]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
62f7fd21 MM |
2 | Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010 |
3 | Free Software Foundation, Inc. | |
b8698a0f | 4 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
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" | |
cf835838 JM |
31 | #include "tree-pretty-print.h" |
32 | #include "gimple-pretty-print.h" | |
ebfd146a IR |
33 | #include "tree-flow.h" |
34 | #include "tree-dump.h" | |
35 | #include "cfgloop.h" | |
36 | #include "cfglayout.h" | |
37 | #include "expr.h" | |
38 | #include "recog.h" | |
39 | #include "optabs.h" | |
40 | #include "toplev.h" | |
41 | #include "tree-vectorizer.h" | |
42 | #include "langhooks.h" | |
43 | ||
44 | ||
45 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ | |
46 | ||
47 | /* Function vect_mark_relevant. | |
48 | ||
49 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
50 | ||
51 | static void | |
52 | vect_mark_relevant (VEC(gimple,heap) **worklist, gimple stmt, | |
53 | enum vect_relevant relevant, bool live_p) | |
54 | { | |
55 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
56 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
57 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
58 | ||
59 | if (vect_print_dump_info (REPORT_DETAILS)) | |
60 | fprintf (vect_dump, "mark relevant %d, live %d.", relevant, live_p); | |
61 | ||
62 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
63 | { | |
64 | gimple pattern_stmt; | |
65 | ||
b8698a0f | 66 | /* This is the last stmt in a sequence that was detected as a |
ebfd146a IR |
67 | pattern that can potentially be vectorized. Don't mark the stmt |
68 | as relevant/live because it's not going to be vectorized. | |
69 | Instead mark the pattern-stmt that replaces it. */ | |
70 | ||
71 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
72 | ||
73 | if (vect_print_dump_info (REPORT_DETAILS)) | |
74 | fprintf (vect_dump, "last stmt in pattern. don't mark relevant/live."); | |
75 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
76 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
77 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
78 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
79 | stmt = pattern_stmt; | |
80 | } | |
81 | ||
82 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
83 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
84 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
85 | ||
86 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
87 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
88 | { | |
89 | if (vect_print_dump_info (REPORT_DETAILS)) | |
90 | fprintf (vect_dump, "already marked relevant/live."); | |
91 | return; | |
92 | } | |
93 | ||
94 | VEC_safe_push (gimple, heap, *worklist, stmt); | |
95 | } | |
96 | ||
97 | ||
98 | /* Function vect_stmt_relevant_p. | |
99 | ||
100 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
101 | "relevant for vectorization". | |
102 | ||
103 | A stmt is considered "relevant for vectorization" if: | |
104 | - it has uses outside the loop. | |
105 | - it has vdefs (it alters memory). | |
106 | - control stmts in the loop (except for the exit condition). | |
107 | ||
108 | CHECKME: what other side effects would the vectorizer allow? */ | |
109 | ||
110 | static bool | |
111 | vect_stmt_relevant_p (gimple stmt, loop_vec_info loop_vinfo, | |
112 | enum vect_relevant *relevant, bool *live_p) | |
113 | { | |
114 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
115 | ssa_op_iter op_iter; | |
116 | imm_use_iterator imm_iter; | |
117 | use_operand_p use_p; | |
118 | def_operand_p def_p; | |
119 | ||
8644a673 | 120 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
121 | *live_p = false; |
122 | ||
123 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
124 | if (is_ctrl_stmt (stmt) |
125 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
126 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 127 | *relevant = vect_used_in_scope; |
ebfd146a IR |
128 | |
129 | /* changing memory. */ | |
130 | if (gimple_code (stmt) != GIMPLE_PHI) | |
5006671f | 131 | if (gimple_vdef (stmt)) |
ebfd146a IR |
132 | { |
133 | if (vect_print_dump_info (REPORT_DETAILS)) | |
134 | fprintf (vect_dump, "vec_stmt_relevant_p: stmt has vdefs."); | |
8644a673 | 135 | *relevant = vect_used_in_scope; |
ebfd146a IR |
136 | } |
137 | ||
138 | /* uses outside the loop. */ | |
139 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
140 | { | |
141 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
142 | { | |
143 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
144 | if (!flow_bb_inside_loop_p (loop, bb)) | |
145 | { | |
146 | if (vect_print_dump_info (REPORT_DETAILS)) | |
147 | fprintf (vect_dump, "vec_stmt_relevant_p: used out of loop."); | |
148 | ||
3157b0c2 AO |
149 | if (is_gimple_debug (USE_STMT (use_p))) |
150 | continue; | |
151 | ||
ebfd146a IR |
152 | /* We expect all such uses to be in the loop exit phis |
153 | (because of loop closed form) */ | |
154 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
155 | gcc_assert (bb == single_exit (loop)->dest); | |
156 | ||
157 | *live_p = true; | |
158 | } | |
159 | } | |
160 | } | |
161 | ||
162 | return (*live_p || *relevant); | |
163 | } | |
164 | ||
165 | ||
b8698a0f | 166 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 167 | |
b8698a0f | 168 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
169 | used in STMT for anything other than indexing an array. */ |
170 | ||
171 | static bool | |
172 | exist_non_indexing_operands_for_use_p (tree use, gimple stmt) | |
173 | { | |
174 | tree operand; | |
175 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 176 | |
ebfd146a IR |
177 | /* USE corresponds to some operand in STMT. If there is no data |
178 | reference in STMT, then any operand that corresponds to USE | |
179 | is not indexing an array. */ | |
180 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
181 | return true; | |
59a05b0c | 182 | |
ebfd146a IR |
183 | /* STMT has a data_ref. FORNOW this means that its of one of |
184 | the following forms: | |
185 | -1- ARRAY_REF = var | |
186 | -2- var = ARRAY_REF | |
187 | (This should have been verified in analyze_data_refs). | |
188 | ||
189 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 190 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
191 | for array indexing. |
192 | ||
193 | Therefore, all we need to check is if STMT falls into the | |
194 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
195 | |
196 | if (!gimple_assign_copy_p (stmt)) | |
197 | return false; | |
59a05b0c EB |
198 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
199 | return false; | |
ebfd146a | 200 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
201 | if (TREE_CODE (operand) != SSA_NAME) |
202 | return false; | |
203 | ||
204 | if (operand == use) | |
205 | return true; | |
206 | ||
207 | return false; | |
208 | } | |
209 | ||
210 | ||
b8698a0f | 211 | /* |
ebfd146a IR |
212 | Function process_use. |
213 | ||
214 | Inputs: | |
215 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 216 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ebfd146a IR |
217 | that defined USE. This is done by calling mark_relevant and passing it |
218 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). | |
219 | ||
220 | Outputs: | |
221 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
222 | relevance info of the DEF_STMT of this USE: | |
223 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
224 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
225 | Exceptions: | |
226 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 227 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 228 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
229 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
230 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
231 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
232 | be modified accordingly. | |
233 | ||
234 | Return true if everything is as expected. Return false otherwise. */ | |
235 | ||
236 | static bool | |
b8698a0f | 237 | process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
ebfd146a IR |
238 | enum vect_relevant relevant, VEC(gimple,heap) **worklist) |
239 | { | |
240 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
241 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
242 | stmt_vec_info dstmt_vinfo; | |
243 | basic_block bb, def_bb; | |
244 | tree def; | |
245 | gimple def_stmt; | |
246 | enum vect_def_type dt; | |
247 | ||
b8698a0f | 248 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a IR |
249 | that are used for address computation are not considered relevant. */ |
250 | if (!exist_non_indexing_operands_for_use_p (use, stmt)) | |
251 | return true; | |
252 | ||
a70d6342 | 253 | if (!vect_is_simple_use (use, loop_vinfo, NULL, &def_stmt, &def, &dt)) |
b8698a0f | 254 | { |
8644a673 | 255 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
ebfd146a IR |
256 | fprintf (vect_dump, "not vectorized: unsupported use in stmt."); |
257 | return false; | |
258 | } | |
259 | ||
260 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
261 | return true; | |
262 | ||
263 | def_bb = gimple_bb (def_stmt); | |
264 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
265 | { | |
266 | if (vect_print_dump_info (REPORT_DETAILS)) | |
267 | fprintf (vect_dump, "def_stmt is out of loop."); | |
268 | return true; | |
269 | } | |
270 | ||
b8698a0f L |
271 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
272 | DEF_STMT must have already been processed, because this should be the | |
273 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
274 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
275 | check that everything is as expected, and we are done. */ |
276 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
277 | bb = gimple_bb (stmt); | |
278 | if (gimple_code (stmt) == GIMPLE_PHI | |
279 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
280 | && gimple_code (def_stmt) != GIMPLE_PHI | |
281 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
282 | && bb->loop_father == def_bb->loop_father) | |
283 | { | |
284 | if (vect_print_dump_info (REPORT_DETAILS)) | |
285 | fprintf (vect_dump, "reduc-stmt defining reduc-phi in the same nest."); | |
286 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) | |
287 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
288 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 289 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 290 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
291 | return true; |
292 | } | |
293 | ||
294 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
295 | outer-loop-header-bb: | |
296 | d = def_stmt | |
297 | inner-loop: | |
298 | stmt # use (d) | |
299 | outer-loop-tail-bb: | |
300 | ... */ | |
301 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
302 | { | |
303 | if (vect_print_dump_info (REPORT_DETAILS)) | |
304 | fprintf (vect_dump, "outer-loop def-stmt defining inner-loop stmt."); | |
7c5222ff | 305 | |
ebfd146a IR |
306 | switch (relevant) |
307 | { | |
8644a673 | 308 | case vect_unused_in_scope: |
7c5222ff IR |
309 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
310 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 311 | break; |
7c5222ff | 312 | |
ebfd146a | 313 | case vect_used_in_outer_by_reduction: |
7c5222ff | 314 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
315 | relevant = vect_used_by_reduction; |
316 | break; | |
7c5222ff | 317 | |
ebfd146a | 318 | case vect_used_in_outer: |
7c5222ff | 319 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 320 | relevant = vect_used_in_scope; |
ebfd146a | 321 | break; |
7c5222ff | 322 | |
8644a673 | 323 | case vect_used_in_scope: |
ebfd146a IR |
324 | break; |
325 | ||
326 | default: | |
327 | gcc_unreachable (); | |
b8698a0f | 328 | } |
ebfd146a IR |
329 | } |
330 | ||
331 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
332 | outer-loop-header-bb: | |
333 | ... | |
334 | inner-loop: | |
335 | d = def_stmt | |
06066f92 | 336 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
337 | stmt # use (d) */ |
338 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
339 | { | |
340 | if (vect_print_dump_info (REPORT_DETAILS)) | |
341 | fprintf (vect_dump, "inner-loop def-stmt defining outer-loop stmt."); | |
7c5222ff | 342 | |
ebfd146a IR |
343 | switch (relevant) |
344 | { | |
8644a673 | 345 | case vect_unused_in_scope: |
b8698a0f | 346 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 347 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 348 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
349 | break; |
350 | ||
ebfd146a IR |
351 | case vect_used_by_reduction: |
352 | relevant = vect_used_in_outer_by_reduction; | |
353 | break; | |
354 | ||
8644a673 | 355 | case vect_used_in_scope: |
ebfd146a IR |
356 | relevant = vect_used_in_outer; |
357 | break; | |
358 | ||
359 | default: | |
360 | gcc_unreachable (); | |
361 | } | |
362 | } | |
363 | ||
364 | vect_mark_relevant (worklist, def_stmt, relevant, live_p); | |
365 | return true; | |
366 | } | |
367 | ||
368 | ||
369 | /* Function vect_mark_stmts_to_be_vectorized. | |
370 | ||
371 | Not all stmts in the loop need to be vectorized. For example: | |
372 | ||
373 | for i... | |
374 | for j... | |
375 | 1. T0 = i + j | |
376 | 2. T1 = a[T0] | |
377 | ||
378 | 3. j = j + 1 | |
379 | ||
380 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
381 | addressing of vectorized data-refs are handled differently. | |
382 | ||
383 | This pass detects such stmts. */ | |
384 | ||
385 | bool | |
386 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
387 | { | |
388 | VEC(gimple,heap) *worklist; | |
389 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
390 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
391 | unsigned int nbbs = loop->num_nodes; | |
392 | gimple_stmt_iterator si; | |
393 | gimple stmt; | |
394 | unsigned int i; | |
395 | stmt_vec_info stmt_vinfo; | |
396 | basic_block bb; | |
397 | gimple phi; | |
398 | bool live_p; | |
06066f92 IR |
399 | enum vect_relevant relevant, tmp_relevant; |
400 | enum vect_def_type def_type; | |
ebfd146a IR |
401 | |
402 | if (vect_print_dump_info (REPORT_DETAILS)) | |
403 | fprintf (vect_dump, "=== vect_mark_stmts_to_be_vectorized ==="); | |
404 | ||
405 | worklist = VEC_alloc (gimple, heap, 64); | |
406 | ||
407 | /* 1. Init worklist. */ | |
408 | for (i = 0; i < nbbs; i++) | |
409 | { | |
410 | bb = bbs[i]; | |
411 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 412 | { |
ebfd146a IR |
413 | phi = gsi_stmt (si); |
414 | if (vect_print_dump_info (REPORT_DETAILS)) | |
415 | { | |
416 | fprintf (vect_dump, "init: phi relevant? "); | |
417 | print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM); | |
418 | } | |
419 | ||
420 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
421 | vect_mark_relevant (&worklist, phi, relevant, live_p); | |
422 | } | |
423 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
424 | { | |
425 | stmt = gsi_stmt (si); | |
426 | if (vect_print_dump_info (REPORT_DETAILS)) | |
427 | { | |
428 | fprintf (vect_dump, "init: stmt relevant? "); | |
429 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
b8698a0f | 430 | } |
ebfd146a IR |
431 | |
432 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
433 | vect_mark_relevant (&worklist, stmt, relevant, live_p); | |
434 | } | |
435 | } | |
436 | ||
437 | /* 2. Process_worklist */ | |
438 | while (VEC_length (gimple, worklist) > 0) | |
439 | { | |
440 | use_operand_p use_p; | |
441 | ssa_op_iter iter; | |
442 | ||
443 | stmt = VEC_pop (gimple, worklist); | |
444 | if (vect_print_dump_info (REPORT_DETAILS)) | |
445 | { | |
446 | fprintf (vect_dump, "worklist: examine stmt: "); | |
447 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
448 | } | |
449 | ||
b8698a0f L |
450 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
451 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
452 | liveness and relevance properties of STMT. */ |
453 | stmt_vinfo = vinfo_for_stmt (stmt); | |
454 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
455 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
456 | ||
457 | /* Generally, the liveness and relevance properties of STMT are | |
458 | propagated as is to the DEF_STMTs of its USEs: | |
459 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
460 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
461 | ||
462 | One exception is when STMT has been identified as defining a reduction | |
463 | variable; in this case we set the liveness/relevance as follows: | |
464 | live_p = false | |
465 | relevant = vect_used_by_reduction | |
466 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f L |
467 | those that are used by a reduction computation, and those that are |
468 | (also) used by a regular computation. This allows us later on to | |
469 | identify stmts that are used solely by a reduction, and therefore the | |
7c5222ff | 470 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 471 | |
06066f92 IR |
472 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
473 | tmp_relevant = relevant; | |
474 | switch (def_type) | |
ebfd146a | 475 | { |
06066f92 IR |
476 | case vect_reduction_def: |
477 | switch (tmp_relevant) | |
478 | { | |
479 | case vect_unused_in_scope: | |
480 | relevant = vect_used_by_reduction; | |
481 | break; | |
482 | ||
483 | case vect_used_by_reduction: | |
484 | if (gimple_code (stmt) == GIMPLE_PHI) | |
485 | break; | |
486 | /* fall through */ | |
487 | ||
488 | default: | |
489 | if (vect_print_dump_info (REPORT_DETAILS)) | |
490 | fprintf (vect_dump, "unsupported use of reduction."); | |
491 | ||
492 | VEC_free (gimple, heap, worklist); | |
493 | return false; | |
494 | } | |
495 | ||
b8698a0f | 496 | live_p = false; |
06066f92 | 497 | break; |
b8698a0f | 498 | |
06066f92 IR |
499 | case vect_nested_cycle: |
500 | if (tmp_relevant != vect_unused_in_scope | |
501 | && tmp_relevant != vect_used_in_outer_by_reduction | |
502 | && tmp_relevant != vect_used_in_outer) | |
503 | { | |
504 | if (vect_print_dump_info (REPORT_DETAILS)) | |
505 | fprintf (vect_dump, "unsupported use of nested cycle."); | |
7c5222ff | 506 | |
06066f92 IR |
507 | VEC_free (gimple, heap, worklist); |
508 | return false; | |
509 | } | |
7c5222ff | 510 | |
b8698a0f L |
511 | live_p = false; |
512 | break; | |
513 | ||
06066f92 IR |
514 | case vect_double_reduction_def: |
515 | if (tmp_relevant != vect_unused_in_scope | |
516 | && tmp_relevant != vect_used_by_reduction) | |
517 | { | |
7c5222ff | 518 | if (vect_print_dump_info (REPORT_DETAILS)) |
06066f92 | 519 | fprintf (vect_dump, "unsupported use of double reduction."); |
7c5222ff IR |
520 | |
521 | VEC_free (gimple, heap, worklist); | |
522 | return false; | |
06066f92 IR |
523 | } |
524 | ||
525 | live_p = false; | |
b8698a0f | 526 | break; |
7c5222ff | 527 | |
06066f92 IR |
528 | default: |
529 | break; | |
7c5222ff | 530 | } |
b8698a0f | 531 | |
ebfd146a IR |
532 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) |
533 | { | |
534 | tree op = USE_FROM_PTR (use_p); | |
535 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, &worklist)) | |
536 | { | |
537 | VEC_free (gimple, heap, worklist); | |
538 | return false; | |
539 | } | |
540 | } | |
541 | } /* while worklist */ | |
542 | ||
543 | VEC_free (gimple, heap, worklist); | |
544 | return true; | |
545 | } | |
546 | ||
547 | ||
548 | int | |
549 | cost_for_stmt (gimple stmt) | |
550 | { | |
551 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
552 | ||
553 | switch (STMT_VINFO_TYPE (stmt_info)) | |
554 | { | |
555 | case load_vec_info_type: | |
35e1a5e7 | 556 | return targetm.vectorize.builtin_vectorization_cost (scalar_load); |
ebfd146a | 557 | case store_vec_info_type: |
35e1a5e7 | 558 | return targetm.vectorize.builtin_vectorization_cost (scalar_store); |
ebfd146a IR |
559 | case op_vec_info_type: |
560 | case condition_vec_info_type: | |
561 | case assignment_vec_info_type: | |
562 | case reduc_vec_info_type: | |
563 | case induc_vec_info_type: | |
564 | case type_promotion_vec_info_type: | |
565 | case type_demotion_vec_info_type: | |
566 | case type_conversion_vec_info_type: | |
567 | case call_vec_info_type: | |
35e1a5e7 | 568 | return targetm.vectorize.builtin_vectorization_cost (scalar_stmt); |
ebfd146a IR |
569 | case undef_vec_info_type: |
570 | default: | |
571 | gcc_unreachable (); | |
572 | } | |
573 | } | |
574 | ||
b8698a0f | 575 | /* Function vect_model_simple_cost. |
ebfd146a | 576 | |
b8698a0f | 577 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
578 | single op. Right now, this does not account for multiple insns that could |
579 | be generated for the single vector op. We will handle that shortly. */ | |
580 | ||
581 | void | |
b8698a0f | 582 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
ebfd146a IR |
583 | enum vect_def_type *dt, slp_tree slp_node) |
584 | { | |
585 | int i; | |
586 | int inside_cost = 0, outside_cost = 0; | |
587 | ||
588 | /* The SLP costs were already calculated during SLP tree build. */ | |
589 | if (PURE_SLP_STMT (stmt_info)) | |
590 | return; | |
591 | ||
35e1a5e7 IR |
592 | inside_cost = ncopies |
593 | * targetm.vectorize.builtin_vectorization_cost (vector_stmt); | |
ebfd146a IR |
594 | |
595 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
596 | for (i = 0; i < 2; i++) | |
597 | { | |
8644a673 | 598 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
35e1a5e7 IR |
599 | outside_cost |
600 | += targetm.vectorize.builtin_vectorization_cost (vector_stmt); | |
ebfd146a | 601 | } |
b8698a0f | 602 | |
ebfd146a IR |
603 | if (vect_print_dump_info (REPORT_COST)) |
604 | fprintf (vect_dump, "vect_model_simple_cost: inside_cost = %d, " | |
605 | "outside_cost = %d .", inside_cost, outside_cost); | |
606 | ||
607 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
608 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
609 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
610 | } | |
611 | ||
612 | ||
b8698a0f L |
613 | /* Function vect_cost_strided_group_size |
614 | ||
ebfd146a IR |
615 | For strided load or store, return the group_size only if it is the first |
616 | load or store of a group, else return 1. This ensures that group size is | |
617 | only returned once per group. */ | |
618 | ||
619 | static int | |
620 | vect_cost_strided_group_size (stmt_vec_info stmt_info) | |
621 | { | |
622 | gimple first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
623 | ||
624 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
625 | return DR_GROUP_SIZE (stmt_info); | |
626 | ||
627 | return 1; | |
628 | } | |
629 | ||
630 | ||
631 | /* Function vect_model_store_cost | |
632 | ||
633 | Models cost for stores. In the case of strided accesses, one access | |
634 | has the overhead of the strided access attributed to it. */ | |
635 | ||
636 | void | |
b8698a0f | 637 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
ebfd146a IR |
638 | enum vect_def_type dt, slp_tree slp_node) |
639 | { | |
640 | int group_size; | |
641 | int inside_cost = 0, outside_cost = 0; | |
642 | ||
643 | /* The SLP costs were already calculated during SLP tree build. */ | |
644 | if (PURE_SLP_STMT (stmt_info)) | |
645 | return; | |
646 | ||
8644a673 | 647 | if (dt == vect_constant_def || dt == vect_external_def) |
35e1a5e7 IR |
648 | outside_cost |
649 | = targetm.vectorize.builtin_vectorization_cost (scalar_to_vec); | |
ebfd146a IR |
650 | |
651 | /* Strided access? */ | |
b8698a0f | 652 | if (DR_GROUP_FIRST_DR (stmt_info) && !slp_node) |
ebfd146a IR |
653 | group_size = vect_cost_strided_group_size (stmt_info); |
654 | /* Not a strided access. */ | |
655 | else | |
656 | group_size = 1; | |
657 | ||
b8698a0f | 658 | /* Is this an access in a group of stores, which provide strided access? |
ebfd146a | 659 | If so, add in the cost of the permutes. */ |
b8698a0f | 660 | if (group_size > 1) |
ebfd146a IR |
661 | { |
662 | /* Uses a high and low interleave operation for each needed permute. */ | |
b8698a0f | 663 | inside_cost = ncopies * exact_log2(group_size) * group_size |
35e1a5e7 | 664 | * targetm.vectorize.builtin_vectorization_cost (vector_stmt); |
ebfd146a IR |
665 | |
666 | if (vect_print_dump_info (REPORT_COST)) | |
667 | fprintf (vect_dump, "vect_model_store_cost: strided group_size = %d .", | |
668 | group_size); | |
669 | ||
670 | } | |
671 | ||
672 | /* Costs of the stores. */ | |
35e1a5e7 IR |
673 | inside_cost += ncopies |
674 | * targetm.vectorize.builtin_vectorization_cost (vector_store); | |
ebfd146a IR |
675 | |
676 | if (vect_print_dump_info (REPORT_COST)) | |
677 | fprintf (vect_dump, "vect_model_store_cost: inside_cost = %d, " | |
678 | "outside_cost = %d .", inside_cost, outside_cost); | |
679 | ||
680 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
681 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
682 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
683 | } | |
684 | ||
685 | ||
686 | /* Function vect_model_load_cost | |
687 | ||
688 | Models cost for loads. In the case of strided accesses, the last access | |
689 | has the overhead of the strided access attributed to it. Since unaligned | |
b8698a0f | 690 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
691 | access scheme chosen. */ |
692 | ||
693 | void | |
694 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, slp_tree slp_node) | |
b8698a0f | 695 | |
ebfd146a IR |
696 | { |
697 | int group_size; | |
698 | int alignment_support_cheme; | |
699 | gimple first_stmt; | |
700 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; | |
701 | int inside_cost = 0, outside_cost = 0; | |
702 | ||
703 | /* The SLP costs were already calculated during SLP tree build. */ | |
704 | if (PURE_SLP_STMT (stmt_info)) | |
705 | return; | |
706 | ||
707 | /* Strided accesses? */ | |
708 | first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
709 | if (first_stmt && !slp_node) | |
710 | { | |
711 | group_size = vect_cost_strided_group_size (stmt_info); | |
712 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
713 | } | |
714 | /* Not a strided access. */ | |
715 | else | |
716 | { | |
717 | group_size = 1; | |
718 | first_dr = dr; | |
719 | } | |
720 | ||
721 | alignment_support_cheme = vect_supportable_dr_alignment (first_dr); | |
722 | ||
b8698a0f | 723 | /* Is this an access in a group of loads providing strided access? |
ebfd146a | 724 | If so, add in the cost of the permutes. */ |
b8698a0f | 725 | if (group_size > 1) |
ebfd146a IR |
726 | { |
727 | /* Uses an even and odd extract operations for each needed permute. */ | |
728 | inside_cost = ncopies * exact_log2(group_size) * group_size | |
35e1a5e7 | 729 | * targetm.vectorize.builtin_vectorization_cost (vector_stmt); |
ebfd146a IR |
730 | |
731 | if (vect_print_dump_info (REPORT_COST)) | |
732 | fprintf (vect_dump, "vect_model_load_cost: strided group_size = %d .", | |
733 | group_size); | |
734 | ||
735 | } | |
736 | ||
737 | /* The loads themselves. */ | |
738 | switch (alignment_support_cheme) | |
739 | { | |
740 | case dr_aligned: | |
741 | { | |
35e1a5e7 IR |
742 | inside_cost += ncopies |
743 | * targetm.vectorize.builtin_vectorization_cost (vector_load); | |
ebfd146a IR |
744 | |
745 | if (vect_print_dump_info (REPORT_COST)) | |
746 | fprintf (vect_dump, "vect_model_load_cost: aligned."); | |
747 | ||
748 | break; | |
749 | } | |
750 | case dr_unaligned_supported: | |
751 | { | |
752 | /* Here, we assign an additional cost for the unaligned load. */ | |
35e1a5e7 IR |
753 | inside_cost += ncopies |
754 | * targetm.vectorize.builtin_vectorization_cost (unaligned_load); | |
ebfd146a IR |
755 | |
756 | if (vect_print_dump_info (REPORT_COST)) | |
757 | fprintf (vect_dump, "vect_model_load_cost: unaligned supported by " | |
758 | "hardware."); | |
759 | ||
760 | break; | |
761 | } | |
762 | case dr_explicit_realign: | |
763 | { | |
35e1a5e7 IR |
764 | inside_cost += ncopies * (2 |
765 | * targetm.vectorize.builtin_vectorization_cost (vector_load) | |
766 | + targetm.vectorize.builtin_vectorization_cost (vector_stmt)); | |
ebfd146a IR |
767 | |
768 | /* FIXME: If the misalignment remains fixed across the iterations of | |
769 | the containing loop, the following cost should be added to the | |
770 | outside costs. */ | |
771 | if (targetm.vectorize.builtin_mask_for_load) | |
35e1a5e7 IR |
772 | inside_cost |
773 | += targetm.vectorize.builtin_vectorization_cost (vector_stmt); | |
ebfd146a IR |
774 | |
775 | break; | |
776 | } | |
777 | case dr_explicit_realign_optimized: | |
778 | { | |
779 | if (vect_print_dump_info (REPORT_COST)) | |
780 | fprintf (vect_dump, "vect_model_load_cost: unaligned software " | |
781 | "pipelined."); | |
782 | ||
783 | /* Unaligned software pipeline has a load of an address, an initial | |
784 | load, and possibly a mask operation to "prime" the loop. However, | |
785 | if this is an access in a group of loads, which provide strided | |
786 | access, then the above cost should only be considered for one | |
787 | access in the group. Inside the loop, there is a load op | |
788 | and a realignment op. */ | |
789 | ||
790 | if ((!DR_GROUP_FIRST_DR (stmt_info)) || group_size > 1 || slp_node) | |
791 | { | |
35e1a5e7 IR |
792 | outside_cost = 2 |
793 | * targetm.vectorize.builtin_vectorization_cost (vector_stmt); | |
ebfd146a | 794 | if (targetm.vectorize.builtin_mask_for_load) |
35e1a5e7 IR |
795 | outside_cost |
796 | += targetm.vectorize.builtin_vectorization_cost (vector_stmt); | |
ebfd146a IR |
797 | } |
798 | ||
35e1a5e7 IR |
799 | inside_cost += ncopies |
800 | * (targetm.vectorize.builtin_vectorization_cost (vector_load) | |
801 | + targetm.vectorize.builtin_vectorization_cost (vector_stmt)); | |
ebfd146a IR |
802 | break; |
803 | } | |
804 | ||
805 | default: | |
806 | gcc_unreachable (); | |
807 | } | |
b8698a0f | 808 | |
ebfd146a IR |
809 | if (vect_print_dump_info (REPORT_COST)) |
810 | fprintf (vect_dump, "vect_model_load_cost: inside_cost = %d, " | |
811 | "outside_cost = %d .", inside_cost, outside_cost); | |
812 | ||
813 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
814 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
815 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
816 | } | |
817 | ||
818 | ||
819 | /* Function vect_init_vector. | |
820 | ||
821 | Insert a new stmt (INIT_STMT) that initializes a new vector variable with | |
822 | the vector elements of VECTOR_VAR. Place the initialization at BSI if it | |
823 | is not NULL. Otherwise, place the initialization at the loop preheader. | |
b8698a0f | 824 | Return the DEF of INIT_STMT. |
ebfd146a IR |
825 | It will be used in the vectorization of STMT. */ |
826 | ||
827 | tree | |
828 | vect_init_vector (gimple stmt, tree vector_var, tree vector_type, | |
829 | gimple_stmt_iterator *gsi) | |
830 | { | |
831 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
832 | tree new_var; | |
833 | gimple init_stmt; | |
834 | tree vec_oprnd; | |
835 | edge pe; | |
836 | tree new_temp; | |
837 | basic_block new_bb; | |
b8698a0f | 838 | |
ebfd146a | 839 | new_var = vect_get_new_vect_var (vector_type, vect_simple_var, "cst_"); |
b8698a0f | 840 | add_referenced_var (new_var); |
ebfd146a IR |
841 | init_stmt = gimple_build_assign (new_var, vector_var); |
842 | new_temp = make_ssa_name (new_var, init_stmt); | |
843 | gimple_assign_set_lhs (init_stmt, new_temp); | |
844 | ||
845 | if (gsi) | |
846 | vect_finish_stmt_generation (stmt, init_stmt, gsi); | |
847 | else | |
848 | { | |
849 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
b8698a0f | 850 | |
a70d6342 IR |
851 | if (loop_vinfo) |
852 | { | |
853 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
854 | ||
855 | if (nested_in_vect_loop_p (loop, stmt)) | |
856 | loop = loop->inner; | |
b8698a0f | 857 | |
a70d6342 IR |
858 | pe = loop_preheader_edge (loop); |
859 | new_bb = gsi_insert_on_edge_immediate (pe, init_stmt); | |
860 | gcc_assert (!new_bb); | |
861 | } | |
862 | else | |
863 | { | |
864 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
865 | basic_block bb; | |
866 | gimple_stmt_iterator gsi_bb_start; | |
867 | ||
868 | gcc_assert (bb_vinfo); | |
869 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 870 | gsi_bb_start = gsi_after_labels (bb); |
a70d6342 IR |
871 | gsi_insert_before (&gsi_bb_start, init_stmt, GSI_SAME_STMT); |
872 | } | |
ebfd146a IR |
873 | } |
874 | ||
875 | if (vect_print_dump_info (REPORT_DETAILS)) | |
876 | { | |
877 | fprintf (vect_dump, "created new init_stmt: "); | |
878 | print_gimple_stmt (vect_dump, init_stmt, 0, TDF_SLIM); | |
879 | } | |
880 | ||
881 | vec_oprnd = gimple_assign_lhs (init_stmt); | |
882 | return vec_oprnd; | |
883 | } | |
884 | ||
a70d6342 | 885 | |
ebfd146a IR |
886 | /* Function vect_get_vec_def_for_operand. |
887 | ||
888 | OP is an operand in STMT. This function returns a (vector) def that will be | |
889 | used in the vectorized stmt for STMT. | |
890 | ||
891 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
892 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
893 | ||
894 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
895 | needs to be introduced. */ | |
896 | ||
897 | tree | |
898 | vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def) | |
899 | { | |
900 | tree vec_oprnd; | |
901 | gimple vec_stmt; | |
902 | gimple def_stmt; | |
903 | stmt_vec_info def_stmt_info = NULL; | |
904 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
905 | tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
906 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
907 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
908 | tree vec_inv; | |
909 | tree vec_cst; | |
910 | tree t = NULL_TREE; | |
911 | tree def; | |
912 | int i; | |
913 | enum vect_def_type dt; | |
914 | bool is_simple_use; | |
915 | tree vector_type; | |
916 | ||
917 | if (vect_print_dump_info (REPORT_DETAILS)) | |
918 | { | |
919 | fprintf (vect_dump, "vect_get_vec_def_for_operand: "); | |
920 | print_generic_expr (vect_dump, op, TDF_SLIM); | |
921 | } | |
922 | ||
b8698a0f | 923 | is_simple_use = vect_is_simple_use (op, loop_vinfo, NULL, &def_stmt, &def, |
a70d6342 | 924 | &dt); |
ebfd146a IR |
925 | gcc_assert (is_simple_use); |
926 | if (vect_print_dump_info (REPORT_DETAILS)) | |
927 | { | |
928 | if (def) | |
929 | { | |
930 | fprintf (vect_dump, "def = "); | |
931 | print_generic_expr (vect_dump, def, TDF_SLIM); | |
932 | } | |
933 | if (def_stmt) | |
934 | { | |
935 | fprintf (vect_dump, " def_stmt = "); | |
936 | print_gimple_stmt (vect_dump, def_stmt, 0, TDF_SLIM); | |
937 | } | |
938 | } | |
939 | ||
940 | switch (dt) | |
941 | { | |
942 | /* Case 1: operand is a constant. */ | |
943 | case vect_constant_def: | |
944 | { | |
7569a6cc RG |
945 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
946 | gcc_assert (vector_type); | |
947 | ||
b8698a0f | 948 | if (scalar_def) |
ebfd146a IR |
949 | *scalar_def = op; |
950 | ||
951 | /* Create 'vect_cst_ = {cst,cst,...,cst}' */ | |
952 | if (vect_print_dump_info (REPORT_DETAILS)) | |
953 | fprintf (vect_dump, "Create vector_cst. nunits = %d", nunits); | |
954 | ||
955 | for (i = nunits - 1; i >= 0; --i) | |
956 | { | |
957 | t = tree_cons (NULL_TREE, op, t); | |
958 | } | |
7569a6cc RG |
959 | vec_cst = build_vector (vector_type, t); |
960 | return vect_init_vector (stmt, vec_cst, vector_type, NULL); | |
ebfd146a IR |
961 | } |
962 | ||
963 | /* Case 2: operand is defined outside the loop - loop invariant. */ | |
8644a673 | 964 | case vect_external_def: |
ebfd146a IR |
965 | { |
966 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); | |
967 | gcc_assert (vector_type); | |
968 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); | |
969 | ||
b8698a0f | 970 | if (scalar_def) |
ebfd146a IR |
971 | *scalar_def = def; |
972 | ||
973 | /* Create 'vec_inv = {inv,inv,..,inv}' */ | |
974 | if (vect_print_dump_info (REPORT_DETAILS)) | |
975 | fprintf (vect_dump, "Create vector_inv."); | |
976 | ||
977 | for (i = nunits - 1; i >= 0; --i) | |
978 | { | |
979 | t = tree_cons (NULL_TREE, def, t); | |
980 | } | |
981 | ||
982 | /* FIXME: use build_constructor directly. */ | |
983 | vec_inv = build_constructor_from_list (vector_type, t); | |
984 | return vect_init_vector (stmt, vec_inv, vector_type, NULL); | |
985 | } | |
986 | ||
987 | /* Case 3: operand is defined inside the loop. */ | |
8644a673 | 988 | case vect_internal_def: |
ebfd146a | 989 | { |
b8698a0f | 990 | if (scalar_def) |
ebfd146a IR |
991 | *scalar_def = NULL/* FIXME tuples: def_stmt*/; |
992 | ||
993 | /* Get the def from the vectorized stmt. */ | |
994 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
995 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
996 | gcc_assert (vec_stmt); | |
997 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
998 | vec_oprnd = PHI_RESULT (vec_stmt); | |
999 | else if (is_gimple_call (vec_stmt)) | |
1000 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1001 | else | |
1002 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1003 | return vec_oprnd; | |
1004 | } | |
1005 | ||
1006 | /* Case 4: operand is defined by a loop header phi - reduction */ | |
1007 | case vect_reduction_def: | |
06066f92 | 1008 | case vect_double_reduction_def: |
7c5222ff | 1009 | case vect_nested_cycle: |
ebfd146a IR |
1010 | { |
1011 | struct loop *loop; | |
1012 | ||
1013 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
b8698a0f | 1014 | loop = (gimple_bb (def_stmt))->loop_father; |
ebfd146a IR |
1015 | |
1016 | /* Get the def before the loop */ | |
1017 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); | |
1018 | return get_initial_def_for_reduction (stmt, op, scalar_def); | |
1019 | } | |
1020 | ||
1021 | /* Case 5: operand is defined by loop-header phi - induction. */ | |
1022 | case vect_induction_def: | |
1023 | { | |
1024 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1025 | ||
1026 | /* Get the def from the vectorized stmt. */ | |
1027 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1028 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
1029 | gcc_assert (vec_stmt && gimple_code (vec_stmt) == GIMPLE_PHI); | |
1030 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1031 | return vec_oprnd; | |
1032 | } | |
1033 | ||
1034 | default: | |
1035 | gcc_unreachable (); | |
1036 | } | |
1037 | } | |
1038 | ||
1039 | ||
1040 | /* Function vect_get_vec_def_for_stmt_copy | |
1041 | ||
b8698a0f L |
1042 | Return a vector-def for an operand. This function is used when the |
1043 | vectorized stmt to be created (by the caller to this function) is a "copy" | |
1044 | created in case the vectorized result cannot fit in one vector, and several | |
1045 | copies of the vector-stmt are required. In this case the vector-def is | |
ebfd146a | 1046 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1047 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1048 | DT is the type of the vector def VEC_OPRND. |
1049 | ||
1050 | Context: | |
1051 | In case the vectorization factor (VF) is bigger than the number | |
1052 | of elements that can fit in a vectype (nunits), we have to generate | |
1053 | more than one vector stmt to vectorize the scalar stmt. This situation | |
b8698a0f | 1054 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1055 | smallest data-type determines the VF, and as a result, when vectorizing |
1056 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1057 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1058 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1059 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1060 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1061 | ||
1062 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1063 | |
ebfd146a IR |
1064 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1065 | VS1.1: vx.1 = memref1 VS1.2 | |
1066 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1067 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1068 | |
1069 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1070 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1071 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1072 | VSnew.3: vz3 = vx.3 + ... | |
1073 | ||
1074 | The vectorization of S1 is explained in vectorizable_load. | |
1075 | The vectorization of S2: | |
b8698a0f L |
1076 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1077 | the function 'vect_get_vec_def_for_operand' is called to | |
ebfd146a IR |
1078 | get the relevant vector-def for each operand of S2. For operand x it |
1079 | returns the vector-def 'vx.0'. | |
1080 | ||
b8698a0f L |
1081 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1082 | function is called to get the relevant vector-def for each operand. It is | |
1083 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1084 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1085 | ||
b8698a0f L |
1086 | For example, to obtain the vector-def 'vx.1' in order to create the |
1087 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1088 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1089 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1090 | and return its def ('vx.1'). | |
1091 | Overall, to create the above sequence this function will be called 3 times: | |
1092 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1093 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1094 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1095 | ||
1096 | tree | |
1097 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1098 | { | |
1099 | gimple vec_stmt_for_operand; | |
1100 | stmt_vec_info def_stmt_info; | |
1101 | ||
1102 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1103 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1104 | return vec_oprnd; |
1105 | ||
1106 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1107 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1108 | gcc_assert (def_stmt_info); | |
1109 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1110 | gcc_assert (vec_stmt_for_operand); | |
1111 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1112 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) | |
1113 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1114 | else | |
1115 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1116 | return vec_oprnd; | |
1117 | } | |
1118 | ||
1119 | ||
1120 | /* Get vectorized definitions for the operands to create a copy of an original | |
1121 | stmt. See vect_get_vec_def_for_stmt_copy() for details. */ | |
1122 | ||
1123 | static void | |
b8698a0f L |
1124 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
1125 | VEC(tree,heap) **vec_oprnds0, | |
ebfd146a IR |
1126 | VEC(tree,heap) **vec_oprnds1) |
1127 | { | |
1128 | tree vec_oprnd = VEC_pop (tree, *vec_oprnds0); | |
1129 | ||
1130 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
1131 | VEC_quick_push (tree, *vec_oprnds0, vec_oprnd); | |
1132 | ||
1133 | if (vec_oprnds1 && *vec_oprnds1) | |
1134 | { | |
1135 | vec_oprnd = VEC_pop (tree, *vec_oprnds1); | |
1136 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); | |
1137 | VEC_quick_push (tree, *vec_oprnds1, vec_oprnd); | |
1138 | } | |
1139 | } | |
1140 | ||
1141 | ||
1142 | /* Get vectorized definitions for OP0 and OP1, or SLP_NODE if it is not NULL. */ | |
1143 | ||
1144 | static void | |
1145 | vect_get_vec_defs (tree op0, tree op1, gimple stmt, | |
1146 | VEC(tree,heap) **vec_oprnds0, VEC(tree,heap) **vec_oprnds1, | |
1147 | slp_tree slp_node) | |
1148 | { | |
1149 | if (slp_node) | |
b5aeb3bb | 1150 | vect_get_slp_defs (slp_node, vec_oprnds0, vec_oprnds1, -1); |
ebfd146a IR |
1151 | else |
1152 | { | |
1153 | tree vec_oprnd; | |
1154 | ||
b8698a0f L |
1155 | *vec_oprnds0 = VEC_alloc (tree, heap, 1); |
1156 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
ebfd146a IR |
1157 | VEC_quick_push (tree, *vec_oprnds0, vec_oprnd); |
1158 | ||
1159 | if (op1) | |
1160 | { | |
b8698a0f L |
1161 | *vec_oprnds1 = VEC_alloc (tree, heap, 1); |
1162 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL); | |
ebfd146a IR |
1163 | VEC_quick_push (tree, *vec_oprnds1, vec_oprnd); |
1164 | } | |
1165 | } | |
1166 | } | |
1167 | ||
1168 | ||
1169 | /* Function vect_finish_stmt_generation. | |
1170 | ||
1171 | Insert a new stmt. */ | |
1172 | ||
1173 | void | |
1174 | vect_finish_stmt_generation (gimple stmt, gimple vec_stmt, | |
1175 | gimple_stmt_iterator *gsi) | |
1176 | { | |
1177 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1178 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1179 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1180 | |
1181 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1182 | ||
1183 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); | |
1184 | ||
b8698a0f | 1185 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo, |
a70d6342 | 1186 | bb_vinfo)); |
ebfd146a IR |
1187 | |
1188 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1189 | { | |
1190 | fprintf (vect_dump, "add new stmt: "); | |
1191 | print_gimple_stmt (vect_dump, vec_stmt, 0, TDF_SLIM); | |
1192 | } | |
1193 | ||
1194 | gimple_set_location (vec_stmt, gimple_location (gsi_stmt (*gsi))); | |
1195 | } | |
1196 | ||
1197 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1198 | a function declaration if the target has a vectorized version | |
1199 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1200 | ||
1201 | tree | |
1202 | vectorizable_function (gimple call, tree vectype_out, tree vectype_in) | |
1203 | { | |
1204 | tree fndecl = gimple_call_fndecl (call); | |
ebfd146a IR |
1205 | |
1206 | /* We only handle functions that do not read or clobber memory -- i.e. | |
1207 | const or novops ones. */ | |
1208 | if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS))) | |
1209 | return NULL_TREE; | |
1210 | ||
1211 | if (!fndecl | |
1212 | || TREE_CODE (fndecl) != FUNCTION_DECL | |
1213 | || !DECL_BUILT_IN (fndecl)) | |
1214 | return NULL_TREE; | |
1215 | ||
62f7fd21 | 1216 | return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out, |
ebfd146a IR |
1217 | vectype_in); |
1218 | } | |
1219 | ||
1220 | /* Function vectorizable_call. | |
1221 | ||
b8698a0f L |
1222 | Check if STMT performs a function call that can be vectorized. |
1223 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1224 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1225 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1226 | ||
1227 | static bool | |
1228 | vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt) | |
1229 | { | |
1230 | tree vec_dest; | |
1231 | tree scalar_dest; | |
1232 | tree op, type; | |
1233 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1234 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
1235 | tree vectype_out, vectype_in; | |
1236 | int nunits_in; | |
1237 | int nunits_out; | |
1238 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
b690cc0f | 1239 | tree fndecl, new_temp, def, rhs_type; |
ebfd146a IR |
1240 | gimple def_stmt; |
1241 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
63827fb8 | 1242 | gimple new_stmt = NULL; |
ebfd146a IR |
1243 | int ncopies, j; |
1244 | VEC(tree, heap) *vargs = NULL; | |
1245 | enum { NARROW, NONE, WIDEN } modifier; | |
1246 | size_t i, nargs; | |
1247 | ||
a70d6342 IR |
1248 | /* FORNOW: unsupported in basic block SLP. */ |
1249 | gcc_assert (loop_vinfo); | |
b8698a0f | 1250 | |
ebfd146a IR |
1251 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
1252 | return false; | |
1253 | ||
8644a673 | 1254 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1255 | return false; |
1256 | ||
1257 | /* FORNOW: SLP not supported. */ | |
1258 | if (STMT_SLP_TYPE (stmt_info)) | |
1259 | return false; | |
1260 | ||
1261 | /* Is STMT a vectorizable call? */ | |
1262 | if (!is_gimple_call (stmt)) | |
1263 | return false; | |
1264 | ||
1265 | if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
1266 | return false; | |
1267 | ||
b690cc0f RG |
1268 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
1269 | ||
ebfd146a IR |
1270 | /* Process function arguments. */ |
1271 | rhs_type = NULL_TREE; | |
b690cc0f | 1272 | vectype_in = NULL_TREE; |
ebfd146a IR |
1273 | nargs = gimple_call_num_args (stmt); |
1274 | ||
1275 | /* Bail out if the function has more than two arguments, we | |
1276 | do not have interesting builtin functions to vectorize with | |
1277 | more than two arguments. No arguments is also not good. */ | |
1278 | if (nargs == 0 || nargs > 2) | |
1279 | return false; | |
1280 | ||
1281 | for (i = 0; i < nargs; i++) | |
1282 | { | |
b690cc0f RG |
1283 | tree opvectype; |
1284 | ||
ebfd146a IR |
1285 | op = gimple_call_arg (stmt, i); |
1286 | ||
1287 | /* We can only handle calls with arguments of the same type. */ | |
1288 | if (rhs_type | |
8533c9d8 | 1289 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a IR |
1290 | { |
1291 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1292 | fprintf (vect_dump, "argument types differ."); | |
1293 | return false; | |
1294 | } | |
b690cc0f RG |
1295 | if (!rhs_type) |
1296 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 1297 | |
b690cc0f RG |
1298 | if (!vect_is_simple_use_1 (op, loop_vinfo, NULL, |
1299 | &def_stmt, &def, &dt[i], &opvectype)) | |
ebfd146a IR |
1300 | { |
1301 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1302 | fprintf (vect_dump, "use not simple."); | |
1303 | return false; | |
1304 | } | |
ebfd146a | 1305 | |
b690cc0f RG |
1306 | if (!vectype_in) |
1307 | vectype_in = opvectype; | |
1308 | else if (opvectype | |
1309 | && opvectype != vectype_in) | |
1310 | { | |
1311 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1312 | fprintf (vect_dump, "argument vector types differ."); | |
1313 | return false; | |
1314 | } | |
1315 | } | |
1316 | /* If all arguments are external or constant defs use a vector type with | |
1317 | the same size as the output vector type. */ | |
ebfd146a | 1318 | if (!vectype_in) |
b690cc0f | 1319 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
1320 | if (vec_stmt) |
1321 | gcc_assert (vectype_in); | |
1322 | if (!vectype_in) | |
1323 | { | |
1324 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1325 | { | |
1326 | fprintf (vect_dump, "no vectype for scalar type "); | |
1327 | print_generic_expr (vect_dump, rhs_type, TDF_SLIM); | |
1328 | } | |
1329 | ||
1330 | return false; | |
1331 | } | |
ebfd146a IR |
1332 | |
1333 | /* FORNOW */ | |
b690cc0f RG |
1334 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
1335 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
1336 | if (nunits_in == nunits_out / 2) |
1337 | modifier = NARROW; | |
1338 | else if (nunits_out == nunits_in) | |
1339 | modifier = NONE; | |
1340 | else if (nunits_out == nunits_in / 2) | |
1341 | modifier = WIDEN; | |
1342 | else | |
1343 | return false; | |
1344 | ||
1345 | /* For now, we only vectorize functions if a target specific builtin | |
1346 | is available. TODO -- in some cases, it might be profitable to | |
1347 | insert the calls for pieces of the vector, in order to be able | |
1348 | to vectorize other operations in the loop. */ | |
1349 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
1350 | if (fndecl == NULL_TREE) | |
1351 | { | |
1352 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1353 | fprintf (vect_dump, "function is not vectorizable."); | |
1354 | ||
1355 | return false; | |
1356 | } | |
1357 | ||
5006671f | 1358 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a IR |
1359 | |
1360 | if (modifier == NARROW) | |
1361 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
1362 | else | |
1363 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1364 | ||
1365 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
1366 | needs to be generated. */ | |
1367 | gcc_assert (ncopies >= 1); | |
1368 | ||
1369 | if (!vec_stmt) /* transformation not required. */ | |
1370 | { | |
1371 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1372 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1373 | fprintf (vect_dump, "=== vectorizable_call ==="); | |
1374 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
1375 | return true; | |
1376 | } | |
1377 | ||
1378 | /** Transform. **/ | |
1379 | ||
1380 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1381 | fprintf (vect_dump, "transform operation."); | |
1382 | ||
1383 | /* Handle def. */ | |
1384 | scalar_dest = gimple_call_lhs (stmt); | |
1385 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1386 | ||
1387 | prev_stmt_info = NULL; | |
1388 | switch (modifier) | |
1389 | { | |
1390 | case NONE: | |
1391 | for (j = 0; j < ncopies; ++j) | |
1392 | { | |
1393 | /* Build argument list for the vectorized call. */ | |
1394 | if (j == 0) | |
1395 | vargs = VEC_alloc (tree, heap, nargs); | |
1396 | else | |
1397 | VEC_truncate (tree, vargs, 0); | |
1398 | ||
1399 | for (i = 0; i < nargs; i++) | |
1400 | { | |
1401 | op = gimple_call_arg (stmt, i); | |
1402 | if (j == 0) | |
1403 | vec_oprnd0 | |
1404 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1405 | else | |
63827fb8 IR |
1406 | { |
1407 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
1408 | vec_oprnd0 | |
1409 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
1410 | } | |
ebfd146a IR |
1411 | |
1412 | VEC_quick_push (tree, vargs, vec_oprnd0); | |
1413 | } | |
1414 | ||
1415 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1416 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1417 | gimple_call_set_lhs (new_stmt, new_temp); | |
1418 | ||
1419 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7411b8f0 | 1420 | mark_symbols_for_renaming (new_stmt); |
ebfd146a IR |
1421 | |
1422 | if (j == 0) | |
1423 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1424 | else | |
1425 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1426 | ||
1427 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1428 | } | |
1429 | ||
1430 | break; | |
1431 | ||
1432 | case NARROW: | |
1433 | for (j = 0; j < ncopies; ++j) | |
1434 | { | |
1435 | /* Build argument list for the vectorized call. */ | |
1436 | if (j == 0) | |
1437 | vargs = VEC_alloc (tree, heap, nargs * 2); | |
1438 | else | |
1439 | VEC_truncate (tree, vargs, 0); | |
1440 | ||
1441 | for (i = 0; i < nargs; i++) | |
1442 | { | |
1443 | op = gimple_call_arg (stmt, i); | |
1444 | if (j == 0) | |
1445 | { | |
1446 | vec_oprnd0 | |
1447 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1448 | vec_oprnd1 | |
63827fb8 | 1449 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
1450 | } |
1451 | else | |
1452 | { | |
63827fb8 | 1453 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i); |
ebfd146a | 1454 | vec_oprnd0 |
63827fb8 | 1455 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 1456 | vec_oprnd1 |
63827fb8 | 1457 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
1458 | } |
1459 | ||
1460 | VEC_quick_push (tree, vargs, vec_oprnd0); | |
1461 | VEC_quick_push (tree, vargs, vec_oprnd1); | |
1462 | } | |
1463 | ||
1464 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1465 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1466 | gimple_call_set_lhs (new_stmt, new_temp); | |
1467 | ||
1468 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7411b8f0 | 1469 | mark_symbols_for_renaming (new_stmt); |
ebfd146a IR |
1470 | |
1471 | if (j == 0) | |
1472 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1473 | else | |
1474 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1475 | ||
1476 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1477 | } | |
1478 | ||
1479 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
1480 | ||
1481 | break; | |
1482 | ||
1483 | case WIDEN: | |
1484 | /* No current target implements this case. */ | |
1485 | return false; | |
1486 | } | |
1487 | ||
1488 | VEC_free (tree, heap, vargs); | |
1489 | ||
1490 | /* Update the exception handling table with the vector stmt if necessary. */ | |
1491 | if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt)) | |
1492 | gimple_purge_dead_eh_edges (gimple_bb (stmt)); | |
1493 | ||
1494 | /* The call in STMT might prevent it from being removed in dce. | |
1495 | We however cannot remove it here, due to the way the ssa name | |
1496 | it defines is mapped to the new definition. So just replace | |
1497 | rhs of the statement with something harmless. */ | |
1498 | ||
1499 | type = TREE_TYPE (scalar_dest); | |
1500 | new_stmt = gimple_build_assign (gimple_call_lhs (stmt), | |
1501 | fold_convert (type, integer_zero_node)); | |
1502 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
1503 | set_vinfo_for_stmt (stmt, NULL); | |
1504 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
1505 | gsi_replace (gsi, new_stmt, false); | |
1506 | SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt; | |
1507 | ||
1508 | return true; | |
1509 | } | |
1510 | ||
1511 | ||
1512 | /* Function vect_gen_widened_results_half | |
1513 | ||
1514 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 1515 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ebfd146a IR |
1516 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
1517 | In the case that CODE is a CALL_EXPR, this means that a call to DECL | |
1518 | needs to be created (DECL is a function-decl of a target-builtin). | |
1519 | STMT is the original scalar stmt that we are vectorizing. */ | |
1520 | ||
1521 | static gimple | |
1522 | vect_gen_widened_results_half (enum tree_code code, | |
1523 | tree decl, | |
1524 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
1525 | tree vec_dest, gimple_stmt_iterator *gsi, | |
1526 | gimple stmt) | |
b8698a0f | 1527 | { |
ebfd146a | 1528 | gimple new_stmt; |
b8698a0f L |
1529 | tree new_temp; |
1530 | ||
1531 | /* Generate half of the widened result: */ | |
1532 | if (code == CALL_EXPR) | |
1533 | { | |
1534 | /* Target specific support */ | |
ebfd146a IR |
1535 | if (op_type == binary_op) |
1536 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
1537 | else | |
1538 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
1539 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1540 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
1541 | } |
1542 | else | |
ebfd146a | 1543 | { |
b8698a0f L |
1544 | /* Generic support */ |
1545 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
1546 | if (op_type != binary_op) |
1547 | vec_oprnd1 = NULL; | |
1548 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0, | |
1549 | vec_oprnd1); | |
1550 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1551 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 1552 | } |
ebfd146a IR |
1553 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1554 | ||
ebfd146a IR |
1555 | return new_stmt; |
1556 | } | |
1557 | ||
1558 | ||
b8698a0f L |
1559 | /* Check if STMT performs a conversion operation, that can be vectorized. |
1560 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1561 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1562 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1563 | ||
1564 | static bool | |
1565 | vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi, | |
1566 | gimple *vec_stmt, slp_tree slp_node) | |
1567 | { | |
1568 | tree vec_dest; | |
1569 | tree scalar_dest; | |
1570 | tree op0; | |
1571 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1572 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1573 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1574 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
1575 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; | |
1576 | tree new_temp; | |
1577 | tree def; | |
1578 | gimple def_stmt; | |
1579 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
1580 | gimple new_stmt = NULL; | |
1581 | stmt_vec_info prev_stmt_info; | |
1582 | int nunits_in; | |
1583 | int nunits_out; | |
1584 | tree vectype_out, vectype_in; | |
1585 | int ncopies, j; | |
b690cc0f | 1586 | tree rhs_type; |
ebfd146a IR |
1587 | tree builtin_decl; |
1588 | enum { NARROW, NONE, WIDEN } modifier; | |
1589 | int i; | |
1590 | VEC(tree,heap) *vec_oprnds0 = NULL; | |
1591 | tree vop0; | |
ebfd146a IR |
1592 | VEC(tree,heap) *dummy = NULL; |
1593 | int dummy_int; | |
1594 | ||
1595 | /* Is STMT a vectorizable conversion? */ | |
1596 | ||
a70d6342 IR |
1597 | /* FORNOW: unsupported in basic block SLP. */ |
1598 | gcc_assert (loop_vinfo); | |
b8698a0f | 1599 | |
ebfd146a IR |
1600 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
1601 | return false; | |
1602 | ||
8644a673 | 1603 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1604 | return false; |
1605 | ||
1606 | if (!is_gimple_assign (stmt)) | |
1607 | return false; | |
1608 | ||
1609 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
1610 | return false; | |
1611 | ||
1612 | code = gimple_assign_rhs_code (stmt); | |
1613 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
1614 | return false; | |
1615 | ||
1616 | /* Check types of lhs and rhs. */ | |
b690cc0f RG |
1617 | scalar_dest = gimple_assign_lhs (stmt); |
1618 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
1619 | ||
ebfd146a IR |
1620 | op0 = gimple_assign_rhs1 (stmt); |
1621 | rhs_type = TREE_TYPE (op0); | |
b690cc0f RG |
1622 | /* Check the operands of the operation. */ |
1623 | if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL, | |
1624 | &def_stmt, &def, &dt[0], &vectype_in)) | |
1625 | { | |
1626 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1627 | fprintf (vect_dump, "use not simple."); | |
1628 | return false; | |
1629 | } | |
1630 | /* If op0 is an external or constant defs use a vector type of | |
1631 | the same size as the output vector type. */ | |
ebfd146a | 1632 | if (!vectype_in) |
b690cc0f | 1633 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
1634 | if (vec_stmt) |
1635 | gcc_assert (vectype_in); | |
1636 | if (!vectype_in) | |
1637 | { | |
1638 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1639 | { | |
1640 | fprintf (vect_dump, "no vectype for scalar type "); | |
1641 | print_generic_expr (vect_dump, rhs_type, TDF_SLIM); | |
1642 | } | |
1643 | ||
1644 | return false; | |
1645 | } | |
ebfd146a IR |
1646 | |
1647 | /* FORNOW */ | |
b690cc0f RG |
1648 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
1649 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
1650 | if (nunits_in == nunits_out / 2) |
1651 | modifier = NARROW; | |
1652 | else if (nunits_out == nunits_in) | |
1653 | modifier = NONE; | |
1654 | else if (nunits_out == nunits_in / 2) | |
1655 | modifier = WIDEN; | |
1656 | else | |
1657 | return false; | |
1658 | ||
ebfd146a IR |
1659 | if (modifier == NARROW) |
1660 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
1661 | else | |
1662 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1663 | ||
1664 | /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies | |
1665 | this, so we can safely override NCOPIES with 1 here. */ | |
1666 | if (slp_node) | |
1667 | ncopies = 1; | |
b8698a0f | 1668 | |
ebfd146a IR |
1669 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
1670 | needs to be generated. */ | |
1671 | gcc_assert (ncopies >= 1); | |
1672 | ||
ebfd146a IR |
1673 | /* Supportable by target? */ |
1674 | if ((modifier == NONE | |
88dd7150 | 1675 | && !targetm.vectorize.builtin_conversion (code, vectype_out, vectype_in)) |
ebfd146a | 1676 | || (modifier == WIDEN |
b690cc0f RG |
1677 | && !supportable_widening_operation (code, stmt, |
1678 | vectype_out, vectype_in, | |
ebfd146a IR |
1679 | &decl1, &decl2, |
1680 | &code1, &code2, | |
1681 | &dummy_int, &dummy)) | |
1682 | || (modifier == NARROW | |
b690cc0f | 1683 | && !supportable_narrowing_operation (code, vectype_out, vectype_in, |
ebfd146a IR |
1684 | &code1, &dummy_int, &dummy))) |
1685 | { | |
1686 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1687 | fprintf (vect_dump, "conversion not supported by target."); | |
1688 | return false; | |
1689 | } | |
1690 | ||
1691 | if (modifier != NONE) | |
1692 | { | |
ebfd146a IR |
1693 | /* FORNOW: SLP not supported. */ |
1694 | if (STMT_SLP_TYPE (stmt_info)) | |
b8698a0f | 1695 | return false; |
ebfd146a IR |
1696 | } |
1697 | ||
1698 | if (!vec_stmt) /* transformation not required. */ | |
1699 | { | |
1700 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
1701 | return true; | |
1702 | } | |
1703 | ||
1704 | /** Transform. **/ | |
1705 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1706 | fprintf (vect_dump, "transform conversion."); | |
1707 | ||
1708 | /* Handle def. */ | |
1709 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1710 | ||
1711 | if (modifier == NONE && !slp_node) | |
1712 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
1713 | ||
1714 | prev_stmt_info = NULL; | |
1715 | switch (modifier) | |
1716 | { | |
1717 | case NONE: | |
1718 | for (j = 0; j < ncopies; j++) | |
1719 | { | |
ebfd146a | 1720 | if (j == 0) |
b8698a0f | 1721 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node); |
ebfd146a IR |
1722 | else |
1723 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
1724 | ||
1725 | builtin_decl = | |
88dd7150 RG |
1726 | targetm.vectorize.builtin_conversion (code, |
1727 | vectype_out, vectype_in); | |
ebfd146a | 1728 | for (i = 0; VEC_iterate (tree, vec_oprnds0, i, vop0); i++) |
b8698a0f | 1729 | { |
ebfd146a IR |
1730 | /* Arguments are ready. create the new vector stmt. */ |
1731 | new_stmt = gimple_build_call (builtin_decl, 1, vop0); | |
1732 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1733 | gimple_call_set_lhs (new_stmt, new_temp); | |
1734 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a IR |
1735 | if (slp_node) |
1736 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
1737 | } | |
1738 | ||
1739 | if (j == 0) | |
1740 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1741 | else | |
1742 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1743 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1744 | } | |
1745 | break; | |
1746 | ||
1747 | case WIDEN: | |
1748 | /* In case the vectorization factor (VF) is bigger than the number | |
1749 | of elements that we can fit in a vectype (nunits), we have to | |
1750 | generate more than one vector stmt - i.e - we need to "unroll" | |
1751 | the vector stmt by a factor VF/nunits. */ | |
1752 | for (j = 0; j < ncopies; j++) | |
1753 | { | |
1754 | if (j == 0) | |
1755 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
1756 | else | |
1757 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1758 | ||
ebfd146a IR |
1759 | /* Generate first half of the widened result: */ |
1760 | new_stmt | |
b8698a0f | 1761 | = vect_gen_widened_results_half (code1, decl1, |
ebfd146a IR |
1762 | vec_oprnd0, vec_oprnd1, |
1763 | unary_op, vec_dest, gsi, stmt); | |
1764 | if (j == 0) | |
1765 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1766 | else | |
1767 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1768 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1769 | ||
1770 | /* Generate second half of the widened result: */ | |
1771 | new_stmt | |
1772 | = vect_gen_widened_results_half (code2, decl2, | |
1773 | vec_oprnd0, vec_oprnd1, | |
1774 | unary_op, vec_dest, gsi, stmt); | |
1775 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1776 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1777 | } | |
1778 | break; | |
1779 | ||
1780 | case NARROW: | |
1781 | /* In case the vectorization factor (VF) is bigger than the number | |
1782 | of elements that we can fit in a vectype (nunits), we have to | |
1783 | generate more than one vector stmt - i.e - we need to "unroll" | |
1784 | the vector stmt by a factor VF/nunits. */ | |
1785 | for (j = 0; j < ncopies; j++) | |
1786 | { | |
1787 | /* Handle uses. */ | |
1788 | if (j == 0) | |
1789 | { | |
1790 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
1791 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1792 | } | |
1793 | else | |
1794 | { | |
1795 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd1); | |
1796 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1797 | } | |
1798 | ||
1799 | /* Arguments are ready. Create the new vector stmt. */ | |
ebfd146a IR |
1800 | new_stmt = gimple_build_assign_with_ops (code1, vec_dest, vec_oprnd0, |
1801 | vec_oprnd1); | |
1802 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1803 | gimple_assign_set_lhs (new_stmt, new_temp); | |
1804 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1805 | ||
1806 | if (j == 0) | |
1807 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1808 | else | |
1809 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1810 | ||
1811 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1812 | } | |
1813 | ||
1814 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
1815 | } | |
1816 | ||
1817 | if (vec_oprnds0) | |
b8698a0f | 1818 | VEC_free (tree, heap, vec_oprnds0); |
ebfd146a IR |
1819 | |
1820 | return true; | |
1821 | } | |
1822 | /* Function vectorizable_assignment. | |
1823 | ||
b8698a0f L |
1824 | Check if STMT performs an assignment (copy) that can be vectorized. |
1825 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1826 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1827 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1828 | ||
1829 | static bool | |
1830 | vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi, | |
1831 | gimple *vec_stmt, slp_tree slp_node) | |
1832 | { | |
1833 | tree vec_dest; | |
1834 | tree scalar_dest; | |
1835 | tree op; | |
1836 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1837 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
1838 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1839 | tree new_temp; | |
1840 | tree def; | |
1841 | gimple def_stmt; | |
1842 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
fde9c428 | 1843 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
ebfd146a | 1844 | int ncopies; |
f18b55bd | 1845 | int i, j; |
ebfd146a IR |
1846 | VEC(tree,heap) *vec_oprnds = NULL; |
1847 | tree vop; | |
a70d6342 | 1848 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
f18b55bd IR |
1849 | gimple new_stmt = NULL; |
1850 | stmt_vec_info prev_stmt_info = NULL; | |
fde9c428 RG |
1851 | enum tree_code code; |
1852 | tree vectype_in; | |
ebfd146a IR |
1853 | |
1854 | /* Multiple types in SLP are handled by creating the appropriate number of | |
1855 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
1856 | case of SLP. */ | |
1857 | if (slp_node) | |
1858 | ncopies = 1; | |
1859 | else | |
1860 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
1861 | ||
1862 | gcc_assert (ncopies >= 1); | |
ebfd146a | 1863 | |
a70d6342 | 1864 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
1865 | return false; |
1866 | ||
8644a673 | 1867 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1868 | return false; |
1869 | ||
1870 | /* Is vectorizable assignment? */ | |
1871 | if (!is_gimple_assign (stmt)) | |
1872 | return false; | |
1873 | ||
1874 | scalar_dest = gimple_assign_lhs (stmt); | |
1875 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
1876 | return false; | |
1877 | ||
fde9c428 | 1878 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 1879 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
1880 | || code == PAREN_EXPR |
1881 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
1882 | op = gimple_assign_rhs1 (stmt); |
1883 | else | |
1884 | return false; | |
1885 | ||
fde9c428 RG |
1886 | if (!vect_is_simple_use_1 (op, loop_vinfo, bb_vinfo, |
1887 | &def_stmt, &def, &dt[0], &vectype_in)) | |
ebfd146a IR |
1888 | { |
1889 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1890 | fprintf (vect_dump, "use not simple."); | |
1891 | return false; | |
1892 | } | |
1893 | ||
fde9c428 RG |
1894 | /* We can handle NOP_EXPR conversions that do not change the number |
1895 | of elements or the vector size. */ | |
1896 | if (CONVERT_EXPR_CODE_P (code) | |
1897 | && (!vectype_in | |
1898 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
1899 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
1900 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
1901 | return false; | |
1902 | ||
ebfd146a IR |
1903 | if (!vec_stmt) /* transformation not required. */ |
1904 | { | |
1905 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
1906 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1907 | fprintf (vect_dump, "=== vectorizable_assignment ==="); | |
1908 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
1909 | return true; | |
1910 | } | |
1911 | ||
1912 | /** Transform. **/ | |
1913 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1914 | fprintf (vect_dump, "transform assignment."); | |
1915 | ||
1916 | /* Handle def. */ | |
1917 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
1918 | ||
1919 | /* Handle use. */ | |
f18b55bd | 1920 | for (j = 0; j < ncopies; j++) |
ebfd146a | 1921 | { |
f18b55bd IR |
1922 | /* Handle uses. */ |
1923 | if (j == 0) | |
1924 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); | |
1925 | else | |
1926 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
1927 | ||
1928 | /* Arguments are ready. create the new vector stmt. */ | |
1929 | for (i = 0; VEC_iterate (tree, vec_oprnds, i, vop); i++) | |
1930 | { | |
fde9c428 | 1931 | if (CONVERT_EXPR_CODE_P (code)) |
4a73490d | 1932 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
1933 | new_stmt = gimple_build_assign (vec_dest, vop); |
1934 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1935 | gimple_assign_set_lhs (new_stmt, new_temp); | |
1936 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1937 | if (slp_node) | |
1938 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
1939 | } | |
ebfd146a IR |
1940 | |
1941 | if (slp_node) | |
f18b55bd IR |
1942 | continue; |
1943 | ||
1944 | if (j == 0) | |
1945 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1946 | else | |
1947 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1948 | ||
1949 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1950 | } | |
b8698a0f L |
1951 | |
1952 | VEC_free (tree, heap, vec_oprnds); | |
ebfd146a IR |
1953 | return true; |
1954 | } | |
1955 | ||
1956 | /* Function vectorizable_operation. | |
1957 | ||
b8698a0f L |
1958 | Check if STMT performs a binary or unary operation that can be vectorized. |
1959 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1960 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1961 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1962 | ||
1963 | static bool | |
1964 | vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi, | |
1965 | gimple *vec_stmt, slp_tree slp_node) | |
1966 | { | |
1967 | tree vec_dest; | |
1968 | tree scalar_dest; | |
1969 | tree op0, op1 = NULL; | |
1970 | tree vec_oprnd1 = NULL_TREE; | |
1971 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b690cc0f | 1972 | tree vectype; |
ebfd146a IR |
1973 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
1974 | enum tree_code code; | |
1975 | enum machine_mode vec_mode; | |
1976 | tree new_temp; | |
1977 | int op_type; | |
1978 | optab optab; | |
1979 | int icode; | |
1980 | enum machine_mode optab_op2_mode; | |
1981 | tree def; | |
1982 | gimple def_stmt; | |
1983 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
1984 | gimple new_stmt = NULL; | |
1985 | stmt_vec_info prev_stmt_info; | |
b690cc0f | 1986 | int nunits_in; |
ebfd146a IR |
1987 | int nunits_out; |
1988 | tree vectype_out; | |
1989 | int ncopies; | |
1990 | int j, i; | |
1991 | VEC(tree,heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; | |
1992 | tree vop0, vop1; | |
1993 | unsigned int k; | |
ebfd146a | 1994 | bool scalar_shift_arg = false; |
a70d6342 IR |
1995 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
1996 | int vf; | |
1997 | ||
a70d6342 | 1998 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
1999 | return false; |
2000 | ||
8644a673 | 2001 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2002 | return false; |
2003 | ||
2004 | /* Is STMT a vectorizable binary/unary operation? */ | |
2005 | if (!is_gimple_assign (stmt)) | |
2006 | return false; | |
2007 | ||
2008 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2009 | return false; | |
2010 | ||
ebfd146a IR |
2011 | code = gimple_assign_rhs_code (stmt); |
2012 | ||
2013 | /* For pointer addition, we should use the normal plus for | |
2014 | the vector addition. */ | |
2015 | if (code == POINTER_PLUS_EXPR) | |
2016 | code = PLUS_EXPR; | |
2017 | ||
2018 | /* Support only unary or binary operations. */ | |
2019 | op_type = TREE_CODE_LENGTH (code); | |
2020 | if (op_type != unary_op && op_type != binary_op) | |
2021 | { | |
2022 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2023 | fprintf (vect_dump, "num. args = %d (not unary/binary op).", op_type); | |
2024 | return false; | |
2025 | } | |
2026 | ||
b690cc0f RG |
2027 | scalar_dest = gimple_assign_lhs (stmt); |
2028 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2029 | ||
ebfd146a | 2030 | op0 = gimple_assign_rhs1 (stmt); |
b690cc0f RG |
2031 | if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo, |
2032 | &def_stmt, &def, &dt[0], &vectype)) | |
ebfd146a IR |
2033 | { |
2034 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2035 | fprintf (vect_dump, "use not simple."); | |
2036 | return false; | |
2037 | } | |
b690cc0f RG |
2038 | /* If op0 is an external or constant def use a vector type with |
2039 | the same size as the output vector type. */ | |
2040 | if (!vectype) | |
2041 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
2042 | if (vec_stmt) |
2043 | gcc_assert (vectype); | |
2044 | if (!vectype) | |
2045 | { | |
2046 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2047 | { | |
2048 | fprintf (vect_dump, "no vectype for scalar type "); | |
2049 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
2050 | } | |
2051 | ||
2052 | return false; | |
2053 | } | |
b690cc0f RG |
2054 | |
2055 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
2056 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
2057 | if (nunits_out != nunits_in) | |
2058 | return false; | |
ebfd146a IR |
2059 | |
2060 | if (op_type == binary_op) | |
2061 | { | |
2062 | op1 = gimple_assign_rhs2 (stmt); | |
b8698a0f | 2063 | if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def, |
a70d6342 | 2064 | &dt[1])) |
ebfd146a IR |
2065 | { |
2066 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2067 | fprintf (vect_dump, "use not simple."); | |
2068 | return false; | |
2069 | } | |
2070 | } | |
2071 | ||
b690cc0f RG |
2072 | if (loop_vinfo) |
2073 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
2074 | else | |
2075 | vf = 1; | |
2076 | ||
2077 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2078 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2079 | case of SLP. */ | |
2080 | if (slp_node) | |
2081 | ncopies = 1; | |
2082 | else | |
2083 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2084 | ||
2085 | gcc_assert (ncopies >= 1); | |
2086 | ||
ebfd146a IR |
2087 | /* If this is a shift/rotate, determine whether the shift amount is a vector, |
2088 | or scalar. If the shift/rotate amount is a vector, use the vector/vector | |
2089 | shift optabs. */ | |
2090 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
2091 | || code == RROTATE_EXPR) | |
2092 | { | |
ebfd146a | 2093 | /* vector shifted by vector */ |
8644a673 | 2094 | if (dt[1] == vect_internal_def) |
ebfd146a IR |
2095 | { |
2096 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
2097 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2098 | fprintf (vect_dump, "vector/vector shift/rotate found."); | |
2099 | } | |
2100 | ||
2101 | /* See if the machine has a vector shifted by scalar insn and if not | |
2102 | then see if it has a vector shifted by vector insn */ | |
8644a673 | 2103 | else if (dt[1] == vect_constant_def || dt[1] == vect_external_def) |
ebfd146a IR |
2104 | { |
2105 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
2106 | if (optab | |
2107 | && (optab_handler (optab, TYPE_MODE (vectype))->insn_code | |
2108 | != CODE_FOR_nothing)) | |
2109 | { | |
2110 | scalar_shift_arg = true; | |
2111 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2112 | fprintf (vect_dump, "vector/scalar shift/rotate found."); | |
2113 | } | |
2114 | else | |
2115 | { | |
2116 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
ad6c0864 | 2117 | if (optab |
ebfd146a IR |
2118 | && (optab_handler (optab, TYPE_MODE (vectype))->insn_code |
2119 | != CODE_FOR_nothing)) | |
ad6c0864 MM |
2120 | { |
2121 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2122 | fprintf (vect_dump, "vector/vector shift/rotate found."); | |
2123 | ||
2124 | /* Unlike the other binary operators, shifts/rotates have | |
2125 | the rhs being int, instead of the same type as the lhs, | |
2126 | so make sure the scalar is the right type if we are | |
2127 | dealing with vectors of short/char. */ | |
2128 | if (dt[1] == vect_constant_def) | |
2129 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
2130 | } | |
ebfd146a IR |
2131 | } |
2132 | } | |
2133 | ||
2134 | else | |
2135 | { | |
2136 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2137 | fprintf (vect_dump, "operand mode requires invariant argument."); | |
2138 | return false; | |
2139 | } | |
2140 | } | |
2141 | else | |
2142 | optab = optab_for_tree_code (code, vectype, optab_default); | |
2143 | ||
2144 | /* Supportable by target? */ | |
2145 | if (!optab) | |
2146 | { | |
2147 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2148 | fprintf (vect_dump, "no optab."); | |
2149 | return false; | |
2150 | } | |
2151 | vec_mode = TYPE_MODE (vectype); | |
2152 | icode = (int) optab_handler (optab, vec_mode)->insn_code; | |
2153 | if (icode == CODE_FOR_nothing) | |
2154 | { | |
2155 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2156 | fprintf (vect_dump, "op not supported by target."); | |
2157 | /* Check only during analysis. */ | |
2158 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
a70d6342 | 2159 | || (vf < vect_min_worthwhile_factor (code) |
ebfd146a IR |
2160 | && !vec_stmt)) |
2161 | return false; | |
2162 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2163 | fprintf (vect_dump, "proceeding using word mode."); | |
2164 | } | |
2165 | ||
2166 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
2167 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
a70d6342 | 2168 | && vf < vect_min_worthwhile_factor (code) |
ebfd146a IR |
2169 | && !vec_stmt) |
2170 | { | |
2171 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2172 | fprintf (vect_dump, "not worthwhile without SIMD support."); | |
2173 | return false; | |
2174 | } | |
2175 | ||
2176 | if (!vec_stmt) /* transformation not required. */ | |
2177 | { | |
2178 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; | |
2179 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2180 | fprintf (vect_dump, "=== vectorizable_operation ==="); | |
2181 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2182 | return true; | |
2183 | } | |
2184 | ||
2185 | /** Transform. **/ | |
2186 | ||
2187 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2188 | fprintf (vect_dump, "transform binary/unary operation."); | |
2189 | ||
2190 | /* Handle def. */ | |
2191 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2192 | ||
b8698a0f | 2193 | /* Allocate VECs for vector operands. In case of SLP, vector operands are |
ebfd146a IR |
2194 | created in the previous stages of the recursion, so no allocation is |
2195 | needed, except for the case of shift with scalar shift argument. In that | |
2196 | case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to | |
2197 | be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE. | |
b8698a0f L |
2198 | In case of loop-based vectorization we allocate VECs of size 1. We |
2199 | allocate VEC_OPRNDS1 only in case of binary operation. */ | |
ebfd146a IR |
2200 | if (!slp_node) |
2201 | { | |
2202 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
2203 | if (op_type == binary_op) | |
2204 | vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
2205 | } | |
2206 | else if (scalar_shift_arg) | |
b8698a0f | 2207 | vec_oprnds1 = VEC_alloc (tree, heap, slp_node->vec_stmts_size); |
ebfd146a IR |
2208 | |
2209 | /* In case the vectorization factor (VF) is bigger than the number | |
2210 | of elements that we can fit in a vectype (nunits), we have to generate | |
2211 | more than one vector stmt - i.e - we need to "unroll" the | |
2212 | vector stmt by a factor VF/nunits. In doing so, we record a pointer | |
2213 | from one copy of the vector stmt to the next, in the field | |
2214 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
2215 | stages to find the correct vector defs to be used when vectorizing | |
2216 | stmts that use the defs of the current stmt. The example below illustrates | |
2217 | the vectorization process when VF=16 and nunits=4 (i.e - we need to create | |
2218 | 4 vectorized stmts): | |
2219 | ||
2220 | before vectorization: | |
2221 | RELATED_STMT VEC_STMT | |
2222 | S1: x = memref - - | |
2223 | S2: z = x + 1 - - | |
2224 | ||
2225 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
2226 | there): | |
2227 | RELATED_STMT VEC_STMT | |
2228 | VS1_0: vx0 = memref0 VS1_1 - | |
2229 | VS1_1: vx1 = memref1 VS1_2 - | |
2230 | VS1_2: vx2 = memref2 VS1_3 - | |
2231 | VS1_3: vx3 = memref3 - - | |
2232 | S1: x = load - VS1_0 | |
2233 | S2: z = x + 1 - - | |
2234 | ||
2235 | step2: vectorize stmt S2 (done here): | |
2236 | To vectorize stmt S2 we first need to find the relevant vector | |
2237 | def for the first operand 'x'. This is, as usual, obtained from | |
2238 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
2239 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
2240 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
2241 | the vector stmt VS2_0, and as usual, record it in the | |
2242 | STMT_VINFO_VEC_STMT of stmt S2. | |
2243 | When creating the second copy (VS2_1), we obtain the relevant vector | |
2244 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
2245 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
2246 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
2247 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
2248 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
2249 | chain of stmts and pointers: | |
2250 | RELATED_STMT VEC_STMT | |
2251 | VS1_0: vx0 = memref0 VS1_1 - | |
2252 | VS1_1: vx1 = memref1 VS1_2 - | |
2253 | VS1_2: vx2 = memref2 VS1_3 - | |
2254 | VS1_3: vx3 = memref3 - - | |
2255 | S1: x = load - VS1_0 | |
2256 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
2257 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
2258 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
2259 | VS2_3: vz3 = vx3 + v1 - - | |
2260 | S2: z = x + 1 - VS2_0 */ | |
2261 | ||
2262 | prev_stmt_info = NULL; | |
2263 | for (j = 0; j < ncopies; j++) | |
2264 | { | |
2265 | /* Handle uses. */ | |
2266 | if (j == 0) | |
2267 | { | |
2268 | if (op_type == binary_op && scalar_shift_arg) | |
2269 | { | |
b8698a0f L |
2270 | /* Vector shl and shr insn patterns can be defined with scalar |
2271 | operand 2 (shift operand). In this case, use constant or loop | |
2272 | invariant op1 directly, without extending it to vector mode | |
ebfd146a IR |
2273 | first. */ |
2274 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
2275 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
2276 | { | |
2277 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2278 | fprintf (vect_dump, "operand 1 using scalar mode."); | |
2279 | vec_oprnd1 = op1; | |
2280 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2281 | if (slp_node) | |
2282 | { | |
2283 | /* Store vec_oprnd1 for every vector stmt to be created | |
2284 | for SLP_NODE. We check during the analysis that all the | |
b8698a0f L |
2285 | shift arguments are the same. |
2286 | TODO: Allow different constants for different vector | |
2287 | stmts generated for an SLP instance. */ | |
ebfd146a IR |
2288 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) |
2289 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2290 | } | |
2291 | } | |
2292 | } | |
b8698a0f L |
2293 | |
2294 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
2295 | (a special case for certain kind of vector shifts); otherwise, | |
ebfd146a IR |
2296 | operand 1 should be of a vector type (the usual case). */ |
2297 | if (op_type == binary_op && !vec_oprnd1) | |
b8698a0f | 2298 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, |
ebfd146a IR |
2299 | slp_node); |
2300 | else | |
b8698a0f | 2301 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, |
ebfd146a IR |
2302 | slp_node); |
2303 | } | |
2304 | else | |
2305 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
2306 | ||
2307 | /* Arguments are ready. Create the new vector stmt. */ | |
2308 | for (i = 0; VEC_iterate (tree, vec_oprnds0, i, vop0); i++) | |
2309 | { | |
2310 | vop1 = ((op_type == binary_op) | |
2311 | ? VEC_index (tree, vec_oprnds1, i) : NULL); | |
2312 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); | |
2313 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2314 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2315 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2316 | if (slp_node) | |
2317 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2318 | } | |
2319 | ||
2320 | if (slp_node) | |
2321 | continue; | |
2322 | ||
2323 | if (j == 0) | |
2324 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2325 | else | |
2326 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2327 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2328 | } | |
2329 | ||
2330 | VEC_free (tree, heap, vec_oprnds0); | |
2331 | if (vec_oprnds1) | |
2332 | VEC_free (tree, heap, vec_oprnds1); | |
2333 | ||
2334 | return true; | |
2335 | } | |
2336 | ||
2337 | ||
2338 | /* Get vectorized definitions for loop-based vectorization. For the first | |
b8698a0f L |
2339 | operand we call vect_get_vec_def_for_operand() (with OPRND containing |
2340 | scalar operand), and for the rest we get a copy with | |
ebfd146a IR |
2341 | vect_get_vec_def_for_stmt_copy() using the previous vector definition |
2342 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
2343 | The vectors are collected into VEC_OPRNDS. */ | |
2344 | ||
2345 | static void | |
b8698a0f | 2346 | vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt, |
ebfd146a IR |
2347 | VEC (tree, heap) **vec_oprnds, int multi_step_cvt) |
2348 | { | |
2349 | tree vec_oprnd; | |
2350 | ||
2351 | /* Get first vector operand. */ | |
2352 | /* All the vector operands except the very first one (that is scalar oprnd) | |
2353 | are stmt copies. */ | |
b8698a0f | 2354 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) |
ebfd146a IR |
2355 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL); |
2356 | else | |
2357 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
2358 | ||
2359 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
2360 | ||
2361 | /* Get second vector operand. */ | |
2362 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
2363 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
b8698a0f | 2364 | |
ebfd146a IR |
2365 | *oprnd = vec_oprnd; |
2366 | ||
b8698a0f | 2367 | /* For conversion in multiple steps, continue to get operands |
ebfd146a IR |
2368 | recursively. */ |
2369 | if (multi_step_cvt) | |
b8698a0f | 2370 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); |
ebfd146a IR |
2371 | } |
2372 | ||
2373 | ||
2374 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
b8698a0f | 2375 | For multi-step conversions store the resulting vectors and call the function |
ebfd146a IR |
2376 | recursively. */ |
2377 | ||
2378 | static void | |
2379 | vect_create_vectorized_demotion_stmts (VEC (tree, heap) **vec_oprnds, | |
2380 | int multi_step_cvt, gimple stmt, | |
2381 | VEC (tree, heap) *vec_dsts, | |
2382 | gimple_stmt_iterator *gsi, | |
2383 | slp_tree slp_node, enum tree_code code, | |
2384 | stmt_vec_info *prev_stmt_info) | |
2385 | { | |
2386 | unsigned int i; | |
2387 | tree vop0, vop1, new_tmp, vec_dest; | |
2388 | gimple new_stmt; | |
2389 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2390 | ||
b8698a0f | 2391 | vec_dest = VEC_pop (tree, vec_dsts); |
ebfd146a IR |
2392 | |
2393 | for (i = 0; i < VEC_length (tree, *vec_oprnds); i += 2) | |
2394 | { | |
2395 | /* Create demotion operation. */ | |
2396 | vop0 = VEC_index (tree, *vec_oprnds, i); | |
2397 | vop1 = VEC_index (tree, *vec_oprnds, i + 1); | |
2398 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); | |
2399 | new_tmp = make_ssa_name (vec_dest, new_stmt); | |
2400 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
2401 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2402 | ||
2403 | if (multi_step_cvt) | |
2404 | /* Store the resulting vector for next recursive call. */ | |
b8698a0f | 2405 | VEC_replace (tree, *vec_oprnds, i/2, new_tmp); |
ebfd146a IR |
2406 | else |
2407 | { | |
b8698a0f | 2408 | /* This is the last step of the conversion sequence. Store the |
ebfd146a IR |
2409 | vectors in SLP_NODE or in vector info of the scalar statement |
2410 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
2411 | if (slp_node) | |
2412 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2413 | else | |
2414 | { | |
2415 | if (!*prev_stmt_info) | |
2416 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2417 | else | |
2418 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
2419 | ||
2420 | *prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2421 | } | |
2422 | } | |
2423 | } | |
2424 | ||
2425 | /* For multi-step demotion operations we first generate demotion operations | |
b8698a0f | 2426 | from the source type to the intermediate types, and then combine the |
ebfd146a IR |
2427 | results (stored in VEC_OPRNDS) in demotion operation to the destination |
2428 | type. */ | |
2429 | if (multi_step_cvt) | |
2430 | { | |
2431 | /* At each level of recursion we have have of the operands we had at the | |
2432 | previous level. */ | |
2433 | VEC_truncate (tree, *vec_oprnds, (i+1)/2); | |
b8698a0f | 2434 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
ebfd146a IR |
2435 | stmt, vec_dsts, gsi, slp_node, |
2436 | code, prev_stmt_info); | |
2437 | } | |
2438 | } | |
2439 | ||
2440 | ||
2441 | /* Function vectorizable_type_demotion | |
2442 | ||
2443 | Check if STMT performs a binary or unary operation that involves | |
2444 | type demotion, and if it can be vectorized. | |
2445 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2446 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2447 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2448 | ||
2449 | static bool | |
2450 | vectorizable_type_demotion (gimple stmt, gimple_stmt_iterator *gsi, | |
2451 | gimple *vec_stmt, slp_tree slp_node) | |
2452 | { | |
2453 | tree vec_dest; | |
2454 | tree scalar_dest; | |
2455 | tree op0; | |
2456 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2457 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2458 | enum tree_code code, code1 = ERROR_MARK; | |
2459 | tree def; | |
2460 | gimple def_stmt; | |
2461 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2462 | stmt_vec_info prev_stmt_info; | |
2463 | int nunits_in; | |
2464 | int nunits_out; | |
2465 | tree vectype_out; | |
2466 | int ncopies; | |
2467 | int j, i; | |
2468 | tree vectype_in; | |
2469 | int multi_step_cvt = 0; | |
2470 | VEC (tree, heap) *vec_oprnds0 = NULL; | |
2471 | VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL; | |
2472 | tree last_oprnd, intermediate_type; | |
2473 | ||
a70d6342 IR |
2474 | /* FORNOW: not supported by basic block SLP vectorization. */ |
2475 | gcc_assert (loop_vinfo); | |
2476 | ||
ebfd146a IR |
2477 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
2478 | return false; | |
2479 | ||
8644a673 | 2480 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2481 | return false; |
2482 | ||
2483 | /* Is STMT a vectorizable type-demotion operation? */ | |
2484 | if (!is_gimple_assign (stmt)) | |
2485 | return false; | |
2486 | ||
2487 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2488 | return false; | |
2489 | ||
2490 | code = gimple_assign_rhs_code (stmt); | |
2491 | if (!CONVERT_EXPR_CODE_P (code)) | |
2492 | return false; | |
2493 | ||
b690cc0f RG |
2494 | scalar_dest = gimple_assign_lhs (stmt); |
2495 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2496 | ||
2497 | /* Check the operands of the operation. */ | |
ebfd146a | 2498 | op0 = gimple_assign_rhs1 (stmt); |
b690cc0f RG |
2499 | if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) |
2500 | && INTEGRAL_TYPE_P (TREE_TYPE (op0))) | |
2501 | || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest)) | |
2502 | && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0)) | |
2503 | && CONVERT_EXPR_CODE_P (code)))) | |
2504 | return false; | |
2505 | if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL, | |
2506 | &def_stmt, &def, &dt[0], &vectype_in)) | |
2507 | { | |
2508 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2509 | fprintf (vect_dump, "use not simple."); | |
2510 | return false; | |
2511 | } | |
2512 | /* If op0 is an external def use a vector type with the | |
2513 | same size as the output vector type if possible. */ | |
2514 | if (!vectype_in) | |
2515 | vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
2516 | if (vec_stmt) |
2517 | gcc_assert (vectype_in); | |
ebfd146a | 2518 | if (!vectype_in) |
7d8930a0 IR |
2519 | { |
2520 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2521 | { | |
2522 | fprintf (vect_dump, "no vectype for scalar type "); | |
2523 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
2524 | } | |
2525 | ||
2526 | return false; | |
2527 | } | |
ebfd146a | 2528 | |
b690cc0f | 2529 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
ebfd146a IR |
2530 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); |
2531 | if (nunits_in >= nunits_out) | |
2532 | return false; | |
2533 | ||
2534 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2535 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2536 | case of SLP. */ | |
2537 | if (slp_node) | |
2538 | ncopies = 1; | |
2539 | else | |
2540 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
ebfd146a IR |
2541 | gcc_assert (ncopies >= 1); |
2542 | ||
ebfd146a | 2543 | /* Supportable by target? */ |
b690cc0f RG |
2544 | if (!supportable_narrowing_operation (code, vectype_out, vectype_in, |
2545 | &code1, &multi_step_cvt, &interm_types)) | |
ebfd146a IR |
2546 | return false; |
2547 | ||
ebfd146a IR |
2548 | if (!vec_stmt) /* transformation not required. */ |
2549 | { | |
2550 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
2551 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2552 | fprintf (vect_dump, "=== vectorizable_demotion ==="); | |
2553 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2554 | return true; | |
2555 | } | |
2556 | ||
2557 | /** Transform. **/ | |
2558 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2559 | fprintf (vect_dump, "transform type demotion operation. ncopies = %d.", | |
2560 | ncopies); | |
2561 | ||
b8698a0f L |
2562 | /* In case of multi-step demotion, we first generate demotion operations to |
2563 | the intermediate types, and then from that types to the final one. | |
ebfd146a | 2564 | We create vector destinations for the intermediate type (TYPES) received |
b8698a0f | 2565 | from supportable_narrowing_operation, and store them in the correct order |
ebfd146a IR |
2566 | for future use in vect_create_vectorized_demotion_stmts(). */ |
2567 | if (multi_step_cvt) | |
2568 | vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1); | |
2569 | else | |
2570 | vec_dsts = VEC_alloc (tree, heap, 1); | |
b8698a0f | 2571 | |
ebfd146a IR |
2572 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); |
2573 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2574 | ||
2575 | if (multi_step_cvt) | |
2576 | { | |
b8698a0f | 2577 | for (i = VEC_length (tree, interm_types) - 1; |
ebfd146a IR |
2578 | VEC_iterate (tree, interm_types, i, intermediate_type); i--) |
2579 | { | |
b8698a0f | 2580 | vec_dest = vect_create_destination_var (scalar_dest, |
ebfd146a IR |
2581 | intermediate_type); |
2582 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2583 | } | |
2584 | } | |
2585 | ||
2586 | /* In case the vectorization factor (VF) is bigger than the number | |
2587 | of elements that we can fit in a vectype (nunits), we have to generate | |
2588 | more than one vector stmt - i.e - we need to "unroll" the | |
2589 | vector stmt by a factor VF/nunits. */ | |
2590 | last_oprnd = op0; | |
2591 | prev_stmt_info = NULL; | |
2592 | for (j = 0; j < ncopies; j++) | |
2593 | { | |
2594 | /* Handle uses. */ | |
2595 | if (slp_node) | |
b5aeb3bb | 2596 | vect_get_slp_defs (slp_node, &vec_oprnds0, NULL, -1); |
ebfd146a IR |
2597 | else |
2598 | { | |
2599 | VEC_free (tree, heap, vec_oprnds0); | |
2600 | vec_oprnds0 = VEC_alloc (tree, heap, | |
2601 | (multi_step_cvt ? vect_pow2 (multi_step_cvt) * 2 : 2)); | |
b8698a0f | 2602 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
ebfd146a IR |
2603 | vect_pow2 (multi_step_cvt) - 1); |
2604 | } | |
2605 | ||
2606 | /* Arguments are ready. Create the new vector stmts. */ | |
2607 | tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts); | |
b8698a0f | 2608 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, |
ebfd146a | 2609 | multi_step_cvt, stmt, tmp_vec_dsts, |
b8698a0f | 2610 | gsi, slp_node, code1, |
ebfd146a IR |
2611 | &prev_stmt_info); |
2612 | } | |
2613 | ||
2614 | VEC_free (tree, heap, vec_oprnds0); | |
2615 | VEC_free (tree, heap, vec_dsts); | |
2616 | VEC_free (tree, heap, tmp_vec_dsts); | |
2617 | VEC_free (tree, heap, interm_types); | |
2618 | ||
2619 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2620 | return true; | |
2621 | } | |
2622 | ||
2623 | ||
2624 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
b8698a0f | 2625 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store |
ebfd146a IR |
2626 | the resulting vectors and call the function recursively. */ |
2627 | ||
2628 | static void | |
2629 | vect_create_vectorized_promotion_stmts (VEC (tree, heap) **vec_oprnds0, | |
2630 | VEC (tree, heap) **vec_oprnds1, | |
2631 | int multi_step_cvt, gimple stmt, | |
2632 | VEC (tree, heap) *vec_dsts, | |
2633 | gimple_stmt_iterator *gsi, | |
2634 | slp_tree slp_node, enum tree_code code1, | |
b8698a0f | 2635 | enum tree_code code2, tree decl1, |
ebfd146a IR |
2636 | tree decl2, int op_type, |
2637 | stmt_vec_info *prev_stmt_info) | |
2638 | { | |
2639 | int i; | |
2640 | tree vop0, vop1, new_tmp1, new_tmp2, vec_dest; | |
2641 | gimple new_stmt1, new_stmt2; | |
2642 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2643 | VEC (tree, heap) *vec_tmp; | |
2644 | ||
2645 | vec_dest = VEC_pop (tree, vec_dsts); | |
2646 | vec_tmp = VEC_alloc (tree, heap, VEC_length (tree, *vec_oprnds0) * 2); | |
2647 | ||
2648 | for (i = 0; VEC_iterate (tree, *vec_oprnds0, i, vop0); i++) | |
2649 | { | |
2650 | if (op_type == binary_op) | |
2651 | vop1 = VEC_index (tree, *vec_oprnds1, i); | |
2652 | else | |
2653 | vop1 = NULL_TREE; | |
2654 | ||
2655 | /* Generate the two halves of promotion operation. */ | |
b8698a0f | 2656 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, |
ebfd146a IR |
2657 | op_type, vec_dest, gsi, stmt); |
2658 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
2659 | op_type, vec_dest, gsi, stmt); | |
2660 | if (is_gimple_call (new_stmt1)) | |
2661 | { | |
2662 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
2663 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
2664 | } | |
2665 | else | |
2666 | { | |
2667 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
2668 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
2669 | } | |
2670 | ||
2671 | if (multi_step_cvt) | |
2672 | { | |
2673 | /* Store the results for the recursive call. */ | |
2674 | VEC_quick_push (tree, vec_tmp, new_tmp1); | |
2675 | VEC_quick_push (tree, vec_tmp, new_tmp2); | |
2676 | } | |
2677 | else | |
2678 | { | |
2679 | /* Last step of promotion sequience - store the results. */ | |
2680 | if (slp_node) | |
2681 | { | |
2682 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt1); | |
2683 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt2); | |
2684 | } | |
2685 | else | |
2686 | { | |
2687 | if (!*prev_stmt_info) | |
2688 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt1; | |
2689 | else | |
2690 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt1; | |
2691 | ||
2692 | *prev_stmt_info = vinfo_for_stmt (new_stmt1); | |
2693 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt2; | |
2694 | *prev_stmt_info = vinfo_for_stmt (new_stmt2); | |
2695 | } | |
2696 | } | |
2697 | } | |
2698 | ||
2699 | if (multi_step_cvt) | |
2700 | { | |
b8698a0f | 2701 | /* For multi-step promotion operation we first generate we call the |
ebfd146a IR |
2702 | function recurcively for every stage. We start from the input type, |
2703 | create promotion operations to the intermediate types, and then | |
2704 | create promotions to the output type. */ | |
2705 | *vec_oprnds0 = VEC_copy (tree, heap, vec_tmp); | |
2706 | VEC_free (tree, heap, vec_tmp); | |
2707 | vect_create_vectorized_promotion_stmts (vec_oprnds0, vec_oprnds1, | |
2708 | multi_step_cvt - 1, stmt, | |
2709 | vec_dsts, gsi, slp_node, code1, | |
2710 | code2, decl2, decl2, op_type, | |
2711 | prev_stmt_info); | |
2712 | } | |
2713 | } | |
b8698a0f | 2714 | |
ebfd146a IR |
2715 | |
2716 | /* Function vectorizable_type_promotion | |
2717 | ||
2718 | Check if STMT performs a binary or unary operation that involves | |
2719 | type promotion, and if it can be vectorized. | |
2720 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2721 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2722 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2723 | ||
2724 | static bool | |
2725 | vectorizable_type_promotion (gimple stmt, gimple_stmt_iterator *gsi, | |
2726 | gimple *vec_stmt, slp_tree slp_node) | |
2727 | { | |
2728 | tree vec_dest; | |
2729 | tree scalar_dest; | |
2730 | tree op0, op1 = NULL; | |
2731 | tree vec_oprnd0=NULL, vec_oprnd1=NULL; | |
2732 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2733 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2734 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
2735 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; | |
b8698a0f | 2736 | int op_type; |
ebfd146a IR |
2737 | tree def; |
2738 | gimple def_stmt; | |
2739 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2740 | stmt_vec_info prev_stmt_info; | |
2741 | int nunits_in; | |
2742 | int nunits_out; | |
2743 | tree vectype_out; | |
2744 | int ncopies; | |
2745 | int j, i; | |
2746 | tree vectype_in; | |
2747 | tree intermediate_type = NULL_TREE; | |
2748 | int multi_step_cvt = 0; | |
2749 | VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; | |
2750 | VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL; | |
b8698a0f | 2751 | |
a70d6342 IR |
2752 | /* FORNOW: not supported by basic block SLP vectorization. */ |
2753 | gcc_assert (loop_vinfo); | |
b8698a0f | 2754 | |
ebfd146a IR |
2755 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
2756 | return false; | |
2757 | ||
8644a673 | 2758 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2759 | return false; |
2760 | ||
2761 | /* Is STMT a vectorizable type-promotion operation? */ | |
2762 | if (!is_gimple_assign (stmt)) | |
2763 | return false; | |
2764 | ||
2765 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2766 | return false; | |
2767 | ||
2768 | code = gimple_assign_rhs_code (stmt); | |
2769 | if (!CONVERT_EXPR_CODE_P (code) | |
2770 | && code != WIDEN_MULT_EXPR) | |
2771 | return false; | |
2772 | ||
b690cc0f RG |
2773 | scalar_dest = gimple_assign_lhs (stmt); |
2774 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2775 | ||
2776 | /* Check the operands of the operation. */ | |
ebfd146a | 2777 | op0 = gimple_assign_rhs1 (stmt); |
b690cc0f RG |
2778 | if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) |
2779 | && INTEGRAL_TYPE_P (TREE_TYPE (op0))) | |
2780 | || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest)) | |
2781 | && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0)) | |
2782 | && CONVERT_EXPR_CODE_P (code)))) | |
2783 | return false; | |
2784 | if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL, | |
2785 | &def_stmt, &def, &dt[0], &vectype_in)) | |
2786 | { | |
2787 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2788 | fprintf (vect_dump, "use not simple."); | |
2789 | return false; | |
2790 | } | |
2791 | /* If op0 is an external or constant def use a vector type with | |
2792 | the same size as the output vector type. */ | |
2793 | if (!vectype_in) | |
2794 | vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
2795 | if (vec_stmt) |
2796 | gcc_assert (vectype_in); | |
ebfd146a | 2797 | if (!vectype_in) |
7d8930a0 IR |
2798 | { |
2799 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2800 | { | |
2801 | fprintf (vect_dump, "no vectype for scalar type "); | |
2802 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
2803 | } | |
2804 | ||
2805 | return false; | |
2806 | } | |
ebfd146a | 2807 | |
b690cc0f | 2808 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
ebfd146a IR |
2809 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); |
2810 | if (nunits_in <= nunits_out) | |
2811 | return false; | |
2812 | ||
2813 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2814 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2815 | case of SLP. */ | |
2816 | if (slp_node) | |
2817 | ncopies = 1; | |
2818 | else | |
2819 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2820 | ||
2821 | gcc_assert (ncopies >= 1); | |
2822 | ||
ebfd146a IR |
2823 | op_type = TREE_CODE_LENGTH (code); |
2824 | if (op_type == binary_op) | |
2825 | { | |
2826 | op1 = gimple_assign_rhs2 (stmt); | |
a70d6342 | 2827 | if (!vect_is_simple_use (op1, loop_vinfo, NULL, &def_stmt, &def, &dt[1])) |
ebfd146a IR |
2828 | { |
2829 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2830 | fprintf (vect_dump, "use not simple."); | |
2831 | return false; | |
2832 | } | |
2833 | } | |
2834 | ||
2835 | /* Supportable by target? */ | |
b690cc0f | 2836 | if (!supportable_widening_operation (code, stmt, vectype_out, vectype_in, |
ebfd146a IR |
2837 | &decl1, &decl2, &code1, &code2, |
2838 | &multi_step_cvt, &interm_types)) | |
2839 | return false; | |
2840 | ||
2841 | /* Binary widening operation can only be supported directly by the | |
2842 | architecture. */ | |
2843 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
2844 | ||
ebfd146a IR |
2845 | if (!vec_stmt) /* transformation not required. */ |
2846 | { | |
2847 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
2848 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2849 | fprintf (vect_dump, "=== vectorizable_promotion ==="); | |
2850 | vect_model_simple_cost (stmt_info, 2*ncopies, dt, NULL); | |
2851 | return true; | |
2852 | } | |
2853 | ||
2854 | /** Transform. **/ | |
2855 | ||
2856 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2857 | fprintf (vect_dump, "transform type promotion operation. ncopies = %d.", | |
2858 | ncopies); | |
2859 | ||
2860 | /* Handle def. */ | |
b8698a0f | 2861 | /* In case of multi-step promotion, we first generate promotion operations |
ebfd146a | 2862 | to the intermediate types, and then from that types to the final one. |
b8698a0f L |
2863 | We store vector destination in VEC_DSTS in the correct order for |
2864 | recursive creation of promotion operations in | |
ebfd146a IR |
2865 | vect_create_vectorized_promotion_stmts(). Vector destinations are created |
2866 | according to TYPES recieved from supportable_widening_operation(). */ | |
2867 | if (multi_step_cvt) | |
2868 | vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1); | |
2869 | else | |
2870 | vec_dsts = VEC_alloc (tree, heap, 1); | |
2871 | ||
2872 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2873 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2874 | ||
2875 | if (multi_step_cvt) | |
2876 | { | |
2877 | for (i = VEC_length (tree, interm_types) - 1; | |
2878 | VEC_iterate (tree, interm_types, i, intermediate_type); i--) | |
2879 | { | |
2880 | vec_dest = vect_create_destination_var (scalar_dest, | |
2881 | intermediate_type); | |
2882 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2883 | } | |
2884 | } | |
b8698a0f | 2885 | |
ebfd146a IR |
2886 | if (!slp_node) |
2887 | { | |
b8698a0f | 2888 | vec_oprnds0 = VEC_alloc (tree, heap, |
ebfd146a IR |
2889 | (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); |
2890 | if (op_type == binary_op) | |
2891 | vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
2892 | } | |
2893 | ||
2894 | /* In case the vectorization factor (VF) is bigger than the number | |
2895 | of elements that we can fit in a vectype (nunits), we have to generate | |
2896 | more than one vector stmt - i.e - we need to "unroll" the | |
2897 | vector stmt by a factor VF/nunits. */ | |
2898 | ||
2899 | prev_stmt_info = NULL; | |
2900 | for (j = 0; j < ncopies; j++) | |
2901 | { | |
2902 | /* Handle uses. */ | |
2903 | if (j == 0) | |
2904 | { | |
2905 | if (slp_node) | |
b5aeb3bb | 2906 | vect_get_slp_defs (slp_node, &vec_oprnds0, &vec_oprnds1, -1); |
ebfd146a IR |
2907 | else |
2908 | { | |
2909 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
2910 | VEC_quick_push (tree, vec_oprnds0, vec_oprnd0); | |
2911 | if (op_type == binary_op) | |
2912 | { | |
2913 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, NULL); | |
2914 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2915 | } | |
2916 | } | |
2917 | } | |
2918 | else | |
2919 | { | |
2920 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
2921 | VEC_replace (tree, vec_oprnds0, 0, vec_oprnd0); | |
2922 | if (op_type == binary_op) | |
2923 | { | |
2924 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd1); | |
2925 | VEC_replace (tree, vec_oprnds1, 0, vec_oprnd1); | |
2926 | } | |
2927 | } | |
2928 | ||
2929 | /* Arguments are ready. Create the new vector stmts. */ | |
2930 | tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts); | |
2931 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, &vec_oprnds1, | |
b8698a0f | 2932 | multi_step_cvt, stmt, |
ebfd146a IR |
2933 | tmp_vec_dsts, |
2934 | gsi, slp_node, code1, code2, | |
2935 | decl1, decl2, op_type, | |
2936 | &prev_stmt_info); | |
2937 | } | |
2938 | ||
2939 | VEC_free (tree, heap, vec_dsts); | |
2940 | VEC_free (tree, heap, tmp_vec_dsts); | |
2941 | VEC_free (tree, heap, interm_types); | |
2942 | VEC_free (tree, heap, vec_oprnds0); | |
2943 | VEC_free (tree, heap, vec_oprnds1); | |
2944 | ||
2945 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2946 | return true; | |
2947 | } | |
2948 | ||
2949 | ||
2950 | /* Function vectorizable_store. | |
2951 | ||
b8698a0f L |
2952 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
2953 | can be vectorized. | |
2954 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
2955 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2956 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2957 | ||
2958 | static bool | |
2959 | vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
2960 | slp_tree slp_node) | |
2961 | { | |
2962 | tree scalar_dest; | |
2963 | tree data_ref; | |
2964 | tree op; | |
2965 | tree vec_oprnd = NULL_TREE; | |
2966 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2967 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
2968 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2969 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 2970 | struct loop *loop = NULL; |
ebfd146a IR |
2971 | enum machine_mode vec_mode; |
2972 | tree dummy; | |
2973 | enum dr_alignment_support alignment_support_scheme; | |
2974 | tree def; | |
2975 | gimple def_stmt; | |
2976 | enum vect_def_type dt; | |
2977 | stmt_vec_info prev_stmt_info = NULL; | |
2978 | tree dataref_ptr = NULL_TREE; | |
2979 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2980 | int ncopies; | |
2981 | int j; | |
2982 | gimple next_stmt, first_stmt = NULL; | |
2983 | bool strided_store = false; | |
2984 | unsigned int group_size, i; | |
2985 | VEC(tree,heap) *dr_chain = NULL, *oprnds = NULL, *result_chain = NULL; | |
2986 | bool inv_p; | |
2987 | VEC(tree,heap) *vec_oprnds = NULL; | |
2988 | bool slp = (slp_node != NULL); | |
ebfd146a | 2989 | unsigned int vec_num; |
a70d6342 IR |
2990 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2991 | ||
2992 | if (loop_vinfo) | |
2993 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
2994 | |
2995 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2996 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2997 | case of SLP. */ | |
2998 | if (slp) | |
2999 | ncopies = 1; | |
3000 | else | |
3001 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3002 | ||
3003 | gcc_assert (ncopies >= 1); | |
3004 | ||
3005 | /* FORNOW. This restriction should be relaxed. */ | |
a70d6342 | 3006 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) |
ebfd146a IR |
3007 | { |
3008 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3009 | fprintf (vect_dump, "multiple types in nested loop."); | |
3010 | return false; | |
3011 | } | |
3012 | ||
a70d6342 | 3013 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3014 | return false; |
3015 | ||
8644a673 | 3016 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3017 | return false; |
3018 | ||
3019 | /* Is vectorizable store? */ | |
3020 | ||
3021 | if (!is_gimple_assign (stmt)) | |
3022 | return false; | |
3023 | ||
3024 | scalar_dest = gimple_assign_lhs (stmt); | |
3025 | if (TREE_CODE (scalar_dest) != ARRAY_REF | |
3026 | && TREE_CODE (scalar_dest) != INDIRECT_REF | |
e9dbe7bb IR |
3027 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
3028 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
3029 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
3030 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
3031 | return false; |
3032 | ||
3033 | gcc_assert (gimple_assign_single_p (stmt)); | |
3034 | op = gimple_assign_rhs1 (stmt); | |
a70d6342 | 3035 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt)) |
ebfd146a IR |
3036 | { |
3037 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3038 | fprintf (vect_dump, "use not simple."); | |
3039 | return false; | |
3040 | } | |
3041 | ||
3042 | /* The scalar rhs type needs to be trivially convertible to the vector | |
3043 | component type. This should always be the case. */ | |
3044 | if (!useless_type_conversion_p (TREE_TYPE (vectype), TREE_TYPE (op))) | |
b8698a0f | 3045 | { |
ebfd146a IR |
3046 | if (vect_print_dump_info (REPORT_DETAILS)) |
3047 | fprintf (vect_dump, "??? operands of different types"); | |
3048 | return false; | |
3049 | } | |
3050 | ||
3051 | vec_mode = TYPE_MODE (vectype); | |
3052 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
3053 | (e.g. - array initialization with 0). */ | |
3054 | if (optab_handler (mov_optab, (int)vec_mode)->insn_code == CODE_FOR_nothing) | |
3055 | return false; | |
3056 | ||
3057 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
3058 | return false; | |
3059 | ||
3060 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) | |
3061 | { | |
3062 | strided_store = true; | |
3063 | first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
3064 | if (!vect_strided_store_supported (vectype) | |
3065 | && !PURE_SLP_STMT (stmt_info) && !slp) | |
3066 | return false; | |
b8698a0f | 3067 | |
ebfd146a IR |
3068 | if (first_stmt == stmt) |
3069 | { | |
3070 | /* STMT is the leader of the group. Check the operands of all the | |
3071 | stmts of the group. */ | |
3072 | next_stmt = DR_GROUP_NEXT_DR (stmt_info); | |
3073 | while (next_stmt) | |
3074 | { | |
3075 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
3076 | op = gimple_assign_rhs1 (next_stmt); | |
b8698a0f | 3077 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, |
a70d6342 | 3078 | &def, &dt)) |
ebfd146a IR |
3079 | { |
3080 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3081 | fprintf (vect_dump, "use not simple."); | |
3082 | return false; | |
3083 | } | |
3084 | next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); | |
3085 | } | |
3086 | } | |
3087 | } | |
3088 | ||
3089 | if (!vec_stmt) /* transformation not required. */ | |
3090 | { | |
3091 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
3092 | vect_model_store_cost (stmt_info, ncopies, dt, NULL); | |
3093 | return true; | |
3094 | } | |
3095 | ||
3096 | /** Transform. **/ | |
3097 | ||
3098 | if (strided_store) | |
3099 | { | |
3100 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
3101 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
3102 | ||
3103 | DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; | |
3104 | ||
3105 | /* FORNOW */ | |
a70d6342 | 3106 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
3107 | |
3108 | /* We vectorize all the stmts of the interleaving group when we | |
3109 | reach the last stmt in the group. */ | |
b8698a0f | 3110 | if (DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
ebfd146a IR |
3111 | < DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)) |
3112 | && !slp) | |
3113 | { | |
3114 | *vec_stmt = NULL; | |
3115 | return true; | |
3116 | } | |
3117 | ||
3118 | if (slp) | |
4b5caab7 IR |
3119 | { |
3120 | strided_store = false; | |
3121 | /* VEC_NUM is the number of vect stmts to be created for this | |
3122 | group. */ | |
3123 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
3124 | first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0); | |
3125 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
3126 | } | |
ebfd146a | 3127 | else |
4b5caab7 IR |
3128 | /* VEC_NUM is the number of vect stmts to be created for this |
3129 | group. */ | |
ebfd146a IR |
3130 | vec_num = group_size; |
3131 | } | |
b8698a0f | 3132 | else |
ebfd146a IR |
3133 | { |
3134 | first_stmt = stmt; | |
3135 | first_dr = dr; | |
3136 | group_size = vec_num = 1; | |
ebfd146a | 3137 | } |
b8698a0f | 3138 | |
ebfd146a IR |
3139 | if (vect_print_dump_info (REPORT_DETAILS)) |
3140 | fprintf (vect_dump, "transform store. ncopies = %d",ncopies); | |
3141 | ||
3142 | dr_chain = VEC_alloc (tree, heap, group_size); | |
3143 | oprnds = VEC_alloc (tree, heap, group_size); | |
3144 | ||
3145 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr); | |
3146 | gcc_assert (alignment_support_scheme); | |
ebfd146a IR |
3147 | |
3148 | /* In case the vectorization factor (VF) is bigger than the number | |
3149 | of elements that we can fit in a vectype (nunits), we have to generate | |
3150 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 3151 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
3152 | vect_get_vec_def_for_copy_stmt. */ |
3153 | ||
3154 | /* In case of interleaving (non-unit strided access): | |
3155 | ||
3156 | S1: &base + 2 = x2 | |
3157 | S2: &base = x0 | |
3158 | S3: &base + 1 = x1 | |
3159 | S4: &base + 3 = x3 | |
3160 | ||
3161 | We create vectorized stores starting from base address (the access of the | |
3162 | first stmt in the chain (S2 in the above example), when the last store stmt | |
3163 | of the chain (S4) is reached: | |
3164 | ||
3165 | VS1: &base = vx2 | |
3166 | VS2: &base + vec_size*1 = vx0 | |
3167 | VS3: &base + vec_size*2 = vx1 | |
3168 | VS4: &base + vec_size*3 = vx3 | |
3169 | ||
3170 | Then permutation statements are generated: | |
3171 | ||
3172 | VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 > | |
3173 | VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 > | |
3174 | ... | |
b8698a0f | 3175 | |
ebfd146a IR |
3176 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
3177 | (the order of the data-refs in the output of vect_permute_store_chain | |
3178 | corresponds to the order of scalar stmts in the interleaving chain - see | |
3179 | the documentation of vect_permute_store_chain()). | |
3180 | ||
3181 | In case of both multiple types and interleaving, above vector stores and | |
3182 | permutation stmts are created for every copy. The result vector stmts are | |
3183 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding | |
b8698a0f | 3184 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
3185 | */ |
3186 | ||
3187 | prev_stmt_info = NULL; | |
3188 | for (j = 0; j < ncopies; j++) | |
3189 | { | |
3190 | gimple new_stmt; | |
3191 | gimple ptr_incr; | |
3192 | ||
3193 | if (j == 0) | |
3194 | { | |
3195 | if (slp) | |
3196 | { | |
3197 | /* Get vectorized arguments for SLP_NODE. */ | |
b5aeb3bb | 3198 | vect_get_slp_defs (slp_node, &vec_oprnds, NULL, -1); |
ebfd146a IR |
3199 | |
3200 | vec_oprnd = VEC_index (tree, vec_oprnds, 0); | |
3201 | } | |
3202 | else | |
3203 | { | |
b8698a0f L |
3204 | /* For interleaved stores we collect vectorized defs for all the |
3205 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
3206 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
3207 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
3208 | ||
3209 | If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and | |
3210 | OPRNDS are of size 1. */ | |
b8698a0f | 3211 | next_stmt = first_stmt; |
ebfd146a IR |
3212 | for (i = 0; i < group_size; i++) |
3213 | { | |
b8698a0f L |
3214 | /* Since gaps are not supported for interleaved stores, |
3215 | GROUP_SIZE is the exact number of stmts in the chain. | |
3216 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
3217 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
3218 | iteration of the loop will be executed. */ |
3219 | gcc_assert (next_stmt | |
3220 | && gimple_assign_single_p (next_stmt)); | |
3221 | op = gimple_assign_rhs1 (next_stmt); | |
3222 | ||
b8698a0f | 3223 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt, |
ebfd146a | 3224 | NULL); |
b8698a0f L |
3225 | VEC_quick_push(tree, dr_chain, vec_oprnd); |
3226 | VEC_quick_push(tree, oprnds, vec_oprnd); | |
ebfd146a IR |
3227 | next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); |
3228 | } | |
3229 | } | |
3230 | ||
3231 | /* We should have catched mismatched types earlier. */ | |
3232 | gcc_assert (useless_type_conversion_p (vectype, | |
3233 | TREE_TYPE (vec_oprnd))); | |
b8698a0f L |
3234 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, NULL, NULL_TREE, |
3235 | &dummy, &ptr_incr, false, | |
5006671f | 3236 | &inv_p); |
a70d6342 | 3237 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 3238 | } |
b8698a0f | 3239 | else |
ebfd146a | 3240 | { |
b8698a0f L |
3241 | /* For interleaved stores we created vectorized defs for all the |
3242 | defs stored in OPRNDS in the previous iteration (previous copy). | |
3243 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
3244 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
3245 | next copy. | |
3246 | If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and | |
3247 | OPRNDS are of size 1. */ | |
3248 | for (i = 0; i < group_size; i++) | |
3249 | { | |
3250 | op = VEC_index (tree, oprnds, i); | |
b8698a0f | 3251 | vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, |
a70d6342 | 3252 | &dt); |
b8698a0f | 3253 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
ebfd146a IR |
3254 | VEC_replace(tree, dr_chain, i, vec_oprnd); |
3255 | VEC_replace(tree, oprnds, i, vec_oprnd); | |
3256 | } | |
b8698a0f | 3257 | dataref_ptr = |
ebfd146a IR |
3258 | bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE); |
3259 | } | |
3260 | ||
3261 | if (strided_store) | |
3262 | { | |
b8698a0f | 3263 | result_chain = VEC_alloc (tree, heap, group_size); |
ebfd146a IR |
3264 | /* Permute. */ |
3265 | if (!vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
3266 | &result_chain)) | |
3267 | return false; | |
3268 | } | |
3269 | ||
3270 | next_stmt = first_stmt; | |
3271 | for (i = 0; i < vec_num; i++) | |
3272 | { | |
3273 | if (i > 0) | |
3274 | /* Bump the vector pointer. */ | |
3275 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
3276 | NULL_TREE); | |
3277 | ||
3278 | if (slp) | |
3279 | vec_oprnd = VEC_index (tree, vec_oprnds, i); | |
3280 | else if (strided_store) | |
b8698a0f | 3281 | /* For strided stores vectorized defs are interleaved in |
ebfd146a IR |
3282 | vect_permute_store_chain(). */ |
3283 | vec_oprnd = VEC_index (tree, result_chain, i); | |
3284 | ||
8f439681 | 3285 | if (aligned_access_p (first_dr)) |
70f34814 | 3286 | data_ref = build_simple_mem_ref (dataref_ptr); |
8f439681 RE |
3287 | else |
3288 | { | |
3289 | int mis = DR_MISALIGNMENT (first_dr); | |
3290 | tree tmis = (mis == -1 ? size_zero_node : size_int (mis)); | |
3291 | tmis = size_binop (MULT_EXPR, tmis, size_int (BITS_PER_UNIT)); | |
3292 | data_ref = build2 (MISALIGNED_INDIRECT_REF, vectype, dataref_ptr, tmis); | |
3293 | } | |
3294 | ||
5006671f RG |
3295 | /* If accesses through a pointer to vectype do not alias the original |
3296 | memory reference we have a problem. This should never happen. */ | |
3297 | gcc_assert (alias_sets_conflict_p (get_alias_set (data_ref), | |
3298 | get_alias_set (gimple_assign_lhs (stmt)))); | |
ebfd146a IR |
3299 | |
3300 | /* Arguments are ready. Create the new vector stmt. */ | |
3301 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
3302 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3303 | mark_symbols_for_renaming (new_stmt); | |
3304 | ||
3305 | if (slp) | |
3306 | continue; | |
b8698a0f | 3307 | |
ebfd146a IR |
3308 | if (j == 0) |
3309 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3310 | else | |
3311 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3312 | ||
3313 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3314 | next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); | |
3315 | if (!next_stmt) | |
3316 | break; | |
3317 | } | |
3318 | } | |
3319 | ||
b8698a0f L |
3320 | VEC_free (tree, heap, dr_chain); |
3321 | VEC_free (tree, heap, oprnds); | |
ebfd146a | 3322 | if (result_chain) |
b8698a0f | 3323 | VEC_free (tree, heap, result_chain); |
ebfd146a IR |
3324 | |
3325 | return true; | |
3326 | } | |
3327 | ||
3328 | /* vectorizable_load. | |
3329 | ||
b8698a0f L |
3330 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
3331 | can be vectorized. | |
3332 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
3333 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3334 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3335 | ||
3336 | static bool | |
3337 | vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
3338 | slp_tree slp_node, slp_instance slp_node_instance) | |
3339 | { | |
3340 | tree scalar_dest; | |
3341 | tree vec_dest = NULL; | |
3342 | tree data_ref = NULL; | |
3343 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 3344 | stmt_vec_info prev_stmt_info; |
ebfd146a | 3345 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 3346 | struct loop *loop = NULL; |
ebfd146a | 3347 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 3348 | bool nested_in_vect_loop = false; |
ebfd146a IR |
3349 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; |
3350 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3351 | tree new_temp; | |
3352 | int mode; | |
3353 | gimple new_stmt = NULL; | |
3354 | tree dummy; | |
3355 | enum dr_alignment_support alignment_support_scheme; | |
3356 | tree dataref_ptr = NULL_TREE; | |
3357 | gimple ptr_incr; | |
3358 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
3359 | int ncopies; | |
3360 | int i, j, group_size; | |
3361 | tree msq = NULL_TREE, lsq; | |
3362 | tree offset = NULL_TREE; | |
3363 | tree realignment_token = NULL_TREE; | |
3364 | gimple phi = NULL; | |
3365 | VEC(tree,heap) *dr_chain = NULL; | |
3366 | bool strided_load = false; | |
3367 | gimple first_stmt; | |
3368 | tree scalar_type; | |
3369 | bool inv_p; | |
3370 | bool compute_in_loop = false; | |
3371 | struct loop *at_loop; | |
3372 | int vec_num; | |
3373 | bool slp = (slp_node != NULL); | |
3374 | bool slp_perm = false; | |
3375 | enum tree_code code; | |
a70d6342 IR |
3376 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3377 | int vf; | |
3378 | ||
3379 | if (loop_vinfo) | |
3380 | { | |
3381 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
3382 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
3383 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3384 | } | |
3385 | else | |
3533e503 | 3386 | vf = 1; |
ebfd146a IR |
3387 | |
3388 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3389 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3390 | case of SLP. */ | |
3391 | if (slp) | |
3392 | ncopies = 1; | |
3393 | else | |
3394 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3395 | ||
3396 | gcc_assert (ncopies >= 1); | |
3397 | ||
3398 | /* FORNOW. This restriction should be relaxed. */ | |
3399 | if (nested_in_vect_loop && ncopies > 1) | |
3400 | { | |
3401 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3402 | fprintf (vect_dump, "multiple types in nested loop."); | |
3403 | return false; | |
3404 | } | |
3405 | ||
a70d6342 | 3406 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3407 | return false; |
3408 | ||
8644a673 | 3409 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3410 | return false; |
3411 | ||
3412 | /* Is vectorizable load? */ | |
3413 | if (!is_gimple_assign (stmt)) | |
3414 | return false; | |
3415 | ||
3416 | scalar_dest = gimple_assign_lhs (stmt); | |
3417 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
3418 | return false; | |
3419 | ||
3420 | code = gimple_assign_rhs_code (stmt); | |
3421 | if (code != ARRAY_REF | |
3422 | && code != INDIRECT_REF | |
e9dbe7bb IR |
3423 | && code != COMPONENT_REF |
3424 | && code != IMAGPART_EXPR | |
70f34814 RG |
3425 | && code != REALPART_EXPR |
3426 | && code != MEM_REF) | |
ebfd146a IR |
3427 | return false; |
3428 | ||
3429 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
3430 | return false; | |
3431 | ||
3432 | scalar_type = TREE_TYPE (DR_REF (dr)); | |
3433 | mode = (int) TYPE_MODE (vectype); | |
3434 | ||
3435 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
3436 | (e.g. - data copies). */ | |
3437 | if (optab_handler (mov_optab, mode)->insn_code == CODE_FOR_nothing) | |
3438 | { | |
3439 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3440 | fprintf (vect_dump, "Aligned load, but unsupported type."); | |
3441 | return false; | |
3442 | } | |
3443 | ||
3444 | /* The vector component type needs to be trivially convertible to the | |
3445 | scalar lhs. This should always be the case. */ | |
3446 | if (!useless_type_conversion_p (TREE_TYPE (scalar_dest), TREE_TYPE (vectype))) | |
b8698a0f | 3447 | { |
ebfd146a IR |
3448 | if (vect_print_dump_info (REPORT_DETAILS)) |
3449 | fprintf (vect_dump, "??? operands of different types"); | |
3450 | return false; | |
3451 | } | |
3452 | ||
3453 | /* Check if the load is a part of an interleaving chain. */ | |
3454 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) | |
3455 | { | |
3456 | strided_load = true; | |
3457 | /* FORNOW */ | |
3458 | gcc_assert (! nested_in_vect_loop); | |
3459 | ||
3460 | /* Check if interleaving is supported. */ | |
3461 | if (!vect_strided_load_supported (vectype) | |
3462 | && !PURE_SLP_STMT (stmt_info) && !slp) | |
3463 | return false; | |
3464 | } | |
3465 | ||
3466 | if (!vec_stmt) /* transformation not required. */ | |
3467 | { | |
3468 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
3469 | vect_model_load_cost (stmt_info, ncopies, NULL); | |
3470 | return true; | |
3471 | } | |
3472 | ||
3473 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3474 | fprintf (vect_dump, "transform load."); | |
3475 | ||
3476 | /** Transform. **/ | |
3477 | ||
3478 | if (strided_load) | |
3479 | { | |
3480 | first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
3481 | /* Check if the chain of loads is already vectorized. */ | |
3482 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt))) | |
3483 | { | |
3484 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
3485 | return true; | |
3486 | } | |
3487 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
3488 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
3489 | ||
3490 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
3491 | if (slp) | |
3492 | { | |
3493 | strided_load = false; | |
3494 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
a70d6342 IR |
3495 | if (SLP_INSTANCE_LOAD_PERMUTATION (slp_node_instance)) |
3496 | slp_perm = true; | |
3497 | } | |
ebfd146a IR |
3498 | else |
3499 | vec_num = group_size; | |
3500 | ||
3501 | dr_chain = VEC_alloc (tree, heap, vec_num); | |
3502 | } | |
3503 | else | |
3504 | { | |
3505 | first_stmt = stmt; | |
3506 | first_dr = dr; | |
3507 | group_size = vec_num = 1; | |
3508 | } | |
3509 | ||
3510 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr); | |
3511 | gcc_assert (alignment_support_scheme); | |
3512 | ||
3513 | /* In case the vectorization factor (VF) is bigger than the number | |
3514 | of elements that we can fit in a vectype (nunits), we have to generate | |
3515 | more than one vector stmt - i.e - we need to "unroll" the | |
3516 | vector stmt by a factor VF/nunits. In doing so, we record a pointer | |
3517 | from one copy of the vector stmt to the next, in the field | |
3518 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
3519 | stages to find the correct vector defs to be used when vectorizing | |
3520 | stmts that use the defs of the current stmt. The example below illustrates | |
3521 | the vectorization process when VF=16 and nunits=4 (i.e - we need to create | |
3522 | 4 vectorized stmts): | |
3523 | ||
3524 | before vectorization: | |
3525 | RELATED_STMT VEC_STMT | |
3526 | S1: x = memref - - | |
3527 | S2: z = x + 1 - - | |
3528 | ||
3529 | step 1: vectorize stmt S1: | |
3530 | We first create the vector stmt VS1_0, and, as usual, record a | |
3531 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
3532 | Next, we create the vector stmt VS1_1, and record a pointer to | |
3533 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
3534 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of | |
3535 | stmts and pointers: | |
3536 | RELATED_STMT VEC_STMT | |
3537 | VS1_0: vx0 = memref0 VS1_1 - | |
3538 | VS1_1: vx1 = memref1 VS1_2 - | |
3539 | VS1_2: vx2 = memref2 VS1_3 - | |
3540 | VS1_3: vx3 = memref3 - - | |
3541 | S1: x = load - VS1_0 | |
3542 | S2: z = x + 1 - - | |
3543 | ||
b8698a0f L |
3544 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
3545 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
3546 | stmt S2. */ |
3547 | ||
3548 | /* In case of interleaving (non-unit strided access): | |
3549 | ||
3550 | S1: x2 = &base + 2 | |
3551 | S2: x0 = &base | |
3552 | S3: x1 = &base + 1 | |
3553 | S4: x3 = &base + 3 | |
3554 | ||
b8698a0f | 3555 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
3556 | starting from the access of the first stmt of the chain: |
3557 | ||
3558 | VS1: vx0 = &base | |
3559 | VS2: vx1 = &base + vec_size*1 | |
3560 | VS3: vx3 = &base + vec_size*2 | |
3561 | VS4: vx4 = &base + vec_size*3 | |
3562 | ||
3563 | Then permutation statements are generated: | |
3564 | ||
3565 | VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 > | |
3566 | VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 > | |
3567 | ... | |
3568 | ||
3569 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
3570 | (the order of the data-refs in the output of vect_permute_load_chain | |
3571 | corresponds to the order of scalar stmts in the interleaving chain - see | |
3572 | the documentation of vect_permute_load_chain()). | |
3573 | The generation of permutation stmts and recording them in | |
3574 | STMT_VINFO_VEC_STMT is done in vect_transform_strided_load(). | |
3575 | ||
b8698a0f | 3576 | In case of both multiple types and interleaving, the vector loads and |
ebfd146a IR |
3577 | permutation stmts above are created for every copy. The result vector stmts |
3578 | are put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding | |
3579 | STMT_VINFO_RELATED_STMT for the next copies. */ | |
3580 | ||
3581 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
3582 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
3583 | we generate the following code: | |
3584 | p = initial_addr; | |
3585 | indx = 0; | |
3586 | loop { | |
3587 | p = p + indx * vectype_size; | |
3588 | vec_dest = *(p); | |
3589 | indx = indx + 1; | |
3590 | } | |
3591 | ||
3592 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 3593 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
3594 | then generate the following code, in which the data in each iteration is |
3595 | obtained by two vector loads, one from the previous iteration, and one | |
3596 | from the current iteration: | |
3597 | p1 = initial_addr; | |
3598 | msq_init = *(floor(p1)) | |
3599 | p2 = initial_addr + VS - 1; | |
3600 | realignment_token = call target_builtin; | |
3601 | indx = 0; | |
3602 | loop { | |
3603 | p2 = p2 + indx * vectype_size | |
3604 | lsq = *(floor(p2)) | |
3605 | vec_dest = realign_load (msq, lsq, realignment_token) | |
3606 | indx = indx + 1; | |
3607 | msq = lsq; | |
3608 | } */ | |
3609 | ||
3610 | /* If the misalignment remains the same throughout the execution of the | |
3611 | loop, we can create the init_addr and permutation mask at the loop | |
3612 | preheader. Otherwise, it needs to be created inside the loop. | |
3613 | This can only occur when vectorizing memory accesses in the inner-loop | |
3614 | nested within an outer-loop that is being vectorized. */ | |
3615 | ||
a70d6342 | 3616 | if (loop && nested_in_vect_loop_p (loop, stmt) |
ebfd146a IR |
3617 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
3618 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) | |
3619 | { | |
3620 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
3621 | compute_in_loop = true; | |
3622 | } | |
3623 | ||
3624 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
3625 | || alignment_support_scheme == dr_explicit_realign) | |
3626 | && !compute_in_loop) | |
3627 | { | |
3628 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
3629 | alignment_support_scheme, NULL_TREE, | |
3630 | &at_loop); | |
3631 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
3632 | { | |
3633 | phi = SSA_NAME_DEF_STMT (msq); | |
3634 | offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
3635 | } | |
3636 | } | |
3637 | else | |
3638 | at_loop = loop; | |
3639 | ||
3640 | prev_stmt_info = NULL; | |
3641 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 3642 | { |
ebfd146a IR |
3643 | /* 1. Create the vector pointer update chain. */ |
3644 | if (j == 0) | |
3645 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, | |
b8698a0f L |
3646 | at_loop, offset, |
3647 | &dummy, &ptr_incr, false, | |
5006671f | 3648 | &inv_p); |
ebfd146a | 3649 | else |
b8698a0f | 3650 | dataref_ptr = |
ebfd146a IR |
3651 | bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE); |
3652 | ||
3653 | for (i = 0; i < vec_num; i++) | |
3654 | { | |
3655 | if (i > 0) | |
3656 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
3657 | NULL_TREE); | |
3658 | ||
3659 | /* 2. Create the vector-load in the loop. */ | |
3660 | switch (alignment_support_scheme) | |
3661 | { | |
3662 | case dr_aligned: | |
3663 | gcc_assert (aligned_access_p (first_dr)); | |
70f34814 | 3664 | data_ref = build_simple_mem_ref (dataref_ptr); |
ebfd146a IR |
3665 | break; |
3666 | case dr_unaligned_supported: | |
3667 | { | |
3668 | int mis = DR_MISALIGNMENT (first_dr); | |
3669 | tree tmis = (mis == -1 ? size_zero_node : size_int (mis)); | |
3670 | ||
3671 | tmis = size_binop (MULT_EXPR, tmis, size_int(BITS_PER_UNIT)); | |
3672 | data_ref = | |
3673 | build2 (MISALIGNED_INDIRECT_REF, vectype, dataref_ptr, tmis); | |
3674 | break; | |
3675 | } | |
3676 | case dr_explicit_realign: | |
3677 | { | |
3678 | tree ptr, bump; | |
3679 | tree vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
3680 | ||
3681 | if (compute_in_loop) | |
3682 | msq = vect_setup_realignment (first_stmt, gsi, | |
3683 | &realignment_token, | |
b8698a0f | 3684 | dr_explicit_realign, |
ebfd146a IR |
3685 | dataref_ptr, NULL); |
3686 | ||
3687 | data_ref = build1 (ALIGN_INDIRECT_REF, vectype, dataref_ptr); | |
3688 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3689 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
3690 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3691 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5006671f RG |
3692 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); |
3693 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
ebfd146a | 3694 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
3695 | msq = new_temp; |
3696 | ||
3697 | bump = size_binop (MULT_EXPR, vs_minus_1, | |
3698 | TYPE_SIZE_UNIT (scalar_type)); | |
3699 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); | |
3700 | data_ref = build1 (ALIGN_INDIRECT_REF, vectype, ptr); | |
3701 | break; | |
3702 | } | |
3703 | case dr_explicit_realign_optimized: | |
3704 | data_ref = build1 (ALIGN_INDIRECT_REF, vectype, dataref_ptr); | |
3705 | break; | |
3706 | default: | |
3707 | gcc_unreachable (); | |
3708 | } | |
5006671f RG |
3709 | /* If accesses through a pointer to vectype do not alias the original |
3710 | memory reference we have a problem. This should never happen. */ | |
3711 | gcc_assert (alias_sets_conflict_p (get_alias_set (data_ref), | |
3712 | get_alias_set (gimple_assign_rhs1 (stmt)))); | |
ebfd146a IR |
3713 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
3714 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
3715 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3716 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3717 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3718 | mark_symbols_for_renaming (new_stmt); | |
3719 | ||
3720 | /* 3. Handle explicit realignment if necessary/supported. Create in | |
3721 | loop: vec_dest = realign_load (msq, lsq, realignment_token) */ | |
3722 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
3723 | || alignment_support_scheme == dr_explicit_realign) | |
3724 | { | |
3725 | tree tmp; | |
3726 | ||
3727 | lsq = gimple_assign_lhs (new_stmt); | |
3728 | if (!realignment_token) | |
3729 | realignment_token = dataref_ptr; | |
3730 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3731 | tmp = build3 (REALIGN_LOAD_EXPR, vectype, msq, lsq, | |
3732 | realignment_token); | |
3733 | new_stmt = gimple_build_assign (vec_dest, tmp); | |
3734 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3735 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3736 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3737 | ||
3738 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
3739 | { | |
3740 | gcc_assert (phi); | |
3741 | if (i == vec_num - 1 && j == ncopies - 1) | |
f5045c96 AM |
3742 | add_phi_arg (phi, lsq, loop_latch_edge (containing_loop), |
3743 | UNKNOWN_LOCATION); | |
ebfd146a IR |
3744 | msq = lsq; |
3745 | } | |
3746 | } | |
3747 | ||
3748 | /* 4. Handle invariant-load. */ | |
a70d6342 | 3749 | if (inv_p && !bb_vinfo) |
ebfd146a IR |
3750 | { |
3751 | gcc_assert (!strided_load); | |
3752 | gcc_assert (nested_in_vect_loop_p (loop, stmt)); | |
3753 | if (j == 0) | |
3754 | { | |
3755 | int k; | |
3756 | tree t = NULL_TREE; | |
3757 | tree vec_inv, bitpos, bitsize = TYPE_SIZE (scalar_type); | |
3758 | ||
3759 | /* CHECKME: bitpos depends on endianess? */ | |
3760 | bitpos = bitsize_zero_node; | |
b8698a0f | 3761 | vec_inv = build3 (BIT_FIELD_REF, scalar_type, new_temp, |
ebfd146a | 3762 | bitsize, bitpos); |
b8698a0f | 3763 | vec_dest = |
ebfd146a IR |
3764 | vect_create_destination_var (scalar_dest, NULL_TREE); |
3765 | new_stmt = gimple_build_assign (vec_dest, vec_inv); | |
3766 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3767 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3768 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3769 | ||
3770 | for (k = nunits - 1; k >= 0; --k) | |
3771 | t = tree_cons (NULL_TREE, new_temp, t); | |
3772 | /* FIXME: use build_constructor directly. */ | |
3773 | vec_inv = build_constructor_from_list (vectype, t); | |
3774 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
3775 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
3776 | } | |
3777 | else | |
3778 | gcc_unreachable (); /* FORNOW. */ | |
3779 | } | |
3780 | ||
3781 | /* Collect vector loads and later create their permutation in | |
3782 | vect_transform_strided_load (). */ | |
3783 | if (strided_load || slp_perm) | |
3784 | VEC_quick_push (tree, dr_chain, new_temp); | |
3785 | ||
3786 | /* Store vector loads in the corresponding SLP_NODE. */ | |
3787 | if (slp && !slp_perm) | |
3788 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
3789 | } | |
3790 | ||
3791 | if (slp && !slp_perm) | |
3792 | continue; | |
3793 | ||
3794 | if (slp_perm) | |
3795 | { | |
a70d6342 | 3796 | if (!vect_transform_slp_perm_load (stmt, dr_chain, gsi, vf, |
ebfd146a IR |
3797 | slp_node_instance, false)) |
3798 | { | |
3799 | VEC_free (tree, heap, dr_chain); | |
3800 | return false; | |
3801 | } | |
3802 | } | |
3803 | else | |
3804 | { | |
3805 | if (strided_load) | |
3806 | { | |
3807 | if (!vect_transform_strided_load (stmt, dr_chain, group_size, gsi)) | |
b8698a0f | 3808 | return false; |
ebfd146a IR |
3809 | |
3810 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
3811 | VEC_free (tree, heap, dr_chain); | |
3812 | dr_chain = VEC_alloc (tree, heap, group_size); | |
3813 | } | |
3814 | else | |
3815 | { | |
3816 | if (j == 0) | |
3817 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3818 | else | |
3819 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3820 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3821 | } | |
3822 | } | |
3823 | } | |
3824 | ||
3825 | if (dr_chain) | |
3826 | VEC_free (tree, heap, dr_chain); | |
3827 | ||
3828 | return true; | |
3829 | } | |
3830 | ||
3831 | /* Function vect_is_simple_cond. | |
b8698a0f | 3832 | |
ebfd146a IR |
3833 | Input: |
3834 | LOOP - the loop that is being vectorized. | |
3835 | COND - Condition that is checked for simple use. | |
3836 | ||
3837 | Returns whether a COND can be vectorized. Checks whether | |
3838 | condition operands are supportable using vec_is_simple_use. */ | |
3839 | ||
3840 | static bool | |
3841 | vect_is_simple_cond (tree cond, loop_vec_info loop_vinfo) | |
3842 | { | |
3843 | tree lhs, rhs; | |
3844 | tree def; | |
3845 | enum vect_def_type dt; | |
3846 | ||
3847 | if (!COMPARISON_CLASS_P (cond)) | |
3848 | return false; | |
3849 | ||
3850 | lhs = TREE_OPERAND (cond, 0); | |
3851 | rhs = TREE_OPERAND (cond, 1); | |
3852 | ||
3853 | if (TREE_CODE (lhs) == SSA_NAME) | |
3854 | { | |
3855 | gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); | |
b8698a0f | 3856 | if (!vect_is_simple_use (lhs, loop_vinfo, NULL, &lhs_def_stmt, &def, |
a70d6342 | 3857 | &dt)) |
ebfd146a IR |
3858 | return false; |
3859 | } | |
3860 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
3861 | && TREE_CODE (lhs) != FIXED_CST) | |
3862 | return false; | |
3863 | ||
3864 | if (TREE_CODE (rhs) == SSA_NAME) | |
3865 | { | |
3866 | gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); | |
b8698a0f | 3867 | if (!vect_is_simple_use (rhs, loop_vinfo, NULL, &rhs_def_stmt, &def, |
a70d6342 | 3868 | &dt)) |
ebfd146a IR |
3869 | return false; |
3870 | } | |
3871 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST | |
3872 | && TREE_CODE (rhs) != FIXED_CST) | |
3873 | return false; | |
3874 | ||
3875 | return true; | |
3876 | } | |
3877 | ||
3878 | /* vectorizable_condition. | |
3879 | ||
b8698a0f L |
3880 | Check if STMT is conditional modify expression that can be vectorized. |
3881 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3882 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
3883 | at GSI. |
3884 | ||
3885 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
3886 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
3887 | else caluse if it is 2). | |
ebfd146a IR |
3888 | |
3889 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3890 | ||
4bbe8262 | 3891 | bool |
ebfd146a | 3892 | vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi, |
4bbe8262 | 3893 | gimple *vec_stmt, tree reduc_def, int reduc_index) |
ebfd146a IR |
3894 | { |
3895 | tree scalar_dest = NULL_TREE; | |
3896 | tree vec_dest = NULL_TREE; | |
3897 | tree op = NULL_TREE; | |
3898 | tree cond_expr, then_clause, else_clause; | |
3899 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3900 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3901 | tree vec_cond_lhs, vec_cond_rhs, vec_then_clause, vec_else_clause; | |
3902 | tree vec_compare, vec_cond_expr; | |
3903 | tree new_temp; | |
3904 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3905 | enum machine_mode vec_mode; | |
3906 | tree def; | |
3907 | enum vect_def_type dt; | |
3908 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
3909 | int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3910 | enum tree_code code; | |
3911 | ||
a70d6342 IR |
3912 | /* FORNOW: unsupported in basic block SLP. */ |
3913 | gcc_assert (loop_vinfo); | |
b8698a0f | 3914 | |
ebfd146a IR |
3915 | gcc_assert (ncopies >= 1); |
3916 | if (ncopies > 1) | |
3917 | return false; /* FORNOW */ | |
3918 | ||
3919 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
3920 | return false; | |
3921 | ||
4bbe8262 IR |
3922 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3923 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
3924 | && reduc_def)) | |
ebfd146a IR |
3925 | return false; |
3926 | ||
3927 | /* FORNOW: SLP not supported. */ | |
3928 | if (STMT_SLP_TYPE (stmt_info)) | |
3929 | return false; | |
3930 | ||
3931 | /* FORNOW: not yet supported. */ | |
b8698a0f | 3932 | if (STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a IR |
3933 | { |
3934 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3935 | fprintf (vect_dump, "value used after loop."); | |
3936 | return false; | |
3937 | } | |
3938 | ||
3939 | /* Is vectorizable conditional operation? */ | |
3940 | if (!is_gimple_assign (stmt)) | |
3941 | return false; | |
3942 | ||
3943 | code = gimple_assign_rhs_code (stmt); | |
3944 | ||
3945 | if (code != COND_EXPR) | |
3946 | return false; | |
3947 | ||
3948 | gcc_assert (gimple_assign_single_p (stmt)); | |
3949 | op = gimple_assign_rhs1 (stmt); | |
3950 | cond_expr = TREE_OPERAND (op, 0); | |
3951 | then_clause = TREE_OPERAND (op, 1); | |
3952 | else_clause = TREE_OPERAND (op, 2); | |
3953 | ||
3954 | if (!vect_is_simple_cond (cond_expr, loop_vinfo)) | |
3955 | return false; | |
3956 | ||
3957 | /* We do not handle two different vector types for the condition | |
3958 | and the values. */ | |
8533c9d8 SP |
3959 | if (!types_compatible_p (TREE_TYPE (TREE_OPERAND (cond_expr, 0)), |
3960 | TREE_TYPE (vectype))) | |
ebfd146a IR |
3961 | return false; |
3962 | ||
3963 | if (TREE_CODE (then_clause) == SSA_NAME) | |
3964 | { | |
3965 | gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause); | |
b8698a0f | 3966 | if (!vect_is_simple_use (then_clause, loop_vinfo, NULL, |
ebfd146a IR |
3967 | &then_def_stmt, &def, &dt)) |
3968 | return false; | |
3969 | } | |
b8698a0f | 3970 | else if (TREE_CODE (then_clause) != INTEGER_CST |
ebfd146a IR |
3971 | && TREE_CODE (then_clause) != REAL_CST |
3972 | && TREE_CODE (then_clause) != FIXED_CST) | |
3973 | return false; | |
3974 | ||
3975 | if (TREE_CODE (else_clause) == SSA_NAME) | |
3976 | { | |
3977 | gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause); | |
a70d6342 | 3978 | if (!vect_is_simple_use (else_clause, loop_vinfo, NULL, |
ebfd146a IR |
3979 | &else_def_stmt, &def, &dt)) |
3980 | return false; | |
3981 | } | |
b8698a0f | 3982 | else if (TREE_CODE (else_clause) != INTEGER_CST |
ebfd146a IR |
3983 | && TREE_CODE (else_clause) != REAL_CST |
3984 | && TREE_CODE (else_clause) != FIXED_CST) | |
3985 | return false; | |
3986 | ||
3987 | ||
3988 | vec_mode = TYPE_MODE (vectype); | |
3989 | ||
b8698a0f | 3990 | if (!vec_stmt) |
ebfd146a IR |
3991 | { |
3992 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
8e7aa1f9 | 3993 | return expand_vec_cond_expr_p (TREE_TYPE (op), vec_mode); |
ebfd146a IR |
3994 | } |
3995 | ||
3996 | /* Transform */ | |
3997 | ||
3998 | /* Handle def. */ | |
3999 | scalar_dest = gimple_assign_lhs (stmt); | |
4000 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4001 | ||
4002 | /* Handle cond expr. */ | |
b8698a0f | 4003 | vec_cond_lhs = |
ebfd146a | 4004 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), stmt, NULL); |
b8698a0f | 4005 | vec_cond_rhs = |
ebfd146a | 4006 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), stmt, NULL); |
4bbe8262 IR |
4007 | if (reduc_index == 1) |
4008 | vec_then_clause = reduc_def; | |
4009 | else | |
4010 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, stmt, NULL); | |
4011 | if (reduc_index == 2) | |
4012 | vec_else_clause = reduc_def; | |
4013 | else | |
4014 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, stmt, NULL); | |
ebfd146a IR |
4015 | |
4016 | /* Arguments are ready. Create the new vector stmt. */ | |
b8698a0f | 4017 | vec_compare = build2 (TREE_CODE (cond_expr), vectype, |
ebfd146a | 4018 | vec_cond_lhs, vec_cond_rhs); |
b8698a0f | 4019 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, |
ebfd146a IR |
4020 | vec_compare, vec_then_clause, vec_else_clause); |
4021 | ||
4022 | *vec_stmt = gimple_build_assign (vec_dest, vec_cond_expr); | |
4023 | new_temp = make_ssa_name (vec_dest, *vec_stmt); | |
4024 | gimple_assign_set_lhs (*vec_stmt, new_temp); | |
4025 | vect_finish_stmt_generation (stmt, *vec_stmt, gsi); | |
b8698a0f | 4026 | |
ebfd146a IR |
4027 | return true; |
4028 | } | |
4029 | ||
4030 | ||
8644a673 | 4031 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
4032 | |
4033 | bool | |
a70d6342 | 4034 | vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 4035 | { |
8644a673 | 4036 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 4037 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 4038 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 4039 | bool ok; |
a70d6342 | 4040 | tree scalar_type, vectype; |
ebfd146a IR |
4041 | |
4042 | if (vect_print_dump_info (REPORT_DETAILS)) | |
ebfd146a | 4043 | { |
8644a673 IR |
4044 | fprintf (vect_dump, "==> examining statement: "); |
4045 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4046 | } | |
ebfd146a | 4047 | |
1825a1f3 | 4048 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 4049 | { |
1825a1f3 IR |
4050 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
4051 | fprintf (vect_dump, "not vectorized: stmt has volatile operands"); | |
4052 | ||
4053 | return false; | |
4054 | } | |
b8698a0f L |
4055 | |
4056 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
4057 | to include: |
4058 | - the COND_EXPR which is the loop exit condition | |
4059 | - any LABEL_EXPRs in the loop | |
b8698a0f | 4060 | - computations that are used only for array indexing or loop control. |
8644a673 IR |
4061 | In basic blocks we only analyze statements that are a part of some SLP |
4062 | instance, therefore, all the statements are relevant. */ | |
ebfd146a | 4063 | |
b8698a0f | 4064 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 4065 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a IR |
4066 | { |
4067 | if (vect_print_dump_info (REPORT_DETAILS)) | |
8644a673 | 4068 | fprintf (vect_dump, "irrelevant."); |
ebfd146a | 4069 | |
8644a673 IR |
4070 | return true; |
4071 | } | |
ebfd146a | 4072 | |
8644a673 IR |
4073 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
4074 | { | |
4075 | case vect_internal_def: | |
4076 | break; | |
ebfd146a | 4077 | |
8644a673 | 4078 | case vect_reduction_def: |
7c5222ff | 4079 | case vect_nested_cycle: |
a70d6342 | 4080 | gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer |
8644a673 | 4081 | || relevance == vect_used_in_outer_by_reduction |
a70d6342 | 4082 | || relevance == vect_unused_in_scope)); |
8644a673 IR |
4083 | break; |
4084 | ||
4085 | case vect_induction_def: | |
4086 | case vect_constant_def: | |
4087 | case vect_external_def: | |
4088 | case vect_unknown_def_type: | |
4089 | default: | |
4090 | gcc_unreachable (); | |
4091 | } | |
ebfd146a | 4092 | |
a70d6342 IR |
4093 | if (bb_vinfo) |
4094 | { | |
4095 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
4096 | ||
b690cc0f | 4097 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
a70d6342 IR |
4098 | if (vect_print_dump_info (REPORT_DETAILS)) |
4099 | { | |
4100 | fprintf (vect_dump, "get vectype for scalar type: "); | |
4101 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
4102 | } | |
4103 | ||
4104 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4105 | if (!vectype) | |
4106 | { | |
4107 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4108 | { | |
4109 | fprintf (vect_dump, "not SLPed: unsupported data-type "); | |
4110 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
4111 | } | |
4112 | return false; | |
4113 | } | |
4114 | ||
4115 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4116 | { | |
4117 | fprintf (vect_dump, "vectype: "); | |
4118 | print_generic_expr (vect_dump, vectype, TDF_SLIM); | |
4119 | } | |
4120 | ||
4121 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
4122 | } | |
4123 | ||
8644a673 | 4124 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 4125 | { |
8644a673 IR |
4126 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
4127 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info)); | |
4128 | *need_to_vectorize = true; | |
ebfd146a IR |
4129 | } |
4130 | ||
8644a673 | 4131 | ok = true; |
b8698a0f | 4132 | if (!bb_vinfo |
a70d6342 IR |
4133 | && (STMT_VINFO_RELEVANT_P (stmt_info) |
4134 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
8644a673 IR |
4135 | ok = (vectorizable_type_promotion (stmt, NULL, NULL, NULL) |
4136 | || vectorizable_type_demotion (stmt, NULL, NULL, NULL) | |
4137 | || vectorizable_conversion (stmt, NULL, NULL, NULL) | |
4138 | || vectorizable_operation (stmt, NULL, NULL, NULL) | |
4139 | || vectorizable_assignment (stmt, NULL, NULL, NULL) | |
4140 | || vectorizable_load (stmt, NULL, NULL, NULL, NULL) | |
4141 | || vectorizable_call (stmt, NULL, NULL) | |
4142 | || vectorizable_store (stmt, NULL, NULL, NULL) | |
b5aeb3bb | 4143 | || vectorizable_reduction (stmt, NULL, NULL, NULL) |
4bbe8262 | 4144 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0)); |
a70d6342 IR |
4145 | else |
4146 | { | |
4147 | if (bb_vinfo) | |
4148 | ok = (vectorizable_operation (stmt, NULL, NULL, node) | |
4149 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
4150 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
4151 | || vectorizable_store (stmt, NULL, NULL, node)); | |
b8698a0f | 4152 | } |
8644a673 IR |
4153 | |
4154 | if (!ok) | |
ebfd146a | 4155 | { |
8644a673 IR |
4156 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
4157 | { | |
4158 | fprintf (vect_dump, "not vectorized: relevant stmt not "); | |
4159 | fprintf (vect_dump, "supported: "); | |
4160 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4161 | } | |
b8698a0f | 4162 | |
ebfd146a IR |
4163 | return false; |
4164 | } | |
4165 | ||
a70d6342 IR |
4166 | if (bb_vinfo) |
4167 | return true; | |
4168 | ||
8644a673 IR |
4169 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
4170 | need extra handling, except for vectorizable reductions. */ | |
4171 | if (STMT_VINFO_LIVE_P (stmt_info) | |
4172 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
4173 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 4174 | |
8644a673 | 4175 | if (!ok) |
ebfd146a | 4176 | { |
8644a673 IR |
4177 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
4178 | { | |
4179 | fprintf (vect_dump, "not vectorized: live stmt not "); | |
4180 | fprintf (vect_dump, "supported: "); | |
4181 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4182 | } | |
b8698a0f | 4183 | |
8644a673 | 4184 | return false; |
ebfd146a IR |
4185 | } |
4186 | ||
8644a673 | 4187 | if (!PURE_SLP_STMT (stmt_info)) |
ebfd146a | 4188 | { |
b8698a0f L |
4189 | /* Groups of strided accesses whose size is not a power of 2 are not |
4190 | vectorizable yet using loop-vectorization. Therefore, if this stmt | |
4191 | feeds non-SLP-able stmts (i.e., this stmt has to be both SLPed and | |
a70d6342 | 4192 | loop-based vectorized), the loop cannot be vectorized. */ |
8644a673 IR |
4193 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) |
4194 | && exact_log2 (DR_GROUP_SIZE (vinfo_for_stmt ( | |
4195 | DR_GROUP_FIRST_DR (stmt_info)))) == -1) | |
ebfd146a | 4196 | { |
8644a673 IR |
4197 | if (vect_print_dump_info (REPORT_DETAILS)) |
4198 | { | |
4199 | fprintf (vect_dump, "not vectorized: the size of group " | |
4200 | "of strided accesses is not a power of 2"); | |
4201 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4202 | } | |
4203 | ||
ebfd146a IR |
4204 | return false; |
4205 | } | |
4206 | } | |
b8698a0f | 4207 | |
ebfd146a IR |
4208 | return true; |
4209 | } | |
4210 | ||
4211 | ||
4212 | /* Function vect_transform_stmt. | |
4213 | ||
4214 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
4215 | ||
4216 | bool | |
4217 | vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi, | |
b8698a0f | 4218 | bool *strided_store, slp_tree slp_node, |
ebfd146a IR |
4219 | slp_instance slp_node_instance) |
4220 | { | |
4221 | bool is_store = false; | |
4222 | gimple vec_stmt = NULL; | |
4223 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4224 | gimple orig_stmt_in_pattern; | |
4225 | bool done; | |
ebfd146a IR |
4226 | |
4227 | switch (STMT_VINFO_TYPE (stmt_info)) | |
4228 | { | |
4229 | case type_demotion_vec_info_type: | |
4230 | done = vectorizable_type_demotion (stmt, gsi, &vec_stmt, slp_node); | |
4231 | gcc_assert (done); | |
4232 | break; | |
4233 | ||
4234 | case type_promotion_vec_info_type: | |
4235 | done = vectorizable_type_promotion (stmt, gsi, &vec_stmt, slp_node); | |
4236 | gcc_assert (done); | |
4237 | break; | |
4238 | ||
4239 | case type_conversion_vec_info_type: | |
4240 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
4241 | gcc_assert (done); | |
4242 | break; | |
4243 | ||
4244 | case induc_vec_info_type: | |
4245 | gcc_assert (!slp_node); | |
4246 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
4247 | gcc_assert (done); | |
4248 | break; | |
4249 | ||
4250 | case op_vec_info_type: | |
4251 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
4252 | gcc_assert (done); | |
4253 | break; | |
4254 | ||
4255 | case assignment_vec_info_type: | |
4256 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
4257 | gcc_assert (done); | |
4258 | break; | |
4259 | ||
4260 | case load_vec_info_type: | |
b8698a0f | 4261 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
4262 | slp_node_instance); |
4263 | gcc_assert (done); | |
4264 | break; | |
4265 | ||
4266 | case store_vec_info_type: | |
4267 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
4268 | gcc_assert (done); | |
4269 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) && !slp_node) | |
4270 | { | |
4271 | /* In case of interleaving, the whole chain is vectorized when the | |
4272 | last store in the chain is reached. Store stmts before the last | |
4273 | one are skipped, and there vec_stmt_info shouldn't be freed | |
4274 | meanwhile. */ | |
4275 | *strided_store = true; | |
4276 | if (STMT_VINFO_VEC_STMT (stmt_info)) | |
4277 | is_store = true; | |
4278 | } | |
4279 | else | |
4280 | is_store = true; | |
4281 | break; | |
4282 | ||
4283 | case condition_vec_info_type: | |
4284 | gcc_assert (!slp_node); | |
4bbe8262 | 4285 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0); |
ebfd146a IR |
4286 | gcc_assert (done); |
4287 | break; | |
4288 | ||
4289 | case call_vec_info_type: | |
4290 | gcc_assert (!slp_node); | |
4291 | done = vectorizable_call (stmt, gsi, &vec_stmt); | |
4292 | break; | |
4293 | ||
4294 | case reduc_vec_info_type: | |
b5aeb3bb | 4295 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
4296 | gcc_assert (done); |
4297 | break; | |
4298 | ||
4299 | default: | |
4300 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
4301 | { | |
4302 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4303 | fprintf (vect_dump, "stmt not supported."); | |
4304 | gcc_unreachable (); | |
4305 | } | |
4306 | } | |
4307 | ||
4308 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that | |
4309 | is being vectorized, but outside the immediately enclosing loop. */ | |
4310 | if (vec_stmt | |
a70d6342 IR |
4311 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
4312 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
4313 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
4314 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
4315 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 4316 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 4317 | vect_used_in_outer_by_reduction)) |
ebfd146a | 4318 | { |
a70d6342 IR |
4319 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
4320 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
4321 | imm_use_iterator imm_iter; |
4322 | use_operand_p use_p; | |
4323 | tree scalar_dest; | |
4324 | gimple exit_phi; | |
4325 | ||
4326 | if (vect_print_dump_info (REPORT_DETAILS)) | |
a70d6342 | 4327 | fprintf (vect_dump, "Record the vdef for outer-loop vectorization."); |
ebfd146a IR |
4328 | |
4329 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
4330 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
4331 | STMT). */ | |
4332 | if (gimple_code (stmt) == GIMPLE_PHI) | |
4333 | scalar_dest = PHI_RESULT (stmt); | |
4334 | else | |
4335 | scalar_dest = gimple_assign_lhs (stmt); | |
4336 | ||
4337 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
4338 | { | |
4339 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
4340 | { | |
4341 | exit_phi = USE_STMT (use_p); | |
4342 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
4343 | } | |
4344 | } | |
4345 | } | |
4346 | ||
4347 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
4348 | being vectorized. */ | |
4349 | if (STMT_VINFO_LIVE_P (stmt_info) | |
4350 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
4351 | { | |
4352 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
4353 | gcc_assert (done); | |
4354 | } | |
4355 | ||
4356 | if (vec_stmt) | |
4357 | { | |
4358 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; | |
4359 | orig_stmt_in_pattern = STMT_VINFO_RELATED_STMT (stmt_info); | |
4360 | if (orig_stmt_in_pattern) | |
4361 | { | |
4362 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt_in_pattern); | |
4363 | /* STMT was inserted by the vectorizer to replace a computation idiom. | |
b8698a0f L |
4364 | ORIG_STMT_IN_PATTERN is a stmt in the original sequence that |
4365 | computed this idiom. We need to record a pointer to VEC_STMT in | |
4366 | the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the | |
ebfd146a IR |
4367 | documentation of vect_pattern_recog. */ |
4368 | if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo)) | |
4369 | { | |
4370 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) == stmt); | |
4371 | STMT_VINFO_VEC_STMT (stmt_vinfo) = vec_stmt; | |
4372 | } | |
4373 | } | |
4374 | } | |
4375 | ||
b8698a0f | 4376 | return is_store; |
ebfd146a IR |
4377 | } |
4378 | ||
4379 | ||
b8698a0f | 4380 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
4381 | stmt_vec_info. */ |
4382 | ||
4383 | void | |
4384 | vect_remove_stores (gimple first_stmt) | |
4385 | { | |
4386 | gimple next = first_stmt; | |
4387 | gimple tmp; | |
4388 | gimple_stmt_iterator next_si; | |
4389 | ||
4390 | while (next) | |
4391 | { | |
4392 | /* Free the attached stmt_vec_info and remove the stmt. */ | |
4393 | next_si = gsi_for_stmt (next); | |
4394 | gsi_remove (&next_si, true); | |
4395 | tmp = DR_GROUP_NEXT_DR (vinfo_for_stmt (next)); | |
4396 | free_stmt_vec_info (next); | |
4397 | next = tmp; | |
4398 | } | |
4399 | } | |
4400 | ||
4401 | ||
4402 | /* Function new_stmt_vec_info. | |
4403 | ||
4404 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
4405 | ||
4406 | stmt_vec_info | |
b8698a0f | 4407 | new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 4408 | bb_vec_info bb_vinfo) |
ebfd146a IR |
4409 | { |
4410 | stmt_vec_info res; | |
4411 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
4412 | ||
4413 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
4414 | STMT_VINFO_STMT (res) = stmt; | |
4415 | STMT_VINFO_LOOP_VINFO (res) = loop_vinfo; | |
a70d6342 | 4416 | STMT_VINFO_BB_VINFO (res) = bb_vinfo; |
8644a673 | 4417 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
4418 | STMT_VINFO_LIVE_P (res) = false; |
4419 | STMT_VINFO_VECTYPE (res) = NULL; | |
4420 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 4421 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
4422 | STMT_VINFO_IN_PATTERN_P (res) = false; |
4423 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
4424 | STMT_VINFO_DATA_REF (res) = NULL; | |
4425 | ||
4426 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
4427 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
4428 | STMT_VINFO_DR_INIT (res) = NULL; | |
4429 | STMT_VINFO_DR_STEP (res) = NULL; | |
4430 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
4431 | ||
4432 | if (gimple_code (stmt) == GIMPLE_PHI | |
4433 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
4434 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
4435 | else | |
8644a673 IR |
4436 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
4437 | ||
ebfd146a IR |
4438 | STMT_VINFO_SAME_ALIGN_REFS (res) = VEC_alloc (dr_p, heap, 5); |
4439 | STMT_VINFO_INSIDE_OF_LOOP_COST (res) = 0; | |
4440 | STMT_VINFO_OUTSIDE_OF_LOOP_COST (res) = 0; | |
32e8bb8e | 4441 | STMT_SLP_TYPE (res) = loop_vect; |
ebfd146a IR |
4442 | DR_GROUP_FIRST_DR (res) = NULL; |
4443 | DR_GROUP_NEXT_DR (res) = NULL; | |
4444 | DR_GROUP_SIZE (res) = 0; | |
4445 | DR_GROUP_STORE_COUNT (res) = 0; | |
4446 | DR_GROUP_GAP (res) = 0; | |
4447 | DR_GROUP_SAME_DR_STMT (res) = NULL; | |
4448 | DR_GROUP_READ_WRITE_DEPENDENCE (res) = false; | |
4449 | ||
4450 | return res; | |
4451 | } | |
4452 | ||
4453 | ||
4454 | /* Create a hash table for stmt_vec_info. */ | |
4455 | ||
4456 | void | |
4457 | init_stmt_vec_info_vec (void) | |
4458 | { | |
4459 | gcc_assert (!stmt_vec_info_vec); | |
4460 | stmt_vec_info_vec = VEC_alloc (vec_void_p, heap, 50); | |
4461 | } | |
4462 | ||
4463 | ||
4464 | /* Free hash table for stmt_vec_info. */ | |
4465 | ||
4466 | void | |
4467 | free_stmt_vec_info_vec (void) | |
4468 | { | |
4469 | gcc_assert (stmt_vec_info_vec); | |
4470 | VEC_free (vec_void_p, heap, stmt_vec_info_vec); | |
4471 | } | |
4472 | ||
4473 | ||
4474 | /* Free stmt vectorization related info. */ | |
4475 | ||
4476 | void | |
4477 | free_stmt_vec_info (gimple stmt) | |
4478 | { | |
4479 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4480 | ||
4481 | if (!stmt_info) | |
4482 | return; | |
4483 | ||
4484 | VEC_free (dr_p, heap, STMT_VINFO_SAME_ALIGN_REFS (stmt_info)); | |
4485 | set_vinfo_for_stmt (stmt, NULL); | |
4486 | free (stmt_info); | |
4487 | } | |
4488 | ||
4489 | ||
4490 | /* Function get_vectype_for_scalar_type. | |
4491 | ||
4492 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
4493 | by the target. */ | |
4494 | ||
4495 | tree | |
4496 | get_vectype_for_scalar_type (tree scalar_type) | |
4497 | { | |
4498 | enum machine_mode inner_mode = TYPE_MODE (scalar_type); | |
2f816591 | 4499 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
4500 | int nunits; |
4501 | tree vectype; | |
4502 | ||
4503 | if (nbytes == 0 || nbytes >= UNITS_PER_SIMD_WORD (inner_mode)) | |
4504 | return NULL_TREE; | |
4505 | ||
2f816591 RG |
4506 | /* We can't build a vector type of elements with alignment bigger than |
4507 | their size. */ | |
4508 | if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) | |
4509 | return NULL_TREE; | |
4510 | ||
6d7971b8 RG |
4511 | /* If we'd build a vector type of elements whose mode precision doesn't |
4512 | match their types precision we'll get mismatched types on vector | |
4513 | extracts via BIT_FIELD_REFs. This effectively means we disable | |
4514 | vectorization of bool and/or enum types in some languages. */ | |
4515 | if (INTEGRAL_TYPE_P (scalar_type) | |
4516 | && GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type)) | |
4517 | return NULL_TREE; | |
4518 | ||
ebfd146a IR |
4519 | /* FORNOW: Only a single vector size per mode (UNITS_PER_SIMD_WORD) |
4520 | is expected. */ | |
4521 | nunits = UNITS_PER_SIMD_WORD (inner_mode) / nbytes; | |
4522 | ||
4523 | vectype = build_vector_type (scalar_type, nunits); | |
4524 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4525 | { | |
4526 | fprintf (vect_dump, "get vectype with %d units of type ", nunits); | |
4527 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
4528 | } | |
4529 | ||
4530 | if (!vectype) | |
4531 | return NULL_TREE; | |
4532 | ||
4533 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4534 | { | |
4535 | fprintf (vect_dump, "vectype: "); | |
4536 | print_generic_expr (vect_dump, vectype, TDF_SLIM); | |
4537 | } | |
4538 | ||
4539 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
4540 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
4541 | { | |
4542 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4543 | fprintf (vect_dump, "mode not supported by target."); | |
4544 | return NULL_TREE; | |
4545 | } | |
4546 | ||
4547 | return vectype; | |
4548 | } | |
4549 | ||
b690cc0f RG |
4550 | /* Function get_same_sized_vectype |
4551 | ||
4552 | Returns a vector type corresponding to SCALAR_TYPE of size | |
4553 | VECTOR_TYPE if supported by the target. */ | |
4554 | ||
4555 | tree | |
4556 | get_same_sized_vectype (tree scalar_type, tree vector_type ATTRIBUTE_UNUSED) | |
4557 | { | |
4558 | return get_vectype_for_scalar_type (scalar_type); | |
4559 | } | |
4560 | ||
ebfd146a IR |
4561 | /* Function vect_is_simple_use. |
4562 | ||
4563 | Input: | |
a70d6342 IR |
4564 | LOOP_VINFO - the vect info of the loop that is being vectorized. |
4565 | BB_VINFO - the vect info of the basic block that is being vectorized. | |
4566 | OPERAND - operand of a stmt in the loop or bb. | |
ebfd146a IR |
4567 | DEF - the defining stmt in case OPERAND is an SSA_NAME. |
4568 | ||
4569 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f L |
4570 | For loops, supportable operands are constants, loop invariants, and operands |
4571 | that are defined by the current iteration of the loop. Unsupportable | |
4572 | operands are those that are defined by a previous iteration of the loop (as | |
a70d6342 IR |
4573 | is the case in reduction/induction computations). |
4574 | For basic blocks, supportable operands are constants and bb invariants. | |
4575 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
4576 | |
4577 | bool | |
b8698a0f | 4578 | vect_is_simple_use (tree operand, loop_vec_info loop_vinfo, |
a70d6342 | 4579 | bb_vec_info bb_vinfo, gimple *def_stmt, |
ebfd146a | 4580 | tree *def, enum vect_def_type *dt) |
b8698a0f | 4581 | { |
ebfd146a IR |
4582 | basic_block bb; |
4583 | stmt_vec_info stmt_vinfo; | |
a70d6342 | 4584 | struct loop *loop = NULL; |
b8698a0f | 4585 | |
a70d6342 IR |
4586 | if (loop_vinfo) |
4587 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
4588 | |
4589 | *def_stmt = NULL; | |
4590 | *def = NULL_TREE; | |
b8698a0f | 4591 | |
ebfd146a IR |
4592 | if (vect_print_dump_info (REPORT_DETAILS)) |
4593 | { | |
4594 | fprintf (vect_dump, "vect_is_simple_use: operand "); | |
4595 | print_generic_expr (vect_dump, operand, TDF_SLIM); | |
4596 | } | |
b8698a0f | 4597 | |
ebfd146a IR |
4598 | if (TREE_CODE (operand) == INTEGER_CST || TREE_CODE (operand) == REAL_CST) |
4599 | { | |
4600 | *dt = vect_constant_def; | |
4601 | return true; | |
4602 | } | |
b8698a0f | 4603 | |
ebfd146a IR |
4604 | if (is_gimple_min_invariant (operand)) |
4605 | { | |
4606 | *def = operand; | |
8644a673 | 4607 | *dt = vect_external_def; |
ebfd146a IR |
4608 | return true; |
4609 | } | |
4610 | ||
4611 | if (TREE_CODE (operand) == PAREN_EXPR) | |
4612 | { | |
4613 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4614 | fprintf (vect_dump, "non-associatable copy."); | |
4615 | operand = TREE_OPERAND (operand, 0); | |
4616 | } | |
b8698a0f | 4617 | |
ebfd146a IR |
4618 | if (TREE_CODE (operand) != SSA_NAME) |
4619 | { | |
4620 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4621 | fprintf (vect_dump, "not ssa-name."); | |
4622 | return false; | |
4623 | } | |
b8698a0f | 4624 | |
ebfd146a IR |
4625 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
4626 | if (*def_stmt == NULL) | |
4627 | { | |
4628 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4629 | fprintf (vect_dump, "no def_stmt."); | |
4630 | return false; | |
4631 | } | |
4632 | ||
4633 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4634 | { | |
4635 | fprintf (vect_dump, "def_stmt: "); | |
4636 | print_gimple_stmt (vect_dump, *def_stmt, 0, TDF_SLIM); | |
4637 | } | |
4638 | ||
8644a673 | 4639 | /* Empty stmt is expected only in case of a function argument. |
ebfd146a IR |
4640 | (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */ |
4641 | if (gimple_nop_p (*def_stmt)) | |
4642 | { | |
4643 | *def = operand; | |
8644a673 | 4644 | *dt = vect_external_def; |
ebfd146a IR |
4645 | return true; |
4646 | } | |
4647 | ||
4648 | bb = gimple_bb (*def_stmt); | |
a70d6342 IR |
4649 | |
4650 | if ((loop && !flow_bb_inside_loop_p (loop, bb)) | |
4651 | || (!loop && bb != BB_VINFO_BB (bb_vinfo)) | |
b8698a0f | 4652 | || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI)) |
8644a673 | 4653 | *dt = vect_external_def; |
ebfd146a IR |
4654 | else |
4655 | { | |
4656 | stmt_vinfo = vinfo_for_stmt (*def_stmt); | |
4657 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
4658 | } | |
4659 | ||
4660 | if (*dt == vect_unknown_def_type) | |
4661 | { | |
4662 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4663 | fprintf (vect_dump, "Unsupported pattern."); | |
4664 | return false; | |
4665 | } | |
4666 | ||
4667 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4668 | fprintf (vect_dump, "type of def: %d.",*dt); | |
4669 | ||
4670 | switch (gimple_code (*def_stmt)) | |
4671 | { | |
4672 | case GIMPLE_PHI: | |
4673 | *def = gimple_phi_result (*def_stmt); | |
4674 | break; | |
4675 | ||
4676 | case GIMPLE_ASSIGN: | |
4677 | *def = gimple_assign_lhs (*def_stmt); | |
4678 | break; | |
4679 | ||
4680 | case GIMPLE_CALL: | |
4681 | *def = gimple_call_lhs (*def_stmt); | |
4682 | if (*def != NULL) | |
4683 | break; | |
4684 | /* FALLTHRU */ | |
4685 | default: | |
4686 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4687 | fprintf (vect_dump, "unsupported defining stmt: "); | |
4688 | return false; | |
4689 | } | |
4690 | ||
4691 | return true; | |
4692 | } | |
4693 | ||
b690cc0f RG |
4694 | /* Function vect_is_simple_use_1. |
4695 | ||
4696 | Same as vect_is_simple_use_1 but also determines the vector operand | |
4697 | type of OPERAND and stores it to *VECTYPE. If the definition of | |
4698 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
4699 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
4700 | is responsible to compute the best suited vector type for the | |
4701 | scalar operand. */ | |
4702 | ||
4703 | bool | |
4704 | vect_is_simple_use_1 (tree operand, loop_vec_info loop_vinfo, | |
4705 | bb_vec_info bb_vinfo, gimple *def_stmt, | |
4706 | tree *def, enum vect_def_type *dt, tree *vectype) | |
4707 | { | |
4708 | if (!vect_is_simple_use (operand, loop_vinfo, bb_vinfo, def_stmt, def, dt)) | |
4709 | return false; | |
4710 | ||
4711 | /* Now get a vector type if the def is internal, otherwise supply | |
4712 | NULL_TREE and leave it up to the caller to figure out a proper | |
4713 | type for the use stmt. */ | |
4714 | if (*dt == vect_internal_def | |
4715 | || *dt == vect_induction_def | |
4716 | || *dt == vect_reduction_def | |
4717 | || *dt == vect_double_reduction_def | |
4718 | || *dt == vect_nested_cycle) | |
4719 | { | |
4720 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
4721 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
4722 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
4723 | *vectype = STMT_VINFO_VECTYPE (stmt_info); | |
4724 | gcc_assert (*vectype != NULL_TREE); | |
4725 | } | |
4726 | else if (*dt == vect_uninitialized_def | |
4727 | || *dt == vect_constant_def | |
4728 | || *dt == vect_external_def) | |
4729 | *vectype = NULL_TREE; | |
4730 | else | |
4731 | gcc_unreachable (); | |
4732 | ||
4733 | return true; | |
4734 | } | |
4735 | ||
ebfd146a IR |
4736 | |
4737 | /* Function supportable_widening_operation | |
4738 | ||
b8698a0f L |
4739 | Check whether an operation represented by the code CODE is a |
4740 | widening operation that is supported by the target platform in | |
b690cc0f RG |
4741 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
4742 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 4743 | |
ebfd146a IR |
4744 | Widening operations we currently support are NOP (CONVERT), FLOAT |
4745 | and WIDEN_MULT. This function checks if these operations are supported | |
4746 | by the target platform either directly (via vector tree-codes), or via | |
4747 | target builtins. | |
4748 | ||
4749 | Output: | |
b8698a0f L |
4750 | - CODE1 and CODE2 are codes of vector operations to be used when |
4751 | vectorizing the operation, if available. | |
ebfd146a IR |
4752 | - DECL1 and DECL2 are decls of target builtin functions to be used |
4753 | when vectorizing the operation, if available. In this case, | |
b8698a0f | 4754 | CODE1 and CODE2 are CALL_EXPR. |
ebfd146a IR |
4755 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
4756 | case of multi-step conversion (like char->short->int - in that case | |
4757 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
4758 | - INTERM_TYPES contains the intermediate type required to perform the |
4759 | widening operation (short in the above example). */ | |
ebfd146a IR |
4760 | |
4761 | bool | |
b690cc0f RG |
4762 | supportable_widening_operation (enum tree_code code, gimple stmt, |
4763 | tree vectype_out, tree vectype_in, | |
ebfd146a IR |
4764 | tree *decl1, tree *decl2, |
4765 | enum tree_code *code1, enum tree_code *code2, | |
4766 | int *multi_step_cvt, | |
4767 | VEC (tree, heap) **interm_types) | |
4768 | { | |
4769 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4770 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4771 | struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info); | |
4772 | bool ordered_p; | |
4773 | enum machine_mode vec_mode; | |
81f40b79 | 4774 | enum insn_code icode1, icode2; |
ebfd146a | 4775 | optab optab1, optab2; |
b690cc0f RG |
4776 | tree vectype = vectype_in; |
4777 | tree wide_vectype = vectype_out; | |
ebfd146a IR |
4778 | enum tree_code c1, c2; |
4779 | ||
4780 | /* The result of a vectorized widening operation usually requires two vectors | |
b8698a0f L |
4781 | (because the widened results do not fit int one vector). The generated |
4782 | vector results would normally be expected to be generated in the same | |
ebfd146a IR |
4783 | order as in the original scalar computation, i.e. if 8 results are |
4784 | generated in each vector iteration, they are to be organized as follows: | |
b8698a0f | 4785 | vect1: [res1,res2,res3,res4], vect2: [res5,res6,res7,res8]. |
ebfd146a | 4786 | |
b8698a0f | 4787 | However, in the special case that the result of the widening operation is |
ebfd146a | 4788 | used in a reduction computation only, the order doesn't matter (because |
b8698a0f | 4789 | when vectorizing a reduction we change the order of the computation). |
ebfd146a IR |
4790 | Some targets can take advantage of this and generate more efficient code. |
4791 | For example, targets like Altivec, that support widen_mult using a sequence | |
4792 | of {mult_even,mult_odd} generate the following vectors: | |
4793 | vect1: [res1,res3,res5,res7], vect2: [res2,res4,res6,res8]. | |
4794 | ||
4795 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
b8698a0f L |
4796 | (each vectorized inner-loop iteration contributes to VF outer-loop |
4797 | iterations in parallel). We therefore don't allow to change the order | |
ebfd146a IR |
4798 | of the computation in the inner-loop during outer-loop vectorization. */ |
4799 | ||
4800 | if (STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
4801 | && !nested_in_vect_loop_p (vect_loop, stmt)) | |
4802 | ordered_p = false; | |
4803 | else | |
4804 | ordered_p = true; | |
4805 | ||
4806 | if (!ordered_p | |
4807 | && code == WIDEN_MULT_EXPR | |
4808 | && targetm.vectorize.builtin_mul_widen_even | |
4809 | && targetm.vectorize.builtin_mul_widen_even (vectype) | |
4810 | && targetm.vectorize.builtin_mul_widen_odd | |
4811 | && targetm.vectorize.builtin_mul_widen_odd (vectype)) | |
4812 | { | |
4813 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4814 | fprintf (vect_dump, "Unordered widening operation detected."); | |
4815 | ||
4816 | *code1 = *code2 = CALL_EXPR; | |
4817 | *decl1 = targetm.vectorize.builtin_mul_widen_even (vectype); | |
4818 | *decl2 = targetm.vectorize.builtin_mul_widen_odd (vectype); | |
4819 | return true; | |
4820 | } | |
4821 | ||
4822 | switch (code) | |
4823 | { | |
4824 | case WIDEN_MULT_EXPR: | |
4825 | if (BYTES_BIG_ENDIAN) | |
4826 | { | |
4827 | c1 = VEC_WIDEN_MULT_HI_EXPR; | |
4828 | c2 = VEC_WIDEN_MULT_LO_EXPR; | |
4829 | } | |
4830 | else | |
4831 | { | |
4832 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
4833 | c1 = VEC_WIDEN_MULT_LO_EXPR; | |
4834 | } | |
4835 | break; | |
4836 | ||
4837 | CASE_CONVERT: | |
4838 | if (BYTES_BIG_ENDIAN) | |
4839 | { | |
4840 | c1 = VEC_UNPACK_HI_EXPR; | |
4841 | c2 = VEC_UNPACK_LO_EXPR; | |
4842 | } | |
4843 | else | |
4844 | { | |
4845 | c2 = VEC_UNPACK_HI_EXPR; | |
4846 | c1 = VEC_UNPACK_LO_EXPR; | |
4847 | } | |
4848 | break; | |
4849 | ||
4850 | case FLOAT_EXPR: | |
4851 | if (BYTES_BIG_ENDIAN) | |
4852 | { | |
4853 | c1 = VEC_UNPACK_FLOAT_HI_EXPR; | |
4854 | c2 = VEC_UNPACK_FLOAT_LO_EXPR; | |
4855 | } | |
4856 | else | |
4857 | { | |
4858 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
4859 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; | |
4860 | } | |
4861 | break; | |
4862 | ||
4863 | case FIX_TRUNC_EXPR: | |
4864 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
4865 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
4866 | computing the operation. */ | |
4867 | return false; | |
4868 | ||
4869 | default: | |
4870 | gcc_unreachable (); | |
4871 | } | |
4872 | ||
4873 | if (code == FIX_TRUNC_EXPR) | |
4874 | { | |
4875 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
4876 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
4877 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
4878 | } |
4879 | else | |
4880 | { | |
4881 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
4882 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
4883 | } | |
4884 | ||
4885 | if (!optab1 || !optab2) | |
4886 | return false; | |
4887 | ||
4888 | vec_mode = TYPE_MODE (vectype); | |
4889 | if ((icode1 = optab_handler (optab1, vec_mode)->insn_code) == CODE_FOR_nothing | |
4890 | || (icode2 = optab_handler (optab2, vec_mode)->insn_code) | |
4891 | == CODE_FOR_nothing) | |
4892 | return false; | |
4893 | ||
b8698a0f | 4894 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a IR |
4895 | types. */ |
4896 | if (insn_data[icode1].operand[0].mode != TYPE_MODE (wide_vectype) | |
4897 | || insn_data[icode2].operand[0].mode != TYPE_MODE (wide_vectype)) | |
4898 | { | |
4899 | int i; | |
4900 | tree prev_type = vectype, intermediate_type; | |
4901 | enum machine_mode intermediate_mode, prev_mode = vec_mode; | |
4902 | optab optab3, optab4; | |
4903 | ||
4904 | if (!CONVERT_EXPR_CODE_P (code)) | |
4905 | return false; | |
b8698a0f | 4906 | |
ebfd146a IR |
4907 | *code1 = c1; |
4908 | *code2 = c2; | |
b8698a0f | 4909 | |
ebfd146a IR |
4910 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
4911 | intermediate steps in promotion sequence. We try MAX_INTERM_CVT_STEPS | |
4912 | to get to NARROW_VECTYPE, and fail if we do not. */ | |
4913 | *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS); | |
4914 | for (i = 0; i < 3; i++) | |
4915 | { | |
4916 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
4917 | intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode, | |
4918 | TYPE_UNSIGNED (prev_type)); | |
4919 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
4920 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
4921 | ||
4922 | if (!optab3 || !optab4 | |
4923 | || (icode1 = optab1->handlers[(int) prev_mode].insn_code) | |
4924 | == CODE_FOR_nothing | |
4925 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
4926 | || (icode2 = optab2->handlers[(int) prev_mode].insn_code) | |
4927 | == CODE_FOR_nothing | |
4928 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
b8698a0f | 4929 | || (icode1 = optab3->handlers[(int) intermediate_mode].insn_code) |
ebfd146a IR |
4930 | == CODE_FOR_nothing |
4931 | || (icode2 = optab4->handlers[(int) intermediate_mode].insn_code) | |
4932 | == CODE_FOR_nothing) | |
4933 | return false; | |
4934 | ||
4935 | VEC_quick_push (tree, *interm_types, intermediate_type); | |
4936 | (*multi_step_cvt)++; | |
4937 | ||
4938 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
4939 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
4940 | return true; | |
4941 | ||
4942 | prev_type = intermediate_type; | |
4943 | prev_mode = intermediate_mode; | |
4944 | } | |
4945 | ||
4946 | return false; | |
4947 | } | |
4948 | ||
4949 | *code1 = c1; | |
4950 | *code2 = c2; | |
4951 | return true; | |
4952 | } | |
4953 | ||
4954 | ||
4955 | /* Function supportable_narrowing_operation | |
4956 | ||
b8698a0f L |
4957 | Check whether an operation represented by the code CODE is a |
4958 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
4959 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
4960 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 4961 | |
ebfd146a IR |
4962 | Narrowing operations we currently support are NOP (CONVERT) and |
4963 | FIX_TRUNC. This function checks if these operations are supported by | |
4964 | the target platform directly via vector tree-codes. | |
4965 | ||
4966 | Output: | |
b8698a0f L |
4967 | - CODE1 is the code of a vector operation to be used when |
4968 | vectorizing the operation, if available. | |
ebfd146a IR |
4969 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
4970 | case of multi-step conversion (like int->short->char - in that case | |
4971 | MULTI_STEP_CVT will be 1). | |
4972 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 4973 | narrowing operation (short in the above example). */ |
ebfd146a IR |
4974 | |
4975 | bool | |
4976 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 4977 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
4978 | enum tree_code *code1, int *multi_step_cvt, |
4979 | VEC (tree, heap) **interm_types) | |
4980 | { | |
4981 | enum machine_mode vec_mode; | |
4982 | enum insn_code icode1; | |
4983 | optab optab1, interm_optab; | |
b690cc0f RG |
4984 | tree vectype = vectype_in; |
4985 | tree narrow_vectype = vectype_out; | |
ebfd146a IR |
4986 | enum tree_code c1; |
4987 | tree intermediate_type, prev_type; | |
4988 | int i; | |
4989 | ||
4990 | switch (code) | |
4991 | { | |
4992 | CASE_CONVERT: | |
4993 | c1 = VEC_PACK_TRUNC_EXPR; | |
4994 | break; | |
4995 | ||
4996 | case FIX_TRUNC_EXPR: | |
4997 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
4998 | break; | |
4999 | ||
5000 | case FLOAT_EXPR: | |
5001 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
5002 | tree code and optabs used for computing the operation. */ | |
5003 | return false; | |
5004 | ||
5005 | default: | |
5006 | gcc_unreachable (); | |
5007 | } | |
5008 | ||
5009 | if (code == FIX_TRUNC_EXPR) | |
5010 | /* The signedness is determined from output operand. */ | |
b690cc0f | 5011 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
5012 | else |
5013 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
5014 | ||
5015 | if (!optab1) | |
5016 | return false; | |
5017 | ||
5018 | vec_mode = TYPE_MODE (vectype); | |
b8698a0f | 5019 | if ((icode1 = optab_handler (optab1, vec_mode)->insn_code) |
ebfd146a IR |
5020 | == CODE_FOR_nothing) |
5021 | return false; | |
5022 | ||
5023 | /* Check if it's a multi-step conversion that can be done using intermediate | |
5024 | types. */ | |
5025 | if (insn_data[icode1].operand[0].mode != TYPE_MODE (narrow_vectype)) | |
5026 | { | |
5027 | enum machine_mode intermediate_mode, prev_mode = vec_mode; | |
5028 | ||
5029 | *code1 = c1; | |
5030 | prev_type = vectype; | |
5031 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS | |
5032 | intermediate steps in promotion sequence. We try MAX_INTERM_CVT_STEPS | |
5033 | to get to NARROW_VECTYPE, and fail if we do not. */ | |
5034 | *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS); | |
5035 | for (i = 0; i < 3; i++) | |
5036 | { | |
5037 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
5038 | intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode, | |
5039 | TYPE_UNSIGNED (prev_type)); | |
b8698a0f | 5040 | interm_optab = optab_for_tree_code (c1, intermediate_type, |
ebfd146a | 5041 | optab_default); |
b8698a0f | 5042 | if (!interm_optab |
ebfd146a IR |
5043 | || (icode1 = optab1->handlers[(int) prev_mode].insn_code) |
5044 | == CODE_FOR_nothing | |
5045 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
b8698a0f | 5046 | || (icode1 |
ebfd146a IR |
5047 | = interm_optab->handlers[(int) intermediate_mode].insn_code) |
5048 | == CODE_FOR_nothing) | |
5049 | return false; | |
5050 | ||
5051 | VEC_quick_push (tree, *interm_types, intermediate_type); | |
5052 | (*multi_step_cvt)++; | |
5053 | ||
5054 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5055 | return true; | |
5056 | ||
5057 | prev_type = intermediate_type; | |
5058 | prev_mode = intermediate_mode; | |
5059 | } | |
5060 | ||
5061 | return false; | |
5062 | } | |
5063 | ||
5064 | *code1 = c1; | |
5065 | return true; | |
5066 | } |