]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
62f7fd21 MM |
2 | Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010 |
3 | Free Software Foundation, Inc. | |
b8698a0f | 4 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
5 | and Ira Rosen <irar@il.ibm.com> |
6 | ||
7 | This file is part of GCC. | |
8 | ||
9 | GCC is free software; you can redistribute it and/or modify it under | |
10 | the terms of the GNU General Public License as published by the Free | |
11 | Software Foundation; either version 3, or (at your option) any later | |
12 | version. | |
13 | ||
14 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
15 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
16 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
17 | for more details. | |
18 | ||
19 | You should have received a copy of the GNU General Public License | |
20 | along with GCC; see the file COPYING3. If not see | |
21 | <http://www.gnu.org/licenses/>. */ | |
22 | ||
23 | #include "config.h" | |
24 | #include "system.h" | |
25 | #include "coretypes.h" | |
26 | #include "tm.h" | |
27 | #include "ggc.h" | |
28 | #include "tree.h" | |
29 | #include "target.h" | |
30 | #include "basic-block.h" | |
cf835838 JM |
31 | #include "tree-pretty-print.h" |
32 | #include "gimple-pretty-print.h" | |
ebfd146a IR |
33 | #include "tree-flow.h" |
34 | #include "tree-dump.h" | |
35 | #include "cfgloop.h" | |
36 | #include "cfglayout.h" | |
37 | #include "expr.h" | |
38 | #include "recog.h" | |
39 | #include "optabs.h" | |
718f9c0f | 40 | #include "diagnostic-core.h" |
ebfd146a IR |
41 | #include "tree-vectorizer.h" |
42 | #include "langhooks.h" | |
43 | ||
44 | ||
45 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ | |
46 | ||
47 | /* Function vect_mark_relevant. | |
48 | ||
49 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
50 | ||
51 | static void | |
52 | vect_mark_relevant (VEC(gimple,heap) **worklist, gimple stmt, | |
53 | enum vect_relevant relevant, bool live_p) | |
54 | { | |
55 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
56 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
57 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
58 | ||
59 | if (vect_print_dump_info (REPORT_DETAILS)) | |
60 | fprintf (vect_dump, "mark relevant %d, live %d.", relevant, live_p); | |
61 | ||
62 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
63 | { | |
64 | gimple pattern_stmt; | |
65 | ||
b8698a0f | 66 | /* This is the last stmt in a sequence that was detected as a |
ebfd146a IR |
67 | pattern that can potentially be vectorized. Don't mark the stmt |
68 | as relevant/live because it's not going to be vectorized. | |
69 | Instead mark the pattern-stmt that replaces it. */ | |
70 | ||
71 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
72 | ||
73 | if (vect_print_dump_info (REPORT_DETAILS)) | |
74 | fprintf (vect_dump, "last stmt in pattern. don't mark relevant/live."); | |
75 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
76 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
77 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
78 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
79 | stmt = pattern_stmt; | |
80 | } | |
81 | ||
82 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
83 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
84 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
85 | ||
86 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
87 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
88 | { | |
89 | if (vect_print_dump_info (REPORT_DETAILS)) | |
90 | fprintf (vect_dump, "already marked relevant/live."); | |
91 | return; | |
92 | } | |
93 | ||
94 | VEC_safe_push (gimple, heap, *worklist, stmt); | |
95 | } | |
96 | ||
97 | ||
98 | /* Function vect_stmt_relevant_p. | |
99 | ||
100 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
101 | "relevant for vectorization". | |
102 | ||
103 | A stmt is considered "relevant for vectorization" if: | |
104 | - it has uses outside the loop. | |
105 | - it has vdefs (it alters memory). | |
106 | - control stmts in the loop (except for the exit condition). | |
107 | ||
108 | CHECKME: what other side effects would the vectorizer allow? */ | |
109 | ||
110 | static bool | |
111 | vect_stmt_relevant_p (gimple stmt, loop_vec_info loop_vinfo, | |
112 | enum vect_relevant *relevant, bool *live_p) | |
113 | { | |
114 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
115 | ssa_op_iter op_iter; | |
116 | imm_use_iterator imm_iter; | |
117 | use_operand_p use_p; | |
118 | def_operand_p def_p; | |
119 | ||
8644a673 | 120 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
121 | *live_p = false; |
122 | ||
123 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
124 | if (is_ctrl_stmt (stmt) |
125 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
126 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 127 | *relevant = vect_used_in_scope; |
ebfd146a IR |
128 | |
129 | /* changing memory. */ | |
130 | if (gimple_code (stmt) != GIMPLE_PHI) | |
5006671f | 131 | if (gimple_vdef (stmt)) |
ebfd146a IR |
132 | { |
133 | if (vect_print_dump_info (REPORT_DETAILS)) | |
134 | fprintf (vect_dump, "vec_stmt_relevant_p: stmt has vdefs."); | |
8644a673 | 135 | *relevant = vect_used_in_scope; |
ebfd146a IR |
136 | } |
137 | ||
138 | /* uses outside the loop. */ | |
139 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
140 | { | |
141 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
142 | { | |
143 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
144 | if (!flow_bb_inside_loop_p (loop, bb)) | |
145 | { | |
146 | if (vect_print_dump_info (REPORT_DETAILS)) | |
147 | fprintf (vect_dump, "vec_stmt_relevant_p: used out of loop."); | |
148 | ||
3157b0c2 AO |
149 | if (is_gimple_debug (USE_STMT (use_p))) |
150 | continue; | |
151 | ||
ebfd146a IR |
152 | /* We expect all such uses to be in the loop exit phis |
153 | (because of loop closed form) */ | |
154 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
155 | gcc_assert (bb == single_exit (loop)->dest); | |
156 | ||
157 | *live_p = true; | |
158 | } | |
159 | } | |
160 | } | |
161 | ||
162 | return (*live_p || *relevant); | |
163 | } | |
164 | ||
165 | ||
b8698a0f | 166 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 167 | |
ff802fa1 | 168 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
169 | used in STMT for anything other than indexing an array. */ |
170 | ||
171 | static bool | |
172 | exist_non_indexing_operands_for_use_p (tree use, gimple stmt) | |
173 | { | |
174 | tree operand; | |
175 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 176 | |
ff802fa1 | 177 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
178 | reference in STMT, then any operand that corresponds to USE |
179 | is not indexing an array. */ | |
180 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
181 | return true; | |
59a05b0c | 182 | |
ebfd146a IR |
183 | /* STMT has a data_ref. FORNOW this means that its of one of |
184 | the following forms: | |
185 | -1- ARRAY_REF = var | |
186 | -2- var = ARRAY_REF | |
187 | (This should have been verified in analyze_data_refs). | |
188 | ||
189 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 190 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
191 | for array indexing. |
192 | ||
193 | Therefore, all we need to check is if STMT falls into the | |
194 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
195 | |
196 | if (!gimple_assign_copy_p (stmt)) | |
197 | return false; | |
59a05b0c EB |
198 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
199 | return false; | |
ebfd146a | 200 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
201 | if (TREE_CODE (operand) != SSA_NAME) |
202 | return false; | |
203 | ||
204 | if (operand == use) | |
205 | return true; | |
206 | ||
207 | return false; | |
208 | } | |
209 | ||
210 | ||
b8698a0f | 211 | /* |
ebfd146a IR |
212 | Function process_use. |
213 | ||
214 | Inputs: | |
215 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 216 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ff802fa1 | 217 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a IR |
218 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
219 | ||
220 | Outputs: | |
221 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
222 | relevance info of the DEF_STMT of this USE: | |
223 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
224 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
225 | Exceptions: | |
226 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 227 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 228 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
229 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
230 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
231 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
232 | be modified accordingly. | |
233 | ||
234 | Return true if everything is as expected. Return false otherwise. */ | |
235 | ||
236 | static bool | |
b8698a0f | 237 | process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
ebfd146a IR |
238 | enum vect_relevant relevant, VEC(gimple,heap) **worklist) |
239 | { | |
240 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
241 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
242 | stmt_vec_info dstmt_vinfo; | |
243 | basic_block bb, def_bb; | |
244 | tree def; | |
245 | gimple def_stmt; | |
246 | enum vect_def_type dt; | |
247 | ||
b8698a0f | 248 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a IR |
249 | that are used for address computation are not considered relevant. */ |
250 | if (!exist_non_indexing_operands_for_use_p (use, stmt)) | |
251 | return true; | |
252 | ||
a70d6342 | 253 | if (!vect_is_simple_use (use, loop_vinfo, NULL, &def_stmt, &def, &dt)) |
b8698a0f | 254 | { |
8644a673 | 255 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
ebfd146a IR |
256 | fprintf (vect_dump, "not vectorized: unsupported use in stmt."); |
257 | return false; | |
258 | } | |
259 | ||
260 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
261 | return true; | |
262 | ||
263 | def_bb = gimple_bb (def_stmt); | |
264 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
265 | { | |
266 | if (vect_print_dump_info (REPORT_DETAILS)) | |
267 | fprintf (vect_dump, "def_stmt is out of loop."); | |
268 | return true; | |
269 | } | |
270 | ||
b8698a0f L |
271 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
272 | DEF_STMT must have already been processed, because this should be the | |
273 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
274 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
275 | check that everything is as expected, and we are done. */ |
276 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
277 | bb = gimple_bb (stmt); | |
278 | if (gimple_code (stmt) == GIMPLE_PHI | |
279 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
280 | && gimple_code (def_stmt) != GIMPLE_PHI | |
281 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
282 | && bb->loop_father == def_bb->loop_father) | |
283 | { | |
284 | if (vect_print_dump_info (REPORT_DETAILS)) | |
285 | fprintf (vect_dump, "reduc-stmt defining reduc-phi in the same nest."); | |
286 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) | |
287 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
288 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 289 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 290 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
291 | return true; |
292 | } | |
293 | ||
294 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
295 | outer-loop-header-bb: | |
296 | d = def_stmt | |
297 | inner-loop: | |
298 | stmt # use (d) | |
299 | outer-loop-tail-bb: | |
300 | ... */ | |
301 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
302 | { | |
303 | if (vect_print_dump_info (REPORT_DETAILS)) | |
304 | fprintf (vect_dump, "outer-loop def-stmt defining inner-loop stmt."); | |
7c5222ff | 305 | |
ebfd146a IR |
306 | switch (relevant) |
307 | { | |
8644a673 | 308 | case vect_unused_in_scope: |
7c5222ff IR |
309 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
310 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 311 | break; |
7c5222ff | 312 | |
ebfd146a | 313 | case vect_used_in_outer_by_reduction: |
7c5222ff | 314 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
315 | relevant = vect_used_by_reduction; |
316 | break; | |
7c5222ff | 317 | |
ebfd146a | 318 | case vect_used_in_outer: |
7c5222ff | 319 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 320 | relevant = vect_used_in_scope; |
ebfd146a | 321 | break; |
7c5222ff | 322 | |
8644a673 | 323 | case vect_used_in_scope: |
ebfd146a IR |
324 | break; |
325 | ||
326 | default: | |
327 | gcc_unreachable (); | |
b8698a0f | 328 | } |
ebfd146a IR |
329 | } |
330 | ||
331 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
332 | outer-loop-header-bb: | |
333 | ... | |
334 | inner-loop: | |
335 | d = def_stmt | |
06066f92 | 336 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
337 | stmt # use (d) */ |
338 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
339 | { | |
340 | if (vect_print_dump_info (REPORT_DETAILS)) | |
341 | fprintf (vect_dump, "inner-loop def-stmt defining outer-loop stmt."); | |
7c5222ff | 342 | |
ebfd146a IR |
343 | switch (relevant) |
344 | { | |
8644a673 | 345 | case vect_unused_in_scope: |
b8698a0f | 346 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 347 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 348 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
349 | break; |
350 | ||
ebfd146a IR |
351 | case vect_used_by_reduction: |
352 | relevant = vect_used_in_outer_by_reduction; | |
353 | break; | |
354 | ||
8644a673 | 355 | case vect_used_in_scope: |
ebfd146a IR |
356 | relevant = vect_used_in_outer; |
357 | break; | |
358 | ||
359 | default: | |
360 | gcc_unreachable (); | |
361 | } | |
362 | } | |
363 | ||
364 | vect_mark_relevant (worklist, def_stmt, relevant, live_p); | |
365 | return true; | |
366 | } | |
367 | ||
368 | ||
369 | /* Function vect_mark_stmts_to_be_vectorized. | |
370 | ||
371 | Not all stmts in the loop need to be vectorized. For example: | |
372 | ||
373 | for i... | |
374 | for j... | |
375 | 1. T0 = i + j | |
376 | 2. T1 = a[T0] | |
377 | ||
378 | 3. j = j + 1 | |
379 | ||
380 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
381 | addressing of vectorized data-refs are handled differently. | |
382 | ||
383 | This pass detects such stmts. */ | |
384 | ||
385 | bool | |
386 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
387 | { | |
388 | VEC(gimple,heap) *worklist; | |
389 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
390 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
391 | unsigned int nbbs = loop->num_nodes; | |
392 | gimple_stmt_iterator si; | |
393 | gimple stmt; | |
394 | unsigned int i; | |
395 | stmt_vec_info stmt_vinfo; | |
396 | basic_block bb; | |
397 | gimple phi; | |
398 | bool live_p; | |
06066f92 IR |
399 | enum vect_relevant relevant, tmp_relevant; |
400 | enum vect_def_type def_type; | |
ebfd146a IR |
401 | |
402 | if (vect_print_dump_info (REPORT_DETAILS)) | |
403 | fprintf (vect_dump, "=== vect_mark_stmts_to_be_vectorized ==="); | |
404 | ||
405 | worklist = VEC_alloc (gimple, heap, 64); | |
406 | ||
407 | /* 1. Init worklist. */ | |
408 | for (i = 0; i < nbbs; i++) | |
409 | { | |
410 | bb = bbs[i]; | |
411 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 412 | { |
ebfd146a IR |
413 | phi = gsi_stmt (si); |
414 | if (vect_print_dump_info (REPORT_DETAILS)) | |
415 | { | |
416 | fprintf (vect_dump, "init: phi relevant? "); | |
417 | print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM); | |
418 | } | |
419 | ||
420 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
421 | vect_mark_relevant (&worklist, phi, relevant, live_p); | |
422 | } | |
423 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
424 | { | |
425 | stmt = gsi_stmt (si); | |
426 | if (vect_print_dump_info (REPORT_DETAILS)) | |
427 | { | |
428 | fprintf (vect_dump, "init: stmt relevant? "); | |
429 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
b8698a0f | 430 | } |
ebfd146a IR |
431 | |
432 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
433 | vect_mark_relevant (&worklist, stmt, relevant, live_p); | |
434 | } | |
435 | } | |
436 | ||
437 | /* 2. Process_worklist */ | |
438 | while (VEC_length (gimple, worklist) > 0) | |
439 | { | |
440 | use_operand_p use_p; | |
441 | ssa_op_iter iter; | |
442 | ||
443 | stmt = VEC_pop (gimple, worklist); | |
444 | if (vect_print_dump_info (REPORT_DETAILS)) | |
445 | { | |
446 | fprintf (vect_dump, "worklist: examine stmt: "); | |
447 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
448 | } | |
449 | ||
b8698a0f L |
450 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
451 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
452 | liveness and relevance properties of STMT. */ |
453 | stmt_vinfo = vinfo_for_stmt (stmt); | |
454 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
455 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
456 | ||
457 | /* Generally, the liveness and relevance properties of STMT are | |
458 | propagated as is to the DEF_STMTs of its USEs: | |
459 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
460 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
461 | ||
462 | One exception is when STMT has been identified as defining a reduction | |
463 | variable; in this case we set the liveness/relevance as follows: | |
464 | live_p = false | |
465 | relevant = vect_used_by_reduction | |
466 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f | 467 | those that are used by a reduction computation, and those that are |
ff802fa1 | 468 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 469 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 470 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 471 | |
06066f92 IR |
472 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
473 | tmp_relevant = relevant; | |
474 | switch (def_type) | |
ebfd146a | 475 | { |
06066f92 IR |
476 | case vect_reduction_def: |
477 | switch (tmp_relevant) | |
478 | { | |
479 | case vect_unused_in_scope: | |
480 | relevant = vect_used_by_reduction; | |
481 | break; | |
482 | ||
483 | case vect_used_by_reduction: | |
484 | if (gimple_code (stmt) == GIMPLE_PHI) | |
485 | break; | |
486 | /* fall through */ | |
487 | ||
488 | default: | |
489 | if (vect_print_dump_info (REPORT_DETAILS)) | |
490 | fprintf (vect_dump, "unsupported use of reduction."); | |
491 | ||
492 | VEC_free (gimple, heap, worklist); | |
493 | return false; | |
494 | } | |
495 | ||
b8698a0f | 496 | live_p = false; |
06066f92 | 497 | break; |
b8698a0f | 498 | |
06066f92 IR |
499 | case vect_nested_cycle: |
500 | if (tmp_relevant != vect_unused_in_scope | |
501 | && tmp_relevant != vect_used_in_outer_by_reduction | |
502 | && tmp_relevant != vect_used_in_outer) | |
503 | { | |
504 | if (vect_print_dump_info (REPORT_DETAILS)) | |
505 | fprintf (vect_dump, "unsupported use of nested cycle."); | |
7c5222ff | 506 | |
06066f92 IR |
507 | VEC_free (gimple, heap, worklist); |
508 | return false; | |
509 | } | |
7c5222ff | 510 | |
b8698a0f L |
511 | live_p = false; |
512 | break; | |
513 | ||
06066f92 IR |
514 | case vect_double_reduction_def: |
515 | if (tmp_relevant != vect_unused_in_scope | |
516 | && tmp_relevant != vect_used_by_reduction) | |
517 | { | |
7c5222ff | 518 | if (vect_print_dump_info (REPORT_DETAILS)) |
06066f92 | 519 | fprintf (vect_dump, "unsupported use of double reduction."); |
7c5222ff IR |
520 | |
521 | VEC_free (gimple, heap, worklist); | |
522 | return false; | |
06066f92 IR |
523 | } |
524 | ||
525 | live_p = false; | |
b8698a0f | 526 | break; |
7c5222ff | 527 | |
06066f92 IR |
528 | default: |
529 | break; | |
7c5222ff | 530 | } |
b8698a0f | 531 | |
ebfd146a IR |
532 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) |
533 | { | |
534 | tree op = USE_FROM_PTR (use_p); | |
535 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, &worklist)) | |
536 | { | |
537 | VEC_free (gimple, heap, worklist); | |
538 | return false; | |
539 | } | |
540 | } | |
541 | } /* while worklist */ | |
542 | ||
543 | VEC_free (gimple, heap, worklist); | |
544 | return true; | |
545 | } | |
546 | ||
547 | ||
720f5239 IR |
548 | /* Get cost by calling cost target builtin. */ |
549 | ||
550 | static inline | |
551 | int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost) | |
552 | { | |
553 | tree dummy_type = NULL; | |
554 | int dummy = 0; | |
555 | ||
556 | return targetm.vectorize.builtin_vectorization_cost (type_of_cost, | |
557 | dummy_type, dummy); | |
558 | } | |
559 | ||
ff802fa1 IR |
560 | |
561 | /* Get cost for STMT. */ | |
562 | ||
ebfd146a IR |
563 | int |
564 | cost_for_stmt (gimple stmt) | |
565 | { | |
566 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
567 | ||
568 | switch (STMT_VINFO_TYPE (stmt_info)) | |
569 | { | |
570 | case load_vec_info_type: | |
720f5239 | 571 | return vect_get_stmt_cost (scalar_load); |
ebfd146a | 572 | case store_vec_info_type: |
720f5239 | 573 | return vect_get_stmt_cost (scalar_store); |
ebfd146a IR |
574 | case op_vec_info_type: |
575 | case condition_vec_info_type: | |
576 | case assignment_vec_info_type: | |
577 | case reduc_vec_info_type: | |
578 | case induc_vec_info_type: | |
579 | case type_promotion_vec_info_type: | |
580 | case type_demotion_vec_info_type: | |
581 | case type_conversion_vec_info_type: | |
582 | case call_vec_info_type: | |
720f5239 | 583 | return vect_get_stmt_cost (scalar_stmt); |
ebfd146a IR |
584 | case undef_vec_info_type: |
585 | default: | |
586 | gcc_unreachable (); | |
587 | } | |
588 | } | |
589 | ||
b8698a0f | 590 | /* Function vect_model_simple_cost. |
ebfd146a | 591 | |
b8698a0f | 592 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
593 | single op. Right now, this does not account for multiple insns that could |
594 | be generated for the single vector op. We will handle that shortly. */ | |
595 | ||
596 | void | |
b8698a0f | 597 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
ebfd146a IR |
598 | enum vect_def_type *dt, slp_tree slp_node) |
599 | { | |
600 | int i; | |
601 | int inside_cost = 0, outside_cost = 0; | |
602 | ||
603 | /* The SLP costs were already calculated during SLP tree build. */ | |
604 | if (PURE_SLP_STMT (stmt_info)) | |
605 | return; | |
606 | ||
720f5239 | 607 | inside_cost = ncopies * vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
608 | |
609 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
610 | for (i = 0; i < 2; i++) | |
611 | { | |
8644a673 | 612 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
720f5239 | 613 | outside_cost += vect_get_stmt_cost (vector_stmt); |
ebfd146a | 614 | } |
b8698a0f | 615 | |
ebfd146a IR |
616 | if (vect_print_dump_info (REPORT_COST)) |
617 | fprintf (vect_dump, "vect_model_simple_cost: inside_cost = %d, " | |
618 | "outside_cost = %d .", inside_cost, outside_cost); | |
619 | ||
620 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
621 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
622 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
623 | } | |
624 | ||
625 | ||
b8698a0f L |
626 | /* Function vect_cost_strided_group_size |
627 | ||
ebfd146a IR |
628 | For strided load or store, return the group_size only if it is the first |
629 | load or store of a group, else return 1. This ensures that group size is | |
630 | only returned once per group. */ | |
631 | ||
632 | static int | |
633 | vect_cost_strided_group_size (stmt_vec_info stmt_info) | |
634 | { | |
635 | gimple first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
636 | ||
637 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
638 | return DR_GROUP_SIZE (stmt_info); | |
639 | ||
640 | return 1; | |
641 | } | |
642 | ||
643 | ||
644 | /* Function vect_model_store_cost | |
645 | ||
646 | Models cost for stores. In the case of strided accesses, one access | |
647 | has the overhead of the strided access attributed to it. */ | |
648 | ||
649 | void | |
b8698a0f | 650 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
ebfd146a IR |
651 | enum vect_def_type dt, slp_tree slp_node) |
652 | { | |
653 | int group_size; | |
720f5239 IR |
654 | unsigned int inside_cost = 0, outside_cost = 0; |
655 | struct data_reference *first_dr; | |
656 | gimple first_stmt; | |
ebfd146a IR |
657 | |
658 | /* The SLP costs were already calculated during SLP tree build. */ | |
659 | if (PURE_SLP_STMT (stmt_info)) | |
660 | return; | |
661 | ||
8644a673 | 662 | if (dt == vect_constant_def || dt == vect_external_def) |
720f5239 | 663 | outside_cost = vect_get_stmt_cost (scalar_to_vec); |
ebfd146a IR |
664 | |
665 | /* Strided access? */ | |
720f5239 IR |
666 | if (DR_GROUP_FIRST_DR (stmt_info)) |
667 | { | |
668 | if (slp_node) | |
669 | { | |
670 | first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0); | |
671 | group_size = 1; | |
672 | } | |
673 | else | |
674 | { | |
675 | first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
676 | group_size = vect_cost_strided_group_size (stmt_info); | |
677 | } | |
678 | ||
679 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
680 | } | |
ebfd146a IR |
681 | /* Not a strided access. */ |
682 | else | |
720f5239 IR |
683 | { |
684 | group_size = 1; | |
685 | first_dr = STMT_VINFO_DATA_REF (stmt_info); | |
686 | } | |
ebfd146a | 687 | |
b8698a0f | 688 | /* Is this an access in a group of stores, which provide strided access? |
ebfd146a | 689 | If so, add in the cost of the permutes. */ |
b8698a0f | 690 | if (group_size > 1) |
ebfd146a IR |
691 | { |
692 | /* Uses a high and low interleave operation for each needed permute. */ | |
b8698a0f | 693 | inside_cost = ncopies * exact_log2(group_size) * group_size |
720f5239 | 694 | * vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
695 | |
696 | if (vect_print_dump_info (REPORT_COST)) | |
697 | fprintf (vect_dump, "vect_model_store_cost: strided group_size = %d .", | |
698 | group_size); | |
699 | ||
700 | } | |
701 | ||
702 | /* Costs of the stores. */ | |
720f5239 | 703 | vect_get_store_cost (first_dr, ncopies, &inside_cost); |
ebfd146a IR |
704 | |
705 | if (vect_print_dump_info (REPORT_COST)) | |
706 | fprintf (vect_dump, "vect_model_store_cost: inside_cost = %d, " | |
707 | "outside_cost = %d .", inside_cost, outside_cost); | |
708 | ||
709 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
710 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
711 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
712 | } | |
713 | ||
714 | ||
720f5239 IR |
715 | /* Calculate cost of DR's memory access. */ |
716 | void | |
717 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
718 | unsigned int *inside_cost) | |
719 | { | |
720 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
721 | ||
722 | switch (alignment_support_scheme) | |
723 | { | |
724 | case dr_aligned: | |
725 | { | |
726 | *inside_cost += ncopies * vect_get_stmt_cost (vector_store); | |
727 | ||
728 | if (vect_print_dump_info (REPORT_COST)) | |
729 | fprintf (vect_dump, "vect_model_store_cost: aligned."); | |
730 | ||
731 | break; | |
732 | } | |
733 | ||
734 | case dr_unaligned_supported: | |
735 | { | |
736 | gimple stmt = DR_STMT (dr); | |
737 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
738 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
739 | ||
740 | /* Here, we assign an additional cost for the unaligned store. */ | |
741 | *inside_cost += ncopies | |
742 | * targetm.vectorize.builtin_vectorization_cost (unaligned_store, | |
743 | vectype, DR_MISALIGNMENT (dr)); | |
744 | ||
745 | if (vect_print_dump_info (REPORT_COST)) | |
746 | fprintf (vect_dump, "vect_model_store_cost: unaligned supported by " | |
747 | "hardware."); | |
748 | ||
749 | break; | |
750 | } | |
751 | ||
752 | default: | |
753 | gcc_unreachable (); | |
754 | } | |
755 | } | |
756 | ||
757 | ||
ebfd146a IR |
758 | /* Function vect_model_load_cost |
759 | ||
760 | Models cost for loads. In the case of strided accesses, the last access | |
761 | has the overhead of the strided access attributed to it. Since unaligned | |
b8698a0f | 762 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
763 | access scheme chosen. */ |
764 | ||
765 | void | |
766 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, slp_tree slp_node) | |
b8698a0f | 767 | |
ebfd146a IR |
768 | { |
769 | int group_size; | |
ebfd146a IR |
770 | gimple first_stmt; |
771 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; | |
720f5239 | 772 | unsigned int inside_cost = 0, outside_cost = 0; |
ebfd146a IR |
773 | |
774 | /* The SLP costs were already calculated during SLP tree build. */ | |
775 | if (PURE_SLP_STMT (stmt_info)) | |
776 | return; | |
777 | ||
778 | /* Strided accesses? */ | |
779 | first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
780 | if (first_stmt && !slp_node) | |
781 | { | |
782 | group_size = vect_cost_strided_group_size (stmt_info); | |
783 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
784 | } | |
785 | /* Not a strided access. */ | |
786 | else | |
787 | { | |
788 | group_size = 1; | |
789 | first_dr = dr; | |
790 | } | |
791 | ||
b8698a0f | 792 | /* Is this an access in a group of loads providing strided access? |
ebfd146a | 793 | If so, add in the cost of the permutes. */ |
b8698a0f | 794 | if (group_size > 1) |
ebfd146a IR |
795 | { |
796 | /* Uses an even and odd extract operations for each needed permute. */ | |
797 | inside_cost = ncopies * exact_log2(group_size) * group_size | |
720f5239 | 798 | * vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
799 | |
800 | if (vect_print_dump_info (REPORT_COST)) | |
801 | fprintf (vect_dump, "vect_model_load_cost: strided group_size = %d .", | |
802 | group_size); | |
ebfd146a IR |
803 | } |
804 | ||
805 | /* The loads themselves. */ | |
720f5239 IR |
806 | vect_get_load_cost (first_dr, ncopies, |
807 | ((!DR_GROUP_FIRST_DR (stmt_info)) || group_size > 1 || slp_node), | |
808 | &inside_cost, &outside_cost); | |
809 | ||
810 | if (vect_print_dump_info (REPORT_COST)) | |
811 | fprintf (vect_dump, "vect_model_load_cost: inside_cost = %d, " | |
812 | "outside_cost = %d .", inside_cost, outside_cost); | |
813 | ||
814 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
815 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
816 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
817 | } | |
818 | ||
819 | ||
820 | /* Calculate cost of DR's memory access. */ | |
821 | void | |
822 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
823 | bool add_realign_cost, unsigned int *inside_cost, | |
824 | unsigned int *outside_cost) | |
825 | { | |
826 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
827 | ||
828 | switch (alignment_support_scheme) | |
ebfd146a IR |
829 | { |
830 | case dr_aligned: | |
831 | { | |
9940b13c | 832 | *inside_cost += ncopies * vect_get_stmt_cost (vector_load); |
ebfd146a IR |
833 | |
834 | if (vect_print_dump_info (REPORT_COST)) | |
835 | fprintf (vect_dump, "vect_model_load_cost: aligned."); | |
836 | ||
837 | break; | |
838 | } | |
839 | case dr_unaligned_supported: | |
840 | { | |
720f5239 IR |
841 | gimple stmt = DR_STMT (dr); |
842 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
843 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
ebfd146a | 844 | |
720f5239 IR |
845 | /* Here, we assign an additional cost for the unaligned load. */ |
846 | *inside_cost += ncopies | |
847 | * targetm.vectorize.builtin_vectorization_cost (unaligned_load, | |
848 | vectype, DR_MISALIGNMENT (dr)); | |
ebfd146a IR |
849 | if (vect_print_dump_info (REPORT_COST)) |
850 | fprintf (vect_dump, "vect_model_load_cost: unaligned supported by " | |
851 | "hardware."); | |
852 | ||
853 | break; | |
854 | } | |
855 | case dr_explicit_realign: | |
856 | { | |
720f5239 IR |
857 | *inside_cost += ncopies * (2 * vect_get_stmt_cost (vector_load) |
858 | + vect_get_stmt_cost (vector_stmt)); | |
ebfd146a IR |
859 | |
860 | /* FIXME: If the misalignment remains fixed across the iterations of | |
861 | the containing loop, the following cost should be added to the | |
862 | outside costs. */ | |
863 | if (targetm.vectorize.builtin_mask_for_load) | |
720f5239 | 864 | *inside_cost += vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
865 | |
866 | break; | |
867 | } | |
868 | case dr_explicit_realign_optimized: | |
869 | { | |
870 | if (vect_print_dump_info (REPORT_COST)) | |
871 | fprintf (vect_dump, "vect_model_load_cost: unaligned software " | |
872 | "pipelined."); | |
873 | ||
874 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 875 | load, and possibly a mask operation to "prime" the loop. However, |
ebfd146a IR |
876 | if this is an access in a group of loads, which provide strided |
877 | access, then the above cost should only be considered for one | |
ff802fa1 | 878 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
879 | and a realignment op. */ |
880 | ||
720f5239 | 881 | if (add_realign_cost) |
ebfd146a | 882 | { |
720f5239 | 883 | *outside_cost = 2 * vect_get_stmt_cost (vector_stmt); |
ebfd146a | 884 | if (targetm.vectorize.builtin_mask_for_load) |
720f5239 | 885 | *outside_cost += vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
886 | } |
887 | ||
720f5239 IR |
888 | *inside_cost += ncopies * (vect_get_stmt_cost (vector_load) |
889 | + vect_get_stmt_cost (vector_stmt)); | |
ebfd146a IR |
890 | break; |
891 | } | |
892 | ||
893 | default: | |
894 | gcc_unreachable (); | |
895 | } | |
ebfd146a IR |
896 | } |
897 | ||
898 | ||
899 | /* Function vect_init_vector. | |
900 | ||
901 | Insert a new stmt (INIT_STMT) that initializes a new vector variable with | |
ff802fa1 IR |
902 | the vector elements of VECTOR_VAR. Place the initialization at BSI if it |
903 | is not NULL. Otherwise, place the initialization at the loop preheader. | |
b8698a0f | 904 | Return the DEF of INIT_STMT. |
ebfd146a IR |
905 | It will be used in the vectorization of STMT. */ |
906 | ||
907 | tree | |
908 | vect_init_vector (gimple stmt, tree vector_var, tree vector_type, | |
909 | gimple_stmt_iterator *gsi) | |
910 | { | |
911 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
912 | tree new_var; | |
913 | gimple init_stmt; | |
914 | tree vec_oprnd; | |
915 | edge pe; | |
916 | tree new_temp; | |
917 | basic_block new_bb; | |
b8698a0f | 918 | |
ebfd146a | 919 | new_var = vect_get_new_vect_var (vector_type, vect_simple_var, "cst_"); |
b8698a0f | 920 | add_referenced_var (new_var); |
ebfd146a IR |
921 | init_stmt = gimple_build_assign (new_var, vector_var); |
922 | new_temp = make_ssa_name (new_var, init_stmt); | |
923 | gimple_assign_set_lhs (init_stmt, new_temp); | |
924 | ||
925 | if (gsi) | |
926 | vect_finish_stmt_generation (stmt, init_stmt, gsi); | |
927 | else | |
928 | { | |
929 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
b8698a0f | 930 | |
a70d6342 IR |
931 | if (loop_vinfo) |
932 | { | |
933 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
934 | ||
935 | if (nested_in_vect_loop_p (loop, stmt)) | |
936 | loop = loop->inner; | |
b8698a0f | 937 | |
a70d6342 IR |
938 | pe = loop_preheader_edge (loop); |
939 | new_bb = gsi_insert_on_edge_immediate (pe, init_stmt); | |
940 | gcc_assert (!new_bb); | |
941 | } | |
942 | else | |
943 | { | |
944 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
945 | basic_block bb; | |
946 | gimple_stmt_iterator gsi_bb_start; | |
947 | ||
948 | gcc_assert (bb_vinfo); | |
949 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 950 | gsi_bb_start = gsi_after_labels (bb); |
a70d6342 IR |
951 | gsi_insert_before (&gsi_bb_start, init_stmt, GSI_SAME_STMT); |
952 | } | |
ebfd146a IR |
953 | } |
954 | ||
955 | if (vect_print_dump_info (REPORT_DETAILS)) | |
956 | { | |
957 | fprintf (vect_dump, "created new init_stmt: "); | |
958 | print_gimple_stmt (vect_dump, init_stmt, 0, TDF_SLIM); | |
959 | } | |
960 | ||
961 | vec_oprnd = gimple_assign_lhs (init_stmt); | |
962 | return vec_oprnd; | |
963 | } | |
964 | ||
a70d6342 | 965 | |
ebfd146a IR |
966 | /* Function vect_get_vec_def_for_operand. |
967 | ||
ff802fa1 | 968 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
969 | used in the vectorized stmt for STMT. |
970 | ||
971 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
972 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
973 | ||
974 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
975 | needs to be introduced. */ | |
976 | ||
977 | tree | |
978 | vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def) | |
979 | { | |
980 | tree vec_oprnd; | |
981 | gimple vec_stmt; | |
982 | gimple def_stmt; | |
983 | stmt_vec_info def_stmt_info = NULL; | |
984 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
9dc3f7de | 985 | unsigned int nunits; |
ebfd146a IR |
986 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
987 | tree vec_inv; | |
988 | tree vec_cst; | |
e7e9eb2f | 989 | tree t = NULL_TREE; |
ebfd146a | 990 | tree def; |
e7e9eb2f | 991 | int i; |
ebfd146a IR |
992 | enum vect_def_type dt; |
993 | bool is_simple_use; | |
994 | tree vector_type; | |
995 | ||
996 | if (vect_print_dump_info (REPORT_DETAILS)) | |
997 | { | |
998 | fprintf (vect_dump, "vect_get_vec_def_for_operand: "); | |
999 | print_generic_expr (vect_dump, op, TDF_SLIM); | |
1000 | } | |
1001 | ||
b8698a0f | 1002 | is_simple_use = vect_is_simple_use (op, loop_vinfo, NULL, &def_stmt, &def, |
a70d6342 | 1003 | &dt); |
ebfd146a IR |
1004 | gcc_assert (is_simple_use); |
1005 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1006 | { | |
1007 | if (def) | |
1008 | { | |
1009 | fprintf (vect_dump, "def = "); | |
1010 | print_generic_expr (vect_dump, def, TDF_SLIM); | |
1011 | } | |
1012 | if (def_stmt) | |
1013 | { | |
1014 | fprintf (vect_dump, " def_stmt = "); | |
1015 | print_gimple_stmt (vect_dump, def_stmt, 0, TDF_SLIM); | |
1016 | } | |
1017 | } | |
1018 | ||
1019 | switch (dt) | |
1020 | { | |
1021 | /* Case 1: operand is a constant. */ | |
1022 | case vect_constant_def: | |
1023 | { | |
7569a6cc RG |
1024 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
1025 | gcc_assert (vector_type); | |
9dc3f7de | 1026 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); |
7569a6cc | 1027 | |
b8698a0f | 1028 | if (scalar_def) |
ebfd146a IR |
1029 | *scalar_def = op; |
1030 | ||
1031 | /* Create 'vect_cst_ = {cst,cst,...,cst}' */ | |
1032 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1033 | fprintf (vect_dump, "Create vector_cst. nunits = %d", nunits); | |
1034 | ||
b9acc9f1 | 1035 | vec_cst = build_vector_from_val (vector_type, op); |
7569a6cc | 1036 | return vect_init_vector (stmt, vec_cst, vector_type, NULL); |
ebfd146a IR |
1037 | } |
1038 | ||
1039 | /* Case 2: operand is defined outside the loop - loop invariant. */ | |
8644a673 | 1040 | case vect_external_def: |
ebfd146a IR |
1041 | { |
1042 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); | |
1043 | gcc_assert (vector_type); | |
1044 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); | |
1045 | ||
b8698a0f | 1046 | if (scalar_def) |
ebfd146a IR |
1047 | *scalar_def = def; |
1048 | ||
1049 | /* Create 'vec_inv = {inv,inv,..,inv}' */ | |
1050 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1051 | fprintf (vect_dump, "Create vector_inv."); | |
1052 | ||
e7e9eb2f NF |
1053 | for (i = nunits - 1; i >= 0; --i) |
1054 | { | |
1055 | t = tree_cons (NULL_TREE, def, t); | |
1056 | } | |
1057 | ||
1058 | /* FIXME: use build_constructor directly. */ | |
1059 | vec_inv = build_constructor_from_list (vector_type, t); | |
ebfd146a IR |
1060 | return vect_init_vector (stmt, vec_inv, vector_type, NULL); |
1061 | } | |
1062 | ||
1063 | /* Case 3: operand is defined inside the loop. */ | |
8644a673 | 1064 | case vect_internal_def: |
ebfd146a | 1065 | { |
b8698a0f | 1066 | if (scalar_def) |
ebfd146a IR |
1067 | *scalar_def = NULL/* FIXME tuples: def_stmt*/; |
1068 | ||
1069 | /* Get the def from the vectorized stmt. */ | |
1070 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1071 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
1072 | gcc_assert (vec_stmt); | |
1073 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1074 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1075 | else if (is_gimple_call (vec_stmt)) | |
1076 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1077 | else | |
1078 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1079 | return vec_oprnd; | |
1080 | } | |
1081 | ||
1082 | /* Case 4: operand is defined by a loop header phi - reduction */ | |
1083 | case vect_reduction_def: | |
06066f92 | 1084 | case vect_double_reduction_def: |
7c5222ff | 1085 | case vect_nested_cycle: |
ebfd146a IR |
1086 | { |
1087 | struct loop *loop; | |
1088 | ||
1089 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
b8698a0f | 1090 | loop = (gimple_bb (def_stmt))->loop_father; |
ebfd146a IR |
1091 | |
1092 | /* Get the def before the loop */ | |
1093 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); | |
1094 | return get_initial_def_for_reduction (stmt, op, scalar_def); | |
1095 | } | |
1096 | ||
1097 | /* Case 5: operand is defined by loop-header phi - induction. */ | |
1098 | case vect_induction_def: | |
1099 | { | |
1100 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1101 | ||
1102 | /* Get the def from the vectorized stmt. */ | |
1103 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1104 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
1105 | gcc_assert (vec_stmt && gimple_code (vec_stmt) == GIMPLE_PHI); | |
1106 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1107 | return vec_oprnd; | |
1108 | } | |
1109 | ||
1110 | default: | |
1111 | gcc_unreachable (); | |
1112 | } | |
1113 | } | |
1114 | ||
1115 | ||
1116 | /* Function vect_get_vec_def_for_stmt_copy | |
1117 | ||
ff802fa1 | 1118 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1119 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1120 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1121 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1122 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1123 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1124 | DT is the type of the vector def VEC_OPRND. |
1125 | ||
1126 | Context: | |
1127 | In case the vectorization factor (VF) is bigger than the number | |
1128 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1129 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1130 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1131 | smallest data-type determines the VF, and as a result, when vectorizing |
1132 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1133 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1134 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1135 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1136 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1137 | ||
1138 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1139 | |
ebfd146a IR |
1140 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1141 | VS1.1: vx.1 = memref1 VS1.2 | |
1142 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1143 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1144 | |
1145 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1146 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1147 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1148 | VSnew.3: vz3 = vx.3 + ... | |
1149 | ||
1150 | The vectorization of S1 is explained in vectorizable_load. | |
1151 | The vectorization of S2: | |
b8698a0f L |
1152 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1153 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1154 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1155 | returns the vector-def 'vx.0'. |
1156 | ||
b8698a0f L |
1157 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1158 | function is called to get the relevant vector-def for each operand. It is | |
1159 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1160 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1161 | ||
b8698a0f L |
1162 | For example, to obtain the vector-def 'vx.1' in order to create the |
1163 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1164 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1165 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1166 | and return its def ('vx.1'). | |
1167 | Overall, to create the above sequence this function will be called 3 times: | |
1168 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1169 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1170 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1171 | ||
1172 | tree | |
1173 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1174 | { | |
1175 | gimple vec_stmt_for_operand; | |
1176 | stmt_vec_info def_stmt_info; | |
1177 | ||
1178 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1179 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1180 | return vec_oprnd; |
1181 | ||
1182 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1183 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1184 | gcc_assert (def_stmt_info); | |
1185 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1186 | gcc_assert (vec_stmt_for_operand); | |
1187 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1188 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) | |
1189 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1190 | else | |
1191 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1192 | return vec_oprnd; | |
1193 | } | |
1194 | ||
1195 | ||
1196 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1197 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1198 | |
1199 | static void | |
b8698a0f L |
1200 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
1201 | VEC(tree,heap) **vec_oprnds0, | |
ebfd146a IR |
1202 | VEC(tree,heap) **vec_oprnds1) |
1203 | { | |
1204 | tree vec_oprnd = VEC_pop (tree, *vec_oprnds0); | |
1205 | ||
1206 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
1207 | VEC_quick_push (tree, *vec_oprnds0, vec_oprnd); | |
1208 | ||
1209 | if (vec_oprnds1 && *vec_oprnds1) | |
1210 | { | |
1211 | vec_oprnd = VEC_pop (tree, *vec_oprnds1); | |
1212 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); | |
1213 | VEC_quick_push (tree, *vec_oprnds1, vec_oprnd); | |
1214 | } | |
1215 | } | |
1216 | ||
1217 | ||
ff802fa1 IR |
1218 | /* Get vectorized definitions for OP0 and OP1, or SLP_NODE if it is not |
1219 | NULL. */ | |
ebfd146a IR |
1220 | |
1221 | static void | |
1222 | vect_get_vec_defs (tree op0, tree op1, gimple stmt, | |
1223 | VEC(tree,heap) **vec_oprnds0, VEC(tree,heap) **vec_oprnds1, | |
1224 | slp_tree slp_node) | |
1225 | { | |
1226 | if (slp_node) | |
9dc3f7de | 1227 | vect_get_slp_defs (op0, op1, slp_node, vec_oprnds0, vec_oprnds1, -1); |
ebfd146a IR |
1228 | else |
1229 | { | |
1230 | tree vec_oprnd; | |
1231 | ||
b8698a0f L |
1232 | *vec_oprnds0 = VEC_alloc (tree, heap, 1); |
1233 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
ebfd146a IR |
1234 | VEC_quick_push (tree, *vec_oprnds0, vec_oprnd); |
1235 | ||
1236 | if (op1) | |
1237 | { | |
b8698a0f L |
1238 | *vec_oprnds1 = VEC_alloc (tree, heap, 1); |
1239 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL); | |
ebfd146a IR |
1240 | VEC_quick_push (tree, *vec_oprnds1, vec_oprnd); |
1241 | } | |
1242 | } | |
1243 | } | |
1244 | ||
1245 | ||
1246 | /* Function vect_finish_stmt_generation. | |
1247 | ||
1248 | Insert a new stmt. */ | |
1249 | ||
1250 | void | |
1251 | vect_finish_stmt_generation (gimple stmt, gimple vec_stmt, | |
1252 | gimple_stmt_iterator *gsi) | |
1253 | { | |
1254 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1255 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1256 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1257 | |
1258 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1259 | ||
1260 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); | |
1261 | ||
b8698a0f | 1262 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo, |
a70d6342 | 1263 | bb_vinfo)); |
ebfd146a IR |
1264 | |
1265 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1266 | { | |
1267 | fprintf (vect_dump, "add new stmt: "); | |
1268 | print_gimple_stmt (vect_dump, vec_stmt, 0, TDF_SLIM); | |
1269 | } | |
1270 | ||
1271 | gimple_set_location (vec_stmt, gimple_location (gsi_stmt (*gsi))); | |
1272 | } | |
1273 | ||
1274 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1275 | a function declaration if the target has a vectorized version | |
1276 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1277 | ||
1278 | tree | |
1279 | vectorizable_function (gimple call, tree vectype_out, tree vectype_in) | |
1280 | { | |
1281 | tree fndecl = gimple_call_fndecl (call); | |
ebfd146a IR |
1282 | |
1283 | /* We only handle functions that do not read or clobber memory -- i.e. | |
1284 | const or novops ones. */ | |
1285 | if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS))) | |
1286 | return NULL_TREE; | |
1287 | ||
1288 | if (!fndecl | |
1289 | || TREE_CODE (fndecl) != FUNCTION_DECL | |
1290 | || !DECL_BUILT_IN (fndecl)) | |
1291 | return NULL_TREE; | |
1292 | ||
62f7fd21 | 1293 | return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out, |
ebfd146a IR |
1294 | vectype_in); |
1295 | } | |
1296 | ||
1297 | /* Function vectorizable_call. | |
1298 | ||
b8698a0f L |
1299 | Check if STMT performs a function call that can be vectorized. |
1300 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1301 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1302 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1303 | ||
1304 | static bool | |
1305 | vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt) | |
1306 | { | |
1307 | tree vec_dest; | |
1308 | tree scalar_dest; | |
1309 | tree op, type; | |
1310 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1311 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
1312 | tree vectype_out, vectype_in; | |
1313 | int nunits_in; | |
1314 | int nunits_out; | |
1315 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
b690cc0f | 1316 | tree fndecl, new_temp, def, rhs_type; |
ebfd146a | 1317 | gimple def_stmt; |
0502fb85 UB |
1318 | enum vect_def_type dt[3] |
1319 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
63827fb8 | 1320 | gimple new_stmt = NULL; |
ebfd146a IR |
1321 | int ncopies, j; |
1322 | VEC(tree, heap) *vargs = NULL; | |
1323 | enum { NARROW, NONE, WIDEN } modifier; | |
1324 | size_t i, nargs; | |
1325 | ||
a70d6342 IR |
1326 | /* FORNOW: unsupported in basic block SLP. */ |
1327 | gcc_assert (loop_vinfo); | |
b8698a0f | 1328 | |
ebfd146a IR |
1329 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
1330 | return false; | |
1331 | ||
8644a673 | 1332 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1333 | return false; |
1334 | ||
1335 | /* FORNOW: SLP not supported. */ | |
1336 | if (STMT_SLP_TYPE (stmt_info)) | |
1337 | return false; | |
1338 | ||
1339 | /* Is STMT a vectorizable call? */ | |
1340 | if (!is_gimple_call (stmt)) | |
1341 | return false; | |
1342 | ||
1343 | if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
1344 | return false; | |
1345 | ||
5a2c1986 IR |
1346 | if (stmt_could_throw_p (stmt)) |
1347 | return false; | |
1348 | ||
b690cc0f RG |
1349 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
1350 | ||
ebfd146a IR |
1351 | /* Process function arguments. */ |
1352 | rhs_type = NULL_TREE; | |
b690cc0f | 1353 | vectype_in = NULL_TREE; |
ebfd146a IR |
1354 | nargs = gimple_call_num_args (stmt); |
1355 | ||
1b1562a5 MM |
1356 | /* Bail out if the function has more than three arguments, we do not have |
1357 | interesting builtin functions to vectorize with more than two arguments | |
1358 | except for fma. No arguments is also not good. */ | |
1359 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
1360 | return false; |
1361 | ||
1362 | for (i = 0; i < nargs; i++) | |
1363 | { | |
b690cc0f RG |
1364 | tree opvectype; |
1365 | ||
ebfd146a IR |
1366 | op = gimple_call_arg (stmt, i); |
1367 | ||
1368 | /* We can only handle calls with arguments of the same type. */ | |
1369 | if (rhs_type | |
8533c9d8 | 1370 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a IR |
1371 | { |
1372 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1373 | fprintf (vect_dump, "argument types differ."); | |
1374 | return false; | |
1375 | } | |
b690cc0f RG |
1376 | if (!rhs_type) |
1377 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 1378 | |
b690cc0f RG |
1379 | if (!vect_is_simple_use_1 (op, loop_vinfo, NULL, |
1380 | &def_stmt, &def, &dt[i], &opvectype)) | |
ebfd146a IR |
1381 | { |
1382 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1383 | fprintf (vect_dump, "use not simple."); | |
1384 | return false; | |
1385 | } | |
ebfd146a | 1386 | |
b690cc0f RG |
1387 | if (!vectype_in) |
1388 | vectype_in = opvectype; | |
1389 | else if (opvectype | |
1390 | && opvectype != vectype_in) | |
1391 | { | |
1392 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1393 | fprintf (vect_dump, "argument vector types differ."); | |
1394 | return false; | |
1395 | } | |
1396 | } | |
1397 | /* If all arguments are external or constant defs use a vector type with | |
1398 | the same size as the output vector type. */ | |
ebfd146a | 1399 | if (!vectype_in) |
b690cc0f | 1400 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
1401 | if (vec_stmt) |
1402 | gcc_assert (vectype_in); | |
1403 | if (!vectype_in) | |
1404 | { | |
1405 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1406 | { | |
1407 | fprintf (vect_dump, "no vectype for scalar type "); | |
1408 | print_generic_expr (vect_dump, rhs_type, TDF_SLIM); | |
1409 | } | |
1410 | ||
1411 | return false; | |
1412 | } | |
ebfd146a IR |
1413 | |
1414 | /* FORNOW */ | |
b690cc0f RG |
1415 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
1416 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
1417 | if (nunits_in == nunits_out / 2) |
1418 | modifier = NARROW; | |
1419 | else if (nunits_out == nunits_in) | |
1420 | modifier = NONE; | |
1421 | else if (nunits_out == nunits_in / 2) | |
1422 | modifier = WIDEN; | |
1423 | else | |
1424 | return false; | |
1425 | ||
1426 | /* For now, we only vectorize functions if a target specific builtin | |
1427 | is available. TODO -- in some cases, it might be profitable to | |
1428 | insert the calls for pieces of the vector, in order to be able | |
1429 | to vectorize other operations in the loop. */ | |
1430 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
1431 | if (fndecl == NULL_TREE) | |
1432 | { | |
1433 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1434 | fprintf (vect_dump, "function is not vectorizable."); | |
1435 | ||
1436 | return false; | |
1437 | } | |
1438 | ||
5006671f | 1439 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a IR |
1440 | |
1441 | if (modifier == NARROW) | |
1442 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
1443 | else | |
1444 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1445 | ||
1446 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
1447 | needs to be generated. */ | |
1448 | gcc_assert (ncopies >= 1); | |
1449 | ||
1450 | if (!vec_stmt) /* transformation not required. */ | |
1451 | { | |
1452 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1453 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1454 | fprintf (vect_dump, "=== vectorizable_call ==="); | |
1455 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
1456 | return true; | |
1457 | } | |
1458 | ||
1459 | /** Transform. **/ | |
1460 | ||
1461 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1462 | fprintf (vect_dump, "transform operation."); | |
1463 | ||
1464 | /* Handle def. */ | |
1465 | scalar_dest = gimple_call_lhs (stmt); | |
1466 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1467 | ||
1468 | prev_stmt_info = NULL; | |
1469 | switch (modifier) | |
1470 | { | |
1471 | case NONE: | |
1472 | for (j = 0; j < ncopies; ++j) | |
1473 | { | |
1474 | /* Build argument list for the vectorized call. */ | |
1475 | if (j == 0) | |
1476 | vargs = VEC_alloc (tree, heap, nargs); | |
1477 | else | |
1478 | VEC_truncate (tree, vargs, 0); | |
1479 | ||
1480 | for (i = 0; i < nargs; i++) | |
1481 | { | |
1482 | op = gimple_call_arg (stmt, i); | |
1483 | if (j == 0) | |
1484 | vec_oprnd0 | |
1485 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1486 | else | |
63827fb8 IR |
1487 | { |
1488 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
1489 | vec_oprnd0 | |
1490 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
1491 | } | |
ebfd146a IR |
1492 | |
1493 | VEC_quick_push (tree, vargs, vec_oprnd0); | |
1494 | } | |
1495 | ||
1496 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1497 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1498 | gimple_call_set_lhs (new_stmt, new_temp); | |
1499 | ||
1500 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7411b8f0 | 1501 | mark_symbols_for_renaming (new_stmt); |
ebfd146a IR |
1502 | |
1503 | if (j == 0) | |
1504 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1505 | else | |
1506 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1507 | ||
1508 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1509 | } | |
1510 | ||
1511 | break; | |
1512 | ||
1513 | case NARROW: | |
1514 | for (j = 0; j < ncopies; ++j) | |
1515 | { | |
1516 | /* Build argument list for the vectorized call. */ | |
1517 | if (j == 0) | |
1518 | vargs = VEC_alloc (tree, heap, nargs * 2); | |
1519 | else | |
1520 | VEC_truncate (tree, vargs, 0); | |
1521 | ||
1522 | for (i = 0; i < nargs; i++) | |
1523 | { | |
1524 | op = gimple_call_arg (stmt, i); | |
1525 | if (j == 0) | |
1526 | { | |
1527 | vec_oprnd0 | |
1528 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1529 | vec_oprnd1 | |
63827fb8 | 1530 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
1531 | } |
1532 | else | |
1533 | { | |
63827fb8 | 1534 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i); |
ebfd146a | 1535 | vec_oprnd0 |
63827fb8 | 1536 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 1537 | vec_oprnd1 |
63827fb8 | 1538 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
1539 | } |
1540 | ||
1541 | VEC_quick_push (tree, vargs, vec_oprnd0); | |
1542 | VEC_quick_push (tree, vargs, vec_oprnd1); | |
1543 | } | |
1544 | ||
1545 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1546 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1547 | gimple_call_set_lhs (new_stmt, new_temp); | |
1548 | ||
1549 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7411b8f0 | 1550 | mark_symbols_for_renaming (new_stmt); |
ebfd146a IR |
1551 | |
1552 | if (j == 0) | |
1553 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1554 | else | |
1555 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1556 | ||
1557 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1558 | } | |
1559 | ||
1560 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
1561 | ||
1562 | break; | |
1563 | ||
1564 | case WIDEN: | |
1565 | /* No current target implements this case. */ | |
1566 | return false; | |
1567 | } | |
1568 | ||
1569 | VEC_free (tree, heap, vargs); | |
1570 | ||
1571 | /* Update the exception handling table with the vector stmt if necessary. */ | |
1572 | if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt)) | |
1573 | gimple_purge_dead_eh_edges (gimple_bb (stmt)); | |
1574 | ||
1575 | /* The call in STMT might prevent it from being removed in dce. | |
1576 | We however cannot remove it here, due to the way the ssa name | |
1577 | it defines is mapped to the new definition. So just replace | |
1578 | rhs of the statement with something harmless. */ | |
1579 | ||
1580 | type = TREE_TYPE (scalar_dest); | |
1581 | new_stmt = gimple_build_assign (gimple_call_lhs (stmt), | |
e8160c9a | 1582 | build_zero_cst (type)); |
ebfd146a IR |
1583 | set_vinfo_for_stmt (new_stmt, stmt_info); |
1584 | set_vinfo_for_stmt (stmt, NULL); | |
1585 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
1586 | gsi_replace (gsi, new_stmt, false); | |
1587 | SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt; | |
1588 | ||
1589 | return true; | |
1590 | } | |
1591 | ||
1592 | ||
1593 | /* Function vect_gen_widened_results_half | |
1594 | ||
1595 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 1596 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 1597 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
1598 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
1599 | needs to be created (DECL is a function-decl of a target-builtin). | |
1600 | STMT is the original scalar stmt that we are vectorizing. */ | |
1601 | ||
1602 | static gimple | |
1603 | vect_gen_widened_results_half (enum tree_code code, | |
1604 | tree decl, | |
1605 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
1606 | tree vec_dest, gimple_stmt_iterator *gsi, | |
1607 | gimple stmt) | |
b8698a0f | 1608 | { |
ebfd146a | 1609 | gimple new_stmt; |
b8698a0f L |
1610 | tree new_temp; |
1611 | ||
1612 | /* Generate half of the widened result: */ | |
1613 | if (code == CALL_EXPR) | |
1614 | { | |
1615 | /* Target specific support */ | |
ebfd146a IR |
1616 | if (op_type == binary_op) |
1617 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
1618 | else | |
1619 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
1620 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1621 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
1622 | } |
1623 | else | |
ebfd146a | 1624 | { |
b8698a0f L |
1625 | /* Generic support */ |
1626 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
1627 | if (op_type != binary_op) |
1628 | vec_oprnd1 = NULL; | |
1629 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0, | |
1630 | vec_oprnd1); | |
1631 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1632 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 1633 | } |
ebfd146a IR |
1634 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1635 | ||
ebfd146a IR |
1636 | return new_stmt; |
1637 | } | |
1638 | ||
1639 | ||
b8698a0f L |
1640 | /* Check if STMT performs a conversion operation, that can be vectorized. |
1641 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1642 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1643 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1644 | ||
1645 | static bool | |
1646 | vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi, | |
1647 | gimple *vec_stmt, slp_tree slp_node) | |
1648 | { | |
1649 | tree vec_dest; | |
1650 | tree scalar_dest; | |
1651 | tree op0; | |
1652 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1653 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1654 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1655 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
1656 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; | |
1657 | tree new_temp; | |
1658 | tree def; | |
1659 | gimple def_stmt; | |
1660 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
1661 | gimple new_stmt = NULL; | |
1662 | stmt_vec_info prev_stmt_info; | |
1663 | int nunits_in; | |
1664 | int nunits_out; | |
1665 | tree vectype_out, vectype_in; | |
1666 | int ncopies, j; | |
b690cc0f | 1667 | tree rhs_type; |
ebfd146a IR |
1668 | tree builtin_decl; |
1669 | enum { NARROW, NONE, WIDEN } modifier; | |
1670 | int i; | |
1671 | VEC(tree,heap) *vec_oprnds0 = NULL; | |
1672 | tree vop0; | |
ebfd146a IR |
1673 | VEC(tree,heap) *dummy = NULL; |
1674 | int dummy_int; | |
1675 | ||
1676 | /* Is STMT a vectorizable conversion? */ | |
1677 | ||
a70d6342 IR |
1678 | /* FORNOW: unsupported in basic block SLP. */ |
1679 | gcc_assert (loop_vinfo); | |
b8698a0f | 1680 | |
ebfd146a IR |
1681 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
1682 | return false; | |
1683 | ||
8644a673 | 1684 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1685 | return false; |
1686 | ||
1687 | if (!is_gimple_assign (stmt)) | |
1688 | return false; | |
1689 | ||
1690 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
1691 | return false; | |
1692 | ||
1693 | code = gimple_assign_rhs_code (stmt); | |
1694 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
1695 | return false; | |
1696 | ||
1697 | /* Check types of lhs and rhs. */ | |
b690cc0f RG |
1698 | scalar_dest = gimple_assign_lhs (stmt); |
1699 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
1700 | ||
ebfd146a IR |
1701 | op0 = gimple_assign_rhs1 (stmt); |
1702 | rhs_type = TREE_TYPE (op0); | |
b690cc0f RG |
1703 | /* Check the operands of the operation. */ |
1704 | if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL, | |
1705 | &def_stmt, &def, &dt[0], &vectype_in)) | |
1706 | { | |
1707 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1708 | fprintf (vect_dump, "use not simple."); | |
1709 | return false; | |
1710 | } | |
1711 | /* If op0 is an external or constant defs use a vector type of | |
1712 | the same size as the output vector type. */ | |
ebfd146a | 1713 | if (!vectype_in) |
b690cc0f | 1714 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
1715 | if (vec_stmt) |
1716 | gcc_assert (vectype_in); | |
1717 | if (!vectype_in) | |
1718 | { | |
1719 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1720 | { | |
1721 | fprintf (vect_dump, "no vectype for scalar type "); | |
1722 | print_generic_expr (vect_dump, rhs_type, TDF_SLIM); | |
1723 | } | |
1724 | ||
1725 | return false; | |
1726 | } | |
ebfd146a IR |
1727 | |
1728 | /* FORNOW */ | |
b690cc0f RG |
1729 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
1730 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
1731 | if (nunits_in == nunits_out / 2) |
1732 | modifier = NARROW; | |
1733 | else if (nunits_out == nunits_in) | |
1734 | modifier = NONE; | |
1735 | else if (nunits_out == nunits_in / 2) | |
1736 | modifier = WIDEN; | |
1737 | else | |
1738 | return false; | |
1739 | ||
ebfd146a IR |
1740 | if (modifier == NARROW) |
1741 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
1742 | else | |
1743 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1744 | ||
ff802fa1 IR |
1745 | /* Multiple types in SLP are handled by creating the appropriate number of |
1746 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
1747 | case of SLP. */ | |
ebfd146a IR |
1748 | if (slp_node) |
1749 | ncopies = 1; | |
b8698a0f | 1750 | |
ebfd146a IR |
1751 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
1752 | needs to be generated. */ | |
1753 | gcc_assert (ncopies >= 1); | |
1754 | ||
ebfd146a IR |
1755 | /* Supportable by target? */ |
1756 | if ((modifier == NONE | |
88dd7150 | 1757 | && !targetm.vectorize.builtin_conversion (code, vectype_out, vectype_in)) |
ebfd146a | 1758 | || (modifier == WIDEN |
b690cc0f RG |
1759 | && !supportable_widening_operation (code, stmt, |
1760 | vectype_out, vectype_in, | |
ebfd146a IR |
1761 | &decl1, &decl2, |
1762 | &code1, &code2, | |
1763 | &dummy_int, &dummy)) | |
1764 | || (modifier == NARROW | |
b690cc0f | 1765 | && !supportable_narrowing_operation (code, vectype_out, vectype_in, |
ebfd146a IR |
1766 | &code1, &dummy_int, &dummy))) |
1767 | { | |
1768 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1769 | fprintf (vect_dump, "conversion not supported by target."); | |
1770 | return false; | |
1771 | } | |
1772 | ||
1773 | if (modifier != NONE) | |
1774 | { | |
ebfd146a IR |
1775 | /* FORNOW: SLP not supported. */ |
1776 | if (STMT_SLP_TYPE (stmt_info)) | |
b8698a0f | 1777 | return false; |
ebfd146a IR |
1778 | } |
1779 | ||
1780 | if (!vec_stmt) /* transformation not required. */ | |
1781 | { | |
1782 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
1783 | return true; | |
1784 | } | |
1785 | ||
1786 | /** Transform. **/ | |
1787 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1788 | fprintf (vect_dump, "transform conversion."); | |
1789 | ||
1790 | /* Handle def. */ | |
1791 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1792 | ||
1793 | if (modifier == NONE && !slp_node) | |
1794 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
1795 | ||
1796 | prev_stmt_info = NULL; | |
1797 | switch (modifier) | |
1798 | { | |
1799 | case NONE: | |
1800 | for (j = 0; j < ncopies; j++) | |
1801 | { | |
ebfd146a | 1802 | if (j == 0) |
b8698a0f | 1803 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node); |
ebfd146a IR |
1804 | else |
1805 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
1806 | ||
1807 | builtin_decl = | |
88dd7150 RG |
1808 | targetm.vectorize.builtin_conversion (code, |
1809 | vectype_out, vectype_in); | |
ac47786e | 1810 | FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0) |
b8698a0f | 1811 | { |
ebfd146a IR |
1812 | /* Arguments are ready. create the new vector stmt. */ |
1813 | new_stmt = gimple_build_call (builtin_decl, 1, vop0); | |
1814 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1815 | gimple_call_set_lhs (new_stmt, new_temp); | |
1816 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a IR |
1817 | if (slp_node) |
1818 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
1819 | } | |
1820 | ||
1821 | if (j == 0) | |
1822 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1823 | else | |
1824 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1825 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1826 | } | |
1827 | break; | |
1828 | ||
1829 | case WIDEN: | |
1830 | /* In case the vectorization factor (VF) is bigger than the number | |
1831 | of elements that we can fit in a vectype (nunits), we have to | |
1832 | generate more than one vector stmt - i.e - we need to "unroll" | |
1833 | the vector stmt by a factor VF/nunits. */ | |
1834 | for (j = 0; j < ncopies; j++) | |
1835 | { | |
1836 | if (j == 0) | |
1837 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
1838 | else | |
1839 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1840 | ||
ebfd146a IR |
1841 | /* Generate first half of the widened result: */ |
1842 | new_stmt | |
b8698a0f | 1843 | = vect_gen_widened_results_half (code1, decl1, |
ebfd146a IR |
1844 | vec_oprnd0, vec_oprnd1, |
1845 | unary_op, vec_dest, gsi, stmt); | |
1846 | if (j == 0) | |
1847 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1848 | else | |
1849 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1850 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1851 | ||
1852 | /* Generate second half of the widened result: */ | |
1853 | new_stmt | |
1854 | = vect_gen_widened_results_half (code2, decl2, | |
1855 | vec_oprnd0, vec_oprnd1, | |
1856 | unary_op, vec_dest, gsi, stmt); | |
1857 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1858 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1859 | } | |
1860 | break; | |
1861 | ||
1862 | case NARROW: | |
1863 | /* In case the vectorization factor (VF) is bigger than the number | |
1864 | of elements that we can fit in a vectype (nunits), we have to | |
1865 | generate more than one vector stmt - i.e - we need to "unroll" | |
1866 | the vector stmt by a factor VF/nunits. */ | |
1867 | for (j = 0; j < ncopies; j++) | |
1868 | { | |
1869 | /* Handle uses. */ | |
1870 | if (j == 0) | |
1871 | { | |
1872 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
1873 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1874 | } | |
1875 | else | |
1876 | { | |
1877 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd1); | |
1878 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1879 | } | |
1880 | ||
9dc3f7de | 1881 | /* Arguments are ready. Create the new vector stmt. */ |
ebfd146a IR |
1882 | new_stmt = gimple_build_assign_with_ops (code1, vec_dest, vec_oprnd0, |
1883 | vec_oprnd1); | |
1884 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1885 | gimple_assign_set_lhs (new_stmt, new_temp); | |
1886 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1887 | ||
1888 | if (j == 0) | |
1889 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1890 | else | |
1891 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1892 | ||
1893 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1894 | } | |
1895 | ||
1896 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
1897 | } | |
1898 | ||
1899 | if (vec_oprnds0) | |
b8698a0f | 1900 | VEC_free (tree, heap, vec_oprnds0); |
ebfd146a IR |
1901 | |
1902 | return true; | |
1903 | } | |
ff802fa1 IR |
1904 | |
1905 | ||
ebfd146a IR |
1906 | /* Function vectorizable_assignment. |
1907 | ||
b8698a0f L |
1908 | Check if STMT performs an assignment (copy) that can be vectorized. |
1909 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1910 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1911 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1912 | ||
1913 | static bool | |
1914 | vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi, | |
1915 | gimple *vec_stmt, slp_tree slp_node) | |
1916 | { | |
1917 | tree vec_dest; | |
1918 | tree scalar_dest; | |
1919 | tree op; | |
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 | tree new_temp; | |
1924 | tree def; | |
1925 | gimple def_stmt; | |
1926 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
fde9c428 | 1927 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
ebfd146a | 1928 | int ncopies; |
f18b55bd | 1929 | int i, j; |
ebfd146a IR |
1930 | VEC(tree,heap) *vec_oprnds = NULL; |
1931 | tree vop; | |
a70d6342 | 1932 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
f18b55bd IR |
1933 | gimple new_stmt = NULL; |
1934 | stmt_vec_info prev_stmt_info = NULL; | |
fde9c428 RG |
1935 | enum tree_code code; |
1936 | tree vectype_in; | |
ebfd146a IR |
1937 | |
1938 | /* Multiple types in SLP are handled by creating the appropriate number of | |
1939 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
1940 | case of SLP. */ | |
1941 | if (slp_node) | |
1942 | ncopies = 1; | |
1943 | else | |
1944 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
1945 | ||
1946 | gcc_assert (ncopies >= 1); | |
ebfd146a | 1947 | |
a70d6342 | 1948 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
1949 | return false; |
1950 | ||
8644a673 | 1951 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1952 | return false; |
1953 | ||
1954 | /* Is vectorizable assignment? */ | |
1955 | if (!is_gimple_assign (stmt)) | |
1956 | return false; | |
1957 | ||
1958 | scalar_dest = gimple_assign_lhs (stmt); | |
1959 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
1960 | return false; | |
1961 | ||
fde9c428 | 1962 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 1963 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
1964 | || code == PAREN_EXPR |
1965 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
1966 | op = gimple_assign_rhs1 (stmt); |
1967 | else | |
1968 | return false; | |
1969 | ||
fde9c428 RG |
1970 | if (!vect_is_simple_use_1 (op, loop_vinfo, bb_vinfo, |
1971 | &def_stmt, &def, &dt[0], &vectype_in)) | |
ebfd146a IR |
1972 | { |
1973 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1974 | fprintf (vect_dump, "use not simple."); | |
1975 | return false; | |
1976 | } | |
1977 | ||
fde9c428 RG |
1978 | /* We can handle NOP_EXPR conversions that do not change the number |
1979 | of elements or the vector size. */ | |
1980 | if (CONVERT_EXPR_CODE_P (code) | |
1981 | && (!vectype_in | |
1982 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
1983 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
1984 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
1985 | return false; | |
1986 | ||
ebfd146a IR |
1987 | if (!vec_stmt) /* transformation not required. */ |
1988 | { | |
1989 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
1990 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1991 | fprintf (vect_dump, "=== vectorizable_assignment ==="); | |
1992 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
1993 | return true; | |
1994 | } | |
1995 | ||
1996 | /** Transform. **/ | |
1997 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1998 | fprintf (vect_dump, "transform assignment."); | |
1999 | ||
2000 | /* Handle def. */ | |
2001 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2002 | ||
2003 | /* Handle use. */ | |
f18b55bd | 2004 | for (j = 0; j < ncopies; j++) |
ebfd146a | 2005 | { |
f18b55bd IR |
2006 | /* Handle uses. */ |
2007 | if (j == 0) | |
2008 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); | |
2009 | else | |
2010 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
2011 | ||
2012 | /* Arguments are ready. create the new vector stmt. */ | |
ac47786e | 2013 | FOR_EACH_VEC_ELT (tree, vec_oprnds, i, vop) |
f18b55bd | 2014 | { |
fde9c428 | 2015 | if (CONVERT_EXPR_CODE_P (code)) |
4a73490d | 2016 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
2017 | new_stmt = gimple_build_assign (vec_dest, vop); |
2018 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2019 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2020 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2021 | if (slp_node) | |
2022 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2023 | } | |
ebfd146a IR |
2024 | |
2025 | if (slp_node) | |
f18b55bd IR |
2026 | continue; |
2027 | ||
2028 | if (j == 0) | |
2029 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2030 | else | |
2031 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2032 | ||
2033 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2034 | } | |
b8698a0f L |
2035 | |
2036 | VEC_free (tree, heap, vec_oprnds); | |
ebfd146a IR |
2037 | return true; |
2038 | } | |
2039 | ||
9dc3f7de IR |
2040 | |
2041 | /* Function vectorizable_shift. | |
2042 | ||
2043 | Check if STMT performs a shift operation that can be vectorized. | |
2044 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2045 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2046 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2047 | ||
2048 | static bool | |
2049 | vectorizable_shift (gimple stmt, gimple_stmt_iterator *gsi, | |
2050 | gimple *vec_stmt, slp_tree slp_node) | |
2051 | { | |
2052 | tree vec_dest; | |
2053 | tree scalar_dest; | |
2054 | tree op0, op1 = NULL; | |
2055 | tree vec_oprnd1 = NULL_TREE; | |
2056 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2057 | tree vectype; | |
2058 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2059 | enum tree_code code; | |
2060 | enum machine_mode vec_mode; | |
2061 | tree new_temp; | |
2062 | optab optab; | |
2063 | int icode; | |
2064 | enum machine_mode optab_op2_mode; | |
2065 | tree def; | |
2066 | gimple def_stmt; | |
2067 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2068 | gimple new_stmt = NULL; | |
2069 | stmt_vec_info prev_stmt_info; | |
2070 | int nunits_in; | |
2071 | int nunits_out; | |
2072 | tree vectype_out; | |
2073 | int ncopies; | |
2074 | int j, i; | |
2075 | VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; | |
2076 | tree vop0, vop1; | |
2077 | unsigned int k; | |
2078 | bool scalar_shift_arg = false; | |
2079 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
2080 | int vf; | |
2081 | ||
2082 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
2083 | return false; | |
2084 | ||
2085 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
2086 | return false; | |
2087 | ||
2088 | /* Is STMT a vectorizable binary/unary operation? */ | |
2089 | if (!is_gimple_assign (stmt)) | |
2090 | return false; | |
2091 | ||
2092 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2093 | return false; | |
2094 | ||
2095 | code = gimple_assign_rhs_code (stmt); | |
2096 | ||
2097 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
2098 | || code == RROTATE_EXPR)) | |
2099 | return false; | |
2100 | ||
2101 | scalar_dest = gimple_assign_lhs (stmt); | |
2102 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2103 | ||
2104 | op0 = gimple_assign_rhs1 (stmt); | |
2105 | if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo, | |
2106 | &def_stmt, &def, &dt[0], &vectype)) | |
2107 | { | |
2108 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2109 | fprintf (vect_dump, "use not simple."); | |
2110 | return false; | |
2111 | } | |
2112 | /* If op0 is an external or constant def use a vector type with | |
2113 | the same size as the output vector type. */ | |
2114 | if (!vectype) | |
2115 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
2116 | if (vec_stmt) | |
2117 | gcc_assert (vectype); | |
2118 | if (!vectype) | |
2119 | { | |
2120 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2121 | { | |
2122 | fprintf (vect_dump, "no vectype for scalar type "); | |
2123 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
2124 | } | |
2125 | ||
2126 | return false; | |
2127 | } | |
2128 | ||
2129 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
2130 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
2131 | if (nunits_out != nunits_in) | |
2132 | return false; | |
2133 | ||
2134 | op1 = gimple_assign_rhs2 (stmt); | |
2135 | if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt[1])) | |
2136 | { | |
2137 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2138 | fprintf (vect_dump, "use not simple."); | |
2139 | return false; | |
2140 | } | |
2141 | ||
2142 | if (loop_vinfo) | |
2143 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
2144 | else | |
2145 | vf = 1; | |
2146 | ||
2147 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2148 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2149 | case of SLP. */ | |
2150 | if (slp_node) | |
2151 | ncopies = 1; | |
2152 | else | |
2153 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2154 | ||
2155 | gcc_assert (ncopies >= 1); | |
2156 | ||
2157 | /* Determine whether the shift amount is a vector, or scalar. If the | |
2158 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
2159 | ||
2160 | /* Vector shifted by vector. */ | |
2161 | if (dt[1] == vect_internal_def) | |
2162 | { | |
2163 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
2164 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2165 | fprintf (vect_dump, "vector/vector shift/rotate found."); | |
2166 | } | |
2167 | /* See if the machine has a vector shifted by scalar insn and if not | |
2168 | then see if it has a vector shifted by vector insn. */ | |
2169 | else if (dt[1] == vect_constant_def || dt[1] == vect_external_def) | |
2170 | { | |
2171 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
2172 | if (optab | |
2173 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
2174 | { | |
2175 | scalar_shift_arg = true; | |
2176 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2177 | fprintf (vect_dump, "vector/scalar shift/rotate found."); | |
2178 | } | |
2179 | else | |
2180 | { | |
2181 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
2182 | if (optab | |
2183 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
2184 | != CODE_FOR_nothing)) | |
2185 | { | |
2186 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2187 | fprintf (vect_dump, "vector/vector shift/rotate found."); | |
2188 | ||
2189 | /* Unlike the other binary operators, shifts/rotates have | |
2190 | the rhs being int, instead of the same type as the lhs, | |
2191 | so make sure the scalar is the right type if we are | |
2192 | dealing with vectors of short/char. */ | |
2193 | if (dt[1] == vect_constant_def) | |
2194 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
2195 | } | |
2196 | } | |
2197 | } | |
2198 | else | |
2199 | { | |
2200 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2201 | fprintf (vect_dump, "operand mode requires invariant argument."); | |
2202 | return false; | |
2203 | } | |
2204 | ||
2205 | /* Supportable by target? */ | |
2206 | if (!optab) | |
2207 | { | |
2208 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2209 | fprintf (vect_dump, "no optab."); | |
2210 | return false; | |
2211 | } | |
2212 | vec_mode = TYPE_MODE (vectype); | |
2213 | icode = (int) optab_handler (optab, vec_mode); | |
2214 | if (icode == CODE_FOR_nothing) | |
2215 | { | |
2216 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2217 | fprintf (vect_dump, "op not supported by target."); | |
2218 | /* Check only during analysis. */ | |
2219 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
2220 | || (vf < vect_min_worthwhile_factor (code) | |
2221 | && !vec_stmt)) | |
2222 | return false; | |
2223 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2224 | fprintf (vect_dump, "proceeding using word mode."); | |
2225 | } | |
2226 | ||
2227 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
2228 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
2229 | && vf < vect_min_worthwhile_factor (code) | |
2230 | && !vec_stmt) | |
2231 | { | |
2232 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2233 | fprintf (vect_dump, "not worthwhile without SIMD support."); | |
2234 | return false; | |
2235 | } | |
2236 | ||
2237 | if (!vec_stmt) /* transformation not required. */ | |
2238 | { | |
2239 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
2240 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2241 | fprintf (vect_dump, "=== vectorizable_shift ==="); | |
2242 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2243 | return true; | |
2244 | } | |
2245 | ||
2246 | /** Transform. **/ | |
2247 | ||
2248 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2249 | fprintf (vect_dump, "transform binary/unary operation."); | |
2250 | ||
2251 | /* Handle def. */ | |
2252 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2253 | ||
2254 | /* Allocate VECs for vector operands. In case of SLP, vector operands are | |
2255 | created in the previous stages of the recursion, so no allocation is | |
2256 | needed, except for the case of shift with scalar shift argument. In that | |
2257 | case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to | |
2258 | be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE. | |
2259 | In case of loop-based vectorization we allocate VECs of size 1. We | |
2260 | allocate VEC_OPRNDS1 only in case of binary operation. */ | |
2261 | if (!slp_node) | |
2262 | { | |
2263 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
2264 | vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
2265 | } | |
2266 | else if (scalar_shift_arg) | |
2267 | vec_oprnds1 = VEC_alloc (tree, heap, slp_node->vec_stmts_size); | |
2268 | ||
2269 | prev_stmt_info = NULL; | |
2270 | for (j = 0; j < ncopies; j++) | |
2271 | { | |
2272 | /* Handle uses. */ | |
2273 | if (j == 0) | |
2274 | { | |
2275 | if (scalar_shift_arg) | |
2276 | { | |
2277 | /* Vector shl and shr insn patterns can be defined with scalar | |
2278 | operand 2 (shift operand). In this case, use constant or loop | |
2279 | invariant op1 directly, without extending it to vector mode | |
2280 | first. */ | |
2281 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
2282 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
2283 | { | |
2284 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2285 | fprintf (vect_dump, "operand 1 using scalar mode."); | |
2286 | vec_oprnd1 = op1; | |
2287 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2288 | if (slp_node) | |
2289 | { | |
2290 | /* Store vec_oprnd1 for every vector stmt to be created | |
2291 | for SLP_NODE. We check during the analysis that all | |
2292 | the shift arguments are the same. | |
2293 | TODO: Allow different constants for different vector | |
2294 | stmts generated for an SLP instance. */ | |
2295 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
2296 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2297 | } | |
2298 | } | |
2299 | } | |
2300 | ||
2301 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
2302 | (a special case for certain kind of vector shifts); otherwise, | |
2303 | operand 1 should be of a vector type (the usual case). */ | |
2304 | if (vec_oprnd1) | |
2305 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2306 | slp_node); | |
2307 | else | |
2308 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
2309 | slp_node); | |
2310 | } | |
2311 | else | |
2312 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
2313 | ||
2314 | /* Arguments are ready. Create the new vector stmt. */ | |
2315 | FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0) | |
2316 | { | |
2317 | vop1 = VEC_index (tree, vec_oprnds1, i); | |
2318 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); | |
2319 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2320 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2321 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2322 | if (slp_node) | |
2323 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2324 | } | |
2325 | ||
2326 | if (slp_node) | |
2327 | continue; | |
2328 | ||
2329 | if (j == 0) | |
2330 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2331 | else | |
2332 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2333 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2334 | } | |
2335 | ||
2336 | VEC_free (tree, heap, vec_oprnds0); | |
2337 | VEC_free (tree, heap, vec_oprnds1); | |
2338 | ||
2339 | return true; | |
2340 | } | |
2341 | ||
2342 | ||
ebfd146a IR |
2343 | /* Function vectorizable_operation. |
2344 | ||
16949072 RG |
2345 | Check if STMT performs a binary, unary or ternary operation that can |
2346 | be vectorized. | |
b8698a0f | 2347 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2348 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2349 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2350 | ||
2351 | static bool | |
2352 | vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi, | |
2353 | gimple *vec_stmt, slp_tree slp_node) | |
2354 | { | |
2355 | tree vec_dest; | |
2356 | tree scalar_dest; | |
16949072 | 2357 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 2358 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
b690cc0f | 2359 | tree vectype; |
ebfd146a IR |
2360 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
2361 | enum tree_code code; | |
2362 | enum machine_mode vec_mode; | |
2363 | tree new_temp; | |
2364 | int op_type; | |
2365 | optab optab; | |
2366 | int icode; | |
ebfd146a IR |
2367 | tree def; |
2368 | gimple def_stmt; | |
16949072 RG |
2369 | enum vect_def_type dt[3] |
2370 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
ebfd146a IR |
2371 | gimple new_stmt = NULL; |
2372 | stmt_vec_info prev_stmt_info; | |
b690cc0f | 2373 | int nunits_in; |
ebfd146a IR |
2374 | int nunits_out; |
2375 | tree vectype_out; | |
2376 | int ncopies; | |
2377 | int j, i; | |
16949072 RG |
2378 | VEC(tree,heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL, *vec_oprnds2 = NULL; |
2379 | tree vop0, vop1, vop2; | |
a70d6342 IR |
2380 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2381 | int vf; | |
2382 | ||
a70d6342 | 2383 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2384 | return false; |
2385 | ||
8644a673 | 2386 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2387 | return false; |
2388 | ||
2389 | /* Is STMT a vectorizable binary/unary operation? */ | |
2390 | if (!is_gimple_assign (stmt)) | |
2391 | return false; | |
2392 | ||
2393 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2394 | return false; | |
2395 | ||
ebfd146a IR |
2396 | code = gimple_assign_rhs_code (stmt); |
2397 | ||
2398 | /* For pointer addition, we should use the normal plus for | |
2399 | the vector addition. */ | |
2400 | if (code == POINTER_PLUS_EXPR) | |
2401 | code = PLUS_EXPR; | |
2402 | ||
2403 | /* Support only unary or binary operations. */ | |
2404 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 2405 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a IR |
2406 | { |
2407 | if (vect_print_dump_info (REPORT_DETAILS)) | |
16949072 RG |
2408 | fprintf (vect_dump, "num. args = %d (not unary/binary/ternary op).", |
2409 | op_type); | |
ebfd146a IR |
2410 | return false; |
2411 | } | |
2412 | ||
b690cc0f RG |
2413 | scalar_dest = gimple_assign_lhs (stmt); |
2414 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2415 | ||
ebfd146a | 2416 | op0 = gimple_assign_rhs1 (stmt); |
b690cc0f RG |
2417 | if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo, |
2418 | &def_stmt, &def, &dt[0], &vectype)) | |
ebfd146a IR |
2419 | { |
2420 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2421 | fprintf (vect_dump, "use not simple."); | |
2422 | return false; | |
2423 | } | |
b690cc0f RG |
2424 | /* If op0 is an external or constant def use a vector type with |
2425 | the same size as the output vector type. */ | |
2426 | if (!vectype) | |
2427 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
2428 | if (vec_stmt) |
2429 | gcc_assert (vectype); | |
2430 | if (!vectype) | |
2431 | { | |
2432 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2433 | { | |
2434 | fprintf (vect_dump, "no vectype for scalar type "); | |
2435 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
2436 | } | |
2437 | ||
2438 | return false; | |
2439 | } | |
b690cc0f RG |
2440 | |
2441 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
2442 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
2443 | if (nunits_out != nunits_in) | |
2444 | return false; | |
ebfd146a | 2445 | |
16949072 | 2446 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
2447 | { |
2448 | op1 = gimple_assign_rhs2 (stmt); | |
b8698a0f | 2449 | if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def, |
a70d6342 | 2450 | &dt[1])) |
ebfd146a IR |
2451 | { |
2452 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2453 | fprintf (vect_dump, "use not simple."); | |
2454 | return false; | |
2455 | } | |
2456 | } | |
16949072 RG |
2457 | if (op_type == ternary_op) |
2458 | { | |
2459 | op2 = gimple_assign_rhs3 (stmt); | |
2460 | if (!vect_is_simple_use (op2, loop_vinfo, bb_vinfo, &def_stmt, &def, | |
2461 | &dt[2])) | |
2462 | { | |
2463 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2464 | fprintf (vect_dump, "use not simple."); | |
2465 | return false; | |
2466 | } | |
2467 | } | |
ebfd146a | 2468 | |
b690cc0f RG |
2469 | if (loop_vinfo) |
2470 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
2471 | else | |
2472 | vf = 1; | |
2473 | ||
2474 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 2475 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f RG |
2476 | case of SLP. */ |
2477 | if (slp_node) | |
2478 | ncopies = 1; | |
2479 | else | |
2480 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2481 | ||
2482 | gcc_assert (ncopies >= 1); | |
2483 | ||
9dc3f7de | 2484 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
2485 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
2486 | || code == RROTATE_EXPR) | |
9dc3f7de | 2487 | return false; |
ebfd146a | 2488 | |
16949072 | 2489 | optab = optab_for_tree_code (code, vectype, optab_default); |
ebfd146a IR |
2490 | |
2491 | /* Supportable by target? */ | |
2492 | if (!optab) | |
2493 | { | |
2494 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2495 | fprintf (vect_dump, "no optab."); | |
2496 | return false; | |
2497 | } | |
2498 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 2499 | icode = (int) optab_handler (optab, vec_mode); |
ebfd146a IR |
2500 | if (icode == CODE_FOR_nothing) |
2501 | { | |
2502 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2503 | fprintf (vect_dump, "op not supported by target."); | |
2504 | /* Check only during analysis. */ | |
2505 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
a70d6342 | 2506 | || (vf < vect_min_worthwhile_factor (code) |
ebfd146a IR |
2507 | && !vec_stmt)) |
2508 | return false; | |
2509 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2510 | fprintf (vect_dump, "proceeding using word mode."); | |
2511 | } | |
2512 | ||
ff802fa1 | 2513 | /* Worthwhile without SIMD support? Check only during analysis. */ |
ebfd146a | 2514 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) |
a70d6342 | 2515 | && vf < vect_min_worthwhile_factor (code) |
ebfd146a IR |
2516 | && !vec_stmt) |
2517 | { | |
2518 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2519 | fprintf (vect_dump, "not worthwhile without SIMD support."); | |
2520 | return false; | |
2521 | } | |
2522 | ||
2523 | if (!vec_stmt) /* transformation not required. */ | |
2524 | { | |
2525 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; | |
2526 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2527 | fprintf (vect_dump, "=== vectorizable_operation ==="); | |
2528 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2529 | return true; | |
2530 | } | |
2531 | ||
2532 | /** Transform. **/ | |
2533 | ||
2534 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2535 | fprintf (vect_dump, "transform binary/unary operation."); | |
2536 | ||
2537 | /* Handle def. */ | |
2538 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2539 | ||
ff802fa1 | 2540 | /* Allocate VECs for vector operands. In case of SLP, vector operands are |
ebfd146a | 2541 | created in the previous stages of the recursion, so no allocation is |
ff802fa1 | 2542 | needed, except for the case of shift with scalar shift argument. In that |
ebfd146a IR |
2543 | case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to |
2544 | be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE. | |
ff802fa1 | 2545 | In case of loop-based vectorization we allocate VECs of size 1. We |
b8698a0f | 2546 | allocate VEC_OPRNDS1 only in case of binary operation. */ |
ebfd146a IR |
2547 | if (!slp_node) |
2548 | { | |
2549 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
16949072 | 2550 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a | 2551 | vec_oprnds1 = VEC_alloc (tree, heap, 1); |
16949072 RG |
2552 | if (op_type == ternary_op) |
2553 | vec_oprnds2 = VEC_alloc (tree, heap, 1); | |
ebfd146a | 2554 | } |
ebfd146a IR |
2555 | |
2556 | /* In case the vectorization factor (VF) is bigger than the number | |
2557 | of elements that we can fit in a vectype (nunits), we have to generate | |
2558 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 2559 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 2560 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 2561 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 2562 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
2563 | stmts that use the defs of the current stmt. The example below |
2564 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
2565 | we need to create 4 vectorized stmts): | |
ebfd146a IR |
2566 | |
2567 | before vectorization: | |
2568 | RELATED_STMT VEC_STMT | |
2569 | S1: x = memref - - | |
2570 | S2: z = x + 1 - - | |
2571 | ||
2572 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
2573 | there): | |
2574 | RELATED_STMT VEC_STMT | |
2575 | VS1_0: vx0 = memref0 VS1_1 - | |
2576 | VS1_1: vx1 = memref1 VS1_2 - | |
2577 | VS1_2: vx2 = memref2 VS1_3 - | |
2578 | VS1_3: vx3 = memref3 - - | |
2579 | S1: x = load - VS1_0 | |
2580 | S2: z = x + 1 - - | |
2581 | ||
2582 | step2: vectorize stmt S2 (done here): | |
2583 | To vectorize stmt S2 we first need to find the relevant vector | |
ff802fa1 | 2584 | def for the first operand 'x'. This is, as usual, obtained from |
ebfd146a | 2585 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt |
ff802fa1 IR |
2586 | that defines 'x' (S1). This way we find the stmt VS1_0, and the |
2587 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
ebfd146a IR |
2588 | the vector stmt VS2_0, and as usual, record it in the |
2589 | STMT_VINFO_VEC_STMT of stmt S2. | |
2590 | When creating the second copy (VS2_1), we obtain the relevant vector | |
2591 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
ff802fa1 IR |
2592 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant |
2593 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
ebfd146a | 2594 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. |
ff802fa1 | 2595 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting |
ebfd146a IR |
2596 | chain of stmts and pointers: |
2597 | RELATED_STMT VEC_STMT | |
2598 | VS1_0: vx0 = memref0 VS1_1 - | |
2599 | VS1_1: vx1 = memref1 VS1_2 - | |
2600 | VS1_2: vx2 = memref2 VS1_3 - | |
2601 | VS1_3: vx3 = memref3 - - | |
2602 | S1: x = load - VS1_0 | |
2603 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
2604 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
2605 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
2606 | VS2_3: vz3 = vx3 + v1 - - | |
2607 | S2: z = x + 1 - VS2_0 */ | |
2608 | ||
2609 | prev_stmt_info = NULL; | |
2610 | for (j = 0; j < ncopies; j++) | |
2611 | { | |
2612 | /* Handle uses. */ | |
2613 | if (j == 0) | |
2614 | { | |
16949072 | 2615 | if (op_type == binary_op || op_type == ternary_op) |
b8698a0f | 2616 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, |
ebfd146a IR |
2617 | slp_node); |
2618 | else | |
b8698a0f | 2619 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, |
ebfd146a | 2620 | slp_node); |
16949072 RG |
2621 | if (op_type == ternary_op) |
2622 | { | |
2623 | vec_oprnds2 = VEC_alloc (tree, heap, 1); | |
2624 | VEC_quick_push (tree, vec_oprnds2, | |
2625 | vect_get_vec_def_for_operand (op2, stmt, NULL)); | |
2626 | } | |
ebfd146a IR |
2627 | } |
2628 | else | |
16949072 RG |
2629 | { |
2630 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
2631 | if (op_type == ternary_op) | |
2632 | { | |
2633 | tree vec_oprnd = VEC_pop (tree, vec_oprnds2); | |
2634 | VEC_quick_push (tree, vec_oprnds2, | |
2635 | vect_get_vec_def_for_stmt_copy (dt[2], | |
2636 | vec_oprnd)); | |
2637 | } | |
2638 | } | |
ebfd146a | 2639 | |
9dc3f7de | 2640 | /* Arguments are ready. Create the new vector stmt. */ |
ac47786e | 2641 | FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0) |
ebfd146a | 2642 | { |
16949072 RG |
2643 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
2644 | ? VEC_index (tree, vec_oprnds1, i) : NULL_TREE); | |
2645 | vop2 = ((op_type == ternary_op) | |
2646 | ? VEC_index (tree, vec_oprnds2, i) : NULL_TREE); | |
2647 | new_stmt = gimple_build_assign_with_ops3 (code, vec_dest, | |
2648 | vop0, vop1, vop2); | |
ebfd146a IR |
2649 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2650 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2651 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2652 | if (slp_node) | |
2653 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2654 | } | |
2655 | ||
2656 | if (slp_node) | |
2657 | continue; | |
2658 | ||
2659 | if (j == 0) | |
2660 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2661 | else | |
2662 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2663 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2664 | } | |
2665 | ||
2666 | VEC_free (tree, heap, vec_oprnds0); | |
2667 | if (vec_oprnds1) | |
2668 | VEC_free (tree, heap, vec_oprnds1); | |
16949072 RG |
2669 | if (vec_oprnds2) |
2670 | VEC_free (tree, heap, vec_oprnds2); | |
ebfd146a IR |
2671 | |
2672 | return true; | |
2673 | } | |
2674 | ||
2675 | ||
ff802fa1 | 2676 | /* Get vectorized definitions for loop-based vectorization. For the first |
b8698a0f L |
2677 | operand we call vect_get_vec_def_for_operand() (with OPRND containing |
2678 | scalar operand), and for the rest we get a copy with | |
ebfd146a IR |
2679 | vect_get_vec_def_for_stmt_copy() using the previous vector definition |
2680 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
2681 | The vectors are collected into VEC_OPRNDS. */ | |
2682 | ||
2683 | static void | |
b8698a0f | 2684 | vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt, |
ebfd146a IR |
2685 | VEC (tree, heap) **vec_oprnds, int multi_step_cvt) |
2686 | { | |
2687 | tree vec_oprnd; | |
2688 | ||
2689 | /* Get first vector operand. */ | |
2690 | /* All the vector operands except the very first one (that is scalar oprnd) | |
2691 | are stmt copies. */ | |
b8698a0f | 2692 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) |
ebfd146a IR |
2693 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL); |
2694 | else | |
2695 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
2696 | ||
2697 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
2698 | ||
2699 | /* Get second vector operand. */ | |
2700 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
2701 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
b8698a0f | 2702 | |
ebfd146a IR |
2703 | *oprnd = vec_oprnd; |
2704 | ||
b8698a0f | 2705 | /* For conversion in multiple steps, continue to get operands |
ebfd146a IR |
2706 | recursively. */ |
2707 | if (multi_step_cvt) | |
b8698a0f | 2708 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); |
ebfd146a IR |
2709 | } |
2710 | ||
2711 | ||
2712 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
b8698a0f | 2713 | For multi-step conversions store the resulting vectors and call the function |
ebfd146a IR |
2714 | recursively. */ |
2715 | ||
2716 | static void | |
2717 | vect_create_vectorized_demotion_stmts (VEC (tree, heap) **vec_oprnds, | |
2718 | int multi_step_cvt, gimple stmt, | |
2719 | VEC (tree, heap) *vec_dsts, | |
2720 | gimple_stmt_iterator *gsi, | |
2721 | slp_tree slp_node, enum tree_code code, | |
2722 | stmt_vec_info *prev_stmt_info) | |
2723 | { | |
2724 | unsigned int i; | |
2725 | tree vop0, vop1, new_tmp, vec_dest; | |
2726 | gimple new_stmt; | |
2727 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2728 | ||
b8698a0f | 2729 | vec_dest = VEC_pop (tree, vec_dsts); |
ebfd146a IR |
2730 | |
2731 | for (i = 0; i < VEC_length (tree, *vec_oprnds); i += 2) | |
2732 | { | |
2733 | /* Create demotion operation. */ | |
2734 | vop0 = VEC_index (tree, *vec_oprnds, i); | |
2735 | vop1 = VEC_index (tree, *vec_oprnds, i + 1); | |
2736 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); | |
2737 | new_tmp = make_ssa_name (vec_dest, new_stmt); | |
2738 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
2739 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2740 | ||
2741 | if (multi_step_cvt) | |
2742 | /* Store the resulting vector for next recursive call. */ | |
b8698a0f | 2743 | VEC_replace (tree, *vec_oprnds, i/2, new_tmp); |
ebfd146a IR |
2744 | else |
2745 | { | |
b8698a0f | 2746 | /* This is the last step of the conversion sequence. Store the |
ebfd146a IR |
2747 | vectors in SLP_NODE or in vector info of the scalar statement |
2748 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
2749 | if (slp_node) | |
2750 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2751 | else | |
2752 | { | |
2753 | if (!*prev_stmt_info) | |
2754 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2755 | else | |
2756 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
2757 | ||
2758 | *prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2759 | } | |
2760 | } | |
2761 | } | |
2762 | ||
2763 | /* For multi-step demotion operations we first generate demotion operations | |
b8698a0f | 2764 | from the source type to the intermediate types, and then combine the |
ebfd146a IR |
2765 | results (stored in VEC_OPRNDS) in demotion operation to the destination |
2766 | type. */ | |
2767 | if (multi_step_cvt) | |
2768 | { | |
2769 | /* At each level of recursion we have have of the operands we had at the | |
2770 | previous level. */ | |
2771 | VEC_truncate (tree, *vec_oprnds, (i+1)/2); | |
b8698a0f | 2772 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
ebfd146a IR |
2773 | stmt, vec_dsts, gsi, slp_node, |
2774 | code, prev_stmt_info); | |
2775 | } | |
2776 | } | |
2777 | ||
2778 | ||
2779 | /* Function vectorizable_type_demotion | |
2780 | ||
2781 | Check if STMT performs a binary or unary operation that involves | |
2782 | type demotion, and if it can be vectorized. | |
2783 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2784 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2785 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2786 | ||
2787 | static bool | |
2788 | vectorizable_type_demotion (gimple stmt, gimple_stmt_iterator *gsi, | |
2789 | gimple *vec_stmt, slp_tree slp_node) | |
2790 | { | |
2791 | tree vec_dest; | |
2792 | tree scalar_dest; | |
2793 | tree op0; | |
2794 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2795 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2796 | enum tree_code code, code1 = ERROR_MARK; | |
2797 | tree def; | |
2798 | gimple def_stmt; | |
2799 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2800 | stmt_vec_info prev_stmt_info; | |
2801 | int nunits_in; | |
2802 | int nunits_out; | |
2803 | tree vectype_out; | |
2804 | int ncopies; | |
2805 | int j, i; | |
2806 | tree vectype_in; | |
2807 | int multi_step_cvt = 0; | |
2808 | VEC (tree, heap) *vec_oprnds0 = NULL; | |
2809 | VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL; | |
2810 | tree last_oprnd, intermediate_type; | |
2811 | ||
a70d6342 IR |
2812 | /* FORNOW: not supported by basic block SLP vectorization. */ |
2813 | gcc_assert (loop_vinfo); | |
2814 | ||
ebfd146a IR |
2815 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
2816 | return false; | |
2817 | ||
8644a673 | 2818 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2819 | return false; |
2820 | ||
2821 | /* Is STMT a vectorizable type-demotion operation? */ | |
2822 | if (!is_gimple_assign (stmt)) | |
2823 | return false; | |
2824 | ||
2825 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2826 | return false; | |
2827 | ||
2828 | code = gimple_assign_rhs_code (stmt); | |
2829 | if (!CONVERT_EXPR_CODE_P (code)) | |
2830 | return false; | |
2831 | ||
b690cc0f RG |
2832 | scalar_dest = gimple_assign_lhs (stmt); |
2833 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2834 | ||
2835 | /* Check the operands of the operation. */ | |
ebfd146a | 2836 | op0 = gimple_assign_rhs1 (stmt); |
b690cc0f RG |
2837 | if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) |
2838 | && INTEGRAL_TYPE_P (TREE_TYPE (op0))) | |
2839 | || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest)) | |
2840 | && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0)) | |
2841 | && CONVERT_EXPR_CODE_P (code)))) | |
2842 | return false; | |
2843 | if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL, | |
2844 | &def_stmt, &def, &dt[0], &vectype_in)) | |
2845 | { | |
2846 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2847 | fprintf (vect_dump, "use not simple."); | |
2848 | return false; | |
2849 | } | |
2850 | /* If op0 is an external def use a vector type with the | |
2851 | same size as the output vector type if possible. */ | |
2852 | if (!vectype_in) | |
2853 | vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
2854 | if (vec_stmt) |
2855 | gcc_assert (vectype_in); | |
ebfd146a | 2856 | if (!vectype_in) |
7d8930a0 IR |
2857 | { |
2858 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2859 | { | |
2860 | fprintf (vect_dump, "no vectype for scalar type "); | |
2861 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
2862 | } | |
2863 | ||
2864 | return false; | |
2865 | } | |
ebfd146a | 2866 | |
b690cc0f | 2867 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
ebfd146a IR |
2868 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); |
2869 | if (nunits_in >= nunits_out) | |
2870 | return false; | |
2871 | ||
2872 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 2873 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a IR |
2874 | case of SLP. */ |
2875 | if (slp_node) | |
2876 | ncopies = 1; | |
2877 | else | |
2878 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
ebfd146a IR |
2879 | gcc_assert (ncopies >= 1); |
2880 | ||
ebfd146a | 2881 | /* Supportable by target? */ |
b690cc0f RG |
2882 | if (!supportable_narrowing_operation (code, vectype_out, vectype_in, |
2883 | &code1, &multi_step_cvt, &interm_types)) | |
ebfd146a IR |
2884 | return false; |
2885 | ||
ebfd146a IR |
2886 | if (!vec_stmt) /* transformation not required. */ |
2887 | { | |
2888 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
2889 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2890 | fprintf (vect_dump, "=== vectorizable_demotion ==="); | |
2891 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2892 | return true; | |
2893 | } | |
2894 | ||
2895 | /** Transform. **/ | |
2896 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2897 | fprintf (vect_dump, "transform type demotion operation. ncopies = %d.", | |
2898 | ncopies); | |
2899 | ||
b8698a0f L |
2900 | /* In case of multi-step demotion, we first generate demotion operations to |
2901 | the intermediate types, and then from that types to the final one. | |
ebfd146a | 2902 | We create vector destinations for the intermediate type (TYPES) received |
b8698a0f | 2903 | from supportable_narrowing_operation, and store them in the correct order |
ebfd146a IR |
2904 | for future use in vect_create_vectorized_demotion_stmts(). */ |
2905 | if (multi_step_cvt) | |
2906 | vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1); | |
2907 | else | |
2908 | vec_dsts = VEC_alloc (tree, heap, 1); | |
b8698a0f | 2909 | |
ebfd146a IR |
2910 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); |
2911 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2912 | ||
2913 | if (multi_step_cvt) | |
2914 | { | |
b8698a0f | 2915 | for (i = VEC_length (tree, interm_types) - 1; |
ebfd146a IR |
2916 | VEC_iterate (tree, interm_types, i, intermediate_type); i--) |
2917 | { | |
b8698a0f | 2918 | vec_dest = vect_create_destination_var (scalar_dest, |
ebfd146a IR |
2919 | intermediate_type); |
2920 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2921 | } | |
2922 | } | |
2923 | ||
2924 | /* In case the vectorization factor (VF) is bigger than the number | |
2925 | of elements that we can fit in a vectype (nunits), we have to generate | |
2926 | more than one vector stmt - i.e - we need to "unroll" the | |
2927 | vector stmt by a factor VF/nunits. */ | |
2928 | last_oprnd = op0; | |
2929 | prev_stmt_info = NULL; | |
2930 | for (j = 0; j < ncopies; j++) | |
2931 | { | |
2932 | /* Handle uses. */ | |
2933 | if (slp_node) | |
9dc3f7de | 2934 | vect_get_slp_defs (op0, NULL_TREE, slp_node, &vec_oprnds0, NULL, -1); |
ebfd146a IR |
2935 | else |
2936 | { | |
2937 | VEC_free (tree, heap, vec_oprnds0); | |
2938 | vec_oprnds0 = VEC_alloc (tree, heap, | |
2939 | (multi_step_cvt ? vect_pow2 (multi_step_cvt) * 2 : 2)); | |
b8698a0f | 2940 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
ebfd146a IR |
2941 | vect_pow2 (multi_step_cvt) - 1); |
2942 | } | |
2943 | ||
9dc3f7de | 2944 | /* Arguments are ready. Create the new vector stmts. */ |
ebfd146a | 2945 | tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts); |
b8698a0f | 2946 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, |
ebfd146a | 2947 | multi_step_cvt, stmt, tmp_vec_dsts, |
b8698a0f | 2948 | gsi, slp_node, code1, |
ebfd146a IR |
2949 | &prev_stmt_info); |
2950 | } | |
2951 | ||
2952 | VEC_free (tree, heap, vec_oprnds0); | |
2953 | VEC_free (tree, heap, vec_dsts); | |
2954 | VEC_free (tree, heap, tmp_vec_dsts); | |
2955 | VEC_free (tree, heap, interm_types); | |
2956 | ||
2957 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2958 | return true; | |
2959 | } | |
2960 | ||
2961 | ||
2962 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
ff802fa1 | 2963 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store |
ebfd146a IR |
2964 | the resulting vectors and call the function recursively. */ |
2965 | ||
2966 | static void | |
2967 | vect_create_vectorized_promotion_stmts (VEC (tree, heap) **vec_oprnds0, | |
2968 | VEC (tree, heap) **vec_oprnds1, | |
2969 | int multi_step_cvt, gimple stmt, | |
2970 | VEC (tree, heap) *vec_dsts, | |
2971 | gimple_stmt_iterator *gsi, | |
2972 | slp_tree slp_node, enum tree_code code1, | |
b8698a0f | 2973 | enum tree_code code2, tree decl1, |
ebfd146a IR |
2974 | tree decl2, int op_type, |
2975 | stmt_vec_info *prev_stmt_info) | |
2976 | { | |
2977 | int i; | |
2978 | tree vop0, vop1, new_tmp1, new_tmp2, vec_dest; | |
2979 | gimple new_stmt1, new_stmt2; | |
2980 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2981 | VEC (tree, heap) *vec_tmp; | |
2982 | ||
2983 | vec_dest = VEC_pop (tree, vec_dsts); | |
2984 | vec_tmp = VEC_alloc (tree, heap, VEC_length (tree, *vec_oprnds0) * 2); | |
2985 | ||
ac47786e | 2986 | FOR_EACH_VEC_ELT (tree, *vec_oprnds0, i, vop0) |
ebfd146a IR |
2987 | { |
2988 | if (op_type == binary_op) | |
2989 | vop1 = VEC_index (tree, *vec_oprnds1, i); | |
2990 | else | |
2991 | vop1 = NULL_TREE; | |
2992 | ||
2993 | /* Generate the two halves of promotion operation. */ | |
b8698a0f | 2994 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, |
ebfd146a IR |
2995 | op_type, vec_dest, gsi, stmt); |
2996 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
2997 | op_type, vec_dest, gsi, stmt); | |
2998 | if (is_gimple_call (new_stmt1)) | |
2999 | { | |
3000 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
3001 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
3002 | } | |
3003 | else | |
3004 | { | |
3005 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
3006 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
3007 | } | |
3008 | ||
3009 | if (multi_step_cvt) | |
3010 | { | |
3011 | /* Store the results for the recursive call. */ | |
3012 | VEC_quick_push (tree, vec_tmp, new_tmp1); | |
3013 | VEC_quick_push (tree, vec_tmp, new_tmp2); | |
3014 | } | |
3015 | else | |
3016 | { | |
3017 | /* Last step of promotion sequience - store the results. */ | |
3018 | if (slp_node) | |
3019 | { | |
3020 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt1); | |
3021 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt2); | |
3022 | } | |
3023 | else | |
3024 | { | |
3025 | if (!*prev_stmt_info) | |
3026 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt1; | |
3027 | else | |
3028 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt1; | |
3029 | ||
3030 | *prev_stmt_info = vinfo_for_stmt (new_stmt1); | |
3031 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt2; | |
3032 | *prev_stmt_info = vinfo_for_stmt (new_stmt2); | |
3033 | } | |
3034 | } | |
3035 | } | |
3036 | ||
3037 | if (multi_step_cvt) | |
3038 | { | |
b8698a0f | 3039 | /* For multi-step promotion operation we first generate we call the |
ff802fa1 | 3040 | function recurcively for every stage. We start from the input type, |
ebfd146a IR |
3041 | create promotion operations to the intermediate types, and then |
3042 | create promotions to the output type. */ | |
3043 | *vec_oprnds0 = VEC_copy (tree, heap, vec_tmp); | |
ebfd146a IR |
3044 | vect_create_vectorized_promotion_stmts (vec_oprnds0, vec_oprnds1, |
3045 | multi_step_cvt - 1, stmt, | |
3046 | vec_dsts, gsi, slp_node, code1, | |
3047 | code2, decl2, decl2, op_type, | |
3048 | prev_stmt_info); | |
3049 | } | |
ff802fa1 IR |
3050 | |
3051 | VEC_free (tree, heap, vec_tmp); | |
ebfd146a | 3052 | } |
b8698a0f | 3053 | |
ebfd146a IR |
3054 | |
3055 | /* Function vectorizable_type_promotion | |
3056 | ||
3057 | Check if STMT performs a binary or unary operation that involves | |
3058 | type promotion, and if it can be vectorized. | |
3059 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3060 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3061 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3062 | ||
3063 | static bool | |
3064 | vectorizable_type_promotion (gimple stmt, gimple_stmt_iterator *gsi, | |
3065 | gimple *vec_stmt, slp_tree slp_node) | |
3066 | { | |
3067 | tree vec_dest; | |
3068 | tree scalar_dest; | |
3069 | tree op0, op1 = NULL; | |
3070 | tree vec_oprnd0=NULL, vec_oprnd1=NULL; | |
3071 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3072 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3073 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
3074 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; | |
b8698a0f | 3075 | int op_type; |
ebfd146a IR |
3076 | tree def; |
3077 | gimple def_stmt; | |
3078 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
3079 | stmt_vec_info prev_stmt_info; | |
3080 | int nunits_in; | |
3081 | int nunits_out; | |
3082 | tree vectype_out; | |
3083 | int ncopies; | |
3084 | int j, i; | |
3085 | tree vectype_in; | |
3086 | tree intermediate_type = NULL_TREE; | |
3087 | int multi_step_cvt = 0; | |
3088 | VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; | |
3089 | VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL; | |
b8698a0f | 3090 | |
a70d6342 IR |
3091 | /* FORNOW: not supported by basic block SLP vectorization. */ |
3092 | gcc_assert (loop_vinfo); | |
b8698a0f | 3093 | |
ebfd146a IR |
3094 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
3095 | return false; | |
3096 | ||
8644a673 | 3097 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3098 | return false; |
3099 | ||
3100 | /* Is STMT a vectorizable type-promotion operation? */ | |
3101 | if (!is_gimple_assign (stmt)) | |
3102 | return false; | |
3103 | ||
3104 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3105 | return false; | |
3106 | ||
3107 | code = gimple_assign_rhs_code (stmt); | |
3108 | if (!CONVERT_EXPR_CODE_P (code) | |
3109 | && code != WIDEN_MULT_EXPR) | |
3110 | return false; | |
3111 | ||
b690cc0f RG |
3112 | scalar_dest = gimple_assign_lhs (stmt); |
3113 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
3114 | ||
3115 | /* Check the operands of the operation. */ | |
ebfd146a | 3116 | op0 = gimple_assign_rhs1 (stmt); |
b690cc0f RG |
3117 | if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) |
3118 | && INTEGRAL_TYPE_P (TREE_TYPE (op0))) | |
3119 | || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest)) | |
3120 | && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0)) | |
3121 | && CONVERT_EXPR_CODE_P (code)))) | |
3122 | return false; | |
3123 | if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL, | |
3124 | &def_stmt, &def, &dt[0], &vectype_in)) | |
3125 | { | |
3126 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3127 | fprintf (vect_dump, "use not simple."); | |
3128 | return false; | |
3129 | } | |
3130 | /* If op0 is an external or constant def use a vector type with | |
3131 | the same size as the output vector type. */ | |
3132 | if (!vectype_in) | |
3133 | vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
3134 | if (vec_stmt) |
3135 | gcc_assert (vectype_in); | |
ebfd146a | 3136 | if (!vectype_in) |
7d8930a0 IR |
3137 | { |
3138 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3139 | { | |
3140 | fprintf (vect_dump, "no vectype for scalar type "); | |
3141 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
3142 | } | |
3143 | ||
3144 | return false; | |
3145 | } | |
ebfd146a | 3146 | |
b690cc0f | 3147 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
ebfd146a IR |
3148 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); |
3149 | if (nunits_in <= nunits_out) | |
3150 | return false; | |
3151 | ||
3152 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 3153 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a IR |
3154 | case of SLP. */ |
3155 | if (slp_node) | |
3156 | ncopies = 1; | |
3157 | else | |
3158 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
3159 | ||
3160 | gcc_assert (ncopies >= 1); | |
3161 | ||
ebfd146a IR |
3162 | op_type = TREE_CODE_LENGTH (code); |
3163 | if (op_type == binary_op) | |
3164 | { | |
3165 | op1 = gimple_assign_rhs2 (stmt); | |
a70d6342 | 3166 | if (!vect_is_simple_use (op1, loop_vinfo, NULL, &def_stmt, &def, &dt[1])) |
ebfd146a IR |
3167 | { |
3168 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3169 | fprintf (vect_dump, "use not simple."); | |
3170 | return false; | |
3171 | } | |
3172 | } | |
3173 | ||
3174 | /* Supportable by target? */ | |
b690cc0f | 3175 | if (!supportable_widening_operation (code, stmt, vectype_out, vectype_in, |
ebfd146a IR |
3176 | &decl1, &decl2, &code1, &code2, |
3177 | &multi_step_cvt, &interm_types)) | |
3178 | return false; | |
3179 | ||
3180 | /* Binary widening operation can only be supported directly by the | |
3181 | architecture. */ | |
3182 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
3183 | ||
ebfd146a IR |
3184 | if (!vec_stmt) /* transformation not required. */ |
3185 | { | |
3186 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
3187 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3188 | fprintf (vect_dump, "=== vectorizable_promotion ==="); | |
3189 | vect_model_simple_cost (stmt_info, 2*ncopies, dt, NULL); | |
3190 | return true; | |
3191 | } | |
3192 | ||
3193 | /** Transform. **/ | |
3194 | ||
3195 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3196 | fprintf (vect_dump, "transform type promotion operation. ncopies = %d.", | |
3197 | ncopies); | |
3198 | ||
3199 | /* Handle def. */ | |
b8698a0f | 3200 | /* In case of multi-step promotion, we first generate promotion operations |
ebfd146a | 3201 | to the intermediate types, and then from that types to the final one. |
b8698a0f L |
3202 | We store vector destination in VEC_DSTS in the correct order for |
3203 | recursive creation of promotion operations in | |
ebfd146a IR |
3204 | vect_create_vectorized_promotion_stmts(). Vector destinations are created |
3205 | according to TYPES recieved from supportable_widening_operation(). */ | |
3206 | if (multi_step_cvt) | |
3207 | vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1); | |
3208 | else | |
3209 | vec_dsts = VEC_alloc (tree, heap, 1); | |
3210 | ||
3211 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
3212 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
3213 | ||
3214 | if (multi_step_cvt) | |
3215 | { | |
3216 | for (i = VEC_length (tree, interm_types) - 1; | |
3217 | VEC_iterate (tree, interm_types, i, intermediate_type); i--) | |
3218 | { | |
3219 | vec_dest = vect_create_destination_var (scalar_dest, | |
3220 | intermediate_type); | |
3221 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
3222 | } | |
3223 | } | |
b8698a0f | 3224 | |
ebfd146a IR |
3225 | if (!slp_node) |
3226 | { | |
b8698a0f | 3227 | vec_oprnds0 = VEC_alloc (tree, heap, |
ebfd146a IR |
3228 | (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); |
3229 | if (op_type == binary_op) | |
3230 | vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
3231 | } | |
3232 | ||
3233 | /* In case the vectorization factor (VF) is bigger than the number | |
3234 | of elements that we can fit in a vectype (nunits), we have to generate | |
3235 | more than one vector stmt - i.e - we need to "unroll" the | |
3236 | vector stmt by a factor VF/nunits. */ | |
3237 | ||
3238 | prev_stmt_info = NULL; | |
3239 | for (j = 0; j < ncopies; j++) | |
3240 | { | |
3241 | /* Handle uses. */ | |
3242 | if (j == 0) | |
3243 | { | |
3244 | if (slp_node) | |
9dc3f7de IR |
3245 | vect_get_slp_defs (op0, op1, slp_node, &vec_oprnds0, |
3246 | &vec_oprnds1, -1); | |
ebfd146a IR |
3247 | else |
3248 | { | |
3249 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
3250 | VEC_quick_push (tree, vec_oprnds0, vec_oprnd0); | |
3251 | if (op_type == binary_op) | |
3252 | { | |
3253 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, NULL); | |
3254 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
3255 | } | |
3256 | } | |
3257 | } | |
3258 | else | |
3259 | { | |
3260 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
3261 | VEC_replace (tree, vec_oprnds0, 0, vec_oprnd0); | |
3262 | if (op_type == binary_op) | |
3263 | { | |
3264 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd1); | |
3265 | VEC_replace (tree, vec_oprnds1, 0, vec_oprnd1); | |
3266 | } | |
3267 | } | |
3268 | ||
9dc3f7de | 3269 | /* Arguments are ready. Create the new vector stmts. */ |
ebfd146a IR |
3270 | tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts); |
3271 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, &vec_oprnds1, | |
b8698a0f | 3272 | multi_step_cvt, stmt, |
ebfd146a IR |
3273 | tmp_vec_dsts, |
3274 | gsi, slp_node, code1, code2, | |
3275 | decl1, decl2, op_type, | |
3276 | &prev_stmt_info); | |
3277 | } | |
3278 | ||
3279 | VEC_free (tree, heap, vec_dsts); | |
3280 | VEC_free (tree, heap, tmp_vec_dsts); | |
3281 | VEC_free (tree, heap, interm_types); | |
3282 | VEC_free (tree, heap, vec_oprnds0); | |
3283 | VEC_free (tree, heap, vec_oprnds1); | |
3284 | ||
3285 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
3286 | return true; | |
3287 | } | |
3288 | ||
3289 | ||
3290 | /* Function vectorizable_store. | |
3291 | ||
b8698a0f L |
3292 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
3293 | can be vectorized. | |
3294 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
3295 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3296 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3297 | ||
3298 | static bool | |
3299 | vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
3300 | slp_tree slp_node) | |
3301 | { | |
3302 | tree scalar_dest; | |
3303 | tree data_ref; | |
3304 | tree op; | |
3305 | tree vec_oprnd = NULL_TREE; | |
3306 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3307 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
3308 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3309 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 3310 | struct loop *loop = NULL; |
ebfd146a IR |
3311 | enum machine_mode vec_mode; |
3312 | tree dummy; | |
3313 | enum dr_alignment_support alignment_support_scheme; | |
3314 | tree def; | |
3315 | gimple def_stmt; | |
3316 | enum vect_def_type dt; | |
3317 | stmt_vec_info prev_stmt_info = NULL; | |
3318 | tree dataref_ptr = NULL_TREE; | |
3319 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
3320 | int ncopies; | |
3321 | int j; | |
3322 | gimple next_stmt, first_stmt = NULL; | |
3323 | bool strided_store = false; | |
3324 | unsigned int group_size, i; | |
3325 | VEC(tree,heap) *dr_chain = NULL, *oprnds = NULL, *result_chain = NULL; | |
3326 | bool inv_p; | |
3327 | VEC(tree,heap) *vec_oprnds = NULL; | |
3328 | bool slp = (slp_node != NULL); | |
ebfd146a | 3329 | unsigned int vec_num; |
a70d6342 IR |
3330 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3331 | ||
3332 | if (loop_vinfo) | |
3333 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
3334 | |
3335 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3336 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3337 | case of SLP. */ | |
3338 | if (slp) | |
3339 | ncopies = 1; | |
3340 | else | |
3341 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3342 | ||
3343 | gcc_assert (ncopies >= 1); | |
3344 | ||
3345 | /* FORNOW. This restriction should be relaxed. */ | |
a70d6342 | 3346 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) |
ebfd146a IR |
3347 | { |
3348 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3349 | fprintf (vect_dump, "multiple types in nested loop."); | |
3350 | return false; | |
3351 | } | |
3352 | ||
a70d6342 | 3353 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3354 | return false; |
3355 | ||
8644a673 | 3356 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3357 | return false; |
3358 | ||
3359 | /* Is vectorizable store? */ | |
3360 | ||
3361 | if (!is_gimple_assign (stmt)) | |
3362 | return false; | |
3363 | ||
3364 | scalar_dest = gimple_assign_lhs (stmt); | |
3365 | if (TREE_CODE (scalar_dest) != ARRAY_REF | |
3366 | && TREE_CODE (scalar_dest) != INDIRECT_REF | |
e9dbe7bb IR |
3367 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
3368 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
3369 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
3370 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
3371 | return false; |
3372 | ||
3373 | gcc_assert (gimple_assign_single_p (stmt)); | |
3374 | op = gimple_assign_rhs1 (stmt); | |
a70d6342 | 3375 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt)) |
ebfd146a IR |
3376 | { |
3377 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3378 | fprintf (vect_dump, "use not simple."); | |
3379 | return false; | |
3380 | } | |
3381 | ||
3382 | /* The scalar rhs type needs to be trivially convertible to the vector | |
3383 | component type. This should always be the case. */ | |
3384 | if (!useless_type_conversion_p (TREE_TYPE (vectype), TREE_TYPE (op))) | |
b8698a0f | 3385 | { |
ebfd146a IR |
3386 | if (vect_print_dump_info (REPORT_DETAILS)) |
3387 | fprintf (vect_dump, "??? operands of different types"); | |
3388 | return false; | |
3389 | } | |
3390 | ||
3391 | vec_mode = TYPE_MODE (vectype); | |
3392 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
3393 | (e.g. - array initialization with 0). */ | |
947131ba | 3394 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
3395 | return false; |
3396 | ||
3397 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
3398 | return false; | |
3399 | ||
a1e53f3f L |
3400 | if (tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0) |
3401 | { | |
3402 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3403 | fprintf (vect_dump, "negative step for store."); | |
3404 | return false; | |
3405 | } | |
3406 | ||
ebfd146a IR |
3407 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) |
3408 | { | |
3409 | strided_store = true; | |
3410 | first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
3411 | if (!vect_strided_store_supported (vectype) | |
3412 | && !PURE_SLP_STMT (stmt_info) && !slp) | |
3413 | return false; | |
b8698a0f | 3414 | |
ebfd146a IR |
3415 | if (first_stmt == stmt) |
3416 | { | |
3417 | /* STMT is the leader of the group. Check the operands of all the | |
3418 | stmts of the group. */ | |
3419 | next_stmt = DR_GROUP_NEXT_DR (stmt_info); | |
3420 | while (next_stmt) | |
3421 | { | |
3422 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
3423 | op = gimple_assign_rhs1 (next_stmt); | |
b8698a0f | 3424 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, |
a70d6342 | 3425 | &def, &dt)) |
ebfd146a IR |
3426 | { |
3427 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3428 | fprintf (vect_dump, "use not simple."); | |
3429 | return false; | |
3430 | } | |
3431 | next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); | |
3432 | } | |
3433 | } | |
3434 | } | |
3435 | ||
3436 | if (!vec_stmt) /* transformation not required. */ | |
3437 | { | |
3438 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
3439 | vect_model_store_cost (stmt_info, ncopies, dt, NULL); | |
3440 | return true; | |
3441 | } | |
3442 | ||
3443 | /** Transform. **/ | |
3444 | ||
3445 | if (strided_store) | |
3446 | { | |
3447 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
3448 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
3449 | ||
3450 | DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; | |
3451 | ||
3452 | /* FORNOW */ | |
a70d6342 | 3453 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
3454 | |
3455 | /* We vectorize all the stmts of the interleaving group when we | |
3456 | reach the last stmt in the group. */ | |
b8698a0f | 3457 | if (DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
ebfd146a IR |
3458 | < DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)) |
3459 | && !slp) | |
3460 | { | |
3461 | *vec_stmt = NULL; | |
3462 | return true; | |
3463 | } | |
3464 | ||
3465 | if (slp) | |
4b5caab7 IR |
3466 | { |
3467 | strided_store = false; | |
3468 | /* VEC_NUM is the number of vect stmts to be created for this | |
3469 | group. */ | |
3470 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
3471 | first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0); | |
3472 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
3473 | } | |
ebfd146a | 3474 | else |
4b5caab7 IR |
3475 | /* VEC_NUM is the number of vect stmts to be created for this |
3476 | group. */ | |
ebfd146a IR |
3477 | vec_num = group_size; |
3478 | } | |
b8698a0f | 3479 | else |
ebfd146a IR |
3480 | { |
3481 | first_stmt = stmt; | |
3482 | first_dr = dr; | |
3483 | group_size = vec_num = 1; | |
ebfd146a | 3484 | } |
b8698a0f | 3485 | |
ebfd146a IR |
3486 | if (vect_print_dump_info (REPORT_DETAILS)) |
3487 | fprintf (vect_dump, "transform store. ncopies = %d",ncopies); | |
3488 | ||
3489 | dr_chain = VEC_alloc (tree, heap, group_size); | |
3490 | oprnds = VEC_alloc (tree, heap, group_size); | |
3491 | ||
720f5239 | 3492 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 3493 | gcc_assert (alignment_support_scheme); |
ebfd146a IR |
3494 | |
3495 | /* In case the vectorization factor (VF) is bigger than the number | |
3496 | of elements that we can fit in a vectype (nunits), we have to generate | |
3497 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 3498 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
3499 | vect_get_vec_def_for_copy_stmt. */ |
3500 | ||
3501 | /* In case of interleaving (non-unit strided access): | |
3502 | ||
3503 | S1: &base + 2 = x2 | |
3504 | S2: &base = x0 | |
3505 | S3: &base + 1 = x1 | |
3506 | S4: &base + 3 = x3 | |
3507 | ||
3508 | We create vectorized stores starting from base address (the access of the | |
3509 | first stmt in the chain (S2 in the above example), when the last store stmt | |
3510 | of the chain (S4) is reached: | |
3511 | ||
3512 | VS1: &base = vx2 | |
3513 | VS2: &base + vec_size*1 = vx0 | |
3514 | VS3: &base + vec_size*2 = vx1 | |
3515 | VS4: &base + vec_size*3 = vx3 | |
3516 | ||
3517 | Then permutation statements are generated: | |
3518 | ||
3519 | VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 > | |
3520 | VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 > | |
3521 | ... | |
b8698a0f | 3522 | |
ebfd146a IR |
3523 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
3524 | (the order of the data-refs in the output of vect_permute_store_chain | |
3525 | corresponds to the order of scalar stmts in the interleaving chain - see | |
3526 | the documentation of vect_permute_store_chain()). | |
3527 | ||
3528 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 3529 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 3530 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 3531 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
3532 | */ |
3533 | ||
3534 | prev_stmt_info = NULL; | |
3535 | for (j = 0; j < ncopies; j++) | |
3536 | { | |
3537 | gimple new_stmt; | |
3538 | gimple ptr_incr; | |
3539 | ||
3540 | if (j == 0) | |
3541 | { | |
3542 | if (slp) | |
3543 | { | |
3544 | /* Get vectorized arguments for SLP_NODE. */ | |
9dc3f7de IR |
3545 | vect_get_slp_defs (NULL_TREE, NULL_TREE, slp_node, &vec_oprnds, |
3546 | NULL, -1); | |
ebfd146a IR |
3547 | |
3548 | vec_oprnd = VEC_index (tree, vec_oprnds, 0); | |
3549 | } | |
3550 | else | |
3551 | { | |
b8698a0f L |
3552 | /* For interleaved stores we collect vectorized defs for all the |
3553 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
3554 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
3555 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
3556 | ||
3557 | If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and | |
3558 | OPRNDS are of size 1. */ | |
b8698a0f | 3559 | next_stmt = first_stmt; |
ebfd146a IR |
3560 | for (i = 0; i < group_size; i++) |
3561 | { | |
b8698a0f L |
3562 | /* Since gaps are not supported for interleaved stores, |
3563 | GROUP_SIZE is the exact number of stmts in the chain. | |
3564 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
3565 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
3566 | iteration of the loop will be executed. */ |
3567 | gcc_assert (next_stmt | |
3568 | && gimple_assign_single_p (next_stmt)); | |
3569 | op = gimple_assign_rhs1 (next_stmt); | |
3570 | ||
b8698a0f | 3571 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt, |
ebfd146a | 3572 | NULL); |
b8698a0f L |
3573 | VEC_quick_push(tree, dr_chain, vec_oprnd); |
3574 | VEC_quick_push(tree, oprnds, vec_oprnd); | |
ebfd146a IR |
3575 | next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); |
3576 | } | |
3577 | } | |
3578 | ||
3579 | /* We should have catched mismatched types earlier. */ | |
3580 | gcc_assert (useless_type_conversion_p (vectype, | |
3581 | TREE_TYPE (vec_oprnd))); | |
b8698a0f L |
3582 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, NULL, NULL_TREE, |
3583 | &dummy, &ptr_incr, false, | |
5006671f | 3584 | &inv_p); |
a70d6342 | 3585 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 3586 | } |
b8698a0f | 3587 | else |
ebfd146a | 3588 | { |
b8698a0f L |
3589 | /* For interleaved stores we created vectorized defs for all the |
3590 | defs stored in OPRNDS in the previous iteration (previous copy). | |
3591 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
3592 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
3593 | next copy. | |
3594 | If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and | |
3595 | OPRNDS are of size 1. */ | |
3596 | for (i = 0; i < group_size; i++) | |
3597 | { | |
3598 | op = VEC_index (tree, oprnds, i); | |
b8698a0f | 3599 | vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, |
a70d6342 | 3600 | &dt); |
b8698a0f | 3601 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
ebfd146a IR |
3602 | VEC_replace(tree, dr_chain, i, vec_oprnd); |
3603 | VEC_replace(tree, oprnds, i, vec_oprnd); | |
3604 | } | |
b8698a0f | 3605 | dataref_ptr = |
ebfd146a IR |
3606 | bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE); |
3607 | } | |
3608 | ||
3609 | if (strided_store) | |
3610 | { | |
b8698a0f | 3611 | result_chain = VEC_alloc (tree, heap, group_size); |
ebfd146a IR |
3612 | /* Permute. */ |
3613 | if (!vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
3614 | &result_chain)) | |
3615 | return false; | |
3616 | } | |
3617 | ||
3618 | next_stmt = first_stmt; | |
3619 | for (i = 0; i < vec_num; i++) | |
3620 | { | |
be1ac4ec RG |
3621 | struct ptr_info_def *pi; |
3622 | ||
ebfd146a IR |
3623 | if (i > 0) |
3624 | /* Bump the vector pointer. */ | |
3625 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
3626 | NULL_TREE); | |
3627 | ||
3628 | if (slp) | |
3629 | vec_oprnd = VEC_index (tree, vec_oprnds, i); | |
3630 | else if (strided_store) | |
b8698a0f | 3631 | /* For strided stores vectorized defs are interleaved in |
ebfd146a IR |
3632 | vect_permute_store_chain(). */ |
3633 | vec_oprnd = VEC_index (tree, result_chain, i); | |
3634 | ||
be1ac4ec RG |
3635 | data_ref = build2 (MEM_REF, TREE_TYPE (vec_oprnd), dataref_ptr, |
3636 | build_int_cst (reference_alias_ptr_type | |
3637 | (DR_REF (first_dr)), 0)); | |
3638 | pi = get_ptr_info (dataref_ptr); | |
3639 | pi->align = TYPE_ALIGN_UNIT (vectype); | |
8f439681 | 3640 | if (aligned_access_p (first_dr)) |
be1ac4ec RG |
3641 | pi->misalign = 0; |
3642 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
3643 | { | |
3644 | TREE_TYPE (data_ref) | |
3645 | = build_aligned_type (TREE_TYPE (data_ref), | |
3646 | TYPE_ALIGN (TREE_TYPE (vectype))); | |
3647 | pi->align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype)); | |
3648 | pi->misalign = 0; | |
3649 | } | |
3650 | else | |
3651 | { | |
3652 | TREE_TYPE (data_ref) | |
3653 | = build_aligned_type (TREE_TYPE (data_ref), | |
3654 | TYPE_ALIGN (TREE_TYPE (vectype))); | |
3655 | pi->misalign = DR_MISALIGNMENT (first_dr); | |
3656 | } | |
8f439681 | 3657 | |
9dc3f7de | 3658 | /* Arguments are ready. Create the new vector stmt. */ |
ebfd146a IR |
3659 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); |
3660 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3661 | mark_symbols_for_renaming (new_stmt); | |
3662 | ||
3663 | if (slp) | |
3664 | continue; | |
b8698a0f | 3665 | |
ebfd146a IR |
3666 | if (j == 0) |
3667 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3668 | else | |
3669 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3670 | ||
3671 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3672 | next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); | |
3673 | if (!next_stmt) | |
3674 | break; | |
3675 | } | |
3676 | } | |
3677 | ||
b8698a0f L |
3678 | VEC_free (tree, heap, dr_chain); |
3679 | VEC_free (tree, heap, oprnds); | |
ebfd146a | 3680 | if (result_chain) |
b8698a0f | 3681 | VEC_free (tree, heap, result_chain); |
ff802fa1 IR |
3682 | if (vec_oprnds) |
3683 | VEC_free (tree, heap, vec_oprnds); | |
ebfd146a IR |
3684 | |
3685 | return true; | |
3686 | } | |
3687 | ||
a1e53f3f L |
3688 | /* Given a vector type VECTYPE returns a builtin DECL to be used |
3689 | for vector permutation and stores a mask into *MASK that implements | |
3690 | reversal of the vector elements. If that is impossible to do | |
3691 | returns NULL (and *MASK is unchanged). */ | |
3692 | ||
3693 | static tree | |
3694 | perm_mask_for_reverse (tree vectype, tree *mask) | |
3695 | { | |
3696 | tree builtin_decl; | |
3697 | tree mask_element_type, mask_type; | |
3698 | tree mask_vec = NULL; | |
3699 | int i; | |
3700 | int nunits; | |
3701 | if (!targetm.vectorize.builtin_vec_perm) | |
3702 | return NULL; | |
3703 | ||
3704 | builtin_decl = targetm.vectorize.builtin_vec_perm (vectype, | |
3705 | &mask_element_type); | |
3706 | if (!builtin_decl || !mask_element_type) | |
3707 | return NULL; | |
3708 | ||
3709 | mask_type = get_vectype_for_scalar_type (mask_element_type); | |
3710 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
bb67d9c7 RG |
3711 | if (!mask_type |
3712 | || TYPE_VECTOR_SUBPARTS (vectype) != TYPE_VECTOR_SUBPARTS (mask_type)) | |
a1e53f3f L |
3713 | return NULL; |
3714 | ||
3715 | for (i = 0; i < nunits; i++) | |
3716 | mask_vec = tree_cons (NULL, build_int_cst (mask_element_type, i), mask_vec); | |
3717 | mask_vec = build_vector (mask_type, mask_vec); | |
3718 | ||
3719 | if (!targetm.vectorize.builtin_vec_perm_ok (vectype, mask_vec)) | |
3720 | return NULL; | |
3721 | if (mask) | |
3722 | *mask = mask_vec; | |
3723 | return builtin_decl; | |
3724 | } | |
3725 | ||
3726 | /* Given a vector variable X, that was generated for the scalar LHS of | |
3727 | STMT, generate instructions to reverse the vector elements of X, | |
3728 | insert them a *GSI and return the permuted vector variable. */ | |
3729 | ||
3730 | static tree | |
3731 | reverse_vec_elements (tree x, gimple stmt, gimple_stmt_iterator *gsi) | |
3732 | { | |
3733 | tree vectype = TREE_TYPE (x); | |
3734 | tree mask_vec, builtin_decl; | |
3735 | tree perm_dest, data_ref; | |
3736 | gimple perm_stmt; | |
3737 | ||
3738 | builtin_decl = perm_mask_for_reverse (vectype, &mask_vec); | |
3739 | ||
3740 | perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype); | |
3741 | ||
3742 | /* Generate the permute statement. */ | |
3743 | perm_stmt = gimple_build_call (builtin_decl, 3, x, x, mask_vec); | |
2a2651b7 RG |
3744 | if (!useless_type_conversion_p (vectype, |
3745 | TREE_TYPE (TREE_TYPE (builtin_decl)))) | |
3746 | { | |
3747 | tree tem = create_tmp_reg (TREE_TYPE (TREE_TYPE (builtin_decl)), NULL); | |
3748 | tem = make_ssa_name (tem, perm_stmt); | |
3749 | gimple_call_set_lhs (perm_stmt, tem); | |
3750 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); | |
3751 | perm_stmt = gimple_build_assign (NULL_TREE, | |
3752 | build1 (VIEW_CONVERT_EXPR, | |
3753 | vectype, tem)); | |
3754 | } | |
a1e53f3f | 3755 | data_ref = make_ssa_name (perm_dest, perm_stmt); |
2a2651b7 | 3756 | gimple_set_lhs (perm_stmt, data_ref); |
a1e53f3f L |
3757 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
3758 | ||
3759 | return data_ref; | |
3760 | } | |
3761 | ||
ebfd146a IR |
3762 | /* vectorizable_load. |
3763 | ||
b8698a0f L |
3764 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
3765 | can be vectorized. | |
3766 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
3767 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3768 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3769 | ||
3770 | static bool | |
3771 | vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
3772 | slp_tree slp_node, slp_instance slp_node_instance) | |
3773 | { | |
3774 | tree scalar_dest; | |
3775 | tree vec_dest = NULL; | |
3776 | tree data_ref = NULL; | |
3777 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 3778 | stmt_vec_info prev_stmt_info; |
ebfd146a | 3779 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 3780 | struct loop *loop = NULL; |
ebfd146a | 3781 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 3782 | bool nested_in_vect_loop = false; |
ebfd146a IR |
3783 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; |
3784 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3785 | tree new_temp; | |
947131ba | 3786 | enum machine_mode mode; |
ebfd146a IR |
3787 | gimple new_stmt = NULL; |
3788 | tree dummy; | |
3789 | enum dr_alignment_support alignment_support_scheme; | |
3790 | tree dataref_ptr = NULL_TREE; | |
3791 | gimple ptr_incr; | |
3792 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
3793 | int ncopies; | |
3794 | int i, j, group_size; | |
3795 | tree msq = NULL_TREE, lsq; | |
3796 | tree offset = NULL_TREE; | |
3797 | tree realignment_token = NULL_TREE; | |
3798 | gimple phi = NULL; | |
3799 | VEC(tree,heap) *dr_chain = NULL; | |
3800 | bool strided_load = false; | |
3801 | gimple first_stmt; | |
3802 | tree scalar_type; | |
3803 | bool inv_p; | |
a1e53f3f | 3804 | bool negative; |
ebfd146a IR |
3805 | bool compute_in_loop = false; |
3806 | struct loop *at_loop; | |
3807 | int vec_num; | |
3808 | bool slp = (slp_node != NULL); | |
3809 | bool slp_perm = false; | |
3810 | enum tree_code code; | |
a70d6342 IR |
3811 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3812 | int vf; | |
3813 | ||
3814 | if (loop_vinfo) | |
3815 | { | |
3816 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
3817 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
3818 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3819 | } | |
3820 | else | |
3533e503 | 3821 | vf = 1; |
ebfd146a IR |
3822 | |
3823 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 3824 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a IR |
3825 | case of SLP. */ |
3826 | if (slp) | |
3827 | ncopies = 1; | |
3828 | else | |
3829 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3830 | ||
3831 | gcc_assert (ncopies >= 1); | |
3832 | ||
3833 | /* FORNOW. This restriction should be relaxed. */ | |
3834 | if (nested_in_vect_loop && ncopies > 1) | |
3835 | { | |
3836 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3837 | fprintf (vect_dump, "multiple types in nested loop."); | |
3838 | return false; | |
3839 | } | |
3840 | ||
a70d6342 | 3841 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3842 | return false; |
3843 | ||
8644a673 | 3844 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3845 | return false; |
3846 | ||
3847 | /* Is vectorizable load? */ | |
3848 | if (!is_gimple_assign (stmt)) | |
3849 | return false; | |
3850 | ||
3851 | scalar_dest = gimple_assign_lhs (stmt); | |
3852 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
3853 | return false; | |
3854 | ||
3855 | code = gimple_assign_rhs_code (stmt); | |
3856 | if (code != ARRAY_REF | |
3857 | && code != INDIRECT_REF | |
e9dbe7bb IR |
3858 | && code != COMPONENT_REF |
3859 | && code != IMAGPART_EXPR | |
70f34814 RG |
3860 | && code != REALPART_EXPR |
3861 | && code != MEM_REF) | |
ebfd146a IR |
3862 | return false; |
3863 | ||
3864 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
3865 | return false; | |
3866 | ||
a1e53f3f L |
3867 | negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0; |
3868 | if (negative && ncopies > 1) | |
3869 | { | |
3870 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3871 | fprintf (vect_dump, "multiple types with negative step."); | |
3872 | return false; | |
3873 | } | |
3874 | ||
ebfd146a | 3875 | scalar_type = TREE_TYPE (DR_REF (dr)); |
947131ba | 3876 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
3877 | |
3878 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
3879 | (e.g. - data copies). */ | |
947131ba | 3880 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a IR |
3881 | { |
3882 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3883 | fprintf (vect_dump, "Aligned load, but unsupported type."); | |
3884 | return false; | |
3885 | } | |
3886 | ||
3887 | /* The vector component type needs to be trivially convertible to the | |
3888 | scalar lhs. This should always be the case. */ | |
3889 | if (!useless_type_conversion_p (TREE_TYPE (scalar_dest), TREE_TYPE (vectype))) | |
b8698a0f | 3890 | { |
ebfd146a IR |
3891 | if (vect_print_dump_info (REPORT_DETAILS)) |
3892 | fprintf (vect_dump, "??? operands of different types"); | |
3893 | return false; | |
3894 | } | |
3895 | ||
3896 | /* Check if the load is a part of an interleaving chain. */ | |
3897 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) | |
3898 | { | |
3899 | strided_load = true; | |
3900 | /* FORNOW */ | |
3901 | gcc_assert (! nested_in_vect_loop); | |
3902 | ||
3903 | /* Check if interleaving is supported. */ | |
3904 | if (!vect_strided_load_supported (vectype) | |
3905 | && !PURE_SLP_STMT (stmt_info) && !slp) | |
3906 | return false; | |
3907 | } | |
3908 | ||
a1e53f3f L |
3909 | if (negative) |
3910 | { | |
3911 | gcc_assert (!strided_load); | |
3912 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
3913 | if (alignment_support_scheme != dr_aligned | |
3914 | && alignment_support_scheme != dr_unaligned_supported) | |
3915 | { | |
3916 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3917 | fprintf (vect_dump, "negative step but alignment required."); | |
3918 | return false; | |
3919 | } | |
3920 | if (!perm_mask_for_reverse (vectype, NULL)) | |
3921 | { | |
3922 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3923 | fprintf (vect_dump, "negative step and reversing not supported."); | |
3924 | return false; | |
3925 | } | |
3926 | } | |
3927 | ||
ebfd146a IR |
3928 | if (!vec_stmt) /* transformation not required. */ |
3929 | { | |
3930 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
3931 | vect_model_load_cost (stmt_info, ncopies, NULL); | |
3932 | return true; | |
3933 | } | |
3934 | ||
3935 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3936 | fprintf (vect_dump, "transform load."); | |
3937 | ||
3938 | /** Transform. **/ | |
3939 | ||
3940 | if (strided_load) | |
3941 | { | |
3942 | first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
3943 | /* Check if the chain of loads is already vectorized. */ | |
3944 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt))) | |
3945 | { | |
3946 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
3947 | return true; | |
3948 | } | |
3949 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
3950 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
3951 | ||
3952 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
3953 | if (slp) | |
3954 | { | |
3955 | strided_load = false; | |
3956 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
a70d6342 IR |
3957 | if (SLP_INSTANCE_LOAD_PERMUTATION (slp_node_instance)) |
3958 | slp_perm = true; | |
3959 | } | |
ebfd146a IR |
3960 | else |
3961 | vec_num = group_size; | |
3962 | ||
3963 | dr_chain = VEC_alloc (tree, heap, vec_num); | |
3964 | } | |
3965 | else | |
3966 | { | |
3967 | first_stmt = stmt; | |
3968 | first_dr = dr; | |
3969 | group_size = vec_num = 1; | |
3970 | } | |
3971 | ||
720f5239 | 3972 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a IR |
3973 | gcc_assert (alignment_support_scheme); |
3974 | ||
3975 | /* In case the vectorization factor (VF) is bigger than the number | |
3976 | of elements that we can fit in a vectype (nunits), we have to generate | |
3977 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 3978 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 3979 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 3980 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 3981 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
3982 | stmts that use the defs of the current stmt. The example below |
3983 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
3984 | need to create 4 vectorized stmts): | |
ebfd146a IR |
3985 | |
3986 | before vectorization: | |
3987 | RELATED_STMT VEC_STMT | |
3988 | S1: x = memref - - | |
3989 | S2: z = x + 1 - - | |
3990 | ||
3991 | step 1: vectorize stmt S1: | |
3992 | We first create the vector stmt VS1_0, and, as usual, record a | |
3993 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
3994 | Next, we create the vector stmt VS1_1, and record a pointer to | |
3995 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 3996 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
3997 | stmts and pointers: |
3998 | RELATED_STMT VEC_STMT | |
3999 | VS1_0: vx0 = memref0 VS1_1 - | |
4000 | VS1_1: vx1 = memref1 VS1_2 - | |
4001 | VS1_2: vx2 = memref2 VS1_3 - | |
4002 | VS1_3: vx3 = memref3 - - | |
4003 | S1: x = load - VS1_0 | |
4004 | S2: z = x + 1 - - | |
4005 | ||
b8698a0f L |
4006 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
4007 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
4008 | stmt S2. */ |
4009 | ||
4010 | /* In case of interleaving (non-unit strided access): | |
4011 | ||
4012 | S1: x2 = &base + 2 | |
4013 | S2: x0 = &base | |
4014 | S3: x1 = &base + 1 | |
4015 | S4: x3 = &base + 3 | |
4016 | ||
b8698a0f | 4017 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
4018 | starting from the access of the first stmt of the chain: |
4019 | ||
4020 | VS1: vx0 = &base | |
4021 | VS2: vx1 = &base + vec_size*1 | |
4022 | VS3: vx3 = &base + vec_size*2 | |
4023 | VS4: vx4 = &base + vec_size*3 | |
4024 | ||
4025 | Then permutation statements are generated: | |
4026 | ||
4027 | VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 > | |
4028 | VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 > | |
4029 | ... | |
4030 | ||
4031 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
4032 | (the order of the data-refs in the output of vect_permute_load_chain | |
4033 | corresponds to the order of scalar stmts in the interleaving chain - see | |
4034 | the documentation of vect_permute_load_chain()). | |
4035 | The generation of permutation stmts and recording them in | |
4036 | STMT_VINFO_VEC_STMT is done in vect_transform_strided_load(). | |
4037 | ||
b8698a0f | 4038 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
4039 | permutation stmts above are created for every copy. The result vector |
4040 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
4041 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
4042 | |
4043 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
4044 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
4045 | we generate the following code: | |
4046 | p = initial_addr; | |
4047 | indx = 0; | |
4048 | loop { | |
4049 | p = p + indx * vectype_size; | |
4050 | vec_dest = *(p); | |
4051 | indx = indx + 1; | |
4052 | } | |
4053 | ||
4054 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 4055 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
4056 | then generate the following code, in which the data in each iteration is |
4057 | obtained by two vector loads, one from the previous iteration, and one | |
4058 | from the current iteration: | |
4059 | p1 = initial_addr; | |
4060 | msq_init = *(floor(p1)) | |
4061 | p2 = initial_addr + VS - 1; | |
4062 | realignment_token = call target_builtin; | |
4063 | indx = 0; | |
4064 | loop { | |
4065 | p2 = p2 + indx * vectype_size | |
4066 | lsq = *(floor(p2)) | |
4067 | vec_dest = realign_load (msq, lsq, realignment_token) | |
4068 | indx = indx + 1; | |
4069 | msq = lsq; | |
4070 | } */ | |
4071 | ||
4072 | /* If the misalignment remains the same throughout the execution of the | |
4073 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 4074 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
4075 | This can only occur when vectorizing memory accesses in the inner-loop |
4076 | nested within an outer-loop that is being vectorized. */ | |
4077 | ||
a70d6342 | 4078 | if (loop && nested_in_vect_loop_p (loop, stmt) |
ebfd146a IR |
4079 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
4080 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) | |
4081 | { | |
4082 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
4083 | compute_in_loop = true; | |
4084 | } | |
4085 | ||
4086 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
4087 | || alignment_support_scheme == dr_explicit_realign) | |
4088 | && !compute_in_loop) | |
4089 | { | |
4090 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
4091 | alignment_support_scheme, NULL_TREE, | |
4092 | &at_loop); | |
4093 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
4094 | { | |
4095 | phi = SSA_NAME_DEF_STMT (msq); | |
4096 | offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
4097 | } | |
4098 | } | |
4099 | else | |
4100 | at_loop = loop; | |
4101 | ||
a1e53f3f L |
4102 | if (negative) |
4103 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
4104 | ||
ebfd146a IR |
4105 | prev_stmt_info = NULL; |
4106 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 4107 | { |
ebfd146a IR |
4108 | /* 1. Create the vector pointer update chain. */ |
4109 | if (j == 0) | |
4110 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, | |
b8698a0f L |
4111 | at_loop, offset, |
4112 | &dummy, &ptr_incr, false, | |
5006671f | 4113 | &inv_p); |
ebfd146a | 4114 | else |
b8698a0f | 4115 | dataref_ptr = |
ebfd146a IR |
4116 | bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE); |
4117 | ||
4118 | for (i = 0; i < vec_num; i++) | |
4119 | { | |
4120 | if (i > 0) | |
4121 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
4122 | NULL_TREE); | |
4123 | ||
4124 | /* 2. Create the vector-load in the loop. */ | |
4125 | switch (alignment_support_scheme) | |
4126 | { | |
4127 | case dr_aligned: | |
ebfd146a IR |
4128 | case dr_unaligned_supported: |
4129 | { | |
be1ac4ec RG |
4130 | struct ptr_info_def *pi; |
4131 | data_ref | |
4132 | = build2 (MEM_REF, vectype, dataref_ptr, | |
4133 | build_int_cst (reference_alias_ptr_type | |
4134 | (DR_REF (first_dr)), 0)); | |
4135 | pi = get_ptr_info (dataref_ptr); | |
4136 | pi->align = TYPE_ALIGN_UNIT (vectype); | |
4137 | if (alignment_support_scheme == dr_aligned) | |
4138 | { | |
4139 | gcc_assert (aligned_access_p (first_dr)); | |
4140 | pi->misalign = 0; | |
4141 | } | |
4142 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
4143 | { | |
4144 | TREE_TYPE (data_ref) | |
4145 | = build_aligned_type (TREE_TYPE (data_ref), | |
4146 | TYPE_ALIGN (TREE_TYPE (vectype))); | |
4147 | pi->align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype)); | |
4148 | pi->misalign = 0; | |
4149 | } | |
4150 | else | |
4151 | { | |
4152 | TREE_TYPE (data_ref) | |
4153 | = build_aligned_type (TREE_TYPE (data_ref), | |
4154 | TYPE_ALIGN (TREE_TYPE (vectype))); | |
4155 | pi->misalign = DR_MISALIGNMENT (first_dr); | |
4156 | } | |
ebfd146a IR |
4157 | break; |
4158 | } | |
4159 | case dr_explicit_realign: | |
4160 | { | |
4161 | tree ptr, bump; | |
4162 | tree vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
4163 | ||
4164 | if (compute_in_loop) | |
4165 | msq = vect_setup_realignment (first_stmt, gsi, | |
4166 | &realignment_token, | |
b8698a0f | 4167 | dr_explicit_realign, |
ebfd146a IR |
4168 | dataref_ptr, NULL); |
4169 | ||
75421dcd RG |
4170 | new_stmt = gimple_build_assign_with_ops |
4171 | (BIT_AND_EXPR, NULL_TREE, dataref_ptr, | |
4172 | build_int_cst | |
4173 | (TREE_TYPE (dataref_ptr), | |
4174 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
4175 | ptr = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt); | |
4176 | gimple_assign_set_lhs (new_stmt, ptr); | |
4177 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
20ede5c6 RG |
4178 | data_ref |
4179 | = build2 (MEM_REF, vectype, ptr, | |
4180 | build_int_cst (reference_alias_ptr_type | |
4181 | (DR_REF (first_dr)), 0)); | |
ebfd146a IR |
4182 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
4183 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
4184 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4185 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5006671f RG |
4186 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); |
4187 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
ebfd146a | 4188 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
4189 | msq = new_temp; |
4190 | ||
4191 | bump = size_binop (MULT_EXPR, vs_minus_1, | |
4192 | TYPE_SIZE_UNIT (scalar_type)); | |
4193 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); | |
75421dcd RG |
4194 | new_stmt = gimple_build_assign_with_ops |
4195 | (BIT_AND_EXPR, NULL_TREE, ptr, | |
4196 | build_int_cst | |
4197 | (TREE_TYPE (ptr), | |
4198 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
4199 | ptr = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt); | |
4200 | gimple_assign_set_lhs (new_stmt, ptr); | |
4201 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
20ede5c6 RG |
4202 | data_ref |
4203 | = build2 (MEM_REF, vectype, ptr, | |
4204 | build_int_cst (reference_alias_ptr_type | |
4205 | (DR_REF (first_dr)), 0)); | |
ebfd146a IR |
4206 | break; |
4207 | } | |
4208 | case dr_explicit_realign_optimized: | |
75421dcd RG |
4209 | new_stmt = gimple_build_assign_with_ops |
4210 | (BIT_AND_EXPR, NULL_TREE, dataref_ptr, | |
4211 | build_int_cst | |
4212 | (TREE_TYPE (dataref_ptr), | |
4213 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
4214 | new_temp = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt); | |
4215 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4216 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
20ede5c6 RG |
4217 | data_ref |
4218 | = build2 (MEM_REF, vectype, new_temp, | |
4219 | build_int_cst (reference_alias_ptr_type | |
4220 | (DR_REF (first_dr)), 0)); | |
ebfd146a IR |
4221 | break; |
4222 | default: | |
4223 | gcc_unreachable (); | |
4224 | } | |
4225 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4226 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
4227 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4228 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4229 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4230 | mark_symbols_for_renaming (new_stmt); | |
4231 | ||
ff802fa1 | 4232 | /* 3. Handle explicit realignment if necessary/supported. Create in |
ebfd146a IR |
4233 | loop: vec_dest = realign_load (msq, lsq, realignment_token) */ |
4234 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
4235 | || alignment_support_scheme == dr_explicit_realign) | |
4236 | { | |
4237 | tree tmp; | |
4238 | ||
4239 | lsq = gimple_assign_lhs (new_stmt); | |
4240 | if (!realignment_token) | |
4241 | realignment_token = dataref_ptr; | |
4242 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4243 | tmp = build3 (REALIGN_LOAD_EXPR, vectype, msq, lsq, | |
4244 | realignment_token); | |
4245 | new_stmt = gimple_build_assign (vec_dest, tmp); | |
4246 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4247 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4248 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4249 | ||
4250 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
4251 | { | |
4252 | gcc_assert (phi); | |
4253 | if (i == vec_num - 1 && j == ncopies - 1) | |
f5045c96 AM |
4254 | add_phi_arg (phi, lsq, loop_latch_edge (containing_loop), |
4255 | UNKNOWN_LOCATION); | |
ebfd146a IR |
4256 | msq = lsq; |
4257 | } | |
4258 | } | |
4259 | ||
4260 | /* 4. Handle invariant-load. */ | |
a70d6342 | 4261 | if (inv_p && !bb_vinfo) |
ebfd146a IR |
4262 | { |
4263 | gcc_assert (!strided_load); | |
4264 | gcc_assert (nested_in_vect_loop_p (loop, stmt)); | |
4265 | if (j == 0) | |
4266 | { | |
4267 | int k; | |
4268 | tree t = NULL_TREE; | |
4269 | tree vec_inv, bitpos, bitsize = TYPE_SIZE (scalar_type); | |
4270 | ||
4271 | /* CHECKME: bitpos depends on endianess? */ | |
4272 | bitpos = bitsize_zero_node; | |
b8698a0f | 4273 | vec_inv = build3 (BIT_FIELD_REF, scalar_type, new_temp, |
ebfd146a | 4274 | bitsize, bitpos); |
b8698a0f | 4275 | vec_dest = |
ebfd146a IR |
4276 | vect_create_destination_var (scalar_dest, NULL_TREE); |
4277 | new_stmt = gimple_build_assign (vec_dest, vec_inv); | |
4278 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4279 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4280 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4281 | ||
4282 | for (k = nunits - 1; k >= 0; --k) | |
4283 | t = tree_cons (NULL_TREE, new_temp, t); | |
4284 | /* FIXME: use build_constructor directly. */ | |
4285 | vec_inv = build_constructor_from_list (vectype, t); | |
4286 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
4287 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
4288 | } | |
4289 | else | |
4290 | gcc_unreachable (); /* FORNOW. */ | |
4291 | } | |
4292 | ||
a1e53f3f L |
4293 | if (negative) |
4294 | { | |
4295 | new_temp = reverse_vec_elements (new_temp, stmt, gsi); | |
4296 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
4297 | } | |
4298 | ||
ebfd146a IR |
4299 | /* Collect vector loads and later create their permutation in |
4300 | vect_transform_strided_load (). */ | |
4301 | if (strided_load || slp_perm) | |
4302 | VEC_quick_push (tree, dr_chain, new_temp); | |
4303 | ||
4304 | /* Store vector loads in the corresponding SLP_NODE. */ | |
4305 | if (slp && !slp_perm) | |
4306 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
4307 | } | |
4308 | ||
4309 | if (slp && !slp_perm) | |
4310 | continue; | |
4311 | ||
4312 | if (slp_perm) | |
4313 | { | |
a70d6342 | 4314 | if (!vect_transform_slp_perm_load (stmt, dr_chain, gsi, vf, |
ebfd146a IR |
4315 | slp_node_instance, false)) |
4316 | { | |
4317 | VEC_free (tree, heap, dr_chain); | |
4318 | return false; | |
4319 | } | |
4320 | } | |
4321 | else | |
4322 | { | |
4323 | if (strided_load) | |
4324 | { | |
4325 | if (!vect_transform_strided_load (stmt, dr_chain, group_size, gsi)) | |
b8698a0f | 4326 | return false; |
ebfd146a IR |
4327 | |
4328 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4329 | VEC_free (tree, heap, dr_chain); | |
4330 | dr_chain = VEC_alloc (tree, heap, group_size); | |
4331 | } | |
4332 | else | |
4333 | { | |
4334 | if (j == 0) | |
4335 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4336 | else | |
4337 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4338 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4339 | } | |
4340 | } | |
4341 | } | |
4342 | ||
4343 | if (dr_chain) | |
4344 | VEC_free (tree, heap, dr_chain); | |
4345 | ||
4346 | return true; | |
4347 | } | |
4348 | ||
4349 | /* Function vect_is_simple_cond. | |
b8698a0f | 4350 | |
ebfd146a IR |
4351 | Input: |
4352 | LOOP - the loop that is being vectorized. | |
4353 | COND - Condition that is checked for simple use. | |
4354 | ||
4355 | Returns whether a COND can be vectorized. Checks whether | |
4356 | condition operands are supportable using vec_is_simple_use. */ | |
4357 | ||
4358 | static bool | |
4359 | vect_is_simple_cond (tree cond, loop_vec_info loop_vinfo) | |
4360 | { | |
4361 | tree lhs, rhs; | |
4362 | tree def; | |
4363 | enum vect_def_type dt; | |
4364 | ||
4365 | if (!COMPARISON_CLASS_P (cond)) | |
4366 | return false; | |
4367 | ||
4368 | lhs = TREE_OPERAND (cond, 0); | |
4369 | rhs = TREE_OPERAND (cond, 1); | |
4370 | ||
4371 | if (TREE_CODE (lhs) == SSA_NAME) | |
4372 | { | |
4373 | gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); | |
b8698a0f | 4374 | if (!vect_is_simple_use (lhs, loop_vinfo, NULL, &lhs_def_stmt, &def, |
a70d6342 | 4375 | &dt)) |
ebfd146a IR |
4376 | return false; |
4377 | } | |
4378 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
4379 | && TREE_CODE (lhs) != FIXED_CST) | |
4380 | return false; | |
4381 | ||
4382 | if (TREE_CODE (rhs) == SSA_NAME) | |
4383 | { | |
4384 | gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); | |
b8698a0f | 4385 | if (!vect_is_simple_use (rhs, loop_vinfo, NULL, &rhs_def_stmt, &def, |
a70d6342 | 4386 | &dt)) |
ebfd146a IR |
4387 | return false; |
4388 | } | |
4389 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST | |
4390 | && TREE_CODE (rhs) != FIXED_CST) | |
4391 | return false; | |
4392 | ||
4393 | return true; | |
4394 | } | |
4395 | ||
4396 | /* vectorizable_condition. | |
4397 | ||
b8698a0f L |
4398 | Check if STMT is conditional modify expression that can be vectorized. |
4399 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4400 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
4401 | at GSI. |
4402 | ||
4403 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
4404 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
4405 | else caluse if it is 2). | |
ebfd146a IR |
4406 | |
4407 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4408 | ||
4bbe8262 | 4409 | bool |
ebfd146a | 4410 | vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi, |
4bbe8262 | 4411 | gimple *vec_stmt, tree reduc_def, int reduc_index) |
ebfd146a IR |
4412 | { |
4413 | tree scalar_dest = NULL_TREE; | |
4414 | tree vec_dest = NULL_TREE; | |
4415 | tree op = NULL_TREE; | |
4416 | tree cond_expr, then_clause, else_clause; | |
4417 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4418 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
ff802fa1 IR |
4419 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
4420 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
ebfd146a IR |
4421 | tree vec_compare, vec_cond_expr; |
4422 | tree new_temp; | |
4423 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4424 | enum machine_mode vec_mode; | |
4425 | tree def; | |
a855b1b1 | 4426 | enum vect_def_type dt, dts[4]; |
ebfd146a IR |
4427 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
4428 | int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4429 | enum tree_code code; | |
a855b1b1 MM |
4430 | stmt_vec_info prev_stmt_info = NULL; |
4431 | int j; | |
ebfd146a | 4432 | |
a70d6342 IR |
4433 | /* FORNOW: unsupported in basic block SLP. */ |
4434 | gcc_assert (loop_vinfo); | |
b8698a0f | 4435 | |
ebfd146a | 4436 | gcc_assert (ncopies >= 1); |
a855b1b1 | 4437 | if (reduc_index && ncopies > 1) |
ebfd146a IR |
4438 | return false; /* FORNOW */ |
4439 | ||
4440 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
4441 | return false; | |
4442 | ||
4bbe8262 IR |
4443 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4444 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
4445 | && reduc_def)) | |
ebfd146a IR |
4446 | return false; |
4447 | ||
4448 | /* FORNOW: SLP not supported. */ | |
4449 | if (STMT_SLP_TYPE (stmt_info)) | |
4450 | return false; | |
4451 | ||
4452 | /* FORNOW: not yet supported. */ | |
b8698a0f | 4453 | if (STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a IR |
4454 | { |
4455 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4456 | fprintf (vect_dump, "value used after loop."); | |
4457 | return false; | |
4458 | } | |
4459 | ||
4460 | /* Is vectorizable conditional operation? */ | |
4461 | if (!is_gimple_assign (stmt)) | |
4462 | return false; | |
4463 | ||
4464 | code = gimple_assign_rhs_code (stmt); | |
4465 | ||
4466 | if (code != COND_EXPR) | |
4467 | return false; | |
4468 | ||
4469 | gcc_assert (gimple_assign_single_p (stmt)); | |
4470 | op = gimple_assign_rhs1 (stmt); | |
4471 | cond_expr = TREE_OPERAND (op, 0); | |
4472 | then_clause = TREE_OPERAND (op, 1); | |
4473 | else_clause = TREE_OPERAND (op, 2); | |
4474 | ||
4475 | if (!vect_is_simple_cond (cond_expr, loop_vinfo)) | |
4476 | return false; | |
4477 | ||
4478 | /* We do not handle two different vector types for the condition | |
4479 | and the values. */ | |
8533c9d8 SP |
4480 | if (!types_compatible_p (TREE_TYPE (TREE_OPERAND (cond_expr, 0)), |
4481 | TREE_TYPE (vectype))) | |
ebfd146a IR |
4482 | return false; |
4483 | ||
4484 | if (TREE_CODE (then_clause) == SSA_NAME) | |
4485 | { | |
4486 | gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause); | |
b8698a0f | 4487 | if (!vect_is_simple_use (then_clause, loop_vinfo, NULL, |
ebfd146a IR |
4488 | &then_def_stmt, &def, &dt)) |
4489 | return false; | |
4490 | } | |
b8698a0f | 4491 | else if (TREE_CODE (then_clause) != INTEGER_CST |
ebfd146a IR |
4492 | && TREE_CODE (then_clause) != REAL_CST |
4493 | && TREE_CODE (then_clause) != FIXED_CST) | |
4494 | return false; | |
4495 | ||
4496 | if (TREE_CODE (else_clause) == SSA_NAME) | |
4497 | { | |
4498 | gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause); | |
a70d6342 | 4499 | if (!vect_is_simple_use (else_clause, loop_vinfo, NULL, |
ebfd146a IR |
4500 | &else_def_stmt, &def, &dt)) |
4501 | return false; | |
4502 | } | |
b8698a0f | 4503 | else if (TREE_CODE (else_clause) != INTEGER_CST |
ebfd146a IR |
4504 | && TREE_CODE (else_clause) != REAL_CST |
4505 | && TREE_CODE (else_clause) != FIXED_CST) | |
4506 | return false; | |
4507 | ||
4508 | ||
4509 | vec_mode = TYPE_MODE (vectype); | |
4510 | ||
b8698a0f | 4511 | if (!vec_stmt) |
ebfd146a IR |
4512 | { |
4513 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
8e7aa1f9 | 4514 | return expand_vec_cond_expr_p (TREE_TYPE (op), vec_mode); |
ebfd146a IR |
4515 | } |
4516 | ||
4517 | /* Transform */ | |
4518 | ||
4519 | /* Handle def. */ | |
4520 | scalar_dest = gimple_assign_lhs (stmt); | |
4521 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4522 | ||
4523 | /* Handle cond expr. */ | |
a855b1b1 MM |
4524 | for (j = 0; j < ncopies; j++) |
4525 | { | |
4526 | gimple new_stmt; | |
4527 | if (j == 0) | |
4528 | { | |
4529 | gimple gtemp; | |
4530 | vec_cond_lhs = | |
4531 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), | |
4532 | stmt, NULL); | |
4533 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), loop_vinfo, | |
4534 | NULL, >emp, &def, &dts[0]); | |
4535 | vec_cond_rhs = | |
4536 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
4537 | stmt, NULL); | |
4538 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), loop_vinfo, | |
4539 | NULL, >emp, &def, &dts[1]); | |
4540 | if (reduc_index == 1) | |
4541 | vec_then_clause = reduc_def; | |
4542 | else | |
4543 | { | |
4544 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
4545 | stmt, NULL); | |
4546 | vect_is_simple_use (then_clause, loop_vinfo, | |
4547 | NULL, >emp, &def, &dts[2]); | |
4548 | } | |
4549 | if (reduc_index == 2) | |
4550 | vec_else_clause = reduc_def; | |
4551 | else | |
4552 | { | |
4553 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
4554 | stmt, NULL); | |
4555 | vect_is_simple_use (else_clause, loop_vinfo, | |
4556 | NULL, >emp, &def, &dts[3]); | |
4557 | } | |
4558 | } | |
4559 | else | |
4560 | { | |
4561 | vec_cond_lhs = vect_get_vec_def_for_stmt_copy (dts[0], vec_cond_lhs); | |
4562 | vec_cond_rhs = vect_get_vec_def_for_stmt_copy (dts[1], vec_cond_rhs); | |
4563 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], | |
4564 | vec_then_clause); | |
4565 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], | |
4566 | vec_else_clause); | |
4567 | } | |
4568 | ||
9dc3f7de | 4569 | /* Arguments are ready. Create the new vector stmt. */ |
a855b1b1 MM |
4570 | vec_compare = build2 (TREE_CODE (cond_expr), vectype, |
4571 | vec_cond_lhs, vec_cond_rhs); | |
4572 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, | |
4573 | vec_compare, vec_then_clause, vec_else_clause); | |
4574 | ||
4575 | new_stmt = gimple_build_assign (vec_dest, vec_cond_expr); | |
4576 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4577 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4578 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4579 | if (j == 0) | |
4580 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4581 | else | |
4582 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4583 | ||
4584 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4585 | } | |
b8698a0f | 4586 | |
ebfd146a IR |
4587 | return true; |
4588 | } | |
4589 | ||
4590 | ||
8644a673 | 4591 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
4592 | |
4593 | bool | |
a70d6342 | 4594 | vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 4595 | { |
8644a673 | 4596 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 4597 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 4598 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 4599 | bool ok; |
a70d6342 | 4600 | tree scalar_type, vectype; |
ebfd146a IR |
4601 | |
4602 | if (vect_print_dump_info (REPORT_DETAILS)) | |
ebfd146a | 4603 | { |
8644a673 IR |
4604 | fprintf (vect_dump, "==> examining statement: "); |
4605 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4606 | } | |
ebfd146a | 4607 | |
1825a1f3 | 4608 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 4609 | { |
1825a1f3 IR |
4610 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
4611 | fprintf (vect_dump, "not vectorized: stmt has volatile operands"); | |
4612 | ||
4613 | return false; | |
4614 | } | |
b8698a0f L |
4615 | |
4616 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
4617 | to include: |
4618 | - the COND_EXPR which is the loop exit condition | |
4619 | - any LABEL_EXPRs in the loop | |
b8698a0f | 4620 | - computations that are used only for array indexing or loop control. |
8644a673 IR |
4621 | In basic blocks we only analyze statements that are a part of some SLP |
4622 | instance, therefore, all the statements are relevant. */ | |
ebfd146a | 4623 | |
b8698a0f | 4624 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 4625 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a IR |
4626 | { |
4627 | if (vect_print_dump_info (REPORT_DETAILS)) | |
8644a673 | 4628 | fprintf (vect_dump, "irrelevant."); |
ebfd146a | 4629 | |
8644a673 IR |
4630 | return true; |
4631 | } | |
ebfd146a | 4632 | |
8644a673 IR |
4633 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
4634 | { | |
4635 | case vect_internal_def: | |
4636 | break; | |
ebfd146a | 4637 | |
8644a673 | 4638 | case vect_reduction_def: |
7c5222ff | 4639 | case vect_nested_cycle: |
a70d6342 | 4640 | gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer |
8644a673 | 4641 | || relevance == vect_used_in_outer_by_reduction |
a70d6342 | 4642 | || relevance == vect_unused_in_scope)); |
8644a673 IR |
4643 | break; |
4644 | ||
4645 | case vect_induction_def: | |
4646 | case vect_constant_def: | |
4647 | case vect_external_def: | |
4648 | case vect_unknown_def_type: | |
4649 | default: | |
4650 | gcc_unreachable (); | |
4651 | } | |
ebfd146a | 4652 | |
a70d6342 IR |
4653 | if (bb_vinfo) |
4654 | { | |
4655 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
4656 | ||
b690cc0f | 4657 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
a70d6342 IR |
4658 | if (vect_print_dump_info (REPORT_DETAILS)) |
4659 | { | |
4660 | fprintf (vect_dump, "get vectype for scalar type: "); | |
4661 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
4662 | } | |
4663 | ||
4664 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4665 | if (!vectype) | |
4666 | { | |
4667 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4668 | { | |
4669 | fprintf (vect_dump, "not SLPed: unsupported data-type "); | |
4670 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
4671 | } | |
4672 | return false; | |
4673 | } | |
4674 | ||
4675 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4676 | { | |
4677 | fprintf (vect_dump, "vectype: "); | |
4678 | print_generic_expr (vect_dump, vectype, TDF_SLIM); | |
4679 | } | |
4680 | ||
4681 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
4682 | } | |
4683 | ||
8644a673 | 4684 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 4685 | { |
8644a673 IR |
4686 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
4687 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info)); | |
4688 | *need_to_vectorize = true; | |
ebfd146a IR |
4689 | } |
4690 | ||
8644a673 | 4691 | ok = true; |
b8698a0f | 4692 | if (!bb_vinfo |
a70d6342 IR |
4693 | && (STMT_VINFO_RELEVANT_P (stmt_info) |
4694 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
8644a673 IR |
4695 | ok = (vectorizable_type_promotion (stmt, NULL, NULL, NULL) |
4696 | || vectorizable_type_demotion (stmt, NULL, NULL, NULL) | |
4697 | || vectorizable_conversion (stmt, NULL, NULL, NULL) | |
9dc3f7de | 4698 | || vectorizable_shift (stmt, NULL, NULL, NULL) |
8644a673 IR |
4699 | || vectorizable_operation (stmt, NULL, NULL, NULL) |
4700 | || vectorizable_assignment (stmt, NULL, NULL, NULL) | |
4701 | || vectorizable_load (stmt, NULL, NULL, NULL, NULL) | |
4702 | || vectorizable_call (stmt, NULL, NULL) | |
4703 | || vectorizable_store (stmt, NULL, NULL, NULL) | |
b5aeb3bb | 4704 | || vectorizable_reduction (stmt, NULL, NULL, NULL) |
4bbe8262 | 4705 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0)); |
a70d6342 IR |
4706 | else |
4707 | { | |
4708 | if (bb_vinfo) | |
57416708 | 4709 | ok = (vectorizable_shift (stmt, NULL, NULL, node) |
9dc3f7de | 4710 | || vectorizable_operation (stmt, NULL, NULL, node) |
a70d6342 IR |
4711 | || vectorizable_assignment (stmt, NULL, NULL, node) |
4712 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
4713 | || vectorizable_store (stmt, NULL, NULL, node)); | |
b8698a0f | 4714 | } |
8644a673 IR |
4715 | |
4716 | if (!ok) | |
ebfd146a | 4717 | { |
8644a673 IR |
4718 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
4719 | { | |
4720 | fprintf (vect_dump, "not vectorized: relevant stmt not "); | |
4721 | fprintf (vect_dump, "supported: "); | |
4722 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4723 | } | |
b8698a0f | 4724 | |
ebfd146a IR |
4725 | return false; |
4726 | } | |
4727 | ||
a70d6342 IR |
4728 | if (bb_vinfo) |
4729 | return true; | |
4730 | ||
8644a673 IR |
4731 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
4732 | need extra handling, except for vectorizable reductions. */ | |
4733 | if (STMT_VINFO_LIVE_P (stmt_info) | |
4734 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
4735 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 4736 | |
8644a673 | 4737 | if (!ok) |
ebfd146a | 4738 | { |
8644a673 IR |
4739 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
4740 | { | |
4741 | fprintf (vect_dump, "not vectorized: live stmt not "); | |
4742 | fprintf (vect_dump, "supported: "); | |
4743 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4744 | } | |
b8698a0f | 4745 | |
8644a673 | 4746 | return false; |
ebfd146a IR |
4747 | } |
4748 | ||
8644a673 | 4749 | if (!PURE_SLP_STMT (stmt_info)) |
ebfd146a | 4750 | { |
b8698a0f | 4751 | /* Groups of strided accesses whose size is not a power of 2 are not |
ff802fa1 | 4752 | vectorizable yet using loop-vectorization. Therefore, if this stmt |
b8698a0f | 4753 | feeds non-SLP-able stmts (i.e., this stmt has to be both SLPed and |
a70d6342 | 4754 | loop-based vectorized), the loop cannot be vectorized. */ |
8644a673 IR |
4755 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) |
4756 | && exact_log2 (DR_GROUP_SIZE (vinfo_for_stmt ( | |
4757 | DR_GROUP_FIRST_DR (stmt_info)))) == -1) | |
ebfd146a | 4758 | { |
8644a673 IR |
4759 | if (vect_print_dump_info (REPORT_DETAILS)) |
4760 | { | |
4761 | fprintf (vect_dump, "not vectorized: the size of group " | |
4762 | "of strided accesses is not a power of 2"); | |
4763 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4764 | } | |
4765 | ||
ebfd146a IR |
4766 | return false; |
4767 | } | |
4768 | } | |
b8698a0f | 4769 | |
ebfd146a IR |
4770 | return true; |
4771 | } | |
4772 | ||
4773 | ||
4774 | /* Function vect_transform_stmt. | |
4775 | ||
4776 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
4777 | ||
4778 | bool | |
4779 | vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi, | |
b8698a0f | 4780 | bool *strided_store, slp_tree slp_node, |
ebfd146a IR |
4781 | slp_instance slp_node_instance) |
4782 | { | |
4783 | bool is_store = false; | |
4784 | gimple vec_stmt = NULL; | |
4785 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
039d9ea1 | 4786 | gimple orig_stmt_in_pattern, orig_scalar_stmt = stmt; |
ebfd146a | 4787 | bool done; |
ebfd146a IR |
4788 | |
4789 | switch (STMT_VINFO_TYPE (stmt_info)) | |
4790 | { | |
4791 | case type_demotion_vec_info_type: | |
4792 | done = vectorizable_type_demotion (stmt, gsi, &vec_stmt, slp_node); | |
4793 | gcc_assert (done); | |
4794 | break; | |
4795 | ||
4796 | case type_promotion_vec_info_type: | |
4797 | done = vectorizable_type_promotion (stmt, gsi, &vec_stmt, slp_node); | |
4798 | gcc_assert (done); | |
4799 | break; | |
4800 | ||
4801 | case type_conversion_vec_info_type: | |
4802 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
4803 | gcc_assert (done); | |
4804 | break; | |
4805 | ||
4806 | case induc_vec_info_type: | |
4807 | gcc_assert (!slp_node); | |
4808 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
4809 | gcc_assert (done); | |
4810 | break; | |
4811 | ||
9dc3f7de IR |
4812 | case shift_vec_info_type: |
4813 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
4814 | gcc_assert (done); | |
4815 | break; | |
4816 | ||
ebfd146a IR |
4817 | case op_vec_info_type: |
4818 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
4819 | gcc_assert (done); | |
4820 | break; | |
4821 | ||
4822 | case assignment_vec_info_type: | |
4823 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
4824 | gcc_assert (done); | |
4825 | break; | |
4826 | ||
4827 | case load_vec_info_type: | |
b8698a0f | 4828 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
4829 | slp_node_instance); |
4830 | gcc_assert (done); | |
4831 | break; | |
4832 | ||
4833 | case store_vec_info_type: | |
4834 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
4835 | gcc_assert (done); | |
4836 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) && !slp_node) | |
4837 | { | |
4838 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 4839 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
4840 | one are skipped, and there vec_stmt_info shouldn't be freed |
4841 | meanwhile. */ | |
4842 | *strided_store = true; | |
4843 | if (STMT_VINFO_VEC_STMT (stmt_info)) | |
4844 | is_store = true; | |
4845 | } | |
4846 | else | |
4847 | is_store = true; | |
4848 | break; | |
4849 | ||
4850 | case condition_vec_info_type: | |
4851 | gcc_assert (!slp_node); | |
4bbe8262 | 4852 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0); |
ebfd146a IR |
4853 | gcc_assert (done); |
4854 | break; | |
4855 | ||
4856 | case call_vec_info_type: | |
4857 | gcc_assert (!slp_node); | |
4858 | done = vectorizable_call (stmt, gsi, &vec_stmt); | |
039d9ea1 | 4859 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
4860 | break; |
4861 | ||
4862 | case reduc_vec_info_type: | |
b5aeb3bb | 4863 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
4864 | gcc_assert (done); |
4865 | break; | |
4866 | ||
4867 | default: | |
4868 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
4869 | { | |
4870 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4871 | fprintf (vect_dump, "stmt not supported."); | |
4872 | gcc_unreachable (); | |
4873 | } | |
4874 | } | |
4875 | ||
4876 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that | |
4877 | is being vectorized, but outside the immediately enclosing loop. */ | |
4878 | if (vec_stmt | |
a70d6342 IR |
4879 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
4880 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
4881 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
4882 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
4883 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 4884 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 4885 | vect_used_in_outer_by_reduction)) |
ebfd146a | 4886 | { |
a70d6342 IR |
4887 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
4888 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
4889 | imm_use_iterator imm_iter; |
4890 | use_operand_p use_p; | |
4891 | tree scalar_dest; | |
4892 | gimple exit_phi; | |
4893 | ||
4894 | if (vect_print_dump_info (REPORT_DETAILS)) | |
a70d6342 | 4895 | fprintf (vect_dump, "Record the vdef for outer-loop vectorization."); |
ebfd146a IR |
4896 | |
4897 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
4898 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
4899 | STMT). */ | |
4900 | if (gimple_code (stmt) == GIMPLE_PHI) | |
4901 | scalar_dest = PHI_RESULT (stmt); | |
4902 | else | |
4903 | scalar_dest = gimple_assign_lhs (stmt); | |
4904 | ||
4905 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
4906 | { | |
4907 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
4908 | { | |
4909 | exit_phi = USE_STMT (use_p); | |
4910 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
4911 | } | |
4912 | } | |
4913 | } | |
4914 | ||
4915 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
4916 | being vectorized. */ | |
4917 | if (STMT_VINFO_LIVE_P (stmt_info) | |
4918 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
4919 | { | |
4920 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
4921 | gcc_assert (done); | |
4922 | } | |
4923 | ||
4924 | if (vec_stmt) | |
4925 | { | |
4926 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; | |
4927 | orig_stmt_in_pattern = STMT_VINFO_RELATED_STMT (stmt_info); | |
4928 | if (orig_stmt_in_pattern) | |
4929 | { | |
4930 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt_in_pattern); | |
4931 | /* STMT was inserted by the vectorizer to replace a computation idiom. | |
b8698a0f L |
4932 | ORIG_STMT_IN_PATTERN is a stmt in the original sequence that |
4933 | computed this idiom. We need to record a pointer to VEC_STMT in | |
4934 | the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the | |
ebfd146a IR |
4935 | documentation of vect_pattern_recog. */ |
4936 | if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo)) | |
4937 | { | |
039d9ea1 IR |
4938 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) |
4939 | == orig_scalar_stmt); | |
ebfd146a IR |
4940 | STMT_VINFO_VEC_STMT (stmt_vinfo) = vec_stmt; |
4941 | } | |
4942 | } | |
4943 | } | |
4944 | ||
b8698a0f | 4945 | return is_store; |
ebfd146a IR |
4946 | } |
4947 | ||
4948 | ||
b8698a0f | 4949 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
4950 | stmt_vec_info. */ |
4951 | ||
4952 | void | |
4953 | vect_remove_stores (gimple first_stmt) | |
4954 | { | |
4955 | gimple next = first_stmt; | |
4956 | gimple tmp; | |
4957 | gimple_stmt_iterator next_si; | |
4958 | ||
4959 | while (next) | |
4960 | { | |
4961 | /* Free the attached stmt_vec_info and remove the stmt. */ | |
4962 | next_si = gsi_for_stmt (next); | |
4963 | gsi_remove (&next_si, true); | |
4964 | tmp = DR_GROUP_NEXT_DR (vinfo_for_stmt (next)); | |
4965 | free_stmt_vec_info (next); | |
4966 | next = tmp; | |
4967 | } | |
4968 | } | |
4969 | ||
4970 | ||
4971 | /* Function new_stmt_vec_info. | |
4972 | ||
4973 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
4974 | ||
4975 | stmt_vec_info | |
b8698a0f | 4976 | new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 4977 | bb_vec_info bb_vinfo) |
ebfd146a IR |
4978 | { |
4979 | stmt_vec_info res; | |
4980 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
4981 | ||
4982 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
4983 | STMT_VINFO_STMT (res) = stmt; | |
4984 | STMT_VINFO_LOOP_VINFO (res) = loop_vinfo; | |
a70d6342 | 4985 | STMT_VINFO_BB_VINFO (res) = bb_vinfo; |
8644a673 | 4986 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
4987 | STMT_VINFO_LIVE_P (res) = false; |
4988 | STMT_VINFO_VECTYPE (res) = NULL; | |
4989 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 4990 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
4991 | STMT_VINFO_IN_PATTERN_P (res) = false; |
4992 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
4993 | STMT_VINFO_DATA_REF (res) = NULL; | |
4994 | ||
4995 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
4996 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
4997 | STMT_VINFO_DR_INIT (res) = NULL; | |
4998 | STMT_VINFO_DR_STEP (res) = NULL; | |
4999 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
5000 | ||
5001 | if (gimple_code (stmt) == GIMPLE_PHI | |
5002 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
5003 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
5004 | else | |
8644a673 IR |
5005 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
5006 | ||
ebfd146a IR |
5007 | STMT_VINFO_SAME_ALIGN_REFS (res) = VEC_alloc (dr_p, heap, 5); |
5008 | STMT_VINFO_INSIDE_OF_LOOP_COST (res) = 0; | |
5009 | STMT_VINFO_OUTSIDE_OF_LOOP_COST (res) = 0; | |
32e8bb8e | 5010 | STMT_SLP_TYPE (res) = loop_vect; |
ebfd146a IR |
5011 | DR_GROUP_FIRST_DR (res) = NULL; |
5012 | DR_GROUP_NEXT_DR (res) = NULL; | |
5013 | DR_GROUP_SIZE (res) = 0; | |
5014 | DR_GROUP_STORE_COUNT (res) = 0; | |
5015 | DR_GROUP_GAP (res) = 0; | |
5016 | DR_GROUP_SAME_DR_STMT (res) = NULL; | |
5017 | DR_GROUP_READ_WRITE_DEPENDENCE (res) = false; | |
5018 | ||
5019 | return res; | |
5020 | } | |
5021 | ||
5022 | ||
5023 | /* Create a hash table for stmt_vec_info. */ | |
5024 | ||
5025 | void | |
5026 | init_stmt_vec_info_vec (void) | |
5027 | { | |
5028 | gcc_assert (!stmt_vec_info_vec); | |
5029 | stmt_vec_info_vec = VEC_alloc (vec_void_p, heap, 50); | |
5030 | } | |
5031 | ||
5032 | ||
5033 | /* Free hash table for stmt_vec_info. */ | |
5034 | ||
5035 | void | |
5036 | free_stmt_vec_info_vec (void) | |
5037 | { | |
5038 | gcc_assert (stmt_vec_info_vec); | |
5039 | VEC_free (vec_void_p, heap, stmt_vec_info_vec); | |
5040 | } | |
5041 | ||
5042 | ||
5043 | /* Free stmt vectorization related info. */ | |
5044 | ||
5045 | void | |
5046 | free_stmt_vec_info (gimple stmt) | |
5047 | { | |
5048 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5049 | ||
5050 | if (!stmt_info) | |
5051 | return; | |
5052 | ||
5053 | VEC_free (dr_p, heap, STMT_VINFO_SAME_ALIGN_REFS (stmt_info)); | |
5054 | set_vinfo_for_stmt (stmt, NULL); | |
5055 | free (stmt_info); | |
5056 | } | |
5057 | ||
5058 | ||
bb67d9c7 | 5059 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 5060 | |
bb67d9c7 | 5061 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
5062 | by the target. */ |
5063 | ||
bb67d9c7 RG |
5064 | static tree |
5065 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a IR |
5066 | { |
5067 | enum machine_mode inner_mode = TYPE_MODE (scalar_type); | |
cc4b5170 | 5068 | enum machine_mode simd_mode; |
2f816591 | 5069 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
5070 | int nunits; |
5071 | tree vectype; | |
5072 | ||
cc4b5170 | 5073 | if (nbytes == 0) |
ebfd146a IR |
5074 | return NULL_TREE; |
5075 | ||
2f816591 RG |
5076 | /* We can't build a vector type of elements with alignment bigger than |
5077 | their size. */ | |
5078 | if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) | |
5079 | return NULL_TREE; | |
5080 | ||
6d7971b8 RG |
5081 | /* If we'd build a vector type of elements whose mode precision doesn't |
5082 | match their types precision we'll get mismatched types on vector | |
5083 | extracts via BIT_FIELD_REFs. This effectively means we disable | |
5084 | vectorization of bool and/or enum types in some languages. */ | |
5085 | if (INTEGRAL_TYPE_P (scalar_type) | |
5086 | && GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type)) | |
5087 | return NULL_TREE; | |
5088 | ||
cc4b5170 RG |
5089 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
5090 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
5091 | return NULL_TREE; | |
5092 | ||
bb67d9c7 RG |
5093 | /* If no size was supplied use the mode the target prefers. Otherwise |
5094 | lookup a vector mode of the specified size. */ | |
5095 | if (size == 0) | |
5096 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
5097 | else | |
5098 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
5099 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
5100 | if (nunits <= 1) | |
5101 | return NULL_TREE; | |
ebfd146a IR |
5102 | |
5103 | vectype = build_vector_type (scalar_type, nunits); | |
5104 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5105 | { | |
5106 | fprintf (vect_dump, "get vectype with %d units of type ", nunits); | |
5107 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
5108 | } | |
5109 | ||
5110 | if (!vectype) | |
5111 | return NULL_TREE; | |
5112 | ||
5113 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5114 | { | |
5115 | fprintf (vect_dump, "vectype: "); | |
5116 | print_generic_expr (vect_dump, vectype, TDF_SLIM); | |
5117 | } | |
5118 | ||
5119 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
5120 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
5121 | { | |
5122 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5123 | fprintf (vect_dump, "mode not supported by target."); | |
5124 | return NULL_TREE; | |
5125 | } | |
5126 | ||
5127 | return vectype; | |
5128 | } | |
5129 | ||
bb67d9c7 RG |
5130 | unsigned int current_vector_size; |
5131 | ||
5132 | /* Function get_vectype_for_scalar_type. | |
5133 | ||
5134 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
5135 | by the target. */ | |
5136 | ||
5137 | tree | |
5138 | get_vectype_for_scalar_type (tree scalar_type) | |
5139 | { | |
5140 | tree vectype; | |
5141 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
5142 | current_vector_size); | |
5143 | if (vectype | |
5144 | && current_vector_size == 0) | |
5145 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
5146 | return vectype; | |
5147 | } | |
5148 | ||
b690cc0f RG |
5149 | /* Function get_same_sized_vectype |
5150 | ||
5151 | Returns a vector type corresponding to SCALAR_TYPE of size | |
5152 | VECTOR_TYPE if supported by the target. */ | |
5153 | ||
5154 | tree | |
bb67d9c7 | 5155 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 5156 | { |
bb67d9c7 RG |
5157 | return get_vectype_for_scalar_type_and_size |
5158 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
5159 | } |
5160 | ||
ebfd146a IR |
5161 | /* Function vect_is_simple_use. |
5162 | ||
5163 | Input: | |
a70d6342 IR |
5164 | LOOP_VINFO - the vect info of the loop that is being vectorized. |
5165 | BB_VINFO - the vect info of the basic block that is being vectorized. | |
5166 | OPERAND - operand of a stmt in the loop or bb. | |
ebfd146a IR |
5167 | DEF - the defining stmt in case OPERAND is an SSA_NAME. |
5168 | ||
5169 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 5170 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 5171 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 5172 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
5173 | is the case in reduction/induction computations). |
5174 | For basic blocks, supportable operands are constants and bb invariants. | |
5175 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
5176 | |
5177 | bool | |
b8698a0f | 5178 | vect_is_simple_use (tree operand, loop_vec_info loop_vinfo, |
a70d6342 | 5179 | bb_vec_info bb_vinfo, gimple *def_stmt, |
ebfd146a | 5180 | tree *def, enum vect_def_type *dt) |
b8698a0f | 5181 | { |
ebfd146a IR |
5182 | basic_block bb; |
5183 | stmt_vec_info stmt_vinfo; | |
a70d6342 | 5184 | struct loop *loop = NULL; |
b8698a0f | 5185 | |
a70d6342 IR |
5186 | if (loop_vinfo) |
5187 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
5188 | |
5189 | *def_stmt = NULL; | |
5190 | *def = NULL_TREE; | |
b8698a0f | 5191 | |
ebfd146a IR |
5192 | if (vect_print_dump_info (REPORT_DETAILS)) |
5193 | { | |
5194 | fprintf (vect_dump, "vect_is_simple_use: operand "); | |
5195 | print_generic_expr (vect_dump, operand, TDF_SLIM); | |
5196 | } | |
b8698a0f | 5197 | |
ebfd146a IR |
5198 | if (TREE_CODE (operand) == INTEGER_CST || TREE_CODE (operand) == REAL_CST) |
5199 | { | |
5200 | *dt = vect_constant_def; | |
5201 | return true; | |
5202 | } | |
b8698a0f | 5203 | |
ebfd146a IR |
5204 | if (is_gimple_min_invariant (operand)) |
5205 | { | |
5206 | *def = operand; | |
8644a673 | 5207 | *dt = vect_external_def; |
ebfd146a IR |
5208 | return true; |
5209 | } | |
5210 | ||
5211 | if (TREE_CODE (operand) == PAREN_EXPR) | |
5212 | { | |
5213 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5214 | fprintf (vect_dump, "non-associatable copy."); | |
5215 | operand = TREE_OPERAND (operand, 0); | |
5216 | } | |
b8698a0f | 5217 | |
ebfd146a IR |
5218 | if (TREE_CODE (operand) != SSA_NAME) |
5219 | { | |
5220 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5221 | fprintf (vect_dump, "not ssa-name."); | |
5222 | return false; | |
5223 | } | |
b8698a0f | 5224 | |
ebfd146a IR |
5225 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
5226 | if (*def_stmt == NULL) | |
5227 | { | |
5228 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5229 | fprintf (vect_dump, "no def_stmt."); | |
5230 | return false; | |
5231 | } | |
5232 | ||
5233 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5234 | { | |
5235 | fprintf (vect_dump, "def_stmt: "); | |
5236 | print_gimple_stmt (vect_dump, *def_stmt, 0, TDF_SLIM); | |
5237 | } | |
5238 | ||
8644a673 | 5239 | /* Empty stmt is expected only in case of a function argument. |
ebfd146a IR |
5240 | (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */ |
5241 | if (gimple_nop_p (*def_stmt)) | |
5242 | { | |
5243 | *def = operand; | |
8644a673 | 5244 | *dt = vect_external_def; |
ebfd146a IR |
5245 | return true; |
5246 | } | |
5247 | ||
5248 | bb = gimple_bb (*def_stmt); | |
a70d6342 IR |
5249 | |
5250 | if ((loop && !flow_bb_inside_loop_p (loop, bb)) | |
5251 | || (!loop && bb != BB_VINFO_BB (bb_vinfo)) | |
b8698a0f | 5252 | || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI)) |
8644a673 | 5253 | *dt = vect_external_def; |
ebfd146a IR |
5254 | else |
5255 | { | |
5256 | stmt_vinfo = vinfo_for_stmt (*def_stmt); | |
5257 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
5258 | } | |
5259 | ||
5260 | if (*dt == vect_unknown_def_type) | |
5261 | { | |
5262 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5263 | fprintf (vect_dump, "Unsupported pattern."); | |
5264 | return false; | |
5265 | } | |
5266 | ||
5267 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5268 | fprintf (vect_dump, "type of def: %d.",*dt); | |
5269 | ||
5270 | switch (gimple_code (*def_stmt)) | |
5271 | { | |
5272 | case GIMPLE_PHI: | |
5273 | *def = gimple_phi_result (*def_stmt); | |
5274 | break; | |
5275 | ||
5276 | case GIMPLE_ASSIGN: | |
5277 | *def = gimple_assign_lhs (*def_stmt); | |
5278 | break; | |
5279 | ||
5280 | case GIMPLE_CALL: | |
5281 | *def = gimple_call_lhs (*def_stmt); | |
5282 | if (*def != NULL) | |
5283 | break; | |
5284 | /* FALLTHRU */ | |
5285 | default: | |
5286 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5287 | fprintf (vect_dump, "unsupported defining stmt: "); | |
5288 | return false; | |
5289 | } | |
5290 | ||
5291 | return true; | |
5292 | } | |
5293 | ||
b690cc0f RG |
5294 | /* Function vect_is_simple_use_1. |
5295 | ||
5296 | Same as vect_is_simple_use_1 but also determines the vector operand | |
5297 | type of OPERAND and stores it to *VECTYPE. If the definition of | |
5298 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
5299 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
5300 | is responsible to compute the best suited vector type for the | |
5301 | scalar operand. */ | |
5302 | ||
5303 | bool | |
5304 | vect_is_simple_use_1 (tree operand, loop_vec_info loop_vinfo, | |
5305 | bb_vec_info bb_vinfo, gimple *def_stmt, | |
5306 | tree *def, enum vect_def_type *dt, tree *vectype) | |
5307 | { | |
5308 | if (!vect_is_simple_use (operand, loop_vinfo, bb_vinfo, def_stmt, def, dt)) | |
5309 | return false; | |
5310 | ||
5311 | /* Now get a vector type if the def is internal, otherwise supply | |
5312 | NULL_TREE and leave it up to the caller to figure out a proper | |
5313 | type for the use stmt. */ | |
5314 | if (*dt == vect_internal_def | |
5315 | || *dt == vect_induction_def | |
5316 | || *dt == vect_reduction_def | |
5317 | || *dt == vect_double_reduction_def | |
5318 | || *dt == vect_nested_cycle) | |
5319 | { | |
5320 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
5321 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
5322 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
5323 | *vectype = STMT_VINFO_VECTYPE (stmt_info); | |
5324 | gcc_assert (*vectype != NULL_TREE); | |
5325 | } | |
5326 | else if (*dt == vect_uninitialized_def | |
5327 | || *dt == vect_constant_def | |
5328 | || *dt == vect_external_def) | |
5329 | *vectype = NULL_TREE; | |
5330 | else | |
5331 | gcc_unreachable (); | |
5332 | ||
5333 | return true; | |
5334 | } | |
5335 | ||
ebfd146a IR |
5336 | |
5337 | /* Function supportable_widening_operation | |
5338 | ||
b8698a0f L |
5339 | Check whether an operation represented by the code CODE is a |
5340 | widening operation that is supported by the target platform in | |
b690cc0f RG |
5341 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
5342 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 5343 | |
ebfd146a IR |
5344 | Widening operations we currently support are NOP (CONVERT), FLOAT |
5345 | and WIDEN_MULT. This function checks if these operations are supported | |
5346 | by the target platform either directly (via vector tree-codes), or via | |
5347 | target builtins. | |
5348 | ||
5349 | Output: | |
b8698a0f L |
5350 | - CODE1 and CODE2 are codes of vector operations to be used when |
5351 | vectorizing the operation, if available. | |
ebfd146a | 5352 | - DECL1 and DECL2 are decls of target builtin functions to be used |
ff802fa1 | 5353 | when vectorizing the operation, if available. In this case, |
b8698a0f | 5354 | CODE1 and CODE2 are CALL_EXPR. |
ebfd146a IR |
5355 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
5356 | case of multi-step conversion (like char->short->int - in that case | |
5357 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
5358 | - INTERM_TYPES contains the intermediate type required to perform the |
5359 | widening operation (short in the above example). */ | |
ebfd146a IR |
5360 | |
5361 | bool | |
b690cc0f RG |
5362 | supportable_widening_operation (enum tree_code code, gimple stmt, |
5363 | tree vectype_out, tree vectype_in, | |
ebfd146a IR |
5364 | tree *decl1, tree *decl2, |
5365 | enum tree_code *code1, enum tree_code *code2, | |
5366 | int *multi_step_cvt, | |
5367 | VEC (tree, heap) **interm_types) | |
5368 | { | |
5369 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5370 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
5371 | struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info); | |
5372 | bool ordered_p; | |
5373 | enum machine_mode vec_mode; | |
81f40b79 | 5374 | enum insn_code icode1, icode2; |
ebfd146a | 5375 | optab optab1, optab2; |
b690cc0f RG |
5376 | tree vectype = vectype_in; |
5377 | tree wide_vectype = vectype_out; | |
ebfd146a IR |
5378 | enum tree_code c1, c2; |
5379 | ||
5380 | /* The result of a vectorized widening operation usually requires two vectors | |
b8698a0f L |
5381 | (because the widened results do not fit int one vector). The generated |
5382 | vector results would normally be expected to be generated in the same | |
ebfd146a IR |
5383 | order as in the original scalar computation, i.e. if 8 results are |
5384 | generated in each vector iteration, they are to be organized as follows: | |
b8698a0f | 5385 | vect1: [res1,res2,res3,res4], vect2: [res5,res6,res7,res8]. |
ebfd146a | 5386 | |
b8698a0f | 5387 | However, in the special case that the result of the widening operation is |
ebfd146a | 5388 | used in a reduction computation only, the order doesn't matter (because |
b8698a0f | 5389 | when vectorizing a reduction we change the order of the computation). |
ebfd146a IR |
5390 | Some targets can take advantage of this and generate more efficient code. |
5391 | For example, targets like Altivec, that support widen_mult using a sequence | |
5392 | of {mult_even,mult_odd} generate the following vectors: | |
5393 | vect1: [res1,res3,res5,res7], vect2: [res2,res4,res6,res8]. | |
5394 | ||
5395 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
b8698a0f | 5396 | (each vectorized inner-loop iteration contributes to VF outer-loop |
ff802fa1 | 5397 | iterations in parallel). We therefore don't allow to change the order |
ebfd146a IR |
5398 | of the computation in the inner-loop during outer-loop vectorization. */ |
5399 | ||
5400 | if (STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
5401 | && !nested_in_vect_loop_p (vect_loop, stmt)) | |
5402 | ordered_p = false; | |
5403 | else | |
5404 | ordered_p = true; | |
5405 | ||
5406 | if (!ordered_p | |
5407 | && code == WIDEN_MULT_EXPR | |
5408 | && targetm.vectorize.builtin_mul_widen_even | |
5409 | && targetm.vectorize.builtin_mul_widen_even (vectype) | |
5410 | && targetm.vectorize.builtin_mul_widen_odd | |
5411 | && targetm.vectorize.builtin_mul_widen_odd (vectype)) | |
5412 | { | |
5413 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5414 | fprintf (vect_dump, "Unordered widening operation detected."); | |
5415 | ||
5416 | *code1 = *code2 = CALL_EXPR; | |
5417 | *decl1 = targetm.vectorize.builtin_mul_widen_even (vectype); | |
5418 | *decl2 = targetm.vectorize.builtin_mul_widen_odd (vectype); | |
5419 | return true; | |
5420 | } | |
5421 | ||
5422 | switch (code) | |
5423 | { | |
5424 | case WIDEN_MULT_EXPR: | |
5425 | if (BYTES_BIG_ENDIAN) | |
5426 | { | |
5427 | c1 = VEC_WIDEN_MULT_HI_EXPR; | |
5428 | c2 = VEC_WIDEN_MULT_LO_EXPR; | |
5429 | } | |
5430 | else | |
5431 | { | |
5432 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
5433 | c1 = VEC_WIDEN_MULT_LO_EXPR; | |
5434 | } | |
5435 | break; | |
5436 | ||
5437 | CASE_CONVERT: | |
5438 | if (BYTES_BIG_ENDIAN) | |
5439 | { | |
5440 | c1 = VEC_UNPACK_HI_EXPR; | |
5441 | c2 = VEC_UNPACK_LO_EXPR; | |
5442 | } | |
5443 | else | |
5444 | { | |
5445 | c2 = VEC_UNPACK_HI_EXPR; | |
5446 | c1 = VEC_UNPACK_LO_EXPR; | |
5447 | } | |
5448 | break; | |
5449 | ||
5450 | case FLOAT_EXPR: | |
5451 | if (BYTES_BIG_ENDIAN) | |
5452 | { | |
5453 | c1 = VEC_UNPACK_FLOAT_HI_EXPR; | |
5454 | c2 = VEC_UNPACK_FLOAT_LO_EXPR; | |
5455 | } | |
5456 | else | |
5457 | { | |
5458 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
5459 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; | |
5460 | } | |
5461 | break; | |
5462 | ||
5463 | case FIX_TRUNC_EXPR: | |
5464 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
5465 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
5466 | computing the operation. */ | |
5467 | return false; | |
5468 | ||
5469 | default: | |
5470 | gcc_unreachable (); | |
5471 | } | |
5472 | ||
5473 | if (code == FIX_TRUNC_EXPR) | |
5474 | { | |
5475 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
5476 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
5477 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
5478 | } |
5479 | else | |
5480 | { | |
5481 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
5482 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
5483 | } | |
5484 | ||
5485 | if (!optab1 || !optab2) | |
5486 | return false; | |
5487 | ||
5488 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
5489 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
5490 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
5491 | return false; |
5492 | ||
b8698a0f | 5493 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a IR |
5494 | types. */ |
5495 | if (insn_data[icode1].operand[0].mode != TYPE_MODE (wide_vectype) | |
5496 | || insn_data[icode2].operand[0].mode != TYPE_MODE (wide_vectype)) | |
5497 | { | |
5498 | int i; | |
5499 | tree prev_type = vectype, intermediate_type; | |
5500 | enum machine_mode intermediate_mode, prev_mode = vec_mode; | |
5501 | optab optab3, optab4; | |
5502 | ||
5503 | if (!CONVERT_EXPR_CODE_P (code)) | |
5504 | return false; | |
b8698a0f | 5505 | |
ebfd146a IR |
5506 | *code1 = c1; |
5507 | *code2 = c2; | |
b8698a0f | 5508 | |
ebfd146a | 5509 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
ff802fa1 IR |
5510 | intermediate steps in promotion sequence. We try |
5511 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
5512 | not. */ | |
ebfd146a IR |
5513 | *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS); |
5514 | for (i = 0; i < 3; i++) | |
5515 | { | |
5516 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
5517 | intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode, | |
5518 | TYPE_UNSIGNED (prev_type)); | |
5519 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
5520 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
5521 | ||
5522 | if (!optab3 || !optab4 | |
947131ba RS |
5523 | || ((icode1 = optab_handler (optab1, prev_mode)) |
5524 | == CODE_FOR_nothing) | |
ebfd146a | 5525 | || insn_data[icode1].operand[0].mode != intermediate_mode |
947131ba RS |
5526 | || ((icode2 = optab_handler (optab2, prev_mode)) |
5527 | == CODE_FOR_nothing) | |
ebfd146a | 5528 | || insn_data[icode2].operand[0].mode != intermediate_mode |
947131ba RS |
5529 | || ((icode1 = optab_handler (optab3, intermediate_mode)) |
5530 | == CODE_FOR_nothing) | |
5531 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
5532 | == CODE_FOR_nothing)) | |
ebfd146a IR |
5533 | return false; |
5534 | ||
5535 | VEC_quick_push (tree, *interm_types, intermediate_type); | |
5536 | (*multi_step_cvt)++; | |
5537 | ||
5538 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
5539 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5540 | return true; | |
5541 | ||
5542 | prev_type = intermediate_type; | |
5543 | prev_mode = intermediate_mode; | |
5544 | } | |
5545 | ||
5546 | return false; | |
5547 | } | |
5548 | ||
5549 | *code1 = c1; | |
5550 | *code2 = c2; | |
5551 | return true; | |
5552 | } | |
5553 | ||
5554 | ||
5555 | /* Function supportable_narrowing_operation | |
5556 | ||
b8698a0f L |
5557 | Check whether an operation represented by the code CODE is a |
5558 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
5559 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
5560 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 5561 | |
ebfd146a | 5562 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 5563 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
5564 | the target platform directly via vector tree-codes. |
5565 | ||
5566 | Output: | |
b8698a0f L |
5567 | - CODE1 is the code of a vector operation to be used when |
5568 | vectorizing the operation, if available. | |
ebfd146a IR |
5569 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
5570 | case of multi-step conversion (like int->short->char - in that case | |
5571 | MULTI_STEP_CVT will be 1). | |
5572 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 5573 | narrowing operation (short in the above example). */ |
ebfd146a IR |
5574 | |
5575 | bool | |
5576 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 5577 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
5578 | enum tree_code *code1, int *multi_step_cvt, |
5579 | VEC (tree, heap) **interm_types) | |
5580 | { | |
5581 | enum machine_mode vec_mode; | |
5582 | enum insn_code icode1; | |
5583 | optab optab1, interm_optab; | |
b690cc0f RG |
5584 | tree vectype = vectype_in; |
5585 | tree narrow_vectype = vectype_out; | |
ebfd146a IR |
5586 | enum tree_code c1; |
5587 | tree intermediate_type, prev_type; | |
5588 | int i; | |
5589 | ||
5590 | switch (code) | |
5591 | { | |
5592 | CASE_CONVERT: | |
5593 | c1 = VEC_PACK_TRUNC_EXPR; | |
5594 | break; | |
5595 | ||
5596 | case FIX_TRUNC_EXPR: | |
5597 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
5598 | break; | |
5599 | ||
5600 | case FLOAT_EXPR: | |
5601 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
5602 | tree code and optabs used for computing the operation. */ | |
5603 | return false; | |
5604 | ||
5605 | default: | |
5606 | gcc_unreachable (); | |
5607 | } | |
5608 | ||
5609 | if (code == FIX_TRUNC_EXPR) | |
5610 | /* The signedness is determined from output operand. */ | |
b690cc0f | 5611 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
5612 | else |
5613 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
5614 | ||
5615 | if (!optab1) | |
5616 | return false; | |
5617 | ||
5618 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 5619 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
5620 | return false; |
5621 | ||
5622 | /* Check if it's a multi-step conversion that can be done using intermediate | |
5623 | types. */ | |
5624 | if (insn_data[icode1].operand[0].mode != TYPE_MODE (narrow_vectype)) | |
5625 | { | |
5626 | enum machine_mode intermediate_mode, prev_mode = vec_mode; | |
5627 | ||
5628 | *code1 = c1; | |
5629 | prev_type = vectype; | |
5630 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS | |
ff802fa1 IR |
5631 | intermediate steps in promotion sequence. We try |
5632 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
5633 | not. */ | |
ebfd146a IR |
5634 | *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS); |
5635 | for (i = 0; i < 3; i++) | |
5636 | { | |
5637 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
5638 | intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode, | |
5639 | TYPE_UNSIGNED (prev_type)); | |
b8698a0f | 5640 | interm_optab = optab_for_tree_code (c1, intermediate_type, |
ebfd146a | 5641 | optab_default); |
b8698a0f | 5642 | if (!interm_optab |
947131ba RS |
5643 | || ((icode1 = optab_handler (optab1, prev_mode)) |
5644 | == CODE_FOR_nothing) | |
ebfd146a | 5645 | || insn_data[icode1].operand[0].mode != intermediate_mode |
947131ba RS |
5646 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) |
5647 | == CODE_FOR_nothing)) | |
ebfd146a IR |
5648 | return false; |
5649 | ||
5650 | VEC_quick_push (tree, *interm_types, intermediate_type); | |
5651 | (*multi_step_cvt)++; | |
5652 | ||
5653 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5654 | return true; | |
5655 | ||
5656 | prev_type = intermediate_type; | |
5657 | prev_mode = intermediate_mode; | |
5658 | } | |
5659 | ||
5660 | return false; | |
5661 | } | |
5662 | ||
5663 | *code1 = c1; | |
5664 | return true; | |
5665 | } |