]>
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" |
45b0be94 | 32 | #include "gimplify.h" |
5be5c238 | 33 | #include "gimple-iterator.h" |
442b4905 AM |
34 | #include "gimple-ssa.h" |
35 | #include "tree-cfg.h" | |
36 | #include "tree-phinodes.h" | |
37 | #include "ssa-iterators.h" | |
38 | #include "tree-ssanames.h" | |
e28030cf | 39 | #include "tree-ssa-loop-manip.h" |
ebfd146a | 40 | #include "cfgloop.h" |
ebfd146a | 41 | #include "expr.h" |
7ee2468b | 42 | #include "recog.h" /* FIXME: for insn_data */ |
ebfd146a | 43 | #include "optabs.h" |
718f9c0f | 44 | #include "diagnostic-core.h" |
ebfd146a | 45 | #include "tree-vectorizer.h" |
7ee2468b | 46 | #include "dumpfile.h" |
ebfd146a | 47 | |
7ee2468b SB |
48 | /* For lang_hooks.types.type_for_mode. */ |
49 | #include "langhooks.h" | |
ebfd146a | 50 | |
c3e7ee41 BS |
51 | /* Return the vectorized type for the given statement. */ |
52 | ||
53 | tree | |
54 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
55 | { | |
56 | return STMT_VINFO_VECTYPE (stmt_info); | |
57 | } | |
58 | ||
59 | /* Return TRUE iff the given statement is in an inner loop relative to | |
60 | the loop being vectorized. */ | |
61 | bool | |
62 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
63 | { | |
64 | gimple stmt = STMT_VINFO_STMT (stmt_info); | |
65 | basic_block bb = gimple_bb (stmt); | |
66 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
67 | struct loop* loop; | |
68 | ||
69 | if (!loop_vinfo) | |
70 | return false; | |
71 | ||
72 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
73 | ||
74 | return (bb->loop_father == loop->inner); | |
75 | } | |
76 | ||
77 | /* Record the cost of a statement, either by directly informing the | |
78 | target model or by saving it in a vector for later processing. | |
79 | Return a preliminary estimate of the statement's cost. */ | |
80 | ||
81 | unsigned | |
92345349 | 82 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 83 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 84 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 85 | { |
92345349 | 86 | if (body_cost_vec) |
c3e7ee41 | 87 | { |
92345349 BS |
88 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
89 | add_stmt_info_to_vec (body_cost_vec, count, kind, | |
90 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
91 | misalign); | |
c3e7ee41 | 92 | return (unsigned) |
92345349 | 93 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
94 | |
95 | } | |
96 | else | |
97 | { | |
98 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
99 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
100 | void *target_cost_data; | |
101 | ||
102 | if (loop_vinfo) | |
103 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
104 | else | |
105 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
106 | ||
92345349 BS |
107 | return add_stmt_cost (target_cost_data, count, kind, stmt_info, |
108 | misalign, where); | |
c3e7ee41 BS |
109 | } |
110 | } | |
111 | ||
272c6793 RS |
112 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
113 | ||
114 | static tree | |
115 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
116 | { | |
117 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
118 | "vect_array"); | |
119 | } | |
120 | ||
121 | /* ARRAY is an array of vectors created by create_vector_array. | |
122 | Return an SSA_NAME for the vector in index N. The reference | |
123 | is part of the vectorization of STMT and the vector is associated | |
124 | with scalar destination SCALAR_DEST. */ | |
125 | ||
126 | static tree | |
127 | read_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree scalar_dest, | |
128 | tree array, unsigned HOST_WIDE_INT n) | |
129 | { | |
130 | tree vect_type, vect, vect_name, array_ref; | |
131 | gimple new_stmt; | |
132 | ||
133 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
134 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
135 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
136 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
137 | build_int_cst (size_type_node, n), | |
138 | NULL_TREE, NULL_TREE); | |
139 | ||
140 | new_stmt = gimple_build_assign (vect, array_ref); | |
141 | vect_name = make_ssa_name (vect, new_stmt); | |
142 | gimple_assign_set_lhs (new_stmt, vect_name); | |
143 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
144 | |
145 | return vect_name; | |
146 | } | |
147 | ||
148 | /* ARRAY is an array of vectors created by create_vector_array. | |
149 | Emit code to store SSA_NAME VECT in index N of the array. | |
150 | The store is part of the vectorization of STMT. */ | |
151 | ||
152 | static void | |
153 | write_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree vect, | |
154 | tree array, unsigned HOST_WIDE_INT n) | |
155 | { | |
156 | tree array_ref; | |
157 | gimple new_stmt; | |
158 | ||
159 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
160 | build_int_cst (size_type_node, n), | |
161 | NULL_TREE, NULL_TREE); | |
162 | ||
163 | new_stmt = gimple_build_assign (array_ref, vect); | |
164 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
165 | } |
166 | ||
167 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
168 | of *PTR. The memory reference replaces those in FIRST_DR | |
169 | (and its group). */ | |
170 | ||
171 | static tree | |
172 | create_array_ref (tree type, tree ptr, struct data_reference *first_dr) | |
173 | { | |
272c6793 RS |
174 | tree mem_ref, alias_ptr_type; |
175 | ||
176 | alias_ptr_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
177 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); | |
178 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 179 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
180 | return mem_ref; |
181 | } | |
182 | ||
ebfd146a IR |
183 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
184 | ||
185 | /* Function vect_mark_relevant. | |
186 | ||
187 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
188 | ||
189 | static void | |
9771b263 | 190 | vect_mark_relevant (vec<gimple> *worklist, gimple stmt, |
83197f37 IR |
191 | enum vect_relevant relevant, bool live_p, |
192 | bool used_in_pattern) | |
ebfd146a IR |
193 | { |
194 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
195 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
196 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
83197f37 | 197 | gimple pattern_stmt; |
ebfd146a | 198 | |
73fbfcad | 199 | if (dump_enabled_p ()) |
78c60e3d | 200 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 201 | "mark relevant %d, live %d.\n", relevant, live_p); |
ebfd146a | 202 | |
83197f37 IR |
203 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
204 | related pattern stmt instead of the original stmt. However, such stmts | |
205 | may have their own uses that are not in any pattern, in such cases the | |
206 | stmt itself should be marked. */ | |
ebfd146a IR |
207 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
208 | { | |
83197f37 IR |
209 | bool found = false; |
210 | if (!used_in_pattern) | |
211 | { | |
212 | imm_use_iterator imm_iter; | |
213 | use_operand_p use_p; | |
214 | gimple use_stmt; | |
215 | tree lhs; | |
13c931c9 JJ |
216 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
217 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a | 218 | |
83197f37 IR |
219 | if (is_gimple_assign (stmt)) |
220 | lhs = gimple_assign_lhs (stmt); | |
221 | else | |
222 | lhs = gimple_call_lhs (stmt); | |
ebfd146a | 223 | |
83197f37 IR |
224 | /* This use is out of pattern use, if LHS has other uses that are |
225 | pattern uses, we should mark the stmt itself, and not the pattern | |
226 | stmt. */ | |
ab0ef706 JJ |
227 | if (TREE_CODE (lhs) == SSA_NAME) |
228 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs) | |
229 | { | |
230 | if (is_gimple_debug (USE_STMT (use_p))) | |
231 | continue; | |
232 | use_stmt = USE_STMT (use_p); | |
233 | ||
13c931c9 JJ |
234 | if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) |
235 | continue; | |
236 | ||
ab0ef706 JJ |
237 | if (vinfo_for_stmt (use_stmt) |
238 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt))) | |
239 | { | |
240 | found = true; | |
241 | break; | |
242 | } | |
243 | } | |
83197f37 IR |
244 | } |
245 | ||
246 | if (!found) | |
247 | { | |
248 | /* This is the last stmt in a sequence that was detected as a | |
249 | pattern that can potentially be vectorized. Don't mark the stmt | |
250 | as relevant/live because it's not going to be vectorized. | |
251 | Instead mark the pattern-stmt that replaces it. */ | |
252 | ||
253 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
254 | ||
73fbfcad | 255 | if (dump_enabled_p ()) |
78c60e3d SS |
256 | dump_printf_loc (MSG_NOTE, vect_location, |
257 | "last stmt in pattern. don't mark" | |
e645e942 | 258 | " relevant/live.\n"); |
83197f37 IR |
259 | stmt_info = vinfo_for_stmt (pattern_stmt); |
260 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
261 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
262 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
263 | stmt = pattern_stmt; | |
264 | } | |
ebfd146a IR |
265 | } |
266 | ||
267 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
268 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
269 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
270 | ||
271 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
272 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
273 | { | |
73fbfcad | 274 | if (dump_enabled_p ()) |
78c60e3d | 275 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 276 | "already marked relevant/live.\n"); |
ebfd146a IR |
277 | return; |
278 | } | |
279 | ||
9771b263 | 280 | worklist->safe_push (stmt); |
ebfd146a IR |
281 | } |
282 | ||
283 | ||
284 | /* Function vect_stmt_relevant_p. | |
285 | ||
286 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
287 | "relevant for vectorization". | |
288 | ||
289 | A stmt is considered "relevant for vectorization" if: | |
290 | - it has uses outside the loop. | |
291 | - it has vdefs (it alters memory). | |
292 | - control stmts in the loop (except for the exit condition). | |
293 | ||
294 | CHECKME: what other side effects would the vectorizer allow? */ | |
295 | ||
296 | static bool | |
297 | vect_stmt_relevant_p (gimple stmt, loop_vec_info loop_vinfo, | |
298 | enum vect_relevant *relevant, bool *live_p) | |
299 | { | |
300 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
301 | ssa_op_iter op_iter; | |
302 | imm_use_iterator imm_iter; | |
303 | use_operand_p use_p; | |
304 | def_operand_p def_p; | |
305 | ||
8644a673 | 306 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
307 | *live_p = false; |
308 | ||
309 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
310 | if (is_ctrl_stmt (stmt) |
311 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
312 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 313 | *relevant = vect_used_in_scope; |
ebfd146a IR |
314 | |
315 | /* changing memory. */ | |
316 | if (gimple_code (stmt) != GIMPLE_PHI) | |
5006671f | 317 | if (gimple_vdef (stmt)) |
ebfd146a | 318 | { |
73fbfcad | 319 | if (dump_enabled_p ()) |
78c60e3d | 320 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 321 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 322 | *relevant = vect_used_in_scope; |
ebfd146a IR |
323 | } |
324 | ||
325 | /* uses outside the loop. */ | |
326 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
327 | { | |
328 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
329 | { | |
330 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
331 | if (!flow_bb_inside_loop_p (loop, bb)) | |
332 | { | |
73fbfcad | 333 | if (dump_enabled_p ()) |
78c60e3d | 334 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 335 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 336 | |
3157b0c2 AO |
337 | if (is_gimple_debug (USE_STMT (use_p))) |
338 | continue; | |
339 | ||
ebfd146a IR |
340 | /* We expect all such uses to be in the loop exit phis |
341 | (because of loop closed form) */ | |
342 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
343 | gcc_assert (bb == single_exit (loop)->dest); | |
344 | ||
345 | *live_p = true; | |
346 | } | |
347 | } | |
348 | } | |
349 | ||
350 | return (*live_p || *relevant); | |
351 | } | |
352 | ||
353 | ||
b8698a0f | 354 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 355 | |
ff802fa1 | 356 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
357 | used in STMT for anything other than indexing an array. */ |
358 | ||
359 | static bool | |
360 | exist_non_indexing_operands_for_use_p (tree use, gimple stmt) | |
361 | { | |
362 | tree operand; | |
363 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 364 | |
ff802fa1 | 365 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
366 | reference in STMT, then any operand that corresponds to USE |
367 | is not indexing an array. */ | |
368 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
369 | return true; | |
59a05b0c | 370 | |
ebfd146a IR |
371 | /* STMT has a data_ref. FORNOW this means that its of one of |
372 | the following forms: | |
373 | -1- ARRAY_REF = var | |
374 | -2- var = ARRAY_REF | |
375 | (This should have been verified in analyze_data_refs). | |
376 | ||
377 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 378 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
379 | for array indexing. |
380 | ||
381 | Therefore, all we need to check is if STMT falls into the | |
382 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
383 | |
384 | if (!gimple_assign_copy_p (stmt)) | |
385 | return false; | |
59a05b0c EB |
386 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
387 | return false; | |
ebfd146a | 388 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
389 | if (TREE_CODE (operand) != SSA_NAME) |
390 | return false; | |
391 | ||
392 | if (operand == use) | |
393 | return true; | |
394 | ||
395 | return false; | |
396 | } | |
397 | ||
398 | ||
b8698a0f | 399 | /* |
ebfd146a IR |
400 | Function process_use. |
401 | ||
402 | Inputs: | |
403 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 404 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ff802fa1 | 405 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 406 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
407 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
408 | be performed. | |
ebfd146a IR |
409 | |
410 | Outputs: | |
411 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
412 | relevance info of the DEF_STMT of this USE: | |
413 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
414 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
415 | Exceptions: | |
416 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 417 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 418 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
419 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
420 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
421 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
422 | be modified accordingly. | |
423 | ||
424 | Return true if everything is as expected. Return false otherwise. */ | |
425 | ||
426 | static bool | |
b8698a0f | 427 | process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
9771b263 | 428 | enum vect_relevant relevant, vec<gimple> *worklist, |
aec7ae7d | 429 | bool force) |
ebfd146a IR |
430 | { |
431 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
432 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
433 | stmt_vec_info dstmt_vinfo; | |
434 | basic_block bb, def_bb; | |
435 | tree def; | |
436 | gimple def_stmt; | |
437 | enum vect_def_type dt; | |
438 | ||
b8698a0f | 439 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 440 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 441 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
442 | return true; |
443 | ||
24ee1384 | 444 | if (!vect_is_simple_use (use, stmt, loop_vinfo, NULL, &def_stmt, &def, &dt)) |
b8698a0f | 445 | { |
73fbfcad | 446 | if (dump_enabled_p ()) |
78c60e3d | 447 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 448 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
449 | return false; |
450 | } | |
451 | ||
452 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
453 | return true; | |
454 | ||
455 | def_bb = gimple_bb (def_stmt); | |
456 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
457 | { | |
73fbfcad | 458 | if (dump_enabled_p ()) |
e645e942 | 459 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
460 | return true; |
461 | } | |
462 | ||
b8698a0f L |
463 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
464 | DEF_STMT must have already been processed, because this should be the | |
465 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
466 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
467 | check that everything is as expected, and we are done. */ |
468 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
469 | bb = gimple_bb (stmt); | |
470 | if (gimple_code (stmt) == GIMPLE_PHI | |
471 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
472 | && gimple_code (def_stmt) != GIMPLE_PHI | |
473 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
474 | && bb->loop_father == def_bb->loop_father) | |
475 | { | |
73fbfcad | 476 | if (dump_enabled_p ()) |
78c60e3d | 477 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 478 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
479 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
480 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
481 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 482 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 483 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
484 | return true; |
485 | } | |
486 | ||
487 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
488 | outer-loop-header-bb: | |
489 | d = def_stmt | |
490 | inner-loop: | |
491 | stmt # use (d) | |
492 | outer-loop-tail-bb: | |
493 | ... */ | |
494 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
495 | { | |
73fbfcad | 496 | if (dump_enabled_p ()) |
78c60e3d | 497 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 498 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 499 | |
ebfd146a IR |
500 | switch (relevant) |
501 | { | |
8644a673 | 502 | case vect_unused_in_scope: |
7c5222ff IR |
503 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
504 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 505 | break; |
7c5222ff | 506 | |
ebfd146a | 507 | case vect_used_in_outer_by_reduction: |
7c5222ff | 508 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
509 | relevant = vect_used_by_reduction; |
510 | break; | |
7c5222ff | 511 | |
ebfd146a | 512 | case vect_used_in_outer: |
7c5222ff | 513 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 514 | relevant = vect_used_in_scope; |
ebfd146a | 515 | break; |
7c5222ff | 516 | |
8644a673 | 517 | case vect_used_in_scope: |
ebfd146a IR |
518 | break; |
519 | ||
520 | default: | |
521 | gcc_unreachable (); | |
b8698a0f | 522 | } |
ebfd146a IR |
523 | } |
524 | ||
525 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
526 | outer-loop-header-bb: | |
527 | ... | |
528 | inner-loop: | |
529 | d = def_stmt | |
06066f92 | 530 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
531 | stmt # use (d) */ |
532 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
533 | { | |
73fbfcad | 534 | if (dump_enabled_p ()) |
78c60e3d | 535 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 536 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 537 | |
ebfd146a IR |
538 | switch (relevant) |
539 | { | |
8644a673 | 540 | case vect_unused_in_scope: |
b8698a0f | 541 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 542 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 543 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
544 | break; |
545 | ||
ebfd146a IR |
546 | case vect_used_by_reduction: |
547 | relevant = vect_used_in_outer_by_reduction; | |
548 | break; | |
549 | ||
8644a673 | 550 | case vect_used_in_scope: |
ebfd146a IR |
551 | relevant = vect_used_in_outer; |
552 | break; | |
553 | ||
554 | default: | |
555 | gcc_unreachable (); | |
556 | } | |
557 | } | |
558 | ||
83197f37 IR |
559 | vect_mark_relevant (worklist, def_stmt, relevant, live_p, |
560 | is_pattern_stmt_p (stmt_vinfo)); | |
ebfd146a IR |
561 | return true; |
562 | } | |
563 | ||
564 | ||
565 | /* Function vect_mark_stmts_to_be_vectorized. | |
566 | ||
567 | Not all stmts in the loop need to be vectorized. For example: | |
568 | ||
569 | for i... | |
570 | for j... | |
571 | 1. T0 = i + j | |
572 | 2. T1 = a[T0] | |
573 | ||
574 | 3. j = j + 1 | |
575 | ||
576 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
577 | addressing of vectorized data-refs are handled differently. | |
578 | ||
579 | This pass detects such stmts. */ | |
580 | ||
581 | bool | |
582 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
583 | { | |
ebfd146a IR |
584 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
585 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
586 | unsigned int nbbs = loop->num_nodes; | |
587 | gimple_stmt_iterator si; | |
588 | gimple stmt; | |
589 | unsigned int i; | |
590 | stmt_vec_info stmt_vinfo; | |
591 | basic_block bb; | |
592 | gimple phi; | |
593 | bool live_p; | |
06066f92 IR |
594 | enum vect_relevant relevant, tmp_relevant; |
595 | enum vect_def_type def_type; | |
ebfd146a | 596 | |
73fbfcad | 597 | if (dump_enabled_p ()) |
78c60e3d | 598 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 599 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 600 | |
07687835 | 601 | stack_vec<gimple, 64> worklist; |
ebfd146a IR |
602 | |
603 | /* 1. Init worklist. */ | |
604 | for (i = 0; i < nbbs; i++) | |
605 | { | |
606 | bb = bbs[i]; | |
607 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 608 | { |
ebfd146a | 609 | phi = gsi_stmt (si); |
73fbfcad | 610 | if (dump_enabled_p ()) |
ebfd146a | 611 | { |
78c60e3d SS |
612 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
613 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
e645e942 | 614 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
615 | } |
616 | ||
617 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 618 | vect_mark_relevant (&worklist, phi, relevant, live_p, false); |
ebfd146a IR |
619 | } |
620 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
621 | { | |
622 | stmt = gsi_stmt (si); | |
73fbfcad | 623 | if (dump_enabled_p ()) |
ebfd146a | 624 | { |
78c60e3d SS |
625 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
626 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 627 | dump_printf (MSG_NOTE, "\n"); |
b8698a0f | 628 | } |
ebfd146a IR |
629 | |
630 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 631 | vect_mark_relevant (&worklist, stmt, relevant, live_p, false); |
ebfd146a IR |
632 | } |
633 | } | |
634 | ||
635 | /* 2. Process_worklist */ | |
9771b263 | 636 | while (worklist.length () > 0) |
ebfd146a IR |
637 | { |
638 | use_operand_p use_p; | |
639 | ssa_op_iter iter; | |
640 | ||
9771b263 | 641 | stmt = worklist.pop (); |
73fbfcad | 642 | if (dump_enabled_p ()) |
ebfd146a | 643 | { |
78c60e3d SS |
644 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
645 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 646 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
647 | } |
648 | ||
b8698a0f L |
649 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
650 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
651 | liveness and relevance properties of STMT. */ |
652 | stmt_vinfo = vinfo_for_stmt (stmt); | |
653 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
654 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
655 | ||
656 | /* Generally, the liveness and relevance properties of STMT are | |
657 | propagated as is to the DEF_STMTs of its USEs: | |
658 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
659 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
660 | ||
661 | One exception is when STMT has been identified as defining a reduction | |
662 | variable; in this case we set the liveness/relevance as follows: | |
663 | live_p = false | |
664 | relevant = vect_used_by_reduction | |
665 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f | 666 | those that are used by a reduction computation, and those that are |
ff802fa1 | 667 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 668 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 669 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 670 | |
06066f92 IR |
671 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
672 | tmp_relevant = relevant; | |
673 | switch (def_type) | |
ebfd146a | 674 | { |
06066f92 IR |
675 | case vect_reduction_def: |
676 | switch (tmp_relevant) | |
677 | { | |
678 | case vect_unused_in_scope: | |
679 | relevant = vect_used_by_reduction; | |
680 | break; | |
681 | ||
682 | case vect_used_by_reduction: | |
683 | if (gimple_code (stmt) == GIMPLE_PHI) | |
684 | break; | |
685 | /* fall through */ | |
686 | ||
687 | default: | |
73fbfcad | 688 | if (dump_enabled_p ()) |
78c60e3d | 689 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 690 | "unsupported use of reduction.\n"); |
06066f92 IR |
691 | return false; |
692 | } | |
693 | ||
b8698a0f | 694 | live_p = false; |
06066f92 | 695 | break; |
b8698a0f | 696 | |
06066f92 IR |
697 | case vect_nested_cycle: |
698 | if (tmp_relevant != vect_unused_in_scope | |
699 | && tmp_relevant != vect_used_in_outer_by_reduction | |
700 | && tmp_relevant != vect_used_in_outer) | |
701 | { | |
73fbfcad | 702 | if (dump_enabled_p ()) |
78c60e3d | 703 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 704 | "unsupported use of nested cycle.\n"); |
7c5222ff | 705 | |
06066f92 IR |
706 | return false; |
707 | } | |
7c5222ff | 708 | |
b8698a0f L |
709 | live_p = false; |
710 | break; | |
711 | ||
06066f92 IR |
712 | case vect_double_reduction_def: |
713 | if (tmp_relevant != vect_unused_in_scope | |
714 | && tmp_relevant != vect_used_by_reduction) | |
715 | { | |
73fbfcad | 716 | if (dump_enabled_p ()) |
78c60e3d | 717 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 718 | "unsupported use of double reduction.\n"); |
7c5222ff | 719 | |
7c5222ff | 720 | return false; |
06066f92 IR |
721 | } |
722 | ||
723 | live_p = false; | |
b8698a0f | 724 | break; |
7c5222ff | 725 | |
06066f92 IR |
726 | default: |
727 | break; | |
7c5222ff | 728 | } |
b8698a0f | 729 | |
aec7ae7d | 730 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
731 | { |
732 | /* Pattern statements are not inserted into the code, so | |
733 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
734 | have to scan the RHS or function arguments instead. */ | |
735 | if (is_gimple_assign (stmt)) | |
736 | { | |
69d2aade JJ |
737 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
738 | tree op = gimple_assign_rhs1 (stmt); | |
739 | ||
740 | i = 1; | |
741 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
742 | { | |
743 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
aec7ae7d | 744 | live_p, relevant, &worklist, false) |
69d2aade | 745 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
aec7ae7d | 746 | live_p, relevant, &worklist, false)) |
07687835 | 747 | return false; |
69d2aade JJ |
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)) |
07687835 | 755 | return false; |
9d5e7640 IR |
756 | } |
757 | } | |
758 | else if (is_gimple_call (stmt)) | |
759 | { | |
760 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
761 | { | |
762 | tree arg = gimple_call_arg (stmt, i); | |
763 | if (!process_use (stmt, arg, loop_vinfo, live_p, relevant, | |
aec7ae7d | 764 | &worklist, false)) |
07687835 | 765 | return false; |
9d5e7640 IR |
766 | } |
767 | } | |
768 | } | |
769 | else | |
770 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
771 | { | |
772 | tree op = USE_FROM_PTR (use_p); | |
773 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, | |
aec7ae7d | 774 | &worklist, false)) |
07687835 | 775 | return false; |
9d5e7640 | 776 | } |
aec7ae7d JJ |
777 | |
778 | if (STMT_VINFO_GATHER_P (stmt_vinfo)) | |
779 | { | |
780 | tree off; | |
781 | tree decl = vect_check_gather (stmt, loop_vinfo, NULL, &off, NULL); | |
782 | gcc_assert (decl); | |
783 | if (!process_use (stmt, off, loop_vinfo, live_p, relevant, | |
784 | &worklist, true)) | |
07687835 | 785 | return false; |
aec7ae7d | 786 | } |
ebfd146a IR |
787 | } /* while worklist */ |
788 | ||
ebfd146a IR |
789 | return true; |
790 | } | |
791 | ||
792 | ||
b8698a0f | 793 | /* Function vect_model_simple_cost. |
ebfd146a | 794 | |
b8698a0f | 795 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
796 | single op. Right now, this does not account for multiple insns that could |
797 | be generated for the single vector op. We will handle that shortly. */ | |
798 | ||
799 | void | |
b8698a0f | 800 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 BS |
801 | enum vect_def_type *dt, |
802 | stmt_vector_for_cost *prologue_cost_vec, | |
803 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
804 | { |
805 | int i; | |
92345349 | 806 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
807 | |
808 | /* The SLP costs were already calculated during SLP tree build. */ | |
809 | if (PURE_SLP_STMT (stmt_info)) | |
810 | return; | |
811 | ||
ebfd146a IR |
812 | /* FORNOW: Assuming maximum 2 args per stmts. */ |
813 | for (i = 0; i < 2; i++) | |
92345349 BS |
814 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
815 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, vector_stmt, | |
816 | stmt_info, 0, vect_prologue); | |
c3e7ee41 BS |
817 | |
818 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
819 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
820 | stmt_info, 0, vect_body); | |
c3e7ee41 | 821 | |
73fbfcad | 822 | if (dump_enabled_p ()) |
78c60e3d SS |
823 | dump_printf_loc (MSG_NOTE, vect_location, |
824 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 825 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
826 | } |
827 | ||
828 | ||
8bd37302 BS |
829 | /* Model cost for type demotion and promotion operations. PWR is normally |
830 | zero for single-step promotions and demotions. It will be one if | |
831 | two-step promotion/demotion is required, and so on. Each additional | |
832 | step doubles the number of instructions required. */ | |
833 | ||
834 | static void | |
835 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
836 | enum vect_def_type *dt, int pwr) | |
837 | { | |
838 | int i, tmp; | |
92345349 | 839 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
840 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
841 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
842 | void *target_cost_data; | |
8bd37302 BS |
843 | |
844 | /* The SLP costs were already calculated during SLP tree build. */ | |
845 | if (PURE_SLP_STMT (stmt_info)) | |
846 | return; | |
847 | ||
c3e7ee41 BS |
848 | if (loop_vinfo) |
849 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
850 | else | |
851 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
852 | ||
8bd37302 BS |
853 | for (i = 0; i < pwr + 1; i++) |
854 | { | |
855 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
856 | (i + 1) : i; | |
c3e7ee41 | 857 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
858 | vec_promote_demote, stmt_info, 0, |
859 | vect_body); | |
8bd37302 BS |
860 | } |
861 | ||
862 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
863 | for (i = 0; i < 2; i++) | |
92345349 BS |
864 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
865 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
866 | stmt_info, 0, vect_prologue); | |
8bd37302 | 867 | |
73fbfcad | 868 | if (dump_enabled_p ()) |
78c60e3d SS |
869 | dump_printf_loc (MSG_NOTE, vect_location, |
870 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 871 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
872 | } |
873 | ||
0d0293ac | 874 | /* Function vect_cost_group_size |
b8698a0f | 875 | |
0d0293ac | 876 | For grouped load or store, return the group_size only if it is the first |
ebfd146a IR |
877 | load or store of a group, else return 1. This ensures that group size is |
878 | only returned once per group. */ | |
879 | ||
880 | static int | |
0d0293ac | 881 | vect_cost_group_size (stmt_vec_info stmt_info) |
ebfd146a | 882 | { |
e14c1050 | 883 | gimple first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
884 | |
885 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
e14c1050 | 886 | return GROUP_SIZE (stmt_info); |
ebfd146a IR |
887 | |
888 | return 1; | |
889 | } | |
890 | ||
891 | ||
892 | /* Function vect_model_store_cost | |
893 | ||
0d0293ac MM |
894 | Models cost for stores. In the case of grouped accesses, one access |
895 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
896 | |
897 | void | |
b8698a0f | 898 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
272c6793 | 899 | bool store_lanes_p, enum vect_def_type dt, |
92345349 BS |
900 | slp_tree slp_node, |
901 | stmt_vector_for_cost *prologue_cost_vec, | |
902 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
903 | { |
904 | int group_size; | |
92345349 | 905 | unsigned int inside_cost = 0, prologue_cost = 0; |
720f5239 IR |
906 | struct data_reference *first_dr; |
907 | gimple first_stmt; | |
ebfd146a IR |
908 | |
909 | /* The SLP costs were already calculated during SLP tree build. */ | |
910 | if (PURE_SLP_STMT (stmt_info)) | |
911 | return; | |
912 | ||
8644a673 | 913 | if (dt == vect_constant_def || dt == vect_external_def) |
92345349 BS |
914 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
915 | stmt_info, 0, vect_prologue); | |
ebfd146a | 916 | |
0d0293ac MM |
917 | /* Grouped access? */ |
918 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
720f5239 IR |
919 | { |
920 | if (slp_node) | |
921 | { | |
9771b263 | 922 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
720f5239 IR |
923 | group_size = 1; |
924 | } | |
925 | else | |
926 | { | |
e14c1050 | 927 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 928 | group_size = vect_cost_group_size (stmt_info); |
720f5239 IR |
929 | } |
930 | ||
931 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
932 | } | |
0d0293ac | 933 | /* Not a grouped access. */ |
ebfd146a | 934 | else |
720f5239 IR |
935 | { |
936 | group_size = 1; | |
937 | first_dr = STMT_VINFO_DATA_REF (stmt_info); | |
938 | } | |
ebfd146a | 939 | |
272c6793 | 940 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 941 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 RS |
942 | access is instead being provided by a permute-and-store operation, |
943 | include the cost of the permutes. */ | |
944 | if (!store_lanes_p && group_size > 1) | |
ebfd146a IR |
945 | { |
946 | /* Uses a high and low interleave operation for each needed permute. */ | |
c3e7ee41 BS |
947 | |
948 | int nstmts = ncopies * exact_log2 (group_size) * group_size; | |
92345349 BS |
949 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
950 | stmt_info, 0, vect_body); | |
ebfd146a | 951 | |
73fbfcad | 952 | if (dump_enabled_p ()) |
78c60e3d | 953 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 954 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 955 | group_size); |
ebfd146a IR |
956 | } |
957 | ||
958 | /* Costs of the stores. */ | |
92345349 | 959 | vect_get_store_cost (first_dr, ncopies, &inside_cost, body_cost_vec); |
ebfd146a | 960 | |
73fbfcad | 961 | if (dump_enabled_p ()) |
78c60e3d SS |
962 | dump_printf_loc (MSG_NOTE, vect_location, |
963 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 964 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
965 | } |
966 | ||
967 | ||
720f5239 IR |
968 | /* Calculate cost of DR's memory access. */ |
969 | void | |
970 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 971 | unsigned int *inside_cost, |
92345349 | 972 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
973 | { |
974 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
c3e7ee41 BS |
975 | gimple stmt = DR_STMT (dr); |
976 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
720f5239 IR |
977 | |
978 | switch (alignment_support_scheme) | |
979 | { | |
980 | case dr_aligned: | |
981 | { | |
92345349 BS |
982 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
983 | vector_store, stmt_info, 0, | |
984 | vect_body); | |
720f5239 | 985 | |
73fbfcad | 986 | if (dump_enabled_p ()) |
78c60e3d | 987 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 988 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
989 | break; |
990 | } | |
991 | ||
992 | case dr_unaligned_supported: | |
993 | { | |
720f5239 | 994 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 995 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 996 | unaligned_store, stmt_info, |
92345349 | 997 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 998 | if (dump_enabled_p ()) |
78c60e3d SS |
999 | dump_printf_loc (MSG_NOTE, vect_location, |
1000 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1001 | "hardware.\n"); |
720f5239 IR |
1002 | break; |
1003 | } | |
1004 | ||
38eec4c6 UW |
1005 | case dr_unaligned_unsupported: |
1006 | { | |
1007 | *inside_cost = VECT_MAX_COST; | |
1008 | ||
73fbfcad | 1009 | if (dump_enabled_p ()) |
78c60e3d | 1010 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1011 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1012 | break; |
1013 | } | |
1014 | ||
720f5239 IR |
1015 | default: |
1016 | gcc_unreachable (); | |
1017 | } | |
1018 | } | |
1019 | ||
1020 | ||
ebfd146a IR |
1021 | /* Function vect_model_load_cost |
1022 | ||
0d0293ac MM |
1023 | Models cost for loads. In the case of grouped accesses, the last access |
1024 | has the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1025 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1026 | access scheme chosen. */ |
1027 | ||
1028 | void | |
92345349 BS |
1029 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
1030 | bool load_lanes_p, slp_tree slp_node, | |
1031 | stmt_vector_for_cost *prologue_cost_vec, | |
1032 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
1033 | { |
1034 | int group_size; | |
ebfd146a IR |
1035 | gimple first_stmt; |
1036 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; | |
92345349 | 1037 | unsigned int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
1038 | |
1039 | /* The SLP costs were already calculated during SLP tree build. */ | |
1040 | if (PURE_SLP_STMT (stmt_info)) | |
1041 | return; | |
1042 | ||
0d0293ac | 1043 | /* Grouped accesses? */ |
e14c1050 | 1044 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 1045 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && first_stmt && !slp_node) |
ebfd146a | 1046 | { |
0d0293ac | 1047 | group_size = vect_cost_group_size (stmt_info); |
ebfd146a IR |
1048 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
1049 | } | |
0d0293ac | 1050 | /* Not a grouped access. */ |
ebfd146a IR |
1051 | else |
1052 | { | |
1053 | group_size = 1; | |
1054 | first_dr = dr; | |
1055 | } | |
1056 | ||
272c6793 | 1057 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1058 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 RS |
1059 | access is instead being provided by a load-and-permute operation, |
1060 | include the cost of the permutes. */ | |
1061 | if (!load_lanes_p && group_size > 1) | |
ebfd146a IR |
1062 | { |
1063 | /* Uses an even and odd extract operations for each needed permute. */ | |
c3e7ee41 | 1064 | int nstmts = ncopies * exact_log2 (group_size) * group_size; |
92345349 BS |
1065 | inside_cost += record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
1066 | stmt_info, 0, vect_body); | |
ebfd146a | 1067 | |
73fbfcad | 1068 | if (dump_enabled_p ()) |
e645e942 TJ |
1069 | dump_printf_loc (MSG_NOTE, vect_location, |
1070 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1071 | group_size); |
ebfd146a IR |
1072 | } |
1073 | ||
1074 | /* The loads themselves. */ | |
a82960aa RG |
1075 | if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
1076 | { | |
a21892ad BS |
1077 | /* N scalar loads plus gathering them into a vector. */ |
1078 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
92345349 | 1079 | inside_cost += record_stmt_cost (body_cost_vec, |
c3e7ee41 | 1080 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), |
92345349 BS |
1081 | scalar_load, stmt_info, 0, vect_body); |
1082 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, | |
1083 | stmt_info, 0, vect_body); | |
a82960aa RG |
1084 | } |
1085 | else | |
1086 | vect_get_load_cost (first_dr, ncopies, | |
1087 | ((!STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1088 | || group_size > 1 || slp_node), | |
92345349 BS |
1089 | &inside_cost, &prologue_cost, |
1090 | prologue_cost_vec, body_cost_vec, true); | |
720f5239 | 1091 | |
73fbfcad | 1092 | if (dump_enabled_p ()) |
78c60e3d SS |
1093 | dump_printf_loc (MSG_NOTE, vect_location, |
1094 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1095 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1096 | } |
1097 | ||
1098 | ||
1099 | /* Calculate cost of DR's memory access. */ | |
1100 | void | |
1101 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1102 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1103 | unsigned int *prologue_cost, |
1104 | stmt_vector_for_cost *prologue_cost_vec, | |
1105 | stmt_vector_for_cost *body_cost_vec, | |
1106 | bool record_prologue_costs) | |
720f5239 IR |
1107 | { |
1108 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
c3e7ee41 BS |
1109 | gimple stmt = DR_STMT (dr); |
1110 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
720f5239 IR |
1111 | |
1112 | switch (alignment_support_scheme) | |
ebfd146a IR |
1113 | { |
1114 | case dr_aligned: | |
1115 | { | |
92345349 BS |
1116 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1117 | stmt_info, 0, vect_body); | |
ebfd146a | 1118 | |
73fbfcad | 1119 | if (dump_enabled_p ()) |
78c60e3d | 1120 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1121 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1122 | |
1123 | break; | |
1124 | } | |
1125 | case dr_unaligned_supported: | |
1126 | { | |
720f5239 | 1127 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1128 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1129 | unaligned_load, stmt_info, |
92345349 | 1130 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1131 | |
73fbfcad | 1132 | if (dump_enabled_p ()) |
78c60e3d SS |
1133 | dump_printf_loc (MSG_NOTE, vect_location, |
1134 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1135 | "hardware.\n"); |
ebfd146a IR |
1136 | |
1137 | break; | |
1138 | } | |
1139 | case dr_explicit_realign: | |
1140 | { | |
92345349 BS |
1141 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1142 | vector_load, stmt_info, 0, vect_body); | |
1143 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1144 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1145 | |
1146 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1147 | the containing loop, the following cost should be added to the | |
92345349 | 1148 | prologue costs. */ |
ebfd146a | 1149 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1150 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1151 | stmt_info, 0, vect_body); | |
ebfd146a | 1152 | |
73fbfcad | 1153 | if (dump_enabled_p ()) |
e645e942 TJ |
1154 | dump_printf_loc (MSG_NOTE, vect_location, |
1155 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1156 | |
ebfd146a IR |
1157 | break; |
1158 | } | |
1159 | case dr_explicit_realign_optimized: | |
1160 | { | |
73fbfcad | 1161 | if (dump_enabled_p ()) |
e645e942 | 1162 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1163 | "vect_model_load_cost: unaligned software " |
e645e942 | 1164 | "pipelined.\n"); |
ebfd146a IR |
1165 | |
1166 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1167 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1168 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1169 | access, then the above cost should only be considered for one |
ff802fa1 | 1170 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1171 | and a realignment op. */ |
1172 | ||
92345349 | 1173 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1174 | { |
92345349 BS |
1175 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1176 | vector_stmt, stmt_info, | |
1177 | 0, vect_prologue); | |
ebfd146a | 1178 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1179 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1180 | vector_stmt, stmt_info, | |
1181 | 0, vect_prologue); | |
ebfd146a IR |
1182 | } |
1183 | ||
92345349 BS |
1184 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1185 | stmt_info, 0, vect_body); | |
1186 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1187 | stmt_info, 0, vect_body); | |
8bd37302 | 1188 | |
73fbfcad | 1189 | if (dump_enabled_p ()) |
78c60e3d | 1190 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1191 | "vect_model_load_cost: explicit realign optimized" |
1192 | "\n"); | |
8bd37302 | 1193 | |
ebfd146a IR |
1194 | break; |
1195 | } | |
1196 | ||
38eec4c6 UW |
1197 | case dr_unaligned_unsupported: |
1198 | { | |
1199 | *inside_cost = VECT_MAX_COST; | |
1200 | ||
73fbfcad | 1201 | if (dump_enabled_p ()) |
78c60e3d | 1202 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1203 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1204 | break; |
1205 | } | |
1206 | ||
ebfd146a IR |
1207 | default: |
1208 | gcc_unreachable (); | |
1209 | } | |
ebfd146a IR |
1210 | } |
1211 | ||
418b7df3 RG |
1212 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1213 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1214 | |
418b7df3 RG |
1215 | static void |
1216 | vect_init_vector_1 (gimple stmt, gimple new_stmt, gimple_stmt_iterator *gsi) | |
ebfd146a | 1217 | { |
ebfd146a | 1218 | if (gsi) |
418b7df3 | 1219 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1220 | else |
1221 | { | |
418b7df3 | 1222 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1223 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1224 | |
a70d6342 IR |
1225 | if (loop_vinfo) |
1226 | { | |
1227 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1228 | basic_block new_bb; |
1229 | edge pe; | |
a70d6342 IR |
1230 | |
1231 | if (nested_in_vect_loop_p (loop, stmt)) | |
1232 | loop = loop->inner; | |
b8698a0f | 1233 | |
a70d6342 | 1234 | pe = loop_preheader_edge (loop); |
418b7df3 | 1235 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1236 | gcc_assert (!new_bb); |
1237 | } | |
1238 | else | |
1239 | { | |
1240 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1241 | basic_block bb; | |
1242 | gimple_stmt_iterator gsi_bb_start; | |
1243 | ||
1244 | gcc_assert (bb_vinfo); | |
1245 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1246 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1247 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1248 | } |
ebfd146a IR |
1249 | } |
1250 | ||
73fbfcad | 1251 | if (dump_enabled_p ()) |
ebfd146a | 1252 | { |
78c60e3d SS |
1253 | dump_printf_loc (MSG_NOTE, vect_location, |
1254 | "created new init_stmt: "); | |
1255 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
e645e942 | 1256 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 1257 | } |
418b7df3 RG |
1258 | } |
1259 | ||
1260 | /* Function vect_init_vector. | |
ebfd146a | 1261 | |
5467ee52 RG |
1262 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1263 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1264 | vector type a vector with all elements equal to VAL is created first. | |
1265 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1266 | initialization at the loop preheader. | |
418b7df3 RG |
1267 | Return the DEF of INIT_STMT. |
1268 | It will be used in the vectorization of STMT. */ | |
1269 | ||
1270 | tree | |
5467ee52 | 1271 | vect_init_vector (gimple stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 RG |
1272 | { |
1273 | tree new_var; | |
1274 | gimple init_stmt; | |
1275 | tree vec_oprnd; | |
1276 | tree new_temp; | |
1277 | ||
5467ee52 RG |
1278 | if (TREE_CODE (type) == VECTOR_TYPE |
1279 | && TREE_CODE (TREE_TYPE (val)) != VECTOR_TYPE) | |
418b7df3 | 1280 | { |
5467ee52 | 1281 | if (!types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) |
418b7df3 | 1282 | { |
5467ee52 RG |
1283 | if (CONSTANT_CLASS_P (val)) |
1284 | val = fold_unary (VIEW_CONVERT_EXPR, TREE_TYPE (type), val); | |
418b7df3 RG |
1285 | else |
1286 | { | |
83d5977e | 1287 | new_temp = make_ssa_name (TREE_TYPE (type), NULL); |
418b7df3 | 1288 | init_stmt = gimple_build_assign_with_ops (NOP_EXPR, |
83d5977e | 1289 | new_temp, val, |
418b7df3 | 1290 | NULL_TREE); |
418b7df3 | 1291 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1292 | val = new_temp; |
418b7df3 RG |
1293 | } |
1294 | } | |
5467ee52 | 1295 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1296 | } |
1297 | ||
5467ee52 | 1298 | new_var = vect_get_new_vect_var (type, vect_simple_var, "cst_"); |
5467ee52 | 1299 | init_stmt = gimple_build_assign (new_var, val); |
418b7df3 RG |
1300 | new_temp = make_ssa_name (new_var, init_stmt); |
1301 | gimple_assign_set_lhs (init_stmt, new_temp); | |
1302 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
ebfd146a IR |
1303 | vec_oprnd = gimple_assign_lhs (init_stmt); |
1304 | return vec_oprnd; | |
1305 | } | |
1306 | ||
a70d6342 | 1307 | |
ebfd146a IR |
1308 | /* Function vect_get_vec_def_for_operand. |
1309 | ||
ff802fa1 | 1310 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
1311 | used in the vectorized stmt for STMT. |
1312 | ||
1313 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1314 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1315 | ||
1316 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1317 | needs to be introduced. */ | |
1318 | ||
1319 | tree | |
1320 | vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def) | |
1321 | { | |
1322 | tree vec_oprnd; | |
1323 | gimple vec_stmt; | |
1324 | gimple def_stmt; | |
1325 | stmt_vec_info def_stmt_info = NULL; | |
1326 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
9dc3f7de | 1327 | unsigned int nunits; |
ebfd146a | 1328 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
ebfd146a | 1329 | tree def; |
ebfd146a IR |
1330 | enum vect_def_type dt; |
1331 | bool is_simple_use; | |
1332 | tree vector_type; | |
1333 | ||
73fbfcad | 1334 | if (dump_enabled_p ()) |
ebfd146a | 1335 | { |
78c60e3d SS |
1336 | dump_printf_loc (MSG_NOTE, vect_location, |
1337 | "vect_get_vec_def_for_operand: "); | |
1338 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
e645e942 | 1339 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1340 | } |
1341 | ||
24ee1384 IR |
1342 | is_simple_use = vect_is_simple_use (op, stmt, loop_vinfo, NULL, |
1343 | &def_stmt, &def, &dt); | |
ebfd146a | 1344 | gcc_assert (is_simple_use); |
73fbfcad | 1345 | if (dump_enabled_p ()) |
ebfd146a | 1346 | { |
78c60e3d | 1347 | int loc_printed = 0; |
ebfd146a IR |
1348 | if (def) |
1349 | { | |
78c60e3d SS |
1350 | dump_printf_loc (MSG_NOTE, vect_location, "def = "); |
1351 | loc_printed = 1; | |
1352 | dump_generic_expr (MSG_NOTE, TDF_SLIM, def); | |
e645e942 | 1353 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1354 | } |
1355 | if (def_stmt) | |
1356 | { | |
78c60e3d SS |
1357 | if (loc_printed) |
1358 | dump_printf (MSG_NOTE, " def_stmt = "); | |
1359 | else | |
1360 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1361 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
e645e942 | 1362 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1363 | } |
1364 | } | |
1365 | ||
1366 | switch (dt) | |
1367 | { | |
1368 | /* Case 1: operand is a constant. */ | |
1369 | case vect_constant_def: | |
1370 | { | |
7569a6cc RG |
1371 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
1372 | gcc_assert (vector_type); | |
9dc3f7de | 1373 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); |
7569a6cc | 1374 | |
b8698a0f | 1375 | if (scalar_def) |
ebfd146a IR |
1376 | *scalar_def = op; |
1377 | ||
1378 | /* Create 'vect_cst_ = {cst,cst,...,cst}' */ | |
73fbfcad | 1379 | if (dump_enabled_p ()) |
78c60e3d | 1380 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1381 | "Create vector_cst. nunits = %d\n", nunits); |
ebfd146a | 1382 | |
418b7df3 | 1383 | return vect_init_vector (stmt, op, vector_type, NULL); |
ebfd146a IR |
1384 | } |
1385 | ||
1386 | /* Case 2: operand is defined outside the loop - loop invariant. */ | |
8644a673 | 1387 | case vect_external_def: |
ebfd146a IR |
1388 | { |
1389 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); | |
1390 | gcc_assert (vector_type); | |
ebfd146a | 1391 | |
b8698a0f | 1392 | if (scalar_def) |
ebfd146a IR |
1393 | *scalar_def = def; |
1394 | ||
1395 | /* Create 'vec_inv = {inv,inv,..,inv}' */ | |
73fbfcad | 1396 | if (dump_enabled_p ()) |
e645e942 | 1397 | dump_printf_loc (MSG_NOTE, vect_location, "Create vector_inv.\n"); |
ebfd146a | 1398 | |
418b7df3 | 1399 | return vect_init_vector (stmt, def, vector_type, NULL); |
ebfd146a IR |
1400 | } |
1401 | ||
1402 | /* Case 3: operand is defined inside the loop. */ | |
8644a673 | 1403 | case vect_internal_def: |
ebfd146a | 1404 | { |
b8698a0f | 1405 | if (scalar_def) |
ebfd146a IR |
1406 | *scalar_def = NULL/* FIXME tuples: def_stmt*/; |
1407 | ||
1408 | /* Get the def from the vectorized stmt. */ | |
1409 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1410 | |
ebfd146a | 1411 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1412 | /* Get vectorized pattern statement. */ |
1413 | if (!vec_stmt | |
1414 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1415 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1416 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1417 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1418 | gcc_assert (vec_stmt); |
1419 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1420 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1421 | else if (is_gimple_call (vec_stmt)) | |
1422 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1423 | else | |
1424 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1425 | return vec_oprnd; | |
1426 | } | |
1427 | ||
1428 | /* Case 4: operand is defined by a loop header phi - reduction */ | |
1429 | case vect_reduction_def: | |
06066f92 | 1430 | case vect_double_reduction_def: |
7c5222ff | 1431 | case vect_nested_cycle: |
ebfd146a IR |
1432 | { |
1433 | struct loop *loop; | |
1434 | ||
1435 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
b8698a0f | 1436 | loop = (gimple_bb (def_stmt))->loop_father; |
ebfd146a IR |
1437 | |
1438 | /* Get the def before the loop */ | |
1439 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); | |
1440 | return get_initial_def_for_reduction (stmt, op, scalar_def); | |
1441 | } | |
1442 | ||
1443 | /* Case 5: operand is defined by loop-header phi - induction. */ | |
1444 | case vect_induction_def: | |
1445 | { | |
1446 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1447 | ||
1448 | /* Get the def from the vectorized stmt. */ | |
1449 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1450 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1451 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1452 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1453 | else | |
1454 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1455 | return vec_oprnd; |
1456 | } | |
1457 | ||
1458 | default: | |
1459 | gcc_unreachable (); | |
1460 | } | |
1461 | } | |
1462 | ||
1463 | ||
1464 | /* Function vect_get_vec_def_for_stmt_copy | |
1465 | ||
ff802fa1 | 1466 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1467 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1468 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1469 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1470 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1471 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1472 | DT is the type of the vector def VEC_OPRND. |
1473 | ||
1474 | Context: | |
1475 | In case the vectorization factor (VF) is bigger than the number | |
1476 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1477 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1478 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1479 | smallest data-type determines the VF, and as a result, when vectorizing |
1480 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1481 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1482 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1483 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1484 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1485 | ||
1486 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1487 | |
ebfd146a IR |
1488 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1489 | VS1.1: vx.1 = memref1 VS1.2 | |
1490 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1491 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1492 | |
1493 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1494 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1495 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1496 | VSnew.3: vz3 = vx.3 + ... | |
1497 | ||
1498 | The vectorization of S1 is explained in vectorizable_load. | |
1499 | The vectorization of S2: | |
b8698a0f L |
1500 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1501 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1502 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1503 | returns the vector-def 'vx.0'. |
1504 | ||
b8698a0f L |
1505 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1506 | function is called to get the relevant vector-def for each operand. It is | |
1507 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1508 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1509 | ||
b8698a0f L |
1510 | For example, to obtain the vector-def 'vx.1' in order to create the |
1511 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1512 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1513 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1514 | and return its def ('vx.1'). | |
1515 | Overall, to create the above sequence this function will be called 3 times: | |
1516 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1517 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1518 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1519 | ||
1520 | tree | |
1521 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1522 | { | |
1523 | gimple vec_stmt_for_operand; | |
1524 | stmt_vec_info def_stmt_info; | |
1525 | ||
1526 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1527 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1528 | return vec_oprnd; |
1529 | ||
1530 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1531 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1532 | gcc_assert (def_stmt_info); | |
1533 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1534 | gcc_assert (vec_stmt_for_operand); | |
1535 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1536 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) | |
1537 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1538 | else | |
1539 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1540 | return vec_oprnd; | |
1541 | } | |
1542 | ||
1543 | ||
1544 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1545 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1546 | |
1547 | static void | |
b8698a0f | 1548 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1549 | vec<tree> *vec_oprnds0, |
1550 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1551 | { |
9771b263 | 1552 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1553 | |
1554 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1555 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1556 | |
9771b263 | 1557 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1558 | { |
9771b263 | 1559 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1560 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1561 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1562 | } |
1563 | } | |
1564 | ||
1565 | ||
d092494c IR |
1566 | /* Get vectorized definitions for OP0 and OP1. |
1567 | REDUC_INDEX is the index of reduction operand in case of reduction, | |
1568 | and -1 otherwise. */ | |
ebfd146a | 1569 | |
d092494c | 1570 | void |
ebfd146a | 1571 | vect_get_vec_defs (tree op0, tree op1, gimple stmt, |
9771b263 DN |
1572 | vec<tree> *vec_oprnds0, |
1573 | vec<tree> *vec_oprnds1, | |
d092494c | 1574 | slp_tree slp_node, int reduc_index) |
ebfd146a IR |
1575 | { |
1576 | if (slp_node) | |
d092494c IR |
1577 | { |
1578 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
9771b263 DN |
1579 | vec<tree> ops; |
1580 | ops.create (nops); | |
37b5ec8f | 1581 | vec<vec<tree> > vec_defs; |
9771b263 | 1582 | vec_defs.create (nops); |
d092494c | 1583 | |
9771b263 | 1584 | ops.quick_push (op0); |
d092494c | 1585 | if (op1) |
9771b263 | 1586 | ops.quick_push (op1); |
d092494c IR |
1587 | |
1588 | vect_get_slp_defs (ops, slp_node, &vec_defs, reduc_index); | |
1589 | ||
37b5ec8f | 1590 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1591 | if (op1) |
37b5ec8f | 1592 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1593 | |
9771b263 DN |
1594 | ops.release (); |
1595 | vec_defs.release (); | |
d092494c | 1596 | } |
ebfd146a IR |
1597 | else |
1598 | { | |
1599 | tree vec_oprnd; | |
1600 | ||
9771b263 | 1601 | vec_oprnds0->create (1); |
b8698a0f | 1602 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL); |
9771b263 | 1603 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1604 | |
1605 | if (op1) | |
1606 | { | |
9771b263 | 1607 | vec_oprnds1->create (1); |
b8698a0f | 1608 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL); |
9771b263 | 1609 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1610 | } |
1611 | } | |
1612 | } | |
1613 | ||
1614 | ||
1615 | /* Function vect_finish_stmt_generation. | |
1616 | ||
1617 | Insert a new stmt. */ | |
1618 | ||
1619 | void | |
1620 | vect_finish_stmt_generation (gimple stmt, gimple vec_stmt, | |
1621 | gimple_stmt_iterator *gsi) | |
1622 | { | |
1623 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1624 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1625 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1626 | |
1627 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1628 | ||
54e8e2c3 RG |
1629 | if (!gsi_end_p (*gsi) |
1630 | && gimple_has_mem_ops (vec_stmt)) | |
1631 | { | |
1632 | gimple at_stmt = gsi_stmt (*gsi); | |
1633 | tree vuse = gimple_vuse (at_stmt); | |
1634 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1635 | { | |
1636 | tree vdef = gimple_vdef (at_stmt); | |
1637 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1638 | /* If we have an SSA vuse and insert a store, update virtual | |
1639 | SSA form to avoid triggering the renamer. Do so only | |
1640 | if we can easily see all uses - which is what almost always | |
1641 | happens with the way vectorized stmts are inserted. */ | |
1642 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1643 | && ((is_gimple_assign (vec_stmt) | |
1644 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1645 | || (is_gimple_call (vec_stmt) | |
1646 | && !(gimple_call_flags (vec_stmt) | |
1647 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1648 | { | |
1649 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1650 | gimple_set_vdef (vec_stmt, new_vdef); | |
1651 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1652 | } | |
1653 | } | |
1654 | } | |
ebfd146a IR |
1655 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1656 | ||
b8698a0f | 1657 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo, |
a70d6342 | 1658 | bb_vinfo)); |
ebfd146a | 1659 | |
73fbfcad | 1660 | if (dump_enabled_p ()) |
ebfd146a | 1661 | { |
78c60e3d SS |
1662 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1663 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
e645e942 | 1664 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1665 | } |
1666 | ||
ad885386 | 1667 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
ebfd146a IR |
1668 | } |
1669 | ||
1670 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1671 | a function declaration if the target has a vectorized version | |
1672 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1673 | ||
1674 | tree | |
1675 | vectorizable_function (gimple call, tree vectype_out, tree vectype_in) | |
1676 | { | |
1677 | tree fndecl = gimple_call_fndecl (call); | |
ebfd146a IR |
1678 | |
1679 | /* We only handle functions that do not read or clobber memory -- i.e. | |
1680 | const or novops ones. */ | |
1681 | if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS))) | |
1682 | return NULL_TREE; | |
1683 | ||
1684 | if (!fndecl | |
1685 | || TREE_CODE (fndecl) != FUNCTION_DECL | |
1686 | || !DECL_BUILT_IN (fndecl)) | |
1687 | return NULL_TREE; | |
1688 | ||
62f7fd21 | 1689 | return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out, |
ebfd146a IR |
1690 | vectype_in); |
1691 | } | |
1692 | ||
1693 | /* Function vectorizable_call. | |
1694 | ||
b8698a0f L |
1695 | Check if STMT performs a function call that can be vectorized. |
1696 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1697 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1698 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1699 | ||
1700 | static bool | |
190c2236 JJ |
1701 | vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, |
1702 | slp_tree slp_node) | |
ebfd146a IR |
1703 | { |
1704 | tree vec_dest; | |
1705 | tree scalar_dest; | |
1706 | tree op, type; | |
1707 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1708 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
1709 | tree vectype_out, vectype_in; | |
1710 | int nunits_in; | |
1711 | int nunits_out; | |
1712 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 1713 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b690cc0f | 1714 | tree fndecl, new_temp, def, rhs_type; |
ebfd146a | 1715 | gimple def_stmt; |
0502fb85 UB |
1716 | enum vect_def_type dt[3] |
1717 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
63827fb8 | 1718 | gimple new_stmt = NULL; |
ebfd146a | 1719 | int ncopies, j; |
6e1aa848 | 1720 | vec<tree> vargs = vNULL; |
ebfd146a IR |
1721 | enum { NARROW, NONE, WIDEN } modifier; |
1722 | size_t i, nargs; | |
9d5e7640 | 1723 | tree lhs; |
ebfd146a | 1724 | |
190c2236 | 1725 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
1726 | return false; |
1727 | ||
8644a673 | 1728 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1729 | return false; |
1730 | ||
ebfd146a IR |
1731 | /* Is STMT a vectorizable call? */ |
1732 | if (!is_gimple_call (stmt)) | |
1733 | return false; | |
1734 | ||
1735 | if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
1736 | return false; | |
1737 | ||
822ba6d7 | 1738 | if (stmt_can_throw_internal (stmt)) |
5a2c1986 IR |
1739 | return false; |
1740 | ||
b690cc0f RG |
1741 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
1742 | ||
ebfd146a IR |
1743 | /* Process function arguments. */ |
1744 | rhs_type = NULL_TREE; | |
b690cc0f | 1745 | vectype_in = NULL_TREE; |
ebfd146a IR |
1746 | nargs = gimple_call_num_args (stmt); |
1747 | ||
1b1562a5 MM |
1748 | /* Bail out if the function has more than three arguments, we do not have |
1749 | interesting builtin functions to vectorize with more than two arguments | |
1750 | except for fma. No arguments is also not good. */ | |
1751 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
1752 | return false; |
1753 | ||
74bf76ed JJ |
1754 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
1755 | if (gimple_call_internal_p (stmt) | |
1756 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
1757 | { | |
1758 | nargs = 0; | |
1759 | rhs_type = unsigned_type_node; | |
1760 | } | |
1761 | ||
ebfd146a IR |
1762 | for (i = 0; i < nargs; i++) |
1763 | { | |
b690cc0f RG |
1764 | tree opvectype; |
1765 | ||
ebfd146a IR |
1766 | op = gimple_call_arg (stmt, i); |
1767 | ||
1768 | /* We can only handle calls with arguments of the same type. */ | |
1769 | if (rhs_type | |
8533c9d8 | 1770 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 1771 | { |
73fbfcad | 1772 | if (dump_enabled_p ()) |
78c60e3d | 1773 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1774 | "argument types differ.\n"); |
ebfd146a IR |
1775 | return false; |
1776 | } | |
b690cc0f RG |
1777 | if (!rhs_type) |
1778 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 1779 | |
24ee1384 | 1780 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, |
b690cc0f | 1781 | &def_stmt, &def, &dt[i], &opvectype)) |
ebfd146a | 1782 | { |
73fbfcad | 1783 | if (dump_enabled_p ()) |
78c60e3d | 1784 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1785 | "use not simple.\n"); |
ebfd146a IR |
1786 | return false; |
1787 | } | |
ebfd146a | 1788 | |
b690cc0f RG |
1789 | if (!vectype_in) |
1790 | vectype_in = opvectype; | |
1791 | else if (opvectype | |
1792 | && opvectype != vectype_in) | |
1793 | { | |
73fbfcad | 1794 | if (dump_enabled_p ()) |
78c60e3d | 1795 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1796 | "argument vector types differ.\n"); |
b690cc0f RG |
1797 | return false; |
1798 | } | |
1799 | } | |
1800 | /* If all arguments are external or constant defs use a vector type with | |
1801 | the same size as the output vector type. */ | |
ebfd146a | 1802 | if (!vectype_in) |
b690cc0f | 1803 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
1804 | if (vec_stmt) |
1805 | gcc_assert (vectype_in); | |
1806 | if (!vectype_in) | |
1807 | { | |
73fbfcad | 1808 | if (dump_enabled_p ()) |
7d8930a0 | 1809 | { |
78c60e3d SS |
1810 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1811 | "no vectype for scalar type "); | |
1812 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 1813 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
1814 | } |
1815 | ||
1816 | return false; | |
1817 | } | |
ebfd146a IR |
1818 | |
1819 | /* FORNOW */ | |
b690cc0f RG |
1820 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
1821 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
1822 | if (nunits_in == nunits_out / 2) |
1823 | modifier = NARROW; | |
1824 | else if (nunits_out == nunits_in) | |
1825 | modifier = NONE; | |
1826 | else if (nunits_out == nunits_in / 2) | |
1827 | modifier = WIDEN; | |
1828 | else | |
1829 | return false; | |
1830 | ||
1831 | /* For now, we only vectorize functions if a target specific builtin | |
1832 | is available. TODO -- in some cases, it might be profitable to | |
1833 | insert the calls for pieces of the vector, in order to be able | |
1834 | to vectorize other operations in the loop. */ | |
1835 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
1836 | if (fndecl == NULL_TREE) | |
1837 | { | |
74bf76ed JJ |
1838 | if (gimple_call_internal_p (stmt) |
1839 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE | |
1840 | && !slp_node | |
1841 | && loop_vinfo | |
1842 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
1843 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
1844 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
1845 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
1846 | { | |
1847 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
1848 | { 0, 1, 2, ... vf - 1 } vector. */ | |
1849 | gcc_assert (nargs == 0); | |
1850 | } | |
1851 | else | |
1852 | { | |
1853 | if (dump_enabled_p ()) | |
1854 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 1855 | "function is not vectorizable.\n"); |
74bf76ed JJ |
1856 | return false; |
1857 | } | |
ebfd146a IR |
1858 | } |
1859 | ||
5006671f | 1860 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a | 1861 | |
190c2236 JJ |
1862 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
1863 | ncopies = 1; | |
1864 | else if (modifier == NARROW) | |
ebfd146a IR |
1865 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
1866 | else | |
1867 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1868 | ||
1869 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
1870 | needs to be generated. */ | |
1871 | gcc_assert (ncopies >= 1); | |
1872 | ||
1873 | if (!vec_stmt) /* transformation not required. */ | |
1874 | { | |
1875 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 1876 | if (dump_enabled_p ()) |
e645e942 TJ |
1877 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
1878 | "\n"); | |
c3e7ee41 | 1879 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
1880 | return true; |
1881 | } | |
1882 | ||
1883 | /** Transform. **/ | |
1884 | ||
73fbfcad | 1885 | if (dump_enabled_p ()) |
e645e942 | 1886 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
1887 | |
1888 | /* Handle def. */ | |
1889 | scalar_dest = gimple_call_lhs (stmt); | |
1890 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1891 | ||
1892 | prev_stmt_info = NULL; | |
1893 | switch (modifier) | |
1894 | { | |
1895 | case NONE: | |
1896 | for (j = 0; j < ncopies; ++j) | |
1897 | { | |
1898 | /* Build argument list for the vectorized call. */ | |
1899 | if (j == 0) | |
9771b263 | 1900 | vargs.create (nargs); |
ebfd146a | 1901 | else |
9771b263 | 1902 | vargs.truncate (0); |
ebfd146a | 1903 | |
190c2236 JJ |
1904 | if (slp_node) |
1905 | { | |
37b5ec8f | 1906 | vec<vec<tree> > vec_defs; |
9771b263 DN |
1907 | vec_defs.create (nargs); |
1908 | vec<tree> vec_oprnds0; | |
190c2236 JJ |
1909 | |
1910 | for (i = 0; i < nargs; i++) | |
9771b263 | 1911 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 1912 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 1913 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
1914 | |
1915 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 1916 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
1917 | { |
1918 | size_t k; | |
1919 | for (k = 0; k < nargs; k++) | |
1920 | { | |
37b5ec8f | 1921 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 1922 | vargs[k] = vec_oprndsk[i]; |
190c2236 JJ |
1923 | } |
1924 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1925 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1926 | gimple_call_set_lhs (new_stmt, new_temp); | |
1927 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 1928 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
1929 | } |
1930 | ||
1931 | for (i = 0; i < nargs; i++) | |
1932 | { | |
37b5ec8f | 1933 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 1934 | vec_oprndsi.release (); |
190c2236 | 1935 | } |
9771b263 | 1936 | vec_defs.release (); |
190c2236 JJ |
1937 | continue; |
1938 | } | |
1939 | ||
ebfd146a IR |
1940 | for (i = 0; i < nargs; i++) |
1941 | { | |
1942 | op = gimple_call_arg (stmt, i); | |
1943 | if (j == 0) | |
1944 | vec_oprnd0 | |
1945 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1946 | else | |
63827fb8 IR |
1947 | { |
1948 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
1949 | vec_oprnd0 | |
1950 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
1951 | } | |
ebfd146a | 1952 | |
9771b263 | 1953 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
1954 | } |
1955 | ||
74bf76ed JJ |
1956 | if (gimple_call_internal_p (stmt) |
1957 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
1958 | { | |
1959 | tree *v = XALLOCAVEC (tree, nunits_out); | |
1960 | int k; | |
1961 | for (k = 0; k < nunits_out; ++k) | |
1962 | v[k] = build_int_cst (unsigned_type_node, j * nunits_out + k); | |
1963 | tree cst = build_vector (vectype_out, v); | |
1964 | tree new_var | |
1965 | = vect_get_new_vect_var (vectype_out, vect_simple_var, "cst_"); | |
1966 | gimple init_stmt = gimple_build_assign (new_var, cst); | |
1967 | new_temp = make_ssa_name (new_var, init_stmt); | |
1968 | gimple_assign_set_lhs (init_stmt, new_temp); | |
1969 | vect_init_vector_1 (stmt, init_stmt, NULL); | |
1970 | new_temp = make_ssa_name (vec_dest, NULL); | |
1971 | new_stmt = gimple_build_assign (new_temp, | |
1972 | gimple_assign_lhs (init_stmt)); | |
1973 | } | |
1974 | else | |
1975 | { | |
1976 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1977 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1978 | gimple_call_set_lhs (new_stmt, new_temp); | |
1979 | } | |
ebfd146a IR |
1980 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1981 | ||
1982 | if (j == 0) | |
1983 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1984 | else | |
1985 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1986 | ||
1987 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1988 | } | |
1989 | ||
1990 | break; | |
1991 | ||
1992 | case NARROW: | |
1993 | for (j = 0; j < ncopies; ++j) | |
1994 | { | |
1995 | /* Build argument list for the vectorized call. */ | |
1996 | if (j == 0) | |
9771b263 | 1997 | vargs.create (nargs * 2); |
ebfd146a | 1998 | else |
9771b263 | 1999 | vargs.truncate (0); |
ebfd146a | 2000 | |
190c2236 JJ |
2001 | if (slp_node) |
2002 | { | |
37b5ec8f | 2003 | vec<vec<tree> > vec_defs; |
9771b263 DN |
2004 | vec_defs.create (nargs); |
2005 | vec<tree> vec_oprnds0; | |
190c2236 JJ |
2006 | |
2007 | for (i = 0; i < nargs; i++) | |
9771b263 | 2008 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2009 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2010 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2011 | |
2012 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2013 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
2014 | { |
2015 | size_t k; | |
9771b263 | 2016 | vargs.truncate (0); |
190c2236 JJ |
2017 | for (k = 0; k < nargs; k++) |
2018 | { | |
37b5ec8f | 2019 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
2020 | vargs.quick_push (vec_oprndsk[i]); |
2021 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 JJ |
2022 | } |
2023 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2024 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2025 | gimple_call_set_lhs (new_stmt, new_temp); | |
2026 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2027 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2028 | } |
2029 | ||
2030 | for (i = 0; i < nargs; i++) | |
2031 | { | |
37b5ec8f | 2032 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2033 | vec_oprndsi.release (); |
190c2236 | 2034 | } |
9771b263 | 2035 | vec_defs.release (); |
190c2236 JJ |
2036 | continue; |
2037 | } | |
2038 | ||
ebfd146a IR |
2039 | for (i = 0; i < nargs; i++) |
2040 | { | |
2041 | op = gimple_call_arg (stmt, i); | |
2042 | if (j == 0) | |
2043 | { | |
2044 | vec_oprnd0 | |
2045 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
2046 | vec_oprnd1 | |
63827fb8 | 2047 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2048 | } |
2049 | else | |
2050 | { | |
336ecb65 | 2051 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 2052 | vec_oprnd0 |
63827fb8 | 2053 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 2054 | vec_oprnd1 |
63827fb8 | 2055 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2056 | } |
2057 | ||
9771b263 DN |
2058 | vargs.quick_push (vec_oprnd0); |
2059 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
2060 | } |
2061 | ||
2062 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2063 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2064 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
2065 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2066 | ||
2067 | if (j == 0) | |
2068 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2069 | else | |
2070 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2071 | ||
2072 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2073 | } | |
2074 | ||
2075 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2076 | ||
2077 | break; | |
2078 | ||
2079 | case WIDEN: | |
2080 | /* No current target implements this case. */ | |
2081 | return false; | |
2082 | } | |
2083 | ||
9771b263 | 2084 | vargs.release (); |
ebfd146a IR |
2085 | |
2086 | /* Update the exception handling table with the vector stmt if necessary. */ | |
2087 | if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt)) | |
2088 | gimple_purge_dead_eh_edges (gimple_bb (stmt)); | |
2089 | ||
2090 | /* The call in STMT might prevent it from being removed in dce. | |
2091 | We however cannot remove it here, due to the way the ssa name | |
2092 | it defines is mapped to the new definition. So just replace | |
2093 | rhs of the statement with something harmless. */ | |
2094 | ||
dd34c087 JJ |
2095 | if (slp_node) |
2096 | return true; | |
2097 | ||
ebfd146a | 2098 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
2099 | if (is_pattern_stmt_p (stmt_info)) |
2100 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
2101 | else | |
2102 | lhs = gimple_call_lhs (stmt); | |
2103 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
ebfd146a | 2104 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 2105 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
2106 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
2107 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
2108 | |
2109 | return true; | |
2110 | } | |
2111 | ||
2112 | ||
2113 | /* Function vect_gen_widened_results_half | |
2114 | ||
2115 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 2116 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 2117 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
2118 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
2119 | needs to be created (DECL is a function-decl of a target-builtin). | |
2120 | STMT is the original scalar stmt that we are vectorizing. */ | |
2121 | ||
2122 | static gimple | |
2123 | vect_gen_widened_results_half (enum tree_code code, | |
2124 | tree decl, | |
2125 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
2126 | tree vec_dest, gimple_stmt_iterator *gsi, | |
2127 | gimple stmt) | |
b8698a0f | 2128 | { |
ebfd146a | 2129 | gimple new_stmt; |
b8698a0f L |
2130 | tree new_temp; |
2131 | ||
2132 | /* Generate half of the widened result: */ | |
2133 | if (code == CALL_EXPR) | |
2134 | { | |
2135 | /* Target specific support */ | |
ebfd146a IR |
2136 | if (op_type == binary_op) |
2137 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
2138 | else | |
2139 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
2140 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2141 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
2142 | } |
2143 | else | |
ebfd146a | 2144 | { |
b8698a0f L |
2145 | /* Generic support */ |
2146 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
2147 | if (op_type != binary_op) |
2148 | vec_oprnd1 = NULL; | |
2149 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0, | |
2150 | vec_oprnd1); | |
2151 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2152 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 2153 | } |
ebfd146a IR |
2154 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2155 | ||
ebfd146a IR |
2156 | return new_stmt; |
2157 | } | |
2158 | ||
4a00c761 JJ |
2159 | |
2160 | /* Get vectorized definitions for loop-based vectorization. For the first | |
2161 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
2162 | scalar operand), and for the rest we get a copy with | |
2163 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
2164 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
2165 | The vectors are collected into VEC_OPRNDS. */ | |
2166 | ||
2167 | static void | |
2168 | vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt, | |
9771b263 | 2169 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
2170 | { |
2171 | tree vec_oprnd; | |
2172 | ||
2173 | /* Get first vector operand. */ | |
2174 | /* All the vector operands except the very first one (that is scalar oprnd) | |
2175 | are stmt copies. */ | |
2176 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
2177 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL); | |
2178 | else | |
2179 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
2180 | ||
9771b263 | 2181 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
2182 | |
2183 | /* Get second vector operand. */ | |
2184 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 2185 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
2186 | |
2187 | *oprnd = vec_oprnd; | |
2188 | ||
2189 | /* For conversion in multiple steps, continue to get operands | |
2190 | recursively. */ | |
2191 | if (multi_step_cvt) | |
2192 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
2193 | } | |
2194 | ||
2195 | ||
2196 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
2197 | For multi-step conversions store the resulting vectors and call the function | |
2198 | recursively. */ | |
2199 | ||
2200 | static void | |
9771b263 | 2201 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
4a00c761 | 2202 | int multi_step_cvt, gimple stmt, |
9771b263 | 2203 | vec<tree> vec_dsts, |
4a00c761 JJ |
2204 | gimple_stmt_iterator *gsi, |
2205 | slp_tree slp_node, enum tree_code code, | |
2206 | stmt_vec_info *prev_stmt_info) | |
2207 | { | |
2208 | unsigned int i; | |
2209 | tree vop0, vop1, new_tmp, vec_dest; | |
2210 | gimple new_stmt; | |
2211 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2212 | ||
9771b263 | 2213 | vec_dest = vec_dsts.pop (); |
4a00c761 | 2214 | |
9771b263 | 2215 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
2216 | { |
2217 | /* Create demotion operation. */ | |
9771b263 DN |
2218 | vop0 = (*vec_oprnds)[i]; |
2219 | vop1 = (*vec_oprnds)[i + 1]; | |
4a00c761 JJ |
2220 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); |
2221 | new_tmp = make_ssa_name (vec_dest, new_stmt); | |
2222 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
2223 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2224 | ||
2225 | if (multi_step_cvt) | |
2226 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 2227 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
2228 | else |
2229 | { | |
2230 | /* This is the last step of the conversion sequence. Store the | |
2231 | vectors in SLP_NODE or in vector info of the scalar statement | |
2232 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
2233 | if (slp_node) | |
9771b263 | 2234 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2235 | else |
2236 | { | |
2237 | if (!*prev_stmt_info) | |
2238 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2239 | else | |
2240 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
2241 | ||
2242 | *prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2243 | } | |
2244 | } | |
2245 | } | |
2246 | ||
2247 | /* For multi-step demotion operations we first generate demotion operations | |
2248 | from the source type to the intermediate types, and then combine the | |
2249 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
2250 | type. */ | |
2251 | if (multi_step_cvt) | |
2252 | { | |
2253 | /* At each level of recursion we have half of the operands we had at the | |
2254 | previous level. */ | |
9771b263 | 2255 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
2256 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
2257 | stmt, vec_dsts, gsi, slp_node, | |
2258 | VEC_PACK_TRUNC_EXPR, | |
2259 | prev_stmt_info); | |
2260 | } | |
2261 | ||
9771b263 | 2262 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2263 | } |
2264 | ||
2265 | ||
2266 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
2267 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
2268 | the resulting vectors and call the function recursively. */ | |
2269 | ||
2270 | static void | |
9771b263 DN |
2271 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
2272 | vec<tree> *vec_oprnds1, | |
4a00c761 JJ |
2273 | gimple stmt, tree vec_dest, |
2274 | gimple_stmt_iterator *gsi, | |
2275 | enum tree_code code1, | |
2276 | enum tree_code code2, tree decl1, | |
2277 | tree decl2, int op_type) | |
2278 | { | |
2279 | int i; | |
2280 | tree vop0, vop1, new_tmp1, new_tmp2; | |
2281 | gimple new_stmt1, new_stmt2; | |
6e1aa848 | 2282 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 2283 | |
9771b263 DN |
2284 | vec_tmp.create (vec_oprnds0->length () * 2); |
2285 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
2286 | { |
2287 | if (op_type == binary_op) | |
9771b263 | 2288 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
2289 | else |
2290 | vop1 = NULL_TREE; | |
2291 | ||
2292 | /* Generate the two halves of promotion operation. */ | |
2293 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
2294 | op_type, vec_dest, gsi, stmt); | |
2295 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
2296 | op_type, vec_dest, gsi, stmt); | |
2297 | if (is_gimple_call (new_stmt1)) | |
2298 | { | |
2299 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
2300 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
2301 | } | |
2302 | else | |
2303 | { | |
2304 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
2305 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
2306 | } | |
2307 | ||
2308 | /* Store the results for the next step. */ | |
9771b263 DN |
2309 | vec_tmp.quick_push (new_tmp1); |
2310 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
2311 | } |
2312 | ||
689eaba3 | 2313 | vec_oprnds0->release (); |
4a00c761 JJ |
2314 | *vec_oprnds0 = vec_tmp; |
2315 | } | |
2316 | ||
2317 | ||
b8698a0f L |
2318 | /* Check if STMT performs a conversion operation, that can be vectorized. |
2319 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 2320 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
2321 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
2322 | ||
2323 | static bool | |
2324 | vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi, | |
2325 | gimple *vec_stmt, slp_tree slp_node) | |
2326 | { | |
2327 | tree vec_dest; | |
2328 | tree scalar_dest; | |
4a00c761 | 2329 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
2330 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
2331 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2332 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2333 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 2334 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
2335 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
2336 | tree new_temp; | |
2337 | tree def; | |
2338 | gimple def_stmt; | |
2339 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2340 | gimple new_stmt = NULL; | |
2341 | stmt_vec_info prev_stmt_info; | |
2342 | int nunits_in; | |
2343 | int nunits_out; | |
2344 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
2345 | int ncopies, i, j; |
2346 | tree lhs_type, rhs_type; | |
ebfd146a | 2347 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
2348 | vec<tree> vec_oprnds0 = vNULL; |
2349 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 2350 | tree vop0; |
4a00c761 JJ |
2351 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2352 | int multi_step_cvt = 0; | |
6e1aa848 DN |
2353 | vec<tree> vec_dsts = vNULL; |
2354 | vec<tree> interm_types = vNULL; | |
4a00c761 JJ |
2355 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
2356 | int op_type; | |
2357 | enum machine_mode rhs_mode; | |
2358 | unsigned short fltsz; | |
ebfd146a IR |
2359 | |
2360 | /* Is STMT a vectorizable conversion? */ | |
2361 | ||
4a00c761 | 2362 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2363 | return false; |
2364 | ||
8644a673 | 2365 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2366 | return false; |
2367 | ||
2368 | if (!is_gimple_assign (stmt)) | |
2369 | return false; | |
2370 | ||
2371 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2372 | return false; | |
2373 | ||
2374 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
2375 | if (!CONVERT_EXPR_CODE_P (code) |
2376 | && code != FIX_TRUNC_EXPR | |
2377 | && code != FLOAT_EXPR | |
2378 | && code != WIDEN_MULT_EXPR | |
2379 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
2380 | return false; |
2381 | ||
4a00c761 JJ |
2382 | op_type = TREE_CODE_LENGTH (code); |
2383 | ||
ebfd146a | 2384 | /* Check types of lhs and rhs. */ |
b690cc0f | 2385 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 2386 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
2387 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2388 | ||
ebfd146a IR |
2389 | op0 = gimple_assign_rhs1 (stmt); |
2390 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
2391 | |
2392 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
2393 | && !((INTEGRAL_TYPE_P (lhs_type) | |
2394 | && INTEGRAL_TYPE_P (rhs_type)) | |
2395 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
2396 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
2397 | return false; | |
2398 | ||
2399 | if ((INTEGRAL_TYPE_P (lhs_type) | |
2400 | && (TYPE_PRECISION (lhs_type) | |
2401 | != GET_MODE_PRECISION (TYPE_MODE (lhs_type)))) | |
2402 | || (INTEGRAL_TYPE_P (rhs_type) | |
2403 | && (TYPE_PRECISION (rhs_type) | |
2404 | != GET_MODE_PRECISION (TYPE_MODE (rhs_type))))) | |
2405 | { | |
73fbfcad | 2406 | if (dump_enabled_p ()) |
78c60e3d | 2407 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
2408 | "type conversion to/from bit-precision unsupported." |
2409 | "\n"); | |
4a00c761 JJ |
2410 | return false; |
2411 | } | |
2412 | ||
b690cc0f | 2413 | /* Check the operands of the operation. */ |
24ee1384 | 2414 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
b690cc0f RG |
2415 | &def_stmt, &def, &dt[0], &vectype_in)) |
2416 | { | |
73fbfcad | 2417 | if (dump_enabled_p ()) |
78c60e3d | 2418 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2419 | "use not simple.\n"); |
b690cc0f RG |
2420 | return false; |
2421 | } | |
4a00c761 JJ |
2422 | if (op_type == binary_op) |
2423 | { | |
2424 | bool ok; | |
2425 | ||
2426 | op1 = gimple_assign_rhs2 (stmt); | |
2427 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
2428 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
2429 | OP1. */ | |
2430 | if (CONSTANT_CLASS_P (op0)) | |
f5709183 | 2431 | ok = vect_is_simple_use_1 (op1, stmt, loop_vinfo, bb_vinfo, |
4a00c761 JJ |
2432 | &def_stmt, &def, &dt[1], &vectype_in); |
2433 | else | |
f5709183 | 2434 | ok = vect_is_simple_use (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
24ee1384 | 2435 | &def, &dt[1]); |
4a00c761 JJ |
2436 | |
2437 | if (!ok) | |
2438 | { | |
73fbfcad | 2439 | if (dump_enabled_p ()) |
78c60e3d | 2440 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2441 | "use not simple.\n"); |
4a00c761 JJ |
2442 | return false; |
2443 | } | |
2444 | } | |
2445 | ||
b690cc0f RG |
2446 | /* If op0 is an external or constant defs use a vector type of |
2447 | the same size as the output vector type. */ | |
ebfd146a | 2448 | if (!vectype_in) |
b690cc0f | 2449 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2450 | if (vec_stmt) |
2451 | gcc_assert (vectype_in); | |
2452 | if (!vectype_in) | |
2453 | { | |
73fbfcad | 2454 | if (dump_enabled_p ()) |
4a00c761 | 2455 | { |
78c60e3d SS |
2456 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2457 | "no vectype for scalar type "); | |
2458 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2459 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 2460 | } |
7d8930a0 IR |
2461 | |
2462 | return false; | |
2463 | } | |
ebfd146a | 2464 | |
b690cc0f RG |
2465 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2466 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 2467 | if (nunits_in < nunits_out) |
ebfd146a IR |
2468 | modifier = NARROW; |
2469 | else if (nunits_out == nunits_in) | |
2470 | modifier = NONE; | |
ebfd146a | 2471 | else |
4a00c761 | 2472 | modifier = WIDEN; |
ebfd146a | 2473 | |
ff802fa1 IR |
2474 | /* Multiple types in SLP are handled by creating the appropriate number of |
2475 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2476 | case of SLP. */ | |
437f4a00 | 2477 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a | 2478 | ncopies = 1; |
4a00c761 JJ |
2479 | else if (modifier == NARROW) |
2480 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
2481 | else | |
2482 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 2483 | |
ebfd146a IR |
2484 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
2485 | needs to be generated. */ | |
2486 | gcc_assert (ncopies >= 1); | |
2487 | ||
ebfd146a | 2488 | /* Supportable by target? */ |
4a00c761 | 2489 | switch (modifier) |
ebfd146a | 2490 | { |
4a00c761 JJ |
2491 | case NONE: |
2492 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
2493 | return false; | |
2494 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
2495 | &decl1, &code1)) | |
2496 | break; | |
2497 | /* FALLTHRU */ | |
2498 | unsupported: | |
73fbfcad | 2499 | if (dump_enabled_p ()) |
78c60e3d | 2500 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2501 | "conversion not supported by target.\n"); |
ebfd146a | 2502 | return false; |
ebfd146a | 2503 | |
4a00c761 JJ |
2504 | case WIDEN: |
2505 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
2506 | &code1, &code2, &multi_step_cvt, |
2507 | &interm_types)) | |
4a00c761 JJ |
2508 | { |
2509 | /* Binary widening operation can only be supported directly by the | |
2510 | architecture. */ | |
2511 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
2512 | break; | |
2513 | } | |
2514 | ||
2515 | if (code != FLOAT_EXPR | |
2516 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
2517 | <= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
2518 | goto unsupported; | |
2519 | ||
2520 | rhs_mode = TYPE_MODE (rhs_type); | |
2521 | fltsz = GET_MODE_SIZE (TYPE_MODE (lhs_type)); | |
2522 | for (rhs_mode = GET_MODE_2XWIDER_MODE (TYPE_MODE (rhs_type)); | |
2523 | rhs_mode != VOIDmode && GET_MODE_SIZE (rhs_mode) <= fltsz; | |
2524 | rhs_mode = GET_MODE_2XWIDER_MODE (rhs_mode)) | |
2525 | { | |
2526 | cvt_type | |
2527 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
2528 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
2529 | if (cvt_type == NULL_TREE) | |
2530 | goto unsupported; | |
2531 | ||
2532 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
2533 | { | |
2534 | if (!supportable_convert_operation (code, vectype_out, | |
2535 | cvt_type, &decl1, &codecvt1)) | |
2536 | goto unsupported; | |
2537 | } | |
2538 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
2539 | cvt_type, &codecvt1, |
2540 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
2541 | &interm_types)) |
2542 | continue; | |
2543 | else | |
2544 | gcc_assert (multi_step_cvt == 0); | |
2545 | ||
2546 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
2547 | vectype_in, &code1, &code2, |
2548 | &multi_step_cvt, &interm_types)) | |
4a00c761 JJ |
2549 | break; |
2550 | } | |
2551 | ||
2552 | if (rhs_mode == VOIDmode || GET_MODE_SIZE (rhs_mode) > fltsz) | |
2553 | goto unsupported; | |
2554 | ||
2555 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
2556 | codecvt2 = ERROR_MARK; | |
2557 | else | |
2558 | { | |
2559 | multi_step_cvt++; | |
9771b263 | 2560 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
2561 | cvt_type = NULL_TREE; |
2562 | } | |
2563 | break; | |
2564 | ||
2565 | case NARROW: | |
2566 | gcc_assert (op_type == unary_op); | |
2567 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
2568 | &code1, &multi_step_cvt, | |
2569 | &interm_types)) | |
2570 | break; | |
2571 | ||
2572 | if (code != FIX_TRUNC_EXPR | |
2573 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
2574 | >= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
2575 | goto unsupported; | |
2576 | ||
2577 | rhs_mode = TYPE_MODE (rhs_type); | |
2578 | cvt_type | |
2579 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
2580 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
2581 | if (cvt_type == NULL_TREE) | |
2582 | goto unsupported; | |
2583 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
2584 | &decl1, &codecvt1)) | |
2585 | goto unsupported; | |
2586 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
2587 | &code1, &multi_step_cvt, | |
2588 | &interm_types)) | |
2589 | break; | |
2590 | goto unsupported; | |
2591 | ||
2592 | default: | |
2593 | gcc_unreachable (); | |
ebfd146a IR |
2594 | } |
2595 | ||
2596 | if (!vec_stmt) /* transformation not required. */ | |
2597 | { | |
73fbfcad | 2598 | if (dump_enabled_p ()) |
78c60e3d | 2599 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 2600 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 2601 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
2602 | { |
2603 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
c3e7ee41 | 2604 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
8bd37302 | 2605 | } |
4a00c761 JJ |
2606 | else if (modifier == NARROW) |
2607 | { | |
2608 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 2609 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
2610 | } |
2611 | else | |
2612 | { | |
2613 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 2614 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 2615 | } |
9771b263 | 2616 | interm_types.release (); |
ebfd146a IR |
2617 | return true; |
2618 | } | |
2619 | ||
2620 | /** Transform. **/ | |
73fbfcad | 2621 | if (dump_enabled_p ()) |
78c60e3d | 2622 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 2623 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 2624 | |
4a00c761 JJ |
2625 | if (op_type == binary_op) |
2626 | { | |
2627 | if (CONSTANT_CLASS_P (op0)) | |
2628 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
2629 | else if (CONSTANT_CLASS_P (op1)) | |
2630 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
2631 | } | |
2632 | ||
2633 | /* In case of multi-step conversion, we first generate conversion operations | |
2634 | to the intermediate types, and then from that types to the final one. | |
2635 | We create vector destinations for the intermediate type (TYPES) received | |
2636 | from supportable_*_operation, and store them in the correct order | |
2637 | for future use in vect_create_vectorized_*_stmts (). */ | |
9771b263 | 2638 | vec_dsts.create (multi_step_cvt + 1); |
82294ec1 JJ |
2639 | vec_dest = vect_create_destination_var (scalar_dest, |
2640 | (cvt_type && modifier == WIDEN) | |
2641 | ? cvt_type : vectype_out); | |
9771b263 | 2642 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2643 | |
2644 | if (multi_step_cvt) | |
2645 | { | |
9771b263 DN |
2646 | for (i = interm_types.length () - 1; |
2647 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
2648 | { |
2649 | vec_dest = vect_create_destination_var (scalar_dest, | |
2650 | intermediate_type); | |
9771b263 | 2651 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2652 | } |
2653 | } | |
ebfd146a | 2654 | |
4a00c761 | 2655 | if (cvt_type) |
82294ec1 JJ |
2656 | vec_dest = vect_create_destination_var (scalar_dest, |
2657 | modifier == WIDEN | |
2658 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
2659 | |
2660 | if (!slp_node) | |
2661 | { | |
30862efc | 2662 | if (modifier == WIDEN) |
4a00c761 | 2663 | { |
c3284718 | 2664 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 2665 | if (op_type == binary_op) |
9771b263 | 2666 | vec_oprnds1.create (1); |
4a00c761 | 2667 | } |
30862efc | 2668 | else if (modifier == NARROW) |
9771b263 DN |
2669 | vec_oprnds0.create ( |
2670 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
2671 | } |
2672 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 2673 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 2674 | |
4a00c761 | 2675 | last_oprnd = op0; |
ebfd146a IR |
2676 | prev_stmt_info = NULL; |
2677 | switch (modifier) | |
2678 | { | |
2679 | case NONE: | |
2680 | for (j = 0; j < ncopies; j++) | |
2681 | { | |
ebfd146a | 2682 | if (j == 0) |
d092494c IR |
2683 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node, |
2684 | -1); | |
ebfd146a IR |
2685 | else |
2686 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
2687 | ||
9771b263 | 2688 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2689 | { |
2690 | /* Arguments are ready, create the new vector stmt. */ | |
2691 | if (code1 == CALL_EXPR) | |
2692 | { | |
2693 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2694 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2695 | gimple_call_set_lhs (new_stmt, new_temp); | |
2696 | } | |
2697 | else | |
2698 | { | |
2699 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
2700 | new_stmt = gimple_build_assign_with_ops (code1, vec_dest, | |
2701 | vop0, NULL); | |
2702 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2703 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2704 | } | |
2705 | ||
2706 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2707 | if (slp_node) | |
9771b263 | 2708 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2709 | } |
2710 | ||
ebfd146a IR |
2711 | if (j == 0) |
2712 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2713 | else | |
2714 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2715 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2716 | } | |
2717 | break; | |
2718 | ||
2719 | case WIDEN: | |
2720 | /* In case the vectorization factor (VF) is bigger than the number | |
2721 | of elements that we can fit in a vectype (nunits), we have to | |
2722 | generate more than one vector stmt - i.e - we need to "unroll" | |
2723 | the vector stmt by a factor VF/nunits. */ | |
2724 | for (j = 0; j < ncopies; j++) | |
2725 | { | |
4a00c761 | 2726 | /* Handle uses. */ |
ebfd146a | 2727 | if (j == 0) |
4a00c761 JJ |
2728 | { |
2729 | if (slp_node) | |
2730 | { | |
2731 | if (code == WIDEN_LSHIFT_EXPR) | |
2732 | { | |
2733 | unsigned int k; | |
ebfd146a | 2734 | |
4a00c761 JJ |
2735 | vec_oprnd1 = op1; |
2736 | /* Store vec_oprnd1 for every vector stmt to be created | |
2737 | for SLP_NODE. We check during the analysis that all | |
2738 | the shift arguments are the same. */ | |
2739 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 2740 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
2741 | |
2742 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2743 | slp_node, -1); | |
2744 | } | |
2745 | else | |
2746 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
2747 | &vec_oprnds1, slp_node, -1); | |
2748 | } | |
2749 | else | |
2750 | { | |
2751 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
9771b263 | 2752 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
2753 | if (op_type == binary_op) |
2754 | { | |
2755 | if (code == WIDEN_LSHIFT_EXPR) | |
2756 | vec_oprnd1 = op1; | |
2757 | else | |
2758 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, | |
2759 | NULL); | |
9771b263 | 2760 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
2761 | } |
2762 | } | |
2763 | } | |
ebfd146a | 2764 | else |
4a00c761 JJ |
2765 | { |
2766 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
2767 | vec_oprnds0.truncate (0); |
2768 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
2769 | if (op_type == binary_op) |
2770 | { | |
2771 | if (code == WIDEN_LSHIFT_EXPR) | |
2772 | vec_oprnd1 = op1; | |
2773 | else | |
2774 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
2775 | vec_oprnd1); | |
9771b263 DN |
2776 | vec_oprnds1.truncate (0); |
2777 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
2778 | } |
2779 | } | |
ebfd146a | 2780 | |
4a00c761 JJ |
2781 | /* Arguments are ready. Create the new vector stmts. */ |
2782 | for (i = multi_step_cvt; i >= 0; i--) | |
2783 | { | |
9771b263 | 2784 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
2785 | enum tree_code c1 = code1, c2 = code2; |
2786 | if (i == 0 && codecvt2 != ERROR_MARK) | |
2787 | { | |
2788 | c1 = codecvt1; | |
2789 | c2 = codecvt2; | |
2790 | } | |
2791 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
2792 | &vec_oprnds1, | |
2793 | stmt, this_dest, gsi, | |
2794 | c1, c2, decl1, decl2, | |
2795 | op_type); | |
2796 | } | |
2797 | ||
9771b263 | 2798 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2799 | { |
2800 | if (cvt_type) | |
2801 | { | |
2802 | if (codecvt1 == CALL_EXPR) | |
2803 | { | |
2804 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2805 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2806 | gimple_call_set_lhs (new_stmt, new_temp); | |
2807 | } | |
2808 | else | |
2809 | { | |
2810 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
2811 | new_temp = make_ssa_name (vec_dest, NULL); | |
2812 | new_stmt = gimple_build_assign_with_ops (codecvt1, | |
2813 | new_temp, | |
2814 | vop0, NULL); | |
2815 | } | |
2816 | ||
2817 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2818 | } | |
2819 | else | |
2820 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
2821 | ||
2822 | if (slp_node) | |
9771b263 | 2823 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2824 | else |
2825 | { | |
2826 | if (!prev_stmt_info) | |
2827 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2828 | else | |
2829 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2830 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2831 | } | |
2832 | } | |
ebfd146a | 2833 | } |
4a00c761 JJ |
2834 | |
2835 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
2836 | break; |
2837 | ||
2838 | case NARROW: | |
2839 | /* In case the vectorization factor (VF) is bigger than the number | |
2840 | of elements that we can fit in a vectype (nunits), we have to | |
2841 | generate more than one vector stmt - i.e - we need to "unroll" | |
2842 | the vector stmt by a factor VF/nunits. */ | |
2843 | for (j = 0; j < ncopies; j++) | |
2844 | { | |
2845 | /* Handle uses. */ | |
4a00c761 JJ |
2846 | if (slp_node) |
2847 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2848 | slp_node, -1); | |
ebfd146a IR |
2849 | else |
2850 | { | |
9771b263 | 2851 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
2852 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
2853 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
2854 | } |
2855 | ||
4a00c761 JJ |
2856 | /* Arguments are ready. Create the new vector stmts. */ |
2857 | if (cvt_type) | |
9771b263 | 2858 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2859 | { |
2860 | if (codecvt1 == CALL_EXPR) | |
2861 | { | |
2862 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2863 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2864 | gimple_call_set_lhs (new_stmt, new_temp); | |
2865 | } | |
2866 | else | |
2867 | { | |
2868 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
2869 | new_temp = make_ssa_name (vec_dest, NULL); | |
2870 | new_stmt = gimple_build_assign_with_ops (codecvt1, new_temp, | |
2871 | vop0, NULL); | |
2872 | } | |
ebfd146a | 2873 | |
4a00c761 | 2874 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 2875 | vec_oprnds0[i] = new_temp; |
4a00c761 | 2876 | } |
ebfd146a | 2877 | |
4a00c761 JJ |
2878 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
2879 | stmt, vec_dsts, gsi, | |
2880 | slp_node, code1, | |
2881 | &prev_stmt_info); | |
ebfd146a IR |
2882 | } |
2883 | ||
2884 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 2885 | break; |
ebfd146a IR |
2886 | } |
2887 | ||
9771b263 DN |
2888 | vec_oprnds0.release (); |
2889 | vec_oprnds1.release (); | |
2890 | vec_dsts.release (); | |
2891 | interm_types.release (); | |
ebfd146a IR |
2892 | |
2893 | return true; | |
2894 | } | |
ff802fa1 IR |
2895 | |
2896 | ||
ebfd146a IR |
2897 | /* Function vectorizable_assignment. |
2898 | ||
b8698a0f L |
2899 | Check if STMT performs an assignment (copy) that can be vectorized. |
2900 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
2901 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2902 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2903 | ||
2904 | static bool | |
2905 | vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi, | |
2906 | gimple *vec_stmt, slp_tree slp_node) | |
2907 | { | |
2908 | tree vec_dest; | |
2909 | tree scalar_dest; | |
2910 | tree op; | |
2911 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2912 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2913 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2914 | tree new_temp; | |
2915 | tree def; | |
2916 | gimple def_stmt; | |
2917 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
fde9c428 | 2918 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
ebfd146a | 2919 | int ncopies; |
f18b55bd | 2920 | int i, j; |
6e1aa848 | 2921 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 2922 | tree vop; |
a70d6342 | 2923 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
f18b55bd IR |
2924 | gimple new_stmt = NULL; |
2925 | stmt_vec_info prev_stmt_info = NULL; | |
fde9c428 RG |
2926 | enum tree_code code; |
2927 | tree vectype_in; | |
ebfd146a IR |
2928 | |
2929 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2930 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2931 | case of SLP. */ | |
437f4a00 | 2932 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
2933 | ncopies = 1; |
2934 | else | |
2935 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2936 | ||
2937 | gcc_assert (ncopies >= 1); | |
ebfd146a | 2938 | |
a70d6342 | 2939 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2940 | return false; |
2941 | ||
8644a673 | 2942 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2943 | return false; |
2944 | ||
2945 | /* Is vectorizable assignment? */ | |
2946 | if (!is_gimple_assign (stmt)) | |
2947 | return false; | |
2948 | ||
2949 | scalar_dest = gimple_assign_lhs (stmt); | |
2950 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
2951 | return false; | |
2952 | ||
fde9c428 | 2953 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 2954 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
2955 | || code == PAREN_EXPR |
2956 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
2957 | op = gimple_assign_rhs1 (stmt); |
2958 | else | |
2959 | return false; | |
2960 | ||
7b7ec6c5 RG |
2961 | if (code == VIEW_CONVERT_EXPR) |
2962 | op = TREE_OPERAND (op, 0); | |
2963 | ||
24ee1384 | 2964 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, |
fde9c428 | 2965 | &def_stmt, &def, &dt[0], &vectype_in)) |
ebfd146a | 2966 | { |
73fbfcad | 2967 | if (dump_enabled_p ()) |
78c60e3d | 2968 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2969 | "use not simple.\n"); |
ebfd146a IR |
2970 | return false; |
2971 | } | |
2972 | ||
fde9c428 RG |
2973 | /* We can handle NOP_EXPR conversions that do not change the number |
2974 | of elements or the vector size. */ | |
7b7ec6c5 RG |
2975 | if ((CONVERT_EXPR_CODE_P (code) |
2976 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
2977 | && (!vectype_in |
2978 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
2979 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
2980 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
2981 | return false; | |
2982 | ||
7b7b1813 RG |
2983 | /* We do not handle bit-precision changes. */ |
2984 | if ((CONVERT_EXPR_CODE_P (code) | |
2985 | || code == VIEW_CONVERT_EXPR) | |
2986 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2987 | && ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
2988 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
2989 | || ((TYPE_PRECISION (TREE_TYPE (op)) | |
2990 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))) | |
2991 | /* But a conversion that does not change the bit-pattern is ok. */ | |
2992 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
2993 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2994 | && TYPE_UNSIGNED (TREE_TYPE (op)))) | |
2995 | { | |
73fbfcad | 2996 | if (dump_enabled_p ()) |
78c60e3d SS |
2997 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2998 | "type conversion to/from bit-precision " | |
e645e942 | 2999 | "unsupported.\n"); |
7b7b1813 RG |
3000 | return false; |
3001 | } | |
3002 | ||
ebfd146a IR |
3003 | if (!vec_stmt) /* transformation not required. */ |
3004 | { | |
3005 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 3006 | if (dump_enabled_p ()) |
78c60e3d | 3007 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3008 | "=== vectorizable_assignment ===\n"); |
c3e7ee41 | 3009 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
3010 | return true; |
3011 | } | |
3012 | ||
3013 | /** Transform. **/ | |
73fbfcad | 3014 | if (dump_enabled_p ()) |
e645e942 | 3015 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
3016 | |
3017 | /* Handle def. */ | |
3018 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3019 | ||
3020 | /* Handle use. */ | |
f18b55bd | 3021 | for (j = 0; j < ncopies; j++) |
ebfd146a | 3022 | { |
f18b55bd IR |
3023 | /* Handle uses. */ |
3024 | if (j == 0) | |
d092494c | 3025 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node, -1); |
f18b55bd IR |
3026 | else |
3027 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
3028 | ||
3029 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 3030 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 3031 | { |
7b7ec6c5 RG |
3032 | if (CONVERT_EXPR_CODE_P (code) |
3033 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 3034 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
3035 | new_stmt = gimple_build_assign (vec_dest, vop); |
3036 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3037 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3038 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3039 | if (slp_node) | |
9771b263 | 3040 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 3041 | } |
ebfd146a IR |
3042 | |
3043 | if (slp_node) | |
f18b55bd IR |
3044 | continue; |
3045 | ||
3046 | if (j == 0) | |
3047 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3048 | else | |
3049 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3050 | ||
3051 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3052 | } | |
b8698a0f | 3053 | |
9771b263 | 3054 | vec_oprnds.release (); |
ebfd146a IR |
3055 | return true; |
3056 | } | |
3057 | ||
9dc3f7de | 3058 | |
1107f3ae IR |
3059 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
3060 | either as shift by a scalar or by a vector. */ | |
3061 | ||
3062 | bool | |
3063 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
3064 | { | |
3065 | ||
3066 | enum machine_mode vec_mode; | |
3067 | optab optab; | |
3068 | int icode; | |
3069 | tree vectype; | |
3070 | ||
3071 | vectype = get_vectype_for_scalar_type (scalar_type); | |
3072 | if (!vectype) | |
3073 | return false; | |
3074 | ||
3075 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
3076 | if (!optab | |
3077 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
3078 | { | |
3079 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
3080 | if (!optab | |
3081 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
3082 | == CODE_FOR_nothing)) | |
3083 | return false; | |
3084 | } | |
3085 | ||
3086 | vec_mode = TYPE_MODE (vectype); | |
3087 | icode = (int) optab_handler (optab, vec_mode); | |
3088 | if (icode == CODE_FOR_nothing) | |
3089 | return false; | |
3090 | ||
3091 | return true; | |
3092 | } | |
3093 | ||
3094 | ||
9dc3f7de IR |
3095 | /* Function vectorizable_shift. |
3096 | ||
3097 | Check if STMT performs a shift operation that can be vectorized. | |
3098 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3099 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3100 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3101 | ||
3102 | static bool | |
3103 | vectorizable_shift (gimple stmt, gimple_stmt_iterator *gsi, | |
3104 | gimple *vec_stmt, slp_tree slp_node) | |
3105 | { | |
3106 | tree vec_dest; | |
3107 | tree scalar_dest; | |
3108 | tree op0, op1 = NULL; | |
3109 | tree vec_oprnd1 = NULL_TREE; | |
3110 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3111 | tree vectype; | |
3112 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3113 | enum tree_code code; | |
3114 | enum machine_mode vec_mode; | |
3115 | tree new_temp; | |
3116 | optab optab; | |
3117 | int icode; | |
3118 | enum machine_mode optab_op2_mode; | |
3119 | tree def; | |
3120 | gimple def_stmt; | |
3121 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
3122 | gimple new_stmt = NULL; | |
3123 | stmt_vec_info prev_stmt_info; | |
3124 | int nunits_in; | |
3125 | int nunits_out; | |
3126 | tree vectype_out; | |
cede2577 | 3127 | tree op1_vectype; |
9dc3f7de IR |
3128 | int ncopies; |
3129 | int j, i; | |
6e1aa848 DN |
3130 | vec<tree> vec_oprnds0 = vNULL; |
3131 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
3132 | tree vop0, vop1; |
3133 | unsigned int k; | |
49eab32e | 3134 | bool scalar_shift_arg = true; |
9dc3f7de IR |
3135 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3136 | int vf; | |
3137 | ||
3138 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
3139 | return false; | |
3140 | ||
3141 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
3142 | return false; | |
3143 | ||
3144 | /* Is STMT a vectorizable binary/unary operation? */ | |
3145 | if (!is_gimple_assign (stmt)) | |
3146 | return false; | |
3147 | ||
3148 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3149 | return false; | |
3150 | ||
3151 | code = gimple_assign_rhs_code (stmt); | |
3152 | ||
3153 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
3154 | || code == RROTATE_EXPR)) | |
3155 | return false; | |
3156 | ||
3157 | scalar_dest = gimple_assign_lhs (stmt); | |
3158 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
7b7b1813 RG |
3159 | if (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
3160 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
3161 | { | |
73fbfcad | 3162 | if (dump_enabled_p ()) |
78c60e3d | 3163 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3164 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
3165 | return false; |
3166 | } | |
9dc3f7de IR |
3167 | |
3168 | op0 = gimple_assign_rhs1 (stmt); | |
24ee1384 | 3169 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
9dc3f7de IR |
3170 | &def_stmt, &def, &dt[0], &vectype)) |
3171 | { | |
73fbfcad | 3172 | if (dump_enabled_p ()) |
78c60e3d | 3173 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3174 | "use not simple.\n"); |
9dc3f7de IR |
3175 | return false; |
3176 | } | |
3177 | /* If op0 is an external or constant def use a vector type with | |
3178 | the same size as the output vector type. */ | |
3179 | if (!vectype) | |
3180 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
3181 | if (vec_stmt) | |
3182 | gcc_assert (vectype); | |
3183 | if (!vectype) | |
3184 | { | |
73fbfcad | 3185 | if (dump_enabled_p ()) |
78c60e3d | 3186 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3187 | "no vectype for scalar type\n"); |
9dc3f7de IR |
3188 | return false; |
3189 | } | |
3190 | ||
3191 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
3192 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
3193 | if (nunits_out != nunits_in) | |
3194 | return false; | |
3195 | ||
3196 | op1 = gimple_assign_rhs2 (stmt); | |
24ee1384 IR |
3197 | if (!vect_is_simple_use_1 (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3198 | &def, &dt[1], &op1_vectype)) | |
9dc3f7de | 3199 | { |
73fbfcad | 3200 | if (dump_enabled_p ()) |
78c60e3d | 3201 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3202 | "use not simple.\n"); |
9dc3f7de IR |
3203 | return false; |
3204 | } | |
3205 | ||
3206 | if (loop_vinfo) | |
3207 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3208 | else | |
3209 | vf = 1; | |
3210 | ||
3211 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3212 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3213 | case of SLP. */ | |
437f4a00 | 3214 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
9dc3f7de IR |
3215 | ncopies = 1; |
3216 | else | |
3217 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
3218 | ||
3219 | gcc_assert (ncopies >= 1); | |
3220 | ||
3221 | /* Determine whether the shift amount is a vector, or scalar. If the | |
3222 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
3223 | ||
49eab32e JJ |
3224 | if (dt[1] == vect_internal_def && !slp_node) |
3225 | scalar_shift_arg = false; | |
3226 | else if (dt[1] == vect_constant_def | |
3227 | || dt[1] == vect_external_def | |
3228 | || dt[1] == vect_internal_def) | |
3229 | { | |
3230 | /* In SLP, need to check whether the shift count is the same, | |
3231 | in loops if it is a constant or invariant, it is always | |
3232 | a scalar shift. */ | |
3233 | if (slp_node) | |
3234 | { | |
9771b263 | 3235 | vec<gimple> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
49eab32e JJ |
3236 | gimple slpstmt; |
3237 | ||
9771b263 | 3238 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
3239 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
3240 | scalar_shift_arg = false; | |
3241 | } | |
3242 | } | |
3243 | else | |
3244 | { | |
73fbfcad | 3245 | if (dump_enabled_p ()) |
78c60e3d | 3246 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3247 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
3248 | return false; |
3249 | } | |
3250 | ||
9dc3f7de | 3251 | /* Vector shifted by vector. */ |
49eab32e | 3252 | if (!scalar_shift_arg) |
9dc3f7de IR |
3253 | { |
3254 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 3255 | if (dump_enabled_p ()) |
78c60e3d | 3256 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3257 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 3258 | |
aa948027 JJ |
3259 | if (!op1_vectype) |
3260 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
3261 | if (op1_vectype == NULL_TREE | |
3262 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 3263 | { |
73fbfcad | 3264 | if (dump_enabled_p ()) |
78c60e3d SS |
3265 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3266 | "unusable type for last operand in" | |
e645e942 | 3267 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
3268 | return false; |
3269 | } | |
9dc3f7de IR |
3270 | } |
3271 | /* See if the machine has a vector shifted by scalar insn and if not | |
3272 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 3273 | else |
9dc3f7de IR |
3274 | { |
3275 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
3276 | if (optab | |
3277 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
3278 | { | |
73fbfcad | 3279 | if (dump_enabled_p ()) |
78c60e3d | 3280 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3281 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
3282 | } |
3283 | else | |
3284 | { | |
3285 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
3286 | if (optab | |
3287 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
3288 | != CODE_FOR_nothing)) | |
3289 | { | |
49eab32e JJ |
3290 | scalar_shift_arg = false; |
3291 | ||
73fbfcad | 3292 | if (dump_enabled_p ()) |
78c60e3d | 3293 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3294 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
3295 | |
3296 | /* Unlike the other binary operators, shifts/rotates have | |
3297 | the rhs being int, instead of the same type as the lhs, | |
3298 | so make sure the scalar is the right type if we are | |
aa948027 | 3299 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
3300 | if (dt[1] == vect_constant_def) |
3301 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
3302 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
3303 | TREE_TYPE (op1))) | |
3304 | { | |
3305 | if (slp_node | |
3306 | && TYPE_MODE (TREE_TYPE (vectype)) | |
3307 | != TYPE_MODE (TREE_TYPE (op1))) | |
3308 | { | |
73fbfcad | 3309 | if (dump_enabled_p ()) |
78c60e3d SS |
3310 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3311 | "unusable type for last operand in" | |
e645e942 | 3312 | " vector/vector shift/rotate.\n"); |
aa948027 JJ |
3313 | return false; |
3314 | } | |
3315 | if (vec_stmt && !slp_node) | |
3316 | { | |
3317 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
3318 | op1 = vect_init_vector (stmt, op1, | |
3319 | TREE_TYPE (vectype), NULL); | |
3320 | } | |
3321 | } | |
9dc3f7de IR |
3322 | } |
3323 | } | |
3324 | } | |
9dc3f7de IR |
3325 | |
3326 | /* Supportable by target? */ | |
3327 | if (!optab) | |
3328 | { | |
73fbfcad | 3329 | if (dump_enabled_p ()) |
78c60e3d | 3330 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3331 | "no optab.\n"); |
9dc3f7de IR |
3332 | return false; |
3333 | } | |
3334 | vec_mode = TYPE_MODE (vectype); | |
3335 | icode = (int) optab_handler (optab, vec_mode); | |
3336 | if (icode == CODE_FOR_nothing) | |
3337 | { | |
73fbfcad | 3338 | if (dump_enabled_p ()) |
78c60e3d | 3339 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3340 | "op not supported by target.\n"); |
9dc3f7de IR |
3341 | /* Check only during analysis. */ |
3342 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
3343 | || (vf < vect_min_worthwhile_factor (code) | |
3344 | && !vec_stmt)) | |
3345 | return false; | |
73fbfcad | 3346 | if (dump_enabled_p ()) |
e645e942 TJ |
3347 | dump_printf_loc (MSG_NOTE, vect_location, |
3348 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
3349 | } |
3350 | ||
3351 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
3352 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
3353 | && vf < vect_min_worthwhile_factor (code) | |
3354 | && !vec_stmt) | |
3355 | { | |
73fbfcad | 3356 | if (dump_enabled_p ()) |
78c60e3d | 3357 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3358 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
3359 | return false; |
3360 | } | |
3361 | ||
3362 | if (!vec_stmt) /* transformation not required. */ | |
3363 | { | |
3364 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 3365 | if (dump_enabled_p ()) |
e645e942 TJ |
3366 | dump_printf_loc (MSG_NOTE, vect_location, |
3367 | "=== vectorizable_shift ===\n"); | |
c3e7ee41 | 3368 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
9dc3f7de IR |
3369 | return true; |
3370 | } | |
3371 | ||
3372 | /** Transform. **/ | |
3373 | ||
73fbfcad | 3374 | if (dump_enabled_p ()) |
78c60e3d | 3375 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3376 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
3377 | |
3378 | /* Handle def. */ | |
3379 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3380 | ||
9dc3f7de IR |
3381 | prev_stmt_info = NULL; |
3382 | for (j = 0; j < ncopies; j++) | |
3383 | { | |
3384 | /* Handle uses. */ | |
3385 | if (j == 0) | |
3386 | { | |
3387 | if (scalar_shift_arg) | |
3388 | { | |
3389 | /* Vector shl and shr insn patterns can be defined with scalar | |
3390 | operand 2 (shift operand). In this case, use constant or loop | |
3391 | invariant op1 directly, without extending it to vector mode | |
3392 | first. */ | |
3393 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
3394 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
3395 | { | |
73fbfcad | 3396 | if (dump_enabled_p ()) |
78c60e3d | 3397 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3398 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 3399 | vec_oprnd1 = op1; |
8930f723 | 3400 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 3401 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
3402 | if (slp_node) |
3403 | { | |
3404 | /* Store vec_oprnd1 for every vector stmt to be created | |
3405 | for SLP_NODE. We check during the analysis that all | |
3406 | the shift arguments are the same. | |
3407 | TODO: Allow different constants for different vector | |
3408 | stmts generated for an SLP instance. */ | |
3409 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 3410 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
3411 | } |
3412 | } | |
3413 | } | |
3414 | ||
3415 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
3416 | (a special case for certain kind of vector shifts); otherwise, | |
3417 | operand 1 should be of a vector type (the usual case). */ | |
3418 | if (vec_oprnd1) | |
3419 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
d092494c | 3420 | slp_node, -1); |
9dc3f7de IR |
3421 | else |
3422 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
d092494c | 3423 | slp_node, -1); |
9dc3f7de IR |
3424 | } |
3425 | else | |
3426 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
3427 | ||
3428 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3429 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 3430 | { |
9771b263 | 3431 | vop1 = vec_oprnds1[i]; |
9dc3f7de IR |
3432 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); |
3433 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3434 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3435 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3436 | if (slp_node) | |
9771b263 | 3437 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
3438 | } |
3439 | ||
3440 | if (slp_node) | |
3441 | continue; | |
3442 | ||
3443 | if (j == 0) | |
3444 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3445 | else | |
3446 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3447 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3448 | } | |
3449 | ||
9771b263 DN |
3450 | vec_oprnds0.release (); |
3451 | vec_oprnds1.release (); | |
9dc3f7de IR |
3452 | |
3453 | return true; | |
3454 | } | |
3455 | ||
3456 | ||
5deb57cb JJ |
3457 | static tree permute_vec_elements (tree, tree, tree, gimple, |
3458 | gimple_stmt_iterator *); | |
3459 | ||
3460 | ||
ebfd146a IR |
3461 | /* Function vectorizable_operation. |
3462 | ||
16949072 RG |
3463 | Check if STMT performs a binary, unary or ternary operation that can |
3464 | be vectorized. | |
b8698a0f | 3465 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
3466 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3467 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3468 | ||
3469 | static bool | |
3470 | vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi, | |
3471 | gimple *vec_stmt, slp_tree slp_node) | |
3472 | { | |
00f07b86 | 3473 | tree vec_dest; |
ebfd146a | 3474 | tree scalar_dest; |
16949072 | 3475 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 3476 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 3477 | tree vectype; |
ebfd146a IR |
3478 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
3479 | enum tree_code code; | |
3480 | enum machine_mode vec_mode; | |
3481 | tree new_temp; | |
3482 | int op_type; | |
00f07b86 | 3483 | optab optab; |
ebfd146a | 3484 | int icode; |
ebfd146a IR |
3485 | tree def; |
3486 | gimple def_stmt; | |
16949072 RG |
3487 | enum vect_def_type dt[3] |
3488 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
ebfd146a IR |
3489 | gimple new_stmt = NULL; |
3490 | stmt_vec_info prev_stmt_info; | |
b690cc0f | 3491 | int nunits_in; |
ebfd146a IR |
3492 | int nunits_out; |
3493 | tree vectype_out; | |
3494 | int ncopies; | |
3495 | int j, i; | |
6e1aa848 DN |
3496 | vec<tree> vec_oprnds0 = vNULL; |
3497 | vec<tree> vec_oprnds1 = vNULL; | |
3498 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 3499 | tree vop0, vop1, vop2; |
a70d6342 IR |
3500 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3501 | int vf; | |
3502 | ||
a70d6342 | 3503 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3504 | return false; |
3505 | ||
8644a673 | 3506 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3507 | return false; |
3508 | ||
3509 | /* Is STMT a vectorizable binary/unary operation? */ | |
3510 | if (!is_gimple_assign (stmt)) | |
3511 | return false; | |
3512 | ||
3513 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3514 | return false; | |
3515 | ||
ebfd146a IR |
3516 | code = gimple_assign_rhs_code (stmt); |
3517 | ||
3518 | /* For pointer addition, we should use the normal plus for | |
3519 | the vector addition. */ | |
3520 | if (code == POINTER_PLUS_EXPR) | |
3521 | code = PLUS_EXPR; | |
3522 | ||
3523 | /* Support only unary or binary operations. */ | |
3524 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 3525 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 3526 | { |
73fbfcad | 3527 | if (dump_enabled_p ()) |
78c60e3d | 3528 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3529 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 3530 | op_type); |
ebfd146a IR |
3531 | return false; |
3532 | } | |
3533 | ||
b690cc0f RG |
3534 | scalar_dest = gimple_assign_lhs (stmt); |
3535 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
3536 | ||
7b7b1813 RG |
3537 | /* Most operations cannot handle bit-precision types without extra |
3538 | truncations. */ | |
3539 | if ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
3540 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
3541 | /* Exception are bitwise binary operations. */ | |
3542 | && code != BIT_IOR_EXPR | |
3543 | && code != BIT_XOR_EXPR | |
3544 | && code != BIT_AND_EXPR) | |
3545 | { | |
73fbfcad | 3546 | if (dump_enabled_p ()) |
78c60e3d | 3547 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3548 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
3549 | return false; |
3550 | } | |
3551 | ||
ebfd146a | 3552 | op0 = gimple_assign_rhs1 (stmt); |
24ee1384 | 3553 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
b690cc0f | 3554 | &def_stmt, &def, &dt[0], &vectype)) |
ebfd146a | 3555 | { |
73fbfcad | 3556 | if (dump_enabled_p ()) |
78c60e3d | 3557 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3558 | "use not simple.\n"); |
ebfd146a IR |
3559 | return false; |
3560 | } | |
b690cc0f RG |
3561 | /* If op0 is an external or constant def use a vector type with |
3562 | the same size as the output vector type. */ | |
3563 | if (!vectype) | |
3564 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
3565 | if (vec_stmt) |
3566 | gcc_assert (vectype); | |
3567 | if (!vectype) | |
3568 | { | |
73fbfcad | 3569 | if (dump_enabled_p ()) |
7d8930a0 | 3570 | { |
78c60e3d SS |
3571 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3572 | "no vectype for scalar type "); | |
3573 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
3574 | TREE_TYPE (op0)); | |
e645e942 | 3575 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
3576 | } |
3577 | ||
3578 | return false; | |
3579 | } | |
b690cc0f RG |
3580 | |
3581 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
3582 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
3583 | if (nunits_out != nunits_in) | |
3584 | return false; | |
ebfd146a | 3585 | |
16949072 | 3586 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
3587 | { |
3588 | op1 = gimple_assign_rhs2 (stmt); | |
24ee1384 IR |
3589 | if (!vect_is_simple_use (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3590 | &def, &dt[1])) | |
ebfd146a | 3591 | { |
73fbfcad | 3592 | if (dump_enabled_p ()) |
78c60e3d | 3593 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3594 | "use not simple.\n"); |
ebfd146a IR |
3595 | return false; |
3596 | } | |
3597 | } | |
16949072 RG |
3598 | if (op_type == ternary_op) |
3599 | { | |
3600 | op2 = gimple_assign_rhs3 (stmt); | |
24ee1384 IR |
3601 | if (!vect_is_simple_use (op2, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3602 | &def, &dt[2])) | |
16949072 | 3603 | { |
73fbfcad | 3604 | if (dump_enabled_p ()) |
78c60e3d | 3605 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3606 | "use not simple.\n"); |
16949072 RG |
3607 | return false; |
3608 | } | |
3609 | } | |
ebfd146a | 3610 | |
b690cc0f RG |
3611 | if (loop_vinfo) |
3612 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3613 | else | |
3614 | vf = 1; | |
3615 | ||
3616 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 3617 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 3618 | case of SLP. */ |
437f4a00 | 3619 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
b690cc0f RG |
3620 | ncopies = 1; |
3621 | else | |
3622 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
3623 | ||
3624 | gcc_assert (ncopies >= 1); | |
3625 | ||
9dc3f7de | 3626 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
3627 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
3628 | || code == RROTATE_EXPR) | |
9dc3f7de | 3629 | return false; |
ebfd146a | 3630 | |
ebfd146a | 3631 | /* Supportable by target? */ |
00f07b86 RH |
3632 | |
3633 | vec_mode = TYPE_MODE (vectype); | |
3634 | if (code == MULT_HIGHPART_EXPR) | |
ebfd146a | 3635 | { |
00f07b86 | 3636 | if (can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype))) |
dee54b6e | 3637 | icode = LAST_INSN_CODE; |
00f07b86 RH |
3638 | else |
3639 | icode = CODE_FOR_nothing; | |
ebfd146a | 3640 | } |
00f07b86 RH |
3641 | else |
3642 | { | |
3643 | optab = optab_for_tree_code (code, vectype, optab_default); | |
3644 | if (!optab) | |
5deb57cb | 3645 | { |
73fbfcad | 3646 | if (dump_enabled_p ()) |
78c60e3d | 3647 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3648 | "no optab.\n"); |
00f07b86 | 3649 | return false; |
5deb57cb | 3650 | } |
00f07b86 | 3651 | icode = (int) optab_handler (optab, vec_mode); |
5deb57cb JJ |
3652 | } |
3653 | ||
ebfd146a IR |
3654 | if (icode == CODE_FOR_nothing) |
3655 | { | |
73fbfcad | 3656 | if (dump_enabled_p ()) |
78c60e3d | 3657 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3658 | "op not supported by target.\n"); |
ebfd146a IR |
3659 | /* Check only during analysis. */ |
3660 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5deb57cb | 3661 | || (!vec_stmt && vf < vect_min_worthwhile_factor (code))) |
ebfd146a | 3662 | return false; |
73fbfcad | 3663 | if (dump_enabled_p ()) |
e645e942 TJ |
3664 | dump_printf_loc (MSG_NOTE, vect_location, |
3665 | "proceeding using word mode.\n"); | |
383d9c83 IR |
3666 | } |
3667 | ||
4a00c761 | 3668 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
3669 | if (!VECTOR_MODE_P (vec_mode) |
3670 | && !vec_stmt | |
3671 | && vf < vect_min_worthwhile_factor (code)) | |
7d8930a0 | 3672 | { |
73fbfcad | 3673 | if (dump_enabled_p ()) |
78c60e3d | 3674 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3675 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 3676 | return false; |
7d8930a0 | 3677 | } |
ebfd146a | 3678 | |
ebfd146a IR |
3679 | if (!vec_stmt) /* transformation not required. */ |
3680 | { | |
4a00c761 | 3681 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 3682 | if (dump_enabled_p ()) |
78c60e3d | 3683 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3684 | "=== vectorizable_operation ===\n"); |
c3e7ee41 | 3685 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
3686 | return true; |
3687 | } | |
3688 | ||
3689 | /** Transform. **/ | |
3690 | ||
73fbfcad | 3691 | if (dump_enabled_p ()) |
78c60e3d | 3692 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3693 | "transform binary/unary operation.\n"); |
383d9c83 | 3694 | |
ebfd146a | 3695 | /* Handle def. */ |
00f07b86 | 3696 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 3697 | |
ebfd146a IR |
3698 | /* In case the vectorization factor (VF) is bigger than the number |
3699 | of elements that we can fit in a vectype (nunits), we have to generate | |
3700 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
3701 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
3702 | from one copy of the vector stmt to the next, in the field | |
3703 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
3704 | stages to find the correct vector defs to be used when vectorizing | |
3705 | stmts that use the defs of the current stmt. The example below | |
3706 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
3707 | we need to create 4 vectorized stmts): | |
3708 | ||
3709 | before vectorization: | |
3710 | RELATED_STMT VEC_STMT | |
3711 | S1: x = memref - - | |
3712 | S2: z = x + 1 - - | |
3713 | ||
3714 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
3715 | there): | |
3716 | RELATED_STMT VEC_STMT | |
3717 | VS1_0: vx0 = memref0 VS1_1 - | |
3718 | VS1_1: vx1 = memref1 VS1_2 - | |
3719 | VS1_2: vx2 = memref2 VS1_3 - | |
3720 | VS1_3: vx3 = memref3 - - | |
3721 | S1: x = load - VS1_0 | |
3722 | S2: z = x + 1 - - | |
3723 | ||
3724 | step2: vectorize stmt S2 (done here): | |
3725 | To vectorize stmt S2 we first need to find the relevant vector | |
3726 | def for the first operand 'x'. This is, as usual, obtained from | |
3727 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
3728 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
3729 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
3730 | the vector stmt VS2_0, and as usual, record it in the | |
3731 | STMT_VINFO_VEC_STMT of stmt S2. | |
3732 | When creating the second copy (VS2_1), we obtain the relevant vector | |
3733 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
3734 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
3735 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
3736 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
3737 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
3738 | chain of stmts and pointers: | |
3739 | RELATED_STMT VEC_STMT | |
3740 | VS1_0: vx0 = memref0 VS1_1 - | |
3741 | VS1_1: vx1 = memref1 VS1_2 - | |
3742 | VS1_2: vx2 = memref2 VS1_3 - | |
3743 | VS1_3: vx3 = memref3 - - | |
3744 | S1: x = load - VS1_0 | |
3745 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
3746 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
3747 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
3748 | VS2_3: vz3 = vx3 + v1 - - | |
3749 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
3750 | |
3751 | prev_stmt_info = NULL; | |
3752 | for (j = 0; j < ncopies; j++) | |
3753 | { | |
3754 | /* Handle uses. */ | |
3755 | if (j == 0) | |
4a00c761 JJ |
3756 | { |
3757 | if (op_type == binary_op || op_type == ternary_op) | |
3758 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
3759 | slp_node, -1); | |
3760 | else | |
3761 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
3762 | slp_node, -1); | |
3763 | if (op_type == ternary_op) | |
36ba4aae | 3764 | { |
9771b263 DN |
3765 | vec_oprnds2.create (1); |
3766 | vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2, | |
3767 | stmt, | |
3768 | NULL)); | |
36ba4aae | 3769 | } |
4a00c761 | 3770 | } |
ebfd146a | 3771 | else |
4a00c761 JJ |
3772 | { |
3773 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
3774 | if (op_type == ternary_op) | |
3775 | { | |
9771b263 DN |
3776 | tree vec_oprnd = vec_oprnds2.pop (); |
3777 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
3778 | vec_oprnd)); | |
4a00c761 JJ |
3779 | } |
3780 | } | |
3781 | ||
3782 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3783 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 3784 | { |
4a00c761 | 3785 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 3786 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 3787 | vop2 = ((op_type == ternary_op) |
9771b263 | 3788 | ? vec_oprnds2[i] : NULL_TREE); |
73804b12 RG |
3789 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, |
3790 | vop0, vop1, vop2); | |
4a00c761 JJ |
3791 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3792 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3793 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3794 | if (slp_node) | |
9771b263 | 3795 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
3796 | } |
3797 | ||
4a00c761 JJ |
3798 | if (slp_node) |
3799 | continue; | |
3800 | ||
3801 | if (j == 0) | |
3802 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3803 | else | |
3804 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3805 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
3806 | } |
3807 | ||
9771b263 DN |
3808 | vec_oprnds0.release (); |
3809 | vec_oprnds1.release (); | |
3810 | vec_oprnds2.release (); | |
ebfd146a | 3811 | |
ebfd146a IR |
3812 | return true; |
3813 | } | |
3814 | ||
c716e67f XDL |
3815 | /* A helper function to ensure data reference DR's base alignment |
3816 | for STMT_INFO. */ | |
3817 | ||
3818 | static void | |
3819 | ensure_base_align (stmt_vec_info stmt_info, struct data_reference *dr) | |
3820 | { | |
3821 | if (!dr->aux) | |
3822 | return; | |
3823 | ||
3824 | if (((dataref_aux *)dr->aux)->base_misaligned) | |
3825 | { | |
3826 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3827 | tree base_decl = ((dataref_aux *)dr->aux)->base_decl; | |
3828 | ||
3829 | DECL_ALIGN (base_decl) = TYPE_ALIGN (vectype); | |
3830 | DECL_USER_ALIGN (base_decl) = 1; | |
3831 | ((dataref_aux *)dr->aux)->base_misaligned = false; | |
3832 | } | |
3833 | } | |
3834 | ||
ebfd146a IR |
3835 | |
3836 | /* Function vectorizable_store. | |
3837 | ||
b8698a0f L |
3838 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
3839 | can be vectorized. | |
3840 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
3841 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3842 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3843 | ||
3844 | static bool | |
3845 | vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
c716e67f | 3846 | slp_tree slp_node) |
ebfd146a IR |
3847 | { |
3848 | tree scalar_dest; | |
3849 | tree data_ref; | |
3850 | tree op; | |
3851 | tree vec_oprnd = NULL_TREE; | |
3852 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3853 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
3854 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
272c6793 | 3855 | tree elem_type; |
ebfd146a | 3856 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 3857 | struct loop *loop = NULL; |
ebfd146a IR |
3858 | enum machine_mode vec_mode; |
3859 | tree dummy; | |
3860 | enum dr_alignment_support alignment_support_scheme; | |
3861 | tree def; | |
3862 | gimple def_stmt; | |
3863 | enum vect_def_type dt; | |
3864 | stmt_vec_info prev_stmt_info = NULL; | |
3865 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 3866 | tree dataref_offset = NULL_TREE; |
fef4d2b3 | 3867 | gimple ptr_incr = NULL; |
ebfd146a IR |
3868 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
3869 | int ncopies; | |
3870 | int j; | |
3871 | gimple next_stmt, first_stmt = NULL; | |
0d0293ac | 3872 | bool grouped_store = false; |
272c6793 | 3873 | bool store_lanes_p = false; |
ebfd146a | 3874 | unsigned int group_size, i; |
6e1aa848 DN |
3875 | vec<tree> dr_chain = vNULL; |
3876 | vec<tree> oprnds = vNULL; | |
3877 | vec<tree> result_chain = vNULL; | |
ebfd146a | 3878 | bool inv_p; |
6e1aa848 | 3879 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 3880 | bool slp = (slp_node != NULL); |
ebfd146a | 3881 | unsigned int vec_num; |
a70d6342 | 3882 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
272c6793 | 3883 | tree aggr_type; |
a70d6342 IR |
3884 | |
3885 | if (loop_vinfo) | |
3886 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
3887 | |
3888 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3889 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3890 | case of SLP. */ | |
437f4a00 | 3891 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
3892 | ncopies = 1; |
3893 | else | |
3894 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3895 | ||
3896 | gcc_assert (ncopies >= 1); | |
3897 | ||
3898 | /* FORNOW. This restriction should be relaxed. */ | |
a70d6342 | 3899 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) |
ebfd146a | 3900 | { |
73fbfcad | 3901 | if (dump_enabled_p ()) |
78c60e3d | 3902 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3903 | "multiple types in nested loop.\n"); |
ebfd146a IR |
3904 | return false; |
3905 | } | |
3906 | ||
a70d6342 | 3907 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3908 | return false; |
3909 | ||
8644a673 | 3910 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3911 | return false; |
3912 | ||
3913 | /* Is vectorizable store? */ | |
3914 | ||
3915 | if (!is_gimple_assign (stmt)) | |
3916 | return false; | |
3917 | ||
3918 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
3919 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
3920 | && is_pattern_stmt_p (stmt_info)) | |
3921 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 3922 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 3923 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 3924 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
3925 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
3926 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
3927 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
3928 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
3929 | return false; |
3930 | ||
3931 | gcc_assert (gimple_assign_single_p (stmt)); | |
3932 | op = gimple_assign_rhs1 (stmt); | |
24ee1384 IR |
3933 | if (!vect_is_simple_use (op, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3934 | &def, &dt)) | |
ebfd146a | 3935 | { |
73fbfcad | 3936 | if (dump_enabled_p ()) |
78c60e3d | 3937 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3938 | "use not simple.\n"); |
ebfd146a IR |
3939 | return false; |
3940 | } | |
3941 | ||
272c6793 | 3942 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 3943 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 3944 | |
ebfd146a IR |
3945 | /* FORNOW. In some cases can vectorize even if data-type not supported |
3946 | (e.g. - array initialization with 0). */ | |
947131ba | 3947 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
3948 | return false; |
3949 | ||
3950 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
3951 | return false; | |
3952 | ||
a7ce6ec3 RG |
3953 | if (tree_int_cst_compare (loop && nested_in_vect_loop_p (loop, stmt) |
3954 | ? STMT_VINFO_DR_STEP (stmt_info) : DR_STEP (dr), | |
3955 | size_zero_node) < 0) | |
a1e53f3f | 3956 | { |
73fbfcad | 3957 | if (dump_enabled_p ()) |
78c60e3d | 3958 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3959 | "negative step for store.\n"); |
a1e53f3f L |
3960 | return false; |
3961 | } | |
3962 | ||
0d0293ac | 3963 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 3964 | { |
0d0293ac | 3965 | grouped_store = true; |
e14c1050 | 3966 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
3967 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
3968 | { | |
e14c1050 | 3969 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
3970 | if (vect_store_lanes_supported (vectype, group_size)) |
3971 | store_lanes_p = true; | |
0d0293ac | 3972 | else if (!vect_grouped_store_supported (vectype, group_size)) |
b602d918 RS |
3973 | return false; |
3974 | } | |
b8698a0f | 3975 | |
ebfd146a IR |
3976 | if (first_stmt == stmt) |
3977 | { | |
3978 | /* STMT is the leader of the group. Check the operands of all the | |
3979 | stmts of the group. */ | |
e14c1050 | 3980 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
3981 | while (next_stmt) |
3982 | { | |
3983 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
3984 | op = gimple_assign_rhs1 (next_stmt); | |
24ee1384 IR |
3985 | if (!vect_is_simple_use (op, next_stmt, loop_vinfo, bb_vinfo, |
3986 | &def_stmt, &def, &dt)) | |
ebfd146a | 3987 | { |
73fbfcad | 3988 | if (dump_enabled_p ()) |
78c60e3d | 3989 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3990 | "use not simple.\n"); |
ebfd146a IR |
3991 | return false; |
3992 | } | |
e14c1050 | 3993 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
3994 | } |
3995 | } | |
3996 | } | |
3997 | ||
3998 | if (!vec_stmt) /* transformation not required. */ | |
3999 | { | |
4000 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
92345349 BS |
4001 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, |
4002 | NULL, NULL, NULL); | |
ebfd146a IR |
4003 | return true; |
4004 | } | |
4005 | ||
4006 | /** Transform. **/ | |
4007 | ||
c716e67f XDL |
4008 | ensure_base_align (stmt_info, dr); |
4009 | ||
0d0293ac | 4010 | if (grouped_store) |
ebfd146a IR |
4011 | { |
4012 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 4013 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 4014 | |
e14c1050 | 4015 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
4016 | |
4017 | /* FORNOW */ | |
a70d6342 | 4018 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
4019 | |
4020 | /* We vectorize all the stmts of the interleaving group when we | |
4021 | reach the last stmt in the group. */ | |
e14c1050 IR |
4022 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
4023 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
4024 | && !slp) |
4025 | { | |
4026 | *vec_stmt = NULL; | |
4027 | return true; | |
4028 | } | |
4029 | ||
4030 | if (slp) | |
4b5caab7 | 4031 | { |
0d0293ac | 4032 | grouped_store = false; |
4b5caab7 IR |
4033 | /* VEC_NUM is the number of vect stmts to be created for this |
4034 | group. */ | |
4035 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 4036 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
4b5caab7 | 4037 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 4038 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 4039 | } |
ebfd146a | 4040 | else |
4b5caab7 IR |
4041 | /* VEC_NUM is the number of vect stmts to be created for this |
4042 | group. */ | |
ebfd146a IR |
4043 | vec_num = group_size; |
4044 | } | |
b8698a0f | 4045 | else |
ebfd146a IR |
4046 | { |
4047 | first_stmt = stmt; | |
4048 | first_dr = dr; | |
4049 | group_size = vec_num = 1; | |
ebfd146a | 4050 | } |
b8698a0f | 4051 | |
73fbfcad | 4052 | if (dump_enabled_p ()) |
78c60e3d | 4053 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4054 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 4055 | |
9771b263 DN |
4056 | dr_chain.create (group_size); |
4057 | oprnds.create (group_size); | |
ebfd146a | 4058 | |
720f5239 | 4059 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 4060 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
4061 | /* Targets with store-lane instructions must not require explicit |
4062 | realignment. */ | |
4063 | gcc_assert (!store_lanes_p | |
4064 | || alignment_support_scheme == dr_aligned | |
4065 | || alignment_support_scheme == dr_unaligned_supported); | |
4066 | ||
4067 | if (store_lanes_p) | |
4068 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
4069 | else | |
4070 | aggr_type = vectype; | |
ebfd146a IR |
4071 | |
4072 | /* In case the vectorization factor (VF) is bigger than the number | |
4073 | of elements that we can fit in a vectype (nunits), we have to generate | |
4074 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 4075 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
4076 | vect_get_vec_def_for_copy_stmt. */ |
4077 | ||
0d0293ac | 4078 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
4079 | |
4080 | S1: &base + 2 = x2 | |
4081 | S2: &base = x0 | |
4082 | S3: &base + 1 = x1 | |
4083 | S4: &base + 3 = x3 | |
4084 | ||
4085 | We create vectorized stores starting from base address (the access of the | |
4086 | first stmt in the chain (S2 in the above example), when the last store stmt | |
4087 | of the chain (S4) is reached: | |
4088 | ||
4089 | VS1: &base = vx2 | |
4090 | VS2: &base + vec_size*1 = vx0 | |
4091 | VS3: &base + vec_size*2 = vx1 | |
4092 | VS4: &base + vec_size*3 = vx3 | |
4093 | ||
4094 | Then permutation statements are generated: | |
4095 | ||
3fcc1b55 JJ |
4096 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
4097 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 4098 | ... |
b8698a0f | 4099 | |
ebfd146a IR |
4100 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
4101 | (the order of the data-refs in the output of vect_permute_store_chain | |
4102 | corresponds to the order of scalar stmts in the interleaving chain - see | |
4103 | the documentation of vect_permute_store_chain()). | |
4104 | ||
4105 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 4106 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 4107 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 4108 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
4109 | */ |
4110 | ||
4111 | prev_stmt_info = NULL; | |
4112 | for (j = 0; j < ncopies; j++) | |
4113 | { | |
4114 | gimple new_stmt; | |
ebfd146a IR |
4115 | |
4116 | if (j == 0) | |
4117 | { | |
4118 | if (slp) | |
4119 | { | |
4120 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c IR |
4121 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
4122 | NULL, slp_node, -1); | |
ebfd146a | 4123 | |
9771b263 | 4124 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
4125 | } |
4126 | else | |
4127 | { | |
b8698a0f L |
4128 | /* For interleaved stores we collect vectorized defs for all the |
4129 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
4130 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
4131 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
4132 | ||
0d0293ac | 4133 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 4134 | OPRNDS are of size 1. */ |
b8698a0f | 4135 | next_stmt = first_stmt; |
ebfd146a IR |
4136 | for (i = 0; i < group_size; i++) |
4137 | { | |
b8698a0f L |
4138 | /* Since gaps are not supported for interleaved stores, |
4139 | GROUP_SIZE is the exact number of stmts in the chain. | |
4140 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
4141 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
4142 | iteration of the loop will be executed. */ |
4143 | gcc_assert (next_stmt | |
4144 | && gimple_assign_single_p (next_stmt)); | |
4145 | op = gimple_assign_rhs1 (next_stmt); | |
4146 | ||
b8698a0f | 4147 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt, |
ebfd146a | 4148 | NULL); |
9771b263 DN |
4149 | dr_chain.quick_push (vec_oprnd); |
4150 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 4151 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
4152 | } |
4153 | } | |
4154 | ||
4155 | /* We should have catched mismatched types earlier. */ | |
4156 | gcc_assert (useless_type_conversion_p (vectype, | |
4157 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
4158 | bool simd_lane_access_p |
4159 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
4160 | if (simd_lane_access_p | |
4161 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
4162 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
4163 | && integer_zerop (DR_OFFSET (first_dr)) | |
4164 | && integer_zerop (DR_INIT (first_dr)) | |
4165 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
4166 | get_alias_set (DR_REF (first_dr)))) | |
4167 | { | |
4168 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
4169 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
4170 | (DR_REF (first_dr)), 0); | |
8928eff3 | 4171 | inv_p = false; |
74bf76ed JJ |
4172 | } |
4173 | else | |
4174 | dataref_ptr | |
4175 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
4176 | simd_lane_access_p ? loop : NULL, | |
4177 | NULL_TREE, &dummy, gsi, &ptr_incr, | |
4178 | simd_lane_access_p, &inv_p); | |
a70d6342 | 4179 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 4180 | } |
b8698a0f | 4181 | else |
ebfd146a | 4182 | { |
b8698a0f L |
4183 | /* For interleaved stores we created vectorized defs for all the |
4184 | defs stored in OPRNDS in the previous iteration (previous copy). | |
4185 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
4186 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
4187 | next copy. | |
0d0293ac | 4188 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
4189 | OPRNDS are of size 1. */ |
4190 | for (i = 0; i < group_size; i++) | |
4191 | { | |
9771b263 | 4192 | op = oprnds[i]; |
24ee1384 IR |
4193 | vect_is_simple_use (op, NULL, loop_vinfo, bb_vinfo, &def_stmt, |
4194 | &def, &dt); | |
b8698a0f | 4195 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
4196 | dr_chain[i] = vec_oprnd; |
4197 | oprnds[i] = vec_oprnd; | |
ebfd146a | 4198 | } |
74bf76ed JJ |
4199 | if (dataref_offset) |
4200 | dataref_offset | |
4201 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
4202 | TYPE_SIZE_UNIT (aggr_type)); | |
4203 | else | |
4204 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
4205 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
4206 | } |
4207 | ||
272c6793 | 4208 | if (store_lanes_p) |
ebfd146a | 4209 | { |
272c6793 | 4210 | tree vec_array; |
267d3070 | 4211 | |
272c6793 RS |
4212 | /* Combine all the vectors into an array. */ |
4213 | vec_array = create_vector_array (vectype, vec_num); | |
4214 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 4215 | { |
9771b263 | 4216 | vec_oprnd = dr_chain[i]; |
272c6793 | 4217 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 4218 | } |
b8698a0f | 4219 | |
272c6793 RS |
4220 | /* Emit: |
4221 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
4222 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
4223 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
4224 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 | 4225 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
4226 | } |
4227 | else | |
4228 | { | |
4229 | new_stmt = NULL; | |
0d0293ac | 4230 | if (grouped_store) |
272c6793 | 4231 | { |
b6b9227d JJ |
4232 | if (j == 0) |
4233 | result_chain.create (group_size); | |
272c6793 RS |
4234 | /* Permute. */ |
4235 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
4236 | &result_chain); | |
4237 | } | |
c2d7ab2a | 4238 | |
272c6793 RS |
4239 | next_stmt = first_stmt; |
4240 | for (i = 0; i < vec_num; i++) | |
4241 | { | |
644ffefd | 4242 | unsigned align, misalign; |
272c6793 RS |
4243 | |
4244 | if (i > 0) | |
4245 | /* Bump the vector pointer. */ | |
4246 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
4247 | stmt, NULL_TREE); | |
4248 | ||
4249 | if (slp) | |
9771b263 | 4250 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
4251 | else if (grouped_store) |
4252 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 4253 | vect_permute_store_chain(). */ |
9771b263 | 4254 | vec_oprnd = result_chain[i]; |
272c6793 RS |
4255 | |
4256 | data_ref = build2 (MEM_REF, TREE_TYPE (vec_oprnd), dataref_ptr, | |
74bf76ed JJ |
4257 | dataref_offset |
4258 | ? dataref_offset | |
4259 | : build_int_cst (reference_alias_ptr_type | |
4260 | (DR_REF (first_dr)), 0)); | |
644ffefd | 4261 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 | 4262 | if (aligned_access_p (first_dr)) |
644ffefd | 4263 | misalign = 0; |
272c6793 RS |
4264 | else if (DR_MISALIGNMENT (first_dr) == -1) |
4265 | { | |
4266 | TREE_TYPE (data_ref) | |
4267 | = build_aligned_type (TREE_TYPE (data_ref), | |
4268 | TYPE_ALIGN (elem_type)); | |
644ffefd MJ |
4269 | align = TYPE_ALIGN_UNIT (elem_type); |
4270 | misalign = 0; | |
272c6793 RS |
4271 | } |
4272 | else | |
4273 | { | |
4274 | TREE_TYPE (data_ref) | |
4275 | = build_aligned_type (TREE_TYPE (data_ref), | |
4276 | TYPE_ALIGN (elem_type)); | |
644ffefd | 4277 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 4278 | } |
74bf76ed JJ |
4279 | if (dataref_offset == NULL_TREE) |
4280 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
4281 | misalign); | |
c2d7ab2a | 4282 | |
272c6793 RS |
4283 | /* Arguments are ready. Create the new vector stmt. */ |
4284 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
4285 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
4286 | |
4287 | if (slp) | |
4288 | continue; | |
4289 | ||
e14c1050 | 4290 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
4291 | if (!next_stmt) |
4292 | break; | |
4293 | } | |
ebfd146a | 4294 | } |
1da0876c RS |
4295 | if (!slp) |
4296 | { | |
4297 | if (j == 0) | |
4298 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4299 | else | |
4300 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4301 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4302 | } | |
ebfd146a IR |
4303 | } |
4304 | ||
9771b263 DN |
4305 | dr_chain.release (); |
4306 | oprnds.release (); | |
4307 | result_chain.release (); | |
4308 | vec_oprnds.release (); | |
ebfd146a IR |
4309 | |
4310 | return true; | |
4311 | } | |
4312 | ||
aec7ae7d JJ |
4313 | /* Given a vector type VECTYPE and permutation SEL returns |
4314 | the VECTOR_CST mask that implements the permutation of the | |
4315 | vector elements. If that is impossible to do, returns NULL. */ | |
a1e53f3f | 4316 | |
3fcc1b55 JJ |
4317 | tree |
4318 | vect_gen_perm_mask (tree vectype, unsigned char *sel) | |
a1e53f3f | 4319 | { |
d2a12ae7 | 4320 | tree mask_elt_type, mask_type, mask_vec, *mask_elts; |
2635892a | 4321 | int i, nunits; |
a1e53f3f | 4322 | |
22e4dee7 | 4323 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 RH |
4324 | |
4325 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) | |
a1e53f3f L |
4326 | return NULL; |
4327 | ||
96f9265a RG |
4328 | mask_elt_type = lang_hooks.types.type_for_mode |
4329 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
22e4dee7 | 4330 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 4331 | |
d2a12ae7 | 4332 | mask_elts = XALLOCAVEC (tree, nunits); |
aec7ae7d | 4333 | for (i = nunits - 1; i >= 0; i--) |
d2a12ae7 RG |
4334 | mask_elts[i] = build_int_cst (mask_elt_type, sel[i]); |
4335 | mask_vec = build_vector (mask_type, mask_elts); | |
a1e53f3f | 4336 | |
2635892a | 4337 | return mask_vec; |
a1e53f3f L |
4338 | } |
4339 | ||
aec7ae7d JJ |
4340 | /* Given a vector type VECTYPE returns the VECTOR_CST mask that implements |
4341 | reversal of the vector elements. If that is impossible to do, | |
4342 | returns NULL. */ | |
4343 | ||
4344 | static tree | |
4345 | perm_mask_for_reverse (tree vectype) | |
4346 | { | |
4347 | int i, nunits; | |
4348 | unsigned char *sel; | |
4349 | ||
4350 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4351 | sel = XALLOCAVEC (unsigned char, nunits); | |
4352 | ||
4353 | for (i = 0; i < nunits; ++i) | |
4354 | sel[i] = nunits - 1 - i; | |
4355 | ||
3fcc1b55 | 4356 | return vect_gen_perm_mask (vectype, sel); |
aec7ae7d JJ |
4357 | } |
4358 | ||
4359 | /* Given a vector variable X and Y, that was generated for the scalar | |
4360 | STMT, generate instructions to permute the vector elements of X and Y | |
4361 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
4362 | permuted vector variable. */ | |
a1e53f3f L |
4363 | |
4364 | static tree | |
aec7ae7d JJ |
4365 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple stmt, |
4366 | gimple_stmt_iterator *gsi) | |
a1e53f3f L |
4367 | { |
4368 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 4369 | tree perm_dest, data_ref; |
a1e53f3f L |
4370 | gimple perm_stmt; |
4371 | ||
a1e53f3f | 4372 | perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype); |
aec7ae7d | 4373 | data_ref = make_ssa_name (perm_dest, NULL); |
a1e53f3f L |
4374 | |
4375 | /* Generate the permute statement. */ | |
73804b12 RG |
4376 | perm_stmt = gimple_build_assign_with_ops (VEC_PERM_EXPR, data_ref, |
4377 | x, y, mask_vec); | |
a1e53f3f L |
4378 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
4379 | ||
4380 | return data_ref; | |
4381 | } | |
4382 | ||
ebfd146a IR |
4383 | /* vectorizable_load. |
4384 | ||
b8698a0f L |
4385 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
4386 | can be vectorized. | |
4387 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4388 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4389 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4390 | ||
4391 | static bool | |
4392 | vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
c716e67f | 4393 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
4394 | { |
4395 | tree scalar_dest; | |
4396 | tree vec_dest = NULL; | |
4397 | tree data_ref = NULL; | |
4398 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 4399 | stmt_vec_info prev_stmt_info; |
ebfd146a | 4400 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 4401 | struct loop *loop = NULL; |
ebfd146a | 4402 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 4403 | bool nested_in_vect_loop = false; |
c716e67f | 4404 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
ebfd146a | 4405 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
272c6793 | 4406 | tree elem_type; |
ebfd146a | 4407 | tree new_temp; |
947131ba | 4408 | enum machine_mode mode; |
ebfd146a IR |
4409 | gimple new_stmt = NULL; |
4410 | tree dummy; | |
4411 | enum dr_alignment_support alignment_support_scheme; | |
4412 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 4413 | tree dataref_offset = NULL_TREE; |
fef4d2b3 | 4414 | gimple ptr_incr = NULL; |
ebfd146a IR |
4415 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
4416 | int ncopies; | |
a64b9c26 | 4417 | int i, j, group_size, group_gap; |
ebfd146a IR |
4418 | tree msq = NULL_TREE, lsq; |
4419 | tree offset = NULL_TREE; | |
4420 | tree realignment_token = NULL_TREE; | |
4421 | gimple phi = NULL; | |
6e1aa848 | 4422 | vec<tree> dr_chain = vNULL; |
0d0293ac | 4423 | bool grouped_load = false; |
272c6793 | 4424 | bool load_lanes_p = false; |
ebfd146a | 4425 | gimple first_stmt; |
ebfd146a | 4426 | bool inv_p; |
319e6439 | 4427 | bool negative = false; |
ebfd146a IR |
4428 | bool compute_in_loop = false; |
4429 | struct loop *at_loop; | |
4430 | int vec_num; | |
4431 | bool slp = (slp_node != NULL); | |
4432 | bool slp_perm = false; | |
4433 | enum tree_code code; | |
a70d6342 IR |
4434 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
4435 | int vf; | |
272c6793 | 4436 | tree aggr_type; |
aec7ae7d JJ |
4437 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; |
4438 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
4439 | int gather_scale = 1; | |
4440 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
a70d6342 IR |
4441 | |
4442 | if (loop_vinfo) | |
4443 | { | |
4444 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
4445 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
4446 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4447 | } | |
4448 | else | |
3533e503 | 4449 | vf = 1; |
ebfd146a IR |
4450 | |
4451 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4452 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 4453 | case of SLP. */ |
437f4a00 | 4454 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
4455 | ncopies = 1; |
4456 | else | |
4457 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4458 | ||
4459 | gcc_assert (ncopies >= 1); | |
4460 | ||
4461 | /* FORNOW. This restriction should be relaxed. */ | |
4462 | if (nested_in_vect_loop && ncopies > 1) | |
4463 | { | |
73fbfcad | 4464 | if (dump_enabled_p ()) |
78c60e3d | 4465 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4466 | "multiple types in nested loop.\n"); |
ebfd146a IR |
4467 | return false; |
4468 | } | |
4469 | ||
a70d6342 | 4470 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4471 | return false; |
4472 | ||
8644a673 | 4473 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4474 | return false; |
4475 | ||
4476 | /* Is vectorizable load? */ | |
4477 | if (!is_gimple_assign (stmt)) | |
4478 | return false; | |
4479 | ||
4480 | scalar_dest = gimple_assign_lhs (stmt); | |
4481 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4482 | return false; | |
4483 | ||
4484 | code = gimple_assign_rhs_code (stmt); | |
4485 | if (code != ARRAY_REF | |
38000232 | 4486 | && code != BIT_FIELD_REF |
ebfd146a | 4487 | && code != INDIRECT_REF |
e9dbe7bb IR |
4488 | && code != COMPONENT_REF |
4489 | && code != IMAGPART_EXPR | |
70f34814 | 4490 | && code != REALPART_EXPR |
42373e0b RG |
4491 | && code != MEM_REF |
4492 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
ebfd146a IR |
4493 | return false; |
4494 | ||
4495 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
4496 | return false; | |
4497 | ||
7b7b1813 | 4498 | elem_type = TREE_TYPE (vectype); |
947131ba | 4499 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
4500 | |
4501 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
4502 | (e.g. - data copies). */ | |
947131ba | 4503 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 4504 | { |
73fbfcad | 4505 | if (dump_enabled_p ()) |
78c60e3d | 4506 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4507 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
4508 | return false; |
4509 | } | |
4510 | ||
ebfd146a | 4511 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 4512 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 4513 | { |
0d0293ac | 4514 | grouped_load = true; |
ebfd146a | 4515 | /* FORNOW */ |
aec7ae7d | 4516 | gcc_assert (! nested_in_vect_loop && !STMT_VINFO_GATHER_P (stmt_info)); |
ebfd146a | 4517 | |
e14c1050 | 4518 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
4519 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
4520 | { | |
e14c1050 | 4521 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
4522 | if (vect_load_lanes_supported (vectype, group_size)) |
4523 | load_lanes_p = true; | |
0d0293ac | 4524 | else if (!vect_grouped_load_supported (vectype, group_size)) |
b602d918 RS |
4525 | return false; |
4526 | } | |
ebfd146a IR |
4527 | } |
4528 | ||
a1e53f3f | 4529 | |
aec7ae7d JJ |
4530 | if (STMT_VINFO_GATHER_P (stmt_info)) |
4531 | { | |
4532 | gimple def_stmt; | |
4533 | tree def; | |
4534 | gather_decl = vect_check_gather (stmt, loop_vinfo, &gather_base, | |
4535 | &gather_off, &gather_scale); | |
4536 | gcc_assert (gather_decl); | |
24ee1384 | 4537 | if (!vect_is_simple_use_1 (gather_off, NULL, loop_vinfo, bb_vinfo, |
aec7ae7d JJ |
4538 | &def_stmt, &def, &gather_dt, |
4539 | &gather_off_vectype)) | |
4540 | { | |
73fbfcad | 4541 | if (dump_enabled_p ()) |
78c60e3d | 4542 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4543 | "gather index use not simple.\n"); |
aec7ae7d JJ |
4544 | return false; |
4545 | } | |
4546 | } | |
7d75abc8 | 4547 | else if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
14ac6aa2 | 4548 | ; |
319e6439 RG |
4549 | else |
4550 | { | |
4551 | negative = tree_int_cst_compare (nested_in_vect_loop | |
4552 | ? STMT_VINFO_DR_STEP (stmt_info) | |
4553 | : DR_STEP (dr), | |
4554 | size_zero_node) < 0; | |
4555 | if (negative && ncopies > 1) | |
4556 | { | |
73fbfcad | 4557 | if (dump_enabled_p ()) |
78c60e3d | 4558 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4559 | "multiple types with negative step.\n"); |
319e6439 RG |
4560 | return false; |
4561 | } | |
4562 | ||
4563 | if (negative) | |
4564 | { | |
08940f33 RB |
4565 | if (grouped_load) |
4566 | { | |
4567 | if (dump_enabled_p ()) | |
4568 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 TJ |
4569 | "negative step for group load not supported" |
4570 | "\n"); | |
08940f33 RB |
4571 | return false; |
4572 | } | |
319e6439 RG |
4573 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
4574 | if (alignment_support_scheme != dr_aligned | |
4575 | && alignment_support_scheme != dr_unaligned_supported) | |
4576 | { | |
73fbfcad | 4577 | if (dump_enabled_p ()) |
78c60e3d | 4578 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4579 | "negative step but alignment required.\n"); |
319e6439 RG |
4580 | return false; |
4581 | } | |
4582 | if (!perm_mask_for_reverse (vectype)) | |
4583 | { | |
73fbfcad | 4584 | if (dump_enabled_p ()) |
78c60e3d | 4585 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
4586 | "negative step and reversing not supported." |
4587 | "\n"); | |
319e6439 RG |
4588 | return false; |
4589 | } | |
4590 | } | |
7d75abc8 | 4591 | } |
aec7ae7d | 4592 | |
ebfd146a IR |
4593 | if (!vec_stmt) /* transformation not required. */ |
4594 | { | |
4595 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
92345349 | 4596 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, NULL, NULL, NULL); |
ebfd146a IR |
4597 | return true; |
4598 | } | |
4599 | ||
73fbfcad | 4600 | if (dump_enabled_p ()) |
78c60e3d | 4601 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4602 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a IR |
4603 | |
4604 | /** Transform. **/ | |
4605 | ||
c716e67f XDL |
4606 | ensure_base_align (stmt_info, dr); |
4607 | ||
aec7ae7d JJ |
4608 | if (STMT_VINFO_GATHER_P (stmt_info)) |
4609 | { | |
4610 | tree vec_oprnd0 = NULL_TREE, op; | |
4611 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
4612 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
4613 | tree ptr, mask, var, scale, perm_mask = NULL_TREE, prev_res = NULL_TREE; | |
4614 | edge pe = loop_preheader_edge (loop); | |
4615 | gimple_seq seq; | |
4616 | basic_block new_bb; | |
4617 | enum { NARROW, NONE, WIDEN } modifier; | |
4618 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
4619 | ||
4620 | if (nunits == gather_off_nunits) | |
4621 | modifier = NONE; | |
4622 | else if (nunits == gather_off_nunits / 2) | |
4623 | { | |
4624 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
4625 | modifier = WIDEN; | |
4626 | ||
4627 | for (i = 0; i < gather_off_nunits; ++i) | |
4628 | sel[i] = i | nunits; | |
4629 | ||
3fcc1b55 | 4630 | perm_mask = vect_gen_perm_mask (gather_off_vectype, sel); |
aec7ae7d JJ |
4631 | gcc_assert (perm_mask != NULL_TREE); |
4632 | } | |
4633 | else if (nunits == gather_off_nunits * 2) | |
4634 | { | |
4635 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
4636 | modifier = NARROW; | |
4637 | ||
4638 | for (i = 0; i < nunits; ++i) | |
4639 | sel[i] = i < gather_off_nunits | |
4640 | ? i : i + nunits - gather_off_nunits; | |
4641 | ||
3fcc1b55 | 4642 | perm_mask = vect_gen_perm_mask (vectype, sel); |
aec7ae7d JJ |
4643 | gcc_assert (perm_mask != NULL_TREE); |
4644 | ncopies *= 2; | |
4645 | } | |
4646 | else | |
4647 | gcc_unreachable (); | |
4648 | ||
4649 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
4650 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4651 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4652 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4653 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4654 | scaletype = TREE_VALUE (arglist); | |
4655 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
4656 | && types_compatible_p (srctype, masktype)); | |
4657 | ||
4658 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4659 | ||
4660 | ptr = fold_convert (ptrtype, gather_base); | |
4661 | if (!is_gimple_min_invariant (ptr)) | |
4662 | { | |
4663 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
4664 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
4665 | gcc_assert (!new_bb); | |
4666 | } | |
4667 | ||
4668 | /* Currently we support only unconditional gather loads, | |
4669 | so mask should be all ones. */ | |
4670 | if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
4671 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
4672 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) | |
4673 | { | |
4674 | REAL_VALUE_TYPE r; | |
4675 | long tmp[6]; | |
4676 | for (j = 0; j < 6; ++j) | |
4677 | tmp[j] = -1; | |
4678 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
4679 | mask = build_real (TREE_TYPE (masktype), r); | |
4680 | } | |
4681 | else | |
4682 | gcc_unreachable (); | |
4683 | mask = build_vector_from_val (masktype, mask); | |
4684 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
4685 | ||
4686 | scale = build_int_cst (scaletype, gather_scale); | |
4687 | ||
4688 | prev_stmt_info = NULL; | |
4689 | for (j = 0; j < ncopies; ++j) | |
4690 | { | |
4691 | if (modifier == WIDEN && (j & 1)) | |
4692 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
4693 | perm_mask, stmt, gsi); | |
4694 | else if (j == 0) | |
4695 | op = vec_oprnd0 | |
4696 | = vect_get_vec_def_for_operand (gather_off, stmt, NULL); | |
4697 | else | |
4698 | op = vec_oprnd0 | |
4699 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
4700 | ||
4701 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
4702 | { | |
4703 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
4704 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
4705 | var = vect_get_new_vect_var (idxtype, vect_simple_var, NULL); | |
aec7ae7d JJ |
4706 | var = make_ssa_name (var, NULL); |
4707 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); | |
4708 | new_stmt | |
4709 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, | |
4710 | op, NULL_TREE); | |
4711 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4712 | op = var; | |
4713 | } | |
4714 | ||
4715 | new_stmt | |
4716 | = gimple_build_call (gather_decl, 5, mask, ptr, op, mask, scale); | |
4717 | ||
4718 | if (!useless_type_conversion_p (vectype, rettype)) | |
4719 | { | |
4720 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
4721 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
4722 | var = vect_get_new_vect_var (rettype, vect_simple_var, NULL); | |
aec7ae7d JJ |
4723 | op = make_ssa_name (var, new_stmt); |
4724 | gimple_call_set_lhs (new_stmt, op); | |
4725 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4726 | var = make_ssa_name (vec_dest, NULL); | |
4727 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); | |
4728 | new_stmt | |
4729 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, op, | |
4730 | NULL_TREE); | |
4731 | } | |
4732 | else | |
4733 | { | |
4734 | var = make_ssa_name (vec_dest, new_stmt); | |
4735 | gimple_call_set_lhs (new_stmt, var); | |
4736 | } | |
4737 | ||
4738 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4739 | ||
4740 | if (modifier == NARROW) | |
4741 | { | |
4742 | if ((j & 1) == 0) | |
4743 | { | |
4744 | prev_res = var; | |
4745 | continue; | |
4746 | } | |
4747 | var = permute_vec_elements (prev_res, var, | |
4748 | perm_mask, stmt, gsi); | |
4749 | new_stmt = SSA_NAME_DEF_STMT (var); | |
4750 | } | |
4751 | ||
4752 | if (prev_stmt_info == NULL) | |
4753 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4754 | else | |
4755 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4756 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4757 | } | |
4758 | return true; | |
4759 | } | |
7d75abc8 MM |
4760 | else if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
4761 | { | |
4762 | gimple_stmt_iterator incr_gsi; | |
4763 | bool insert_after; | |
4764 | gimple incr; | |
4765 | tree offvar; | |
7d75abc8 MM |
4766 | tree ivstep; |
4767 | tree running_off; | |
9771b263 | 4768 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 4769 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
4770 | tree stride_base, stride_step, alias_off; |
4771 | ||
4772 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 4773 | |
14ac6aa2 RB |
4774 | stride_base |
4775 | = fold_build_pointer_plus | |
4776 | (unshare_expr (DR_BASE_ADDRESS (dr)), | |
4777 | size_binop (PLUS_EXPR, | |
4778 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (dr))), | |
c3284718 | 4779 | convert_to_ptrofftype (DR_INIT (dr)))); |
14ac6aa2 | 4780 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (dr))); |
7d75abc8 MM |
4781 | |
4782 | /* For a load with loop-invariant (but other than power-of-2) | |
4783 | stride (i.e. not a grouped access) like so: | |
4784 | ||
4785 | for (i = 0; i < n; i += stride) | |
4786 | ... = array[i]; | |
4787 | ||
4788 | we generate a new induction variable and new accesses to | |
4789 | form a new vector (or vectors, depending on ncopies): | |
4790 | ||
4791 | for (j = 0; ; j += VF*stride) | |
4792 | tmp1 = array[j]; | |
4793 | tmp2 = array[j + stride]; | |
4794 | ... | |
4795 | vectemp = {tmp1, tmp2, ...} | |
4796 | */ | |
4797 | ||
4798 | ivstep = stride_step; | |
4799 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
4800 | build_int_cst (TREE_TYPE (ivstep), vf)); | |
4801 | ||
4802 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
4803 | ||
4804 | create_iv (stride_base, ivstep, NULL, | |
4805 | loop, &incr_gsi, insert_after, | |
4806 | &offvar, NULL); | |
4807 | incr = gsi_stmt (incr_gsi); | |
4808 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo, NULL)); | |
4809 | ||
4810 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
4811 | if (stmts) | |
4812 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
4813 | ||
4814 | prev_stmt_info = NULL; | |
4815 | running_off = offvar; | |
14ac6aa2 | 4816 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (dr)), 0); |
7d75abc8 MM |
4817 | for (j = 0; j < ncopies; j++) |
4818 | { | |
4819 | tree vec_inv; | |
4820 | ||
9771b263 | 4821 | vec_alloc (v, nunits); |
7d75abc8 MM |
4822 | for (i = 0; i < nunits; i++) |
4823 | { | |
4824 | tree newref, newoff; | |
4825 | gimple incr; | |
14ac6aa2 RB |
4826 | newref = build2 (MEM_REF, TREE_TYPE (vectype), |
4827 | running_off, alias_off); | |
7d75abc8 MM |
4828 | |
4829 | newref = force_gimple_operand_gsi (gsi, newref, true, | |
4830 | NULL_TREE, true, | |
4831 | GSI_SAME_STMT); | |
4832 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, newref); | |
070ecdfd | 4833 | newoff = copy_ssa_name (running_off, NULL); |
14ac6aa2 RB |
4834 | incr = gimple_build_assign_with_ops (POINTER_PLUS_EXPR, newoff, |
4835 | running_off, stride_step); | |
7d75abc8 MM |
4836 | vect_finish_stmt_generation (stmt, incr, gsi); |
4837 | ||
4838 | running_off = newoff; | |
4839 | } | |
4840 | ||
4841 | vec_inv = build_constructor (vectype, v); | |
4842 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
4843 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
7d75abc8 MM |
4844 | |
4845 | if (j == 0) | |
4846 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4847 | else | |
4848 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4849 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4850 | } | |
4851 | return true; | |
4852 | } | |
aec7ae7d | 4853 | |
0d0293ac | 4854 | if (grouped_load) |
ebfd146a | 4855 | { |
e14c1050 | 4856 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
6aa904c4 | 4857 | if (slp |
01d8bf07 | 4858 | && !SLP_TREE_LOAD_PERMUTATION (slp_node).exists () |
9771b263 DN |
4859 | && first_stmt != SLP_TREE_SCALAR_STMTS (slp_node)[0]) |
4860 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 4861 | |
ebfd146a | 4862 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
4863 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
4864 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
4865 | ??? But we can only do so if there is exactly one | |
4866 | as we have no way to get at the rest. Leave the CSE | |
4867 | opportunity alone. | |
4868 | ??? With the group load eventually participating | |
4869 | in multiple different permutations (having multiple | |
4870 | slp nodes which refer to the same group) the CSE | |
4871 | is even wrong code. See PR56270. */ | |
4872 | && !slp) | |
ebfd146a IR |
4873 | { |
4874 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4875 | return true; | |
4876 | } | |
4877 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 4878 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a IR |
4879 | |
4880 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
4881 | if (slp) | |
4882 | { | |
0d0293ac | 4883 | grouped_load = false; |
ebfd146a | 4884 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
01d8bf07 | 4885 | if (SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
a70d6342 | 4886 | slp_perm = true; |
a64b9c26 | 4887 | group_gap = GROUP_GAP (vinfo_for_stmt (first_stmt)); |
a70d6342 | 4888 | } |
ebfd146a | 4889 | else |
a64b9c26 RB |
4890 | { |
4891 | vec_num = group_size; | |
4892 | group_gap = 0; | |
4893 | } | |
ebfd146a IR |
4894 | } |
4895 | else | |
4896 | { | |
4897 | first_stmt = stmt; | |
4898 | first_dr = dr; | |
4899 | group_size = vec_num = 1; | |
a64b9c26 | 4900 | group_gap = 0; |
ebfd146a IR |
4901 | } |
4902 | ||
720f5239 | 4903 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 4904 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
4905 | /* Targets with load-lane instructions must not require explicit |
4906 | realignment. */ | |
4907 | gcc_assert (!load_lanes_p | |
4908 | || alignment_support_scheme == dr_aligned | |
4909 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
4910 | |
4911 | /* In case the vectorization factor (VF) is bigger than the number | |
4912 | of elements that we can fit in a vectype (nunits), we have to generate | |
4913 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 4914 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 4915 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 4916 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 4917 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
4918 | stmts that use the defs of the current stmt. The example below |
4919 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
4920 | need to create 4 vectorized stmts): | |
ebfd146a IR |
4921 | |
4922 | before vectorization: | |
4923 | RELATED_STMT VEC_STMT | |
4924 | S1: x = memref - - | |
4925 | S2: z = x + 1 - - | |
4926 | ||
4927 | step 1: vectorize stmt S1: | |
4928 | We first create the vector stmt VS1_0, and, as usual, record a | |
4929 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
4930 | Next, we create the vector stmt VS1_1, and record a pointer to | |
4931 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 4932 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
4933 | stmts and pointers: |
4934 | RELATED_STMT VEC_STMT | |
4935 | VS1_0: vx0 = memref0 VS1_1 - | |
4936 | VS1_1: vx1 = memref1 VS1_2 - | |
4937 | VS1_2: vx2 = memref2 VS1_3 - | |
4938 | VS1_3: vx3 = memref3 - - | |
4939 | S1: x = load - VS1_0 | |
4940 | S2: z = x + 1 - - | |
4941 | ||
b8698a0f L |
4942 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
4943 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
4944 | stmt S2. */ |
4945 | ||
0d0293ac | 4946 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
4947 | |
4948 | S1: x2 = &base + 2 | |
4949 | S2: x0 = &base | |
4950 | S3: x1 = &base + 1 | |
4951 | S4: x3 = &base + 3 | |
4952 | ||
b8698a0f | 4953 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
4954 | starting from the access of the first stmt of the chain: |
4955 | ||
4956 | VS1: vx0 = &base | |
4957 | VS2: vx1 = &base + vec_size*1 | |
4958 | VS3: vx3 = &base + vec_size*2 | |
4959 | VS4: vx4 = &base + vec_size*3 | |
4960 | ||
4961 | Then permutation statements are generated: | |
4962 | ||
e2c83630 RH |
4963 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
4964 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
4965 | ... |
4966 | ||
4967 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
4968 | (the order of the data-refs in the output of vect_permute_load_chain | |
4969 | corresponds to the order of scalar stmts in the interleaving chain - see | |
4970 | the documentation of vect_permute_load_chain()). | |
4971 | The generation of permutation stmts and recording them in | |
0d0293ac | 4972 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 4973 | |
b8698a0f | 4974 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
4975 | permutation stmts above are created for every copy. The result vector |
4976 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
4977 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
4978 | |
4979 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
4980 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
4981 | we generate the following code: | |
4982 | p = initial_addr; | |
4983 | indx = 0; | |
4984 | loop { | |
4985 | p = p + indx * vectype_size; | |
4986 | vec_dest = *(p); | |
4987 | indx = indx + 1; | |
4988 | } | |
4989 | ||
4990 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 4991 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
4992 | then generate the following code, in which the data in each iteration is |
4993 | obtained by two vector loads, one from the previous iteration, and one | |
4994 | from the current iteration: | |
4995 | p1 = initial_addr; | |
4996 | msq_init = *(floor(p1)) | |
4997 | p2 = initial_addr + VS - 1; | |
4998 | realignment_token = call target_builtin; | |
4999 | indx = 0; | |
5000 | loop { | |
5001 | p2 = p2 + indx * vectype_size | |
5002 | lsq = *(floor(p2)) | |
5003 | vec_dest = realign_load (msq, lsq, realignment_token) | |
5004 | indx = indx + 1; | |
5005 | msq = lsq; | |
5006 | } */ | |
5007 | ||
5008 | /* If the misalignment remains the same throughout the execution of the | |
5009 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 5010 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
5011 | This can only occur when vectorizing memory accesses in the inner-loop |
5012 | nested within an outer-loop that is being vectorized. */ | |
5013 | ||
d1e4b493 | 5014 | if (nested_in_vect_loop |
211bea38 | 5015 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
ebfd146a IR |
5016 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) |
5017 | { | |
5018 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
5019 | compute_in_loop = true; | |
5020 | } | |
5021 | ||
5022 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
5023 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 5024 | && !compute_in_loop) |
ebfd146a IR |
5025 | { |
5026 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
5027 | alignment_support_scheme, NULL_TREE, | |
5028 | &at_loop); | |
5029 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
5030 | { | |
5031 | phi = SSA_NAME_DEF_STMT (msq); | |
5032 | offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
5033 | } | |
5034 | } | |
5035 | else | |
5036 | at_loop = loop; | |
5037 | ||
a1e53f3f L |
5038 | if (negative) |
5039 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
5040 | ||
272c6793 RS |
5041 | if (load_lanes_p) |
5042 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
5043 | else | |
5044 | aggr_type = vectype; | |
5045 | ||
ebfd146a IR |
5046 | prev_stmt_info = NULL; |
5047 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 5048 | { |
272c6793 | 5049 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 5050 | if (j == 0) |
74bf76ed JJ |
5051 | { |
5052 | bool simd_lane_access_p | |
5053 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
5054 | if (simd_lane_access_p | |
5055 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
5056 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
5057 | && integer_zerop (DR_OFFSET (first_dr)) | |
5058 | && integer_zerop (DR_INIT (first_dr)) | |
5059 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
5060 | get_alias_set (DR_REF (first_dr))) | |
5061 | && (alignment_support_scheme == dr_aligned | |
5062 | || alignment_support_scheme == dr_unaligned_supported)) | |
5063 | { | |
5064 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
5065 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
5066 | (DR_REF (first_dr)), 0); | |
8928eff3 | 5067 | inv_p = false; |
74bf76ed JJ |
5068 | } |
5069 | else | |
5070 | dataref_ptr | |
5071 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
5072 | offset, &dummy, gsi, &ptr_incr, | |
5073 | simd_lane_access_p, &inv_p); | |
5074 | } | |
5075 | else if (dataref_offset) | |
5076 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
5077 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 5078 | else |
272c6793 RS |
5079 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
5080 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 5081 | |
0d0293ac | 5082 | if (grouped_load || slp_perm) |
9771b263 | 5083 | dr_chain.create (vec_num); |
5ce1ee7f | 5084 | |
272c6793 | 5085 | if (load_lanes_p) |
ebfd146a | 5086 | { |
272c6793 RS |
5087 | tree vec_array; |
5088 | ||
5089 | vec_array = create_vector_array (vectype, vec_num); | |
5090 | ||
5091 | /* Emit: | |
5092 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
5093 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
5094 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
5095 | gimple_call_set_lhs (new_stmt, vec_array); | |
5096 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 5097 | |
272c6793 RS |
5098 | /* Extract each vector into an SSA_NAME. */ |
5099 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 5100 | { |
272c6793 RS |
5101 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
5102 | vec_array, i); | |
9771b263 | 5103 | dr_chain.quick_push (new_temp); |
272c6793 RS |
5104 | } |
5105 | ||
5106 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 5107 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
5108 | } |
5109 | else | |
5110 | { | |
5111 | for (i = 0; i < vec_num; i++) | |
5112 | { | |
5113 | if (i > 0) | |
5114 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5115 | stmt, NULL_TREE); | |
5116 | ||
5117 | /* 2. Create the vector-load in the loop. */ | |
5118 | switch (alignment_support_scheme) | |
5119 | { | |
5120 | case dr_aligned: | |
5121 | case dr_unaligned_supported: | |
be1ac4ec | 5122 | { |
644ffefd MJ |
5123 | unsigned int align, misalign; |
5124 | ||
272c6793 RS |
5125 | data_ref |
5126 | = build2 (MEM_REF, vectype, dataref_ptr, | |
74bf76ed JJ |
5127 | dataref_offset |
5128 | ? dataref_offset | |
5129 | : build_int_cst (reference_alias_ptr_type | |
5130 | (DR_REF (first_dr)), 0)); | |
644ffefd | 5131 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 RS |
5132 | if (alignment_support_scheme == dr_aligned) |
5133 | { | |
5134 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 5135 | misalign = 0; |
272c6793 RS |
5136 | } |
5137 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
5138 | { | |
5139 | TREE_TYPE (data_ref) | |
5140 | = build_aligned_type (TREE_TYPE (data_ref), | |
5141 | TYPE_ALIGN (elem_type)); | |
644ffefd MJ |
5142 | align = TYPE_ALIGN_UNIT (elem_type); |
5143 | misalign = 0; | |
272c6793 RS |
5144 | } |
5145 | else | |
5146 | { | |
5147 | TREE_TYPE (data_ref) | |
5148 | = build_aligned_type (TREE_TYPE (data_ref), | |
5149 | TYPE_ALIGN (elem_type)); | |
644ffefd | 5150 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 5151 | } |
74bf76ed JJ |
5152 | if (dataref_offset == NULL_TREE) |
5153 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), | |
5154 | align, misalign); | |
272c6793 | 5155 | break; |
be1ac4ec | 5156 | } |
272c6793 | 5157 | case dr_explicit_realign: |
267d3070 | 5158 | { |
272c6793 RS |
5159 | tree ptr, bump; |
5160 | tree vs_minus_1; | |
5161 | ||
5162 | vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
5163 | ||
5164 | if (compute_in_loop) | |
5165 | msq = vect_setup_realignment (first_stmt, gsi, | |
5166 | &realignment_token, | |
5167 | dr_explicit_realign, | |
5168 | dataref_ptr, NULL); | |
5169 | ||
070ecdfd | 5170 | ptr = copy_ssa_name (dataref_ptr, NULL); |
272c6793 | 5171 | new_stmt = gimple_build_assign_with_ops |
070ecdfd | 5172 | (BIT_AND_EXPR, ptr, dataref_ptr, |
272c6793 RS |
5173 | build_int_cst |
5174 | (TREE_TYPE (dataref_ptr), | |
5175 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
5176 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
5177 | data_ref | |
5178 | = build2 (MEM_REF, vectype, ptr, | |
5179 | build_int_cst (reference_alias_ptr_type | |
5180 | (DR_REF (first_dr)), 0)); | |
5181 | vec_dest = vect_create_destination_var (scalar_dest, | |
5182 | vectype); | |
5183 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
5184 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5185 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5186 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
5187 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
5188 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5189 | msq = new_temp; | |
5190 | ||
5191 | bump = size_binop (MULT_EXPR, vs_minus_1, | |
7b7b1813 | 5192 | TYPE_SIZE_UNIT (elem_type)); |
272c6793 RS |
5193 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
5194 | new_stmt = gimple_build_assign_with_ops | |
5195 | (BIT_AND_EXPR, NULL_TREE, ptr, | |
5196 | build_int_cst | |
5197 | (TREE_TYPE (ptr), | |
5198 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
070ecdfd | 5199 | ptr = copy_ssa_name (dataref_ptr, new_stmt); |
272c6793 RS |
5200 | gimple_assign_set_lhs (new_stmt, ptr); |
5201 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5202 | data_ref | |
5203 | = build2 (MEM_REF, vectype, ptr, | |
5204 | build_int_cst (reference_alias_ptr_type | |
5205 | (DR_REF (first_dr)), 0)); | |
5206 | break; | |
267d3070 | 5207 | } |
272c6793 | 5208 | case dr_explicit_realign_optimized: |
070ecdfd | 5209 | new_temp = copy_ssa_name (dataref_ptr, NULL); |
272c6793 | 5210 | new_stmt = gimple_build_assign_with_ops |
070ecdfd | 5211 | (BIT_AND_EXPR, new_temp, dataref_ptr, |
272c6793 RS |
5212 | build_int_cst |
5213 | (TREE_TYPE (dataref_ptr), | |
5214 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
5215 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
5216 | data_ref | |
5217 | = build2 (MEM_REF, vectype, new_temp, | |
5218 | build_int_cst (reference_alias_ptr_type | |
5219 | (DR_REF (first_dr)), 0)); | |
5220 | break; | |
5221 | default: | |
5222 | gcc_unreachable (); | |
5223 | } | |
ebfd146a | 5224 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 5225 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
5226 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5227 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5228 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5229 | ||
272c6793 RS |
5230 | /* 3. Handle explicit realignment if necessary/supported. |
5231 | Create in loop: | |
5232 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
5233 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
5234 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 5235 | { |
272c6793 RS |
5236 | lsq = gimple_assign_lhs (new_stmt); |
5237 | if (!realignment_token) | |
5238 | realignment_token = dataref_ptr; | |
5239 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5240 | new_stmt | |
73804b12 RG |
5241 | = gimple_build_assign_with_ops (REALIGN_LOAD_EXPR, |
5242 | vec_dest, msq, lsq, | |
5243 | realignment_token); | |
272c6793 RS |
5244 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5245 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5246 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5247 | ||
5248 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
5249 | { | |
5250 | gcc_assert (phi); | |
5251 | if (i == vec_num - 1 && j == ncopies - 1) | |
5252 | add_phi_arg (phi, lsq, | |
5253 | loop_latch_edge (containing_loop), | |
9e227d60 | 5254 | UNKNOWN_LOCATION); |
272c6793 RS |
5255 | msq = lsq; |
5256 | } | |
ebfd146a | 5257 | } |
ebfd146a | 5258 | |
59fd17e3 RB |
5259 | /* 4. Handle invariant-load. */ |
5260 | if (inv_p && !bb_vinfo) | |
5261 | { | |
5262 | gimple_stmt_iterator gsi2 = *gsi; | |
5263 | gcc_assert (!grouped_load); | |
5264 | gsi_next (&gsi2); | |
5265 | new_temp = vect_init_vector (stmt, scalar_dest, | |
5266 | vectype, &gsi2); | |
5267 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
5268 | } | |
5269 | ||
272c6793 RS |
5270 | if (negative) |
5271 | { | |
aec7ae7d JJ |
5272 | tree perm_mask = perm_mask_for_reverse (vectype); |
5273 | new_temp = permute_vec_elements (new_temp, new_temp, | |
5274 | perm_mask, stmt, gsi); | |
ebfd146a IR |
5275 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
5276 | } | |
267d3070 | 5277 | |
272c6793 | 5278 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
5279 | vect_transform_grouped_load (). */ |
5280 | if (grouped_load || slp_perm) | |
9771b263 | 5281 | dr_chain.quick_push (new_temp); |
267d3070 | 5282 | |
272c6793 RS |
5283 | /* Store vector loads in the corresponding SLP_NODE. */ |
5284 | if (slp && !slp_perm) | |
9771b263 | 5285 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
272c6793 | 5286 | } |
a64b9c26 RB |
5287 | /* Bump the vector pointer to account for a gap. */ |
5288 | if (slp && group_gap != 0) | |
5289 | { | |
5290 | tree bump = size_binop (MULT_EXPR, | |
5291 | TYPE_SIZE_UNIT (elem_type), | |
5292 | size_int (group_gap)); | |
5293 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5294 | stmt, bump); | |
5295 | } | |
ebfd146a IR |
5296 | } |
5297 | ||
5298 | if (slp && !slp_perm) | |
5299 | continue; | |
5300 | ||
5301 | if (slp_perm) | |
5302 | { | |
01d8bf07 | 5303 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
ebfd146a IR |
5304 | slp_node_instance, false)) |
5305 | { | |
9771b263 | 5306 | dr_chain.release (); |
ebfd146a IR |
5307 | return false; |
5308 | } | |
5309 | } | |
5310 | else | |
5311 | { | |
0d0293ac | 5312 | if (grouped_load) |
ebfd146a | 5313 | { |
272c6793 | 5314 | if (!load_lanes_p) |
0d0293ac | 5315 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 5316 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
5317 | } |
5318 | else | |
5319 | { | |
5320 | if (j == 0) | |
5321 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5322 | else | |
5323 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5324 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5325 | } | |
5326 | } | |
9771b263 | 5327 | dr_chain.release (); |
ebfd146a IR |
5328 | } |
5329 | ||
ebfd146a IR |
5330 | return true; |
5331 | } | |
5332 | ||
5333 | /* Function vect_is_simple_cond. | |
b8698a0f | 5334 | |
ebfd146a IR |
5335 | Input: |
5336 | LOOP - the loop that is being vectorized. | |
5337 | COND - Condition that is checked for simple use. | |
5338 | ||
e9e1d143 RG |
5339 | Output: |
5340 | *COMP_VECTYPE - the vector type for the comparison. | |
5341 | ||
ebfd146a IR |
5342 | Returns whether a COND can be vectorized. Checks whether |
5343 | condition operands are supportable using vec_is_simple_use. */ | |
5344 | ||
87aab9b2 | 5345 | static bool |
24ee1384 IR |
5346 | vect_is_simple_cond (tree cond, gimple stmt, loop_vec_info loop_vinfo, |
5347 | bb_vec_info bb_vinfo, tree *comp_vectype) | |
ebfd146a IR |
5348 | { |
5349 | tree lhs, rhs; | |
5350 | tree def; | |
5351 | enum vect_def_type dt; | |
e9e1d143 | 5352 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a IR |
5353 | |
5354 | if (!COMPARISON_CLASS_P (cond)) | |
5355 | return false; | |
5356 | ||
5357 | lhs = TREE_OPERAND (cond, 0); | |
5358 | rhs = TREE_OPERAND (cond, 1); | |
5359 | ||
5360 | if (TREE_CODE (lhs) == SSA_NAME) | |
5361 | { | |
5362 | gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); | |
24ee1384 IR |
5363 | if (!vect_is_simple_use_1 (lhs, stmt, loop_vinfo, bb_vinfo, |
5364 | &lhs_def_stmt, &def, &dt, &vectype1)) | |
ebfd146a IR |
5365 | return false; |
5366 | } | |
5367 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
5368 | && TREE_CODE (lhs) != FIXED_CST) | |
5369 | return false; | |
5370 | ||
5371 | if (TREE_CODE (rhs) == SSA_NAME) | |
5372 | { | |
5373 | gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); | |
24ee1384 IR |
5374 | if (!vect_is_simple_use_1 (rhs, stmt, loop_vinfo, bb_vinfo, |
5375 | &rhs_def_stmt, &def, &dt, &vectype2)) | |
ebfd146a IR |
5376 | return false; |
5377 | } | |
f7e531cf | 5378 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST |
ebfd146a IR |
5379 | && TREE_CODE (rhs) != FIXED_CST) |
5380 | return false; | |
5381 | ||
e9e1d143 | 5382 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
5383 | return true; |
5384 | } | |
5385 | ||
5386 | /* vectorizable_condition. | |
5387 | ||
b8698a0f L |
5388 | Check if STMT is conditional modify expression that can be vectorized. |
5389 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
5390 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
5391 | at GSI. |
5392 | ||
5393 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
5394 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
5395 | else caluse if it is 2). | |
ebfd146a IR |
5396 | |
5397 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5398 | ||
4bbe8262 | 5399 | bool |
ebfd146a | 5400 | vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi, |
f7e531cf IR |
5401 | gimple *vec_stmt, tree reduc_def, int reduc_index, |
5402 | slp_tree slp_node) | |
ebfd146a IR |
5403 | { |
5404 | tree scalar_dest = NULL_TREE; | |
5405 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
5406 | tree cond_expr, then_clause, else_clause; |
5407 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5408 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
df11cc78 | 5409 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
5410 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
5411 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
ebfd146a IR |
5412 | tree vec_compare, vec_cond_expr; |
5413 | tree new_temp; | |
5414 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
ebfd146a | 5415 | tree def; |
a855b1b1 | 5416 | enum vect_def_type dt, dts[4]; |
ebfd146a | 5417 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
f7e531cf | 5418 | int ncopies; |
ebfd146a | 5419 | enum tree_code code; |
a855b1b1 | 5420 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
5421 | int i, j; |
5422 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
5423 | vec<tree> vec_oprnds0 = vNULL; |
5424 | vec<tree> vec_oprnds1 = vNULL; | |
5425 | vec<tree> vec_oprnds2 = vNULL; | |
5426 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 5427 | tree vec_cmp_type; |
b8698a0f | 5428 | |
f7e531cf IR |
5429 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
5430 | ncopies = 1; | |
5431 | else | |
5432 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
437f4a00 | 5433 | |
ebfd146a | 5434 | gcc_assert (ncopies >= 1); |
a855b1b1 | 5435 | if (reduc_index && ncopies > 1) |
ebfd146a IR |
5436 | return false; /* FORNOW */ |
5437 | ||
f7e531cf IR |
5438 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
5439 | return false; | |
5440 | ||
5441 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
ebfd146a IR |
5442 | return false; |
5443 | ||
4bbe8262 IR |
5444 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5445 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
5446 | && reduc_def)) | |
ebfd146a IR |
5447 | return false; |
5448 | ||
ebfd146a | 5449 | /* FORNOW: not yet supported. */ |
b8698a0f | 5450 | if (STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 5451 | { |
73fbfcad | 5452 | if (dump_enabled_p ()) |
78c60e3d | 5453 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5454 | "value used after loop.\n"); |
ebfd146a IR |
5455 | return false; |
5456 | } | |
5457 | ||
5458 | /* Is vectorizable conditional operation? */ | |
5459 | if (!is_gimple_assign (stmt)) | |
5460 | return false; | |
5461 | ||
5462 | code = gimple_assign_rhs_code (stmt); | |
5463 | ||
5464 | if (code != COND_EXPR) | |
5465 | return false; | |
5466 | ||
4e71066d RG |
5467 | cond_expr = gimple_assign_rhs1 (stmt); |
5468 | then_clause = gimple_assign_rhs2 (stmt); | |
5469 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 5470 | |
24ee1384 IR |
5471 | if (!vect_is_simple_cond (cond_expr, stmt, loop_vinfo, bb_vinfo, |
5472 | &comp_vectype) | |
e9e1d143 | 5473 | || !comp_vectype) |
ebfd146a IR |
5474 | return false; |
5475 | ||
5476 | if (TREE_CODE (then_clause) == SSA_NAME) | |
5477 | { | |
5478 | gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause); | |
24ee1384 | 5479 | if (!vect_is_simple_use (then_clause, stmt, loop_vinfo, bb_vinfo, |
ebfd146a IR |
5480 | &then_def_stmt, &def, &dt)) |
5481 | return false; | |
5482 | } | |
b8698a0f | 5483 | else if (TREE_CODE (then_clause) != INTEGER_CST |
ebfd146a IR |
5484 | && TREE_CODE (then_clause) != REAL_CST |
5485 | && TREE_CODE (then_clause) != FIXED_CST) | |
5486 | return false; | |
5487 | ||
5488 | if (TREE_CODE (else_clause) == SSA_NAME) | |
5489 | { | |
5490 | gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause); | |
24ee1384 | 5491 | if (!vect_is_simple_use (else_clause, stmt, loop_vinfo, bb_vinfo, |
ebfd146a IR |
5492 | &else_def_stmt, &def, &dt)) |
5493 | return false; | |
5494 | } | |
b8698a0f | 5495 | else if (TREE_CODE (else_clause) != INTEGER_CST |
ebfd146a IR |
5496 | && TREE_CODE (else_clause) != REAL_CST |
5497 | && TREE_CODE (else_clause) != FIXED_CST) | |
5498 | return false; | |
5499 | ||
74946978 MP |
5500 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (vectype))); |
5501 | /* The result of a vector comparison should be signed type. */ | |
5502 | tree cmp_type = build_nonstandard_integer_type (prec, 0); | |
5503 | vec_cmp_type = get_same_sized_vectype (cmp_type, vectype); | |
5504 | if (vec_cmp_type == NULL_TREE) | |
5505 | return false; | |
784fb9b3 | 5506 | |
b8698a0f | 5507 | if (!vec_stmt) |
ebfd146a IR |
5508 | { |
5509 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
e9e1d143 | 5510 | return expand_vec_cond_expr_p (vectype, comp_vectype); |
ebfd146a IR |
5511 | } |
5512 | ||
f7e531cf IR |
5513 | /* Transform. */ |
5514 | ||
5515 | if (!slp_node) | |
5516 | { | |
9771b263 DN |
5517 | vec_oprnds0.create (1); |
5518 | vec_oprnds1.create (1); | |
5519 | vec_oprnds2.create (1); | |
5520 | vec_oprnds3.create (1); | |
f7e531cf | 5521 | } |
ebfd146a IR |
5522 | |
5523 | /* Handle def. */ | |
5524 | scalar_dest = gimple_assign_lhs (stmt); | |
5525 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5526 | ||
5527 | /* Handle cond expr. */ | |
a855b1b1 MM |
5528 | for (j = 0; j < ncopies; j++) |
5529 | { | |
f7e531cf | 5530 | gimple new_stmt = NULL; |
a855b1b1 MM |
5531 | if (j == 0) |
5532 | { | |
f7e531cf IR |
5533 | if (slp_node) |
5534 | { | |
07687835 TS |
5535 | stack_vec<tree, 4> ops; |
5536 | stack_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 5537 | |
9771b263 DN |
5538 | ops.safe_push (TREE_OPERAND (cond_expr, 0)); |
5539 | ops.safe_push (TREE_OPERAND (cond_expr, 1)); | |
5540 | ops.safe_push (then_clause); | |
5541 | ops.safe_push (else_clause); | |
f7e531cf | 5542 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); |
37b5ec8f JJ |
5543 | vec_oprnds3 = vec_defs.pop (); |
5544 | vec_oprnds2 = vec_defs.pop (); | |
5545 | vec_oprnds1 = vec_defs.pop (); | |
5546 | vec_oprnds0 = vec_defs.pop (); | |
f7e531cf | 5547 | |
9771b263 DN |
5548 | ops.release (); |
5549 | vec_defs.release (); | |
f7e531cf IR |
5550 | } |
5551 | else | |
5552 | { | |
5553 | gimple gtemp; | |
5554 | vec_cond_lhs = | |
a855b1b1 MM |
5555 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), |
5556 | stmt, NULL); | |
24ee1384 IR |
5557 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), stmt, |
5558 | loop_vinfo, NULL, >emp, &def, &dts[0]); | |
f7e531cf IR |
5559 | |
5560 | vec_cond_rhs = | |
5561 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
5562 | stmt, NULL); | |
24ee1384 IR |
5563 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), stmt, |
5564 | loop_vinfo, NULL, >emp, &def, &dts[1]); | |
f7e531cf IR |
5565 | if (reduc_index == 1) |
5566 | vec_then_clause = reduc_def; | |
5567 | else | |
5568 | { | |
5569 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
5570 | stmt, NULL); | |
24ee1384 | 5571 | vect_is_simple_use (then_clause, stmt, loop_vinfo, |
f7e531cf IR |
5572 | NULL, >emp, &def, &dts[2]); |
5573 | } | |
5574 | if (reduc_index == 2) | |
5575 | vec_else_clause = reduc_def; | |
5576 | else | |
5577 | { | |
5578 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
a855b1b1 | 5579 | stmt, NULL); |
24ee1384 | 5580 | vect_is_simple_use (else_clause, stmt, loop_vinfo, |
a855b1b1 | 5581 | NULL, >emp, &def, &dts[3]); |
f7e531cf | 5582 | } |
a855b1b1 MM |
5583 | } |
5584 | } | |
5585 | else | |
5586 | { | |
f7e531cf | 5587 | vec_cond_lhs = vect_get_vec_def_for_stmt_copy (dts[0], |
9771b263 | 5588 | vec_oprnds0.pop ()); |
f7e531cf | 5589 | vec_cond_rhs = vect_get_vec_def_for_stmt_copy (dts[1], |
9771b263 | 5590 | vec_oprnds1.pop ()); |
a855b1b1 | 5591 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 5592 | vec_oprnds2.pop ()); |
a855b1b1 | 5593 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 5594 | vec_oprnds3.pop ()); |
f7e531cf IR |
5595 | } |
5596 | ||
5597 | if (!slp_node) | |
5598 | { | |
9771b263 DN |
5599 | vec_oprnds0.quick_push (vec_cond_lhs); |
5600 | vec_oprnds1.quick_push (vec_cond_rhs); | |
5601 | vec_oprnds2.quick_push (vec_then_clause); | |
5602 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
5603 | } |
5604 | ||
9dc3f7de | 5605 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 5606 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 5607 | { |
9771b263 DN |
5608 | vec_cond_rhs = vec_oprnds1[i]; |
5609 | vec_then_clause = vec_oprnds2[i]; | |
5610 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 5611 | |
784fb9b3 JJ |
5612 | vec_compare = build2 (TREE_CODE (cond_expr), vec_cmp_type, |
5613 | vec_cond_lhs, vec_cond_rhs); | |
f7e531cf IR |
5614 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, |
5615 | vec_compare, vec_then_clause, vec_else_clause); | |
a855b1b1 | 5616 | |
f7e531cf IR |
5617 | new_stmt = gimple_build_assign (vec_dest, vec_cond_expr); |
5618 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5619 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5620 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5621 | if (slp_node) | |
9771b263 | 5622 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
5623 | } |
5624 | ||
5625 | if (slp_node) | |
5626 | continue; | |
5627 | ||
5628 | if (j == 0) | |
5629 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5630 | else | |
5631 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5632 | ||
5633 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 5634 | } |
b8698a0f | 5635 | |
9771b263 DN |
5636 | vec_oprnds0.release (); |
5637 | vec_oprnds1.release (); | |
5638 | vec_oprnds2.release (); | |
5639 | vec_oprnds3.release (); | |
f7e531cf | 5640 | |
ebfd146a IR |
5641 | return true; |
5642 | } | |
5643 | ||
5644 | ||
8644a673 | 5645 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
5646 | |
5647 | bool | |
a70d6342 | 5648 | vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 5649 | { |
8644a673 | 5650 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 5651 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 5652 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 5653 | bool ok; |
a70d6342 | 5654 | tree scalar_type, vectype; |
363477c0 JJ |
5655 | gimple pattern_stmt; |
5656 | gimple_seq pattern_def_seq; | |
ebfd146a | 5657 | |
73fbfcad | 5658 | if (dump_enabled_p ()) |
ebfd146a | 5659 | { |
78c60e3d SS |
5660 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
5661 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 5662 | dump_printf (MSG_NOTE, "\n"); |
8644a673 | 5663 | } |
ebfd146a | 5664 | |
1825a1f3 | 5665 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 5666 | { |
73fbfcad | 5667 | if (dump_enabled_p ()) |
78c60e3d | 5668 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5669 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
5670 | |
5671 | return false; | |
5672 | } | |
b8698a0f L |
5673 | |
5674 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
5675 | to include: |
5676 | - the COND_EXPR which is the loop exit condition | |
5677 | - any LABEL_EXPRs in the loop | |
b8698a0f | 5678 | - computations that are used only for array indexing or loop control. |
8644a673 | 5679 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 5680 | instance, therefore, all the statements are relevant. |
ebfd146a | 5681 | |
d092494c | 5682 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 5683 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
5684 | statements. In basic blocks we are called from some SLP instance |
5685 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
5686 | already will be part of SLP instance. */ | |
83197f37 IR |
5687 | |
5688 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 5689 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 5690 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 5691 | { |
9d5e7640 | 5692 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 5693 | && pattern_stmt |
9d5e7640 IR |
5694 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
5695 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
5696 | { | |
83197f37 | 5697 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
5698 | stmt = pattern_stmt; |
5699 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 5700 | if (dump_enabled_p ()) |
9d5e7640 | 5701 | { |
78c60e3d SS |
5702 | dump_printf_loc (MSG_NOTE, vect_location, |
5703 | "==> examining pattern statement: "); | |
5704 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 5705 | dump_printf (MSG_NOTE, "\n"); |
9d5e7640 IR |
5706 | } |
5707 | } | |
5708 | else | |
5709 | { | |
73fbfcad | 5710 | if (dump_enabled_p ()) |
e645e942 | 5711 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 5712 | |
9d5e7640 IR |
5713 | return true; |
5714 | } | |
8644a673 | 5715 | } |
83197f37 | 5716 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 5717 | && node == NULL |
83197f37 IR |
5718 | && pattern_stmt |
5719 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
5720 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
5721 | { | |
5722 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 5723 | if (dump_enabled_p ()) |
83197f37 | 5724 | { |
78c60e3d SS |
5725 | dump_printf_loc (MSG_NOTE, vect_location, |
5726 | "==> examining pattern statement: "); | |
5727 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 5728 | dump_printf (MSG_NOTE, "\n"); |
83197f37 IR |
5729 | } |
5730 | ||
5731 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
5732 | return false; | |
5733 | } | |
ebfd146a | 5734 | |
1107f3ae | 5735 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 5736 | && node == NULL |
363477c0 | 5737 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 5738 | { |
363477c0 | 5739 | gimple_stmt_iterator si; |
1107f3ae | 5740 | |
363477c0 JJ |
5741 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
5742 | { | |
5743 | gimple pattern_def_stmt = gsi_stmt (si); | |
5744 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) | |
5745 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
5746 | { | |
5747 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 5748 | if (dump_enabled_p ()) |
363477c0 | 5749 | { |
78c60e3d SS |
5750 | dump_printf_loc (MSG_NOTE, vect_location, |
5751 | "==> examining pattern def statement: "); | |
5752 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
e645e942 | 5753 | dump_printf (MSG_NOTE, "\n"); |
363477c0 | 5754 | } |
1107f3ae | 5755 | |
363477c0 JJ |
5756 | if (!vect_analyze_stmt (pattern_def_stmt, |
5757 | need_to_vectorize, node)) | |
5758 | return false; | |
5759 | } | |
5760 | } | |
5761 | } | |
1107f3ae | 5762 | |
8644a673 IR |
5763 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
5764 | { | |
5765 | case vect_internal_def: | |
5766 | break; | |
ebfd146a | 5767 | |
8644a673 | 5768 | case vect_reduction_def: |
7c5222ff | 5769 | case vect_nested_cycle: |
a70d6342 | 5770 | gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer |
8644a673 | 5771 | || relevance == vect_used_in_outer_by_reduction |
a70d6342 | 5772 | || relevance == vect_unused_in_scope)); |
8644a673 IR |
5773 | break; |
5774 | ||
5775 | case vect_induction_def: | |
5776 | case vect_constant_def: | |
5777 | case vect_external_def: | |
5778 | case vect_unknown_def_type: | |
5779 | default: | |
5780 | gcc_unreachable (); | |
5781 | } | |
ebfd146a | 5782 | |
a70d6342 IR |
5783 | if (bb_vinfo) |
5784 | { | |
5785 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
5786 | ||
b690cc0f | 5787 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
73fbfcad | 5788 | if (dump_enabled_p ()) |
a70d6342 | 5789 | { |
78c60e3d SS |
5790 | dump_printf_loc (MSG_NOTE, vect_location, |
5791 | "get vectype for scalar type: "); | |
5792 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
e645e942 | 5793 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
5794 | } |
5795 | ||
5796 | vectype = get_vectype_for_scalar_type (scalar_type); | |
5797 | if (!vectype) | |
5798 | { | |
73fbfcad | 5799 | if (dump_enabled_p ()) |
a70d6342 | 5800 | { |
78c60e3d SS |
5801 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5802 | "not SLPed: unsupported data-type "); | |
5803 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
5804 | scalar_type); | |
e645e942 | 5805 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
a70d6342 IR |
5806 | } |
5807 | return false; | |
5808 | } | |
5809 | ||
73fbfcad | 5810 | if (dump_enabled_p ()) |
a70d6342 | 5811 | { |
78c60e3d SS |
5812 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); |
5813 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
e645e942 | 5814 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
5815 | } |
5816 | ||
5817 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
5818 | } | |
5819 | ||
8644a673 | 5820 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 5821 | { |
8644a673 IR |
5822 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
5823 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info)); | |
5824 | *need_to_vectorize = true; | |
ebfd146a IR |
5825 | } |
5826 | ||
8644a673 | 5827 | ok = true; |
b8698a0f | 5828 | if (!bb_vinfo |
a70d6342 IR |
5829 | && (STMT_VINFO_RELEVANT_P (stmt_info) |
5830 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
4a00c761 | 5831 | ok = (vectorizable_conversion (stmt, NULL, NULL, NULL) |
9dc3f7de | 5832 | || vectorizable_shift (stmt, NULL, NULL, NULL) |
8644a673 IR |
5833 | || vectorizable_operation (stmt, NULL, NULL, NULL) |
5834 | || vectorizable_assignment (stmt, NULL, NULL, NULL) | |
5835 | || vectorizable_load (stmt, NULL, NULL, NULL, NULL) | |
190c2236 | 5836 | || vectorizable_call (stmt, NULL, NULL, NULL) |
8644a673 | 5837 | || vectorizable_store (stmt, NULL, NULL, NULL) |
b5aeb3bb | 5838 | || vectorizable_reduction (stmt, NULL, NULL, NULL) |
f7e531cf | 5839 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, NULL)); |
a70d6342 IR |
5840 | else |
5841 | { | |
5842 | if (bb_vinfo) | |
4a00c761 JJ |
5843 | ok = (vectorizable_conversion (stmt, NULL, NULL, node) |
5844 | || vectorizable_shift (stmt, NULL, NULL, node) | |
9dc3f7de | 5845 | || vectorizable_operation (stmt, NULL, NULL, node) |
a70d6342 IR |
5846 | || vectorizable_assignment (stmt, NULL, NULL, node) |
5847 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
190c2236 | 5848 | || vectorizable_call (stmt, NULL, NULL, node) |
f7e531cf IR |
5849 | || vectorizable_store (stmt, NULL, NULL, node) |
5850 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node)); | |
b8698a0f | 5851 | } |
8644a673 IR |
5852 | |
5853 | if (!ok) | |
ebfd146a | 5854 | { |
73fbfcad | 5855 | if (dump_enabled_p ()) |
8644a673 | 5856 | { |
78c60e3d SS |
5857 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5858 | "not vectorized: relevant stmt not "); | |
5859 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
5860 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 5861 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
8644a673 | 5862 | } |
b8698a0f | 5863 | |
ebfd146a IR |
5864 | return false; |
5865 | } | |
5866 | ||
a70d6342 IR |
5867 | if (bb_vinfo) |
5868 | return true; | |
5869 | ||
8644a673 IR |
5870 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
5871 | need extra handling, except for vectorizable reductions. */ | |
5872 | if (STMT_VINFO_LIVE_P (stmt_info) | |
5873 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
5874 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 5875 | |
8644a673 | 5876 | if (!ok) |
ebfd146a | 5877 | { |
73fbfcad | 5878 | if (dump_enabled_p ()) |
8644a673 | 5879 | { |
78c60e3d SS |
5880 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5881 | "not vectorized: live stmt not "); | |
5882 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
5883 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 5884 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
8644a673 | 5885 | } |
b8698a0f | 5886 | |
8644a673 | 5887 | return false; |
ebfd146a IR |
5888 | } |
5889 | ||
ebfd146a IR |
5890 | return true; |
5891 | } | |
5892 | ||
5893 | ||
5894 | /* Function vect_transform_stmt. | |
5895 | ||
5896 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
5897 | ||
5898 | bool | |
5899 | vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi, | |
0d0293ac | 5900 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
5901 | slp_instance slp_node_instance) |
5902 | { | |
5903 | bool is_store = false; | |
5904 | gimple vec_stmt = NULL; | |
5905 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a | 5906 | bool done; |
ebfd146a IR |
5907 | |
5908 | switch (STMT_VINFO_TYPE (stmt_info)) | |
5909 | { | |
5910 | case type_demotion_vec_info_type: | |
ebfd146a | 5911 | case type_promotion_vec_info_type: |
ebfd146a IR |
5912 | case type_conversion_vec_info_type: |
5913 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
5914 | gcc_assert (done); | |
5915 | break; | |
5916 | ||
5917 | case induc_vec_info_type: | |
5918 | gcc_assert (!slp_node); | |
5919 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
5920 | gcc_assert (done); | |
5921 | break; | |
5922 | ||
9dc3f7de IR |
5923 | case shift_vec_info_type: |
5924 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
5925 | gcc_assert (done); | |
5926 | break; | |
5927 | ||
ebfd146a IR |
5928 | case op_vec_info_type: |
5929 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
5930 | gcc_assert (done); | |
5931 | break; | |
5932 | ||
5933 | case assignment_vec_info_type: | |
5934 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
5935 | gcc_assert (done); | |
5936 | break; | |
5937 | ||
5938 | case load_vec_info_type: | |
b8698a0f | 5939 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
5940 | slp_node_instance); |
5941 | gcc_assert (done); | |
5942 | break; | |
5943 | ||
5944 | case store_vec_info_type: | |
5945 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
5946 | gcc_assert (done); | |
0d0293ac | 5947 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
5948 | { |
5949 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 5950 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
5951 | one are skipped, and there vec_stmt_info shouldn't be freed |
5952 | meanwhile. */ | |
0d0293ac | 5953 | *grouped_store = true; |
ebfd146a IR |
5954 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
5955 | is_store = true; | |
5956 | } | |
5957 | else | |
5958 | is_store = true; | |
5959 | break; | |
5960 | ||
5961 | case condition_vec_info_type: | |
f7e531cf | 5962 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
5963 | gcc_assert (done); |
5964 | break; | |
5965 | ||
5966 | case call_vec_info_type: | |
190c2236 | 5967 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 5968 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
5969 | break; |
5970 | ||
5971 | case reduc_vec_info_type: | |
b5aeb3bb | 5972 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
5973 | gcc_assert (done); |
5974 | break; | |
5975 | ||
5976 | default: | |
5977 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
5978 | { | |
73fbfcad | 5979 | if (dump_enabled_p ()) |
78c60e3d | 5980 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5981 | "stmt not supported.\n"); |
ebfd146a IR |
5982 | gcc_unreachable (); |
5983 | } | |
5984 | } | |
5985 | ||
5986 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that | |
5987 | is being vectorized, but outside the immediately enclosing loop. */ | |
5988 | if (vec_stmt | |
a70d6342 IR |
5989 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
5990 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
5991 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
5992 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
5993 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 5994 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 5995 | vect_used_in_outer_by_reduction)) |
ebfd146a | 5996 | { |
a70d6342 IR |
5997 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
5998 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
5999 | imm_use_iterator imm_iter; |
6000 | use_operand_p use_p; | |
6001 | tree scalar_dest; | |
6002 | gimple exit_phi; | |
6003 | ||
73fbfcad | 6004 | if (dump_enabled_p ()) |
78c60e3d | 6005 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6006 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
6007 | |
6008 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
6009 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
6010 | STMT). */ | |
6011 | if (gimple_code (stmt) == GIMPLE_PHI) | |
6012 | scalar_dest = PHI_RESULT (stmt); | |
6013 | else | |
6014 | scalar_dest = gimple_assign_lhs (stmt); | |
6015 | ||
6016 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
6017 | { | |
6018 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
6019 | { | |
6020 | exit_phi = USE_STMT (use_p); | |
6021 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
6022 | } | |
6023 | } | |
6024 | } | |
6025 | ||
6026 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
6027 | being vectorized. */ | |
6028 | if (STMT_VINFO_LIVE_P (stmt_info) | |
6029 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
6030 | { | |
6031 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
6032 | gcc_assert (done); | |
6033 | } | |
6034 | ||
6035 | if (vec_stmt) | |
83197f37 | 6036 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 6037 | |
b8698a0f | 6038 | return is_store; |
ebfd146a IR |
6039 | } |
6040 | ||
6041 | ||
b8698a0f | 6042 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
6043 | stmt_vec_info. */ |
6044 | ||
6045 | void | |
6046 | vect_remove_stores (gimple first_stmt) | |
6047 | { | |
6048 | gimple next = first_stmt; | |
6049 | gimple tmp; | |
6050 | gimple_stmt_iterator next_si; | |
6051 | ||
6052 | while (next) | |
6053 | { | |
78048b1c JJ |
6054 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
6055 | ||
6056 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
6057 | if (is_pattern_stmt_p (stmt_info)) | |
6058 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
6059 | /* Free the attached stmt_vec_info and remove the stmt. */ |
6060 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 6061 | unlink_stmt_vdef (next); |
ebfd146a | 6062 | gsi_remove (&next_si, true); |
3d3f2249 | 6063 | release_defs (next); |
ebfd146a IR |
6064 | free_stmt_vec_info (next); |
6065 | next = tmp; | |
6066 | } | |
6067 | } | |
6068 | ||
6069 | ||
6070 | /* Function new_stmt_vec_info. | |
6071 | ||
6072 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
6073 | ||
6074 | stmt_vec_info | |
b8698a0f | 6075 | new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 6076 | bb_vec_info bb_vinfo) |
ebfd146a IR |
6077 | { |
6078 | stmt_vec_info res; | |
6079 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
6080 | ||
6081 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
6082 | STMT_VINFO_STMT (res) = stmt; | |
6083 | STMT_VINFO_LOOP_VINFO (res) = loop_vinfo; | |
a70d6342 | 6084 | STMT_VINFO_BB_VINFO (res) = bb_vinfo; |
8644a673 | 6085 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
6086 | STMT_VINFO_LIVE_P (res) = false; |
6087 | STMT_VINFO_VECTYPE (res) = NULL; | |
6088 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 6089 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
6090 | STMT_VINFO_IN_PATTERN_P (res) = false; |
6091 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 6092 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a IR |
6093 | STMT_VINFO_DATA_REF (res) = NULL; |
6094 | ||
6095 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
6096 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
6097 | STMT_VINFO_DR_INIT (res) = NULL; | |
6098 | STMT_VINFO_DR_STEP (res) = NULL; | |
6099 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
6100 | ||
6101 | if (gimple_code (stmt) == GIMPLE_PHI | |
6102 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
6103 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
6104 | else | |
8644a673 IR |
6105 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
6106 | ||
9771b263 | 6107 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 6108 | STMT_SLP_TYPE (res) = loop_vect; |
e14c1050 IR |
6109 | GROUP_FIRST_ELEMENT (res) = NULL; |
6110 | GROUP_NEXT_ELEMENT (res) = NULL; | |
6111 | GROUP_SIZE (res) = 0; | |
6112 | GROUP_STORE_COUNT (res) = 0; | |
6113 | GROUP_GAP (res) = 0; | |
6114 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
6115 | |
6116 | return res; | |
6117 | } | |
6118 | ||
6119 | ||
6120 | /* Create a hash table for stmt_vec_info. */ | |
6121 | ||
6122 | void | |
6123 | init_stmt_vec_info_vec (void) | |
6124 | { | |
9771b263 DN |
6125 | gcc_assert (!stmt_vec_info_vec.exists ()); |
6126 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
6127 | } |
6128 | ||
6129 | ||
6130 | /* Free hash table for stmt_vec_info. */ | |
6131 | ||
6132 | void | |
6133 | free_stmt_vec_info_vec (void) | |
6134 | { | |
93675444 JJ |
6135 | unsigned int i; |
6136 | vec_void_p info; | |
6137 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) | |
6138 | if (info != NULL) | |
6139 | free_stmt_vec_info (STMT_VINFO_STMT ((stmt_vec_info) info)); | |
9771b263 DN |
6140 | gcc_assert (stmt_vec_info_vec.exists ()); |
6141 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
6142 | } |
6143 | ||
6144 | ||
6145 | /* Free stmt vectorization related info. */ | |
6146 | ||
6147 | void | |
6148 | free_stmt_vec_info (gimple stmt) | |
6149 | { | |
6150 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6151 | ||
6152 | if (!stmt_info) | |
6153 | return; | |
6154 | ||
78048b1c JJ |
6155 | /* Check if this statement has a related "pattern stmt" |
6156 | (introduced by the vectorizer during the pattern recognition | |
6157 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
6158 | too. */ | |
6159 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
6160 | { | |
6161 | stmt_vec_info patt_info | |
6162 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
6163 | if (patt_info) | |
6164 | { | |
363477c0 JJ |
6165 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
6166 | if (seq) | |
6167 | { | |
6168 | gimple_stmt_iterator si; | |
6169 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
6170 | free_stmt_vec_info (gsi_stmt (si)); | |
6171 | } | |
78048b1c JJ |
6172 | free_stmt_vec_info (STMT_VINFO_RELATED_STMT (stmt_info)); |
6173 | } | |
6174 | } | |
6175 | ||
9771b263 | 6176 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
ebfd146a IR |
6177 | set_vinfo_for_stmt (stmt, NULL); |
6178 | free (stmt_info); | |
6179 | } | |
6180 | ||
6181 | ||
bb67d9c7 | 6182 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 6183 | |
bb67d9c7 | 6184 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
6185 | by the target. */ |
6186 | ||
bb67d9c7 RG |
6187 | static tree |
6188 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a IR |
6189 | { |
6190 | enum machine_mode inner_mode = TYPE_MODE (scalar_type); | |
cc4b5170 | 6191 | enum machine_mode simd_mode; |
2f816591 | 6192 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
6193 | int nunits; |
6194 | tree vectype; | |
6195 | ||
cc4b5170 | 6196 | if (nbytes == 0) |
ebfd146a IR |
6197 | return NULL_TREE; |
6198 | ||
48f2e373 RB |
6199 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
6200 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
6201 | return NULL_TREE; | |
6202 | ||
7b7b1813 RG |
6203 | /* For vector types of elements whose mode precision doesn't |
6204 | match their types precision we use a element type of mode | |
6205 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
6206 | they support the proper result truncation/extension. |
6207 | We also make sure to build vector types with INTEGER_TYPE | |
6208 | component type only. */ | |
6d7971b8 | 6209 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
6210 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
6211 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
6212 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
6213 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 6214 | |
ccbf5bb4 RG |
6215 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
6216 | When the component mode passes the above test simply use a type | |
6217 | corresponding to that mode. The theory is that any use that | |
6218 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 6219 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 6220 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
6221 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
6222 | ||
6223 | /* We can't build a vector type of elements with alignment bigger than | |
6224 | their size. */ | |
dfc2e2ac | 6225 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
6226 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
6227 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 6228 | |
dfc2e2ac RB |
6229 | /* If we felt back to using the mode fail if there was |
6230 | no scalar type for it. */ | |
6231 | if (scalar_type == NULL_TREE) | |
6232 | return NULL_TREE; | |
6233 | ||
bb67d9c7 RG |
6234 | /* If no size was supplied use the mode the target prefers. Otherwise |
6235 | lookup a vector mode of the specified size. */ | |
6236 | if (size == 0) | |
6237 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
6238 | else | |
6239 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
6240 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
6241 | if (nunits <= 1) | |
6242 | return NULL_TREE; | |
ebfd146a IR |
6243 | |
6244 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
6245 | |
6246 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
6247 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 6248 | return NULL_TREE; |
ebfd146a IR |
6249 | |
6250 | return vectype; | |
6251 | } | |
6252 | ||
bb67d9c7 RG |
6253 | unsigned int current_vector_size; |
6254 | ||
6255 | /* Function get_vectype_for_scalar_type. | |
6256 | ||
6257 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
6258 | by the target. */ | |
6259 | ||
6260 | tree | |
6261 | get_vectype_for_scalar_type (tree scalar_type) | |
6262 | { | |
6263 | tree vectype; | |
6264 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
6265 | current_vector_size); | |
6266 | if (vectype | |
6267 | && current_vector_size == 0) | |
6268 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
6269 | return vectype; | |
6270 | } | |
6271 | ||
b690cc0f RG |
6272 | /* Function get_same_sized_vectype |
6273 | ||
6274 | Returns a vector type corresponding to SCALAR_TYPE of size | |
6275 | VECTOR_TYPE if supported by the target. */ | |
6276 | ||
6277 | tree | |
bb67d9c7 | 6278 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 6279 | { |
bb67d9c7 RG |
6280 | return get_vectype_for_scalar_type_and_size |
6281 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
6282 | } |
6283 | ||
ebfd146a IR |
6284 | /* Function vect_is_simple_use. |
6285 | ||
6286 | Input: | |
a70d6342 IR |
6287 | LOOP_VINFO - the vect info of the loop that is being vectorized. |
6288 | BB_VINFO - the vect info of the basic block that is being vectorized. | |
24ee1384 | 6289 | OPERAND - operand of STMT in the loop or bb. |
ebfd146a IR |
6290 | DEF - the defining stmt in case OPERAND is an SSA_NAME. |
6291 | ||
6292 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 6293 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 6294 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 6295 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
6296 | is the case in reduction/induction computations). |
6297 | For basic blocks, supportable operands are constants and bb invariants. | |
6298 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
6299 | |
6300 | bool | |
24ee1384 | 6301 | vect_is_simple_use (tree operand, gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 6302 | bb_vec_info bb_vinfo, gimple *def_stmt, |
ebfd146a | 6303 | tree *def, enum vect_def_type *dt) |
b8698a0f | 6304 | { |
ebfd146a IR |
6305 | basic_block bb; |
6306 | stmt_vec_info stmt_vinfo; | |
a70d6342 | 6307 | struct loop *loop = NULL; |
b8698a0f | 6308 | |
a70d6342 IR |
6309 | if (loop_vinfo) |
6310 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
6311 | |
6312 | *def_stmt = NULL; | |
6313 | *def = NULL_TREE; | |
b8698a0f | 6314 | |
73fbfcad | 6315 | if (dump_enabled_p ()) |
ebfd146a | 6316 | { |
78c60e3d SS |
6317 | dump_printf_loc (MSG_NOTE, vect_location, |
6318 | "vect_is_simple_use: operand "); | |
6319 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 6320 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 6321 | } |
b8698a0f | 6322 | |
b758f602 | 6323 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
6324 | { |
6325 | *dt = vect_constant_def; | |
6326 | return true; | |
6327 | } | |
b8698a0f | 6328 | |
ebfd146a IR |
6329 | if (is_gimple_min_invariant (operand)) |
6330 | { | |
6331 | *def = operand; | |
8644a673 | 6332 | *dt = vect_external_def; |
ebfd146a IR |
6333 | return true; |
6334 | } | |
6335 | ||
6336 | if (TREE_CODE (operand) == PAREN_EXPR) | |
6337 | { | |
73fbfcad | 6338 | if (dump_enabled_p ()) |
e645e942 | 6339 | dump_printf_loc (MSG_NOTE, vect_location, "non-associatable copy.\n"); |
ebfd146a IR |
6340 | operand = TREE_OPERAND (operand, 0); |
6341 | } | |
b8698a0f | 6342 | |
ebfd146a IR |
6343 | if (TREE_CODE (operand) != SSA_NAME) |
6344 | { | |
73fbfcad | 6345 | if (dump_enabled_p ()) |
78c60e3d | 6346 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6347 | "not ssa-name.\n"); |
ebfd146a IR |
6348 | return false; |
6349 | } | |
b8698a0f | 6350 | |
ebfd146a IR |
6351 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
6352 | if (*def_stmt == NULL) | |
6353 | { | |
73fbfcad | 6354 | if (dump_enabled_p ()) |
78c60e3d | 6355 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6356 | "no def_stmt.\n"); |
ebfd146a IR |
6357 | return false; |
6358 | } | |
6359 | ||
73fbfcad | 6360 | if (dump_enabled_p ()) |
ebfd146a | 6361 | { |
78c60e3d SS |
6362 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
6363 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
e645e942 | 6364 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
6365 | } |
6366 | ||
8644a673 | 6367 | /* Empty stmt is expected only in case of a function argument. |
ebfd146a IR |
6368 | (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */ |
6369 | if (gimple_nop_p (*def_stmt)) | |
6370 | { | |
6371 | *def = operand; | |
8644a673 | 6372 | *dt = vect_external_def; |
ebfd146a IR |
6373 | return true; |
6374 | } | |
6375 | ||
6376 | bb = gimple_bb (*def_stmt); | |
a70d6342 IR |
6377 | |
6378 | if ((loop && !flow_bb_inside_loop_p (loop, bb)) | |
6379 | || (!loop && bb != BB_VINFO_BB (bb_vinfo)) | |
b8698a0f | 6380 | || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI)) |
8644a673 | 6381 | *dt = vect_external_def; |
ebfd146a IR |
6382 | else |
6383 | { | |
6384 | stmt_vinfo = vinfo_for_stmt (*def_stmt); | |
6385 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
6386 | } | |
6387 | ||
24ee1384 IR |
6388 | if (*dt == vect_unknown_def_type |
6389 | || (stmt | |
6390 | && *dt == vect_double_reduction_def | |
6391 | && gimple_code (stmt) != GIMPLE_PHI)) | |
ebfd146a | 6392 | { |
73fbfcad | 6393 | if (dump_enabled_p ()) |
78c60e3d | 6394 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6395 | "Unsupported pattern.\n"); |
ebfd146a IR |
6396 | return false; |
6397 | } | |
6398 | ||
73fbfcad | 6399 | if (dump_enabled_p ()) |
e645e942 | 6400 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: %d.\n", *dt); |
ebfd146a IR |
6401 | |
6402 | switch (gimple_code (*def_stmt)) | |
6403 | { | |
6404 | case GIMPLE_PHI: | |
6405 | *def = gimple_phi_result (*def_stmt); | |
6406 | break; | |
6407 | ||
6408 | case GIMPLE_ASSIGN: | |
6409 | *def = gimple_assign_lhs (*def_stmt); | |
6410 | break; | |
6411 | ||
6412 | case GIMPLE_CALL: | |
6413 | *def = gimple_call_lhs (*def_stmt); | |
6414 | if (*def != NULL) | |
6415 | break; | |
6416 | /* FALLTHRU */ | |
6417 | default: | |
73fbfcad | 6418 | if (dump_enabled_p ()) |
78c60e3d | 6419 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6420 | "unsupported defining stmt:\n"); |
ebfd146a IR |
6421 | return false; |
6422 | } | |
6423 | ||
6424 | return true; | |
6425 | } | |
6426 | ||
b690cc0f RG |
6427 | /* Function vect_is_simple_use_1. |
6428 | ||
6429 | Same as vect_is_simple_use_1 but also determines the vector operand | |
6430 | type of OPERAND and stores it to *VECTYPE. If the definition of | |
6431 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
6432 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
6433 | is responsible to compute the best suited vector type for the | |
6434 | scalar operand. */ | |
6435 | ||
6436 | bool | |
24ee1384 | 6437 | vect_is_simple_use_1 (tree operand, gimple stmt, loop_vec_info loop_vinfo, |
b690cc0f RG |
6438 | bb_vec_info bb_vinfo, gimple *def_stmt, |
6439 | tree *def, enum vect_def_type *dt, tree *vectype) | |
6440 | { | |
24ee1384 IR |
6441 | if (!vect_is_simple_use (operand, stmt, loop_vinfo, bb_vinfo, def_stmt, |
6442 | def, dt)) | |
b690cc0f RG |
6443 | return false; |
6444 | ||
6445 | /* Now get a vector type if the def is internal, otherwise supply | |
6446 | NULL_TREE and leave it up to the caller to figure out a proper | |
6447 | type for the use stmt. */ | |
6448 | if (*dt == vect_internal_def | |
6449 | || *dt == vect_induction_def | |
6450 | || *dt == vect_reduction_def | |
6451 | || *dt == vect_double_reduction_def | |
6452 | || *dt == vect_nested_cycle) | |
6453 | { | |
6454 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
6455 | |
6456 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
6457 | && !STMT_VINFO_RELEVANT (stmt_info) | |
6458 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 6459 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 6460 | |
b690cc0f RG |
6461 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
6462 | gcc_assert (*vectype != NULL_TREE); | |
6463 | } | |
6464 | else if (*dt == vect_uninitialized_def | |
6465 | || *dt == vect_constant_def | |
6466 | || *dt == vect_external_def) | |
6467 | *vectype = NULL_TREE; | |
6468 | else | |
6469 | gcc_unreachable (); | |
6470 | ||
6471 | return true; | |
6472 | } | |
6473 | ||
ebfd146a IR |
6474 | |
6475 | /* Function supportable_widening_operation | |
6476 | ||
b8698a0f L |
6477 | Check whether an operation represented by the code CODE is a |
6478 | widening operation that is supported by the target platform in | |
b690cc0f RG |
6479 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
6480 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 6481 | |
ebfd146a IR |
6482 | Widening operations we currently support are NOP (CONVERT), FLOAT |
6483 | and WIDEN_MULT. This function checks if these operations are supported | |
6484 | by the target platform either directly (via vector tree-codes), or via | |
6485 | target builtins. | |
6486 | ||
6487 | Output: | |
b8698a0f L |
6488 | - CODE1 and CODE2 are codes of vector operations to be used when |
6489 | vectorizing the operation, if available. | |
ebfd146a IR |
6490 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
6491 | case of multi-step conversion (like char->short->int - in that case | |
6492 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
6493 | - INTERM_TYPES contains the intermediate type required to perform the |
6494 | widening operation (short in the above example). */ | |
ebfd146a IR |
6495 | |
6496 | bool | |
b690cc0f RG |
6497 | supportable_widening_operation (enum tree_code code, gimple stmt, |
6498 | tree vectype_out, tree vectype_in, | |
ebfd146a IR |
6499 | enum tree_code *code1, enum tree_code *code2, |
6500 | int *multi_step_cvt, | |
9771b263 | 6501 | vec<tree> *interm_types) |
ebfd146a IR |
6502 | { |
6503 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6504 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 6505 | struct loop *vect_loop = NULL; |
ebfd146a | 6506 | enum machine_mode vec_mode; |
81f40b79 | 6507 | enum insn_code icode1, icode2; |
ebfd146a | 6508 | optab optab1, optab2; |
b690cc0f RG |
6509 | tree vectype = vectype_in; |
6510 | tree wide_vectype = vectype_out; | |
ebfd146a | 6511 | enum tree_code c1, c2; |
4a00c761 JJ |
6512 | int i; |
6513 | tree prev_type, intermediate_type; | |
6514 | enum machine_mode intermediate_mode, prev_mode; | |
6515 | optab optab3, optab4; | |
ebfd146a | 6516 | |
4a00c761 | 6517 | *multi_step_cvt = 0; |
4ef69dfc IR |
6518 | if (loop_info) |
6519 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
6520 | ||
ebfd146a IR |
6521 | switch (code) |
6522 | { | |
6523 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
6524 | /* The result of a vectorized widening operation usually requires |
6525 | two vectors (because the widened results do not fit into one vector). | |
6526 | The generated vector results would normally be expected to be | |
6527 | generated in the same order as in the original scalar computation, | |
6528 | i.e. if 8 results are generated in each vector iteration, they are | |
6529 | to be organized as follows: | |
6530 | vect1: [res1,res2,res3,res4], | |
6531 | vect2: [res5,res6,res7,res8]. | |
6532 | ||
6533 | However, in the special case that the result of the widening | |
6534 | operation is used in a reduction computation only, the order doesn't | |
6535 | matter (because when vectorizing a reduction we change the order of | |
6536 | the computation). Some targets can take advantage of this and | |
6537 | generate more efficient code. For example, targets like Altivec, | |
6538 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
6539 | generate the following vectors: | |
6540 | vect1: [res1,res3,res5,res7], | |
6541 | vect2: [res2,res4,res6,res8]. | |
6542 | ||
6543 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
6544 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
6545 | iterations in parallel). We therefore don't allow to change the | |
6546 | order of the computation in the inner-loop during outer-loop | |
6547 | vectorization. */ | |
6548 | /* TODO: Another case in which order doesn't *really* matter is when we | |
6549 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
6550 | Normally, pack_trunc performs an even/odd permute, whereas the | |
6551 | repack from an even/odd expansion would be an interleave, which | |
6552 | would be significantly simpler for e.g. AVX2. */ | |
6553 | /* In any case, in order to avoid duplicating the code below, recurse | |
6554 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
6555 | are properly set up for the caller. If we fail, we'll continue with | |
6556 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
6557 | if (vect_loop | |
6558 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
6559 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
6560 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
6561 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
6562 | code1, code2, multi_step_cvt, |
6563 | interm_types)) | |
6ae6116f | 6564 | return true; |
4a00c761 JJ |
6565 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
6566 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
6567 | break; |
6568 | ||
6ae6116f RH |
6569 | case VEC_WIDEN_MULT_EVEN_EXPR: |
6570 | /* Support the recursion induced just above. */ | |
6571 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
6572 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
6573 | break; | |
6574 | ||
36ba4aae | 6575 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
6576 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
6577 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
6578 | break; |
6579 | ||
ebfd146a | 6580 | CASE_CONVERT: |
4a00c761 JJ |
6581 | c1 = VEC_UNPACK_LO_EXPR; |
6582 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
6583 | break; |
6584 | ||
6585 | case FLOAT_EXPR: | |
4a00c761 JJ |
6586 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
6587 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
6588 | break; |
6589 | ||
6590 | case FIX_TRUNC_EXPR: | |
6591 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
6592 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
6593 | computing the operation. */ | |
6594 | return false; | |
6595 | ||
6596 | default: | |
6597 | gcc_unreachable (); | |
6598 | } | |
6599 | ||
6ae6116f | 6600 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
4a00c761 JJ |
6601 | { |
6602 | enum tree_code ctmp = c1; | |
6603 | c1 = c2; | |
6604 | c2 = ctmp; | |
6605 | } | |
6606 | ||
ebfd146a IR |
6607 | if (code == FIX_TRUNC_EXPR) |
6608 | { | |
6609 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
6610 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
6611 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
6612 | } |
6613 | else | |
6614 | { | |
6615 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
6616 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
6617 | } | |
6618 | ||
6619 | if (!optab1 || !optab2) | |
6620 | return false; | |
6621 | ||
6622 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
6623 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
6624 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
6625 | return false; |
6626 | ||
4a00c761 JJ |
6627 | *code1 = c1; |
6628 | *code2 = c2; | |
6629 | ||
6630 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
6631 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
6632 | return true; | |
6633 | ||
b8698a0f | 6634 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 6635 | types. */ |
ebfd146a | 6636 | |
4a00c761 JJ |
6637 | prev_type = vectype; |
6638 | prev_mode = vec_mode; | |
b8698a0f | 6639 | |
4a00c761 JJ |
6640 | if (!CONVERT_EXPR_CODE_P (code)) |
6641 | return false; | |
b8698a0f | 6642 | |
4a00c761 JJ |
6643 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
6644 | intermediate steps in promotion sequence. We try | |
6645 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
6646 | not. */ | |
9771b263 | 6647 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
6648 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
6649 | { | |
6650 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
6651 | intermediate_type | |
6652 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
6653 | TYPE_UNSIGNED (prev_type)); | |
6654 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
6655 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
6656 | ||
6657 | if (!optab3 || !optab4 | |
6658 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
6659 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
6660 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
6661 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
6662 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
6663 | == CODE_FOR_nothing) | |
6664 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
6665 | == CODE_FOR_nothing)) | |
6666 | break; | |
ebfd146a | 6667 | |
9771b263 | 6668 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
6669 | (*multi_step_cvt)++; |
6670 | ||
6671 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
6672 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
6673 | return true; | |
6674 | ||
6675 | prev_type = intermediate_type; | |
6676 | prev_mode = intermediate_mode; | |
ebfd146a IR |
6677 | } |
6678 | ||
9771b263 | 6679 | interm_types->release (); |
4a00c761 | 6680 | return false; |
ebfd146a IR |
6681 | } |
6682 | ||
6683 | ||
6684 | /* Function supportable_narrowing_operation | |
6685 | ||
b8698a0f L |
6686 | Check whether an operation represented by the code CODE is a |
6687 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
6688 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
6689 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 6690 | |
ebfd146a | 6691 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 6692 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
6693 | the target platform directly via vector tree-codes. |
6694 | ||
6695 | Output: | |
b8698a0f L |
6696 | - CODE1 is the code of a vector operation to be used when |
6697 | vectorizing the operation, if available. | |
ebfd146a IR |
6698 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
6699 | case of multi-step conversion (like int->short->char - in that case | |
6700 | MULTI_STEP_CVT will be 1). | |
6701 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 6702 | narrowing operation (short in the above example). */ |
ebfd146a IR |
6703 | |
6704 | bool | |
6705 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 6706 | tree vectype_out, tree vectype_in, |
ebfd146a | 6707 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 6708 | vec<tree> *interm_types) |
ebfd146a IR |
6709 | { |
6710 | enum machine_mode vec_mode; | |
6711 | enum insn_code icode1; | |
6712 | optab optab1, interm_optab; | |
b690cc0f RG |
6713 | tree vectype = vectype_in; |
6714 | tree narrow_vectype = vectype_out; | |
ebfd146a | 6715 | enum tree_code c1; |
4a00c761 JJ |
6716 | tree intermediate_type; |
6717 | enum machine_mode intermediate_mode, prev_mode; | |
ebfd146a | 6718 | int i; |
4a00c761 | 6719 | bool uns; |
ebfd146a | 6720 | |
4a00c761 | 6721 | *multi_step_cvt = 0; |
ebfd146a IR |
6722 | switch (code) |
6723 | { | |
6724 | CASE_CONVERT: | |
6725 | c1 = VEC_PACK_TRUNC_EXPR; | |
6726 | break; | |
6727 | ||
6728 | case FIX_TRUNC_EXPR: | |
6729 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
6730 | break; | |
6731 | ||
6732 | case FLOAT_EXPR: | |
6733 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
6734 | tree code and optabs used for computing the operation. */ | |
6735 | return false; | |
6736 | ||
6737 | default: | |
6738 | gcc_unreachable (); | |
6739 | } | |
6740 | ||
6741 | if (code == FIX_TRUNC_EXPR) | |
6742 | /* The signedness is determined from output operand. */ | |
b690cc0f | 6743 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
6744 | else |
6745 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
6746 | ||
6747 | if (!optab1) | |
6748 | return false; | |
6749 | ||
6750 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 6751 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
6752 | return false; |
6753 | ||
4a00c761 JJ |
6754 | *code1 = c1; |
6755 | ||
6756 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
6757 | return true; | |
6758 | ||
ebfd146a IR |
6759 | /* Check if it's a multi-step conversion that can be done using intermediate |
6760 | types. */ | |
4a00c761 JJ |
6761 | prev_mode = vec_mode; |
6762 | if (code == FIX_TRUNC_EXPR) | |
6763 | uns = TYPE_UNSIGNED (vectype_out); | |
6764 | else | |
6765 | uns = TYPE_UNSIGNED (vectype); | |
6766 | ||
6767 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
6768 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
6769 | costly than signed. */ | |
6770 | if (code == FIX_TRUNC_EXPR && uns) | |
6771 | { | |
6772 | enum insn_code icode2; | |
6773 | ||
6774 | intermediate_type | |
6775 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
6776 | interm_optab | |
6777 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 6778 | if (interm_optab != unknown_optab |
4a00c761 JJ |
6779 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
6780 | && insn_data[icode1].operand[0].mode | |
6781 | == insn_data[icode2].operand[0].mode) | |
6782 | { | |
6783 | uns = false; | |
6784 | optab1 = interm_optab; | |
6785 | icode1 = icode2; | |
6786 | } | |
6787 | } | |
ebfd146a | 6788 | |
4a00c761 JJ |
6789 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
6790 | intermediate steps in promotion sequence. We try | |
6791 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 6792 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
6793 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
6794 | { | |
6795 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
6796 | intermediate_type | |
6797 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
6798 | interm_optab | |
6799 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
6800 | optab_default); | |
6801 | if (!interm_optab | |
6802 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
6803 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
6804 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
6805 | == CODE_FOR_nothing)) | |
6806 | break; | |
6807 | ||
9771b263 | 6808 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
6809 | (*multi_step_cvt)++; |
6810 | ||
6811 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
6812 | return true; | |
6813 | ||
6814 | prev_mode = intermediate_mode; | |
6815 | optab1 = interm_optab; | |
ebfd146a IR |
6816 | } |
6817 | ||
9771b263 | 6818 | interm_types->release (); |
4a00c761 | 6819 | return false; |
ebfd146a | 6820 | } |