: vx2 = memref2 VS1_3 - | |
2179 | VS1_3: vx3 = memref3 - - | |
2180 | S1: x = load - VS1_0 | |
2181 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
2182 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
2183 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
2184 | VS2_3: vz3 = vx3 + v1 - - | |
2185 | S2: z = x + 1 - VS2_0 */ | |
2186 | ||
2187 | prev_stmt_info = NULL; | |
2188 | for (j = 0; j < ncopies; j++) | |
2189 | { | |
2190 | /* Handle uses. */ | |
2191 | if (j == 0) | |
2192 | { | |
2193 | if (op_type == binary_op && scalar_shift_arg) | |
2194 | { | |
2195 | /* Vector shl and shr insn patterns can be defined with scalar | |
2196 | operand 2 (shift operand). In this case, use constant or loop | |
2197 | invariant op1 directly, without extending it to vector mode | |
2198 | first. */ | |
2199 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
2200 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
2201 | { | |
2202 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2203 | fprintf (vect_dump, "operand 1 using scalar mode."); | |
2204 | vec_oprnd1 = op1; | |
2205 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2206 | if (slp_node) | |
2207 | { | |
2208 | /* Store vec_oprnd1 for every vector stmt to be created | |
2209 | for SLP_NODE. We check during the analysis that all the | |
2210 | shift arguments are the same. | |
2211 | TODO: Allow different constants for different vector | |
2212 | stmts generated for an SLP instance. */ | |
2213 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
2214 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2215 | } | |
2216 | } | |
2217 | } | |
2218 | ||
2219 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
2220 | (a special case for certain kind of vector shifts); otherwise, | |
2221 | operand 1 should be of a vector type (the usual case). */ | |
2222 | if (op_type == binary_op && !vec_oprnd1) | |
2223 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
2224 | slp_node); | |
2225 | else | |
2226 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2227 | slp_node); | |
2228 | } | |
2229 | else | |
2230 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
2231 | ||
2232 | /* Arguments are ready. Create the new vector stmt. */ | |
2233 | for (i = 0; VEC_iterate (tree, vec_oprnds0, i, vop0); i++) | |
2234 | { | |
2235 | vop1 = ((op_type == binary_op) | |
2236 | ? VEC_index (tree, vec_oprnds1, i) : NULL); | |
2237 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); | |
2238 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2239 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2240 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2241 | if (slp_node) | |
2242 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2243 | } | |
2244 | ||
2245 | if (slp_node) | |
2246 | continue; | |
2247 | ||
2248 | if (j == 0) | |
2249 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2250 | else | |
2251 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2252 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2253 | } | |
2254 | ||
2255 | VEC_free (tree, heap, vec_oprnds0); | |
2256 | if (vec_oprnds1) | |
2257 | VEC_free (tree, heap, vec_oprnds1); | |
2258 | ||
2259 | return true; | |
2260 | } | |
2261 | ||
2262 | ||
2263 | /* Get vectorized definitions for loop-based vectorization. For the first | |
2264 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
2265 | scalar operand), and for the rest we get a copy with | |
2266 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
2267 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
2268 | The vectors are collected into VEC_OPRNDS. */ | |
2269 | ||
2270 | static void | |
2271 | vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt, | |
2272 | VEC (tree, heap) **vec_oprnds, int multi_step_cvt) | |
2273 | { | |
2274 | tree vec_oprnd; | |
2275 | ||
2276 | /* Get first vector operand. */ | |
2277 | /* All the vector operands except the very first one (that is scalar oprnd) | |
2278 | are stmt copies. */ | |
2279 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
2280 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL); | |
2281 | else | |
2282 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
2283 | ||
2284 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
2285 | ||
2286 | /* Get second vector operand. */ | |
2287 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
2288 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
2289 | ||
2290 | *oprnd = vec_oprnd; | |
2291 | ||
2292 | /* For conversion in multiple steps, continue to get operands | |
2293 | recursively. */ | |
2294 | if (multi_step_cvt) | |
2295 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
2296 | } | |
2297 | ||
2298 | ||
2299 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
2300 | For multi-step conversions store the resulting vectors and call the function | |
2301 | recursively. */ | |
2302 | ||
2303 | static void | |
2304 | vect_create_vectorized_demotion_stmts (VEC (tree, heap) **vec_oprnds, | |
2305 | int multi_step_cvt, gimple stmt, | |
2306 | VEC (tree, heap) *vec_dsts, | |
2307 | gimple_stmt_iterator *gsi, | |
2308 | slp_tree slp_node, enum tree_code code, | |
2309 | stmt_vec_info *prev_stmt_info) | |
2310 | { | |
2311 | unsigned int i; | |
2312 | tree vop0, vop1, new_tmp, vec_dest; | |
2313 | gimple new_stmt; | |
2314 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2315 | ||
2316 | vec_dest = VEC_pop (tree, vec_dsts); | |
2317 | ||
2318 | for (i = 0; i < VEC_length (tree, *vec_oprnds); i += 2) | |
2319 | { | |
2320 | /* Create demotion operation. */ | |
2321 | vop0 = VEC_index (tree, *vec_oprnds, i); | |
2322 | vop1 = VEC_index (tree, *vec_oprnds, i + 1); | |
2323 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); | |
2324 | new_tmp = make_ssa_name (vec_dest, new_stmt); | |
2325 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
2326 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2327 | ||
2328 | if (multi_step_cvt) | |
2329 | /* Store the resulting vector for next recursive call. */ | |
2330 | VEC_replace (tree, *vec_oprnds, i/2, new_tmp); | |
2331 | else | |
2332 | { | |
2333 | /* This is the last step of the conversion sequence. Store the | |
2334 | vectors in SLP_NODE or in vector info of the scalar statement | |
2335 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
2336 | if (slp_node) | |
2337 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2338 | else | |
2339 | { | |
2340 | if (!*prev_stmt_info) | |
2341 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2342 | else | |
2343 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
2344 | ||
2345 | *prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2346 | } | |
2347 | } | |
2348 | } | |
2349 | ||
2350 | /* For multi-step demotion operations we first generate demotion operations | |
2351 | from the source type to the intermediate types, and then combine the | |
2352 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
2353 | type. */ | |
2354 | if (multi_step_cvt) | |
2355 | { | |
2356 | /* At each level of recursion we have have of the operands we had at the | |
2357 | previous level. */ | |
2358 | VEC_truncate (tree, *vec_oprnds, (i+1)/2); | |
2359 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, | |
2360 | stmt, vec_dsts, gsi, slp_node, | |
2361 | code, prev_stmt_info); | |
2362 | } | |
2363 | } | |
2364 | ||
2365 | ||
2366 | /* Function vectorizable_type_demotion | |
2367 | ||
2368 | Check if STMT performs a binary or unary operation that involves | |
2369 | type demotion, and if it can be vectorized. | |
2370 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2371 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2372 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2373 | ||
2374 | static bool | |
2375 | vectorizable_type_demotion (gimple stmt, gimple_stmt_iterator *gsi, | |
2376 | gimple *vec_stmt, slp_tree slp_node) | |
2377 | { | |
2378 | tree vec_dest; | |
2379 | tree scalar_dest; | |
2380 | tree op0; | |
2381 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2382 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2383 | enum tree_code code, code1 = ERROR_MARK; | |
2384 | tree def; | |
2385 | gimple def_stmt; | |
2386 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2387 | stmt_vec_info prev_stmt_info; | |
2388 | int nunits_in; | |
2389 | int nunits_out; | |
2390 | tree vectype_out; | |
2391 | int ncopies; | |
2392 | int j, i; | |
2393 | tree vectype_in; | |
2394 | int multi_step_cvt = 0; | |
2395 | VEC (tree, heap) *vec_oprnds0 = NULL; | |
2396 | VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL; | |
2397 | tree last_oprnd, intermediate_type; | |
2398 | ||
a70d6342 IR |
2399 | /* FORNOW: not supported by basic block SLP vectorization. */ |
2400 | gcc_assert (loop_vinfo); | |
2401 | ||
ebfd146a IR |
2402 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
2403 | return false; | |
2404 | ||
8644a673 | 2405 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2406 | return false; |
2407 | ||
2408 | /* Is STMT a vectorizable type-demotion operation? */ | |
2409 | if (!is_gimple_assign (stmt)) | |
2410 | return false; | |
2411 | ||
2412 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2413 | return false; | |
2414 | ||
2415 | code = gimple_assign_rhs_code (stmt); | |
2416 | if (!CONVERT_EXPR_CODE_P (code)) | |
2417 | return false; | |
2418 | ||
2419 | op0 = gimple_assign_rhs1 (stmt); | |
2420 | vectype_in = get_vectype_for_scalar_type (TREE_TYPE (op0)); | |
2421 | if (!vectype_in) | |
2422 | return false; | |
2423 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); | |
2424 | ||
2425 | scalar_dest = gimple_assign_lhs (stmt); | |
2426 | vectype_out = get_vectype_for_scalar_type (TREE_TYPE (scalar_dest)); | |
2427 | if (!vectype_out) | |
2428 | return false; | |
2429 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
2430 | if (nunits_in >= nunits_out) | |
2431 | return false; | |
2432 | ||
2433 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2434 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2435 | case of SLP. */ | |
2436 | if (slp_node) | |
2437 | ncopies = 1; | |
2438 | else | |
2439 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
ebfd146a IR |
2440 | gcc_assert (ncopies >= 1); |
2441 | ||
2442 | if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2443 | && INTEGRAL_TYPE_P (TREE_TYPE (op0))) | |
2444 | || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest)) | |
2445 | && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0)) | |
2446 | && CONVERT_EXPR_CODE_P (code)))) | |
2447 | return false; | |
2448 | ||
2449 | /* Check the operands of the operation. */ | |
a70d6342 | 2450 | if (!vect_is_simple_use (op0, loop_vinfo, NULL, &def_stmt, &def, &dt[0])) |
ebfd146a IR |
2451 | { |
2452 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2453 | fprintf (vect_dump, "use not simple."); | |
2454 | return false; | |
2455 | } | |
2456 | ||
2457 | /* Supportable by target? */ | |
2458 | if (!supportable_narrowing_operation (code, stmt, vectype_in, &code1, | |
2459 | &multi_step_cvt, &interm_types)) | |
2460 | return false; | |
2461 | ||
2462 | STMT_VINFO_VECTYPE (stmt_info) = vectype_in; | |
2463 | ||
2464 | if (!vec_stmt) /* transformation not required. */ | |
2465 | { | |
2466 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
2467 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2468 | fprintf (vect_dump, "=== vectorizable_demotion ==="); | |
2469 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2470 | return true; | |
2471 | } | |
2472 | ||
2473 | /** Transform. **/ | |
2474 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2475 | fprintf (vect_dump, "transform type demotion operation. ncopies = %d.", | |
2476 | ncopies); | |
2477 | ||
2478 | /* In case of multi-step demotion, we first generate demotion operations to | |
2479 | the intermediate types, and then from that types to the final one. | |
2480 | We create vector destinations for the intermediate type (TYPES) received | |
2481 | from supportable_narrowing_operation, and store them in the correct order | |
2482 | for future use in vect_create_vectorized_demotion_stmts(). */ | |
2483 | if (multi_step_cvt) | |
2484 | vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1); | |
2485 | else | |
2486 | vec_dsts = VEC_alloc (tree, heap, 1); | |
2487 | ||
2488 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2489 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2490 | ||
2491 | if (multi_step_cvt) | |
2492 | { | |
2493 | for (i = VEC_length (tree, interm_types) - 1; | |
2494 | VEC_iterate (tree, interm_types, i, intermediate_type); i--) | |
2495 | { | |
2496 | vec_dest = vect_create_destination_var (scalar_dest, | |
2497 | intermediate_type); | |
2498 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2499 | } | |
2500 | } | |
2501 | ||
2502 | /* In case the vectorization factor (VF) is bigger than the number | |
2503 | of elements that we can fit in a vectype (nunits), we have to generate | |
2504 | more than one vector stmt - i.e - we need to "unroll" the | |
2505 | vector stmt by a factor VF/nunits. */ | |
2506 | last_oprnd = op0; | |
2507 | prev_stmt_info = NULL; | |
2508 | for (j = 0; j < ncopies; j++) | |
2509 | { | |
2510 | /* Handle uses. */ | |
2511 | if (slp_node) | |
2512 | vect_get_slp_defs (slp_node, &vec_oprnds0, NULL); | |
2513 | else | |
2514 | { | |
2515 | VEC_free (tree, heap, vec_oprnds0); | |
2516 | vec_oprnds0 = VEC_alloc (tree, heap, | |
2517 | (multi_step_cvt ? vect_pow2 (multi_step_cvt) * 2 : 2)); | |
2518 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, | |
2519 | vect_pow2 (multi_step_cvt) - 1); | |
2520 | } | |
2521 | ||
2522 | /* Arguments are ready. Create the new vector stmts. */ | |
2523 | tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts); | |
2524 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, | |
2525 | multi_step_cvt, stmt, tmp_vec_dsts, | |
2526 | gsi, slp_node, code1, | |
2527 | &prev_stmt_info); | |
2528 | } | |
2529 | ||
2530 | VEC_free (tree, heap, vec_oprnds0); | |
2531 | VEC_free (tree, heap, vec_dsts); | |
2532 | VEC_free (tree, heap, tmp_vec_dsts); | |
2533 | VEC_free (tree, heap, interm_types); | |
2534 | ||
2535 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2536 | return true; | |
2537 | } | |
2538 | ||
2539 | ||
2540 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
2541 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
2542 | the resulting vectors and call the function recursively. */ | |
2543 | ||
2544 | static void | |
2545 | vect_create_vectorized_promotion_stmts (VEC (tree, heap) **vec_oprnds0, | |
2546 | VEC (tree, heap) **vec_oprnds1, | |
2547 | int multi_step_cvt, gimple stmt, | |
2548 | VEC (tree, heap) *vec_dsts, | |
2549 | gimple_stmt_iterator *gsi, | |
2550 | slp_tree slp_node, enum tree_code code1, | |
2551 | enum tree_code code2, tree decl1, | |
2552 | tree decl2, int op_type, | |
2553 | stmt_vec_info *prev_stmt_info) | |
2554 | { | |
2555 | int i; | |
2556 | tree vop0, vop1, new_tmp1, new_tmp2, vec_dest; | |
2557 | gimple new_stmt1, new_stmt2; | |
2558 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2559 | VEC (tree, heap) *vec_tmp; | |
2560 | ||
2561 | vec_dest = VEC_pop (tree, vec_dsts); | |
2562 | vec_tmp = VEC_alloc (tree, heap, VEC_length (tree, *vec_oprnds0) * 2); | |
2563 | ||
2564 | for (i = 0; VEC_iterate (tree, *vec_oprnds0, i, vop0); i++) | |
2565 | { | |
2566 | if (op_type == binary_op) | |
2567 | vop1 = VEC_index (tree, *vec_oprnds1, i); | |
2568 | else | |
2569 | vop1 = NULL_TREE; | |
2570 | ||
2571 | /* Generate the two halves of promotion operation. */ | |
2572 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
2573 | op_type, vec_dest, gsi, stmt); | |
2574 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
2575 | op_type, vec_dest, gsi, stmt); | |
2576 | if (is_gimple_call (new_stmt1)) | |
2577 | { | |
2578 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
2579 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
2580 | } | |
2581 | else | |
2582 | { | |
2583 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
2584 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
2585 | } | |
2586 | ||
2587 | if (multi_step_cvt) | |
2588 | { | |
2589 | /* Store the results for the recursive call. */ | |
2590 | VEC_quick_push (tree, vec_tmp, new_tmp1); | |
2591 | VEC_quick_push (tree, vec_tmp, new_tmp2); | |
2592 | } | |
2593 | else | |
2594 | { | |
2595 | /* Last step of promotion sequience - store the results. */ | |
2596 | if (slp_node) | |
2597 | { | |
2598 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt1); | |
2599 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt2); | |
2600 | } | |
2601 | else | |
2602 | { | |
2603 | if (!*prev_stmt_info) | |
2604 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt1; | |
2605 | else | |
2606 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt1; | |
2607 | ||
2608 | *prev_stmt_info = vinfo_for_stmt (new_stmt1); | |
2609 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt2; | |
2610 | *prev_stmt_info = vinfo_for_stmt (new_stmt2); | |
2611 | } | |
2612 | } | |
2613 | } | |
2614 | ||
2615 | if (multi_step_cvt) | |
2616 | { | |
2617 | /* For multi-step promotion operation we first generate we call the | |
2618 | function recurcively for every stage. We start from the input type, | |
2619 | create promotion operations to the intermediate types, and then | |
2620 | create promotions to the output type. */ | |
2621 | *vec_oprnds0 = VEC_copy (tree, heap, vec_tmp); | |
2622 | VEC_free (tree, heap, vec_tmp); | |
2623 | vect_create_vectorized_promotion_stmts (vec_oprnds0, vec_oprnds1, | |
2624 | multi_step_cvt - 1, stmt, | |
2625 | vec_dsts, gsi, slp_node, code1, | |
2626 | code2, decl2, decl2, op_type, | |
2627 | prev_stmt_info); | |
2628 | } | |
2629 | } | |
a70d6342 | 2630 | |
ebfd146a IR |
2631 | |
2632 | /* Function vectorizable_type_promotion | |
2633 | ||
2634 | Check if STMT performs a binary or unary operation that involves | |
2635 | type promotion, and if it can be vectorized. | |
2636 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2637 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2638 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2639 | ||
2640 | static bool | |
2641 | vectorizable_type_promotion (gimple stmt, gimple_stmt_iterator *gsi, | |
2642 | gimple *vec_stmt, slp_tree slp_node) | |
2643 | { | |
2644 | tree vec_dest; | |
2645 | tree scalar_dest; | |
2646 | tree op0, op1 = NULL; | |
2647 | tree vec_oprnd0=NULL, vec_oprnd1=NULL; | |
2648 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2649 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2650 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
2651 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; | |
2652 | int op_type; | |
2653 | tree def; | |
2654 | gimple def_stmt; | |
2655 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2656 | stmt_vec_info prev_stmt_info; | |
2657 | int nunits_in; | |
2658 | int nunits_out; | |
2659 | tree vectype_out; | |
2660 | int ncopies; | |
2661 | int j, i; | |
2662 | tree vectype_in; | |
2663 | tree intermediate_type = NULL_TREE; | |
2664 | int multi_step_cvt = 0; | |
2665 | VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; | |
2666 | VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL; | |
2667 | ||
a70d6342 IR |
2668 | /* FORNOW: not supported by basic block SLP vectorization. */ |
2669 | gcc_assert (loop_vinfo); | |
2670 | ||
ebfd146a IR |
2671 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
2672 | return false; | |
2673 | ||
8644a673 | 2674 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2675 | return false; |
2676 | ||
2677 | /* Is STMT a vectorizable type-promotion operation? */ | |
2678 | if (!is_gimple_assign (stmt)) | |
2679 | return false; | |
2680 | ||
2681 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2682 | return false; | |
2683 | ||
2684 | code = gimple_assign_rhs_code (stmt); | |
2685 | if (!CONVERT_EXPR_CODE_P (code) | |
2686 | && code != WIDEN_MULT_EXPR) | |
2687 | return false; | |
2688 | ||
2689 | op0 = gimple_assign_rhs1 (stmt); | |
2690 | vectype_in = get_vectype_for_scalar_type (TREE_TYPE (op0)); | |
2691 | if (!vectype_in) | |
2692 | return false; | |
2693 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); | |
2694 | ||
2695 | scalar_dest = gimple_assign_lhs (stmt); | |
2696 | vectype_out = get_vectype_for_scalar_type (TREE_TYPE (scalar_dest)); | |
2697 | if (!vectype_out) | |
2698 | return false; | |
2699 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
2700 | if (nunits_in <= nunits_out) | |
2701 | return false; | |
2702 | ||
2703 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2704 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2705 | case of SLP. */ | |
2706 | if (slp_node) | |
2707 | ncopies = 1; | |
2708 | else | |
2709 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2710 | ||
2711 | gcc_assert (ncopies >= 1); | |
2712 | ||
2713 | if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2714 | && INTEGRAL_TYPE_P (TREE_TYPE (op0))) | |
2715 | || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest)) | |
2716 | && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0)) | |
2717 | && CONVERT_EXPR_CODE_P (code)))) | |
2718 | return false; | |
2719 | ||
2720 | /* Check the operands of the operation. */ | |
a70d6342 | 2721 | if (!vect_is_simple_use (op0, loop_vinfo, NULL, &def_stmt, &def, &dt[0])) |
ebfd146a IR |
2722 | { |
2723 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2724 | fprintf (vect_dump, "use not simple."); | |
2725 | return false; | |
2726 | } | |
2727 | ||
2728 | op_type = TREE_CODE_LENGTH (code); | |
2729 | if (op_type == binary_op) | |
2730 | { | |
2731 | op1 = gimple_assign_rhs2 (stmt); | |
a70d6342 | 2732 | if (!vect_is_simple_use (op1, loop_vinfo, NULL, &def_stmt, &def, &dt[1])) |
ebfd146a IR |
2733 | { |
2734 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2735 | fprintf (vect_dump, "use not simple."); | |
2736 | return false; | |
2737 | } | |
2738 | } | |
2739 | ||
2740 | /* Supportable by target? */ | |
2741 | if (!supportable_widening_operation (code, stmt, vectype_in, | |
2742 | &decl1, &decl2, &code1, &code2, | |
2743 | &multi_step_cvt, &interm_types)) | |
2744 | return false; | |
2745 | ||
2746 | /* Binary widening operation can only be supported directly by the | |
2747 | architecture. */ | |
2748 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
2749 | ||
2750 | STMT_VINFO_VECTYPE (stmt_info) = vectype_in; | |
2751 | ||
2752 | if (!vec_stmt) /* transformation not required. */ | |
2753 | { | |
2754 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
2755 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2756 | fprintf (vect_dump, "=== vectorizable_promotion ==="); | |
2757 | vect_model_simple_cost (stmt_info, 2*ncopies, dt, NULL); | |
2758 | return true; | |
2759 | } | |
2760 | ||
2761 | /** Transform. **/ | |
2762 | ||
2763 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2764 | fprintf (vect_dump, "transform type promotion operation. ncopies = %d.", | |
2765 | ncopies); | |
2766 | ||
2767 | /* Handle def. */ | |
2768 | /* In case of multi-step promotion, we first generate promotion operations | |
2769 | to the intermediate types, and then from that types to the final one. | |
2770 | We store vector destination in VEC_DSTS in the correct order for | |
2771 | recursive creation of promotion operations in | |
2772 | vect_create_vectorized_promotion_stmts(). Vector destinations are created | |
2773 | according to TYPES recieved from supportable_widening_operation(). */ | |
2774 | if (multi_step_cvt) | |
2775 | vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1); | |
2776 | else | |
2777 | vec_dsts = VEC_alloc (tree, heap, 1); | |
2778 | ||
2779 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2780 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2781 | ||
2782 | if (multi_step_cvt) | |
2783 | { | |
2784 | for (i = VEC_length (tree, interm_types) - 1; | |
2785 | VEC_iterate (tree, interm_types, i, intermediate_type); i--) | |
2786 | { | |
2787 | vec_dest = vect_create_destination_var (scalar_dest, | |
2788 | intermediate_type); | |
2789 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2790 | } | |
2791 | } | |
2792 | ||
2793 | if (!slp_node) | |
2794 | { | |
2795 | vec_oprnds0 = VEC_alloc (tree, heap, | |
2796 | (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
2797 | if (op_type == binary_op) | |
2798 | vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
2799 | } | |
2800 | ||
2801 | /* In case the vectorization factor (VF) is bigger than the number | |
2802 | of elements that we can fit in a vectype (nunits), we have to generate | |
2803 | more than one vector stmt - i.e - we need to "unroll" the | |
2804 | vector stmt by a factor VF/nunits. */ | |
2805 | ||
2806 | prev_stmt_info = NULL; | |
2807 | for (j = 0; j < ncopies; j++) | |
2808 | { | |
2809 | /* Handle uses. */ | |
2810 | if (j == 0) | |
2811 | { | |
2812 | if (slp_node) | |
2813 | vect_get_slp_defs (slp_node, &vec_oprnds0, &vec_oprnds1); | |
2814 | else | |
2815 | { | |
2816 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
2817 | VEC_quick_push (tree, vec_oprnds0, vec_oprnd0); | |
2818 | if (op_type == binary_op) | |
2819 | { | |
2820 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, NULL); | |
2821 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2822 | } | |
2823 | } | |
2824 | } | |
2825 | else | |
2826 | { | |
2827 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
2828 | VEC_replace (tree, vec_oprnds0, 0, vec_oprnd0); | |
2829 | if (op_type == binary_op) | |
2830 | { | |
2831 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd1); | |
2832 | VEC_replace (tree, vec_oprnds1, 0, vec_oprnd1); | |
2833 | } | |
2834 | } | |
2835 | ||
2836 | /* Arguments are ready. Create the new vector stmts. */ | |
2837 | tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts); | |
2838 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, &vec_oprnds1, | |
2839 | multi_step_cvt, stmt, | |
2840 | tmp_vec_dsts, | |
2841 | gsi, slp_node, code1, code2, | |
2842 | decl1, decl2, op_type, | |
2843 | &prev_stmt_info); | |
2844 | } | |
2845 | ||
2846 | VEC_free (tree, heap, vec_dsts); | |
2847 | VEC_free (tree, heap, tmp_vec_dsts); | |
2848 | VEC_free (tree, heap, interm_types); | |
2849 | VEC_free (tree, heap, vec_oprnds0); | |
2850 | VEC_free (tree, heap, vec_oprnds1); | |
2851 | ||
2852 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2853 | return true; | |
2854 | } | |
2855 | ||
2856 | ||
2857 | /* Function vectorizable_store. | |
2858 | ||
2859 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that | |
2860 | can be vectorized. | |
2861 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2862 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2863 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2864 | ||
2865 | static bool | |
2866 | vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
2867 | slp_tree slp_node) | |
2868 | { | |
2869 | tree scalar_dest; | |
2870 | tree data_ref; | |
2871 | tree op; | |
2872 | tree vec_oprnd = NULL_TREE; | |
2873 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2874 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
2875 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2876 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 2877 | struct loop *loop = NULL; |
ebfd146a IR |
2878 | enum machine_mode vec_mode; |
2879 | tree dummy; | |
2880 | enum dr_alignment_support alignment_support_scheme; | |
2881 | tree def; | |
2882 | gimple def_stmt; | |
2883 | enum vect_def_type dt; | |
2884 | stmt_vec_info prev_stmt_info = NULL; | |
2885 | tree dataref_ptr = NULL_TREE; | |
2886 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2887 | int ncopies; | |
2888 | int j; | |
2889 | gimple next_stmt, first_stmt = NULL; | |
2890 | bool strided_store = false; | |
2891 | unsigned int group_size, i; | |
2892 | VEC(tree,heap) *dr_chain = NULL, *oprnds = NULL, *result_chain = NULL; | |
2893 | bool inv_p; | |
2894 | VEC(tree,heap) *vec_oprnds = NULL; | |
2895 | bool slp = (slp_node != NULL); | |
2896 | stmt_vec_info first_stmt_vinfo; | |
2897 | unsigned int vec_num; | |
a70d6342 IR |
2898 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2899 | ||
2900 | if (loop_vinfo) | |
2901 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
2902 | |
2903 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2904 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2905 | case of SLP. */ | |
2906 | if (slp) | |
2907 | ncopies = 1; | |
2908 | else | |
2909 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2910 | ||
2911 | gcc_assert (ncopies >= 1); | |
2912 | ||
2913 | /* FORNOW. This restriction should be relaxed. */ | |
a70d6342 | 2914 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) |
ebfd146a IR |
2915 | { |
2916 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2917 | fprintf (vect_dump, "multiple types in nested loop."); | |
2918 | return false; | |
2919 | } | |
2920 | ||
a70d6342 | 2921 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2922 | return false; |
2923 | ||
8644a673 | 2924 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2925 | return false; |
2926 | ||
2927 | /* Is vectorizable store? */ | |
2928 | ||
2929 | if (!is_gimple_assign (stmt)) | |
2930 | return false; | |
2931 | ||
2932 | scalar_dest = gimple_assign_lhs (stmt); | |
2933 | if (TREE_CODE (scalar_dest) != ARRAY_REF | |
2934 | && TREE_CODE (scalar_dest) != INDIRECT_REF | |
e9dbe7bb IR |
2935 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
2936 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
2937 | && TREE_CODE (scalar_dest) != REALPART_EXPR) | |
ebfd146a IR |
2938 | return false; |
2939 | ||
2940 | gcc_assert (gimple_assign_single_p (stmt)); | |
2941 | op = gimple_assign_rhs1 (stmt); | |
a70d6342 | 2942 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt)) |
ebfd146a IR |
2943 | { |
2944 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2945 | fprintf (vect_dump, "use not simple."); | |
2946 | return false; | |
2947 | } | |
2948 | ||
2949 | /* The scalar rhs type needs to be trivially convertible to the vector | |
2950 | component type. This should always be the case. */ | |
2951 | if (!useless_type_conversion_p (TREE_TYPE (vectype), TREE_TYPE (op))) | |
2952 | { | |
2953 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2954 | fprintf (vect_dump, "??? operands of different types"); | |
2955 | return false; | |
2956 | } | |
2957 | ||
2958 | vec_mode = TYPE_MODE (vectype); | |
2959 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
2960 | (e.g. - array initialization with 0). */ | |
2961 | if (optab_handler (mov_optab, (int)vec_mode)->insn_code == CODE_FOR_nothing) | |
2962 | return false; | |
2963 | ||
2964 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
2965 | return false; | |
2966 | ||
2967 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) | |
2968 | { | |
2969 | strided_store = true; | |
2970 | first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
2971 | if (!vect_strided_store_supported (vectype) | |
2972 | && !PURE_SLP_STMT (stmt_info) && !slp) | |
2973 | return false; | |
2974 | ||
2975 | if (first_stmt == stmt) | |
2976 | { | |
2977 | /* STMT is the leader of the group. Check the operands of all the | |
2978 | stmts of the group. */ | |
2979 | next_stmt = DR_GROUP_NEXT_DR (stmt_info); | |
2980 | while (next_stmt) | |
2981 | { | |
2982 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
2983 | op = gimple_assign_rhs1 (next_stmt); | |
a70d6342 IR |
2984 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, |
2985 | &def, &dt)) | |
ebfd146a IR |
2986 | { |
2987 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2988 | fprintf (vect_dump, "use not simple."); | |
2989 | return false; | |
2990 | } | |
2991 | next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); | |
2992 | } | |
2993 | } | |
2994 | } | |
2995 | ||
2996 | if (!vec_stmt) /* transformation not required. */ | |
2997 | { | |
2998 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
2999 | vect_model_store_cost (stmt_info, ncopies, dt, NULL); | |
3000 | return true; | |
3001 | } | |
3002 | ||
3003 | /** Transform. **/ | |
3004 | ||
3005 | if (strided_store) | |
3006 | { | |
3007 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
3008 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
3009 | ||
3010 | DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; | |
3011 | ||
3012 | /* FORNOW */ | |
a70d6342 | 3013 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
3014 | |
3015 | /* We vectorize all the stmts of the interleaving group when we | |
3016 | reach the last stmt in the group. */ | |
3017 | if (DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) | |
3018 | < DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
3019 | && !slp) | |
3020 | { | |
3021 | *vec_stmt = NULL; | |
3022 | return true; | |
3023 | } | |
3024 | ||
3025 | if (slp) | |
3026 | strided_store = false; | |
3027 | ||
3028 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
3029 | if (slp) | |
3030 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
3031 | else | |
3032 | vec_num = group_size; | |
3033 | } | |
3034 | else | |
3035 | { | |
3036 | first_stmt = stmt; | |
3037 | first_dr = dr; | |
3038 | group_size = vec_num = 1; | |
3039 | first_stmt_vinfo = stmt_info; | |
3040 | } | |
3041 | ||
3042 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3043 | fprintf (vect_dump, "transform store. ncopies = %d",ncopies); | |
3044 | ||
3045 | dr_chain = VEC_alloc (tree, heap, group_size); | |
3046 | oprnds = VEC_alloc (tree, heap, group_size); | |
3047 | ||
3048 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr); | |
3049 | gcc_assert (alignment_support_scheme); | |
ebfd146a IR |
3050 | |
3051 | /* In case the vectorization factor (VF) is bigger than the number | |
3052 | of elements that we can fit in a vectype (nunits), we have to generate | |
3053 | more than one vector stmt - i.e - we need to "unroll" the | |
3054 | vector stmt by a factor VF/nunits. For more details see documentation in | |
3055 | vect_get_vec_def_for_copy_stmt. */ | |
3056 | ||
3057 | /* In case of interleaving (non-unit strided access): | |
3058 | ||
3059 | S1: &base + 2 = x2 | |
3060 | S2: &base = x0 | |
3061 | S3: &base + 1 = x1 | |
3062 | S4: &base + 3 = x3 | |
3063 | ||
3064 | We create vectorized stores starting from base address (the access of the | |
3065 | first stmt in the chain (S2 in the above example), when the last store stmt | |
3066 | of the chain (S4) is reached: | |
3067 | ||
3068 | VS1: &base = vx2 | |
3069 | VS2: &base + vec_size*1 = vx0 | |
3070 | VS3: &base + vec_size*2 = vx1 | |
3071 | VS4: &base + vec_size*3 = vx3 | |
3072 | ||
3073 | Then permutation statements are generated: | |
3074 | ||
3075 | VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 > | |
3076 | VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 > | |
3077 | ... | |
3078 | ||
3079 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
3080 | (the order of the data-refs in the output of vect_permute_store_chain | |
3081 | corresponds to the order of scalar stmts in the interleaving chain - see | |
3082 | the documentation of vect_permute_store_chain()). | |
3083 | ||
3084 | In case of both multiple types and interleaving, above vector stores and | |
3085 | permutation stmts are created for every copy. The result vector stmts are | |
3086 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding | |
3087 | STMT_VINFO_RELATED_STMT for the next copies. | |
3088 | */ | |
3089 | ||
3090 | prev_stmt_info = NULL; | |
3091 | for (j = 0; j < ncopies; j++) | |
3092 | { | |
3093 | gimple new_stmt; | |
3094 | gimple ptr_incr; | |
3095 | ||
3096 | if (j == 0) | |
3097 | { | |
3098 | if (slp) | |
3099 | { | |
3100 | /* Get vectorized arguments for SLP_NODE. */ | |
3101 | vect_get_slp_defs (slp_node, &vec_oprnds, NULL); | |
3102 | ||
3103 | vec_oprnd = VEC_index (tree, vec_oprnds, 0); | |
3104 | } | |
3105 | else | |
3106 | { | |
3107 | /* For interleaved stores we collect vectorized defs for all the | |
3108 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
3109 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
3110 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. | |
3111 | ||
3112 | If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and | |
3113 | OPRNDS are of size 1. */ | |
3114 | next_stmt = first_stmt; | |
3115 | for (i = 0; i < group_size; i++) | |
3116 | { | |
3117 | /* Since gaps are not supported for interleaved stores, | |
3118 | GROUP_SIZE is the exact number of stmts in the chain. | |
3119 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
3120 | there is no interleaving, GROUP_SIZE is 1, and only one | |
3121 | iteration of the loop will be executed. */ | |
3122 | gcc_assert (next_stmt | |
3123 | && gimple_assign_single_p (next_stmt)); | |
3124 | op = gimple_assign_rhs1 (next_stmt); | |
3125 | ||
3126 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt, | |
3127 | NULL); | |
3128 | VEC_quick_push(tree, dr_chain, vec_oprnd); | |
3129 | VEC_quick_push(tree, oprnds, vec_oprnd); | |
3130 | next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); | |
3131 | } | |
3132 | } | |
3133 | ||
3134 | /* We should have catched mismatched types earlier. */ | |
3135 | gcc_assert (useless_type_conversion_p (vectype, | |
3136 | TREE_TYPE (vec_oprnd))); | |
3137 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, NULL, NULL_TREE, | |
3138 | &dummy, &ptr_incr, false, | |
5006671f | 3139 | &inv_p); |
a70d6342 | 3140 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a IR |
3141 | } |
3142 | else | |
3143 | { | |
3144 | /* For interleaved stores we created vectorized defs for all the | |
3145 | defs stored in OPRNDS in the previous iteration (previous copy). | |
3146 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
3147 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the | |
3148 | next copy. | |
3149 | If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and | |
3150 | OPRNDS are of size 1. */ | |
3151 | for (i = 0; i < group_size; i++) | |
3152 | { | |
3153 | op = VEC_index (tree, oprnds, i); | |
a70d6342 IR |
3154 | vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, |
3155 | &dt); | |
ebfd146a IR |
3156 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
3157 | VEC_replace(tree, dr_chain, i, vec_oprnd); | |
3158 | VEC_replace(tree, oprnds, i, vec_oprnd); | |
3159 | } | |
3160 | dataref_ptr = | |
3161 | bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE); | |
3162 | } | |
3163 | ||
3164 | if (strided_store) | |
3165 | { | |
3166 | result_chain = VEC_alloc (tree, heap, group_size); | |
3167 | /* Permute. */ | |
3168 | if (!vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
3169 | &result_chain)) | |
3170 | return false; | |
3171 | } | |
3172 | ||
3173 | next_stmt = first_stmt; | |
3174 | for (i = 0; i < vec_num; i++) | |
3175 | { | |
3176 | if (i > 0) | |
3177 | /* Bump the vector pointer. */ | |
3178 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
3179 | NULL_TREE); | |
3180 | ||
3181 | if (slp) | |
3182 | vec_oprnd = VEC_index (tree, vec_oprnds, i); | |
3183 | else if (strided_store) | |
3184 | /* For strided stores vectorized defs are interleaved in | |
3185 | vect_permute_store_chain(). */ | |
3186 | vec_oprnd = VEC_index (tree, result_chain, i); | |
3187 | ||
8f439681 RE |
3188 | if (aligned_access_p (first_dr)) |
3189 | data_ref = build_fold_indirect_ref (dataref_ptr); | |
3190 | else | |
3191 | { | |
3192 | int mis = DR_MISALIGNMENT (first_dr); | |
3193 | tree tmis = (mis == -1 ? size_zero_node : size_int (mis)); | |
3194 | tmis = size_binop (MULT_EXPR, tmis, size_int (BITS_PER_UNIT)); | |
3195 | data_ref = build2 (MISALIGNED_INDIRECT_REF, vectype, dataref_ptr, tmis); | |
3196 | } | |
3197 | ||
5006671f RG |
3198 | /* If accesses through a pointer to vectype do not alias the original |
3199 | memory reference we have a problem. This should never happen. */ | |
3200 | gcc_assert (alias_sets_conflict_p (get_alias_set (data_ref), | |
3201 | get_alias_set (gimple_assign_lhs (stmt)))); | |
ebfd146a IR |
3202 | |
3203 | /* Arguments are ready. Create the new vector stmt. */ | |
3204 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
3205 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3206 | mark_symbols_for_renaming (new_stmt); | |
3207 | ||
3208 | if (slp) | |
3209 | continue; | |
3210 | ||
3211 | if (j == 0) | |
3212 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3213 | else | |
3214 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3215 | ||
3216 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3217 | next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); | |
3218 | if (!next_stmt) | |
3219 | break; | |
3220 | } | |
3221 | } | |
3222 | ||
3223 | VEC_free (tree, heap, dr_chain); | |
3224 | VEC_free (tree, heap, oprnds); | |
3225 | if (result_chain) | |
3226 | VEC_free (tree, heap, result_chain); | |
3227 | ||
3228 | return true; | |
3229 | } | |
3230 | ||
3231 | /* vectorizable_load. | |
3232 | ||
3233 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that | |
3234 | can be vectorized. | |
3235 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3236 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3237 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3238 | ||
3239 | static bool | |
3240 | vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
3241 | slp_tree slp_node, slp_instance slp_node_instance) | |
3242 | { | |
3243 | tree scalar_dest; | |
3244 | tree vec_dest = NULL; | |
3245 | tree data_ref = NULL; | |
3246 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3247 | stmt_vec_info prev_stmt_info; | |
3248 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 3249 | struct loop *loop = NULL; |
ebfd146a | 3250 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 3251 | bool nested_in_vect_loop = false; |
ebfd146a IR |
3252 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; |
3253 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3254 | tree new_temp; | |
3255 | int mode; | |
3256 | gimple new_stmt = NULL; | |
3257 | tree dummy; | |
3258 | enum dr_alignment_support alignment_support_scheme; | |
3259 | tree dataref_ptr = NULL_TREE; | |
3260 | gimple ptr_incr; | |
3261 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
3262 | int ncopies; | |
3263 | int i, j, group_size; | |
3264 | tree msq = NULL_TREE, lsq; | |
3265 | tree offset = NULL_TREE; | |
3266 | tree realignment_token = NULL_TREE; | |
3267 | gimple phi = NULL; | |
3268 | VEC(tree,heap) *dr_chain = NULL; | |
3269 | bool strided_load = false; | |
3270 | gimple first_stmt; | |
3271 | tree scalar_type; | |
3272 | bool inv_p; | |
3273 | bool compute_in_loop = false; | |
3274 | struct loop *at_loop; | |
3275 | int vec_num; | |
3276 | bool slp = (slp_node != NULL); | |
3277 | bool slp_perm = false; | |
3278 | enum tree_code code; | |
a70d6342 IR |
3279 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3280 | int vf; | |
3281 | ||
3282 | if (loop_vinfo) | |
3283 | { | |
3284 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
3285 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
3286 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3287 | } | |
3288 | else | |
3289 | /* FORNOW: multiple types are not supported in basic block SLP. */ | |
3290 | vf = nunits; | |
ebfd146a IR |
3291 | |
3292 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3293 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3294 | case of SLP. */ | |
3295 | if (slp) | |
3296 | ncopies = 1; | |
3297 | else | |
3298 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3299 | ||
3300 | gcc_assert (ncopies >= 1); | |
3301 | ||
3302 | /* FORNOW. This restriction should be relaxed. */ | |
3303 | if (nested_in_vect_loop && ncopies > 1) | |
3304 | { | |
3305 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3306 | fprintf (vect_dump, "multiple types in nested loop."); | |
3307 | return false; | |
3308 | } | |
3309 | ||
a70d6342 | 3310 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3311 | return false; |
3312 | ||
8644a673 | 3313 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3314 | return false; |
3315 | ||
3316 | /* Is vectorizable load? */ | |
3317 | if (!is_gimple_assign (stmt)) | |
3318 | return false; | |
3319 | ||
3320 | scalar_dest = gimple_assign_lhs (stmt); | |
3321 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
3322 | return false; | |
3323 | ||
3324 | code = gimple_assign_rhs_code (stmt); | |
3325 | if (code != ARRAY_REF | |
3326 | && code != INDIRECT_REF | |
e9dbe7bb IR |
3327 | && code != COMPONENT_REF |
3328 | && code != IMAGPART_EXPR | |
3329 | && code != REALPART_EXPR) | |
ebfd146a IR |
3330 | return false; |
3331 | ||
3332 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
3333 | return false; | |
3334 | ||
3335 | scalar_type = TREE_TYPE (DR_REF (dr)); | |
3336 | mode = (int) TYPE_MODE (vectype); | |
3337 | ||
3338 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
3339 | (e.g. - data copies). */ | |
3340 | if (optab_handler (mov_optab, mode)->insn_code == CODE_FOR_nothing) | |
3341 | { | |
3342 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3343 | fprintf (vect_dump, "Aligned load, but unsupported type."); | |
3344 | return false; | |
3345 | } | |
3346 | ||
3347 | /* The vector component type needs to be trivially convertible to the | |
3348 | scalar lhs. This should always be the case. */ | |
3349 | if (!useless_type_conversion_p (TREE_TYPE (scalar_dest), TREE_TYPE (vectype))) | |
3350 | { | |
3351 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3352 | fprintf (vect_dump, "??? operands of different types"); | |
3353 | return false; | |
3354 | } | |
3355 | ||
3356 | /* Check if the load is a part of an interleaving chain. */ | |
3357 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) | |
3358 | { | |
3359 | strided_load = true; | |
3360 | /* FORNOW */ | |
3361 | gcc_assert (! nested_in_vect_loop); | |
3362 | ||
3363 | /* Check if interleaving is supported. */ | |
3364 | if (!vect_strided_load_supported (vectype) | |
3365 | && !PURE_SLP_STMT (stmt_info) && !slp) | |
3366 | return false; | |
3367 | } | |
3368 | ||
3369 | if (!vec_stmt) /* transformation not required. */ | |
3370 | { | |
3371 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
3372 | vect_model_load_cost (stmt_info, ncopies, NULL); | |
3373 | return true; | |
3374 | } | |
3375 | ||
3376 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3377 | fprintf (vect_dump, "transform load."); | |
3378 | ||
3379 | /** Transform. **/ | |
3380 | ||
3381 | if (strided_load) | |
3382 | { | |
3383 | first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
3384 | /* Check if the chain of loads is already vectorized. */ | |
3385 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt))) | |
3386 | { | |
3387 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
3388 | return true; | |
3389 | } | |
3390 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
3391 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
3392 | ||
3393 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
3394 | if (slp) | |
3395 | { | |
3396 | strided_load = false; | |
3397 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
a70d6342 IR |
3398 | if (SLP_INSTANCE_LOAD_PERMUTATION (slp_node_instance)) |
3399 | slp_perm = true; | |
3400 | } | |
ebfd146a IR |
3401 | else |
3402 | vec_num = group_size; | |
3403 | ||
3404 | dr_chain = VEC_alloc (tree, heap, vec_num); | |
3405 | } | |
3406 | else | |
3407 | { | |
3408 | first_stmt = stmt; | |
3409 | first_dr = dr; | |
3410 | group_size = vec_num = 1; | |
3411 | } | |
3412 | ||
3413 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr); | |
3414 | gcc_assert (alignment_support_scheme); | |
3415 | ||
3416 | /* In case the vectorization factor (VF) is bigger than the number | |
3417 | of elements that we can fit in a vectype (nunits), we have to generate | |
3418 | more than one vector stmt - i.e - we need to "unroll" the | |
3419 | vector stmt by a factor VF/nunits. In doing so, we record a pointer | |
3420 | from one copy of the vector stmt to the next, in the field | |
3421 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
3422 | stages to find the correct vector defs to be used when vectorizing | |
3423 | stmts that use the defs of the current stmt. The example below illustrates | |
3424 | the vectorization process when VF=16 and nunits=4 (i.e - we need to create | |
3425 | 4 vectorized stmts): | |
3426 | ||
3427 | before vectorization: | |
3428 | RELATED_STMT VEC_STMT | |
3429 | S1: x = memref - - | |
3430 | S2: z = x + 1 - - | |
3431 | ||
3432 | step 1: vectorize stmt S1: | |
3433 | We first create the vector stmt VS1_0, and, as usual, record a | |
3434 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
3435 | Next, we create the vector stmt VS1_1, and record a pointer to | |
3436 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
3437 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of | |
3438 | stmts and pointers: | |
3439 | RELATED_STMT VEC_STMT | |
3440 | VS1_0: vx0 = memref0 VS1_1 - | |
3441 | VS1_1: vx1 = memref1 VS1_2 - | |
3442 | VS1_2: vx2 = memref2 VS1_3 - | |
3443 | VS1_3: vx3 = memref3 - - | |
3444 | S1: x = load - VS1_0 | |
3445 | S2: z = x + 1 - - | |
3446 | ||
3447 | See in documentation in vect_get_vec_def_for_stmt_copy for how the | |
3448 | information we recorded in RELATED_STMT field is used to vectorize | |
3449 | stmt S2. */ | |
3450 | ||
3451 | /* In case of interleaving (non-unit strided access): | |
3452 | ||
3453 | S1: x2 = &base + 2 | |
3454 | S2: x0 = &base | |
3455 | S3: x1 = &base + 1 | |
3456 | S4: x3 = &base + 3 | |
3457 | ||
3458 | Vectorized loads are created in the order of memory accesses | |
3459 | starting from the access of the first stmt of the chain: | |
3460 | ||
3461 | VS1: vx0 = &base | |
3462 | VS2: vx1 = &base + vec_size*1 | |
3463 | VS3: vx3 = &base + vec_size*2 | |
3464 | VS4: vx4 = &base + vec_size*3 | |
3465 | ||
3466 | Then permutation statements are generated: | |
3467 | ||
3468 | VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 > | |
3469 | VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 > | |
3470 | ... | |
3471 | ||
3472 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
3473 | (the order of the data-refs in the output of vect_permute_load_chain | |
3474 | corresponds to the order of scalar stmts in the interleaving chain - see | |
3475 | the documentation of vect_permute_load_chain()). | |
3476 | The generation of permutation stmts and recording them in | |
3477 | STMT_VINFO_VEC_STMT is done in vect_transform_strided_load(). | |
3478 | ||
3479 | In case of both multiple types and interleaving, the vector loads and | |
3480 | permutation stmts above are created for every copy. The result vector stmts | |
3481 | are put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding | |
3482 | STMT_VINFO_RELATED_STMT for the next copies. */ | |
3483 | ||
3484 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
3485 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
3486 | we generate the following code: | |
3487 | p = initial_addr; | |
3488 | indx = 0; | |
3489 | loop { | |
3490 | p = p + indx * vectype_size; | |
3491 | vec_dest = *(p); | |
3492 | indx = indx + 1; | |
3493 | } | |
3494 | ||
3495 | Otherwise, the data reference is potentially unaligned on a target that | |
3496 | does not support unaligned accesses (dr_explicit_realign_optimized) - | |
3497 | then generate the following code, in which the data in each iteration is | |
3498 | obtained by two vector loads, one from the previous iteration, and one | |
3499 | from the current iteration: | |
3500 | p1 = initial_addr; | |
3501 | msq_init = *(floor(p1)) | |
3502 | p2 = initial_addr + VS - 1; | |
3503 | realignment_token = call target_builtin; | |
3504 | indx = 0; | |
3505 | loop { | |
3506 | p2 = p2 + indx * vectype_size | |
3507 | lsq = *(floor(p2)) | |
3508 | vec_dest = realign_load (msq, lsq, realignment_token) | |
3509 | indx = indx + 1; | |
3510 | msq = lsq; | |
3511 | } */ | |
3512 | ||
3513 | /* If the misalignment remains the same throughout the execution of the | |
3514 | loop, we can create the init_addr and permutation mask at the loop | |
3515 | preheader. Otherwise, it needs to be created inside the loop. | |
3516 | This can only occur when vectorizing memory accesses in the inner-loop | |
3517 | nested within an outer-loop that is being vectorized. */ | |
3518 | ||
a70d6342 | 3519 | if (loop && nested_in_vect_loop_p (loop, stmt) |
ebfd146a IR |
3520 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
3521 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) | |
3522 | { | |
3523 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
3524 | compute_in_loop = true; | |
3525 | } | |
3526 | ||
3527 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
3528 | || alignment_support_scheme == dr_explicit_realign) | |
3529 | && !compute_in_loop) | |
3530 | { | |
3531 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
3532 | alignment_support_scheme, NULL_TREE, | |
3533 | &at_loop); | |
3534 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
3535 | { | |
3536 | phi = SSA_NAME_DEF_STMT (msq); | |
3537 | offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
3538 | } | |
3539 | } | |
3540 | else | |
3541 | at_loop = loop; | |
3542 | ||
3543 | prev_stmt_info = NULL; | |
3544 | for (j = 0; j < ncopies; j++) | |
3545 | { | |
3546 | /* 1. Create the vector pointer update chain. */ | |
3547 | if (j == 0) | |
3548 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, | |
3549 | at_loop, offset, | |
3550 | &dummy, &ptr_incr, false, | |
5006671f | 3551 | &inv_p); |
ebfd146a IR |
3552 | else |
3553 | dataref_ptr = | |
3554 | bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE); | |
3555 | ||
3556 | for (i = 0; i < vec_num; i++) | |
3557 | { | |
3558 | if (i > 0) | |
3559 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
3560 | NULL_TREE); | |
3561 | ||
3562 | /* 2. Create the vector-load in the loop. */ | |
3563 | switch (alignment_support_scheme) | |
3564 | { | |
3565 | case dr_aligned: | |
3566 | gcc_assert (aligned_access_p (first_dr)); | |
3567 | data_ref = build_fold_indirect_ref (dataref_ptr); | |
3568 | break; | |
3569 | case dr_unaligned_supported: | |
3570 | { | |
3571 | int mis = DR_MISALIGNMENT (first_dr); | |
3572 | tree tmis = (mis == -1 ? size_zero_node : size_int (mis)); | |
3573 | ||
3574 | tmis = size_binop (MULT_EXPR, tmis, size_int(BITS_PER_UNIT)); | |
3575 | data_ref = | |
3576 | build2 (MISALIGNED_INDIRECT_REF, vectype, dataref_ptr, tmis); | |
3577 | break; | |
3578 | } | |
3579 | case dr_explicit_realign: | |
3580 | { | |
3581 | tree ptr, bump; | |
3582 | tree vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
3583 | ||
3584 | if (compute_in_loop) | |
3585 | msq = vect_setup_realignment (first_stmt, gsi, | |
3586 | &realignment_token, | |
3587 | dr_explicit_realign, | |
3588 | dataref_ptr, NULL); | |
3589 | ||
3590 | data_ref = build1 (ALIGN_INDIRECT_REF, vectype, dataref_ptr); | |
3591 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3592 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
3593 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3594 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5006671f RG |
3595 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); |
3596 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
ebfd146a | 3597 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
3598 | msq = new_temp; |
3599 | ||
3600 | bump = size_binop (MULT_EXPR, vs_minus_1, | |
3601 | TYPE_SIZE_UNIT (scalar_type)); | |
3602 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); | |
3603 | data_ref = build1 (ALIGN_INDIRECT_REF, vectype, ptr); | |
3604 | break; | |
3605 | } | |
3606 | case dr_explicit_realign_optimized: | |
3607 | data_ref = build1 (ALIGN_INDIRECT_REF, vectype, dataref_ptr); | |
3608 | break; | |
3609 | default: | |
3610 | gcc_unreachable (); | |
3611 | } | |
5006671f RG |
3612 | /* If accesses through a pointer to vectype do not alias the original |
3613 | memory reference we have a problem. This should never happen. */ | |
3614 | gcc_assert (alias_sets_conflict_p (get_alias_set (data_ref), | |
3615 | get_alias_set (gimple_assign_rhs1 (stmt)))); | |
ebfd146a IR |
3616 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
3617 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
3618 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3619 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3620 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3621 | mark_symbols_for_renaming (new_stmt); | |
3622 | ||
3623 | /* 3. Handle explicit realignment if necessary/supported. Create in | |
3624 | loop: vec_dest = realign_load (msq, lsq, realignment_token) */ | |
3625 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
3626 | || alignment_support_scheme == dr_explicit_realign) | |
3627 | { | |
3628 | tree tmp; | |
3629 | ||
3630 | lsq = gimple_assign_lhs (new_stmt); | |
3631 | if (!realignment_token) | |
3632 | realignment_token = dataref_ptr; | |
3633 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3634 | tmp = build3 (REALIGN_LOAD_EXPR, vectype, msq, lsq, | |
3635 | realignment_token); | |
3636 | new_stmt = gimple_build_assign (vec_dest, tmp); | |
3637 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3638 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3639 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3640 | ||
3641 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
3642 | { | |
3643 | gcc_assert (phi); | |
3644 | if (i == vec_num - 1 && j == ncopies - 1) | |
f5045c96 AM |
3645 | add_phi_arg (phi, lsq, loop_latch_edge (containing_loop), |
3646 | UNKNOWN_LOCATION); | |
ebfd146a IR |
3647 | msq = lsq; |
3648 | } | |
3649 | } | |
3650 | ||
3651 | /* 4. Handle invariant-load. */ | |
a70d6342 | 3652 | if (inv_p && !bb_vinfo) |
ebfd146a IR |
3653 | { |
3654 | gcc_assert (!strided_load); | |
3655 | gcc_assert (nested_in_vect_loop_p (loop, stmt)); | |
3656 | if (j == 0) | |
3657 | { | |
3658 | int k; | |
3659 | tree t = NULL_TREE; | |
3660 | tree vec_inv, bitpos, bitsize = TYPE_SIZE (scalar_type); | |
3661 | ||
3662 | /* CHECKME: bitpos depends on endianess? */ | |
3663 | bitpos = bitsize_zero_node; | |
3664 | vec_inv = build3 (BIT_FIELD_REF, scalar_type, new_temp, | |
3665 | bitsize, bitpos); | |
3666 | vec_dest = | |
3667 | vect_create_destination_var (scalar_dest, NULL_TREE); | |
3668 | new_stmt = gimple_build_assign (vec_dest, vec_inv); | |
3669 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3670 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3671 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3672 | ||
3673 | for (k = nunits - 1; k >= 0; --k) | |
3674 | t = tree_cons (NULL_TREE, new_temp, t); | |
3675 | /* FIXME: use build_constructor directly. */ | |
3676 | vec_inv = build_constructor_from_list (vectype, t); | |
3677 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
3678 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
3679 | } | |
3680 | else | |
3681 | gcc_unreachable (); /* FORNOW. */ | |
3682 | } | |
3683 | ||
3684 | /* Collect vector loads and later create their permutation in | |
3685 | vect_transform_strided_load (). */ | |
3686 | if (strided_load || slp_perm) | |
3687 | VEC_quick_push (tree, dr_chain, new_temp); | |
3688 | ||
3689 | /* Store vector loads in the corresponding SLP_NODE. */ | |
3690 | if (slp && !slp_perm) | |
3691 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
3692 | } | |
3693 | ||
3694 | if (slp && !slp_perm) | |
3695 | continue; | |
3696 | ||
3697 | if (slp_perm) | |
3698 | { | |
a70d6342 | 3699 | if (!vect_transform_slp_perm_load (stmt, dr_chain, gsi, vf, |
ebfd146a IR |
3700 | slp_node_instance, false)) |
3701 | { | |
3702 | VEC_free (tree, heap, dr_chain); | |
3703 | return false; | |
3704 | } | |
3705 | } | |
3706 | else | |
3707 | { | |
3708 | if (strided_load) | |
3709 | { | |
3710 | if (!vect_transform_strided_load (stmt, dr_chain, group_size, gsi)) | |
3711 | return false; | |
3712 | ||
3713 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
3714 | VEC_free (tree, heap, dr_chain); | |
3715 | dr_chain = VEC_alloc (tree, heap, group_size); | |
3716 | } | |
3717 | else | |
3718 | { | |
3719 | if (j == 0) | |
3720 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3721 | else | |
3722 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3723 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3724 | } | |
3725 | } | |
3726 | } | |
3727 | ||
3728 | if (dr_chain) | |
3729 | VEC_free (tree, heap, dr_chain); | |
3730 | ||
3731 | return true; | |
3732 | } | |
3733 | ||
3734 | /* Function vect_is_simple_cond. | |
3735 | ||
3736 | Input: | |
3737 | LOOP - the loop that is being vectorized. | |
3738 | COND - Condition that is checked for simple use. | |
3739 | ||
3740 | Returns whether a COND can be vectorized. Checks whether | |
3741 | condition operands are supportable using vec_is_simple_use. */ | |
3742 | ||
3743 | static bool | |
3744 | vect_is_simple_cond (tree cond, loop_vec_info loop_vinfo) | |
3745 | { | |
3746 | tree lhs, rhs; | |
3747 | tree def; | |
3748 | enum vect_def_type dt; | |
3749 | ||
3750 | if (!COMPARISON_CLASS_P (cond)) | |
3751 | return false; | |
3752 | ||
3753 | lhs = TREE_OPERAND (cond, 0); | |
3754 | rhs = TREE_OPERAND (cond, 1); | |
3755 | ||
3756 | if (TREE_CODE (lhs) == SSA_NAME) | |
3757 | { | |
3758 | gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); | |
a70d6342 IR |
3759 | if (!vect_is_simple_use (lhs, loop_vinfo, NULL, &lhs_def_stmt, &def, |
3760 | &dt)) | |
ebfd146a IR |
3761 | return false; |
3762 | } | |
3763 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
3764 | && TREE_CODE (lhs) != FIXED_CST) | |
3765 | return false; | |
3766 | ||
3767 | if (TREE_CODE (rhs) == SSA_NAME) | |
3768 | { | |
3769 | gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); | |
a70d6342 IR |
3770 | if (!vect_is_simple_use (rhs, loop_vinfo, NULL, &rhs_def_stmt, &def, |
3771 | &dt)) | |
ebfd146a IR |
3772 | return false; |
3773 | } | |
3774 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST | |
3775 | && TREE_CODE (rhs) != FIXED_CST) | |
3776 | return false; | |
3777 | ||
3778 | return true; | |
3779 | } | |
3780 | ||
3781 | /* vectorizable_condition. | |
3782 | ||
3783 | Check if STMT is conditional modify expression that can be vectorized. | |
3784 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3785 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
3786 | at GSI. |
3787 | ||
3788 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
3789 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
3790 | else caluse if it is 2). | |
ebfd146a IR |
3791 | |
3792 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3793 | ||
4bbe8262 | 3794 | bool |
ebfd146a | 3795 | vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi, |
4bbe8262 | 3796 | gimple *vec_stmt, tree reduc_def, int reduc_index) |
ebfd146a IR |
3797 | { |
3798 | tree scalar_dest = NULL_TREE; | |
3799 | tree vec_dest = NULL_TREE; | |
3800 | tree op = NULL_TREE; | |
3801 | tree cond_expr, then_clause, else_clause; | |
3802 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3803 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3804 | tree vec_cond_lhs, vec_cond_rhs, vec_then_clause, vec_else_clause; | |
3805 | tree vec_compare, vec_cond_expr; | |
3806 | tree new_temp; | |
3807 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3808 | enum machine_mode vec_mode; | |
3809 | tree def; | |
3810 | enum vect_def_type dt; | |
3811 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
3812 | int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3813 | enum tree_code code; | |
3814 | ||
a70d6342 IR |
3815 | /* FORNOW: unsupported in basic block SLP. */ |
3816 | gcc_assert (loop_vinfo); | |
3817 | ||
ebfd146a IR |
3818 | gcc_assert (ncopies >= 1); |
3819 | if (ncopies > 1) | |
3820 | return false; /* FORNOW */ | |
3821 | ||
3822 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
3823 | return false; | |
3824 | ||
4bbe8262 IR |
3825 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3826 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
3827 | && reduc_def)) | |
ebfd146a IR |
3828 | return false; |
3829 | ||
3830 | /* FORNOW: SLP not supported. */ | |
3831 | if (STMT_SLP_TYPE (stmt_info)) | |
3832 | return false; | |
3833 | ||
3834 | /* FORNOW: not yet supported. */ | |
4bbe8262 | 3835 | if (STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a IR |
3836 | { |
3837 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3838 | fprintf (vect_dump, "value used after loop."); | |
3839 | return false; | |
3840 | } | |
3841 | ||
3842 | /* Is vectorizable conditional operation? */ | |
3843 | if (!is_gimple_assign (stmt)) | |
3844 | return false; | |
3845 | ||
3846 | code = gimple_assign_rhs_code (stmt); | |
3847 | ||
3848 | if (code != COND_EXPR) | |
3849 | return false; | |
3850 | ||
3851 | gcc_assert (gimple_assign_single_p (stmt)); | |
3852 | op = gimple_assign_rhs1 (stmt); | |
3853 | cond_expr = TREE_OPERAND (op, 0); | |
3854 | then_clause = TREE_OPERAND (op, 1); | |
3855 | else_clause = TREE_OPERAND (op, 2); | |
3856 | ||
3857 | if (!vect_is_simple_cond (cond_expr, loop_vinfo)) | |
3858 | return false; | |
3859 | ||
3860 | /* We do not handle two different vector types for the condition | |
3861 | and the values. */ | |
3862 | if (TREE_TYPE (TREE_OPERAND (cond_expr, 0)) != TREE_TYPE (vectype)) | |
3863 | return false; | |
3864 | ||
3865 | if (TREE_CODE (then_clause) == SSA_NAME) | |
3866 | { | |
3867 | gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause); | |
a70d6342 | 3868 | if (!vect_is_simple_use (then_clause, loop_vinfo, NULL, |
ebfd146a IR |
3869 | &then_def_stmt, &def, &dt)) |
3870 | return false; | |
3871 | } | |
3872 | else if (TREE_CODE (then_clause) != INTEGER_CST | |
3873 | && TREE_CODE (then_clause) != REAL_CST | |
3874 | && TREE_CODE (then_clause) != FIXED_CST) | |
3875 | return false; | |
3876 | ||
3877 | if (TREE_CODE (else_clause) == SSA_NAME) | |
3878 | { | |
3879 | gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause); | |
a70d6342 | 3880 | if (!vect_is_simple_use (else_clause, loop_vinfo, NULL, |
ebfd146a IR |
3881 | &else_def_stmt, &def, &dt)) |
3882 | return false; | |
3883 | } | |
3884 | else if (TREE_CODE (else_clause) != INTEGER_CST | |
3885 | && TREE_CODE (else_clause) != REAL_CST | |
3886 | && TREE_CODE (else_clause) != FIXED_CST) | |
3887 | return false; | |
3888 | ||
3889 | ||
3890 | vec_mode = TYPE_MODE (vectype); | |
3891 | ||
3892 | if (!vec_stmt) | |
3893 | { | |
3894 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
8e7aa1f9 | 3895 | return expand_vec_cond_expr_p (TREE_TYPE (op), vec_mode); |
ebfd146a IR |
3896 | } |
3897 | ||
3898 | /* Transform */ | |
3899 | ||
3900 | /* Handle def. */ | |
3901 | scalar_dest = gimple_assign_lhs (stmt); | |
3902 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3903 | ||
3904 | /* Handle cond expr. */ | |
3905 | vec_cond_lhs = | |
3906 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), stmt, NULL); | |
3907 | vec_cond_rhs = | |
3908 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), stmt, NULL); | |
4bbe8262 IR |
3909 | if (reduc_index == 1) |
3910 | vec_then_clause = reduc_def; | |
3911 | else | |
3912 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, stmt, NULL); | |
3913 | if (reduc_index == 2) | |
3914 | vec_else_clause = reduc_def; | |
3915 | else | |
3916 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, stmt, NULL); | |
ebfd146a IR |
3917 | |
3918 | /* Arguments are ready. Create the new vector stmt. */ | |
3919 | vec_compare = build2 (TREE_CODE (cond_expr), vectype, | |
3920 | vec_cond_lhs, vec_cond_rhs); | |
3921 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, | |
3922 | vec_compare, vec_then_clause, vec_else_clause); | |
3923 | ||
3924 | *vec_stmt = gimple_build_assign (vec_dest, vec_cond_expr); | |
3925 | new_temp = make_ssa_name (vec_dest, *vec_stmt); | |
3926 | gimple_assign_set_lhs (*vec_stmt, new_temp); | |
3927 | vect_finish_stmt_generation (stmt, *vec_stmt, gsi); | |
3928 | ||
3929 | return true; | |
3930 | } | |
3931 | ||
3932 | ||
8644a673 | 3933 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
3934 | |
3935 | bool | |
a70d6342 | 3936 | vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 3937 | { |
8644a673 | 3938 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 IR |
3939 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3940 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); | |
ebfd146a | 3941 | bool ok; |
a70d6342 IR |
3942 | HOST_WIDE_INT dummy; |
3943 | tree scalar_type, vectype; | |
ebfd146a IR |
3944 | |
3945 | if (vect_print_dump_info (REPORT_DETAILS)) | |
ebfd146a | 3946 | { |
8644a673 IR |
3947 | fprintf (vect_dump, "==> examining statement: "); |
3948 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
3949 | } | |
ebfd146a | 3950 | |
1825a1f3 IR |
3951 | if (gimple_has_volatile_ops (stmt)) |
3952 | { | |
3953 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
3954 | fprintf (vect_dump, "not vectorized: stmt has volatile operands"); | |
3955 | ||
3956 | return false; | |
3957 | } | |
3958 | ||
a70d6342 | 3959 | /* Skip stmts that do not need to be vectorized. In loops this is expected |
8644a673 IR |
3960 | to include: |
3961 | - the COND_EXPR which is the loop exit condition | |
3962 | - any LABEL_EXPRs in the loop | |
a70d6342 | 3963 | - computations that are used only for array indexing or loop control. |
8644a673 IR |
3964 | In basic blocks we only analyze statements that are a part of some SLP |
3965 | instance, therefore, all the statements are relevant. */ | |
ebfd146a | 3966 | |
a70d6342 | 3967 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 3968 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a IR |
3969 | { |
3970 | if (vect_print_dump_info (REPORT_DETAILS)) | |
8644a673 | 3971 | fprintf (vect_dump, "irrelevant."); |
ebfd146a | 3972 | |
8644a673 IR |
3973 | return true; |
3974 | } | |
ebfd146a | 3975 | |
8644a673 IR |
3976 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
3977 | { | |
3978 | case vect_internal_def: | |
3979 | break; | |
ebfd146a | 3980 | |
8644a673 | 3981 | case vect_reduction_def: |
7c5222ff | 3982 | case vect_nested_cycle: |
a70d6342 | 3983 | gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer |
8644a673 | 3984 | || relevance == vect_used_in_outer_by_reduction |
a70d6342 | 3985 | || relevance == vect_unused_in_scope)); |
8644a673 IR |
3986 | break; |
3987 | ||
3988 | case vect_induction_def: | |
3989 | case vect_constant_def: | |
3990 | case vect_external_def: | |
3991 | case vect_unknown_def_type: | |
3992 | default: | |
3993 | gcc_unreachable (); | |
3994 | } | |
ebfd146a | 3995 | |
a70d6342 IR |
3996 | if (bb_vinfo) |
3997 | { | |
3998 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
3999 | ||
4000 | scalar_type = vect_get_smallest_scalar_type (stmt, &dummy, &dummy); | |
4001 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4002 | { | |
4003 | fprintf (vect_dump, "get vectype for scalar type: "); | |
4004 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
4005 | } | |
4006 | ||
4007 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4008 | if (!vectype) | |
4009 | { | |
4010 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4011 | { | |
4012 | fprintf (vect_dump, "not SLPed: unsupported data-type "); | |
4013 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
4014 | } | |
4015 | return false; | |
4016 | } | |
4017 | ||
4018 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4019 | { | |
4020 | fprintf (vect_dump, "vectype: "); | |
4021 | print_generic_expr (vect_dump, vectype, TDF_SLIM); | |
4022 | } | |
4023 | ||
4024 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
4025 | } | |
4026 | ||
8644a673 | 4027 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 4028 | { |
8644a673 IR |
4029 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
4030 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info)); | |
4031 | *need_to_vectorize = true; | |
ebfd146a IR |
4032 | } |
4033 | ||
8644a673 | 4034 | ok = true; |
a70d6342 IR |
4035 | if (!bb_vinfo |
4036 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
4037 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
8644a673 IR |
4038 | ok = (vectorizable_type_promotion (stmt, NULL, NULL, NULL) |
4039 | || vectorizable_type_demotion (stmt, NULL, NULL, NULL) | |
4040 | || vectorizable_conversion (stmt, NULL, NULL, NULL) | |
4041 | || vectorizable_operation (stmt, NULL, NULL, NULL) | |
4042 | || vectorizable_assignment (stmt, NULL, NULL, NULL) | |
4043 | || vectorizable_load (stmt, NULL, NULL, NULL, NULL) | |
4044 | || vectorizable_call (stmt, NULL, NULL) | |
4045 | || vectorizable_store (stmt, NULL, NULL, NULL) | |
4bbe8262 IR |
4046 | || vectorizable_reduction (stmt, NULL, NULL) |
4047 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0)); | |
a70d6342 IR |
4048 | else |
4049 | { | |
4050 | if (bb_vinfo) | |
4051 | ok = (vectorizable_operation (stmt, NULL, NULL, node) | |
4052 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
4053 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
4054 | || vectorizable_store (stmt, NULL, NULL, node)); | |
4055 | } | |
8644a673 IR |
4056 | |
4057 | if (!ok) | |
ebfd146a | 4058 | { |
8644a673 IR |
4059 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
4060 | { | |
4061 | fprintf (vect_dump, "not vectorized: relevant stmt not "); | |
4062 | fprintf (vect_dump, "supported: "); | |
4063 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4064 | } | |
a70d6342 | 4065 | |
ebfd146a IR |
4066 | return false; |
4067 | } | |
4068 | ||
a70d6342 IR |
4069 | if (bb_vinfo) |
4070 | return true; | |
4071 | ||
8644a673 IR |
4072 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
4073 | need extra handling, except for vectorizable reductions. */ | |
4074 | if (STMT_VINFO_LIVE_P (stmt_info) | |
4075 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
4076 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 4077 | |
8644a673 | 4078 | if (!ok) |
ebfd146a | 4079 | { |
8644a673 IR |
4080 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
4081 | { | |
4082 | fprintf (vect_dump, "not vectorized: live stmt not "); | |
4083 | fprintf (vect_dump, "supported: "); | |
4084 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4085 | } | |
a70d6342 | 4086 | |
8644a673 | 4087 | return false; |
ebfd146a IR |
4088 | } |
4089 | ||
8644a673 | 4090 | if (!PURE_SLP_STMT (stmt_info)) |
ebfd146a | 4091 | { |
a70d6342 IR |
4092 | /* Groups of strided accesses whose size is not a power of 2 are not |
4093 | vectorizable yet using loop-vectorization. Therefore, if this stmt | |
4094 | feeds non-SLP-able stmts (i.e., this stmt has to be both SLPed and | |
4095 | loop-based vectorized), the loop cannot be vectorized. */ | |
8644a673 IR |
4096 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) |
4097 | && exact_log2 (DR_GROUP_SIZE (vinfo_for_stmt ( | |
4098 | DR_GROUP_FIRST_DR (stmt_info)))) == -1) | |
ebfd146a | 4099 | { |
8644a673 IR |
4100 | if (vect_print_dump_info (REPORT_DETAILS)) |
4101 | { | |
4102 | fprintf (vect_dump, "not vectorized: the size of group " | |
4103 | "of strided accesses is not a power of 2"); | |
4104 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4105 | } | |
4106 | ||
ebfd146a IR |
4107 | return false; |
4108 | } | |
4109 | } | |
a70d6342 | 4110 | |
ebfd146a IR |
4111 | return true; |
4112 | } | |
4113 | ||
4114 | ||
4115 | /* Function vect_transform_stmt. | |
4116 | ||
4117 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
4118 | ||
4119 | bool | |
4120 | vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi, | |
4121 | bool *strided_store, slp_tree slp_node, | |
4122 | slp_instance slp_node_instance) | |
4123 | { | |
4124 | bool is_store = false; | |
4125 | gimple vec_stmt = NULL; | |
4126 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4127 | gimple orig_stmt_in_pattern; | |
4128 | bool done; | |
ebfd146a IR |
4129 | |
4130 | switch (STMT_VINFO_TYPE (stmt_info)) | |
4131 | { | |
4132 | case type_demotion_vec_info_type: | |
4133 | done = vectorizable_type_demotion (stmt, gsi, &vec_stmt, slp_node); | |
4134 | gcc_assert (done); | |
4135 | break; | |
4136 | ||
4137 | case type_promotion_vec_info_type: | |
4138 | done = vectorizable_type_promotion (stmt, gsi, &vec_stmt, slp_node); | |
4139 | gcc_assert (done); | |
4140 | break; | |
4141 | ||
4142 | case type_conversion_vec_info_type: | |
4143 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
4144 | gcc_assert (done); | |
4145 | break; | |
4146 | ||
4147 | case induc_vec_info_type: | |
4148 | gcc_assert (!slp_node); | |
4149 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
4150 | gcc_assert (done); | |
4151 | break; | |
4152 | ||
4153 | case op_vec_info_type: | |
4154 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
4155 | gcc_assert (done); | |
4156 | break; | |
4157 | ||
4158 | case assignment_vec_info_type: | |
4159 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
4160 | gcc_assert (done); | |
4161 | break; | |
4162 | ||
4163 | case load_vec_info_type: | |
4164 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, | |
4165 | slp_node_instance); | |
4166 | gcc_assert (done); | |
4167 | break; | |
4168 | ||
4169 | case store_vec_info_type: | |
4170 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
4171 | gcc_assert (done); | |
4172 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) && !slp_node) | |
4173 | { | |
4174 | /* In case of interleaving, the whole chain is vectorized when the | |
4175 | last store in the chain is reached. Store stmts before the last | |
4176 | one are skipped, and there vec_stmt_info shouldn't be freed | |
4177 | meanwhile. */ | |
4178 | *strided_store = true; | |
4179 | if (STMT_VINFO_VEC_STMT (stmt_info)) | |
4180 | is_store = true; | |
4181 | } | |
4182 | else | |
4183 | is_store = true; | |
4184 | break; | |
4185 | ||
4186 | case condition_vec_info_type: | |
4187 | gcc_assert (!slp_node); | |
4bbe8262 | 4188 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0); |
ebfd146a IR |
4189 | gcc_assert (done); |
4190 | break; | |
4191 | ||
4192 | case call_vec_info_type: | |
4193 | gcc_assert (!slp_node); | |
4194 | done = vectorizable_call (stmt, gsi, &vec_stmt); | |
4195 | break; | |
4196 | ||
4197 | case reduc_vec_info_type: | |
4198 | gcc_assert (!slp_node); | |
4199 | done = vectorizable_reduction (stmt, gsi, &vec_stmt); | |
4200 | gcc_assert (done); | |
4201 | break; | |
4202 | ||
4203 | default: | |
4204 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
4205 | { | |
4206 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4207 | fprintf (vect_dump, "stmt not supported."); | |
4208 | gcc_unreachable (); | |
4209 | } | |
4210 | } | |
4211 | ||
4212 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that | |
4213 | is being vectorized, but outside the immediately enclosing loop. */ | |
4214 | if (vec_stmt | |
a70d6342 IR |
4215 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
4216 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
4217 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
4218 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
4219 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
a70d6342 IR |
4220 | || STMT_VINFO_RELEVANT (stmt_info) == |
4221 | vect_used_in_outer_by_reduction)) | |
ebfd146a | 4222 | { |
a70d6342 IR |
4223 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
4224 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
4225 | imm_use_iterator imm_iter; |
4226 | use_operand_p use_p; | |
4227 | tree scalar_dest; | |
4228 | gimple exit_phi; | |
4229 | ||
4230 | if (vect_print_dump_info (REPORT_DETAILS)) | |
a70d6342 | 4231 | fprintf (vect_dump, "Record the vdef for outer-loop vectorization."); |
ebfd146a IR |
4232 | |
4233 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
4234 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
4235 | STMT). */ | |
4236 | if (gimple_code (stmt) == GIMPLE_PHI) | |
4237 | scalar_dest = PHI_RESULT (stmt); | |
4238 | else | |
4239 | scalar_dest = gimple_assign_lhs (stmt); | |
4240 | ||
4241 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
4242 | { | |
4243 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
4244 | { | |
4245 | exit_phi = USE_STMT (use_p); | |
4246 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
4247 | } | |
4248 | } | |
4249 | } | |
4250 | ||
4251 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
4252 | being vectorized. */ | |
4253 | if (STMT_VINFO_LIVE_P (stmt_info) | |
4254 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
4255 | { | |
4256 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
4257 | gcc_assert (done); | |
4258 | } | |
4259 | ||
4260 | if (vec_stmt) | |
4261 | { | |
4262 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; | |
4263 | orig_stmt_in_pattern = STMT_VINFO_RELATED_STMT (stmt_info); | |
4264 | if (orig_stmt_in_pattern) | |
4265 | { | |
4266 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt_in_pattern); | |
4267 | /* STMT was inserted by the vectorizer to replace a computation idiom. | |
4268 | ORIG_STMT_IN_PATTERN is a stmt in the original sequence that | |
4269 | computed this idiom. We need to record a pointer to VEC_STMT in | |
4270 | the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the | |
4271 | documentation of vect_pattern_recog. */ | |
4272 | if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo)) | |
4273 | { | |
4274 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) == stmt); | |
4275 | STMT_VINFO_VEC_STMT (stmt_vinfo) = vec_stmt; | |
4276 | } | |
4277 | } | |
4278 | } | |
4279 | ||
4280 | return is_store; | |
4281 | } | |
4282 | ||
4283 | ||
4284 | /* Remove a group of stores (for SLP or interleaving), free their | |
4285 | stmt_vec_info. */ | |
4286 | ||
4287 | void | |
4288 | vect_remove_stores (gimple first_stmt) | |
4289 | { | |
4290 | gimple next = first_stmt; | |
4291 | gimple tmp; | |
4292 | gimple_stmt_iterator next_si; | |
4293 | ||
4294 | while (next) | |
4295 | { | |
4296 | /* Free the attached stmt_vec_info and remove the stmt. */ | |
4297 | next_si = gsi_for_stmt (next); | |
4298 | gsi_remove (&next_si, true); | |
4299 | tmp = DR_GROUP_NEXT_DR (vinfo_for_stmt (next)); | |
4300 | free_stmt_vec_info (next); | |
4301 | next = tmp; | |
4302 | } | |
4303 | } | |
4304 | ||
4305 | ||
4306 | /* Function new_stmt_vec_info. | |
4307 | ||
4308 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
4309 | ||
4310 | stmt_vec_info | |
a70d6342 IR |
4311 | new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo, |
4312 | bb_vec_info bb_vinfo) | |
ebfd146a IR |
4313 | { |
4314 | stmt_vec_info res; | |
4315 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
4316 | ||
4317 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
4318 | STMT_VINFO_STMT (res) = stmt; | |
4319 | STMT_VINFO_LOOP_VINFO (res) = loop_vinfo; | |
a70d6342 | 4320 | STMT_VINFO_BB_VINFO (res) = bb_vinfo; |
8644a673 | 4321 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
4322 | STMT_VINFO_LIVE_P (res) = false; |
4323 | STMT_VINFO_VECTYPE (res) = NULL; | |
4324 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4325 | STMT_VINFO_IN_PATTERN_P (res) = false; | |
4326 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
4327 | STMT_VINFO_DATA_REF (res) = NULL; | |
4328 | ||
4329 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
4330 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
4331 | STMT_VINFO_DR_INIT (res) = NULL; | |
4332 | STMT_VINFO_DR_STEP (res) = NULL; | |
4333 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
4334 | ||
4335 | if (gimple_code (stmt) == GIMPLE_PHI | |
4336 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
4337 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
4338 | else | |
8644a673 IR |
4339 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
4340 | ||
ebfd146a IR |
4341 | STMT_VINFO_SAME_ALIGN_REFS (res) = VEC_alloc (dr_p, heap, 5); |
4342 | STMT_VINFO_INSIDE_OF_LOOP_COST (res) = 0; | |
4343 | STMT_VINFO_OUTSIDE_OF_LOOP_COST (res) = 0; | |
32e8bb8e | 4344 | STMT_SLP_TYPE (res) = loop_vect; |
ebfd146a IR |
4345 | DR_GROUP_FIRST_DR (res) = NULL; |
4346 | DR_GROUP_NEXT_DR (res) = NULL; | |
4347 | DR_GROUP_SIZE (res) = 0; | |
4348 | DR_GROUP_STORE_COUNT (res) = 0; | |
4349 | DR_GROUP_GAP (res) = 0; | |
4350 | DR_GROUP_SAME_DR_STMT (res) = NULL; | |
4351 | DR_GROUP_READ_WRITE_DEPENDENCE (res) = false; | |
4352 | ||
4353 | return res; | |
4354 | } | |
4355 | ||
4356 | ||
4357 | /* Create a hash table for stmt_vec_info. */ | |
4358 | ||
4359 | void | |
4360 | init_stmt_vec_info_vec (void) | |
4361 | { | |
4362 | gcc_assert (!stmt_vec_info_vec); | |
4363 | stmt_vec_info_vec = VEC_alloc (vec_void_p, heap, 50); | |
4364 | } | |
4365 | ||
4366 | ||
4367 | /* Free hash table for stmt_vec_info. */ | |
4368 | ||
4369 | void | |
4370 | free_stmt_vec_info_vec (void) | |
4371 | { | |
4372 | gcc_assert (stmt_vec_info_vec); | |
4373 | VEC_free (vec_void_p, heap, stmt_vec_info_vec); | |
4374 | } | |
4375 | ||
4376 | ||
4377 | /* Free stmt vectorization related info. */ | |
4378 | ||
4379 | void | |
4380 | free_stmt_vec_info (gimple stmt) | |
4381 | { | |
4382 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4383 | ||
4384 | if (!stmt_info) | |
4385 | return; | |
4386 | ||
4387 | VEC_free (dr_p, heap, STMT_VINFO_SAME_ALIGN_REFS (stmt_info)); | |
4388 | set_vinfo_for_stmt (stmt, NULL); | |
4389 | free (stmt_info); | |
4390 | } | |
4391 | ||
4392 | ||
4393 | /* Function get_vectype_for_scalar_type. | |
4394 | ||
4395 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
4396 | by the target. */ | |
4397 | ||
4398 | tree | |
4399 | get_vectype_for_scalar_type (tree scalar_type) | |
4400 | { | |
4401 | enum machine_mode inner_mode = TYPE_MODE (scalar_type); | |
4402 | int nbytes = GET_MODE_SIZE (inner_mode); | |
4403 | int nunits; | |
4404 | tree vectype; | |
4405 | ||
4406 | if (nbytes == 0 || nbytes >= UNITS_PER_SIMD_WORD (inner_mode)) | |
4407 | return NULL_TREE; | |
4408 | ||
4409 | /* FORNOW: Only a single vector size per mode (UNITS_PER_SIMD_WORD) | |
4410 | is expected. */ | |
4411 | nunits = UNITS_PER_SIMD_WORD (inner_mode) / nbytes; | |
4412 | ||
4413 | vectype = build_vector_type (scalar_type, nunits); | |
4414 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4415 | { | |
4416 | fprintf (vect_dump, "get vectype with %d units of type ", nunits); | |
4417 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
4418 | } | |
4419 | ||
4420 | if (!vectype) | |
4421 | return NULL_TREE; | |
4422 | ||
4423 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4424 | { | |
4425 | fprintf (vect_dump, "vectype: "); | |
4426 | print_generic_expr (vect_dump, vectype, TDF_SLIM); | |
4427 | } | |
4428 | ||
4429 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
4430 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
4431 | { | |
4432 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4433 | fprintf (vect_dump, "mode not supported by target."); | |
4434 | return NULL_TREE; | |
4435 | } | |
4436 | ||
4437 | return vectype; | |
4438 | } | |
4439 | ||
4440 | /* Function vect_is_simple_use. | |
4441 | ||
4442 | Input: | |
a70d6342 IR |
4443 | LOOP_VINFO - the vect info of the loop that is being vectorized. |
4444 | BB_VINFO - the vect info of the basic block that is being vectorized. | |
4445 | OPERAND - operand of a stmt in the loop or bb. | |
ebfd146a IR |
4446 | DEF - the defining stmt in case OPERAND is an SSA_NAME. |
4447 | ||
4448 | Returns whether a stmt with OPERAND can be vectorized. | |
a70d6342 IR |
4449 | For loops, supportable operands are constants, loop invariants, and operands |
4450 | that are defined by the current iteration of the loop. Unsupportable | |
4451 | operands are those that are defined by a previous iteration of the loop (as | |
4452 | is the case in reduction/induction computations). | |
4453 | For basic blocks, supportable operands are constants and bb invariants. | |
4454 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
4455 | |
4456 | bool | |
a70d6342 IR |
4457 | vect_is_simple_use (tree operand, loop_vec_info loop_vinfo, |
4458 | bb_vec_info bb_vinfo, gimple *def_stmt, | |
ebfd146a IR |
4459 | tree *def, enum vect_def_type *dt) |
4460 | { | |
4461 | basic_block bb; | |
4462 | stmt_vec_info stmt_vinfo; | |
a70d6342 IR |
4463 | struct loop *loop = NULL; |
4464 | ||
4465 | if (loop_vinfo) | |
4466 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
4467 | |
4468 | *def_stmt = NULL; | |
4469 | *def = NULL_TREE; | |
4470 | ||
4471 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4472 | { | |
4473 | fprintf (vect_dump, "vect_is_simple_use: operand "); | |
4474 | print_generic_expr (vect_dump, operand, TDF_SLIM); | |
4475 | } | |
4476 | ||
4477 | if (TREE_CODE (operand) == INTEGER_CST || TREE_CODE (operand) == REAL_CST) | |
4478 | { | |
4479 | *dt = vect_constant_def; | |
4480 | return true; | |
4481 | } | |
a70d6342 | 4482 | |
ebfd146a IR |
4483 | if (is_gimple_min_invariant (operand)) |
4484 | { | |
4485 | *def = operand; | |
8644a673 | 4486 | *dt = vect_external_def; |
ebfd146a IR |
4487 | return true; |
4488 | } | |
4489 | ||
4490 | if (TREE_CODE (operand) == PAREN_EXPR) | |
4491 | { | |
4492 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4493 | fprintf (vect_dump, "non-associatable copy."); | |
4494 | operand = TREE_OPERAND (operand, 0); | |
4495 | } | |
a70d6342 | 4496 | |
ebfd146a IR |
4497 | if (TREE_CODE (operand) != SSA_NAME) |
4498 | { | |
4499 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4500 | fprintf (vect_dump, "not ssa-name."); | |
4501 | return false; | |
4502 | } | |
4503 | ||
4504 | *def_stmt = SSA_NAME_DEF_STMT (operand); | |
4505 | if (*def_stmt == NULL) | |
4506 | { | |
4507 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4508 | fprintf (vect_dump, "no def_stmt."); | |
4509 | return false; | |
4510 | } | |
4511 | ||
4512 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4513 | { | |
4514 | fprintf (vect_dump, "def_stmt: "); | |
4515 | print_gimple_stmt (vect_dump, *def_stmt, 0, TDF_SLIM); | |
4516 | } | |
4517 | ||
8644a673 | 4518 | /* Empty stmt is expected only in case of a function argument. |
ebfd146a IR |
4519 | (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */ |
4520 | if (gimple_nop_p (*def_stmt)) | |
4521 | { | |
4522 | *def = operand; | |
8644a673 | 4523 | *dt = vect_external_def; |
ebfd146a IR |
4524 | return true; |
4525 | } | |
4526 | ||
4527 | bb = gimple_bb (*def_stmt); | |
a70d6342 IR |
4528 | |
4529 | if ((loop && !flow_bb_inside_loop_p (loop, bb)) | |
4530 | || (!loop && bb != BB_VINFO_BB (bb_vinfo)) | |
4531 | || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI)) | |
8644a673 | 4532 | *dt = vect_external_def; |
ebfd146a IR |
4533 | else |
4534 | { | |
4535 | stmt_vinfo = vinfo_for_stmt (*def_stmt); | |
4536 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
4537 | } | |
4538 | ||
4539 | if (*dt == vect_unknown_def_type) | |
4540 | { | |
4541 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4542 | fprintf (vect_dump, "Unsupported pattern."); | |
4543 | return false; | |
4544 | } | |
4545 | ||
4546 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4547 | fprintf (vect_dump, "type of def: %d.",*dt); | |
4548 | ||
4549 | switch (gimple_code (*def_stmt)) | |
4550 | { | |
4551 | case GIMPLE_PHI: | |
4552 | *def = gimple_phi_result (*def_stmt); | |
4553 | break; | |
4554 | ||
4555 | case GIMPLE_ASSIGN: | |
4556 | *def = gimple_assign_lhs (*def_stmt); | |
4557 | break; | |
4558 | ||
4559 | case GIMPLE_CALL: | |
4560 | *def = gimple_call_lhs (*def_stmt); | |
4561 | if (*def != NULL) | |
4562 | break; | |
4563 | /* FALLTHRU */ | |
4564 | default: | |
4565 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4566 | fprintf (vect_dump, "unsupported defining stmt: "); | |
4567 | return false; | |
4568 | } | |
4569 | ||
4570 | return true; | |
4571 | } | |
4572 | ||
4573 | ||
4574 | /* Function supportable_widening_operation | |
4575 | ||
4576 | Check whether an operation represented by the code CODE is a | |
4577 | widening operation that is supported by the target platform in | |
4578 | vector form (i.e., when operating on arguments of type VECTYPE). | |
4579 | ||
4580 | Widening operations we currently support are NOP (CONVERT), FLOAT | |
4581 | and WIDEN_MULT. This function checks if these operations are supported | |
4582 | by the target platform either directly (via vector tree-codes), or via | |
4583 | target builtins. | |
4584 | ||
4585 | Output: | |
4586 | - CODE1 and CODE2 are codes of vector operations to be used when | |
4587 | vectorizing the operation, if available. | |
4588 | - DECL1 and DECL2 are decls of target builtin functions to be used | |
4589 | when vectorizing the operation, if available. In this case, | |
4590 | CODE1 and CODE2 are CALL_EXPR. | |
4591 | - MULTI_STEP_CVT determines the number of required intermediate steps in | |
4592 | case of multi-step conversion (like char->short->int - in that case | |
4593 | MULTI_STEP_CVT will be 1). | |
4594 | - INTERM_TYPES contains the intermediate type required to perform the | |
4595 | widening operation (short in the above example). */ | |
4596 | ||
4597 | bool | |
4598 | supportable_widening_operation (enum tree_code code, gimple stmt, tree vectype, | |
4599 | tree *decl1, tree *decl2, | |
4600 | enum tree_code *code1, enum tree_code *code2, | |
4601 | int *multi_step_cvt, | |
4602 | VEC (tree, heap) **interm_types) | |
4603 | { | |
4604 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4605 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4606 | struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info); | |
4607 | bool ordered_p; | |
4608 | enum machine_mode vec_mode; | |
81f40b79 | 4609 | enum insn_code icode1, icode2; |
ebfd146a IR |
4610 | optab optab1, optab2; |
4611 | tree type = gimple_expr_type (stmt); | |
4612 | tree wide_vectype = get_vectype_for_scalar_type (type); | |
4613 | enum tree_code c1, c2; | |
4614 | ||
4615 | /* The result of a vectorized widening operation usually requires two vectors | |
4616 | (because the widened results do not fit int one vector). The generated | |
4617 | vector results would normally be expected to be generated in the same | |
4618 | order as in the original scalar computation, i.e. if 8 results are | |
4619 | generated in each vector iteration, they are to be organized as follows: | |
4620 | vect1: [res1,res2,res3,res4], vect2: [res5,res6,res7,res8]. | |
4621 | ||
4622 | However, in the special case that the result of the widening operation is | |
4623 | used in a reduction computation only, the order doesn't matter (because | |
4624 | when vectorizing a reduction we change the order of the computation). | |
4625 | Some targets can take advantage of this and generate more efficient code. | |
4626 | For example, targets like Altivec, that support widen_mult using a sequence | |
4627 | of {mult_even,mult_odd} generate the following vectors: | |
4628 | vect1: [res1,res3,res5,res7], vect2: [res2,res4,res6,res8]. | |
4629 | ||
4630 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
4631 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
4632 | iterations in parallel). We therefore don't allow to change the order | |
4633 | of the computation in the inner-loop during outer-loop vectorization. */ | |
4634 | ||
4635 | if (STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
4636 | && !nested_in_vect_loop_p (vect_loop, stmt)) | |
4637 | ordered_p = false; | |
4638 | else | |
4639 | ordered_p = true; | |
4640 | ||
4641 | if (!ordered_p | |
4642 | && code == WIDEN_MULT_EXPR | |
4643 | && targetm.vectorize.builtin_mul_widen_even | |
4644 | && targetm.vectorize.builtin_mul_widen_even (vectype) | |
4645 | && targetm.vectorize.builtin_mul_widen_odd | |
4646 | && targetm.vectorize.builtin_mul_widen_odd (vectype)) | |
4647 | { | |
4648 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4649 | fprintf (vect_dump, "Unordered widening operation detected."); | |
4650 | ||
4651 | *code1 = *code2 = CALL_EXPR; | |
4652 | *decl1 = targetm.vectorize.builtin_mul_widen_even (vectype); | |
4653 | *decl2 = targetm.vectorize.builtin_mul_widen_odd (vectype); | |
4654 | return true; | |
4655 | } | |
4656 | ||
4657 | switch (code) | |
4658 | { | |
4659 | case WIDEN_MULT_EXPR: | |
4660 | if (BYTES_BIG_ENDIAN) | |
4661 | { | |
4662 | c1 = VEC_WIDEN_MULT_HI_EXPR; | |
4663 | c2 = VEC_WIDEN_MULT_LO_EXPR; | |
4664 | } | |
4665 | else | |
4666 | { | |
4667 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
4668 | c1 = VEC_WIDEN_MULT_LO_EXPR; | |
4669 | } | |
4670 | break; | |
4671 | ||
4672 | CASE_CONVERT: | |
4673 | if (BYTES_BIG_ENDIAN) | |
4674 | { | |
4675 | c1 = VEC_UNPACK_HI_EXPR; | |
4676 | c2 = VEC_UNPACK_LO_EXPR; | |
4677 | } | |
4678 | else | |
4679 | { | |
4680 | c2 = VEC_UNPACK_HI_EXPR; | |
4681 | c1 = VEC_UNPACK_LO_EXPR; | |
4682 | } | |
4683 | break; | |
4684 | ||
4685 | case FLOAT_EXPR: | |
4686 | if (BYTES_BIG_ENDIAN) | |
4687 | { | |
4688 | c1 = VEC_UNPACK_FLOAT_HI_EXPR; | |
4689 | c2 = VEC_UNPACK_FLOAT_LO_EXPR; | |
4690 | } | |
4691 | else | |
4692 | { | |
4693 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
4694 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; | |
4695 | } | |
4696 | break; | |
4697 | ||
4698 | case FIX_TRUNC_EXPR: | |
4699 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
4700 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
4701 | computing the operation. */ | |
4702 | return false; | |
4703 | ||
4704 | default: | |
4705 | gcc_unreachable (); | |
4706 | } | |
4707 | ||
4708 | if (code == FIX_TRUNC_EXPR) | |
4709 | { | |
4710 | /* The signedness is determined from output operand. */ | |
4711 | optab1 = optab_for_tree_code (c1, type, optab_default); | |
4712 | optab2 = optab_for_tree_code (c2, type, optab_default); | |
4713 | } | |
4714 | else | |
4715 | { | |
4716 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
4717 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
4718 | } | |
4719 | ||
4720 | if (!optab1 || !optab2) | |
4721 | return false; | |
4722 | ||
4723 | vec_mode = TYPE_MODE (vectype); | |
4724 | if ((icode1 = optab_handler (optab1, vec_mode)->insn_code) == CODE_FOR_nothing | |
4725 | || (icode2 = optab_handler (optab2, vec_mode)->insn_code) | |
4726 | == CODE_FOR_nothing) | |
4727 | return false; | |
4728 | ||
4729 | /* Check if it's a multi-step conversion that can be done using intermediate | |
4730 | types. */ | |
4731 | if (insn_data[icode1].operand[0].mode != TYPE_MODE (wide_vectype) | |
4732 | || insn_data[icode2].operand[0].mode != TYPE_MODE (wide_vectype)) | |
4733 | { | |
4734 | int i; | |
4735 | tree prev_type = vectype, intermediate_type; | |
4736 | enum machine_mode intermediate_mode, prev_mode = vec_mode; | |
4737 | optab optab3, optab4; | |
4738 | ||
4739 | if (!CONVERT_EXPR_CODE_P (code)) | |
4740 | return false; | |
4741 | ||
4742 | *code1 = c1; | |
4743 | *code2 = c2; | |
4744 | ||
4745 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS | |
4746 | intermediate steps in promotion sequence. We try MAX_INTERM_CVT_STEPS | |
4747 | to get to NARROW_VECTYPE, and fail if we do not. */ | |
4748 | *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS); | |
4749 | for (i = 0; i < 3; i++) | |
4750 | { | |
4751 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
4752 | intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode, | |
4753 | TYPE_UNSIGNED (prev_type)); | |
4754 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
4755 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
4756 | ||
4757 | if (!optab3 || !optab4 | |
4758 | || (icode1 = optab1->handlers[(int) prev_mode].insn_code) | |
4759 | == CODE_FOR_nothing | |
4760 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
4761 | || (icode2 = optab2->handlers[(int) prev_mode].insn_code) | |
4762 | == CODE_FOR_nothing | |
4763 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
4764 | || (icode1 = optab3->handlers[(int) intermediate_mode].insn_code) | |
4765 | == CODE_FOR_nothing | |
4766 | || (icode2 = optab4->handlers[(int) intermediate_mode].insn_code) | |
4767 | == CODE_FOR_nothing) | |
4768 | return false; | |
4769 | ||
4770 | VEC_quick_push (tree, *interm_types, intermediate_type); | |
4771 | (*multi_step_cvt)++; | |
4772 | ||
4773 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
4774 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
4775 | return true; | |
4776 | ||
4777 | prev_type = intermediate_type; | |
4778 | prev_mode = intermediate_mode; | |
4779 | } | |
4780 | ||
4781 | return false; | |
4782 | } | |
4783 | ||
4784 | *code1 = c1; | |
4785 | *code2 = c2; | |
4786 | return true; | |
4787 | } | |
4788 | ||
4789 | ||
4790 | /* Function supportable_narrowing_operation | |
4791 | ||
4792 | Check whether an operation represented by the code CODE is a | |
4793 | narrowing operation that is supported by the target platform in | |
4794 | vector form (i.e., when operating on arguments of type VECTYPE). | |
4795 | ||
4796 | Narrowing operations we currently support are NOP (CONVERT) and | |
4797 | FIX_TRUNC. This function checks if these operations are supported by | |
4798 | the target platform directly via vector tree-codes. | |
4799 | ||
4800 | Output: | |
4801 | - CODE1 is the code of a vector operation to be used when | |
4802 | vectorizing the operation, if available. | |
4803 | - MULTI_STEP_CVT determines the number of required intermediate steps in | |
4804 | case of multi-step conversion (like int->short->char - in that case | |
4805 | MULTI_STEP_CVT will be 1). | |
4806 | - INTERM_TYPES contains the intermediate type required to perform the | |
4807 | narrowing operation (short in the above example). */ | |
4808 | ||
4809 | bool | |
4810 | supportable_narrowing_operation (enum tree_code code, | |
4811 | const_gimple stmt, tree vectype, | |
4812 | enum tree_code *code1, int *multi_step_cvt, | |
4813 | VEC (tree, heap) **interm_types) | |
4814 | { | |
4815 | enum machine_mode vec_mode; | |
4816 | enum insn_code icode1; | |
4817 | optab optab1, interm_optab; | |
4818 | tree type = gimple_expr_type (stmt); | |
4819 | tree narrow_vectype = get_vectype_for_scalar_type (type); | |
4820 | enum tree_code c1; | |
4821 | tree intermediate_type, prev_type; | |
4822 | int i; | |
4823 | ||
4824 | switch (code) | |
4825 | { | |
4826 | CASE_CONVERT: | |
4827 | c1 = VEC_PACK_TRUNC_EXPR; | |
4828 | break; | |
4829 | ||
4830 | case FIX_TRUNC_EXPR: | |
4831 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
4832 | break; | |
4833 | ||
4834 | case FLOAT_EXPR: | |
4835 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
4836 | tree code and optabs used for computing the operation. */ | |
4837 | return false; | |
4838 | ||
4839 | default: | |
4840 | gcc_unreachable (); | |
4841 | } | |
4842 | ||
4843 | if (code == FIX_TRUNC_EXPR) | |
4844 | /* The signedness is determined from output operand. */ | |
4845 | optab1 = optab_for_tree_code (c1, type, optab_default); | |
4846 | else | |
4847 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
4848 | ||
4849 | if (!optab1) | |
4850 | return false; | |
4851 | ||
4852 | vec_mode = TYPE_MODE (vectype); | |
4853 | if ((icode1 = optab_handler (optab1, vec_mode)->insn_code) | |
4854 | == CODE_FOR_nothing) | |
4855 | return false; | |
4856 | ||
4857 | /* Check if it's a multi-step conversion that can be done using intermediate | |
4858 | types. */ | |
4859 | if (insn_data[icode1].operand[0].mode != TYPE_MODE (narrow_vectype)) | |
4860 | { | |
4861 | enum machine_mode intermediate_mode, prev_mode = vec_mode; | |
4862 | ||
4863 | *code1 = c1; | |
4864 | prev_type = vectype; | |
4865 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS | |
4866 | intermediate steps in promotion sequence. We try MAX_INTERM_CVT_STEPS | |
4867 | to get to NARROW_VECTYPE, and fail if we do not. */ | |
4868 | *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS); | |
4869 | for (i = 0; i < 3; i++) | |
4870 | { | |
4871 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
4872 | intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode, | |
4873 | TYPE_UNSIGNED (prev_type)); | |
4874 | interm_optab = optab_for_tree_code (c1, intermediate_type, | |
4875 | optab_default); | |
4876 | if (!interm_optab | |
4877 | || (icode1 = optab1->handlers[(int) prev_mode].insn_code) | |
4878 | == CODE_FOR_nothing | |
4879 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
4880 | || (icode1 | |
4881 | = interm_optab->handlers[(int) intermediate_mode].insn_code) | |
4882 | == CODE_FOR_nothing) | |
4883 | return false; | |
4884 | ||
4885 | VEC_quick_push (tree, *interm_types, intermediate_type); | |
4886 | (*multi_step_cvt)++; | |
4887 | ||
4888 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
4889 | return true; | |
4890 | ||
4891 | prev_type = intermediate_type; | |
4892 | prev_mode = intermediate_mode; | |
4893 | } | |
4894 | ||
4895 | return false; | |
4896 | } | |
4897 | ||
4898 | *code1 = c1; | |
4899 | return true; | |
4900 | } |