]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
4dee9718 | 2 | Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011 |
62f7fd21 | 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, | |
83197f37 IR |
129 | enum vect_relevant relevant, bool live_p, |
130 | bool used_in_pattern) | |
ebfd146a IR |
131 | { |
132 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
133 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
134 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
83197f37 | 135 | gimple pattern_stmt; |
ebfd146a IR |
136 | |
137 | if (vect_print_dump_info (REPORT_DETAILS)) | |
138 | fprintf (vect_dump, "mark relevant %d, live %d.", relevant, live_p); | |
139 | ||
83197f37 IR |
140 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
141 | related pattern stmt instead of the original stmt. However, such stmts | |
142 | may have their own uses that are not in any pattern, in such cases the | |
143 | stmt itself should be marked. */ | |
ebfd146a IR |
144 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
145 | { | |
83197f37 IR |
146 | bool found = false; |
147 | if (!used_in_pattern) | |
148 | { | |
149 | imm_use_iterator imm_iter; | |
150 | use_operand_p use_p; | |
151 | gimple use_stmt; | |
152 | tree lhs; | |
ebfd146a | 153 | |
83197f37 IR |
154 | if (is_gimple_assign (stmt)) |
155 | lhs = gimple_assign_lhs (stmt); | |
156 | else | |
157 | lhs = gimple_call_lhs (stmt); | |
ebfd146a | 158 | |
83197f37 IR |
159 | /* This use is out of pattern use, if LHS has other uses that are |
160 | pattern uses, we should mark the stmt itself, and not the pattern | |
161 | stmt. */ | |
ab0ef706 JJ |
162 | if (TREE_CODE (lhs) == SSA_NAME) |
163 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs) | |
164 | { | |
165 | if (is_gimple_debug (USE_STMT (use_p))) | |
166 | continue; | |
167 | use_stmt = USE_STMT (use_p); | |
168 | ||
169 | if (vinfo_for_stmt (use_stmt) | |
170 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt))) | |
171 | { | |
172 | found = true; | |
173 | break; | |
174 | } | |
175 | } | |
83197f37 IR |
176 | } |
177 | ||
178 | if (!found) | |
179 | { | |
180 | /* This is the last stmt in a sequence that was detected as a | |
181 | pattern that can potentially be vectorized. Don't mark the stmt | |
182 | as relevant/live because it's not going to be vectorized. | |
183 | Instead mark the pattern-stmt that replaces it. */ | |
184 | ||
185 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
186 | ||
187 | if (vect_print_dump_info (REPORT_DETAILS)) | |
188 | fprintf (vect_dump, "last stmt in pattern. don't mark" | |
189 | " relevant/live."); | |
190 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
191 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
192 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
193 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
194 | stmt = pattern_stmt; | |
195 | } | |
ebfd146a IR |
196 | } |
197 | ||
198 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
199 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
200 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
201 | ||
202 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
203 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
204 | { | |
205 | if (vect_print_dump_info (REPORT_DETAILS)) | |
206 | fprintf (vect_dump, "already marked relevant/live."); | |
207 | return; | |
208 | } | |
209 | ||
210 | VEC_safe_push (gimple, heap, *worklist, stmt); | |
211 | } | |
212 | ||
213 | ||
214 | /* Function vect_stmt_relevant_p. | |
215 | ||
216 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
217 | "relevant for vectorization". | |
218 | ||
219 | A stmt is considered "relevant for vectorization" if: | |
220 | - it has uses outside the loop. | |
221 | - it has vdefs (it alters memory). | |
222 | - control stmts in the loop (except for the exit condition). | |
223 | ||
224 | CHECKME: what other side effects would the vectorizer allow? */ | |
225 | ||
226 | static bool | |
227 | vect_stmt_relevant_p (gimple stmt, loop_vec_info loop_vinfo, | |
228 | enum vect_relevant *relevant, bool *live_p) | |
229 | { | |
230 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
231 | ssa_op_iter op_iter; | |
232 | imm_use_iterator imm_iter; | |
233 | use_operand_p use_p; | |
234 | def_operand_p def_p; | |
235 | ||
8644a673 | 236 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
237 | *live_p = false; |
238 | ||
239 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
240 | if (is_ctrl_stmt (stmt) |
241 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
242 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 243 | *relevant = vect_used_in_scope; |
ebfd146a IR |
244 | |
245 | /* changing memory. */ | |
246 | if (gimple_code (stmt) != GIMPLE_PHI) | |
5006671f | 247 | if (gimple_vdef (stmt)) |
ebfd146a IR |
248 | { |
249 | if (vect_print_dump_info (REPORT_DETAILS)) | |
250 | fprintf (vect_dump, "vec_stmt_relevant_p: stmt has vdefs."); | |
8644a673 | 251 | *relevant = vect_used_in_scope; |
ebfd146a IR |
252 | } |
253 | ||
254 | /* uses outside the loop. */ | |
255 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
256 | { | |
257 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
258 | { | |
259 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
260 | if (!flow_bb_inside_loop_p (loop, bb)) | |
261 | { | |
262 | if (vect_print_dump_info (REPORT_DETAILS)) | |
263 | fprintf (vect_dump, "vec_stmt_relevant_p: used out of loop."); | |
264 | ||
3157b0c2 AO |
265 | if (is_gimple_debug (USE_STMT (use_p))) |
266 | continue; | |
267 | ||
ebfd146a IR |
268 | /* We expect all such uses to be in the loop exit phis |
269 | (because of loop closed form) */ | |
270 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
271 | gcc_assert (bb == single_exit (loop)->dest); | |
272 | ||
273 | *live_p = true; | |
274 | } | |
275 | } | |
276 | } | |
277 | ||
278 | return (*live_p || *relevant); | |
279 | } | |
280 | ||
281 | ||
b8698a0f | 282 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 283 | |
ff802fa1 | 284 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
285 | used in STMT for anything other than indexing an array. */ |
286 | ||
287 | static bool | |
288 | exist_non_indexing_operands_for_use_p (tree use, gimple stmt) | |
289 | { | |
290 | tree operand; | |
291 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 292 | |
ff802fa1 | 293 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
294 | reference in STMT, then any operand that corresponds to USE |
295 | is not indexing an array. */ | |
296 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
297 | return true; | |
59a05b0c | 298 | |
ebfd146a IR |
299 | /* STMT has a data_ref. FORNOW this means that its of one of |
300 | the following forms: | |
301 | -1- ARRAY_REF = var | |
302 | -2- var = ARRAY_REF | |
303 | (This should have been verified in analyze_data_refs). | |
304 | ||
305 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 306 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
307 | for array indexing. |
308 | ||
309 | Therefore, all we need to check is if STMT falls into the | |
310 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
311 | |
312 | if (!gimple_assign_copy_p (stmt)) | |
313 | return false; | |
59a05b0c EB |
314 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
315 | return false; | |
ebfd146a | 316 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
317 | if (TREE_CODE (operand) != SSA_NAME) |
318 | return false; | |
319 | ||
320 | if (operand == use) | |
321 | return true; | |
322 | ||
323 | return false; | |
324 | } | |
325 | ||
326 | ||
b8698a0f | 327 | /* |
ebfd146a IR |
328 | Function process_use. |
329 | ||
330 | Inputs: | |
331 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 332 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ff802fa1 | 333 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 334 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
335 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
336 | be performed. | |
ebfd146a IR |
337 | |
338 | Outputs: | |
339 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
340 | relevance info of the DEF_STMT of this USE: | |
341 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
342 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
343 | Exceptions: | |
344 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 345 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 346 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
347 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
348 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
349 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
350 | be modified accordingly. | |
351 | ||
352 | Return true if everything is as expected. Return false otherwise. */ | |
353 | ||
354 | static bool | |
b8698a0f | 355 | process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
aec7ae7d JJ |
356 | enum vect_relevant relevant, VEC(gimple,heap) **worklist, |
357 | bool force) | |
ebfd146a IR |
358 | { |
359 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
360 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
361 | stmt_vec_info dstmt_vinfo; | |
362 | basic_block bb, def_bb; | |
363 | tree def; | |
364 | gimple def_stmt; | |
365 | enum vect_def_type dt; | |
366 | ||
b8698a0f | 367 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 368 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 369 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
370 | return true; |
371 | ||
a70d6342 | 372 | if (!vect_is_simple_use (use, loop_vinfo, NULL, &def_stmt, &def, &dt)) |
b8698a0f | 373 | { |
8644a673 | 374 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
ebfd146a IR |
375 | fprintf (vect_dump, "not vectorized: unsupported use in stmt."); |
376 | return false; | |
377 | } | |
378 | ||
379 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
380 | return true; | |
381 | ||
382 | def_bb = gimple_bb (def_stmt); | |
383 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
384 | { | |
385 | if (vect_print_dump_info (REPORT_DETAILS)) | |
386 | fprintf (vect_dump, "def_stmt is out of loop."); | |
387 | return true; | |
388 | } | |
389 | ||
b8698a0f L |
390 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
391 | DEF_STMT must have already been processed, because this should be the | |
392 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
393 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
394 | check that everything is as expected, and we are done. */ |
395 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
396 | bb = gimple_bb (stmt); | |
397 | if (gimple_code (stmt) == GIMPLE_PHI | |
398 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
399 | && gimple_code (def_stmt) != GIMPLE_PHI | |
400 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
401 | && bb->loop_father == def_bb->loop_father) | |
402 | { | |
403 | if (vect_print_dump_info (REPORT_DETAILS)) | |
404 | fprintf (vect_dump, "reduc-stmt defining reduc-phi in the same nest."); | |
405 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) | |
406 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
407 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 408 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 409 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
410 | return true; |
411 | } | |
412 | ||
413 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
414 | outer-loop-header-bb: | |
415 | d = def_stmt | |
416 | inner-loop: | |
417 | stmt # use (d) | |
418 | outer-loop-tail-bb: | |
419 | ... */ | |
420 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
421 | { | |
422 | if (vect_print_dump_info (REPORT_DETAILS)) | |
423 | fprintf (vect_dump, "outer-loop def-stmt defining inner-loop stmt."); | |
7c5222ff | 424 | |
ebfd146a IR |
425 | switch (relevant) |
426 | { | |
8644a673 | 427 | case vect_unused_in_scope: |
7c5222ff IR |
428 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
429 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 430 | break; |
7c5222ff | 431 | |
ebfd146a | 432 | case vect_used_in_outer_by_reduction: |
7c5222ff | 433 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
434 | relevant = vect_used_by_reduction; |
435 | break; | |
7c5222ff | 436 | |
ebfd146a | 437 | case vect_used_in_outer: |
7c5222ff | 438 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 439 | relevant = vect_used_in_scope; |
ebfd146a | 440 | break; |
7c5222ff | 441 | |
8644a673 | 442 | case vect_used_in_scope: |
ebfd146a IR |
443 | break; |
444 | ||
445 | default: | |
446 | gcc_unreachable (); | |
b8698a0f | 447 | } |
ebfd146a IR |
448 | } |
449 | ||
450 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
451 | outer-loop-header-bb: | |
452 | ... | |
453 | inner-loop: | |
454 | d = def_stmt | |
06066f92 | 455 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
456 | stmt # use (d) */ |
457 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
458 | { | |
459 | if (vect_print_dump_info (REPORT_DETAILS)) | |
460 | fprintf (vect_dump, "inner-loop def-stmt defining outer-loop stmt."); | |
7c5222ff | 461 | |
ebfd146a IR |
462 | switch (relevant) |
463 | { | |
8644a673 | 464 | case vect_unused_in_scope: |
b8698a0f | 465 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 466 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 467 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
468 | break; |
469 | ||
ebfd146a IR |
470 | case vect_used_by_reduction: |
471 | relevant = vect_used_in_outer_by_reduction; | |
472 | break; | |
473 | ||
8644a673 | 474 | case vect_used_in_scope: |
ebfd146a IR |
475 | relevant = vect_used_in_outer; |
476 | break; | |
477 | ||
478 | default: | |
479 | gcc_unreachable (); | |
480 | } | |
481 | } | |
482 | ||
83197f37 IR |
483 | vect_mark_relevant (worklist, def_stmt, relevant, live_p, |
484 | is_pattern_stmt_p (stmt_vinfo)); | |
ebfd146a IR |
485 | return true; |
486 | } | |
487 | ||
488 | ||
489 | /* Function vect_mark_stmts_to_be_vectorized. | |
490 | ||
491 | Not all stmts in the loop need to be vectorized. For example: | |
492 | ||
493 | for i... | |
494 | for j... | |
495 | 1. T0 = i + j | |
496 | 2. T1 = a[T0] | |
497 | ||
498 | 3. j = j + 1 | |
499 | ||
500 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
501 | addressing of vectorized data-refs are handled differently. | |
502 | ||
503 | This pass detects such stmts. */ | |
504 | ||
505 | bool | |
506 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
507 | { | |
508 | VEC(gimple,heap) *worklist; | |
509 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
510 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
511 | unsigned int nbbs = loop->num_nodes; | |
512 | gimple_stmt_iterator si; | |
513 | gimple stmt; | |
514 | unsigned int i; | |
515 | stmt_vec_info stmt_vinfo; | |
516 | basic_block bb; | |
517 | gimple phi; | |
518 | bool live_p; | |
06066f92 IR |
519 | enum vect_relevant relevant, tmp_relevant; |
520 | enum vect_def_type def_type; | |
ebfd146a IR |
521 | |
522 | if (vect_print_dump_info (REPORT_DETAILS)) | |
523 | fprintf (vect_dump, "=== vect_mark_stmts_to_be_vectorized ==="); | |
524 | ||
525 | worklist = VEC_alloc (gimple, heap, 64); | |
526 | ||
527 | /* 1. Init worklist. */ | |
528 | for (i = 0; i < nbbs; i++) | |
529 | { | |
530 | bb = bbs[i]; | |
531 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 532 | { |
ebfd146a IR |
533 | phi = gsi_stmt (si); |
534 | if (vect_print_dump_info (REPORT_DETAILS)) | |
535 | { | |
536 | fprintf (vect_dump, "init: phi relevant? "); | |
537 | print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM); | |
538 | } | |
539 | ||
540 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 541 | vect_mark_relevant (&worklist, phi, relevant, live_p, false); |
ebfd146a IR |
542 | } |
543 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
544 | { | |
545 | stmt = gsi_stmt (si); | |
546 | if (vect_print_dump_info (REPORT_DETAILS)) | |
547 | { | |
548 | fprintf (vect_dump, "init: stmt relevant? "); | |
549 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
b8698a0f | 550 | } |
ebfd146a IR |
551 | |
552 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 553 | vect_mark_relevant (&worklist, stmt, relevant, live_p, false); |
ebfd146a IR |
554 | } |
555 | } | |
556 | ||
557 | /* 2. Process_worklist */ | |
558 | while (VEC_length (gimple, worklist) > 0) | |
559 | { | |
560 | use_operand_p use_p; | |
561 | ssa_op_iter iter; | |
562 | ||
563 | stmt = VEC_pop (gimple, worklist); | |
564 | if (vect_print_dump_info (REPORT_DETAILS)) | |
565 | { | |
566 | fprintf (vect_dump, "worklist: examine stmt: "); | |
567 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
568 | } | |
569 | ||
b8698a0f L |
570 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
571 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
572 | liveness and relevance properties of STMT. */ |
573 | stmt_vinfo = vinfo_for_stmt (stmt); | |
574 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
575 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
576 | ||
577 | /* Generally, the liveness and relevance properties of STMT are | |
578 | propagated as is to the DEF_STMTs of its USEs: | |
579 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
580 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
581 | ||
582 | One exception is when STMT has been identified as defining a reduction | |
583 | variable; in this case we set the liveness/relevance as follows: | |
584 | live_p = false | |
585 | relevant = vect_used_by_reduction | |
586 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f | 587 | those that are used by a reduction computation, and those that are |
ff802fa1 | 588 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 589 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 590 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 591 | |
06066f92 IR |
592 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
593 | tmp_relevant = relevant; | |
594 | switch (def_type) | |
ebfd146a | 595 | { |
06066f92 IR |
596 | case vect_reduction_def: |
597 | switch (tmp_relevant) | |
598 | { | |
599 | case vect_unused_in_scope: | |
600 | relevant = vect_used_by_reduction; | |
601 | break; | |
602 | ||
603 | case vect_used_by_reduction: | |
604 | if (gimple_code (stmt) == GIMPLE_PHI) | |
605 | break; | |
606 | /* fall through */ | |
607 | ||
608 | default: | |
609 | if (vect_print_dump_info (REPORT_DETAILS)) | |
610 | fprintf (vect_dump, "unsupported use of reduction."); | |
611 | ||
612 | VEC_free (gimple, heap, worklist); | |
613 | return false; | |
614 | } | |
615 | ||
b8698a0f | 616 | live_p = false; |
06066f92 | 617 | break; |
b8698a0f | 618 | |
06066f92 IR |
619 | case vect_nested_cycle: |
620 | if (tmp_relevant != vect_unused_in_scope | |
621 | && tmp_relevant != vect_used_in_outer_by_reduction | |
622 | && tmp_relevant != vect_used_in_outer) | |
623 | { | |
624 | if (vect_print_dump_info (REPORT_DETAILS)) | |
625 | fprintf (vect_dump, "unsupported use of nested cycle."); | |
7c5222ff | 626 | |
06066f92 IR |
627 | VEC_free (gimple, heap, worklist); |
628 | return false; | |
629 | } | |
7c5222ff | 630 | |
b8698a0f L |
631 | live_p = false; |
632 | break; | |
633 | ||
06066f92 IR |
634 | case vect_double_reduction_def: |
635 | if (tmp_relevant != vect_unused_in_scope | |
636 | && tmp_relevant != vect_used_by_reduction) | |
637 | { | |
7c5222ff | 638 | if (vect_print_dump_info (REPORT_DETAILS)) |
06066f92 | 639 | fprintf (vect_dump, "unsupported use of double reduction."); |
7c5222ff IR |
640 | |
641 | VEC_free (gimple, heap, worklist); | |
642 | return false; | |
06066f92 IR |
643 | } |
644 | ||
645 | live_p = false; | |
b8698a0f | 646 | break; |
7c5222ff | 647 | |
06066f92 IR |
648 | default: |
649 | break; | |
7c5222ff | 650 | } |
b8698a0f | 651 | |
aec7ae7d | 652 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
653 | { |
654 | /* Pattern statements are not inserted into the code, so | |
655 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
656 | have to scan the RHS or function arguments instead. */ | |
657 | if (is_gimple_assign (stmt)) | |
658 | { | |
69d2aade JJ |
659 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
660 | tree op = gimple_assign_rhs1 (stmt); | |
661 | ||
662 | i = 1; | |
663 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
664 | { | |
665 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
aec7ae7d | 666 | live_p, relevant, &worklist, false) |
69d2aade | 667 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
aec7ae7d | 668 | live_p, relevant, &worklist, false)) |
69d2aade JJ |
669 | { |
670 | VEC_free (gimple, heap, worklist); | |
671 | return false; | |
672 | } | |
673 | i = 2; | |
674 | } | |
675 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 676 | { |
69d2aade | 677 | op = gimple_op (stmt, i); |
9d5e7640 | 678 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, |
aec7ae7d | 679 | &worklist, false)) |
9d5e7640 IR |
680 | { |
681 | VEC_free (gimple, heap, worklist); | |
682 | return false; | |
683 | } | |
684 | } | |
685 | } | |
686 | else if (is_gimple_call (stmt)) | |
687 | { | |
688 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
689 | { | |
690 | tree arg = gimple_call_arg (stmt, i); | |
691 | if (!process_use (stmt, arg, loop_vinfo, live_p, relevant, | |
aec7ae7d | 692 | &worklist, false)) |
9d5e7640 IR |
693 | { |
694 | VEC_free (gimple, heap, worklist); | |
695 | return false; | |
696 | } | |
697 | } | |
698 | } | |
699 | } | |
700 | else | |
701 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
702 | { | |
703 | tree op = USE_FROM_PTR (use_p); | |
704 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, | |
aec7ae7d | 705 | &worklist, false)) |
9d5e7640 IR |
706 | { |
707 | VEC_free (gimple, heap, worklist); | |
708 | return false; | |
709 | } | |
710 | } | |
aec7ae7d JJ |
711 | |
712 | if (STMT_VINFO_GATHER_P (stmt_vinfo)) | |
713 | { | |
714 | tree off; | |
715 | tree decl = vect_check_gather (stmt, loop_vinfo, NULL, &off, NULL); | |
716 | gcc_assert (decl); | |
717 | if (!process_use (stmt, off, loop_vinfo, live_p, relevant, | |
718 | &worklist, true)) | |
719 | { | |
720 | VEC_free (gimple, heap, worklist); | |
721 | return false; | |
722 | } | |
723 | } | |
ebfd146a IR |
724 | } /* while worklist */ |
725 | ||
726 | VEC_free (gimple, heap, worklist); | |
727 | return true; | |
728 | } | |
729 | ||
730 | ||
720f5239 IR |
731 | /* Get cost by calling cost target builtin. */ |
732 | ||
733 | static inline | |
734 | int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost) | |
735 | { | |
736 | tree dummy_type = NULL; | |
737 | int dummy = 0; | |
738 | ||
739 | return targetm.vectorize.builtin_vectorization_cost (type_of_cost, | |
740 | dummy_type, dummy); | |
741 | } | |
742 | ||
ff802fa1 IR |
743 | |
744 | /* Get cost for STMT. */ | |
745 | ||
ebfd146a IR |
746 | int |
747 | cost_for_stmt (gimple stmt) | |
748 | { | |
749 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
750 | ||
751 | switch (STMT_VINFO_TYPE (stmt_info)) | |
752 | { | |
753 | case load_vec_info_type: | |
720f5239 | 754 | return vect_get_stmt_cost (scalar_load); |
ebfd146a | 755 | case store_vec_info_type: |
720f5239 | 756 | return vect_get_stmt_cost (scalar_store); |
ebfd146a IR |
757 | case op_vec_info_type: |
758 | case condition_vec_info_type: | |
759 | case assignment_vec_info_type: | |
760 | case reduc_vec_info_type: | |
761 | case induc_vec_info_type: | |
762 | case type_promotion_vec_info_type: | |
763 | case type_demotion_vec_info_type: | |
764 | case type_conversion_vec_info_type: | |
765 | case call_vec_info_type: | |
720f5239 | 766 | return vect_get_stmt_cost (scalar_stmt); |
ebfd146a IR |
767 | case undef_vec_info_type: |
768 | default: | |
769 | gcc_unreachable (); | |
770 | } | |
771 | } | |
772 | ||
b8698a0f | 773 | /* Function vect_model_simple_cost. |
ebfd146a | 774 | |
b8698a0f | 775 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
776 | single op. Right now, this does not account for multiple insns that could |
777 | be generated for the single vector op. We will handle that shortly. */ | |
778 | ||
779 | void | |
b8698a0f | 780 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
ebfd146a IR |
781 | enum vect_def_type *dt, slp_tree slp_node) |
782 | { | |
783 | int i; | |
784 | int inside_cost = 0, outside_cost = 0; | |
785 | ||
786 | /* The SLP costs were already calculated during SLP tree build. */ | |
787 | if (PURE_SLP_STMT (stmt_info)) | |
788 | return; | |
789 | ||
720f5239 | 790 | inside_cost = ncopies * vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
791 | |
792 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
793 | for (i = 0; i < 2; i++) | |
794 | { | |
8644a673 | 795 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
720f5239 | 796 | outside_cost += vect_get_stmt_cost (vector_stmt); |
ebfd146a | 797 | } |
b8698a0f | 798 | |
ebfd146a IR |
799 | if (vect_print_dump_info (REPORT_COST)) |
800 | fprintf (vect_dump, "vect_model_simple_cost: inside_cost = %d, " | |
801 | "outside_cost = %d .", inside_cost, outside_cost); | |
802 | ||
803 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
804 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
805 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
806 | } | |
807 | ||
808 | ||
b8698a0f L |
809 | /* Function vect_cost_strided_group_size |
810 | ||
ebfd146a IR |
811 | For strided load or store, return the group_size only if it is the first |
812 | load or store of a group, else return 1. This ensures that group size is | |
813 | only returned once per group. */ | |
814 | ||
815 | static int | |
816 | vect_cost_strided_group_size (stmt_vec_info stmt_info) | |
817 | { | |
e14c1050 | 818 | gimple first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
819 | |
820 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
e14c1050 | 821 | return GROUP_SIZE (stmt_info); |
ebfd146a IR |
822 | |
823 | return 1; | |
824 | } | |
825 | ||
826 | ||
827 | /* Function vect_model_store_cost | |
828 | ||
829 | Models cost for stores. In the case of strided accesses, one access | |
830 | has the overhead of the strided access attributed to it. */ | |
831 | ||
832 | void | |
b8698a0f | 833 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
272c6793 RS |
834 | bool store_lanes_p, enum vect_def_type dt, |
835 | slp_tree slp_node) | |
ebfd146a IR |
836 | { |
837 | int group_size; | |
720f5239 IR |
838 | unsigned int inside_cost = 0, outside_cost = 0; |
839 | struct data_reference *first_dr; | |
840 | gimple first_stmt; | |
ebfd146a IR |
841 | |
842 | /* The SLP costs were already calculated during SLP tree build. */ | |
843 | if (PURE_SLP_STMT (stmt_info)) | |
844 | return; | |
845 | ||
8644a673 | 846 | if (dt == vect_constant_def || dt == vect_external_def) |
720f5239 | 847 | outside_cost = vect_get_stmt_cost (scalar_to_vec); |
ebfd146a IR |
848 | |
849 | /* Strided access? */ | |
e14c1050 | 850 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) |
720f5239 IR |
851 | { |
852 | if (slp_node) | |
853 | { | |
854 | first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0); | |
855 | group_size = 1; | |
856 | } | |
857 | else | |
858 | { | |
e14c1050 | 859 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
720f5239 IR |
860 | group_size = vect_cost_strided_group_size (stmt_info); |
861 | } | |
862 | ||
863 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
864 | } | |
ebfd146a IR |
865 | /* Not a strided access. */ |
866 | else | |
720f5239 IR |
867 | { |
868 | group_size = 1; | |
869 | first_dr = STMT_VINFO_DATA_REF (stmt_info); | |
870 | } | |
ebfd146a | 871 | |
272c6793 RS |
872 | /* We assume that the cost of a single store-lanes instruction is |
873 | equivalent to the cost of GROUP_SIZE separate stores. If a strided | |
874 | access is instead being provided by a permute-and-store operation, | |
875 | include the cost of the permutes. */ | |
876 | if (!store_lanes_p && group_size > 1) | |
ebfd146a IR |
877 | { |
878 | /* Uses a high and low interleave operation for each needed permute. */ | |
b8698a0f | 879 | inside_cost = ncopies * exact_log2(group_size) * group_size |
720f5239 | 880 | * vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
881 | |
882 | if (vect_print_dump_info (REPORT_COST)) | |
883 | fprintf (vect_dump, "vect_model_store_cost: strided group_size = %d .", | |
884 | group_size); | |
885 | ||
886 | } | |
887 | ||
888 | /* Costs of the stores. */ | |
720f5239 | 889 | vect_get_store_cost (first_dr, ncopies, &inside_cost); |
ebfd146a IR |
890 | |
891 | if (vect_print_dump_info (REPORT_COST)) | |
892 | fprintf (vect_dump, "vect_model_store_cost: inside_cost = %d, " | |
893 | "outside_cost = %d .", inside_cost, outside_cost); | |
894 | ||
895 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
896 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
897 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
898 | } | |
899 | ||
900 | ||
720f5239 IR |
901 | /* Calculate cost of DR's memory access. */ |
902 | void | |
903 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
904 | unsigned int *inside_cost) | |
905 | { | |
906 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
907 | ||
908 | switch (alignment_support_scheme) | |
909 | { | |
910 | case dr_aligned: | |
911 | { | |
912 | *inside_cost += ncopies * vect_get_stmt_cost (vector_store); | |
913 | ||
914 | if (vect_print_dump_info (REPORT_COST)) | |
915 | fprintf (vect_dump, "vect_model_store_cost: aligned."); | |
916 | ||
917 | break; | |
918 | } | |
919 | ||
920 | case dr_unaligned_supported: | |
921 | { | |
922 | gimple stmt = DR_STMT (dr); | |
923 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
924 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
925 | ||
926 | /* Here, we assign an additional cost for the unaligned store. */ | |
927 | *inside_cost += ncopies | |
928 | * targetm.vectorize.builtin_vectorization_cost (unaligned_store, | |
929 | vectype, DR_MISALIGNMENT (dr)); | |
930 | ||
931 | if (vect_print_dump_info (REPORT_COST)) | |
932 | fprintf (vect_dump, "vect_model_store_cost: unaligned supported by " | |
933 | "hardware."); | |
934 | ||
935 | break; | |
936 | } | |
937 | ||
938 | default: | |
939 | gcc_unreachable (); | |
940 | } | |
941 | } | |
942 | ||
943 | ||
ebfd146a IR |
944 | /* Function vect_model_load_cost |
945 | ||
946 | Models cost for loads. In the case of strided accesses, the last access | |
947 | has the overhead of the strided access attributed to it. Since unaligned | |
b8698a0f | 948 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
949 | access scheme chosen. */ |
950 | ||
951 | void | |
272c6793 RS |
952 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, bool load_lanes_p, |
953 | slp_tree slp_node) | |
ebfd146a IR |
954 | { |
955 | int group_size; | |
ebfd146a IR |
956 | gimple first_stmt; |
957 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; | |
720f5239 | 958 | unsigned int inside_cost = 0, outside_cost = 0; |
ebfd146a IR |
959 | |
960 | /* The SLP costs were already calculated during SLP tree build. */ | |
961 | if (PURE_SLP_STMT (stmt_info)) | |
962 | return; | |
963 | ||
964 | /* Strided accesses? */ | |
e14c1050 IR |
965 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
966 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) && first_stmt && !slp_node) | |
ebfd146a IR |
967 | { |
968 | group_size = vect_cost_strided_group_size (stmt_info); | |
969 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
970 | } | |
971 | /* Not a strided access. */ | |
972 | else | |
973 | { | |
974 | group_size = 1; | |
975 | first_dr = dr; | |
976 | } | |
977 | ||
272c6793 RS |
978 | /* We assume that the cost of a single load-lanes instruction is |
979 | equivalent to the cost of GROUP_SIZE separate loads. If a strided | |
980 | access is instead being provided by a load-and-permute operation, | |
981 | include the cost of the permutes. */ | |
982 | if (!load_lanes_p && group_size > 1) | |
ebfd146a IR |
983 | { |
984 | /* Uses an even and odd extract operations for each needed permute. */ | |
985 | inside_cost = ncopies * exact_log2(group_size) * group_size | |
720f5239 | 986 | * vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
987 | |
988 | if (vect_print_dump_info (REPORT_COST)) | |
989 | fprintf (vect_dump, "vect_model_load_cost: strided group_size = %d .", | |
990 | group_size); | |
ebfd146a IR |
991 | } |
992 | ||
993 | /* The loads themselves. */ | |
720f5239 | 994 | vect_get_load_cost (first_dr, ncopies, |
e14c1050 IR |
995 | ((!STMT_VINFO_STRIDED_ACCESS (stmt_info)) || group_size > 1 |
996 | || slp_node), | |
720f5239 IR |
997 | &inside_cost, &outside_cost); |
998 | ||
999 | if (vect_print_dump_info (REPORT_COST)) | |
1000 | fprintf (vect_dump, "vect_model_load_cost: inside_cost = %d, " | |
1001 | "outside_cost = %d .", inside_cost, outside_cost); | |
1002 | ||
1003 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
1004 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
1005 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
1006 | } | |
1007 | ||
1008 | ||
1009 | /* Calculate cost of DR's memory access. */ | |
1010 | void | |
1011 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
1012 | bool add_realign_cost, unsigned int *inside_cost, | |
1013 | unsigned int *outside_cost) | |
1014 | { | |
1015 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
1016 | ||
1017 | switch (alignment_support_scheme) | |
ebfd146a IR |
1018 | { |
1019 | case dr_aligned: | |
1020 | { | |
9940b13c | 1021 | *inside_cost += ncopies * vect_get_stmt_cost (vector_load); |
ebfd146a IR |
1022 | |
1023 | if (vect_print_dump_info (REPORT_COST)) | |
1024 | fprintf (vect_dump, "vect_model_load_cost: aligned."); | |
1025 | ||
1026 | break; | |
1027 | } | |
1028 | case dr_unaligned_supported: | |
1029 | { | |
720f5239 IR |
1030 | gimple stmt = DR_STMT (dr); |
1031 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1032 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
ebfd146a | 1033 | |
720f5239 IR |
1034 | /* Here, we assign an additional cost for the unaligned load. */ |
1035 | *inside_cost += ncopies | |
1036 | * targetm.vectorize.builtin_vectorization_cost (unaligned_load, | |
1037 | vectype, DR_MISALIGNMENT (dr)); | |
ebfd146a IR |
1038 | if (vect_print_dump_info (REPORT_COST)) |
1039 | fprintf (vect_dump, "vect_model_load_cost: unaligned supported by " | |
1040 | "hardware."); | |
1041 | ||
1042 | break; | |
1043 | } | |
1044 | case dr_explicit_realign: | |
1045 | { | |
720f5239 IR |
1046 | *inside_cost += ncopies * (2 * vect_get_stmt_cost (vector_load) |
1047 | + vect_get_stmt_cost (vector_stmt)); | |
ebfd146a IR |
1048 | |
1049 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1050 | the containing loop, the following cost should be added to the | |
1051 | outside costs. */ | |
1052 | if (targetm.vectorize.builtin_mask_for_load) | |
720f5239 | 1053 | *inside_cost += vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
1054 | |
1055 | break; | |
1056 | } | |
1057 | case dr_explicit_realign_optimized: | |
1058 | { | |
1059 | if (vect_print_dump_info (REPORT_COST)) | |
1060 | fprintf (vect_dump, "vect_model_load_cost: unaligned software " | |
1061 | "pipelined."); | |
1062 | ||
1063 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1064 | load, and possibly a mask operation to "prime" the loop. However, |
ebfd146a IR |
1065 | if this is an access in a group of loads, which provide strided |
1066 | access, then the above cost should only be considered for one | |
ff802fa1 | 1067 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1068 | and a realignment op. */ |
1069 | ||
720f5239 | 1070 | if (add_realign_cost) |
ebfd146a | 1071 | { |
720f5239 | 1072 | *outside_cost = 2 * vect_get_stmt_cost (vector_stmt); |
ebfd146a | 1073 | if (targetm.vectorize.builtin_mask_for_load) |
720f5239 | 1074 | *outside_cost += vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
1075 | } |
1076 | ||
720f5239 IR |
1077 | *inside_cost += ncopies * (vect_get_stmt_cost (vector_load) |
1078 | + vect_get_stmt_cost (vector_stmt)); | |
ebfd146a IR |
1079 | break; |
1080 | } | |
1081 | ||
1082 | default: | |
1083 | gcc_unreachable (); | |
1084 | } | |
ebfd146a IR |
1085 | } |
1086 | ||
1087 | ||
1088 | /* Function vect_init_vector. | |
1089 | ||
1090 | Insert a new stmt (INIT_STMT) that initializes a new vector variable with | |
ff802fa1 IR |
1091 | the vector elements of VECTOR_VAR. Place the initialization at BSI if it |
1092 | is not NULL. Otherwise, place the initialization at the loop preheader. | |
b8698a0f | 1093 | Return the DEF of INIT_STMT. |
ebfd146a IR |
1094 | It will be used in the vectorization of STMT. */ |
1095 | ||
1096 | tree | |
1097 | vect_init_vector (gimple stmt, tree vector_var, tree vector_type, | |
1098 | gimple_stmt_iterator *gsi) | |
1099 | { | |
1100 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
1101 | tree new_var; | |
1102 | gimple init_stmt; | |
1103 | tree vec_oprnd; | |
1104 | edge pe; | |
1105 | tree new_temp; | |
1106 | basic_block new_bb; | |
b8698a0f | 1107 | |
ebfd146a | 1108 | new_var = vect_get_new_vect_var (vector_type, vect_simple_var, "cst_"); |
b8698a0f | 1109 | add_referenced_var (new_var); |
ebfd146a IR |
1110 | init_stmt = gimple_build_assign (new_var, vector_var); |
1111 | new_temp = make_ssa_name (new_var, init_stmt); | |
1112 | gimple_assign_set_lhs (init_stmt, new_temp); | |
1113 | ||
1114 | if (gsi) | |
1115 | vect_finish_stmt_generation (stmt, init_stmt, gsi); | |
1116 | else | |
1117 | { | |
1118 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
b8698a0f | 1119 | |
a70d6342 IR |
1120 | if (loop_vinfo) |
1121 | { | |
1122 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1123 | ||
1124 | if (nested_in_vect_loop_p (loop, stmt)) | |
1125 | loop = loop->inner; | |
b8698a0f | 1126 | |
a70d6342 IR |
1127 | pe = loop_preheader_edge (loop); |
1128 | new_bb = gsi_insert_on_edge_immediate (pe, init_stmt); | |
1129 | gcc_assert (!new_bb); | |
1130 | } | |
1131 | else | |
1132 | { | |
1133 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1134 | basic_block bb; | |
1135 | gimple_stmt_iterator gsi_bb_start; | |
1136 | ||
1137 | gcc_assert (bb_vinfo); | |
1138 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1139 | gsi_bb_start = gsi_after_labels (bb); |
a70d6342 IR |
1140 | gsi_insert_before (&gsi_bb_start, init_stmt, GSI_SAME_STMT); |
1141 | } | |
ebfd146a IR |
1142 | } |
1143 | ||
1144 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1145 | { | |
1146 | fprintf (vect_dump, "created new init_stmt: "); | |
1147 | print_gimple_stmt (vect_dump, init_stmt, 0, TDF_SLIM); | |
1148 | } | |
1149 | ||
1150 | vec_oprnd = gimple_assign_lhs (init_stmt); | |
1151 | return vec_oprnd; | |
1152 | } | |
1153 | ||
a70d6342 | 1154 | |
ebfd146a IR |
1155 | /* Function vect_get_vec_def_for_operand. |
1156 | ||
ff802fa1 | 1157 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
1158 | used in the vectorized stmt for STMT. |
1159 | ||
1160 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1161 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1162 | ||
1163 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1164 | needs to be introduced. */ | |
1165 | ||
1166 | tree | |
1167 | vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def) | |
1168 | { | |
1169 | tree vec_oprnd; | |
1170 | gimple vec_stmt; | |
1171 | gimple def_stmt; | |
1172 | stmt_vec_info def_stmt_info = NULL; | |
1173 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
9dc3f7de | 1174 | unsigned int nunits; |
ebfd146a IR |
1175 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
1176 | tree vec_inv; | |
1177 | tree vec_cst; | |
e7e9eb2f | 1178 | tree t = NULL_TREE; |
ebfd146a | 1179 | tree def; |
e7e9eb2f | 1180 | int i; |
ebfd146a IR |
1181 | enum vect_def_type dt; |
1182 | bool is_simple_use; | |
1183 | tree vector_type; | |
1184 | ||
1185 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1186 | { | |
1187 | fprintf (vect_dump, "vect_get_vec_def_for_operand: "); | |
1188 | print_generic_expr (vect_dump, op, TDF_SLIM); | |
1189 | } | |
1190 | ||
b8698a0f | 1191 | is_simple_use = vect_is_simple_use (op, loop_vinfo, NULL, &def_stmt, &def, |
a70d6342 | 1192 | &dt); |
ebfd146a IR |
1193 | gcc_assert (is_simple_use); |
1194 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1195 | { | |
1196 | if (def) | |
1197 | { | |
1198 | fprintf (vect_dump, "def = "); | |
1199 | print_generic_expr (vect_dump, def, TDF_SLIM); | |
1200 | } | |
1201 | if (def_stmt) | |
1202 | { | |
1203 | fprintf (vect_dump, " def_stmt = "); | |
1204 | print_gimple_stmt (vect_dump, def_stmt, 0, TDF_SLIM); | |
1205 | } | |
1206 | } | |
1207 | ||
1208 | switch (dt) | |
1209 | { | |
1210 | /* Case 1: operand is a constant. */ | |
1211 | case vect_constant_def: | |
1212 | { | |
7569a6cc RG |
1213 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
1214 | gcc_assert (vector_type); | |
9dc3f7de | 1215 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); |
7569a6cc | 1216 | |
b8698a0f | 1217 | if (scalar_def) |
ebfd146a IR |
1218 | *scalar_def = op; |
1219 | ||
1220 | /* Create 'vect_cst_ = {cst,cst,...,cst}' */ | |
1221 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1222 | fprintf (vect_dump, "Create vector_cst. nunits = %d", nunits); | |
1223 | ||
7b7b1813 RG |
1224 | vec_cst = build_vector_from_val (vector_type, |
1225 | fold_convert (TREE_TYPE (vector_type), | |
1226 | op)); | |
7569a6cc | 1227 | return vect_init_vector (stmt, vec_cst, vector_type, NULL); |
ebfd146a IR |
1228 | } |
1229 | ||
1230 | /* Case 2: operand is defined outside the loop - loop invariant. */ | |
8644a673 | 1231 | case vect_external_def: |
ebfd146a IR |
1232 | { |
1233 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); | |
1234 | gcc_assert (vector_type); | |
1235 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); | |
1236 | ||
b8698a0f | 1237 | if (scalar_def) |
ebfd146a IR |
1238 | *scalar_def = def; |
1239 | ||
1240 | /* Create 'vec_inv = {inv,inv,..,inv}' */ | |
1241 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1242 | fprintf (vect_dump, "Create vector_inv."); | |
1243 | ||
e7e9eb2f NF |
1244 | for (i = nunits - 1; i >= 0; --i) |
1245 | { | |
1246 | t = tree_cons (NULL_TREE, def, t); | |
1247 | } | |
1248 | ||
1249 | /* FIXME: use build_constructor directly. */ | |
1250 | vec_inv = build_constructor_from_list (vector_type, t); | |
ebfd146a IR |
1251 | return vect_init_vector (stmt, vec_inv, vector_type, NULL); |
1252 | } | |
1253 | ||
1254 | /* Case 3: operand is defined inside the loop. */ | |
8644a673 | 1255 | case vect_internal_def: |
ebfd146a | 1256 | { |
b8698a0f | 1257 | if (scalar_def) |
ebfd146a IR |
1258 | *scalar_def = NULL/* FIXME tuples: def_stmt*/; |
1259 | ||
1260 | /* Get the def from the vectorized stmt. */ | |
1261 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1262 | |
ebfd146a | 1263 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1264 | /* Get vectorized pattern statement. */ |
1265 | if (!vec_stmt | |
1266 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1267 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1268 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1269 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1270 | gcc_assert (vec_stmt); |
1271 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1272 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1273 | else if (is_gimple_call (vec_stmt)) | |
1274 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1275 | else | |
1276 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1277 | return vec_oprnd; | |
1278 | } | |
1279 | ||
1280 | /* Case 4: operand is defined by a loop header phi - reduction */ | |
1281 | case vect_reduction_def: | |
06066f92 | 1282 | case vect_double_reduction_def: |
7c5222ff | 1283 | case vect_nested_cycle: |
ebfd146a IR |
1284 | { |
1285 | struct loop *loop; | |
1286 | ||
1287 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
b8698a0f | 1288 | loop = (gimple_bb (def_stmt))->loop_father; |
ebfd146a IR |
1289 | |
1290 | /* Get the def before the loop */ | |
1291 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); | |
1292 | return get_initial_def_for_reduction (stmt, op, scalar_def); | |
1293 | } | |
1294 | ||
1295 | /* Case 5: operand is defined by loop-header phi - induction. */ | |
1296 | case vect_induction_def: | |
1297 | { | |
1298 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1299 | ||
1300 | /* Get the def from the vectorized stmt. */ | |
1301 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1302 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1303 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1304 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1305 | else | |
1306 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1307 | return vec_oprnd; |
1308 | } | |
1309 | ||
1310 | default: | |
1311 | gcc_unreachable (); | |
1312 | } | |
1313 | } | |
1314 | ||
1315 | ||
1316 | /* Function vect_get_vec_def_for_stmt_copy | |
1317 | ||
ff802fa1 | 1318 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1319 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1320 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1321 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1322 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1323 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1324 | DT is the type of the vector def VEC_OPRND. |
1325 | ||
1326 | Context: | |
1327 | In case the vectorization factor (VF) is bigger than the number | |
1328 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1329 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1330 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1331 | smallest data-type determines the VF, and as a result, when vectorizing |
1332 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1333 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1334 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1335 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1336 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1337 | ||
1338 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1339 | |
ebfd146a IR |
1340 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1341 | VS1.1: vx.1 = memref1 VS1.2 | |
1342 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1343 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1344 | |
1345 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1346 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1347 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1348 | VSnew.3: vz3 = vx.3 + ... | |
1349 | ||
1350 | The vectorization of S1 is explained in vectorizable_load. | |
1351 | The vectorization of S2: | |
b8698a0f L |
1352 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1353 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1354 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1355 | returns the vector-def 'vx.0'. |
1356 | ||
b8698a0f L |
1357 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1358 | function is called to get the relevant vector-def for each operand. It is | |
1359 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1360 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1361 | ||
b8698a0f L |
1362 | For example, to obtain the vector-def 'vx.1' in order to create the |
1363 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1364 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1365 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1366 | and return its def ('vx.1'). | |
1367 | Overall, to create the above sequence this function will be called 3 times: | |
1368 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1369 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1370 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1371 | ||
1372 | tree | |
1373 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1374 | { | |
1375 | gimple vec_stmt_for_operand; | |
1376 | stmt_vec_info def_stmt_info; | |
1377 | ||
1378 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1379 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1380 | return vec_oprnd; |
1381 | ||
1382 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1383 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1384 | gcc_assert (def_stmt_info); | |
1385 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1386 | gcc_assert (vec_stmt_for_operand); | |
1387 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1388 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) | |
1389 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1390 | else | |
1391 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1392 | return vec_oprnd; | |
1393 | } | |
1394 | ||
1395 | ||
1396 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1397 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1398 | |
1399 | static void | |
b8698a0f L |
1400 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
1401 | VEC(tree,heap) **vec_oprnds0, | |
ebfd146a IR |
1402 | VEC(tree,heap) **vec_oprnds1) |
1403 | { | |
1404 | tree vec_oprnd = VEC_pop (tree, *vec_oprnds0); | |
1405 | ||
1406 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
1407 | VEC_quick_push (tree, *vec_oprnds0, vec_oprnd); | |
1408 | ||
1409 | if (vec_oprnds1 && *vec_oprnds1) | |
1410 | { | |
1411 | vec_oprnd = VEC_pop (tree, *vec_oprnds1); | |
1412 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); | |
1413 | VEC_quick_push (tree, *vec_oprnds1, vec_oprnd); | |
1414 | } | |
1415 | } | |
1416 | ||
1417 | ||
d092494c IR |
1418 | /* Get vectorized definitions for OP0 and OP1. |
1419 | REDUC_INDEX is the index of reduction operand in case of reduction, | |
1420 | and -1 otherwise. */ | |
ebfd146a | 1421 | |
d092494c | 1422 | void |
ebfd146a | 1423 | vect_get_vec_defs (tree op0, tree op1, gimple stmt, |
d092494c IR |
1424 | VEC (tree, heap) **vec_oprnds0, |
1425 | VEC (tree, heap) **vec_oprnds1, | |
1426 | slp_tree slp_node, int reduc_index) | |
ebfd146a IR |
1427 | { |
1428 | if (slp_node) | |
d092494c IR |
1429 | { |
1430 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
1431 | VEC (tree, heap) *ops = VEC_alloc (tree, heap, nops); | |
1432 | VEC (slp_void_p, heap) *vec_defs = VEC_alloc (slp_void_p, heap, nops); | |
1433 | ||
1434 | VEC_quick_push (tree, ops, op0); | |
1435 | if (op1) | |
1436 | VEC_quick_push (tree, ops, op1); | |
1437 | ||
1438 | vect_get_slp_defs (ops, slp_node, &vec_defs, reduc_index); | |
1439 | ||
1440 | *vec_oprnds0 = (VEC (tree, heap) *) VEC_index (slp_void_p, vec_defs, 0); | |
1441 | if (op1) | |
1442 | *vec_oprnds1 = (VEC (tree, heap) *) VEC_index (slp_void_p, vec_defs, 1); | |
1443 | ||
1444 | VEC_free (tree, heap, ops); | |
1445 | VEC_free (slp_void_p, heap, vec_defs); | |
1446 | } | |
ebfd146a IR |
1447 | else |
1448 | { | |
1449 | tree vec_oprnd; | |
1450 | ||
b8698a0f L |
1451 | *vec_oprnds0 = VEC_alloc (tree, heap, 1); |
1452 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
ebfd146a IR |
1453 | VEC_quick_push (tree, *vec_oprnds0, vec_oprnd); |
1454 | ||
1455 | if (op1) | |
1456 | { | |
b8698a0f L |
1457 | *vec_oprnds1 = VEC_alloc (tree, heap, 1); |
1458 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL); | |
ebfd146a IR |
1459 | VEC_quick_push (tree, *vec_oprnds1, vec_oprnd); |
1460 | } | |
1461 | } | |
1462 | } | |
1463 | ||
1464 | ||
1465 | /* Function vect_finish_stmt_generation. | |
1466 | ||
1467 | Insert a new stmt. */ | |
1468 | ||
1469 | void | |
1470 | vect_finish_stmt_generation (gimple stmt, gimple vec_stmt, | |
1471 | gimple_stmt_iterator *gsi) | |
1472 | { | |
1473 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1474 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1475 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1476 | |
1477 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1478 | ||
1479 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); | |
1480 | ||
b8698a0f | 1481 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo, |
a70d6342 | 1482 | bb_vinfo)); |
ebfd146a IR |
1483 | |
1484 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1485 | { | |
1486 | fprintf (vect_dump, "add new stmt: "); | |
1487 | print_gimple_stmt (vect_dump, vec_stmt, 0, TDF_SLIM); | |
1488 | } | |
1489 | ||
ad885386 | 1490 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
ebfd146a IR |
1491 | } |
1492 | ||
1493 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1494 | a function declaration if the target has a vectorized version | |
1495 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1496 | ||
1497 | tree | |
1498 | vectorizable_function (gimple call, tree vectype_out, tree vectype_in) | |
1499 | { | |
1500 | tree fndecl = gimple_call_fndecl (call); | |
ebfd146a IR |
1501 | |
1502 | /* We only handle functions that do not read or clobber memory -- i.e. | |
1503 | const or novops ones. */ | |
1504 | if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS))) | |
1505 | return NULL_TREE; | |
1506 | ||
1507 | if (!fndecl | |
1508 | || TREE_CODE (fndecl) != FUNCTION_DECL | |
1509 | || !DECL_BUILT_IN (fndecl)) | |
1510 | return NULL_TREE; | |
1511 | ||
62f7fd21 | 1512 | return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out, |
ebfd146a IR |
1513 | vectype_in); |
1514 | } | |
1515 | ||
1516 | /* Function vectorizable_call. | |
1517 | ||
b8698a0f L |
1518 | Check if STMT performs a function call that can be vectorized. |
1519 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1520 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1521 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1522 | ||
1523 | static bool | |
190c2236 JJ |
1524 | vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, |
1525 | slp_tree slp_node) | |
ebfd146a IR |
1526 | { |
1527 | tree vec_dest; | |
1528 | tree scalar_dest; | |
1529 | tree op, type; | |
1530 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1531 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
1532 | tree vectype_out, vectype_in; | |
1533 | int nunits_in; | |
1534 | int nunits_out; | |
1535 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 1536 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b690cc0f | 1537 | tree fndecl, new_temp, def, rhs_type; |
ebfd146a | 1538 | gimple def_stmt; |
0502fb85 UB |
1539 | enum vect_def_type dt[3] |
1540 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
63827fb8 | 1541 | gimple new_stmt = NULL; |
ebfd146a IR |
1542 | int ncopies, j; |
1543 | VEC(tree, heap) *vargs = NULL; | |
1544 | enum { NARROW, NONE, WIDEN } modifier; | |
1545 | size_t i, nargs; | |
9d5e7640 | 1546 | tree lhs; |
ebfd146a | 1547 | |
190c2236 | 1548 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
1549 | return false; |
1550 | ||
8644a673 | 1551 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1552 | return false; |
1553 | ||
ebfd146a IR |
1554 | /* Is STMT a vectorizable call? */ |
1555 | if (!is_gimple_call (stmt)) | |
1556 | return false; | |
1557 | ||
1558 | if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
1559 | return false; | |
1560 | ||
822ba6d7 | 1561 | if (stmt_can_throw_internal (stmt)) |
5a2c1986 IR |
1562 | return false; |
1563 | ||
b690cc0f RG |
1564 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
1565 | ||
ebfd146a IR |
1566 | /* Process function arguments. */ |
1567 | rhs_type = NULL_TREE; | |
b690cc0f | 1568 | vectype_in = NULL_TREE; |
ebfd146a IR |
1569 | nargs = gimple_call_num_args (stmt); |
1570 | ||
1b1562a5 MM |
1571 | /* Bail out if the function has more than three arguments, we do not have |
1572 | interesting builtin functions to vectorize with more than two arguments | |
1573 | except for fma. No arguments is also not good. */ | |
1574 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
1575 | return false; |
1576 | ||
1577 | for (i = 0; i < nargs; i++) | |
1578 | { | |
b690cc0f RG |
1579 | tree opvectype; |
1580 | ||
ebfd146a IR |
1581 | op = gimple_call_arg (stmt, i); |
1582 | ||
1583 | /* We can only handle calls with arguments of the same type. */ | |
1584 | if (rhs_type | |
8533c9d8 | 1585 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a IR |
1586 | { |
1587 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1588 | fprintf (vect_dump, "argument types differ."); | |
1589 | return false; | |
1590 | } | |
b690cc0f RG |
1591 | if (!rhs_type) |
1592 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 1593 | |
190c2236 | 1594 | if (!vect_is_simple_use_1 (op, loop_vinfo, bb_vinfo, |
b690cc0f | 1595 | &def_stmt, &def, &dt[i], &opvectype)) |
ebfd146a IR |
1596 | { |
1597 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1598 | fprintf (vect_dump, "use not simple."); | |
1599 | return false; | |
1600 | } | |
ebfd146a | 1601 | |
b690cc0f RG |
1602 | if (!vectype_in) |
1603 | vectype_in = opvectype; | |
1604 | else if (opvectype | |
1605 | && opvectype != vectype_in) | |
1606 | { | |
1607 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1608 | fprintf (vect_dump, "argument vector types differ."); | |
1609 | return false; | |
1610 | } | |
1611 | } | |
1612 | /* If all arguments are external or constant defs use a vector type with | |
1613 | the same size as the output vector type. */ | |
ebfd146a | 1614 | if (!vectype_in) |
b690cc0f | 1615 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
1616 | if (vec_stmt) |
1617 | gcc_assert (vectype_in); | |
1618 | if (!vectype_in) | |
1619 | { | |
1620 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1621 | { | |
1622 | fprintf (vect_dump, "no vectype for scalar type "); | |
1623 | print_generic_expr (vect_dump, rhs_type, TDF_SLIM); | |
1624 | } | |
1625 | ||
1626 | return false; | |
1627 | } | |
ebfd146a IR |
1628 | |
1629 | /* FORNOW */ | |
b690cc0f RG |
1630 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
1631 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
1632 | if (nunits_in == nunits_out / 2) |
1633 | modifier = NARROW; | |
1634 | else if (nunits_out == nunits_in) | |
1635 | modifier = NONE; | |
1636 | else if (nunits_out == nunits_in / 2) | |
1637 | modifier = WIDEN; | |
1638 | else | |
1639 | return false; | |
1640 | ||
1641 | /* For now, we only vectorize functions if a target specific builtin | |
1642 | is available. TODO -- in some cases, it might be profitable to | |
1643 | insert the calls for pieces of the vector, in order to be able | |
1644 | to vectorize other operations in the loop. */ | |
1645 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
1646 | if (fndecl == NULL_TREE) | |
1647 | { | |
1648 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1649 | fprintf (vect_dump, "function is not vectorizable."); | |
1650 | ||
1651 | return false; | |
1652 | } | |
1653 | ||
5006671f | 1654 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a | 1655 | |
190c2236 JJ |
1656 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
1657 | ncopies = 1; | |
1658 | else if (modifier == NARROW) | |
ebfd146a IR |
1659 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
1660 | else | |
1661 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1662 | ||
1663 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
1664 | needs to be generated. */ | |
1665 | gcc_assert (ncopies >= 1); | |
1666 | ||
1667 | if (!vec_stmt) /* transformation not required. */ | |
1668 | { | |
1669 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1670 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1671 | fprintf (vect_dump, "=== vectorizable_call ==="); | |
1672 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
1673 | return true; | |
1674 | } | |
1675 | ||
1676 | /** Transform. **/ | |
1677 | ||
1678 | if (vect_print_dump_info (REPORT_DETAILS)) | |
9d5e7640 | 1679 | fprintf (vect_dump, "transform call."); |
ebfd146a IR |
1680 | |
1681 | /* Handle def. */ | |
1682 | scalar_dest = gimple_call_lhs (stmt); | |
1683 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1684 | ||
1685 | prev_stmt_info = NULL; | |
1686 | switch (modifier) | |
1687 | { | |
1688 | case NONE: | |
1689 | for (j = 0; j < ncopies; ++j) | |
1690 | { | |
1691 | /* Build argument list for the vectorized call. */ | |
1692 | if (j == 0) | |
1693 | vargs = VEC_alloc (tree, heap, nargs); | |
1694 | else | |
1695 | VEC_truncate (tree, vargs, 0); | |
1696 | ||
190c2236 JJ |
1697 | if (slp_node) |
1698 | { | |
1699 | VEC (slp_void_p, heap) *vec_defs | |
1700 | = VEC_alloc (slp_void_p, heap, nargs); | |
1701 | VEC (tree, heap) *vec_oprnds0; | |
1702 | ||
1703 | for (i = 0; i < nargs; i++) | |
1704 | VEC_quick_push (tree, vargs, gimple_call_arg (stmt, i)); | |
1705 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); | |
1706 | vec_oprnds0 | |
1707 | = (VEC (tree, heap) *) VEC_index (slp_void_p, vec_defs, 0); | |
1708 | ||
1709 | /* Arguments are ready. Create the new vector stmt. */ | |
1710 | FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vec_oprnd0) | |
1711 | { | |
1712 | size_t k; | |
1713 | for (k = 0; k < nargs; k++) | |
1714 | { | |
1715 | VEC (tree, heap) *vec_oprndsk | |
1716 | = (VEC (tree, heap) *) | |
1717 | VEC_index (slp_void_p, vec_defs, k); | |
1718 | VEC_replace (tree, vargs, k, | |
1719 | VEC_index (tree, vec_oprndsk, i)); | |
1720 | } | |
1721 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1722 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1723 | gimple_call_set_lhs (new_stmt, new_temp); | |
1724 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1725 | mark_symbols_for_renaming (new_stmt); | |
1726 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), | |
1727 | new_stmt); | |
1728 | } | |
1729 | ||
1730 | for (i = 0; i < nargs; i++) | |
1731 | { | |
1732 | VEC (tree, heap) *vec_oprndsi | |
1733 | = (VEC (tree, heap) *) | |
1734 | VEC_index (slp_void_p, vec_defs, i); | |
1735 | VEC_free (tree, heap, vec_oprndsi); | |
1736 | } | |
1737 | VEC_free (slp_void_p, heap, vec_defs); | |
1738 | continue; | |
1739 | } | |
1740 | ||
ebfd146a IR |
1741 | for (i = 0; i < nargs; i++) |
1742 | { | |
1743 | op = gimple_call_arg (stmt, i); | |
1744 | if (j == 0) | |
1745 | vec_oprnd0 | |
1746 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1747 | else | |
63827fb8 IR |
1748 | { |
1749 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
1750 | vec_oprnd0 | |
1751 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
1752 | } | |
ebfd146a IR |
1753 | |
1754 | VEC_quick_push (tree, vargs, vec_oprnd0); | |
1755 | } | |
1756 | ||
1757 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1758 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1759 | gimple_call_set_lhs (new_stmt, new_temp); | |
1760 | ||
1761 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7411b8f0 | 1762 | mark_symbols_for_renaming (new_stmt); |
ebfd146a IR |
1763 | |
1764 | if (j == 0) | |
1765 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1766 | else | |
1767 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1768 | ||
1769 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1770 | } | |
1771 | ||
1772 | break; | |
1773 | ||
1774 | case NARROW: | |
1775 | for (j = 0; j < ncopies; ++j) | |
1776 | { | |
1777 | /* Build argument list for the vectorized call. */ | |
1778 | if (j == 0) | |
1779 | vargs = VEC_alloc (tree, heap, nargs * 2); | |
1780 | else | |
1781 | VEC_truncate (tree, vargs, 0); | |
1782 | ||
190c2236 JJ |
1783 | if (slp_node) |
1784 | { | |
1785 | VEC (slp_void_p, heap) *vec_defs | |
1786 | = VEC_alloc (slp_void_p, heap, nargs); | |
1787 | VEC (tree, heap) *vec_oprnds0; | |
1788 | ||
1789 | for (i = 0; i < nargs; i++) | |
1790 | VEC_quick_push (tree, vargs, gimple_call_arg (stmt, i)); | |
1791 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); | |
1792 | vec_oprnds0 | |
1793 | = (VEC (tree, heap) *) VEC_index (slp_void_p, vec_defs, 0); | |
1794 | ||
1795 | /* Arguments are ready. Create the new vector stmt. */ | |
1796 | for (i = 0; VEC_iterate (tree, vec_oprnds0, i, vec_oprnd0); | |
1797 | i += 2) | |
1798 | { | |
1799 | size_t k; | |
1800 | VEC_truncate (tree, vargs, 0); | |
1801 | for (k = 0; k < nargs; k++) | |
1802 | { | |
1803 | VEC (tree, heap) *vec_oprndsk | |
1804 | = (VEC (tree, heap) *) | |
1805 | VEC_index (slp_void_p, vec_defs, k); | |
1806 | VEC_quick_push (tree, vargs, | |
1807 | VEC_index (tree, vec_oprndsk, i)); | |
1808 | VEC_quick_push (tree, vargs, | |
1809 | VEC_index (tree, vec_oprndsk, i + 1)); | |
1810 | } | |
1811 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1812 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1813 | gimple_call_set_lhs (new_stmt, new_temp); | |
1814 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1815 | mark_symbols_for_renaming (new_stmt); | |
1816 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), | |
1817 | new_stmt); | |
1818 | } | |
1819 | ||
1820 | for (i = 0; i < nargs; i++) | |
1821 | { | |
1822 | VEC (tree, heap) *vec_oprndsi | |
1823 | = (VEC (tree, heap) *) | |
1824 | VEC_index (slp_void_p, vec_defs, i); | |
1825 | VEC_free (tree, heap, vec_oprndsi); | |
1826 | } | |
1827 | VEC_free (slp_void_p, heap, vec_defs); | |
1828 | continue; | |
1829 | } | |
1830 | ||
ebfd146a IR |
1831 | for (i = 0; i < nargs; i++) |
1832 | { | |
1833 | op = gimple_call_arg (stmt, i); | |
1834 | if (j == 0) | |
1835 | { | |
1836 | vec_oprnd0 | |
1837 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1838 | vec_oprnd1 | |
63827fb8 | 1839 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
1840 | } |
1841 | else | |
1842 | { | |
336ecb65 | 1843 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 1844 | vec_oprnd0 |
63827fb8 | 1845 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 1846 | vec_oprnd1 |
63827fb8 | 1847 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
1848 | } |
1849 | ||
1850 | VEC_quick_push (tree, vargs, vec_oprnd0); | |
1851 | VEC_quick_push (tree, vargs, vec_oprnd1); | |
1852 | } | |
1853 | ||
1854 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1855 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1856 | gimple_call_set_lhs (new_stmt, new_temp); | |
1857 | ||
1858 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7411b8f0 | 1859 | mark_symbols_for_renaming (new_stmt); |
ebfd146a IR |
1860 | |
1861 | if (j == 0) | |
1862 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1863 | else | |
1864 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1865 | ||
1866 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1867 | } | |
1868 | ||
1869 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
1870 | ||
1871 | break; | |
1872 | ||
1873 | case WIDEN: | |
1874 | /* No current target implements this case. */ | |
1875 | return false; | |
1876 | } | |
1877 | ||
1878 | VEC_free (tree, heap, vargs); | |
1879 | ||
1880 | /* Update the exception handling table with the vector stmt if necessary. */ | |
1881 | if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt)) | |
1882 | gimple_purge_dead_eh_edges (gimple_bb (stmt)); | |
1883 | ||
1884 | /* The call in STMT might prevent it from being removed in dce. | |
1885 | We however cannot remove it here, due to the way the ssa name | |
1886 | it defines is mapped to the new definition. So just replace | |
1887 | rhs of the statement with something harmless. */ | |
1888 | ||
dd34c087 JJ |
1889 | if (slp_node) |
1890 | return true; | |
1891 | ||
ebfd146a | 1892 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
1893 | if (is_pattern_stmt_p (stmt_info)) |
1894 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
1895 | else | |
1896 | lhs = gimple_call_lhs (stmt); | |
1897 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
ebfd146a | 1898 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 1899 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
1900 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
1901 | gsi_replace (gsi, new_stmt, false); | |
1902 | SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt; | |
1903 | ||
1904 | return true; | |
1905 | } | |
1906 | ||
1907 | ||
1908 | /* Function vect_gen_widened_results_half | |
1909 | ||
1910 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 1911 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 1912 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
1913 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
1914 | needs to be created (DECL is a function-decl of a target-builtin). | |
1915 | STMT is the original scalar stmt that we are vectorizing. */ | |
1916 | ||
1917 | static gimple | |
1918 | vect_gen_widened_results_half (enum tree_code code, | |
1919 | tree decl, | |
1920 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
1921 | tree vec_dest, gimple_stmt_iterator *gsi, | |
1922 | gimple stmt) | |
b8698a0f | 1923 | { |
ebfd146a | 1924 | gimple new_stmt; |
b8698a0f L |
1925 | tree new_temp; |
1926 | ||
1927 | /* Generate half of the widened result: */ | |
1928 | if (code == CALL_EXPR) | |
1929 | { | |
1930 | /* Target specific support */ | |
ebfd146a IR |
1931 | if (op_type == binary_op) |
1932 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
1933 | else | |
1934 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
1935 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1936 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
1937 | } |
1938 | else | |
ebfd146a | 1939 | { |
b8698a0f L |
1940 | /* Generic support */ |
1941 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
1942 | if (op_type != binary_op) |
1943 | vec_oprnd1 = NULL; | |
1944 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0, | |
1945 | vec_oprnd1); | |
1946 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1947 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 1948 | } |
ebfd146a IR |
1949 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1950 | ||
ebfd146a IR |
1951 | return new_stmt; |
1952 | } | |
1953 | ||
4a00c761 JJ |
1954 | |
1955 | /* Get vectorized definitions for loop-based vectorization. For the first | |
1956 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
1957 | scalar operand), and for the rest we get a copy with | |
1958 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
1959 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
1960 | The vectors are collected into VEC_OPRNDS. */ | |
1961 | ||
1962 | static void | |
1963 | vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt, | |
1964 | VEC (tree, heap) **vec_oprnds, int multi_step_cvt) | |
1965 | { | |
1966 | tree vec_oprnd; | |
1967 | ||
1968 | /* Get first vector operand. */ | |
1969 | /* All the vector operands except the very first one (that is scalar oprnd) | |
1970 | are stmt copies. */ | |
1971 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
1972 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL); | |
1973 | else | |
1974 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
1975 | ||
1976 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
1977 | ||
1978 | /* Get second vector operand. */ | |
1979 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
1980 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
1981 | ||
1982 | *oprnd = vec_oprnd; | |
1983 | ||
1984 | /* For conversion in multiple steps, continue to get operands | |
1985 | recursively. */ | |
1986 | if (multi_step_cvt) | |
1987 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
1988 | } | |
1989 | ||
1990 | ||
1991 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
1992 | For multi-step conversions store the resulting vectors and call the function | |
1993 | recursively. */ | |
1994 | ||
1995 | static void | |
1996 | vect_create_vectorized_demotion_stmts (VEC (tree, heap) **vec_oprnds, | |
1997 | int multi_step_cvt, gimple stmt, | |
1998 | VEC (tree, heap) *vec_dsts, | |
1999 | gimple_stmt_iterator *gsi, | |
2000 | slp_tree slp_node, enum tree_code code, | |
2001 | stmt_vec_info *prev_stmt_info) | |
2002 | { | |
2003 | unsigned int i; | |
2004 | tree vop0, vop1, new_tmp, vec_dest; | |
2005 | gimple new_stmt; | |
2006 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2007 | ||
2008 | vec_dest = VEC_pop (tree, vec_dsts); | |
2009 | ||
2010 | for (i = 0; i < VEC_length (tree, *vec_oprnds); i += 2) | |
2011 | { | |
2012 | /* Create demotion operation. */ | |
2013 | vop0 = VEC_index (tree, *vec_oprnds, i); | |
2014 | vop1 = VEC_index (tree, *vec_oprnds, i + 1); | |
2015 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); | |
2016 | new_tmp = make_ssa_name (vec_dest, new_stmt); | |
2017 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
2018 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2019 | ||
2020 | if (multi_step_cvt) | |
2021 | /* Store the resulting vector for next recursive call. */ | |
2022 | VEC_replace (tree, *vec_oprnds, i/2, new_tmp); | |
2023 | else | |
2024 | { | |
2025 | /* This is the last step of the conversion sequence. Store the | |
2026 | vectors in SLP_NODE or in vector info of the scalar statement | |
2027 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
2028 | if (slp_node) | |
2029 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2030 | else | |
2031 | { | |
2032 | if (!*prev_stmt_info) | |
2033 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2034 | else | |
2035 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
2036 | ||
2037 | *prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2038 | } | |
2039 | } | |
2040 | } | |
2041 | ||
2042 | /* For multi-step demotion operations we first generate demotion operations | |
2043 | from the source type to the intermediate types, and then combine the | |
2044 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
2045 | type. */ | |
2046 | if (multi_step_cvt) | |
2047 | { | |
2048 | /* At each level of recursion we have half of the operands we had at the | |
2049 | previous level. */ | |
2050 | VEC_truncate (tree, *vec_oprnds, (i+1)/2); | |
2051 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, | |
2052 | stmt, vec_dsts, gsi, slp_node, | |
2053 | VEC_PACK_TRUNC_EXPR, | |
2054 | prev_stmt_info); | |
2055 | } | |
2056 | ||
2057 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2058 | } | |
2059 | ||
2060 | ||
2061 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
2062 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
2063 | the resulting vectors and call the function recursively. */ | |
2064 | ||
2065 | static void | |
2066 | vect_create_vectorized_promotion_stmts (VEC (tree, heap) **vec_oprnds0, | |
2067 | VEC (tree, heap) **vec_oprnds1, | |
2068 | gimple stmt, tree vec_dest, | |
2069 | gimple_stmt_iterator *gsi, | |
2070 | enum tree_code code1, | |
2071 | enum tree_code code2, tree decl1, | |
2072 | tree decl2, int op_type) | |
2073 | { | |
2074 | int i; | |
2075 | tree vop0, vop1, new_tmp1, new_tmp2; | |
2076 | gimple new_stmt1, new_stmt2; | |
2077 | VEC (tree, heap) *vec_tmp = NULL; | |
2078 | ||
2079 | vec_tmp = VEC_alloc (tree, heap, VEC_length (tree, *vec_oprnds0) * 2); | |
2080 | FOR_EACH_VEC_ELT (tree, *vec_oprnds0, i, vop0) | |
2081 | { | |
2082 | if (op_type == binary_op) | |
2083 | vop1 = VEC_index (tree, *vec_oprnds1, i); | |
2084 | else | |
2085 | vop1 = NULL_TREE; | |
2086 | ||
2087 | /* Generate the two halves of promotion operation. */ | |
2088 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
2089 | op_type, vec_dest, gsi, stmt); | |
2090 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
2091 | op_type, vec_dest, gsi, stmt); | |
2092 | if (is_gimple_call (new_stmt1)) | |
2093 | { | |
2094 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
2095 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
2096 | } | |
2097 | else | |
2098 | { | |
2099 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
2100 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
2101 | } | |
2102 | ||
2103 | /* Store the results for the next step. */ | |
2104 | VEC_quick_push (tree, vec_tmp, new_tmp1); | |
2105 | VEC_quick_push (tree, vec_tmp, new_tmp2); | |
2106 | } | |
2107 | ||
2108 | VEC_free (tree, heap, *vec_oprnds0); | |
2109 | *vec_oprnds0 = vec_tmp; | |
2110 | } | |
2111 | ||
2112 | ||
b8698a0f L |
2113 | /* Check if STMT performs a conversion operation, that can be vectorized. |
2114 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 2115 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
2116 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
2117 | ||
2118 | static bool | |
2119 | vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi, | |
2120 | gimple *vec_stmt, slp_tree slp_node) | |
2121 | { | |
2122 | tree vec_dest; | |
2123 | tree scalar_dest; | |
4a00c761 | 2124 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
2125 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
2126 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2127 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2128 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 2129 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
2130 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
2131 | tree new_temp; | |
2132 | tree def; | |
2133 | gimple def_stmt; | |
2134 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2135 | gimple new_stmt = NULL; | |
2136 | stmt_vec_info prev_stmt_info; | |
2137 | int nunits_in; | |
2138 | int nunits_out; | |
2139 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
2140 | int ncopies, i, j; |
2141 | tree lhs_type, rhs_type; | |
ebfd146a | 2142 | enum { NARROW, NONE, WIDEN } modifier; |
4a00c761 | 2143 | VEC (tree,heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; |
ebfd146a | 2144 | tree vop0; |
4a00c761 JJ |
2145 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2146 | int multi_step_cvt = 0; | |
2147 | VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL; | |
2148 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; | |
2149 | int op_type; | |
2150 | enum machine_mode rhs_mode; | |
2151 | unsigned short fltsz; | |
ebfd146a IR |
2152 | |
2153 | /* Is STMT a vectorizable conversion? */ | |
2154 | ||
4a00c761 | 2155 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2156 | return false; |
2157 | ||
8644a673 | 2158 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2159 | return false; |
2160 | ||
2161 | if (!is_gimple_assign (stmt)) | |
2162 | return false; | |
2163 | ||
2164 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2165 | return false; | |
2166 | ||
2167 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
2168 | if (!CONVERT_EXPR_CODE_P (code) |
2169 | && code != FIX_TRUNC_EXPR | |
2170 | && code != FLOAT_EXPR | |
2171 | && code != WIDEN_MULT_EXPR | |
2172 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
2173 | return false; |
2174 | ||
4a00c761 JJ |
2175 | op_type = TREE_CODE_LENGTH (code); |
2176 | ||
ebfd146a | 2177 | /* Check types of lhs and rhs. */ |
b690cc0f | 2178 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 2179 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
2180 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2181 | ||
ebfd146a IR |
2182 | op0 = gimple_assign_rhs1 (stmt); |
2183 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
2184 | |
2185 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
2186 | && !((INTEGRAL_TYPE_P (lhs_type) | |
2187 | && INTEGRAL_TYPE_P (rhs_type)) | |
2188 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
2189 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
2190 | return false; | |
2191 | ||
2192 | if ((INTEGRAL_TYPE_P (lhs_type) | |
2193 | && (TYPE_PRECISION (lhs_type) | |
2194 | != GET_MODE_PRECISION (TYPE_MODE (lhs_type)))) | |
2195 | || (INTEGRAL_TYPE_P (rhs_type) | |
2196 | && (TYPE_PRECISION (rhs_type) | |
2197 | != GET_MODE_PRECISION (TYPE_MODE (rhs_type))))) | |
2198 | { | |
2199 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2200 | fprintf (vect_dump, | |
2201 | "type conversion to/from bit-precision unsupported."); | |
2202 | return false; | |
2203 | } | |
2204 | ||
b690cc0f | 2205 | /* Check the operands of the operation. */ |
4a00c761 | 2206 | if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo, |
b690cc0f RG |
2207 | &def_stmt, &def, &dt[0], &vectype_in)) |
2208 | { | |
2209 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2210 | fprintf (vect_dump, "use not simple."); | |
2211 | return false; | |
2212 | } | |
4a00c761 JJ |
2213 | if (op_type == binary_op) |
2214 | { | |
2215 | bool ok; | |
2216 | ||
2217 | op1 = gimple_assign_rhs2 (stmt); | |
2218 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
2219 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
2220 | OP1. */ | |
2221 | if (CONSTANT_CLASS_P (op0)) | |
2222 | ok = vect_is_simple_use_1 (op1, loop_vinfo, NULL, | |
2223 | &def_stmt, &def, &dt[1], &vectype_in); | |
2224 | else | |
2225 | ok = vect_is_simple_use (op1, loop_vinfo, NULL, &def_stmt, &def, | |
2226 | &dt[1]); | |
2227 | ||
2228 | if (!ok) | |
2229 | { | |
2230 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2231 | fprintf (vect_dump, "use not simple."); | |
2232 | return false; | |
2233 | } | |
2234 | } | |
2235 | ||
b690cc0f RG |
2236 | /* If op0 is an external or constant defs use a vector type of |
2237 | the same size as the output vector type. */ | |
ebfd146a | 2238 | if (!vectype_in) |
b690cc0f | 2239 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2240 | if (vec_stmt) |
2241 | gcc_assert (vectype_in); | |
2242 | if (!vectype_in) | |
2243 | { | |
2244 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4a00c761 JJ |
2245 | { |
2246 | fprintf (vect_dump, "no vectype for scalar type "); | |
2247 | print_generic_expr (vect_dump, rhs_type, TDF_SLIM); | |
2248 | } | |
7d8930a0 IR |
2249 | |
2250 | return false; | |
2251 | } | |
ebfd146a | 2252 | |
b690cc0f RG |
2253 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2254 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 2255 | if (nunits_in < nunits_out) |
ebfd146a IR |
2256 | modifier = NARROW; |
2257 | else if (nunits_out == nunits_in) | |
2258 | modifier = NONE; | |
ebfd146a | 2259 | else |
4a00c761 | 2260 | modifier = WIDEN; |
ebfd146a | 2261 | |
ff802fa1 IR |
2262 | /* Multiple types in SLP are handled by creating the appropriate number of |
2263 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2264 | case of SLP. */ | |
437f4a00 | 2265 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a | 2266 | ncopies = 1; |
4a00c761 JJ |
2267 | else if (modifier == NARROW) |
2268 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
2269 | else | |
2270 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 2271 | |
ebfd146a IR |
2272 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
2273 | needs to be generated. */ | |
2274 | gcc_assert (ncopies >= 1); | |
2275 | ||
ebfd146a | 2276 | /* Supportable by target? */ |
4a00c761 | 2277 | switch (modifier) |
ebfd146a | 2278 | { |
4a00c761 JJ |
2279 | case NONE: |
2280 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
2281 | return false; | |
2282 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
2283 | &decl1, &code1)) | |
2284 | break; | |
2285 | /* FALLTHRU */ | |
2286 | unsupported: | |
ebfd146a | 2287 | if (vect_print_dump_info (REPORT_DETAILS)) |
4a00c761 | 2288 | fprintf (vect_dump, "conversion not supported by target."); |
ebfd146a | 2289 | return false; |
ebfd146a | 2290 | |
4a00c761 JJ |
2291 | case WIDEN: |
2292 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
2293 | &decl1, &decl2, &code1, &code2, | |
2294 | &multi_step_cvt, &interm_types)) | |
2295 | { | |
2296 | /* Binary widening operation can only be supported directly by the | |
2297 | architecture. */ | |
2298 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
2299 | break; | |
2300 | } | |
2301 | ||
2302 | if (code != FLOAT_EXPR | |
2303 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
2304 | <= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
2305 | goto unsupported; | |
2306 | ||
2307 | rhs_mode = TYPE_MODE (rhs_type); | |
2308 | fltsz = GET_MODE_SIZE (TYPE_MODE (lhs_type)); | |
2309 | for (rhs_mode = GET_MODE_2XWIDER_MODE (TYPE_MODE (rhs_type)); | |
2310 | rhs_mode != VOIDmode && GET_MODE_SIZE (rhs_mode) <= fltsz; | |
2311 | rhs_mode = GET_MODE_2XWIDER_MODE (rhs_mode)) | |
2312 | { | |
2313 | cvt_type | |
2314 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
2315 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
2316 | if (cvt_type == NULL_TREE) | |
2317 | goto unsupported; | |
2318 | ||
2319 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
2320 | { | |
2321 | if (!supportable_convert_operation (code, vectype_out, | |
2322 | cvt_type, &decl1, &codecvt1)) | |
2323 | goto unsupported; | |
2324 | } | |
2325 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
2326 | cvt_type, &decl1, &decl2, | |
2327 | &codecvt1, &codecvt2, | |
2328 | &multi_step_cvt, | |
2329 | &interm_types)) | |
2330 | continue; | |
2331 | else | |
2332 | gcc_assert (multi_step_cvt == 0); | |
2333 | ||
2334 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
2335 | vectype_in, NULL, NULL, &code1, | |
2336 | &code2, &multi_step_cvt, | |
2337 | &interm_types)) | |
2338 | break; | |
2339 | } | |
2340 | ||
2341 | if (rhs_mode == VOIDmode || GET_MODE_SIZE (rhs_mode) > fltsz) | |
2342 | goto unsupported; | |
2343 | ||
2344 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
2345 | codecvt2 = ERROR_MARK; | |
2346 | else | |
2347 | { | |
2348 | multi_step_cvt++; | |
2349 | VEC_safe_push (tree, heap, interm_types, cvt_type); | |
2350 | cvt_type = NULL_TREE; | |
2351 | } | |
2352 | break; | |
2353 | ||
2354 | case NARROW: | |
2355 | gcc_assert (op_type == unary_op); | |
2356 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
2357 | &code1, &multi_step_cvt, | |
2358 | &interm_types)) | |
2359 | break; | |
2360 | ||
2361 | if (code != FIX_TRUNC_EXPR | |
2362 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
2363 | >= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
2364 | goto unsupported; | |
2365 | ||
2366 | rhs_mode = TYPE_MODE (rhs_type); | |
2367 | cvt_type | |
2368 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
2369 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
2370 | if (cvt_type == NULL_TREE) | |
2371 | goto unsupported; | |
2372 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
2373 | &decl1, &codecvt1)) | |
2374 | goto unsupported; | |
2375 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
2376 | &code1, &multi_step_cvt, | |
2377 | &interm_types)) | |
2378 | break; | |
2379 | goto unsupported; | |
2380 | ||
2381 | default: | |
2382 | gcc_unreachable (); | |
ebfd146a IR |
2383 | } |
2384 | ||
2385 | if (!vec_stmt) /* transformation not required. */ | |
2386 | { | |
4a00c761 JJ |
2387 | if (vect_print_dump_info (REPORT_DETAILS)) |
2388 | fprintf (vect_dump, "=== vectorizable_conversion ==="); | |
2389 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) | |
2390 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
2391 | else if (modifier == NARROW) | |
2392 | { | |
2393 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
2394 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2395 | } | |
2396 | else | |
2397 | { | |
2398 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
2399 | vect_model_simple_cost (stmt_info, 2 * ncopies, dt, NULL); | |
2400 | } | |
2401 | VEC_free (tree, heap, interm_types); | |
ebfd146a IR |
2402 | return true; |
2403 | } | |
2404 | ||
2405 | /** Transform. **/ | |
2406 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4a00c761 | 2407 | fprintf (vect_dump, "transform conversion. ncopies = %d.", ncopies); |
ebfd146a | 2408 | |
4a00c761 JJ |
2409 | if (op_type == binary_op) |
2410 | { | |
2411 | if (CONSTANT_CLASS_P (op0)) | |
2412 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
2413 | else if (CONSTANT_CLASS_P (op1)) | |
2414 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
2415 | } | |
2416 | ||
2417 | /* In case of multi-step conversion, we first generate conversion operations | |
2418 | to the intermediate types, and then from that types to the final one. | |
2419 | We create vector destinations for the intermediate type (TYPES) received | |
2420 | from supportable_*_operation, and store them in the correct order | |
2421 | for future use in vect_create_vectorized_*_stmts (). */ | |
2422 | vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1); | |
ebfd146a | 2423 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); |
4a00c761 JJ |
2424 | VEC_quick_push (tree, vec_dsts, vec_dest); |
2425 | ||
2426 | if (multi_step_cvt) | |
2427 | { | |
2428 | for (i = VEC_length (tree, interm_types) - 1; | |
2429 | VEC_iterate (tree, interm_types, i, intermediate_type); i--) | |
2430 | { | |
2431 | vec_dest = vect_create_destination_var (scalar_dest, | |
2432 | intermediate_type); | |
2433 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2434 | } | |
2435 | } | |
ebfd146a | 2436 | |
4a00c761 JJ |
2437 | if (cvt_type) |
2438 | vec_dest = vect_create_destination_var (scalar_dest, cvt_type); | |
2439 | ||
2440 | if (!slp_node) | |
2441 | { | |
2442 | if (modifier == NONE) | |
2443 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
2444 | else if (modifier == WIDEN) | |
2445 | { | |
2446 | vec_oprnds0 = VEC_alloc (tree, heap, | |
2447 | (multi_step_cvt | |
2448 | ? vect_pow2 (multi_step_cvt) : 1)); | |
2449 | if (op_type == binary_op) | |
2450 | vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
2451 | } | |
2452 | else | |
2453 | vec_oprnds0 = VEC_alloc (tree, heap, | |
2454 | 2 * (multi_step_cvt | |
2455 | ? vect_pow2 (multi_step_cvt) : 1)); | |
2456 | } | |
2457 | else if (code == WIDEN_LSHIFT_EXPR) | |
2458 | vec_oprnds1 = VEC_alloc (tree, heap, slp_node->vec_stmts_size); | |
ebfd146a | 2459 | |
4a00c761 | 2460 | last_oprnd = op0; |
ebfd146a IR |
2461 | prev_stmt_info = NULL; |
2462 | switch (modifier) | |
2463 | { | |
2464 | case NONE: | |
2465 | for (j = 0; j < ncopies; j++) | |
2466 | { | |
ebfd146a | 2467 | if (j == 0) |
d092494c IR |
2468 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node, |
2469 | -1); | |
ebfd146a IR |
2470 | else |
2471 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
2472 | ||
ac47786e | 2473 | FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0) |
4a00c761 JJ |
2474 | { |
2475 | /* Arguments are ready, create the new vector stmt. */ | |
2476 | if (code1 == CALL_EXPR) | |
2477 | { | |
2478 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2479 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2480 | gimple_call_set_lhs (new_stmt, new_temp); | |
2481 | } | |
2482 | else | |
2483 | { | |
2484 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
2485 | new_stmt = gimple_build_assign_with_ops (code1, vec_dest, | |
2486 | vop0, NULL); | |
2487 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2488 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2489 | } | |
2490 | ||
2491 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2492 | if (slp_node) | |
2493 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), | |
2494 | new_stmt); | |
2495 | } | |
2496 | ||
ebfd146a IR |
2497 | if (j == 0) |
2498 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2499 | else | |
2500 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2501 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2502 | } | |
2503 | break; | |
2504 | ||
2505 | case WIDEN: | |
2506 | /* In case the vectorization factor (VF) is bigger than the number | |
2507 | of elements that we can fit in a vectype (nunits), we have to | |
2508 | generate more than one vector stmt - i.e - we need to "unroll" | |
2509 | the vector stmt by a factor VF/nunits. */ | |
2510 | for (j = 0; j < ncopies; j++) | |
2511 | { | |
4a00c761 | 2512 | /* Handle uses. */ |
ebfd146a | 2513 | if (j == 0) |
4a00c761 JJ |
2514 | { |
2515 | if (slp_node) | |
2516 | { | |
2517 | if (code == WIDEN_LSHIFT_EXPR) | |
2518 | { | |
2519 | unsigned int k; | |
ebfd146a | 2520 | |
4a00c761 JJ |
2521 | vec_oprnd1 = op1; |
2522 | /* Store vec_oprnd1 for every vector stmt to be created | |
2523 | for SLP_NODE. We check during the analysis that all | |
2524 | the shift arguments are the same. */ | |
2525 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
2526 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2527 | ||
2528 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2529 | slp_node, -1); | |
2530 | } | |
2531 | else | |
2532 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
2533 | &vec_oprnds1, slp_node, -1); | |
2534 | } | |
2535 | else | |
2536 | { | |
2537 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
2538 | VEC_quick_push (tree, vec_oprnds0, vec_oprnd0); | |
2539 | if (op_type == binary_op) | |
2540 | { | |
2541 | if (code == WIDEN_LSHIFT_EXPR) | |
2542 | vec_oprnd1 = op1; | |
2543 | else | |
2544 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, | |
2545 | NULL); | |
2546 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2547 | } | |
2548 | } | |
2549 | } | |
ebfd146a | 2550 | else |
4a00c761 JJ |
2551 | { |
2552 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
2553 | VEC_truncate (tree, vec_oprnds0, 0); | |
2554 | VEC_quick_push (tree, vec_oprnds0, vec_oprnd0); | |
2555 | if (op_type == binary_op) | |
2556 | { | |
2557 | if (code == WIDEN_LSHIFT_EXPR) | |
2558 | vec_oprnd1 = op1; | |
2559 | else | |
2560 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
2561 | vec_oprnd1); | |
2562 | VEC_truncate (tree, vec_oprnds1, 0); | |
2563 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2564 | } | |
2565 | } | |
ebfd146a | 2566 | |
4a00c761 JJ |
2567 | /* Arguments are ready. Create the new vector stmts. */ |
2568 | for (i = multi_step_cvt; i >= 0; i--) | |
2569 | { | |
2570 | tree this_dest = VEC_index (tree, vec_dsts, i); | |
2571 | enum tree_code c1 = code1, c2 = code2; | |
2572 | if (i == 0 && codecvt2 != ERROR_MARK) | |
2573 | { | |
2574 | c1 = codecvt1; | |
2575 | c2 = codecvt2; | |
2576 | } | |
2577 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
2578 | &vec_oprnds1, | |
2579 | stmt, this_dest, gsi, | |
2580 | c1, c2, decl1, decl2, | |
2581 | op_type); | |
2582 | } | |
2583 | ||
2584 | FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0) | |
2585 | { | |
2586 | if (cvt_type) | |
2587 | { | |
2588 | if (codecvt1 == CALL_EXPR) | |
2589 | { | |
2590 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2591 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2592 | gimple_call_set_lhs (new_stmt, new_temp); | |
2593 | } | |
2594 | else | |
2595 | { | |
2596 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
2597 | new_temp = make_ssa_name (vec_dest, NULL); | |
2598 | new_stmt = gimple_build_assign_with_ops (codecvt1, | |
2599 | new_temp, | |
2600 | vop0, NULL); | |
2601 | } | |
2602 | ||
2603 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2604 | } | |
2605 | else | |
2606 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
2607 | ||
2608 | if (slp_node) | |
2609 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), | |
2610 | new_stmt); | |
2611 | else | |
2612 | { | |
2613 | if (!prev_stmt_info) | |
2614 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2615 | else | |
2616 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2617 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2618 | } | |
2619 | } | |
ebfd146a | 2620 | } |
4a00c761 JJ |
2621 | |
2622 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
2623 | break; |
2624 | ||
2625 | case NARROW: | |
2626 | /* In case the vectorization factor (VF) is bigger than the number | |
2627 | of elements that we can fit in a vectype (nunits), we have to | |
2628 | generate more than one vector stmt - i.e - we need to "unroll" | |
2629 | the vector stmt by a factor VF/nunits. */ | |
2630 | for (j = 0; j < ncopies; j++) | |
2631 | { | |
2632 | /* Handle uses. */ | |
4a00c761 JJ |
2633 | if (slp_node) |
2634 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2635 | slp_node, -1); | |
ebfd146a IR |
2636 | else |
2637 | { | |
4a00c761 JJ |
2638 | VEC_truncate (tree, vec_oprnds0, 0); |
2639 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, | |
2640 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
2641 | } |
2642 | ||
4a00c761 JJ |
2643 | /* Arguments are ready. Create the new vector stmts. */ |
2644 | if (cvt_type) | |
2645 | FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0) | |
2646 | { | |
2647 | if (codecvt1 == CALL_EXPR) | |
2648 | { | |
2649 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2650 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2651 | gimple_call_set_lhs (new_stmt, new_temp); | |
2652 | } | |
2653 | else | |
2654 | { | |
2655 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
2656 | new_temp = make_ssa_name (vec_dest, NULL); | |
2657 | new_stmt = gimple_build_assign_with_ops (codecvt1, new_temp, | |
2658 | vop0, NULL); | |
2659 | } | |
ebfd146a | 2660 | |
4a00c761 JJ |
2661 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2662 | VEC_replace (tree, vec_oprnds0, i, new_temp); | |
2663 | } | |
ebfd146a | 2664 | |
4a00c761 JJ |
2665 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
2666 | stmt, vec_dsts, gsi, | |
2667 | slp_node, code1, | |
2668 | &prev_stmt_info); | |
ebfd146a IR |
2669 | } |
2670 | ||
2671 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 2672 | break; |
ebfd146a IR |
2673 | } |
2674 | ||
4a00c761 JJ |
2675 | VEC_free (tree, heap, vec_oprnds0); |
2676 | VEC_free (tree, heap, vec_oprnds1); | |
2677 | VEC_free (tree, heap, vec_dsts); | |
2678 | VEC_free (tree, heap, interm_types); | |
ebfd146a IR |
2679 | |
2680 | return true; | |
2681 | } | |
ff802fa1 IR |
2682 | |
2683 | ||
ebfd146a IR |
2684 | /* Function vectorizable_assignment. |
2685 | ||
b8698a0f L |
2686 | Check if STMT performs an assignment (copy) that can be vectorized. |
2687 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
2688 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2689 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2690 | ||
2691 | static bool | |
2692 | vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi, | |
2693 | gimple *vec_stmt, slp_tree slp_node) | |
2694 | { | |
2695 | tree vec_dest; | |
2696 | tree scalar_dest; | |
2697 | tree op; | |
2698 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2699 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2700 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2701 | tree new_temp; | |
2702 | tree def; | |
2703 | gimple def_stmt; | |
2704 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
fde9c428 | 2705 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
ebfd146a | 2706 | int ncopies; |
f18b55bd | 2707 | int i, j; |
ebfd146a IR |
2708 | VEC(tree,heap) *vec_oprnds = NULL; |
2709 | tree vop; | |
a70d6342 | 2710 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
f18b55bd IR |
2711 | gimple new_stmt = NULL; |
2712 | stmt_vec_info prev_stmt_info = NULL; | |
fde9c428 RG |
2713 | enum tree_code code; |
2714 | tree vectype_in; | |
ebfd146a IR |
2715 | |
2716 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2717 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2718 | case of SLP. */ | |
437f4a00 | 2719 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
2720 | ncopies = 1; |
2721 | else | |
2722 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2723 | ||
2724 | gcc_assert (ncopies >= 1); | |
ebfd146a | 2725 | |
a70d6342 | 2726 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2727 | return false; |
2728 | ||
8644a673 | 2729 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2730 | return false; |
2731 | ||
2732 | /* Is vectorizable assignment? */ | |
2733 | if (!is_gimple_assign (stmt)) | |
2734 | return false; | |
2735 | ||
2736 | scalar_dest = gimple_assign_lhs (stmt); | |
2737 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
2738 | return false; | |
2739 | ||
fde9c428 | 2740 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 2741 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
2742 | || code == PAREN_EXPR |
2743 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
2744 | op = gimple_assign_rhs1 (stmt); |
2745 | else | |
2746 | return false; | |
2747 | ||
7b7ec6c5 RG |
2748 | if (code == VIEW_CONVERT_EXPR) |
2749 | op = TREE_OPERAND (op, 0); | |
2750 | ||
fde9c428 RG |
2751 | if (!vect_is_simple_use_1 (op, loop_vinfo, bb_vinfo, |
2752 | &def_stmt, &def, &dt[0], &vectype_in)) | |
ebfd146a IR |
2753 | { |
2754 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2755 | fprintf (vect_dump, "use not simple."); | |
2756 | return false; | |
2757 | } | |
2758 | ||
fde9c428 RG |
2759 | /* We can handle NOP_EXPR conversions that do not change the number |
2760 | of elements or the vector size. */ | |
7b7ec6c5 RG |
2761 | if ((CONVERT_EXPR_CODE_P (code) |
2762 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
2763 | && (!vectype_in |
2764 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
2765 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
2766 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
2767 | return false; | |
2768 | ||
7b7b1813 RG |
2769 | /* We do not handle bit-precision changes. */ |
2770 | if ((CONVERT_EXPR_CODE_P (code) | |
2771 | || code == VIEW_CONVERT_EXPR) | |
2772 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2773 | && ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
2774 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
2775 | || ((TYPE_PRECISION (TREE_TYPE (op)) | |
2776 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))) | |
2777 | /* But a conversion that does not change the bit-pattern is ok. */ | |
2778 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
2779 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2780 | && TYPE_UNSIGNED (TREE_TYPE (op)))) | |
2781 | { | |
2782 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2783 | fprintf (vect_dump, "type conversion to/from bit-precision " | |
2784 | "unsupported."); | |
2785 | return false; | |
2786 | } | |
2787 | ||
ebfd146a IR |
2788 | if (!vec_stmt) /* transformation not required. */ |
2789 | { | |
2790 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
2791 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2792 | fprintf (vect_dump, "=== vectorizable_assignment ==="); | |
2793 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2794 | return true; | |
2795 | } | |
2796 | ||
2797 | /** Transform. **/ | |
2798 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2799 | fprintf (vect_dump, "transform assignment."); | |
2800 | ||
2801 | /* Handle def. */ | |
2802 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2803 | ||
2804 | /* Handle use. */ | |
f18b55bd | 2805 | for (j = 0; j < ncopies; j++) |
ebfd146a | 2806 | { |
f18b55bd IR |
2807 | /* Handle uses. */ |
2808 | if (j == 0) | |
d092494c | 2809 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node, -1); |
f18b55bd IR |
2810 | else |
2811 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
2812 | ||
2813 | /* Arguments are ready. create the new vector stmt. */ | |
ac47786e | 2814 | FOR_EACH_VEC_ELT (tree, vec_oprnds, i, vop) |
f18b55bd | 2815 | { |
7b7ec6c5 RG |
2816 | if (CONVERT_EXPR_CODE_P (code) |
2817 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 2818 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
2819 | new_stmt = gimple_build_assign (vec_dest, vop); |
2820 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2821 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2822 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2823 | if (slp_node) | |
2824 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2825 | } | |
ebfd146a IR |
2826 | |
2827 | if (slp_node) | |
f18b55bd IR |
2828 | continue; |
2829 | ||
2830 | if (j == 0) | |
2831 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2832 | else | |
2833 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2834 | ||
2835 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2836 | } | |
b8698a0f L |
2837 | |
2838 | VEC_free (tree, heap, vec_oprnds); | |
ebfd146a IR |
2839 | return true; |
2840 | } | |
2841 | ||
9dc3f7de | 2842 | |
1107f3ae IR |
2843 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
2844 | either as shift by a scalar or by a vector. */ | |
2845 | ||
2846 | bool | |
2847 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
2848 | { | |
2849 | ||
2850 | enum machine_mode vec_mode; | |
2851 | optab optab; | |
2852 | int icode; | |
2853 | tree vectype; | |
2854 | ||
2855 | vectype = get_vectype_for_scalar_type (scalar_type); | |
2856 | if (!vectype) | |
2857 | return false; | |
2858 | ||
2859 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
2860 | if (!optab | |
2861 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
2862 | { | |
2863 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
2864 | if (!optab | |
2865 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
2866 | == CODE_FOR_nothing)) | |
2867 | return false; | |
2868 | } | |
2869 | ||
2870 | vec_mode = TYPE_MODE (vectype); | |
2871 | icode = (int) optab_handler (optab, vec_mode); | |
2872 | if (icode == CODE_FOR_nothing) | |
2873 | return false; | |
2874 | ||
2875 | return true; | |
2876 | } | |
2877 | ||
2878 | ||
9dc3f7de IR |
2879 | /* Function vectorizable_shift. |
2880 | ||
2881 | Check if STMT performs a shift operation that can be vectorized. | |
2882 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2883 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2884 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2885 | ||
2886 | static bool | |
2887 | vectorizable_shift (gimple stmt, gimple_stmt_iterator *gsi, | |
2888 | gimple *vec_stmt, slp_tree slp_node) | |
2889 | { | |
2890 | tree vec_dest; | |
2891 | tree scalar_dest; | |
2892 | tree op0, op1 = NULL; | |
2893 | tree vec_oprnd1 = NULL_TREE; | |
2894 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2895 | tree vectype; | |
2896 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2897 | enum tree_code code; | |
2898 | enum machine_mode vec_mode; | |
2899 | tree new_temp; | |
2900 | optab optab; | |
2901 | int icode; | |
2902 | enum machine_mode optab_op2_mode; | |
2903 | tree def; | |
2904 | gimple def_stmt; | |
2905 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2906 | gimple new_stmt = NULL; | |
2907 | stmt_vec_info prev_stmt_info; | |
2908 | int nunits_in; | |
2909 | int nunits_out; | |
2910 | tree vectype_out; | |
cede2577 | 2911 | tree op1_vectype; |
9dc3f7de IR |
2912 | int ncopies; |
2913 | int j, i; | |
2914 | VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; | |
2915 | tree vop0, vop1; | |
2916 | unsigned int k; | |
49eab32e | 2917 | bool scalar_shift_arg = true; |
9dc3f7de IR |
2918 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2919 | int vf; | |
2920 | ||
2921 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
2922 | return false; | |
2923 | ||
2924 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
2925 | return false; | |
2926 | ||
2927 | /* Is STMT a vectorizable binary/unary operation? */ | |
2928 | if (!is_gimple_assign (stmt)) | |
2929 | return false; | |
2930 | ||
2931 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2932 | return false; | |
2933 | ||
2934 | code = gimple_assign_rhs_code (stmt); | |
2935 | ||
2936 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
2937 | || code == RROTATE_EXPR)) | |
2938 | return false; | |
2939 | ||
2940 | scalar_dest = gimple_assign_lhs (stmt); | |
2941 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
7b7b1813 RG |
2942 | if (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
2943 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
2944 | { | |
2945 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2946 | fprintf (vect_dump, "bit-precision shifts not supported."); | |
2947 | return false; | |
2948 | } | |
9dc3f7de IR |
2949 | |
2950 | op0 = gimple_assign_rhs1 (stmt); | |
2951 | if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo, | |
2952 | &def_stmt, &def, &dt[0], &vectype)) | |
2953 | { | |
2954 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2955 | fprintf (vect_dump, "use not simple."); | |
2956 | return false; | |
2957 | } | |
2958 | /* If op0 is an external or constant def use a vector type with | |
2959 | the same size as the output vector type. */ | |
2960 | if (!vectype) | |
2961 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
2962 | if (vec_stmt) | |
2963 | gcc_assert (vectype); | |
2964 | if (!vectype) | |
2965 | { | |
2966 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2967 | { | |
2968 | fprintf (vect_dump, "no vectype for scalar type "); | |
2969 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
2970 | } | |
2971 | ||
2972 | return false; | |
2973 | } | |
2974 | ||
2975 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
2976 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
2977 | if (nunits_out != nunits_in) | |
2978 | return false; | |
2979 | ||
2980 | op1 = gimple_assign_rhs2 (stmt); | |
cede2577 JJ |
2981 | if (!vect_is_simple_use_1 (op1, loop_vinfo, bb_vinfo, &def_stmt, &def, |
2982 | &dt[1], &op1_vectype)) | |
9dc3f7de IR |
2983 | { |
2984 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2985 | fprintf (vect_dump, "use not simple."); | |
2986 | return false; | |
2987 | } | |
2988 | ||
2989 | if (loop_vinfo) | |
2990 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
2991 | else | |
2992 | vf = 1; | |
2993 | ||
2994 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2995 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2996 | case of SLP. */ | |
437f4a00 | 2997 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
9dc3f7de IR |
2998 | ncopies = 1; |
2999 | else | |
3000 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
3001 | ||
3002 | gcc_assert (ncopies >= 1); | |
3003 | ||
3004 | /* Determine whether the shift amount is a vector, or scalar. If the | |
3005 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
3006 | ||
49eab32e JJ |
3007 | if (dt[1] == vect_internal_def && !slp_node) |
3008 | scalar_shift_arg = false; | |
3009 | else if (dt[1] == vect_constant_def | |
3010 | || dt[1] == vect_external_def | |
3011 | || dt[1] == vect_internal_def) | |
3012 | { | |
3013 | /* In SLP, need to check whether the shift count is the same, | |
3014 | in loops if it is a constant or invariant, it is always | |
3015 | a scalar shift. */ | |
3016 | if (slp_node) | |
3017 | { | |
3018 | VEC (gimple, heap) *stmts = SLP_TREE_SCALAR_STMTS (slp_node); | |
3019 | gimple slpstmt; | |
3020 | ||
3021 | FOR_EACH_VEC_ELT (gimple, stmts, k, slpstmt) | |
3022 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) | |
3023 | scalar_shift_arg = false; | |
3024 | } | |
3025 | } | |
3026 | else | |
3027 | { | |
3028 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3029 | fprintf (vect_dump, "operand mode requires invariant argument."); | |
3030 | return false; | |
3031 | } | |
3032 | ||
9dc3f7de | 3033 | /* Vector shifted by vector. */ |
49eab32e | 3034 | if (!scalar_shift_arg) |
9dc3f7de IR |
3035 | { |
3036 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
3037 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3038 | fprintf (vect_dump, "vector/vector shift/rotate found."); | |
aa948027 JJ |
3039 | if (!op1_vectype) |
3040 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
3041 | if (op1_vectype == NULL_TREE | |
3042 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 JJ |
3043 | { |
3044 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3045 | fprintf (vect_dump, "unusable type for last operand in" | |
3046 | " vector/vector shift/rotate."); | |
3047 | return false; | |
3048 | } | |
9dc3f7de IR |
3049 | } |
3050 | /* See if the machine has a vector shifted by scalar insn and if not | |
3051 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 3052 | else |
9dc3f7de IR |
3053 | { |
3054 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
3055 | if (optab | |
3056 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
3057 | { | |
9dc3f7de IR |
3058 | if (vect_print_dump_info (REPORT_DETAILS)) |
3059 | fprintf (vect_dump, "vector/scalar shift/rotate found."); | |
3060 | } | |
3061 | else | |
3062 | { | |
3063 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
3064 | if (optab | |
3065 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
3066 | != CODE_FOR_nothing)) | |
3067 | { | |
49eab32e JJ |
3068 | scalar_shift_arg = false; |
3069 | ||
9dc3f7de IR |
3070 | if (vect_print_dump_info (REPORT_DETAILS)) |
3071 | fprintf (vect_dump, "vector/vector shift/rotate found."); | |
3072 | ||
3073 | /* Unlike the other binary operators, shifts/rotates have | |
3074 | the rhs being int, instead of the same type as the lhs, | |
3075 | so make sure the scalar is the right type if we are | |
aa948027 | 3076 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
3077 | if (dt[1] == vect_constant_def) |
3078 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
3079 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
3080 | TREE_TYPE (op1))) | |
3081 | { | |
3082 | if (slp_node | |
3083 | && TYPE_MODE (TREE_TYPE (vectype)) | |
3084 | != TYPE_MODE (TREE_TYPE (op1))) | |
3085 | { | |
3086 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3087 | fprintf (vect_dump, "unusable type for last operand in" | |
3088 | " vector/vector shift/rotate."); | |
3089 | return false; | |
3090 | } | |
3091 | if (vec_stmt && !slp_node) | |
3092 | { | |
3093 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
3094 | op1 = vect_init_vector (stmt, op1, | |
3095 | TREE_TYPE (vectype), NULL); | |
3096 | } | |
3097 | } | |
9dc3f7de IR |
3098 | } |
3099 | } | |
3100 | } | |
9dc3f7de IR |
3101 | |
3102 | /* Supportable by target? */ | |
3103 | if (!optab) | |
3104 | { | |
3105 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3106 | fprintf (vect_dump, "no optab."); | |
3107 | return false; | |
3108 | } | |
3109 | vec_mode = TYPE_MODE (vectype); | |
3110 | icode = (int) optab_handler (optab, vec_mode); | |
3111 | if (icode == CODE_FOR_nothing) | |
3112 | { | |
3113 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3114 | fprintf (vect_dump, "op not supported by target."); | |
3115 | /* Check only during analysis. */ | |
3116 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
3117 | || (vf < vect_min_worthwhile_factor (code) | |
3118 | && !vec_stmt)) | |
3119 | return false; | |
3120 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3121 | fprintf (vect_dump, "proceeding using word mode."); | |
3122 | } | |
3123 | ||
3124 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
3125 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
3126 | && vf < vect_min_worthwhile_factor (code) | |
3127 | && !vec_stmt) | |
3128 | { | |
3129 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3130 | fprintf (vect_dump, "not worthwhile without SIMD support."); | |
3131 | return false; | |
3132 | } | |
3133 | ||
3134 | if (!vec_stmt) /* transformation not required. */ | |
3135 | { | |
3136 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
3137 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3138 | fprintf (vect_dump, "=== vectorizable_shift ==="); | |
3139 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
3140 | return true; | |
3141 | } | |
3142 | ||
3143 | /** Transform. **/ | |
3144 | ||
3145 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3146 | fprintf (vect_dump, "transform binary/unary operation."); | |
3147 | ||
3148 | /* Handle def. */ | |
3149 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3150 | ||
3151 | /* Allocate VECs for vector operands. In case of SLP, vector operands are | |
3152 | created in the previous stages of the recursion, so no allocation is | |
3153 | needed, except for the case of shift with scalar shift argument. In that | |
3154 | case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to | |
3155 | be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE. | |
3156 | In case of loop-based vectorization we allocate VECs of size 1. We | |
3157 | allocate VEC_OPRNDS1 only in case of binary operation. */ | |
3158 | if (!slp_node) | |
3159 | { | |
3160 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
3161 | vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
3162 | } | |
3163 | else if (scalar_shift_arg) | |
3164 | vec_oprnds1 = VEC_alloc (tree, heap, slp_node->vec_stmts_size); | |
3165 | ||
3166 | prev_stmt_info = NULL; | |
3167 | for (j = 0; j < ncopies; j++) | |
3168 | { | |
3169 | /* Handle uses. */ | |
3170 | if (j == 0) | |
3171 | { | |
3172 | if (scalar_shift_arg) | |
3173 | { | |
3174 | /* Vector shl and shr insn patterns can be defined with scalar | |
3175 | operand 2 (shift operand). In this case, use constant or loop | |
3176 | invariant op1 directly, without extending it to vector mode | |
3177 | first. */ | |
3178 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
3179 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
3180 | { | |
3181 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3182 | fprintf (vect_dump, "operand 1 using scalar mode."); | |
3183 | vec_oprnd1 = op1; | |
3184 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
3185 | if (slp_node) | |
3186 | { | |
3187 | /* Store vec_oprnd1 for every vector stmt to be created | |
3188 | for SLP_NODE. We check during the analysis that all | |
3189 | the shift arguments are the same. | |
3190 | TODO: Allow different constants for different vector | |
3191 | stmts generated for an SLP instance. */ | |
3192 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
3193 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
3194 | } | |
3195 | } | |
3196 | } | |
3197 | ||
3198 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
3199 | (a special case for certain kind of vector shifts); otherwise, | |
3200 | operand 1 should be of a vector type (the usual case). */ | |
3201 | if (vec_oprnd1) | |
3202 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
d092494c | 3203 | slp_node, -1); |
9dc3f7de IR |
3204 | else |
3205 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
d092494c | 3206 | slp_node, -1); |
9dc3f7de IR |
3207 | } |
3208 | else | |
3209 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
3210 | ||
3211 | /* Arguments are ready. Create the new vector stmt. */ | |
3212 | FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0) | |
3213 | { | |
3214 | vop1 = VEC_index (tree, vec_oprnds1, i); | |
3215 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); | |
3216 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3217 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3218 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3219 | if (slp_node) | |
3220 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
3221 | } | |
3222 | ||
3223 | if (slp_node) | |
3224 | continue; | |
3225 | ||
3226 | if (j == 0) | |
3227 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3228 | else | |
3229 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3230 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3231 | } | |
3232 | ||
3233 | VEC_free (tree, heap, vec_oprnds0); | |
3234 | VEC_free (tree, heap, vec_oprnds1); | |
3235 | ||
3236 | return true; | |
3237 | } | |
3238 | ||
3239 | ||
ebfd146a IR |
3240 | /* Function vectorizable_operation. |
3241 | ||
16949072 RG |
3242 | Check if STMT performs a binary, unary or ternary operation that can |
3243 | be vectorized. | |
b8698a0f | 3244 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
3245 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3246 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3247 | ||
3248 | static bool | |
3249 | vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi, | |
3250 | gimple *vec_stmt, slp_tree slp_node) | |
3251 | { | |
3252 | tree vec_dest; | |
3253 | tree scalar_dest; | |
16949072 | 3254 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 3255 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
b690cc0f | 3256 | tree vectype; |
ebfd146a IR |
3257 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
3258 | enum tree_code code; | |
3259 | enum machine_mode vec_mode; | |
3260 | tree new_temp; | |
3261 | int op_type; | |
3262 | optab optab; | |
3263 | int icode; | |
ebfd146a IR |
3264 | tree def; |
3265 | gimple def_stmt; | |
16949072 RG |
3266 | enum vect_def_type dt[3] |
3267 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
ebfd146a IR |
3268 | gimple new_stmt = NULL; |
3269 | stmt_vec_info prev_stmt_info; | |
b690cc0f | 3270 | int nunits_in; |
ebfd146a IR |
3271 | int nunits_out; |
3272 | tree vectype_out; | |
3273 | int ncopies; | |
3274 | int j, i; | |
16949072 RG |
3275 | VEC(tree,heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL, *vec_oprnds2 = NULL; |
3276 | tree vop0, vop1, vop2; | |
a70d6342 IR |
3277 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3278 | int vf; | |
3279 | ||
a70d6342 | 3280 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3281 | return false; |
3282 | ||
8644a673 | 3283 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3284 | return false; |
3285 | ||
3286 | /* Is STMT a vectorizable binary/unary operation? */ | |
3287 | if (!is_gimple_assign (stmt)) | |
3288 | return false; | |
3289 | ||
3290 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3291 | return false; | |
3292 | ||
ebfd146a IR |
3293 | code = gimple_assign_rhs_code (stmt); |
3294 | ||
3295 | /* For pointer addition, we should use the normal plus for | |
3296 | the vector addition. */ | |
3297 | if (code == POINTER_PLUS_EXPR) | |
3298 | code = PLUS_EXPR; | |
3299 | ||
3300 | /* Support only unary or binary operations. */ | |
3301 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 3302 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a IR |
3303 | { |
3304 | if (vect_print_dump_info (REPORT_DETAILS)) | |
16949072 RG |
3305 | fprintf (vect_dump, "num. args = %d (not unary/binary/ternary op).", |
3306 | op_type); | |
ebfd146a IR |
3307 | return false; |
3308 | } | |
3309 | ||
b690cc0f RG |
3310 | scalar_dest = gimple_assign_lhs (stmt); |
3311 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
3312 | ||
7b7b1813 RG |
3313 | /* Most operations cannot handle bit-precision types without extra |
3314 | truncations. */ | |
3315 | if ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
3316 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
3317 | /* Exception are bitwise binary operations. */ | |
3318 | && code != BIT_IOR_EXPR | |
3319 | && code != BIT_XOR_EXPR | |
3320 | && code != BIT_AND_EXPR) | |
3321 | { | |
3322 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3323 | fprintf (vect_dump, "bit-precision arithmetic not supported."); | |
3324 | return false; | |
3325 | } | |
3326 | ||
ebfd146a | 3327 | op0 = gimple_assign_rhs1 (stmt); |
b690cc0f RG |
3328 | if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo, |
3329 | &def_stmt, &def, &dt[0], &vectype)) | |
ebfd146a IR |
3330 | { |
3331 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3332 | fprintf (vect_dump, "use not simple."); | |
3333 | return false; | |
3334 | } | |
b690cc0f RG |
3335 | /* If op0 is an external or constant def use a vector type with |
3336 | the same size as the output vector type. */ | |
3337 | if (!vectype) | |
3338 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
3339 | if (vec_stmt) |
3340 | gcc_assert (vectype); | |
3341 | if (!vectype) | |
3342 | { | |
3343 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3344 | { | |
3345 | fprintf (vect_dump, "no vectype for scalar type "); | |
3346 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
3347 | } | |
3348 | ||
3349 | return false; | |
3350 | } | |
b690cc0f RG |
3351 | |
3352 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
3353 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
3354 | if (nunits_out != nunits_in) | |
3355 | return false; | |
ebfd146a | 3356 | |
16949072 | 3357 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
3358 | { |
3359 | op1 = gimple_assign_rhs2 (stmt); | |
b8698a0f | 3360 | if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def, |
a70d6342 | 3361 | &dt[1])) |
ebfd146a IR |
3362 | { |
3363 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3364 | fprintf (vect_dump, "use not simple."); | |
3365 | return false; | |
3366 | } | |
3367 | } | |
16949072 RG |
3368 | if (op_type == ternary_op) |
3369 | { | |
3370 | op2 = gimple_assign_rhs3 (stmt); | |
3371 | if (!vect_is_simple_use (op2, loop_vinfo, bb_vinfo, &def_stmt, &def, | |
3372 | &dt[2])) | |
3373 | { | |
3374 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3375 | fprintf (vect_dump, "use not simple."); | |
3376 | return false; | |
3377 | } | |
3378 | } | |
ebfd146a | 3379 | |
b690cc0f RG |
3380 | if (loop_vinfo) |
3381 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3382 | else | |
3383 | vf = 1; | |
3384 | ||
3385 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 3386 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 3387 | case of SLP. */ |
437f4a00 | 3388 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
b690cc0f RG |
3389 | ncopies = 1; |
3390 | else | |
3391 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
3392 | ||
3393 | gcc_assert (ncopies >= 1); | |
3394 | ||
9dc3f7de | 3395 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
3396 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
3397 | || code == RROTATE_EXPR) | |
9dc3f7de | 3398 | return false; |
ebfd146a | 3399 | |
16949072 | 3400 | optab = optab_for_tree_code (code, vectype, optab_default); |
ebfd146a IR |
3401 | |
3402 | /* Supportable by target? */ | |
3403 | if (!optab) | |
3404 | { | |
3405 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3406 | fprintf (vect_dump, "no optab."); | |
3407 | return false; | |
3408 | } | |
3409 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 3410 | icode = (int) optab_handler (optab, vec_mode); |
ebfd146a IR |
3411 | if (icode == CODE_FOR_nothing) |
3412 | { | |
3413 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3414 | fprintf (vect_dump, "op not supported by target."); | |
3415 | /* Check only during analysis. */ | |
3416 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
a70d6342 | 3417 | || (vf < vect_min_worthwhile_factor (code) |
ebfd146a IR |
3418 | && !vec_stmt)) |
3419 | return false; | |
3420 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4a00c761 | 3421 | fprintf (vect_dump, "proceeding using word mode."); |
383d9c83 IR |
3422 | } |
3423 | ||
4a00c761 JJ |
3424 | /* Worthwhile without SIMD support? Check only during analysis. */ |
3425 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
3426 | && vf < vect_min_worthwhile_factor (code) | |
3427 | && !vec_stmt) | |
7d8930a0 IR |
3428 | { |
3429 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4a00c761 | 3430 | fprintf (vect_dump, "not worthwhile without SIMD support."); |
7d8930a0 IR |
3431 | return false; |
3432 | } | |
ebfd146a | 3433 | |
ebfd146a IR |
3434 | if (!vec_stmt) /* transformation not required. */ |
3435 | { | |
4a00c761 | 3436 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
ebfd146a | 3437 | if (vect_print_dump_info (REPORT_DETAILS)) |
4a00c761 JJ |
3438 | fprintf (vect_dump, "=== vectorizable_operation ==="); |
3439 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
ebfd146a IR |
3440 | return true; |
3441 | } | |
3442 | ||
3443 | /** Transform. **/ | |
3444 | ||
3445 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4a00c761 | 3446 | fprintf (vect_dump, "transform binary/unary operation."); |
383d9c83 | 3447 | |
ebfd146a | 3448 | /* Handle def. */ |
4a00c761 | 3449 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 3450 | |
4a00c761 JJ |
3451 | /* Allocate VECs for vector operands. In case of SLP, vector operands are |
3452 | created in the previous stages of the recursion, so no allocation is | |
3453 | needed, except for the case of shift with scalar shift argument. In that | |
3454 | case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to | |
3455 | be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE. | |
3456 | In case of loop-based vectorization we allocate VECs of size 1. We | |
3457 | allocate VEC_OPRNDS1 only in case of binary operation. */ | |
ebfd146a IR |
3458 | if (!slp_node) |
3459 | { | |
4a00c761 JJ |
3460 | vec_oprnds0 = VEC_alloc (tree, heap, 1); |
3461 | if (op_type == binary_op || op_type == ternary_op) | |
ebfd146a | 3462 | vec_oprnds1 = VEC_alloc (tree, heap, 1); |
4a00c761 JJ |
3463 | if (op_type == ternary_op) |
3464 | vec_oprnds2 = VEC_alloc (tree, heap, 1); | |
ebfd146a IR |
3465 | } |
3466 | ||
3467 | /* In case the vectorization factor (VF) is bigger than the number | |
3468 | of elements that we can fit in a vectype (nunits), we have to generate | |
3469 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
3470 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
3471 | from one copy of the vector stmt to the next, in the field | |
3472 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
3473 | stages to find the correct vector defs to be used when vectorizing | |
3474 | stmts that use the defs of the current stmt. The example below | |
3475 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
3476 | we need to create 4 vectorized stmts): | |
3477 | ||
3478 | before vectorization: | |
3479 | RELATED_STMT VEC_STMT | |
3480 | S1: x = memref - - | |
3481 | S2: z = x + 1 - - | |
3482 | ||
3483 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
3484 | there): | |
3485 | RELATED_STMT VEC_STMT | |
3486 | VS1_0: vx0 = memref0 VS1_1 - | |
3487 | VS1_1: vx1 = memref1 VS1_2 - | |
3488 | VS1_2: vx2 = memref2 VS1_3 - | |
3489 | VS1_3: vx3 = memref3 - - | |
3490 | S1: x = load - VS1_0 | |
3491 | S2: z = x + 1 - - | |
3492 | ||
3493 | step2: vectorize stmt S2 (done here): | |
3494 | To vectorize stmt S2 we first need to find the relevant vector | |
3495 | def for the first operand 'x'. This is, as usual, obtained from | |
3496 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
3497 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
3498 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
3499 | the vector stmt VS2_0, and as usual, record it in the | |
3500 | STMT_VINFO_VEC_STMT of stmt S2. | |
3501 | When creating the second copy (VS2_1), we obtain the relevant vector | |
3502 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
3503 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
3504 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
3505 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
3506 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
3507 | chain of stmts and pointers: | |
3508 | RELATED_STMT VEC_STMT | |
3509 | VS1_0: vx0 = memref0 VS1_1 - | |
3510 | VS1_1: vx1 = memref1 VS1_2 - | |
3511 | VS1_2: vx2 = memref2 VS1_3 - | |
3512 | VS1_3: vx3 = memref3 - - | |
3513 | S1: x = load - VS1_0 | |
3514 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
3515 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
3516 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
3517 | VS2_3: vz3 = vx3 + v1 - - | |
3518 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
3519 | |
3520 | prev_stmt_info = NULL; | |
3521 | for (j = 0; j < ncopies; j++) | |
3522 | { | |
3523 | /* Handle uses. */ | |
3524 | if (j == 0) | |
4a00c761 JJ |
3525 | { |
3526 | if (op_type == binary_op || op_type == ternary_op) | |
3527 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
3528 | slp_node, -1); | |
3529 | else | |
3530 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
3531 | slp_node, -1); | |
3532 | if (op_type == ternary_op) | |
36ba4aae | 3533 | { |
4a00c761 JJ |
3534 | vec_oprnds2 = VEC_alloc (tree, heap, 1); |
3535 | VEC_quick_push (tree, vec_oprnds2, | |
3536 | vect_get_vec_def_for_operand (op2, stmt, NULL)); | |
36ba4aae | 3537 | } |
4a00c761 | 3538 | } |
ebfd146a | 3539 | else |
4a00c761 JJ |
3540 | { |
3541 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
3542 | if (op_type == ternary_op) | |
3543 | { | |
3544 | tree vec_oprnd = VEC_pop (tree, vec_oprnds2); | |
3545 | VEC_quick_push (tree, vec_oprnds2, | |
3546 | vect_get_vec_def_for_stmt_copy (dt[2], | |
3547 | vec_oprnd)); | |
3548 | } | |
3549 | } | |
3550 | ||
3551 | /* Arguments are ready. Create the new vector stmt. */ | |
3552 | FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0) | |
ebfd146a | 3553 | { |
4a00c761 JJ |
3554 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
3555 | ? VEC_index (tree, vec_oprnds1, i) : NULL_TREE); | |
3556 | vop2 = ((op_type == ternary_op) | |
3557 | ? VEC_index (tree, vec_oprnds2, i) : NULL_TREE); | |
3558 | new_stmt = gimple_build_assign_with_ops3 (code, vec_dest, | |
3559 | vop0, vop1, vop2); | |
3560 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3561 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3562 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3563 | if (slp_node) | |
3564 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
ebfd146a IR |
3565 | } |
3566 | ||
4a00c761 JJ |
3567 | if (slp_node) |
3568 | continue; | |
3569 | ||
3570 | if (j == 0) | |
3571 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3572 | else | |
3573 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3574 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
3575 | } |
3576 | ||
ebfd146a | 3577 | VEC_free (tree, heap, vec_oprnds0); |
4a00c761 JJ |
3578 | if (vec_oprnds1) |
3579 | VEC_free (tree, heap, vec_oprnds1); | |
3580 | if (vec_oprnds2) | |
3581 | VEC_free (tree, heap, vec_oprnds2); | |
ebfd146a | 3582 | |
ebfd146a IR |
3583 | return true; |
3584 | } | |
3585 | ||
3586 | ||
3587 | /* Function vectorizable_store. | |
3588 | ||
b8698a0f L |
3589 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
3590 | can be vectorized. | |
3591 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
3592 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3593 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3594 | ||
3595 | static bool | |
3596 | vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
3597 | slp_tree slp_node) | |
3598 | { | |
3599 | tree scalar_dest; | |
3600 | tree data_ref; | |
3601 | tree op; | |
3602 | tree vec_oprnd = NULL_TREE; | |
3603 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3604 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
3605 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
272c6793 | 3606 | tree elem_type; |
ebfd146a | 3607 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 3608 | struct loop *loop = NULL; |
ebfd146a IR |
3609 | enum machine_mode vec_mode; |
3610 | tree dummy; | |
3611 | enum dr_alignment_support alignment_support_scheme; | |
3612 | tree def; | |
3613 | gimple def_stmt; | |
3614 | enum vect_def_type dt; | |
3615 | stmt_vec_info prev_stmt_info = NULL; | |
3616 | tree dataref_ptr = NULL_TREE; | |
3617 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
3618 | int ncopies; | |
3619 | int j; | |
3620 | gimple next_stmt, first_stmt = NULL; | |
3621 | bool strided_store = false; | |
272c6793 | 3622 | bool store_lanes_p = false; |
ebfd146a IR |
3623 | unsigned int group_size, i; |
3624 | VEC(tree,heap) *dr_chain = NULL, *oprnds = NULL, *result_chain = NULL; | |
3625 | bool inv_p; | |
3626 | VEC(tree,heap) *vec_oprnds = NULL; | |
3627 | bool slp = (slp_node != NULL); | |
ebfd146a | 3628 | unsigned int vec_num; |
a70d6342 | 3629 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
272c6793 | 3630 | tree aggr_type; |
a70d6342 IR |
3631 | |
3632 | if (loop_vinfo) | |
3633 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
3634 | |
3635 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3636 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3637 | case of SLP. */ | |
437f4a00 | 3638 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
3639 | ncopies = 1; |
3640 | else | |
3641 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3642 | ||
3643 | gcc_assert (ncopies >= 1); | |
3644 | ||
3645 | /* FORNOW. This restriction should be relaxed. */ | |
a70d6342 | 3646 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) |
ebfd146a IR |
3647 | { |
3648 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3649 | fprintf (vect_dump, "multiple types in nested loop."); | |
3650 | return false; | |
3651 | } | |
3652 | ||
a70d6342 | 3653 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3654 | return false; |
3655 | ||
8644a673 | 3656 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3657 | return false; |
3658 | ||
3659 | /* Is vectorizable store? */ | |
3660 | ||
3661 | if (!is_gimple_assign (stmt)) | |
3662 | return false; | |
3663 | ||
3664 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
3665 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
3666 | && is_pattern_stmt_p (stmt_info)) | |
3667 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a IR |
3668 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
3669 | && TREE_CODE (scalar_dest) != INDIRECT_REF | |
e9dbe7bb IR |
3670 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
3671 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
3672 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
3673 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
3674 | return false; |
3675 | ||
3676 | gcc_assert (gimple_assign_single_p (stmt)); | |
3677 | op = gimple_assign_rhs1 (stmt); | |
a70d6342 | 3678 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt)) |
ebfd146a IR |
3679 | { |
3680 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3681 | fprintf (vect_dump, "use not simple."); | |
3682 | return false; | |
3683 | } | |
3684 | ||
272c6793 | 3685 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 3686 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 3687 | |
ebfd146a IR |
3688 | /* FORNOW. In some cases can vectorize even if data-type not supported |
3689 | (e.g. - array initialization with 0). */ | |
947131ba | 3690 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
3691 | return false; |
3692 | ||
3693 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
3694 | return false; | |
3695 | ||
a1e53f3f L |
3696 | if (tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0) |
3697 | { | |
3698 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3699 | fprintf (vect_dump, "negative step for store."); | |
3700 | return false; | |
3701 | } | |
3702 | ||
ebfd146a IR |
3703 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) |
3704 | { | |
3705 | strided_store = true; | |
e14c1050 | 3706 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
3707 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
3708 | { | |
e14c1050 | 3709 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
3710 | if (vect_store_lanes_supported (vectype, group_size)) |
3711 | store_lanes_p = true; | |
3712 | else if (!vect_strided_store_supported (vectype, group_size)) | |
b602d918 RS |
3713 | return false; |
3714 | } | |
b8698a0f | 3715 | |
ebfd146a IR |
3716 | if (first_stmt == stmt) |
3717 | { | |
3718 | /* STMT is the leader of the group. Check the operands of all the | |
3719 | stmts of the group. */ | |
e14c1050 | 3720 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
3721 | while (next_stmt) |
3722 | { | |
3723 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
3724 | op = gimple_assign_rhs1 (next_stmt); | |
b8698a0f | 3725 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, |
a70d6342 | 3726 | &def, &dt)) |
ebfd146a IR |
3727 | { |
3728 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3729 | fprintf (vect_dump, "use not simple."); | |
3730 | return false; | |
3731 | } | |
e14c1050 | 3732 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
3733 | } |
3734 | } | |
3735 | } | |
3736 | ||
3737 | if (!vec_stmt) /* transformation not required. */ | |
3738 | { | |
3739 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
272c6793 | 3740 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, NULL); |
ebfd146a IR |
3741 | return true; |
3742 | } | |
3743 | ||
3744 | /** Transform. **/ | |
3745 | ||
3746 | if (strided_store) | |
3747 | { | |
3748 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 3749 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 3750 | |
e14c1050 | 3751 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
3752 | |
3753 | /* FORNOW */ | |
a70d6342 | 3754 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
3755 | |
3756 | /* We vectorize all the stmts of the interleaving group when we | |
3757 | reach the last stmt in the group. */ | |
e14c1050 IR |
3758 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
3759 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
3760 | && !slp) |
3761 | { | |
3762 | *vec_stmt = NULL; | |
3763 | return true; | |
3764 | } | |
3765 | ||
3766 | if (slp) | |
4b5caab7 IR |
3767 | { |
3768 | strided_store = false; | |
3769 | /* VEC_NUM is the number of vect stmts to be created for this | |
3770 | group. */ | |
3771 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
3772 | first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0); | |
3773 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
d092494c | 3774 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 3775 | } |
ebfd146a | 3776 | else |
4b5caab7 IR |
3777 | /* VEC_NUM is the number of vect stmts to be created for this |
3778 | group. */ | |
ebfd146a IR |
3779 | vec_num = group_size; |
3780 | } | |
b8698a0f | 3781 | else |
ebfd146a IR |
3782 | { |
3783 | first_stmt = stmt; | |
3784 | first_dr = dr; | |
3785 | group_size = vec_num = 1; | |
ebfd146a | 3786 | } |
b8698a0f | 3787 | |
ebfd146a IR |
3788 | if (vect_print_dump_info (REPORT_DETAILS)) |
3789 | fprintf (vect_dump, "transform store. ncopies = %d",ncopies); | |
3790 | ||
3791 | dr_chain = VEC_alloc (tree, heap, group_size); | |
3792 | oprnds = VEC_alloc (tree, heap, group_size); | |
3793 | ||
720f5239 | 3794 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 3795 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
3796 | /* Targets with store-lane instructions must not require explicit |
3797 | realignment. */ | |
3798 | gcc_assert (!store_lanes_p | |
3799 | || alignment_support_scheme == dr_aligned | |
3800 | || alignment_support_scheme == dr_unaligned_supported); | |
3801 | ||
3802 | if (store_lanes_p) | |
3803 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
3804 | else | |
3805 | aggr_type = vectype; | |
ebfd146a IR |
3806 | |
3807 | /* In case the vectorization factor (VF) is bigger than the number | |
3808 | of elements that we can fit in a vectype (nunits), we have to generate | |
3809 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 3810 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
3811 | vect_get_vec_def_for_copy_stmt. */ |
3812 | ||
3813 | /* In case of interleaving (non-unit strided access): | |
3814 | ||
3815 | S1: &base + 2 = x2 | |
3816 | S2: &base = x0 | |
3817 | S3: &base + 1 = x1 | |
3818 | S4: &base + 3 = x3 | |
3819 | ||
3820 | We create vectorized stores starting from base address (the access of the | |
3821 | first stmt in the chain (S2 in the above example), when the last store stmt | |
3822 | of the chain (S4) is reached: | |
3823 | ||
3824 | VS1: &base = vx2 | |
3825 | VS2: &base + vec_size*1 = vx0 | |
3826 | VS3: &base + vec_size*2 = vx1 | |
3827 | VS4: &base + vec_size*3 = vx3 | |
3828 | ||
3829 | Then permutation statements are generated: | |
3830 | ||
b826bea7 RH |
3831 | VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 > |
3832 | VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 > | |
ebfd146a | 3833 | ... |
b8698a0f | 3834 | |
ebfd146a IR |
3835 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
3836 | (the order of the data-refs in the output of vect_permute_store_chain | |
3837 | corresponds to the order of scalar stmts in the interleaving chain - see | |
3838 | the documentation of vect_permute_store_chain()). | |
3839 | ||
3840 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 3841 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 3842 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 3843 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
3844 | */ |
3845 | ||
3846 | prev_stmt_info = NULL; | |
3847 | for (j = 0; j < ncopies; j++) | |
3848 | { | |
3849 | gimple new_stmt; | |
3850 | gimple ptr_incr; | |
3851 | ||
3852 | if (j == 0) | |
3853 | { | |
3854 | if (slp) | |
3855 | { | |
3856 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c IR |
3857 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
3858 | NULL, slp_node, -1); | |
ebfd146a IR |
3859 | |
3860 | vec_oprnd = VEC_index (tree, vec_oprnds, 0); | |
3861 | } | |
3862 | else | |
3863 | { | |
b8698a0f L |
3864 | /* For interleaved stores we collect vectorized defs for all the |
3865 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
3866 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
3867 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
3868 | ||
3869 | If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and | |
3870 | OPRNDS are of size 1. */ | |
b8698a0f | 3871 | next_stmt = first_stmt; |
ebfd146a IR |
3872 | for (i = 0; i < group_size; i++) |
3873 | { | |
b8698a0f L |
3874 | /* Since gaps are not supported for interleaved stores, |
3875 | GROUP_SIZE is the exact number of stmts in the chain. | |
3876 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
3877 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
3878 | iteration of the loop will be executed. */ |
3879 | gcc_assert (next_stmt | |
3880 | && gimple_assign_single_p (next_stmt)); | |
3881 | op = gimple_assign_rhs1 (next_stmt); | |
3882 | ||
b8698a0f | 3883 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt, |
ebfd146a | 3884 | NULL); |
b8698a0f L |
3885 | VEC_quick_push(tree, dr_chain, vec_oprnd); |
3886 | VEC_quick_push(tree, oprnds, vec_oprnd); | |
e14c1050 | 3887 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
3888 | } |
3889 | } | |
3890 | ||
3891 | /* We should have catched mismatched types earlier. */ | |
3892 | gcc_assert (useless_type_conversion_p (vectype, | |
3893 | TREE_TYPE (vec_oprnd))); | |
272c6793 | 3894 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, aggr_type, NULL, |
920e8172 RS |
3895 | NULL_TREE, &dummy, gsi, |
3896 | &ptr_incr, false, &inv_p); | |
a70d6342 | 3897 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 3898 | } |
b8698a0f | 3899 | else |
ebfd146a | 3900 | { |
b8698a0f L |
3901 | /* For interleaved stores we created vectorized defs for all the |
3902 | defs stored in OPRNDS in the previous iteration (previous copy). | |
3903 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
3904 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
3905 | next copy. | |
3906 | If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and | |
3907 | OPRNDS are of size 1. */ | |
3908 | for (i = 0; i < group_size; i++) | |
3909 | { | |
3910 | op = VEC_index (tree, oprnds, i); | |
b8698a0f | 3911 | vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, |
a70d6342 | 3912 | &dt); |
b8698a0f | 3913 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
ebfd146a IR |
3914 | VEC_replace(tree, dr_chain, i, vec_oprnd); |
3915 | VEC_replace(tree, oprnds, i, vec_oprnd); | |
3916 | } | |
272c6793 RS |
3917 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
3918 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
3919 | } |
3920 | ||
272c6793 | 3921 | if (store_lanes_p) |
ebfd146a | 3922 | { |
272c6793 | 3923 | tree vec_array; |
267d3070 | 3924 | |
272c6793 RS |
3925 | /* Combine all the vectors into an array. */ |
3926 | vec_array = create_vector_array (vectype, vec_num); | |
3927 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 3928 | { |
272c6793 RS |
3929 | vec_oprnd = VEC_index (tree, dr_chain, i); |
3930 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); | |
267d3070 | 3931 | } |
b8698a0f | 3932 | |
272c6793 RS |
3933 | /* Emit: |
3934 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
3935 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
3936 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
3937 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 RS |
3938 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3939 | mark_symbols_for_renaming (new_stmt); | |
272c6793 RS |
3940 | } |
3941 | else | |
3942 | { | |
3943 | new_stmt = NULL; | |
3944 | if (strided_store) | |
3945 | { | |
3946 | result_chain = VEC_alloc (tree, heap, group_size); | |
3947 | /* Permute. */ | |
3948 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
3949 | &result_chain); | |
3950 | } | |
c2d7ab2a | 3951 | |
272c6793 RS |
3952 | next_stmt = first_stmt; |
3953 | for (i = 0; i < vec_num; i++) | |
3954 | { | |
3955 | struct ptr_info_def *pi; | |
3956 | ||
3957 | if (i > 0) | |
3958 | /* Bump the vector pointer. */ | |
3959 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
3960 | stmt, NULL_TREE); | |
3961 | ||
3962 | if (slp) | |
3963 | vec_oprnd = VEC_index (tree, vec_oprnds, i); | |
3964 | else if (strided_store) | |
3965 | /* For strided stores vectorized defs are interleaved in | |
3966 | vect_permute_store_chain(). */ | |
3967 | vec_oprnd = VEC_index (tree, result_chain, i); | |
3968 | ||
3969 | data_ref = build2 (MEM_REF, TREE_TYPE (vec_oprnd), dataref_ptr, | |
3970 | build_int_cst (reference_alias_ptr_type | |
3971 | (DR_REF (first_dr)), 0)); | |
3972 | pi = get_ptr_info (dataref_ptr); | |
3973 | pi->align = TYPE_ALIGN_UNIT (vectype); | |
3974 | if (aligned_access_p (first_dr)) | |
3975 | pi->misalign = 0; | |
3976 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
3977 | { | |
3978 | TREE_TYPE (data_ref) | |
3979 | = build_aligned_type (TREE_TYPE (data_ref), | |
3980 | TYPE_ALIGN (elem_type)); | |
3981 | pi->align = TYPE_ALIGN_UNIT (elem_type); | |
3982 | pi->misalign = 0; | |
3983 | } | |
3984 | else | |
3985 | { | |
3986 | TREE_TYPE (data_ref) | |
3987 | = build_aligned_type (TREE_TYPE (data_ref), | |
3988 | TYPE_ALIGN (elem_type)); | |
3989 | pi->misalign = DR_MISALIGNMENT (first_dr); | |
3990 | } | |
c2d7ab2a | 3991 | |
272c6793 RS |
3992 | /* Arguments are ready. Create the new vector stmt. */ |
3993 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
3994 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3995 | mark_symbols_for_renaming (new_stmt); | |
3996 | ||
3997 | if (slp) | |
3998 | continue; | |
3999 | ||
e14c1050 | 4000 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
4001 | if (!next_stmt) |
4002 | break; | |
4003 | } | |
ebfd146a | 4004 | } |
1da0876c RS |
4005 | if (!slp) |
4006 | { | |
4007 | if (j == 0) | |
4008 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4009 | else | |
4010 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4011 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4012 | } | |
ebfd146a IR |
4013 | } |
4014 | ||
b8698a0f L |
4015 | VEC_free (tree, heap, dr_chain); |
4016 | VEC_free (tree, heap, oprnds); | |
ebfd146a | 4017 | if (result_chain) |
b8698a0f | 4018 | VEC_free (tree, heap, result_chain); |
ff802fa1 IR |
4019 | if (vec_oprnds) |
4020 | VEC_free (tree, heap, vec_oprnds); | |
ebfd146a IR |
4021 | |
4022 | return true; | |
4023 | } | |
4024 | ||
aec7ae7d JJ |
4025 | /* Given a vector type VECTYPE and permutation SEL returns |
4026 | the VECTOR_CST mask that implements the permutation of the | |
4027 | vector elements. If that is impossible to do, returns NULL. */ | |
a1e53f3f | 4028 | |
b826bea7 RH |
4029 | static tree |
4030 | gen_perm_mask (tree vectype, unsigned char *sel) | |
a1e53f3f | 4031 | { |
22e4dee7 | 4032 | tree mask_elt_type, mask_type, mask_vec; |
2635892a | 4033 | int i, nunits; |
a1e53f3f | 4034 | |
22e4dee7 | 4035 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 RH |
4036 | |
4037 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) | |
a1e53f3f L |
4038 | return NULL; |
4039 | ||
22e4dee7 | 4040 | mask_elt_type |
2635892a RH |
4041 | = lang_hooks.types.type_for_size |
4042 | (TREE_INT_CST_LOW (TYPE_SIZE (TREE_TYPE (vectype))), 1); | |
22e4dee7 | 4043 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 4044 | |
22e4dee7 | 4045 | mask_vec = NULL; |
aec7ae7d JJ |
4046 | for (i = nunits - 1; i >= 0; i--) |
4047 | mask_vec = tree_cons (NULL, build_int_cst (mask_elt_type, sel[i]), | |
4048 | mask_vec); | |
a1e53f3f L |
4049 | mask_vec = build_vector (mask_type, mask_vec); |
4050 | ||
2635892a | 4051 | return mask_vec; |
a1e53f3f L |
4052 | } |
4053 | ||
aec7ae7d JJ |
4054 | /* Given a vector type VECTYPE returns the VECTOR_CST mask that implements |
4055 | reversal of the vector elements. If that is impossible to do, | |
4056 | returns NULL. */ | |
4057 | ||
4058 | static tree | |
4059 | perm_mask_for_reverse (tree vectype) | |
4060 | { | |
4061 | int i, nunits; | |
4062 | unsigned char *sel; | |
4063 | ||
4064 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4065 | sel = XALLOCAVEC (unsigned char, nunits); | |
4066 | ||
4067 | for (i = 0; i < nunits; ++i) | |
4068 | sel[i] = nunits - 1 - i; | |
4069 | ||
b826bea7 | 4070 | return gen_perm_mask (vectype, sel); |
aec7ae7d JJ |
4071 | } |
4072 | ||
4073 | /* Given a vector variable X and Y, that was generated for the scalar | |
4074 | STMT, generate instructions to permute the vector elements of X and Y | |
4075 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
4076 | permuted vector variable. */ | |
a1e53f3f L |
4077 | |
4078 | static tree | |
aec7ae7d JJ |
4079 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple stmt, |
4080 | gimple_stmt_iterator *gsi) | |
a1e53f3f L |
4081 | { |
4082 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 4083 | tree perm_dest, data_ref; |
a1e53f3f L |
4084 | gimple perm_stmt; |
4085 | ||
a1e53f3f | 4086 | perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype); |
aec7ae7d | 4087 | data_ref = make_ssa_name (perm_dest, NULL); |
a1e53f3f L |
4088 | |
4089 | /* Generate the permute statement. */ | |
aec7ae7d JJ |
4090 | perm_stmt = gimple_build_assign_with_ops3 (VEC_PERM_EXPR, data_ref, |
4091 | x, y, mask_vec); | |
a1e53f3f L |
4092 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
4093 | ||
4094 | return data_ref; | |
4095 | } | |
4096 | ||
ebfd146a IR |
4097 | /* vectorizable_load. |
4098 | ||
b8698a0f L |
4099 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
4100 | can be vectorized. | |
4101 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4102 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4103 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4104 | ||
4105 | static bool | |
4106 | vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
4107 | slp_tree slp_node, slp_instance slp_node_instance) | |
4108 | { | |
4109 | tree scalar_dest; | |
4110 | tree vec_dest = NULL; | |
4111 | tree data_ref = NULL; | |
4112 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 4113 | stmt_vec_info prev_stmt_info; |
ebfd146a | 4114 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 4115 | struct loop *loop = NULL; |
ebfd146a | 4116 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 4117 | bool nested_in_vect_loop = false; |
ebfd146a IR |
4118 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; |
4119 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
272c6793 | 4120 | tree elem_type; |
ebfd146a | 4121 | tree new_temp; |
947131ba | 4122 | enum machine_mode mode; |
ebfd146a IR |
4123 | gimple new_stmt = NULL; |
4124 | tree dummy; | |
4125 | enum dr_alignment_support alignment_support_scheme; | |
4126 | tree dataref_ptr = NULL_TREE; | |
4127 | gimple ptr_incr; | |
4128 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4129 | int ncopies; | |
4130 | int i, j, group_size; | |
4131 | tree msq = NULL_TREE, lsq; | |
4132 | tree offset = NULL_TREE; | |
4133 | tree realignment_token = NULL_TREE; | |
4134 | gimple phi = NULL; | |
4135 | VEC(tree,heap) *dr_chain = NULL; | |
4136 | bool strided_load = false; | |
272c6793 | 4137 | bool load_lanes_p = false; |
ebfd146a | 4138 | gimple first_stmt; |
ebfd146a | 4139 | bool inv_p; |
a1e53f3f | 4140 | bool negative; |
ebfd146a IR |
4141 | bool compute_in_loop = false; |
4142 | struct loop *at_loop; | |
4143 | int vec_num; | |
4144 | bool slp = (slp_node != NULL); | |
4145 | bool slp_perm = false; | |
4146 | enum tree_code code; | |
a70d6342 IR |
4147 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
4148 | int vf; | |
272c6793 | 4149 | tree aggr_type; |
aec7ae7d JJ |
4150 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; |
4151 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
4152 | int gather_scale = 1; | |
4153 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
a70d6342 IR |
4154 | |
4155 | if (loop_vinfo) | |
4156 | { | |
4157 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
4158 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
4159 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4160 | } | |
4161 | else | |
3533e503 | 4162 | vf = 1; |
ebfd146a IR |
4163 | |
4164 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4165 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 4166 | case of SLP. */ |
437f4a00 | 4167 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
4168 | ncopies = 1; |
4169 | else | |
4170 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4171 | ||
4172 | gcc_assert (ncopies >= 1); | |
4173 | ||
4174 | /* FORNOW. This restriction should be relaxed. */ | |
4175 | if (nested_in_vect_loop && ncopies > 1) | |
4176 | { | |
4177 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4178 | fprintf (vect_dump, "multiple types in nested loop."); | |
4179 | return false; | |
4180 | } | |
4181 | ||
a70d6342 | 4182 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4183 | return false; |
4184 | ||
8644a673 | 4185 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4186 | return false; |
4187 | ||
4188 | /* Is vectorizable load? */ | |
4189 | if (!is_gimple_assign (stmt)) | |
4190 | return false; | |
4191 | ||
4192 | scalar_dest = gimple_assign_lhs (stmt); | |
4193 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4194 | return false; | |
4195 | ||
4196 | code = gimple_assign_rhs_code (stmt); | |
4197 | if (code != ARRAY_REF | |
4198 | && code != INDIRECT_REF | |
e9dbe7bb IR |
4199 | && code != COMPONENT_REF |
4200 | && code != IMAGPART_EXPR | |
70f34814 | 4201 | && code != REALPART_EXPR |
42373e0b RG |
4202 | && code != MEM_REF |
4203 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
ebfd146a IR |
4204 | return false; |
4205 | ||
4206 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
4207 | return false; | |
4208 | ||
a1e53f3f L |
4209 | negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0; |
4210 | if (negative && ncopies > 1) | |
4211 | { | |
4212 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4213 | fprintf (vect_dump, "multiple types with negative step."); | |
4214 | return false; | |
4215 | } | |
4216 | ||
7b7b1813 | 4217 | elem_type = TREE_TYPE (vectype); |
947131ba | 4218 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
4219 | |
4220 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
4221 | (e.g. - data copies). */ | |
947131ba | 4222 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a IR |
4223 | { |
4224 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4225 | fprintf (vect_dump, "Aligned load, but unsupported type."); | |
4226 | return false; | |
4227 | } | |
4228 | ||
ebfd146a IR |
4229 | /* Check if the load is a part of an interleaving chain. */ |
4230 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) | |
4231 | { | |
4232 | strided_load = true; | |
4233 | /* FORNOW */ | |
aec7ae7d | 4234 | gcc_assert (! nested_in_vect_loop && !STMT_VINFO_GATHER_P (stmt_info)); |
ebfd146a | 4235 | |
e14c1050 | 4236 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
4237 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
4238 | { | |
e14c1050 | 4239 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
4240 | if (vect_load_lanes_supported (vectype, group_size)) |
4241 | load_lanes_p = true; | |
4242 | else if (!vect_strided_load_supported (vectype, group_size)) | |
b602d918 RS |
4243 | return false; |
4244 | } | |
ebfd146a IR |
4245 | } |
4246 | ||
a1e53f3f L |
4247 | if (negative) |
4248 | { | |
aec7ae7d | 4249 | gcc_assert (!strided_load && !STMT_VINFO_GATHER_P (stmt_info)); |
a1e53f3f L |
4250 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
4251 | if (alignment_support_scheme != dr_aligned | |
4252 | && alignment_support_scheme != dr_unaligned_supported) | |
4253 | { | |
4254 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4255 | fprintf (vect_dump, "negative step but alignment required."); | |
4256 | return false; | |
4257 | } | |
2635892a | 4258 | if (!perm_mask_for_reverse (vectype)) |
a1e53f3f L |
4259 | { |
4260 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4261 | fprintf (vect_dump, "negative step and reversing not supported."); | |
4262 | return false; | |
4263 | } | |
4264 | } | |
4265 | ||
aec7ae7d JJ |
4266 | if (STMT_VINFO_GATHER_P (stmt_info)) |
4267 | { | |
4268 | gimple def_stmt; | |
4269 | tree def; | |
4270 | gather_decl = vect_check_gather (stmt, loop_vinfo, &gather_base, | |
4271 | &gather_off, &gather_scale); | |
4272 | gcc_assert (gather_decl); | |
4273 | if (!vect_is_simple_use_1 (gather_off, loop_vinfo, bb_vinfo, | |
4274 | &def_stmt, &def, &gather_dt, | |
4275 | &gather_off_vectype)) | |
4276 | { | |
4277 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4278 | fprintf (vect_dump, "gather index use not simple."); | |
4279 | return false; | |
4280 | } | |
4281 | } | |
4282 | ||
ebfd146a IR |
4283 | if (!vec_stmt) /* transformation not required. */ |
4284 | { | |
4285 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
272c6793 | 4286 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, NULL); |
ebfd146a IR |
4287 | return true; |
4288 | } | |
4289 | ||
4290 | if (vect_print_dump_info (REPORT_DETAILS)) | |
0ea25ecd | 4291 | fprintf (vect_dump, "transform load. ncopies = %d", ncopies); |
ebfd146a IR |
4292 | |
4293 | /** Transform. **/ | |
4294 | ||
aec7ae7d JJ |
4295 | if (STMT_VINFO_GATHER_P (stmt_info)) |
4296 | { | |
4297 | tree vec_oprnd0 = NULL_TREE, op; | |
4298 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
4299 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
4300 | tree ptr, mask, var, scale, perm_mask = NULL_TREE, prev_res = NULL_TREE; | |
4301 | edge pe = loop_preheader_edge (loop); | |
4302 | gimple_seq seq; | |
4303 | basic_block new_bb; | |
4304 | enum { NARROW, NONE, WIDEN } modifier; | |
4305 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
4306 | ||
4307 | if (nunits == gather_off_nunits) | |
4308 | modifier = NONE; | |
4309 | else if (nunits == gather_off_nunits / 2) | |
4310 | { | |
4311 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
4312 | modifier = WIDEN; | |
4313 | ||
4314 | for (i = 0; i < gather_off_nunits; ++i) | |
4315 | sel[i] = i | nunits; | |
4316 | ||
b826bea7 | 4317 | perm_mask = gen_perm_mask (gather_off_vectype, sel); |
aec7ae7d JJ |
4318 | gcc_assert (perm_mask != NULL_TREE); |
4319 | } | |
4320 | else if (nunits == gather_off_nunits * 2) | |
4321 | { | |
4322 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
4323 | modifier = NARROW; | |
4324 | ||
4325 | for (i = 0; i < nunits; ++i) | |
4326 | sel[i] = i < gather_off_nunits | |
4327 | ? i : i + nunits - gather_off_nunits; | |
4328 | ||
b826bea7 | 4329 | perm_mask = gen_perm_mask (vectype, sel); |
aec7ae7d JJ |
4330 | gcc_assert (perm_mask != NULL_TREE); |
4331 | ncopies *= 2; | |
4332 | } | |
4333 | else | |
4334 | gcc_unreachable (); | |
4335 | ||
4336 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
4337 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4338 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4339 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4340 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4341 | scaletype = TREE_VALUE (arglist); | |
4342 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
4343 | && types_compatible_p (srctype, masktype)); | |
4344 | ||
4345 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4346 | ||
4347 | ptr = fold_convert (ptrtype, gather_base); | |
4348 | if (!is_gimple_min_invariant (ptr)) | |
4349 | { | |
4350 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
4351 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
4352 | gcc_assert (!new_bb); | |
4353 | } | |
4354 | ||
4355 | /* Currently we support only unconditional gather loads, | |
4356 | so mask should be all ones. */ | |
4357 | if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
4358 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
4359 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) | |
4360 | { | |
4361 | REAL_VALUE_TYPE r; | |
4362 | long tmp[6]; | |
4363 | for (j = 0; j < 6; ++j) | |
4364 | tmp[j] = -1; | |
4365 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
4366 | mask = build_real (TREE_TYPE (masktype), r); | |
4367 | } | |
4368 | else | |
4369 | gcc_unreachable (); | |
4370 | mask = build_vector_from_val (masktype, mask); | |
4371 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
4372 | ||
4373 | scale = build_int_cst (scaletype, gather_scale); | |
4374 | ||
4375 | prev_stmt_info = NULL; | |
4376 | for (j = 0; j < ncopies; ++j) | |
4377 | { | |
4378 | if (modifier == WIDEN && (j & 1)) | |
4379 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
4380 | perm_mask, stmt, gsi); | |
4381 | else if (j == 0) | |
4382 | op = vec_oprnd0 | |
4383 | = vect_get_vec_def_for_operand (gather_off, stmt, NULL); | |
4384 | else | |
4385 | op = vec_oprnd0 | |
4386 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
4387 | ||
4388 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
4389 | { | |
4390 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
4391 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
4392 | var = vect_get_new_vect_var (idxtype, vect_simple_var, NULL); | |
4393 | add_referenced_var (var); | |
4394 | var = make_ssa_name (var, NULL); | |
4395 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); | |
4396 | new_stmt | |
4397 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, | |
4398 | op, NULL_TREE); | |
4399 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4400 | op = var; | |
4401 | } | |
4402 | ||
4403 | new_stmt | |
4404 | = gimple_build_call (gather_decl, 5, mask, ptr, op, mask, scale); | |
4405 | ||
4406 | if (!useless_type_conversion_p (vectype, rettype)) | |
4407 | { | |
4408 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
4409 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
4410 | var = vect_get_new_vect_var (rettype, vect_simple_var, NULL); | |
4411 | add_referenced_var (var); | |
4412 | op = make_ssa_name (var, new_stmt); | |
4413 | gimple_call_set_lhs (new_stmt, op); | |
4414 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4415 | var = make_ssa_name (vec_dest, NULL); | |
4416 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); | |
4417 | new_stmt | |
4418 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, op, | |
4419 | NULL_TREE); | |
4420 | } | |
4421 | else | |
4422 | { | |
4423 | var = make_ssa_name (vec_dest, new_stmt); | |
4424 | gimple_call_set_lhs (new_stmt, var); | |
4425 | } | |
4426 | ||
4427 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4428 | ||
4429 | if (modifier == NARROW) | |
4430 | { | |
4431 | if ((j & 1) == 0) | |
4432 | { | |
4433 | prev_res = var; | |
4434 | continue; | |
4435 | } | |
4436 | var = permute_vec_elements (prev_res, var, | |
4437 | perm_mask, stmt, gsi); | |
4438 | new_stmt = SSA_NAME_DEF_STMT (var); | |
4439 | } | |
4440 | ||
4441 | if (prev_stmt_info == NULL) | |
4442 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4443 | else | |
4444 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4445 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4446 | } | |
4447 | return true; | |
4448 | } | |
4449 | ||
ebfd146a IR |
4450 | if (strided_load) |
4451 | { | |
e14c1050 | 4452 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
6aa904c4 IR |
4453 | if (slp |
4454 | && !SLP_INSTANCE_LOAD_PERMUTATION (slp_node_instance) | |
4455 | && first_stmt != VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0)) | |
4456 | first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0); | |
4457 | ||
ebfd146a IR |
4458 | /* Check if the chain of loads is already vectorized. */ |
4459 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt))) | |
4460 | { | |
4461 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4462 | return true; | |
4463 | } | |
4464 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 4465 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a IR |
4466 | |
4467 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
4468 | if (slp) | |
4469 | { | |
4470 | strided_load = false; | |
4471 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
a70d6342 IR |
4472 | if (SLP_INSTANCE_LOAD_PERMUTATION (slp_node_instance)) |
4473 | slp_perm = true; | |
4474 | } | |
ebfd146a IR |
4475 | else |
4476 | vec_num = group_size; | |
ebfd146a IR |
4477 | } |
4478 | else | |
4479 | { | |
4480 | first_stmt = stmt; | |
4481 | first_dr = dr; | |
4482 | group_size = vec_num = 1; | |
4483 | } | |
4484 | ||
720f5239 | 4485 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 4486 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
4487 | /* Targets with load-lane instructions must not require explicit |
4488 | realignment. */ | |
4489 | gcc_assert (!load_lanes_p | |
4490 | || alignment_support_scheme == dr_aligned | |
4491 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
4492 | |
4493 | /* In case the vectorization factor (VF) is bigger than the number | |
4494 | of elements that we can fit in a vectype (nunits), we have to generate | |
4495 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 4496 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 4497 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 4498 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 4499 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
4500 | stmts that use the defs of the current stmt. The example below |
4501 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
4502 | need to create 4 vectorized stmts): | |
ebfd146a IR |
4503 | |
4504 | before vectorization: | |
4505 | RELATED_STMT VEC_STMT | |
4506 | S1: x = memref - - | |
4507 | S2: z = x + 1 - - | |
4508 | ||
4509 | step 1: vectorize stmt S1: | |
4510 | We first create the vector stmt VS1_0, and, as usual, record a | |
4511 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
4512 | Next, we create the vector stmt VS1_1, and record a pointer to | |
4513 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 4514 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
4515 | stmts and pointers: |
4516 | RELATED_STMT VEC_STMT | |
4517 | VS1_0: vx0 = memref0 VS1_1 - | |
4518 | VS1_1: vx1 = memref1 VS1_2 - | |
4519 | VS1_2: vx2 = memref2 VS1_3 - | |
4520 | VS1_3: vx3 = memref3 - - | |
4521 | S1: x = load - VS1_0 | |
4522 | S2: z = x + 1 - - | |
4523 | ||
b8698a0f L |
4524 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
4525 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
4526 | stmt S2. */ |
4527 | ||
4528 | /* In case of interleaving (non-unit strided access): | |
4529 | ||
4530 | S1: x2 = &base + 2 | |
4531 | S2: x0 = &base | |
4532 | S3: x1 = &base + 1 | |
4533 | S4: x3 = &base + 3 | |
4534 | ||
b8698a0f | 4535 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
4536 | starting from the access of the first stmt of the chain: |
4537 | ||
4538 | VS1: vx0 = &base | |
4539 | VS2: vx1 = &base + vec_size*1 | |
4540 | VS3: vx3 = &base + vec_size*2 | |
4541 | VS4: vx4 = &base + vec_size*3 | |
4542 | ||
4543 | Then permutation statements are generated: | |
4544 | ||
4545 | VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 > | |
4546 | VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 > | |
4547 | ... | |
4548 | ||
4549 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
4550 | (the order of the data-refs in the output of vect_permute_load_chain | |
4551 | corresponds to the order of scalar stmts in the interleaving chain - see | |
4552 | the documentation of vect_permute_load_chain()). | |
4553 | The generation of permutation stmts and recording them in | |
4554 | STMT_VINFO_VEC_STMT is done in vect_transform_strided_load(). | |
4555 | ||
b8698a0f | 4556 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
4557 | permutation stmts above are created for every copy. The result vector |
4558 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
4559 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
4560 | |
4561 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
4562 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
4563 | we generate the following code: | |
4564 | p = initial_addr; | |
4565 | indx = 0; | |
4566 | loop { | |
4567 | p = p + indx * vectype_size; | |
4568 | vec_dest = *(p); | |
4569 | indx = indx + 1; | |
4570 | } | |
4571 | ||
4572 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 4573 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
4574 | then generate the following code, in which the data in each iteration is |
4575 | obtained by two vector loads, one from the previous iteration, and one | |
4576 | from the current iteration: | |
4577 | p1 = initial_addr; | |
4578 | msq_init = *(floor(p1)) | |
4579 | p2 = initial_addr + VS - 1; | |
4580 | realignment_token = call target_builtin; | |
4581 | indx = 0; | |
4582 | loop { | |
4583 | p2 = p2 + indx * vectype_size | |
4584 | lsq = *(floor(p2)) | |
4585 | vec_dest = realign_load (msq, lsq, realignment_token) | |
4586 | indx = indx + 1; | |
4587 | msq = lsq; | |
4588 | } */ | |
4589 | ||
4590 | /* If the misalignment remains the same throughout the execution of the | |
4591 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 4592 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
4593 | This can only occur when vectorizing memory accesses in the inner-loop |
4594 | nested within an outer-loop that is being vectorized. */ | |
4595 | ||
a70d6342 | 4596 | if (loop && nested_in_vect_loop_p (loop, stmt) |
ebfd146a IR |
4597 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
4598 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) | |
4599 | { | |
4600 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
4601 | compute_in_loop = true; | |
4602 | } | |
4603 | ||
4604 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
4605 | || alignment_support_scheme == dr_explicit_realign) | |
4606 | && !compute_in_loop) | |
4607 | { | |
4608 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
4609 | alignment_support_scheme, NULL_TREE, | |
4610 | &at_loop); | |
4611 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
4612 | { | |
4613 | phi = SSA_NAME_DEF_STMT (msq); | |
4614 | offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
4615 | } | |
4616 | } | |
4617 | else | |
4618 | at_loop = loop; | |
4619 | ||
a1e53f3f L |
4620 | if (negative) |
4621 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
4622 | ||
272c6793 RS |
4623 | if (load_lanes_p) |
4624 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
4625 | else | |
4626 | aggr_type = vectype; | |
4627 | ||
ebfd146a IR |
4628 | prev_stmt_info = NULL; |
4629 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 4630 | { |
272c6793 | 4631 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 4632 | if (j == 0) |
272c6793 | 4633 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, |
920e8172 RS |
4634 | offset, &dummy, gsi, |
4635 | &ptr_incr, false, &inv_p); | |
ebfd146a | 4636 | else |
272c6793 RS |
4637 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
4638 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 4639 | |
5ce1ee7f RS |
4640 | if (strided_load || slp_perm) |
4641 | dr_chain = VEC_alloc (tree, heap, vec_num); | |
4642 | ||
272c6793 | 4643 | if (load_lanes_p) |
ebfd146a | 4644 | { |
272c6793 RS |
4645 | tree vec_array; |
4646 | ||
4647 | vec_array = create_vector_array (vectype, vec_num); | |
4648 | ||
4649 | /* Emit: | |
4650 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
4651 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
4652 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
4653 | gimple_call_set_lhs (new_stmt, vec_array); | |
4654 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4655 | mark_symbols_for_renaming (new_stmt); | |
ebfd146a | 4656 | |
272c6793 RS |
4657 | /* Extract each vector into an SSA_NAME. */ |
4658 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 4659 | { |
272c6793 RS |
4660 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
4661 | vec_array, i); | |
4662 | VEC_quick_push (tree, dr_chain, new_temp); | |
4663 | } | |
4664 | ||
4665 | /* Record the mapping between SSA_NAMEs and statements. */ | |
4666 | vect_record_strided_load_vectors (stmt, dr_chain); | |
4667 | } | |
4668 | else | |
4669 | { | |
4670 | for (i = 0; i < vec_num; i++) | |
4671 | { | |
4672 | if (i > 0) | |
4673 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
4674 | stmt, NULL_TREE); | |
4675 | ||
4676 | /* 2. Create the vector-load in the loop. */ | |
4677 | switch (alignment_support_scheme) | |
4678 | { | |
4679 | case dr_aligned: | |
4680 | case dr_unaligned_supported: | |
be1ac4ec | 4681 | { |
272c6793 RS |
4682 | struct ptr_info_def *pi; |
4683 | data_ref | |
4684 | = build2 (MEM_REF, vectype, dataref_ptr, | |
4685 | build_int_cst (reference_alias_ptr_type | |
4686 | (DR_REF (first_dr)), 0)); | |
4687 | pi = get_ptr_info (dataref_ptr); | |
4688 | pi->align = TYPE_ALIGN_UNIT (vectype); | |
4689 | if (alignment_support_scheme == dr_aligned) | |
4690 | { | |
4691 | gcc_assert (aligned_access_p (first_dr)); | |
4692 | pi->misalign = 0; | |
4693 | } | |
4694 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
4695 | { | |
4696 | TREE_TYPE (data_ref) | |
4697 | = build_aligned_type (TREE_TYPE (data_ref), | |
4698 | TYPE_ALIGN (elem_type)); | |
4699 | pi->align = TYPE_ALIGN_UNIT (elem_type); | |
4700 | pi->misalign = 0; | |
4701 | } | |
4702 | else | |
4703 | { | |
4704 | TREE_TYPE (data_ref) | |
4705 | = build_aligned_type (TREE_TYPE (data_ref), | |
4706 | TYPE_ALIGN (elem_type)); | |
4707 | pi->misalign = DR_MISALIGNMENT (first_dr); | |
4708 | } | |
4709 | break; | |
be1ac4ec | 4710 | } |
272c6793 | 4711 | case dr_explicit_realign: |
267d3070 | 4712 | { |
272c6793 RS |
4713 | tree ptr, bump; |
4714 | tree vs_minus_1; | |
4715 | ||
4716 | vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
4717 | ||
4718 | if (compute_in_loop) | |
4719 | msq = vect_setup_realignment (first_stmt, gsi, | |
4720 | &realignment_token, | |
4721 | dr_explicit_realign, | |
4722 | dataref_ptr, NULL); | |
4723 | ||
4724 | new_stmt = gimple_build_assign_with_ops | |
4725 | (BIT_AND_EXPR, NULL_TREE, dataref_ptr, | |
4726 | build_int_cst | |
4727 | (TREE_TYPE (dataref_ptr), | |
4728 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
4729 | ptr = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt); | |
4730 | gimple_assign_set_lhs (new_stmt, ptr); | |
4731 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4732 | data_ref | |
4733 | = build2 (MEM_REF, vectype, ptr, | |
4734 | build_int_cst (reference_alias_ptr_type | |
4735 | (DR_REF (first_dr)), 0)); | |
4736 | vec_dest = vect_create_destination_var (scalar_dest, | |
4737 | vectype); | |
4738 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
4739 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4740 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4741 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
4742 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
4743 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4744 | msq = new_temp; | |
4745 | ||
4746 | bump = size_binop (MULT_EXPR, vs_minus_1, | |
7b7b1813 | 4747 | TYPE_SIZE_UNIT (elem_type)); |
272c6793 RS |
4748 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
4749 | new_stmt = gimple_build_assign_with_ops | |
4750 | (BIT_AND_EXPR, NULL_TREE, ptr, | |
4751 | build_int_cst | |
4752 | (TREE_TYPE (ptr), | |
4753 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
4754 | ptr = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt); | |
4755 | gimple_assign_set_lhs (new_stmt, ptr); | |
4756 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4757 | data_ref | |
4758 | = build2 (MEM_REF, vectype, ptr, | |
4759 | build_int_cst (reference_alias_ptr_type | |
4760 | (DR_REF (first_dr)), 0)); | |
4761 | break; | |
267d3070 | 4762 | } |
272c6793 RS |
4763 | case dr_explicit_realign_optimized: |
4764 | new_stmt = gimple_build_assign_with_ops | |
4765 | (BIT_AND_EXPR, NULL_TREE, dataref_ptr, | |
4766 | build_int_cst | |
4767 | (TREE_TYPE (dataref_ptr), | |
4768 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
4769 | new_temp = make_ssa_name (SSA_NAME_VAR (dataref_ptr), | |
4770 | new_stmt); | |
4771 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4772 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4773 | data_ref | |
4774 | = build2 (MEM_REF, vectype, new_temp, | |
4775 | build_int_cst (reference_alias_ptr_type | |
4776 | (DR_REF (first_dr)), 0)); | |
4777 | break; | |
4778 | default: | |
4779 | gcc_unreachable (); | |
4780 | } | |
ebfd146a | 4781 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 4782 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
4783 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4784 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4785 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 | 4786 | mark_symbols_for_renaming (new_stmt); |
ebfd146a | 4787 | |
272c6793 RS |
4788 | /* 3. Handle explicit realignment if necessary/supported. |
4789 | Create in loop: | |
4790 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
4791 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
4792 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 4793 | { |
272c6793 RS |
4794 | lsq = gimple_assign_lhs (new_stmt); |
4795 | if (!realignment_token) | |
4796 | realignment_token = dataref_ptr; | |
4797 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4798 | new_stmt | |
4799 | = gimple_build_assign_with_ops3 (REALIGN_LOAD_EXPR, | |
4800 | vec_dest, msq, lsq, | |
4801 | realignment_token); | |
4802 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4803 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4804 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4805 | ||
4806 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
4807 | { | |
4808 | gcc_assert (phi); | |
4809 | if (i == vec_num - 1 && j == ncopies - 1) | |
4810 | add_phi_arg (phi, lsq, | |
4811 | loop_latch_edge (containing_loop), | |
4812 | UNKNOWN_LOCATION); | |
4813 | msq = lsq; | |
4814 | } | |
ebfd146a | 4815 | } |
ebfd146a | 4816 | |
272c6793 RS |
4817 | /* 4. Handle invariant-load. */ |
4818 | if (inv_p && !bb_vinfo) | |
ebfd146a | 4819 | { |
98f4fb34 | 4820 | tree tem, vec_inv; |
ab70d825 | 4821 | gimple_stmt_iterator gsi2 = *gsi; |
272c6793 | 4822 | gcc_assert (!strided_load); |
ab70d825 | 4823 | gsi_next (&gsi2); |
98f4fb34 RG |
4824 | tem = scalar_dest; |
4825 | if (!useless_type_conversion_p (TREE_TYPE (vectype), | |
4826 | TREE_TYPE (tem))) | |
4827 | { | |
4828 | tem = fold_convert (TREE_TYPE (vectype), tem); | |
4829 | tem = force_gimple_operand_gsi (&gsi2, tem, true, | |
4830 | NULL_TREE, true, | |
4831 | GSI_SAME_STMT); | |
4832 | } | |
4833 | vec_inv = build_vector_from_val (vectype, tem); | |
ab70d825 RG |
4834 | new_temp = vect_init_vector (stmt, vec_inv, |
4835 | vectype, &gsi2); | |
4836 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
272c6793 | 4837 | } |
ebfd146a | 4838 | |
272c6793 RS |
4839 | if (negative) |
4840 | { | |
aec7ae7d JJ |
4841 | tree perm_mask = perm_mask_for_reverse (vectype); |
4842 | new_temp = permute_vec_elements (new_temp, new_temp, | |
4843 | perm_mask, stmt, gsi); | |
ebfd146a IR |
4844 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
4845 | } | |
267d3070 | 4846 | |
272c6793 RS |
4847 | /* Collect vector loads and later create their permutation in |
4848 | vect_transform_strided_load (). */ | |
4849 | if (strided_load || slp_perm) | |
4850 | VEC_quick_push (tree, dr_chain, new_temp); | |
267d3070 | 4851 | |
272c6793 RS |
4852 | /* Store vector loads in the corresponding SLP_NODE. */ |
4853 | if (slp && !slp_perm) | |
4854 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), | |
4855 | new_stmt); | |
4856 | } | |
ebfd146a IR |
4857 | } |
4858 | ||
4859 | if (slp && !slp_perm) | |
4860 | continue; | |
4861 | ||
4862 | if (slp_perm) | |
4863 | { | |
a70d6342 | 4864 | if (!vect_transform_slp_perm_load (stmt, dr_chain, gsi, vf, |
ebfd146a IR |
4865 | slp_node_instance, false)) |
4866 | { | |
4867 | VEC_free (tree, heap, dr_chain); | |
4868 | return false; | |
4869 | } | |
4870 | } | |
4871 | else | |
4872 | { | |
4873 | if (strided_load) | |
4874 | { | |
272c6793 RS |
4875 | if (!load_lanes_p) |
4876 | vect_transform_strided_load (stmt, dr_chain, group_size, gsi); | |
ebfd146a | 4877 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
4878 | } |
4879 | else | |
4880 | { | |
4881 | if (j == 0) | |
4882 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4883 | else | |
4884 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4885 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4886 | } | |
4887 | } | |
5ce1ee7f RS |
4888 | if (dr_chain) |
4889 | VEC_free (tree, heap, dr_chain); | |
ebfd146a IR |
4890 | } |
4891 | ||
ebfd146a IR |
4892 | return true; |
4893 | } | |
4894 | ||
4895 | /* Function vect_is_simple_cond. | |
b8698a0f | 4896 | |
ebfd146a IR |
4897 | Input: |
4898 | LOOP - the loop that is being vectorized. | |
4899 | COND - Condition that is checked for simple use. | |
4900 | ||
e9e1d143 RG |
4901 | Output: |
4902 | *COMP_VECTYPE - the vector type for the comparison. | |
4903 | ||
ebfd146a IR |
4904 | Returns whether a COND can be vectorized. Checks whether |
4905 | condition operands are supportable using vec_is_simple_use. */ | |
4906 | ||
87aab9b2 | 4907 | static bool |
f7e531cf IR |
4908 | vect_is_simple_cond (tree cond, loop_vec_info loop_vinfo, bb_vec_info bb_vinfo, |
4909 | tree *comp_vectype) | |
ebfd146a IR |
4910 | { |
4911 | tree lhs, rhs; | |
4912 | tree def; | |
4913 | enum vect_def_type dt; | |
e9e1d143 | 4914 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a IR |
4915 | |
4916 | if (!COMPARISON_CLASS_P (cond)) | |
4917 | return false; | |
4918 | ||
4919 | lhs = TREE_OPERAND (cond, 0); | |
4920 | rhs = TREE_OPERAND (cond, 1); | |
4921 | ||
4922 | if (TREE_CODE (lhs) == SSA_NAME) | |
4923 | { | |
4924 | gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); | |
f7e531cf | 4925 | if (!vect_is_simple_use_1 (lhs, loop_vinfo, bb_vinfo, &lhs_def_stmt, &def, |
e9e1d143 | 4926 | &dt, &vectype1)) |
ebfd146a IR |
4927 | return false; |
4928 | } | |
4929 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
4930 | && TREE_CODE (lhs) != FIXED_CST) | |
4931 | return false; | |
4932 | ||
4933 | if (TREE_CODE (rhs) == SSA_NAME) | |
4934 | { | |
4935 | gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); | |
f7e531cf | 4936 | if (!vect_is_simple_use_1 (rhs, loop_vinfo, bb_vinfo, &rhs_def_stmt, &def, |
e9e1d143 | 4937 | &dt, &vectype2)) |
ebfd146a IR |
4938 | return false; |
4939 | } | |
f7e531cf | 4940 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST |
ebfd146a IR |
4941 | && TREE_CODE (rhs) != FIXED_CST) |
4942 | return false; | |
4943 | ||
e9e1d143 | 4944 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
4945 | return true; |
4946 | } | |
4947 | ||
4948 | /* vectorizable_condition. | |
4949 | ||
b8698a0f L |
4950 | Check if STMT is conditional modify expression that can be vectorized. |
4951 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4952 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
4953 | at GSI. |
4954 | ||
4955 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
4956 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
4957 | else caluse if it is 2). | |
ebfd146a IR |
4958 | |
4959 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4960 | ||
4bbe8262 | 4961 | bool |
ebfd146a | 4962 | vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi, |
f7e531cf IR |
4963 | gimple *vec_stmt, tree reduc_def, int reduc_index, |
4964 | slp_tree slp_node) | |
ebfd146a IR |
4965 | { |
4966 | tree scalar_dest = NULL_TREE; | |
4967 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
4968 | tree cond_expr, then_clause, else_clause; |
4969 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4970 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
df11cc78 | 4971 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
4972 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
4973 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
ebfd146a IR |
4974 | tree vec_compare, vec_cond_expr; |
4975 | tree new_temp; | |
4976 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
ebfd146a | 4977 | tree def; |
a855b1b1 | 4978 | enum vect_def_type dt, dts[4]; |
ebfd146a | 4979 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
f7e531cf | 4980 | int ncopies; |
ebfd146a | 4981 | enum tree_code code; |
a855b1b1 | 4982 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
4983 | int i, j; |
4984 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
4985 | VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; | |
4986 | VEC (tree, heap) *vec_oprnds2 = NULL, *vec_oprnds3 = NULL; | |
b8698a0f | 4987 | |
f7e531cf IR |
4988 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
4989 | ncopies = 1; | |
4990 | else | |
4991 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
437f4a00 | 4992 | |
ebfd146a | 4993 | gcc_assert (ncopies >= 1); |
a855b1b1 | 4994 | if (reduc_index && ncopies > 1) |
ebfd146a IR |
4995 | return false; /* FORNOW */ |
4996 | ||
f7e531cf IR |
4997 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
4998 | return false; | |
4999 | ||
5000 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
ebfd146a IR |
5001 | return false; |
5002 | ||
4bbe8262 IR |
5003 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5004 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
5005 | && reduc_def)) | |
ebfd146a IR |
5006 | return false; |
5007 | ||
ebfd146a | 5008 | /* FORNOW: not yet supported. */ |
b8698a0f | 5009 | if (STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a IR |
5010 | { |
5011 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5012 | fprintf (vect_dump, "value used after loop."); | |
5013 | return false; | |
5014 | } | |
5015 | ||
5016 | /* Is vectorizable conditional operation? */ | |
5017 | if (!is_gimple_assign (stmt)) | |
5018 | return false; | |
5019 | ||
5020 | code = gimple_assign_rhs_code (stmt); | |
5021 | ||
5022 | if (code != COND_EXPR) | |
5023 | return false; | |
5024 | ||
4e71066d RG |
5025 | cond_expr = gimple_assign_rhs1 (stmt); |
5026 | then_clause = gimple_assign_rhs2 (stmt); | |
5027 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 5028 | |
f7e531cf | 5029 | if (!vect_is_simple_cond (cond_expr, loop_vinfo, bb_vinfo, &comp_vectype) |
e9e1d143 | 5030 | || !comp_vectype) |
ebfd146a IR |
5031 | return false; |
5032 | ||
5033 | if (TREE_CODE (then_clause) == SSA_NAME) | |
5034 | { | |
5035 | gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause); | |
f7e531cf | 5036 | if (!vect_is_simple_use (then_clause, loop_vinfo, bb_vinfo, |
ebfd146a IR |
5037 | &then_def_stmt, &def, &dt)) |
5038 | return false; | |
5039 | } | |
b8698a0f | 5040 | else if (TREE_CODE (then_clause) != INTEGER_CST |
ebfd146a IR |
5041 | && TREE_CODE (then_clause) != REAL_CST |
5042 | && TREE_CODE (then_clause) != FIXED_CST) | |
5043 | return false; | |
5044 | ||
5045 | if (TREE_CODE (else_clause) == SSA_NAME) | |
5046 | { | |
5047 | gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause); | |
f7e531cf | 5048 | if (!vect_is_simple_use (else_clause, loop_vinfo, bb_vinfo, |
ebfd146a IR |
5049 | &else_def_stmt, &def, &dt)) |
5050 | return false; | |
5051 | } | |
b8698a0f | 5052 | else if (TREE_CODE (else_clause) != INTEGER_CST |
ebfd146a IR |
5053 | && TREE_CODE (else_clause) != REAL_CST |
5054 | && TREE_CODE (else_clause) != FIXED_CST) | |
5055 | return false; | |
5056 | ||
b8698a0f | 5057 | if (!vec_stmt) |
ebfd146a IR |
5058 | { |
5059 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
e9e1d143 | 5060 | return expand_vec_cond_expr_p (vectype, comp_vectype); |
ebfd146a IR |
5061 | } |
5062 | ||
f7e531cf IR |
5063 | /* Transform. */ |
5064 | ||
5065 | if (!slp_node) | |
5066 | { | |
5067 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
5068 | vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
5069 | vec_oprnds2 = VEC_alloc (tree, heap, 1); | |
5070 | vec_oprnds3 = VEC_alloc (tree, heap, 1); | |
5071 | } | |
ebfd146a IR |
5072 | |
5073 | /* Handle def. */ | |
5074 | scalar_dest = gimple_assign_lhs (stmt); | |
5075 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5076 | ||
5077 | /* Handle cond expr. */ | |
a855b1b1 MM |
5078 | for (j = 0; j < ncopies; j++) |
5079 | { | |
f7e531cf | 5080 | gimple new_stmt = NULL; |
a855b1b1 MM |
5081 | if (j == 0) |
5082 | { | |
f7e531cf IR |
5083 | if (slp_node) |
5084 | { | |
5085 | VEC (tree, heap) *ops = VEC_alloc (tree, heap, 4); | |
5086 | VEC (slp_void_p, heap) *vec_defs; | |
5087 | ||
5088 | vec_defs = VEC_alloc (slp_void_p, heap, 4); | |
5089 | VEC_safe_push (tree, heap, ops, TREE_OPERAND (cond_expr, 0)); | |
5090 | VEC_safe_push (tree, heap, ops, TREE_OPERAND (cond_expr, 1)); | |
5091 | VEC_safe_push (tree, heap, ops, then_clause); | |
5092 | VEC_safe_push (tree, heap, ops, else_clause); | |
5093 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); | |
5094 | vec_oprnds3 = (VEC (tree, heap) *) VEC_pop (slp_void_p, vec_defs); | |
5095 | vec_oprnds2 = (VEC (tree, heap) *) VEC_pop (slp_void_p, vec_defs); | |
5096 | vec_oprnds1 = (VEC (tree, heap) *) VEC_pop (slp_void_p, vec_defs); | |
5097 | vec_oprnds0 = (VEC (tree, heap) *) VEC_pop (slp_void_p, vec_defs); | |
5098 | ||
5099 | VEC_free (tree, heap, ops); | |
5100 | VEC_free (slp_void_p, heap, vec_defs); | |
5101 | } | |
5102 | else | |
5103 | { | |
5104 | gimple gtemp; | |
5105 | vec_cond_lhs = | |
a855b1b1 MM |
5106 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), |
5107 | stmt, NULL); | |
f7e531cf | 5108 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), loop_vinfo, |
a855b1b1 | 5109 | NULL, >emp, &def, &dts[0]); |
f7e531cf IR |
5110 | |
5111 | vec_cond_rhs = | |
5112 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
5113 | stmt, NULL); | |
5114 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), loop_vinfo, | |
5115 | NULL, >emp, &def, &dts[1]); | |
5116 | if (reduc_index == 1) | |
5117 | vec_then_clause = reduc_def; | |
5118 | else | |
5119 | { | |
5120 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
5121 | stmt, NULL); | |
5122 | vect_is_simple_use (then_clause, loop_vinfo, | |
5123 | NULL, >emp, &def, &dts[2]); | |
5124 | } | |
5125 | if (reduc_index == 2) | |
5126 | vec_else_clause = reduc_def; | |
5127 | else | |
5128 | { | |
5129 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
a855b1b1 | 5130 | stmt, NULL); |
f7e531cf | 5131 | vect_is_simple_use (else_clause, loop_vinfo, |
a855b1b1 | 5132 | NULL, >emp, &def, &dts[3]); |
f7e531cf | 5133 | } |
a855b1b1 MM |
5134 | } |
5135 | } | |
5136 | else | |
5137 | { | |
f7e531cf IR |
5138 | vec_cond_lhs = vect_get_vec_def_for_stmt_copy (dts[0], |
5139 | VEC_pop (tree, vec_oprnds0)); | |
5140 | vec_cond_rhs = vect_get_vec_def_for_stmt_copy (dts[1], | |
5141 | VEC_pop (tree, vec_oprnds1)); | |
a855b1b1 | 5142 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
f7e531cf | 5143 | VEC_pop (tree, vec_oprnds2)); |
a855b1b1 | 5144 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
f7e531cf IR |
5145 | VEC_pop (tree, vec_oprnds3)); |
5146 | } | |
5147 | ||
5148 | if (!slp_node) | |
5149 | { | |
5150 | VEC_quick_push (tree, vec_oprnds0, vec_cond_lhs); | |
5151 | VEC_quick_push (tree, vec_oprnds1, vec_cond_rhs); | |
5152 | VEC_quick_push (tree, vec_oprnds2, vec_then_clause); | |
5153 | VEC_quick_push (tree, vec_oprnds3, vec_else_clause); | |
a855b1b1 MM |
5154 | } |
5155 | ||
9dc3f7de | 5156 | /* Arguments are ready. Create the new vector stmt. */ |
f7e531cf IR |
5157 | FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vec_cond_lhs) |
5158 | { | |
5159 | vec_cond_rhs = VEC_index (tree, vec_oprnds1, i); | |
5160 | vec_then_clause = VEC_index (tree, vec_oprnds2, i); | |
5161 | vec_else_clause = VEC_index (tree, vec_oprnds3, i); | |
a855b1b1 | 5162 | |
f7e531cf IR |
5163 | vec_compare = build2 (TREE_CODE (cond_expr), vectype, |
5164 | vec_cond_lhs, vec_cond_rhs); | |
5165 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, | |
5166 | vec_compare, vec_then_clause, vec_else_clause); | |
a855b1b1 | 5167 | |
f7e531cf IR |
5168 | new_stmt = gimple_build_assign (vec_dest, vec_cond_expr); |
5169 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5170 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5171 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5172 | if (slp_node) | |
5173 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
5174 | } | |
5175 | ||
5176 | if (slp_node) | |
5177 | continue; | |
5178 | ||
5179 | if (j == 0) | |
5180 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5181 | else | |
5182 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5183 | ||
5184 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 5185 | } |
b8698a0f | 5186 | |
f7e531cf IR |
5187 | VEC_free (tree, heap, vec_oprnds0); |
5188 | VEC_free (tree, heap, vec_oprnds1); | |
5189 | VEC_free (tree, heap, vec_oprnds2); | |
5190 | VEC_free (tree, heap, vec_oprnds3); | |
5191 | ||
ebfd146a IR |
5192 | return true; |
5193 | } | |
5194 | ||
5195 | ||
8644a673 | 5196 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
5197 | |
5198 | bool | |
a70d6342 | 5199 | vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 5200 | { |
8644a673 | 5201 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 5202 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 5203 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 5204 | bool ok; |
a70d6342 | 5205 | tree scalar_type, vectype; |
363477c0 JJ |
5206 | gimple pattern_stmt; |
5207 | gimple_seq pattern_def_seq; | |
ebfd146a IR |
5208 | |
5209 | if (vect_print_dump_info (REPORT_DETAILS)) | |
ebfd146a | 5210 | { |
8644a673 IR |
5211 | fprintf (vect_dump, "==> examining statement: "); |
5212 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
5213 | } | |
ebfd146a | 5214 | |
1825a1f3 | 5215 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 5216 | { |
1825a1f3 IR |
5217 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
5218 | fprintf (vect_dump, "not vectorized: stmt has volatile operands"); | |
5219 | ||
5220 | return false; | |
5221 | } | |
b8698a0f L |
5222 | |
5223 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
5224 | to include: |
5225 | - the COND_EXPR which is the loop exit condition | |
5226 | - any LABEL_EXPRs in the loop | |
b8698a0f | 5227 | - computations that are used only for array indexing or loop control. |
8644a673 | 5228 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 5229 | instance, therefore, all the statements are relevant. |
ebfd146a | 5230 | |
d092494c | 5231 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 IR |
5232 | if the original statement is not relevant. Otherwise, we analyze both |
5233 | statements. */ | |
5234 | ||
5235 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 5236 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 5237 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 5238 | { |
9d5e7640 | 5239 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 5240 | && pattern_stmt |
9d5e7640 IR |
5241 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
5242 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
5243 | { | |
83197f37 | 5244 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
5245 | stmt = pattern_stmt; |
5246 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
5247 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5248 | { | |
5249 | fprintf (vect_dump, "==> examining pattern statement: "); | |
5250 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
5251 | } | |
5252 | } | |
5253 | else | |
5254 | { | |
5255 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5256 | fprintf (vect_dump, "irrelevant."); | |
ebfd146a | 5257 | |
9d5e7640 IR |
5258 | return true; |
5259 | } | |
8644a673 | 5260 | } |
83197f37 IR |
5261 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
5262 | && pattern_stmt | |
5263 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
5264 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
5265 | { | |
5266 | /* Analyze PATTERN_STMT too. */ | |
5267 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5268 | { | |
5269 | fprintf (vect_dump, "==> examining pattern statement: "); | |
5270 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
5271 | } | |
5272 | ||
5273 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
5274 | return false; | |
5275 | } | |
ebfd146a | 5276 | |
1107f3ae | 5277 | if (is_pattern_stmt_p (stmt_info) |
363477c0 | 5278 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 5279 | { |
363477c0 | 5280 | gimple_stmt_iterator si; |
1107f3ae | 5281 | |
363477c0 JJ |
5282 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
5283 | { | |
5284 | gimple pattern_def_stmt = gsi_stmt (si); | |
5285 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) | |
5286 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
5287 | { | |
5288 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
5289 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5290 | { | |
5291 | fprintf (vect_dump, "==> examining pattern def statement: "); | |
5292 | print_gimple_stmt (vect_dump, pattern_def_stmt, 0, TDF_SLIM); | |
5293 | } | |
1107f3ae | 5294 | |
363477c0 JJ |
5295 | if (!vect_analyze_stmt (pattern_def_stmt, |
5296 | need_to_vectorize, node)) | |
5297 | return false; | |
5298 | } | |
5299 | } | |
5300 | } | |
1107f3ae | 5301 | |
8644a673 IR |
5302 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
5303 | { | |
5304 | case vect_internal_def: | |
5305 | break; | |
ebfd146a | 5306 | |
8644a673 | 5307 | case vect_reduction_def: |
7c5222ff | 5308 | case vect_nested_cycle: |
a70d6342 | 5309 | gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer |
8644a673 | 5310 | || relevance == vect_used_in_outer_by_reduction |
a70d6342 | 5311 | || relevance == vect_unused_in_scope)); |
8644a673 IR |
5312 | break; |
5313 | ||
5314 | case vect_induction_def: | |
5315 | case vect_constant_def: | |
5316 | case vect_external_def: | |
5317 | case vect_unknown_def_type: | |
5318 | default: | |
5319 | gcc_unreachable (); | |
5320 | } | |
ebfd146a | 5321 | |
a70d6342 IR |
5322 | if (bb_vinfo) |
5323 | { | |
5324 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
5325 | ||
b690cc0f | 5326 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
a70d6342 IR |
5327 | if (vect_print_dump_info (REPORT_DETAILS)) |
5328 | { | |
5329 | fprintf (vect_dump, "get vectype for scalar type: "); | |
5330 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
5331 | } | |
5332 | ||
5333 | vectype = get_vectype_for_scalar_type (scalar_type); | |
5334 | if (!vectype) | |
5335 | { | |
5336 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5337 | { | |
5338 | fprintf (vect_dump, "not SLPed: unsupported data-type "); | |
5339 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
5340 | } | |
5341 | return false; | |
5342 | } | |
5343 | ||
5344 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5345 | { | |
5346 | fprintf (vect_dump, "vectype: "); | |
5347 | print_generic_expr (vect_dump, vectype, TDF_SLIM); | |
5348 | } | |
5349 | ||
5350 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
5351 | } | |
5352 | ||
8644a673 | 5353 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 5354 | { |
8644a673 IR |
5355 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
5356 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info)); | |
5357 | *need_to_vectorize = true; | |
ebfd146a IR |
5358 | } |
5359 | ||
8644a673 | 5360 | ok = true; |
b8698a0f | 5361 | if (!bb_vinfo |
a70d6342 IR |
5362 | && (STMT_VINFO_RELEVANT_P (stmt_info) |
5363 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
4a00c761 | 5364 | ok = (vectorizable_conversion (stmt, NULL, NULL, NULL) |
9dc3f7de | 5365 | || vectorizable_shift (stmt, NULL, NULL, NULL) |
8644a673 IR |
5366 | || vectorizable_operation (stmt, NULL, NULL, NULL) |
5367 | || vectorizable_assignment (stmt, NULL, NULL, NULL) | |
5368 | || vectorizable_load (stmt, NULL, NULL, NULL, NULL) | |
190c2236 | 5369 | || vectorizable_call (stmt, NULL, NULL, NULL) |
8644a673 | 5370 | || vectorizable_store (stmt, NULL, NULL, NULL) |
b5aeb3bb | 5371 | || vectorizable_reduction (stmt, NULL, NULL, NULL) |
f7e531cf | 5372 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, NULL)); |
a70d6342 IR |
5373 | else |
5374 | { | |
5375 | if (bb_vinfo) | |
4a00c761 JJ |
5376 | ok = (vectorizable_conversion (stmt, NULL, NULL, node) |
5377 | || vectorizable_shift (stmt, NULL, NULL, node) | |
9dc3f7de | 5378 | || vectorizable_operation (stmt, NULL, NULL, node) |
a70d6342 IR |
5379 | || vectorizable_assignment (stmt, NULL, NULL, node) |
5380 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
190c2236 | 5381 | || vectorizable_call (stmt, NULL, NULL, node) |
f7e531cf IR |
5382 | || vectorizable_store (stmt, NULL, NULL, node) |
5383 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node)); | |
b8698a0f | 5384 | } |
8644a673 IR |
5385 | |
5386 | if (!ok) | |
ebfd146a | 5387 | { |
8644a673 IR |
5388 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
5389 | { | |
5390 | fprintf (vect_dump, "not vectorized: relevant stmt not "); | |
5391 | fprintf (vect_dump, "supported: "); | |
5392 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
5393 | } | |
b8698a0f | 5394 | |
ebfd146a IR |
5395 | return false; |
5396 | } | |
5397 | ||
a70d6342 IR |
5398 | if (bb_vinfo) |
5399 | return true; | |
5400 | ||
8644a673 IR |
5401 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
5402 | need extra handling, except for vectorizable reductions. */ | |
5403 | if (STMT_VINFO_LIVE_P (stmt_info) | |
5404 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
5405 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 5406 | |
8644a673 | 5407 | if (!ok) |
ebfd146a | 5408 | { |
8644a673 IR |
5409 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
5410 | { | |
5411 | fprintf (vect_dump, "not vectorized: live stmt not "); | |
5412 | fprintf (vect_dump, "supported: "); | |
5413 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
5414 | } | |
b8698a0f | 5415 | |
8644a673 | 5416 | return false; |
ebfd146a IR |
5417 | } |
5418 | ||
ebfd146a IR |
5419 | return true; |
5420 | } | |
5421 | ||
5422 | ||
5423 | /* Function vect_transform_stmt. | |
5424 | ||
5425 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
5426 | ||
5427 | bool | |
5428 | vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi, | |
b8698a0f | 5429 | bool *strided_store, slp_tree slp_node, |
ebfd146a IR |
5430 | slp_instance slp_node_instance) |
5431 | { | |
5432 | bool is_store = false; | |
5433 | gimple vec_stmt = NULL; | |
5434 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a | 5435 | bool done; |
ebfd146a IR |
5436 | |
5437 | switch (STMT_VINFO_TYPE (stmt_info)) | |
5438 | { | |
5439 | case type_demotion_vec_info_type: | |
ebfd146a | 5440 | case type_promotion_vec_info_type: |
ebfd146a IR |
5441 | case type_conversion_vec_info_type: |
5442 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
5443 | gcc_assert (done); | |
5444 | break; | |
5445 | ||
5446 | case induc_vec_info_type: | |
5447 | gcc_assert (!slp_node); | |
5448 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
5449 | gcc_assert (done); | |
5450 | break; | |
5451 | ||
9dc3f7de IR |
5452 | case shift_vec_info_type: |
5453 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
5454 | gcc_assert (done); | |
5455 | break; | |
5456 | ||
ebfd146a IR |
5457 | case op_vec_info_type: |
5458 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
5459 | gcc_assert (done); | |
5460 | break; | |
5461 | ||
5462 | case assignment_vec_info_type: | |
5463 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
5464 | gcc_assert (done); | |
5465 | break; | |
5466 | ||
5467 | case load_vec_info_type: | |
b8698a0f | 5468 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
5469 | slp_node_instance); |
5470 | gcc_assert (done); | |
5471 | break; | |
5472 | ||
5473 | case store_vec_info_type: | |
5474 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
5475 | gcc_assert (done); | |
5476 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) && !slp_node) | |
5477 | { | |
5478 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 5479 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
5480 | one are skipped, and there vec_stmt_info shouldn't be freed |
5481 | meanwhile. */ | |
5482 | *strided_store = true; | |
5483 | if (STMT_VINFO_VEC_STMT (stmt_info)) | |
5484 | is_store = true; | |
5485 | } | |
5486 | else | |
5487 | is_store = true; | |
5488 | break; | |
5489 | ||
5490 | case condition_vec_info_type: | |
f7e531cf | 5491 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
5492 | gcc_assert (done); |
5493 | break; | |
5494 | ||
5495 | case call_vec_info_type: | |
190c2236 | 5496 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 5497 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
5498 | break; |
5499 | ||
5500 | case reduc_vec_info_type: | |
b5aeb3bb | 5501 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
5502 | gcc_assert (done); |
5503 | break; | |
5504 | ||
5505 | default: | |
5506 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
5507 | { | |
5508 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5509 | fprintf (vect_dump, "stmt not supported."); | |
5510 | gcc_unreachable (); | |
5511 | } | |
5512 | } | |
5513 | ||
5514 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that | |
5515 | is being vectorized, but outside the immediately enclosing loop. */ | |
5516 | if (vec_stmt | |
a70d6342 IR |
5517 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
5518 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
5519 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
5520 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
5521 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 5522 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 5523 | vect_used_in_outer_by_reduction)) |
ebfd146a | 5524 | { |
a70d6342 IR |
5525 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
5526 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
5527 | imm_use_iterator imm_iter; |
5528 | use_operand_p use_p; | |
5529 | tree scalar_dest; | |
5530 | gimple exit_phi; | |
5531 | ||
5532 | if (vect_print_dump_info (REPORT_DETAILS)) | |
a70d6342 | 5533 | fprintf (vect_dump, "Record the vdef for outer-loop vectorization."); |
ebfd146a IR |
5534 | |
5535 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
5536 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
5537 | STMT). */ | |
5538 | if (gimple_code (stmt) == GIMPLE_PHI) | |
5539 | scalar_dest = PHI_RESULT (stmt); | |
5540 | else | |
5541 | scalar_dest = gimple_assign_lhs (stmt); | |
5542 | ||
5543 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
5544 | { | |
5545 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
5546 | { | |
5547 | exit_phi = USE_STMT (use_p); | |
5548 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
5549 | } | |
5550 | } | |
5551 | } | |
5552 | ||
5553 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
5554 | being vectorized. */ | |
5555 | if (STMT_VINFO_LIVE_P (stmt_info) | |
5556 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
5557 | { | |
5558 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
5559 | gcc_assert (done); | |
5560 | } | |
5561 | ||
5562 | if (vec_stmt) | |
83197f37 | 5563 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 5564 | |
b8698a0f | 5565 | return is_store; |
ebfd146a IR |
5566 | } |
5567 | ||
5568 | ||
b8698a0f | 5569 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
5570 | stmt_vec_info. */ |
5571 | ||
5572 | void | |
5573 | vect_remove_stores (gimple first_stmt) | |
5574 | { | |
5575 | gimple next = first_stmt; | |
5576 | gimple tmp; | |
5577 | gimple_stmt_iterator next_si; | |
5578 | ||
5579 | while (next) | |
5580 | { | |
78048b1c JJ |
5581 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
5582 | ||
5583 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
5584 | if (is_pattern_stmt_p (stmt_info)) | |
5585 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
5586 | /* Free the attached stmt_vec_info and remove the stmt. */ |
5587 | next_si = gsi_for_stmt (next); | |
5588 | gsi_remove (&next_si, true); | |
ebfd146a IR |
5589 | free_stmt_vec_info (next); |
5590 | next = tmp; | |
5591 | } | |
5592 | } | |
5593 | ||
5594 | ||
5595 | /* Function new_stmt_vec_info. | |
5596 | ||
5597 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
5598 | ||
5599 | stmt_vec_info | |
b8698a0f | 5600 | new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 5601 | bb_vec_info bb_vinfo) |
ebfd146a IR |
5602 | { |
5603 | stmt_vec_info res; | |
5604 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
5605 | ||
5606 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
5607 | STMT_VINFO_STMT (res) = stmt; | |
5608 | STMT_VINFO_LOOP_VINFO (res) = loop_vinfo; | |
a70d6342 | 5609 | STMT_VINFO_BB_VINFO (res) = bb_vinfo; |
8644a673 | 5610 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
5611 | STMT_VINFO_LIVE_P (res) = false; |
5612 | STMT_VINFO_VECTYPE (res) = NULL; | |
5613 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 5614 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
5615 | STMT_VINFO_IN_PATTERN_P (res) = false; |
5616 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 5617 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a IR |
5618 | STMT_VINFO_DATA_REF (res) = NULL; |
5619 | ||
5620 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
5621 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
5622 | STMT_VINFO_DR_INIT (res) = NULL; | |
5623 | STMT_VINFO_DR_STEP (res) = NULL; | |
5624 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
5625 | ||
5626 | if (gimple_code (stmt) == GIMPLE_PHI | |
5627 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
5628 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
5629 | else | |
8644a673 IR |
5630 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
5631 | ||
ebfd146a IR |
5632 | STMT_VINFO_SAME_ALIGN_REFS (res) = VEC_alloc (dr_p, heap, 5); |
5633 | STMT_VINFO_INSIDE_OF_LOOP_COST (res) = 0; | |
5634 | STMT_VINFO_OUTSIDE_OF_LOOP_COST (res) = 0; | |
32e8bb8e | 5635 | STMT_SLP_TYPE (res) = loop_vect; |
e14c1050 IR |
5636 | GROUP_FIRST_ELEMENT (res) = NULL; |
5637 | GROUP_NEXT_ELEMENT (res) = NULL; | |
5638 | GROUP_SIZE (res) = 0; | |
5639 | GROUP_STORE_COUNT (res) = 0; | |
5640 | GROUP_GAP (res) = 0; | |
5641 | GROUP_SAME_DR_STMT (res) = NULL; | |
5642 | GROUP_READ_WRITE_DEPENDENCE (res) = false; | |
ebfd146a IR |
5643 | |
5644 | return res; | |
5645 | } | |
5646 | ||
5647 | ||
5648 | /* Create a hash table for stmt_vec_info. */ | |
5649 | ||
5650 | void | |
5651 | init_stmt_vec_info_vec (void) | |
5652 | { | |
5653 | gcc_assert (!stmt_vec_info_vec); | |
5654 | stmt_vec_info_vec = VEC_alloc (vec_void_p, heap, 50); | |
5655 | } | |
5656 | ||
5657 | ||
5658 | /* Free hash table for stmt_vec_info. */ | |
5659 | ||
5660 | void | |
5661 | free_stmt_vec_info_vec (void) | |
5662 | { | |
5663 | gcc_assert (stmt_vec_info_vec); | |
5664 | VEC_free (vec_void_p, heap, stmt_vec_info_vec); | |
5665 | } | |
5666 | ||
5667 | ||
5668 | /* Free stmt vectorization related info. */ | |
5669 | ||
5670 | void | |
5671 | free_stmt_vec_info (gimple stmt) | |
5672 | { | |
5673 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5674 | ||
5675 | if (!stmt_info) | |
5676 | return; | |
5677 | ||
78048b1c JJ |
5678 | /* Check if this statement has a related "pattern stmt" |
5679 | (introduced by the vectorizer during the pattern recognition | |
5680 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
5681 | too. */ | |
5682 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
5683 | { | |
5684 | stmt_vec_info patt_info | |
5685 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
5686 | if (patt_info) | |
5687 | { | |
363477c0 JJ |
5688 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
5689 | if (seq) | |
5690 | { | |
5691 | gimple_stmt_iterator si; | |
5692 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
5693 | free_stmt_vec_info (gsi_stmt (si)); | |
5694 | } | |
78048b1c JJ |
5695 | free_stmt_vec_info (STMT_VINFO_RELATED_STMT (stmt_info)); |
5696 | } | |
5697 | } | |
5698 | ||
ebfd146a IR |
5699 | VEC_free (dr_p, heap, STMT_VINFO_SAME_ALIGN_REFS (stmt_info)); |
5700 | set_vinfo_for_stmt (stmt, NULL); | |
5701 | free (stmt_info); | |
5702 | } | |
5703 | ||
5704 | ||
bb67d9c7 | 5705 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 5706 | |
bb67d9c7 | 5707 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
5708 | by the target. */ |
5709 | ||
bb67d9c7 RG |
5710 | static tree |
5711 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a IR |
5712 | { |
5713 | enum machine_mode inner_mode = TYPE_MODE (scalar_type); | |
cc4b5170 | 5714 | enum machine_mode simd_mode; |
2f816591 | 5715 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
5716 | int nunits; |
5717 | tree vectype; | |
5718 | ||
cc4b5170 | 5719 | if (nbytes == 0) |
ebfd146a IR |
5720 | return NULL_TREE; |
5721 | ||
48f2e373 RB |
5722 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
5723 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
5724 | return NULL_TREE; | |
5725 | ||
2f816591 RG |
5726 | /* We can't build a vector type of elements with alignment bigger than |
5727 | their size. */ | |
5728 | if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) | |
5729 | return NULL_TREE; | |
5730 | ||
7b7b1813 RG |
5731 | /* For vector types of elements whose mode precision doesn't |
5732 | match their types precision we use a element type of mode | |
5733 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
5734 | they support the proper result truncation/extension. |
5735 | We also make sure to build vector types with INTEGER_TYPE | |
5736 | component type only. */ | |
6d7971b8 | 5737 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
5738 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
5739 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
5740 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
5741 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 5742 | |
ccbf5bb4 RG |
5743 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
5744 | When the component mode passes the above test simply use a type | |
5745 | corresponding to that mode. The theory is that any use that | |
5746 | would cause problems with this will disable vectorization anyway. */ | |
5747 | if (!SCALAR_FLOAT_TYPE_P (scalar_type) | |
5748 | && !INTEGRAL_TYPE_P (scalar_type) | |
5749 | && !POINTER_TYPE_P (scalar_type)) | |
5750 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); | |
5751 | ||
bb67d9c7 RG |
5752 | /* If no size was supplied use the mode the target prefers. Otherwise |
5753 | lookup a vector mode of the specified size. */ | |
5754 | if (size == 0) | |
5755 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
5756 | else | |
5757 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
5758 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
5759 | if (nunits <= 1) | |
5760 | return NULL_TREE; | |
ebfd146a IR |
5761 | |
5762 | vectype = build_vector_type (scalar_type, nunits); | |
5763 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5764 | { | |
5765 | fprintf (vect_dump, "get vectype with %d units of type ", nunits); | |
5766 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
5767 | } | |
5768 | ||
5769 | if (!vectype) | |
5770 | return NULL_TREE; | |
5771 | ||
5772 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5773 | { | |
5774 | fprintf (vect_dump, "vectype: "); | |
5775 | print_generic_expr (vect_dump, vectype, TDF_SLIM); | |
5776 | } | |
5777 | ||
5778 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
5779 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
5780 | { | |
5781 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5782 | fprintf (vect_dump, "mode not supported by target."); | |
5783 | return NULL_TREE; | |
5784 | } | |
5785 | ||
5786 | return vectype; | |
5787 | } | |
5788 | ||
bb67d9c7 RG |
5789 | unsigned int current_vector_size; |
5790 | ||
5791 | /* Function get_vectype_for_scalar_type. | |
5792 | ||
5793 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
5794 | by the target. */ | |
5795 | ||
5796 | tree | |
5797 | get_vectype_for_scalar_type (tree scalar_type) | |
5798 | { | |
5799 | tree vectype; | |
5800 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
5801 | current_vector_size); | |
5802 | if (vectype | |
5803 | && current_vector_size == 0) | |
5804 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
5805 | return vectype; | |
5806 | } | |
5807 | ||
b690cc0f RG |
5808 | /* Function get_same_sized_vectype |
5809 | ||
5810 | Returns a vector type corresponding to SCALAR_TYPE of size | |
5811 | VECTOR_TYPE if supported by the target. */ | |
5812 | ||
5813 | tree | |
bb67d9c7 | 5814 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 5815 | { |
bb67d9c7 RG |
5816 | return get_vectype_for_scalar_type_and_size |
5817 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
5818 | } |
5819 | ||
ebfd146a IR |
5820 | /* Function vect_is_simple_use. |
5821 | ||
5822 | Input: | |
a70d6342 IR |
5823 | LOOP_VINFO - the vect info of the loop that is being vectorized. |
5824 | BB_VINFO - the vect info of the basic block that is being vectorized. | |
5825 | OPERAND - operand of a stmt in the loop or bb. | |
ebfd146a IR |
5826 | DEF - the defining stmt in case OPERAND is an SSA_NAME. |
5827 | ||
5828 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 5829 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 5830 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 5831 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
5832 | is the case in reduction/induction computations). |
5833 | For basic blocks, supportable operands are constants and bb invariants. | |
5834 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
5835 | |
5836 | bool | |
b8698a0f | 5837 | vect_is_simple_use (tree operand, loop_vec_info loop_vinfo, |
a70d6342 | 5838 | bb_vec_info bb_vinfo, gimple *def_stmt, |
ebfd146a | 5839 | tree *def, enum vect_def_type *dt) |
b8698a0f | 5840 | { |
ebfd146a IR |
5841 | basic_block bb; |
5842 | stmt_vec_info stmt_vinfo; | |
a70d6342 | 5843 | struct loop *loop = NULL; |
b8698a0f | 5844 | |
a70d6342 IR |
5845 | if (loop_vinfo) |
5846 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
5847 | |
5848 | *def_stmt = NULL; | |
5849 | *def = NULL_TREE; | |
b8698a0f | 5850 | |
ebfd146a IR |
5851 | if (vect_print_dump_info (REPORT_DETAILS)) |
5852 | { | |
5853 | fprintf (vect_dump, "vect_is_simple_use: operand "); | |
5854 | print_generic_expr (vect_dump, operand, TDF_SLIM); | |
5855 | } | |
b8698a0f | 5856 | |
ebfd146a IR |
5857 | if (TREE_CODE (operand) == INTEGER_CST || TREE_CODE (operand) == REAL_CST) |
5858 | { | |
5859 | *dt = vect_constant_def; | |
5860 | return true; | |
5861 | } | |
b8698a0f | 5862 | |
ebfd146a IR |
5863 | if (is_gimple_min_invariant (operand)) |
5864 | { | |
5865 | *def = operand; | |
8644a673 | 5866 | *dt = vect_external_def; |
ebfd146a IR |
5867 | return true; |
5868 | } | |
5869 | ||
5870 | if (TREE_CODE (operand) == PAREN_EXPR) | |
5871 | { | |
5872 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5873 | fprintf (vect_dump, "non-associatable copy."); | |
5874 | operand = TREE_OPERAND (operand, 0); | |
5875 | } | |
b8698a0f | 5876 | |
ebfd146a IR |
5877 | if (TREE_CODE (operand) != SSA_NAME) |
5878 | { | |
5879 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5880 | fprintf (vect_dump, "not ssa-name."); | |
5881 | return false; | |
5882 | } | |
b8698a0f | 5883 | |
ebfd146a IR |
5884 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
5885 | if (*def_stmt == NULL) | |
5886 | { | |
5887 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5888 | fprintf (vect_dump, "no def_stmt."); | |
5889 | return false; | |
5890 | } | |
5891 | ||
5892 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5893 | { | |
5894 | fprintf (vect_dump, "def_stmt: "); | |
5895 | print_gimple_stmt (vect_dump, *def_stmt, 0, TDF_SLIM); | |
5896 | } | |
5897 | ||
8644a673 | 5898 | /* Empty stmt is expected only in case of a function argument. |
ebfd146a IR |
5899 | (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */ |
5900 | if (gimple_nop_p (*def_stmt)) | |
5901 | { | |
5902 | *def = operand; | |
8644a673 | 5903 | *dt = vect_external_def; |
ebfd146a IR |
5904 | return true; |
5905 | } | |
5906 | ||
5907 | bb = gimple_bb (*def_stmt); | |
a70d6342 IR |
5908 | |
5909 | if ((loop && !flow_bb_inside_loop_p (loop, bb)) | |
5910 | || (!loop && bb != BB_VINFO_BB (bb_vinfo)) | |
b8698a0f | 5911 | || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI)) |
8644a673 | 5912 | *dt = vect_external_def; |
ebfd146a IR |
5913 | else |
5914 | { | |
5915 | stmt_vinfo = vinfo_for_stmt (*def_stmt); | |
5916 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
5917 | } | |
5918 | ||
5919 | if (*dt == vect_unknown_def_type) | |
5920 | { | |
5921 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5922 | fprintf (vect_dump, "Unsupported pattern."); | |
5923 | return false; | |
5924 | } | |
5925 | ||
5926 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5927 | fprintf (vect_dump, "type of def: %d.",*dt); | |
5928 | ||
5929 | switch (gimple_code (*def_stmt)) | |
5930 | { | |
5931 | case GIMPLE_PHI: | |
5932 | *def = gimple_phi_result (*def_stmt); | |
5933 | break; | |
5934 | ||
5935 | case GIMPLE_ASSIGN: | |
5936 | *def = gimple_assign_lhs (*def_stmt); | |
5937 | break; | |
5938 | ||
5939 | case GIMPLE_CALL: | |
5940 | *def = gimple_call_lhs (*def_stmt); | |
5941 | if (*def != NULL) | |
5942 | break; | |
5943 | /* FALLTHRU */ | |
5944 | default: | |
5945 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5946 | fprintf (vect_dump, "unsupported defining stmt: "); | |
5947 | return false; | |
5948 | } | |
5949 | ||
5950 | return true; | |
5951 | } | |
5952 | ||
b690cc0f RG |
5953 | /* Function vect_is_simple_use_1. |
5954 | ||
5955 | Same as vect_is_simple_use_1 but also determines the vector operand | |
5956 | type of OPERAND and stores it to *VECTYPE. If the definition of | |
5957 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
5958 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
5959 | is responsible to compute the best suited vector type for the | |
5960 | scalar operand. */ | |
5961 | ||
5962 | bool | |
5963 | vect_is_simple_use_1 (tree operand, loop_vec_info loop_vinfo, | |
5964 | bb_vec_info bb_vinfo, gimple *def_stmt, | |
5965 | tree *def, enum vect_def_type *dt, tree *vectype) | |
5966 | { | |
5967 | if (!vect_is_simple_use (operand, loop_vinfo, bb_vinfo, def_stmt, def, dt)) | |
5968 | return false; | |
5969 | ||
5970 | /* Now get a vector type if the def is internal, otherwise supply | |
5971 | NULL_TREE and leave it up to the caller to figure out a proper | |
5972 | type for the use stmt. */ | |
5973 | if (*dt == vect_internal_def | |
5974 | || *dt == vect_induction_def | |
5975 | || *dt == vect_reduction_def | |
5976 | || *dt == vect_double_reduction_def | |
5977 | || *dt == vect_nested_cycle) | |
5978 | { | |
5979 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
5980 | |
5981 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
5982 | && !STMT_VINFO_RELEVANT (stmt_info) | |
5983 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 5984 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 5985 | |
b690cc0f RG |
5986 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
5987 | gcc_assert (*vectype != NULL_TREE); | |
5988 | } | |
5989 | else if (*dt == vect_uninitialized_def | |
5990 | || *dt == vect_constant_def | |
5991 | || *dt == vect_external_def) | |
5992 | *vectype = NULL_TREE; | |
5993 | else | |
5994 | gcc_unreachable (); | |
5995 | ||
5996 | return true; | |
5997 | } | |
5998 | ||
ebfd146a IR |
5999 | |
6000 | /* Function supportable_widening_operation | |
6001 | ||
b8698a0f L |
6002 | Check whether an operation represented by the code CODE is a |
6003 | widening operation that is supported by the target platform in | |
b690cc0f RG |
6004 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
6005 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 6006 | |
ebfd146a IR |
6007 | Widening operations we currently support are NOP (CONVERT), FLOAT |
6008 | and WIDEN_MULT. This function checks if these operations are supported | |
6009 | by the target platform either directly (via vector tree-codes), or via | |
6010 | target builtins. | |
6011 | ||
6012 | Output: | |
b8698a0f L |
6013 | - CODE1 and CODE2 are codes of vector operations to be used when |
6014 | vectorizing the operation, if available. | |
ebfd146a | 6015 | - DECL1 and DECL2 are decls of target builtin functions to be used |
ff802fa1 | 6016 | when vectorizing the operation, if available. In this case, |
b8698a0f | 6017 | CODE1 and CODE2 are CALL_EXPR. |
ebfd146a IR |
6018 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
6019 | case of multi-step conversion (like char->short->int - in that case | |
6020 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
6021 | - INTERM_TYPES contains the intermediate type required to perform the |
6022 | widening operation (short in the above example). */ | |
ebfd146a IR |
6023 | |
6024 | bool | |
b690cc0f RG |
6025 | supportable_widening_operation (enum tree_code code, gimple stmt, |
6026 | tree vectype_out, tree vectype_in, | |
ebfd146a IR |
6027 | tree *decl1, tree *decl2, |
6028 | enum tree_code *code1, enum tree_code *code2, | |
6029 | int *multi_step_cvt, | |
6030 | VEC (tree, heap) **interm_types) | |
6031 | { | |
6032 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6033 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 6034 | struct loop *vect_loop = NULL; |
ebfd146a IR |
6035 | bool ordered_p; |
6036 | enum machine_mode vec_mode; | |
81f40b79 | 6037 | enum insn_code icode1, icode2; |
ebfd146a | 6038 | optab optab1, optab2; |
b690cc0f RG |
6039 | tree vectype = vectype_in; |
6040 | tree wide_vectype = vectype_out; | |
ebfd146a | 6041 | enum tree_code c1, c2; |
4a00c761 JJ |
6042 | int i; |
6043 | tree prev_type, intermediate_type; | |
6044 | enum machine_mode intermediate_mode, prev_mode; | |
6045 | optab optab3, optab4; | |
ebfd146a | 6046 | |
4a00c761 | 6047 | *multi_step_cvt = 0; |
4ef69dfc IR |
6048 | if (loop_info) |
6049 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
6050 | ||
ebfd146a | 6051 | /* The result of a vectorized widening operation usually requires two vectors |
4a00c761 | 6052 | (because the widened results do not fit into one vector). The generated |
b8698a0f | 6053 | vector results would normally be expected to be generated in the same |
ebfd146a IR |
6054 | order as in the original scalar computation, i.e. if 8 results are |
6055 | generated in each vector iteration, they are to be organized as follows: | |
b8698a0f | 6056 | vect1: [res1,res2,res3,res4], vect2: [res5,res6,res7,res8]. |
ebfd146a | 6057 | |
b8698a0f | 6058 | However, in the special case that the result of the widening operation is |
ebfd146a | 6059 | used in a reduction computation only, the order doesn't matter (because |
b8698a0f | 6060 | when vectorizing a reduction we change the order of the computation). |
ebfd146a IR |
6061 | Some targets can take advantage of this and generate more efficient code. |
6062 | For example, targets like Altivec, that support widen_mult using a sequence | |
6063 | of {mult_even,mult_odd} generate the following vectors: | |
6064 | vect1: [res1,res3,res5,res7], vect2: [res2,res4,res6,res8]. | |
6065 | ||
6066 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
b8698a0f | 6067 | (each vectorized inner-loop iteration contributes to VF outer-loop |
ff802fa1 | 6068 | iterations in parallel). We therefore don't allow to change the order |
ebfd146a IR |
6069 | of the computation in the inner-loop during outer-loop vectorization. */ |
6070 | ||
4ef69dfc IR |
6071 | if (vect_loop |
6072 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
ebfd146a IR |
6073 | && !nested_in_vect_loop_p (vect_loop, stmt)) |
6074 | ordered_p = false; | |
6075 | else | |
6076 | ordered_p = true; | |
6077 | ||
6078 | if (!ordered_p | |
6079 | && code == WIDEN_MULT_EXPR | |
6080 | && targetm.vectorize.builtin_mul_widen_even | |
6081 | && targetm.vectorize.builtin_mul_widen_even (vectype) | |
6082 | && targetm.vectorize.builtin_mul_widen_odd | |
6083 | && targetm.vectorize.builtin_mul_widen_odd (vectype)) | |
6084 | { | |
6085 | if (vect_print_dump_info (REPORT_DETAILS)) | |
6086 | fprintf (vect_dump, "Unordered widening operation detected."); | |
6087 | ||
6088 | *code1 = *code2 = CALL_EXPR; | |
6089 | *decl1 = targetm.vectorize.builtin_mul_widen_even (vectype); | |
6090 | *decl2 = targetm.vectorize.builtin_mul_widen_odd (vectype); | |
6091 | return true; | |
6092 | } | |
6093 | ||
6094 | switch (code) | |
6095 | { | |
6096 | case WIDEN_MULT_EXPR: | |
4a00c761 JJ |
6097 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
6098 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
6099 | break; |
6100 | ||
36ba4aae | 6101 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
6102 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
6103 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
6104 | break; |
6105 | ||
ebfd146a | 6106 | CASE_CONVERT: |
4a00c761 JJ |
6107 | c1 = VEC_UNPACK_LO_EXPR; |
6108 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
6109 | break; |
6110 | ||
6111 | case FLOAT_EXPR: | |
4a00c761 JJ |
6112 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
6113 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
6114 | break; |
6115 | ||
6116 | case FIX_TRUNC_EXPR: | |
6117 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
6118 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
6119 | computing the operation. */ | |
6120 | return false; | |
6121 | ||
6122 | default: | |
6123 | gcc_unreachable (); | |
6124 | } | |
6125 | ||
4a00c761 JJ |
6126 | if (BYTES_BIG_ENDIAN) |
6127 | { | |
6128 | enum tree_code ctmp = c1; | |
6129 | c1 = c2; | |
6130 | c2 = ctmp; | |
6131 | } | |
6132 | ||
ebfd146a IR |
6133 | if (code == FIX_TRUNC_EXPR) |
6134 | { | |
6135 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
6136 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
6137 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
6138 | } |
6139 | else | |
6140 | { | |
6141 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
6142 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
6143 | } | |
6144 | ||
6145 | if (!optab1 || !optab2) | |
6146 | return false; | |
6147 | ||
6148 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
6149 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
6150 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
6151 | return false; |
6152 | ||
4a00c761 JJ |
6153 | *code1 = c1; |
6154 | *code2 = c2; | |
6155 | ||
6156 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
6157 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
6158 | return true; | |
6159 | ||
b8698a0f | 6160 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 6161 | types. */ |
ebfd146a | 6162 | |
4a00c761 JJ |
6163 | prev_type = vectype; |
6164 | prev_mode = vec_mode; | |
b8698a0f | 6165 | |
4a00c761 JJ |
6166 | if (!CONVERT_EXPR_CODE_P (code)) |
6167 | return false; | |
b8698a0f | 6168 | |
4a00c761 JJ |
6169 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
6170 | intermediate steps in promotion sequence. We try | |
6171 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
6172 | not. */ | |
6173 | *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS); | |
6174 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) | |
6175 | { | |
6176 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
6177 | intermediate_type | |
6178 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
6179 | TYPE_UNSIGNED (prev_type)); | |
6180 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
6181 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
6182 | ||
6183 | if (!optab3 || !optab4 | |
6184 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
6185 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
6186 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
6187 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
6188 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
6189 | == CODE_FOR_nothing) | |
6190 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
6191 | == CODE_FOR_nothing)) | |
6192 | break; | |
ebfd146a | 6193 | |
4a00c761 JJ |
6194 | VEC_quick_push (tree, *interm_types, intermediate_type); |
6195 | (*multi_step_cvt)++; | |
6196 | ||
6197 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
6198 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
6199 | return true; | |
6200 | ||
6201 | prev_type = intermediate_type; | |
6202 | prev_mode = intermediate_mode; | |
ebfd146a IR |
6203 | } |
6204 | ||
4a00c761 JJ |
6205 | VEC_free (tree, heap, *interm_types); |
6206 | return false; | |
ebfd146a IR |
6207 | } |
6208 | ||
6209 | ||
6210 | /* Function supportable_narrowing_operation | |
6211 | ||
b8698a0f L |
6212 | Check whether an operation represented by the code CODE is a |
6213 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
6214 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
6215 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 6216 | |
ebfd146a | 6217 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 6218 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
6219 | the target platform directly via vector tree-codes. |
6220 | ||
6221 | Output: | |
b8698a0f L |
6222 | - CODE1 is the code of a vector operation to be used when |
6223 | vectorizing the operation, if available. | |
ebfd146a IR |
6224 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
6225 | case of multi-step conversion (like int->short->char - in that case | |
6226 | MULTI_STEP_CVT will be 1). | |
6227 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 6228 | narrowing operation (short in the above example). */ |
ebfd146a IR |
6229 | |
6230 | bool | |
6231 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 6232 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
6233 | enum tree_code *code1, int *multi_step_cvt, |
6234 | VEC (tree, heap) **interm_types) | |
6235 | { | |
6236 | enum machine_mode vec_mode; | |
6237 | enum insn_code icode1; | |
6238 | optab optab1, interm_optab; | |
b690cc0f RG |
6239 | tree vectype = vectype_in; |
6240 | tree narrow_vectype = vectype_out; | |
ebfd146a | 6241 | enum tree_code c1; |
4a00c761 JJ |
6242 | tree intermediate_type; |
6243 | enum machine_mode intermediate_mode, prev_mode; | |
ebfd146a | 6244 | int i; |
4a00c761 | 6245 | bool uns; |
ebfd146a | 6246 | |
4a00c761 | 6247 | *multi_step_cvt = 0; |
ebfd146a IR |
6248 | switch (code) |
6249 | { | |
6250 | CASE_CONVERT: | |
6251 | c1 = VEC_PACK_TRUNC_EXPR; | |
6252 | break; | |
6253 | ||
6254 | case FIX_TRUNC_EXPR: | |
6255 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
6256 | break; | |
6257 | ||
6258 | case FLOAT_EXPR: | |
6259 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
6260 | tree code and optabs used for computing the operation. */ | |
6261 | return false; | |
6262 | ||
6263 | default: | |
6264 | gcc_unreachable (); | |
6265 | } | |
6266 | ||
6267 | if (code == FIX_TRUNC_EXPR) | |
6268 | /* The signedness is determined from output operand. */ | |
b690cc0f | 6269 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
6270 | else |
6271 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
6272 | ||
6273 | if (!optab1) | |
6274 | return false; | |
6275 | ||
6276 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 6277 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
6278 | return false; |
6279 | ||
4a00c761 JJ |
6280 | *code1 = c1; |
6281 | ||
6282 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
6283 | return true; | |
6284 | ||
ebfd146a IR |
6285 | /* Check if it's a multi-step conversion that can be done using intermediate |
6286 | types. */ | |
4a00c761 JJ |
6287 | prev_mode = vec_mode; |
6288 | if (code == FIX_TRUNC_EXPR) | |
6289 | uns = TYPE_UNSIGNED (vectype_out); | |
6290 | else | |
6291 | uns = TYPE_UNSIGNED (vectype); | |
6292 | ||
6293 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
6294 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
6295 | costly than signed. */ | |
6296 | if (code == FIX_TRUNC_EXPR && uns) | |
6297 | { | |
6298 | enum insn_code icode2; | |
6299 | ||
6300 | intermediate_type | |
6301 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
6302 | interm_optab | |
6303 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
6304 | if (interm_optab != NULL | |
6305 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing | |
6306 | && insn_data[icode1].operand[0].mode | |
6307 | == insn_data[icode2].operand[0].mode) | |
6308 | { | |
6309 | uns = false; | |
6310 | optab1 = interm_optab; | |
6311 | icode1 = icode2; | |
6312 | } | |
6313 | } | |
ebfd146a | 6314 | |
4a00c761 JJ |
6315 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
6316 | intermediate steps in promotion sequence. We try | |
6317 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
6318 | *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS); | |
6319 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) | |
6320 | { | |
6321 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
6322 | intermediate_type | |
6323 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
6324 | interm_optab | |
6325 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
6326 | optab_default); | |
6327 | if (!interm_optab | |
6328 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
6329 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
6330 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
6331 | == CODE_FOR_nothing)) | |
6332 | break; | |
6333 | ||
6334 | VEC_quick_push (tree, *interm_types, intermediate_type); | |
6335 | (*multi_step_cvt)++; | |
6336 | ||
6337 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
6338 | return true; | |
6339 | ||
6340 | prev_mode = intermediate_mode; | |
6341 | optab1 = interm_optab; | |
ebfd146a IR |
6342 | } |
6343 | ||
4a00c761 JJ |
6344 | VEC_free (tree, heap, *interm_types); |
6345 | return false; | |
ebfd146a | 6346 | } |