]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
5624e564 | 2 | Copyright (C) 2003-2015 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" | |
c7131fb2 | 25 | #include "backend.h" |
957060b5 AM |
26 | #include "target.h" |
27 | #include "rtl.h" | |
ebfd146a | 28 | #include "tree.h" |
c7131fb2 | 29 | #include "gimple.h" |
c7131fb2 | 30 | #include "ssa.h" |
957060b5 AM |
31 | #include "optabs-tree.h" |
32 | #include "insn-config.h" | |
33 | #include "recog.h" /* FIXME: for insn_data */ | |
34 | #include "cgraph.h" | |
957060b5 | 35 | #include "dumpfile.h" |
c7131fb2 | 36 | #include "alias.h" |
40e23961 | 37 | #include "fold-const.h" |
d8a2d370 | 38 | #include "stor-layout.h" |
2fb9a547 | 39 | #include "tree-eh.h" |
45b0be94 | 40 | #include "gimplify.h" |
5be5c238 | 41 | #include "gimple-iterator.h" |
18f429e2 | 42 | #include "gimplify-me.h" |
442b4905 | 43 | #include "tree-cfg.h" |
e28030cf | 44 | #include "tree-ssa-loop-manip.h" |
ebfd146a | 45 | #include "cfgloop.h" |
0136f8f0 AH |
46 | #include "tree-ssa-loop.h" |
47 | #include "tree-scalar-evolution.h" | |
ebfd146a | 48 | #include "tree-vectorizer.h" |
9b2b7279 | 49 | #include "builtins.h" |
ebfd146a | 50 | |
7ee2468b SB |
51 | /* For lang_hooks.types.type_for_mode. */ |
52 | #include "langhooks.h" | |
ebfd146a | 53 | |
c3e7ee41 BS |
54 | /* Return the vectorized type for the given statement. */ |
55 | ||
56 | tree | |
57 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
58 | { | |
59 | return STMT_VINFO_VECTYPE (stmt_info); | |
60 | } | |
61 | ||
62 | /* Return TRUE iff the given statement is in an inner loop relative to | |
63 | the loop being vectorized. */ | |
64 | bool | |
65 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
66 | { | |
355fe088 | 67 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
c3e7ee41 BS |
68 | basic_block bb = gimple_bb (stmt); |
69 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
70 | struct loop* loop; | |
71 | ||
72 | if (!loop_vinfo) | |
73 | return false; | |
74 | ||
75 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
76 | ||
77 | return (bb->loop_father == loop->inner); | |
78 | } | |
79 | ||
80 | /* Record the cost of a statement, either by directly informing the | |
81 | target model or by saving it in a vector for later processing. | |
82 | Return a preliminary estimate of the statement's cost. */ | |
83 | ||
84 | unsigned | |
92345349 | 85 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 86 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 87 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 88 | { |
92345349 | 89 | if (body_cost_vec) |
c3e7ee41 | 90 | { |
92345349 | 91 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
ddf56386 RB |
92 | stmt_info_for_cost si = { count, kind, |
93 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
94 | misalign }; | |
95 | body_cost_vec->safe_push (si); | |
c3e7ee41 | 96 | return (unsigned) |
92345349 | 97 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
98 | } |
99 | else | |
310213d4 RB |
100 | return add_stmt_cost (stmt_info->vinfo->target_cost_data, |
101 | count, kind, stmt_info, misalign, where); | |
c3e7ee41 BS |
102 | } |
103 | ||
272c6793 RS |
104 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
105 | ||
106 | static tree | |
107 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
108 | { | |
109 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
110 | "vect_array"); | |
111 | } | |
112 | ||
113 | /* ARRAY is an array of vectors created by create_vector_array. | |
114 | Return an SSA_NAME for the vector in index N. The reference | |
115 | is part of the vectorization of STMT and the vector is associated | |
116 | with scalar destination SCALAR_DEST. */ | |
117 | ||
118 | static tree | |
355fe088 | 119 | read_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree scalar_dest, |
272c6793 RS |
120 | tree array, unsigned HOST_WIDE_INT n) |
121 | { | |
122 | tree vect_type, vect, vect_name, array_ref; | |
355fe088 | 123 | gimple *new_stmt; |
272c6793 RS |
124 | |
125 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
126 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
127 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
128 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
129 | build_int_cst (size_type_node, n), | |
130 | NULL_TREE, NULL_TREE); | |
131 | ||
132 | new_stmt = gimple_build_assign (vect, array_ref); | |
133 | vect_name = make_ssa_name (vect, new_stmt); | |
134 | gimple_assign_set_lhs (new_stmt, vect_name); | |
135 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
136 | |
137 | return vect_name; | |
138 | } | |
139 | ||
140 | /* ARRAY is an array of vectors created by create_vector_array. | |
141 | Emit code to store SSA_NAME VECT in index N of the array. | |
142 | The store is part of the vectorization of STMT. */ | |
143 | ||
144 | static void | |
355fe088 | 145 | write_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree vect, |
272c6793 RS |
146 | tree array, unsigned HOST_WIDE_INT n) |
147 | { | |
148 | tree array_ref; | |
355fe088 | 149 | gimple *new_stmt; |
272c6793 RS |
150 | |
151 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
152 | build_int_cst (size_type_node, n), | |
153 | NULL_TREE, NULL_TREE); | |
154 | ||
155 | new_stmt = gimple_build_assign (array_ref, vect); | |
156 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
157 | } |
158 | ||
159 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
160 | of *PTR. The memory reference replaces those in FIRST_DR | |
161 | (and its group). */ | |
162 | ||
163 | static tree | |
164 | create_array_ref (tree type, tree ptr, struct data_reference *first_dr) | |
165 | { | |
272c6793 RS |
166 | tree mem_ref, alias_ptr_type; |
167 | ||
168 | alias_ptr_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
169 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); | |
170 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 171 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
172 | return mem_ref; |
173 | } | |
174 | ||
ebfd146a IR |
175 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
176 | ||
177 | /* Function vect_mark_relevant. | |
178 | ||
179 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
180 | ||
181 | static void | |
355fe088 | 182 | vect_mark_relevant (vec<gimple *> *worklist, gimple *stmt, |
83197f37 IR |
183 | enum vect_relevant relevant, bool live_p, |
184 | bool used_in_pattern) | |
ebfd146a IR |
185 | { |
186 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
187 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
188 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
355fe088 | 189 | gimple *pattern_stmt; |
ebfd146a | 190 | |
73fbfcad | 191 | if (dump_enabled_p ()) |
78c60e3d | 192 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 193 | "mark relevant %d, live %d.\n", relevant, live_p); |
ebfd146a | 194 | |
83197f37 IR |
195 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
196 | related pattern stmt instead of the original stmt. However, such stmts | |
197 | may have their own uses that are not in any pattern, in such cases the | |
198 | stmt itself should be marked. */ | |
ebfd146a IR |
199 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
200 | { | |
83197f37 IR |
201 | bool found = false; |
202 | if (!used_in_pattern) | |
203 | { | |
204 | imm_use_iterator imm_iter; | |
205 | use_operand_p use_p; | |
355fe088 | 206 | gimple *use_stmt; |
83197f37 | 207 | tree lhs; |
13c931c9 JJ |
208 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
209 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a | 210 | |
83197f37 IR |
211 | if (is_gimple_assign (stmt)) |
212 | lhs = gimple_assign_lhs (stmt); | |
213 | else | |
214 | lhs = gimple_call_lhs (stmt); | |
ebfd146a | 215 | |
83197f37 IR |
216 | /* This use is out of pattern use, if LHS has other uses that are |
217 | pattern uses, we should mark the stmt itself, and not the pattern | |
218 | stmt. */ | |
5ce9450f | 219 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
ab0ef706 JJ |
220 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs) |
221 | { | |
222 | if (is_gimple_debug (USE_STMT (use_p))) | |
223 | continue; | |
224 | use_stmt = USE_STMT (use_p); | |
225 | ||
13c931c9 JJ |
226 | if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) |
227 | continue; | |
228 | ||
ab0ef706 JJ |
229 | if (vinfo_for_stmt (use_stmt) |
230 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt))) | |
231 | { | |
232 | found = true; | |
233 | break; | |
234 | } | |
235 | } | |
83197f37 IR |
236 | } |
237 | ||
238 | if (!found) | |
239 | { | |
240 | /* This is the last stmt in a sequence that was detected as a | |
241 | pattern that can potentially be vectorized. Don't mark the stmt | |
242 | as relevant/live because it's not going to be vectorized. | |
243 | Instead mark the pattern-stmt that replaces it. */ | |
244 | ||
245 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
246 | ||
73fbfcad | 247 | if (dump_enabled_p ()) |
78c60e3d SS |
248 | dump_printf_loc (MSG_NOTE, vect_location, |
249 | "last stmt in pattern. don't mark" | |
e645e942 | 250 | " relevant/live.\n"); |
83197f37 IR |
251 | stmt_info = vinfo_for_stmt (pattern_stmt); |
252 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
253 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
254 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
255 | stmt = pattern_stmt; | |
256 | } | |
ebfd146a IR |
257 | } |
258 | ||
259 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
260 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
261 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
262 | ||
263 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
264 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
265 | { | |
73fbfcad | 266 | if (dump_enabled_p ()) |
78c60e3d | 267 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 268 | "already marked relevant/live.\n"); |
ebfd146a IR |
269 | return; |
270 | } | |
271 | ||
9771b263 | 272 | worklist->safe_push (stmt); |
ebfd146a IR |
273 | } |
274 | ||
275 | ||
276 | /* Function vect_stmt_relevant_p. | |
277 | ||
278 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
279 | "relevant for vectorization". | |
280 | ||
281 | A stmt is considered "relevant for vectorization" if: | |
282 | - it has uses outside the loop. | |
283 | - it has vdefs (it alters memory). | |
284 | - control stmts in the loop (except for the exit condition). | |
285 | ||
286 | CHECKME: what other side effects would the vectorizer allow? */ | |
287 | ||
288 | static bool | |
355fe088 | 289 | vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo, |
ebfd146a IR |
290 | enum vect_relevant *relevant, bool *live_p) |
291 | { | |
292 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
293 | ssa_op_iter op_iter; | |
294 | imm_use_iterator imm_iter; | |
295 | use_operand_p use_p; | |
296 | def_operand_p def_p; | |
297 | ||
8644a673 | 298 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
299 | *live_p = false; |
300 | ||
301 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
302 | if (is_ctrl_stmt (stmt) |
303 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
304 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 305 | *relevant = vect_used_in_scope; |
ebfd146a IR |
306 | |
307 | /* changing memory. */ | |
308 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
309 | if (gimple_vdef (stmt) |
310 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 311 | { |
73fbfcad | 312 | if (dump_enabled_p ()) |
78c60e3d | 313 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 314 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 315 | *relevant = vect_used_in_scope; |
ebfd146a IR |
316 | } |
317 | ||
318 | /* uses outside the loop. */ | |
319 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
320 | { | |
321 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
322 | { | |
323 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
324 | if (!flow_bb_inside_loop_p (loop, bb)) | |
325 | { | |
73fbfcad | 326 | if (dump_enabled_p ()) |
78c60e3d | 327 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 328 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 329 | |
3157b0c2 AO |
330 | if (is_gimple_debug (USE_STMT (use_p))) |
331 | continue; | |
332 | ||
ebfd146a IR |
333 | /* We expect all such uses to be in the loop exit phis |
334 | (because of loop closed form) */ | |
335 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
336 | gcc_assert (bb == single_exit (loop)->dest); | |
337 | ||
338 | *live_p = true; | |
339 | } | |
340 | } | |
341 | } | |
342 | ||
343 | return (*live_p || *relevant); | |
344 | } | |
345 | ||
346 | ||
b8698a0f | 347 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 348 | |
ff802fa1 | 349 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
350 | used in STMT for anything other than indexing an array. */ |
351 | ||
352 | static bool | |
355fe088 | 353 | exist_non_indexing_operands_for_use_p (tree use, gimple *stmt) |
ebfd146a IR |
354 | { |
355 | tree operand; | |
356 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 357 | |
ff802fa1 | 358 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
359 | reference in STMT, then any operand that corresponds to USE |
360 | is not indexing an array. */ | |
361 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
362 | return true; | |
59a05b0c | 363 | |
ebfd146a IR |
364 | /* STMT has a data_ref. FORNOW this means that its of one of |
365 | the following forms: | |
366 | -1- ARRAY_REF = var | |
367 | -2- var = ARRAY_REF | |
368 | (This should have been verified in analyze_data_refs). | |
369 | ||
370 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 371 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
372 | for array indexing. |
373 | ||
374 | Therefore, all we need to check is if STMT falls into the | |
375 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
376 | |
377 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
378 | { |
379 | if (is_gimple_call (stmt) | |
380 | && gimple_call_internal_p (stmt)) | |
381 | switch (gimple_call_internal_fn (stmt)) | |
382 | { | |
383 | case IFN_MASK_STORE: | |
384 | operand = gimple_call_arg (stmt, 3); | |
385 | if (operand == use) | |
386 | return true; | |
387 | /* FALLTHRU */ | |
388 | case IFN_MASK_LOAD: | |
389 | operand = gimple_call_arg (stmt, 2); | |
390 | if (operand == use) | |
391 | return true; | |
392 | break; | |
393 | default: | |
394 | break; | |
395 | } | |
396 | return false; | |
397 | } | |
398 | ||
59a05b0c EB |
399 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
400 | return false; | |
ebfd146a | 401 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
402 | if (TREE_CODE (operand) != SSA_NAME) |
403 | return false; | |
404 | ||
405 | if (operand == use) | |
406 | return true; | |
407 | ||
408 | return false; | |
409 | } | |
410 | ||
411 | ||
b8698a0f | 412 | /* |
ebfd146a IR |
413 | Function process_use. |
414 | ||
415 | Inputs: | |
416 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 417 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ff802fa1 | 418 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 419 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
420 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
421 | be performed. | |
ebfd146a IR |
422 | |
423 | Outputs: | |
424 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
425 | relevance info of the DEF_STMT of this USE: | |
426 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
427 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
428 | Exceptions: | |
429 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 430 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 431 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
432 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
433 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
434 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
435 | be modified accordingly. | |
436 | ||
437 | Return true if everything is as expected. Return false otherwise. */ | |
438 | ||
439 | static bool | |
355fe088 TS |
440 | process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
441 | enum vect_relevant relevant, vec<gimple *> *worklist, | |
aec7ae7d | 442 | bool force) |
ebfd146a IR |
443 | { |
444 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
445 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
446 | stmt_vec_info dstmt_vinfo; | |
447 | basic_block bb, def_bb; | |
355fe088 | 448 | gimple *def_stmt; |
ebfd146a IR |
449 | enum vect_def_type dt; |
450 | ||
b8698a0f | 451 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 452 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 453 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
454 | return true; |
455 | ||
81c40241 | 456 | if (!vect_is_simple_use (use, loop_vinfo, &def_stmt, &dt)) |
b8698a0f | 457 | { |
73fbfcad | 458 | if (dump_enabled_p ()) |
78c60e3d | 459 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 460 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
461 | return false; |
462 | } | |
463 | ||
464 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
465 | return true; | |
466 | ||
467 | def_bb = gimple_bb (def_stmt); | |
468 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
469 | { | |
73fbfcad | 470 | if (dump_enabled_p ()) |
e645e942 | 471 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
472 | return true; |
473 | } | |
474 | ||
b8698a0f L |
475 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
476 | DEF_STMT must have already been processed, because this should be the | |
477 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
478 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
479 | check that everything is as expected, and we are done. */ |
480 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
481 | bb = gimple_bb (stmt); | |
482 | if (gimple_code (stmt) == GIMPLE_PHI | |
483 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
484 | && gimple_code (def_stmt) != GIMPLE_PHI | |
485 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
486 | && bb->loop_father == def_bb->loop_father) | |
487 | { | |
73fbfcad | 488 | if (dump_enabled_p ()) |
78c60e3d | 489 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 490 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
491 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
492 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
493 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 494 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 495 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
496 | return true; |
497 | } | |
498 | ||
499 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
500 | outer-loop-header-bb: | |
501 | d = def_stmt | |
502 | inner-loop: | |
503 | stmt # use (d) | |
504 | outer-loop-tail-bb: | |
505 | ... */ | |
506 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
507 | { | |
73fbfcad | 508 | if (dump_enabled_p ()) |
78c60e3d | 509 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 510 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 511 | |
ebfd146a IR |
512 | switch (relevant) |
513 | { | |
8644a673 | 514 | case vect_unused_in_scope: |
7c5222ff IR |
515 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
516 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 517 | break; |
7c5222ff | 518 | |
ebfd146a | 519 | case vect_used_in_outer_by_reduction: |
7c5222ff | 520 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
521 | relevant = vect_used_by_reduction; |
522 | break; | |
7c5222ff | 523 | |
ebfd146a | 524 | case vect_used_in_outer: |
7c5222ff | 525 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 526 | relevant = vect_used_in_scope; |
ebfd146a | 527 | break; |
7c5222ff | 528 | |
8644a673 | 529 | case vect_used_in_scope: |
ebfd146a IR |
530 | break; |
531 | ||
532 | default: | |
533 | gcc_unreachable (); | |
b8698a0f | 534 | } |
ebfd146a IR |
535 | } |
536 | ||
537 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
538 | outer-loop-header-bb: | |
539 | ... | |
540 | inner-loop: | |
541 | d = def_stmt | |
06066f92 | 542 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
543 | stmt # use (d) */ |
544 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
545 | { | |
73fbfcad | 546 | if (dump_enabled_p ()) |
78c60e3d | 547 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 548 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 549 | |
ebfd146a IR |
550 | switch (relevant) |
551 | { | |
8644a673 | 552 | case vect_unused_in_scope: |
b8698a0f | 553 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 554 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 555 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
556 | break; |
557 | ||
ebfd146a IR |
558 | case vect_used_by_reduction: |
559 | relevant = vect_used_in_outer_by_reduction; | |
560 | break; | |
561 | ||
8644a673 | 562 | case vect_used_in_scope: |
ebfd146a IR |
563 | relevant = vect_used_in_outer; |
564 | break; | |
565 | ||
566 | default: | |
567 | gcc_unreachable (); | |
568 | } | |
569 | } | |
570 | ||
83197f37 IR |
571 | vect_mark_relevant (worklist, def_stmt, relevant, live_p, |
572 | is_pattern_stmt_p (stmt_vinfo)); | |
ebfd146a IR |
573 | return true; |
574 | } | |
575 | ||
576 | ||
577 | /* Function vect_mark_stmts_to_be_vectorized. | |
578 | ||
579 | Not all stmts in the loop need to be vectorized. For example: | |
580 | ||
581 | for i... | |
582 | for j... | |
583 | 1. T0 = i + j | |
584 | 2. T1 = a[T0] | |
585 | ||
586 | 3. j = j + 1 | |
587 | ||
588 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
589 | addressing of vectorized data-refs are handled differently. | |
590 | ||
591 | This pass detects such stmts. */ | |
592 | ||
593 | bool | |
594 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
595 | { | |
ebfd146a IR |
596 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
597 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
598 | unsigned int nbbs = loop->num_nodes; | |
599 | gimple_stmt_iterator si; | |
355fe088 | 600 | gimple *stmt; |
ebfd146a IR |
601 | unsigned int i; |
602 | stmt_vec_info stmt_vinfo; | |
603 | basic_block bb; | |
355fe088 | 604 | gimple *phi; |
ebfd146a | 605 | bool live_p; |
06066f92 IR |
606 | enum vect_relevant relevant, tmp_relevant; |
607 | enum vect_def_type def_type; | |
ebfd146a | 608 | |
73fbfcad | 609 | if (dump_enabled_p ()) |
78c60e3d | 610 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 611 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 612 | |
355fe088 | 613 | auto_vec<gimple *, 64> worklist; |
ebfd146a IR |
614 | |
615 | /* 1. Init worklist. */ | |
616 | for (i = 0; i < nbbs; i++) | |
617 | { | |
618 | bb = bbs[i]; | |
619 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 620 | { |
ebfd146a | 621 | phi = gsi_stmt (si); |
73fbfcad | 622 | if (dump_enabled_p ()) |
ebfd146a | 623 | { |
78c60e3d SS |
624 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
625 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
626 | } |
627 | ||
628 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 629 | vect_mark_relevant (&worklist, phi, relevant, live_p, false); |
ebfd146a IR |
630 | } |
631 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
632 | { | |
633 | stmt = gsi_stmt (si); | |
73fbfcad | 634 | if (dump_enabled_p ()) |
ebfd146a | 635 | { |
78c60e3d SS |
636 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
637 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 638 | } |
ebfd146a IR |
639 | |
640 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 641 | vect_mark_relevant (&worklist, stmt, relevant, live_p, false); |
ebfd146a IR |
642 | } |
643 | } | |
644 | ||
645 | /* 2. Process_worklist */ | |
9771b263 | 646 | while (worklist.length () > 0) |
ebfd146a IR |
647 | { |
648 | use_operand_p use_p; | |
649 | ssa_op_iter iter; | |
650 | ||
9771b263 | 651 | stmt = worklist.pop (); |
73fbfcad | 652 | if (dump_enabled_p ()) |
ebfd146a | 653 | { |
78c60e3d SS |
654 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
655 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
656 | } |
657 | ||
b8698a0f L |
658 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
659 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
660 | liveness and relevance properties of STMT. */ |
661 | stmt_vinfo = vinfo_for_stmt (stmt); | |
662 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
663 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
664 | ||
665 | /* Generally, the liveness and relevance properties of STMT are | |
666 | propagated as is to the DEF_STMTs of its USEs: | |
667 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
668 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
669 | ||
670 | One exception is when STMT has been identified as defining a reduction | |
671 | variable; in this case we set the liveness/relevance as follows: | |
672 | live_p = false | |
673 | relevant = vect_used_by_reduction | |
674 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f | 675 | those that are used by a reduction computation, and those that are |
ff802fa1 | 676 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 677 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 678 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 679 | |
06066f92 IR |
680 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
681 | tmp_relevant = relevant; | |
682 | switch (def_type) | |
ebfd146a | 683 | { |
06066f92 IR |
684 | case vect_reduction_def: |
685 | switch (tmp_relevant) | |
686 | { | |
687 | case vect_unused_in_scope: | |
688 | relevant = vect_used_by_reduction; | |
689 | break; | |
690 | ||
691 | case vect_used_by_reduction: | |
692 | if (gimple_code (stmt) == GIMPLE_PHI) | |
693 | break; | |
694 | /* fall through */ | |
695 | ||
696 | default: | |
73fbfcad | 697 | if (dump_enabled_p ()) |
78c60e3d | 698 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 699 | "unsupported use of reduction.\n"); |
06066f92 IR |
700 | return false; |
701 | } | |
702 | ||
b8698a0f | 703 | live_p = false; |
06066f92 | 704 | break; |
b8698a0f | 705 | |
06066f92 IR |
706 | case vect_nested_cycle: |
707 | if (tmp_relevant != vect_unused_in_scope | |
708 | && tmp_relevant != vect_used_in_outer_by_reduction | |
709 | && tmp_relevant != vect_used_in_outer) | |
710 | { | |
73fbfcad | 711 | if (dump_enabled_p ()) |
78c60e3d | 712 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 713 | "unsupported use of nested cycle.\n"); |
7c5222ff | 714 | |
06066f92 IR |
715 | return false; |
716 | } | |
7c5222ff | 717 | |
b8698a0f L |
718 | live_p = false; |
719 | break; | |
720 | ||
06066f92 IR |
721 | case vect_double_reduction_def: |
722 | if (tmp_relevant != vect_unused_in_scope | |
723 | && tmp_relevant != vect_used_by_reduction) | |
724 | { | |
73fbfcad | 725 | if (dump_enabled_p ()) |
78c60e3d | 726 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 727 | "unsupported use of double reduction.\n"); |
7c5222ff | 728 | |
7c5222ff | 729 | return false; |
06066f92 IR |
730 | } |
731 | ||
732 | live_p = false; | |
b8698a0f | 733 | break; |
7c5222ff | 734 | |
06066f92 IR |
735 | default: |
736 | break; | |
7c5222ff | 737 | } |
b8698a0f | 738 | |
aec7ae7d | 739 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
740 | { |
741 | /* Pattern statements are not inserted into the code, so | |
742 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
743 | have to scan the RHS or function arguments instead. */ | |
744 | if (is_gimple_assign (stmt)) | |
745 | { | |
69d2aade JJ |
746 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
747 | tree op = gimple_assign_rhs1 (stmt); | |
748 | ||
749 | i = 1; | |
750 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
751 | { | |
752 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
aec7ae7d | 753 | live_p, relevant, &worklist, false) |
69d2aade | 754 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
aec7ae7d | 755 | live_p, relevant, &worklist, false)) |
566d377a | 756 | return false; |
69d2aade JJ |
757 | i = 2; |
758 | } | |
759 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 760 | { |
69d2aade | 761 | op = gimple_op (stmt, i); |
afbe6325 RB |
762 | if (TREE_CODE (op) == SSA_NAME |
763 | && !process_use (stmt, op, loop_vinfo, live_p, relevant, | |
764 | &worklist, false)) | |
07687835 | 765 | return false; |
9d5e7640 IR |
766 | } |
767 | } | |
768 | else if (is_gimple_call (stmt)) | |
769 | { | |
770 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
771 | { | |
772 | tree arg = gimple_call_arg (stmt, i); | |
773 | if (!process_use (stmt, arg, loop_vinfo, live_p, relevant, | |
aec7ae7d | 774 | &worklist, false)) |
07687835 | 775 | return false; |
9d5e7640 IR |
776 | } |
777 | } | |
778 | } | |
779 | else | |
780 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
781 | { | |
782 | tree op = USE_FROM_PTR (use_p); | |
783 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, | |
aec7ae7d | 784 | &worklist, false)) |
07687835 | 785 | return false; |
9d5e7640 | 786 | } |
aec7ae7d | 787 | |
3bab6342 | 788 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d JJ |
789 | { |
790 | tree off; | |
3bab6342 | 791 | tree decl = vect_check_gather_scatter (stmt, loop_vinfo, NULL, &off, NULL); |
aec7ae7d JJ |
792 | gcc_assert (decl); |
793 | if (!process_use (stmt, off, loop_vinfo, live_p, relevant, | |
794 | &worklist, true)) | |
566d377a | 795 | return false; |
aec7ae7d | 796 | } |
ebfd146a IR |
797 | } /* while worklist */ |
798 | ||
ebfd146a IR |
799 | return true; |
800 | } | |
801 | ||
802 | ||
b8698a0f | 803 | /* Function vect_model_simple_cost. |
ebfd146a | 804 | |
b8698a0f | 805 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
806 | single op. Right now, this does not account for multiple insns that could |
807 | be generated for the single vector op. We will handle that shortly. */ | |
808 | ||
809 | void | |
b8698a0f | 810 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 BS |
811 | enum vect_def_type *dt, |
812 | stmt_vector_for_cost *prologue_cost_vec, | |
813 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
814 | { |
815 | int i; | |
92345349 | 816 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
817 | |
818 | /* The SLP costs were already calculated during SLP tree build. */ | |
819 | if (PURE_SLP_STMT (stmt_info)) | |
820 | return; | |
821 | ||
ebfd146a IR |
822 | /* FORNOW: Assuming maximum 2 args per stmts. */ |
823 | for (i = 0; i < 2; i++) | |
92345349 BS |
824 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
825 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, vector_stmt, | |
826 | stmt_info, 0, vect_prologue); | |
c3e7ee41 BS |
827 | |
828 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
829 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
830 | stmt_info, 0, vect_body); | |
c3e7ee41 | 831 | |
73fbfcad | 832 | if (dump_enabled_p ()) |
78c60e3d SS |
833 | dump_printf_loc (MSG_NOTE, vect_location, |
834 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 835 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
836 | } |
837 | ||
838 | ||
8bd37302 BS |
839 | /* Model cost for type demotion and promotion operations. PWR is normally |
840 | zero for single-step promotions and demotions. It will be one if | |
841 | two-step promotion/demotion is required, and so on. Each additional | |
842 | step doubles the number of instructions required. */ | |
843 | ||
844 | static void | |
845 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
846 | enum vect_def_type *dt, int pwr) | |
847 | { | |
848 | int i, tmp; | |
92345349 | 849 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
850 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
851 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
852 | void *target_cost_data; | |
8bd37302 BS |
853 | |
854 | /* The SLP costs were already calculated during SLP tree build. */ | |
855 | if (PURE_SLP_STMT (stmt_info)) | |
856 | return; | |
857 | ||
c3e7ee41 BS |
858 | if (loop_vinfo) |
859 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
860 | else | |
861 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
862 | ||
8bd37302 BS |
863 | for (i = 0; i < pwr + 1; i++) |
864 | { | |
865 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
866 | (i + 1) : i; | |
c3e7ee41 | 867 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
868 | vec_promote_demote, stmt_info, 0, |
869 | vect_body); | |
8bd37302 BS |
870 | } |
871 | ||
872 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
873 | for (i = 0; i < 2; i++) | |
92345349 BS |
874 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
875 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
876 | stmt_info, 0, vect_prologue); | |
8bd37302 | 877 | |
73fbfcad | 878 | if (dump_enabled_p ()) |
78c60e3d SS |
879 | dump_printf_loc (MSG_NOTE, vect_location, |
880 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 881 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
882 | } |
883 | ||
0d0293ac | 884 | /* Function vect_cost_group_size |
b8698a0f | 885 | |
0d0293ac | 886 | For grouped load or store, return the group_size only if it is the first |
ebfd146a IR |
887 | load or store of a group, else return 1. This ensures that group size is |
888 | only returned once per group. */ | |
889 | ||
890 | static int | |
0d0293ac | 891 | vect_cost_group_size (stmt_vec_info stmt_info) |
ebfd146a | 892 | { |
355fe088 | 893 | gimple *first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
894 | |
895 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
e14c1050 | 896 | return GROUP_SIZE (stmt_info); |
ebfd146a IR |
897 | |
898 | return 1; | |
899 | } | |
900 | ||
901 | ||
902 | /* Function vect_model_store_cost | |
903 | ||
0d0293ac MM |
904 | Models cost for stores. In the case of grouped accesses, one access |
905 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
906 | |
907 | void | |
b8698a0f | 908 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
272c6793 | 909 | bool store_lanes_p, enum vect_def_type dt, |
92345349 BS |
910 | slp_tree slp_node, |
911 | stmt_vector_for_cost *prologue_cost_vec, | |
912 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
913 | { |
914 | int group_size; | |
92345349 | 915 | unsigned int inside_cost = 0, prologue_cost = 0; |
720f5239 | 916 | struct data_reference *first_dr; |
355fe088 | 917 | gimple *first_stmt; |
ebfd146a | 918 | |
8644a673 | 919 | if (dt == vect_constant_def || dt == vect_external_def) |
92345349 BS |
920 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
921 | stmt_info, 0, vect_prologue); | |
ebfd146a | 922 | |
0d0293ac MM |
923 | /* Grouped access? */ |
924 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
720f5239 IR |
925 | { |
926 | if (slp_node) | |
927 | { | |
9771b263 | 928 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
720f5239 IR |
929 | group_size = 1; |
930 | } | |
931 | else | |
932 | { | |
e14c1050 | 933 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 934 | group_size = vect_cost_group_size (stmt_info); |
720f5239 IR |
935 | } |
936 | ||
937 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
938 | } | |
0d0293ac | 939 | /* Not a grouped access. */ |
ebfd146a | 940 | else |
720f5239 IR |
941 | { |
942 | group_size = 1; | |
943 | first_dr = STMT_VINFO_DATA_REF (stmt_info); | |
944 | } | |
ebfd146a | 945 | |
272c6793 | 946 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 947 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 RS |
948 | access is instead being provided by a permute-and-store operation, |
949 | include the cost of the permutes. */ | |
cee62fee MM |
950 | if (!store_lanes_p && group_size > 1 |
951 | && !STMT_VINFO_STRIDED_P (stmt_info)) | |
ebfd146a | 952 | { |
e1377713 ES |
953 | /* Uses a high and low interleave or shuffle operations for each |
954 | needed permute. */ | |
955 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; | |
92345349 BS |
956 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
957 | stmt_info, 0, vect_body); | |
ebfd146a | 958 | |
73fbfcad | 959 | if (dump_enabled_p ()) |
78c60e3d | 960 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 961 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 962 | group_size); |
ebfd146a IR |
963 | } |
964 | ||
cee62fee | 965 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 966 | /* Costs of the stores. */ |
cee62fee MM |
967 | if (STMT_VINFO_STRIDED_P (stmt_info) |
968 | && !STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
f2e2a985 MM |
969 | { |
970 | /* N scalar stores plus extracting the elements. */ | |
f2e2a985 MM |
971 | inside_cost += record_stmt_cost (body_cost_vec, |
972 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
973 | scalar_store, stmt_info, 0, vect_body); | |
f2e2a985 MM |
974 | } |
975 | else | |
976 | vect_get_store_cost (first_dr, ncopies, &inside_cost, body_cost_vec); | |
ebfd146a | 977 | |
cee62fee MM |
978 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
979 | inside_cost += record_stmt_cost (body_cost_vec, | |
980 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
981 | vec_to_scalar, stmt_info, 0, vect_body); | |
982 | ||
73fbfcad | 983 | if (dump_enabled_p ()) |
78c60e3d SS |
984 | dump_printf_loc (MSG_NOTE, vect_location, |
985 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 986 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
987 | } |
988 | ||
989 | ||
720f5239 IR |
990 | /* Calculate cost of DR's memory access. */ |
991 | void | |
992 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 993 | unsigned int *inside_cost, |
92345349 | 994 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
995 | { |
996 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 997 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 998 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
999 | |
1000 | switch (alignment_support_scheme) | |
1001 | { | |
1002 | case dr_aligned: | |
1003 | { | |
92345349 BS |
1004 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
1005 | vector_store, stmt_info, 0, | |
1006 | vect_body); | |
720f5239 | 1007 | |
73fbfcad | 1008 | if (dump_enabled_p ()) |
78c60e3d | 1009 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1010 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
1011 | break; |
1012 | } | |
1013 | ||
1014 | case dr_unaligned_supported: | |
1015 | { | |
720f5239 | 1016 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1017 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1018 | unaligned_store, stmt_info, |
92345349 | 1019 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1020 | if (dump_enabled_p ()) |
78c60e3d SS |
1021 | dump_printf_loc (MSG_NOTE, vect_location, |
1022 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1023 | "hardware.\n"); |
720f5239 IR |
1024 | break; |
1025 | } | |
1026 | ||
38eec4c6 UW |
1027 | case dr_unaligned_unsupported: |
1028 | { | |
1029 | *inside_cost = VECT_MAX_COST; | |
1030 | ||
73fbfcad | 1031 | if (dump_enabled_p ()) |
78c60e3d | 1032 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1033 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1034 | break; |
1035 | } | |
1036 | ||
720f5239 IR |
1037 | default: |
1038 | gcc_unreachable (); | |
1039 | } | |
1040 | } | |
1041 | ||
1042 | ||
ebfd146a IR |
1043 | /* Function vect_model_load_cost |
1044 | ||
0d0293ac MM |
1045 | Models cost for loads. In the case of grouped accesses, the last access |
1046 | has the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1047 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1048 | access scheme chosen. */ |
1049 | ||
1050 | void | |
92345349 BS |
1051 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
1052 | bool load_lanes_p, slp_tree slp_node, | |
1053 | stmt_vector_for_cost *prologue_cost_vec, | |
1054 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
1055 | { |
1056 | int group_size; | |
355fe088 | 1057 | gimple *first_stmt; |
ebfd146a | 1058 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; |
92345349 | 1059 | unsigned int inside_cost = 0, prologue_cost = 0; |
ebfd146a | 1060 | |
0d0293ac | 1061 | /* Grouped accesses? */ |
e14c1050 | 1062 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 1063 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && first_stmt && !slp_node) |
ebfd146a | 1064 | { |
0d0293ac | 1065 | group_size = vect_cost_group_size (stmt_info); |
ebfd146a IR |
1066 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
1067 | } | |
0d0293ac | 1068 | /* Not a grouped access. */ |
ebfd146a IR |
1069 | else |
1070 | { | |
1071 | group_size = 1; | |
1072 | first_dr = dr; | |
1073 | } | |
1074 | ||
272c6793 | 1075 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1076 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 RS |
1077 | access is instead being provided by a load-and-permute operation, |
1078 | include the cost of the permutes. */ | |
7b5fc413 | 1079 | if (!load_lanes_p && group_size > 1 |
f2e2a985 | 1080 | && !STMT_VINFO_STRIDED_P (stmt_info)) |
ebfd146a | 1081 | { |
2c23db6d ES |
1082 | /* Uses an even and odd extract operations or shuffle operations |
1083 | for each needed permute. */ | |
1084 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; | |
1085 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, | |
1086 | stmt_info, 0, vect_body); | |
ebfd146a | 1087 | |
73fbfcad | 1088 | if (dump_enabled_p ()) |
e645e942 TJ |
1089 | dump_printf_loc (MSG_NOTE, vect_location, |
1090 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1091 | group_size); |
ebfd146a IR |
1092 | } |
1093 | ||
1094 | /* The loads themselves. */ | |
f2e2a985 | 1095 | if (STMT_VINFO_STRIDED_P (stmt_info) |
7b5fc413 | 1096 | && !STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
a82960aa | 1097 | { |
a21892ad BS |
1098 | /* N scalar loads plus gathering them into a vector. */ |
1099 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
92345349 | 1100 | inside_cost += record_stmt_cost (body_cost_vec, |
c3e7ee41 | 1101 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), |
92345349 | 1102 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1103 | } |
1104 | else | |
1105 | vect_get_load_cost (first_dr, ncopies, | |
1106 | ((!STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1107 | || group_size > 1 || slp_node), | |
92345349 BS |
1108 | &inside_cost, &prologue_cost, |
1109 | prologue_cost_vec, body_cost_vec, true); | |
f2e2a985 | 1110 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
7b5fc413 RB |
1111 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, |
1112 | stmt_info, 0, vect_body); | |
720f5239 | 1113 | |
73fbfcad | 1114 | if (dump_enabled_p ()) |
78c60e3d SS |
1115 | dump_printf_loc (MSG_NOTE, vect_location, |
1116 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1117 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1118 | } |
1119 | ||
1120 | ||
1121 | /* Calculate cost of DR's memory access. */ | |
1122 | void | |
1123 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1124 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1125 | unsigned int *prologue_cost, |
1126 | stmt_vector_for_cost *prologue_cost_vec, | |
1127 | stmt_vector_for_cost *body_cost_vec, | |
1128 | bool record_prologue_costs) | |
720f5239 IR |
1129 | { |
1130 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 1131 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 1132 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
1133 | |
1134 | switch (alignment_support_scheme) | |
ebfd146a IR |
1135 | { |
1136 | case dr_aligned: | |
1137 | { | |
92345349 BS |
1138 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1139 | stmt_info, 0, vect_body); | |
ebfd146a | 1140 | |
73fbfcad | 1141 | if (dump_enabled_p ()) |
78c60e3d | 1142 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1143 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1144 | |
1145 | break; | |
1146 | } | |
1147 | case dr_unaligned_supported: | |
1148 | { | |
720f5239 | 1149 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1150 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1151 | unaligned_load, stmt_info, |
92345349 | 1152 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1153 | |
73fbfcad | 1154 | if (dump_enabled_p ()) |
78c60e3d SS |
1155 | dump_printf_loc (MSG_NOTE, vect_location, |
1156 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1157 | "hardware.\n"); |
ebfd146a IR |
1158 | |
1159 | break; | |
1160 | } | |
1161 | case dr_explicit_realign: | |
1162 | { | |
92345349 BS |
1163 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1164 | vector_load, stmt_info, 0, vect_body); | |
1165 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1166 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1167 | |
1168 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1169 | the containing loop, the following cost should be added to the | |
92345349 | 1170 | prologue costs. */ |
ebfd146a | 1171 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1172 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1173 | stmt_info, 0, vect_body); | |
ebfd146a | 1174 | |
73fbfcad | 1175 | if (dump_enabled_p ()) |
e645e942 TJ |
1176 | dump_printf_loc (MSG_NOTE, vect_location, |
1177 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1178 | |
ebfd146a IR |
1179 | break; |
1180 | } | |
1181 | case dr_explicit_realign_optimized: | |
1182 | { | |
73fbfcad | 1183 | if (dump_enabled_p ()) |
e645e942 | 1184 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1185 | "vect_model_load_cost: unaligned software " |
e645e942 | 1186 | "pipelined.\n"); |
ebfd146a IR |
1187 | |
1188 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1189 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1190 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1191 | access, then the above cost should only be considered for one |
ff802fa1 | 1192 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1193 | and a realignment op. */ |
1194 | ||
92345349 | 1195 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1196 | { |
92345349 BS |
1197 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1198 | vector_stmt, stmt_info, | |
1199 | 0, vect_prologue); | |
ebfd146a | 1200 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1201 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1202 | vector_stmt, stmt_info, | |
1203 | 0, vect_prologue); | |
ebfd146a IR |
1204 | } |
1205 | ||
92345349 BS |
1206 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1207 | stmt_info, 0, vect_body); | |
1208 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1209 | stmt_info, 0, vect_body); | |
8bd37302 | 1210 | |
73fbfcad | 1211 | if (dump_enabled_p ()) |
78c60e3d | 1212 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1213 | "vect_model_load_cost: explicit realign optimized" |
1214 | "\n"); | |
8bd37302 | 1215 | |
ebfd146a IR |
1216 | break; |
1217 | } | |
1218 | ||
38eec4c6 UW |
1219 | case dr_unaligned_unsupported: |
1220 | { | |
1221 | *inside_cost = VECT_MAX_COST; | |
1222 | ||
73fbfcad | 1223 | if (dump_enabled_p ()) |
78c60e3d | 1224 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1225 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1226 | break; |
1227 | } | |
1228 | ||
ebfd146a IR |
1229 | default: |
1230 | gcc_unreachable (); | |
1231 | } | |
ebfd146a IR |
1232 | } |
1233 | ||
418b7df3 RG |
1234 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1235 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1236 | |
418b7df3 | 1237 | static void |
355fe088 | 1238 | vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi) |
ebfd146a | 1239 | { |
ebfd146a | 1240 | if (gsi) |
418b7df3 | 1241 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1242 | else |
1243 | { | |
418b7df3 | 1244 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1245 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1246 | |
a70d6342 IR |
1247 | if (loop_vinfo) |
1248 | { | |
1249 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1250 | basic_block new_bb; |
1251 | edge pe; | |
a70d6342 IR |
1252 | |
1253 | if (nested_in_vect_loop_p (loop, stmt)) | |
1254 | loop = loop->inner; | |
b8698a0f | 1255 | |
a70d6342 | 1256 | pe = loop_preheader_edge (loop); |
418b7df3 | 1257 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1258 | gcc_assert (!new_bb); |
1259 | } | |
1260 | else | |
1261 | { | |
1262 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1263 | basic_block bb; | |
1264 | gimple_stmt_iterator gsi_bb_start; | |
1265 | ||
1266 | gcc_assert (bb_vinfo); | |
1267 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1268 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1269 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1270 | } |
ebfd146a IR |
1271 | } |
1272 | ||
73fbfcad | 1273 | if (dump_enabled_p ()) |
ebfd146a | 1274 | { |
78c60e3d SS |
1275 | dump_printf_loc (MSG_NOTE, vect_location, |
1276 | "created new init_stmt: "); | |
1277 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1278 | } |
418b7df3 RG |
1279 | } |
1280 | ||
1281 | /* Function vect_init_vector. | |
ebfd146a | 1282 | |
5467ee52 RG |
1283 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1284 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1285 | vector type a vector with all elements equal to VAL is created first. | |
1286 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1287 | initialization at the loop preheader. | |
418b7df3 RG |
1288 | Return the DEF of INIT_STMT. |
1289 | It will be used in the vectorization of STMT. */ | |
1290 | ||
1291 | tree | |
355fe088 | 1292 | vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 | 1293 | { |
355fe088 | 1294 | gimple *init_stmt; |
418b7df3 RG |
1295 | tree new_temp; |
1296 | ||
5467ee52 RG |
1297 | if (TREE_CODE (type) == VECTOR_TYPE |
1298 | && TREE_CODE (TREE_TYPE (val)) != VECTOR_TYPE) | |
418b7df3 | 1299 | { |
5467ee52 | 1300 | if (!types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) |
418b7df3 | 1301 | { |
5467ee52 | 1302 | if (CONSTANT_CLASS_P (val)) |
42fd8198 | 1303 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1304 | else |
1305 | { | |
b731b390 | 1306 | new_temp = make_ssa_name (TREE_TYPE (type)); |
0d0e4a03 | 1307 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); |
418b7df3 | 1308 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1309 | val = new_temp; |
418b7df3 RG |
1310 | } |
1311 | } | |
5467ee52 | 1312 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1313 | } |
1314 | ||
0e22bb5a RB |
1315 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1316 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1317 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1318 | return new_temp; |
ebfd146a IR |
1319 | } |
1320 | ||
a70d6342 | 1321 | |
ebfd146a IR |
1322 | /* Function vect_get_vec_def_for_operand. |
1323 | ||
ff802fa1 | 1324 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
1325 | used in the vectorized stmt for STMT. |
1326 | ||
1327 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1328 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1329 | ||
1330 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
42fd8198 IE |
1331 | needs to be introduced. VECTYPE may be used to specify a required type for |
1332 | vector invariant. */ | |
ebfd146a IR |
1333 | |
1334 | tree | |
42fd8198 | 1335 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) |
ebfd146a IR |
1336 | { |
1337 | tree vec_oprnd; | |
355fe088 TS |
1338 | gimple *vec_stmt; |
1339 | gimple *def_stmt; | |
ebfd146a IR |
1340 | stmt_vec_info def_stmt_info = NULL; |
1341 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
42fd8198 | 1342 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); |
ebfd146a | 1343 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
ebfd146a IR |
1344 | enum vect_def_type dt; |
1345 | bool is_simple_use; | |
1346 | tree vector_type; | |
1347 | ||
73fbfcad | 1348 | if (dump_enabled_p ()) |
ebfd146a | 1349 | { |
78c60e3d SS |
1350 | dump_printf_loc (MSG_NOTE, vect_location, |
1351 | "vect_get_vec_def_for_operand: "); | |
1352 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
e645e942 | 1353 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1354 | } |
1355 | ||
81c40241 | 1356 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt); |
ebfd146a | 1357 | gcc_assert (is_simple_use); |
73fbfcad | 1358 | if (dump_enabled_p ()) |
ebfd146a | 1359 | { |
78c60e3d | 1360 | int loc_printed = 0; |
ebfd146a IR |
1361 | if (def_stmt) |
1362 | { | |
78c60e3d SS |
1363 | if (loc_printed) |
1364 | dump_printf (MSG_NOTE, " def_stmt = "); | |
1365 | else | |
1366 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1367 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
ebfd146a IR |
1368 | } |
1369 | } | |
1370 | ||
1371 | switch (dt) | |
1372 | { | |
81c40241 | 1373 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1374 | case vect_constant_def: |
81c40241 | 1375 | case vect_external_def: |
ebfd146a | 1376 | { |
42fd8198 IE |
1377 | if (vectype) |
1378 | vector_type = vectype; | |
1379 | else if (TREE_CODE (TREE_TYPE (op)) == BOOLEAN_TYPE | |
1380 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) | |
1381 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1382 | else | |
1383 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1384 | ||
7569a6cc | 1385 | gcc_assert (vector_type); |
418b7df3 | 1386 | return vect_init_vector (stmt, op, vector_type, NULL); |
ebfd146a IR |
1387 | } |
1388 | ||
81c40241 | 1389 | /* operand is defined inside the loop. */ |
8644a673 | 1390 | case vect_internal_def: |
ebfd146a | 1391 | { |
ebfd146a IR |
1392 | /* Get the def from the vectorized stmt. */ |
1393 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1394 | |
ebfd146a | 1395 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1396 | /* Get vectorized pattern statement. */ |
1397 | if (!vec_stmt | |
1398 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1399 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1400 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1401 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1402 | gcc_assert (vec_stmt); |
1403 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1404 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1405 | else if (is_gimple_call (vec_stmt)) | |
1406 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1407 | else | |
1408 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1409 | return vec_oprnd; | |
1410 | } | |
1411 | ||
81c40241 | 1412 | /* operand is defined by a loop header phi - reduction */ |
ebfd146a | 1413 | case vect_reduction_def: |
06066f92 | 1414 | case vect_double_reduction_def: |
7c5222ff | 1415 | case vect_nested_cycle: |
81c40241 RB |
1416 | /* Code should use get_initial_def_for_reduction. */ |
1417 | gcc_unreachable (); | |
ebfd146a | 1418 | |
81c40241 | 1419 | /* operand is defined by loop-header phi - induction. */ |
ebfd146a IR |
1420 | case vect_induction_def: |
1421 | { | |
1422 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1423 | ||
1424 | /* Get the def from the vectorized stmt. */ | |
1425 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1426 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1427 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1428 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1429 | else | |
1430 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1431 | return vec_oprnd; |
1432 | } | |
1433 | ||
1434 | default: | |
1435 | gcc_unreachable (); | |
1436 | } | |
1437 | } | |
1438 | ||
1439 | ||
1440 | /* Function vect_get_vec_def_for_stmt_copy | |
1441 | ||
ff802fa1 | 1442 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1443 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1444 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1445 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1446 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1447 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1448 | DT is the type of the vector def VEC_OPRND. |
1449 | ||
1450 | Context: | |
1451 | In case the vectorization factor (VF) is bigger than the number | |
1452 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1453 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1454 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1455 | smallest data-type determines the VF, and as a result, when vectorizing |
1456 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1457 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1458 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1459 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1460 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1461 | ||
1462 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1463 | |
ebfd146a IR |
1464 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1465 | VS1.1: vx.1 = memref1 VS1.2 | |
1466 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1467 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1468 | |
1469 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1470 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1471 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1472 | VSnew.3: vz3 = vx.3 + ... | |
1473 | ||
1474 | The vectorization of S1 is explained in vectorizable_load. | |
1475 | The vectorization of S2: | |
b8698a0f L |
1476 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1477 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1478 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1479 | returns the vector-def 'vx.0'. |
1480 | ||
b8698a0f L |
1481 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1482 | function is called to get the relevant vector-def for each operand. It is | |
1483 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1484 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1485 | ||
b8698a0f L |
1486 | For example, to obtain the vector-def 'vx.1' in order to create the |
1487 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1488 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1489 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1490 | and return its def ('vx.1'). | |
1491 | Overall, to create the above sequence this function will be called 3 times: | |
1492 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1493 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1494 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1495 | ||
1496 | tree | |
1497 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1498 | { | |
355fe088 | 1499 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1500 | stmt_vec_info def_stmt_info; |
1501 | ||
1502 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1503 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1504 | return vec_oprnd; |
1505 | ||
1506 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1507 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1508 | gcc_assert (def_stmt_info); | |
1509 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1510 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1511 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1512 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1513 | else | |
1514 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1515 | return vec_oprnd; | |
1516 | } | |
1517 | ||
1518 | ||
1519 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1520 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1521 | |
1522 | static void | |
b8698a0f | 1523 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1524 | vec<tree> *vec_oprnds0, |
1525 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1526 | { |
9771b263 | 1527 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1528 | |
1529 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1530 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1531 | |
9771b263 | 1532 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1533 | { |
9771b263 | 1534 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1535 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1536 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1537 | } |
1538 | } | |
1539 | ||
1540 | ||
d092494c IR |
1541 | /* Get vectorized definitions for OP0 and OP1. |
1542 | REDUC_INDEX is the index of reduction operand in case of reduction, | |
1543 | and -1 otherwise. */ | |
ebfd146a | 1544 | |
d092494c | 1545 | void |
355fe088 | 1546 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1547 | vec<tree> *vec_oprnds0, |
1548 | vec<tree> *vec_oprnds1, | |
d092494c | 1549 | slp_tree slp_node, int reduc_index) |
ebfd146a IR |
1550 | { |
1551 | if (slp_node) | |
d092494c IR |
1552 | { |
1553 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1554 | auto_vec<tree> ops (nops); |
1555 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1556 | |
9771b263 | 1557 | ops.quick_push (op0); |
d092494c | 1558 | if (op1) |
9771b263 | 1559 | ops.quick_push (op1); |
d092494c IR |
1560 | |
1561 | vect_get_slp_defs (ops, slp_node, &vec_defs, reduc_index); | |
1562 | ||
37b5ec8f | 1563 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1564 | if (op1) |
37b5ec8f | 1565 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1566 | } |
ebfd146a IR |
1567 | else |
1568 | { | |
1569 | tree vec_oprnd; | |
1570 | ||
9771b263 | 1571 | vec_oprnds0->create (1); |
81c40241 | 1572 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1573 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1574 | |
1575 | if (op1) | |
1576 | { | |
9771b263 | 1577 | vec_oprnds1->create (1); |
81c40241 | 1578 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1579 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1580 | } |
1581 | } | |
1582 | } | |
1583 | ||
1584 | ||
1585 | /* Function vect_finish_stmt_generation. | |
1586 | ||
1587 | Insert a new stmt. */ | |
1588 | ||
1589 | void | |
355fe088 | 1590 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1591 | gimple_stmt_iterator *gsi) |
1592 | { | |
1593 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
310213d4 | 1594 | vec_info *vinfo = stmt_info->vinfo; |
ebfd146a IR |
1595 | |
1596 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1597 | ||
54e8e2c3 RG |
1598 | if (!gsi_end_p (*gsi) |
1599 | && gimple_has_mem_ops (vec_stmt)) | |
1600 | { | |
355fe088 | 1601 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1602 | tree vuse = gimple_vuse (at_stmt); |
1603 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1604 | { | |
1605 | tree vdef = gimple_vdef (at_stmt); | |
1606 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1607 | /* If we have an SSA vuse and insert a store, update virtual | |
1608 | SSA form to avoid triggering the renamer. Do so only | |
1609 | if we can easily see all uses - which is what almost always | |
1610 | happens with the way vectorized stmts are inserted. */ | |
1611 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1612 | && ((is_gimple_assign (vec_stmt) | |
1613 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1614 | || (is_gimple_call (vec_stmt) | |
1615 | && !(gimple_call_flags (vec_stmt) | |
1616 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1617 | { | |
1618 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1619 | gimple_set_vdef (vec_stmt, new_vdef); | |
1620 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1621 | } | |
1622 | } | |
1623 | } | |
ebfd146a IR |
1624 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1625 | ||
310213d4 | 1626 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, vinfo)); |
ebfd146a | 1627 | |
73fbfcad | 1628 | if (dump_enabled_p ()) |
ebfd146a | 1629 | { |
78c60e3d SS |
1630 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1631 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
ebfd146a IR |
1632 | } |
1633 | ||
ad885386 | 1634 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
8e91d222 JJ |
1635 | |
1636 | /* While EH edges will generally prevent vectorization, stmt might | |
1637 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1638 | that could throw are part of the same region. */ | |
1639 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1640 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1641 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
ebfd146a IR |
1642 | } |
1643 | ||
1644 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1645 | a function declaration if the target has a vectorized version | |
1646 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1647 | ||
1648 | tree | |
538dd0b7 | 1649 | vectorizable_function (gcall *call, tree vectype_out, tree vectype_in) |
ebfd146a | 1650 | { |
10766209 RS |
1651 | /* We only handle functions that do not read or clobber memory. */ |
1652 | if (gimple_vuse (call)) | |
ebfd146a IR |
1653 | return NULL_TREE; |
1654 | ||
10766209 RS |
1655 | combined_fn fn = gimple_call_combined_fn (call); |
1656 | if (fn != CFN_LAST) | |
1657 | return targetm.vectorize.builtin_vectorized_function | |
1658 | (fn, vectype_out, vectype_in); | |
1659 | ||
1660 | if (gimple_call_builtin_p (call, BUILT_IN_MD)) | |
1661 | return targetm.vectorize.builtin_md_vectorized_function | |
1662 | (gimple_call_fndecl (call), vectype_out, vectype_in); | |
ebfd146a | 1663 | |
10766209 | 1664 | return NULL_TREE; |
ebfd146a IR |
1665 | } |
1666 | ||
5ce9450f | 1667 | |
355fe088 | 1668 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1669 | gimple_stmt_iterator *); |
1670 | ||
1671 | ||
1672 | /* Function vectorizable_mask_load_store. | |
1673 | ||
1674 | Check if STMT performs a conditional load or store that can be vectorized. | |
1675 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1676 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
1677 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1678 | ||
1679 | static bool | |
355fe088 TS |
1680 | vectorizable_mask_load_store (gimple *stmt, gimple_stmt_iterator *gsi, |
1681 | gimple **vec_stmt, slp_tree slp_node) | |
5ce9450f JJ |
1682 | { |
1683 | tree vec_dest = NULL; | |
1684 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1685 | stmt_vec_info prev_stmt_info; | |
1686 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1687 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1688 | bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
1689 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
1690 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
57e2f6ad | 1691 | tree rhs_vectype = NULL_TREE; |
045c1278 | 1692 | tree mask_vectype; |
5ce9450f | 1693 | tree elem_type; |
355fe088 | 1694 | gimple *new_stmt; |
5ce9450f JJ |
1695 | tree dummy; |
1696 | tree dataref_ptr = NULL_TREE; | |
355fe088 | 1697 | gimple *ptr_incr; |
5ce9450f JJ |
1698 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
1699 | int ncopies; | |
1700 | int i, j; | |
1701 | bool inv_p; | |
1702 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; | |
1703 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
1704 | int gather_scale = 1; | |
1705 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
1706 | bool is_store; | |
1707 | tree mask; | |
355fe088 | 1708 | gimple *def_stmt; |
5ce9450f JJ |
1709 | enum vect_def_type dt; |
1710 | ||
1711 | if (slp_node != NULL) | |
1712 | return false; | |
1713 | ||
1714 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
1715 | gcc_assert (ncopies >= 1); | |
1716 | ||
1717 | is_store = gimple_call_internal_fn (stmt) == IFN_MASK_STORE; | |
1718 | mask = gimple_call_arg (stmt, 2); | |
045c1278 IE |
1719 | |
1720 | if (TREE_CODE (TREE_TYPE (mask)) != BOOLEAN_TYPE) | |
5ce9450f JJ |
1721 | return false; |
1722 | ||
1723 | /* FORNOW. This restriction should be relaxed. */ | |
1724 | if (nested_in_vect_loop && ncopies > 1) | |
1725 | { | |
1726 | if (dump_enabled_p ()) | |
1727 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1728 | "multiple types in nested loop."); | |
1729 | return false; | |
1730 | } | |
1731 | ||
1732 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
1733 | return false; | |
1734 | ||
1735 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
1736 | return false; | |
1737 | ||
1738 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
1739 | return false; | |
1740 | ||
1741 | elem_type = TREE_TYPE (vectype); | |
1742 | ||
1743 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1744 | return false; | |
1745 | ||
f2e2a985 | 1746 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
5ce9450f JJ |
1747 | return false; |
1748 | ||
045c1278 IE |
1749 | if (TREE_CODE (mask) != SSA_NAME) |
1750 | return false; | |
1751 | ||
1752 | if (!vect_is_simple_use (mask, loop_vinfo, &def_stmt, &dt, &mask_vectype)) | |
1753 | return false; | |
1754 | ||
1755 | if (!mask_vectype) | |
1756 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
1757 | ||
1758 | if (!mask_vectype) | |
1759 | return false; | |
1760 | ||
57e2f6ad IE |
1761 | if (is_store) |
1762 | { | |
1763 | tree rhs = gimple_call_arg (stmt, 3); | |
1764 | if (!vect_is_simple_use (rhs, loop_vinfo, &def_stmt, &dt, &rhs_vectype)) | |
1765 | return false; | |
1766 | } | |
1767 | ||
3bab6342 | 1768 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5ce9450f | 1769 | { |
355fe088 | 1770 | gimple *def_stmt; |
3bab6342 | 1771 | gather_decl = vect_check_gather_scatter (stmt, loop_vinfo, &gather_base, |
5ce9450f JJ |
1772 | &gather_off, &gather_scale); |
1773 | gcc_assert (gather_decl); | |
81c40241 RB |
1774 | if (!vect_is_simple_use (gather_off, loop_vinfo, &def_stmt, &gather_dt, |
1775 | &gather_off_vectype)) | |
5ce9450f JJ |
1776 | { |
1777 | if (dump_enabled_p ()) | |
1778 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1779 | "gather index use not simple."); | |
1780 | return false; | |
1781 | } | |
03b9e8e4 JJ |
1782 | |
1783 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1784 | tree masktype | |
1785 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
1786 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
1787 | { | |
1788 | if (dump_enabled_p ()) | |
1789 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1790 | "masked gather with integer mask not supported."); | |
1791 | return false; | |
1792 | } | |
5ce9450f JJ |
1793 | } |
1794 | else if (tree_int_cst_compare (nested_in_vect_loop | |
1795 | ? STMT_VINFO_DR_STEP (stmt_info) | |
1796 | : DR_STEP (dr), size_zero_node) <= 0) | |
1797 | return false; | |
1798 | else if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
045c1278 IE |
1799 | || !can_vec_mask_load_store_p (TYPE_MODE (vectype), |
1800 | TYPE_MODE (mask_vectype), | |
57e2f6ad IE |
1801 | !is_store) |
1802 | || (rhs_vectype | |
1803 | && !useless_type_conversion_p (vectype, rhs_vectype))) | |
5ce9450f JJ |
1804 | return false; |
1805 | ||
5ce9450f JJ |
1806 | if (!vec_stmt) /* transformation not required. */ |
1807 | { | |
1808 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1809 | if (is_store) | |
1810 | vect_model_store_cost (stmt_info, ncopies, false, dt, | |
1811 | NULL, NULL, NULL); | |
1812 | else | |
1813 | vect_model_load_cost (stmt_info, ncopies, false, NULL, NULL, NULL); | |
1814 | return true; | |
1815 | } | |
1816 | ||
1817 | /** Transform. **/ | |
1818 | ||
3bab6342 | 1819 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5ce9450f JJ |
1820 | { |
1821 | tree vec_oprnd0 = NULL_TREE, op; | |
1822 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1823 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
acdcd61b | 1824 | tree ptr, vec_mask = NULL_TREE, mask_op = NULL_TREE, var, scale; |
5ce9450f | 1825 | tree perm_mask = NULL_TREE, prev_res = NULL_TREE; |
acdcd61b | 1826 | tree mask_perm_mask = NULL_TREE; |
5ce9450f JJ |
1827 | edge pe = loop_preheader_edge (loop); |
1828 | gimple_seq seq; | |
1829 | basic_block new_bb; | |
1830 | enum { NARROW, NONE, WIDEN } modifier; | |
1831 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
1832 | ||
acdcd61b JJ |
1833 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); |
1834 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1835 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1836 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1837 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1838 | scaletype = TREE_VALUE (arglist); | |
1839 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
1840 | && types_compatible_p (srctype, masktype)); | |
1841 | ||
5ce9450f JJ |
1842 | if (nunits == gather_off_nunits) |
1843 | modifier = NONE; | |
1844 | else if (nunits == gather_off_nunits / 2) | |
1845 | { | |
1846 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
1847 | modifier = WIDEN; | |
1848 | ||
1849 | for (i = 0; i < gather_off_nunits; ++i) | |
1850 | sel[i] = i | nunits; | |
1851 | ||
557be5a8 | 1852 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
5ce9450f JJ |
1853 | } |
1854 | else if (nunits == gather_off_nunits * 2) | |
1855 | { | |
1856 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
1857 | modifier = NARROW; | |
1858 | ||
1859 | for (i = 0; i < nunits; ++i) | |
1860 | sel[i] = i < gather_off_nunits | |
1861 | ? i : i + nunits - gather_off_nunits; | |
1862 | ||
557be5a8 | 1863 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
5ce9450f | 1864 | ncopies *= 2; |
acdcd61b JJ |
1865 | for (i = 0; i < nunits; ++i) |
1866 | sel[i] = i | gather_off_nunits; | |
557be5a8 | 1867 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, sel); |
5ce9450f JJ |
1868 | } |
1869 | else | |
1870 | gcc_unreachable (); | |
1871 | ||
5ce9450f JJ |
1872 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); |
1873 | ||
1874 | ptr = fold_convert (ptrtype, gather_base); | |
1875 | if (!is_gimple_min_invariant (ptr)) | |
1876 | { | |
1877 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
1878 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
1879 | gcc_assert (!new_bb); | |
1880 | } | |
1881 | ||
1882 | scale = build_int_cst (scaletype, gather_scale); | |
1883 | ||
1884 | prev_stmt_info = NULL; | |
1885 | for (j = 0; j < ncopies; ++j) | |
1886 | { | |
1887 | if (modifier == WIDEN && (j & 1)) | |
1888 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
1889 | perm_mask, stmt, gsi); | |
1890 | else if (j == 0) | |
1891 | op = vec_oprnd0 | |
81c40241 | 1892 | = vect_get_vec_def_for_operand (gather_off, stmt); |
5ce9450f JJ |
1893 | else |
1894 | op = vec_oprnd0 | |
1895 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
1896 | ||
1897 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
1898 | { | |
1899 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
1900 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 1901 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
5ce9450f JJ |
1902 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
1903 | new_stmt | |
0d0e4a03 | 1904 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
1905 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1906 | op = var; | |
1907 | } | |
1908 | ||
acdcd61b JJ |
1909 | if (mask_perm_mask && (j & 1)) |
1910 | mask_op = permute_vec_elements (mask_op, mask_op, | |
1911 | mask_perm_mask, stmt, gsi); | |
5ce9450f JJ |
1912 | else |
1913 | { | |
acdcd61b | 1914 | if (j == 0) |
81c40241 | 1915 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
acdcd61b JJ |
1916 | else |
1917 | { | |
81c40241 | 1918 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
acdcd61b JJ |
1919 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
1920 | } | |
5ce9450f | 1921 | |
acdcd61b JJ |
1922 | mask_op = vec_mask; |
1923 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
1924 | { | |
1925 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)) | |
1926 | == TYPE_VECTOR_SUBPARTS (masktype)); | |
0e22bb5a | 1927 | var = vect_get_new_ssa_name (masktype, vect_simple_var); |
acdcd61b JJ |
1928 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); |
1929 | new_stmt | |
0d0e4a03 | 1930 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_op); |
acdcd61b JJ |
1931 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1932 | mask_op = var; | |
1933 | } | |
5ce9450f JJ |
1934 | } |
1935 | ||
1936 | new_stmt | |
1937 | = gimple_build_call (gather_decl, 5, mask_op, ptr, op, mask_op, | |
1938 | scale); | |
1939 | ||
1940 | if (!useless_type_conversion_p (vectype, rettype)) | |
1941 | { | |
1942 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
1943 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 1944 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
5ce9450f JJ |
1945 | gimple_call_set_lhs (new_stmt, op); |
1946 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 1947 | var = make_ssa_name (vec_dest); |
5ce9450f | 1948 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
0d0e4a03 | 1949 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
1950 | } |
1951 | else | |
1952 | { | |
1953 | var = make_ssa_name (vec_dest, new_stmt); | |
1954 | gimple_call_set_lhs (new_stmt, var); | |
1955 | } | |
1956 | ||
1957 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1958 | ||
1959 | if (modifier == NARROW) | |
1960 | { | |
1961 | if ((j & 1) == 0) | |
1962 | { | |
1963 | prev_res = var; | |
1964 | continue; | |
1965 | } | |
1966 | var = permute_vec_elements (prev_res, var, | |
1967 | perm_mask, stmt, gsi); | |
1968 | new_stmt = SSA_NAME_DEF_STMT (var); | |
1969 | } | |
1970 | ||
1971 | if (prev_stmt_info == NULL) | |
1972 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1973 | else | |
1974 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1975 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1976 | } | |
3efe2e2c JJ |
1977 | |
1978 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
1979 | from the IL. */ | |
e6f5c25d IE |
1980 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
1981 | { | |
1982 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
1983 | stmt_info = vinfo_for_stmt (stmt); | |
1984 | } | |
3efe2e2c JJ |
1985 | tree lhs = gimple_call_lhs (stmt); |
1986 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
1987 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
1988 | set_vinfo_for_stmt (stmt, NULL); | |
1989 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
1990 | gsi_replace (gsi, new_stmt, true); | |
5ce9450f JJ |
1991 | return true; |
1992 | } | |
1993 | else if (is_store) | |
1994 | { | |
1995 | tree vec_rhs = NULL_TREE, vec_mask = NULL_TREE; | |
1996 | prev_stmt_info = NULL; | |
1997 | for (i = 0; i < ncopies; i++) | |
1998 | { | |
1999 | unsigned align, misalign; | |
2000 | ||
2001 | if (i == 0) | |
2002 | { | |
2003 | tree rhs = gimple_call_arg (stmt, 3); | |
81c40241 RB |
2004 | vec_rhs = vect_get_vec_def_for_operand (rhs, stmt); |
2005 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); | |
5ce9450f JJ |
2006 | /* We should have catched mismatched types earlier. */ |
2007 | gcc_assert (useless_type_conversion_p (vectype, | |
2008 | TREE_TYPE (vec_rhs))); | |
2009 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, | |
2010 | NULL_TREE, &dummy, gsi, | |
2011 | &ptr_incr, false, &inv_p); | |
2012 | gcc_assert (!inv_p); | |
2013 | } | |
2014 | else | |
2015 | { | |
81c40241 | 2016 | vect_is_simple_use (vec_rhs, loop_vinfo, &def_stmt, &dt); |
5ce9450f | 2017 | vec_rhs = vect_get_vec_def_for_stmt_copy (dt, vec_rhs); |
81c40241 | 2018 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2019 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2020 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2021 | TYPE_SIZE_UNIT (vectype)); | |
2022 | } | |
2023 | ||
2024 | align = TYPE_ALIGN_UNIT (vectype); | |
2025 | if (aligned_access_p (dr)) | |
2026 | misalign = 0; | |
2027 | else if (DR_MISALIGNMENT (dr) == -1) | |
2028 | { | |
2029 | align = TYPE_ALIGN_UNIT (elem_type); | |
2030 | misalign = 0; | |
2031 | } | |
2032 | else | |
2033 | misalign = DR_MISALIGNMENT (dr); | |
2034 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2035 | misalign); | |
2036 | new_stmt | |
2037 | = gimple_build_call_internal (IFN_MASK_STORE, 4, dataref_ptr, | |
2038 | gimple_call_arg (stmt, 1), | |
2039 | vec_mask, vec_rhs); | |
2040 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2041 | if (i == 0) | |
2042 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2043 | else | |
2044 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2045 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2046 | } | |
2047 | } | |
2048 | else | |
2049 | { | |
2050 | tree vec_mask = NULL_TREE; | |
2051 | prev_stmt_info = NULL; | |
2052 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); | |
2053 | for (i = 0; i < ncopies; i++) | |
2054 | { | |
2055 | unsigned align, misalign; | |
2056 | ||
2057 | if (i == 0) | |
2058 | { | |
81c40241 | 2059 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
5ce9450f JJ |
2060 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, |
2061 | NULL_TREE, &dummy, gsi, | |
2062 | &ptr_incr, false, &inv_p); | |
2063 | gcc_assert (!inv_p); | |
2064 | } | |
2065 | else | |
2066 | { | |
81c40241 | 2067 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2068 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2069 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2070 | TYPE_SIZE_UNIT (vectype)); | |
2071 | } | |
2072 | ||
2073 | align = TYPE_ALIGN_UNIT (vectype); | |
2074 | if (aligned_access_p (dr)) | |
2075 | misalign = 0; | |
2076 | else if (DR_MISALIGNMENT (dr) == -1) | |
2077 | { | |
2078 | align = TYPE_ALIGN_UNIT (elem_type); | |
2079 | misalign = 0; | |
2080 | } | |
2081 | else | |
2082 | misalign = DR_MISALIGNMENT (dr); | |
2083 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2084 | misalign); | |
2085 | new_stmt | |
2086 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, dataref_ptr, | |
2087 | gimple_call_arg (stmt, 1), | |
2088 | vec_mask); | |
b731b390 | 2089 | gimple_call_set_lhs (new_stmt, make_ssa_name (vec_dest)); |
5ce9450f JJ |
2090 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2091 | if (i == 0) | |
2092 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2093 | else | |
2094 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2095 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2096 | } | |
2097 | } | |
2098 | ||
3efe2e2c JJ |
2099 | if (!is_store) |
2100 | { | |
2101 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2102 | from the IL. */ | |
e6f5c25d IE |
2103 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
2104 | { | |
2105 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
2106 | stmt_info = vinfo_for_stmt (stmt); | |
2107 | } | |
3efe2e2c JJ |
2108 | tree lhs = gimple_call_lhs (stmt); |
2109 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2110 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2111 | set_vinfo_for_stmt (stmt, NULL); | |
2112 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2113 | gsi_replace (gsi, new_stmt, true); | |
2114 | } | |
2115 | ||
5ce9450f JJ |
2116 | return true; |
2117 | } | |
2118 | ||
2119 | ||
ebfd146a IR |
2120 | /* Function vectorizable_call. |
2121 | ||
538dd0b7 | 2122 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 2123 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2124 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2125 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2126 | ||
2127 | static bool | |
355fe088 | 2128 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
190c2236 | 2129 | slp_tree slp_node) |
ebfd146a | 2130 | { |
538dd0b7 | 2131 | gcall *stmt; |
ebfd146a IR |
2132 | tree vec_dest; |
2133 | tree scalar_dest; | |
2134 | tree op, type; | |
2135 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
538dd0b7 | 2136 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a IR |
2137 | tree vectype_out, vectype_in; |
2138 | int nunits_in; | |
2139 | int nunits_out; | |
2140 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 2141 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 2142 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 2143 | tree fndecl, new_temp, rhs_type; |
355fe088 | 2144 | gimple *def_stmt; |
0502fb85 UB |
2145 | enum vect_def_type dt[3] |
2146 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
355fe088 | 2147 | gimple *new_stmt = NULL; |
ebfd146a | 2148 | int ncopies, j; |
6e1aa848 | 2149 | vec<tree> vargs = vNULL; |
ebfd146a IR |
2150 | enum { NARROW, NONE, WIDEN } modifier; |
2151 | size_t i, nargs; | |
9d5e7640 | 2152 | tree lhs; |
ebfd146a | 2153 | |
190c2236 | 2154 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2155 | return false; |
2156 | ||
8644a673 | 2157 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2158 | return false; |
2159 | ||
538dd0b7 DM |
2160 | /* Is GS a vectorizable call? */ |
2161 | stmt = dyn_cast <gcall *> (gs); | |
2162 | if (!stmt) | |
ebfd146a IR |
2163 | return false; |
2164 | ||
5ce9450f JJ |
2165 | if (gimple_call_internal_p (stmt) |
2166 | && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD | |
2167 | || gimple_call_internal_fn (stmt) == IFN_MASK_STORE)) | |
2168 | return vectorizable_mask_load_store (stmt, gsi, vec_stmt, | |
2169 | slp_node); | |
2170 | ||
0136f8f0 AH |
2171 | if (gimple_call_lhs (stmt) == NULL_TREE |
2172 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
2173 | return false; |
2174 | ||
0136f8f0 | 2175 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 2176 | |
b690cc0f RG |
2177 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2178 | ||
ebfd146a IR |
2179 | /* Process function arguments. */ |
2180 | rhs_type = NULL_TREE; | |
b690cc0f | 2181 | vectype_in = NULL_TREE; |
ebfd146a IR |
2182 | nargs = gimple_call_num_args (stmt); |
2183 | ||
1b1562a5 MM |
2184 | /* Bail out if the function has more than three arguments, we do not have |
2185 | interesting builtin functions to vectorize with more than two arguments | |
2186 | except for fma. No arguments is also not good. */ | |
2187 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
2188 | return false; |
2189 | ||
74bf76ed JJ |
2190 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2191 | if (gimple_call_internal_p (stmt) | |
2192 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2193 | { | |
2194 | nargs = 0; | |
2195 | rhs_type = unsigned_type_node; | |
2196 | } | |
2197 | ||
ebfd146a IR |
2198 | for (i = 0; i < nargs; i++) |
2199 | { | |
b690cc0f RG |
2200 | tree opvectype; |
2201 | ||
ebfd146a IR |
2202 | op = gimple_call_arg (stmt, i); |
2203 | ||
2204 | /* We can only handle calls with arguments of the same type. */ | |
2205 | if (rhs_type | |
8533c9d8 | 2206 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 2207 | { |
73fbfcad | 2208 | if (dump_enabled_p ()) |
78c60e3d | 2209 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2210 | "argument types differ.\n"); |
ebfd146a IR |
2211 | return false; |
2212 | } | |
b690cc0f RG |
2213 | if (!rhs_type) |
2214 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 2215 | |
81c40241 | 2216 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[i], &opvectype)) |
ebfd146a | 2217 | { |
73fbfcad | 2218 | if (dump_enabled_p ()) |
78c60e3d | 2219 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2220 | "use not simple.\n"); |
ebfd146a IR |
2221 | return false; |
2222 | } | |
ebfd146a | 2223 | |
b690cc0f RG |
2224 | if (!vectype_in) |
2225 | vectype_in = opvectype; | |
2226 | else if (opvectype | |
2227 | && opvectype != vectype_in) | |
2228 | { | |
73fbfcad | 2229 | if (dump_enabled_p ()) |
78c60e3d | 2230 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2231 | "argument vector types differ.\n"); |
b690cc0f RG |
2232 | return false; |
2233 | } | |
2234 | } | |
2235 | /* If all arguments are external or constant defs use a vector type with | |
2236 | the same size as the output vector type. */ | |
ebfd146a | 2237 | if (!vectype_in) |
b690cc0f | 2238 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2239 | if (vec_stmt) |
2240 | gcc_assert (vectype_in); | |
2241 | if (!vectype_in) | |
2242 | { | |
73fbfcad | 2243 | if (dump_enabled_p ()) |
7d8930a0 | 2244 | { |
78c60e3d SS |
2245 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2246 | "no vectype for scalar type "); | |
2247 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2248 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
2249 | } |
2250 | ||
2251 | return false; | |
2252 | } | |
ebfd146a IR |
2253 | |
2254 | /* FORNOW */ | |
b690cc0f RG |
2255 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2256 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
2257 | if (nunits_in == nunits_out / 2) |
2258 | modifier = NARROW; | |
2259 | else if (nunits_out == nunits_in) | |
2260 | modifier = NONE; | |
2261 | else if (nunits_out == nunits_in / 2) | |
2262 | modifier = WIDEN; | |
2263 | else | |
2264 | return false; | |
2265 | ||
2266 | /* For now, we only vectorize functions if a target specific builtin | |
2267 | is available. TODO -- in some cases, it might be profitable to | |
2268 | insert the calls for pieces of the vector, in order to be able | |
2269 | to vectorize other operations in the loop. */ | |
2270 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
2271 | if (fndecl == NULL_TREE) | |
2272 | { | |
74bf76ed JJ |
2273 | if (gimple_call_internal_p (stmt) |
2274 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE | |
2275 | && !slp_node | |
2276 | && loop_vinfo | |
2277 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2278 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
2279 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2280 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
2281 | { | |
2282 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
2283 | { 0, 1, 2, ... vf - 1 } vector. */ | |
2284 | gcc_assert (nargs == 0); | |
2285 | } | |
2286 | else | |
2287 | { | |
2288 | if (dump_enabled_p ()) | |
2289 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 2290 | "function is not vectorizable.\n"); |
74bf76ed JJ |
2291 | return false; |
2292 | } | |
ebfd146a IR |
2293 | } |
2294 | ||
5006671f | 2295 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a | 2296 | |
190c2236 JJ |
2297 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
2298 | ncopies = 1; | |
2299 | else if (modifier == NARROW) | |
ebfd146a IR |
2300 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
2301 | else | |
2302 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2303 | ||
2304 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2305 | needs to be generated. */ | |
2306 | gcc_assert (ncopies >= 1); | |
2307 | ||
2308 | if (!vec_stmt) /* transformation not required. */ | |
2309 | { | |
2310 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 2311 | if (dump_enabled_p ()) |
e645e942 TJ |
2312 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
2313 | "\n"); | |
c3e7ee41 | 2314 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
2315 | return true; |
2316 | } | |
2317 | ||
2318 | /** Transform. **/ | |
2319 | ||
73fbfcad | 2320 | if (dump_enabled_p ()) |
e645e942 | 2321 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
2322 | |
2323 | /* Handle def. */ | |
2324 | scalar_dest = gimple_call_lhs (stmt); | |
2325 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2326 | ||
2327 | prev_stmt_info = NULL; | |
2328 | switch (modifier) | |
2329 | { | |
2330 | case NONE: | |
2331 | for (j = 0; j < ncopies; ++j) | |
2332 | { | |
2333 | /* Build argument list for the vectorized call. */ | |
2334 | if (j == 0) | |
9771b263 | 2335 | vargs.create (nargs); |
ebfd146a | 2336 | else |
9771b263 | 2337 | vargs.truncate (0); |
ebfd146a | 2338 | |
190c2236 JJ |
2339 | if (slp_node) |
2340 | { | |
ef062b13 | 2341 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2342 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2343 | |
2344 | for (i = 0; i < nargs; i++) | |
9771b263 | 2345 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2346 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2347 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2348 | |
2349 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2350 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
2351 | { |
2352 | size_t k; | |
2353 | for (k = 0; k < nargs; k++) | |
2354 | { | |
37b5ec8f | 2355 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 2356 | vargs[k] = vec_oprndsk[i]; |
190c2236 JJ |
2357 | } |
2358 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2359 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2360 | gimple_call_set_lhs (new_stmt, new_temp); | |
2361 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2362 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2363 | } |
2364 | ||
2365 | for (i = 0; i < nargs; i++) | |
2366 | { | |
37b5ec8f | 2367 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2368 | vec_oprndsi.release (); |
190c2236 | 2369 | } |
190c2236 JJ |
2370 | continue; |
2371 | } | |
2372 | ||
ebfd146a IR |
2373 | for (i = 0; i < nargs; i++) |
2374 | { | |
2375 | op = gimple_call_arg (stmt, i); | |
2376 | if (j == 0) | |
2377 | vec_oprnd0 | |
81c40241 | 2378 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2379 | else |
63827fb8 IR |
2380 | { |
2381 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
2382 | vec_oprnd0 | |
2383 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
2384 | } | |
ebfd146a | 2385 | |
9771b263 | 2386 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
2387 | } |
2388 | ||
74bf76ed JJ |
2389 | if (gimple_call_internal_p (stmt) |
2390 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2391 | { | |
2392 | tree *v = XALLOCAVEC (tree, nunits_out); | |
2393 | int k; | |
2394 | for (k = 0; k < nunits_out; ++k) | |
2395 | v[k] = build_int_cst (unsigned_type_node, j * nunits_out + k); | |
2396 | tree cst = build_vector (vectype_out, v); | |
2397 | tree new_var | |
0e22bb5a | 2398 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 2399 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 2400 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 2401 | new_temp = make_ssa_name (vec_dest); |
0e22bb5a | 2402 | new_stmt = gimple_build_assign (new_temp, new_var); |
74bf76ed JJ |
2403 | } |
2404 | else | |
2405 | { | |
2406 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2407 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2408 | gimple_call_set_lhs (new_stmt, new_temp); | |
2409 | } | |
ebfd146a IR |
2410 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2411 | ||
2412 | if (j == 0) | |
2413 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2414 | else | |
2415 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2416 | ||
2417 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2418 | } | |
2419 | ||
2420 | break; | |
2421 | ||
2422 | case NARROW: | |
2423 | for (j = 0; j < ncopies; ++j) | |
2424 | { | |
2425 | /* Build argument list for the vectorized call. */ | |
2426 | if (j == 0) | |
9771b263 | 2427 | vargs.create (nargs * 2); |
ebfd146a | 2428 | else |
9771b263 | 2429 | vargs.truncate (0); |
ebfd146a | 2430 | |
190c2236 JJ |
2431 | if (slp_node) |
2432 | { | |
ef062b13 | 2433 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2434 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2435 | |
2436 | for (i = 0; i < nargs; i++) | |
9771b263 | 2437 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2438 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2439 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2440 | |
2441 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2442 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
2443 | { |
2444 | size_t k; | |
9771b263 | 2445 | vargs.truncate (0); |
190c2236 JJ |
2446 | for (k = 0; k < nargs; k++) |
2447 | { | |
37b5ec8f | 2448 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
2449 | vargs.quick_push (vec_oprndsk[i]); |
2450 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 JJ |
2451 | } |
2452 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2453 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2454 | gimple_call_set_lhs (new_stmt, new_temp); | |
2455 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2456 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2457 | } |
2458 | ||
2459 | for (i = 0; i < nargs; i++) | |
2460 | { | |
37b5ec8f | 2461 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2462 | vec_oprndsi.release (); |
190c2236 | 2463 | } |
190c2236 JJ |
2464 | continue; |
2465 | } | |
2466 | ||
ebfd146a IR |
2467 | for (i = 0; i < nargs; i++) |
2468 | { | |
2469 | op = gimple_call_arg (stmt, i); | |
2470 | if (j == 0) | |
2471 | { | |
2472 | vec_oprnd0 | |
81c40241 | 2473 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2474 | vec_oprnd1 |
63827fb8 | 2475 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2476 | } |
2477 | else | |
2478 | { | |
336ecb65 | 2479 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 2480 | vec_oprnd0 |
63827fb8 | 2481 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 2482 | vec_oprnd1 |
63827fb8 | 2483 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2484 | } |
2485 | ||
9771b263 DN |
2486 | vargs.quick_push (vec_oprnd0); |
2487 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
2488 | } |
2489 | ||
2490 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2491 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2492 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
2493 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2494 | ||
2495 | if (j == 0) | |
2496 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2497 | else | |
2498 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2499 | ||
2500 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2501 | } | |
2502 | ||
2503 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2504 | ||
2505 | break; | |
2506 | ||
2507 | case WIDEN: | |
2508 | /* No current target implements this case. */ | |
2509 | return false; | |
2510 | } | |
2511 | ||
9771b263 | 2512 | vargs.release (); |
ebfd146a | 2513 | |
ebfd146a IR |
2514 | /* The call in STMT might prevent it from being removed in dce. |
2515 | We however cannot remove it here, due to the way the ssa name | |
2516 | it defines is mapped to the new definition. So just replace | |
2517 | rhs of the statement with something harmless. */ | |
2518 | ||
dd34c087 JJ |
2519 | if (slp_node) |
2520 | return true; | |
2521 | ||
ebfd146a | 2522 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
2523 | if (is_pattern_stmt_p (stmt_info)) |
2524 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
2525 | else | |
2526 | lhs = gimple_call_lhs (stmt); | |
3cc2fa2a JJ |
2527 | |
2528 | if (gimple_call_internal_p (stmt) | |
2529 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2530 | { | |
2531 | /* Replace uses of the lhs of GOMP_SIMD_LANE call outside the loop | |
2532 | with vf - 1 rather than 0, that is the last iteration of the | |
2533 | vectorized loop. */ | |
2534 | imm_use_iterator iter; | |
2535 | use_operand_p use_p; | |
355fe088 | 2536 | gimple *use_stmt; |
3cc2fa2a JJ |
2537 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs) |
2538 | { | |
2539 | basic_block use_bb = gimple_bb (use_stmt); | |
2540 | if (use_bb | |
2541 | && !flow_bb_inside_loop_p (LOOP_VINFO_LOOP (loop_vinfo), use_bb)) | |
2542 | { | |
2543 | FOR_EACH_IMM_USE_ON_STMT (use_p, iter) | |
2544 | SET_USE (use_p, build_int_cst (TREE_TYPE (lhs), | |
2545 | ncopies * nunits_out - 1)); | |
2546 | update_stmt (use_stmt); | |
2547 | } | |
2548 | } | |
2549 | } | |
2550 | ||
9d5e7640 | 2551 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
ebfd146a | 2552 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 2553 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
2554 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
2555 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
2556 | |
2557 | return true; | |
2558 | } | |
2559 | ||
2560 | ||
0136f8f0 AH |
2561 | struct simd_call_arg_info |
2562 | { | |
2563 | tree vectype; | |
2564 | tree op; | |
2565 | enum vect_def_type dt; | |
2566 | HOST_WIDE_INT linear_step; | |
2567 | unsigned int align; | |
17b658af | 2568 | bool simd_lane_linear; |
0136f8f0 AH |
2569 | }; |
2570 | ||
17b658af JJ |
2571 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
2572 | is linear within simd lane (but not within whole loop), note it in | |
2573 | *ARGINFO. */ | |
2574 | ||
2575 | static void | |
2576 | vect_simd_lane_linear (tree op, struct loop *loop, | |
2577 | struct simd_call_arg_info *arginfo) | |
2578 | { | |
355fe088 | 2579 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
2580 | |
2581 | if (!is_gimple_assign (def_stmt) | |
2582 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
2583 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
2584 | return; | |
2585 | ||
2586 | tree base = gimple_assign_rhs1 (def_stmt); | |
2587 | HOST_WIDE_INT linear_step = 0; | |
2588 | tree v = gimple_assign_rhs2 (def_stmt); | |
2589 | while (TREE_CODE (v) == SSA_NAME) | |
2590 | { | |
2591 | tree t; | |
2592 | def_stmt = SSA_NAME_DEF_STMT (v); | |
2593 | if (is_gimple_assign (def_stmt)) | |
2594 | switch (gimple_assign_rhs_code (def_stmt)) | |
2595 | { | |
2596 | case PLUS_EXPR: | |
2597 | t = gimple_assign_rhs2 (def_stmt); | |
2598 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
2599 | return; | |
2600 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
2601 | v = gimple_assign_rhs1 (def_stmt); | |
2602 | continue; | |
2603 | case MULT_EXPR: | |
2604 | t = gimple_assign_rhs2 (def_stmt); | |
2605 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
2606 | return; | |
2607 | linear_step = tree_to_shwi (t); | |
2608 | v = gimple_assign_rhs1 (def_stmt); | |
2609 | continue; | |
2610 | CASE_CONVERT: | |
2611 | t = gimple_assign_rhs1 (def_stmt); | |
2612 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
2613 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
2614 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
2615 | return; | |
2616 | if (!linear_step) | |
2617 | linear_step = 1; | |
2618 | v = t; | |
2619 | continue; | |
2620 | default: | |
2621 | return; | |
2622 | } | |
2623 | else if (is_gimple_call (def_stmt) | |
2624 | && gimple_call_internal_p (def_stmt) | |
2625 | && gimple_call_internal_fn (def_stmt) == IFN_GOMP_SIMD_LANE | |
2626 | && loop->simduid | |
2627 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
2628 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
2629 | == loop->simduid)) | |
2630 | { | |
2631 | if (!linear_step) | |
2632 | linear_step = 1; | |
2633 | arginfo->linear_step = linear_step; | |
2634 | arginfo->op = base; | |
2635 | arginfo->simd_lane_linear = true; | |
2636 | return; | |
2637 | } | |
2638 | } | |
2639 | } | |
2640 | ||
0136f8f0 AH |
2641 | /* Function vectorizable_simd_clone_call. |
2642 | ||
2643 | Check if STMT performs a function call that can be vectorized | |
2644 | by calling a simd clone of the function. | |
2645 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2646 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2647 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2648 | ||
2649 | static bool | |
355fe088 TS |
2650 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
2651 | gimple **vec_stmt, slp_tree slp_node) | |
0136f8f0 AH |
2652 | { |
2653 | tree vec_dest; | |
2654 | tree scalar_dest; | |
2655 | tree op, type; | |
2656 | tree vec_oprnd0 = NULL_TREE; | |
2657 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
2658 | tree vectype; | |
2659 | unsigned int nunits; | |
2660 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2661 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 2662 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 2663 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 2664 | tree fndecl, new_temp; |
355fe088 TS |
2665 | gimple *def_stmt; |
2666 | gimple *new_stmt = NULL; | |
0136f8f0 AH |
2667 | int ncopies, j; |
2668 | vec<simd_call_arg_info> arginfo = vNULL; | |
2669 | vec<tree> vargs = vNULL; | |
2670 | size_t i, nargs; | |
2671 | tree lhs, rtype, ratype; | |
2672 | vec<constructor_elt, va_gc> *ret_ctor_elts; | |
2673 | ||
2674 | /* Is STMT a vectorizable call? */ | |
2675 | if (!is_gimple_call (stmt)) | |
2676 | return false; | |
2677 | ||
2678 | fndecl = gimple_call_fndecl (stmt); | |
2679 | if (fndecl == NULL_TREE) | |
2680 | return false; | |
2681 | ||
d52f5295 | 2682 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
2683 | if (node == NULL || node->simd_clones == NULL) |
2684 | return false; | |
2685 | ||
2686 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
2687 | return false; | |
2688 | ||
2689 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
2690 | return false; | |
2691 | ||
2692 | if (gimple_call_lhs (stmt) | |
2693 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
2694 | return false; | |
2695 | ||
2696 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
2697 | ||
2698 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2699 | ||
2700 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
2701 | return false; | |
2702 | ||
2703 | /* FORNOW */ | |
2704 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
2705 | return false; | |
2706 | ||
2707 | /* Process function arguments. */ | |
2708 | nargs = gimple_call_num_args (stmt); | |
2709 | ||
2710 | /* Bail out if the function has zero arguments. */ | |
2711 | if (nargs == 0) | |
2712 | return false; | |
2713 | ||
2714 | arginfo.create (nargs); | |
2715 | ||
2716 | for (i = 0; i < nargs; i++) | |
2717 | { | |
2718 | simd_call_arg_info thisarginfo; | |
2719 | affine_iv iv; | |
2720 | ||
2721 | thisarginfo.linear_step = 0; | |
2722 | thisarginfo.align = 0; | |
2723 | thisarginfo.op = NULL_TREE; | |
17b658af | 2724 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
2725 | |
2726 | op = gimple_call_arg (stmt, i); | |
81c40241 RB |
2727 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &thisarginfo.dt, |
2728 | &thisarginfo.vectype) | |
0136f8f0 AH |
2729 | || thisarginfo.dt == vect_uninitialized_def) |
2730 | { | |
2731 | if (dump_enabled_p ()) | |
2732 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2733 | "use not simple.\n"); | |
2734 | arginfo.release (); | |
2735 | return false; | |
2736 | } | |
2737 | ||
2738 | if (thisarginfo.dt == vect_constant_def | |
2739 | || thisarginfo.dt == vect_external_def) | |
2740 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
2741 | else | |
2742 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
2743 | ||
6c9e85fb JJ |
2744 | /* For linear arguments, the analyze phase should have saved |
2745 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
2746 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
2747 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
2748 | { |
2749 | gcc_assert (vec_stmt); | |
2750 | thisarginfo.linear_step | |
17b658af | 2751 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 2752 | thisarginfo.op |
17b658af JJ |
2753 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
2754 | thisarginfo.simd_lane_linear | |
2755 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
2756 | == boolean_true_node); | |
6c9e85fb JJ |
2757 | /* If loop has been peeled for alignment, we need to adjust it. */ |
2758 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
2759 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 2760 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
2761 | { |
2762 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 2763 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
2764 | tree opt = TREE_TYPE (thisarginfo.op); |
2765 | bias = fold_convert (TREE_TYPE (step), bias); | |
2766 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
2767 | thisarginfo.op | |
2768 | = fold_build2 (POINTER_TYPE_P (opt) | |
2769 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
2770 | thisarginfo.op, bias); | |
2771 | } | |
2772 | } | |
2773 | else if (!vec_stmt | |
2774 | && thisarginfo.dt != vect_constant_def | |
2775 | && thisarginfo.dt != vect_external_def | |
2776 | && loop_vinfo | |
2777 | && TREE_CODE (op) == SSA_NAME | |
2778 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
2779 | &iv, false) | |
2780 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
2781 | { |
2782 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
2783 | thisarginfo.op = iv.base; | |
2784 | } | |
2785 | else if ((thisarginfo.dt == vect_constant_def | |
2786 | || thisarginfo.dt == vect_external_def) | |
2787 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
2788 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
2789 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
2790 | linear too. */ | |
2791 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
2792 | && !thisarginfo.linear_step | |
2793 | && !vec_stmt | |
2794 | && thisarginfo.dt != vect_constant_def | |
2795 | && thisarginfo.dt != vect_external_def | |
2796 | && loop_vinfo | |
2797 | && !slp_node | |
2798 | && TREE_CODE (op) == SSA_NAME) | |
2799 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
2800 | |
2801 | arginfo.quick_push (thisarginfo); | |
2802 | } | |
2803 | ||
2804 | unsigned int badness = 0; | |
2805 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
2806 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
2807 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
2808 | else |
2809 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
2810 | n = n->simdclone->next_clone) | |
2811 | { | |
2812 | unsigned int this_badness = 0; | |
2813 | if (n->simdclone->simdlen | |
2814 | > (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
2815 | || n->simdclone->nargs != nargs) | |
2816 | continue; | |
2817 | if (n->simdclone->simdlen | |
2818 | < (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2819 | this_badness += (exact_log2 (LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2820 | - exact_log2 (n->simdclone->simdlen)) * 1024; | |
2821 | if (n->simdclone->inbranch) | |
2822 | this_badness += 2048; | |
2823 | int target_badness = targetm.simd_clone.usable (n); | |
2824 | if (target_badness < 0) | |
2825 | continue; | |
2826 | this_badness += target_badness * 512; | |
2827 | /* FORNOW: Have to add code to add the mask argument. */ | |
2828 | if (n->simdclone->inbranch) | |
2829 | continue; | |
2830 | for (i = 0; i < nargs; i++) | |
2831 | { | |
2832 | switch (n->simdclone->args[i].arg_type) | |
2833 | { | |
2834 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
2835 | if (!useless_type_conversion_p | |
2836 | (n->simdclone->args[i].orig_type, | |
2837 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
2838 | i = -1; | |
2839 | else if (arginfo[i].dt == vect_constant_def | |
2840 | || arginfo[i].dt == vect_external_def | |
2841 | || arginfo[i].linear_step) | |
2842 | this_badness += 64; | |
2843 | break; | |
2844 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
2845 | if (arginfo[i].dt != vect_constant_def | |
2846 | && arginfo[i].dt != vect_external_def) | |
2847 | i = -1; | |
2848 | break; | |
2849 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 2850 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
2851 | if (arginfo[i].dt == vect_constant_def |
2852 | || arginfo[i].dt == vect_external_def | |
2853 | || (arginfo[i].linear_step | |
2854 | != n->simdclone->args[i].linear_step)) | |
2855 | i = -1; | |
2856 | break; | |
2857 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
2858 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
2859 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
2860 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
2861 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
2862 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
2863 | /* FORNOW */ |
2864 | i = -1; | |
2865 | break; | |
2866 | case SIMD_CLONE_ARG_TYPE_MASK: | |
2867 | gcc_unreachable (); | |
2868 | } | |
2869 | if (i == (size_t) -1) | |
2870 | break; | |
2871 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
2872 | { | |
2873 | i = -1; | |
2874 | break; | |
2875 | } | |
2876 | if (arginfo[i].align) | |
2877 | this_badness += (exact_log2 (arginfo[i].align) | |
2878 | - exact_log2 (n->simdclone->args[i].alignment)); | |
2879 | } | |
2880 | if (i == (size_t) -1) | |
2881 | continue; | |
2882 | if (bestn == NULL || this_badness < badness) | |
2883 | { | |
2884 | bestn = n; | |
2885 | badness = this_badness; | |
2886 | } | |
2887 | } | |
2888 | ||
2889 | if (bestn == NULL) | |
2890 | { | |
2891 | arginfo.release (); | |
2892 | return false; | |
2893 | } | |
2894 | ||
2895 | for (i = 0; i < nargs; i++) | |
2896 | if ((arginfo[i].dt == vect_constant_def | |
2897 | || arginfo[i].dt == vect_external_def) | |
2898 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
2899 | { | |
2900 | arginfo[i].vectype | |
2901 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
2902 | i))); | |
2903 | if (arginfo[i].vectype == NULL | |
2904 | || (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
2905 | > bestn->simdclone->simdlen)) | |
2906 | { | |
2907 | arginfo.release (); | |
2908 | return false; | |
2909 | } | |
2910 | } | |
2911 | ||
2912 | fndecl = bestn->decl; | |
2913 | nunits = bestn->simdclone->simdlen; | |
2914 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2915 | ||
2916 | /* If the function isn't const, only allow it in simd loops where user | |
2917 | has asserted that at least nunits consecutive iterations can be | |
2918 | performed using SIMD instructions. */ | |
2919 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
2920 | && gimple_vuse (stmt)) | |
2921 | { | |
2922 | arginfo.release (); | |
2923 | return false; | |
2924 | } | |
2925 | ||
2926 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2927 | needs to be generated. */ | |
2928 | gcc_assert (ncopies >= 1); | |
2929 | ||
2930 | if (!vec_stmt) /* transformation not required. */ | |
2931 | { | |
6c9e85fb JJ |
2932 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
2933 | for (i = 0; i < nargs; i++) | |
2934 | if (bestn->simdclone->args[i].arg_type | |
2935 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
2936 | { | |
17b658af | 2937 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
2938 | + 1); |
2939 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
2940 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
2941 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
2942 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
2943 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
2944 | tree sll = arginfo[i].simd_lane_linear |
2945 | ? boolean_true_node : boolean_false_node; | |
2946 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 2947 | } |
0136f8f0 AH |
2948 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
2949 | if (dump_enabled_p ()) | |
2950 | dump_printf_loc (MSG_NOTE, vect_location, | |
2951 | "=== vectorizable_simd_clone_call ===\n"); | |
2952 | /* vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); */ | |
2953 | arginfo.release (); | |
2954 | return true; | |
2955 | } | |
2956 | ||
2957 | /** Transform. **/ | |
2958 | ||
2959 | if (dump_enabled_p ()) | |
2960 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
2961 | ||
2962 | /* Handle def. */ | |
2963 | scalar_dest = gimple_call_lhs (stmt); | |
2964 | vec_dest = NULL_TREE; | |
2965 | rtype = NULL_TREE; | |
2966 | ratype = NULL_TREE; | |
2967 | if (scalar_dest) | |
2968 | { | |
2969 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2970 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
2971 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
2972 | { | |
2973 | ratype = rtype; | |
2974 | rtype = TREE_TYPE (ratype); | |
2975 | } | |
2976 | } | |
2977 | ||
2978 | prev_stmt_info = NULL; | |
2979 | for (j = 0; j < ncopies; ++j) | |
2980 | { | |
2981 | /* Build argument list for the vectorized call. */ | |
2982 | if (j == 0) | |
2983 | vargs.create (nargs); | |
2984 | else | |
2985 | vargs.truncate (0); | |
2986 | ||
2987 | for (i = 0; i < nargs; i++) | |
2988 | { | |
2989 | unsigned int k, l, m, o; | |
2990 | tree atype; | |
2991 | op = gimple_call_arg (stmt, i); | |
2992 | switch (bestn->simdclone->args[i].arg_type) | |
2993 | { | |
2994 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
2995 | atype = bestn->simdclone->args[i].vector_type; | |
2996 | o = nunits / TYPE_VECTOR_SUBPARTS (atype); | |
2997 | for (m = j * o; m < (j + 1) * o; m++) | |
2998 | { | |
2999 | if (TYPE_VECTOR_SUBPARTS (atype) | |
3000 | < TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)) | |
3001 | { | |
3002 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); | |
3003 | k = (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
3004 | / TYPE_VECTOR_SUBPARTS (atype)); | |
3005 | gcc_assert ((k & (k - 1)) == 0); | |
3006 | if (m == 0) | |
3007 | vec_oprnd0 | |
81c40241 | 3008 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3009 | else |
3010 | { | |
3011 | vec_oprnd0 = arginfo[i].op; | |
3012 | if ((m & (k - 1)) == 0) | |
3013 | vec_oprnd0 | |
3014 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3015 | vec_oprnd0); | |
3016 | } | |
3017 | arginfo[i].op = vec_oprnd0; | |
3018 | vec_oprnd0 | |
3019 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
3020 | size_int (prec), | |
3021 | bitsize_int ((m & (k - 1)) * prec)); | |
3022 | new_stmt | |
b731b390 | 3023 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3024 | vec_oprnd0); |
3025 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3026 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3027 | } | |
3028 | else | |
3029 | { | |
3030 | k = (TYPE_VECTOR_SUBPARTS (atype) | |
3031 | / TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)); | |
3032 | gcc_assert ((k & (k - 1)) == 0); | |
3033 | vec<constructor_elt, va_gc> *ctor_elts; | |
3034 | if (k != 1) | |
3035 | vec_alloc (ctor_elts, k); | |
3036 | else | |
3037 | ctor_elts = NULL; | |
3038 | for (l = 0; l < k; l++) | |
3039 | { | |
3040 | if (m == 0 && l == 0) | |
3041 | vec_oprnd0 | |
81c40241 | 3042 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3043 | else |
3044 | vec_oprnd0 | |
3045 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3046 | arginfo[i].op); | |
3047 | arginfo[i].op = vec_oprnd0; | |
3048 | if (k == 1) | |
3049 | break; | |
3050 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
3051 | vec_oprnd0); | |
3052 | } | |
3053 | if (k == 1) | |
3054 | vargs.safe_push (vec_oprnd0); | |
3055 | else | |
3056 | { | |
3057 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
3058 | new_stmt | |
b731b390 | 3059 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3060 | vec_oprnd0); |
3061 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3062 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3063 | } | |
3064 | } | |
3065 | } | |
3066 | break; | |
3067 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3068 | vargs.safe_push (op); | |
3069 | break; | |
3070 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
3071 | if (j == 0) | |
3072 | { | |
3073 | gimple_seq stmts; | |
3074 | arginfo[i].op | |
3075 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
3076 | NULL_TREE); | |
3077 | if (stmts != NULL) | |
3078 | { | |
3079 | basic_block new_bb; | |
3080 | edge pe = loop_preheader_edge (loop); | |
3081 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
3082 | gcc_assert (!new_bb); | |
3083 | } | |
17b658af JJ |
3084 | if (arginfo[i].simd_lane_linear) |
3085 | { | |
3086 | vargs.safe_push (arginfo[i].op); | |
3087 | break; | |
3088 | } | |
b731b390 | 3089 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 3090 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
0136f8f0 | 3091 | set_vinfo_for_stmt (new_phi, |
310213d4 | 3092 | new_stmt_vec_info (new_phi, loop_vinfo)); |
0136f8f0 AH |
3093 | add_phi_arg (new_phi, arginfo[i].op, |
3094 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
3095 | enum tree_code code | |
3096 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3097 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3098 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3099 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3100 | widest_int cst |
3101 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3102 | ncopies * nunits); | |
3103 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3104 | tree phi_arg = copy_ssa_name (op); |
0d0e4a03 JJ |
3105 | new_stmt |
3106 | = gimple_build_assign (phi_arg, code, phi_res, tcst); | |
0136f8f0 AH |
3107 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
3108 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
3109 | set_vinfo_for_stmt (new_stmt, | |
310213d4 | 3110 | new_stmt_vec_info (new_stmt, loop_vinfo)); |
0136f8f0 AH |
3111 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
3112 | UNKNOWN_LOCATION); | |
3113 | arginfo[i].op = phi_res; | |
3114 | vargs.safe_push (phi_res); | |
3115 | } | |
3116 | else | |
3117 | { | |
3118 | enum tree_code code | |
3119 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3120 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3121 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3122 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3123 | widest_int cst |
3124 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3125 | j * nunits); | |
3126 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3127 | new_temp = make_ssa_name (TREE_TYPE (op)); |
0d0e4a03 JJ |
3128 | new_stmt = gimple_build_assign (new_temp, code, |
3129 | arginfo[i].op, tcst); | |
0136f8f0 AH |
3130 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3131 | vargs.safe_push (new_temp); | |
3132 | } | |
3133 | break; | |
3134 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
e01d41e5 JJ |
3135 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3136 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3137 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3138 | default: |
3139 | gcc_unreachable (); | |
3140 | } | |
3141 | } | |
3142 | ||
3143 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
3144 | if (vec_dest) | |
3145 | { | |
3146 | gcc_assert (ratype || TYPE_VECTOR_SUBPARTS (rtype) == nunits); | |
3147 | if (ratype) | |
b731b390 | 3148 | new_temp = create_tmp_var (ratype); |
0136f8f0 AH |
3149 | else if (TYPE_VECTOR_SUBPARTS (vectype) |
3150 | == TYPE_VECTOR_SUBPARTS (rtype)) | |
3151 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3152 | else | |
3153 | new_temp = make_ssa_name (rtype, new_stmt); | |
3154 | gimple_call_set_lhs (new_stmt, new_temp); | |
3155 | } | |
3156 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3157 | ||
3158 | if (vec_dest) | |
3159 | { | |
3160 | if (TYPE_VECTOR_SUBPARTS (vectype) < nunits) | |
3161 | { | |
3162 | unsigned int k, l; | |
3163 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); | |
3164 | k = nunits / TYPE_VECTOR_SUBPARTS (vectype); | |
3165 | gcc_assert ((k & (k - 1)) == 0); | |
3166 | for (l = 0; l < k; l++) | |
3167 | { | |
3168 | tree t; | |
3169 | if (ratype) | |
3170 | { | |
3171 | t = build_fold_addr_expr (new_temp); | |
3172 | t = build2 (MEM_REF, vectype, t, | |
3173 | build_int_cst (TREE_TYPE (t), | |
3174 | l * prec / BITS_PER_UNIT)); | |
3175 | } | |
3176 | else | |
3177 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
3178 | size_int (prec), bitsize_int (l * prec)); | |
3179 | new_stmt | |
b731b390 | 3180 | = gimple_build_assign (make_ssa_name (vectype), t); |
0136f8f0 AH |
3181 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3182 | if (j == 0 && l == 0) | |
3183 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3184 | else | |
3185 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3186 | ||
3187 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3188 | } | |
3189 | ||
3190 | if (ratype) | |
3191 | { | |
3192 | tree clobber = build_constructor (ratype, NULL); | |
3193 | TREE_THIS_VOLATILE (clobber) = 1; | |
3194 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3195 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3196 | } | |
3197 | continue; | |
3198 | } | |
3199 | else if (TYPE_VECTOR_SUBPARTS (vectype) > nunits) | |
3200 | { | |
3201 | unsigned int k = (TYPE_VECTOR_SUBPARTS (vectype) | |
3202 | / TYPE_VECTOR_SUBPARTS (rtype)); | |
3203 | gcc_assert ((k & (k - 1)) == 0); | |
3204 | if ((j & (k - 1)) == 0) | |
3205 | vec_alloc (ret_ctor_elts, k); | |
3206 | if (ratype) | |
3207 | { | |
3208 | unsigned int m, o = nunits / TYPE_VECTOR_SUBPARTS (rtype); | |
3209 | for (m = 0; m < o; m++) | |
3210 | { | |
3211 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
3212 | size_int (m), NULL_TREE, NULL_TREE); | |
3213 | new_stmt | |
b731b390 | 3214 | = gimple_build_assign (make_ssa_name (rtype), tem); |
0136f8f0 AH |
3215 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3216 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
3217 | gimple_assign_lhs (new_stmt)); | |
3218 | } | |
3219 | tree clobber = build_constructor (ratype, NULL); | |
3220 | TREE_THIS_VOLATILE (clobber) = 1; | |
3221 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3222 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3223 | } | |
3224 | else | |
3225 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
3226 | if ((j & (k - 1)) != k - 1) | |
3227 | continue; | |
3228 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
3229 | new_stmt | |
b731b390 | 3230 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
0136f8f0 AH |
3231 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3232 | ||
3233 | if ((unsigned) j == k - 1) | |
3234 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3235 | else | |
3236 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3237 | ||
3238 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3239 | continue; | |
3240 | } | |
3241 | else if (ratype) | |
3242 | { | |
3243 | tree t = build_fold_addr_expr (new_temp); | |
3244 | t = build2 (MEM_REF, vectype, t, | |
3245 | build_int_cst (TREE_TYPE (t), 0)); | |
3246 | new_stmt | |
b731b390 | 3247 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
0136f8f0 AH |
3248 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3249 | tree clobber = build_constructor (ratype, NULL); | |
3250 | TREE_THIS_VOLATILE (clobber) = 1; | |
3251 | vect_finish_stmt_generation (stmt, | |
3252 | gimple_build_assign (new_temp, | |
3253 | clobber), gsi); | |
3254 | } | |
3255 | } | |
3256 | ||
3257 | if (j == 0) | |
3258 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3259 | else | |
3260 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3261 | ||
3262 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3263 | } | |
3264 | ||
3265 | vargs.release (); | |
3266 | ||
3267 | /* The call in STMT might prevent it from being removed in dce. | |
3268 | We however cannot remove it here, due to the way the ssa name | |
3269 | it defines is mapped to the new definition. So just replace | |
3270 | rhs of the statement with something harmless. */ | |
3271 | ||
3272 | if (slp_node) | |
3273 | return true; | |
3274 | ||
3275 | if (scalar_dest) | |
3276 | { | |
3277 | type = TREE_TYPE (scalar_dest); | |
3278 | if (is_pattern_stmt_p (stmt_info)) | |
3279 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3280 | else | |
3281 | lhs = gimple_call_lhs (stmt); | |
3282 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
3283 | } | |
3284 | else | |
3285 | new_stmt = gimple_build_nop (); | |
3286 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
3287 | set_vinfo_for_stmt (stmt, NULL); | |
3288 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 3289 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
3290 | unlink_stmt_vdef (stmt); |
3291 | ||
3292 | return true; | |
3293 | } | |
3294 | ||
3295 | ||
ebfd146a IR |
3296 | /* Function vect_gen_widened_results_half |
3297 | ||
3298 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 3299 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 3300 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
3301 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
3302 | needs to be created (DECL is a function-decl of a target-builtin). | |
3303 | STMT is the original scalar stmt that we are vectorizing. */ | |
3304 | ||
355fe088 | 3305 | static gimple * |
ebfd146a IR |
3306 | vect_gen_widened_results_half (enum tree_code code, |
3307 | tree decl, | |
3308 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
3309 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 3310 | gimple *stmt) |
b8698a0f | 3311 | { |
355fe088 | 3312 | gimple *new_stmt; |
b8698a0f L |
3313 | tree new_temp; |
3314 | ||
3315 | /* Generate half of the widened result: */ | |
3316 | if (code == CALL_EXPR) | |
3317 | { | |
3318 | /* Target specific support */ | |
ebfd146a IR |
3319 | if (op_type == binary_op) |
3320 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
3321 | else | |
3322 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
3323 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3324 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
3325 | } |
3326 | else | |
ebfd146a | 3327 | { |
b8698a0f L |
3328 | /* Generic support */ |
3329 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
3330 | if (op_type != binary_op) |
3331 | vec_oprnd1 = NULL; | |
0d0e4a03 | 3332 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
3333 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3334 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 3335 | } |
ebfd146a IR |
3336 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3337 | ||
ebfd146a IR |
3338 | return new_stmt; |
3339 | } | |
3340 | ||
4a00c761 JJ |
3341 | |
3342 | /* Get vectorized definitions for loop-based vectorization. For the first | |
3343 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
3344 | scalar operand), and for the rest we get a copy with | |
3345 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
3346 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
3347 | The vectors are collected into VEC_OPRNDS. */ | |
3348 | ||
3349 | static void | |
355fe088 | 3350 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 3351 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
3352 | { |
3353 | tree vec_oprnd; | |
3354 | ||
3355 | /* Get first vector operand. */ | |
3356 | /* All the vector operands except the very first one (that is scalar oprnd) | |
3357 | are stmt copies. */ | |
3358 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 3359 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
3360 | else |
3361 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
3362 | ||
9771b263 | 3363 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3364 | |
3365 | /* Get second vector operand. */ | |
3366 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 3367 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3368 | |
3369 | *oprnd = vec_oprnd; | |
3370 | ||
3371 | /* For conversion in multiple steps, continue to get operands | |
3372 | recursively. */ | |
3373 | if (multi_step_cvt) | |
3374 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
3375 | } | |
3376 | ||
3377 | ||
3378 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
3379 | For multi-step conversions store the resulting vectors and call the function | |
3380 | recursively. */ | |
3381 | ||
3382 | static void | |
9771b263 | 3383 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 3384 | int multi_step_cvt, gimple *stmt, |
9771b263 | 3385 | vec<tree> vec_dsts, |
4a00c761 JJ |
3386 | gimple_stmt_iterator *gsi, |
3387 | slp_tree slp_node, enum tree_code code, | |
3388 | stmt_vec_info *prev_stmt_info) | |
3389 | { | |
3390 | unsigned int i; | |
3391 | tree vop0, vop1, new_tmp, vec_dest; | |
355fe088 | 3392 | gimple *new_stmt; |
4a00c761 JJ |
3393 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
3394 | ||
9771b263 | 3395 | vec_dest = vec_dsts.pop (); |
4a00c761 | 3396 | |
9771b263 | 3397 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
3398 | { |
3399 | /* Create demotion operation. */ | |
9771b263 DN |
3400 | vop0 = (*vec_oprnds)[i]; |
3401 | vop1 = (*vec_oprnds)[i + 1]; | |
0d0e4a03 | 3402 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
3403 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
3404 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
3405 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3406 | ||
3407 | if (multi_step_cvt) | |
3408 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 3409 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
3410 | else |
3411 | { | |
3412 | /* This is the last step of the conversion sequence. Store the | |
3413 | vectors in SLP_NODE or in vector info of the scalar statement | |
3414 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
3415 | if (slp_node) | |
9771b263 | 3416 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 3417 | else |
c689ce1e RB |
3418 | { |
3419 | if (!*prev_stmt_info) | |
3420 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3421 | else | |
3422 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
4a00c761 | 3423 | |
c689ce1e RB |
3424 | *prev_stmt_info = vinfo_for_stmt (new_stmt); |
3425 | } | |
4a00c761 JJ |
3426 | } |
3427 | } | |
3428 | ||
3429 | /* For multi-step demotion operations we first generate demotion operations | |
3430 | from the source type to the intermediate types, and then combine the | |
3431 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
3432 | type. */ | |
3433 | if (multi_step_cvt) | |
3434 | { | |
3435 | /* At each level of recursion we have half of the operands we had at the | |
3436 | previous level. */ | |
9771b263 | 3437 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
3438 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
3439 | stmt, vec_dsts, gsi, slp_node, | |
3440 | VEC_PACK_TRUNC_EXPR, | |
3441 | prev_stmt_info); | |
3442 | } | |
3443 | ||
9771b263 | 3444 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3445 | } |
3446 | ||
3447 | ||
3448 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
3449 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
3450 | the resulting vectors and call the function recursively. */ | |
3451 | ||
3452 | static void | |
9771b263 DN |
3453 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
3454 | vec<tree> *vec_oprnds1, | |
355fe088 | 3455 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
3456 | gimple_stmt_iterator *gsi, |
3457 | enum tree_code code1, | |
3458 | enum tree_code code2, tree decl1, | |
3459 | tree decl2, int op_type) | |
3460 | { | |
3461 | int i; | |
3462 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 3463 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 3464 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 3465 | |
9771b263 DN |
3466 | vec_tmp.create (vec_oprnds0->length () * 2); |
3467 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
3468 | { |
3469 | if (op_type == binary_op) | |
9771b263 | 3470 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
3471 | else |
3472 | vop1 = NULL_TREE; | |
3473 | ||
3474 | /* Generate the two halves of promotion operation. */ | |
3475 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
3476 | op_type, vec_dest, gsi, stmt); | |
3477 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
3478 | op_type, vec_dest, gsi, stmt); | |
3479 | if (is_gimple_call (new_stmt1)) | |
3480 | { | |
3481 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
3482 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
3483 | } | |
3484 | else | |
3485 | { | |
3486 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
3487 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
3488 | } | |
3489 | ||
3490 | /* Store the results for the next step. */ | |
9771b263 DN |
3491 | vec_tmp.quick_push (new_tmp1); |
3492 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
3493 | } |
3494 | ||
689eaba3 | 3495 | vec_oprnds0->release (); |
4a00c761 JJ |
3496 | *vec_oprnds0 = vec_tmp; |
3497 | } | |
3498 | ||
3499 | ||
b8698a0f L |
3500 | /* Check if STMT performs a conversion operation, that can be vectorized. |
3501 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 3502 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
3503 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
3504 | ||
3505 | static bool | |
355fe088 TS |
3506 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
3507 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
3508 | { |
3509 | tree vec_dest; | |
3510 | tree scalar_dest; | |
4a00c761 | 3511 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
3512 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
3513 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3514 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3515 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 3516 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
3517 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
3518 | tree new_temp; | |
355fe088 | 3519 | gimple *def_stmt; |
ebfd146a | 3520 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
355fe088 | 3521 | gimple *new_stmt = NULL; |
ebfd146a IR |
3522 | stmt_vec_info prev_stmt_info; |
3523 | int nunits_in; | |
3524 | int nunits_out; | |
3525 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
3526 | int ncopies, i, j; |
3527 | tree lhs_type, rhs_type; | |
ebfd146a | 3528 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
3529 | vec<tree> vec_oprnds0 = vNULL; |
3530 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 3531 | tree vop0; |
4a00c761 | 3532 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 3533 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 3534 | int multi_step_cvt = 0; |
6e1aa848 DN |
3535 | vec<tree> vec_dsts = vNULL; |
3536 | vec<tree> interm_types = vNULL; | |
4a00c761 JJ |
3537 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
3538 | int op_type; | |
ef4bddc2 | 3539 | machine_mode rhs_mode; |
4a00c761 | 3540 | unsigned short fltsz; |
ebfd146a IR |
3541 | |
3542 | /* Is STMT a vectorizable conversion? */ | |
3543 | ||
4a00c761 | 3544 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3545 | return false; |
3546 | ||
8644a673 | 3547 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3548 | return false; |
3549 | ||
3550 | if (!is_gimple_assign (stmt)) | |
3551 | return false; | |
3552 | ||
3553 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3554 | return false; | |
3555 | ||
3556 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
3557 | if (!CONVERT_EXPR_CODE_P (code) |
3558 | && code != FIX_TRUNC_EXPR | |
3559 | && code != FLOAT_EXPR | |
3560 | && code != WIDEN_MULT_EXPR | |
3561 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
3562 | return false; |
3563 | ||
4a00c761 JJ |
3564 | op_type = TREE_CODE_LENGTH (code); |
3565 | ||
ebfd146a | 3566 | /* Check types of lhs and rhs. */ |
b690cc0f | 3567 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 3568 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
3569 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3570 | ||
ebfd146a IR |
3571 | op0 = gimple_assign_rhs1 (stmt); |
3572 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
3573 | |
3574 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3575 | && !((INTEGRAL_TYPE_P (lhs_type) | |
3576 | && INTEGRAL_TYPE_P (rhs_type)) | |
3577 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
3578 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
3579 | return false; | |
3580 | ||
e6f5c25d IE |
3581 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
3582 | && ((INTEGRAL_TYPE_P (lhs_type) | |
3583 | && (TYPE_PRECISION (lhs_type) | |
3584 | != GET_MODE_PRECISION (TYPE_MODE (lhs_type)))) | |
3585 | || (INTEGRAL_TYPE_P (rhs_type) | |
3586 | && (TYPE_PRECISION (rhs_type) | |
3587 | != GET_MODE_PRECISION (TYPE_MODE (rhs_type)))))) | |
4a00c761 | 3588 | { |
73fbfcad | 3589 | if (dump_enabled_p ()) |
78c60e3d | 3590 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
3591 | "type conversion to/from bit-precision unsupported." |
3592 | "\n"); | |
4a00c761 JJ |
3593 | return false; |
3594 | } | |
3595 | ||
b690cc0f | 3596 | /* Check the operands of the operation. */ |
81c40241 | 3597 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype_in)) |
b690cc0f | 3598 | { |
73fbfcad | 3599 | if (dump_enabled_p ()) |
78c60e3d | 3600 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3601 | "use not simple.\n"); |
b690cc0f RG |
3602 | return false; |
3603 | } | |
4a00c761 JJ |
3604 | if (op_type == binary_op) |
3605 | { | |
3606 | bool ok; | |
3607 | ||
3608 | op1 = gimple_assign_rhs2 (stmt); | |
3609 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
3610 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
3611 | OP1. */ | |
3612 | if (CONSTANT_CLASS_P (op0)) | |
81c40241 | 3613 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &vectype_in); |
4a00c761 | 3614 | else |
81c40241 | 3615 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1]); |
4a00c761 JJ |
3616 | |
3617 | if (!ok) | |
3618 | { | |
73fbfcad | 3619 | if (dump_enabled_p ()) |
78c60e3d | 3620 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3621 | "use not simple.\n"); |
4a00c761 JJ |
3622 | return false; |
3623 | } | |
3624 | } | |
3625 | ||
b690cc0f RG |
3626 | /* If op0 is an external or constant defs use a vector type of |
3627 | the same size as the output vector type. */ | |
ebfd146a | 3628 | if (!vectype_in) |
b690cc0f | 3629 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
3630 | if (vec_stmt) |
3631 | gcc_assert (vectype_in); | |
3632 | if (!vectype_in) | |
3633 | { | |
73fbfcad | 3634 | if (dump_enabled_p ()) |
4a00c761 | 3635 | { |
78c60e3d SS |
3636 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3637 | "no vectype for scalar type "); | |
3638 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 3639 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 3640 | } |
7d8930a0 IR |
3641 | |
3642 | return false; | |
3643 | } | |
ebfd146a | 3644 | |
e6f5c25d IE |
3645 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
3646 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
3647 | { | |
3648 | if (dump_enabled_p ()) | |
3649 | { | |
3650 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3651 | "can't convert between boolean and non " | |
3652 | "boolean vectors"); | |
3653 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
3654 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
3655 | } | |
3656 | ||
3657 | return false; | |
3658 | } | |
3659 | ||
b690cc0f RG |
3660 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
3661 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 3662 | if (nunits_in < nunits_out) |
ebfd146a IR |
3663 | modifier = NARROW; |
3664 | else if (nunits_out == nunits_in) | |
3665 | modifier = NONE; | |
ebfd146a | 3666 | else |
4a00c761 | 3667 | modifier = WIDEN; |
ebfd146a | 3668 | |
ff802fa1 IR |
3669 | /* Multiple types in SLP are handled by creating the appropriate number of |
3670 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3671 | case of SLP. */ | |
437f4a00 | 3672 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a | 3673 | ncopies = 1; |
4a00c761 JJ |
3674 | else if (modifier == NARROW) |
3675 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
3676 | else | |
3677 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 3678 | |
ebfd146a IR |
3679 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
3680 | needs to be generated. */ | |
3681 | gcc_assert (ncopies >= 1); | |
3682 | ||
ebfd146a | 3683 | /* Supportable by target? */ |
4a00c761 | 3684 | switch (modifier) |
ebfd146a | 3685 | { |
4a00c761 JJ |
3686 | case NONE: |
3687 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3688 | return false; | |
3689 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
3690 | &decl1, &code1)) | |
3691 | break; | |
3692 | /* FALLTHRU */ | |
3693 | unsupported: | |
73fbfcad | 3694 | if (dump_enabled_p ()) |
78c60e3d | 3695 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3696 | "conversion not supported by target.\n"); |
ebfd146a | 3697 | return false; |
ebfd146a | 3698 | |
4a00c761 JJ |
3699 | case WIDEN: |
3700 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
3701 | &code1, &code2, &multi_step_cvt, |
3702 | &interm_types)) | |
4a00c761 JJ |
3703 | { |
3704 | /* Binary widening operation can only be supported directly by the | |
3705 | architecture. */ | |
3706 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
3707 | break; | |
3708 | } | |
3709 | ||
3710 | if (code != FLOAT_EXPR | |
3711 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3712 | <= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3713 | goto unsupported; | |
3714 | ||
3715 | rhs_mode = TYPE_MODE (rhs_type); | |
3716 | fltsz = GET_MODE_SIZE (TYPE_MODE (lhs_type)); | |
3717 | for (rhs_mode = GET_MODE_2XWIDER_MODE (TYPE_MODE (rhs_type)); | |
3718 | rhs_mode != VOIDmode && GET_MODE_SIZE (rhs_mode) <= fltsz; | |
3719 | rhs_mode = GET_MODE_2XWIDER_MODE (rhs_mode)) | |
3720 | { | |
3721 | cvt_type | |
3722 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3723 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3724 | if (cvt_type == NULL_TREE) | |
3725 | goto unsupported; | |
3726 | ||
3727 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3728 | { | |
3729 | if (!supportable_convert_operation (code, vectype_out, | |
3730 | cvt_type, &decl1, &codecvt1)) | |
3731 | goto unsupported; | |
3732 | } | |
3733 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
3734 | cvt_type, &codecvt1, |
3735 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
3736 | &interm_types)) |
3737 | continue; | |
3738 | else | |
3739 | gcc_assert (multi_step_cvt == 0); | |
3740 | ||
3741 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
3742 | vectype_in, &code1, &code2, |
3743 | &multi_step_cvt, &interm_types)) | |
4a00c761 JJ |
3744 | break; |
3745 | } | |
3746 | ||
3747 | if (rhs_mode == VOIDmode || GET_MODE_SIZE (rhs_mode) > fltsz) | |
3748 | goto unsupported; | |
3749 | ||
3750 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3751 | codecvt2 = ERROR_MARK; | |
3752 | else | |
3753 | { | |
3754 | multi_step_cvt++; | |
9771b263 | 3755 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
3756 | cvt_type = NULL_TREE; |
3757 | } | |
3758 | break; | |
3759 | ||
3760 | case NARROW: | |
3761 | gcc_assert (op_type == unary_op); | |
3762 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
3763 | &code1, &multi_step_cvt, | |
3764 | &interm_types)) | |
3765 | break; | |
3766 | ||
3767 | if (code != FIX_TRUNC_EXPR | |
3768 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3769 | >= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3770 | goto unsupported; | |
3771 | ||
3772 | rhs_mode = TYPE_MODE (rhs_type); | |
3773 | cvt_type | |
3774 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3775 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3776 | if (cvt_type == NULL_TREE) | |
3777 | goto unsupported; | |
3778 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
3779 | &decl1, &codecvt1)) | |
3780 | goto unsupported; | |
3781 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
3782 | &code1, &multi_step_cvt, | |
3783 | &interm_types)) | |
3784 | break; | |
3785 | goto unsupported; | |
3786 | ||
3787 | default: | |
3788 | gcc_unreachable (); | |
ebfd146a IR |
3789 | } |
3790 | ||
3791 | if (!vec_stmt) /* transformation not required. */ | |
3792 | { | |
73fbfcad | 3793 | if (dump_enabled_p ()) |
78c60e3d | 3794 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3795 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 3796 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
3797 | { |
3798 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
c3e7ee41 | 3799 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
8bd37302 | 3800 | } |
4a00c761 JJ |
3801 | else if (modifier == NARROW) |
3802 | { | |
3803 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 3804 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
3805 | } |
3806 | else | |
3807 | { | |
3808 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 3809 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 3810 | } |
9771b263 | 3811 | interm_types.release (); |
ebfd146a IR |
3812 | return true; |
3813 | } | |
3814 | ||
3815 | /** Transform. **/ | |
73fbfcad | 3816 | if (dump_enabled_p ()) |
78c60e3d | 3817 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3818 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 3819 | |
4a00c761 JJ |
3820 | if (op_type == binary_op) |
3821 | { | |
3822 | if (CONSTANT_CLASS_P (op0)) | |
3823 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
3824 | else if (CONSTANT_CLASS_P (op1)) | |
3825 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
3826 | } | |
3827 | ||
3828 | /* In case of multi-step conversion, we first generate conversion operations | |
3829 | to the intermediate types, and then from that types to the final one. | |
3830 | We create vector destinations for the intermediate type (TYPES) received | |
3831 | from supportable_*_operation, and store them in the correct order | |
3832 | for future use in vect_create_vectorized_*_stmts (). */ | |
9771b263 | 3833 | vec_dsts.create (multi_step_cvt + 1); |
82294ec1 JJ |
3834 | vec_dest = vect_create_destination_var (scalar_dest, |
3835 | (cvt_type && modifier == WIDEN) | |
3836 | ? cvt_type : vectype_out); | |
9771b263 | 3837 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3838 | |
3839 | if (multi_step_cvt) | |
3840 | { | |
9771b263 DN |
3841 | for (i = interm_types.length () - 1; |
3842 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
3843 | { |
3844 | vec_dest = vect_create_destination_var (scalar_dest, | |
3845 | intermediate_type); | |
9771b263 | 3846 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3847 | } |
3848 | } | |
ebfd146a | 3849 | |
4a00c761 | 3850 | if (cvt_type) |
82294ec1 JJ |
3851 | vec_dest = vect_create_destination_var (scalar_dest, |
3852 | modifier == WIDEN | |
3853 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
3854 | |
3855 | if (!slp_node) | |
3856 | { | |
30862efc | 3857 | if (modifier == WIDEN) |
4a00c761 | 3858 | { |
c3284718 | 3859 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 3860 | if (op_type == binary_op) |
9771b263 | 3861 | vec_oprnds1.create (1); |
4a00c761 | 3862 | } |
30862efc | 3863 | else if (modifier == NARROW) |
9771b263 DN |
3864 | vec_oprnds0.create ( |
3865 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
3866 | } |
3867 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 3868 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 3869 | |
4a00c761 | 3870 | last_oprnd = op0; |
ebfd146a IR |
3871 | prev_stmt_info = NULL; |
3872 | switch (modifier) | |
3873 | { | |
3874 | case NONE: | |
3875 | for (j = 0; j < ncopies; j++) | |
3876 | { | |
ebfd146a | 3877 | if (j == 0) |
d092494c IR |
3878 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node, |
3879 | -1); | |
ebfd146a IR |
3880 | else |
3881 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
3882 | ||
9771b263 | 3883 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
3884 | { |
3885 | /* Arguments are ready, create the new vector stmt. */ | |
3886 | if (code1 == CALL_EXPR) | |
3887 | { | |
3888 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
3889 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3890 | gimple_call_set_lhs (new_stmt, new_temp); | |
3891 | } | |
3892 | else | |
3893 | { | |
3894 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
0d0e4a03 | 3895 | new_stmt = gimple_build_assign (vec_dest, code1, vop0); |
4a00c761 JJ |
3896 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3897 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3898 | } | |
3899 | ||
3900 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3901 | if (slp_node) | |
9771b263 | 3902 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
225ce44b RB |
3903 | else |
3904 | { | |
3905 | if (!prev_stmt_info) | |
3906 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3907 | else | |
3908 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3909 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3910 | } | |
4a00c761 | 3911 | } |
ebfd146a IR |
3912 | } |
3913 | break; | |
3914 | ||
3915 | case WIDEN: | |
3916 | /* In case the vectorization factor (VF) is bigger than the number | |
3917 | of elements that we can fit in a vectype (nunits), we have to | |
3918 | generate more than one vector stmt - i.e - we need to "unroll" | |
3919 | the vector stmt by a factor VF/nunits. */ | |
3920 | for (j = 0; j < ncopies; j++) | |
3921 | { | |
4a00c761 | 3922 | /* Handle uses. */ |
ebfd146a | 3923 | if (j == 0) |
4a00c761 JJ |
3924 | { |
3925 | if (slp_node) | |
3926 | { | |
3927 | if (code == WIDEN_LSHIFT_EXPR) | |
3928 | { | |
3929 | unsigned int k; | |
ebfd146a | 3930 | |
4a00c761 JJ |
3931 | vec_oprnd1 = op1; |
3932 | /* Store vec_oprnd1 for every vector stmt to be created | |
3933 | for SLP_NODE. We check during the analysis that all | |
3934 | the shift arguments are the same. */ | |
3935 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 3936 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
3937 | |
3938 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
3939 | slp_node, -1); | |
3940 | } | |
3941 | else | |
3942 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
3943 | &vec_oprnds1, slp_node, -1); | |
3944 | } | |
3945 | else | |
3946 | { | |
81c40241 | 3947 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 3948 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
3949 | if (op_type == binary_op) |
3950 | { | |
3951 | if (code == WIDEN_LSHIFT_EXPR) | |
3952 | vec_oprnd1 = op1; | |
3953 | else | |
81c40241 | 3954 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 3955 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
3956 | } |
3957 | } | |
3958 | } | |
ebfd146a | 3959 | else |
4a00c761 JJ |
3960 | { |
3961 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
3962 | vec_oprnds0.truncate (0); |
3963 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
3964 | if (op_type == binary_op) |
3965 | { | |
3966 | if (code == WIDEN_LSHIFT_EXPR) | |
3967 | vec_oprnd1 = op1; | |
3968 | else | |
3969 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
3970 | vec_oprnd1); | |
9771b263 DN |
3971 | vec_oprnds1.truncate (0); |
3972 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
3973 | } |
3974 | } | |
ebfd146a | 3975 | |
4a00c761 JJ |
3976 | /* Arguments are ready. Create the new vector stmts. */ |
3977 | for (i = multi_step_cvt; i >= 0; i--) | |
3978 | { | |
9771b263 | 3979 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
3980 | enum tree_code c1 = code1, c2 = code2; |
3981 | if (i == 0 && codecvt2 != ERROR_MARK) | |
3982 | { | |
3983 | c1 = codecvt1; | |
3984 | c2 = codecvt2; | |
3985 | } | |
3986 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
3987 | &vec_oprnds1, | |
3988 | stmt, this_dest, gsi, | |
3989 | c1, c2, decl1, decl2, | |
3990 | op_type); | |
3991 | } | |
3992 | ||
9771b263 | 3993 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
3994 | { |
3995 | if (cvt_type) | |
3996 | { | |
3997 | if (codecvt1 == CALL_EXPR) | |
3998 | { | |
3999 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4000 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4001 | gimple_call_set_lhs (new_stmt, new_temp); | |
4002 | } | |
4003 | else | |
4004 | { | |
4005 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4006 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4007 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4008 | vop0); | |
4a00c761 JJ |
4009 | } |
4010 | ||
4011 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4012 | } | |
4013 | else | |
4014 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
4015 | ||
4016 | if (slp_node) | |
9771b263 | 4017 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4018 | else |
c689ce1e RB |
4019 | { |
4020 | if (!prev_stmt_info) | |
4021 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4022 | else | |
4023 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4024 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4025 | } | |
4a00c761 | 4026 | } |
ebfd146a | 4027 | } |
4a00c761 JJ |
4028 | |
4029 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
4030 | break; |
4031 | ||
4032 | case NARROW: | |
4033 | /* In case the vectorization factor (VF) is bigger than the number | |
4034 | of elements that we can fit in a vectype (nunits), we have to | |
4035 | generate more than one vector stmt - i.e - we need to "unroll" | |
4036 | the vector stmt by a factor VF/nunits. */ | |
4037 | for (j = 0; j < ncopies; j++) | |
4038 | { | |
4039 | /* Handle uses. */ | |
4a00c761 JJ |
4040 | if (slp_node) |
4041 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
4042 | slp_node, -1); | |
ebfd146a IR |
4043 | else |
4044 | { | |
9771b263 | 4045 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
4046 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
4047 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
4048 | } |
4049 | ||
4a00c761 JJ |
4050 | /* Arguments are ready. Create the new vector stmts. */ |
4051 | if (cvt_type) | |
9771b263 | 4052 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4053 | { |
4054 | if (codecvt1 == CALL_EXPR) | |
4055 | { | |
4056 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4057 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4058 | gimple_call_set_lhs (new_stmt, new_temp); | |
4059 | } | |
4060 | else | |
4061 | { | |
4062 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4063 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4064 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4065 | vop0); | |
4a00c761 | 4066 | } |
ebfd146a | 4067 | |
4a00c761 | 4068 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 4069 | vec_oprnds0[i] = new_temp; |
4a00c761 | 4070 | } |
ebfd146a | 4071 | |
4a00c761 JJ |
4072 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
4073 | stmt, vec_dsts, gsi, | |
4074 | slp_node, code1, | |
4075 | &prev_stmt_info); | |
ebfd146a IR |
4076 | } |
4077 | ||
4078 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 4079 | break; |
ebfd146a IR |
4080 | } |
4081 | ||
9771b263 DN |
4082 | vec_oprnds0.release (); |
4083 | vec_oprnds1.release (); | |
4084 | vec_dsts.release (); | |
4085 | interm_types.release (); | |
ebfd146a IR |
4086 | |
4087 | return true; | |
4088 | } | |
ff802fa1 IR |
4089 | |
4090 | ||
ebfd146a IR |
4091 | /* Function vectorizable_assignment. |
4092 | ||
b8698a0f L |
4093 | Check if STMT performs an assignment (copy) that can be vectorized. |
4094 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4095 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4096 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4097 | ||
4098 | static bool | |
355fe088 TS |
4099 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
4100 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4101 | { |
4102 | tree vec_dest; | |
4103 | tree scalar_dest; | |
4104 | tree op; | |
4105 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
4106 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4107 | tree new_temp; | |
355fe088 | 4108 | gimple *def_stmt; |
ebfd146a | 4109 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
ebfd146a | 4110 | int ncopies; |
f18b55bd | 4111 | int i, j; |
6e1aa848 | 4112 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 4113 | tree vop; |
a70d6342 | 4114 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4115 | vec_info *vinfo = stmt_info->vinfo; |
355fe088 | 4116 | gimple *new_stmt = NULL; |
f18b55bd | 4117 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
4118 | enum tree_code code; |
4119 | tree vectype_in; | |
ebfd146a | 4120 | |
a70d6342 | 4121 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4122 | return false; |
4123 | ||
8644a673 | 4124 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4125 | return false; |
4126 | ||
4127 | /* Is vectorizable assignment? */ | |
4128 | if (!is_gimple_assign (stmt)) | |
4129 | return false; | |
4130 | ||
4131 | scalar_dest = gimple_assign_lhs (stmt); | |
4132 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4133 | return false; | |
4134 | ||
fde9c428 | 4135 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 4136 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
4137 | || code == PAREN_EXPR |
4138 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
4139 | op = gimple_assign_rhs1 (stmt); |
4140 | else | |
4141 | return false; | |
4142 | ||
7b7ec6c5 RG |
4143 | if (code == VIEW_CONVERT_EXPR) |
4144 | op = TREE_OPERAND (op, 0); | |
4145 | ||
465c8c19 JJ |
4146 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
4147 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4148 | ||
4149 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4150 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4151 | case of SLP. */ | |
4152 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
4153 | ncopies = 1; | |
4154 | else | |
4155 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4156 | ||
4157 | gcc_assert (ncopies >= 1); | |
4158 | ||
81c40241 | 4159 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[0], &vectype_in)) |
ebfd146a | 4160 | { |
73fbfcad | 4161 | if (dump_enabled_p ()) |
78c60e3d | 4162 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4163 | "use not simple.\n"); |
ebfd146a IR |
4164 | return false; |
4165 | } | |
4166 | ||
fde9c428 RG |
4167 | /* We can handle NOP_EXPR conversions that do not change the number |
4168 | of elements or the vector size. */ | |
7b7ec6c5 RG |
4169 | if ((CONVERT_EXPR_CODE_P (code) |
4170 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
4171 | && (!vectype_in |
4172 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
4173 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
4174 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
4175 | return false; | |
4176 | ||
7b7b1813 RG |
4177 | /* We do not handle bit-precision changes. */ |
4178 | if ((CONVERT_EXPR_CODE_P (code) | |
4179 | || code == VIEW_CONVERT_EXPR) | |
4180 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
4181 | && ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4182 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4183 | || ((TYPE_PRECISION (TREE_TYPE (op)) | |
4184 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))) | |
4185 | /* But a conversion that does not change the bit-pattern is ok. */ | |
4186 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4187 | > TYPE_PRECISION (TREE_TYPE (op))) | |
4188 | && TYPE_UNSIGNED (TREE_TYPE (op)))) | |
4189 | { | |
73fbfcad | 4190 | if (dump_enabled_p ()) |
78c60e3d SS |
4191 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4192 | "type conversion to/from bit-precision " | |
e645e942 | 4193 | "unsupported.\n"); |
7b7b1813 RG |
4194 | return false; |
4195 | } | |
4196 | ||
ebfd146a IR |
4197 | if (!vec_stmt) /* transformation not required. */ |
4198 | { | |
4199 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 4200 | if (dump_enabled_p ()) |
78c60e3d | 4201 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4202 | "=== vectorizable_assignment ===\n"); |
c3e7ee41 | 4203 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
4204 | return true; |
4205 | } | |
4206 | ||
4207 | /** Transform. **/ | |
73fbfcad | 4208 | if (dump_enabled_p ()) |
e645e942 | 4209 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
4210 | |
4211 | /* Handle def. */ | |
4212 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4213 | ||
4214 | /* Handle use. */ | |
f18b55bd | 4215 | for (j = 0; j < ncopies; j++) |
ebfd146a | 4216 | { |
f18b55bd IR |
4217 | /* Handle uses. */ |
4218 | if (j == 0) | |
d092494c | 4219 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node, -1); |
f18b55bd IR |
4220 | else |
4221 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
4222 | ||
4223 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 4224 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 4225 | { |
7b7ec6c5 RG |
4226 | if (CONVERT_EXPR_CODE_P (code) |
4227 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 4228 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
4229 | new_stmt = gimple_build_assign (vec_dest, vop); |
4230 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4231 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4232 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4233 | if (slp_node) | |
9771b263 | 4234 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 4235 | } |
ebfd146a IR |
4236 | |
4237 | if (slp_node) | |
f18b55bd IR |
4238 | continue; |
4239 | ||
4240 | if (j == 0) | |
4241 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4242 | else | |
4243 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4244 | ||
4245 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4246 | } | |
b8698a0f | 4247 | |
9771b263 | 4248 | vec_oprnds.release (); |
ebfd146a IR |
4249 | return true; |
4250 | } | |
4251 | ||
9dc3f7de | 4252 | |
1107f3ae IR |
4253 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
4254 | either as shift by a scalar or by a vector. */ | |
4255 | ||
4256 | bool | |
4257 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
4258 | { | |
4259 | ||
ef4bddc2 | 4260 | machine_mode vec_mode; |
1107f3ae IR |
4261 | optab optab; |
4262 | int icode; | |
4263 | tree vectype; | |
4264 | ||
4265 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4266 | if (!vectype) | |
4267 | return false; | |
4268 | ||
4269 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4270 | if (!optab | |
4271 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
4272 | { | |
4273 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4274 | if (!optab | |
4275 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
4276 | == CODE_FOR_nothing)) | |
4277 | return false; | |
4278 | } | |
4279 | ||
4280 | vec_mode = TYPE_MODE (vectype); | |
4281 | icode = (int) optab_handler (optab, vec_mode); | |
4282 | if (icode == CODE_FOR_nothing) | |
4283 | return false; | |
4284 | ||
4285 | return true; | |
4286 | } | |
4287 | ||
4288 | ||
9dc3f7de IR |
4289 | /* Function vectorizable_shift. |
4290 | ||
4291 | Check if STMT performs a shift operation that can be vectorized. | |
4292 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4293 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
4294 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4295 | ||
4296 | static bool | |
355fe088 TS |
4297 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
4298 | gimple **vec_stmt, slp_tree slp_node) | |
9dc3f7de IR |
4299 | { |
4300 | tree vec_dest; | |
4301 | tree scalar_dest; | |
4302 | tree op0, op1 = NULL; | |
4303 | tree vec_oprnd1 = NULL_TREE; | |
4304 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4305 | tree vectype; | |
4306 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4307 | enum tree_code code; | |
ef4bddc2 | 4308 | machine_mode vec_mode; |
9dc3f7de IR |
4309 | tree new_temp; |
4310 | optab optab; | |
4311 | int icode; | |
ef4bddc2 | 4312 | machine_mode optab_op2_mode; |
355fe088 | 4313 | gimple *def_stmt; |
9dc3f7de | 4314 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
355fe088 | 4315 | gimple *new_stmt = NULL; |
9dc3f7de IR |
4316 | stmt_vec_info prev_stmt_info; |
4317 | int nunits_in; | |
4318 | int nunits_out; | |
4319 | tree vectype_out; | |
cede2577 | 4320 | tree op1_vectype; |
9dc3f7de IR |
4321 | int ncopies; |
4322 | int j, i; | |
6e1aa848 DN |
4323 | vec<tree> vec_oprnds0 = vNULL; |
4324 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
4325 | tree vop0, vop1; |
4326 | unsigned int k; | |
49eab32e | 4327 | bool scalar_shift_arg = true; |
9dc3f7de | 4328 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4329 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
4330 | int vf; |
4331 | ||
4332 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
4333 | return false; | |
4334 | ||
4335 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
4336 | return false; | |
4337 | ||
4338 | /* Is STMT a vectorizable binary/unary operation? */ | |
4339 | if (!is_gimple_assign (stmt)) | |
4340 | return false; | |
4341 | ||
4342 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4343 | return false; | |
4344 | ||
4345 | code = gimple_assign_rhs_code (stmt); | |
4346 | ||
4347 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
4348 | || code == RROTATE_EXPR)) | |
4349 | return false; | |
4350 | ||
4351 | scalar_dest = gimple_assign_lhs (stmt); | |
4352 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
7b7b1813 RG |
4353 | if (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
4354 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4355 | { | |
73fbfcad | 4356 | if (dump_enabled_p ()) |
78c60e3d | 4357 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4358 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
4359 | return false; |
4360 | } | |
9dc3f7de IR |
4361 | |
4362 | op0 = gimple_assign_rhs1 (stmt); | |
81c40241 | 4363 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
9dc3f7de | 4364 | { |
73fbfcad | 4365 | if (dump_enabled_p ()) |
78c60e3d | 4366 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4367 | "use not simple.\n"); |
9dc3f7de IR |
4368 | return false; |
4369 | } | |
4370 | /* If op0 is an external or constant def use a vector type with | |
4371 | the same size as the output vector type. */ | |
4372 | if (!vectype) | |
4373 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4374 | if (vec_stmt) | |
4375 | gcc_assert (vectype); | |
4376 | if (!vectype) | |
4377 | { | |
73fbfcad | 4378 | if (dump_enabled_p ()) |
78c60e3d | 4379 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4380 | "no vectype for scalar type\n"); |
9dc3f7de IR |
4381 | return false; |
4382 | } | |
4383 | ||
4384 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4385 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4386 | if (nunits_out != nunits_in) | |
4387 | return false; | |
4388 | ||
4389 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4390 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &op1_vectype)) |
9dc3f7de | 4391 | { |
73fbfcad | 4392 | if (dump_enabled_p ()) |
78c60e3d | 4393 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4394 | "use not simple.\n"); |
9dc3f7de IR |
4395 | return false; |
4396 | } | |
4397 | ||
4398 | if (loop_vinfo) | |
4399 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4400 | else | |
4401 | vf = 1; | |
4402 | ||
4403 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4404 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4405 | case of SLP. */ | |
437f4a00 | 4406 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
9dc3f7de IR |
4407 | ncopies = 1; |
4408 | else | |
4409 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4410 | ||
4411 | gcc_assert (ncopies >= 1); | |
4412 | ||
4413 | /* Determine whether the shift amount is a vector, or scalar. If the | |
4414 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
4415 | ||
dbfa87aa YR |
4416 | if ((dt[1] == vect_internal_def |
4417 | || dt[1] == vect_induction_def) | |
4418 | && !slp_node) | |
49eab32e JJ |
4419 | scalar_shift_arg = false; |
4420 | else if (dt[1] == vect_constant_def | |
4421 | || dt[1] == vect_external_def | |
4422 | || dt[1] == vect_internal_def) | |
4423 | { | |
4424 | /* In SLP, need to check whether the shift count is the same, | |
4425 | in loops if it is a constant or invariant, it is always | |
4426 | a scalar shift. */ | |
4427 | if (slp_node) | |
4428 | { | |
355fe088 TS |
4429 | vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
4430 | gimple *slpstmt; | |
49eab32e | 4431 | |
9771b263 | 4432 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
4433 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
4434 | scalar_shift_arg = false; | |
4435 | } | |
4436 | } | |
4437 | else | |
4438 | { | |
73fbfcad | 4439 | if (dump_enabled_p ()) |
78c60e3d | 4440 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4441 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
4442 | return false; |
4443 | } | |
4444 | ||
9dc3f7de | 4445 | /* Vector shifted by vector. */ |
49eab32e | 4446 | if (!scalar_shift_arg) |
9dc3f7de IR |
4447 | { |
4448 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 4449 | if (dump_enabled_p ()) |
78c60e3d | 4450 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4451 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 4452 | |
aa948027 JJ |
4453 | if (!op1_vectype) |
4454 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
4455 | if (op1_vectype == NULL_TREE | |
4456 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 4457 | { |
73fbfcad | 4458 | if (dump_enabled_p ()) |
78c60e3d SS |
4459 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4460 | "unusable type for last operand in" | |
e645e942 | 4461 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
4462 | return false; |
4463 | } | |
9dc3f7de IR |
4464 | } |
4465 | /* See if the machine has a vector shifted by scalar insn and if not | |
4466 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 4467 | else |
9dc3f7de IR |
4468 | { |
4469 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4470 | if (optab | |
4471 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
4472 | { | |
73fbfcad | 4473 | if (dump_enabled_p ()) |
78c60e3d | 4474 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4475 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
4476 | } |
4477 | else | |
4478 | { | |
4479 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4480 | if (optab | |
4481 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
4482 | != CODE_FOR_nothing)) | |
4483 | { | |
49eab32e JJ |
4484 | scalar_shift_arg = false; |
4485 | ||
73fbfcad | 4486 | if (dump_enabled_p ()) |
78c60e3d | 4487 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4488 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
4489 | |
4490 | /* Unlike the other binary operators, shifts/rotates have | |
4491 | the rhs being int, instead of the same type as the lhs, | |
4492 | so make sure the scalar is the right type if we are | |
aa948027 | 4493 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
4494 | if (dt[1] == vect_constant_def) |
4495 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
4496 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
4497 | TREE_TYPE (op1))) | |
4498 | { | |
4499 | if (slp_node | |
4500 | && TYPE_MODE (TREE_TYPE (vectype)) | |
4501 | != TYPE_MODE (TREE_TYPE (op1))) | |
4502 | { | |
73fbfcad | 4503 | if (dump_enabled_p ()) |
78c60e3d SS |
4504 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4505 | "unusable type for last operand in" | |
e645e942 | 4506 | " vector/vector shift/rotate.\n"); |
21c0a521 | 4507 | return false; |
aa948027 JJ |
4508 | } |
4509 | if (vec_stmt && !slp_node) | |
4510 | { | |
4511 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
4512 | op1 = vect_init_vector (stmt, op1, | |
4513 | TREE_TYPE (vectype), NULL); | |
4514 | } | |
4515 | } | |
9dc3f7de IR |
4516 | } |
4517 | } | |
4518 | } | |
9dc3f7de IR |
4519 | |
4520 | /* Supportable by target? */ | |
4521 | if (!optab) | |
4522 | { | |
73fbfcad | 4523 | if (dump_enabled_p ()) |
78c60e3d | 4524 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4525 | "no optab.\n"); |
9dc3f7de IR |
4526 | return false; |
4527 | } | |
4528 | vec_mode = TYPE_MODE (vectype); | |
4529 | icode = (int) optab_handler (optab, vec_mode); | |
4530 | if (icode == CODE_FOR_nothing) | |
4531 | { | |
73fbfcad | 4532 | if (dump_enabled_p ()) |
78c60e3d | 4533 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4534 | "op not supported by target.\n"); |
9dc3f7de IR |
4535 | /* Check only during analysis. */ |
4536 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
4537 | || (vf < vect_min_worthwhile_factor (code) | |
4538 | && !vec_stmt)) | |
4539 | return false; | |
73fbfcad | 4540 | if (dump_enabled_p ()) |
e645e942 TJ |
4541 | dump_printf_loc (MSG_NOTE, vect_location, |
4542 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
4543 | } |
4544 | ||
4545 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
4546 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
4547 | && vf < vect_min_worthwhile_factor (code) | |
4548 | && !vec_stmt) | |
4549 | { | |
73fbfcad | 4550 | if (dump_enabled_p ()) |
78c60e3d | 4551 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4552 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
4553 | return false; |
4554 | } | |
4555 | ||
4556 | if (!vec_stmt) /* transformation not required. */ | |
4557 | { | |
4558 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 4559 | if (dump_enabled_p ()) |
e645e942 TJ |
4560 | dump_printf_loc (MSG_NOTE, vect_location, |
4561 | "=== vectorizable_shift ===\n"); | |
c3e7ee41 | 4562 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
9dc3f7de IR |
4563 | return true; |
4564 | } | |
4565 | ||
4566 | /** Transform. **/ | |
4567 | ||
73fbfcad | 4568 | if (dump_enabled_p ()) |
78c60e3d | 4569 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4570 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
4571 | |
4572 | /* Handle def. */ | |
4573 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4574 | ||
9dc3f7de IR |
4575 | prev_stmt_info = NULL; |
4576 | for (j = 0; j < ncopies; j++) | |
4577 | { | |
4578 | /* Handle uses. */ | |
4579 | if (j == 0) | |
4580 | { | |
4581 | if (scalar_shift_arg) | |
4582 | { | |
4583 | /* Vector shl and shr insn patterns can be defined with scalar | |
4584 | operand 2 (shift operand). In this case, use constant or loop | |
4585 | invariant op1 directly, without extending it to vector mode | |
4586 | first. */ | |
4587 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
4588 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
4589 | { | |
73fbfcad | 4590 | if (dump_enabled_p ()) |
78c60e3d | 4591 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4592 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 4593 | vec_oprnd1 = op1; |
8930f723 | 4594 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 4595 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4596 | if (slp_node) |
4597 | { | |
4598 | /* Store vec_oprnd1 for every vector stmt to be created | |
4599 | for SLP_NODE. We check during the analysis that all | |
4600 | the shift arguments are the same. | |
4601 | TODO: Allow different constants for different vector | |
4602 | stmts generated for an SLP instance. */ | |
4603 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4604 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4605 | } |
4606 | } | |
4607 | } | |
4608 | ||
4609 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
4610 | (a special case for certain kind of vector shifts); otherwise, | |
4611 | operand 1 should be of a vector type (the usual case). */ | |
4612 | if (vec_oprnd1) | |
4613 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
d092494c | 4614 | slp_node, -1); |
9dc3f7de IR |
4615 | else |
4616 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
d092494c | 4617 | slp_node, -1); |
9dc3f7de IR |
4618 | } |
4619 | else | |
4620 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
4621 | ||
4622 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 4623 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 4624 | { |
9771b263 | 4625 | vop1 = vec_oprnds1[i]; |
0d0e4a03 | 4626 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
4627 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4628 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4629 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4630 | if (slp_node) | |
9771b263 | 4631 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
4632 | } |
4633 | ||
4634 | if (slp_node) | |
4635 | continue; | |
4636 | ||
4637 | if (j == 0) | |
4638 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4639 | else | |
4640 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4641 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4642 | } | |
4643 | ||
9771b263 DN |
4644 | vec_oprnds0.release (); |
4645 | vec_oprnds1.release (); | |
9dc3f7de IR |
4646 | |
4647 | return true; | |
4648 | } | |
4649 | ||
4650 | ||
ebfd146a IR |
4651 | /* Function vectorizable_operation. |
4652 | ||
16949072 RG |
4653 | Check if STMT performs a binary, unary or ternary operation that can |
4654 | be vectorized. | |
b8698a0f | 4655 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
4656 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4657 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4658 | ||
4659 | static bool | |
355fe088 TS |
4660 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
4661 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a | 4662 | { |
00f07b86 | 4663 | tree vec_dest; |
ebfd146a | 4664 | tree scalar_dest; |
16949072 | 4665 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 4666 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 4667 | tree vectype; |
ebfd146a IR |
4668 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4669 | enum tree_code code; | |
ef4bddc2 | 4670 | machine_mode vec_mode; |
ebfd146a IR |
4671 | tree new_temp; |
4672 | int op_type; | |
00f07b86 | 4673 | optab optab; |
523ba738 | 4674 | bool target_support_p; |
355fe088 | 4675 | gimple *def_stmt; |
16949072 RG |
4676 | enum vect_def_type dt[3] |
4677 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
355fe088 | 4678 | gimple *new_stmt = NULL; |
ebfd146a | 4679 | stmt_vec_info prev_stmt_info; |
b690cc0f | 4680 | int nunits_in; |
ebfd146a IR |
4681 | int nunits_out; |
4682 | tree vectype_out; | |
4683 | int ncopies; | |
4684 | int j, i; | |
6e1aa848 DN |
4685 | vec<tree> vec_oprnds0 = vNULL; |
4686 | vec<tree> vec_oprnds1 = vNULL; | |
4687 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 4688 | tree vop0, vop1, vop2; |
a70d6342 | 4689 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4690 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 IR |
4691 | int vf; |
4692 | ||
a70d6342 | 4693 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4694 | return false; |
4695 | ||
8644a673 | 4696 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4697 | return false; |
4698 | ||
4699 | /* Is STMT a vectorizable binary/unary operation? */ | |
4700 | if (!is_gimple_assign (stmt)) | |
4701 | return false; | |
4702 | ||
4703 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4704 | return false; | |
4705 | ||
ebfd146a IR |
4706 | code = gimple_assign_rhs_code (stmt); |
4707 | ||
4708 | /* For pointer addition, we should use the normal plus for | |
4709 | the vector addition. */ | |
4710 | if (code == POINTER_PLUS_EXPR) | |
4711 | code = PLUS_EXPR; | |
4712 | ||
4713 | /* Support only unary or binary operations. */ | |
4714 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 4715 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 4716 | { |
73fbfcad | 4717 | if (dump_enabled_p ()) |
78c60e3d | 4718 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4719 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 4720 | op_type); |
ebfd146a IR |
4721 | return false; |
4722 | } | |
4723 | ||
b690cc0f RG |
4724 | scalar_dest = gimple_assign_lhs (stmt); |
4725 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
4726 | ||
7b7b1813 RG |
4727 | /* Most operations cannot handle bit-precision types without extra |
4728 | truncations. */ | |
045c1278 IE |
4729 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4730 | && (TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4731 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
7b7b1813 RG |
4732 | /* Exception are bitwise binary operations. */ |
4733 | && code != BIT_IOR_EXPR | |
4734 | && code != BIT_XOR_EXPR | |
4735 | && code != BIT_AND_EXPR) | |
4736 | { | |
73fbfcad | 4737 | if (dump_enabled_p ()) |
78c60e3d | 4738 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4739 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
4740 | return false; |
4741 | } | |
4742 | ||
ebfd146a | 4743 | op0 = gimple_assign_rhs1 (stmt); |
81c40241 | 4744 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
ebfd146a | 4745 | { |
73fbfcad | 4746 | if (dump_enabled_p ()) |
78c60e3d | 4747 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4748 | "use not simple.\n"); |
ebfd146a IR |
4749 | return false; |
4750 | } | |
b690cc0f RG |
4751 | /* If op0 is an external or constant def use a vector type with |
4752 | the same size as the output vector type. */ | |
4753 | if (!vectype) | |
b036c6c5 IE |
4754 | { |
4755 | /* For boolean type we cannot determine vectype by | |
4756 | invariant value (don't know whether it is a vector | |
4757 | of booleans or vector of integers). We use output | |
4758 | vectype because operations on boolean don't change | |
4759 | type. */ | |
4760 | if (TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE) | |
4761 | { | |
4762 | if (TREE_CODE (TREE_TYPE (scalar_dest)) != BOOLEAN_TYPE) | |
4763 | { | |
4764 | if (dump_enabled_p ()) | |
4765 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4766 | "not supported operation on bool value.\n"); | |
4767 | return false; | |
4768 | } | |
4769 | vectype = vectype_out; | |
4770 | } | |
4771 | else | |
4772 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4773 | } | |
7d8930a0 IR |
4774 | if (vec_stmt) |
4775 | gcc_assert (vectype); | |
4776 | if (!vectype) | |
4777 | { | |
73fbfcad | 4778 | if (dump_enabled_p ()) |
7d8930a0 | 4779 | { |
78c60e3d SS |
4780 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4781 | "no vectype for scalar type "); | |
4782 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
4783 | TREE_TYPE (op0)); | |
e645e942 | 4784 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
4785 | } |
4786 | ||
4787 | return false; | |
4788 | } | |
b690cc0f RG |
4789 | |
4790 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4791 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4792 | if (nunits_out != nunits_in) | |
4793 | return false; | |
ebfd146a | 4794 | |
16949072 | 4795 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
4796 | { |
4797 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4798 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1])) |
ebfd146a | 4799 | { |
73fbfcad | 4800 | if (dump_enabled_p ()) |
78c60e3d | 4801 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4802 | "use not simple.\n"); |
ebfd146a IR |
4803 | return false; |
4804 | } | |
4805 | } | |
16949072 RG |
4806 | if (op_type == ternary_op) |
4807 | { | |
4808 | op2 = gimple_assign_rhs3 (stmt); | |
81c40241 | 4809 | if (!vect_is_simple_use (op2, vinfo, &def_stmt, &dt[2])) |
16949072 | 4810 | { |
73fbfcad | 4811 | if (dump_enabled_p ()) |
78c60e3d | 4812 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4813 | "use not simple.\n"); |
16949072 RG |
4814 | return false; |
4815 | } | |
4816 | } | |
ebfd146a | 4817 | |
b690cc0f RG |
4818 | if (loop_vinfo) |
4819 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4820 | else | |
4821 | vf = 1; | |
4822 | ||
4823 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4824 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 4825 | case of SLP. */ |
437f4a00 | 4826 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
b690cc0f RG |
4827 | ncopies = 1; |
4828 | else | |
4829 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4830 | ||
4831 | gcc_assert (ncopies >= 1); | |
4832 | ||
9dc3f7de | 4833 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
4834 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
4835 | || code == RROTATE_EXPR) | |
9dc3f7de | 4836 | return false; |
ebfd146a | 4837 | |
ebfd146a | 4838 | /* Supportable by target? */ |
00f07b86 RH |
4839 | |
4840 | vec_mode = TYPE_MODE (vectype); | |
4841 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 4842 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
4843 | else |
4844 | { | |
4845 | optab = optab_for_tree_code (code, vectype, optab_default); | |
4846 | if (!optab) | |
5deb57cb | 4847 | { |
73fbfcad | 4848 | if (dump_enabled_p ()) |
78c60e3d | 4849 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4850 | "no optab.\n"); |
00f07b86 | 4851 | return false; |
5deb57cb | 4852 | } |
523ba738 RS |
4853 | target_support_p = (optab_handler (optab, vec_mode) |
4854 | != CODE_FOR_nothing); | |
5deb57cb JJ |
4855 | } |
4856 | ||
523ba738 | 4857 | if (!target_support_p) |
ebfd146a | 4858 | { |
73fbfcad | 4859 | if (dump_enabled_p ()) |
78c60e3d | 4860 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4861 | "op not supported by target.\n"); |
ebfd146a IR |
4862 | /* Check only during analysis. */ |
4863 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5deb57cb | 4864 | || (!vec_stmt && vf < vect_min_worthwhile_factor (code))) |
ebfd146a | 4865 | return false; |
73fbfcad | 4866 | if (dump_enabled_p ()) |
e645e942 TJ |
4867 | dump_printf_loc (MSG_NOTE, vect_location, |
4868 | "proceeding using word mode.\n"); | |
383d9c83 IR |
4869 | } |
4870 | ||
4a00c761 | 4871 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
4872 | if (!VECTOR_MODE_P (vec_mode) |
4873 | && !vec_stmt | |
4874 | && vf < vect_min_worthwhile_factor (code)) | |
7d8930a0 | 4875 | { |
73fbfcad | 4876 | if (dump_enabled_p ()) |
78c60e3d | 4877 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4878 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 4879 | return false; |
7d8930a0 | 4880 | } |
ebfd146a | 4881 | |
ebfd146a IR |
4882 | if (!vec_stmt) /* transformation not required. */ |
4883 | { | |
4a00c761 | 4884 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 4885 | if (dump_enabled_p ()) |
78c60e3d | 4886 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4887 | "=== vectorizable_operation ===\n"); |
c3e7ee41 | 4888 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
4889 | return true; |
4890 | } | |
4891 | ||
4892 | /** Transform. **/ | |
4893 | ||
73fbfcad | 4894 | if (dump_enabled_p ()) |
78c60e3d | 4895 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4896 | "transform binary/unary operation.\n"); |
383d9c83 | 4897 | |
ebfd146a | 4898 | /* Handle def. */ |
00f07b86 | 4899 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 4900 | |
ebfd146a IR |
4901 | /* In case the vectorization factor (VF) is bigger than the number |
4902 | of elements that we can fit in a vectype (nunits), we have to generate | |
4903 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
4904 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
4905 | from one copy of the vector stmt to the next, in the field | |
4906 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
4907 | stages to find the correct vector defs to be used when vectorizing | |
4908 | stmts that use the defs of the current stmt. The example below | |
4909 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
4910 | we need to create 4 vectorized stmts): | |
4911 | ||
4912 | before vectorization: | |
4913 | RELATED_STMT VEC_STMT | |
4914 | S1: x = memref - - | |
4915 | S2: z = x + 1 - - | |
4916 | ||
4917 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
4918 | there): | |
4919 | RELATED_STMT VEC_STMT | |
4920 | VS1_0: vx0 = memref0 VS1_1 - | |
4921 | VS1_1: vx1 = memref1 VS1_2 - | |
4922 | VS1_2: vx2 = memref2 VS1_3 - | |
4923 | VS1_3: vx3 = memref3 - - | |
4924 | S1: x = load - VS1_0 | |
4925 | S2: z = x + 1 - - | |
4926 | ||
4927 | step2: vectorize stmt S2 (done here): | |
4928 | To vectorize stmt S2 we first need to find the relevant vector | |
4929 | def for the first operand 'x'. This is, as usual, obtained from | |
4930 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
4931 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
4932 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
4933 | the vector stmt VS2_0, and as usual, record it in the | |
4934 | STMT_VINFO_VEC_STMT of stmt S2. | |
4935 | When creating the second copy (VS2_1), we obtain the relevant vector | |
4936 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
4937 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
4938 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
4939 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
4940 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
4941 | chain of stmts and pointers: | |
4942 | RELATED_STMT VEC_STMT | |
4943 | VS1_0: vx0 = memref0 VS1_1 - | |
4944 | VS1_1: vx1 = memref1 VS1_2 - | |
4945 | VS1_2: vx2 = memref2 VS1_3 - | |
4946 | VS1_3: vx3 = memref3 - - | |
4947 | S1: x = load - VS1_0 | |
4948 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
4949 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
4950 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
4951 | VS2_3: vz3 = vx3 + v1 - - | |
4952 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
4953 | |
4954 | prev_stmt_info = NULL; | |
4955 | for (j = 0; j < ncopies; j++) | |
4956 | { | |
4957 | /* Handle uses. */ | |
4958 | if (j == 0) | |
4a00c761 JJ |
4959 | { |
4960 | if (op_type == binary_op || op_type == ternary_op) | |
4961 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
4962 | slp_node, -1); | |
4963 | else | |
4964 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
4965 | slp_node, -1); | |
4966 | if (op_type == ternary_op) | |
36ba4aae | 4967 | { |
9771b263 DN |
4968 | vec_oprnds2.create (1); |
4969 | vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2, | |
81c40241 | 4970 | stmt)); |
36ba4aae | 4971 | } |
4a00c761 | 4972 | } |
ebfd146a | 4973 | else |
4a00c761 JJ |
4974 | { |
4975 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
4976 | if (op_type == ternary_op) | |
4977 | { | |
9771b263 DN |
4978 | tree vec_oprnd = vec_oprnds2.pop (); |
4979 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
4980 | vec_oprnd)); | |
4a00c761 JJ |
4981 | } |
4982 | } | |
4983 | ||
4984 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 4985 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 4986 | { |
4a00c761 | 4987 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 4988 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 4989 | vop2 = ((op_type == ternary_op) |
9771b263 | 4990 | ? vec_oprnds2[i] : NULL_TREE); |
0d0e4a03 | 4991 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
4a00c761 JJ |
4992 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4993 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4994 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4995 | if (slp_node) | |
9771b263 | 4996 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
4997 | } |
4998 | ||
4a00c761 JJ |
4999 | if (slp_node) |
5000 | continue; | |
5001 | ||
5002 | if (j == 0) | |
5003 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5004 | else | |
5005 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5006 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
5007 | } |
5008 | ||
9771b263 DN |
5009 | vec_oprnds0.release (); |
5010 | vec_oprnds1.release (); | |
5011 | vec_oprnds2.release (); | |
ebfd146a | 5012 | |
ebfd146a IR |
5013 | return true; |
5014 | } | |
5015 | ||
c716e67f XDL |
5016 | /* A helper function to ensure data reference DR's base alignment |
5017 | for STMT_INFO. */ | |
5018 | ||
5019 | static void | |
5020 | ensure_base_align (stmt_vec_info stmt_info, struct data_reference *dr) | |
5021 | { | |
5022 | if (!dr->aux) | |
5023 | return; | |
5024 | ||
52639a61 | 5025 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f XDL |
5026 | { |
5027 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
52639a61 | 5028 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 5029 | |
428f0c67 JH |
5030 | if (decl_in_symtab_p (base_decl)) |
5031 | symtab_node::get (base_decl)->increase_alignment (TYPE_ALIGN (vectype)); | |
5032 | else | |
5033 | { | |
5034 | DECL_ALIGN (base_decl) = TYPE_ALIGN (vectype); | |
5035 | DECL_USER_ALIGN (base_decl) = 1; | |
5036 | } | |
52639a61 | 5037 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
5038 | } |
5039 | } | |
5040 | ||
ebfd146a | 5041 | |
09dfa495 BM |
5042 | /* Given a vector type VECTYPE returns the VECTOR_CST mask that implements |
5043 | reversal of the vector elements. If that is impossible to do, | |
5044 | returns NULL. */ | |
5045 | ||
5046 | static tree | |
5047 | perm_mask_for_reverse (tree vectype) | |
5048 | { | |
5049 | int i, nunits; | |
5050 | unsigned char *sel; | |
5051 | ||
5052 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
5053 | sel = XALLOCAVEC (unsigned char, nunits); | |
5054 | ||
5055 | for (i = 0; i < nunits; ++i) | |
5056 | sel[i] = nunits - 1 - i; | |
5057 | ||
557be5a8 AL |
5058 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) |
5059 | return NULL_TREE; | |
5060 | return vect_gen_perm_mask_checked (vectype, sel); | |
09dfa495 BM |
5061 | } |
5062 | ||
ebfd146a IR |
5063 | /* Function vectorizable_store. |
5064 | ||
b8698a0f L |
5065 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
5066 | can be vectorized. | |
5067 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5068 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5069 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5070 | ||
5071 | static bool | |
355fe088 | 5072 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 5073 | slp_tree slp_node) |
ebfd146a IR |
5074 | { |
5075 | tree scalar_dest; | |
5076 | tree data_ref; | |
5077 | tree op; | |
5078 | tree vec_oprnd = NULL_TREE; | |
5079 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5080 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 5081 | tree elem_type; |
ebfd146a | 5082 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 5083 | struct loop *loop = NULL; |
ef4bddc2 | 5084 | machine_mode vec_mode; |
ebfd146a IR |
5085 | tree dummy; |
5086 | enum dr_alignment_support alignment_support_scheme; | |
355fe088 | 5087 | gimple *def_stmt; |
ebfd146a IR |
5088 | enum vect_def_type dt; |
5089 | stmt_vec_info prev_stmt_info = NULL; | |
5090 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 5091 | tree dataref_offset = NULL_TREE; |
355fe088 | 5092 | gimple *ptr_incr = NULL; |
ebfd146a IR |
5093 | int ncopies; |
5094 | int j; | |
355fe088 | 5095 | gimple *next_stmt, *first_stmt = NULL; |
0d0293ac | 5096 | bool grouped_store = false; |
272c6793 | 5097 | bool store_lanes_p = false; |
ebfd146a | 5098 | unsigned int group_size, i; |
6e1aa848 DN |
5099 | vec<tree> dr_chain = vNULL; |
5100 | vec<tree> oprnds = vNULL; | |
5101 | vec<tree> result_chain = vNULL; | |
ebfd146a | 5102 | bool inv_p; |
09dfa495 BM |
5103 | bool negative = false; |
5104 | tree offset = NULL_TREE; | |
6e1aa848 | 5105 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5106 | bool slp = (slp_node != NULL); |
ebfd146a | 5107 | unsigned int vec_num; |
a70d6342 | 5108 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5109 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 5110 | tree aggr_type; |
3bab6342 AT |
5111 | tree scatter_base = NULL_TREE, scatter_off = NULL_TREE; |
5112 | tree scatter_off_vectype = NULL_TREE, scatter_decl = NULL_TREE; | |
5113 | int scatter_scale = 1; | |
5114 | enum vect_def_type scatter_idx_dt = vect_unknown_def_type; | |
5115 | enum vect_def_type scatter_src_dt = vect_unknown_def_type; | |
355fe088 | 5116 | gimple *new_stmt; |
a70d6342 | 5117 | |
a70d6342 | 5118 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5119 | return false; |
5120 | ||
8644a673 | 5121 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
5122 | return false; |
5123 | ||
5124 | /* Is vectorizable store? */ | |
5125 | ||
5126 | if (!is_gimple_assign (stmt)) | |
5127 | return false; | |
5128 | ||
5129 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
5130 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
5131 | && is_pattern_stmt_p (stmt_info)) | |
5132 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 5133 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 5134 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 5135 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
5136 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
5137 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
5138 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
5139 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
5140 | return false; |
5141 | ||
5142 | gcc_assert (gimple_assign_single_p (stmt)); | |
465c8c19 JJ |
5143 | |
5144 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
5145 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
5146 | ||
5147 | if (loop_vinfo) | |
5148 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
5149 | ||
5150 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5151 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5152 | case of SLP. */ | |
5153 | if (slp || PURE_SLP_STMT (stmt_info)) | |
5154 | ncopies = 1; | |
5155 | else | |
5156 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
5157 | ||
5158 | gcc_assert (ncopies >= 1); | |
5159 | ||
5160 | /* FORNOW. This restriction should be relaxed. */ | |
5161 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
5162 | { | |
5163 | if (dump_enabled_p ()) | |
5164 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5165 | "multiple types in nested loop.\n"); | |
5166 | return false; | |
5167 | } | |
5168 | ||
ebfd146a | 5169 | op = gimple_assign_rhs1 (stmt); |
81c40241 | 5170 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) |
ebfd146a | 5171 | { |
73fbfcad | 5172 | if (dump_enabled_p ()) |
78c60e3d | 5173 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5174 | "use not simple.\n"); |
ebfd146a IR |
5175 | return false; |
5176 | } | |
5177 | ||
272c6793 | 5178 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 5179 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 5180 | |
ebfd146a IR |
5181 | /* FORNOW. In some cases can vectorize even if data-type not supported |
5182 | (e.g. - array initialization with 0). */ | |
947131ba | 5183 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
5184 | return false; |
5185 | ||
5186 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
5187 | return false; | |
5188 | ||
f2e2a985 | 5189 | if (!STMT_VINFO_STRIDED_P (stmt_info)) |
09dfa495 | 5190 | { |
f2e2a985 MM |
5191 | negative = |
5192 | tree_int_cst_compare (loop && nested_in_vect_loop_p (loop, stmt) | |
5193 | ? STMT_VINFO_DR_STEP (stmt_info) : DR_STEP (dr), | |
5194 | size_zero_node) < 0; | |
5195 | if (negative && ncopies > 1) | |
09dfa495 BM |
5196 | { |
5197 | if (dump_enabled_p ()) | |
5198 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
f2e2a985 | 5199 | "multiple types with negative step.\n"); |
09dfa495 BM |
5200 | return false; |
5201 | } | |
f2e2a985 | 5202 | if (negative) |
09dfa495 | 5203 | { |
f2e2a985 MM |
5204 | gcc_assert (!grouped_store); |
5205 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
5206 | if (alignment_support_scheme != dr_aligned | |
5207 | && alignment_support_scheme != dr_unaligned_supported) | |
5208 | { | |
5209 | if (dump_enabled_p ()) | |
5210 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5211 | "negative step but alignment required.\n"); | |
5212 | return false; | |
5213 | } | |
5214 | if (dt != vect_constant_def | |
5215 | && dt != vect_external_def | |
5216 | && !perm_mask_for_reverse (vectype)) | |
5217 | { | |
5218 | if (dump_enabled_p ()) | |
5219 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5220 | "negative step and reversing not supported.\n"); | |
5221 | return false; | |
5222 | } | |
09dfa495 BM |
5223 | } |
5224 | } | |
5225 | ||
0d0293ac | 5226 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 5227 | { |
0d0293ac | 5228 | grouped_store = true; |
e14c1050 | 5229 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
cee62fee MM |
5230 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
5231 | if (!slp | |
5232 | && !PURE_SLP_STMT (stmt_info) | |
5233 | && !STMT_VINFO_STRIDED_P (stmt_info)) | |
b602d918 | 5234 | { |
272c6793 RS |
5235 | if (vect_store_lanes_supported (vectype, group_size)) |
5236 | store_lanes_p = true; | |
0d0293ac | 5237 | else if (!vect_grouped_store_supported (vectype, group_size)) |
b602d918 RS |
5238 | return false; |
5239 | } | |
b8698a0f | 5240 | |
cee62fee MM |
5241 | if (STMT_VINFO_STRIDED_P (stmt_info) |
5242 | && (slp || PURE_SLP_STMT (stmt_info)) | |
5243 | && (group_size > nunits | |
5244 | || nunits % group_size != 0)) | |
5245 | { | |
5246 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5247 | "unhandled strided group store\n"); | |
5248 | return false; | |
5249 | } | |
5250 | ||
ebfd146a IR |
5251 | if (first_stmt == stmt) |
5252 | { | |
5253 | /* STMT is the leader of the group. Check the operands of all the | |
5254 | stmts of the group. */ | |
e14c1050 | 5255 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
5256 | while (next_stmt) |
5257 | { | |
5258 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5259 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 5260 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) |
ebfd146a | 5261 | { |
73fbfcad | 5262 | if (dump_enabled_p ()) |
78c60e3d | 5263 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5264 | "use not simple.\n"); |
ebfd146a IR |
5265 | return false; |
5266 | } | |
e14c1050 | 5267 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5268 | } |
5269 | } | |
5270 | } | |
5271 | ||
3bab6342 AT |
5272 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5273 | { | |
355fe088 | 5274 | gimple *def_stmt; |
3bab6342 AT |
5275 | scatter_decl = vect_check_gather_scatter (stmt, loop_vinfo, &scatter_base, |
5276 | &scatter_off, &scatter_scale); | |
5277 | gcc_assert (scatter_decl); | |
81c40241 RB |
5278 | if (!vect_is_simple_use (scatter_off, vinfo, &def_stmt, &scatter_idx_dt, |
5279 | &scatter_off_vectype)) | |
3bab6342 AT |
5280 | { |
5281 | if (dump_enabled_p ()) | |
5282 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5283 | "scatter index use not simple."); | |
5284 | return false; | |
5285 | } | |
5286 | } | |
5287 | ||
ebfd146a IR |
5288 | if (!vec_stmt) /* transformation not required. */ |
5289 | { | |
5290 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
2e8ab70c RB |
5291 | /* The SLP costs are calculated during SLP analysis. */ |
5292 | if (!PURE_SLP_STMT (stmt_info)) | |
5293 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, | |
5294 | NULL, NULL, NULL); | |
ebfd146a IR |
5295 | return true; |
5296 | } | |
5297 | ||
5298 | /** Transform. **/ | |
5299 | ||
c716e67f XDL |
5300 | ensure_base_align (stmt_info, dr); |
5301 | ||
3bab6342 AT |
5302 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5303 | { | |
5304 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, op, src; | |
5305 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (scatter_decl)); | |
5306 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
5307 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
5308 | edge pe = loop_preheader_edge (loop); | |
5309 | gimple_seq seq; | |
5310 | basic_block new_bb; | |
5311 | enum { NARROW, NONE, WIDEN } modifier; | |
5312 | int scatter_off_nunits = TYPE_VECTOR_SUBPARTS (scatter_off_vectype); | |
5313 | ||
5314 | if (nunits == (unsigned int) scatter_off_nunits) | |
5315 | modifier = NONE; | |
5316 | else if (nunits == (unsigned int) scatter_off_nunits / 2) | |
5317 | { | |
5318 | unsigned char *sel = XALLOCAVEC (unsigned char, scatter_off_nunits); | |
5319 | modifier = WIDEN; | |
5320 | ||
5321 | for (i = 0; i < (unsigned int) scatter_off_nunits; ++i) | |
5322 | sel[i] = i | nunits; | |
5323 | ||
5324 | perm_mask = vect_gen_perm_mask_checked (scatter_off_vectype, sel); | |
5325 | gcc_assert (perm_mask != NULL_TREE); | |
5326 | } | |
5327 | else if (nunits == (unsigned int) scatter_off_nunits * 2) | |
5328 | { | |
5329 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
5330 | modifier = NARROW; | |
5331 | ||
5332 | for (i = 0; i < (unsigned int) nunits; ++i) | |
5333 | sel[i] = i | scatter_off_nunits; | |
5334 | ||
5335 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); | |
5336 | gcc_assert (perm_mask != NULL_TREE); | |
5337 | ncopies *= 2; | |
5338 | } | |
5339 | else | |
5340 | gcc_unreachable (); | |
5341 | ||
5342 | rettype = TREE_TYPE (TREE_TYPE (scatter_decl)); | |
5343 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5344 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5345 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5346 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5347 | scaletype = TREE_VALUE (arglist); | |
5348 | ||
5349 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
5350 | && TREE_CODE (rettype) == VOID_TYPE); | |
5351 | ||
5352 | ptr = fold_convert (ptrtype, scatter_base); | |
5353 | if (!is_gimple_min_invariant (ptr)) | |
5354 | { | |
5355 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
5356 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
5357 | gcc_assert (!new_bb); | |
5358 | } | |
5359 | ||
5360 | /* Currently we support only unconditional scatter stores, | |
5361 | so mask should be all ones. */ | |
5362 | mask = build_int_cst (masktype, -1); | |
5363 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
5364 | ||
5365 | scale = build_int_cst (scaletype, scatter_scale); | |
5366 | ||
5367 | prev_stmt_info = NULL; | |
5368 | for (j = 0; j < ncopies; ++j) | |
5369 | { | |
5370 | if (j == 0) | |
5371 | { | |
5372 | src = vec_oprnd1 | |
81c40241 | 5373 | = vect_get_vec_def_for_operand (gimple_assign_rhs1 (stmt), stmt); |
3bab6342 | 5374 | op = vec_oprnd0 |
81c40241 | 5375 | = vect_get_vec_def_for_operand (scatter_off, stmt); |
3bab6342 AT |
5376 | } |
5377 | else if (modifier != NONE && (j & 1)) | |
5378 | { | |
5379 | if (modifier == WIDEN) | |
5380 | { | |
5381 | src = vec_oprnd1 | |
5382 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5383 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, | |
5384 | stmt, gsi); | |
5385 | } | |
5386 | else if (modifier == NARROW) | |
5387 | { | |
5388 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
5389 | stmt, gsi); | |
5390 | op = vec_oprnd0 | |
5391 | = vect_get_vec_def_for_stmt_copy (scatter_idx_dt, vec_oprnd0); | |
5392 | } | |
5393 | else | |
5394 | gcc_unreachable (); | |
5395 | } | |
5396 | else | |
5397 | { | |
5398 | src = vec_oprnd1 | |
5399 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5400 | op = vec_oprnd0 | |
5401 | = vect_get_vec_def_for_stmt_copy (scatter_idx_dt, vec_oprnd0); | |
5402 | } | |
5403 | ||
5404 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
5405 | { | |
5406 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)) | |
5407 | == TYPE_VECTOR_SUBPARTS (srctype)); | |
0e22bb5a | 5408 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 AT |
5409 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
5410 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
5411 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5412 | src = var; | |
5413 | } | |
5414 | ||
5415 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
5416 | { | |
5417 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
5418 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 5419 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 AT |
5420 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
5421 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
5422 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5423 | op = var; | |
5424 | } | |
5425 | ||
5426 | new_stmt | |
5427 | = gimple_build_call (scatter_decl, 5, ptr, mask, op, src, scale); | |
5428 | ||
5429 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5430 | ||
5431 | if (prev_stmt_info == NULL) | |
5432 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5433 | else | |
5434 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5435 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5436 | } | |
5437 | return true; | |
5438 | } | |
5439 | ||
0d0293ac | 5440 | if (grouped_store) |
ebfd146a IR |
5441 | { |
5442 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 5443 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 5444 | |
e14c1050 | 5445 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
5446 | |
5447 | /* FORNOW */ | |
a70d6342 | 5448 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
5449 | |
5450 | /* We vectorize all the stmts of the interleaving group when we | |
5451 | reach the last stmt in the group. */ | |
e14c1050 IR |
5452 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
5453 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
5454 | && !slp) |
5455 | { | |
5456 | *vec_stmt = NULL; | |
5457 | return true; | |
5458 | } | |
5459 | ||
5460 | if (slp) | |
4b5caab7 | 5461 | { |
0d0293ac | 5462 | grouped_store = false; |
4b5caab7 IR |
5463 | /* VEC_NUM is the number of vect stmts to be created for this |
5464 | group. */ | |
5465 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 5466 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
52eab378 | 5467 | gcc_assert (GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt)) == first_stmt); |
4b5caab7 | 5468 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 5469 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 5470 | } |
ebfd146a | 5471 | else |
4b5caab7 IR |
5472 | /* VEC_NUM is the number of vect stmts to be created for this |
5473 | group. */ | |
ebfd146a IR |
5474 | vec_num = group_size; |
5475 | } | |
b8698a0f | 5476 | else |
ebfd146a IR |
5477 | { |
5478 | first_stmt = stmt; | |
5479 | first_dr = dr; | |
5480 | group_size = vec_num = 1; | |
ebfd146a | 5481 | } |
b8698a0f | 5482 | |
73fbfcad | 5483 | if (dump_enabled_p ()) |
78c60e3d | 5484 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5485 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 5486 | |
f2e2a985 MM |
5487 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
5488 | { | |
5489 | gimple_stmt_iterator incr_gsi; | |
5490 | bool insert_after; | |
355fe088 | 5491 | gimple *incr; |
f2e2a985 MM |
5492 | tree offvar; |
5493 | tree ivstep; | |
5494 | tree running_off; | |
5495 | gimple_seq stmts = NULL; | |
5496 | tree stride_base, stride_step, alias_off; | |
5497 | tree vec_oprnd; | |
f502d50e | 5498 | unsigned int g; |
f2e2a985 MM |
5499 | |
5500 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); | |
5501 | ||
5502 | stride_base | |
5503 | = fold_build_pointer_plus | |
f502d50e | 5504 | (unshare_expr (DR_BASE_ADDRESS (first_dr)), |
f2e2a985 | 5505 | size_binop (PLUS_EXPR, |
f502d50e MM |
5506 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (first_dr))), |
5507 | convert_to_ptrofftype (DR_INIT(first_dr)))); | |
5508 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (first_dr))); | |
f2e2a985 MM |
5509 | |
5510 | /* For a store with loop-invariant (but other than power-of-2) | |
5511 | stride (i.e. not a grouped access) like so: | |
5512 | ||
5513 | for (i = 0; i < n; i += stride) | |
5514 | array[i] = ...; | |
5515 | ||
5516 | we generate a new induction variable and new stores from | |
5517 | the components of the (vectorized) rhs: | |
5518 | ||
5519 | for (j = 0; ; j += VF*stride) | |
5520 | vectemp = ...; | |
5521 | tmp1 = vectemp[0]; | |
5522 | array[j] = tmp1; | |
5523 | tmp2 = vectemp[1]; | |
5524 | array[j + stride] = tmp2; | |
5525 | ... | |
5526 | */ | |
5527 | ||
cee62fee MM |
5528 | unsigned nstores = nunits; |
5529 | tree ltype = elem_type; | |
5530 | if (slp) | |
5531 | { | |
5532 | nstores = nunits / group_size; | |
5533 | if (group_size < nunits) | |
5534 | ltype = build_vector_type (elem_type, group_size); | |
5535 | else | |
5536 | ltype = vectype; | |
5537 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); | |
5538 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
f502d50e | 5539 | group_size = 1; |
cee62fee MM |
5540 | } |
5541 | ||
f2e2a985 MM |
5542 | ivstep = stride_step; |
5543 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
5544 | build_int_cst (TREE_TYPE (ivstep), | |
cee62fee | 5545 | ncopies * nstores)); |
f2e2a985 MM |
5546 | |
5547 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
5548 | ||
5549 | create_iv (stride_base, ivstep, NULL, | |
5550 | loop, &incr_gsi, insert_after, | |
5551 | &offvar, NULL); | |
5552 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 5553 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
f2e2a985 MM |
5554 | |
5555 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
5556 | if (stmts) | |
5557 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
5558 | ||
5559 | prev_stmt_info = NULL; | |
f502d50e MM |
5560 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (first_dr)), 0); |
5561 | next_stmt = first_stmt; | |
5562 | for (g = 0; g < group_size; g++) | |
f2e2a985 | 5563 | { |
f502d50e MM |
5564 | running_off = offvar; |
5565 | if (g) | |
f2e2a985 | 5566 | { |
f502d50e MM |
5567 | tree size = TYPE_SIZE_UNIT (ltype); |
5568 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 5569 | size); |
f502d50e | 5570 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 5571 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 5572 | running_off, pos); |
f2e2a985 | 5573 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 5574 | running_off = newoff; |
f502d50e MM |
5575 | } |
5576 | for (j = 0; j < ncopies; j++) | |
5577 | { | |
5578 | /* We've set op and dt above, from gimple_assign_rhs1(stmt), | |
5579 | and first_stmt == stmt. */ | |
5580 | if (j == 0) | |
5581 | { | |
5582 | if (slp) | |
5583 | { | |
5584 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
5585 | slp_node, -1); | |
5586 | vec_oprnd = vec_oprnds[0]; | |
5587 | } | |
5588 | else | |
5589 | { | |
5590 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5591 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 5592 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
f502d50e MM |
5593 | } |
5594 | } | |
f2e2a985 | 5595 | else |
f502d50e MM |
5596 | { |
5597 | if (slp) | |
5598 | vec_oprnd = vec_oprnds[j]; | |
5599 | else | |
c079cbac | 5600 | { |
81c40241 | 5601 | vect_is_simple_use (vec_oprnd, vinfo, &def_stmt, &dt); |
c079cbac RB |
5602 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); |
5603 | } | |
f502d50e MM |
5604 | } |
5605 | ||
5606 | for (i = 0; i < nstores; i++) | |
5607 | { | |
5608 | tree newref, newoff; | |
355fe088 | 5609 | gimple *incr, *assign; |
f502d50e MM |
5610 | tree size = TYPE_SIZE (ltype); |
5611 | /* Extract the i'th component. */ | |
5612 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
5613 | bitsize_int (i), size); | |
5614 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
5615 | size, pos); | |
5616 | ||
5617 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
5618 | NULL_TREE, true, | |
5619 | GSI_SAME_STMT); | |
5620 | ||
5621 | newref = build2 (MEM_REF, ltype, | |
5622 | running_off, alias_off); | |
5623 | ||
5624 | /* And store it to *running_off. */ | |
5625 | assign = gimple_build_assign (newref, elem); | |
5626 | vect_finish_stmt_generation (stmt, assign, gsi); | |
5627 | ||
5628 | newoff = copy_ssa_name (running_off, NULL); | |
5629 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
5630 | running_off, stride_step); | |
5631 | vect_finish_stmt_generation (stmt, incr, gsi); | |
5632 | ||
5633 | running_off = newoff; | |
225ce44b RB |
5634 | if (g == group_size - 1 |
5635 | && !slp) | |
f502d50e MM |
5636 | { |
5637 | if (j == 0 && i == 0) | |
225ce44b RB |
5638 | STMT_VINFO_VEC_STMT (stmt_info) |
5639 | = *vec_stmt = assign; | |
f502d50e MM |
5640 | else |
5641 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign; | |
5642 | prev_stmt_info = vinfo_for_stmt (assign); | |
5643 | } | |
5644 | } | |
f2e2a985 | 5645 | } |
f502d50e | 5646 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
f2e2a985 MM |
5647 | } |
5648 | return true; | |
5649 | } | |
5650 | ||
9771b263 DN |
5651 | dr_chain.create (group_size); |
5652 | oprnds.create (group_size); | |
ebfd146a | 5653 | |
720f5239 | 5654 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 5655 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
5656 | /* Targets with store-lane instructions must not require explicit |
5657 | realignment. */ | |
5658 | gcc_assert (!store_lanes_p | |
5659 | || alignment_support_scheme == dr_aligned | |
5660 | || alignment_support_scheme == dr_unaligned_supported); | |
5661 | ||
09dfa495 BM |
5662 | if (negative) |
5663 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
5664 | ||
272c6793 RS |
5665 | if (store_lanes_p) |
5666 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
5667 | else | |
5668 | aggr_type = vectype; | |
ebfd146a IR |
5669 | |
5670 | /* In case the vectorization factor (VF) is bigger than the number | |
5671 | of elements that we can fit in a vectype (nunits), we have to generate | |
5672 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 5673 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
5674 | vect_get_vec_def_for_copy_stmt. */ |
5675 | ||
0d0293ac | 5676 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
5677 | |
5678 | S1: &base + 2 = x2 | |
5679 | S2: &base = x0 | |
5680 | S3: &base + 1 = x1 | |
5681 | S4: &base + 3 = x3 | |
5682 | ||
5683 | We create vectorized stores starting from base address (the access of the | |
5684 | first stmt in the chain (S2 in the above example), when the last store stmt | |
5685 | of the chain (S4) is reached: | |
5686 | ||
5687 | VS1: &base = vx2 | |
5688 | VS2: &base + vec_size*1 = vx0 | |
5689 | VS3: &base + vec_size*2 = vx1 | |
5690 | VS4: &base + vec_size*3 = vx3 | |
5691 | ||
5692 | Then permutation statements are generated: | |
5693 | ||
3fcc1b55 JJ |
5694 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
5695 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 5696 | ... |
b8698a0f | 5697 | |
ebfd146a IR |
5698 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
5699 | (the order of the data-refs in the output of vect_permute_store_chain | |
5700 | corresponds to the order of scalar stmts in the interleaving chain - see | |
5701 | the documentation of vect_permute_store_chain()). | |
5702 | ||
5703 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 5704 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 5705 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 5706 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
5707 | */ |
5708 | ||
5709 | prev_stmt_info = NULL; | |
5710 | for (j = 0; j < ncopies; j++) | |
5711 | { | |
ebfd146a IR |
5712 | |
5713 | if (j == 0) | |
5714 | { | |
5715 | if (slp) | |
5716 | { | |
5717 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c IR |
5718 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
5719 | NULL, slp_node, -1); | |
ebfd146a | 5720 | |
9771b263 | 5721 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
5722 | } |
5723 | else | |
5724 | { | |
b8698a0f L |
5725 | /* For interleaved stores we collect vectorized defs for all the |
5726 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
5727 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
5728 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
5729 | ||
0d0293ac | 5730 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 5731 | OPRNDS are of size 1. */ |
b8698a0f | 5732 | next_stmt = first_stmt; |
ebfd146a IR |
5733 | for (i = 0; i < group_size; i++) |
5734 | { | |
b8698a0f L |
5735 | /* Since gaps are not supported for interleaved stores, |
5736 | GROUP_SIZE is the exact number of stmts in the chain. | |
5737 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
5738 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
5739 | iteration of the loop will be executed. */ |
5740 | gcc_assert (next_stmt | |
5741 | && gimple_assign_single_p (next_stmt)); | |
5742 | op = gimple_assign_rhs1 (next_stmt); | |
5743 | ||
81c40241 | 5744 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
9771b263 DN |
5745 | dr_chain.quick_push (vec_oprnd); |
5746 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 5747 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5748 | } |
5749 | } | |
5750 | ||
5751 | /* We should have catched mismatched types earlier. */ | |
5752 | gcc_assert (useless_type_conversion_p (vectype, | |
5753 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
5754 | bool simd_lane_access_p |
5755 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
5756 | if (simd_lane_access_p | |
5757 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
5758 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
5759 | && integer_zerop (DR_OFFSET (first_dr)) | |
5760 | && integer_zerop (DR_INIT (first_dr)) | |
5761 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
5762 | get_alias_set (DR_REF (first_dr)))) | |
5763 | { | |
5764 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
5765 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
5766 | (DR_REF (first_dr)), 0); | |
8928eff3 | 5767 | inv_p = false; |
74bf76ed JJ |
5768 | } |
5769 | else | |
5770 | dataref_ptr | |
5771 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
5772 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 5773 | offset, &dummy, gsi, &ptr_incr, |
74bf76ed | 5774 | simd_lane_access_p, &inv_p); |
a70d6342 | 5775 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 5776 | } |
b8698a0f | 5777 | else |
ebfd146a | 5778 | { |
b8698a0f L |
5779 | /* For interleaved stores we created vectorized defs for all the |
5780 | defs stored in OPRNDS in the previous iteration (previous copy). | |
5781 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
5782 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
5783 | next copy. | |
0d0293ac | 5784 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
5785 | OPRNDS are of size 1. */ |
5786 | for (i = 0; i < group_size; i++) | |
5787 | { | |
9771b263 | 5788 | op = oprnds[i]; |
81c40241 | 5789 | vect_is_simple_use (op, vinfo, &def_stmt, &dt); |
b8698a0f | 5790 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
5791 | dr_chain[i] = vec_oprnd; |
5792 | oprnds[i] = vec_oprnd; | |
ebfd146a | 5793 | } |
74bf76ed JJ |
5794 | if (dataref_offset) |
5795 | dataref_offset | |
5796 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
5797 | TYPE_SIZE_UNIT (aggr_type)); | |
5798 | else | |
5799 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
5800 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
5801 | } |
5802 | ||
272c6793 | 5803 | if (store_lanes_p) |
ebfd146a | 5804 | { |
272c6793 | 5805 | tree vec_array; |
267d3070 | 5806 | |
272c6793 RS |
5807 | /* Combine all the vectors into an array. */ |
5808 | vec_array = create_vector_array (vectype, vec_num); | |
5809 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 5810 | { |
9771b263 | 5811 | vec_oprnd = dr_chain[i]; |
272c6793 | 5812 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 5813 | } |
b8698a0f | 5814 | |
272c6793 RS |
5815 | /* Emit: |
5816 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
5817 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
5818 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
5819 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 | 5820 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
5821 | } |
5822 | else | |
5823 | { | |
5824 | new_stmt = NULL; | |
0d0293ac | 5825 | if (grouped_store) |
272c6793 | 5826 | { |
b6b9227d JJ |
5827 | if (j == 0) |
5828 | result_chain.create (group_size); | |
272c6793 RS |
5829 | /* Permute. */ |
5830 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
5831 | &result_chain); | |
5832 | } | |
c2d7ab2a | 5833 | |
272c6793 RS |
5834 | next_stmt = first_stmt; |
5835 | for (i = 0; i < vec_num; i++) | |
5836 | { | |
644ffefd | 5837 | unsigned align, misalign; |
272c6793 RS |
5838 | |
5839 | if (i > 0) | |
5840 | /* Bump the vector pointer. */ | |
5841 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5842 | stmt, NULL_TREE); | |
5843 | ||
5844 | if (slp) | |
9771b263 | 5845 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
5846 | else if (grouped_store) |
5847 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 5848 | vect_permute_store_chain(). */ |
9771b263 | 5849 | vec_oprnd = result_chain[i]; |
272c6793 | 5850 | |
aed93b23 RB |
5851 | data_ref = fold_build2 (MEM_REF, TREE_TYPE (vec_oprnd), |
5852 | dataref_ptr, | |
5853 | dataref_offset | |
5854 | ? dataref_offset | |
5855 | : build_int_cst (reference_alias_ptr_type | |
5856 | (DR_REF (first_dr)), 0)); | |
644ffefd | 5857 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 | 5858 | if (aligned_access_p (first_dr)) |
644ffefd | 5859 | misalign = 0; |
272c6793 RS |
5860 | else if (DR_MISALIGNMENT (first_dr) == -1) |
5861 | { | |
52639a61 RB |
5862 | if (DR_VECT_AUX (first_dr)->base_element_aligned) |
5863 | align = TYPE_ALIGN_UNIT (elem_type); | |
5864 | else | |
5865 | align = get_object_alignment (DR_REF (first_dr)) | |
5866 | / BITS_PER_UNIT; | |
5867 | misalign = 0; | |
272c6793 RS |
5868 | TREE_TYPE (data_ref) |
5869 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 5870 | align * BITS_PER_UNIT); |
272c6793 RS |
5871 | } |
5872 | else | |
5873 | { | |
5874 | TREE_TYPE (data_ref) | |
5875 | = build_aligned_type (TREE_TYPE (data_ref), | |
5876 | TYPE_ALIGN (elem_type)); | |
644ffefd | 5877 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 5878 | } |
aed93b23 RB |
5879 | if (dataref_offset == NULL_TREE |
5880 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
5881 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
5882 | misalign); | |
c2d7ab2a | 5883 | |
f234d260 BM |
5884 | if (negative |
5885 | && dt != vect_constant_def | |
5886 | && dt != vect_external_def) | |
09dfa495 BM |
5887 | { |
5888 | tree perm_mask = perm_mask_for_reverse (vectype); | |
5889 | tree perm_dest | |
5890 | = vect_create_destination_var (gimple_assign_rhs1 (stmt), | |
5891 | vectype); | |
b731b390 | 5892 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
5893 | |
5894 | /* Generate the permute statement. */ | |
355fe088 | 5895 | gimple *perm_stmt |
0d0e4a03 JJ |
5896 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
5897 | vec_oprnd, perm_mask); | |
09dfa495 BM |
5898 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
5899 | ||
5900 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
5901 | vec_oprnd = new_temp; | |
5902 | } | |
5903 | ||
272c6793 RS |
5904 | /* Arguments are ready. Create the new vector stmt. */ |
5905 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
5906 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
5907 | |
5908 | if (slp) | |
5909 | continue; | |
5910 | ||
e14c1050 | 5911 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
5912 | if (!next_stmt) |
5913 | break; | |
5914 | } | |
ebfd146a | 5915 | } |
1da0876c RS |
5916 | if (!slp) |
5917 | { | |
5918 | if (j == 0) | |
5919 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5920 | else | |
5921 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5922 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5923 | } | |
ebfd146a IR |
5924 | } |
5925 | ||
9771b263 DN |
5926 | dr_chain.release (); |
5927 | oprnds.release (); | |
5928 | result_chain.release (); | |
5929 | vec_oprnds.release (); | |
ebfd146a IR |
5930 | |
5931 | return true; | |
5932 | } | |
5933 | ||
557be5a8 AL |
5934 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
5935 | VECTOR_CST mask. No checks are made that the target platform supports the | |
5936 | mask, so callers may wish to test can_vec_perm_p separately, or use | |
5937 | vect_gen_perm_mask_checked. */ | |
a1e53f3f | 5938 | |
3fcc1b55 | 5939 | tree |
557be5a8 | 5940 | vect_gen_perm_mask_any (tree vectype, const unsigned char *sel) |
a1e53f3f | 5941 | { |
d2a12ae7 | 5942 | tree mask_elt_type, mask_type, mask_vec, *mask_elts; |
2635892a | 5943 | int i, nunits; |
a1e53f3f | 5944 | |
22e4dee7 | 5945 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 | 5946 | |
96f9265a RG |
5947 | mask_elt_type = lang_hooks.types.type_for_mode |
5948 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
22e4dee7 | 5949 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 5950 | |
d2a12ae7 | 5951 | mask_elts = XALLOCAVEC (tree, nunits); |
aec7ae7d | 5952 | for (i = nunits - 1; i >= 0; i--) |
d2a12ae7 RG |
5953 | mask_elts[i] = build_int_cst (mask_elt_type, sel[i]); |
5954 | mask_vec = build_vector (mask_type, mask_elts); | |
a1e53f3f | 5955 | |
2635892a | 5956 | return mask_vec; |
a1e53f3f L |
5957 | } |
5958 | ||
cf7aa6a3 AL |
5959 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_p, |
5960 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ | |
557be5a8 AL |
5961 | |
5962 | tree | |
5963 | vect_gen_perm_mask_checked (tree vectype, const unsigned char *sel) | |
5964 | { | |
5965 | gcc_assert (can_vec_perm_p (TYPE_MODE (vectype), false, sel)); | |
5966 | return vect_gen_perm_mask_any (vectype, sel); | |
5967 | } | |
5968 | ||
aec7ae7d JJ |
5969 | /* Given a vector variable X and Y, that was generated for the scalar |
5970 | STMT, generate instructions to permute the vector elements of X and Y | |
5971 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
5972 | permuted vector variable. */ | |
a1e53f3f L |
5973 | |
5974 | static tree | |
355fe088 | 5975 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 5976 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
5977 | { |
5978 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 5979 | tree perm_dest, data_ref; |
355fe088 | 5980 | gimple *perm_stmt; |
a1e53f3f | 5981 | |
acdcd61b | 5982 | perm_dest = vect_create_destination_var (gimple_get_lhs (stmt), vectype); |
b731b390 | 5983 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
5984 | |
5985 | /* Generate the permute statement. */ | |
0d0e4a03 | 5986 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
5987 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
5988 | ||
5989 | return data_ref; | |
5990 | } | |
5991 | ||
6b916b36 RB |
5992 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
5993 | inserting them on the loops preheader edge. Returns true if we | |
5994 | were successful in doing so (and thus STMT can be moved then), | |
5995 | otherwise returns false. */ | |
5996 | ||
5997 | static bool | |
355fe088 | 5998 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
5999 | { |
6000 | ssa_op_iter i; | |
6001 | tree op; | |
6002 | bool any = false; | |
6003 | ||
6004 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6005 | { | |
355fe088 | 6006 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6007 | if (!gimple_nop_p (def_stmt) |
6008 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6009 | { | |
6010 | /* Make sure we don't need to recurse. While we could do | |
6011 | so in simple cases when there are more complex use webs | |
6012 | we don't have an easy way to preserve stmt order to fulfil | |
6013 | dependencies within them. */ | |
6014 | tree op2; | |
6015 | ssa_op_iter i2; | |
d1417442 JJ |
6016 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
6017 | return false; | |
6b916b36 RB |
6018 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
6019 | { | |
355fe088 | 6020 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
6021 | if (!gimple_nop_p (def_stmt2) |
6022 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
6023 | return false; | |
6024 | } | |
6025 | any = true; | |
6026 | } | |
6027 | } | |
6028 | ||
6029 | if (!any) | |
6030 | return true; | |
6031 | ||
6032 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6033 | { | |
355fe088 | 6034 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6035 | if (!gimple_nop_p (def_stmt) |
6036 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6037 | { | |
6038 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
6039 | gsi_remove (&gsi, false); | |
6040 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
6041 | } | |
6042 | } | |
6043 | ||
6044 | return true; | |
6045 | } | |
6046 | ||
ebfd146a IR |
6047 | /* vectorizable_load. |
6048 | ||
b8698a0f L |
6049 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
6050 | can be vectorized. | |
6051 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6052 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6053 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6054 | ||
6055 | static bool | |
355fe088 | 6056 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 6057 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
6058 | { |
6059 | tree scalar_dest; | |
6060 | tree vec_dest = NULL; | |
6061 | tree data_ref = NULL; | |
6062 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 6063 | stmt_vec_info prev_stmt_info; |
ebfd146a | 6064 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6065 | struct loop *loop = NULL; |
ebfd146a | 6066 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 6067 | bool nested_in_vect_loop = false; |
c716e67f | 6068 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 6069 | tree elem_type; |
ebfd146a | 6070 | tree new_temp; |
ef4bddc2 | 6071 | machine_mode mode; |
355fe088 | 6072 | gimple *new_stmt = NULL; |
ebfd146a IR |
6073 | tree dummy; |
6074 | enum dr_alignment_support alignment_support_scheme; | |
6075 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6076 | tree dataref_offset = NULL_TREE; |
355fe088 | 6077 | gimple *ptr_incr = NULL; |
ebfd146a | 6078 | int ncopies; |
9b999e8c | 6079 | int i, j, group_size = -1, group_gap_adj; |
ebfd146a IR |
6080 | tree msq = NULL_TREE, lsq; |
6081 | tree offset = NULL_TREE; | |
356bbc4c | 6082 | tree byte_offset = NULL_TREE; |
ebfd146a | 6083 | tree realignment_token = NULL_TREE; |
538dd0b7 | 6084 | gphi *phi = NULL; |
6e1aa848 | 6085 | vec<tree> dr_chain = vNULL; |
0d0293ac | 6086 | bool grouped_load = false; |
272c6793 | 6087 | bool load_lanes_p = false; |
355fe088 | 6088 | gimple *first_stmt; |
ebfd146a | 6089 | bool inv_p; |
319e6439 | 6090 | bool negative = false; |
ebfd146a IR |
6091 | bool compute_in_loop = false; |
6092 | struct loop *at_loop; | |
6093 | int vec_num; | |
6094 | bool slp = (slp_node != NULL); | |
6095 | bool slp_perm = false; | |
6096 | enum tree_code code; | |
a70d6342 IR |
6097 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
6098 | int vf; | |
272c6793 | 6099 | tree aggr_type; |
aec7ae7d JJ |
6100 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; |
6101 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
6102 | int gather_scale = 1; | |
6103 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
310213d4 | 6104 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 6105 | |
465c8c19 JJ |
6106 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
6107 | return false; | |
6108 | ||
6109 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
6110 | return false; | |
6111 | ||
6112 | /* Is vectorizable load? */ | |
6113 | if (!is_gimple_assign (stmt)) | |
6114 | return false; | |
6115 | ||
6116 | scalar_dest = gimple_assign_lhs (stmt); | |
6117 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
6118 | return false; | |
6119 | ||
6120 | code = gimple_assign_rhs_code (stmt); | |
6121 | if (code != ARRAY_REF | |
6122 | && code != BIT_FIELD_REF | |
6123 | && code != INDIRECT_REF | |
6124 | && code != COMPONENT_REF | |
6125 | && code != IMAGPART_EXPR | |
6126 | && code != REALPART_EXPR | |
6127 | && code != MEM_REF | |
6128 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
6129 | return false; | |
6130 | ||
6131 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
6132 | return false; | |
6133 | ||
6134 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
6135 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
6136 | ||
a70d6342 IR |
6137 | if (loop_vinfo) |
6138 | { | |
6139 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6140 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
6141 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6142 | } | |
6143 | else | |
3533e503 | 6144 | vf = 1; |
ebfd146a IR |
6145 | |
6146 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 6147 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 6148 | case of SLP. */ |
437f4a00 | 6149 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
6150 | ncopies = 1; |
6151 | else | |
6152 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
6153 | ||
6154 | gcc_assert (ncopies >= 1); | |
6155 | ||
6156 | /* FORNOW. This restriction should be relaxed. */ | |
6157 | if (nested_in_vect_loop && ncopies > 1) | |
6158 | { | |
73fbfcad | 6159 | if (dump_enabled_p ()) |
78c60e3d | 6160 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6161 | "multiple types in nested loop.\n"); |
ebfd146a IR |
6162 | return false; |
6163 | } | |
6164 | ||
f2556b68 RB |
6165 | /* Invalidate assumptions made by dependence analysis when vectorization |
6166 | on the unrolled body effectively re-orders stmts. */ | |
6167 | if (ncopies > 1 | |
6168 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6169 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6170 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6171 | { | |
6172 | if (dump_enabled_p ()) | |
6173 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6174 | "cannot perform implicit CSE when unrolling " | |
6175 | "with negative dependence distance\n"); | |
6176 | return false; | |
6177 | } | |
6178 | ||
7b7b1813 | 6179 | elem_type = TREE_TYPE (vectype); |
947131ba | 6180 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
6181 | |
6182 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6183 | (e.g. - data copies). */ | |
947131ba | 6184 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 6185 | { |
73fbfcad | 6186 | if (dump_enabled_p ()) |
78c60e3d | 6187 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6188 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
6189 | return false; |
6190 | } | |
6191 | ||
ebfd146a | 6192 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 6193 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 6194 | { |
0d0293ac | 6195 | grouped_load = true; |
ebfd146a | 6196 | /* FORNOW */ |
3bab6342 | 6197 | gcc_assert (!nested_in_vect_loop && !STMT_VINFO_GATHER_SCATTER_P (stmt_info)); |
ebfd146a | 6198 | |
e14c1050 | 6199 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
d5f035ea RB |
6200 | |
6201 | /* If this is single-element interleaving with an element distance | |
6202 | that leaves unused vector loads around punt - we at least create | |
6203 | very sub-optimal code in that case (and blow up memory, | |
6204 | see PR65518). */ | |
6205 | if (first_stmt == stmt | |
6206 | && !GROUP_NEXT_ELEMENT (stmt_info) | |
6207 | && GROUP_SIZE (stmt_info) > TYPE_VECTOR_SUBPARTS (vectype)) | |
6208 | { | |
6209 | if (dump_enabled_p ()) | |
6210 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6211 | "single-element interleaving not supported " | |
6212 | "for not adjacent vector loads\n"); | |
6213 | return false; | |
6214 | } | |
6215 | ||
b1af7da6 RB |
6216 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
6217 | slp_perm = true; | |
6218 | ||
7b5fc413 RB |
6219 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
6220 | if (!slp | |
6221 | && !PURE_SLP_STMT (stmt_info) | |
f2e2a985 | 6222 | && !STMT_VINFO_STRIDED_P (stmt_info)) |
b602d918 | 6223 | { |
272c6793 RS |
6224 | if (vect_load_lanes_supported (vectype, group_size)) |
6225 | load_lanes_p = true; | |
0d0293ac | 6226 | else if (!vect_grouped_load_supported (vectype, group_size)) |
b602d918 RS |
6227 | return false; |
6228 | } | |
f2556b68 RB |
6229 | |
6230 | /* Invalidate assumptions made by dependence analysis when vectorization | |
6231 | on the unrolled body effectively re-orders stmts. */ | |
6232 | if (!PURE_SLP_STMT (stmt_info) | |
6233 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6234 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6235 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6236 | { | |
6237 | if (dump_enabled_p ()) | |
6238 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6239 | "cannot perform implicit CSE when performing " | |
6240 | "group loads with negative dependence distance\n"); | |
6241 | return false; | |
6242 | } | |
96bb56b2 RB |
6243 | |
6244 | /* Similarly when the stmt is a load that is both part of a SLP | |
6245 | instance and a loop vectorized stmt via the same-dr mechanism | |
6246 | we have to give up. */ | |
6247 | if (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info) | |
6248 | && (STMT_SLP_TYPE (stmt_info) | |
6249 | != STMT_SLP_TYPE (vinfo_for_stmt | |
6250 | (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info))))) | |
6251 | { | |
6252 | if (dump_enabled_p ()) | |
6253 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6254 | "conflicting SLP types for CSEd load\n"); | |
6255 | return false; | |
6256 | } | |
ebfd146a IR |
6257 | } |
6258 | ||
a1e53f3f | 6259 | |
3bab6342 | 6260 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
aec7ae7d | 6261 | { |
355fe088 | 6262 | gimple *def_stmt; |
3bab6342 AT |
6263 | gather_decl = vect_check_gather_scatter (stmt, loop_vinfo, &gather_base, |
6264 | &gather_off, &gather_scale); | |
aec7ae7d | 6265 | gcc_assert (gather_decl); |
81c40241 RB |
6266 | if (!vect_is_simple_use (gather_off, vinfo, &def_stmt, &gather_dt, |
6267 | &gather_off_vectype)) | |
aec7ae7d | 6268 | { |
73fbfcad | 6269 | if (dump_enabled_p ()) |
78c60e3d | 6270 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6271 | "gather index use not simple.\n"); |
aec7ae7d JJ |
6272 | return false; |
6273 | } | |
6274 | } | |
f2e2a985 | 6275 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
7b5fc413 RB |
6276 | { |
6277 | if ((grouped_load | |
6278 | && (slp || PURE_SLP_STMT (stmt_info))) | |
6279 | && (group_size > nunits | |
b266b968 | 6280 | || nunits % group_size != 0)) |
7b5fc413 RB |
6281 | { |
6282 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6283 | "unhandled strided group load\n"); | |
6284 | return false; | |
6285 | } | |
6286 | } | |
319e6439 RG |
6287 | else |
6288 | { | |
6289 | negative = tree_int_cst_compare (nested_in_vect_loop | |
6290 | ? STMT_VINFO_DR_STEP (stmt_info) | |
6291 | : DR_STEP (dr), | |
6292 | size_zero_node) < 0; | |
6293 | if (negative && ncopies > 1) | |
6294 | { | |
73fbfcad | 6295 | if (dump_enabled_p ()) |
78c60e3d | 6296 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6297 | "multiple types with negative step.\n"); |
319e6439 RG |
6298 | return false; |
6299 | } | |
6300 | ||
6301 | if (negative) | |
6302 | { | |
08940f33 RB |
6303 | if (grouped_load) |
6304 | { | |
6305 | if (dump_enabled_p ()) | |
6306 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 TJ |
6307 | "negative step for group load not supported" |
6308 | "\n"); | |
08940f33 RB |
6309 | return false; |
6310 | } | |
319e6439 RG |
6311 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
6312 | if (alignment_support_scheme != dr_aligned | |
6313 | && alignment_support_scheme != dr_unaligned_supported) | |
6314 | { | |
73fbfcad | 6315 | if (dump_enabled_p ()) |
78c60e3d | 6316 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6317 | "negative step but alignment required.\n"); |
319e6439 RG |
6318 | return false; |
6319 | } | |
6320 | if (!perm_mask_for_reverse (vectype)) | |
6321 | { | |
73fbfcad | 6322 | if (dump_enabled_p ()) |
78c60e3d | 6323 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
6324 | "negative step and reversing not supported." |
6325 | "\n"); | |
319e6439 RG |
6326 | return false; |
6327 | } | |
6328 | } | |
7d75abc8 | 6329 | } |
aec7ae7d | 6330 | |
ebfd146a IR |
6331 | if (!vec_stmt) /* transformation not required. */ |
6332 | { | |
6333 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
2e8ab70c RB |
6334 | /* The SLP costs are calculated during SLP analysis. */ |
6335 | if (!PURE_SLP_STMT (stmt_info)) | |
6336 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, | |
6337 | NULL, NULL, NULL); | |
ebfd146a IR |
6338 | return true; |
6339 | } | |
6340 | ||
73fbfcad | 6341 | if (dump_enabled_p ()) |
78c60e3d | 6342 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6343 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a IR |
6344 | |
6345 | /** Transform. **/ | |
6346 | ||
c716e67f XDL |
6347 | ensure_base_align (stmt_info, dr); |
6348 | ||
3bab6342 | 6349 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
aec7ae7d JJ |
6350 | { |
6351 | tree vec_oprnd0 = NULL_TREE, op; | |
6352 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
6353 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
d3c2fee0 | 6354 | tree ptr, mask, var, scale, merge, perm_mask = NULL_TREE, prev_res = NULL_TREE; |
aec7ae7d JJ |
6355 | edge pe = loop_preheader_edge (loop); |
6356 | gimple_seq seq; | |
6357 | basic_block new_bb; | |
6358 | enum { NARROW, NONE, WIDEN } modifier; | |
6359 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
6360 | ||
6361 | if (nunits == gather_off_nunits) | |
6362 | modifier = NONE; | |
6363 | else if (nunits == gather_off_nunits / 2) | |
6364 | { | |
6365 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
6366 | modifier = WIDEN; | |
6367 | ||
6368 | for (i = 0; i < gather_off_nunits; ++i) | |
6369 | sel[i] = i | nunits; | |
6370 | ||
557be5a8 | 6371 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
aec7ae7d JJ |
6372 | } |
6373 | else if (nunits == gather_off_nunits * 2) | |
6374 | { | |
6375 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
6376 | modifier = NARROW; | |
6377 | ||
6378 | for (i = 0; i < nunits; ++i) | |
6379 | sel[i] = i < gather_off_nunits | |
6380 | ? i : i + nunits - gather_off_nunits; | |
6381 | ||
557be5a8 | 6382 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
aec7ae7d JJ |
6383 | ncopies *= 2; |
6384 | } | |
6385 | else | |
6386 | gcc_unreachable (); | |
6387 | ||
6388 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
6389 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6390 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6391 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6392 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6393 | scaletype = TREE_VALUE (arglist); | |
d3c2fee0 | 6394 | gcc_checking_assert (types_compatible_p (srctype, rettype)); |
aec7ae7d JJ |
6395 | |
6396 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
6397 | ||
6398 | ptr = fold_convert (ptrtype, gather_base); | |
6399 | if (!is_gimple_min_invariant (ptr)) | |
6400 | { | |
6401 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6402 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6403 | gcc_assert (!new_bb); | |
6404 | } | |
6405 | ||
6406 | /* Currently we support only unconditional gather loads, | |
6407 | so mask should be all ones. */ | |
d3c2fee0 AI |
6408 | if (TREE_CODE (masktype) == INTEGER_TYPE) |
6409 | mask = build_int_cst (masktype, -1); | |
6410 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
6411 | { | |
6412 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
6413 | mask = build_vector_from_val (masktype, mask); | |
03b9e8e4 | 6414 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
d3c2fee0 | 6415 | } |
aec7ae7d JJ |
6416 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) |
6417 | { | |
6418 | REAL_VALUE_TYPE r; | |
6419 | long tmp[6]; | |
6420 | for (j = 0; j < 6; ++j) | |
6421 | tmp[j] = -1; | |
6422 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
6423 | mask = build_real (TREE_TYPE (masktype), r); | |
d3c2fee0 | 6424 | mask = build_vector_from_val (masktype, mask); |
03b9e8e4 | 6425 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
aec7ae7d JJ |
6426 | } |
6427 | else | |
6428 | gcc_unreachable (); | |
aec7ae7d JJ |
6429 | |
6430 | scale = build_int_cst (scaletype, gather_scale); | |
6431 | ||
d3c2fee0 AI |
6432 | if (TREE_CODE (TREE_TYPE (rettype)) == INTEGER_TYPE) |
6433 | merge = build_int_cst (TREE_TYPE (rettype), 0); | |
6434 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (rettype))) | |
6435 | { | |
6436 | REAL_VALUE_TYPE r; | |
6437 | long tmp[6]; | |
6438 | for (j = 0; j < 6; ++j) | |
6439 | tmp[j] = 0; | |
6440 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (rettype))); | |
6441 | merge = build_real (TREE_TYPE (rettype), r); | |
6442 | } | |
6443 | else | |
6444 | gcc_unreachable (); | |
6445 | merge = build_vector_from_val (rettype, merge); | |
6446 | merge = vect_init_vector (stmt, merge, rettype, NULL); | |
6447 | ||
aec7ae7d JJ |
6448 | prev_stmt_info = NULL; |
6449 | for (j = 0; j < ncopies; ++j) | |
6450 | { | |
6451 | if (modifier == WIDEN && (j & 1)) | |
6452 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
6453 | perm_mask, stmt, gsi); | |
6454 | else if (j == 0) | |
6455 | op = vec_oprnd0 | |
81c40241 | 6456 | = vect_get_vec_def_for_operand (gather_off, stmt); |
aec7ae7d JJ |
6457 | else |
6458 | op = vec_oprnd0 | |
6459 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
6460 | ||
6461 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6462 | { | |
6463 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
6464 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 6465 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
aec7ae7d JJ |
6466 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
6467 | new_stmt | |
0d0e4a03 | 6468 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6469 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6470 | op = var; | |
6471 | } | |
6472 | ||
6473 | new_stmt | |
d3c2fee0 | 6474 | = gimple_build_call (gather_decl, 5, merge, ptr, op, mask, scale); |
aec7ae7d JJ |
6475 | |
6476 | if (!useless_type_conversion_p (vectype, rettype)) | |
6477 | { | |
6478 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
6479 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 6480 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
aec7ae7d JJ |
6481 | gimple_call_set_lhs (new_stmt, op); |
6482 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 6483 | var = make_ssa_name (vec_dest); |
aec7ae7d JJ |
6484 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
6485 | new_stmt | |
0d0e4a03 | 6486 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6487 | } |
6488 | else | |
6489 | { | |
6490 | var = make_ssa_name (vec_dest, new_stmt); | |
6491 | gimple_call_set_lhs (new_stmt, var); | |
6492 | } | |
6493 | ||
6494 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6495 | ||
6496 | if (modifier == NARROW) | |
6497 | { | |
6498 | if ((j & 1) == 0) | |
6499 | { | |
6500 | prev_res = var; | |
6501 | continue; | |
6502 | } | |
6503 | var = permute_vec_elements (prev_res, var, | |
6504 | perm_mask, stmt, gsi); | |
6505 | new_stmt = SSA_NAME_DEF_STMT (var); | |
6506 | } | |
6507 | ||
6508 | if (prev_stmt_info == NULL) | |
6509 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6510 | else | |
6511 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6512 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6513 | } | |
6514 | return true; | |
6515 | } | |
f2e2a985 | 6516 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
7d75abc8 MM |
6517 | { |
6518 | gimple_stmt_iterator incr_gsi; | |
6519 | bool insert_after; | |
355fe088 | 6520 | gimple *incr; |
7d75abc8 | 6521 | tree offvar; |
7d75abc8 MM |
6522 | tree ivstep; |
6523 | tree running_off; | |
9771b263 | 6524 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 6525 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
6526 | tree stride_base, stride_step, alias_off; |
6527 | ||
6528 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 6529 | |
f502d50e | 6530 | if (slp && grouped_load) |
ab313a8c RB |
6531 | first_dr = STMT_VINFO_DATA_REF |
6532 | (vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info))); | |
6533 | else | |
6534 | first_dr = dr; | |
6535 | ||
14ac6aa2 RB |
6536 | stride_base |
6537 | = fold_build_pointer_plus | |
ab313a8c | 6538 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 6539 | size_binop (PLUS_EXPR, |
ab313a8c RB |
6540 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
6541 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
6542 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
6543 | |
6544 | /* For a load with loop-invariant (but other than power-of-2) | |
6545 | stride (i.e. not a grouped access) like so: | |
6546 | ||
6547 | for (i = 0; i < n; i += stride) | |
6548 | ... = array[i]; | |
6549 | ||
6550 | we generate a new induction variable and new accesses to | |
6551 | form a new vector (or vectors, depending on ncopies): | |
6552 | ||
6553 | for (j = 0; ; j += VF*stride) | |
6554 | tmp1 = array[j]; | |
6555 | tmp2 = array[j + stride]; | |
6556 | ... | |
6557 | vectemp = {tmp1, tmp2, ...} | |
6558 | */ | |
6559 | ||
ab313a8c RB |
6560 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
6561 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
6562 | |
6563 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6564 | ||
ab313a8c | 6565 | create_iv (unshare_expr (stride_base), unshare_expr (ivstep), NULL, |
7d75abc8 MM |
6566 | loop, &incr_gsi, insert_after, |
6567 | &offvar, NULL); | |
6568 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 6569 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
7d75abc8 | 6570 | |
ab313a8c RB |
6571 | stride_step = force_gimple_operand (unshare_expr (stride_step), |
6572 | &stmts, true, NULL_TREE); | |
7d75abc8 MM |
6573 | if (stmts) |
6574 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
6575 | ||
6576 | prev_stmt_info = NULL; | |
6577 | running_off = offvar; | |
ab313a8c | 6578 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (first_dr)), 0); |
7b5fc413 RB |
6579 | int nloads = nunits; |
6580 | tree ltype = TREE_TYPE (vectype); | |
b266b968 | 6581 | auto_vec<tree> dr_chain; |
7b5fc413 RB |
6582 | if (slp) |
6583 | { | |
6584 | nloads = nunits / group_size; | |
6585 | if (group_size < nunits) | |
6586 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); | |
6587 | else | |
6588 | ltype = vectype; | |
6589 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); | |
6590 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
b266b968 RB |
6591 | if (slp_perm) |
6592 | dr_chain.create (ncopies); | |
7b5fc413 | 6593 | } |
7d75abc8 MM |
6594 | for (j = 0; j < ncopies; j++) |
6595 | { | |
6596 | tree vec_inv; | |
6597 | ||
7b5fc413 RB |
6598 | if (nloads > 1) |
6599 | { | |
6600 | vec_alloc (v, nloads); | |
6601 | for (i = 0; i < nloads; i++) | |
6602 | { | |
6603 | tree newref, newoff; | |
355fe088 | 6604 | gimple *incr; |
7b5fc413 RB |
6605 | newref = build2 (MEM_REF, ltype, running_off, alias_off); |
6606 | ||
6607 | newref = force_gimple_operand_gsi (gsi, newref, true, | |
6608 | NULL_TREE, true, | |
6609 | GSI_SAME_STMT); | |
6610 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, newref); | |
6611 | newoff = copy_ssa_name (running_off); | |
6612 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6613 | running_off, stride_step); | |
6614 | vect_finish_stmt_generation (stmt, incr, gsi); | |
6615 | ||
6616 | running_off = newoff; | |
6617 | } | |
6618 | ||
6619 | vec_inv = build_constructor (vectype, v); | |
6620 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
6621 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
6622 | } | |
6623 | else | |
7d75abc8 | 6624 | { |
7b5fc413 RB |
6625 | new_stmt = gimple_build_assign (make_ssa_name (ltype), |
6626 | build2 (MEM_REF, ltype, | |
6627 | running_off, alias_off)); | |
6628 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6629 | ||
6630 | tree newoff = copy_ssa_name (running_off); | |
355fe088 | 6631 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
0d0e4a03 | 6632 | running_off, stride_step); |
7d75abc8 MM |
6633 | vect_finish_stmt_generation (stmt, incr, gsi); |
6634 | ||
6635 | running_off = newoff; | |
6636 | } | |
6637 | ||
7b5fc413 | 6638 | if (slp) |
b266b968 RB |
6639 | { |
6640 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
6641 | if (slp_perm) | |
6642 | dr_chain.quick_push (gimple_assign_lhs (new_stmt)); | |
6643 | } | |
7d75abc8 | 6644 | else |
225ce44b RB |
6645 | { |
6646 | if (j == 0) | |
6647 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6648 | else | |
6649 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6650 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6651 | } | |
7d75abc8 | 6652 | } |
b266b968 RB |
6653 | if (slp_perm) |
6654 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
6655 | slp_node_instance, false); | |
7d75abc8 MM |
6656 | return true; |
6657 | } | |
aec7ae7d | 6658 | |
0d0293ac | 6659 | if (grouped_load) |
ebfd146a | 6660 | { |
e14c1050 | 6661 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
52eab378 RB |
6662 | /* For BB vectorization we directly vectorize a subchain |
6663 | without permutation. */ | |
6664 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
9771b263 | 6665 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
6aa904c4 | 6666 | |
ebfd146a | 6667 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
6668 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
6669 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
6670 | ??? But we can only do so if there is exactly one | |
6671 | as we have no way to get at the rest. Leave the CSE | |
6672 | opportunity alone. | |
6673 | ??? With the group load eventually participating | |
6674 | in multiple different permutations (having multiple | |
6675 | slp nodes which refer to the same group) the CSE | |
6676 | is even wrong code. See PR56270. */ | |
6677 | && !slp) | |
ebfd146a IR |
6678 | { |
6679 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
6680 | return true; | |
6681 | } | |
6682 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 6683 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
9b999e8c | 6684 | group_gap_adj = 0; |
ebfd146a IR |
6685 | |
6686 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
6687 | if (slp) | |
6688 | { | |
0d0293ac | 6689 | grouped_load = false; |
91ff1504 RB |
6690 | /* For SLP permutation support we need to load the whole group, |
6691 | not only the number of vector stmts the permutation result | |
6692 | fits in. */ | |
6693 | if (slp_perm) | |
6694 | vec_num = (group_size * vf + nunits - 1) / nunits; | |
6695 | else | |
6696 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9b999e8c | 6697 | group_gap_adj = vf * group_size - nunits * vec_num; |
a70d6342 | 6698 | } |
ebfd146a | 6699 | else |
9b999e8c | 6700 | vec_num = group_size; |
ebfd146a IR |
6701 | } |
6702 | else | |
6703 | { | |
6704 | first_stmt = stmt; | |
6705 | first_dr = dr; | |
6706 | group_size = vec_num = 1; | |
9b999e8c | 6707 | group_gap_adj = 0; |
ebfd146a IR |
6708 | } |
6709 | ||
720f5239 | 6710 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6711 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
6712 | /* Targets with load-lane instructions must not require explicit |
6713 | realignment. */ | |
6714 | gcc_assert (!load_lanes_p | |
6715 | || alignment_support_scheme == dr_aligned | |
6716 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
6717 | |
6718 | /* In case the vectorization factor (VF) is bigger than the number | |
6719 | of elements that we can fit in a vectype (nunits), we have to generate | |
6720 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 6721 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 6722 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 6723 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 6724 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
6725 | stmts that use the defs of the current stmt. The example below |
6726 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
6727 | need to create 4 vectorized stmts): | |
ebfd146a IR |
6728 | |
6729 | before vectorization: | |
6730 | RELATED_STMT VEC_STMT | |
6731 | S1: x = memref - - | |
6732 | S2: z = x + 1 - - | |
6733 | ||
6734 | step 1: vectorize stmt S1: | |
6735 | We first create the vector stmt VS1_0, and, as usual, record a | |
6736 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
6737 | Next, we create the vector stmt VS1_1, and record a pointer to | |
6738 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 6739 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
6740 | stmts and pointers: |
6741 | RELATED_STMT VEC_STMT | |
6742 | VS1_0: vx0 = memref0 VS1_1 - | |
6743 | VS1_1: vx1 = memref1 VS1_2 - | |
6744 | VS1_2: vx2 = memref2 VS1_3 - | |
6745 | VS1_3: vx3 = memref3 - - | |
6746 | S1: x = load - VS1_0 | |
6747 | S2: z = x + 1 - - | |
6748 | ||
b8698a0f L |
6749 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
6750 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
6751 | stmt S2. */ |
6752 | ||
0d0293ac | 6753 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6754 | |
6755 | S1: x2 = &base + 2 | |
6756 | S2: x0 = &base | |
6757 | S3: x1 = &base + 1 | |
6758 | S4: x3 = &base + 3 | |
6759 | ||
b8698a0f | 6760 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
6761 | starting from the access of the first stmt of the chain: |
6762 | ||
6763 | VS1: vx0 = &base | |
6764 | VS2: vx1 = &base + vec_size*1 | |
6765 | VS3: vx3 = &base + vec_size*2 | |
6766 | VS4: vx4 = &base + vec_size*3 | |
6767 | ||
6768 | Then permutation statements are generated: | |
6769 | ||
e2c83630 RH |
6770 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
6771 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
6772 | ... |
6773 | ||
6774 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
6775 | (the order of the data-refs in the output of vect_permute_load_chain | |
6776 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6777 | the documentation of vect_permute_load_chain()). | |
6778 | The generation of permutation stmts and recording them in | |
0d0293ac | 6779 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 6780 | |
b8698a0f | 6781 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
6782 | permutation stmts above are created for every copy. The result vector |
6783 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
6784 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
6785 | |
6786 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
6787 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
6788 | we generate the following code: | |
6789 | p = initial_addr; | |
6790 | indx = 0; | |
6791 | loop { | |
6792 | p = p + indx * vectype_size; | |
6793 | vec_dest = *(p); | |
6794 | indx = indx + 1; | |
6795 | } | |
6796 | ||
6797 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 6798 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
6799 | then generate the following code, in which the data in each iteration is |
6800 | obtained by two vector loads, one from the previous iteration, and one | |
6801 | from the current iteration: | |
6802 | p1 = initial_addr; | |
6803 | msq_init = *(floor(p1)) | |
6804 | p2 = initial_addr + VS - 1; | |
6805 | realignment_token = call target_builtin; | |
6806 | indx = 0; | |
6807 | loop { | |
6808 | p2 = p2 + indx * vectype_size | |
6809 | lsq = *(floor(p2)) | |
6810 | vec_dest = realign_load (msq, lsq, realignment_token) | |
6811 | indx = indx + 1; | |
6812 | msq = lsq; | |
6813 | } */ | |
6814 | ||
6815 | /* If the misalignment remains the same throughout the execution of the | |
6816 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 6817 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
6818 | This can only occur when vectorizing memory accesses in the inner-loop |
6819 | nested within an outer-loop that is being vectorized. */ | |
6820 | ||
d1e4b493 | 6821 | if (nested_in_vect_loop |
211bea38 | 6822 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
ebfd146a IR |
6823 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) |
6824 | { | |
6825 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
6826 | compute_in_loop = true; | |
6827 | } | |
6828 | ||
6829 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
6830 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 6831 | && !compute_in_loop) |
ebfd146a IR |
6832 | { |
6833 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
6834 | alignment_support_scheme, NULL_TREE, | |
6835 | &at_loop); | |
6836 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
6837 | { | |
538dd0b7 | 6838 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
6839 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
6840 | size_one_node); | |
ebfd146a IR |
6841 | } |
6842 | } | |
6843 | else | |
6844 | at_loop = loop; | |
6845 | ||
a1e53f3f L |
6846 | if (negative) |
6847 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
6848 | ||
272c6793 RS |
6849 | if (load_lanes_p) |
6850 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
6851 | else | |
6852 | aggr_type = vectype; | |
6853 | ||
ebfd146a IR |
6854 | prev_stmt_info = NULL; |
6855 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 6856 | { |
272c6793 | 6857 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 6858 | if (j == 0) |
74bf76ed JJ |
6859 | { |
6860 | bool simd_lane_access_p | |
6861 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
6862 | if (simd_lane_access_p | |
6863 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
6864 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
6865 | && integer_zerop (DR_OFFSET (first_dr)) | |
6866 | && integer_zerop (DR_INIT (first_dr)) | |
6867 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
6868 | get_alias_set (DR_REF (first_dr))) | |
6869 | && (alignment_support_scheme == dr_aligned | |
6870 | || alignment_support_scheme == dr_unaligned_supported)) | |
6871 | { | |
6872 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
6873 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
6874 | (DR_REF (first_dr)), 0); | |
8928eff3 | 6875 | inv_p = false; |
74bf76ed JJ |
6876 | } |
6877 | else | |
6878 | dataref_ptr | |
6879 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
6880 | offset, &dummy, gsi, &ptr_incr, | |
356bbc4c JJ |
6881 | simd_lane_access_p, &inv_p, |
6882 | byte_offset); | |
74bf76ed JJ |
6883 | } |
6884 | else if (dataref_offset) | |
6885 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
6886 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 6887 | else |
272c6793 RS |
6888 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
6889 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 6890 | |
0d0293ac | 6891 | if (grouped_load || slp_perm) |
9771b263 | 6892 | dr_chain.create (vec_num); |
5ce1ee7f | 6893 | |
272c6793 | 6894 | if (load_lanes_p) |
ebfd146a | 6895 | { |
272c6793 RS |
6896 | tree vec_array; |
6897 | ||
6898 | vec_array = create_vector_array (vectype, vec_num); | |
6899 | ||
6900 | /* Emit: | |
6901 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
6902 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
6903 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
6904 | gimple_call_set_lhs (new_stmt, vec_array); | |
6905 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 6906 | |
272c6793 RS |
6907 | /* Extract each vector into an SSA_NAME. */ |
6908 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 6909 | { |
272c6793 RS |
6910 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
6911 | vec_array, i); | |
9771b263 | 6912 | dr_chain.quick_push (new_temp); |
272c6793 RS |
6913 | } |
6914 | ||
6915 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 6916 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
6917 | } |
6918 | else | |
6919 | { | |
6920 | for (i = 0; i < vec_num; i++) | |
6921 | { | |
6922 | if (i > 0) | |
6923 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
6924 | stmt, NULL_TREE); | |
6925 | ||
6926 | /* 2. Create the vector-load in the loop. */ | |
6927 | switch (alignment_support_scheme) | |
6928 | { | |
6929 | case dr_aligned: | |
6930 | case dr_unaligned_supported: | |
be1ac4ec | 6931 | { |
644ffefd MJ |
6932 | unsigned int align, misalign; |
6933 | ||
272c6793 | 6934 | data_ref |
aed93b23 RB |
6935 | = fold_build2 (MEM_REF, vectype, dataref_ptr, |
6936 | dataref_offset | |
6937 | ? dataref_offset | |
6938 | : build_int_cst (reference_alias_ptr_type | |
6939 | (DR_REF (first_dr)), 0)); | |
644ffefd | 6940 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 RS |
6941 | if (alignment_support_scheme == dr_aligned) |
6942 | { | |
6943 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 6944 | misalign = 0; |
272c6793 RS |
6945 | } |
6946 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
6947 | { | |
52639a61 RB |
6948 | if (DR_VECT_AUX (first_dr)->base_element_aligned) |
6949 | align = TYPE_ALIGN_UNIT (elem_type); | |
6950 | else | |
6951 | align = (get_object_alignment (DR_REF (first_dr)) | |
6952 | / BITS_PER_UNIT); | |
6953 | misalign = 0; | |
272c6793 RS |
6954 | TREE_TYPE (data_ref) |
6955 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 6956 | align * BITS_PER_UNIT); |
272c6793 RS |
6957 | } |
6958 | else | |
6959 | { | |
6960 | TREE_TYPE (data_ref) | |
6961 | = build_aligned_type (TREE_TYPE (data_ref), | |
6962 | TYPE_ALIGN (elem_type)); | |
644ffefd | 6963 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 6964 | } |
aed93b23 RB |
6965 | if (dataref_offset == NULL_TREE |
6966 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
6967 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
6968 | align, misalign); | |
272c6793 | 6969 | break; |
be1ac4ec | 6970 | } |
272c6793 | 6971 | case dr_explicit_realign: |
267d3070 | 6972 | { |
272c6793 | 6973 | tree ptr, bump; |
272c6793 | 6974 | |
d88981fc | 6975 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
6976 | |
6977 | if (compute_in_loop) | |
6978 | msq = vect_setup_realignment (first_stmt, gsi, | |
6979 | &realignment_token, | |
6980 | dr_explicit_realign, | |
6981 | dataref_ptr, NULL); | |
6982 | ||
aed93b23 RB |
6983 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
6984 | ptr = copy_ssa_name (dataref_ptr); | |
6985 | else | |
6986 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
6987 | new_stmt = gimple_build_assign |
6988 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
6989 | build_int_cst |
6990 | (TREE_TYPE (dataref_ptr), | |
6991 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
6992 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6993 | data_ref | |
6994 | = build2 (MEM_REF, vectype, ptr, | |
6995 | build_int_cst (reference_alias_ptr_type | |
6996 | (DR_REF (first_dr)), 0)); | |
6997 | vec_dest = vect_create_destination_var (scalar_dest, | |
6998 | vectype); | |
6999 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
7000 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
7001 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7002 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
7003 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
7004 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7005 | msq = new_temp; | |
7006 | ||
d88981fc | 7007 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 7008 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 7009 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 7010 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
7011 | new_stmt = gimple_build_assign |
7012 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 RS |
7013 | build_int_cst |
7014 | (TREE_TYPE (ptr), | |
7015 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
aed93b23 | 7016 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
7017 | gimple_assign_set_lhs (new_stmt, ptr); |
7018 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7019 | data_ref | |
7020 | = build2 (MEM_REF, vectype, ptr, | |
7021 | build_int_cst (reference_alias_ptr_type | |
7022 | (DR_REF (first_dr)), 0)); | |
7023 | break; | |
267d3070 | 7024 | } |
272c6793 | 7025 | case dr_explicit_realign_optimized: |
aed93b23 RB |
7026 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7027 | new_temp = copy_ssa_name (dataref_ptr); | |
7028 | else | |
7029 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
7030 | new_stmt = gimple_build_assign |
7031 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7032 | build_int_cst |
7033 | (TREE_TYPE (dataref_ptr), | |
7034 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
7035 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7036 | data_ref | |
7037 | = build2 (MEM_REF, vectype, new_temp, | |
7038 | build_int_cst (reference_alias_ptr_type | |
7039 | (DR_REF (first_dr)), 0)); | |
7040 | break; | |
7041 | default: | |
7042 | gcc_unreachable (); | |
7043 | } | |
ebfd146a | 7044 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 7045 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
7046 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7047 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7048 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7049 | ||
272c6793 RS |
7050 | /* 3. Handle explicit realignment if necessary/supported. |
7051 | Create in loop: | |
7052 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
7053 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
7054 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 7055 | { |
272c6793 RS |
7056 | lsq = gimple_assign_lhs (new_stmt); |
7057 | if (!realignment_token) | |
7058 | realignment_token = dataref_ptr; | |
7059 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
7060 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
7061 | msq, lsq, realignment_token); | |
272c6793 RS |
7062 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7063 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7064 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7065 | ||
7066 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7067 | { | |
7068 | gcc_assert (phi); | |
7069 | if (i == vec_num - 1 && j == ncopies - 1) | |
7070 | add_phi_arg (phi, lsq, | |
7071 | loop_latch_edge (containing_loop), | |
9e227d60 | 7072 | UNKNOWN_LOCATION); |
272c6793 RS |
7073 | msq = lsq; |
7074 | } | |
ebfd146a | 7075 | } |
ebfd146a | 7076 | |
59fd17e3 RB |
7077 | /* 4. Handle invariant-load. */ |
7078 | if (inv_p && !bb_vinfo) | |
7079 | { | |
59fd17e3 | 7080 | gcc_assert (!grouped_load); |
d1417442 JJ |
7081 | /* If we have versioned for aliasing or the loop doesn't |
7082 | have any data dependencies that would preclude this, | |
7083 | then we are sure this is a loop invariant load and | |
7084 | thus we can insert it on the preheader edge. */ | |
7085 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
7086 | && !nested_in_vect_loop | |
6b916b36 | 7087 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
7088 | { |
7089 | if (dump_enabled_p ()) | |
7090 | { | |
7091 | dump_printf_loc (MSG_NOTE, vect_location, | |
7092 | "hoisting out of the vectorized " | |
7093 | "loop: "); | |
7094 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 7095 | } |
b731b390 | 7096 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
7097 | gsi_insert_on_edge_immediate |
7098 | (loop_preheader_edge (loop), | |
7099 | gimple_build_assign (tem, | |
7100 | unshare_expr | |
7101 | (gimple_assign_rhs1 (stmt)))); | |
7102 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
7103 | } | |
7104 | else | |
7105 | { | |
7106 | gimple_stmt_iterator gsi2 = *gsi; | |
7107 | gsi_next (&gsi2); | |
7108 | new_temp = vect_init_vector (stmt, scalar_dest, | |
7109 | vectype, &gsi2); | |
7110 | } | |
59fd17e3 | 7111 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
a0e35eb0 | 7112 | set_vinfo_for_stmt (new_stmt, |
310213d4 | 7113 | new_stmt_vec_info (new_stmt, vinfo)); |
59fd17e3 RB |
7114 | } |
7115 | ||
272c6793 RS |
7116 | if (negative) |
7117 | { | |
aec7ae7d JJ |
7118 | tree perm_mask = perm_mask_for_reverse (vectype); |
7119 | new_temp = permute_vec_elements (new_temp, new_temp, | |
7120 | perm_mask, stmt, gsi); | |
ebfd146a IR |
7121 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7122 | } | |
267d3070 | 7123 | |
272c6793 | 7124 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
7125 | vect_transform_grouped_load (). */ |
7126 | if (grouped_load || slp_perm) | |
9771b263 | 7127 | dr_chain.quick_push (new_temp); |
267d3070 | 7128 | |
272c6793 RS |
7129 | /* Store vector loads in the corresponding SLP_NODE. */ |
7130 | if (slp && !slp_perm) | |
9771b263 | 7131 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
272c6793 | 7132 | } |
9b999e8c RB |
7133 | /* Bump the vector pointer to account for a gap or for excess |
7134 | elements loaded for a permuted SLP load. */ | |
7135 | if (group_gap_adj != 0) | |
a64b9c26 | 7136 | { |
9b999e8c RB |
7137 | bool ovf; |
7138 | tree bump | |
7139 | = wide_int_to_tree (sizetype, | |
7140 | wi::smul (TYPE_SIZE_UNIT (elem_type), | |
7141 | group_gap_adj, &ovf)); | |
a64b9c26 RB |
7142 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
7143 | stmt, bump); | |
7144 | } | |
ebfd146a IR |
7145 | } |
7146 | ||
7147 | if (slp && !slp_perm) | |
7148 | continue; | |
7149 | ||
7150 | if (slp_perm) | |
7151 | { | |
01d8bf07 | 7152 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
ebfd146a IR |
7153 | slp_node_instance, false)) |
7154 | { | |
9771b263 | 7155 | dr_chain.release (); |
ebfd146a IR |
7156 | return false; |
7157 | } | |
7158 | } | |
7159 | else | |
7160 | { | |
0d0293ac | 7161 | if (grouped_load) |
ebfd146a | 7162 | { |
272c6793 | 7163 | if (!load_lanes_p) |
0d0293ac | 7164 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 7165 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
7166 | } |
7167 | else | |
7168 | { | |
7169 | if (j == 0) | |
7170 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7171 | else | |
7172 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7173 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7174 | } | |
7175 | } | |
9771b263 | 7176 | dr_chain.release (); |
ebfd146a IR |
7177 | } |
7178 | ||
ebfd146a IR |
7179 | return true; |
7180 | } | |
7181 | ||
7182 | /* Function vect_is_simple_cond. | |
b8698a0f | 7183 | |
ebfd146a IR |
7184 | Input: |
7185 | LOOP - the loop that is being vectorized. | |
7186 | COND - Condition that is checked for simple use. | |
7187 | ||
e9e1d143 RG |
7188 | Output: |
7189 | *COMP_VECTYPE - the vector type for the comparison. | |
7190 | ||
ebfd146a IR |
7191 | Returns whether a COND can be vectorized. Checks whether |
7192 | condition operands are supportable using vec_is_simple_use. */ | |
7193 | ||
87aab9b2 | 7194 | static bool |
81c40241 | 7195 | vect_is_simple_cond (tree cond, vec_info *vinfo, tree *comp_vectype) |
ebfd146a IR |
7196 | { |
7197 | tree lhs, rhs; | |
ebfd146a | 7198 | enum vect_def_type dt; |
e9e1d143 | 7199 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 7200 | |
a414c77f IE |
7201 | /* Mask case. */ |
7202 | if (TREE_CODE (cond) == SSA_NAME | |
7203 | && TREE_CODE (TREE_TYPE (cond)) == BOOLEAN_TYPE) | |
7204 | { | |
7205 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (cond); | |
7206 | if (!vect_is_simple_use (cond, vinfo, &lhs_def_stmt, | |
7207 | &dt, comp_vectype) | |
7208 | || !*comp_vectype | |
7209 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
7210 | return false; | |
7211 | return true; | |
7212 | } | |
7213 | ||
ebfd146a IR |
7214 | if (!COMPARISON_CLASS_P (cond)) |
7215 | return false; | |
7216 | ||
7217 | lhs = TREE_OPERAND (cond, 0); | |
7218 | rhs = TREE_OPERAND (cond, 1); | |
7219 | ||
7220 | if (TREE_CODE (lhs) == SSA_NAME) | |
7221 | { | |
355fe088 | 7222 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); |
81c40241 | 7223 | if (!vect_is_simple_use (lhs, vinfo, &lhs_def_stmt, &dt, &vectype1)) |
ebfd146a IR |
7224 | return false; |
7225 | } | |
7226 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
7227 | && TREE_CODE (lhs) != FIXED_CST) | |
7228 | return false; | |
7229 | ||
7230 | if (TREE_CODE (rhs) == SSA_NAME) | |
7231 | { | |
355fe088 | 7232 | gimple *rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); |
81c40241 | 7233 | if (!vect_is_simple_use (rhs, vinfo, &rhs_def_stmt, &dt, &vectype2)) |
ebfd146a IR |
7234 | return false; |
7235 | } | |
f7e531cf | 7236 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST |
ebfd146a IR |
7237 | && TREE_CODE (rhs) != FIXED_CST) |
7238 | return false; | |
7239 | ||
e9e1d143 | 7240 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
7241 | return true; |
7242 | } | |
7243 | ||
7244 | /* vectorizable_condition. | |
7245 | ||
b8698a0f L |
7246 | Check if STMT is conditional modify expression that can be vectorized. |
7247 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7248 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
7249 | at GSI. |
7250 | ||
7251 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
7252 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 7253 | else clause if it is 2). |
ebfd146a IR |
7254 | |
7255 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7256 | ||
4bbe8262 | 7257 | bool |
355fe088 TS |
7258 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
7259 | gimple **vec_stmt, tree reduc_def, int reduc_index, | |
f7e531cf | 7260 | slp_tree slp_node) |
ebfd146a IR |
7261 | { |
7262 | tree scalar_dest = NULL_TREE; | |
7263 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
7264 | tree cond_expr, then_clause, else_clause; |
7265 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
df11cc78 | 7266 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
7267 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
7268 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
ebfd146a IR |
7269 | tree vec_compare, vec_cond_expr; |
7270 | tree new_temp; | |
7271 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a855b1b1 | 7272 | enum vect_def_type dt, dts[4]; |
f7e531cf | 7273 | int ncopies; |
ebfd146a | 7274 | enum tree_code code; |
a855b1b1 | 7275 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
7276 | int i, j; |
7277 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
7278 | vec<tree> vec_oprnds0 = vNULL; |
7279 | vec<tree> vec_oprnds1 = vNULL; | |
7280 | vec<tree> vec_oprnds2 = vNULL; | |
7281 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 7282 | tree vec_cmp_type; |
a414c77f | 7283 | bool masked = false; |
b8698a0f | 7284 | |
f7e531cf IR |
7285 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
7286 | return false; | |
7287 | ||
af29617a AH |
7288 | if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == TREE_CODE_REDUCTION) |
7289 | { | |
7290 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
7291 | return false; | |
ebfd146a | 7292 | |
af29617a AH |
7293 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7294 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7295 | && reduc_def)) | |
7296 | return false; | |
ebfd146a | 7297 | |
af29617a AH |
7298 | /* FORNOW: not yet supported. */ |
7299 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7300 | { | |
7301 | if (dump_enabled_p ()) | |
7302 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7303 | "value used after loop.\n"); | |
7304 | return false; | |
7305 | } | |
ebfd146a IR |
7306 | } |
7307 | ||
7308 | /* Is vectorizable conditional operation? */ | |
7309 | if (!is_gimple_assign (stmt)) | |
7310 | return false; | |
7311 | ||
7312 | code = gimple_assign_rhs_code (stmt); | |
7313 | ||
7314 | if (code != COND_EXPR) | |
7315 | return false; | |
7316 | ||
465c8c19 JJ |
7317 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
7318 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
7319 | ||
7320 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
7321 | ncopies = 1; | |
7322 | else | |
7323 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
7324 | ||
7325 | gcc_assert (ncopies >= 1); | |
7326 | if (reduc_index && ncopies > 1) | |
7327 | return false; /* FORNOW */ | |
7328 | ||
4e71066d RG |
7329 | cond_expr = gimple_assign_rhs1 (stmt); |
7330 | then_clause = gimple_assign_rhs2 (stmt); | |
7331 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 7332 | |
81c40241 | 7333 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, &comp_vectype) |
e9e1d143 | 7334 | || !comp_vectype) |
ebfd146a IR |
7335 | return false; |
7336 | ||
81c40241 RB |
7337 | gimple *def_stmt; |
7338 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &def_stmt, &dt)) | |
ebfd146a | 7339 | return false; |
81c40241 | 7340 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &def_stmt, &dt)) |
ebfd146a IR |
7341 | return false; |
7342 | ||
a414c77f IE |
7343 | if (VECTOR_BOOLEAN_TYPE_P (comp_vectype)) |
7344 | { | |
7345 | vec_cmp_type = comp_vectype; | |
7346 | masked = true; | |
7347 | } | |
7348 | else | |
7349 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
74946978 MP |
7350 | if (vec_cmp_type == NULL_TREE) |
7351 | return false; | |
784fb9b3 | 7352 | |
b8698a0f | 7353 | if (!vec_stmt) |
ebfd146a IR |
7354 | { |
7355 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
e9e1d143 | 7356 | return expand_vec_cond_expr_p (vectype, comp_vectype); |
ebfd146a IR |
7357 | } |
7358 | ||
f7e531cf IR |
7359 | /* Transform. */ |
7360 | ||
7361 | if (!slp_node) | |
7362 | { | |
9771b263 DN |
7363 | vec_oprnds0.create (1); |
7364 | vec_oprnds1.create (1); | |
7365 | vec_oprnds2.create (1); | |
7366 | vec_oprnds3.create (1); | |
f7e531cf | 7367 | } |
ebfd146a IR |
7368 | |
7369 | /* Handle def. */ | |
7370 | scalar_dest = gimple_assign_lhs (stmt); | |
7371 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
7372 | ||
7373 | /* Handle cond expr. */ | |
a855b1b1 MM |
7374 | for (j = 0; j < ncopies; j++) |
7375 | { | |
538dd0b7 | 7376 | gassign *new_stmt = NULL; |
a855b1b1 MM |
7377 | if (j == 0) |
7378 | { | |
f7e531cf IR |
7379 | if (slp_node) |
7380 | { | |
00f96dc9 TS |
7381 | auto_vec<tree, 4> ops; |
7382 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 7383 | |
a414c77f IE |
7384 | if (masked) |
7385 | ops.safe_push (cond_expr); | |
7386 | else | |
7387 | { | |
7388 | ops.safe_push (TREE_OPERAND (cond_expr, 0)); | |
7389 | ops.safe_push (TREE_OPERAND (cond_expr, 1)); | |
7390 | } | |
9771b263 DN |
7391 | ops.safe_push (then_clause); |
7392 | ops.safe_push (else_clause); | |
f7e531cf | 7393 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); |
37b5ec8f JJ |
7394 | vec_oprnds3 = vec_defs.pop (); |
7395 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
7396 | if (!masked) |
7397 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 7398 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf | 7399 | |
9771b263 DN |
7400 | ops.release (); |
7401 | vec_defs.release (); | |
f7e531cf IR |
7402 | } |
7403 | else | |
7404 | { | |
355fe088 | 7405 | gimple *gtemp; |
a414c77f IE |
7406 | if (masked) |
7407 | { | |
7408 | vec_cond_lhs | |
7409 | = vect_get_vec_def_for_operand (cond_expr, stmt, | |
7410 | comp_vectype); | |
7411 | vect_is_simple_use (cond_expr, stmt_info->vinfo, | |
7412 | >emp, &dts[0]); | |
7413 | } | |
7414 | else | |
7415 | { | |
7416 | vec_cond_lhs = | |
7417 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), | |
7418 | stmt, comp_vectype); | |
7419 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), | |
7420 | loop_vinfo, >emp, &dts[0]); | |
7421 | ||
7422 | vec_cond_rhs = | |
7423 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
7424 | stmt, comp_vectype); | |
7425 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), | |
7426 | loop_vinfo, >emp, &dts[1]); | |
7427 | } | |
f7e531cf IR |
7428 | if (reduc_index == 1) |
7429 | vec_then_clause = reduc_def; | |
7430 | else | |
7431 | { | |
7432 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 RB |
7433 | stmt); |
7434 | vect_is_simple_use (then_clause, loop_vinfo, | |
7435 | >emp, &dts[2]); | |
f7e531cf IR |
7436 | } |
7437 | if (reduc_index == 2) | |
7438 | vec_else_clause = reduc_def; | |
7439 | else | |
7440 | { | |
7441 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 RB |
7442 | stmt); |
7443 | vect_is_simple_use (else_clause, loop_vinfo, >emp, &dts[3]); | |
f7e531cf | 7444 | } |
a855b1b1 MM |
7445 | } |
7446 | } | |
7447 | else | |
7448 | { | |
a414c77f IE |
7449 | vec_cond_lhs |
7450 | = vect_get_vec_def_for_stmt_copy (dts[0], | |
7451 | vec_oprnds0.pop ()); | |
7452 | if (!masked) | |
7453 | vec_cond_rhs | |
7454 | = vect_get_vec_def_for_stmt_copy (dts[1], | |
7455 | vec_oprnds1.pop ()); | |
7456 | ||
a855b1b1 | 7457 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 7458 | vec_oprnds2.pop ()); |
a855b1b1 | 7459 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 7460 | vec_oprnds3.pop ()); |
f7e531cf IR |
7461 | } |
7462 | ||
7463 | if (!slp_node) | |
7464 | { | |
9771b263 | 7465 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
7466 | if (!masked) |
7467 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
7468 | vec_oprnds2.quick_push (vec_then_clause); |
7469 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
7470 | } |
7471 | ||
9dc3f7de | 7472 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 7473 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 7474 | { |
9771b263 DN |
7475 | vec_then_clause = vec_oprnds2[i]; |
7476 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 7477 | |
a414c77f IE |
7478 | if (masked) |
7479 | vec_compare = vec_cond_lhs; | |
7480 | else | |
7481 | { | |
7482 | vec_cond_rhs = vec_oprnds1[i]; | |
7483 | vec_compare = build2 (TREE_CODE (cond_expr), vec_cmp_type, | |
7484 | vec_cond_lhs, vec_cond_rhs); | |
7485 | } | |
f7e531cf IR |
7486 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, |
7487 | vec_compare, vec_then_clause, vec_else_clause); | |
a855b1b1 | 7488 | |
f7e531cf IR |
7489 | new_stmt = gimple_build_assign (vec_dest, vec_cond_expr); |
7490 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
7491 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7492 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7493 | if (slp_node) | |
9771b263 | 7494 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
7495 | } |
7496 | ||
7497 | if (slp_node) | |
7498 | continue; | |
7499 | ||
7500 | if (j == 0) | |
7501 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7502 | else | |
7503 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7504 | ||
7505 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 7506 | } |
b8698a0f | 7507 | |
9771b263 DN |
7508 | vec_oprnds0.release (); |
7509 | vec_oprnds1.release (); | |
7510 | vec_oprnds2.release (); | |
7511 | vec_oprnds3.release (); | |
f7e531cf | 7512 | |
ebfd146a IR |
7513 | return true; |
7514 | } | |
7515 | ||
42fd8198 IE |
7516 | /* vectorizable_comparison. |
7517 | ||
7518 | Check if STMT is comparison expression that can be vectorized. | |
7519 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7520 | comparison, put it in VEC_STMT, and insert it at GSI. | |
7521 | ||
7522 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7523 | ||
7524 | bool | |
7525 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, | |
7526 | gimple **vec_stmt, tree reduc_def, | |
7527 | slp_tree slp_node) | |
7528 | { | |
7529 | tree lhs, rhs1, rhs2; | |
7530 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
7531 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
7532 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
7533 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
7534 | tree new_temp; | |
7535 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
7536 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
7537 | unsigned nunits; | |
7538 | int ncopies; | |
7539 | enum tree_code code; | |
7540 | stmt_vec_info prev_stmt_info = NULL; | |
7541 | int i, j; | |
7542 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
7543 | vec<tree> vec_oprnds0 = vNULL; | |
7544 | vec<tree> vec_oprnds1 = vNULL; | |
7545 | gimple *def_stmt; | |
7546 | tree mask_type; | |
7547 | tree mask; | |
7548 | ||
7549 | if (!VECTOR_BOOLEAN_TYPE_P (vectype)) | |
7550 | return false; | |
7551 | ||
7552 | mask_type = vectype; | |
7553 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
7554 | ||
7555 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
7556 | ncopies = 1; | |
7557 | else | |
7558 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
7559 | ||
7560 | gcc_assert (ncopies >= 1); | |
7561 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
7562 | return false; | |
7563 | ||
7564 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def | |
7565 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7566 | && reduc_def)) | |
7567 | return false; | |
7568 | ||
7569 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7570 | { | |
7571 | if (dump_enabled_p ()) | |
7572 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7573 | "value used after loop.\n"); | |
7574 | return false; | |
7575 | } | |
7576 | ||
7577 | if (!is_gimple_assign (stmt)) | |
7578 | return false; | |
7579 | ||
7580 | code = gimple_assign_rhs_code (stmt); | |
7581 | ||
7582 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
7583 | return false; | |
7584 | ||
7585 | rhs1 = gimple_assign_rhs1 (stmt); | |
7586 | rhs2 = gimple_assign_rhs2 (stmt); | |
7587 | ||
7588 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &def_stmt, | |
7589 | &dts[0], &vectype1)) | |
7590 | return false; | |
7591 | ||
7592 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &def_stmt, | |
7593 | &dts[1], &vectype2)) | |
7594 | return false; | |
7595 | ||
7596 | if (vectype1 && vectype2 | |
7597 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
7598 | return false; | |
7599 | ||
7600 | vectype = vectype1 ? vectype1 : vectype2; | |
7601 | ||
7602 | /* Invariant comparison. */ | |
7603 | if (!vectype) | |
7604 | { | |
7605 | vectype = build_vector_type (TREE_TYPE (rhs1), nunits); | |
7606 | if (tree_to_shwi (TYPE_SIZE_UNIT (vectype)) != current_vector_size) | |
7607 | return false; | |
7608 | } | |
7609 | else if (nunits != TYPE_VECTOR_SUBPARTS (vectype)) | |
7610 | return false; | |
7611 | ||
7612 | if (!vec_stmt) | |
7613 | { | |
7614 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
7615 | vect_model_simple_cost (stmt_info, ncopies, dts, NULL, NULL); | |
7616 | return expand_vec_cmp_expr_p (vectype, mask_type); | |
7617 | } | |
7618 | ||
7619 | /* Transform. */ | |
7620 | if (!slp_node) | |
7621 | { | |
7622 | vec_oprnds0.create (1); | |
7623 | vec_oprnds1.create (1); | |
7624 | } | |
7625 | ||
7626 | /* Handle def. */ | |
7627 | lhs = gimple_assign_lhs (stmt); | |
7628 | mask = vect_create_destination_var (lhs, mask_type); | |
7629 | ||
7630 | /* Handle cmp expr. */ | |
7631 | for (j = 0; j < ncopies; j++) | |
7632 | { | |
7633 | gassign *new_stmt = NULL; | |
7634 | if (j == 0) | |
7635 | { | |
7636 | if (slp_node) | |
7637 | { | |
7638 | auto_vec<tree, 2> ops; | |
7639 | auto_vec<vec<tree>, 2> vec_defs; | |
7640 | ||
7641 | ops.safe_push (rhs1); | |
7642 | ops.safe_push (rhs2); | |
7643 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); | |
7644 | vec_oprnds1 = vec_defs.pop (); | |
7645 | vec_oprnds0 = vec_defs.pop (); | |
7646 | } | |
7647 | else | |
7648 | { | |
e4af0bc4 IE |
7649 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, vectype); |
7650 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, vectype); | |
42fd8198 IE |
7651 | } |
7652 | } | |
7653 | else | |
7654 | { | |
7655 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
7656 | vec_oprnds0.pop ()); | |
7657 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
7658 | vec_oprnds1.pop ()); | |
7659 | } | |
7660 | ||
7661 | if (!slp_node) | |
7662 | { | |
7663 | vec_oprnds0.quick_push (vec_rhs1); | |
7664 | vec_oprnds1.quick_push (vec_rhs2); | |
7665 | } | |
7666 | ||
7667 | /* Arguments are ready. Create the new vector stmt. */ | |
7668 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
7669 | { | |
7670 | vec_rhs2 = vec_oprnds1[i]; | |
7671 | ||
7672 | new_temp = make_ssa_name (mask); | |
7673 | new_stmt = gimple_build_assign (new_temp, code, vec_rhs1, vec_rhs2); | |
7674 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7675 | if (slp_node) | |
7676 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
7677 | } | |
7678 | ||
7679 | if (slp_node) | |
7680 | continue; | |
7681 | ||
7682 | if (j == 0) | |
7683 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7684 | else | |
7685 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7686 | ||
7687 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7688 | } | |
7689 | ||
7690 | vec_oprnds0.release (); | |
7691 | vec_oprnds1.release (); | |
7692 | ||
7693 | return true; | |
7694 | } | |
ebfd146a | 7695 | |
8644a673 | 7696 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
7697 | |
7698 | bool | |
355fe088 | 7699 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 7700 | { |
8644a673 | 7701 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 7702 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 7703 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 7704 | bool ok; |
a70d6342 | 7705 | tree scalar_type, vectype; |
355fe088 | 7706 | gimple *pattern_stmt; |
363477c0 | 7707 | gimple_seq pattern_def_seq; |
ebfd146a | 7708 | |
73fbfcad | 7709 | if (dump_enabled_p ()) |
ebfd146a | 7710 | { |
78c60e3d SS |
7711 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
7712 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 7713 | } |
ebfd146a | 7714 | |
1825a1f3 | 7715 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 7716 | { |
73fbfcad | 7717 | if (dump_enabled_p ()) |
78c60e3d | 7718 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7719 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
7720 | |
7721 | return false; | |
7722 | } | |
b8698a0f L |
7723 | |
7724 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
7725 | to include: |
7726 | - the COND_EXPR which is the loop exit condition | |
7727 | - any LABEL_EXPRs in the loop | |
b8698a0f | 7728 | - computations that are used only for array indexing or loop control. |
8644a673 | 7729 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 7730 | instance, therefore, all the statements are relevant. |
ebfd146a | 7731 | |
d092494c | 7732 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 7733 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
7734 | statements. In basic blocks we are called from some SLP instance |
7735 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
7736 | already will be part of SLP instance. */ | |
83197f37 IR |
7737 | |
7738 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 7739 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 7740 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 7741 | { |
9d5e7640 | 7742 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 7743 | && pattern_stmt |
9d5e7640 IR |
7744 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
7745 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
7746 | { | |
83197f37 | 7747 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
7748 | stmt = pattern_stmt; |
7749 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 7750 | if (dump_enabled_p ()) |
9d5e7640 | 7751 | { |
78c60e3d SS |
7752 | dump_printf_loc (MSG_NOTE, vect_location, |
7753 | "==> examining pattern statement: "); | |
7754 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
7755 | } |
7756 | } | |
7757 | else | |
7758 | { | |
73fbfcad | 7759 | if (dump_enabled_p ()) |
e645e942 | 7760 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 7761 | |
9d5e7640 IR |
7762 | return true; |
7763 | } | |
8644a673 | 7764 | } |
83197f37 | 7765 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 7766 | && node == NULL |
83197f37 IR |
7767 | && pattern_stmt |
7768 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
7769 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
7770 | { | |
7771 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 7772 | if (dump_enabled_p ()) |
83197f37 | 7773 | { |
78c60e3d SS |
7774 | dump_printf_loc (MSG_NOTE, vect_location, |
7775 | "==> examining pattern statement: "); | |
7776 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
7777 | } |
7778 | ||
7779 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
7780 | return false; | |
7781 | } | |
ebfd146a | 7782 | |
1107f3ae | 7783 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 7784 | && node == NULL |
363477c0 | 7785 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 7786 | { |
363477c0 | 7787 | gimple_stmt_iterator si; |
1107f3ae | 7788 | |
363477c0 JJ |
7789 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
7790 | { | |
355fe088 | 7791 | gimple *pattern_def_stmt = gsi_stmt (si); |
363477c0 JJ |
7792 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) |
7793 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
7794 | { | |
7795 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 7796 | if (dump_enabled_p ()) |
363477c0 | 7797 | { |
78c60e3d SS |
7798 | dump_printf_loc (MSG_NOTE, vect_location, |
7799 | "==> examining pattern def statement: "); | |
7800 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 7801 | } |
1107f3ae | 7802 | |
363477c0 JJ |
7803 | if (!vect_analyze_stmt (pattern_def_stmt, |
7804 | need_to_vectorize, node)) | |
7805 | return false; | |
7806 | } | |
7807 | } | |
7808 | } | |
1107f3ae | 7809 | |
8644a673 IR |
7810 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
7811 | { | |
7812 | case vect_internal_def: | |
7813 | break; | |
ebfd146a | 7814 | |
8644a673 | 7815 | case vect_reduction_def: |
7c5222ff | 7816 | case vect_nested_cycle: |
14a61437 RB |
7817 | gcc_assert (!bb_vinfo |
7818 | && (relevance == vect_used_in_outer | |
7819 | || relevance == vect_used_in_outer_by_reduction | |
7820 | || relevance == vect_used_by_reduction | |
7821 | || relevance == vect_unused_in_scope)); | |
8644a673 IR |
7822 | break; |
7823 | ||
7824 | case vect_induction_def: | |
7825 | case vect_constant_def: | |
7826 | case vect_external_def: | |
7827 | case vect_unknown_def_type: | |
7828 | default: | |
7829 | gcc_unreachable (); | |
7830 | } | |
ebfd146a | 7831 | |
a70d6342 IR |
7832 | if (bb_vinfo) |
7833 | { | |
7834 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
7835 | ||
b690cc0f | 7836 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
73fbfcad | 7837 | if (dump_enabled_p ()) |
a70d6342 | 7838 | { |
78c60e3d SS |
7839 | dump_printf_loc (MSG_NOTE, vect_location, |
7840 | "get vectype for scalar type: "); | |
7841 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
e645e942 | 7842 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
7843 | } |
7844 | ||
7845 | vectype = get_vectype_for_scalar_type (scalar_type); | |
7846 | if (!vectype) | |
7847 | { | |
73fbfcad | 7848 | if (dump_enabled_p ()) |
a70d6342 | 7849 | { |
78c60e3d SS |
7850 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7851 | "not SLPed: unsupported data-type "); | |
7852 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
7853 | scalar_type); | |
e645e942 | 7854 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
a70d6342 IR |
7855 | } |
7856 | return false; | |
7857 | } | |
7858 | ||
73fbfcad | 7859 | if (dump_enabled_p ()) |
a70d6342 | 7860 | { |
78c60e3d SS |
7861 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); |
7862 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
e645e942 | 7863 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
7864 | } |
7865 | ||
7866 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
7867 | } | |
7868 | ||
8644a673 | 7869 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 7870 | { |
8644a673 | 7871 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
7872 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
7873 | || (is_gimple_call (stmt) | |
7874 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 7875 | *need_to_vectorize = true; |
ebfd146a IR |
7876 | } |
7877 | ||
b1af7da6 RB |
7878 | if (PURE_SLP_STMT (stmt_info) && !node) |
7879 | { | |
7880 | dump_printf_loc (MSG_NOTE, vect_location, | |
7881 | "handled only by SLP analysis\n"); | |
7882 | return true; | |
7883 | } | |
7884 | ||
7885 | ok = true; | |
7886 | if (!bb_vinfo | |
7887 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
7888 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
7889 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
7890 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
7891 | || vectorizable_shift (stmt, NULL, NULL, node) | |
7892 | || vectorizable_operation (stmt, NULL, NULL, node) | |
7893 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
7894 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
7895 | || vectorizable_call (stmt, NULL, NULL, node) | |
7896 | || vectorizable_store (stmt, NULL, NULL, node) | |
7897 | || vectorizable_reduction (stmt, NULL, NULL, node) | |
42fd8198 IE |
7898 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
7899 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 RB |
7900 | else |
7901 | { | |
7902 | if (bb_vinfo) | |
7903 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
7904 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
7905 | || vectorizable_shift (stmt, NULL, NULL, node) | |
7906 | || vectorizable_operation (stmt, NULL, NULL, node) | |
7907 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
7908 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
7909 | || vectorizable_call (stmt, NULL, NULL, node) | |
7910 | || vectorizable_store (stmt, NULL, NULL, node) | |
42fd8198 IE |
7911 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
7912 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 | 7913 | } |
8644a673 IR |
7914 | |
7915 | if (!ok) | |
ebfd146a | 7916 | { |
73fbfcad | 7917 | if (dump_enabled_p ()) |
8644a673 | 7918 | { |
78c60e3d SS |
7919 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7920 | "not vectorized: relevant stmt not "); | |
7921 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
7922 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 7923 | } |
b8698a0f | 7924 | |
ebfd146a IR |
7925 | return false; |
7926 | } | |
7927 | ||
a70d6342 IR |
7928 | if (bb_vinfo) |
7929 | return true; | |
7930 | ||
8644a673 IR |
7931 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
7932 | need extra handling, except for vectorizable reductions. */ | |
7933 | if (STMT_VINFO_LIVE_P (stmt_info) | |
7934 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
7935 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 7936 | |
8644a673 | 7937 | if (!ok) |
ebfd146a | 7938 | { |
73fbfcad | 7939 | if (dump_enabled_p ()) |
8644a673 | 7940 | { |
78c60e3d SS |
7941 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7942 | "not vectorized: live stmt not "); | |
7943 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
7944 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 7945 | } |
b8698a0f | 7946 | |
8644a673 | 7947 | return false; |
ebfd146a IR |
7948 | } |
7949 | ||
ebfd146a IR |
7950 | return true; |
7951 | } | |
7952 | ||
7953 | ||
7954 | /* Function vect_transform_stmt. | |
7955 | ||
7956 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
7957 | ||
7958 | bool | |
355fe088 | 7959 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 7960 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
7961 | slp_instance slp_node_instance) |
7962 | { | |
7963 | bool is_store = false; | |
355fe088 | 7964 | gimple *vec_stmt = NULL; |
ebfd146a | 7965 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
ebfd146a | 7966 | bool done; |
ebfd146a | 7967 | |
355fe088 | 7968 | gimple *old_vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 7969 | |
ebfd146a IR |
7970 | switch (STMT_VINFO_TYPE (stmt_info)) |
7971 | { | |
7972 | case type_demotion_vec_info_type: | |
ebfd146a | 7973 | case type_promotion_vec_info_type: |
ebfd146a IR |
7974 | case type_conversion_vec_info_type: |
7975 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
7976 | gcc_assert (done); | |
7977 | break; | |
7978 | ||
7979 | case induc_vec_info_type: | |
7980 | gcc_assert (!slp_node); | |
7981 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
7982 | gcc_assert (done); | |
7983 | break; | |
7984 | ||
9dc3f7de IR |
7985 | case shift_vec_info_type: |
7986 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
7987 | gcc_assert (done); | |
7988 | break; | |
7989 | ||
ebfd146a IR |
7990 | case op_vec_info_type: |
7991 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
7992 | gcc_assert (done); | |
7993 | break; | |
7994 | ||
7995 | case assignment_vec_info_type: | |
7996 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
7997 | gcc_assert (done); | |
7998 | break; | |
7999 | ||
8000 | case load_vec_info_type: | |
b8698a0f | 8001 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
8002 | slp_node_instance); |
8003 | gcc_assert (done); | |
8004 | break; | |
8005 | ||
8006 | case store_vec_info_type: | |
8007 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
8008 | gcc_assert (done); | |
0d0293ac | 8009 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
8010 | { |
8011 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 8012 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
8013 | one are skipped, and there vec_stmt_info shouldn't be freed |
8014 | meanwhile. */ | |
0d0293ac | 8015 | *grouped_store = true; |
ebfd146a IR |
8016 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
8017 | is_store = true; | |
8018 | } | |
8019 | else | |
8020 | is_store = true; | |
8021 | break; | |
8022 | ||
8023 | case condition_vec_info_type: | |
f7e531cf | 8024 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
8025 | gcc_assert (done); |
8026 | break; | |
8027 | ||
42fd8198 IE |
8028 | case comparison_vec_info_type: |
8029 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node); | |
8030 | gcc_assert (done); | |
8031 | break; | |
8032 | ||
ebfd146a | 8033 | case call_vec_info_type: |
190c2236 | 8034 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 8035 | stmt = gsi_stmt (*gsi); |
5ce9450f JJ |
8036 | if (is_gimple_call (stmt) |
8037 | && gimple_call_internal_p (stmt) | |
8038 | && gimple_call_internal_fn (stmt) == IFN_MASK_STORE) | |
8039 | is_store = true; | |
ebfd146a IR |
8040 | break; |
8041 | ||
0136f8f0 AH |
8042 | case call_simd_clone_vec_info_type: |
8043 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node); | |
8044 | stmt = gsi_stmt (*gsi); | |
8045 | break; | |
8046 | ||
ebfd146a | 8047 | case reduc_vec_info_type: |
b5aeb3bb | 8048 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
8049 | gcc_assert (done); |
8050 | break; | |
8051 | ||
8052 | default: | |
8053 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
8054 | { | |
73fbfcad | 8055 | if (dump_enabled_p ()) |
78c60e3d | 8056 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8057 | "stmt not supported.\n"); |
ebfd146a IR |
8058 | gcc_unreachable (); |
8059 | } | |
8060 | } | |
8061 | ||
225ce44b RB |
8062 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
8063 | This would break hybrid SLP vectorization. */ | |
8064 | if (slp_node) | |
d90f8440 RB |
8065 | gcc_assert (!vec_stmt |
8066 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt); | |
225ce44b | 8067 | |
ebfd146a IR |
8068 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
8069 | is being vectorized, but outside the immediately enclosing loop. */ | |
8070 | if (vec_stmt | |
a70d6342 IR |
8071 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
8072 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
8073 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
8074 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
8075 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 8076 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 8077 | vect_used_in_outer_by_reduction)) |
ebfd146a | 8078 | { |
a70d6342 IR |
8079 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
8080 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
8081 | imm_use_iterator imm_iter; |
8082 | use_operand_p use_p; | |
8083 | tree scalar_dest; | |
355fe088 | 8084 | gimple *exit_phi; |
ebfd146a | 8085 | |
73fbfcad | 8086 | if (dump_enabled_p ()) |
78c60e3d | 8087 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 8088 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
8089 | |
8090 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
8091 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
8092 | STMT). */ | |
8093 | if (gimple_code (stmt) == GIMPLE_PHI) | |
8094 | scalar_dest = PHI_RESULT (stmt); | |
8095 | else | |
8096 | scalar_dest = gimple_assign_lhs (stmt); | |
8097 | ||
8098 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
8099 | { | |
8100 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
8101 | { | |
8102 | exit_phi = USE_STMT (use_p); | |
8103 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
8104 | } | |
8105 | } | |
8106 | } | |
8107 | ||
8108 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
8109 | being vectorized. */ | |
8110 | if (STMT_VINFO_LIVE_P (stmt_info) | |
8111 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
8112 | { | |
8113 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
8114 | gcc_assert (done); | |
8115 | } | |
8116 | ||
8117 | if (vec_stmt) | |
83197f37 | 8118 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 8119 | |
b8698a0f | 8120 | return is_store; |
ebfd146a IR |
8121 | } |
8122 | ||
8123 | ||
b8698a0f | 8124 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
8125 | stmt_vec_info. */ |
8126 | ||
8127 | void | |
355fe088 | 8128 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 8129 | { |
355fe088 TS |
8130 | gimple *next = first_stmt; |
8131 | gimple *tmp; | |
ebfd146a IR |
8132 | gimple_stmt_iterator next_si; |
8133 | ||
8134 | while (next) | |
8135 | { | |
78048b1c JJ |
8136 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
8137 | ||
8138 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
8139 | if (is_pattern_stmt_p (stmt_info)) | |
8140 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
8141 | /* Free the attached stmt_vec_info and remove the stmt. */ |
8142 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 8143 | unlink_stmt_vdef (next); |
ebfd146a | 8144 | gsi_remove (&next_si, true); |
3d3f2249 | 8145 | release_defs (next); |
ebfd146a IR |
8146 | free_stmt_vec_info (next); |
8147 | next = tmp; | |
8148 | } | |
8149 | } | |
8150 | ||
8151 | ||
8152 | /* Function new_stmt_vec_info. | |
8153 | ||
8154 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
8155 | ||
8156 | stmt_vec_info | |
310213d4 | 8157 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
8158 | { |
8159 | stmt_vec_info res; | |
8160 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
8161 | ||
8162 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
8163 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 8164 | res->vinfo = vinfo; |
8644a673 | 8165 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
8166 | STMT_VINFO_LIVE_P (res) = false; |
8167 | STMT_VINFO_VECTYPE (res) = NULL; | |
8168 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 8169 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
8170 | STMT_VINFO_IN_PATTERN_P (res) = false; |
8171 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 8172 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 8173 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 8174 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
ebfd146a IR |
8175 | |
8176 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
8177 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
8178 | STMT_VINFO_DR_INIT (res) = NULL; | |
8179 | STMT_VINFO_DR_STEP (res) = NULL; | |
8180 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
8181 | ||
8182 | if (gimple_code (stmt) == GIMPLE_PHI | |
8183 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
8184 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
8185 | else | |
8644a673 IR |
8186 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
8187 | ||
9771b263 | 8188 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 8189 | STMT_SLP_TYPE (res) = loop_vect; |
e14c1050 IR |
8190 | GROUP_FIRST_ELEMENT (res) = NULL; |
8191 | GROUP_NEXT_ELEMENT (res) = NULL; | |
8192 | GROUP_SIZE (res) = 0; | |
8193 | GROUP_STORE_COUNT (res) = 0; | |
8194 | GROUP_GAP (res) = 0; | |
8195 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
8196 | |
8197 | return res; | |
8198 | } | |
8199 | ||
8200 | ||
8201 | /* Create a hash table for stmt_vec_info. */ | |
8202 | ||
8203 | void | |
8204 | init_stmt_vec_info_vec (void) | |
8205 | { | |
9771b263 DN |
8206 | gcc_assert (!stmt_vec_info_vec.exists ()); |
8207 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
8208 | } |
8209 | ||
8210 | ||
8211 | /* Free hash table for stmt_vec_info. */ | |
8212 | ||
8213 | void | |
8214 | free_stmt_vec_info_vec (void) | |
8215 | { | |
93675444 | 8216 | unsigned int i; |
3161455c | 8217 | stmt_vec_info info; |
93675444 JJ |
8218 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) |
8219 | if (info != NULL) | |
3161455c | 8220 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
9771b263 DN |
8221 | gcc_assert (stmt_vec_info_vec.exists ()); |
8222 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
8223 | } |
8224 | ||
8225 | ||
8226 | /* Free stmt vectorization related info. */ | |
8227 | ||
8228 | void | |
355fe088 | 8229 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
8230 | { |
8231 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8232 | ||
8233 | if (!stmt_info) | |
8234 | return; | |
8235 | ||
78048b1c JJ |
8236 | /* Check if this statement has a related "pattern stmt" |
8237 | (introduced by the vectorizer during the pattern recognition | |
8238 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
8239 | too. */ | |
8240 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
8241 | { | |
8242 | stmt_vec_info patt_info | |
8243 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
8244 | if (patt_info) | |
8245 | { | |
363477c0 | 8246 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
355fe088 | 8247 | gimple *patt_stmt = STMT_VINFO_STMT (patt_info); |
f0281fde RB |
8248 | gimple_set_bb (patt_stmt, NULL); |
8249 | tree lhs = gimple_get_lhs (patt_stmt); | |
e6f5c25d | 8250 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde | 8251 | release_ssa_name (lhs); |
363477c0 JJ |
8252 | if (seq) |
8253 | { | |
8254 | gimple_stmt_iterator si; | |
8255 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
f0281fde | 8256 | { |
355fe088 | 8257 | gimple *seq_stmt = gsi_stmt (si); |
f0281fde | 8258 | gimple_set_bb (seq_stmt, NULL); |
7532abf2 | 8259 | lhs = gimple_get_lhs (seq_stmt); |
e6f5c25d | 8260 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde RB |
8261 | release_ssa_name (lhs); |
8262 | free_stmt_vec_info (seq_stmt); | |
8263 | } | |
363477c0 | 8264 | } |
f0281fde | 8265 | free_stmt_vec_info (patt_stmt); |
78048b1c JJ |
8266 | } |
8267 | } | |
8268 | ||
9771b263 | 8269 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 8270 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
8271 | set_vinfo_for_stmt (stmt, NULL); |
8272 | free (stmt_info); | |
8273 | } | |
8274 | ||
8275 | ||
bb67d9c7 | 8276 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 8277 | |
bb67d9c7 | 8278 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
8279 | by the target. */ |
8280 | ||
bb67d9c7 RG |
8281 | static tree |
8282 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a | 8283 | { |
ef4bddc2 RS |
8284 | machine_mode inner_mode = TYPE_MODE (scalar_type); |
8285 | machine_mode simd_mode; | |
2f816591 | 8286 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
8287 | int nunits; |
8288 | tree vectype; | |
8289 | ||
cc4b5170 | 8290 | if (nbytes == 0) |
ebfd146a IR |
8291 | return NULL_TREE; |
8292 | ||
48f2e373 RB |
8293 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
8294 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
8295 | return NULL_TREE; | |
8296 | ||
7b7b1813 RG |
8297 | /* For vector types of elements whose mode precision doesn't |
8298 | match their types precision we use a element type of mode | |
8299 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
8300 | they support the proper result truncation/extension. |
8301 | We also make sure to build vector types with INTEGER_TYPE | |
8302 | component type only. */ | |
6d7971b8 | 8303 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
8304 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
8305 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
8306 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
8307 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 8308 | |
ccbf5bb4 RG |
8309 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
8310 | When the component mode passes the above test simply use a type | |
8311 | corresponding to that mode. The theory is that any use that | |
8312 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 8313 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 8314 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
8315 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
8316 | ||
8317 | /* We can't build a vector type of elements with alignment bigger than | |
8318 | their size. */ | |
dfc2e2ac | 8319 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
8320 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
8321 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 8322 | |
dfc2e2ac RB |
8323 | /* If we felt back to using the mode fail if there was |
8324 | no scalar type for it. */ | |
8325 | if (scalar_type == NULL_TREE) | |
8326 | return NULL_TREE; | |
8327 | ||
bb67d9c7 RG |
8328 | /* If no size was supplied use the mode the target prefers. Otherwise |
8329 | lookup a vector mode of the specified size. */ | |
8330 | if (size == 0) | |
8331 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
8332 | else | |
8333 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
8334 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
8335 | if (nunits <= 1) | |
8336 | return NULL_TREE; | |
ebfd146a IR |
8337 | |
8338 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
8339 | |
8340 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
8341 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 8342 | return NULL_TREE; |
ebfd146a IR |
8343 | |
8344 | return vectype; | |
8345 | } | |
8346 | ||
bb67d9c7 RG |
8347 | unsigned int current_vector_size; |
8348 | ||
8349 | /* Function get_vectype_for_scalar_type. | |
8350 | ||
8351 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
8352 | by the target. */ | |
8353 | ||
8354 | tree | |
8355 | get_vectype_for_scalar_type (tree scalar_type) | |
8356 | { | |
8357 | tree vectype; | |
8358 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
8359 | current_vector_size); | |
8360 | if (vectype | |
8361 | && current_vector_size == 0) | |
8362 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
8363 | return vectype; | |
8364 | } | |
8365 | ||
42fd8198 IE |
8366 | /* Function get_mask_type_for_scalar_type. |
8367 | ||
8368 | Returns the mask type corresponding to a result of comparison | |
8369 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
8370 | ||
8371 | tree | |
8372 | get_mask_type_for_scalar_type (tree scalar_type) | |
8373 | { | |
8374 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
8375 | ||
8376 | if (!vectype) | |
8377 | return NULL; | |
8378 | ||
8379 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
8380 | current_vector_size); | |
8381 | } | |
8382 | ||
b690cc0f RG |
8383 | /* Function get_same_sized_vectype |
8384 | ||
8385 | Returns a vector type corresponding to SCALAR_TYPE of size | |
8386 | VECTOR_TYPE if supported by the target. */ | |
8387 | ||
8388 | tree | |
bb67d9c7 | 8389 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 8390 | { |
9f47c7e5 IE |
8391 | if (TREE_CODE (scalar_type) == BOOLEAN_TYPE) |
8392 | return build_same_sized_truth_vector_type (vector_type); | |
8393 | ||
bb67d9c7 RG |
8394 | return get_vectype_for_scalar_type_and_size |
8395 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
8396 | } |
8397 | ||
ebfd146a IR |
8398 | /* Function vect_is_simple_use. |
8399 | ||
8400 | Input: | |
81c40241 RB |
8401 | VINFO - the vect info of the loop or basic block that is being vectorized. |
8402 | OPERAND - operand in the loop or bb. | |
8403 | Output: | |
8404 | DEF_STMT - the defining stmt in case OPERAND is an SSA_NAME. | |
8405 | DT - the type of definition | |
ebfd146a IR |
8406 | |
8407 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 8408 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 8409 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 8410 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
8411 | is the case in reduction/induction computations). |
8412 | For basic blocks, supportable operands are constants and bb invariants. | |
8413 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
8414 | |
8415 | bool | |
81c40241 RB |
8416 | vect_is_simple_use (tree operand, vec_info *vinfo, |
8417 | gimple **def_stmt, enum vect_def_type *dt) | |
b8698a0f | 8418 | { |
ebfd146a | 8419 | *def_stmt = NULL; |
3fc356dc | 8420 | *dt = vect_unknown_def_type; |
b8698a0f | 8421 | |
73fbfcad | 8422 | if (dump_enabled_p ()) |
ebfd146a | 8423 | { |
78c60e3d SS |
8424 | dump_printf_loc (MSG_NOTE, vect_location, |
8425 | "vect_is_simple_use: operand "); | |
8426 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 8427 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 8428 | } |
b8698a0f | 8429 | |
b758f602 | 8430 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
8431 | { |
8432 | *dt = vect_constant_def; | |
8433 | return true; | |
8434 | } | |
b8698a0f | 8435 | |
ebfd146a IR |
8436 | if (is_gimple_min_invariant (operand)) |
8437 | { | |
8644a673 | 8438 | *dt = vect_external_def; |
ebfd146a IR |
8439 | return true; |
8440 | } | |
8441 | ||
ebfd146a IR |
8442 | if (TREE_CODE (operand) != SSA_NAME) |
8443 | { | |
73fbfcad | 8444 | if (dump_enabled_p ()) |
af29617a AH |
8445 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8446 | "not ssa-name.\n"); | |
ebfd146a IR |
8447 | return false; |
8448 | } | |
b8698a0f | 8449 | |
3fc356dc | 8450 | if (SSA_NAME_IS_DEFAULT_DEF (operand)) |
ebfd146a | 8451 | { |
3fc356dc RB |
8452 | *dt = vect_external_def; |
8453 | return true; | |
ebfd146a IR |
8454 | } |
8455 | ||
3fc356dc | 8456 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
73fbfcad | 8457 | if (dump_enabled_p ()) |
ebfd146a | 8458 | { |
78c60e3d SS |
8459 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
8460 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
8461 | } |
8462 | ||
61d371eb | 8463 | if (! vect_stmt_in_region_p (vinfo, *def_stmt)) |
8644a673 | 8464 | *dt = vect_external_def; |
ebfd146a IR |
8465 | else |
8466 | { | |
3fc356dc | 8467 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (*def_stmt); |
310213d4 | 8468 | if (is_a <bb_vec_info> (vinfo) && !STMT_VINFO_VECTORIZABLE (stmt_vinfo)) |
90dd6e3d RB |
8469 | *dt = vect_external_def; |
8470 | else | |
8471 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
ebfd146a IR |
8472 | } |
8473 | ||
2e8ab70c RB |
8474 | if (dump_enabled_p ()) |
8475 | { | |
8476 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: "); | |
8477 | switch (*dt) | |
8478 | { | |
8479 | case vect_uninitialized_def: | |
8480 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
8481 | break; | |
8482 | case vect_constant_def: | |
8483 | dump_printf (MSG_NOTE, "constant\n"); | |
8484 | break; | |
8485 | case vect_external_def: | |
8486 | dump_printf (MSG_NOTE, "external\n"); | |
8487 | break; | |
8488 | case vect_internal_def: | |
8489 | dump_printf (MSG_NOTE, "internal\n"); | |
8490 | break; | |
8491 | case vect_induction_def: | |
8492 | dump_printf (MSG_NOTE, "induction\n"); | |
8493 | break; | |
8494 | case vect_reduction_def: | |
8495 | dump_printf (MSG_NOTE, "reduction\n"); | |
8496 | break; | |
8497 | case vect_double_reduction_def: | |
8498 | dump_printf (MSG_NOTE, "double reduction\n"); | |
8499 | break; | |
8500 | case vect_nested_cycle: | |
8501 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
8502 | break; | |
8503 | case vect_unknown_def_type: | |
8504 | dump_printf (MSG_NOTE, "unknown\n"); | |
8505 | break; | |
8506 | } | |
8507 | } | |
8508 | ||
81c40241 | 8509 | if (*dt == vect_unknown_def_type) |
ebfd146a | 8510 | { |
73fbfcad | 8511 | if (dump_enabled_p ()) |
78c60e3d | 8512 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8513 | "Unsupported pattern.\n"); |
ebfd146a IR |
8514 | return false; |
8515 | } | |
8516 | ||
ebfd146a IR |
8517 | switch (gimple_code (*def_stmt)) |
8518 | { | |
8519 | case GIMPLE_PHI: | |
ebfd146a | 8520 | case GIMPLE_ASSIGN: |
ebfd146a | 8521 | case GIMPLE_CALL: |
81c40241 | 8522 | break; |
ebfd146a | 8523 | default: |
73fbfcad | 8524 | if (dump_enabled_p ()) |
78c60e3d | 8525 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8526 | "unsupported defining stmt:\n"); |
ebfd146a IR |
8527 | return false; |
8528 | } | |
8529 | ||
8530 | return true; | |
8531 | } | |
8532 | ||
81c40241 | 8533 | /* Function vect_is_simple_use. |
b690cc0f | 8534 | |
81c40241 | 8535 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
8536 | type of OPERAND and stores it to *VECTYPE. If the definition of |
8537 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
8538 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
8539 | is responsible to compute the best suited vector type for the | |
8540 | scalar operand. */ | |
8541 | ||
8542 | bool | |
81c40241 RB |
8543 | vect_is_simple_use (tree operand, vec_info *vinfo, |
8544 | gimple **def_stmt, enum vect_def_type *dt, tree *vectype) | |
b690cc0f | 8545 | { |
81c40241 | 8546 | if (!vect_is_simple_use (operand, vinfo, def_stmt, dt)) |
b690cc0f RG |
8547 | return false; |
8548 | ||
8549 | /* Now get a vector type if the def is internal, otherwise supply | |
8550 | NULL_TREE and leave it up to the caller to figure out a proper | |
8551 | type for the use stmt. */ | |
8552 | if (*dt == vect_internal_def | |
8553 | || *dt == vect_induction_def | |
8554 | || *dt == vect_reduction_def | |
8555 | || *dt == vect_double_reduction_def | |
8556 | || *dt == vect_nested_cycle) | |
8557 | { | |
8558 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
8559 | |
8560 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
8561 | && !STMT_VINFO_RELEVANT (stmt_info) | |
8562 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 8563 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 8564 | |
b690cc0f RG |
8565 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
8566 | gcc_assert (*vectype != NULL_TREE); | |
8567 | } | |
8568 | else if (*dt == vect_uninitialized_def | |
8569 | || *dt == vect_constant_def | |
8570 | || *dt == vect_external_def) | |
8571 | *vectype = NULL_TREE; | |
8572 | else | |
8573 | gcc_unreachable (); | |
8574 | ||
8575 | return true; | |
8576 | } | |
8577 | ||
ebfd146a IR |
8578 | |
8579 | /* Function supportable_widening_operation | |
8580 | ||
b8698a0f L |
8581 | Check whether an operation represented by the code CODE is a |
8582 | widening operation that is supported by the target platform in | |
b690cc0f RG |
8583 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
8584 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 8585 | |
ebfd146a IR |
8586 | Widening operations we currently support are NOP (CONVERT), FLOAT |
8587 | and WIDEN_MULT. This function checks if these operations are supported | |
8588 | by the target platform either directly (via vector tree-codes), or via | |
8589 | target builtins. | |
8590 | ||
8591 | Output: | |
b8698a0f L |
8592 | - CODE1 and CODE2 are codes of vector operations to be used when |
8593 | vectorizing the operation, if available. | |
ebfd146a IR |
8594 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
8595 | case of multi-step conversion (like char->short->int - in that case | |
8596 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
8597 | - INTERM_TYPES contains the intermediate type required to perform the |
8598 | widening operation (short in the above example). */ | |
ebfd146a IR |
8599 | |
8600 | bool | |
355fe088 | 8601 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 8602 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
8603 | enum tree_code *code1, enum tree_code *code2, |
8604 | int *multi_step_cvt, | |
9771b263 | 8605 | vec<tree> *interm_types) |
ebfd146a IR |
8606 | { |
8607 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8608 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 8609 | struct loop *vect_loop = NULL; |
ef4bddc2 | 8610 | machine_mode vec_mode; |
81f40b79 | 8611 | enum insn_code icode1, icode2; |
ebfd146a | 8612 | optab optab1, optab2; |
b690cc0f RG |
8613 | tree vectype = vectype_in; |
8614 | tree wide_vectype = vectype_out; | |
ebfd146a | 8615 | enum tree_code c1, c2; |
4a00c761 JJ |
8616 | int i; |
8617 | tree prev_type, intermediate_type; | |
ef4bddc2 | 8618 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 8619 | optab optab3, optab4; |
ebfd146a | 8620 | |
4a00c761 | 8621 | *multi_step_cvt = 0; |
4ef69dfc IR |
8622 | if (loop_info) |
8623 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
8624 | ||
ebfd146a IR |
8625 | switch (code) |
8626 | { | |
8627 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
8628 | /* The result of a vectorized widening operation usually requires |
8629 | two vectors (because the widened results do not fit into one vector). | |
8630 | The generated vector results would normally be expected to be | |
8631 | generated in the same order as in the original scalar computation, | |
8632 | i.e. if 8 results are generated in each vector iteration, they are | |
8633 | to be organized as follows: | |
8634 | vect1: [res1,res2,res3,res4], | |
8635 | vect2: [res5,res6,res7,res8]. | |
8636 | ||
8637 | However, in the special case that the result of the widening | |
8638 | operation is used in a reduction computation only, the order doesn't | |
8639 | matter (because when vectorizing a reduction we change the order of | |
8640 | the computation). Some targets can take advantage of this and | |
8641 | generate more efficient code. For example, targets like Altivec, | |
8642 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
8643 | generate the following vectors: | |
8644 | vect1: [res1,res3,res5,res7], | |
8645 | vect2: [res2,res4,res6,res8]. | |
8646 | ||
8647 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
8648 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
8649 | iterations in parallel). We therefore don't allow to change the | |
8650 | order of the computation in the inner-loop during outer-loop | |
8651 | vectorization. */ | |
8652 | /* TODO: Another case in which order doesn't *really* matter is when we | |
8653 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
8654 | Normally, pack_trunc performs an even/odd permute, whereas the | |
8655 | repack from an even/odd expansion would be an interleave, which | |
8656 | would be significantly simpler for e.g. AVX2. */ | |
8657 | /* In any case, in order to avoid duplicating the code below, recurse | |
8658 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
8659 | are properly set up for the caller. If we fail, we'll continue with | |
8660 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
8661 | if (vect_loop | |
8662 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
8663 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
8664 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
8665 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
8666 | code1, code2, multi_step_cvt, |
8667 | interm_types)) | |
ebc047a2 CH |
8668 | { |
8669 | /* Elements in a vector with vect_used_by_reduction property cannot | |
8670 | be reordered if the use chain with this property does not have the | |
8671 | same operation. One such an example is s += a * b, where elements | |
8672 | in a and b cannot be reordered. Here we check if the vector defined | |
8673 | by STMT is only directly used in the reduction statement. */ | |
8674 | tree lhs = gimple_assign_lhs (stmt); | |
8675 | use_operand_p dummy; | |
355fe088 | 8676 | gimple *use_stmt; |
ebc047a2 CH |
8677 | stmt_vec_info use_stmt_info = NULL; |
8678 | if (single_imm_use (lhs, &dummy, &use_stmt) | |
8679 | && (use_stmt_info = vinfo_for_stmt (use_stmt)) | |
8680 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
8681 | return true; | |
8682 | } | |
4a00c761 JJ |
8683 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
8684 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
8685 | break; |
8686 | ||
81c40241 RB |
8687 | case DOT_PROD_EXPR: |
8688 | c1 = DOT_PROD_EXPR; | |
8689 | c2 = DOT_PROD_EXPR; | |
8690 | break; | |
8691 | ||
8692 | case SAD_EXPR: | |
8693 | c1 = SAD_EXPR; | |
8694 | c2 = SAD_EXPR; | |
8695 | break; | |
8696 | ||
6ae6116f RH |
8697 | case VEC_WIDEN_MULT_EVEN_EXPR: |
8698 | /* Support the recursion induced just above. */ | |
8699 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
8700 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
8701 | break; | |
8702 | ||
36ba4aae | 8703 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
8704 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
8705 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
8706 | break; |
8707 | ||
ebfd146a | 8708 | CASE_CONVERT: |
4a00c761 JJ |
8709 | c1 = VEC_UNPACK_LO_EXPR; |
8710 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
8711 | break; |
8712 | ||
8713 | case FLOAT_EXPR: | |
4a00c761 JJ |
8714 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
8715 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
8716 | break; |
8717 | ||
8718 | case FIX_TRUNC_EXPR: | |
8719 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
8720 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
8721 | computing the operation. */ | |
8722 | return false; | |
8723 | ||
8724 | default: | |
8725 | gcc_unreachable (); | |
8726 | } | |
8727 | ||
6ae6116f | 8728 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 8729 | std::swap (c1, c2); |
4a00c761 | 8730 | |
ebfd146a IR |
8731 | if (code == FIX_TRUNC_EXPR) |
8732 | { | |
8733 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
8734 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
8735 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
8736 | } |
8737 | else | |
8738 | { | |
8739 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
8740 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
8741 | } | |
8742 | ||
8743 | if (!optab1 || !optab2) | |
8744 | return false; | |
8745 | ||
8746 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
8747 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
8748 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
8749 | return false; |
8750 | ||
4a00c761 JJ |
8751 | *code1 = c1; |
8752 | *code2 = c2; | |
8753 | ||
8754 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
8755 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
8756 | return true; | |
8757 | ||
b8698a0f | 8758 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 8759 | types. */ |
ebfd146a | 8760 | |
4a00c761 JJ |
8761 | prev_type = vectype; |
8762 | prev_mode = vec_mode; | |
b8698a0f | 8763 | |
4a00c761 JJ |
8764 | if (!CONVERT_EXPR_CODE_P (code)) |
8765 | return false; | |
b8698a0f | 8766 | |
4a00c761 JJ |
8767 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
8768 | intermediate steps in promotion sequence. We try | |
8769 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
8770 | not. */ | |
9771b263 | 8771 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
8772 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
8773 | { | |
8774 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
8775 | intermediate_type | |
8776 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
8777 | TYPE_UNSIGNED (prev_type)); | |
8778 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
8779 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
8780 | ||
8781 | if (!optab3 || !optab4 | |
8782 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
8783 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
8784 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
8785 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
8786 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
8787 | == CODE_FOR_nothing) | |
8788 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
8789 | == CODE_FOR_nothing)) | |
8790 | break; | |
ebfd146a | 8791 | |
9771b263 | 8792 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
8793 | (*multi_step_cvt)++; |
8794 | ||
8795 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
8796 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
8797 | return true; | |
8798 | ||
8799 | prev_type = intermediate_type; | |
8800 | prev_mode = intermediate_mode; | |
ebfd146a IR |
8801 | } |
8802 | ||
9771b263 | 8803 | interm_types->release (); |
4a00c761 | 8804 | return false; |
ebfd146a IR |
8805 | } |
8806 | ||
8807 | ||
8808 | /* Function supportable_narrowing_operation | |
8809 | ||
b8698a0f L |
8810 | Check whether an operation represented by the code CODE is a |
8811 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
8812 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
8813 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 8814 | |
ebfd146a | 8815 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 8816 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
8817 | the target platform directly via vector tree-codes. |
8818 | ||
8819 | Output: | |
b8698a0f L |
8820 | - CODE1 is the code of a vector operation to be used when |
8821 | vectorizing the operation, if available. | |
ebfd146a IR |
8822 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
8823 | case of multi-step conversion (like int->short->char - in that case | |
8824 | MULTI_STEP_CVT will be 1). | |
8825 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 8826 | narrowing operation (short in the above example). */ |
ebfd146a IR |
8827 | |
8828 | bool | |
8829 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 8830 | tree vectype_out, tree vectype_in, |
ebfd146a | 8831 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 8832 | vec<tree> *interm_types) |
ebfd146a | 8833 | { |
ef4bddc2 | 8834 | machine_mode vec_mode; |
ebfd146a IR |
8835 | enum insn_code icode1; |
8836 | optab optab1, interm_optab; | |
b690cc0f RG |
8837 | tree vectype = vectype_in; |
8838 | tree narrow_vectype = vectype_out; | |
ebfd146a | 8839 | enum tree_code c1; |
4a00c761 | 8840 | tree intermediate_type; |
ef4bddc2 | 8841 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 8842 | int i; |
4a00c761 | 8843 | bool uns; |
ebfd146a | 8844 | |
4a00c761 | 8845 | *multi_step_cvt = 0; |
ebfd146a IR |
8846 | switch (code) |
8847 | { | |
8848 | CASE_CONVERT: | |
8849 | c1 = VEC_PACK_TRUNC_EXPR; | |
8850 | break; | |
8851 | ||
8852 | case FIX_TRUNC_EXPR: | |
8853 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
8854 | break; | |
8855 | ||
8856 | case FLOAT_EXPR: | |
8857 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
8858 | tree code and optabs used for computing the operation. */ | |
8859 | return false; | |
8860 | ||
8861 | default: | |
8862 | gcc_unreachable (); | |
8863 | } | |
8864 | ||
8865 | if (code == FIX_TRUNC_EXPR) | |
8866 | /* The signedness is determined from output operand. */ | |
b690cc0f | 8867 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
8868 | else |
8869 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
8870 | ||
8871 | if (!optab1) | |
8872 | return false; | |
8873 | ||
8874 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 8875 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
8876 | return false; |
8877 | ||
4a00c761 JJ |
8878 | *code1 = c1; |
8879 | ||
8880 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
8881 | return true; | |
8882 | ||
ebfd146a IR |
8883 | /* Check if it's a multi-step conversion that can be done using intermediate |
8884 | types. */ | |
4a00c761 JJ |
8885 | prev_mode = vec_mode; |
8886 | if (code == FIX_TRUNC_EXPR) | |
8887 | uns = TYPE_UNSIGNED (vectype_out); | |
8888 | else | |
8889 | uns = TYPE_UNSIGNED (vectype); | |
8890 | ||
8891 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
8892 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
8893 | costly than signed. */ | |
8894 | if (code == FIX_TRUNC_EXPR && uns) | |
8895 | { | |
8896 | enum insn_code icode2; | |
8897 | ||
8898 | intermediate_type | |
8899 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
8900 | interm_optab | |
8901 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 8902 | if (interm_optab != unknown_optab |
4a00c761 JJ |
8903 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
8904 | && insn_data[icode1].operand[0].mode | |
8905 | == insn_data[icode2].operand[0].mode) | |
8906 | { | |
8907 | uns = false; | |
8908 | optab1 = interm_optab; | |
8909 | icode1 = icode2; | |
8910 | } | |
8911 | } | |
ebfd146a | 8912 | |
4a00c761 JJ |
8913 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
8914 | intermediate steps in promotion sequence. We try | |
8915 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 8916 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
8917 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
8918 | { | |
8919 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
8920 | intermediate_type | |
8921 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
8922 | interm_optab | |
8923 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
8924 | optab_default); | |
8925 | if (!interm_optab | |
8926 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
8927 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
8928 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
8929 | == CODE_FOR_nothing)) | |
8930 | break; | |
8931 | ||
9771b263 | 8932 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
8933 | (*multi_step_cvt)++; |
8934 | ||
8935 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
8936 | return true; | |
8937 | ||
8938 | prev_mode = intermediate_mode; | |
8939 | optab1 = interm_optab; | |
ebfd146a IR |
8940 | } |
8941 | ||
9771b263 | 8942 | interm_types->release (); |
4a00c761 | 8943 | return false; |
ebfd146a | 8944 | } |