]>
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. */ | |
162 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs) | |
163 | { | |
164 | if (is_gimple_debug (USE_STMT (use_p))) | |
165 | continue; | |
166 | use_stmt = USE_STMT (use_p); | |
ebfd146a | 167 | |
83197f37 IR |
168 | if (vinfo_for_stmt (use_stmt) |
169 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt))) | |
170 | { | |
171 | found = true; | |
172 | break; | |
173 | } | |
174 | } | |
175 | } | |
176 | ||
177 | if (!found) | |
178 | { | |
179 | /* This is the last stmt in a sequence that was detected as a | |
180 | pattern that can potentially be vectorized. Don't mark the stmt | |
181 | as relevant/live because it's not going to be vectorized. | |
182 | Instead mark the pattern-stmt that replaces it. */ | |
183 | ||
184 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
185 | ||
186 | if (vect_print_dump_info (REPORT_DETAILS)) | |
187 | fprintf (vect_dump, "last stmt in pattern. don't mark" | |
188 | " relevant/live."); | |
189 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
190 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
191 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
192 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
193 | stmt = pattern_stmt; | |
194 | } | |
ebfd146a IR |
195 | } |
196 | ||
197 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
198 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
199 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
200 | ||
201 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
202 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
203 | { | |
204 | if (vect_print_dump_info (REPORT_DETAILS)) | |
205 | fprintf (vect_dump, "already marked relevant/live."); | |
206 | return; | |
207 | } | |
208 | ||
209 | VEC_safe_push (gimple, heap, *worklist, stmt); | |
210 | } | |
211 | ||
212 | ||
213 | /* Function vect_stmt_relevant_p. | |
214 | ||
215 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
216 | "relevant for vectorization". | |
217 | ||
218 | A stmt is considered "relevant for vectorization" if: | |
219 | - it has uses outside the loop. | |
220 | - it has vdefs (it alters memory). | |
221 | - control stmts in the loop (except for the exit condition). | |
222 | ||
223 | CHECKME: what other side effects would the vectorizer allow? */ | |
224 | ||
225 | static bool | |
226 | vect_stmt_relevant_p (gimple stmt, loop_vec_info loop_vinfo, | |
227 | enum vect_relevant *relevant, bool *live_p) | |
228 | { | |
229 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
230 | ssa_op_iter op_iter; | |
231 | imm_use_iterator imm_iter; | |
232 | use_operand_p use_p; | |
233 | def_operand_p def_p; | |
234 | ||
8644a673 | 235 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
236 | *live_p = false; |
237 | ||
238 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
239 | if (is_ctrl_stmt (stmt) |
240 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
241 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 242 | *relevant = vect_used_in_scope; |
ebfd146a IR |
243 | |
244 | /* changing memory. */ | |
245 | if (gimple_code (stmt) != GIMPLE_PHI) | |
5006671f | 246 | if (gimple_vdef (stmt)) |
ebfd146a IR |
247 | { |
248 | if (vect_print_dump_info (REPORT_DETAILS)) | |
249 | fprintf (vect_dump, "vec_stmt_relevant_p: stmt has vdefs."); | |
8644a673 | 250 | *relevant = vect_used_in_scope; |
ebfd146a IR |
251 | } |
252 | ||
253 | /* uses outside the loop. */ | |
254 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
255 | { | |
256 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
257 | { | |
258 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
259 | if (!flow_bb_inside_loop_p (loop, bb)) | |
260 | { | |
261 | if (vect_print_dump_info (REPORT_DETAILS)) | |
262 | fprintf (vect_dump, "vec_stmt_relevant_p: used out of loop."); | |
263 | ||
3157b0c2 AO |
264 | if (is_gimple_debug (USE_STMT (use_p))) |
265 | continue; | |
266 | ||
ebfd146a IR |
267 | /* We expect all such uses to be in the loop exit phis |
268 | (because of loop closed form) */ | |
269 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
270 | gcc_assert (bb == single_exit (loop)->dest); | |
271 | ||
272 | *live_p = true; | |
273 | } | |
274 | } | |
275 | } | |
276 | ||
277 | return (*live_p || *relevant); | |
278 | } | |
279 | ||
280 | ||
b8698a0f | 281 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 282 | |
ff802fa1 | 283 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
284 | used in STMT for anything other than indexing an array. */ |
285 | ||
286 | static bool | |
287 | exist_non_indexing_operands_for_use_p (tree use, gimple stmt) | |
288 | { | |
289 | tree operand; | |
290 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 291 | |
ff802fa1 | 292 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
293 | reference in STMT, then any operand that corresponds to USE |
294 | is not indexing an array. */ | |
295 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
296 | return true; | |
59a05b0c | 297 | |
ebfd146a IR |
298 | /* STMT has a data_ref. FORNOW this means that its of one of |
299 | the following forms: | |
300 | -1- ARRAY_REF = var | |
301 | -2- var = ARRAY_REF | |
302 | (This should have been verified in analyze_data_refs). | |
303 | ||
304 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 305 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
306 | for array indexing. |
307 | ||
308 | Therefore, all we need to check is if STMT falls into the | |
309 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
310 | |
311 | if (!gimple_assign_copy_p (stmt)) | |
312 | return false; | |
59a05b0c EB |
313 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
314 | return false; | |
ebfd146a | 315 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
316 | if (TREE_CODE (operand) != SSA_NAME) |
317 | return false; | |
318 | ||
319 | if (operand == use) | |
320 | return true; | |
321 | ||
322 | return false; | |
323 | } | |
324 | ||
325 | ||
b8698a0f | 326 | /* |
ebfd146a IR |
327 | Function process_use. |
328 | ||
329 | Inputs: | |
330 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 331 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ff802fa1 | 332 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a IR |
333 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
334 | ||
335 | Outputs: | |
336 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
337 | relevance info of the DEF_STMT of this USE: | |
338 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
339 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
340 | Exceptions: | |
341 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 342 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 343 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
344 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
345 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
346 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
347 | be modified accordingly. | |
348 | ||
349 | Return true if everything is as expected. Return false otherwise. */ | |
350 | ||
351 | static bool | |
b8698a0f | 352 | process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
ebfd146a IR |
353 | enum vect_relevant relevant, VEC(gimple,heap) **worklist) |
354 | { | |
355 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
356 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
357 | stmt_vec_info dstmt_vinfo; | |
358 | basic_block bb, def_bb; | |
359 | tree def; | |
360 | gimple def_stmt; | |
361 | enum vect_def_type dt; | |
362 | ||
b8698a0f | 363 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a IR |
364 | that are used for address computation are not considered relevant. */ |
365 | if (!exist_non_indexing_operands_for_use_p (use, stmt)) | |
366 | return true; | |
367 | ||
a70d6342 | 368 | if (!vect_is_simple_use (use, loop_vinfo, NULL, &def_stmt, &def, &dt)) |
b8698a0f | 369 | { |
8644a673 | 370 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
ebfd146a IR |
371 | fprintf (vect_dump, "not vectorized: unsupported use in stmt."); |
372 | return false; | |
373 | } | |
374 | ||
375 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
376 | return true; | |
377 | ||
378 | def_bb = gimple_bb (def_stmt); | |
379 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
380 | { | |
381 | if (vect_print_dump_info (REPORT_DETAILS)) | |
382 | fprintf (vect_dump, "def_stmt is out of loop."); | |
383 | return true; | |
384 | } | |
385 | ||
b8698a0f L |
386 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
387 | DEF_STMT must have already been processed, because this should be the | |
388 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
389 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
390 | check that everything is as expected, and we are done. */ |
391 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
392 | bb = gimple_bb (stmt); | |
393 | if (gimple_code (stmt) == GIMPLE_PHI | |
394 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
395 | && gimple_code (def_stmt) != GIMPLE_PHI | |
396 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
397 | && bb->loop_father == def_bb->loop_father) | |
398 | { | |
399 | if (vect_print_dump_info (REPORT_DETAILS)) | |
400 | fprintf (vect_dump, "reduc-stmt defining reduc-phi in the same nest."); | |
401 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) | |
402 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
403 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 404 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 405 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
406 | return true; |
407 | } | |
408 | ||
409 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
410 | outer-loop-header-bb: | |
411 | d = def_stmt | |
412 | inner-loop: | |
413 | stmt # use (d) | |
414 | outer-loop-tail-bb: | |
415 | ... */ | |
416 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
417 | { | |
418 | if (vect_print_dump_info (REPORT_DETAILS)) | |
419 | fprintf (vect_dump, "outer-loop def-stmt defining inner-loop stmt."); | |
7c5222ff | 420 | |
ebfd146a IR |
421 | switch (relevant) |
422 | { | |
8644a673 | 423 | case vect_unused_in_scope: |
7c5222ff IR |
424 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
425 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 426 | break; |
7c5222ff | 427 | |
ebfd146a | 428 | case vect_used_in_outer_by_reduction: |
7c5222ff | 429 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
430 | relevant = vect_used_by_reduction; |
431 | break; | |
7c5222ff | 432 | |
ebfd146a | 433 | case vect_used_in_outer: |
7c5222ff | 434 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 435 | relevant = vect_used_in_scope; |
ebfd146a | 436 | break; |
7c5222ff | 437 | |
8644a673 | 438 | case vect_used_in_scope: |
ebfd146a IR |
439 | break; |
440 | ||
441 | default: | |
442 | gcc_unreachable (); | |
b8698a0f | 443 | } |
ebfd146a IR |
444 | } |
445 | ||
446 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
447 | outer-loop-header-bb: | |
448 | ... | |
449 | inner-loop: | |
450 | d = def_stmt | |
06066f92 | 451 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
452 | stmt # use (d) */ |
453 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
454 | { | |
455 | if (vect_print_dump_info (REPORT_DETAILS)) | |
456 | fprintf (vect_dump, "inner-loop def-stmt defining outer-loop stmt."); | |
7c5222ff | 457 | |
ebfd146a IR |
458 | switch (relevant) |
459 | { | |
8644a673 | 460 | case vect_unused_in_scope: |
b8698a0f | 461 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 462 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 463 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
464 | break; |
465 | ||
ebfd146a IR |
466 | case vect_used_by_reduction: |
467 | relevant = vect_used_in_outer_by_reduction; | |
468 | break; | |
469 | ||
8644a673 | 470 | case vect_used_in_scope: |
ebfd146a IR |
471 | relevant = vect_used_in_outer; |
472 | break; | |
473 | ||
474 | default: | |
475 | gcc_unreachable (); | |
476 | } | |
477 | } | |
478 | ||
83197f37 IR |
479 | vect_mark_relevant (worklist, def_stmt, relevant, live_p, |
480 | is_pattern_stmt_p (stmt_vinfo)); | |
ebfd146a IR |
481 | return true; |
482 | } | |
483 | ||
484 | ||
485 | /* Function vect_mark_stmts_to_be_vectorized. | |
486 | ||
487 | Not all stmts in the loop need to be vectorized. For example: | |
488 | ||
489 | for i... | |
490 | for j... | |
491 | 1. T0 = i + j | |
492 | 2. T1 = a[T0] | |
493 | ||
494 | 3. j = j + 1 | |
495 | ||
496 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
497 | addressing of vectorized data-refs are handled differently. | |
498 | ||
499 | This pass detects such stmts. */ | |
500 | ||
501 | bool | |
502 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
503 | { | |
504 | VEC(gimple,heap) *worklist; | |
505 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
506 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
507 | unsigned int nbbs = loop->num_nodes; | |
508 | gimple_stmt_iterator si; | |
509 | gimple stmt; | |
510 | unsigned int i; | |
511 | stmt_vec_info stmt_vinfo; | |
512 | basic_block bb; | |
513 | gimple phi; | |
514 | bool live_p; | |
06066f92 IR |
515 | enum vect_relevant relevant, tmp_relevant; |
516 | enum vect_def_type def_type; | |
ebfd146a IR |
517 | |
518 | if (vect_print_dump_info (REPORT_DETAILS)) | |
519 | fprintf (vect_dump, "=== vect_mark_stmts_to_be_vectorized ==="); | |
520 | ||
521 | worklist = VEC_alloc (gimple, heap, 64); | |
522 | ||
523 | /* 1. Init worklist. */ | |
524 | for (i = 0; i < nbbs; i++) | |
525 | { | |
526 | bb = bbs[i]; | |
527 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 528 | { |
ebfd146a IR |
529 | phi = gsi_stmt (si); |
530 | if (vect_print_dump_info (REPORT_DETAILS)) | |
531 | { | |
532 | fprintf (vect_dump, "init: phi relevant? "); | |
533 | print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM); | |
534 | } | |
535 | ||
536 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 537 | vect_mark_relevant (&worklist, phi, relevant, live_p, false); |
ebfd146a IR |
538 | } |
539 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
540 | { | |
541 | stmt = gsi_stmt (si); | |
542 | if (vect_print_dump_info (REPORT_DETAILS)) | |
543 | { | |
544 | fprintf (vect_dump, "init: stmt relevant? "); | |
545 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
b8698a0f | 546 | } |
ebfd146a IR |
547 | |
548 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 549 | vect_mark_relevant (&worklist, stmt, relevant, live_p, false); |
ebfd146a IR |
550 | } |
551 | } | |
552 | ||
553 | /* 2. Process_worklist */ | |
554 | while (VEC_length (gimple, worklist) > 0) | |
555 | { | |
556 | use_operand_p use_p; | |
557 | ssa_op_iter iter; | |
558 | ||
559 | stmt = VEC_pop (gimple, worklist); | |
560 | if (vect_print_dump_info (REPORT_DETAILS)) | |
561 | { | |
562 | fprintf (vect_dump, "worklist: examine stmt: "); | |
563 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
564 | } | |
565 | ||
b8698a0f L |
566 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
567 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
568 | liveness and relevance properties of STMT. */ |
569 | stmt_vinfo = vinfo_for_stmt (stmt); | |
570 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
571 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
572 | ||
573 | /* Generally, the liveness and relevance properties of STMT are | |
574 | propagated as is to the DEF_STMTs of its USEs: | |
575 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
576 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
577 | ||
578 | One exception is when STMT has been identified as defining a reduction | |
579 | variable; in this case we set the liveness/relevance as follows: | |
580 | live_p = false | |
581 | relevant = vect_used_by_reduction | |
582 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f | 583 | those that are used by a reduction computation, and those that are |
ff802fa1 | 584 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 585 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 586 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 587 | |
06066f92 IR |
588 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
589 | tmp_relevant = relevant; | |
590 | switch (def_type) | |
ebfd146a | 591 | { |
06066f92 IR |
592 | case vect_reduction_def: |
593 | switch (tmp_relevant) | |
594 | { | |
595 | case vect_unused_in_scope: | |
596 | relevant = vect_used_by_reduction; | |
597 | break; | |
598 | ||
599 | case vect_used_by_reduction: | |
600 | if (gimple_code (stmt) == GIMPLE_PHI) | |
601 | break; | |
602 | /* fall through */ | |
603 | ||
604 | default: | |
605 | if (vect_print_dump_info (REPORT_DETAILS)) | |
606 | fprintf (vect_dump, "unsupported use of reduction."); | |
607 | ||
608 | VEC_free (gimple, heap, worklist); | |
609 | return false; | |
610 | } | |
611 | ||
b8698a0f | 612 | live_p = false; |
06066f92 | 613 | break; |
b8698a0f | 614 | |
06066f92 IR |
615 | case vect_nested_cycle: |
616 | if (tmp_relevant != vect_unused_in_scope | |
617 | && tmp_relevant != vect_used_in_outer_by_reduction | |
618 | && tmp_relevant != vect_used_in_outer) | |
619 | { | |
620 | if (vect_print_dump_info (REPORT_DETAILS)) | |
621 | fprintf (vect_dump, "unsupported use of nested cycle."); | |
7c5222ff | 622 | |
06066f92 IR |
623 | VEC_free (gimple, heap, worklist); |
624 | return false; | |
625 | } | |
7c5222ff | 626 | |
b8698a0f L |
627 | live_p = false; |
628 | break; | |
629 | ||
06066f92 IR |
630 | case vect_double_reduction_def: |
631 | if (tmp_relevant != vect_unused_in_scope | |
632 | && tmp_relevant != vect_used_by_reduction) | |
633 | { | |
7c5222ff | 634 | if (vect_print_dump_info (REPORT_DETAILS)) |
06066f92 | 635 | fprintf (vect_dump, "unsupported use of double reduction."); |
7c5222ff IR |
636 | |
637 | VEC_free (gimple, heap, worklist); | |
638 | return false; | |
06066f92 IR |
639 | } |
640 | ||
641 | live_p = false; | |
b8698a0f | 642 | break; |
7c5222ff | 643 | |
06066f92 IR |
644 | default: |
645 | break; | |
7c5222ff | 646 | } |
b8698a0f | 647 | |
9d5e7640 IR |
648 | if (is_pattern_stmt_p (vinfo_for_stmt (stmt))) |
649 | { | |
650 | /* Pattern statements are not inserted into the code, so | |
651 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
652 | have to scan the RHS or function arguments instead. */ | |
653 | if (is_gimple_assign (stmt)) | |
654 | { | |
655 | for (i = 1; i < gimple_num_ops (stmt); i++) | |
656 | { | |
657 | tree op = gimple_op (stmt, i); | |
658 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, | |
659 | &worklist)) | |
660 | { | |
661 | VEC_free (gimple, heap, worklist); | |
662 | return false; | |
663 | } | |
664 | } | |
665 | } | |
666 | else if (is_gimple_call (stmt)) | |
667 | { | |
668 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
669 | { | |
670 | tree arg = gimple_call_arg (stmt, i); | |
671 | if (!process_use (stmt, arg, loop_vinfo, live_p, relevant, | |
672 | &worklist)) | |
673 | { | |
674 | VEC_free (gimple, heap, worklist); | |
675 | return false; | |
676 | } | |
677 | } | |
678 | } | |
679 | } | |
680 | else | |
681 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
682 | { | |
683 | tree op = USE_FROM_PTR (use_p); | |
684 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, | |
685 | &worklist)) | |
686 | { | |
687 | VEC_free (gimple, heap, worklist); | |
688 | return false; | |
689 | } | |
690 | } | |
ebfd146a IR |
691 | } /* while worklist */ |
692 | ||
693 | VEC_free (gimple, heap, worklist); | |
694 | return true; | |
695 | } | |
696 | ||
697 | ||
720f5239 IR |
698 | /* Get cost by calling cost target builtin. */ |
699 | ||
700 | static inline | |
701 | int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost) | |
702 | { | |
703 | tree dummy_type = NULL; | |
704 | int dummy = 0; | |
705 | ||
706 | return targetm.vectorize.builtin_vectorization_cost (type_of_cost, | |
707 | dummy_type, dummy); | |
708 | } | |
709 | ||
ff802fa1 IR |
710 | |
711 | /* Get cost for STMT. */ | |
712 | ||
ebfd146a IR |
713 | int |
714 | cost_for_stmt (gimple stmt) | |
715 | { | |
716 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
717 | ||
718 | switch (STMT_VINFO_TYPE (stmt_info)) | |
719 | { | |
720 | case load_vec_info_type: | |
720f5239 | 721 | return vect_get_stmt_cost (scalar_load); |
ebfd146a | 722 | case store_vec_info_type: |
720f5239 | 723 | return vect_get_stmt_cost (scalar_store); |
ebfd146a IR |
724 | case op_vec_info_type: |
725 | case condition_vec_info_type: | |
726 | case assignment_vec_info_type: | |
727 | case reduc_vec_info_type: | |
728 | case induc_vec_info_type: | |
729 | case type_promotion_vec_info_type: | |
730 | case type_demotion_vec_info_type: | |
731 | case type_conversion_vec_info_type: | |
732 | case call_vec_info_type: | |
720f5239 | 733 | return vect_get_stmt_cost (scalar_stmt); |
ebfd146a IR |
734 | case undef_vec_info_type: |
735 | default: | |
736 | gcc_unreachable (); | |
737 | } | |
738 | } | |
739 | ||
b8698a0f | 740 | /* Function vect_model_simple_cost. |
ebfd146a | 741 | |
b8698a0f | 742 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
743 | single op. Right now, this does not account for multiple insns that could |
744 | be generated for the single vector op. We will handle that shortly. */ | |
745 | ||
746 | void | |
b8698a0f | 747 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
ebfd146a IR |
748 | enum vect_def_type *dt, slp_tree slp_node) |
749 | { | |
750 | int i; | |
751 | int inside_cost = 0, outside_cost = 0; | |
752 | ||
753 | /* The SLP costs were already calculated during SLP tree build. */ | |
754 | if (PURE_SLP_STMT (stmt_info)) | |
755 | return; | |
756 | ||
720f5239 | 757 | inside_cost = ncopies * vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
758 | |
759 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
760 | for (i = 0; i < 2; i++) | |
761 | { | |
8644a673 | 762 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
720f5239 | 763 | outside_cost += vect_get_stmt_cost (vector_stmt); |
ebfd146a | 764 | } |
b8698a0f | 765 | |
ebfd146a IR |
766 | if (vect_print_dump_info (REPORT_COST)) |
767 | fprintf (vect_dump, "vect_model_simple_cost: inside_cost = %d, " | |
768 | "outside_cost = %d .", inside_cost, outside_cost); | |
769 | ||
770 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
771 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
772 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
773 | } | |
774 | ||
775 | ||
b8698a0f L |
776 | /* Function vect_cost_strided_group_size |
777 | ||
ebfd146a IR |
778 | For strided load or store, return the group_size only if it is the first |
779 | load or store of a group, else return 1. This ensures that group size is | |
780 | only returned once per group. */ | |
781 | ||
782 | static int | |
783 | vect_cost_strided_group_size (stmt_vec_info stmt_info) | |
784 | { | |
e14c1050 | 785 | gimple first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
786 | |
787 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
e14c1050 | 788 | return GROUP_SIZE (stmt_info); |
ebfd146a IR |
789 | |
790 | return 1; | |
791 | } | |
792 | ||
793 | ||
794 | /* Function vect_model_store_cost | |
795 | ||
796 | Models cost for stores. In the case of strided accesses, one access | |
797 | has the overhead of the strided access attributed to it. */ | |
798 | ||
799 | void | |
b8698a0f | 800 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
272c6793 RS |
801 | bool store_lanes_p, enum vect_def_type dt, |
802 | slp_tree slp_node) | |
ebfd146a IR |
803 | { |
804 | int group_size; | |
720f5239 IR |
805 | unsigned int inside_cost = 0, outside_cost = 0; |
806 | struct data_reference *first_dr; | |
807 | gimple first_stmt; | |
ebfd146a IR |
808 | |
809 | /* The SLP costs were already calculated during SLP tree build. */ | |
810 | if (PURE_SLP_STMT (stmt_info)) | |
811 | return; | |
812 | ||
8644a673 | 813 | if (dt == vect_constant_def || dt == vect_external_def) |
720f5239 | 814 | outside_cost = vect_get_stmt_cost (scalar_to_vec); |
ebfd146a IR |
815 | |
816 | /* Strided access? */ | |
e14c1050 | 817 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) |
720f5239 IR |
818 | { |
819 | if (slp_node) | |
820 | { | |
821 | first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0); | |
822 | group_size = 1; | |
823 | } | |
824 | else | |
825 | { | |
e14c1050 | 826 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
720f5239 IR |
827 | group_size = vect_cost_strided_group_size (stmt_info); |
828 | } | |
829 | ||
830 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
831 | } | |
ebfd146a IR |
832 | /* Not a strided access. */ |
833 | else | |
720f5239 IR |
834 | { |
835 | group_size = 1; | |
836 | first_dr = STMT_VINFO_DATA_REF (stmt_info); | |
837 | } | |
ebfd146a | 838 | |
272c6793 RS |
839 | /* We assume that the cost of a single store-lanes instruction is |
840 | equivalent to the cost of GROUP_SIZE separate stores. If a strided | |
841 | access is instead being provided by a permute-and-store operation, | |
842 | include the cost of the permutes. */ | |
843 | if (!store_lanes_p && group_size > 1) | |
ebfd146a IR |
844 | { |
845 | /* Uses a high and low interleave operation for each needed permute. */ | |
b8698a0f | 846 | inside_cost = ncopies * exact_log2(group_size) * group_size |
720f5239 | 847 | * vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
848 | |
849 | if (vect_print_dump_info (REPORT_COST)) | |
850 | fprintf (vect_dump, "vect_model_store_cost: strided group_size = %d .", | |
851 | group_size); | |
852 | ||
853 | } | |
854 | ||
855 | /* Costs of the stores. */ | |
720f5239 | 856 | vect_get_store_cost (first_dr, ncopies, &inside_cost); |
ebfd146a IR |
857 | |
858 | if (vect_print_dump_info (REPORT_COST)) | |
859 | fprintf (vect_dump, "vect_model_store_cost: inside_cost = %d, " | |
860 | "outside_cost = %d .", inside_cost, outside_cost); | |
861 | ||
862 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
863 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
864 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
865 | } | |
866 | ||
867 | ||
720f5239 IR |
868 | /* Calculate cost of DR's memory access. */ |
869 | void | |
870 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
871 | unsigned int *inside_cost) | |
872 | { | |
873 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
874 | ||
875 | switch (alignment_support_scheme) | |
876 | { | |
877 | case dr_aligned: | |
878 | { | |
879 | *inside_cost += ncopies * vect_get_stmt_cost (vector_store); | |
880 | ||
881 | if (vect_print_dump_info (REPORT_COST)) | |
882 | fprintf (vect_dump, "vect_model_store_cost: aligned."); | |
883 | ||
884 | break; | |
885 | } | |
886 | ||
887 | case dr_unaligned_supported: | |
888 | { | |
889 | gimple stmt = DR_STMT (dr); | |
890 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
891 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
892 | ||
893 | /* Here, we assign an additional cost for the unaligned store. */ | |
894 | *inside_cost += ncopies | |
895 | * targetm.vectorize.builtin_vectorization_cost (unaligned_store, | |
896 | vectype, DR_MISALIGNMENT (dr)); | |
897 | ||
898 | if (vect_print_dump_info (REPORT_COST)) | |
899 | fprintf (vect_dump, "vect_model_store_cost: unaligned supported by " | |
900 | "hardware."); | |
901 | ||
902 | break; | |
903 | } | |
904 | ||
905 | default: | |
906 | gcc_unreachable (); | |
907 | } | |
908 | } | |
909 | ||
910 | ||
ebfd146a IR |
911 | /* Function vect_model_load_cost |
912 | ||
913 | Models cost for loads. In the case of strided accesses, the last access | |
914 | has the overhead of the strided access attributed to it. Since unaligned | |
b8698a0f | 915 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
916 | access scheme chosen. */ |
917 | ||
918 | void | |
272c6793 RS |
919 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, bool load_lanes_p, |
920 | slp_tree slp_node) | |
ebfd146a IR |
921 | { |
922 | int group_size; | |
ebfd146a IR |
923 | gimple first_stmt; |
924 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; | |
720f5239 | 925 | unsigned int inside_cost = 0, outside_cost = 0; |
ebfd146a IR |
926 | |
927 | /* The SLP costs were already calculated during SLP tree build. */ | |
928 | if (PURE_SLP_STMT (stmt_info)) | |
929 | return; | |
930 | ||
931 | /* Strided accesses? */ | |
e14c1050 IR |
932 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
933 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) && first_stmt && !slp_node) | |
ebfd146a IR |
934 | { |
935 | group_size = vect_cost_strided_group_size (stmt_info); | |
936 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
937 | } | |
938 | /* Not a strided access. */ | |
939 | else | |
940 | { | |
941 | group_size = 1; | |
942 | first_dr = dr; | |
943 | } | |
944 | ||
272c6793 RS |
945 | /* We assume that the cost of a single load-lanes instruction is |
946 | equivalent to the cost of GROUP_SIZE separate loads. If a strided | |
947 | access is instead being provided by a load-and-permute operation, | |
948 | include the cost of the permutes. */ | |
949 | if (!load_lanes_p && group_size > 1) | |
ebfd146a IR |
950 | { |
951 | /* Uses an even and odd extract operations for each needed permute. */ | |
952 | inside_cost = ncopies * exact_log2(group_size) * group_size | |
720f5239 | 953 | * vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
954 | |
955 | if (vect_print_dump_info (REPORT_COST)) | |
956 | fprintf (vect_dump, "vect_model_load_cost: strided group_size = %d .", | |
957 | group_size); | |
ebfd146a IR |
958 | } |
959 | ||
960 | /* The loads themselves. */ | |
720f5239 | 961 | vect_get_load_cost (first_dr, ncopies, |
e14c1050 IR |
962 | ((!STMT_VINFO_STRIDED_ACCESS (stmt_info)) || group_size > 1 |
963 | || slp_node), | |
720f5239 IR |
964 | &inside_cost, &outside_cost); |
965 | ||
966 | if (vect_print_dump_info (REPORT_COST)) | |
967 | fprintf (vect_dump, "vect_model_load_cost: inside_cost = %d, " | |
968 | "outside_cost = %d .", inside_cost, outside_cost); | |
969 | ||
970 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
971 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
972 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
973 | } | |
974 | ||
975 | ||
976 | /* Calculate cost of DR's memory access. */ | |
977 | void | |
978 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
979 | bool add_realign_cost, unsigned int *inside_cost, | |
980 | unsigned int *outside_cost) | |
981 | { | |
982 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
983 | ||
984 | switch (alignment_support_scheme) | |
ebfd146a IR |
985 | { |
986 | case dr_aligned: | |
987 | { | |
9940b13c | 988 | *inside_cost += ncopies * vect_get_stmt_cost (vector_load); |
ebfd146a IR |
989 | |
990 | if (vect_print_dump_info (REPORT_COST)) | |
991 | fprintf (vect_dump, "vect_model_load_cost: aligned."); | |
992 | ||
993 | break; | |
994 | } | |
995 | case dr_unaligned_supported: | |
996 | { | |
720f5239 IR |
997 | gimple stmt = DR_STMT (dr); |
998 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
999 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
ebfd146a | 1000 | |
720f5239 IR |
1001 | /* Here, we assign an additional cost for the unaligned load. */ |
1002 | *inside_cost += ncopies | |
1003 | * targetm.vectorize.builtin_vectorization_cost (unaligned_load, | |
1004 | vectype, DR_MISALIGNMENT (dr)); | |
ebfd146a IR |
1005 | if (vect_print_dump_info (REPORT_COST)) |
1006 | fprintf (vect_dump, "vect_model_load_cost: unaligned supported by " | |
1007 | "hardware."); | |
1008 | ||
1009 | break; | |
1010 | } | |
1011 | case dr_explicit_realign: | |
1012 | { | |
720f5239 IR |
1013 | *inside_cost += ncopies * (2 * vect_get_stmt_cost (vector_load) |
1014 | + vect_get_stmt_cost (vector_stmt)); | |
ebfd146a IR |
1015 | |
1016 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1017 | the containing loop, the following cost should be added to the | |
1018 | outside costs. */ | |
1019 | if (targetm.vectorize.builtin_mask_for_load) | |
720f5239 | 1020 | *inside_cost += vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
1021 | |
1022 | break; | |
1023 | } | |
1024 | case dr_explicit_realign_optimized: | |
1025 | { | |
1026 | if (vect_print_dump_info (REPORT_COST)) | |
1027 | fprintf (vect_dump, "vect_model_load_cost: unaligned software " | |
1028 | "pipelined."); | |
1029 | ||
1030 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1031 | load, and possibly a mask operation to "prime" the loop. However, |
ebfd146a IR |
1032 | if this is an access in a group of loads, which provide strided |
1033 | access, then the above cost should only be considered for one | |
ff802fa1 | 1034 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1035 | and a realignment op. */ |
1036 | ||
720f5239 | 1037 | if (add_realign_cost) |
ebfd146a | 1038 | { |
720f5239 | 1039 | *outside_cost = 2 * vect_get_stmt_cost (vector_stmt); |
ebfd146a | 1040 | if (targetm.vectorize.builtin_mask_for_load) |
720f5239 | 1041 | *outside_cost += vect_get_stmt_cost (vector_stmt); |
ebfd146a IR |
1042 | } |
1043 | ||
720f5239 IR |
1044 | *inside_cost += ncopies * (vect_get_stmt_cost (vector_load) |
1045 | + vect_get_stmt_cost (vector_stmt)); | |
ebfd146a IR |
1046 | break; |
1047 | } | |
1048 | ||
1049 | default: | |
1050 | gcc_unreachable (); | |
1051 | } | |
ebfd146a IR |
1052 | } |
1053 | ||
1054 | ||
1055 | /* Function vect_init_vector. | |
1056 | ||
1057 | Insert a new stmt (INIT_STMT) that initializes a new vector variable with | |
ff802fa1 IR |
1058 | the vector elements of VECTOR_VAR. Place the initialization at BSI if it |
1059 | is not NULL. Otherwise, place the initialization at the loop preheader. | |
b8698a0f | 1060 | Return the DEF of INIT_STMT. |
ebfd146a IR |
1061 | It will be used in the vectorization of STMT. */ |
1062 | ||
1063 | tree | |
1064 | vect_init_vector (gimple stmt, tree vector_var, tree vector_type, | |
1065 | gimple_stmt_iterator *gsi) | |
1066 | { | |
1067 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
1068 | tree new_var; | |
1069 | gimple init_stmt; | |
1070 | tree vec_oprnd; | |
1071 | edge pe; | |
1072 | tree new_temp; | |
1073 | basic_block new_bb; | |
b8698a0f | 1074 | |
ebfd146a | 1075 | new_var = vect_get_new_vect_var (vector_type, vect_simple_var, "cst_"); |
b8698a0f | 1076 | add_referenced_var (new_var); |
ebfd146a IR |
1077 | init_stmt = gimple_build_assign (new_var, vector_var); |
1078 | new_temp = make_ssa_name (new_var, init_stmt); | |
1079 | gimple_assign_set_lhs (init_stmt, new_temp); | |
1080 | ||
1081 | if (gsi) | |
1082 | vect_finish_stmt_generation (stmt, init_stmt, gsi); | |
1083 | else | |
1084 | { | |
1085 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
b8698a0f | 1086 | |
a70d6342 IR |
1087 | if (loop_vinfo) |
1088 | { | |
1089 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1090 | ||
1091 | if (nested_in_vect_loop_p (loop, stmt)) | |
1092 | loop = loop->inner; | |
b8698a0f | 1093 | |
a70d6342 IR |
1094 | pe = loop_preheader_edge (loop); |
1095 | new_bb = gsi_insert_on_edge_immediate (pe, init_stmt); | |
1096 | gcc_assert (!new_bb); | |
1097 | } | |
1098 | else | |
1099 | { | |
1100 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1101 | basic_block bb; | |
1102 | gimple_stmt_iterator gsi_bb_start; | |
1103 | ||
1104 | gcc_assert (bb_vinfo); | |
1105 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1106 | gsi_bb_start = gsi_after_labels (bb); |
a70d6342 IR |
1107 | gsi_insert_before (&gsi_bb_start, init_stmt, GSI_SAME_STMT); |
1108 | } | |
ebfd146a IR |
1109 | } |
1110 | ||
1111 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1112 | { | |
1113 | fprintf (vect_dump, "created new init_stmt: "); | |
1114 | print_gimple_stmt (vect_dump, init_stmt, 0, TDF_SLIM); | |
1115 | } | |
1116 | ||
1117 | vec_oprnd = gimple_assign_lhs (init_stmt); | |
1118 | return vec_oprnd; | |
1119 | } | |
1120 | ||
a70d6342 | 1121 | |
ebfd146a IR |
1122 | /* Function vect_get_vec_def_for_operand. |
1123 | ||
ff802fa1 | 1124 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
1125 | used in the vectorized stmt for STMT. |
1126 | ||
1127 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1128 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1129 | ||
1130 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1131 | needs to be introduced. */ | |
1132 | ||
1133 | tree | |
1134 | vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def) | |
1135 | { | |
1136 | tree vec_oprnd; | |
1137 | gimple vec_stmt; | |
1138 | gimple def_stmt; | |
1139 | stmt_vec_info def_stmt_info = NULL; | |
1140 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
9dc3f7de | 1141 | unsigned int nunits; |
ebfd146a IR |
1142 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
1143 | tree vec_inv; | |
1144 | tree vec_cst; | |
e7e9eb2f | 1145 | tree t = NULL_TREE; |
ebfd146a | 1146 | tree def; |
e7e9eb2f | 1147 | int i; |
ebfd146a IR |
1148 | enum vect_def_type dt; |
1149 | bool is_simple_use; | |
1150 | tree vector_type; | |
1151 | ||
1152 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1153 | { | |
1154 | fprintf (vect_dump, "vect_get_vec_def_for_operand: "); | |
1155 | print_generic_expr (vect_dump, op, TDF_SLIM); | |
1156 | } | |
1157 | ||
b8698a0f | 1158 | is_simple_use = vect_is_simple_use (op, loop_vinfo, NULL, &def_stmt, &def, |
a70d6342 | 1159 | &dt); |
ebfd146a IR |
1160 | gcc_assert (is_simple_use); |
1161 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1162 | { | |
1163 | if (def) | |
1164 | { | |
1165 | fprintf (vect_dump, "def = "); | |
1166 | print_generic_expr (vect_dump, def, TDF_SLIM); | |
1167 | } | |
1168 | if (def_stmt) | |
1169 | { | |
1170 | fprintf (vect_dump, " def_stmt = "); | |
1171 | print_gimple_stmt (vect_dump, def_stmt, 0, TDF_SLIM); | |
1172 | } | |
1173 | } | |
1174 | ||
1175 | switch (dt) | |
1176 | { | |
1177 | /* Case 1: operand is a constant. */ | |
1178 | case vect_constant_def: | |
1179 | { | |
7569a6cc RG |
1180 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
1181 | gcc_assert (vector_type); | |
9dc3f7de | 1182 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); |
7569a6cc | 1183 | |
b8698a0f | 1184 | if (scalar_def) |
ebfd146a IR |
1185 | *scalar_def = op; |
1186 | ||
1187 | /* Create 'vect_cst_ = {cst,cst,...,cst}' */ | |
1188 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1189 | fprintf (vect_dump, "Create vector_cst. nunits = %d", nunits); | |
1190 | ||
b9acc9f1 | 1191 | vec_cst = build_vector_from_val (vector_type, op); |
7569a6cc | 1192 | return vect_init_vector (stmt, vec_cst, vector_type, NULL); |
ebfd146a IR |
1193 | } |
1194 | ||
1195 | /* Case 2: operand is defined outside the loop - loop invariant. */ | |
8644a673 | 1196 | case vect_external_def: |
ebfd146a IR |
1197 | { |
1198 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); | |
1199 | gcc_assert (vector_type); | |
1200 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); | |
1201 | ||
b8698a0f | 1202 | if (scalar_def) |
ebfd146a IR |
1203 | *scalar_def = def; |
1204 | ||
1205 | /* Create 'vec_inv = {inv,inv,..,inv}' */ | |
1206 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1207 | fprintf (vect_dump, "Create vector_inv."); | |
1208 | ||
e7e9eb2f NF |
1209 | for (i = nunits - 1; i >= 0; --i) |
1210 | { | |
1211 | t = tree_cons (NULL_TREE, def, t); | |
1212 | } | |
1213 | ||
1214 | /* FIXME: use build_constructor directly. */ | |
1215 | vec_inv = build_constructor_from_list (vector_type, t); | |
ebfd146a IR |
1216 | return vect_init_vector (stmt, vec_inv, vector_type, NULL); |
1217 | } | |
1218 | ||
1219 | /* Case 3: operand is defined inside the loop. */ | |
8644a673 | 1220 | case vect_internal_def: |
ebfd146a | 1221 | { |
b8698a0f | 1222 | if (scalar_def) |
ebfd146a IR |
1223 | *scalar_def = NULL/* FIXME tuples: def_stmt*/; |
1224 | ||
1225 | /* Get the def from the vectorized stmt. */ | |
1226 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1227 | |
ebfd146a | 1228 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1229 | /* Get vectorized pattern statement. */ |
1230 | if (!vec_stmt | |
1231 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1232 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1233 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1234 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1235 | gcc_assert (vec_stmt); |
1236 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1237 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1238 | else if (is_gimple_call (vec_stmt)) | |
1239 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1240 | else | |
1241 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1242 | return vec_oprnd; | |
1243 | } | |
1244 | ||
1245 | /* Case 4: operand is defined by a loop header phi - reduction */ | |
1246 | case vect_reduction_def: | |
06066f92 | 1247 | case vect_double_reduction_def: |
7c5222ff | 1248 | case vect_nested_cycle: |
ebfd146a IR |
1249 | { |
1250 | struct loop *loop; | |
1251 | ||
1252 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
b8698a0f | 1253 | loop = (gimple_bb (def_stmt))->loop_father; |
ebfd146a IR |
1254 | |
1255 | /* Get the def before the loop */ | |
1256 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); | |
1257 | return get_initial_def_for_reduction (stmt, op, scalar_def); | |
1258 | } | |
1259 | ||
1260 | /* Case 5: operand is defined by loop-header phi - induction. */ | |
1261 | case vect_induction_def: | |
1262 | { | |
1263 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1264 | ||
1265 | /* Get the def from the vectorized stmt. */ | |
1266 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1267 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1268 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1269 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1270 | else | |
1271 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1272 | return vec_oprnd; |
1273 | } | |
1274 | ||
1275 | default: | |
1276 | gcc_unreachable (); | |
1277 | } | |
1278 | } | |
1279 | ||
1280 | ||
1281 | /* Function vect_get_vec_def_for_stmt_copy | |
1282 | ||
ff802fa1 | 1283 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1284 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1285 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1286 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1287 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1288 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1289 | DT is the type of the vector def VEC_OPRND. |
1290 | ||
1291 | Context: | |
1292 | In case the vectorization factor (VF) is bigger than the number | |
1293 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1294 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1295 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1296 | smallest data-type determines the VF, and as a result, when vectorizing |
1297 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1298 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1299 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1300 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1301 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1302 | ||
1303 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1304 | |
ebfd146a IR |
1305 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1306 | VS1.1: vx.1 = memref1 VS1.2 | |
1307 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1308 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1309 | |
1310 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1311 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1312 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1313 | VSnew.3: vz3 = vx.3 + ... | |
1314 | ||
1315 | The vectorization of S1 is explained in vectorizable_load. | |
1316 | The vectorization of S2: | |
b8698a0f L |
1317 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1318 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1319 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1320 | returns the vector-def 'vx.0'. |
1321 | ||
b8698a0f L |
1322 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1323 | function is called to get the relevant vector-def for each operand. It is | |
1324 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1325 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1326 | ||
b8698a0f L |
1327 | For example, to obtain the vector-def 'vx.1' in order to create the |
1328 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1329 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1330 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1331 | and return its def ('vx.1'). | |
1332 | Overall, to create the above sequence this function will be called 3 times: | |
1333 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1334 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1335 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1336 | ||
1337 | tree | |
1338 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1339 | { | |
1340 | gimple vec_stmt_for_operand; | |
1341 | stmt_vec_info def_stmt_info; | |
1342 | ||
1343 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1344 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1345 | return vec_oprnd; |
1346 | ||
1347 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1348 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1349 | gcc_assert (def_stmt_info); | |
1350 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1351 | gcc_assert (vec_stmt_for_operand); | |
1352 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1353 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) | |
1354 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1355 | else | |
1356 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1357 | return vec_oprnd; | |
1358 | } | |
1359 | ||
1360 | ||
1361 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1362 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1363 | |
1364 | static void | |
b8698a0f L |
1365 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
1366 | VEC(tree,heap) **vec_oprnds0, | |
ebfd146a IR |
1367 | VEC(tree,heap) **vec_oprnds1) |
1368 | { | |
1369 | tree vec_oprnd = VEC_pop (tree, *vec_oprnds0); | |
1370 | ||
1371 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
1372 | VEC_quick_push (tree, *vec_oprnds0, vec_oprnd); | |
1373 | ||
1374 | if (vec_oprnds1 && *vec_oprnds1) | |
1375 | { | |
1376 | vec_oprnd = VEC_pop (tree, *vec_oprnds1); | |
1377 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); | |
1378 | VEC_quick_push (tree, *vec_oprnds1, vec_oprnd); | |
1379 | } | |
1380 | } | |
1381 | ||
1382 | ||
ff802fa1 IR |
1383 | /* Get vectorized definitions for OP0 and OP1, or SLP_NODE if it is not |
1384 | NULL. */ | |
ebfd146a IR |
1385 | |
1386 | static void | |
1387 | vect_get_vec_defs (tree op0, tree op1, gimple stmt, | |
1388 | VEC(tree,heap) **vec_oprnds0, VEC(tree,heap) **vec_oprnds1, | |
1389 | slp_tree slp_node) | |
1390 | { | |
1391 | if (slp_node) | |
9dc3f7de | 1392 | vect_get_slp_defs (op0, op1, slp_node, vec_oprnds0, vec_oprnds1, -1); |
ebfd146a IR |
1393 | else |
1394 | { | |
1395 | tree vec_oprnd; | |
1396 | ||
b8698a0f L |
1397 | *vec_oprnds0 = VEC_alloc (tree, heap, 1); |
1398 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
ebfd146a IR |
1399 | VEC_quick_push (tree, *vec_oprnds0, vec_oprnd); |
1400 | ||
1401 | if (op1) | |
1402 | { | |
b8698a0f L |
1403 | *vec_oprnds1 = VEC_alloc (tree, heap, 1); |
1404 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL); | |
ebfd146a IR |
1405 | VEC_quick_push (tree, *vec_oprnds1, vec_oprnd); |
1406 | } | |
1407 | } | |
1408 | } | |
1409 | ||
1410 | ||
1411 | /* Function vect_finish_stmt_generation. | |
1412 | ||
1413 | Insert a new stmt. */ | |
1414 | ||
1415 | void | |
1416 | vect_finish_stmt_generation (gimple stmt, gimple vec_stmt, | |
1417 | gimple_stmt_iterator *gsi) | |
1418 | { | |
1419 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1420 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1421 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1422 | |
1423 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1424 | ||
1425 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); | |
1426 | ||
b8698a0f | 1427 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo, |
a70d6342 | 1428 | bb_vinfo)); |
ebfd146a IR |
1429 | |
1430 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1431 | { | |
1432 | fprintf (vect_dump, "add new stmt: "); | |
1433 | print_gimple_stmt (vect_dump, vec_stmt, 0, TDF_SLIM); | |
1434 | } | |
1435 | ||
ad885386 | 1436 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
ebfd146a IR |
1437 | } |
1438 | ||
1439 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1440 | a function declaration if the target has a vectorized version | |
1441 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1442 | ||
1443 | tree | |
1444 | vectorizable_function (gimple call, tree vectype_out, tree vectype_in) | |
1445 | { | |
1446 | tree fndecl = gimple_call_fndecl (call); | |
ebfd146a IR |
1447 | |
1448 | /* We only handle functions that do not read or clobber memory -- i.e. | |
1449 | const or novops ones. */ | |
1450 | if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS))) | |
1451 | return NULL_TREE; | |
1452 | ||
1453 | if (!fndecl | |
1454 | || TREE_CODE (fndecl) != FUNCTION_DECL | |
1455 | || !DECL_BUILT_IN (fndecl)) | |
1456 | return NULL_TREE; | |
1457 | ||
62f7fd21 | 1458 | return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out, |
ebfd146a IR |
1459 | vectype_in); |
1460 | } | |
1461 | ||
1462 | /* Function vectorizable_call. | |
1463 | ||
b8698a0f L |
1464 | Check if STMT performs a function call that can be vectorized. |
1465 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1466 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1467 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1468 | ||
1469 | static bool | |
1470 | vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt) | |
1471 | { | |
1472 | tree vec_dest; | |
1473 | tree scalar_dest; | |
1474 | tree op, type; | |
1475 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1476 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
1477 | tree vectype_out, vectype_in; | |
1478 | int nunits_in; | |
1479 | int nunits_out; | |
1480 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
b690cc0f | 1481 | tree fndecl, new_temp, def, rhs_type; |
ebfd146a | 1482 | gimple def_stmt; |
0502fb85 UB |
1483 | enum vect_def_type dt[3] |
1484 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
63827fb8 | 1485 | gimple new_stmt = NULL; |
ebfd146a IR |
1486 | int ncopies, j; |
1487 | VEC(tree, heap) *vargs = NULL; | |
1488 | enum { NARROW, NONE, WIDEN } modifier; | |
1489 | size_t i, nargs; | |
9d5e7640 | 1490 | tree lhs; |
ebfd146a | 1491 | |
a70d6342 IR |
1492 | /* FORNOW: unsupported in basic block SLP. */ |
1493 | gcc_assert (loop_vinfo); | |
b8698a0f | 1494 | |
ebfd146a IR |
1495 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
1496 | return false; | |
1497 | ||
8644a673 | 1498 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1499 | return false; |
1500 | ||
1501 | /* FORNOW: SLP not supported. */ | |
1502 | if (STMT_SLP_TYPE (stmt_info)) | |
1503 | return false; | |
1504 | ||
1505 | /* Is STMT a vectorizable call? */ | |
1506 | if (!is_gimple_call (stmt)) | |
1507 | return false; | |
1508 | ||
1509 | if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
1510 | return false; | |
1511 | ||
822ba6d7 | 1512 | if (stmt_can_throw_internal (stmt)) |
5a2c1986 IR |
1513 | return false; |
1514 | ||
b690cc0f RG |
1515 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
1516 | ||
ebfd146a IR |
1517 | /* Process function arguments. */ |
1518 | rhs_type = NULL_TREE; | |
b690cc0f | 1519 | vectype_in = NULL_TREE; |
ebfd146a IR |
1520 | nargs = gimple_call_num_args (stmt); |
1521 | ||
1b1562a5 MM |
1522 | /* Bail out if the function has more than three arguments, we do not have |
1523 | interesting builtin functions to vectorize with more than two arguments | |
1524 | except for fma. No arguments is also not good. */ | |
1525 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
1526 | return false; |
1527 | ||
1528 | for (i = 0; i < nargs; i++) | |
1529 | { | |
b690cc0f RG |
1530 | tree opvectype; |
1531 | ||
ebfd146a IR |
1532 | op = gimple_call_arg (stmt, i); |
1533 | ||
1534 | /* We can only handle calls with arguments of the same type. */ | |
1535 | if (rhs_type | |
8533c9d8 | 1536 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a IR |
1537 | { |
1538 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1539 | fprintf (vect_dump, "argument types differ."); | |
1540 | return false; | |
1541 | } | |
b690cc0f RG |
1542 | if (!rhs_type) |
1543 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 1544 | |
b690cc0f RG |
1545 | if (!vect_is_simple_use_1 (op, loop_vinfo, NULL, |
1546 | &def_stmt, &def, &dt[i], &opvectype)) | |
ebfd146a IR |
1547 | { |
1548 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1549 | fprintf (vect_dump, "use not simple."); | |
1550 | return false; | |
1551 | } | |
ebfd146a | 1552 | |
b690cc0f RG |
1553 | if (!vectype_in) |
1554 | vectype_in = opvectype; | |
1555 | else if (opvectype | |
1556 | && opvectype != vectype_in) | |
1557 | { | |
1558 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1559 | fprintf (vect_dump, "argument vector types differ."); | |
1560 | return false; | |
1561 | } | |
1562 | } | |
1563 | /* If all arguments are external or constant defs use a vector type with | |
1564 | the same size as the output vector type. */ | |
ebfd146a | 1565 | if (!vectype_in) |
b690cc0f | 1566 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
1567 | if (vec_stmt) |
1568 | gcc_assert (vectype_in); | |
1569 | if (!vectype_in) | |
1570 | { | |
1571 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1572 | { | |
1573 | fprintf (vect_dump, "no vectype for scalar type "); | |
1574 | print_generic_expr (vect_dump, rhs_type, TDF_SLIM); | |
1575 | } | |
1576 | ||
1577 | return false; | |
1578 | } | |
ebfd146a IR |
1579 | |
1580 | /* FORNOW */ | |
b690cc0f RG |
1581 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
1582 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
1583 | if (nunits_in == nunits_out / 2) |
1584 | modifier = NARROW; | |
1585 | else if (nunits_out == nunits_in) | |
1586 | modifier = NONE; | |
1587 | else if (nunits_out == nunits_in / 2) | |
1588 | modifier = WIDEN; | |
1589 | else | |
1590 | return false; | |
1591 | ||
1592 | /* For now, we only vectorize functions if a target specific builtin | |
1593 | is available. TODO -- in some cases, it might be profitable to | |
1594 | insert the calls for pieces of the vector, in order to be able | |
1595 | to vectorize other operations in the loop. */ | |
1596 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
1597 | if (fndecl == NULL_TREE) | |
1598 | { | |
1599 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1600 | fprintf (vect_dump, "function is not vectorizable."); | |
1601 | ||
1602 | return false; | |
1603 | } | |
1604 | ||
5006671f | 1605 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a IR |
1606 | |
1607 | if (modifier == NARROW) | |
1608 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
1609 | else | |
1610 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1611 | ||
1612 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
1613 | needs to be generated. */ | |
1614 | gcc_assert (ncopies >= 1); | |
1615 | ||
1616 | if (!vec_stmt) /* transformation not required. */ | |
1617 | { | |
1618 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1619 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1620 | fprintf (vect_dump, "=== vectorizable_call ==="); | |
1621 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
1622 | return true; | |
1623 | } | |
1624 | ||
1625 | /** Transform. **/ | |
1626 | ||
1627 | if (vect_print_dump_info (REPORT_DETAILS)) | |
9d5e7640 | 1628 | fprintf (vect_dump, "transform call."); |
ebfd146a IR |
1629 | |
1630 | /* Handle def. */ | |
1631 | scalar_dest = gimple_call_lhs (stmt); | |
1632 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1633 | ||
1634 | prev_stmt_info = NULL; | |
1635 | switch (modifier) | |
1636 | { | |
1637 | case NONE: | |
1638 | for (j = 0; j < ncopies; ++j) | |
1639 | { | |
1640 | /* Build argument list for the vectorized call. */ | |
1641 | if (j == 0) | |
1642 | vargs = VEC_alloc (tree, heap, nargs); | |
1643 | else | |
1644 | VEC_truncate (tree, vargs, 0); | |
1645 | ||
1646 | for (i = 0; i < nargs; i++) | |
1647 | { | |
1648 | op = gimple_call_arg (stmt, i); | |
1649 | if (j == 0) | |
1650 | vec_oprnd0 | |
1651 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1652 | else | |
63827fb8 IR |
1653 | { |
1654 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
1655 | vec_oprnd0 | |
1656 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
1657 | } | |
ebfd146a IR |
1658 | |
1659 | VEC_quick_push (tree, vargs, vec_oprnd0); | |
1660 | } | |
1661 | ||
1662 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1663 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1664 | gimple_call_set_lhs (new_stmt, new_temp); | |
1665 | ||
1666 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7411b8f0 | 1667 | mark_symbols_for_renaming (new_stmt); |
ebfd146a IR |
1668 | |
1669 | if (j == 0) | |
1670 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1671 | else | |
1672 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1673 | ||
1674 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1675 | } | |
1676 | ||
1677 | break; | |
1678 | ||
1679 | case NARROW: | |
1680 | for (j = 0; j < ncopies; ++j) | |
1681 | { | |
1682 | /* Build argument list for the vectorized call. */ | |
1683 | if (j == 0) | |
1684 | vargs = VEC_alloc (tree, heap, nargs * 2); | |
1685 | else | |
1686 | VEC_truncate (tree, vargs, 0); | |
1687 | ||
1688 | for (i = 0; i < nargs; i++) | |
1689 | { | |
1690 | op = gimple_call_arg (stmt, i); | |
1691 | if (j == 0) | |
1692 | { | |
1693 | vec_oprnd0 | |
1694 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1695 | vec_oprnd1 | |
63827fb8 | 1696 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
1697 | } |
1698 | else | |
1699 | { | |
336ecb65 | 1700 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 1701 | vec_oprnd0 |
63827fb8 | 1702 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 1703 | vec_oprnd1 |
63827fb8 | 1704 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
1705 | } |
1706 | ||
1707 | VEC_quick_push (tree, vargs, vec_oprnd0); | |
1708 | VEC_quick_push (tree, vargs, vec_oprnd1); | |
1709 | } | |
1710 | ||
1711 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1712 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1713 | gimple_call_set_lhs (new_stmt, new_temp); | |
1714 | ||
1715 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7411b8f0 | 1716 | mark_symbols_for_renaming (new_stmt); |
ebfd146a IR |
1717 | |
1718 | if (j == 0) | |
1719 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1720 | else | |
1721 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1722 | ||
1723 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1724 | } | |
1725 | ||
1726 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
1727 | ||
1728 | break; | |
1729 | ||
1730 | case WIDEN: | |
1731 | /* No current target implements this case. */ | |
1732 | return false; | |
1733 | } | |
1734 | ||
1735 | VEC_free (tree, heap, vargs); | |
1736 | ||
1737 | /* Update the exception handling table with the vector stmt if necessary. */ | |
1738 | if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt)) | |
1739 | gimple_purge_dead_eh_edges (gimple_bb (stmt)); | |
1740 | ||
1741 | /* The call in STMT might prevent it from being removed in dce. | |
1742 | We however cannot remove it here, due to the way the ssa name | |
1743 | it defines is mapped to the new definition. So just replace | |
1744 | rhs of the statement with something harmless. */ | |
1745 | ||
1746 | type = TREE_TYPE (scalar_dest); | |
9d5e7640 IR |
1747 | if (is_pattern_stmt_p (stmt_info)) |
1748 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
1749 | else | |
1750 | lhs = gimple_call_lhs (stmt); | |
1751 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
ebfd146a IR |
1752 | set_vinfo_for_stmt (new_stmt, stmt_info); |
1753 | set_vinfo_for_stmt (stmt, NULL); | |
1754 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
1755 | gsi_replace (gsi, new_stmt, false); | |
1756 | SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt; | |
1757 | ||
1758 | return true; | |
1759 | } | |
1760 | ||
1761 | ||
1762 | /* Function vect_gen_widened_results_half | |
1763 | ||
1764 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 1765 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 1766 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
1767 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
1768 | needs to be created (DECL is a function-decl of a target-builtin). | |
1769 | STMT is the original scalar stmt that we are vectorizing. */ | |
1770 | ||
1771 | static gimple | |
1772 | vect_gen_widened_results_half (enum tree_code code, | |
1773 | tree decl, | |
1774 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
1775 | tree vec_dest, gimple_stmt_iterator *gsi, | |
1776 | gimple stmt) | |
b8698a0f | 1777 | { |
ebfd146a | 1778 | gimple new_stmt; |
b8698a0f L |
1779 | tree new_temp; |
1780 | ||
1781 | /* Generate half of the widened result: */ | |
1782 | if (code == CALL_EXPR) | |
1783 | { | |
1784 | /* Target specific support */ | |
ebfd146a IR |
1785 | if (op_type == binary_op) |
1786 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
1787 | else | |
1788 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
1789 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1790 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
1791 | } |
1792 | else | |
ebfd146a | 1793 | { |
b8698a0f L |
1794 | /* Generic support */ |
1795 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
1796 | if (op_type != binary_op) |
1797 | vec_oprnd1 = NULL; | |
1798 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0, | |
1799 | vec_oprnd1); | |
1800 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1801 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 1802 | } |
ebfd146a IR |
1803 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1804 | ||
ebfd146a IR |
1805 | return new_stmt; |
1806 | } | |
1807 | ||
1808 | ||
b8698a0f L |
1809 | /* Check if STMT performs a conversion operation, that can be vectorized. |
1810 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1811 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1812 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1813 | ||
1814 | static bool | |
1815 | vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi, | |
1816 | gimple *vec_stmt, slp_tree slp_node) | |
1817 | { | |
1818 | tree vec_dest; | |
1819 | tree scalar_dest; | |
1820 | tree op0; | |
1821 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1822 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1823 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1824 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
1825 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; | |
1826 | tree new_temp; | |
1827 | tree def; | |
1828 | gimple def_stmt; | |
1829 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
1830 | gimple new_stmt = NULL; | |
1831 | stmt_vec_info prev_stmt_info; | |
1832 | int nunits_in; | |
1833 | int nunits_out; | |
1834 | tree vectype_out, vectype_in; | |
1835 | int ncopies, j; | |
b690cc0f | 1836 | tree rhs_type; |
ebfd146a IR |
1837 | tree builtin_decl; |
1838 | enum { NARROW, NONE, WIDEN } modifier; | |
1839 | int i; | |
1840 | VEC(tree,heap) *vec_oprnds0 = NULL; | |
1841 | tree vop0; | |
ebfd146a IR |
1842 | VEC(tree,heap) *dummy = NULL; |
1843 | int dummy_int; | |
1844 | ||
1845 | /* Is STMT a vectorizable conversion? */ | |
1846 | ||
a70d6342 IR |
1847 | /* FORNOW: unsupported in basic block SLP. */ |
1848 | gcc_assert (loop_vinfo); | |
b8698a0f | 1849 | |
ebfd146a IR |
1850 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
1851 | return false; | |
1852 | ||
8644a673 | 1853 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1854 | return false; |
1855 | ||
1856 | if (!is_gimple_assign (stmt)) | |
1857 | return false; | |
1858 | ||
1859 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
1860 | return false; | |
1861 | ||
1862 | code = gimple_assign_rhs_code (stmt); | |
1863 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
1864 | return false; | |
1865 | ||
1866 | /* Check types of lhs and rhs. */ | |
b690cc0f RG |
1867 | scalar_dest = gimple_assign_lhs (stmt); |
1868 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
1869 | ||
ebfd146a IR |
1870 | op0 = gimple_assign_rhs1 (stmt); |
1871 | rhs_type = TREE_TYPE (op0); | |
b690cc0f RG |
1872 | /* Check the operands of the operation. */ |
1873 | if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL, | |
1874 | &def_stmt, &def, &dt[0], &vectype_in)) | |
1875 | { | |
1876 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1877 | fprintf (vect_dump, "use not simple."); | |
1878 | return false; | |
1879 | } | |
1880 | /* If op0 is an external or constant defs use a vector type of | |
1881 | the same size as the output vector type. */ | |
ebfd146a | 1882 | if (!vectype_in) |
b690cc0f | 1883 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
1884 | if (vec_stmt) |
1885 | gcc_assert (vectype_in); | |
1886 | if (!vectype_in) | |
1887 | { | |
1888 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1889 | { | |
1890 | fprintf (vect_dump, "no vectype for scalar type "); | |
1891 | print_generic_expr (vect_dump, rhs_type, TDF_SLIM); | |
1892 | } | |
1893 | ||
1894 | return false; | |
1895 | } | |
ebfd146a IR |
1896 | |
1897 | /* FORNOW */ | |
b690cc0f RG |
1898 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
1899 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
1900 | if (nunits_in == nunits_out / 2) |
1901 | modifier = NARROW; | |
1902 | else if (nunits_out == nunits_in) | |
1903 | modifier = NONE; | |
1904 | else if (nunits_out == nunits_in / 2) | |
1905 | modifier = WIDEN; | |
1906 | else | |
1907 | return false; | |
1908 | ||
ebfd146a IR |
1909 | if (modifier == NARROW) |
1910 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
1911 | else | |
1912 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1913 | ||
ff802fa1 IR |
1914 | /* Multiple types in SLP are handled by creating the appropriate number of |
1915 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
1916 | case of SLP. */ | |
437f4a00 | 1917 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a | 1918 | ncopies = 1; |
b8698a0f | 1919 | |
ebfd146a IR |
1920 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
1921 | needs to be generated. */ | |
1922 | gcc_assert (ncopies >= 1); | |
1923 | ||
ebfd146a IR |
1924 | /* Supportable by target? */ |
1925 | if ((modifier == NONE | |
88dd7150 | 1926 | && !targetm.vectorize.builtin_conversion (code, vectype_out, vectype_in)) |
ebfd146a | 1927 | || (modifier == WIDEN |
b690cc0f RG |
1928 | && !supportable_widening_operation (code, stmt, |
1929 | vectype_out, vectype_in, | |
ebfd146a IR |
1930 | &decl1, &decl2, |
1931 | &code1, &code2, | |
1932 | &dummy_int, &dummy)) | |
1933 | || (modifier == NARROW | |
b690cc0f | 1934 | && !supportable_narrowing_operation (code, vectype_out, vectype_in, |
ebfd146a IR |
1935 | &code1, &dummy_int, &dummy))) |
1936 | { | |
1937 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1938 | fprintf (vect_dump, "conversion not supported by target."); | |
1939 | return false; | |
1940 | } | |
1941 | ||
1942 | if (modifier != NONE) | |
1943 | { | |
ebfd146a IR |
1944 | /* FORNOW: SLP not supported. */ |
1945 | if (STMT_SLP_TYPE (stmt_info)) | |
b8698a0f | 1946 | return false; |
ebfd146a IR |
1947 | } |
1948 | ||
1949 | if (!vec_stmt) /* transformation not required. */ | |
1950 | { | |
1951 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
1952 | return true; | |
1953 | } | |
1954 | ||
1955 | /** Transform. **/ | |
1956 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1957 | fprintf (vect_dump, "transform conversion."); | |
1958 | ||
1959 | /* Handle def. */ | |
1960 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1961 | ||
1962 | if (modifier == NONE && !slp_node) | |
1963 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
1964 | ||
1965 | prev_stmt_info = NULL; | |
1966 | switch (modifier) | |
1967 | { | |
1968 | case NONE: | |
1969 | for (j = 0; j < ncopies; j++) | |
1970 | { | |
ebfd146a | 1971 | if (j == 0) |
b8698a0f | 1972 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node); |
ebfd146a IR |
1973 | else |
1974 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
1975 | ||
1976 | builtin_decl = | |
88dd7150 RG |
1977 | targetm.vectorize.builtin_conversion (code, |
1978 | vectype_out, vectype_in); | |
ac47786e | 1979 | FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0) |
b8698a0f | 1980 | { |
ebfd146a IR |
1981 | /* Arguments are ready. create the new vector stmt. */ |
1982 | new_stmt = gimple_build_call (builtin_decl, 1, vop0); | |
1983 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1984 | gimple_call_set_lhs (new_stmt, new_temp); | |
1985 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a IR |
1986 | if (slp_node) |
1987 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
1988 | } | |
1989 | ||
1990 | if (j == 0) | |
1991 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1992 | else | |
1993 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1994 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1995 | } | |
1996 | break; | |
1997 | ||
1998 | case WIDEN: | |
1999 | /* In case the vectorization factor (VF) is bigger than the number | |
2000 | of elements that we can fit in a vectype (nunits), we have to | |
2001 | generate more than one vector stmt - i.e - we need to "unroll" | |
2002 | the vector stmt by a factor VF/nunits. */ | |
2003 | for (j = 0; j < ncopies; j++) | |
2004 | { | |
2005 | if (j == 0) | |
2006 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
2007 | else | |
2008 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
2009 | ||
ebfd146a IR |
2010 | /* Generate first half of the widened result: */ |
2011 | new_stmt | |
b8698a0f | 2012 | = vect_gen_widened_results_half (code1, decl1, |
ebfd146a IR |
2013 | vec_oprnd0, vec_oprnd1, |
2014 | unary_op, vec_dest, gsi, stmt); | |
2015 | if (j == 0) | |
2016 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2017 | else | |
2018 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2019 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2020 | ||
2021 | /* Generate second half of the widened result: */ | |
2022 | new_stmt | |
2023 | = vect_gen_widened_results_half (code2, decl2, | |
2024 | vec_oprnd0, vec_oprnd1, | |
2025 | unary_op, vec_dest, gsi, stmt); | |
2026 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2027 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2028 | } | |
2029 | break; | |
2030 | ||
2031 | case NARROW: | |
2032 | /* In case the vectorization factor (VF) is bigger than the number | |
2033 | of elements that we can fit in a vectype (nunits), we have to | |
2034 | generate more than one vector stmt - i.e - we need to "unroll" | |
2035 | the vector stmt by a factor VF/nunits. */ | |
2036 | for (j = 0; j < ncopies; j++) | |
2037 | { | |
2038 | /* Handle uses. */ | |
2039 | if (j == 0) | |
2040 | { | |
2041 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
2042 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
2043 | } | |
2044 | else | |
2045 | { | |
2046 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd1); | |
2047 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
2048 | } | |
2049 | ||
9dc3f7de | 2050 | /* Arguments are ready. Create the new vector stmt. */ |
ebfd146a IR |
2051 | new_stmt = gimple_build_assign_with_ops (code1, vec_dest, vec_oprnd0, |
2052 | vec_oprnd1); | |
2053 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2054 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2055 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2056 | ||
2057 | if (j == 0) | |
2058 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2059 | else | |
2060 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2061 | ||
2062 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2063 | } | |
2064 | ||
2065 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2066 | } | |
2067 | ||
2068 | if (vec_oprnds0) | |
b8698a0f | 2069 | VEC_free (tree, heap, vec_oprnds0); |
ebfd146a IR |
2070 | |
2071 | return true; | |
2072 | } | |
ff802fa1 IR |
2073 | |
2074 | ||
ebfd146a IR |
2075 | /* Function vectorizable_assignment. |
2076 | ||
b8698a0f L |
2077 | Check if STMT performs an assignment (copy) that can be vectorized. |
2078 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
2079 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2080 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2081 | ||
2082 | static bool | |
2083 | vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi, | |
2084 | gimple *vec_stmt, slp_tree slp_node) | |
2085 | { | |
2086 | tree vec_dest; | |
2087 | tree scalar_dest; | |
2088 | tree op; | |
2089 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2090 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2091 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2092 | tree new_temp; | |
2093 | tree def; | |
2094 | gimple def_stmt; | |
2095 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
fde9c428 | 2096 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
ebfd146a | 2097 | int ncopies; |
f18b55bd | 2098 | int i, j; |
ebfd146a IR |
2099 | VEC(tree,heap) *vec_oprnds = NULL; |
2100 | tree vop; | |
a70d6342 | 2101 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
f18b55bd IR |
2102 | gimple new_stmt = NULL; |
2103 | stmt_vec_info prev_stmt_info = NULL; | |
fde9c428 RG |
2104 | enum tree_code code; |
2105 | tree vectype_in; | |
ebfd146a IR |
2106 | |
2107 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2108 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2109 | case of SLP. */ | |
437f4a00 | 2110 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
2111 | ncopies = 1; |
2112 | else | |
2113 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2114 | ||
2115 | gcc_assert (ncopies >= 1); | |
ebfd146a | 2116 | |
a70d6342 | 2117 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2118 | return false; |
2119 | ||
8644a673 | 2120 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2121 | return false; |
2122 | ||
2123 | /* Is vectorizable assignment? */ | |
2124 | if (!is_gimple_assign (stmt)) | |
2125 | return false; | |
2126 | ||
2127 | scalar_dest = gimple_assign_lhs (stmt); | |
2128 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
2129 | return false; | |
2130 | ||
fde9c428 | 2131 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 2132 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
2133 | || code == PAREN_EXPR |
2134 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
2135 | op = gimple_assign_rhs1 (stmt); |
2136 | else | |
2137 | return false; | |
2138 | ||
7b7ec6c5 RG |
2139 | if (code == VIEW_CONVERT_EXPR) |
2140 | op = TREE_OPERAND (op, 0); | |
2141 | ||
fde9c428 RG |
2142 | if (!vect_is_simple_use_1 (op, loop_vinfo, bb_vinfo, |
2143 | &def_stmt, &def, &dt[0], &vectype_in)) | |
ebfd146a IR |
2144 | { |
2145 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2146 | fprintf (vect_dump, "use not simple."); | |
2147 | return false; | |
2148 | } | |
2149 | ||
fde9c428 RG |
2150 | /* We can handle NOP_EXPR conversions that do not change the number |
2151 | of elements or the vector size. */ | |
7b7ec6c5 RG |
2152 | if ((CONVERT_EXPR_CODE_P (code) |
2153 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
2154 | && (!vectype_in |
2155 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
2156 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
2157 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
2158 | return false; | |
2159 | ||
ebfd146a IR |
2160 | if (!vec_stmt) /* transformation not required. */ |
2161 | { | |
2162 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
2163 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2164 | fprintf (vect_dump, "=== vectorizable_assignment ==="); | |
2165 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2166 | return true; | |
2167 | } | |
2168 | ||
2169 | /** Transform. **/ | |
2170 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2171 | fprintf (vect_dump, "transform assignment."); | |
2172 | ||
2173 | /* Handle def. */ | |
2174 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2175 | ||
2176 | /* Handle use. */ | |
f18b55bd | 2177 | for (j = 0; j < ncopies; j++) |
ebfd146a | 2178 | { |
f18b55bd IR |
2179 | /* Handle uses. */ |
2180 | if (j == 0) | |
2181 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); | |
2182 | else | |
2183 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
2184 | ||
2185 | /* Arguments are ready. create the new vector stmt. */ | |
ac47786e | 2186 | FOR_EACH_VEC_ELT (tree, vec_oprnds, i, vop) |
f18b55bd | 2187 | { |
7b7ec6c5 RG |
2188 | if (CONVERT_EXPR_CODE_P (code) |
2189 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 2190 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
2191 | new_stmt = gimple_build_assign (vec_dest, vop); |
2192 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2193 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2194 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2195 | if (slp_node) | |
2196 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2197 | } | |
ebfd146a IR |
2198 | |
2199 | if (slp_node) | |
f18b55bd IR |
2200 | continue; |
2201 | ||
2202 | if (j == 0) | |
2203 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2204 | else | |
2205 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2206 | ||
2207 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2208 | } | |
b8698a0f L |
2209 | |
2210 | VEC_free (tree, heap, vec_oprnds); | |
ebfd146a IR |
2211 | return true; |
2212 | } | |
2213 | ||
9dc3f7de | 2214 | |
1107f3ae IR |
2215 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
2216 | either as shift by a scalar or by a vector. */ | |
2217 | ||
2218 | bool | |
2219 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
2220 | { | |
2221 | ||
2222 | enum machine_mode vec_mode; | |
2223 | optab optab; | |
2224 | int icode; | |
2225 | tree vectype; | |
2226 | ||
2227 | vectype = get_vectype_for_scalar_type (scalar_type); | |
2228 | if (!vectype) | |
2229 | return false; | |
2230 | ||
2231 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
2232 | if (!optab | |
2233 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
2234 | { | |
2235 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
2236 | if (!optab | |
2237 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
2238 | == CODE_FOR_nothing)) | |
2239 | return false; | |
2240 | } | |
2241 | ||
2242 | vec_mode = TYPE_MODE (vectype); | |
2243 | icode = (int) optab_handler (optab, vec_mode); | |
2244 | if (icode == CODE_FOR_nothing) | |
2245 | return false; | |
2246 | ||
2247 | return true; | |
2248 | } | |
2249 | ||
2250 | ||
9dc3f7de IR |
2251 | /* Function vectorizable_shift. |
2252 | ||
2253 | Check if STMT performs a shift operation that can be vectorized. | |
2254 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2255 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2256 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2257 | ||
2258 | static bool | |
2259 | vectorizable_shift (gimple stmt, gimple_stmt_iterator *gsi, | |
2260 | gimple *vec_stmt, slp_tree slp_node) | |
2261 | { | |
2262 | tree vec_dest; | |
2263 | tree scalar_dest; | |
2264 | tree op0, op1 = NULL; | |
2265 | tree vec_oprnd1 = NULL_TREE; | |
2266 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2267 | tree vectype; | |
2268 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2269 | enum tree_code code; | |
2270 | enum machine_mode vec_mode; | |
2271 | tree new_temp; | |
2272 | optab optab; | |
2273 | int icode; | |
2274 | enum machine_mode optab_op2_mode; | |
2275 | tree def; | |
2276 | gimple def_stmt; | |
2277 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2278 | gimple new_stmt = NULL; | |
2279 | stmt_vec_info prev_stmt_info; | |
2280 | int nunits_in; | |
2281 | int nunits_out; | |
2282 | tree vectype_out; | |
2283 | int ncopies; | |
2284 | int j, i; | |
2285 | VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; | |
2286 | tree vop0, vop1; | |
2287 | unsigned int k; | |
49eab32e | 2288 | bool scalar_shift_arg = true; |
9dc3f7de IR |
2289 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2290 | int vf; | |
2291 | ||
2292 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
2293 | return false; | |
2294 | ||
2295 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
2296 | return false; | |
2297 | ||
2298 | /* Is STMT a vectorizable binary/unary operation? */ | |
2299 | if (!is_gimple_assign (stmt)) | |
2300 | return false; | |
2301 | ||
2302 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2303 | return false; | |
2304 | ||
2305 | code = gimple_assign_rhs_code (stmt); | |
2306 | ||
2307 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
2308 | || code == RROTATE_EXPR)) | |
2309 | return false; | |
2310 | ||
2311 | scalar_dest = gimple_assign_lhs (stmt); | |
2312 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2313 | ||
2314 | op0 = gimple_assign_rhs1 (stmt); | |
2315 | if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo, | |
2316 | &def_stmt, &def, &dt[0], &vectype)) | |
2317 | { | |
2318 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2319 | fprintf (vect_dump, "use not simple."); | |
2320 | return false; | |
2321 | } | |
2322 | /* If op0 is an external or constant def use a vector type with | |
2323 | the same size as the output vector type. */ | |
2324 | if (!vectype) | |
2325 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
2326 | if (vec_stmt) | |
2327 | gcc_assert (vectype); | |
2328 | if (!vectype) | |
2329 | { | |
2330 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2331 | { | |
2332 | fprintf (vect_dump, "no vectype for scalar type "); | |
2333 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
2334 | } | |
2335 | ||
2336 | return false; | |
2337 | } | |
2338 | ||
2339 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
2340 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
2341 | if (nunits_out != nunits_in) | |
2342 | return false; | |
2343 | ||
2344 | op1 = gimple_assign_rhs2 (stmt); | |
2345 | if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt[1])) | |
2346 | { | |
2347 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2348 | fprintf (vect_dump, "use not simple."); | |
2349 | return false; | |
2350 | } | |
2351 | ||
2352 | if (loop_vinfo) | |
2353 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
2354 | else | |
2355 | vf = 1; | |
2356 | ||
2357 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2358 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2359 | case of SLP. */ | |
437f4a00 | 2360 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
9dc3f7de IR |
2361 | ncopies = 1; |
2362 | else | |
2363 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2364 | ||
2365 | gcc_assert (ncopies >= 1); | |
2366 | ||
2367 | /* Determine whether the shift amount is a vector, or scalar. If the | |
2368 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
2369 | ||
49eab32e JJ |
2370 | if (dt[1] == vect_internal_def && !slp_node) |
2371 | scalar_shift_arg = false; | |
2372 | else if (dt[1] == vect_constant_def | |
2373 | || dt[1] == vect_external_def | |
2374 | || dt[1] == vect_internal_def) | |
2375 | { | |
2376 | /* In SLP, need to check whether the shift count is the same, | |
2377 | in loops if it is a constant or invariant, it is always | |
2378 | a scalar shift. */ | |
2379 | if (slp_node) | |
2380 | { | |
2381 | VEC (gimple, heap) *stmts = SLP_TREE_SCALAR_STMTS (slp_node); | |
2382 | gimple slpstmt; | |
2383 | ||
2384 | FOR_EACH_VEC_ELT (gimple, stmts, k, slpstmt) | |
2385 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) | |
2386 | scalar_shift_arg = false; | |
2387 | } | |
2388 | } | |
2389 | else | |
2390 | { | |
2391 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2392 | fprintf (vect_dump, "operand mode requires invariant argument."); | |
2393 | return false; | |
2394 | } | |
2395 | ||
9dc3f7de | 2396 | /* Vector shifted by vector. */ |
49eab32e | 2397 | if (!scalar_shift_arg) |
9dc3f7de IR |
2398 | { |
2399 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
2400 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2401 | fprintf (vect_dump, "vector/vector shift/rotate found."); | |
2402 | } | |
2403 | /* See if the machine has a vector shifted by scalar insn and if not | |
2404 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 2405 | else |
9dc3f7de IR |
2406 | { |
2407 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
2408 | if (optab | |
2409 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
2410 | { | |
9dc3f7de IR |
2411 | if (vect_print_dump_info (REPORT_DETAILS)) |
2412 | fprintf (vect_dump, "vector/scalar shift/rotate found."); | |
2413 | } | |
2414 | else | |
2415 | { | |
2416 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
2417 | if (optab | |
2418 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
2419 | != CODE_FOR_nothing)) | |
2420 | { | |
49eab32e JJ |
2421 | scalar_shift_arg = false; |
2422 | ||
9dc3f7de IR |
2423 | if (vect_print_dump_info (REPORT_DETAILS)) |
2424 | fprintf (vect_dump, "vector/vector shift/rotate found."); | |
2425 | ||
2426 | /* Unlike the other binary operators, shifts/rotates have | |
2427 | the rhs being int, instead of the same type as the lhs, | |
2428 | so make sure the scalar is the right type if we are | |
2429 | dealing with vectors of short/char. */ | |
2430 | if (dt[1] == vect_constant_def) | |
2431 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
2432 | } | |
2433 | } | |
2434 | } | |
9dc3f7de IR |
2435 | |
2436 | /* Supportable by target? */ | |
2437 | if (!optab) | |
2438 | { | |
2439 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2440 | fprintf (vect_dump, "no optab."); | |
2441 | return false; | |
2442 | } | |
2443 | vec_mode = TYPE_MODE (vectype); | |
2444 | icode = (int) optab_handler (optab, vec_mode); | |
2445 | if (icode == CODE_FOR_nothing) | |
2446 | { | |
2447 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2448 | fprintf (vect_dump, "op not supported by target."); | |
2449 | /* Check only during analysis. */ | |
2450 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
2451 | || (vf < vect_min_worthwhile_factor (code) | |
2452 | && !vec_stmt)) | |
2453 | return false; | |
2454 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2455 | fprintf (vect_dump, "proceeding using word mode."); | |
2456 | } | |
2457 | ||
2458 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
2459 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
2460 | && vf < vect_min_worthwhile_factor (code) | |
2461 | && !vec_stmt) | |
2462 | { | |
2463 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2464 | fprintf (vect_dump, "not worthwhile without SIMD support."); | |
2465 | return false; | |
2466 | } | |
2467 | ||
2468 | if (!vec_stmt) /* transformation not required. */ | |
2469 | { | |
2470 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
2471 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2472 | fprintf (vect_dump, "=== vectorizable_shift ==="); | |
2473 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2474 | return true; | |
2475 | } | |
2476 | ||
2477 | /** Transform. **/ | |
2478 | ||
2479 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2480 | fprintf (vect_dump, "transform binary/unary operation."); | |
2481 | ||
2482 | /* Handle def. */ | |
2483 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2484 | ||
2485 | /* Allocate VECs for vector operands. In case of SLP, vector operands are | |
2486 | created in the previous stages of the recursion, so no allocation is | |
2487 | needed, except for the case of shift with scalar shift argument. In that | |
2488 | case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to | |
2489 | be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE. | |
2490 | In case of loop-based vectorization we allocate VECs of size 1. We | |
2491 | allocate VEC_OPRNDS1 only in case of binary operation. */ | |
2492 | if (!slp_node) | |
2493 | { | |
2494 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
2495 | vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
2496 | } | |
2497 | else if (scalar_shift_arg) | |
2498 | vec_oprnds1 = VEC_alloc (tree, heap, slp_node->vec_stmts_size); | |
2499 | ||
2500 | prev_stmt_info = NULL; | |
2501 | for (j = 0; j < ncopies; j++) | |
2502 | { | |
2503 | /* Handle uses. */ | |
2504 | if (j == 0) | |
2505 | { | |
2506 | if (scalar_shift_arg) | |
2507 | { | |
2508 | /* Vector shl and shr insn patterns can be defined with scalar | |
2509 | operand 2 (shift operand). In this case, use constant or loop | |
2510 | invariant op1 directly, without extending it to vector mode | |
2511 | first. */ | |
2512 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
2513 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
2514 | { | |
2515 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2516 | fprintf (vect_dump, "operand 1 using scalar mode."); | |
2517 | vec_oprnd1 = op1; | |
2518 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2519 | if (slp_node) | |
2520 | { | |
2521 | /* Store vec_oprnd1 for every vector stmt to be created | |
2522 | for SLP_NODE. We check during the analysis that all | |
2523 | the shift arguments are the same. | |
2524 | TODO: Allow different constants for different vector | |
2525 | stmts generated for an SLP instance. */ | |
2526 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
2527 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2528 | } | |
2529 | } | |
2530 | } | |
2531 | ||
2532 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
2533 | (a special case for certain kind of vector shifts); otherwise, | |
2534 | operand 1 should be of a vector type (the usual case). */ | |
2535 | if (vec_oprnd1) | |
2536 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2537 | slp_node); | |
2538 | else | |
2539 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
2540 | slp_node); | |
2541 | } | |
2542 | else | |
2543 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
2544 | ||
2545 | /* Arguments are ready. Create the new vector stmt. */ | |
2546 | FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0) | |
2547 | { | |
2548 | vop1 = VEC_index (tree, vec_oprnds1, i); | |
2549 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); | |
2550 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2551 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2552 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2553 | if (slp_node) | |
2554 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2555 | } | |
2556 | ||
2557 | if (slp_node) | |
2558 | continue; | |
2559 | ||
2560 | if (j == 0) | |
2561 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2562 | else | |
2563 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2564 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2565 | } | |
2566 | ||
2567 | VEC_free (tree, heap, vec_oprnds0); | |
2568 | VEC_free (tree, heap, vec_oprnds1); | |
2569 | ||
2570 | return true; | |
2571 | } | |
2572 | ||
2573 | ||
ebfd146a IR |
2574 | /* Function vectorizable_operation. |
2575 | ||
16949072 RG |
2576 | Check if STMT performs a binary, unary or ternary operation that can |
2577 | be vectorized. | |
b8698a0f | 2578 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2579 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2580 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2581 | ||
2582 | static bool | |
2583 | vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi, | |
2584 | gimple *vec_stmt, slp_tree slp_node) | |
2585 | { | |
2586 | tree vec_dest; | |
2587 | tree scalar_dest; | |
16949072 | 2588 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 2589 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
b690cc0f | 2590 | tree vectype; |
ebfd146a IR |
2591 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
2592 | enum tree_code code; | |
2593 | enum machine_mode vec_mode; | |
2594 | tree new_temp; | |
2595 | int op_type; | |
2596 | optab optab; | |
2597 | int icode; | |
ebfd146a IR |
2598 | tree def; |
2599 | gimple def_stmt; | |
16949072 RG |
2600 | enum vect_def_type dt[3] |
2601 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
ebfd146a IR |
2602 | gimple new_stmt = NULL; |
2603 | stmt_vec_info prev_stmt_info; | |
b690cc0f | 2604 | int nunits_in; |
ebfd146a IR |
2605 | int nunits_out; |
2606 | tree vectype_out; | |
2607 | int ncopies; | |
2608 | int j, i; | |
16949072 RG |
2609 | VEC(tree,heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL, *vec_oprnds2 = NULL; |
2610 | tree vop0, vop1, vop2; | |
a70d6342 IR |
2611 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2612 | int vf; | |
2613 | ||
a70d6342 | 2614 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2615 | return false; |
2616 | ||
8644a673 | 2617 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2618 | return false; |
2619 | ||
2620 | /* Is STMT a vectorizable binary/unary operation? */ | |
2621 | if (!is_gimple_assign (stmt)) | |
2622 | return false; | |
2623 | ||
2624 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2625 | return false; | |
2626 | ||
ebfd146a IR |
2627 | code = gimple_assign_rhs_code (stmt); |
2628 | ||
2629 | /* For pointer addition, we should use the normal plus for | |
2630 | the vector addition. */ | |
2631 | if (code == POINTER_PLUS_EXPR) | |
2632 | code = PLUS_EXPR; | |
2633 | ||
2634 | /* Support only unary or binary operations. */ | |
2635 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 2636 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a IR |
2637 | { |
2638 | if (vect_print_dump_info (REPORT_DETAILS)) | |
16949072 RG |
2639 | fprintf (vect_dump, "num. args = %d (not unary/binary/ternary op).", |
2640 | op_type); | |
ebfd146a IR |
2641 | return false; |
2642 | } | |
2643 | ||
b690cc0f RG |
2644 | scalar_dest = gimple_assign_lhs (stmt); |
2645 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2646 | ||
ebfd146a | 2647 | op0 = gimple_assign_rhs1 (stmt); |
b690cc0f RG |
2648 | if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo, |
2649 | &def_stmt, &def, &dt[0], &vectype)) | |
ebfd146a IR |
2650 | { |
2651 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2652 | fprintf (vect_dump, "use not simple."); | |
2653 | return false; | |
2654 | } | |
b690cc0f RG |
2655 | /* If op0 is an external or constant def use a vector type with |
2656 | the same size as the output vector type. */ | |
2657 | if (!vectype) | |
2658 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
2659 | if (vec_stmt) |
2660 | gcc_assert (vectype); | |
2661 | if (!vectype) | |
2662 | { | |
2663 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2664 | { | |
2665 | fprintf (vect_dump, "no vectype for scalar type "); | |
2666 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
2667 | } | |
2668 | ||
2669 | return false; | |
2670 | } | |
b690cc0f RG |
2671 | |
2672 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
2673 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
2674 | if (nunits_out != nunits_in) | |
2675 | return false; | |
ebfd146a | 2676 | |
16949072 | 2677 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
2678 | { |
2679 | op1 = gimple_assign_rhs2 (stmt); | |
b8698a0f | 2680 | if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def, |
a70d6342 | 2681 | &dt[1])) |
ebfd146a IR |
2682 | { |
2683 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2684 | fprintf (vect_dump, "use not simple."); | |
2685 | return false; | |
2686 | } | |
2687 | } | |
16949072 RG |
2688 | if (op_type == ternary_op) |
2689 | { | |
2690 | op2 = gimple_assign_rhs3 (stmt); | |
2691 | if (!vect_is_simple_use (op2, loop_vinfo, bb_vinfo, &def_stmt, &def, | |
2692 | &dt[2])) | |
2693 | { | |
2694 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2695 | fprintf (vect_dump, "use not simple."); | |
2696 | return false; | |
2697 | } | |
2698 | } | |
ebfd146a | 2699 | |
b690cc0f RG |
2700 | if (loop_vinfo) |
2701 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
2702 | else | |
2703 | vf = 1; | |
2704 | ||
2705 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 2706 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 2707 | case of SLP. */ |
437f4a00 | 2708 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
b690cc0f RG |
2709 | ncopies = 1; |
2710 | else | |
2711 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2712 | ||
2713 | gcc_assert (ncopies >= 1); | |
2714 | ||
9dc3f7de | 2715 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
2716 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
2717 | || code == RROTATE_EXPR) | |
9dc3f7de | 2718 | return false; |
ebfd146a | 2719 | |
16949072 | 2720 | optab = optab_for_tree_code (code, vectype, optab_default); |
ebfd146a IR |
2721 | |
2722 | /* Supportable by target? */ | |
2723 | if (!optab) | |
2724 | { | |
2725 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2726 | fprintf (vect_dump, "no optab."); | |
2727 | return false; | |
2728 | } | |
2729 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 2730 | icode = (int) optab_handler (optab, vec_mode); |
ebfd146a IR |
2731 | if (icode == CODE_FOR_nothing) |
2732 | { | |
2733 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2734 | fprintf (vect_dump, "op not supported by target."); | |
2735 | /* Check only during analysis. */ | |
2736 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
a70d6342 | 2737 | || (vf < vect_min_worthwhile_factor (code) |
ebfd146a IR |
2738 | && !vec_stmt)) |
2739 | return false; | |
2740 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2741 | fprintf (vect_dump, "proceeding using word mode."); | |
2742 | } | |
2743 | ||
ff802fa1 | 2744 | /* Worthwhile without SIMD support? Check only during analysis. */ |
ebfd146a | 2745 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) |
a70d6342 | 2746 | && vf < vect_min_worthwhile_factor (code) |
ebfd146a IR |
2747 | && !vec_stmt) |
2748 | { | |
2749 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2750 | fprintf (vect_dump, "not worthwhile without SIMD support."); | |
2751 | return false; | |
2752 | } | |
2753 | ||
2754 | if (!vec_stmt) /* transformation not required. */ | |
2755 | { | |
2756 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; | |
2757 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2758 | fprintf (vect_dump, "=== vectorizable_operation ==="); | |
2759 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2760 | return true; | |
2761 | } | |
2762 | ||
2763 | /** Transform. **/ | |
2764 | ||
2765 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2766 | fprintf (vect_dump, "transform binary/unary operation."); | |
2767 | ||
2768 | /* Handle def. */ | |
2769 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2770 | ||
ff802fa1 | 2771 | /* Allocate VECs for vector operands. In case of SLP, vector operands are |
ebfd146a | 2772 | created in the previous stages of the recursion, so no allocation is |
ff802fa1 | 2773 | needed, except for the case of shift with scalar shift argument. In that |
ebfd146a IR |
2774 | case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to |
2775 | be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE. | |
ff802fa1 | 2776 | In case of loop-based vectorization we allocate VECs of size 1. We |
b8698a0f | 2777 | allocate VEC_OPRNDS1 only in case of binary operation. */ |
ebfd146a IR |
2778 | if (!slp_node) |
2779 | { | |
2780 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
16949072 | 2781 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a | 2782 | vec_oprnds1 = VEC_alloc (tree, heap, 1); |
16949072 RG |
2783 | if (op_type == ternary_op) |
2784 | vec_oprnds2 = VEC_alloc (tree, heap, 1); | |
ebfd146a | 2785 | } |
ebfd146a IR |
2786 | |
2787 | /* In case the vectorization factor (VF) is bigger than the number | |
2788 | of elements that we can fit in a vectype (nunits), we have to generate | |
2789 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 2790 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 2791 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 2792 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 2793 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
2794 | stmts that use the defs of the current stmt. The example below |
2795 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
2796 | we need to create 4 vectorized stmts): | |
ebfd146a IR |
2797 | |
2798 | before vectorization: | |
2799 | RELATED_STMT VEC_STMT | |
2800 | S1: x = memref - - | |
2801 | S2: z = x + 1 - - | |
2802 | ||
2803 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
2804 | there): | |
2805 | RELATED_STMT VEC_STMT | |
2806 | VS1_0: vx0 = memref0 VS1_1 - | |
2807 | VS1_1: vx1 = memref1 VS1_2 - | |
2808 | VS1_2: vx2 = memref2 VS1_3 - | |
2809 | VS1_3: vx3 = memref3 - - | |
2810 | S1: x = load - VS1_0 | |
2811 | S2: z = x + 1 - - | |
2812 | ||
2813 | step2: vectorize stmt S2 (done here): | |
2814 | To vectorize stmt S2 we first need to find the relevant vector | |
ff802fa1 | 2815 | def for the first operand 'x'. This is, as usual, obtained from |
ebfd146a | 2816 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt |
ff802fa1 IR |
2817 | that defines 'x' (S1). This way we find the stmt VS1_0, and the |
2818 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
ebfd146a IR |
2819 | the vector stmt VS2_0, and as usual, record it in the |
2820 | STMT_VINFO_VEC_STMT of stmt S2. | |
2821 | When creating the second copy (VS2_1), we obtain the relevant vector | |
2822 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
ff802fa1 IR |
2823 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant |
2824 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
ebfd146a | 2825 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. |
ff802fa1 | 2826 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting |
ebfd146a IR |
2827 | chain of stmts and pointers: |
2828 | RELATED_STMT VEC_STMT | |
2829 | VS1_0: vx0 = memref0 VS1_1 - | |
2830 | VS1_1: vx1 = memref1 VS1_2 - | |
2831 | VS1_2: vx2 = memref2 VS1_3 - | |
2832 | VS1_3: vx3 = memref3 - - | |
2833 | S1: x = load - VS1_0 | |
2834 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
2835 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
2836 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
2837 | VS2_3: vz3 = vx3 + v1 - - | |
2838 | S2: z = x + 1 - VS2_0 */ | |
2839 | ||
2840 | prev_stmt_info = NULL; | |
2841 | for (j = 0; j < ncopies; j++) | |
2842 | { | |
2843 | /* Handle uses. */ | |
2844 | if (j == 0) | |
2845 | { | |
16949072 | 2846 | if (op_type == binary_op || op_type == ternary_op) |
b8698a0f | 2847 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, |
ebfd146a IR |
2848 | slp_node); |
2849 | else | |
b8698a0f | 2850 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, |
ebfd146a | 2851 | slp_node); |
16949072 RG |
2852 | if (op_type == ternary_op) |
2853 | { | |
2854 | vec_oprnds2 = VEC_alloc (tree, heap, 1); | |
2855 | VEC_quick_push (tree, vec_oprnds2, | |
2856 | vect_get_vec_def_for_operand (op2, stmt, NULL)); | |
2857 | } | |
ebfd146a IR |
2858 | } |
2859 | else | |
16949072 RG |
2860 | { |
2861 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
2862 | if (op_type == ternary_op) | |
2863 | { | |
2864 | tree vec_oprnd = VEC_pop (tree, vec_oprnds2); | |
2865 | VEC_quick_push (tree, vec_oprnds2, | |
2866 | vect_get_vec_def_for_stmt_copy (dt[2], | |
2867 | vec_oprnd)); | |
2868 | } | |
2869 | } | |
ebfd146a | 2870 | |
9dc3f7de | 2871 | /* Arguments are ready. Create the new vector stmt. */ |
ac47786e | 2872 | FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0) |
ebfd146a | 2873 | { |
16949072 RG |
2874 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
2875 | ? VEC_index (tree, vec_oprnds1, i) : NULL_TREE); | |
2876 | vop2 = ((op_type == ternary_op) | |
2877 | ? VEC_index (tree, vec_oprnds2, i) : NULL_TREE); | |
2878 | new_stmt = gimple_build_assign_with_ops3 (code, vec_dest, | |
2879 | vop0, vop1, vop2); | |
ebfd146a IR |
2880 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2881 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2882 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2883 | if (slp_node) | |
2884 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2885 | } | |
2886 | ||
2887 | if (slp_node) | |
2888 | continue; | |
2889 | ||
2890 | if (j == 0) | |
2891 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2892 | else | |
2893 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2894 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2895 | } | |
2896 | ||
2897 | VEC_free (tree, heap, vec_oprnds0); | |
2898 | if (vec_oprnds1) | |
2899 | VEC_free (tree, heap, vec_oprnds1); | |
16949072 RG |
2900 | if (vec_oprnds2) |
2901 | VEC_free (tree, heap, vec_oprnds2); | |
ebfd146a IR |
2902 | |
2903 | return true; | |
2904 | } | |
2905 | ||
2906 | ||
ff802fa1 | 2907 | /* Get vectorized definitions for loop-based vectorization. For the first |
b8698a0f L |
2908 | operand we call vect_get_vec_def_for_operand() (with OPRND containing |
2909 | scalar operand), and for the rest we get a copy with | |
ebfd146a IR |
2910 | vect_get_vec_def_for_stmt_copy() using the previous vector definition |
2911 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
2912 | The vectors are collected into VEC_OPRNDS. */ | |
2913 | ||
2914 | static void | |
b8698a0f | 2915 | vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt, |
ebfd146a IR |
2916 | VEC (tree, heap) **vec_oprnds, int multi_step_cvt) |
2917 | { | |
2918 | tree vec_oprnd; | |
2919 | ||
2920 | /* Get first vector operand. */ | |
2921 | /* All the vector operands except the very first one (that is scalar oprnd) | |
2922 | are stmt copies. */ | |
b8698a0f | 2923 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) |
ebfd146a IR |
2924 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL); |
2925 | else | |
2926 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
2927 | ||
2928 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
2929 | ||
2930 | /* Get second vector operand. */ | |
2931 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
2932 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
b8698a0f | 2933 | |
ebfd146a IR |
2934 | *oprnd = vec_oprnd; |
2935 | ||
b8698a0f | 2936 | /* For conversion in multiple steps, continue to get operands |
ebfd146a IR |
2937 | recursively. */ |
2938 | if (multi_step_cvt) | |
b8698a0f | 2939 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); |
ebfd146a IR |
2940 | } |
2941 | ||
2942 | ||
2943 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
b8698a0f | 2944 | For multi-step conversions store the resulting vectors and call the function |
ebfd146a IR |
2945 | recursively. */ |
2946 | ||
2947 | static void | |
2948 | vect_create_vectorized_demotion_stmts (VEC (tree, heap) **vec_oprnds, | |
2949 | int multi_step_cvt, gimple stmt, | |
2950 | VEC (tree, heap) *vec_dsts, | |
2951 | gimple_stmt_iterator *gsi, | |
2952 | slp_tree slp_node, enum tree_code code, | |
2953 | stmt_vec_info *prev_stmt_info) | |
2954 | { | |
2955 | unsigned int i; | |
2956 | tree vop0, vop1, new_tmp, vec_dest; | |
2957 | gimple new_stmt; | |
2958 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2959 | ||
b8698a0f | 2960 | vec_dest = VEC_pop (tree, vec_dsts); |
ebfd146a IR |
2961 | |
2962 | for (i = 0; i < VEC_length (tree, *vec_oprnds); i += 2) | |
2963 | { | |
2964 | /* Create demotion operation. */ | |
2965 | vop0 = VEC_index (tree, *vec_oprnds, i); | |
2966 | vop1 = VEC_index (tree, *vec_oprnds, i + 1); | |
2967 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); | |
2968 | new_tmp = make_ssa_name (vec_dest, new_stmt); | |
2969 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
2970 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2971 | ||
2972 | if (multi_step_cvt) | |
2973 | /* Store the resulting vector for next recursive call. */ | |
b8698a0f | 2974 | VEC_replace (tree, *vec_oprnds, i/2, new_tmp); |
ebfd146a IR |
2975 | else |
2976 | { | |
b8698a0f | 2977 | /* This is the last step of the conversion sequence. Store the |
ebfd146a IR |
2978 | vectors in SLP_NODE or in vector info of the scalar statement |
2979 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
2980 | if (slp_node) | |
2981 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2982 | else | |
2983 | { | |
2984 | if (!*prev_stmt_info) | |
2985 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2986 | else | |
2987 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
2988 | ||
2989 | *prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2990 | } | |
2991 | } | |
2992 | } | |
2993 | ||
2994 | /* For multi-step demotion operations we first generate demotion operations | |
b8698a0f | 2995 | from the source type to the intermediate types, and then combine the |
ebfd146a IR |
2996 | results (stored in VEC_OPRNDS) in demotion operation to the destination |
2997 | type. */ | |
2998 | if (multi_step_cvt) | |
2999 | { | |
3000 | /* At each level of recursion we have have of the operands we had at the | |
3001 | previous level. */ | |
3002 | VEC_truncate (tree, *vec_oprnds, (i+1)/2); | |
b8698a0f | 3003 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
ebfd146a IR |
3004 | stmt, vec_dsts, gsi, slp_node, |
3005 | code, prev_stmt_info); | |
3006 | } | |
3007 | } | |
3008 | ||
3009 | ||
3010 | /* Function vectorizable_type_demotion | |
3011 | ||
3012 | Check if STMT performs a binary or unary operation that involves | |
3013 | type demotion, and if it can be vectorized. | |
3014 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3015 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3016 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3017 | ||
3018 | static bool | |
3019 | vectorizable_type_demotion (gimple stmt, gimple_stmt_iterator *gsi, | |
3020 | gimple *vec_stmt, slp_tree slp_node) | |
3021 | { | |
3022 | tree vec_dest; | |
3023 | tree scalar_dest; | |
3024 | tree op0; | |
3025 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3026 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3027 | enum tree_code code, code1 = ERROR_MARK; | |
3028 | tree def; | |
3029 | gimple def_stmt; | |
3030 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
3031 | stmt_vec_info prev_stmt_info; | |
3032 | int nunits_in; | |
3033 | int nunits_out; | |
3034 | tree vectype_out; | |
3035 | int ncopies; | |
3036 | int j, i; | |
3037 | tree vectype_in; | |
3038 | int multi_step_cvt = 0; | |
3039 | VEC (tree, heap) *vec_oprnds0 = NULL; | |
3040 | VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL; | |
3041 | tree last_oprnd, intermediate_type; | |
3042 | ||
a70d6342 IR |
3043 | /* FORNOW: not supported by basic block SLP vectorization. */ |
3044 | gcc_assert (loop_vinfo); | |
3045 | ||
ebfd146a IR |
3046 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
3047 | return false; | |
3048 | ||
8644a673 | 3049 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3050 | return false; |
3051 | ||
3052 | /* Is STMT a vectorizable type-demotion operation? */ | |
3053 | if (!is_gimple_assign (stmt)) | |
3054 | return false; | |
3055 | ||
3056 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3057 | return false; | |
3058 | ||
3059 | code = gimple_assign_rhs_code (stmt); | |
3060 | if (!CONVERT_EXPR_CODE_P (code)) | |
3061 | return false; | |
3062 | ||
b690cc0f RG |
3063 | scalar_dest = gimple_assign_lhs (stmt); |
3064 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
3065 | ||
3066 | /* Check the operands of the operation. */ | |
ebfd146a | 3067 | op0 = gimple_assign_rhs1 (stmt); |
b690cc0f RG |
3068 | if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) |
3069 | && INTEGRAL_TYPE_P (TREE_TYPE (op0))) | |
3070 | || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest)) | |
3071 | && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0)) | |
3072 | && CONVERT_EXPR_CODE_P (code)))) | |
3073 | return false; | |
3074 | if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL, | |
3075 | &def_stmt, &def, &dt[0], &vectype_in)) | |
3076 | { | |
3077 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3078 | fprintf (vect_dump, "use not simple."); | |
3079 | return false; | |
3080 | } | |
3081 | /* If op0 is an external def use a vector type with the | |
3082 | same size as the output vector type if possible. */ | |
3083 | if (!vectype_in) | |
3084 | vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
3085 | if (vec_stmt) |
3086 | gcc_assert (vectype_in); | |
ebfd146a | 3087 | if (!vectype_in) |
7d8930a0 IR |
3088 | { |
3089 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3090 | { | |
3091 | fprintf (vect_dump, "no vectype for scalar type "); | |
3092 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
3093 | } | |
3094 | ||
3095 | return false; | |
3096 | } | |
ebfd146a | 3097 | |
b690cc0f | 3098 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
ebfd146a IR |
3099 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); |
3100 | if (nunits_in >= nunits_out) | |
3101 | return false; | |
3102 | ||
3103 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 3104 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 3105 | case of SLP. */ |
437f4a00 | 3106 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
3107 | ncopies = 1; |
3108 | else | |
3109 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
ebfd146a IR |
3110 | gcc_assert (ncopies >= 1); |
3111 | ||
ebfd146a | 3112 | /* Supportable by target? */ |
b690cc0f RG |
3113 | if (!supportable_narrowing_operation (code, vectype_out, vectype_in, |
3114 | &code1, &multi_step_cvt, &interm_types)) | |
ebfd146a IR |
3115 | return false; |
3116 | ||
ebfd146a IR |
3117 | if (!vec_stmt) /* transformation not required. */ |
3118 | { | |
3119 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
3120 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3121 | fprintf (vect_dump, "=== vectorizable_demotion ==="); | |
3122 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
3123 | return true; | |
3124 | } | |
3125 | ||
3126 | /** Transform. **/ | |
3127 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3128 | fprintf (vect_dump, "transform type demotion operation. ncopies = %d.", | |
3129 | ncopies); | |
3130 | ||
b8698a0f L |
3131 | /* In case of multi-step demotion, we first generate demotion operations to |
3132 | the intermediate types, and then from that types to the final one. | |
ebfd146a | 3133 | We create vector destinations for the intermediate type (TYPES) received |
b8698a0f | 3134 | from supportable_narrowing_operation, and store them in the correct order |
ebfd146a IR |
3135 | for future use in vect_create_vectorized_demotion_stmts(). */ |
3136 | if (multi_step_cvt) | |
3137 | vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1); | |
3138 | else | |
3139 | vec_dsts = VEC_alloc (tree, heap, 1); | |
b8698a0f | 3140 | |
ebfd146a IR |
3141 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); |
3142 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
3143 | ||
3144 | if (multi_step_cvt) | |
3145 | { | |
b8698a0f | 3146 | for (i = VEC_length (tree, interm_types) - 1; |
ebfd146a IR |
3147 | VEC_iterate (tree, interm_types, i, intermediate_type); i--) |
3148 | { | |
b8698a0f | 3149 | vec_dest = vect_create_destination_var (scalar_dest, |
ebfd146a IR |
3150 | intermediate_type); |
3151 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
3152 | } | |
3153 | } | |
3154 | ||
3155 | /* In case the vectorization factor (VF) is bigger than the number | |
3156 | of elements that we can fit in a vectype (nunits), we have to generate | |
3157 | more than one vector stmt - i.e - we need to "unroll" the | |
3158 | vector stmt by a factor VF/nunits. */ | |
3159 | last_oprnd = op0; | |
3160 | prev_stmt_info = NULL; | |
3161 | for (j = 0; j < ncopies; j++) | |
3162 | { | |
3163 | /* Handle uses. */ | |
3164 | if (slp_node) | |
9dc3f7de | 3165 | vect_get_slp_defs (op0, NULL_TREE, slp_node, &vec_oprnds0, NULL, -1); |
ebfd146a IR |
3166 | else |
3167 | { | |
3168 | VEC_free (tree, heap, vec_oprnds0); | |
3169 | vec_oprnds0 = VEC_alloc (tree, heap, | |
3170 | (multi_step_cvt ? vect_pow2 (multi_step_cvt) * 2 : 2)); | |
b8698a0f | 3171 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
ebfd146a IR |
3172 | vect_pow2 (multi_step_cvt) - 1); |
3173 | } | |
3174 | ||
9dc3f7de | 3175 | /* Arguments are ready. Create the new vector stmts. */ |
ebfd146a | 3176 | tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts); |
b8698a0f | 3177 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, |
ebfd146a | 3178 | multi_step_cvt, stmt, tmp_vec_dsts, |
b8698a0f | 3179 | gsi, slp_node, code1, |
ebfd146a IR |
3180 | &prev_stmt_info); |
3181 | } | |
3182 | ||
3183 | VEC_free (tree, heap, vec_oprnds0); | |
3184 | VEC_free (tree, heap, vec_dsts); | |
3185 | VEC_free (tree, heap, tmp_vec_dsts); | |
3186 | VEC_free (tree, heap, interm_types); | |
3187 | ||
3188 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
3189 | return true; | |
3190 | } | |
3191 | ||
3192 | ||
3193 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
ff802fa1 | 3194 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store |
ebfd146a IR |
3195 | the resulting vectors and call the function recursively. */ |
3196 | ||
3197 | static void | |
3198 | vect_create_vectorized_promotion_stmts (VEC (tree, heap) **vec_oprnds0, | |
3199 | VEC (tree, heap) **vec_oprnds1, | |
3200 | int multi_step_cvt, gimple stmt, | |
3201 | VEC (tree, heap) *vec_dsts, | |
3202 | gimple_stmt_iterator *gsi, | |
3203 | slp_tree slp_node, enum tree_code code1, | |
b8698a0f | 3204 | enum tree_code code2, tree decl1, |
ebfd146a IR |
3205 | tree decl2, int op_type, |
3206 | stmt_vec_info *prev_stmt_info) | |
3207 | { | |
3208 | int i; | |
3209 | tree vop0, vop1, new_tmp1, new_tmp2, vec_dest; | |
3210 | gimple new_stmt1, new_stmt2; | |
3211 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3212 | VEC (tree, heap) *vec_tmp; | |
3213 | ||
3214 | vec_dest = VEC_pop (tree, vec_dsts); | |
3215 | vec_tmp = VEC_alloc (tree, heap, VEC_length (tree, *vec_oprnds0) * 2); | |
3216 | ||
ac47786e | 3217 | FOR_EACH_VEC_ELT (tree, *vec_oprnds0, i, vop0) |
ebfd146a IR |
3218 | { |
3219 | if (op_type == binary_op) | |
3220 | vop1 = VEC_index (tree, *vec_oprnds1, i); | |
3221 | else | |
3222 | vop1 = NULL_TREE; | |
3223 | ||
3224 | /* Generate the two halves of promotion operation. */ | |
b8698a0f | 3225 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, |
ebfd146a IR |
3226 | op_type, vec_dest, gsi, stmt); |
3227 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
3228 | op_type, vec_dest, gsi, stmt); | |
3229 | if (is_gimple_call (new_stmt1)) | |
3230 | { | |
3231 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
3232 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
3233 | } | |
3234 | else | |
3235 | { | |
3236 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
3237 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
3238 | } | |
3239 | ||
3240 | if (multi_step_cvt) | |
3241 | { | |
3242 | /* Store the results for the recursive call. */ | |
3243 | VEC_quick_push (tree, vec_tmp, new_tmp1); | |
3244 | VEC_quick_push (tree, vec_tmp, new_tmp2); | |
3245 | } | |
3246 | else | |
3247 | { | |
3248 | /* Last step of promotion sequience - store the results. */ | |
3249 | if (slp_node) | |
3250 | { | |
3251 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt1); | |
3252 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt2); | |
3253 | } | |
3254 | else | |
3255 | { | |
3256 | if (!*prev_stmt_info) | |
3257 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt1; | |
3258 | else | |
3259 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt1; | |
3260 | ||
3261 | *prev_stmt_info = vinfo_for_stmt (new_stmt1); | |
3262 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt2; | |
3263 | *prev_stmt_info = vinfo_for_stmt (new_stmt2); | |
3264 | } | |
3265 | } | |
3266 | } | |
3267 | ||
3268 | if (multi_step_cvt) | |
3269 | { | |
b8698a0f | 3270 | /* For multi-step promotion operation we first generate we call the |
ff802fa1 | 3271 | function recurcively for every stage. We start from the input type, |
ebfd146a IR |
3272 | create promotion operations to the intermediate types, and then |
3273 | create promotions to the output type. */ | |
3274 | *vec_oprnds0 = VEC_copy (tree, heap, vec_tmp); | |
ebfd146a IR |
3275 | vect_create_vectorized_promotion_stmts (vec_oprnds0, vec_oprnds1, |
3276 | multi_step_cvt - 1, stmt, | |
3277 | vec_dsts, gsi, slp_node, code1, | |
3278 | code2, decl2, decl2, op_type, | |
3279 | prev_stmt_info); | |
3280 | } | |
ff802fa1 IR |
3281 | |
3282 | VEC_free (tree, heap, vec_tmp); | |
ebfd146a | 3283 | } |
b8698a0f | 3284 | |
ebfd146a IR |
3285 | |
3286 | /* Function vectorizable_type_promotion | |
3287 | ||
3288 | Check if STMT performs a binary or unary operation that involves | |
3289 | type promotion, and if it can be vectorized. | |
3290 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3291 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3292 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3293 | ||
3294 | static bool | |
3295 | vectorizable_type_promotion (gimple stmt, gimple_stmt_iterator *gsi, | |
3296 | gimple *vec_stmt, slp_tree slp_node) | |
3297 | { | |
3298 | tree vec_dest; | |
3299 | tree scalar_dest; | |
3300 | tree op0, op1 = NULL; | |
3301 | tree vec_oprnd0=NULL, vec_oprnd1=NULL; | |
3302 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3303 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3304 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
3305 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; | |
b8698a0f | 3306 | int op_type; |
ebfd146a IR |
3307 | tree def; |
3308 | gimple def_stmt; | |
3309 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
3310 | stmt_vec_info prev_stmt_info; | |
3311 | int nunits_in; | |
3312 | int nunits_out; | |
3313 | tree vectype_out; | |
3314 | int ncopies; | |
3315 | int j, i; | |
3316 | tree vectype_in; | |
3317 | tree intermediate_type = NULL_TREE; | |
3318 | int multi_step_cvt = 0; | |
3319 | VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; | |
3320 | VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL; | |
b8698a0f | 3321 | |
a70d6342 IR |
3322 | /* FORNOW: not supported by basic block SLP vectorization. */ |
3323 | gcc_assert (loop_vinfo); | |
b8698a0f | 3324 | |
ebfd146a IR |
3325 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
3326 | return false; | |
3327 | ||
8644a673 | 3328 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3329 | return false; |
3330 | ||
3331 | /* Is STMT a vectorizable type-promotion operation? */ | |
3332 | if (!is_gimple_assign (stmt)) | |
3333 | return false; | |
3334 | ||
3335 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3336 | return false; | |
3337 | ||
3338 | code = gimple_assign_rhs_code (stmt); | |
3339 | if (!CONVERT_EXPR_CODE_P (code) | |
3340 | && code != WIDEN_MULT_EXPR) | |
3341 | return false; | |
3342 | ||
b690cc0f RG |
3343 | scalar_dest = gimple_assign_lhs (stmt); |
3344 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
3345 | ||
3346 | /* Check the operands of the operation. */ | |
ebfd146a | 3347 | op0 = gimple_assign_rhs1 (stmt); |
b690cc0f RG |
3348 | if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) |
3349 | && INTEGRAL_TYPE_P (TREE_TYPE (op0))) | |
3350 | || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest)) | |
3351 | && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0)) | |
3352 | && CONVERT_EXPR_CODE_P (code)))) | |
3353 | return false; | |
3354 | if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL, | |
3355 | &def_stmt, &def, &dt[0], &vectype_in)) | |
3356 | { | |
3357 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3358 | fprintf (vect_dump, "use not simple."); | |
3359 | return false; | |
3360 | } | |
383d9c83 IR |
3361 | |
3362 | op_type = TREE_CODE_LENGTH (code); | |
3363 | if (op_type == binary_op) | |
3364 | { | |
3365 | bool ok; | |
3366 | ||
3367 | op1 = gimple_assign_rhs2 (stmt); | |
3368 | if (code == WIDEN_MULT_EXPR) | |
3369 | { | |
3370 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
3371 | OP1. */ | |
3372 | if (CONSTANT_CLASS_P (op0)) | |
3373 | ok = vect_is_simple_use_1 (op1, loop_vinfo, NULL, | |
3374 | &def_stmt, &def, &dt[1], &vectype_in); | |
3375 | else | |
3376 | ok = vect_is_simple_use (op1, loop_vinfo, NULL, &def_stmt, &def, | |
3377 | &dt[1]); | |
3378 | ||
3379 | if (!ok) | |
3380 | { | |
3381 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3382 | fprintf (vect_dump, "use not simple."); | |
3383 | return false; | |
3384 | } | |
3385 | } | |
3386 | } | |
3387 | ||
b690cc0f RG |
3388 | /* If op0 is an external or constant def use a vector type with |
3389 | the same size as the output vector type. */ | |
3390 | if (!vectype_in) | |
3391 | vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
3392 | if (vec_stmt) |
3393 | gcc_assert (vectype_in); | |
ebfd146a | 3394 | if (!vectype_in) |
7d8930a0 IR |
3395 | { |
3396 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3397 | { | |
3398 | fprintf (vect_dump, "no vectype for scalar type "); | |
3399 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
3400 | } | |
3401 | ||
3402 | return false; | |
3403 | } | |
ebfd146a | 3404 | |
b690cc0f | 3405 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
ebfd146a IR |
3406 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); |
3407 | if (nunits_in <= nunits_out) | |
3408 | return false; | |
3409 | ||
3410 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 3411 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 3412 | case of SLP. */ |
437f4a00 | 3413 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
3414 | ncopies = 1; |
3415 | else | |
3416 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
3417 | ||
3418 | gcc_assert (ncopies >= 1); | |
3419 | ||
ebfd146a | 3420 | /* Supportable by target? */ |
b690cc0f | 3421 | if (!supportable_widening_operation (code, stmt, vectype_out, vectype_in, |
ebfd146a IR |
3422 | &decl1, &decl2, &code1, &code2, |
3423 | &multi_step_cvt, &interm_types)) | |
3424 | return false; | |
3425 | ||
3426 | /* Binary widening operation can only be supported directly by the | |
3427 | architecture. */ | |
3428 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
3429 | ||
ebfd146a IR |
3430 | if (!vec_stmt) /* transformation not required. */ |
3431 | { | |
3432 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
3433 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3434 | fprintf (vect_dump, "=== vectorizable_promotion ==="); | |
3435 | vect_model_simple_cost (stmt_info, 2*ncopies, dt, NULL); | |
3436 | return true; | |
3437 | } | |
3438 | ||
3439 | /** Transform. **/ | |
3440 | ||
3441 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3442 | fprintf (vect_dump, "transform type promotion operation. ncopies = %d.", | |
3443 | ncopies); | |
3444 | ||
383d9c83 IR |
3445 | if (code == WIDEN_MULT_EXPR) |
3446 | { | |
3447 | if (CONSTANT_CLASS_P (op0)) | |
3448 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
3449 | else if (CONSTANT_CLASS_P (op1)) | |
3450 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
3451 | } | |
3452 | ||
ebfd146a | 3453 | /* Handle def. */ |
b8698a0f | 3454 | /* In case of multi-step promotion, we first generate promotion operations |
ebfd146a | 3455 | to the intermediate types, and then from that types to the final one. |
b8698a0f L |
3456 | We store vector destination in VEC_DSTS in the correct order for |
3457 | recursive creation of promotion operations in | |
ebfd146a IR |
3458 | vect_create_vectorized_promotion_stmts(). Vector destinations are created |
3459 | according to TYPES recieved from supportable_widening_operation(). */ | |
3460 | if (multi_step_cvt) | |
3461 | vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1); | |
3462 | else | |
3463 | vec_dsts = VEC_alloc (tree, heap, 1); | |
3464 | ||
3465 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
3466 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
3467 | ||
3468 | if (multi_step_cvt) | |
3469 | { | |
3470 | for (i = VEC_length (tree, interm_types) - 1; | |
3471 | VEC_iterate (tree, interm_types, i, intermediate_type); i--) | |
3472 | { | |
3473 | vec_dest = vect_create_destination_var (scalar_dest, | |
3474 | intermediate_type); | |
3475 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
3476 | } | |
3477 | } | |
b8698a0f | 3478 | |
ebfd146a IR |
3479 | if (!slp_node) |
3480 | { | |
b8698a0f | 3481 | vec_oprnds0 = VEC_alloc (tree, heap, |
ebfd146a IR |
3482 | (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); |
3483 | if (op_type == binary_op) | |
3484 | vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
3485 | } | |
3486 | ||
3487 | /* In case the vectorization factor (VF) is bigger than the number | |
3488 | of elements that we can fit in a vectype (nunits), we have to generate | |
3489 | more than one vector stmt - i.e - we need to "unroll" the | |
3490 | vector stmt by a factor VF/nunits. */ | |
3491 | ||
3492 | prev_stmt_info = NULL; | |
3493 | for (j = 0; j < ncopies; j++) | |
3494 | { | |
3495 | /* Handle uses. */ | |
3496 | if (j == 0) | |
3497 | { | |
3498 | if (slp_node) | |
9dc3f7de IR |
3499 | vect_get_slp_defs (op0, op1, slp_node, &vec_oprnds0, |
3500 | &vec_oprnds1, -1); | |
ebfd146a IR |
3501 | else |
3502 | { | |
3503 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
3504 | VEC_quick_push (tree, vec_oprnds0, vec_oprnd0); | |
3505 | if (op_type == binary_op) | |
3506 | { | |
3507 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, NULL); | |
3508 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
3509 | } | |
3510 | } | |
3511 | } | |
3512 | else | |
3513 | { | |
3514 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
3515 | VEC_replace (tree, vec_oprnds0, 0, vec_oprnd0); | |
3516 | if (op_type == binary_op) | |
3517 | { | |
3518 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd1); | |
3519 | VEC_replace (tree, vec_oprnds1, 0, vec_oprnd1); | |
3520 | } | |
3521 | } | |
3522 | ||
9dc3f7de | 3523 | /* Arguments are ready. Create the new vector stmts. */ |
ebfd146a IR |
3524 | tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts); |
3525 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, &vec_oprnds1, | |
b8698a0f | 3526 | multi_step_cvt, stmt, |
ebfd146a IR |
3527 | tmp_vec_dsts, |
3528 | gsi, slp_node, code1, code2, | |
3529 | decl1, decl2, op_type, | |
3530 | &prev_stmt_info); | |
3531 | } | |
3532 | ||
3533 | VEC_free (tree, heap, vec_dsts); | |
3534 | VEC_free (tree, heap, tmp_vec_dsts); | |
3535 | VEC_free (tree, heap, interm_types); | |
3536 | VEC_free (tree, heap, vec_oprnds0); | |
3537 | VEC_free (tree, heap, vec_oprnds1); | |
3538 | ||
3539 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
3540 | return true; | |
3541 | } | |
3542 | ||
3543 | ||
3544 | /* Function vectorizable_store. | |
3545 | ||
b8698a0f L |
3546 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
3547 | can be vectorized. | |
3548 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
3549 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3550 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3551 | ||
3552 | static bool | |
3553 | vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
3554 | slp_tree slp_node) | |
3555 | { | |
3556 | tree scalar_dest; | |
3557 | tree data_ref; | |
3558 | tree op; | |
3559 | tree vec_oprnd = NULL_TREE; | |
3560 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3561 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
3562 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
272c6793 | 3563 | tree elem_type; |
ebfd146a | 3564 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 3565 | struct loop *loop = NULL; |
ebfd146a IR |
3566 | enum machine_mode vec_mode; |
3567 | tree dummy; | |
3568 | enum dr_alignment_support alignment_support_scheme; | |
3569 | tree def; | |
3570 | gimple def_stmt; | |
3571 | enum vect_def_type dt; | |
3572 | stmt_vec_info prev_stmt_info = NULL; | |
3573 | tree dataref_ptr = NULL_TREE; | |
3574 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
3575 | int ncopies; | |
3576 | int j; | |
3577 | gimple next_stmt, first_stmt = NULL; | |
3578 | bool strided_store = false; | |
272c6793 | 3579 | bool store_lanes_p = false; |
ebfd146a IR |
3580 | unsigned int group_size, i; |
3581 | VEC(tree,heap) *dr_chain = NULL, *oprnds = NULL, *result_chain = NULL; | |
3582 | bool inv_p; | |
3583 | VEC(tree,heap) *vec_oprnds = NULL; | |
3584 | bool slp = (slp_node != NULL); | |
ebfd146a | 3585 | unsigned int vec_num; |
a70d6342 | 3586 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
272c6793 | 3587 | tree aggr_type; |
a70d6342 IR |
3588 | |
3589 | if (loop_vinfo) | |
3590 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
3591 | |
3592 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3593 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3594 | case of SLP. */ | |
437f4a00 | 3595 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
3596 | ncopies = 1; |
3597 | else | |
3598 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3599 | ||
3600 | gcc_assert (ncopies >= 1); | |
3601 | ||
3602 | /* FORNOW. This restriction should be relaxed. */ | |
a70d6342 | 3603 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) |
ebfd146a IR |
3604 | { |
3605 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3606 | fprintf (vect_dump, "multiple types in nested loop."); | |
3607 | return false; | |
3608 | } | |
3609 | ||
a70d6342 | 3610 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3611 | return false; |
3612 | ||
8644a673 | 3613 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3614 | return false; |
3615 | ||
3616 | /* Is vectorizable store? */ | |
3617 | ||
3618 | if (!is_gimple_assign (stmt)) | |
3619 | return false; | |
3620 | ||
3621 | scalar_dest = gimple_assign_lhs (stmt); | |
3622 | if (TREE_CODE (scalar_dest) != ARRAY_REF | |
3623 | && TREE_CODE (scalar_dest) != INDIRECT_REF | |
e9dbe7bb IR |
3624 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
3625 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
3626 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
3627 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
3628 | return false; |
3629 | ||
3630 | gcc_assert (gimple_assign_single_p (stmt)); | |
3631 | op = gimple_assign_rhs1 (stmt); | |
a70d6342 | 3632 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt)) |
ebfd146a IR |
3633 | { |
3634 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3635 | fprintf (vect_dump, "use not simple."); | |
3636 | return false; | |
3637 | } | |
3638 | ||
3639 | /* The scalar rhs type needs to be trivially convertible to the vector | |
3640 | component type. This should always be the case. */ | |
272c6793 RS |
3641 | elem_type = TREE_TYPE (vectype); |
3642 | if (!useless_type_conversion_p (elem_type, TREE_TYPE (op))) | |
b8698a0f | 3643 | { |
ebfd146a IR |
3644 | if (vect_print_dump_info (REPORT_DETAILS)) |
3645 | fprintf (vect_dump, "??? operands of different types"); | |
3646 | return false; | |
3647 | } | |
3648 | ||
3649 | vec_mode = TYPE_MODE (vectype); | |
3650 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
3651 | (e.g. - array initialization with 0). */ | |
947131ba | 3652 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
3653 | return false; |
3654 | ||
3655 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
3656 | return false; | |
3657 | ||
a1e53f3f L |
3658 | if (tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0) |
3659 | { | |
3660 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3661 | fprintf (vect_dump, "negative step for store."); | |
3662 | return false; | |
3663 | } | |
3664 | ||
ebfd146a IR |
3665 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) |
3666 | { | |
3667 | strided_store = true; | |
e14c1050 | 3668 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
3669 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
3670 | { | |
e14c1050 | 3671 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
3672 | if (vect_store_lanes_supported (vectype, group_size)) |
3673 | store_lanes_p = true; | |
3674 | else if (!vect_strided_store_supported (vectype, group_size)) | |
b602d918 RS |
3675 | return false; |
3676 | } | |
b8698a0f | 3677 | |
ebfd146a IR |
3678 | if (first_stmt == stmt) |
3679 | { | |
3680 | /* STMT is the leader of the group. Check the operands of all the | |
3681 | stmts of the group. */ | |
e14c1050 | 3682 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
3683 | while (next_stmt) |
3684 | { | |
3685 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
3686 | op = gimple_assign_rhs1 (next_stmt); | |
b8698a0f | 3687 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, |
a70d6342 | 3688 | &def, &dt)) |
ebfd146a IR |
3689 | { |
3690 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3691 | fprintf (vect_dump, "use not simple."); | |
3692 | return false; | |
3693 | } | |
e14c1050 | 3694 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
3695 | } |
3696 | } | |
3697 | } | |
3698 | ||
3699 | if (!vec_stmt) /* transformation not required. */ | |
3700 | { | |
3701 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
272c6793 | 3702 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, NULL); |
ebfd146a IR |
3703 | return true; |
3704 | } | |
3705 | ||
3706 | /** Transform. **/ | |
3707 | ||
3708 | if (strided_store) | |
3709 | { | |
3710 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 3711 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 3712 | |
e14c1050 | 3713 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
3714 | |
3715 | /* FORNOW */ | |
a70d6342 | 3716 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
3717 | |
3718 | /* We vectorize all the stmts of the interleaving group when we | |
3719 | reach the last stmt in the group. */ | |
e14c1050 IR |
3720 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
3721 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
3722 | && !slp) |
3723 | { | |
3724 | *vec_stmt = NULL; | |
3725 | return true; | |
3726 | } | |
3727 | ||
3728 | if (slp) | |
4b5caab7 IR |
3729 | { |
3730 | strided_store = false; | |
3731 | /* VEC_NUM is the number of vect stmts to be created for this | |
3732 | group. */ | |
3733 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
3734 | first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0); | |
3735 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
3736 | } | |
ebfd146a | 3737 | else |
4b5caab7 IR |
3738 | /* VEC_NUM is the number of vect stmts to be created for this |
3739 | group. */ | |
ebfd146a IR |
3740 | vec_num = group_size; |
3741 | } | |
b8698a0f | 3742 | else |
ebfd146a IR |
3743 | { |
3744 | first_stmt = stmt; | |
3745 | first_dr = dr; | |
3746 | group_size = vec_num = 1; | |
ebfd146a | 3747 | } |
b8698a0f | 3748 | |
ebfd146a IR |
3749 | if (vect_print_dump_info (REPORT_DETAILS)) |
3750 | fprintf (vect_dump, "transform store. ncopies = %d",ncopies); | |
3751 | ||
3752 | dr_chain = VEC_alloc (tree, heap, group_size); | |
3753 | oprnds = VEC_alloc (tree, heap, group_size); | |
3754 | ||
720f5239 | 3755 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 3756 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
3757 | /* Targets with store-lane instructions must not require explicit |
3758 | realignment. */ | |
3759 | gcc_assert (!store_lanes_p | |
3760 | || alignment_support_scheme == dr_aligned | |
3761 | || alignment_support_scheme == dr_unaligned_supported); | |
3762 | ||
3763 | if (store_lanes_p) | |
3764 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
3765 | else | |
3766 | aggr_type = vectype; | |
ebfd146a IR |
3767 | |
3768 | /* In case the vectorization factor (VF) is bigger than the number | |
3769 | of elements that we can fit in a vectype (nunits), we have to generate | |
3770 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 3771 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
3772 | vect_get_vec_def_for_copy_stmt. */ |
3773 | ||
3774 | /* In case of interleaving (non-unit strided access): | |
3775 | ||
3776 | S1: &base + 2 = x2 | |
3777 | S2: &base = x0 | |
3778 | S3: &base + 1 = x1 | |
3779 | S4: &base + 3 = x3 | |
3780 | ||
3781 | We create vectorized stores starting from base address (the access of the | |
3782 | first stmt in the chain (S2 in the above example), when the last store stmt | |
3783 | of the chain (S4) is reached: | |
3784 | ||
3785 | VS1: &base = vx2 | |
3786 | VS2: &base + vec_size*1 = vx0 | |
3787 | VS3: &base + vec_size*2 = vx1 | |
3788 | VS4: &base + vec_size*3 = vx3 | |
3789 | ||
3790 | Then permutation statements are generated: | |
3791 | ||
3792 | VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 > | |
3793 | VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 > | |
3794 | ... | |
b8698a0f | 3795 | |
ebfd146a IR |
3796 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
3797 | (the order of the data-refs in the output of vect_permute_store_chain | |
3798 | corresponds to the order of scalar stmts in the interleaving chain - see | |
3799 | the documentation of vect_permute_store_chain()). | |
3800 | ||
3801 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 3802 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 3803 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 3804 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
3805 | */ |
3806 | ||
3807 | prev_stmt_info = NULL; | |
3808 | for (j = 0; j < ncopies; j++) | |
3809 | { | |
3810 | gimple new_stmt; | |
3811 | gimple ptr_incr; | |
3812 | ||
3813 | if (j == 0) | |
3814 | { | |
3815 | if (slp) | |
3816 | { | |
3817 | /* Get vectorized arguments for SLP_NODE. */ | |
9dc3f7de IR |
3818 | vect_get_slp_defs (NULL_TREE, NULL_TREE, slp_node, &vec_oprnds, |
3819 | NULL, -1); | |
ebfd146a IR |
3820 | |
3821 | vec_oprnd = VEC_index (tree, vec_oprnds, 0); | |
3822 | } | |
3823 | else | |
3824 | { | |
b8698a0f L |
3825 | /* For interleaved stores we collect vectorized defs for all the |
3826 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
3827 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
3828 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
3829 | ||
3830 | If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and | |
3831 | OPRNDS are of size 1. */ | |
b8698a0f | 3832 | next_stmt = first_stmt; |
ebfd146a IR |
3833 | for (i = 0; i < group_size; i++) |
3834 | { | |
b8698a0f L |
3835 | /* Since gaps are not supported for interleaved stores, |
3836 | GROUP_SIZE is the exact number of stmts in the chain. | |
3837 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
3838 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
3839 | iteration of the loop will be executed. */ |
3840 | gcc_assert (next_stmt | |
3841 | && gimple_assign_single_p (next_stmt)); | |
3842 | op = gimple_assign_rhs1 (next_stmt); | |
3843 | ||
b8698a0f | 3844 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt, |
ebfd146a | 3845 | NULL); |
b8698a0f L |
3846 | VEC_quick_push(tree, dr_chain, vec_oprnd); |
3847 | VEC_quick_push(tree, oprnds, vec_oprnd); | |
e14c1050 | 3848 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
3849 | } |
3850 | } | |
3851 | ||
3852 | /* We should have catched mismatched types earlier. */ | |
3853 | gcc_assert (useless_type_conversion_p (vectype, | |
3854 | TREE_TYPE (vec_oprnd))); | |
272c6793 | 3855 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, aggr_type, NULL, |
920e8172 RS |
3856 | NULL_TREE, &dummy, gsi, |
3857 | &ptr_incr, false, &inv_p); | |
a70d6342 | 3858 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 3859 | } |
b8698a0f | 3860 | else |
ebfd146a | 3861 | { |
b8698a0f L |
3862 | /* For interleaved stores we created vectorized defs for all the |
3863 | defs stored in OPRNDS in the previous iteration (previous copy). | |
3864 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
3865 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
3866 | next copy. | |
3867 | If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and | |
3868 | OPRNDS are of size 1. */ | |
3869 | for (i = 0; i < group_size; i++) | |
3870 | { | |
3871 | op = VEC_index (tree, oprnds, i); | |
b8698a0f | 3872 | vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, |
a70d6342 | 3873 | &dt); |
b8698a0f | 3874 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
ebfd146a IR |
3875 | VEC_replace(tree, dr_chain, i, vec_oprnd); |
3876 | VEC_replace(tree, oprnds, i, vec_oprnd); | |
3877 | } | |
272c6793 RS |
3878 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
3879 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
3880 | } |
3881 | ||
272c6793 | 3882 | if (store_lanes_p) |
ebfd146a | 3883 | { |
272c6793 | 3884 | tree vec_array; |
267d3070 | 3885 | |
272c6793 RS |
3886 | /* Combine all the vectors into an array. */ |
3887 | vec_array = create_vector_array (vectype, vec_num); | |
3888 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 3889 | { |
272c6793 RS |
3890 | vec_oprnd = VEC_index (tree, dr_chain, i); |
3891 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); | |
267d3070 | 3892 | } |
b8698a0f | 3893 | |
272c6793 RS |
3894 | /* Emit: |
3895 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
3896 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
3897 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
3898 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 RS |
3899 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3900 | mark_symbols_for_renaming (new_stmt); | |
272c6793 RS |
3901 | } |
3902 | else | |
3903 | { | |
3904 | new_stmt = NULL; | |
3905 | if (strided_store) | |
3906 | { | |
3907 | result_chain = VEC_alloc (tree, heap, group_size); | |
3908 | /* Permute. */ | |
3909 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
3910 | &result_chain); | |
3911 | } | |
c2d7ab2a | 3912 | |
272c6793 RS |
3913 | next_stmt = first_stmt; |
3914 | for (i = 0; i < vec_num; i++) | |
3915 | { | |
3916 | struct ptr_info_def *pi; | |
3917 | ||
3918 | if (i > 0) | |
3919 | /* Bump the vector pointer. */ | |
3920 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
3921 | stmt, NULL_TREE); | |
3922 | ||
3923 | if (slp) | |
3924 | vec_oprnd = VEC_index (tree, vec_oprnds, i); | |
3925 | else if (strided_store) | |
3926 | /* For strided stores vectorized defs are interleaved in | |
3927 | vect_permute_store_chain(). */ | |
3928 | vec_oprnd = VEC_index (tree, result_chain, i); | |
3929 | ||
3930 | data_ref = build2 (MEM_REF, TREE_TYPE (vec_oprnd), dataref_ptr, | |
3931 | build_int_cst (reference_alias_ptr_type | |
3932 | (DR_REF (first_dr)), 0)); | |
3933 | pi = get_ptr_info (dataref_ptr); | |
3934 | pi->align = TYPE_ALIGN_UNIT (vectype); | |
3935 | if (aligned_access_p (first_dr)) | |
3936 | pi->misalign = 0; | |
3937 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
3938 | { | |
3939 | TREE_TYPE (data_ref) | |
3940 | = build_aligned_type (TREE_TYPE (data_ref), | |
3941 | TYPE_ALIGN (elem_type)); | |
3942 | pi->align = TYPE_ALIGN_UNIT (elem_type); | |
3943 | pi->misalign = 0; | |
3944 | } | |
3945 | else | |
3946 | { | |
3947 | TREE_TYPE (data_ref) | |
3948 | = build_aligned_type (TREE_TYPE (data_ref), | |
3949 | TYPE_ALIGN (elem_type)); | |
3950 | pi->misalign = DR_MISALIGNMENT (first_dr); | |
3951 | } | |
c2d7ab2a | 3952 | |
272c6793 RS |
3953 | /* Arguments are ready. Create the new vector stmt. */ |
3954 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
3955 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3956 | mark_symbols_for_renaming (new_stmt); | |
3957 | ||
3958 | if (slp) | |
3959 | continue; | |
3960 | ||
e14c1050 | 3961 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
3962 | if (!next_stmt) |
3963 | break; | |
3964 | } | |
ebfd146a | 3965 | } |
1da0876c RS |
3966 | if (!slp) |
3967 | { | |
3968 | if (j == 0) | |
3969 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3970 | else | |
3971 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3972 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3973 | } | |
ebfd146a IR |
3974 | } |
3975 | ||
b8698a0f L |
3976 | VEC_free (tree, heap, dr_chain); |
3977 | VEC_free (tree, heap, oprnds); | |
ebfd146a | 3978 | if (result_chain) |
b8698a0f | 3979 | VEC_free (tree, heap, result_chain); |
ff802fa1 IR |
3980 | if (vec_oprnds) |
3981 | VEC_free (tree, heap, vec_oprnds); | |
ebfd146a IR |
3982 | |
3983 | return true; | |
3984 | } | |
3985 | ||
a1e53f3f L |
3986 | /* Given a vector type VECTYPE returns a builtin DECL to be used |
3987 | for vector permutation and stores a mask into *MASK that implements | |
3988 | reversal of the vector elements. If that is impossible to do | |
3989 | returns NULL (and *MASK is unchanged). */ | |
3990 | ||
3991 | static tree | |
3992 | perm_mask_for_reverse (tree vectype, tree *mask) | |
3993 | { | |
3994 | tree builtin_decl; | |
3995 | tree mask_element_type, mask_type; | |
3996 | tree mask_vec = NULL; | |
3997 | int i; | |
3998 | int nunits; | |
3999 | if (!targetm.vectorize.builtin_vec_perm) | |
4000 | return NULL; | |
4001 | ||
4002 | builtin_decl = targetm.vectorize.builtin_vec_perm (vectype, | |
4003 | &mask_element_type); | |
4004 | if (!builtin_decl || !mask_element_type) | |
4005 | return NULL; | |
4006 | ||
4007 | mask_type = get_vectype_for_scalar_type (mask_element_type); | |
4008 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
bb67d9c7 RG |
4009 | if (!mask_type |
4010 | || TYPE_VECTOR_SUBPARTS (vectype) != TYPE_VECTOR_SUBPARTS (mask_type)) | |
a1e53f3f L |
4011 | return NULL; |
4012 | ||
4013 | for (i = 0; i < nunits; i++) | |
4014 | mask_vec = tree_cons (NULL, build_int_cst (mask_element_type, i), mask_vec); | |
4015 | mask_vec = build_vector (mask_type, mask_vec); | |
4016 | ||
4017 | if (!targetm.vectorize.builtin_vec_perm_ok (vectype, mask_vec)) | |
4018 | return NULL; | |
4019 | if (mask) | |
4020 | *mask = mask_vec; | |
4021 | return builtin_decl; | |
4022 | } | |
4023 | ||
4024 | /* Given a vector variable X, that was generated for the scalar LHS of | |
4025 | STMT, generate instructions to reverse the vector elements of X, | |
4026 | insert them a *GSI and return the permuted vector variable. */ | |
4027 | ||
4028 | static tree | |
4029 | reverse_vec_elements (tree x, gimple stmt, gimple_stmt_iterator *gsi) | |
4030 | { | |
4031 | tree vectype = TREE_TYPE (x); | |
4032 | tree mask_vec, builtin_decl; | |
4033 | tree perm_dest, data_ref; | |
4034 | gimple perm_stmt; | |
4035 | ||
4036 | builtin_decl = perm_mask_for_reverse (vectype, &mask_vec); | |
4037 | ||
4038 | perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype); | |
4039 | ||
4040 | /* Generate the permute statement. */ | |
4041 | perm_stmt = gimple_build_call (builtin_decl, 3, x, x, mask_vec); | |
2a2651b7 RG |
4042 | if (!useless_type_conversion_p (vectype, |
4043 | TREE_TYPE (TREE_TYPE (builtin_decl)))) | |
4044 | { | |
4045 | tree tem = create_tmp_reg (TREE_TYPE (TREE_TYPE (builtin_decl)), NULL); | |
4046 | tem = make_ssa_name (tem, perm_stmt); | |
4047 | gimple_call_set_lhs (perm_stmt, tem); | |
4048 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); | |
4049 | perm_stmt = gimple_build_assign (NULL_TREE, | |
4050 | build1 (VIEW_CONVERT_EXPR, | |
4051 | vectype, tem)); | |
4052 | } | |
a1e53f3f | 4053 | data_ref = make_ssa_name (perm_dest, perm_stmt); |
2a2651b7 | 4054 | gimple_set_lhs (perm_stmt, data_ref); |
a1e53f3f L |
4055 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
4056 | ||
4057 | return data_ref; | |
4058 | } | |
4059 | ||
ebfd146a IR |
4060 | /* vectorizable_load. |
4061 | ||
b8698a0f L |
4062 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
4063 | can be vectorized. | |
4064 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4065 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4066 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4067 | ||
4068 | static bool | |
4069 | vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
4070 | slp_tree slp_node, slp_instance slp_node_instance) | |
4071 | { | |
4072 | tree scalar_dest; | |
4073 | tree vec_dest = NULL; | |
4074 | tree data_ref = NULL; | |
4075 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 4076 | stmt_vec_info prev_stmt_info; |
ebfd146a | 4077 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 4078 | struct loop *loop = NULL; |
ebfd146a | 4079 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 4080 | bool nested_in_vect_loop = false; |
ebfd146a IR |
4081 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; |
4082 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
272c6793 | 4083 | tree elem_type; |
ebfd146a | 4084 | tree new_temp; |
947131ba | 4085 | enum machine_mode mode; |
ebfd146a IR |
4086 | gimple new_stmt = NULL; |
4087 | tree dummy; | |
4088 | enum dr_alignment_support alignment_support_scheme; | |
4089 | tree dataref_ptr = NULL_TREE; | |
4090 | gimple ptr_incr; | |
4091 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4092 | int ncopies; | |
4093 | int i, j, group_size; | |
4094 | tree msq = NULL_TREE, lsq; | |
4095 | tree offset = NULL_TREE; | |
4096 | tree realignment_token = NULL_TREE; | |
4097 | gimple phi = NULL; | |
4098 | VEC(tree,heap) *dr_chain = NULL; | |
4099 | bool strided_load = false; | |
272c6793 | 4100 | bool load_lanes_p = false; |
ebfd146a IR |
4101 | gimple first_stmt; |
4102 | tree scalar_type; | |
4103 | bool inv_p; | |
a1e53f3f | 4104 | bool negative; |
ebfd146a IR |
4105 | bool compute_in_loop = false; |
4106 | struct loop *at_loop; | |
4107 | int vec_num; | |
4108 | bool slp = (slp_node != NULL); | |
4109 | bool slp_perm = false; | |
4110 | enum tree_code code; | |
a70d6342 IR |
4111 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
4112 | int vf; | |
272c6793 | 4113 | tree aggr_type; |
a70d6342 IR |
4114 | |
4115 | if (loop_vinfo) | |
4116 | { | |
4117 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
4118 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
4119 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4120 | } | |
4121 | else | |
3533e503 | 4122 | vf = 1; |
ebfd146a IR |
4123 | |
4124 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4125 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 4126 | case of SLP. */ |
437f4a00 | 4127 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
4128 | ncopies = 1; |
4129 | else | |
4130 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4131 | ||
4132 | gcc_assert (ncopies >= 1); | |
4133 | ||
4134 | /* FORNOW. This restriction should be relaxed. */ | |
4135 | if (nested_in_vect_loop && ncopies > 1) | |
4136 | { | |
4137 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4138 | fprintf (vect_dump, "multiple types in nested loop."); | |
4139 | return false; | |
4140 | } | |
4141 | ||
a70d6342 | 4142 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4143 | return false; |
4144 | ||
8644a673 | 4145 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4146 | return false; |
4147 | ||
4148 | /* Is vectorizable load? */ | |
4149 | if (!is_gimple_assign (stmt)) | |
4150 | return false; | |
4151 | ||
4152 | scalar_dest = gimple_assign_lhs (stmt); | |
4153 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4154 | return false; | |
4155 | ||
4156 | code = gimple_assign_rhs_code (stmt); | |
4157 | if (code != ARRAY_REF | |
4158 | && code != INDIRECT_REF | |
e9dbe7bb IR |
4159 | && code != COMPONENT_REF |
4160 | && code != IMAGPART_EXPR | |
70f34814 | 4161 | && code != REALPART_EXPR |
42373e0b RG |
4162 | && code != MEM_REF |
4163 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
ebfd146a IR |
4164 | return false; |
4165 | ||
4166 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
4167 | return false; | |
4168 | ||
a1e53f3f L |
4169 | negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0; |
4170 | if (negative && ncopies > 1) | |
4171 | { | |
4172 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4173 | fprintf (vect_dump, "multiple types with negative step."); | |
4174 | return false; | |
4175 | } | |
4176 | ||
ebfd146a | 4177 | scalar_type = TREE_TYPE (DR_REF (dr)); |
947131ba | 4178 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
4179 | |
4180 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
4181 | (e.g. - data copies). */ | |
947131ba | 4182 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a IR |
4183 | { |
4184 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4185 | fprintf (vect_dump, "Aligned load, but unsupported type."); | |
4186 | return false; | |
4187 | } | |
4188 | ||
4189 | /* The vector component type needs to be trivially convertible to the | |
4190 | scalar lhs. This should always be the case. */ | |
272c6793 RS |
4191 | elem_type = TREE_TYPE (vectype); |
4192 | if (!useless_type_conversion_p (TREE_TYPE (scalar_dest), elem_type)) | |
b8698a0f | 4193 | { |
ebfd146a IR |
4194 | if (vect_print_dump_info (REPORT_DETAILS)) |
4195 | fprintf (vect_dump, "??? operands of different types"); | |
4196 | return false; | |
4197 | } | |
4198 | ||
4199 | /* Check if the load is a part of an interleaving chain. */ | |
4200 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) | |
4201 | { | |
4202 | strided_load = true; | |
4203 | /* FORNOW */ | |
4204 | gcc_assert (! nested_in_vect_loop); | |
4205 | ||
e14c1050 | 4206 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
4207 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
4208 | { | |
e14c1050 | 4209 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
4210 | if (vect_load_lanes_supported (vectype, group_size)) |
4211 | load_lanes_p = true; | |
4212 | else if (!vect_strided_load_supported (vectype, group_size)) | |
b602d918 RS |
4213 | return false; |
4214 | } | |
ebfd146a IR |
4215 | } |
4216 | ||
a1e53f3f L |
4217 | if (negative) |
4218 | { | |
4219 | gcc_assert (!strided_load); | |
4220 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
4221 | if (alignment_support_scheme != dr_aligned | |
4222 | && alignment_support_scheme != dr_unaligned_supported) | |
4223 | { | |
4224 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4225 | fprintf (vect_dump, "negative step but alignment required."); | |
4226 | return false; | |
4227 | } | |
4228 | if (!perm_mask_for_reverse (vectype, NULL)) | |
4229 | { | |
4230 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4231 | fprintf (vect_dump, "negative step and reversing not supported."); | |
4232 | return false; | |
4233 | } | |
4234 | } | |
4235 | ||
ebfd146a IR |
4236 | if (!vec_stmt) /* transformation not required. */ |
4237 | { | |
4238 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
272c6793 | 4239 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, NULL); |
ebfd146a IR |
4240 | return true; |
4241 | } | |
4242 | ||
4243 | if (vect_print_dump_info (REPORT_DETAILS)) | |
0ea25ecd | 4244 | fprintf (vect_dump, "transform load. ncopies = %d", ncopies); |
ebfd146a IR |
4245 | |
4246 | /** Transform. **/ | |
4247 | ||
4248 | if (strided_load) | |
4249 | { | |
e14c1050 | 4250 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
4251 | /* Check if the chain of loads is already vectorized. */ |
4252 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt))) | |
4253 | { | |
4254 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4255 | return true; | |
4256 | } | |
4257 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 4258 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a IR |
4259 | |
4260 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
4261 | if (slp) | |
4262 | { | |
4263 | strided_load = false; | |
4264 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
a70d6342 IR |
4265 | if (SLP_INSTANCE_LOAD_PERMUTATION (slp_node_instance)) |
4266 | slp_perm = true; | |
4267 | } | |
ebfd146a IR |
4268 | else |
4269 | vec_num = group_size; | |
ebfd146a IR |
4270 | } |
4271 | else | |
4272 | { | |
4273 | first_stmt = stmt; | |
4274 | first_dr = dr; | |
4275 | group_size = vec_num = 1; | |
4276 | } | |
4277 | ||
720f5239 | 4278 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 4279 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
4280 | /* Targets with load-lane instructions must not require explicit |
4281 | realignment. */ | |
4282 | gcc_assert (!load_lanes_p | |
4283 | || alignment_support_scheme == dr_aligned | |
4284 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
4285 | |
4286 | /* In case the vectorization factor (VF) is bigger than the number | |
4287 | of elements that we can fit in a vectype (nunits), we have to generate | |
4288 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 4289 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 4290 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 4291 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 4292 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
4293 | stmts that use the defs of the current stmt. The example below |
4294 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
4295 | need to create 4 vectorized stmts): | |
ebfd146a IR |
4296 | |
4297 | before vectorization: | |
4298 | RELATED_STMT VEC_STMT | |
4299 | S1: x = memref - - | |
4300 | S2: z = x + 1 - - | |
4301 | ||
4302 | step 1: vectorize stmt S1: | |
4303 | We first create the vector stmt VS1_0, and, as usual, record a | |
4304 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
4305 | Next, we create the vector stmt VS1_1, and record a pointer to | |
4306 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 4307 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
4308 | stmts and pointers: |
4309 | RELATED_STMT VEC_STMT | |
4310 | VS1_0: vx0 = memref0 VS1_1 - | |
4311 | VS1_1: vx1 = memref1 VS1_2 - | |
4312 | VS1_2: vx2 = memref2 VS1_3 - | |
4313 | VS1_3: vx3 = memref3 - - | |
4314 | S1: x = load - VS1_0 | |
4315 | S2: z = x + 1 - - | |
4316 | ||
b8698a0f L |
4317 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
4318 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
4319 | stmt S2. */ |
4320 | ||
4321 | /* In case of interleaving (non-unit strided access): | |
4322 | ||
4323 | S1: x2 = &base + 2 | |
4324 | S2: x0 = &base | |
4325 | S3: x1 = &base + 1 | |
4326 | S4: x3 = &base + 3 | |
4327 | ||
b8698a0f | 4328 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
4329 | starting from the access of the first stmt of the chain: |
4330 | ||
4331 | VS1: vx0 = &base | |
4332 | VS2: vx1 = &base + vec_size*1 | |
4333 | VS3: vx3 = &base + vec_size*2 | |
4334 | VS4: vx4 = &base + vec_size*3 | |
4335 | ||
4336 | Then permutation statements are generated: | |
4337 | ||
4338 | VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 > | |
4339 | VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 > | |
4340 | ... | |
4341 | ||
4342 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
4343 | (the order of the data-refs in the output of vect_permute_load_chain | |
4344 | corresponds to the order of scalar stmts in the interleaving chain - see | |
4345 | the documentation of vect_permute_load_chain()). | |
4346 | The generation of permutation stmts and recording them in | |
4347 | STMT_VINFO_VEC_STMT is done in vect_transform_strided_load(). | |
4348 | ||
b8698a0f | 4349 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
4350 | permutation stmts above are created for every copy. The result vector |
4351 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
4352 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
4353 | |
4354 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
4355 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
4356 | we generate the following code: | |
4357 | p = initial_addr; | |
4358 | indx = 0; | |
4359 | loop { | |
4360 | p = p + indx * vectype_size; | |
4361 | vec_dest = *(p); | |
4362 | indx = indx + 1; | |
4363 | } | |
4364 | ||
4365 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 4366 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
4367 | then generate the following code, in which the data in each iteration is |
4368 | obtained by two vector loads, one from the previous iteration, and one | |
4369 | from the current iteration: | |
4370 | p1 = initial_addr; | |
4371 | msq_init = *(floor(p1)) | |
4372 | p2 = initial_addr + VS - 1; | |
4373 | realignment_token = call target_builtin; | |
4374 | indx = 0; | |
4375 | loop { | |
4376 | p2 = p2 + indx * vectype_size | |
4377 | lsq = *(floor(p2)) | |
4378 | vec_dest = realign_load (msq, lsq, realignment_token) | |
4379 | indx = indx + 1; | |
4380 | msq = lsq; | |
4381 | } */ | |
4382 | ||
4383 | /* If the misalignment remains the same throughout the execution of the | |
4384 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 4385 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
4386 | This can only occur when vectorizing memory accesses in the inner-loop |
4387 | nested within an outer-loop that is being vectorized. */ | |
4388 | ||
a70d6342 | 4389 | if (loop && nested_in_vect_loop_p (loop, stmt) |
ebfd146a IR |
4390 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
4391 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) | |
4392 | { | |
4393 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
4394 | compute_in_loop = true; | |
4395 | } | |
4396 | ||
4397 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
4398 | || alignment_support_scheme == dr_explicit_realign) | |
4399 | && !compute_in_loop) | |
4400 | { | |
4401 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
4402 | alignment_support_scheme, NULL_TREE, | |
4403 | &at_loop); | |
4404 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
4405 | { | |
4406 | phi = SSA_NAME_DEF_STMT (msq); | |
4407 | offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
4408 | } | |
4409 | } | |
4410 | else | |
4411 | at_loop = loop; | |
4412 | ||
a1e53f3f L |
4413 | if (negative) |
4414 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
4415 | ||
272c6793 RS |
4416 | if (load_lanes_p) |
4417 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
4418 | else | |
4419 | aggr_type = vectype; | |
4420 | ||
ebfd146a IR |
4421 | prev_stmt_info = NULL; |
4422 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 4423 | { |
272c6793 | 4424 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 4425 | if (j == 0) |
272c6793 | 4426 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, |
920e8172 RS |
4427 | offset, &dummy, gsi, |
4428 | &ptr_incr, false, &inv_p); | |
ebfd146a | 4429 | else |
272c6793 RS |
4430 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
4431 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 4432 | |
5ce1ee7f RS |
4433 | if (strided_load || slp_perm) |
4434 | dr_chain = VEC_alloc (tree, heap, vec_num); | |
4435 | ||
272c6793 | 4436 | if (load_lanes_p) |
ebfd146a | 4437 | { |
272c6793 RS |
4438 | tree vec_array; |
4439 | ||
4440 | vec_array = create_vector_array (vectype, vec_num); | |
4441 | ||
4442 | /* Emit: | |
4443 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
4444 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
4445 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
4446 | gimple_call_set_lhs (new_stmt, vec_array); | |
4447 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4448 | mark_symbols_for_renaming (new_stmt); | |
ebfd146a | 4449 | |
272c6793 RS |
4450 | /* Extract each vector into an SSA_NAME. */ |
4451 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 4452 | { |
272c6793 RS |
4453 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
4454 | vec_array, i); | |
4455 | VEC_quick_push (tree, dr_chain, new_temp); | |
4456 | } | |
4457 | ||
4458 | /* Record the mapping between SSA_NAMEs and statements. */ | |
4459 | vect_record_strided_load_vectors (stmt, dr_chain); | |
4460 | } | |
4461 | else | |
4462 | { | |
4463 | for (i = 0; i < vec_num; i++) | |
4464 | { | |
4465 | if (i > 0) | |
4466 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
4467 | stmt, NULL_TREE); | |
4468 | ||
4469 | /* 2. Create the vector-load in the loop. */ | |
4470 | switch (alignment_support_scheme) | |
4471 | { | |
4472 | case dr_aligned: | |
4473 | case dr_unaligned_supported: | |
be1ac4ec | 4474 | { |
272c6793 RS |
4475 | struct ptr_info_def *pi; |
4476 | data_ref | |
4477 | = build2 (MEM_REF, vectype, dataref_ptr, | |
4478 | build_int_cst (reference_alias_ptr_type | |
4479 | (DR_REF (first_dr)), 0)); | |
4480 | pi = get_ptr_info (dataref_ptr); | |
4481 | pi->align = TYPE_ALIGN_UNIT (vectype); | |
4482 | if (alignment_support_scheme == dr_aligned) | |
4483 | { | |
4484 | gcc_assert (aligned_access_p (first_dr)); | |
4485 | pi->misalign = 0; | |
4486 | } | |
4487 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
4488 | { | |
4489 | TREE_TYPE (data_ref) | |
4490 | = build_aligned_type (TREE_TYPE (data_ref), | |
4491 | TYPE_ALIGN (elem_type)); | |
4492 | pi->align = TYPE_ALIGN_UNIT (elem_type); | |
4493 | pi->misalign = 0; | |
4494 | } | |
4495 | else | |
4496 | { | |
4497 | TREE_TYPE (data_ref) | |
4498 | = build_aligned_type (TREE_TYPE (data_ref), | |
4499 | TYPE_ALIGN (elem_type)); | |
4500 | pi->misalign = DR_MISALIGNMENT (first_dr); | |
4501 | } | |
4502 | break; | |
be1ac4ec | 4503 | } |
272c6793 | 4504 | case dr_explicit_realign: |
267d3070 | 4505 | { |
272c6793 RS |
4506 | tree ptr, bump; |
4507 | tree vs_minus_1; | |
4508 | ||
4509 | vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
4510 | ||
4511 | if (compute_in_loop) | |
4512 | msq = vect_setup_realignment (first_stmt, gsi, | |
4513 | &realignment_token, | |
4514 | dr_explicit_realign, | |
4515 | dataref_ptr, NULL); | |
4516 | ||
4517 | new_stmt = gimple_build_assign_with_ops | |
4518 | (BIT_AND_EXPR, NULL_TREE, dataref_ptr, | |
4519 | build_int_cst | |
4520 | (TREE_TYPE (dataref_ptr), | |
4521 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
4522 | ptr = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt); | |
4523 | gimple_assign_set_lhs (new_stmt, ptr); | |
4524 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4525 | data_ref | |
4526 | = build2 (MEM_REF, vectype, ptr, | |
4527 | build_int_cst (reference_alias_ptr_type | |
4528 | (DR_REF (first_dr)), 0)); | |
4529 | vec_dest = vect_create_destination_var (scalar_dest, | |
4530 | vectype); | |
4531 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
4532 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4533 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4534 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
4535 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
4536 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4537 | msq = new_temp; | |
4538 | ||
4539 | bump = size_binop (MULT_EXPR, vs_minus_1, | |
4540 | TYPE_SIZE_UNIT (scalar_type)); | |
4541 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); | |
4542 | new_stmt = gimple_build_assign_with_ops | |
4543 | (BIT_AND_EXPR, NULL_TREE, ptr, | |
4544 | build_int_cst | |
4545 | (TREE_TYPE (ptr), | |
4546 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
4547 | ptr = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt); | |
4548 | gimple_assign_set_lhs (new_stmt, ptr); | |
4549 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4550 | data_ref | |
4551 | = build2 (MEM_REF, vectype, ptr, | |
4552 | build_int_cst (reference_alias_ptr_type | |
4553 | (DR_REF (first_dr)), 0)); | |
4554 | break; | |
267d3070 | 4555 | } |
272c6793 RS |
4556 | case dr_explicit_realign_optimized: |
4557 | new_stmt = gimple_build_assign_with_ops | |
4558 | (BIT_AND_EXPR, NULL_TREE, dataref_ptr, | |
4559 | build_int_cst | |
4560 | (TREE_TYPE (dataref_ptr), | |
4561 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
4562 | new_temp = make_ssa_name (SSA_NAME_VAR (dataref_ptr), | |
4563 | new_stmt); | |
4564 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4565 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4566 | data_ref | |
4567 | = build2 (MEM_REF, vectype, new_temp, | |
4568 | build_int_cst (reference_alias_ptr_type | |
4569 | (DR_REF (first_dr)), 0)); | |
4570 | break; | |
4571 | default: | |
4572 | gcc_unreachable (); | |
4573 | } | |
ebfd146a | 4574 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 4575 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
4576 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4577 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4578 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 | 4579 | mark_symbols_for_renaming (new_stmt); |
ebfd146a | 4580 | |
272c6793 RS |
4581 | /* 3. Handle explicit realignment if necessary/supported. |
4582 | Create in loop: | |
4583 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
4584 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
4585 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 4586 | { |
272c6793 RS |
4587 | lsq = gimple_assign_lhs (new_stmt); |
4588 | if (!realignment_token) | |
4589 | realignment_token = dataref_ptr; | |
4590 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4591 | new_stmt | |
4592 | = gimple_build_assign_with_ops3 (REALIGN_LOAD_EXPR, | |
4593 | vec_dest, msq, lsq, | |
4594 | realignment_token); | |
4595 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4596 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4597 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4598 | ||
4599 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
4600 | { | |
4601 | gcc_assert (phi); | |
4602 | if (i == vec_num - 1 && j == ncopies - 1) | |
4603 | add_phi_arg (phi, lsq, | |
4604 | loop_latch_edge (containing_loop), | |
4605 | UNKNOWN_LOCATION); | |
4606 | msq = lsq; | |
4607 | } | |
ebfd146a | 4608 | } |
ebfd146a | 4609 | |
272c6793 RS |
4610 | /* 4. Handle invariant-load. */ |
4611 | if (inv_p && !bb_vinfo) | |
ebfd146a | 4612 | { |
ab70d825 RG |
4613 | tree vec_inv; |
4614 | gimple_stmt_iterator gsi2 = *gsi; | |
272c6793 | 4615 | gcc_assert (!strided_load); |
ab70d825 RG |
4616 | gsi_next (&gsi2); |
4617 | vec_inv = build_vector_from_val (vectype, scalar_dest); | |
4618 | new_temp = vect_init_vector (stmt, vec_inv, | |
4619 | vectype, &gsi2); | |
4620 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
272c6793 | 4621 | } |
ebfd146a | 4622 | |
272c6793 RS |
4623 | if (negative) |
4624 | { | |
4625 | new_temp = reverse_vec_elements (new_temp, stmt, gsi); | |
ebfd146a IR |
4626 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
4627 | } | |
267d3070 | 4628 | |
272c6793 RS |
4629 | /* Collect vector loads and later create their permutation in |
4630 | vect_transform_strided_load (). */ | |
4631 | if (strided_load || slp_perm) | |
4632 | VEC_quick_push (tree, dr_chain, new_temp); | |
267d3070 | 4633 | |
272c6793 RS |
4634 | /* Store vector loads in the corresponding SLP_NODE. */ |
4635 | if (slp && !slp_perm) | |
4636 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), | |
4637 | new_stmt); | |
4638 | } | |
ebfd146a IR |
4639 | } |
4640 | ||
4641 | if (slp && !slp_perm) | |
4642 | continue; | |
4643 | ||
4644 | if (slp_perm) | |
4645 | { | |
a70d6342 | 4646 | if (!vect_transform_slp_perm_load (stmt, dr_chain, gsi, vf, |
ebfd146a IR |
4647 | slp_node_instance, false)) |
4648 | { | |
4649 | VEC_free (tree, heap, dr_chain); | |
4650 | return false; | |
4651 | } | |
4652 | } | |
4653 | else | |
4654 | { | |
4655 | if (strided_load) | |
4656 | { | |
272c6793 RS |
4657 | if (!load_lanes_p) |
4658 | vect_transform_strided_load (stmt, dr_chain, group_size, gsi); | |
ebfd146a | 4659 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
4660 | } |
4661 | else | |
4662 | { | |
4663 | if (j == 0) | |
4664 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4665 | else | |
4666 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4667 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4668 | } | |
4669 | } | |
5ce1ee7f RS |
4670 | if (dr_chain) |
4671 | VEC_free (tree, heap, dr_chain); | |
ebfd146a IR |
4672 | } |
4673 | ||
ebfd146a IR |
4674 | return true; |
4675 | } | |
4676 | ||
4677 | /* Function vect_is_simple_cond. | |
b8698a0f | 4678 | |
ebfd146a IR |
4679 | Input: |
4680 | LOOP - the loop that is being vectorized. | |
4681 | COND - Condition that is checked for simple use. | |
4682 | ||
4683 | Returns whether a COND can be vectorized. Checks whether | |
4684 | condition operands are supportable using vec_is_simple_use. */ | |
4685 | ||
4686 | static bool | |
4687 | vect_is_simple_cond (tree cond, loop_vec_info loop_vinfo) | |
4688 | { | |
4689 | tree lhs, rhs; | |
4690 | tree def; | |
4691 | enum vect_def_type dt; | |
4692 | ||
4693 | if (!COMPARISON_CLASS_P (cond)) | |
4694 | return false; | |
4695 | ||
4696 | lhs = TREE_OPERAND (cond, 0); | |
4697 | rhs = TREE_OPERAND (cond, 1); | |
4698 | ||
4699 | if (TREE_CODE (lhs) == SSA_NAME) | |
4700 | { | |
4701 | gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); | |
b8698a0f | 4702 | if (!vect_is_simple_use (lhs, loop_vinfo, NULL, &lhs_def_stmt, &def, |
a70d6342 | 4703 | &dt)) |
ebfd146a IR |
4704 | return false; |
4705 | } | |
4706 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
4707 | && TREE_CODE (lhs) != FIXED_CST) | |
4708 | return false; | |
4709 | ||
4710 | if (TREE_CODE (rhs) == SSA_NAME) | |
4711 | { | |
4712 | gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); | |
b8698a0f | 4713 | if (!vect_is_simple_use (rhs, loop_vinfo, NULL, &rhs_def_stmt, &def, |
a70d6342 | 4714 | &dt)) |
ebfd146a IR |
4715 | return false; |
4716 | } | |
4717 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST | |
4718 | && TREE_CODE (rhs) != FIXED_CST) | |
4719 | return false; | |
4720 | ||
4721 | return true; | |
4722 | } | |
4723 | ||
4724 | /* vectorizable_condition. | |
4725 | ||
b8698a0f L |
4726 | Check if STMT is conditional modify expression that can be vectorized. |
4727 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4728 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
4729 | at GSI. |
4730 | ||
4731 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
4732 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
4733 | else caluse if it is 2). | |
ebfd146a IR |
4734 | |
4735 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4736 | ||
4bbe8262 | 4737 | bool |
ebfd146a | 4738 | vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi, |
4bbe8262 | 4739 | gimple *vec_stmt, tree reduc_def, int reduc_index) |
ebfd146a IR |
4740 | { |
4741 | tree scalar_dest = NULL_TREE; | |
4742 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
4743 | tree cond_expr, then_clause, else_clause; |
4744 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4745 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
ff802fa1 IR |
4746 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
4747 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
ebfd146a IR |
4748 | tree vec_compare, vec_cond_expr; |
4749 | tree new_temp; | |
4750 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4751 | enum machine_mode vec_mode; | |
4752 | tree def; | |
a855b1b1 | 4753 | enum vect_def_type dt, dts[4]; |
ebfd146a IR |
4754 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
4755 | int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4756 | enum tree_code code; | |
a855b1b1 MM |
4757 | stmt_vec_info prev_stmt_info = NULL; |
4758 | int j; | |
ebfd146a | 4759 | |
a70d6342 IR |
4760 | /* FORNOW: unsupported in basic block SLP. */ |
4761 | gcc_assert (loop_vinfo); | |
b8698a0f | 4762 | |
437f4a00 IR |
4763 | /* FORNOW: SLP not supported. */ |
4764 | if (STMT_SLP_TYPE (stmt_info)) | |
4765 | return false; | |
4766 | ||
ebfd146a | 4767 | gcc_assert (ncopies >= 1); |
a855b1b1 | 4768 | if (reduc_index && ncopies > 1) |
ebfd146a IR |
4769 | return false; /* FORNOW */ |
4770 | ||
4771 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
4772 | return false; | |
4773 | ||
4bbe8262 IR |
4774 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4775 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
4776 | && reduc_def)) | |
ebfd146a IR |
4777 | return false; |
4778 | ||
ebfd146a | 4779 | /* FORNOW: not yet supported. */ |
b8698a0f | 4780 | if (STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a IR |
4781 | { |
4782 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4783 | fprintf (vect_dump, "value used after loop."); | |
4784 | return false; | |
4785 | } | |
4786 | ||
4787 | /* Is vectorizable conditional operation? */ | |
4788 | if (!is_gimple_assign (stmt)) | |
4789 | return false; | |
4790 | ||
4791 | code = gimple_assign_rhs_code (stmt); | |
4792 | ||
4793 | if (code != COND_EXPR) | |
4794 | return false; | |
4795 | ||
4e71066d RG |
4796 | cond_expr = gimple_assign_rhs1 (stmt); |
4797 | then_clause = gimple_assign_rhs2 (stmt); | |
4798 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a IR |
4799 | |
4800 | if (!vect_is_simple_cond (cond_expr, loop_vinfo)) | |
4801 | return false; | |
4802 | ||
4803 | /* We do not handle two different vector types for the condition | |
4804 | and the values. */ | |
8533c9d8 SP |
4805 | if (!types_compatible_p (TREE_TYPE (TREE_OPERAND (cond_expr, 0)), |
4806 | TREE_TYPE (vectype))) | |
ebfd146a IR |
4807 | return false; |
4808 | ||
4809 | if (TREE_CODE (then_clause) == SSA_NAME) | |
4810 | { | |
4811 | gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause); | |
b8698a0f | 4812 | if (!vect_is_simple_use (then_clause, loop_vinfo, NULL, |
ebfd146a IR |
4813 | &then_def_stmt, &def, &dt)) |
4814 | return false; | |
4815 | } | |
b8698a0f | 4816 | else if (TREE_CODE (then_clause) != INTEGER_CST |
ebfd146a IR |
4817 | && TREE_CODE (then_clause) != REAL_CST |
4818 | && TREE_CODE (then_clause) != FIXED_CST) | |
4819 | return false; | |
4820 | ||
4821 | if (TREE_CODE (else_clause) == SSA_NAME) | |
4822 | { | |
4823 | gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause); | |
a70d6342 | 4824 | if (!vect_is_simple_use (else_clause, loop_vinfo, NULL, |
ebfd146a IR |
4825 | &else_def_stmt, &def, &dt)) |
4826 | return false; | |
4827 | } | |
b8698a0f | 4828 | else if (TREE_CODE (else_clause) != INTEGER_CST |
ebfd146a IR |
4829 | && TREE_CODE (else_clause) != REAL_CST |
4830 | && TREE_CODE (else_clause) != FIXED_CST) | |
4831 | return false; | |
4832 | ||
4833 | ||
4834 | vec_mode = TYPE_MODE (vectype); | |
4835 | ||
b8698a0f | 4836 | if (!vec_stmt) |
ebfd146a IR |
4837 | { |
4838 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
4e71066d RG |
4839 | return expand_vec_cond_expr_p (TREE_TYPE (gimple_assign_lhs (stmt)), |
4840 | vec_mode); | |
ebfd146a IR |
4841 | } |
4842 | ||
4843 | /* Transform */ | |
4844 | ||
4845 | /* Handle def. */ | |
4846 | scalar_dest = gimple_assign_lhs (stmt); | |
4847 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4848 | ||
4849 | /* Handle cond expr. */ | |
a855b1b1 MM |
4850 | for (j = 0; j < ncopies; j++) |
4851 | { | |
4852 | gimple new_stmt; | |
4853 | if (j == 0) | |
4854 | { | |
4855 | gimple gtemp; | |
4856 | vec_cond_lhs = | |
4857 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), | |
4858 | stmt, NULL); | |
4859 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), loop_vinfo, | |
4860 | NULL, >emp, &def, &dts[0]); | |
4861 | vec_cond_rhs = | |
4862 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
4863 | stmt, NULL); | |
4864 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), loop_vinfo, | |
4865 | NULL, >emp, &def, &dts[1]); | |
4866 | if (reduc_index == 1) | |
4867 | vec_then_clause = reduc_def; | |
4868 | else | |
4869 | { | |
4870 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
4871 | stmt, NULL); | |
4872 | vect_is_simple_use (then_clause, loop_vinfo, | |
4873 | NULL, >emp, &def, &dts[2]); | |
4874 | } | |
4875 | if (reduc_index == 2) | |
4876 | vec_else_clause = reduc_def; | |
4877 | else | |
4878 | { | |
4879 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
4880 | stmt, NULL); | |
4881 | vect_is_simple_use (else_clause, loop_vinfo, | |
4882 | NULL, >emp, &def, &dts[3]); | |
4883 | } | |
4884 | } | |
4885 | else | |
4886 | { | |
4887 | vec_cond_lhs = vect_get_vec_def_for_stmt_copy (dts[0], vec_cond_lhs); | |
4888 | vec_cond_rhs = vect_get_vec_def_for_stmt_copy (dts[1], vec_cond_rhs); | |
4889 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], | |
4890 | vec_then_clause); | |
4891 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], | |
4892 | vec_else_clause); | |
4893 | } | |
4894 | ||
9dc3f7de | 4895 | /* Arguments are ready. Create the new vector stmt. */ |
a855b1b1 MM |
4896 | vec_compare = build2 (TREE_CODE (cond_expr), vectype, |
4897 | vec_cond_lhs, vec_cond_rhs); | |
4898 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, | |
4899 | vec_compare, vec_then_clause, vec_else_clause); | |
4900 | ||
4901 | new_stmt = gimple_build_assign (vec_dest, vec_cond_expr); | |
4902 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4903 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4904 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4905 | if (j == 0) | |
4906 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4907 | else | |
4908 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4909 | ||
4910 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4911 | } | |
b8698a0f | 4912 | |
ebfd146a IR |
4913 | return true; |
4914 | } | |
4915 | ||
4916 | ||
8644a673 | 4917 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
4918 | |
4919 | bool | |
a70d6342 | 4920 | vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 4921 | { |
8644a673 | 4922 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 4923 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 4924 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 4925 | bool ok; |
a70d6342 | 4926 | tree scalar_type, vectype; |
1107f3ae | 4927 | gimple pattern_stmt, pattern_def_stmt; |
ebfd146a IR |
4928 | |
4929 | if (vect_print_dump_info (REPORT_DETAILS)) | |
ebfd146a | 4930 | { |
8644a673 IR |
4931 | fprintf (vect_dump, "==> examining statement: "); |
4932 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4933 | } | |
ebfd146a | 4934 | |
1825a1f3 | 4935 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 4936 | { |
1825a1f3 IR |
4937 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
4938 | fprintf (vect_dump, "not vectorized: stmt has volatile operands"); | |
4939 | ||
4940 | return false; | |
4941 | } | |
b8698a0f L |
4942 | |
4943 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
4944 | to include: |
4945 | - the COND_EXPR which is the loop exit condition | |
4946 | - any LABEL_EXPRs in the loop | |
b8698a0f | 4947 | - computations that are used only for array indexing or loop control. |
8644a673 | 4948 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 4949 | instance, therefore, all the statements are relevant. |
ebfd146a | 4950 | |
83197f37 IR |
4951 | Pattern statement need to be analyzed instead of the original statement |
4952 | if the original statement is not relevant. Otherwise, we analyze both | |
4953 | statements. */ | |
4954 | ||
4955 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 4956 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 4957 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 4958 | { |
9d5e7640 | 4959 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 4960 | && pattern_stmt |
9d5e7640 IR |
4961 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
4962 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
4963 | { | |
83197f37 | 4964 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
4965 | stmt = pattern_stmt; |
4966 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
4967 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4968 | { | |
4969 | fprintf (vect_dump, "==> examining pattern statement: "); | |
4970 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4971 | } | |
4972 | } | |
4973 | else | |
4974 | { | |
4975 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4976 | fprintf (vect_dump, "irrelevant."); | |
ebfd146a | 4977 | |
9d5e7640 IR |
4978 | return true; |
4979 | } | |
8644a673 | 4980 | } |
83197f37 IR |
4981 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
4982 | && pattern_stmt | |
4983 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
4984 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
4985 | { | |
4986 | /* Analyze PATTERN_STMT too. */ | |
4987 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4988 | { | |
4989 | fprintf (vect_dump, "==> examining pattern statement: "); | |
4990 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4991 | } | |
4992 | ||
4993 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
4994 | return false; | |
4995 | } | |
ebfd146a | 4996 | |
1107f3ae IR |
4997 | if (is_pattern_stmt_p (stmt_info) |
4998 | && (pattern_def_stmt = STMT_VINFO_PATTERN_DEF_STMT (stmt_info)) | |
4999 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) | |
5000 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt)))) | |
5001 | { | |
5002 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
5003 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5004 | { | |
5005 | fprintf (vect_dump, "==> examining pattern def statement: "); | |
5006 | print_gimple_stmt (vect_dump, pattern_def_stmt, 0, TDF_SLIM); | |
5007 | } | |
5008 | ||
5009 | if (!vect_analyze_stmt (pattern_def_stmt, need_to_vectorize, node)) | |
5010 | return false; | |
5011 | } | |
5012 | ||
5013 | ||
8644a673 IR |
5014 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
5015 | { | |
5016 | case vect_internal_def: | |
5017 | break; | |
ebfd146a | 5018 | |
8644a673 | 5019 | case vect_reduction_def: |
7c5222ff | 5020 | case vect_nested_cycle: |
a70d6342 | 5021 | gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer |
8644a673 | 5022 | || relevance == vect_used_in_outer_by_reduction |
a70d6342 | 5023 | || relevance == vect_unused_in_scope)); |
8644a673 IR |
5024 | break; |
5025 | ||
5026 | case vect_induction_def: | |
5027 | case vect_constant_def: | |
5028 | case vect_external_def: | |
5029 | case vect_unknown_def_type: | |
5030 | default: | |
5031 | gcc_unreachable (); | |
5032 | } | |
ebfd146a | 5033 | |
a70d6342 IR |
5034 | if (bb_vinfo) |
5035 | { | |
5036 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
5037 | ||
b690cc0f | 5038 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
a70d6342 IR |
5039 | if (vect_print_dump_info (REPORT_DETAILS)) |
5040 | { | |
5041 | fprintf (vect_dump, "get vectype for scalar type: "); | |
5042 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
5043 | } | |
5044 | ||
5045 | vectype = get_vectype_for_scalar_type (scalar_type); | |
5046 | if (!vectype) | |
5047 | { | |
5048 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5049 | { | |
5050 | fprintf (vect_dump, "not SLPed: unsupported data-type "); | |
5051 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
5052 | } | |
5053 | return false; | |
5054 | } | |
5055 | ||
5056 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5057 | { | |
5058 | fprintf (vect_dump, "vectype: "); | |
5059 | print_generic_expr (vect_dump, vectype, TDF_SLIM); | |
5060 | } | |
5061 | ||
5062 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
5063 | } | |
5064 | ||
8644a673 | 5065 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 5066 | { |
8644a673 IR |
5067 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
5068 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info)); | |
5069 | *need_to_vectorize = true; | |
ebfd146a IR |
5070 | } |
5071 | ||
8644a673 | 5072 | ok = true; |
b8698a0f | 5073 | if (!bb_vinfo |
a70d6342 IR |
5074 | && (STMT_VINFO_RELEVANT_P (stmt_info) |
5075 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
8644a673 IR |
5076 | ok = (vectorizable_type_promotion (stmt, NULL, NULL, NULL) |
5077 | || vectorizable_type_demotion (stmt, NULL, NULL, NULL) | |
5078 | || vectorizable_conversion (stmt, NULL, NULL, NULL) | |
9dc3f7de | 5079 | || vectorizable_shift (stmt, NULL, NULL, NULL) |
8644a673 IR |
5080 | || vectorizable_operation (stmt, NULL, NULL, NULL) |
5081 | || vectorizable_assignment (stmt, NULL, NULL, NULL) | |
5082 | || vectorizable_load (stmt, NULL, NULL, NULL, NULL) | |
5083 | || vectorizable_call (stmt, NULL, NULL) | |
5084 | || vectorizable_store (stmt, NULL, NULL, NULL) | |
b5aeb3bb | 5085 | || vectorizable_reduction (stmt, NULL, NULL, NULL) |
4bbe8262 | 5086 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0)); |
a70d6342 IR |
5087 | else |
5088 | { | |
5089 | if (bb_vinfo) | |
57416708 | 5090 | ok = (vectorizable_shift (stmt, NULL, NULL, node) |
9dc3f7de | 5091 | || vectorizable_operation (stmt, NULL, NULL, node) |
a70d6342 IR |
5092 | || vectorizable_assignment (stmt, NULL, NULL, node) |
5093 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
5094 | || vectorizable_store (stmt, NULL, NULL, node)); | |
b8698a0f | 5095 | } |
8644a673 IR |
5096 | |
5097 | if (!ok) | |
ebfd146a | 5098 | { |
8644a673 IR |
5099 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
5100 | { | |
5101 | fprintf (vect_dump, "not vectorized: relevant stmt not "); | |
5102 | fprintf (vect_dump, "supported: "); | |
5103 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
5104 | } | |
b8698a0f | 5105 | |
ebfd146a IR |
5106 | return false; |
5107 | } | |
5108 | ||
a70d6342 IR |
5109 | if (bb_vinfo) |
5110 | return true; | |
5111 | ||
8644a673 IR |
5112 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
5113 | need extra handling, except for vectorizable reductions. */ | |
5114 | if (STMT_VINFO_LIVE_P (stmt_info) | |
5115 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
5116 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 5117 | |
8644a673 | 5118 | if (!ok) |
ebfd146a | 5119 | { |
8644a673 IR |
5120 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
5121 | { | |
5122 | fprintf (vect_dump, "not vectorized: live stmt not "); | |
5123 | fprintf (vect_dump, "supported: "); | |
5124 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
5125 | } | |
b8698a0f | 5126 | |
8644a673 | 5127 | return false; |
ebfd146a IR |
5128 | } |
5129 | ||
ebfd146a IR |
5130 | return true; |
5131 | } | |
5132 | ||
5133 | ||
5134 | /* Function vect_transform_stmt. | |
5135 | ||
5136 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
5137 | ||
5138 | bool | |
5139 | vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi, | |
b8698a0f | 5140 | bool *strided_store, slp_tree slp_node, |
ebfd146a IR |
5141 | slp_instance slp_node_instance) |
5142 | { | |
5143 | bool is_store = false; | |
5144 | gimple vec_stmt = NULL; | |
5145 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a | 5146 | bool done; |
ebfd146a IR |
5147 | |
5148 | switch (STMT_VINFO_TYPE (stmt_info)) | |
5149 | { | |
5150 | case type_demotion_vec_info_type: | |
5151 | done = vectorizable_type_demotion (stmt, gsi, &vec_stmt, slp_node); | |
5152 | gcc_assert (done); | |
5153 | break; | |
5154 | ||
5155 | case type_promotion_vec_info_type: | |
5156 | done = vectorizable_type_promotion (stmt, gsi, &vec_stmt, slp_node); | |
5157 | gcc_assert (done); | |
5158 | break; | |
5159 | ||
5160 | case type_conversion_vec_info_type: | |
5161 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
5162 | gcc_assert (done); | |
5163 | break; | |
5164 | ||
5165 | case induc_vec_info_type: | |
5166 | gcc_assert (!slp_node); | |
5167 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
5168 | gcc_assert (done); | |
5169 | break; | |
5170 | ||
9dc3f7de IR |
5171 | case shift_vec_info_type: |
5172 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
5173 | gcc_assert (done); | |
5174 | break; | |
5175 | ||
ebfd146a IR |
5176 | case op_vec_info_type: |
5177 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
5178 | gcc_assert (done); | |
5179 | break; | |
5180 | ||
5181 | case assignment_vec_info_type: | |
5182 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
5183 | gcc_assert (done); | |
5184 | break; | |
5185 | ||
5186 | case load_vec_info_type: | |
b8698a0f | 5187 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
5188 | slp_node_instance); |
5189 | gcc_assert (done); | |
5190 | break; | |
5191 | ||
5192 | case store_vec_info_type: | |
5193 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
5194 | gcc_assert (done); | |
5195 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) && !slp_node) | |
5196 | { | |
5197 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 5198 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
5199 | one are skipped, and there vec_stmt_info shouldn't be freed |
5200 | meanwhile. */ | |
5201 | *strided_store = true; | |
5202 | if (STMT_VINFO_VEC_STMT (stmt_info)) | |
5203 | is_store = true; | |
5204 | } | |
5205 | else | |
5206 | is_store = true; | |
5207 | break; | |
5208 | ||
5209 | case condition_vec_info_type: | |
5210 | gcc_assert (!slp_node); | |
4bbe8262 | 5211 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0); |
ebfd146a IR |
5212 | gcc_assert (done); |
5213 | break; | |
5214 | ||
5215 | case call_vec_info_type: | |
5216 | gcc_assert (!slp_node); | |
5217 | done = vectorizable_call (stmt, gsi, &vec_stmt); | |
039d9ea1 | 5218 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
5219 | break; |
5220 | ||
5221 | case reduc_vec_info_type: | |
b5aeb3bb | 5222 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
5223 | gcc_assert (done); |
5224 | break; | |
5225 | ||
5226 | default: | |
5227 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
5228 | { | |
5229 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5230 | fprintf (vect_dump, "stmt not supported."); | |
5231 | gcc_unreachable (); | |
5232 | } | |
5233 | } | |
5234 | ||
5235 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that | |
5236 | is being vectorized, but outside the immediately enclosing loop. */ | |
5237 | if (vec_stmt | |
a70d6342 IR |
5238 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
5239 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
5240 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
5241 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
5242 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 5243 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 5244 | vect_used_in_outer_by_reduction)) |
ebfd146a | 5245 | { |
a70d6342 IR |
5246 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
5247 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
5248 | imm_use_iterator imm_iter; |
5249 | use_operand_p use_p; | |
5250 | tree scalar_dest; | |
5251 | gimple exit_phi; | |
5252 | ||
5253 | if (vect_print_dump_info (REPORT_DETAILS)) | |
a70d6342 | 5254 | fprintf (vect_dump, "Record the vdef for outer-loop vectorization."); |
ebfd146a IR |
5255 | |
5256 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
5257 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
5258 | STMT). */ | |
5259 | if (gimple_code (stmt) == GIMPLE_PHI) | |
5260 | scalar_dest = PHI_RESULT (stmt); | |
5261 | else | |
5262 | scalar_dest = gimple_assign_lhs (stmt); | |
5263 | ||
5264 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
5265 | { | |
5266 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
5267 | { | |
5268 | exit_phi = USE_STMT (use_p); | |
5269 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
5270 | } | |
5271 | } | |
5272 | } | |
5273 | ||
5274 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
5275 | being vectorized. */ | |
5276 | if (STMT_VINFO_LIVE_P (stmt_info) | |
5277 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
5278 | { | |
5279 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
5280 | gcc_assert (done); | |
5281 | } | |
5282 | ||
5283 | if (vec_stmt) | |
83197f37 | 5284 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 5285 | |
b8698a0f | 5286 | return is_store; |
ebfd146a IR |
5287 | } |
5288 | ||
5289 | ||
b8698a0f | 5290 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
5291 | stmt_vec_info. */ |
5292 | ||
5293 | void | |
5294 | vect_remove_stores (gimple first_stmt) | |
5295 | { | |
5296 | gimple next = first_stmt; | |
5297 | gimple tmp; | |
5298 | gimple_stmt_iterator next_si; | |
5299 | ||
5300 | while (next) | |
5301 | { | |
5302 | /* Free the attached stmt_vec_info and remove the stmt. */ | |
5303 | next_si = gsi_for_stmt (next); | |
5304 | gsi_remove (&next_si, true); | |
e14c1050 | 5305 | tmp = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next)); |
ebfd146a IR |
5306 | free_stmt_vec_info (next); |
5307 | next = tmp; | |
5308 | } | |
5309 | } | |
5310 | ||
5311 | ||
5312 | /* Function new_stmt_vec_info. | |
5313 | ||
5314 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
5315 | ||
5316 | stmt_vec_info | |
b8698a0f | 5317 | new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 5318 | bb_vec_info bb_vinfo) |
ebfd146a IR |
5319 | { |
5320 | stmt_vec_info res; | |
5321 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
5322 | ||
5323 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
5324 | STMT_VINFO_STMT (res) = stmt; | |
5325 | STMT_VINFO_LOOP_VINFO (res) = loop_vinfo; | |
a70d6342 | 5326 | STMT_VINFO_BB_VINFO (res) = bb_vinfo; |
8644a673 | 5327 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
5328 | STMT_VINFO_LIVE_P (res) = false; |
5329 | STMT_VINFO_VECTYPE (res) = NULL; | |
5330 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 5331 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
5332 | STMT_VINFO_IN_PATTERN_P (res) = false; |
5333 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
1107f3ae | 5334 | STMT_VINFO_PATTERN_DEF_STMT (res) = NULL; |
ebfd146a IR |
5335 | STMT_VINFO_DATA_REF (res) = NULL; |
5336 | ||
5337 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
5338 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
5339 | STMT_VINFO_DR_INIT (res) = NULL; | |
5340 | STMT_VINFO_DR_STEP (res) = NULL; | |
5341 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
5342 | ||
5343 | if (gimple_code (stmt) == GIMPLE_PHI | |
5344 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
5345 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
5346 | else | |
8644a673 IR |
5347 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
5348 | ||
ebfd146a IR |
5349 | STMT_VINFO_SAME_ALIGN_REFS (res) = VEC_alloc (dr_p, heap, 5); |
5350 | STMT_VINFO_INSIDE_OF_LOOP_COST (res) = 0; | |
5351 | STMT_VINFO_OUTSIDE_OF_LOOP_COST (res) = 0; | |
32e8bb8e | 5352 | STMT_SLP_TYPE (res) = loop_vect; |
e14c1050 IR |
5353 | GROUP_FIRST_ELEMENT (res) = NULL; |
5354 | GROUP_NEXT_ELEMENT (res) = NULL; | |
5355 | GROUP_SIZE (res) = 0; | |
5356 | GROUP_STORE_COUNT (res) = 0; | |
5357 | GROUP_GAP (res) = 0; | |
5358 | GROUP_SAME_DR_STMT (res) = NULL; | |
5359 | GROUP_READ_WRITE_DEPENDENCE (res) = false; | |
ebfd146a IR |
5360 | |
5361 | return res; | |
5362 | } | |
5363 | ||
5364 | ||
5365 | /* Create a hash table for stmt_vec_info. */ | |
5366 | ||
5367 | void | |
5368 | init_stmt_vec_info_vec (void) | |
5369 | { | |
5370 | gcc_assert (!stmt_vec_info_vec); | |
5371 | stmt_vec_info_vec = VEC_alloc (vec_void_p, heap, 50); | |
5372 | } | |
5373 | ||
5374 | ||
5375 | /* Free hash table for stmt_vec_info. */ | |
5376 | ||
5377 | void | |
5378 | free_stmt_vec_info_vec (void) | |
5379 | { | |
5380 | gcc_assert (stmt_vec_info_vec); | |
5381 | VEC_free (vec_void_p, heap, stmt_vec_info_vec); | |
5382 | } | |
5383 | ||
5384 | ||
5385 | /* Free stmt vectorization related info. */ | |
5386 | ||
5387 | void | |
5388 | free_stmt_vec_info (gimple stmt) | |
5389 | { | |
5390 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5391 | ||
5392 | if (!stmt_info) | |
5393 | return; | |
5394 | ||
5395 | VEC_free (dr_p, heap, STMT_VINFO_SAME_ALIGN_REFS (stmt_info)); | |
5396 | set_vinfo_for_stmt (stmt, NULL); | |
5397 | free (stmt_info); | |
5398 | } | |
5399 | ||
5400 | ||
bb67d9c7 | 5401 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 5402 | |
bb67d9c7 | 5403 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
5404 | by the target. */ |
5405 | ||
bb67d9c7 RG |
5406 | static tree |
5407 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a IR |
5408 | { |
5409 | enum machine_mode inner_mode = TYPE_MODE (scalar_type); | |
cc4b5170 | 5410 | enum machine_mode simd_mode; |
2f816591 | 5411 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
5412 | int nunits; |
5413 | tree vectype; | |
5414 | ||
cc4b5170 | 5415 | if (nbytes == 0) |
ebfd146a IR |
5416 | return NULL_TREE; |
5417 | ||
2f816591 RG |
5418 | /* We can't build a vector type of elements with alignment bigger than |
5419 | their size. */ | |
5420 | if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) | |
5421 | return NULL_TREE; | |
5422 | ||
6d7971b8 RG |
5423 | /* If we'd build a vector type of elements whose mode precision doesn't |
5424 | match their types precision we'll get mismatched types on vector | |
5425 | extracts via BIT_FIELD_REFs. This effectively means we disable | |
5426 | vectorization of bool and/or enum types in some languages. */ | |
5427 | if (INTEGRAL_TYPE_P (scalar_type) | |
5428 | && GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type)) | |
5429 | return NULL_TREE; | |
5430 | ||
cc4b5170 RG |
5431 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
5432 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
5433 | return NULL_TREE; | |
5434 | ||
ccbf5bb4 RG |
5435 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
5436 | When the component mode passes the above test simply use a type | |
5437 | corresponding to that mode. The theory is that any use that | |
5438 | would cause problems with this will disable vectorization anyway. */ | |
5439 | if (!SCALAR_FLOAT_TYPE_P (scalar_type) | |
5440 | && !INTEGRAL_TYPE_P (scalar_type) | |
5441 | && !POINTER_TYPE_P (scalar_type)) | |
5442 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); | |
5443 | ||
bb67d9c7 RG |
5444 | /* If no size was supplied use the mode the target prefers. Otherwise |
5445 | lookup a vector mode of the specified size. */ | |
5446 | if (size == 0) | |
5447 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
5448 | else | |
5449 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
5450 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
5451 | if (nunits <= 1) | |
5452 | return NULL_TREE; | |
ebfd146a IR |
5453 | |
5454 | vectype = build_vector_type (scalar_type, nunits); | |
5455 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5456 | { | |
5457 | fprintf (vect_dump, "get vectype with %d units of type ", nunits); | |
5458 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
5459 | } | |
5460 | ||
5461 | if (!vectype) | |
5462 | return NULL_TREE; | |
5463 | ||
5464 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5465 | { | |
5466 | fprintf (vect_dump, "vectype: "); | |
5467 | print_generic_expr (vect_dump, vectype, TDF_SLIM); | |
5468 | } | |
5469 | ||
5470 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
5471 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
5472 | { | |
5473 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5474 | fprintf (vect_dump, "mode not supported by target."); | |
5475 | return NULL_TREE; | |
5476 | } | |
5477 | ||
5478 | return vectype; | |
5479 | } | |
5480 | ||
bb67d9c7 RG |
5481 | unsigned int current_vector_size; |
5482 | ||
5483 | /* Function get_vectype_for_scalar_type. | |
5484 | ||
5485 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
5486 | by the target. */ | |
5487 | ||
5488 | tree | |
5489 | get_vectype_for_scalar_type (tree scalar_type) | |
5490 | { | |
5491 | tree vectype; | |
5492 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
5493 | current_vector_size); | |
5494 | if (vectype | |
5495 | && current_vector_size == 0) | |
5496 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
5497 | return vectype; | |
5498 | } | |
5499 | ||
b690cc0f RG |
5500 | /* Function get_same_sized_vectype |
5501 | ||
5502 | Returns a vector type corresponding to SCALAR_TYPE of size | |
5503 | VECTOR_TYPE if supported by the target. */ | |
5504 | ||
5505 | tree | |
bb67d9c7 | 5506 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 5507 | { |
bb67d9c7 RG |
5508 | return get_vectype_for_scalar_type_and_size |
5509 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
5510 | } |
5511 | ||
ebfd146a IR |
5512 | /* Function vect_is_simple_use. |
5513 | ||
5514 | Input: | |
a70d6342 IR |
5515 | LOOP_VINFO - the vect info of the loop that is being vectorized. |
5516 | BB_VINFO - the vect info of the basic block that is being vectorized. | |
5517 | OPERAND - operand of a stmt in the loop or bb. | |
ebfd146a IR |
5518 | DEF - the defining stmt in case OPERAND is an SSA_NAME. |
5519 | ||
5520 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 5521 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 5522 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 5523 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
5524 | is the case in reduction/induction computations). |
5525 | For basic blocks, supportable operands are constants and bb invariants. | |
5526 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
5527 | |
5528 | bool | |
b8698a0f | 5529 | vect_is_simple_use (tree operand, loop_vec_info loop_vinfo, |
a70d6342 | 5530 | bb_vec_info bb_vinfo, gimple *def_stmt, |
ebfd146a | 5531 | tree *def, enum vect_def_type *dt) |
b8698a0f | 5532 | { |
ebfd146a IR |
5533 | basic_block bb; |
5534 | stmt_vec_info stmt_vinfo; | |
a70d6342 | 5535 | struct loop *loop = NULL; |
b8698a0f | 5536 | |
a70d6342 IR |
5537 | if (loop_vinfo) |
5538 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
5539 | |
5540 | *def_stmt = NULL; | |
5541 | *def = NULL_TREE; | |
b8698a0f | 5542 | |
ebfd146a IR |
5543 | if (vect_print_dump_info (REPORT_DETAILS)) |
5544 | { | |
5545 | fprintf (vect_dump, "vect_is_simple_use: operand "); | |
5546 | print_generic_expr (vect_dump, operand, TDF_SLIM); | |
5547 | } | |
b8698a0f | 5548 | |
ebfd146a IR |
5549 | if (TREE_CODE (operand) == INTEGER_CST || TREE_CODE (operand) == REAL_CST) |
5550 | { | |
5551 | *dt = vect_constant_def; | |
5552 | return true; | |
5553 | } | |
b8698a0f | 5554 | |
ebfd146a IR |
5555 | if (is_gimple_min_invariant (operand)) |
5556 | { | |
5557 | *def = operand; | |
8644a673 | 5558 | *dt = vect_external_def; |
ebfd146a IR |
5559 | return true; |
5560 | } | |
5561 | ||
5562 | if (TREE_CODE (operand) == PAREN_EXPR) | |
5563 | { | |
5564 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5565 | fprintf (vect_dump, "non-associatable copy."); | |
5566 | operand = TREE_OPERAND (operand, 0); | |
5567 | } | |
b8698a0f | 5568 | |
ebfd146a IR |
5569 | if (TREE_CODE (operand) != SSA_NAME) |
5570 | { | |
5571 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5572 | fprintf (vect_dump, "not ssa-name."); | |
5573 | return false; | |
5574 | } | |
b8698a0f | 5575 | |
ebfd146a IR |
5576 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
5577 | if (*def_stmt == NULL) | |
5578 | { | |
5579 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5580 | fprintf (vect_dump, "no def_stmt."); | |
5581 | return false; | |
5582 | } | |
5583 | ||
5584 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5585 | { | |
5586 | fprintf (vect_dump, "def_stmt: "); | |
5587 | print_gimple_stmt (vect_dump, *def_stmt, 0, TDF_SLIM); | |
5588 | } | |
5589 | ||
8644a673 | 5590 | /* Empty stmt is expected only in case of a function argument. |
ebfd146a IR |
5591 | (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */ |
5592 | if (gimple_nop_p (*def_stmt)) | |
5593 | { | |
5594 | *def = operand; | |
8644a673 | 5595 | *dt = vect_external_def; |
ebfd146a IR |
5596 | return true; |
5597 | } | |
5598 | ||
5599 | bb = gimple_bb (*def_stmt); | |
a70d6342 IR |
5600 | |
5601 | if ((loop && !flow_bb_inside_loop_p (loop, bb)) | |
5602 | || (!loop && bb != BB_VINFO_BB (bb_vinfo)) | |
b8698a0f | 5603 | || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI)) |
8644a673 | 5604 | *dt = vect_external_def; |
ebfd146a IR |
5605 | else |
5606 | { | |
5607 | stmt_vinfo = vinfo_for_stmt (*def_stmt); | |
5608 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
5609 | } | |
5610 | ||
5611 | if (*dt == vect_unknown_def_type) | |
5612 | { | |
5613 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5614 | fprintf (vect_dump, "Unsupported pattern."); | |
5615 | return false; | |
5616 | } | |
5617 | ||
5618 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5619 | fprintf (vect_dump, "type of def: %d.",*dt); | |
5620 | ||
5621 | switch (gimple_code (*def_stmt)) | |
5622 | { | |
5623 | case GIMPLE_PHI: | |
5624 | *def = gimple_phi_result (*def_stmt); | |
5625 | break; | |
5626 | ||
5627 | case GIMPLE_ASSIGN: | |
5628 | *def = gimple_assign_lhs (*def_stmt); | |
5629 | break; | |
5630 | ||
5631 | case GIMPLE_CALL: | |
5632 | *def = gimple_call_lhs (*def_stmt); | |
5633 | if (*def != NULL) | |
5634 | break; | |
5635 | /* FALLTHRU */ | |
5636 | default: | |
5637 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5638 | fprintf (vect_dump, "unsupported defining stmt: "); | |
5639 | return false; | |
5640 | } | |
5641 | ||
5642 | return true; | |
5643 | } | |
5644 | ||
b690cc0f RG |
5645 | /* Function vect_is_simple_use_1. |
5646 | ||
5647 | Same as vect_is_simple_use_1 but also determines the vector operand | |
5648 | type of OPERAND and stores it to *VECTYPE. If the definition of | |
5649 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
5650 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
5651 | is responsible to compute the best suited vector type for the | |
5652 | scalar operand. */ | |
5653 | ||
5654 | bool | |
5655 | vect_is_simple_use_1 (tree operand, loop_vec_info loop_vinfo, | |
5656 | bb_vec_info bb_vinfo, gimple *def_stmt, | |
5657 | tree *def, enum vect_def_type *dt, tree *vectype) | |
5658 | { | |
5659 | if (!vect_is_simple_use (operand, loop_vinfo, bb_vinfo, def_stmt, def, dt)) | |
5660 | return false; | |
5661 | ||
5662 | /* Now get a vector type if the def is internal, otherwise supply | |
5663 | NULL_TREE and leave it up to the caller to figure out a proper | |
5664 | type for the use stmt. */ | |
5665 | if (*dt == vect_internal_def | |
5666 | || *dt == vect_induction_def | |
5667 | || *dt == vect_reduction_def | |
5668 | || *dt == vect_double_reduction_def | |
5669 | || *dt == vect_nested_cycle) | |
5670 | { | |
5671 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
5672 | |
5673 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
5674 | && !STMT_VINFO_RELEVANT (stmt_info) | |
5675 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 5676 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 5677 | |
b690cc0f RG |
5678 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
5679 | gcc_assert (*vectype != NULL_TREE); | |
5680 | } | |
5681 | else if (*dt == vect_uninitialized_def | |
5682 | || *dt == vect_constant_def | |
5683 | || *dt == vect_external_def) | |
5684 | *vectype = NULL_TREE; | |
5685 | else | |
5686 | gcc_unreachable (); | |
5687 | ||
5688 | return true; | |
5689 | } | |
5690 | ||
ebfd146a IR |
5691 | |
5692 | /* Function supportable_widening_operation | |
5693 | ||
b8698a0f L |
5694 | Check whether an operation represented by the code CODE is a |
5695 | widening operation that is supported by the target platform in | |
b690cc0f RG |
5696 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
5697 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 5698 | |
ebfd146a IR |
5699 | Widening operations we currently support are NOP (CONVERT), FLOAT |
5700 | and WIDEN_MULT. This function checks if these operations are supported | |
5701 | by the target platform either directly (via vector tree-codes), or via | |
5702 | target builtins. | |
5703 | ||
5704 | Output: | |
b8698a0f L |
5705 | - CODE1 and CODE2 are codes of vector operations to be used when |
5706 | vectorizing the operation, if available. | |
ebfd146a | 5707 | - DECL1 and DECL2 are decls of target builtin functions to be used |
ff802fa1 | 5708 | when vectorizing the operation, if available. In this case, |
b8698a0f | 5709 | CODE1 and CODE2 are CALL_EXPR. |
ebfd146a IR |
5710 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
5711 | case of multi-step conversion (like char->short->int - in that case | |
5712 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
5713 | - INTERM_TYPES contains the intermediate type required to perform the |
5714 | widening operation (short in the above example). */ | |
ebfd146a IR |
5715 | |
5716 | bool | |
b690cc0f RG |
5717 | supportable_widening_operation (enum tree_code code, gimple stmt, |
5718 | tree vectype_out, tree vectype_in, | |
ebfd146a IR |
5719 | tree *decl1, tree *decl2, |
5720 | enum tree_code *code1, enum tree_code *code2, | |
5721 | int *multi_step_cvt, | |
5722 | VEC (tree, heap) **interm_types) | |
5723 | { | |
5724 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5725 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
5726 | struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info); | |
5727 | bool ordered_p; | |
5728 | enum machine_mode vec_mode; | |
81f40b79 | 5729 | enum insn_code icode1, icode2; |
ebfd146a | 5730 | optab optab1, optab2; |
b690cc0f RG |
5731 | tree vectype = vectype_in; |
5732 | tree wide_vectype = vectype_out; | |
ebfd146a IR |
5733 | enum tree_code c1, c2; |
5734 | ||
5735 | /* The result of a vectorized widening operation usually requires two vectors | |
b8698a0f L |
5736 | (because the widened results do not fit int one vector). The generated |
5737 | vector results would normally be expected to be generated in the same | |
ebfd146a IR |
5738 | order as in the original scalar computation, i.e. if 8 results are |
5739 | generated in each vector iteration, they are to be organized as follows: | |
b8698a0f | 5740 | vect1: [res1,res2,res3,res4], vect2: [res5,res6,res7,res8]. |
ebfd146a | 5741 | |
b8698a0f | 5742 | However, in the special case that the result of the widening operation is |
ebfd146a | 5743 | used in a reduction computation only, the order doesn't matter (because |
b8698a0f | 5744 | when vectorizing a reduction we change the order of the computation). |
ebfd146a IR |
5745 | Some targets can take advantage of this and generate more efficient code. |
5746 | For example, targets like Altivec, that support widen_mult using a sequence | |
5747 | of {mult_even,mult_odd} generate the following vectors: | |
5748 | vect1: [res1,res3,res5,res7], vect2: [res2,res4,res6,res8]. | |
5749 | ||
5750 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
b8698a0f | 5751 | (each vectorized inner-loop iteration contributes to VF outer-loop |
ff802fa1 | 5752 | iterations in parallel). We therefore don't allow to change the order |
ebfd146a IR |
5753 | of the computation in the inner-loop during outer-loop vectorization. */ |
5754 | ||
5755 | if (STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
5756 | && !nested_in_vect_loop_p (vect_loop, stmt)) | |
5757 | ordered_p = false; | |
5758 | else | |
5759 | ordered_p = true; | |
5760 | ||
5761 | if (!ordered_p | |
5762 | && code == WIDEN_MULT_EXPR | |
5763 | && targetm.vectorize.builtin_mul_widen_even | |
5764 | && targetm.vectorize.builtin_mul_widen_even (vectype) | |
5765 | && targetm.vectorize.builtin_mul_widen_odd | |
5766 | && targetm.vectorize.builtin_mul_widen_odd (vectype)) | |
5767 | { | |
5768 | if (vect_print_dump_info (REPORT_DETAILS)) | |
5769 | fprintf (vect_dump, "Unordered widening operation detected."); | |
5770 | ||
5771 | *code1 = *code2 = CALL_EXPR; | |
5772 | *decl1 = targetm.vectorize.builtin_mul_widen_even (vectype); | |
5773 | *decl2 = targetm.vectorize.builtin_mul_widen_odd (vectype); | |
5774 | return true; | |
5775 | } | |
5776 | ||
5777 | switch (code) | |
5778 | { | |
5779 | case WIDEN_MULT_EXPR: | |
5780 | if (BYTES_BIG_ENDIAN) | |
5781 | { | |
5782 | c1 = VEC_WIDEN_MULT_HI_EXPR; | |
5783 | c2 = VEC_WIDEN_MULT_LO_EXPR; | |
5784 | } | |
5785 | else | |
5786 | { | |
5787 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
5788 | c1 = VEC_WIDEN_MULT_LO_EXPR; | |
5789 | } | |
5790 | break; | |
5791 | ||
5792 | CASE_CONVERT: | |
5793 | if (BYTES_BIG_ENDIAN) | |
5794 | { | |
5795 | c1 = VEC_UNPACK_HI_EXPR; | |
5796 | c2 = VEC_UNPACK_LO_EXPR; | |
5797 | } | |
5798 | else | |
5799 | { | |
5800 | c2 = VEC_UNPACK_HI_EXPR; | |
5801 | c1 = VEC_UNPACK_LO_EXPR; | |
5802 | } | |
5803 | break; | |
5804 | ||
5805 | case FLOAT_EXPR: | |
5806 | if (BYTES_BIG_ENDIAN) | |
5807 | { | |
5808 | c1 = VEC_UNPACK_FLOAT_HI_EXPR; | |
5809 | c2 = VEC_UNPACK_FLOAT_LO_EXPR; | |
5810 | } | |
5811 | else | |
5812 | { | |
5813 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
5814 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; | |
5815 | } | |
5816 | break; | |
5817 | ||
5818 | case FIX_TRUNC_EXPR: | |
5819 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
5820 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
5821 | computing the operation. */ | |
5822 | return false; | |
5823 | ||
5824 | default: | |
5825 | gcc_unreachable (); | |
5826 | } | |
5827 | ||
5828 | if (code == FIX_TRUNC_EXPR) | |
5829 | { | |
5830 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
5831 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
5832 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
5833 | } |
5834 | else | |
5835 | { | |
5836 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
5837 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
5838 | } | |
5839 | ||
5840 | if (!optab1 || !optab2) | |
5841 | return false; | |
5842 | ||
5843 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
5844 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
5845 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
5846 | return false; |
5847 | ||
b8698a0f | 5848 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a IR |
5849 | types. */ |
5850 | if (insn_data[icode1].operand[0].mode != TYPE_MODE (wide_vectype) | |
5851 | || insn_data[icode2].operand[0].mode != TYPE_MODE (wide_vectype)) | |
5852 | { | |
5853 | int i; | |
5854 | tree prev_type = vectype, intermediate_type; | |
5855 | enum machine_mode intermediate_mode, prev_mode = vec_mode; | |
5856 | optab optab3, optab4; | |
5857 | ||
5858 | if (!CONVERT_EXPR_CODE_P (code)) | |
5859 | return false; | |
b8698a0f | 5860 | |
ebfd146a IR |
5861 | *code1 = c1; |
5862 | *code2 = c2; | |
b8698a0f | 5863 | |
ebfd146a | 5864 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
ff802fa1 IR |
5865 | intermediate steps in promotion sequence. We try |
5866 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
5867 | not. */ | |
ebfd146a IR |
5868 | *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS); |
5869 | for (i = 0; i < 3; i++) | |
5870 | { | |
5871 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
5872 | intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode, | |
5873 | TYPE_UNSIGNED (prev_type)); | |
5874 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
5875 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
5876 | ||
5877 | if (!optab3 || !optab4 | |
947131ba RS |
5878 | || ((icode1 = optab_handler (optab1, prev_mode)) |
5879 | == CODE_FOR_nothing) | |
ebfd146a | 5880 | || insn_data[icode1].operand[0].mode != intermediate_mode |
947131ba RS |
5881 | || ((icode2 = optab_handler (optab2, prev_mode)) |
5882 | == CODE_FOR_nothing) | |
ebfd146a | 5883 | || insn_data[icode2].operand[0].mode != intermediate_mode |
947131ba RS |
5884 | || ((icode1 = optab_handler (optab3, intermediate_mode)) |
5885 | == CODE_FOR_nothing) | |
5886 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
5887 | == CODE_FOR_nothing)) | |
ebfd146a IR |
5888 | return false; |
5889 | ||
5890 | VEC_quick_push (tree, *interm_types, intermediate_type); | |
5891 | (*multi_step_cvt)++; | |
5892 | ||
5893 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
5894 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5895 | return true; | |
5896 | ||
5897 | prev_type = intermediate_type; | |
5898 | prev_mode = intermediate_mode; | |
5899 | } | |
5900 | ||
5901 | return false; | |
5902 | } | |
5903 | ||
5904 | *code1 = c1; | |
5905 | *code2 = c2; | |
5906 | return true; | |
5907 | } | |
5908 | ||
5909 | ||
5910 | /* Function supportable_narrowing_operation | |
5911 | ||
b8698a0f L |
5912 | Check whether an operation represented by the code CODE is a |
5913 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
5914 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
5915 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 5916 | |
ebfd146a | 5917 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 5918 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
5919 | the target platform directly via vector tree-codes. |
5920 | ||
5921 | Output: | |
b8698a0f L |
5922 | - CODE1 is the code of a vector operation to be used when |
5923 | vectorizing the operation, if available. | |
ebfd146a IR |
5924 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
5925 | case of multi-step conversion (like int->short->char - in that case | |
5926 | MULTI_STEP_CVT will be 1). | |
5927 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 5928 | narrowing operation (short in the above example). */ |
ebfd146a IR |
5929 | |
5930 | bool | |
5931 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 5932 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
5933 | enum tree_code *code1, int *multi_step_cvt, |
5934 | VEC (tree, heap) **interm_types) | |
5935 | { | |
5936 | enum machine_mode vec_mode; | |
5937 | enum insn_code icode1; | |
5938 | optab optab1, interm_optab; | |
b690cc0f RG |
5939 | tree vectype = vectype_in; |
5940 | tree narrow_vectype = vectype_out; | |
ebfd146a IR |
5941 | enum tree_code c1; |
5942 | tree intermediate_type, prev_type; | |
5943 | int i; | |
5944 | ||
5945 | switch (code) | |
5946 | { | |
5947 | CASE_CONVERT: | |
5948 | c1 = VEC_PACK_TRUNC_EXPR; | |
5949 | break; | |
5950 | ||
5951 | case FIX_TRUNC_EXPR: | |
5952 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
5953 | break; | |
5954 | ||
5955 | case FLOAT_EXPR: | |
5956 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
5957 | tree code and optabs used for computing the operation. */ | |
5958 | return false; | |
5959 | ||
5960 | default: | |
5961 | gcc_unreachable (); | |
5962 | } | |
5963 | ||
5964 | if (code == FIX_TRUNC_EXPR) | |
5965 | /* The signedness is determined from output operand. */ | |
b690cc0f | 5966 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
5967 | else |
5968 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
5969 | ||
5970 | if (!optab1) | |
5971 | return false; | |
5972 | ||
5973 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 5974 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
5975 | return false; |
5976 | ||
5977 | /* Check if it's a multi-step conversion that can be done using intermediate | |
5978 | types. */ | |
5979 | if (insn_data[icode1].operand[0].mode != TYPE_MODE (narrow_vectype)) | |
5980 | { | |
5981 | enum machine_mode intermediate_mode, prev_mode = vec_mode; | |
5982 | ||
5983 | *code1 = c1; | |
5984 | prev_type = vectype; | |
5985 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS | |
ff802fa1 IR |
5986 | intermediate steps in promotion sequence. We try |
5987 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
5988 | not. */ | |
ebfd146a IR |
5989 | *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS); |
5990 | for (i = 0; i < 3; i++) | |
5991 | { | |
5992 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
5993 | intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode, | |
5994 | TYPE_UNSIGNED (prev_type)); | |
b8698a0f | 5995 | interm_optab = optab_for_tree_code (c1, intermediate_type, |
ebfd146a | 5996 | optab_default); |
b8698a0f | 5997 | if (!interm_optab |
947131ba RS |
5998 | || ((icode1 = optab_handler (optab1, prev_mode)) |
5999 | == CODE_FOR_nothing) | |
ebfd146a | 6000 | || insn_data[icode1].operand[0].mode != intermediate_mode |
947131ba RS |
6001 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) |
6002 | == CODE_FOR_nothing)) | |
ebfd146a IR |
6003 | return false; |
6004 | ||
6005 | VEC_quick_push (tree, *interm_types, intermediate_type); | |
6006 | (*multi_step_cvt)++; | |
6007 | ||
6008 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
6009 | return true; | |
6010 | ||
6011 | prev_type = intermediate_type; | |
6012 | prev_mode = intermediate_mode; | |
6013 | } | |
6014 | ||
6015 | return false; | |
6016 | } | |
6017 | ||
6018 | *code1 = c1; | |
6019 | return true; | |
6020 | } |