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