]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
d1e082c2 | 2 | Copyright (C) 2003-2013 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
78c60e3d | 25 | #include "dumpfile.h" |
ebfd146a IR |
26 | #include "tm.h" |
27 | #include "ggc.h" | |
28 | #include "tree.h" | |
29 | #include "target.h" | |
30 | #include "basic-block.h" | |
cf835838 | 31 | #include "gimple-pretty-print.h" |
ebfd146a | 32 | #include "tree-flow.h" |
ebfd146a | 33 | #include "cfgloop.h" |
ebfd146a | 34 | #include "expr.h" |
7ee2468b | 35 | #include "recog.h" /* FIXME: for insn_data */ |
ebfd146a | 36 | #include "optabs.h" |
718f9c0f | 37 | #include "diagnostic-core.h" |
ebfd146a | 38 | #include "tree-vectorizer.h" |
7ee2468b | 39 | #include "dumpfile.h" |
ebfd146a | 40 | |
7ee2468b SB |
41 | /* For lang_hooks.types.type_for_mode. */ |
42 | #include "langhooks.h" | |
ebfd146a | 43 | |
c3e7ee41 BS |
44 | /* Return the vectorized type for the given statement. */ |
45 | ||
46 | tree | |
47 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
48 | { | |
49 | return STMT_VINFO_VECTYPE (stmt_info); | |
50 | } | |
51 | ||
52 | /* Return TRUE iff the given statement is in an inner loop relative to | |
53 | the loop being vectorized. */ | |
54 | bool | |
55 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
56 | { | |
57 | gimple stmt = STMT_VINFO_STMT (stmt_info); | |
58 | basic_block bb = gimple_bb (stmt); | |
59 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
60 | struct loop* loop; | |
61 | ||
62 | if (!loop_vinfo) | |
63 | return false; | |
64 | ||
65 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
66 | ||
67 | return (bb->loop_father == loop->inner); | |
68 | } | |
69 | ||
70 | /* Record the cost of a statement, either by directly informing the | |
71 | target model or by saving it in a vector for later processing. | |
72 | Return a preliminary estimate of the statement's cost. */ | |
73 | ||
74 | unsigned | |
92345349 | 75 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 76 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 77 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 78 | { |
92345349 | 79 | if (body_cost_vec) |
c3e7ee41 | 80 | { |
92345349 BS |
81 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
82 | add_stmt_info_to_vec (body_cost_vec, count, kind, | |
83 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
84 | misalign); | |
c3e7ee41 | 85 | return (unsigned) |
92345349 | 86 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
87 | |
88 | } | |
89 | else | |
90 | { | |
91 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
92 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
93 | void *target_cost_data; | |
94 | ||
95 | if (loop_vinfo) | |
96 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
97 | else | |
98 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
99 | ||
92345349 BS |
100 | return add_stmt_cost (target_cost_data, count, kind, stmt_info, |
101 | misalign, where); | |
c3e7ee41 BS |
102 | } |
103 | } | |
104 | ||
272c6793 RS |
105 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
106 | ||
107 | static tree | |
108 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
109 | { | |
110 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
111 | "vect_array"); | |
112 | } | |
113 | ||
114 | /* ARRAY is an array of vectors created by create_vector_array. | |
115 | Return an SSA_NAME for the vector in index N. The reference | |
116 | is part of the vectorization of STMT and the vector is associated | |
117 | with scalar destination SCALAR_DEST. */ | |
118 | ||
119 | static tree | |
120 | read_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree scalar_dest, | |
121 | tree array, unsigned HOST_WIDE_INT n) | |
122 | { | |
123 | tree vect_type, vect, vect_name, array_ref; | |
124 | gimple new_stmt; | |
125 | ||
126 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
127 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
128 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
129 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
130 | build_int_cst (size_type_node, n), | |
131 | NULL_TREE, NULL_TREE); | |
132 | ||
133 | new_stmt = gimple_build_assign (vect, array_ref); | |
134 | vect_name = make_ssa_name (vect, new_stmt); | |
135 | gimple_assign_set_lhs (new_stmt, vect_name); | |
136 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
137 | |
138 | return vect_name; | |
139 | } | |
140 | ||
141 | /* ARRAY is an array of vectors created by create_vector_array. | |
142 | Emit code to store SSA_NAME VECT in index N of the array. | |
143 | The store is part of the vectorization of STMT. */ | |
144 | ||
145 | static void | |
146 | write_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree vect, | |
147 | tree array, unsigned HOST_WIDE_INT n) | |
148 | { | |
149 | tree array_ref; | |
150 | gimple new_stmt; | |
151 | ||
152 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
153 | build_int_cst (size_type_node, n), | |
154 | NULL_TREE, NULL_TREE); | |
155 | ||
156 | new_stmt = gimple_build_assign (array_ref, vect); | |
157 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
158 | } |
159 | ||
160 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
161 | of *PTR. The memory reference replaces those in FIRST_DR | |
162 | (and its group). */ | |
163 | ||
164 | static tree | |
165 | create_array_ref (tree type, tree ptr, struct data_reference *first_dr) | |
166 | { | |
272c6793 RS |
167 | tree mem_ref, alias_ptr_type; |
168 | ||
169 | alias_ptr_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
170 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); | |
171 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 172 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
173 | return mem_ref; |
174 | } | |
175 | ||
ebfd146a IR |
176 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
177 | ||
178 | /* Function vect_mark_relevant. | |
179 | ||
180 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
181 | ||
182 | static void | |
9771b263 | 183 | vect_mark_relevant (vec<gimple> *worklist, gimple stmt, |
83197f37 IR |
184 | enum vect_relevant relevant, bool live_p, |
185 | bool used_in_pattern) | |
ebfd146a IR |
186 | { |
187 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
188 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
189 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
83197f37 | 190 | gimple pattern_stmt; |
ebfd146a | 191 | |
73fbfcad | 192 | if (dump_enabled_p ()) |
78c60e3d SS |
193 | dump_printf_loc (MSG_NOTE, vect_location, |
194 | "mark relevant %d, live %d.", relevant, live_p); | |
ebfd146a | 195 | |
83197f37 IR |
196 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
197 | related pattern stmt instead of the original stmt. However, such stmts | |
198 | may have their own uses that are not in any pattern, in such cases the | |
199 | stmt itself should be marked. */ | |
ebfd146a IR |
200 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
201 | { | |
83197f37 IR |
202 | bool found = false; |
203 | if (!used_in_pattern) | |
204 | { | |
205 | imm_use_iterator imm_iter; | |
206 | use_operand_p use_p; | |
207 | gimple use_stmt; | |
208 | tree lhs; | |
13c931c9 JJ |
209 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
210 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a | 211 | |
83197f37 IR |
212 | if (is_gimple_assign (stmt)) |
213 | lhs = gimple_assign_lhs (stmt); | |
214 | else | |
215 | lhs = gimple_call_lhs (stmt); | |
ebfd146a | 216 | |
83197f37 IR |
217 | /* This use is out of pattern use, if LHS has other uses that are |
218 | pattern uses, we should mark the stmt itself, and not the pattern | |
219 | stmt. */ | |
ab0ef706 JJ |
220 | if (TREE_CODE (lhs) == SSA_NAME) |
221 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs) | |
222 | { | |
223 | if (is_gimple_debug (USE_STMT (use_p))) | |
224 | continue; | |
225 | use_stmt = USE_STMT (use_p); | |
226 | ||
13c931c9 JJ |
227 | if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) |
228 | continue; | |
229 | ||
ab0ef706 JJ |
230 | if (vinfo_for_stmt (use_stmt) |
231 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt))) | |
232 | { | |
233 | found = true; | |
234 | break; | |
235 | } | |
236 | } | |
83197f37 IR |
237 | } |
238 | ||
239 | if (!found) | |
240 | { | |
241 | /* This is the last stmt in a sequence that was detected as a | |
242 | pattern that can potentially be vectorized. Don't mark the stmt | |
243 | as relevant/live because it's not going to be vectorized. | |
244 | Instead mark the pattern-stmt that replaces it. */ | |
245 | ||
246 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
247 | ||
73fbfcad | 248 | if (dump_enabled_p ()) |
78c60e3d SS |
249 | dump_printf_loc (MSG_NOTE, vect_location, |
250 | "last stmt in pattern. don't mark" | |
251 | " relevant/live."); | |
83197f37 IR |
252 | stmt_info = vinfo_for_stmt (pattern_stmt); |
253 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
254 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
255 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
256 | stmt = pattern_stmt; | |
257 | } | |
ebfd146a IR |
258 | } |
259 | ||
260 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
261 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
262 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
263 | ||
264 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
265 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
266 | { | |
73fbfcad | 267 | if (dump_enabled_p ()) |
78c60e3d SS |
268 | dump_printf_loc (MSG_NOTE, vect_location, |
269 | "already marked relevant/live."); | |
ebfd146a IR |
270 | return; |
271 | } | |
272 | ||
9771b263 | 273 | worklist->safe_push (stmt); |
ebfd146a IR |
274 | } |
275 | ||
276 | ||
277 | /* Function vect_stmt_relevant_p. | |
278 | ||
279 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
280 | "relevant for vectorization". | |
281 | ||
282 | A stmt is considered "relevant for vectorization" if: | |
283 | - it has uses outside the loop. | |
284 | - it has vdefs (it alters memory). | |
285 | - control stmts in the loop (except for the exit condition). | |
286 | ||
287 | CHECKME: what other side effects would the vectorizer allow? */ | |
288 | ||
289 | static bool | |
290 | vect_stmt_relevant_p (gimple stmt, loop_vec_info loop_vinfo, | |
291 | enum vect_relevant *relevant, bool *live_p) | |
292 | { | |
293 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
294 | ssa_op_iter op_iter; | |
295 | imm_use_iterator imm_iter; | |
296 | use_operand_p use_p; | |
297 | def_operand_p def_p; | |
298 | ||
8644a673 | 299 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
300 | *live_p = false; |
301 | ||
302 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
303 | if (is_ctrl_stmt (stmt) |
304 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
305 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 306 | *relevant = vect_used_in_scope; |
ebfd146a IR |
307 | |
308 | /* changing memory. */ | |
309 | if (gimple_code (stmt) != GIMPLE_PHI) | |
5006671f | 310 | if (gimple_vdef (stmt)) |
ebfd146a | 311 | { |
73fbfcad | 312 | if (dump_enabled_p ()) |
78c60e3d SS |
313 | dump_printf_loc (MSG_NOTE, vect_location, |
314 | "vec_stmt_relevant_p: stmt has vdefs."); | |
8644a673 | 315 | *relevant = vect_used_in_scope; |
ebfd146a IR |
316 | } |
317 | ||
318 | /* uses outside the loop. */ | |
319 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
320 | { | |
321 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
322 | { | |
323 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
324 | if (!flow_bb_inside_loop_p (loop, bb)) | |
325 | { | |
73fbfcad | 326 | if (dump_enabled_p ()) |
78c60e3d SS |
327 | dump_printf_loc (MSG_NOTE, vect_location, |
328 | "vec_stmt_relevant_p: used out of loop."); | |
ebfd146a | 329 | |
3157b0c2 AO |
330 | if (is_gimple_debug (USE_STMT (use_p))) |
331 | continue; | |
332 | ||
ebfd146a IR |
333 | /* We expect all such uses to be in the loop exit phis |
334 | (because of loop closed form) */ | |
335 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
336 | gcc_assert (bb == single_exit (loop)->dest); | |
337 | ||
338 | *live_p = true; | |
339 | } | |
340 | } | |
341 | } | |
342 | ||
343 | return (*live_p || *relevant); | |
344 | } | |
345 | ||
346 | ||
b8698a0f | 347 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 348 | |
ff802fa1 | 349 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
350 | used in STMT for anything other than indexing an array. */ |
351 | ||
352 | static bool | |
353 | exist_non_indexing_operands_for_use_p (tree use, gimple stmt) | |
354 | { | |
355 | tree operand; | |
356 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 357 | |
ff802fa1 | 358 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
359 | reference in STMT, then any operand that corresponds to USE |
360 | is not indexing an array. */ | |
361 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
362 | return true; | |
59a05b0c | 363 | |
ebfd146a IR |
364 | /* STMT has a data_ref. FORNOW this means that its of one of |
365 | the following forms: | |
366 | -1- ARRAY_REF = var | |
367 | -2- var = ARRAY_REF | |
368 | (This should have been verified in analyze_data_refs). | |
369 | ||
370 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 371 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
372 | for array indexing. |
373 | ||
374 | Therefore, all we need to check is if STMT falls into the | |
375 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
376 | |
377 | if (!gimple_assign_copy_p (stmt)) | |
378 | return false; | |
59a05b0c EB |
379 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
380 | return false; | |
ebfd146a | 381 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
382 | if (TREE_CODE (operand) != SSA_NAME) |
383 | return false; | |
384 | ||
385 | if (operand == use) | |
386 | return true; | |
387 | ||
388 | return false; | |
389 | } | |
390 | ||
391 | ||
b8698a0f | 392 | /* |
ebfd146a IR |
393 | Function process_use. |
394 | ||
395 | Inputs: | |
396 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 397 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ff802fa1 | 398 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 399 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
400 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
401 | be performed. | |
ebfd146a IR |
402 | |
403 | Outputs: | |
404 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
405 | relevance info of the DEF_STMT of this USE: | |
406 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
407 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
408 | Exceptions: | |
409 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 410 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 411 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
412 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
413 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
414 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
415 | be modified accordingly. | |
416 | ||
417 | Return true if everything is as expected. Return false otherwise. */ | |
418 | ||
419 | static bool | |
b8698a0f | 420 | process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
9771b263 | 421 | enum vect_relevant relevant, vec<gimple> *worklist, |
aec7ae7d | 422 | bool force) |
ebfd146a IR |
423 | { |
424 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
425 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
426 | stmt_vec_info dstmt_vinfo; | |
427 | basic_block bb, def_bb; | |
428 | tree def; | |
429 | gimple def_stmt; | |
430 | enum vect_def_type dt; | |
431 | ||
b8698a0f | 432 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 433 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 434 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
435 | return true; |
436 | ||
24ee1384 | 437 | if (!vect_is_simple_use (use, stmt, loop_vinfo, NULL, &def_stmt, &def, &dt)) |
b8698a0f | 438 | { |
73fbfcad | 439 | if (dump_enabled_p ()) |
78c60e3d SS |
440 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
441 | "not vectorized: unsupported use in stmt."); | |
ebfd146a IR |
442 | return false; |
443 | } | |
444 | ||
445 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
446 | return true; | |
447 | ||
448 | def_bb = gimple_bb (def_stmt); | |
449 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
450 | { | |
73fbfcad | 451 | if (dump_enabled_p ()) |
78c60e3d | 452 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop."); |
ebfd146a IR |
453 | return true; |
454 | } | |
455 | ||
b8698a0f L |
456 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
457 | DEF_STMT must have already been processed, because this should be the | |
458 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
459 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
460 | check that everything is as expected, and we are done. */ |
461 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
462 | bb = gimple_bb (stmt); | |
463 | if (gimple_code (stmt) == GIMPLE_PHI | |
464 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
465 | && gimple_code (def_stmt) != GIMPLE_PHI | |
466 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
467 | && bb->loop_father == def_bb->loop_father) | |
468 | { | |
73fbfcad | 469 | if (dump_enabled_p ()) |
78c60e3d SS |
470 | dump_printf_loc (MSG_NOTE, vect_location, |
471 | "reduc-stmt defining reduc-phi in the same nest."); | |
ebfd146a IR |
472 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
473 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
474 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 475 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 476 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
477 | return true; |
478 | } | |
479 | ||
480 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
481 | outer-loop-header-bb: | |
482 | d = def_stmt | |
483 | inner-loop: | |
484 | stmt # use (d) | |
485 | outer-loop-tail-bb: | |
486 | ... */ | |
487 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
488 | { | |
73fbfcad | 489 | if (dump_enabled_p ()) |
78c60e3d SS |
490 | dump_printf_loc (MSG_NOTE, vect_location, |
491 | "outer-loop def-stmt defining inner-loop stmt."); | |
7c5222ff | 492 | |
ebfd146a IR |
493 | switch (relevant) |
494 | { | |
8644a673 | 495 | case vect_unused_in_scope: |
7c5222ff IR |
496 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
497 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 498 | break; |
7c5222ff | 499 | |
ebfd146a | 500 | case vect_used_in_outer_by_reduction: |
7c5222ff | 501 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
502 | relevant = vect_used_by_reduction; |
503 | break; | |
7c5222ff | 504 | |
ebfd146a | 505 | case vect_used_in_outer: |
7c5222ff | 506 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 507 | relevant = vect_used_in_scope; |
ebfd146a | 508 | break; |
7c5222ff | 509 | |
8644a673 | 510 | case vect_used_in_scope: |
ebfd146a IR |
511 | break; |
512 | ||
513 | default: | |
514 | gcc_unreachable (); | |
b8698a0f | 515 | } |
ebfd146a IR |
516 | } |
517 | ||
518 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
519 | outer-loop-header-bb: | |
520 | ... | |
521 | inner-loop: | |
522 | d = def_stmt | |
06066f92 | 523 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
524 | stmt # use (d) */ |
525 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
526 | { | |
73fbfcad | 527 | if (dump_enabled_p ()) |
78c60e3d SS |
528 | dump_printf_loc (MSG_NOTE, vect_location, |
529 | "inner-loop def-stmt defining outer-loop stmt."); | |
7c5222ff | 530 | |
ebfd146a IR |
531 | switch (relevant) |
532 | { | |
8644a673 | 533 | case vect_unused_in_scope: |
b8698a0f | 534 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 535 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 536 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
537 | break; |
538 | ||
ebfd146a IR |
539 | case vect_used_by_reduction: |
540 | relevant = vect_used_in_outer_by_reduction; | |
541 | break; | |
542 | ||
8644a673 | 543 | case vect_used_in_scope: |
ebfd146a IR |
544 | relevant = vect_used_in_outer; |
545 | break; | |
546 | ||
547 | default: | |
548 | gcc_unreachable (); | |
549 | } | |
550 | } | |
551 | ||
83197f37 IR |
552 | vect_mark_relevant (worklist, def_stmt, relevant, live_p, |
553 | is_pattern_stmt_p (stmt_vinfo)); | |
ebfd146a IR |
554 | return true; |
555 | } | |
556 | ||
557 | ||
558 | /* Function vect_mark_stmts_to_be_vectorized. | |
559 | ||
560 | Not all stmts in the loop need to be vectorized. For example: | |
561 | ||
562 | for i... | |
563 | for j... | |
564 | 1. T0 = i + j | |
565 | 2. T1 = a[T0] | |
566 | ||
567 | 3. j = j + 1 | |
568 | ||
569 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
570 | addressing of vectorized data-refs are handled differently. | |
571 | ||
572 | This pass detects such stmts. */ | |
573 | ||
574 | bool | |
575 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
576 | { | |
9771b263 | 577 | vec<gimple> worklist; |
ebfd146a IR |
578 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
579 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
580 | unsigned int nbbs = loop->num_nodes; | |
581 | gimple_stmt_iterator si; | |
582 | gimple stmt; | |
583 | unsigned int i; | |
584 | stmt_vec_info stmt_vinfo; | |
585 | basic_block bb; | |
586 | gimple phi; | |
587 | bool live_p; | |
06066f92 IR |
588 | enum vect_relevant relevant, tmp_relevant; |
589 | enum vect_def_type def_type; | |
ebfd146a | 590 | |
73fbfcad | 591 | if (dump_enabled_p ()) |
78c60e3d SS |
592 | dump_printf_loc (MSG_NOTE, vect_location, |
593 | "=== vect_mark_stmts_to_be_vectorized ==="); | |
ebfd146a | 594 | |
9771b263 | 595 | worklist.create (64); |
ebfd146a IR |
596 | |
597 | /* 1. Init worklist. */ | |
598 | for (i = 0; i < nbbs; i++) | |
599 | { | |
600 | bb = bbs[i]; | |
601 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 602 | { |
ebfd146a | 603 | phi = gsi_stmt (si); |
73fbfcad | 604 | if (dump_enabled_p ()) |
ebfd146a | 605 | { |
78c60e3d SS |
606 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
607 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
608 | } |
609 | ||
610 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 611 | vect_mark_relevant (&worklist, phi, relevant, live_p, false); |
ebfd146a IR |
612 | } |
613 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
614 | { | |
615 | stmt = gsi_stmt (si); | |
73fbfcad | 616 | if (dump_enabled_p ()) |
ebfd146a | 617 | { |
78c60e3d SS |
618 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
619 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 620 | } |
ebfd146a IR |
621 | |
622 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 623 | vect_mark_relevant (&worklist, stmt, relevant, live_p, false); |
ebfd146a IR |
624 | } |
625 | } | |
626 | ||
627 | /* 2. Process_worklist */ | |
9771b263 | 628 | while (worklist.length () > 0) |
ebfd146a IR |
629 | { |
630 | use_operand_p use_p; | |
631 | ssa_op_iter iter; | |
632 | ||
9771b263 | 633 | stmt = worklist.pop (); |
73fbfcad | 634 | if (dump_enabled_p ()) |
ebfd146a | 635 | { |
78c60e3d SS |
636 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
637 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
638 | } |
639 | ||
b8698a0f L |
640 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
641 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
642 | liveness and relevance properties of STMT. */ |
643 | stmt_vinfo = vinfo_for_stmt (stmt); | |
644 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
645 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
646 | ||
647 | /* Generally, the liveness and relevance properties of STMT are | |
648 | propagated as is to the DEF_STMTs of its USEs: | |
649 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
650 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
651 | ||
652 | One exception is when STMT has been identified as defining a reduction | |
653 | variable; in this case we set the liveness/relevance as follows: | |
654 | live_p = false | |
655 | relevant = vect_used_by_reduction | |
656 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f | 657 | those that are used by a reduction computation, and those that are |
ff802fa1 | 658 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 659 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 660 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 661 | |
06066f92 IR |
662 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
663 | tmp_relevant = relevant; | |
664 | switch (def_type) | |
ebfd146a | 665 | { |
06066f92 IR |
666 | case vect_reduction_def: |
667 | switch (tmp_relevant) | |
668 | { | |
669 | case vect_unused_in_scope: | |
670 | relevant = vect_used_by_reduction; | |
671 | break; | |
672 | ||
673 | case vect_used_by_reduction: | |
674 | if (gimple_code (stmt) == GIMPLE_PHI) | |
675 | break; | |
676 | /* fall through */ | |
677 | ||
678 | default: | |
73fbfcad | 679 | if (dump_enabled_p ()) |
78c60e3d SS |
680 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
681 | "unsupported use of reduction."); | |
9771b263 | 682 | worklist.release (); |
06066f92 IR |
683 | return false; |
684 | } | |
685 | ||
b8698a0f | 686 | live_p = false; |
06066f92 | 687 | break; |
b8698a0f | 688 | |
06066f92 IR |
689 | case vect_nested_cycle: |
690 | if (tmp_relevant != vect_unused_in_scope | |
691 | && tmp_relevant != vect_used_in_outer_by_reduction | |
692 | && tmp_relevant != vect_used_in_outer) | |
693 | { | |
73fbfcad | 694 | if (dump_enabled_p ()) |
78c60e3d SS |
695 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
696 | "unsupported use of nested cycle."); | |
7c5222ff | 697 | |
9771b263 | 698 | worklist.release (); |
06066f92 IR |
699 | return false; |
700 | } | |
7c5222ff | 701 | |
b8698a0f L |
702 | live_p = false; |
703 | break; | |
704 | ||
06066f92 IR |
705 | case vect_double_reduction_def: |
706 | if (tmp_relevant != vect_unused_in_scope | |
707 | && tmp_relevant != vect_used_by_reduction) | |
708 | { | |
73fbfcad | 709 | if (dump_enabled_p ()) |
78c60e3d SS |
710 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
711 | "unsupported use of double reduction."); | |
7c5222ff | 712 | |
9771b263 | 713 | worklist.release (); |
7c5222ff | 714 | return false; |
06066f92 IR |
715 | } |
716 | ||
717 | live_p = false; | |
b8698a0f | 718 | break; |
7c5222ff | 719 | |
06066f92 IR |
720 | default: |
721 | break; | |
7c5222ff | 722 | } |
b8698a0f | 723 | |
aec7ae7d | 724 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
725 | { |
726 | /* Pattern statements are not inserted into the code, so | |
727 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
728 | have to scan the RHS or function arguments instead. */ | |
729 | if (is_gimple_assign (stmt)) | |
730 | { | |
69d2aade JJ |
731 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
732 | tree op = gimple_assign_rhs1 (stmt); | |
733 | ||
734 | i = 1; | |
735 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
736 | { | |
737 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
aec7ae7d | 738 | live_p, relevant, &worklist, false) |
69d2aade | 739 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
aec7ae7d | 740 | live_p, relevant, &worklist, false)) |
69d2aade | 741 | { |
9771b263 | 742 | worklist.release (); |
69d2aade JJ |
743 | return false; |
744 | } | |
745 | i = 2; | |
746 | } | |
747 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 748 | { |
69d2aade | 749 | op = gimple_op (stmt, i); |
9d5e7640 | 750 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, |
aec7ae7d | 751 | &worklist, false)) |
9d5e7640 | 752 | { |
9771b263 | 753 | worklist.release (); |
9d5e7640 IR |
754 | return false; |
755 | } | |
756 | } | |
757 | } | |
758 | else if (is_gimple_call (stmt)) | |
759 | { | |
760 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
761 | { | |
762 | tree arg = gimple_call_arg (stmt, i); | |
763 | if (!process_use (stmt, arg, loop_vinfo, live_p, relevant, | |
aec7ae7d | 764 | &worklist, false)) |
9d5e7640 | 765 | { |
9771b263 | 766 | worklist.release (); |
9d5e7640 IR |
767 | return false; |
768 | } | |
769 | } | |
770 | } | |
771 | } | |
772 | else | |
773 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
774 | { | |
775 | tree op = USE_FROM_PTR (use_p); | |
776 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, | |
aec7ae7d | 777 | &worklist, false)) |
9d5e7640 | 778 | { |
9771b263 | 779 | worklist.release (); |
9d5e7640 IR |
780 | return false; |
781 | } | |
782 | } | |
aec7ae7d JJ |
783 | |
784 | if (STMT_VINFO_GATHER_P (stmt_vinfo)) | |
785 | { | |
786 | tree off; | |
787 | tree decl = vect_check_gather (stmt, loop_vinfo, NULL, &off, NULL); | |
788 | gcc_assert (decl); | |
789 | if (!process_use (stmt, off, loop_vinfo, live_p, relevant, | |
790 | &worklist, true)) | |
791 | { | |
9771b263 | 792 | worklist.release (); |
aec7ae7d JJ |
793 | return false; |
794 | } | |
795 | } | |
ebfd146a IR |
796 | } /* while worklist */ |
797 | ||
9771b263 | 798 | worklist.release (); |
ebfd146a IR |
799 | return true; |
800 | } | |
801 | ||
802 | ||
b8698a0f | 803 | /* Function vect_model_simple_cost. |
ebfd146a | 804 | |
b8698a0f | 805 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
806 | single op. Right now, this does not account for multiple insns that could |
807 | be generated for the single vector op. We will handle that shortly. */ | |
808 | ||
809 | void | |
b8698a0f | 810 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 BS |
811 | enum vect_def_type *dt, |
812 | stmt_vector_for_cost *prologue_cost_vec, | |
813 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
814 | { |
815 | int i; | |
92345349 | 816 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
817 | |
818 | /* The SLP costs were already calculated during SLP tree build. */ | |
819 | if (PURE_SLP_STMT (stmt_info)) | |
820 | return; | |
821 | ||
ebfd146a IR |
822 | /* FORNOW: Assuming maximum 2 args per stmts. */ |
823 | for (i = 0; i < 2; i++) | |
92345349 BS |
824 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
825 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, vector_stmt, | |
826 | stmt_info, 0, vect_prologue); | |
c3e7ee41 BS |
827 | |
828 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
829 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
830 | stmt_info, 0, vect_body); | |
c3e7ee41 | 831 | |
73fbfcad | 832 | if (dump_enabled_p ()) |
78c60e3d SS |
833 | dump_printf_loc (MSG_NOTE, vect_location, |
834 | "vect_model_simple_cost: inside_cost = %d, " | |
835 | "prologue_cost = %d .", inside_cost, prologue_cost); | |
ebfd146a IR |
836 | } |
837 | ||
838 | ||
8bd37302 BS |
839 | /* Model cost for type demotion and promotion operations. PWR is normally |
840 | zero for single-step promotions and demotions. It will be one if | |
841 | two-step promotion/demotion is required, and so on. Each additional | |
842 | step doubles the number of instructions required. */ | |
843 | ||
844 | static void | |
845 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
846 | enum vect_def_type *dt, int pwr) | |
847 | { | |
848 | int i, tmp; | |
92345349 | 849 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
850 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
851 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
852 | void *target_cost_data; | |
8bd37302 BS |
853 | |
854 | /* The SLP costs were already calculated during SLP tree build. */ | |
855 | if (PURE_SLP_STMT (stmt_info)) | |
856 | return; | |
857 | ||
c3e7ee41 BS |
858 | if (loop_vinfo) |
859 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
860 | else | |
861 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
862 | ||
8bd37302 BS |
863 | for (i = 0; i < pwr + 1; i++) |
864 | { | |
865 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
866 | (i + 1) : i; | |
c3e7ee41 | 867 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
868 | vec_promote_demote, stmt_info, 0, |
869 | vect_body); | |
8bd37302 BS |
870 | } |
871 | ||
872 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
873 | for (i = 0; i < 2; i++) | |
92345349 BS |
874 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
875 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
876 | stmt_info, 0, vect_prologue); | |
8bd37302 | 877 | |
73fbfcad | 878 | if (dump_enabled_p ()) |
78c60e3d SS |
879 | dump_printf_loc (MSG_NOTE, vect_location, |
880 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
881 | "prologue_cost = %d .", inside_cost, prologue_cost); | |
8bd37302 BS |
882 | } |
883 | ||
0d0293ac | 884 | /* Function vect_cost_group_size |
b8698a0f | 885 | |
0d0293ac | 886 | For grouped load or store, return the group_size only if it is the first |
ebfd146a IR |
887 | load or store of a group, else return 1. This ensures that group size is |
888 | only returned once per group. */ | |
889 | ||
890 | static int | |
0d0293ac | 891 | vect_cost_group_size (stmt_vec_info stmt_info) |
ebfd146a | 892 | { |
e14c1050 | 893 | gimple first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
894 | |
895 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
e14c1050 | 896 | return GROUP_SIZE (stmt_info); |
ebfd146a IR |
897 | |
898 | return 1; | |
899 | } | |
900 | ||
901 | ||
902 | /* Function vect_model_store_cost | |
903 | ||
0d0293ac MM |
904 | Models cost for stores. In the case of grouped accesses, one access |
905 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
906 | |
907 | void | |
b8698a0f | 908 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
272c6793 | 909 | bool store_lanes_p, enum vect_def_type dt, |
92345349 BS |
910 | slp_tree slp_node, |
911 | stmt_vector_for_cost *prologue_cost_vec, | |
912 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
913 | { |
914 | int group_size; | |
92345349 | 915 | unsigned int inside_cost = 0, prologue_cost = 0; |
720f5239 IR |
916 | struct data_reference *first_dr; |
917 | gimple first_stmt; | |
ebfd146a IR |
918 | |
919 | /* The SLP costs were already calculated during SLP tree build. */ | |
920 | if (PURE_SLP_STMT (stmt_info)) | |
921 | return; | |
922 | ||
8644a673 | 923 | if (dt == vect_constant_def || dt == vect_external_def) |
92345349 BS |
924 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
925 | stmt_info, 0, vect_prologue); | |
ebfd146a | 926 | |
0d0293ac MM |
927 | /* Grouped access? */ |
928 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
720f5239 IR |
929 | { |
930 | if (slp_node) | |
931 | { | |
9771b263 | 932 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
720f5239 IR |
933 | group_size = 1; |
934 | } | |
935 | else | |
936 | { | |
e14c1050 | 937 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 938 | group_size = vect_cost_group_size (stmt_info); |
720f5239 IR |
939 | } |
940 | ||
941 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
942 | } | |
0d0293ac | 943 | /* Not a grouped access. */ |
ebfd146a | 944 | else |
720f5239 IR |
945 | { |
946 | group_size = 1; | |
947 | first_dr = STMT_VINFO_DATA_REF (stmt_info); | |
948 | } | |
ebfd146a | 949 | |
272c6793 | 950 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 951 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 RS |
952 | access is instead being provided by a permute-and-store operation, |
953 | include the cost of the permutes. */ | |
954 | if (!store_lanes_p && group_size > 1) | |
ebfd146a IR |
955 | { |
956 | /* Uses a high and low interleave operation for each needed permute. */ | |
c3e7ee41 BS |
957 | |
958 | int nstmts = ncopies * exact_log2 (group_size) * group_size; | |
92345349 BS |
959 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
960 | stmt_info, 0, vect_body); | |
ebfd146a | 961 | |
73fbfcad | 962 | if (dump_enabled_p ()) |
78c60e3d SS |
963 | dump_printf_loc (MSG_NOTE, vect_location, |
964 | "vect_model_store_cost: strided group_size = %d .", | |
965 | group_size); | |
ebfd146a IR |
966 | } |
967 | ||
968 | /* Costs of the stores. */ | |
92345349 | 969 | vect_get_store_cost (first_dr, ncopies, &inside_cost, body_cost_vec); |
ebfd146a | 970 | |
73fbfcad | 971 | if (dump_enabled_p ()) |
78c60e3d SS |
972 | dump_printf_loc (MSG_NOTE, vect_location, |
973 | "vect_model_store_cost: inside_cost = %d, " | |
974 | "prologue_cost = %d .", inside_cost, prologue_cost); | |
ebfd146a IR |
975 | } |
976 | ||
977 | ||
720f5239 IR |
978 | /* Calculate cost of DR's memory access. */ |
979 | void | |
980 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 981 | unsigned int *inside_cost, |
92345349 | 982 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
983 | { |
984 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
c3e7ee41 BS |
985 | gimple stmt = DR_STMT (dr); |
986 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
720f5239 IR |
987 | |
988 | switch (alignment_support_scheme) | |
989 | { | |
990 | case dr_aligned: | |
991 | { | |
92345349 BS |
992 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
993 | vector_store, stmt_info, 0, | |
994 | vect_body); | |
720f5239 | 995 | |
73fbfcad | 996 | if (dump_enabled_p ()) |
78c60e3d SS |
997 | dump_printf_loc (MSG_NOTE, vect_location, |
998 | "vect_model_store_cost: aligned."); | |
720f5239 IR |
999 | break; |
1000 | } | |
1001 | ||
1002 | case dr_unaligned_supported: | |
1003 | { | |
720f5239 | 1004 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1005 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1006 | unaligned_store, stmt_info, |
92345349 | 1007 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1008 | if (dump_enabled_p ()) |
78c60e3d SS |
1009 | dump_printf_loc (MSG_NOTE, vect_location, |
1010 | "vect_model_store_cost: unaligned supported by " | |
1011 | "hardware."); | |
720f5239 IR |
1012 | break; |
1013 | } | |
1014 | ||
38eec4c6 UW |
1015 | case dr_unaligned_unsupported: |
1016 | { | |
1017 | *inside_cost = VECT_MAX_COST; | |
1018 | ||
73fbfcad | 1019 | if (dump_enabled_p ()) |
78c60e3d SS |
1020 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1021 | "vect_model_store_cost: unsupported access."); | |
38eec4c6 UW |
1022 | break; |
1023 | } | |
1024 | ||
720f5239 IR |
1025 | default: |
1026 | gcc_unreachable (); | |
1027 | } | |
1028 | } | |
1029 | ||
1030 | ||
ebfd146a IR |
1031 | /* Function vect_model_load_cost |
1032 | ||
0d0293ac MM |
1033 | Models cost for loads. In the case of grouped accesses, the last access |
1034 | has the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1035 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1036 | access scheme chosen. */ |
1037 | ||
1038 | void | |
92345349 BS |
1039 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
1040 | bool load_lanes_p, slp_tree slp_node, | |
1041 | stmt_vector_for_cost *prologue_cost_vec, | |
1042 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
1043 | { |
1044 | int group_size; | |
ebfd146a IR |
1045 | gimple first_stmt; |
1046 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; | |
92345349 | 1047 | unsigned int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
1048 | |
1049 | /* The SLP costs were already calculated during SLP tree build. */ | |
1050 | if (PURE_SLP_STMT (stmt_info)) | |
1051 | return; | |
1052 | ||
0d0293ac | 1053 | /* Grouped accesses? */ |
e14c1050 | 1054 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 1055 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && first_stmt && !slp_node) |
ebfd146a | 1056 | { |
0d0293ac | 1057 | group_size = vect_cost_group_size (stmt_info); |
ebfd146a IR |
1058 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
1059 | } | |
0d0293ac | 1060 | /* Not a grouped access. */ |
ebfd146a IR |
1061 | else |
1062 | { | |
1063 | group_size = 1; | |
1064 | first_dr = dr; | |
1065 | } | |
1066 | ||
272c6793 | 1067 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1068 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 RS |
1069 | access is instead being provided by a load-and-permute operation, |
1070 | include the cost of the permutes. */ | |
1071 | if (!load_lanes_p && group_size > 1) | |
ebfd146a IR |
1072 | { |
1073 | /* Uses an even and odd extract operations for each needed permute. */ | |
c3e7ee41 | 1074 | int nstmts = ncopies * exact_log2 (group_size) * group_size; |
92345349 BS |
1075 | inside_cost += record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
1076 | stmt_info, 0, vect_body); | |
ebfd146a | 1077 | |
73fbfcad | 1078 | if (dump_enabled_p ()) |
78c60e3d SS |
1079 | dump_printf_loc (MSG_NOTE, vect_location, |
1080 | "vect_model_load_cost: strided group_size = %d .", | |
1081 | group_size); | |
ebfd146a IR |
1082 | } |
1083 | ||
1084 | /* The loads themselves. */ | |
a82960aa RG |
1085 | if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
1086 | { | |
a21892ad BS |
1087 | /* N scalar loads plus gathering them into a vector. */ |
1088 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
92345349 | 1089 | inside_cost += record_stmt_cost (body_cost_vec, |
c3e7ee41 | 1090 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), |
92345349 BS |
1091 | scalar_load, stmt_info, 0, vect_body); |
1092 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, | |
1093 | stmt_info, 0, vect_body); | |
a82960aa RG |
1094 | } |
1095 | else | |
1096 | vect_get_load_cost (first_dr, ncopies, | |
1097 | ((!STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1098 | || group_size > 1 || slp_node), | |
92345349 BS |
1099 | &inside_cost, &prologue_cost, |
1100 | prologue_cost_vec, body_cost_vec, true); | |
720f5239 | 1101 | |
73fbfcad | 1102 | if (dump_enabled_p ()) |
78c60e3d SS |
1103 | dump_printf_loc (MSG_NOTE, vect_location, |
1104 | "vect_model_load_cost: inside_cost = %d, " | |
1105 | "prologue_cost = %d .", inside_cost, prologue_cost); | |
720f5239 IR |
1106 | } |
1107 | ||
1108 | ||
1109 | /* Calculate cost of DR's memory access. */ | |
1110 | void | |
1111 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1112 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1113 | unsigned int *prologue_cost, |
1114 | stmt_vector_for_cost *prologue_cost_vec, | |
1115 | stmt_vector_for_cost *body_cost_vec, | |
1116 | bool record_prologue_costs) | |
720f5239 IR |
1117 | { |
1118 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
c3e7ee41 BS |
1119 | gimple stmt = DR_STMT (dr); |
1120 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
720f5239 IR |
1121 | |
1122 | switch (alignment_support_scheme) | |
ebfd146a IR |
1123 | { |
1124 | case dr_aligned: | |
1125 | { | |
92345349 BS |
1126 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1127 | stmt_info, 0, vect_body); | |
ebfd146a | 1128 | |
73fbfcad | 1129 | if (dump_enabled_p ()) |
78c60e3d SS |
1130 | dump_printf_loc (MSG_NOTE, vect_location, |
1131 | "vect_model_load_cost: aligned."); | |
ebfd146a IR |
1132 | |
1133 | break; | |
1134 | } | |
1135 | case dr_unaligned_supported: | |
1136 | { | |
720f5239 | 1137 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1138 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1139 | unaligned_load, stmt_info, |
92345349 | 1140 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1141 | |
73fbfcad | 1142 | if (dump_enabled_p ()) |
78c60e3d SS |
1143 | dump_printf_loc (MSG_NOTE, vect_location, |
1144 | "vect_model_load_cost: unaligned supported by " | |
1145 | "hardware."); | |
ebfd146a IR |
1146 | |
1147 | break; | |
1148 | } | |
1149 | case dr_explicit_realign: | |
1150 | { | |
92345349 BS |
1151 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1152 | vector_load, stmt_info, 0, vect_body); | |
1153 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1154 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1155 | |
1156 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1157 | the containing loop, the following cost should be added to the | |
92345349 | 1158 | prologue costs. */ |
ebfd146a | 1159 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1160 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1161 | stmt_info, 0, vect_body); | |
ebfd146a | 1162 | |
73fbfcad | 1163 | if (dump_enabled_p ()) |
78c60e3d SS |
1164 | dump_printf_loc (MSG_NOTE, vect_location, |
1165 | "vect_model_load_cost: explicit realign"); | |
8bd37302 | 1166 | |
ebfd146a IR |
1167 | break; |
1168 | } | |
1169 | case dr_explicit_realign_optimized: | |
1170 | { | |
73fbfcad | 1171 | if (dump_enabled_p ()) |
78c60e3d SS |
1172 | dump_printf_loc (MSG_NOTE, vect_location, |
1173 | "vect_model_load_cost: unaligned software " | |
1174 | "pipelined."); | |
ebfd146a IR |
1175 | |
1176 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1177 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1178 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1179 | access, then the above cost should only be considered for one |
ff802fa1 | 1180 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1181 | and a realignment op. */ |
1182 | ||
92345349 | 1183 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1184 | { |
92345349 BS |
1185 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1186 | vector_stmt, stmt_info, | |
1187 | 0, vect_prologue); | |
ebfd146a | 1188 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1189 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1190 | vector_stmt, stmt_info, | |
1191 | 0, vect_prologue); | |
ebfd146a IR |
1192 | } |
1193 | ||
92345349 BS |
1194 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1195 | stmt_info, 0, vect_body); | |
1196 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1197 | stmt_info, 0, vect_body); | |
8bd37302 | 1198 | |
73fbfcad | 1199 | if (dump_enabled_p ()) |
78c60e3d SS |
1200 | dump_printf_loc (MSG_NOTE, vect_location, |
1201 | "vect_model_load_cost: explicit realign optimized"); | |
8bd37302 | 1202 | |
ebfd146a IR |
1203 | break; |
1204 | } | |
1205 | ||
38eec4c6 UW |
1206 | case dr_unaligned_unsupported: |
1207 | { | |
1208 | *inside_cost = VECT_MAX_COST; | |
1209 | ||
73fbfcad | 1210 | if (dump_enabled_p ()) |
78c60e3d SS |
1211 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1212 | "vect_model_load_cost: unsupported access."); | |
38eec4c6 UW |
1213 | break; |
1214 | } | |
1215 | ||
ebfd146a IR |
1216 | default: |
1217 | gcc_unreachable (); | |
1218 | } | |
ebfd146a IR |
1219 | } |
1220 | ||
418b7df3 RG |
1221 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1222 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1223 | |
418b7df3 RG |
1224 | static void |
1225 | vect_init_vector_1 (gimple stmt, gimple new_stmt, gimple_stmt_iterator *gsi) | |
ebfd146a | 1226 | { |
ebfd146a | 1227 | if (gsi) |
418b7df3 | 1228 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1229 | else |
1230 | { | |
418b7df3 | 1231 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1232 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1233 | |
a70d6342 IR |
1234 | if (loop_vinfo) |
1235 | { | |
1236 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1237 | basic_block new_bb; |
1238 | edge pe; | |
a70d6342 IR |
1239 | |
1240 | if (nested_in_vect_loop_p (loop, stmt)) | |
1241 | loop = loop->inner; | |
b8698a0f | 1242 | |
a70d6342 | 1243 | pe = loop_preheader_edge (loop); |
418b7df3 | 1244 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1245 | gcc_assert (!new_bb); |
1246 | } | |
1247 | else | |
1248 | { | |
1249 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1250 | basic_block bb; | |
1251 | gimple_stmt_iterator gsi_bb_start; | |
1252 | ||
1253 | gcc_assert (bb_vinfo); | |
1254 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1255 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1256 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1257 | } |
ebfd146a IR |
1258 | } |
1259 | ||
73fbfcad | 1260 | if (dump_enabled_p ()) |
ebfd146a | 1261 | { |
78c60e3d SS |
1262 | dump_printf_loc (MSG_NOTE, vect_location, |
1263 | "created new init_stmt: "); | |
1264 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1265 | } |
418b7df3 RG |
1266 | } |
1267 | ||
1268 | /* Function vect_init_vector. | |
ebfd146a | 1269 | |
5467ee52 RG |
1270 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1271 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1272 | vector type a vector with all elements equal to VAL is created first. | |
1273 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1274 | initialization at the loop preheader. | |
418b7df3 RG |
1275 | Return the DEF of INIT_STMT. |
1276 | It will be used in the vectorization of STMT. */ | |
1277 | ||
1278 | tree | |
5467ee52 | 1279 | vect_init_vector (gimple stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 RG |
1280 | { |
1281 | tree new_var; | |
1282 | gimple init_stmt; | |
1283 | tree vec_oprnd; | |
1284 | tree new_temp; | |
1285 | ||
5467ee52 RG |
1286 | if (TREE_CODE (type) == VECTOR_TYPE |
1287 | && TREE_CODE (TREE_TYPE (val)) != VECTOR_TYPE) | |
418b7df3 | 1288 | { |
5467ee52 | 1289 | if (!types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) |
418b7df3 | 1290 | { |
5467ee52 RG |
1291 | if (CONSTANT_CLASS_P (val)) |
1292 | val = fold_unary (VIEW_CONVERT_EXPR, TREE_TYPE (type), val); | |
418b7df3 RG |
1293 | else |
1294 | { | |
83d5977e | 1295 | new_temp = make_ssa_name (TREE_TYPE (type), NULL); |
418b7df3 | 1296 | init_stmt = gimple_build_assign_with_ops (NOP_EXPR, |
83d5977e | 1297 | new_temp, val, |
418b7df3 | 1298 | NULL_TREE); |
418b7df3 | 1299 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1300 | val = new_temp; |
418b7df3 RG |
1301 | } |
1302 | } | |
5467ee52 | 1303 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1304 | } |
1305 | ||
5467ee52 | 1306 | new_var = vect_get_new_vect_var (type, vect_simple_var, "cst_"); |
5467ee52 | 1307 | init_stmt = gimple_build_assign (new_var, val); |
418b7df3 RG |
1308 | new_temp = make_ssa_name (new_var, init_stmt); |
1309 | gimple_assign_set_lhs (init_stmt, new_temp); | |
1310 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
ebfd146a IR |
1311 | vec_oprnd = gimple_assign_lhs (init_stmt); |
1312 | return vec_oprnd; | |
1313 | } | |
1314 | ||
a70d6342 | 1315 | |
ebfd146a IR |
1316 | /* Function vect_get_vec_def_for_operand. |
1317 | ||
ff802fa1 | 1318 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
1319 | used in the vectorized stmt for STMT. |
1320 | ||
1321 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1322 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1323 | ||
1324 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1325 | needs to be introduced. */ | |
1326 | ||
1327 | tree | |
1328 | vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def) | |
1329 | { | |
1330 | tree vec_oprnd; | |
1331 | gimple vec_stmt; | |
1332 | gimple def_stmt; | |
1333 | stmt_vec_info def_stmt_info = NULL; | |
1334 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
9dc3f7de | 1335 | unsigned int nunits; |
ebfd146a | 1336 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
ebfd146a | 1337 | tree def; |
ebfd146a IR |
1338 | enum vect_def_type dt; |
1339 | bool is_simple_use; | |
1340 | tree vector_type; | |
1341 | ||
73fbfcad | 1342 | if (dump_enabled_p ()) |
ebfd146a | 1343 | { |
78c60e3d SS |
1344 | dump_printf_loc (MSG_NOTE, vect_location, |
1345 | "vect_get_vec_def_for_operand: "); | |
1346 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
ebfd146a IR |
1347 | } |
1348 | ||
24ee1384 IR |
1349 | is_simple_use = vect_is_simple_use (op, stmt, loop_vinfo, NULL, |
1350 | &def_stmt, &def, &dt); | |
ebfd146a | 1351 | gcc_assert (is_simple_use); |
73fbfcad | 1352 | if (dump_enabled_p ()) |
ebfd146a | 1353 | { |
78c60e3d | 1354 | int loc_printed = 0; |
ebfd146a IR |
1355 | if (def) |
1356 | { | |
78c60e3d SS |
1357 | dump_printf_loc (MSG_NOTE, vect_location, "def = "); |
1358 | loc_printed = 1; | |
1359 | dump_generic_expr (MSG_NOTE, TDF_SLIM, def); | |
ebfd146a IR |
1360 | } |
1361 | if (def_stmt) | |
1362 | { | |
78c60e3d SS |
1363 | if (loc_printed) |
1364 | dump_printf (MSG_NOTE, " def_stmt = "); | |
1365 | else | |
1366 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1367 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
ebfd146a IR |
1368 | } |
1369 | } | |
1370 | ||
1371 | switch (dt) | |
1372 | { | |
1373 | /* Case 1: operand is a constant. */ | |
1374 | case vect_constant_def: | |
1375 | { | |
7569a6cc RG |
1376 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
1377 | gcc_assert (vector_type); | |
9dc3f7de | 1378 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); |
7569a6cc | 1379 | |
b8698a0f | 1380 | if (scalar_def) |
ebfd146a IR |
1381 | *scalar_def = op; |
1382 | ||
1383 | /* Create 'vect_cst_ = {cst,cst,...,cst}' */ | |
73fbfcad | 1384 | if (dump_enabled_p ()) |
78c60e3d SS |
1385 | dump_printf_loc (MSG_NOTE, vect_location, |
1386 | "Create vector_cst. nunits = %d", nunits); | |
ebfd146a | 1387 | |
418b7df3 | 1388 | return vect_init_vector (stmt, op, vector_type, NULL); |
ebfd146a IR |
1389 | } |
1390 | ||
1391 | /* Case 2: operand is defined outside the loop - loop invariant. */ | |
8644a673 | 1392 | case vect_external_def: |
ebfd146a IR |
1393 | { |
1394 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); | |
1395 | gcc_assert (vector_type); | |
ebfd146a | 1396 | |
b8698a0f | 1397 | if (scalar_def) |
ebfd146a IR |
1398 | *scalar_def = def; |
1399 | ||
1400 | /* Create 'vec_inv = {inv,inv,..,inv}' */ | |
73fbfcad | 1401 | if (dump_enabled_p ()) |
78c60e3d | 1402 | dump_printf_loc (MSG_NOTE, vect_location, "Create vector_inv."); |
ebfd146a | 1403 | |
418b7df3 | 1404 | return vect_init_vector (stmt, def, vector_type, NULL); |
ebfd146a IR |
1405 | } |
1406 | ||
1407 | /* Case 3: operand is defined inside the loop. */ | |
8644a673 | 1408 | case vect_internal_def: |
ebfd146a | 1409 | { |
b8698a0f | 1410 | if (scalar_def) |
ebfd146a IR |
1411 | *scalar_def = NULL/* FIXME tuples: def_stmt*/; |
1412 | ||
1413 | /* Get the def from the vectorized stmt. */ | |
1414 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1415 | |
ebfd146a | 1416 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1417 | /* Get vectorized pattern statement. */ |
1418 | if (!vec_stmt | |
1419 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1420 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1421 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1422 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1423 | gcc_assert (vec_stmt); |
1424 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1425 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1426 | else if (is_gimple_call (vec_stmt)) | |
1427 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1428 | else | |
1429 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1430 | return vec_oprnd; | |
1431 | } | |
1432 | ||
1433 | /* Case 4: operand is defined by a loop header phi - reduction */ | |
1434 | case vect_reduction_def: | |
06066f92 | 1435 | case vect_double_reduction_def: |
7c5222ff | 1436 | case vect_nested_cycle: |
ebfd146a IR |
1437 | { |
1438 | struct loop *loop; | |
1439 | ||
1440 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
b8698a0f | 1441 | loop = (gimple_bb (def_stmt))->loop_father; |
ebfd146a IR |
1442 | |
1443 | /* Get the def before the loop */ | |
1444 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); | |
1445 | return get_initial_def_for_reduction (stmt, op, scalar_def); | |
1446 | } | |
1447 | ||
1448 | /* Case 5: operand is defined by loop-header phi - induction. */ | |
1449 | case vect_induction_def: | |
1450 | { | |
1451 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1452 | ||
1453 | /* Get the def from the vectorized stmt. */ | |
1454 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1455 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1456 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1457 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1458 | else | |
1459 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1460 | return vec_oprnd; |
1461 | } | |
1462 | ||
1463 | default: | |
1464 | gcc_unreachable (); | |
1465 | } | |
1466 | } | |
1467 | ||
1468 | ||
1469 | /* Function vect_get_vec_def_for_stmt_copy | |
1470 | ||
ff802fa1 | 1471 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1472 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1473 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1474 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1475 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1476 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1477 | DT is the type of the vector def VEC_OPRND. |
1478 | ||
1479 | Context: | |
1480 | In case the vectorization factor (VF) is bigger than the number | |
1481 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1482 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1483 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1484 | smallest data-type determines the VF, and as a result, when vectorizing |
1485 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1486 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1487 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1488 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1489 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1490 | ||
1491 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1492 | |
ebfd146a IR |
1493 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1494 | VS1.1: vx.1 = memref1 VS1.2 | |
1495 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1496 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1497 | |
1498 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1499 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1500 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1501 | VSnew.3: vz3 = vx.3 + ... | |
1502 | ||
1503 | The vectorization of S1 is explained in vectorizable_load. | |
1504 | The vectorization of S2: | |
b8698a0f L |
1505 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1506 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1507 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1508 | returns the vector-def 'vx.0'. |
1509 | ||
b8698a0f L |
1510 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1511 | function is called to get the relevant vector-def for each operand. It is | |
1512 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1513 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1514 | ||
b8698a0f L |
1515 | For example, to obtain the vector-def 'vx.1' in order to create the |
1516 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1517 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1518 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1519 | and return its def ('vx.1'). | |
1520 | Overall, to create the above sequence this function will be called 3 times: | |
1521 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1522 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1523 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1524 | ||
1525 | tree | |
1526 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1527 | { | |
1528 | gimple vec_stmt_for_operand; | |
1529 | stmt_vec_info def_stmt_info; | |
1530 | ||
1531 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1532 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1533 | return vec_oprnd; |
1534 | ||
1535 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1536 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1537 | gcc_assert (def_stmt_info); | |
1538 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1539 | gcc_assert (vec_stmt_for_operand); | |
1540 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1541 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) | |
1542 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1543 | else | |
1544 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1545 | return vec_oprnd; | |
1546 | } | |
1547 | ||
1548 | ||
1549 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1550 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1551 | |
1552 | static void | |
b8698a0f | 1553 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1554 | vec<tree> *vec_oprnds0, |
1555 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1556 | { |
9771b263 | 1557 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1558 | |
1559 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1560 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1561 | |
9771b263 | 1562 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1563 | { |
9771b263 | 1564 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1565 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1566 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1567 | } |
1568 | } | |
1569 | ||
1570 | ||
d092494c IR |
1571 | /* Get vectorized definitions for OP0 and OP1. |
1572 | REDUC_INDEX is the index of reduction operand in case of reduction, | |
1573 | and -1 otherwise. */ | |
ebfd146a | 1574 | |
d092494c | 1575 | void |
ebfd146a | 1576 | vect_get_vec_defs (tree op0, tree op1, gimple stmt, |
9771b263 DN |
1577 | vec<tree> *vec_oprnds0, |
1578 | vec<tree> *vec_oprnds1, | |
d092494c | 1579 | slp_tree slp_node, int reduc_index) |
ebfd146a IR |
1580 | { |
1581 | if (slp_node) | |
d092494c IR |
1582 | { |
1583 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
9771b263 DN |
1584 | vec<tree> ops; |
1585 | ops.create (nops); | |
37b5ec8f | 1586 | vec<vec<tree> > vec_defs; |
9771b263 | 1587 | vec_defs.create (nops); |
d092494c | 1588 | |
9771b263 | 1589 | ops.quick_push (op0); |
d092494c | 1590 | if (op1) |
9771b263 | 1591 | ops.quick_push (op1); |
d092494c IR |
1592 | |
1593 | vect_get_slp_defs (ops, slp_node, &vec_defs, reduc_index); | |
1594 | ||
37b5ec8f | 1595 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1596 | if (op1) |
37b5ec8f | 1597 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1598 | |
9771b263 DN |
1599 | ops.release (); |
1600 | vec_defs.release (); | |
d092494c | 1601 | } |
ebfd146a IR |
1602 | else |
1603 | { | |
1604 | tree vec_oprnd; | |
1605 | ||
9771b263 | 1606 | vec_oprnds0->create (1); |
b8698a0f | 1607 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL); |
9771b263 | 1608 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1609 | |
1610 | if (op1) | |
1611 | { | |
9771b263 | 1612 | vec_oprnds1->create (1); |
b8698a0f | 1613 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL); |
9771b263 | 1614 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1615 | } |
1616 | } | |
1617 | } | |
1618 | ||
1619 | ||
1620 | /* Function vect_finish_stmt_generation. | |
1621 | ||
1622 | Insert a new stmt. */ | |
1623 | ||
1624 | void | |
1625 | vect_finish_stmt_generation (gimple stmt, gimple vec_stmt, | |
1626 | gimple_stmt_iterator *gsi) | |
1627 | { | |
1628 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1629 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1630 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1631 | |
1632 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1633 | ||
54e8e2c3 RG |
1634 | if (!gsi_end_p (*gsi) |
1635 | && gimple_has_mem_ops (vec_stmt)) | |
1636 | { | |
1637 | gimple at_stmt = gsi_stmt (*gsi); | |
1638 | tree vuse = gimple_vuse (at_stmt); | |
1639 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1640 | { | |
1641 | tree vdef = gimple_vdef (at_stmt); | |
1642 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1643 | /* If we have an SSA vuse and insert a store, update virtual | |
1644 | SSA form to avoid triggering the renamer. Do so only | |
1645 | if we can easily see all uses - which is what almost always | |
1646 | happens with the way vectorized stmts are inserted. */ | |
1647 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1648 | && ((is_gimple_assign (vec_stmt) | |
1649 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1650 | || (is_gimple_call (vec_stmt) | |
1651 | && !(gimple_call_flags (vec_stmt) | |
1652 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1653 | { | |
1654 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1655 | gimple_set_vdef (vec_stmt, new_vdef); | |
1656 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1657 | } | |
1658 | } | |
1659 | } | |
ebfd146a IR |
1660 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1661 | ||
b8698a0f | 1662 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo, |
a70d6342 | 1663 | bb_vinfo)); |
ebfd146a | 1664 | |
73fbfcad | 1665 | if (dump_enabled_p ()) |
ebfd146a | 1666 | { |
78c60e3d SS |
1667 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1668 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
ebfd146a IR |
1669 | } |
1670 | ||
ad885386 | 1671 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
ebfd146a IR |
1672 | } |
1673 | ||
1674 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1675 | a function declaration if the target has a vectorized version | |
1676 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1677 | ||
1678 | tree | |
1679 | vectorizable_function (gimple call, tree vectype_out, tree vectype_in) | |
1680 | { | |
1681 | tree fndecl = gimple_call_fndecl (call); | |
ebfd146a IR |
1682 | |
1683 | /* We only handle functions that do not read or clobber memory -- i.e. | |
1684 | const or novops ones. */ | |
1685 | if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS))) | |
1686 | return NULL_TREE; | |
1687 | ||
1688 | if (!fndecl | |
1689 | || TREE_CODE (fndecl) != FUNCTION_DECL | |
1690 | || !DECL_BUILT_IN (fndecl)) | |
1691 | return NULL_TREE; | |
1692 | ||
62f7fd21 | 1693 | return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out, |
ebfd146a IR |
1694 | vectype_in); |
1695 | } | |
1696 | ||
1697 | /* Function vectorizable_call. | |
1698 | ||
b8698a0f L |
1699 | Check if STMT performs a function call that can be vectorized. |
1700 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1701 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1702 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1703 | ||
1704 | static bool | |
190c2236 JJ |
1705 | vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, |
1706 | slp_tree slp_node) | |
ebfd146a IR |
1707 | { |
1708 | tree vec_dest; | |
1709 | tree scalar_dest; | |
1710 | tree op, type; | |
1711 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1712 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
1713 | tree vectype_out, vectype_in; | |
1714 | int nunits_in; | |
1715 | int nunits_out; | |
1716 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 1717 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b690cc0f | 1718 | tree fndecl, new_temp, def, rhs_type; |
ebfd146a | 1719 | gimple def_stmt; |
0502fb85 UB |
1720 | enum vect_def_type dt[3] |
1721 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
63827fb8 | 1722 | gimple new_stmt = NULL; |
ebfd146a | 1723 | int ncopies, j; |
6e1aa848 | 1724 | vec<tree> vargs = vNULL; |
ebfd146a IR |
1725 | enum { NARROW, NONE, WIDEN } modifier; |
1726 | size_t i, nargs; | |
9d5e7640 | 1727 | tree lhs; |
ebfd146a | 1728 | |
190c2236 | 1729 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
1730 | return false; |
1731 | ||
8644a673 | 1732 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1733 | return false; |
1734 | ||
ebfd146a IR |
1735 | /* Is STMT a vectorizable call? */ |
1736 | if (!is_gimple_call (stmt)) | |
1737 | return false; | |
1738 | ||
1739 | if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
1740 | return false; | |
1741 | ||
822ba6d7 | 1742 | if (stmt_can_throw_internal (stmt)) |
5a2c1986 IR |
1743 | return false; |
1744 | ||
b690cc0f RG |
1745 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
1746 | ||
ebfd146a IR |
1747 | /* Process function arguments. */ |
1748 | rhs_type = NULL_TREE; | |
b690cc0f | 1749 | vectype_in = NULL_TREE; |
ebfd146a IR |
1750 | nargs = gimple_call_num_args (stmt); |
1751 | ||
1b1562a5 MM |
1752 | /* Bail out if the function has more than three arguments, we do not have |
1753 | interesting builtin functions to vectorize with more than two arguments | |
1754 | except for fma. No arguments is also not good. */ | |
1755 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
1756 | return false; |
1757 | ||
1758 | for (i = 0; i < nargs; i++) | |
1759 | { | |
b690cc0f RG |
1760 | tree opvectype; |
1761 | ||
ebfd146a IR |
1762 | op = gimple_call_arg (stmt, i); |
1763 | ||
1764 | /* We can only handle calls with arguments of the same type. */ | |
1765 | if (rhs_type | |
8533c9d8 | 1766 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 1767 | { |
73fbfcad | 1768 | if (dump_enabled_p ()) |
78c60e3d SS |
1769 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1770 | "argument types differ."); | |
ebfd146a IR |
1771 | return false; |
1772 | } | |
b690cc0f RG |
1773 | if (!rhs_type) |
1774 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 1775 | |
24ee1384 | 1776 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, |
b690cc0f | 1777 | &def_stmt, &def, &dt[i], &opvectype)) |
ebfd146a | 1778 | { |
73fbfcad | 1779 | if (dump_enabled_p ()) |
78c60e3d SS |
1780 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1781 | "use not simple."); | |
ebfd146a IR |
1782 | return false; |
1783 | } | |
ebfd146a | 1784 | |
b690cc0f RG |
1785 | if (!vectype_in) |
1786 | vectype_in = opvectype; | |
1787 | else if (opvectype | |
1788 | && opvectype != vectype_in) | |
1789 | { | |
73fbfcad | 1790 | if (dump_enabled_p ()) |
78c60e3d SS |
1791 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1792 | "argument vector types differ."); | |
b690cc0f RG |
1793 | return false; |
1794 | } | |
1795 | } | |
1796 | /* If all arguments are external or constant defs use a vector type with | |
1797 | the same size as the output vector type. */ | |
ebfd146a | 1798 | if (!vectype_in) |
b690cc0f | 1799 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
1800 | if (vec_stmt) |
1801 | gcc_assert (vectype_in); | |
1802 | if (!vectype_in) | |
1803 | { | |
73fbfcad | 1804 | if (dump_enabled_p ()) |
7d8930a0 | 1805 | { |
78c60e3d SS |
1806 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1807 | "no vectype for scalar type "); | |
1808 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
7d8930a0 IR |
1809 | } |
1810 | ||
1811 | return false; | |
1812 | } | |
ebfd146a IR |
1813 | |
1814 | /* FORNOW */ | |
b690cc0f RG |
1815 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
1816 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
1817 | if (nunits_in == nunits_out / 2) |
1818 | modifier = NARROW; | |
1819 | else if (nunits_out == nunits_in) | |
1820 | modifier = NONE; | |
1821 | else if (nunits_out == nunits_in / 2) | |
1822 | modifier = WIDEN; | |
1823 | else | |
1824 | return false; | |
1825 | ||
1826 | /* For now, we only vectorize functions if a target specific builtin | |
1827 | is available. TODO -- in some cases, it might be profitable to | |
1828 | insert the calls for pieces of the vector, in order to be able | |
1829 | to vectorize other operations in the loop. */ | |
1830 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
1831 | if (fndecl == NULL_TREE) | |
1832 | { | |
73fbfcad | 1833 | if (dump_enabled_p ()) |
78c60e3d SS |
1834 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1835 | "function is not vectorizable."); | |
ebfd146a IR |
1836 | |
1837 | return false; | |
1838 | } | |
1839 | ||
5006671f | 1840 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a | 1841 | |
190c2236 JJ |
1842 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
1843 | ncopies = 1; | |
1844 | else if (modifier == NARROW) | |
ebfd146a IR |
1845 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
1846 | else | |
1847 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1848 | ||
1849 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
1850 | needs to be generated. */ | |
1851 | gcc_assert (ncopies >= 1); | |
1852 | ||
1853 | if (!vec_stmt) /* transformation not required. */ | |
1854 | { | |
1855 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 1856 | if (dump_enabled_p ()) |
78c60e3d | 1857 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ==="); |
c3e7ee41 | 1858 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
1859 | return true; |
1860 | } | |
1861 | ||
1862 | /** Transform. **/ | |
1863 | ||
73fbfcad | 1864 | if (dump_enabled_p ()) |
78c60e3d | 1865 | dump_printf_loc (MSG_NOTE, vect_location, "transform call."); |
ebfd146a IR |
1866 | |
1867 | /* Handle def. */ | |
1868 | scalar_dest = gimple_call_lhs (stmt); | |
1869 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1870 | ||
1871 | prev_stmt_info = NULL; | |
1872 | switch (modifier) | |
1873 | { | |
1874 | case NONE: | |
1875 | for (j = 0; j < ncopies; ++j) | |
1876 | { | |
1877 | /* Build argument list for the vectorized call. */ | |
1878 | if (j == 0) | |
9771b263 | 1879 | vargs.create (nargs); |
ebfd146a | 1880 | else |
9771b263 | 1881 | vargs.truncate (0); |
ebfd146a | 1882 | |
190c2236 JJ |
1883 | if (slp_node) |
1884 | { | |
37b5ec8f | 1885 | vec<vec<tree> > vec_defs; |
9771b263 DN |
1886 | vec_defs.create (nargs); |
1887 | vec<tree> vec_oprnds0; | |
190c2236 JJ |
1888 | |
1889 | for (i = 0; i < nargs; i++) | |
9771b263 | 1890 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 1891 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 1892 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
1893 | |
1894 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 1895 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
1896 | { |
1897 | size_t k; | |
1898 | for (k = 0; k < nargs; k++) | |
1899 | { | |
37b5ec8f | 1900 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 1901 | vargs[k] = vec_oprndsk[i]; |
190c2236 JJ |
1902 | } |
1903 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1904 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1905 | gimple_call_set_lhs (new_stmt, new_temp); | |
1906 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 1907 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
1908 | } |
1909 | ||
1910 | for (i = 0; i < nargs; i++) | |
1911 | { | |
37b5ec8f | 1912 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 1913 | vec_oprndsi.release (); |
190c2236 | 1914 | } |
9771b263 | 1915 | vec_defs.release (); |
190c2236 JJ |
1916 | continue; |
1917 | } | |
1918 | ||
ebfd146a IR |
1919 | for (i = 0; i < nargs; i++) |
1920 | { | |
1921 | op = gimple_call_arg (stmt, i); | |
1922 | if (j == 0) | |
1923 | vec_oprnd0 | |
1924 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1925 | else | |
63827fb8 IR |
1926 | { |
1927 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
1928 | vec_oprnd0 | |
1929 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
1930 | } | |
ebfd146a | 1931 | |
9771b263 | 1932 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
1933 | } |
1934 | ||
1935 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1936 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1937 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
1938 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1939 | ||
1940 | if (j == 0) | |
1941 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1942 | else | |
1943 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1944 | ||
1945 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1946 | } | |
1947 | ||
1948 | break; | |
1949 | ||
1950 | case NARROW: | |
1951 | for (j = 0; j < ncopies; ++j) | |
1952 | { | |
1953 | /* Build argument list for the vectorized call. */ | |
1954 | if (j == 0) | |
9771b263 | 1955 | vargs.create (nargs * 2); |
ebfd146a | 1956 | else |
9771b263 | 1957 | vargs.truncate (0); |
ebfd146a | 1958 | |
190c2236 JJ |
1959 | if (slp_node) |
1960 | { | |
37b5ec8f | 1961 | vec<vec<tree> > vec_defs; |
9771b263 DN |
1962 | vec_defs.create (nargs); |
1963 | vec<tree> vec_oprnds0; | |
190c2236 JJ |
1964 | |
1965 | for (i = 0; i < nargs; i++) | |
9771b263 | 1966 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 1967 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 1968 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
1969 | |
1970 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 1971 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
1972 | { |
1973 | size_t k; | |
9771b263 | 1974 | vargs.truncate (0); |
190c2236 JJ |
1975 | for (k = 0; k < nargs; k++) |
1976 | { | |
37b5ec8f | 1977 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
1978 | vargs.quick_push (vec_oprndsk[i]); |
1979 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 JJ |
1980 | } |
1981 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1982 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1983 | gimple_call_set_lhs (new_stmt, new_temp); | |
1984 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 1985 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
1986 | } |
1987 | ||
1988 | for (i = 0; i < nargs; i++) | |
1989 | { | |
37b5ec8f | 1990 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 1991 | vec_oprndsi.release (); |
190c2236 | 1992 | } |
9771b263 | 1993 | vec_defs.release (); |
190c2236 JJ |
1994 | continue; |
1995 | } | |
1996 | ||
ebfd146a IR |
1997 | for (i = 0; i < nargs; i++) |
1998 | { | |
1999 | op = gimple_call_arg (stmt, i); | |
2000 | if (j == 0) | |
2001 | { | |
2002 | vec_oprnd0 | |
2003 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
2004 | vec_oprnd1 | |
63827fb8 | 2005 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2006 | } |
2007 | else | |
2008 | { | |
336ecb65 | 2009 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 2010 | vec_oprnd0 |
63827fb8 | 2011 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 2012 | vec_oprnd1 |
63827fb8 | 2013 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2014 | } |
2015 | ||
9771b263 DN |
2016 | vargs.quick_push (vec_oprnd0); |
2017 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
2018 | } |
2019 | ||
2020 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2021 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2022 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
2023 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2024 | ||
2025 | if (j == 0) | |
2026 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2027 | else | |
2028 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2029 | ||
2030 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2031 | } | |
2032 | ||
2033 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2034 | ||
2035 | break; | |
2036 | ||
2037 | case WIDEN: | |
2038 | /* No current target implements this case. */ | |
2039 | return false; | |
2040 | } | |
2041 | ||
9771b263 | 2042 | vargs.release (); |
ebfd146a IR |
2043 | |
2044 | /* Update the exception handling table with the vector stmt if necessary. */ | |
2045 | if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt)) | |
2046 | gimple_purge_dead_eh_edges (gimple_bb (stmt)); | |
2047 | ||
2048 | /* The call in STMT might prevent it from being removed in dce. | |
2049 | We however cannot remove it here, due to the way the ssa name | |
2050 | it defines is mapped to the new definition. So just replace | |
2051 | rhs of the statement with something harmless. */ | |
2052 | ||
dd34c087 JJ |
2053 | if (slp_node) |
2054 | return true; | |
2055 | ||
ebfd146a | 2056 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
2057 | if (is_pattern_stmt_p (stmt_info)) |
2058 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
2059 | else | |
2060 | lhs = gimple_call_lhs (stmt); | |
2061 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
ebfd146a | 2062 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 2063 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
2064 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
2065 | gsi_replace (gsi, new_stmt, false); | |
2066 | SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt; | |
2067 | ||
2068 | return true; | |
2069 | } | |
2070 | ||
2071 | ||
2072 | /* Function vect_gen_widened_results_half | |
2073 | ||
2074 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 2075 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 2076 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
2077 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
2078 | needs to be created (DECL is a function-decl of a target-builtin). | |
2079 | STMT is the original scalar stmt that we are vectorizing. */ | |
2080 | ||
2081 | static gimple | |
2082 | vect_gen_widened_results_half (enum tree_code code, | |
2083 | tree decl, | |
2084 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
2085 | tree vec_dest, gimple_stmt_iterator *gsi, | |
2086 | gimple stmt) | |
b8698a0f | 2087 | { |
ebfd146a | 2088 | gimple new_stmt; |
b8698a0f L |
2089 | tree new_temp; |
2090 | ||
2091 | /* Generate half of the widened result: */ | |
2092 | if (code == CALL_EXPR) | |
2093 | { | |
2094 | /* Target specific support */ | |
ebfd146a IR |
2095 | if (op_type == binary_op) |
2096 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
2097 | else | |
2098 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
2099 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2100 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
2101 | } |
2102 | else | |
ebfd146a | 2103 | { |
b8698a0f L |
2104 | /* Generic support */ |
2105 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
2106 | if (op_type != binary_op) |
2107 | vec_oprnd1 = NULL; | |
2108 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0, | |
2109 | vec_oprnd1); | |
2110 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2111 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 2112 | } |
ebfd146a IR |
2113 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2114 | ||
ebfd146a IR |
2115 | return new_stmt; |
2116 | } | |
2117 | ||
4a00c761 JJ |
2118 | |
2119 | /* Get vectorized definitions for loop-based vectorization. For the first | |
2120 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
2121 | scalar operand), and for the rest we get a copy with | |
2122 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
2123 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
2124 | The vectors are collected into VEC_OPRNDS. */ | |
2125 | ||
2126 | static void | |
2127 | vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt, | |
9771b263 | 2128 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
2129 | { |
2130 | tree vec_oprnd; | |
2131 | ||
2132 | /* Get first vector operand. */ | |
2133 | /* All the vector operands except the very first one (that is scalar oprnd) | |
2134 | are stmt copies. */ | |
2135 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
2136 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL); | |
2137 | else | |
2138 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
2139 | ||
9771b263 | 2140 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
2141 | |
2142 | /* Get second vector operand. */ | |
2143 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 2144 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
2145 | |
2146 | *oprnd = vec_oprnd; | |
2147 | ||
2148 | /* For conversion in multiple steps, continue to get operands | |
2149 | recursively. */ | |
2150 | if (multi_step_cvt) | |
2151 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
2152 | } | |
2153 | ||
2154 | ||
2155 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
2156 | For multi-step conversions store the resulting vectors and call the function | |
2157 | recursively. */ | |
2158 | ||
2159 | static void | |
9771b263 | 2160 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
4a00c761 | 2161 | int multi_step_cvt, gimple stmt, |
9771b263 | 2162 | vec<tree> vec_dsts, |
4a00c761 JJ |
2163 | gimple_stmt_iterator *gsi, |
2164 | slp_tree slp_node, enum tree_code code, | |
2165 | stmt_vec_info *prev_stmt_info) | |
2166 | { | |
2167 | unsigned int i; | |
2168 | tree vop0, vop1, new_tmp, vec_dest; | |
2169 | gimple new_stmt; | |
2170 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2171 | ||
9771b263 | 2172 | vec_dest = vec_dsts.pop (); |
4a00c761 | 2173 | |
9771b263 | 2174 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
2175 | { |
2176 | /* Create demotion operation. */ | |
9771b263 DN |
2177 | vop0 = (*vec_oprnds)[i]; |
2178 | vop1 = (*vec_oprnds)[i + 1]; | |
4a00c761 JJ |
2179 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); |
2180 | new_tmp = make_ssa_name (vec_dest, new_stmt); | |
2181 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
2182 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2183 | ||
2184 | if (multi_step_cvt) | |
2185 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 2186 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
2187 | else |
2188 | { | |
2189 | /* This is the last step of the conversion sequence. Store the | |
2190 | vectors in SLP_NODE or in vector info of the scalar statement | |
2191 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
2192 | if (slp_node) | |
9771b263 | 2193 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2194 | else |
2195 | { | |
2196 | if (!*prev_stmt_info) | |
2197 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2198 | else | |
2199 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
2200 | ||
2201 | *prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2202 | } | |
2203 | } | |
2204 | } | |
2205 | ||
2206 | /* For multi-step demotion operations we first generate demotion operations | |
2207 | from the source type to the intermediate types, and then combine the | |
2208 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
2209 | type. */ | |
2210 | if (multi_step_cvt) | |
2211 | { | |
2212 | /* At each level of recursion we have half of the operands we had at the | |
2213 | previous level. */ | |
9771b263 | 2214 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
2215 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
2216 | stmt, vec_dsts, gsi, slp_node, | |
2217 | VEC_PACK_TRUNC_EXPR, | |
2218 | prev_stmt_info); | |
2219 | } | |
2220 | ||
9771b263 | 2221 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2222 | } |
2223 | ||
2224 | ||
2225 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
2226 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
2227 | the resulting vectors and call the function recursively. */ | |
2228 | ||
2229 | static void | |
9771b263 DN |
2230 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
2231 | vec<tree> *vec_oprnds1, | |
4a00c761 JJ |
2232 | gimple stmt, tree vec_dest, |
2233 | gimple_stmt_iterator *gsi, | |
2234 | enum tree_code code1, | |
2235 | enum tree_code code2, tree decl1, | |
2236 | tree decl2, int op_type) | |
2237 | { | |
2238 | int i; | |
2239 | tree vop0, vop1, new_tmp1, new_tmp2; | |
2240 | gimple new_stmt1, new_stmt2; | |
6e1aa848 | 2241 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 2242 | |
9771b263 DN |
2243 | vec_tmp.create (vec_oprnds0->length () * 2); |
2244 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
2245 | { |
2246 | if (op_type == binary_op) | |
9771b263 | 2247 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
2248 | else |
2249 | vop1 = NULL_TREE; | |
2250 | ||
2251 | /* Generate the two halves of promotion operation. */ | |
2252 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
2253 | op_type, vec_dest, gsi, stmt); | |
2254 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
2255 | op_type, vec_dest, gsi, stmt); | |
2256 | if (is_gimple_call (new_stmt1)) | |
2257 | { | |
2258 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
2259 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
2260 | } | |
2261 | else | |
2262 | { | |
2263 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
2264 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
2265 | } | |
2266 | ||
2267 | /* Store the results for the next step. */ | |
9771b263 DN |
2268 | vec_tmp.quick_push (new_tmp1); |
2269 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
2270 | } |
2271 | ||
689eaba3 | 2272 | vec_oprnds0->release (); |
4a00c761 JJ |
2273 | *vec_oprnds0 = vec_tmp; |
2274 | } | |
2275 | ||
2276 | ||
b8698a0f L |
2277 | /* Check if STMT performs a conversion operation, that can be vectorized. |
2278 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 2279 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
2280 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
2281 | ||
2282 | static bool | |
2283 | vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi, | |
2284 | gimple *vec_stmt, slp_tree slp_node) | |
2285 | { | |
2286 | tree vec_dest; | |
2287 | tree scalar_dest; | |
4a00c761 | 2288 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
2289 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
2290 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2291 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2292 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 2293 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
2294 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
2295 | tree new_temp; | |
2296 | tree def; | |
2297 | gimple def_stmt; | |
2298 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2299 | gimple new_stmt = NULL; | |
2300 | stmt_vec_info prev_stmt_info; | |
2301 | int nunits_in; | |
2302 | int nunits_out; | |
2303 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
2304 | int ncopies, i, j; |
2305 | tree lhs_type, rhs_type; | |
ebfd146a | 2306 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
2307 | vec<tree> vec_oprnds0 = vNULL; |
2308 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 2309 | tree vop0; |
4a00c761 JJ |
2310 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2311 | int multi_step_cvt = 0; | |
6e1aa848 DN |
2312 | vec<tree> vec_dsts = vNULL; |
2313 | vec<tree> interm_types = vNULL; | |
4a00c761 JJ |
2314 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
2315 | int op_type; | |
2316 | enum machine_mode rhs_mode; | |
2317 | unsigned short fltsz; | |
ebfd146a IR |
2318 | |
2319 | /* Is STMT a vectorizable conversion? */ | |
2320 | ||
4a00c761 | 2321 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2322 | return false; |
2323 | ||
8644a673 | 2324 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2325 | return false; |
2326 | ||
2327 | if (!is_gimple_assign (stmt)) | |
2328 | return false; | |
2329 | ||
2330 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2331 | return false; | |
2332 | ||
2333 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
2334 | if (!CONVERT_EXPR_CODE_P (code) |
2335 | && code != FIX_TRUNC_EXPR | |
2336 | && code != FLOAT_EXPR | |
2337 | && code != WIDEN_MULT_EXPR | |
2338 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
2339 | return false; |
2340 | ||
4a00c761 JJ |
2341 | op_type = TREE_CODE_LENGTH (code); |
2342 | ||
ebfd146a | 2343 | /* Check types of lhs and rhs. */ |
b690cc0f | 2344 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 2345 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
2346 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2347 | ||
ebfd146a IR |
2348 | op0 = gimple_assign_rhs1 (stmt); |
2349 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
2350 | |
2351 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
2352 | && !((INTEGRAL_TYPE_P (lhs_type) | |
2353 | && INTEGRAL_TYPE_P (rhs_type)) | |
2354 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
2355 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
2356 | return false; | |
2357 | ||
2358 | if ((INTEGRAL_TYPE_P (lhs_type) | |
2359 | && (TYPE_PRECISION (lhs_type) | |
2360 | != GET_MODE_PRECISION (TYPE_MODE (lhs_type)))) | |
2361 | || (INTEGRAL_TYPE_P (rhs_type) | |
2362 | && (TYPE_PRECISION (rhs_type) | |
2363 | != GET_MODE_PRECISION (TYPE_MODE (rhs_type))))) | |
2364 | { | |
73fbfcad | 2365 | if (dump_enabled_p ()) |
78c60e3d SS |
2366 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2367 | "type conversion to/from bit-precision unsupported."); | |
4a00c761 JJ |
2368 | return false; |
2369 | } | |
2370 | ||
b690cc0f | 2371 | /* Check the operands of the operation. */ |
24ee1384 | 2372 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
b690cc0f RG |
2373 | &def_stmt, &def, &dt[0], &vectype_in)) |
2374 | { | |
73fbfcad | 2375 | if (dump_enabled_p ()) |
78c60e3d SS |
2376 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2377 | "use not simple."); | |
b690cc0f RG |
2378 | return false; |
2379 | } | |
4a00c761 JJ |
2380 | if (op_type == binary_op) |
2381 | { | |
2382 | bool ok; | |
2383 | ||
2384 | op1 = gimple_assign_rhs2 (stmt); | |
2385 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
2386 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
2387 | OP1. */ | |
2388 | if (CONSTANT_CLASS_P (op0)) | |
f5709183 | 2389 | ok = vect_is_simple_use_1 (op1, stmt, loop_vinfo, bb_vinfo, |
4a00c761 JJ |
2390 | &def_stmt, &def, &dt[1], &vectype_in); |
2391 | else | |
f5709183 | 2392 | ok = vect_is_simple_use (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
24ee1384 | 2393 | &def, &dt[1]); |
4a00c761 JJ |
2394 | |
2395 | if (!ok) | |
2396 | { | |
73fbfcad | 2397 | if (dump_enabled_p ()) |
78c60e3d SS |
2398 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2399 | "use not simple."); | |
4a00c761 JJ |
2400 | return false; |
2401 | } | |
2402 | } | |
2403 | ||
b690cc0f RG |
2404 | /* If op0 is an external or constant defs use a vector type of |
2405 | the same size as the output vector type. */ | |
ebfd146a | 2406 | if (!vectype_in) |
b690cc0f | 2407 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2408 | if (vec_stmt) |
2409 | gcc_assert (vectype_in); | |
2410 | if (!vectype_in) | |
2411 | { | |
73fbfcad | 2412 | if (dump_enabled_p ()) |
4a00c761 | 2413 | { |
78c60e3d SS |
2414 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2415 | "no vectype for scalar type "); | |
2416 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
4a00c761 | 2417 | } |
7d8930a0 IR |
2418 | |
2419 | return false; | |
2420 | } | |
ebfd146a | 2421 | |
b690cc0f RG |
2422 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2423 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 2424 | if (nunits_in < nunits_out) |
ebfd146a IR |
2425 | modifier = NARROW; |
2426 | else if (nunits_out == nunits_in) | |
2427 | modifier = NONE; | |
ebfd146a | 2428 | else |
4a00c761 | 2429 | modifier = WIDEN; |
ebfd146a | 2430 | |
ff802fa1 IR |
2431 | /* Multiple types in SLP are handled by creating the appropriate number of |
2432 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2433 | case of SLP. */ | |
437f4a00 | 2434 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a | 2435 | ncopies = 1; |
4a00c761 JJ |
2436 | else if (modifier == NARROW) |
2437 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
2438 | else | |
2439 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 2440 | |
ebfd146a IR |
2441 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
2442 | needs to be generated. */ | |
2443 | gcc_assert (ncopies >= 1); | |
2444 | ||
ebfd146a | 2445 | /* Supportable by target? */ |
4a00c761 | 2446 | switch (modifier) |
ebfd146a | 2447 | { |
4a00c761 JJ |
2448 | case NONE: |
2449 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
2450 | return false; | |
2451 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
2452 | &decl1, &code1)) | |
2453 | break; | |
2454 | /* FALLTHRU */ | |
2455 | unsupported: | |
73fbfcad | 2456 | if (dump_enabled_p ()) |
78c60e3d SS |
2457 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2458 | "conversion not supported by target."); | |
ebfd146a | 2459 | return false; |
ebfd146a | 2460 | |
4a00c761 JJ |
2461 | case WIDEN: |
2462 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
2463 | &code1, &code2, &multi_step_cvt, |
2464 | &interm_types)) | |
4a00c761 JJ |
2465 | { |
2466 | /* Binary widening operation can only be supported directly by the | |
2467 | architecture. */ | |
2468 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
2469 | break; | |
2470 | } | |
2471 | ||
2472 | if (code != FLOAT_EXPR | |
2473 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
2474 | <= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
2475 | goto unsupported; | |
2476 | ||
2477 | rhs_mode = TYPE_MODE (rhs_type); | |
2478 | fltsz = GET_MODE_SIZE (TYPE_MODE (lhs_type)); | |
2479 | for (rhs_mode = GET_MODE_2XWIDER_MODE (TYPE_MODE (rhs_type)); | |
2480 | rhs_mode != VOIDmode && GET_MODE_SIZE (rhs_mode) <= fltsz; | |
2481 | rhs_mode = GET_MODE_2XWIDER_MODE (rhs_mode)) | |
2482 | { | |
2483 | cvt_type | |
2484 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
2485 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
2486 | if (cvt_type == NULL_TREE) | |
2487 | goto unsupported; | |
2488 | ||
2489 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
2490 | { | |
2491 | if (!supportable_convert_operation (code, vectype_out, | |
2492 | cvt_type, &decl1, &codecvt1)) | |
2493 | goto unsupported; | |
2494 | } | |
2495 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
2496 | cvt_type, &codecvt1, |
2497 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
2498 | &interm_types)) |
2499 | continue; | |
2500 | else | |
2501 | gcc_assert (multi_step_cvt == 0); | |
2502 | ||
2503 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
2504 | vectype_in, &code1, &code2, |
2505 | &multi_step_cvt, &interm_types)) | |
4a00c761 JJ |
2506 | break; |
2507 | } | |
2508 | ||
2509 | if (rhs_mode == VOIDmode || GET_MODE_SIZE (rhs_mode) > fltsz) | |
2510 | goto unsupported; | |
2511 | ||
2512 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
2513 | codecvt2 = ERROR_MARK; | |
2514 | else | |
2515 | { | |
2516 | multi_step_cvt++; | |
9771b263 | 2517 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
2518 | cvt_type = NULL_TREE; |
2519 | } | |
2520 | break; | |
2521 | ||
2522 | case NARROW: | |
2523 | gcc_assert (op_type == unary_op); | |
2524 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
2525 | &code1, &multi_step_cvt, | |
2526 | &interm_types)) | |
2527 | break; | |
2528 | ||
2529 | if (code != FIX_TRUNC_EXPR | |
2530 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
2531 | >= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
2532 | goto unsupported; | |
2533 | ||
2534 | rhs_mode = TYPE_MODE (rhs_type); | |
2535 | cvt_type | |
2536 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
2537 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
2538 | if (cvt_type == NULL_TREE) | |
2539 | goto unsupported; | |
2540 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
2541 | &decl1, &codecvt1)) | |
2542 | goto unsupported; | |
2543 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
2544 | &code1, &multi_step_cvt, | |
2545 | &interm_types)) | |
2546 | break; | |
2547 | goto unsupported; | |
2548 | ||
2549 | default: | |
2550 | gcc_unreachable (); | |
ebfd146a IR |
2551 | } |
2552 | ||
2553 | if (!vec_stmt) /* transformation not required. */ | |
2554 | { | |
73fbfcad | 2555 | if (dump_enabled_p ()) |
78c60e3d SS |
2556 | dump_printf_loc (MSG_NOTE, vect_location, |
2557 | "=== vectorizable_conversion ==="); | |
4a00c761 | 2558 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
2559 | { |
2560 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
c3e7ee41 | 2561 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
8bd37302 | 2562 | } |
4a00c761 JJ |
2563 | else if (modifier == NARROW) |
2564 | { | |
2565 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 2566 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
2567 | } |
2568 | else | |
2569 | { | |
2570 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 2571 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 2572 | } |
9771b263 | 2573 | interm_types.release (); |
ebfd146a IR |
2574 | return true; |
2575 | } | |
2576 | ||
2577 | /** Transform. **/ | |
73fbfcad | 2578 | if (dump_enabled_p ()) |
78c60e3d SS |
2579 | dump_printf_loc (MSG_NOTE, vect_location, |
2580 | "transform conversion. ncopies = %d.", ncopies); | |
ebfd146a | 2581 | |
4a00c761 JJ |
2582 | if (op_type == binary_op) |
2583 | { | |
2584 | if (CONSTANT_CLASS_P (op0)) | |
2585 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
2586 | else if (CONSTANT_CLASS_P (op1)) | |
2587 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
2588 | } | |
2589 | ||
2590 | /* In case of multi-step conversion, we first generate conversion operations | |
2591 | to the intermediate types, and then from that types to the final one. | |
2592 | We create vector destinations for the intermediate type (TYPES) received | |
2593 | from supportable_*_operation, and store them in the correct order | |
2594 | for future use in vect_create_vectorized_*_stmts (). */ | |
9771b263 | 2595 | vec_dsts.create (multi_step_cvt + 1); |
82294ec1 JJ |
2596 | vec_dest = vect_create_destination_var (scalar_dest, |
2597 | (cvt_type && modifier == WIDEN) | |
2598 | ? cvt_type : vectype_out); | |
9771b263 | 2599 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2600 | |
2601 | if (multi_step_cvt) | |
2602 | { | |
9771b263 DN |
2603 | for (i = interm_types.length () - 1; |
2604 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
2605 | { |
2606 | vec_dest = vect_create_destination_var (scalar_dest, | |
2607 | intermediate_type); | |
9771b263 | 2608 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2609 | } |
2610 | } | |
ebfd146a | 2611 | |
4a00c761 | 2612 | if (cvt_type) |
82294ec1 JJ |
2613 | vec_dest = vect_create_destination_var (scalar_dest, |
2614 | modifier == WIDEN | |
2615 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
2616 | |
2617 | if (!slp_node) | |
2618 | { | |
30862efc | 2619 | if (modifier == WIDEN) |
4a00c761 | 2620 | { |
9771b263 | 2621 | vec_oprnds0.create (multi_step_cvt ? vect_pow2(multi_step_cvt) : 1); |
4a00c761 | 2622 | if (op_type == binary_op) |
9771b263 | 2623 | vec_oprnds1.create (1); |
4a00c761 | 2624 | } |
30862efc | 2625 | else if (modifier == NARROW) |
9771b263 DN |
2626 | vec_oprnds0.create ( |
2627 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
2628 | } |
2629 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 2630 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 2631 | |
4a00c761 | 2632 | last_oprnd = op0; |
ebfd146a IR |
2633 | prev_stmt_info = NULL; |
2634 | switch (modifier) | |
2635 | { | |
2636 | case NONE: | |
2637 | for (j = 0; j < ncopies; j++) | |
2638 | { | |
ebfd146a | 2639 | if (j == 0) |
d092494c IR |
2640 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node, |
2641 | -1); | |
ebfd146a IR |
2642 | else |
2643 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
2644 | ||
9771b263 | 2645 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2646 | { |
2647 | /* Arguments are ready, create the new vector stmt. */ | |
2648 | if (code1 == CALL_EXPR) | |
2649 | { | |
2650 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2651 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2652 | gimple_call_set_lhs (new_stmt, new_temp); | |
2653 | } | |
2654 | else | |
2655 | { | |
2656 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
2657 | new_stmt = gimple_build_assign_with_ops (code1, vec_dest, | |
2658 | vop0, NULL); | |
2659 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2660 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2661 | } | |
2662 | ||
2663 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2664 | if (slp_node) | |
9771b263 | 2665 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2666 | } |
2667 | ||
ebfd146a IR |
2668 | if (j == 0) |
2669 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2670 | else | |
2671 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2672 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2673 | } | |
2674 | break; | |
2675 | ||
2676 | case WIDEN: | |
2677 | /* In case the vectorization factor (VF) is bigger than the number | |
2678 | of elements that we can fit in a vectype (nunits), we have to | |
2679 | generate more than one vector stmt - i.e - we need to "unroll" | |
2680 | the vector stmt by a factor VF/nunits. */ | |
2681 | for (j = 0; j < ncopies; j++) | |
2682 | { | |
4a00c761 | 2683 | /* Handle uses. */ |
ebfd146a | 2684 | if (j == 0) |
4a00c761 JJ |
2685 | { |
2686 | if (slp_node) | |
2687 | { | |
2688 | if (code == WIDEN_LSHIFT_EXPR) | |
2689 | { | |
2690 | unsigned int k; | |
ebfd146a | 2691 | |
4a00c761 JJ |
2692 | vec_oprnd1 = op1; |
2693 | /* Store vec_oprnd1 for every vector stmt to be created | |
2694 | for SLP_NODE. We check during the analysis that all | |
2695 | the shift arguments are the same. */ | |
2696 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 2697 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
2698 | |
2699 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2700 | slp_node, -1); | |
2701 | } | |
2702 | else | |
2703 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
2704 | &vec_oprnds1, slp_node, -1); | |
2705 | } | |
2706 | else | |
2707 | { | |
2708 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
9771b263 | 2709 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
2710 | if (op_type == binary_op) |
2711 | { | |
2712 | if (code == WIDEN_LSHIFT_EXPR) | |
2713 | vec_oprnd1 = op1; | |
2714 | else | |
2715 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, | |
2716 | NULL); | |
9771b263 | 2717 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
2718 | } |
2719 | } | |
2720 | } | |
ebfd146a | 2721 | else |
4a00c761 JJ |
2722 | { |
2723 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
2724 | vec_oprnds0.truncate (0); |
2725 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
2726 | if (op_type == binary_op) |
2727 | { | |
2728 | if (code == WIDEN_LSHIFT_EXPR) | |
2729 | vec_oprnd1 = op1; | |
2730 | else | |
2731 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
2732 | vec_oprnd1); | |
9771b263 DN |
2733 | vec_oprnds1.truncate (0); |
2734 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
2735 | } |
2736 | } | |
ebfd146a | 2737 | |
4a00c761 JJ |
2738 | /* Arguments are ready. Create the new vector stmts. */ |
2739 | for (i = multi_step_cvt; i >= 0; i--) | |
2740 | { | |
9771b263 | 2741 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
2742 | enum tree_code c1 = code1, c2 = code2; |
2743 | if (i == 0 && codecvt2 != ERROR_MARK) | |
2744 | { | |
2745 | c1 = codecvt1; | |
2746 | c2 = codecvt2; | |
2747 | } | |
2748 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
2749 | &vec_oprnds1, | |
2750 | stmt, this_dest, gsi, | |
2751 | c1, c2, decl1, decl2, | |
2752 | op_type); | |
2753 | } | |
2754 | ||
9771b263 | 2755 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2756 | { |
2757 | if (cvt_type) | |
2758 | { | |
2759 | if (codecvt1 == CALL_EXPR) | |
2760 | { | |
2761 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2762 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2763 | gimple_call_set_lhs (new_stmt, new_temp); | |
2764 | } | |
2765 | else | |
2766 | { | |
2767 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
2768 | new_temp = make_ssa_name (vec_dest, NULL); | |
2769 | new_stmt = gimple_build_assign_with_ops (codecvt1, | |
2770 | new_temp, | |
2771 | vop0, NULL); | |
2772 | } | |
2773 | ||
2774 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2775 | } | |
2776 | else | |
2777 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
2778 | ||
2779 | if (slp_node) | |
9771b263 | 2780 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2781 | else |
2782 | { | |
2783 | if (!prev_stmt_info) | |
2784 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2785 | else | |
2786 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2787 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2788 | } | |
2789 | } | |
ebfd146a | 2790 | } |
4a00c761 JJ |
2791 | |
2792 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
2793 | break; |
2794 | ||
2795 | case NARROW: | |
2796 | /* In case the vectorization factor (VF) is bigger than the number | |
2797 | of elements that we can fit in a vectype (nunits), we have to | |
2798 | generate more than one vector stmt - i.e - we need to "unroll" | |
2799 | the vector stmt by a factor VF/nunits. */ | |
2800 | for (j = 0; j < ncopies; j++) | |
2801 | { | |
2802 | /* Handle uses. */ | |
4a00c761 JJ |
2803 | if (slp_node) |
2804 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2805 | slp_node, -1); | |
ebfd146a IR |
2806 | else |
2807 | { | |
9771b263 | 2808 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
2809 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
2810 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
2811 | } |
2812 | ||
4a00c761 JJ |
2813 | /* Arguments are ready. Create the new vector stmts. */ |
2814 | if (cvt_type) | |
9771b263 | 2815 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2816 | { |
2817 | if (codecvt1 == CALL_EXPR) | |
2818 | { | |
2819 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2820 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2821 | gimple_call_set_lhs (new_stmt, new_temp); | |
2822 | } | |
2823 | else | |
2824 | { | |
2825 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
2826 | new_temp = make_ssa_name (vec_dest, NULL); | |
2827 | new_stmt = gimple_build_assign_with_ops (codecvt1, new_temp, | |
2828 | vop0, NULL); | |
2829 | } | |
ebfd146a | 2830 | |
4a00c761 | 2831 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 2832 | vec_oprnds0[i] = new_temp; |
4a00c761 | 2833 | } |
ebfd146a | 2834 | |
4a00c761 JJ |
2835 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
2836 | stmt, vec_dsts, gsi, | |
2837 | slp_node, code1, | |
2838 | &prev_stmt_info); | |
ebfd146a IR |
2839 | } |
2840 | ||
2841 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 2842 | break; |
ebfd146a IR |
2843 | } |
2844 | ||
9771b263 DN |
2845 | vec_oprnds0.release (); |
2846 | vec_oprnds1.release (); | |
2847 | vec_dsts.release (); | |
2848 | interm_types.release (); | |
ebfd146a IR |
2849 | |
2850 | return true; | |
2851 | } | |
ff802fa1 IR |
2852 | |
2853 | ||
ebfd146a IR |
2854 | /* Function vectorizable_assignment. |
2855 | ||
b8698a0f L |
2856 | Check if STMT performs an assignment (copy) that can be vectorized. |
2857 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
2858 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2859 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2860 | ||
2861 | static bool | |
2862 | vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi, | |
2863 | gimple *vec_stmt, slp_tree slp_node) | |
2864 | { | |
2865 | tree vec_dest; | |
2866 | tree scalar_dest; | |
2867 | tree op; | |
2868 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2869 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2870 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2871 | tree new_temp; | |
2872 | tree def; | |
2873 | gimple def_stmt; | |
2874 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
fde9c428 | 2875 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
ebfd146a | 2876 | int ncopies; |
f18b55bd | 2877 | int i, j; |
6e1aa848 | 2878 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 2879 | tree vop; |
a70d6342 | 2880 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
f18b55bd IR |
2881 | gimple new_stmt = NULL; |
2882 | stmt_vec_info prev_stmt_info = NULL; | |
fde9c428 RG |
2883 | enum tree_code code; |
2884 | tree vectype_in; | |
ebfd146a IR |
2885 | |
2886 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2887 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2888 | case of SLP. */ | |
437f4a00 | 2889 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
2890 | ncopies = 1; |
2891 | else | |
2892 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2893 | ||
2894 | gcc_assert (ncopies >= 1); | |
ebfd146a | 2895 | |
a70d6342 | 2896 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2897 | return false; |
2898 | ||
8644a673 | 2899 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2900 | return false; |
2901 | ||
2902 | /* Is vectorizable assignment? */ | |
2903 | if (!is_gimple_assign (stmt)) | |
2904 | return false; | |
2905 | ||
2906 | scalar_dest = gimple_assign_lhs (stmt); | |
2907 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
2908 | return false; | |
2909 | ||
fde9c428 | 2910 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 2911 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
2912 | || code == PAREN_EXPR |
2913 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
2914 | op = gimple_assign_rhs1 (stmt); |
2915 | else | |
2916 | return false; | |
2917 | ||
7b7ec6c5 RG |
2918 | if (code == VIEW_CONVERT_EXPR) |
2919 | op = TREE_OPERAND (op, 0); | |
2920 | ||
24ee1384 | 2921 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, |
fde9c428 | 2922 | &def_stmt, &def, &dt[0], &vectype_in)) |
ebfd146a | 2923 | { |
73fbfcad | 2924 | if (dump_enabled_p ()) |
78c60e3d SS |
2925 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2926 | "use not simple."); | |
ebfd146a IR |
2927 | return false; |
2928 | } | |
2929 | ||
fde9c428 RG |
2930 | /* We can handle NOP_EXPR conversions that do not change the number |
2931 | of elements or the vector size. */ | |
7b7ec6c5 RG |
2932 | if ((CONVERT_EXPR_CODE_P (code) |
2933 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
2934 | && (!vectype_in |
2935 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
2936 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
2937 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
2938 | return false; | |
2939 | ||
7b7b1813 RG |
2940 | /* We do not handle bit-precision changes. */ |
2941 | if ((CONVERT_EXPR_CODE_P (code) | |
2942 | || code == VIEW_CONVERT_EXPR) | |
2943 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2944 | && ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
2945 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
2946 | || ((TYPE_PRECISION (TREE_TYPE (op)) | |
2947 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))) | |
2948 | /* But a conversion that does not change the bit-pattern is ok. */ | |
2949 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
2950 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2951 | && TYPE_UNSIGNED (TREE_TYPE (op)))) | |
2952 | { | |
73fbfcad | 2953 | if (dump_enabled_p ()) |
78c60e3d SS |
2954 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2955 | "type conversion to/from bit-precision " | |
2956 | "unsupported."); | |
7b7b1813 RG |
2957 | return false; |
2958 | } | |
2959 | ||
ebfd146a IR |
2960 | if (!vec_stmt) /* transformation not required. */ |
2961 | { | |
2962 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 2963 | if (dump_enabled_p ()) |
78c60e3d SS |
2964 | dump_printf_loc (MSG_NOTE, vect_location, |
2965 | "=== vectorizable_assignment ==="); | |
c3e7ee41 | 2966 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
2967 | return true; |
2968 | } | |
2969 | ||
2970 | /** Transform. **/ | |
73fbfcad | 2971 | if (dump_enabled_p ()) |
78c60e3d | 2972 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment."); |
ebfd146a IR |
2973 | |
2974 | /* Handle def. */ | |
2975 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2976 | ||
2977 | /* Handle use. */ | |
f18b55bd | 2978 | for (j = 0; j < ncopies; j++) |
ebfd146a | 2979 | { |
f18b55bd IR |
2980 | /* Handle uses. */ |
2981 | if (j == 0) | |
d092494c | 2982 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node, -1); |
f18b55bd IR |
2983 | else |
2984 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
2985 | ||
2986 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 2987 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 2988 | { |
7b7ec6c5 RG |
2989 | if (CONVERT_EXPR_CODE_P (code) |
2990 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 2991 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
2992 | new_stmt = gimple_build_assign (vec_dest, vop); |
2993 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2994 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2995 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2996 | if (slp_node) | |
9771b263 | 2997 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 2998 | } |
ebfd146a IR |
2999 | |
3000 | if (slp_node) | |
f18b55bd IR |
3001 | continue; |
3002 | ||
3003 | if (j == 0) | |
3004 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3005 | else | |
3006 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3007 | ||
3008 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3009 | } | |
b8698a0f | 3010 | |
9771b263 | 3011 | vec_oprnds.release (); |
ebfd146a IR |
3012 | return true; |
3013 | } | |
3014 | ||
9dc3f7de | 3015 | |
1107f3ae IR |
3016 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
3017 | either as shift by a scalar or by a vector. */ | |
3018 | ||
3019 | bool | |
3020 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
3021 | { | |
3022 | ||
3023 | enum machine_mode vec_mode; | |
3024 | optab optab; | |
3025 | int icode; | |
3026 | tree vectype; | |
3027 | ||
3028 | vectype = get_vectype_for_scalar_type (scalar_type); | |
3029 | if (!vectype) | |
3030 | return false; | |
3031 | ||
3032 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
3033 | if (!optab | |
3034 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
3035 | { | |
3036 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
3037 | if (!optab | |
3038 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
3039 | == CODE_FOR_nothing)) | |
3040 | return false; | |
3041 | } | |
3042 | ||
3043 | vec_mode = TYPE_MODE (vectype); | |
3044 | icode = (int) optab_handler (optab, vec_mode); | |
3045 | if (icode == CODE_FOR_nothing) | |
3046 | return false; | |
3047 | ||
3048 | return true; | |
3049 | } | |
3050 | ||
3051 | ||
9dc3f7de IR |
3052 | /* Function vectorizable_shift. |
3053 | ||
3054 | Check if STMT performs a shift operation that can be vectorized. | |
3055 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3056 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3057 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3058 | ||
3059 | static bool | |
3060 | vectorizable_shift (gimple stmt, gimple_stmt_iterator *gsi, | |
3061 | gimple *vec_stmt, slp_tree slp_node) | |
3062 | { | |
3063 | tree vec_dest; | |
3064 | tree scalar_dest; | |
3065 | tree op0, op1 = NULL; | |
3066 | tree vec_oprnd1 = NULL_TREE; | |
3067 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3068 | tree vectype; | |
3069 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3070 | enum tree_code code; | |
3071 | enum machine_mode vec_mode; | |
3072 | tree new_temp; | |
3073 | optab optab; | |
3074 | int icode; | |
3075 | enum machine_mode optab_op2_mode; | |
3076 | tree def; | |
3077 | gimple def_stmt; | |
3078 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
3079 | gimple new_stmt = NULL; | |
3080 | stmt_vec_info prev_stmt_info; | |
3081 | int nunits_in; | |
3082 | int nunits_out; | |
3083 | tree vectype_out; | |
cede2577 | 3084 | tree op1_vectype; |
9dc3f7de IR |
3085 | int ncopies; |
3086 | int j, i; | |
6e1aa848 DN |
3087 | vec<tree> vec_oprnds0 = vNULL; |
3088 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
3089 | tree vop0, vop1; |
3090 | unsigned int k; | |
49eab32e | 3091 | bool scalar_shift_arg = true; |
9dc3f7de IR |
3092 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3093 | int vf; | |
3094 | ||
3095 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
3096 | return false; | |
3097 | ||
3098 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
3099 | return false; | |
3100 | ||
3101 | /* Is STMT a vectorizable binary/unary operation? */ | |
3102 | if (!is_gimple_assign (stmt)) | |
3103 | return false; | |
3104 | ||
3105 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3106 | return false; | |
3107 | ||
3108 | code = gimple_assign_rhs_code (stmt); | |
3109 | ||
3110 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
3111 | || code == RROTATE_EXPR)) | |
3112 | return false; | |
3113 | ||
3114 | scalar_dest = gimple_assign_lhs (stmt); | |
3115 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
7b7b1813 RG |
3116 | if (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
3117 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
3118 | { | |
73fbfcad | 3119 | if (dump_enabled_p ()) |
78c60e3d SS |
3120 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3121 | "bit-precision shifts not supported."); | |
7b7b1813 RG |
3122 | return false; |
3123 | } | |
9dc3f7de IR |
3124 | |
3125 | op0 = gimple_assign_rhs1 (stmt); | |
24ee1384 | 3126 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
9dc3f7de IR |
3127 | &def_stmt, &def, &dt[0], &vectype)) |
3128 | { | |
73fbfcad | 3129 | if (dump_enabled_p ()) |
78c60e3d SS |
3130 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3131 | "use not simple."); | |
9dc3f7de IR |
3132 | return false; |
3133 | } | |
3134 | /* If op0 is an external or constant def use a vector type with | |
3135 | the same size as the output vector type. */ | |
3136 | if (!vectype) | |
3137 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
3138 | if (vec_stmt) | |
3139 | gcc_assert (vectype); | |
3140 | if (!vectype) | |
3141 | { | |
73fbfcad | 3142 | if (dump_enabled_p ()) |
78c60e3d SS |
3143 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3144 | "no vectype for scalar type "); | |
9dc3f7de IR |
3145 | return false; |
3146 | } | |
3147 | ||
3148 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
3149 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
3150 | if (nunits_out != nunits_in) | |
3151 | return false; | |
3152 | ||
3153 | op1 = gimple_assign_rhs2 (stmt); | |
24ee1384 IR |
3154 | if (!vect_is_simple_use_1 (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3155 | &def, &dt[1], &op1_vectype)) | |
9dc3f7de | 3156 | { |
73fbfcad | 3157 | if (dump_enabled_p ()) |
78c60e3d SS |
3158 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3159 | "use not simple."); | |
9dc3f7de IR |
3160 | return false; |
3161 | } | |
3162 | ||
3163 | if (loop_vinfo) | |
3164 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3165 | else | |
3166 | vf = 1; | |
3167 | ||
3168 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3169 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3170 | case of SLP. */ | |
437f4a00 | 3171 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
9dc3f7de IR |
3172 | ncopies = 1; |
3173 | else | |
3174 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
3175 | ||
3176 | gcc_assert (ncopies >= 1); | |
3177 | ||
3178 | /* Determine whether the shift amount is a vector, or scalar. If the | |
3179 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
3180 | ||
49eab32e JJ |
3181 | if (dt[1] == vect_internal_def && !slp_node) |
3182 | scalar_shift_arg = false; | |
3183 | else if (dt[1] == vect_constant_def | |
3184 | || dt[1] == vect_external_def | |
3185 | || dt[1] == vect_internal_def) | |
3186 | { | |
3187 | /* In SLP, need to check whether the shift count is the same, | |
3188 | in loops if it is a constant or invariant, it is always | |
3189 | a scalar shift. */ | |
3190 | if (slp_node) | |
3191 | { | |
9771b263 | 3192 | vec<gimple> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
49eab32e JJ |
3193 | gimple slpstmt; |
3194 | ||
9771b263 | 3195 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
3196 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
3197 | scalar_shift_arg = false; | |
3198 | } | |
3199 | } | |
3200 | else | |
3201 | { | |
73fbfcad | 3202 | if (dump_enabled_p ()) |
78c60e3d SS |
3203 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3204 | "operand mode requires invariant argument."); | |
49eab32e JJ |
3205 | return false; |
3206 | } | |
3207 | ||
9dc3f7de | 3208 | /* Vector shifted by vector. */ |
49eab32e | 3209 | if (!scalar_shift_arg) |
9dc3f7de IR |
3210 | { |
3211 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 3212 | if (dump_enabled_p ()) |
78c60e3d SS |
3213 | dump_printf_loc (MSG_NOTE, vect_location, |
3214 | "vector/vector shift/rotate found."); | |
3215 | ||
aa948027 JJ |
3216 | if (!op1_vectype) |
3217 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
3218 | if (op1_vectype == NULL_TREE | |
3219 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 3220 | { |
73fbfcad | 3221 | if (dump_enabled_p ()) |
78c60e3d SS |
3222 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3223 | "unusable type for last operand in" | |
3224 | " vector/vector shift/rotate."); | |
cede2577 JJ |
3225 | return false; |
3226 | } | |
9dc3f7de IR |
3227 | } |
3228 | /* See if the machine has a vector shifted by scalar insn and if not | |
3229 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 3230 | else |
9dc3f7de IR |
3231 | { |
3232 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
3233 | if (optab | |
3234 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
3235 | { | |
73fbfcad | 3236 | if (dump_enabled_p ()) |
78c60e3d SS |
3237 | dump_printf_loc (MSG_NOTE, vect_location, |
3238 | "vector/scalar shift/rotate found."); | |
9dc3f7de IR |
3239 | } |
3240 | else | |
3241 | { | |
3242 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
3243 | if (optab | |
3244 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
3245 | != CODE_FOR_nothing)) | |
3246 | { | |
49eab32e JJ |
3247 | scalar_shift_arg = false; |
3248 | ||
73fbfcad | 3249 | if (dump_enabled_p ()) |
78c60e3d SS |
3250 | dump_printf_loc (MSG_NOTE, vect_location, |
3251 | "vector/vector shift/rotate found."); | |
9dc3f7de IR |
3252 | |
3253 | /* Unlike the other binary operators, shifts/rotates have | |
3254 | the rhs being int, instead of the same type as the lhs, | |
3255 | so make sure the scalar is the right type if we are | |
aa948027 | 3256 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
3257 | if (dt[1] == vect_constant_def) |
3258 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
3259 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
3260 | TREE_TYPE (op1))) | |
3261 | { | |
3262 | if (slp_node | |
3263 | && TYPE_MODE (TREE_TYPE (vectype)) | |
3264 | != TYPE_MODE (TREE_TYPE (op1))) | |
3265 | { | |
73fbfcad | 3266 | if (dump_enabled_p ()) |
78c60e3d SS |
3267 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3268 | "unusable type for last operand in" | |
3269 | " vector/vector shift/rotate."); | |
aa948027 JJ |
3270 | return false; |
3271 | } | |
3272 | if (vec_stmt && !slp_node) | |
3273 | { | |
3274 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
3275 | op1 = vect_init_vector (stmt, op1, | |
3276 | TREE_TYPE (vectype), NULL); | |
3277 | } | |
3278 | } | |
9dc3f7de IR |
3279 | } |
3280 | } | |
3281 | } | |
9dc3f7de IR |
3282 | |
3283 | /* Supportable by target? */ | |
3284 | if (!optab) | |
3285 | { | |
73fbfcad | 3286 | if (dump_enabled_p ()) |
78c60e3d SS |
3287 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3288 | "no optab."); | |
9dc3f7de IR |
3289 | return false; |
3290 | } | |
3291 | vec_mode = TYPE_MODE (vectype); | |
3292 | icode = (int) optab_handler (optab, vec_mode); | |
3293 | if (icode == CODE_FOR_nothing) | |
3294 | { | |
73fbfcad | 3295 | if (dump_enabled_p ()) |
78c60e3d SS |
3296 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3297 | "op not supported by target."); | |
9dc3f7de IR |
3298 | /* Check only during analysis. */ |
3299 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
3300 | || (vf < vect_min_worthwhile_factor (code) | |
3301 | && !vec_stmt)) | |
3302 | return false; | |
73fbfcad | 3303 | if (dump_enabled_p ()) |
78c60e3d | 3304 | dump_printf_loc (MSG_NOTE, vect_location, "proceeding using word mode."); |
9dc3f7de IR |
3305 | } |
3306 | ||
3307 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
3308 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
3309 | && vf < vect_min_worthwhile_factor (code) | |
3310 | && !vec_stmt) | |
3311 | { | |
73fbfcad | 3312 | if (dump_enabled_p ()) |
78c60e3d SS |
3313 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3314 | "not worthwhile without SIMD support."); | |
9dc3f7de IR |
3315 | return false; |
3316 | } | |
3317 | ||
3318 | if (!vec_stmt) /* transformation not required. */ | |
3319 | { | |
3320 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 3321 | if (dump_enabled_p ()) |
78c60e3d | 3322 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_shift ==="); |
c3e7ee41 | 3323 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
9dc3f7de IR |
3324 | return true; |
3325 | } | |
3326 | ||
3327 | /** Transform. **/ | |
3328 | ||
73fbfcad | 3329 | if (dump_enabled_p ()) |
78c60e3d SS |
3330 | dump_printf_loc (MSG_NOTE, vect_location, |
3331 | "transform binary/unary operation."); | |
9dc3f7de IR |
3332 | |
3333 | /* Handle def. */ | |
3334 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3335 | ||
9dc3f7de IR |
3336 | prev_stmt_info = NULL; |
3337 | for (j = 0; j < ncopies; j++) | |
3338 | { | |
3339 | /* Handle uses. */ | |
3340 | if (j == 0) | |
3341 | { | |
3342 | if (scalar_shift_arg) | |
3343 | { | |
3344 | /* Vector shl and shr insn patterns can be defined with scalar | |
3345 | operand 2 (shift operand). In this case, use constant or loop | |
3346 | invariant op1 directly, without extending it to vector mode | |
3347 | first. */ | |
3348 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
3349 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
3350 | { | |
73fbfcad | 3351 | if (dump_enabled_p ()) |
78c60e3d SS |
3352 | dump_printf_loc (MSG_NOTE, vect_location, |
3353 | "operand 1 using scalar mode."); | |
9dc3f7de | 3354 | vec_oprnd1 = op1; |
8930f723 | 3355 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 3356 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
3357 | if (slp_node) |
3358 | { | |
3359 | /* Store vec_oprnd1 for every vector stmt to be created | |
3360 | for SLP_NODE. We check during the analysis that all | |
3361 | the shift arguments are the same. | |
3362 | TODO: Allow different constants for different vector | |
3363 | stmts generated for an SLP instance. */ | |
3364 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 3365 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
3366 | } |
3367 | } | |
3368 | } | |
3369 | ||
3370 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
3371 | (a special case for certain kind of vector shifts); otherwise, | |
3372 | operand 1 should be of a vector type (the usual case). */ | |
3373 | if (vec_oprnd1) | |
3374 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
d092494c | 3375 | slp_node, -1); |
9dc3f7de IR |
3376 | else |
3377 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
d092494c | 3378 | slp_node, -1); |
9dc3f7de IR |
3379 | } |
3380 | else | |
3381 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
3382 | ||
3383 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3384 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 3385 | { |
9771b263 | 3386 | vop1 = vec_oprnds1[i]; |
9dc3f7de IR |
3387 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); |
3388 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3389 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3390 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3391 | if (slp_node) | |
9771b263 | 3392 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
3393 | } |
3394 | ||
3395 | if (slp_node) | |
3396 | continue; | |
3397 | ||
3398 | if (j == 0) | |
3399 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3400 | else | |
3401 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3402 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3403 | } | |
3404 | ||
9771b263 DN |
3405 | vec_oprnds0.release (); |
3406 | vec_oprnds1.release (); | |
9dc3f7de IR |
3407 | |
3408 | return true; | |
3409 | } | |
3410 | ||
3411 | ||
5deb57cb JJ |
3412 | static tree permute_vec_elements (tree, tree, tree, gimple, |
3413 | gimple_stmt_iterator *); | |
3414 | ||
3415 | ||
ebfd146a IR |
3416 | /* Function vectorizable_operation. |
3417 | ||
16949072 RG |
3418 | Check if STMT performs a binary, unary or ternary operation that can |
3419 | be vectorized. | |
b8698a0f | 3420 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
3421 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3422 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3423 | ||
3424 | static bool | |
3425 | vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi, | |
3426 | gimple *vec_stmt, slp_tree slp_node) | |
3427 | { | |
00f07b86 | 3428 | tree vec_dest; |
ebfd146a | 3429 | tree scalar_dest; |
16949072 | 3430 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 3431 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 3432 | tree vectype; |
ebfd146a IR |
3433 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
3434 | enum tree_code code; | |
3435 | enum machine_mode vec_mode; | |
3436 | tree new_temp; | |
3437 | int op_type; | |
00f07b86 | 3438 | optab optab; |
ebfd146a | 3439 | int icode; |
ebfd146a IR |
3440 | tree def; |
3441 | gimple def_stmt; | |
16949072 RG |
3442 | enum vect_def_type dt[3] |
3443 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
ebfd146a IR |
3444 | gimple new_stmt = NULL; |
3445 | stmt_vec_info prev_stmt_info; | |
b690cc0f | 3446 | int nunits_in; |
ebfd146a IR |
3447 | int nunits_out; |
3448 | tree vectype_out; | |
3449 | int ncopies; | |
3450 | int j, i; | |
6e1aa848 DN |
3451 | vec<tree> vec_oprnds0 = vNULL; |
3452 | vec<tree> vec_oprnds1 = vNULL; | |
3453 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 3454 | tree vop0, vop1, vop2; |
a70d6342 IR |
3455 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3456 | int vf; | |
3457 | ||
a70d6342 | 3458 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3459 | return false; |
3460 | ||
8644a673 | 3461 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3462 | return false; |
3463 | ||
3464 | /* Is STMT a vectorizable binary/unary operation? */ | |
3465 | if (!is_gimple_assign (stmt)) | |
3466 | return false; | |
3467 | ||
3468 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3469 | return false; | |
3470 | ||
ebfd146a IR |
3471 | code = gimple_assign_rhs_code (stmt); |
3472 | ||
3473 | /* For pointer addition, we should use the normal plus for | |
3474 | the vector addition. */ | |
3475 | if (code == POINTER_PLUS_EXPR) | |
3476 | code = PLUS_EXPR; | |
3477 | ||
3478 | /* Support only unary or binary operations. */ | |
3479 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 3480 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 3481 | { |
73fbfcad | 3482 | if (dump_enabled_p ()) |
78c60e3d SS |
3483 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3484 | "num. args = %d (not unary/binary/ternary op).", | |
3485 | op_type); | |
ebfd146a IR |
3486 | return false; |
3487 | } | |
3488 | ||
b690cc0f RG |
3489 | scalar_dest = gimple_assign_lhs (stmt); |
3490 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
3491 | ||
7b7b1813 RG |
3492 | /* Most operations cannot handle bit-precision types without extra |
3493 | truncations. */ | |
3494 | if ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
3495 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
3496 | /* Exception are bitwise binary operations. */ | |
3497 | && code != BIT_IOR_EXPR | |
3498 | && code != BIT_XOR_EXPR | |
3499 | && code != BIT_AND_EXPR) | |
3500 | { | |
73fbfcad | 3501 | if (dump_enabled_p ()) |
78c60e3d SS |
3502 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3503 | "bit-precision arithmetic not supported."); | |
7b7b1813 RG |
3504 | return false; |
3505 | } | |
3506 | ||
ebfd146a | 3507 | op0 = gimple_assign_rhs1 (stmt); |
24ee1384 | 3508 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
b690cc0f | 3509 | &def_stmt, &def, &dt[0], &vectype)) |
ebfd146a | 3510 | { |
73fbfcad | 3511 | if (dump_enabled_p ()) |
78c60e3d SS |
3512 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3513 | "use not simple."); | |
ebfd146a IR |
3514 | return false; |
3515 | } | |
b690cc0f RG |
3516 | /* If op0 is an external or constant def use a vector type with |
3517 | the same size as the output vector type. */ | |
3518 | if (!vectype) | |
3519 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
3520 | if (vec_stmt) |
3521 | gcc_assert (vectype); | |
3522 | if (!vectype) | |
3523 | { | |
73fbfcad | 3524 | if (dump_enabled_p ()) |
7d8930a0 | 3525 | { |
78c60e3d SS |
3526 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3527 | "no vectype for scalar type "); | |
3528 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
3529 | TREE_TYPE (op0)); | |
7d8930a0 IR |
3530 | } |
3531 | ||
3532 | return false; | |
3533 | } | |
b690cc0f RG |
3534 | |
3535 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
3536 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
3537 | if (nunits_out != nunits_in) | |
3538 | return false; | |
ebfd146a | 3539 | |
16949072 | 3540 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
3541 | { |
3542 | op1 = gimple_assign_rhs2 (stmt); | |
24ee1384 IR |
3543 | if (!vect_is_simple_use (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3544 | &def, &dt[1])) | |
ebfd146a | 3545 | { |
73fbfcad | 3546 | if (dump_enabled_p ()) |
78c60e3d SS |
3547 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3548 | "use not simple."); | |
ebfd146a IR |
3549 | return false; |
3550 | } | |
3551 | } | |
16949072 RG |
3552 | if (op_type == ternary_op) |
3553 | { | |
3554 | op2 = gimple_assign_rhs3 (stmt); | |
24ee1384 IR |
3555 | if (!vect_is_simple_use (op2, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3556 | &def, &dt[2])) | |
16949072 | 3557 | { |
73fbfcad | 3558 | if (dump_enabled_p ()) |
78c60e3d SS |
3559 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3560 | "use not simple."); | |
16949072 RG |
3561 | return false; |
3562 | } | |
3563 | } | |
ebfd146a | 3564 | |
b690cc0f RG |
3565 | if (loop_vinfo) |
3566 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3567 | else | |
3568 | vf = 1; | |
3569 | ||
3570 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 3571 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 3572 | case of SLP. */ |
437f4a00 | 3573 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
b690cc0f RG |
3574 | ncopies = 1; |
3575 | else | |
3576 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
3577 | ||
3578 | gcc_assert (ncopies >= 1); | |
3579 | ||
9dc3f7de | 3580 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
3581 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
3582 | || code == RROTATE_EXPR) | |
9dc3f7de | 3583 | return false; |
ebfd146a | 3584 | |
ebfd146a | 3585 | /* Supportable by target? */ |
00f07b86 RH |
3586 | |
3587 | vec_mode = TYPE_MODE (vectype); | |
3588 | if (code == MULT_HIGHPART_EXPR) | |
ebfd146a | 3589 | { |
00f07b86 | 3590 | if (can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype))) |
dee54b6e | 3591 | icode = LAST_INSN_CODE; |
00f07b86 RH |
3592 | else |
3593 | icode = CODE_FOR_nothing; | |
ebfd146a | 3594 | } |
00f07b86 RH |
3595 | else |
3596 | { | |
3597 | optab = optab_for_tree_code (code, vectype, optab_default); | |
3598 | if (!optab) | |
5deb57cb | 3599 | { |
73fbfcad | 3600 | if (dump_enabled_p ()) |
78c60e3d SS |
3601 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3602 | "no optab."); | |
00f07b86 | 3603 | return false; |
5deb57cb | 3604 | } |
00f07b86 | 3605 | icode = (int) optab_handler (optab, vec_mode); |
5deb57cb JJ |
3606 | } |
3607 | ||
ebfd146a IR |
3608 | if (icode == CODE_FOR_nothing) |
3609 | { | |
73fbfcad | 3610 | if (dump_enabled_p ()) |
78c60e3d SS |
3611 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3612 | "op not supported by target."); | |
ebfd146a IR |
3613 | /* Check only during analysis. */ |
3614 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5deb57cb | 3615 | || (!vec_stmt && vf < vect_min_worthwhile_factor (code))) |
ebfd146a | 3616 | return false; |
73fbfcad | 3617 | if (dump_enabled_p ()) |
78c60e3d | 3618 | dump_printf_loc (MSG_NOTE, vect_location, "proceeding using word mode."); |
383d9c83 IR |
3619 | } |
3620 | ||
4a00c761 | 3621 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
3622 | if (!VECTOR_MODE_P (vec_mode) |
3623 | && !vec_stmt | |
3624 | && vf < vect_min_worthwhile_factor (code)) | |
7d8930a0 | 3625 | { |
73fbfcad | 3626 | if (dump_enabled_p ()) |
78c60e3d SS |
3627 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3628 | "not worthwhile without SIMD support."); | |
e34842c6 | 3629 | return false; |
7d8930a0 | 3630 | } |
ebfd146a | 3631 | |
ebfd146a IR |
3632 | if (!vec_stmt) /* transformation not required. */ |
3633 | { | |
4a00c761 | 3634 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 3635 | if (dump_enabled_p ()) |
78c60e3d SS |
3636 | dump_printf_loc (MSG_NOTE, vect_location, |
3637 | "=== vectorizable_operation ==="); | |
c3e7ee41 | 3638 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
3639 | return true; |
3640 | } | |
3641 | ||
3642 | /** Transform. **/ | |
3643 | ||
73fbfcad | 3644 | if (dump_enabled_p ()) |
78c60e3d SS |
3645 | dump_printf_loc (MSG_NOTE, vect_location, |
3646 | "transform binary/unary operation."); | |
383d9c83 | 3647 | |
ebfd146a | 3648 | /* Handle def. */ |
00f07b86 | 3649 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 3650 | |
ebfd146a IR |
3651 | /* In case the vectorization factor (VF) is bigger than the number |
3652 | of elements that we can fit in a vectype (nunits), we have to generate | |
3653 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
3654 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
3655 | from one copy of the vector stmt to the next, in the field | |
3656 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
3657 | stages to find the correct vector defs to be used when vectorizing | |
3658 | stmts that use the defs of the current stmt. The example below | |
3659 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
3660 | we need to create 4 vectorized stmts): | |
3661 | ||
3662 | before vectorization: | |
3663 | RELATED_STMT VEC_STMT | |
3664 | S1: x = memref - - | |
3665 | S2: z = x + 1 - - | |
3666 | ||
3667 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
3668 | there): | |
3669 | RELATED_STMT VEC_STMT | |
3670 | VS1_0: vx0 = memref0 VS1_1 - | |
3671 | VS1_1: vx1 = memref1 VS1_2 - | |
3672 | VS1_2: vx2 = memref2 VS1_3 - | |
3673 | VS1_3: vx3 = memref3 - - | |
3674 | S1: x = load - VS1_0 | |
3675 | S2: z = x + 1 - - | |
3676 | ||
3677 | step2: vectorize stmt S2 (done here): | |
3678 | To vectorize stmt S2 we first need to find the relevant vector | |
3679 | def for the first operand 'x'. This is, as usual, obtained from | |
3680 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
3681 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
3682 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
3683 | the vector stmt VS2_0, and as usual, record it in the | |
3684 | STMT_VINFO_VEC_STMT of stmt S2. | |
3685 | When creating the second copy (VS2_1), we obtain the relevant vector | |
3686 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
3687 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
3688 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
3689 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
3690 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
3691 | chain of stmts and pointers: | |
3692 | RELATED_STMT VEC_STMT | |
3693 | VS1_0: vx0 = memref0 VS1_1 - | |
3694 | VS1_1: vx1 = memref1 VS1_2 - | |
3695 | VS1_2: vx2 = memref2 VS1_3 - | |
3696 | VS1_3: vx3 = memref3 - - | |
3697 | S1: x = load - VS1_0 | |
3698 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
3699 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
3700 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
3701 | VS2_3: vz3 = vx3 + v1 - - | |
3702 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
3703 | |
3704 | prev_stmt_info = NULL; | |
3705 | for (j = 0; j < ncopies; j++) | |
3706 | { | |
3707 | /* Handle uses. */ | |
3708 | if (j == 0) | |
4a00c761 JJ |
3709 | { |
3710 | if (op_type == binary_op || op_type == ternary_op) | |
3711 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
3712 | slp_node, -1); | |
3713 | else | |
3714 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
3715 | slp_node, -1); | |
3716 | if (op_type == ternary_op) | |
36ba4aae | 3717 | { |
9771b263 DN |
3718 | vec_oprnds2.create (1); |
3719 | vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2, | |
3720 | stmt, | |
3721 | NULL)); | |
36ba4aae | 3722 | } |
4a00c761 | 3723 | } |
ebfd146a | 3724 | else |
4a00c761 JJ |
3725 | { |
3726 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
3727 | if (op_type == ternary_op) | |
3728 | { | |
9771b263 DN |
3729 | tree vec_oprnd = vec_oprnds2.pop (); |
3730 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
3731 | vec_oprnd)); | |
4a00c761 JJ |
3732 | } |
3733 | } | |
3734 | ||
3735 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3736 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 3737 | { |
4a00c761 | 3738 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 3739 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 3740 | vop2 = ((op_type == ternary_op) |
9771b263 | 3741 | ? vec_oprnds2[i] : NULL_TREE); |
73804b12 RG |
3742 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, |
3743 | vop0, vop1, vop2); | |
4a00c761 JJ |
3744 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3745 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3746 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3747 | if (slp_node) | |
9771b263 | 3748 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
3749 | } |
3750 | ||
4a00c761 JJ |
3751 | if (slp_node) |
3752 | continue; | |
3753 | ||
3754 | if (j == 0) | |
3755 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3756 | else | |
3757 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3758 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
3759 | } |
3760 | ||
9771b263 DN |
3761 | vec_oprnds0.release (); |
3762 | vec_oprnds1.release (); | |
3763 | vec_oprnds2.release (); | |
ebfd146a | 3764 | |
ebfd146a IR |
3765 | return true; |
3766 | } | |
3767 | ||
3768 | ||
3769 | /* Function vectorizable_store. | |
3770 | ||
b8698a0f L |
3771 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
3772 | can be vectorized. | |
3773 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
3774 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3775 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3776 | ||
3777 | static bool | |
3778 | vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
3779 | slp_tree slp_node) | |
3780 | { | |
3781 | tree scalar_dest; | |
3782 | tree data_ref; | |
3783 | tree op; | |
3784 | tree vec_oprnd = NULL_TREE; | |
3785 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3786 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
3787 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
272c6793 | 3788 | tree elem_type; |
ebfd146a | 3789 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 3790 | struct loop *loop = NULL; |
ebfd146a IR |
3791 | enum machine_mode vec_mode; |
3792 | tree dummy; | |
3793 | enum dr_alignment_support alignment_support_scheme; | |
3794 | tree def; | |
3795 | gimple def_stmt; | |
3796 | enum vect_def_type dt; | |
3797 | stmt_vec_info prev_stmt_info = NULL; | |
3798 | tree dataref_ptr = NULL_TREE; | |
3799 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
3800 | int ncopies; | |
3801 | int j; | |
3802 | gimple next_stmt, first_stmt = NULL; | |
0d0293ac | 3803 | bool grouped_store = false; |
272c6793 | 3804 | bool store_lanes_p = false; |
ebfd146a | 3805 | unsigned int group_size, i; |
6e1aa848 DN |
3806 | vec<tree> dr_chain = vNULL; |
3807 | vec<tree> oprnds = vNULL; | |
3808 | vec<tree> result_chain = vNULL; | |
ebfd146a | 3809 | bool inv_p; |
6e1aa848 | 3810 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 3811 | bool slp = (slp_node != NULL); |
ebfd146a | 3812 | unsigned int vec_num; |
a70d6342 | 3813 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
272c6793 | 3814 | tree aggr_type; |
a70d6342 IR |
3815 | |
3816 | if (loop_vinfo) | |
3817 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
3818 | |
3819 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3820 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3821 | case of SLP. */ | |
437f4a00 | 3822 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
3823 | ncopies = 1; |
3824 | else | |
3825 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3826 | ||
3827 | gcc_assert (ncopies >= 1); | |
3828 | ||
3829 | /* FORNOW. This restriction should be relaxed. */ | |
a70d6342 | 3830 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) |
ebfd146a | 3831 | { |
73fbfcad | 3832 | if (dump_enabled_p ()) |
78c60e3d SS |
3833 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3834 | "multiple types in nested loop."); | |
ebfd146a IR |
3835 | return false; |
3836 | } | |
3837 | ||
a70d6342 | 3838 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3839 | return false; |
3840 | ||
8644a673 | 3841 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3842 | return false; |
3843 | ||
3844 | /* Is vectorizable store? */ | |
3845 | ||
3846 | if (!is_gimple_assign (stmt)) | |
3847 | return false; | |
3848 | ||
3849 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
3850 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
3851 | && is_pattern_stmt_p (stmt_info)) | |
3852 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a IR |
3853 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
3854 | && TREE_CODE (scalar_dest) != INDIRECT_REF | |
e9dbe7bb IR |
3855 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
3856 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
3857 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
3858 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
3859 | return false; |
3860 | ||
3861 | gcc_assert (gimple_assign_single_p (stmt)); | |
3862 | op = gimple_assign_rhs1 (stmt); | |
24ee1384 IR |
3863 | if (!vect_is_simple_use (op, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3864 | &def, &dt)) | |
ebfd146a | 3865 | { |
73fbfcad | 3866 | if (dump_enabled_p ()) |
78c60e3d SS |
3867 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3868 | "use not simple."); | |
ebfd146a IR |
3869 | return false; |
3870 | } | |
3871 | ||
272c6793 | 3872 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 3873 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 3874 | |
ebfd146a IR |
3875 | /* FORNOW. In some cases can vectorize even if data-type not supported |
3876 | (e.g. - array initialization with 0). */ | |
947131ba | 3877 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
3878 | return false; |
3879 | ||
3880 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
3881 | return false; | |
3882 | ||
a7ce6ec3 RG |
3883 | if (tree_int_cst_compare (loop && nested_in_vect_loop_p (loop, stmt) |
3884 | ? STMT_VINFO_DR_STEP (stmt_info) : DR_STEP (dr), | |
3885 | size_zero_node) < 0) | |
a1e53f3f | 3886 | { |
73fbfcad | 3887 | if (dump_enabled_p ()) |
78c60e3d SS |
3888 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3889 | "negative step for store."); | |
a1e53f3f L |
3890 | return false; |
3891 | } | |
3892 | ||
0d0293ac | 3893 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 3894 | { |
0d0293ac | 3895 | grouped_store = true; |
e14c1050 | 3896 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
3897 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
3898 | { | |
e14c1050 | 3899 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
3900 | if (vect_store_lanes_supported (vectype, group_size)) |
3901 | store_lanes_p = true; | |
0d0293ac | 3902 | else if (!vect_grouped_store_supported (vectype, group_size)) |
b602d918 RS |
3903 | return false; |
3904 | } | |
b8698a0f | 3905 | |
ebfd146a IR |
3906 | if (first_stmt == stmt) |
3907 | { | |
3908 | /* STMT is the leader of the group. Check the operands of all the | |
3909 | stmts of the group. */ | |
e14c1050 | 3910 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
3911 | while (next_stmt) |
3912 | { | |
3913 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
3914 | op = gimple_assign_rhs1 (next_stmt); | |
24ee1384 IR |
3915 | if (!vect_is_simple_use (op, next_stmt, loop_vinfo, bb_vinfo, |
3916 | &def_stmt, &def, &dt)) | |
ebfd146a | 3917 | { |
73fbfcad | 3918 | if (dump_enabled_p ()) |
78c60e3d SS |
3919 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3920 | "use not simple."); | |
ebfd146a IR |
3921 | return false; |
3922 | } | |
e14c1050 | 3923 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
3924 | } |
3925 | } | |
3926 | } | |
3927 | ||
3928 | if (!vec_stmt) /* transformation not required. */ | |
3929 | { | |
3930 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
92345349 BS |
3931 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, |
3932 | NULL, NULL, NULL); | |
ebfd146a IR |
3933 | return true; |
3934 | } | |
3935 | ||
3936 | /** Transform. **/ | |
3937 | ||
0d0293ac | 3938 | if (grouped_store) |
ebfd146a IR |
3939 | { |
3940 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 3941 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 3942 | |
e14c1050 | 3943 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
3944 | |
3945 | /* FORNOW */ | |
a70d6342 | 3946 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
3947 | |
3948 | /* We vectorize all the stmts of the interleaving group when we | |
3949 | reach the last stmt in the group. */ | |
e14c1050 IR |
3950 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
3951 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
3952 | && !slp) |
3953 | { | |
3954 | *vec_stmt = NULL; | |
3955 | return true; | |
3956 | } | |
3957 | ||
3958 | if (slp) | |
4b5caab7 | 3959 | { |
0d0293ac | 3960 | grouped_store = false; |
4b5caab7 IR |
3961 | /* VEC_NUM is the number of vect stmts to be created for this |
3962 | group. */ | |
3963 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 3964 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
4b5caab7 | 3965 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 3966 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 3967 | } |
ebfd146a | 3968 | else |
4b5caab7 IR |
3969 | /* VEC_NUM is the number of vect stmts to be created for this |
3970 | group. */ | |
ebfd146a IR |
3971 | vec_num = group_size; |
3972 | } | |
b8698a0f | 3973 | else |
ebfd146a IR |
3974 | { |
3975 | first_stmt = stmt; | |
3976 | first_dr = dr; | |
3977 | group_size = vec_num = 1; | |
ebfd146a | 3978 | } |
b8698a0f | 3979 | |
73fbfcad | 3980 | if (dump_enabled_p ()) |
78c60e3d SS |
3981 | dump_printf_loc (MSG_NOTE, vect_location, |
3982 | "transform store. ncopies = %d", ncopies); | |
ebfd146a | 3983 | |
9771b263 DN |
3984 | dr_chain.create (group_size); |
3985 | oprnds.create (group_size); | |
ebfd146a | 3986 | |
720f5239 | 3987 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 3988 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
3989 | /* Targets with store-lane instructions must not require explicit |
3990 | realignment. */ | |
3991 | gcc_assert (!store_lanes_p | |
3992 | || alignment_support_scheme == dr_aligned | |
3993 | || alignment_support_scheme == dr_unaligned_supported); | |
3994 | ||
3995 | if (store_lanes_p) | |
3996 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
3997 | else | |
3998 | aggr_type = vectype; | |
ebfd146a IR |
3999 | |
4000 | /* In case the vectorization factor (VF) is bigger than the number | |
4001 | of elements that we can fit in a vectype (nunits), we have to generate | |
4002 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 4003 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
4004 | vect_get_vec_def_for_copy_stmt. */ |
4005 | ||
0d0293ac | 4006 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
4007 | |
4008 | S1: &base + 2 = x2 | |
4009 | S2: &base = x0 | |
4010 | S3: &base + 1 = x1 | |
4011 | S4: &base + 3 = x3 | |
4012 | ||
4013 | We create vectorized stores starting from base address (the access of the | |
4014 | first stmt in the chain (S2 in the above example), when the last store stmt | |
4015 | of the chain (S4) is reached: | |
4016 | ||
4017 | VS1: &base = vx2 | |
4018 | VS2: &base + vec_size*1 = vx0 | |
4019 | VS3: &base + vec_size*2 = vx1 | |
4020 | VS4: &base + vec_size*3 = vx3 | |
4021 | ||
4022 | Then permutation statements are generated: | |
4023 | ||
3fcc1b55 JJ |
4024 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
4025 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 4026 | ... |
b8698a0f | 4027 | |
ebfd146a IR |
4028 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
4029 | (the order of the data-refs in the output of vect_permute_store_chain | |
4030 | corresponds to the order of scalar stmts in the interleaving chain - see | |
4031 | the documentation of vect_permute_store_chain()). | |
4032 | ||
4033 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 4034 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 4035 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 4036 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
4037 | */ |
4038 | ||
4039 | prev_stmt_info = NULL; | |
4040 | for (j = 0; j < ncopies; j++) | |
4041 | { | |
4042 | gimple new_stmt; | |
4043 | gimple ptr_incr; | |
4044 | ||
4045 | if (j == 0) | |
4046 | { | |
4047 | if (slp) | |
4048 | { | |
4049 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c IR |
4050 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
4051 | NULL, slp_node, -1); | |
ebfd146a | 4052 | |
9771b263 | 4053 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
4054 | } |
4055 | else | |
4056 | { | |
b8698a0f L |
4057 | /* For interleaved stores we collect vectorized defs for all the |
4058 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
4059 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
4060 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
4061 | ||
0d0293ac | 4062 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 4063 | OPRNDS are of size 1. */ |
b8698a0f | 4064 | next_stmt = first_stmt; |
ebfd146a IR |
4065 | for (i = 0; i < group_size; i++) |
4066 | { | |
b8698a0f L |
4067 | /* Since gaps are not supported for interleaved stores, |
4068 | GROUP_SIZE is the exact number of stmts in the chain. | |
4069 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
4070 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
4071 | iteration of the loop will be executed. */ |
4072 | gcc_assert (next_stmt | |
4073 | && gimple_assign_single_p (next_stmt)); | |
4074 | op = gimple_assign_rhs1 (next_stmt); | |
4075 | ||
b8698a0f | 4076 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt, |
ebfd146a | 4077 | NULL); |
9771b263 DN |
4078 | dr_chain.quick_push (vec_oprnd); |
4079 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 4080 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
4081 | } |
4082 | } | |
4083 | ||
4084 | /* We should have catched mismatched types earlier. */ | |
4085 | gcc_assert (useless_type_conversion_p (vectype, | |
4086 | TREE_TYPE (vec_oprnd))); | |
272c6793 | 4087 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, aggr_type, NULL, |
920e8172 RS |
4088 | NULL_TREE, &dummy, gsi, |
4089 | &ptr_incr, false, &inv_p); | |
a70d6342 | 4090 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 4091 | } |
b8698a0f | 4092 | else |
ebfd146a | 4093 | { |
b8698a0f L |
4094 | /* For interleaved stores we created vectorized defs for all the |
4095 | defs stored in OPRNDS in the previous iteration (previous copy). | |
4096 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
4097 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
4098 | next copy. | |
0d0293ac | 4099 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
4100 | OPRNDS are of size 1. */ |
4101 | for (i = 0; i < group_size; i++) | |
4102 | { | |
9771b263 | 4103 | op = oprnds[i]; |
24ee1384 IR |
4104 | vect_is_simple_use (op, NULL, loop_vinfo, bb_vinfo, &def_stmt, |
4105 | &def, &dt); | |
b8698a0f | 4106 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
4107 | dr_chain[i] = vec_oprnd; |
4108 | oprnds[i] = vec_oprnd; | |
ebfd146a | 4109 | } |
272c6793 RS |
4110 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
4111 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
4112 | } |
4113 | ||
272c6793 | 4114 | if (store_lanes_p) |
ebfd146a | 4115 | { |
272c6793 | 4116 | tree vec_array; |
267d3070 | 4117 | |
272c6793 RS |
4118 | /* Combine all the vectors into an array. */ |
4119 | vec_array = create_vector_array (vectype, vec_num); | |
4120 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 4121 | { |
9771b263 | 4122 | vec_oprnd = dr_chain[i]; |
272c6793 | 4123 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 4124 | } |
b8698a0f | 4125 | |
272c6793 RS |
4126 | /* Emit: |
4127 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
4128 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
4129 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
4130 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 | 4131 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
4132 | } |
4133 | else | |
4134 | { | |
4135 | new_stmt = NULL; | |
0d0293ac | 4136 | if (grouped_store) |
272c6793 | 4137 | { |
b6b9227d JJ |
4138 | if (j == 0) |
4139 | result_chain.create (group_size); | |
272c6793 RS |
4140 | /* Permute. */ |
4141 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
4142 | &result_chain); | |
4143 | } | |
c2d7ab2a | 4144 | |
272c6793 RS |
4145 | next_stmt = first_stmt; |
4146 | for (i = 0; i < vec_num; i++) | |
4147 | { | |
644ffefd | 4148 | unsigned align, misalign; |
272c6793 RS |
4149 | |
4150 | if (i > 0) | |
4151 | /* Bump the vector pointer. */ | |
4152 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
4153 | stmt, NULL_TREE); | |
4154 | ||
4155 | if (slp) | |
9771b263 | 4156 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
4157 | else if (grouped_store) |
4158 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 4159 | vect_permute_store_chain(). */ |
9771b263 | 4160 | vec_oprnd = result_chain[i]; |
272c6793 RS |
4161 | |
4162 | data_ref = build2 (MEM_REF, TREE_TYPE (vec_oprnd), dataref_ptr, | |
4163 | build_int_cst (reference_alias_ptr_type | |
4164 | (DR_REF (first_dr)), 0)); | |
644ffefd | 4165 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 | 4166 | if (aligned_access_p (first_dr)) |
644ffefd | 4167 | misalign = 0; |
272c6793 RS |
4168 | else if (DR_MISALIGNMENT (first_dr) == -1) |
4169 | { | |
4170 | TREE_TYPE (data_ref) | |
4171 | = build_aligned_type (TREE_TYPE (data_ref), | |
4172 | TYPE_ALIGN (elem_type)); | |
644ffefd MJ |
4173 | align = TYPE_ALIGN_UNIT (elem_type); |
4174 | misalign = 0; | |
272c6793 RS |
4175 | } |
4176 | else | |
4177 | { | |
4178 | TREE_TYPE (data_ref) | |
4179 | = build_aligned_type (TREE_TYPE (data_ref), | |
4180 | TYPE_ALIGN (elem_type)); | |
644ffefd | 4181 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 4182 | } |
644ffefd MJ |
4183 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
4184 | misalign); | |
c2d7ab2a | 4185 | |
272c6793 RS |
4186 | /* Arguments are ready. Create the new vector stmt. */ |
4187 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
4188 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
4189 | |
4190 | if (slp) | |
4191 | continue; | |
4192 | ||
e14c1050 | 4193 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
4194 | if (!next_stmt) |
4195 | break; | |
4196 | } | |
ebfd146a | 4197 | } |
1da0876c RS |
4198 | if (!slp) |
4199 | { | |
4200 | if (j == 0) | |
4201 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4202 | else | |
4203 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4204 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4205 | } | |
ebfd146a IR |
4206 | } |
4207 | ||
9771b263 DN |
4208 | dr_chain.release (); |
4209 | oprnds.release (); | |
4210 | result_chain.release (); | |
4211 | vec_oprnds.release (); | |
ebfd146a IR |
4212 | |
4213 | return true; | |
4214 | } | |
4215 | ||
aec7ae7d JJ |
4216 | /* Given a vector type VECTYPE and permutation SEL returns |
4217 | the VECTOR_CST mask that implements the permutation of the | |
4218 | vector elements. If that is impossible to do, returns NULL. */ | |
a1e53f3f | 4219 | |
3fcc1b55 JJ |
4220 | tree |
4221 | vect_gen_perm_mask (tree vectype, unsigned char *sel) | |
a1e53f3f | 4222 | { |
d2a12ae7 | 4223 | tree mask_elt_type, mask_type, mask_vec, *mask_elts; |
2635892a | 4224 | int i, nunits; |
a1e53f3f | 4225 | |
22e4dee7 | 4226 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 RH |
4227 | |
4228 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) | |
a1e53f3f L |
4229 | return NULL; |
4230 | ||
96f9265a RG |
4231 | mask_elt_type = lang_hooks.types.type_for_mode |
4232 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
22e4dee7 | 4233 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 4234 | |
d2a12ae7 | 4235 | mask_elts = XALLOCAVEC (tree, nunits); |
aec7ae7d | 4236 | for (i = nunits - 1; i >= 0; i--) |
d2a12ae7 RG |
4237 | mask_elts[i] = build_int_cst (mask_elt_type, sel[i]); |
4238 | mask_vec = build_vector (mask_type, mask_elts); | |
a1e53f3f | 4239 | |
2635892a | 4240 | return mask_vec; |
a1e53f3f L |
4241 | } |
4242 | ||
aec7ae7d JJ |
4243 | /* Given a vector type VECTYPE returns the VECTOR_CST mask that implements |
4244 | reversal of the vector elements. If that is impossible to do, | |
4245 | returns NULL. */ | |
4246 | ||
4247 | static tree | |
4248 | perm_mask_for_reverse (tree vectype) | |
4249 | { | |
4250 | int i, nunits; | |
4251 | unsigned char *sel; | |
4252 | ||
4253 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4254 | sel = XALLOCAVEC (unsigned char, nunits); | |
4255 | ||
4256 | for (i = 0; i < nunits; ++i) | |
4257 | sel[i] = nunits - 1 - i; | |
4258 | ||
3fcc1b55 | 4259 | return vect_gen_perm_mask (vectype, sel); |
aec7ae7d JJ |
4260 | } |
4261 | ||
4262 | /* Given a vector variable X and Y, that was generated for the scalar | |
4263 | STMT, generate instructions to permute the vector elements of X and Y | |
4264 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
4265 | permuted vector variable. */ | |
a1e53f3f L |
4266 | |
4267 | static tree | |
aec7ae7d JJ |
4268 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple stmt, |
4269 | gimple_stmt_iterator *gsi) | |
a1e53f3f L |
4270 | { |
4271 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 4272 | tree perm_dest, data_ref; |
a1e53f3f L |
4273 | gimple perm_stmt; |
4274 | ||
a1e53f3f | 4275 | perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype); |
aec7ae7d | 4276 | data_ref = make_ssa_name (perm_dest, NULL); |
a1e53f3f L |
4277 | |
4278 | /* Generate the permute statement. */ | |
73804b12 RG |
4279 | perm_stmt = gimple_build_assign_with_ops (VEC_PERM_EXPR, data_ref, |
4280 | x, y, mask_vec); | |
a1e53f3f L |
4281 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
4282 | ||
4283 | return data_ref; | |
4284 | } | |
4285 | ||
ebfd146a IR |
4286 | /* vectorizable_load. |
4287 | ||
b8698a0f L |
4288 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
4289 | can be vectorized. | |
4290 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4291 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4292 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4293 | ||
4294 | static bool | |
4295 | vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
4296 | slp_tree slp_node, slp_instance slp_node_instance) | |
4297 | { | |
4298 | tree scalar_dest; | |
4299 | tree vec_dest = NULL; | |
4300 | tree data_ref = NULL; | |
4301 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 4302 | stmt_vec_info prev_stmt_info; |
ebfd146a | 4303 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 4304 | struct loop *loop = NULL; |
ebfd146a | 4305 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 4306 | bool nested_in_vect_loop = false; |
ebfd146a IR |
4307 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; |
4308 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
272c6793 | 4309 | tree elem_type; |
ebfd146a | 4310 | tree new_temp; |
947131ba | 4311 | enum machine_mode mode; |
ebfd146a IR |
4312 | gimple new_stmt = NULL; |
4313 | tree dummy; | |
4314 | enum dr_alignment_support alignment_support_scheme; | |
4315 | tree dataref_ptr = NULL_TREE; | |
4316 | gimple ptr_incr; | |
4317 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4318 | int ncopies; | |
4319 | int i, j, group_size; | |
4320 | tree msq = NULL_TREE, lsq; | |
4321 | tree offset = NULL_TREE; | |
4322 | tree realignment_token = NULL_TREE; | |
4323 | gimple phi = NULL; | |
6e1aa848 | 4324 | vec<tree> dr_chain = vNULL; |
0d0293ac | 4325 | bool grouped_load = false; |
272c6793 | 4326 | bool load_lanes_p = false; |
ebfd146a | 4327 | gimple first_stmt; |
ebfd146a | 4328 | bool inv_p; |
319e6439 | 4329 | bool negative = false; |
ebfd146a IR |
4330 | bool compute_in_loop = false; |
4331 | struct loop *at_loop; | |
4332 | int vec_num; | |
4333 | bool slp = (slp_node != NULL); | |
4334 | bool slp_perm = false; | |
4335 | enum tree_code code; | |
a70d6342 IR |
4336 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
4337 | int vf; | |
272c6793 | 4338 | tree aggr_type; |
aec7ae7d JJ |
4339 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; |
4340 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
4341 | int gather_scale = 1; | |
4342 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
a70d6342 IR |
4343 | |
4344 | if (loop_vinfo) | |
4345 | { | |
4346 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
4347 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
4348 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4349 | } | |
4350 | else | |
3533e503 | 4351 | vf = 1; |
ebfd146a IR |
4352 | |
4353 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4354 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 4355 | case of SLP. */ |
437f4a00 | 4356 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
4357 | ncopies = 1; |
4358 | else | |
4359 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4360 | ||
4361 | gcc_assert (ncopies >= 1); | |
4362 | ||
4363 | /* FORNOW. This restriction should be relaxed. */ | |
4364 | if (nested_in_vect_loop && ncopies > 1) | |
4365 | { | |
73fbfcad | 4366 | if (dump_enabled_p ()) |
78c60e3d SS |
4367 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4368 | "multiple types in nested loop."); | |
ebfd146a IR |
4369 | return false; |
4370 | } | |
4371 | ||
a70d6342 | 4372 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4373 | return false; |
4374 | ||
8644a673 | 4375 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4376 | return false; |
4377 | ||
4378 | /* Is vectorizable load? */ | |
4379 | if (!is_gimple_assign (stmt)) | |
4380 | return false; | |
4381 | ||
4382 | scalar_dest = gimple_assign_lhs (stmt); | |
4383 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4384 | return false; | |
4385 | ||
4386 | code = gimple_assign_rhs_code (stmt); | |
4387 | if (code != ARRAY_REF | |
4388 | && code != INDIRECT_REF | |
e9dbe7bb IR |
4389 | && code != COMPONENT_REF |
4390 | && code != IMAGPART_EXPR | |
70f34814 | 4391 | && code != REALPART_EXPR |
42373e0b RG |
4392 | && code != MEM_REF |
4393 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
ebfd146a IR |
4394 | return false; |
4395 | ||
4396 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
4397 | return false; | |
4398 | ||
7b7b1813 | 4399 | elem_type = TREE_TYPE (vectype); |
947131ba | 4400 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
4401 | |
4402 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
4403 | (e.g. - data copies). */ | |
947131ba | 4404 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 4405 | { |
73fbfcad | 4406 | if (dump_enabled_p ()) |
78c60e3d SS |
4407 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4408 | "Aligned load, but unsupported type."); | |
ebfd146a IR |
4409 | return false; |
4410 | } | |
4411 | ||
ebfd146a | 4412 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 4413 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 4414 | { |
0d0293ac | 4415 | grouped_load = true; |
ebfd146a | 4416 | /* FORNOW */ |
aec7ae7d | 4417 | gcc_assert (! nested_in_vect_loop && !STMT_VINFO_GATHER_P (stmt_info)); |
ebfd146a | 4418 | |
e14c1050 | 4419 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
4420 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
4421 | { | |
e14c1050 | 4422 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
4423 | if (vect_load_lanes_supported (vectype, group_size)) |
4424 | load_lanes_p = true; | |
0d0293ac | 4425 | else if (!vect_grouped_load_supported (vectype, group_size)) |
b602d918 RS |
4426 | return false; |
4427 | } | |
ebfd146a IR |
4428 | } |
4429 | ||
a1e53f3f | 4430 | |
aec7ae7d JJ |
4431 | if (STMT_VINFO_GATHER_P (stmt_info)) |
4432 | { | |
4433 | gimple def_stmt; | |
4434 | tree def; | |
4435 | gather_decl = vect_check_gather (stmt, loop_vinfo, &gather_base, | |
4436 | &gather_off, &gather_scale); | |
4437 | gcc_assert (gather_decl); | |
24ee1384 | 4438 | if (!vect_is_simple_use_1 (gather_off, NULL, loop_vinfo, bb_vinfo, |
aec7ae7d JJ |
4439 | &def_stmt, &def, &gather_dt, |
4440 | &gather_off_vectype)) | |
4441 | { | |
73fbfcad | 4442 | if (dump_enabled_p ()) |
78c60e3d SS |
4443 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4444 | "gather index use not simple."); | |
aec7ae7d JJ |
4445 | return false; |
4446 | } | |
4447 | } | |
7d75abc8 | 4448 | else if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
14ac6aa2 | 4449 | ; |
319e6439 RG |
4450 | else |
4451 | { | |
4452 | negative = tree_int_cst_compare (nested_in_vect_loop | |
4453 | ? STMT_VINFO_DR_STEP (stmt_info) | |
4454 | : DR_STEP (dr), | |
4455 | size_zero_node) < 0; | |
4456 | if (negative && ncopies > 1) | |
4457 | { | |
73fbfcad | 4458 | if (dump_enabled_p ()) |
78c60e3d SS |
4459 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4460 | "multiple types with negative step."); | |
319e6439 RG |
4461 | return false; |
4462 | } | |
4463 | ||
4464 | if (negative) | |
4465 | { | |
4466 | gcc_assert (!grouped_load); | |
4467 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
4468 | if (alignment_support_scheme != dr_aligned | |
4469 | && alignment_support_scheme != dr_unaligned_supported) | |
4470 | { | |
73fbfcad | 4471 | if (dump_enabled_p ()) |
78c60e3d SS |
4472 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4473 | "negative step but alignment required."); | |
319e6439 RG |
4474 | return false; |
4475 | } | |
4476 | if (!perm_mask_for_reverse (vectype)) | |
4477 | { | |
73fbfcad | 4478 | if (dump_enabled_p ()) |
78c60e3d SS |
4479 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4480 | "negative step and reversing not supported."); | |
319e6439 RG |
4481 | return false; |
4482 | } | |
4483 | } | |
7d75abc8 | 4484 | } |
aec7ae7d | 4485 | |
ebfd146a IR |
4486 | if (!vec_stmt) /* transformation not required. */ |
4487 | { | |
4488 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
92345349 | 4489 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, NULL, NULL, NULL); |
ebfd146a IR |
4490 | return true; |
4491 | } | |
4492 | ||
73fbfcad | 4493 | if (dump_enabled_p ()) |
78c60e3d SS |
4494 | dump_printf_loc (MSG_NOTE, vect_location, |
4495 | "transform load. ncopies = %d", ncopies); | |
ebfd146a IR |
4496 | |
4497 | /** Transform. **/ | |
4498 | ||
aec7ae7d JJ |
4499 | if (STMT_VINFO_GATHER_P (stmt_info)) |
4500 | { | |
4501 | tree vec_oprnd0 = NULL_TREE, op; | |
4502 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
4503 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
4504 | tree ptr, mask, var, scale, perm_mask = NULL_TREE, prev_res = NULL_TREE; | |
4505 | edge pe = loop_preheader_edge (loop); | |
4506 | gimple_seq seq; | |
4507 | basic_block new_bb; | |
4508 | enum { NARROW, NONE, WIDEN } modifier; | |
4509 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
4510 | ||
4511 | if (nunits == gather_off_nunits) | |
4512 | modifier = NONE; | |
4513 | else if (nunits == gather_off_nunits / 2) | |
4514 | { | |
4515 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
4516 | modifier = WIDEN; | |
4517 | ||
4518 | for (i = 0; i < gather_off_nunits; ++i) | |
4519 | sel[i] = i | nunits; | |
4520 | ||
3fcc1b55 | 4521 | perm_mask = vect_gen_perm_mask (gather_off_vectype, sel); |
aec7ae7d JJ |
4522 | gcc_assert (perm_mask != NULL_TREE); |
4523 | } | |
4524 | else if (nunits == gather_off_nunits * 2) | |
4525 | { | |
4526 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
4527 | modifier = NARROW; | |
4528 | ||
4529 | for (i = 0; i < nunits; ++i) | |
4530 | sel[i] = i < gather_off_nunits | |
4531 | ? i : i + nunits - gather_off_nunits; | |
4532 | ||
3fcc1b55 | 4533 | perm_mask = vect_gen_perm_mask (vectype, sel); |
aec7ae7d JJ |
4534 | gcc_assert (perm_mask != NULL_TREE); |
4535 | ncopies *= 2; | |
4536 | } | |
4537 | else | |
4538 | gcc_unreachable (); | |
4539 | ||
4540 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
4541 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4542 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4543 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4544 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4545 | scaletype = TREE_VALUE (arglist); | |
4546 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
4547 | && types_compatible_p (srctype, masktype)); | |
4548 | ||
4549 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4550 | ||
4551 | ptr = fold_convert (ptrtype, gather_base); | |
4552 | if (!is_gimple_min_invariant (ptr)) | |
4553 | { | |
4554 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
4555 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
4556 | gcc_assert (!new_bb); | |
4557 | } | |
4558 | ||
4559 | /* Currently we support only unconditional gather loads, | |
4560 | so mask should be all ones. */ | |
4561 | if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
4562 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
4563 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) | |
4564 | { | |
4565 | REAL_VALUE_TYPE r; | |
4566 | long tmp[6]; | |
4567 | for (j = 0; j < 6; ++j) | |
4568 | tmp[j] = -1; | |
4569 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
4570 | mask = build_real (TREE_TYPE (masktype), r); | |
4571 | } | |
4572 | else | |
4573 | gcc_unreachable (); | |
4574 | mask = build_vector_from_val (masktype, mask); | |
4575 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
4576 | ||
4577 | scale = build_int_cst (scaletype, gather_scale); | |
4578 | ||
4579 | prev_stmt_info = NULL; | |
4580 | for (j = 0; j < ncopies; ++j) | |
4581 | { | |
4582 | if (modifier == WIDEN && (j & 1)) | |
4583 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
4584 | perm_mask, stmt, gsi); | |
4585 | else if (j == 0) | |
4586 | op = vec_oprnd0 | |
4587 | = vect_get_vec_def_for_operand (gather_off, stmt, NULL); | |
4588 | else | |
4589 | op = vec_oprnd0 | |
4590 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
4591 | ||
4592 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
4593 | { | |
4594 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
4595 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
4596 | var = vect_get_new_vect_var (idxtype, vect_simple_var, NULL); | |
aec7ae7d JJ |
4597 | var = make_ssa_name (var, NULL); |
4598 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); | |
4599 | new_stmt | |
4600 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, | |
4601 | op, NULL_TREE); | |
4602 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4603 | op = var; | |
4604 | } | |
4605 | ||
4606 | new_stmt | |
4607 | = gimple_build_call (gather_decl, 5, mask, ptr, op, mask, scale); | |
4608 | ||
4609 | if (!useless_type_conversion_p (vectype, rettype)) | |
4610 | { | |
4611 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
4612 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
4613 | var = vect_get_new_vect_var (rettype, vect_simple_var, NULL); | |
aec7ae7d JJ |
4614 | op = make_ssa_name (var, new_stmt); |
4615 | gimple_call_set_lhs (new_stmt, op); | |
4616 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4617 | var = make_ssa_name (vec_dest, NULL); | |
4618 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); | |
4619 | new_stmt | |
4620 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, op, | |
4621 | NULL_TREE); | |
4622 | } | |
4623 | else | |
4624 | { | |
4625 | var = make_ssa_name (vec_dest, new_stmt); | |
4626 | gimple_call_set_lhs (new_stmt, var); | |
4627 | } | |
4628 | ||
4629 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4630 | ||
4631 | if (modifier == NARROW) | |
4632 | { | |
4633 | if ((j & 1) == 0) | |
4634 | { | |
4635 | prev_res = var; | |
4636 | continue; | |
4637 | } | |
4638 | var = permute_vec_elements (prev_res, var, | |
4639 | perm_mask, stmt, gsi); | |
4640 | new_stmt = SSA_NAME_DEF_STMT (var); | |
4641 | } | |
4642 | ||
4643 | if (prev_stmt_info == NULL) | |
4644 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4645 | else | |
4646 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4647 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4648 | } | |
4649 | return true; | |
4650 | } | |
7d75abc8 MM |
4651 | else if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
4652 | { | |
4653 | gimple_stmt_iterator incr_gsi; | |
4654 | bool insert_after; | |
4655 | gimple incr; | |
4656 | tree offvar; | |
7d75abc8 MM |
4657 | tree ivstep; |
4658 | tree running_off; | |
9771b263 | 4659 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 4660 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
4661 | tree stride_base, stride_step, alias_off; |
4662 | ||
4663 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 4664 | |
14ac6aa2 RB |
4665 | stride_base |
4666 | = fold_build_pointer_plus | |
4667 | (unshare_expr (DR_BASE_ADDRESS (dr)), | |
4668 | size_binop (PLUS_EXPR, | |
4669 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (dr))), | |
4670 | convert_to_ptrofftype (DR_INIT(dr)))); | |
4671 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (dr))); | |
7d75abc8 MM |
4672 | |
4673 | /* For a load with loop-invariant (but other than power-of-2) | |
4674 | stride (i.e. not a grouped access) like so: | |
4675 | ||
4676 | for (i = 0; i < n; i += stride) | |
4677 | ... = array[i]; | |
4678 | ||
4679 | we generate a new induction variable and new accesses to | |
4680 | form a new vector (or vectors, depending on ncopies): | |
4681 | ||
4682 | for (j = 0; ; j += VF*stride) | |
4683 | tmp1 = array[j]; | |
4684 | tmp2 = array[j + stride]; | |
4685 | ... | |
4686 | vectemp = {tmp1, tmp2, ...} | |
4687 | */ | |
4688 | ||
4689 | ivstep = stride_step; | |
4690 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
4691 | build_int_cst (TREE_TYPE (ivstep), vf)); | |
4692 | ||
4693 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
4694 | ||
4695 | create_iv (stride_base, ivstep, NULL, | |
4696 | loop, &incr_gsi, insert_after, | |
4697 | &offvar, NULL); | |
4698 | incr = gsi_stmt (incr_gsi); | |
4699 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo, NULL)); | |
4700 | ||
4701 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
4702 | if (stmts) | |
4703 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
4704 | ||
4705 | prev_stmt_info = NULL; | |
4706 | running_off = offvar; | |
14ac6aa2 | 4707 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (dr)), 0); |
7d75abc8 MM |
4708 | for (j = 0; j < ncopies; j++) |
4709 | { | |
4710 | tree vec_inv; | |
4711 | ||
9771b263 | 4712 | vec_alloc (v, nunits); |
7d75abc8 MM |
4713 | for (i = 0; i < nunits; i++) |
4714 | { | |
4715 | tree newref, newoff; | |
4716 | gimple incr; | |
14ac6aa2 RB |
4717 | newref = build2 (MEM_REF, TREE_TYPE (vectype), |
4718 | running_off, alias_off); | |
7d75abc8 MM |
4719 | |
4720 | newref = force_gimple_operand_gsi (gsi, newref, true, | |
4721 | NULL_TREE, true, | |
4722 | GSI_SAME_STMT); | |
4723 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, newref); | |
070ecdfd | 4724 | newoff = copy_ssa_name (running_off, NULL); |
14ac6aa2 RB |
4725 | incr = gimple_build_assign_with_ops (POINTER_PLUS_EXPR, newoff, |
4726 | running_off, stride_step); | |
7d75abc8 MM |
4727 | vect_finish_stmt_generation (stmt, incr, gsi); |
4728 | ||
4729 | running_off = newoff; | |
4730 | } | |
4731 | ||
4732 | vec_inv = build_constructor (vectype, v); | |
4733 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
4734 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
7d75abc8 MM |
4735 | |
4736 | if (j == 0) | |
4737 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4738 | else | |
4739 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4740 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4741 | } | |
4742 | return true; | |
4743 | } | |
aec7ae7d | 4744 | |
0d0293ac | 4745 | if (grouped_load) |
ebfd146a | 4746 | { |
e14c1050 | 4747 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
6aa904c4 | 4748 | if (slp |
9771b263 DN |
4749 | && !SLP_INSTANCE_LOAD_PERMUTATION (slp_node_instance).exists () |
4750 | && first_stmt != SLP_TREE_SCALAR_STMTS (slp_node)[0]) | |
4751 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 4752 | |
ebfd146a IR |
4753 | /* Check if the chain of loads is already vectorized. */ |
4754 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt))) | |
4755 | { | |
4756 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4757 | return true; | |
4758 | } | |
4759 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 4760 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a IR |
4761 | |
4762 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
4763 | if (slp) | |
4764 | { | |
0d0293ac | 4765 | grouped_load = false; |
ebfd146a | 4766 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
9771b263 | 4767 | if (SLP_INSTANCE_LOAD_PERMUTATION (slp_node_instance).exists ()) |
a70d6342 IR |
4768 | slp_perm = true; |
4769 | } | |
ebfd146a IR |
4770 | else |
4771 | vec_num = group_size; | |
ebfd146a IR |
4772 | } |
4773 | else | |
4774 | { | |
4775 | first_stmt = stmt; | |
4776 | first_dr = dr; | |
4777 | group_size = vec_num = 1; | |
4778 | } | |
4779 | ||
720f5239 | 4780 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 4781 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
4782 | /* Targets with load-lane instructions must not require explicit |
4783 | realignment. */ | |
4784 | gcc_assert (!load_lanes_p | |
4785 | || alignment_support_scheme == dr_aligned | |
4786 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
4787 | |
4788 | /* In case the vectorization factor (VF) is bigger than the number | |
4789 | of elements that we can fit in a vectype (nunits), we have to generate | |
4790 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 4791 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 4792 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 4793 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 4794 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
4795 | stmts that use the defs of the current stmt. The example below |
4796 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
4797 | need to create 4 vectorized stmts): | |
ebfd146a IR |
4798 | |
4799 | before vectorization: | |
4800 | RELATED_STMT VEC_STMT | |
4801 | S1: x = memref - - | |
4802 | S2: z = x + 1 - - | |
4803 | ||
4804 | step 1: vectorize stmt S1: | |
4805 | We first create the vector stmt VS1_0, and, as usual, record a | |
4806 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
4807 | Next, we create the vector stmt VS1_1, and record a pointer to | |
4808 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 4809 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
4810 | stmts and pointers: |
4811 | RELATED_STMT VEC_STMT | |
4812 | VS1_0: vx0 = memref0 VS1_1 - | |
4813 | VS1_1: vx1 = memref1 VS1_2 - | |
4814 | VS1_2: vx2 = memref2 VS1_3 - | |
4815 | VS1_3: vx3 = memref3 - - | |
4816 | S1: x = load - VS1_0 | |
4817 | S2: z = x + 1 - - | |
4818 | ||
b8698a0f L |
4819 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
4820 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
4821 | stmt S2. */ |
4822 | ||
0d0293ac | 4823 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
4824 | |
4825 | S1: x2 = &base + 2 | |
4826 | S2: x0 = &base | |
4827 | S3: x1 = &base + 1 | |
4828 | S4: x3 = &base + 3 | |
4829 | ||
b8698a0f | 4830 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
4831 | starting from the access of the first stmt of the chain: |
4832 | ||
4833 | VS1: vx0 = &base | |
4834 | VS2: vx1 = &base + vec_size*1 | |
4835 | VS3: vx3 = &base + vec_size*2 | |
4836 | VS4: vx4 = &base + vec_size*3 | |
4837 | ||
4838 | Then permutation statements are generated: | |
4839 | ||
e2c83630 RH |
4840 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
4841 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
4842 | ... |
4843 | ||
4844 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
4845 | (the order of the data-refs in the output of vect_permute_load_chain | |
4846 | corresponds to the order of scalar stmts in the interleaving chain - see | |
4847 | the documentation of vect_permute_load_chain()). | |
4848 | The generation of permutation stmts and recording them in | |
0d0293ac | 4849 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 4850 | |
b8698a0f | 4851 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
4852 | permutation stmts above are created for every copy. The result vector |
4853 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
4854 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
4855 | |
4856 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
4857 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
4858 | we generate the following code: | |
4859 | p = initial_addr; | |
4860 | indx = 0; | |
4861 | loop { | |
4862 | p = p + indx * vectype_size; | |
4863 | vec_dest = *(p); | |
4864 | indx = indx + 1; | |
4865 | } | |
4866 | ||
4867 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 4868 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
4869 | then generate the following code, in which the data in each iteration is |
4870 | obtained by two vector loads, one from the previous iteration, and one | |
4871 | from the current iteration: | |
4872 | p1 = initial_addr; | |
4873 | msq_init = *(floor(p1)) | |
4874 | p2 = initial_addr + VS - 1; | |
4875 | realignment_token = call target_builtin; | |
4876 | indx = 0; | |
4877 | loop { | |
4878 | p2 = p2 + indx * vectype_size | |
4879 | lsq = *(floor(p2)) | |
4880 | vec_dest = realign_load (msq, lsq, realignment_token) | |
4881 | indx = indx + 1; | |
4882 | msq = lsq; | |
4883 | } */ | |
4884 | ||
4885 | /* If the misalignment remains the same throughout the execution of the | |
4886 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 4887 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
4888 | This can only occur when vectorizing memory accesses in the inner-loop |
4889 | nested within an outer-loop that is being vectorized. */ | |
4890 | ||
d1e4b493 | 4891 | if (nested_in_vect_loop |
211bea38 | 4892 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
ebfd146a IR |
4893 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) |
4894 | { | |
4895 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
4896 | compute_in_loop = true; | |
4897 | } | |
4898 | ||
4899 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
4900 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 4901 | && !compute_in_loop) |
ebfd146a IR |
4902 | { |
4903 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
4904 | alignment_support_scheme, NULL_TREE, | |
4905 | &at_loop); | |
4906 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
4907 | { | |
4908 | phi = SSA_NAME_DEF_STMT (msq); | |
4909 | offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
4910 | } | |
4911 | } | |
4912 | else | |
4913 | at_loop = loop; | |
4914 | ||
a1e53f3f L |
4915 | if (negative) |
4916 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
4917 | ||
272c6793 RS |
4918 | if (load_lanes_p) |
4919 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
4920 | else | |
4921 | aggr_type = vectype; | |
4922 | ||
ebfd146a IR |
4923 | prev_stmt_info = NULL; |
4924 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 4925 | { |
272c6793 | 4926 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 4927 | if (j == 0) |
272c6793 | 4928 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, |
920e8172 RS |
4929 | offset, &dummy, gsi, |
4930 | &ptr_incr, false, &inv_p); | |
ebfd146a | 4931 | else |
272c6793 RS |
4932 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
4933 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 4934 | |
0d0293ac | 4935 | if (grouped_load || slp_perm) |
9771b263 | 4936 | dr_chain.create (vec_num); |
5ce1ee7f | 4937 | |
272c6793 | 4938 | if (load_lanes_p) |
ebfd146a | 4939 | { |
272c6793 RS |
4940 | tree vec_array; |
4941 | ||
4942 | vec_array = create_vector_array (vectype, vec_num); | |
4943 | ||
4944 | /* Emit: | |
4945 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
4946 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
4947 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
4948 | gimple_call_set_lhs (new_stmt, vec_array); | |
4949 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 4950 | |
272c6793 RS |
4951 | /* Extract each vector into an SSA_NAME. */ |
4952 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 4953 | { |
272c6793 RS |
4954 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
4955 | vec_array, i); | |
9771b263 | 4956 | dr_chain.quick_push (new_temp); |
272c6793 RS |
4957 | } |
4958 | ||
4959 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 4960 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
4961 | } |
4962 | else | |
4963 | { | |
4964 | for (i = 0; i < vec_num; i++) | |
4965 | { | |
4966 | if (i > 0) | |
4967 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
4968 | stmt, NULL_TREE); | |
4969 | ||
4970 | /* 2. Create the vector-load in the loop. */ | |
4971 | switch (alignment_support_scheme) | |
4972 | { | |
4973 | case dr_aligned: | |
4974 | case dr_unaligned_supported: | |
be1ac4ec | 4975 | { |
644ffefd MJ |
4976 | unsigned int align, misalign; |
4977 | ||
272c6793 RS |
4978 | data_ref |
4979 | = build2 (MEM_REF, vectype, dataref_ptr, | |
4980 | build_int_cst (reference_alias_ptr_type | |
4981 | (DR_REF (first_dr)), 0)); | |
644ffefd | 4982 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 RS |
4983 | if (alignment_support_scheme == dr_aligned) |
4984 | { | |
4985 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 4986 | misalign = 0; |
272c6793 RS |
4987 | } |
4988 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
4989 | { | |
4990 | TREE_TYPE (data_ref) | |
4991 | = build_aligned_type (TREE_TYPE (data_ref), | |
4992 | TYPE_ALIGN (elem_type)); | |
644ffefd MJ |
4993 | align = TYPE_ALIGN_UNIT (elem_type); |
4994 | misalign = 0; | |
272c6793 RS |
4995 | } |
4996 | else | |
4997 | { | |
4998 | TREE_TYPE (data_ref) | |
4999 | = build_aligned_type (TREE_TYPE (data_ref), | |
5000 | TYPE_ALIGN (elem_type)); | |
644ffefd | 5001 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 5002 | } |
644ffefd MJ |
5003 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
5004 | align, misalign); | |
272c6793 | 5005 | break; |
be1ac4ec | 5006 | } |
272c6793 | 5007 | case dr_explicit_realign: |
267d3070 | 5008 | { |
272c6793 RS |
5009 | tree ptr, bump; |
5010 | tree vs_minus_1; | |
5011 | ||
5012 | vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
5013 | ||
5014 | if (compute_in_loop) | |
5015 | msq = vect_setup_realignment (first_stmt, gsi, | |
5016 | &realignment_token, | |
5017 | dr_explicit_realign, | |
5018 | dataref_ptr, NULL); | |
5019 | ||
070ecdfd | 5020 | ptr = copy_ssa_name (dataref_ptr, NULL); |
272c6793 | 5021 | new_stmt = gimple_build_assign_with_ops |
070ecdfd | 5022 | (BIT_AND_EXPR, ptr, dataref_ptr, |
272c6793 RS |
5023 | build_int_cst |
5024 | (TREE_TYPE (dataref_ptr), | |
5025 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
5026 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
5027 | data_ref | |
5028 | = build2 (MEM_REF, vectype, ptr, | |
5029 | build_int_cst (reference_alias_ptr_type | |
5030 | (DR_REF (first_dr)), 0)); | |
5031 | vec_dest = vect_create_destination_var (scalar_dest, | |
5032 | vectype); | |
5033 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
5034 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5035 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5036 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
5037 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
5038 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5039 | msq = new_temp; | |
5040 | ||
5041 | bump = size_binop (MULT_EXPR, vs_minus_1, | |
7b7b1813 | 5042 | TYPE_SIZE_UNIT (elem_type)); |
272c6793 RS |
5043 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
5044 | new_stmt = gimple_build_assign_with_ops | |
5045 | (BIT_AND_EXPR, NULL_TREE, ptr, | |
5046 | build_int_cst | |
5047 | (TREE_TYPE (ptr), | |
5048 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
070ecdfd | 5049 | ptr = copy_ssa_name (dataref_ptr, new_stmt); |
272c6793 RS |
5050 | gimple_assign_set_lhs (new_stmt, ptr); |
5051 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5052 | data_ref | |
5053 | = build2 (MEM_REF, vectype, ptr, | |
5054 | build_int_cst (reference_alias_ptr_type | |
5055 | (DR_REF (first_dr)), 0)); | |
5056 | break; | |
267d3070 | 5057 | } |
272c6793 | 5058 | case dr_explicit_realign_optimized: |
070ecdfd | 5059 | new_temp = copy_ssa_name (dataref_ptr, NULL); |
272c6793 | 5060 | new_stmt = gimple_build_assign_with_ops |
070ecdfd | 5061 | (BIT_AND_EXPR, new_temp, dataref_ptr, |
272c6793 RS |
5062 | build_int_cst |
5063 | (TREE_TYPE (dataref_ptr), | |
5064 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
5065 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
5066 | data_ref | |
5067 | = build2 (MEM_REF, vectype, new_temp, | |
5068 | build_int_cst (reference_alias_ptr_type | |
5069 | (DR_REF (first_dr)), 0)); | |
5070 | break; | |
5071 | default: | |
5072 | gcc_unreachable (); | |
5073 | } | |
ebfd146a | 5074 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 5075 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
5076 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5077 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5078 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5079 | ||
272c6793 RS |
5080 | /* 3. Handle explicit realignment if necessary/supported. |
5081 | Create in loop: | |
5082 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
5083 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
5084 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 5085 | { |
272c6793 RS |
5086 | lsq = gimple_assign_lhs (new_stmt); |
5087 | if (!realignment_token) | |
5088 | realignment_token = dataref_ptr; | |
5089 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5090 | new_stmt | |
73804b12 RG |
5091 | = gimple_build_assign_with_ops (REALIGN_LOAD_EXPR, |
5092 | vec_dest, msq, lsq, | |
5093 | realignment_token); | |
272c6793 RS |
5094 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5095 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5096 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5097 | ||
5098 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
5099 | { | |
5100 | gcc_assert (phi); | |
5101 | if (i == vec_num - 1 && j == ncopies - 1) | |
5102 | add_phi_arg (phi, lsq, | |
5103 | loop_latch_edge (containing_loop), | |
9e227d60 | 5104 | UNKNOWN_LOCATION); |
272c6793 RS |
5105 | msq = lsq; |
5106 | } | |
ebfd146a | 5107 | } |
ebfd146a | 5108 | |
59fd17e3 RB |
5109 | /* 4. Handle invariant-load. */ |
5110 | if (inv_p && !bb_vinfo) | |
5111 | { | |
5112 | gimple_stmt_iterator gsi2 = *gsi; | |
5113 | gcc_assert (!grouped_load); | |
5114 | gsi_next (&gsi2); | |
5115 | new_temp = vect_init_vector (stmt, scalar_dest, | |
5116 | vectype, &gsi2); | |
5117 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
5118 | } | |
5119 | ||
272c6793 RS |
5120 | if (negative) |
5121 | { | |
aec7ae7d JJ |
5122 | tree perm_mask = perm_mask_for_reverse (vectype); |
5123 | new_temp = permute_vec_elements (new_temp, new_temp, | |
5124 | perm_mask, stmt, gsi); | |
ebfd146a IR |
5125 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
5126 | } | |
267d3070 | 5127 | |
272c6793 | 5128 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
5129 | vect_transform_grouped_load (). */ |
5130 | if (grouped_load || slp_perm) | |
9771b263 | 5131 | dr_chain.quick_push (new_temp); |
267d3070 | 5132 | |
272c6793 RS |
5133 | /* Store vector loads in the corresponding SLP_NODE. */ |
5134 | if (slp && !slp_perm) | |
9771b263 | 5135 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
272c6793 | 5136 | } |
ebfd146a IR |
5137 | } |
5138 | ||
5139 | if (slp && !slp_perm) | |
5140 | continue; | |
5141 | ||
5142 | if (slp_perm) | |
5143 | { | |
a70d6342 | 5144 | if (!vect_transform_slp_perm_load (stmt, dr_chain, gsi, vf, |
ebfd146a IR |
5145 | slp_node_instance, false)) |
5146 | { | |
9771b263 | 5147 | dr_chain.release (); |
ebfd146a IR |
5148 | return false; |
5149 | } | |
5150 | } | |
5151 | else | |
5152 | { | |
0d0293ac | 5153 | if (grouped_load) |
ebfd146a | 5154 | { |
272c6793 | 5155 | if (!load_lanes_p) |
0d0293ac | 5156 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 5157 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
5158 | } |
5159 | else | |
5160 | { | |
5161 | if (j == 0) | |
5162 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5163 | else | |
5164 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5165 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5166 | } | |
5167 | } | |
9771b263 | 5168 | dr_chain.release (); |
ebfd146a IR |
5169 | } |
5170 | ||
ebfd146a IR |
5171 | return true; |
5172 | } | |
5173 | ||
5174 | /* Function vect_is_simple_cond. | |
b8698a0f | 5175 | |
ebfd146a IR |
5176 | Input: |
5177 | LOOP - the loop that is being vectorized. | |
5178 | COND - Condition that is checked for simple use. | |
5179 | ||
e9e1d143 RG |
5180 | Output: |
5181 | *COMP_VECTYPE - the vector type for the comparison. | |
5182 | ||
ebfd146a IR |
5183 | Returns whether a COND can be vectorized. Checks whether |
5184 | condition operands are supportable using vec_is_simple_use. */ | |
5185 | ||
87aab9b2 | 5186 | static bool |
24ee1384 IR |
5187 | vect_is_simple_cond (tree cond, gimple stmt, loop_vec_info loop_vinfo, |
5188 | bb_vec_info bb_vinfo, tree *comp_vectype) | |
ebfd146a IR |
5189 | { |
5190 | tree lhs, rhs; | |
5191 | tree def; | |
5192 | enum vect_def_type dt; | |
e9e1d143 | 5193 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a IR |
5194 | |
5195 | if (!COMPARISON_CLASS_P (cond)) | |
5196 | return false; | |
5197 | ||
5198 | lhs = TREE_OPERAND (cond, 0); | |
5199 | rhs = TREE_OPERAND (cond, 1); | |
5200 | ||
5201 | if (TREE_CODE (lhs) == SSA_NAME) | |
5202 | { | |
5203 | gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); | |
24ee1384 IR |
5204 | if (!vect_is_simple_use_1 (lhs, stmt, loop_vinfo, bb_vinfo, |
5205 | &lhs_def_stmt, &def, &dt, &vectype1)) | |
ebfd146a IR |
5206 | return false; |
5207 | } | |
5208 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
5209 | && TREE_CODE (lhs) != FIXED_CST) | |
5210 | return false; | |
5211 | ||
5212 | if (TREE_CODE (rhs) == SSA_NAME) | |
5213 | { | |
5214 | gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); | |
24ee1384 IR |
5215 | if (!vect_is_simple_use_1 (rhs, stmt, loop_vinfo, bb_vinfo, |
5216 | &rhs_def_stmt, &def, &dt, &vectype2)) | |
ebfd146a IR |
5217 | return false; |
5218 | } | |
f7e531cf | 5219 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST |
ebfd146a IR |
5220 | && TREE_CODE (rhs) != FIXED_CST) |
5221 | return false; | |
5222 | ||
e9e1d143 | 5223 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
5224 | return true; |
5225 | } | |
5226 | ||
5227 | /* vectorizable_condition. | |
5228 | ||
b8698a0f L |
5229 | Check if STMT is conditional modify expression that can be vectorized. |
5230 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
5231 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
5232 | at GSI. |
5233 | ||
5234 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
5235 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
5236 | else caluse if it is 2). | |
ebfd146a IR |
5237 | |
5238 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5239 | ||
4bbe8262 | 5240 | bool |
ebfd146a | 5241 | vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi, |
f7e531cf IR |
5242 | gimple *vec_stmt, tree reduc_def, int reduc_index, |
5243 | slp_tree slp_node) | |
ebfd146a IR |
5244 | { |
5245 | tree scalar_dest = NULL_TREE; | |
5246 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
5247 | tree cond_expr, then_clause, else_clause; |
5248 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5249 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
df11cc78 | 5250 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
5251 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
5252 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
ebfd146a IR |
5253 | tree vec_compare, vec_cond_expr; |
5254 | tree new_temp; | |
5255 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
ebfd146a | 5256 | tree def; |
a855b1b1 | 5257 | enum vect_def_type dt, dts[4]; |
ebfd146a | 5258 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
f7e531cf | 5259 | int ncopies; |
ebfd146a | 5260 | enum tree_code code; |
a855b1b1 | 5261 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
5262 | int i, j; |
5263 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
5264 | vec<tree> vec_oprnds0 = vNULL; |
5265 | vec<tree> vec_oprnds1 = vNULL; | |
5266 | vec<tree> vec_oprnds2 = vNULL; | |
5267 | vec<tree> vec_oprnds3 = vNULL; | |
784fb9b3 | 5268 | tree vec_cmp_type = vectype; |
b8698a0f | 5269 | |
f7e531cf IR |
5270 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
5271 | ncopies = 1; | |
5272 | else | |
5273 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
437f4a00 | 5274 | |
ebfd146a | 5275 | gcc_assert (ncopies >= 1); |
a855b1b1 | 5276 | if (reduc_index && ncopies > 1) |
ebfd146a IR |
5277 | return false; /* FORNOW */ |
5278 | ||
f7e531cf IR |
5279 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
5280 | return false; | |
5281 | ||
5282 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
ebfd146a IR |
5283 | return false; |
5284 | ||
4bbe8262 IR |
5285 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5286 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
5287 | && reduc_def)) | |
ebfd146a IR |
5288 | return false; |
5289 | ||
ebfd146a | 5290 | /* FORNOW: not yet supported. */ |
b8698a0f | 5291 | if (STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 5292 | { |
73fbfcad | 5293 | if (dump_enabled_p ()) |
78c60e3d SS |
5294 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5295 | "value used after loop."); | |
ebfd146a IR |
5296 | return false; |
5297 | } | |
5298 | ||
5299 | /* Is vectorizable conditional operation? */ | |
5300 | if (!is_gimple_assign (stmt)) | |
5301 | return false; | |
5302 | ||
5303 | code = gimple_assign_rhs_code (stmt); | |
5304 | ||
5305 | if (code != COND_EXPR) | |
5306 | return false; | |
5307 | ||
4e71066d RG |
5308 | cond_expr = gimple_assign_rhs1 (stmt); |
5309 | then_clause = gimple_assign_rhs2 (stmt); | |
5310 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 5311 | |
24ee1384 IR |
5312 | if (!vect_is_simple_cond (cond_expr, stmt, loop_vinfo, bb_vinfo, |
5313 | &comp_vectype) | |
e9e1d143 | 5314 | || !comp_vectype) |
ebfd146a IR |
5315 | return false; |
5316 | ||
5317 | if (TREE_CODE (then_clause) == SSA_NAME) | |
5318 | { | |
5319 | gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause); | |
24ee1384 | 5320 | if (!vect_is_simple_use (then_clause, stmt, loop_vinfo, bb_vinfo, |
ebfd146a IR |
5321 | &then_def_stmt, &def, &dt)) |
5322 | return false; | |
5323 | } | |
b8698a0f | 5324 | else if (TREE_CODE (then_clause) != INTEGER_CST |
ebfd146a IR |
5325 | && TREE_CODE (then_clause) != REAL_CST |
5326 | && TREE_CODE (then_clause) != FIXED_CST) | |
5327 | return false; | |
5328 | ||
5329 | if (TREE_CODE (else_clause) == SSA_NAME) | |
5330 | { | |
5331 | gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause); | |
24ee1384 | 5332 | if (!vect_is_simple_use (else_clause, stmt, loop_vinfo, bb_vinfo, |
ebfd146a IR |
5333 | &else_def_stmt, &def, &dt)) |
5334 | return false; | |
5335 | } | |
b8698a0f | 5336 | else if (TREE_CODE (else_clause) != INTEGER_CST |
ebfd146a IR |
5337 | && TREE_CODE (else_clause) != REAL_CST |
5338 | && TREE_CODE (else_clause) != FIXED_CST) | |
5339 | return false; | |
5340 | ||
784fb9b3 JJ |
5341 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype))) |
5342 | { | |
5343 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (vectype))); | |
5344 | tree cmp_type = build_nonstandard_integer_type (prec, 1); | |
5345 | vec_cmp_type = get_same_sized_vectype (cmp_type, vectype); | |
5346 | if (vec_cmp_type == NULL_TREE) | |
5347 | return false; | |
5348 | } | |
5349 | ||
b8698a0f | 5350 | if (!vec_stmt) |
ebfd146a IR |
5351 | { |
5352 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
e9e1d143 | 5353 | return expand_vec_cond_expr_p (vectype, comp_vectype); |
ebfd146a IR |
5354 | } |
5355 | ||
f7e531cf IR |
5356 | /* Transform. */ |
5357 | ||
5358 | if (!slp_node) | |
5359 | { | |
9771b263 DN |
5360 | vec_oprnds0.create (1); |
5361 | vec_oprnds1.create (1); | |
5362 | vec_oprnds2.create (1); | |
5363 | vec_oprnds3.create (1); | |
f7e531cf | 5364 | } |
ebfd146a IR |
5365 | |
5366 | /* Handle def. */ | |
5367 | scalar_dest = gimple_assign_lhs (stmt); | |
5368 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5369 | ||
5370 | /* Handle cond expr. */ | |
a855b1b1 MM |
5371 | for (j = 0; j < ncopies; j++) |
5372 | { | |
f7e531cf | 5373 | gimple new_stmt = NULL; |
a855b1b1 MM |
5374 | if (j == 0) |
5375 | { | |
f7e531cf IR |
5376 | if (slp_node) |
5377 | { | |
9771b263 DN |
5378 | vec<tree> ops; |
5379 | ops.create (4); | |
37b5ec8f | 5380 | vec<vec<tree> > vec_defs; |
9771b263 DN |
5381 | |
5382 | vec_defs.create (4); | |
5383 | ops.safe_push (TREE_OPERAND (cond_expr, 0)); | |
5384 | ops.safe_push (TREE_OPERAND (cond_expr, 1)); | |
5385 | ops.safe_push (then_clause); | |
5386 | ops.safe_push (else_clause); | |
f7e531cf | 5387 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); |
37b5ec8f JJ |
5388 | vec_oprnds3 = vec_defs.pop (); |
5389 | vec_oprnds2 = vec_defs.pop (); | |
5390 | vec_oprnds1 = vec_defs.pop (); | |
5391 | vec_oprnds0 = vec_defs.pop (); | |
f7e531cf | 5392 | |
9771b263 DN |
5393 | ops.release (); |
5394 | vec_defs.release (); | |
f7e531cf IR |
5395 | } |
5396 | else | |
5397 | { | |
5398 | gimple gtemp; | |
5399 | vec_cond_lhs = | |
a855b1b1 MM |
5400 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), |
5401 | stmt, NULL); | |
24ee1384 IR |
5402 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), stmt, |
5403 | loop_vinfo, NULL, >emp, &def, &dts[0]); | |
f7e531cf IR |
5404 | |
5405 | vec_cond_rhs = | |
5406 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
5407 | stmt, NULL); | |
24ee1384 IR |
5408 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), stmt, |
5409 | loop_vinfo, NULL, >emp, &def, &dts[1]); | |
f7e531cf IR |
5410 | if (reduc_index == 1) |
5411 | vec_then_clause = reduc_def; | |
5412 | else | |
5413 | { | |
5414 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
5415 | stmt, NULL); | |
24ee1384 | 5416 | vect_is_simple_use (then_clause, stmt, loop_vinfo, |
f7e531cf IR |
5417 | NULL, >emp, &def, &dts[2]); |
5418 | } | |
5419 | if (reduc_index == 2) | |
5420 | vec_else_clause = reduc_def; | |
5421 | else | |
5422 | { | |
5423 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
a855b1b1 | 5424 | stmt, NULL); |
24ee1384 | 5425 | vect_is_simple_use (else_clause, stmt, loop_vinfo, |
a855b1b1 | 5426 | NULL, >emp, &def, &dts[3]); |
f7e531cf | 5427 | } |
a855b1b1 MM |
5428 | } |
5429 | } | |
5430 | else | |
5431 | { | |
f7e531cf | 5432 | vec_cond_lhs = vect_get_vec_def_for_stmt_copy (dts[0], |
9771b263 | 5433 | vec_oprnds0.pop ()); |
f7e531cf | 5434 | vec_cond_rhs = vect_get_vec_def_for_stmt_copy (dts[1], |
9771b263 | 5435 | vec_oprnds1.pop ()); |
a855b1b1 | 5436 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 5437 | vec_oprnds2.pop ()); |
a855b1b1 | 5438 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 5439 | vec_oprnds3.pop ()); |
f7e531cf IR |
5440 | } |
5441 | ||
5442 | if (!slp_node) | |
5443 | { | |
9771b263 DN |
5444 | vec_oprnds0.quick_push (vec_cond_lhs); |
5445 | vec_oprnds1.quick_push (vec_cond_rhs); | |
5446 | vec_oprnds2.quick_push (vec_then_clause); | |
5447 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
5448 | } |
5449 | ||
9dc3f7de | 5450 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 5451 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 5452 | { |
9771b263 DN |
5453 | vec_cond_rhs = vec_oprnds1[i]; |
5454 | vec_then_clause = vec_oprnds2[i]; | |
5455 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 5456 | |
784fb9b3 JJ |
5457 | vec_compare = build2 (TREE_CODE (cond_expr), vec_cmp_type, |
5458 | vec_cond_lhs, vec_cond_rhs); | |
f7e531cf IR |
5459 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, |
5460 | vec_compare, vec_then_clause, vec_else_clause); | |
a855b1b1 | 5461 | |
f7e531cf IR |
5462 | new_stmt = gimple_build_assign (vec_dest, vec_cond_expr); |
5463 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5464 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5465 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5466 | if (slp_node) | |
9771b263 | 5467 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
5468 | } |
5469 | ||
5470 | if (slp_node) | |
5471 | continue; | |
5472 | ||
5473 | if (j == 0) | |
5474 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5475 | else | |
5476 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5477 | ||
5478 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 5479 | } |
b8698a0f | 5480 | |
9771b263 DN |
5481 | vec_oprnds0.release (); |
5482 | vec_oprnds1.release (); | |
5483 | vec_oprnds2.release (); | |
5484 | vec_oprnds3.release (); | |
f7e531cf | 5485 | |
ebfd146a IR |
5486 | return true; |
5487 | } | |
5488 | ||
5489 | ||
8644a673 | 5490 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
5491 | |
5492 | bool | |
a70d6342 | 5493 | vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 5494 | { |
8644a673 | 5495 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 5496 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 5497 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 5498 | bool ok; |
a70d6342 | 5499 | tree scalar_type, vectype; |
363477c0 JJ |
5500 | gimple pattern_stmt; |
5501 | gimple_seq pattern_def_seq; | |
ebfd146a | 5502 | |
73fbfcad | 5503 | if (dump_enabled_p ()) |
ebfd146a | 5504 | { |
78c60e3d SS |
5505 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
5506 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 5507 | } |
ebfd146a | 5508 | |
1825a1f3 | 5509 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 5510 | { |
73fbfcad | 5511 | if (dump_enabled_p ()) |
78c60e3d SS |
5512 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5513 | "not vectorized: stmt has volatile operands"); | |
1825a1f3 IR |
5514 | |
5515 | return false; | |
5516 | } | |
b8698a0f L |
5517 | |
5518 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
5519 | to include: |
5520 | - the COND_EXPR which is the loop exit condition | |
5521 | - any LABEL_EXPRs in the loop | |
b8698a0f | 5522 | - computations that are used only for array indexing or loop control. |
8644a673 | 5523 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 5524 | instance, therefore, all the statements are relevant. |
ebfd146a | 5525 | |
d092494c | 5526 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 5527 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
5528 | statements. In basic blocks we are called from some SLP instance |
5529 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
5530 | already will be part of SLP instance. */ | |
83197f37 IR |
5531 | |
5532 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 5533 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 5534 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 5535 | { |
9d5e7640 | 5536 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 5537 | && pattern_stmt |
9d5e7640 IR |
5538 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
5539 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
5540 | { | |
83197f37 | 5541 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
5542 | stmt = pattern_stmt; |
5543 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 5544 | if (dump_enabled_p ()) |
9d5e7640 | 5545 | { |
78c60e3d SS |
5546 | dump_printf_loc (MSG_NOTE, vect_location, |
5547 | "==> examining pattern statement: "); | |
5548 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
5549 | } |
5550 | } | |
5551 | else | |
5552 | { | |
73fbfcad | 5553 | if (dump_enabled_p ()) |
78c60e3d | 5554 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant."); |
ebfd146a | 5555 | |
9d5e7640 IR |
5556 | return true; |
5557 | } | |
8644a673 | 5558 | } |
83197f37 | 5559 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 5560 | && node == NULL |
83197f37 IR |
5561 | && pattern_stmt |
5562 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
5563 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
5564 | { | |
5565 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 5566 | if (dump_enabled_p ()) |
83197f37 | 5567 | { |
78c60e3d SS |
5568 | dump_printf_loc (MSG_NOTE, vect_location, |
5569 | "==> examining pattern statement: "); | |
5570 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
5571 | } |
5572 | ||
5573 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
5574 | return false; | |
5575 | } | |
ebfd146a | 5576 | |
1107f3ae | 5577 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 5578 | && node == NULL |
363477c0 | 5579 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 5580 | { |
363477c0 | 5581 | gimple_stmt_iterator si; |
1107f3ae | 5582 | |
363477c0 JJ |
5583 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
5584 | { | |
5585 | gimple pattern_def_stmt = gsi_stmt (si); | |
5586 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) | |
5587 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
5588 | { | |
5589 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 5590 | if (dump_enabled_p ()) |
363477c0 | 5591 | { |
78c60e3d SS |
5592 | dump_printf_loc (MSG_NOTE, vect_location, |
5593 | "==> examining pattern def statement: "); | |
5594 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 5595 | } |
1107f3ae | 5596 | |
363477c0 JJ |
5597 | if (!vect_analyze_stmt (pattern_def_stmt, |
5598 | need_to_vectorize, node)) | |
5599 | return false; | |
5600 | } | |
5601 | } | |
5602 | } | |
1107f3ae | 5603 | |
8644a673 IR |
5604 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
5605 | { | |
5606 | case vect_internal_def: | |
5607 | break; | |
ebfd146a | 5608 | |
8644a673 | 5609 | case vect_reduction_def: |
7c5222ff | 5610 | case vect_nested_cycle: |
a70d6342 | 5611 | gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer |
8644a673 | 5612 | || relevance == vect_used_in_outer_by_reduction |
a70d6342 | 5613 | || relevance == vect_unused_in_scope)); |
8644a673 IR |
5614 | break; |
5615 | ||
5616 | case vect_induction_def: | |
5617 | case vect_constant_def: | |
5618 | case vect_external_def: | |
5619 | case vect_unknown_def_type: | |
5620 | default: | |
5621 | gcc_unreachable (); | |
5622 | } | |
ebfd146a | 5623 | |
a70d6342 IR |
5624 | if (bb_vinfo) |
5625 | { | |
5626 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
5627 | ||
b690cc0f | 5628 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
73fbfcad | 5629 | if (dump_enabled_p ()) |
a70d6342 | 5630 | { |
78c60e3d SS |
5631 | dump_printf_loc (MSG_NOTE, vect_location, |
5632 | "get vectype for scalar type: "); | |
5633 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
a70d6342 IR |
5634 | } |
5635 | ||
5636 | vectype = get_vectype_for_scalar_type (scalar_type); | |
5637 | if (!vectype) | |
5638 | { | |
73fbfcad | 5639 | if (dump_enabled_p ()) |
a70d6342 | 5640 | { |
78c60e3d SS |
5641 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5642 | "not SLPed: unsupported data-type "); | |
5643 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
5644 | scalar_type); | |
a70d6342 IR |
5645 | } |
5646 | return false; | |
5647 | } | |
5648 | ||
73fbfcad | 5649 | if (dump_enabled_p ()) |
a70d6342 | 5650 | { |
78c60e3d SS |
5651 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); |
5652 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
a70d6342 IR |
5653 | } |
5654 | ||
5655 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
5656 | } | |
5657 | ||
8644a673 | 5658 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 5659 | { |
8644a673 IR |
5660 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
5661 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info)); | |
5662 | *need_to_vectorize = true; | |
ebfd146a IR |
5663 | } |
5664 | ||
8644a673 | 5665 | ok = true; |
b8698a0f | 5666 | if (!bb_vinfo |
a70d6342 IR |
5667 | && (STMT_VINFO_RELEVANT_P (stmt_info) |
5668 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
4a00c761 | 5669 | ok = (vectorizable_conversion (stmt, NULL, NULL, NULL) |
9dc3f7de | 5670 | || vectorizable_shift (stmt, NULL, NULL, NULL) |
8644a673 IR |
5671 | || vectorizable_operation (stmt, NULL, NULL, NULL) |
5672 | || vectorizable_assignment (stmt, NULL, NULL, NULL) | |
5673 | || vectorizable_load (stmt, NULL, NULL, NULL, NULL) | |
190c2236 | 5674 | || vectorizable_call (stmt, NULL, NULL, NULL) |
8644a673 | 5675 | || vectorizable_store (stmt, NULL, NULL, NULL) |
b5aeb3bb | 5676 | || vectorizable_reduction (stmt, NULL, NULL, NULL) |
f7e531cf | 5677 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, NULL)); |
a70d6342 IR |
5678 | else |
5679 | { | |
5680 | if (bb_vinfo) | |
4a00c761 JJ |
5681 | ok = (vectorizable_conversion (stmt, NULL, NULL, node) |
5682 | || vectorizable_shift (stmt, NULL, NULL, node) | |
9dc3f7de | 5683 | || vectorizable_operation (stmt, NULL, NULL, node) |
a70d6342 IR |
5684 | || vectorizable_assignment (stmt, NULL, NULL, node) |
5685 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
190c2236 | 5686 | || vectorizable_call (stmt, NULL, NULL, node) |
f7e531cf IR |
5687 | || vectorizable_store (stmt, NULL, NULL, node) |
5688 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node)); | |
b8698a0f | 5689 | } |
8644a673 IR |
5690 | |
5691 | if (!ok) | |
ebfd146a | 5692 | { |
73fbfcad | 5693 | if (dump_enabled_p ()) |
8644a673 | 5694 | { |
78c60e3d SS |
5695 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5696 | "not vectorized: relevant stmt not "); | |
5697 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
5698 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 5699 | } |
b8698a0f | 5700 | |
ebfd146a IR |
5701 | return false; |
5702 | } | |
5703 | ||
a70d6342 IR |
5704 | if (bb_vinfo) |
5705 | return true; | |
5706 | ||
8644a673 IR |
5707 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
5708 | need extra handling, except for vectorizable reductions. */ | |
5709 | if (STMT_VINFO_LIVE_P (stmt_info) | |
5710 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
5711 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 5712 | |
8644a673 | 5713 | if (!ok) |
ebfd146a | 5714 | { |
73fbfcad | 5715 | if (dump_enabled_p ()) |
8644a673 | 5716 | { |
78c60e3d SS |
5717 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5718 | "not vectorized: live stmt not "); | |
5719 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
5720 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 5721 | } |
b8698a0f | 5722 | |
8644a673 | 5723 | return false; |
ebfd146a IR |
5724 | } |
5725 | ||
ebfd146a IR |
5726 | return true; |
5727 | } | |
5728 | ||
5729 | ||
5730 | /* Function vect_transform_stmt. | |
5731 | ||
5732 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
5733 | ||
5734 | bool | |
5735 | vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi, | |
0d0293ac | 5736 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
5737 | slp_instance slp_node_instance) |
5738 | { | |
5739 | bool is_store = false; | |
5740 | gimple vec_stmt = NULL; | |
5741 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a | 5742 | bool done; |
ebfd146a IR |
5743 | |
5744 | switch (STMT_VINFO_TYPE (stmt_info)) | |
5745 | { | |
5746 | case type_demotion_vec_info_type: | |
ebfd146a | 5747 | case type_promotion_vec_info_type: |
ebfd146a IR |
5748 | case type_conversion_vec_info_type: |
5749 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
5750 | gcc_assert (done); | |
5751 | break; | |
5752 | ||
5753 | case induc_vec_info_type: | |
5754 | gcc_assert (!slp_node); | |
5755 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
5756 | gcc_assert (done); | |
5757 | break; | |
5758 | ||
9dc3f7de IR |
5759 | case shift_vec_info_type: |
5760 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
5761 | gcc_assert (done); | |
5762 | break; | |
5763 | ||
ebfd146a IR |
5764 | case op_vec_info_type: |
5765 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
5766 | gcc_assert (done); | |
5767 | break; | |
5768 | ||
5769 | case assignment_vec_info_type: | |
5770 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
5771 | gcc_assert (done); | |
5772 | break; | |
5773 | ||
5774 | case load_vec_info_type: | |
b8698a0f | 5775 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
5776 | slp_node_instance); |
5777 | gcc_assert (done); | |
5778 | break; | |
5779 | ||
5780 | case store_vec_info_type: | |
5781 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
5782 | gcc_assert (done); | |
0d0293ac | 5783 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
5784 | { |
5785 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 5786 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
5787 | one are skipped, and there vec_stmt_info shouldn't be freed |
5788 | meanwhile. */ | |
0d0293ac | 5789 | *grouped_store = true; |
ebfd146a IR |
5790 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
5791 | is_store = true; | |
5792 | } | |
5793 | else | |
5794 | is_store = true; | |
5795 | break; | |
5796 | ||
5797 | case condition_vec_info_type: | |
f7e531cf | 5798 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
5799 | gcc_assert (done); |
5800 | break; | |
5801 | ||
5802 | case call_vec_info_type: | |
190c2236 | 5803 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 5804 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
5805 | break; |
5806 | ||
5807 | case reduc_vec_info_type: | |
b5aeb3bb | 5808 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
5809 | gcc_assert (done); |
5810 | break; | |
5811 | ||
5812 | default: | |
5813 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
5814 | { | |
73fbfcad | 5815 | if (dump_enabled_p ()) |
78c60e3d SS |
5816 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5817 | "stmt not supported."); | |
ebfd146a IR |
5818 | gcc_unreachable (); |
5819 | } | |
5820 | } | |
5821 | ||
5822 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that | |
5823 | is being vectorized, but outside the immediately enclosing loop. */ | |
5824 | if (vec_stmt | |
a70d6342 IR |
5825 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
5826 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
5827 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
5828 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
5829 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 5830 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 5831 | vect_used_in_outer_by_reduction)) |
ebfd146a | 5832 | { |
a70d6342 IR |
5833 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
5834 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
5835 | imm_use_iterator imm_iter; |
5836 | use_operand_p use_p; | |
5837 | tree scalar_dest; | |
5838 | gimple exit_phi; | |
5839 | ||
73fbfcad | 5840 | if (dump_enabled_p ()) |
78c60e3d SS |
5841 | dump_printf_loc (MSG_NOTE, vect_location, |
5842 | "Record the vdef for outer-loop vectorization."); | |
ebfd146a IR |
5843 | |
5844 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
5845 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
5846 | STMT). */ | |
5847 | if (gimple_code (stmt) == GIMPLE_PHI) | |
5848 | scalar_dest = PHI_RESULT (stmt); | |
5849 | else | |
5850 | scalar_dest = gimple_assign_lhs (stmt); | |
5851 | ||
5852 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
5853 | { | |
5854 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
5855 | { | |
5856 | exit_phi = USE_STMT (use_p); | |
5857 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
5858 | } | |
5859 | } | |
5860 | } | |
5861 | ||
5862 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
5863 | being vectorized. */ | |
5864 | if (STMT_VINFO_LIVE_P (stmt_info) | |
5865 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
5866 | { | |
5867 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
5868 | gcc_assert (done); | |
5869 | } | |
5870 | ||
5871 | if (vec_stmt) | |
83197f37 | 5872 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 5873 | |
b8698a0f | 5874 | return is_store; |
ebfd146a IR |
5875 | } |
5876 | ||
5877 | ||
b8698a0f | 5878 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
5879 | stmt_vec_info. */ |
5880 | ||
5881 | void | |
5882 | vect_remove_stores (gimple first_stmt) | |
5883 | { | |
5884 | gimple next = first_stmt; | |
5885 | gimple tmp; | |
5886 | gimple_stmt_iterator next_si; | |
5887 | ||
5888 | while (next) | |
5889 | { | |
78048b1c JJ |
5890 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
5891 | ||
5892 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
5893 | if (is_pattern_stmt_p (stmt_info)) | |
5894 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
5895 | /* Free the attached stmt_vec_info and remove the stmt. */ |
5896 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 5897 | unlink_stmt_vdef (next); |
ebfd146a | 5898 | gsi_remove (&next_si, true); |
3d3f2249 | 5899 | release_defs (next); |
ebfd146a IR |
5900 | free_stmt_vec_info (next); |
5901 | next = tmp; | |
5902 | } | |
5903 | } | |
5904 | ||
5905 | ||
5906 | /* Function new_stmt_vec_info. | |
5907 | ||
5908 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
5909 | ||
5910 | stmt_vec_info | |
b8698a0f | 5911 | new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 5912 | bb_vec_info bb_vinfo) |
ebfd146a IR |
5913 | { |
5914 | stmt_vec_info res; | |
5915 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
5916 | ||
5917 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
5918 | STMT_VINFO_STMT (res) = stmt; | |
5919 | STMT_VINFO_LOOP_VINFO (res) = loop_vinfo; | |
a70d6342 | 5920 | STMT_VINFO_BB_VINFO (res) = bb_vinfo; |
8644a673 | 5921 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
5922 | STMT_VINFO_LIVE_P (res) = false; |
5923 | STMT_VINFO_VECTYPE (res) = NULL; | |
5924 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 5925 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
5926 | STMT_VINFO_IN_PATTERN_P (res) = false; |
5927 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 5928 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a IR |
5929 | STMT_VINFO_DATA_REF (res) = NULL; |
5930 | ||
5931 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
5932 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
5933 | STMT_VINFO_DR_INIT (res) = NULL; | |
5934 | STMT_VINFO_DR_STEP (res) = NULL; | |
5935 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
5936 | ||
5937 | if (gimple_code (stmt) == GIMPLE_PHI | |
5938 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
5939 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
5940 | else | |
8644a673 IR |
5941 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
5942 | ||
9771b263 | 5943 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 5944 | STMT_SLP_TYPE (res) = loop_vect; |
e14c1050 IR |
5945 | GROUP_FIRST_ELEMENT (res) = NULL; |
5946 | GROUP_NEXT_ELEMENT (res) = NULL; | |
5947 | GROUP_SIZE (res) = 0; | |
5948 | GROUP_STORE_COUNT (res) = 0; | |
5949 | GROUP_GAP (res) = 0; | |
5950 | GROUP_SAME_DR_STMT (res) = NULL; | |
5951 | GROUP_READ_WRITE_DEPENDENCE (res) = false; | |
ebfd146a IR |
5952 | |
5953 | return res; | |
5954 | } | |
5955 | ||
5956 | ||
5957 | /* Create a hash table for stmt_vec_info. */ | |
5958 | ||
5959 | void | |
5960 | init_stmt_vec_info_vec (void) | |
5961 | { | |
9771b263 DN |
5962 | gcc_assert (!stmt_vec_info_vec.exists ()); |
5963 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
5964 | } |
5965 | ||
5966 | ||
5967 | /* Free hash table for stmt_vec_info. */ | |
5968 | ||
5969 | void | |
5970 | free_stmt_vec_info_vec (void) | |
5971 | { | |
93675444 JJ |
5972 | unsigned int i; |
5973 | vec_void_p info; | |
5974 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) | |
5975 | if (info != NULL) | |
5976 | free_stmt_vec_info (STMT_VINFO_STMT ((stmt_vec_info) info)); | |
9771b263 DN |
5977 | gcc_assert (stmt_vec_info_vec.exists ()); |
5978 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
5979 | } |
5980 | ||
5981 | ||
5982 | /* Free stmt vectorization related info. */ | |
5983 | ||
5984 | void | |
5985 | free_stmt_vec_info (gimple stmt) | |
5986 | { | |
5987 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5988 | ||
5989 | if (!stmt_info) | |
5990 | return; | |
5991 | ||
78048b1c JJ |
5992 | /* Check if this statement has a related "pattern stmt" |
5993 | (introduced by the vectorizer during the pattern recognition | |
5994 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
5995 | too. */ | |
5996 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
5997 | { | |
5998 | stmt_vec_info patt_info | |
5999 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
6000 | if (patt_info) | |
6001 | { | |
363477c0 JJ |
6002 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
6003 | if (seq) | |
6004 | { | |
6005 | gimple_stmt_iterator si; | |
6006 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
6007 | free_stmt_vec_info (gsi_stmt (si)); | |
6008 | } | |
78048b1c JJ |
6009 | free_stmt_vec_info (STMT_VINFO_RELATED_STMT (stmt_info)); |
6010 | } | |
6011 | } | |
6012 | ||
9771b263 | 6013 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
ebfd146a IR |
6014 | set_vinfo_for_stmt (stmt, NULL); |
6015 | free (stmt_info); | |
6016 | } | |
6017 | ||
6018 | ||
bb67d9c7 | 6019 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 6020 | |
bb67d9c7 | 6021 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
6022 | by the target. */ |
6023 | ||
bb67d9c7 RG |
6024 | static tree |
6025 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a IR |
6026 | { |
6027 | enum machine_mode inner_mode = TYPE_MODE (scalar_type); | |
cc4b5170 | 6028 | enum machine_mode simd_mode; |
2f816591 | 6029 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
6030 | int nunits; |
6031 | tree vectype; | |
6032 | ||
cc4b5170 | 6033 | if (nbytes == 0) |
ebfd146a IR |
6034 | return NULL_TREE; |
6035 | ||
48f2e373 RB |
6036 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
6037 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
6038 | return NULL_TREE; | |
6039 | ||
7b7b1813 RG |
6040 | /* For vector types of elements whose mode precision doesn't |
6041 | match their types precision we use a element type of mode | |
6042 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
6043 | they support the proper result truncation/extension. |
6044 | We also make sure to build vector types with INTEGER_TYPE | |
6045 | component type only. */ | |
6d7971b8 | 6046 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
6047 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
6048 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
6049 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
6050 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 6051 | |
ccbf5bb4 RG |
6052 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
6053 | When the component mode passes the above test simply use a type | |
6054 | corresponding to that mode. The theory is that any use that | |
6055 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac RB |
6056 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
6057 | && !INTEGRAL_TYPE_P (scalar_type) | |
6058 | && !POINTER_TYPE_P (scalar_type)) | |
60b95d28 RB |
6059 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
6060 | ||
6061 | /* We can't build a vector type of elements with alignment bigger than | |
6062 | their size. */ | |
dfc2e2ac | 6063 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
6064 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
6065 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 6066 | |
dfc2e2ac RB |
6067 | /* If we felt back to using the mode fail if there was |
6068 | no scalar type for it. */ | |
6069 | if (scalar_type == NULL_TREE) | |
6070 | return NULL_TREE; | |
6071 | ||
bb67d9c7 RG |
6072 | /* If no size was supplied use the mode the target prefers. Otherwise |
6073 | lookup a vector mode of the specified size. */ | |
6074 | if (size == 0) | |
6075 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
6076 | else | |
6077 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
6078 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
6079 | if (nunits <= 1) | |
6080 | return NULL_TREE; | |
ebfd146a IR |
6081 | |
6082 | vectype = build_vector_type (scalar_type, nunits); | |
73fbfcad | 6083 | if (dump_enabled_p ()) |
ebfd146a | 6084 | { |
78c60e3d SS |
6085 | dump_printf_loc (MSG_NOTE, vect_location, |
6086 | "get vectype with %d units of type ", nunits); | |
6087 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
ebfd146a IR |
6088 | } |
6089 | ||
6090 | if (!vectype) | |
6091 | return NULL_TREE; | |
6092 | ||
73fbfcad | 6093 | if (dump_enabled_p ()) |
ebfd146a | 6094 | { |
78c60e3d SS |
6095 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); |
6096 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
ebfd146a IR |
6097 | } |
6098 | ||
6099 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
6100 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
6101 | { | |
73fbfcad | 6102 | if (dump_enabled_p ()) |
78c60e3d SS |
6103 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6104 | "mode not supported by target."); | |
ebfd146a IR |
6105 | return NULL_TREE; |
6106 | } | |
6107 | ||
6108 | return vectype; | |
6109 | } | |
6110 | ||
bb67d9c7 RG |
6111 | unsigned int current_vector_size; |
6112 | ||
6113 | /* Function get_vectype_for_scalar_type. | |
6114 | ||
6115 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
6116 | by the target. */ | |
6117 | ||
6118 | tree | |
6119 | get_vectype_for_scalar_type (tree scalar_type) | |
6120 | { | |
6121 | tree vectype; | |
6122 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
6123 | current_vector_size); | |
6124 | if (vectype | |
6125 | && current_vector_size == 0) | |
6126 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
6127 | return vectype; | |
6128 | } | |
6129 | ||
b690cc0f RG |
6130 | /* Function get_same_sized_vectype |
6131 | ||
6132 | Returns a vector type corresponding to SCALAR_TYPE of size | |
6133 | VECTOR_TYPE if supported by the target. */ | |
6134 | ||
6135 | tree | |
bb67d9c7 | 6136 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 6137 | { |
bb67d9c7 RG |
6138 | return get_vectype_for_scalar_type_and_size |
6139 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
6140 | } |
6141 | ||
ebfd146a IR |
6142 | /* Function vect_is_simple_use. |
6143 | ||
6144 | Input: | |
a70d6342 IR |
6145 | LOOP_VINFO - the vect info of the loop that is being vectorized. |
6146 | BB_VINFO - the vect info of the basic block that is being vectorized. | |
24ee1384 | 6147 | OPERAND - operand of STMT in the loop or bb. |
ebfd146a IR |
6148 | DEF - the defining stmt in case OPERAND is an SSA_NAME. |
6149 | ||
6150 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 6151 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 6152 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 6153 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
6154 | is the case in reduction/induction computations). |
6155 | For basic blocks, supportable operands are constants and bb invariants. | |
6156 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
6157 | |
6158 | bool | |
24ee1384 | 6159 | vect_is_simple_use (tree operand, gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 6160 | bb_vec_info bb_vinfo, gimple *def_stmt, |
ebfd146a | 6161 | tree *def, enum vect_def_type *dt) |
b8698a0f | 6162 | { |
ebfd146a IR |
6163 | basic_block bb; |
6164 | stmt_vec_info stmt_vinfo; | |
a70d6342 | 6165 | struct loop *loop = NULL; |
b8698a0f | 6166 | |
a70d6342 IR |
6167 | if (loop_vinfo) |
6168 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
6169 | |
6170 | *def_stmt = NULL; | |
6171 | *def = NULL_TREE; | |
b8698a0f | 6172 | |
73fbfcad | 6173 | if (dump_enabled_p ()) |
ebfd146a | 6174 | { |
78c60e3d SS |
6175 | dump_printf_loc (MSG_NOTE, vect_location, |
6176 | "vect_is_simple_use: operand "); | |
6177 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
ebfd146a | 6178 | } |
b8698a0f | 6179 | |
b758f602 | 6180 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
6181 | { |
6182 | *dt = vect_constant_def; | |
6183 | return true; | |
6184 | } | |
b8698a0f | 6185 | |
ebfd146a IR |
6186 | if (is_gimple_min_invariant (operand)) |
6187 | { | |
6188 | *def = operand; | |
8644a673 | 6189 | *dt = vect_external_def; |
ebfd146a IR |
6190 | return true; |
6191 | } | |
6192 | ||
6193 | if (TREE_CODE (operand) == PAREN_EXPR) | |
6194 | { | |
73fbfcad | 6195 | if (dump_enabled_p ()) |
78c60e3d | 6196 | dump_printf_loc (MSG_NOTE, vect_location, "non-associatable copy."); |
ebfd146a IR |
6197 | operand = TREE_OPERAND (operand, 0); |
6198 | } | |
b8698a0f | 6199 | |
ebfd146a IR |
6200 | if (TREE_CODE (operand) != SSA_NAME) |
6201 | { | |
73fbfcad | 6202 | if (dump_enabled_p ()) |
78c60e3d SS |
6203 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6204 | "not ssa-name."); | |
ebfd146a IR |
6205 | return false; |
6206 | } | |
b8698a0f | 6207 | |
ebfd146a IR |
6208 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
6209 | if (*def_stmt == NULL) | |
6210 | { | |
73fbfcad | 6211 | if (dump_enabled_p ()) |
78c60e3d SS |
6212 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6213 | "no def_stmt."); | |
ebfd146a IR |
6214 | return false; |
6215 | } | |
6216 | ||
73fbfcad | 6217 | if (dump_enabled_p ()) |
ebfd146a | 6218 | { |
78c60e3d SS |
6219 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
6220 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
6221 | } |
6222 | ||
8644a673 | 6223 | /* Empty stmt is expected only in case of a function argument. |
ebfd146a IR |
6224 | (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */ |
6225 | if (gimple_nop_p (*def_stmt)) | |
6226 | { | |
6227 | *def = operand; | |
8644a673 | 6228 | *dt = vect_external_def; |
ebfd146a IR |
6229 | return true; |
6230 | } | |
6231 | ||
6232 | bb = gimple_bb (*def_stmt); | |
a70d6342 IR |
6233 | |
6234 | if ((loop && !flow_bb_inside_loop_p (loop, bb)) | |
6235 | || (!loop && bb != BB_VINFO_BB (bb_vinfo)) | |
b8698a0f | 6236 | || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI)) |
8644a673 | 6237 | *dt = vect_external_def; |
ebfd146a IR |
6238 | else |
6239 | { | |
6240 | stmt_vinfo = vinfo_for_stmt (*def_stmt); | |
6241 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
6242 | } | |
6243 | ||
24ee1384 IR |
6244 | if (*dt == vect_unknown_def_type |
6245 | || (stmt | |
6246 | && *dt == vect_double_reduction_def | |
6247 | && gimple_code (stmt) != GIMPLE_PHI)) | |
ebfd146a | 6248 | { |
73fbfcad | 6249 | if (dump_enabled_p ()) |
78c60e3d SS |
6250 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6251 | "Unsupported pattern."); | |
ebfd146a IR |
6252 | return false; |
6253 | } | |
6254 | ||
73fbfcad | 6255 | if (dump_enabled_p ()) |
78c60e3d | 6256 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: %d.", *dt); |
ebfd146a IR |
6257 | |
6258 | switch (gimple_code (*def_stmt)) | |
6259 | { | |
6260 | case GIMPLE_PHI: | |
6261 | *def = gimple_phi_result (*def_stmt); | |
6262 | break; | |
6263 | ||
6264 | case GIMPLE_ASSIGN: | |
6265 | *def = gimple_assign_lhs (*def_stmt); | |
6266 | break; | |
6267 | ||
6268 | case GIMPLE_CALL: | |
6269 | *def = gimple_call_lhs (*def_stmt); | |
6270 | if (*def != NULL) | |
6271 | break; | |
6272 | /* FALLTHRU */ | |
6273 | default: | |
73fbfcad | 6274 | if (dump_enabled_p ()) |
78c60e3d SS |
6275 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6276 | "unsupported defining stmt: "); | |
ebfd146a IR |
6277 | return false; |
6278 | } | |
6279 | ||
6280 | return true; | |
6281 | } | |
6282 | ||
b690cc0f RG |
6283 | /* Function vect_is_simple_use_1. |
6284 | ||
6285 | Same as vect_is_simple_use_1 but also determines the vector operand | |
6286 | type of OPERAND and stores it to *VECTYPE. If the definition of | |
6287 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
6288 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
6289 | is responsible to compute the best suited vector type for the | |
6290 | scalar operand. */ | |
6291 | ||
6292 | bool | |
24ee1384 | 6293 | vect_is_simple_use_1 (tree operand, gimple stmt, loop_vec_info loop_vinfo, |
b690cc0f RG |
6294 | bb_vec_info bb_vinfo, gimple *def_stmt, |
6295 | tree *def, enum vect_def_type *dt, tree *vectype) | |
6296 | { | |
24ee1384 IR |
6297 | if (!vect_is_simple_use (operand, stmt, loop_vinfo, bb_vinfo, def_stmt, |
6298 | def, dt)) | |
b690cc0f RG |
6299 | return false; |
6300 | ||
6301 | /* Now get a vector type if the def is internal, otherwise supply | |
6302 | NULL_TREE and leave it up to the caller to figure out a proper | |
6303 | type for the use stmt. */ | |
6304 | if (*dt == vect_internal_def | |
6305 | || *dt == vect_induction_def | |
6306 | || *dt == vect_reduction_def | |
6307 | || *dt == vect_double_reduction_def | |
6308 | || *dt == vect_nested_cycle) | |
6309 | { | |
6310 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
6311 | |
6312 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
6313 | && !STMT_VINFO_RELEVANT (stmt_info) | |
6314 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 6315 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 6316 | |
b690cc0f RG |
6317 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
6318 | gcc_assert (*vectype != NULL_TREE); | |
6319 | } | |
6320 | else if (*dt == vect_uninitialized_def | |
6321 | || *dt == vect_constant_def | |
6322 | || *dt == vect_external_def) | |
6323 | *vectype = NULL_TREE; | |
6324 | else | |
6325 | gcc_unreachable (); | |
6326 | ||
6327 | return true; | |
6328 | } | |
6329 | ||
ebfd146a IR |
6330 | |
6331 | /* Function supportable_widening_operation | |
6332 | ||
b8698a0f L |
6333 | Check whether an operation represented by the code CODE is a |
6334 | widening operation that is supported by the target platform in | |
b690cc0f RG |
6335 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
6336 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 6337 | |
ebfd146a IR |
6338 | Widening operations we currently support are NOP (CONVERT), FLOAT |
6339 | and WIDEN_MULT. This function checks if these operations are supported | |
6340 | by the target platform either directly (via vector tree-codes), or via | |
6341 | target builtins. | |
6342 | ||
6343 | Output: | |
b8698a0f L |
6344 | - CODE1 and CODE2 are codes of vector operations to be used when |
6345 | vectorizing the operation, if available. | |
ebfd146a IR |
6346 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
6347 | case of multi-step conversion (like char->short->int - in that case | |
6348 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
6349 | - INTERM_TYPES contains the intermediate type required to perform the |
6350 | widening operation (short in the above example). */ | |
ebfd146a IR |
6351 | |
6352 | bool | |
b690cc0f RG |
6353 | supportable_widening_operation (enum tree_code code, gimple stmt, |
6354 | tree vectype_out, tree vectype_in, | |
ebfd146a IR |
6355 | enum tree_code *code1, enum tree_code *code2, |
6356 | int *multi_step_cvt, | |
9771b263 | 6357 | vec<tree> *interm_types) |
ebfd146a IR |
6358 | { |
6359 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6360 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 6361 | struct loop *vect_loop = NULL; |
ebfd146a | 6362 | enum machine_mode vec_mode; |
81f40b79 | 6363 | enum insn_code icode1, icode2; |
ebfd146a | 6364 | optab optab1, optab2; |
b690cc0f RG |
6365 | tree vectype = vectype_in; |
6366 | tree wide_vectype = vectype_out; | |
ebfd146a | 6367 | enum tree_code c1, c2; |
4a00c761 JJ |
6368 | int i; |
6369 | tree prev_type, intermediate_type; | |
6370 | enum machine_mode intermediate_mode, prev_mode; | |
6371 | optab optab3, optab4; | |
ebfd146a | 6372 | |
4a00c761 | 6373 | *multi_step_cvt = 0; |
4ef69dfc IR |
6374 | if (loop_info) |
6375 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
6376 | ||
ebfd146a IR |
6377 | switch (code) |
6378 | { | |
6379 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
6380 | /* The result of a vectorized widening operation usually requires |
6381 | two vectors (because the widened results do not fit into one vector). | |
6382 | The generated vector results would normally be expected to be | |
6383 | generated in the same order as in the original scalar computation, | |
6384 | i.e. if 8 results are generated in each vector iteration, they are | |
6385 | to be organized as follows: | |
6386 | vect1: [res1,res2,res3,res4], | |
6387 | vect2: [res5,res6,res7,res8]. | |
6388 | ||
6389 | However, in the special case that the result of the widening | |
6390 | operation is used in a reduction computation only, the order doesn't | |
6391 | matter (because when vectorizing a reduction we change the order of | |
6392 | the computation). Some targets can take advantage of this and | |
6393 | generate more efficient code. For example, targets like Altivec, | |
6394 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
6395 | generate the following vectors: | |
6396 | vect1: [res1,res3,res5,res7], | |
6397 | vect2: [res2,res4,res6,res8]. | |
6398 | ||
6399 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
6400 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
6401 | iterations in parallel). We therefore don't allow to change the | |
6402 | order of the computation in the inner-loop during outer-loop | |
6403 | vectorization. */ | |
6404 | /* TODO: Another case in which order doesn't *really* matter is when we | |
6405 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
6406 | Normally, pack_trunc performs an even/odd permute, whereas the | |
6407 | repack from an even/odd expansion would be an interleave, which | |
6408 | would be significantly simpler for e.g. AVX2. */ | |
6409 | /* In any case, in order to avoid duplicating the code below, recurse | |
6410 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
6411 | are properly set up for the caller. If we fail, we'll continue with | |
6412 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
6413 | if (vect_loop | |
6414 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
6415 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
6416 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
6417 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
6418 | code1, code2, multi_step_cvt, |
6419 | interm_types)) | |
6ae6116f | 6420 | return true; |
4a00c761 JJ |
6421 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
6422 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
6423 | break; |
6424 | ||
6ae6116f RH |
6425 | case VEC_WIDEN_MULT_EVEN_EXPR: |
6426 | /* Support the recursion induced just above. */ | |
6427 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
6428 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
6429 | break; | |
6430 | ||
36ba4aae | 6431 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
6432 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
6433 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
6434 | break; |
6435 | ||
ebfd146a | 6436 | CASE_CONVERT: |
4a00c761 JJ |
6437 | c1 = VEC_UNPACK_LO_EXPR; |
6438 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
6439 | break; |
6440 | ||
6441 | case FLOAT_EXPR: | |
4a00c761 JJ |
6442 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
6443 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
6444 | break; |
6445 | ||
6446 | case FIX_TRUNC_EXPR: | |
6447 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
6448 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
6449 | computing the operation. */ | |
6450 | return false; | |
6451 | ||
6452 | default: | |
6453 | gcc_unreachable (); | |
6454 | } | |
6455 | ||
6ae6116f | 6456 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
4a00c761 JJ |
6457 | { |
6458 | enum tree_code ctmp = c1; | |
6459 | c1 = c2; | |
6460 | c2 = ctmp; | |
6461 | } | |
6462 | ||
ebfd146a IR |
6463 | if (code == FIX_TRUNC_EXPR) |
6464 | { | |
6465 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
6466 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
6467 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
6468 | } |
6469 | else | |
6470 | { | |
6471 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
6472 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
6473 | } | |
6474 | ||
6475 | if (!optab1 || !optab2) | |
6476 | return false; | |
6477 | ||
6478 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
6479 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
6480 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
6481 | return false; |
6482 | ||
4a00c761 JJ |
6483 | *code1 = c1; |
6484 | *code2 = c2; | |
6485 | ||
6486 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
6487 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
6488 | return true; | |
6489 | ||
b8698a0f | 6490 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 6491 | types. */ |
ebfd146a | 6492 | |
4a00c761 JJ |
6493 | prev_type = vectype; |
6494 | prev_mode = vec_mode; | |
b8698a0f | 6495 | |
4a00c761 JJ |
6496 | if (!CONVERT_EXPR_CODE_P (code)) |
6497 | return false; | |
b8698a0f | 6498 | |
4a00c761 JJ |
6499 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
6500 | intermediate steps in promotion sequence. We try | |
6501 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
6502 | not. */ | |
9771b263 | 6503 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
6504 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
6505 | { | |
6506 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
6507 | intermediate_type | |
6508 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
6509 | TYPE_UNSIGNED (prev_type)); | |
6510 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
6511 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
6512 | ||
6513 | if (!optab3 || !optab4 | |
6514 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
6515 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
6516 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
6517 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
6518 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
6519 | == CODE_FOR_nothing) | |
6520 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
6521 | == CODE_FOR_nothing)) | |
6522 | break; | |
ebfd146a | 6523 | |
9771b263 | 6524 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
6525 | (*multi_step_cvt)++; |
6526 | ||
6527 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
6528 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
6529 | return true; | |
6530 | ||
6531 | prev_type = intermediate_type; | |
6532 | prev_mode = intermediate_mode; | |
ebfd146a IR |
6533 | } |
6534 | ||
9771b263 | 6535 | interm_types->release (); |
4a00c761 | 6536 | return false; |
ebfd146a IR |
6537 | } |
6538 | ||
6539 | ||
6540 | /* Function supportable_narrowing_operation | |
6541 | ||
b8698a0f L |
6542 | Check whether an operation represented by the code CODE is a |
6543 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
6544 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
6545 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 6546 | |
ebfd146a | 6547 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 6548 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
6549 | the target platform directly via vector tree-codes. |
6550 | ||
6551 | Output: | |
b8698a0f L |
6552 | - CODE1 is the code of a vector operation to be used when |
6553 | vectorizing the operation, if available. | |
ebfd146a IR |
6554 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
6555 | case of multi-step conversion (like int->short->char - in that case | |
6556 | MULTI_STEP_CVT will be 1). | |
6557 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 6558 | narrowing operation (short in the above example). */ |
ebfd146a IR |
6559 | |
6560 | bool | |
6561 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 6562 | tree vectype_out, tree vectype_in, |
ebfd146a | 6563 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 6564 | vec<tree> *interm_types) |
ebfd146a IR |
6565 | { |
6566 | enum machine_mode vec_mode; | |
6567 | enum insn_code icode1; | |
6568 | optab optab1, interm_optab; | |
b690cc0f RG |
6569 | tree vectype = vectype_in; |
6570 | tree narrow_vectype = vectype_out; | |
ebfd146a | 6571 | enum tree_code c1; |
4a00c761 JJ |
6572 | tree intermediate_type; |
6573 | enum machine_mode intermediate_mode, prev_mode; | |
ebfd146a | 6574 | int i; |
4a00c761 | 6575 | bool uns; |
ebfd146a | 6576 | |
4a00c761 | 6577 | *multi_step_cvt = 0; |
ebfd146a IR |
6578 | switch (code) |
6579 | { | |
6580 | CASE_CONVERT: | |
6581 | c1 = VEC_PACK_TRUNC_EXPR; | |
6582 | break; | |
6583 | ||
6584 | case FIX_TRUNC_EXPR: | |
6585 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
6586 | break; | |
6587 | ||
6588 | case FLOAT_EXPR: | |
6589 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
6590 | tree code and optabs used for computing the operation. */ | |
6591 | return false; | |
6592 | ||
6593 | default: | |
6594 | gcc_unreachable (); | |
6595 | } | |
6596 | ||
6597 | if (code == FIX_TRUNC_EXPR) | |
6598 | /* The signedness is determined from output operand. */ | |
b690cc0f | 6599 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
6600 | else |
6601 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
6602 | ||
6603 | if (!optab1) | |
6604 | return false; | |
6605 | ||
6606 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 6607 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
6608 | return false; |
6609 | ||
4a00c761 JJ |
6610 | *code1 = c1; |
6611 | ||
6612 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
6613 | return true; | |
6614 | ||
ebfd146a IR |
6615 | /* Check if it's a multi-step conversion that can be done using intermediate |
6616 | types. */ | |
4a00c761 JJ |
6617 | prev_mode = vec_mode; |
6618 | if (code == FIX_TRUNC_EXPR) | |
6619 | uns = TYPE_UNSIGNED (vectype_out); | |
6620 | else | |
6621 | uns = TYPE_UNSIGNED (vectype); | |
6622 | ||
6623 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
6624 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
6625 | costly than signed. */ | |
6626 | if (code == FIX_TRUNC_EXPR && uns) | |
6627 | { | |
6628 | enum insn_code icode2; | |
6629 | ||
6630 | intermediate_type | |
6631 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
6632 | interm_optab | |
6633 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 6634 | if (interm_optab != unknown_optab |
4a00c761 JJ |
6635 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
6636 | && insn_data[icode1].operand[0].mode | |
6637 | == insn_data[icode2].operand[0].mode) | |
6638 | { | |
6639 | uns = false; | |
6640 | optab1 = interm_optab; | |
6641 | icode1 = icode2; | |
6642 | } | |
6643 | } | |
ebfd146a | 6644 | |
4a00c761 JJ |
6645 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
6646 | intermediate steps in promotion sequence. We try | |
6647 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 6648 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
6649 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
6650 | { | |
6651 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
6652 | intermediate_type | |
6653 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
6654 | interm_optab | |
6655 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
6656 | optab_default); | |
6657 | if (!interm_optab | |
6658 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
6659 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
6660 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
6661 | == CODE_FOR_nothing)) | |
6662 | break; | |
6663 | ||
9771b263 | 6664 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
6665 | (*multi_step_cvt)++; |
6666 | ||
6667 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
6668 | return true; | |
6669 | ||
6670 | prev_mode = intermediate_mode; | |
6671 | optab1 = interm_optab; | |
ebfd146a IR |
6672 | } |
6673 | ||
9771b263 | 6674 | interm_types->release (); |
4a00c761 | 6675 | return false; |
ebfd146a | 6676 | } |