]>
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 | 26 | #include "tm.h" |
ebfd146a | 27 | #include "tree.h" |
d8a2d370 | 28 | #include "stor-layout.h" |
ebfd146a IR |
29 | #include "target.h" |
30 | #include "basic-block.h" | |
cf835838 | 31 | #include "gimple-pretty-print.h" |
2fb9a547 AM |
32 | #include "tree-ssa-alias.h" |
33 | #include "internal-fn.h" | |
34 | #include "tree-eh.h" | |
35 | #include "gimple-expr.h" | |
36 | #include "is-a.h" | |
18f429e2 | 37 | #include "gimple.h" |
45b0be94 | 38 | #include "gimplify.h" |
5be5c238 | 39 | #include "gimple-iterator.h" |
18f429e2 | 40 | #include "gimplify-me.h" |
442b4905 AM |
41 | #include "gimple-ssa.h" |
42 | #include "tree-cfg.h" | |
43 | #include "tree-phinodes.h" | |
44 | #include "ssa-iterators.h" | |
d8a2d370 | 45 | #include "stringpool.h" |
442b4905 | 46 | #include "tree-ssanames.h" |
e28030cf | 47 | #include "tree-ssa-loop-manip.h" |
ebfd146a | 48 | #include "cfgloop.h" |
0136f8f0 AH |
49 | #include "tree-ssa-loop.h" |
50 | #include "tree-scalar-evolution.h" | |
ebfd146a | 51 | #include "expr.h" |
7ee2468b | 52 | #include "recog.h" /* FIXME: for insn_data */ |
ebfd146a | 53 | #include "optabs.h" |
718f9c0f | 54 | #include "diagnostic-core.h" |
ebfd146a | 55 | #include "tree-vectorizer.h" |
7ee2468b | 56 | #include "dumpfile.h" |
0136f8f0 | 57 | #include "cgraph.h" |
ebfd146a | 58 | |
7ee2468b SB |
59 | /* For lang_hooks.types.type_for_mode. */ |
60 | #include "langhooks.h" | |
ebfd146a | 61 | |
c3e7ee41 BS |
62 | /* Return the vectorized type for the given statement. */ |
63 | ||
64 | tree | |
65 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
66 | { | |
67 | return STMT_VINFO_VECTYPE (stmt_info); | |
68 | } | |
69 | ||
70 | /* Return TRUE iff the given statement is in an inner loop relative to | |
71 | the loop being vectorized. */ | |
72 | bool | |
73 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
74 | { | |
75 | gimple stmt = STMT_VINFO_STMT (stmt_info); | |
76 | basic_block bb = gimple_bb (stmt); | |
77 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
78 | struct loop* loop; | |
79 | ||
80 | if (!loop_vinfo) | |
81 | return false; | |
82 | ||
83 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
84 | ||
85 | return (bb->loop_father == loop->inner); | |
86 | } | |
87 | ||
88 | /* Record the cost of a statement, either by directly informing the | |
89 | target model or by saving it in a vector for later processing. | |
90 | Return a preliminary estimate of the statement's cost. */ | |
91 | ||
92 | unsigned | |
92345349 | 93 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 94 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 95 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 96 | { |
92345349 | 97 | if (body_cost_vec) |
c3e7ee41 | 98 | { |
92345349 BS |
99 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
100 | add_stmt_info_to_vec (body_cost_vec, count, kind, | |
101 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
102 | misalign); | |
c3e7ee41 | 103 | return (unsigned) |
92345349 | 104 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
105 | |
106 | } | |
107 | else | |
108 | { | |
109 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
110 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
111 | void *target_cost_data; | |
112 | ||
113 | if (loop_vinfo) | |
114 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
115 | else | |
116 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
117 | ||
92345349 BS |
118 | return add_stmt_cost (target_cost_data, count, kind, stmt_info, |
119 | misalign, where); | |
c3e7ee41 BS |
120 | } |
121 | } | |
122 | ||
272c6793 RS |
123 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
124 | ||
125 | static tree | |
126 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
127 | { | |
128 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
129 | "vect_array"); | |
130 | } | |
131 | ||
132 | /* ARRAY is an array of vectors created by create_vector_array. | |
133 | Return an SSA_NAME for the vector in index N. The reference | |
134 | is part of the vectorization of STMT and the vector is associated | |
135 | with scalar destination SCALAR_DEST. */ | |
136 | ||
137 | static tree | |
138 | read_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree scalar_dest, | |
139 | tree array, unsigned HOST_WIDE_INT n) | |
140 | { | |
141 | tree vect_type, vect, vect_name, array_ref; | |
142 | gimple new_stmt; | |
143 | ||
144 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
145 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
146 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
147 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
148 | build_int_cst (size_type_node, n), | |
149 | NULL_TREE, NULL_TREE); | |
150 | ||
151 | new_stmt = gimple_build_assign (vect, array_ref); | |
152 | vect_name = make_ssa_name (vect, new_stmt); | |
153 | gimple_assign_set_lhs (new_stmt, vect_name); | |
154 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
155 | |
156 | return vect_name; | |
157 | } | |
158 | ||
159 | /* ARRAY is an array of vectors created by create_vector_array. | |
160 | Emit code to store SSA_NAME VECT in index N of the array. | |
161 | The store is part of the vectorization of STMT. */ | |
162 | ||
163 | static void | |
164 | write_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree vect, | |
165 | tree array, unsigned HOST_WIDE_INT n) | |
166 | { | |
167 | tree array_ref; | |
168 | gimple new_stmt; | |
169 | ||
170 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
171 | build_int_cst (size_type_node, n), | |
172 | NULL_TREE, NULL_TREE); | |
173 | ||
174 | new_stmt = gimple_build_assign (array_ref, vect); | |
175 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
176 | } |
177 | ||
178 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
179 | of *PTR. The memory reference replaces those in FIRST_DR | |
180 | (and its group). */ | |
181 | ||
182 | static tree | |
183 | create_array_ref (tree type, tree ptr, struct data_reference *first_dr) | |
184 | { | |
272c6793 RS |
185 | tree mem_ref, alias_ptr_type; |
186 | ||
187 | alias_ptr_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
188 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); | |
189 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 190 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
191 | return mem_ref; |
192 | } | |
193 | ||
ebfd146a IR |
194 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
195 | ||
196 | /* Function vect_mark_relevant. | |
197 | ||
198 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
199 | ||
200 | static void | |
9771b263 | 201 | vect_mark_relevant (vec<gimple> *worklist, gimple stmt, |
83197f37 IR |
202 | enum vect_relevant relevant, bool live_p, |
203 | bool used_in_pattern) | |
ebfd146a IR |
204 | { |
205 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
206 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
207 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
83197f37 | 208 | gimple pattern_stmt; |
ebfd146a | 209 | |
73fbfcad | 210 | if (dump_enabled_p ()) |
78c60e3d | 211 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 212 | "mark relevant %d, live %d.\n", relevant, live_p); |
ebfd146a | 213 | |
83197f37 IR |
214 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
215 | related pattern stmt instead of the original stmt. However, such stmts | |
216 | may have their own uses that are not in any pattern, in such cases the | |
217 | stmt itself should be marked. */ | |
ebfd146a IR |
218 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
219 | { | |
83197f37 IR |
220 | bool found = false; |
221 | if (!used_in_pattern) | |
222 | { | |
223 | imm_use_iterator imm_iter; | |
224 | use_operand_p use_p; | |
225 | gimple use_stmt; | |
226 | tree lhs; | |
13c931c9 JJ |
227 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
228 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a | 229 | |
83197f37 IR |
230 | if (is_gimple_assign (stmt)) |
231 | lhs = gimple_assign_lhs (stmt); | |
232 | else | |
233 | lhs = gimple_call_lhs (stmt); | |
ebfd146a | 234 | |
83197f37 IR |
235 | /* This use is out of pattern use, if LHS has other uses that are |
236 | pattern uses, we should mark the stmt itself, and not the pattern | |
237 | stmt. */ | |
5ce9450f | 238 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
ab0ef706 JJ |
239 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs) |
240 | { | |
241 | if (is_gimple_debug (USE_STMT (use_p))) | |
242 | continue; | |
243 | use_stmt = USE_STMT (use_p); | |
244 | ||
13c931c9 JJ |
245 | if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) |
246 | continue; | |
247 | ||
ab0ef706 JJ |
248 | if (vinfo_for_stmt (use_stmt) |
249 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt))) | |
250 | { | |
251 | found = true; | |
252 | break; | |
253 | } | |
254 | } | |
83197f37 IR |
255 | } |
256 | ||
257 | if (!found) | |
258 | { | |
259 | /* This is the last stmt in a sequence that was detected as a | |
260 | pattern that can potentially be vectorized. Don't mark the stmt | |
261 | as relevant/live because it's not going to be vectorized. | |
262 | Instead mark the pattern-stmt that replaces it. */ | |
263 | ||
264 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
265 | ||
73fbfcad | 266 | if (dump_enabled_p ()) |
78c60e3d SS |
267 | dump_printf_loc (MSG_NOTE, vect_location, |
268 | "last stmt in pattern. don't mark" | |
e645e942 | 269 | " relevant/live.\n"); |
83197f37 IR |
270 | stmt_info = vinfo_for_stmt (pattern_stmt); |
271 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
272 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
273 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
274 | stmt = pattern_stmt; | |
275 | } | |
ebfd146a IR |
276 | } |
277 | ||
278 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
279 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
280 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
281 | ||
282 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
283 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
284 | { | |
73fbfcad | 285 | if (dump_enabled_p ()) |
78c60e3d | 286 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 287 | "already marked relevant/live.\n"); |
ebfd146a IR |
288 | return; |
289 | } | |
290 | ||
9771b263 | 291 | worklist->safe_push (stmt); |
ebfd146a IR |
292 | } |
293 | ||
294 | ||
295 | /* Function vect_stmt_relevant_p. | |
296 | ||
297 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
298 | "relevant for vectorization". | |
299 | ||
300 | A stmt is considered "relevant for vectorization" if: | |
301 | - it has uses outside the loop. | |
302 | - it has vdefs (it alters memory). | |
303 | - control stmts in the loop (except for the exit condition). | |
304 | ||
305 | CHECKME: what other side effects would the vectorizer allow? */ | |
306 | ||
307 | static bool | |
308 | vect_stmt_relevant_p (gimple stmt, loop_vec_info loop_vinfo, | |
309 | enum vect_relevant *relevant, bool *live_p) | |
310 | { | |
311 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
312 | ssa_op_iter op_iter; | |
313 | imm_use_iterator imm_iter; | |
314 | use_operand_p use_p; | |
315 | def_operand_p def_p; | |
316 | ||
8644a673 | 317 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
318 | *live_p = false; |
319 | ||
320 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
321 | if (is_ctrl_stmt (stmt) |
322 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
323 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 324 | *relevant = vect_used_in_scope; |
ebfd146a IR |
325 | |
326 | /* changing memory. */ | |
327 | if (gimple_code (stmt) != GIMPLE_PHI) | |
5006671f | 328 | if (gimple_vdef (stmt)) |
ebfd146a | 329 | { |
73fbfcad | 330 | if (dump_enabled_p ()) |
78c60e3d | 331 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 332 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 333 | *relevant = vect_used_in_scope; |
ebfd146a IR |
334 | } |
335 | ||
336 | /* uses outside the loop. */ | |
337 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
338 | { | |
339 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
340 | { | |
341 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
342 | if (!flow_bb_inside_loop_p (loop, bb)) | |
343 | { | |
73fbfcad | 344 | if (dump_enabled_p ()) |
78c60e3d | 345 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 346 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 347 | |
3157b0c2 AO |
348 | if (is_gimple_debug (USE_STMT (use_p))) |
349 | continue; | |
350 | ||
ebfd146a IR |
351 | /* We expect all such uses to be in the loop exit phis |
352 | (because of loop closed form) */ | |
353 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
354 | gcc_assert (bb == single_exit (loop)->dest); | |
355 | ||
356 | *live_p = true; | |
357 | } | |
358 | } | |
359 | } | |
360 | ||
361 | return (*live_p || *relevant); | |
362 | } | |
363 | ||
364 | ||
b8698a0f | 365 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 366 | |
ff802fa1 | 367 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
368 | used in STMT for anything other than indexing an array. */ |
369 | ||
370 | static bool | |
371 | exist_non_indexing_operands_for_use_p (tree use, gimple stmt) | |
372 | { | |
373 | tree operand; | |
374 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 375 | |
ff802fa1 | 376 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
377 | reference in STMT, then any operand that corresponds to USE |
378 | is not indexing an array. */ | |
379 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
380 | return true; | |
59a05b0c | 381 | |
ebfd146a IR |
382 | /* STMT has a data_ref. FORNOW this means that its of one of |
383 | the following forms: | |
384 | -1- ARRAY_REF = var | |
385 | -2- var = ARRAY_REF | |
386 | (This should have been verified in analyze_data_refs). | |
387 | ||
388 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 389 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
390 | for array indexing. |
391 | ||
392 | Therefore, all we need to check is if STMT falls into the | |
393 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
394 | |
395 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
396 | { |
397 | if (is_gimple_call (stmt) | |
398 | && gimple_call_internal_p (stmt)) | |
399 | switch (gimple_call_internal_fn (stmt)) | |
400 | { | |
401 | case IFN_MASK_STORE: | |
402 | operand = gimple_call_arg (stmt, 3); | |
403 | if (operand == use) | |
404 | return true; | |
405 | /* FALLTHRU */ | |
406 | case IFN_MASK_LOAD: | |
407 | operand = gimple_call_arg (stmt, 2); | |
408 | if (operand == use) | |
409 | return true; | |
410 | break; | |
411 | default: | |
412 | break; | |
413 | } | |
414 | return false; | |
415 | } | |
416 | ||
59a05b0c EB |
417 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
418 | return false; | |
ebfd146a | 419 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
420 | if (TREE_CODE (operand) != SSA_NAME) |
421 | return false; | |
422 | ||
423 | if (operand == use) | |
424 | return true; | |
425 | ||
426 | return false; | |
427 | } | |
428 | ||
429 | ||
b8698a0f | 430 | /* |
ebfd146a IR |
431 | Function process_use. |
432 | ||
433 | Inputs: | |
434 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 435 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ff802fa1 | 436 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 437 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
438 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
439 | be performed. | |
ebfd146a IR |
440 | |
441 | Outputs: | |
442 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
443 | relevance info of the DEF_STMT of this USE: | |
444 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
445 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
446 | Exceptions: | |
447 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 448 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 449 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
450 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
451 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
452 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
453 | be modified accordingly. | |
454 | ||
455 | Return true if everything is as expected. Return false otherwise. */ | |
456 | ||
457 | static bool | |
b8698a0f | 458 | process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
9771b263 | 459 | enum vect_relevant relevant, vec<gimple> *worklist, |
aec7ae7d | 460 | bool force) |
ebfd146a IR |
461 | { |
462 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
463 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
464 | stmt_vec_info dstmt_vinfo; | |
465 | basic_block bb, def_bb; | |
466 | tree def; | |
467 | gimple def_stmt; | |
468 | enum vect_def_type dt; | |
469 | ||
b8698a0f | 470 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 471 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 472 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
473 | return true; |
474 | ||
24ee1384 | 475 | if (!vect_is_simple_use (use, stmt, loop_vinfo, NULL, &def_stmt, &def, &dt)) |
b8698a0f | 476 | { |
73fbfcad | 477 | if (dump_enabled_p ()) |
78c60e3d | 478 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 479 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
480 | return false; |
481 | } | |
482 | ||
483 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
484 | return true; | |
485 | ||
486 | def_bb = gimple_bb (def_stmt); | |
487 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
488 | { | |
73fbfcad | 489 | if (dump_enabled_p ()) |
e645e942 | 490 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
491 | return true; |
492 | } | |
493 | ||
b8698a0f L |
494 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
495 | DEF_STMT must have already been processed, because this should be the | |
496 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
497 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
498 | check that everything is as expected, and we are done. */ |
499 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
500 | bb = gimple_bb (stmt); | |
501 | if (gimple_code (stmt) == GIMPLE_PHI | |
502 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
503 | && gimple_code (def_stmt) != GIMPLE_PHI | |
504 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
505 | && bb->loop_father == def_bb->loop_father) | |
506 | { | |
73fbfcad | 507 | if (dump_enabled_p ()) |
78c60e3d | 508 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 509 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
510 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
511 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
512 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 513 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 514 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
515 | return true; |
516 | } | |
517 | ||
518 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
519 | outer-loop-header-bb: | |
520 | d = def_stmt | |
521 | inner-loop: | |
522 | stmt # use (d) | |
523 | outer-loop-tail-bb: | |
524 | ... */ | |
525 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
526 | { | |
73fbfcad | 527 | if (dump_enabled_p ()) |
78c60e3d | 528 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 529 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 530 | |
ebfd146a IR |
531 | switch (relevant) |
532 | { | |
8644a673 | 533 | case vect_unused_in_scope: |
7c5222ff IR |
534 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
535 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 536 | break; |
7c5222ff | 537 | |
ebfd146a | 538 | case vect_used_in_outer_by_reduction: |
7c5222ff | 539 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
540 | relevant = vect_used_by_reduction; |
541 | break; | |
7c5222ff | 542 | |
ebfd146a | 543 | case vect_used_in_outer: |
7c5222ff | 544 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 545 | relevant = vect_used_in_scope; |
ebfd146a | 546 | break; |
7c5222ff | 547 | |
8644a673 | 548 | case vect_used_in_scope: |
ebfd146a IR |
549 | break; |
550 | ||
551 | default: | |
552 | gcc_unreachable (); | |
b8698a0f | 553 | } |
ebfd146a IR |
554 | } |
555 | ||
556 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
557 | outer-loop-header-bb: | |
558 | ... | |
559 | inner-loop: | |
560 | d = def_stmt | |
06066f92 | 561 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
562 | stmt # use (d) */ |
563 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
564 | { | |
73fbfcad | 565 | if (dump_enabled_p ()) |
78c60e3d | 566 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 567 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 568 | |
ebfd146a IR |
569 | switch (relevant) |
570 | { | |
8644a673 | 571 | case vect_unused_in_scope: |
b8698a0f | 572 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 573 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 574 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
575 | break; |
576 | ||
ebfd146a IR |
577 | case vect_used_by_reduction: |
578 | relevant = vect_used_in_outer_by_reduction; | |
579 | break; | |
580 | ||
8644a673 | 581 | case vect_used_in_scope: |
ebfd146a IR |
582 | relevant = vect_used_in_outer; |
583 | break; | |
584 | ||
585 | default: | |
586 | gcc_unreachable (); | |
587 | } | |
588 | } | |
589 | ||
83197f37 IR |
590 | vect_mark_relevant (worklist, def_stmt, relevant, live_p, |
591 | is_pattern_stmt_p (stmt_vinfo)); | |
ebfd146a IR |
592 | return true; |
593 | } | |
594 | ||
595 | ||
596 | /* Function vect_mark_stmts_to_be_vectorized. | |
597 | ||
598 | Not all stmts in the loop need to be vectorized. For example: | |
599 | ||
600 | for i... | |
601 | for j... | |
602 | 1. T0 = i + j | |
603 | 2. T1 = a[T0] | |
604 | ||
605 | 3. j = j + 1 | |
606 | ||
607 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
608 | addressing of vectorized data-refs are handled differently. | |
609 | ||
610 | This pass detects such stmts. */ | |
611 | ||
612 | bool | |
613 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
614 | { | |
ebfd146a IR |
615 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
616 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
617 | unsigned int nbbs = loop->num_nodes; | |
618 | gimple_stmt_iterator si; | |
619 | gimple stmt; | |
620 | unsigned int i; | |
621 | stmt_vec_info stmt_vinfo; | |
622 | basic_block bb; | |
623 | gimple phi; | |
624 | bool live_p; | |
06066f92 IR |
625 | enum vect_relevant relevant, tmp_relevant; |
626 | enum vect_def_type def_type; | |
ebfd146a | 627 | |
73fbfcad | 628 | if (dump_enabled_p ()) |
78c60e3d | 629 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 630 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 631 | |
07687835 | 632 | stack_vec<gimple, 64> worklist; |
ebfd146a IR |
633 | |
634 | /* 1. Init worklist. */ | |
635 | for (i = 0; i < nbbs; i++) | |
636 | { | |
637 | bb = bbs[i]; | |
638 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 639 | { |
ebfd146a | 640 | phi = gsi_stmt (si); |
73fbfcad | 641 | if (dump_enabled_p ()) |
ebfd146a | 642 | { |
78c60e3d SS |
643 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
644 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
e645e942 | 645 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
646 | } |
647 | ||
648 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 649 | vect_mark_relevant (&worklist, phi, relevant, live_p, false); |
ebfd146a IR |
650 | } |
651 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
652 | { | |
653 | stmt = gsi_stmt (si); | |
73fbfcad | 654 | if (dump_enabled_p ()) |
ebfd146a | 655 | { |
78c60e3d SS |
656 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
657 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 658 | dump_printf (MSG_NOTE, "\n"); |
b8698a0f | 659 | } |
ebfd146a IR |
660 | |
661 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 662 | vect_mark_relevant (&worklist, stmt, relevant, live_p, false); |
ebfd146a IR |
663 | } |
664 | } | |
665 | ||
666 | /* 2. Process_worklist */ | |
9771b263 | 667 | while (worklist.length () > 0) |
ebfd146a IR |
668 | { |
669 | use_operand_p use_p; | |
670 | ssa_op_iter iter; | |
671 | ||
9771b263 | 672 | stmt = worklist.pop (); |
73fbfcad | 673 | if (dump_enabled_p ()) |
ebfd146a | 674 | { |
78c60e3d SS |
675 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
676 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 677 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
678 | } |
679 | ||
b8698a0f L |
680 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
681 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
682 | liveness and relevance properties of STMT. */ |
683 | stmt_vinfo = vinfo_for_stmt (stmt); | |
684 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
685 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
686 | ||
687 | /* Generally, the liveness and relevance properties of STMT are | |
688 | propagated as is to the DEF_STMTs of its USEs: | |
689 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
690 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
691 | ||
692 | One exception is when STMT has been identified as defining a reduction | |
693 | variable; in this case we set the liveness/relevance as follows: | |
694 | live_p = false | |
695 | relevant = vect_used_by_reduction | |
696 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f | 697 | those that are used by a reduction computation, and those that are |
ff802fa1 | 698 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 699 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 700 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 701 | |
06066f92 IR |
702 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
703 | tmp_relevant = relevant; | |
704 | switch (def_type) | |
ebfd146a | 705 | { |
06066f92 IR |
706 | case vect_reduction_def: |
707 | switch (tmp_relevant) | |
708 | { | |
709 | case vect_unused_in_scope: | |
710 | relevant = vect_used_by_reduction; | |
711 | break; | |
712 | ||
713 | case vect_used_by_reduction: | |
714 | if (gimple_code (stmt) == GIMPLE_PHI) | |
715 | break; | |
716 | /* fall through */ | |
717 | ||
718 | default: | |
73fbfcad | 719 | if (dump_enabled_p ()) |
78c60e3d | 720 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 721 | "unsupported use of reduction.\n"); |
06066f92 IR |
722 | return false; |
723 | } | |
724 | ||
b8698a0f | 725 | live_p = false; |
06066f92 | 726 | break; |
b8698a0f | 727 | |
06066f92 IR |
728 | case vect_nested_cycle: |
729 | if (tmp_relevant != vect_unused_in_scope | |
730 | && tmp_relevant != vect_used_in_outer_by_reduction | |
731 | && tmp_relevant != vect_used_in_outer) | |
732 | { | |
73fbfcad | 733 | if (dump_enabled_p ()) |
78c60e3d | 734 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 735 | "unsupported use of nested cycle.\n"); |
7c5222ff | 736 | |
06066f92 IR |
737 | return false; |
738 | } | |
7c5222ff | 739 | |
b8698a0f L |
740 | live_p = false; |
741 | break; | |
742 | ||
06066f92 IR |
743 | case vect_double_reduction_def: |
744 | if (tmp_relevant != vect_unused_in_scope | |
745 | && tmp_relevant != vect_used_by_reduction) | |
746 | { | |
73fbfcad | 747 | if (dump_enabled_p ()) |
78c60e3d | 748 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 749 | "unsupported use of double reduction.\n"); |
7c5222ff | 750 | |
7c5222ff | 751 | return false; |
06066f92 IR |
752 | } |
753 | ||
754 | live_p = false; | |
b8698a0f | 755 | break; |
7c5222ff | 756 | |
06066f92 IR |
757 | default: |
758 | break; | |
7c5222ff | 759 | } |
b8698a0f | 760 | |
aec7ae7d | 761 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
762 | { |
763 | /* Pattern statements are not inserted into the code, so | |
764 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
765 | have to scan the RHS or function arguments instead. */ | |
766 | if (is_gimple_assign (stmt)) | |
767 | { | |
69d2aade JJ |
768 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
769 | tree op = gimple_assign_rhs1 (stmt); | |
770 | ||
771 | i = 1; | |
772 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
773 | { | |
774 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
aec7ae7d | 775 | live_p, relevant, &worklist, false) |
69d2aade | 776 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
aec7ae7d | 777 | live_p, relevant, &worklist, false)) |
07687835 | 778 | return false; |
69d2aade JJ |
779 | i = 2; |
780 | } | |
781 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 782 | { |
69d2aade | 783 | op = gimple_op (stmt, i); |
9d5e7640 | 784 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, |
aec7ae7d | 785 | &worklist, false)) |
07687835 | 786 | return false; |
9d5e7640 IR |
787 | } |
788 | } | |
789 | else if (is_gimple_call (stmt)) | |
790 | { | |
791 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
792 | { | |
793 | tree arg = gimple_call_arg (stmt, i); | |
794 | if (!process_use (stmt, arg, loop_vinfo, live_p, relevant, | |
aec7ae7d | 795 | &worklist, false)) |
07687835 | 796 | return false; |
9d5e7640 IR |
797 | } |
798 | } | |
799 | } | |
800 | else | |
801 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
802 | { | |
803 | tree op = USE_FROM_PTR (use_p); | |
804 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, | |
aec7ae7d | 805 | &worklist, false)) |
07687835 | 806 | return false; |
9d5e7640 | 807 | } |
aec7ae7d JJ |
808 | |
809 | if (STMT_VINFO_GATHER_P (stmt_vinfo)) | |
810 | { | |
811 | tree off; | |
812 | tree decl = vect_check_gather (stmt, loop_vinfo, NULL, &off, NULL); | |
813 | gcc_assert (decl); | |
814 | if (!process_use (stmt, off, loop_vinfo, live_p, relevant, | |
815 | &worklist, true)) | |
07687835 | 816 | return false; |
aec7ae7d | 817 | } |
ebfd146a IR |
818 | } /* while worklist */ |
819 | ||
ebfd146a IR |
820 | return true; |
821 | } | |
822 | ||
823 | ||
b8698a0f | 824 | /* Function vect_model_simple_cost. |
ebfd146a | 825 | |
b8698a0f | 826 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
827 | single op. Right now, this does not account for multiple insns that could |
828 | be generated for the single vector op. We will handle that shortly. */ | |
829 | ||
830 | void | |
b8698a0f | 831 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 BS |
832 | enum vect_def_type *dt, |
833 | stmt_vector_for_cost *prologue_cost_vec, | |
834 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
835 | { |
836 | int i; | |
92345349 | 837 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
838 | |
839 | /* The SLP costs were already calculated during SLP tree build. */ | |
840 | if (PURE_SLP_STMT (stmt_info)) | |
841 | return; | |
842 | ||
ebfd146a IR |
843 | /* FORNOW: Assuming maximum 2 args per stmts. */ |
844 | for (i = 0; i < 2; i++) | |
92345349 BS |
845 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
846 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, vector_stmt, | |
847 | stmt_info, 0, vect_prologue); | |
c3e7ee41 BS |
848 | |
849 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
850 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
851 | stmt_info, 0, vect_body); | |
c3e7ee41 | 852 | |
73fbfcad | 853 | if (dump_enabled_p ()) |
78c60e3d SS |
854 | dump_printf_loc (MSG_NOTE, vect_location, |
855 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 856 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
857 | } |
858 | ||
859 | ||
8bd37302 BS |
860 | /* Model cost for type demotion and promotion operations. PWR is normally |
861 | zero for single-step promotions and demotions. It will be one if | |
862 | two-step promotion/demotion is required, and so on. Each additional | |
863 | step doubles the number of instructions required. */ | |
864 | ||
865 | static void | |
866 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
867 | enum vect_def_type *dt, int pwr) | |
868 | { | |
869 | int i, tmp; | |
92345349 | 870 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
871 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
872 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
873 | void *target_cost_data; | |
8bd37302 BS |
874 | |
875 | /* The SLP costs were already calculated during SLP tree build. */ | |
876 | if (PURE_SLP_STMT (stmt_info)) | |
877 | return; | |
878 | ||
c3e7ee41 BS |
879 | if (loop_vinfo) |
880 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
881 | else | |
882 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
883 | ||
8bd37302 BS |
884 | for (i = 0; i < pwr + 1; i++) |
885 | { | |
886 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
887 | (i + 1) : i; | |
c3e7ee41 | 888 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
889 | vec_promote_demote, stmt_info, 0, |
890 | vect_body); | |
8bd37302 BS |
891 | } |
892 | ||
893 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
894 | for (i = 0; i < 2; i++) | |
92345349 BS |
895 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
896 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
897 | stmt_info, 0, vect_prologue); | |
8bd37302 | 898 | |
73fbfcad | 899 | if (dump_enabled_p ()) |
78c60e3d SS |
900 | dump_printf_loc (MSG_NOTE, vect_location, |
901 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 902 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
903 | } |
904 | ||
0d0293ac | 905 | /* Function vect_cost_group_size |
b8698a0f | 906 | |
0d0293ac | 907 | For grouped load or store, return the group_size only if it is the first |
ebfd146a IR |
908 | load or store of a group, else return 1. This ensures that group size is |
909 | only returned once per group. */ | |
910 | ||
911 | static int | |
0d0293ac | 912 | vect_cost_group_size (stmt_vec_info stmt_info) |
ebfd146a | 913 | { |
e14c1050 | 914 | gimple first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
915 | |
916 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
e14c1050 | 917 | return GROUP_SIZE (stmt_info); |
ebfd146a IR |
918 | |
919 | return 1; | |
920 | } | |
921 | ||
922 | ||
923 | /* Function vect_model_store_cost | |
924 | ||
0d0293ac MM |
925 | Models cost for stores. In the case of grouped accesses, one access |
926 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
927 | |
928 | void | |
b8698a0f | 929 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
272c6793 | 930 | bool store_lanes_p, enum vect_def_type dt, |
92345349 BS |
931 | slp_tree slp_node, |
932 | stmt_vector_for_cost *prologue_cost_vec, | |
933 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
934 | { |
935 | int group_size; | |
92345349 | 936 | unsigned int inside_cost = 0, prologue_cost = 0; |
720f5239 IR |
937 | struct data_reference *first_dr; |
938 | gimple first_stmt; | |
ebfd146a IR |
939 | |
940 | /* The SLP costs were already calculated during SLP tree build. */ | |
941 | if (PURE_SLP_STMT (stmt_info)) | |
942 | return; | |
943 | ||
8644a673 | 944 | if (dt == vect_constant_def || dt == vect_external_def) |
92345349 BS |
945 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
946 | stmt_info, 0, vect_prologue); | |
ebfd146a | 947 | |
0d0293ac MM |
948 | /* Grouped access? */ |
949 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
720f5239 IR |
950 | { |
951 | if (slp_node) | |
952 | { | |
9771b263 | 953 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
720f5239 IR |
954 | group_size = 1; |
955 | } | |
956 | else | |
957 | { | |
e14c1050 | 958 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 959 | group_size = vect_cost_group_size (stmt_info); |
720f5239 IR |
960 | } |
961 | ||
962 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
963 | } | |
0d0293ac | 964 | /* Not a grouped access. */ |
ebfd146a | 965 | else |
720f5239 IR |
966 | { |
967 | group_size = 1; | |
968 | first_dr = STMT_VINFO_DATA_REF (stmt_info); | |
969 | } | |
ebfd146a | 970 | |
272c6793 | 971 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 972 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 RS |
973 | access is instead being provided by a permute-and-store operation, |
974 | include the cost of the permutes. */ | |
975 | if (!store_lanes_p && group_size > 1) | |
ebfd146a IR |
976 | { |
977 | /* Uses a high and low interleave operation for each needed permute. */ | |
c3e7ee41 BS |
978 | |
979 | int nstmts = ncopies * exact_log2 (group_size) * group_size; | |
92345349 BS |
980 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
981 | stmt_info, 0, vect_body); | |
ebfd146a | 982 | |
73fbfcad | 983 | if (dump_enabled_p ()) |
78c60e3d | 984 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 985 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 986 | group_size); |
ebfd146a IR |
987 | } |
988 | ||
989 | /* Costs of the stores. */ | |
92345349 | 990 | vect_get_store_cost (first_dr, ncopies, &inside_cost, body_cost_vec); |
ebfd146a | 991 | |
73fbfcad | 992 | if (dump_enabled_p ()) |
78c60e3d SS |
993 | dump_printf_loc (MSG_NOTE, vect_location, |
994 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 995 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
996 | } |
997 | ||
998 | ||
720f5239 IR |
999 | /* Calculate cost of DR's memory access. */ |
1000 | void | |
1001 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1002 | unsigned int *inside_cost, |
92345349 | 1003 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
1004 | { |
1005 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
c3e7ee41 BS |
1006 | gimple stmt = DR_STMT (dr); |
1007 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
720f5239 IR |
1008 | |
1009 | switch (alignment_support_scheme) | |
1010 | { | |
1011 | case dr_aligned: | |
1012 | { | |
92345349 BS |
1013 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
1014 | vector_store, stmt_info, 0, | |
1015 | vect_body); | |
720f5239 | 1016 | |
73fbfcad | 1017 | if (dump_enabled_p ()) |
78c60e3d | 1018 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1019 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
1020 | break; |
1021 | } | |
1022 | ||
1023 | case dr_unaligned_supported: | |
1024 | { | |
720f5239 | 1025 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1026 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1027 | unaligned_store, stmt_info, |
92345349 | 1028 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1029 | if (dump_enabled_p ()) |
78c60e3d SS |
1030 | dump_printf_loc (MSG_NOTE, vect_location, |
1031 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1032 | "hardware.\n"); |
720f5239 IR |
1033 | break; |
1034 | } | |
1035 | ||
38eec4c6 UW |
1036 | case dr_unaligned_unsupported: |
1037 | { | |
1038 | *inside_cost = VECT_MAX_COST; | |
1039 | ||
73fbfcad | 1040 | if (dump_enabled_p ()) |
78c60e3d | 1041 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1042 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1043 | break; |
1044 | } | |
1045 | ||
720f5239 IR |
1046 | default: |
1047 | gcc_unreachable (); | |
1048 | } | |
1049 | } | |
1050 | ||
1051 | ||
ebfd146a IR |
1052 | /* Function vect_model_load_cost |
1053 | ||
0d0293ac MM |
1054 | Models cost for loads. In the case of grouped accesses, the last access |
1055 | has the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1056 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1057 | access scheme chosen. */ |
1058 | ||
1059 | void | |
92345349 BS |
1060 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
1061 | bool load_lanes_p, slp_tree slp_node, | |
1062 | stmt_vector_for_cost *prologue_cost_vec, | |
1063 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
1064 | { |
1065 | int group_size; | |
ebfd146a IR |
1066 | gimple first_stmt; |
1067 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; | |
92345349 | 1068 | unsigned int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
1069 | |
1070 | /* The SLP costs were already calculated during SLP tree build. */ | |
1071 | if (PURE_SLP_STMT (stmt_info)) | |
1072 | return; | |
1073 | ||
0d0293ac | 1074 | /* Grouped accesses? */ |
e14c1050 | 1075 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 1076 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && first_stmt && !slp_node) |
ebfd146a | 1077 | { |
0d0293ac | 1078 | group_size = vect_cost_group_size (stmt_info); |
ebfd146a IR |
1079 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
1080 | } | |
0d0293ac | 1081 | /* Not a grouped access. */ |
ebfd146a IR |
1082 | else |
1083 | { | |
1084 | group_size = 1; | |
1085 | first_dr = dr; | |
1086 | } | |
1087 | ||
272c6793 | 1088 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1089 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 RS |
1090 | access is instead being provided by a load-and-permute operation, |
1091 | include the cost of the permutes. */ | |
1092 | if (!load_lanes_p && group_size > 1) | |
ebfd146a IR |
1093 | { |
1094 | /* Uses an even and odd extract operations for each needed permute. */ | |
c3e7ee41 | 1095 | int nstmts = ncopies * exact_log2 (group_size) * group_size; |
92345349 BS |
1096 | inside_cost += record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
1097 | stmt_info, 0, vect_body); | |
ebfd146a | 1098 | |
73fbfcad | 1099 | if (dump_enabled_p ()) |
e645e942 TJ |
1100 | dump_printf_loc (MSG_NOTE, vect_location, |
1101 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1102 | group_size); |
ebfd146a IR |
1103 | } |
1104 | ||
1105 | /* The loads themselves. */ | |
a82960aa RG |
1106 | if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
1107 | { | |
a21892ad BS |
1108 | /* N scalar loads plus gathering them into a vector. */ |
1109 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
92345349 | 1110 | inside_cost += record_stmt_cost (body_cost_vec, |
c3e7ee41 | 1111 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), |
92345349 BS |
1112 | scalar_load, stmt_info, 0, vect_body); |
1113 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, | |
1114 | stmt_info, 0, vect_body); | |
a82960aa RG |
1115 | } |
1116 | else | |
1117 | vect_get_load_cost (first_dr, ncopies, | |
1118 | ((!STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1119 | || group_size > 1 || slp_node), | |
92345349 BS |
1120 | &inside_cost, &prologue_cost, |
1121 | prologue_cost_vec, body_cost_vec, true); | |
720f5239 | 1122 | |
73fbfcad | 1123 | if (dump_enabled_p ()) |
78c60e3d SS |
1124 | dump_printf_loc (MSG_NOTE, vect_location, |
1125 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1126 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1127 | } |
1128 | ||
1129 | ||
1130 | /* Calculate cost of DR's memory access. */ | |
1131 | void | |
1132 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1133 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1134 | unsigned int *prologue_cost, |
1135 | stmt_vector_for_cost *prologue_cost_vec, | |
1136 | stmt_vector_for_cost *body_cost_vec, | |
1137 | bool record_prologue_costs) | |
720f5239 IR |
1138 | { |
1139 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
c3e7ee41 BS |
1140 | gimple stmt = DR_STMT (dr); |
1141 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
720f5239 IR |
1142 | |
1143 | switch (alignment_support_scheme) | |
ebfd146a IR |
1144 | { |
1145 | case dr_aligned: | |
1146 | { | |
92345349 BS |
1147 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1148 | stmt_info, 0, vect_body); | |
ebfd146a | 1149 | |
73fbfcad | 1150 | if (dump_enabled_p ()) |
78c60e3d | 1151 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1152 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1153 | |
1154 | break; | |
1155 | } | |
1156 | case dr_unaligned_supported: | |
1157 | { | |
720f5239 | 1158 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1159 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1160 | unaligned_load, stmt_info, |
92345349 | 1161 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1162 | |
73fbfcad | 1163 | if (dump_enabled_p ()) |
78c60e3d SS |
1164 | dump_printf_loc (MSG_NOTE, vect_location, |
1165 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1166 | "hardware.\n"); |
ebfd146a IR |
1167 | |
1168 | break; | |
1169 | } | |
1170 | case dr_explicit_realign: | |
1171 | { | |
92345349 BS |
1172 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1173 | vector_load, stmt_info, 0, vect_body); | |
1174 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1175 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1176 | |
1177 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1178 | the containing loop, the following cost should be added to the | |
92345349 | 1179 | prologue costs. */ |
ebfd146a | 1180 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1181 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1182 | stmt_info, 0, vect_body); | |
ebfd146a | 1183 | |
73fbfcad | 1184 | if (dump_enabled_p ()) |
e645e942 TJ |
1185 | dump_printf_loc (MSG_NOTE, vect_location, |
1186 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1187 | |
ebfd146a IR |
1188 | break; |
1189 | } | |
1190 | case dr_explicit_realign_optimized: | |
1191 | { | |
73fbfcad | 1192 | if (dump_enabled_p ()) |
e645e942 | 1193 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1194 | "vect_model_load_cost: unaligned software " |
e645e942 | 1195 | "pipelined.\n"); |
ebfd146a IR |
1196 | |
1197 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1198 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1199 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1200 | access, then the above cost should only be considered for one |
ff802fa1 | 1201 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1202 | and a realignment op. */ |
1203 | ||
92345349 | 1204 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1205 | { |
92345349 BS |
1206 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1207 | vector_stmt, stmt_info, | |
1208 | 0, vect_prologue); | |
ebfd146a | 1209 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1210 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1211 | vector_stmt, stmt_info, | |
1212 | 0, vect_prologue); | |
ebfd146a IR |
1213 | } |
1214 | ||
92345349 BS |
1215 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1216 | stmt_info, 0, vect_body); | |
1217 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1218 | stmt_info, 0, vect_body); | |
8bd37302 | 1219 | |
73fbfcad | 1220 | if (dump_enabled_p ()) |
78c60e3d | 1221 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1222 | "vect_model_load_cost: explicit realign optimized" |
1223 | "\n"); | |
8bd37302 | 1224 | |
ebfd146a IR |
1225 | break; |
1226 | } | |
1227 | ||
38eec4c6 UW |
1228 | case dr_unaligned_unsupported: |
1229 | { | |
1230 | *inside_cost = VECT_MAX_COST; | |
1231 | ||
73fbfcad | 1232 | if (dump_enabled_p ()) |
78c60e3d | 1233 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1234 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1235 | break; |
1236 | } | |
1237 | ||
ebfd146a IR |
1238 | default: |
1239 | gcc_unreachable (); | |
1240 | } | |
ebfd146a IR |
1241 | } |
1242 | ||
418b7df3 RG |
1243 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1244 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1245 | |
418b7df3 RG |
1246 | static void |
1247 | vect_init_vector_1 (gimple stmt, gimple new_stmt, gimple_stmt_iterator *gsi) | |
ebfd146a | 1248 | { |
ebfd146a | 1249 | if (gsi) |
418b7df3 | 1250 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1251 | else |
1252 | { | |
418b7df3 | 1253 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1254 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1255 | |
a70d6342 IR |
1256 | if (loop_vinfo) |
1257 | { | |
1258 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1259 | basic_block new_bb; |
1260 | edge pe; | |
a70d6342 IR |
1261 | |
1262 | if (nested_in_vect_loop_p (loop, stmt)) | |
1263 | loop = loop->inner; | |
b8698a0f | 1264 | |
a70d6342 | 1265 | pe = loop_preheader_edge (loop); |
418b7df3 | 1266 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1267 | gcc_assert (!new_bb); |
1268 | } | |
1269 | else | |
1270 | { | |
1271 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1272 | basic_block bb; | |
1273 | gimple_stmt_iterator gsi_bb_start; | |
1274 | ||
1275 | gcc_assert (bb_vinfo); | |
1276 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1277 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1278 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1279 | } |
ebfd146a IR |
1280 | } |
1281 | ||
73fbfcad | 1282 | if (dump_enabled_p ()) |
ebfd146a | 1283 | { |
78c60e3d SS |
1284 | dump_printf_loc (MSG_NOTE, vect_location, |
1285 | "created new init_stmt: "); | |
1286 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
e645e942 | 1287 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 1288 | } |
418b7df3 RG |
1289 | } |
1290 | ||
1291 | /* Function vect_init_vector. | |
ebfd146a | 1292 | |
5467ee52 RG |
1293 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1294 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1295 | vector type a vector with all elements equal to VAL is created first. | |
1296 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1297 | initialization at the loop preheader. | |
418b7df3 RG |
1298 | Return the DEF of INIT_STMT. |
1299 | It will be used in the vectorization of STMT. */ | |
1300 | ||
1301 | tree | |
5467ee52 | 1302 | vect_init_vector (gimple stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 RG |
1303 | { |
1304 | tree new_var; | |
1305 | gimple init_stmt; | |
1306 | tree vec_oprnd; | |
1307 | tree new_temp; | |
1308 | ||
5467ee52 RG |
1309 | if (TREE_CODE (type) == VECTOR_TYPE |
1310 | && TREE_CODE (TREE_TYPE (val)) != VECTOR_TYPE) | |
418b7df3 | 1311 | { |
5467ee52 | 1312 | if (!types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) |
418b7df3 | 1313 | { |
5467ee52 RG |
1314 | if (CONSTANT_CLASS_P (val)) |
1315 | val = fold_unary (VIEW_CONVERT_EXPR, TREE_TYPE (type), val); | |
418b7df3 RG |
1316 | else |
1317 | { | |
83d5977e | 1318 | new_temp = make_ssa_name (TREE_TYPE (type), NULL); |
418b7df3 | 1319 | init_stmt = gimple_build_assign_with_ops (NOP_EXPR, |
83d5977e | 1320 | new_temp, val, |
418b7df3 | 1321 | NULL_TREE); |
418b7df3 | 1322 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1323 | val = new_temp; |
418b7df3 RG |
1324 | } |
1325 | } | |
5467ee52 | 1326 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1327 | } |
1328 | ||
5467ee52 | 1329 | new_var = vect_get_new_vect_var (type, vect_simple_var, "cst_"); |
5467ee52 | 1330 | init_stmt = gimple_build_assign (new_var, val); |
418b7df3 RG |
1331 | new_temp = make_ssa_name (new_var, init_stmt); |
1332 | gimple_assign_set_lhs (init_stmt, new_temp); | |
1333 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
ebfd146a IR |
1334 | vec_oprnd = gimple_assign_lhs (init_stmt); |
1335 | return vec_oprnd; | |
1336 | } | |
1337 | ||
a70d6342 | 1338 | |
ebfd146a IR |
1339 | /* Function vect_get_vec_def_for_operand. |
1340 | ||
ff802fa1 | 1341 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
1342 | used in the vectorized stmt for STMT. |
1343 | ||
1344 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1345 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1346 | ||
1347 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1348 | needs to be introduced. */ | |
1349 | ||
1350 | tree | |
1351 | vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def) | |
1352 | { | |
1353 | tree vec_oprnd; | |
1354 | gimple vec_stmt; | |
1355 | gimple def_stmt; | |
1356 | stmt_vec_info def_stmt_info = NULL; | |
1357 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
9dc3f7de | 1358 | unsigned int nunits; |
ebfd146a | 1359 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
ebfd146a | 1360 | tree def; |
ebfd146a IR |
1361 | enum vect_def_type dt; |
1362 | bool is_simple_use; | |
1363 | tree vector_type; | |
1364 | ||
73fbfcad | 1365 | if (dump_enabled_p ()) |
ebfd146a | 1366 | { |
78c60e3d SS |
1367 | dump_printf_loc (MSG_NOTE, vect_location, |
1368 | "vect_get_vec_def_for_operand: "); | |
1369 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
e645e942 | 1370 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1371 | } |
1372 | ||
24ee1384 IR |
1373 | is_simple_use = vect_is_simple_use (op, stmt, loop_vinfo, NULL, |
1374 | &def_stmt, &def, &dt); | |
ebfd146a | 1375 | gcc_assert (is_simple_use); |
73fbfcad | 1376 | if (dump_enabled_p ()) |
ebfd146a | 1377 | { |
78c60e3d | 1378 | int loc_printed = 0; |
ebfd146a IR |
1379 | if (def) |
1380 | { | |
78c60e3d SS |
1381 | dump_printf_loc (MSG_NOTE, vect_location, "def = "); |
1382 | loc_printed = 1; | |
1383 | dump_generic_expr (MSG_NOTE, TDF_SLIM, def); | |
e645e942 | 1384 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1385 | } |
1386 | if (def_stmt) | |
1387 | { | |
78c60e3d SS |
1388 | if (loc_printed) |
1389 | dump_printf (MSG_NOTE, " def_stmt = "); | |
1390 | else | |
1391 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1392 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
e645e942 | 1393 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1394 | } |
1395 | } | |
1396 | ||
1397 | switch (dt) | |
1398 | { | |
1399 | /* Case 1: operand is a constant. */ | |
1400 | case vect_constant_def: | |
1401 | { | |
7569a6cc RG |
1402 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
1403 | gcc_assert (vector_type); | |
9dc3f7de | 1404 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); |
7569a6cc | 1405 | |
b8698a0f | 1406 | if (scalar_def) |
ebfd146a IR |
1407 | *scalar_def = op; |
1408 | ||
1409 | /* Create 'vect_cst_ = {cst,cst,...,cst}' */ | |
73fbfcad | 1410 | if (dump_enabled_p ()) |
78c60e3d | 1411 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1412 | "Create vector_cst. nunits = %d\n", nunits); |
ebfd146a | 1413 | |
418b7df3 | 1414 | return vect_init_vector (stmt, op, vector_type, NULL); |
ebfd146a IR |
1415 | } |
1416 | ||
1417 | /* Case 2: operand is defined outside the loop - loop invariant. */ | |
8644a673 | 1418 | case vect_external_def: |
ebfd146a IR |
1419 | { |
1420 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); | |
1421 | gcc_assert (vector_type); | |
ebfd146a | 1422 | |
b8698a0f | 1423 | if (scalar_def) |
ebfd146a IR |
1424 | *scalar_def = def; |
1425 | ||
1426 | /* Create 'vec_inv = {inv,inv,..,inv}' */ | |
73fbfcad | 1427 | if (dump_enabled_p ()) |
e645e942 | 1428 | dump_printf_loc (MSG_NOTE, vect_location, "Create vector_inv.\n"); |
ebfd146a | 1429 | |
418b7df3 | 1430 | return vect_init_vector (stmt, def, vector_type, NULL); |
ebfd146a IR |
1431 | } |
1432 | ||
1433 | /* Case 3: operand is defined inside the loop. */ | |
8644a673 | 1434 | case vect_internal_def: |
ebfd146a | 1435 | { |
b8698a0f | 1436 | if (scalar_def) |
ebfd146a IR |
1437 | *scalar_def = NULL/* FIXME tuples: def_stmt*/; |
1438 | ||
1439 | /* Get the def from the vectorized stmt. */ | |
1440 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1441 | |
ebfd146a | 1442 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1443 | /* Get vectorized pattern statement. */ |
1444 | if (!vec_stmt | |
1445 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1446 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1447 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1448 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1449 | gcc_assert (vec_stmt); |
1450 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1451 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1452 | else if (is_gimple_call (vec_stmt)) | |
1453 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1454 | else | |
1455 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1456 | return vec_oprnd; | |
1457 | } | |
1458 | ||
1459 | /* Case 4: operand is defined by a loop header phi - reduction */ | |
1460 | case vect_reduction_def: | |
06066f92 | 1461 | case vect_double_reduction_def: |
7c5222ff | 1462 | case vect_nested_cycle: |
ebfd146a IR |
1463 | { |
1464 | struct loop *loop; | |
1465 | ||
1466 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
b8698a0f | 1467 | loop = (gimple_bb (def_stmt))->loop_father; |
ebfd146a IR |
1468 | |
1469 | /* Get the def before the loop */ | |
1470 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); | |
1471 | return get_initial_def_for_reduction (stmt, op, scalar_def); | |
1472 | } | |
1473 | ||
1474 | /* Case 5: operand is defined by loop-header phi - induction. */ | |
1475 | case vect_induction_def: | |
1476 | { | |
1477 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1478 | ||
1479 | /* Get the def from the vectorized stmt. */ | |
1480 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1481 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1482 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1483 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1484 | else | |
1485 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1486 | return vec_oprnd; |
1487 | } | |
1488 | ||
1489 | default: | |
1490 | gcc_unreachable (); | |
1491 | } | |
1492 | } | |
1493 | ||
1494 | ||
1495 | /* Function vect_get_vec_def_for_stmt_copy | |
1496 | ||
ff802fa1 | 1497 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1498 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1499 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1500 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1501 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1502 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1503 | DT is the type of the vector def VEC_OPRND. |
1504 | ||
1505 | Context: | |
1506 | In case the vectorization factor (VF) is bigger than the number | |
1507 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1508 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1509 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1510 | smallest data-type determines the VF, and as a result, when vectorizing |
1511 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1512 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1513 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1514 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1515 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1516 | ||
1517 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1518 | |
ebfd146a IR |
1519 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1520 | VS1.1: vx.1 = memref1 VS1.2 | |
1521 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1522 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1523 | |
1524 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1525 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1526 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1527 | VSnew.3: vz3 = vx.3 + ... | |
1528 | ||
1529 | The vectorization of S1 is explained in vectorizable_load. | |
1530 | The vectorization of S2: | |
b8698a0f L |
1531 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1532 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1533 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1534 | returns the vector-def 'vx.0'. |
1535 | ||
b8698a0f L |
1536 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1537 | function is called to get the relevant vector-def for each operand. It is | |
1538 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1539 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1540 | ||
b8698a0f L |
1541 | For example, to obtain the vector-def 'vx.1' in order to create the |
1542 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1543 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1544 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1545 | and return its def ('vx.1'). | |
1546 | Overall, to create the above sequence this function will be called 3 times: | |
1547 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1548 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1549 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1550 | ||
1551 | tree | |
1552 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1553 | { | |
1554 | gimple vec_stmt_for_operand; | |
1555 | stmt_vec_info def_stmt_info; | |
1556 | ||
1557 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1558 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1559 | return vec_oprnd; |
1560 | ||
1561 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1562 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1563 | gcc_assert (def_stmt_info); | |
1564 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1565 | gcc_assert (vec_stmt_for_operand); | |
1566 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1567 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) | |
1568 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1569 | else | |
1570 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1571 | return vec_oprnd; | |
1572 | } | |
1573 | ||
1574 | ||
1575 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1576 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1577 | |
1578 | static void | |
b8698a0f | 1579 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1580 | vec<tree> *vec_oprnds0, |
1581 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1582 | { |
9771b263 | 1583 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1584 | |
1585 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1586 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1587 | |
9771b263 | 1588 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1589 | { |
9771b263 | 1590 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1591 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1592 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1593 | } |
1594 | } | |
1595 | ||
1596 | ||
d092494c IR |
1597 | /* Get vectorized definitions for OP0 and OP1. |
1598 | REDUC_INDEX is the index of reduction operand in case of reduction, | |
1599 | and -1 otherwise. */ | |
ebfd146a | 1600 | |
d092494c | 1601 | void |
ebfd146a | 1602 | vect_get_vec_defs (tree op0, tree op1, gimple stmt, |
9771b263 DN |
1603 | vec<tree> *vec_oprnds0, |
1604 | vec<tree> *vec_oprnds1, | |
d092494c | 1605 | slp_tree slp_node, int reduc_index) |
ebfd146a IR |
1606 | { |
1607 | if (slp_node) | |
d092494c IR |
1608 | { |
1609 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1610 | auto_vec<tree> ops (nops); |
1611 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1612 | |
9771b263 | 1613 | ops.quick_push (op0); |
d092494c | 1614 | if (op1) |
9771b263 | 1615 | ops.quick_push (op1); |
d092494c IR |
1616 | |
1617 | vect_get_slp_defs (ops, slp_node, &vec_defs, reduc_index); | |
1618 | ||
37b5ec8f | 1619 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1620 | if (op1) |
37b5ec8f | 1621 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1622 | } |
ebfd146a IR |
1623 | else |
1624 | { | |
1625 | tree vec_oprnd; | |
1626 | ||
9771b263 | 1627 | vec_oprnds0->create (1); |
b8698a0f | 1628 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL); |
9771b263 | 1629 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1630 | |
1631 | if (op1) | |
1632 | { | |
9771b263 | 1633 | vec_oprnds1->create (1); |
b8698a0f | 1634 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL); |
9771b263 | 1635 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1636 | } |
1637 | } | |
1638 | } | |
1639 | ||
1640 | ||
1641 | /* Function vect_finish_stmt_generation. | |
1642 | ||
1643 | Insert a new stmt. */ | |
1644 | ||
1645 | void | |
1646 | vect_finish_stmt_generation (gimple stmt, gimple vec_stmt, | |
1647 | gimple_stmt_iterator *gsi) | |
1648 | { | |
1649 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1650 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1651 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1652 | |
1653 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1654 | ||
54e8e2c3 RG |
1655 | if (!gsi_end_p (*gsi) |
1656 | && gimple_has_mem_ops (vec_stmt)) | |
1657 | { | |
1658 | gimple at_stmt = gsi_stmt (*gsi); | |
1659 | tree vuse = gimple_vuse (at_stmt); | |
1660 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1661 | { | |
1662 | tree vdef = gimple_vdef (at_stmt); | |
1663 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1664 | /* If we have an SSA vuse and insert a store, update virtual | |
1665 | SSA form to avoid triggering the renamer. Do so only | |
1666 | if we can easily see all uses - which is what almost always | |
1667 | happens with the way vectorized stmts are inserted. */ | |
1668 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1669 | && ((is_gimple_assign (vec_stmt) | |
1670 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1671 | || (is_gimple_call (vec_stmt) | |
1672 | && !(gimple_call_flags (vec_stmt) | |
1673 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1674 | { | |
1675 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1676 | gimple_set_vdef (vec_stmt, new_vdef); | |
1677 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1678 | } | |
1679 | } | |
1680 | } | |
ebfd146a IR |
1681 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1682 | ||
b8698a0f | 1683 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo, |
a70d6342 | 1684 | bb_vinfo)); |
ebfd146a | 1685 | |
73fbfcad | 1686 | if (dump_enabled_p ()) |
ebfd146a | 1687 | { |
78c60e3d SS |
1688 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1689 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
e645e942 | 1690 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1691 | } |
1692 | ||
ad885386 | 1693 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
ebfd146a IR |
1694 | } |
1695 | ||
1696 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1697 | a function declaration if the target has a vectorized version | |
1698 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1699 | ||
1700 | tree | |
1701 | vectorizable_function (gimple call, tree vectype_out, tree vectype_in) | |
1702 | { | |
1703 | tree fndecl = gimple_call_fndecl (call); | |
ebfd146a IR |
1704 | |
1705 | /* We only handle functions that do not read or clobber memory -- i.e. | |
1706 | const or novops ones. */ | |
1707 | if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS))) | |
1708 | return NULL_TREE; | |
1709 | ||
1710 | if (!fndecl | |
1711 | || TREE_CODE (fndecl) != FUNCTION_DECL | |
1712 | || !DECL_BUILT_IN (fndecl)) | |
1713 | return NULL_TREE; | |
1714 | ||
62f7fd21 | 1715 | return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out, |
ebfd146a IR |
1716 | vectype_in); |
1717 | } | |
1718 | ||
5ce9450f JJ |
1719 | |
1720 | static tree permute_vec_elements (tree, tree, tree, gimple, | |
1721 | gimple_stmt_iterator *); | |
1722 | ||
1723 | ||
1724 | /* Function vectorizable_mask_load_store. | |
1725 | ||
1726 | Check if STMT performs a conditional load or store that can be vectorized. | |
1727 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1728 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
1729 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1730 | ||
1731 | static bool | |
1732 | vectorizable_mask_load_store (gimple stmt, gimple_stmt_iterator *gsi, | |
1733 | gimple *vec_stmt, slp_tree slp_node) | |
1734 | { | |
1735 | tree vec_dest = NULL; | |
1736 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1737 | stmt_vec_info prev_stmt_info; | |
1738 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1739 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1740 | bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
1741 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
1742 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
1743 | tree elem_type; | |
1744 | gimple new_stmt; | |
1745 | tree dummy; | |
1746 | tree dataref_ptr = NULL_TREE; | |
1747 | gimple ptr_incr; | |
1748 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
1749 | int ncopies; | |
1750 | int i, j; | |
1751 | bool inv_p; | |
1752 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; | |
1753 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
1754 | int gather_scale = 1; | |
1755 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
1756 | bool is_store; | |
1757 | tree mask; | |
1758 | gimple def_stmt; | |
1759 | tree def; | |
1760 | enum vect_def_type dt; | |
1761 | ||
1762 | if (slp_node != NULL) | |
1763 | return false; | |
1764 | ||
1765 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
1766 | gcc_assert (ncopies >= 1); | |
1767 | ||
1768 | is_store = gimple_call_internal_fn (stmt) == IFN_MASK_STORE; | |
1769 | mask = gimple_call_arg (stmt, 2); | |
1770 | if (TYPE_PRECISION (TREE_TYPE (mask)) | |
1771 | != GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (vectype)))) | |
1772 | return false; | |
1773 | ||
1774 | /* FORNOW. This restriction should be relaxed. */ | |
1775 | if (nested_in_vect_loop && ncopies > 1) | |
1776 | { | |
1777 | if (dump_enabled_p ()) | |
1778 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1779 | "multiple types in nested loop."); | |
1780 | return false; | |
1781 | } | |
1782 | ||
1783 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
1784 | return false; | |
1785 | ||
1786 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
1787 | return false; | |
1788 | ||
1789 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
1790 | return false; | |
1791 | ||
1792 | elem_type = TREE_TYPE (vectype); | |
1793 | ||
1794 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1795 | return false; | |
1796 | ||
1797 | if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) | |
1798 | return false; | |
1799 | ||
1800 | if (STMT_VINFO_GATHER_P (stmt_info)) | |
1801 | { | |
1802 | gimple def_stmt; | |
1803 | tree def; | |
1804 | gather_decl = vect_check_gather (stmt, loop_vinfo, &gather_base, | |
1805 | &gather_off, &gather_scale); | |
1806 | gcc_assert (gather_decl); | |
1807 | if (!vect_is_simple_use_1 (gather_off, NULL, loop_vinfo, NULL, | |
1808 | &def_stmt, &def, &gather_dt, | |
1809 | &gather_off_vectype)) | |
1810 | { | |
1811 | if (dump_enabled_p ()) | |
1812 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1813 | "gather index use not simple."); | |
1814 | return false; | |
1815 | } | |
1816 | } | |
1817 | else if (tree_int_cst_compare (nested_in_vect_loop | |
1818 | ? STMT_VINFO_DR_STEP (stmt_info) | |
1819 | : DR_STEP (dr), size_zero_node) <= 0) | |
1820 | return false; | |
1821 | else if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
1822 | || !can_vec_mask_load_store_p (TYPE_MODE (vectype), !is_store)) | |
1823 | return false; | |
1824 | ||
1825 | if (TREE_CODE (mask) != SSA_NAME) | |
1826 | return false; | |
1827 | ||
1828 | if (!vect_is_simple_use (mask, stmt, loop_vinfo, NULL, | |
1829 | &def_stmt, &def, &dt)) | |
1830 | return false; | |
1831 | ||
1832 | if (is_store) | |
1833 | { | |
1834 | tree rhs = gimple_call_arg (stmt, 3); | |
1835 | if (!vect_is_simple_use (rhs, stmt, loop_vinfo, NULL, | |
1836 | &def_stmt, &def, &dt)) | |
1837 | return false; | |
1838 | } | |
1839 | ||
1840 | if (!vec_stmt) /* transformation not required. */ | |
1841 | { | |
1842 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1843 | if (is_store) | |
1844 | vect_model_store_cost (stmt_info, ncopies, false, dt, | |
1845 | NULL, NULL, NULL); | |
1846 | else | |
1847 | vect_model_load_cost (stmt_info, ncopies, false, NULL, NULL, NULL); | |
1848 | return true; | |
1849 | } | |
1850 | ||
1851 | /** Transform. **/ | |
1852 | ||
1853 | if (STMT_VINFO_GATHER_P (stmt_info)) | |
1854 | { | |
1855 | tree vec_oprnd0 = NULL_TREE, op; | |
1856 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1857 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
1858 | tree ptr, vec_mask = NULL_TREE, mask_op, var, scale; | |
1859 | tree perm_mask = NULL_TREE, prev_res = NULL_TREE; | |
1860 | edge pe = loop_preheader_edge (loop); | |
1861 | gimple_seq seq; | |
1862 | basic_block new_bb; | |
1863 | enum { NARROW, NONE, WIDEN } modifier; | |
1864 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
1865 | ||
1866 | if (nunits == gather_off_nunits) | |
1867 | modifier = NONE; | |
1868 | else if (nunits == gather_off_nunits / 2) | |
1869 | { | |
1870 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
1871 | modifier = WIDEN; | |
1872 | ||
1873 | for (i = 0; i < gather_off_nunits; ++i) | |
1874 | sel[i] = i | nunits; | |
1875 | ||
1876 | perm_mask = vect_gen_perm_mask (gather_off_vectype, sel); | |
1877 | gcc_assert (perm_mask != NULL_TREE); | |
1878 | } | |
1879 | else if (nunits == gather_off_nunits * 2) | |
1880 | { | |
1881 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
1882 | modifier = NARROW; | |
1883 | ||
1884 | for (i = 0; i < nunits; ++i) | |
1885 | sel[i] = i < gather_off_nunits | |
1886 | ? i : i + nunits - gather_off_nunits; | |
1887 | ||
1888 | perm_mask = vect_gen_perm_mask (vectype, sel); | |
1889 | gcc_assert (perm_mask != NULL_TREE); | |
1890 | ncopies *= 2; | |
1891 | } | |
1892 | else | |
1893 | gcc_unreachable (); | |
1894 | ||
1895 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
1896 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1897 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1898 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1899 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1900 | scaletype = TREE_VALUE (arglist); | |
1901 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
1902 | && types_compatible_p (srctype, masktype)); | |
1903 | ||
1904 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); | |
1905 | ||
1906 | ptr = fold_convert (ptrtype, gather_base); | |
1907 | if (!is_gimple_min_invariant (ptr)) | |
1908 | { | |
1909 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
1910 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
1911 | gcc_assert (!new_bb); | |
1912 | } | |
1913 | ||
1914 | scale = build_int_cst (scaletype, gather_scale); | |
1915 | ||
1916 | prev_stmt_info = NULL; | |
1917 | for (j = 0; j < ncopies; ++j) | |
1918 | { | |
1919 | if (modifier == WIDEN && (j & 1)) | |
1920 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
1921 | perm_mask, stmt, gsi); | |
1922 | else if (j == 0) | |
1923 | op = vec_oprnd0 | |
1924 | = vect_get_vec_def_for_operand (gather_off, stmt, NULL); | |
1925 | else | |
1926 | op = vec_oprnd0 | |
1927 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
1928 | ||
1929 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
1930 | { | |
1931 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
1932 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
1933 | var = vect_get_new_vect_var (idxtype, vect_simple_var, NULL); | |
1934 | var = make_ssa_name (var, NULL); | |
1935 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); | |
1936 | new_stmt | |
1937 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, | |
1938 | op, NULL_TREE); | |
1939 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1940 | op = var; | |
1941 | } | |
1942 | ||
1943 | if (j == 0) | |
1944 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, NULL); | |
1945 | else | |
1946 | { | |
1947 | vect_is_simple_use (vec_mask, NULL, loop_vinfo, NULL, &def_stmt, | |
1948 | &def, &dt); | |
1949 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); | |
1950 | } | |
1951 | ||
1952 | mask_op = vec_mask; | |
1953 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
1954 | { | |
1955 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)) | |
1956 | == TYPE_VECTOR_SUBPARTS (masktype)); | |
1957 | var = vect_get_new_vect_var (masktype, vect_simple_var, NULL); | |
1958 | var = make_ssa_name (var, NULL); | |
1959 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); | |
1960 | new_stmt | |
1961 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, | |
1962 | mask_op, NULL_TREE); | |
1963 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1964 | mask_op = var; | |
1965 | } | |
1966 | ||
1967 | new_stmt | |
1968 | = gimple_build_call (gather_decl, 5, mask_op, ptr, op, mask_op, | |
1969 | scale); | |
1970 | ||
1971 | if (!useless_type_conversion_p (vectype, rettype)) | |
1972 | { | |
1973 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
1974 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
1975 | var = vect_get_new_vect_var (rettype, vect_simple_var, NULL); | |
1976 | op = make_ssa_name (var, new_stmt); | |
1977 | gimple_call_set_lhs (new_stmt, op); | |
1978 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1979 | var = make_ssa_name (vec_dest, NULL); | |
1980 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); | |
1981 | new_stmt | |
1982 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, op, | |
1983 | NULL_TREE); | |
1984 | } | |
1985 | else | |
1986 | { | |
1987 | var = make_ssa_name (vec_dest, new_stmt); | |
1988 | gimple_call_set_lhs (new_stmt, var); | |
1989 | } | |
1990 | ||
1991 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1992 | ||
1993 | if (modifier == NARROW) | |
1994 | { | |
1995 | if ((j & 1) == 0) | |
1996 | { | |
1997 | prev_res = var; | |
1998 | continue; | |
1999 | } | |
2000 | var = permute_vec_elements (prev_res, var, | |
2001 | perm_mask, stmt, gsi); | |
2002 | new_stmt = SSA_NAME_DEF_STMT (var); | |
2003 | } | |
2004 | ||
2005 | if (prev_stmt_info == NULL) | |
2006 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2007 | else | |
2008 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2009 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2010 | } | |
2011 | return true; | |
2012 | } | |
2013 | else if (is_store) | |
2014 | { | |
2015 | tree vec_rhs = NULL_TREE, vec_mask = NULL_TREE; | |
2016 | prev_stmt_info = NULL; | |
2017 | for (i = 0; i < ncopies; i++) | |
2018 | { | |
2019 | unsigned align, misalign; | |
2020 | ||
2021 | if (i == 0) | |
2022 | { | |
2023 | tree rhs = gimple_call_arg (stmt, 3); | |
2024 | vec_rhs = vect_get_vec_def_for_operand (rhs, stmt, NULL); | |
2025 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, NULL); | |
2026 | /* We should have catched mismatched types earlier. */ | |
2027 | gcc_assert (useless_type_conversion_p (vectype, | |
2028 | TREE_TYPE (vec_rhs))); | |
2029 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, | |
2030 | NULL_TREE, &dummy, gsi, | |
2031 | &ptr_incr, false, &inv_p); | |
2032 | gcc_assert (!inv_p); | |
2033 | } | |
2034 | else | |
2035 | { | |
2036 | vect_is_simple_use (vec_rhs, NULL, loop_vinfo, NULL, &def_stmt, | |
2037 | &def, &dt); | |
2038 | vec_rhs = vect_get_vec_def_for_stmt_copy (dt, vec_rhs); | |
2039 | vect_is_simple_use (vec_mask, NULL, loop_vinfo, NULL, &def_stmt, | |
2040 | &def, &dt); | |
2041 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); | |
2042 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2043 | TYPE_SIZE_UNIT (vectype)); | |
2044 | } | |
2045 | ||
2046 | align = TYPE_ALIGN_UNIT (vectype); | |
2047 | if (aligned_access_p (dr)) | |
2048 | misalign = 0; | |
2049 | else if (DR_MISALIGNMENT (dr) == -1) | |
2050 | { | |
2051 | align = TYPE_ALIGN_UNIT (elem_type); | |
2052 | misalign = 0; | |
2053 | } | |
2054 | else | |
2055 | misalign = DR_MISALIGNMENT (dr); | |
2056 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2057 | misalign); | |
2058 | new_stmt | |
2059 | = gimple_build_call_internal (IFN_MASK_STORE, 4, dataref_ptr, | |
2060 | gimple_call_arg (stmt, 1), | |
2061 | vec_mask, vec_rhs); | |
2062 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2063 | if (i == 0) | |
2064 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2065 | else | |
2066 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2067 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2068 | } | |
2069 | } | |
2070 | else | |
2071 | { | |
2072 | tree vec_mask = NULL_TREE; | |
2073 | prev_stmt_info = NULL; | |
2074 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); | |
2075 | for (i = 0; i < ncopies; i++) | |
2076 | { | |
2077 | unsigned align, misalign; | |
2078 | ||
2079 | if (i == 0) | |
2080 | { | |
2081 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, NULL); | |
2082 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, | |
2083 | NULL_TREE, &dummy, gsi, | |
2084 | &ptr_incr, false, &inv_p); | |
2085 | gcc_assert (!inv_p); | |
2086 | } | |
2087 | else | |
2088 | { | |
2089 | vect_is_simple_use (vec_mask, NULL, loop_vinfo, NULL, &def_stmt, | |
2090 | &def, &dt); | |
2091 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); | |
2092 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2093 | TYPE_SIZE_UNIT (vectype)); | |
2094 | } | |
2095 | ||
2096 | align = TYPE_ALIGN_UNIT (vectype); | |
2097 | if (aligned_access_p (dr)) | |
2098 | misalign = 0; | |
2099 | else if (DR_MISALIGNMENT (dr) == -1) | |
2100 | { | |
2101 | align = TYPE_ALIGN_UNIT (elem_type); | |
2102 | misalign = 0; | |
2103 | } | |
2104 | else | |
2105 | misalign = DR_MISALIGNMENT (dr); | |
2106 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2107 | misalign); | |
2108 | new_stmt | |
2109 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, dataref_ptr, | |
2110 | gimple_call_arg (stmt, 1), | |
2111 | vec_mask); | |
2112 | gimple_call_set_lhs (new_stmt, make_ssa_name (vec_dest, NULL)); | |
2113 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2114 | if (i == 0) | |
2115 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2116 | else | |
2117 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2118 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2119 | } | |
2120 | } | |
2121 | ||
2122 | return true; | |
2123 | } | |
2124 | ||
2125 | ||
ebfd146a IR |
2126 | /* Function vectorizable_call. |
2127 | ||
b8698a0f L |
2128 | Check if STMT performs a function call that can be vectorized. |
2129 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
2130 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2131 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2132 | ||
2133 | static bool | |
190c2236 JJ |
2134 | vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, |
2135 | slp_tree slp_node) | |
ebfd146a IR |
2136 | { |
2137 | tree vec_dest; | |
2138 | tree scalar_dest; | |
2139 | tree op, type; | |
2140 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
2141 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
2142 | tree vectype_out, vectype_in; | |
2143 | int nunits_in; | |
2144 | int nunits_out; | |
2145 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 2146 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b690cc0f | 2147 | tree fndecl, new_temp, def, rhs_type; |
ebfd146a | 2148 | gimple def_stmt; |
0502fb85 UB |
2149 | enum vect_def_type dt[3] |
2150 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
63827fb8 | 2151 | gimple new_stmt = NULL; |
ebfd146a | 2152 | int ncopies, j; |
6e1aa848 | 2153 | vec<tree> vargs = vNULL; |
ebfd146a IR |
2154 | enum { NARROW, NONE, WIDEN } modifier; |
2155 | size_t i, nargs; | |
9d5e7640 | 2156 | tree lhs; |
ebfd146a | 2157 | |
190c2236 | 2158 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2159 | return false; |
2160 | ||
8644a673 | 2161 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2162 | return false; |
2163 | ||
ebfd146a IR |
2164 | /* Is STMT a vectorizable call? */ |
2165 | if (!is_gimple_call (stmt)) | |
2166 | return false; | |
2167 | ||
5ce9450f JJ |
2168 | if (gimple_call_internal_p (stmt) |
2169 | && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD | |
2170 | || gimple_call_internal_fn (stmt) == IFN_MASK_STORE)) | |
2171 | return vectorizable_mask_load_store (stmt, gsi, vec_stmt, | |
2172 | slp_node); | |
2173 | ||
0136f8f0 AH |
2174 | if (gimple_call_lhs (stmt) == NULL_TREE |
2175 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
2176 | return false; |
2177 | ||
0136f8f0 | 2178 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 2179 | |
b690cc0f RG |
2180 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2181 | ||
ebfd146a IR |
2182 | /* Process function arguments. */ |
2183 | rhs_type = NULL_TREE; | |
b690cc0f | 2184 | vectype_in = NULL_TREE; |
ebfd146a IR |
2185 | nargs = gimple_call_num_args (stmt); |
2186 | ||
1b1562a5 MM |
2187 | /* Bail out if the function has more than three arguments, we do not have |
2188 | interesting builtin functions to vectorize with more than two arguments | |
2189 | except for fma. No arguments is also not good. */ | |
2190 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
2191 | return false; |
2192 | ||
74bf76ed JJ |
2193 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2194 | if (gimple_call_internal_p (stmt) | |
2195 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2196 | { | |
2197 | nargs = 0; | |
2198 | rhs_type = unsigned_type_node; | |
2199 | } | |
2200 | ||
ebfd146a IR |
2201 | for (i = 0; i < nargs; i++) |
2202 | { | |
b690cc0f RG |
2203 | tree opvectype; |
2204 | ||
ebfd146a IR |
2205 | op = gimple_call_arg (stmt, i); |
2206 | ||
2207 | /* We can only handle calls with arguments of the same type. */ | |
2208 | if (rhs_type | |
8533c9d8 | 2209 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 2210 | { |
73fbfcad | 2211 | if (dump_enabled_p ()) |
78c60e3d | 2212 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2213 | "argument types differ.\n"); |
ebfd146a IR |
2214 | return false; |
2215 | } | |
b690cc0f RG |
2216 | if (!rhs_type) |
2217 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 2218 | |
24ee1384 | 2219 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, |
b690cc0f | 2220 | &def_stmt, &def, &dt[i], &opvectype)) |
ebfd146a | 2221 | { |
73fbfcad | 2222 | if (dump_enabled_p ()) |
78c60e3d | 2223 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2224 | "use not simple.\n"); |
ebfd146a IR |
2225 | return false; |
2226 | } | |
ebfd146a | 2227 | |
b690cc0f RG |
2228 | if (!vectype_in) |
2229 | vectype_in = opvectype; | |
2230 | else if (opvectype | |
2231 | && opvectype != vectype_in) | |
2232 | { | |
73fbfcad | 2233 | if (dump_enabled_p ()) |
78c60e3d | 2234 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2235 | "argument vector types differ.\n"); |
b690cc0f RG |
2236 | return false; |
2237 | } | |
2238 | } | |
2239 | /* If all arguments are external or constant defs use a vector type with | |
2240 | the same size as the output vector type. */ | |
ebfd146a | 2241 | if (!vectype_in) |
b690cc0f | 2242 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2243 | if (vec_stmt) |
2244 | gcc_assert (vectype_in); | |
2245 | if (!vectype_in) | |
2246 | { | |
73fbfcad | 2247 | if (dump_enabled_p ()) |
7d8930a0 | 2248 | { |
78c60e3d SS |
2249 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2250 | "no vectype for scalar type "); | |
2251 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2252 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
2253 | } |
2254 | ||
2255 | return false; | |
2256 | } | |
ebfd146a IR |
2257 | |
2258 | /* FORNOW */ | |
b690cc0f RG |
2259 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2260 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
2261 | if (nunits_in == nunits_out / 2) |
2262 | modifier = NARROW; | |
2263 | else if (nunits_out == nunits_in) | |
2264 | modifier = NONE; | |
2265 | else if (nunits_out == nunits_in / 2) | |
2266 | modifier = WIDEN; | |
2267 | else | |
2268 | return false; | |
2269 | ||
2270 | /* For now, we only vectorize functions if a target specific builtin | |
2271 | is available. TODO -- in some cases, it might be profitable to | |
2272 | insert the calls for pieces of the vector, in order to be able | |
2273 | to vectorize other operations in the loop. */ | |
2274 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
2275 | if (fndecl == NULL_TREE) | |
2276 | { | |
74bf76ed JJ |
2277 | if (gimple_call_internal_p (stmt) |
2278 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE | |
2279 | && !slp_node | |
2280 | && loop_vinfo | |
2281 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2282 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
2283 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2284 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
2285 | { | |
2286 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
2287 | { 0, 1, 2, ... vf - 1 } vector. */ | |
2288 | gcc_assert (nargs == 0); | |
2289 | } | |
2290 | else | |
2291 | { | |
2292 | if (dump_enabled_p ()) | |
2293 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 2294 | "function is not vectorizable.\n"); |
74bf76ed JJ |
2295 | return false; |
2296 | } | |
ebfd146a IR |
2297 | } |
2298 | ||
5006671f | 2299 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a | 2300 | |
190c2236 JJ |
2301 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
2302 | ncopies = 1; | |
2303 | else if (modifier == NARROW) | |
ebfd146a IR |
2304 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
2305 | else | |
2306 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2307 | ||
2308 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2309 | needs to be generated. */ | |
2310 | gcc_assert (ncopies >= 1); | |
2311 | ||
2312 | if (!vec_stmt) /* transformation not required. */ | |
2313 | { | |
2314 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 2315 | if (dump_enabled_p ()) |
e645e942 TJ |
2316 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
2317 | "\n"); | |
c3e7ee41 | 2318 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
2319 | return true; |
2320 | } | |
2321 | ||
2322 | /** Transform. **/ | |
2323 | ||
73fbfcad | 2324 | if (dump_enabled_p ()) |
e645e942 | 2325 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
2326 | |
2327 | /* Handle def. */ | |
2328 | scalar_dest = gimple_call_lhs (stmt); | |
2329 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2330 | ||
2331 | prev_stmt_info = NULL; | |
2332 | switch (modifier) | |
2333 | { | |
2334 | case NONE: | |
2335 | for (j = 0; j < ncopies; ++j) | |
2336 | { | |
2337 | /* Build argument list for the vectorized call. */ | |
2338 | if (j == 0) | |
9771b263 | 2339 | vargs.create (nargs); |
ebfd146a | 2340 | else |
9771b263 | 2341 | vargs.truncate (0); |
ebfd146a | 2342 | |
190c2236 JJ |
2343 | if (slp_node) |
2344 | { | |
ef062b13 | 2345 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2346 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2347 | |
2348 | for (i = 0; i < nargs; i++) | |
9771b263 | 2349 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2350 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2351 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2352 | |
2353 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2354 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
2355 | { |
2356 | size_t k; | |
2357 | for (k = 0; k < nargs; k++) | |
2358 | { | |
37b5ec8f | 2359 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 2360 | vargs[k] = vec_oprndsk[i]; |
190c2236 JJ |
2361 | } |
2362 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2363 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2364 | gimple_call_set_lhs (new_stmt, new_temp); | |
2365 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2366 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2367 | } |
2368 | ||
2369 | for (i = 0; i < nargs; i++) | |
2370 | { | |
37b5ec8f | 2371 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2372 | vec_oprndsi.release (); |
190c2236 | 2373 | } |
190c2236 JJ |
2374 | continue; |
2375 | } | |
2376 | ||
ebfd146a IR |
2377 | for (i = 0; i < nargs; i++) |
2378 | { | |
2379 | op = gimple_call_arg (stmt, i); | |
2380 | if (j == 0) | |
2381 | vec_oprnd0 | |
2382 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
2383 | else | |
63827fb8 IR |
2384 | { |
2385 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
2386 | vec_oprnd0 | |
2387 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
2388 | } | |
ebfd146a | 2389 | |
9771b263 | 2390 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
2391 | } |
2392 | ||
74bf76ed JJ |
2393 | if (gimple_call_internal_p (stmt) |
2394 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2395 | { | |
2396 | tree *v = XALLOCAVEC (tree, nunits_out); | |
2397 | int k; | |
2398 | for (k = 0; k < nunits_out; ++k) | |
2399 | v[k] = build_int_cst (unsigned_type_node, j * nunits_out + k); | |
2400 | tree cst = build_vector (vectype_out, v); | |
2401 | tree new_var | |
2402 | = vect_get_new_vect_var (vectype_out, vect_simple_var, "cst_"); | |
2403 | gimple init_stmt = gimple_build_assign (new_var, cst); | |
2404 | new_temp = make_ssa_name (new_var, init_stmt); | |
2405 | gimple_assign_set_lhs (init_stmt, new_temp); | |
2406 | vect_init_vector_1 (stmt, init_stmt, NULL); | |
2407 | new_temp = make_ssa_name (vec_dest, NULL); | |
2408 | new_stmt = gimple_build_assign (new_temp, | |
2409 | gimple_assign_lhs (init_stmt)); | |
2410 | } | |
2411 | else | |
2412 | { | |
2413 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2414 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2415 | gimple_call_set_lhs (new_stmt, new_temp); | |
2416 | } | |
ebfd146a IR |
2417 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2418 | ||
2419 | if (j == 0) | |
2420 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2421 | else | |
2422 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2423 | ||
2424 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2425 | } | |
2426 | ||
2427 | break; | |
2428 | ||
2429 | case NARROW: | |
2430 | for (j = 0; j < ncopies; ++j) | |
2431 | { | |
2432 | /* Build argument list for the vectorized call. */ | |
2433 | if (j == 0) | |
9771b263 | 2434 | vargs.create (nargs * 2); |
ebfd146a | 2435 | else |
9771b263 | 2436 | vargs.truncate (0); |
ebfd146a | 2437 | |
190c2236 JJ |
2438 | if (slp_node) |
2439 | { | |
ef062b13 | 2440 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2441 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2442 | |
2443 | for (i = 0; i < nargs; i++) | |
9771b263 | 2444 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2445 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2446 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2447 | |
2448 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2449 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
2450 | { |
2451 | size_t k; | |
9771b263 | 2452 | vargs.truncate (0); |
190c2236 JJ |
2453 | for (k = 0; k < nargs; k++) |
2454 | { | |
37b5ec8f | 2455 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
2456 | vargs.quick_push (vec_oprndsk[i]); |
2457 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 JJ |
2458 | } |
2459 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2460 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2461 | gimple_call_set_lhs (new_stmt, new_temp); | |
2462 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2463 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2464 | } |
2465 | ||
2466 | for (i = 0; i < nargs; i++) | |
2467 | { | |
37b5ec8f | 2468 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2469 | vec_oprndsi.release (); |
190c2236 | 2470 | } |
190c2236 JJ |
2471 | continue; |
2472 | } | |
2473 | ||
ebfd146a IR |
2474 | for (i = 0; i < nargs; i++) |
2475 | { | |
2476 | op = gimple_call_arg (stmt, i); | |
2477 | if (j == 0) | |
2478 | { | |
2479 | vec_oprnd0 | |
2480 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
2481 | vec_oprnd1 | |
63827fb8 | 2482 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2483 | } |
2484 | else | |
2485 | { | |
336ecb65 | 2486 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 2487 | vec_oprnd0 |
63827fb8 | 2488 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 2489 | vec_oprnd1 |
63827fb8 | 2490 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2491 | } |
2492 | ||
9771b263 DN |
2493 | vargs.quick_push (vec_oprnd0); |
2494 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
2495 | } |
2496 | ||
2497 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2498 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2499 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
2500 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2501 | ||
2502 | if (j == 0) | |
2503 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2504 | else | |
2505 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2506 | ||
2507 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2508 | } | |
2509 | ||
2510 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2511 | ||
2512 | break; | |
2513 | ||
2514 | case WIDEN: | |
2515 | /* No current target implements this case. */ | |
2516 | return false; | |
2517 | } | |
2518 | ||
9771b263 | 2519 | vargs.release (); |
ebfd146a | 2520 | |
ebfd146a IR |
2521 | /* The call in STMT might prevent it from being removed in dce. |
2522 | We however cannot remove it here, due to the way the ssa name | |
2523 | it defines is mapped to the new definition. So just replace | |
2524 | rhs of the statement with something harmless. */ | |
2525 | ||
dd34c087 JJ |
2526 | if (slp_node) |
2527 | return true; | |
2528 | ||
ebfd146a | 2529 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
2530 | if (is_pattern_stmt_p (stmt_info)) |
2531 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
2532 | else | |
2533 | lhs = gimple_call_lhs (stmt); | |
2534 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
ebfd146a | 2535 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 2536 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
2537 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
2538 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
2539 | |
2540 | return true; | |
2541 | } | |
2542 | ||
2543 | ||
0136f8f0 AH |
2544 | struct simd_call_arg_info |
2545 | { | |
2546 | tree vectype; | |
2547 | tree op; | |
2548 | enum vect_def_type dt; | |
2549 | HOST_WIDE_INT linear_step; | |
2550 | unsigned int align; | |
2551 | }; | |
2552 | ||
2553 | /* Function vectorizable_simd_clone_call. | |
2554 | ||
2555 | Check if STMT performs a function call that can be vectorized | |
2556 | by calling a simd clone of the function. | |
2557 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2558 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2559 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2560 | ||
2561 | static bool | |
2562 | vectorizable_simd_clone_call (gimple stmt, gimple_stmt_iterator *gsi, | |
2563 | gimple *vec_stmt, slp_tree slp_node) | |
2564 | { | |
2565 | tree vec_dest; | |
2566 | tree scalar_dest; | |
2567 | tree op, type; | |
2568 | tree vec_oprnd0 = NULL_TREE; | |
2569 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
2570 | tree vectype; | |
2571 | unsigned int nunits; | |
2572 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2573 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
2574 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; | |
2575 | tree fndecl, new_temp, def; | |
2576 | gimple def_stmt; | |
2577 | gimple new_stmt = NULL; | |
2578 | int ncopies, j; | |
2579 | vec<simd_call_arg_info> arginfo = vNULL; | |
2580 | vec<tree> vargs = vNULL; | |
2581 | size_t i, nargs; | |
2582 | tree lhs, rtype, ratype; | |
2583 | vec<constructor_elt, va_gc> *ret_ctor_elts; | |
2584 | ||
2585 | /* Is STMT a vectorizable call? */ | |
2586 | if (!is_gimple_call (stmt)) | |
2587 | return false; | |
2588 | ||
2589 | fndecl = gimple_call_fndecl (stmt); | |
2590 | if (fndecl == NULL_TREE) | |
2591 | return false; | |
2592 | ||
2593 | struct cgraph_node *node = cgraph_get_node (fndecl); | |
2594 | if (node == NULL || node->simd_clones == NULL) | |
2595 | return false; | |
2596 | ||
2597 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
2598 | return false; | |
2599 | ||
2600 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
2601 | return false; | |
2602 | ||
2603 | if (gimple_call_lhs (stmt) | |
2604 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
2605 | return false; | |
2606 | ||
2607 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
2608 | ||
2609 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2610 | ||
2611 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
2612 | return false; | |
2613 | ||
2614 | /* FORNOW */ | |
2615 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
2616 | return false; | |
2617 | ||
2618 | /* Process function arguments. */ | |
2619 | nargs = gimple_call_num_args (stmt); | |
2620 | ||
2621 | /* Bail out if the function has zero arguments. */ | |
2622 | if (nargs == 0) | |
2623 | return false; | |
2624 | ||
2625 | arginfo.create (nargs); | |
2626 | ||
2627 | for (i = 0; i < nargs; i++) | |
2628 | { | |
2629 | simd_call_arg_info thisarginfo; | |
2630 | affine_iv iv; | |
2631 | ||
2632 | thisarginfo.linear_step = 0; | |
2633 | thisarginfo.align = 0; | |
2634 | thisarginfo.op = NULL_TREE; | |
2635 | ||
2636 | op = gimple_call_arg (stmt, i); | |
2637 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, | |
2638 | &def_stmt, &def, &thisarginfo.dt, | |
2639 | &thisarginfo.vectype) | |
2640 | || thisarginfo.dt == vect_uninitialized_def) | |
2641 | { | |
2642 | if (dump_enabled_p ()) | |
2643 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2644 | "use not simple.\n"); | |
2645 | arginfo.release (); | |
2646 | return false; | |
2647 | } | |
2648 | ||
2649 | if (thisarginfo.dt == vect_constant_def | |
2650 | || thisarginfo.dt == vect_external_def) | |
2651 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
2652 | else | |
2653 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
2654 | ||
2655 | if (thisarginfo.dt != vect_constant_def | |
2656 | && thisarginfo.dt != vect_external_def | |
2657 | && loop_vinfo | |
2658 | && TREE_CODE (op) == SSA_NAME | |
2659 | && simple_iv (loop, loop_containing_stmt (stmt), op, &iv, false) | |
2660 | && tree_fits_shwi_p (iv.step)) | |
2661 | { | |
2662 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
2663 | thisarginfo.op = iv.base; | |
2664 | } | |
2665 | else if ((thisarginfo.dt == vect_constant_def | |
2666 | || thisarginfo.dt == vect_external_def) | |
2667 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
2668 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
2669 | ||
2670 | arginfo.quick_push (thisarginfo); | |
2671 | } | |
2672 | ||
2673 | unsigned int badness = 0; | |
2674 | struct cgraph_node *bestn = NULL; | |
2675 | if (STMT_VINFO_SIMD_CLONE_FNDECL (stmt_info)) | |
2676 | bestn = cgraph_get_node (STMT_VINFO_SIMD_CLONE_FNDECL (stmt_info)); | |
2677 | else | |
2678 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
2679 | n = n->simdclone->next_clone) | |
2680 | { | |
2681 | unsigned int this_badness = 0; | |
2682 | if (n->simdclone->simdlen | |
2683 | > (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
2684 | || n->simdclone->nargs != nargs) | |
2685 | continue; | |
2686 | if (n->simdclone->simdlen | |
2687 | < (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2688 | this_badness += (exact_log2 (LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2689 | - exact_log2 (n->simdclone->simdlen)) * 1024; | |
2690 | if (n->simdclone->inbranch) | |
2691 | this_badness += 2048; | |
2692 | int target_badness = targetm.simd_clone.usable (n); | |
2693 | if (target_badness < 0) | |
2694 | continue; | |
2695 | this_badness += target_badness * 512; | |
2696 | /* FORNOW: Have to add code to add the mask argument. */ | |
2697 | if (n->simdclone->inbranch) | |
2698 | continue; | |
2699 | for (i = 0; i < nargs; i++) | |
2700 | { | |
2701 | switch (n->simdclone->args[i].arg_type) | |
2702 | { | |
2703 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
2704 | if (!useless_type_conversion_p | |
2705 | (n->simdclone->args[i].orig_type, | |
2706 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
2707 | i = -1; | |
2708 | else if (arginfo[i].dt == vect_constant_def | |
2709 | || arginfo[i].dt == vect_external_def | |
2710 | || arginfo[i].linear_step) | |
2711 | this_badness += 64; | |
2712 | break; | |
2713 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
2714 | if (arginfo[i].dt != vect_constant_def | |
2715 | && arginfo[i].dt != vect_external_def) | |
2716 | i = -1; | |
2717 | break; | |
2718 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
2719 | if (arginfo[i].dt == vect_constant_def | |
2720 | || arginfo[i].dt == vect_external_def | |
2721 | || (arginfo[i].linear_step | |
2722 | != n->simdclone->args[i].linear_step)) | |
2723 | i = -1; | |
2724 | break; | |
2725 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
2726 | /* FORNOW */ | |
2727 | i = -1; | |
2728 | break; | |
2729 | case SIMD_CLONE_ARG_TYPE_MASK: | |
2730 | gcc_unreachable (); | |
2731 | } | |
2732 | if (i == (size_t) -1) | |
2733 | break; | |
2734 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
2735 | { | |
2736 | i = -1; | |
2737 | break; | |
2738 | } | |
2739 | if (arginfo[i].align) | |
2740 | this_badness += (exact_log2 (arginfo[i].align) | |
2741 | - exact_log2 (n->simdclone->args[i].alignment)); | |
2742 | } | |
2743 | if (i == (size_t) -1) | |
2744 | continue; | |
2745 | if (bestn == NULL || this_badness < badness) | |
2746 | { | |
2747 | bestn = n; | |
2748 | badness = this_badness; | |
2749 | } | |
2750 | } | |
2751 | ||
2752 | if (bestn == NULL) | |
2753 | { | |
2754 | arginfo.release (); | |
2755 | return false; | |
2756 | } | |
2757 | ||
2758 | for (i = 0; i < nargs; i++) | |
2759 | if ((arginfo[i].dt == vect_constant_def | |
2760 | || arginfo[i].dt == vect_external_def) | |
2761 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
2762 | { | |
2763 | arginfo[i].vectype | |
2764 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
2765 | i))); | |
2766 | if (arginfo[i].vectype == NULL | |
2767 | || (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
2768 | > bestn->simdclone->simdlen)) | |
2769 | { | |
2770 | arginfo.release (); | |
2771 | return false; | |
2772 | } | |
2773 | } | |
2774 | ||
2775 | fndecl = bestn->decl; | |
2776 | nunits = bestn->simdclone->simdlen; | |
2777 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2778 | ||
2779 | /* If the function isn't const, only allow it in simd loops where user | |
2780 | has asserted that at least nunits consecutive iterations can be | |
2781 | performed using SIMD instructions. */ | |
2782 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
2783 | && gimple_vuse (stmt)) | |
2784 | { | |
2785 | arginfo.release (); | |
2786 | return false; | |
2787 | } | |
2788 | ||
2789 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2790 | needs to be generated. */ | |
2791 | gcc_assert (ncopies >= 1); | |
2792 | ||
2793 | if (!vec_stmt) /* transformation not required. */ | |
2794 | { | |
2795 | STMT_VINFO_SIMD_CLONE_FNDECL (stmt_info) = bestn->decl; | |
2796 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; | |
2797 | if (dump_enabled_p ()) | |
2798 | dump_printf_loc (MSG_NOTE, vect_location, | |
2799 | "=== vectorizable_simd_clone_call ===\n"); | |
2800 | /* vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); */ | |
2801 | arginfo.release (); | |
2802 | return true; | |
2803 | } | |
2804 | ||
2805 | /** Transform. **/ | |
2806 | ||
2807 | if (dump_enabled_p ()) | |
2808 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
2809 | ||
2810 | /* Handle def. */ | |
2811 | scalar_dest = gimple_call_lhs (stmt); | |
2812 | vec_dest = NULL_TREE; | |
2813 | rtype = NULL_TREE; | |
2814 | ratype = NULL_TREE; | |
2815 | if (scalar_dest) | |
2816 | { | |
2817 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2818 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
2819 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
2820 | { | |
2821 | ratype = rtype; | |
2822 | rtype = TREE_TYPE (ratype); | |
2823 | } | |
2824 | } | |
2825 | ||
2826 | prev_stmt_info = NULL; | |
2827 | for (j = 0; j < ncopies; ++j) | |
2828 | { | |
2829 | /* Build argument list for the vectorized call. */ | |
2830 | if (j == 0) | |
2831 | vargs.create (nargs); | |
2832 | else | |
2833 | vargs.truncate (0); | |
2834 | ||
2835 | for (i = 0; i < nargs; i++) | |
2836 | { | |
2837 | unsigned int k, l, m, o; | |
2838 | tree atype; | |
2839 | op = gimple_call_arg (stmt, i); | |
2840 | switch (bestn->simdclone->args[i].arg_type) | |
2841 | { | |
2842 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
2843 | atype = bestn->simdclone->args[i].vector_type; | |
2844 | o = nunits / TYPE_VECTOR_SUBPARTS (atype); | |
2845 | for (m = j * o; m < (j + 1) * o; m++) | |
2846 | { | |
2847 | if (TYPE_VECTOR_SUBPARTS (atype) | |
2848 | < TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)) | |
2849 | { | |
2850 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); | |
2851 | k = (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
2852 | / TYPE_VECTOR_SUBPARTS (atype)); | |
2853 | gcc_assert ((k & (k - 1)) == 0); | |
2854 | if (m == 0) | |
2855 | vec_oprnd0 | |
2856 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
2857 | else | |
2858 | { | |
2859 | vec_oprnd0 = arginfo[i].op; | |
2860 | if ((m & (k - 1)) == 0) | |
2861 | vec_oprnd0 | |
2862 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
2863 | vec_oprnd0); | |
2864 | } | |
2865 | arginfo[i].op = vec_oprnd0; | |
2866 | vec_oprnd0 | |
2867 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
2868 | size_int (prec), | |
2869 | bitsize_int ((m & (k - 1)) * prec)); | |
2870 | new_stmt | |
2871 | = gimple_build_assign (make_ssa_name (atype, NULL), | |
2872 | vec_oprnd0); | |
2873 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2874 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
2875 | } | |
2876 | else | |
2877 | { | |
2878 | k = (TYPE_VECTOR_SUBPARTS (atype) | |
2879 | / TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)); | |
2880 | gcc_assert ((k & (k - 1)) == 0); | |
2881 | vec<constructor_elt, va_gc> *ctor_elts; | |
2882 | if (k != 1) | |
2883 | vec_alloc (ctor_elts, k); | |
2884 | else | |
2885 | ctor_elts = NULL; | |
2886 | for (l = 0; l < k; l++) | |
2887 | { | |
2888 | if (m == 0 && l == 0) | |
2889 | vec_oprnd0 | |
2890 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
2891 | else | |
2892 | vec_oprnd0 | |
2893 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
2894 | arginfo[i].op); | |
2895 | arginfo[i].op = vec_oprnd0; | |
2896 | if (k == 1) | |
2897 | break; | |
2898 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
2899 | vec_oprnd0); | |
2900 | } | |
2901 | if (k == 1) | |
2902 | vargs.safe_push (vec_oprnd0); | |
2903 | else | |
2904 | { | |
2905 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
2906 | new_stmt | |
2907 | = gimple_build_assign (make_ssa_name (atype, NULL), | |
2908 | vec_oprnd0); | |
2909 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2910 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
2911 | } | |
2912 | } | |
2913 | } | |
2914 | break; | |
2915 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
2916 | vargs.safe_push (op); | |
2917 | break; | |
2918 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
2919 | if (j == 0) | |
2920 | { | |
2921 | gimple_seq stmts; | |
2922 | arginfo[i].op | |
2923 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
2924 | NULL_TREE); | |
2925 | if (stmts != NULL) | |
2926 | { | |
2927 | basic_block new_bb; | |
2928 | edge pe = loop_preheader_edge (loop); | |
2929 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
2930 | gcc_assert (!new_bb); | |
2931 | } | |
2932 | tree phi_res = copy_ssa_name (op, NULL); | |
2933 | gimple new_phi = create_phi_node (phi_res, loop->header); | |
2934 | set_vinfo_for_stmt (new_phi, | |
2935 | new_stmt_vec_info (new_phi, loop_vinfo, | |
2936 | NULL)); | |
2937 | add_phi_arg (new_phi, arginfo[i].op, | |
2938 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
2939 | enum tree_code code | |
2940 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
2941 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
2942 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
2943 | ? sizetype : TREE_TYPE (op); | |
2944 | double_int cst | |
2945 | = double_int::from_shwi | |
2946 | (bestn->simdclone->args[i].linear_step); | |
2947 | cst *= double_int::from_uhwi (ncopies * nunits); | |
2948 | tree tcst = double_int_to_tree (type, cst); | |
2949 | tree phi_arg = copy_ssa_name (op, NULL); | |
2950 | new_stmt = gimple_build_assign_with_ops (code, phi_arg, | |
2951 | phi_res, tcst); | |
2952 | gimple_stmt_iterator si = gsi_after_labels (loop->header); | |
2953 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
2954 | set_vinfo_for_stmt (new_stmt, | |
2955 | new_stmt_vec_info (new_stmt, loop_vinfo, | |
2956 | NULL)); | |
2957 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), | |
2958 | UNKNOWN_LOCATION); | |
2959 | arginfo[i].op = phi_res; | |
2960 | vargs.safe_push (phi_res); | |
2961 | } | |
2962 | else | |
2963 | { | |
2964 | enum tree_code code | |
2965 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
2966 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
2967 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
2968 | ? sizetype : TREE_TYPE (op); | |
2969 | double_int cst | |
2970 | = double_int::from_shwi | |
2971 | (bestn->simdclone->args[i].linear_step); | |
2972 | cst *= double_int::from_uhwi (j * nunits); | |
2973 | tree tcst = double_int_to_tree (type, cst); | |
2974 | new_temp = make_ssa_name (TREE_TYPE (op), NULL); | |
2975 | new_stmt | |
2976 | = gimple_build_assign_with_ops (code, new_temp, | |
2977 | arginfo[i].op, tcst); | |
2978 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2979 | vargs.safe_push (new_temp); | |
2980 | } | |
2981 | break; | |
2982 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
2983 | default: | |
2984 | gcc_unreachable (); | |
2985 | } | |
2986 | } | |
2987 | ||
2988 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2989 | if (vec_dest) | |
2990 | { | |
2991 | gcc_assert (ratype || TYPE_VECTOR_SUBPARTS (rtype) == nunits); | |
2992 | if (ratype) | |
2993 | new_temp = create_tmp_var (ratype, NULL); | |
2994 | else if (TYPE_VECTOR_SUBPARTS (vectype) | |
2995 | == TYPE_VECTOR_SUBPARTS (rtype)) | |
2996 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2997 | else | |
2998 | new_temp = make_ssa_name (rtype, new_stmt); | |
2999 | gimple_call_set_lhs (new_stmt, new_temp); | |
3000 | } | |
3001 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3002 | ||
3003 | if (vec_dest) | |
3004 | { | |
3005 | if (TYPE_VECTOR_SUBPARTS (vectype) < nunits) | |
3006 | { | |
3007 | unsigned int k, l; | |
3008 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); | |
3009 | k = nunits / TYPE_VECTOR_SUBPARTS (vectype); | |
3010 | gcc_assert ((k & (k - 1)) == 0); | |
3011 | for (l = 0; l < k; l++) | |
3012 | { | |
3013 | tree t; | |
3014 | if (ratype) | |
3015 | { | |
3016 | t = build_fold_addr_expr (new_temp); | |
3017 | t = build2 (MEM_REF, vectype, t, | |
3018 | build_int_cst (TREE_TYPE (t), | |
3019 | l * prec / BITS_PER_UNIT)); | |
3020 | } | |
3021 | else | |
3022 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
3023 | size_int (prec), bitsize_int (l * prec)); | |
3024 | new_stmt | |
3025 | = gimple_build_assign (make_ssa_name (vectype, NULL), t); | |
3026 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3027 | if (j == 0 && l == 0) | |
3028 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3029 | else | |
3030 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3031 | ||
3032 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3033 | } | |
3034 | ||
3035 | if (ratype) | |
3036 | { | |
3037 | tree clobber = build_constructor (ratype, NULL); | |
3038 | TREE_THIS_VOLATILE (clobber) = 1; | |
3039 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3040 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3041 | } | |
3042 | continue; | |
3043 | } | |
3044 | else if (TYPE_VECTOR_SUBPARTS (vectype) > nunits) | |
3045 | { | |
3046 | unsigned int k = (TYPE_VECTOR_SUBPARTS (vectype) | |
3047 | / TYPE_VECTOR_SUBPARTS (rtype)); | |
3048 | gcc_assert ((k & (k - 1)) == 0); | |
3049 | if ((j & (k - 1)) == 0) | |
3050 | vec_alloc (ret_ctor_elts, k); | |
3051 | if (ratype) | |
3052 | { | |
3053 | unsigned int m, o = nunits / TYPE_VECTOR_SUBPARTS (rtype); | |
3054 | for (m = 0; m < o; m++) | |
3055 | { | |
3056 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
3057 | size_int (m), NULL_TREE, NULL_TREE); | |
3058 | new_stmt | |
3059 | = gimple_build_assign (make_ssa_name (rtype, NULL), | |
3060 | tem); | |
3061 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3062 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
3063 | gimple_assign_lhs (new_stmt)); | |
3064 | } | |
3065 | tree clobber = build_constructor (ratype, NULL); | |
3066 | TREE_THIS_VOLATILE (clobber) = 1; | |
3067 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3068 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3069 | } | |
3070 | else | |
3071 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
3072 | if ((j & (k - 1)) != k - 1) | |
3073 | continue; | |
3074 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
3075 | new_stmt | |
3076 | = gimple_build_assign (make_ssa_name (vec_dest, NULL), | |
3077 | vec_oprnd0); | |
3078 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3079 | ||
3080 | if ((unsigned) j == k - 1) | |
3081 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3082 | else | |
3083 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3084 | ||
3085 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3086 | continue; | |
3087 | } | |
3088 | else if (ratype) | |
3089 | { | |
3090 | tree t = build_fold_addr_expr (new_temp); | |
3091 | t = build2 (MEM_REF, vectype, t, | |
3092 | build_int_cst (TREE_TYPE (t), 0)); | |
3093 | new_stmt | |
3094 | = gimple_build_assign (make_ssa_name (vec_dest, NULL), t); | |
3095 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3096 | tree clobber = build_constructor (ratype, NULL); | |
3097 | TREE_THIS_VOLATILE (clobber) = 1; | |
3098 | vect_finish_stmt_generation (stmt, | |
3099 | gimple_build_assign (new_temp, | |
3100 | clobber), gsi); | |
3101 | } | |
3102 | } | |
3103 | ||
3104 | if (j == 0) | |
3105 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3106 | else | |
3107 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3108 | ||
3109 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3110 | } | |
3111 | ||
3112 | vargs.release (); | |
3113 | ||
3114 | /* The call in STMT might prevent it from being removed in dce. | |
3115 | We however cannot remove it here, due to the way the ssa name | |
3116 | it defines is mapped to the new definition. So just replace | |
3117 | rhs of the statement with something harmless. */ | |
3118 | ||
3119 | if (slp_node) | |
3120 | return true; | |
3121 | ||
3122 | if (scalar_dest) | |
3123 | { | |
3124 | type = TREE_TYPE (scalar_dest); | |
3125 | if (is_pattern_stmt_p (stmt_info)) | |
3126 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3127 | else | |
3128 | lhs = gimple_call_lhs (stmt); | |
3129 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
3130 | } | |
3131 | else | |
3132 | new_stmt = gimple_build_nop (); | |
3133 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
3134 | set_vinfo_for_stmt (stmt, NULL); | |
3135 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
3136 | gsi_replace (gsi, new_stmt, false); | |
3137 | unlink_stmt_vdef (stmt); | |
3138 | ||
3139 | return true; | |
3140 | } | |
3141 | ||
3142 | ||
ebfd146a IR |
3143 | /* Function vect_gen_widened_results_half |
3144 | ||
3145 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 3146 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 3147 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
3148 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
3149 | needs to be created (DECL is a function-decl of a target-builtin). | |
3150 | STMT is the original scalar stmt that we are vectorizing. */ | |
3151 | ||
3152 | static gimple | |
3153 | vect_gen_widened_results_half (enum tree_code code, | |
3154 | tree decl, | |
3155 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
3156 | tree vec_dest, gimple_stmt_iterator *gsi, | |
3157 | gimple stmt) | |
b8698a0f | 3158 | { |
ebfd146a | 3159 | gimple new_stmt; |
b8698a0f L |
3160 | tree new_temp; |
3161 | ||
3162 | /* Generate half of the widened result: */ | |
3163 | if (code == CALL_EXPR) | |
3164 | { | |
3165 | /* Target specific support */ | |
ebfd146a IR |
3166 | if (op_type == binary_op) |
3167 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
3168 | else | |
3169 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
3170 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3171 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
3172 | } |
3173 | else | |
ebfd146a | 3174 | { |
b8698a0f L |
3175 | /* Generic support */ |
3176 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
3177 | if (op_type != binary_op) |
3178 | vec_oprnd1 = NULL; | |
3179 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0, | |
3180 | vec_oprnd1); | |
3181 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3182 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 3183 | } |
ebfd146a IR |
3184 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3185 | ||
ebfd146a IR |
3186 | return new_stmt; |
3187 | } | |
3188 | ||
4a00c761 JJ |
3189 | |
3190 | /* Get vectorized definitions for loop-based vectorization. For the first | |
3191 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
3192 | scalar operand), and for the rest we get a copy with | |
3193 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
3194 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
3195 | The vectors are collected into VEC_OPRNDS. */ | |
3196 | ||
3197 | static void | |
3198 | vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt, | |
9771b263 | 3199 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
3200 | { |
3201 | tree vec_oprnd; | |
3202 | ||
3203 | /* Get first vector operand. */ | |
3204 | /* All the vector operands except the very first one (that is scalar oprnd) | |
3205 | are stmt copies. */ | |
3206 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
3207 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL); | |
3208 | else | |
3209 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
3210 | ||
9771b263 | 3211 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3212 | |
3213 | /* Get second vector operand. */ | |
3214 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 3215 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3216 | |
3217 | *oprnd = vec_oprnd; | |
3218 | ||
3219 | /* For conversion in multiple steps, continue to get operands | |
3220 | recursively. */ | |
3221 | if (multi_step_cvt) | |
3222 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
3223 | } | |
3224 | ||
3225 | ||
3226 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
3227 | For multi-step conversions store the resulting vectors and call the function | |
3228 | recursively. */ | |
3229 | ||
3230 | static void | |
9771b263 | 3231 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
4a00c761 | 3232 | int multi_step_cvt, gimple stmt, |
9771b263 | 3233 | vec<tree> vec_dsts, |
4a00c761 JJ |
3234 | gimple_stmt_iterator *gsi, |
3235 | slp_tree slp_node, enum tree_code code, | |
3236 | stmt_vec_info *prev_stmt_info) | |
3237 | { | |
3238 | unsigned int i; | |
3239 | tree vop0, vop1, new_tmp, vec_dest; | |
3240 | gimple new_stmt; | |
3241 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3242 | ||
9771b263 | 3243 | vec_dest = vec_dsts.pop (); |
4a00c761 | 3244 | |
9771b263 | 3245 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
3246 | { |
3247 | /* Create demotion operation. */ | |
9771b263 DN |
3248 | vop0 = (*vec_oprnds)[i]; |
3249 | vop1 = (*vec_oprnds)[i + 1]; | |
4a00c761 JJ |
3250 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); |
3251 | new_tmp = make_ssa_name (vec_dest, new_stmt); | |
3252 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
3253 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3254 | ||
3255 | if (multi_step_cvt) | |
3256 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 3257 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
3258 | else |
3259 | { | |
3260 | /* This is the last step of the conversion sequence. Store the | |
3261 | vectors in SLP_NODE or in vector info of the scalar statement | |
3262 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
3263 | if (slp_node) | |
9771b263 | 3264 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
3265 | else |
3266 | { | |
3267 | if (!*prev_stmt_info) | |
3268 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3269 | else | |
3270 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
3271 | ||
3272 | *prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3273 | } | |
3274 | } | |
3275 | } | |
3276 | ||
3277 | /* For multi-step demotion operations we first generate demotion operations | |
3278 | from the source type to the intermediate types, and then combine the | |
3279 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
3280 | type. */ | |
3281 | if (multi_step_cvt) | |
3282 | { | |
3283 | /* At each level of recursion we have half of the operands we had at the | |
3284 | previous level. */ | |
9771b263 | 3285 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
3286 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
3287 | stmt, vec_dsts, gsi, slp_node, | |
3288 | VEC_PACK_TRUNC_EXPR, | |
3289 | prev_stmt_info); | |
3290 | } | |
3291 | ||
9771b263 | 3292 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3293 | } |
3294 | ||
3295 | ||
3296 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
3297 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
3298 | the resulting vectors and call the function recursively. */ | |
3299 | ||
3300 | static void | |
9771b263 DN |
3301 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
3302 | vec<tree> *vec_oprnds1, | |
4a00c761 JJ |
3303 | gimple stmt, tree vec_dest, |
3304 | gimple_stmt_iterator *gsi, | |
3305 | enum tree_code code1, | |
3306 | enum tree_code code2, tree decl1, | |
3307 | tree decl2, int op_type) | |
3308 | { | |
3309 | int i; | |
3310 | tree vop0, vop1, new_tmp1, new_tmp2; | |
3311 | gimple new_stmt1, new_stmt2; | |
6e1aa848 | 3312 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 3313 | |
9771b263 DN |
3314 | vec_tmp.create (vec_oprnds0->length () * 2); |
3315 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
3316 | { |
3317 | if (op_type == binary_op) | |
9771b263 | 3318 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
3319 | else |
3320 | vop1 = NULL_TREE; | |
3321 | ||
3322 | /* Generate the two halves of promotion operation. */ | |
3323 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
3324 | op_type, vec_dest, gsi, stmt); | |
3325 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
3326 | op_type, vec_dest, gsi, stmt); | |
3327 | if (is_gimple_call (new_stmt1)) | |
3328 | { | |
3329 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
3330 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
3331 | } | |
3332 | else | |
3333 | { | |
3334 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
3335 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
3336 | } | |
3337 | ||
3338 | /* Store the results for the next step. */ | |
9771b263 DN |
3339 | vec_tmp.quick_push (new_tmp1); |
3340 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
3341 | } |
3342 | ||
689eaba3 | 3343 | vec_oprnds0->release (); |
4a00c761 JJ |
3344 | *vec_oprnds0 = vec_tmp; |
3345 | } | |
3346 | ||
3347 | ||
b8698a0f L |
3348 | /* Check if STMT performs a conversion operation, that can be vectorized. |
3349 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 3350 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
3351 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
3352 | ||
3353 | static bool | |
3354 | vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi, | |
3355 | gimple *vec_stmt, slp_tree slp_node) | |
3356 | { | |
3357 | tree vec_dest; | |
3358 | tree scalar_dest; | |
4a00c761 | 3359 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
3360 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
3361 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3362 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3363 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 3364 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
3365 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
3366 | tree new_temp; | |
3367 | tree def; | |
3368 | gimple def_stmt; | |
3369 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
3370 | gimple new_stmt = NULL; | |
3371 | stmt_vec_info prev_stmt_info; | |
3372 | int nunits_in; | |
3373 | int nunits_out; | |
3374 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
3375 | int ncopies, i, j; |
3376 | tree lhs_type, rhs_type; | |
ebfd146a | 3377 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
3378 | vec<tree> vec_oprnds0 = vNULL; |
3379 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 3380 | tree vop0; |
4a00c761 JJ |
3381 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3382 | int multi_step_cvt = 0; | |
6e1aa848 DN |
3383 | vec<tree> vec_dsts = vNULL; |
3384 | vec<tree> interm_types = vNULL; | |
4a00c761 JJ |
3385 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
3386 | int op_type; | |
3387 | enum machine_mode rhs_mode; | |
3388 | unsigned short fltsz; | |
ebfd146a IR |
3389 | |
3390 | /* Is STMT a vectorizable conversion? */ | |
3391 | ||
4a00c761 | 3392 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3393 | return false; |
3394 | ||
8644a673 | 3395 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3396 | return false; |
3397 | ||
3398 | if (!is_gimple_assign (stmt)) | |
3399 | return false; | |
3400 | ||
3401 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3402 | return false; | |
3403 | ||
3404 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
3405 | if (!CONVERT_EXPR_CODE_P (code) |
3406 | && code != FIX_TRUNC_EXPR | |
3407 | && code != FLOAT_EXPR | |
3408 | && code != WIDEN_MULT_EXPR | |
3409 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
3410 | return false; |
3411 | ||
4a00c761 JJ |
3412 | op_type = TREE_CODE_LENGTH (code); |
3413 | ||
ebfd146a | 3414 | /* Check types of lhs and rhs. */ |
b690cc0f | 3415 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 3416 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
3417 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3418 | ||
ebfd146a IR |
3419 | op0 = gimple_assign_rhs1 (stmt); |
3420 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
3421 | |
3422 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3423 | && !((INTEGRAL_TYPE_P (lhs_type) | |
3424 | && INTEGRAL_TYPE_P (rhs_type)) | |
3425 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
3426 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
3427 | return false; | |
3428 | ||
3429 | if ((INTEGRAL_TYPE_P (lhs_type) | |
3430 | && (TYPE_PRECISION (lhs_type) | |
3431 | != GET_MODE_PRECISION (TYPE_MODE (lhs_type)))) | |
3432 | || (INTEGRAL_TYPE_P (rhs_type) | |
3433 | && (TYPE_PRECISION (rhs_type) | |
3434 | != GET_MODE_PRECISION (TYPE_MODE (rhs_type))))) | |
3435 | { | |
73fbfcad | 3436 | if (dump_enabled_p ()) |
78c60e3d | 3437 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
3438 | "type conversion to/from bit-precision unsupported." |
3439 | "\n"); | |
4a00c761 JJ |
3440 | return false; |
3441 | } | |
3442 | ||
b690cc0f | 3443 | /* Check the operands of the operation. */ |
24ee1384 | 3444 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
b690cc0f RG |
3445 | &def_stmt, &def, &dt[0], &vectype_in)) |
3446 | { | |
73fbfcad | 3447 | if (dump_enabled_p ()) |
78c60e3d | 3448 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3449 | "use not simple.\n"); |
b690cc0f RG |
3450 | return false; |
3451 | } | |
4a00c761 JJ |
3452 | if (op_type == binary_op) |
3453 | { | |
3454 | bool ok; | |
3455 | ||
3456 | op1 = gimple_assign_rhs2 (stmt); | |
3457 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
3458 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
3459 | OP1. */ | |
3460 | if (CONSTANT_CLASS_P (op0)) | |
f5709183 | 3461 | ok = vect_is_simple_use_1 (op1, stmt, loop_vinfo, bb_vinfo, |
4a00c761 JJ |
3462 | &def_stmt, &def, &dt[1], &vectype_in); |
3463 | else | |
f5709183 | 3464 | ok = vect_is_simple_use (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
24ee1384 | 3465 | &def, &dt[1]); |
4a00c761 JJ |
3466 | |
3467 | if (!ok) | |
3468 | { | |
73fbfcad | 3469 | if (dump_enabled_p ()) |
78c60e3d | 3470 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3471 | "use not simple.\n"); |
4a00c761 JJ |
3472 | return false; |
3473 | } | |
3474 | } | |
3475 | ||
b690cc0f RG |
3476 | /* If op0 is an external or constant defs use a vector type of |
3477 | the same size as the output vector type. */ | |
ebfd146a | 3478 | if (!vectype_in) |
b690cc0f | 3479 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
3480 | if (vec_stmt) |
3481 | gcc_assert (vectype_in); | |
3482 | if (!vectype_in) | |
3483 | { | |
73fbfcad | 3484 | if (dump_enabled_p ()) |
4a00c761 | 3485 | { |
78c60e3d SS |
3486 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3487 | "no vectype for scalar type "); | |
3488 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 3489 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 3490 | } |
7d8930a0 IR |
3491 | |
3492 | return false; | |
3493 | } | |
ebfd146a | 3494 | |
b690cc0f RG |
3495 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
3496 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 3497 | if (nunits_in < nunits_out) |
ebfd146a IR |
3498 | modifier = NARROW; |
3499 | else if (nunits_out == nunits_in) | |
3500 | modifier = NONE; | |
ebfd146a | 3501 | else |
4a00c761 | 3502 | modifier = WIDEN; |
ebfd146a | 3503 | |
ff802fa1 IR |
3504 | /* Multiple types in SLP are handled by creating the appropriate number of |
3505 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3506 | case of SLP. */ | |
437f4a00 | 3507 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a | 3508 | ncopies = 1; |
4a00c761 JJ |
3509 | else if (modifier == NARROW) |
3510 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
3511 | else | |
3512 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 3513 | |
ebfd146a IR |
3514 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
3515 | needs to be generated. */ | |
3516 | gcc_assert (ncopies >= 1); | |
3517 | ||
ebfd146a | 3518 | /* Supportable by target? */ |
4a00c761 | 3519 | switch (modifier) |
ebfd146a | 3520 | { |
4a00c761 JJ |
3521 | case NONE: |
3522 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3523 | return false; | |
3524 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
3525 | &decl1, &code1)) | |
3526 | break; | |
3527 | /* FALLTHRU */ | |
3528 | unsupported: | |
73fbfcad | 3529 | if (dump_enabled_p ()) |
78c60e3d | 3530 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3531 | "conversion not supported by target.\n"); |
ebfd146a | 3532 | return false; |
ebfd146a | 3533 | |
4a00c761 JJ |
3534 | case WIDEN: |
3535 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
3536 | &code1, &code2, &multi_step_cvt, |
3537 | &interm_types)) | |
4a00c761 JJ |
3538 | { |
3539 | /* Binary widening operation can only be supported directly by the | |
3540 | architecture. */ | |
3541 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
3542 | break; | |
3543 | } | |
3544 | ||
3545 | if (code != FLOAT_EXPR | |
3546 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3547 | <= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3548 | goto unsupported; | |
3549 | ||
3550 | rhs_mode = TYPE_MODE (rhs_type); | |
3551 | fltsz = GET_MODE_SIZE (TYPE_MODE (lhs_type)); | |
3552 | for (rhs_mode = GET_MODE_2XWIDER_MODE (TYPE_MODE (rhs_type)); | |
3553 | rhs_mode != VOIDmode && GET_MODE_SIZE (rhs_mode) <= fltsz; | |
3554 | rhs_mode = GET_MODE_2XWIDER_MODE (rhs_mode)) | |
3555 | { | |
3556 | cvt_type | |
3557 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3558 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3559 | if (cvt_type == NULL_TREE) | |
3560 | goto unsupported; | |
3561 | ||
3562 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3563 | { | |
3564 | if (!supportable_convert_operation (code, vectype_out, | |
3565 | cvt_type, &decl1, &codecvt1)) | |
3566 | goto unsupported; | |
3567 | } | |
3568 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
3569 | cvt_type, &codecvt1, |
3570 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
3571 | &interm_types)) |
3572 | continue; | |
3573 | else | |
3574 | gcc_assert (multi_step_cvt == 0); | |
3575 | ||
3576 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
3577 | vectype_in, &code1, &code2, |
3578 | &multi_step_cvt, &interm_types)) | |
4a00c761 JJ |
3579 | break; |
3580 | } | |
3581 | ||
3582 | if (rhs_mode == VOIDmode || GET_MODE_SIZE (rhs_mode) > fltsz) | |
3583 | goto unsupported; | |
3584 | ||
3585 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3586 | codecvt2 = ERROR_MARK; | |
3587 | else | |
3588 | { | |
3589 | multi_step_cvt++; | |
9771b263 | 3590 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
3591 | cvt_type = NULL_TREE; |
3592 | } | |
3593 | break; | |
3594 | ||
3595 | case NARROW: | |
3596 | gcc_assert (op_type == unary_op); | |
3597 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
3598 | &code1, &multi_step_cvt, | |
3599 | &interm_types)) | |
3600 | break; | |
3601 | ||
3602 | if (code != FIX_TRUNC_EXPR | |
3603 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3604 | >= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3605 | goto unsupported; | |
3606 | ||
3607 | rhs_mode = TYPE_MODE (rhs_type); | |
3608 | cvt_type | |
3609 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3610 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3611 | if (cvt_type == NULL_TREE) | |
3612 | goto unsupported; | |
3613 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
3614 | &decl1, &codecvt1)) | |
3615 | goto unsupported; | |
3616 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
3617 | &code1, &multi_step_cvt, | |
3618 | &interm_types)) | |
3619 | break; | |
3620 | goto unsupported; | |
3621 | ||
3622 | default: | |
3623 | gcc_unreachable (); | |
ebfd146a IR |
3624 | } |
3625 | ||
3626 | if (!vec_stmt) /* transformation not required. */ | |
3627 | { | |
73fbfcad | 3628 | if (dump_enabled_p ()) |
78c60e3d | 3629 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3630 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 3631 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
3632 | { |
3633 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
c3e7ee41 | 3634 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
8bd37302 | 3635 | } |
4a00c761 JJ |
3636 | else if (modifier == NARROW) |
3637 | { | |
3638 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 3639 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
3640 | } |
3641 | else | |
3642 | { | |
3643 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 3644 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 3645 | } |
9771b263 | 3646 | interm_types.release (); |
ebfd146a IR |
3647 | return true; |
3648 | } | |
3649 | ||
3650 | /** Transform. **/ | |
73fbfcad | 3651 | if (dump_enabled_p ()) |
78c60e3d | 3652 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3653 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 3654 | |
4a00c761 JJ |
3655 | if (op_type == binary_op) |
3656 | { | |
3657 | if (CONSTANT_CLASS_P (op0)) | |
3658 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
3659 | else if (CONSTANT_CLASS_P (op1)) | |
3660 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
3661 | } | |
3662 | ||
3663 | /* In case of multi-step conversion, we first generate conversion operations | |
3664 | to the intermediate types, and then from that types to the final one. | |
3665 | We create vector destinations for the intermediate type (TYPES) received | |
3666 | from supportable_*_operation, and store them in the correct order | |
3667 | for future use in vect_create_vectorized_*_stmts (). */ | |
9771b263 | 3668 | vec_dsts.create (multi_step_cvt + 1); |
82294ec1 JJ |
3669 | vec_dest = vect_create_destination_var (scalar_dest, |
3670 | (cvt_type && modifier == WIDEN) | |
3671 | ? cvt_type : vectype_out); | |
9771b263 | 3672 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3673 | |
3674 | if (multi_step_cvt) | |
3675 | { | |
9771b263 DN |
3676 | for (i = interm_types.length () - 1; |
3677 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
3678 | { |
3679 | vec_dest = vect_create_destination_var (scalar_dest, | |
3680 | intermediate_type); | |
9771b263 | 3681 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3682 | } |
3683 | } | |
ebfd146a | 3684 | |
4a00c761 | 3685 | if (cvt_type) |
82294ec1 JJ |
3686 | vec_dest = vect_create_destination_var (scalar_dest, |
3687 | modifier == WIDEN | |
3688 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
3689 | |
3690 | if (!slp_node) | |
3691 | { | |
30862efc | 3692 | if (modifier == WIDEN) |
4a00c761 | 3693 | { |
c3284718 | 3694 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 3695 | if (op_type == binary_op) |
9771b263 | 3696 | vec_oprnds1.create (1); |
4a00c761 | 3697 | } |
30862efc | 3698 | else if (modifier == NARROW) |
9771b263 DN |
3699 | vec_oprnds0.create ( |
3700 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
3701 | } |
3702 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 3703 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 3704 | |
4a00c761 | 3705 | last_oprnd = op0; |
ebfd146a IR |
3706 | prev_stmt_info = NULL; |
3707 | switch (modifier) | |
3708 | { | |
3709 | case NONE: | |
3710 | for (j = 0; j < ncopies; j++) | |
3711 | { | |
ebfd146a | 3712 | if (j == 0) |
d092494c IR |
3713 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node, |
3714 | -1); | |
ebfd146a IR |
3715 | else |
3716 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
3717 | ||
9771b263 | 3718 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
3719 | { |
3720 | /* Arguments are ready, create the new vector stmt. */ | |
3721 | if (code1 == CALL_EXPR) | |
3722 | { | |
3723 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
3724 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3725 | gimple_call_set_lhs (new_stmt, new_temp); | |
3726 | } | |
3727 | else | |
3728 | { | |
3729 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
3730 | new_stmt = gimple_build_assign_with_ops (code1, vec_dest, | |
3731 | vop0, NULL); | |
3732 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3733 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3734 | } | |
3735 | ||
3736 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3737 | if (slp_node) | |
9771b263 | 3738 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
3739 | } |
3740 | ||
ebfd146a IR |
3741 | if (j == 0) |
3742 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3743 | else | |
3744 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3745 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3746 | } | |
3747 | break; | |
3748 | ||
3749 | case WIDEN: | |
3750 | /* In case the vectorization factor (VF) is bigger than the number | |
3751 | of elements that we can fit in a vectype (nunits), we have to | |
3752 | generate more than one vector stmt - i.e - we need to "unroll" | |
3753 | the vector stmt by a factor VF/nunits. */ | |
3754 | for (j = 0; j < ncopies; j++) | |
3755 | { | |
4a00c761 | 3756 | /* Handle uses. */ |
ebfd146a | 3757 | if (j == 0) |
4a00c761 JJ |
3758 | { |
3759 | if (slp_node) | |
3760 | { | |
3761 | if (code == WIDEN_LSHIFT_EXPR) | |
3762 | { | |
3763 | unsigned int k; | |
ebfd146a | 3764 | |
4a00c761 JJ |
3765 | vec_oprnd1 = op1; |
3766 | /* Store vec_oprnd1 for every vector stmt to be created | |
3767 | for SLP_NODE. We check during the analysis that all | |
3768 | the shift arguments are the same. */ | |
3769 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 3770 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
3771 | |
3772 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
3773 | slp_node, -1); | |
3774 | } | |
3775 | else | |
3776 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
3777 | &vec_oprnds1, slp_node, -1); | |
3778 | } | |
3779 | else | |
3780 | { | |
3781 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
9771b263 | 3782 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
3783 | if (op_type == binary_op) |
3784 | { | |
3785 | if (code == WIDEN_LSHIFT_EXPR) | |
3786 | vec_oprnd1 = op1; | |
3787 | else | |
3788 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, | |
3789 | NULL); | |
9771b263 | 3790 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
3791 | } |
3792 | } | |
3793 | } | |
ebfd146a | 3794 | else |
4a00c761 JJ |
3795 | { |
3796 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
3797 | vec_oprnds0.truncate (0); |
3798 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
3799 | if (op_type == binary_op) |
3800 | { | |
3801 | if (code == WIDEN_LSHIFT_EXPR) | |
3802 | vec_oprnd1 = op1; | |
3803 | else | |
3804 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
3805 | vec_oprnd1); | |
9771b263 DN |
3806 | vec_oprnds1.truncate (0); |
3807 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
3808 | } |
3809 | } | |
ebfd146a | 3810 | |
4a00c761 JJ |
3811 | /* Arguments are ready. Create the new vector stmts. */ |
3812 | for (i = multi_step_cvt; i >= 0; i--) | |
3813 | { | |
9771b263 | 3814 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
3815 | enum tree_code c1 = code1, c2 = code2; |
3816 | if (i == 0 && codecvt2 != ERROR_MARK) | |
3817 | { | |
3818 | c1 = codecvt1; | |
3819 | c2 = codecvt2; | |
3820 | } | |
3821 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
3822 | &vec_oprnds1, | |
3823 | stmt, this_dest, gsi, | |
3824 | c1, c2, decl1, decl2, | |
3825 | op_type); | |
3826 | } | |
3827 | ||
9771b263 | 3828 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
3829 | { |
3830 | if (cvt_type) | |
3831 | { | |
3832 | if (codecvt1 == CALL_EXPR) | |
3833 | { | |
3834 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
3835 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3836 | gimple_call_set_lhs (new_stmt, new_temp); | |
3837 | } | |
3838 | else | |
3839 | { | |
3840 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
3841 | new_temp = make_ssa_name (vec_dest, NULL); | |
3842 | new_stmt = gimple_build_assign_with_ops (codecvt1, | |
3843 | new_temp, | |
3844 | vop0, NULL); | |
3845 | } | |
3846 | ||
3847 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3848 | } | |
3849 | else | |
3850 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
3851 | ||
3852 | if (slp_node) | |
9771b263 | 3853 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
3854 | else |
3855 | { | |
3856 | if (!prev_stmt_info) | |
3857 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3858 | else | |
3859 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3860 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3861 | } | |
3862 | } | |
ebfd146a | 3863 | } |
4a00c761 JJ |
3864 | |
3865 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
3866 | break; |
3867 | ||
3868 | case NARROW: | |
3869 | /* In case the vectorization factor (VF) is bigger than the number | |
3870 | of elements that we can fit in a vectype (nunits), we have to | |
3871 | generate more than one vector stmt - i.e - we need to "unroll" | |
3872 | the vector stmt by a factor VF/nunits. */ | |
3873 | for (j = 0; j < ncopies; j++) | |
3874 | { | |
3875 | /* Handle uses. */ | |
4a00c761 JJ |
3876 | if (slp_node) |
3877 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
3878 | slp_node, -1); | |
ebfd146a IR |
3879 | else |
3880 | { | |
9771b263 | 3881 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
3882 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
3883 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
3884 | } |
3885 | ||
4a00c761 JJ |
3886 | /* Arguments are ready. Create the new vector stmts. */ |
3887 | if (cvt_type) | |
9771b263 | 3888 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
3889 | { |
3890 | if (codecvt1 == CALL_EXPR) | |
3891 | { | |
3892 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
3893 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3894 | gimple_call_set_lhs (new_stmt, new_temp); | |
3895 | } | |
3896 | else | |
3897 | { | |
3898 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
3899 | new_temp = make_ssa_name (vec_dest, NULL); | |
3900 | new_stmt = gimple_build_assign_with_ops (codecvt1, new_temp, | |
3901 | vop0, NULL); | |
3902 | } | |
ebfd146a | 3903 | |
4a00c761 | 3904 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 3905 | vec_oprnds0[i] = new_temp; |
4a00c761 | 3906 | } |
ebfd146a | 3907 | |
4a00c761 JJ |
3908 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
3909 | stmt, vec_dsts, gsi, | |
3910 | slp_node, code1, | |
3911 | &prev_stmt_info); | |
ebfd146a IR |
3912 | } |
3913 | ||
3914 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 3915 | break; |
ebfd146a IR |
3916 | } |
3917 | ||
9771b263 DN |
3918 | vec_oprnds0.release (); |
3919 | vec_oprnds1.release (); | |
3920 | vec_dsts.release (); | |
3921 | interm_types.release (); | |
ebfd146a IR |
3922 | |
3923 | return true; | |
3924 | } | |
ff802fa1 IR |
3925 | |
3926 | ||
ebfd146a IR |
3927 | /* Function vectorizable_assignment. |
3928 | ||
b8698a0f L |
3929 | Check if STMT performs an assignment (copy) that can be vectorized. |
3930 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
3931 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3932 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3933 | ||
3934 | static bool | |
3935 | vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi, | |
3936 | gimple *vec_stmt, slp_tree slp_node) | |
3937 | { | |
3938 | tree vec_dest; | |
3939 | tree scalar_dest; | |
3940 | tree op; | |
3941 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3942 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3943 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3944 | tree new_temp; | |
3945 | tree def; | |
3946 | gimple def_stmt; | |
3947 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
fde9c428 | 3948 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
ebfd146a | 3949 | int ncopies; |
f18b55bd | 3950 | int i, j; |
6e1aa848 | 3951 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 3952 | tree vop; |
a70d6342 | 3953 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
f18b55bd IR |
3954 | gimple new_stmt = NULL; |
3955 | stmt_vec_info prev_stmt_info = NULL; | |
fde9c428 RG |
3956 | enum tree_code code; |
3957 | tree vectype_in; | |
ebfd146a IR |
3958 | |
3959 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3960 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3961 | case of SLP. */ | |
437f4a00 | 3962 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
3963 | ncopies = 1; |
3964 | else | |
3965 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3966 | ||
3967 | gcc_assert (ncopies >= 1); | |
ebfd146a | 3968 | |
a70d6342 | 3969 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3970 | return false; |
3971 | ||
8644a673 | 3972 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3973 | return false; |
3974 | ||
3975 | /* Is vectorizable assignment? */ | |
3976 | if (!is_gimple_assign (stmt)) | |
3977 | return false; | |
3978 | ||
3979 | scalar_dest = gimple_assign_lhs (stmt); | |
3980 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
3981 | return false; | |
3982 | ||
fde9c428 | 3983 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 3984 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
3985 | || code == PAREN_EXPR |
3986 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
3987 | op = gimple_assign_rhs1 (stmt); |
3988 | else | |
3989 | return false; | |
3990 | ||
7b7ec6c5 RG |
3991 | if (code == VIEW_CONVERT_EXPR) |
3992 | op = TREE_OPERAND (op, 0); | |
3993 | ||
24ee1384 | 3994 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, |
fde9c428 | 3995 | &def_stmt, &def, &dt[0], &vectype_in)) |
ebfd146a | 3996 | { |
73fbfcad | 3997 | if (dump_enabled_p ()) |
78c60e3d | 3998 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3999 | "use not simple.\n"); |
ebfd146a IR |
4000 | return false; |
4001 | } | |
4002 | ||
fde9c428 RG |
4003 | /* We can handle NOP_EXPR conversions that do not change the number |
4004 | of elements or the vector size. */ | |
7b7ec6c5 RG |
4005 | if ((CONVERT_EXPR_CODE_P (code) |
4006 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
4007 | && (!vectype_in |
4008 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
4009 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
4010 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
4011 | return false; | |
4012 | ||
7b7b1813 RG |
4013 | /* We do not handle bit-precision changes. */ |
4014 | if ((CONVERT_EXPR_CODE_P (code) | |
4015 | || code == VIEW_CONVERT_EXPR) | |
4016 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
4017 | && ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4018 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4019 | || ((TYPE_PRECISION (TREE_TYPE (op)) | |
4020 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))) | |
4021 | /* But a conversion that does not change the bit-pattern is ok. */ | |
4022 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4023 | > TYPE_PRECISION (TREE_TYPE (op))) | |
4024 | && TYPE_UNSIGNED (TREE_TYPE (op)))) | |
4025 | { | |
73fbfcad | 4026 | if (dump_enabled_p ()) |
78c60e3d SS |
4027 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4028 | "type conversion to/from bit-precision " | |
e645e942 | 4029 | "unsupported.\n"); |
7b7b1813 RG |
4030 | return false; |
4031 | } | |
4032 | ||
ebfd146a IR |
4033 | if (!vec_stmt) /* transformation not required. */ |
4034 | { | |
4035 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 4036 | if (dump_enabled_p ()) |
78c60e3d | 4037 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4038 | "=== vectorizable_assignment ===\n"); |
c3e7ee41 | 4039 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
4040 | return true; |
4041 | } | |
4042 | ||
4043 | /** Transform. **/ | |
73fbfcad | 4044 | if (dump_enabled_p ()) |
e645e942 | 4045 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
4046 | |
4047 | /* Handle def. */ | |
4048 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4049 | ||
4050 | /* Handle use. */ | |
f18b55bd | 4051 | for (j = 0; j < ncopies; j++) |
ebfd146a | 4052 | { |
f18b55bd IR |
4053 | /* Handle uses. */ |
4054 | if (j == 0) | |
d092494c | 4055 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node, -1); |
f18b55bd IR |
4056 | else |
4057 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
4058 | ||
4059 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 4060 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 4061 | { |
7b7ec6c5 RG |
4062 | if (CONVERT_EXPR_CODE_P (code) |
4063 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 4064 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
4065 | new_stmt = gimple_build_assign (vec_dest, vop); |
4066 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4067 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4068 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4069 | if (slp_node) | |
9771b263 | 4070 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 4071 | } |
ebfd146a IR |
4072 | |
4073 | if (slp_node) | |
f18b55bd IR |
4074 | continue; |
4075 | ||
4076 | if (j == 0) | |
4077 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4078 | else | |
4079 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4080 | ||
4081 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4082 | } | |
b8698a0f | 4083 | |
9771b263 | 4084 | vec_oprnds.release (); |
ebfd146a IR |
4085 | return true; |
4086 | } | |
4087 | ||
9dc3f7de | 4088 | |
1107f3ae IR |
4089 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
4090 | either as shift by a scalar or by a vector. */ | |
4091 | ||
4092 | bool | |
4093 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
4094 | { | |
4095 | ||
4096 | enum machine_mode vec_mode; | |
4097 | optab optab; | |
4098 | int icode; | |
4099 | tree vectype; | |
4100 | ||
4101 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4102 | if (!vectype) | |
4103 | return false; | |
4104 | ||
4105 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4106 | if (!optab | |
4107 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
4108 | { | |
4109 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4110 | if (!optab | |
4111 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
4112 | == CODE_FOR_nothing)) | |
4113 | return false; | |
4114 | } | |
4115 | ||
4116 | vec_mode = TYPE_MODE (vectype); | |
4117 | icode = (int) optab_handler (optab, vec_mode); | |
4118 | if (icode == CODE_FOR_nothing) | |
4119 | return false; | |
4120 | ||
4121 | return true; | |
4122 | } | |
4123 | ||
4124 | ||
9dc3f7de IR |
4125 | /* Function vectorizable_shift. |
4126 | ||
4127 | Check if STMT performs a shift operation that can be vectorized. | |
4128 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4129 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
4130 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4131 | ||
4132 | static bool | |
4133 | vectorizable_shift (gimple stmt, gimple_stmt_iterator *gsi, | |
4134 | gimple *vec_stmt, slp_tree slp_node) | |
4135 | { | |
4136 | tree vec_dest; | |
4137 | tree scalar_dest; | |
4138 | tree op0, op1 = NULL; | |
4139 | tree vec_oprnd1 = NULL_TREE; | |
4140 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4141 | tree vectype; | |
4142 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4143 | enum tree_code code; | |
4144 | enum machine_mode vec_mode; | |
4145 | tree new_temp; | |
4146 | optab optab; | |
4147 | int icode; | |
4148 | enum machine_mode optab_op2_mode; | |
4149 | tree def; | |
4150 | gimple def_stmt; | |
4151 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
4152 | gimple new_stmt = NULL; | |
4153 | stmt_vec_info prev_stmt_info; | |
4154 | int nunits_in; | |
4155 | int nunits_out; | |
4156 | tree vectype_out; | |
cede2577 | 4157 | tree op1_vectype; |
9dc3f7de IR |
4158 | int ncopies; |
4159 | int j, i; | |
6e1aa848 DN |
4160 | vec<tree> vec_oprnds0 = vNULL; |
4161 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
4162 | tree vop0, vop1; |
4163 | unsigned int k; | |
49eab32e | 4164 | bool scalar_shift_arg = true; |
9dc3f7de IR |
4165 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
4166 | int vf; | |
4167 | ||
4168 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
4169 | return false; | |
4170 | ||
4171 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
4172 | return false; | |
4173 | ||
4174 | /* Is STMT a vectorizable binary/unary operation? */ | |
4175 | if (!is_gimple_assign (stmt)) | |
4176 | return false; | |
4177 | ||
4178 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4179 | return false; | |
4180 | ||
4181 | code = gimple_assign_rhs_code (stmt); | |
4182 | ||
4183 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
4184 | || code == RROTATE_EXPR)) | |
4185 | return false; | |
4186 | ||
4187 | scalar_dest = gimple_assign_lhs (stmt); | |
4188 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
7b7b1813 RG |
4189 | if (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
4190 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4191 | { | |
73fbfcad | 4192 | if (dump_enabled_p ()) |
78c60e3d | 4193 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4194 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
4195 | return false; |
4196 | } | |
9dc3f7de IR |
4197 | |
4198 | op0 = gimple_assign_rhs1 (stmt); | |
24ee1384 | 4199 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
9dc3f7de IR |
4200 | &def_stmt, &def, &dt[0], &vectype)) |
4201 | { | |
73fbfcad | 4202 | if (dump_enabled_p ()) |
78c60e3d | 4203 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4204 | "use not simple.\n"); |
9dc3f7de IR |
4205 | return false; |
4206 | } | |
4207 | /* If op0 is an external or constant def use a vector type with | |
4208 | the same size as the output vector type. */ | |
4209 | if (!vectype) | |
4210 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4211 | if (vec_stmt) | |
4212 | gcc_assert (vectype); | |
4213 | if (!vectype) | |
4214 | { | |
73fbfcad | 4215 | if (dump_enabled_p ()) |
78c60e3d | 4216 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4217 | "no vectype for scalar type\n"); |
9dc3f7de IR |
4218 | return false; |
4219 | } | |
4220 | ||
4221 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4222 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4223 | if (nunits_out != nunits_in) | |
4224 | return false; | |
4225 | ||
4226 | op1 = gimple_assign_rhs2 (stmt); | |
24ee1384 IR |
4227 | if (!vect_is_simple_use_1 (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
4228 | &def, &dt[1], &op1_vectype)) | |
9dc3f7de | 4229 | { |
73fbfcad | 4230 | if (dump_enabled_p ()) |
78c60e3d | 4231 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4232 | "use not simple.\n"); |
9dc3f7de IR |
4233 | return false; |
4234 | } | |
4235 | ||
4236 | if (loop_vinfo) | |
4237 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4238 | else | |
4239 | vf = 1; | |
4240 | ||
4241 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4242 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4243 | case of SLP. */ | |
437f4a00 | 4244 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
9dc3f7de IR |
4245 | ncopies = 1; |
4246 | else | |
4247 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4248 | ||
4249 | gcc_assert (ncopies >= 1); | |
4250 | ||
4251 | /* Determine whether the shift amount is a vector, or scalar. If the | |
4252 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
4253 | ||
49eab32e JJ |
4254 | if (dt[1] == vect_internal_def && !slp_node) |
4255 | scalar_shift_arg = false; | |
4256 | else if (dt[1] == vect_constant_def | |
4257 | || dt[1] == vect_external_def | |
4258 | || dt[1] == vect_internal_def) | |
4259 | { | |
4260 | /* In SLP, need to check whether the shift count is the same, | |
4261 | in loops if it is a constant or invariant, it is always | |
4262 | a scalar shift. */ | |
4263 | if (slp_node) | |
4264 | { | |
9771b263 | 4265 | vec<gimple> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
49eab32e JJ |
4266 | gimple slpstmt; |
4267 | ||
9771b263 | 4268 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
4269 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
4270 | scalar_shift_arg = false; | |
4271 | } | |
4272 | } | |
4273 | else | |
4274 | { | |
73fbfcad | 4275 | if (dump_enabled_p ()) |
78c60e3d | 4276 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4277 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
4278 | return false; |
4279 | } | |
4280 | ||
9dc3f7de | 4281 | /* Vector shifted by vector. */ |
49eab32e | 4282 | if (!scalar_shift_arg) |
9dc3f7de IR |
4283 | { |
4284 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 4285 | if (dump_enabled_p ()) |
78c60e3d | 4286 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4287 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 4288 | |
aa948027 JJ |
4289 | if (!op1_vectype) |
4290 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
4291 | if (op1_vectype == NULL_TREE | |
4292 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 4293 | { |
73fbfcad | 4294 | if (dump_enabled_p ()) |
78c60e3d SS |
4295 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4296 | "unusable type for last operand in" | |
e645e942 | 4297 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
4298 | return false; |
4299 | } | |
9dc3f7de IR |
4300 | } |
4301 | /* See if the machine has a vector shifted by scalar insn and if not | |
4302 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 4303 | else |
9dc3f7de IR |
4304 | { |
4305 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4306 | if (optab | |
4307 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
4308 | { | |
73fbfcad | 4309 | if (dump_enabled_p ()) |
78c60e3d | 4310 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4311 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
4312 | } |
4313 | else | |
4314 | { | |
4315 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4316 | if (optab | |
4317 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
4318 | != CODE_FOR_nothing)) | |
4319 | { | |
49eab32e JJ |
4320 | scalar_shift_arg = false; |
4321 | ||
73fbfcad | 4322 | if (dump_enabled_p ()) |
78c60e3d | 4323 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4324 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
4325 | |
4326 | /* Unlike the other binary operators, shifts/rotates have | |
4327 | the rhs being int, instead of the same type as the lhs, | |
4328 | so make sure the scalar is the right type if we are | |
aa948027 | 4329 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
4330 | if (dt[1] == vect_constant_def) |
4331 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
4332 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
4333 | TREE_TYPE (op1))) | |
4334 | { | |
4335 | if (slp_node | |
4336 | && TYPE_MODE (TREE_TYPE (vectype)) | |
4337 | != TYPE_MODE (TREE_TYPE (op1))) | |
4338 | { | |
73fbfcad | 4339 | if (dump_enabled_p ()) |
78c60e3d SS |
4340 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4341 | "unusable type for last operand in" | |
e645e942 | 4342 | " vector/vector shift/rotate.\n"); |
aa948027 JJ |
4343 | return false; |
4344 | } | |
4345 | if (vec_stmt && !slp_node) | |
4346 | { | |
4347 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
4348 | op1 = vect_init_vector (stmt, op1, | |
4349 | TREE_TYPE (vectype), NULL); | |
4350 | } | |
4351 | } | |
9dc3f7de IR |
4352 | } |
4353 | } | |
4354 | } | |
9dc3f7de IR |
4355 | |
4356 | /* Supportable by target? */ | |
4357 | if (!optab) | |
4358 | { | |
73fbfcad | 4359 | if (dump_enabled_p ()) |
78c60e3d | 4360 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4361 | "no optab.\n"); |
9dc3f7de IR |
4362 | return false; |
4363 | } | |
4364 | vec_mode = TYPE_MODE (vectype); | |
4365 | icode = (int) optab_handler (optab, vec_mode); | |
4366 | if (icode == CODE_FOR_nothing) | |
4367 | { | |
73fbfcad | 4368 | if (dump_enabled_p ()) |
78c60e3d | 4369 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4370 | "op not supported by target.\n"); |
9dc3f7de IR |
4371 | /* Check only during analysis. */ |
4372 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
4373 | || (vf < vect_min_worthwhile_factor (code) | |
4374 | && !vec_stmt)) | |
4375 | return false; | |
73fbfcad | 4376 | if (dump_enabled_p ()) |
e645e942 TJ |
4377 | dump_printf_loc (MSG_NOTE, vect_location, |
4378 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
4379 | } |
4380 | ||
4381 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
4382 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
4383 | && vf < vect_min_worthwhile_factor (code) | |
4384 | && !vec_stmt) | |
4385 | { | |
73fbfcad | 4386 | if (dump_enabled_p ()) |
78c60e3d | 4387 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4388 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
4389 | return false; |
4390 | } | |
4391 | ||
4392 | if (!vec_stmt) /* transformation not required. */ | |
4393 | { | |
4394 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 4395 | if (dump_enabled_p ()) |
e645e942 TJ |
4396 | dump_printf_loc (MSG_NOTE, vect_location, |
4397 | "=== vectorizable_shift ===\n"); | |
c3e7ee41 | 4398 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
9dc3f7de IR |
4399 | return true; |
4400 | } | |
4401 | ||
4402 | /** Transform. **/ | |
4403 | ||
73fbfcad | 4404 | if (dump_enabled_p ()) |
78c60e3d | 4405 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4406 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
4407 | |
4408 | /* Handle def. */ | |
4409 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4410 | ||
9dc3f7de IR |
4411 | prev_stmt_info = NULL; |
4412 | for (j = 0; j < ncopies; j++) | |
4413 | { | |
4414 | /* Handle uses. */ | |
4415 | if (j == 0) | |
4416 | { | |
4417 | if (scalar_shift_arg) | |
4418 | { | |
4419 | /* Vector shl and shr insn patterns can be defined with scalar | |
4420 | operand 2 (shift operand). In this case, use constant or loop | |
4421 | invariant op1 directly, without extending it to vector mode | |
4422 | first. */ | |
4423 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
4424 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
4425 | { | |
73fbfcad | 4426 | if (dump_enabled_p ()) |
78c60e3d | 4427 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4428 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 4429 | vec_oprnd1 = op1; |
8930f723 | 4430 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 4431 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4432 | if (slp_node) |
4433 | { | |
4434 | /* Store vec_oprnd1 for every vector stmt to be created | |
4435 | for SLP_NODE. We check during the analysis that all | |
4436 | the shift arguments are the same. | |
4437 | TODO: Allow different constants for different vector | |
4438 | stmts generated for an SLP instance. */ | |
4439 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4440 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4441 | } |
4442 | } | |
4443 | } | |
4444 | ||
4445 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
4446 | (a special case for certain kind of vector shifts); otherwise, | |
4447 | operand 1 should be of a vector type (the usual case). */ | |
4448 | if (vec_oprnd1) | |
4449 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
d092494c | 4450 | slp_node, -1); |
9dc3f7de IR |
4451 | else |
4452 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
d092494c | 4453 | slp_node, -1); |
9dc3f7de IR |
4454 | } |
4455 | else | |
4456 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
4457 | ||
4458 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 4459 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 4460 | { |
9771b263 | 4461 | vop1 = vec_oprnds1[i]; |
9dc3f7de IR |
4462 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); |
4463 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4464 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4465 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4466 | if (slp_node) | |
9771b263 | 4467 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
4468 | } |
4469 | ||
4470 | if (slp_node) | |
4471 | continue; | |
4472 | ||
4473 | if (j == 0) | |
4474 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4475 | else | |
4476 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4477 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4478 | } | |
4479 | ||
9771b263 DN |
4480 | vec_oprnds0.release (); |
4481 | vec_oprnds1.release (); | |
9dc3f7de IR |
4482 | |
4483 | return true; | |
4484 | } | |
4485 | ||
4486 | ||
ebfd146a IR |
4487 | /* Function vectorizable_operation. |
4488 | ||
16949072 RG |
4489 | Check if STMT performs a binary, unary or ternary operation that can |
4490 | be vectorized. | |
b8698a0f | 4491 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
4492 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4493 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4494 | ||
4495 | static bool | |
4496 | vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi, | |
4497 | gimple *vec_stmt, slp_tree slp_node) | |
4498 | { | |
00f07b86 | 4499 | tree vec_dest; |
ebfd146a | 4500 | tree scalar_dest; |
16949072 | 4501 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 4502 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 4503 | tree vectype; |
ebfd146a IR |
4504 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4505 | enum tree_code code; | |
4506 | enum machine_mode vec_mode; | |
4507 | tree new_temp; | |
4508 | int op_type; | |
00f07b86 | 4509 | optab optab; |
ebfd146a | 4510 | int icode; |
ebfd146a IR |
4511 | tree def; |
4512 | gimple def_stmt; | |
16949072 RG |
4513 | enum vect_def_type dt[3] |
4514 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
ebfd146a IR |
4515 | gimple new_stmt = NULL; |
4516 | stmt_vec_info prev_stmt_info; | |
b690cc0f | 4517 | int nunits_in; |
ebfd146a IR |
4518 | int nunits_out; |
4519 | tree vectype_out; | |
4520 | int ncopies; | |
4521 | int j, i; | |
6e1aa848 DN |
4522 | vec<tree> vec_oprnds0 = vNULL; |
4523 | vec<tree> vec_oprnds1 = vNULL; | |
4524 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 4525 | tree vop0, vop1, vop2; |
a70d6342 IR |
4526 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
4527 | int vf; | |
4528 | ||
a70d6342 | 4529 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4530 | return false; |
4531 | ||
8644a673 | 4532 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4533 | return false; |
4534 | ||
4535 | /* Is STMT a vectorizable binary/unary operation? */ | |
4536 | if (!is_gimple_assign (stmt)) | |
4537 | return false; | |
4538 | ||
4539 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4540 | return false; | |
4541 | ||
ebfd146a IR |
4542 | code = gimple_assign_rhs_code (stmt); |
4543 | ||
4544 | /* For pointer addition, we should use the normal plus for | |
4545 | the vector addition. */ | |
4546 | if (code == POINTER_PLUS_EXPR) | |
4547 | code = PLUS_EXPR; | |
4548 | ||
4549 | /* Support only unary or binary operations. */ | |
4550 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 4551 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 4552 | { |
73fbfcad | 4553 | if (dump_enabled_p ()) |
78c60e3d | 4554 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4555 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 4556 | op_type); |
ebfd146a IR |
4557 | return false; |
4558 | } | |
4559 | ||
b690cc0f RG |
4560 | scalar_dest = gimple_assign_lhs (stmt); |
4561 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
4562 | ||
7b7b1813 RG |
4563 | /* Most operations cannot handle bit-precision types without extra |
4564 | truncations. */ | |
4565 | if ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4566 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4567 | /* Exception are bitwise binary operations. */ | |
4568 | && code != BIT_IOR_EXPR | |
4569 | && code != BIT_XOR_EXPR | |
4570 | && code != BIT_AND_EXPR) | |
4571 | { | |
73fbfcad | 4572 | if (dump_enabled_p ()) |
78c60e3d | 4573 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4574 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
4575 | return false; |
4576 | } | |
4577 | ||
ebfd146a | 4578 | op0 = gimple_assign_rhs1 (stmt); |
24ee1384 | 4579 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
b690cc0f | 4580 | &def_stmt, &def, &dt[0], &vectype)) |
ebfd146a | 4581 | { |
73fbfcad | 4582 | if (dump_enabled_p ()) |
78c60e3d | 4583 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4584 | "use not simple.\n"); |
ebfd146a IR |
4585 | return false; |
4586 | } | |
b690cc0f RG |
4587 | /* If op0 is an external or constant def use a vector type with |
4588 | the same size as the output vector type. */ | |
4589 | if (!vectype) | |
4590 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
4591 | if (vec_stmt) |
4592 | gcc_assert (vectype); | |
4593 | if (!vectype) | |
4594 | { | |
73fbfcad | 4595 | if (dump_enabled_p ()) |
7d8930a0 | 4596 | { |
78c60e3d SS |
4597 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4598 | "no vectype for scalar type "); | |
4599 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
4600 | TREE_TYPE (op0)); | |
e645e942 | 4601 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
4602 | } |
4603 | ||
4604 | return false; | |
4605 | } | |
b690cc0f RG |
4606 | |
4607 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4608 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4609 | if (nunits_out != nunits_in) | |
4610 | return false; | |
ebfd146a | 4611 | |
16949072 | 4612 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
4613 | { |
4614 | op1 = gimple_assign_rhs2 (stmt); | |
24ee1384 IR |
4615 | if (!vect_is_simple_use (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
4616 | &def, &dt[1])) | |
ebfd146a | 4617 | { |
73fbfcad | 4618 | if (dump_enabled_p ()) |
78c60e3d | 4619 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4620 | "use not simple.\n"); |
ebfd146a IR |
4621 | return false; |
4622 | } | |
4623 | } | |
16949072 RG |
4624 | if (op_type == ternary_op) |
4625 | { | |
4626 | op2 = gimple_assign_rhs3 (stmt); | |
24ee1384 IR |
4627 | if (!vect_is_simple_use (op2, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
4628 | &def, &dt[2])) | |
16949072 | 4629 | { |
73fbfcad | 4630 | if (dump_enabled_p ()) |
78c60e3d | 4631 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4632 | "use not simple.\n"); |
16949072 RG |
4633 | return false; |
4634 | } | |
4635 | } | |
ebfd146a | 4636 | |
b690cc0f RG |
4637 | if (loop_vinfo) |
4638 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4639 | else | |
4640 | vf = 1; | |
4641 | ||
4642 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4643 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 4644 | case of SLP. */ |
437f4a00 | 4645 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
b690cc0f RG |
4646 | ncopies = 1; |
4647 | else | |
4648 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4649 | ||
4650 | gcc_assert (ncopies >= 1); | |
4651 | ||
9dc3f7de | 4652 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
4653 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
4654 | || code == RROTATE_EXPR) | |
9dc3f7de | 4655 | return false; |
ebfd146a | 4656 | |
ebfd146a | 4657 | /* Supportable by target? */ |
00f07b86 RH |
4658 | |
4659 | vec_mode = TYPE_MODE (vectype); | |
4660 | if (code == MULT_HIGHPART_EXPR) | |
ebfd146a | 4661 | { |
00f07b86 | 4662 | if (can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype))) |
dee54b6e | 4663 | icode = LAST_INSN_CODE; |
00f07b86 RH |
4664 | else |
4665 | icode = CODE_FOR_nothing; | |
ebfd146a | 4666 | } |
00f07b86 RH |
4667 | else |
4668 | { | |
4669 | optab = optab_for_tree_code (code, vectype, optab_default); | |
4670 | if (!optab) | |
5deb57cb | 4671 | { |
73fbfcad | 4672 | if (dump_enabled_p ()) |
78c60e3d | 4673 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4674 | "no optab.\n"); |
00f07b86 | 4675 | return false; |
5deb57cb | 4676 | } |
00f07b86 | 4677 | icode = (int) optab_handler (optab, vec_mode); |
5deb57cb JJ |
4678 | } |
4679 | ||
ebfd146a IR |
4680 | if (icode == CODE_FOR_nothing) |
4681 | { | |
73fbfcad | 4682 | if (dump_enabled_p ()) |
78c60e3d | 4683 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4684 | "op not supported by target.\n"); |
ebfd146a IR |
4685 | /* Check only during analysis. */ |
4686 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5deb57cb | 4687 | || (!vec_stmt && vf < vect_min_worthwhile_factor (code))) |
ebfd146a | 4688 | return false; |
73fbfcad | 4689 | if (dump_enabled_p ()) |
e645e942 TJ |
4690 | dump_printf_loc (MSG_NOTE, vect_location, |
4691 | "proceeding using word mode.\n"); | |
383d9c83 IR |
4692 | } |
4693 | ||
4a00c761 | 4694 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
4695 | if (!VECTOR_MODE_P (vec_mode) |
4696 | && !vec_stmt | |
4697 | && vf < vect_min_worthwhile_factor (code)) | |
7d8930a0 | 4698 | { |
73fbfcad | 4699 | if (dump_enabled_p ()) |
78c60e3d | 4700 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4701 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 4702 | return false; |
7d8930a0 | 4703 | } |
ebfd146a | 4704 | |
ebfd146a IR |
4705 | if (!vec_stmt) /* transformation not required. */ |
4706 | { | |
4a00c761 | 4707 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 4708 | if (dump_enabled_p ()) |
78c60e3d | 4709 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4710 | "=== vectorizable_operation ===\n"); |
c3e7ee41 | 4711 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
4712 | return true; |
4713 | } | |
4714 | ||
4715 | /** Transform. **/ | |
4716 | ||
73fbfcad | 4717 | if (dump_enabled_p ()) |
78c60e3d | 4718 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4719 | "transform binary/unary operation.\n"); |
383d9c83 | 4720 | |
ebfd146a | 4721 | /* Handle def. */ |
00f07b86 | 4722 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 4723 | |
ebfd146a IR |
4724 | /* In case the vectorization factor (VF) is bigger than the number |
4725 | of elements that we can fit in a vectype (nunits), we have to generate | |
4726 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
4727 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
4728 | from one copy of the vector stmt to the next, in the field | |
4729 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
4730 | stages to find the correct vector defs to be used when vectorizing | |
4731 | stmts that use the defs of the current stmt. The example below | |
4732 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
4733 | we need to create 4 vectorized stmts): | |
4734 | ||
4735 | before vectorization: | |
4736 | RELATED_STMT VEC_STMT | |
4737 | S1: x = memref - - | |
4738 | S2: z = x + 1 - - | |
4739 | ||
4740 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
4741 | there): | |
4742 | RELATED_STMT VEC_STMT | |
4743 | VS1_0: vx0 = memref0 VS1_1 - | |
4744 | VS1_1: vx1 = memref1 VS1_2 - | |
4745 | VS1_2: vx2 = memref2 VS1_3 - | |
4746 | VS1_3: vx3 = memref3 - - | |
4747 | S1: x = load - VS1_0 | |
4748 | S2: z = x + 1 - - | |
4749 | ||
4750 | step2: vectorize stmt S2 (done here): | |
4751 | To vectorize stmt S2 we first need to find the relevant vector | |
4752 | def for the first operand 'x'. This is, as usual, obtained from | |
4753 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
4754 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
4755 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
4756 | the vector stmt VS2_0, and as usual, record it in the | |
4757 | STMT_VINFO_VEC_STMT of stmt S2. | |
4758 | When creating the second copy (VS2_1), we obtain the relevant vector | |
4759 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
4760 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
4761 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
4762 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
4763 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
4764 | chain of stmts and pointers: | |
4765 | RELATED_STMT VEC_STMT | |
4766 | VS1_0: vx0 = memref0 VS1_1 - | |
4767 | VS1_1: vx1 = memref1 VS1_2 - | |
4768 | VS1_2: vx2 = memref2 VS1_3 - | |
4769 | VS1_3: vx3 = memref3 - - | |
4770 | S1: x = load - VS1_0 | |
4771 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
4772 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
4773 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
4774 | VS2_3: vz3 = vx3 + v1 - - | |
4775 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
4776 | |
4777 | prev_stmt_info = NULL; | |
4778 | for (j = 0; j < ncopies; j++) | |
4779 | { | |
4780 | /* Handle uses. */ | |
4781 | if (j == 0) | |
4a00c761 JJ |
4782 | { |
4783 | if (op_type == binary_op || op_type == ternary_op) | |
4784 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
4785 | slp_node, -1); | |
4786 | else | |
4787 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
4788 | slp_node, -1); | |
4789 | if (op_type == ternary_op) | |
36ba4aae | 4790 | { |
9771b263 DN |
4791 | vec_oprnds2.create (1); |
4792 | vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2, | |
4793 | stmt, | |
4794 | NULL)); | |
36ba4aae | 4795 | } |
4a00c761 | 4796 | } |
ebfd146a | 4797 | else |
4a00c761 JJ |
4798 | { |
4799 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
4800 | if (op_type == ternary_op) | |
4801 | { | |
9771b263 DN |
4802 | tree vec_oprnd = vec_oprnds2.pop (); |
4803 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
4804 | vec_oprnd)); | |
4a00c761 JJ |
4805 | } |
4806 | } | |
4807 | ||
4808 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 4809 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 4810 | { |
4a00c761 | 4811 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 4812 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 4813 | vop2 = ((op_type == ternary_op) |
9771b263 | 4814 | ? vec_oprnds2[i] : NULL_TREE); |
73804b12 RG |
4815 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, |
4816 | vop0, vop1, vop2); | |
4a00c761 JJ |
4817 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4818 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4819 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4820 | if (slp_node) | |
9771b263 | 4821 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
4822 | } |
4823 | ||
4a00c761 JJ |
4824 | if (slp_node) |
4825 | continue; | |
4826 | ||
4827 | if (j == 0) | |
4828 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4829 | else | |
4830 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4831 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
4832 | } |
4833 | ||
9771b263 DN |
4834 | vec_oprnds0.release (); |
4835 | vec_oprnds1.release (); | |
4836 | vec_oprnds2.release (); | |
ebfd146a | 4837 | |
ebfd146a IR |
4838 | return true; |
4839 | } | |
4840 | ||
c716e67f XDL |
4841 | /* A helper function to ensure data reference DR's base alignment |
4842 | for STMT_INFO. */ | |
4843 | ||
4844 | static void | |
4845 | ensure_base_align (stmt_vec_info stmt_info, struct data_reference *dr) | |
4846 | { | |
4847 | if (!dr->aux) | |
4848 | return; | |
4849 | ||
4850 | if (((dataref_aux *)dr->aux)->base_misaligned) | |
4851 | { | |
4852 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
4853 | tree base_decl = ((dataref_aux *)dr->aux)->base_decl; | |
4854 | ||
4855 | DECL_ALIGN (base_decl) = TYPE_ALIGN (vectype); | |
4856 | DECL_USER_ALIGN (base_decl) = 1; | |
4857 | ((dataref_aux *)dr->aux)->base_misaligned = false; | |
4858 | } | |
4859 | } | |
4860 | ||
ebfd146a | 4861 | |
09dfa495 BM |
4862 | /* Given a vector type VECTYPE returns the VECTOR_CST mask that implements |
4863 | reversal of the vector elements. If that is impossible to do, | |
4864 | returns NULL. */ | |
4865 | ||
4866 | static tree | |
4867 | perm_mask_for_reverse (tree vectype) | |
4868 | { | |
4869 | int i, nunits; | |
4870 | unsigned char *sel; | |
4871 | ||
4872 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4873 | sel = XALLOCAVEC (unsigned char, nunits); | |
4874 | ||
4875 | for (i = 0; i < nunits; ++i) | |
4876 | sel[i] = nunits - 1 - i; | |
4877 | ||
4878 | return vect_gen_perm_mask (vectype, sel); | |
4879 | } | |
4880 | ||
ebfd146a IR |
4881 | /* Function vectorizable_store. |
4882 | ||
b8698a0f L |
4883 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
4884 | can be vectorized. | |
4885 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4886 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4887 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4888 | ||
4889 | static bool | |
4890 | vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
c716e67f | 4891 | slp_tree slp_node) |
ebfd146a IR |
4892 | { |
4893 | tree scalar_dest; | |
4894 | tree data_ref; | |
4895 | tree op; | |
4896 | tree vec_oprnd = NULL_TREE; | |
4897 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4898 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
4899 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
272c6793 | 4900 | tree elem_type; |
ebfd146a | 4901 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 4902 | struct loop *loop = NULL; |
ebfd146a IR |
4903 | enum machine_mode vec_mode; |
4904 | tree dummy; | |
4905 | enum dr_alignment_support alignment_support_scheme; | |
4906 | tree def; | |
4907 | gimple def_stmt; | |
4908 | enum vect_def_type dt; | |
4909 | stmt_vec_info prev_stmt_info = NULL; | |
4910 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 4911 | tree dataref_offset = NULL_TREE; |
fef4d2b3 | 4912 | gimple ptr_incr = NULL; |
ebfd146a IR |
4913 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
4914 | int ncopies; | |
4915 | int j; | |
4916 | gimple next_stmt, first_stmt = NULL; | |
0d0293ac | 4917 | bool grouped_store = false; |
272c6793 | 4918 | bool store_lanes_p = false; |
ebfd146a | 4919 | unsigned int group_size, i; |
6e1aa848 DN |
4920 | vec<tree> dr_chain = vNULL; |
4921 | vec<tree> oprnds = vNULL; | |
4922 | vec<tree> result_chain = vNULL; | |
ebfd146a | 4923 | bool inv_p; |
09dfa495 BM |
4924 | bool negative = false; |
4925 | tree offset = NULL_TREE; | |
6e1aa848 | 4926 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 4927 | bool slp = (slp_node != NULL); |
ebfd146a | 4928 | unsigned int vec_num; |
a70d6342 | 4929 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
272c6793 | 4930 | tree aggr_type; |
a70d6342 IR |
4931 | |
4932 | if (loop_vinfo) | |
4933 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
4934 | |
4935 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4936 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4937 | case of SLP. */ | |
437f4a00 | 4938 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
4939 | ncopies = 1; |
4940 | else | |
4941 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4942 | ||
4943 | gcc_assert (ncopies >= 1); | |
4944 | ||
4945 | /* FORNOW. This restriction should be relaxed. */ | |
a70d6342 | 4946 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) |
ebfd146a | 4947 | { |
73fbfcad | 4948 | if (dump_enabled_p ()) |
78c60e3d | 4949 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4950 | "multiple types in nested loop.\n"); |
ebfd146a IR |
4951 | return false; |
4952 | } | |
4953 | ||
a70d6342 | 4954 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4955 | return false; |
4956 | ||
8644a673 | 4957 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4958 | return false; |
4959 | ||
4960 | /* Is vectorizable store? */ | |
4961 | ||
4962 | if (!is_gimple_assign (stmt)) | |
4963 | return false; | |
4964 | ||
4965 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
4966 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
4967 | && is_pattern_stmt_p (stmt_info)) | |
4968 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 4969 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 4970 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 4971 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
4972 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
4973 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
4974 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
4975 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
4976 | return false; |
4977 | ||
4978 | gcc_assert (gimple_assign_single_p (stmt)); | |
4979 | op = gimple_assign_rhs1 (stmt); | |
24ee1384 IR |
4980 | if (!vect_is_simple_use (op, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
4981 | &def, &dt)) | |
ebfd146a | 4982 | { |
73fbfcad | 4983 | if (dump_enabled_p ()) |
78c60e3d | 4984 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4985 | "use not simple.\n"); |
ebfd146a IR |
4986 | return false; |
4987 | } | |
4988 | ||
272c6793 | 4989 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 4990 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 4991 | |
ebfd146a IR |
4992 | /* FORNOW. In some cases can vectorize even if data-type not supported |
4993 | (e.g. - array initialization with 0). */ | |
947131ba | 4994 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
4995 | return false; |
4996 | ||
4997 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
4998 | return false; | |
4999 | ||
09dfa495 BM |
5000 | negative = |
5001 | tree_int_cst_compare (loop && nested_in_vect_loop_p (loop, stmt) | |
5002 | ? STMT_VINFO_DR_STEP (stmt_info) : DR_STEP (dr), | |
5003 | size_zero_node) < 0; | |
5004 | if (negative && ncopies > 1) | |
a1e53f3f | 5005 | { |
73fbfcad | 5006 | if (dump_enabled_p ()) |
78c60e3d | 5007 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
09dfa495 | 5008 | "multiple types with negative step."); |
a1e53f3f L |
5009 | return false; |
5010 | } | |
5011 | ||
09dfa495 BM |
5012 | if (negative) |
5013 | { | |
5014 | gcc_assert (!grouped_store); | |
5015 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
5016 | if (alignment_support_scheme != dr_aligned | |
5017 | && alignment_support_scheme != dr_unaligned_supported) | |
5018 | { | |
5019 | if (dump_enabled_p ()) | |
5020 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5021 | "negative step but alignment required."); | |
5022 | return false; | |
5023 | } | |
5024 | if (!perm_mask_for_reverse (vectype)) | |
5025 | { | |
5026 | if (dump_enabled_p ()) | |
5027 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5028 | "negative step and reversing not supported."); | |
5029 | return false; | |
5030 | } | |
5031 | } | |
5032 | ||
0d0293ac | 5033 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 5034 | { |
0d0293ac | 5035 | grouped_store = true; |
e14c1050 | 5036 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
5037 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
5038 | { | |
e14c1050 | 5039 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
5040 | if (vect_store_lanes_supported (vectype, group_size)) |
5041 | store_lanes_p = true; | |
0d0293ac | 5042 | else if (!vect_grouped_store_supported (vectype, group_size)) |
b602d918 RS |
5043 | return false; |
5044 | } | |
b8698a0f | 5045 | |
ebfd146a IR |
5046 | if (first_stmt == stmt) |
5047 | { | |
5048 | /* STMT is the leader of the group. Check the operands of all the | |
5049 | stmts of the group. */ | |
e14c1050 | 5050 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
5051 | while (next_stmt) |
5052 | { | |
5053 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5054 | op = gimple_assign_rhs1 (next_stmt); | |
24ee1384 IR |
5055 | if (!vect_is_simple_use (op, next_stmt, loop_vinfo, bb_vinfo, |
5056 | &def_stmt, &def, &dt)) | |
ebfd146a | 5057 | { |
73fbfcad | 5058 | if (dump_enabled_p ()) |
78c60e3d | 5059 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5060 | "use not simple.\n"); |
ebfd146a IR |
5061 | return false; |
5062 | } | |
e14c1050 | 5063 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5064 | } |
5065 | } | |
5066 | } | |
5067 | ||
5068 | if (!vec_stmt) /* transformation not required. */ | |
5069 | { | |
5070 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
92345349 BS |
5071 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, |
5072 | NULL, NULL, NULL); | |
ebfd146a IR |
5073 | return true; |
5074 | } | |
5075 | ||
5076 | /** Transform. **/ | |
5077 | ||
c716e67f XDL |
5078 | ensure_base_align (stmt_info, dr); |
5079 | ||
0d0293ac | 5080 | if (grouped_store) |
ebfd146a IR |
5081 | { |
5082 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 5083 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 5084 | |
e14c1050 | 5085 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
5086 | |
5087 | /* FORNOW */ | |
a70d6342 | 5088 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
5089 | |
5090 | /* We vectorize all the stmts of the interleaving group when we | |
5091 | reach the last stmt in the group. */ | |
e14c1050 IR |
5092 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
5093 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
5094 | && !slp) |
5095 | { | |
5096 | *vec_stmt = NULL; | |
5097 | return true; | |
5098 | } | |
5099 | ||
5100 | if (slp) | |
4b5caab7 | 5101 | { |
0d0293ac | 5102 | grouped_store = false; |
4b5caab7 IR |
5103 | /* VEC_NUM is the number of vect stmts to be created for this |
5104 | group. */ | |
5105 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 5106 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
4b5caab7 | 5107 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 5108 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 5109 | } |
ebfd146a | 5110 | else |
4b5caab7 IR |
5111 | /* VEC_NUM is the number of vect stmts to be created for this |
5112 | group. */ | |
ebfd146a IR |
5113 | vec_num = group_size; |
5114 | } | |
b8698a0f | 5115 | else |
ebfd146a IR |
5116 | { |
5117 | first_stmt = stmt; | |
5118 | first_dr = dr; | |
5119 | group_size = vec_num = 1; | |
ebfd146a | 5120 | } |
b8698a0f | 5121 | |
73fbfcad | 5122 | if (dump_enabled_p ()) |
78c60e3d | 5123 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5124 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 5125 | |
9771b263 DN |
5126 | dr_chain.create (group_size); |
5127 | oprnds.create (group_size); | |
ebfd146a | 5128 | |
720f5239 | 5129 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 5130 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
5131 | /* Targets with store-lane instructions must not require explicit |
5132 | realignment. */ | |
5133 | gcc_assert (!store_lanes_p | |
5134 | || alignment_support_scheme == dr_aligned | |
5135 | || alignment_support_scheme == dr_unaligned_supported); | |
5136 | ||
09dfa495 BM |
5137 | if (negative) |
5138 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
5139 | ||
272c6793 RS |
5140 | if (store_lanes_p) |
5141 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
5142 | else | |
5143 | aggr_type = vectype; | |
ebfd146a IR |
5144 | |
5145 | /* In case the vectorization factor (VF) is bigger than the number | |
5146 | of elements that we can fit in a vectype (nunits), we have to generate | |
5147 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 5148 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
5149 | vect_get_vec_def_for_copy_stmt. */ |
5150 | ||
0d0293ac | 5151 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
5152 | |
5153 | S1: &base + 2 = x2 | |
5154 | S2: &base = x0 | |
5155 | S3: &base + 1 = x1 | |
5156 | S4: &base + 3 = x3 | |
5157 | ||
5158 | We create vectorized stores starting from base address (the access of the | |
5159 | first stmt in the chain (S2 in the above example), when the last store stmt | |
5160 | of the chain (S4) is reached: | |
5161 | ||
5162 | VS1: &base = vx2 | |
5163 | VS2: &base + vec_size*1 = vx0 | |
5164 | VS3: &base + vec_size*2 = vx1 | |
5165 | VS4: &base + vec_size*3 = vx3 | |
5166 | ||
5167 | Then permutation statements are generated: | |
5168 | ||
3fcc1b55 JJ |
5169 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
5170 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 5171 | ... |
b8698a0f | 5172 | |
ebfd146a IR |
5173 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
5174 | (the order of the data-refs in the output of vect_permute_store_chain | |
5175 | corresponds to the order of scalar stmts in the interleaving chain - see | |
5176 | the documentation of vect_permute_store_chain()). | |
5177 | ||
5178 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 5179 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 5180 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 5181 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
5182 | */ |
5183 | ||
5184 | prev_stmt_info = NULL; | |
5185 | for (j = 0; j < ncopies; j++) | |
5186 | { | |
5187 | gimple new_stmt; | |
ebfd146a IR |
5188 | |
5189 | if (j == 0) | |
5190 | { | |
5191 | if (slp) | |
5192 | { | |
5193 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c IR |
5194 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
5195 | NULL, slp_node, -1); | |
ebfd146a | 5196 | |
9771b263 | 5197 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
5198 | } |
5199 | else | |
5200 | { | |
b8698a0f L |
5201 | /* For interleaved stores we collect vectorized defs for all the |
5202 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
5203 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
5204 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
5205 | ||
0d0293ac | 5206 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 5207 | OPRNDS are of size 1. */ |
b8698a0f | 5208 | next_stmt = first_stmt; |
ebfd146a IR |
5209 | for (i = 0; i < group_size; i++) |
5210 | { | |
b8698a0f L |
5211 | /* Since gaps are not supported for interleaved stores, |
5212 | GROUP_SIZE is the exact number of stmts in the chain. | |
5213 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
5214 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
5215 | iteration of the loop will be executed. */ |
5216 | gcc_assert (next_stmt | |
5217 | && gimple_assign_single_p (next_stmt)); | |
5218 | op = gimple_assign_rhs1 (next_stmt); | |
5219 | ||
b8698a0f | 5220 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt, |
ebfd146a | 5221 | NULL); |
9771b263 DN |
5222 | dr_chain.quick_push (vec_oprnd); |
5223 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 5224 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5225 | } |
5226 | } | |
5227 | ||
5228 | /* We should have catched mismatched types earlier. */ | |
5229 | gcc_assert (useless_type_conversion_p (vectype, | |
5230 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
5231 | bool simd_lane_access_p |
5232 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
5233 | if (simd_lane_access_p | |
5234 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
5235 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
5236 | && integer_zerop (DR_OFFSET (first_dr)) | |
5237 | && integer_zerop (DR_INIT (first_dr)) | |
5238 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
5239 | get_alias_set (DR_REF (first_dr)))) | |
5240 | { | |
5241 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
5242 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
5243 | (DR_REF (first_dr)), 0); | |
8928eff3 | 5244 | inv_p = false; |
74bf76ed JJ |
5245 | } |
5246 | else | |
5247 | dataref_ptr | |
5248 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
5249 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 5250 | offset, &dummy, gsi, &ptr_incr, |
74bf76ed | 5251 | simd_lane_access_p, &inv_p); |
a70d6342 | 5252 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 5253 | } |
b8698a0f | 5254 | else |
ebfd146a | 5255 | { |
b8698a0f L |
5256 | /* For interleaved stores we created vectorized defs for all the |
5257 | defs stored in OPRNDS in the previous iteration (previous copy). | |
5258 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
5259 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
5260 | next copy. | |
0d0293ac | 5261 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
5262 | OPRNDS are of size 1. */ |
5263 | for (i = 0; i < group_size; i++) | |
5264 | { | |
9771b263 | 5265 | op = oprnds[i]; |
24ee1384 IR |
5266 | vect_is_simple_use (op, NULL, loop_vinfo, bb_vinfo, &def_stmt, |
5267 | &def, &dt); | |
b8698a0f | 5268 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
5269 | dr_chain[i] = vec_oprnd; |
5270 | oprnds[i] = vec_oprnd; | |
ebfd146a | 5271 | } |
74bf76ed JJ |
5272 | if (dataref_offset) |
5273 | dataref_offset | |
5274 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
5275 | TYPE_SIZE_UNIT (aggr_type)); | |
5276 | else | |
5277 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
5278 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
5279 | } |
5280 | ||
272c6793 | 5281 | if (store_lanes_p) |
ebfd146a | 5282 | { |
272c6793 | 5283 | tree vec_array; |
267d3070 | 5284 | |
272c6793 RS |
5285 | /* Combine all the vectors into an array. */ |
5286 | vec_array = create_vector_array (vectype, vec_num); | |
5287 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 5288 | { |
9771b263 | 5289 | vec_oprnd = dr_chain[i]; |
272c6793 | 5290 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 5291 | } |
b8698a0f | 5292 | |
272c6793 RS |
5293 | /* Emit: |
5294 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
5295 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
5296 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
5297 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 | 5298 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
5299 | } |
5300 | else | |
5301 | { | |
5302 | new_stmt = NULL; | |
0d0293ac | 5303 | if (grouped_store) |
272c6793 | 5304 | { |
b6b9227d JJ |
5305 | if (j == 0) |
5306 | result_chain.create (group_size); | |
272c6793 RS |
5307 | /* Permute. */ |
5308 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
5309 | &result_chain); | |
5310 | } | |
c2d7ab2a | 5311 | |
272c6793 RS |
5312 | next_stmt = first_stmt; |
5313 | for (i = 0; i < vec_num; i++) | |
5314 | { | |
644ffefd | 5315 | unsigned align, misalign; |
272c6793 RS |
5316 | |
5317 | if (i > 0) | |
5318 | /* Bump the vector pointer. */ | |
5319 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5320 | stmt, NULL_TREE); | |
5321 | ||
5322 | if (slp) | |
9771b263 | 5323 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
5324 | else if (grouped_store) |
5325 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 5326 | vect_permute_store_chain(). */ |
9771b263 | 5327 | vec_oprnd = result_chain[i]; |
272c6793 RS |
5328 | |
5329 | data_ref = build2 (MEM_REF, TREE_TYPE (vec_oprnd), dataref_ptr, | |
74bf76ed JJ |
5330 | dataref_offset |
5331 | ? dataref_offset | |
5332 | : build_int_cst (reference_alias_ptr_type | |
5333 | (DR_REF (first_dr)), 0)); | |
644ffefd | 5334 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 | 5335 | if (aligned_access_p (first_dr)) |
644ffefd | 5336 | misalign = 0; |
272c6793 RS |
5337 | else if (DR_MISALIGNMENT (first_dr) == -1) |
5338 | { | |
5339 | TREE_TYPE (data_ref) | |
5340 | = build_aligned_type (TREE_TYPE (data_ref), | |
5341 | TYPE_ALIGN (elem_type)); | |
644ffefd MJ |
5342 | align = TYPE_ALIGN_UNIT (elem_type); |
5343 | misalign = 0; | |
272c6793 RS |
5344 | } |
5345 | else | |
5346 | { | |
5347 | TREE_TYPE (data_ref) | |
5348 | = build_aligned_type (TREE_TYPE (data_ref), | |
5349 | TYPE_ALIGN (elem_type)); | |
644ffefd | 5350 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 5351 | } |
74bf76ed JJ |
5352 | if (dataref_offset == NULL_TREE) |
5353 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
5354 | misalign); | |
c2d7ab2a | 5355 | |
09dfa495 BM |
5356 | if (negative) |
5357 | { | |
5358 | tree perm_mask = perm_mask_for_reverse (vectype); | |
5359 | tree perm_dest | |
5360 | = vect_create_destination_var (gimple_assign_rhs1 (stmt), | |
5361 | vectype); | |
5362 | tree new_temp = make_ssa_name (perm_dest, NULL); | |
5363 | ||
5364 | /* Generate the permute statement. */ | |
5365 | gimple perm_stmt | |
5366 | = gimple_build_assign_with_ops (VEC_PERM_EXPR, new_temp, | |
5367 | vec_oprnd, vec_oprnd, | |
5368 | perm_mask); | |
5369 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); | |
5370 | ||
5371 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
5372 | vec_oprnd = new_temp; | |
5373 | } | |
5374 | ||
272c6793 RS |
5375 | /* Arguments are ready. Create the new vector stmt. */ |
5376 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
5377 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
5378 | |
5379 | if (slp) | |
5380 | continue; | |
5381 | ||
e14c1050 | 5382 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
5383 | if (!next_stmt) |
5384 | break; | |
5385 | } | |
ebfd146a | 5386 | } |
1da0876c RS |
5387 | if (!slp) |
5388 | { | |
5389 | if (j == 0) | |
5390 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5391 | else | |
5392 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5393 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5394 | } | |
ebfd146a IR |
5395 | } |
5396 | ||
9771b263 DN |
5397 | dr_chain.release (); |
5398 | oprnds.release (); | |
5399 | result_chain.release (); | |
5400 | vec_oprnds.release (); | |
ebfd146a IR |
5401 | |
5402 | return true; | |
5403 | } | |
5404 | ||
aec7ae7d JJ |
5405 | /* Given a vector type VECTYPE and permutation SEL returns |
5406 | the VECTOR_CST mask that implements the permutation of the | |
5407 | vector elements. If that is impossible to do, returns NULL. */ | |
a1e53f3f | 5408 | |
3fcc1b55 JJ |
5409 | tree |
5410 | vect_gen_perm_mask (tree vectype, unsigned char *sel) | |
a1e53f3f | 5411 | { |
d2a12ae7 | 5412 | tree mask_elt_type, mask_type, mask_vec, *mask_elts; |
2635892a | 5413 | int i, nunits; |
a1e53f3f | 5414 | |
22e4dee7 | 5415 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 RH |
5416 | |
5417 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) | |
a1e53f3f L |
5418 | return NULL; |
5419 | ||
96f9265a RG |
5420 | mask_elt_type = lang_hooks.types.type_for_mode |
5421 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
22e4dee7 | 5422 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 5423 | |
d2a12ae7 | 5424 | mask_elts = XALLOCAVEC (tree, nunits); |
aec7ae7d | 5425 | for (i = nunits - 1; i >= 0; i--) |
d2a12ae7 RG |
5426 | mask_elts[i] = build_int_cst (mask_elt_type, sel[i]); |
5427 | mask_vec = build_vector (mask_type, mask_elts); | |
a1e53f3f | 5428 | |
2635892a | 5429 | return mask_vec; |
a1e53f3f L |
5430 | } |
5431 | ||
aec7ae7d JJ |
5432 | /* Given a vector variable X and Y, that was generated for the scalar |
5433 | STMT, generate instructions to permute the vector elements of X and Y | |
5434 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
5435 | permuted vector variable. */ | |
a1e53f3f L |
5436 | |
5437 | static tree | |
aec7ae7d JJ |
5438 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple stmt, |
5439 | gimple_stmt_iterator *gsi) | |
a1e53f3f L |
5440 | { |
5441 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 5442 | tree perm_dest, data_ref; |
a1e53f3f L |
5443 | gimple perm_stmt; |
5444 | ||
a1e53f3f | 5445 | perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype); |
aec7ae7d | 5446 | data_ref = make_ssa_name (perm_dest, NULL); |
a1e53f3f L |
5447 | |
5448 | /* Generate the permute statement. */ | |
73804b12 RG |
5449 | perm_stmt = gimple_build_assign_with_ops (VEC_PERM_EXPR, data_ref, |
5450 | x, y, mask_vec); | |
a1e53f3f L |
5451 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
5452 | ||
5453 | return data_ref; | |
5454 | } | |
5455 | ||
ebfd146a IR |
5456 | /* vectorizable_load. |
5457 | ||
b8698a0f L |
5458 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
5459 | can be vectorized. | |
5460 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5461 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5462 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5463 | ||
5464 | static bool | |
5465 | vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
c716e67f | 5466 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
5467 | { |
5468 | tree scalar_dest; | |
5469 | tree vec_dest = NULL; | |
5470 | tree data_ref = NULL; | |
5471 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 5472 | stmt_vec_info prev_stmt_info; |
ebfd146a | 5473 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 5474 | struct loop *loop = NULL; |
ebfd146a | 5475 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 5476 | bool nested_in_vect_loop = false; |
c716e67f | 5477 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
ebfd146a | 5478 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
272c6793 | 5479 | tree elem_type; |
ebfd146a | 5480 | tree new_temp; |
947131ba | 5481 | enum machine_mode mode; |
ebfd146a IR |
5482 | gimple new_stmt = NULL; |
5483 | tree dummy; | |
5484 | enum dr_alignment_support alignment_support_scheme; | |
5485 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 5486 | tree dataref_offset = NULL_TREE; |
fef4d2b3 | 5487 | gimple ptr_incr = NULL; |
ebfd146a IR |
5488 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
5489 | int ncopies; | |
a64b9c26 | 5490 | int i, j, group_size, group_gap; |
ebfd146a IR |
5491 | tree msq = NULL_TREE, lsq; |
5492 | tree offset = NULL_TREE; | |
5493 | tree realignment_token = NULL_TREE; | |
5494 | gimple phi = NULL; | |
6e1aa848 | 5495 | vec<tree> dr_chain = vNULL; |
0d0293ac | 5496 | bool grouped_load = false; |
272c6793 | 5497 | bool load_lanes_p = false; |
ebfd146a | 5498 | gimple first_stmt; |
ebfd146a | 5499 | bool inv_p; |
319e6439 | 5500 | bool negative = false; |
ebfd146a IR |
5501 | bool compute_in_loop = false; |
5502 | struct loop *at_loop; | |
5503 | int vec_num; | |
5504 | bool slp = (slp_node != NULL); | |
5505 | bool slp_perm = false; | |
5506 | enum tree_code code; | |
a70d6342 IR |
5507 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
5508 | int vf; | |
272c6793 | 5509 | tree aggr_type; |
aec7ae7d JJ |
5510 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; |
5511 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
5512 | int gather_scale = 1; | |
5513 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
a70d6342 IR |
5514 | |
5515 | if (loop_vinfo) | |
5516 | { | |
5517 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
5518 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
5519 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
5520 | } | |
5521 | else | |
3533e503 | 5522 | vf = 1; |
ebfd146a IR |
5523 | |
5524 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 5525 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 5526 | case of SLP. */ |
437f4a00 | 5527 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
5528 | ncopies = 1; |
5529 | else | |
5530 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
5531 | ||
5532 | gcc_assert (ncopies >= 1); | |
5533 | ||
5534 | /* FORNOW. This restriction should be relaxed. */ | |
5535 | if (nested_in_vect_loop && ncopies > 1) | |
5536 | { | |
73fbfcad | 5537 | if (dump_enabled_p ()) |
78c60e3d | 5538 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5539 | "multiple types in nested loop.\n"); |
ebfd146a IR |
5540 | return false; |
5541 | } | |
5542 | ||
a70d6342 | 5543 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5544 | return false; |
5545 | ||
8644a673 | 5546 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
5547 | return false; |
5548 | ||
5549 | /* Is vectorizable load? */ | |
5550 | if (!is_gimple_assign (stmt)) | |
5551 | return false; | |
5552 | ||
5553 | scalar_dest = gimple_assign_lhs (stmt); | |
5554 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
5555 | return false; | |
5556 | ||
5557 | code = gimple_assign_rhs_code (stmt); | |
5558 | if (code != ARRAY_REF | |
38000232 | 5559 | && code != BIT_FIELD_REF |
ebfd146a | 5560 | && code != INDIRECT_REF |
e9dbe7bb IR |
5561 | && code != COMPONENT_REF |
5562 | && code != IMAGPART_EXPR | |
70f34814 | 5563 | && code != REALPART_EXPR |
42373e0b RG |
5564 | && code != MEM_REF |
5565 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
ebfd146a IR |
5566 | return false; |
5567 | ||
5568 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
5569 | return false; | |
5570 | ||
7b7b1813 | 5571 | elem_type = TREE_TYPE (vectype); |
947131ba | 5572 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
5573 | |
5574 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
5575 | (e.g. - data copies). */ | |
947131ba | 5576 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 5577 | { |
73fbfcad | 5578 | if (dump_enabled_p ()) |
78c60e3d | 5579 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5580 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
5581 | return false; |
5582 | } | |
5583 | ||
ebfd146a | 5584 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 5585 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 5586 | { |
0d0293ac | 5587 | grouped_load = true; |
ebfd146a | 5588 | /* FORNOW */ |
aec7ae7d | 5589 | gcc_assert (! nested_in_vect_loop && !STMT_VINFO_GATHER_P (stmt_info)); |
ebfd146a | 5590 | |
e14c1050 | 5591 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
5592 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
5593 | { | |
e14c1050 | 5594 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
5595 | if (vect_load_lanes_supported (vectype, group_size)) |
5596 | load_lanes_p = true; | |
0d0293ac | 5597 | else if (!vect_grouped_load_supported (vectype, group_size)) |
b602d918 RS |
5598 | return false; |
5599 | } | |
ebfd146a IR |
5600 | } |
5601 | ||
a1e53f3f | 5602 | |
aec7ae7d JJ |
5603 | if (STMT_VINFO_GATHER_P (stmt_info)) |
5604 | { | |
5605 | gimple def_stmt; | |
5606 | tree def; | |
5607 | gather_decl = vect_check_gather (stmt, loop_vinfo, &gather_base, | |
5608 | &gather_off, &gather_scale); | |
5609 | gcc_assert (gather_decl); | |
24ee1384 | 5610 | if (!vect_is_simple_use_1 (gather_off, NULL, loop_vinfo, bb_vinfo, |
aec7ae7d JJ |
5611 | &def_stmt, &def, &gather_dt, |
5612 | &gather_off_vectype)) | |
5613 | { | |
73fbfcad | 5614 | if (dump_enabled_p ()) |
78c60e3d | 5615 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5616 | "gather index use not simple.\n"); |
aec7ae7d JJ |
5617 | return false; |
5618 | } | |
5619 | } | |
7d75abc8 | 5620 | else if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
14ac6aa2 | 5621 | ; |
319e6439 RG |
5622 | else |
5623 | { | |
5624 | negative = tree_int_cst_compare (nested_in_vect_loop | |
5625 | ? STMT_VINFO_DR_STEP (stmt_info) | |
5626 | : DR_STEP (dr), | |
5627 | size_zero_node) < 0; | |
5628 | if (negative && ncopies > 1) | |
5629 | { | |
73fbfcad | 5630 | if (dump_enabled_p ()) |
78c60e3d | 5631 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5632 | "multiple types with negative step.\n"); |
319e6439 RG |
5633 | return false; |
5634 | } | |
5635 | ||
5636 | if (negative) | |
5637 | { | |
08940f33 RB |
5638 | if (grouped_load) |
5639 | { | |
5640 | if (dump_enabled_p ()) | |
5641 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 TJ |
5642 | "negative step for group load not supported" |
5643 | "\n"); | |
08940f33 RB |
5644 | return false; |
5645 | } | |
319e6439 RG |
5646 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
5647 | if (alignment_support_scheme != dr_aligned | |
5648 | && alignment_support_scheme != dr_unaligned_supported) | |
5649 | { | |
73fbfcad | 5650 | if (dump_enabled_p ()) |
78c60e3d | 5651 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5652 | "negative step but alignment required.\n"); |
319e6439 RG |
5653 | return false; |
5654 | } | |
5655 | if (!perm_mask_for_reverse (vectype)) | |
5656 | { | |
73fbfcad | 5657 | if (dump_enabled_p ()) |
78c60e3d | 5658 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
5659 | "negative step and reversing not supported." |
5660 | "\n"); | |
319e6439 RG |
5661 | return false; |
5662 | } | |
5663 | } | |
7d75abc8 | 5664 | } |
aec7ae7d | 5665 | |
ebfd146a IR |
5666 | if (!vec_stmt) /* transformation not required. */ |
5667 | { | |
5668 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
92345349 | 5669 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, NULL, NULL, NULL); |
ebfd146a IR |
5670 | return true; |
5671 | } | |
5672 | ||
73fbfcad | 5673 | if (dump_enabled_p ()) |
78c60e3d | 5674 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5675 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a IR |
5676 | |
5677 | /** Transform. **/ | |
5678 | ||
c716e67f XDL |
5679 | ensure_base_align (stmt_info, dr); |
5680 | ||
aec7ae7d JJ |
5681 | if (STMT_VINFO_GATHER_P (stmt_info)) |
5682 | { | |
5683 | tree vec_oprnd0 = NULL_TREE, op; | |
5684 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
5685 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
5686 | tree ptr, mask, var, scale, perm_mask = NULL_TREE, prev_res = NULL_TREE; | |
5687 | edge pe = loop_preheader_edge (loop); | |
5688 | gimple_seq seq; | |
5689 | basic_block new_bb; | |
5690 | enum { NARROW, NONE, WIDEN } modifier; | |
5691 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
5692 | ||
5693 | if (nunits == gather_off_nunits) | |
5694 | modifier = NONE; | |
5695 | else if (nunits == gather_off_nunits / 2) | |
5696 | { | |
5697 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
5698 | modifier = WIDEN; | |
5699 | ||
5700 | for (i = 0; i < gather_off_nunits; ++i) | |
5701 | sel[i] = i | nunits; | |
5702 | ||
3fcc1b55 | 5703 | perm_mask = vect_gen_perm_mask (gather_off_vectype, sel); |
aec7ae7d JJ |
5704 | gcc_assert (perm_mask != NULL_TREE); |
5705 | } | |
5706 | else if (nunits == gather_off_nunits * 2) | |
5707 | { | |
5708 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
5709 | modifier = NARROW; | |
5710 | ||
5711 | for (i = 0; i < nunits; ++i) | |
5712 | sel[i] = i < gather_off_nunits | |
5713 | ? i : i + nunits - gather_off_nunits; | |
5714 | ||
3fcc1b55 | 5715 | perm_mask = vect_gen_perm_mask (vectype, sel); |
aec7ae7d JJ |
5716 | gcc_assert (perm_mask != NULL_TREE); |
5717 | ncopies *= 2; | |
5718 | } | |
5719 | else | |
5720 | gcc_unreachable (); | |
5721 | ||
5722 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
5723 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5724 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5725 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5726 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5727 | scaletype = TREE_VALUE (arglist); | |
5728 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
5729 | && types_compatible_p (srctype, masktype)); | |
5730 | ||
5731 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5732 | ||
5733 | ptr = fold_convert (ptrtype, gather_base); | |
5734 | if (!is_gimple_min_invariant (ptr)) | |
5735 | { | |
5736 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
5737 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
5738 | gcc_assert (!new_bb); | |
5739 | } | |
5740 | ||
5741 | /* Currently we support only unconditional gather loads, | |
5742 | so mask should be all ones. */ | |
5743 | if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
5744 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
5745 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) | |
5746 | { | |
5747 | REAL_VALUE_TYPE r; | |
5748 | long tmp[6]; | |
5749 | for (j = 0; j < 6; ++j) | |
5750 | tmp[j] = -1; | |
5751 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
5752 | mask = build_real (TREE_TYPE (masktype), r); | |
5753 | } | |
5754 | else | |
5755 | gcc_unreachable (); | |
5756 | mask = build_vector_from_val (masktype, mask); | |
5757 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
5758 | ||
5759 | scale = build_int_cst (scaletype, gather_scale); | |
5760 | ||
5761 | prev_stmt_info = NULL; | |
5762 | for (j = 0; j < ncopies; ++j) | |
5763 | { | |
5764 | if (modifier == WIDEN && (j & 1)) | |
5765 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
5766 | perm_mask, stmt, gsi); | |
5767 | else if (j == 0) | |
5768 | op = vec_oprnd0 | |
5769 | = vect_get_vec_def_for_operand (gather_off, stmt, NULL); | |
5770 | else | |
5771 | op = vec_oprnd0 | |
5772 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
5773 | ||
5774 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
5775 | { | |
5776 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
5777 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
5778 | var = vect_get_new_vect_var (idxtype, vect_simple_var, NULL); | |
aec7ae7d JJ |
5779 | var = make_ssa_name (var, NULL); |
5780 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); | |
5781 | new_stmt | |
5782 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, | |
5783 | op, NULL_TREE); | |
5784 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5785 | op = var; | |
5786 | } | |
5787 | ||
5788 | new_stmt | |
5789 | = gimple_build_call (gather_decl, 5, mask, ptr, op, mask, scale); | |
5790 | ||
5791 | if (!useless_type_conversion_p (vectype, rettype)) | |
5792 | { | |
5793 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
5794 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
5795 | var = vect_get_new_vect_var (rettype, vect_simple_var, NULL); | |
aec7ae7d JJ |
5796 | op = make_ssa_name (var, new_stmt); |
5797 | gimple_call_set_lhs (new_stmt, op); | |
5798 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5799 | var = make_ssa_name (vec_dest, NULL); | |
5800 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); | |
5801 | new_stmt | |
5802 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, op, | |
5803 | NULL_TREE); | |
5804 | } | |
5805 | else | |
5806 | { | |
5807 | var = make_ssa_name (vec_dest, new_stmt); | |
5808 | gimple_call_set_lhs (new_stmt, var); | |
5809 | } | |
5810 | ||
5811 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5812 | ||
5813 | if (modifier == NARROW) | |
5814 | { | |
5815 | if ((j & 1) == 0) | |
5816 | { | |
5817 | prev_res = var; | |
5818 | continue; | |
5819 | } | |
5820 | var = permute_vec_elements (prev_res, var, | |
5821 | perm_mask, stmt, gsi); | |
5822 | new_stmt = SSA_NAME_DEF_STMT (var); | |
5823 | } | |
5824 | ||
5825 | if (prev_stmt_info == NULL) | |
5826 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5827 | else | |
5828 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5829 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5830 | } | |
5831 | return true; | |
5832 | } | |
7d75abc8 MM |
5833 | else if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
5834 | { | |
5835 | gimple_stmt_iterator incr_gsi; | |
5836 | bool insert_after; | |
5837 | gimple incr; | |
5838 | tree offvar; | |
7d75abc8 MM |
5839 | tree ivstep; |
5840 | tree running_off; | |
9771b263 | 5841 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 5842 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
5843 | tree stride_base, stride_step, alias_off; |
5844 | ||
5845 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 5846 | |
14ac6aa2 RB |
5847 | stride_base |
5848 | = fold_build_pointer_plus | |
5849 | (unshare_expr (DR_BASE_ADDRESS (dr)), | |
5850 | size_binop (PLUS_EXPR, | |
5851 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (dr))), | |
c3284718 | 5852 | convert_to_ptrofftype (DR_INIT (dr)))); |
14ac6aa2 | 5853 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (dr))); |
7d75abc8 MM |
5854 | |
5855 | /* For a load with loop-invariant (but other than power-of-2) | |
5856 | stride (i.e. not a grouped access) like so: | |
5857 | ||
5858 | for (i = 0; i < n; i += stride) | |
5859 | ... = array[i]; | |
5860 | ||
5861 | we generate a new induction variable and new accesses to | |
5862 | form a new vector (or vectors, depending on ncopies): | |
5863 | ||
5864 | for (j = 0; ; j += VF*stride) | |
5865 | tmp1 = array[j]; | |
5866 | tmp2 = array[j + stride]; | |
5867 | ... | |
5868 | vectemp = {tmp1, tmp2, ...} | |
5869 | */ | |
5870 | ||
5871 | ivstep = stride_step; | |
5872 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
5873 | build_int_cst (TREE_TYPE (ivstep), vf)); | |
5874 | ||
5875 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
5876 | ||
5877 | create_iv (stride_base, ivstep, NULL, | |
5878 | loop, &incr_gsi, insert_after, | |
5879 | &offvar, NULL); | |
5880 | incr = gsi_stmt (incr_gsi); | |
5881 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo, NULL)); | |
5882 | ||
5883 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
5884 | if (stmts) | |
5885 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
5886 | ||
5887 | prev_stmt_info = NULL; | |
5888 | running_off = offvar; | |
14ac6aa2 | 5889 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (dr)), 0); |
7d75abc8 MM |
5890 | for (j = 0; j < ncopies; j++) |
5891 | { | |
5892 | tree vec_inv; | |
5893 | ||
9771b263 | 5894 | vec_alloc (v, nunits); |
7d75abc8 MM |
5895 | for (i = 0; i < nunits; i++) |
5896 | { | |
5897 | tree newref, newoff; | |
5898 | gimple incr; | |
14ac6aa2 RB |
5899 | newref = build2 (MEM_REF, TREE_TYPE (vectype), |
5900 | running_off, alias_off); | |
7d75abc8 MM |
5901 | |
5902 | newref = force_gimple_operand_gsi (gsi, newref, true, | |
5903 | NULL_TREE, true, | |
5904 | GSI_SAME_STMT); | |
5905 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, newref); | |
070ecdfd | 5906 | newoff = copy_ssa_name (running_off, NULL); |
14ac6aa2 RB |
5907 | incr = gimple_build_assign_with_ops (POINTER_PLUS_EXPR, newoff, |
5908 | running_off, stride_step); | |
7d75abc8 MM |
5909 | vect_finish_stmt_generation (stmt, incr, gsi); |
5910 | ||
5911 | running_off = newoff; | |
5912 | } | |
5913 | ||
5914 | vec_inv = build_constructor (vectype, v); | |
5915 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
5916 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
7d75abc8 MM |
5917 | |
5918 | if (j == 0) | |
5919 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5920 | else | |
5921 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5922 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5923 | } | |
5924 | return true; | |
5925 | } | |
aec7ae7d | 5926 | |
0d0293ac | 5927 | if (grouped_load) |
ebfd146a | 5928 | { |
e14c1050 | 5929 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
6aa904c4 | 5930 | if (slp |
01d8bf07 | 5931 | && !SLP_TREE_LOAD_PERMUTATION (slp_node).exists () |
9771b263 DN |
5932 | && first_stmt != SLP_TREE_SCALAR_STMTS (slp_node)[0]) |
5933 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 5934 | |
ebfd146a | 5935 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
5936 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
5937 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
5938 | ??? But we can only do so if there is exactly one | |
5939 | as we have no way to get at the rest. Leave the CSE | |
5940 | opportunity alone. | |
5941 | ??? With the group load eventually participating | |
5942 | in multiple different permutations (having multiple | |
5943 | slp nodes which refer to the same group) the CSE | |
5944 | is even wrong code. See PR56270. */ | |
5945 | && !slp) | |
ebfd146a IR |
5946 | { |
5947 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
5948 | return true; | |
5949 | } | |
5950 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 5951 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a IR |
5952 | |
5953 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
5954 | if (slp) | |
5955 | { | |
0d0293ac | 5956 | grouped_load = false; |
ebfd146a | 5957 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
01d8bf07 | 5958 | if (SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
a70d6342 | 5959 | slp_perm = true; |
a64b9c26 | 5960 | group_gap = GROUP_GAP (vinfo_for_stmt (first_stmt)); |
a70d6342 | 5961 | } |
ebfd146a | 5962 | else |
a64b9c26 RB |
5963 | { |
5964 | vec_num = group_size; | |
5965 | group_gap = 0; | |
5966 | } | |
ebfd146a IR |
5967 | } |
5968 | else | |
5969 | { | |
5970 | first_stmt = stmt; | |
5971 | first_dr = dr; | |
5972 | group_size = vec_num = 1; | |
a64b9c26 | 5973 | group_gap = 0; |
ebfd146a IR |
5974 | } |
5975 | ||
720f5239 | 5976 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 5977 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
5978 | /* Targets with load-lane instructions must not require explicit |
5979 | realignment. */ | |
5980 | gcc_assert (!load_lanes_p | |
5981 | || alignment_support_scheme == dr_aligned | |
5982 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
5983 | |
5984 | /* In case the vectorization factor (VF) is bigger than the number | |
5985 | of elements that we can fit in a vectype (nunits), we have to generate | |
5986 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 5987 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 5988 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 5989 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 5990 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
5991 | stmts that use the defs of the current stmt. The example below |
5992 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
5993 | need to create 4 vectorized stmts): | |
ebfd146a IR |
5994 | |
5995 | before vectorization: | |
5996 | RELATED_STMT VEC_STMT | |
5997 | S1: x = memref - - | |
5998 | S2: z = x + 1 - - | |
5999 | ||
6000 | step 1: vectorize stmt S1: | |
6001 | We first create the vector stmt VS1_0, and, as usual, record a | |
6002 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
6003 | Next, we create the vector stmt VS1_1, and record a pointer to | |
6004 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 6005 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
6006 | stmts and pointers: |
6007 | RELATED_STMT VEC_STMT | |
6008 | VS1_0: vx0 = memref0 VS1_1 - | |
6009 | VS1_1: vx1 = memref1 VS1_2 - | |
6010 | VS1_2: vx2 = memref2 VS1_3 - | |
6011 | VS1_3: vx3 = memref3 - - | |
6012 | S1: x = load - VS1_0 | |
6013 | S2: z = x + 1 - - | |
6014 | ||
b8698a0f L |
6015 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
6016 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
6017 | stmt S2. */ |
6018 | ||
0d0293ac | 6019 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6020 | |
6021 | S1: x2 = &base + 2 | |
6022 | S2: x0 = &base | |
6023 | S3: x1 = &base + 1 | |
6024 | S4: x3 = &base + 3 | |
6025 | ||
b8698a0f | 6026 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
6027 | starting from the access of the first stmt of the chain: |
6028 | ||
6029 | VS1: vx0 = &base | |
6030 | VS2: vx1 = &base + vec_size*1 | |
6031 | VS3: vx3 = &base + vec_size*2 | |
6032 | VS4: vx4 = &base + vec_size*3 | |
6033 | ||
6034 | Then permutation statements are generated: | |
6035 | ||
e2c83630 RH |
6036 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
6037 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
6038 | ... |
6039 | ||
6040 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
6041 | (the order of the data-refs in the output of vect_permute_load_chain | |
6042 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6043 | the documentation of vect_permute_load_chain()). | |
6044 | The generation of permutation stmts and recording them in | |
0d0293ac | 6045 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 6046 | |
b8698a0f | 6047 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
6048 | permutation stmts above are created for every copy. The result vector |
6049 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
6050 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
6051 | |
6052 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
6053 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
6054 | we generate the following code: | |
6055 | p = initial_addr; | |
6056 | indx = 0; | |
6057 | loop { | |
6058 | p = p + indx * vectype_size; | |
6059 | vec_dest = *(p); | |
6060 | indx = indx + 1; | |
6061 | } | |
6062 | ||
6063 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 6064 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
6065 | then generate the following code, in which the data in each iteration is |
6066 | obtained by two vector loads, one from the previous iteration, and one | |
6067 | from the current iteration: | |
6068 | p1 = initial_addr; | |
6069 | msq_init = *(floor(p1)) | |
6070 | p2 = initial_addr + VS - 1; | |
6071 | realignment_token = call target_builtin; | |
6072 | indx = 0; | |
6073 | loop { | |
6074 | p2 = p2 + indx * vectype_size | |
6075 | lsq = *(floor(p2)) | |
6076 | vec_dest = realign_load (msq, lsq, realignment_token) | |
6077 | indx = indx + 1; | |
6078 | msq = lsq; | |
6079 | } */ | |
6080 | ||
6081 | /* If the misalignment remains the same throughout the execution of the | |
6082 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 6083 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
6084 | This can only occur when vectorizing memory accesses in the inner-loop |
6085 | nested within an outer-loop that is being vectorized. */ | |
6086 | ||
d1e4b493 | 6087 | if (nested_in_vect_loop |
211bea38 | 6088 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
ebfd146a IR |
6089 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) |
6090 | { | |
6091 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
6092 | compute_in_loop = true; | |
6093 | } | |
6094 | ||
6095 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
6096 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 6097 | && !compute_in_loop) |
ebfd146a IR |
6098 | { |
6099 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
6100 | alignment_support_scheme, NULL_TREE, | |
6101 | &at_loop); | |
6102 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
6103 | { | |
6104 | phi = SSA_NAME_DEF_STMT (msq); | |
6105 | offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
6106 | } | |
6107 | } | |
6108 | else | |
6109 | at_loop = loop; | |
6110 | ||
a1e53f3f L |
6111 | if (negative) |
6112 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
6113 | ||
272c6793 RS |
6114 | if (load_lanes_p) |
6115 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
6116 | else | |
6117 | aggr_type = vectype; | |
6118 | ||
ebfd146a IR |
6119 | prev_stmt_info = NULL; |
6120 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 6121 | { |
272c6793 | 6122 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 6123 | if (j == 0) |
74bf76ed JJ |
6124 | { |
6125 | bool simd_lane_access_p | |
6126 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
6127 | if (simd_lane_access_p | |
6128 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
6129 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
6130 | && integer_zerop (DR_OFFSET (first_dr)) | |
6131 | && integer_zerop (DR_INIT (first_dr)) | |
6132 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
6133 | get_alias_set (DR_REF (first_dr))) | |
6134 | && (alignment_support_scheme == dr_aligned | |
6135 | || alignment_support_scheme == dr_unaligned_supported)) | |
6136 | { | |
6137 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
6138 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
6139 | (DR_REF (first_dr)), 0); | |
8928eff3 | 6140 | inv_p = false; |
74bf76ed JJ |
6141 | } |
6142 | else | |
6143 | dataref_ptr | |
6144 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
6145 | offset, &dummy, gsi, &ptr_incr, | |
6146 | simd_lane_access_p, &inv_p); | |
6147 | } | |
6148 | else if (dataref_offset) | |
6149 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
6150 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 6151 | else |
272c6793 RS |
6152 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
6153 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 6154 | |
0d0293ac | 6155 | if (grouped_load || slp_perm) |
9771b263 | 6156 | dr_chain.create (vec_num); |
5ce1ee7f | 6157 | |
272c6793 | 6158 | if (load_lanes_p) |
ebfd146a | 6159 | { |
272c6793 RS |
6160 | tree vec_array; |
6161 | ||
6162 | vec_array = create_vector_array (vectype, vec_num); | |
6163 | ||
6164 | /* Emit: | |
6165 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
6166 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
6167 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
6168 | gimple_call_set_lhs (new_stmt, vec_array); | |
6169 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 6170 | |
272c6793 RS |
6171 | /* Extract each vector into an SSA_NAME. */ |
6172 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 6173 | { |
272c6793 RS |
6174 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
6175 | vec_array, i); | |
9771b263 | 6176 | dr_chain.quick_push (new_temp); |
272c6793 RS |
6177 | } |
6178 | ||
6179 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 6180 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
6181 | } |
6182 | else | |
6183 | { | |
6184 | for (i = 0; i < vec_num; i++) | |
6185 | { | |
6186 | if (i > 0) | |
6187 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
6188 | stmt, NULL_TREE); | |
6189 | ||
6190 | /* 2. Create the vector-load in the loop. */ | |
6191 | switch (alignment_support_scheme) | |
6192 | { | |
6193 | case dr_aligned: | |
6194 | case dr_unaligned_supported: | |
be1ac4ec | 6195 | { |
644ffefd MJ |
6196 | unsigned int align, misalign; |
6197 | ||
272c6793 RS |
6198 | data_ref |
6199 | = build2 (MEM_REF, vectype, dataref_ptr, | |
74bf76ed JJ |
6200 | dataref_offset |
6201 | ? dataref_offset | |
6202 | : build_int_cst (reference_alias_ptr_type | |
6203 | (DR_REF (first_dr)), 0)); | |
644ffefd | 6204 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 RS |
6205 | if (alignment_support_scheme == dr_aligned) |
6206 | { | |
6207 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 6208 | misalign = 0; |
272c6793 RS |
6209 | } |
6210 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
6211 | { | |
6212 | TREE_TYPE (data_ref) | |
6213 | = build_aligned_type (TREE_TYPE (data_ref), | |
6214 | TYPE_ALIGN (elem_type)); | |
644ffefd MJ |
6215 | align = TYPE_ALIGN_UNIT (elem_type); |
6216 | misalign = 0; | |
272c6793 RS |
6217 | } |
6218 | else | |
6219 | { | |
6220 | TREE_TYPE (data_ref) | |
6221 | = build_aligned_type (TREE_TYPE (data_ref), | |
6222 | TYPE_ALIGN (elem_type)); | |
644ffefd | 6223 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 6224 | } |
74bf76ed JJ |
6225 | if (dataref_offset == NULL_TREE) |
6226 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), | |
6227 | align, misalign); | |
272c6793 | 6228 | break; |
be1ac4ec | 6229 | } |
272c6793 | 6230 | case dr_explicit_realign: |
267d3070 | 6231 | { |
272c6793 RS |
6232 | tree ptr, bump; |
6233 | tree vs_minus_1; | |
6234 | ||
6235 | vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
6236 | ||
6237 | if (compute_in_loop) | |
6238 | msq = vect_setup_realignment (first_stmt, gsi, | |
6239 | &realignment_token, | |
6240 | dr_explicit_realign, | |
6241 | dataref_ptr, NULL); | |
6242 | ||
070ecdfd | 6243 | ptr = copy_ssa_name (dataref_ptr, NULL); |
272c6793 | 6244 | new_stmt = gimple_build_assign_with_ops |
070ecdfd | 6245 | (BIT_AND_EXPR, ptr, dataref_ptr, |
272c6793 RS |
6246 | build_int_cst |
6247 | (TREE_TYPE (dataref_ptr), | |
6248 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
6249 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6250 | data_ref | |
6251 | = build2 (MEM_REF, vectype, ptr, | |
6252 | build_int_cst (reference_alias_ptr_type | |
6253 | (DR_REF (first_dr)), 0)); | |
6254 | vec_dest = vect_create_destination_var (scalar_dest, | |
6255 | vectype); | |
6256 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
6257 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
6258 | gimple_assign_set_lhs (new_stmt, new_temp); | |
6259 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
6260 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
6261 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6262 | msq = new_temp; | |
6263 | ||
6264 | bump = size_binop (MULT_EXPR, vs_minus_1, | |
7b7b1813 | 6265 | TYPE_SIZE_UNIT (elem_type)); |
272c6793 RS |
6266 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
6267 | new_stmt = gimple_build_assign_with_ops | |
6268 | (BIT_AND_EXPR, NULL_TREE, ptr, | |
6269 | build_int_cst | |
6270 | (TREE_TYPE (ptr), | |
6271 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
070ecdfd | 6272 | ptr = copy_ssa_name (dataref_ptr, new_stmt); |
272c6793 RS |
6273 | gimple_assign_set_lhs (new_stmt, ptr); |
6274 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6275 | data_ref | |
6276 | = build2 (MEM_REF, vectype, ptr, | |
6277 | build_int_cst (reference_alias_ptr_type | |
6278 | (DR_REF (first_dr)), 0)); | |
6279 | break; | |
267d3070 | 6280 | } |
272c6793 | 6281 | case dr_explicit_realign_optimized: |
070ecdfd | 6282 | new_temp = copy_ssa_name (dataref_ptr, NULL); |
272c6793 | 6283 | new_stmt = gimple_build_assign_with_ops |
070ecdfd | 6284 | (BIT_AND_EXPR, new_temp, dataref_ptr, |
272c6793 RS |
6285 | build_int_cst |
6286 | (TREE_TYPE (dataref_ptr), | |
6287 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
6288 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6289 | data_ref | |
6290 | = build2 (MEM_REF, vectype, new_temp, | |
6291 | build_int_cst (reference_alias_ptr_type | |
6292 | (DR_REF (first_dr)), 0)); | |
6293 | break; | |
6294 | default: | |
6295 | gcc_unreachable (); | |
6296 | } | |
ebfd146a | 6297 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 6298 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
6299 | new_temp = make_ssa_name (vec_dest, new_stmt); |
6300 | gimple_assign_set_lhs (new_stmt, new_temp); | |
6301 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6302 | ||
272c6793 RS |
6303 | /* 3. Handle explicit realignment if necessary/supported. |
6304 | Create in loop: | |
6305 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
6306 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
6307 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 6308 | { |
272c6793 RS |
6309 | lsq = gimple_assign_lhs (new_stmt); |
6310 | if (!realignment_token) | |
6311 | realignment_token = dataref_ptr; | |
6312 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
6313 | new_stmt | |
73804b12 RG |
6314 | = gimple_build_assign_with_ops (REALIGN_LOAD_EXPR, |
6315 | vec_dest, msq, lsq, | |
6316 | realignment_token); | |
272c6793 RS |
6317 | new_temp = make_ssa_name (vec_dest, new_stmt); |
6318 | gimple_assign_set_lhs (new_stmt, new_temp); | |
6319 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6320 | ||
6321 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
6322 | { | |
6323 | gcc_assert (phi); | |
6324 | if (i == vec_num - 1 && j == ncopies - 1) | |
6325 | add_phi_arg (phi, lsq, | |
6326 | loop_latch_edge (containing_loop), | |
9e227d60 | 6327 | UNKNOWN_LOCATION); |
272c6793 RS |
6328 | msq = lsq; |
6329 | } | |
ebfd146a | 6330 | } |
ebfd146a | 6331 | |
59fd17e3 RB |
6332 | /* 4. Handle invariant-load. */ |
6333 | if (inv_p && !bb_vinfo) | |
6334 | { | |
6335 | gimple_stmt_iterator gsi2 = *gsi; | |
6336 | gcc_assert (!grouped_load); | |
6337 | gsi_next (&gsi2); | |
6338 | new_temp = vect_init_vector (stmt, scalar_dest, | |
6339 | vectype, &gsi2); | |
6340 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
6341 | } | |
6342 | ||
272c6793 RS |
6343 | if (negative) |
6344 | { | |
aec7ae7d JJ |
6345 | tree perm_mask = perm_mask_for_reverse (vectype); |
6346 | new_temp = permute_vec_elements (new_temp, new_temp, | |
6347 | perm_mask, stmt, gsi); | |
ebfd146a IR |
6348 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
6349 | } | |
267d3070 | 6350 | |
272c6793 | 6351 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
6352 | vect_transform_grouped_load (). */ |
6353 | if (grouped_load || slp_perm) | |
9771b263 | 6354 | dr_chain.quick_push (new_temp); |
267d3070 | 6355 | |
272c6793 RS |
6356 | /* Store vector loads in the corresponding SLP_NODE. */ |
6357 | if (slp && !slp_perm) | |
9771b263 | 6358 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
272c6793 | 6359 | } |
a64b9c26 RB |
6360 | /* Bump the vector pointer to account for a gap. */ |
6361 | if (slp && group_gap != 0) | |
6362 | { | |
6363 | tree bump = size_binop (MULT_EXPR, | |
6364 | TYPE_SIZE_UNIT (elem_type), | |
6365 | size_int (group_gap)); | |
6366 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
6367 | stmt, bump); | |
6368 | } | |
ebfd146a IR |
6369 | } |
6370 | ||
6371 | if (slp && !slp_perm) | |
6372 | continue; | |
6373 | ||
6374 | if (slp_perm) | |
6375 | { | |
01d8bf07 | 6376 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
ebfd146a IR |
6377 | slp_node_instance, false)) |
6378 | { | |
9771b263 | 6379 | dr_chain.release (); |
ebfd146a IR |
6380 | return false; |
6381 | } | |
6382 | } | |
6383 | else | |
6384 | { | |
0d0293ac | 6385 | if (grouped_load) |
ebfd146a | 6386 | { |
272c6793 | 6387 | if (!load_lanes_p) |
0d0293ac | 6388 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 6389 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
6390 | } |
6391 | else | |
6392 | { | |
6393 | if (j == 0) | |
6394 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6395 | else | |
6396 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6397 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6398 | } | |
6399 | } | |
9771b263 | 6400 | dr_chain.release (); |
ebfd146a IR |
6401 | } |
6402 | ||
ebfd146a IR |
6403 | return true; |
6404 | } | |
6405 | ||
6406 | /* Function vect_is_simple_cond. | |
b8698a0f | 6407 | |
ebfd146a IR |
6408 | Input: |
6409 | LOOP - the loop that is being vectorized. | |
6410 | COND - Condition that is checked for simple use. | |
6411 | ||
e9e1d143 RG |
6412 | Output: |
6413 | *COMP_VECTYPE - the vector type for the comparison. | |
6414 | ||
ebfd146a IR |
6415 | Returns whether a COND can be vectorized. Checks whether |
6416 | condition operands are supportable using vec_is_simple_use. */ | |
6417 | ||
87aab9b2 | 6418 | static bool |
24ee1384 IR |
6419 | vect_is_simple_cond (tree cond, gimple stmt, loop_vec_info loop_vinfo, |
6420 | bb_vec_info bb_vinfo, tree *comp_vectype) | |
ebfd146a IR |
6421 | { |
6422 | tree lhs, rhs; | |
6423 | tree def; | |
6424 | enum vect_def_type dt; | |
e9e1d143 | 6425 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a IR |
6426 | |
6427 | if (!COMPARISON_CLASS_P (cond)) | |
6428 | return false; | |
6429 | ||
6430 | lhs = TREE_OPERAND (cond, 0); | |
6431 | rhs = TREE_OPERAND (cond, 1); | |
6432 | ||
6433 | if (TREE_CODE (lhs) == SSA_NAME) | |
6434 | { | |
6435 | gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); | |
24ee1384 IR |
6436 | if (!vect_is_simple_use_1 (lhs, stmt, loop_vinfo, bb_vinfo, |
6437 | &lhs_def_stmt, &def, &dt, &vectype1)) | |
ebfd146a IR |
6438 | return false; |
6439 | } | |
6440 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
6441 | && TREE_CODE (lhs) != FIXED_CST) | |
6442 | return false; | |
6443 | ||
6444 | if (TREE_CODE (rhs) == SSA_NAME) | |
6445 | { | |
6446 | gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); | |
24ee1384 IR |
6447 | if (!vect_is_simple_use_1 (rhs, stmt, loop_vinfo, bb_vinfo, |
6448 | &rhs_def_stmt, &def, &dt, &vectype2)) | |
ebfd146a IR |
6449 | return false; |
6450 | } | |
f7e531cf | 6451 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST |
ebfd146a IR |
6452 | && TREE_CODE (rhs) != FIXED_CST) |
6453 | return false; | |
6454 | ||
e9e1d143 | 6455 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
6456 | return true; |
6457 | } | |
6458 | ||
6459 | /* vectorizable_condition. | |
6460 | ||
b8698a0f L |
6461 | Check if STMT is conditional modify expression that can be vectorized. |
6462 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
6463 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
6464 | at GSI. |
6465 | ||
6466 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
6467 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
6468 | else caluse if it is 2). | |
ebfd146a IR |
6469 | |
6470 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6471 | ||
4bbe8262 | 6472 | bool |
ebfd146a | 6473 | vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi, |
f7e531cf IR |
6474 | gimple *vec_stmt, tree reduc_def, int reduc_index, |
6475 | slp_tree slp_node) | |
ebfd146a IR |
6476 | { |
6477 | tree scalar_dest = NULL_TREE; | |
6478 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
6479 | tree cond_expr, then_clause, else_clause; |
6480 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6481 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
df11cc78 | 6482 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
6483 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
6484 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
ebfd146a IR |
6485 | tree vec_compare, vec_cond_expr; |
6486 | tree new_temp; | |
6487 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
ebfd146a | 6488 | tree def; |
a855b1b1 | 6489 | enum vect_def_type dt, dts[4]; |
ebfd146a | 6490 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
f7e531cf | 6491 | int ncopies; |
ebfd146a | 6492 | enum tree_code code; |
a855b1b1 | 6493 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
6494 | int i, j; |
6495 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
6496 | vec<tree> vec_oprnds0 = vNULL; |
6497 | vec<tree> vec_oprnds1 = vNULL; | |
6498 | vec<tree> vec_oprnds2 = vNULL; | |
6499 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 6500 | tree vec_cmp_type; |
b8698a0f | 6501 | |
f7e531cf IR |
6502 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
6503 | ncopies = 1; | |
6504 | else | |
6505 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
437f4a00 | 6506 | |
ebfd146a | 6507 | gcc_assert (ncopies >= 1); |
a855b1b1 | 6508 | if (reduc_index && ncopies > 1) |
ebfd146a IR |
6509 | return false; /* FORNOW */ |
6510 | ||
f7e531cf IR |
6511 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
6512 | return false; | |
6513 | ||
6514 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
ebfd146a IR |
6515 | return false; |
6516 | ||
4bbe8262 IR |
6517 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6518 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
6519 | && reduc_def)) | |
ebfd146a IR |
6520 | return false; |
6521 | ||
ebfd146a | 6522 | /* FORNOW: not yet supported. */ |
b8698a0f | 6523 | if (STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 6524 | { |
73fbfcad | 6525 | if (dump_enabled_p ()) |
78c60e3d | 6526 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6527 | "value used after loop.\n"); |
ebfd146a IR |
6528 | return false; |
6529 | } | |
6530 | ||
6531 | /* Is vectorizable conditional operation? */ | |
6532 | if (!is_gimple_assign (stmt)) | |
6533 | return false; | |
6534 | ||
6535 | code = gimple_assign_rhs_code (stmt); | |
6536 | ||
6537 | if (code != COND_EXPR) | |
6538 | return false; | |
6539 | ||
4e71066d RG |
6540 | cond_expr = gimple_assign_rhs1 (stmt); |
6541 | then_clause = gimple_assign_rhs2 (stmt); | |
6542 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 6543 | |
24ee1384 IR |
6544 | if (!vect_is_simple_cond (cond_expr, stmt, loop_vinfo, bb_vinfo, |
6545 | &comp_vectype) | |
e9e1d143 | 6546 | || !comp_vectype) |
ebfd146a IR |
6547 | return false; |
6548 | ||
6549 | if (TREE_CODE (then_clause) == SSA_NAME) | |
6550 | { | |
6551 | gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause); | |
24ee1384 | 6552 | if (!vect_is_simple_use (then_clause, stmt, loop_vinfo, bb_vinfo, |
ebfd146a IR |
6553 | &then_def_stmt, &def, &dt)) |
6554 | return false; | |
6555 | } | |
b8698a0f | 6556 | else if (TREE_CODE (then_clause) != INTEGER_CST |
ebfd146a IR |
6557 | && TREE_CODE (then_clause) != REAL_CST |
6558 | && TREE_CODE (then_clause) != FIXED_CST) | |
6559 | return false; | |
6560 | ||
6561 | if (TREE_CODE (else_clause) == SSA_NAME) | |
6562 | { | |
6563 | gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause); | |
24ee1384 | 6564 | if (!vect_is_simple_use (else_clause, stmt, loop_vinfo, bb_vinfo, |
ebfd146a IR |
6565 | &else_def_stmt, &def, &dt)) |
6566 | return false; | |
6567 | } | |
b8698a0f | 6568 | else if (TREE_CODE (else_clause) != INTEGER_CST |
ebfd146a IR |
6569 | && TREE_CODE (else_clause) != REAL_CST |
6570 | && TREE_CODE (else_clause) != FIXED_CST) | |
6571 | return false; | |
6572 | ||
74946978 MP |
6573 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (vectype))); |
6574 | /* The result of a vector comparison should be signed type. */ | |
6575 | tree cmp_type = build_nonstandard_integer_type (prec, 0); | |
6576 | vec_cmp_type = get_same_sized_vectype (cmp_type, vectype); | |
6577 | if (vec_cmp_type == NULL_TREE) | |
6578 | return false; | |
784fb9b3 | 6579 | |
b8698a0f | 6580 | if (!vec_stmt) |
ebfd146a IR |
6581 | { |
6582 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
e9e1d143 | 6583 | return expand_vec_cond_expr_p (vectype, comp_vectype); |
ebfd146a IR |
6584 | } |
6585 | ||
f7e531cf IR |
6586 | /* Transform. */ |
6587 | ||
6588 | if (!slp_node) | |
6589 | { | |
9771b263 DN |
6590 | vec_oprnds0.create (1); |
6591 | vec_oprnds1.create (1); | |
6592 | vec_oprnds2.create (1); | |
6593 | vec_oprnds3.create (1); | |
f7e531cf | 6594 | } |
ebfd146a IR |
6595 | |
6596 | /* Handle def. */ | |
6597 | scalar_dest = gimple_assign_lhs (stmt); | |
6598 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
6599 | ||
6600 | /* Handle cond expr. */ | |
a855b1b1 MM |
6601 | for (j = 0; j < ncopies; j++) |
6602 | { | |
f7e531cf | 6603 | gimple new_stmt = NULL; |
a855b1b1 MM |
6604 | if (j == 0) |
6605 | { | |
f7e531cf IR |
6606 | if (slp_node) |
6607 | { | |
07687835 TS |
6608 | stack_vec<tree, 4> ops; |
6609 | stack_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 6610 | |
9771b263 DN |
6611 | ops.safe_push (TREE_OPERAND (cond_expr, 0)); |
6612 | ops.safe_push (TREE_OPERAND (cond_expr, 1)); | |
6613 | ops.safe_push (then_clause); | |
6614 | ops.safe_push (else_clause); | |
f7e531cf | 6615 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); |
37b5ec8f JJ |
6616 | vec_oprnds3 = vec_defs.pop (); |
6617 | vec_oprnds2 = vec_defs.pop (); | |
6618 | vec_oprnds1 = vec_defs.pop (); | |
6619 | vec_oprnds0 = vec_defs.pop (); | |
f7e531cf | 6620 | |
9771b263 DN |
6621 | ops.release (); |
6622 | vec_defs.release (); | |
f7e531cf IR |
6623 | } |
6624 | else | |
6625 | { | |
6626 | gimple gtemp; | |
6627 | vec_cond_lhs = | |
a855b1b1 MM |
6628 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), |
6629 | stmt, NULL); | |
24ee1384 IR |
6630 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), stmt, |
6631 | loop_vinfo, NULL, >emp, &def, &dts[0]); | |
f7e531cf IR |
6632 | |
6633 | vec_cond_rhs = | |
6634 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
6635 | stmt, NULL); | |
24ee1384 IR |
6636 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), stmt, |
6637 | loop_vinfo, NULL, >emp, &def, &dts[1]); | |
f7e531cf IR |
6638 | if (reduc_index == 1) |
6639 | vec_then_clause = reduc_def; | |
6640 | else | |
6641 | { | |
6642 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
6643 | stmt, NULL); | |
24ee1384 | 6644 | vect_is_simple_use (then_clause, stmt, loop_vinfo, |
f7e531cf IR |
6645 | NULL, >emp, &def, &dts[2]); |
6646 | } | |
6647 | if (reduc_index == 2) | |
6648 | vec_else_clause = reduc_def; | |
6649 | else | |
6650 | { | |
6651 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
a855b1b1 | 6652 | stmt, NULL); |
24ee1384 | 6653 | vect_is_simple_use (else_clause, stmt, loop_vinfo, |
a855b1b1 | 6654 | NULL, >emp, &def, &dts[3]); |
f7e531cf | 6655 | } |
a855b1b1 MM |
6656 | } |
6657 | } | |
6658 | else | |
6659 | { | |
f7e531cf | 6660 | vec_cond_lhs = vect_get_vec_def_for_stmt_copy (dts[0], |
9771b263 | 6661 | vec_oprnds0.pop ()); |
f7e531cf | 6662 | vec_cond_rhs = vect_get_vec_def_for_stmt_copy (dts[1], |
9771b263 | 6663 | vec_oprnds1.pop ()); |
a855b1b1 | 6664 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 6665 | vec_oprnds2.pop ()); |
a855b1b1 | 6666 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 6667 | vec_oprnds3.pop ()); |
f7e531cf IR |
6668 | } |
6669 | ||
6670 | if (!slp_node) | |
6671 | { | |
9771b263 DN |
6672 | vec_oprnds0.quick_push (vec_cond_lhs); |
6673 | vec_oprnds1.quick_push (vec_cond_rhs); | |
6674 | vec_oprnds2.quick_push (vec_then_clause); | |
6675 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
6676 | } |
6677 | ||
9dc3f7de | 6678 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 6679 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 6680 | { |
9771b263 DN |
6681 | vec_cond_rhs = vec_oprnds1[i]; |
6682 | vec_then_clause = vec_oprnds2[i]; | |
6683 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 6684 | |
784fb9b3 JJ |
6685 | vec_compare = build2 (TREE_CODE (cond_expr), vec_cmp_type, |
6686 | vec_cond_lhs, vec_cond_rhs); | |
f7e531cf IR |
6687 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, |
6688 | vec_compare, vec_then_clause, vec_else_clause); | |
a855b1b1 | 6689 | |
f7e531cf IR |
6690 | new_stmt = gimple_build_assign (vec_dest, vec_cond_expr); |
6691 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
6692 | gimple_assign_set_lhs (new_stmt, new_temp); | |
6693 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6694 | if (slp_node) | |
9771b263 | 6695 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
6696 | } |
6697 | ||
6698 | if (slp_node) | |
6699 | continue; | |
6700 | ||
6701 | if (j == 0) | |
6702 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6703 | else | |
6704 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6705 | ||
6706 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 6707 | } |
b8698a0f | 6708 | |
9771b263 DN |
6709 | vec_oprnds0.release (); |
6710 | vec_oprnds1.release (); | |
6711 | vec_oprnds2.release (); | |
6712 | vec_oprnds3.release (); | |
f7e531cf | 6713 | |
ebfd146a IR |
6714 | return true; |
6715 | } | |
6716 | ||
6717 | ||
8644a673 | 6718 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
6719 | |
6720 | bool | |
a70d6342 | 6721 | vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 6722 | { |
8644a673 | 6723 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 6724 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 6725 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 6726 | bool ok; |
a70d6342 | 6727 | tree scalar_type, vectype; |
363477c0 JJ |
6728 | gimple pattern_stmt; |
6729 | gimple_seq pattern_def_seq; | |
ebfd146a | 6730 | |
73fbfcad | 6731 | if (dump_enabled_p ()) |
ebfd146a | 6732 | { |
78c60e3d SS |
6733 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
6734 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 6735 | dump_printf (MSG_NOTE, "\n"); |
8644a673 | 6736 | } |
ebfd146a | 6737 | |
1825a1f3 | 6738 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 6739 | { |
73fbfcad | 6740 | if (dump_enabled_p ()) |
78c60e3d | 6741 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6742 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
6743 | |
6744 | return false; | |
6745 | } | |
b8698a0f L |
6746 | |
6747 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
6748 | to include: |
6749 | - the COND_EXPR which is the loop exit condition | |
6750 | - any LABEL_EXPRs in the loop | |
b8698a0f | 6751 | - computations that are used only for array indexing or loop control. |
8644a673 | 6752 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 6753 | instance, therefore, all the statements are relevant. |
ebfd146a | 6754 | |
d092494c | 6755 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 6756 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
6757 | statements. In basic blocks we are called from some SLP instance |
6758 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
6759 | already will be part of SLP instance. */ | |
83197f37 IR |
6760 | |
6761 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 6762 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 6763 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 6764 | { |
9d5e7640 | 6765 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 6766 | && pattern_stmt |
9d5e7640 IR |
6767 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
6768 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
6769 | { | |
83197f37 | 6770 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
6771 | stmt = pattern_stmt; |
6772 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 6773 | if (dump_enabled_p ()) |
9d5e7640 | 6774 | { |
78c60e3d SS |
6775 | dump_printf_loc (MSG_NOTE, vect_location, |
6776 | "==> examining pattern statement: "); | |
6777 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 6778 | dump_printf (MSG_NOTE, "\n"); |
9d5e7640 IR |
6779 | } |
6780 | } | |
6781 | else | |
6782 | { | |
73fbfcad | 6783 | if (dump_enabled_p ()) |
e645e942 | 6784 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 6785 | |
9d5e7640 IR |
6786 | return true; |
6787 | } | |
8644a673 | 6788 | } |
83197f37 | 6789 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 6790 | && node == NULL |
83197f37 IR |
6791 | && pattern_stmt |
6792 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
6793 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
6794 | { | |
6795 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 6796 | if (dump_enabled_p ()) |
83197f37 | 6797 | { |
78c60e3d SS |
6798 | dump_printf_loc (MSG_NOTE, vect_location, |
6799 | "==> examining pattern statement: "); | |
6800 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 6801 | dump_printf (MSG_NOTE, "\n"); |
83197f37 IR |
6802 | } |
6803 | ||
6804 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
6805 | return false; | |
6806 | } | |
ebfd146a | 6807 | |
1107f3ae | 6808 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 6809 | && node == NULL |
363477c0 | 6810 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 6811 | { |
363477c0 | 6812 | gimple_stmt_iterator si; |
1107f3ae | 6813 | |
363477c0 JJ |
6814 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
6815 | { | |
6816 | gimple pattern_def_stmt = gsi_stmt (si); | |
6817 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) | |
6818 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
6819 | { | |
6820 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 6821 | if (dump_enabled_p ()) |
363477c0 | 6822 | { |
78c60e3d SS |
6823 | dump_printf_loc (MSG_NOTE, vect_location, |
6824 | "==> examining pattern def statement: "); | |
6825 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
e645e942 | 6826 | dump_printf (MSG_NOTE, "\n"); |
363477c0 | 6827 | } |
1107f3ae | 6828 | |
363477c0 JJ |
6829 | if (!vect_analyze_stmt (pattern_def_stmt, |
6830 | need_to_vectorize, node)) | |
6831 | return false; | |
6832 | } | |
6833 | } | |
6834 | } | |
1107f3ae | 6835 | |
8644a673 IR |
6836 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
6837 | { | |
6838 | case vect_internal_def: | |
6839 | break; | |
ebfd146a | 6840 | |
8644a673 | 6841 | case vect_reduction_def: |
7c5222ff | 6842 | case vect_nested_cycle: |
a70d6342 | 6843 | gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer |
8644a673 | 6844 | || relevance == vect_used_in_outer_by_reduction |
a70d6342 | 6845 | || relevance == vect_unused_in_scope)); |
8644a673 IR |
6846 | break; |
6847 | ||
6848 | case vect_induction_def: | |
6849 | case vect_constant_def: | |
6850 | case vect_external_def: | |
6851 | case vect_unknown_def_type: | |
6852 | default: | |
6853 | gcc_unreachable (); | |
6854 | } | |
ebfd146a | 6855 | |
a70d6342 IR |
6856 | if (bb_vinfo) |
6857 | { | |
6858 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
6859 | ||
b690cc0f | 6860 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
73fbfcad | 6861 | if (dump_enabled_p ()) |
a70d6342 | 6862 | { |
78c60e3d SS |
6863 | dump_printf_loc (MSG_NOTE, vect_location, |
6864 | "get vectype for scalar type: "); | |
6865 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
e645e942 | 6866 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
6867 | } |
6868 | ||
6869 | vectype = get_vectype_for_scalar_type (scalar_type); | |
6870 | if (!vectype) | |
6871 | { | |
73fbfcad | 6872 | if (dump_enabled_p ()) |
a70d6342 | 6873 | { |
78c60e3d SS |
6874 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6875 | "not SLPed: unsupported data-type "); | |
6876 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
6877 | scalar_type); | |
e645e942 | 6878 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
a70d6342 IR |
6879 | } |
6880 | return false; | |
6881 | } | |
6882 | ||
73fbfcad | 6883 | if (dump_enabled_p ()) |
a70d6342 | 6884 | { |
78c60e3d SS |
6885 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); |
6886 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
e645e942 | 6887 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
6888 | } |
6889 | ||
6890 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
6891 | } | |
6892 | ||
8644a673 | 6893 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 6894 | { |
8644a673 | 6895 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
6896 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
6897 | || (is_gimple_call (stmt) | |
6898 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 6899 | *need_to_vectorize = true; |
ebfd146a IR |
6900 | } |
6901 | ||
8644a673 | 6902 | ok = true; |
b8698a0f | 6903 | if (!bb_vinfo |
a70d6342 IR |
6904 | && (STMT_VINFO_RELEVANT_P (stmt_info) |
6905 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
0136f8f0 AH |
6906 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, NULL) |
6907 | || vectorizable_conversion (stmt, NULL, NULL, NULL) | |
9dc3f7de | 6908 | || vectorizable_shift (stmt, NULL, NULL, NULL) |
8644a673 IR |
6909 | || vectorizable_operation (stmt, NULL, NULL, NULL) |
6910 | || vectorizable_assignment (stmt, NULL, NULL, NULL) | |
6911 | || vectorizable_load (stmt, NULL, NULL, NULL, NULL) | |
190c2236 | 6912 | || vectorizable_call (stmt, NULL, NULL, NULL) |
8644a673 | 6913 | || vectorizable_store (stmt, NULL, NULL, NULL) |
b5aeb3bb | 6914 | || vectorizable_reduction (stmt, NULL, NULL, NULL) |
f7e531cf | 6915 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, NULL)); |
a70d6342 IR |
6916 | else |
6917 | { | |
6918 | if (bb_vinfo) | |
0136f8f0 AH |
6919 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) |
6920 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
4a00c761 | 6921 | || vectorizable_shift (stmt, NULL, NULL, node) |
9dc3f7de | 6922 | || vectorizable_operation (stmt, NULL, NULL, node) |
a70d6342 IR |
6923 | || vectorizable_assignment (stmt, NULL, NULL, node) |
6924 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
190c2236 | 6925 | || vectorizable_call (stmt, NULL, NULL, node) |
f7e531cf IR |
6926 | || vectorizable_store (stmt, NULL, NULL, node) |
6927 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node)); | |
b8698a0f | 6928 | } |
8644a673 IR |
6929 | |
6930 | if (!ok) | |
ebfd146a | 6931 | { |
73fbfcad | 6932 | if (dump_enabled_p ()) |
8644a673 | 6933 | { |
78c60e3d SS |
6934 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6935 | "not vectorized: relevant stmt not "); | |
6936 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
6937 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 6938 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
8644a673 | 6939 | } |
b8698a0f | 6940 | |
ebfd146a IR |
6941 | return false; |
6942 | } | |
6943 | ||
a70d6342 IR |
6944 | if (bb_vinfo) |
6945 | return true; | |
6946 | ||
8644a673 IR |
6947 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
6948 | need extra handling, except for vectorizable reductions. */ | |
6949 | if (STMT_VINFO_LIVE_P (stmt_info) | |
6950 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
6951 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 6952 | |
8644a673 | 6953 | if (!ok) |
ebfd146a | 6954 | { |
73fbfcad | 6955 | if (dump_enabled_p ()) |
8644a673 | 6956 | { |
78c60e3d SS |
6957 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6958 | "not vectorized: live stmt not "); | |
6959 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
6960 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 6961 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
8644a673 | 6962 | } |
b8698a0f | 6963 | |
8644a673 | 6964 | return false; |
ebfd146a IR |
6965 | } |
6966 | ||
ebfd146a IR |
6967 | return true; |
6968 | } | |
6969 | ||
6970 | ||
6971 | /* Function vect_transform_stmt. | |
6972 | ||
6973 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
6974 | ||
6975 | bool | |
6976 | vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi, | |
0d0293ac | 6977 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
6978 | slp_instance slp_node_instance) |
6979 | { | |
6980 | bool is_store = false; | |
6981 | gimple vec_stmt = NULL; | |
6982 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a | 6983 | bool done; |
ebfd146a IR |
6984 | |
6985 | switch (STMT_VINFO_TYPE (stmt_info)) | |
6986 | { | |
6987 | case type_demotion_vec_info_type: | |
ebfd146a | 6988 | case type_promotion_vec_info_type: |
ebfd146a IR |
6989 | case type_conversion_vec_info_type: |
6990 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
6991 | gcc_assert (done); | |
6992 | break; | |
6993 | ||
6994 | case induc_vec_info_type: | |
6995 | gcc_assert (!slp_node); | |
6996 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
6997 | gcc_assert (done); | |
6998 | break; | |
6999 | ||
9dc3f7de IR |
7000 | case shift_vec_info_type: |
7001 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
7002 | gcc_assert (done); | |
7003 | break; | |
7004 | ||
ebfd146a IR |
7005 | case op_vec_info_type: |
7006 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
7007 | gcc_assert (done); | |
7008 | break; | |
7009 | ||
7010 | case assignment_vec_info_type: | |
7011 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
7012 | gcc_assert (done); | |
7013 | break; | |
7014 | ||
7015 | case load_vec_info_type: | |
b8698a0f | 7016 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
7017 | slp_node_instance); |
7018 | gcc_assert (done); | |
7019 | break; | |
7020 | ||
7021 | case store_vec_info_type: | |
7022 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
7023 | gcc_assert (done); | |
0d0293ac | 7024 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
7025 | { |
7026 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 7027 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
7028 | one are skipped, and there vec_stmt_info shouldn't be freed |
7029 | meanwhile. */ | |
0d0293ac | 7030 | *grouped_store = true; |
ebfd146a IR |
7031 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
7032 | is_store = true; | |
7033 | } | |
7034 | else | |
7035 | is_store = true; | |
7036 | break; | |
7037 | ||
7038 | case condition_vec_info_type: | |
f7e531cf | 7039 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
7040 | gcc_assert (done); |
7041 | break; | |
7042 | ||
7043 | case call_vec_info_type: | |
190c2236 | 7044 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 7045 | stmt = gsi_stmt (*gsi); |
5ce9450f JJ |
7046 | if (is_gimple_call (stmt) |
7047 | && gimple_call_internal_p (stmt) | |
7048 | && gimple_call_internal_fn (stmt) == IFN_MASK_STORE) | |
7049 | is_store = true; | |
ebfd146a IR |
7050 | break; |
7051 | ||
0136f8f0 AH |
7052 | case call_simd_clone_vec_info_type: |
7053 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node); | |
7054 | stmt = gsi_stmt (*gsi); | |
7055 | break; | |
7056 | ||
ebfd146a | 7057 | case reduc_vec_info_type: |
b5aeb3bb | 7058 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
7059 | gcc_assert (done); |
7060 | break; | |
7061 | ||
7062 | default: | |
7063 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
7064 | { | |
73fbfcad | 7065 | if (dump_enabled_p ()) |
78c60e3d | 7066 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7067 | "stmt not supported.\n"); |
ebfd146a IR |
7068 | gcc_unreachable (); |
7069 | } | |
7070 | } | |
7071 | ||
7072 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that | |
7073 | is being vectorized, but outside the immediately enclosing loop. */ | |
7074 | if (vec_stmt | |
a70d6342 IR |
7075 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
7076 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
7077 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
7078 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
7079 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 7080 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 7081 | vect_used_in_outer_by_reduction)) |
ebfd146a | 7082 | { |
a70d6342 IR |
7083 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
7084 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
7085 | imm_use_iterator imm_iter; |
7086 | use_operand_p use_p; | |
7087 | tree scalar_dest; | |
7088 | gimple exit_phi; | |
7089 | ||
73fbfcad | 7090 | if (dump_enabled_p ()) |
78c60e3d | 7091 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 7092 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
7093 | |
7094 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
7095 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
7096 | STMT). */ | |
7097 | if (gimple_code (stmt) == GIMPLE_PHI) | |
7098 | scalar_dest = PHI_RESULT (stmt); | |
7099 | else | |
7100 | scalar_dest = gimple_assign_lhs (stmt); | |
7101 | ||
7102 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
7103 | { | |
7104 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
7105 | { | |
7106 | exit_phi = USE_STMT (use_p); | |
7107 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
7108 | } | |
7109 | } | |
7110 | } | |
7111 | ||
7112 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
7113 | being vectorized. */ | |
7114 | if (STMT_VINFO_LIVE_P (stmt_info) | |
7115 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
7116 | { | |
7117 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
7118 | gcc_assert (done); | |
7119 | } | |
7120 | ||
7121 | if (vec_stmt) | |
83197f37 | 7122 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 7123 | |
b8698a0f | 7124 | return is_store; |
ebfd146a IR |
7125 | } |
7126 | ||
7127 | ||
b8698a0f | 7128 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
7129 | stmt_vec_info. */ |
7130 | ||
7131 | void | |
7132 | vect_remove_stores (gimple first_stmt) | |
7133 | { | |
7134 | gimple next = first_stmt; | |
7135 | gimple tmp; | |
7136 | gimple_stmt_iterator next_si; | |
7137 | ||
7138 | while (next) | |
7139 | { | |
78048b1c JJ |
7140 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
7141 | ||
7142 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
7143 | if (is_pattern_stmt_p (stmt_info)) | |
7144 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
7145 | /* Free the attached stmt_vec_info and remove the stmt. */ |
7146 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 7147 | unlink_stmt_vdef (next); |
ebfd146a | 7148 | gsi_remove (&next_si, true); |
3d3f2249 | 7149 | release_defs (next); |
ebfd146a IR |
7150 | free_stmt_vec_info (next); |
7151 | next = tmp; | |
7152 | } | |
7153 | } | |
7154 | ||
7155 | ||
7156 | /* Function new_stmt_vec_info. | |
7157 | ||
7158 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
7159 | ||
7160 | stmt_vec_info | |
b8698a0f | 7161 | new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 7162 | bb_vec_info bb_vinfo) |
ebfd146a IR |
7163 | { |
7164 | stmt_vec_info res; | |
7165 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
7166 | ||
7167 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
7168 | STMT_VINFO_STMT (res) = stmt; | |
7169 | STMT_VINFO_LOOP_VINFO (res) = loop_vinfo; | |
a70d6342 | 7170 | STMT_VINFO_BB_VINFO (res) = bb_vinfo; |
8644a673 | 7171 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
7172 | STMT_VINFO_LIVE_P (res) = false; |
7173 | STMT_VINFO_VECTYPE (res) = NULL; | |
7174 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 7175 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
7176 | STMT_VINFO_IN_PATTERN_P (res) = false; |
7177 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 7178 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a IR |
7179 | STMT_VINFO_DATA_REF (res) = NULL; |
7180 | ||
7181 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
7182 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
7183 | STMT_VINFO_DR_INIT (res) = NULL; | |
7184 | STMT_VINFO_DR_STEP (res) = NULL; | |
7185 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
7186 | ||
7187 | if (gimple_code (stmt) == GIMPLE_PHI | |
7188 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
7189 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
7190 | else | |
8644a673 IR |
7191 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
7192 | ||
9771b263 | 7193 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 7194 | STMT_SLP_TYPE (res) = loop_vect; |
e14c1050 IR |
7195 | GROUP_FIRST_ELEMENT (res) = NULL; |
7196 | GROUP_NEXT_ELEMENT (res) = NULL; | |
7197 | GROUP_SIZE (res) = 0; | |
7198 | GROUP_STORE_COUNT (res) = 0; | |
7199 | GROUP_GAP (res) = 0; | |
7200 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
7201 | |
7202 | return res; | |
7203 | } | |
7204 | ||
7205 | ||
7206 | /* Create a hash table for stmt_vec_info. */ | |
7207 | ||
7208 | void | |
7209 | init_stmt_vec_info_vec (void) | |
7210 | { | |
9771b263 DN |
7211 | gcc_assert (!stmt_vec_info_vec.exists ()); |
7212 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
7213 | } |
7214 | ||
7215 | ||
7216 | /* Free hash table for stmt_vec_info. */ | |
7217 | ||
7218 | void | |
7219 | free_stmt_vec_info_vec (void) | |
7220 | { | |
93675444 JJ |
7221 | unsigned int i; |
7222 | vec_void_p info; | |
7223 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) | |
7224 | if (info != NULL) | |
7225 | free_stmt_vec_info (STMT_VINFO_STMT ((stmt_vec_info) info)); | |
9771b263 DN |
7226 | gcc_assert (stmt_vec_info_vec.exists ()); |
7227 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
7228 | } |
7229 | ||
7230 | ||
7231 | /* Free stmt vectorization related info. */ | |
7232 | ||
7233 | void | |
7234 | free_stmt_vec_info (gimple stmt) | |
7235 | { | |
7236 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
7237 | ||
7238 | if (!stmt_info) | |
7239 | return; | |
7240 | ||
78048b1c JJ |
7241 | /* Check if this statement has a related "pattern stmt" |
7242 | (introduced by the vectorizer during the pattern recognition | |
7243 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
7244 | too. */ | |
7245 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
7246 | { | |
7247 | stmt_vec_info patt_info | |
7248 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
7249 | if (patt_info) | |
7250 | { | |
363477c0 JJ |
7251 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
7252 | if (seq) | |
7253 | { | |
7254 | gimple_stmt_iterator si; | |
7255 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
7256 | free_stmt_vec_info (gsi_stmt (si)); | |
7257 | } | |
78048b1c JJ |
7258 | free_stmt_vec_info (STMT_VINFO_RELATED_STMT (stmt_info)); |
7259 | } | |
7260 | } | |
7261 | ||
9771b263 | 7262 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
ebfd146a IR |
7263 | set_vinfo_for_stmt (stmt, NULL); |
7264 | free (stmt_info); | |
7265 | } | |
7266 | ||
7267 | ||
bb67d9c7 | 7268 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 7269 | |
bb67d9c7 | 7270 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
7271 | by the target. */ |
7272 | ||
bb67d9c7 RG |
7273 | static tree |
7274 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a IR |
7275 | { |
7276 | enum machine_mode inner_mode = TYPE_MODE (scalar_type); | |
cc4b5170 | 7277 | enum machine_mode simd_mode; |
2f816591 | 7278 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
7279 | int nunits; |
7280 | tree vectype; | |
7281 | ||
cc4b5170 | 7282 | if (nbytes == 0) |
ebfd146a IR |
7283 | return NULL_TREE; |
7284 | ||
48f2e373 RB |
7285 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
7286 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
7287 | return NULL_TREE; | |
7288 | ||
7b7b1813 RG |
7289 | /* For vector types of elements whose mode precision doesn't |
7290 | match their types precision we use a element type of mode | |
7291 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
7292 | they support the proper result truncation/extension. |
7293 | We also make sure to build vector types with INTEGER_TYPE | |
7294 | component type only. */ | |
6d7971b8 | 7295 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
7296 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
7297 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
7298 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
7299 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 7300 | |
ccbf5bb4 RG |
7301 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
7302 | When the component mode passes the above test simply use a type | |
7303 | corresponding to that mode. The theory is that any use that | |
7304 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 7305 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 7306 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
7307 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
7308 | ||
7309 | /* We can't build a vector type of elements with alignment bigger than | |
7310 | their size. */ | |
dfc2e2ac | 7311 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
7312 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
7313 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 7314 | |
dfc2e2ac RB |
7315 | /* If we felt back to using the mode fail if there was |
7316 | no scalar type for it. */ | |
7317 | if (scalar_type == NULL_TREE) | |
7318 | return NULL_TREE; | |
7319 | ||
bb67d9c7 RG |
7320 | /* If no size was supplied use the mode the target prefers. Otherwise |
7321 | lookup a vector mode of the specified size. */ | |
7322 | if (size == 0) | |
7323 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
7324 | else | |
7325 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
7326 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
7327 | if (nunits <= 1) | |
7328 | return NULL_TREE; | |
ebfd146a IR |
7329 | |
7330 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
7331 | |
7332 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
7333 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 7334 | return NULL_TREE; |
ebfd146a IR |
7335 | |
7336 | return vectype; | |
7337 | } | |
7338 | ||
bb67d9c7 RG |
7339 | unsigned int current_vector_size; |
7340 | ||
7341 | /* Function get_vectype_for_scalar_type. | |
7342 | ||
7343 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
7344 | by the target. */ | |
7345 | ||
7346 | tree | |
7347 | get_vectype_for_scalar_type (tree scalar_type) | |
7348 | { | |
7349 | tree vectype; | |
7350 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
7351 | current_vector_size); | |
7352 | if (vectype | |
7353 | && current_vector_size == 0) | |
7354 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
7355 | return vectype; | |
7356 | } | |
7357 | ||
b690cc0f RG |
7358 | /* Function get_same_sized_vectype |
7359 | ||
7360 | Returns a vector type corresponding to SCALAR_TYPE of size | |
7361 | VECTOR_TYPE if supported by the target. */ | |
7362 | ||
7363 | tree | |
bb67d9c7 | 7364 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 7365 | { |
bb67d9c7 RG |
7366 | return get_vectype_for_scalar_type_and_size |
7367 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
7368 | } |
7369 | ||
ebfd146a IR |
7370 | /* Function vect_is_simple_use. |
7371 | ||
7372 | Input: | |
a70d6342 IR |
7373 | LOOP_VINFO - the vect info of the loop that is being vectorized. |
7374 | BB_VINFO - the vect info of the basic block that is being vectorized. | |
24ee1384 | 7375 | OPERAND - operand of STMT in the loop or bb. |
ebfd146a IR |
7376 | DEF - the defining stmt in case OPERAND is an SSA_NAME. |
7377 | ||
7378 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 7379 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 7380 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 7381 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
7382 | is the case in reduction/induction computations). |
7383 | For basic blocks, supportable operands are constants and bb invariants. | |
7384 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
7385 | |
7386 | bool | |
24ee1384 | 7387 | vect_is_simple_use (tree operand, gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 7388 | bb_vec_info bb_vinfo, gimple *def_stmt, |
ebfd146a | 7389 | tree *def, enum vect_def_type *dt) |
b8698a0f | 7390 | { |
ebfd146a IR |
7391 | basic_block bb; |
7392 | stmt_vec_info stmt_vinfo; | |
a70d6342 | 7393 | struct loop *loop = NULL; |
b8698a0f | 7394 | |
a70d6342 IR |
7395 | if (loop_vinfo) |
7396 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
7397 | |
7398 | *def_stmt = NULL; | |
7399 | *def = NULL_TREE; | |
b8698a0f | 7400 | |
73fbfcad | 7401 | if (dump_enabled_p ()) |
ebfd146a | 7402 | { |
78c60e3d SS |
7403 | dump_printf_loc (MSG_NOTE, vect_location, |
7404 | "vect_is_simple_use: operand "); | |
7405 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 7406 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 7407 | } |
b8698a0f | 7408 | |
b758f602 | 7409 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
7410 | { |
7411 | *dt = vect_constant_def; | |
7412 | return true; | |
7413 | } | |
b8698a0f | 7414 | |
ebfd146a IR |
7415 | if (is_gimple_min_invariant (operand)) |
7416 | { | |
7417 | *def = operand; | |
8644a673 | 7418 | *dt = vect_external_def; |
ebfd146a IR |
7419 | return true; |
7420 | } | |
7421 | ||
7422 | if (TREE_CODE (operand) == PAREN_EXPR) | |
7423 | { | |
73fbfcad | 7424 | if (dump_enabled_p ()) |
e645e942 | 7425 | dump_printf_loc (MSG_NOTE, vect_location, "non-associatable copy.\n"); |
ebfd146a IR |
7426 | operand = TREE_OPERAND (operand, 0); |
7427 | } | |
b8698a0f | 7428 | |
ebfd146a IR |
7429 | if (TREE_CODE (operand) != SSA_NAME) |
7430 | { | |
73fbfcad | 7431 | if (dump_enabled_p ()) |
78c60e3d | 7432 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7433 | "not ssa-name.\n"); |
ebfd146a IR |
7434 | return false; |
7435 | } | |
b8698a0f | 7436 | |
ebfd146a IR |
7437 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
7438 | if (*def_stmt == NULL) | |
7439 | { | |
73fbfcad | 7440 | if (dump_enabled_p ()) |
78c60e3d | 7441 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7442 | "no def_stmt.\n"); |
ebfd146a IR |
7443 | return false; |
7444 | } | |
7445 | ||
73fbfcad | 7446 | if (dump_enabled_p ()) |
ebfd146a | 7447 | { |
78c60e3d SS |
7448 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
7449 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
e645e942 | 7450 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
7451 | } |
7452 | ||
8644a673 | 7453 | /* Empty stmt is expected only in case of a function argument. |
ebfd146a IR |
7454 | (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */ |
7455 | if (gimple_nop_p (*def_stmt)) | |
7456 | { | |
7457 | *def = operand; | |
8644a673 | 7458 | *dt = vect_external_def; |
ebfd146a IR |
7459 | return true; |
7460 | } | |
7461 | ||
7462 | bb = gimple_bb (*def_stmt); | |
a70d6342 IR |
7463 | |
7464 | if ((loop && !flow_bb_inside_loop_p (loop, bb)) | |
7465 | || (!loop && bb != BB_VINFO_BB (bb_vinfo)) | |
b8698a0f | 7466 | || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI)) |
8644a673 | 7467 | *dt = vect_external_def; |
ebfd146a IR |
7468 | else |
7469 | { | |
7470 | stmt_vinfo = vinfo_for_stmt (*def_stmt); | |
7471 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
7472 | } | |
7473 | ||
24ee1384 IR |
7474 | if (*dt == vect_unknown_def_type |
7475 | || (stmt | |
7476 | && *dt == vect_double_reduction_def | |
7477 | && gimple_code (stmt) != GIMPLE_PHI)) | |
ebfd146a | 7478 | { |
73fbfcad | 7479 | if (dump_enabled_p ()) |
78c60e3d | 7480 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7481 | "Unsupported pattern.\n"); |
ebfd146a IR |
7482 | return false; |
7483 | } | |
7484 | ||
73fbfcad | 7485 | if (dump_enabled_p ()) |
e645e942 | 7486 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: %d.\n", *dt); |
ebfd146a IR |
7487 | |
7488 | switch (gimple_code (*def_stmt)) | |
7489 | { | |
7490 | case GIMPLE_PHI: | |
7491 | *def = gimple_phi_result (*def_stmt); | |
7492 | break; | |
7493 | ||
7494 | case GIMPLE_ASSIGN: | |
7495 | *def = gimple_assign_lhs (*def_stmt); | |
7496 | break; | |
7497 | ||
7498 | case GIMPLE_CALL: | |
7499 | *def = gimple_call_lhs (*def_stmt); | |
7500 | if (*def != NULL) | |
7501 | break; | |
7502 | /* FALLTHRU */ | |
7503 | default: | |
73fbfcad | 7504 | if (dump_enabled_p ()) |
78c60e3d | 7505 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7506 | "unsupported defining stmt:\n"); |
ebfd146a IR |
7507 | return false; |
7508 | } | |
7509 | ||
7510 | return true; | |
7511 | } | |
7512 | ||
b690cc0f RG |
7513 | /* Function vect_is_simple_use_1. |
7514 | ||
7515 | Same as vect_is_simple_use_1 but also determines the vector operand | |
7516 | type of OPERAND and stores it to *VECTYPE. If the definition of | |
7517 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
7518 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
7519 | is responsible to compute the best suited vector type for the | |
7520 | scalar operand. */ | |
7521 | ||
7522 | bool | |
24ee1384 | 7523 | vect_is_simple_use_1 (tree operand, gimple stmt, loop_vec_info loop_vinfo, |
b690cc0f RG |
7524 | bb_vec_info bb_vinfo, gimple *def_stmt, |
7525 | tree *def, enum vect_def_type *dt, tree *vectype) | |
7526 | { | |
24ee1384 IR |
7527 | if (!vect_is_simple_use (operand, stmt, loop_vinfo, bb_vinfo, def_stmt, |
7528 | def, dt)) | |
b690cc0f RG |
7529 | return false; |
7530 | ||
7531 | /* Now get a vector type if the def is internal, otherwise supply | |
7532 | NULL_TREE and leave it up to the caller to figure out a proper | |
7533 | type for the use stmt. */ | |
7534 | if (*dt == vect_internal_def | |
7535 | || *dt == vect_induction_def | |
7536 | || *dt == vect_reduction_def | |
7537 | || *dt == vect_double_reduction_def | |
7538 | || *dt == vect_nested_cycle) | |
7539 | { | |
7540 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
7541 | |
7542 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
7543 | && !STMT_VINFO_RELEVANT (stmt_info) | |
7544 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 7545 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 7546 | |
b690cc0f RG |
7547 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
7548 | gcc_assert (*vectype != NULL_TREE); | |
7549 | } | |
7550 | else if (*dt == vect_uninitialized_def | |
7551 | || *dt == vect_constant_def | |
7552 | || *dt == vect_external_def) | |
7553 | *vectype = NULL_TREE; | |
7554 | else | |
7555 | gcc_unreachable (); | |
7556 | ||
7557 | return true; | |
7558 | } | |
7559 | ||
ebfd146a IR |
7560 | |
7561 | /* Function supportable_widening_operation | |
7562 | ||
b8698a0f L |
7563 | Check whether an operation represented by the code CODE is a |
7564 | widening operation that is supported by the target platform in | |
b690cc0f RG |
7565 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
7566 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 7567 | |
ebfd146a IR |
7568 | Widening operations we currently support are NOP (CONVERT), FLOAT |
7569 | and WIDEN_MULT. This function checks if these operations are supported | |
7570 | by the target platform either directly (via vector tree-codes), or via | |
7571 | target builtins. | |
7572 | ||
7573 | Output: | |
b8698a0f L |
7574 | - CODE1 and CODE2 are codes of vector operations to be used when |
7575 | vectorizing the operation, if available. | |
ebfd146a IR |
7576 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
7577 | case of multi-step conversion (like char->short->int - in that case | |
7578 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
7579 | - INTERM_TYPES contains the intermediate type required to perform the |
7580 | widening operation (short in the above example). */ | |
ebfd146a IR |
7581 | |
7582 | bool | |
b690cc0f RG |
7583 | supportable_widening_operation (enum tree_code code, gimple stmt, |
7584 | tree vectype_out, tree vectype_in, | |
ebfd146a IR |
7585 | enum tree_code *code1, enum tree_code *code2, |
7586 | int *multi_step_cvt, | |
9771b263 | 7587 | vec<tree> *interm_types) |
ebfd146a IR |
7588 | { |
7589 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
7590 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 7591 | struct loop *vect_loop = NULL; |
ebfd146a | 7592 | enum machine_mode vec_mode; |
81f40b79 | 7593 | enum insn_code icode1, icode2; |
ebfd146a | 7594 | optab optab1, optab2; |
b690cc0f RG |
7595 | tree vectype = vectype_in; |
7596 | tree wide_vectype = vectype_out; | |
ebfd146a | 7597 | enum tree_code c1, c2; |
4a00c761 JJ |
7598 | int i; |
7599 | tree prev_type, intermediate_type; | |
7600 | enum machine_mode intermediate_mode, prev_mode; | |
7601 | optab optab3, optab4; | |
ebfd146a | 7602 | |
4a00c761 | 7603 | *multi_step_cvt = 0; |
4ef69dfc IR |
7604 | if (loop_info) |
7605 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
7606 | ||
ebfd146a IR |
7607 | switch (code) |
7608 | { | |
7609 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
7610 | /* The result of a vectorized widening operation usually requires |
7611 | two vectors (because the widened results do not fit into one vector). | |
7612 | The generated vector results would normally be expected to be | |
7613 | generated in the same order as in the original scalar computation, | |
7614 | i.e. if 8 results are generated in each vector iteration, they are | |
7615 | to be organized as follows: | |
7616 | vect1: [res1,res2,res3,res4], | |
7617 | vect2: [res5,res6,res7,res8]. | |
7618 | ||
7619 | However, in the special case that the result of the widening | |
7620 | operation is used in a reduction computation only, the order doesn't | |
7621 | matter (because when vectorizing a reduction we change the order of | |
7622 | the computation). Some targets can take advantage of this and | |
7623 | generate more efficient code. For example, targets like Altivec, | |
7624 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
7625 | generate the following vectors: | |
7626 | vect1: [res1,res3,res5,res7], | |
7627 | vect2: [res2,res4,res6,res8]. | |
7628 | ||
7629 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
7630 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
7631 | iterations in parallel). We therefore don't allow to change the | |
7632 | order of the computation in the inner-loop during outer-loop | |
7633 | vectorization. */ | |
7634 | /* TODO: Another case in which order doesn't *really* matter is when we | |
7635 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
7636 | Normally, pack_trunc performs an even/odd permute, whereas the | |
7637 | repack from an even/odd expansion would be an interleave, which | |
7638 | would be significantly simpler for e.g. AVX2. */ | |
7639 | /* In any case, in order to avoid duplicating the code below, recurse | |
7640 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
7641 | are properly set up for the caller. If we fail, we'll continue with | |
7642 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
7643 | if (vect_loop | |
7644 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
7645 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
7646 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
7647 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
7648 | code1, code2, multi_step_cvt, |
7649 | interm_types)) | |
6ae6116f | 7650 | return true; |
4a00c761 JJ |
7651 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
7652 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
7653 | break; |
7654 | ||
6ae6116f RH |
7655 | case VEC_WIDEN_MULT_EVEN_EXPR: |
7656 | /* Support the recursion induced just above. */ | |
7657 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
7658 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
7659 | break; | |
7660 | ||
36ba4aae | 7661 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
7662 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
7663 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
7664 | break; |
7665 | ||
ebfd146a | 7666 | CASE_CONVERT: |
4a00c761 JJ |
7667 | c1 = VEC_UNPACK_LO_EXPR; |
7668 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
7669 | break; |
7670 | ||
7671 | case FLOAT_EXPR: | |
4a00c761 JJ |
7672 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
7673 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
7674 | break; |
7675 | ||
7676 | case FIX_TRUNC_EXPR: | |
7677 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
7678 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
7679 | computing the operation. */ | |
7680 | return false; | |
7681 | ||
7682 | default: | |
7683 | gcc_unreachable (); | |
7684 | } | |
7685 | ||
6ae6116f | 7686 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
4a00c761 JJ |
7687 | { |
7688 | enum tree_code ctmp = c1; | |
7689 | c1 = c2; | |
7690 | c2 = ctmp; | |
7691 | } | |
7692 | ||
ebfd146a IR |
7693 | if (code == FIX_TRUNC_EXPR) |
7694 | { | |
7695 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
7696 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
7697 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
7698 | } |
7699 | else | |
7700 | { | |
7701 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
7702 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
7703 | } | |
7704 | ||
7705 | if (!optab1 || !optab2) | |
7706 | return false; | |
7707 | ||
7708 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
7709 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
7710 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
7711 | return false; |
7712 | ||
4a00c761 JJ |
7713 | *code1 = c1; |
7714 | *code2 = c2; | |
7715 | ||
7716 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
7717 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
7718 | return true; | |
7719 | ||
b8698a0f | 7720 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 7721 | types. */ |
ebfd146a | 7722 | |
4a00c761 JJ |
7723 | prev_type = vectype; |
7724 | prev_mode = vec_mode; | |
b8698a0f | 7725 | |
4a00c761 JJ |
7726 | if (!CONVERT_EXPR_CODE_P (code)) |
7727 | return false; | |
b8698a0f | 7728 | |
4a00c761 JJ |
7729 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
7730 | intermediate steps in promotion sequence. We try | |
7731 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
7732 | not. */ | |
9771b263 | 7733 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
7734 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
7735 | { | |
7736 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
7737 | intermediate_type | |
7738 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
7739 | TYPE_UNSIGNED (prev_type)); | |
7740 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
7741 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
7742 | ||
7743 | if (!optab3 || !optab4 | |
7744 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
7745 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
7746 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
7747 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
7748 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
7749 | == CODE_FOR_nothing) | |
7750 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
7751 | == CODE_FOR_nothing)) | |
7752 | break; | |
ebfd146a | 7753 | |
9771b263 | 7754 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
7755 | (*multi_step_cvt)++; |
7756 | ||
7757 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
7758 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
7759 | return true; | |
7760 | ||
7761 | prev_type = intermediate_type; | |
7762 | prev_mode = intermediate_mode; | |
ebfd146a IR |
7763 | } |
7764 | ||
9771b263 | 7765 | interm_types->release (); |
4a00c761 | 7766 | return false; |
ebfd146a IR |
7767 | } |
7768 | ||
7769 | ||
7770 | /* Function supportable_narrowing_operation | |
7771 | ||
b8698a0f L |
7772 | Check whether an operation represented by the code CODE is a |
7773 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
7774 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
7775 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 7776 | |
ebfd146a | 7777 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 7778 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
7779 | the target platform directly via vector tree-codes. |
7780 | ||
7781 | Output: | |
b8698a0f L |
7782 | - CODE1 is the code of a vector operation to be used when |
7783 | vectorizing the operation, if available. | |
ebfd146a IR |
7784 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
7785 | case of multi-step conversion (like int->short->char - in that case | |
7786 | MULTI_STEP_CVT will be 1). | |
7787 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 7788 | narrowing operation (short in the above example). */ |
ebfd146a IR |
7789 | |
7790 | bool | |
7791 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 7792 | tree vectype_out, tree vectype_in, |
ebfd146a | 7793 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 7794 | vec<tree> *interm_types) |
ebfd146a IR |
7795 | { |
7796 | enum machine_mode vec_mode; | |
7797 | enum insn_code icode1; | |
7798 | optab optab1, interm_optab; | |
b690cc0f RG |
7799 | tree vectype = vectype_in; |
7800 | tree narrow_vectype = vectype_out; | |
ebfd146a | 7801 | enum tree_code c1; |
4a00c761 JJ |
7802 | tree intermediate_type; |
7803 | enum machine_mode intermediate_mode, prev_mode; | |
ebfd146a | 7804 | int i; |
4a00c761 | 7805 | bool uns; |
ebfd146a | 7806 | |
4a00c761 | 7807 | *multi_step_cvt = 0; |
ebfd146a IR |
7808 | switch (code) |
7809 | { | |
7810 | CASE_CONVERT: | |
7811 | c1 = VEC_PACK_TRUNC_EXPR; | |
7812 | break; | |
7813 | ||
7814 | case FIX_TRUNC_EXPR: | |
7815 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
7816 | break; | |
7817 | ||
7818 | case FLOAT_EXPR: | |
7819 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
7820 | tree code and optabs used for computing the operation. */ | |
7821 | return false; | |
7822 | ||
7823 | default: | |
7824 | gcc_unreachable (); | |
7825 | } | |
7826 | ||
7827 | if (code == FIX_TRUNC_EXPR) | |
7828 | /* The signedness is determined from output operand. */ | |
b690cc0f | 7829 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
7830 | else |
7831 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
7832 | ||
7833 | if (!optab1) | |
7834 | return false; | |
7835 | ||
7836 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 7837 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
7838 | return false; |
7839 | ||
4a00c761 JJ |
7840 | *code1 = c1; |
7841 | ||
7842 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
7843 | return true; | |
7844 | ||
ebfd146a IR |
7845 | /* Check if it's a multi-step conversion that can be done using intermediate |
7846 | types. */ | |
4a00c761 JJ |
7847 | prev_mode = vec_mode; |
7848 | if (code == FIX_TRUNC_EXPR) | |
7849 | uns = TYPE_UNSIGNED (vectype_out); | |
7850 | else | |
7851 | uns = TYPE_UNSIGNED (vectype); | |
7852 | ||
7853 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
7854 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
7855 | costly than signed. */ | |
7856 | if (code == FIX_TRUNC_EXPR && uns) | |
7857 | { | |
7858 | enum insn_code icode2; | |
7859 | ||
7860 | intermediate_type | |
7861 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
7862 | interm_optab | |
7863 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 7864 | if (interm_optab != unknown_optab |
4a00c761 JJ |
7865 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
7866 | && insn_data[icode1].operand[0].mode | |
7867 | == insn_data[icode2].operand[0].mode) | |
7868 | { | |
7869 | uns = false; | |
7870 | optab1 = interm_optab; | |
7871 | icode1 = icode2; | |
7872 | } | |
7873 | } | |
ebfd146a | 7874 | |
4a00c761 JJ |
7875 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
7876 | intermediate steps in promotion sequence. We try | |
7877 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 7878 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
7879 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
7880 | { | |
7881 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
7882 | intermediate_type | |
7883 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
7884 | interm_optab | |
7885 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
7886 | optab_default); | |
7887 | if (!interm_optab | |
7888 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
7889 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
7890 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
7891 | == CODE_FOR_nothing)) | |
7892 | break; | |
7893 | ||
9771b263 | 7894 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
7895 | (*multi_step_cvt)++; |
7896 | ||
7897 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
7898 | return true; | |
7899 | ||
7900 | prev_mode = intermediate_mode; | |
7901 | optab1 = interm_optab; | |
ebfd146a IR |
7902 | } |
7903 | ||
9771b263 | 7904 | interm_types->release (); |
4a00c761 | 7905 | return false; |
ebfd146a | 7906 | } |