]>
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 | ||
272c6793 RS |
45 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
46 | ||
47 | static tree | |
48 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
49 | { | |
50 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
51 | "vect_array"); | |
52 | } | |
53 | ||
54 | /* ARRAY is an array of vectors created by create_vector_array. | |
55 | Return an SSA_NAME for the vector in index N. The reference | |
56 | is part of the vectorization of STMT and the vector is associated | |
57 | with scalar destination SCALAR_DEST. */ | |
58 | ||
59 | static tree | |
60 | read_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree scalar_dest, | |
61 | tree array, unsigned HOST_WIDE_INT n) | |
62 | { | |
63 | tree vect_type, vect, vect_name, array_ref; | |
64 | gimple new_stmt; | |
65 | ||
66 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
67 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
68 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
69 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
70 | build_int_cst (size_type_node, n), | |
71 | NULL_TREE, NULL_TREE); | |
72 | ||
73 | new_stmt = gimple_build_assign (vect, array_ref); | |
74 | vect_name = make_ssa_name (vect, new_stmt); | |
75 | gimple_assign_set_lhs (new_stmt, vect_name); | |
76 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
77 | mark_symbols_for_renaming (new_stmt); | |
78 | ||
79 | return vect_name; | |
80 | } | |
81 | ||
82 | /* ARRAY is an array of vectors created by create_vector_array. | |
83 | Emit code to store SSA_NAME VECT in index N of the array. | |
84 | The store is part of the vectorization of STMT. */ | |
85 | ||
86 | static void | |
87 | write_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree vect, | |
88 | tree array, unsigned HOST_WIDE_INT n) | |
89 | { | |
90 | tree array_ref; | |
91 | gimple new_stmt; | |
92 | ||
93 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
94 | build_int_cst (size_type_node, n), | |
95 | NULL_TREE, NULL_TREE); | |
96 | ||
97 | new_stmt = gimple_build_assign (array_ref, vect); | |
98 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
99 | mark_symbols_for_renaming (new_stmt); | |
100 | } | |
101 | ||
102 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
103 | of *PTR. The memory reference replaces those in FIRST_DR | |
104 | (and its group). */ | |
105 | ||
106 | static tree | |
107 | create_array_ref (tree type, tree ptr, struct data_reference *first_dr) | |
108 | { | |
109 | struct ptr_info_def *pi; | |
110 | tree mem_ref, alias_ptr_type; | |
111 | ||
112 | alias_ptr_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
113 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); | |
114 | /* Arrays have the same alignment as their type. */ | |
115 | pi = get_ptr_info (ptr); | |
116 | pi->align = TYPE_ALIGN_UNIT (type); | |
117 | pi->misalign = 0; | |
118 | return mem_ref; | |
119 | } | |
120 | ||
ebfd146a IR |
121 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
122 | ||
123 | /* Function vect_mark_relevant. | |
124 | ||
125 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
126 | ||
127 | static void | |
128 | vect_mark_relevant (VEC(gimple,heap) **worklist, gimple stmt, | |
129 | enum vect_relevant relevant, bool live_p) | |
130 | { | |
131 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
132 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
133 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
134 | ||
135 | if (vect_print_dump_info (REPORT_DETAILS)) | |
136 | fprintf (vect_dump, "mark relevant %d, live %d.", relevant, live_p); | |
137 | ||
138 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
139 | { | |
140 | gimple pattern_stmt; | |
141 | ||
b8698a0f | 142 | /* This is the last stmt in a sequence that was detected as a |
ebfd146a IR |
143 | pattern that can potentially be vectorized. Don't mark the stmt |
144 | as relevant/live because it's not going to be vectorized. | |
145 | Instead mark the pattern-stmt that replaces it. */ | |
146 | ||
147 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
148 | ||
149 | if (vect_print_dump_info (REPORT_DETAILS)) | |
150 | fprintf (vect_dump, "last stmt in pattern. don't mark relevant/live."); | |
151 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
152 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
153 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
154 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
155 | stmt = pattern_stmt; | |
156 | } | |
157 | ||
158 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
159 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
160 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
161 | ||
162 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
163 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
164 | { | |
165 | if (vect_print_dump_info (REPORT_DETAILS)) | |
166 | fprintf (vect_dump, "already marked relevant/live."); | |
167 | return; | |
168 | } | |
169 | ||
170 | VEC_safe_push (gimple, heap, *worklist, stmt); | |
171 | } | |
172 | ||
173 | ||
174 | /* Function vect_stmt_relevant_p. | |
175 | ||
176 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
177 | "relevant for vectorization". | |
178 | ||
179 | A stmt is considered "relevant for vectorization" if: | |
180 | - it has uses outside the loop. | |
181 | - it has vdefs (it alters memory). | |
182 | - control stmts in the loop (except for the exit condition). | |
183 | ||
184 | CHECKME: what other side effects would the vectorizer allow? */ | |
185 | ||
186 | static bool | |
187 | vect_stmt_relevant_p (gimple stmt, loop_vec_info loop_vinfo, | |
188 | enum vect_relevant *relevant, bool *live_p) | |
189 | { | |
190 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
191 | ssa_op_iter op_iter; | |
192 | imm_use_iterator imm_iter; | |
193 | use_operand_p use_p; | |
194 | def_operand_p def_p; | |
195 | ||
8644a673 | 196 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
197 | *live_p = false; |
198 | ||
199 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
200 | if (is_ctrl_stmt (stmt) |
201 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
202 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 203 | *relevant = vect_used_in_scope; |
ebfd146a IR |
204 | |
205 | /* changing memory. */ | |
206 | if (gimple_code (stmt) != GIMPLE_PHI) | |
5006671f | 207 | if (gimple_vdef (stmt)) |
ebfd146a IR |
208 | { |
209 | if (vect_print_dump_info (REPORT_DETAILS)) | |
210 | fprintf (vect_dump, "vec_stmt_relevant_p: stmt has vdefs."); | |
8644a673 | 211 | *relevant = vect_used_in_scope; |
ebfd146a IR |
212 | } |
213 | ||
214 | /* uses outside the loop. */ | |
215 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
216 | { | |
217 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
218 | { | |
219 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
220 | if (!flow_bb_inside_loop_p (loop, bb)) | |
221 | { | |
222 | if (vect_print_dump_info (REPORT_DETAILS)) | |
223 | fprintf (vect_dump, "vec_stmt_relevant_p: used out of loop."); | |
224 | ||
3157b0c2 AO |
225 | if (is_gimple_debug (USE_STMT (use_p))) |
226 | continue; | |
227 | ||
ebfd146a IR |
228 | /* We expect all such uses to be in the loop exit phis |
229 | (because of loop closed form) */ | |
230 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
231 | gcc_assert (bb == single_exit (loop)->dest); | |
232 | ||
233 | *live_p = true; | |
234 | } | |
235 | } | |
236 | } | |
237 | ||
238 | return (*live_p || *relevant); | |
239 | } | |
240 | ||
241 | ||
b8698a0f | 242 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 243 | |
ff802fa1 | 244 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
245 | used in STMT for anything other than indexing an array. */ |
246 | ||
247 | static bool | |
248 | exist_non_indexing_operands_for_use_p (tree use, gimple stmt) | |
249 | { | |
250 | tree operand; | |
251 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 252 | |
ff802fa1 | 253 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
254 | reference in STMT, then any operand that corresponds to USE |
255 | is not indexing an array. */ | |
256 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
257 | return true; | |
59a05b0c | 258 | |
ebfd146a IR |
259 | /* STMT has a data_ref. FORNOW this means that its of one of |
260 | the following forms: | |
261 | -1- ARRAY_REF = var | |
262 | -2- var = ARRAY_REF | |
263 | (This should have been verified in analyze_data_refs). | |
264 | ||
265 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 266 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
267 | for array indexing. |
268 | ||
269 | Therefore, all we need to check is if STMT falls into the | |
270 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
271 | |
272 | if (!gimple_assign_copy_p (stmt)) | |
273 | return false; | |
59a05b0c EB |
274 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
275 | return false; | |
ebfd146a | 276 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
277 | if (TREE_CODE (operand) != SSA_NAME) |
278 | return false; | |
279 | ||
280 | if (operand == use) | |
281 | return true; | |
282 | ||
283 | return false; | |
284 | } | |
285 | ||
286 | ||
b8698a0f | 287 | /* |
ebfd146a IR |
288 | Function process_use. |
289 | ||
290 | Inputs: | |
291 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 292 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ff802fa1 | 293 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a IR |
294 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
295 | ||
296 | Outputs: | |
297 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
298 | relevance info of the DEF_STMT of this USE: | |
299 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
300 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
301 | Exceptions: | |
302 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 303 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 304 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
305 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
306 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
307 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
308 | be modified accordingly. | |
309 | ||
310 | Return true if everything is as expected. Return false otherwise. */ | |
311 | ||
312 | static bool | |
b8698a0f | 313 | process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
ebfd146a IR |
314 | enum vect_relevant relevant, VEC(gimple,heap) **worklist) |
315 | { | |
316 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
317 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
318 | stmt_vec_info dstmt_vinfo; | |
319 | basic_block bb, def_bb; | |
320 | tree def; | |
321 | gimple def_stmt; | |
322 | enum vect_def_type dt; | |
323 | ||
b8698a0f | 324 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a IR |
325 | that are used for address computation are not considered relevant. */ |
326 | if (!exist_non_indexing_operands_for_use_p (use, stmt)) | |
327 | return true; | |
328 | ||
a70d6342 | 329 | if (!vect_is_simple_use (use, loop_vinfo, NULL, &def_stmt, &def, &dt)) |
b8698a0f | 330 | { |
8644a673 | 331 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
ebfd146a IR |
332 | fprintf (vect_dump, "not vectorized: unsupported use in stmt."); |
333 | return false; | |
334 | } | |
335 | ||
336 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
337 | return true; | |
338 | ||
339 | def_bb = gimple_bb (def_stmt); | |
340 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
341 | { | |
342 | if (vect_print_dump_info (REPORT_DETAILS)) | |
343 | fprintf (vect_dump, "def_stmt is out of loop."); | |
344 | return true; | |
345 | } | |
346 | ||
b8698a0f L |
347 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
348 | DEF_STMT must have already been processed, because this should be the | |
349 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
350 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
351 | check that everything is as expected, and we are done. */ |
352 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
353 | bb = gimple_bb (stmt); | |
354 | if (gimple_code (stmt) == GIMPLE_PHI | |
355 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
356 | && gimple_code (def_stmt) != GIMPLE_PHI | |
357 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
358 | && bb->loop_father == def_bb->loop_father) | |
359 | { | |
360 | if (vect_print_dump_info (REPORT_DETAILS)) | |
361 | fprintf (vect_dump, "reduc-stmt defining reduc-phi in the same nest."); | |
362 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) | |
363 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
364 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 365 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 366 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
367 | return true; |
368 | } | |
369 | ||
370 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
371 | outer-loop-header-bb: | |
372 | d = def_stmt | |
373 | inner-loop: | |
374 | stmt # use (d) | |
375 | outer-loop-tail-bb: | |
376 | ... */ | |
377 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
378 | { | |
379 | if (vect_print_dump_info (REPORT_DETAILS)) | |
380 | fprintf (vect_dump, "outer-loop def-stmt defining inner-loop stmt."); | |
7c5222ff | 381 | |
ebfd146a IR |
382 | switch (relevant) |
383 | { | |
8644a673 | 384 | case vect_unused_in_scope: |
7c5222ff IR |
385 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
386 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 387 | break; |
7c5222ff | 388 | |
ebfd146a | 389 | case vect_used_in_outer_by_reduction: |
7c5222ff | 390 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
391 | relevant = vect_used_by_reduction; |
392 | break; | |
7c5222ff | 393 | |
ebfd146a | 394 | case vect_used_in_outer: |
7c5222ff | 395 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 396 | relevant = vect_used_in_scope; |
ebfd146a | 397 | break; |
7c5222ff | 398 | |
8644a673 | 399 | case vect_used_in_scope: |
ebfd146a IR |
400 | break; |
401 | ||
402 | default: | |
403 | gcc_unreachable (); | |
b8698a0f | 404 | } |
ebfd146a IR |
405 | } |
406 | ||
407 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
408 | outer-loop-header-bb: | |
409 | ... | |
410 | inner-loop: | |
411 | d = def_stmt | |
06066f92 | 412 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
413 | stmt # use (d) */ |
414 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
415 | { | |
416 | if (vect_print_dump_info (REPORT_DETAILS)) | |
417 | fprintf (vect_dump, "inner-loop def-stmt defining outer-loop stmt."); | |
7c5222ff | 418 | |
ebfd146a IR |
419 | switch (relevant) |
420 | { | |
8644a673 | 421 | case vect_unused_in_scope: |
b8698a0f | 422 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 423 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 424 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
425 | break; |
426 | ||
ebfd146a IR |
427 | case vect_used_by_reduction: |
428 | relevant = vect_used_in_outer_by_reduction; | |
429 | break; | |
430 | ||
8644a673 | 431 | case vect_used_in_scope: |
ebfd146a IR |
432 | relevant = vect_used_in_outer; |
433 | break; | |
434 | ||
435 | default: | |
436 | gcc_unreachable (); | |
437 | } | |
438 | } | |
439 | ||
440 | vect_mark_relevant (worklist, def_stmt, relevant, live_p); | |
441 | return true; | |
442 | } | |
443 | ||
444 | ||
445 | /* Function vect_mark_stmts_to_be_vectorized. | |
446 | ||
447 | Not all stmts in the loop need to be vectorized. For example: | |
448 | ||
449 | for i... | |
450 | for j... | |
451 | 1. T0 = i + j | |
452 | 2. T1 = a[T0] | |
453 | ||
454 | 3. j = j + 1 | |
455 | ||
456 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
457 | addressing of vectorized data-refs are handled differently. | |
458 | ||
459 | This pass detects such stmts. */ | |
460 | ||
461 | bool | |
462 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
463 | { | |
464 | VEC(gimple,heap) *worklist; | |
465 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
466 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
467 | unsigned int nbbs = loop->num_nodes; | |
468 | gimple_stmt_iterator si; | |
469 | gimple stmt; | |
470 | unsigned int i; | |
471 | stmt_vec_info stmt_vinfo; | |
472 | basic_block bb; | |
473 | gimple phi; | |
474 | bool live_p; | |
06066f92 IR |
475 | enum vect_relevant relevant, tmp_relevant; |
476 | enum vect_def_type def_type; | |
ebfd146a IR |
477 | |
478 | if (vect_print_dump_info (REPORT_DETAILS)) | |
479 | fprintf (vect_dump, "=== vect_mark_stmts_to_be_vectorized ==="); | |
480 | ||
481 | worklist = VEC_alloc (gimple, heap, 64); | |
482 | ||
483 | /* 1. Init worklist. */ | |
484 | for (i = 0; i < nbbs; i++) | |
485 | { | |
486 | bb = bbs[i]; | |
487 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 488 | { |
ebfd146a IR |
489 | phi = gsi_stmt (si); |
490 | if (vect_print_dump_info (REPORT_DETAILS)) | |
491 | { | |
492 | fprintf (vect_dump, "init: phi relevant? "); | |
493 | print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM); | |
494 | } | |
495 | ||
496 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
497 | vect_mark_relevant (&worklist, phi, relevant, live_p); | |
498 | } | |
499 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
500 | { | |
501 | stmt = gsi_stmt (si); | |
502 | if (vect_print_dump_info (REPORT_DETAILS)) | |
503 | { | |
504 | fprintf (vect_dump, "init: stmt relevant? "); | |
505 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
b8698a0f | 506 | } |
ebfd146a IR |
507 | |
508 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
509 | vect_mark_relevant (&worklist, stmt, relevant, live_p); | |
510 | } | |
511 | } | |
512 | ||
513 | /* 2. Process_worklist */ | |
514 | while (VEC_length (gimple, worklist) > 0) | |
515 | { | |
516 | use_operand_p use_p; | |
517 | ssa_op_iter iter; | |
518 | ||
519 | stmt = VEC_pop (gimple, worklist); | |
520 | if (vect_print_dump_info (REPORT_DETAILS)) | |
521 | { | |
522 | fprintf (vect_dump, "worklist: examine stmt: "); | |
523 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
524 | } | |
525 | ||
b8698a0f L |
526 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
527 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
528 | liveness and relevance properties of STMT. */ |
529 | stmt_vinfo = vinfo_for_stmt (stmt); | |
530 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
531 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
532 | ||
533 | /* Generally, the liveness and relevance properties of STMT are | |
534 | propagated as is to the DEF_STMTs of its USEs: | |
535 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
536 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
537 | ||
538 | One exception is when STMT has been identified as defining a reduction | |
539 | variable; in this case we set the liveness/relevance as follows: | |
540 | live_p = false | |
541 | relevant = vect_used_by_reduction | |
542 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f | 543 | those that are used by a reduction computation, and those that are |
ff802fa1 | 544 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 545 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 546 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 547 | |
06066f92 IR |
548 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
549 | tmp_relevant = relevant; | |
550 | switch (def_type) | |
ebfd146a | 551 | { |
06066f92 IR |
552 | case vect_reduction_def: |
553 | switch (tmp_relevant) | |
554 | { | |
555 | case vect_unused_in_scope: | |
556 | relevant = vect_used_by_reduction; | |
557 | break; | |
558 | ||
559 | case vect_used_by_reduction: | |
560 | if (gimple_code (stmt) == GIMPLE_PHI) | |
561 | break; | |
562 | /* fall through */ | |
563 | ||
564 | default: | |
565 | if (vect_print_dump_info (REPORT_DETAILS)) | |
566 | fprintf (vect_dump, "unsupported use of reduction."); | |
567 | ||
568 | VEC_free (gimple, heap, worklist); | |
569 | return false; | |
570 | } | |
571 | ||
b8698a0f | 572 | live_p = false; |
06066f92 | 573 | break; |
b8698a0f | 574 | |
06066f92 IR |
575 | case vect_nested_cycle: |
576 | if (tmp_relevant != vect_unused_in_scope | |
577 | && tmp_relevant != vect_used_in_outer_by_reduction | |
578 | && tmp_relevant != vect_used_in_outer) | |
579 | { | |
580 | if (vect_print_dump_info (REPORT_DETAILS)) | |
581 | fprintf (vect_dump, "unsupported use of nested cycle."); | |
7c5222ff | 582 | |
06066f92 IR |
583 | VEC_free (gimple, heap, worklist); |
584 | return false; | |
585 | } | |
7c5222ff | 586 | |
b8698a0f L |
587 | live_p = false; |
588 | break; | |
589 | ||
06066f92 IR |
590 | case vect_double_reduction_def: |
591 | if (tmp_relevant != vect_unused_in_scope | |
592 | && tmp_relevant != vect_used_by_reduction) | |
593 | { | |
7c5222ff | 594 | if (vect_print_dump_info (REPORT_DETAILS)) |
06066f92 | 595 | fprintf (vect_dump, "unsupported use of double reduction."); |
7c5222ff IR |
596 | |
597 | VEC_free (gimple, heap, worklist); | |
598 | return false; | |
06066f92 IR |
599 | } |
600 | ||
601 | live_p = false; | |
b8698a0f | 602 | break; |
7c5222ff | 603 | |
06066f92 IR |
604 | default: |
605 | break; | |
7c5222ff | 606 | } |
b8698a0f | 607 | |
ebfd146a IR |
608 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) |
609 | { | |
610 | tree op = USE_FROM_PTR (use_p); | |
611 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, &worklist)) | |
612 | { | |
613 | VEC_free (gimple, heap, worklist); | |
614 | return false; | |
615 | } | |
616 | } | |
617 | } /* while worklist */ | |
618 | ||
619 | VEC_free (gimple, heap, worklist); | |
620 | return true; | |
621 | } | |
622 | ||
623 | ||
720f5239 IR |
624 | /* Get cost by calling cost target builtin. */ |
625 | ||
626 | static inline | |
627 | int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost) | |
628 | { | |
629 | tree dummy_type = NULL; | |
630 | int dummy = 0; | |
631 | ||
632 | return targetm.vectorize.builtin_vectorization_cost (type_of_cost, | |
633 | dummy_type, dummy); | |
634 | } | |
635 | ||
ff802fa1 IR |
636 | |
637 | /* Get cost for STMT. */ | |
638 | ||
ebfd146a IR |
639 | int |
640 | cost_for_stmt (gimple stmt) | |
641 | { | |
642 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
643 | ||
644 | switch (STMT_VINFO_TYPE (stmt_info)) | |
645 | { | |
646 | case load_vec_info_type: | |
720f5239 | 647 | return vect_get_stmt_cost (scalar_load); |
ebfd146a | 648 | case store_vec_info_type: |
720f5239 | 649 | return vect_get_stmt_cost (scalar_store); |
ebfd146a IR |
650 | case op_vec_info_type: |
651 | case condition_vec_info_type: | |
652 | case assignment_vec_info_type: | |
653 | case reduc_vec_info_type: | |
654 | case induc_vec_info_type: | |
655 | case type_promotion_vec_info_type: | |
656 | case type_demotion_vec_info_type: | |
657 | case type_conversion_vec_info_type: | |
658 | case call_vec_info_type: | |
720f5239 | 659 | return vect_get_stmt_cost (scalar_stmt); |
ebfd146a IR |
660 | case undef_vec_info_type: |
661 | default: | |
662 | gcc_unreachable (); | |
663 | } | |
664 | } | |
665 | ||
b8698a0f | 666 | /* Function vect_model_simple_cost. |
ebfd146a | 667 | |
b8698a0f | 668 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
669 | single op. Right now, this does not account for multiple insns that could |
670 | be generated for the single vector op. We will handle that shortly. */ | |
671 | ||
672 | void | |
b8698a0f | 673 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
ebfd146a IR |
674 | enum vect_def_type *dt, slp_tree slp_node) |
675 | { | |
676 | int i; | |
677 | int inside_cost = 0, outside_cost = 0; | |
678 | ||
679 | /* The SLP costs were already calculated during SLP tree build. */ | |
680 | if (PURE_SLP_STMT (stmt_info)) | |
681 | return; | |
682 | ||
720f5239 | 683 | inside_cost = ncopies * vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
684 | |
685 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
686 | for (i = 0; i < 2; i++) | |
687 | { | |
8644a673 | 688 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
720f5239 | 689 | outside_cost += vect_get_stmt_cost (vector_stmt); |
ebfd146a | 690 | } |
b8698a0f | 691 | |
ebfd146a IR |
692 | if (vect_print_dump_info (REPORT_COST)) |
693 | fprintf (vect_dump, "vect_model_simple_cost: inside_cost = %d, " | |
694 | "outside_cost = %d .", inside_cost, outside_cost); | |
695 | ||
696 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
697 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
698 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
699 | } | |
700 | ||
701 | ||
b8698a0f L |
702 | /* Function vect_cost_strided_group_size |
703 | ||
ebfd146a IR |
704 | For strided load or store, return the group_size only if it is the first |
705 | load or store of a group, else return 1. This ensures that group size is | |
706 | only returned once per group. */ | |
707 | ||
708 | static int | |
709 | vect_cost_strided_group_size (stmt_vec_info stmt_info) | |
710 | { | |
e14c1050 | 711 | gimple first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
712 | |
713 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
e14c1050 | 714 | return GROUP_SIZE (stmt_info); |
ebfd146a IR |
715 | |
716 | return 1; | |
717 | } | |
718 | ||
719 | ||
720 | /* Function vect_model_store_cost | |
721 | ||
722 | Models cost for stores. In the case of strided accesses, one access | |
723 | has the overhead of the strided access attributed to it. */ | |
724 | ||
725 | void | |
b8698a0f | 726 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
272c6793 RS |
727 | bool store_lanes_p, enum vect_def_type dt, |
728 | slp_tree slp_node) | |
ebfd146a IR |
729 | { |
730 | int group_size; | |
720f5239 IR |
731 | unsigned int inside_cost = 0, outside_cost = 0; |
732 | struct data_reference *first_dr; | |
733 | gimple first_stmt; | |
ebfd146a IR |
734 | |
735 | /* The SLP costs were already calculated during SLP tree build. */ | |
736 | if (PURE_SLP_STMT (stmt_info)) | |
737 | return; | |
738 | ||
8644a673 | 739 | if (dt == vect_constant_def || dt == vect_external_def) |
720f5239 | 740 | outside_cost = vect_get_stmt_cost (scalar_to_vec); |
ebfd146a IR |
741 | |
742 | /* Strided access? */ | |
e14c1050 | 743 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) |
720f5239 IR |
744 | { |
745 | if (slp_node) | |
746 | { | |
747 | first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0); | |
748 | group_size = 1; | |
749 | } | |
750 | else | |
751 | { | |
e14c1050 | 752 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
720f5239 IR |
753 | group_size = vect_cost_strided_group_size (stmt_info); |
754 | } | |
755 | ||
756 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
757 | } | |
ebfd146a IR |
758 | /* Not a strided access. */ |
759 | else | |
720f5239 IR |
760 | { |
761 | group_size = 1; | |
762 | first_dr = STMT_VINFO_DATA_REF (stmt_info); | |
763 | } | |
ebfd146a | 764 | |
272c6793 RS |
765 | /* We assume that the cost of a single store-lanes instruction is |
766 | equivalent to the cost of GROUP_SIZE separate stores. If a strided | |
767 | access is instead being provided by a permute-and-store operation, | |
768 | include the cost of the permutes. */ | |
769 | if (!store_lanes_p && group_size > 1) | |
ebfd146a IR |
770 | { |
771 | /* Uses a high and low interleave operation for each needed permute. */ | |
b8698a0f | 772 | inside_cost = ncopies * exact_log2(group_size) * group_size |
720f5239 | 773 | * vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
774 | |
775 | if (vect_print_dump_info (REPORT_COST)) | |
776 | fprintf (vect_dump, "vect_model_store_cost: strided group_size = %d .", | |
777 | group_size); | |
778 | ||
779 | } | |
780 | ||
781 | /* Costs of the stores. */ | |
720f5239 | 782 | vect_get_store_cost (first_dr, ncopies, &inside_cost); |
ebfd146a IR |
783 | |
784 | if (vect_print_dump_info (REPORT_COST)) | |
785 | fprintf (vect_dump, "vect_model_store_cost: inside_cost = %d, " | |
786 | "outside_cost = %d .", inside_cost, outside_cost); | |
787 | ||
788 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
789 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
790 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
791 | } | |
792 | ||
793 | ||
720f5239 IR |
794 | /* Calculate cost of DR's memory access. */ |
795 | void | |
796 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
797 | unsigned int *inside_cost) | |
798 | { | |
799 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
800 | ||
801 | switch (alignment_support_scheme) | |
802 | { | |
803 | case dr_aligned: | |
804 | { | |
805 | *inside_cost += ncopies * vect_get_stmt_cost (vector_store); | |
806 | ||
807 | if (vect_print_dump_info (REPORT_COST)) | |
808 | fprintf (vect_dump, "vect_model_store_cost: aligned."); | |
809 | ||
810 | break; | |
811 | } | |
812 | ||
813 | case dr_unaligned_supported: | |
814 | { | |
815 | gimple stmt = DR_STMT (dr); | |
816 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
817 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
818 | ||
819 | /* Here, we assign an additional cost for the unaligned store. */ | |
820 | *inside_cost += ncopies | |
821 | * targetm.vectorize.builtin_vectorization_cost (unaligned_store, | |
822 | vectype, DR_MISALIGNMENT (dr)); | |
823 | ||
824 | if (vect_print_dump_info (REPORT_COST)) | |
825 | fprintf (vect_dump, "vect_model_store_cost: unaligned supported by " | |
826 | "hardware."); | |
827 | ||
828 | break; | |
829 | } | |
830 | ||
831 | default: | |
832 | gcc_unreachable (); | |
833 | } | |
834 | } | |
835 | ||
836 | ||
ebfd146a IR |
837 | /* Function vect_model_load_cost |
838 | ||
839 | Models cost for loads. In the case of strided accesses, the last access | |
840 | has the overhead of the strided access attributed to it. Since unaligned | |
b8698a0f | 841 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
842 | access scheme chosen. */ |
843 | ||
844 | void | |
272c6793 RS |
845 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, bool load_lanes_p, |
846 | slp_tree slp_node) | |
ebfd146a IR |
847 | { |
848 | int group_size; | |
ebfd146a IR |
849 | gimple first_stmt; |
850 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; | |
720f5239 | 851 | unsigned int inside_cost = 0, outside_cost = 0; |
ebfd146a IR |
852 | |
853 | /* The SLP costs were already calculated during SLP tree build. */ | |
854 | if (PURE_SLP_STMT (stmt_info)) | |
855 | return; | |
856 | ||
857 | /* Strided accesses? */ | |
e14c1050 IR |
858 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
859 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) && first_stmt && !slp_node) | |
ebfd146a IR |
860 | { |
861 | group_size = vect_cost_strided_group_size (stmt_info); | |
862 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
863 | } | |
864 | /* Not a strided access. */ | |
865 | else | |
866 | { | |
867 | group_size = 1; | |
868 | first_dr = dr; | |
869 | } | |
870 | ||
272c6793 RS |
871 | /* We assume that the cost of a single load-lanes instruction is |
872 | equivalent to the cost of GROUP_SIZE separate loads. If a strided | |
873 | access is instead being provided by a load-and-permute operation, | |
874 | include the cost of the permutes. */ | |
875 | if (!load_lanes_p && group_size > 1) | |
ebfd146a IR |
876 | { |
877 | /* Uses an even and odd extract operations for each needed permute. */ | |
878 | inside_cost = ncopies * exact_log2(group_size) * group_size | |
720f5239 | 879 | * vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
880 | |
881 | if (vect_print_dump_info (REPORT_COST)) | |
882 | fprintf (vect_dump, "vect_model_load_cost: strided group_size = %d .", | |
883 | group_size); | |
ebfd146a IR |
884 | } |
885 | ||
886 | /* The loads themselves. */ | |
720f5239 | 887 | vect_get_load_cost (first_dr, ncopies, |
e14c1050 IR |
888 | ((!STMT_VINFO_STRIDED_ACCESS (stmt_info)) || group_size > 1 |
889 | || slp_node), | |
720f5239 IR |
890 | &inside_cost, &outside_cost); |
891 | ||
892 | if (vect_print_dump_info (REPORT_COST)) | |
893 | fprintf (vect_dump, "vect_model_load_cost: inside_cost = %d, " | |
894 | "outside_cost = %d .", inside_cost, outside_cost); | |
895 | ||
896 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
897 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
898 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
899 | } | |
900 | ||
901 | ||
902 | /* Calculate cost of DR's memory access. */ | |
903 | void | |
904 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
905 | bool add_realign_cost, unsigned int *inside_cost, | |
906 | unsigned int *outside_cost) | |
907 | { | |
908 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
909 | ||
910 | switch (alignment_support_scheme) | |
ebfd146a IR |
911 | { |
912 | case dr_aligned: | |
913 | { | |
9940b13c | 914 | *inside_cost += ncopies * vect_get_stmt_cost (vector_load); |
ebfd146a IR |
915 | |
916 | if (vect_print_dump_info (REPORT_COST)) | |
917 | fprintf (vect_dump, "vect_model_load_cost: aligned."); | |
918 | ||
919 | break; | |
920 | } | |
921 | case dr_unaligned_supported: | |
922 | { | |
720f5239 IR |
923 | gimple stmt = DR_STMT (dr); |
924 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
925 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
ebfd146a | 926 | |
720f5239 IR |
927 | /* Here, we assign an additional cost for the unaligned load. */ |
928 | *inside_cost += ncopies | |
929 | * targetm.vectorize.builtin_vectorization_cost (unaligned_load, | |
930 | vectype, DR_MISALIGNMENT (dr)); | |
ebfd146a IR |
931 | if (vect_print_dump_info (REPORT_COST)) |
932 | fprintf (vect_dump, "vect_model_load_cost: unaligned supported by " | |
933 | "hardware."); | |
934 | ||
935 | break; | |
936 | } | |
937 | case dr_explicit_realign: | |
938 | { | |
720f5239 IR |
939 | *inside_cost += ncopies * (2 * vect_get_stmt_cost (vector_load) |
940 | + vect_get_stmt_cost (vector_stmt)); | |
ebfd146a IR |
941 | |
942 | /* FIXME: If the misalignment remains fixed across the iterations of | |
943 | the containing loop, the following cost should be added to the | |
944 | outside costs. */ | |
945 | if (targetm.vectorize.builtin_mask_for_load) | |
720f5239 | 946 | *inside_cost += vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
947 | |
948 | break; | |
949 | } | |
950 | case dr_explicit_realign_optimized: | |
951 | { | |
952 | if (vect_print_dump_info (REPORT_COST)) | |
953 | fprintf (vect_dump, "vect_model_load_cost: unaligned software " | |
954 | "pipelined."); | |
955 | ||
956 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 957 | load, and possibly a mask operation to "prime" the loop. However, |
ebfd146a IR |
958 | if this is an access in a group of loads, which provide strided |
959 | access, then the above cost should only be considered for one | |
ff802fa1 | 960 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
961 | and a realignment op. */ |
962 | ||
720f5239 | 963 | if (add_realign_cost) |
ebfd146a | 964 | { |
720f5239 | 965 | *outside_cost = 2 * vect_get_stmt_cost (vector_stmt); |
ebfd146a | 966 | if (targetm.vectorize.builtin_mask_for_load) |
720f5239 | 967 | *outside_cost += vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
968 | } |
969 | ||
720f5239 IR |
970 | *inside_cost += ncopies * (vect_get_stmt_cost (vector_load) |
971 | + vect_get_stmt_cost (vector_stmt)); | |
ebfd146a IR |
972 | break; |
973 | } | |
974 | ||
975 | default: | |
976 | gcc_unreachable (); | |
977 | } | |
ebfd146a IR |
978 | } |
979 | ||
980 | ||
981 | /* Function vect_init_vector. | |
982 | ||
983 | Insert a new stmt (INIT_STMT) that initializes a new vector variable with | |
ff802fa1 IR |
984 | the vector elements of VECTOR_VAR. Place the initialization at BSI if it |
985 | is not NULL. Otherwise, place the initialization at the loop preheader. | |
b8698a0f | 986 | Return the DEF of INIT_STMT. |
ebfd146a IR |
987 | It will be used in the vectorization of STMT. */ |
988 | ||
989 | tree | |
990 | vect_init_vector (gimple stmt, tree vector_var, tree vector_type, | |
991 | gimple_stmt_iterator *gsi) | |
992 | { | |
993 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
994 | tree new_var; | |
995 | gimple init_stmt; | |
996 | tree vec_oprnd; | |
997 | edge pe; | |
998 | tree new_temp; | |
999 | basic_block new_bb; | |
b8698a0f | 1000 | |
ebfd146a | 1001 | new_var = vect_get_new_vect_var (vector_type, vect_simple_var, "cst_"); |
b8698a0f | 1002 | add_referenced_var (new_var); |
ebfd146a IR |
1003 | init_stmt = gimple_build_assign (new_var, vector_var); |
1004 | new_temp = make_ssa_name (new_var, init_stmt); | |
1005 | gimple_assign_set_lhs (init_stmt, new_temp); | |
1006 | ||
1007 | if (gsi) | |
1008 | vect_finish_stmt_generation (stmt, init_stmt, gsi); | |
1009 | else | |
1010 | { | |
1011 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
b8698a0f | 1012 | |
a70d6342 IR |
1013 | if (loop_vinfo) |
1014 | { | |
1015 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1016 | ||
1017 | if (nested_in_vect_loop_p (loop, stmt)) | |
1018 | loop = loop->inner; | |
b8698a0f | 1019 | |
a70d6342 IR |
1020 | pe = loop_preheader_edge (loop); |
1021 | new_bb = gsi_insert_on_edge_immediate (pe, init_stmt); | |
1022 | gcc_assert (!new_bb); | |
1023 | } | |
1024 | else | |
1025 | { | |
1026 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1027 | basic_block bb; | |
1028 | gimple_stmt_iterator gsi_bb_start; | |
1029 | ||
1030 | gcc_assert (bb_vinfo); | |
1031 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1032 | gsi_bb_start = gsi_after_labels (bb); |
a70d6342 IR |
1033 | gsi_insert_before (&gsi_bb_start, init_stmt, GSI_SAME_STMT); |
1034 | } | |
ebfd146a IR |
1035 | } |
1036 | ||
1037 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1038 | { | |
1039 | fprintf (vect_dump, "created new init_stmt: "); | |
1040 | print_gimple_stmt (vect_dump, init_stmt, 0, TDF_SLIM); | |
1041 | } | |
1042 | ||
1043 | vec_oprnd = gimple_assign_lhs (init_stmt); | |
1044 | return vec_oprnd; | |
1045 | } | |
1046 | ||
a70d6342 | 1047 | |
ebfd146a IR |
1048 | /* Function vect_get_vec_def_for_operand. |
1049 | ||
ff802fa1 | 1050 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
1051 | used in the vectorized stmt for STMT. |
1052 | ||
1053 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1054 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1055 | ||
1056 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1057 | needs to be introduced. */ | |
1058 | ||
1059 | tree | |
1060 | vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def) | |
1061 | { | |
1062 | tree vec_oprnd; | |
1063 | gimple vec_stmt; | |
1064 | gimple def_stmt; | |
1065 | stmt_vec_info def_stmt_info = NULL; | |
1066 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
9dc3f7de | 1067 | unsigned int nunits; |
ebfd146a IR |
1068 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
1069 | tree vec_inv; | |
1070 | tree vec_cst; | |
e7e9eb2f | 1071 | tree t = NULL_TREE; |
ebfd146a | 1072 | tree def; |
e7e9eb2f | 1073 | int i; |
ebfd146a IR |
1074 | enum vect_def_type dt; |
1075 | bool is_simple_use; | |
1076 | tree vector_type; | |
1077 | ||
1078 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1079 | { | |
1080 | fprintf (vect_dump, "vect_get_vec_def_for_operand: "); | |
1081 | print_generic_expr (vect_dump, op, TDF_SLIM); | |
1082 | } | |
1083 | ||
b8698a0f | 1084 | is_simple_use = vect_is_simple_use (op, loop_vinfo, NULL, &def_stmt, &def, |
a70d6342 | 1085 | &dt); |
ebfd146a IR |
1086 | gcc_assert (is_simple_use); |
1087 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1088 | { | |
1089 | if (def) | |
1090 | { | |
1091 | fprintf (vect_dump, "def = "); | |
1092 | print_generic_expr (vect_dump, def, TDF_SLIM); | |
1093 | } | |
1094 | if (def_stmt) | |
1095 | { | |
1096 | fprintf (vect_dump, " def_stmt = "); | |
1097 | print_gimple_stmt (vect_dump, def_stmt, 0, TDF_SLIM); | |
1098 | } | |
1099 | } | |
1100 | ||
1101 | switch (dt) | |
1102 | { | |
1103 | /* Case 1: operand is a constant. */ | |
1104 | case vect_constant_def: | |
1105 | { | |
7569a6cc RG |
1106 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
1107 | gcc_assert (vector_type); | |
9dc3f7de | 1108 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); |
7569a6cc | 1109 | |
b8698a0f | 1110 | if (scalar_def) |
ebfd146a IR |
1111 | *scalar_def = op; |
1112 | ||
1113 | /* Create 'vect_cst_ = {cst,cst,...,cst}' */ | |
1114 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1115 | fprintf (vect_dump, "Create vector_cst. nunits = %d", nunits); | |
1116 | ||
b9acc9f1 | 1117 | vec_cst = build_vector_from_val (vector_type, op); |
7569a6cc | 1118 | return vect_init_vector (stmt, vec_cst, vector_type, NULL); |
ebfd146a IR |
1119 | } |
1120 | ||
1121 | /* Case 2: operand is defined outside the loop - loop invariant. */ | |
8644a673 | 1122 | case vect_external_def: |
ebfd146a IR |
1123 | { |
1124 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); | |
1125 | gcc_assert (vector_type); | |
1126 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); | |
1127 | ||
b8698a0f | 1128 | if (scalar_def) |
ebfd146a IR |
1129 | *scalar_def = def; |
1130 | ||
1131 | /* Create 'vec_inv = {inv,inv,..,inv}' */ | |
1132 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1133 | fprintf (vect_dump, "Create vector_inv."); | |
1134 | ||
e7e9eb2f NF |
1135 | for (i = nunits - 1; i >= 0; --i) |
1136 | { | |
1137 | t = tree_cons (NULL_TREE, def, t); | |
1138 | } | |
1139 | ||
1140 | /* FIXME: use build_constructor directly. */ | |
1141 | vec_inv = build_constructor_from_list (vector_type, t); | |
ebfd146a IR |
1142 | return vect_init_vector (stmt, vec_inv, vector_type, NULL); |
1143 | } | |
1144 | ||
1145 | /* Case 3: operand is defined inside the loop. */ | |
8644a673 | 1146 | case vect_internal_def: |
ebfd146a | 1147 | { |
b8698a0f | 1148 | if (scalar_def) |
ebfd146a IR |
1149 | *scalar_def = NULL/* FIXME tuples: def_stmt*/; |
1150 | ||
1151 | /* Get the def from the vectorized stmt. */ | |
1152 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1153 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
1154 | gcc_assert (vec_stmt); | |
1155 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1156 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1157 | else if (is_gimple_call (vec_stmt)) | |
1158 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1159 | else | |
1160 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1161 | return vec_oprnd; | |
1162 | } | |
1163 | ||
1164 | /* Case 4: operand is defined by a loop header phi - reduction */ | |
1165 | case vect_reduction_def: | |
06066f92 | 1166 | case vect_double_reduction_def: |
7c5222ff | 1167 | case vect_nested_cycle: |
ebfd146a IR |
1168 | { |
1169 | struct loop *loop; | |
1170 | ||
1171 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
b8698a0f | 1172 | loop = (gimple_bb (def_stmt))->loop_father; |
ebfd146a IR |
1173 | |
1174 | /* Get the def before the loop */ | |
1175 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); | |
1176 | return get_initial_def_for_reduction (stmt, op, scalar_def); | |
1177 | } | |
1178 | ||
1179 | /* Case 5: operand is defined by loop-header phi - induction. */ | |
1180 | case vect_induction_def: | |
1181 | { | |
1182 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1183 | ||
1184 | /* Get the def from the vectorized stmt. */ | |
1185 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1186 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1187 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1188 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1189 | else | |
1190 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1191 | return vec_oprnd; |
1192 | } | |
1193 | ||
1194 | default: | |
1195 | gcc_unreachable (); | |
1196 | } | |
1197 | } | |
1198 | ||
1199 | ||
1200 | /* Function vect_get_vec_def_for_stmt_copy | |
1201 | ||
ff802fa1 | 1202 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1203 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1204 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1205 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1206 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1207 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1208 | DT is the type of the vector def VEC_OPRND. |
1209 | ||
1210 | Context: | |
1211 | In case the vectorization factor (VF) is bigger than the number | |
1212 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1213 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1214 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1215 | smallest data-type determines the VF, and as a result, when vectorizing |
1216 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1217 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1218 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1219 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1220 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1221 | ||
1222 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1223 | |
ebfd146a IR |
1224 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1225 | VS1.1: vx.1 = memref1 VS1.2 | |
1226 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1227 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1228 | |
1229 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1230 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1231 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1232 | VSnew.3: vz3 = vx.3 + ... | |
1233 | ||
1234 | The vectorization of S1 is explained in vectorizable_load. | |
1235 | The vectorization of S2: | |
b8698a0f L |
1236 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1237 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1238 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1239 | returns the vector-def 'vx.0'. |
1240 | ||
b8698a0f L |
1241 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1242 | function is called to get the relevant vector-def for each operand. It is | |
1243 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1244 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1245 | ||
b8698a0f L |
1246 | For example, to obtain the vector-def 'vx.1' in order to create the |
1247 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1248 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1249 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1250 | and return its def ('vx.1'). | |
1251 | Overall, to create the above sequence this function will be called 3 times: | |
1252 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1253 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1254 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1255 | ||
1256 | tree | |
1257 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1258 | { | |
1259 | gimple vec_stmt_for_operand; | |
1260 | stmt_vec_info def_stmt_info; | |
1261 | ||
1262 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1263 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1264 | return vec_oprnd; |
1265 | ||
1266 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1267 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1268 | gcc_assert (def_stmt_info); | |
1269 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1270 | gcc_assert (vec_stmt_for_operand); | |
1271 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1272 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) | |
1273 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1274 | else | |
1275 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1276 | return vec_oprnd; | |
1277 | } | |
1278 | ||
1279 | ||
1280 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1281 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1282 | |
1283 | static void | |
b8698a0f L |
1284 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
1285 | VEC(tree,heap) **vec_oprnds0, | |
ebfd146a IR |
1286 | VEC(tree,heap) **vec_oprnds1) |
1287 | { | |
1288 | tree vec_oprnd = VEC_pop (tree, *vec_oprnds0); | |
1289 | ||
1290 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
1291 | VEC_quick_push (tree, *vec_oprnds0, vec_oprnd); | |
1292 | ||
1293 | if (vec_oprnds1 && *vec_oprnds1) | |
1294 | { | |
1295 | vec_oprnd = VEC_pop (tree, *vec_oprnds1); | |
1296 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); | |
1297 | VEC_quick_push (tree, *vec_oprnds1, vec_oprnd); | |
1298 | } | |
1299 | } | |
1300 | ||
1301 | ||
ff802fa1 IR |
1302 | /* Get vectorized definitions for OP0 and OP1, or SLP_NODE if it is not |
1303 | NULL. */ | |
ebfd146a IR |
1304 | |
1305 | static void | |
1306 | vect_get_vec_defs (tree op0, tree op1, gimple stmt, | |
1307 | VEC(tree,heap) **vec_oprnds0, VEC(tree,heap) **vec_oprnds1, | |
1308 | slp_tree slp_node) | |
1309 | { | |
1310 | if (slp_node) | |
9dc3f7de | 1311 | vect_get_slp_defs (op0, op1, slp_node, vec_oprnds0, vec_oprnds1, -1); |
ebfd146a IR |
1312 | else |
1313 | { | |
1314 | tree vec_oprnd; | |
1315 | ||
b8698a0f L |
1316 | *vec_oprnds0 = VEC_alloc (tree, heap, 1); |
1317 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
ebfd146a IR |
1318 | VEC_quick_push (tree, *vec_oprnds0, vec_oprnd); |
1319 | ||
1320 | if (op1) | |
1321 | { | |
b8698a0f L |
1322 | *vec_oprnds1 = VEC_alloc (tree, heap, 1); |
1323 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL); | |
ebfd146a IR |
1324 | VEC_quick_push (tree, *vec_oprnds1, vec_oprnd); |
1325 | } | |
1326 | } | |
1327 | } | |
1328 | ||
1329 | ||
1330 | /* Function vect_finish_stmt_generation. | |
1331 | ||
1332 | Insert a new stmt. */ | |
1333 | ||
1334 | void | |
1335 | vect_finish_stmt_generation (gimple stmt, gimple vec_stmt, | |
1336 | gimple_stmt_iterator *gsi) | |
1337 | { | |
1338 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1339 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1340 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1341 | |
1342 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1343 | ||
1344 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); | |
1345 | ||
b8698a0f | 1346 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo, |
a70d6342 | 1347 | bb_vinfo)); |
ebfd146a IR |
1348 | |
1349 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1350 | { | |
1351 | fprintf (vect_dump, "add new stmt: "); | |
1352 | print_gimple_stmt (vect_dump, vec_stmt, 0, TDF_SLIM); | |
1353 | } | |
1354 | ||
1355 | gimple_set_location (vec_stmt, gimple_location (gsi_stmt (*gsi))); | |
1356 | } | |
1357 | ||
1358 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1359 | a function declaration if the target has a vectorized version | |
1360 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1361 | ||
1362 | tree | |
1363 | vectorizable_function (gimple call, tree vectype_out, tree vectype_in) | |
1364 | { | |
1365 | tree fndecl = gimple_call_fndecl (call); | |
ebfd146a IR |
1366 | |
1367 | /* We only handle functions that do not read or clobber memory -- i.e. | |
1368 | const or novops ones. */ | |
1369 | if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS))) | |
1370 | return NULL_TREE; | |
1371 | ||
1372 | if (!fndecl | |
1373 | || TREE_CODE (fndecl) != FUNCTION_DECL | |
1374 | || !DECL_BUILT_IN (fndecl)) | |
1375 | return NULL_TREE; | |
1376 | ||
62f7fd21 | 1377 | return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out, |
ebfd146a IR |
1378 | vectype_in); |
1379 | } | |
1380 | ||
1381 | /* Function vectorizable_call. | |
1382 | ||
b8698a0f L |
1383 | Check if STMT performs a function call that can be vectorized. |
1384 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1385 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1386 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1387 | ||
1388 | static bool | |
1389 | vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt) | |
1390 | { | |
1391 | tree vec_dest; | |
1392 | tree scalar_dest; | |
1393 | tree op, type; | |
1394 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1395 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
1396 | tree vectype_out, vectype_in; | |
1397 | int nunits_in; | |
1398 | int nunits_out; | |
1399 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
b690cc0f | 1400 | tree fndecl, new_temp, def, rhs_type; |
ebfd146a | 1401 | gimple def_stmt; |
0502fb85 UB |
1402 | enum vect_def_type dt[3] |
1403 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
63827fb8 | 1404 | gimple new_stmt = NULL; |
ebfd146a IR |
1405 | int ncopies, j; |
1406 | VEC(tree, heap) *vargs = NULL; | |
1407 | enum { NARROW, NONE, WIDEN } modifier; | |
1408 | size_t i, nargs; | |
1409 | ||
a70d6342 IR |
1410 | /* FORNOW: unsupported in basic block SLP. */ |
1411 | gcc_assert (loop_vinfo); | |
b8698a0f | 1412 | |
ebfd146a IR |
1413 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
1414 | return false; | |
1415 | ||
8644a673 | 1416 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1417 | return false; |
1418 | ||
1419 | /* FORNOW: SLP not supported. */ | |
1420 | if (STMT_SLP_TYPE (stmt_info)) | |
1421 | return false; | |
1422 | ||
1423 | /* Is STMT a vectorizable call? */ | |
1424 | if (!is_gimple_call (stmt)) | |
1425 | return false; | |
1426 | ||
1427 | if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
1428 | return false; | |
1429 | ||
822ba6d7 | 1430 | if (stmt_can_throw_internal (stmt)) |
5a2c1986 IR |
1431 | return false; |
1432 | ||
b690cc0f RG |
1433 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
1434 | ||
ebfd146a IR |
1435 | /* Process function arguments. */ |
1436 | rhs_type = NULL_TREE; | |
b690cc0f | 1437 | vectype_in = NULL_TREE; |
ebfd146a IR |
1438 | nargs = gimple_call_num_args (stmt); |
1439 | ||
1b1562a5 MM |
1440 | /* Bail out if the function has more than three arguments, we do not have |
1441 | interesting builtin functions to vectorize with more than two arguments | |
1442 | except for fma. No arguments is also not good. */ | |
1443 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
1444 | return false; |
1445 | ||
1446 | for (i = 0; i < nargs; i++) | |
1447 | { | |
b690cc0f RG |
1448 | tree opvectype; |
1449 | ||
ebfd146a IR |
1450 | op = gimple_call_arg (stmt, i); |
1451 | ||
1452 | /* We can only handle calls with arguments of the same type. */ | |
1453 | if (rhs_type | |
8533c9d8 | 1454 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a IR |
1455 | { |
1456 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1457 | fprintf (vect_dump, "argument types differ."); | |
1458 | return false; | |
1459 | } | |
b690cc0f RG |
1460 | if (!rhs_type) |
1461 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 1462 | |
b690cc0f RG |
1463 | if (!vect_is_simple_use_1 (op, loop_vinfo, NULL, |
1464 | &def_stmt, &def, &dt[i], &opvectype)) | |
ebfd146a IR |
1465 | { |
1466 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1467 | fprintf (vect_dump, "use not simple."); | |
1468 | return false; | |
1469 | } | |
ebfd146a | 1470 | |
b690cc0f RG |
1471 | if (!vectype_in) |
1472 | vectype_in = opvectype; | |
1473 | else if (opvectype | |
1474 | && opvectype != vectype_in) | |
1475 | { | |
1476 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1477 | fprintf (vect_dump, "argument vector types differ."); | |
1478 | return false; | |
1479 | } | |
1480 | } | |
1481 | /* If all arguments are external or constant defs use a vector type with | |
1482 | the same size as the output vector type. */ | |
ebfd146a | 1483 | if (!vectype_in) |
b690cc0f | 1484 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
1485 | if (vec_stmt) |
1486 | gcc_assert (vectype_in); | |
1487 | if (!vectype_in) | |
1488 | { | |
1489 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1490 | { | |
1491 | fprintf (vect_dump, "no vectype for scalar type "); | |
1492 | print_generic_expr (vect_dump, rhs_type, TDF_SLIM); | |
1493 | } | |
1494 | ||
1495 | return false; | |
1496 | } | |
ebfd146a IR |
1497 | |
1498 | /* FORNOW */ | |
b690cc0f RG |
1499 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
1500 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
1501 | if (nunits_in == nunits_out / 2) |
1502 | modifier = NARROW; | |
1503 | else if (nunits_out == nunits_in) | |
1504 | modifier = NONE; | |
1505 | else if (nunits_out == nunits_in / 2) | |
1506 | modifier = WIDEN; | |
1507 | else | |
1508 | return false; | |
1509 | ||
1510 | /* For now, we only vectorize functions if a target specific builtin | |
1511 | is available. TODO -- in some cases, it might be profitable to | |
1512 | insert the calls for pieces of the vector, in order to be able | |
1513 | to vectorize other operations in the loop. */ | |
1514 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
1515 | if (fndecl == NULL_TREE) | |
1516 | { | |
1517 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1518 | fprintf (vect_dump, "function is not vectorizable."); | |
1519 | ||
1520 | return false; | |
1521 | } | |
1522 | ||
5006671f | 1523 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a IR |
1524 | |
1525 | if (modifier == NARROW) | |
1526 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
1527 | else | |
1528 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1529 | ||
1530 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
1531 | needs to be generated. */ | |
1532 | gcc_assert (ncopies >= 1); | |
1533 | ||
1534 | if (!vec_stmt) /* transformation not required. */ | |
1535 | { | |
1536 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1537 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1538 | fprintf (vect_dump, "=== vectorizable_call ==="); | |
1539 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
1540 | return true; | |
1541 | } | |
1542 | ||
1543 | /** Transform. **/ | |
1544 | ||
1545 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1546 | fprintf (vect_dump, "transform operation."); | |
1547 | ||
1548 | /* Handle def. */ | |
1549 | scalar_dest = gimple_call_lhs (stmt); | |
1550 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1551 | ||
1552 | prev_stmt_info = NULL; | |
1553 | switch (modifier) | |
1554 | { | |
1555 | case NONE: | |
1556 | for (j = 0; j < ncopies; ++j) | |
1557 | { | |
1558 | /* Build argument list for the vectorized call. */ | |
1559 | if (j == 0) | |
1560 | vargs = VEC_alloc (tree, heap, nargs); | |
1561 | else | |
1562 | VEC_truncate (tree, vargs, 0); | |
1563 | ||
1564 | for (i = 0; i < nargs; i++) | |
1565 | { | |
1566 | op = gimple_call_arg (stmt, i); | |
1567 | if (j == 0) | |
1568 | vec_oprnd0 | |
1569 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1570 | else | |
63827fb8 IR |
1571 | { |
1572 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
1573 | vec_oprnd0 | |
1574 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
1575 | } | |
ebfd146a IR |
1576 | |
1577 | VEC_quick_push (tree, vargs, vec_oprnd0); | |
1578 | } | |
1579 | ||
1580 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1581 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1582 | gimple_call_set_lhs (new_stmt, new_temp); | |
1583 | ||
1584 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7411b8f0 | 1585 | mark_symbols_for_renaming (new_stmt); |
ebfd146a IR |
1586 | |
1587 | if (j == 0) | |
1588 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1589 | else | |
1590 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1591 | ||
1592 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1593 | } | |
1594 | ||
1595 | break; | |
1596 | ||
1597 | case NARROW: | |
1598 | for (j = 0; j < ncopies; ++j) | |
1599 | { | |
1600 | /* Build argument list for the vectorized call. */ | |
1601 | if (j == 0) | |
1602 | vargs = VEC_alloc (tree, heap, nargs * 2); | |
1603 | else | |
1604 | VEC_truncate (tree, vargs, 0); | |
1605 | ||
1606 | for (i = 0; i < nargs; i++) | |
1607 | { | |
1608 | op = gimple_call_arg (stmt, i); | |
1609 | if (j == 0) | |
1610 | { | |
1611 | vec_oprnd0 | |
1612 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1613 | vec_oprnd1 | |
63827fb8 | 1614 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
1615 | } |
1616 | else | |
1617 | { | |
63827fb8 | 1618 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i); |
ebfd146a | 1619 | vec_oprnd0 |
63827fb8 | 1620 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 1621 | vec_oprnd1 |
63827fb8 | 1622 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
1623 | } |
1624 | ||
1625 | VEC_quick_push (tree, vargs, vec_oprnd0); | |
1626 | VEC_quick_push (tree, vargs, vec_oprnd1); | |
1627 | } | |
1628 | ||
1629 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1630 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1631 | gimple_call_set_lhs (new_stmt, new_temp); | |
1632 | ||
1633 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7411b8f0 | 1634 | mark_symbols_for_renaming (new_stmt); |
ebfd146a IR |
1635 | |
1636 | if (j == 0) | |
1637 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1638 | else | |
1639 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1640 | ||
1641 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1642 | } | |
1643 | ||
1644 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
1645 | ||
1646 | break; | |
1647 | ||
1648 | case WIDEN: | |
1649 | /* No current target implements this case. */ | |
1650 | return false; | |
1651 | } | |
1652 | ||
1653 | VEC_free (tree, heap, vargs); | |
1654 | ||
1655 | /* Update the exception handling table with the vector stmt if necessary. */ | |
1656 | if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt)) | |
1657 | gimple_purge_dead_eh_edges (gimple_bb (stmt)); | |
1658 | ||
1659 | /* The call in STMT might prevent it from being removed in dce. | |
1660 | We however cannot remove it here, due to the way the ssa name | |
1661 | it defines is mapped to the new definition. So just replace | |
1662 | rhs of the statement with something harmless. */ | |
1663 | ||
1664 | type = TREE_TYPE (scalar_dest); | |
1665 | new_stmt = gimple_build_assign (gimple_call_lhs (stmt), | |
e8160c9a | 1666 | build_zero_cst (type)); |
ebfd146a IR |
1667 | set_vinfo_for_stmt (new_stmt, stmt_info); |
1668 | set_vinfo_for_stmt (stmt, NULL); | |
1669 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
1670 | gsi_replace (gsi, new_stmt, false); | |
1671 | SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt; | |
1672 | ||
1673 | return true; | |
1674 | } | |
1675 | ||
1676 | ||
1677 | /* Function vect_gen_widened_results_half | |
1678 | ||
1679 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 1680 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 1681 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
1682 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
1683 | needs to be created (DECL is a function-decl of a target-builtin). | |
1684 | STMT is the original scalar stmt that we are vectorizing. */ | |
1685 | ||
1686 | static gimple | |
1687 | vect_gen_widened_results_half (enum tree_code code, | |
1688 | tree decl, | |
1689 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
1690 | tree vec_dest, gimple_stmt_iterator *gsi, | |
1691 | gimple stmt) | |
b8698a0f | 1692 | { |
ebfd146a | 1693 | gimple new_stmt; |
b8698a0f L |
1694 | tree new_temp; |
1695 | ||
1696 | /* Generate half of the widened result: */ | |
1697 | if (code == CALL_EXPR) | |
1698 | { | |
1699 | /* Target specific support */ | |
ebfd146a IR |
1700 | if (op_type == binary_op) |
1701 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
1702 | else | |
1703 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
1704 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1705 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
1706 | } |
1707 | else | |
ebfd146a | 1708 | { |
b8698a0f L |
1709 | /* Generic support */ |
1710 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
1711 | if (op_type != binary_op) |
1712 | vec_oprnd1 = NULL; | |
1713 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0, | |
1714 | vec_oprnd1); | |
1715 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1716 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 1717 | } |
ebfd146a IR |
1718 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1719 | ||
ebfd146a IR |
1720 | return new_stmt; |
1721 | } | |
1722 | ||
1723 | ||
b8698a0f L |
1724 | /* Check if STMT performs a conversion operation, that can be vectorized. |
1725 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1726 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1727 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1728 | ||
1729 | static bool | |
1730 | vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi, | |
1731 | gimple *vec_stmt, slp_tree slp_node) | |
1732 | { | |
1733 | tree vec_dest; | |
1734 | tree scalar_dest; | |
1735 | tree op0; | |
1736 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1737 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1738 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1739 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
1740 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; | |
1741 | tree new_temp; | |
1742 | tree def; | |
1743 | gimple def_stmt; | |
1744 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
1745 | gimple new_stmt = NULL; | |
1746 | stmt_vec_info prev_stmt_info; | |
1747 | int nunits_in; | |
1748 | int nunits_out; | |
1749 | tree vectype_out, vectype_in; | |
1750 | int ncopies, j; | |
b690cc0f | 1751 | tree rhs_type; |
ebfd146a IR |
1752 | tree builtin_decl; |
1753 | enum { NARROW, NONE, WIDEN } modifier; | |
1754 | int i; | |
1755 | VEC(tree,heap) *vec_oprnds0 = NULL; | |
1756 | tree vop0; | |
ebfd146a IR |
1757 | VEC(tree,heap) *dummy = NULL; |
1758 | int dummy_int; | |
1759 | ||
1760 | /* Is STMT a vectorizable conversion? */ | |
1761 | ||
a70d6342 IR |
1762 | /* FORNOW: unsupported in basic block SLP. */ |
1763 | gcc_assert (loop_vinfo); | |
b8698a0f | 1764 | |
ebfd146a IR |
1765 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
1766 | return false; | |
1767 | ||
8644a673 | 1768 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1769 | return false; |
1770 | ||
1771 | if (!is_gimple_assign (stmt)) | |
1772 | return false; | |
1773 | ||
1774 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
1775 | return false; | |
1776 | ||
1777 | code = gimple_assign_rhs_code (stmt); | |
1778 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
1779 | return false; | |
1780 | ||
1781 | /* Check types of lhs and rhs. */ | |
b690cc0f RG |
1782 | scalar_dest = gimple_assign_lhs (stmt); |
1783 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
1784 | ||
ebfd146a IR |
1785 | op0 = gimple_assign_rhs1 (stmt); |
1786 | rhs_type = TREE_TYPE (op0); | |
b690cc0f RG |
1787 | /* Check the operands of the operation. */ |
1788 | if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL, | |
1789 | &def_stmt, &def, &dt[0], &vectype_in)) | |
1790 | { | |
1791 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1792 | fprintf (vect_dump, "use not simple."); | |
1793 | return false; | |
1794 | } | |
1795 | /* If op0 is an external or constant defs use a vector type of | |
1796 | the same size as the output vector type. */ | |
ebfd146a | 1797 | if (!vectype_in) |
b690cc0f | 1798 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
1799 | if (vec_stmt) |
1800 | gcc_assert (vectype_in); | |
1801 | if (!vectype_in) | |
1802 | { | |
1803 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1804 | { | |
1805 | fprintf (vect_dump, "no vectype for scalar type "); | |
1806 | print_generic_expr (vect_dump, rhs_type, TDF_SLIM); | |
1807 | } | |
1808 | ||
1809 | return false; | |
1810 | } | |
ebfd146a IR |
1811 | |
1812 | /* FORNOW */ | |
b690cc0f RG |
1813 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
1814 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
1815 | if (nunits_in == nunits_out / 2) |
1816 | modifier = NARROW; | |
1817 | else if (nunits_out == nunits_in) | |
1818 | modifier = NONE; | |
1819 | else if (nunits_out == nunits_in / 2) | |
1820 | modifier = WIDEN; | |
1821 | else | |
1822 | return false; | |
1823 | ||
ebfd146a IR |
1824 | if (modifier == NARROW) |
1825 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
1826 | else | |
1827 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1828 | ||
ff802fa1 IR |
1829 | /* Multiple types in SLP are handled by creating the appropriate number of |
1830 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
1831 | case of SLP. */ | |
437f4a00 | 1832 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a | 1833 | ncopies = 1; |
b8698a0f | 1834 | |
ebfd146a IR |
1835 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
1836 | needs to be generated. */ | |
1837 | gcc_assert (ncopies >= 1); | |
1838 | ||
ebfd146a IR |
1839 | /* Supportable by target? */ |
1840 | if ((modifier == NONE | |
88dd7150 | 1841 | && !targetm.vectorize.builtin_conversion (code, vectype_out, vectype_in)) |
ebfd146a | 1842 | || (modifier == WIDEN |
b690cc0f RG |
1843 | && !supportable_widening_operation (code, stmt, |
1844 | vectype_out, vectype_in, | |
ebfd146a IR |
1845 | &decl1, &decl2, |
1846 | &code1, &code2, | |
1847 | &dummy_int, &dummy)) | |
1848 | || (modifier == NARROW | |
b690cc0f | 1849 | && !supportable_narrowing_operation (code, vectype_out, vectype_in, |
ebfd146a IR |
1850 | &code1, &dummy_int, &dummy))) |
1851 | { | |
1852 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1853 | fprintf (vect_dump, "conversion not supported by target."); | |
1854 | return false; | |
1855 | } | |
1856 | ||
1857 | if (modifier != NONE) | |
1858 | { | |
ebfd146a IR |
1859 | /* FORNOW: SLP not supported. */ |
1860 | if (STMT_SLP_TYPE (stmt_info)) | |
b8698a0f | 1861 | return false; |
ebfd146a IR |
1862 | } |
1863 | ||
1864 | if (!vec_stmt) /* transformation not required. */ | |
1865 | { | |
1866 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
1867 | return true; | |
1868 | } | |
1869 | ||
1870 | /** Transform. **/ | |
1871 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1872 | fprintf (vect_dump, "transform conversion."); | |
1873 | ||
1874 | /* Handle def. */ | |
1875 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1876 | ||
1877 | if (modifier == NONE && !slp_node) | |
1878 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
1879 | ||
1880 | prev_stmt_info = NULL; | |
1881 | switch (modifier) | |
1882 | { | |
1883 | case NONE: | |
1884 | for (j = 0; j < ncopies; j++) | |
1885 | { | |
ebfd146a | 1886 | if (j == 0) |
b8698a0f | 1887 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node); |
ebfd146a IR |
1888 | else |
1889 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
1890 | ||
1891 | builtin_decl = | |
88dd7150 RG |
1892 | targetm.vectorize.builtin_conversion (code, |
1893 | vectype_out, vectype_in); | |
ac47786e | 1894 | FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0) |
b8698a0f | 1895 | { |
ebfd146a IR |
1896 | /* Arguments are ready. create the new vector stmt. */ |
1897 | new_stmt = gimple_build_call (builtin_decl, 1, vop0); | |
1898 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1899 | gimple_call_set_lhs (new_stmt, new_temp); | |
1900 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a IR |
1901 | if (slp_node) |
1902 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
1903 | } | |
1904 | ||
1905 | if (j == 0) | |
1906 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1907 | else | |
1908 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1909 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1910 | } | |
1911 | break; | |
1912 | ||
1913 | case WIDEN: | |
1914 | /* In case the vectorization factor (VF) is bigger than the number | |
1915 | of elements that we can fit in a vectype (nunits), we have to | |
1916 | generate more than one vector stmt - i.e - we need to "unroll" | |
1917 | the vector stmt by a factor VF/nunits. */ | |
1918 | for (j = 0; j < ncopies; j++) | |
1919 | { | |
1920 | if (j == 0) | |
1921 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
1922 | else | |
1923 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1924 | ||
ebfd146a IR |
1925 | /* Generate first half of the widened result: */ |
1926 | new_stmt | |
b8698a0f | 1927 | = vect_gen_widened_results_half (code1, decl1, |
ebfd146a IR |
1928 | vec_oprnd0, vec_oprnd1, |
1929 | unary_op, vec_dest, gsi, stmt); | |
1930 | if (j == 0) | |
1931 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1932 | else | |
1933 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1934 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1935 | ||
1936 | /* Generate second half of the widened result: */ | |
1937 | new_stmt | |
1938 | = vect_gen_widened_results_half (code2, decl2, | |
1939 | vec_oprnd0, vec_oprnd1, | |
1940 | unary_op, vec_dest, gsi, stmt); | |
1941 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1942 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1943 | } | |
1944 | break; | |
1945 | ||
1946 | case NARROW: | |
1947 | /* In case the vectorization factor (VF) is bigger than the number | |
1948 | of elements that we can fit in a vectype (nunits), we have to | |
1949 | generate more than one vector stmt - i.e - we need to "unroll" | |
1950 | the vector stmt by a factor VF/nunits. */ | |
1951 | for (j = 0; j < ncopies; j++) | |
1952 | { | |
1953 | /* Handle uses. */ | |
1954 | if (j == 0) | |
1955 | { | |
1956 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
1957 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1958 | } | |
1959 | else | |
1960 | { | |
1961 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd1); | |
1962 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1963 | } | |
1964 | ||
9dc3f7de | 1965 | /* Arguments are ready. Create the new vector stmt. */ |
ebfd146a IR |
1966 | new_stmt = gimple_build_assign_with_ops (code1, vec_dest, vec_oprnd0, |
1967 | vec_oprnd1); | |
1968 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1969 | gimple_assign_set_lhs (new_stmt, new_temp); | |
1970 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1971 | ||
1972 | if (j == 0) | |
1973 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1974 | else | |
1975 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1976 | ||
1977 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1978 | } | |
1979 | ||
1980 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
1981 | } | |
1982 | ||
1983 | if (vec_oprnds0) | |
b8698a0f | 1984 | VEC_free (tree, heap, vec_oprnds0); |
ebfd146a IR |
1985 | |
1986 | return true; | |
1987 | } | |
ff802fa1 IR |
1988 | |
1989 | ||
ebfd146a IR |
1990 | /* Function vectorizable_assignment. |
1991 | ||
b8698a0f L |
1992 | Check if STMT performs an assignment (copy) that can be vectorized. |
1993 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1994 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1995 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1996 | ||
1997 | static bool | |
1998 | vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi, | |
1999 | gimple *vec_stmt, slp_tree slp_node) | |
2000 | { | |
2001 | tree vec_dest; | |
2002 | tree scalar_dest; | |
2003 | tree op; | |
2004 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2005 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2006 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2007 | tree new_temp; | |
2008 | tree def; | |
2009 | gimple def_stmt; | |
2010 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
fde9c428 | 2011 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
ebfd146a | 2012 | int ncopies; |
f18b55bd | 2013 | int i, j; |
ebfd146a IR |
2014 | VEC(tree,heap) *vec_oprnds = NULL; |
2015 | tree vop; | |
a70d6342 | 2016 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
f18b55bd IR |
2017 | gimple new_stmt = NULL; |
2018 | stmt_vec_info prev_stmt_info = NULL; | |
fde9c428 RG |
2019 | enum tree_code code; |
2020 | tree vectype_in; | |
ebfd146a IR |
2021 | |
2022 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2023 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2024 | case of SLP. */ | |
437f4a00 | 2025 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
2026 | ncopies = 1; |
2027 | else | |
2028 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2029 | ||
2030 | gcc_assert (ncopies >= 1); | |
ebfd146a | 2031 | |
a70d6342 | 2032 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2033 | return false; |
2034 | ||
8644a673 | 2035 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2036 | return false; |
2037 | ||
2038 | /* Is vectorizable assignment? */ | |
2039 | if (!is_gimple_assign (stmt)) | |
2040 | return false; | |
2041 | ||
2042 | scalar_dest = gimple_assign_lhs (stmt); | |
2043 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
2044 | return false; | |
2045 | ||
fde9c428 | 2046 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 2047 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
2048 | || code == PAREN_EXPR |
2049 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
2050 | op = gimple_assign_rhs1 (stmt); |
2051 | else | |
2052 | return false; | |
2053 | ||
fde9c428 RG |
2054 | if (!vect_is_simple_use_1 (op, loop_vinfo, bb_vinfo, |
2055 | &def_stmt, &def, &dt[0], &vectype_in)) | |
ebfd146a IR |
2056 | { |
2057 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2058 | fprintf (vect_dump, "use not simple."); | |
2059 | return false; | |
2060 | } | |
2061 | ||
fde9c428 RG |
2062 | /* We can handle NOP_EXPR conversions that do not change the number |
2063 | of elements or the vector size. */ | |
2064 | if (CONVERT_EXPR_CODE_P (code) | |
2065 | && (!vectype_in | |
2066 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
2067 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
2068 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
2069 | return false; | |
2070 | ||
ebfd146a IR |
2071 | if (!vec_stmt) /* transformation not required. */ |
2072 | { | |
2073 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
2074 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2075 | fprintf (vect_dump, "=== vectorizable_assignment ==="); | |
2076 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2077 | return true; | |
2078 | } | |
2079 | ||
2080 | /** Transform. **/ | |
2081 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2082 | fprintf (vect_dump, "transform assignment."); | |
2083 | ||
2084 | /* Handle def. */ | |
2085 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2086 | ||
2087 | /* Handle use. */ | |
f18b55bd | 2088 | for (j = 0; j < ncopies; j++) |
ebfd146a | 2089 | { |
f18b55bd IR |
2090 | /* Handle uses. */ |
2091 | if (j == 0) | |
2092 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); | |
2093 | else | |
2094 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
2095 | ||
2096 | /* Arguments are ready. create the new vector stmt. */ | |
ac47786e | 2097 | FOR_EACH_VEC_ELT (tree, vec_oprnds, i, vop) |
f18b55bd | 2098 | { |
fde9c428 | 2099 | if (CONVERT_EXPR_CODE_P (code)) |
4a73490d | 2100 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
2101 | new_stmt = gimple_build_assign (vec_dest, vop); |
2102 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2103 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2104 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2105 | if (slp_node) | |
2106 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2107 | } | |
ebfd146a IR |
2108 | |
2109 | if (slp_node) | |
f18b55bd IR |
2110 | continue; |
2111 | ||
2112 | if (j == 0) | |
2113 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2114 | else | |
2115 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2116 | ||
2117 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2118 | } | |
b8698a0f L |
2119 | |
2120 | VEC_free (tree, heap, vec_oprnds); | |
ebfd146a IR |
2121 | return true; |
2122 | } | |
2123 | ||
9dc3f7de IR |
2124 | |
2125 | /* Function vectorizable_shift. | |
2126 | ||
2127 | Check if STMT performs a shift operation that can be vectorized. | |
2128 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2129 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2130 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2131 | ||
2132 | static bool | |
2133 | vectorizable_shift (gimple stmt, gimple_stmt_iterator *gsi, | |
2134 | gimple *vec_stmt, slp_tree slp_node) | |
2135 | { | |
2136 | tree vec_dest; | |
2137 | tree scalar_dest; | |
2138 | tree op0, op1 = NULL; | |
2139 | tree vec_oprnd1 = NULL_TREE; | |
2140 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2141 | tree vectype; | |
2142 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2143 | enum tree_code code; | |
2144 | enum machine_mode vec_mode; | |
2145 | tree new_temp; | |
2146 | optab optab; | |
2147 | int icode; | |
2148 | enum machine_mode optab_op2_mode; | |
2149 | tree def; | |
2150 | gimple def_stmt; | |
2151 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2152 | gimple new_stmt = NULL; | |
2153 | stmt_vec_info prev_stmt_info; | |
2154 | int nunits_in; | |
2155 | int nunits_out; | |
2156 | tree vectype_out; | |
2157 | int ncopies; | |
2158 | int j, i; | |
2159 | VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; | |
2160 | tree vop0, vop1; | |
2161 | unsigned int k; | |
49eab32e | 2162 | bool scalar_shift_arg = true; |
9dc3f7de IR |
2163 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2164 | int vf; | |
2165 | ||
2166 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
2167 | return false; | |
2168 | ||
2169 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
2170 | return false; | |
2171 | ||
2172 | /* Is STMT a vectorizable binary/unary operation? */ | |
2173 | if (!is_gimple_assign (stmt)) | |
2174 | return false; | |
2175 | ||
2176 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2177 | return false; | |
2178 | ||
2179 | code = gimple_assign_rhs_code (stmt); | |
2180 | ||
2181 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
2182 | || code == RROTATE_EXPR)) | |
2183 | return false; | |
2184 | ||
2185 | scalar_dest = gimple_assign_lhs (stmt); | |
2186 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2187 | ||
2188 | op0 = gimple_assign_rhs1 (stmt); | |
2189 | if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo, | |
2190 | &def_stmt, &def, &dt[0], &vectype)) | |
2191 | { | |
2192 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2193 | fprintf (vect_dump, "use not simple."); | |
2194 | return false; | |
2195 | } | |
2196 | /* If op0 is an external or constant def use a vector type with | |
2197 | the same size as the output vector type. */ | |
2198 | if (!vectype) | |
2199 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
2200 | if (vec_stmt) | |
2201 | gcc_assert (vectype); | |
2202 | if (!vectype) | |
2203 | { | |
2204 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2205 | { | |
2206 | fprintf (vect_dump, "no vectype for scalar type "); | |
2207 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
2208 | } | |
2209 | ||
2210 | return false; | |
2211 | } | |
2212 | ||
2213 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
2214 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
2215 | if (nunits_out != nunits_in) | |
2216 | return false; | |
2217 | ||
2218 | op1 = gimple_assign_rhs2 (stmt); | |
2219 | if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt[1])) | |
2220 | { | |
2221 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2222 | fprintf (vect_dump, "use not simple."); | |
2223 | return false; | |
2224 | } | |
2225 | ||
2226 | if (loop_vinfo) | |
2227 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
2228 | else | |
2229 | vf = 1; | |
2230 | ||
2231 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2232 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2233 | case of SLP. */ | |
437f4a00 | 2234 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
9dc3f7de IR |
2235 | ncopies = 1; |
2236 | else | |
2237 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2238 | ||
2239 | gcc_assert (ncopies >= 1); | |
2240 | ||
2241 | /* Determine whether the shift amount is a vector, or scalar. If the | |
2242 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
2243 | ||
49eab32e JJ |
2244 | if (dt[1] == vect_internal_def && !slp_node) |
2245 | scalar_shift_arg = false; | |
2246 | else if (dt[1] == vect_constant_def | |
2247 | || dt[1] == vect_external_def | |
2248 | || dt[1] == vect_internal_def) | |
2249 | { | |
2250 | /* In SLP, need to check whether the shift count is the same, | |
2251 | in loops if it is a constant or invariant, it is always | |
2252 | a scalar shift. */ | |
2253 | if (slp_node) | |
2254 | { | |
2255 | VEC (gimple, heap) *stmts = SLP_TREE_SCALAR_STMTS (slp_node); | |
2256 | gimple slpstmt; | |
2257 | ||
2258 | FOR_EACH_VEC_ELT (gimple, stmts, k, slpstmt) | |
2259 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) | |
2260 | scalar_shift_arg = false; | |
2261 | } | |
2262 | } | |
2263 | else | |
2264 | { | |
2265 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2266 | fprintf (vect_dump, "operand mode requires invariant argument."); | |
2267 | return false; | |
2268 | } | |
2269 | ||
9dc3f7de | 2270 | /* Vector shifted by vector. */ |
49eab32e | 2271 | if (!scalar_shift_arg) |
9dc3f7de IR |
2272 | { |
2273 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
2274 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2275 | fprintf (vect_dump, "vector/vector shift/rotate found."); | |
2276 | } | |
2277 | /* See if the machine has a vector shifted by scalar insn and if not | |
2278 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 2279 | else |
9dc3f7de IR |
2280 | { |
2281 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
2282 | if (optab | |
2283 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
2284 | { | |
9dc3f7de IR |
2285 | if (vect_print_dump_info (REPORT_DETAILS)) |
2286 | fprintf (vect_dump, "vector/scalar shift/rotate found."); | |
2287 | } | |
2288 | else | |
2289 | { | |
2290 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
2291 | if (optab | |
2292 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
2293 | != CODE_FOR_nothing)) | |
2294 | { | |
49eab32e JJ |
2295 | scalar_shift_arg = false; |
2296 | ||
9dc3f7de IR |
2297 | if (vect_print_dump_info (REPORT_DETAILS)) |
2298 | fprintf (vect_dump, "vector/vector shift/rotate found."); | |
2299 | ||
2300 | /* Unlike the other binary operators, shifts/rotates have | |
2301 | the rhs being int, instead of the same type as the lhs, | |
2302 | so make sure the scalar is the right type if we are | |
2303 | dealing with vectors of short/char. */ | |
2304 | if (dt[1] == vect_constant_def) | |
2305 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
2306 | } | |
2307 | } | |
2308 | } | |
9dc3f7de IR |
2309 | |
2310 | /* Supportable by target? */ | |
2311 | if (!optab) | |
2312 | { | |
2313 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2314 | fprintf (vect_dump, "no optab."); | |
2315 | return false; | |
2316 | } | |
2317 | vec_mode = TYPE_MODE (vectype); | |
2318 | icode = (int) optab_handler (optab, vec_mode); | |
2319 | if (icode == CODE_FOR_nothing) | |
2320 | { | |
2321 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2322 | fprintf (vect_dump, "op not supported by target."); | |
2323 | /* Check only during analysis. */ | |
2324 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
2325 | || (vf < vect_min_worthwhile_factor (code) | |
2326 | && !vec_stmt)) | |
2327 | return false; | |
2328 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2329 | fprintf (vect_dump, "proceeding using word mode."); | |
2330 | } | |
2331 | ||
2332 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
2333 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
2334 | && vf < vect_min_worthwhile_factor (code) | |
2335 | && !vec_stmt) | |
2336 | { | |
2337 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2338 | fprintf (vect_dump, "not worthwhile without SIMD support."); | |
2339 | return false; | |
2340 | } | |
2341 | ||
2342 | if (!vec_stmt) /* transformation not required. */ | |
2343 | { | |
2344 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
2345 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2346 | fprintf (vect_dump, "=== vectorizable_shift ==="); | |
2347 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2348 | return true; | |
2349 | } | |
2350 | ||
2351 | /** Transform. **/ | |
2352 | ||
2353 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2354 | fprintf (vect_dump, "transform binary/unary operation."); | |
2355 | ||
2356 | /* Handle def. */ | |
2357 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2358 | ||
2359 | /* Allocate VECs for vector operands. In case of SLP, vector operands are | |
2360 | created in the previous stages of the recursion, so no allocation is | |
2361 | needed, except for the case of shift with scalar shift argument. In that | |
2362 | case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to | |
2363 | be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE. | |
2364 | In case of loop-based vectorization we allocate VECs of size 1. We | |
2365 | allocate VEC_OPRNDS1 only in case of binary operation. */ | |
2366 | if (!slp_node) | |
2367 | { | |
2368 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
2369 | vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
2370 | } | |
2371 | else if (scalar_shift_arg) | |
2372 | vec_oprnds1 = VEC_alloc (tree, heap, slp_node->vec_stmts_size); | |
2373 | ||
2374 | prev_stmt_info = NULL; | |
2375 | for (j = 0; j < ncopies; j++) | |
2376 | { | |
2377 | /* Handle uses. */ | |
2378 | if (j == 0) | |
2379 | { | |
2380 | if (scalar_shift_arg) | |
2381 | { | |
2382 | /* Vector shl and shr insn patterns can be defined with scalar | |
2383 | operand 2 (shift operand). In this case, use constant or loop | |
2384 | invariant op1 directly, without extending it to vector mode | |
2385 | first. */ | |
2386 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
2387 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
2388 | { | |
2389 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2390 | fprintf (vect_dump, "operand 1 using scalar mode."); | |
2391 | vec_oprnd1 = op1; | |
2392 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2393 | if (slp_node) | |
2394 | { | |
2395 | /* Store vec_oprnd1 for every vector stmt to be created | |
2396 | for SLP_NODE. We check during the analysis that all | |
2397 | the shift arguments are the same. | |
2398 | TODO: Allow different constants for different vector | |
2399 | stmts generated for an SLP instance. */ | |
2400 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
2401 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2402 | } | |
2403 | } | |
2404 | } | |
2405 | ||
2406 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
2407 | (a special case for certain kind of vector shifts); otherwise, | |
2408 | operand 1 should be of a vector type (the usual case). */ | |
2409 | if (vec_oprnd1) | |
2410 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2411 | slp_node); | |
2412 | else | |
2413 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
2414 | slp_node); | |
2415 | } | |
2416 | else | |
2417 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
2418 | ||
2419 | /* Arguments are ready. Create the new vector stmt. */ | |
2420 | FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0) | |
2421 | { | |
2422 | vop1 = VEC_index (tree, vec_oprnds1, i); | |
2423 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); | |
2424 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2425 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2426 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2427 | if (slp_node) | |
2428 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2429 | } | |
2430 | ||
2431 | if (slp_node) | |
2432 | continue; | |
2433 | ||
2434 | if (j == 0) | |
2435 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2436 | else | |
2437 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2438 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2439 | } | |
2440 | ||
2441 | VEC_free (tree, heap, vec_oprnds0); | |
2442 | VEC_free (tree, heap, vec_oprnds1); | |
2443 | ||
2444 | return true; | |
2445 | } | |
2446 | ||
2447 | ||
ebfd146a IR |
2448 | /* Function vectorizable_operation. |
2449 | ||
16949072 RG |
2450 | Check if STMT performs a binary, unary or ternary operation that can |
2451 | be vectorized. | |
b8698a0f | 2452 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2453 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2454 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2455 | ||
2456 | static bool | |
2457 | vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi, | |
2458 | gimple *vec_stmt, slp_tree slp_node) | |
2459 | { | |
2460 | tree vec_dest; | |
2461 | tree scalar_dest; | |
16949072 | 2462 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 2463 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
b690cc0f | 2464 | tree vectype; |
ebfd146a IR |
2465 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
2466 | enum tree_code code; | |
2467 | enum machine_mode vec_mode; | |
2468 | tree new_temp; | |
2469 | int op_type; | |
2470 | optab optab; | |
2471 | int icode; | |
ebfd146a IR |
2472 | tree def; |
2473 | gimple def_stmt; | |
16949072 RG |
2474 | enum vect_def_type dt[3] |
2475 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
ebfd146a IR |
2476 | gimple new_stmt = NULL; |
2477 | stmt_vec_info prev_stmt_info; | |
b690cc0f | 2478 | int nunits_in; |
ebfd146a IR |
2479 | int nunits_out; |
2480 | tree vectype_out; | |
2481 | int ncopies; | |
2482 | int j, i; | |
16949072 RG |
2483 | VEC(tree,heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL, *vec_oprnds2 = NULL; |
2484 | tree vop0, vop1, vop2; | |
a70d6342 IR |
2485 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2486 | int vf; | |
2487 | ||
a70d6342 | 2488 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2489 | return false; |
2490 | ||
8644a673 | 2491 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2492 | return false; |
2493 | ||
2494 | /* Is STMT a vectorizable binary/unary operation? */ | |
2495 | if (!is_gimple_assign (stmt)) | |
2496 | return false; | |
2497 | ||
2498 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2499 | return false; | |
2500 | ||
ebfd146a IR |
2501 | code = gimple_assign_rhs_code (stmt); |
2502 | ||
2503 | /* For pointer addition, we should use the normal plus for | |
2504 | the vector addition. */ | |
2505 | if (code == POINTER_PLUS_EXPR) | |
2506 | code = PLUS_EXPR; | |
2507 | ||
2508 | /* Support only unary or binary operations. */ | |
2509 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 2510 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a IR |
2511 | { |
2512 | if (vect_print_dump_info (REPORT_DETAILS)) | |
16949072 RG |
2513 | fprintf (vect_dump, "num. args = %d (not unary/binary/ternary op).", |
2514 | op_type); | |
ebfd146a IR |
2515 | return false; |
2516 | } | |
2517 | ||
b690cc0f RG |
2518 | scalar_dest = gimple_assign_lhs (stmt); |
2519 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2520 | ||
ebfd146a | 2521 | op0 = gimple_assign_rhs1 (stmt); |
b690cc0f RG |
2522 | if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo, |
2523 | &def_stmt, &def, &dt[0], &vectype)) | |
ebfd146a IR |
2524 | { |
2525 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2526 | fprintf (vect_dump, "use not simple."); | |
2527 | return false; | |
2528 | } | |
b690cc0f RG |
2529 | /* If op0 is an external or constant def use a vector type with |
2530 | the same size as the output vector type. */ | |
2531 | if (!vectype) | |
2532 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
2533 | if (vec_stmt) |
2534 | gcc_assert (vectype); | |
2535 | if (!vectype) | |
2536 | { | |
2537 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2538 | { | |
2539 | fprintf (vect_dump, "no vectype for scalar type "); | |
2540 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
2541 | } | |
2542 | ||
2543 | return false; | |
2544 | } | |
b690cc0f RG |
2545 | |
2546 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
2547 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
2548 | if (nunits_out != nunits_in) | |
2549 | return false; | |
ebfd146a | 2550 | |
16949072 | 2551 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
2552 | { |
2553 | op1 = gimple_assign_rhs2 (stmt); | |
b8698a0f | 2554 | if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def, |
a70d6342 | 2555 | &dt[1])) |
ebfd146a IR |
2556 | { |
2557 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2558 | fprintf (vect_dump, "use not simple."); | |
2559 | return false; | |
2560 | } | |
2561 | } | |
16949072 RG |
2562 | if (op_type == ternary_op) |
2563 | { | |
2564 | op2 = gimple_assign_rhs3 (stmt); | |
2565 | if (!vect_is_simple_use (op2, loop_vinfo, bb_vinfo, &def_stmt, &def, | |
2566 | &dt[2])) | |
2567 | { | |
2568 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2569 | fprintf (vect_dump, "use not simple."); | |
2570 | return false; | |
2571 | } | |
2572 | } | |
ebfd146a | 2573 | |
b690cc0f RG |
2574 | if (loop_vinfo) |
2575 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
2576 | else | |
2577 | vf = 1; | |
2578 | ||
2579 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 2580 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 2581 | case of SLP. */ |
437f4a00 | 2582 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
b690cc0f RG |
2583 | ncopies = 1; |
2584 | else | |
2585 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2586 | ||
2587 | gcc_assert (ncopies >= 1); | |
2588 | ||
9dc3f7de | 2589 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
2590 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
2591 | || code == RROTATE_EXPR) | |
9dc3f7de | 2592 | return false; |
ebfd146a | 2593 | |
16949072 | 2594 | optab = optab_for_tree_code (code, vectype, optab_default); |
ebfd146a IR |
2595 | |
2596 | /* Supportable by target? */ | |
2597 | if (!optab) | |
2598 | { | |
2599 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2600 | fprintf (vect_dump, "no optab."); | |
2601 | return false; | |
2602 | } | |
2603 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 2604 | icode = (int) optab_handler (optab, vec_mode); |
ebfd146a IR |
2605 | if (icode == CODE_FOR_nothing) |
2606 | { | |
2607 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2608 | fprintf (vect_dump, "op not supported by target."); | |
2609 | /* Check only during analysis. */ | |
2610 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
a70d6342 | 2611 | || (vf < vect_min_worthwhile_factor (code) |
ebfd146a IR |
2612 | && !vec_stmt)) |
2613 | return false; | |
2614 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2615 | fprintf (vect_dump, "proceeding using word mode."); | |
2616 | } | |
2617 | ||
ff802fa1 | 2618 | /* Worthwhile without SIMD support? Check only during analysis. */ |
ebfd146a | 2619 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) |
a70d6342 | 2620 | && vf < vect_min_worthwhile_factor (code) |
ebfd146a IR |
2621 | && !vec_stmt) |
2622 | { | |
2623 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2624 | fprintf (vect_dump, "not worthwhile without SIMD support."); | |
2625 | return false; | |
2626 | } | |
2627 | ||
2628 | if (!vec_stmt) /* transformation not required. */ | |
2629 | { | |
2630 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; | |
2631 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2632 | fprintf (vect_dump, "=== vectorizable_operation ==="); | |
2633 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2634 | return true; | |
2635 | } | |
2636 | ||
2637 | /** Transform. **/ | |
2638 | ||
2639 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2640 | fprintf (vect_dump, "transform binary/unary operation."); | |
2641 | ||
2642 | /* Handle def. */ | |
2643 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2644 | ||
ff802fa1 | 2645 | /* Allocate VECs for vector operands. In case of SLP, vector operands are |
ebfd146a | 2646 | created in the previous stages of the recursion, so no allocation is |
ff802fa1 | 2647 | needed, except for the case of shift with scalar shift argument. In that |
ebfd146a IR |
2648 | case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to |
2649 | be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE. | |
ff802fa1 | 2650 | In case of loop-based vectorization we allocate VECs of size 1. We |
b8698a0f | 2651 | allocate VEC_OPRNDS1 only in case of binary operation. */ |
ebfd146a IR |
2652 | if (!slp_node) |
2653 | { | |
2654 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
16949072 | 2655 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a | 2656 | vec_oprnds1 = VEC_alloc (tree, heap, 1); |
16949072 RG |
2657 | if (op_type == ternary_op) |
2658 | vec_oprnds2 = VEC_alloc (tree, heap, 1); | |
ebfd146a | 2659 | } |
ebfd146a IR |
2660 | |
2661 | /* In case the vectorization factor (VF) is bigger than the number | |
2662 | of elements that we can fit in a vectype (nunits), we have to generate | |
2663 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 2664 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 2665 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 2666 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 2667 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
2668 | stmts that use the defs of the current stmt. The example below |
2669 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
2670 | we need to create 4 vectorized stmts): | |
ebfd146a IR |
2671 | |
2672 | before vectorization: | |
2673 | RELATED_STMT VEC_STMT | |
2674 | S1: x = memref - - | |
2675 | S2: z = x + 1 - - | |
2676 | ||
2677 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
2678 | there): | |
2679 | RELATED_STMT VEC_STMT | |
2680 | VS1_0: vx0 = memref0 VS1_1 - | |
2681 | VS1_1: vx1 = memref1 VS1_2 - | |
2682 | VS1_2: vx2 = memref2 VS1_3 - | |
2683 | VS1_3: vx3 = memref3 - - | |
2684 | S1: x = load - VS1_0 | |
2685 | S2: z = x + 1 - - | |
2686 | ||
2687 | step2: vectorize stmt S2 (done here): | |
2688 | To vectorize stmt S2 we first need to find the relevant vector | |
ff802fa1 | 2689 | def for the first operand 'x'. This is, as usual, obtained from |
ebfd146a | 2690 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt |
ff802fa1 IR |
2691 | that defines 'x' (S1). This way we find the stmt VS1_0, and the |
2692 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
ebfd146a IR |
2693 | the vector stmt VS2_0, and as usual, record it in the |
2694 | STMT_VINFO_VEC_STMT of stmt S2. | |
2695 | When creating the second copy (VS2_1), we obtain the relevant vector | |
2696 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
ff802fa1 IR |
2697 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant |
2698 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
ebfd146a | 2699 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. |
ff802fa1 | 2700 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting |
ebfd146a IR |
2701 | chain of stmts and pointers: |
2702 | RELATED_STMT VEC_STMT | |
2703 | VS1_0: vx0 = memref0 VS1_1 - | |
2704 | VS1_1: vx1 = memref1 VS1_2 - | |
2705 | VS1_2: vx2 = memref2 VS1_3 - | |
2706 | VS1_3: vx3 = memref3 - - | |
2707 | S1: x = load - VS1_0 | |
2708 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
2709 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
2710 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
2711 | VS2_3: vz3 = vx3 + v1 - - | |
2712 | S2: z = x + 1 - VS2_0 */ | |
2713 | ||
2714 | prev_stmt_info = NULL; | |
2715 | for (j = 0; j < ncopies; j++) | |
2716 | { | |
2717 | /* Handle uses. */ | |
2718 | if (j == 0) | |
2719 | { | |
16949072 | 2720 | if (op_type == binary_op || op_type == ternary_op) |
b8698a0f | 2721 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, |
ebfd146a IR |
2722 | slp_node); |
2723 | else | |
b8698a0f | 2724 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, |
ebfd146a | 2725 | slp_node); |
16949072 RG |
2726 | if (op_type == ternary_op) |
2727 | { | |
2728 | vec_oprnds2 = VEC_alloc (tree, heap, 1); | |
2729 | VEC_quick_push (tree, vec_oprnds2, | |
2730 | vect_get_vec_def_for_operand (op2, stmt, NULL)); | |
2731 | } | |
ebfd146a IR |
2732 | } |
2733 | else | |
16949072 RG |
2734 | { |
2735 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
2736 | if (op_type == ternary_op) | |
2737 | { | |
2738 | tree vec_oprnd = VEC_pop (tree, vec_oprnds2); | |
2739 | VEC_quick_push (tree, vec_oprnds2, | |
2740 | vect_get_vec_def_for_stmt_copy (dt[2], | |
2741 | vec_oprnd)); | |
2742 | } | |
2743 | } | |
ebfd146a | 2744 | |
9dc3f7de | 2745 | /* Arguments are ready. Create the new vector stmt. */ |
ac47786e | 2746 | FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0) |
ebfd146a | 2747 | { |
16949072 RG |
2748 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
2749 | ? VEC_index (tree, vec_oprnds1, i) : NULL_TREE); | |
2750 | vop2 = ((op_type == ternary_op) | |
2751 | ? VEC_index (tree, vec_oprnds2, i) : NULL_TREE); | |
2752 | new_stmt = gimple_build_assign_with_ops3 (code, vec_dest, | |
2753 | vop0, vop1, vop2); | |
ebfd146a IR |
2754 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2755 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2756 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2757 | if (slp_node) | |
2758 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2759 | } | |
2760 | ||
2761 | if (slp_node) | |
2762 | continue; | |
2763 | ||
2764 | if (j == 0) | |
2765 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2766 | else | |
2767 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2768 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2769 | } | |
2770 | ||
2771 | VEC_free (tree, heap, vec_oprnds0); | |
2772 | if (vec_oprnds1) | |
2773 | VEC_free (tree, heap, vec_oprnds1); | |
16949072 RG |
2774 | if (vec_oprnds2) |
2775 | VEC_free (tree, heap, vec_oprnds2); | |
ebfd146a IR |
2776 | |
2777 | return true; | |
2778 | } | |
2779 | ||
2780 | ||
ff802fa1 | 2781 | /* Get vectorized definitions for loop-based vectorization. For the first |
b8698a0f L |
2782 | operand we call vect_get_vec_def_for_operand() (with OPRND containing |
2783 | scalar operand), and for the rest we get a copy with | |
ebfd146a IR |
2784 | vect_get_vec_def_for_stmt_copy() using the previous vector definition |
2785 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
2786 | The vectors are collected into VEC_OPRNDS. */ | |
2787 | ||
2788 | static void | |
b8698a0f | 2789 | vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt, |
ebfd146a IR |
2790 | VEC (tree, heap) **vec_oprnds, int multi_step_cvt) |
2791 | { | |
2792 | tree vec_oprnd; | |
2793 | ||
2794 | /* Get first vector operand. */ | |
2795 | /* All the vector operands except the very first one (that is scalar oprnd) | |
2796 | are stmt copies. */ | |
b8698a0f | 2797 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) |
ebfd146a IR |
2798 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL); |
2799 | else | |
2800 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
2801 | ||
2802 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
2803 | ||
2804 | /* Get second vector operand. */ | |
2805 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
2806 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
b8698a0f | 2807 | |
ebfd146a IR |
2808 | *oprnd = vec_oprnd; |
2809 | ||
b8698a0f | 2810 | /* For conversion in multiple steps, continue to get operands |
ebfd146a IR |
2811 | recursively. */ |
2812 | if (multi_step_cvt) | |
b8698a0f | 2813 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); |
ebfd146a IR |
2814 | } |
2815 | ||
2816 | ||
2817 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
b8698a0f | 2818 | For multi-step conversions store the resulting vectors and call the function |
ebfd146a IR |
2819 | recursively. */ |
2820 | ||
2821 | static void | |
2822 | vect_create_vectorized_demotion_stmts (VEC (tree, heap) **vec_oprnds, | |
2823 | int multi_step_cvt, gimple stmt, | |
2824 | VEC (tree, heap) *vec_dsts, | |
2825 | gimple_stmt_iterator *gsi, | |
2826 | slp_tree slp_node, enum tree_code code, | |
2827 | stmt_vec_info *prev_stmt_info) | |
2828 | { | |
2829 | unsigned int i; | |
2830 | tree vop0, vop1, new_tmp, vec_dest; | |
2831 | gimple new_stmt; | |
2832 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2833 | ||
b8698a0f | 2834 | vec_dest = VEC_pop (tree, vec_dsts); |
ebfd146a IR |
2835 | |
2836 | for (i = 0; i < VEC_length (tree, *vec_oprnds); i += 2) | |
2837 | { | |
2838 | /* Create demotion operation. */ | |
2839 | vop0 = VEC_index (tree, *vec_oprnds, i); | |
2840 | vop1 = VEC_index (tree, *vec_oprnds, i + 1); | |
2841 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); | |
2842 | new_tmp = make_ssa_name (vec_dest, new_stmt); | |
2843 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
2844 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2845 | ||
2846 | if (multi_step_cvt) | |
2847 | /* Store the resulting vector for next recursive call. */ | |
b8698a0f | 2848 | VEC_replace (tree, *vec_oprnds, i/2, new_tmp); |
ebfd146a IR |
2849 | else |
2850 | { | |
b8698a0f | 2851 | /* This is the last step of the conversion sequence. Store the |
ebfd146a IR |
2852 | vectors in SLP_NODE or in vector info of the scalar statement |
2853 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
2854 | if (slp_node) | |
2855 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2856 | else | |
2857 | { | |
2858 | if (!*prev_stmt_info) | |
2859 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2860 | else | |
2861 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
2862 | ||
2863 | *prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2864 | } | |
2865 | } | |
2866 | } | |
2867 | ||
2868 | /* For multi-step demotion operations we first generate demotion operations | |
b8698a0f | 2869 | from the source type to the intermediate types, and then combine the |
ebfd146a IR |
2870 | results (stored in VEC_OPRNDS) in demotion operation to the destination |
2871 | type. */ | |
2872 | if (multi_step_cvt) | |
2873 | { | |
2874 | /* At each level of recursion we have have of the operands we had at the | |
2875 | previous level. */ | |
2876 | VEC_truncate (tree, *vec_oprnds, (i+1)/2); | |
b8698a0f | 2877 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
ebfd146a IR |
2878 | stmt, vec_dsts, gsi, slp_node, |
2879 | code, prev_stmt_info); | |
2880 | } | |
2881 | } | |
2882 | ||
2883 | ||
2884 | /* Function vectorizable_type_demotion | |
2885 | ||
2886 | Check if STMT performs a binary or unary operation that involves | |
2887 | type demotion, and if it can be vectorized. | |
2888 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2889 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2890 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2891 | ||
2892 | static bool | |
2893 | vectorizable_type_demotion (gimple stmt, gimple_stmt_iterator *gsi, | |
2894 | gimple *vec_stmt, slp_tree slp_node) | |
2895 | { | |
2896 | tree vec_dest; | |
2897 | tree scalar_dest; | |
2898 | tree op0; | |
2899 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2900 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2901 | enum tree_code code, code1 = ERROR_MARK; | |
2902 | tree def; | |
2903 | gimple def_stmt; | |
2904 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2905 | stmt_vec_info prev_stmt_info; | |
2906 | int nunits_in; | |
2907 | int nunits_out; | |
2908 | tree vectype_out; | |
2909 | int ncopies; | |
2910 | int j, i; | |
2911 | tree vectype_in; | |
2912 | int multi_step_cvt = 0; | |
2913 | VEC (tree, heap) *vec_oprnds0 = NULL; | |
2914 | VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL; | |
2915 | tree last_oprnd, intermediate_type; | |
2916 | ||
a70d6342 IR |
2917 | /* FORNOW: not supported by basic block SLP vectorization. */ |
2918 | gcc_assert (loop_vinfo); | |
2919 | ||
ebfd146a IR |
2920 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
2921 | return false; | |
2922 | ||
8644a673 | 2923 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2924 | return false; |
2925 | ||
2926 | /* Is STMT a vectorizable type-demotion operation? */ | |
2927 | if (!is_gimple_assign (stmt)) | |
2928 | return false; | |
2929 | ||
2930 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2931 | return false; | |
2932 | ||
2933 | code = gimple_assign_rhs_code (stmt); | |
2934 | if (!CONVERT_EXPR_CODE_P (code)) | |
2935 | return false; | |
2936 | ||
b690cc0f RG |
2937 | scalar_dest = gimple_assign_lhs (stmt); |
2938 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2939 | ||
2940 | /* Check the operands of the operation. */ | |
ebfd146a | 2941 | op0 = gimple_assign_rhs1 (stmt); |
b690cc0f RG |
2942 | if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) |
2943 | && INTEGRAL_TYPE_P (TREE_TYPE (op0))) | |
2944 | || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest)) | |
2945 | && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0)) | |
2946 | && CONVERT_EXPR_CODE_P (code)))) | |
2947 | return false; | |
2948 | if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL, | |
2949 | &def_stmt, &def, &dt[0], &vectype_in)) | |
2950 | { | |
2951 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2952 | fprintf (vect_dump, "use not simple."); | |
2953 | return false; | |
2954 | } | |
2955 | /* If op0 is an external def use a vector type with the | |
2956 | same size as the output vector type if possible. */ | |
2957 | if (!vectype_in) | |
2958 | vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
2959 | if (vec_stmt) |
2960 | gcc_assert (vectype_in); | |
ebfd146a | 2961 | if (!vectype_in) |
7d8930a0 IR |
2962 | { |
2963 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2964 | { | |
2965 | fprintf (vect_dump, "no vectype for scalar type "); | |
2966 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
2967 | } | |
2968 | ||
2969 | return false; | |
2970 | } | |
ebfd146a | 2971 | |
b690cc0f | 2972 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
ebfd146a IR |
2973 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); |
2974 | if (nunits_in >= nunits_out) | |
2975 | return false; | |
2976 | ||
2977 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 2978 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 2979 | case of SLP. */ |
437f4a00 | 2980 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
2981 | ncopies = 1; |
2982 | else | |
2983 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
ebfd146a IR |
2984 | gcc_assert (ncopies >= 1); |
2985 | ||
ebfd146a | 2986 | /* Supportable by target? */ |
b690cc0f RG |
2987 | if (!supportable_narrowing_operation (code, vectype_out, vectype_in, |
2988 | &code1, &multi_step_cvt, &interm_types)) | |
ebfd146a IR |
2989 | return false; |
2990 | ||
ebfd146a IR |
2991 | if (!vec_stmt) /* transformation not required. */ |
2992 | { | |
2993 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
2994 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2995 | fprintf (vect_dump, "=== vectorizable_demotion ==="); | |
2996 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2997 | return true; | |
2998 | } | |
2999 | ||
3000 | /** Transform. **/ | |
3001 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3002 | fprintf (vect_dump, "transform type demotion operation. ncopies = %d.", | |
3003 | ncopies); | |
3004 | ||
b8698a0f L |
3005 | /* In case of multi-step demotion, we first generate demotion operations to |
3006 | the intermediate types, and then from that types to the final one. | |
ebfd146a | 3007 | We create vector destinations for the intermediate type (TYPES) received |
b8698a0f | 3008 | from supportable_narrowing_operation, and store them in the correct order |
ebfd146a IR |
3009 | for future use in vect_create_vectorized_demotion_stmts(). */ |
3010 | if (multi_step_cvt) | |
3011 | vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1); | |
3012 | else | |
3013 | vec_dsts = VEC_alloc (tree, heap, 1); | |
b8698a0f | 3014 | |
ebfd146a IR |
3015 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); |
3016 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
3017 | ||
3018 | if (multi_step_cvt) | |
3019 | { | |
b8698a0f | 3020 | for (i = VEC_length (tree, interm_types) - 1; |
ebfd146a IR |
3021 | VEC_iterate (tree, interm_types, i, intermediate_type); i--) |
3022 | { | |
b8698a0f | 3023 | vec_dest = vect_create_destination_var (scalar_dest, |
ebfd146a IR |
3024 | intermediate_type); |
3025 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
3026 | } | |
3027 | } | |
3028 | ||
3029 | /* In case the vectorization factor (VF) is bigger than the number | |
3030 | of elements that we can fit in a vectype (nunits), we have to generate | |
3031 | more than one vector stmt - i.e - we need to "unroll" the | |
3032 | vector stmt by a factor VF/nunits. */ | |
3033 | last_oprnd = op0; | |
3034 | prev_stmt_info = NULL; | |
3035 | for (j = 0; j < ncopies; j++) | |
3036 | { | |
3037 | /* Handle uses. */ | |
3038 | if (slp_node) | |
9dc3f7de | 3039 | vect_get_slp_defs (op0, NULL_TREE, slp_node, &vec_oprnds0, NULL, -1); |
ebfd146a IR |
3040 | else |
3041 | { | |
3042 | VEC_free (tree, heap, vec_oprnds0); | |
3043 | vec_oprnds0 = VEC_alloc (tree, heap, | |
3044 | (multi_step_cvt ? vect_pow2 (multi_step_cvt) * 2 : 2)); | |
b8698a0f | 3045 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
ebfd146a IR |
3046 | vect_pow2 (multi_step_cvt) - 1); |
3047 | } | |
3048 | ||
9dc3f7de | 3049 | /* Arguments are ready. Create the new vector stmts. */ |
ebfd146a | 3050 | tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts); |
b8698a0f | 3051 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, |
ebfd146a | 3052 | multi_step_cvt, stmt, tmp_vec_dsts, |
b8698a0f | 3053 | gsi, slp_node, code1, |
ebfd146a IR |
3054 | &prev_stmt_info); |
3055 | } | |
3056 | ||
3057 | VEC_free (tree, heap, vec_oprnds0); | |
3058 | VEC_free (tree, heap, vec_dsts); | |
3059 | VEC_free (tree, heap, tmp_vec_dsts); | |
3060 | VEC_free (tree, heap, interm_types); | |
3061 | ||
3062 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
3063 | return true; | |
3064 | } | |
3065 | ||
3066 | ||
3067 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
ff802fa1 | 3068 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store |
ebfd146a IR |
3069 | the resulting vectors and call the function recursively. */ |
3070 | ||
3071 | static void | |
3072 | vect_create_vectorized_promotion_stmts (VEC (tree, heap) **vec_oprnds0, | |
3073 | VEC (tree, heap) **vec_oprnds1, | |
3074 | int multi_step_cvt, gimple stmt, | |
3075 | VEC (tree, heap) *vec_dsts, | |
3076 | gimple_stmt_iterator *gsi, | |
3077 | slp_tree slp_node, enum tree_code code1, | |
b8698a0f | 3078 | enum tree_code code2, tree decl1, |
ebfd146a IR |
3079 | tree decl2, int op_type, |
3080 | stmt_vec_info *prev_stmt_info) | |
3081 | { | |
3082 | int i; | |
3083 | tree vop0, vop1, new_tmp1, new_tmp2, vec_dest; | |
3084 | gimple new_stmt1, new_stmt2; | |
3085 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3086 | VEC (tree, heap) *vec_tmp; | |
3087 | ||
3088 | vec_dest = VEC_pop (tree, vec_dsts); | |
3089 | vec_tmp = VEC_alloc (tree, heap, VEC_length (tree, *vec_oprnds0) * 2); | |
3090 | ||
ac47786e | 3091 | FOR_EACH_VEC_ELT (tree, *vec_oprnds0, i, vop0) |
ebfd146a IR |
3092 | { |
3093 | if (op_type == binary_op) | |
3094 | vop1 = VEC_index (tree, *vec_oprnds1, i); | |
3095 | else | |
3096 | vop1 = NULL_TREE; | |
3097 | ||
3098 | /* Generate the two halves of promotion operation. */ | |
b8698a0f | 3099 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, |
ebfd146a IR |
3100 | op_type, vec_dest, gsi, stmt); |
3101 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
3102 | op_type, vec_dest, gsi, stmt); | |
3103 | if (is_gimple_call (new_stmt1)) | |
3104 | { | |
3105 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
3106 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
3107 | } | |
3108 | else | |
3109 | { | |
3110 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
3111 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
3112 | } | |
3113 | ||
3114 | if (multi_step_cvt) | |
3115 | { | |
3116 | /* Store the results for the recursive call. */ | |
3117 | VEC_quick_push (tree, vec_tmp, new_tmp1); | |
3118 | VEC_quick_push (tree, vec_tmp, new_tmp2); | |
3119 | } | |
3120 | else | |
3121 | { | |
3122 | /* Last step of promotion sequience - store the results. */ | |
3123 | if (slp_node) | |
3124 | { | |
3125 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt1); | |
3126 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt2); | |
3127 | } | |
3128 | else | |
3129 | { | |
3130 | if (!*prev_stmt_info) | |
3131 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt1; | |
3132 | else | |
3133 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt1; | |
3134 | ||
3135 | *prev_stmt_info = vinfo_for_stmt (new_stmt1); | |
3136 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt2; | |
3137 | *prev_stmt_info = vinfo_for_stmt (new_stmt2); | |
3138 | } | |
3139 | } | |
3140 | } | |
3141 | ||
3142 | if (multi_step_cvt) | |
3143 | { | |
b8698a0f | 3144 | /* For multi-step promotion operation we first generate we call the |
ff802fa1 | 3145 | function recurcively for every stage. We start from the input type, |
ebfd146a IR |
3146 | create promotion operations to the intermediate types, and then |
3147 | create promotions to the output type. */ | |
3148 | *vec_oprnds0 = VEC_copy (tree, heap, vec_tmp); | |
ebfd146a IR |
3149 | vect_create_vectorized_promotion_stmts (vec_oprnds0, vec_oprnds1, |
3150 | multi_step_cvt - 1, stmt, | |
3151 | vec_dsts, gsi, slp_node, code1, | |
3152 | code2, decl2, decl2, op_type, | |
3153 | prev_stmt_info); | |
3154 | } | |
ff802fa1 IR |
3155 | |
3156 | VEC_free (tree, heap, vec_tmp); | |
ebfd146a | 3157 | } |
b8698a0f | 3158 | |
ebfd146a IR |
3159 | |
3160 | /* Function vectorizable_type_promotion | |
3161 | ||
3162 | Check if STMT performs a binary or unary operation that involves | |
3163 | type promotion, and if it can be vectorized. | |
3164 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3165 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3166 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3167 | ||
3168 | static bool | |
3169 | vectorizable_type_promotion (gimple stmt, gimple_stmt_iterator *gsi, | |
3170 | gimple *vec_stmt, slp_tree slp_node) | |
3171 | { | |
3172 | tree vec_dest; | |
3173 | tree scalar_dest; | |
3174 | tree op0, op1 = NULL; | |
3175 | tree vec_oprnd0=NULL, vec_oprnd1=NULL; | |
3176 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3177 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3178 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
3179 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; | |
b8698a0f | 3180 | int op_type; |
ebfd146a IR |
3181 | tree def; |
3182 | gimple def_stmt; | |
3183 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
3184 | stmt_vec_info prev_stmt_info; | |
3185 | int nunits_in; | |
3186 | int nunits_out; | |
3187 | tree vectype_out; | |
3188 | int ncopies; | |
3189 | int j, i; | |
3190 | tree vectype_in; | |
3191 | tree intermediate_type = NULL_TREE; | |
3192 | int multi_step_cvt = 0; | |
3193 | VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; | |
3194 | VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL; | |
b8698a0f | 3195 | |
a70d6342 IR |
3196 | /* FORNOW: not supported by basic block SLP vectorization. */ |
3197 | gcc_assert (loop_vinfo); | |
b8698a0f | 3198 | |
ebfd146a IR |
3199 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
3200 | return false; | |
3201 | ||
8644a673 | 3202 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3203 | return false; |
3204 | ||
3205 | /* Is STMT a vectorizable type-promotion operation? */ | |
3206 | if (!is_gimple_assign (stmt)) | |
3207 | return false; | |
3208 | ||
3209 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3210 | return false; | |
3211 | ||
3212 | code = gimple_assign_rhs_code (stmt); | |
3213 | if (!CONVERT_EXPR_CODE_P (code) | |
3214 | && code != WIDEN_MULT_EXPR) | |
3215 | return false; | |
3216 | ||
b690cc0f RG |
3217 | scalar_dest = gimple_assign_lhs (stmt); |
3218 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
3219 | ||
3220 | /* Check the operands of the operation. */ | |
ebfd146a | 3221 | op0 = gimple_assign_rhs1 (stmt); |
b690cc0f RG |
3222 | if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) |
3223 | && INTEGRAL_TYPE_P (TREE_TYPE (op0))) | |
3224 | || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest)) | |
3225 | && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0)) | |
3226 | && CONVERT_EXPR_CODE_P (code)))) | |
3227 | return false; | |
3228 | if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL, | |
3229 | &def_stmt, &def, &dt[0], &vectype_in)) | |
3230 | { | |
3231 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3232 | fprintf (vect_dump, "use not simple."); | |
3233 | return false; | |
3234 | } | |
3235 | /* If op0 is an external or constant def use a vector type with | |
3236 | the same size as the output vector type. */ | |
3237 | if (!vectype_in) | |
3238 | vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
3239 | if (vec_stmt) |
3240 | gcc_assert (vectype_in); | |
ebfd146a | 3241 | if (!vectype_in) |
7d8930a0 IR |
3242 | { |
3243 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3244 | { | |
3245 | fprintf (vect_dump, "no vectype for scalar type "); | |
3246 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
3247 | } | |
3248 | ||
3249 | return false; | |
3250 | } | |
ebfd146a | 3251 | |
b690cc0f | 3252 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
ebfd146a IR |
3253 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); |
3254 | if (nunits_in <= nunits_out) | |
3255 | return false; | |
3256 | ||
3257 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 3258 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 3259 | case of SLP. */ |
437f4a00 | 3260 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
3261 | ncopies = 1; |
3262 | else | |
3263 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
3264 | ||
3265 | gcc_assert (ncopies >= 1); | |
3266 | ||
ebfd146a IR |
3267 | op_type = TREE_CODE_LENGTH (code); |
3268 | if (op_type == binary_op) | |
3269 | { | |
3270 | op1 = gimple_assign_rhs2 (stmt); | |
a70d6342 | 3271 | if (!vect_is_simple_use (op1, loop_vinfo, NULL, &def_stmt, &def, &dt[1])) |
ebfd146a IR |
3272 | { |
3273 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3274 | fprintf (vect_dump, "use not simple."); | |
3275 | return false; | |
3276 | } | |
3277 | } | |
3278 | ||
3279 | /* Supportable by target? */ | |
b690cc0f | 3280 | if (!supportable_widening_operation (code, stmt, vectype_out, vectype_in, |
ebfd146a IR |
3281 | &decl1, &decl2, &code1, &code2, |
3282 | &multi_step_cvt, &interm_types)) | |
3283 | return false; | |
3284 | ||
3285 | /* Binary widening operation can only be supported directly by the | |
3286 | architecture. */ | |
3287 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
3288 | ||
ebfd146a IR |
3289 | if (!vec_stmt) /* transformation not required. */ |
3290 | { | |
3291 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
3292 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3293 | fprintf (vect_dump, "=== vectorizable_promotion ==="); | |
3294 | vect_model_simple_cost (stmt_info, 2*ncopies, dt, NULL); | |
3295 | return true; | |
3296 | } | |
3297 | ||
3298 | /** Transform. **/ | |
3299 | ||
3300 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3301 | fprintf (vect_dump, "transform type promotion operation. ncopies = %d.", | |
3302 | ncopies); | |
3303 | ||
3304 | /* Handle def. */ | |
b8698a0f | 3305 | /* In case of multi-step promotion, we first generate promotion operations |
ebfd146a | 3306 | to the intermediate types, and then from that types to the final one. |
b8698a0f L |
3307 | We store vector destination in VEC_DSTS in the correct order for |
3308 | recursive creation of promotion operations in | |
ebfd146a IR |
3309 | vect_create_vectorized_promotion_stmts(). Vector destinations are created |
3310 | according to TYPES recieved from supportable_widening_operation(). */ | |
3311 | if (multi_step_cvt) | |
3312 | vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1); | |
3313 | else | |
3314 | vec_dsts = VEC_alloc (tree, heap, 1); | |
3315 | ||
3316 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
3317 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
3318 | ||
3319 | if (multi_step_cvt) | |
3320 | { | |
3321 | for (i = VEC_length (tree, interm_types) - 1; | |
3322 | VEC_iterate (tree, interm_types, i, intermediate_type); i--) | |
3323 | { | |
3324 | vec_dest = vect_create_destination_var (scalar_dest, | |
3325 | intermediate_type); | |
3326 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
3327 | } | |
3328 | } | |
b8698a0f | 3329 | |
ebfd146a IR |
3330 | if (!slp_node) |
3331 | { | |
b8698a0f | 3332 | vec_oprnds0 = VEC_alloc (tree, heap, |
ebfd146a IR |
3333 | (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); |
3334 | if (op_type == binary_op) | |
3335 | vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
3336 | } | |
3337 | ||
3338 | /* In case the vectorization factor (VF) is bigger than the number | |
3339 | of elements that we can fit in a vectype (nunits), we have to generate | |
3340 | more than one vector stmt - i.e - we need to "unroll" the | |
3341 | vector stmt by a factor VF/nunits. */ | |
3342 | ||
3343 | prev_stmt_info = NULL; | |
3344 | for (j = 0; j < ncopies; j++) | |
3345 | { | |
3346 | /* Handle uses. */ | |
3347 | if (j == 0) | |
3348 | { | |
3349 | if (slp_node) | |
9dc3f7de IR |
3350 | vect_get_slp_defs (op0, op1, slp_node, &vec_oprnds0, |
3351 | &vec_oprnds1, -1); | |
ebfd146a IR |
3352 | else |
3353 | { | |
3354 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
3355 | VEC_quick_push (tree, vec_oprnds0, vec_oprnd0); | |
3356 | if (op_type == binary_op) | |
3357 | { | |
3358 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, NULL); | |
3359 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
3360 | } | |
3361 | } | |
3362 | } | |
3363 | else | |
3364 | { | |
3365 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
3366 | VEC_replace (tree, vec_oprnds0, 0, vec_oprnd0); | |
3367 | if (op_type == binary_op) | |
3368 | { | |
3369 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd1); | |
3370 | VEC_replace (tree, vec_oprnds1, 0, vec_oprnd1); | |
3371 | } | |
3372 | } | |
3373 | ||
9dc3f7de | 3374 | /* Arguments are ready. Create the new vector stmts. */ |
ebfd146a IR |
3375 | tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts); |
3376 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, &vec_oprnds1, | |
b8698a0f | 3377 | multi_step_cvt, stmt, |
ebfd146a IR |
3378 | tmp_vec_dsts, |
3379 | gsi, slp_node, code1, code2, | |
3380 | decl1, decl2, op_type, | |
3381 | &prev_stmt_info); | |
3382 | } | |
3383 | ||
3384 | VEC_free (tree, heap, vec_dsts); | |
3385 | VEC_free (tree, heap, tmp_vec_dsts); | |
3386 | VEC_free (tree, heap, interm_types); | |
3387 | VEC_free (tree, heap, vec_oprnds0); | |
3388 | VEC_free (tree, heap, vec_oprnds1); | |
3389 | ||
3390 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
3391 | return true; | |
3392 | } | |
3393 | ||
3394 | ||
3395 | /* Function vectorizable_store. | |
3396 | ||
b8698a0f L |
3397 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
3398 | can be vectorized. | |
3399 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
3400 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3401 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3402 | ||
3403 | static bool | |
3404 | vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
3405 | slp_tree slp_node) | |
3406 | { | |
3407 | tree scalar_dest; | |
3408 | tree data_ref; | |
3409 | tree op; | |
3410 | tree vec_oprnd = NULL_TREE; | |
3411 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3412 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
3413 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
272c6793 | 3414 | tree elem_type; |
ebfd146a | 3415 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 3416 | struct loop *loop = NULL; |
ebfd146a IR |
3417 | enum machine_mode vec_mode; |
3418 | tree dummy; | |
3419 | enum dr_alignment_support alignment_support_scheme; | |
3420 | tree def; | |
3421 | gimple def_stmt; | |
3422 | enum vect_def_type dt; | |
3423 | stmt_vec_info prev_stmt_info = NULL; | |
3424 | tree dataref_ptr = NULL_TREE; | |
3425 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
3426 | int ncopies; | |
3427 | int j; | |
3428 | gimple next_stmt, first_stmt = NULL; | |
3429 | bool strided_store = false; | |
272c6793 | 3430 | bool store_lanes_p = false; |
ebfd146a IR |
3431 | unsigned int group_size, i; |
3432 | VEC(tree,heap) *dr_chain = NULL, *oprnds = NULL, *result_chain = NULL; | |
3433 | bool inv_p; | |
3434 | VEC(tree,heap) *vec_oprnds = NULL; | |
3435 | bool slp = (slp_node != NULL); | |
ebfd146a | 3436 | unsigned int vec_num; |
a70d6342 | 3437 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
272c6793 | 3438 | tree aggr_type; |
a70d6342 IR |
3439 | |
3440 | if (loop_vinfo) | |
3441 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
3442 | |
3443 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3444 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3445 | case of SLP. */ | |
437f4a00 | 3446 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
3447 | ncopies = 1; |
3448 | else | |
3449 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3450 | ||
3451 | gcc_assert (ncopies >= 1); | |
3452 | ||
3453 | /* FORNOW. This restriction should be relaxed. */ | |
a70d6342 | 3454 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) |
ebfd146a IR |
3455 | { |
3456 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3457 | fprintf (vect_dump, "multiple types in nested loop."); | |
3458 | return false; | |
3459 | } | |
3460 | ||
a70d6342 | 3461 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3462 | return false; |
3463 | ||
8644a673 | 3464 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3465 | return false; |
3466 | ||
3467 | /* Is vectorizable store? */ | |
3468 | ||
3469 | if (!is_gimple_assign (stmt)) | |
3470 | return false; | |
3471 | ||
3472 | scalar_dest = gimple_assign_lhs (stmt); | |
3473 | if (TREE_CODE (scalar_dest) != ARRAY_REF | |
3474 | && TREE_CODE (scalar_dest) != INDIRECT_REF | |
e9dbe7bb IR |
3475 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
3476 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
3477 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
3478 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
3479 | return false; |
3480 | ||
3481 | gcc_assert (gimple_assign_single_p (stmt)); | |
3482 | op = gimple_assign_rhs1 (stmt); | |
a70d6342 | 3483 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt)) |
ebfd146a IR |
3484 | { |
3485 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3486 | fprintf (vect_dump, "use not simple."); | |
3487 | return false; | |
3488 | } | |
3489 | ||
3490 | /* The scalar rhs type needs to be trivially convertible to the vector | |
3491 | component type. This should always be the case. */ | |
272c6793 RS |
3492 | elem_type = TREE_TYPE (vectype); |
3493 | if (!useless_type_conversion_p (elem_type, TREE_TYPE (op))) | |
b8698a0f | 3494 | { |
ebfd146a IR |
3495 | if (vect_print_dump_info (REPORT_DETAILS)) |
3496 | fprintf (vect_dump, "??? operands of different types"); | |
3497 | return false; | |
3498 | } | |
3499 | ||
3500 | vec_mode = TYPE_MODE (vectype); | |
3501 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
3502 | (e.g. - array initialization with 0). */ | |
947131ba | 3503 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
3504 | return false; |
3505 | ||
3506 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
3507 | return false; | |
3508 | ||
a1e53f3f L |
3509 | if (tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0) |
3510 | { | |
3511 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3512 | fprintf (vect_dump, "negative step for store."); | |
3513 | return false; | |
3514 | } | |
3515 | ||
ebfd146a IR |
3516 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) |
3517 | { | |
3518 | strided_store = true; | |
e14c1050 | 3519 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
3520 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
3521 | { | |
e14c1050 | 3522 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
3523 | if (vect_store_lanes_supported (vectype, group_size)) |
3524 | store_lanes_p = true; | |
3525 | else if (!vect_strided_store_supported (vectype, group_size)) | |
b602d918 RS |
3526 | return false; |
3527 | } | |
b8698a0f | 3528 | |
ebfd146a IR |
3529 | if (first_stmt == stmt) |
3530 | { | |
3531 | /* STMT is the leader of the group. Check the operands of all the | |
3532 | stmts of the group. */ | |
e14c1050 | 3533 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
3534 | while (next_stmt) |
3535 | { | |
3536 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
3537 | op = gimple_assign_rhs1 (next_stmt); | |
b8698a0f | 3538 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, |
a70d6342 | 3539 | &def, &dt)) |
ebfd146a IR |
3540 | { |
3541 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3542 | fprintf (vect_dump, "use not simple."); | |
3543 | return false; | |
3544 | } | |
e14c1050 | 3545 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
3546 | } |
3547 | } | |
3548 | } | |
3549 | ||
3550 | if (!vec_stmt) /* transformation not required. */ | |
3551 | { | |
3552 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
272c6793 | 3553 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, NULL); |
ebfd146a IR |
3554 | return true; |
3555 | } | |
3556 | ||
3557 | /** Transform. **/ | |
3558 | ||
3559 | if (strided_store) | |
3560 | { | |
3561 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 3562 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 3563 | |
e14c1050 | 3564 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
3565 | |
3566 | /* FORNOW */ | |
a70d6342 | 3567 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
3568 | |
3569 | /* We vectorize all the stmts of the interleaving group when we | |
3570 | reach the last stmt in the group. */ | |
e14c1050 IR |
3571 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
3572 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
3573 | && !slp) |
3574 | { | |
3575 | *vec_stmt = NULL; | |
3576 | return true; | |
3577 | } | |
3578 | ||
3579 | if (slp) | |
4b5caab7 IR |
3580 | { |
3581 | strided_store = false; | |
3582 | /* VEC_NUM is the number of vect stmts to be created for this | |
3583 | group. */ | |
3584 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
3585 | first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0); | |
3586 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
3587 | } | |
ebfd146a | 3588 | else |
4b5caab7 IR |
3589 | /* VEC_NUM is the number of vect stmts to be created for this |
3590 | group. */ | |
ebfd146a IR |
3591 | vec_num = group_size; |
3592 | } | |
b8698a0f | 3593 | else |
ebfd146a IR |
3594 | { |
3595 | first_stmt = stmt; | |
3596 | first_dr = dr; | |
3597 | group_size = vec_num = 1; | |
ebfd146a | 3598 | } |
b8698a0f | 3599 | |
ebfd146a IR |
3600 | if (vect_print_dump_info (REPORT_DETAILS)) |
3601 | fprintf (vect_dump, "transform store. ncopies = %d",ncopies); | |
3602 | ||
3603 | dr_chain = VEC_alloc (tree, heap, group_size); | |
3604 | oprnds = VEC_alloc (tree, heap, group_size); | |
3605 | ||
720f5239 | 3606 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 3607 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
3608 | /* Targets with store-lane instructions must not require explicit |
3609 | realignment. */ | |
3610 | gcc_assert (!store_lanes_p | |
3611 | || alignment_support_scheme == dr_aligned | |
3612 | || alignment_support_scheme == dr_unaligned_supported); | |
3613 | ||
3614 | if (store_lanes_p) | |
3615 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
3616 | else | |
3617 | aggr_type = vectype; | |
ebfd146a IR |
3618 | |
3619 | /* In case the vectorization factor (VF) is bigger than the number | |
3620 | of elements that we can fit in a vectype (nunits), we have to generate | |
3621 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 3622 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
3623 | vect_get_vec_def_for_copy_stmt. */ |
3624 | ||
3625 | /* In case of interleaving (non-unit strided access): | |
3626 | ||
3627 | S1: &base + 2 = x2 | |
3628 | S2: &base = x0 | |
3629 | S3: &base + 1 = x1 | |
3630 | S4: &base + 3 = x3 | |
3631 | ||
3632 | We create vectorized stores starting from base address (the access of the | |
3633 | first stmt in the chain (S2 in the above example), when the last store stmt | |
3634 | of the chain (S4) is reached: | |
3635 | ||
3636 | VS1: &base = vx2 | |
3637 | VS2: &base + vec_size*1 = vx0 | |
3638 | VS3: &base + vec_size*2 = vx1 | |
3639 | VS4: &base + vec_size*3 = vx3 | |
3640 | ||
3641 | Then permutation statements are generated: | |
3642 | ||
3643 | VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 > | |
3644 | VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 > | |
3645 | ... | |
b8698a0f | 3646 | |
ebfd146a IR |
3647 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
3648 | (the order of the data-refs in the output of vect_permute_store_chain | |
3649 | corresponds to the order of scalar stmts in the interleaving chain - see | |
3650 | the documentation of vect_permute_store_chain()). | |
3651 | ||
3652 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 3653 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 3654 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 3655 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
3656 | */ |
3657 | ||
3658 | prev_stmt_info = NULL; | |
3659 | for (j = 0; j < ncopies; j++) | |
3660 | { | |
3661 | gimple new_stmt; | |
3662 | gimple ptr_incr; | |
3663 | ||
3664 | if (j == 0) | |
3665 | { | |
3666 | if (slp) | |
3667 | { | |
3668 | /* Get vectorized arguments for SLP_NODE. */ | |
9dc3f7de IR |
3669 | vect_get_slp_defs (NULL_TREE, NULL_TREE, slp_node, &vec_oprnds, |
3670 | NULL, -1); | |
ebfd146a IR |
3671 | |
3672 | vec_oprnd = VEC_index (tree, vec_oprnds, 0); | |
3673 | } | |
3674 | else | |
3675 | { | |
b8698a0f L |
3676 | /* For interleaved stores we collect vectorized defs for all the |
3677 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
3678 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
3679 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
3680 | ||
3681 | If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and | |
3682 | OPRNDS are of size 1. */ | |
b8698a0f | 3683 | next_stmt = first_stmt; |
ebfd146a IR |
3684 | for (i = 0; i < group_size; i++) |
3685 | { | |
b8698a0f L |
3686 | /* Since gaps are not supported for interleaved stores, |
3687 | GROUP_SIZE is the exact number of stmts in the chain. | |
3688 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
3689 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
3690 | iteration of the loop will be executed. */ |
3691 | gcc_assert (next_stmt | |
3692 | && gimple_assign_single_p (next_stmt)); | |
3693 | op = gimple_assign_rhs1 (next_stmt); | |
3694 | ||
b8698a0f | 3695 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt, |
ebfd146a | 3696 | NULL); |
b8698a0f L |
3697 | VEC_quick_push(tree, dr_chain, vec_oprnd); |
3698 | VEC_quick_push(tree, oprnds, vec_oprnd); | |
e14c1050 | 3699 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
3700 | } |
3701 | } | |
3702 | ||
3703 | /* We should have catched mismatched types earlier. */ | |
3704 | gcc_assert (useless_type_conversion_p (vectype, | |
3705 | TREE_TYPE (vec_oprnd))); | |
272c6793 | 3706 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, aggr_type, NULL, |
920e8172 RS |
3707 | NULL_TREE, &dummy, gsi, |
3708 | &ptr_incr, false, &inv_p); | |
a70d6342 | 3709 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 3710 | } |
b8698a0f | 3711 | else |
ebfd146a | 3712 | { |
b8698a0f L |
3713 | /* For interleaved stores we created vectorized defs for all the |
3714 | defs stored in OPRNDS in the previous iteration (previous copy). | |
3715 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
3716 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
3717 | next copy. | |
3718 | If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and | |
3719 | OPRNDS are of size 1. */ | |
3720 | for (i = 0; i < group_size; i++) | |
3721 | { | |
3722 | op = VEC_index (tree, oprnds, i); | |
b8698a0f | 3723 | vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, |
a70d6342 | 3724 | &dt); |
b8698a0f | 3725 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
ebfd146a IR |
3726 | VEC_replace(tree, dr_chain, i, vec_oprnd); |
3727 | VEC_replace(tree, oprnds, i, vec_oprnd); | |
3728 | } | |
272c6793 RS |
3729 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
3730 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
3731 | } |
3732 | ||
272c6793 | 3733 | if (store_lanes_p) |
ebfd146a | 3734 | { |
272c6793 | 3735 | tree vec_array; |
267d3070 | 3736 | |
272c6793 RS |
3737 | /* Combine all the vectors into an array. */ |
3738 | vec_array = create_vector_array (vectype, vec_num); | |
3739 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 3740 | { |
272c6793 RS |
3741 | vec_oprnd = VEC_index (tree, dr_chain, i); |
3742 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); | |
267d3070 | 3743 | } |
b8698a0f | 3744 | |
272c6793 RS |
3745 | /* Emit: |
3746 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
3747 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
3748 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
3749 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 RS |
3750 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3751 | mark_symbols_for_renaming (new_stmt); | |
272c6793 RS |
3752 | } |
3753 | else | |
3754 | { | |
3755 | new_stmt = NULL; | |
3756 | if (strided_store) | |
3757 | { | |
3758 | result_chain = VEC_alloc (tree, heap, group_size); | |
3759 | /* Permute. */ | |
3760 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
3761 | &result_chain); | |
3762 | } | |
c2d7ab2a | 3763 | |
272c6793 RS |
3764 | next_stmt = first_stmt; |
3765 | for (i = 0; i < vec_num; i++) | |
3766 | { | |
3767 | struct ptr_info_def *pi; | |
3768 | ||
3769 | if (i > 0) | |
3770 | /* Bump the vector pointer. */ | |
3771 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
3772 | stmt, NULL_TREE); | |
3773 | ||
3774 | if (slp) | |
3775 | vec_oprnd = VEC_index (tree, vec_oprnds, i); | |
3776 | else if (strided_store) | |
3777 | /* For strided stores vectorized defs are interleaved in | |
3778 | vect_permute_store_chain(). */ | |
3779 | vec_oprnd = VEC_index (tree, result_chain, i); | |
3780 | ||
3781 | data_ref = build2 (MEM_REF, TREE_TYPE (vec_oprnd), dataref_ptr, | |
3782 | build_int_cst (reference_alias_ptr_type | |
3783 | (DR_REF (first_dr)), 0)); | |
3784 | pi = get_ptr_info (dataref_ptr); | |
3785 | pi->align = TYPE_ALIGN_UNIT (vectype); | |
3786 | if (aligned_access_p (first_dr)) | |
3787 | pi->misalign = 0; | |
3788 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
3789 | { | |
3790 | TREE_TYPE (data_ref) | |
3791 | = build_aligned_type (TREE_TYPE (data_ref), | |
3792 | TYPE_ALIGN (elem_type)); | |
3793 | pi->align = TYPE_ALIGN_UNIT (elem_type); | |
3794 | pi->misalign = 0; | |
3795 | } | |
3796 | else | |
3797 | { | |
3798 | TREE_TYPE (data_ref) | |
3799 | = build_aligned_type (TREE_TYPE (data_ref), | |
3800 | TYPE_ALIGN (elem_type)); | |
3801 | pi->misalign = DR_MISALIGNMENT (first_dr); | |
3802 | } | |
c2d7ab2a | 3803 | |
272c6793 RS |
3804 | /* Arguments are ready. Create the new vector stmt. */ |
3805 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
3806 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3807 | mark_symbols_for_renaming (new_stmt); | |
3808 | ||
3809 | if (slp) | |
3810 | continue; | |
3811 | ||
e14c1050 | 3812 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
3813 | if (!next_stmt) |
3814 | break; | |
3815 | } | |
ebfd146a | 3816 | } |
1da0876c RS |
3817 | if (!slp) |
3818 | { | |
3819 | if (j == 0) | |
3820 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3821 | else | |
3822 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3823 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3824 | } | |
ebfd146a IR |
3825 | } |
3826 | ||
b8698a0f L |
3827 | VEC_free (tree, heap, dr_chain); |
3828 | VEC_free (tree, heap, oprnds); | |
ebfd146a | 3829 | if (result_chain) |
b8698a0f | 3830 | VEC_free (tree, heap, result_chain); |
ff802fa1 IR |
3831 | if (vec_oprnds) |
3832 | VEC_free (tree, heap, vec_oprnds); | |
ebfd146a IR |
3833 | |
3834 | return true; | |
3835 | } | |
3836 | ||
a1e53f3f L |
3837 | /* Given a vector type VECTYPE returns a builtin DECL to be used |
3838 | for vector permutation and stores a mask into *MASK that implements | |
3839 | reversal of the vector elements. If that is impossible to do | |
3840 | returns NULL (and *MASK is unchanged). */ | |
3841 | ||
3842 | static tree | |
3843 | perm_mask_for_reverse (tree vectype, tree *mask) | |
3844 | { | |
3845 | tree builtin_decl; | |
3846 | tree mask_element_type, mask_type; | |
3847 | tree mask_vec = NULL; | |
3848 | int i; | |
3849 | int nunits; | |
3850 | if (!targetm.vectorize.builtin_vec_perm) | |
3851 | return NULL; | |
3852 | ||
3853 | builtin_decl = targetm.vectorize.builtin_vec_perm (vectype, | |
3854 | &mask_element_type); | |
3855 | if (!builtin_decl || !mask_element_type) | |
3856 | return NULL; | |
3857 | ||
3858 | mask_type = get_vectype_for_scalar_type (mask_element_type); | |
3859 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
bb67d9c7 RG |
3860 | if (!mask_type |
3861 | || TYPE_VECTOR_SUBPARTS (vectype) != TYPE_VECTOR_SUBPARTS (mask_type)) | |
a1e53f3f L |
3862 | return NULL; |
3863 | ||
3864 | for (i = 0; i < nunits; i++) | |
3865 | mask_vec = tree_cons (NULL, build_int_cst (mask_element_type, i), mask_vec); | |
3866 | mask_vec = build_vector (mask_type, mask_vec); | |
3867 | ||
3868 | if (!targetm.vectorize.builtin_vec_perm_ok (vectype, mask_vec)) | |
3869 | return NULL; | |
3870 | if (mask) | |
3871 | *mask = mask_vec; | |
3872 | return builtin_decl; | |
3873 | } | |
3874 | ||
3875 | /* Given a vector variable X, that was generated for the scalar LHS of | |
3876 | STMT, generate instructions to reverse the vector elements of X, | |
3877 | insert them a *GSI and return the permuted vector variable. */ | |
3878 | ||
3879 | static tree | |
3880 | reverse_vec_elements (tree x, gimple stmt, gimple_stmt_iterator *gsi) | |
3881 | { | |
3882 | tree vectype = TREE_TYPE (x); | |
3883 | tree mask_vec, builtin_decl; | |
3884 | tree perm_dest, data_ref; | |
3885 | gimple perm_stmt; | |
3886 | ||
3887 | builtin_decl = perm_mask_for_reverse (vectype, &mask_vec); | |
3888 | ||
3889 | perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype); | |
3890 | ||
3891 | /* Generate the permute statement. */ | |
3892 | perm_stmt = gimple_build_call (builtin_decl, 3, x, x, mask_vec); | |
2a2651b7 RG |
3893 | if (!useless_type_conversion_p (vectype, |
3894 | TREE_TYPE (TREE_TYPE (builtin_decl)))) | |
3895 | { | |
3896 | tree tem = create_tmp_reg (TREE_TYPE (TREE_TYPE (builtin_decl)), NULL); | |
3897 | tem = make_ssa_name (tem, perm_stmt); | |
3898 | gimple_call_set_lhs (perm_stmt, tem); | |
3899 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); | |
3900 | perm_stmt = gimple_build_assign (NULL_TREE, | |
3901 | build1 (VIEW_CONVERT_EXPR, | |
3902 | vectype, tem)); | |
3903 | } | |
a1e53f3f | 3904 | data_ref = make_ssa_name (perm_dest, perm_stmt); |
2a2651b7 | 3905 | gimple_set_lhs (perm_stmt, data_ref); |
a1e53f3f L |
3906 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
3907 | ||
3908 | return data_ref; | |
3909 | } | |
3910 | ||
ebfd146a IR |
3911 | /* vectorizable_load. |
3912 | ||
b8698a0f L |
3913 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
3914 | can be vectorized. | |
3915 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
3916 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3917 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3918 | ||
3919 | static bool | |
3920 | vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
3921 | slp_tree slp_node, slp_instance slp_node_instance) | |
3922 | { | |
3923 | tree scalar_dest; | |
3924 | tree vec_dest = NULL; | |
3925 | tree data_ref = NULL; | |
3926 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 3927 | stmt_vec_info prev_stmt_info; |
ebfd146a | 3928 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 3929 | struct loop *loop = NULL; |
ebfd146a | 3930 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 3931 | bool nested_in_vect_loop = false; |
ebfd146a IR |
3932 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; |
3933 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
272c6793 | 3934 | tree elem_type; |
ebfd146a | 3935 | tree new_temp; |
947131ba | 3936 | enum machine_mode mode; |
ebfd146a IR |
3937 | gimple new_stmt = NULL; |
3938 | tree dummy; | |
3939 | enum dr_alignment_support alignment_support_scheme; | |
3940 | tree dataref_ptr = NULL_TREE; | |
3941 | gimple ptr_incr; | |
3942 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
3943 | int ncopies; | |
3944 | int i, j, group_size; | |
3945 | tree msq = NULL_TREE, lsq; | |
3946 | tree offset = NULL_TREE; | |
3947 | tree realignment_token = NULL_TREE; | |
3948 | gimple phi = NULL; | |
3949 | VEC(tree,heap) *dr_chain = NULL; | |
3950 | bool strided_load = false; | |
272c6793 | 3951 | bool load_lanes_p = false; |
ebfd146a IR |
3952 | gimple first_stmt; |
3953 | tree scalar_type; | |
3954 | bool inv_p; | |
a1e53f3f | 3955 | bool negative; |
ebfd146a IR |
3956 | bool compute_in_loop = false; |
3957 | struct loop *at_loop; | |
3958 | int vec_num; | |
3959 | bool slp = (slp_node != NULL); | |
3960 | bool slp_perm = false; | |
3961 | enum tree_code code; | |
a70d6342 IR |
3962 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3963 | int vf; | |
272c6793 | 3964 | tree aggr_type; |
a70d6342 IR |
3965 | |
3966 | if (loop_vinfo) | |
3967 | { | |
3968 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
3969 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
3970 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3971 | } | |
3972 | else | |
3533e503 | 3973 | vf = 1; |
ebfd146a IR |
3974 | |
3975 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 3976 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 3977 | case of SLP. */ |
437f4a00 | 3978 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
3979 | ncopies = 1; |
3980 | else | |
3981 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3982 | ||
3983 | gcc_assert (ncopies >= 1); | |
3984 | ||
3985 | /* FORNOW. This restriction should be relaxed. */ | |
3986 | if (nested_in_vect_loop && ncopies > 1) | |
3987 | { | |
3988 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3989 | fprintf (vect_dump, "multiple types in nested loop."); | |
3990 | return false; | |
3991 | } | |
3992 | ||
a70d6342 | 3993 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3994 | return false; |
3995 | ||
8644a673 | 3996 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3997 | return false; |
3998 | ||
3999 | /* Is vectorizable load? */ | |
4000 | if (!is_gimple_assign (stmt)) | |
4001 | return false; | |
4002 | ||
4003 | scalar_dest = gimple_assign_lhs (stmt); | |
4004 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4005 | return false; | |
4006 | ||
4007 | code = gimple_assign_rhs_code (stmt); | |
4008 | if (code != ARRAY_REF | |
4009 | && code != INDIRECT_REF | |
e9dbe7bb IR |
4010 | && code != COMPONENT_REF |
4011 | && code != IMAGPART_EXPR | |
70f34814 RG |
4012 | && code != REALPART_EXPR |
4013 | && code != MEM_REF) | |
ebfd146a IR |
4014 | return false; |
4015 | ||
4016 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
4017 | return false; | |
4018 | ||
a1e53f3f L |
4019 | negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0; |
4020 | if (negative && ncopies > 1) | |
4021 | { | |
4022 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4023 | fprintf (vect_dump, "multiple types with negative step."); | |
4024 | return false; | |
4025 | } | |
4026 | ||
ebfd146a | 4027 | scalar_type = TREE_TYPE (DR_REF (dr)); |
947131ba | 4028 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
4029 | |
4030 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
4031 | (e.g. - data copies). */ | |
947131ba | 4032 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a IR |
4033 | { |
4034 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4035 | fprintf (vect_dump, "Aligned load, but unsupported type."); | |
4036 | return false; | |
4037 | } | |
4038 | ||
4039 | /* The vector component type needs to be trivially convertible to the | |
4040 | scalar lhs. This should always be the case. */ | |
272c6793 RS |
4041 | elem_type = TREE_TYPE (vectype); |
4042 | if (!useless_type_conversion_p (TREE_TYPE (scalar_dest), elem_type)) | |
b8698a0f | 4043 | { |
ebfd146a IR |
4044 | if (vect_print_dump_info (REPORT_DETAILS)) |
4045 | fprintf (vect_dump, "??? operands of different types"); | |
4046 | return false; | |
4047 | } | |
4048 | ||
4049 | /* Check if the load is a part of an interleaving chain. */ | |
4050 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) | |
4051 | { | |
4052 | strided_load = true; | |
4053 | /* FORNOW */ | |
4054 | gcc_assert (! nested_in_vect_loop); | |
4055 | ||
e14c1050 | 4056 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
4057 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
4058 | { | |
e14c1050 | 4059 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
4060 | if (vect_load_lanes_supported (vectype, group_size)) |
4061 | load_lanes_p = true; | |
4062 | else if (!vect_strided_load_supported (vectype, group_size)) | |
b602d918 RS |
4063 | return false; |
4064 | } | |
ebfd146a IR |
4065 | } |
4066 | ||
a1e53f3f L |
4067 | if (negative) |
4068 | { | |
4069 | gcc_assert (!strided_load); | |
4070 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
4071 | if (alignment_support_scheme != dr_aligned | |
4072 | && alignment_support_scheme != dr_unaligned_supported) | |
4073 | { | |
4074 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4075 | fprintf (vect_dump, "negative step but alignment required."); | |
4076 | return false; | |
4077 | } | |
4078 | if (!perm_mask_for_reverse (vectype, NULL)) | |
4079 | { | |
4080 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4081 | fprintf (vect_dump, "negative step and reversing not supported."); | |
4082 | return false; | |
4083 | } | |
4084 | } | |
4085 | ||
ebfd146a IR |
4086 | if (!vec_stmt) /* transformation not required. */ |
4087 | { | |
4088 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
272c6793 | 4089 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, NULL); |
ebfd146a IR |
4090 | return true; |
4091 | } | |
4092 | ||
4093 | if (vect_print_dump_info (REPORT_DETAILS)) | |
0ea25ecd | 4094 | fprintf (vect_dump, "transform load. ncopies = %d", ncopies); |
ebfd146a IR |
4095 | |
4096 | /** Transform. **/ | |
4097 | ||
4098 | if (strided_load) | |
4099 | { | |
e14c1050 | 4100 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
4101 | /* Check if the chain of loads is already vectorized. */ |
4102 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt))) | |
4103 | { | |
4104 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4105 | return true; | |
4106 | } | |
4107 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 4108 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a IR |
4109 | |
4110 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
4111 | if (slp) | |
4112 | { | |
4113 | strided_load = false; | |
4114 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
a70d6342 IR |
4115 | if (SLP_INSTANCE_LOAD_PERMUTATION (slp_node_instance)) |
4116 | slp_perm = true; | |
4117 | } | |
ebfd146a IR |
4118 | else |
4119 | vec_num = group_size; | |
ebfd146a IR |
4120 | } |
4121 | else | |
4122 | { | |
4123 | first_stmt = stmt; | |
4124 | first_dr = dr; | |
4125 | group_size = vec_num = 1; | |
4126 | } | |
4127 | ||
720f5239 | 4128 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 4129 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
4130 | /* Targets with load-lane instructions must not require explicit |
4131 | realignment. */ | |
4132 | gcc_assert (!load_lanes_p | |
4133 | || alignment_support_scheme == dr_aligned | |
4134 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
4135 | |
4136 | /* In case the vectorization factor (VF) is bigger than the number | |
4137 | of elements that we can fit in a vectype (nunits), we have to generate | |
4138 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 4139 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 4140 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 4141 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 4142 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
4143 | stmts that use the defs of the current stmt. The example below |
4144 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
4145 | need to create 4 vectorized stmts): | |
ebfd146a IR |
4146 | |
4147 | before vectorization: | |
4148 | RELATED_STMT VEC_STMT | |
4149 | S1: x = memref - - | |
4150 | S2: z = x + 1 - - | |
4151 | ||
4152 | step 1: vectorize stmt S1: | |
4153 | We first create the vector stmt VS1_0, and, as usual, record a | |
4154 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
4155 | Next, we create the vector stmt VS1_1, and record a pointer to | |
4156 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 4157 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
4158 | stmts and pointers: |
4159 | RELATED_STMT VEC_STMT | |
4160 | VS1_0: vx0 = memref0 VS1_1 - | |
4161 | VS1_1: vx1 = memref1 VS1_2 - | |
4162 | VS1_2: vx2 = memref2 VS1_3 - | |
4163 | VS1_3: vx3 = memref3 - - | |
4164 | S1: x = load - VS1_0 | |
4165 | S2: z = x + 1 - - | |
4166 | ||
b8698a0f L |
4167 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
4168 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
4169 | stmt S2. */ |
4170 | ||
4171 | /* In case of interleaving (non-unit strided access): | |
4172 | ||
4173 | S1: x2 = &base + 2 | |
4174 | S2: x0 = &base | |
4175 | S3: x1 = &base + 1 | |
4176 | S4: x3 = &base + 3 | |
4177 | ||
b8698a0f | 4178 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
4179 | starting from the access of the first stmt of the chain: |
4180 | ||
4181 | VS1: vx0 = &base | |
4182 | VS2: vx1 = &base + vec_size*1 | |
4183 | VS3: vx3 = &base + vec_size*2 | |
4184 | VS4: vx4 = &base + vec_size*3 | |
4185 | ||
4186 | Then permutation statements are generated: | |
4187 | ||
4188 | VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 > | |
4189 | VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 > | |
4190 | ... | |
4191 | ||
4192 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
4193 | (the order of the data-refs in the output of vect_permute_load_chain | |
4194 | corresponds to the order of scalar stmts in the interleaving chain - see | |
4195 | the documentation of vect_permute_load_chain()). | |
4196 | The generation of permutation stmts and recording them in | |
4197 | STMT_VINFO_VEC_STMT is done in vect_transform_strided_load(). | |
4198 | ||
b8698a0f | 4199 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
4200 | permutation stmts above are created for every copy. The result vector |
4201 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
4202 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
4203 | |
4204 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
4205 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
4206 | we generate the following code: | |
4207 | p = initial_addr; | |
4208 | indx = 0; | |
4209 | loop { | |
4210 | p = p + indx * vectype_size; | |
4211 | vec_dest = *(p); | |
4212 | indx = indx + 1; | |
4213 | } | |
4214 | ||
4215 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 4216 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
4217 | then generate the following code, in which the data in each iteration is |
4218 | obtained by two vector loads, one from the previous iteration, and one | |
4219 | from the current iteration: | |
4220 | p1 = initial_addr; | |
4221 | msq_init = *(floor(p1)) | |
4222 | p2 = initial_addr + VS - 1; | |
4223 | realignment_token = call target_builtin; | |
4224 | indx = 0; | |
4225 | loop { | |
4226 | p2 = p2 + indx * vectype_size | |
4227 | lsq = *(floor(p2)) | |
4228 | vec_dest = realign_load (msq, lsq, realignment_token) | |
4229 | indx = indx + 1; | |
4230 | msq = lsq; | |
4231 | } */ | |
4232 | ||
4233 | /* If the misalignment remains the same throughout the execution of the | |
4234 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 4235 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
4236 | This can only occur when vectorizing memory accesses in the inner-loop |
4237 | nested within an outer-loop that is being vectorized. */ | |
4238 | ||
a70d6342 | 4239 | if (loop && nested_in_vect_loop_p (loop, stmt) |
ebfd146a IR |
4240 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
4241 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) | |
4242 | { | |
4243 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
4244 | compute_in_loop = true; | |
4245 | } | |
4246 | ||
4247 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
4248 | || alignment_support_scheme == dr_explicit_realign) | |
4249 | && !compute_in_loop) | |
4250 | { | |
4251 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
4252 | alignment_support_scheme, NULL_TREE, | |
4253 | &at_loop); | |
4254 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
4255 | { | |
4256 | phi = SSA_NAME_DEF_STMT (msq); | |
4257 | offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
4258 | } | |
4259 | } | |
4260 | else | |
4261 | at_loop = loop; | |
4262 | ||
a1e53f3f L |
4263 | if (negative) |
4264 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
4265 | ||
272c6793 RS |
4266 | if (load_lanes_p) |
4267 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
4268 | else | |
4269 | aggr_type = vectype; | |
4270 | ||
ebfd146a IR |
4271 | prev_stmt_info = NULL; |
4272 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 4273 | { |
272c6793 | 4274 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 4275 | if (j == 0) |
272c6793 | 4276 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, |
920e8172 RS |
4277 | offset, &dummy, gsi, |
4278 | &ptr_incr, false, &inv_p); | |
ebfd146a | 4279 | else |
272c6793 RS |
4280 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
4281 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 4282 | |
5ce1ee7f RS |
4283 | if (strided_load || slp_perm) |
4284 | dr_chain = VEC_alloc (tree, heap, vec_num); | |
4285 | ||
272c6793 | 4286 | if (load_lanes_p) |
ebfd146a | 4287 | { |
272c6793 RS |
4288 | tree vec_array; |
4289 | ||
4290 | vec_array = create_vector_array (vectype, vec_num); | |
4291 | ||
4292 | /* Emit: | |
4293 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
4294 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
4295 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
4296 | gimple_call_set_lhs (new_stmt, vec_array); | |
4297 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4298 | mark_symbols_for_renaming (new_stmt); | |
ebfd146a | 4299 | |
272c6793 RS |
4300 | /* Extract each vector into an SSA_NAME. */ |
4301 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 4302 | { |
272c6793 RS |
4303 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
4304 | vec_array, i); | |
4305 | VEC_quick_push (tree, dr_chain, new_temp); | |
4306 | } | |
4307 | ||
4308 | /* Record the mapping between SSA_NAMEs and statements. */ | |
4309 | vect_record_strided_load_vectors (stmt, dr_chain); | |
4310 | } | |
4311 | else | |
4312 | { | |
4313 | for (i = 0; i < vec_num; i++) | |
4314 | { | |
4315 | if (i > 0) | |
4316 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
4317 | stmt, NULL_TREE); | |
4318 | ||
4319 | /* 2. Create the vector-load in the loop. */ | |
4320 | switch (alignment_support_scheme) | |
4321 | { | |
4322 | case dr_aligned: | |
4323 | case dr_unaligned_supported: | |
be1ac4ec | 4324 | { |
272c6793 RS |
4325 | struct ptr_info_def *pi; |
4326 | data_ref | |
4327 | = build2 (MEM_REF, vectype, dataref_ptr, | |
4328 | build_int_cst (reference_alias_ptr_type | |
4329 | (DR_REF (first_dr)), 0)); | |
4330 | pi = get_ptr_info (dataref_ptr); | |
4331 | pi->align = TYPE_ALIGN_UNIT (vectype); | |
4332 | if (alignment_support_scheme == dr_aligned) | |
4333 | { | |
4334 | gcc_assert (aligned_access_p (first_dr)); | |
4335 | pi->misalign = 0; | |
4336 | } | |
4337 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
4338 | { | |
4339 | TREE_TYPE (data_ref) | |
4340 | = build_aligned_type (TREE_TYPE (data_ref), | |
4341 | TYPE_ALIGN (elem_type)); | |
4342 | pi->align = TYPE_ALIGN_UNIT (elem_type); | |
4343 | pi->misalign = 0; | |
4344 | } | |
4345 | else | |
4346 | { | |
4347 | TREE_TYPE (data_ref) | |
4348 | = build_aligned_type (TREE_TYPE (data_ref), | |
4349 | TYPE_ALIGN (elem_type)); | |
4350 | pi->misalign = DR_MISALIGNMENT (first_dr); | |
4351 | } | |
4352 | break; | |
be1ac4ec | 4353 | } |
272c6793 | 4354 | case dr_explicit_realign: |
267d3070 | 4355 | { |
272c6793 RS |
4356 | tree ptr, bump; |
4357 | tree vs_minus_1; | |
4358 | ||
4359 | vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
4360 | ||
4361 | if (compute_in_loop) | |
4362 | msq = vect_setup_realignment (first_stmt, gsi, | |
4363 | &realignment_token, | |
4364 | dr_explicit_realign, | |
4365 | dataref_ptr, NULL); | |
4366 | ||
4367 | new_stmt = gimple_build_assign_with_ops | |
4368 | (BIT_AND_EXPR, NULL_TREE, dataref_ptr, | |
4369 | build_int_cst | |
4370 | (TREE_TYPE (dataref_ptr), | |
4371 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
4372 | ptr = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt); | |
4373 | gimple_assign_set_lhs (new_stmt, ptr); | |
4374 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4375 | data_ref | |
4376 | = build2 (MEM_REF, vectype, ptr, | |
4377 | build_int_cst (reference_alias_ptr_type | |
4378 | (DR_REF (first_dr)), 0)); | |
4379 | vec_dest = vect_create_destination_var (scalar_dest, | |
4380 | vectype); | |
4381 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
4382 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4383 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4384 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
4385 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
4386 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4387 | msq = new_temp; | |
4388 | ||
4389 | bump = size_binop (MULT_EXPR, vs_minus_1, | |
4390 | TYPE_SIZE_UNIT (scalar_type)); | |
4391 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); | |
4392 | new_stmt = gimple_build_assign_with_ops | |
4393 | (BIT_AND_EXPR, NULL_TREE, ptr, | |
4394 | build_int_cst | |
4395 | (TREE_TYPE (ptr), | |
4396 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
4397 | ptr = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt); | |
4398 | gimple_assign_set_lhs (new_stmt, ptr); | |
4399 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4400 | data_ref | |
4401 | = build2 (MEM_REF, vectype, ptr, | |
4402 | build_int_cst (reference_alias_ptr_type | |
4403 | (DR_REF (first_dr)), 0)); | |
4404 | break; | |
267d3070 | 4405 | } |
272c6793 RS |
4406 | case dr_explicit_realign_optimized: |
4407 | new_stmt = gimple_build_assign_with_ops | |
4408 | (BIT_AND_EXPR, NULL_TREE, dataref_ptr, | |
4409 | build_int_cst | |
4410 | (TREE_TYPE (dataref_ptr), | |
4411 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
4412 | new_temp = make_ssa_name (SSA_NAME_VAR (dataref_ptr), | |
4413 | new_stmt); | |
4414 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4415 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4416 | data_ref | |
4417 | = build2 (MEM_REF, vectype, new_temp, | |
4418 | build_int_cst (reference_alias_ptr_type | |
4419 | (DR_REF (first_dr)), 0)); | |
4420 | break; | |
4421 | default: | |
4422 | gcc_unreachable (); | |
4423 | } | |
ebfd146a | 4424 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 4425 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
4426 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4427 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4428 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 | 4429 | mark_symbols_for_renaming (new_stmt); |
ebfd146a | 4430 | |
272c6793 RS |
4431 | /* 3. Handle explicit realignment if necessary/supported. |
4432 | Create in loop: | |
4433 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
4434 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
4435 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 4436 | { |
272c6793 RS |
4437 | lsq = gimple_assign_lhs (new_stmt); |
4438 | if (!realignment_token) | |
4439 | realignment_token = dataref_ptr; | |
4440 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4441 | new_stmt | |
4442 | = gimple_build_assign_with_ops3 (REALIGN_LOAD_EXPR, | |
4443 | vec_dest, msq, lsq, | |
4444 | realignment_token); | |
4445 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4446 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4447 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4448 | ||
4449 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
4450 | { | |
4451 | gcc_assert (phi); | |
4452 | if (i == vec_num - 1 && j == ncopies - 1) | |
4453 | add_phi_arg (phi, lsq, | |
4454 | loop_latch_edge (containing_loop), | |
4455 | UNKNOWN_LOCATION); | |
4456 | msq = lsq; | |
4457 | } | |
ebfd146a | 4458 | } |
ebfd146a | 4459 | |
272c6793 RS |
4460 | /* 4. Handle invariant-load. */ |
4461 | if (inv_p && !bb_vinfo) | |
ebfd146a | 4462 | { |
272c6793 RS |
4463 | gcc_assert (!strided_load); |
4464 | gcc_assert (nested_in_vect_loop_p (loop, stmt)); | |
4465 | if (j == 0) | |
4466 | { | |
4467 | int k; | |
4468 | tree t = NULL_TREE; | |
4469 | tree vec_inv, bitpos, bitsize = TYPE_SIZE (scalar_type); | |
4470 | ||
4471 | /* CHECKME: bitpos depends on endianess? */ | |
4472 | bitpos = bitsize_zero_node; | |
4473 | vec_inv = build3 (BIT_FIELD_REF, scalar_type, new_temp, | |
4474 | bitsize, bitpos); | |
4475 | vec_dest = vect_create_destination_var (scalar_dest, | |
4476 | NULL_TREE); | |
4477 | new_stmt = gimple_build_assign (vec_dest, vec_inv); | |
4478 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4479 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4480 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4481 | ||
4482 | for (k = nunits - 1; k >= 0; --k) | |
4483 | t = tree_cons (NULL_TREE, new_temp, t); | |
4484 | /* FIXME: use build_constructor directly. */ | |
4485 | vec_inv = build_constructor_from_list (vectype, t); | |
4486 | new_temp = vect_init_vector (stmt, vec_inv, | |
4487 | vectype, gsi); | |
4488 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
4489 | } | |
4490 | else | |
4491 | gcc_unreachable (); /* FORNOW. */ | |
4492 | } | |
ebfd146a | 4493 | |
272c6793 RS |
4494 | if (negative) |
4495 | { | |
4496 | new_temp = reverse_vec_elements (new_temp, stmt, gsi); | |
ebfd146a IR |
4497 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
4498 | } | |
267d3070 | 4499 | |
272c6793 RS |
4500 | /* Collect vector loads and later create their permutation in |
4501 | vect_transform_strided_load (). */ | |
4502 | if (strided_load || slp_perm) | |
4503 | VEC_quick_push (tree, dr_chain, new_temp); | |
267d3070 | 4504 | |
272c6793 RS |
4505 | /* Store vector loads in the corresponding SLP_NODE. */ |
4506 | if (slp && !slp_perm) | |
4507 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), | |
4508 | new_stmt); | |
4509 | } | |
ebfd146a IR |
4510 | } |
4511 | ||
4512 | if (slp && !slp_perm) | |
4513 | continue; | |
4514 | ||
4515 | if (slp_perm) | |
4516 | { | |
a70d6342 | 4517 | if (!vect_transform_slp_perm_load (stmt, dr_chain, gsi, vf, |
ebfd146a IR |
4518 | slp_node_instance, false)) |
4519 | { | |
4520 | VEC_free (tree, heap, dr_chain); | |
4521 | return false; | |
4522 | } | |
4523 | } | |
4524 | else | |
4525 | { | |
4526 | if (strided_load) | |
4527 | { | |
272c6793 RS |
4528 | if (!load_lanes_p) |
4529 | vect_transform_strided_load (stmt, dr_chain, group_size, gsi); | |
ebfd146a | 4530 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
4531 | } |
4532 | else | |
4533 | { | |
4534 | if (j == 0) | |
4535 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4536 | else | |
4537 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4538 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4539 | } | |
4540 | } | |
5ce1ee7f RS |
4541 | if (dr_chain) |
4542 | VEC_free (tree, heap, dr_chain); | |
ebfd146a IR |
4543 | } |
4544 | ||
ebfd146a IR |
4545 | return true; |
4546 | } | |
4547 | ||
4548 | /* Function vect_is_simple_cond. | |
b8698a0f | 4549 | |
ebfd146a IR |
4550 | Input: |
4551 | LOOP - the loop that is being vectorized. | |
4552 | COND - Condition that is checked for simple use. | |
4553 | ||
4554 | Returns whether a COND can be vectorized. Checks whether | |
4555 | condition operands are supportable using vec_is_simple_use. */ | |
4556 | ||
4557 | static bool | |
4558 | vect_is_simple_cond (tree cond, loop_vec_info loop_vinfo) | |
4559 | { | |
4560 | tree lhs, rhs; | |
4561 | tree def; | |
4562 | enum vect_def_type dt; | |
4563 | ||
4564 | if (!COMPARISON_CLASS_P (cond)) | |
4565 | return false; | |
4566 | ||
4567 | lhs = TREE_OPERAND (cond, 0); | |
4568 | rhs = TREE_OPERAND (cond, 1); | |
4569 | ||
4570 | if (TREE_CODE (lhs) == SSA_NAME) | |
4571 | { | |
4572 | gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); | |
b8698a0f | 4573 | if (!vect_is_simple_use (lhs, loop_vinfo, NULL, &lhs_def_stmt, &def, |
a70d6342 | 4574 | &dt)) |
ebfd146a IR |
4575 | return false; |
4576 | } | |
4577 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
4578 | && TREE_CODE (lhs) != FIXED_CST) | |
4579 | return false; | |
4580 | ||
4581 | if (TREE_CODE (rhs) == SSA_NAME) | |
4582 | { | |
4583 | gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); | |
b8698a0f | 4584 | if (!vect_is_simple_use (rhs, loop_vinfo, NULL, &rhs_def_stmt, &def, |
a70d6342 | 4585 | &dt)) |
ebfd146a IR |
4586 | return false; |
4587 | } | |
4588 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST | |
4589 | && TREE_CODE (rhs) != FIXED_CST) | |
4590 | return false; | |
4591 | ||
4592 | return true; | |
4593 | } | |
4594 | ||
4595 | /* vectorizable_condition. | |
4596 | ||
b8698a0f L |
4597 | Check if STMT is conditional modify expression that can be vectorized. |
4598 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4599 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
4600 | at GSI. |
4601 | ||
4602 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
4603 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
4604 | else caluse if it is 2). | |
ebfd146a IR |
4605 | |
4606 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4607 | ||
4bbe8262 | 4608 | bool |
ebfd146a | 4609 | vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi, |
4bbe8262 | 4610 | gimple *vec_stmt, tree reduc_def, int reduc_index) |
ebfd146a IR |
4611 | { |
4612 | tree scalar_dest = NULL_TREE; | |
4613 | tree vec_dest = NULL_TREE; | |
4614 | tree op = NULL_TREE; | |
4615 | tree cond_expr, then_clause, else_clause; | |
4616 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4617 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
ff802fa1 IR |
4618 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
4619 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
ebfd146a IR |
4620 | tree vec_compare, vec_cond_expr; |
4621 | tree new_temp; | |
4622 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4623 | enum machine_mode vec_mode; | |
4624 | tree def; | |
a855b1b1 | 4625 | enum vect_def_type dt, dts[4]; |
ebfd146a IR |
4626 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
4627 | int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4628 | enum tree_code code; | |
a855b1b1 MM |
4629 | stmt_vec_info prev_stmt_info = NULL; |
4630 | int j; | |
ebfd146a | 4631 | |
a70d6342 IR |
4632 | /* FORNOW: unsupported in basic block SLP. */ |
4633 | gcc_assert (loop_vinfo); | |
b8698a0f | 4634 | |
437f4a00 IR |
4635 | /* FORNOW: SLP not supported. */ |
4636 | if (STMT_SLP_TYPE (stmt_info)) | |
4637 | return false; | |
4638 | ||
ebfd146a | 4639 | gcc_assert (ncopies >= 1); |
a855b1b1 | 4640 | if (reduc_index && ncopies > 1) |
ebfd146a IR |
4641 | return false; /* FORNOW */ |
4642 | ||
4643 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
4644 | return false; | |
4645 | ||
4bbe8262 IR |
4646 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4647 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
4648 | && reduc_def)) | |
ebfd146a IR |
4649 | return false; |
4650 | ||
ebfd146a | 4651 | /* FORNOW: not yet supported. */ |
b8698a0f | 4652 | if (STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a IR |
4653 | { |
4654 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4655 | fprintf (vect_dump, "value used after loop."); | |
4656 | return false; | |
4657 | } | |
4658 | ||
4659 | /* Is vectorizable conditional operation? */ | |
4660 | if (!is_gimple_assign (stmt)) | |
4661 | return false; | |
4662 | ||
4663 | code = gimple_assign_rhs_code (stmt); | |
4664 | ||
4665 | if (code != COND_EXPR) | |
4666 | return false; | |
4667 | ||
4668 | gcc_assert (gimple_assign_single_p (stmt)); | |
4669 | op = gimple_assign_rhs1 (stmt); | |
4670 | cond_expr = TREE_OPERAND (op, 0); | |
4671 | then_clause = TREE_OPERAND (op, 1); | |
4672 | else_clause = TREE_OPERAND (op, 2); | |
4673 | ||
4674 | if (!vect_is_simple_cond (cond_expr, loop_vinfo)) | |
4675 | return false; | |
4676 | ||
4677 | /* We do not handle two different vector types for the condition | |
4678 | and the values. */ | |
8533c9d8 SP |
4679 | if (!types_compatible_p (TREE_TYPE (TREE_OPERAND (cond_expr, 0)), |
4680 | TREE_TYPE (vectype))) | |
ebfd146a IR |
4681 | return false; |
4682 | ||
4683 | if (TREE_CODE (then_clause) == SSA_NAME) | |
4684 | { | |
4685 | gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause); | |
b8698a0f | 4686 | if (!vect_is_simple_use (then_clause, loop_vinfo, NULL, |
ebfd146a IR |
4687 | &then_def_stmt, &def, &dt)) |
4688 | return false; | |
4689 | } | |
b8698a0f | 4690 | else if (TREE_CODE (then_clause) != INTEGER_CST |
ebfd146a IR |
4691 | && TREE_CODE (then_clause) != REAL_CST |
4692 | && TREE_CODE (then_clause) != FIXED_CST) | |
4693 | return false; | |
4694 | ||
4695 | if (TREE_CODE (else_clause) == SSA_NAME) | |
4696 | { | |
4697 | gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause); | |
a70d6342 | 4698 | if (!vect_is_simple_use (else_clause, loop_vinfo, NULL, |
ebfd146a IR |
4699 | &else_def_stmt, &def, &dt)) |
4700 | return false; | |
4701 | } | |
b8698a0f | 4702 | else if (TREE_CODE (else_clause) != INTEGER_CST |
ebfd146a IR |
4703 | && TREE_CODE (else_clause) != REAL_CST |
4704 | && TREE_CODE (else_clause) != FIXED_CST) | |
4705 | return false; | |
4706 | ||
4707 | ||
4708 | vec_mode = TYPE_MODE (vectype); | |
4709 | ||
b8698a0f | 4710 | if (!vec_stmt) |
ebfd146a IR |
4711 | { |
4712 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
8e7aa1f9 | 4713 | return expand_vec_cond_expr_p (TREE_TYPE (op), vec_mode); |
ebfd146a IR |
4714 | } |
4715 | ||
4716 | /* Transform */ | |
4717 | ||
4718 | /* Handle def. */ | |
4719 | scalar_dest = gimple_assign_lhs (stmt); | |
4720 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4721 | ||
4722 | /* Handle cond expr. */ | |
a855b1b1 MM |
4723 | for (j = 0; j < ncopies; j++) |
4724 | { | |
4725 | gimple new_stmt; | |
4726 | if (j == 0) | |
4727 | { | |
4728 | gimple gtemp; | |
4729 | vec_cond_lhs = | |
4730 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), | |
4731 | stmt, NULL); | |
4732 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), loop_vinfo, | |
4733 | NULL, >emp, &def, &dts[0]); | |
4734 | vec_cond_rhs = | |
4735 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
4736 | stmt, NULL); | |
4737 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), loop_vinfo, | |
4738 | NULL, >emp, &def, &dts[1]); | |
4739 | if (reduc_index == 1) | |
4740 | vec_then_clause = reduc_def; | |
4741 | else | |
4742 | { | |
4743 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
4744 | stmt, NULL); | |
4745 | vect_is_simple_use (then_clause, loop_vinfo, | |
4746 | NULL, >emp, &def, &dts[2]); | |
4747 | } | |
4748 | if (reduc_index == 2) | |
4749 | vec_else_clause = reduc_def; | |
4750 | else | |
4751 | { | |
4752 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
4753 | stmt, NULL); | |
4754 | vect_is_simple_use (else_clause, loop_vinfo, | |
4755 | NULL, >emp, &def, &dts[3]); | |
4756 | } | |
4757 | } | |
4758 | else | |
4759 | { | |
4760 | vec_cond_lhs = vect_get_vec_def_for_stmt_copy (dts[0], vec_cond_lhs); | |
4761 | vec_cond_rhs = vect_get_vec_def_for_stmt_copy (dts[1], vec_cond_rhs); | |
4762 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], | |
4763 | vec_then_clause); | |
4764 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], | |
4765 | vec_else_clause); | |
4766 | } | |
4767 | ||
9dc3f7de | 4768 | /* Arguments are ready. Create the new vector stmt. */ |
a855b1b1 MM |
4769 | vec_compare = build2 (TREE_CODE (cond_expr), vectype, |
4770 | vec_cond_lhs, vec_cond_rhs); | |
4771 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, | |
4772 | vec_compare, vec_then_clause, vec_else_clause); | |
4773 | ||
4774 | new_stmt = gimple_build_assign (vec_dest, vec_cond_expr); | |
4775 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4776 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4777 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4778 | if (j == 0) | |
4779 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4780 | else | |
4781 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4782 | ||
4783 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4784 | } | |
b8698a0f | 4785 | |
ebfd146a IR |
4786 | return true; |
4787 | } | |
4788 | ||
4789 | ||
8644a673 | 4790 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
4791 | |
4792 | bool | |
a70d6342 | 4793 | vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 4794 | { |
8644a673 | 4795 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 4796 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 4797 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 4798 | bool ok; |
a70d6342 | 4799 | tree scalar_type, vectype; |
ebfd146a IR |
4800 | |
4801 | if (vect_print_dump_info (REPORT_DETAILS)) | |
ebfd146a | 4802 | { |
8644a673 IR |
4803 | fprintf (vect_dump, "==> examining statement: "); |
4804 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4805 | } | |
ebfd146a | 4806 | |
1825a1f3 | 4807 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 4808 | { |
1825a1f3 IR |
4809 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
4810 | fprintf (vect_dump, "not vectorized: stmt has volatile operands"); | |
4811 | ||
4812 | return false; | |
4813 | } | |
b8698a0f L |
4814 | |
4815 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
4816 | to include: |
4817 | - the COND_EXPR which is the loop exit condition | |
4818 | - any LABEL_EXPRs in the loop | |
b8698a0f | 4819 | - computations that are used only for array indexing or loop control. |
8644a673 IR |
4820 | In basic blocks we only analyze statements that are a part of some SLP |
4821 | instance, therefore, all the statements are relevant. */ | |
ebfd146a | 4822 | |
b8698a0f | 4823 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 4824 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a IR |
4825 | { |
4826 | if (vect_print_dump_info (REPORT_DETAILS)) | |
8644a673 | 4827 | fprintf (vect_dump, "irrelevant."); |
ebfd146a | 4828 | |
8644a673 IR |
4829 | return true; |
4830 | } | |
ebfd146a | 4831 | |
8644a673 IR |
4832 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
4833 | { | |
4834 | case vect_internal_def: | |
4835 | break; | |
ebfd146a | 4836 | |
8644a673 | 4837 | case vect_reduction_def: |
7c5222ff | 4838 | case vect_nested_cycle: |
a70d6342 | 4839 | gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer |
8644a673 | 4840 | || relevance == vect_used_in_outer_by_reduction |
a70d6342 | 4841 | || relevance == vect_unused_in_scope)); |
8644a673 IR |
4842 | break; |
4843 | ||
4844 | case vect_induction_def: | |
4845 | case vect_constant_def: | |
4846 | case vect_external_def: | |
4847 | case vect_unknown_def_type: | |
4848 | default: | |
4849 | gcc_unreachable (); | |
4850 | } | |
ebfd146a | 4851 | |
a70d6342 IR |
4852 | if (bb_vinfo) |
4853 | { | |
4854 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
4855 | ||
b690cc0f | 4856 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
a70d6342 IR |
4857 | if (vect_print_dump_info (REPORT_DETAILS)) |
4858 | { | |
4859 | fprintf (vect_dump, "get vectype for scalar type: "); | |
4860 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
4861 | } | |
4862 | ||
4863 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4864 | if (!vectype) | |
4865 | { | |
4866 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4867 | { | |
4868 | fprintf (vect_dump, "not SLPed: unsupported data-type "); | |
4869 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
4870 | } | |
4871 | return false; | |
4872 | } | |
4873 | ||
4874 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4875 | { | |
4876 | fprintf (vect_dump, "vectype: "); | |
4877 | print_generic_expr (vect_dump, vectype, TDF_SLIM); | |
4878 | } | |
4879 | ||
4880 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
4881 | } | |
4882 | ||
8644a673 | 4883 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 4884 | { |
8644a673 IR |
4885 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
4886 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info)); | |
4887 | *need_to_vectorize = true; | |
ebfd146a IR |
4888 | } |
4889 | ||
8644a673 | 4890 | ok = true; |
b8698a0f | 4891 | if (!bb_vinfo |
a70d6342 IR |
4892 | && (STMT_VINFO_RELEVANT_P (stmt_info) |
4893 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
8644a673 IR |
4894 | ok = (vectorizable_type_promotion (stmt, NULL, NULL, NULL) |
4895 | || vectorizable_type_demotion (stmt, NULL, NULL, NULL) | |
4896 | || vectorizable_conversion (stmt, NULL, NULL, NULL) | |
9dc3f7de | 4897 | || vectorizable_shift (stmt, NULL, NULL, NULL) |
8644a673 IR |
4898 | || vectorizable_operation (stmt, NULL, NULL, NULL) |
4899 | || vectorizable_assignment (stmt, NULL, NULL, NULL) | |
4900 | || vectorizable_load (stmt, NULL, NULL, NULL, NULL) | |
4901 | || vectorizable_call (stmt, NULL, NULL) | |
4902 | || vectorizable_store (stmt, NULL, NULL, NULL) | |
b5aeb3bb | 4903 | || vectorizable_reduction (stmt, NULL, NULL, NULL) |
4bbe8262 | 4904 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0)); |
a70d6342 IR |
4905 | else |
4906 | { | |
4907 | if (bb_vinfo) | |
57416708 | 4908 | ok = (vectorizable_shift (stmt, NULL, NULL, node) |
9dc3f7de | 4909 | || vectorizable_operation (stmt, NULL, NULL, node) |
a70d6342 IR |
4910 | || vectorizable_assignment (stmt, NULL, NULL, node) |
4911 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
4912 | || vectorizable_store (stmt, NULL, NULL, node)); | |
b8698a0f | 4913 | } |
8644a673 IR |
4914 | |
4915 | if (!ok) | |
ebfd146a | 4916 | { |
8644a673 IR |
4917 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
4918 | { | |
4919 | fprintf (vect_dump, "not vectorized: relevant stmt not "); | |
4920 | fprintf (vect_dump, "supported: "); | |
4921 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4922 | } | |
b8698a0f | 4923 | |
ebfd146a IR |
4924 | return false; |
4925 | } | |
4926 | ||
a70d6342 IR |
4927 | if (bb_vinfo) |
4928 | return true; | |
4929 | ||
8644a673 IR |
4930 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
4931 | need extra handling, except for vectorizable reductions. */ | |
4932 | if (STMT_VINFO_LIVE_P (stmt_info) | |
4933 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
4934 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 4935 | |
8644a673 | 4936 | if (!ok) |
ebfd146a | 4937 | { |
8644a673 IR |
4938 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
4939 | { | |
4940 | fprintf (vect_dump, "not vectorized: live stmt not "); | |
4941 | fprintf (vect_dump, "supported: "); | |
4942 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4943 | } | |
b8698a0f | 4944 | |
8644a673 | 4945 | return false; |
ebfd146a IR |
4946 | } |
4947 | ||
ebfd146a IR |
4948 | return true; |
4949 | } | |
4950 | ||
4951 | ||
4952 | /* Function vect_transform_stmt. | |
4953 | ||
4954 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
4955 | ||
4956 | bool | |
4957 | vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi, | |
b8698a0f | 4958 | bool *strided_store, slp_tree slp_node, |
ebfd146a IR |
4959 | slp_instance slp_node_instance) |
4960 | { | |
4961 | bool is_store = false; | |
4962 | gimple vec_stmt = NULL; | |
4963 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
039d9ea1 | 4964 | gimple orig_stmt_in_pattern, orig_scalar_stmt = stmt; |
ebfd146a | 4965 | bool done; |
ebfd146a IR |
4966 | |
4967 | switch (STMT_VINFO_TYPE (stmt_info)) | |
4968 | { | |
4969 | case type_demotion_vec_info_type: | |
4970 | done = vectorizable_type_demotion (stmt, gsi, &vec_stmt, slp_node); | |
4971 | gcc_assert (done); | |
4972 | break; | |
4973 | ||
4974 | case type_promotion_vec_info_type: | |
4975 | done = vectorizable_type_promotion (stmt, gsi, &vec_stmt, slp_node); | |
4976 | gcc_assert (done); | |
4977 | break; | |
4978 | ||
4979 | case type_conversion_vec_info_type: | |
4980 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
4981 | gcc_assert (done); | |
4982 | break; | |
4983 | ||
4984 | case induc_vec_info_type: | |
4985 | gcc_assert (!slp_node); | |
4986 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
4987 | gcc_assert (done); | |
4988 | break; | |
4989 | ||
9dc3f7de IR |
4990 | case shift_vec_info_type: |
4991 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
4992 | gcc_assert (done); | |
4993 | break; | |
4994 | ||
ebfd146a IR |
4995 | case op_vec_info_type: |
4996 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
4997 | gcc_assert (done); | |
4998 | break; | |
4999 | ||
5000 | case assignment_vec_info_type: | |
5001 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
5002 | gcc_assert (done); | |
5003 | break; | |
5004 | ||
5005 | case load_vec_info_type: | |
b8698a0f | 5006 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
5007 | slp_node_instance); |
5008 | gcc_assert (done); | |
5009 | break; | |
5010 | ||
5011 | case store_vec_info_type: | |
5012 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
5013 | gcc_assert (done); | |
5014 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) && !slp_node) | |
5015 | { | |
5016 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 5017 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
5018 | one are skipped, and there vec_stmt_info shouldn't be freed |
5019 | meanwhile. */ | |
5020 | *strided_store = true; | |
5021 | if (STMT_VINFO_VEC_STMT (stmt_info)) | |
5022 | is_store = true; | |
5023 | } | |
5024 | else | |
5025 | is_store = true; | |
5026 | break; | |
5027 | ||
5028 | case condition_vec_info_type: | |
5029 | gcc_assert (!slp_node); | |
4bbe8262 | 5030 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0); |
ebfd146a IR |
5031 | gcc_assert (done); |
5032 | break; | |
5033 | ||
5034 | case call_vec_info_type: | |
5035 | gcc_assert (!slp_node); | |
5036 | done = vectorizable_call (stmt, gsi, &vec_stmt); | |
039d9ea1 | 5037 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
5038 | break; |
5039 | ||
5040 | case reduc_vec_info_type: | |
b5aeb3bb | 5041 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
5042 | gcc_assert (done); |
5043 | break; | |
5044 | ||
5045 | default: | |
5046 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
5047 | { | |
5048 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5049 | fprintf (vect_dump, "stmt not supported."); | |
5050 | gcc_unreachable (); | |
5051 | } | |
5052 | } | |
5053 | ||
5054 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that | |
5055 | is being vectorized, but outside the immediately enclosing loop. */ | |
5056 | if (vec_stmt | |
a70d6342 IR |
5057 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
5058 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
5059 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
5060 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
5061 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 5062 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 5063 | vect_used_in_outer_by_reduction)) |
ebfd146a | 5064 | { |
a70d6342 IR |
5065 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
5066 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
5067 | imm_use_iterator imm_iter; |
5068 | use_operand_p use_p; | |
5069 | tree scalar_dest; | |
5070 | gimple exit_phi; | |
5071 | ||
5072 | if (vect_print_dump_info (REPORT_DETAILS)) | |
a70d6342 | 5073 | fprintf (vect_dump, "Record the vdef for outer-loop vectorization."); |
ebfd146a IR |
5074 | |
5075 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
5076 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
5077 | STMT). */ | |
5078 | if (gimple_code (stmt) == GIMPLE_PHI) | |
5079 | scalar_dest = PHI_RESULT (stmt); | |
5080 | else | |
5081 | scalar_dest = gimple_assign_lhs (stmt); | |
5082 | ||
5083 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
5084 | { | |
5085 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
5086 | { | |
5087 | exit_phi = USE_STMT (use_p); | |
5088 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
5089 | } | |
5090 | } | |
5091 | } | |
5092 | ||
5093 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
5094 | being vectorized. */ | |
5095 | if (STMT_VINFO_LIVE_P (stmt_info) | |
5096 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
5097 | { | |
5098 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
5099 | gcc_assert (done); | |
5100 | } | |
5101 | ||
5102 | if (vec_stmt) | |
5103 | { | |
5104 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; | |
5105 | orig_stmt_in_pattern = STMT_VINFO_RELATED_STMT (stmt_info); | |
5106 | if (orig_stmt_in_pattern) | |
5107 | { | |
5108 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt_in_pattern); | |
5109 | /* STMT was inserted by the vectorizer to replace a computation idiom. | |
b8698a0f L |
5110 | ORIG_STMT_IN_PATTERN is a stmt in the original sequence that |
5111 | computed this idiom. We need to record a pointer to VEC_STMT in | |
5112 | the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the | |
ebfd146a IR |
5113 | documentation of vect_pattern_recog. */ |
5114 | if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo)) | |
5115 | { | |
039d9ea1 IR |
5116 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) |
5117 | == orig_scalar_stmt); | |
ebfd146a IR |
5118 | STMT_VINFO_VEC_STMT (stmt_vinfo) = vec_stmt; |
5119 | } | |
5120 | } | |
5121 | } | |
5122 | ||
b8698a0f | 5123 | return is_store; |
ebfd146a IR |
5124 | } |
5125 | ||
5126 | ||
b8698a0f | 5127 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
5128 | stmt_vec_info. */ |
5129 | ||
5130 | void | |
5131 | vect_remove_stores (gimple first_stmt) | |
5132 | { | |
5133 | gimple next = first_stmt; | |
5134 | gimple tmp; | |
5135 | gimple_stmt_iterator next_si; | |
5136 | ||
5137 | while (next) | |
5138 | { | |
5139 | /* Free the attached stmt_vec_info and remove the stmt. */ | |
5140 | next_si = gsi_for_stmt (next); | |
5141 | gsi_remove (&next_si, true); | |
e14c1050 | 5142 | tmp = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next)); |
ebfd146a IR |
5143 | free_stmt_vec_info (next); |
5144 | next = tmp; | |
5145 | } | |
5146 | } | |
5147 | ||
5148 | ||
5149 | /* Function new_stmt_vec_info. | |
5150 | ||
5151 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
5152 | ||
5153 | stmt_vec_info | |
b8698a0f | 5154 | new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 5155 | bb_vec_info bb_vinfo) |
ebfd146a IR |
5156 | { |
5157 | stmt_vec_info res; | |
5158 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
5159 | ||
5160 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
5161 | STMT_VINFO_STMT (res) = stmt; | |
5162 | STMT_VINFO_LOOP_VINFO (res) = loop_vinfo; | |
a70d6342 | 5163 | STMT_VINFO_BB_VINFO (res) = bb_vinfo; |
8644a673 | 5164 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
5165 | STMT_VINFO_LIVE_P (res) = false; |
5166 | STMT_VINFO_VECTYPE (res) = NULL; | |
5167 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 5168 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
5169 | STMT_VINFO_IN_PATTERN_P (res) = false; |
5170 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
5171 | STMT_VINFO_DATA_REF (res) = NULL; | |
5172 | ||
5173 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
5174 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
5175 | STMT_VINFO_DR_INIT (res) = NULL; | |
5176 | STMT_VINFO_DR_STEP (res) = NULL; | |
5177 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
5178 | ||
5179 | if (gimple_code (stmt) == GIMPLE_PHI | |
5180 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
5181 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
5182 | else | |
8644a673 IR |
5183 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
5184 | ||
ebfd146a IR |
5185 | STMT_VINFO_SAME_ALIGN_REFS (res) = VEC_alloc (dr_p, heap, 5); |
5186 | STMT_VINFO_INSIDE_OF_LOOP_COST (res) = 0; | |
5187 | STMT_VINFO_OUTSIDE_OF_LOOP_COST (res) = 0; | |
32e8bb8e | 5188 | STMT_SLP_TYPE (res) = loop_vect; |
e14c1050 IR |
5189 | GROUP_FIRST_ELEMENT (res) = NULL; |
5190 | GROUP_NEXT_ELEMENT (res) = NULL; | |
5191 | GROUP_SIZE (res) = 0; | |
5192 | GROUP_STORE_COUNT (res) = 0; | |
5193 | GROUP_GAP (res) = 0; | |
5194 | GROUP_SAME_DR_STMT (res) = NULL; | |
5195 | GROUP_READ_WRITE_DEPENDENCE (res) = false; | |
ebfd146a IR |
5196 | |
5197 | return res; | |
5198 | } | |
5199 | ||
5200 | ||
5201 | /* Create a hash table for stmt_vec_info. */ | |
5202 | ||
5203 | void | |
5204 | init_stmt_vec_info_vec (void) | |
5205 | { | |
5206 | gcc_assert (!stmt_vec_info_vec); | |
5207 | stmt_vec_info_vec = VEC_alloc (vec_void_p, heap, 50); | |
5208 | } | |
5209 | ||
5210 | ||
5211 | /* Free hash table for stmt_vec_info. */ | |
5212 | ||
5213 | void | |
5214 | free_stmt_vec_info_vec (void) | |
5215 | { | |
5216 | gcc_assert (stmt_vec_info_vec); | |
5217 | VEC_free (vec_void_p, heap, stmt_vec_info_vec); | |
5218 | } | |
5219 | ||
5220 | ||
5221 | /* Free stmt vectorization related info. */ | |
5222 | ||
5223 | void | |
5224 | free_stmt_vec_info (gimple stmt) | |
5225 | { | |
5226 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5227 | ||
5228 | if (!stmt_info) | |
5229 | return; | |
5230 | ||
5231 | VEC_free (dr_p, heap, STMT_VINFO_SAME_ALIGN_REFS (stmt_info)); | |
5232 | set_vinfo_for_stmt (stmt, NULL); | |
5233 | free (stmt_info); | |
5234 | } | |
5235 | ||
5236 | ||
bb67d9c7 | 5237 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 5238 | |
bb67d9c7 | 5239 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
5240 | by the target. */ |
5241 | ||
bb67d9c7 RG |
5242 | static tree |
5243 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a IR |
5244 | { |
5245 | enum machine_mode inner_mode = TYPE_MODE (scalar_type); | |
cc4b5170 | 5246 | enum machine_mode simd_mode; |
2f816591 | 5247 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
5248 | int nunits; |
5249 | tree vectype; | |
5250 | ||
cc4b5170 | 5251 | if (nbytes == 0) |
ebfd146a IR |
5252 | return NULL_TREE; |
5253 | ||
2f816591 RG |
5254 | /* We can't build a vector type of elements with alignment bigger than |
5255 | their size. */ | |
5256 | if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) | |
5257 | return NULL_TREE; | |
5258 | ||
6d7971b8 RG |
5259 | /* If we'd build a vector type of elements whose mode precision doesn't |
5260 | match their types precision we'll get mismatched types on vector | |
5261 | extracts via BIT_FIELD_REFs. This effectively means we disable | |
5262 | vectorization of bool and/or enum types in some languages. */ | |
5263 | if (INTEGRAL_TYPE_P (scalar_type) | |
5264 | && GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type)) | |
5265 | return NULL_TREE; | |
5266 | ||
cc4b5170 RG |
5267 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
5268 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
5269 | return NULL_TREE; | |
5270 | ||
bb67d9c7 RG |
5271 | /* If no size was supplied use the mode the target prefers. Otherwise |
5272 | lookup a vector mode of the specified size. */ | |
5273 | if (size == 0) | |
5274 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
5275 | else | |
5276 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
5277 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
5278 | if (nunits <= 1) | |
5279 | return NULL_TREE; | |
ebfd146a IR |
5280 | |
5281 | vectype = build_vector_type (scalar_type, nunits); | |
5282 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5283 | { | |
5284 | fprintf (vect_dump, "get vectype with %d units of type ", nunits); | |
5285 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
5286 | } | |
5287 | ||
5288 | if (!vectype) | |
5289 | return NULL_TREE; | |
5290 | ||
5291 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5292 | { | |
5293 | fprintf (vect_dump, "vectype: "); | |
5294 | print_generic_expr (vect_dump, vectype, TDF_SLIM); | |
5295 | } | |
5296 | ||
5297 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
5298 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
5299 | { | |
5300 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5301 | fprintf (vect_dump, "mode not supported by target."); | |
5302 | return NULL_TREE; | |
5303 | } | |
5304 | ||
5305 | return vectype; | |
5306 | } | |
5307 | ||
bb67d9c7 RG |
5308 | unsigned int current_vector_size; |
5309 | ||
5310 | /* Function get_vectype_for_scalar_type. | |
5311 | ||
5312 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
5313 | by the target. */ | |
5314 | ||
5315 | tree | |
5316 | get_vectype_for_scalar_type (tree scalar_type) | |
5317 | { | |
5318 | tree vectype; | |
5319 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
5320 | current_vector_size); | |
5321 | if (vectype | |
5322 | && current_vector_size == 0) | |
5323 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
5324 | return vectype; | |
5325 | } | |
5326 | ||
b690cc0f RG |
5327 | /* Function get_same_sized_vectype |
5328 | ||
5329 | Returns a vector type corresponding to SCALAR_TYPE of size | |
5330 | VECTOR_TYPE if supported by the target. */ | |
5331 | ||
5332 | tree | |
bb67d9c7 | 5333 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 5334 | { |
bb67d9c7 RG |
5335 | return get_vectype_for_scalar_type_and_size |
5336 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
5337 | } |
5338 | ||
ebfd146a IR |
5339 | /* Function vect_is_simple_use. |
5340 | ||
5341 | Input: | |
a70d6342 IR |
5342 | LOOP_VINFO - the vect info of the loop that is being vectorized. |
5343 | BB_VINFO - the vect info of the basic block that is being vectorized. | |
5344 | OPERAND - operand of a stmt in the loop or bb. | |
ebfd146a IR |
5345 | DEF - the defining stmt in case OPERAND is an SSA_NAME. |
5346 | ||
5347 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 5348 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 5349 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 5350 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
5351 | is the case in reduction/induction computations). |
5352 | For basic blocks, supportable operands are constants and bb invariants. | |
5353 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
5354 | |
5355 | bool | |
b8698a0f | 5356 | vect_is_simple_use (tree operand, loop_vec_info loop_vinfo, |
a70d6342 | 5357 | bb_vec_info bb_vinfo, gimple *def_stmt, |
ebfd146a | 5358 | tree *def, enum vect_def_type *dt) |
b8698a0f | 5359 | { |
ebfd146a IR |
5360 | basic_block bb; |
5361 | stmt_vec_info stmt_vinfo; | |
a70d6342 | 5362 | struct loop *loop = NULL; |
b8698a0f | 5363 | |
a70d6342 IR |
5364 | if (loop_vinfo) |
5365 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
5366 | |
5367 | *def_stmt = NULL; | |
5368 | *def = NULL_TREE; | |
b8698a0f | 5369 | |
ebfd146a IR |
5370 | if (vect_print_dump_info (REPORT_DETAILS)) |
5371 | { | |
5372 | fprintf (vect_dump, "vect_is_simple_use: operand "); | |
5373 | print_generic_expr (vect_dump, operand, TDF_SLIM); | |
5374 | } | |
b8698a0f | 5375 | |
ebfd146a IR |
5376 | if (TREE_CODE (operand) == INTEGER_CST || TREE_CODE (operand) == REAL_CST) |
5377 | { | |
5378 | *dt = vect_constant_def; | |
5379 | return true; | |
5380 | } | |
b8698a0f | 5381 | |
ebfd146a IR |
5382 | if (is_gimple_min_invariant (operand)) |
5383 | { | |
5384 | *def = operand; | |
8644a673 | 5385 | *dt = vect_external_def; |
ebfd146a IR |
5386 | return true; |
5387 | } | |
5388 | ||
5389 | if (TREE_CODE (operand) == PAREN_EXPR) | |
5390 | { | |
5391 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5392 | fprintf (vect_dump, "non-associatable copy."); | |
5393 | operand = TREE_OPERAND (operand, 0); | |
5394 | } | |
b8698a0f | 5395 | |
ebfd146a IR |
5396 | if (TREE_CODE (operand) != SSA_NAME) |
5397 | { | |
5398 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5399 | fprintf (vect_dump, "not ssa-name."); | |
5400 | return false; | |
5401 | } | |
b8698a0f | 5402 | |
ebfd146a IR |
5403 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
5404 | if (*def_stmt == NULL) | |
5405 | { | |
5406 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5407 | fprintf (vect_dump, "no def_stmt."); | |
5408 | return false; | |
5409 | } | |
5410 | ||
5411 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5412 | { | |
5413 | fprintf (vect_dump, "def_stmt: "); | |
5414 | print_gimple_stmt (vect_dump, *def_stmt, 0, TDF_SLIM); | |
5415 | } | |
5416 | ||
8644a673 | 5417 | /* Empty stmt is expected only in case of a function argument. |
ebfd146a IR |
5418 | (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */ |
5419 | if (gimple_nop_p (*def_stmt)) | |
5420 | { | |
5421 | *def = operand; | |
8644a673 | 5422 | *dt = vect_external_def; |
ebfd146a IR |
5423 | return true; |
5424 | } | |
5425 | ||
5426 | bb = gimple_bb (*def_stmt); | |
a70d6342 IR |
5427 | |
5428 | if ((loop && !flow_bb_inside_loop_p (loop, bb)) | |
5429 | || (!loop && bb != BB_VINFO_BB (bb_vinfo)) | |
b8698a0f | 5430 | || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI)) |
8644a673 | 5431 | *dt = vect_external_def; |
ebfd146a IR |
5432 | else |
5433 | { | |
5434 | stmt_vinfo = vinfo_for_stmt (*def_stmt); | |
5435 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
5436 | } | |
5437 | ||
5438 | if (*dt == vect_unknown_def_type) | |
5439 | { | |
5440 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5441 | fprintf (vect_dump, "Unsupported pattern."); | |
5442 | return false; | |
5443 | } | |
5444 | ||
5445 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5446 | fprintf (vect_dump, "type of def: %d.",*dt); | |
5447 | ||
5448 | switch (gimple_code (*def_stmt)) | |
5449 | { | |
5450 | case GIMPLE_PHI: | |
5451 | *def = gimple_phi_result (*def_stmt); | |
5452 | break; | |
5453 | ||
5454 | case GIMPLE_ASSIGN: | |
5455 | *def = gimple_assign_lhs (*def_stmt); | |
5456 | break; | |
5457 | ||
5458 | case GIMPLE_CALL: | |
5459 | *def = gimple_call_lhs (*def_stmt); | |
5460 | if (*def != NULL) | |
5461 | break; | |
5462 | /* FALLTHRU */ | |
5463 | default: | |
5464 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5465 | fprintf (vect_dump, "unsupported defining stmt: "); | |
5466 | return false; | |
5467 | } | |
5468 | ||
5469 | return true; | |
5470 | } | |
5471 | ||
b690cc0f RG |
5472 | /* Function vect_is_simple_use_1. |
5473 | ||
5474 | Same as vect_is_simple_use_1 but also determines the vector operand | |
5475 | type of OPERAND and stores it to *VECTYPE. If the definition of | |
5476 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
5477 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
5478 | is responsible to compute the best suited vector type for the | |
5479 | scalar operand. */ | |
5480 | ||
5481 | bool | |
5482 | vect_is_simple_use_1 (tree operand, loop_vec_info loop_vinfo, | |
5483 | bb_vec_info bb_vinfo, gimple *def_stmt, | |
5484 | tree *def, enum vect_def_type *dt, tree *vectype) | |
5485 | { | |
5486 | if (!vect_is_simple_use (operand, loop_vinfo, bb_vinfo, def_stmt, def, dt)) | |
5487 | return false; | |
5488 | ||
5489 | /* Now get a vector type if the def is internal, otherwise supply | |
5490 | NULL_TREE and leave it up to the caller to figure out a proper | |
5491 | type for the use stmt. */ | |
5492 | if (*dt == vect_internal_def | |
5493 | || *dt == vect_induction_def | |
5494 | || *dt == vect_reduction_def | |
5495 | || *dt == vect_double_reduction_def | |
5496 | || *dt == vect_nested_cycle) | |
5497 | { | |
5498 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
5499 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
5500 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
5501 | *vectype = STMT_VINFO_VECTYPE (stmt_info); | |
5502 | gcc_assert (*vectype != NULL_TREE); | |
5503 | } | |
5504 | else if (*dt == vect_uninitialized_def | |
5505 | || *dt == vect_constant_def | |
5506 | || *dt == vect_external_def) | |
5507 | *vectype = NULL_TREE; | |
5508 | else | |
5509 | gcc_unreachable (); | |
5510 | ||
5511 | return true; | |
5512 | } | |
5513 | ||
ebfd146a IR |
5514 | |
5515 | /* Function supportable_widening_operation | |
5516 | ||
b8698a0f L |
5517 | Check whether an operation represented by the code CODE is a |
5518 | widening operation that is supported by the target platform in | |
b690cc0f RG |
5519 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
5520 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 5521 | |
ebfd146a IR |
5522 | Widening operations we currently support are NOP (CONVERT), FLOAT |
5523 | and WIDEN_MULT. This function checks if these operations are supported | |
5524 | by the target platform either directly (via vector tree-codes), or via | |
5525 | target builtins. | |
5526 | ||
5527 | Output: | |
b8698a0f L |
5528 | - CODE1 and CODE2 are codes of vector operations to be used when |
5529 | vectorizing the operation, if available. | |
ebfd146a | 5530 | - DECL1 and DECL2 are decls of target builtin functions to be used |
ff802fa1 | 5531 | when vectorizing the operation, if available. In this case, |
b8698a0f | 5532 | CODE1 and CODE2 are CALL_EXPR. |
ebfd146a IR |
5533 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
5534 | case of multi-step conversion (like char->short->int - in that case | |
5535 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
5536 | - INTERM_TYPES contains the intermediate type required to perform the |
5537 | widening operation (short in the above example). */ | |
ebfd146a IR |
5538 | |
5539 | bool | |
b690cc0f RG |
5540 | supportable_widening_operation (enum tree_code code, gimple stmt, |
5541 | tree vectype_out, tree vectype_in, | |
ebfd146a IR |
5542 | tree *decl1, tree *decl2, |
5543 | enum tree_code *code1, enum tree_code *code2, | |
5544 | int *multi_step_cvt, | |
5545 | VEC (tree, heap) **interm_types) | |
5546 | { | |
5547 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5548 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
5549 | struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info); | |
5550 | bool ordered_p; | |
5551 | enum machine_mode vec_mode; | |
81f40b79 | 5552 | enum insn_code icode1, icode2; |
ebfd146a | 5553 | optab optab1, optab2; |
b690cc0f RG |
5554 | tree vectype = vectype_in; |
5555 | tree wide_vectype = vectype_out; | |
ebfd146a IR |
5556 | enum tree_code c1, c2; |
5557 | ||
5558 | /* The result of a vectorized widening operation usually requires two vectors | |
b8698a0f L |
5559 | (because the widened results do not fit int one vector). The generated |
5560 | vector results would normally be expected to be generated in the same | |
ebfd146a IR |
5561 | order as in the original scalar computation, i.e. if 8 results are |
5562 | generated in each vector iteration, they are to be organized as follows: | |
b8698a0f | 5563 | vect1: [res1,res2,res3,res4], vect2: [res5,res6,res7,res8]. |
ebfd146a | 5564 | |
b8698a0f | 5565 | However, in the special case that the result of the widening operation is |
ebfd146a | 5566 | used in a reduction computation only, the order doesn't matter (because |
b8698a0f | 5567 | when vectorizing a reduction we change the order of the computation). |
ebfd146a IR |
5568 | Some targets can take advantage of this and generate more efficient code. |
5569 | For example, targets like Altivec, that support widen_mult using a sequence | |
5570 | of {mult_even,mult_odd} generate the following vectors: | |
5571 | vect1: [res1,res3,res5,res7], vect2: [res2,res4,res6,res8]. | |
5572 | ||
5573 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
b8698a0f | 5574 | (each vectorized inner-loop iteration contributes to VF outer-loop |
ff802fa1 | 5575 | iterations in parallel). We therefore don't allow to change the order |
ebfd146a IR |
5576 | of the computation in the inner-loop during outer-loop vectorization. */ |
5577 | ||
5578 | if (STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
5579 | && !nested_in_vect_loop_p (vect_loop, stmt)) | |
5580 | ordered_p = false; | |
5581 | else | |
5582 | ordered_p = true; | |
5583 | ||
5584 | if (!ordered_p | |
5585 | && code == WIDEN_MULT_EXPR | |
5586 | && targetm.vectorize.builtin_mul_widen_even | |
5587 | && targetm.vectorize.builtin_mul_widen_even (vectype) | |
5588 | && targetm.vectorize.builtin_mul_widen_odd | |
5589 | && targetm.vectorize.builtin_mul_widen_odd (vectype)) | |
5590 | { | |
5591 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5592 | fprintf (vect_dump, "Unordered widening operation detected."); | |
5593 | ||
5594 | *code1 = *code2 = CALL_EXPR; | |
5595 | *decl1 = targetm.vectorize.builtin_mul_widen_even (vectype); | |
5596 | *decl2 = targetm.vectorize.builtin_mul_widen_odd (vectype); | |
5597 | return true; | |
5598 | } | |
5599 | ||
5600 | switch (code) | |
5601 | { | |
5602 | case WIDEN_MULT_EXPR: | |
5603 | if (BYTES_BIG_ENDIAN) | |
5604 | { | |
5605 | c1 = VEC_WIDEN_MULT_HI_EXPR; | |
5606 | c2 = VEC_WIDEN_MULT_LO_EXPR; | |
5607 | } | |
5608 | else | |
5609 | { | |
5610 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
5611 | c1 = VEC_WIDEN_MULT_LO_EXPR; | |
5612 | } | |
5613 | break; | |
5614 | ||
5615 | CASE_CONVERT: | |
5616 | if (BYTES_BIG_ENDIAN) | |
5617 | { | |
5618 | c1 = VEC_UNPACK_HI_EXPR; | |
5619 | c2 = VEC_UNPACK_LO_EXPR; | |
5620 | } | |
5621 | else | |
5622 | { | |
5623 | c2 = VEC_UNPACK_HI_EXPR; | |
5624 | c1 = VEC_UNPACK_LO_EXPR; | |
5625 | } | |
5626 | break; | |
5627 | ||
5628 | case FLOAT_EXPR: | |
5629 | if (BYTES_BIG_ENDIAN) | |
5630 | { | |
5631 | c1 = VEC_UNPACK_FLOAT_HI_EXPR; | |
5632 | c2 = VEC_UNPACK_FLOAT_LO_EXPR; | |
5633 | } | |
5634 | else | |
5635 | { | |
5636 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
5637 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; | |
5638 | } | |
5639 | break; | |
5640 | ||
5641 | case FIX_TRUNC_EXPR: | |
5642 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
5643 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
5644 | computing the operation. */ | |
5645 | return false; | |
5646 | ||
5647 | default: | |
5648 | gcc_unreachable (); | |
5649 | } | |
5650 | ||
5651 | if (code == FIX_TRUNC_EXPR) | |
5652 | { | |
5653 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
5654 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
5655 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
5656 | } |
5657 | else | |
5658 | { | |
5659 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
5660 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
5661 | } | |
5662 | ||
5663 | if (!optab1 || !optab2) | |
5664 | return false; | |
5665 | ||
5666 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
5667 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
5668 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
5669 | return false; |
5670 | ||
b8698a0f | 5671 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a IR |
5672 | types. */ |
5673 | if (insn_data[icode1].operand[0].mode != TYPE_MODE (wide_vectype) | |
5674 | || insn_data[icode2].operand[0].mode != TYPE_MODE (wide_vectype)) | |
5675 | { | |
5676 | int i; | |
5677 | tree prev_type = vectype, intermediate_type; | |
5678 | enum machine_mode intermediate_mode, prev_mode = vec_mode; | |
5679 | optab optab3, optab4; | |
5680 | ||
5681 | if (!CONVERT_EXPR_CODE_P (code)) | |
5682 | return false; | |
b8698a0f | 5683 | |
ebfd146a IR |
5684 | *code1 = c1; |
5685 | *code2 = c2; | |
b8698a0f | 5686 | |
ebfd146a | 5687 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
ff802fa1 IR |
5688 | intermediate steps in promotion sequence. We try |
5689 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
5690 | not. */ | |
ebfd146a IR |
5691 | *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS); |
5692 | for (i = 0; i < 3; i++) | |
5693 | { | |
5694 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
5695 | intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode, | |
5696 | TYPE_UNSIGNED (prev_type)); | |
5697 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
5698 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
5699 | ||
5700 | if (!optab3 || !optab4 | |
947131ba RS |
5701 | || ((icode1 = optab_handler (optab1, prev_mode)) |
5702 | == CODE_FOR_nothing) | |
ebfd146a | 5703 | || insn_data[icode1].operand[0].mode != intermediate_mode |
947131ba RS |
5704 | || ((icode2 = optab_handler (optab2, prev_mode)) |
5705 | == CODE_FOR_nothing) | |
ebfd146a | 5706 | || insn_data[icode2].operand[0].mode != intermediate_mode |
947131ba RS |
5707 | || ((icode1 = optab_handler (optab3, intermediate_mode)) |
5708 | == CODE_FOR_nothing) | |
5709 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
5710 | == CODE_FOR_nothing)) | |
ebfd146a IR |
5711 | return false; |
5712 | ||
5713 | VEC_quick_push (tree, *interm_types, intermediate_type); | |
5714 | (*multi_step_cvt)++; | |
5715 | ||
5716 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
5717 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5718 | return true; | |
5719 | ||
5720 | prev_type = intermediate_type; | |
5721 | prev_mode = intermediate_mode; | |
5722 | } | |
5723 | ||
5724 | return false; | |
5725 | } | |
5726 | ||
5727 | *code1 = c1; | |
5728 | *code2 = c2; | |
5729 | return true; | |
5730 | } | |
5731 | ||
5732 | ||
5733 | /* Function supportable_narrowing_operation | |
5734 | ||
b8698a0f L |
5735 | Check whether an operation represented by the code CODE is a |
5736 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
5737 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
5738 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 5739 | |
ebfd146a | 5740 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 5741 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
5742 | the target platform directly via vector tree-codes. |
5743 | ||
5744 | Output: | |
b8698a0f L |
5745 | - CODE1 is the code of a vector operation to be used when |
5746 | vectorizing the operation, if available. | |
ebfd146a IR |
5747 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
5748 | case of multi-step conversion (like int->short->char - in that case | |
5749 | MULTI_STEP_CVT will be 1). | |
5750 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 5751 | narrowing operation (short in the above example). */ |
ebfd146a IR |
5752 | |
5753 | bool | |
5754 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 5755 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
5756 | enum tree_code *code1, int *multi_step_cvt, |
5757 | VEC (tree, heap) **interm_types) | |
5758 | { | |
5759 | enum machine_mode vec_mode; | |
5760 | enum insn_code icode1; | |
5761 | optab optab1, interm_optab; | |
b690cc0f RG |
5762 | tree vectype = vectype_in; |
5763 | tree narrow_vectype = vectype_out; | |
ebfd146a IR |
5764 | enum tree_code c1; |
5765 | tree intermediate_type, prev_type; | |
5766 | int i; | |
5767 | ||
5768 | switch (code) | |
5769 | { | |
5770 | CASE_CONVERT: | |
5771 | c1 = VEC_PACK_TRUNC_EXPR; | |
5772 | break; | |
5773 | ||
5774 | case FIX_TRUNC_EXPR: | |
5775 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
5776 | break; | |
5777 | ||
5778 | case FLOAT_EXPR: | |
5779 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
5780 | tree code and optabs used for computing the operation. */ | |
5781 | return false; | |
5782 | ||
5783 | default: | |
5784 | gcc_unreachable (); | |
5785 | } | |
5786 | ||
5787 | if (code == FIX_TRUNC_EXPR) | |
5788 | /* The signedness is determined from output operand. */ | |
b690cc0f | 5789 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
5790 | else |
5791 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
5792 | ||
5793 | if (!optab1) | |
5794 | return false; | |
5795 | ||
5796 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 5797 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
5798 | return false; |
5799 | ||
5800 | /* Check if it's a multi-step conversion that can be done using intermediate | |
5801 | types. */ | |
5802 | if (insn_data[icode1].operand[0].mode != TYPE_MODE (narrow_vectype)) | |
5803 | { | |
5804 | enum machine_mode intermediate_mode, prev_mode = vec_mode; | |
5805 | ||
5806 | *code1 = c1; | |
5807 | prev_type = vectype; | |
5808 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS | |
ff802fa1 IR |
5809 | intermediate steps in promotion sequence. We try |
5810 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
5811 | not. */ | |
ebfd146a IR |
5812 | *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS); |
5813 | for (i = 0; i < 3; i++) | |
5814 | { | |
5815 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
5816 | intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode, | |
5817 | TYPE_UNSIGNED (prev_type)); | |
b8698a0f | 5818 | interm_optab = optab_for_tree_code (c1, intermediate_type, |
ebfd146a | 5819 | optab_default); |
b8698a0f | 5820 | if (!interm_optab |
947131ba RS |
5821 | || ((icode1 = optab_handler (optab1, prev_mode)) |
5822 | == CODE_FOR_nothing) | |
ebfd146a | 5823 | || insn_data[icode1].operand[0].mode != intermediate_mode |
947131ba RS |
5824 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) |
5825 | == CODE_FOR_nothing)) | |
ebfd146a IR |
5826 | return false; |
5827 | ||
5828 | VEC_quick_push (tree, *interm_types, intermediate_type); | |
5829 | (*multi_step_cvt)++; | |
5830 | ||
5831 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5832 | return true; | |
5833 | ||
5834 | prev_type = intermediate_type; | |
5835 | prev_mode = intermediate_mode; | |
5836 | } | |
5837 | ||
5838 | return false; | |
5839 | } | |
5840 | ||
5841 | *code1 = c1; | |
5842 | return true; | |
5843 | } |