]>
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" |
ebfd146a | 49 | #include "expr.h" |
7ee2468b | 50 | #include "recog.h" /* FIXME: for insn_data */ |
ebfd146a | 51 | #include "optabs.h" |
718f9c0f | 52 | #include "diagnostic-core.h" |
ebfd146a | 53 | #include "tree-vectorizer.h" |
7ee2468b | 54 | #include "dumpfile.h" |
ebfd146a | 55 | |
7ee2468b SB |
56 | /* For lang_hooks.types.type_for_mode. */ |
57 | #include "langhooks.h" | |
ebfd146a | 58 | |
c3e7ee41 BS |
59 | /* Return the vectorized type for the given statement. */ |
60 | ||
61 | tree | |
62 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
63 | { | |
64 | return STMT_VINFO_VECTYPE (stmt_info); | |
65 | } | |
66 | ||
67 | /* Return TRUE iff the given statement is in an inner loop relative to | |
68 | the loop being vectorized. */ | |
69 | bool | |
70 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
71 | { | |
72 | gimple stmt = STMT_VINFO_STMT (stmt_info); | |
73 | basic_block bb = gimple_bb (stmt); | |
74 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
75 | struct loop* loop; | |
76 | ||
77 | if (!loop_vinfo) | |
78 | return false; | |
79 | ||
80 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
81 | ||
82 | return (bb->loop_father == loop->inner); | |
83 | } | |
84 | ||
85 | /* Record the cost of a statement, either by directly informing the | |
86 | target model or by saving it in a vector for later processing. | |
87 | Return a preliminary estimate of the statement's cost. */ | |
88 | ||
89 | unsigned | |
92345349 | 90 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 91 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 92 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 93 | { |
92345349 | 94 | if (body_cost_vec) |
c3e7ee41 | 95 | { |
92345349 BS |
96 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
97 | add_stmt_info_to_vec (body_cost_vec, count, kind, | |
98 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
99 | misalign); | |
c3e7ee41 | 100 | return (unsigned) |
92345349 | 101 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
102 | |
103 | } | |
104 | else | |
105 | { | |
106 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
107 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
108 | void *target_cost_data; | |
109 | ||
110 | if (loop_vinfo) | |
111 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
112 | else | |
113 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
114 | ||
92345349 BS |
115 | return add_stmt_cost (target_cost_data, count, kind, stmt_info, |
116 | misalign, where); | |
c3e7ee41 BS |
117 | } |
118 | } | |
119 | ||
272c6793 RS |
120 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
121 | ||
122 | static tree | |
123 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
124 | { | |
125 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
126 | "vect_array"); | |
127 | } | |
128 | ||
129 | /* ARRAY is an array of vectors created by create_vector_array. | |
130 | Return an SSA_NAME for the vector in index N. The reference | |
131 | is part of the vectorization of STMT and the vector is associated | |
132 | with scalar destination SCALAR_DEST. */ | |
133 | ||
134 | static tree | |
135 | read_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree scalar_dest, | |
136 | tree array, unsigned HOST_WIDE_INT n) | |
137 | { | |
138 | tree vect_type, vect, vect_name, array_ref; | |
139 | gimple new_stmt; | |
140 | ||
141 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
142 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
143 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
144 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
145 | build_int_cst (size_type_node, n), | |
146 | NULL_TREE, NULL_TREE); | |
147 | ||
148 | new_stmt = gimple_build_assign (vect, array_ref); | |
149 | vect_name = make_ssa_name (vect, new_stmt); | |
150 | gimple_assign_set_lhs (new_stmt, vect_name); | |
151 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
152 | |
153 | return vect_name; | |
154 | } | |
155 | ||
156 | /* ARRAY is an array of vectors created by create_vector_array. | |
157 | Emit code to store SSA_NAME VECT in index N of the array. | |
158 | The store is part of the vectorization of STMT. */ | |
159 | ||
160 | static void | |
161 | write_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree vect, | |
162 | tree array, unsigned HOST_WIDE_INT n) | |
163 | { | |
164 | tree array_ref; | |
165 | gimple new_stmt; | |
166 | ||
167 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
168 | build_int_cst (size_type_node, n), | |
169 | NULL_TREE, NULL_TREE); | |
170 | ||
171 | new_stmt = gimple_build_assign (array_ref, vect); | |
172 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
173 | } |
174 | ||
175 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
176 | of *PTR. The memory reference replaces those in FIRST_DR | |
177 | (and its group). */ | |
178 | ||
179 | static tree | |
180 | create_array_ref (tree type, tree ptr, struct data_reference *first_dr) | |
181 | { | |
272c6793 RS |
182 | tree mem_ref, alias_ptr_type; |
183 | ||
184 | alias_ptr_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
185 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); | |
186 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 187 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
188 | return mem_ref; |
189 | } | |
190 | ||
ebfd146a IR |
191 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
192 | ||
193 | /* Function vect_mark_relevant. | |
194 | ||
195 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
196 | ||
197 | static void | |
9771b263 | 198 | vect_mark_relevant (vec<gimple> *worklist, gimple stmt, |
83197f37 IR |
199 | enum vect_relevant relevant, bool live_p, |
200 | bool used_in_pattern) | |
ebfd146a IR |
201 | { |
202 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
203 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
204 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
83197f37 | 205 | gimple pattern_stmt; |
ebfd146a | 206 | |
73fbfcad | 207 | if (dump_enabled_p ()) |
78c60e3d | 208 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 209 | "mark relevant %d, live %d.\n", relevant, live_p); |
ebfd146a | 210 | |
83197f37 IR |
211 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
212 | related pattern stmt instead of the original stmt. However, such stmts | |
213 | may have their own uses that are not in any pattern, in such cases the | |
214 | stmt itself should be marked. */ | |
ebfd146a IR |
215 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
216 | { | |
83197f37 IR |
217 | bool found = false; |
218 | if (!used_in_pattern) | |
219 | { | |
220 | imm_use_iterator imm_iter; | |
221 | use_operand_p use_p; | |
222 | gimple use_stmt; | |
223 | tree lhs; | |
13c931c9 JJ |
224 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
225 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a | 226 | |
83197f37 IR |
227 | if (is_gimple_assign (stmt)) |
228 | lhs = gimple_assign_lhs (stmt); | |
229 | else | |
230 | lhs = gimple_call_lhs (stmt); | |
ebfd146a | 231 | |
83197f37 IR |
232 | /* This use is out of pattern use, if LHS has other uses that are |
233 | pattern uses, we should mark the stmt itself, and not the pattern | |
234 | stmt. */ | |
ab0ef706 JJ |
235 | if (TREE_CODE (lhs) == SSA_NAME) |
236 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs) | |
237 | { | |
238 | if (is_gimple_debug (USE_STMT (use_p))) | |
239 | continue; | |
240 | use_stmt = USE_STMT (use_p); | |
241 | ||
13c931c9 JJ |
242 | if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) |
243 | continue; | |
244 | ||
ab0ef706 JJ |
245 | if (vinfo_for_stmt (use_stmt) |
246 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt))) | |
247 | { | |
248 | found = true; | |
249 | break; | |
250 | } | |
251 | } | |
83197f37 IR |
252 | } |
253 | ||
254 | if (!found) | |
255 | { | |
256 | /* This is the last stmt in a sequence that was detected as a | |
257 | pattern that can potentially be vectorized. Don't mark the stmt | |
258 | as relevant/live because it's not going to be vectorized. | |
259 | Instead mark the pattern-stmt that replaces it. */ | |
260 | ||
261 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
262 | ||
73fbfcad | 263 | if (dump_enabled_p ()) |
78c60e3d SS |
264 | dump_printf_loc (MSG_NOTE, vect_location, |
265 | "last stmt in pattern. don't mark" | |
e645e942 | 266 | " relevant/live.\n"); |
83197f37 IR |
267 | stmt_info = vinfo_for_stmt (pattern_stmt); |
268 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
269 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
270 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
271 | stmt = pattern_stmt; | |
272 | } | |
ebfd146a IR |
273 | } |
274 | ||
275 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
276 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
277 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
278 | ||
279 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
280 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
281 | { | |
73fbfcad | 282 | if (dump_enabled_p ()) |
78c60e3d | 283 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 284 | "already marked relevant/live.\n"); |
ebfd146a IR |
285 | return; |
286 | } | |
287 | ||
9771b263 | 288 | worklist->safe_push (stmt); |
ebfd146a IR |
289 | } |
290 | ||
291 | ||
292 | /* Function vect_stmt_relevant_p. | |
293 | ||
294 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
295 | "relevant for vectorization". | |
296 | ||
297 | A stmt is considered "relevant for vectorization" if: | |
298 | - it has uses outside the loop. | |
299 | - it has vdefs (it alters memory). | |
300 | - control stmts in the loop (except for the exit condition). | |
301 | ||
302 | CHECKME: what other side effects would the vectorizer allow? */ | |
303 | ||
304 | static bool | |
305 | vect_stmt_relevant_p (gimple stmt, loop_vec_info loop_vinfo, | |
306 | enum vect_relevant *relevant, bool *live_p) | |
307 | { | |
308 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
309 | ssa_op_iter op_iter; | |
310 | imm_use_iterator imm_iter; | |
311 | use_operand_p use_p; | |
312 | def_operand_p def_p; | |
313 | ||
8644a673 | 314 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
315 | *live_p = false; |
316 | ||
317 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
318 | if (is_ctrl_stmt (stmt) |
319 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
320 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 321 | *relevant = vect_used_in_scope; |
ebfd146a IR |
322 | |
323 | /* changing memory. */ | |
324 | if (gimple_code (stmt) != GIMPLE_PHI) | |
5006671f | 325 | if (gimple_vdef (stmt)) |
ebfd146a | 326 | { |
73fbfcad | 327 | if (dump_enabled_p ()) |
78c60e3d | 328 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 329 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 330 | *relevant = vect_used_in_scope; |
ebfd146a IR |
331 | } |
332 | ||
333 | /* uses outside the loop. */ | |
334 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
335 | { | |
336 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
337 | { | |
338 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
339 | if (!flow_bb_inside_loop_p (loop, bb)) | |
340 | { | |
73fbfcad | 341 | if (dump_enabled_p ()) |
78c60e3d | 342 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 343 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 344 | |
3157b0c2 AO |
345 | if (is_gimple_debug (USE_STMT (use_p))) |
346 | continue; | |
347 | ||
ebfd146a IR |
348 | /* We expect all such uses to be in the loop exit phis |
349 | (because of loop closed form) */ | |
350 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
351 | gcc_assert (bb == single_exit (loop)->dest); | |
352 | ||
353 | *live_p = true; | |
354 | } | |
355 | } | |
356 | } | |
357 | ||
358 | return (*live_p || *relevant); | |
359 | } | |
360 | ||
361 | ||
b8698a0f | 362 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 363 | |
ff802fa1 | 364 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
365 | used in STMT for anything other than indexing an array. */ |
366 | ||
367 | static bool | |
368 | exist_non_indexing_operands_for_use_p (tree use, gimple stmt) | |
369 | { | |
370 | tree operand; | |
371 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 372 | |
ff802fa1 | 373 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
374 | reference in STMT, then any operand that corresponds to USE |
375 | is not indexing an array. */ | |
376 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
377 | return true; | |
59a05b0c | 378 | |
ebfd146a IR |
379 | /* STMT has a data_ref. FORNOW this means that its of one of |
380 | the following forms: | |
381 | -1- ARRAY_REF = var | |
382 | -2- var = ARRAY_REF | |
383 | (This should have been verified in analyze_data_refs). | |
384 | ||
385 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 386 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
387 | for array indexing. |
388 | ||
389 | Therefore, all we need to check is if STMT falls into the | |
390 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
391 | |
392 | if (!gimple_assign_copy_p (stmt)) | |
393 | return false; | |
59a05b0c EB |
394 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
395 | return false; | |
ebfd146a | 396 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
397 | if (TREE_CODE (operand) != SSA_NAME) |
398 | return false; | |
399 | ||
400 | if (operand == use) | |
401 | return true; | |
402 | ||
403 | return false; | |
404 | } | |
405 | ||
406 | ||
b8698a0f | 407 | /* |
ebfd146a IR |
408 | Function process_use. |
409 | ||
410 | Inputs: | |
411 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 412 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ff802fa1 | 413 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 414 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
415 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
416 | be performed. | |
ebfd146a IR |
417 | |
418 | Outputs: | |
419 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
420 | relevance info of the DEF_STMT of this USE: | |
421 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
422 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
423 | Exceptions: | |
424 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 425 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 426 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
427 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
428 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
429 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
430 | be modified accordingly. | |
431 | ||
432 | Return true if everything is as expected. Return false otherwise. */ | |
433 | ||
434 | static bool | |
b8698a0f | 435 | process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
9771b263 | 436 | enum vect_relevant relevant, vec<gimple> *worklist, |
aec7ae7d | 437 | bool force) |
ebfd146a IR |
438 | { |
439 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
440 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
441 | stmt_vec_info dstmt_vinfo; | |
442 | basic_block bb, def_bb; | |
443 | tree def; | |
444 | gimple def_stmt; | |
445 | enum vect_def_type dt; | |
446 | ||
b8698a0f | 447 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 448 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 449 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
450 | return true; |
451 | ||
24ee1384 | 452 | if (!vect_is_simple_use (use, stmt, loop_vinfo, NULL, &def_stmt, &def, &dt)) |
b8698a0f | 453 | { |
73fbfcad | 454 | if (dump_enabled_p ()) |
78c60e3d | 455 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 456 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
457 | return false; |
458 | } | |
459 | ||
460 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
461 | return true; | |
462 | ||
463 | def_bb = gimple_bb (def_stmt); | |
464 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
465 | { | |
73fbfcad | 466 | if (dump_enabled_p ()) |
e645e942 | 467 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
468 | return true; |
469 | } | |
470 | ||
b8698a0f L |
471 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
472 | DEF_STMT must have already been processed, because this should be the | |
473 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
474 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
475 | check that everything is as expected, and we are done. */ |
476 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
477 | bb = gimple_bb (stmt); | |
478 | if (gimple_code (stmt) == GIMPLE_PHI | |
479 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
480 | && gimple_code (def_stmt) != GIMPLE_PHI | |
481 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
482 | && bb->loop_father == def_bb->loop_father) | |
483 | { | |
73fbfcad | 484 | if (dump_enabled_p ()) |
78c60e3d | 485 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 486 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
487 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
488 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
489 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 490 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 491 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
492 | return true; |
493 | } | |
494 | ||
495 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
496 | outer-loop-header-bb: | |
497 | d = def_stmt | |
498 | inner-loop: | |
499 | stmt # use (d) | |
500 | outer-loop-tail-bb: | |
501 | ... */ | |
502 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
503 | { | |
73fbfcad | 504 | if (dump_enabled_p ()) |
78c60e3d | 505 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 506 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 507 | |
ebfd146a IR |
508 | switch (relevant) |
509 | { | |
8644a673 | 510 | case vect_unused_in_scope: |
7c5222ff IR |
511 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
512 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 513 | break; |
7c5222ff | 514 | |
ebfd146a | 515 | case vect_used_in_outer_by_reduction: |
7c5222ff | 516 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
517 | relevant = vect_used_by_reduction; |
518 | break; | |
7c5222ff | 519 | |
ebfd146a | 520 | case vect_used_in_outer: |
7c5222ff | 521 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 522 | relevant = vect_used_in_scope; |
ebfd146a | 523 | break; |
7c5222ff | 524 | |
8644a673 | 525 | case vect_used_in_scope: |
ebfd146a IR |
526 | break; |
527 | ||
528 | default: | |
529 | gcc_unreachable (); | |
b8698a0f | 530 | } |
ebfd146a IR |
531 | } |
532 | ||
533 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
534 | outer-loop-header-bb: | |
535 | ... | |
536 | inner-loop: | |
537 | d = def_stmt | |
06066f92 | 538 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
539 | stmt # use (d) */ |
540 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
541 | { | |
73fbfcad | 542 | if (dump_enabled_p ()) |
78c60e3d | 543 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 544 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 545 | |
ebfd146a IR |
546 | switch (relevant) |
547 | { | |
8644a673 | 548 | case vect_unused_in_scope: |
b8698a0f | 549 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 550 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 551 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
552 | break; |
553 | ||
ebfd146a IR |
554 | case vect_used_by_reduction: |
555 | relevant = vect_used_in_outer_by_reduction; | |
556 | break; | |
557 | ||
8644a673 | 558 | case vect_used_in_scope: |
ebfd146a IR |
559 | relevant = vect_used_in_outer; |
560 | break; | |
561 | ||
562 | default: | |
563 | gcc_unreachable (); | |
564 | } | |
565 | } | |
566 | ||
83197f37 IR |
567 | vect_mark_relevant (worklist, def_stmt, relevant, live_p, |
568 | is_pattern_stmt_p (stmt_vinfo)); | |
ebfd146a IR |
569 | return true; |
570 | } | |
571 | ||
572 | ||
573 | /* Function vect_mark_stmts_to_be_vectorized. | |
574 | ||
575 | Not all stmts in the loop need to be vectorized. For example: | |
576 | ||
577 | for i... | |
578 | for j... | |
579 | 1. T0 = i + j | |
580 | 2. T1 = a[T0] | |
581 | ||
582 | 3. j = j + 1 | |
583 | ||
584 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
585 | addressing of vectorized data-refs are handled differently. | |
586 | ||
587 | This pass detects such stmts. */ | |
588 | ||
589 | bool | |
590 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
591 | { | |
ebfd146a IR |
592 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
593 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
594 | unsigned int nbbs = loop->num_nodes; | |
595 | gimple_stmt_iterator si; | |
596 | gimple stmt; | |
597 | unsigned int i; | |
598 | stmt_vec_info stmt_vinfo; | |
599 | basic_block bb; | |
600 | gimple phi; | |
601 | bool live_p; | |
06066f92 IR |
602 | enum vect_relevant relevant, tmp_relevant; |
603 | enum vect_def_type def_type; | |
ebfd146a | 604 | |
73fbfcad | 605 | if (dump_enabled_p ()) |
78c60e3d | 606 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 607 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 608 | |
07687835 | 609 | stack_vec<gimple, 64> worklist; |
ebfd146a IR |
610 | |
611 | /* 1. Init worklist. */ | |
612 | for (i = 0; i < nbbs; i++) | |
613 | { | |
614 | bb = bbs[i]; | |
615 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 616 | { |
ebfd146a | 617 | phi = gsi_stmt (si); |
73fbfcad | 618 | if (dump_enabled_p ()) |
ebfd146a | 619 | { |
78c60e3d SS |
620 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
621 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
e645e942 | 622 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
623 | } |
624 | ||
625 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 626 | vect_mark_relevant (&worklist, phi, relevant, live_p, false); |
ebfd146a IR |
627 | } |
628 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
629 | { | |
630 | stmt = gsi_stmt (si); | |
73fbfcad | 631 | if (dump_enabled_p ()) |
ebfd146a | 632 | { |
78c60e3d SS |
633 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
634 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 635 | dump_printf (MSG_NOTE, "\n"); |
b8698a0f | 636 | } |
ebfd146a IR |
637 | |
638 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 639 | vect_mark_relevant (&worklist, stmt, relevant, live_p, false); |
ebfd146a IR |
640 | } |
641 | } | |
642 | ||
643 | /* 2. Process_worklist */ | |
9771b263 | 644 | while (worklist.length () > 0) |
ebfd146a IR |
645 | { |
646 | use_operand_p use_p; | |
647 | ssa_op_iter iter; | |
648 | ||
9771b263 | 649 | stmt = worklist.pop (); |
73fbfcad | 650 | if (dump_enabled_p ()) |
ebfd146a | 651 | { |
78c60e3d SS |
652 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
653 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 654 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
655 | } |
656 | ||
b8698a0f L |
657 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
658 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
659 | liveness and relevance properties of STMT. */ |
660 | stmt_vinfo = vinfo_for_stmt (stmt); | |
661 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
662 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
663 | ||
664 | /* Generally, the liveness and relevance properties of STMT are | |
665 | propagated as is to the DEF_STMTs of its USEs: | |
666 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
667 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
668 | ||
669 | One exception is when STMT has been identified as defining a reduction | |
670 | variable; in this case we set the liveness/relevance as follows: | |
671 | live_p = false | |
672 | relevant = vect_used_by_reduction | |
673 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f | 674 | those that are used by a reduction computation, and those that are |
ff802fa1 | 675 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 676 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 677 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 678 | |
06066f92 IR |
679 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
680 | tmp_relevant = relevant; | |
681 | switch (def_type) | |
ebfd146a | 682 | { |
06066f92 IR |
683 | case vect_reduction_def: |
684 | switch (tmp_relevant) | |
685 | { | |
686 | case vect_unused_in_scope: | |
687 | relevant = vect_used_by_reduction; | |
688 | break; | |
689 | ||
690 | case vect_used_by_reduction: | |
691 | if (gimple_code (stmt) == GIMPLE_PHI) | |
692 | break; | |
693 | /* fall through */ | |
694 | ||
695 | default: | |
73fbfcad | 696 | if (dump_enabled_p ()) |
78c60e3d | 697 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 698 | "unsupported use of reduction.\n"); |
06066f92 IR |
699 | return false; |
700 | } | |
701 | ||
b8698a0f | 702 | live_p = false; |
06066f92 | 703 | break; |
b8698a0f | 704 | |
06066f92 IR |
705 | case vect_nested_cycle: |
706 | if (tmp_relevant != vect_unused_in_scope | |
707 | && tmp_relevant != vect_used_in_outer_by_reduction | |
708 | && tmp_relevant != vect_used_in_outer) | |
709 | { | |
73fbfcad | 710 | if (dump_enabled_p ()) |
78c60e3d | 711 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 712 | "unsupported use of nested cycle.\n"); |
7c5222ff | 713 | |
06066f92 IR |
714 | return false; |
715 | } | |
7c5222ff | 716 | |
b8698a0f L |
717 | live_p = false; |
718 | break; | |
719 | ||
06066f92 IR |
720 | case vect_double_reduction_def: |
721 | if (tmp_relevant != vect_unused_in_scope | |
722 | && tmp_relevant != vect_used_by_reduction) | |
723 | { | |
73fbfcad | 724 | if (dump_enabled_p ()) |
78c60e3d | 725 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 726 | "unsupported use of double reduction.\n"); |
7c5222ff | 727 | |
7c5222ff | 728 | return false; |
06066f92 IR |
729 | } |
730 | ||
731 | live_p = false; | |
b8698a0f | 732 | break; |
7c5222ff | 733 | |
06066f92 IR |
734 | default: |
735 | break; | |
7c5222ff | 736 | } |
b8698a0f | 737 | |
aec7ae7d | 738 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
739 | { |
740 | /* Pattern statements are not inserted into the code, so | |
741 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
742 | have to scan the RHS or function arguments instead. */ | |
743 | if (is_gimple_assign (stmt)) | |
744 | { | |
69d2aade JJ |
745 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
746 | tree op = gimple_assign_rhs1 (stmt); | |
747 | ||
748 | i = 1; | |
749 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
750 | { | |
751 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
aec7ae7d | 752 | live_p, relevant, &worklist, false) |
69d2aade | 753 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
aec7ae7d | 754 | live_p, relevant, &worklist, false)) |
07687835 | 755 | return false; |
69d2aade JJ |
756 | i = 2; |
757 | } | |
758 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 759 | { |
69d2aade | 760 | op = gimple_op (stmt, i); |
9d5e7640 | 761 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, |
aec7ae7d | 762 | &worklist, false)) |
07687835 | 763 | return false; |
9d5e7640 IR |
764 | } |
765 | } | |
766 | else if (is_gimple_call (stmt)) | |
767 | { | |
768 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
769 | { | |
770 | tree arg = gimple_call_arg (stmt, i); | |
771 | if (!process_use (stmt, arg, loop_vinfo, live_p, relevant, | |
aec7ae7d | 772 | &worklist, false)) |
07687835 | 773 | return false; |
9d5e7640 IR |
774 | } |
775 | } | |
776 | } | |
777 | else | |
778 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
779 | { | |
780 | tree op = USE_FROM_PTR (use_p); | |
781 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, | |
aec7ae7d | 782 | &worklist, false)) |
07687835 | 783 | return false; |
9d5e7640 | 784 | } |
aec7ae7d JJ |
785 | |
786 | if (STMT_VINFO_GATHER_P (stmt_vinfo)) | |
787 | { | |
788 | tree off; | |
789 | tree decl = vect_check_gather (stmt, loop_vinfo, NULL, &off, NULL); | |
790 | gcc_assert (decl); | |
791 | if (!process_use (stmt, off, loop_vinfo, live_p, relevant, | |
792 | &worklist, true)) | |
07687835 | 793 | return false; |
aec7ae7d | 794 | } |
ebfd146a IR |
795 | } /* while worklist */ |
796 | ||
ebfd146a IR |
797 | return true; |
798 | } | |
799 | ||
800 | ||
b8698a0f | 801 | /* Function vect_model_simple_cost. |
ebfd146a | 802 | |
b8698a0f | 803 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
804 | single op. Right now, this does not account for multiple insns that could |
805 | be generated for the single vector op. We will handle that shortly. */ | |
806 | ||
807 | void | |
b8698a0f | 808 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 BS |
809 | enum vect_def_type *dt, |
810 | stmt_vector_for_cost *prologue_cost_vec, | |
811 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
812 | { |
813 | int i; | |
92345349 | 814 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
815 | |
816 | /* The SLP costs were already calculated during SLP tree build. */ | |
817 | if (PURE_SLP_STMT (stmt_info)) | |
818 | return; | |
819 | ||
ebfd146a IR |
820 | /* FORNOW: Assuming maximum 2 args per stmts. */ |
821 | for (i = 0; i < 2; i++) | |
92345349 BS |
822 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
823 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, vector_stmt, | |
824 | stmt_info, 0, vect_prologue); | |
c3e7ee41 BS |
825 | |
826 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
827 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
828 | stmt_info, 0, vect_body); | |
c3e7ee41 | 829 | |
73fbfcad | 830 | if (dump_enabled_p ()) |
78c60e3d SS |
831 | dump_printf_loc (MSG_NOTE, vect_location, |
832 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 833 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
834 | } |
835 | ||
836 | ||
8bd37302 BS |
837 | /* Model cost for type demotion and promotion operations. PWR is normally |
838 | zero for single-step promotions and demotions. It will be one if | |
839 | two-step promotion/demotion is required, and so on. Each additional | |
840 | step doubles the number of instructions required. */ | |
841 | ||
842 | static void | |
843 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
844 | enum vect_def_type *dt, int pwr) | |
845 | { | |
846 | int i, tmp; | |
92345349 | 847 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
848 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
849 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
850 | void *target_cost_data; | |
8bd37302 BS |
851 | |
852 | /* The SLP costs were already calculated during SLP tree build. */ | |
853 | if (PURE_SLP_STMT (stmt_info)) | |
854 | return; | |
855 | ||
c3e7ee41 BS |
856 | if (loop_vinfo) |
857 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
858 | else | |
859 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
860 | ||
8bd37302 BS |
861 | for (i = 0; i < pwr + 1; i++) |
862 | { | |
863 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
864 | (i + 1) : i; | |
c3e7ee41 | 865 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
866 | vec_promote_demote, stmt_info, 0, |
867 | vect_body); | |
8bd37302 BS |
868 | } |
869 | ||
870 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
871 | for (i = 0; i < 2; i++) | |
92345349 BS |
872 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
873 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
874 | stmt_info, 0, vect_prologue); | |
8bd37302 | 875 | |
73fbfcad | 876 | if (dump_enabled_p ()) |
78c60e3d SS |
877 | dump_printf_loc (MSG_NOTE, vect_location, |
878 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 879 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
880 | } |
881 | ||
0d0293ac | 882 | /* Function vect_cost_group_size |
b8698a0f | 883 | |
0d0293ac | 884 | For grouped load or store, return the group_size only if it is the first |
ebfd146a IR |
885 | load or store of a group, else return 1. This ensures that group size is |
886 | only returned once per group. */ | |
887 | ||
888 | static int | |
0d0293ac | 889 | vect_cost_group_size (stmt_vec_info stmt_info) |
ebfd146a | 890 | { |
e14c1050 | 891 | gimple first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
892 | |
893 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
e14c1050 | 894 | return GROUP_SIZE (stmt_info); |
ebfd146a IR |
895 | |
896 | return 1; | |
897 | } | |
898 | ||
899 | ||
900 | /* Function vect_model_store_cost | |
901 | ||
0d0293ac MM |
902 | Models cost for stores. In the case of grouped accesses, one access |
903 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
904 | |
905 | void | |
b8698a0f | 906 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
272c6793 | 907 | bool store_lanes_p, enum vect_def_type dt, |
92345349 BS |
908 | slp_tree slp_node, |
909 | stmt_vector_for_cost *prologue_cost_vec, | |
910 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
911 | { |
912 | int group_size; | |
92345349 | 913 | unsigned int inside_cost = 0, prologue_cost = 0; |
720f5239 IR |
914 | struct data_reference *first_dr; |
915 | gimple first_stmt; | |
ebfd146a IR |
916 | |
917 | /* The SLP costs were already calculated during SLP tree build. */ | |
918 | if (PURE_SLP_STMT (stmt_info)) | |
919 | return; | |
920 | ||
8644a673 | 921 | if (dt == vect_constant_def || dt == vect_external_def) |
92345349 BS |
922 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
923 | stmt_info, 0, vect_prologue); | |
ebfd146a | 924 | |
0d0293ac MM |
925 | /* Grouped access? */ |
926 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
720f5239 IR |
927 | { |
928 | if (slp_node) | |
929 | { | |
9771b263 | 930 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
720f5239 IR |
931 | group_size = 1; |
932 | } | |
933 | else | |
934 | { | |
e14c1050 | 935 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 936 | group_size = vect_cost_group_size (stmt_info); |
720f5239 IR |
937 | } |
938 | ||
939 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
940 | } | |
0d0293ac | 941 | /* Not a grouped access. */ |
ebfd146a | 942 | else |
720f5239 IR |
943 | { |
944 | group_size = 1; | |
945 | first_dr = STMT_VINFO_DATA_REF (stmt_info); | |
946 | } | |
ebfd146a | 947 | |
272c6793 | 948 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 949 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 RS |
950 | access is instead being provided by a permute-and-store operation, |
951 | include the cost of the permutes. */ | |
952 | if (!store_lanes_p && group_size > 1) | |
ebfd146a IR |
953 | { |
954 | /* Uses a high and low interleave operation for each needed permute. */ | |
c3e7ee41 BS |
955 | |
956 | int nstmts = ncopies * exact_log2 (group_size) * group_size; | |
92345349 BS |
957 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
958 | stmt_info, 0, vect_body); | |
ebfd146a | 959 | |
73fbfcad | 960 | if (dump_enabled_p ()) |
78c60e3d | 961 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 962 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 963 | group_size); |
ebfd146a IR |
964 | } |
965 | ||
966 | /* Costs of the stores. */ | |
92345349 | 967 | vect_get_store_cost (first_dr, ncopies, &inside_cost, body_cost_vec); |
ebfd146a | 968 | |
73fbfcad | 969 | if (dump_enabled_p ()) |
78c60e3d SS |
970 | dump_printf_loc (MSG_NOTE, vect_location, |
971 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 972 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
973 | } |
974 | ||
975 | ||
720f5239 IR |
976 | /* Calculate cost of DR's memory access. */ |
977 | void | |
978 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 979 | unsigned int *inside_cost, |
92345349 | 980 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
981 | { |
982 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
c3e7ee41 BS |
983 | gimple stmt = DR_STMT (dr); |
984 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
720f5239 IR |
985 | |
986 | switch (alignment_support_scheme) | |
987 | { | |
988 | case dr_aligned: | |
989 | { | |
92345349 BS |
990 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
991 | vector_store, stmt_info, 0, | |
992 | vect_body); | |
720f5239 | 993 | |
73fbfcad | 994 | if (dump_enabled_p ()) |
78c60e3d | 995 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 996 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
997 | break; |
998 | } | |
999 | ||
1000 | case dr_unaligned_supported: | |
1001 | { | |
720f5239 | 1002 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1003 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1004 | unaligned_store, stmt_info, |
92345349 | 1005 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1006 | if (dump_enabled_p ()) |
78c60e3d SS |
1007 | dump_printf_loc (MSG_NOTE, vect_location, |
1008 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1009 | "hardware.\n"); |
720f5239 IR |
1010 | break; |
1011 | } | |
1012 | ||
38eec4c6 UW |
1013 | case dr_unaligned_unsupported: |
1014 | { | |
1015 | *inside_cost = VECT_MAX_COST; | |
1016 | ||
73fbfcad | 1017 | if (dump_enabled_p ()) |
78c60e3d | 1018 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1019 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1020 | break; |
1021 | } | |
1022 | ||
720f5239 IR |
1023 | default: |
1024 | gcc_unreachable (); | |
1025 | } | |
1026 | } | |
1027 | ||
1028 | ||
ebfd146a IR |
1029 | /* Function vect_model_load_cost |
1030 | ||
0d0293ac MM |
1031 | Models cost for loads. In the case of grouped accesses, the last access |
1032 | has the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1033 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1034 | access scheme chosen. */ |
1035 | ||
1036 | void | |
92345349 BS |
1037 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
1038 | bool load_lanes_p, slp_tree slp_node, | |
1039 | stmt_vector_for_cost *prologue_cost_vec, | |
1040 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
1041 | { |
1042 | int group_size; | |
ebfd146a IR |
1043 | gimple first_stmt; |
1044 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; | |
92345349 | 1045 | unsigned int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
1046 | |
1047 | /* The SLP costs were already calculated during SLP tree build. */ | |
1048 | if (PURE_SLP_STMT (stmt_info)) | |
1049 | return; | |
1050 | ||
0d0293ac | 1051 | /* Grouped accesses? */ |
e14c1050 | 1052 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 1053 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && first_stmt && !slp_node) |
ebfd146a | 1054 | { |
0d0293ac | 1055 | group_size = vect_cost_group_size (stmt_info); |
ebfd146a IR |
1056 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
1057 | } | |
0d0293ac | 1058 | /* Not a grouped access. */ |
ebfd146a IR |
1059 | else |
1060 | { | |
1061 | group_size = 1; | |
1062 | first_dr = dr; | |
1063 | } | |
1064 | ||
272c6793 | 1065 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1066 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 RS |
1067 | access is instead being provided by a load-and-permute operation, |
1068 | include the cost of the permutes. */ | |
1069 | if (!load_lanes_p && group_size > 1) | |
ebfd146a IR |
1070 | { |
1071 | /* Uses an even and odd extract operations for each needed permute. */ | |
c3e7ee41 | 1072 | int nstmts = ncopies * exact_log2 (group_size) * group_size; |
92345349 BS |
1073 | inside_cost += record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
1074 | stmt_info, 0, vect_body); | |
ebfd146a | 1075 | |
73fbfcad | 1076 | if (dump_enabled_p ()) |
e645e942 TJ |
1077 | dump_printf_loc (MSG_NOTE, vect_location, |
1078 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1079 | group_size); |
ebfd146a IR |
1080 | } |
1081 | ||
1082 | /* The loads themselves. */ | |
a82960aa RG |
1083 | if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
1084 | { | |
a21892ad BS |
1085 | /* N scalar loads plus gathering them into a vector. */ |
1086 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
92345349 | 1087 | inside_cost += record_stmt_cost (body_cost_vec, |
c3e7ee41 | 1088 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), |
92345349 BS |
1089 | scalar_load, stmt_info, 0, vect_body); |
1090 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, | |
1091 | stmt_info, 0, vect_body); | |
a82960aa RG |
1092 | } |
1093 | else | |
1094 | vect_get_load_cost (first_dr, ncopies, | |
1095 | ((!STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1096 | || group_size > 1 || slp_node), | |
92345349 BS |
1097 | &inside_cost, &prologue_cost, |
1098 | prologue_cost_vec, body_cost_vec, true); | |
720f5239 | 1099 | |
73fbfcad | 1100 | if (dump_enabled_p ()) |
78c60e3d SS |
1101 | dump_printf_loc (MSG_NOTE, vect_location, |
1102 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1103 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1104 | } |
1105 | ||
1106 | ||
1107 | /* Calculate cost of DR's memory access. */ | |
1108 | void | |
1109 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1110 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1111 | unsigned int *prologue_cost, |
1112 | stmt_vector_for_cost *prologue_cost_vec, | |
1113 | stmt_vector_for_cost *body_cost_vec, | |
1114 | bool record_prologue_costs) | |
720f5239 IR |
1115 | { |
1116 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
c3e7ee41 BS |
1117 | gimple stmt = DR_STMT (dr); |
1118 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
720f5239 IR |
1119 | |
1120 | switch (alignment_support_scheme) | |
ebfd146a IR |
1121 | { |
1122 | case dr_aligned: | |
1123 | { | |
92345349 BS |
1124 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1125 | stmt_info, 0, vect_body); | |
ebfd146a | 1126 | |
73fbfcad | 1127 | if (dump_enabled_p ()) |
78c60e3d | 1128 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1129 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1130 | |
1131 | break; | |
1132 | } | |
1133 | case dr_unaligned_supported: | |
1134 | { | |
720f5239 | 1135 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1136 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1137 | unaligned_load, stmt_info, |
92345349 | 1138 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1139 | |
73fbfcad | 1140 | if (dump_enabled_p ()) |
78c60e3d SS |
1141 | dump_printf_loc (MSG_NOTE, vect_location, |
1142 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1143 | "hardware.\n"); |
ebfd146a IR |
1144 | |
1145 | break; | |
1146 | } | |
1147 | case dr_explicit_realign: | |
1148 | { | |
92345349 BS |
1149 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1150 | vector_load, stmt_info, 0, vect_body); | |
1151 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1152 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1153 | |
1154 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1155 | the containing loop, the following cost should be added to the | |
92345349 | 1156 | prologue costs. */ |
ebfd146a | 1157 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1158 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1159 | stmt_info, 0, vect_body); | |
ebfd146a | 1160 | |
73fbfcad | 1161 | if (dump_enabled_p ()) |
e645e942 TJ |
1162 | dump_printf_loc (MSG_NOTE, vect_location, |
1163 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1164 | |
ebfd146a IR |
1165 | break; |
1166 | } | |
1167 | case dr_explicit_realign_optimized: | |
1168 | { | |
73fbfcad | 1169 | if (dump_enabled_p ()) |
e645e942 | 1170 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1171 | "vect_model_load_cost: unaligned software " |
e645e942 | 1172 | "pipelined.\n"); |
ebfd146a IR |
1173 | |
1174 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1175 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1176 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1177 | access, then the above cost should only be considered for one |
ff802fa1 | 1178 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1179 | and a realignment op. */ |
1180 | ||
92345349 | 1181 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1182 | { |
92345349 BS |
1183 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1184 | vector_stmt, stmt_info, | |
1185 | 0, vect_prologue); | |
ebfd146a | 1186 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1187 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1188 | vector_stmt, stmt_info, | |
1189 | 0, vect_prologue); | |
ebfd146a IR |
1190 | } |
1191 | ||
92345349 BS |
1192 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1193 | stmt_info, 0, vect_body); | |
1194 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1195 | stmt_info, 0, vect_body); | |
8bd37302 | 1196 | |
73fbfcad | 1197 | if (dump_enabled_p ()) |
78c60e3d | 1198 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1199 | "vect_model_load_cost: explicit realign optimized" |
1200 | "\n"); | |
8bd37302 | 1201 | |
ebfd146a IR |
1202 | break; |
1203 | } | |
1204 | ||
38eec4c6 UW |
1205 | case dr_unaligned_unsupported: |
1206 | { | |
1207 | *inside_cost = VECT_MAX_COST; | |
1208 | ||
73fbfcad | 1209 | if (dump_enabled_p ()) |
78c60e3d | 1210 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1211 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1212 | break; |
1213 | } | |
1214 | ||
ebfd146a IR |
1215 | default: |
1216 | gcc_unreachable (); | |
1217 | } | |
ebfd146a IR |
1218 | } |
1219 | ||
418b7df3 RG |
1220 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1221 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1222 | |
418b7df3 RG |
1223 | static void |
1224 | vect_init_vector_1 (gimple stmt, gimple new_stmt, gimple_stmt_iterator *gsi) | |
ebfd146a | 1225 | { |
ebfd146a | 1226 | if (gsi) |
418b7df3 | 1227 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1228 | else |
1229 | { | |
418b7df3 | 1230 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1231 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1232 | |
a70d6342 IR |
1233 | if (loop_vinfo) |
1234 | { | |
1235 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1236 | basic_block new_bb; |
1237 | edge pe; | |
a70d6342 IR |
1238 | |
1239 | if (nested_in_vect_loop_p (loop, stmt)) | |
1240 | loop = loop->inner; | |
b8698a0f | 1241 | |
a70d6342 | 1242 | pe = loop_preheader_edge (loop); |
418b7df3 | 1243 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1244 | gcc_assert (!new_bb); |
1245 | } | |
1246 | else | |
1247 | { | |
1248 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1249 | basic_block bb; | |
1250 | gimple_stmt_iterator gsi_bb_start; | |
1251 | ||
1252 | gcc_assert (bb_vinfo); | |
1253 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1254 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1255 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1256 | } |
ebfd146a IR |
1257 | } |
1258 | ||
73fbfcad | 1259 | if (dump_enabled_p ()) |
ebfd146a | 1260 | { |
78c60e3d SS |
1261 | dump_printf_loc (MSG_NOTE, vect_location, |
1262 | "created new init_stmt: "); | |
1263 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
e645e942 | 1264 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 1265 | } |
418b7df3 RG |
1266 | } |
1267 | ||
1268 | /* Function vect_init_vector. | |
ebfd146a | 1269 | |
5467ee52 RG |
1270 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1271 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1272 | vector type a vector with all elements equal to VAL is created first. | |
1273 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1274 | initialization at the loop preheader. | |
418b7df3 RG |
1275 | Return the DEF of INIT_STMT. |
1276 | It will be used in the vectorization of STMT. */ | |
1277 | ||
1278 | tree | |
5467ee52 | 1279 | vect_init_vector (gimple stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 RG |
1280 | { |
1281 | tree new_var; | |
1282 | gimple init_stmt; | |
1283 | tree vec_oprnd; | |
1284 | tree new_temp; | |
1285 | ||
5467ee52 RG |
1286 | if (TREE_CODE (type) == VECTOR_TYPE |
1287 | && TREE_CODE (TREE_TYPE (val)) != VECTOR_TYPE) | |
418b7df3 | 1288 | { |
5467ee52 | 1289 | if (!types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) |
418b7df3 | 1290 | { |
5467ee52 RG |
1291 | if (CONSTANT_CLASS_P (val)) |
1292 | val = fold_unary (VIEW_CONVERT_EXPR, TREE_TYPE (type), val); | |
418b7df3 RG |
1293 | else |
1294 | { | |
83d5977e | 1295 | new_temp = make_ssa_name (TREE_TYPE (type), NULL); |
418b7df3 | 1296 | init_stmt = gimple_build_assign_with_ops (NOP_EXPR, |
83d5977e | 1297 | new_temp, val, |
418b7df3 | 1298 | NULL_TREE); |
418b7df3 | 1299 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1300 | val = new_temp; |
418b7df3 RG |
1301 | } |
1302 | } | |
5467ee52 | 1303 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1304 | } |
1305 | ||
5467ee52 | 1306 | new_var = vect_get_new_vect_var (type, vect_simple_var, "cst_"); |
5467ee52 | 1307 | init_stmt = gimple_build_assign (new_var, val); |
418b7df3 RG |
1308 | new_temp = make_ssa_name (new_var, init_stmt); |
1309 | gimple_assign_set_lhs (init_stmt, new_temp); | |
1310 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
ebfd146a IR |
1311 | vec_oprnd = gimple_assign_lhs (init_stmt); |
1312 | return vec_oprnd; | |
1313 | } | |
1314 | ||
a70d6342 | 1315 | |
ebfd146a IR |
1316 | /* Function vect_get_vec_def_for_operand. |
1317 | ||
ff802fa1 | 1318 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
1319 | used in the vectorized stmt for STMT. |
1320 | ||
1321 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1322 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1323 | ||
1324 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1325 | needs to be introduced. */ | |
1326 | ||
1327 | tree | |
1328 | vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def) | |
1329 | { | |
1330 | tree vec_oprnd; | |
1331 | gimple vec_stmt; | |
1332 | gimple def_stmt; | |
1333 | stmt_vec_info def_stmt_info = NULL; | |
1334 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
9dc3f7de | 1335 | unsigned int nunits; |
ebfd146a | 1336 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
ebfd146a | 1337 | tree def; |
ebfd146a IR |
1338 | enum vect_def_type dt; |
1339 | bool is_simple_use; | |
1340 | tree vector_type; | |
1341 | ||
73fbfcad | 1342 | if (dump_enabled_p ()) |
ebfd146a | 1343 | { |
78c60e3d SS |
1344 | dump_printf_loc (MSG_NOTE, vect_location, |
1345 | "vect_get_vec_def_for_operand: "); | |
1346 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
e645e942 | 1347 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1348 | } |
1349 | ||
24ee1384 IR |
1350 | is_simple_use = vect_is_simple_use (op, stmt, loop_vinfo, NULL, |
1351 | &def_stmt, &def, &dt); | |
ebfd146a | 1352 | gcc_assert (is_simple_use); |
73fbfcad | 1353 | if (dump_enabled_p ()) |
ebfd146a | 1354 | { |
78c60e3d | 1355 | int loc_printed = 0; |
ebfd146a IR |
1356 | if (def) |
1357 | { | |
78c60e3d SS |
1358 | dump_printf_loc (MSG_NOTE, vect_location, "def = "); |
1359 | loc_printed = 1; | |
1360 | dump_generic_expr (MSG_NOTE, TDF_SLIM, def); | |
e645e942 | 1361 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1362 | } |
1363 | if (def_stmt) | |
1364 | { | |
78c60e3d SS |
1365 | if (loc_printed) |
1366 | dump_printf (MSG_NOTE, " def_stmt = "); | |
1367 | else | |
1368 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1369 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
e645e942 | 1370 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1371 | } |
1372 | } | |
1373 | ||
1374 | switch (dt) | |
1375 | { | |
1376 | /* Case 1: operand is a constant. */ | |
1377 | case vect_constant_def: | |
1378 | { | |
7569a6cc RG |
1379 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
1380 | gcc_assert (vector_type); | |
9dc3f7de | 1381 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); |
7569a6cc | 1382 | |
b8698a0f | 1383 | if (scalar_def) |
ebfd146a IR |
1384 | *scalar_def = op; |
1385 | ||
1386 | /* Create 'vect_cst_ = {cst,cst,...,cst}' */ | |
73fbfcad | 1387 | if (dump_enabled_p ()) |
78c60e3d | 1388 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1389 | "Create vector_cst. nunits = %d\n", nunits); |
ebfd146a | 1390 | |
418b7df3 | 1391 | return vect_init_vector (stmt, op, vector_type, NULL); |
ebfd146a IR |
1392 | } |
1393 | ||
1394 | /* Case 2: operand is defined outside the loop - loop invariant. */ | |
8644a673 | 1395 | case vect_external_def: |
ebfd146a IR |
1396 | { |
1397 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); | |
1398 | gcc_assert (vector_type); | |
ebfd146a | 1399 | |
b8698a0f | 1400 | if (scalar_def) |
ebfd146a IR |
1401 | *scalar_def = def; |
1402 | ||
1403 | /* Create 'vec_inv = {inv,inv,..,inv}' */ | |
73fbfcad | 1404 | if (dump_enabled_p ()) |
e645e942 | 1405 | dump_printf_loc (MSG_NOTE, vect_location, "Create vector_inv.\n"); |
ebfd146a | 1406 | |
418b7df3 | 1407 | return vect_init_vector (stmt, def, vector_type, NULL); |
ebfd146a IR |
1408 | } |
1409 | ||
1410 | /* Case 3: operand is defined inside the loop. */ | |
8644a673 | 1411 | case vect_internal_def: |
ebfd146a | 1412 | { |
b8698a0f | 1413 | if (scalar_def) |
ebfd146a IR |
1414 | *scalar_def = NULL/* FIXME tuples: def_stmt*/; |
1415 | ||
1416 | /* Get the def from the vectorized stmt. */ | |
1417 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1418 | |
ebfd146a | 1419 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1420 | /* Get vectorized pattern statement. */ |
1421 | if (!vec_stmt | |
1422 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1423 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1424 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1425 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1426 | gcc_assert (vec_stmt); |
1427 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1428 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1429 | else if (is_gimple_call (vec_stmt)) | |
1430 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1431 | else | |
1432 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1433 | return vec_oprnd; | |
1434 | } | |
1435 | ||
1436 | /* Case 4: operand is defined by a loop header phi - reduction */ | |
1437 | case vect_reduction_def: | |
06066f92 | 1438 | case vect_double_reduction_def: |
7c5222ff | 1439 | case vect_nested_cycle: |
ebfd146a IR |
1440 | { |
1441 | struct loop *loop; | |
1442 | ||
1443 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
b8698a0f | 1444 | loop = (gimple_bb (def_stmt))->loop_father; |
ebfd146a IR |
1445 | |
1446 | /* Get the def before the loop */ | |
1447 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); | |
1448 | return get_initial_def_for_reduction (stmt, op, scalar_def); | |
1449 | } | |
1450 | ||
1451 | /* Case 5: operand is defined by loop-header phi - induction. */ | |
1452 | case vect_induction_def: | |
1453 | { | |
1454 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1455 | ||
1456 | /* Get the def from the vectorized stmt. */ | |
1457 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1458 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1459 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1460 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1461 | else | |
1462 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1463 | return vec_oprnd; |
1464 | } | |
1465 | ||
1466 | default: | |
1467 | gcc_unreachable (); | |
1468 | } | |
1469 | } | |
1470 | ||
1471 | ||
1472 | /* Function vect_get_vec_def_for_stmt_copy | |
1473 | ||
ff802fa1 | 1474 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1475 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1476 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1477 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1478 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1479 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1480 | DT is the type of the vector def VEC_OPRND. |
1481 | ||
1482 | Context: | |
1483 | In case the vectorization factor (VF) is bigger than the number | |
1484 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1485 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1486 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1487 | smallest data-type determines the VF, and as a result, when vectorizing |
1488 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1489 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1490 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1491 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1492 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1493 | ||
1494 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1495 | |
ebfd146a IR |
1496 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1497 | VS1.1: vx.1 = memref1 VS1.2 | |
1498 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1499 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1500 | |
1501 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1502 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1503 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1504 | VSnew.3: vz3 = vx.3 + ... | |
1505 | ||
1506 | The vectorization of S1 is explained in vectorizable_load. | |
1507 | The vectorization of S2: | |
b8698a0f L |
1508 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1509 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1510 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1511 | returns the vector-def 'vx.0'. |
1512 | ||
b8698a0f L |
1513 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1514 | function is called to get the relevant vector-def for each operand. It is | |
1515 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1516 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1517 | ||
b8698a0f L |
1518 | For example, to obtain the vector-def 'vx.1' in order to create the |
1519 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1520 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1521 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1522 | and return its def ('vx.1'). | |
1523 | Overall, to create the above sequence this function will be called 3 times: | |
1524 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1525 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1526 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1527 | ||
1528 | tree | |
1529 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1530 | { | |
1531 | gimple vec_stmt_for_operand; | |
1532 | stmt_vec_info def_stmt_info; | |
1533 | ||
1534 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1535 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1536 | return vec_oprnd; |
1537 | ||
1538 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1539 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1540 | gcc_assert (def_stmt_info); | |
1541 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1542 | gcc_assert (vec_stmt_for_operand); | |
1543 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1544 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) | |
1545 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1546 | else | |
1547 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1548 | return vec_oprnd; | |
1549 | } | |
1550 | ||
1551 | ||
1552 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1553 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1554 | |
1555 | static void | |
b8698a0f | 1556 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1557 | vec<tree> *vec_oprnds0, |
1558 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1559 | { |
9771b263 | 1560 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1561 | |
1562 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1563 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1564 | |
9771b263 | 1565 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1566 | { |
9771b263 | 1567 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1568 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1569 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1570 | } |
1571 | } | |
1572 | ||
1573 | ||
d092494c IR |
1574 | /* Get vectorized definitions for OP0 and OP1. |
1575 | REDUC_INDEX is the index of reduction operand in case of reduction, | |
1576 | and -1 otherwise. */ | |
ebfd146a | 1577 | |
d092494c | 1578 | void |
ebfd146a | 1579 | vect_get_vec_defs (tree op0, tree op1, gimple stmt, |
9771b263 DN |
1580 | vec<tree> *vec_oprnds0, |
1581 | vec<tree> *vec_oprnds1, | |
d092494c | 1582 | slp_tree slp_node, int reduc_index) |
ebfd146a IR |
1583 | { |
1584 | if (slp_node) | |
d092494c IR |
1585 | { |
1586 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1587 | auto_vec<tree> ops (nops); |
1588 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1589 | |
9771b263 | 1590 | ops.quick_push (op0); |
d092494c | 1591 | if (op1) |
9771b263 | 1592 | ops.quick_push (op1); |
d092494c IR |
1593 | |
1594 | vect_get_slp_defs (ops, slp_node, &vec_defs, reduc_index); | |
1595 | ||
37b5ec8f | 1596 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1597 | if (op1) |
37b5ec8f | 1598 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1599 | } |
ebfd146a IR |
1600 | else |
1601 | { | |
1602 | tree vec_oprnd; | |
1603 | ||
9771b263 | 1604 | vec_oprnds0->create (1); |
b8698a0f | 1605 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL); |
9771b263 | 1606 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1607 | |
1608 | if (op1) | |
1609 | { | |
9771b263 | 1610 | vec_oprnds1->create (1); |
b8698a0f | 1611 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL); |
9771b263 | 1612 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1613 | } |
1614 | } | |
1615 | } | |
1616 | ||
1617 | ||
1618 | /* Function vect_finish_stmt_generation. | |
1619 | ||
1620 | Insert a new stmt. */ | |
1621 | ||
1622 | void | |
1623 | vect_finish_stmt_generation (gimple stmt, gimple vec_stmt, | |
1624 | gimple_stmt_iterator *gsi) | |
1625 | { | |
1626 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1627 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1628 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1629 | |
1630 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1631 | ||
54e8e2c3 RG |
1632 | if (!gsi_end_p (*gsi) |
1633 | && gimple_has_mem_ops (vec_stmt)) | |
1634 | { | |
1635 | gimple at_stmt = gsi_stmt (*gsi); | |
1636 | tree vuse = gimple_vuse (at_stmt); | |
1637 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1638 | { | |
1639 | tree vdef = gimple_vdef (at_stmt); | |
1640 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1641 | /* If we have an SSA vuse and insert a store, update virtual | |
1642 | SSA form to avoid triggering the renamer. Do so only | |
1643 | if we can easily see all uses - which is what almost always | |
1644 | happens with the way vectorized stmts are inserted. */ | |
1645 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1646 | && ((is_gimple_assign (vec_stmt) | |
1647 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1648 | || (is_gimple_call (vec_stmt) | |
1649 | && !(gimple_call_flags (vec_stmt) | |
1650 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1651 | { | |
1652 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1653 | gimple_set_vdef (vec_stmt, new_vdef); | |
1654 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1655 | } | |
1656 | } | |
1657 | } | |
ebfd146a IR |
1658 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1659 | ||
b8698a0f | 1660 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo, |
a70d6342 | 1661 | bb_vinfo)); |
ebfd146a | 1662 | |
73fbfcad | 1663 | if (dump_enabled_p ()) |
ebfd146a | 1664 | { |
78c60e3d SS |
1665 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1666 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
e645e942 | 1667 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1668 | } |
1669 | ||
ad885386 | 1670 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
ebfd146a IR |
1671 | } |
1672 | ||
1673 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1674 | a function declaration if the target has a vectorized version | |
1675 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1676 | ||
1677 | tree | |
1678 | vectorizable_function (gimple call, tree vectype_out, tree vectype_in) | |
1679 | { | |
1680 | tree fndecl = gimple_call_fndecl (call); | |
ebfd146a IR |
1681 | |
1682 | /* We only handle functions that do not read or clobber memory -- i.e. | |
1683 | const or novops ones. */ | |
1684 | if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS))) | |
1685 | return NULL_TREE; | |
1686 | ||
1687 | if (!fndecl | |
1688 | || TREE_CODE (fndecl) != FUNCTION_DECL | |
1689 | || !DECL_BUILT_IN (fndecl)) | |
1690 | return NULL_TREE; | |
1691 | ||
62f7fd21 | 1692 | return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out, |
ebfd146a IR |
1693 | vectype_in); |
1694 | } | |
1695 | ||
1696 | /* Function vectorizable_call. | |
1697 | ||
b8698a0f L |
1698 | Check if STMT performs a function call that can be vectorized. |
1699 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1700 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1701 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1702 | ||
1703 | static bool | |
190c2236 JJ |
1704 | vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, |
1705 | slp_tree slp_node) | |
ebfd146a IR |
1706 | { |
1707 | tree vec_dest; | |
1708 | tree scalar_dest; | |
1709 | tree op, type; | |
1710 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1711 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
1712 | tree vectype_out, vectype_in; | |
1713 | int nunits_in; | |
1714 | int nunits_out; | |
1715 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 1716 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b690cc0f | 1717 | tree fndecl, new_temp, def, rhs_type; |
ebfd146a | 1718 | gimple def_stmt; |
0502fb85 UB |
1719 | enum vect_def_type dt[3] |
1720 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
63827fb8 | 1721 | gimple new_stmt = NULL; |
ebfd146a | 1722 | int ncopies, j; |
6e1aa848 | 1723 | vec<tree> vargs = vNULL; |
ebfd146a IR |
1724 | enum { NARROW, NONE, WIDEN } modifier; |
1725 | size_t i, nargs; | |
9d5e7640 | 1726 | tree lhs; |
ebfd146a | 1727 | |
190c2236 | 1728 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
1729 | return false; |
1730 | ||
8644a673 | 1731 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1732 | return false; |
1733 | ||
ebfd146a IR |
1734 | /* Is STMT a vectorizable call? */ |
1735 | if (!is_gimple_call (stmt)) | |
1736 | return false; | |
1737 | ||
1738 | if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
1739 | return false; | |
1740 | ||
822ba6d7 | 1741 | if (stmt_can_throw_internal (stmt)) |
5a2c1986 IR |
1742 | return false; |
1743 | ||
b690cc0f RG |
1744 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
1745 | ||
ebfd146a IR |
1746 | /* Process function arguments. */ |
1747 | rhs_type = NULL_TREE; | |
b690cc0f | 1748 | vectype_in = NULL_TREE; |
ebfd146a IR |
1749 | nargs = gimple_call_num_args (stmt); |
1750 | ||
1b1562a5 MM |
1751 | /* Bail out if the function has more than three arguments, we do not have |
1752 | interesting builtin functions to vectorize with more than two arguments | |
1753 | except for fma. No arguments is also not good. */ | |
1754 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
1755 | return false; |
1756 | ||
74bf76ed JJ |
1757 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
1758 | if (gimple_call_internal_p (stmt) | |
1759 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
1760 | { | |
1761 | nargs = 0; | |
1762 | rhs_type = unsigned_type_node; | |
1763 | } | |
1764 | ||
ebfd146a IR |
1765 | for (i = 0; i < nargs; i++) |
1766 | { | |
b690cc0f RG |
1767 | tree opvectype; |
1768 | ||
ebfd146a IR |
1769 | op = gimple_call_arg (stmt, i); |
1770 | ||
1771 | /* We can only handle calls with arguments of the same type. */ | |
1772 | if (rhs_type | |
8533c9d8 | 1773 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 1774 | { |
73fbfcad | 1775 | if (dump_enabled_p ()) |
78c60e3d | 1776 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1777 | "argument types differ.\n"); |
ebfd146a IR |
1778 | return false; |
1779 | } | |
b690cc0f RG |
1780 | if (!rhs_type) |
1781 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 1782 | |
24ee1384 | 1783 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, |
b690cc0f | 1784 | &def_stmt, &def, &dt[i], &opvectype)) |
ebfd146a | 1785 | { |
73fbfcad | 1786 | if (dump_enabled_p ()) |
78c60e3d | 1787 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1788 | "use not simple.\n"); |
ebfd146a IR |
1789 | return false; |
1790 | } | |
ebfd146a | 1791 | |
b690cc0f RG |
1792 | if (!vectype_in) |
1793 | vectype_in = opvectype; | |
1794 | else if (opvectype | |
1795 | && opvectype != vectype_in) | |
1796 | { | |
73fbfcad | 1797 | if (dump_enabled_p ()) |
78c60e3d | 1798 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1799 | "argument vector types differ.\n"); |
b690cc0f RG |
1800 | return false; |
1801 | } | |
1802 | } | |
1803 | /* If all arguments are external or constant defs use a vector type with | |
1804 | the same size as the output vector type. */ | |
ebfd146a | 1805 | if (!vectype_in) |
b690cc0f | 1806 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
1807 | if (vec_stmt) |
1808 | gcc_assert (vectype_in); | |
1809 | if (!vectype_in) | |
1810 | { | |
73fbfcad | 1811 | if (dump_enabled_p ()) |
7d8930a0 | 1812 | { |
78c60e3d SS |
1813 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1814 | "no vectype for scalar type "); | |
1815 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 1816 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
1817 | } |
1818 | ||
1819 | return false; | |
1820 | } | |
ebfd146a IR |
1821 | |
1822 | /* FORNOW */ | |
b690cc0f RG |
1823 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
1824 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
1825 | if (nunits_in == nunits_out / 2) |
1826 | modifier = NARROW; | |
1827 | else if (nunits_out == nunits_in) | |
1828 | modifier = NONE; | |
1829 | else if (nunits_out == nunits_in / 2) | |
1830 | modifier = WIDEN; | |
1831 | else | |
1832 | return false; | |
1833 | ||
1834 | /* For now, we only vectorize functions if a target specific builtin | |
1835 | is available. TODO -- in some cases, it might be profitable to | |
1836 | insert the calls for pieces of the vector, in order to be able | |
1837 | to vectorize other operations in the loop. */ | |
1838 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
1839 | if (fndecl == NULL_TREE) | |
1840 | { | |
74bf76ed JJ |
1841 | if (gimple_call_internal_p (stmt) |
1842 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE | |
1843 | && !slp_node | |
1844 | && loop_vinfo | |
1845 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
1846 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
1847 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
1848 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
1849 | { | |
1850 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
1851 | { 0, 1, 2, ... vf - 1 } vector. */ | |
1852 | gcc_assert (nargs == 0); | |
1853 | } | |
1854 | else | |
1855 | { | |
1856 | if (dump_enabled_p ()) | |
1857 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 1858 | "function is not vectorizable.\n"); |
74bf76ed JJ |
1859 | return false; |
1860 | } | |
ebfd146a IR |
1861 | } |
1862 | ||
5006671f | 1863 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a | 1864 | |
190c2236 JJ |
1865 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
1866 | ncopies = 1; | |
1867 | else if (modifier == NARROW) | |
ebfd146a IR |
1868 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
1869 | else | |
1870 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1871 | ||
1872 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
1873 | needs to be generated. */ | |
1874 | gcc_assert (ncopies >= 1); | |
1875 | ||
1876 | if (!vec_stmt) /* transformation not required. */ | |
1877 | { | |
1878 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 1879 | if (dump_enabled_p ()) |
e645e942 TJ |
1880 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
1881 | "\n"); | |
c3e7ee41 | 1882 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
1883 | return true; |
1884 | } | |
1885 | ||
1886 | /** Transform. **/ | |
1887 | ||
73fbfcad | 1888 | if (dump_enabled_p ()) |
e645e942 | 1889 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
1890 | |
1891 | /* Handle def. */ | |
1892 | scalar_dest = gimple_call_lhs (stmt); | |
1893 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1894 | ||
1895 | prev_stmt_info = NULL; | |
1896 | switch (modifier) | |
1897 | { | |
1898 | case NONE: | |
1899 | for (j = 0; j < ncopies; ++j) | |
1900 | { | |
1901 | /* Build argument list for the vectorized call. */ | |
1902 | if (j == 0) | |
9771b263 | 1903 | vargs.create (nargs); |
ebfd146a | 1904 | else |
9771b263 | 1905 | vargs.truncate (0); |
ebfd146a | 1906 | |
190c2236 JJ |
1907 | if (slp_node) |
1908 | { | |
ef062b13 | 1909 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 1910 | vec<tree> vec_oprnds0; |
190c2236 JJ |
1911 | |
1912 | for (i = 0; i < nargs; i++) | |
9771b263 | 1913 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 1914 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 1915 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
1916 | |
1917 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 1918 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
1919 | { |
1920 | size_t k; | |
1921 | for (k = 0; k < nargs; k++) | |
1922 | { | |
37b5ec8f | 1923 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 1924 | vargs[k] = vec_oprndsk[i]; |
190c2236 JJ |
1925 | } |
1926 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1927 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1928 | gimple_call_set_lhs (new_stmt, new_temp); | |
1929 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 1930 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
1931 | } |
1932 | ||
1933 | for (i = 0; i < nargs; i++) | |
1934 | { | |
37b5ec8f | 1935 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 1936 | vec_oprndsi.release (); |
190c2236 | 1937 | } |
190c2236 JJ |
1938 | continue; |
1939 | } | |
1940 | ||
ebfd146a IR |
1941 | for (i = 0; i < nargs; i++) |
1942 | { | |
1943 | op = gimple_call_arg (stmt, i); | |
1944 | if (j == 0) | |
1945 | vec_oprnd0 | |
1946 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1947 | else | |
63827fb8 IR |
1948 | { |
1949 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
1950 | vec_oprnd0 | |
1951 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
1952 | } | |
ebfd146a | 1953 | |
9771b263 | 1954 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
1955 | } |
1956 | ||
74bf76ed JJ |
1957 | if (gimple_call_internal_p (stmt) |
1958 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
1959 | { | |
1960 | tree *v = XALLOCAVEC (tree, nunits_out); | |
1961 | int k; | |
1962 | for (k = 0; k < nunits_out; ++k) | |
1963 | v[k] = build_int_cst (unsigned_type_node, j * nunits_out + k); | |
1964 | tree cst = build_vector (vectype_out, v); | |
1965 | tree new_var | |
1966 | = vect_get_new_vect_var (vectype_out, vect_simple_var, "cst_"); | |
1967 | gimple init_stmt = gimple_build_assign (new_var, cst); | |
1968 | new_temp = make_ssa_name (new_var, init_stmt); | |
1969 | gimple_assign_set_lhs (init_stmt, new_temp); | |
1970 | vect_init_vector_1 (stmt, init_stmt, NULL); | |
1971 | new_temp = make_ssa_name (vec_dest, NULL); | |
1972 | new_stmt = gimple_build_assign (new_temp, | |
1973 | gimple_assign_lhs (init_stmt)); | |
1974 | } | |
1975 | else | |
1976 | { | |
1977 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1978 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1979 | gimple_call_set_lhs (new_stmt, new_temp); | |
1980 | } | |
ebfd146a IR |
1981 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1982 | ||
1983 | if (j == 0) | |
1984 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1985 | else | |
1986 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1987 | ||
1988 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1989 | } | |
1990 | ||
1991 | break; | |
1992 | ||
1993 | case NARROW: | |
1994 | for (j = 0; j < ncopies; ++j) | |
1995 | { | |
1996 | /* Build argument list for the vectorized call. */ | |
1997 | if (j == 0) | |
9771b263 | 1998 | vargs.create (nargs * 2); |
ebfd146a | 1999 | else |
9771b263 | 2000 | vargs.truncate (0); |
ebfd146a | 2001 | |
190c2236 JJ |
2002 | if (slp_node) |
2003 | { | |
ef062b13 | 2004 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2005 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2006 | |
2007 | for (i = 0; i < nargs; i++) | |
9771b263 | 2008 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2009 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2010 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2011 | |
2012 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2013 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
2014 | { |
2015 | size_t k; | |
9771b263 | 2016 | vargs.truncate (0); |
190c2236 JJ |
2017 | for (k = 0; k < nargs; k++) |
2018 | { | |
37b5ec8f | 2019 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
2020 | vargs.quick_push (vec_oprndsk[i]); |
2021 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 JJ |
2022 | } |
2023 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2024 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2025 | gimple_call_set_lhs (new_stmt, new_temp); | |
2026 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2027 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2028 | } |
2029 | ||
2030 | for (i = 0; i < nargs; i++) | |
2031 | { | |
37b5ec8f | 2032 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2033 | vec_oprndsi.release (); |
190c2236 | 2034 | } |
190c2236 JJ |
2035 | continue; |
2036 | } | |
2037 | ||
ebfd146a IR |
2038 | for (i = 0; i < nargs; i++) |
2039 | { | |
2040 | op = gimple_call_arg (stmt, i); | |
2041 | if (j == 0) | |
2042 | { | |
2043 | vec_oprnd0 | |
2044 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
2045 | vec_oprnd1 | |
63827fb8 | 2046 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2047 | } |
2048 | else | |
2049 | { | |
336ecb65 | 2050 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 2051 | vec_oprnd0 |
63827fb8 | 2052 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 2053 | vec_oprnd1 |
63827fb8 | 2054 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2055 | } |
2056 | ||
9771b263 DN |
2057 | vargs.quick_push (vec_oprnd0); |
2058 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
2059 | } |
2060 | ||
2061 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2062 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2063 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
2064 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2065 | ||
2066 | if (j == 0) | |
2067 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2068 | else | |
2069 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2070 | ||
2071 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2072 | } | |
2073 | ||
2074 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2075 | ||
2076 | break; | |
2077 | ||
2078 | case WIDEN: | |
2079 | /* No current target implements this case. */ | |
2080 | return false; | |
2081 | } | |
2082 | ||
9771b263 | 2083 | vargs.release (); |
ebfd146a IR |
2084 | |
2085 | /* Update the exception handling table with the vector stmt if necessary. */ | |
2086 | if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt)) | |
2087 | gimple_purge_dead_eh_edges (gimple_bb (stmt)); | |
2088 | ||
2089 | /* The call in STMT might prevent it from being removed in dce. | |
2090 | We however cannot remove it here, due to the way the ssa name | |
2091 | it defines is mapped to the new definition. So just replace | |
2092 | rhs of the statement with something harmless. */ | |
2093 | ||
dd34c087 JJ |
2094 | if (slp_node) |
2095 | return true; | |
2096 | ||
ebfd146a | 2097 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
2098 | if (is_pattern_stmt_p (stmt_info)) |
2099 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
2100 | else | |
2101 | lhs = gimple_call_lhs (stmt); | |
2102 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
ebfd146a | 2103 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 2104 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
2105 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
2106 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
2107 | |
2108 | return true; | |
2109 | } | |
2110 | ||
2111 | ||
2112 | /* Function vect_gen_widened_results_half | |
2113 | ||
2114 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 2115 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 2116 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
2117 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
2118 | needs to be created (DECL is a function-decl of a target-builtin). | |
2119 | STMT is the original scalar stmt that we are vectorizing. */ | |
2120 | ||
2121 | static gimple | |
2122 | vect_gen_widened_results_half (enum tree_code code, | |
2123 | tree decl, | |
2124 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
2125 | tree vec_dest, gimple_stmt_iterator *gsi, | |
2126 | gimple stmt) | |
b8698a0f | 2127 | { |
ebfd146a | 2128 | gimple new_stmt; |
b8698a0f L |
2129 | tree new_temp; |
2130 | ||
2131 | /* Generate half of the widened result: */ | |
2132 | if (code == CALL_EXPR) | |
2133 | { | |
2134 | /* Target specific support */ | |
ebfd146a IR |
2135 | if (op_type == binary_op) |
2136 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
2137 | else | |
2138 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
2139 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2140 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
2141 | } |
2142 | else | |
ebfd146a | 2143 | { |
b8698a0f L |
2144 | /* Generic support */ |
2145 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
2146 | if (op_type != binary_op) |
2147 | vec_oprnd1 = NULL; | |
2148 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0, | |
2149 | vec_oprnd1); | |
2150 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2151 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 2152 | } |
ebfd146a IR |
2153 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2154 | ||
ebfd146a IR |
2155 | return new_stmt; |
2156 | } | |
2157 | ||
4a00c761 JJ |
2158 | |
2159 | /* Get vectorized definitions for loop-based vectorization. For the first | |
2160 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
2161 | scalar operand), and for the rest we get a copy with | |
2162 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
2163 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
2164 | The vectors are collected into VEC_OPRNDS. */ | |
2165 | ||
2166 | static void | |
2167 | vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt, | |
9771b263 | 2168 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
2169 | { |
2170 | tree vec_oprnd; | |
2171 | ||
2172 | /* Get first vector operand. */ | |
2173 | /* All the vector operands except the very first one (that is scalar oprnd) | |
2174 | are stmt copies. */ | |
2175 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
2176 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL); | |
2177 | else | |
2178 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
2179 | ||
9771b263 | 2180 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
2181 | |
2182 | /* Get second vector operand. */ | |
2183 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 2184 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
2185 | |
2186 | *oprnd = vec_oprnd; | |
2187 | ||
2188 | /* For conversion in multiple steps, continue to get operands | |
2189 | recursively. */ | |
2190 | if (multi_step_cvt) | |
2191 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
2192 | } | |
2193 | ||
2194 | ||
2195 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
2196 | For multi-step conversions store the resulting vectors and call the function | |
2197 | recursively. */ | |
2198 | ||
2199 | static void | |
9771b263 | 2200 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
4a00c761 | 2201 | int multi_step_cvt, gimple stmt, |
9771b263 | 2202 | vec<tree> vec_dsts, |
4a00c761 JJ |
2203 | gimple_stmt_iterator *gsi, |
2204 | slp_tree slp_node, enum tree_code code, | |
2205 | stmt_vec_info *prev_stmt_info) | |
2206 | { | |
2207 | unsigned int i; | |
2208 | tree vop0, vop1, new_tmp, vec_dest; | |
2209 | gimple new_stmt; | |
2210 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2211 | ||
9771b263 | 2212 | vec_dest = vec_dsts.pop (); |
4a00c761 | 2213 | |
9771b263 | 2214 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
2215 | { |
2216 | /* Create demotion operation. */ | |
9771b263 DN |
2217 | vop0 = (*vec_oprnds)[i]; |
2218 | vop1 = (*vec_oprnds)[i + 1]; | |
4a00c761 JJ |
2219 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); |
2220 | new_tmp = make_ssa_name (vec_dest, new_stmt); | |
2221 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
2222 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2223 | ||
2224 | if (multi_step_cvt) | |
2225 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 2226 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
2227 | else |
2228 | { | |
2229 | /* This is the last step of the conversion sequence. Store the | |
2230 | vectors in SLP_NODE or in vector info of the scalar statement | |
2231 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
2232 | if (slp_node) | |
9771b263 | 2233 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2234 | else |
2235 | { | |
2236 | if (!*prev_stmt_info) | |
2237 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2238 | else | |
2239 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
2240 | ||
2241 | *prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2242 | } | |
2243 | } | |
2244 | } | |
2245 | ||
2246 | /* For multi-step demotion operations we first generate demotion operations | |
2247 | from the source type to the intermediate types, and then combine the | |
2248 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
2249 | type. */ | |
2250 | if (multi_step_cvt) | |
2251 | { | |
2252 | /* At each level of recursion we have half of the operands we had at the | |
2253 | previous level. */ | |
9771b263 | 2254 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
2255 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
2256 | stmt, vec_dsts, gsi, slp_node, | |
2257 | VEC_PACK_TRUNC_EXPR, | |
2258 | prev_stmt_info); | |
2259 | } | |
2260 | ||
9771b263 | 2261 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2262 | } |
2263 | ||
2264 | ||
2265 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
2266 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
2267 | the resulting vectors and call the function recursively. */ | |
2268 | ||
2269 | static void | |
9771b263 DN |
2270 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
2271 | vec<tree> *vec_oprnds1, | |
4a00c761 JJ |
2272 | gimple stmt, tree vec_dest, |
2273 | gimple_stmt_iterator *gsi, | |
2274 | enum tree_code code1, | |
2275 | enum tree_code code2, tree decl1, | |
2276 | tree decl2, int op_type) | |
2277 | { | |
2278 | int i; | |
2279 | tree vop0, vop1, new_tmp1, new_tmp2; | |
2280 | gimple new_stmt1, new_stmt2; | |
6e1aa848 | 2281 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 2282 | |
9771b263 DN |
2283 | vec_tmp.create (vec_oprnds0->length () * 2); |
2284 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
2285 | { |
2286 | if (op_type == binary_op) | |
9771b263 | 2287 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
2288 | else |
2289 | vop1 = NULL_TREE; | |
2290 | ||
2291 | /* Generate the two halves of promotion operation. */ | |
2292 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
2293 | op_type, vec_dest, gsi, stmt); | |
2294 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
2295 | op_type, vec_dest, gsi, stmt); | |
2296 | if (is_gimple_call (new_stmt1)) | |
2297 | { | |
2298 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
2299 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
2300 | } | |
2301 | else | |
2302 | { | |
2303 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
2304 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
2305 | } | |
2306 | ||
2307 | /* Store the results for the next step. */ | |
9771b263 DN |
2308 | vec_tmp.quick_push (new_tmp1); |
2309 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
2310 | } |
2311 | ||
689eaba3 | 2312 | vec_oprnds0->release (); |
4a00c761 JJ |
2313 | *vec_oprnds0 = vec_tmp; |
2314 | } | |
2315 | ||
2316 | ||
b8698a0f L |
2317 | /* Check if STMT performs a conversion operation, that can be vectorized. |
2318 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 2319 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
2320 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
2321 | ||
2322 | static bool | |
2323 | vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi, | |
2324 | gimple *vec_stmt, slp_tree slp_node) | |
2325 | { | |
2326 | tree vec_dest; | |
2327 | tree scalar_dest; | |
4a00c761 | 2328 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
2329 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
2330 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2331 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2332 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 2333 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
2334 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
2335 | tree new_temp; | |
2336 | tree def; | |
2337 | gimple def_stmt; | |
2338 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2339 | gimple new_stmt = NULL; | |
2340 | stmt_vec_info prev_stmt_info; | |
2341 | int nunits_in; | |
2342 | int nunits_out; | |
2343 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
2344 | int ncopies, i, j; |
2345 | tree lhs_type, rhs_type; | |
ebfd146a | 2346 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
2347 | vec<tree> vec_oprnds0 = vNULL; |
2348 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 2349 | tree vop0; |
4a00c761 JJ |
2350 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2351 | int multi_step_cvt = 0; | |
6e1aa848 DN |
2352 | vec<tree> vec_dsts = vNULL; |
2353 | vec<tree> interm_types = vNULL; | |
4a00c761 JJ |
2354 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
2355 | int op_type; | |
2356 | enum machine_mode rhs_mode; | |
2357 | unsigned short fltsz; | |
ebfd146a IR |
2358 | |
2359 | /* Is STMT a vectorizable conversion? */ | |
2360 | ||
4a00c761 | 2361 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2362 | return false; |
2363 | ||
8644a673 | 2364 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2365 | return false; |
2366 | ||
2367 | if (!is_gimple_assign (stmt)) | |
2368 | return false; | |
2369 | ||
2370 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2371 | return false; | |
2372 | ||
2373 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
2374 | if (!CONVERT_EXPR_CODE_P (code) |
2375 | && code != FIX_TRUNC_EXPR | |
2376 | && code != FLOAT_EXPR | |
2377 | && code != WIDEN_MULT_EXPR | |
2378 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
2379 | return false; |
2380 | ||
4a00c761 JJ |
2381 | op_type = TREE_CODE_LENGTH (code); |
2382 | ||
ebfd146a | 2383 | /* Check types of lhs and rhs. */ |
b690cc0f | 2384 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 2385 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
2386 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2387 | ||
ebfd146a IR |
2388 | op0 = gimple_assign_rhs1 (stmt); |
2389 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
2390 | |
2391 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
2392 | && !((INTEGRAL_TYPE_P (lhs_type) | |
2393 | && INTEGRAL_TYPE_P (rhs_type)) | |
2394 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
2395 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
2396 | return false; | |
2397 | ||
2398 | if ((INTEGRAL_TYPE_P (lhs_type) | |
2399 | && (TYPE_PRECISION (lhs_type) | |
2400 | != GET_MODE_PRECISION (TYPE_MODE (lhs_type)))) | |
2401 | || (INTEGRAL_TYPE_P (rhs_type) | |
2402 | && (TYPE_PRECISION (rhs_type) | |
2403 | != GET_MODE_PRECISION (TYPE_MODE (rhs_type))))) | |
2404 | { | |
73fbfcad | 2405 | if (dump_enabled_p ()) |
78c60e3d | 2406 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
2407 | "type conversion to/from bit-precision unsupported." |
2408 | "\n"); | |
4a00c761 JJ |
2409 | return false; |
2410 | } | |
2411 | ||
b690cc0f | 2412 | /* Check the operands of the operation. */ |
24ee1384 | 2413 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
b690cc0f RG |
2414 | &def_stmt, &def, &dt[0], &vectype_in)) |
2415 | { | |
73fbfcad | 2416 | if (dump_enabled_p ()) |
78c60e3d | 2417 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2418 | "use not simple.\n"); |
b690cc0f RG |
2419 | return false; |
2420 | } | |
4a00c761 JJ |
2421 | if (op_type == binary_op) |
2422 | { | |
2423 | bool ok; | |
2424 | ||
2425 | op1 = gimple_assign_rhs2 (stmt); | |
2426 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
2427 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
2428 | OP1. */ | |
2429 | if (CONSTANT_CLASS_P (op0)) | |
f5709183 | 2430 | ok = vect_is_simple_use_1 (op1, stmt, loop_vinfo, bb_vinfo, |
4a00c761 JJ |
2431 | &def_stmt, &def, &dt[1], &vectype_in); |
2432 | else | |
f5709183 | 2433 | ok = vect_is_simple_use (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
24ee1384 | 2434 | &def, &dt[1]); |
4a00c761 JJ |
2435 | |
2436 | if (!ok) | |
2437 | { | |
73fbfcad | 2438 | if (dump_enabled_p ()) |
78c60e3d | 2439 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2440 | "use not simple.\n"); |
4a00c761 JJ |
2441 | return false; |
2442 | } | |
2443 | } | |
2444 | ||
b690cc0f RG |
2445 | /* If op0 is an external or constant defs use a vector type of |
2446 | the same size as the output vector type. */ | |
ebfd146a | 2447 | if (!vectype_in) |
b690cc0f | 2448 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2449 | if (vec_stmt) |
2450 | gcc_assert (vectype_in); | |
2451 | if (!vectype_in) | |
2452 | { | |
73fbfcad | 2453 | if (dump_enabled_p ()) |
4a00c761 | 2454 | { |
78c60e3d SS |
2455 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2456 | "no vectype for scalar type "); | |
2457 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2458 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 2459 | } |
7d8930a0 IR |
2460 | |
2461 | return false; | |
2462 | } | |
ebfd146a | 2463 | |
b690cc0f RG |
2464 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2465 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 2466 | if (nunits_in < nunits_out) |
ebfd146a IR |
2467 | modifier = NARROW; |
2468 | else if (nunits_out == nunits_in) | |
2469 | modifier = NONE; | |
ebfd146a | 2470 | else |
4a00c761 | 2471 | modifier = WIDEN; |
ebfd146a | 2472 | |
ff802fa1 IR |
2473 | /* Multiple types in SLP are handled by creating the appropriate number of |
2474 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2475 | case of SLP. */ | |
437f4a00 | 2476 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a | 2477 | ncopies = 1; |
4a00c761 JJ |
2478 | else if (modifier == NARROW) |
2479 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
2480 | else | |
2481 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 2482 | |
ebfd146a IR |
2483 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
2484 | needs to be generated. */ | |
2485 | gcc_assert (ncopies >= 1); | |
2486 | ||
ebfd146a | 2487 | /* Supportable by target? */ |
4a00c761 | 2488 | switch (modifier) |
ebfd146a | 2489 | { |
4a00c761 JJ |
2490 | case NONE: |
2491 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
2492 | return false; | |
2493 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
2494 | &decl1, &code1)) | |
2495 | break; | |
2496 | /* FALLTHRU */ | |
2497 | unsupported: | |
73fbfcad | 2498 | if (dump_enabled_p ()) |
78c60e3d | 2499 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2500 | "conversion not supported by target.\n"); |
ebfd146a | 2501 | return false; |
ebfd146a | 2502 | |
4a00c761 JJ |
2503 | case WIDEN: |
2504 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
2505 | &code1, &code2, &multi_step_cvt, |
2506 | &interm_types)) | |
4a00c761 JJ |
2507 | { |
2508 | /* Binary widening operation can only be supported directly by the | |
2509 | architecture. */ | |
2510 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
2511 | break; | |
2512 | } | |
2513 | ||
2514 | if (code != FLOAT_EXPR | |
2515 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
2516 | <= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
2517 | goto unsupported; | |
2518 | ||
2519 | rhs_mode = TYPE_MODE (rhs_type); | |
2520 | fltsz = GET_MODE_SIZE (TYPE_MODE (lhs_type)); | |
2521 | for (rhs_mode = GET_MODE_2XWIDER_MODE (TYPE_MODE (rhs_type)); | |
2522 | rhs_mode != VOIDmode && GET_MODE_SIZE (rhs_mode) <= fltsz; | |
2523 | rhs_mode = GET_MODE_2XWIDER_MODE (rhs_mode)) | |
2524 | { | |
2525 | cvt_type | |
2526 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
2527 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
2528 | if (cvt_type == NULL_TREE) | |
2529 | goto unsupported; | |
2530 | ||
2531 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
2532 | { | |
2533 | if (!supportable_convert_operation (code, vectype_out, | |
2534 | cvt_type, &decl1, &codecvt1)) | |
2535 | goto unsupported; | |
2536 | } | |
2537 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
2538 | cvt_type, &codecvt1, |
2539 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
2540 | &interm_types)) |
2541 | continue; | |
2542 | else | |
2543 | gcc_assert (multi_step_cvt == 0); | |
2544 | ||
2545 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
2546 | vectype_in, &code1, &code2, |
2547 | &multi_step_cvt, &interm_types)) | |
4a00c761 JJ |
2548 | break; |
2549 | } | |
2550 | ||
2551 | if (rhs_mode == VOIDmode || GET_MODE_SIZE (rhs_mode) > fltsz) | |
2552 | goto unsupported; | |
2553 | ||
2554 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
2555 | codecvt2 = ERROR_MARK; | |
2556 | else | |
2557 | { | |
2558 | multi_step_cvt++; | |
9771b263 | 2559 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
2560 | cvt_type = NULL_TREE; |
2561 | } | |
2562 | break; | |
2563 | ||
2564 | case NARROW: | |
2565 | gcc_assert (op_type == unary_op); | |
2566 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
2567 | &code1, &multi_step_cvt, | |
2568 | &interm_types)) | |
2569 | break; | |
2570 | ||
2571 | if (code != FIX_TRUNC_EXPR | |
2572 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
2573 | >= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
2574 | goto unsupported; | |
2575 | ||
2576 | rhs_mode = TYPE_MODE (rhs_type); | |
2577 | cvt_type | |
2578 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
2579 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
2580 | if (cvt_type == NULL_TREE) | |
2581 | goto unsupported; | |
2582 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
2583 | &decl1, &codecvt1)) | |
2584 | goto unsupported; | |
2585 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
2586 | &code1, &multi_step_cvt, | |
2587 | &interm_types)) | |
2588 | break; | |
2589 | goto unsupported; | |
2590 | ||
2591 | default: | |
2592 | gcc_unreachable (); | |
ebfd146a IR |
2593 | } |
2594 | ||
2595 | if (!vec_stmt) /* transformation not required. */ | |
2596 | { | |
73fbfcad | 2597 | if (dump_enabled_p ()) |
78c60e3d | 2598 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 2599 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 2600 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
2601 | { |
2602 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
c3e7ee41 | 2603 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
8bd37302 | 2604 | } |
4a00c761 JJ |
2605 | else if (modifier == NARROW) |
2606 | { | |
2607 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 2608 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
2609 | } |
2610 | else | |
2611 | { | |
2612 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 2613 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 2614 | } |
9771b263 | 2615 | interm_types.release (); |
ebfd146a IR |
2616 | return true; |
2617 | } | |
2618 | ||
2619 | /** Transform. **/ | |
73fbfcad | 2620 | if (dump_enabled_p ()) |
78c60e3d | 2621 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 2622 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 2623 | |
4a00c761 JJ |
2624 | if (op_type == binary_op) |
2625 | { | |
2626 | if (CONSTANT_CLASS_P (op0)) | |
2627 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
2628 | else if (CONSTANT_CLASS_P (op1)) | |
2629 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
2630 | } | |
2631 | ||
2632 | /* In case of multi-step conversion, we first generate conversion operations | |
2633 | to the intermediate types, and then from that types to the final one. | |
2634 | We create vector destinations for the intermediate type (TYPES) received | |
2635 | from supportable_*_operation, and store them in the correct order | |
2636 | for future use in vect_create_vectorized_*_stmts (). */ | |
9771b263 | 2637 | vec_dsts.create (multi_step_cvt + 1); |
82294ec1 JJ |
2638 | vec_dest = vect_create_destination_var (scalar_dest, |
2639 | (cvt_type && modifier == WIDEN) | |
2640 | ? cvt_type : vectype_out); | |
9771b263 | 2641 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2642 | |
2643 | if (multi_step_cvt) | |
2644 | { | |
9771b263 DN |
2645 | for (i = interm_types.length () - 1; |
2646 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
2647 | { |
2648 | vec_dest = vect_create_destination_var (scalar_dest, | |
2649 | intermediate_type); | |
9771b263 | 2650 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2651 | } |
2652 | } | |
ebfd146a | 2653 | |
4a00c761 | 2654 | if (cvt_type) |
82294ec1 JJ |
2655 | vec_dest = vect_create_destination_var (scalar_dest, |
2656 | modifier == WIDEN | |
2657 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
2658 | |
2659 | if (!slp_node) | |
2660 | { | |
30862efc | 2661 | if (modifier == WIDEN) |
4a00c761 | 2662 | { |
c3284718 | 2663 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 2664 | if (op_type == binary_op) |
9771b263 | 2665 | vec_oprnds1.create (1); |
4a00c761 | 2666 | } |
30862efc | 2667 | else if (modifier == NARROW) |
9771b263 DN |
2668 | vec_oprnds0.create ( |
2669 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
2670 | } |
2671 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 2672 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 2673 | |
4a00c761 | 2674 | last_oprnd = op0; |
ebfd146a IR |
2675 | prev_stmt_info = NULL; |
2676 | switch (modifier) | |
2677 | { | |
2678 | case NONE: | |
2679 | for (j = 0; j < ncopies; j++) | |
2680 | { | |
ebfd146a | 2681 | if (j == 0) |
d092494c IR |
2682 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node, |
2683 | -1); | |
ebfd146a IR |
2684 | else |
2685 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
2686 | ||
9771b263 | 2687 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2688 | { |
2689 | /* Arguments are ready, create the new vector stmt. */ | |
2690 | if (code1 == CALL_EXPR) | |
2691 | { | |
2692 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2693 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2694 | gimple_call_set_lhs (new_stmt, new_temp); | |
2695 | } | |
2696 | else | |
2697 | { | |
2698 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
2699 | new_stmt = gimple_build_assign_with_ops (code1, vec_dest, | |
2700 | vop0, NULL); | |
2701 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2702 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2703 | } | |
2704 | ||
2705 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2706 | if (slp_node) | |
9771b263 | 2707 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2708 | } |
2709 | ||
ebfd146a IR |
2710 | if (j == 0) |
2711 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2712 | else | |
2713 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2714 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2715 | } | |
2716 | break; | |
2717 | ||
2718 | case WIDEN: | |
2719 | /* In case the vectorization factor (VF) is bigger than the number | |
2720 | of elements that we can fit in a vectype (nunits), we have to | |
2721 | generate more than one vector stmt - i.e - we need to "unroll" | |
2722 | the vector stmt by a factor VF/nunits. */ | |
2723 | for (j = 0; j < ncopies; j++) | |
2724 | { | |
4a00c761 | 2725 | /* Handle uses. */ |
ebfd146a | 2726 | if (j == 0) |
4a00c761 JJ |
2727 | { |
2728 | if (slp_node) | |
2729 | { | |
2730 | if (code == WIDEN_LSHIFT_EXPR) | |
2731 | { | |
2732 | unsigned int k; | |
ebfd146a | 2733 | |
4a00c761 JJ |
2734 | vec_oprnd1 = op1; |
2735 | /* Store vec_oprnd1 for every vector stmt to be created | |
2736 | for SLP_NODE. We check during the analysis that all | |
2737 | the shift arguments are the same. */ | |
2738 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 2739 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
2740 | |
2741 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2742 | slp_node, -1); | |
2743 | } | |
2744 | else | |
2745 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
2746 | &vec_oprnds1, slp_node, -1); | |
2747 | } | |
2748 | else | |
2749 | { | |
2750 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
9771b263 | 2751 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
2752 | if (op_type == binary_op) |
2753 | { | |
2754 | if (code == WIDEN_LSHIFT_EXPR) | |
2755 | vec_oprnd1 = op1; | |
2756 | else | |
2757 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, | |
2758 | NULL); | |
9771b263 | 2759 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
2760 | } |
2761 | } | |
2762 | } | |
ebfd146a | 2763 | else |
4a00c761 JJ |
2764 | { |
2765 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
2766 | vec_oprnds0.truncate (0); |
2767 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
2768 | if (op_type == binary_op) |
2769 | { | |
2770 | if (code == WIDEN_LSHIFT_EXPR) | |
2771 | vec_oprnd1 = op1; | |
2772 | else | |
2773 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
2774 | vec_oprnd1); | |
9771b263 DN |
2775 | vec_oprnds1.truncate (0); |
2776 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
2777 | } |
2778 | } | |
ebfd146a | 2779 | |
4a00c761 JJ |
2780 | /* Arguments are ready. Create the new vector stmts. */ |
2781 | for (i = multi_step_cvt; i >= 0; i--) | |
2782 | { | |
9771b263 | 2783 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
2784 | enum tree_code c1 = code1, c2 = code2; |
2785 | if (i == 0 && codecvt2 != ERROR_MARK) | |
2786 | { | |
2787 | c1 = codecvt1; | |
2788 | c2 = codecvt2; | |
2789 | } | |
2790 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
2791 | &vec_oprnds1, | |
2792 | stmt, this_dest, gsi, | |
2793 | c1, c2, decl1, decl2, | |
2794 | op_type); | |
2795 | } | |
2796 | ||
9771b263 | 2797 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2798 | { |
2799 | if (cvt_type) | |
2800 | { | |
2801 | if (codecvt1 == CALL_EXPR) | |
2802 | { | |
2803 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2804 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2805 | gimple_call_set_lhs (new_stmt, new_temp); | |
2806 | } | |
2807 | else | |
2808 | { | |
2809 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
2810 | new_temp = make_ssa_name (vec_dest, NULL); | |
2811 | new_stmt = gimple_build_assign_with_ops (codecvt1, | |
2812 | new_temp, | |
2813 | vop0, NULL); | |
2814 | } | |
2815 | ||
2816 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2817 | } | |
2818 | else | |
2819 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
2820 | ||
2821 | if (slp_node) | |
9771b263 | 2822 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2823 | else |
2824 | { | |
2825 | if (!prev_stmt_info) | |
2826 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2827 | else | |
2828 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2829 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2830 | } | |
2831 | } | |
ebfd146a | 2832 | } |
4a00c761 JJ |
2833 | |
2834 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
2835 | break; |
2836 | ||
2837 | case NARROW: | |
2838 | /* In case the vectorization factor (VF) is bigger than the number | |
2839 | of elements that we can fit in a vectype (nunits), we have to | |
2840 | generate more than one vector stmt - i.e - we need to "unroll" | |
2841 | the vector stmt by a factor VF/nunits. */ | |
2842 | for (j = 0; j < ncopies; j++) | |
2843 | { | |
2844 | /* Handle uses. */ | |
4a00c761 JJ |
2845 | if (slp_node) |
2846 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2847 | slp_node, -1); | |
ebfd146a IR |
2848 | else |
2849 | { | |
9771b263 | 2850 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
2851 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
2852 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
2853 | } |
2854 | ||
4a00c761 JJ |
2855 | /* Arguments are ready. Create the new vector stmts. */ |
2856 | if (cvt_type) | |
9771b263 | 2857 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2858 | { |
2859 | if (codecvt1 == CALL_EXPR) | |
2860 | { | |
2861 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2862 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2863 | gimple_call_set_lhs (new_stmt, new_temp); | |
2864 | } | |
2865 | else | |
2866 | { | |
2867 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
2868 | new_temp = make_ssa_name (vec_dest, NULL); | |
2869 | new_stmt = gimple_build_assign_with_ops (codecvt1, new_temp, | |
2870 | vop0, NULL); | |
2871 | } | |
ebfd146a | 2872 | |
4a00c761 | 2873 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 2874 | vec_oprnds0[i] = new_temp; |
4a00c761 | 2875 | } |
ebfd146a | 2876 | |
4a00c761 JJ |
2877 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
2878 | stmt, vec_dsts, gsi, | |
2879 | slp_node, code1, | |
2880 | &prev_stmt_info); | |
ebfd146a IR |
2881 | } |
2882 | ||
2883 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 2884 | break; |
ebfd146a IR |
2885 | } |
2886 | ||
9771b263 DN |
2887 | vec_oprnds0.release (); |
2888 | vec_oprnds1.release (); | |
2889 | vec_dsts.release (); | |
2890 | interm_types.release (); | |
ebfd146a IR |
2891 | |
2892 | return true; | |
2893 | } | |
ff802fa1 IR |
2894 | |
2895 | ||
ebfd146a IR |
2896 | /* Function vectorizable_assignment. |
2897 | ||
b8698a0f L |
2898 | Check if STMT performs an assignment (copy) that can be vectorized. |
2899 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
2900 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2901 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2902 | ||
2903 | static bool | |
2904 | vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi, | |
2905 | gimple *vec_stmt, slp_tree slp_node) | |
2906 | { | |
2907 | tree vec_dest; | |
2908 | tree scalar_dest; | |
2909 | tree op; | |
2910 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2911 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2912 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2913 | tree new_temp; | |
2914 | tree def; | |
2915 | gimple def_stmt; | |
2916 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
fde9c428 | 2917 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
ebfd146a | 2918 | int ncopies; |
f18b55bd | 2919 | int i, j; |
6e1aa848 | 2920 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 2921 | tree vop; |
a70d6342 | 2922 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
f18b55bd IR |
2923 | gimple new_stmt = NULL; |
2924 | stmt_vec_info prev_stmt_info = NULL; | |
fde9c428 RG |
2925 | enum tree_code code; |
2926 | tree vectype_in; | |
ebfd146a IR |
2927 | |
2928 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2929 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2930 | case of SLP. */ | |
437f4a00 | 2931 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
2932 | ncopies = 1; |
2933 | else | |
2934 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2935 | ||
2936 | gcc_assert (ncopies >= 1); | |
ebfd146a | 2937 | |
a70d6342 | 2938 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2939 | return false; |
2940 | ||
8644a673 | 2941 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2942 | return false; |
2943 | ||
2944 | /* Is vectorizable assignment? */ | |
2945 | if (!is_gimple_assign (stmt)) | |
2946 | return false; | |
2947 | ||
2948 | scalar_dest = gimple_assign_lhs (stmt); | |
2949 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
2950 | return false; | |
2951 | ||
fde9c428 | 2952 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 2953 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
2954 | || code == PAREN_EXPR |
2955 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
2956 | op = gimple_assign_rhs1 (stmt); |
2957 | else | |
2958 | return false; | |
2959 | ||
7b7ec6c5 RG |
2960 | if (code == VIEW_CONVERT_EXPR) |
2961 | op = TREE_OPERAND (op, 0); | |
2962 | ||
24ee1384 | 2963 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, |
fde9c428 | 2964 | &def_stmt, &def, &dt[0], &vectype_in)) |
ebfd146a | 2965 | { |
73fbfcad | 2966 | if (dump_enabled_p ()) |
78c60e3d | 2967 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2968 | "use not simple.\n"); |
ebfd146a IR |
2969 | return false; |
2970 | } | |
2971 | ||
fde9c428 RG |
2972 | /* We can handle NOP_EXPR conversions that do not change the number |
2973 | of elements or the vector size. */ | |
7b7ec6c5 RG |
2974 | if ((CONVERT_EXPR_CODE_P (code) |
2975 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
2976 | && (!vectype_in |
2977 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
2978 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
2979 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
2980 | return false; | |
2981 | ||
7b7b1813 RG |
2982 | /* We do not handle bit-precision changes. */ |
2983 | if ((CONVERT_EXPR_CODE_P (code) | |
2984 | || code == VIEW_CONVERT_EXPR) | |
2985 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2986 | && ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
2987 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
2988 | || ((TYPE_PRECISION (TREE_TYPE (op)) | |
2989 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))) | |
2990 | /* But a conversion that does not change the bit-pattern is ok. */ | |
2991 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
2992 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2993 | && TYPE_UNSIGNED (TREE_TYPE (op)))) | |
2994 | { | |
73fbfcad | 2995 | if (dump_enabled_p ()) |
78c60e3d SS |
2996 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2997 | "type conversion to/from bit-precision " | |
e645e942 | 2998 | "unsupported.\n"); |
7b7b1813 RG |
2999 | return false; |
3000 | } | |
3001 | ||
ebfd146a IR |
3002 | if (!vec_stmt) /* transformation not required. */ |
3003 | { | |
3004 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 3005 | if (dump_enabled_p ()) |
78c60e3d | 3006 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3007 | "=== vectorizable_assignment ===\n"); |
c3e7ee41 | 3008 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
3009 | return true; |
3010 | } | |
3011 | ||
3012 | /** Transform. **/ | |
73fbfcad | 3013 | if (dump_enabled_p ()) |
e645e942 | 3014 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
3015 | |
3016 | /* Handle def. */ | |
3017 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3018 | ||
3019 | /* Handle use. */ | |
f18b55bd | 3020 | for (j = 0; j < ncopies; j++) |
ebfd146a | 3021 | { |
f18b55bd IR |
3022 | /* Handle uses. */ |
3023 | if (j == 0) | |
d092494c | 3024 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node, -1); |
f18b55bd IR |
3025 | else |
3026 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
3027 | ||
3028 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 3029 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 3030 | { |
7b7ec6c5 RG |
3031 | if (CONVERT_EXPR_CODE_P (code) |
3032 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 3033 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
3034 | new_stmt = gimple_build_assign (vec_dest, vop); |
3035 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3036 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3037 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3038 | if (slp_node) | |
9771b263 | 3039 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 3040 | } |
ebfd146a IR |
3041 | |
3042 | if (slp_node) | |
f18b55bd IR |
3043 | continue; |
3044 | ||
3045 | if (j == 0) | |
3046 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3047 | else | |
3048 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3049 | ||
3050 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3051 | } | |
b8698a0f | 3052 | |
9771b263 | 3053 | vec_oprnds.release (); |
ebfd146a IR |
3054 | return true; |
3055 | } | |
3056 | ||
9dc3f7de | 3057 | |
1107f3ae IR |
3058 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
3059 | either as shift by a scalar or by a vector. */ | |
3060 | ||
3061 | bool | |
3062 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
3063 | { | |
3064 | ||
3065 | enum machine_mode vec_mode; | |
3066 | optab optab; | |
3067 | int icode; | |
3068 | tree vectype; | |
3069 | ||
3070 | vectype = get_vectype_for_scalar_type (scalar_type); | |
3071 | if (!vectype) | |
3072 | return false; | |
3073 | ||
3074 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
3075 | if (!optab | |
3076 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
3077 | { | |
3078 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
3079 | if (!optab | |
3080 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
3081 | == CODE_FOR_nothing)) | |
3082 | return false; | |
3083 | } | |
3084 | ||
3085 | vec_mode = TYPE_MODE (vectype); | |
3086 | icode = (int) optab_handler (optab, vec_mode); | |
3087 | if (icode == CODE_FOR_nothing) | |
3088 | return false; | |
3089 | ||
3090 | return true; | |
3091 | } | |
3092 | ||
3093 | ||
9dc3f7de IR |
3094 | /* Function vectorizable_shift. |
3095 | ||
3096 | Check if STMT performs a shift operation that can be vectorized. | |
3097 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3098 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3099 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3100 | ||
3101 | static bool | |
3102 | vectorizable_shift (gimple stmt, gimple_stmt_iterator *gsi, | |
3103 | gimple *vec_stmt, slp_tree slp_node) | |
3104 | { | |
3105 | tree vec_dest; | |
3106 | tree scalar_dest; | |
3107 | tree op0, op1 = NULL; | |
3108 | tree vec_oprnd1 = NULL_TREE; | |
3109 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3110 | tree vectype; | |
3111 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3112 | enum tree_code code; | |
3113 | enum machine_mode vec_mode; | |
3114 | tree new_temp; | |
3115 | optab optab; | |
3116 | int icode; | |
3117 | enum machine_mode optab_op2_mode; | |
3118 | tree def; | |
3119 | gimple def_stmt; | |
3120 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
3121 | gimple new_stmt = NULL; | |
3122 | stmt_vec_info prev_stmt_info; | |
3123 | int nunits_in; | |
3124 | int nunits_out; | |
3125 | tree vectype_out; | |
cede2577 | 3126 | tree op1_vectype; |
9dc3f7de IR |
3127 | int ncopies; |
3128 | int j, i; | |
6e1aa848 DN |
3129 | vec<tree> vec_oprnds0 = vNULL; |
3130 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
3131 | tree vop0, vop1; |
3132 | unsigned int k; | |
49eab32e | 3133 | bool scalar_shift_arg = true; |
9dc3f7de IR |
3134 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3135 | int vf; | |
3136 | ||
3137 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
3138 | return false; | |
3139 | ||
3140 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
3141 | return false; | |
3142 | ||
3143 | /* Is STMT a vectorizable binary/unary operation? */ | |
3144 | if (!is_gimple_assign (stmt)) | |
3145 | return false; | |
3146 | ||
3147 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3148 | return false; | |
3149 | ||
3150 | code = gimple_assign_rhs_code (stmt); | |
3151 | ||
3152 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
3153 | || code == RROTATE_EXPR)) | |
3154 | return false; | |
3155 | ||
3156 | scalar_dest = gimple_assign_lhs (stmt); | |
3157 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
7b7b1813 RG |
3158 | if (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
3159 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
3160 | { | |
73fbfcad | 3161 | if (dump_enabled_p ()) |
78c60e3d | 3162 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3163 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
3164 | return false; |
3165 | } | |
9dc3f7de IR |
3166 | |
3167 | op0 = gimple_assign_rhs1 (stmt); | |
24ee1384 | 3168 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
9dc3f7de IR |
3169 | &def_stmt, &def, &dt[0], &vectype)) |
3170 | { | |
73fbfcad | 3171 | if (dump_enabled_p ()) |
78c60e3d | 3172 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3173 | "use not simple.\n"); |
9dc3f7de IR |
3174 | return false; |
3175 | } | |
3176 | /* If op0 is an external or constant def use a vector type with | |
3177 | the same size as the output vector type. */ | |
3178 | if (!vectype) | |
3179 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
3180 | if (vec_stmt) | |
3181 | gcc_assert (vectype); | |
3182 | if (!vectype) | |
3183 | { | |
73fbfcad | 3184 | if (dump_enabled_p ()) |
78c60e3d | 3185 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3186 | "no vectype for scalar type\n"); |
9dc3f7de IR |
3187 | return false; |
3188 | } | |
3189 | ||
3190 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
3191 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
3192 | if (nunits_out != nunits_in) | |
3193 | return false; | |
3194 | ||
3195 | op1 = gimple_assign_rhs2 (stmt); | |
24ee1384 IR |
3196 | if (!vect_is_simple_use_1 (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3197 | &def, &dt[1], &op1_vectype)) | |
9dc3f7de | 3198 | { |
73fbfcad | 3199 | if (dump_enabled_p ()) |
78c60e3d | 3200 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3201 | "use not simple.\n"); |
9dc3f7de IR |
3202 | return false; |
3203 | } | |
3204 | ||
3205 | if (loop_vinfo) | |
3206 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3207 | else | |
3208 | vf = 1; | |
3209 | ||
3210 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3211 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3212 | case of SLP. */ | |
437f4a00 | 3213 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
9dc3f7de IR |
3214 | ncopies = 1; |
3215 | else | |
3216 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
3217 | ||
3218 | gcc_assert (ncopies >= 1); | |
3219 | ||
3220 | /* Determine whether the shift amount is a vector, or scalar. If the | |
3221 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
3222 | ||
49eab32e JJ |
3223 | if (dt[1] == vect_internal_def && !slp_node) |
3224 | scalar_shift_arg = false; | |
3225 | else if (dt[1] == vect_constant_def | |
3226 | || dt[1] == vect_external_def | |
3227 | || dt[1] == vect_internal_def) | |
3228 | { | |
3229 | /* In SLP, need to check whether the shift count is the same, | |
3230 | in loops if it is a constant or invariant, it is always | |
3231 | a scalar shift. */ | |
3232 | if (slp_node) | |
3233 | { | |
9771b263 | 3234 | vec<gimple> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
49eab32e JJ |
3235 | gimple slpstmt; |
3236 | ||
9771b263 | 3237 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
3238 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
3239 | scalar_shift_arg = false; | |
3240 | } | |
3241 | } | |
3242 | else | |
3243 | { | |
73fbfcad | 3244 | if (dump_enabled_p ()) |
78c60e3d | 3245 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3246 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
3247 | return false; |
3248 | } | |
3249 | ||
9dc3f7de | 3250 | /* Vector shifted by vector. */ |
49eab32e | 3251 | if (!scalar_shift_arg) |
9dc3f7de IR |
3252 | { |
3253 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 3254 | if (dump_enabled_p ()) |
78c60e3d | 3255 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3256 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 3257 | |
aa948027 JJ |
3258 | if (!op1_vectype) |
3259 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
3260 | if (op1_vectype == NULL_TREE | |
3261 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 3262 | { |
73fbfcad | 3263 | if (dump_enabled_p ()) |
78c60e3d SS |
3264 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3265 | "unusable type for last operand in" | |
e645e942 | 3266 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
3267 | return false; |
3268 | } | |
9dc3f7de IR |
3269 | } |
3270 | /* See if the machine has a vector shifted by scalar insn and if not | |
3271 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 3272 | else |
9dc3f7de IR |
3273 | { |
3274 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
3275 | if (optab | |
3276 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
3277 | { | |
73fbfcad | 3278 | if (dump_enabled_p ()) |
78c60e3d | 3279 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3280 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
3281 | } |
3282 | else | |
3283 | { | |
3284 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
3285 | if (optab | |
3286 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
3287 | != CODE_FOR_nothing)) | |
3288 | { | |
49eab32e JJ |
3289 | scalar_shift_arg = false; |
3290 | ||
73fbfcad | 3291 | if (dump_enabled_p ()) |
78c60e3d | 3292 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3293 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
3294 | |
3295 | /* Unlike the other binary operators, shifts/rotates have | |
3296 | the rhs being int, instead of the same type as the lhs, | |
3297 | so make sure the scalar is the right type if we are | |
aa948027 | 3298 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
3299 | if (dt[1] == vect_constant_def) |
3300 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
3301 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
3302 | TREE_TYPE (op1))) | |
3303 | { | |
3304 | if (slp_node | |
3305 | && TYPE_MODE (TREE_TYPE (vectype)) | |
3306 | != TYPE_MODE (TREE_TYPE (op1))) | |
3307 | { | |
73fbfcad | 3308 | if (dump_enabled_p ()) |
78c60e3d SS |
3309 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3310 | "unusable type for last operand in" | |
e645e942 | 3311 | " vector/vector shift/rotate.\n"); |
aa948027 JJ |
3312 | return false; |
3313 | } | |
3314 | if (vec_stmt && !slp_node) | |
3315 | { | |
3316 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
3317 | op1 = vect_init_vector (stmt, op1, | |
3318 | TREE_TYPE (vectype), NULL); | |
3319 | } | |
3320 | } | |
9dc3f7de IR |
3321 | } |
3322 | } | |
3323 | } | |
9dc3f7de IR |
3324 | |
3325 | /* Supportable by target? */ | |
3326 | if (!optab) | |
3327 | { | |
73fbfcad | 3328 | if (dump_enabled_p ()) |
78c60e3d | 3329 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3330 | "no optab.\n"); |
9dc3f7de IR |
3331 | return false; |
3332 | } | |
3333 | vec_mode = TYPE_MODE (vectype); | |
3334 | icode = (int) optab_handler (optab, vec_mode); | |
3335 | if (icode == CODE_FOR_nothing) | |
3336 | { | |
73fbfcad | 3337 | if (dump_enabled_p ()) |
78c60e3d | 3338 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3339 | "op not supported by target.\n"); |
9dc3f7de IR |
3340 | /* Check only during analysis. */ |
3341 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
3342 | || (vf < vect_min_worthwhile_factor (code) | |
3343 | && !vec_stmt)) | |
3344 | return false; | |
73fbfcad | 3345 | if (dump_enabled_p ()) |
e645e942 TJ |
3346 | dump_printf_loc (MSG_NOTE, vect_location, |
3347 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
3348 | } |
3349 | ||
3350 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
3351 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
3352 | && vf < vect_min_worthwhile_factor (code) | |
3353 | && !vec_stmt) | |
3354 | { | |
73fbfcad | 3355 | if (dump_enabled_p ()) |
78c60e3d | 3356 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3357 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
3358 | return false; |
3359 | } | |
3360 | ||
3361 | if (!vec_stmt) /* transformation not required. */ | |
3362 | { | |
3363 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 3364 | if (dump_enabled_p ()) |
e645e942 TJ |
3365 | dump_printf_loc (MSG_NOTE, vect_location, |
3366 | "=== vectorizable_shift ===\n"); | |
c3e7ee41 | 3367 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
9dc3f7de IR |
3368 | return true; |
3369 | } | |
3370 | ||
3371 | /** Transform. **/ | |
3372 | ||
73fbfcad | 3373 | if (dump_enabled_p ()) |
78c60e3d | 3374 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3375 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
3376 | |
3377 | /* Handle def. */ | |
3378 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3379 | ||
9dc3f7de IR |
3380 | prev_stmt_info = NULL; |
3381 | for (j = 0; j < ncopies; j++) | |
3382 | { | |
3383 | /* Handle uses. */ | |
3384 | if (j == 0) | |
3385 | { | |
3386 | if (scalar_shift_arg) | |
3387 | { | |
3388 | /* Vector shl and shr insn patterns can be defined with scalar | |
3389 | operand 2 (shift operand). In this case, use constant or loop | |
3390 | invariant op1 directly, without extending it to vector mode | |
3391 | first. */ | |
3392 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
3393 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
3394 | { | |
73fbfcad | 3395 | if (dump_enabled_p ()) |
78c60e3d | 3396 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3397 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 3398 | vec_oprnd1 = op1; |
8930f723 | 3399 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 3400 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
3401 | if (slp_node) |
3402 | { | |
3403 | /* Store vec_oprnd1 for every vector stmt to be created | |
3404 | for SLP_NODE. We check during the analysis that all | |
3405 | the shift arguments are the same. | |
3406 | TODO: Allow different constants for different vector | |
3407 | stmts generated for an SLP instance. */ | |
3408 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 3409 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
3410 | } |
3411 | } | |
3412 | } | |
3413 | ||
3414 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
3415 | (a special case for certain kind of vector shifts); otherwise, | |
3416 | operand 1 should be of a vector type (the usual case). */ | |
3417 | if (vec_oprnd1) | |
3418 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
d092494c | 3419 | slp_node, -1); |
9dc3f7de IR |
3420 | else |
3421 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
d092494c | 3422 | slp_node, -1); |
9dc3f7de IR |
3423 | } |
3424 | else | |
3425 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
3426 | ||
3427 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3428 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 3429 | { |
9771b263 | 3430 | vop1 = vec_oprnds1[i]; |
9dc3f7de IR |
3431 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); |
3432 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3433 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3434 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3435 | if (slp_node) | |
9771b263 | 3436 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
3437 | } |
3438 | ||
3439 | if (slp_node) | |
3440 | continue; | |
3441 | ||
3442 | if (j == 0) | |
3443 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3444 | else | |
3445 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3446 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3447 | } | |
3448 | ||
9771b263 DN |
3449 | vec_oprnds0.release (); |
3450 | vec_oprnds1.release (); | |
9dc3f7de IR |
3451 | |
3452 | return true; | |
3453 | } | |
3454 | ||
3455 | ||
5deb57cb JJ |
3456 | static tree permute_vec_elements (tree, tree, tree, gimple, |
3457 | gimple_stmt_iterator *); | |
3458 | ||
3459 | ||
ebfd146a IR |
3460 | /* Function vectorizable_operation. |
3461 | ||
16949072 RG |
3462 | Check if STMT performs a binary, unary or ternary operation that can |
3463 | be vectorized. | |
b8698a0f | 3464 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
3465 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3466 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3467 | ||
3468 | static bool | |
3469 | vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi, | |
3470 | gimple *vec_stmt, slp_tree slp_node) | |
3471 | { | |
00f07b86 | 3472 | tree vec_dest; |
ebfd146a | 3473 | tree scalar_dest; |
16949072 | 3474 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 3475 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 3476 | tree vectype; |
ebfd146a IR |
3477 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
3478 | enum tree_code code; | |
3479 | enum machine_mode vec_mode; | |
3480 | tree new_temp; | |
3481 | int op_type; | |
00f07b86 | 3482 | optab optab; |
ebfd146a | 3483 | int icode; |
ebfd146a IR |
3484 | tree def; |
3485 | gimple def_stmt; | |
16949072 RG |
3486 | enum vect_def_type dt[3] |
3487 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
ebfd146a IR |
3488 | gimple new_stmt = NULL; |
3489 | stmt_vec_info prev_stmt_info; | |
b690cc0f | 3490 | int nunits_in; |
ebfd146a IR |
3491 | int nunits_out; |
3492 | tree vectype_out; | |
3493 | int ncopies; | |
3494 | int j, i; | |
6e1aa848 DN |
3495 | vec<tree> vec_oprnds0 = vNULL; |
3496 | vec<tree> vec_oprnds1 = vNULL; | |
3497 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 3498 | tree vop0, vop1, vop2; |
a70d6342 IR |
3499 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3500 | int vf; | |
3501 | ||
a70d6342 | 3502 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3503 | return false; |
3504 | ||
8644a673 | 3505 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3506 | return false; |
3507 | ||
3508 | /* Is STMT a vectorizable binary/unary operation? */ | |
3509 | if (!is_gimple_assign (stmt)) | |
3510 | return false; | |
3511 | ||
3512 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3513 | return false; | |
3514 | ||
ebfd146a IR |
3515 | code = gimple_assign_rhs_code (stmt); |
3516 | ||
3517 | /* For pointer addition, we should use the normal plus for | |
3518 | the vector addition. */ | |
3519 | if (code == POINTER_PLUS_EXPR) | |
3520 | code = PLUS_EXPR; | |
3521 | ||
3522 | /* Support only unary or binary operations. */ | |
3523 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 3524 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 3525 | { |
73fbfcad | 3526 | if (dump_enabled_p ()) |
78c60e3d | 3527 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3528 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 3529 | op_type); |
ebfd146a IR |
3530 | return false; |
3531 | } | |
3532 | ||
b690cc0f RG |
3533 | scalar_dest = gimple_assign_lhs (stmt); |
3534 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
3535 | ||
7b7b1813 RG |
3536 | /* Most operations cannot handle bit-precision types without extra |
3537 | truncations. */ | |
3538 | if ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
3539 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
3540 | /* Exception are bitwise binary operations. */ | |
3541 | && code != BIT_IOR_EXPR | |
3542 | && code != BIT_XOR_EXPR | |
3543 | && code != BIT_AND_EXPR) | |
3544 | { | |
73fbfcad | 3545 | if (dump_enabled_p ()) |
78c60e3d | 3546 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3547 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
3548 | return false; |
3549 | } | |
3550 | ||
ebfd146a | 3551 | op0 = gimple_assign_rhs1 (stmt); |
24ee1384 | 3552 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
b690cc0f | 3553 | &def_stmt, &def, &dt[0], &vectype)) |
ebfd146a | 3554 | { |
73fbfcad | 3555 | if (dump_enabled_p ()) |
78c60e3d | 3556 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3557 | "use not simple.\n"); |
ebfd146a IR |
3558 | return false; |
3559 | } | |
b690cc0f RG |
3560 | /* If op0 is an external or constant def use a vector type with |
3561 | the same size as the output vector type. */ | |
3562 | if (!vectype) | |
3563 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
3564 | if (vec_stmt) |
3565 | gcc_assert (vectype); | |
3566 | if (!vectype) | |
3567 | { | |
73fbfcad | 3568 | if (dump_enabled_p ()) |
7d8930a0 | 3569 | { |
78c60e3d SS |
3570 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3571 | "no vectype for scalar type "); | |
3572 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
3573 | TREE_TYPE (op0)); | |
e645e942 | 3574 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
3575 | } |
3576 | ||
3577 | return false; | |
3578 | } | |
b690cc0f RG |
3579 | |
3580 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
3581 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
3582 | if (nunits_out != nunits_in) | |
3583 | return false; | |
ebfd146a | 3584 | |
16949072 | 3585 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
3586 | { |
3587 | op1 = gimple_assign_rhs2 (stmt); | |
24ee1384 IR |
3588 | if (!vect_is_simple_use (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3589 | &def, &dt[1])) | |
ebfd146a | 3590 | { |
73fbfcad | 3591 | if (dump_enabled_p ()) |
78c60e3d | 3592 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3593 | "use not simple.\n"); |
ebfd146a IR |
3594 | return false; |
3595 | } | |
3596 | } | |
16949072 RG |
3597 | if (op_type == ternary_op) |
3598 | { | |
3599 | op2 = gimple_assign_rhs3 (stmt); | |
24ee1384 IR |
3600 | if (!vect_is_simple_use (op2, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3601 | &def, &dt[2])) | |
16949072 | 3602 | { |
73fbfcad | 3603 | if (dump_enabled_p ()) |
78c60e3d | 3604 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3605 | "use not simple.\n"); |
16949072 RG |
3606 | return false; |
3607 | } | |
3608 | } | |
ebfd146a | 3609 | |
b690cc0f RG |
3610 | if (loop_vinfo) |
3611 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3612 | else | |
3613 | vf = 1; | |
3614 | ||
3615 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 3616 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 3617 | case of SLP. */ |
437f4a00 | 3618 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
b690cc0f RG |
3619 | ncopies = 1; |
3620 | else | |
3621 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
3622 | ||
3623 | gcc_assert (ncopies >= 1); | |
3624 | ||
9dc3f7de | 3625 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
3626 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
3627 | || code == RROTATE_EXPR) | |
9dc3f7de | 3628 | return false; |
ebfd146a | 3629 | |
ebfd146a | 3630 | /* Supportable by target? */ |
00f07b86 RH |
3631 | |
3632 | vec_mode = TYPE_MODE (vectype); | |
3633 | if (code == MULT_HIGHPART_EXPR) | |
ebfd146a | 3634 | { |
00f07b86 | 3635 | if (can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype))) |
dee54b6e | 3636 | icode = LAST_INSN_CODE; |
00f07b86 RH |
3637 | else |
3638 | icode = CODE_FOR_nothing; | |
ebfd146a | 3639 | } |
00f07b86 RH |
3640 | else |
3641 | { | |
3642 | optab = optab_for_tree_code (code, vectype, optab_default); | |
3643 | if (!optab) | |
5deb57cb | 3644 | { |
73fbfcad | 3645 | if (dump_enabled_p ()) |
78c60e3d | 3646 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3647 | "no optab.\n"); |
00f07b86 | 3648 | return false; |
5deb57cb | 3649 | } |
00f07b86 | 3650 | icode = (int) optab_handler (optab, vec_mode); |
5deb57cb JJ |
3651 | } |
3652 | ||
ebfd146a IR |
3653 | if (icode == CODE_FOR_nothing) |
3654 | { | |
73fbfcad | 3655 | if (dump_enabled_p ()) |
78c60e3d | 3656 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3657 | "op not supported by target.\n"); |
ebfd146a IR |
3658 | /* Check only during analysis. */ |
3659 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5deb57cb | 3660 | || (!vec_stmt && vf < vect_min_worthwhile_factor (code))) |
ebfd146a | 3661 | return false; |
73fbfcad | 3662 | if (dump_enabled_p ()) |
e645e942 TJ |
3663 | dump_printf_loc (MSG_NOTE, vect_location, |
3664 | "proceeding using word mode.\n"); | |
383d9c83 IR |
3665 | } |
3666 | ||
4a00c761 | 3667 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
3668 | if (!VECTOR_MODE_P (vec_mode) |
3669 | && !vec_stmt | |
3670 | && vf < vect_min_worthwhile_factor (code)) | |
7d8930a0 | 3671 | { |
73fbfcad | 3672 | if (dump_enabled_p ()) |
78c60e3d | 3673 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3674 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 3675 | return false; |
7d8930a0 | 3676 | } |
ebfd146a | 3677 | |
ebfd146a IR |
3678 | if (!vec_stmt) /* transformation not required. */ |
3679 | { | |
4a00c761 | 3680 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 3681 | if (dump_enabled_p ()) |
78c60e3d | 3682 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3683 | "=== vectorizable_operation ===\n"); |
c3e7ee41 | 3684 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
3685 | return true; |
3686 | } | |
3687 | ||
3688 | /** Transform. **/ | |
3689 | ||
73fbfcad | 3690 | if (dump_enabled_p ()) |
78c60e3d | 3691 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3692 | "transform binary/unary operation.\n"); |
383d9c83 | 3693 | |
ebfd146a | 3694 | /* Handle def. */ |
00f07b86 | 3695 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 3696 | |
ebfd146a IR |
3697 | /* In case the vectorization factor (VF) is bigger than the number |
3698 | of elements that we can fit in a vectype (nunits), we have to generate | |
3699 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
3700 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
3701 | from one copy of the vector stmt to the next, in the field | |
3702 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
3703 | stages to find the correct vector defs to be used when vectorizing | |
3704 | stmts that use the defs of the current stmt. The example below | |
3705 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
3706 | we need to create 4 vectorized stmts): | |
3707 | ||
3708 | before vectorization: | |
3709 | RELATED_STMT VEC_STMT | |
3710 | S1: x = memref - - | |
3711 | S2: z = x + 1 - - | |
3712 | ||
3713 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
3714 | there): | |
3715 | RELATED_STMT VEC_STMT | |
3716 | VS1_0: vx0 = memref0 VS1_1 - | |
3717 | VS1_1: vx1 = memref1 VS1_2 - | |
3718 | VS1_2: vx2 = memref2 VS1_3 - | |
3719 | VS1_3: vx3 = memref3 - - | |
3720 | S1: x = load - VS1_0 | |
3721 | S2: z = x + 1 - - | |
3722 | ||
3723 | step2: vectorize stmt S2 (done here): | |
3724 | To vectorize stmt S2 we first need to find the relevant vector | |
3725 | def for the first operand 'x'. This is, as usual, obtained from | |
3726 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
3727 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
3728 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
3729 | the vector stmt VS2_0, and as usual, record it in the | |
3730 | STMT_VINFO_VEC_STMT of stmt S2. | |
3731 | When creating the second copy (VS2_1), we obtain the relevant vector | |
3732 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
3733 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
3734 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
3735 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
3736 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
3737 | chain of stmts and pointers: | |
3738 | RELATED_STMT VEC_STMT | |
3739 | VS1_0: vx0 = memref0 VS1_1 - | |
3740 | VS1_1: vx1 = memref1 VS1_2 - | |
3741 | VS1_2: vx2 = memref2 VS1_3 - | |
3742 | VS1_3: vx3 = memref3 - - | |
3743 | S1: x = load - VS1_0 | |
3744 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
3745 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
3746 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
3747 | VS2_3: vz3 = vx3 + v1 - - | |
3748 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
3749 | |
3750 | prev_stmt_info = NULL; | |
3751 | for (j = 0; j < ncopies; j++) | |
3752 | { | |
3753 | /* Handle uses. */ | |
3754 | if (j == 0) | |
4a00c761 JJ |
3755 | { |
3756 | if (op_type == binary_op || op_type == ternary_op) | |
3757 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
3758 | slp_node, -1); | |
3759 | else | |
3760 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
3761 | slp_node, -1); | |
3762 | if (op_type == ternary_op) | |
36ba4aae | 3763 | { |
9771b263 DN |
3764 | vec_oprnds2.create (1); |
3765 | vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2, | |
3766 | stmt, | |
3767 | NULL)); | |
36ba4aae | 3768 | } |
4a00c761 | 3769 | } |
ebfd146a | 3770 | else |
4a00c761 JJ |
3771 | { |
3772 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
3773 | if (op_type == ternary_op) | |
3774 | { | |
9771b263 DN |
3775 | tree vec_oprnd = vec_oprnds2.pop (); |
3776 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
3777 | vec_oprnd)); | |
4a00c761 JJ |
3778 | } |
3779 | } | |
3780 | ||
3781 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3782 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 3783 | { |
4a00c761 | 3784 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 3785 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 3786 | vop2 = ((op_type == ternary_op) |
9771b263 | 3787 | ? vec_oprnds2[i] : NULL_TREE); |
73804b12 RG |
3788 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, |
3789 | vop0, vop1, vop2); | |
4a00c761 JJ |
3790 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3791 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3792 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3793 | if (slp_node) | |
9771b263 | 3794 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
3795 | } |
3796 | ||
4a00c761 JJ |
3797 | if (slp_node) |
3798 | continue; | |
3799 | ||
3800 | if (j == 0) | |
3801 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3802 | else | |
3803 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3804 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
3805 | } |
3806 | ||
9771b263 DN |
3807 | vec_oprnds0.release (); |
3808 | vec_oprnds1.release (); | |
3809 | vec_oprnds2.release (); | |
ebfd146a | 3810 | |
ebfd146a IR |
3811 | return true; |
3812 | } | |
3813 | ||
c716e67f XDL |
3814 | /* A helper function to ensure data reference DR's base alignment |
3815 | for STMT_INFO. */ | |
3816 | ||
3817 | static void | |
3818 | ensure_base_align (stmt_vec_info stmt_info, struct data_reference *dr) | |
3819 | { | |
3820 | if (!dr->aux) | |
3821 | return; | |
3822 | ||
3823 | if (((dataref_aux *)dr->aux)->base_misaligned) | |
3824 | { | |
3825 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3826 | tree base_decl = ((dataref_aux *)dr->aux)->base_decl; | |
3827 | ||
3828 | DECL_ALIGN (base_decl) = TYPE_ALIGN (vectype); | |
3829 | DECL_USER_ALIGN (base_decl) = 1; | |
3830 | ((dataref_aux *)dr->aux)->base_misaligned = false; | |
3831 | } | |
3832 | } | |
3833 | ||
ebfd146a IR |
3834 | |
3835 | /* Function vectorizable_store. | |
3836 | ||
b8698a0f L |
3837 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
3838 | can be vectorized. | |
3839 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
3840 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3841 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3842 | ||
3843 | static bool | |
3844 | vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
c716e67f | 3845 | slp_tree slp_node) |
ebfd146a IR |
3846 | { |
3847 | tree scalar_dest; | |
3848 | tree data_ref; | |
3849 | tree op; | |
3850 | tree vec_oprnd = NULL_TREE; | |
3851 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3852 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
3853 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
272c6793 | 3854 | tree elem_type; |
ebfd146a | 3855 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 3856 | struct loop *loop = NULL; |
ebfd146a IR |
3857 | enum machine_mode vec_mode; |
3858 | tree dummy; | |
3859 | enum dr_alignment_support alignment_support_scheme; | |
3860 | tree def; | |
3861 | gimple def_stmt; | |
3862 | enum vect_def_type dt; | |
3863 | stmt_vec_info prev_stmt_info = NULL; | |
3864 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 3865 | tree dataref_offset = NULL_TREE; |
fef4d2b3 | 3866 | gimple ptr_incr = NULL; |
ebfd146a IR |
3867 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
3868 | int ncopies; | |
3869 | int j; | |
3870 | gimple next_stmt, first_stmt = NULL; | |
0d0293ac | 3871 | bool grouped_store = false; |
272c6793 | 3872 | bool store_lanes_p = false; |
ebfd146a | 3873 | unsigned int group_size, i; |
6e1aa848 DN |
3874 | vec<tree> dr_chain = vNULL; |
3875 | vec<tree> oprnds = vNULL; | |
3876 | vec<tree> result_chain = vNULL; | |
ebfd146a | 3877 | bool inv_p; |
6e1aa848 | 3878 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 3879 | bool slp = (slp_node != NULL); |
ebfd146a | 3880 | unsigned int vec_num; |
a70d6342 | 3881 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
272c6793 | 3882 | tree aggr_type; |
a70d6342 IR |
3883 | |
3884 | if (loop_vinfo) | |
3885 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
3886 | |
3887 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3888 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3889 | case of SLP. */ | |
437f4a00 | 3890 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
3891 | ncopies = 1; |
3892 | else | |
3893 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3894 | ||
3895 | gcc_assert (ncopies >= 1); | |
3896 | ||
3897 | /* FORNOW. This restriction should be relaxed. */ | |
a70d6342 | 3898 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) |
ebfd146a | 3899 | { |
73fbfcad | 3900 | if (dump_enabled_p ()) |
78c60e3d | 3901 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3902 | "multiple types in nested loop.\n"); |
ebfd146a IR |
3903 | return false; |
3904 | } | |
3905 | ||
a70d6342 | 3906 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3907 | return false; |
3908 | ||
8644a673 | 3909 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3910 | return false; |
3911 | ||
3912 | /* Is vectorizable store? */ | |
3913 | ||
3914 | if (!is_gimple_assign (stmt)) | |
3915 | return false; | |
3916 | ||
3917 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
3918 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
3919 | && is_pattern_stmt_p (stmt_info)) | |
3920 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 3921 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 3922 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 3923 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
3924 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
3925 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
3926 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
3927 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
3928 | return false; |
3929 | ||
3930 | gcc_assert (gimple_assign_single_p (stmt)); | |
3931 | op = gimple_assign_rhs1 (stmt); | |
24ee1384 IR |
3932 | if (!vect_is_simple_use (op, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3933 | &def, &dt)) | |
ebfd146a | 3934 | { |
73fbfcad | 3935 | if (dump_enabled_p ()) |
78c60e3d | 3936 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3937 | "use not simple.\n"); |
ebfd146a IR |
3938 | return false; |
3939 | } | |
3940 | ||
272c6793 | 3941 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 3942 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 3943 | |
ebfd146a IR |
3944 | /* FORNOW. In some cases can vectorize even if data-type not supported |
3945 | (e.g. - array initialization with 0). */ | |
947131ba | 3946 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
3947 | return false; |
3948 | ||
3949 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
3950 | return false; | |
3951 | ||
a7ce6ec3 RG |
3952 | if (tree_int_cst_compare (loop && nested_in_vect_loop_p (loop, stmt) |
3953 | ? STMT_VINFO_DR_STEP (stmt_info) : DR_STEP (dr), | |
3954 | size_zero_node) < 0) | |
a1e53f3f | 3955 | { |
73fbfcad | 3956 | if (dump_enabled_p ()) |
78c60e3d | 3957 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3958 | "negative step for store.\n"); |
a1e53f3f L |
3959 | return false; |
3960 | } | |
3961 | ||
0d0293ac | 3962 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 3963 | { |
0d0293ac | 3964 | grouped_store = true; |
e14c1050 | 3965 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
3966 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
3967 | { | |
e14c1050 | 3968 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
3969 | if (vect_store_lanes_supported (vectype, group_size)) |
3970 | store_lanes_p = true; | |
0d0293ac | 3971 | else if (!vect_grouped_store_supported (vectype, group_size)) |
b602d918 RS |
3972 | return false; |
3973 | } | |
b8698a0f | 3974 | |
ebfd146a IR |
3975 | if (first_stmt == stmt) |
3976 | { | |
3977 | /* STMT is the leader of the group. Check the operands of all the | |
3978 | stmts of the group. */ | |
e14c1050 | 3979 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
3980 | while (next_stmt) |
3981 | { | |
3982 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
3983 | op = gimple_assign_rhs1 (next_stmt); | |
24ee1384 IR |
3984 | if (!vect_is_simple_use (op, next_stmt, loop_vinfo, bb_vinfo, |
3985 | &def_stmt, &def, &dt)) | |
ebfd146a | 3986 | { |
73fbfcad | 3987 | if (dump_enabled_p ()) |
78c60e3d | 3988 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3989 | "use not simple.\n"); |
ebfd146a IR |
3990 | return false; |
3991 | } | |
e14c1050 | 3992 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
3993 | } |
3994 | } | |
3995 | } | |
3996 | ||
3997 | if (!vec_stmt) /* transformation not required. */ | |
3998 | { | |
3999 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
92345349 BS |
4000 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, |
4001 | NULL, NULL, NULL); | |
ebfd146a IR |
4002 | return true; |
4003 | } | |
4004 | ||
4005 | /** Transform. **/ | |
4006 | ||
c716e67f XDL |
4007 | ensure_base_align (stmt_info, dr); |
4008 | ||
0d0293ac | 4009 | if (grouped_store) |
ebfd146a IR |
4010 | { |
4011 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 4012 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 4013 | |
e14c1050 | 4014 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
4015 | |
4016 | /* FORNOW */ | |
a70d6342 | 4017 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
4018 | |
4019 | /* We vectorize all the stmts of the interleaving group when we | |
4020 | reach the last stmt in the group. */ | |
e14c1050 IR |
4021 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
4022 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
4023 | && !slp) |
4024 | { | |
4025 | *vec_stmt = NULL; | |
4026 | return true; | |
4027 | } | |
4028 | ||
4029 | if (slp) | |
4b5caab7 | 4030 | { |
0d0293ac | 4031 | grouped_store = false; |
4b5caab7 IR |
4032 | /* VEC_NUM is the number of vect stmts to be created for this |
4033 | group. */ | |
4034 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 4035 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
4b5caab7 | 4036 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 4037 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 4038 | } |
ebfd146a | 4039 | else |
4b5caab7 IR |
4040 | /* VEC_NUM is the number of vect stmts to be created for this |
4041 | group. */ | |
ebfd146a IR |
4042 | vec_num = group_size; |
4043 | } | |
b8698a0f | 4044 | else |
ebfd146a IR |
4045 | { |
4046 | first_stmt = stmt; | |
4047 | first_dr = dr; | |
4048 | group_size = vec_num = 1; | |
ebfd146a | 4049 | } |
b8698a0f | 4050 | |
73fbfcad | 4051 | if (dump_enabled_p ()) |
78c60e3d | 4052 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4053 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 4054 | |
9771b263 DN |
4055 | dr_chain.create (group_size); |
4056 | oprnds.create (group_size); | |
ebfd146a | 4057 | |
720f5239 | 4058 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 4059 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
4060 | /* Targets with store-lane instructions must not require explicit |
4061 | realignment. */ | |
4062 | gcc_assert (!store_lanes_p | |
4063 | || alignment_support_scheme == dr_aligned | |
4064 | || alignment_support_scheme == dr_unaligned_supported); | |
4065 | ||
4066 | if (store_lanes_p) | |
4067 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
4068 | else | |
4069 | aggr_type = vectype; | |
ebfd146a IR |
4070 | |
4071 | /* In case the vectorization factor (VF) is bigger than the number | |
4072 | of elements that we can fit in a vectype (nunits), we have to generate | |
4073 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 4074 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
4075 | vect_get_vec_def_for_copy_stmt. */ |
4076 | ||
0d0293ac | 4077 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
4078 | |
4079 | S1: &base + 2 = x2 | |
4080 | S2: &base = x0 | |
4081 | S3: &base + 1 = x1 | |
4082 | S4: &base + 3 = x3 | |
4083 | ||
4084 | We create vectorized stores starting from base address (the access of the | |
4085 | first stmt in the chain (S2 in the above example), when the last store stmt | |
4086 | of the chain (S4) is reached: | |
4087 | ||
4088 | VS1: &base = vx2 | |
4089 | VS2: &base + vec_size*1 = vx0 | |
4090 | VS3: &base + vec_size*2 = vx1 | |
4091 | VS4: &base + vec_size*3 = vx3 | |
4092 | ||
4093 | Then permutation statements are generated: | |
4094 | ||
3fcc1b55 JJ |
4095 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
4096 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 4097 | ... |
b8698a0f | 4098 | |
ebfd146a IR |
4099 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
4100 | (the order of the data-refs in the output of vect_permute_store_chain | |
4101 | corresponds to the order of scalar stmts in the interleaving chain - see | |
4102 | the documentation of vect_permute_store_chain()). | |
4103 | ||
4104 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 4105 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 4106 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 4107 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
4108 | */ |
4109 | ||
4110 | prev_stmt_info = NULL; | |
4111 | for (j = 0; j < ncopies; j++) | |
4112 | { | |
4113 | gimple new_stmt; | |
ebfd146a IR |
4114 | |
4115 | if (j == 0) | |
4116 | { | |
4117 | if (slp) | |
4118 | { | |
4119 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c IR |
4120 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
4121 | NULL, slp_node, -1); | |
ebfd146a | 4122 | |
9771b263 | 4123 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
4124 | } |
4125 | else | |
4126 | { | |
b8698a0f L |
4127 | /* For interleaved stores we collect vectorized defs for all the |
4128 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
4129 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
4130 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
4131 | ||
0d0293ac | 4132 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 4133 | OPRNDS are of size 1. */ |
b8698a0f | 4134 | next_stmt = first_stmt; |
ebfd146a IR |
4135 | for (i = 0; i < group_size; i++) |
4136 | { | |
b8698a0f L |
4137 | /* Since gaps are not supported for interleaved stores, |
4138 | GROUP_SIZE is the exact number of stmts in the chain. | |
4139 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
4140 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
4141 | iteration of the loop will be executed. */ |
4142 | gcc_assert (next_stmt | |
4143 | && gimple_assign_single_p (next_stmt)); | |
4144 | op = gimple_assign_rhs1 (next_stmt); | |
4145 | ||
b8698a0f | 4146 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt, |
ebfd146a | 4147 | NULL); |
9771b263 DN |
4148 | dr_chain.quick_push (vec_oprnd); |
4149 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 4150 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
4151 | } |
4152 | } | |
4153 | ||
4154 | /* We should have catched mismatched types earlier. */ | |
4155 | gcc_assert (useless_type_conversion_p (vectype, | |
4156 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
4157 | bool simd_lane_access_p |
4158 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
4159 | if (simd_lane_access_p | |
4160 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
4161 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
4162 | && integer_zerop (DR_OFFSET (first_dr)) | |
4163 | && integer_zerop (DR_INIT (first_dr)) | |
4164 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
4165 | get_alias_set (DR_REF (first_dr)))) | |
4166 | { | |
4167 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
4168 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
4169 | (DR_REF (first_dr)), 0); | |
8928eff3 | 4170 | inv_p = false; |
74bf76ed JJ |
4171 | } |
4172 | else | |
4173 | dataref_ptr | |
4174 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
4175 | simd_lane_access_p ? loop : NULL, | |
4176 | NULL_TREE, &dummy, gsi, &ptr_incr, | |
4177 | simd_lane_access_p, &inv_p); | |
a70d6342 | 4178 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 4179 | } |
b8698a0f | 4180 | else |
ebfd146a | 4181 | { |
b8698a0f L |
4182 | /* For interleaved stores we created vectorized defs for all the |
4183 | defs stored in OPRNDS in the previous iteration (previous copy). | |
4184 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
4185 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
4186 | next copy. | |
0d0293ac | 4187 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
4188 | OPRNDS are of size 1. */ |
4189 | for (i = 0; i < group_size; i++) | |
4190 | { | |
9771b263 | 4191 | op = oprnds[i]; |
24ee1384 IR |
4192 | vect_is_simple_use (op, NULL, loop_vinfo, bb_vinfo, &def_stmt, |
4193 | &def, &dt); | |
b8698a0f | 4194 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
4195 | dr_chain[i] = vec_oprnd; |
4196 | oprnds[i] = vec_oprnd; | |
ebfd146a | 4197 | } |
74bf76ed JJ |
4198 | if (dataref_offset) |
4199 | dataref_offset | |
4200 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
4201 | TYPE_SIZE_UNIT (aggr_type)); | |
4202 | else | |
4203 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
4204 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
4205 | } |
4206 | ||
272c6793 | 4207 | if (store_lanes_p) |
ebfd146a | 4208 | { |
272c6793 | 4209 | tree vec_array; |
267d3070 | 4210 | |
272c6793 RS |
4211 | /* Combine all the vectors into an array. */ |
4212 | vec_array = create_vector_array (vectype, vec_num); | |
4213 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 4214 | { |
9771b263 | 4215 | vec_oprnd = dr_chain[i]; |
272c6793 | 4216 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 4217 | } |
b8698a0f | 4218 | |
272c6793 RS |
4219 | /* Emit: |
4220 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
4221 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
4222 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
4223 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 | 4224 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
4225 | } |
4226 | else | |
4227 | { | |
4228 | new_stmt = NULL; | |
0d0293ac | 4229 | if (grouped_store) |
272c6793 | 4230 | { |
b6b9227d JJ |
4231 | if (j == 0) |
4232 | result_chain.create (group_size); | |
272c6793 RS |
4233 | /* Permute. */ |
4234 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
4235 | &result_chain); | |
4236 | } | |
c2d7ab2a | 4237 | |
272c6793 RS |
4238 | next_stmt = first_stmt; |
4239 | for (i = 0; i < vec_num; i++) | |
4240 | { | |
644ffefd | 4241 | unsigned align, misalign; |
272c6793 RS |
4242 | |
4243 | if (i > 0) | |
4244 | /* Bump the vector pointer. */ | |
4245 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
4246 | stmt, NULL_TREE); | |
4247 | ||
4248 | if (slp) | |
9771b263 | 4249 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
4250 | else if (grouped_store) |
4251 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 4252 | vect_permute_store_chain(). */ |
9771b263 | 4253 | vec_oprnd = result_chain[i]; |
272c6793 RS |
4254 | |
4255 | data_ref = build2 (MEM_REF, TREE_TYPE (vec_oprnd), dataref_ptr, | |
74bf76ed JJ |
4256 | dataref_offset |
4257 | ? dataref_offset | |
4258 | : build_int_cst (reference_alias_ptr_type | |
4259 | (DR_REF (first_dr)), 0)); | |
644ffefd | 4260 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 | 4261 | if (aligned_access_p (first_dr)) |
644ffefd | 4262 | misalign = 0; |
272c6793 RS |
4263 | else if (DR_MISALIGNMENT (first_dr) == -1) |
4264 | { | |
4265 | TREE_TYPE (data_ref) | |
4266 | = build_aligned_type (TREE_TYPE (data_ref), | |
4267 | TYPE_ALIGN (elem_type)); | |
644ffefd MJ |
4268 | align = TYPE_ALIGN_UNIT (elem_type); |
4269 | misalign = 0; | |
272c6793 RS |
4270 | } |
4271 | else | |
4272 | { | |
4273 | TREE_TYPE (data_ref) | |
4274 | = build_aligned_type (TREE_TYPE (data_ref), | |
4275 | TYPE_ALIGN (elem_type)); | |
644ffefd | 4276 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 4277 | } |
74bf76ed JJ |
4278 | if (dataref_offset == NULL_TREE) |
4279 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
4280 | misalign); | |
c2d7ab2a | 4281 | |
272c6793 RS |
4282 | /* Arguments are ready. Create the new vector stmt. */ |
4283 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
4284 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
4285 | |
4286 | if (slp) | |
4287 | continue; | |
4288 | ||
e14c1050 | 4289 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
4290 | if (!next_stmt) |
4291 | break; | |
4292 | } | |
ebfd146a | 4293 | } |
1da0876c RS |
4294 | if (!slp) |
4295 | { | |
4296 | if (j == 0) | |
4297 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4298 | else | |
4299 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4300 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4301 | } | |
ebfd146a IR |
4302 | } |
4303 | ||
9771b263 DN |
4304 | dr_chain.release (); |
4305 | oprnds.release (); | |
4306 | result_chain.release (); | |
4307 | vec_oprnds.release (); | |
ebfd146a IR |
4308 | |
4309 | return true; | |
4310 | } | |
4311 | ||
aec7ae7d JJ |
4312 | /* Given a vector type VECTYPE and permutation SEL returns |
4313 | the VECTOR_CST mask that implements the permutation of the | |
4314 | vector elements. If that is impossible to do, returns NULL. */ | |
a1e53f3f | 4315 | |
3fcc1b55 JJ |
4316 | tree |
4317 | vect_gen_perm_mask (tree vectype, unsigned char *sel) | |
a1e53f3f | 4318 | { |
d2a12ae7 | 4319 | tree mask_elt_type, mask_type, mask_vec, *mask_elts; |
2635892a | 4320 | int i, nunits; |
a1e53f3f | 4321 | |
22e4dee7 | 4322 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 RH |
4323 | |
4324 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) | |
a1e53f3f L |
4325 | return NULL; |
4326 | ||
96f9265a RG |
4327 | mask_elt_type = lang_hooks.types.type_for_mode |
4328 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
22e4dee7 | 4329 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 4330 | |
d2a12ae7 | 4331 | mask_elts = XALLOCAVEC (tree, nunits); |
aec7ae7d | 4332 | for (i = nunits - 1; i >= 0; i--) |
d2a12ae7 RG |
4333 | mask_elts[i] = build_int_cst (mask_elt_type, sel[i]); |
4334 | mask_vec = build_vector (mask_type, mask_elts); | |
a1e53f3f | 4335 | |
2635892a | 4336 | return mask_vec; |
a1e53f3f L |
4337 | } |
4338 | ||
aec7ae7d JJ |
4339 | /* Given a vector type VECTYPE returns the VECTOR_CST mask that implements |
4340 | reversal of the vector elements. If that is impossible to do, | |
4341 | returns NULL. */ | |
4342 | ||
4343 | static tree | |
4344 | perm_mask_for_reverse (tree vectype) | |
4345 | { | |
4346 | int i, nunits; | |
4347 | unsigned char *sel; | |
4348 | ||
4349 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4350 | sel = XALLOCAVEC (unsigned char, nunits); | |
4351 | ||
4352 | for (i = 0; i < nunits; ++i) | |
4353 | sel[i] = nunits - 1 - i; | |
4354 | ||
3fcc1b55 | 4355 | return vect_gen_perm_mask (vectype, sel); |
aec7ae7d JJ |
4356 | } |
4357 | ||
4358 | /* Given a vector variable X and Y, that was generated for the scalar | |
4359 | STMT, generate instructions to permute the vector elements of X and Y | |
4360 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
4361 | permuted vector variable. */ | |
a1e53f3f L |
4362 | |
4363 | static tree | |
aec7ae7d JJ |
4364 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple stmt, |
4365 | gimple_stmt_iterator *gsi) | |
a1e53f3f L |
4366 | { |
4367 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 4368 | tree perm_dest, data_ref; |
a1e53f3f L |
4369 | gimple perm_stmt; |
4370 | ||
a1e53f3f | 4371 | perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype); |
aec7ae7d | 4372 | data_ref = make_ssa_name (perm_dest, NULL); |
a1e53f3f L |
4373 | |
4374 | /* Generate the permute statement. */ | |
73804b12 RG |
4375 | perm_stmt = gimple_build_assign_with_ops (VEC_PERM_EXPR, data_ref, |
4376 | x, y, mask_vec); | |
a1e53f3f L |
4377 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
4378 | ||
4379 | return data_ref; | |
4380 | } | |
4381 | ||
ebfd146a IR |
4382 | /* vectorizable_load. |
4383 | ||
b8698a0f L |
4384 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
4385 | can be vectorized. | |
4386 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4387 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4388 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4389 | ||
4390 | static bool | |
4391 | vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
c716e67f | 4392 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
4393 | { |
4394 | tree scalar_dest; | |
4395 | tree vec_dest = NULL; | |
4396 | tree data_ref = NULL; | |
4397 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 4398 | stmt_vec_info prev_stmt_info; |
ebfd146a | 4399 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 4400 | struct loop *loop = NULL; |
ebfd146a | 4401 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 4402 | bool nested_in_vect_loop = false; |
c716e67f | 4403 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
ebfd146a | 4404 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
272c6793 | 4405 | tree elem_type; |
ebfd146a | 4406 | tree new_temp; |
947131ba | 4407 | enum machine_mode mode; |
ebfd146a IR |
4408 | gimple new_stmt = NULL; |
4409 | tree dummy; | |
4410 | enum dr_alignment_support alignment_support_scheme; | |
4411 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 4412 | tree dataref_offset = NULL_TREE; |
fef4d2b3 | 4413 | gimple ptr_incr = NULL; |
ebfd146a IR |
4414 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
4415 | int ncopies; | |
a64b9c26 | 4416 | int i, j, group_size, group_gap; |
ebfd146a IR |
4417 | tree msq = NULL_TREE, lsq; |
4418 | tree offset = NULL_TREE; | |
4419 | tree realignment_token = NULL_TREE; | |
4420 | gimple phi = NULL; | |
6e1aa848 | 4421 | vec<tree> dr_chain = vNULL; |
0d0293ac | 4422 | bool grouped_load = false; |
272c6793 | 4423 | bool load_lanes_p = false; |
ebfd146a | 4424 | gimple first_stmt; |
ebfd146a | 4425 | bool inv_p; |
319e6439 | 4426 | bool negative = false; |
ebfd146a IR |
4427 | bool compute_in_loop = false; |
4428 | struct loop *at_loop; | |
4429 | int vec_num; | |
4430 | bool slp = (slp_node != NULL); | |
4431 | bool slp_perm = false; | |
4432 | enum tree_code code; | |
a70d6342 IR |
4433 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
4434 | int vf; | |
272c6793 | 4435 | tree aggr_type; |
aec7ae7d JJ |
4436 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; |
4437 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
4438 | int gather_scale = 1; | |
4439 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
a70d6342 IR |
4440 | |
4441 | if (loop_vinfo) | |
4442 | { | |
4443 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
4444 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
4445 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4446 | } | |
4447 | else | |
3533e503 | 4448 | vf = 1; |
ebfd146a IR |
4449 | |
4450 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4451 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 4452 | case of SLP. */ |
437f4a00 | 4453 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
4454 | ncopies = 1; |
4455 | else | |
4456 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4457 | ||
4458 | gcc_assert (ncopies >= 1); | |
4459 | ||
4460 | /* FORNOW. This restriction should be relaxed. */ | |
4461 | if (nested_in_vect_loop && ncopies > 1) | |
4462 | { | |
73fbfcad | 4463 | if (dump_enabled_p ()) |
78c60e3d | 4464 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4465 | "multiple types in nested loop.\n"); |
ebfd146a IR |
4466 | return false; |
4467 | } | |
4468 | ||
a70d6342 | 4469 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4470 | return false; |
4471 | ||
8644a673 | 4472 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4473 | return false; |
4474 | ||
4475 | /* Is vectorizable load? */ | |
4476 | if (!is_gimple_assign (stmt)) | |
4477 | return false; | |
4478 | ||
4479 | scalar_dest = gimple_assign_lhs (stmt); | |
4480 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4481 | return false; | |
4482 | ||
4483 | code = gimple_assign_rhs_code (stmt); | |
4484 | if (code != ARRAY_REF | |
38000232 | 4485 | && code != BIT_FIELD_REF |
ebfd146a | 4486 | && code != INDIRECT_REF |
e9dbe7bb IR |
4487 | && code != COMPONENT_REF |
4488 | && code != IMAGPART_EXPR | |
70f34814 | 4489 | && code != REALPART_EXPR |
42373e0b RG |
4490 | && code != MEM_REF |
4491 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
ebfd146a IR |
4492 | return false; |
4493 | ||
4494 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
4495 | return false; | |
4496 | ||
7b7b1813 | 4497 | elem_type = TREE_TYPE (vectype); |
947131ba | 4498 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
4499 | |
4500 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
4501 | (e.g. - data copies). */ | |
947131ba | 4502 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 4503 | { |
73fbfcad | 4504 | if (dump_enabled_p ()) |
78c60e3d | 4505 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4506 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
4507 | return false; |
4508 | } | |
4509 | ||
ebfd146a | 4510 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 4511 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 4512 | { |
0d0293ac | 4513 | grouped_load = true; |
ebfd146a | 4514 | /* FORNOW */ |
aec7ae7d | 4515 | gcc_assert (! nested_in_vect_loop && !STMT_VINFO_GATHER_P (stmt_info)); |
ebfd146a | 4516 | |
e14c1050 | 4517 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
4518 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
4519 | { | |
e14c1050 | 4520 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
4521 | if (vect_load_lanes_supported (vectype, group_size)) |
4522 | load_lanes_p = true; | |
0d0293ac | 4523 | else if (!vect_grouped_load_supported (vectype, group_size)) |
b602d918 RS |
4524 | return false; |
4525 | } | |
ebfd146a IR |
4526 | } |
4527 | ||
a1e53f3f | 4528 | |
aec7ae7d JJ |
4529 | if (STMT_VINFO_GATHER_P (stmt_info)) |
4530 | { | |
4531 | gimple def_stmt; | |
4532 | tree def; | |
4533 | gather_decl = vect_check_gather (stmt, loop_vinfo, &gather_base, | |
4534 | &gather_off, &gather_scale); | |
4535 | gcc_assert (gather_decl); | |
24ee1384 | 4536 | if (!vect_is_simple_use_1 (gather_off, NULL, loop_vinfo, bb_vinfo, |
aec7ae7d JJ |
4537 | &def_stmt, &def, &gather_dt, |
4538 | &gather_off_vectype)) | |
4539 | { | |
73fbfcad | 4540 | if (dump_enabled_p ()) |
78c60e3d | 4541 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4542 | "gather index use not simple.\n"); |
aec7ae7d JJ |
4543 | return false; |
4544 | } | |
4545 | } | |
7d75abc8 | 4546 | else if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
14ac6aa2 | 4547 | ; |
319e6439 RG |
4548 | else |
4549 | { | |
4550 | negative = tree_int_cst_compare (nested_in_vect_loop | |
4551 | ? STMT_VINFO_DR_STEP (stmt_info) | |
4552 | : DR_STEP (dr), | |
4553 | size_zero_node) < 0; | |
4554 | if (negative && ncopies > 1) | |
4555 | { | |
73fbfcad | 4556 | if (dump_enabled_p ()) |
78c60e3d | 4557 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4558 | "multiple types with negative step.\n"); |
319e6439 RG |
4559 | return false; |
4560 | } | |
4561 | ||
4562 | if (negative) | |
4563 | { | |
08940f33 RB |
4564 | if (grouped_load) |
4565 | { | |
4566 | if (dump_enabled_p ()) | |
4567 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 TJ |
4568 | "negative step for group load not supported" |
4569 | "\n"); | |
08940f33 RB |
4570 | return false; |
4571 | } | |
319e6439 RG |
4572 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
4573 | if (alignment_support_scheme != dr_aligned | |
4574 | && alignment_support_scheme != dr_unaligned_supported) | |
4575 | { | |
73fbfcad | 4576 | if (dump_enabled_p ()) |
78c60e3d | 4577 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4578 | "negative step but alignment required.\n"); |
319e6439 RG |
4579 | return false; |
4580 | } | |
4581 | if (!perm_mask_for_reverse (vectype)) | |
4582 | { | |
73fbfcad | 4583 | if (dump_enabled_p ()) |
78c60e3d | 4584 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
4585 | "negative step and reversing not supported." |
4586 | "\n"); | |
319e6439 RG |
4587 | return false; |
4588 | } | |
4589 | } | |
7d75abc8 | 4590 | } |
aec7ae7d | 4591 | |
ebfd146a IR |
4592 | if (!vec_stmt) /* transformation not required. */ |
4593 | { | |
4594 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
92345349 | 4595 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, NULL, NULL, NULL); |
ebfd146a IR |
4596 | return true; |
4597 | } | |
4598 | ||
73fbfcad | 4599 | if (dump_enabled_p ()) |
78c60e3d | 4600 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4601 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a IR |
4602 | |
4603 | /** Transform. **/ | |
4604 | ||
c716e67f XDL |
4605 | ensure_base_align (stmt_info, dr); |
4606 | ||
aec7ae7d JJ |
4607 | if (STMT_VINFO_GATHER_P (stmt_info)) |
4608 | { | |
4609 | tree vec_oprnd0 = NULL_TREE, op; | |
4610 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
4611 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
4612 | tree ptr, mask, var, scale, perm_mask = NULL_TREE, prev_res = NULL_TREE; | |
4613 | edge pe = loop_preheader_edge (loop); | |
4614 | gimple_seq seq; | |
4615 | basic_block new_bb; | |
4616 | enum { NARROW, NONE, WIDEN } modifier; | |
4617 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
4618 | ||
4619 | if (nunits == gather_off_nunits) | |
4620 | modifier = NONE; | |
4621 | else if (nunits == gather_off_nunits / 2) | |
4622 | { | |
4623 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
4624 | modifier = WIDEN; | |
4625 | ||
4626 | for (i = 0; i < gather_off_nunits; ++i) | |
4627 | sel[i] = i | nunits; | |
4628 | ||
3fcc1b55 | 4629 | perm_mask = vect_gen_perm_mask (gather_off_vectype, sel); |
aec7ae7d JJ |
4630 | gcc_assert (perm_mask != NULL_TREE); |
4631 | } | |
4632 | else if (nunits == gather_off_nunits * 2) | |
4633 | { | |
4634 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
4635 | modifier = NARROW; | |
4636 | ||
4637 | for (i = 0; i < nunits; ++i) | |
4638 | sel[i] = i < gather_off_nunits | |
4639 | ? i : i + nunits - gather_off_nunits; | |
4640 | ||
3fcc1b55 | 4641 | perm_mask = vect_gen_perm_mask (vectype, sel); |
aec7ae7d JJ |
4642 | gcc_assert (perm_mask != NULL_TREE); |
4643 | ncopies *= 2; | |
4644 | } | |
4645 | else | |
4646 | gcc_unreachable (); | |
4647 | ||
4648 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
4649 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4650 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4651 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4652 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4653 | scaletype = TREE_VALUE (arglist); | |
4654 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
4655 | && types_compatible_p (srctype, masktype)); | |
4656 | ||
4657 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4658 | ||
4659 | ptr = fold_convert (ptrtype, gather_base); | |
4660 | if (!is_gimple_min_invariant (ptr)) | |
4661 | { | |
4662 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
4663 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
4664 | gcc_assert (!new_bb); | |
4665 | } | |
4666 | ||
4667 | /* Currently we support only unconditional gather loads, | |
4668 | so mask should be all ones. */ | |
4669 | if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
4670 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
4671 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) | |
4672 | { | |
4673 | REAL_VALUE_TYPE r; | |
4674 | long tmp[6]; | |
4675 | for (j = 0; j < 6; ++j) | |
4676 | tmp[j] = -1; | |
4677 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
4678 | mask = build_real (TREE_TYPE (masktype), r); | |
4679 | } | |
4680 | else | |
4681 | gcc_unreachable (); | |
4682 | mask = build_vector_from_val (masktype, mask); | |
4683 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
4684 | ||
4685 | scale = build_int_cst (scaletype, gather_scale); | |
4686 | ||
4687 | prev_stmt_info = NULL; | |
4688 | for (j = 0; j < ncopies; ++j) | |
4689 | { | |
4690 | if (modifier == WIDEN && (j & 1)) | |
4691 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
4692 | perm_mask, stmt, gsi); | |
4693 | else if (j == 0) | |
4694 | op = vec_oprnd0 | |
4695 | = vect_get_vec_def_for_operand (gather_off, stmt, NULL); | |
4696 | else | |
4697 | op = vec_oprnd0 | |
4698 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
4699 | ||
4700 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
4701 | { | |
4702 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
4703 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
4704 | var = vect_get_new_vect_var (idxtype, vect_simple_var, NULL); | |
aec7ae7d JJ |
4705 | var = make_ssa_name (var, NULL); |
4706 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); | |
4707 | new_stmt | |
4708 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, | |
4709 | op, NULL_TREE); | |
4710 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4711 | op = var; | |
4712 | } | |
4713 | ||
4714 | new_stmt | |
4715 | = gimple_build_call (gather_decl, 5, mask, ptr, op, mask, scale); | |
4716 | ||
4717 | if (!useless_type_conversion_p (vectype, rettype)) | |
4718 | { | |
4719 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
4720 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
4721 | var = vect_get_new_vect_var (rettype, vect_simple_var, NULL); | |
aec7ae7d JJ |
4722 | op = make_ssa_name (var, new_stmt); |
4723 | gimple_call_set_lhs (new_stmt, op); | |
4724 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4725 | var = make_ssa_name (vec_dest, NULL); | |
4726 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); | |
4727 | new_stmt | |
4728 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, op, | |
4729 | NULL_TREE); | |
4730 | } | |
4731 | else | |
4732 | { | |
4733 | var = make_ssa_name (vec_dest, new_stmt); | |
4734 | gimple_call_set_lhs (new_stmt, var); | |
4735 | } | |
4736 | ||
4737 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4738 | ||
4739 | if (modifier == NARROW) | |
4740 | { | |
4741 | if ((j & 1) == 0) | |
4742 | { | |
4743 | prev_res = var; | |
4744 | continue; | |
4745 | } | |
4746 | var = permute_vec_elements (prev_res, var, | |
4747 | perm_mask, stmt, gsi); | |
4748 | new_stmt = SSA_NAME_DEF_STMT (var); | |
4749 | } | |
4750 | ||
4751 | if (prev_stmt_info == NULL) | |
4752 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4753 | else | |
4754 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4755 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4756 | } | |
4757 | return true; | |
4758 | } | |
7d75abc8 MM |
4759 | else if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
4760 | { | |
4761 | gimple_stmt_iterator incr_gsi; | |
4762 | bool insert_after; | |
4763 | gimple incr; | |
4764 | tree offvar; | |
7d75abc8 MM |
4765 | tree ivstep; |
4766 | tree running_off; | |
9771b263 | 4767 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 4768 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
4769 | tree stride_base, stride_step, alias_off; |
4770 | ||
4771 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 4772 | |
14ac6aa2 RB |
4773 | stride_base |
4774 | = fold_build_pointer_plus | |
4775 | (unshare_expr (DR_BASE_ADDRESS (dr)), | |
4776 | size_binop (PLUS_EXPR, | |
4777 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (dr))), | |
c3284718 | 4778 | convert_to_ptrofftype (DR_INIT (dr)))); |
14ac6aa2 | 4779 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (dr))); |
7d75abc8 MM |
4780 | |
4781 | /* For a load with loop-invariant (but other than power-of-2) | |
4782 | stride (i.e. not a grouped access) like so: | |
4783 | ||
4784 | for (i = 0; i < n; i += stride) | |
4785 | ... = array[i]; | |
4786 | ||
4787 | we generate a new induction variable and new accesses to | |
4788 | form a new vector (or vectors, depending on ncopies): | |
4789 | ||
4790 | for (j = 0; ; j += VF*stride) | |
4791 | tmp1 = array[j]; | |
4792 | tmp2 = array[j + stride]; | |
4793 | ... | |
4794 | vectemp = {tmp1, tmp2, ...} | |
4795 | */ | |
4796 | ||
4797 | ivstep = stride_step; | |
4798 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
4799 | build_int_cst (TREE_TYPE (ivstep), vf)); | |
4800 | ||
4801 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
4802 | ||
4803 | create_iv (stride_base, ivstep, NULL, | |
4804 | loop, &incr_gsi, insert_after, | |
4805 | &offvar, NULL); | |
4806 | incr = gsi_stmt (incr_gsi); | |
4807 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo, NULL)); | |
4808 | ||
4809 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
4810 | if (stmts) | |
4811 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
4812 | ||
4813 | prev_stmt_info = NULL; | |
4814 | running_off = offvar; | |
14ac6aa2 | 4815 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (dr)), 0); |
7d75abc8 MM |
4816 | for (j = 0; j < ncopies; j++) |
4817 | { | |
4818 | tree vec_inv; | |
4819 | ||
9771b263 | 4820 | vec_alloc (v, nunits); |
7d75abc8 MM |
4821 | for (i = 0; i < nunits; i++) |
4822 | { | |
4823 | tree newref, newoff; | |
4824 | gimple incr; | |
14ac6aa2 RB |
4825 | newref = build2 (MEM_REF, TREE_TYPE (vectype), |
4826 | running_off, alias_off); | |
7d75abc8 MM |
4827 | |
4828 | newref = force_gimple_operand_gsi (gsi, newref, true, | |
4829 | NULL_TREE, true, | |
4830 | GSI_SAME_STMT); | |
4831 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, newref); | |
070ecdfd | 4832 | newoff = copy_ssa_name (running_off, NULL); |
14ac6aa2 RB |
4833 | incr = gimple_build_assign_with_ops (POINTER_PLUS_EXPR, newoff, |
4834 | running_off, stride_step); | |
7d75abc8 MM |
4835 | vect_finish_stmt_generation (stmt, incr, gsi); |
4836 | ||
4837 | running_off = newoff; | |
4838 | } | |
4839 | ||
4840 | vec_inv = build_constructor (vectype, v); | |
4841 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
4842 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
7d75abc8 MM |
4843 | |
4844 | if (j == 0) | |
4845 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4846 | else | |
4847 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4848 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4849 | } | |
4850 | return true; | |
4851 | } | |
aec7ae7d | 4852 | |
0d0293ac | 4853 | if (grouped_load) |
ebfd146a | 4854 | { |
e14c1050 | 4855 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
6aa904c4 | 4856 | if (slp |
01d8bf07 | 4857 | && !SLP_TREE_LOAD_PERMUTATION (slp_node).exists () |
9771b263 DN |
4858 | && first_stmt != SLP_TREE_SCALAR_STMTS (slp_node)[0]) |
4859 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 4860 | |
ebfd146a | 4861 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
4862 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
4863 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
4864 | ??? But we can only do so if there is exactly one | |
4865 | as we have no way to get at the rest. Leave the CSE | |
4866 | opportunity alone. | |
4867 | ??? With the group load eventually participating | |
4868 | in multiple different permutations (having multiple | |
4869 | slp nodes which refer to the same group) the CSE | |
4870 | is even wrong code. See PR56270. */ | |
4871 | && !slp) | |
ebfd146a IR |
4872 | { |
4873 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4874 | return true; | |
4875 | } | |
4876 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 4877 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a IR |
4878 | |
4879 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
4880 | if (slp) | |
4881 | { | |
0d0293ac | 4882 | grouped_load = false; |
ebfd146a | 4883 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
01d8bf07 | 4884 | if (SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
a70d6342 | 4885 | slp_perm = true; |
a64b9c26 | 4886 | group_gap = GROUP_GAP (vinfo_for_stmt (first_stmt)); |
a70d6342 | 4887 | } |
ebfd146a | 4888 | else |
a64b9c26 RB |
4889 | { |
4890 | vec_num = group_size; | |
4891 | group_gap = 0; | |
4892 | } | |
ebfd146a IR |
4893 | } |
4894 | else | |
4895 | { | |
4896 | first_stmt = stmt; | |
4897 | first_dr = dr; | |
4898 | group_size = vec_num = 1; | |
a64b9c26 | 4899 | group_gap = 0; |
ebfd146a IR |
4900 | } |
4901 | ||
720f5239 | 4902 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 4903 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
4904 | /* Targets with load-lane instructions must not require explicit |
4905 | realignment. */ | |
4906 | gcc_assert (!load_lanes_p | |
4907 | || alignment_support_scheme == dr_aligned | |
4908 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
4909 | |
4910 | /* In case the vectorization factor (VF) is bigger than the number | |
4911 | of elements that we can fit in a vectype (nunits), we have to generate | |
4912 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 4913 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 4914 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 4915 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 4916 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
4917 | stmts that use the defs of the current stmt. The example below |
4918 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
4919 | need to create 4 vectorized stmts): | |
ebfd146a IR |
4920 | |
4921 | before vectorization: | |
4922 | RELATED_STMT VEC_STMT | |
4923 | S1: x = memref - - | |
4924 | S2: z = x + 1 - - | |
4925 | ||
4926 | step 1: vectorize stmt S1: | |
4927 | We first create the vector stmt VS1_0, and, as usual, record a | |
4928 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
4929 | Next, we create the vector stmt VS1_1, and record a pointer to | |
4930 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 4931 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
4932 | stmts and pointers: |
4933 | RELATED_STMT VEC_STMT | |
4934 | VS1_0: vx0 = memref0 VS1_1 - | |
4935 | VS1_1: vx1 = memref1 VS1_2 - | |
4936 | VS1_2: vx2 = memref2 VS1_3 - | |
4937 | VS1_3: vx3 = memref3 - - | |
4938 | S1: x = load - VS1_0 | |
4939 | S2: z = x + 1 - - | |
4940 | ||
b8698a0f L |
4941 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
4942 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
4943 | stmt S2. */ |
4944 | ||
0d0293ac | 4945 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
4946 | |
4947 | S1: x2 = &base + 2 | |
4948 | S2: x0 = &base | |
4949 | S3: x1 = &base + 1 | |
4950 | S4: x3 = &base + 3 | |
4951 | ||
b8698a0f | 4952 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
4953 | starting from the access of the first stmt of the chain: |
4954 | ||
4955 | VS1: vx0 = &base | |
4956 | VS2: vx1 = &base + vec_size*1 | |
4957 | VS3: vx3 = &base + vec_size*2 | |
4958 | VS4: vx4 = &base + vec_size*3 | |
4959 | ||
4960 | Then permutation statements are generated: | |
4961 | ||
e2c83630 RH |
4962 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
4963 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
4964 | ... |
4965 | ||
4966 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
4967 | (the order of the data-refs in the output of vect_permute_load_chain | |
4968 | corresponds to the order of scalar stmts in the interleaving chain - see | |
4969 | the documentation of vect_permute_load_chain()). | |
4970 | The generation of permutation stmts and recording them in | |
0d0293ac | 4971 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 4972 | |
b8698a0f | 4973 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
4974 | permutation stmts above are created for every copy. The result vector |
4975 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
4976 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
4977 | |
4978 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
4979 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
4980 | we generate the following code: | |
4981 | p = initial_addr; | |
4982 | indx = 0; | |
4983 | loop { | |
4984 | p = p + indx * vectype_size; | |
4985 | vec_dest = *(p); | |
4986 | indx = indx + 1; | |
4987 | } | |
4988 | ||
4989 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 4990 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
4991 | then generate the following code, in which the data in each iteration is |
4992 | obtained by two vector loads, one from the previous iteration, and one | |
4993 | from the current iteration: | |
4994 | p1 = initial_addr; | |
4995 | msq_init = *(floor(p1)) | |
4996 | p2 = initial_addr + VS - 1; | |
4997 | realignment_token = call target_builtin; | |
4998 | indx = 0; | |
4999 | loop { | |
5000 | p2 = p2 + indx * vectype_size | |
5001 | lsq = *(floor(p2)) | |
5002 | vec_dest = realign_load (msq, lsq, realignment_token) | |
5003 | indx = indx + 1; | |
5004 | msq = lsq; | |
5005 | } */ | |
5006 | ||
5007 | /* If the misalignment remains the same throughout the execution of the | |
5008 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 5009 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
5010 | This can only occur when vectorizing memory accesses in the inner-loop |
5011 | nested within an outer-loop that is being vectorized. */ | |
5012 | ||
d1e4b493 | 5013 | if (nested_in_vect_loop |
211bea38 | 5014 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
ebfd146a IR |
5015 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) |
5016 | { | |
5017 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
5018 | compute_in_loop = true; | |
5019 | } | |
5020 | ||
5021 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
5022 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 5023 | && !compute_in_loop) |
ebfd146a IR |
5024 | { |
5025 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
5026 | alignment_support_scheme, NULL_TREE, | |
5027 | &at_loop); | |
5028 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
5029 | { | |
5030 | phi = SSA_NAME_DEF_STMT (msq); | |
5031 | offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
5032 | } | |
5033 | } | |
5034 | else | |
5035 | at_loop = loop; | |
5036 | ||
a1e53f3f L |
5037 | if (negative) |
5038 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
5039 | ||
272c6793 RS |
5040 | if (load_lanes_p) |
5041 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
5042 | else | |
5043 | aggr_type = vectype; | |
5044 | ||
ebfd146a IR |
5045 | prev_stmt_info = NULL; |
5046 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 5047 | { |
272c6793 | 5048 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 5049 | if (j == 0) |
74bf76ed JJ |
5050 | { |
5051 | bool simd_lane_access_p | |
5052 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
5053 | if (simd_lane_access_p | |
5054 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
5055 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
5056 | && integer_zerop (DR_OFFSET (first_dr)) | |
5057 | && integer_zerop (DR_INIT (first_dr)) | |
5058 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
5059 | get_alias_set (DR_REF (first_dr))) | |
5060 | && (alignment_support_scheme == dr_aligned | |
5061 | || alignment_support_scheme == dr_unaligned_supported)) | |
5062 | { | |
5063 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
5064 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
5065 | (DR_REF (first_dr)), 0); | |
8928eff3 | 5066 | inv_p = false; |
74bf76ed JJ |
5067 | } |
5068 | else | |
5069 | dataref_ptr | |
5070 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
5071 | offset, &dummy, gsi, &ptr_incr, | |
5072 | simd_lane_access_p, &inv_p); | |
5073 | } | |
5074 | else if (dataref_offset) | |
5075 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
5076 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 5077 | else |
272c6793 RS |
5078 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
5079 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 5080 | |
0d0293ac | 5081 | if (grouped_load || slp_perm) |
9771b263 | 5082 | dr_chain.create (vec_num); |
5ce1ee7f | 5083 | |
272c6793 | 5084 | if (load_lanes_p) |
ebfd146a | 5085 | { |
272c6793 RS |
5086 | tree vec_array; |
5087 | ||
5088 | vec_array = create_vector_array (vectype, vec_num); | |
5089 | ||
5090 | /* Emit: | |
5091 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
5092 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
5093 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
5094 | gimple_call_set_lhs (new_stmt, vec_array); | |
5095 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 5096 | |
272c6793 RS |
5097 | /* Extract each vector into an SSA_NAME. */ |
5098 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 5099 | { |
272c6793 RS |
5100 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
5101 | vec_array, i); | |
9771b263 | 5102 | dr_chain.quick_push (new_temp); |
272c6793 RS |
5103 | } |
5104 | ||
5105 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 5106 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
5107 | } |
5108 | else | |
5109 | { | |
5110 | for (i = 0; i < vec_num; i++) | |
5111 | { | |
5112 | if (i > 0) | |
5113 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5114 | stmt, NULL_TREE); | |
5115 | ||
5116 | /* 2. Create the vector-load in the loop. */ | |
5117 | switch (alignment_support_scheme) | |
5118 | { | |
5119 | case dr_aligned: | |
5120 | case dr_unaligned_supported: | |
be1ac4ec | 5121 | { |
644ffefd MJ |
5122 | unsigned int align, misalign; |
5123 | ||
272c6793 RS |
5124 | data_ref |
5125 | = build2 (MEM_REF, vectype, dataref_ptr, | |
74bf76ed JJ |
5126 | dataref_offset |
5127 | ? dataref_offset | |
5128 | : build_int_cst (reference_alias_ptr_type | |
5129 | (DR_REF (first_dr)), 0)); | |
644ffefd | 5130 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 RS |
5131 | if (alignment_support_scheme == dr_aligned) |
5132 | { | |
5133 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 5134 | misalign = 0; |
272c6793 RS |
5135 | } |
5136 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
5137 | { | |
5138 | TREE_TYPE (data_ref) | |
5139 | = build_aligned_type (TREE_TYPE (data_ref), | |
5140 | TYPE_ALIGN (elem_type)); | |
644ffefd MJ |
5141 | align = TYPE_ALIGN_UNIT (elem_type); |
5142 | misalign = 0; | |
272c6793 RS |
5143 | } |
5144 | else | |
5145 | { | |
5146 | TREE_TYPE (data_ref) | |
5147 | = build_aligned_type (TREE_TYPE (data_ref), | |
5148 | TYPE_ALIGN (elem_type)); | |
644ffefd | 5149 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 5150 | } |
74bf76ed JJ |
5151 | if (dataref_offset == NULL_TREE) |
5152 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), | |
5153 | align, misalign); | |
272c6793 | 5154 | break; |
be1ac4ec | 5155 | } |
272c6793 | 5156 | case dr_explicit_realign: |
267d3070 | 5157 | { |
272c6793 RS |
5158 | tree ptr, bump; |
5159 | tree vs_minus_1; | |
5160 | ||
5161 | vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
5162 | ||
5163 | if (compute_in_loop) | |
5164 | msq = vect_setup_realignment (first_stmt, gsi, | |
5165 | &realignment_token, | |
5166 | dr_explicit_realign, | |
5167 | dataref_ptr, NULL); | |
5168 | ||
070ecdfd | 5169 | ptr = copy_ssa_name (dataref_ptr, NULL); |
272c6793 | 5170 | new_stmt = gimple_build_assign_with_ops |
070ecdfd | 5171 | (BIT_AND_EXPR, ptr, dataref_ptr, |
272c6793 RS |
5172 | build_int_cst |
5173 | (TREE_TYPE (dataref_ptr), | |
5174 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
5175 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
5176 | data_ref | |
5177 | = build2 (MEM_REF, vectype, ptr, | |
5178 | build_int_cst (reference_alias_ptr_type | |
5179 | (DR_REF (first_dr)), 0)); | |
5180 | vec_dest = vect_create_destination_var (scalar_dest, | |
5181 | vectype); | |
5182 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
5183 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5184 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5185 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
5186 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
5187 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5188 | msq = new_temp; | |
5189 | ||
5190 | bump = size_binop (MULT_EXPR, vs_minus_1, | |
7b7b1813 | 5191 | TYPE_SIZE_UNIT (elem_type)); |
272c6793 RS |
5192 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
5193 | new_stmt = gimple_build_assign_with_ops | |
5194 | (BIT_AND_EXPR, NULL_TREE, ptr, | |
5195 | build_int_cst | |
5196 | (TREE_TYPE (ptr), | |
5197 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
070ecdfd | 5198 | ptr = copy_ssa_name (dataref_ptr, new_stmt); |
272c6793 RS |
5199 | gimple_assign_set_lhs (new_stmt, ptr); |
5200 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5201 | data_ref | |
5202 | = build2 (MEM_REF, vectype, ptr, | |
5203 | build_int_cst (reference_alias_ptr_type | |
5204 | (DR_REF (first_dr)), 0)); | |
5205 | break; | |
267d3070 | 5206 | } |
272c6793 | 5207 | case dr_explicit_realign_optimized: |
070ecdfd | 5208 | new_temp = copy_ssa_name (dataref_ptr, NULL); |
272c6793 | 5209 | new_stmt = gimple_build_assign_with_ops |
070ecdfd | 5210 | (BIT_AND_EXPR, new_temp, dataref_ptr, |
272c6793 RS |
5211 | build_int_cst |
5212 | (TREE_TYPE (dataref_ptr), | |
5213 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
5214 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
5215 | data_ref | |
5216 | = build2 (MEM_REF, vectype, new_temp, | |
5217 | build_int_cst (reference_alias_ptr_type | |
5218 | (DR_REF (first_dr)), 0)); | |
5219 | break; | |
5220 | default: | |
5221 | gcc_unreachable (); | |
5222 | } | |
ebfd146a | 5223 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 5224 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
5225 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5226 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5227 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5228 | ||
272c6793 RS |
5229 | /* 3. Handle explicit realignment if necessary/supported. |
5230 | Create in loop: | |
5231 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
5232 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
5233 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 5234 | { |
272c6793 RS |
5235 | lsq = gimple_assign_lhs (new_stmt); |
5236 | if (!realignment_token) | |
5237 | realignment_token = dataref_ptr; | |
5238 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5239 | new_stmt | |
73804b12 RG |
5240 | = gimple_build_assign_with_ops (REALIGN_LOAD_EXPR, |
5241 | vec_dest, msq, lsq, | |
5242 | realignment_token); | |
272c6793 RS |
5243 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5244 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5245 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5246 | ||
5247 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
5248 | { | |
5249 | gcc_assert (phi); | |
5250 | if (i == vec_num - 1 && j == ncopies - 1) | |
5251 | add_phi_arg (phi, lsq, | |
5252 | loop_latch_edge (containing_loop), | |
9e227d60 | 5253 | UNKNOWN_LOCATION); |
272c6793 RS |
5254 | msq = lsq; |
5255 | } | |
ebfd146a | 5256 | } |
ebfd146a | 5257 | |
59fd17e3 RB |
5258 | /* 4. Handle invariant-load. */ |
5259 | if (inv_p && !bb_vinfo) | |
5260 | { | |
5261 | gimple_stmt_iterator gsi2 = *gsi; | |
5262 | gcc_assert (!grouped_load); | |
5263 | gsi_next (&gsi2); | |
5264 | new_temp = vect_init_vector (stmt, scalar_dest, | |
5265 | vectype, &gsi2); | |
5266 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
5267 | } | |
5268 | ||
272c6793 RS |
5269 | if (negative) |
5270 | { | |
aec7ae7d JJ |
5271 | tree perm_mask = perm_mask_for_reverse (vectype); |
5272 | new_temp = permute_vec_elements (new_temp, new_temp, | |
5273 | perm_mask, stmt, gsi); | |
ebfd146a IR |
5274 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
5275 | } | |
267d3070 | 5276 | |
272c6793 | 5277 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
5278 | vect_transform_grouped_load (). */ |
5279 | if (grouped_load || slp_perm) | |
9771b263 | 5280 | dr_chain.quick_push (new_temp); |
267d3070 | 5281 | |
272c6793 RS |
5282 | /* Store vector loads in the corresponding SLP_NODE. */ |
5283 | if (slp && !slp_perm) | |
9771b263 | 5284 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
272c6793 | 5285 | } |
a64b9c26 RB |
5286 | /* Bump the vector pointer to account for a gap. */ |
5287 | if (slp && group_gap != 0) | |
5288 | { | |
5289 | tree bump = size_binop (MULT_EXPR, | |
5290 | TYPE_SIZE_UNIT (elem_type), | |
5291 | size_int (group_gap)); | |
5292 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5293 | stmt, bump); | |
5294 | } | |
ebfd146a IR |
5295 | } |
5296 | ||
5297 | if (slp && !slp_perm) | |
5298 | continue; | |
5299 | ||
5300 | if (slp_perm) | |
5301 | { | |
01d8bf07 | 5302 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
ebfd146a IR |
5303 | slp_node_instance, false)) |
5304 | { | |
9771b263 | 5305 | dr_chain.release (); |
ebfd146a IR |
5306 | return false; |
5307 | } | |
5308 | } | |
5309 | else | |
5310 | { | |
0d0293ac | 5311 | if (grouped_load) |
ebfd146a | 5312 | { |
272c6793 | 5313 | if (!load_lanes_p) |
0d0293ac | 5314 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 5315 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
5316 | } |
5317 | else | |
5318 | { | |
5319 | if (j == 0) | |
5320 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5321 | else | |
5322 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5323 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5324 | } | |
5325 | } | |
9771b263 | 5326 | dr_chain.release (); |
ebfd146a IR |
5327 | } |
5328 | ||
ebfd146a IR |
5329 | return true; |
5330 | } | |
5331 | ||
5332 | /* Function vect_is_simple_cond. | |
b8698a0f | 5333 | |
ebfd146a IR |
5334 | Input: |
5335 | LOOP - the loop that is being vectorized. | |
5336 | COND - Condition that is checked for simple use. | |
5337 | ||
e9e1d143 RG |
5338 | Output: |
5339 | *COMP_VECTYPE - the vector type for the comparison. | |
5340 | ||
ebfd146a IR |
5341 | Returns whether a COND can be vectorized. Checks whether |
5342 | condition operands are supportable using vec_is_simple_use. */ | |
5343 | ||
87aab9b2 | 5344 | static bool |
24ee1384 IR |
5345 | vect_is_simple_cond (tree cond, gimple stmt, loop_vec_info loop_vinfo, |
5346 | bb_vec_info bb_vinfo, tree *comp_vectype) | |
ebfd146a IR |
5347 | { |
5348 | tree lhs, rhs; | |
5349 | tree def; | |
5350 | enum vect_def_type dt; | |
e9e1d143 | 5351 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a IR |
5352 | |
5353 | if (!COMPARISON_CLASS_P (cond)) | |
5354 | return false; | |
5355 | ||
5356 | lhs = TREE_OPERAND (cond, 0); | |
5357 | rhs = TREE_OPERAND (cond, 1); | |
5358 | ||
5359 | if (TREE_CODE (lhs) == SSA_NAME) | |
5360 | { | |
5361 | gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); | |
24ee1384 IR |
5362 | if (!vect_is_simple_use_1 (lhs, stmt, loop_vinfo, bb_vinfo, |
5363 | &lhs_def_stmt, &def, &dt, &vectype1)) | |
ebfd146a IR |
5364 | return false; |
5365 | } | |
5366 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
5367 | && TREE_CODE (lhs) != FIXED_CST) | |
5368 | return false; | |
5369 | ||
5370 | if (TREE_CODE (rhs) == SSA_NAME) | |
5371 | { | |
5372 | gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); | |
24ee1384 IR |
5373 | if (!vect_is_simple_use_1 (rhs, stmt, loop_vinfo, bb_vinfo, |
5374 | &rhs_def_stmt, &def, &dt, &vectype2)) | |
ebfd146a IR |
5375 | return false; |
5376 | } | |
f7e531cf | 5377 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST |
ebfd146a IR |
5378 | && TREE_CODE (rhs) != FIXED_CST) |
5379 | return false; | |
5380 | ||
e9e1d143 | 5381 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
5382 | return true; |
5383 | } | |
5384 | ||
5385 | /* vectorizable_condition. | |
5386 | ||
b8698a0f L |
5387 | Check if STMT is conditional modify expression that can be vectorized. |
5388 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
5389 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
5390 | at GSI. |
5391 | ||
5392 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
5393 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
5394 | else caluse if it is 2). | |
ebfd146a IR |
5395 | |
5396 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5397 | ||
4bbe8262 | 5398 | bool |
ebfd146a | 5399 | vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi, |
f7e531cf IR |
5400 | gimple *vec_stmt, tree reduc_def, int reduc_index, |
5401 | slp_tree slp_node) | |
ebfd146a IR |
5402 | { |
5403 | tree scalar_dest = NULL_TREE; | |
5404 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
5405 | tree cond_expr, then_clause, else_clause; |
5406 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5407 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
df11cc78 | 5408 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
5409 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
5410 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
ebfd146a IR |
5411 | tree vec_compare, vec_cond_expr; |
5412 | tree new_temp; | |
5413 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
ebfd146a | 5414 | tree def; |
a855b1b1 | 5415 | enum vect_def_type dt, dts[4]; |
ebfd146a | 5416 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
f7e531cf | 5417 | int ncopies; |
ebfd146a | 5418 | enum tree_code code; |
a855b1b1 | 5419 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
5420 | int i, j; |
5421 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
5422 | vec<tree> vec_oprnds0 = vNULL; |
5423 | vec<tree> vec_oprnds1 = vNULL; | |
5424 | vec<tree> vec_oprnds2 = vNULL; | |
5425 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 5426 | tree vec_cmp_type; |
b8698a0f | 5427 | |
f7e531cf IR |
5428 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
5429 | ncopies = 1; | |
5430 | else | |
5431 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
437f4a00 | 5432 | |
ebfd146a | 5433 | gcc_assert (ncopies >= 1); |
a855b1b1 | 5434 | if (reduc_index && ncopies > 1) |
ebfd146a IR |
5435 | return false; /* FORNOW */ |
5436 | ||
f7e531cf IR |
5437 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
5438 | return false; | |
5439 | ||
5440 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
ebfd146a IR |
5441 | return false; |
5442 | ||
4bbe8262 IR |
5443 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5444 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
5445 | && reduc_def)) | |
ebfd146a IR |
5446 | return false; |
5447 | ||
ebfd146a | 5448 | /* FORNOW: not yet supported. */ |
b8698a0f | 5449 | if (STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 5450 | { |
73fbfcad | 5451 | if (dump_enabled_p ()) |
78c60e3d | 5452 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5453 | "value used after loop.\n"); |
ebfd146a IR |
5454 | return false; |
5455 | } | |
5456 | ||
5457 | /* Is vectorizable conditional operation? */ | |
5458 | if (!is_gimple_assign (stmt)) | |
5459 | return false; | |
5460 | ||
5461 | code = gimple_assign_rhs_code (stmt); | |
5462 | ||
5463 | if (code != COND_EXPR) | |
5464 | return false; | |
5465 | ||
4e71066d RG |
5466 | cond_expr = gimple_assign_rhs1 (stmt); |
5467 | then_clause = gimple_assign_rhs2 (stmt); | |
5468 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 5469 | |
24ee1384 IR |
5470 | if (!vect_is_simple_cond (cond_expr, stmt, loop_vinfo, bb_vinfo, |
5471 | &comp_vectype) | |
e9e1d143 | 5472 | || !comp_vectype) |
ebfd146a IR |
5473 | return false; |
5474 | ||
5475 | if (TREE_CODE (then_clause) == SSA_NAME) | |
5476 | { | |
5477 | gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause); | |
24ee1384 | 5478 | if (!vect_is_simple_use (then_clause, stmt, loop_vinfo, bb_vinfo, |
ebfd146a IR |
5479 | &then_def_stmt, &def, &dt)) |
5480 | return false; | |
5481 | } | |
b8698a0f | 5482 | else if (TREE_CODE (then_clause) != INTEGER_CST |
ebfd146a IR |
5483 | && TREE_CODE (then_clause) != REAL_CST |
5484 | && TREE_CODE (then_clause) != FIXED_CST) | |
5485 | return false; | |
5486 | ||
5487 | if (TREE_CODE (else_clause) == SSA_NAME) | |
5488 | { | |
5489 | gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause); | |
24ee1384 | 5490 | if (!vect_is_simple_use (else_clause, stmt, loop_vinfo, bb_vinfo, |
ebfd146a IR |
5491 | &else_def_stmt, &def, &dt)) |
5492 | return false; | |
5493 | } | |
b8698a0f | 5494 | else if (TREE_CODE (else_clause) != INTEGER_CST |
ebfd146a IR |
5495 | && TREE_CODE (else_clause) != REAL_CST |
5496 | && TREE_CODE (else_clause) != FIXED_CST) | |
5497 | return false; | |
5498 | ||
74946978 MP |
5499 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (vectype))); |
5500 | /* The result of a vector comparison should be signed type. */ | |
5501 | tree cmp_type = build_nonstandard_integer_type (prec, 0); | |
5502 | vec_cmp_type = get_same_sized_vectype (cmp_type, vectype); | |
5503 | if (vec_cmp_type == NULL_TREE) | |
5504 | return false; | |
784fb9b3 | 5505 | |
b8698a0f | 5506 | if (!vec_stmt) |
ebfd146a IR |
5507 | { |
5508 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
e9e1d143 | 5509 | return expand_vec_cond_expr_p (vectype, comp_vectype); |
ebfd146a IR |
5510 | } |
5511 | ||
f7e531cf IR |
5512 | /* Transform. */ |
5513 | ||
5514 | if (!slp_node) | |
5515 | { | |
9771b263 DN |
5516 | vec_oprnds0.create (1); |
5517 | vec_oprnds1.create (1); | |
5518 | vec_oprnds2.create (1); | |
5519 | vec_oprnds3.create (1); | |
f7e531cf | 5520 | } |
ebfd146a IR |
5521 | |
5522 | /* Handle def. */ | |
5523 | scalar_dest = gimple_assign_lhs (stmt); | |
5524 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5525 | ||
5526 | /* Handle cond expr. */ | |
a855b1b1 MM |
5527 | for (j = 0; j < ncopies; j++) |
5528 | { | |
f7e531cf | 5529 | gimple new_stmt = NULL; |
a855b1b1 MM |
5530 | if (j == 0) |
5531 | { | |
f7e531cf IR |
5532 | if (slp_node) |
5533 | { | |
07687835 TS |
5534 | stack_vec<tree, 4> ops; |
5535 | stack_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 5536 | |
9771b263 DN |
5537 | ops.safe_push (TREE_OPERAND (cond_expr, 0)); |
5538 | ops.safe_push (TREE_OPERAND (cond_expr, 1)); | |
5539 | ops.safe_push (then_clause); | |
5540 | ops.safe_push (else_clause); | |
f7e531cf | 5541 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); |
37b5ec8f JJ |
5542 | vec_oprnds3 = vec_defs.pop (); |
5543 | vec_oprnds2 = vec_defs.pop (); | |
5544 | vec_oprnds1 = vec_defs.pop (); | |
5545 | vec_oprnds0 = vec_defs.pop (); | |
f7e531cf | 5546 | |
9771b263 DN |
5547 | ops.release (); |
5548 | vec_defs.release (); | |
f7e531cf IR |
5549 | } |
5550 | else | |
5551 | { | |
5552 | gimple gtemp; | |
5553 | vec_cond_lhs = | |
a855b1b1 MM |
5554 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), |
5555 | stmt, NULL); | |
24ee1384 IR |
5556 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), stmt, |
5557 | loop_vinfo, NULL, >emp, &def, &dts[0]); | |
f7e531cf IR |
5558 | |
5559 | vec_cond_rhs = | |
5560 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
5561 | stmt, NULL); | |
24ee1384 IR |
5562 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), stmt, |
5563 | loop_vinfo, NULL, >emp, &def, &dts[1]); | |
f7e531cf IR |
5564 | if (reduc_index == 1) |
5565 | vec_then_clause = reduc_def; | |
5566 | else | |
5567 | { | |
5568 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
5569 | stmt, NULL); | |
24ee1384 | 5570 | vect_is_simple_use (then_clause, stmt, loop_vinfo, |
f7e531cf IR |
5571 | NULL, >emp, &def, &dts[2]); |
5572 | } | |
5573 | if (reduc_index == 2) | |
5574 | vec_else_clause = reduc_def; | |
5575 | else | |
5576 | { | |
5577 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
a855b1b1 | 5578 | stmt, NULL); |
24ee1384 | 5579 | vect_is_simple_use (else_clause, stmt, loop_vinfo, |
a855b1b1 | 5580 | NULL, >emp, &def, &dts[3]); |
f7e531cf | 5581 | } |
a855b1b1 MM |
5582 | } |
5583 | } | |
5584 | else | |
5585 | { | |
f7e531cf | 5586 | vec_cond_lhs = vect_get_vec_def_for_stmt_copy (dts[0], |
9771b263 | 5587 | vec_oprnds0.pop ()); |
f7e531cf | 5588 | vec_cond_rhs = vect_get_vec_def_for_stmt_copy (dts[1], |
9771b263 | 5589 | vec_oprnds1.pop ()); |
a855b1b1 | 5590 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 5591 | vec_oprnds2.pop ()); |
a855b1b1 | 5592 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 5593 | vec_oprnds3.pop ()); |
f7e531cf IR |
5594 | } |
5595 | ||
5596 | if (!slp_node) | |
5597 | { | |
9771b263 DN |
5598 | vec_oprnds0.quick_push (vec_cond_lhs); |
5599 | vec_oprnds1.quick_push (vec_cond_rhs); | |
5600 | vec_oprnds2.quick_push (vec_then_clause); | |
5601 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
5602 | } |
5603 | ||
9dc3f7de | 5604 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 5605 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 5606 | { |
9771b263 DN |
5607 | vec_cond_rhs = vec_oprnds1[i]; |
5608 | vec_then_clause = vec_oprnds2[i]; | |
5609 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 5610 | |
784fb9b3 JJ |
5611 | vec_compare = build2 (TREE_CODE (cond_expr), vec_cmp_type, |
5612 | vec_cond_lhs, vec_cond_rhs); | |
f7e531cf IR |
5613 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, |
5614 | vec_compare, vec_then_clause, vec_else_clause); | |
a855b1b1 | 5615 | |
f7e531cf IR |
5616 | new_stmt = gimple_build_assign (vec_dest, vec_cond_expr); |
5617 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5618 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5619 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5620 | if (slp_node) | |
9771b263 | 5621 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
5622 | } |
5623 | ||
5624 | if (slp_node) | |
5625 | continue; | |
5626 | ||
5627 | if (j == 0) | |
5628 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5629 | else | |
5630 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5631 | ||
5632 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 5633 | } |
b8698a0f | 5634 | |
9771b263 DN |
5635 | vec_oprnds0.release (); |
5636 | vec_oprnds1.release (); | |
5637 | vec_oprnds2.release (); | |
5638 | vec_oprnds3.release (); | |
f7e531cf | 5639 | |
ebfd146a IR |
5640 | return true; |
5641 | } | |
5642 | ||
5643 | ||
8644a673 | 5644 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
5645 | |
5646 | bool | |
a70d6342 | 5647 | vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 5648 | { |
8644a673 | 5649 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 5650 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 5651 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 5652 | bool ok; |
a70d6342 | 5653 | tree scalar_type, vectype; |
363477c0 JJ |
5654 | gimple pattern_stmt; |
5655 | gimple_seq pattern_def_seq; | |
ebfd146a | 5656 | |
73fbfcad | 5657 | if (dump_enabled_p ()) |
ebfd146a | 5658 | { |
78c60e3d SS |
5659 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
5660 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 5661 | dump_printf (MSG_NOTE, "\n"); |
8644a673 | 5662 | } |
ebfd146a | 5663 | |
1825a1f3 | 5664 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 5665 | { |
73fbfcad | 5666 | if (dump_enabled_p ()) |
78c60e3d | 5667 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5668 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
5669 | |
5670 | return false; | |
5671 | } | |
b8698a0f L |
5672 | |
5673 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
5674 | to include: |
5675 | - the COND_EXPR which is the loop exit condition | |
5676 | - any LABEL_EXPRs in the loop | |
b8698a0f | 5677 | - computations that are used only for array indexing or loop control. |
8644a673 | 5678 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 5679 | instance, therefore, all the statements are relevant. |
ebfd146a | 5680 | |
d092494c | 5681 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 5682 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
5683 | statements. In basic blocks we are called from some SLP instance |
5684 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
5685 | already will be part of SLP instance. */ | |
83197f37 IR |
5686 | |
5687 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 5688 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 5689 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 5690 | { |
9d5e7640 | 5691 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 5692 | && pattern_stmt |
9d5e7640 IR |
5693 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
5694 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
5695 | { | |
83197f37 | 5696 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
5697 | stmt = pattern_stmt; |
5698 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 5699 | if (dump_enabled_p ()) |
9d5e7640 | 5700 | { |
78c60e3d SS |
5701 | dump_printf_loc (MSG_NOTE, vect_location, |
5702 | "==> examining pattern statement: "); | |
5703 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 5704 | dump_printf (MSG_NOTE, "\n"); |
9d5e7640 IR |
5705 | } |
5706 | } | |
5707 | else | |
5708 | { | |
73fbfcad | 5709 | if (dump_enabled_p ()) |
e645e942 | 5710 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 5711 | |
9d5e7640 IR |
5712 | return true; |
5713 | } | |
8644a673 | 5714 | } |
83197f37 | 5715 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 5716 | && node == NULL |
83197f37 IR |
5717 | && pattern_stmt |
5718 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
5719 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
5720 | { | |
5721 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 5722 | if (dump_enabled_p ()) |
83197f37 | 5723 | { |
78c60e3d SS |
5724 | dump_printf_loc (MSG_NOTE, vect_location, |
5725 | "==> examining pattern statement: "); | |
5726 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 5727 | dump_printf (MSG_NOTE, "\n"); |
83197f37 IR |
5728 | } |
5729 | ||
5730 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
5731 | return false; | |
5732 | } | |
ebfd146a | 5733 | |
1107f3ae | 5734 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 5735 | && node == NULL |
363477c0 | 5736 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 5737 | { |
363477c0 | 5738 | gimple_stmt_iterator si; |
1107f3ae | 5739 | |
363477c0 JJ |
5740 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
5741 | { | |
5742 | gimple pattern_def_stmt = gsi_stmt (si); | |
5743 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) | |
5744 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
5745 | { | |
5746 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 5747 | if (dump_enabled_p ()) |
363477c0 | 5748 | { |
78c60e3d SS |
5749 | dump_printf_loc (MSG_NOTE, vect_location, |
5750 | "==> examining pattern def statement: "); | |
5751 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
e645e942 | 5752 | dump_printf (MSG_NOTE, "\n"); |
363477c0 | 5753 | } |
1107f3ae | 5754 | |
363477c0 JJ |
5755 | if (!vect_analyze_stmt (pattern_def_stmt, |
5756 | need_to_vectorize, node)) | |
5757 | return false; | |
5758 | } | |
5759 | } | |
5760 | } | |
1107f3ae | 5761 | |
8644a673 IR |
5762 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
5763 | { | |
5764 | case vect_internal_def: | |
5765 | break; | |
ebfd146a | 5766 | |
8644a673 | 5767 | case vect_reduction_def: |
7c5222ff | 5768 | case vect_nested_cycle: |
a70d6342 | 5769 | gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer |
8644a673 | 5770 | || relevance == vect_used_in_outer_by_reduction |
a70d6342 | 5771 | || relevance == vect_unused_in_scope)); |
8644a673 IR |
5772 | break; |
5773 | ||
5774 | case vect_induction_def: | |
5775 | case vect_constant_def: | |
5776 | case vect_external_def: | |
5777 | case vect_unknown_def_type: | |
5778 | default: | |
5779 | gcc_unreachable (); | |
5780 | } | |
ebfd146a | 5781 | |
a70d6342 IR |
5782 | if (bb_vinfo) |
5783 | { | |
5784 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
5785 | ||
b690cc0f | 5786 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
73fbfcad | 5787 | if (dump_enabled_p ()) |
a70d6342 | 5788 | { |
78c60e3d SS |
5789 | dump_printf_loc (MSG_NOTE, vect_location, |
5790 | "get vectype for scalar type: "); | |
5791 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
e645e942 | 5792 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
5793 | } |
5794 | ||
5795 | vectype = get_vectype_for_scalar_type (scalar_type); | |
5796 | if (!vectype) | |
5797 | { | |
73fbfcad | 5798 | if (dump_enabled_p ()) |
a70d6342 | 5799 | { |
78c60e3d SS |
5800 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5801 | "not SLPed: unsupported data-type "); | |
5802 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
5803 | scalar_type); | |
e645e942 | 5804 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
a70d6342 IR |
5805 | } |
5806 | return false; | |
5807 | } | |
5808 | ||
73fbfcad | 5809 | if (dump_enabled_p ()) |
a70d6342 | 5810 | { |
78c60e3d SS |
5811 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); |
5812 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
e645e942 | 5813 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
5814 | } |
5815 | ||
5816 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
5817 | } | |
5818 | ||
8644a673 | 5819 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 5820 | { |
8644a673 IR |
5821 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
5822 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info)); | |
5823 | *need_to_vectorize = true; | |
ebfd146a IR |
5824 | } |
5825 | ||
8644a673 | 5826 | ok = true; |
b8698a0f | 5827 | if (!bb_vinfo |
a70d6342 IR |
5828 | && (STMT_VINFO_RELEVANT_P (stmt_info) |
5829 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
4a00c761 | 5830 | ok = (vectorizable_conversion (stmt, NULL, NULL, NULL) |
9dc3f7de | 5831 | || vectorizable_shift (stmt, NULL, NULL, NULL) |
8644a673 IR |
5832 | || vectorizable_operation (stmt, NULL, NULL, NULL) |
5833 | || vectorizable_assignment (stmt, NULL, NULL, NULL) | |
5834 | || vectorizable_load (stmt, NULL, NULL, NULL, NULL) | |
190c2236 | 5835 | || vectorizable_call (stmt, NULL, NULL, NULL) |
8644a673 | 5836 | || vectorizable_store (stmt, NULL, NULL, NULL) |
b5aeb3bb | 5837 | || vectorizable_reduction (stmt, NULL, NULL, NULL) |
f7e531cf | 5838 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, NULL)); |
a70d6342 IR |
5839 | else |
5840 | { | |
5841 | if (bb_vinfo) | |
4a00c761 JJ |
5842 | ok = (vectorizable_conversion (stmt, NULL, NULL, node) |
5843 | || vectorizable_shift (stmt, NULL, NULL, node) | |
9dc3f7de | 5844 | || vectorizable_operation (stmt, NULL, NULL, node) |
a70d6342 IR |
5845 | || vectorizable_assignment (stmt, NULL, NULL, node) |
5846 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
190c2236 | 5847 | || vectorizable_call (stmt, NULL, NULL, node) |
f7e531cf IR |
5848 | || vectorizable_store (stmt, NULL, NULL, node) |
5849 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node)); | |
b8698a0f | 5850 | } |
8644a673 IR |
5851 | |
5852 | if (!ok) | |
ebfd146a | 5853 | { |
73fbfcad | 5854 | if (dump_enabled_p ()) |
8644a673 | 5855 | { |
78c60e3d SS |
5856 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5857 | "not vectorized: relevant stmt not "); | |
5858 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
5859 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 5860 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
8644a673 | 5861 | } |
b8698a0f | 5862 | |
ebfd146a IR |
5863 | return false; |
5864 | } | |
5865 | ||
a70d6342 IR |
5866 | if (bb_vinfo) |
5867 | return true; | |
5868 | ||
8644a673 IR |
5869 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
5870 | need extra handling, except for vectorizable reductions. */ | |
5871 | if (STMT_VINFO_LIVE_P (stmt_info) | |
5872 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
5873 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 5874 | |
8644a673 | 5875 | if (!ok) |
ebfd146a | 5876 | { |
73fbfcad | 5877 | if (dump_enabled_p ()) |
8644a673 | 5878 | { |
78c60e3d SS |
5879 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5880 | "not vectorized: live stmt not "); | |
5881 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
5882 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 5883 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
8644a673 | 5884 | } |
b8698a0f | 5885 | |
8644a673 | 5886 | return false; |
ebfd146a IR |
5887 | } |
5888 | ||
ebfd146a IR |
5889 | return true; |
5890 | } | |
5891 | ||
5892 | ||
5893 | /* Function vect_transform_stmt. | |
5894 | ||
5895 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
5896 | ||
5897 | bool | |
5898 | vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi, | |
0d0293ac | 5899 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
5900 | slp_instance slp_node_instance) |
5901 | { | |
5902 | bool is_store = false; | |
5903 | gimple vec_stmt = NULL; | |
5904 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a | 5905 | bool done; |
ebfd146a IR |
5906 | |
5907 | switch (STMT_VINFO_TYPE (stmt_info)) | |
5908 | { | |
5909 | case type_demotion_vec_info_type: | |
ebfd146a | 5910 | case type_promotion_vec_info_type: |
ebfd146a IR |
5911 | case type_conversion_vec_info_type: |
5912 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
5913 | gcc_assert (done); | |
5914 | break; | |
5915 | ||
5916 | case induc_vec_info_type: | |
5917 | gcc_assert (!slp_node); | |
5918 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
5919 | gcc_assert (done); | |
5920 | break; | |
5921 | ||
9dc3f7de IR |
5922 | case shift_vec_info_type: |
5923 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
5924 | gcc_assert (done); | |
5925 | break; | |
5926 | ||
ebfd146a IR |
5927 | case op_vec_info_type: |
5928 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
5929 | gcc_assert (done); | |
5930 | break; | |
5931 | ||
5932 | case assignment_vec_info_type: | |
5933 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
5934 | gcc_assert (done); | |
5935 | break; | |
5936 | ||
5937 | case load_vec_info_type: | |
b8698a0f | 5938 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
5939 | slp_node_instance); |
5940 | gcc_assert (done); | |
5941 | break; | |
5942 | ||
5943 | case store_vec_info_type: | |
5944 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
5945 | gcc_assert (done); | |
0d0293ac | 5946 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
5947 | { |
5948 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 5949 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
5950 | one are skipped, and there vec_stmt_info shouldn't be freed |
5951 | meanwhile. */ | |
0d0293ac | 5952 | *grouped_store = true; |
ebfd146a IR |
5953 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
5954 | is_store = true; | |
5955 | } | |
5956 | else | |
5957 | is_store = true; | |
5958 | break; | |
5959 | ||
5960 | case condition_vec_info_type: | |
f7e531cf | 5961 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
5962 | gcc_assert (done); |
5963 | break; | |
5964 | ||
5965 | case call_vec_info_type: | |
190c2236 | 5966 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 5967 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
5968 | break; |
5969 | ||
5970 | case reduc_vec_info_type: | |
b5aeb3bb | 5971 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
5972 | gcc_assert (done); |
5973 | break; | |
5974 | ||
5975 | default: | |
5976 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
5977 | { | |
73fbfcad | 5978 | if (dump_enabled_p ()) |
78c60e3d | 5979 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5980 | "stmt not supported.\n"); |
ebfd146a IR |
5981 | gcc_unreachable (); |
5982 | } | |
5983 | } | |
5984 | ||
5985 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that | |
5986 | is being vectorized, but outside the immediately enclosing loop. */ | |
5987 | if (vec_stmt | |
a70d6342 IR |
5988 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
5989 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
5990 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
5991 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
5992 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 5993 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 5994 | vect_used_in_outer_by_reduction)) |
ebfd146a | 5995 | { |
a70d6342 IR |
5996 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
5997 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
5998 | imm_use_iterator imm_iter; |
5999 | use_operand_p use_p; | |
6000 | tree scalar_dest; | |
6001 | gimple exit_phi; | |
6002 | ||
73fbfcad | 6003 | if (dump_enabled_p ()) |
78c60e3d | 6004 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6005 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
6006 | |
6007 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
6008 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
6009 | STMT). */ | |
6010 | if (gimple_code (stmt) == GIMPLE_PHI) | |
6011 | scalar_dest = PHI_RESULT (stmt); | |
6012 | else | |
6013 | scalar_dest = gimple_assign_lhs (stmt); | |
6014 | ||
6015 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
6016 | { | |
6017 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
6018 | { | |
6019 | exit_phi = USE_STMT (use_p); | |
6020 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
6021 | } | |
6022 | } | |
6023 | } | |
6024 | ||
6025 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
6026 | being vectorized. */ | |
6027 | if (STMT_VINFO_LIVE_P (stmt_info) | |
6028 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
6029 | { | |
6030 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
6031 | gcc_assert (done); | |
6032 | } | |
6033 | ||
6034 | if (vec_stmt) | |
83197f37 | 6035 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 6036 | |
b8698a0f | 6037 | return is_store; |
ebfd146a IR |
6038 | } |
6039 | ||
6040 | ||
b8698a0f | 6041 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
6042 | stmt_vec_info. */ |
6043 | ||
6044 | void | |
6045 | vect_remove_stores (gimple first_stmt) | |
6046 | { | |
6047 | gimple next = first_stmt; | |
6048 | gimple tmp; | |
6049 | gimple_stmt_iterator next_si; | |
6050 | ||
6051 | while (next) | |
6052 | { | |
78048b1c JJ |
6053 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
6054 | ||
6055 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
6056 | if (is_pattern_stmt_p (stmt_info)) | |
6057 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
6058 | /* Free the attached stmt_vec_info and remove the stmt. */ |
6059 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 6060 | unlink_stmt_vdef (next); |
ebfd146a | 6061 | gsi_remove (&next_si, true); |
3d3f2249 | 6062 | release_defs (next); |
ebfd146a IR |
6063 | free_stmt_vec_info (next); |
6064 | next = tmp; | |
6065 | } | |
6066 | } | |
6067 | ||
6068 | ||
6069 | /* Function new_stmt_vec_info. | |
6070 | ||
6071 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
6072 | ||
6073 | stmt_vec_info | |
b8698a0f | 6074 | new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 6075 | bb_vec_info bb_vinfo) |
ebfd146a IR |
6076 | { |
6077 | stmt_vec_info res; | |
6078 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
6079 | ||
6080 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
6081 | STMT_VINFO_STMT (res) = stmt; | |
6082 | STMT_VINFO_LOOP_VINFO (res) = loop_vinfo; | |
a70d6342 | 6083 | STMT_VINFO_BB_VINFO (res) = bb_vinfo; |
8644a673 | 6084 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
6085 | STMT_VINFO_LIVE_P (res) = false; |
6086 | STMT_VINFO_VECTYPE (res) = NULL; | |
6087 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 6088 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
6089 | STMT_VINFO_IN_PATTERN_P (res) = false; |
6090 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 6091 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a IR |
6092 | STMT_VINFO_DATA_REF (res) = NULL; |
6093 | ||
6094 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
6095 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
6096 | STMT_VINFO_DR_INIT (res) = NULL; | |
6097 | STMT_VINFO_DR_STEP (res) = NULL; | |
6098 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
6099 | ||
6100 | if (gimple_code (stmt) == GIMPLE_PHI | |
6101 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
6102 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
6103 | else | |
8644a673 IR |
6104 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
6105 | ||
9771b263 | 6106 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 6107 | STMT_SLP_TYPE (res) = loop_vect; |
e14c1050 IR |
6108 | GROUP_FIRST_ELEMENT (res) = NULL; |
6109 | GROUP_NEXT_ELEMENT (res) = NULL; | |
6110 | GROUP_SIZE (res) = 0; | |
6111 | GROUP_STORE_COUNT (res) = 0; | |
6112 | GROUP_GAP (res) = 0; | |
6113 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
6114 | |
6115 | return res; | |
6116 | } | |
6117 | ||
6118 | ||
6119 | /* Create a hash table for stmt_vec_info. */ | |
6120 | ||
6121 | void | |
6122 | init_stmt_vec_info_vec (void) | |
6123 | { | |
9771b263 DN |
6124 | gcc_assert (!stmt_vec_info_vec.exists ()); |
6125 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
6126 | } |
6127 | ||
6128 | ||
6129 | /* Free hash table for stmt_vec_info. */ | |
6130 | ||
6131 | void | |
6132 | free_stmt_vec_info_vec (void) | |
6133 | { | |
93675444 JJ |
6134 | unsigned int i; |
6135 | vec_void_p info; | |
6136 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) | |
6137 | if (info != NULL) | |
6138 | free_stmt_vec_info (STMT_VINFO_STMT ((stmt_vec_info) info)); | |
9771b263 DN |
6139 | gcc_assert (stmt_vec_info_vec.exists ()); |
6140 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
6141 | } |
6142 | ||
6143 | ||
6144 | /* Free stmt vectorization related info. */ | |
6145 | ||
6146 | void | |
6147 | free_stmt_vec_info (gimple stmt) | |
6148 | { | |
6149 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6150 | ||
6151 | if (!stmt_info) | |
6152 | return; | |
6153 | ||
78048b1c JJ |
6154 | /* Check if this statement has a related "pattern stmt" |
6155 | (introduced by the vectorizer during the pattern recognition | |
6156 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
6157 | too. */ | |
6158 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
6159 | { | |
6160 | stmt_vec_info patt_info | |
6161 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
6162 | if (patt_info) | |
6163 | { | |
363477c0 JJ |
6164 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
6165 | if (seq) | |
6166 | { | |
6167 | gimple_stmt_iterator si; | |
6168 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
6169 | free_stmt_vec_info (gsi_stmt (si)); | |
6170 | } | |
78048b1c JJ |
6171 | free_stmt_vec_info (STMT_VINFO_RELATED_STMT (stmt_info)); |
6172 | } | |
6173 | } | |
6174 | ||
9771b263 | 6175 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
ebfd146a IR |
6176 | set_vinfo_for_stmt (stmt, NULL); |
6177 | free (stmt_info); | |
6178 | } | |
6179 | ||
6180 | ||
bb67d9c7 | 6181 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 6182 | |
bb67d9c7 | 6183 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
6184 | by the target. */ |
6185 | ||
bb67d9c7 RG |
6186 | static tree |
6187 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a IR |
6188 | { |
6189 | enum machine_mode inner_mode = TYPE_MODE (scalar_type); | |
cc4b5170 | 6190 | enum machine_mode simd_mode; |
2f816591 | 6191 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
6192 | int nunits; |
6193 | tree vectype; | |
6194 | ||
cc4b5170 | 6195 | if (nbytes == 0) |
ebfd146a IR |
6196 | return NULL_TREE; |
6197 | ||
48f2e373 RB |
6198 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
6199 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
6200 | return NULL_TREE; | |
6201 | ||
7b7b1813 RG |
6202 | /* For vector types of elements whose mode precision doesn't |
6203 | match their types precision we use a element type of mode | |
6204 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
6205 | they support the proper result truncation/extension. |
6206 | We also make sure to build vector types with INTEGER_TYPE | |
6207 | component type only. */ | |
6d7971b8 | 6208 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
6209 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
6210 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
6211 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
6212 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 6213 | |
ccbf5bb4 RG |
6214 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
6215 | When the component mode passes the above test simply use a type | |
6216 | corresponding to that mode. The theory is that any use that | |
6217 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 6218 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 6219 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
6220 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
6221 | ||
6222 | /* We can't build a vector type of elements with alignment bigger than | |
6223 | their size. */ | |
dfc2e2ac | 6224 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
6225 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
6226 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 6227 | |
dfc2e2ac RB |
6228 | /* If we felt back to using the mode fail if there was |
6229 | no scalar type for it. */ | |
6230 | if (scalar_type == NULL_TREE) | |
6231 | return NULL_TREE; | |
6232 | ||
bb67d9c7 RG |
6233 | /* If no size was supplied use the mode the target prefers. Otherwise |
6234 | lookup a vector mode of the specified size. */ | |
6235 | if (size == 0) | |
6236 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
6237 | else | |
6238 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
6239 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
6240 | if (nunits <= 1) | |
6241 | return NULL_TREE; | |
ebfd146a IR |
6242 | |
6243 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
6244 | |
6245 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
6246 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 6247 | return NULL_TREE; |
ebfd146a IR |
6248 | |
6249 | return vectype; | |
6250 | } | |
6251 | ||
bb67d9c7 RG |
6252 | unsigned int current_vector_size; |
6253 | ||
6254 | /* Function get_vectype_for_scalar_type. | |
6255 | ||
6256 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
6257 | by the target. */ | |
6258 | ||
6259 | tree | |
6260 | get_vectype_for_scalar_type (tree scalar_type) | |
6261 | { | |
6262 | tree vectype; | |
6263 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
6264 | current_vector_size); | |
6265 | if (vectype | |
6266 | && current_vector_size == 0) | |
6267 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
6268 | return vectype; | |
6269 | } | |
6270 | ||
b690cc0f RG |
6271 | /* Function get_same_sized_vectype |
6272 | ||
6273 | Returns a vector type corresponding to SCALAR_TYPE of size | |
6274 | VECTOR_TYPE if supported by the target. */ | |
6275 | ||
6276 | tree | |
bb67d9c7 | 6277 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 6278 | { |
bb67d9c7 RG |
6279 | return get_vectype_for_scalar_type_and_size |
6280 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
6281 | } |
6282 | ||
ebfd146a IR |
6283 | /* Function vect_is_simple_use. |
6284 | ||
6285 | Input: | |
a70d6342 IR |
6286 | LOOP_VINFO - the vect info of the loop that is being vectorized. |
6287 | BB_VINFO - the vect info of the basic block that is being vectorized. | |
24ee1384 | 6288 | OPERAND - operand of STMT in the loop or bb. |
ebfd146a IR |
6289 | DEF - the defining stmt in case OPERAND is an SSA_NAME. |
6290 | ||
6291 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 6292 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 6293 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 6294 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
6295 | is the case in reduction/induction computations). |
6296 | For basic blocks, supportable operands are constants and bb invariants. | |
6297 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
6298 | |
6299 | bool | |
24ee1384 | 6300 | vect_is_simple_use (tree operand, gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 6301 | bb_vec_info bb_vinfo, gimple *def_stmt, |
ebfd146a | 6302 | tree *def, enum vect_def_type *dt) |
b8698a0f | 6303 | { |
ebfd146a IR |
6304 | basic_block bb; |
6305 | stmt_vec_info stmt_vinfo; | |
a70d6342 | 6306 | struct loop *loop = NULL; |
b8698a0f | 6307 | |
a70d6342 IR |
6308 | if (loop_vinfo) |
6309 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
6310 | |
6311 | *def_stmt = NULL; | |
6312 | *def = NULL_TREE; | |
b8698a0f | 6313 | |
73fbfcad | 6314 | if (dump_enabled_p ()) |
ebfd146a | 6315 | { |
78c60e3d SS |
6316 | dump_printf_loc (MSG_NOTE, vect_location, |
6317 | "vect_is_simple_use: operand "); | |
6318 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 6319 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 6320 | } |
b8698a0f | 6321 | |
b758f602 | 6322 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
6323 | { |
6324 | *dt = vect_constant_def; | |
6325 | return true; | |
6326 | } | |
b8698a0f | 6327 | |
ebfd146a IR |
6328 | if (is_gimple_min_invariant (operand)) |
6329 | { | |
6330 | *def = operand; | |
8644a673 | 6331 | *dt = vect_external_def; |
ebfd146a IR |
6332 | return true; |
6333 | } | |
6334 | ||
6335 | if (TREE_CODE (operand) == PAREN_EXPR) | |
6336 | { | |
73fbfcad | 6337 | if (dump_enabled_p ()) |
e645e942 | 6338 | dump_printf_loc (MSG_NOTE, vect_location, "non-associatable copy.\n"); |
ebfd146a IR |
6339 | operand = TREE_OPERAND (operand, 0); |
6340 | } | |
b8698a0f | 6341 | |
ebfd146a IR |
6342 | if (TREE_CODE (operand) != SSA_NAME) |
6343 | { | |
73fbfcad | 6344 | if (dump_enabled_p ()) |
78c60e3d | 6345 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6346 | "not ssa-name.\n"); |
ebfd146a IR |
6347 | return false; |
6348 | } | |
b8698a0f | 6349 | |
ebfd146a IR |
6350 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
6351 | if (*def_stmt == NULL) | |
6352 | { | |
73fbfcad | 6353 | if (dump_enabled_p ()) |
78c60e3d | 6354 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6355 | "no def_stmt.\n"); |
ebfd146a IR |
6356 | return false; |
6357 | } | |
6358 | ||
73fbfcad | 6359 | if (dump_enabled_p ()) |
ebfd146a | 6360 | { |
78c60e3d SS |
6361 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
6362 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
e645e942 | 6363 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
6364 | } |
6365 | ||
8644a673 | 6366 | /* Empty stmt is expected only in case of a function argument. |
ebfd146a IR |
6367 | (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */ |
6368 | if (gimple_nop_p (*def_stmt)) | |
6369 | { | |
6370 | *def = operand; | |
8644a673 | 6371 | *dt = vect_external_def; |
ebfd146a IR |
6372 | return true; |
6373 | } | |
6374 | ||
6375 | bb = gimple_bb (*def_stmt); | |
a70d6342 IR |
6376 | |
6377 | if ((loop && !flow_bb_inside_loop_p (loop, bb)) | |
6378 | || (!loop && bb != BB_VINFO_BB (bb_vinfo)) | |
b8698a0f | 6379 | || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI)) |
8644a673 | 6380 | *dt = vect_external_def; |
ebfd146a IR |
6381 | else |
6382 | { | |
6383 | stmt_vinfo = vinfo_for_stmt (*def_stmt); | |
6384 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
6385 | } | |
6386 | ||
24ee1384 IR |
6387 | if (*dt == vect_unknown_def_type |
6388 | || (stmt | |
6389 | && *dt == vect_double_reduction_def | |
6390 | && gimple_code (stmt) != GIMPLE_PHI)) | |
ebfd146a | 6391 | { |
73fbfcad | 6392 | if (dump_enabled_p ()) |
78c60e3d | 6393 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6394 | "Unsupported pattern.\n"); |
ebfd146a IR |
6395 | return false; |
6396 | } | |
6397 | ||
73fbfcad | 6398 | if (dump_enabled_p ()) |
e645e942 | 6399 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: %d.\n", *dt); |
ebfd146a IR |
6400 | |
6401 | switch (gimple_code (*def_stmt)) | |
6402 | { | |
6403 | case GIMPLE_PHI: | |
6404 | *def = gimple_phi_result (*def_stmt); | |
6405 | break; | |
6406 | ||
6407 | case GIMPLE_ASSIGN: | |
6408 | *def = gimple_assign_lhs (*def_stmt); | |
6409 | break; | |
6410 | ||
6411 | case GIMPLE_CALL: | |
6412 | *def = gimple_call_lhs (*def_stmt); | |
6413 | if (*def != NULL) | |
6414 | break; | |
6415 | /* FALLTHRU */ | |
6416 | default: | |
73fbfcad | 6417 | if (dump_enabled_p ()) |
78c60e3d | 6418 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6419 | "unsupported defining stmt:\n"); |
ebfd146a IR |
6420 | return false; |
6421 | } | |
6422 | ||
6423 | return true; | |
6424 | } | |
6425 | ||
b690cc0f RG |
6426 | /* Function vect_is_simple_use_1. |
6427 | ||
6428 | Same as vect_is_simple_use_1 but also determines the vector operand | |
6429 | type of OPERAND and stores it to *VECTYPE. If the definition of | |
6430 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
6431 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
6432 | is responsible to compute the best suited vector type for the | |
6433 | scalar operand. */ | |
6434 | ||
6435 | bool | |
24ee1384 | 6436 | vect_is_simple_use_1 (tree operand, gimple stmt, loop_vec_info loop_vinfo, |
b690cc0f RG |
6437 | bb_vec_info bb_vinfo, gimple *def_stmt, |
6438 | tree *def, enum vect_def_type *dt, tree *vectype) | |
6439 | { | |
24ee1384 IR |
6440 | if (!vect_is_simple_use (operand, stmt, loop_vinfo, bb_vinfo, def_stmt, |
6441 | def, dt)) | |
b690cc0f RG |
6442 | return false; |
6443 | ||
6444 | /* Now get a vector type if the def is internal, otherwise supply | |
6445 | NULL_TREE and leave it up to the caller to figure out a proper | |
6446 | type for the use stmt. */ | |
6447 | if (*dt == vect_internal_def | |
6448 | || *dt == vect_induction_def | |
6449 | || *dt == vect_reduction_def | |
6450 | || *dt == vect_double_reduction_def | |
6451 | || *dt == vect_nested_cycle) | |
6452 | { | |
6453 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
6454 | |
6455 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
6456 | && !STMT_VINFO_RELEVANT (stmt_info) | |
6457 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 6458 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 6459 | |
b690cc0f RG |
6460 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
6461 | gcc_assert (*vectype != NULL_TREE); | |
6462 | } | |
6463 | else if (*dt == vect_uninitialized_def | |
6464 | || *dt == vect_constant_def | |
6465 | || *dt == vect_external_def) | |
6466 | *vectype = NULL_TREE; | |
6467 | else | |
6468 | gcc_unreachable (); | |
6469 | ||
6470 | return true; | |
6471 | } | |
6472 | ||
ebfd146a IR |
6473 | |
6474 | /* Function supportable_widening_operation | |
6475 | ||
b8698a0f L |
6476 | Check whether an operation represented by the code CODE is a |
6477 | widening operation that is supported by the target platform in | |
b690cc0f RG |
6478 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
6479 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 6480 | |
ebfd146a IR |
6481 | Widening operations we currently support are NOP (CONVERT), FLOAT |
6482 | and WIDEN_MULT. This function checks if these operations are supported | |
6483 | by the target platform either directly (via vector tree-codes), or via | |
6484 | target builtins. | |
6485 | ||
6486 | Output: | |
b8698a0f L |
6487 | - CODE1 and CODE2 are codes of vector operations to be used when |
6488 | vectorizing the operation, if available. | |
ebfd146a IR |
6489 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
6490 | case of multi-step conversion (like char->short->int - in that case | |
6491 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
6492 | - INTERM_TYPES contains the intermediate type required to perform the |
6493 | widening operation (short in the above example). */ | |
ebfd146a IR |
6494 | |
6495 | bool | |
b690cc0f RG |
6496 | supportable_widening_operation (enum tree_code code, gimple stmt, |
6497 | tree vectype_out, tree vectype_in, | |
ebfd146a IR |
6498 | enum tree_code *code1, enum tree_code *code2, |
6499 | int *multi_step_cvt, | |
9771b263 | 6500 | vec<tree> *interm_types) |
ebfd146a IR |
6501 | { |
6502 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6503 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 6504 | struct loop *vect_loop = NULL; |
ebfd146a | 6505 | enum machine_mode vec_mode; |
81f40b79 | 6506 | enum insn_code icode1, icode2; |
ebfd146a | 6507 | optab optab1, optab2; |
b690cc0f RG |
6508 | tree vectype = vectype_in; |
6509 | tree wide_vectype = vectype_out; | |
ebfd146a | 6510 | enum tree_code c1, c2; |
4a00c761 JJ |
6511 | int i; |
6512 | tree prev_type, intermediate_type; | |
6513 | enum machine_mode intermediate_mode, prev_mode; | |
6514 | optab optab3, optab4; | |
ebfd146a | 6515 | |
4a00c761 | 6516 | *multi_step_cvt = 0; |
4ef69dfc IR |
6517 | if (loop_info) |
6518 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
6519 | ||
ebfd146a IR |
6520 | switch (code) |
6521 | { | |
6522 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
6523 | /* The result of a vectorized widening operation usually requires |
6524 | two vectors (because the widened results do not fit into one vector). | |
6525 | The generated vector results would normally be expected to be | |
6526 | generated in the same order as in the original scalar computation, | |
6527 | i.e. if 8 results are generated in each vector iteration, they are | |
6528 | to be organized as follows: | |
6529 | vect1: [res1,res2,res3,res4], | |
6530 | vect2: [res5,res6,res7,res8]. | |
6531 | ||
6532 | However, in the special case that the result of the widening | |
6533 | operation is used in a reduction computation only, the order doesn't | |
6534 | matter (because when vectorizing a reduction we change the order of | |
6535 | the computation). Some targets can take advantage of this and | |
6536 | generate more efficient code. For example, targets like Altivec, | |
6537 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
6538 | generate the following vectors: | |
6539 | vect1: [res1,res3,res5,res7], | |
6540 | vect2: [res2,res4,res6,res8]. | |
6541 | ||
6542 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
6543 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
6544 | iterations in parallel). We therefore don't allow to change the | |
6545 | order of the computation in the inner-loop during outer-loop | |
6546 | vectorization. */ | |
6547 | /* TODO: Another case in which order doesn't *really* matter is when we | |
6548 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
6549 | Normally, pack_trunc performs an even/odd permute, whereas the | |
6550 | repack from an even/odd expansion would be an interleave, which | |
6551 | would be significantly simpler for e.g. AVX2. */ | |
6552 | /* In any case, in order to avoid duplicating the code below, recurse | |
6553 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
6554 | are properly set up for the caller. If we fail, we'll continue with | |
6555 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
6556 | if (vect_loop | |
6557 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
6558 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
6559 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
6560 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
6561 | code1, code2, multi_step_cvt, |
6562 | interm_types)) | |
6ae6116f | 6563 | return true; |
4a00c761 JJ |
6564 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
6565 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
6566 | break; |
6567 | ||
6ae6116f RH |
6568 | case VEC_WIDEN_MULT_EVEN_EXPR: |
6569 | /* Support the recursion induced just above. */ | |
6570 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
6571 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
6572 | break; | |
6573 | ||
36ba4aae | 6574 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
6575 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
6576 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
6577 | break; |
6578 | ||
ebfd146a | 6579 | CASE_CONVERT: |
4a00c761 JJ |
6580 | c1 = VEC_UNPACK_LO_EXPR; |
6581 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
6582 | break; |
6583 | ||
6584 | case FLOAT_EXPR: | |
4a00c761 JJ |
6585 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
6586 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
6587 | break; |
6588 | ||
6589 | case FIX_TRUNC_EXPR: | |
6590 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
6591 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
6592 | computing the operation. */ | |
6593 | return false; | |
6594 | ||
6595 | default: | |
6596 | gcc_unreachable (); | |
6597 | } | |
6598 | ||
6ae6116f | 6599 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
4a00c761 JJ |
6600 | { |
6601 | enum tree_code ctmp = c1; | |
6602 | c1 = c2; | |
6603 | c2 = ctmp; | |
6604 | } | |
6605 | ||
ebfd146a IR |
6606 | if (code == FIX_TRUNC_EXPR) |
6607 | { | |
6608 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
6609 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
6610 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
6611 | } |
6612 | else | |
6613 | { | |
6614 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
6615 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
6616 | } | |
6617 | ||
6618 | if (!optab1 || !optab2) | |
6619 | return false; | |
6620 | ||
6621 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
6622 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
6623 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
6624 | return false; |
6625 | ||
4a00c761 JJ |
6626 | *code1 = c1; |
6627 | *code2 = c2; | |
6628 | ||
6629 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
6630 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
6631 | return true; | |
6632 | ||
b8698a0f | 6633 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 6634 | types. */ |
ebfd146a | 6635 | |
4a00c761 JJ |
6636 | prev_type = vectype; |
6637 | prev_mode = vec_mode; | |
b8698a0f | 6638 | |
4a00c761 JJ |
6639 | if (!CONVERT_EXPR_CODE_P (code)) |
6640 | return false; | |
b8698a0f | 6641 | |
4a00c761 JJ |
6642 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
6643 | intermediate steps in promotion sequence. We try | |
6644 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
6645 | not. */ | |
9771b263 | 6646 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
6647 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
6648 | { | |
6649 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
6650 | intermediate_type | |
6651 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
6652 | TYPE_UNSIGNED (prev_type)); | |
6653 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
6654 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
6655 | ||
6656 | if (!optab3 || !optab4 | |
6657 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
6658 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
6659 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
6660 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
6661 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
6662 | == CODE_FOR_nothing) | |
6663 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
6664 | == CODE_FOR_nothing)) | |
6665 | break; | |
ebfd146a | 6666 | |
9771b263 | 6667 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
6668 | (*multi_step_cvt)++; |
6669 | ||
6670 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
6671 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
6672 | return true; | |
6673 | ||
6674 | prev_type = intermediate_type; | |
6675 | prev_mode = intermediate_mode; | |
ebfd146a IR |
6676 | } |
6677 | ||
9771b263 | 6678 | interm_types->release (); |
4a00c761 | 6679 | return false; |
ebfd146a IR |
6680 | } |
6681 | ||
6682 | ||
6683 | /* Function supportable_narrowing_operation | |
6684 | ||
b8698a0f L |
6685 | Check whether an operation represented by the code CODE is a |
6686 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
6687 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
6688 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 6689 | |
ebfd146a | 6690 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 6691 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
6692 | the target platform directly via vector tree-codes. |
6693 | ||
6694 | Output: | |
b8698a0f L |
6695 | - CODE1 is the code of a vector operation to be used when |
6696 | vectorizing the operation, if available. | |
ebfd146a IR |
6697 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
6698 | case of multi-step conversion (like int->short->char - in that case | |
6699 | MULTI_STEP_CVT will be 1). | |
6700 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 6701 | narrowing operation (short in the above example). */ |
ebfd146a IR |
6702 | |
6703 | bool | |
6704 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 6705 | tree vectype_out, tree vectype_in, |
ebfd146a | 6706 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 6707 | vec<tree> *interm_types) |
ebfd146a IR |
6708 | { |
6709 | enum machine_mode vec_mode; | |
6710 | enum insn_code icode1; | |
6711 | optab optab1, interm_optab; | |
b690cc0f RG |
6712 | tree vectype = vectype_in; |
6713 | tree narrow_vectype = vectype_out; | |
ebfd146a | 6714 | enum tree_code c1; |
4a00c761 JJ |
6715 | tree intermediate_type; |
6716 | enum machine_mode intermediate_mode, prev_mode; | |
ebfd146a | 6717 | int i; |
4a00c761 | 6718 | bool uns; |
ebfd146a | 6719 | |
4a00c761 | 6720 | *multi_step_cvt = 0; |
ebfd146a IR |
6721 | switch (code) |
6722 | { | |
6723 | CASE_CONVERT: | |
6724 | c1 = VEC_PACK_TRUNC_EXPR; | |
6725 | break; | |
6726 | ||
6727 | case FIX_TRUNC_EXPR: | |
6728 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
6729 | break; | |
6730 | ||
6731 | case FLOAT_EXPR: | |
6732 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
6733 | tree code and optabs used for computing the operation. */ | |
6734 | return false; | |
6735 | ||
6736 | default: | |
6737 | gcc_unreachable (); | |
6738 | } | |
6739 | ||
6740 | if (code == FIX_TRUNC_EXPR) | |
6741 | /* The signedness is determined from output operand. */ | |
b690cc0f | 6742 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
6743 | else |
6744 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
6745 | ||
6746 | if (!optab1) | |
6747 | return false; | |
6748 | ||
6749 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 6750 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
6751 | return false; |
6752 | ||
4a00c761 JJ |
6753 | *code1 = c1; |
6754 | ||
6755 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
6756 | return true; | |
6757 | ||
ebfd146a IR |
6758 | /* Check if it's a multi-step conversion that can be done using intermediate |
6759 | types. */ | |
4a00c761 JJ |
6760 | prev_mode = vec_mode; |
6761 | if (code == FIX_TRUNC_EXPR) | |
6762 | uns = TYPE_UNSIGNED (vectype_out); | |
6763 | else | |
6764 | uns = TYPE_UNSIGNED (vectype); | |
6765 | ||
6766 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
6767 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
6768 | costly than signed. */ | |
6769 | if (code == FIX_TRUNC_EXPR && uns) | |
6770 | { | |
6771 | enum insn_code icode2; | |
6772 | ||
6773 | intermediate_type | |
6774 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
6775 | interm_optab | |
6776 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 6777 | if (interm_optab != unknown_optab |
4a00c761 JJ |
6778 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
6779 | && insn_data[icode1].operand[0].mode | |
6780 | == insn_data[icode2].operand[0].mode) | |
6781 | { | |
6782 | uns = false; | |
6783 | optab1 = interm_optab; | |
6784 | icode1 = icode2; | |
6785 | } | |
6786 | } | |
ebfd146a | 6787 | |
4a00c761 JJ |
6788 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
6789 | intermediate steps in promotion sequence. We try | |
6790 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 6791 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
6792 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
6793 | { | |
6794 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
6795 | intermediate_type | |
6796 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
6797 | interm_optab | |
6798 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
6799 | optab_default); | |
6800 | if (!interm_optab | |
6801 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
6802 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
6803 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
6804 | == CODE_FOR_nothing)) | |
6805 | break; | |
6806 | ||
9771b263 | 6807 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
6808 | (*multi_step_cvt)++; |
6809 | ||
6810 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
6811 | return true; | |
6812 | ||
6813 | prev_mode = intermediate_mode; | |
6814 | optab1 = interm_optab; | |
ebfd146a IR |
6815 | } |
6816 | ||
9771b263 | 6817 | interm_types->release (); |
4a00c761 | 6818 | return false; |
ebfd146a | 6819 | } |