]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
d1e082c2 | 2 | Copyright (C) 2003-2013 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
78c60e3d | 25 | #include "dumpfile.h" |
ebfd146a IR |
26 | #include "tm.h" |
27 | #include "ggc.h" | |
28 | #include "tree.h" | |
29 | #include "target.h" | |
30 | #include "basic-block.h" | |
cf835838 | 31 | #include "gimple-pretty-print.h" |
7a300452 | 32 | #include "tree-ssa.h" |
ebfd146a | 33 | #include "cfgloop.h" |
ebfd146a | 34 | #include "expr.h" |
7ee2468b | 35 | #include "recog.h" /* FIXME: for insn_data */ |
ebfd146a | 36 | #include "optabs.h" |
718f9c0f | 37 | #include "diagnostic-core.h" |
ebfd146a | 38 | #include "tree-vectorizer.h" |
7ee2468b | 39 | #include "dumpfile.h" |
ebfd146a | 40 | |
7ee2468b SB |
41 | /* For lang_hooks.types.type_for_mode. */ |
42 | #include "langhooks.h" | |
ebfd146a | 43 | |
c3e7ee41 BS |
44 | /* Return the vectorized type for the given statement. */ |
45 | ||
46 | tree | |
47 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
48 | { | |
49 | return STMT_VINFO_VECTYPE (stmt_info); | |
50 | } | |
51 | ||
52 | /* Return TRUE iff the given statement is in an inner loop relative to | |
53 | the loop being vectorized. */ | |
54 | bool | |
55 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
56 | { | |
57 | gimple stmt = STMT_VINFO_STMT (stmt_info); | |
58 | basic_block bb = gimple_bb (stmt); | |
59 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
60 | struct loop* loop; | |
61 | ||
62 | if (!loop_vinfo) | |
63 | return false; | |
64 | ||
65 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
66 | ||
67 | return (bb->loop_father == loop->inner); | |
68 | } | |
69 | ||
70 | /* Record the cost of a statement, either by directly informing the | |
71 | target model or by saving it in a vector for later processing. | |
72 | Return a preliminary estimate of the statement's cost. */ | |
73 | ||
74 | unsigned | |
92345349 | 75 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 76 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 77 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 78 | { |
92345349 | 79 | if (body_cost_vec) |
c3e7ee41 | 80 | { |
92345349 BS |
81 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
82 | add_stmt_info_to_vec (body_cost_vec, count, kind, | |
83 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
84 | misalign); | |
c3e7ee41 | 85 | return (unsigned) |
92345349 | 86 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
87 | |
88 | } | |
89 | else | |
90 | { | |
91 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
92 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
93 | void *target_cost_data; | |
94 | ||
95 | if (loop_vinfo) | |
96 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
97 | else | |
98 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
99 | ||
92345349 BS |
100 | return add_stmt_cost (target_cost_data, count, kind, stmt_info, |
101 | misalign, where); | |
c3e7ee41 BS |
102 | } |
103 | } | |
104 | ||
272c6793 RS |
105 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
106 | ||
107 | static tree | |
108 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
109 | { | |
110 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
111 | "vect_array"); | |
112 | } | |
113 | ||
114 | /* ARRAY is an array of vectors created by create_vector_array. | |
115 | Return an SSA_NAME for the vector in index N. The reference | |
116 | is part of the vectorization of STMT and the vector is associated | |
117 | with scalar destination SCALAR_DEST. */ | |
118 | ||
119 | static tree | |
120 | read_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree scalar_dest, | |
121 | tree array, unsigned HOST_WIDE_INT n) | |
122 | { | |
123 | tree vect_type, vect, vect_name, array_ref; | |
124 | gimple new_stmt; | |
125 | ||
126 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
127 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
128 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
129 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
130 | build_int_cst (size_type_node, n), | |
131 | NULL_TREE, NULL_TREE); | |
132 | ||
133 | new_stmt = gimple_build_assign (vect, array_ref); | |
134 | vect_name = make_ssa_name (vect, new_stmt); | |
135 | gimple_assign_set_lhs (new_stmt, vect_name); | |
136 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
137 | |
138 | return vect_name; | |
139 | } | |
140 | ||
141 | /* ARRAY is an array of vectors created by create_vector_array. | |
142 | Emit code to store SSA_NAME VECT in index N of the array. | |
143 | The store is part of the vectorization of STMT. */ | |
144 | ||
145 | static void | |
146 | write_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree vect, | |
147 | tree array, unsigned HOST_WIDE_INT n) | |
148 | { | |
149 | tree array_ref; | |
150 | gimple new_stmt; | |
151 | ||
152 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
153 | build_int_cst (size_type_node, n), | |
154 | NULL_TREE, NULL_TREE); | |
155 | ||
156 | new_stmt = gimple_build_assign (array_ref, vect); | |
157 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
158 | } |
159 | ||
160 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
161 | of *PTR. The memory reference replaces those in FIRST_DR | |
162 | (and its group). */ | |
163 | ||
164 | static tree | |
165 | create_array_ref (tree type, tree ptr, struct data_reference *first_dr) | |
166 | { | |
272c6793 RS |
167 | tree mem_ref, alias_ptr_type; |
168 | ||
169 | alias_ptr_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
170 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); | |
171 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 172 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
173 | return mem_ref; |
174 | } | |
175 | ||
ebfd146a IR |
176 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
177 | ||
178 | /* Function vect_mark_relevant. | |
179 | ||
180 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
181 | ||
182 | static void | |
9771b263 | 183 | vect_mark_relevant (vec<gimple> *worklist, gimple stmt, |
83197f37 IR |
184 | enum vect_relevant relevant, bool live_p, |
185 | bool used_in_pattern) | |
ebfd146a IR |
186 | { |
187 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
188 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
189 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
83197f37 | 190 | gimple pattern_stmt; |
ebfd146a | 191 | |
73fbfcad | 192 | if (dump_enabled_p ()) |
78c60e3d | 193 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 194 | "mark relevant %d, live %d.\n", relevant, live_p); |
ebfd146a | 195 | |
83197f37 IR |
196 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
197 | related pattern stmt instead of the original stmt. However, such stmts | |
198 | may have their own uses that are not in any pattern, in such cases the | |
199 | stmt itself should be marked. */ | |
ebfd146a IR |
200 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
201 | { | |
83197f37 IR |
202 | bool found = false; |
203 | if (!used_in_pattern) | |
204 | { | |
205 | imm_use_iterator imm_iter; | |
206 | use_operand_p use_p; | |
207 | gimple use_stmt; | |
208 | tree lhs; | |
13c931c9 JJ |
209 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
210 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a | 211 | |
83197f37 IR |
212 | if (is_gimple_assign (stmt)) |
213 | lhs = gimple_assign_lhs (stmt); | |
214 | else | |
215 | lhs = gimple_call_lhs (stmt); | |
ebfd146a | 216 | |
83197f37 IR |
217 | /* This use is out of pattern use, if LHS has other uses that are |
218 | pattern uses, we should mark the stmt itself, and not the pattern | |
219 | stmt. */ | |
ab0ef706 JJ |
220 | if (TREE_CODE (lhs) == SSA_NAME) |
221 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs) | |
222 | { | |
223 | if (is_gimple_debug (USE_STMT (use_p))) | |
224 | continue; | |
225 | use_stmt = USE_STMT (use_p); | |
226 | ||
13c931c9 JJ |
227 | if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) |
228 | continue; | |
229 | ||
ab0ef706 JJ |
230 | if (vinfo_for_stmt (use_stmt) |
231 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt))) | |
232 | { | |
233 | found = true; | |
234 | break; | |
235 | } | |
236 | } | |
83197f37 IR |
237 | } |
238 | ||
239 | if (!found) | |
240 | { | |
241 | /* This is the last stmt in a sequence that was detected as a | |
242 | pattern that can potentially be vectorized. Don't mark the stmt | |
243 | as relevant/live because it's not going to be vectorized. | |
244 | Instead mark the pattern-stmt that replaces it. */ | |
245 | ||
246 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
247 | ||
73fbfcad | 248 | if (dump_enabled_p ()) |
78c60e3d SS |
249 | dump_printf_loc (MSG_NOTE, vect_location, |
250 | "last stmt in pattern. don't mark" | |
e645e942 | 251 | " relevant/live.\n"); |
83197f37 IR |
252 | stmt_info = vinfo_for_stmt (pattern_stmt); |
253 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
254 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
255 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
256 | stmt = pattern_stmt; | |
257 | } | |
ebfd146a IR |
258 | } |
259 | ||
260 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
261 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
262 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
263 | ||
264 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
265 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
266 | { | |
73fbfcad | 267 | if (dump_enabled_p ()) |
78c60e3d | 268 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 269 | "already marked relevant/live.\n"); |
ebfd146a IR |
270 | return; |
271 | } | |
272 | ||
9771b263 | 273 | worklist->safe_push (stmt); |
ebfd146a IR |
274 | } |
275 | ||
276 | ||
277 | /* Function vect_stmt_relevant_p. | |
278 | ||
279 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
280 | "relevant for vectorization". | |
281 | ||
282 | A stmt is considered "relevant for vectorization" if: | |
283 | - it has uses outside the loop. | |
284 | - it has vdefs (it alters memory). | |
285 | - control stmts in the loop (except for the exit condition). | |
286 | ||
287 | CHECKME: what other side effects would the vectorizer allow? */ | |
288 | ||
289 | static bool | |
290 | vect_stmt_relevant_p (gimple stmt, loop_vec_info loop_vinfo, | |
291 | enum vect_relevant *relevant, bool *live_p) | |
292 | { | |
293 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
294 | ssa_op_iter op_iter; | |
295 | imm_use_iterator imm_iter; | |
296 | use_operand_p use_p; | |
297 | def_operand_p def_p; | |
298 | ||
8644a673 | 299 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
300 | *live_p = false; |
301 | ||
302 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
303 | if (is_ctrl_stmt (stmt) |
304 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
305 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 306 | *relevant = vect_used_in_scope; |
ebfd146a IR |
307 | |
308 | /* changing memory. */ | |
309 | if (gimple_code (stmt) != GIMPLE_PHI) | |
5006671f | 310 | if (gimple_vdef (stmt)) |
ebfd146a | 311 | { |
73fbfcad | 312 | if (dump_enabled_p ()) |
78c60e3d | 313 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 314 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 315 | *relevant = vect_used_in_scope; |
ebfd146a IR |
316 | } |
317 | ||
318 | /* uses outside the loop. */ | |
319 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
320 | { | |
321 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
322 | { | |
323 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
324 | if (!flow_bb_inside_loop_p (loop, bb)) | |
325 | { | |
73fbfcad | 326 | if (dump_enabled_p ()) |
78c60e3d | 327 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 328 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 329 | |
3157b0c2 AO |
330 | if (is_gimple_debug (USE_STMT (use_p))) |
331 | continue; | |
332 | ||
ebfd146a IR |
333 | /* We expect all such uses to be in the loop exit phis |
334 | (because of loop closed form) */ | |
335 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
336 | gcc_assert (bb == single_exit (loop)->dest); | |
337 | ||
338 | *live_p = true; | |
339 | } | |
340 | } | |
341 | } | |
342 | ||
343 | return (*live_p || *relevant); | |
344 | } | |
345 | ||
346 | ||
b8698a0f | 347 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 348 | |
ff802fa1 | 349 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
350 | used in STMT for anything other than indexing an array. */ |
351 | ||
352 | static bool | |
353 | exist_non_indexing_operands_for_use_p (tree use, gimple stmt) | |
354 | { | |
355 | tree operand; | |
356 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 357 | |
ff802fa1 | 358 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
359 | reference in STMT, then any operand that corresponds to USE |
360 | is not indexing an array. */ | |
361 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
362 | return true; | |
59a05b0c | 363 | |
ebfd146a IR |
364 | /* STMT has a data_ref. FORNOW this means that its of one of |
365 | the following forms: | |
366 | -1- ARRAY_REF = var | |
367 | -2- var = ARRAY_REF | |
368 | (This should have been verified in analyze_data_refs). | |
369 | ||
370 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 371 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
372 | for array indexing. |
373 | ||
374 | Therefore, all we need to check is if STMT falls into the | |
375 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
376 | |
377 | if (!gimple_assign_copy_p (stmt)) | |
378 | return false; | |
59a05b0c EB |
379 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
380 | return false; | |
ebfd146a | 381 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
382 | if (TREE_CODE (operand) != SSA_NAME) |
383 | return false; | |
384 | ||
385 | if (operand == use) | |
386 | return true; | |
387 | ||
388 | return false; | |
389 | } | |
390 | ||
391 | ||
b8698a0f | 392 | /* |
ebfd146a IR |
393 | Function process_use. |
394 | ||
395 | Inputs: | |
396 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 397 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ff802fa1 | 398 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 399 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
400 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
401 | be performed. | |
ebfd146a IR |
402 | |
403 | Outputs: | |
404 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
405 | relevance info of the DEF_STMT of this USE: | |
406 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
407 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
408 | Exceptions: | |
409 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 410 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 411 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
412 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
413 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
414 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
415 | be modified accordingly. | |
416 | ||
417 | Return true if everything is as expected. Return false otherwise. */ | |
418 | ||
419 | static bool | |
b8698a0f | 420 | process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
9771b263 | 421 | enum vect_relevant relevant, vec<gimple> *worklist, |
aec7ae7d | 422 | bool force) |
ebfd146a IR |
423 | { |
424 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
425 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
426 | stmt_vec_info dstmt_vinfo; | |
427 | basic_block bb, def_bb; | |
428 | tree def; | |
429 | gimple def_stmt; | |
430 | enum vect_def_type dt; | |
431 | ||
b8698a0f | 432 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 433 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 434 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
435 | return true; |
436 | ||
24ee1384 | 437 | if (!vect_is_simple_use (use, stmt, loop_vinfo, NULL, &def_stmt, &def, &dt)) |
b8698a0f | 438 | { |
73fbfcad | 439 | if (dump_enabled_p ()) |
78c60e3d | 440 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 441 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
442 | return false; |
443 | } | |
444 | ||
445 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
446 | return true; | |
447 | ||
448 | def_bb = gimple_bb (def_stmt); | |
449 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
450 | { | |
73fbfcad | 451 | if (dump_enabled_p ()) |
e645e942 | 452 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
453 | return true; |
454 | } | |
455 | ||
b8698a0f L |
456 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
457 | DEF_STMT must have already been processed, because this should be the | |
458 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
459 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
460 | check that everything is as expected, and we are done. */ |
461 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
462 | bb = gimple_bb (stmt); | |
463 | if (gimple_code (stmt) == GIMPLE_PHI | |
464 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
465 | && gimple_code (def_stmt) != GIMPLE_PHI | |
466 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
467 | && bb->loop_father == def_bb->loop_father) | |
468 | { | |
73fbfcad | 469 | if (dump_enabled_p ()) |
78c60e3d | 470 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 471 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
472 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
473 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
474 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 475 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 476 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
477 | return true; |
478 | } | |
479 | ||
480 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
481 | outer-loop-header-bb: | |
482 | d = def_stmt | |
483 | inner-loop: | |
484 | stmt # use (d) | |
485 | outer-loop-tail-bb: | |
486 | ... */ | |
487 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
488 | { | |
73fbfcad | 489 | if (dump_enabled_p ()) |
78c60e3d | 490 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 491 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 492 | |
ebfd146a IR |
493 | switch (relevant) |
494 | { | |
8644a673 | 495 | case vect_unused_in_scope: |
7c5222ff IR |
496 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
497 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 498 | break; |
7c5222ff | 499 | |
ebfd146a | 500 | case vect_used_in_outer_by_reduction: |
7c5222ff | 501 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
502 | relevant = vect_used_by_reduction; |
503 | break; | |
7c5222ff | 504 | |
ebfd146a | 505 | case vect_used_in_outer: |
7c5222ff | 506 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 507 | relevant = vect_used_in_scope; |
ebfd146a | 508 | break; |
7c5222ff | 509 | |
8644a673 | 510 | case vect_used_in_scope: |
ebfd146a IR |
511 | break; |
512 | ||
513 | default: | |
514 | gcc_unreachable (); | |
b8698a0f | 515 | } |
ebfd146a IR |
516 | } |
517 | ||
518 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
519 | outer-loop-header-bb: | |
520 | ... | |
521 | inner-loop: | |
522 | d = def_stmt | |
06066f92 | 523 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
524 | stmt # use (d) */ |
525 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
526 | { | |
73fbfcad | 527 | if (dump_enabled_p ()) |
78c60e3d | 528 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 529 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 530 | |
ebfd146a IR |
531 | switch (relevant) |
532 | { | |
8644a673 | 533 | case vect_unused_in_scope: |
b8698a0f | 534 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 535 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 536 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
537 | break; |
538 | ||
ebfd146a IR |
539 | case vect_used_by_reduction: |
540 | relevant = vect_used_in_outer_by_reduction; | |
541 | break; | |
542 | ||
8644a673 | 543 | case vect_used_in_scope: |
ebfd146a IR |
544 | relevant = vect_used_in_outer; |
545 | break; | |
546 | ||
547 | default: | |
548 | gcc_unreachable (); | |
549 | } | |
550 | } | |
551 | ||
83197f37 IR |
552 | vect_mark_relevant (worklist, def_stmt, relevant, live_p, |
553 | is_pattern_stmt_p (stmt_vinfo)); | |
ebfd146a IR |
554 | return true; |
555 | } | |
556 | ||
557 | ||
558 | /* Function vect_mark_stmts_to_be_vectorized. | |
559 | ||
560 | Not all stmts in the loop need to be vectorized. For example: | |
561 | ||
562 | for i... | |
563 | for j... | |
564 | 1. T0 = i + j | |
565 | 2. T1 = a[T0] | |
566 | ||
567 | 3. j = j + 1 | |
568 | ||
569 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
570 | addressing of vectorized data-refs are handled differently. | |
571 | ||
572 | This pass detects such stmts. */ | |
573 | ||
574 | bool | |
575 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
576 | { | |
9771b263 | 577 | vec<gimple> worklist; |
ebfd146a IR |
578 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
579 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
580 | unsigned int nbbs = loop->num_nodes; | |
581 | gimple_stmt_iterator si; | |
582 | gimple stmt; | |
583 | unsigned int i; | |
584 | stmt_vec_info stmt_vinfo; | |
585 | basic_block bb; | |
586 | gimple phi; | |
587 | bool live_p; | |
06066f92 IR |
588 | enum vect_relevant relevant, tmp_relevant; |
589 | enum vect_def_type def_type; | |
ebfd146a | 590 | |
73fbfcad | 591 | if (dump_enabled_p ()) |
78c60e3d | 592 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 593 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 594 | |
9771b263 | 595 | worklist.create (64); |
ebfd146a IR |
596 | |
597 | /* 1. Init worklist. */ | |
598 | for (i = 0; i < nbbs; i++) | |
599 | { | |
600 | bb = bbs[i]; | |
601 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 602 | { |
ebfd146a | 603 | phi = gsi_stmt (si); |
73fbfcad | 604 | if (dump_enabled_p ()) |
ebfd146a | 605 | { |
78c60e3d SS |
606 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
607 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
e645e942 | 608 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
609 | } |
610 | ||
611 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 612 | vect_mark_relevant (&worklist, phi, relevant, live_p, false); |
ebfd146a IR |
613 | } |
614 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
615 | { | |
616 | stmt = gsi_stmt (si); | |
73fbfcad | 617 | if (dump_enabled_p ()) |
ebfd146a | 618 | { |
78c60e3d SS |
619 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
620 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 621 | dump_printf (MSG_NOTE, "\n"); |
b8698a0f | 622 | } |
ebfd146a IR |
623 | |
624 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 625 | vect_mark_relevant (&worklist, stmt, relevant, live_p, false); |
ebfd146a IR |
626 | } |
627 | } | |
628 | ||
629 | /* 2. Process_worklist */ | |
9771b263 | 630 | while (worklist.length () > 0) |
ebfd146a IR |
631 | { |
632 | use_operand_p use_p; | |
633 | ssa_op_iter iter; | |
634 | ||
9771b263 | 635 | stmt = worklist.pop (); |
73fbfcad | 636 | if (dump_enabled_p ()) |
ebfd146a | 637 | { |
78c60e3d SS |
638 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
639 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 640 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
641 | } |
642 | ||
b8698a0f L |
643 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
644 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
645 | liveness and relevance properties of STMT. */ |
646 | stmt_vinfo = vinfo_for_stmt (stmt); | |
647 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
648 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
649 | ||
650 | /* Generally, the liveness and relevance properties of STMT are | |
651 | propagated as is to the DEF_STMTs of its USEs: | |
652 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
653 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
654 | ||
655 | One exception is when STMT has been identified as defining a reduction | |
656 | variable; in this case we set the liveness/relevance as follows: | |
657 | live_p = false | |
658 | relevant = vect_used_by_reduction | |
659 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f | 660 | those that are used by a reduction computation, and those that are |
ff802fa1 | 661 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 662 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 663 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 664 | |
06066f92 IR |
665 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
666 | tmp_relevant = relevant; | |
667 | switch (def_type) | |
ebfd146a | 668 | { |
06066f92 IR |
669 | case vect_reduction_def: |
670 | switch (tmp_relevant) | |
671 | { | |
672 | case vect_unused_in_scope: | |
673 | relevant = vect_used_by_reduction; | |
674 | break; | |
675 | ||
676 | case vect_used_by_reduction: | |
677 | if (gimple_code (stmt) == GIMPLE_PHI) | |
678 | break; | |
679 | /* fall through */ | |
680 | ||
681 | default: | |
73fbfcad | 682 | if (dump_enabled_p ()) |
78c60e3d | 683 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 684 | "unsupported use of reduction.\n"); |
9771b263 | 685 | worklist.release (); |
06066f92 IR |
686 | return false; |
687 | } | |
688 | ||
b8698a0f | 689 | live_p = false; |
06066f92 | 690 | break; |
b8698a0f | 691 | |
06066f92 IR |
692 | case vect_nested_cycle: |
693 | if (tmp_relevant != vect_unused_in_scope | |
694 | && tmp_relevant != vect_used_in_outer_by_reduction | |
695 | && tmp_relevant != vect_used_in_outer) | |
696 | { | |
73fbfcad | 697 | if (dump_enabled_p ()) |
78c60e3d | 698 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 699 | "unsupported use of nested cycle.\n"); |
7c5222ff | 700 | |
9771b263 | 701 | worklist.release (); |
06066f92 IR |
702 | return false; |
703 | } | |
7c5222ff | 704 | |
b8698a0f L |
705 | live_p = false; |
706 | break; | |
707 | ||
06066f92 IR |
708 | case vect_double_reduction_def: |
709 | if (tmp_relevant != vect_unused_in_scope | |
710 | && tmp_relevant != vect_used_by_reduction) | |
711 | { | |
73fbfcad | 712 | if (dump_enabled_p ()) |
78c60e3d | 713 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 714 | "unsupported use of double reduction.\n"); |
7c5222ff | 715 | |
9771b263 | 716 | worklist.release (); |
7c5222ff | 717 | return false; |
06066f92 IR |
718 | } |
719 | ||
720 | live_p = false; | |
b8698a0f | 721 | break; |
7c5222ff | 722 | |
06066f92 IR |
723 | default: |
724 | break; | |
7c5222ff | 725 | } |
b8698a0f | 726 | |
aec7ae7d | 727 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
728 | { |
729 | /* Pattern statements are not inserted into the code, so | |
730 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
731 | have to scan the RHS or function arguments instead. */ | |
732 | if (is_gimple_assign (stmt)) | |
733 | { | |
69d2aade JJ |
734 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
735 | tree op = gimple_assign_rhs1 (stmt); | |
736 | ||
737 | i = 1; | |
738 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
739 | { | |
740 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
aec7ae7d | 741 | live_p, relevant, &worklist, false) |
69d2aade | 742 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
aec7ae7d | 743 | live_p, relevant, &worklist, false)) |
69d2aade | 744 | { |
9771b263 | 745 | worklist.release (); |
69d2aade JJ |
746 | return false; |
747 | } | |
748 | i = 2; | |
749 | } | |
750 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 751 | { |
69d2aade | 752 | op = gimple_op (stmt, i); |
9d5e7640 | 753 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, |
aec7ae7d | 754 | &worklist, false)) |
9d5e7640 | 755 | { |
9771b263 | 756 | worklist.release (); |
9d5e7640 IR |
757 | return false; |
758 | } | |
759 | } | |
760 | } | |
761 | else if (is_gimple_call (stmt)) | |
762 | { | |
763 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
764 | { | |
765 | tree arg = gimple_call_arg (stmt, i); | |
766 | if (!process_use (stmt, arg, loop_vinfo, live_p, relevant, | |
aec7ae7d | 767 | &worklist, false)) |
9d5e7640 | 768 | { |
9771b263 | 769 | worklist.release (); |
9d5e7640 IR |
770 | return false; |
771 | } | |
772 | } | |
773 | } | |
774 | } | |
775 | else | |
776 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
777 | { | |
778 | tree op = USE_FROM_PTR (use_p); | |
779 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, | |
aec7ae7d | 780 | &worklist, false)) |
9d5e7640 | 781 | { |
9771b263 | 782 | worklist.release (); |
9d5e7640 IR |
783 | return false; |
784 | } | |
785 | } | |
aec7ae7d JJ |
786 | |
787 | if (STMT_VINFO_GATHER_P (stmt_vinfo)) | |
788 | { | |
789 | tree off; | |
790 | tree decl = vect_check_gather (stmt, loop_vinfo, NULL, &off, NULL); | |
791 | gcc_assert (decl); | |
792 | if (!process_use (stmt, off, loop_vinfo, live_p, relevant, | |
793 | &worklist, true)) | |
794 | { | |
9771b263 | 795 | worklist.release (); |
aec7ae7d JJ |
796 | return false; |
797 | } | |
798 | } | |
ebfd146a IR |
799 | } /* while worklist */ |
800 | ||
9771b263 | 801 | worklist.release (); |
ebfd146a IR |
802 | return true; |
803 | } | |
804 | ||
805 | ||
b8698a0f | 806 | /* Function vect_model_simple_cost. |
ebfd146a | 807 | |
b8698a0f | 808 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
809 | single op. Right now, this does not account for multiple insns that could |
810 | be generated for the single vector op. We will handle that shortly. */ | |
811 | ||
812 | void | |
b8698a0f | 813 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 BS |
814 | enum vect_def_type *dt, |
815 | stmt_vector_for_cost *prologue_cost_vec, | |
816 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
817 | { |
818 | int i; | |
92345349 | 819 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
820 | |
821 | /* The SLP costs were already calculated during SLP tree build. */ | |
822 | if (PURE_SLP_STMT (stmt_info)) | |
823 | return; | |
824 | ||
ebfd146a IR |
825 | /* FORNOW: Assuming maximum 2 args per stmts. */ |
826 | for (i = 0; i < 2; i++) | |
92345349 BS |
827 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
828 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, vector_stmt, | |
829 | stmt_info, 0, vect_prologue); | |
c3e7ee41 BS |
830 | |
831 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
832 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
833 | stmt_info, 0, vect_body); | |
c3e7ee41 | 834 | |
73fbfcad | 835 | if (dump_enabled_p ()) |
78c60e3d SS |
836 | dump_printf_loc (MSG_NOTE, vect_location, |
837 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 838 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
839 | } |
840 | ||
841 | ||
8bd37302 BS |
842 | /* Model cost for type demotion and promotion operations. PWR is normally |
843 | zero for single-step promotions and demotions. It will be one if | |
844 | two-step promotion/demotion is required, and so on. Each additional | |
845 | step doubles the number of instructions required. */ | |
846 | ||
847 | static void | |
848 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
849 | enum vect_def_type *dt, int pwr) | |
850 | { | |
851 | int i, tmp; | |
92345349 | 852 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
853 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
854 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
855 | void *target_cost_data; | |
8bd37302 BS |
856 | |
857 | /* The SLP costs were already calculated during SLP tree build. */ | |
858 | if (PURE_SLP_STMT (stmt_info)) | |
859 | return; | |
860 | ||
c3e7ee41 BS |
861 | if (loop_vinfo) |
862 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
863 | else | |
864 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
865 | ||
8bd37302 BS |
866 | for (i = 0; i < pwr + 1; i++) |
867 | { | |
868 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
869 | (i + 1) : i; | |
c3e7ee41 | 870 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
871 | vec_promote_demote, stmt_info, 0, |
872 | vect_body); | |
8bd37302 BS |
873 | } |
874 | ||
875 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
876 | for (i = 0; i < 2; i++) | |
92345349 BS |
877 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
878 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
879 | stmt_info, 0, vect_prologue); | |
8bd37302 | 880 | |
73fbfcad | 881 | if (dump_enabled_p ()) |
78c60e3d SS |
882 | dump_printf_loc (MSG_NOTE, vect_location, |
883 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 884 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
885 | } |
886 | ||
0d0293ac | 887 | /* Function vect_cost_group_size |
b8698a0f | 888 | |
0d0293ac | 889 | For grouped load or store, return the group_size only if it is the first |
ebfd146a IR |
890 | load or store of a group, else return 1. This ensures that group size is |
891 | only returned once per group. */ | |
892 | ||
893 | static int | |
0d0293ac | 894 | vect_cost_group_size (stmt_vec_info stmt_info) |
ebfd146a | 895 | { |
e14c1050 | 896 | gimple first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
897 | |
898 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
e14c1050 | 899 | return GROUP_SIZE (stmt_info); |
ebfd146a IR |
900 | |
901 | return 1; | |
902 | } | |
903 | ||
904 | ||
905 | /* Function vect_model_store_cost | |
906 | ||
0d0293ac MM |
907 | Models cost for stores. In the case of grouped accesses, one access |
908 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
909 | |
910 | void | |
b8698a0f | 911 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
272c6793 | 912 | bool store_lanes_p, enum vect_def_type dt, |
92345349 BS |
913 | slp_tree slp_node, |
914 | stmt_vector_for_cost *prologue_cost_vec, | |
915 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
916 | { |
917 | int group_size; | |
92345349 | 918 | unsigned int inside_cost = 0, prologue_cost = 0; |
720f5239 IR |
919 | struct data_reference *first_dr; |
920 | gimple first_stmt; | |
ebfd146a IR |
921 | |
922 | /* The SLP costs were already calculated during SLP tree build. */ | |
923 | if (PURE_SLP_STMT (stmt_info)) | |
924 | return; | |
925 | ||
8644a673 | 926 | if (dt == vect_constant_def || dt == vect_external_def) |
92345349 BS |
927 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
928 | stmt_info, 0, vect_prologue); | |
ebfd146a | 929 | |
0d0293ac MM |
930 | /* Grouped access? */ |
931 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
720f5239 IR |
932 | { |
933 | if (slp_node) | |
934 | { | |
9771b263 | 935 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
720f5239 IR |
936 | group_size = 1; |
937 | } | |
938 | else | |
939 | { | |
e14c1050 | 940 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 941 | group_size = vect_cost_group_size (stmt_info); |
720f5239 IR |
942 | } |
943 | ||
944 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
945 | } | |
0d0293ac | 946 | /* Not a grouped access. */ |
ebfd146a | 947 | else |
720f5239 IR |
948 | { |
949 | group_size = 1; | |
950 | first_dr = STMT_VINFO_DATA_REF (stmt_info); | |
951 | } | |
ebfd146a | 952 | |
272c6793 | 953 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 954 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 RS |
955 | access is instead being provided by a permute-and-store operation, |
956 | include the cost of the permutes. */ | |
957 | if (!store_lanes_p && group_size > 1) | |
ebfd146a IR |
958 | { |
959 | /* Uses a high and low interleave operation for each needed permute. */ | |
c3e7ee41 BS |
960 | |
961 | int nstmts = ncopies * exact_log2 (group_size) * group_size; | |
92345349 BS |
962 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
963 | stmt_info, 0, vect_body); | |
ebfd146a | 964 | |
73fbfcad | 965 | if (dump_enabled_p ()) |
78c60e3d | 966 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 967 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 968 | group_size); |
ebfd146a IR |
969 | } |
970 | ||
971 | /* Costs of the stores. */ | |
92345349 | 972 | vect_get_store_cost (first_dr, ncopies, &inside_cost, body_cost_vec); |
ebfd146a | 973 | |
73fbfcad | 974 | if (dump_enabled_p ()) |
78c60e3d SS |
975 | dump_printf_loc (MSG_NOTE, vect_location, |
976 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 977 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
978 | } |
979 | ||
980 | ||
720f5239 IR |
981 | /* Calculate cost of DR's memory access. */ |
982 | void | |
983 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 984 | unsigned int *inside_cost, |
92345349 | 985 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
986 | { |
987 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
c3e7ee41 BS |
988 | gimple stmt = DR_STMT (dr); |
989 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
720f5239 IR |
990 | |
991 | switch (alignment_support_scheme) | |
992 | { | |
993 | case dr_aligned: | |
994 | { | |
92345349 BS |
995 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
996 | vector_store, stmt_info, 0, | |
997 | vect_body); | |
720f5239 | 998 | |
73fbfcad | 999 | if (dump_enabled_p ()) |
78c60e3d | 1000 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1001 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
1002 | break; |
1003 | } | |
1004 | ||
1005 | case dr_unaligned_supported: | |
1006 | { | |
720f5239 | 1007 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1008 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1009 | unaligned_store, stmt_info, |
92345349 | 1010 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1011 | if (dump_enabled_p ()) |
78c60e3d SS |
1012 | dump_printf_loc (MSG_NOTE, vect_location, |
1013 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1014 | "hardware.\n"); |
720f5239 IR |
1015 | break; |
1016 | } | |
1017 | ||
38eec4c6 UW |
1018 | case dr_unaligned_unsupported: |
1019 | { | |
1020 | *inside_cost = VECT_MAX_COST; | |
1021 | ||
73fbfcad | 1022 | if (dump_enabled_p ()) |
78c60e3d | 1023 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1024 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1025 | break; |
1026 | } | |
1027 | ||
720f5239 IR |
1028 | default: |
1029 | gcc_unreachable (); | |
1030 | } | |
1031 | } | |
1032 | ||
1033 | ||
ebfd146a IR |
1034 | /* Function vect_model_load_cost |
1035 | ||
0d0293ac MM |
1036 | Models cost for loads. In the case of grouped accesses, the last access |
1037 | has the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1038 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1039 | access scheme chosen. */ |
1040 | ||
1041 | void | |
92345349 BS |
1042 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
1043 | bool load_lanes_p, slp_tree slp_node, | |
1044 | stmt_vector_for_cost *prologue_cost_vec, | |
1045 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
1046 | { |
1047 | int group_size; | |
ebfd146a IR |
1048 | gimple first_stmt; |
1049 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; | |
92345349 | 1050 | unsigned int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
1051 | |
1052 | /* The SLP costs were already calculated during SLP tree build. */ | |
1053 | if (PURE_SLP_STMT (stmt_info)) | |
1054 | return; | |
1055 | ||
0d0293ac | 1056 | /* Grouped accesses? */ |
e14c1050 | 1057 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 1058 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && first_stmt && !slp_node) |
ebfd146a | 1059 | { |
0d0293ac | 1060 | group_size = vect_cost_group_size (stmt_info); |
ebfd146a IR |
1061 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
1062 | } | |
0d0293ac | 1063 | /* Not a grouped access. */ |
ebfd146a IR |
1064 | else |
1065 | { | |
1066 | group_size = 1; | |
1067 | first_dr = dr; | |
1068 | } | |
1069 | ||
272c6793 | 1070 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1071 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 RS |
1072 | access is instead being provided by a load-and-permute operation, |
1073 | include the cost of the permutes. */ | |
1074 | if (!load_lanes_p && group_size > 1) | |
ebfd146a IR |
1075 | { |
1076 | /* Uses an even and odd extract operations for each needed permute. */ | |
c3e7ee41 | 1077 | int nstmts = ncopies * exact_log2 (group_size) * group_size; |
92345349 BS |
1078 | inside_cost += record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
1079 | stmt_info, 0, vect_body); | |
ebfd146a | 1080 | |
73fbfcad | 1081 | if (dump_enabled_p ()) |
e645e942 TJ |
1082 | dump_printf_loc (MSG_NOTE, vect_location, |
1083 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1084 | group_size); |
ebfd146a IR |
1085 | } |
1086 | ||
1087 | /* The loads themselves. */ | |
a82960aa RG |
1088 | if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
1089 | { | |
a21892ad BS |
1090 | /* N scalar loads plus gathering them into a vector. */ |
1091 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
92345349 | 1092 | inside_cost += record_stmt_cost (body_cost_vec, |
c3e7ee41 | 1093 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), |
92345349 BS |
1094 | scalar_load, stmt_info, 0, vect_body); |
1095 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, | |
1096 | stmt_info, 0, vect_body); | |
a82960aa RG |
1097 | } |
1098 | else | |
1099 | vect_get_load_cost (first_dr, ncopies, | |
1100 | ((!STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1101 | || group_size > 1 || slp_node), | |
92345349 BS |
1102 | &inside_cost, &prologue_cost, |
1103 | prologue_cost_vec, body_cost_vec, true); | |
720f5239 | 1104 | |
73fbfcad | 1105 | if (dump_enabled_p ()) |
78c60e3d SS |
1106 | dump_printf_loc (MSG_NOTE, vect_location, |
1107 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1108 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1109 | } |
1110 | ||
1111 | ||
1112 | /* Calculate cost of DR's memory access. */ | |
1113 | void | |
1114 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1115 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1116 | unsigned int *prologue_cost, |
1117 | stmt_vector_for_cost *prologue_cost_vec, | |
1118 | stmt_vector_for_cost *body_cost_vec, | |
1119 | bool record_prologue_costs) | |
720f5239 IR |
1120 | { |
1121 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
c3e7ee41 BS |
1122 | gimple stmt = DR_STMT (dr); |
1123 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
720f5239 IR |
1124 | |
1125 | switch (alignment_support_scheme) | |
ebfd146a IR |
1126 | { |
1127 | case dr_aligned: | |
1128 | { | |
92345349 BS |
1129 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1130 | stmt_info, 0, vect_body); | |
ebfd146a | 1131 | |
73fbfcad | 1132 | if (dump_enabled_p ()) |
78c60e3d | 1133 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1134 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1135 | |
1136 | break; | |
1137 | } | |
1138 | case dr_unaligned_supported: | |
1139 | { | |
720f5239 | 1140 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1141 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1142 | unaligned_load, stmt_info, |
92345349 | 1143 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1144 | |
73fbfcad | 1145 | if (dump_enabled_p ()) |
78c60e3d SS |
1146 | dump_printf_loc (MSG_NOTE, vect_location, |
1147 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1148 | "hardware.\n"); |
ebfd146a IR |
1149 | |
1150 | break; | |
1151 | } | |
1152 | case dr_explicit_realign: | |
1153 | { | |
92345349 BS |
1154 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1155 | vector_load, stmt_info, 0, vect_body); | |
1156 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1157 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1158 | |
1159 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1160 | the containing loop, the following cost should be added to the | |
92345349 | 1161 | prologue costs. */ |
ebfd146a | 1162 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1163 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1164 | stmt_info, 0, vect_body); | |
ebfd146a | 1165 | |
73fbfcad | 1166 | if (dump_enabled_p ()) |
e645e942 TJ |
1167 | dump_printf_loc (MSG_NOTE, vect_location, |
1168 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1169 | |
ebfd146a IR |
1170 | break; |
1171 | } | |
1172 | case dr_explicit_realign_optimized: | |
1173 | { | |
73fbfcad | 1174 | if (dump_enabled_p ()) |
e645e942 | 1175 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1176 | "vect_model_load_cost: unaligned software " |
e645e942 | 1177 | "pipelined.\n"); |
ebfd146a IR |
1178 | |
1179 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1180 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1181 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1182 | access, then the above cost should only be considered for one |
ff802fa1 | 1183 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1184 | and a realignment op. */ |
1185 | ||
92345349 | 1186 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1187 | { |
92345349 BS |
1188 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1189 | vector_stmt, stmt_info, | |
1190 | 0, vect_prologue); | |
ebfd146a | 1191 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1192 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1193 | vector_stmt, stmt_info, | |
1194 | 0, vect_prologue); | |
ebfd146a IR |
1195 | } |
1196 | ||
92345349 BS |
1197 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1198 | stmt_info, 0, vect_body); | |
1199 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1200 | stmt_info, 0, vect_body); | |
8bd37302 | 1201 | |
73fbfcad | 1202 | if (dump_enabled_p ()) |
78c60e3d | 1203 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1204 | "vect_model_load_cost: explicit realign optimized" |
1205 | "\n"); | |
8bd37302 | 1206 | |
ebfd146a IR |
1207 | break; |
1208 | } | |
1209 | ||
38eec4c6 UW |
1210 | case dr_unaligned_unsupported: |
1211 | { | |
1212 | *inside_cost = VECT_MAX_COST; | |
1213 | ||
73fbfcad | 1214 | if (dump_enabled_p ()) |
78c60e3d | 1215 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1216 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1217 | break; |
1218 | } | |
1219 | ||
ebfd146a IR |
1220 | default: |
1221 | gcc_unreachable (); | |
1222 | } | |
ebfd146a IR |
1223 | } |
1224 | ||
418b7df3 RG |
1225 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1226 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1227 | |
418b7df3 RG |
1228 | static void |
1229 | vect_init_vector_1 (gimple stmt, gimple new_stmt, gimple_stmt_iterator *gsi) | |
ebfd146a | 1230 | { |
ebfd146a | 1231 | if (gsi) |
418b7df3 | 1232 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1233 | else |
1234 | { | |
418b7df3 | 1235 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1236 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1237 | |
a70d6342 IR |
1238 | if (loop_vinfo) |
1239 | { | |
1240 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1241 | basic_block new_bb; |
1242 | edge pe; | |
a70d6342 IR |
1243 | |
1244 | if (nested_in_vect_loop_p (loop, stmt)) | |
1245 | loop = loop->inner; | |
b8698a0f | 1246 | |
a70d6342 | 1247 | pe = loop_preheader_edge (loop); |
418b7df3 | 1248 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1249 | gcc_assert (!new_bb); |
1250 | } | |
1251 | else | |
1252 | { | |
1253 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1254 | basic_block bb; | |
1255 | gimple_stmt_iterator gsi_bb_start; | |
1256 | ||
1257 | gcc_assert (bb_vinfo); | |
1258 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1259 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1260 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1261 | } |
ebfd146a IR |
1262 | } |
1263 | ||
73fbfcad | 1264 | if (dump_enabled_p ()) |
ebfd146a | 1265 | { |
78c60e3d SS |
1266 | dump_printf_loc (MSG_NOTE, vect_location, |
1267 | "created new init_stmt: "); | |
1268 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
e645e942 | 1269 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 1270 | } |
418b7df3 RG |
1271 | } |
1272 | ||
1273 | /* Function vect_init_vector. | |
ebfd146a | 1274 | |
5467ee52 RG |
1275 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1276 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1277 | vector type a vector with all elements equal to VAL is created first. | |
1278 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1279 | initialization at the loop preheader. | |
418b7df3 RG |
1280 | Return the DEF of INIT_STMT. |
1281 | It will be used in the vectorization of STMT. */ | |
1282 | ||
1283 | tree | |
5467ee52 | 1284 | vect_init_vector (gimple stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 RG |
1285 | { |
1286 | tree new_var; | |
1287 | gimple init_stmt; | |
1288 | tree vec_oprnd; | |
1289 | tree new_temp; | |
1290 | ||
5467ee52 RG |
1291 | if (TREE_CODE (type) == VECTOR_TYPE |
1292 | && TREE_CODE (TREE_TYPE (val)) != VECTOR_TYPE) | |
418b7df3 | 1293 | { |
5467ee52 | 1294 | if (!types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) |
418b7df3 | 1295 | { |
5467ee52 RG |
1296 | if (CONSTANT_CLASS_P (val)) |
1297 | val = fold_unary (VIEW_CONVERT_EXPR, TREE_TYPE (type), val); | |
418b7df3 RG |
1298 | else |
1299 | { | |
83d5977e | 1300 | new_temp = make_ssa_name (TREE_TYPE (type), NULL); |
418b7df3 | 1301 | init_stmt = gimple_build_assign_with_ops (NOP_EXPR, |
83d5977e | 1302 | new_temp, val, |
418b7df3 | 1303 | NULL_TREE); |
418b7df3 | 1304 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1305 | val = new_temp; |
418b7df3 RG |
1306 | } |
1307 | } | |
5467ee52 | 1308 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1309 | } |
1310 | ||
5467ee52 | 1311 | new_var = vect_get_new_vect_var (type, vect_simple_var, "cst_"); |
5467ee52 | 1312 | init_stmt = gimple_build_assign (new_var, val); |
418b7df3 RG |
1313 | new_temp = make_ssa_name (new_var, init_stmt); |
1314 | gimple_assign_set_lhs (init_stmt, new_temp); | |
1315 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
ebfd146a IR |
1316 | vec_oprnd = gimple_assign_lhs (init_stmt); |
1317 | return vec_oprnd; | |
1318 | } | |
1319 | ||
a70d6342 | 1320 | |
ebfd146a IR |
1321 | /* Function vect_get_vec_def_for_operand. |
1322 | ||
ff802fa1 | 1323 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
1324 | used in the vectorized stmt for STMT. |
1325 | ||
1326 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1327 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1328 | ||
1329 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1330 | needs to be introduced. */ | |
1331 | ||
1332 | tree | |
1333 | vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def) | |
1334 | { | |
1335 | tree vec_oprnd; | |
1336 | gimple vec_stmt; | |
1337 | gimple def_stmt; | |
1338 | stmt_vec_info def_stmt_info = NULL; | |
1339 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
9dc3f7de | 1340 | unsigned int nunits; |
ebfd146a | 1341 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
ebfd146a | 1342 | tree def; |
ebfd146a IR |
1343 | enum vect_def_type dt; |
1344 | bool is_simple_use; | |
1345 | tree vector_type; | |
1346 | ||
73fbfcad | 1347 | if (dump_enabled_p ()) |
ebfd146a | 1348 | { |
78c60e3d SS |
1349 | dump_printf_loc (MSG_NOTE, vect_location, |
1350 | "vect_get_vec_def_for_operand: "); | |
1351 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
e645e942 | 1352 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1353 | } |
1354 | ||
24ee1384 IR |
1355 | is_simple_use = vect_is_simple_use (op, stmt, loop_vinfo, NULL, |
1356 | &def_stmt, &def, &dt); | |
ebfd146a | 1357 | gcc_assert (is_simple_use); |
73fbfcad | 1358 | if (dump_enabled_p ()) |
ebfd146a | 1359 | { |
78c60e3d | 1360 | int loc_printed = 0; |
ebfd146a IR |
1361 | if (def) |
1362 | { | |
78c60e3d SS |
1363 | dump_printf_loc (MSG_NOTE, vect_location, "def = "); |
1364 | loc_printed = 1; | |
1365 | dump_generic_expr (MSG_NOTE, TDF_SLIM, def); | |
e645e942 | 1366 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1367 | } |
1368 | if (def_stmt) | |
1369 | { | |
78c60e3d SS |
1370 | if (loc_printed) |
1371 | dump_printf (MSG_NOTE, " def_stmt = "); | |
1372 | else | |
1373 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1374 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
e645e942 | 1375 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1376 | } |
1377 | } | |
1378 | ||
1379 | switch (dt) | |
1380 | { | |
1381 | /* Case 1: operand is a constant. */ | |
1382 | case vect_constant_def: | |
1383 | { | |
7569a6cc RG |
1384 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
1385 | gcc_assert (vector_type); | |
9dc3f7de | 1386 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); |
7569a6cc | 1387 | |
b8698a0f | 1388 | if (scalar_def) |
ebfd146a IR |
1389 | *scalar_def = op; |
1390 | ||
1391 | /* Create 'vect_cst_ = {cst,cst,...,cst}' */ | |
73fbfcad | 1392 | if (dump_enabled_p ()) |
78c60e3d | 1393 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1394 | "Create vector_cst. nunits = %d\n", nunits); |
ebfd146a | 1395 | |
418b7df3 | 1396 | return vect_init_vector (stmt, op, vector_type, NULL); |
ebfd146a IR |
1397 | } |
1398 | ||
1399 | /* Case 2: operand is defined outside the loop - loop invariant. */ | |
8644a673 | 1400 | case vect_external_def: |
ebfd146a IR |
1401 | { |
1402 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); | |
1403 | gcc_assert (vector_type); | |
ebfd146a | 1404 | |
b8698a0f | 1405 | if (scalar_def) |
ebfd146a IR |
1406 | *scalar_def = def; |
1407 | ||
1408 | /* Create 'vec_inv = {inv,inv,..,inv}' */ | |
73fbfcad | 1409 | if (dump_enabled_p ()) |
e645e942 | 1410 | dump_printf_loc (MSG_NOTE, vect_location, "Create vector_inv.\n"); |
ebfd146a | 1411 | |
418b7df3 | 1412 | return vect_init_vector (stmt, def, vector_type, NULL); |
ebfd146a IR |
1413 | } |
1414 | ||
1415 | /* Case 3: operand is defined inside the loop. */ | |
8644a673 | 1416 | case vect_internal_def: |
ebfd146a | 1417 | { |
b8698a0f | 1418 | if (scalar_def) |
ebfd146a IR |
1419 | *scalar_def = NULL/* FIXME tuples: def_stmt*/; |
1420 | ||
1421 | /* Get the def from the vectorized stmt. */ | |
1422 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1423 | |
ebfd146a | 1424 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1425 | /* Get vectorized pattern statement. */ |
1426 | if (!vec_stmt | |
1427 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1428 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1429 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1430 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1431 | gcc_assert (vec_stmt); |
1432 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1433 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1434 | else if (is_gimple_call (vec_stmt)) | |
1435 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1436 | else | |
1437 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1438 | return vec_oprnd; | |
1439 | } | |
1440 | ||
1441 | /* Case 4: operand is defined by a loop header phi - reduction */ | |
1442 | case vect_reduction_def: | |
06066f92 | 1443 | case vect_double_reduction_def: |
7c5222ff | 1444 | case vect_nested_cycle: |
ebfd146a IR |
1445 | { |
1446 | struct loop *loop; | |
1447 | ||
1448 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
b8698a0f | 1449 | loop = (gimple_bb (def_stmt))->loop_father; |
ebfd146a IR |
1450 | |
1451 | /* Get the def before the loop */ | |
1452 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); | |
1453 | return get_initial_def_for_reduction (stmt, op, scalar_def); | |
1454 | } | |
1455 | ||
1456 | /* Case 5: operand is defined by loop-header phi - induction. */ | |
1457 | case vect_induction_def: | |
1458 | { | |
1459 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1460 | ||
1461 | /* Get the def from the vectorized stmt. */ | |
1462 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1463 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1464 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1465 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1466 | else | |
1467 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1468 | return vec_oprnd; |
1469 | } | |
1470 | ||
1471 | default: | |
1472 | gcc_unreachable (); | |
1473 | } | |
1474 | } | |
1475 | ||
1476 | ||
1477 | /* Function vect_get_vec_def_for_stmt_copy | |
1478 | ||
ff802fa1 | 1479 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1480 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1481 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1482 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1483 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1484 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1485 | DT is the type of the vector def VEC_OPRND. |
1486 | ||
1487 | Context: | |
1488 | In case the vectorization factor (VF) is bigger than the number | |
1489 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1490 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1491 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1492 | smallest data-type determines the VF, and as a result, when vectorizing |
1493 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1494 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1495 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1496 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1497 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1498 | ||
1499 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1500 | |
ebfd146a IR |
1501 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1502 | VS1.1: vx.1 = memref1 VS1.2 | |
1503 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1504 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1505 | |
1506 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1507 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1508 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1509 | VSnew.3: vz3 = vx.3 + ... | |
1510 | ||
1511 | The vectorization of S1 is explained in vectorizable_load. | |
1512 | The vectorization of S2: | |
b8698a0f L |
1513 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1514 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1515 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1516 | returns the vector-def 'vx.0'. |
1517 | ||
b8698a0f L |
1518 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1519 | function is called to get the relevant vector-def for each operand. It is | |
1520 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1521 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1522 | ||
b8698a0f L |
1523 | For example, to obtain the vector-def 'vx.1' in order to create the |
1524 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1525 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1526 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1527 | and return its def ('vx.1'). | |
1528 | Overall, to create the above sequence this function will be called 3 times: | |
1529 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1530 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1531 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1532 | ||
1533 | tree | |
1534 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1535 | { | |
1536 | gimple vec_stmt_for_operand; | |
1537 | stmt_vec_info def_stmt_info; | |
1538 | ||
1539 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1540 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1541 | return vec_oprnd; |
1542 | ||
1543 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1544 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1545 | gcc_assert (def_stmt_info); | |
1546 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1547 | gcc_assert (vec_stmt_for_operand); | |
1548 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1549 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) | |
1550 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1551 | else | |
1552 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1553 | return vec_oprnd; | |
1554 | } | |
1555 | ||
1556 | ||
1557 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1558 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1559 | |
1560 | static void | |
b8698a0f | 1561 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1562 | vec<tree> *vec_oprnds0, |
1563 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1564 | { |
9771b263 | 1565 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1566 | |
1567 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1568 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1569 | |
9771b263 | 1570 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1571 | { |
9771b263 | 1572 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1573 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1574 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1575 | } |
1576 | } | |
1577 | ||
1578 | ||
d092494c IR |
1579 | /* Get vectorized definitions for OP0 and OP1. |
1580 | REDUC_INDEX is the index of reduction operand in case of reduction, | |
1581 | and -1 otherwise. */ | |
ebfd146a | 1582 | |
d092494c | 1583 | void |
ebfd146a | 1584 | vect_get_vec_defs (tree op0, tree op1, gimple stmt, |
9771b263 DN |
1585 | vec<tree> *vec_oprnds0, |
1586 | vec<tree> *vec_oprnds1, | |
d092494c | 1587 | slp_tree slp_node, int reduc_index) |
ebfd146a IR |
1588 | { |
1589 | if (slp_node) | |
d092494c IR |
1590 | { |
1591 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
9771b263 DN |
1592 | vec<tree> ops; |
1593 | ops.create (nops); | |
37b5ec8f | 1594 | vec<vec<tree> > vec_defs; |
9771b263 | 1595 | vec_defs.create (nops); |
d092494c | 1596 | |
9771b263 | 1597 | ops.quick_push (op0); |
d092494c | 1598 | if (op1) |
9771b263 | 1599 | ops.quick_push (op1); |
d092494c IR |
1600 | |
1601 | vect_get_slp_defs (ops, slp_node, &vec_defs, reduc_index); | |
1602 | ||
37b5ec8f | 1603 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1604 | if (op1) |
37b5ec8f | 1605 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1606 | |
9771b263 DN |
1607 | ops.release (); |
1608 | vec_defs.release (); | |
d092494c | 1609 | } |
ebfd146a IR |
1610 | else |
1611 | { | |
1612 | tree vec_oprnd; | |
1613 | ||
9771b263 | 1614 | vec_oprnds0->create (1); |
b8698a0f | 1615 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL); |
9771b263 | 1616 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1617 | |
1618 | if (op1) | |
1619 | { | |
9771b263 | 1620 | vec_oprnds1->create (1); |
b8698a0f | 1621 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL); |
9771b263 | 1622 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1623 | } |
1624 | } | |
1625 | } | |
1626 | ||
1627 | ||
1628 | /* Function vect_finish_stmt_generation. | |
1629 | ||
1630 | Insert a new stmt. */ | |
1631 | ||
1632 | void | |
1633 | vect_finish_stmt_generation (gimple stmt, gimple vec_stmt, | |
1634 | gimple_stmt_iterator *gsi) | |
1635 | { | |
1636 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1637 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1638 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1639 | |
1640 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1641 | ||
54e8e2c3 RG |
1642 | if (!gsi_end_p (*gsi) |
1643 | && gimple_has_mem_ops (vec_stmt)) | |
1644 | { | |
1645 | gimple at_stmt = gsi_stmt (*gsi); | |
1646 | tree vuse = gimple_vuse (at_stmt); | |
1647 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1648 | { | |
1649 | tree vdef = gimple_vdef (at_stmt); | |
1650 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1651 | /* If we have an SSA vuse and insert a store, update virtual | |
1652 | SSA form to avoid triggering the renamer. Do so only | |
1653 | if we can easily see all uses - which is what almost always | |
1654 | happens with the way vectorized stmts are inserted. */ | |
1655 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1656 | && ((is_gimple_assign (vec_stmt) | |
1657 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1658 | || (is_gimple_call (vec_stmt) | |
1659 | && !(gimple_call_flags (vec_stmt) | |
1660 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1661 | { | |
1662 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1663 | gimple_set_vdef (vec_stmt, new_vdef); | |
1664 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1665 | } | |
1666 | } | |
1667 | } | |
ebfd146a IR |
1668 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1669 | ||
b8698a0f | 1670 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo, |
a70d6342 | 1671 | bb_vinfo)); |
ebfd146a | 1672 | |
73fbfcad | 1673 | if (dump_enabled_p ()) |
ebfd146a | 1674 | { |
78c60e3d SS |
1675 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1676 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
e645e942 | 1677 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1678 | } |
1679 | ||
ad885386 | 1680 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
ebfd146a IR |
1681 | } |
1682 | ||
1683 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1684 | a function declaration if the target has a vectorized version | |
1685 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1686 | ||
1687 | tree | |
1688 | vectorizable_function (gimple call, tree vectype_out, tree vectype_in) | |
1689 | { | |
1690 | tree fndecl = gimple_call_fndecl (call); | |
ebfd146a IR |
1691 | |
1692 | /* We only handle functions that do not read or clobber memory -- i.e. | |
1693 | const or novops ones. */ | |
1694 | if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS))) | |
1695 | return NULL_TREE; | |
1696 | ||
1697 | if (!fndecl | |
1698 | || TREE_CODE (fndecl) != FUNCTION_DECL | |
1699 | || !DECL_BUILT_IN (fndecl)) | |
1700 | return NULL_TREE; | |
1701 | ||
62f7fd21 | 1702 | return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out, |
ebfd146a IR |
1703 | vectype_in); |
1704 | } | |
1705 | ||
1706 | /* Function vectorizable_call. | |
1707 | ||
b8698a0f L |
1708 | Check if STMT performs a function call that can be vectorized. |
1709 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1710 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1711 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1712 | ||
1713 | static bool | |
190c2236 JJ |
1714 | vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, |
1715 | slp_tree slp_node) | |
ebfd146a IR |
1716 | { |
1717 | tree vec_dest; | |
1718 | tree scalar_dest; | |
1719 | tree op, type; | |
1720 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1721 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
1722 | tree vectype_out, vectype_in; | |
1723 | int nunits_in; | |
1724 | int nunits_out; | |
1725 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 1726 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b690cc0f | 1727 | tree fndecl, new_temp, def, rhs_type; |
ebfd146a | 1728 | gimple def_stmt; |
0502fb85 UB |
1729 | enum vect_def_type dt[3] |
1730 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
63827fb8 | 1731 | gimple new_stmt = NULL; |
ebfd146a | 1732 | int ncopies, j; |
6e1aa848 | 1733 | vec<tree> vargs = vNULL; |
ebfd146a IR |
1734 | enum { NARROW, NONE, WIDEN } modifier; |
1735 | size_t i, nargs; | |
9d5e7640 | 1736 | tree lhs; |
ebfd146a | 1737 | |
190c2236 | 1738 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
1739 | return false; |
1740 | ||
8644a673 | 1741 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1742 | return false; |
1743 | ||
ebfd146a IR |
1744 | /* Is STMT a vectorizable call? */ |
1745 | if (!is_gimple_call (stmt)) | |
1746 | return false; | |
1747 | ||
1748 | if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
1749 | return false; | |
1750 | ||
822ba6d7 | 1751 | if (stmt_can_throw_internal (stmt)) |
5a2c1986 IR |
1752 | return false; |
1753 | ||
b690cc0f RG |
1754 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
1755 | ||
ebfd146a IR |
1756 | /* Process function arguments. */ |
1757 | rhs_type = NULL_TREE; | |
b690cc0f | 1758 | vectype_in = NULL_TREE; |
ebfd146a IR |
1759 | nargs = gimple_call_num_args (stmt); |
1760 | ||
1b1562a5 MM |
1761 | /* Bail out if the function has more than three arguments, we do not have |
1762 | interesting builtin functions to vectorize with more than two arguments | |
1763 | except for fma. No arguments is also not good. */ | |
1764 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
1765 | return false; |
1766 | ||
74bf76ed JJ |
1767 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
1768 | if (gimple_call_internal_p (stmt) | |
1769 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
1770 | { | |
1771 | nargs = 0; | |
1772 | rhs_type = unsigned_type_node; | |
1773 | } | |
1774 | ||
ebfd146a IR |
1775 | for (i = 0; i < nargs; i++) |
1776 | { | |
b690cc0f RG |
1777 | tree opvectype; |
1778 | ||
ebfd146a IR |
1779 | op = gimple_call_arg (stmt, i); |
1780 | ||
1781 | /* We can only handle calls with arguments of the same type. */ | |
1782 | if (rhs_type | |
8533c9d8 | 1783 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 1784 | { |
73fbfcad | 1785 | if (dump_enabled_p ()) |
78c60e3d | 1786 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1787 | "argument types differ.\n"); |
ebfd146a IR |
1788 | return false; |
1789 | } | |
b690cc0f RG |
1790 | if (!rhs_type) |
1791 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 1792 | |
24ee1384 | 1793 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, |
b690cc0f | 1794 | &def_stmt, &def, &dt[i], &opvectype)) |
ebfd146a | 1795 | { |
73fbfcad | 1796 | if (dump_enabled_p ()) |
78c60e3d | 1797 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1798 | "use not simple.\n"); |
ebfd146a IR |
1799 | return false; |
1800 | } | |
ebfd146a | 1801 | |
b690cc0f RG |
1802 | if (!vectype_in) |
1803 | vectype_in = opvectype; | |
1804 | else if (opvectype | |
1805 | && opvectype != vectype_in) | |
1806 | { | |
73fbfcad | 1807 | if (dump_enabled_p ()) |
78c60e3d | 1808 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1809 | "argument vector types differ.\n"); |
b690cc0f RG |
1810 | return false; |
1811 | } | |
1812 | } | |
1813 | /* If all arguments are external or constant defs use a vector type with | |
1814 | the same size as the output vector type. */ | |
ebfd146a | 1815 | if (!vectype_in) |
b690cc0f | 1816 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
1817 | if (vec_stmt) |
1818 | gcc_assert (vectype_in); | |
1819 | if (!vectype_in) | |
1820 | { | |
73fbfcad | 1821 | if (dump_enabled_p ()) |
7d8930a0 | 1822 | { |
78c60e3d SS |
1823 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1824 | "no vectype for scalar type "); | |
1825 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 1826 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
1827 | } |
1828 | ||
1829 | return false; | |
1830 | } | |
ebfd146a IR |
1831 | |
1832 | /* FORNOW */ | |
b690cc0f RG |
1833 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
1834 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
1835 | if (nunits_in == nunits_out / 2) |
1836 | modifier = NARROW; | |
1837 | else if (nunits_out == nunits_in) | |
1838 | modifier = NONE; | |
1839 | else if (nunits_out == nunits_in / 2) | |
1840 | modifier = WIDEN; | |
1841 | else | |
1842 | return false; | |
1843 | ||
1844 | /* For now, we only vectorize functions if a target specific builtin | |
1845 | is available. TODO -- in some cases, it might be profitable to | |
1846 | insert the calls for pieces of the vector, in order to be able | |
1847 | to vectorize other operations in the loop. */ | |
1848 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
1849 | if (fndecl == NULL_TREE) | |
1850 | { | |
74bf76ed JJ |
1851 | if (gimple_call_internal_p (stmt) |
1852 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE | |
1853 | && !slp_node | |
1854 | && loop_vinfo | |
1855 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
1856 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
1857 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
1858 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
1859 | { | |
1860 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
1861 | { 0, 1, 2, ... vf - 1 } vector. */ | |
1862 | gcc_assert (nargs == 0); | |
1863 | } | |
1864 | else | |
1865 | { | |
1866 | if (dump_enabled_p ()) | |
1867 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 1868 | "function is not vectorizable.\n"); |
74bf76ed JJ |
1869 | return false; |
1870 | } | |
ebfd146a IR |
1871 | } |
1872 | ||
5006671f | 1873 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a | 1874 | |
190c2236 JJ |
1875 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
1876 | ncopies = 1; | |
1877 | else if (modifier == NARROW) | |
ebfd146a IR |
1878 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
1879 | else | |
1880 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1881 | ||
1882 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
1883 | needs to be generated. */ | |
1884 | gcc_assert (ncopies >= 1); | |
1885 | ||
1886 | if (!vec_stmt) /* transformation not required. */ | |
1887 | { | |
1888 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 1889 | if (dump_enabled_p ()) |
e645e942 TJ |
1890 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
1891 | "\n"); | |
c3e7ee41 | 1892 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
1893 | return true; |
1894 | } | |
1895 | ||
1896 | /** Transform. **/ | |
1897 | ||
73fbfcad | 1898 | if (dump_enabled_p ()) |
e645e942 | 1899 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
1900 | |
1901 | /* Handle def. */ | |
1902 | scalar_dest = gimple_call_lhs (stmt); | |
1903 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1904 | ||
1905 | prev_stmt_info = NULL; | |
1906 | switch (modifier) | |
1907 | { | |
1908 | case NONE: | |
1909 | for (j = 0; j < ncopies; ++j) | |
1910 | { | |
1911 | /* Build argument list for the vectorized call. */ | |
1912 | if (j == 0) | |
9771b263 | 1913 | vargs.create (nargs); |
ebfd146a | 1914 | else |
9771b263 | 1915 | vargs.truncate (0); |
ebfd146a | 1916 | |
190c2236 JJ |
1917 | if (slp_node) |
1918 | { | |
37b5ec8f | 1919 | vec<vec<tree> > vec_defs; |
9771b263 DN |
1920 | vec_defs.create (nargs); |
1921 | vec<tree> vec_oprnds0; | |
190c2236 JJ |
1922 | |
1923 | for (i = 0; i < nargs; i++) | |
9771b263 | 1924 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 1925 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 1926 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
1927 | |
1928 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 1929 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
1930 | { |
1931 | size_t k; | |
1932 | for (k = 0; k < nargs; k++) | |
1933 | { | |
37b5ec8f | 1934 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 1935 | vargs[k] = vec_oprndsk[i]; |
190c2236 JJ |
1936 | } |
1937 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1938 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1939 | gimple_call_set_lhs (new_stmt, new_temp); | |
1940 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 1941 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
1942 | } |
1943 | ||
1944 | for (i = 0; i < nargs; i++) | |
1945 | { | |
37b5ec8f | 1946 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 1947 | vec_oprndsi.release (); |
190c2236 | 1948 | } |
9771b263 | 1949 | vec_defs.release (); |
190c2236 JJ |
1950 | continue; |
1951 | } | |
1952 | ||
ebfd146a IR |
1953 | for (i = 0; i < nargs; i++) |
1954 | { | |
1955 | op = gimple_call_arg (stmt, i); | |
1956 | if (j == 0) | |
1957 | vec_oprnd0 | |
1958 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1959 | else | |
63827fb8 IR |
1960 | { |
1961 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
1962 | vec_oprnd0 | |
1963 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
1964 | } | |
ebfd146a | 1965 | |
9771b263 | 1966 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
1967 | } |
1968 | ||
74bf76ed JJ |
1969 | if (gimple_call_internal_p (stmt) |
1970 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
1971 | { | |
1972 | tree *v = XALLOCAVEC (tree, nunits_out); | |
1973 | int k; | |
1974 | for (k = 0; k < nunits_out; ++k) | |
1975 | v[k] = build_int_cst (unsigned_type_node, j * nunits_out + k); | |
1976 | tree cst = build_vector (vectype_out, v); | |
1977 | tree new_var | |
1978 | = vect_get_new_vect_var (vectype_out, vect_simple_var, "cst_"); | |
1979 | gimple init_stmt = gimple_build_assign (new_var, cst); | |
1980 | new_temp = make_ssa_name (new_var, init_stmt); | |
1981 | gimple_assign_set_lhs (init_stmt, new_temp); | |
1982 | vect_init_vector_1 (stmt, init_stmt, NULL); | |
1983 | new_temp = make_ssa_name (vec_dest, NULL); | |
1984 | new_stmt = gimple_build_assign (new_temp, | |
1985 | gimple_assign_lhs (init_stmt)); | |
1986 | } | |
1987 | else | |
1988 | { | |
1989 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1990 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1991 | gimple_call_set_lhs (new_stmt, new_temp); | |
1992 | } | |
ebfd146a IR |
1993 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1994 | ||
1995 | if (j == 0) | |
1996 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1997 | else | |
1998 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1999 | ||
2000 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2001 | } | |
2002 | ||
2003 | break; | |
2004 | ||
2005 | case NARROW: | |
2006 | for (j = 0; j < ncopies; ++j) | |
2007 | { | |
2008 | /* Build argument list for the vectorized call. */ | |
2009 | if (j == 0) | |
9771b263 | 2010 | vargs.create (nargs * 2); |
ebfd146a | 2011 | else |
9771b263 | 2012 | vargs.truncate (0); |
ebfd146a | 2013 | |
190c2236 JJ |
2014 | if (slp_node) |
2015 | { | |
37b5ec8f | 2016 | vec<vec<tree> > vec_defs; |
9771b263 DN |
2017 | vec_defs.create (nargs); |
2018 | vec<tree> vec_oprnds0; | |
190c2236 JJ |
2019 | |
2020 | for (i = 0; i < nargs; i++) | |
9771b263 | 2021 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2022 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2023 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2024 | |
2025 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2026 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
2027 | { |
2028 | size_t k; | |
9771b263 | 2029 | vargs.truncate (0); |
190c2236 JJ |
2030 | for (k = 0; k < nargs; k++) |
2031 | { | |
37b5ec8f | 2032 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
2033 | vargs.quick_push (vec_oprndsk[i]); |
2034 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 JJ |
2035 | } |
2036 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2037 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2038 | gimple_call_set_lhs (new_stmt, new_temp); | |
2039 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2040 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2041 | } |
2042 | ||
2043 | for (i = 0; i < nargs; i++) | |
2044 | { | |
37b5ec8f | 2045 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2046 | vec_oprndsi.release (); |
190c2236 | 2047 | } |
9771b263 | 2048 | vec_defs.release (); |
190c2236 JJ |
2049 | continue; |
2050 | } | |
2051 | ||
ebfd146a IR |
2052 | for (i = 0; i < nargs; i++) |
2053 | { | |
2054 | op = gimple_call_arg (stmt, i); | |
2055 | if (j == 0) | |
2056 | { | |
2057 | vec_oprnd0 | |
2058 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
2059 | vec_oprnd1 | |
63827fb8 | 2060 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2061 | } |
2062 | else | |
2063 | { | |
336ecb65 | 2064 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 2065 | vec_oprnd0 |
63827fb8 | 2066 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 2067 | vec_oprnd1 |
63827fb8 | 2068 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2069 | } |
2070 | ||
9771b263 DN |
2071 | vargs.quick_push (vec_oprnd0); |
2072 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
2073 | } |
2074 | ||
2075 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2076 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2077 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
2078 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2079 | ||
2080 | if (j == 0) | |
2081 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2082 | else | |
2083 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2084 | ||
2085 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2086 | } | |
2087 | ||
2088 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2089 | ||
2090 | break; | |
2091 | ||
2092 | case WIDEN: | |
2093 | /* No current target implements this case. */ | |
2094 | return false; | |
2095 | } | |
2096 | ||
9771b263 | 2097 | vargs.release (); |
ebfd146a IR |
2098 | |
2099 | /* Update the exception handling table with the vector stmt if necessary. */ | |
2100 | if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt)) | |
2101 | gimple_purge_dead_eh_edges (gimple_bb (stmt)); | |
2102 | ||
2103 | /* The call in STMT might prevent it from being removed in dce. | |
2104 | We however cannot remove it here, due to the way the ssa name | |
2105 | it defines is mapped to the new definition. So just replace | |
2106 | rhs of the statement with something harmless. */ | |
2107 | ||
dd34c087 JJ |
2108 | if (slp_node) |
2109 | return true; | |
2110 | ||
ebfd146a | 2111 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
2112 | if (is_pattern_stmt_p (stmt_info)) |
2113 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
2114 | else | |
2115 | lhs = gimple_call_lhs (stmt); | |
2116 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
ebfd146a | 2117 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 2118 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
2119 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
2120 | gsi_replace (gsi, new_stmt, false); | |
2121 | SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt; | |
2122 | ||
2123 | return true; | |
2124 | } | |
2125 | ||
2126 | ||
2127 | /* Function vect_gen_widened_results_half | |
2128 | ||
2129 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 2130 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 2131 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
2132 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
2133 | needs to be created (DECL is a function-decl of a target-builtin). | |
2134 | STMT is the original scalar stmt that we are vectorizing. */ | |
2135 | ||
2136 | static gimple | |
2137 | vect_gen_widened_results_half (enum tree_code code, | |
2138 | tree decl, | |
2139 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
2140 | tree vec_dest, gimple_stmt_iterator *gsi, | |
2141 | gimple stmt) | |
b8698a0f | 2142 | { |
ebfd146a | 2143 | gimple new_stmt; |
b8698a0f L |
2144 | tree new_temp; |
2145 | ||
2146 | /* Generate half of the widened result: */ | |
2147 | if (code == CALL_EXPR) | |
2148 | { | |
2149 | /* Target specific support */ | |
ebfd146a IR |
2150 | if (op_type == binary_op) |
2151 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
2152 | else | |
2153 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
2154 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2155 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
2156 | } |
2157 | else | |
ebfd146a | 2158 | { |
b8698a0f L |
2159 | /* Generic support */ |
2160 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
2161 | if (op_type != binary_op) |
2162 | vec_oprnd1 = NULL; | |
2163 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0, | |
2164 | vec_oprnd1); | |
2165 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2166 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 2167 | } |
ebfd146a IR |
2168 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2169 | ||
ebfd146a IR |
2170 | return new_stmt; |
2171 | } | |
2172 | ||
4a00c761 JJ |
2173 | |
2174 | /* Get vectorized definitions for loop-based vectorization. For the first | |
2175 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
2176 | scalar operand), and for the rest we get a copy with | |
2177 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
2178 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
2179 | The vectors are collected into VEC_OPRNDS. */ | |
2180 | ||
2181 | static void | |
2182 | vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt, | |
9771b263 | 2183 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
2184 | { |
2185 | tree vec_oprnd; | |
2186 | ||
2187 | /* Get first vector operand. */ | |
2188 | /* All the vector operands except the very first one (that is scalar oprnd) | |
2189 | are stmt copies. */ | |
2190 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
2191 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL); | |
2192 | else | |
2193 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
2194 | ||
9771b263 | 2195 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
2196 | |
2197 | /* Get second vector operand. */ | |
2198 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 2199 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
2200 | |
2201 | *oprnd = vec_oprnd; | |
2202 | ||
2203 | /* For conversion in multiple steps, continue to get operands | |
2204 | recursively. */ | |
2205 | if (multi_step_cvt) | |
2206 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
2207 | } | |
2208 | ||
2209 | ||
2210 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
2211 | For multi-step conversions store the resulting vectors and call the function | |
2212 | recursively. */ | |
2213 | ||
2214 | static void | |
9771b263 | 2215 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
4a00c761 | 2216 | int multi_step_cvt, gimple stmt, |
9771b263 | 2217 | vec<tree> vec_dsts, |
4a00c761 JJ |
2218 | gimple_stmt_iterator *gsi, |
2219 | slp_tree slp_node, enum tree_code code, | |
2220 | stmt_vec_info *prev_stmt_info) | |
2221 | { | |
2222 | unsigned int i; | |
2223 | tree vop0, vop1, new_tmp, vec_dest; | |
2224 | gimple new_stmt; | |
2225 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2226 | ||
9771b263 | 2227 | vec_dest = vec_dsts.pop (); |
4a00c761 | 2228 | |
9771b263 | 2229 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
2230 | { |
2231 | /* Create demotion operation. */ | |
9771b263 DN |
2232 | vop0 = (*vec_oprnds)[i]; |
2233 | vop1 = (*vec_oprnds)[i + 1]; | |
4a00c761 JJ |
2234 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); |
2235 | new_tmp = make_ssa_name (vec_dest, new_stmt); | |
2236 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
2237 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2238 | ||
2239 | if (multi_step_cvt) | |
2240 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 2241 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
2242 | else |
2243 | { | |
2244 | /* This is the last step of the conversion sequence. Store the | |
2245 | vectors in SLP_NODE or in vector info of the scalar statement | |
2246 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
2247 | if (slp_node) | |
9771b263 | 2248 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2249 | else |
2250 | { | |
2251 | if (!*prev_stmt_info) | |
2252 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2253 | else | |
2254 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
2255 | ||
2256 | *prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2257 | } | |
2258 | } | |
2259 | } | |
2260 | ||
2261 | /* For multi-step demotion operations we first generate demotion operations | |
2262 | from the source type to the intermediate types, and then combine the | |
2263 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
2264 | type. */ | |
2265 | if (multi_step_cvt) | |
2266 | { | |
2267 | /* At each level of recursion we have half of the operands we had at the | |
2268 | previous level. */ | |
9771b263 | 2269 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
2270 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
2271 | stmt, vec_dsts, gsi, slp_node, | |
2272 | VEC_PACK_TRUNC_EXPR, | |
2273 | prev_stmt_info); | |
2274 | } | |
2275 | ||
9771b263 | 2276 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2277 | } |
2278 | ||
2279 | ||
2280 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
2281 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
2282 | the resulting vectors and call the function recursively. */ | |
2283 | ||
2284 | static void | |
9771b263 DN |
2285 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
2286 | vec<tree> *vec_oprnds1, | |
4a00c761 JJ |
2287 | gimple stmt, tree vec_dest, |
2288 | gimple_stmt_iterator *gsi, | |
2289 | enum tree_code code1, | |
2290 | enum tree_code code2, tree decl1, | |
2291 | tree decl2, int op_type) | |
2292 | { | |
2293 | int i; | |
2294 | tree vop0, vop1, new_tmp1, new_tmp2; | |
2295 | gimple new_stmt1, new_stmt2; | |
6e1aa848 | 2296 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 2297 | |
9771b263 DN |
2298 | vec_tmp.create (vec_oprnds0->length () * 2); |
2299 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
2300 | { |
2301 | if (op_type == binary_op) | |
9771b263 | 2302 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
2303 | else |
2304 | vop1 = NULL_TREE; | |
2305 | ||
2306 | /* Generate the two halves of promotion operation. */ | |
2307 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
2308 | op_type, vec_dest, gsi, stmt); | |
2309 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
2310 | op_type, vec_dest, gsi, stmt); | |
2311 | if (is_gimple_call (new_stmt1)) | |
2312 | { | |
2313 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
2314 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
2315 | } | |
2316 | else | |
2317 | { | |
2318 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
2319 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
2320 | } | |
2321 | ||
2322 | /* Store the results for the next step. */ | |
9771b263 DN |
2323 | vec_tmp.quick_push (new_tmp1); |
2324 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
2325 | } |
2326 | ||
689eaba3 | 2327 | vec_oprnds0->release (); |
4a00c761 JJ |
2328 | *vec_oprnds0 = vec_tmp; |
2329 | } | |
2330 | ||
2331 | ||
b8698a0f L |
2332 | /* Check if STMT performs a conversion operation, that can be vectorized. |
2333 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 2334 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
2335 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
2336 | ||
2337 | static bool | |
2338 | vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi, | |
2339 | gimple *vec_stmt, slp_tree slp_node) | |
2340 | { | |
2341 | tree vec_dest; | |
2342 | tree scalar_dest; | |
4a00c761 | 2343 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
2344 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
2345 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2346 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2347 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 2348 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
2349 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
2350 | tree new_temp; | |
2351 | tree def; | |
2352 | gimple def_stmt; | |
2353 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2354 | gimple new_stmt = NULL; | |
2355 | stmt_vec_info prev_stmt_info; | |
2356 | int nunits_in; | |
2357 | int nunits_out; | |
2358 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
2359 | int ncopies, i, j; |
2360 | tree lhs_type, rhs_type; | |
ebfd146a | 2361 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
2362 | vec<tree> vec_oprnds0 = vNULL; |
2363 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 2364 | tree vop0; |
4a00c761 JJ |
2365 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2366 | int multi_step_cvt = 0; | |
6e1aa848 DN |
2367 | vec<tree> vec_dsts = vNULL; |
2368 | vec<tree> interm_types = vNULL; | |
4a00c761 JJ |
2369 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
2370 | int op_type; | |
2371 | enum machine_mode rhs_mode; | |
2372 | unsigned short fltsz; | |
ebfd146a IR |
2373 | |
2374 | /* Is STMT a vectorizable conversion? */ | |
2375 | ||
4a00c761 | 2376 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2377 | return false; |
2378 | ||
8644a673 | 2379 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2380 | return false; |
2381 | ||
2382 | if (!is_gimple_assign (stmt)) | |
2383 | return false; | |
2384 | ||
2385 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2386 | return false; | |
2387 | ||
2388 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
2389 | if (!CONVERT_EXPR_CODE_P (code) |
2390 | && code != FIX_TRUNC_EXPR | |
2391 | && code != FLOAT_EXPR | |
2392 | && code != WIDEN_MULT_EXPR | |
2393 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
2394 | return false; |
2395 | ||
4a00c761 JJ |
2396 | op_type = TREE_CODE_LENGTH (code); |
2397 | ||
ebfd146a | 2398 | /* Check types of lhs and rhs. */ |
b690cc0f | 2399 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 2400 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
2401 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2402 | ||
ebfd146a IR |
2403 | op0 = gimple_assign_rhs1 (stmt); |
2404 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
2405 | |
2406 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
2407 | && !((INTEGRAL_TYPE_P (lhs_type) | |
2408 | && INTEGRAL_TYPE_P (rhs_type)) | |
2409 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
2410 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
2411 | return false; | |
2412 | ||
2413 | if ((INTEGRAL_TYPE_P (lhs_type) | |
2414 | && (TYPE_PRECISION (lhs_type) | |
2415 | != GET_MODE_PRECISION (TYPE_MODE (lhs_type)))) | |
2416 | || (INTEGRAL_TYPE_P (rhs_type) | |
2417 | && (TYPE_PRECISION (rhs_type) | |
2418 | != GET_MODE_PRECISION (TYPE_MODE (rhs_type))))) | |
2419 | { | |
73fbfcad | 2420 | if (dump_enabled_p ()) |
78c60e3d | 2421 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
2422 | "type conversion to/from bit-precision unsupported." |
2423 | "\n"); | |
4a00c761 JJ |
2424 | return false; |
2425 | } | |
2426 | ||
b690cc0f | 2427 | /* Check the operands of the operation. */ |
24ee1384 | 2428 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
b690cc0f RG |
2429 | &def_stmt, &def, &dt[0], &vectype_in)) |
2430 | { | |
73fbfcad | 2431 | if (dump_enabled_p ()) |
78c60e3d | 2432 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2433 | "use not simple.\n"); |
b690cc0f RG |
2434 | return false; |
2435 | } | |
4a00c761 JJ |
2436 | if (op_type == binary_op) |
2437 | { | |
2438 | bool ok; | |
2439 | ||
2440 | op1 = gimple_assign_rhs2 (stmt); | |
2441 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
2442 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
2443 | OP1. */ | |
2444 | if (CONSTANT_CLASS_P (op0)) | |
f5709183 | 2445 | ok = vect_is_simple_use_1 (op1, stmt, loop_vinfo, bb_vinfo, |
4a00c761 JJ |
2446 | &def_stmt, &def, &dt[1], &vectype_in); |
2447 | else | |
f5709183 | 2448 | ok = vect_is_simple_use (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
24ee1384 | 2449 | &def, &dt[1]); |
4a00c761 JJ |
2450 | |
2451 | if (!ok) | |
2452 | { | |
73fbfcad | 2453 | if (dump_enabled_p ()) |
78c60e3d | 2454 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2455 | "use not simple.\n"); |
4a00c761 JJ |
2456 | return false; |
2457 | } | |
2458 | } | |
2459 | ||
b690cc0f RG |
2460 | /* If op0 is an external or constant defs use a vector type of |
2461 | the same size as the output vector type. */ | |
ebfd146a | 2462 | if (!vectype_in) |
b690cc0f | 2463 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2464 | if (vec_stmt) |
2465 | gcc_assert (vectype_in); | |
2466 | if (!vectype_in) | |
2467 | { | |
73fbfcad | 2468 | if (dump_enabled_p ()) |
4a00c761 | 2469 | { |
78c60e3d SS |
2470 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2471 | "no vectype for scalar type "); | |
2472 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2473 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 2474 | } |
7d8930a0 IR |
2475 | |
2476 | return false; | |
2477 | } | |
ebfd146a | 2478 | |
b690cc0f RG |
2479 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2480 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 2481 | if (nunits_in < nunits_out) |
ebfd146a IR |
2482 | modifier = NARROW; |
2483 | else if (nunits_out == nunits_in) | |
2484 | modifier = NONE; | |
ebfd146a | 2485 | else |
4a00c761 | 2486 | modifier = WIDEN; |
ebfd146a | 2487 | |
ff802fa1 IR |
2488 | /* Multiple types in SLP are handled by creating the appropriate number of |
2489 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2490 | case of SLP. */ | |
437f4a00 | 2491 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a | 2492 | ncopies = 1; |
4a00c761 JJ |
2493 | else if (modifier == NARROW) |
2494 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
2495 | else | |
2496 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 2497 | |
ebfd146a IR |
2498 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
2499 | needs to be generated. */ | |
2500 | gcc_assert (ncopies >= 1); | |
2501 | ||
ebfd146a | 2502 | /* Supportable by target? */ |
4a00c761 | 2503 | switch (modifier) |
ebfd146a | 2504 | { |
4a00c761 JJ |
2505 | case NONE: |
2506 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
2507 | return false; | |
2508 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
2509 | &decl1, &code1)) | |
2510 | break; | |
2511 | /* FALLTHRU */ | |
2512 | unsupported: | |
73fbfcad | 2513 | if (dump_enabled_p ()) |
78c60e3d | 2514 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2515 | "conversion not supported by target.\n"); |
ebfd146a | 2516 | return false; |
ebfd146a | 2517 | |
4a00c761 JJ |
2518 | case WIDEN: |
2519 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
2520 | &code1, &code2, &multi_step_cvt, |
2521 | &interm_types)) | |
4a00c761 JJ |
2522 | { |
2523 | /* Binary widening operation can only be supported directly by the | |
2524 | architecture. */ | |
2525 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
2526 | break; | |
2527 | } | |
2528 | ||
2529 | if (code != FLOAT_EXPR | |
2530 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
2531 | <= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
2532 | goto unsupported; | |
2533 | ||
2534 | rhs_mode = TYPE_MODE (rhs_type); | |
2535 | fltsz = GET_MODE_SIZE (TYPE_MODE (lhs_type)); | |
2536 | for (rhs_mode = GET_MODE_2XWIDER_MODE (TYPE_MODE (rhs_type)); | |
2537 | rhs_mode != VOIDmode && GET_MODE_SIZE (rhs_mode) <= fltsz; | |
2538 | rhs_mode = GET_MODE_2XWIDER_MODE (rhs_mode)) | |
2539 | { | |
2540 | cvt_type | |
2541 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
2542 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
2543 | if (cvt_type == NULL_TREE) | |
2544 | goto unsupported; | |
2545 | ||
2546 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
2547 | { | |
2548 | if (!supportable_convert_operation (code, vectype_out, | |
2549 | cvt_type, &decl1, &codecvt1)) | |
2550 | goto unsupported; | |
2551 | } | |
2552 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
2553 | cvt_type, &codecvt1, |
2554 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
2555 | &interm_types)) |
2556 | continue; | |
2557 | else | |
2558 | gcc_assert (multi_step_cvt == 0); | |
2559 | ||
2560 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
2561 | vectype_in, &code1, &code2, |
2562 | &multi_step_cvt, &interm_types)) | |
4a00c761 JJ |
2563 | break; |
2564 | } | |
2565 | ||
2566 | if (rhs_mode == VOIDmode || GET_MODE_SIZE (rhs_mode) > fltsz) | |
2567 | goto unsupported; | |
2568 | ||
2569 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
2570 | codecvt2 = ERROR_MARK; | |
2571 | else | |
2572 | { | |
2573 | multi_step_cvt++; | |
9771b263 | 2574 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
2575 | cvt_type = NULL_TREE; |
2576 | } | |
2577 | break; | |
2578 | ||
2579 | case NARROW: | |
2580 | gcc_assert (op_type == unary_op); | |
2581 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
2582 | &code1, &multi_step_cvt, | |
2583 | &interm_types)) | |
2584 | break; | |
2585 | ||
2586 | if (code != FIX_TRUNC_EXPR | |
2587 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
2588 | >= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
2589 | goto unsupported; | |
2590 | ||
2591 | rhs_mode = TYPE_MODE (rhs_type); | |
2592 | cvt_type | |
2593 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
2594 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
2595 | if (cvt_type == NULL_TREE) | |
2596 | goto unsupported; | |
2597 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
2598 | &decl1, &codecvt1)) | |
2599 | goto unsupported; | |
2600 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
2601 | &code1, &multi_step_cvt, | |
2602 | &interm_types)) | |
2603 | break; | |
2604 | goto unsupported; | |
2605 | ||
2606 | default: | |
2607 | gcc_unreachable (); | |
ebfd146a IR |
2608 | } |
2609 | ||
2610 | if (!vec_stmt) /* transformation not required. */ | |
2611 | { | |
73fbfcad | 2612 | if (dump_enabled_p ()) |
78c60e3d | 2613 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 2614 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 2615 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
2616 | { |
2617 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
c3e7ee41 | 2618 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
8bd37302 | 2619 | } |
4a00c761 JJ |
2620 | else if (modifier == NARROW) |
2621 | { | |
2622 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 2623 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
2624 | } |
2625 | else | |
2626 | { | |
2627 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 2628 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 2629 | } |
9771b263 | 2630 | interm_types.release (); |
ebfd146a IR |
2631 | return true; |
2632 | } | |
2633 | ||
2634 | /** Transform. **/ | |
73fbfcad | 2635 | if (dump_enabled_p ()) |
78c60e3d | 2636 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 2637 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 2638 | |
4a00c761 JJ |
2639 | if (op_type == binary_op) |
2640 | { | |
2641 | if (CONSTANT_CLASS_P (op0)) | |
2642 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
2643 | else if (CONSTANT_CLASS_P (op1)) | |
2644 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
2645 | } | |
2646 | ||
2647 | /* In case of multi-step conversion, we first generate conversion operations | |
2648 | to the intermediate types, and then from that types to the final one. | |
2649 | We create vector destinations for the intermediate type (TYPES) received | |
2650 | from supportable_*_operation, and store them in the correct order | |
2651 | for future use in vect_create_vectorized_*_stmts (). */ | |
9771b263 | 2652 | vec_dsts.create (multi_step_cvt + 1); |
82294ec1 JJ |
2653 | vec_dest = vect_create_destination_var (scalar_dest, |
2654 | (cvt_type && modifier == WIDEN) | |
2655 | ? cvt_type : vectype_out); | |
9771b263 | 2656 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2657 | |
2658 | if (multi_step_cvt) | |
2659 | { | |
9771b263 DN |
2660 | for (i = interm_types.length () - 1; |
2661 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
2662 | { |
2663 | vec_dest = vect_create_destination_var (scalar_dest, | |
2664 | intermediate_type); | |
9771b263 | 2665 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2666 | } |
2667 | } | |
ebfd146a | 2668 | |
4a00c761 | 2669 | if (cvt_type) |
82294ec1 JJ |
2670 | vec_dest = vect_create_destination_var (scalar_dest, |
2671 | modifier == WIDEN | |
2672 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
2673 | |
2674 | if (!slp_node) | |
2675 | { | |
30862efc | 2676 | if (modifier == WIDEN) |
4a00c761 | 2677 | { |
9771b263 | 2678 | vec_oprnds0.create (multi_step_cvt ? vect_pow2(multi_step_cvt) : 1); |
4a00c761 | 2679 | if (op_type == binary_op) |
9771b263 | 2680 | vec_oprnds1.create (1); |
4a00c761 | 2681 | } |
30862efc | 2682 | else if (modifier == NARROW) |
9771b263 DN |
2683 | vec_oprnds0.create ( |
2684 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
2685 | } |
2686 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 2687 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 2688 | |
4a00c761 | 2689 | last_oprnd = op0; |
ebfd146a IR |
2690 | prev_stmt_info = NULL; |
2691 | switch (modifier) | |
2692 | { | |
2693 | case NONE: | |
2694 | for (j = 0; j < ncopies; j++) | |
2695 | { | |
ebfd146a | 2696 | if (j == 0) |
d092494c IR |
2697 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node, |
2698 | -1); | |
ebfd146a IR |
2699 | else |
2700 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
2701 | ||
9771b263 | 2702 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2703 | { |
2704 | /* Arguments are ready, create the new vector stmt. */ | |
2705 | if (code1 == CALL_EXPR) | |
2706 | { | |
2707 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2708 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2709 | gimple_call_set_lhs (new_stmt, new_temp); | |
2710 | } | |
2711 | else | |
2712 | { | |
2713 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
2714 | new_stmt = gimple_build_assign_with_ops (code1, vec_dest, | |
2715 | vop0, NULL); | |
2716 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2717 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2718 | } | |
2719 | ||
2720 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2721 | if (slp_node) | |
9771b263 | 2722 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2723 | } |
2724 | ||
ebfd146a IR |
2725 | if (j == 0) |
2726 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2727 | else | |
2728 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2729 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2730 | } | |
2731 | break; | |
2732 | ||
2733 | case WIDEN: | |
2734 | /* In case the vectorization factor (VF) is bigger than the number | |
2735 | of elements that we can fit in a vectype (nunits), we have to | |
2736 | generate more than one vector stmt - i.e - we need to "unroll" | |
2737 | the vector stmt by a factor VF/nunits. */ | |
2738 | for (j = 0; j < ncopies; j++) | |
2739 | { | |
4a00c761 | 2740 | /* Handle uses. */ |
ebfd146a | 2741 | if (j == 0) |
4a00c761 JJ |
2742 | { |
2743 | if (slp_node) | |
2744 | { | |
2745 | if (code == WIDEN_LSHIFT_EXPR) | |
2746 | { | |
2747 | unsigned int k; | |
ebfd146a | 2748 | |
4a00c761 JJ |
2749 | vec_oprnd1 = op1; |
2750 | /* Store vec_oprnd1 for every vector stmt to be created | |
2751 | for SLP_NODE. We check during the analysis that all | |
2752 | the shift arguments are the same. */ | |
2753 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 2754 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
2755 | |
2756 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2757 | slp_node, -1); | |
2758 | } | |
2759 | else | |
2760 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
2761 | &vec_oprnds1, slp_node, -1); | |
2762 | } | |
2763 | else | |
2764 | { | |
2765 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
9771b263 | 2766 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
2767 | if (op_type == binary_op) |
2768 | { | |
2769 | if (code == WIDEN_LSHIFT_EXPR) | |
2770 | vec_oprnd1 = op1; | |
2771 | else | |
2772 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, | |
2773 | NULL); | |
9771b263 | 2774 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
2775 | } |
2776 | } | |
2777 | } | |
ebfd146a | 2778 | else |
4a00c761 JJ |
2779 | { |
2780 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
2781 | vec_oprnds0.truncate (0); |
2782 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
2783 | if (op_type == binary_op) |
2784 | { | |
2785 | if (code == WIDEN_LSHIFT_EXPR) | |
2786 | vec_oprnd1 = op1; | |
2787 | else | |
2788 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
2789 | vec_oprnd1); | |
9771b263 DN |
2790 | vec_oprnds1.truncate (0); |
2791 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
2792 | } |
2793 | } | |
ebfd146a | 2794 | |
4a00c761 JJ |
2795 | /* Arguments are ready. Create the new vector stmts. */ |
2796 | for (i = multi_step_cvt; i >= 0; i--) | |
2797 | { | |
9771b263 | 2798 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
2799 | enum tree_code c1 = code1, c2 = code2; |
2800 | if (i == 0 && codecvt2 != ERROR_MARK) | |
2801 | { | |
2802 | c1 = codecvt1; | |
2803 | c2 = codecvt2; | |
2804 | } | |
2805 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
2806 | &vec_oprnds1, | |
2807 | stmt, this_dest, gsi, | |
2808 | c1, c2, decl1, decl2, | |
2809 | op_type); | |
2810 | } | |
2811 | ||
9771b263 | 2812 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2813 | { |
2814 | if (cvt_type) | |
2815 | { | |
2816 | if (codecvt1 == CALL_EXPR) | |
2817 | { | |
2818 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2819 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2820 | gimple_call_set_lhs (new_stmt, new_temp); | |
2821 | } | |
2822 | else | |
2823 | { | |
2824 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
2825 | new_temp = make_ssa_name (vec_dest, NULL); | |
2826 | new_stmt = gimple_build_assign_with_ops (codecvt1, | |
2827 | new_temp, | |
2828 | vop0, NULL); | |
2829 | } | |
2830 | ||
2831 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2832 | } | |
2833 | else | |
2834 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
2835 | ||
2836 | if (slp_node) | |
9771b263 | 2837 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2838 | else |
2839 | { | |
2840 | if (!prev_stmt_info) | |
2841 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2842 | else | |
2843 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2844 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2845 | } | |
2846 | } | |
ebfd146a | 2847 | } |
4a00c761 JJ |
2848 | |
2849 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
2850 | break; |
2851 | ||
2852 | case NARROW: | |
2853 | /* In case the vectorization factor (VF) is bigger than the number | |
2854 | of elements that we can fit in a vectype (nunits), we have to | |
2855 | generate more than one vector stmt - i.e - we need to "unroll" | |
2856 | the vector stmt by a factor VF/nunits. */ | |
2857 | for (j = 0; j < ncopies; j++) | |
2858 | { | |
2859 | /* Handle uses. */ | |
4a00c761 JJ |
2860 | if (slp_node) |
2861 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2862 | slp_node, -1); | |
ebfd146a IR |
2863 | else |
2864 | { | |
9771b263 | 2865 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
2866 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
2867 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
2868 | } |
2869 | ||
4a00c761 JJ |
2870 | /* Arguments are ready. Create the new vector stmts. */ |
2871 | if (cvt_type) | |
9771b263 | 2872 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2873 | { |
2874 | if (codecvt1 == CALL_EXPR) | |
2875 | { | |
2876 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2877 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2878 | gimple_call_set_lhs (new_stmt, new_temp); | |
2879 | } | |
2880 | else | |
2881 | { | |
2882 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
2883 | new_temp = make_ssa_name (vec_dest, NULL); | |
2884 | new_stmt = gimple_build_assign_with_ops (codecvt1, new_temp, | |
2885 | vop0, NULL); | |
2886 | } | |
ebfd146a | 2887 | |
4a00c761 | 2888 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 2889 | vec_oprnds0[i] = new_temp; |
4a00c761 | 2890 | } |
ebfd146a | 2891 | |
4a00c761 JJ |
2892 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
2893 | stmt, vec_dsts, gsi, | |
2894 | slp_node, code1, | |
2895 | &prev_stmt_info); | |
ebfd146a IR |
2896 | } |
2897 | ||
2898 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 2899 | break; |
ebfd146a IR |
2900 | } |
2901 | ||
9771b263 DN |
2902 | vec_oprnds0.release (); |
2903 | vec_oprnds1.release (); | |
2904 | vec_dsts.release (); | |
2905 | interm_types.release (); | |
ebfd146a IR |
2906 | |
2907 | return true; | |
2908 | } | |
ff802fa1 IR |
2909 | |
2910 | ||
ebfd146a IR |
2911 | /* Function vectorizable_assignment. |
2912 | ||
b8698a0f L |
2913 | Check if STMT performs an assignment (copy) that can be vectorized. |
2914 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
2915 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2916 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2917 | ||
2918 | static bool | |
2919 | vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi, | |
2920 | gimple *vec_stmt, slp_tree slp_node) | |
2921 | { | |
2922 | tree vec_dest; | |
2923 | tree scalar_dest; | |
2924 | tree op; | |
2925 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2926 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2927 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2928 | tree new_temp; | |
2929 | tree def; | |
2930 | gimple def_stmt; | |
2931 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
fde9c428 | 2932 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
ebfd146a | 2933 | int ncopies; |
f18b55bd | 2934 | int i, j; |
6e1aa848 | 2935 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 2936 | tree vop; |
a70d6342 | 2937 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
f18b55bd IR |
2938 | gimple new_stmt = NULL; |
2939 | stmt_vec_info prev_stmt_info = NULL; | |
fde9c428 RG |
2940 | enum tree_code code; |
2941 | tree vectype_in; | |
ebfd146a IR |
2942 | |
2943 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2944 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2945 | case of SLP. */ | |
437f4a00 | 2946 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
2947 | ncopies = 1; |
2948 | else | |
2949 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2950 | ||
2951 | gcc_assert (ncopies >= 1); | |
ebfd146a | 2952 | |
a70d6342 | 2953 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2954 | return false; |
2955 | ||
8644a673 | 2956 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2957 | return false; |
2958 | ||
2959 | /* Is vectorizable assignment? */ | |
2960 | if (!is_gimple_assign (stmt)) | |
2961 | return false; | |
2962 | ||
2963 | scalar_dest = gimple_assign_lhs (stmt); | |
2964 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
2965 | return false; | |
2966 | ||
fde9c428 | 2967 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 2968 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
2969 | || code == PAREN_EXPR |
2970 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
2971 | op = gimple_assign_rhs1 (stmt); |
2972 | else | |
2973 | return false; | |
2974 | ||
7b7ec6c5 RG |
2975 | if (code == VIEW_CONVERT_EXPR) |
2976 | op = TREE_OPERAND (op, 0); | |
2977 | ||
24ee1384 | 2978 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, |
fde9c428 | 2979 | &def_stmt, &def, &dt[0], &vectype_in)) |
ebfd146a | 2980 | { |
73fbfcad | 2981 | if (dump_enabled_p ()) |
78c60e3d | 2982 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2983 | "use not simple.\n"); |
ebfd146a IR |
2984 | return false; |
2985 | } | |
2986 | ||
fde9c428 RG |
2987 | /* We can handle NOP_EXPR conversions that do not change the number |
2988 | of elements or the vector size. */ | |
7b7ec6c5 RG |
2989 | if ((CONVERT_EXPR_CODE_P (code) |
2990 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
2991 | && (!vectype_in |
2992 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
2993 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
2994 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
2995 | return false; | |
2996 | ||
7b7b1813 RG |
2997 | /* We do not handle bit-precision changes. */ |
2998 | if ((CONVERT_EXPR_CODE_P (code) | |
2999 | || code == VIEW_CONVERT_EXPR) | |
3000 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
3001 | && ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
3002 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
3003 | || ((TYPE_PRECISION (TREE_TYPE (op)) | |
3004 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))) | |
3005 | /* But a conversion that does not change the bit-pattern is ok. */ | |
3006 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
3007 | > TYPE_PRECISION (TREE_TYPE (op))) | |
3008 | && TYPE_UNSIGNED (TREE_TYPE (op)))) | |
3009 | { | |
73fbfcad | 3010 | if (dump_enabled_p ()) |
78c60e3d SS |
3011 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3012 | "type conversion to/from bit-precision " | |
e645e942 | 3013 | "unsupported.\n"); |
7b7b1813 RG |
3014 | return false; |
3015 | } | |
3016 | ||
ebfd146a IR |
3017 | if (!vec_stmt) /* transformation not required. */ |
3018 | { | |
3019 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 3020 | if (dump_enabled_p ()) |
78c60e3d | 3021 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3022 | "=== vectorizable_assignment ===\n"); |
c3e7ee41 | 3023 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
3024 | return true; |
3025 | } | |
3026 | ||
3027 | /** Transform. **/ | |
73fbfcad | 3028 | if (dump_enabled_p ()) |
e645e942 | 3029 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
3030 | |
3031 | /* Handle def. */ | |
3032 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3033 | ||
3034 | /* Handle use. */ | |
f18b55bd | 3035 | for (j = 0; j < ncopies; j++) |
ebfd146a | 3036 | { |
f18b55bd IR |
3037 | /* Handle uses. */ |
3038 | if (j == 0) | |
d092494c | 3039 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node, -1); |
f18b55bd IR |
3040 | else |
3041 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
3042 | ||
3043 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 3044 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 3045 | { |
7b7ec6c5 RG |
3046 | if (CONVERT_EXPR_CODE_P (code) |
3047 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 3048 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
3049 | new_stmt = gimple_build_assign (vec_dest, vop); |
3050 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3051 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3052 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3053 | if (slp_node) | |
9771b263 | 3054 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 3055 | } |
ebfd146a IR |
3056 | |
3057 | if (slp_node) | |
f18b55bd IR |
3058 | continue; |
3059 | ||
3060 | if (j == 0) | |
3061 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3062 | else | |
3063 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3064 | ||
3065 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3066 | } | |
b8698a0f | 3067 | |
9771b263 | 3068 | vec_oprnds.release (); |
ebfd146a IR |
3069 | return true; |
3070 | } | |
3071 | ||
9dc3f7de | 3072 | |
1107f3ae IR |
3073 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
3074 | either as shift by a scalar or by a vector. */ | |
3075 | ||
3076 | bool | |
3077 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
3078 | { | |
3079 | ||
3080 | enum machine_mode vec_mode; | |
3081 | optab optab; | |
3082 | int icode; | |
3083 | tree vectype; | |
3084 | ||
3085 | vectype = get_vectype_for_scalar_type (scalar_type); | |
3086 | if (!vectype) | |
3087 | return false; | |
3088 | ||
3089 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
3090 | if (!optab | |
3091 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
3092 | { | |
3093 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
3094 | if (!optab | |
3095 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
3096 | == CODE_FOR_nothing)) | |
3097 | return false; | |
3098 | } | |
3099 | ||
3100 | vec_mode = TYPE_MODE (vectype); | |
3101 | icode = (int) optab_handler (optab, vec_mode); | |
3102 | if (icode == CODE_FOR_nothing) | |
3103 | return false; | |
3104 | ||
3105 | return true; | |
3106 | } | |
3107 | ||
3108 | ||
9dc3f7de IR |
3109 | /* Function vectorizable_shift. |
3110 | ||
3111 | Check if STMT performs a shift operation that can be vectorized. | |
3112 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3113 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3114 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3115 | ||
3116 | static bool | |
3117 | vectorizable_shift (gimple stmt, gimple_stmt_iterator *gsi, | |
3118 | gimple *vec_stmt, slp_tree slp_node) | |
3119 | { | |
3120 | tree vec_dest; | |
3121 | tree scalar_dest; | |
3122 | tree op0, op1 = NULL; | |
3123 | tree vec_oprnd1 = NULL_TREE; | |
3124 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3125 | tree vectype; | |
3126 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3127 | enum tree_code code; | |
3128 | enum machine_mode vec_mode; | |
3129 | tree new_temp; | |
3130 | optab optab; | |
3131 | int icode; | |
3132 | enum machine_mode optab_op2_mode; | |
3133 | tree def; | |
3134 | gimple def_stmt; | |
3135 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
3136 | gimple new_stmt = NULL; | |
3137 | stmt_vec_info prev_stmt_info; | |
3138 | int nunits_in; | |
3139 | int nunits_out; | |
3140 | tree vectype_out; | |
cede2577 | 3141 | tree op1_vectype; |
9dc3f7de IR |
3142 | int ncopies; |
3143 | int j, i; | |
6e1aa848 DN |
3144 | vec<tree> vec_oprnds0 = vNULL; |
3145 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
3146 | tree vop0, vop1; |
3147 | unsigned int k; | |
49eab32e | 3148 | bool scalar_shift_arg = true; |
9dc3f7de IR |
3149 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3150 | int vf; | |
3151 | ||
3152 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
3153 | return false; | |
3154 | ||
3155 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
3156 | return false; | |
3157 | ||
3158 | /* Is STMT a vectorizable binary/unary operation? */ | |
3159 | if (!is_gimple_assign (stmt)) | |
3160 | return false; | |
3161 | ||
3162 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3163 | return false; | |
3164 | ||
3165 | code = gimple_assign_rhs_code (stmt); | |
3166 | ||
3167 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
3168 | || code == RROTATE_EXPR)) | |
3169 | return false; | |
3170 | ||
3171 | scalar_dest = gimple_assign_lhs (stmt); | |
3172 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
7b7b1813 RG |
3173 | if (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
3174 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
3175 | { | |
73fbfcad | 3176 | if (dump_enabled_p ()) |
78c60e3d | 3177 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3178 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
3179 | return false; |
3180 | } | |
9dc3f7de IR |
3181 | |
3182 | op0 = gimple_assign_rhs1 (stmt); | |
24ee1384 | 3183 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
9dc3f7de IR |
3184 | &def_stmt, &def, &dt[0], &vectype)) |
3185 | { | |
73fbfcad | 3186 | if (dump_enabled_p ()) |
78c60e3d | 3187 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3188 | "use not simple.\n"); |
9dc3f7de IR |
3189 | return false; |
3190 | } | |
3191 | /* If op0 is an external or constant def use a vector type with | |
3192 | the same size as the output vector type. */ | |
3193 | if (!vectype) | |
3194 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
3195 | if (vec_stmt) | |
3196 | gcc_assert (vectype); | |
3197 | if (!vectype) | |
3198 | { | |
73fbfcad | 3199 | if (dump_enabled_p ()) |
78c60e3d | 3200 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3201 | "no vectype for scalar type\n"); |
9dc3f7de IR |
3202 | return false; |
3203 | } | |
3204 | ||
3205 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
3206 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
3207 | if (nunits_out != nunits_in) | |
3208 | return false; | |
3209 | ||
3210 | op1 = gimple_assign_rhs2 (stmt); | |
24ee1384 IR |
3211 | if (!vect_is_simple_use_1 (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3212 | &def, &dt[1], &op1_vectype)) | |
9dc3f7de | 3213 | { |
73fbfcad | 3214 | if (dump_enabled_p ()) |
78c60e3d | 3215 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3216 | "use not simple.\n"); |
9dc3f7de IR |
3217 | return false; |
3218 | } | |
3219 | ||
3220 | if (loop_vinfo) | |
3221 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3222 | else | |
3223 | vf = 1; | |
3224 | ||
3225 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3226 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3227 | case of SLP. */ | |
437f4a00 | 3228 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
9dc3f7de IR |
3229 | ncopies = 1; |
3230 | else | |
3231 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
3232 | ||
3233 | gcc_assert (ncopies >= 1); | |
3234 | ||
3235 | /* Determine whether the shift amount is a vector, or scalar. If the | |
3236 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
3237 | ||
49eab32e JJ |
3238 | if (dt[1] == vect_internal_def && !slp_node) |
3239 | scalar_shift_arg = false; | |
3240 | else if (dt[1] == vect_constant_def | |
3241 | || dt[1] == vect_external_def | |
3242 | || dt[1] == vect_internal_def) | |
3243 | { | |
3244 | /* In SLP, need to check whether the shift count is the same, | |
3245 | in loops if it is a constant or invariant, it is always | |
3246 | a scalar shift. */ | |
3247 | if (slp_node) | |
3248 | { | |
9771b263 | 3249 | vec<gimple> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
49eab32e JJ |
3250 | gimple slpstmt; |
3251 | ||
9771b263 | 3252 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
3253 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
3254 | scalar_shift_arg = false; | |
3255 | } | |
3256 | } | |
3257 | else | |
3258 | { | |
73fbfcad | 3259 | if (dump_enabled_p ()) |
78c60e3d | 3260 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3261 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
3262 | return false; |
3263 | } | |
3264 | ||
9dc3f7de | 3265 | /* Vector shifted by vector. */ |
49eab32e | 3266 | if (!scalar_shift_arg) |
9dc3f7de IR |
3267 | { |
3268 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 3269 | if (dump_enabled_p ()) |
78c60e3d | 3270 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3271 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 3272 | |
aa948027 JJ |
3273 | if (!op1_vectype) |
3274 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
3275 | if (op1_vectype == NULL_TREE | |
3276 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 3277 | { |
73fbfcad | 3278 | if (dump_enabled_p ()) |
78c60e3d SS |
3279 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3280 | "unusable type for last operand in" | |
e645e942 | 3281 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
3282 | return false; |
3283 | } | |
9dc3f7de IR |
3284 | } |
3285 | /* See if the machine has a vector shifted by scalar insn and if not | |
3286 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 3287 | else |
9dc3f7de IR |
3288 | { |
3289 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
3290 | if (optab | |
3291 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
3292 | { | |
73fbfcad | 3293 | if (dump_enabled_p ()) |
78c60e3d | 3294 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3295 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
3296 | } |
3297 | else | |
3298 | { | |
3299 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
3300 | if (optab | |
3301 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
3302 | != CODE_FOR_nothing)) | |
3303 | { | |
49eab32e JJ |
3304 | scalar_shift_arg = false; |
3305 | ||
73fbfcad | 3306 | if (dump_enabled_p ()) |
78c60e3d | 3307 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3308 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
3309 | |
3310 | /* Unlike the other binary operators, shifts/rotates have | |
3311 | the rhs being int, instead of the same type as the lhs, | |
3312 | so make sure the scalar is the right type if we are | |
aa948027 | 3313 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
3314 | if (dt[1] == vect_constant_def) |
3315 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
3316 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
3317 | TREE_TYPE (op1))) | |
3318 | { | |
3319 | if (slp_node | |
3320 | && TYPE_MODE (TREE_TYPE (vectype)) | |
3321 | != TYPE_MODE (TREE_TYPE (op1))) | |
3322 | { | |
73fbfcad | 3323 | if (dump_enabled_p ()) |
78c60e3d SS |
3324 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3325 | "unusable type for last operand in" | |
e645e942 | 3326 | " vector/vector shift/rotate.\n"); |
aa948027 JJ |
3327 | return false; |
3328 | } | |
3329 | if (vec_stmt && !slp_node) | |
3330 | { | |
3331 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
3332 | op1 = vect_init_vector (stmt, op1, | |
3333 | TREE_TYPE (vectype), NULL); | |
3334 | } | |
3335 | } | |
9dc3f7de IR |
3336 | } |
3337 | } | |
3338 | } | |
9dc3f7de IR |
3339 | |
3340 | /* Supportable by target? */ | |
3341 | if (!optab) | |
3342 | { | |
73fbfcad | 3343 | if (dump_enabled_p ()) |
78c60e3d | 3344 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3345 | "no optab.\n"); |
9dc3f7de IR |
3346 | return false; |
3347 | } | |
3348 | vec_mode = TYPE_MODE (vectype); | |
3349 | icode = (int) optab_handler (optab, vec_mode); | |
3350 | if (icode == CODE_FOR_nothing) | |
3351 | { | |
73fbfcad | 3352 | if (dump_enabled_p ()) |
78c60e3d | 3353 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3354 | "op not supported by target.\n"); |
9dc3f7de IR |
3355 | /* Check only during analysis. */ |
3356 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
3357 | || (vf < vect_min_worthwhile_factor (code) | |
3358 | && !vec_stmt)) | |
3359 | return false; | |
73fbfcad | 3360 | if (dump_enabled_p ()) |
e645e942 TJ |
3361 | dump_printf_loc (MSG_NOTE, vect_location, |
3362 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
3363 | } |
3364 | ||
3365 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
3366 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
3367 | && vf < vect_min_worthwhile_factor (code) | |
3368 | && !vec_stmt) | |
3369 | { | |
73fbfcad | 3370 | if (dump_enabled_p ()) |
78c60e3d | 3371 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3372 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
3373 | return false; |
3374 | } | |
3375 | ||
3376 | if (!vec_stmt) /* transformation not required. */ | |
3377 | { | |
3378 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 3379 | if (dump_enabled_p ()) |
e645e942 TJ |
3380 | dump_printf_loc (MSG_NOTE, vect_location, |
3381 | "=== vectorizable_shift ===\n"); | |
c3e7ee41 | 3382 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
9dc3f7de IR |
3383 | return true; |
3384 | } | |
3385 | ||
3386 | /** Transform. **/ | |
3387 | ||
73fbfcad | 3388 | if (dump_enabled_p ()) |
78c60e3d | 3389 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3390 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
3391 | |
3392 | /* Handle def. */ | |
3393 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3394 | ||
9dc3f7de IR |
3395 | prev_stmt_info = NULL; |
3396 | for (j = 0; j < ncopies; j++) | |
3397 | { | |
3398 | /* Handle uses. */ | |
3399 | if (j == 0) | |
3400 | { | |
3401 | if (scalar_shift_arg) | |
3402 | { | |
3403 | /* Vector shl and shr insn patterns can be defined with scalar | |
3404 | operand 2 (shift operand). In this case, use constant or loop | |
3405 | invariant op1 directly, without extending it to vector mode | |
3406 | first. */ | |
3407 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
3408 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
3409 | { | |
73fbfcad | 3410 | if (dump_enabled_p ()) |
78c60e3d | 3411 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3412 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 3413 | vec_oprnd1 = op1; |
8930f723 | 3414 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 3415 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
3416 | if (slp_node) |
3417 | { | |
3418 | /* Store vec_oprnd1 for every vector stmt to be created | |
3419 | for SLP_NODE. We check during the analysis that all | |
3420 | the shift arguments are the same. | |
3421 | TODO: Allow different constants for different vector | |
3422 | stmts generated for an SLP instance. */ | |
3423 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 3424 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
3425 | } |
3426 | } | |
3427 | } | |
3428 | ||
3429 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
3430 | (a special case for certain kind of vector shifts); otherwise, | |
3431 | operand 1 should be of a vector type (the usual case). */ | |
3432 | if (vec_oprnd1) | |
3433 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
d092494c | 3434 | slp_node, -1); |
9dc3f7de IR |
3435 | else |
3436 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
d092494c | 3437 | slp_node, -1); |
9dc3f7de IR |
3438 | } |
3439 | else | |
3440 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
3441 | ||
3442 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3443 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 3444 | { |
9771b263 | 3445 | vop1 = vec_oprnds1[i]; |
9dc3f7de IR |
3446 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); |
3447 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3448 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3449 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3450 | if (slp_node) | |
9771b263 | 3451 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
3452 | } |
3453 | ||
3454 | if (slp_node) | |
3455 | continue; | |
3456 | ||
3457 | if (j == 0) | |
3458 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3459 | else | |
3460 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3461 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3462 | } | |
3463 | ||
9771b263 DN |
3464 | vec_oprnds0.release (); |
3465 | vec_oprnds1.release (); | |
9dc3f7de IR |
3466 | |
3467 | return true; | |
3468 | } | |
3469 | ||
3470 | ||
5deb57cb JJ |
3471 | static tree permute_vec_elements (tree, tree, tree, gimple, |
3472 | gimple_stmt_iterator *); | |
3473 | ||
3474 | ||
ebfd146a IR |
3475 | /* Function vectorizable_operation. |
3476 | ||
16949072 RG |
3477 | Check if STMT performs a binary, unary or ternary operation that can |
3478 | be vectorized. | |
b8698a0f | 3479 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
3480 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3481 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3482 | ||
3483 | static bool | |
3484 | vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi, | |
3485 | gimple *vec_stmt, slp_tree slp_node) | |
3486 | { | |
00f07b86 | 3487 | tree vec_dest; |
ebfd146a | 3488 | tree scalar_dest; |
16949072 | 3489 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 3490 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 3491 | tree vectype; |
ebfd146a IR |
3492 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
3493 | enum tree_code code; | |
3494 | enum machine_mode vec_mode; | |
3495 | tree new_temp; | |
3496 | int op_type; | |
00f07b86 | 3497 | optab optab; |
ebfd146a | 3498 | int icode; |
ebfd146a IR |
3499 | tree def; |
3500 | gimple def_stmt; | |
16949072 RG |
3501 | enum vect_def_type dt[3] |
3502 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
ebfd146a IR |
3503 | gimple new_stmt = NULL; |
3504 | stmt_vec_info prev_stmt_info; | |
b690cc0f | 3505 | int nunits_in; |
ebfd146a IR |
3506 | int nunits_out; |
3507 | tree vectype_out; | |
3508 | int ncopies; | |
3509 | int j, i; | |
6e1aa848 DN |
3510 | vec<tree> vec_oprnds0 = vNULL; |
3511 | vec<tree> vec_oprnds1 = vNULL; | |
3512 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 3513 | tree vop0, vop1, vop2; |
a70d6342 IR |
3514 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3515 | int vf; | |
3516 | ||
a70d6342 | 3517 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3518 | return false; |
3519 | ||
8644a673 | 3520 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3521 | return false; |
3522 | ||
3523 | /* Is STMT a vectorizable binary/unary operation? */ | |
3524 | if (!is_gimple_assign (stmt)) | |
3525 | return false; | |
3526 | ||
3527 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3528 | return false; | |
3529 | ||
ebfd146a IR |
3530 | code = gimple_assign_rhs_code (stmt); |
3531 | ||
3532 | /* For pointer addition, we should use the normal plus for | |
3533 | the vector addition. */ | |
3534 | if (code == POINTER_PLUS_EXPR) | |
3535 | code = PLUS_EXPR; | |
3536 | ||
3537 | /* Support only unary or binary operations. */ | |
3538 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 3539 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 3540 | { |
73fbfcad | 3541 | if (dump_enabled_p ()) |
78c60e3d | 3542 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3543 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 3544 | op_type); |
ebfd146a IR |
3545 | return false; |
3546 | } | |
3547 | ||
b690cc0f RG |
3548 | scalar_dest = gimple_assign_lhs (stmt); |
3549 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
3550 | ||
7b7b1813 RG |
3551 | /* Most operations cannot handle bit-precision types without extra |
3552 | truncations. */ | |
3553 | if ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
3554 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
3555 | /* Exception are bitwise binary operations. */ | |
3556 | && code != BIT_IOR_EXPR | |
3557 | && code != BIT_XOR_EXPR | |
3558 | && code != BIT_AND_EXPR) | |
3559 | { | |
73fbfcad | 3560 | if (dump_enabled_p ()) |
78c60e3d | 3561 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3562 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
3563 | return false; |
3564 | } | |
3565 | ||
ebfd146a | 3566 | op0 = gimple_assign_rhs1 (stmt); |
24ee1384 | 3567 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
b690cc0f | 3568 | &def_stmt, &def, &dt[0], &vectype)) |
ebfd146a | 3569 | { |
73fbfcad | 3570 | if (dump_enabled_p ()) |
78c60e3d | 3571 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3572 | "use not simple.\n"); |
ebfd146a IR |
3573 | return false; |
3574 | } | |
b690cc0f RG |
3575 | /* If op0 is an external or constant def use a vector type with |
3576 | the same size as the output vector type. */ | |
3577 | if (!vectype) | |
3578 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
3579 | if (vec_stmt) |
3580 | gcc_assert (vectype); | |
3581 | if (!vectype) | |
3582 | { | |
73fbfcad | 3583 | if (dump_enabled_p ()) |
7d8930a0 | 3584 | { |
78c60e3d SS |
3585 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3586 | "no vectype for scalar type "); | |
3587 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
3588 | TREE_TYPE (op0)); | |
e645e942 | 3589 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
3590 | } |
3591 | ||
3592 | return false; | |
3593 | } | |
b690cc0f RG |
3594 | |
3595 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
3596 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
3597 | if (nunits_out != nunits_in) | |
3598 | return false; | |
ebfd146a | 3599 | |
16949072 | 3600 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
3601 | { |
3602 | op1 = gimple_assign_rhs2 (stmt); | |
24ee1384 IR |
3603 | if (!vect_is_simple_use (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3604 | &def, &dt[1])) | |
ebfd146a | 3605 | { |
73fbfcad | 3606 | if (dump_enabled_p ()) |
78c60e3d | 3607 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3608 | "use not simple.\n"); |
ebfd146a IR |
3609 | return false; |
3610 | } | |
3611 | } | |
16949072 RG |
3612 | if (op_type == ternary_op) |
3613 | { | |
3614 | op2 = gimple_assign_rhs3 (stmt); | |
24ee1384 IR |
3615 | if (!vect_is_simple_use (op2, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3616 | &def, &dt[2])) | |
16949072 | 3617 | { |
73fbfcad | 3618 | if (dump_enabled_p ()) |
78c60e3d | 3619 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3620 | "use not simple.\n"); |
16949072 RG |
3621 | return false; |
3622 | } | |
3623 | } | |
ebfd146a | 3624 | |
b690cc0f RG |
3625 | if (loop_vinfo) |
3626 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3627 | else | |
3628 | vf = 1; | |
3629 | ||
3630 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 3631 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 3632 | case of SLP. */ |
437f4a00 | 3633 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
b690cc0f RG |
3634 | ncopies = 1; |
3635 | else | |
3636 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
3637 | ||
3638 | gcc_assert (ncopies >= 1); | |
3639 | ||
9dc3f7de | 3640 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
3641 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
3642 | || code == RROTATE_EXPR) | |
9dc3f7de | 3643 | return false; |
ebfd146a | 3644 | |
ebfd146a | 3645 | /* Supportable by target? */ |
00f07b86 RH |
3646 | |
3647 | vec_mode = TYPE_MODE (vectype); | |
3648 | if (code == MULT_HIGHPART_EXPR) | |
ebfd146a | 3649 | { |
00f07b86 | 3650 | if (can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype))) |
dee54b6e | 3651 | icode = LAST_INSN_CODE; |
00f07b86 RH |
3652 | else |
3653 | icode = CODE_FOR_nothing; | |
ebfd146a | 3654 | } |
00f07b86 RH |
3655 | else |
3656 | { | |
3657 | optab = optab_for_tree_code (code, vectype, optab_default); | |
3658 | if (!optab) | |
5deb57cb | 3659 | { |
73fbfcad | 3660 | if (dump_enabled_p ()) |
78c60e3d | 3661 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3662 | "no optab.\n"); |
00f07b86 | 3663 | return false; |
5deb57cb | 3664 | } |
00f07b86 | 3665 | icode = (int) optab_handler (optab, vec_mode); |
5deb57cb JJ |
3666 | } |
3667 | ||
ebfd146a IR |
3668 | if (icode == CODE_FOR_nothing) |
3669 | { | |
73fbfcad | 3670 | if (dump_enabled_p ()) |
78c60e3d | 3671 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3672 | "op not supported by target.\n"); |
ebfd146a IR |
3673 | /* Check only during analysis. */ |
3674 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5deb57cb | 3675 | || (!vec_stmt && vf < vect_min_worthwhile_factor (code))) |
ebfd146a | 3676 | return false; |
73fbfcad | 3677 | if (dump_enabled_p ()) |
e645e942 TJ |
3678 | dump_printf_loc (MSG_NOTE, vect_location, |
3679 | "proceeding using word mode.\n"); | |
383d9c83 IR |
3680 | } |
3681 | ||
4a00c761 | 3682 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
3683 | if (!VECTOR_MODE_P (vec_mode) |
3684 | && !vec_stmt | |
3685 | && vf < vect_min_worthwhile_factor (code)) | |
7d8930a0 | 3686 | { |
73fbfcad | 3687 | if (dump_enabled_p ()) |
78c60e3d | 3688 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3689 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 3690 | return false; |
7d8930a0 | 3691 | } |
ebfd146a | 3692 | |
ebfd146a IR |
3693 | if (!vec_stmt) /* transformation not required. */ |
3694 | { | |
4a00c761 | 3695 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 3696 | if (dump_enabled_p ()) |
78c60e3d | 3697 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3698 | "=== vectorizable_operation ===\n"); |
c3e7ee41 | 3699 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
3700 | return true; |
3701 | } | |
3702 | ||
3703 | /** Transform. **/ | |
3704 | ||
73fbfcad | 3705 | if (dump_enabled_p ()) |
78c60e3d | 3706 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3707 | "transform binary/unary operation.\n"); |
383d9c83 | 3708 | |
ebfd146a | 3709 | /* Handle def. */ |
00f07b86 | 3710 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 3711 | |
ebfd146a IR |
3712 | /* In case the vectorization factor (VF) is bigger than the number |
3713 | of elements that we can fit in a vectype (nunits), we have to generate | |
3714 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
3715 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
3716 | from one copy of the vector stmt to the next, in the field | |
3717 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
3718 | stages to find the correct vector defs to be used when vectorizing | |
3719 | stmts that use the defs of the current stmt. The example below | |
3720 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
3721 | we need to create 4 vectorized stmts): | |
3722 | ||
3723 | before vectorization: | |
3724 | RELATED_STMT VEC_STMT | |
3725 | S1: x = memref - - | |
3726 | S2: z = x + 1 - - | |
3727 | ||
3728 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
3729 | there): | |
3730 | RELATED_STMT VEC_STMT | |
3731 | VS1_0: vx0 = memref0 VS1_1 - | |
3732 | VS1_1: vx1 = memref1 VS1_2 - | |
3733 | VS1_2: vx2 = memref2 VS1_3 - | |
3734 | VS1_3: vx3 = memref3 - - | |
3735 | S1: x = load - VS1_0 | |
3736 | S2: z = x + 1 - - | |
3737 | ||
3738 | step2: vectorize stmt S2 (done here): | |
3739 | To vectorize stmt S2 we first need to find the relevant vector | |
3740 | def for the first operand 'x'. This is, as usual, obtained from | |
3741 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
3742 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
3743 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
3744 | the vector stmt VS2_0, and as usual, record it in the | |
3745 | STMT_VINFO_VEC_STMT of stmt S2. | |
3746 | When creating the second copy (VS2_1), we obtain the relevant vector | |
3747 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
3748 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
3749 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
3750 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
3751 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
3752 | chain of stmts and pointers: | |
3753 | RELATED_STMT VEC_STMT | |
3754 | VS1_0: vx0 = memref0 VS1_1 - | |
3755 | VS1_1: vx1 = memref1 VS1_2 - | |
3756 | VS1_2: vx2 = memref2 VS1_3 - | |
3757 | VS1_3: vx3 = memref3 - - | |
3758 | S1: x = load - VS1_0 | |
3759 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
3760 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
3761 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
3762 | VS2_3: vz3 = vx3 + v1 - - | |
3763 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
3764 | |
3765 | prev_stmt_info = NULL; | |
3766 | for (j = 0; j < ncopies; j++) | |
3767 | { | |
3768 | /* Handle uses. */ | |
3769 | if (j == 0) | |
4a00c761 JJ |
3770 | { |
3771 | if (op_type == binary_op || op_type == ternary_op) | |
3772 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
3773 | slp_node, -1); | |
3774 | else | |
3775 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
3776 | slp_node, -1); | |
3777 | if (op_type == ternary_op) | |
36ba4aae | 3778 | { |
9771b263 DN |
3779 | vec_oprnds2.create (1); |
3780 | vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2, | |
3781 | stmt, | |
3782 | NULL)); | |
36ba4aae | 3783 | } |
4a00c761 | 3784 | } |
ebfd146a | 3785 | else |
4a00c761 JJ |
3786 | { |
3787 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
3788 | if (op_type == ternary_op) | |
3789 | { | |
9771b263 DN |
3790 | tree vec_oprnd = vec_oprnds2.pop (); |
3791 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
3792 | vec_oprnd)); | |
4a00c761 JJ |
3793 | } |
3794 | } | |
3795 | ||
3796 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3797 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 3798 | { |
4a00c761 | 3799 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 3800 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 3801 | vop2 = ((op_type == ternary_op) |
9771b263 | 3802 | ? vec_oprnds2[i] : NULL_TREE); |
73804b12 RG |
3803 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, |
3804 | vop0, vop1, vop2); | |
4a00c761 JJ |
3805 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3806 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3807 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3808 | if (slp_node) | |
9771b263 | 3809 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
3810 | } |
3811 | ||
4a00c761 JJ |
3812 | if (slp_node) |
3813 | continue; | |
3814 | ||
3815 | if (j == 0) | |
3816 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3817 | else | |
3818 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3819 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
3820 | } |
3821 | ||
9771b263 DN |
3822 | vec_oprnds0.release (); |
3823 | vec_oprnds1.release (); | |
3824 | vec_oprnds2.release (); | |
ebfd146a | 3825 | |
ebfd146a IR |
3826 | return true; |
3827 | } | |
3828 | ||
c716e67f XDL |
3829 | /* A helper function to ensure data reference DR's base alignment |
3830 | for STMT_INFO. */ | |
3831 | ||
3832 | static void | |
3833 | ensure_base_align (stmt_vec_info stmt_info, struct data_reference *dr) | |
3834 | { | |
3835 | if (!dr->aux) | |
3836 | return; | |
3837 | ||
3838 | if (((dataref_aux *)dr->aux)->base_misaligned) | |
3839 | { | |
3840 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3841 | tree base_decl = ((dataref_aux *)dr->aux)->base_decl; | |
3842 | ||
3843 | DECL_ALIGN (base_decl) = TYPE_ALIGN (vectype); | |
3844 | DECL_USER_ALIGN (base_decl) = 1; | |
3845 | ((dataref_aux *)dr->aux)->base_misaligned = false; | |
3846 | } | |
3847 | } | |
3848 | ||
ebfd146a IR |
3849 | |
3850 | /* Function vectorizable_store. | |
3851 | ||
b8698a0f L |
3852 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
3853 | can be vectorized. | |
3854 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
3855 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3856 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3857 | ||
3858 | static bool | |
3859 | vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
c716e67f | 3860 | slp_tree slp_node) |
ebfd146a IR |
3861 | { |
3862 | tree scalar_dest; | |
3863 | tree data_ref; | |
3864 | tree op; | |
3865 | tree vec_oprnd = NULL_TREE; | |
3866 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3867 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
3868 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
272c6793 | 3869 | tree elem_type; |
ebfd146a | 3870 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 3871 | struct loop *loop = NULL; |
ebfd146a IR |
3872 | enum machine_mode vec_mode; |
3873 | tree dummy; | |
3874 | enum dr_alignment_support alignment_support_scheme; | |
3875 | tree def; | |
3876 | gimple def_stmt; | |
3877 | enum vect_def_type dt; | |
3878 | stmt_vec_info prev_stmt_info = NULL; | |
3879 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 3880 | tree dataref_offset = NULL_TREE; |
fef4d2b3 | 3881 | gimple ptr_incr = NULL; |
ebfd146a IR |
3882 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
3883 | int ncopies; | |
3884 | int j; | |
3885 | gimple next_stmt, first_stmt = NULL; | |
0d0293ac | 3886 | bool grouped_store = false; |
272c6793 | 3887 | bool store_lanes_p = false; |
ebfd146a | 3888 | unsigned int group_size, i; |
6e1aa848 DN |
3889 | vec<tree> dr_chain = vNULL; |
3890 | vec<tree> oprnds = vNULL; | |
3891 | vec<tree> result_chain = vNULL; | |
ebfd146a | 3892 | bool inv_p; |
6e1aa848 | 3893 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 3894 | bool slp = (slp_node != NULL); |
ebfd146a | 3895 | unsigned int vec_num; |
a70d6342 | 3896 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
272c6793 | 3897 | tree aggr_type; |
a70d6342 IR |
3898 | |
3899 | if (loop_vinfo) | |
3900 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
3901 | |
3902 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3903 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3904 | case of SLP. */ | |
437f4a00 | 3905 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
3906 | ncopies = 1; |
3907 | else | |
3908 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3909 | ||
3910 | gcc_assert (ncopies >= 1); | |
3911 | ||
3912 | /* FORNOW. This restriction should be relaxed. */ | |
a70d6342 | 3913 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) |
ebfd146a | 3914 | { |
73fbfcad | 3915 | if (dump_enabled_p ()) |
78c60e3d | 3916 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3917 | "multiple types in nested loop.\n"); |
ebfd146a IR |
3918 | return false; |
3919 | } | |
3920 | ||
a70d6342 | 3921 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3922 | return false; |
3923 | ||
8644a673 | 3924 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3925 | return false; |
3926 | ||
3927 | /* Is vectorizable store? */ | |
3928 | ||
3929 | if (!is_gimple_assign (stmt)) | |
3930 | return false; | |
3931 | ||
3932 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
3933 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
3934 | && is_pattern_stmt_p (stmt_info)) | |
3935 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 3936 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 3937 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 3938 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
3939 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
3940 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
3941 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
3942 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
3943 | return false; |
3944 | ||
3945 | gcc_assert (gimple_assign_single_p (stmt)); | |
3946 | op = gimple_assign_rhs1 (stmt); | |
24ee1384 IR |
3947 | if (!vect_is_simple_use (op, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3948 | &def, &dt)) | |
ebfd146a | 3949 | { |
73fbfcad | 3950 | if (dump_enabled_p ()) |
78c60e3d | 3951 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3952 | "use not simple.\n"); |
ebfd146a IR |
3953 | return false; |
3954 | } | |
3955 | ||
272c6793 | 3956 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 3957 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 3958 | |
ebfd146a IR |
3959 | /* FORNOW. In some cases can vectorize even if data-type not supported |
3960 | (e.g. - array initialization with 0). */ | |
947131ba | 3961 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
3962 | return false; |
3963 | ||
3964 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
3965 | return false; | |
3966 | ||
a7ce6ec3 RG |
3967 | if (tree_int_cst_compare (loop && nested_in_vect_loop_p (loop, stmt) |
3968 | ? STMT_VINFO_DR_STEP (stmt_info) : DR_STEP (dr), | |
3969 | size_zero_node) < 0) | |
a1e53f3f | 3970 | { |
73fbfcad | 3971 | if (dump_enabled_p ()) |
78c60e3d | 3972 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3973 | "negative step for store.\n"); |
a1e53f3f L |
3974 | return false; |
3975 | } | |
3976 | ||
0d0293ac | 3977 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 3978 | { |
0d0293ac | 3979 | grouped_store = true; |
e14c1050 | 3980 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
3981 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
3982 | { | |
e14c1050 | 3983 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
3984 | if (vect_store_lanes_supported (vectype, group_size)) |
3985 | store_lanes_p = true; | |
0d0293ac | 3986 | else if (!vect_grouped_store_supported (vectype, group_size)) |
b602d918 RS |
3987 | return false; |
3988 | } | |
b8698a0f | 3989 | |
ebfd146a IR |
3990 | if (first_stmt == stmt) |
3991 | { | |
3992 | /* STMT is the leader of the group. Check the operands of all the | |
3993 | stmts of the group. */ | |
e14c1050 | 3994 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
3995 | while (next_stmt) |
3996 | { | |
3997 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
3998 | op = gimple_assign_rhs1 (next_stmt); | |
24ee1384 IR |
3999 | if (!vect_is_simple_use (op, next_stmt, loop_vinfo, bb_vinfo, |
4000 | &def_stmt, &def, &dt)) | |
ebfd146a | 4001 | { |
73fbfcad | 4002 | if (dump_enabled_p ()) |
78c60e3d | 4003 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4004 | "use not simple.\n"); |
ebfd146a IR |
4005 | return false; |
4006 | } | |
e14c1050 | 4007 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
4008 | } |
4009 | } | |
4010 | } | |
4011 | ||
4012 | if (!vec_stmt) /* transformation not required. */ | |
4013 | { | |
4014 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
92345349 BS |
4015 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, |
4016 | NULL, NULL, NULL); | |
ebfd146a IR |
4017 | return true; |
4018 | } | |
4019 | ||
4020 | /** Transform. **/ | |
4021 | ||
c716e67f XDL |
4022 | ensure_base_align (stmt_info, dr); |
4023 | ||
0d0293ac | 4024 | if (grouped_store) |
ebfd146a IR |
4025 | { |
4026 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 4027 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 4028 | |
e14c1050 | 4029 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
4030 | |
4031 | /* FORNOW */ | |
a70d6342 | 4032 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
4033 | |
4034 | /* We vectorize all the stmts of the interleaving group when we | |
4035 | reach the last stmt in the group. */ | |
e14c1050 IR |
4036 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
4037 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
4038 | && !slp) |
4039 | { | |
4040 | *vec_stmt = NULL; | |
4041 | return true; | |
4042 | } | |
4043 | ||
4044 | if (slp) | |
4b5caab7 | 4045 | { |
0d0293ac | 4046 | grouped_store = false; |
4b5caab7 IR |
4047 | /* VEC_NUM is the number of vect stmts to be created for this |
4048 | group. */ | |
4049 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 4050 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
4b5caab7 | 4051 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 4052 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 4053 | } |
ebfd146a | 4054 | else |
4b5caab7 IR |
4055 | /* VEC_NUM is the number of vect stmts to be created for this |
4056 | group. */ | |
ebfd146a IR |
4057 | vec_num = group_size; |
4058 | } | |
b8698a0f | 4059 | else |
ebfd146a IR |
4060 | { |
4061 | first_stmt = stmt; | |
4062 | first_dr = dr; | |
4063 | group_size = vec_num = 1; | |
ebfd146a | 4064 | } |
b8698a0f | 4065 | |
73fbfcad | 4066 | if (dump_enabled_p ()) |
78c60e3d | 4067 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4068 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 4069 | |
9771b263 DN |
4070 | dr_chain.create (group_size); |
4071 | oprnds.create (group_size); | |
ebfd146a | 4072 | |
720f5239 | 4073 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 4074 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
4075 | /* Targets with store-lane instructions must not require explicit |
4076 | realignment. */ | |
4077 | gcc_assert (!store_lanes_p | |
4078 | || alignment_support_scheme == dr_aligned | |
4079 | || alignment_support_scheme == dr_unaligned_supported); | |
4080 | ||
4081 | if (store_lanes_p) | |
4082 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
4083 | else | |
4084 | aggr_type = vectype; | |
ebfd146a IR |
4085 | |
4086 | /* In case the vectorization factor (VF) is bigger than the number | |
4087 | of elements that we can fit in a vectype (nunits), we have to generate | |
4088 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 4089 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
4090 | vect_get_vec_def_for_copy_stmt. */ |
4091 | ||
0d0293ac | 4092 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
4093 | |
4094 | S1: &base + 2 = x2 | |
4095 | S2: &base = x0 | |
4096 | S3: &base + 1 = x1 | |
4097 | S4: &base + 3 = x3 | |
4098 | ||
4099 | We create vectorized stores starting from base address (the access of the | |
4100 | first stmt in the chain (S2 in the above example), when the last store stmt | |
4101 | of the chain (S4) is reached: | |
4102 | ||
4103 | VS1: &base = vx2 | |
4104 | VS2: &base + vec_size*1 = vx0 | |
4105 | VS3: &base + vec_size*2 = vx1 | |
4106 | VS4: &base + vec_size*3 = vx3 | |
4107 | ||
4108 | Then permutation statements are generated: | |
4109 | ||
3fcc1b55 JJ |
4110 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
4111 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 4112 | ... |
b8698a0f | 4113 | |
ebfd146a IR |
4114 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
4115 | (the order of the data-refs in the output of vect_permute_store_chain | |
4116 | corresponds to the order of scalar stmts in the interleaving chain - see | |
4117 | the documentation of vect_permute_store_chain()). | |
4118 | ||
4119 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 4120 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 4121 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 4122 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
4123 | */ |
4124 | ||
4125 | prev_stmt_info = NULL; | |
4126 | for (j = 0; j < ncopies; j++) | |
4127 | { | |
4128 | gimple new_stmt; | |
ebfd146a IR |
4129 | |
4130 | if (j == 0) | |
4131 | { | |
4132 | if (slp) | |
4133 | { | |
4134 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c IR |
4135 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
4136 | NULL, slp_node, -1); | |
ebfd146a | 4137 | |
9771b263 | 4138 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
4139 | } |
4140 | else | |
4141 | { | |
b8698a0f L |
4142 | /* For interleaved stores we collect vectorized defs for all the |
4143 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
4144 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
4145 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
4146 | ||
0d0293ac | 4147 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 4148 | OPRNDS are of size 1. */ |
b8698a0f | 4149 | next_stmt = first_stmt; |
ebfd146a IR |
4150 | for (i = 0; i < group_size; i++) |
4151 | { | |
b8698a0f L |
4152 | /* Since gaps are not supported for interleaved stores, |
4153 | GROUP_SIZE is the exact number of stmts in the chain. | |
4154 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
4155 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
4156 | iteration of the loop will be executed. */ |
4157 | gcc_assert (next_stmt | |
4158 | && gimple_assign_single_p (next_stmt)); | |
4159 | op = gimple_assign_rhs1 (next_stmt); | |
4160 | ||
b8698a0f | 4161 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt, |
ebfd146a | 4162 | NULL); |
9771b263 DN |
4163 | dr_chain.quick_push (vec_oprnd); |
4164 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 4165 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
4166 | } |
4167 | } | |
4168 | ||
4169 | /* We should have catched mismatched types earlier. */ | |
4170 | gcc_assert (useless_type_conversion_p (vectype, | |
4171 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
4172 | bool simd_lane_access_p |
4173 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
4174 | if (simd_lane_access_p | |
4175 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
4176 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
4177 | && integer_zerop (DR_OFFSET (first_dr)) | |
4178 | && integer_zerop (DR_INIT (first_dr)) | |
4179 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
4180 | get_alias_set (DR_REF (first_dr)))) | |
4181 | { | |
4182 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
4183 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
4184 | (DR_REF (first_dr)), 0); | |
8928eff3 | 4185 | inv_p = false; |
74bf76ed JJ |
4186 | } |
4187 | else | |
4188 | dataref_ptr | |
4189 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
4190 | simd_lane_access_p ? loop : NULL, | |
4191 | NULL_TREE, &dummy, gsi, &ptr_incr, | |
4192 | simd_lane_access_p, &inv_p); | |
a70d6342 | 4193 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 4194 | } |
b8698a0f | 4195 | else |
ebfd146a | 4196 | { |
b8698a0f L |
4197 | /* For interleaved stores we created vectorized defs for all the |
4198 | defs stored in OPRNDS in the previous iteration (previous copy). | |
4199 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
4200 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
4201 | next copy. | |
0d0293ac | 4202 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
4203 | OPRNDS are of size 1. */ |
4204 | for (i = 0; i < group_size; i++) | |
4205 | { | |
9771b263 | 4206 | op = oprnds[i]; |
24ee1384 IR |
4207 | vect_is_simple_use (op, NULL, loop_vinfo, bb_vinfo, &def_stmt, |
4208 | &def, &dt); | |
b8698a0f | 4209 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
4210 | dr_chain[i] = vec_oprnd; |
4211 | oprnds[i] = vec_oprnd; | |
ebfd146a | 4212 | } |
74bf76ed JJ |
4213 | if (dataref_offset) |
4214 | dataref_offset | |
4215 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
4216 | TYPE_SIZE_UNIT (aggr_type)); | |
4217 | else | |
4218 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
4219 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
4220 | } |
4221 | ||
272c6793 | 4222 | if (store_lanes_p) |
ebfd146a | 4223 | { |
272c6793 | 4224 | tree vec_array; |
267d3070 | 4225 | |
272c6793 RS |
4226 | /* Combine all the vectors into an array. */ |
4227 | vec_array = create_vector_array (vectype, vec_num); | |
4228 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 4229 | { |
9771b263 | 4230 | vec_oprnd = dr_chain[i]; |
272c6793 | 4231 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 4232 | } |
b8698a0f | 4233 | |
272c6793 RS |
4234 | /* Emit: |
4235 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
4236 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
4237 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
4238 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 | 4239 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
4240 | } |
4241 | else | |
4242 | { | |
4243 | new_stmt = NULL; | |
0d0293ac | 4244 | if (grouped_store) |
272c6793 | 4245 | { |
b6b9227d JJ |
4246 | if (j == 0) |
4247 | result_chain.create (group_size); | |
272c6793 RS |
4248 | /* Permute. */ |
4249 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
4250 | &result_chain); | |
4251 | } | |
c2d7ab2a | 4252 | |
272c6793 RS |
4253 | next_stmt = first_stmt; |
4254 | for (i = 0; i < vec_num; i++) | |
4255 | { | |
644ffefd | 4256 | unsigned align, misalign; |
272c6793 RS |
4257 | |
4258 | if (i > 0) | |
4259 | /* Bump the vector pointer. */ | |
4260 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
4261 | stmt, NULL_TREE); | |
4262 | ||
4263 | if (slp) | |
9771b263 | 4264 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
4265 | else if (grouped_store) |
4266 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 4267 | vect_permute_store_chain(). */ |
9771b263 | 4268 | vec_oprnd = result_chain[i]; |
272c6793 RS |
4269 | |
4270 | data_ref = build2 (MEM_REF, TREE_TYPE (vec_oprnd), dataref_ptr, | |
74bf76ed JJ |
4271 | dataref_offset |
4272 | ? dataref_offset | |
4273 | : build_int_cst (reference_alias_ptr_type | |
4274 | (DR_REF (first_dr)), 0)); | |
644ffefd | 4275 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 | 4276 | if (aligned_access_p (first_dr)) |
644ffefd | 4277 | misalign = 0; |
272c6793 RS |
4278 | else if (DR_MISALIGNMENT (first_dr) == -1) |
4279 | { | |
4280 | TREE_TYPE (data_ref) | |
4281 | = build_aligned_type (TREE_TYPE (data_ref), | |
4282 | TYPE_ALIGN (elem_type)); | |
644ffefd MJ |
4283 | align = TYPE_ALIGN_UNIT (elem_type); |
4284 | misalign = 0; | |
272c6793 RS |
4285 | } |
4286 | else | |
4287 | { | |
4288 | TREE_TYPE (data_ref) | |
4289 | = build_aligned_type (TREE_TYPE (data_ref), | |
4290 | TYPE_ALIGN (elem_type)); | |
644ffefd | 4291 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 4292 | } |
74bf76ed JJ |
4293 | if (dataref_offset == NULL_TREE) |
4294 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
4295 | misalign); | |
c2d7ab2a | 4296 | |
272c6793 RS |
4297 | /* Arguments are ready. Create the new vector stmt. */ |
4298 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
4299 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
4300 | |
4301 | if (slp) | |
4302 | continue; | |
4303 | ||
e14c1050 | 4304 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
4305 | if (!next_stmt) |
4306 | break; | |
4307 | } | |
ebfd146a | 4308 | } |
1da0876c RS |
4309 | if (!slp) |
4310 | { | |
4311 | if (j == 0) | |
4312 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4313 | else | |
4314 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4315 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4316 | } | |
ebfd146a IR |
4317 | } |
4318 | ||
9771b263 DN |
4319 | dr_chain.release (); |
4320 | oprnds.release (); | |
4321 | result_chain.release (); | |
4322 | vec_oprnds.release (); | |
ebfd146a IR |
4323 | |
4324 | return true; | |
4325 | } | |
4326 | ||
aec7ae7d JJ |
4327 | /* Given a vector type VECTYPE and permutation SEL returns |
4328 | the VECTOR_CST mask that implements the permutation of the | |
4329 | vector elements. If that is impossible to do, returns NULL. */ | |
a1e53f3f | 4330 | |
3fcc1b55 JJ |
4331 | tree |
4332 | vect_gen_perm_mask (tree vectype, unsigned char *sel) | |
a1e53f3f | 4333 | { |
d2a12ae7 | 4334 | tree mask_elt_type, mask_type, mask_vec, *mask_elts; |
2635892a | 4335 | int i, nunits; |
a1e53f3f | 4336 | |
22e4dee7 | 4337 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 RH |
4338 | |
4339 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) | |
a1e53f3f L |
4340 | return NULL; |
4341 | ||
96f9265a RG |
4342 | mask_elt_type = lang_hooks.types.type_for_mode |
4343 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
22e4dee7 | 4344 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 4345 | |
d2a12ae7 | 4346 | mask_elts = XALLOCAVEC (tree, nunits); |
aec7ae7d | 4347 | for (i = nunits - 1; i >= 0; i--) |
d2a12ae7 RG |
4348 | mask_elts[i] = build_int_cst (mask_elt_type, sel[i]); |
4349 | mask_vec = build_vector (mask_type, mask_elts); | |
a1e53f3f | 4350 | |
2635892a | 4351 | return mask_vec; |
a1e53f3f L |
4352 | } |
4353 | ||
aec7ae7d JJ |
4354 | /* Given a vector type VECTYPE returns the VECTOR_CST mask that implements |
4355 | reversal of the vector elements. If that is impossible to do, | |
4356 | returns NULL. */ | |
4357 | ||
4358 | static tree | |
4359 | perm_mask_for_reverse (tree vectype) | |
4360 | { | |
4361 | int i, nunits; | |
4362 | unsigned char *sel; | |
4363 | ||
4364 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4365 | sel = XALLOCAVEC (unsigned char, nunits); | |
4366 | ||
4367 | for (i = 0; i < nunits; ++i) | |
4368 | sel[i] = nunits - 1 - i; | |
4369 | ||
3fcc1b55 | 4370 | return vect_gen_perm_mask (vectype, sel); |
aec7ae7d JJ |
4371 | } |
4372 | ||
4373 | /* Given a vector variable X and Y, that was generated for the scalar | |
4374 | STMT, generate instructions to permute the vector elements of X and Y | |
4375 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
4376 | permuted vector variable. */ | |
a1e53f3f L |
4377 | |
4378 | static tree | |
aec7ae7d JJ |
4379 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple stmt, |
4380 | gimple_stmt_iterator *gsi) | |
a1e53f3f L |
4381 | { |
4382 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 4383 | tree perm_dest, data_ref; |
a1e53f3f L |
4384 | gimple perm_stmt; |
4385 | ||
a1e53f3f | 4386 | perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype); |
aec7ae7d | 4387 | data_ref = make_ssa_name (perm_dest, NULL); |
a1e53f3f L |
4388 | |
4389 | /* Generate the permute statement. */ | |
73804b12 RG |
4390 | perm_stmt = gimple_build_assign_with_ops (VEC_PERM_EXPR, data_ref, |
4391 | x, y, mask_vec); | |
a1e53f3f L |
4392 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
4393 | ||
4394 | return data_ref; | |
4395 | } | |
4396 | ||
ebfd146a IR |
4397 | /* vectorizable_load. |
4398 | ||
b8698a0f L |
4399 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
4400 | can be vectorized. | |
4401 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4402 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4403 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4404 | ||
4405 | static bool | |
4406 | vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
c716e67f | 4407 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
4408 | { |
4409 | tree scalar_dest; | |
4410 | tree vec_dest = NULL; | |
4411 | tree data_ref = NULL; | |
4412 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 4413 | stmt_vec_info prev_stmt_info; |
ebfd146a | 4414 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 4415 | struct loop *loop = NULL; |
ebfd146a | 4416 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 4417 | bool nested_in_vect_loop = false; |
c716e67f | 4418 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
ebfd146a | 4419 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
272c6793 | 4420 | tree elem_type; |
ebfd146a | 4421 | tree new_temp; |
947131ba | 4422 | enum machine_mode mode; |
ebfd146a IR |
4423 | gimple new_stmt = NULL; |
4424 | tree dummy; | |
4425 | enum dr_alignment_support alignment_support_scheme; | |
4426 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 4427 | tree dataref_offset = NULL_TREE; |
fef4d2b3 | 4428 | gimple ptr_incr = NULL; |
ebfd146a IR |
4429 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
4430 | int ncopies; | |
a64b9c26 | 4431 | int i, j, group_size, group_gap; |
ebfd146a IR |
4432 | tree msq = NULL_TREE, lsq; |
4433 | tree offset = NULL_TREE; | |
4434 | tree realignment_token = NULL_TREE; | |
4435 | gimple phi = NULL; | |
6e1aa848 | 4436 | vec<tree> dr_chain = vNULL; |
0d0293ac | 4437 | bool grouped_load = false; |
272c6793 | 4438 | bool load_lanes_p = false; |
ebfd146a | 4439 | gimple first_stmt; |
ebfd146a | 4440 | bool inv_p; |
319e6439 | 4441 | bool negative = false; |
ebfd146a IR |
4442 | bool compute_in_loop = false; |
4443 | struct loop *at_loop; | |
4444 | int vec_num; | |
4445 | bool slp = (slp_node != NULL); | |
4446 | bool slp_perm = false; | |
4447 | enum tree_code code; | |
a70d6342 IR |
4448 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
4449 | int vf; | |
272c6793 | 4450 | tree aggr_type; |
aec7ae7d JJ |
4451 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; |
4452 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
4453 | int gather_scale = 1; | |
4454 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
a70d6342 IR |
4455 | |
4456 | if (loop_vinfo) | |
4457 | { | |
4458 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
4459 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
4460 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4461 | } | |
4462 | else | |
3533e503 | 4463 | vf = 1; |
ebfd146a IR |
4464 | |
4465 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4466 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 4467 | case of SLP. */ |
437f4a00 | 4468 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
4469 | ncopies = 1; |
4470 | else | |
4471 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4472 | ||
4473 | gcc_assert (ncopies >= 1); | |
4474 | ||
4475 | /* FORNOW. This restriction should be relaxed. */ | |
4476 | if (nested_in_vect_loop && ncopies > 1) | |
4477 | { | |
73fbfcad | 4478 | if (dump_enabled_p ()) |
78c60e3d | 4479 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4480 | "multiple types in nested loop.\n"); |
ebfd146a IR |
4481 | return false; |
4482 | } | |
4483 | ||
a70d6342 | 4484 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4485 | return false; |
4486 | ||
8644a673 | 4487 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4488 | return false; |
4489 | ||
4490 | /* Is vectorizable load? */ | |
4491 | if (!is_gimple_assign (stmt)) | |
4492 | return false; | |
4493 | ||
4494 | scalar_dest = gimple_assign_lhs (stmt); | |
4495 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4496 | return false; | |
4497 | ||
4498 | code = gimple_assign_rhs_code (stmt); | |
4499 | if (code != ARRAY_REF | |
38000232 | 4500 | && code != BIT_FIELD_REF |
ebfd146a | 4501 | && code != INDIRECT_REF |
e9dbe7bb IR |
4502 | && code != COMPONENT_REF |
4503 | && code != IMAGPART_EXPR | |
70f34814 | 4504 | && code != REALPART_EXPR |
42373e0b RG |
4505 | && code != MEM_REF |
4506 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
ebfd146a IR |
4507 | return false; |
4508 | ||
4509 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
4510 | return false; | |
4511 | ||
7b7b1813 | 4512 | elem_type = TREE_TYPE (vectype); |
947131ba | 4513 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
4514 | |
4515 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
4516 | (e.g. - data copies). */ | |
947131ba | 4517 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 4518 | { |
73fbfcad | 4519 | if (dump_enabled_p ()) |
78c60e3d | 4520 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4521 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
4522 | return false; |
4523 | } | |
4524 | ||
ebfd146a | 4525 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 4526 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 4527 | { |
0d0293ac | 4528 | grouped_load = true; |
ebfd146a | 4529 | /* FORNOW */ |
aec7ae7d | 4530 | gcc_assert (! nested_in_vect_loop && !STMT_VINFO_GATHER_P (stmt_info)); |
ebfd146a | 4531 | |
e14c1050 | 4532 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
4533 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
4534 | { | |
e14c1050 | 4535 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
4536 | if (vect_load_lanes_supported (vectype, group_size)) |
4537 | load_lanes_p = true; | |
0d0293ac | 4538 | else if (!vect_grouped_load_supported (vectype, group_size)) |
b602d918 RS |
4539 | return false; |
4540 | } | |
ebfd146a IR |
4541 | } |
4542 | ||
a1e53f3f | 4543 | |
aec7ae7d JJ |
4544 | if (STMT_VINFO_GATHER_P (stmt_info)) |
4545 | { | |
4546 | gimple def_stmt; | |
4547 | tree def; | |
4548 | gather_decl = vect_check_gather (stmt, loop_vinfo, &gather_base, | |
4549 | &gather_off, &gather_scale); | |
4550 | gcc_assert (gather_decl); | |
24ee1384 | 4551 | if (!vect_is_simple_use_1 (gather_off, NULL, loop_vinfo, bb_vinfo, |
aec7ae7d JJ |
4552 | &def_stmt, &def, &gather_dt, |
4553 | &gather_off_vectype)) | |
4554 | { | |
73fbfcad | 4555 | if (dump_enabled_p ()) |
78c60e3d | 4556 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4557 | "gather index use not simple.\n"); |
aec7ae7d JJ |
4558 | return false; |
4559 | } | |
4560 | } | |
7d75abc8 | 4561 | else if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
14ac6aa2 | 4562 | ; |
319e6439 RG |
4563 | else |
4564 | { | |
4565 | negative = tree_int_cst_compare (nested_in_vect_loop | |
4566 | ? STMT_VINFO_DR_STEP (stmt_info) | |
4567 | : DR_STEP (dr), | |
4568 | size_zero_node) < 0; | |
4569 | if (negative && ncopies > 1) | |
4570 | { | |
73fbfcad | 4571 | if (dump_enabled_p ()) |
78c60e3d | 4572 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4573 | "multiple types with negative step.\n"); |
319e6439 RG |
4574 | return false; |
4575 | } | |
4576 | ||
4577 | if (negative) | |
4578 | { | |
08940f33 RB |
4579 | if (grouped_load) |
4580 | { | |
4581 | if (dump_enabled_p ()) | |
4582 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 TJ |
4583 | "negative step for group load not supported" |
4584 | "\n"); | |
08940f33 RB |
4585 | return false; |
4586 | } | |
319e6439 RG |
4587 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
4588 | if (alignment_support_scheme != dr_aligned | |
4589 | && alignment_support_scheme != dr_unaligned_supported) | |
4590 | { | |
73fbfcad | 4591 | if (dump_enabled_p ()) |
78c60e3d | 4592 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4593 | "negative step but alignment required.\n"); |
319e6439 RG |
4594 | return false; |
4595 | } | |
4596 | if (!perm_mask_for_reverse (vectype)) | |
4597 | { | |
73fbfcad | 4598 | if (dump_enabled_p ()) |
78c60e3d | 4599 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
4600 | "negative step and reversing not supported." |
4601 | "\n"); | |
319e6439 RG |
4602 | return false; |
4603 | } | |
4604 | } | |
7d75abc8 | 4605 | } |
aec7ae7d | 4606 | |
ebfd146a IR |
4607 | if (!vec_stmt) /* transformation not required. */ |
4608 | { | |
4609 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
92345349 | 4610 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, NULL, NULL, NULL); |
ebfd146a IR |
4611 | return true; |
4612 | } | |
4613 | ||
73fbfcad | 4614 | if (dump_enabled_p ()) |
78c60e3d | 4615 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4616 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a IR |
4617 | |
4618 | /** Transform. **/ | |
4619 | ||
c716e67f XDL |
4620 | ensure_base_align (stmt_info, dr); |
4621 | ||
aec7ae7d JJ |
4622 | if (STMT_VINFO_GATHER_P (stmt_info)) |
4623 | { | |
4624 | tree vec_oprnd0 = NULL_TREE, op; | |
4625 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
4626 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
4627 | tree ptr, mask, var, scale, perm_mask = NULL_TREE, prev_res = NULL_TREE; | |
4628 | edge pe = loop_preheader_edge (loop); | |
4629 | gimple_seq seq; | |
4630 | basic_block new_bb; | |
4631 | enum { NARROW, NONE, WIDEN } modifier; | |
4632 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
4633 | ||
4634 | if (nunits == gather_off_nunits) | |
4635 | modifier = NONE; | |
4636 | else if (nunits == gather_off_nunits / 2) | |
4637 | { | |
4638 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
4639 | modifier = WIDEN; | |
4640 | ||
4641 | for (i = 0; i < gather_off_nunits; ++i) | |
4642 | sel[i] = i | nunits; | |
4643 | ||
3fcc1b55 | 4644 | perm_mask = vect_gen_perm_mask (gather_off_vectype, sel); |
aec7ae7d JJ |
4645 | gcc_assert (perm_mask != NULL_TREE); |
4646 | } | |
4647 | else if (nunits == gather_off_nunits * 2) | |
4648 | { | |
4649 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
4650 | modifier = NARROW; | |
4651 | ||
4652 | for (i = 0; i < nunits; ++i) | |
4653 | sel[i] = i < gather_off_nunits | |
4654 | ? i : i + nunits - gather_off_nunits; | |
4655 | ||
3fcc1b55 | 4656 | perm_mask = vect_gen_perm_mask (vectype, sel); |
aec7ae7d JJ |
4657 | gcc_assert (perm_mask != NULL_TREE); |
4658 | ncopies *= 2; | |
4659 | } | |
4660 | else | |
4661 | gcc_unreachable (); | |
4662 | ||
4663 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
4664 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4665 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4666 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4667 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4668 | scaletype = TREE_VALUE (arglist); | |
4669 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
4670 | && types_compatible_p (srctype, masktype)); | |
4671 | ||
4672 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4673 | ||
4674 | ptr = fold_convert (ptrtype, gather_base); | |
4675 | if (!is_gimple_min_invariant (ptr)) | |
4676 | { | |
4677 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
4678 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
4679 | gcc_assert (!new_bb); | |
4680 | } | |
4681 | ||
4682 | /* Currently we support only unconditional gather loads, | |
4683 | so mask should be all ones. */ | |
4684 | if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
4685 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
4686 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) | |
4687 | { | |
4688 | REAL_VALUE_TYPE r; | |
4689 | long tmp[6]; | |
4690 | for (j = 0; j < 6; ++j) | |
4691 | tmp[j] = -1; | |
4692 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
4693 | mask = build_real (TREE_TYPE (masktype), r); | |
4694 | } | |
4695 | else | |
4696 | gcc_unreachable (); | |
4697 | mask = build_vector_from_val (masktype, mask); | |
4698 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
4699 | ||
4700 | scale = build_int_cst (scaletype, gather_scale); | |
4701 | ||
4702 | prev_stmt_info = NULL; | |
4703 | for (j = 0; j < ncopies; ++j) | |
4704 | { | |
4705 | if (modifier == WIDEN && (j & 1)) | |
4706 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
4707 | perm_mask, stmt, gsi); | |
4708 | else if (j == 0) | |
4709 | op = vec_oprnd0 | |
4710 | = vect_get_vec_def_for_operand (gather_off, stmt, NULL); | |
4711 | else | |
4712 | op = vec_oprnd0 | |
4713 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
4714 | ||
4715 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
4716 | { | |
4717 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
4718 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
4719 | var = vect_get_new_vect_var (idxtype, vect_simple_var, NULL); | |
aec7ae7d JJ |
4720 | var = make_ssa_name (var, NULL); |
4721 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); | |
4722 | new_stmt | |
4723 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, | |
4724 | op, NULL_TREE); | |
4725 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4726 | op = var; | |
4727 | } | |
4728 | ||
4729 | new_stmt | |
4730 | = gimple_build_call (gather_decl, 5, mask, ptr, op, mask, scale); | |
4731 | ||
4732 | if (!useless_type_conversion_p (vectype, rettype)) | |
4733 | { | |
4734 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
4735 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
4736 | var = vect_get_new_vect_var (rettype, vect_simple_var, NULL); | |
aec7ae7d JJ |
4737 | op = make_ssa_name (var, new_stmt); |
4738 | gimple_call_set_lhs (new_stmt, op); | |
4739 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4740 | var = make_ssa_name (vec_dest, NULL); | |
4741 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); | |
4742 | new_stmt | |
4743 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, op, | |
4744 | NULL_TREE); | |
4745 | } | |
4746 | else | |
4747 | { | |
4748 | var = make_ssa_name (vec_dest, new_stmt); | |
4749 | gimple_call_set_lhs (new_stmt, var); | |
4750 | } | |
4751 | ||
4752 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4753 | ||
4754 | if (modifier == NARROW) | |
4755 | { | |
4756 | if ((j & 1) == 0) | |
4757 | { | |
4758 | prev_res = var; | |
4759 | continue; | |
4760 | } | |
4761 | var = permute_vec_elements (prev_res, var, | |
4762 | perm_mask, stmt, gsi); | |
4763 | new_stmt = SSA_NAME_DEF_STMT (var); | |
4764 | } | |
4765 | ||
4766 | if (prev_stmt_info == NULL) | |
4767 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4768 | else | |
4769 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4770 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4771 | } | |
4772 | return true; | |
4773 | } | |
7d75abc8 MM |
4774 | else if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
4775 | { | |
4776 | gimple_stmt_iterator incr_gsi; | |
4777 | bool insert_after; | |
4778 | gimple incr; | |
4779 | tree offvar; | |
7d75abc8 MM |
4780 | tree ivstep; |
4781 | tree running_off; | |
9771b263 | 4782 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 4783 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
4784 | tree stride_base, stride_step, alias_off; |
4785 | ||
4786 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 4787 | |
14ac6aa2 RB |
4788 | stride_base |
4789 | = fold_build_pointer_plus | |
4790 | (unshare_expr (DR_BASE_ADDRESS (dr)), | |
4791 | size_binop (PLUS_EXPR, | |
4792 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (dr))), | |
4793 | convert_to_ptrofftype (DR_INIT(dr)))); | |
4794 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (dr))); | |
7d75abc8 MM |
4795 | |
4796 | /* For a load with loop-invariant (but other than power-of-2) | |
4797 | stride (i.e. not a grouped access) like so: | |
4798 | ||
4799 | for (i = 0; i < n; i += stride) | |
4800 | ... = array[i]; | |
4801 | ||
4802 | we generate a new induction variable and new accesses to | |
4803 | form a new vector (or vectors, depending on ncopies): | |
4804 | ||
4805 | for (j = 0; ; j += VF*stride) | |
4806 | tmp1 = array[j]; | |
4807 | tmp2 = array[j + stride]; | |
4808 | ... | |
4809 | vectemp = {tmp1, tmp2, ...} | |
4810 | */ | |
4811 | ||
4812 | ivstep = stride_step; | |
4813 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
4814 | build_int_cst (TREE_TYPE (ivstep), vf)); | |
4815 | ||
4816 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
4817 | ||
4818 | create_iv (stride_base, ivstep, NULL, | |
4819 | loop, &incr_gsi, insert_after, | |
4820 | &offvar, NULL); | |
4821 | incr = gsi_stmt (incr_gsi); | |
4822 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo, NULL)); | |
4823 | ||
4824 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
4825 | if (stmts) | |
4826 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
4827 | ||
4828 | prev_stmt_info = NULL; | |
4829 | running_off = offvar; | |
14ac6aa2 | 4830 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (dr)), 0); |
7d75abc8 MM |
4831 | for (j = 0; j < ncopies; j++) |
4832 | { | |
4833 | tree vec_inv; | |
4834 | ||
9771b263 | 4835 | vec_alloc (v, nunits); |
7d75abc8 MM |
4836 | for (i = 0; i < nunits; i++) |
4837 | { | |
4838 | tree newref, newoff; | |
4839 | gimple incr; | |
14ac6aa2 RB |
4840 | newref = build2 (MEM_REF, TREE_TYPE (vectype), |
4841 | running_off, alias_off); | |
7d75abc8 MM |
4842 | |
4843 | newref = force_gimple_operand_gsi (gsi, newref, true, | |
4844 | NULL_TREE, true, | |
4845 | GSI_SAME_STMT); | |
4846 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, newref); | |
070ecdfd | 4847 | newoff = copy_ssa_name (running_off, NULL); |
14ac6aa2 RB |
4848 | incr = gimple_build_assign_with_ops (POINTER_PLUS_EXPR, newoff, |
4849 | running_off, stride_step); | |
7d75abc8 MM |
4850 | vect_finish_stmt_generation (stmt, incr, gsi); |
4851 | ||
4852 | running_off = newoff; | |
4853 | } | |
4854 | ||
4855 | vec_inv = build_constructor (vectype, v); | |
4856 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
4857 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
7d75abc8 MM |
4858 | |
4859 | if (j == 0) | |
4860 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4861 | else | |
4862 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4863 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4864 | } | |
4865 | return true; | |
4866 | } | |
aec7ae7d | 4867 | |
0d0293ac | 4868 | if (grouped_load) |
ebfd146a | 4869 | { |
e14c1050 | 4870 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
6aa904c4 | 4871 | if (slp |
01d8bf07 | 4872 | && !SLP_TREE_LOAD_PERMUTATION (slp_node).exists () |
9771b263 DN |
4873 | && first_stmt != SLP_TREE_SCALAR_STMTS (slp_node)[0]) |
4874 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 4875 | |
ebfd146a | 4876 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
4877 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
4878 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
4879 | ??? But we can only do so if there is exactly one | |
4880 | as we have no way to get at the rest. Leave the CSE | |
4881 | opportunity alone. | |
4882 | ??? With the group load eventually participating | |
4883 | in multiple different permutations (having multiple | |
4884 | slp nodes which refer to the same group) the CSE | |
4885 | is even wrong code. See PR56270. */ | |
4886 | && !slp) | |
ebfd146a IR |
4887 | { |
4888 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4889 | return true; | |
4890 | } | |
4891 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 4892 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a IR |
4893 | |
4894 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
4895 | if (slp) | |
4896 | { | |
0d0293ac | 4897 | grouped_load = false; |
ebfd146a | 4898 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
01d8bf07 | 4899 | if (SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
a70d6342 | 4900 | slp_perm = true; |
a64b9c26 | 4901 | group_gap = GROUP_GAP (vinfo_for_stmt (first_stmt)); |
a70d6342 | 4902 | } |
ebfd146a | 4903 | else |
a64b9c26 RB |
4904 | { |
4905 | vec_num = group_size; | |
4906 | group_gap = 0; | |
4907 | } | |
ebfd146a IR |
4908 | } |
4909 | else | |
4910 | { | |
4911 | first_stmt = stmt; | |
4912 | first_dr = dr; | |
4913 | group_size = vec_num = 1; | |
a64b9c26 | 4914 | group_gap = 0; |
ebfd146a IR |
4915 | } |
4916 | ||
720f5239 | 4917 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 4918 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
4919 | /* Targets with load-lane instructions must not require explicit |
4920 | realignment. */ | |
4921 | gcc_assert (!load_lanes_p | |
4922 | || alignment_support_scheme == dr_aligned | |
4923 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
4924 | |
4925 | /* In case the vectorization factor (VF) is bigger than the number | |
4926 | of elements that we can fit in a vectype (nunits), we have to generate | |
4927 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 4928 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 4929 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 4930 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 4931 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
4932 | stmts that use the defs of the current stmt. The example below |
4933 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
4934 | need to create 4 vectorized stmts): | |
ebfd146a IR |
4935 | |
4936 | before vectorization: | |
4937 | RELATED_STMT VEC_STMT | |
4938 | S1: x = memref - - | |
4939 | S2: z = x + 1 - - | |
4940 | ||
4941 | step 1: vectorize stmt S1: | |
4942 | We first create the vector stmt VS1_0, and, as usual, record a | |
4943 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
4944 | Next, we create the vector stmt VS1_1, and record a pointer to | |
4945 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 4946 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
4947 | stmts and pointers: |
4948 | RELATED_STMT VEC_STMT | |
4949 | VS1_0: vx0 = memref0 VS1_1 - | |
4950 | VS1_1: vx1 = memref1 VS1_2 - | |
4951 | VS1_2: vx2 = memref2 VS1_3 - | |
4952 | VS1_3: vx3 = memref3 - - | |
4953 | S1: x = load - VS1_0 | |
4954 | S2: z = x + 1 - - | |
4955 | ||
b8698a0f L |
4956 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
4957 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
4958 | stmt S2. */ |
4959 | ||
0d0293ac | 4960 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
4961 | |
4962 | S1: x2 = &base + 2 | |
4963 | S2: x0 = &base | |
4964 | S3: x1 = &base + 1 | |
4965 | S4: x3 = &base + 3 | |
4966 | ||
b8698a0f | 4967 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
4968 | starting from the access of the first stmt of the chain: |
4969 | ||
4970 | VS1: vx0 = &base | |
4971 | VS2: vx1 = &base + vec_size*1 | |
4972 | VS3: vx3 = &base + vec_size*2 | |
4973 | VS4: vx4 = &base + vec_size*3 | |
4974 | ||
4975 | Then permutation statements are generated: | |
4976 | ||
e2c83630 RH |
4977 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
4978 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
4979 | ... |
4980 | ||
4981 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
4982 | (the order of the data-refs in the output of vect_permute_load_chain | |
4983 | corresponds to the order of scalar stmts in the interleaving chain - see | |
4984 | the documentation of vect_permute_load_chain()). | |
4985 | The generation of permutation stmts and recording them in | |
0d0293ac | 4986 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 4987 | |
b8698a0f | 4988 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
4989 | permutation stmts above are created for every copy. The result vector |
4990 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
4991 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
4992 | |
4993 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
4994 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
4995 | we generate the following code: | |
4996 | p = initial_addr; | |
4997 | indx = 0; | |
4998 | loop { | |
4999 | p = p + indx * vectype_size; | |
5000 | vec_dest = *(p); | |
5001 | indx = indx + 1; | |
5002 | } | |
5003 | ||
5004 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 5005 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
5006 | then generate the following code, in which the data in each iteration is |
5007 | obtained by two vector loads, one from the previous iteration, and one | |
5008 | from the current iteration: | |
5009 | p1 = initial_addr; | |
5010 | msq_init = *(floor(p1)) | |
5011 | p2 = initial_addr + VS - 1; | |
5012 | realignment_token = call target_builtin; | |
5013 | indx = 0; | |
5014 | loop { | |
5015 | p2 = p2 + indx * vectype_size | |
5016 | lsq = *(floor(p2)) | |
5017 | vec_dest = realign_load (msq, lsq, realignment_token) | |
5018 | indx = indx + 1; | |
5019 | msq = lsq; | |
5020 | } */ | |
5021 | ||
5022 | /* If the misalignment remains the same throughout the execution of the | |
5023 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 5024 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
5025 | This can only occur when vectorizing memory accesses in the inner-loop |
5026 | nested within an outer-loop that is being vectorized. */ | |
5027 | ||
d1e4b493 | 5028 | if (nested_in_vect_loop |
211bea38 | 5029 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
ebfd146a IR |
5030 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) |
5031 | { | |
5032 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
5033 | compute_in_loop = true; | |
5034 | } | |
5035 | ||
5036 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
5037 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 5038 | && !compute_in_loop) |
ebfd146a IR |
5039 | { |
5040 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
5041 | alignment_support_scheme, NULL_TREE, | |
5042 | &at_loop); | |
5043 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
5044 | { | |
5045 | phi = SSA_NAME_DEF_STMT (msq); | |
5046 | offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
5047 | } | |
5048 | } | |
5049 | else | |
5050 | at_loop = loop; | |
5051 | ||
a1e53f3f L |
5052 | if (negative) |
5053 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
5054 | ||
272c6793 RS |
5055 | if (load_lanes_p) |
5056 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
5057 | else | |
5058 | aggr_type = vectype; | |
5059 | ||
ebfd146a IR |
5060 | prev_stmt_info = NULL; |
5061 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 5062 | { |
272c6793 | 5063 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 5064 | if (j == 0) |
74bf76ed JJ |
5065 | { |
5066 | bool simd_lane_access_p | |
5067 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
5068 | if (simd_lane_access_p | |
5069 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
5070 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
5071 | && integer_zerop (DR_OFFSET (first_dr)) | |
5072 | && integer_zerop (DR_INIT (first_dr)) | |
5073 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
5074 | get_alias_set (DR_REF (first_dr))) | |
5075 | && (alignment_support_scheme == dr_aligned | |
5076 | || alignment_support_scheme == dr_unaligned_supported)) | |
5077 | { | |
5078 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
5079 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
5080 | (DR_REF (first_dr)), 0); | |
8928eff3 | 5081 | inv_p = false; |
74bf76ed JJ |
5082 | } |
5083 | else | |
5084 | dataref_ptr | |
5085 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
5086 | offset, &dummy, gsi, &ptr_incr, | |
5087 | simd_lane_access_p, &inv_p); | |
5088 | } | |
5089 | else if (dataref_offset) | |
5090 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
5091 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 5092 | else |
272c6793 RS |
5093 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
5094 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 5095 | |
0d0293ac | 5096 | if (grouped_load || slp_perm) |
9771b263 | 5097 | dr_chain.create (vec_num); |
5ce1ee7f | 5098 | |
272c6793 | 5099 | if (load_lanes_p) |
ebfd146a | 5100 | { |
272c6793 RS |
5101 | tree vec_array; |
5102 | ||
5103 | vec_array = create_vector_array (vectype, vec_num); | |
5104 | ||
5105 | /* Emit: | |
5106 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
5107 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
5108 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
5109 | gimple_call_set_lhs (new_stmt, vec_array); | |
5110 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 5111 | |
272c6793 RS |
5112 | /* Extract each vector into an SSA_NAME. */ |
5113 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 5114 | { |
272c6793 RS |
5115 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
5116 | vec_array, i); | |
9771b263 | 5117 | dr_chain.quick_push (new_temp); |
272c6793 RS |
5118 | } |
5119 | ||
5120 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 5121 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
5122 | } |
5123 | else | |
5124 | { | |
5125 | for (i = 0; i < vec_num; i++) | |
5126 | { | |
5127 | if (i > 0) | |
5128 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5129 | stmt, NULL_TREE); | |
5130 | ||
5131 | /* 2. Create the vector-load in the loop. */ | |
5132 | switch (alignment_support_scheme) | |
5133 | { | |
5134 | case dr_aligned: | |
5135 | case dr_unaligned_supported: | |
be1ac4ec | 5136 | { |
644ffefd MJ |
5137 | unsigned int align, misalign; |
5138 | ||
272c6793 RS |
5139 | data_ref |
5140 | = build2 (MEM_REF, vectype, dataref_ptr, | |
74bf76ed JJ |
5141 | dataref_offset |
5142 | ? dataref_offset | |
5143 | : build_int_cst (reference_alias_ptr_type | |
5144 | (DR_REF (first_dr)), 0)); | |
644ffefd | 5145 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 RS |
5146 | if (alignment_support_scheme == dr_aligned) |
5147 | { | |
5148 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 5149 | misalign = 0; |
272c6793 RS |
5150 | } |
5151 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
5152 | { | |
5153 | TREE_TYPE (data_ref) | |
5154 | = build_aligned_type (TREE_TYPE (data_ref), | |
5155 | TYPE_ALIGN (elem_type)); | |
644ffefd MJ |
5156 | align = TYPE_ALIGN_UNIT (elem_type); |
5157 | misalign = 0; | |
272c6793 RS |
5158 | } |
5159 | else | |
5160 | { | |
5161 | TREE_TYPE (data_ref) | |
5162 | = build_aligned_type (TREE_TYPE (data_ref), | |
5163 | TYPE_ALIGN (elem_type)); | |
644ffefd | 5164 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 5165 | } |
74bf76ed JJ |
5166 | if (dataref_offset == NULL_TREE) |
5167 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), | |
5168 | align, misalign); | |
272c6793 | 5169 | break; |
be1ac4ec | 5170 | } |
272c6793 | 5171 | case dr_explicit_realign: |
267d3070 | 5172 | { |
272c6793 RS |
5173 | tree ptr, bump; |
5174 | tree vs_minus_1; | |
5175 | ||
5176 | vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
5177 | ||
5178 | if (compute_in_loop) | |
5179 | msq = vect_setup_realignment (first_stmt, gsi, | |
5180 | &realignment_token, | |
5181 | dr_explicit_realign, | |
5182 | dataref_ptr, NULL); | |
5183 | ||
070ecdfd | 5184 | ptr = copy_ssa_name (dataref_ptr, NULL); |
272c6793 | 5185 | new_stmt = gimple_build_assign_with_ops |
070ecdfd | 5186 | (BIT_AND_EXPR, ptr, dataref_ptr, |
272c6793 RS |
5187 | build_int_cst |
5188 | (TREE_TYPE (dataref_ptr), | |
5189 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
5190 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
5191 | data_ref | |
5192 | = build2 (MEM_REF, vectype, ptr, | |
5193 | build_int_cst (reference_alias_ptr_type | |
5194 | (DR_REF (first_dr)), 0)); | |
5195 | vec_dest = vect_create_destination_var (scalar_dest, | |
5196 | vectype); | |
5197 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
5198 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5199 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5200 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
5201 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
5202 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5203 | msq = new_temp; | |
5204 | ||
5205 | bump = size_binop (MULT_EXPR, vs_minus_1, | |
7b7b1813 | 5206 | TYPE_SIZE_UNIT (elem_type)); |
272c6793 RS |
5207 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
5208 | new_stmt = gimple_build_assign_with_ops | |
5209 | (BIT_AND_EXPR, NULL_TREE, ptr, | |
5210 | build_int_cst | |
5211 | (TREE_TYPE (ptr), | |
5212 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
070ecdfd | 5213 | ptr = copy_ssa_name (dataref_ptr, new_stmt); |
272c6793 RS |
5214 | gimple_assign_set_lhs (new_stmt, ptr); |
5215 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5216 | data_ref | |
5217 | = build2 (MEM_REF, vectype, ptr, | |
5218 | build_int_cst (reference_alias_ptr_type | |
5219 | (DR_REF (first_dr)), 0)); | |
5220 | break; | |
267d3070 | 5221 | } |
272c6793 | 5222 | case dr_explicit_realign_optimized: |
070ecdfd | 5223 | new_temp = copy_ssa_name (dataref_ptr, NULL); |
272c6793 | 5224 | new_stmt = gimple_build_assign_with_ops |
070ecdfd | 5225 | (BIT_AND_EXPR, new_temp, dataref_ptr, |
272c6793 RS |
5226 | build_int_cst |
5227 | (TREE_TYPE (dataref_ptr), | |
5228 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
5229 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
5230 | data_ref | |
5231 | = build2 (MEM_REF, vectype, new_temp, | |
5232 | build_int_cst (reference_alias_ptr_type | |
5233 | (DR_REF (first_dr)), 0)); | |
5234 | break; | |
5235 | default: | |
5236 | gcc_unreachable (); | |
5237 | } | |
ebfd146a | 5238 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 5239 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
5240 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5241 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5242 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5243 | ||
272c6793 RS |
5244 | /* 3. Handle explicit realignment if necessary/supported. |
5245 | Create in loop: | |
5246 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
5247 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
5248 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 5249 | { |
272c6793 RS |
5250 | lsq = gimple_assign_lhs (new_stmt); |
5251 | if (!realignment_token) | |
5252 | realignment_token = dataref_ptr; | |
5253 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5254 | new_stmt | |
73804b12 RG |
5255 | = gimple_build_assign_with_ops (REALIGN_LOAD_EXPR, |
5256 | vec_dest, msq, lsq, | |
5257 | realignment_token); | |
272c6793 RS |
5258 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5259 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5260 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5261 | ||
5262 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
5263 | { | |
5264 | gcc_assert (phi); | |
5265 | if (i == vec_num - 1 && j == ncopies - 1) | |
5266 | add_phi_arg (phi, lsq, | |
5267 | loop_latch_edge (containing_loop), | |
9e227d60 | 5268 | UNKNOWN_LOCATION); |
272c6793 RS |
5269 | msq = lsq; |
5270 | } | |
ebfd146a | 5271 | } |
ebfd146a | 5272 | |
59fd17e3 RB |
5273 | /* 4. Handle invariant-load. */ |
5274 | if (inv_p && !bb_vinfo) | |
5275 | { | |
5276 | gimple_stmt_iterator gsi2 = *gsi; | |
5277 | gcc_assert (!grouped_load); | |
5278 | gsi_next (&gsi2); | |
5279 | new_temp = vect_init_vector (stmt, scalar_dest, | |
5280 | vectype, &gsi2); | |
5281 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
5282 | } | |
5283 | ||
272c6793 RS |
5284 | if (negative) |
5285 | { | |
aec7ae7d JJ |
5286 | tree perm_mask = perm_mask_for_reverse (vectype); |
5287 | new_temp = permute_vec_elements (new_temp, new_temp, | |
5288 | perm_mask, stmt, gsi); | |
ebfd146a IR |
5289 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
5290 | } | |
267d3070 | 5291 | |
272c6793 | 5292 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
5293 | vect_transform_grouped_load (). */ |
5294 | if (grouped_load || slp_perm) | |
9771b263 | 5295 | dr_chain.quick_push (new_temp); |
267d3070 | 5296 | |
272c6793 RS |
5297 | /* Store vector loads in the corresponding SLP_NODE. */ |
5298 | if (slp && !slp_perm) | |
9771b263 | 5299 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
272c6793 | 5300 | } |
a64b9c26 RB |
5301 | /* Bump the vector pointer to account for a gap. */ |
5302 | if (slp && group_gap != 0) | |
5303 | { | |
5304 | tree bump = size_binop (MULT_EXPR, | |
5305 | TYPE_SIZE_UNIT (elem_type), | |
5306 | size_int (group_gap)); | |
5307 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5308 | stmt, bump); | |
5309 | } | |
ebfd146a IR |
5310 | } |
5311 | ||
5312 | if (slp && !slp_perm) | |
5313 | continue; | |
5314 | ||
5315 | if (slp_perm) | |
5316 | { | |
01d8bf07 | 5317 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
ebfd146a IR |
5318 | slp_node_instance, false)) |
5319 | { | |
9771b263 | 5320 | dr_chain.release (); |
ebfd146a IR |
5321 | return false; |
5322 | } | |
5323 | } | |
5324 | else | |
5325 | { | |
0d0293ac | 5326 | if (grouped_load) |
ebfd146a | 5327 | { |
272c6793 | 5328 | if (!load_lanes_p) |
0d0293ac | 5329 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 5330 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
5331 | } |
5332 | else | |
5333 | { | |
5334 | if (j == 0) | |
5335 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5336 | else | |
5337 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5338 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5339 | } | |
5340 | } | |
9771b263 | 5341 | dr_chain.release (); |
ebfd146a IR |
5342 | } |
5343 | ||
ebfd146a IR |
5344 | return true; |
5345 | } | |
5346 | ||
5347 | /* Function vect_is_simple_cond. | |
b8698a0f | 5348 | |
ebfd146a IR |
5349 | Input: |
5350 | LOOP - the loop that is being vectorized. | |
5351 | COND - Condition that is checked for simple use. | |
5352 | ||
e9e1d143 RG |
5353 | Output: |
5354 | *COMP_VECTYPE - the vector type for the comparison. | |
5355 | ||
ebfd146a IR |
5356 | Returns whether a COND can be vectorized. Checks whether |
5357 | condition operands are supportable using vec_is_simple_use. */ | |
5358 | ||
87aab9b2 | 5359 | static bool |
24ee1384 IR |
5360 | vect_is_simple_cond (tree cond, gimple stmt, loop_vec_info loop_vinfo, |
5361 | bb_vec_info bb_vinfo, tree *comp_vectype) | |
ebfd146a IR |
5362 | { |
5363 | tree lhs, rhs; | |
5364 | tree def; | |
5365 | enum vect_def_type dt; | |
e9e1d143 | 5366 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a IR |
5367 | |
5368 | if (!COMPARISON_CLASS_P (cond)) | |
5369 | return false; | |
5370 | ||
5371 | lhs = TREE_OPERAND (cond, 0); | |
5372 | rhs = TREE_OPERAND (cond, 1); | |
5373 | ||
5374 | if (TREE_CODE (lhs) == SSA_NAME) | |
5375 | { | |
5376 | gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); | |
24ee1384 IR |
5377 | if (!vect_is_simple_use_1 (lhs, stmt, loop_vinfo, bb_vinfo, |
5378 | &lhs_def_stmt, &def, &dt, &vectype1)) | |
ebfd146a IR |
5379 | return false; |
5380 | } | |
5381 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
5382 | && TREE_CODE (lhs) != FIXED_CST) | |
5383 | return false; | |
5384 | ||
5385 | if (TREE_CODE (rhs) == SSA_NAME) | |
5386 | { | |
5387 | gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); | |
24ee1384 IR |
5388 | if (!vect_is_simple_use_1 (rhs, stmt, loop_vinfo, bb_vinfo, |
5389 | &rhs_def_stmt, &def, &dt, &vectype2)) | |
ebfd146a IR |
5390 | return false; |
5391 | } | |
f7e531cf | 5392 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST |
ebfd146a IR |
5393 | && TREE_CODE (rhs) != FIXED_CST) |
5394 | return false; | |
5395 | ||
e9e1d143 | 5396 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
5397 | return true; |
5398 | } | |
5399 | ||
5400 | /* vectorizable_condition. | |
5401 | ||
b8698a0f L |
5402 | Check if STMT is conditional modify expression that can be vectorized. |
5403 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
5404 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
5405 | at GSI. |
5406 | ||
5407 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
5408 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
5409 | else caluse if it is 2). | |
ebfd146a IR |
5410 | |
5411 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5412 | ||
4bbe8262 | 5413 | bool |
ebfd146a | 5414 | vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi, |
f7e531cf IR |
5415 | gimple *vec_stmt, tree reduc_def, int reduc_index, |
5416 | slp_tree slp_node) | |
ebfd146a IR |
5417 | { |
5418 | tree scalar_dest = NULL_TREE; | |
5419 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
5420 | tree cond_expr, then_clause, else_clause; |
5421 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5422 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
df11cc78 | 5423 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
5424 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
5425 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
ebfd146a IR |
5426 | tree vec_compare, vec_cond_expr; |
5427 | tree new_temp; | |
5428 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
ebfd146a | 5429 | tree def; |
a855b1b1 | 5430 | enum vect_def_type dt, dts[4]; |
ebfd146a | 5431 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
f7e531cf | 5432 | int ncopies; |
ebfd146a | 5433 | enum tree_code code; |
a855b1b1 | 5434 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
5435 | int i, j; |
5436 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
5437 | vec<tree> vec_oprnds0 = vNULL; |
5438 | vec<tree> vec_oprnds1 = vNULL; | |
5439 | vec<tree> vec_oprnds2 = vNULL; | |
5440 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 5441 | tree vec_cmp_type; |
b8698a0f | 5442 | |
f7e531cf IR |
5443 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
5444 | ncopies = 1; | |
5445 | else | |
5446 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
437f4a00 | 5447 | |
ebfd146a | 5448 | gcc_assert (ncopies >= 1); |
a855b1b1 | 5449 | if (reduc_index && ncopies > 1) |
ebfd146a IR |
5450 | return false; /* FORNOW */ |
5451 | ||
f7e531cf IR |
5452 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
5453 | return false; | |
5454 | ||
5455 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
ebfd146a IR |
5456 | return false; |
5457 | ||
4bbe8262 IR |
5458 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5459 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
5460 | && reduc_def)) | |
ebfd146a IR |
5461 | return false; |
5462 | ||
ebfd146a | 5463 | /* FORNOW: not yet supported. */ |
b8698a0f | 5464 | if (STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 5465 | { |
73fbfcad | 5466 | if (dump_enabled_p ()) |
78c60e3d | 5467 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5468 | "value used after loop.\n"); |
ebfd146a IR |
5469 | return false; |
5470 | } | |
5471 | ||
5472 | /* Is vectorizable conditional operation? */ | |
5473 | if (!is_gimple_assign (stmt)) | |
5474 | return false; | |
5475 | ||
5476 | code = gimple_assign_rhs_code (stmt); | |
5477 | ||
5478 | if (code != COND_EXPR) | |
5479 | return false; | |
5480 | ||
4e71066d RG |
5481 | cond_expr = gimple_assign_rhs1 (stmt); |
5482 | then_clause = gimple_assign_rhs2 (stmt); | |
5483 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 5484 | |
24ee1384 IR |
5485 | if (!vect_is_simple_cond (cond_expr, stmt, loop_vinfo, bb_vinfo, |
5486 | &comp_vectype) | |
e9e1d143 | 5487 | || !comp_vectype) |
ebfd146a IR |
5488 | return false; |
5489 | ||
5490 | if (TREE_CODE (then_clause) == SSA_NAME) | |
5491 | { | |
5492 | gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause); | |
24ee1384 | 5493 | if (!vect_is_simple_use (then_clause, stmt, loop_vinfo, bb_vinfo, |
ebfd146a IR |
5494 | &then_def_stmt, &def, &dt)) |
5495 | return false; | |
5496 | } | |
b8698a0f | 5497 | else if (TREE_CODE (then_clause) != INTEGER_CST |
ebfd146a IR |
5498 | && TREE_CODE (then_clause) != REAL_CST |
5499 | && TREE_CODE (then_clause) != FIXED_CST) | |
5500 | return false; | |
5501 | ||
5502 | if (TREE_CODE (else_clause) == SSA_NAME) | |
5503 | { | |
5504 | gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause); | |
24ee1384 | 5505 | if (!vect_is_simple_use (else_clause, stmt, loop_vinfo, bb_vinfo, |
ebfd146a IR |
5506 | &else_def_stmt, &def, &dt)) |
5507 | return false; | |
5508 | } | |
b8698a0f | 5509 | else if (TREE_CODE (else_clause) != INTEGER_CST |
ebfd146a IR |
5510 | && TREE_CODE (else_clause) != REAL_CST |
5511 | && TREE_CODE (else_clause) != FIXED_CST) | |
5512 | return false; | |
5513 | ||
74946978 MP |
5514 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (vectype))); |
5515 | /* The result of a vector comparison should be signed type. */ | |
5516 | tree cmp_type = build_nonstandard_integer_type (prec, 0); | |
5517 | vec_cmp_type = get_same_sized_vectype (cmp_type, vectype); | |
5518 | if (vec_cmp_type == NULL_TREE) | |
5519 | return false; | |
784fb9b3 | 5520 | |
b8698a0f | 5521 | if (!vec_stmt) |
ebfd146a IR |
5522 | { |
5523 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
e9e1d143 | 5524 | return expand_vec_cond_expr_p (vectype, comp_vectype); |
ebfd146a IR |
5525 | } |
5526 | ||
f7e531cf IR |
5527 | /* Transform. */ |
5528 | ||
5529 | if (!slp_node) | |
5530 | { | |
9771b263 DN |
5531 | vec_oprnds0.create (1); |
5532 | vec_oprnds1.create (1); | |
5533 | vec_oprnds2.create (1); | |
5534 | vec_oprnds3.create (1); | |
f7e531cf | 5535 | } |
ebfd146a IR |
5536 | |
5537 | /* Handle def. */ | |
5538 | scalar_dest = gimple_assign_lhs (stmt); | |
5539 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5540 | ||
5541 | /* Handle cond expr. */ | |
a855b1b1 MM |
5542 | for (j = 0; j < ncopies; j++) |
5543 | { | |
f7e531cf | 5544 | gimple new_stmt = NULL; |
a855b1b1 MM |
5545 | if (j == 0) |
5546 | { | |
f7e531cf IR |
5547 | if (slp_node) |
5548 | { | |
9771b263 DN |
5549 | vec<tree> ops; |
5550 | ops.create (4); | |
37b5ec8f | 5551 | vec<vec<tree> > vec_defs; |
9771b263 DN |
5552 | |
5553 | vec_defs.create (4); | |
5554 | ops.safe_push (TREE_OPERAND (cond_expr, 0)); | |
5555 | ops.safe_push (TREE_OPERAND (cond_expr, 1)); | |
5556 | ops.safe_push (then_clause); | |
5557 | ops.safe_push (else_clause); | |
f7e531cf | 5558 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); |
37b5ec8f JJ |
5559 | vec_oprnds3 = vec_defs.pop (); |
5560 | vec_oprnds2 = vec_defs.pop (); | |
5561 | vec_oprnds1 = vec_defs.pop (); | |
5562 | vec_oprnds0 = vec_defs.pop (); | |
f7e531cf | 5563 | |
9771b263 DN |
5564 | ops.release (); |
5565 | vec_defs.release (); | |
f7e531cf IR |
5566 | } |
5567 | else | |
5568 | { | |
5569 | gimple gtemp; | |
5570 | vec_cond_lhs = | |
a855b1b1 MM |
5571 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), |
5572 | stmt, NULL); | |
24ee1384 IR |
5573 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), stmt, |
5574 | loop_vinfo, NULL, >emp, &def, &dts[0]); | |
f7e531cf IR |
5575 | |
5576 | vec_cond_rhs = | |
5577 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
5578 | stmt, NULL); | |
24ee1384 IR |
5579 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), stmt, |
5580 | loop_vinfo, NULL, >emp, &def, &dts[1]); | |
f7e531cf IR |
5581 | if (reduc_index == 1) |
5582 | vec_then_clause = reduc_def; | |
5583 | else | |
5584 | { | |
5585 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
5586 | stmt, NULL); | |
24ee1384 | 5587 | vect_is_simple_use (then_clause, stmt, loop_vinfo, |
f7e531cf IR |
5588 | NULL, >emp, &def, &dts[2]); |
5589 | } | |
5590 | if (reduc_index == 2) | |
5591 | vec_else_clause = reduc_def; | |
5592 | else | |
5593 | { | |
5594 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
a855b1b1 | 5595 | stmt, NULL); |
24ee1384 | 5596 | vect_is_simple_use (else_clause, stmt, loop_vinfo, |
a855b1b1 | 5597 | NULL, >emp, &def, &dts[3]); |
f7e531cf | 5598 | } |
a855b1b1 MM |
5599 | } |
5600 | } | |
5601 | else | |
5602 | { | |
f7e531cf | 5603 | vec_cond_lhs = vect_get_vec_def_for_stmt_copy (dts[0], |
9771b263 | 5604 | vec_oprnds0.pop ()); |
f7e531cf | 5605 | vec_cond_rhs = vect_get_vec_def_for_stmt_copy (dts[1], |
9771b263 | 5606 | vec_oprnds1.pop ()); |
a855b1b1 | 5607 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 5608 | vec_oprnds2.pop ()); |
a855b1b1 | 5609 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 5610 | vec_oprnds3.pop ()); |
f7e531cf IR |
5611 | } |
5612 | ||
5613 | if (!slp_node) | |
5614 | { | |
9771b263 DN |
5615 | vec_oprnds0.quick_push (vec_cond_lhs); |
5616 | vec_oprnds1.quick_push (vec_cond_rhs); | |
5617 | vec_oprnds2.quick_push (vec_then_clause); | |
5618 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
5619 | } |
5620 | ||
9dc3f7de | 5621 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 5622 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 5623 | { |
9771b263 DN |
5624 | vec_cond_rhs = vec_oprnds1[i]; |
5625 | vec_then_clause = vec_oprnds2[i]; | |
5626 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 5627 | |
784fb9b3 JJ |
5628 | vec_compare = build2 (TREE_CODE (cond_expr), vec_cmp_type, |
5629 | vec_cond_lhs, vec_cond_rhs); | |
f7e531cf IR |
5630 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, |
5631 | vec_compare, vec_then_clause, vec_else_clause); | |
a855b1b1 | 5632 | |
f7e531cf IR |
5633 | new_stmt = gimple_build_assign (vec_dest, vec_cond_expr); |
5634 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5635 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5636 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5637 | if (slp_node) | |
9771b263 | 5638 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
5639 | } |
5640 | ||
5641 | if (slp_node) | |
5642 | continue; | |
5643 | ||
5644 | if (j == 0) | |
5645 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5646 | else | |
5647 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5648 | ||
5649 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 5650 | } |
b8698a0f | 5651 | |
9771b263 DN |
5652 | vec_oprnds0.release (); |
5653 | vec_oprnds1.release (); | |
5654 | vec_oprnds2.release (); | |
5655 | vec_oprnds3.release (); | |
f7e531cf | 5656 | |
ebfd146a IR |
5657 | return true; |
5658 | } | |
5659 | ||
5660 | ||
8644a673 | 5661 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
5662 | |
5663 | bool | |
a70d6342 | 5664 | vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 5665 | { |
8644a673 | 5666 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 5667 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 5668 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 5669 | bool ok; |
a70d6342 | 5670 | tree scalar_type, vectype; |
363477c0 JJ |
5671 | gimple pattern_stmt; |
5672 | gimple_seq pattern_def_seq; | |
ebfd146a | 5673 | |
73fbfcad | 5674 | if (dump_enabled_p ()) |
ebfd146a | 5675 | { |
78c60e3d SS |
5676 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
5677 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 5678 | dump_printf (MSG_NOTE, "\n"); |
8644a673 | 5679 | } |
ebfd146a | 5680 | |
1825a1f3 | 5681 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 5682 | { |
73fbfcad | 5683 | if (dump_enabled_p ()) |
78c60e3d | 5684 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5685 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
5686 | |
5687 | return false; | |
5688 | } | |
b8698a0f L |
5689 | |
5690 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
5691 | to include: |
5692 | - the COND_EXPR which is the loop exit condition | |
5693 | - any LABEL_EXPRs in the loop | |
b8698a0f | 5694 | - computations that are used only for array indexing or loop control. |
8644a673 | 5695 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 5696 | instance, therefore, all the statements are relevant. |
ebfd146a | 5697 | |
d092494c | 5698 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 5699 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
5700 | statements. In basic blocks we are called from some SLP instance |
5701 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
5702 | already will be part of SLP instance. */ | |
83197f37 IR |
5703 | |
5704 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 5705 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 5706 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 5707 | { |
9d5e7640 | 5708 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 5709 | && pattern_stmt |
9d5e7640 IR |
5710 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
5711 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
5712 | { | |
83197f37 | 5713 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
5714 | stmt = pattern_stmt; |
5715 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 5716 | if (dump_enabled_p ()) |
9d5e7640 | 5717 | { |
78c60e3d SS |
5718 | dump_printf_loc (MSG_NOTE, vect_location, |
5719 | "==> examining pattern statement: "); | |
5720 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 5721 | dump_printf (MSG_NOTE, "\n"); |
9d5e7640 IR |
5722 | } |
5723 | } | |
5724 | else | |
5725 | { | |
73fbfcad | 5726 | if (dump_enabled_p ()) |
e645e942 | 5727 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 5728 | |
9d5e7640 IR |
5729 | return true; |
5730 | } | |
8644a673 | 5731 | } |
83197f37 | 5732 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 5733 | && node == NULL |
83197f37 IR |
5734 | && pattern_stmt |
5735 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
5736 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
5737 | { | |
5738 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 5739 | if (dump_enabled_p ()) |
83197f37 | 5740 | { |
78c60e3d SS |
5741 | dump_printf_loc (MSG_NOTE, vect_location, |
5742 | "==> examining pattern statement: "); | |
5743 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 5744 | dump_printf (MSG_NOTE, "\n"); |
83197f37 IR |
5745 | } |
5746 | ||
5747 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
5748 | return false; | |
5749 | } | |
ebfd146a | 5750 | |
1107f3ae | 5751 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 5752 | && node == NULL |
363477c0 | 5753 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 5754 | { |
363477c0 | 5755 | gimple_stmt_iterator si; |
1107f3ae | 5756 | |
363477c0 JJ |
5757 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
5758 | { | |
5759 | gimple pattern_def_stmt = gsi_stmt (si); | |
5760 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) | |
5761 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
5762 | { | |
5763 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 5764 | if (dump_enabled_p ()) |
363477c0 | 5765 | { |
78c60e3d SS |
5766 | dump_printf_loc (MSG_NOTE, vect_location, |
5767 | "==> examining pattern def statement: "); | |
5768 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
e645e942 | 5769 | dump_printf (MSG_NOTE, "\n"); |
363477c0 | 5770 | } |
1107f3ae | 5771 | |
363477c0 JJ |
5772 | if (!vect_analyze_stmt (pattern_def_stmt, |
5773 | need_to_vectorize, node)) | |
5774 | return false; | |
5775 | } | |
5776 | } | |
5777 | } | |
1107f3ae | 5778 | |
8644a673 IR |
5779 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
5780 | { | |
5781 | case vect_internal_def: | |
5782 | break; | |
ebfd146a | 5783 | |
8644a673 | 5784 | case vect_reduction_def: |
7c5222ff | 5785 | case vect_nested_cycle: |
a70d6342 | 5786 | gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer |
8644a673 | 5787 | || relevance == vect_used_in_outer_by_reduction |
a70d6342 | 5788 | || relevance == vect_unused_in_scope)); |
8644a673 IR |
5789 | break; |
5790 | ||
5791 | case vect_induction_def: | |
5792 | case vect_constant_def: | |
5793 | case vect_external_def: | |
5794 | case vect_unknown_def_type: | |
5795 | default: | |
5796 | gcc_unreachable (); | |
5797 | } | |
ebfd146a | 5798 | |
a70d6342 IR |
5799 | if (bb_vinfo) |
5800 | { | |
5801 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
5802 | ||
b690cc0f | 5803 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
73fbfcad | 5804 | if (dump_enabled_p ()) |
a70d6342 | 5805 | { |
78c60e3d SS |
5806 | dump_printf_loc (MSG_NOTE, vect_location, |
5807 | "get vectype for scalar type: "); | |
5808 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
e645e942 | 5809 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
5810 | } |
5811 | ||
5812 | vectype = get_vectype_for_scalar_type (scalar_type); | |
5813 | if (!vectype) | |
5814 | { | |
73fbfcad | 5815 | if (dump_enabled_p ()) |
a70d6342 | 5816 | { |
78c60e3d SS |
5817 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5818 | "not SLPed: unsupported data-type "); | |
5819 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
5820 | scalar_type); | |
e645e942 | 5821 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
a70d6342 IR |
5822 | } |
5823 | return false; | |
5824 | } | |
5825 | ||
73fbfcad | 5826 | if (dump_enabled_p ()) |
a70d6342 | 5827 | { |
78c60e3d SS |
5828 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); |
5829 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
e645e942 | 5830 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
5831 | } |
5832 | ||
5833 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
5834 | } | |
5835 | ||
8644a673 | 5836 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 5837 | { |
8644a673 IR |
5838 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
5839 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info)); | |
5840 | *need_to_vectorize = true; | |
ebfd146a IR |
5841 | } |
5842 | ||
8644a673 | 5843 | ok = true; |
b8698a0f | 5844 | if (!bb_vinfo |
a70d6342 IR |
5845 | && (STMT_VINFO_RELEVANT_P (stmt_info) |
5846 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
4a00c761 | 5847 | ok = (vectorizable_conversion (stmt, NULL, NULL, NULL) |
9dc3f7de | 5848 | || vectorizable_shift (stmt, NULL, NULL, NULL) |
8644a673 IR |
5849 | || vectorizable_operation (stmt, NULL, NULL, NULL) |
5850 | || vectorizable_assignment (stmt, NULL, NULL, NULL) | |
5851 | || vectorizable_load (stmt, NULL, NULL, NULL, NULL) | |
190c2236 | 5852 | || vectorizable_call (stmt, NULL, NULL, NULL) |
8644a673 | 5853 | || vectorizable_store (stmt, NULL, NULL, NULL) |
b5aeb3bb | 5854 | || vectorizable_reduction (stmt, NULL, NULL, NULL) |
f7e531cf | 5855 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, NULL)); |
a70d6342 IR |
5856 | else |
5857 | { | |
5858 | if (bb_vinfo) | |
4a00c761 JJ |
5859 | ok = (vectorizable_conversion (stmt, NULL, NULL, node) |
5860 | || vectorizable_shift (stmt, NULL, NULL, node) | |
9dc3f7de | 5861 | || vectorizable_operation (stmt, NULL, NULL, node) |
a70d6342 IR |
5862 | || vectorizable_assignment (stmt, NULL, NULL, node) |
5863 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
190c2236 | 5864 | || vectorizable_call (stmt, NULL, NULL, node) |
f7e531cf IR |
5865 | || vectorizable_store (stmt, NULL, NULL, node) |
5866 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node)); | |
b8698a0f | 5867 | } |
8644a673 IR |
5868 | |
5869 | if (!ok) | |
ebfd146a | 5870 | { |
73fbfcad | 5871 | if (dump_enabled_p ()) |
8644a673 | 5872 | { |
78c60e3d SS |
5873 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5874 | "not vectorized: relevant stmt not "); | |
5875 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
5876 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 5877 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
8644a673 | 5878 | } |
b8698a0f | 5879 | |
ebfd146a IR |
5880 | return false; |
5881 | } | |
5882 | ||
a70d6342 IR |
5883 | if (bb_vinfo) |
5884 | return true; | |
5885 | ||
8644a673 IR |
5886 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
5887 | need extra handling, except for vectorizable reductions. */ | |
5888 | if (STMT_VINFO_LIVE_P (stmt_info) | |
5889 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
5890 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 5891 | |
8644a673 | 5892 | if (!ok) |
ebfd146a | 5893 | { |
73fbfcad | 5894 | if (dump_enabled_p ()) |
8644a673 | 5895 | { |
78c60e3d SS |
5896 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5897 | "not vectorized: live stmt not "); | |
5898 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
5899 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 5900 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
8644a673 | 5901 | } |
b8698a0f | 5902 | |
8644a673 | 5903 | return false; |
ebfd146a IR |
5904 | } |
5905 | ||
ebfd146a IR |
5906 | return true; |
5907 | } | |
5908 | ||
5909 | ||
5910 | /* Function vect_transform_stmt. | |
5911 | ||
5912 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
5913 | ||
5914 | bool | |
5915 | vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi, | |
0d0293ac | 5916 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
5917 | slp_instance slp_node_instance) |
5918 | { | |
5919 | bool is_store = false; | |
5920 | gimple vec_stmt = NULL; | |
5921 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a | 5922 | bool done; |
ebfd146a IR |
5923 | |
5924 | switch (STMT_VINFO_TYPE (stmt_info)) | |
5925 | { | |
5926 | case type_demotion_vec_info_type: | |
ebfd146a | 5927 | case type_promotion_vec_info_type: |
ebfd146a IR |
5928 | case type_conversion_vec_info_type: |
5929 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
5930 | gcc_assert (done); | |
5931 | break; | |
5932 | ||
5933 | case induc_vec_info_type: | |
5934 | gcc_assert (!slp_node); | |
5935 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
5936 | gcc_assert (done); | |
5937 | break; | |
5938 | ||
9dc3f7de IR |
5939 | case shift_vec_info_type: |
5940 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
5941 | gcc_assert (done); | |
5942 | break; | |
5943 | ||
ebfd146a IR |
5944 | case op_vec_info_type: |
5945 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
5946 | gcc_assert (done); | |
5947 | break; | |
5948 | ||
5949 | case assignment_vec_info_type: | |
5950 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
5951 | gcc_assert (done); | |
5952 | break; | |
5953 | ||
5954 | case load_vec_info_type: | |
b8698a0f | 5955 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
5956 | slp_node_instance); |
5957 | gcc_assert (done); | |
5958 | break; | |
5959 | ||
5960 | case store_vec_info_type: | |
5961 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
5962 | gcc_assert (done); | |
0d0293ac | 5963 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
5964 | { |
5965 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 5966 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
5967 | one are skipped, and there vec_stmt_info shouldn't be freed |
5968 | meanwhile. */ | |
0d0293ac | 5969 | *grouped_store = true; |
ebfd146a IR |
5970 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
5971 | is_store = true; | |
5972 | } | |
5973 | else | |
5974 | is_store = true; | |
5975 | break; | |
5976 | ||
5977 | case condition_vec_info_type: | |
f7e531cf | 5978 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
5979 | gcc_assert (done); |
5980 | break; | |
5981 | ||
5982 | case call_vec_info_type: | |
190c2236 | 5983 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 5984 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
5985 | break; |
5986 | ||
5987 | case reduc_vec_info_type: | |
b5aeb3bb | 5988 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
5989 | gcc_assert (done); |
5990 | break; | |
5991 | ||
5992 | default: | |
5993 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
5994 | { | |
73fbfcad | 5995 | if (dump_enabled_p ()) |
78c60e3d | 5996 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5997 | "stmt not supported.\n"); |
ebfd146a IR |
5998 | gcc_unreachable (); |
5999 | } | |
6000 | } | |
6001 | ||
6002 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that | |
6003 | is being vectorized, but outside the immediately enclosing loop. */ | |
6004 | if (vec_stmt | |
a70d6342 IR |
6005 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
6006 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
6007 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
6008 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
6009 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 6010 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 6011 | vect_used_in_outer_by_reduction)) |
ebfd146a | 6012 | { |
a70d6342 IR |
6013 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
6014 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
6015 | imm_use_iterator imm_iter; |
6016 | use_operand_p use_p; | |
6017 | tree scalar_dest; | |
6018 | gimple exit_phi; | |
6019 | ||
73fbfcad | 6020 | if (dump_enabled_p ()) |
78c60e3d | 6021 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6022 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
6023 | |
6024 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
6025 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
6026 | STMT). */ | |
6027 | if (gimple_code (stmt) == GIMPLE_PHI) | |
6028 | scalar_dest = PHI_RESULT (stmt); | |
6029 | else | |
6030 | scalar_dest = gimple_assign_lhs (stmt); | |
6031 | ||
6032 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
6033 | { | |
6034 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
6035 | { | |
6036 | exit_phi = USE_STMT (use_p); | |
6037 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
6038 | } | |
6039 | } | |
6040 | } | |
6041 | ||
6042 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
6043 | being vectorized. */ | |
6044 | if (STMT_VINFO_LIVE_P (stmt_info) | |
6045 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
6046 | { | |
6047 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
6048 | gcc_assert (done); | |
6049 | } | |
6050 | ||
6051 | if (vec_stmt) | |
83197f37 | 6052 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 6053 | |
b8698a0f | 6054 | return is_store; |
ebfd146a IR |
6055 | } |
6056 | ||
6057 | ||
b8698a0f | 6058 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
6059 | stmt_vec_info. */ |
6060 | ||
6061 | void | |
6062 | vect_remove_stores (gimple first_stmt) | |
6063 | { | |
6064 | gimple next = first_stmt; | |
6065 | gimple tmp; | |
6066 | gimple_stmt_iterator next_si; | |
6067 | ||
6068 | while (next) | |
6069 | { | |
78048b1c JJ |
6070 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
6071 | ||
6072 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
6073 | if (is_pattern_stmt_p (stmt_info)) | |
6074 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
6075 | /* Free the attached stmt_vec_info and remove the stmt. */ |
6076 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 6077 | unlink_stmt_vdef (next); |
ebfd146a | 6078 | gsi_remove (&next_si, true); |
3d3f2249 | 6079 | release_defs (next); |
ebfd146a IR |
6080 | free_stmt_vec_info (next); |
6081 | next = tmp; | |
6082 | } | |
6083 | } | |
6084 | ||
6085 | ||
6086 | /* Function new_stmt_vec_info. | |
6087 | ||
6088 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
6089 | ||
6090 | stmt_vec_info | |
b8698a0f | 6091 | new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 6092 | bb_vec_info bb_vinfo) |
ebfd146a IR |
6093 | { |
6094 | stmt_vec_info res; | |
6095 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
6096 | ||
6097 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
6098 | STMT_VINFO_STMT (res) = stmt; | |
6099 | STMT_VINFO_LOOP_VINFO (res) = loop_vinfo; | |
a70d6342 | 6100 | STMT_VINFO_BB_VINFO (res) = bb_vinfo; |
8644a673 | 6101 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
6102 | STMT_VINFO_LIVE_P (res) = false; |
6103 | STMT_VINFO_VECTYPE (res) = NULL; | |
6104 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 6105 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
6106 | STMT_VINFO_IN_PATTERN_P (res) = false; |
6107 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 6108 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a IR |
6109 | STMT_VINFO_DATA_REF (res) = NULL; |
6110 | ||
6111 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
6112 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
6113 | STMT_VINFO_DR_INIT (res) = NULL; | |
6114 | STMT_VINFO_DR_STEP (res) = NULL; | |
6115 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
6116 | ||
6117 | if (gimple_code (stmt) == GIMPLE_PHI | |
6118 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
6119 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
6120 | else | |
8644a673 IR |
6121 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
6122 | ||
9771b263 | 6123 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 6124 | STMT_SLP_TYPE (res) = loop_vect; |
e14c1050 IR |
6125 | GROUP_FIRST_ELEMENT (res) = NULL; |
6126 | GROUP_NEXT_ELEMENT (res) = NULL; | |
6127 | GROUP_SIZE (res) = 0; | |
6128 | GROUP_STORE_COUNT (res) = 0; | |
6129 | GROUP_GAP (res) = 0; | |
6130 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
6131 | |
6132 | return res; | |
6133 | } | |
6134 | ||
6135 | ||
6136 | /* Create a hash table for stmt_vec_info. */ | |
6137 | ||
6138 | void | |
6139 | init_stmt_vec_info_vec (void) | |
6140 | { | |
9771b263 DN |
6141 | gcc_assert (!stmt_vec_info_vec.exists ()); |
6142 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
6143 | } |
6144 | ||
6145 | ||
6146 | /* Free hash table for stmt_vec_info. */ | |
6147 | ||
6148 | void | |
6149 | free_stmt_vec_info_vec (void) | |
6150 | { | |
93675444 JJ |
6151 | unsigned int i; |
6152 | vec_void_p info; | |
6153 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) | |
6154 | if (info != NULL) | |
6155 | free_stmt_vec_info (STMT_VINFO_STMT ((stmt_vec_info) info)); | |
9771b263 DN |
6156 | gcc_assert (stmt_vec_info_vec.exists ()); |
6157 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
6158 | } |
6159 | ||
6160 | ||
6161 | /* Free stmt vectorization related info. */ | |
6162 | ||
6163 | void | |
6164 | free_stmt_vec_info (gimple stmt) | |
6165 | { | |
6166 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6167 | ||
6168 | if (!stmt_info) | |
6169 | return; | |
6170 | ||
78048b1c JJ |
6171 | /* Check if this statement has a related "pattern stmt" |
6172 | (introduced by the vectorizer during the pattern recognition | |
6173 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
6174 | too. */ | |
6175 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
6176 | { | |
6177 | stmt_vec_info patt_info | |
6178 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
6179 | if (patt_info) | |
6180 | { | |
363477c0 JJ |
6181 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
6182 | if (seq) | |
6183 | { | |
6184 | gimple_stmt_iterator si; | |
6185 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
6186 | free_stmt_vec_info (gsi_stmt (si)); | |
6187 | } | |
78048b1c JJ |
6188 | free_stmt_vec_info (STMT_VINFO_RELATED_STMT (stmt_info)); |
6189 | } | |
6190 | } | |
6191 | ||
9771b263 | 6192 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
ebfd146a IR |
6193 | set_vinfo_for_stmt (stmt, NULL); |
6194 | free (stmt_info); | |
6195 | } | |
6196 | ||
6197 | ||
bb67d9c7 | 6198 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 6199 | |
bb67d9c7 | 6200 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
6201 | by the target. */ |
6202 | ||
bb67d9c7 RG |
6203 | static tree |
6204 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a IR |
6205 | { |
6206 | enum machine_mode inner_mode = TYPE_MODE (scalar_type); | |
cc4b5170 | 6207 | enum machine_mode simd_mode; |
2f816591 | 6208 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
6209 | int nunits; |
6210 | tree vectype; | |
6211 | ||
cc4b5170 | 6212 | if (nbytes == 0) |
ebfd146a IR |
6213 | return NULL_TREE; |
6214 | ||
48f2e373 RB |
6215 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
6216 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
6217 | return NULL_TREE; | |
6218 | ||
7b7b1813 RG |
6219 | /* For vector types of elements whose mode precision doesn't |
6220 | match their types precision we use a element type of mode | |
6221 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
6222 | they support the proper result truncation/extension. |
6223 | We also make sure to build vector types with INTEGER_TYPE | |
6224 | component type only. */ | |
6d7971b8 | 6225 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
6226 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
6227 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
6228 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
6229 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 6230 | |
ccbf5bb4 RG |
6231 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
6232 | When the component mode passes the above test simply use a type | |
6233 | corresponding to that mode. The theory is that any use that | |
6234 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 6235 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 6236 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
6237 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
6238 | ||
6239 | /* We can't build a vector type of elements with alignment bigger than | |
6240 | their size. */ | |
dfc2e2ac | 6241 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
6242 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
6243 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 6244 | |
dfc2e2ac RB |
6245 | /* If we felt back to using the mode fail if there was |
6246 | no scalar type for it. */ | |
6247 | if (scalar_type == NULL_TREE) | |
6248 | return NULL_TREE; | |
6249 | ||
bb67d9c7 RG |
6250 | /* If no size was supplied use the mode the target prefers. Otherwise |
6251 | lookup a vector mode of the specified size. */ | |
6252 | if (size == 0) | |
6253 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
6254 | else | |
6255 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
6256 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
6257 | if (nunits <= 1) | |
6258 | return NULL_TREE; | |
ebfd146a IR |
6259 | |
6260 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
6261 | |
6262 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
6263 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 6264 | return NULL_TREE; |
ebfd146a IR |
6265 | |
6266 | return vectype; | |
6267 | } | |
6268 | ||
bb67d9c7 RG |
6269 | unsigned int current_vector_size; |
6270 | ||
6271 | /* Function get_vectype_for_scalar_type. | |
6272 | ||
6273 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
6274 | by the target. */ | |
6275 | ||
6276 | tree | |
6277 | get_vectype_for_scalar_type (tree scalar_type) | |
6278 | { | |
6279 | tree vectype; | |
6280 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
6281 | current_vector_size); | |
6282 | if (vectype | |
6283 | && current_vector_size == 0) | |
6284 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
6285 | return vectype; | |
6286 | } | |
6287 | ||
b690cc0f RG |
6288 | /* Function get_same_sized_vectype |
6289 | ||
6290 | Returns a vector type corresponding to SCALAR_TYPE of size | |
6291 | VECTOR_TYPE if supported by the target. */ | |
6292 | ||
6293 | tree | |
bb67d9c7 | 6294 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 6295 | { |
bb67d9c7 RG |
6296 | return get_vectype_for_scalar_type_and_size |
6297 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
6298 | } |
6299 | ||
ebfd146a IR |
6300 | /* Function vect_is_simple_use. |
6301 | ||
6302 | Input: | |
a70d6342 IR |
6303 | LOOP_VINFO - the vect info of the loop that is being vectorized. |
6304 | BB_VINFO - the vect info of the basic block that is being vectorized. | |
24ee1384 | 6305 | OPERAND - operand of STMT in the loop or bb. |
ebfd146a IR |
6306 | DEF - the defining stmt in case OPERAND is an SSA_NAME. |
6307 | ||
6308 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 6309 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 6310 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 6311 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
6312 | is the case in reduction/induction computations). |
6313 | For basic blocks, supportable operands are constants and bb invariants. | |
6314 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
6315 | |
6316 | bool | |
24ee1384 | 6317 | vect_is_simple_use (tree operand, gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 6318 | bb_vec_info bb_vinfo, gimple *def_stmt, |
ebfd146a | 6319 | tree *def, enum vect_def_type *dt) |
b8698a0f | 6320 | { |
ebfd146a IR |
6321 | basic_block bb; |
6322 | stmt_vec_info stmt_vinfo; | |
a70d6342 | 6323 | struct loop *loop = NULL; |
b8698a0f | 6324 | |
a70d6342 IR |
6325 | if (loop_vinfo) |
6326 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
6327 | |
6328 | *def_stmt = NULL; | |
6329 | *def = NULL_TREE; | |
b8698a0f | 6330 | |
73fbfcad | 6331 | if (dump_enabled_p ()) |
ebfd146a | 6332 | { |
78c60e3d SS |
6333 | dump_printf_loc (MSG_NOTE, vect_location, |
6334 | "vect_is_simple_use: operand "); | |
6335 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 6336 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 6337 | } |
b8698a0f | 6338 | |
b758f602 | 6339 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
6340 | { |
6341 | *dt = vect_constant_def; | |
6342 | return true; | |
6343 | } | |
b8698a0f | 6344 | |
ebfd146a IR |
6345 | if (is_gimple_min_invariant (operand)) |
6346 | { | |
6347 | *def = operand; | |
8644a673 | 6348 | *dt = vect_external_def; |
ebfd146a IR |
6349 | return true; |
6350 | } | |
6351 | ||
6352 | if (TREE_CODE (operand) == PAREN_EXPR) | |
6353 | { | |
73fbfcad | 6354 | if (dump_enabled_p ()) |
e645e942 | 6355 | dump_printf_loc (MSG_NOTE, vect_location, "non-associatable copy.\n"); |
ebfd146a IR |
6356 | operand = TREE_OPERAND (operand, 0); |
6357 | } | |
b8698a0f | 6358 | |
ebfd146a IR |
6359 | if (TREE_CODE (operand) != SSA_NAME) |
6360 | { | |
73fbfcad | 6361 | if (dump_enabled_p ()) |
78c60e3d | 6362 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6363 | "not ssa-name.\n"); |
ebfd146a IR |
6364 | return false; |
6365 | } | |
b8698a0f | 6366 | |
ebfd146a IR |
6367 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
6368 | if (*def_stmt == NULL) | |
6369 | { | |
73fbfcad | 6370 | if (dump_enabled_p ()) |
78c60e3d | 6371 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6372 | "no def_stmt.\n"); |
ebfd146a IR |
6373 | return false; |
6374 | } | |
6375 | ||
73fbfcad | 6376 | if (dump_enabled_p ()) |
ebfd146a | 6377 | { |
78c60e3d SS |
6378 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
6379 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
e645e942 | 6380 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
6381 | } |
6382 | ||
8644a673 | 6383 | /* Empty stmt is expected only in case of a function argument. |
ebfd146a IR |
6384 | (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */ |
6385 | if (gimple_nop_p (*def_stmt)) | |
6386 | { | |
6387 | *def = operand; | |
8644a673 | 6388 | *dt = vect_external_def; |
ebfd146a IR |
6389 | return true; |
6390 | } | |
6391 | ||
6392 | bb = gimple_bb (*def_stmt); | |
a70d6342 IR |
6393 | |
6394 | if ((loop && !flow_bb_inside_loop_p (loop, bb)) | |
6395 | || (!loop && bb != BB_VINFO_BB (bb_vinfo)) | |
b8698a0f | 6396 | || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI)) |
8644a673 | 6397 | *dt = vect_external_def; |
ebfd146a IR |
6398 | else |
6399 | { | |
6400 | stmt_vinfo = vinfo_for_stmt (*def_stmt); | |
6401 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
6402 | } | |
6403 | ||
24ee1384 IR |
6404 | if (*dt == vect_unknown_def_type |
6405 | || (stmt | |
6406 | && *dt == vect_double_reduction_def | |
6407 | && gimple_code (stmt) != GIMPLE_PHI)) | |
ebfd146a | 6408 | { |
73fbfcad | 6409 | if (dump_enabled_p ()) |
78c60e3d | 6410 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6411 | "Unsupported pattern.\n"); |
ebfd146a IR |
6412 | return false; |
6413 | } | |
6414 | ||
73fbfcad | 6415 | if (dump_enabled_p ()) |
e645e942 | 6416 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: %d.\n", *dt); |
ebfd146a IR |
6417 | |
6418 | switch (gimple_code (*def_stmt)) | |
6419 | { | |
6420 | case GIMPLE_PHI: | |
6421 | *def = gimple_phi_result (*def_stmt); | |
6422 | break; | |
6423 | ||
6424 | case GIMPLE_ASSIGN: | |
6425 | *def = gimple_assign_lhs (*def_stmt); | |
6426 | break; | |
6427 | ||
6428 | case GIMPLE_CALL: | |
6429 | *def = gimple_call_lhs (*def_stmt); | |
6430 | if (*def != NULL) | |
6431 | break; | |
6432 | /* FALLTHRU */ | |
6433 | default: | |
73fbfcad | 6434 | if (dump_enabled_p ()) |
78c60e3d | 6435 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6436 | "unsupported defining stmt:\n"); |
ebfd146a IR |
6437 | return false; |
6438 | } | |
6439 | ||
6440 | return true; | |
6441 | } | |
6442 | ||
b690cc0f RG |
6443 | /* Function vect_is_simple_use_1. |
6444 | ||
6445 | Same as vect_is_simple_use_1 but also determines the vector operand | |
6446 | type of OPERAND and stores it to *VECTYPE. If the definition of | |
6447 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
6448 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
6449 | is responsible to compute the best suited vector type for the | |
6450 | scalar operand. */ | |
6451 | ||
6452 | bool | |
24ee1384 | 6453 | vect_is_simple_use_1 (tree operand, gimple stmt, loop_vec_info loop_vinfo, |
b690cc0f RG |
6454 | bb_vec_info bb_vinfo, gimple *def_stmt, |
6455 | tree *def, enum vect_def_type *dt, tree *vectype) | |
6456 | { | |
24ee1384 IR |
6457 | if (!vect_is_simple_use (operand, stmt, loop_vinfo, bb_vinfo, def_stmt, |
6458 | def, dt)) | |
b690cc0f RG |
6459 | return false; |
6460 | ||
6461 | /* Now get a vector type if the def is internal, otherwise supply | |
6462 | NULL_TREE and leave it up to the caller to figure out a proper | |
6463 | type for the use stmt. */ | |
6464 | if (*dt == vect_internal_def | |
6465 | || *dt == vect_induction_def | |
6466 | || *dt == vect_reduction_def | |
6467 | || *dt == vect_double_reduction_def | |
6468 | || *dt == vect_nested_cycle) | |
6469 | { | |
6470 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
6471 | |
6472 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
6473 | && !STMT_VINFO_RELEVANT (stmt_info) | |
6474 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 6475 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 6476 | |
b690cc0f RG |
6477 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
6478 | gcc_assert (*vectype != NULL_TREE); | |
6479 | } | |
6480 | else if (*dt == vect_uninitialized_def | |
6481 | || *dt == vect_constant_def | |
6482 | || *dt == vect_external_def) | |
6483 | *vectype = NULL_TREE; | |
6484 | else | |
6485 | gcc_unreachable (); | |
6486 | ||
6487 | return true; | |
6488 | } | |
6489 | ||
ebfd146a IR |
6490 | |
6491 | /* Function supportable_widening_operation | |
6492 | ||
b8698a0f L |
6493 | Check whether an operation represented by the code CODE is a |
6494 | widening operation that is supported by the target platform in | |
b690cc0f RG |
6495 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
6496 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 6497 | |
ebfd146a IR |
6498 | Widening operations we currently support are NOP (CONVERT), FLOAT |
6499 | and WIDEN_MULT. This function checks if these operations are supported | |
6500 | by the target platform either directly (via vector tree-codes), or via | |
6501 | target builtins. | |
6502 | ||
6503 | Output: | |
b8698a0f L |
6504 | - CODE1 and CODE2 are codes of vector operations to be used when |
6505 | vectorizing the operation, if available. | |
ebfd146a IR |
6506 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
6507 | case of multi-step conversion (like char->short->int - in that case | |
6508 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
6509 | - INTERM_TYPES contains the intermediate type required to perform the |
6510 | widening operation (short in the above example). */ | |
ebfd146a IR |
6511 | |
6512 | bool | |
b690cc0f RG |
6513 | supportable_widening_operation (enum tree_code code, gimple stmt, |
6514 | tree vectype_out, tree vectype_in, | |
ebfd146a IR |
6515 | enum tree_code *code1, enum tree_code *code2, |
6516 | int *multi_step_cvt, | |
9771b263 | 6517 | vec<tree> *interm_types) |
ebfd146a IR |
6518 | { |
6519 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6520 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 6521 | struct loop *vect_loop = NULL; |
ebfd146a | 6522 | enum machine_mode vec_mode; |
81f40b79 | 6523 | enum insn_code icode1, icode2; |
ebfd146a | 6524 | optab optab1, optab2; |
b690cc0f RG |
6525 | tree vectype = vectype_in; |
6526 | tree wide_vectype = vectype_out; | |
ebfd146a | 6527 | enum tree_code c1, c2; |
4a00c761 JJ |
6528 | int i; |
6529 | tree prev_type, intermediate_type; | |
6530 | enum machine_mode intermediate_mode, prev_mode; | |
6531 | optab optab3, optab4; | |
ebfd146a | 6532 | |
4a00c761 | 6533 | *multi_step_cvt = 0; |
4ef69dfc IR |
6534 | if (loop_info) |
6535 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
6536 | ||
ebfd146a IR |
6537 | switch (code) |
6538 | { | |
6539 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
6540 | /* The result of a vectorized widening operation usually requires |
6541 | two vectors (because the widened results do not fit into one vector). | |
6542 | The generated vector results would normally be expected to be | |
6543 | generated in the same order as in the original scalar computation, | |
6544 | i.e. if 8 results are generated in each vector iteration, they are | |
6545 | to be organized as follows: | |
6546 | vect1: [res1,res2,res3,res4], | |
6547 | vect2: [res5,res6,res7,res8]. | |
6548 | ||
6549 | However, in the special case that the result of the widening | |
6550 | operation is used in a reduction computation only, the order doesn't | |
6551 | matter (because when vectorizing a reduction we change the order of | |
6552 | the computation). Some targets can take advantage of this and | |
6553 | generate more efficient code. For example, targets like Altivec, | |
6554 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
6555 | generate the following vectors: | |
6556 | vect1: [res1,res3,res5,res7], | |
6557 | vect2: [res2,res4,res6,res8]. | |
6558 | ||
6559 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
6560 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
6561 | iterations in parallel). We therefore don't allow to change the | |
6562 | order of the computation in the inner-loop during outer-loop | |
6563 | vectorization. */ | |
6564 | /* TODO: Another case in which order doesn't *really* matter is when we | |
6565 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
6566 | Normally, pack_trunc performs an even/odd permute, whereas the | |
6567 | repack from an even/odd expansion would be an interleave, which | |
6568 | would be significantly simpler for e.g. AVX2. */ | |
6569 | /* In any case, in order to avoid duplicating the code below, recurse | |
6570 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
6571 | are properly set up for the caller. If we fail, we'll continue with | |
6572 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
6573 | if (vect_loop | |
6574 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
6575 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
6576 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
6577 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
6578 | code1, code2, multi_step_cvt, |
6579 | interm_types)) | |
6ae6116f | 6580 | return true; |
4a00c761 JJ |
6581 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
6582 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
6583 | break; |
6584 | ||
6ae6116f RH |
6585 | case VEC_WIDEN_MULT_EVEN_EXPR: |
6586 | /* Support the recursion induced just above. */ | |
6587 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
6588 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
6589 | break; | |
6590 | ||
36ba4aae | 6591 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
6592 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
6593 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
6594 | break; |
6595 | ||
ebfd146a | 6596 | CASE_CONVERT: |
4a00c761 JJ |
6597 | c1 = VEC_UNPACK_LO_EXPR; |
6598 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
6599 | break; |
6600 | ||
6601 | case FLOAT_EXPR: | |
4a00c761 JJ |
6602 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
6603 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
6604 | break; |
6605 | ||
6606 | case FIX_TRUNC_EXPR: | |
6607 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
6608 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
6609 | computing the operation. */ | |
6610 | return false; | |
6611 | ||
6612 | default: | |
6613 | gcc_unreachable (); | |
6614 | } | |
6615 | ||
6ae6116f | 6616 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
4a00c761 JJ |
6617 | { |
6618 | enum tree_code ctmp = c1; | |
6619 | c1 = c2; | |
6620 | c2 = ctmp; | |
6621 | } | |
6622 | ||
ebfd146a IR |
6623 | if (code == FIX_TRUNC_EXPR) |
6624 | { | |
6625 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
6626 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
6627 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
6628 | } |
6629 | else | |
6630 | { | |
6631 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
6632 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
6633 | } | |
6634 | ||
6635 | if (!optab1 || !optab2) | |
6636 | return false; | |
6637 | ||
6638 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
6639 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
6640 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
6641 | return false; |
6642 | ||
4a00c761 JJ |
6643 | *code1 = c1; |
6644 | *code2 = c2; | |
6645 | ||
6646 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
6647 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
6648 | return true; | |
6649 | ||
b8698a0f | 6650 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 6651 | types. */ |
ebfd146a | 6652 | |
4a00c761 JJ |
6653 | prev_type = vectype; |
6654 | prev_mode = vec_mode; | |
b8698a0f | 6655 | |
4a00c761 JJ |
6656 | if (!CONVERT_EXPR_CODE_P (code)) |
6657 | return false; | |
b8698a0f | 6658 | |
4a00c761 JJ |
6659 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
6660 | intermediate steps in promotion sequence. We try | |
6661 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
6662 | not. */ | |
9771b263 | 6663 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
6664 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
6665 | { | |
6666 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
6667 | intermediate_type | |
6668 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
6669 | TYPE_UNSIGNED (prev_type)); | |
6670 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
6671 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
6672 | ||
6673 | if (!optab3 || !optab4 | |
6674 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
6675 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
6676 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
6677 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
6678 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
6679 | == CODE_FOR_nothing) | |
6680 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
6681 | == CODE_FOR_nothing)) | |
6682 | break; | |
ebfd146a | 6683 | |
9771b263 | 6684 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
6685 | (*multi_step_cvt)++; |
6686 | ||
6687 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
6688 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
6689 | return true; | |
6690 | ||
6691 | prev_type = intermediate_type; | |
6692 | prev_mode = intermediate_mode; | |
ebfd146a IR |
6693 | } |
6694 | ||
9771b263 | 6695 | interm_types->release (); |
4a00c761 | 6696 | return false; |
ebfd146a IR |
6697 | } |
6698 | ||
6699 | ||
6700 | /* Function supportable_narrowing_operation | |
6701 | ||
b8698a0f L |
6702 | Check whether an operation represented by the code CODE is a |
6703 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
6704 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
6705 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 6706 | |
ebfd146a | 6707 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 6708 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
6709 | the target platform directly via vector tree-codes. |
6710 | ||
6711 | Output: | |
b8698a0f L |
6712 | - CODE1 is the code of a vector operation to be used when |
6713 | vectorizing the operation, if available. | |
ebfd146a IR |
6714 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
6715 | case of multi-step conversion (like int->short->char - in that case | |
6716 | MULTI_STEP_CVT will be 1). | |
6717 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 6718 | narrowing operation (short in the above example). */ |
ebfd146a IR |
6719 | |
6720 | bool | |
6721 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 6722 | tree vectype_out, tree vectype_in, |
ebfd146a | 6723 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 6724 | vec<tree> *interm_types) |
ebfd146a IR |
6725 | { |
6726 | enum machine_mode vec_mode; | |
6727 | enum insn_code icode1; | |
6728 | optab optab1, interm_optab; | |
b690cc0f RG |
6729 | tree vectype = vectype_in; |
6730 | tree narrow_vectype = vectype_out; | |
ebfd146a | 6731 | enum tree_code c1; |
4a00c761 JJ |
6732 | tree intermediate_type; |
6733 | enum machine_mode intermediate_mode, prev_mode; | |
ebfd146a | 6734 | int i; |
4a00c761 | 6735 | bool uns; |
ebfd146a | 6736 | |
4a00c761 | 6737 | *multi_step_cvt = 0; |
ebfd146a IR |
6738 | switch (code) |
6739 | { | |
6740 | CASE_CONVERT: | |
6741 | c1 = VEC_PACK_TRUNC_EXPR; | |
6742 | break; | |
6743 | ||
6744 | case FIX_TRUNC_EXPR: | |
6745 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
6746 | break; | |
6747 | ||
6748 | case FLOAT_EXPR: | |
6749 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
6750 | tree code and optabs used for computing the operation. */ | |
6751 | return false; | |
6752 | ||
6753 | default: | |
6754 | gcc_unreachable (); | |
6755 | } | |
6756 | ||
6757 | if (code == FIX_TRUNC_EXPR) | |
6758 | /* The signedness is determined from output operand. */ | |
b690cc0f | 6759 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
6760 | else |
6761 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
6762 | ||
6763 | if (!optab1) | |
6764 | return false; | |
6765 | ||
6766 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 6767 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
6768 | return false; |
6769 | ||
4a00c761 JJ |
6770 | *code1 = c1; |
6771 | ||
6772 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
6773 | return true; | |
6774 | ||
ebfd146a IR |
6775 | /* Check if it's a multi-step conversion that can be done using intermediate |
6776 | types. */ | |
4a00c761 JJ |
6777 | prev_mode = vec_mode; |
6778 | if (code == FIX_TRUNC_EXPR) | |
6779 | uns = TYPE_UNSIGNED (vectype_out); | |
6780 | else | |
6781 | uns = TYPE_UNSIGNED (vectype); | |
6782 | ||
6783 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
6784 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
6785 | costly than signed. */ | |
6786 | if (code == FIX_TRUNC_EXPR && uns) | |
6787 | { | |
6788 | enum insn_code icode2; | |
6789 | ||
6790 | intermediate_type | |
6791 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
6792 | interm_optab | |
6793 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 6794 | if (interm_optab != unknown_optab |
4a00c761 JJ |
6795 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
6796 | && insn_data[icode1].operand[0].mode | |
6797 | == insn_data[icode2].operand[0].mode) | |
6798 | { | |
6799 | uns = false; | |
6800 | optab1 = interm_optab; | |
6801 | icode1 = icode2; | |
6802 | } | |
6803 | } | |
ebfd146a | 6804 | |
4a00c761 JJ |
6805 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
6806 | intermediate steps in promotion sequence. We try | |
6807 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 6808 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
6809 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
6810 | { | |
6811 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
6812 | intermediate_type | |
6813 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
6814 | interm_optab | |
6815 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
6816 | optab_default); | |
6817 | if (!interm_optab | |
6818 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
6819 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
6820 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
6821 | == CODE_FOR_nothing)) | |
6822 | break; | |
6823 | ||
9771b263 | 6824 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
6825 | (*multi_step_cvt)++; |
6826 | ||
6827 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
6828 | return true; | |
6829 | ||
6830 | prev_mode = intermediate_mode; | |
6831 | optab1 = interm_optab; | |
ebfd146a IR |
6832 | } |
6833 | ||
9771b263 | 6834 | interm_types->release (); |
4a00c761 | 6835 | return false; |
ebfd146a | 6836 | } |