]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
818ab71a | 2 | Copyright (C) 2003-2016 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" |
70439f0d | 50 | #include "internal-fn.h" |
ebfd146a | 51 | |
7ee2468b SB |
52 | /* For lang_hooks.types.type_for_mode. */ |
53 | #include "langhooks.h" | |
ebfd146a | 54 | |
c3e7ee41 BS |
55 | /* Return the vectorized type for the given statement. */ |
56 | ||
57 | tree | |
58 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
59 | { | |
60 | return STMT_VINFO_VECTYPE (stmt_info); | |
61 | } | |
62 | ||
63 | /* Return TRUE iff the given statement is in an inner loop relative to | |
64 | the loop being vectorized. */ | |
65 | bool | |
66 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
67 | { | |
355fe088 | 68 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
c3e7ee41 BS |
69 | basic_block bb = gimple_bb (stmt); |
70 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
71 | struct loop* loop; | |
72 | ||
73 | if (!loop_vinfo) | |
74 | return false; | |
75 | ||
76 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
77 | ||
78 | return (bb->loop_father == loop->inner); | |
79 | } | |
80 | ||
81 | /* Record the cost of a statement, either by directly informing the | |
82 | target model or by saving it in a vector for later processing. | |
83 | Return a preliminary estimate of the statement's cost. */ | |
84 | ||
85 | unsigned | |
92345349 | 86 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 87 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 88 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 89 | { |
92345349 | 90 | if (body_cost_vec) |
c3e7ee41 | 91 | { |
92345349 | 92 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
ddf56386 RB |
93 | stmt_info_for_cost si = { count, kind, |
94 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
95 | misalign }; | |
96 | body_cost_vec->safe_push (si); | |
c3e7ee41 | 97 | return (unsigned) |
92345349 | 98 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
99 | } |
100 | else | |
310213d4 RB |
101 | return add_stmt_cost (stmt_info->vinfo->target_cost_data, |
102 | count, kind, stmt_info, misalign, where); | |
c3e7ee41 BS |
103 | } |
104 | ||
272c6793 RS |
105 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
106 | ||
107 | static tree | |
108 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
109 | { | |
110 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
111 | "vect_array"); | |
112 | } | |
113 | ||
114 | /* ARRAY is an array of vectors created by create_vector_array. | |
115 | Return an SSA_NAME for the vector in index N. The reference | |
116 | is part of the vectorization of STMT and the vector is associated | |
117 | with scalar destination SCALAR_DEST. */ | |
118 | ||
119 | static tree | |
355fe088 | 120 | read_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree scalar_dest, |
272c6793 RS |
121 | tree array, unsigned HOST_WIDE_INT n) |
122 | { | |
123 | tree vect_type, vect, vect_name, array_ref; | |
355fe088 | 124 | gimple *new_stmt; |
272c6793 RS |
125 | |
126 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
127 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
128 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
129 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
130 | build_int_cst (size_type_node, n), | |
131 | NULL_TREE, NULL_TREE); | |
132 | ||
133 | new_stmt = gimple_build_assign (vect, array_ref); | |
134 | vect_name = make_ssa_name (vect, new_stmt); | |
135 | gimple_assign_set_lhs (new_stmt, vect_name); | |
136 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
137 | |
138 | return vect_name; | |
139 | } | |
140 | ||
141 | /* ARRAY is an array of vectors created by create_vector_array. | |
142 | Emit code to store SSA_NAME VECT in index N of the array. | |
143 | The store is part of the vectorization of STMT. */ | |
144 | ||
145 | static void | |
355fe088 | 146 | write_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree vect, |
272c6793 RS |
147 | tree array, unsigned HOST_WIDE_INT n) |
148 | { | |
149 | tree array_ref; | |
355fe088 | 150 | gimple *new_stmt; |
272c6793 RS |
151 | |
152 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
153 | build_int_cst (size_type_node, n), | |
154 | NULL_TREE, NULL_TREE); | |
155 | ||
156 | new_stmt = gimple_build_assign (array_ref, vect); | |
157 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
158 | } |
159 | ||
160 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
161 | of *PTR. The memory reference replaces those in FIRST_DR | |
162 | (and its group). */ | |
163 | ||
164 | static tree | |
165 | create_array_ref (tree type, tree ptr, struct data_reference *first_dr) | |
166 | { | |
272c6793 RS |
167 | tree mem_ref, alias_ptr_type; |
168 | ||
169 | alias_ptr_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
170 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); | |
171 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 172 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
173 | return mem_ref; |
174 | } | |
175 | ||
ebfd146a IR |
176 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
177 | ||
178 | /* Function vect_mark_relevant. | |
179 | ||
180 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
181 | ||
182 | static void | |
355fe088 | 183 | vect_mark_relevant (vec<gimple *> *worklist, gimple *stmt, |
97ecdb46 | 184 | enum vect_relevant relevant, bool live_p) |
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 ()) |
66c16fd9 RB |
192 | { |
193 | dump_printf_loc (MSG_NOTE, vect_location, | |
194 | "mark relevant %d, live %d: ", relevant, live_p); | |
195 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
196 | } | |
ebfd146a | 197 | |
83197f37 IR |
198 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
199 | related pattern stmt instead of the original stmt. However, such stmts | |
200 | may have their own uses that are not in any pattern, in such cases the | |
201 | stmt itself should be marked. */ | |
ebfd146a IR |
202 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
203 | { | |
97ecdb46 JJ |
204 | /* This is the last stmt in a sequence that was detected as a |
205 | pattern that can potentially be vectorized. Don't mark the stmt | |
206 | as relevant/live because it's not going to be vectorized. | |
207 | Instead mark the pattern-stmt that replaces it. */ | |
83197f37 | 208 | |
97ecdb46 JJ |
209 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); |
210 | ||
211 | if (dump_enabled_p ()) | |
212 | dump_printf_loc (MSG_NOTE, vect_location, | |
213 | "last stmt in pattern. don't mark" | |
214 | " relevant/live.\n"); | |
215 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
216 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
217 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
218 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
219 | stmt = pattern_stmt; | |
ebfd146a IR |
220 | } |
221 | ||
222 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
223 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
224 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
225 | ||
226 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
227 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
228 | { | |
73fbfcad | 229 | if (dump_enabled_p ()) |
78c60e3d | 230 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 231 | "already marked relevant/live.\n"); |
ebfd146a IR |
232 | return; |
233 | } | |
234 | ||
9771b263 | 235 | worklist->safe_push (stmt); |
ebfd146a IR |
236 | } |
237 | ||
238 | ||
b28ead45 AH |
239 | /* Function is_simple_and_all_uses_invariant |
240 | ||
241 | Return true if STMT is simple and all uses of it are invariant. */ | |
242 | ||
243 | bool | |
244 | is_simple_and_all_uses_invariant (gimple *stmt, loop_vec_info loop_vinfo) | |
245 | { | |
246 | tree op; | |
247 | gimple *def_stmt; | |
248 | ssa_op_iter iter; | |
249 | ||
250 | if (!is_gimple_assign (stmt)) | |
251 | return false; | |
252 | ||
253 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
254 | { | |
255 | enum vect_def_type dt = vect_uninitialized_def; | |
256 | ||
257 | if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt)) | |
258 | { | |
259 | if (dump_enabled_p ()) | |
260 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
261 | "use not simple.\n"); | |
262 | return false; | |
263 | } | |
264 | ||
265 | if (dt != vect_external_def && dt != vect_constant_def) | |
266 | return false; | |
267 | } | |
268 | return true; | |
269 | } | |
270 | ||
ebfd146a IR |
271 | /* Function vect_stmt_relevant_p. |
272 | ||
273 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
274 | "relevant for vectorization". | |
275 | ||
276 | A stmt is considered "relevant for vectorization" if: | |
277 | - it has uses outside the loop. | |
278 | - it has vdefs (it alters memory). | |
279 | - control stmts in the loop (except for the exit condition). | |
280 | ||
281 | CHECKME: what other side effects would the vectorizer allow? */ | |
282 | ||
283 | static bool | |
355fe088 | 284 | vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo, |
ebfd146a IR |
285 | enum vect_relevant *relevant, bool *live_p) |
286 | { | |
287 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
288 | ssa_op_iter op_iter; | |
289 | imm_use_iterator imm_iter; | |
290 | use_operand_p use_p; | |
291 | def_operand_p def_p; | |
292 | ||
8644a673 | 293 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
294 | *live_p = false; |
295 | ||
296 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
297 | if (is_ctrl_stmt (stmt) |
298 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
299 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 300 | *relevant = vect_used_in_scope; |
ebfd146a IR |
301 | |
302 | /* changing memory. */ | |
303 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
304 | if (gimple_vdef (stmt) |
305 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 306 | { |
73fbfcad | 307 | if (dump_enabled_p ()) |
78c60e3d | 308 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 309 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 310 | *relevant = vect_used_in_scope; |
ebfd146a IR |
311 | } |
312 | ||
313 | /* uses outside the loop. */ | |
314 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
315 | { | |
316 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
317 | { | |
318 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
319 | if (!flow_bb_inside_loop_p (loop, bb)) | |
320 | { | |
73fbfcad | 321 | if (dump_enabled_p ()) |
78c60e3d | 322 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 323 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 324 | |
3157b0c2 AO |
325 | if (is_gimple_debug (USE_STMT (use_p))) |
326 | continue; | |
327 | ||
ebfd146a IR |
328 | /* We expect all such uses to be in the loop exit phis |
329 | (because of loop closed form) */ | |
330 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
331 | gcc_assert (bb == single_exit (loop)->dest); | |
332 | ||
333 | *live_p = true; | |
334 | } | |
335 | } | |
336 | } | |
337 | ||
3a2edf4c AH |
338 | if (*live_p && *relevant == vect_unused_in_scope |
339 | && !is_simple_and_all_uses_invariant (stmt, loop_vinfo)) | |
b28ead45 AH |
340 | { |
341 | if (dump_enabled_p ()) | |
342 | dump_printf_loc (MSG_NOTE, vect_location, | |
343 | "vec_stmt_relevant_p: stmt live but not relevant.\n"); | |
344 | *relevant = vect_used_only_live; | |
345 | } | |
346 | ||
ebfd146a IR |
347 | return (*live_p || *relevant); |
348 | } | |
349 | ||
350 | ||
b8698a0f | 351 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 352 | |
ff802fa1 | 353 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
354 | used in STMT for anything other than indexing an array. */ |
355 | ||
356 | static bool | |
355fe088 | 357 | exist_non_indexing_operands_for_use_p (tree use, gimple *stmt) |
ebfd146a IR |
358 | { |
359 | tree operand; | |
360 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 361 | |
ff802fa1 | 362 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
363 | reference in STMT, then any operand that corresponds to USE |
364 | is not indexing an array. */ | |
365 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
366 | return true; | |
59a05b0c | 367 | |
ebfd146a IR |
368 | /* STMT has a data_ref. FORNOW this means that its of one of |
369 | the following forms: | |
370 | -1- ARRAY_REF = var | |
371 | -2- var = ARRAY_REF | |
372 | (This should have been verified in analyze_data_refs). | |
373 | ||
374 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 375 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
376 | for array indexing. |
377 | ||
378 | Therefore, all we need to check is if STMT falls into the | |
379 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
380 | |
381 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
382 | { |
383 | if (is_gimple_call (stmt) | |
384 | && gimple_call_internal_p (stmt)) | |
385 | switch (gimple_call_internal_fn (stmt)) | |
386 | { | |
387 | case IFN_MASK_STORE: | |
388 | operand = gimple_call_arg (stmt, 3); | |
389 | if (operand == use) | |
390 | return true; | |
391 | /* FALLTHRU */ | |
392 | case IFN_MASK_LOAD: | |
393 | operand = gimple_call_arg (stmt, 2); | |
394 | if (operand == use) | |
395 | return true; | |
396 | break; | |
397 | default: | |
398 | break; | |
399 | } | |
400 | return false; | |
401 | } | |
402 | ||
59a05b0c EB |
403 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
404 | return false; | |
ebfd146a | 405 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
406 | if (TREE_CODE (operand) != SSA_NAME) |
407 | return false; | |
408 | ||
409 | if (operand == use) | |
410 | return true; | |
411 | ||
412 | return false; | |
413 | } | |
414 | ||
415 | ||
b8698a0f | 416 | /* |
ebfd146a IR |
417 | Function process_use. |
418 | ||
419 | Inputs: | |
420 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b28ead45 | 421 | - RELEVANT - enum value to be set in the STMT_VINFO of the stmt |
ff802fa1 | 422 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 423 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
424 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
425 | be performed. | |
ebfd146a IR |
426 | |
427 | Outputs: | |
428 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
429 | relevance info of the DEF_STMT of this USE: | |
430 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
431 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
432 | Exceptions: | |
433 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 434 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 435 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
436 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
437 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
438 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
439 | be modified accordingly. | |
440 | ||
441 | Return true if everything is as expected. Return false otherwise. */ | |
442 | ||
443 | static bool | |
b28ead45 | 444 | process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo, |
355fe088 | 445 | enum vect_relevant relevant, vec<gimple *> *worklist, |
aec7ae7d | 446 | bool force) |
ebfd146a IR |
447 | { |
448 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
449 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
450 | stmt_vec_info dstmt_vinfo; | |
451 | basic_block bb, def_bb; | |
355fe088 | 452 | gimple *def_stmt; |
ebfd146a IR |
453 | enum vect_def_type dt; |
454 | ||
b8698a0f | 455 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 456 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 457 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
458 | return true; |
459 | ||
81c40241 | 460 | if (!vect_is_simple_use (use, loop_vinfo, &def_stmt, &dt)) |
b8698a0f | 461 | { |
73fbfcad | 462 | if (dump_enabled_p ()) |
78c60e3d | 463 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 464 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
465 | return false; |
466 | } | |
467 | ||
468 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
469 | return true; | |
470 | ||
471 | def_bb = gimple_bb (def_stmt); | |
472 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
473 | { | |
73fbfcad | 474 | if (dump_enabled_p ()) |
e645e942 | 475 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
476 | return true; |
477 | } | |
478 | ||
b8698a0f L |
479 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
480 | DEF_STMT must have already been processed, because this should be the | |
481 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
482 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
483 | check that everything is as expected, and we are done. */ |
484 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
485 | bb = gimple_bb (stmt); | |
486 | if (gimple_code (stmt) == GIMPLE_PHI | |
487 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
488 | && gimple_code (def_stmt) != GIMPLE_PHI | |
489 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
490 | && bb->loop_father == def_bb->loop_father) | |
491 | { | |
73fbfcad | 492 | if (dump_enabled_p ()) |
78c60e3d | 493 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 494 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
495 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
496 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
497 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 498 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 499 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
500 | return true; |
501 | } | |
502 | ||
503 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
504 | outer-loop-header-bb: | |
505 | d = def_stmt | |
506 | inner-loop: | |
507 | stmt # use (d) | |
508 | outer-loop-tail-bb: | |
509 | ... */ | |
510 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
511 | { | |
73fbfcad | 512 | if (dump_enabled_p ()) |
78c60e3d | 513 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 514 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 515 | |
ebfd146a IR |
516 | switch (relevant) |
517 | { | |
8644a673 | 518 | case vect_unused_in_scope: |
7c5222ff IR |
519 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
520 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 521 | break; |
7c5222ff | 522 | |
ebfd146a | 523 | case vect_used_in_outer_by_reduction: |
7c5222ff | 524 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
525 | relevant = vect_used_by_reduction; |
526 | break; | |
7c5222ff | 527 | |
ebfd146a | 528 | case vect_used_in_outer: |
7c5222ff | 529 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 530 | relevant = vect_used_in_scope; |
ebfd146a | 531 | break; |
7c5222ff | 532 | |
8644a673 | 533 | case vect_used_in_scope: |
ebfd146a IR |
534 | break; |
535 | ||
536 | default: | |
537 | gcc_unreachable (); | |
b8698a0f | 538 | } |
ebfd146a IR |
539 | } |
540 | ||
541 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
542 | outer-loop-header-bb: | |
543 | ... | |
544 | inner-loop: | |
545 | d = def_stmt | |
06066f92 | 546 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
547 | stmt # use (d) */ |
548 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
549 | { | |
73fbfcad | 550 | if (dump_enabled_p ()) |
78c60e3d | 551 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 552 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 553 | |
ebfd146a IR |
554 | switch (relevant) |
555 | { | |
8644a673 | 556 | case vect_unused_in_scope: |
b8698a0f | 557 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 558 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 559 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
560 | break; |
561 | ||
ebfd146a | 562 | case vect_used_by_reduction: |
b28ead45 | 563 | case vect_used_only_live: |
ebfd146a IR |
564 | relevant = vect_used_in_outer_by_reduction; |
565 | break; | |
566 | ||
8644a673 | 567 | case vect_used_in_scope: |
ebfd146a IR |
568 | relevant = vect_used_in_outer; |
569 | break; | |
570 | ||
571 | default: | |
572 | gcc_unreachable (); | |
573 | } | |
574 | } | |
575 | ||
b28ead45 | 576 | vect_mark_relevant (worklist, def_stmt, relevant, false); |
ebfd146a IR |
577 | return true; |
578 | } | |
579 | ||
580 | ||
581 | /* Function vect_mark_stmts_to_be_vectorized. | |
582 | ||
583 | Not all stmts in the loop need to be vectorized. For example: | |
584 | ||
585 | for i... | |
586 | for j... | |
587 | 1. T0 = i + j | |
588 | 2. T1 = a[T0] | |
589 | ||
590 | 3. j = j + 1 | |
591 | ||
592 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
593 | addressing of vectorized data-refs are handled differently. | |
594 | ||
595 | This pass detects such stmts. */ | |
596 | ||
597 | bool | |
598 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
599 | { | |
ebfd146a IR |
600 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
601 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
602 | unsigned int nbbs = loop->num_nodes; | |
603 | gimple_stmt_iterator si; | |
355fe088 | 604 | gimple *stmt; |
ebfd146a IR |
605 | unsigned int i; |
606 | stmt_vec_info stmt_vinfo; | |
607 | basic_block bb; | |
355fe088 | 608 | gimple *phi; |
ebfd146a | 609 | bool live_p; |
b28ead45 | 610 | enum vect_relevant relevant; |
ebfd146a | 611 | |
73fbfcad | 612 | if (dump_enabled_p ()) |
78c60e3d | 613 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 614 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 615 | |
355fe088 | 616 | auto_vec<gimple *, 64> worklist; |
ebfd146a IR |
617 | |
618 | /* 1. Init worklist. */ | |
619 | for (i = 0; i < nbbs; i++) | |
620 | { | |
621 | bb = bbs[i]; | |
622 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 623 | { |
ebfd146a | 624 | phi = gsi_stmt (si); |
73fbfcad | 625 | if (dump_enabled_p ()) |
ebfd146a | 626 | { |
78c60e3d SS |
627 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
628 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
629 | } |
630 | ||
631 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 632 | vect_mark_relevant (&worklist, phi, relevant, live_p); |
ebfd146a IR |
633 | } |
634 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
635 | { | |
636 | stmt = gsi_stmt (si); | |
73fbfcad | 637 | if (dump_enabled_p ()) |
ebfd146a | 638 | { |
78c60e3d SS |
639 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
640 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 641 | } |
ebfd146a IR |
642 | |
643 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 644 | vect_mark_relevant (&worklist, stmt, relevant, live_p); |
ebfd146a IR |
645 | } |
646 | } | |
647 | ||
648 | /* 2. Process_worklist */ | |
9771b263 | 649 | while (worklist.length () > 0) |
ebfd146a IR |
650 | { |
651 | use_operand_p use_p; | |
652 | ssa_op_iter iter; | |
653 | ||
9771b263 | 654 | stmt = worklist.pop (); |
73fbfcad | 655 | if (dump_enabled_p ()) |
ebfd146a | 656 | { |
78c60e3d SS |
657 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
658 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
659 | } |
660 | ||
b8698a0f | 661 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
b28ead45 AH |
662 | (DEF_STMT) as relevant/irrelevant according to the relevance property |
663 | of STMT. */ | |
ebfd146a IR |
664 | stmt_vinfo = vinfo_for_stmt (stmt); |
665 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
ebfd146a | 666 | |
b28ead45 AH |
667 | /* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is |
668 | propagated as is to the DEF_STMTs of its USEs. | |
ebfd146a IR |
669 | |
670 | One exception is when STMT has been identified as defining a reduction | |
b28ead45 | 671 | variable; in this case we set the relevance to vect_used_by_reduction. |
ebfd146a | 672 | This is because we distinguish between two kinds of relevant stmts - |
b8698a0f | 673 | those that are used by a reduction computation, and those that are |
ff802fa1 | 674 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 675 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 676 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 677 | |
b28ead45 | 678 | switch (STMT_VINFO_DEF_TYPE (stmt_vinfo)) |
ebfd146a | 679 | { |
06066f92 | 680 | case vect_reduction_def: |
b28ead45 AH |
681 | gcc_assert (relevant != vect_unused_in_scope); |
682 | if (relevant != vect_unused_in_scope | |
683 | && relevant != vect_used_in_scope | |
684 | && relevant != vect_used_by_reduction | |
685 | && relevant != vect_used_only_live) | |
06066f92 | 686 | { |
b28ead45 AH |
687 | if (dump_enabled_p ()) |
688 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
689 | "unsupported use of reduction.\n"); | |
690 | return false; | |
06066f92 | 691 | } |
06066f92 | 692 | break; |
b8698a0f | 693 | |
06066f92 | 694 | case vect_nested_cycle: |
b28ead45 AH |
695 | if (relevant != vect_unused_in_scope |
696 | && relevant != vect_used_in_outer_by_reduction | |
697 | && relevant != vect_used_in_outer) | |
06066f92 | 698 | { |
73fbfcad | 699 | if (dump_enabled_p ()) |
78c60e3d | 700 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 701 | "unsupported use of nested cycle.\n"); |
7c5222ff | 702 | |
06066f92 IR |
703 | return false; |
704 | } | |
b8698a0f L |
705 | break; |
706 | ||
06066f92 | 707 | case vect_double_reduction_def: |
b28ead45 AH |
708 | if (relevant != vect_unused_in_scope |
709 | && relevant != vect_used_by_reduction | |
710 | && relevant != vect_used_only_live) | |
06066f92 | 711 | { |
73fbfcad | 712 | if (dump_enabled_p ()) |
78c60e3d | 713 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 714 | "unsupported use of double reduction.\n"); |
7c5222ff | 715 | |
7c5222ff | 716 | return false; |
06066f92 | 717 | } |
b8698a0f | 718 | break; |
7c5222ff | 719 | |
06066f92 IR |
720 | default: |
721 | break; | |
7c5222ff | 722 | } |
b8698a0f | 723 | |
aec7ae7d | 724 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
725 | { |
726 | /* Pattern statements are not inserted into the code, so | |
727 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
728 | have to scan the RHS or function arguments instead. */ | |
729 | if (is_gimple_assign (stmt)) | |
730 | { | |
69d2aade JJ |
731 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
732 | tree op = gimple_assign_rhs1 (stmt); | |
733 | ||
734 | i = 1; | |
735 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
736 | { | |
737 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
b28ead45 | 738 | relevant, &worklist, false) |
69d2aade | 739 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
b28ead45 | 740 | relevant, &worklist, false)) |
566d377a | 741 | return false; |
69d2aade JJ |
742 | i = 2; |
743 | } | |
744 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 745 | { |
69d2aade | 746 | op = gimple_op (stmt, i); |
afbe6325 | 747 | if (TREE_CODE (op) == SSA_NAME |
b28ead45 | 748 | && !process_use (stmt, op, loop_vinfo, relevant, |
afbe6325 | 749 | &worklist, false)) |
07687835 | 750 | return false; |
9d5e7640 IR |
751 | } |
752 | } | |
753 | else if (is_gimple_call (stmt)) | |
754 | { | |
755 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
756 | { | |
757 | tree arg = gimple_call_arg (stmt, i); | |
b28ead45 | 758 | if (!process_use (stmt, arg, loop_vinfo, relevant, |
aec7ae7d | 759 | &worklist, false)) |
07687835 | 760 | return false; |
9d5e7640 IR |
761 | } |
762 | } | |
763 | } | |
764 | else | |
765 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
766 | { | |
767 | tree op = USE_FROM_PTR (use_p); | |
b28ead45 | 768 | if (!process_use (stmt, op, loop_vinfo, relevant, |
aec7ae7d | 769 | &worklist, false)) |
07687835 | 770 | return false; |
9d5e7640 | 771 | } |
aec7ae7d | 772 | |
3bab6342 | 773 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d JJ |
774 | { |
775 | tree off; | |
3bab6342 | 776 | tree decl = vect_check_gather_scatter (stmt, loop_vinfo, NULL, &off, NULL); |
aec7ae7d | 777 | gcc_assert (decl); |
b28ead45 | 778 | if (!process_use (stmt, off, loop_vinfo, relevant, &worklist, true)) |
566d377a | 779 | return false; |
aec7ae7d | 780 | } |
ebfd146a IR |
781 | } /* while worklist */ |
782 | ||
ebfd146a IR |
783 | return true; |
784 | } | |
785 | ||
786 | ||
b8698a0f | 787 | /* Function vect_model_simple_cost. |
ebfd146a | 788 | |
b8698a0f | 789 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
790 | single op. Right now, this does not account for multiple insns that could |
791 | be generated for the single vector op. We will handle that shortly. */ | |
792 | ||
793 | void | |
b8698a0f | 794 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 BS |
795 | enum vect_def_type *dt, |
796 | stmt_vector_for_cost *prologue_cost_vec, | |
797 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
798 | { |
799 | int i; | |
92345349 | 800 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
801 | |
802 | /* The SLP costs were already calculated during SLP tree build. */ | |
803 | if (PURE_SLP_STMT (stmt_info)) | |
804 | return; | |
805 | ||
ebfd146a IR |
806 | /* FORNOW: Assuming maximum 2 args per stmts. */ |
807 | for (i = 0; i < 2; i++) | |
92345349 BS |
808 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
809 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, vector_stmt, | |
810 | stmt_info, 0, vect_prologue); | |
c3e7ee41 BS |
811 | |
812 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
813 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
814 | stmt_info, 0, vect_body); | |
c3e7ee41 | 815 | |
73fbfcad | 816 | if (dump_enabled_p ()) |
78c60e3d SS |
817 | dump_printf_loc (MSG_NOTE, vect_location, |
818 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 819 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
820 | } |
821 | ||
822 | ||
8bd37302 BS |
823 | /* Model cost for type demotion and promotion operations. PWR is normally |
824 | zero for single-step promotions and demotions. It will be one if | |
825 | two-step promotion/demotion is required, and so on. Each additional | |
826 | step doubles the number of instructions required. */ | |
827 | ||
828 | static void | |
829 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
830 | enum vect_def_type *dt, int pwr) | |
831 | { | |
832 | int i, tmp; | |
92345349 | 833 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
834 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
835 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
836 | void *target_cost_data; | |
8bd37302 BS |
837 | |
838 | /* The SLP costs were already calculated during SLP tree build. */ | |
839 | if (PURE_SLP_STMT (stmt_info)) | |
840 | return; | |
841 | ||
c3e7ee41 BS |
842 | if (loop_vinfo) |
843 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
844 | else | |
845 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
846 | ||
8bd37302 BS |
847 | for (i = 0; i < pwr + 1; i++) |
848 | { | |
849 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
850 | (i + 1) : i; | |
c3e7ee41 | 851 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
852 | vec_promote_demote, stmt_info, 0, |
853 | vect_body); | |
8bd37302 BS |
854 | } |
855 | ||
856 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
857 | for (i = 0; i < 2; i++) | |
92345349 BS |
858 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
859 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
860 | stmt_info, 0, vect_prologue); | |
8bd37302 | 861 | |
73fbfcad | 862 | if (dump_enabled_p ()) |
78c60e3d SS |
863 | dump_printf_loc (MSG_NOTE, vect_location, |
864 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 865 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
866 | } |
867 | ||
0d0293ac | 868 | /* Function vect_cost_group_size |
b8698a0f | 869 | |
0d0293ac | 870 | For grouped load or store, return the group_size only if it is the first |
ebfd146a IR |
871 | load or store of a group, else return 1. This ensures that group size is |
872 | only returned once per group. */ | |
873 | ||
874 | static int | |
0d0293ac | 875 | vect_cost_group_size (stmt_vec_info stmt_info) |
ebfd146a | 876 | { |
355fe088 | 877 | gimple *first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
878 | |
879 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
e14c1050 | 880 | return GROUP_SIZE (stmt_info); |
ebfd146a IR |
881 | |
882 | return 1; | |
883 | } | |
884 | ||
885 | ||
886 | /* Function vect_model_store_cost | |
887 | ||
0d0293ac MM |
888 | Models cost for stores. In the case of grouped accesses, one access |
889 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
890 | |
891 | void | |
b8698a0f | 892 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
272c6793 | 893 | bool store_lanes_p, enum vect_def_type dt, |
92345349 BS |
894 | slp_tree slp_node, |
895 | stmt_vector_for_cost *prologue_cost_vec, | |
896 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
897 | { |
898 | int group_size; | |
92345349 | 899 | unsigned int inside_cost = 0, prologue_cost = 0; |
720f5239 | 900 | struct data_reference *first_dr; |
355fe088 | 901 | gimple *first_stmt; |
ebfd146a | 902 | |
8644a673 | 903 | if (dt == vect_constant_def || dt == vect_external_def) |
92345349 BS |
904 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
905 | stmt_info, 0, vect_prologue); | |
ebfd146a | 906 | |
0d0293ac MM |
907 | /* Grouped access? */ |
908 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
720f5239 IR |
909 | { |
910 | if (slp_node) | |
911 | { | |
9771b263 | 912 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
720f5239 IR |
913 | group_size = 1; |
914 | } | |
915 | else | |
916 | { | |
e14c1050 | 917 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 918 | group_size = vect_cost_group_size (stmt_info); |
720f5239 IR |
919 | } |
920 | ||
921 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
922 | } | |
0d0293ac | 923 | /* Not a grouped access. */ |
ebfd146a | 924 | else |
720f5239 IR |
925 | { |
926 | group_size = 1; | |
927 | first_dr = STMT_VINFO_DATA_REF (stmt_info); | |
928 | } | |
ebfd146a | 929 | |
272c6793 | 930 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 931 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 RS |
932 | access is instead being provided by a permute-and-store operation, |
933 | include the cost of the permutes. */ | |
cee62fee MM |
934 | if (!store_lanes_p && group_size > 1 |
935 | && !STMT_VINFO_STRIDED_P (stmt_info)) | |
ebfd146a | 936 | { |
e1377713 ES |
937 | /* Uses a high and low interleave or shuffle operations for each |
938 | needed permute. */ | |
939 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; | |
92345349 BS |
940 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
941 | stmt_info, 0, vect_body); | |
ebfd146a | 942 | |
73fbfcad | 943 | if (dump_enabled_p ()) |
78c60e3d | 944 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 945 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 946 | group_size); |
ebfd146a IR |
947 | } |
948 | ||
cee62fee | 949 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 950 | /* Costs of the stores. */ |
cee62fee MM |
951 | if (STMT_VINFO_STRIDED_P (stmt_info) |
952 | && !STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
f2e2a985 MM |
953 | { |
954 | /* N scalar stores plus extracting the elements. */ | |
f2e2a985 MM |
955 | inside_cost += record_stmt_cost (body_cost_vec, |
956 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
957 | scalar_store, stmt_info, 0, vect_body); | |
f2e2a985 MM |
958 | } |
959 | else | |
960 | vect_get_store_cost (first_dr, ncopies, &inside_cost, body_cost_vec); | |
ebfd146a | 961 | |
cee62fee MM |
962 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
963 | inside_cost += record_stmt_cost (body_cost_vec, | |
964 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
965 | vec_to_scalar, stmt_info, 0, vect_body); | |
966 | ||
73fbfcad | 967 | if (dump_enabled_p ()) |
78c60e3d SS |
968 | dump_printf_loc (MSG_NOTE, vect_location, |
969 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 970 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
971 | } |
972 | ||
973 | ||
720f5239 IR |
974 | /* Calculate cost of DR's memory access. */ |
975 | void | |
976 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 977 | unsigned int *inside_cost, |
92345349 | 978 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
979 | { |
980 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 981 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 982 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
983 | |
984 | switch (alignment_support_scheme) | |
985 | { | |
986 | case dr_aligned: | |
987 | { | |
92345349 BS |
988 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
989 | vector_store, stmt_info, 0, | |
990 | vect_body); | |
720f5239 | 991 | |
73fbfcad | 992 | if (dump_enabled_p ()) |
78c60e3d | 993 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 994 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
995 | break; |
996 | } | |
997 | ||
998 | case dr_unaligned_supported: | |
999 | { | |
720f5239 | 1000 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1001 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1002 | unaligned_store, stmt_info, |
92345349 | 1003 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1004 | if (dump_enabled_p ()) |
78c60e3d SS |
1005 | dump_printf_loc (MSG_NOTE, vect_location, |
1006 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1007 | "hardware.\n"); |
720f5239 IR |
1008 | break; |
1009 | } | |
1010 | ||
38eec4c6 UW |
1011 | case dr_unaligned_unsupported: |
1012 | { | |
1013 | *inside_cost = VECT_MAX_COST; | |
1014 | ||
73fbfcad | 1015 | if (dump_enabled_p ()) |
78c60e3d | 1016 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1017 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1018 | break; |
1019 | } | |
1020 | ||
720f5239 IR |
1021 | default: |
1022 | gcc_unreachable (); | |
1023 | } | |
1024 | } | |
1025 | ||
1026 | ||
ebfd146a IR |
1027 | /* Function vect_model_load_cost |
1028 | ||
0d0293ac MM |
1029 | Models cost for loads. In the case of grouped accesses, the last access |
1030 | has the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1031 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1032 | access scheme chosen. */ |
1033 | ||
1034 | void | |
92345349 BS |
1035 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
1036 | bool load_lanes_p, slp_tree slp_node, | |
1037 | stmt_vector_for_cost *prologue_cost_vec, | |
1038 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
1039 | { |
1040 | int group_size; | |
355fe088 | 1041 | gimple *first_stmt; |
ebfd146a | 1042 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; |
92345349 | 1043 | unsigned int inside_cost = 0, prologue_cost = 0; |
ebfd146a | 1044 | |
0d0293ac | 1045 | /* Grouped accesses? */ |
e14c1050 | 1046 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 1047 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && first_stmt && !slp_node) |
ebfd146a | 1048 | { |
0d0293ac | 1049 | group_size = vect_cost_group_size (stmt_info); |
ebfd146a IR |
1050 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
1051 | } | |
0d0293ac | 1052 | /* Not a grouped access. */ |
ebfd146a IR |
1053 | else |
1054 | { | |
1055 | group_size = 1; | |
1056 | first_dr = dr; | |
1057 | } | |
1058 | ||
272c6793 | 1059 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1060 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 RS |
1061 | access is instead being provided by a load-and-permute operation, |
1062 | include the cost of the permutes. */ | |
7b5fc413 | 1063 | if (!load_lanes_p && group_size > 1 |
f2e2a985 | 1064 | && !STMT_VINFO_STRIDED_P (stmt_info)) |
ebfd146a | 1065 | { |
2c23db6d ES |
1066 | /* Uses an even and odd extract operations or shuffle operations |
1067 | for each needed permute. */ | |
1068 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; | |
1069 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, | |
1070 | stmt_info, 0, vect_body); | |
ebfd146a | 1071 | |
73fbfcad | 1072 | if (dump_enabled_p ()) |
e645e942 TJ |
1073 | dump_printf_loc (MSG_NOTE, vect_location, |
1074 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1075 | group_size); |
ebfd146a IR |
1076 | } |
1077 | ||
1078 | /* The loads themselves. */ | |
f2e2a985 | 1079 | if (STMT_VINFO_STRIDED_P (stmt_info) |
7b5fc413 | 1080 | && !STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
a82960aa | 1081 | { |
a21892ad BS |
1082 | /* N scalar loads plus gathering them into a vector. */ |
1083 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
92345349 | 1084 | inside_cost += record_stmt_cost (body_cost_vec, |
c3e7ee41 | 1085 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), |
92345349 | 1086 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1087 | } |
1088 | else | |
1089 | vect_get_load_cost (first_dr, ncopies, | |
1090 | ((!STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1091 | || group_size > 1 || slp_node), | |
92345349 BS |
1092 | &inside_cost, &prologue_cost, |
1093 | prologue_cost_vec, body_cost_vec, true); | |
f2e2a985 | 1094 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
7b5fc413 RB |
1095 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, |
1096 | stmt_info, 0, vect_body); | |
720f5239 | 1097 | |
73fbfcad | 1098 | if (dump_enabled_p ()) |
78c60e3d SS |
1099 | dump_printf_loc (MSG_NOTE, vect_location, |
1100 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1101 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1102 | } |
1103 | ||
1104 | ||
1105 | /* Calculate cost of DR's memory access. */ | |
1106 | void | |
1107 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1108 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1109 | unsigned int *prologue_cost, |
1110 | stmt_vector_for_cost *prologue_cost_vec, | |
1111 | stmt_vector_for_cost *body_cost_vec, | |
1112 | bool record_prologue_costs) | |
720f5239 IR |
1113 | { |
1114 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 1115 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 1116 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
1117 | |
1118 | switch (alignment_support_scheme) | |
ebfd146a IR |
1119 | { |
1120 | case dr_aligned: | |
1121 | { | |
92345349 BS |
1122 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1123 | stmt_info, 0, vect_body); | |
ebfd146a | 1124 | |
73fbfcad | 1125 | if (dump_enabled_p ()) |
78c60e3d | 1126 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1127 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1128 | |
1129 | break; | |
1130 | } | |
1131 | case dr_unaligned_supported: | |
1132 | { | |
720f5239 | 1133 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1134 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1135 | unaligned_load, stmt_info, |
92345349 | 1136 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1137 | |
73fbfcad | 1138 | if (dump_enabled_p ()) |
78c60e3d SS |
1139 | dump_printf_loc (MSG_NOTE, vect_location, |
1140 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1141 | "hardware.\n"); |
ebfd146a IR |
1142 | |
1143 | break; | |
1144 | } | |
1145 | case dr_explicit_realign: | |
1146 | { | |
92345349 BS |
1147 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1148 | vector_load, stmt_info, 0, vect_body); | |
1149 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1150 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1151 | |
1152 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1153 | the containing loop, the following cost should be added to the | |
92345349 | 1154 | prologue costs. */ |
ebfd146a | 1155 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1156 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1157 | stmt_info, 0, vect_body); | |
ebfd146a | 1158 | |
73fbfcad | 1159 | if (dump_enabled_p ()) |
e645e942 TJ |
1160 | dump_printf_loc (MSG_NOTE, vect_location, |
1161 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1162 | |
ebfd146a IR |
1163 | break; |
1164 | } | |
1165 | case dr_explicit_realign_optimized: | |
1166 | { | |
73fbfcad | 1167 | if (dump_enabled_p ()) |
e645e942 | 1168 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1169 | "vect_model_load_cost: unaligned software " |
e645e942 | 1170 | "pipelined.\n"); |
ebfd146a IR |
1171 | |
1172 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1173 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1174 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1175 | access, then the above cost should only be considered for one |
ff802fa1 | 1176 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1177 | and a realignment op. */ |
1178 | ||
92345349 | 1179 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1180 | { |
92345349 BS |
1181 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1182 | vector_stmt, stmt_info, | |
1183 | 0, vect_prologue); | |
ebfd146a | 1184 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1185 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1186 | vector_stmt, stmt_info, | |
1187 | 0, vect_prologue); | |
ebfd146a IR |
1188 | } |
1189 | ||
92345349 BS |
1190 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1191 | stmt_info, 0, vect_body); | |
1192 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1193 | stmt_info, 0, vect_body); | |
8bd37302 | 1194 | |
73fbfcad | 1195 | if (dump_enabled_p ()) |
78c60e3d | 1196 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1197 | "vect_model_load_cost: explicit realign optimized" |
1198 | "\n"); | |
8bd37302 | 1199 | |
ebfd146a IR |
1200 | break; |
1201 | } | |
1202 | ||
38eec4c6 UW |
1203 | case dr_unaligned_unsupported: |
1204 | { | |
1205 | *inside_cost = VECT_MAX_COST; | |
1206 | ||
73fbfcad | 1207 | if (dump_enabled_p ()) |
78c60e3d | 1208 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1209 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1210 | break; |
1211 | } | |
1212 | ||
ebfd146a IR |
1213 | default: |
1214 | gcc_unreachable (); | |
1215 | } | |
ebfd146a IR |
1216 | } |
1217 | ||
418b7df3 RG |
1218 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1219 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1220 | |
418b7df3 | 1221 | static void |
355fe088 | 1222 | vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi) |
ebfd146a | 1223 | { |
ebfd146a | 1224 | if (gsi) |
418b7df3 | 1225 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1226 | else |
1227 | { | |
418b7df3 | 1228 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1229 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1230 | |
a70d6342 IR |
1231 | if (loop_vinfo) |
1232 | { | |
1233 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1234 | basic_block new_bb; |
1235 | edge pe; | |
a70d6342 IR |
1236 | |
1237 | if (nested_in_vect_loop_p (loop, stmt)) | |
1238 | loop = loop->inner; | |
b8698a0f | 1239 | |
a70d6342 | 1240 | pe = loop_preheader_edge (loop); |
418b7df3 | 1241 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1242 | gcc_assert (!new_bb); |
1243 | } | |
1244 | else | |
1245 | { | |
1246 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1247 | basic_block bb; | |
1248 | gimple_stmt_iterator gsi_bb_start; | |
1249 | ||
1250 | gcc_assert (bb_vinfo); | |
1251 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1252 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1253 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1254 | } |
ebfd146a IR |
1255 | } |
1256 | ||
73fbfcad | 1257 | if (dump_enabled_p ()) |
ebfd146a | 1258 | { |
78c60e3d SS |
1259 | dump_printf_loc (MSG_NOTE, vect_location, |
1260 | "created new init_stmt: "); | |
1261 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1262 | } |
418b7df3 RG |
1263 | } |
1264 | ||
1265 | /* Function vect_init_vector. | |
ebfd146a | 1266 | |
5467ee52 RG |
1267 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1268 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1269 | vector type a vector with all elements equal to VAL is created first. | |
1270 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1271 | initialization at the loop preheader. | |
418b7df3 RG |
1272 | Return the DEF of INIT_STMT. |
1273 | It will be used in the vectorization of STMT. */ | |
1274 | ||
1275 | tree | |
355fe088 | 1276 | vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 | 1277 | { |
355fe088 | 1278 | gimple *init_stmt; |
418b7df3 RG |
1279 | tree new_temp; |
1280 | ||
e412ece4 RB |
1281 | /* We abuse this function to push sth to a SSA name with initial 'val'. */ |
1282 | if (! useless_type_conversion_p (type, TREE_TYPE (val))) | |
418b7df3 | 1283 | { |
e412ece4 RB |
1284 | gcc_assert (TREE_CODE (type) == VECTOR_TYPE); |
1285 | if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) | |
418b7df3 | 1286 | { |
5a308cf1 IE |
1287 | /* Scalar boolean value should be transformed into |
1288 | all zeros or all ones value before building a vector. */ | |
1289 | if (VECTOR_BOOLEAN_TYPE_P (type)) | |
1290 | { | |
b3d51f23 IE |
1291 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1292 | tree false_val = build_zero_cst (TREE_TYPE (type)); | |
5a308cf1 IE |
1293 | |
1294 | if (CONSTANT_CLASS_P (val)) | |
1295 | val = integer_zerop (val) ? false_val : true_val; | |
1296 | else | |
1297 | { | |
1298 | new_temp = make_ssa_name (TREE_TYPE (type)); | |
1299 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, | |
1300 | val, true_val, false_val); | |
1301 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
1302 | val = new_temp; | |
1303 | } | |
1304 | } | |
1305 | else if (CONSTANT_CLASS_P (val)) | |
42fd8198 | 1306 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1307 | else |
1308 | { | |
b731b390 | 1309 | new_temp = make_ssa_name (TREE_TYPE (type)); |
e412ece4 RB |
1310 | if (! INTEGRAL_TYPE_P (TREE_TYPE (val))) |
1311 | init_stmt = gimple_build_assign (new_temp, | |
1312 | fold_build1 (VIEW_CONVERT_EXPR, | |
1313 | TREE_TYPE (type), | |
1314 | val)); | |
1315 | else | |
1316 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); | |
418b7df3 | 1317 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1318 | val = new_temp; |
418b7df3 RG |
1319 | } |
1320 | } | |
5467ee52 | 1321 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1322 | } |
1323 | ||
0e22bb5a RB |
1324 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1325 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1326 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1327 | return new_temp; |
ebfd146a IR |
1328 | } |
1329 | ||
c83a894c | 1330 | /* Function vect_get_vec_def_for_operand_1. |
a70d6342 | 1331 | |
c83a894c AH |
1332 | For a defining stmt DEF_STMT of a scalar stmt, return a vector def with type |
1333 | DT that will be used in the vectorized stmt. */ | |
ebfd146a IR |
1334 | |
1335 | tree | |
c83a894c | 1336 | vect_get_vec_def_for_operand_1 (gimple *def_stmt, enum vect_def_type dt) |
ebfd146a IR |
1337 | { |
1338 | tree vec_oprnd; | |
355fe088 | 1339 | gimple *vec_stmt; |
ebfd146a | 1340 | stmt_vec_info def_stmt_info = NULL; |
ebfd146a IR |
1341 | |
1342 | switch (dt) | |
1343 | { | |
81c40241 | 1344 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1345 | case vect_constant_def: |
81c40241 | 1346 | case vect_external_def: |
c83a894c AH |
1347 | /* Code should use vect_get_vec_def_for_operand. */ |
1348 | gcc_unreachable (); | |
ebfd146a | 1349 | |
81c40241 | 1350 | /* operand is defined inside the loop. */ |
8644a673 | 1351 | case vect_internal_def: |
ebfd146a | 1352 | { |
ebfd146a IR |
1353 | /* Get the def from the vectorized stmt. */ |
1354 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1355 | |
ebfd146a | 1356 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1357 | /* Get vectorized pattern statement. */ |
1358 | if (!vec_stmt | |
1359 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1360 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1361 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1362 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1363 | gcc_assert (vec_stmt); |
1364 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1365 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1366 | else if (is_gimple_call (vec_stmt)) | |
1367 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1368 | else | |
1369 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1370 | return vec_oprnd; | |
1371 | } | |
1372 | ||
81c40241 | 1373 | /* operand is defined by a loop header phi - reduction */ |
ebfd146a | 1374 | case vect_reduction_def: |
06066f92 | 1375 | case vect_double_reduction_def: |
7c5222ff | 1376 | case vect_nested_cycle: |
81c40241 RB |
1377 | /* Code should use get_initial_def_for_reduction. */ |
1378 | gcc_unreachable (); | |
ebfd146a | 1379 | |
81c40241 | 1380 | /* operand is defined by loop-header phi - induction. */ |
ebfd146a IR |
1381 | case vect_induction_def: |
1382 | { | |
1383 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1384 | ||
1385 | /* Get the def from the vectorized stmt. */ | |
1386 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1387 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1388 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1389 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1390 | else | |
1391 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1392 | return vec_oprnd; |
1393 | } | |
1394 | ||
1395 | default: | |
1396 | gcc_unreachable (); | |
1397 | } | |
1398 | } | |
1399 | ||
1400 | ||
c83a894c AH |
1401 | /* Function vect_get_vec_def_for_operand. |
1402 | ||
1403 | OP is an operand in STMT. This function returns a (vector) def that will be | |
1404 | used in the vectorized stmt for STMT. | |
1405 | ||
1406 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1407 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1408 | ||
1409 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1410 | needs to be introduced. VECTYPE may be used to specify a required type for | |
1411 | vector invariant. */ | |
1412 | ||
1413 | tree | |
1414 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) | |
1415 | { | |
1416 | gimple *def_stmt; | |
1417 | enum vect_def_type dt; | |
1418 | bool is_simple_use; | |
1419 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
1420 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
1421 | ||
1422 | if (dump_enabled_p ()) | |
1423 | { | |
1424 | dump_printf_loc (MSG_NOTE, vect_location, | |
1425 | "vect_get_vec_def_for_operand: "); | |
1426 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
1427 | dump_printf (MSG_NOTE, "\n"); | |
1428 | } | |
1429 | ||
1430 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt); | |
1431 | gcc_assert (is_simple_use); | |
1432 | if (def_stmt && dump_enabled_p ()) | |
1433 | { | |
1434 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1435 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
1436 | } | |
1437 | ||
1438 | if (dt == vect_constant_def || dt == vect_external_def) | |
1439 | { | |
1440 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
1441 | tree vector_type; | |
1442 | ||
1443 | if (vectype) | |
1444 | vector_type = vectype; | |
1445 | else if (TREE_CODE (TREE_TYPE (op)) == BOOLEAN_TYPE | |
1446 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) | |
1447 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1448 | else | |
1449 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1450 | ||
1451 | gcc_assert (vector_type); | |
1452 | return vect_init_vector (stmt, op, vector_type, NULL); | |
1453 | } | |
1454 | else | |
1455 | return vect_get_vec_def_for_operand_1 (def_stmt, dt); | |
1456 | } | |
1457 | ||
1458 | ||
ebfd146a IR |
1459 | /* Function vect_get_vec_def_for_stmt_copy |
1460 | ||
ff802fa1 | 1461 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1462 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1463 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1464 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1465 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1466 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1467 | DT is the type of the vector def VEC_OPRND. |
1468 | ||
1469 | Context: | |
1470 | In case the vectorization factor (VF) is bigger than the number | |
1471 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1472 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1473 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1474 | smallest data-type determines the VF, and as a result, when vectorizing |
1475 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1476 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1477 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1478 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1479 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1480 | ||
1481 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1482 | |
ebfd146a IR |
1483 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1484 | VS1.1: vx.1 = memref1 VS1.2 | |
1485 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1486 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1487 | |
1488 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1489 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1490 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1491 | VSnew.3: vz3 = vx.3 + ... | |
1492 | ||
1493 | The vectorization of S1 is explained in vectorizable_load. | |
1494 | The vectorization of S2: | |
b8698a0f L |
1495 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1496 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1497 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1498 | returns the vector-def 'vx.0'. |
1499 | ||
b8698a0f L |
1500 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1501 | function is called to get the relevant vector-def for each operand. It is | |
1502 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1503 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1504 | ||
b8698a0f L |
1505 | For example, to obtain the vector-def 'vx.1' in order to create the |
1506 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1507 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1508 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1509 | and return its def ('vx.1'). | |
1510 | Overall, to create the above sequence this function will be called 3 times: | |
1511 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1512 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1513 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1514 | ||
1515 | tree | |
1516 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1517 | { | |
355fe088 | 1518 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1519 | stmt_vec_info def_stmt_info; |
1520 | ||
1521 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1522 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1523 | return vec_oprnd; |
1524 | ||
1525 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1526 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1527 | gcc_assert (def_stmt_info); | |
1528 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1529 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1530 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1531 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1532 | else | |
1533 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1534 | return vec_oprnd; | |
1535 | } | |
1536 | ||
1537 | ||
1538 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1539 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1540 | |
1541 | static void | |
b8698a0f | 1542 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1543 | vec<tree> *vec_oprnds0, |
1544 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1545 | { |
9771b263 | 1546 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1547 | |
1548 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1549 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1550 | |
9771b263 | 1551 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1552 | { |
9771b263 | 1553 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1554 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1555 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1556 | } |
1557 | } | |
1558 | ||
1559 | ||
d092494c IR |
1560 | /* Get vectorized definitions for OP0 and OP1. |
1561 | REDUC_INDEX is the index of reduction operand in case of reduction, | |
1562 | and -1 otherwise. */ | |
ebfd146a | 1563 | |
d092494c | 1564 | void |
355fe088 | 1565 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1566 | vec<tree> *vec_oprnds0, |
1567 | vec<tree> *vec_oprnds1, | |
d092494c | 1568 | slp_tree slp_node, int reduc_index) |
ebfd146a IR |
1569 | { |
1570 | if (slp_node) | |
d092494c IR |
1571 | { |
1572 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1573 | auto_vec<tree> ops (nops); |
1574 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1575 | |
9771b263 | 1576 | ops.quick_push (op0); |
d092494c | 1577 | if (op1) |
9771b263 | 1578 | ops.quick_push (op1); |
d092494c IR |
1579 | |
1580 | vect_get_slp_defs (ops, slp_node, &vec_defs, reduc_index); | |
1581 | ||
37b5ec8f | 1582 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1583 | if (op1) |
37b5ec8f | 1584 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1585 | } |
ebfd146a IR |
1586 | else |
1587 | { | |
1588 | tree vec_oprnd; | |
1589 | ||
9771b263 | 1590 | vec_oprnds0->create (1); |
81c40241 | 1591 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1592 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1593 | |
1594 | if (op1) | |
1595 | { | |
9771b263 | 1596 | vec_oprnds1->create (1); |
81c40241 | 1597 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1598 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1599 | } |
1600 | } | |
1601 | } | |
1602 | ||
1603 | ||
1604 | /* Function vect_finish_stmt_generation. | |
1605 | ||
1606 | Insert a new stmt. */ | |
1607 | ||
1608 | void | |
355fe088 | 1609 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1610 | gimple_stmt_iterator *gsi) |
1611 | { | |
1612 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
310213d4 | 1613 | vec_info *vinfo = stmt_info->vinfo; |
ebfd146a IR |
1614 | |
1615 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1616 | ||
54e8e2c3 RG |
1617 | if (!gsi_end_p (*gsi) |
1618 | && gimple_has_mem_ops (vec_stmt)) | |
1619 | { | |
355fe088 | 1620 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1621 | tree vuse = gimple_vuse (at_stmt); |
1622 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1623 | { | |
1624 | tree vdef = gimple_vdef (at_stmt); | |
1625 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1626 | /* If we have an SSA vuse and insert a store, update virtual | |
1627 | SSA form to avoid triggering the renamer. Do so only | |
1628 | if we can easily see all uses - which is what almost always | |
1629 | happens with the way vectorized stmts are inserted. */ | |
1630 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1631 | && ((is_gimple_assign (vec_stmt) | |
1632 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1633 | || (is_gimple_call (vec_stmt) | |
1634 | && !(gimple_call_flags (vec_stmt) | |
1635 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1636 | { | |
1637 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1638 | gimple_set_vdef (vec_stmt, new_vdef); | |
1639 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1640 | } | |
1641 | } | |
1642 | } | |
ebfd146a IR |
1643 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1644 | ||
310213d4 | 1645 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, vinfo)); |
ebfd146a | 1646 | |
73fbfcad | 1647 | if (dump_enabled_p ()) |
ebfd146a | 1648 | { |
78c60e3d SS |
1649 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1650 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
ebfd146a IR |
1651 | } |
1652 | ||
ad885386 | 1653 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
8e91d222 JJ |
1654 | |
1655 | /* While EH edges will generally prevent vectorization, stmt might | |
1656 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1657 | that could throw are part of the same region. */ | |
1658 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1659 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1660 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
ebfd146a IR |
1661 | } |
1662 | ||
70439f0d RS |
1663 | /* We want to vectorize a call to combined function CFN with function |
1664 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN | |
1665 | as the types of all inputs. Check whether this is possible using | |
1666 | an internal function, returning its code if so or IFN_LAST if not. */ | |
ebfd146a | 1667 | |
70439f0d RS |
1668 | static internal_fn |
1669 | vectorizable_internal_function (combined_fn cfn, tree fndecl, | |
1670 | tree vectype_out, tree vectype_in) | |
ebfd146a | 1671 | { |
70439f0d RS |
1672 | internal_fn ifn; |
1673 | if (internal_fn_p (cfn)) | |
1674 | ifn = as_internal_fn (cfn); | |
1675 | else | |
1676 | ifn = associated_internal_fn (fndecl); | |
1677 | if (ifn != IFN_LAST && direct_internal_fn_p (ifn)) | |
1678 | { | |
1679 | const direct_internal_fn_info &info = direct_internal_fn (ifn); | |
1680 | if (info.vectorizable) | |
1681 | { | |
1682 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); | |
1683 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); | |
d95ab70a RS |
1684 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1685 | OPTIMIZE_FOR_SPEED)) | |
70439f0d RS |
1686 | return ifn; |
1687 | } | |
1688 | } | |
1689 | return IFN_LAST; | |
ebfd146a IR |
1690 | } |
1691 | ||
5ce9450f | 1692 | |
355fe088 | 1693 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1694 | gimple_stmt_iterator *); |
1695 | ||
1696 | ||
1697 | /* Function vectorizable_mask_load_store. | |
1698 | ||
1699 | Check if STMT performs a conditional load or store that can be vectorized. | |
1700 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1701 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
1702 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1703 | ||
1704 | static bool | |
355fe088 TS |
1705 | vectorizable_mask_load_store (gimple *stmt, gimple_stmt_iterator *gsi, |
1706 | gimple **vec_stmt, slp_tree slp_node) | |
5ce9450f JJ |
1707 | { |
1708 | tree vec_dest = NULL; | |
1709 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1710 | stmt_vec_info prev_stmt_info; | |
1711 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1712 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1713 | bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
1714 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
1715 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
57e2f6ad | 1716 | tree rhs_vectype = NULL_TREE; |
045c1278 | 1717 | tree mask_vectype; |
5ce9450f | 1718 | tree elem_type; |
355fe088 | 1719 | gimple *new_stmt; |
5ce9450f JJ |
1720 | tree dummy; |
1721 | tree dataref_ptr = NULL_TREE; | |
355fe088 | 1722 | gimple *ptr_incr; |
5ce9450f JJ |
1723 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
1724 | int ncopies; | |
1725 | int i, j; | |
1726 | bool inv_p; | |
1727 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; | |
1728 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
1729 | int gather_scale = 1; | |
1730 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
1731 | bool is_store; | |
1732 | tree mask; | |
355fe088 | 1733 | gimple *def_stmt; |
5ce9450f JJ |
1734 | enum vect_def_type dt; |
1735 | ||
1736 | if (slp_node != NULL) | |
1737 | return false; | |
1738 | ||
1739 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
1740 | gcc_assert (ncopies >= 1); | |
1741 | ||
1742 | is_store = gimple_call_internal_fn (stmt) == IFN_MASK_STORE; | |
1743 | mask = gimple_call_arg (stmt, 2); | |
045c1278 IE |
1744 | |
1745 | if (TREE_CODE (TREE_TYPE (mask)) != BOOLEAN_TYPE) | |
5ce9450f JJ |
1746 | return false; |
1747 | ||
1748 | /* FORNOW. This restriction should be relaxed. */ | |
1749 | if (nested_in_vect_loop && ncopies > 1) | |
1750 | { | |
1751 | if (dump_enabled_p ()) | |
1752 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1753 | "multiple types in nested loop."); | |
1754 | return false; | |
1755 | } | |
1756 | ||
1757 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
1758 | return false; | |
1759 | ||
66c16fd9 RB |
1760 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
1761 | && ! vec_stmt) | |
5ce9450f JJ |
1762 | return false; |
1763 | ||
1764 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
1765 | return false; | |
1766 | ||
1767 | elem_type = TREE_TYPE (vectype); | |
1768 | ||
1769 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1770 | return false; | |
1771 | ||
f2e2a985 | 1772 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
5ce9450f JJ |
1773 | return false; |
1774 | ||
045c1278 IE |
1775 | if (TREE_CODE (mask) != SSA_NAME) |
1776 | return false; | |
1777 | ||
1778 | if (!vect_is_simple_use (mask, loop_vinfo, &def_stmt, &dt, &mask_vectype)) | |
1779 | return false; | |
1780 | ||
1781 | if (!mask_vectype) | |
1782 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
1783 | ||
dc6a3147 IE |
1784 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype) |
1785 | || TYPE_VECTOR_SUBPARTS (mask_vectype) != TYPE_VECTOR_SUBPARTS (vectype)) | |
045c1278 IE |
1786 | return false; |
1787 | ||
57e2f6ad IE |
1788 | if (is_store) |
1789 | { | |
1790 | tree rhs = gimple_call_arg (stmt, 3); | |
1791 | if (!vect_is_simple_use (rhs, loop_vinfo, &def_stmt, &dt, &rhs_vectype)) | |
1792 | return false; | |
1793 | } | |
1794 | ||
3bab6342 | 1795 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5ce9450f | 1796 | { |
355fe088 | 1797 | gimple *def_stmt; |
3bab6342 | 1798 | gather_decl = vect_check_gather_scatter (stmt, loop_vinfo, &gather_base, |
5ce9450f JJ |
1799 | &gather_off, &gather_scale); |
1800 | gcc_assert (gather_decl); | |
81c40241 RB |
1801 | if (!vect_is_simple_use (gather_off, loop_vinfo, &def_stmt, &gather_dt, |
1802 | &gather_off_vectype)) | |
5ce9450f JJ |
1803 | { |
1804 | if (dump_enabled_p ()) | |
1805 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1806 | "gather index use not simple."); | |
1807 | return false; | |
1808 | } | |
03b9e8e4 JJ |
1809 | |
1810 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1811 | tree masktype | |
1812 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
1813 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
1814 | { | |
1815 | if (dump_enabled_p ()) | |
1816 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1817 | "masked gather with integer mask not supported."); | |
1818 | return false; | |
1819 | } | |
5ce9450f JJ |
1820 | } |
1821 | else if (tree_int_cst_compare (nested_in_vect_loop | |
1822 | ? STMT_VINFO_DR_STEP (stmt_info) | |
1823 | : DR_STEP (dr), size_zero_node) <= 0) | |
1824 | return false; | |
1825 | else if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
045c1278 IE |
1826 | || !can_vec_mask_load_store_p (TYPE_MODE (vectype), |
1827 | TYPE_MODE (mask_vectype), | |
57e2f6ad IE |
1828 | !is_store) |
1829 | || (rhs_vectype | |
1830 | && !useless_type_conversion_p (vectype, rhs_vectype))) | |
5ce9450f JJ |
1831 | return false; |
1832 | ||
5ce9450f JJ |
1833 | if (!vec_stmt) /* transformation not required. */ |
1834 | { | |
1835 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1836 | if (is_store) | |
1837 | vect_model_store_cost (stmt_info, ncopies, false, dt, | |
1838 | NULL, NULL, NULL); | |
1839 | else | |
1840 | vect_model_load_cost (stmt_info, ncopies, false, NULL, NULL, NULL); | |
1841 | return true; | |
1842 | } | |
1843 | ||
1844 | /** Transform. **/ | |
1845 | ||
3bab6342 | 1846 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5ce9450f JJ |
1847 | { |
1848 | tree vec_oprnd0 = NULL_TREE, op; | |
1849 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1850 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
acdcd61b | 1851 | tree ptr, vec_mask = NULL_TREE, mask_op = NULL_TREE, var, scale; |
5ce9450f | 1852 | tree perm_mask = NULL_TREE, prev_res = NULL_TREE; |
acdcd61b | 1853 | tree mask_perm_mask = NULL_TREE; |
5ce9450f JJ |
1854 | edge pe = loop_preheader_edge (loop); |
1855 | gimple_seq seq; | |
1856 | basic_block new_bb; | |
1857 | enum { NARROW, NONE, WIDEN } modifier; | |
1858 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
1859 | ||
acdcd61b JJ |
1860 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); |
1861 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1862 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1863 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1864 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1865 | scaletype = TREE_VALUE (arglist); | |
1866 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
1867 | && types_compatible_p (srctype, masktype)); | |
1868 | ||
5ce9450f JJ |
1869 | if (nunits == gather_off_nunits) |
1870 | modifier = NONE; | |
1871 | else if (nunits == gather_off_nunits / 2) | |
1872 | { | |
1873 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
1874 | modifier = WIDEN; | |
1875 | ||
1876 | for (i = 0; i < gather_off_nunits; ++i) | |
1877 | sel[i] = i | nunits; | |
1878 | ||
557be5a8 | 1879 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
5ce9450f JJ |
1880 | } |
1881 | else if (nunits == gather_off_nunits * 2) | |
1882 | { | |
1883 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
1884 | modifier = NARROW; | |
1885 | ||
1886 | for (i = 0; i < nunits; ++i) | |
1887 | sel[i] = i < gather_off_nunits | |
1888 | ? i : i + nunits - gather_off_nunits; | |
1889 | ||
557be5a8 | 1890 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
5ce9450f | 1891 | ncopies *= 2; |
acdcd61b JJ |
1892 | for (i = 0; i < nunits; ++i) |
1893 | sel[i] = i | gather_off_nunits; | |
557be5a8 | 1894 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, sel); |
5ce9450f JJ |
1895 | } |
1896 | else | |
1897 | gcc_unreachable (); | |
1898 | ||
5ce9450f JJ |
1899 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); |
1900 | ||
1901 | ptr = fold_convert (ptrtype, gather_base); | |
1902 | if (!is_gimple_min_invariant (ptr)) | |
1903 | { | |
1904 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
1905 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
1906 | gcc_assert (!new_bb); | |
1907 | } | |
1908 | ||
1909 | scale = build_int_cst (scaletype, gather_scale); | |
1910 | ||
1911 | prev_stmt_info = NULL; | |
1912 | for (j = 0; j < ncopies; ++j) | |
1913 | { | |
1914 | if (modifier == WIDEN && (j & 1)) | |
1915 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
1916 | perm_mask, stmt, gsi); | |
1917 | else if (j == 0) | |
1918 | op = vec_oprnd0 | |
81c40241 | 1919 | = vect_get_vec_def_for_operand (gather_off, stmt); |
5ce9450f JJ |
1920 | else |
1921 | op = vec_oprnd0 | |
1922 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
1923 | ||
1924 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
1925 | { | |
1926 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
1927 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 1928 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
5ce9450f JJ |
1929 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
1930 | new_stmt | |
0d0e4a03 | 1931 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
1932 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1933 | op = var; | |
1934 | } | |
1935 | ||
acdcd61b JJ |
1936 | if (mask_perm_mask && (j & 1)) |
1937 | mask_op = permute_vec_elements (mask_op, mask_op, | |
1938 | mask_perm_mask, stmt, gsi); | |
5ce9450f JJ |
1939 | else |
1940 | { | |
acdcd61b | 1941 | if (j == 0) |
81c40241 | 1942 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
acdcd61b JJ |
1943 | else |
1944 | { | |
81c40241 | 1945 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
acdcd61b JJ |
1946 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
1947 | } | |
5ce9450f | 1948 | |
acdcd61b JJ |
1949 | mask_op = vec_mask; |
1950 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
1951 | { | |
1952 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)) | |
1953 | == TYPE_VECTOR_SUBPARTS (masktype)); | |
0e22bb5a | 1954 | var = vect_get_new_ssa_name (masktype, vect_simple_var); |
acdcd61b JJ |
1955 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); |
1956 | new_stmt | |
0d0e4a03 | 1957 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_op); |
acdcd61b JJ |
1958 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1959 | mask_op = var; | |
1960 | } | |
5ce9450f JJ |
1961 | } |
1962 | ||
1963 | new_stmt | |
1964 | = gimple_build_call (gather_decl, 5, mask_op, ptr, op, mask_op, | |
1965 | scale); | |
1966 | ||
1967 | if (!useless_type_conversion_p (vectype, rettype)) | |
1968 | { | |
1969 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
1970 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 1971 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
5ce9450f JJ |
1972 | gimple_call_set_lhs (new_stmt, op); |
1973 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 1974 | var = make_ssa_name (vec_dest); |
5ce9450f | 1975 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
0d0e4a03 | 1976 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
1977 | } |
1978 | else | |
1979 | { | |
1980 | var = make_ssa_name (vec_dest, new_stmt); | |
1981 | gimple_call_set_lhs (new_stmt, var); | |
1982 | } | |
1983 | ||
1984 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1985 | ||
1986 | if (modifier == NARROW) | |
1987 | { | |
1988 | if ((j & 1) == 0) | |
1989 | { | |
1990 | prev_res = var; | |
1991 | continue; | |
1992 | } | |
1993 | var = permute_vec_elements (prev_res, var, | |
1994 | perm_mask, stmt, gsi); | |
1995 | new_stmt = SSA_NAME_DEF_STMT (var); | |
1996 | } | |
1997 | ||
1998 | if (prev_stmt_info == NULL) | |
1999 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2000 | else | |
2001 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2002 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2003 | } | |
3efe2e2c JJ |
2004 | |
2005 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2006 | from the IL. */ | |
e6f5c25d IE |
2007 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
2008 | { | |
2009 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
2010 | stmt_info = vinfo_for_stmt (stmt); | |
2011 | } | |
3efe2e2c JJ |
2012 | tree lhs = gimple_call_lhs (stmt); |
2013 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2014 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2015 | set_vinfo_for_stmt (stmt, NULL); | |
2016 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2017 | gsi_replace (gsi, new_stmt, true); | |
5ce9450f JJ |
2018 | return true; |
2019 | } | |
2020 | else if (is_store) | |
2021 | { | |
2022 | tree vec_rhs = NULL_TREE, vec_mask = NULL_TREE; | |
2023 | prev_stmt_info = NULL; | |
2d4dc223 | 2024 | LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true; |
5ce9450f JJ |
2025 | for (i = 0; i < ncopies; i++) |
2026 | { | |
2027 | unsigned align, misalign; | |
2028 | ||
2029 | if (i == 0) | |
2030 | { | |
2031 | tree rhs = gimple_call_arg (stmt, 3); | |
81c40241 RB |
2032 | vec_rhs = vect_get_vec_def_for_operand (rhs, stmt); |
2033 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); | |
5ce9450f JJ |
2034 | /* We should have catched mismatched types earlier. */ |
2035 | gcc_assert (useless_type_conversion_p (vectype, | |
2036 | TREE_TYPE (vec_rhs))); | |
2037 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, | |
2038 | NULL_TREE, &dummy, gsi, | |
2039 | &ptr_incr, false, &inv_p); | |
2040 | gcc_assert (!inv_p); | |
2041 | } | |
2042 | else | |
2043 | { | |
81c40241 | 2044 | vect_is_simple_use (vec_rhs, loop_vinfo, &def_stmt, &dt); |
5ce9450f | 2045 | vec_rhs = vect_get_vec_def_for_stmt_copy (dt, vec_rhs); |
81c40241 | 2046 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2047 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2048 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2049 | TYPE_SIZE_UNIT (vectype)); | |
2050 | } | |
2051 | ||
2052 | align = TYPE_ALIGN_UNIT (vectype); | |
2053 | if (aligned_access_p (dr)) | |
2054 | misalign = 0; | |
2055 | else if (DR_MISALIGNMENT (dr) == -1) | |
2056 | { | |
2057 | align = TYPE_ALIGN_UNIT (elem_type); | |
2058 | misalign = 0; | |
2059 | } | |
2060 | else | |
2061 | misalign = DR_MISALIGNMENT (dr); | |
2062 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2063 | misalign); | |
08554c26 JJ |
2064 | tree ptr = build_int_cst (TREE_TYPE (gimple_call_arg (stmt, 1)), |
2065 | misalign ? misalign & -misalign : align); | |
5ce9450f JJ |
2066 | new_stmt |
2067 | = gimple_build_call_internal (IFN_MASK_STORE, 4, dataref_ptr, | |
08554c26 | 2068 | ptr, vec_mask, vec_rhs); |
5ce9450f JJ |
2069 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2070 | if (i == 0) | |
2071 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2072 | else | |
2073 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2074 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2075 | } | |
2076 | } | |
2077 | else | |
2078 | { | |
2079 | tree vec_mask = NULL_TREE; | |
2080 | prev_stmt_info = NULL; | |
2081 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); | |
2082 | for (i = 0; i < ncopies; i++) | |
2083 | { | |
2084 | unsigned align, misalign; | |
2085 | ||
2086 | if (i == 0) | |
2087 | { | |
81c40241 | 2088 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
5ce9450f JJ |
2089 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, |
2090 | NULL_TREE, &dummy, gsi, | |
2091 | &ptr_incr, false, &inv_p); | |
2092 | gcc_assert (!inv_p); | |
2093 | } | |
2094 | else | |
2095 | { | |
81c40241 | 2096 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2097 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2098 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2099 | TYPE_SIZE_UNIT (vectype)); | |
2100 | } | |
2101 | ||
2102 | align = TYPE_ALIGN_UNIT (vectype); | |
2103 | if (aligned_access_p (dr)) | |
2104 | misalign = 0; | |
2105 | else if (DR_MISALIGNMENT (dr) == -1) | |
2106 | { | |
2107 | align = TYPE_ALIGN_UNIT (elem_type); | |
2108 | misalign = 0; | |
2109 | } | |
2110 | else | |
2111 | misalign = DR_MISALIGNMENT (dr); | |
2112 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2113 | misalign); | |
08554c26 JJ |
2114 | tree ptr = build_int_cst (TREE_TYPE (gimple_call_arg (stmt, 1)), |
2115 | misalign ? misalign & -misalign : align); | |
5ce9450f JJ |
2116 | new_stmt |
2117 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, dataref_ptr, | |
08554c26 | 2118 | ptr, vec_mask); |
b731b390 | 2119 | gimple_call_set_lhs (new_stmt, make_ssa_name (vec_dest)); |
5ce9450f JJ |
2120 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2121 | if (i == 0) | |
2122 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2123 | else | |
2124 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2125 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2126 | } | |
2127 | } | |
2128 | ||
3efe2e2c JJ |
2129 | if (!is_store) |
2130 | { | |
2131 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2132 | from the IL. */ | |
e6f5c25d IE |
2133 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
2134 | { | |
2135 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
2136 | stmt_info = vinfo_for_stmt (stmt); | |
2137 | } | |
3efe2e2c JJ |
2138 | tree lhs = gimple_call_lhs (stmt); |
2139 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2140 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2141 | set_vinfo_for_stmt (stmt, NULL); | |
2142 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2143 | gsi_replace (gsi, new_stmt, true); | |
2144 | } | |
2145 | ||
5ce9450f JJ |
2146 | return true; |
2147 | } | |
2148 | ||
b1b6836e RS |
2149 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
2150 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT | |
2151 | in a single step. On success, store the binary pack code in | |
2152 | *CONVERT_CODE. */ | |
2153 | ||
2154 | static bool | |
2155 | simple_integer_narrowing (tree vectype_out, tree vectype_in, | |
2156 | tree_code *convert_code) | |
2157 | { | |
2158 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) | |
2159 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) | |
2160 | return false; | |
2161 | ||
2162 | tree_code code; | |
2163 | int multi_step_cvt = 0; | |
2164 | auto_vec <tree, 8> interm_types; | |
2165 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, | |
2166 | &code, &multi_step_cvt, | |
2167 | &interm_types) | |
2168 | || multi_step_cvt) | |
2169 | return false; | |
2170 | ||
2171 | *convert_code = code; | |
2172 | return true; | |
2173 | } | |
5ce9450f | 2174 | |
ebfd146a IR |
2175 | /* Function vectorizable_call. |
2176 | ||
538dd0b7 | 2177 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 2178 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2179 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2180 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2181 | ||
2182 | static bool | |
355fe088 | 2183 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
190c2236 | 2184 | slp_tree slp_node) |
ebfd146a | 2185 | { |
538dd0b7 | 2186 | gcall *stmt; |
ebfd146a IR |
2187 | tree vec_dest; |
2188 | tree scalar_dest; | |
2189 | tree op, type; | |
2190 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
538dd0b7 | 2191 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a IR |
2192 | tree vectype_out, vectype_in; |
2193 | int nunits_in; | |
2194 | int nunits_out; | |
2195 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 2196 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 2197 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 2198 | tree fndecl, new_temp, rhs_type; |
355fe088 | 2199 | gimple *def_stmt; |
0502fb85 UB |
2200 | enum vect_def_type dt[3] |
2201 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
355fe088 | 2202 | gimple *new_stmt = NULL; |
ebfd146a | 2203 | int ncopies, j; |
6e1aa848 | 2204 | vec<tree> vargs = vNULL; |
ebfd146a IR |
2205 | enum { NARROW, NONE, WIDEN } modifier; |
2206 | size_t i, nargs; | |
9d5e7640 | 2207 | tree lhs; |
ebfd146a | 2208 | |
190c2236 | 2209 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2210 | return false; |
2211 | ||
66c16fd9 RB |
2212 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
2213 | && ! vec_stmt) | |
ebfd146a IR |
2214 | return false; |
2215 | ||
538dd0b7 DM |
2216 | /* Is GS a vectorizable call? */ |
2217 | stmt = dyn_cast <gcall *> (gs); | |
2218 | if (!stmt) | |
ebfd146a IR |
2219 | return false; |
2220 | ||
5ce9450f JJ |
2221 | if (gimple_call_internal_p (stmt) |
2222 | && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD | |
2223 | || gimple_call_internal_fn (stmt) == IFN_MASK_STORE)) | |
2224 | return vectorizable_mask_load_store (stmt, gsi, vec_stmt, | |
2225 | slp_node); | |
2226 | ||
0136f8f0 AH |
2227 | if (gimple_call_lhs (stmt) == NULL_TREE |
2228 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
2229 | return false; |
2230 | ||
0136f8f0 | 2231 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 2232 | |
b690cc0f RG |
2233 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2234 | ||
ebfd146a IR |
2235 | /* Process function arguments. */ |
2236 | rhs_type = NULL_TREE; | |
b690cc0f | 2237 | vectype_in = NULL_TREE; |
ebfd146a IR |
2238 | nargs = gimple_call_num_args (stmt); |
2239 | ||
1b1562a5 MM |
2240 | /* Bail out if the function has more than three arguments, we do not have |
2241 | interesting builtin functions to vectorize with more than two arguments | |
2242 | except for fma. No arguments is also not good. */ | |
2243 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
2244 | return false; |
2245 | ||
74bf76ed JJ |
2246 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2247 | if (gimple_call_internal_p (stmt) | |
2248 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2249 | { | |
2250 | nargs = 0; | |
2251 | rhs_type = unsigned_type_node; | |
2252 | } | |
2253 | ||
ebfd146a IR |
2254 | for (i = 0; i < nargs; i++) |
2255 | { | |
b690cc0f RG |
2256 | tree opvectype; |
2257 | ||
ebfd146a IR |
2258 | op = gimple_call_arg (stmt, i); |
2259 | ||
2260 | /* We can only handle calls with arguments of the same type. */ | |
2261 | if (rhs_type | |
8533c9d8 | 2262 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 2263 | { |
73fbfcad | 2264 | if (dump_enabled_p ()) |
78c60e3d | 2265 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2266 | "argument types differ.\n"); |
ebfd146a IR |
2267 | return false; |
2268 | } | |
b690cc0f RG |
2269 | if (!rhs_type) |
2270 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 2271 | |
81c40241 | 2272 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[i], &opvectype)) |
ebfd146a | 2273 | { |
73fbfcad | 2274 | if (dump_enabled_p ()) |
78c60e3d | 2275 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2276 | "use not simple.\n"); |
ebfd146a IR |
2277 | return false; |
2278 | } | |
ebfd146a | 2279 | |
b690cc0f RG |
2280 | if (!vectype_in) |
2281 | vectype_in = opvectype; | |
2282 | else if (opvectype | |
2283 | && opvectype != vectype_in) | |
2284 | { | |
73fbfcad | 2285 | if (dump_enabled_p ()) |
78c60e3d | 2286 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2287 | "argument vector types differ.\n"); |
b690cc0f RG |
2288 | return false; |
2289 | } | |
2290 | } | |
2291 | /* If all arguments are external or constant defs use a vector type with | |
2292 | the same size as the output vector type. */ | |
ebfd146a | 2293 | if (!vectype_in) |
b690cc0f | 2294 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2295 | if (vec_stmt) |
2296 | gcc_assert (vectype_in); | |
2297 | if (!vectype_in) | |
2298 | { | |
73fbfcad | 2299 | if (dump_enabled_p ()) |
7d8930a0 | 2300 | { |
78c60e3d SS |
2301 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2302 | "no vectype for scalar type "); | |
2303 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2304 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
2305 | } |
2306 | ||
2307 | return false; | |
2308 | } | |
ebfd146a IR |
2309 | |
2310 | /* FORNOW */ | |
b690cc0f RG |
2311 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2312 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
2313 | if (nunits_in == nunits_out / 2) |
2314 | modifier = NARROW; | |
2315 | else if (nunits_out == nunits_in) | |
2316 | modifier = NONE; | |
2317 | else if (nunits_out == nunits_in / 2) | |
2318 | modifier = WIDEN; | |
2319 | else | |
2320 | return false; | |
2321 | ||
70439f0d RS |
2322 | /* We only handle functions that do not read or clobber memory. */ |
2323 | if (gimple_vuse (stmt)) | |
2324 | { | |
2325 | if (dump_enabled_p ()) | |
2326 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2327 | "function reads from or writes to memory.\n"); | |
2328 | return false; | |
2329 | } | |
2330 | ||
ebfd146a IR |
2331 | /* For now, we only vectorize functions if a target specific builtin |
2332 | is available. TODO -- in some cases, it might be profitable to | |
2333 | insert the calls for pieces of the vector, in order to be able | |
2334 | to vectorize other operations in the loop. */ | |
70439f0d RS |
2335 | fndecl = NULL_TREE; |
2336 | internal_fn ifn = IFN_LAST; | |
2337 | combined_fn cfn = gimple_call_combined_fn (stmt); | |
2338 | tree callee = gimple_call_fndecl (stmt); | |
2339 | ||
2340 | /* First try using an internal function. */ | |
b1b6836e RS |
2341 | tree_code convert_code = ERROR_MARK; |
2342 | if (cfn != CFN_LAST | |
2343 | && (modifier == NONE | |
2344 | || (modifier == NARROW | |
2345 | && simple_integer_narrowing (vectype_out, vectype_in, | |
2346 | &convert_code)))) | |
70439f0d RS |
2347 | ifn = vectorizable_internal_function (cfn, callee, vectype_out, |
2348 | vectype_in); | |
2349 | ||
2350 | /* If that fails, try asking for a target-specific built-in function. */ | |
2351 | if (ifn == IFN_LAST) | |
2352 | { | |
2353 | if (cfn != CFN_LAST) | |
2354 | fndecl = targetm.vectorize.builtin_vectorized_function | |
2355 | (cfn, vectype_out, vectype_in); | |
2356 | else | |
2357 | fndecl = targetm.vectorize.builtin_md_vectorized_function | |
2358 | (callee, vectype_out, vectype_in); | |
2359 | } | |
2360 | ||
2361 | if (ifn == IFN_LAST && !fndecl) | |
ebfd146a | 2362 | { |
70439f0d | 2363 | if (cfn == CFN_GOMP_SIMD_LANE |
74bf76ed JJ |
2364 | && !slp_node |
2365 | && loop_vinfo | |
2366 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2367 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
2368 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2369 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
2370 | { | |
2371 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
2372 | { 0, 1, 2, ... vf - 1 } vector. */ | |
2373 | gcc_assert (nargs == 0); | |
2374 | } | |
2375 | else | |
2376 | { | |
2377 | if (dump_enabled_p ()) | |
2378 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 2379 | "function is not vectorizable.\n"); |
74bf76ed JJ |
2380 | return false; |
2381 | } | |
ebfd146a IR |
2382 | } |
2383 | ||
fce57248 | 2384 | if (slp_node) |
190c2236 | 2385 | ncopies = 1; |
b1b6836e | 2386 | else if (modifier == NARROW && ifn == IFN_LAST) |
ebfd146a IR |
2387 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
2388 | else | |
2389 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2390 | ||
2391 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2392 | needs to be generated. */ | |
2393 | gcc_assert (ncopies >= 1); | |
2394 | ||
2395 | if (!vec_stmt) /* transformation not required. */ | |
2396 | { | |
2397 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 2398 | if (dump_enabled_p ()) |
e645e942 TJ |
2399 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
2400 | "\n"); | |
c3e7ee41 | 2401 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
b1b6836e RS |
2402 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) |
2403 | add_stmt_cost (stmt_info->vinfo->target_cost_data, ncopies / 2, | |
2404 | vec_promote_demote, stmt_info, 0, vect_body); | |
2405 | ||
ebfd146a IR |
2406 | return true; |
2407 | } | |
2408 | ||
2409 | /** Transform. **/ | |
2410 | ||
73fbfcad | 2411 | if (dump_enabled_p ()) |
e645e942 | 2412 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
2413 | |
2414 | /* Handle def. */ | |
2415 | scalar_dest = gimple_call_lhs (stmt); | |
2416 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2417 | ||
2418 | prev_stmt_info = NULL; | |
b1b6836e | 2419 | if (modifier == NONE || ifn != IFN_LAST) |
ebfd146a | 2420 | { |
b1b6836e | 2421 | tree prev_res = NULL_TREE; |
ebfd146a IR |
2422 | for (j = 0; j < ncopies; ++j) |
2423 | { | |
2424 | /* Build argument list for the vectorized call. */ | |
2425 | if (j == 0) | |
9771b263 | 2426 | vargs.create (nargs); |
ebfd146a | 2427 | else |
9771b263 | 2428 | vargs.truncate (0); |
ebfd146a | 2429 | |
190c2236 JJ |
2430 | if (slp_node) |
2431 | { | |
ef062b13 | 2432 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2433 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2434 | |
2435 | for (i = 0; i < nargs; i++) | |
9771b263 | 2436 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2437 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2438 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2439 | |
2440 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2441 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
2442 | { |
2443 | size_t k; | |
2444 | for (k = 0; k < nargs; k++) | |
2445 | { | |
37b5ec8f | 2446 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 2447 | vargs[k] = vec_oprndsk[i]; |
190c2236 | 2448 | } |
b1b6836e RS |
2449 | if (modifier == NARROW) |
2450 | { | |
2451 | tree half_res = make_ssa_name (vectype_in); | |
2452 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2453 | gimple_call_set_lhs (new_stmt, half_res); | |
2454 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2455 | if ((i & 1) == 0) | |
2456 | { | |
2457 | prev_res = half_res; | |
2458 | continue; | |
2459 | } | |
2460 | new_temp = make_ssa_name (vec_dest); | |
2461 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
2462 | prev_res, half_res); | |
2463 | } | |
70439f0d | 2464 | else |
b1b6836e RS |
2465 | { |
2466 | if (ifn != IFN_LAST) | |
2467 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2468 | else | |
2469 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2470 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2471 | gimple_call_set_lhs (new_stmt, new_temp); | |
2472 | } | |
190c2236 | 2473 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 2474 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2475 | } |
2476 | ||
2477 | for (i = 0; i < nargs; i++) | |
2478 | { | |
37b5ec8f | 2479 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2480 | vec_oprndsi.release (); |
190c2236 | 2481 | } |
190c2236 JJ |
2482 | continue; |
2483 | } | |
2484 | ||
ebfd146a IR |
2485 | for (i = 0; i < nargs; i++) |
2486 | { | |
2487 | op = gimple_call_arg (stmt, i); | |
2488 | if (j == 0) | |
2489 | vec_oprnd0 | |
81c40241 | 2490 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2491 | else |
63827fb8 IR |
2492 | { |
2493 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
2494 | vec_oprnd0 | |
2495 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
2496 | } | |
ebfd146a | 2497 | |
9771b263 | 2498 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
2499 | } |
2500 | ||
74bf76ed JJ |
2501 | if (gimple_call_internal_p (stmt) |
2502 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2503 | { | |
2504 | tree *v = XALLOCAVEC (tree, nunits_out); | |
2505 | int k; | |
2506 | for (k = 0; k < nunits_out; ++k) | |
2507 | v[k] = build_int_cst (unsigned_type_node, j * nunits_out + k); | |
2508 | tree cst = build_vector (vectype_out, v); | |
2509 | tree new_var | |
0e22bb5a | 2510 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 2511 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 2512 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 2513 | new_temp = make_ssa_name (vec_dest); |
0e22bb5a | 2514 | new_stmt = gimple_build_assign (new_temp, new_var); |
74bf76ed | 2515 | } |
b1b6836e RS |
2516 | else if (modifier == NARROW) |
2517 | { | |
2518 | tree half_res = make_ssa_name (vectype_in); | |
2519 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2520 | gimple_call_set_lhs (new_stmt, half_res); | |
2521 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2522 | if ((j & 1) == 0) | |
2523 | { | |
2524 | prev_res = half_res; | |
2525 | continue; | |
2526 | } | |
2527 | new_temp = make_ssa_name (vec_dest); | |
2528 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
2529 | prev_res, half_res); | |
2530 | } | |
74bf76ed JJ |
2531 | else |
2532 | { | |
70439f0d RS |
2533 | if (ifn != IFN_LAST) |
2534 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2535 | else | |
2536 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
74bf76ed JJ |
2537 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2538 | gimple_call_set_lhs (new_stmt, new_temp); | |
2539 | } | |
ebfd146a IR |
2540 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2541 | ||
b1b6836e | 2542 | if (j == (modifier == NARROW ? 1 : 0)) |
ebfd146a IR |
2543 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; |
2544 | else | |
2545 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2546 | ||
2547 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2548 | } | |
b1b6836e RS |
2549 | } |
2550 | else if (modifier == NARROW) | |
2551 | { | |
ebfd146a IR |
2552 | for (j = 0; j < ncopies; ++j) |
2553 | { | |
2554 | /* Build argument list for the vectorized call. */ | |
2555 | if (j == 0) | |
9771b263 | 2556 | vargs.create (nargs * 2); |
ebfd146a | 2557 | else |
9771b263 | 2558 | vargs.truncate (0); |
ebfd146a | 2559 | |
190c2236 JJ |
2560 | if (slp_node) |
2561 | { | |
ef062b13 | 2562 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2563 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2564 | |
2565 | for (i = 0; i < nargs; i++) | |
9771b263 | 2566 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2567 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2568 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2569 | |
2570 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2571 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
2572 | { |
2573 | size_t k; | |
9771b263 | 2574 | vargs.truncate (0); |
190c2236 JJ |
2575 | for (k = 0; k < nargs; k++) |
2576 | { | |
37b5ec8f | 2577 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
2578 | vargs.quick_push (vec_oprndsk[i]); |
2579 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 | 2580 | } |
70439f0d RS |
2581 | if (ifn != IFN_LAST) |
2582 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2583 | else | |
2584 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
190c2236 JJ |
2585 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2586 | gimple_call_set_lhs (new_stmt, new_temp); | |
2587 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2588 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2589 | } |
2590 | ||
2591 | for (i = 0; i < nargs; i++) | |
2592 | { | |
37b5ec8f | 2593 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2594 | vec_oprndsi.release (); |
190c2236 | 2595 | } |
190c2236 JJ |
2596 | continue; |
2597 | } | |
2598 | ||
ebfd146a IR |
2599 | for (i = 0; i < nargs; i++) |
2600 | { | |
2601 | op = gimple_call_arg (stmt, i); | |
2602 | if (j == 0) | |
2603 | { | |
2604 | vec_oprnd0 | |
81c40241 | 2605 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2606 | vec_oprnd1 |
63827fb8 | 2607 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2608 | } |
2609 | else | |
2610 | { | |
336ecb65 | 2611 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 2612 | vec_oprnd0 |
63827fb8 | 2613 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 2614 | vec_oprnd1 |
63827fb8 | 2615 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2616 | } |
2617 | ||
9771b263 DN |
2618 | vargs.quick_push (vec_oprnd0); |
2619 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
2620 | } |
2621 | ||
b1b6836e | 2622 | new_stmt = gimple_build_call_vec (fndecl, vargs); |
ebfd146a IR |
2623 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2624 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
2625 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2626 | ||
2627 | if (j == 0) | |
2628 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2629 | else | |
2630 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2631 | ||
2632 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2633 | } | |
2634 | ||
2635 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a | 2636 | } |
b1b6836e RS |
2637 | else |
2638 | /* No current target implements this case. */ | |
2639 | return false; | |
ebfd146a | 2640 | |
9771b263 | 2641 | vargs.release (); |
ebfd146a | 2642 | |
ebfd146a IR |
2643 | /* The call in STMT might prevent it from being removed in dce. |
2644 | We however cannot remove it here, due to the way the ssa name | |
2645 | it defines is mapped to the new definition. So just replace | |
2646 | rhs of the statement with something harmless. */ | |
2647 | ||
dd34c087 JJ |
2648 | if (slp_node) |
2649 | return true; | |
2650 | ||
ebfd146a | 2651 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
2652 | if (is_pattern_stmt_p (stmt_info)) |
2653 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
2654 | else | |
2655 | lhs = gimple_call_lhs (stmt); | |
3cc2fa2a | 2656 | |
9d5e7640 | 2657 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
ebfd146a | 2658 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 2659 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
2660 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
2661 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
2662 | |
2663 | return true; | |
2664 | } | |
2665 | ||
2666 | ||
0136f8f0 AH |
2667 | struct simd_call_arg_info |
2668 | { | |
2669 | tree vectype; | |
2670 | tree op; | |
2671 | enum vect_def_type dt; | |
2672 | HOST_WIDE_INT linear_step; | |
2673 | unsigned int align; | |
17b658af | 2674 | bool simd_lane_linear; |
0136f8f0 AH |
2675 | }; |
2676 | ||
17b658af JJ |
2677 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
2678 | is linear within simd lane (but not within whole loop), note it in | |
2679 | *ARGINFO. */ | |
2680 | ||
2681 | static void | |
2682 | vect_simd_lane_linear (tree op, struct loop *loop, | |
2683 | struct simd_call_arg_info *arginfo) | |
2684 | { | |
355fe088 | 2685 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
2686 | |
2687 | if (!is_gimple_assign (def_stmt) | |
2688 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
2689 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
2690 | return; | |
2691 | ||
2692 | tree base = gimple_assign_rhs1 (def_stmt); | |
2693 | HOST_WIDE_INT linear_step = 0; | |
2694 | tree v = gimple_assign_rhs2 (def_stmt); | |
2695 | while (TREE_CODE (v) == SSA_NAME) | |
2696 | { | |
2697 | tree t; | |
2698 | def_stmt = SSA_NAME_DEF_STMT (v); | |
2699 | if (is_gimple_assign (def_stmt)) | |
2700 | switch (gimple_assign_rhs_code (def_stmt)) | |
2701 | { | |
2702 | case PLUS_EXPR: | |
2703 | t = gimple_assign_rhs2 (def_stmt); | |
2704 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
2705 | return; | |
2706 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
2707 | v = gimple_assign_rhs1 (def_stmt); | |
2708 | continue; | |
2709 | case MULT_EXPR: | |
2710 | t = gimple_assign_rhs2 (def_stmt); | |
2711 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
2712 | return; | |
2713 | linear_step = tree_to_shwi (t); | |
2714 | v = gimple_assign_rhs1 (def_stmt); | |
2715 | continue; | |
2716 | CASE_CONVERT: | |
2717 | t = gimple_assign_rhs1 (def_stmt); | |
2718 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
2719 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
2720 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
2721 | return; | |
2722 | if (!linear_step) | |
2723 | linear_step = 1; | |
2724 | v = t; | |
2725 | continue; | |
2726 | default: | |
2727 | return; | |
2728 | } | |
2729 | else if (is_gimple_call (def_stmt) | |
2730 | && gimple_call_internal_p (def_stmt) | |
2731 | && gimple_call_internal_fn (def_stmt) == IFN_GOMP_SIMD_LANE | |
2732 | && loop->simduid | |
2733 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
2734 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
2735 | == loop->simduid)) | |
2736 | { | |
2737 | if (!linear_step) | |
2738 | linear_step = 1; | |
2739 | arginfo->linear_step = linear_step; | |
2740 | arginfo->op = base; | |
2741 | arginfo->simd_lane_linear = true; | |
2742 | return; | |
2743 | } | |
2744 | } | |
2745 | } | |
2746 | ||
0136f8f0 AH |
2747 | /* Function vectorizable_simd_clone_call. |
2748 | ||
2749 | Check if STMT performs a function call that can be vectorized | |
2750 | by calling a simd clone of the function. | |
2751 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2752 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2753 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2754 | ||
2755 | static bool | |
355fe088 TS |
2756 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
2757 | gimple **vec_stmt, slp_tree slp_node) | |
0136f8f0 AH |
2758 | { |
2759 | tree vec_dest; | |
2760 | tree scalar_dest; | |
2761 | tree op, type; | |
2762 | tree vec_oprnd0 = NULL_TREE; | |
2763 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
2764 | tree vectype; | |
2765 | unsigned int nunits; | |
2766 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2767 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 2768 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 2769 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 2770 | tree fndecl, new_temp; |
355fe088 TS |
2771 | gimple *def_stmt; |
2772 | gimple *new_stmt = NULL; | |
0136f8f0 | 2773 | int ncopies, j; |
00426f9a | 2774 | auto_vec<simd_call_arg_info> arginfo; |
0136f8f0 AH |
2775 | vec<tree> vargs = vNULL; |
2776 | size_t i, nargs; | |
2777 | tree lhs, rtype, ratype; | |
2778 | vec<constructor_elt, va_gc> *ret_ctor_elts; | |
2779 | ||
2780 | /* Is STMT a vectorizable call? */ | |
2781 | if (!is_gimple_call (stmt)) | |
2782 | return false; | |
2783 | ||
2784 | fndecl = gimple_call_fndecl (stmt); | |
2785 | if (fndecl == NULL_TREE) | |
2786 | return false; | |
2787 | ||
d52f5295 | 2788 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
2789 | if (node == NULL || node->simd_clones == NULL) |
2790 | return false; | |
2791 | ||
2792 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
2793 | return false; | |
2794 | ||
66c16fd9 RB |
2795 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
2796 | && ! vec_stmt) | |
0136f8f0 AH |
2797 | return false; |
2798 | ||
2799 | if (gimple_call_lhs (stmt) | |
2800 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
2801 | return false; | |
2802 | ||
2803 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
2804 | ||
2805 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2806 | ||
2807 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
2808 | return false; | |
2809 | ||
2810 | /* FORNOW */ | |
fce57248 | 2811 | if (slp_node) |
0136f8f0 AH |
2812 | return false; |
2813 | ||
2814 | /* Process function arguments. */ | |
2815 | nargs = gimple_call_num_args (stmt); | |
2816 | ||
2817 | /* Bail out if the function has zero arguments. */ | |
2818 | if (nargs == 0) | |
2819 | return false; | |
2820 | ||
00426f9a | 2821 | arginfo.reserve (nargs, true); |
0136f8f0 AH |
2822 | |
2823 | for (i = 0; i < nargs; i++) | |
2824 | { | |
2825 | simd_call_arg_info thisarginfo; | |
2826 | affine_iv iv; | |
2827 | ||
2828 | thisarginfo.linear_step = 0; | |
2829 | thisarginfo.align = 0; | |
2830 | thisarginfo.op = NULL_TREE; | |
17b658af | 2831 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
2832 | |
2833 | op = gimple_call_arg (stmt, i); | |
81c40241 RB |
2834 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &thisarginfo.dt, |
2835 | &thisarginfo.vectype) | |
0136f8f0 AH |
2836 | || thisarginfo.dt == vect_uninitialized_def) |
2837 | { | |
2838 | if (dump_enabled_p ()) | |
2839 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2840 | "use not simple.\n"); | |
0136f8f0 AH |
2841 | return false; |
2842 | } | |
2843 | ||
2844 | if (thisarginfo.dt == vect_constant_def | |
2845 | || thisarginfo.dt == vect_external_def) | |
2846 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
2847 | else | |
2848 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
2849 | ||
6c9e85fb JJ |
2850 | /* For linear arguments, the analyze phase should have saved |
2851 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
2852 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
2853 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
2854 | { |
2855 | gcc_assert (vec_stmt); | |
2856 | thisarginfo.linear_step | |
17b658af | 2857 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 2858 | thisarginfo.op |
17b658af JJ |
2859 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
2860 | thisarginfo.simd_lane_linear | |
2861 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
2862 | == boolean_true_node); | |
6c9e85fb JJ |
2863 | /* If loop has been peeled for alignment, we need to adjust it. */ |
2864 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
2865 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 2866 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
2867 | { |
2868 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 2869 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
2870 | tree opt = TREE_TYPE (thisarginfo.op); |
2871 | bias = fold_convert (TREE_TYPE (step), bias); | |
2872 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
2873 | thisarginfo.op | |
2874 | = fold_build2 (POINTER_TYPE_P (opt) | |
2875 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
2876 | thisarginfo.op, bias); | |
2877 | } | |
2878 | } | |
2879 | else if (!vec_stmt | |
2880 | && thisarginfo.dt != vect_constant_def | |
2881 | && thisarginfo.dt != vect_external_def | |
2882 | && loop_vinfo | |
2883 | && TREE_CODE (op) == SSA_NAME | |
2884 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
2885 | &iv, false) | |
2886 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
2887 | { |
2888 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
2889 | thisarginfo.op = iv.base; | |
2890 | } | |
2891 | else if ((thisarginfo.dt == vect_constant_def | |
2892 | || thisarginfo.dt == vect_external_def) | |
2893 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
2894 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
2895 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
2896 | linear too. */ | |
2897 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
2898 | && !thisarginfo.linear_step | |
2899 | && !vec_stmt | |
2900 | && thisarginfo.dt != vect_constant_def | |
2901 | && thisarginfo.dt != vect_external_def | |
2902 | && loop_vinfo | |
2903 | && !slp_node | |
2904 | && TREE_CODE (op) == SSA_NAME) | |
2905 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
2906 | |
2907 | arginfo.quick_push (thisarginfo); | |
2908 | } | |
2909 | ||
2910 | unsigned int badness = 0; | |
2911 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
2912 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
2913 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
2914 | else |
2915 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
2916 | n = n->simdclone->next_clone) | |
2917 | { | |
2918 | unsigned int this_badness = 0; | |
2919 | if (n->simdclone->simdlen | |
2920 | > (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
2921 | || n->simdclone->nargs != nargs) | |
2922 | continue; | |
2923 | if (n->simdclone->simdlen | |
2924 | < (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2925 | this_badness += (exact_log2 (LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2926 | - exact_log2 (n->simdclone->simdlen)) * 1024; | |
2927 | if (n->simdclone->inbranch) | |
2928 | this_badness += 2048; | |
2929 | int target_badness = targetm.simd_clone.usable (n); | |
2930 | if (target_badness < 0) | |
2931 | continue; | |
2932 | this_badness += target_badness * 512; | |
2933 | /* FORNOW: Have to add code to add the mask argument. */ | |
2934 | if (n->simdclone->inbranch) | |
2935 | continue; | |
2936 | for (i = 0; i < nargs; i++) | |
2937 | { | |
2938 | switch (n->simdclone->args[i].arg_type) | |
2939 | { | |
2940 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
2941 | if (!useless_type_conversion_p | |
2942 | (n->simdclone->args[i].orig_type, | |
2943 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
2944 | i = -1; | |
2945 | else if (arginfo[i].dt == vect_constant_def | |
2946 | || arginfo[i].dt == vect_external_def | |
2947 | || arginfo[i].linear_step) | |
2948 | this_badness += 64; | |
2949 | break; | |
2950 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
2951 | if (arginfo[i].dt != vect_constant_def | |
2952 | && arginfo[i].dt != vect_external_def) | |
2953 | i = -1; | |
2954 | break; | |
2955 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 2956 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
2957 | if (arginfo[i].dt == vect_constant_def |
2958 | || arginfo[i].dt == vect_external_def | |
2959 | || (arginfo[i].linear_step | |
2960 | != n->simdclone->args[i].linear_step)) | |
2961 | i = -1; | |
2962 | break; | |
2963 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
2964 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
2965 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
2966 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
2967 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
2968 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
2969 | /* FORNOW */ |
2970 | i = -1; | |
2971 | break; | |
2972 | case SIMD_CLONE_ARG_TYPE_MASK: | |
2973 | gcc_unreachable (); | |
2974 | } | |
2975 | if (i == (size_t) -1) | |
2976 | break; | |
2977 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
2978 | { | |
2979 | i = -1; | |
2980 | break; | |
2981 | } | |
2982 | if (arginfo[i].align) | |
2983 | this_badness += (exact_log2 (arginfo[i].align) | |
2984 | - exact_log2 (n->simdclone->args[i].alignment)); | |
2985 | } | |
2986 | if (i == (size_t) -1) | |
2987 | continue; | |
2988 | if (bestn == NULL || this_badness < badness) | |
2989 | { | |
2990 | bestn = n; | |
2991 | badness = this_badness; | |
2992 | } | |
2993 | } | |
2994 | ||
2995 | if (bestn == NULL) | |
00426f9a | 2996 | return false; |
0136f8f0 AH |
2997 | |
2998 | for (i = 0; i < nargs; i++) | |
2999 | if ((arginfo[i].dt == vect_constant_def | |
3000 | || arginfo[i].dt == vect_external_def) | |
3001 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
3002 | { | |
3003 | arginfo[i].vectype | |
3004 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
3005 | i))); | |
3006 | if (arginfo[i].vectype == NULL | |
3007 | || (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
3008 | > bestn->simdclone->simdlen)) | |
00426f9a | 3009 | return false; |
0136f8f0 AH |
3010 | } |
3011 | ||
3012 | fndecl = bestn->decl; | |
3013 | nunits = bestn->simdclone->simdlen; | |
3014 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3015 | ||
3016 | /* If the function isn't const, only allow it in simd loops where user | |
3017 | has asserted that at least nunits consecutive iterations can be | |
3018 | performed using SIMD instructions. */ | |
3019 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
3020 | && gimple_vuse (stmt)) | |
00426f9a | 3021 | return false; |
0136f8f0 AH |
3022 | |
3023 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3024 | needs to be generated. */ | |
3025 | gcc_assert (ncopies >= 1); | |
3026 | ||
3027 | if (!vec_stmt) /* transformation not required. */ | |
3028 | { | |
6c9e85fb JJ |
3029 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
3030 | for (i = 0; i < nargs; i++) | |
7adb26f2 JJ |
3031 | if ((bestn->simdclone->args[i].arg_type |
3032 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
3033 | || (bestn->simdclone->args[i].arg_type | |
3034 | == SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP)) | |
6c9e85fb | 3035 | { |
17b658af | 3036 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
3037 | + 1); |
3038 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
3039 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
3040 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
3041 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
3042 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
3043 | tree sll = arginfo[i].simd_lane_linear |
3044 | ? boolean_true_node : boolean_false_node; | |
3045 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 3046 | } |
0136f8f0 AH |
3047 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
3048 | if (dump_enabled_p ()) | |
3049 | dump_printf_loc (MSG_NOTE, vect_location, | |
3050 | "=== vectorizable_simd_clone_call ===\n"); | |
3051 | /* vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); */ | |
0136f8f0 AH |
3052 | return true; |
3053 | } | |
3054 | ||
3055 | /** Transform. **/ | |
3056 | ||
3057 | if (dump_enabled_p ()) | |
3058 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
3059 | ||
3060 | /* Handle def. */ | |
3061 | scalar_dest = gimple_call_lhs (stmt); | |
3062 | vec_dest = NULL_TREE; | |
3063 | rtype = NULL_TREE; | |
3064 | ratype = NULL_TREE; | |
3065 | if (scalar_dest) | |
3066 | { | |
3067 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3068 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
3069 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
3070 | { | |
3071 | ratype = rtype; | |
3072 | rtype = TREE_TYPE (ratype); | |
3073 | } | |
3074 | } | |
3075 | ||
3076 | prev_stmt_info = NULL; | |
3077 | for (j = 0; j < ncopies; ++j) | |
3078 | { | |
3079 | /* Build argument list for the vectorized call. */ | |
3080 | if (j == 0) | |
3081 | vargs.create (nargs); | |
3082 | else | |
3083 | vargs.truncate (0); | |
3084 | ||
3085 | for (i = 0; i < nargs; i++) | |
3086 | { | |
3087 | unsigned int k, l, m, o; | |
3088 | tree atype; | |
3089 | op = gimple_call_arg (stmt, i); | |
3090 | switch (bestn->simdclone->args[i].arg_type) | |
3091 | { | |
3092 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3093 | atype = bestn->simdclone->args[i].vector_type; | |
3094 | o = nunits / TYPE_VECTOR_SUBPARTS (atype); | |
3095 | for (m = j * o; m < (j + 1) * o; m++) | |
3096 | { | |
3097 | if (TYPE_VECTOR_SUBPARTS (atype) | |
3098 | < TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)) | |
3099 | { | |
3100 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); | |
3101 | k = (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
3102 | / TYPE_VECTOR_SUBPARTS (atype)); | |
3103 | gcc_assert ((k & (k - 1)) == 0); | |
3104 | if (m == 0) | |
3105 | vec_oprnd0 | |
81c40241 | 3106 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3107 | else |
3108 | { | |
3109 | vec_oprnd0 = arginfo[i].op; | |
3110 | if ((m & (k - 1)) == 0) | |
3111 | vec_oprnd0 | |
3112 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3113 | vec_oprnd0); | |
3114 | } | |
3115 | arginfo[i].op = vec_oprnd0; | |
3116 | vec_oprnd0 | |
3117 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
3118 | size_int (prec), | |
3119 | bitsize_int ((m & (k - 1)) * prec)); | |
3120 | new_stmt | |
b731b390 | 3121 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3122 | vec_oprnd0); |
3123 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3124 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3125 | } | |
3126 | else | |
3127 | { | |
3128 | k = (TYPE_VECTOR_SUBPARTS (atype) | |
3129 | / TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)); | |
3130 | gcc_assert ((k & (k - 1)) == 0); | |
3131 | vec<constructor_elt, va_gc> *ctor_elts; | |
3132 | if (k != 1) | |
3133 | vec_alloc (ctor_elts, k); | |
3134 | else | |
3135 | ctor_elts = NULL; | |
3136 | for (l = 0; l < k; l++) | |
3137 | { | |
3138 | if (m == 0 && l == 0) | |
3139 | vec_oprnd0 | |
81c40241 | 3140 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3141 | else |
3142 | vec_oprnd0 | |
3143 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3144 | arginfo[i].op); | |
3145 | arginfo[i].op = vec_oprnd0; | |
3146 | if (k == 1) | |
3147 | break; | |
3148 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
3149 | vec_oprnd0); | |
3150 | } | |
3151 | if (k == 1) | |
3152 | vargs.safe_push (vec_oprnd0); | |
3153 | else | |
3154 | { | |
3155 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
3156 | new_stmt | |
b731b390 | 3157 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3158 | vec_oprnd0); |
3159 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3160 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3161 | } | |
3162 | } | |
3163 | } | |
3164 | break; | |
3165 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3166 | vargs.safe_push (op); | |
3167 | break; | |
3168 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
7adb26f2 | 3169 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
3170 | if (j == 0) |
3171 | { | |
3172 | gimple_seq stmts; | |
3173 | arginfo[i].op | |
3174 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
3175 | NULL_TREE); | |
3176 | if (stmts != NULL) | |
3177 | { | |
3178 | basic_block new_bb; | |
3179 | edge pe = loop_preheader_edge (loop); | |
3180 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
3181 | gcc_assert (!new_bb); | |
3182 | } | |
17b658af JJ |
3183 | if (arginfo[i].simd_lane_linear) |
3184 | { | |
3185 | vargs.safe_push (arginfo[i].op); | |
3186 | break; | |
3187 | } | |
b731b390 | 3188 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 3189 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
0136f8f0 | 3190 | set_vinfo_for_stmt (new_phi, |
310213d4 | 3191 | new_stmt_vec_info (new_phi, loop_vinfo)); |
0136f8f0 AH |
3192 | add_phi_arg (new_phi, arginfo[i].op, |
3193 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
3194 | enum tree_code code | |
3195 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3196 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3197 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3198 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3199 | widest_int cst |
3200 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3201 | ncopies * nunits); | |
3202 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3203 | tree phi_arg = copy_ssa_name (op); |
0d0e4a03 JJ |
3204 | new_stmt |
3205 | = gimple_build_assign (phi_arg, code, phi_res, tcst); | |
0136f8f0 AH |
3206 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
3207 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
3208 | set_vinfo_for_stmt (new_stmt, | |
310213d4 | 3209 | new_stmt_vec_info (new_stmt, loop_vinfo)); |
0136f8f0 AH |
3210 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
3211 | UNKNOWN_LOCATION); | |
3212 | arginfo[i].op = phi_res; | |
3213 | vargs.safe_push (phi_res); | |
3214 | } | |
3215 | else | |
3216 | { | |
3217 | enum tree_code code | |
3218 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3219 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3220 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3221 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3222 | widest_int cst |
3223 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3224 | j * nunits); | |
3225 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3226 | new_temp = make_ssa_name (TREE_TYPE (op)); |
0d0e4a03 JJ |
3227 | new_stmt = gimple_build_assign (new_temp, code, |
3228 | arginfo[i].op, tcst); | |
0136f8f0 AH |
3229 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3230 | vargs.safe_push (new_temp); | |
3231 | } | |
3232 | break; | |
7adb26f2 JJ |
3233 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
3234 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
0136f8f0 | 3235 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
e01d41e5 JJ |
3236 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3237 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3238 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3239 | default: |
3240 | gcc_unreachable (); | |
3241 | } | |
3242 | } | |
3243 | ||
3244 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
3245 | if (vec_dest) | |
3246 | { | |
3247 | gcc_assert (ratype || TYPE_VECTOR_SUBPARTS (rtype) == nunits); | |
3248 | if (ratype) | |
b731b390 | 3249 | new_temp = create_tmp_var (ratype); |
0136f8f0 AH |
3250 | else if (TYPE_VECTOR_SUBPARTS (vectype) |
3251 | == TYPE_VECTOR_SUBPARTS (rtype)) | |
3252 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3253 | else | |
3254 | new_temp = make_ssa_name (rtype, new_stmt); | |
3255 | gimple_call_set_lhs (new_stmt, new_temp); | |
3256 | } | |
3257 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3258 | ||
3259 | if (vec_dest) | |
3260 | { | |
3261 | if (TYPE_VECTOR_SUBPARTS (vectype) < nunits) | |
3262 | { | |
3263 | unsigned int k, l; | |
3264 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); | |
3265 | k = nunits / TYPE_VECTOR_SUBPARTS (vectype); | |
3266 | gcc_assert ((k & (k - 1)) == 0); | |
3267 | for (l = 0; l < k; l++) | |
3268 | { | |
3269 | tree t; | |
3270 | if (ratype) | |
3271 | { | |
3272 | t = build_fold_addr_expr (new_temp); | |
3273 | t = build2 (MEM_REF, vectype, t, | |
3274 | build_int_cst (TREE_TYPE (t), | |
3275 | l * prec / BITS_PER_UNIT)); | |
3276 | } | |
3277 | else | |
3278 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
3279 | size_int (prec), bitsize_int (l * prec)); | |
3280 | new_stmt | |
b731b390 | 3281 | = gimple_build_assign (make_ssa_name (vectype), t); |
0136f8f0 AH |
3282 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3283 | if (j == 0 && l == 0) | |
3284 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3285 | else | |
3286 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3287 | ||
3288 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3289 | } | |
3290 | ||
3291 | if (ratype) | |
3292 | { | |
3293 | tree clobber = build_constructor (ratype, NULL); | |
3294 | TREE_THIS_VOLATILE (clobber) = 1; | |
3295 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3296 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3297 | } | |
3298 | continue; | |
3299 | } | |
3300 | else if (TYPE_VECTOR_SUBPARTS (vectype) > nunits) | |
3301 | { | |
3302 | unsigned int k = (TYPE_VECTOR_SUBPARTS (vectype) | |
3303 | / TYPE_VECTOR_SUBPARTS (rtype)); | |
3304 | gcc_assert ((k & (k - 1)) == 0); | |
3305 | if ((j & (k - 1)) == 0) | |
3306 | vec_alloc (ret_ctor_elts, k); | |
3307 | if (ratype) | |
3308 | { | |
3309 | unsigned int m, o = nunits / TYPE_VECTOR_SUBPARTS (rtype); | |
3310 | for (m = 0; m < o; m++) | |
3311 | { | |
3312 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
3313 | size_int (m), NULL_TREE, NULL_TREE); | |
3314 | new_stmt | |
b731b390 | 3315 | = gimple_build_assign (make_ssa_name (rtype), tem); |
0136f8f0 AH |
3316 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3317 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
3318 | gimple_assign_lhs (new_stmt)); | |
3319 | } | |
3320 | tree clobber = build_constructor (ratype, NULL); | |
3321 | TREE_THIS_VOLATILE (clobber) = 1; | |
3322 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3323 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3324 | } | |
3325 | else | |
3326 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
3327 | if ((j & (k - 1)) != k - 1) | |
3328 | continue; | |
3329 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
3330 | new_stmt | |
b731b390 | 3331 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
0136f8f0 AH |
3332 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3333 | ||
3334 | if ((unsigned) j == k - 1) | |
3335 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3336 | else | |
3337 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3338 | ||
3339 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3340 | continue; | |
3341 | } | |
3342 | else if (ratype) | |
3343 | { | |
3344 | tree t = build_fold_addr_expr (new_temp); | |
3345 | t = build2 (MEM_REF, vectype, t, | |
3346 | build_int_cst (TREE_TYPE (t), 0)); | |
3347 | new_stmt | |
b731b390 | 3348 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
0136f8f0 AH |
3349 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3350 | tree clobber = build_constructor (ratype, NULL); | |
3351 | TREE_THIS_VOLATILE (clobber) = 1; | |
3352 | vect_finish_stmt_generation (stmt, | |
3353 | gimple_build_assign (new_temp, | |
3354 | clobber), gsi); | |
3355 | } | |
3356 | } | |
3357 | ||
3358 | if (j == 0) | |
3359 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3360 | else | |
3361 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3362 | ||
3363 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3364 | } | |
3365 | ||
3366 | vargs.release (); | |
3367 | ||
3368 | /* The call in STMT might prevent it from being removed in dce. | |
3369 | We however cannot remove it here, due to the way the ssa name | |
3370 | it defines is mapped to the new definition. So just replace | |
3371 | rhs of the statement with something harmless. */ | |
3372 | ||
3373 | if (slp_node) | |
3374 | return true; | |
3375 | ||
3376 | if (scalar_dest) | |
3377 | { | |
3378 | type = TREE_TYPE (scalar_dest); | |
3379 | if (is_pattern_stmt_p (stmt_info)) | |
3380 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3381 | else | |
3382 | lhs = gimple_call_lhs (stmt); | |
3383 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
3384 | } | |
3385 | else | |
3386 | new_stmt = gimple_build_nop (); | |
3387 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
3388 | set_vinfo_for_stmt (stmt, NULL); | |
3389 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 3390 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
3391 | unlink_stmt_vdef (stmt); |
3392 | ||
3393 | return true; | |
3394 | } | |
3395 | ||
3396 | ||
ebfd146a IR |
3397 | /* Function vect_gen_widened_results_half |
3398 | ||
3399 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 3400 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 3401 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
3402 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
3403 | needs to be created (DECL is a function-decl of a target-builtin). | |
3404 | STMT is the original scalar stmt that we are vectorizing. */ | |
3405 | ||
355fe088 | 3406 | static gimple * |
ebfd146a IR |
3407 | vect_gen_widened_results_half (enum tree_code code, |
3408 | tree decl, | |
3409 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
3410 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 3411 | gimple *stmt) |
b8698a0f | 3412 | { |
355fe088 | 3413 | gimple *new_stmt; |
b8698a0f L |
3414 | tree new_temp; |
3415 | ||
3416 | /* Generate half of the widened result: */ | |
3417 | if (code == CALL_EXPR) | |
3418 | { | |
3419 | /* Target specific support */ | |
ebfd146a IR |
3420 | if (op_type == binary_op) |
3421 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
3422 | else | |
3423 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
3424 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3425 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
3426 | } |
3427 | else | |
ebfd146a | 3428 | { |
b8698a0f L |
3429 | /* Generic support */ |
3430 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
3431 | if (op_type != binary_op) |
3432 | vec_oprnd1 = NULL; | |
0d0e4a03 | 3433 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
3434 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3435 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 3436 | } |
ebfd146a IR |
3437 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3438 | ||
ebfd146a IR |
3439 | return new_stmt; |
3440 | } | |
3441 | ||
4a00c761 JJ |
3442 | |
3443 | /* Get vectorized definitions for loop-based vectorization. For the first | |
3444 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
3445 | scalar operand), and for the rest we get a copy with | |
3446 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
3447 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
3448 | The vectors are collected into VEC_OPRNDS. */ | |
3449 | ||
3450 | static void | |
355fe088 | 3451 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 3452 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
3453 | { |
3454 | tree vec_oprnd; | |
3455 | ||
3456 | /* Get first vector operand. */ | |
3457 | /* All the vector operands except the very first one (that is scalar oprnd) | |
3458 | are stmt copies. */ | |
3459 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 3460 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
3461 | else |
3462 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
3463 | ||
9771b263 | 3464 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3465 | |
3466 | /* Get second vector operand. */ | |
3467 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 3468 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3469 | |
3470 | *oprnd = vec_oprnd; | |
3471 | ||
3472 | /* For conversion in multiple steps, continue to get operands | |
3473 | recursively. */ | |
3474 | if (multi_step_cvt) | |
3475 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
3476 | } | |
3477 | ||
3478 | ||
3479 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
3480 | For multi-step conversions store the resulting vectors and call the function | |
3481 | recursively. */ | |
3482 | ||
3483 | static void | |
9771b263 | 3484 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 3485 | int multi_step_cvt, gimple *stmt, |
9771b263 | 3486 | vec<tree> vec_dsts, |
4a00c761 JJ |
3487 | gimple_stmt_iterator *gsi, |
3488 | slp_tree slp_node, enum tree_code code, | |
3489 | stmt_vec_info *prev_stmt_info) | |
3490 | { | |
3491 | unsigned int i; | |
3492 | tree vop0, vop1, new_tmp, vec_dest; | |
355fe088 | 3493 | gimple *new_stmt; |
4a00c761 JJ |
3494 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
3495 | ||
9771b263 | 3496 | vec_dest = vec_dsts.pop (); |
4a00c761 | 3497 | |
9771b263 | 3498 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
3499 | { |
3500 | /* Create demotion operation. */ | |
9771b263 DN |
3501 | vop0 = (*vec_oprnds)[i]; |
3502 | vop1 = (*vec_oprnds)[i + 1]; | |
0d0e4a03 | 3503 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
3504 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
3505 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
3506 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3507 | ||
3508 | if (multi_step_cvt) | |
3509 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 3510 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
3511 | else |
3512 | { | |
3513 | /* This is the last step of the conversion sequence. Store the | |
3514 | vectors in SLP_NODE or in vector info of the scalar statement | |
3515 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
3516 | if (slp_node) | |
9771b263 | 3517 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 3518 | else |
c689ce1e RB |
3519 | { |
3520 | if (!*prev_stmt_info) | |
3521 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3522 | else | |
3523 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
4a00c761 | 3524 | |
c689ce1e RB |
3525 | *prev_stmt_info = vinfo_for_stmt (new_stmt); |
3526 | } | |
4a00c761 JJ |
3527 | } |
3528 | } | |
3529 | ||
3530 | /* For multi-step demotion operations we first generate demotion operations | |
3531 | from the source type to the intermediate types, and then combine the | |
3532 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
3533 | type. */ | |
3534 | if (multi_step_cvt) | |
3535 | { | |
3536 | /* At each level of recursion we have half of the operands we had at the | |
3537 | previous level. */ | |
9771b263 | 3538 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
3539 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
3540 | stmt, vec_dsts, gsi, slp_node, | |
3541 | VEC_PACK_TRUNC_EXPR, | |
3542 | prev_stmt_info); | |
3543 | } | |
3544 | ||
9771b263 | 3545 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3546 | } |
3547 | ||
3548 | ||
3549 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
3550 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
3551 | the resulting vectors and call the function recursively. */ | |
3552 | ||
3553 | static void | |
9771b263 DN |
3554 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
3555 | vec<tree> *vec_oprnds1, | |
355fe088 | 3556 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
3557 | gimple_stmt_iterator *gsi, |
3558 | enum tree_code code1, | |
3559 | enum tree_code code2, tree decl1, | |
3560 | tree decl2, int op_type) | |
3561 | { | |
3562 | int i; | |
3563 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 3564 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 3565 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 3566 | |
9771b263 DN |
3567 | vec_tmp.create (vec_oprnds0->length () * 2); |
3568 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
3569 | { |
3570 | if (op_type == binary_op) | |
9771b263 | 3571 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
3572 | else |
3573 | vop1 = NULL_TREE; | |
3574 | ||
3575 | /* Generate the two halves of promotion operation. */ | |
3576 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
3577 | op_type, vec_dest, gsi, stmt); | |
3578 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
3579 | op_type, vec_dest, gsi, stmt); | |
3580 | if (is_gimple_call (new_stmt1)) | |
3581 | { | |
3582 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
3583 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
3584 | } | |
3585 | else | |
3586 | { | |
3587 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
3588 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
3589 | } | |
3590 | ||
3591 | /* Store the results for the next step. */ | |
9771b263 DN |
3592 | vec_tmp.quick_push (new_tmp1); |
3593 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
3594 | } |
3595 | ||
689eaba3 | 3596 | vec_oprnds0->release (); |
4a00c761 JJ |
3597 | *vec_oprnds0 = vec_tmp; |
3598 | } | |
3599 | ||
3600 | ||
b8698a0f L |
3601 | /* Check if STMT performs a conversion operation, that can be vectorized. |
3602 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 3603 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
3604 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
3605 | ||
3606 | static bool | |
355fe088 TS |
3607 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
3608 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
3609 | { |
3610 | tree vec_dest; | |
3611 | tree scalar_dest; | |
4a00c761 | 3612 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
3613 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
3614 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3615 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3616 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 3617 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
3618 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
3619 | tree new_temp; | |
355fe088 | 3620 | gimple *def_stmt; |
ebfd146a | 3621 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
355fe088 | 3622 | gimple *new_stmt = NULL; |
ebfd146a IR |
3623 | stmt_vec_info prev_stmt_info; |
3624 | int nunits_in; | |
3625 | int nunits_out; | |
3626 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
3627 | int ncopies, i, j; |
3628 | tree lhs_type, rhs_type; | |
ebfd146a | 3629 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
3630 | vec<tree> vec_oprnds0 = vNULL; |
3631 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 3632 | tree vop0; |
4a00c761 | 3633 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 3634 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 3635 | int multi_step_cvt = 0; |
6e1aa848 DN |
3636 | vec<tree> vec_dsts = vNULL; |
3637 | vec<tree> interm_types = vNULL; | |
4a00c761 JJ |
3638 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
3639 | int op_type; | |
ef4bddc2 | 3640 | machine_mode rhs_mode; |
4a00c761 | 3641 | unsigned short fltsz; |
ebfd146a IR |
3642 | |
3643 | /* Is STMT a vectorizable conversion? */ | |
3644 | ||
4a00c761 | 3645 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3646 | return false; |
3647 | ||
66c16fd9 RB |
3648 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3649 | && ! vec_stmt) | |
ebfd146a IR |
3650 | return false; |
3651 | ||
3652 | if (!is_gimple_assign (stmt)) | |
3653 | return false; | |
3654 | ||
3655 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3656 | return false; | |
3657 | ||
3658 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
3659 | if (!CONVERT_EXPR_CODE_P (code) |
3660 | && code != FIX_TRUNC_EXPR | |
3661 | && code != FLOAT_EXPR | |
3662 | && code != WIDEN_MULT_EXPR | |
3663 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
3664 | return false; |
3665 | ||
4a00c761 JJ |
3666 | op_type = TREE_CODE_LENGTH (code); |
3667 | ||
ebfd146a | 3668 | /* Check types of lhs and rhs. */ |
b690cc0f | 3669 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 3670 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
3671 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3672 | ||
ebfd146a IR |
3673 | op0 = gimple_assign_rhs1 (stmt); |
3674 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
3675 | |
3676 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3677 | && !((INTEGRAL_TYPE_P (lhs_type) | |
3678 | && INTEGRAL_TYPE_P (rhs_type)) | |
3679 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
3680 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
3681 | return false; | |
3682 | ||
e6f5c25d IE |
3683 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
3684 | && ((INTEGRAL_TYPE_P (lhs_type) | |
3685 | && (TYPE_PRECISION (lhs_type) | |
3686 | != GET_MODE_PRECISION (TYPE_MODE (lhs_type)))) | |
3687 | || (INTEGRAL_TYPE_P (rhs_type) | |
3688 | && (TYPE_PRECISION (rhs_type) | |
3689 | != GET_MODE_PRECISION (TYPE_MODE (rhs_type)))))) | |
4a00c761 | 3690 | { |
73fbfcad | 3691 | if (dump_enabled_p ()) |
78c60e3d | 3692 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
3693 | "type conversion to/from bit-precision unsupported." |
3694 | "\n"); | |
4a00c761 JJ |
3695 | return false; |
3696 | } | |
3697 | ||
b690cc0f | 3698 | /* Check the operands of the operation. */ |
81c40241 | 3699 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype_in)) |
b690cc0f | 3700 | { |
73fbfcad | 3701 | if (dump_enabled_p ()) |
78c60e3d | 3702 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3703 | "use not simple.\n"); |
b690cc0f RG |
3704 | return false; |
3705 | } | |
4a00c761 JJ |
3706 | if (op_type == binary_op) |
3707 | { | |
3708 | bool ok; | |
3709 | ||
3710 | op1 = gimple_assign_rhs2 (stmt); | |
3711 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
3712 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
3713 | OP1. */ | |
3714 | if (CONSTANT_CLASS_P (op0)) | |
81c40241 | 3715 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &vectype_in); |
4a00c761 | 3716 | else |
81c40241 | 3717 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1]); |
4a00c761 JJ |
3718 | |
3719 | if (!ok) | |
3720 | { | |
73fbfcad | 3721 | if (dump_enabled_p ()) |
78c60e3d | 3722 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3723 | "use not simple.\n"); |
4a00c761 JJ |
3724 | return false; |
3725 | } | |
3726 | } | |
3727 | ||
b690cc0f RG |
3728 | /* If op0 is an external or constant defs use a vector type of |
3729 | the same size as the output vector type. */ | |
ebfd146a | 3730 | if (!vectype_in) |
b690cc0f | 3731 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
3732 | if (vec_stmt) |
3733 | gcc_assert (vectype_in); | |
3734 | if (!vectype_in) | |
3735 | { | |
73fbfcad | 3736 | if (dump_enabled_p ()) |
4a00c761 | 3737 | { |
78c60e3d SS |
3738 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3739 | "no vectype for scalar type "); | |
3740 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 3741 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 3742 | } |
7d8930a0 IR |
3743 | |
3744 | return false; | |
3745 | } | |
ebfd146a | 3746 | |
e6f5c25d IE |
3747 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
3748 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
3749 | { | |
3750 | if (dump_enabled_p ()) | |
3751 | { | |
3752 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3753 | "can't convert between boolean and non " | |
3754 | "boolean vectors"); | |
3755 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
3756 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
3757 | } | |
3758 | ||
3759 | return false; | |
3760 | } | |
3761 | ||
b690cc0f RG |
3762 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
3763 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 3764 | if (nunits_in < nunits_out) |
ebfd146a IR |
3765 | modifier = NARROW; |
3766 | else if (nunits_out == nunits_in) | |
3767 | modifier = NONE; | |
ebfd146a | 3768 | else |
4a00c761 | 3769 | modifier = WIDEN; |
ebfd146a | 3770 | |
ff802fa1 IR |
3771 | /* Multiple types in SLP are handled by creating the appropriate number of |
3772 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3773 | case of SLP. */ | |
fce57248 | 3774 | if (slp_node) |
ebfd146a | 3775 | ncopies = 1; |
4a00c761 JJ |
3776 | else if (modifier == NARROW) |
3777 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
3778 | else | |
3779 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 3780 | |
ebfd146a IR |
3781 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
3782 | needs to be generated. */ | |
3783 | gcc_assert (ncopies >= 1); | |
3784 | ||
ebfd146a | 3785 | /* Supportable by target? */ |
4a00c761 | 3786 | switch (modifier) |
ebfd146a | 3787 | { |
4a00c761 JJ |
3788 | case NONE: |
3789 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3790 | return false; | |
3791 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
3792 | &decl1, &code1)) | |
3793 | break; | |
3794 | /* FALLTHRU */ | |
3795 | unsupported: | |
73fbfcad | 3796 | if (dump_enabled_p ()) |
78c60e3d | 3797 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3798 | "conversion not supported by target.\n"); |
ebfd146a | 3799 | return false; |
ebfd146a | 3800 | |
4a00c761 JJ |
3801 | case WIDEN: |
3802 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
3803 | &code1, &code2, &multi_step_cvt, |
3804 | &interm_types)) | |
4a00c761 JJ |
3805 | { |
3806 | /* Binary widening operation can only be supported directly by the | |
3807 | architecture. */ | |
3808 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
3809 | break; | |
3810 | } | |
3811 | ||
3812 | if (code != FLOAT_EXPR | |
3813 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3814 | <= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3815 | goto unsupported; | |
3816 | ||
3817 | rhs_mode = TYPE_MODE (rhs_type); | |
3818 | fltsz = GET_MODE_SIZE (TYPE_MODE (lhs_type)); | |
3819 | for (rhs_mode = GET_MODE_2XWIDER_MODE (TYPE_MODE (rhs_type)); | |
3820 | rhs_mode != VOIDmode && GET_MODE_SIZE (rhs_mode) <= fltsz; | |
3821 | rhs_mode = GET_MODE_2XWIDER_MODE (rhs_mode)) | |
3822 | { | |
3823 | cvt_type | |
3824 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3825 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3826 | if (cvt_type == NULL_TREE) | |
3827 | goto unsupported; | |
3828 | ||
3829 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3830 | { | |
3831 | if (!supportable_convert_operation (code, vectype_out, | |
3832 | cvt_type, &decl1, &codecvt1)) | |
3833 | goto unsupported; | |
3834 | } | |
3835 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
3836 | cvt_type, &codecvt1, |
3837 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
3838 | &interm_types)) |
3839 | continue; | |
3840 | else | |
3841 | gcc_assert (multi_step_cvt == 0); | |
3842 | ||
3843 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
3844 | vectype_in, &code1, &code2, |
3845 | &multi_step_cvt, &interm_types)) | |
4a00c761 JJ |
3846 | break; |
3847 | } | |
3848 | ||
3849 | if (rhs_mode == VOIDmode || GET_MODE_SIZE (rhs_mode) > fltsz) | |
3850 | goto unsupported; | |
3851 | ||
3852 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3853 | codecvt2 = ERROR_MARK; | |
3854 | else | |
3855 | { | |
3856 | multi_step_cvt++; | |
9771b263 | 3857 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
3858 | cvt_type = NULL_TREE; |
3859 | } | |
3860 | break; | |
3861 | ||
3862 | case NARROW: | |
3863 | gcc_assert (op_type == unary_op); | |
3864 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
3865 | &code1, &multi_step_cvt, | |
3866 | &interm_types)) | |
3867 | break; | |
3868 | ||
3869 | if (code != FIX_TRUNC_EXPR | |
3870 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3871 | >= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3872 | goto unsupported; | |
3873 | ||
3874 | rhs_mode = TYPE_MODE (rhs_type); | |
3875 | cvt_type | |
3876 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3877 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3878 | if (cvt_type == NULL_TREE) | |
3879 | goto unsupported; | |
3880 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
3881 | &decl1, &codecvt1)) | |
3882 | goto unsupported; | |
3883 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
3884 | &code1, &multi_step_cvt, | |
3885 | &interm_types)) | |
3886 | break; | |
3887 | goto unsupported; | |
3888 | ||
3889 | default: | |
3890 | gcc_unreachable (); | |
ebfd146a IR |
3891 | } |
3892 | ||
3893 | if (!vec_stmt) /* transformation not required. */ | |
3894 | { | |
73fbfcad | 3895 | if (dump_enabled_p ()) |
78c60e3d | 3896 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3897 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 3898 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
3899 | { |
3900 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
c3e7ee41 | 3901 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
8bd37302 | 3902 | } |
4a00c761 JJ |
3903 | else if (modifier == NARROW) |
3904 | { | |
3905 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 3906 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
3907 | } |
3908 | else | |
3909 | { | |
3910 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 3911 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 3912 | } |
9771b263 | 3913 | interm_types.release (); |
ebfd146a IR |
3914 | return true; |
3915 | } | |
3916 | ||
3917 | /** Transform. **/ | |
73fbfcad | 3918 | if (dump_enabled_p ()) |
78c60e3d | 3919 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3920 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 3921 | |
4a00c761 JJ |
3922 | if (op_type == binary_op) |
3923 | { | |
3924 | if (CONSTANT_CLASS_P (op0)) | |
3925 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
3926 | else if (CONSTANT_CLASS_P (op1)) | |
3927 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
3928 | } | |
3929 | ||
3930 | /* In case of multi-step conversion, we first generate conversion operations | |
3931 | to the intermediate types, and then from that types to the final one. | |
3932 | We create vector destinations for the intermediate type (TYPES) received | |
3933 | from supportable_*_operation, and store them in the correct order | |
3934 | for future use in vect_create_vectorized_*_stmts (). */ | |
9771b263 | 3935 | vec_dsts.create (multi_step_cvt + 1); |
82294ec1 JJ |
3936 | vec_dest = vect_create_destination_var (scalar_dest, |
3937 | (cvt_type && modifier == WIDEN) | |
3938 | ? cvt_type : vectype_out); | |
9771b263 | 3939 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3940 | |
3941 | if (multi_step_cvt) | |
3942 | { | |
9771b263 DN |
3943 | for (i = interm_types.length () - 1; |
3944 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
3945 | { |
3946 | vec_dest = vect_create_destination_var (scalar_dest, | |
3947 | intermediate_type); | |
9771b263 | 3948 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3949 | } |
3950 | } | |
ebfd146a | 3951 | |
4a00c761 | 3952 | if (cvt_type) |
82294ec1 JJ |
3953 | vec_dest = vect_create_destination_var (scalar_dest, |
3954 | modifier == WIDEN | |
3955 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
3956 | |
3957 | if (!slp_node) | |
3958 | { | |
30862efc | 3959 | if (modifier == WIDEN) |
4a00c761 | 3960 | { |
c3284718 | 3961 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 3962 | if (op_type == binary_op) |
9771b263 | 3963 | vec_oprnds1.create (1); |
4a00c761 | 3964 | } |
30862efc | 3965 | else if (modifier == NARROW) |
9771b263 DN |
3966 | vec_oprnds0.create ( |
3967 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
3968 | } |
3969 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 3970 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 3971 | |
4a00c761 | 3972 | last_oprnd = op0; |
ebfd146a IR |
3973 | prev_stmt_info = NULL; |
3974 | switch (modifier) | |
3975 | { | |
3976 | case NONE: | |
3977 | for (j = 0; j < ncopies; j++) | |
3978 | { | |
ebfd146a | 3979 | if (j == 0) |
d092494c IR |
3980 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node, |
3981 | -1); | |
ebfd146a IR |
3982 | else |
3983 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
3984 | ||
9771b263 | 3985 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
3986 | { |
3987 | /* Arguments are ready, create the new vector stmt. */ | |
3988 | if (code1 == CALL_EXPR) | |
3989 | { | |
3990 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
3991 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3992 | gimple_call_set_lhs (new_stmt, new_temp); | |
3993 | } | |
3994 | else | |
3995 | { | |
3996 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
0d0e4a03 | 3997 | new_stmt = gimple_build_assign (vec_dest, code1, vop0); |
4a00c761 JJ |
3998 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3999 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4000 | } | |
4001 | ||
4002 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4003 | if (slp_node) | |
9771b263 | 4004 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
225ce44b RB |
4005 | else |
4006 | { | |
4007 | if (!prev_stmt_info) | |
4008 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4009 | else | |
4010 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4011 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4012 | } | |
4a00c761 | 4013 | } |
ebfd146a IR |
4014 | } |
4015 | break; | |
4016 | ||
4017 | case WIDEN: | |
4018 | /* In case the vectorization factor (VF) is bigger than the number | |
4019 | of elements that we can fit in a vectype (nunits), we have to | |
4020 | generate more than one vector stmt - i.e - we need to "unroll" | |
4021 | the vector stmt by a factor VF/nunits. */ | |
4022 | for (j = 0; j < ncopies; j++) | |
4023 | { | |
4a00c761 | 4024 | /* Handle uses. */ |
ebfd146a | 4025 | if (j == 0) |
4a00c761 JJ |
4026 | { |
4027 | if (slp_node) | |
4028 | { | |
4029 | if (code == WIDEN_LSHIFT_EXPR) | |
4030 | { | |
4031 | unsigned int k; | |
ebfd146a | 4032 | |
4a00c761 JJ |
4033 | vec_oprnd1 = op1; |
4034 | /* Store vec_oprnd1 for every vector stmt to be created | |
4035 | for SLP_NODE. We check during the analysis that all | |
4036 | the shift arguments are the same. */ | |
4037 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4038 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4039 | |
4040 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
4041 | slp_node, -1); | |
4042 | } | |
4043 | else | |
4044 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
4045 | &vec_oprnds1, slp_node, -1); | |
4046 | } | |
4047 | else | |
4048 | { | |
81c40241 | 4049 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 4050 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
4051 | if (op_type == binary_op) |
4052 | { | |
4053 | if (code == WIDEN_LSHIFT_EXPR) | |
4054 | vec_oprnd1 = op1; | |
4055 | else | |
81c40241 | 4056 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 4057 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4058 | } |
4059 | } | |
4060 | } | |
ebfd146a | 4061 | else |
4a00c761 JJ |
4062 | { |
4063 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
4064 | vec_oprnds0.truncate (0); |
4065 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
4066 | if (op_type == binary_op) |
4067 | { | |
4068 | if (code == WIDEN_LSHIFT_EXPR) | |
4069 | vec_oprnd1 = op1; | |
4070 | else | |
4071 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
4072 | vec_oprnd1); | |
9771b263 DN |
4073 | vec_oprnds1.truncate (0); |
4074 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
4075 | } |
4076 | } | |
ebfd146a | 4077 | |
4a00c761 JJ |
4078 | /* Arguments are ready. Create the new vector stmts. */ |
4079 | for (i = multi_step_cvt; i >= 0; i--) | |
4080 | { | |
9771b263 | 4081 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
4082 | enum tree_code c1 = code1, c2 = code2; |
4083 | if (i == 0 && codecvt2 != ERROR_MARK) | |
4084 | { | |
4085 | c1 = codecvt1; | |
4086 | c2 = codecvt2; | |
4087 | } | |
4088 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
4089 | &vec_oprnds1, | |
4090 | stmt, this_dest, gsi, | |
4091 | c1, c2, decl1, decl2, | |
4092 | op_type); | |
4093 | } | |
4094 | ||
9771b263 | 4095 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4096 | { |
4097 | if (cvt_type) | |
4098 | { | |
4099 | if (codecvt1 == CALL_EXPR) | |
4100 | { | |
4101 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4102 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4103 | gimple_call_set_lhs (new_stmt, new_temp); | |
4104 | } | |
4105 | else | |
4106 | { | |
4107 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4108 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4109 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4110 | vop0); | |
4a00c761 JJ |
4111 | } |
4112 | ||
4113 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4114 | } | |
4115 | else | |
4116 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
4117 | ||
4118 | if (slp_node) | |
9771b263 | 4119 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4120 | else |
c689ce1e RB |
4121 | { |
4122 | if (!prev_stmt_info) | |
4123 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4124 | else | |
4125 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4126 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4127 | } | |
4a00c761 | 4128 | } |
ebfd146a | 4129 | } |
4a00c761 JJ |
4130 | |
4131 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
4132 | break; |
4133 | ||
4134 | case NARROW: | |
4135 | /* In case the vectorization factor (VF) is bigger than the number | |
4136 | of elements that we can fit in a vectype (nunits), we have to | |
4137 | generate more than one vector stmt - i.e - we need to "unroll" | |
4138 | the vector stmt by a factor VF/nunits. */ | |
4139 | for (j = 0; j < ncopies; j++) | |
4140 | { | |
4141 | /* Handle uses. */ | |
4a00c761 JJ |
4142 | if (slp_node) |
4143 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
4144 | slp_node, -1); | |
ebfd146a IR |
4145 | else |
4146 | { | |
9771b263 | 4147 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
4148 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
4149 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
4150 | } |
4151 | ||
4a00c761 JJ |
4152 | /* Arguments are ready. Create the new vector stmts. */ |
4153 | if (cvt_type) | |
9771b263 | 4154 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4155 | { |
4156 | if (codecvt1 == CALL_EXPR) | |
4157 | { | |
4158 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4159 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4160 | gimple_call_set_lhs (new_stmt, new_temp); | |
4161 | } | |
4162 | else | |
4163 | { | |
4164 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4165 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4166 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4167 | vop0); | |
4a00c761 | 4168 | } |
ebfd146a | 4169 | |
4a00c761 | 4170 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 4171 | vec_oprnds0[i] = new_temp; |
4a00c761 | 4172 | } |
ebfd146a | 4173 | |
4a00c761 JJ |
4174 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
4175 | stmt, vec_dsts, gsi, | |
4176 | slp_node, code1, | |
4177 | &prev_stmt_info); | |
ebfd146a IR |
4178 | } |
4179 | ||
4180 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 4181 | break; |
ebfd146a IR |
4182 | } |
4183 | ||
9771b263 DN |
4184 | vec_oprnds0.release (); |
4185 | vec_oprnds1.release (); | |
4186 | vec_dsts.release (); | |
4187 | interm_types.release (); | |
ebfd146a IR |
4188 | |
4189 | return true; | |
4190 | } | |
ff802fa1 IR |
4191 | |
4192 | ||
ebfd146a IR |
4193 | /* Function vectorizable_assignment. |
4194 | ||
b8698a0f L |
4195 | Check if STMT performs an assignment (copy) that can be vectorized. |
4196 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4197 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4198 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4199 | ||
4200 | static bool | |
355fe088 TS |
4201 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
4202 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4203 | { |
4204 | tree vec_dest; | |
4205 | tree scalar_dest; | |
4206 | tree op; | |
4207 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
4208 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4209 | tree new_temp; | |
355fe088 | 4210 | gimple *def_stmt; |
ebfd146a | 4211 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
ebfd146a | 4212 | int ncopies; |
f18b55bd | 4213 | int i, j; |
6e1aa848 | 4214 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 4215 | tree vop; |
a70d6342 | 4216 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4217 | vec_info *vinfo = stmt_info->vinfo; |
355fe088 | 4218 | gimple *new_stmt = NULL; |
f18b55bd | 4219 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
4220 | enum tree_code code; |
4221 | tree vectype_in; | |
ebfd146a | 4222 | |
a70d6342 | 4223 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4224 | return false; |
4225 | ||
66c16fd9 RB |
4226 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4227 | && ! vec_stmt) | |
ebfd146a IR |
4228 | return false; |
4229 | ||
4230 | /* Is vectorizable assignment? */ | |
4231 | if (!is_gimple_assign (stmt)) | |
4232 | return false; | |
4233 | ||
4234 | scalar_dest = gimple_assign_lhs (stmt); | |
4235 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4236 | return false; | |
4237 | ||
fde9c428 | 4238 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 4239 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
4240 | || code == PAREN_EXPR |
4241 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
4242 | op = gimple_assign_rhs1 (stmt); |
4243 | else | |
4244 | return false; | |
4245 | ||
7b7ec6c5 RG |
4246 | if (code == VIEW_CONVERT_EXPR) |
4247 | op = TREE_OPERAND (op, 0); | |
4248 | ||
465c8c19 JJ |
4249 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
4250 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4251 | ||
4252 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4253 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4254 | case of SLP. */ | |
fce57248 | 4255 | if (slp_node) |
465c8c19 JJ |
4256 | ncopies = 1; |
4257 | else | |
4258 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4259 | ||
4260 | gcc_assert (ncopies >= 1); | |
4261 | ||
81c40241 | 4262 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[0], &vectype_in)) |
ebfd146a | 4263 | { |
73fbfcad | 4264 | if (dump_enabled_p ()) |
78c60e3d | 4265 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4266 | "use not simple.\n"); |
ebfd146a IR |
4267 | return false; |
4268 | } | |
4269 | ||
fde9c428 RG |
4270 | /* We can handle NOP_EXPR conversions that do not change the number |
4271 | of elements or the vector size. */ | |
7b7ec6c5 RG |
4272 | if ((CONVERT_EXPR_CODE_P (code) |
4273 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
4274 | && (!vectype_in |
4275 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
4276 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
4277 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
4278 | return false; | |
4279 | ||
7b7b1813 RG |
4280 | /* We do not handle bit-precision changes. */ |
4281 | if ((CONVERT_EXPR_CODE_P (code) | |
4282 | || code == VIEW_CONVERT_EXPR) | |
4283 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
4284 | && ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4285 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4286 | || ((TYPE_PRECISION (TREE_TYPE (op)) | |
4287 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))) | |
4288 | /* But a conversion that does not change the bit-pattern is ok. */ | |
4289 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4290 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2dab46d5 IE |
4291 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
4292 | /* Conversion between boolean types of different sizes is | |
4293 | a simple assignment in case their vectypes are same | |
4294 | boolean vectors. */ | |
4295 | && (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
4296 | || !VECTOR_BOOLEAN_TYPE_P (vectype_in))) | |
7b7b1813 | 4297 | { |
73fbfcad | 4298 | if (dump_enabled_p ()) |
78c60e3d SS |
4299 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4300 | "type conversion to/from bit-precision " | |
e645e942 | 4301 | "unsupported.\n"); |
7b7b1813 RG |
4302 | return false; |
4303 | } | |
4304 | ||
ebfd146a IR |
4305 | if (!vec_stmt) /* transformation not required. */ |
4306 | { | |
4307 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 4308 | if (dump_enabled_p ()) |
78c60e3d | 4309 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4310 | "=== vectorizable_assignment ===\n"); |
c3e7ee41 | 4311 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
4312 | return true; |
4313 | } | |
4314 | ||
4315 | /** Transform. **/ | |
73fbfcad | 4316 | if (dump_enabled_p ()) |
e645e942 | 4317 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
4318 | |
4319 | /* Handle def. */ | |
4320 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4321 | ||
4322 | /* Handle use. */ | |
f18b55bd | 4323 | for (j = 0; j < ncopies; j++) |
ebfd146a | 4324 | { |
f18b55bd IR |
4325 | /* Handle uses. */ |
4326 | if (j == 0) | |
d092494c | 4327 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node, -1); |
f18b55bd IR |
4328 | else |
4329 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
4330 | ||
4331 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 4332 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 4333 | { |
7b7ec6c5 RG |
4334 | if (CONVERT_EXPR_CODE_P (code) |
4335 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 4336 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
4337 | new_stmt = gimple_build_assign (vec_dest, vop); |
4338 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4339 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4340 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4341 | if (slp_node) | |
9771b263 | 4342 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 4343 | } |
ebfd146a IR |
4344 | |
4345 | if (slp_node) | |
f18b55bd IR |
4346 | continue; |
4347 | ||
4348 | if (j == 0) | |
4349 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4350 | else | |
4351 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4352 | ||
4353 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4354 | } | |
b8698a0f | 4355 | |
9771b263 | 4356 | vec_oprnds.release (); |
ebfd146a IR |
4357 | return true; |
4358 | } | |
4359 | ||
9dc3f7de | 4360 | |
1107f3ae IR |
4361 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
4362 | either as shift by a scalar or by a vector. */ | |
4363 | ||
4364 | bool | |
4365 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
4366 | { | |
4367 | ||
ef4bddc2 | 4368 | machine_mode vec_mode; |
1107f3ae IR |
4369 | optab optab; |
4370 | int icode; | |
4371 | tree vectype; | |
4372 | ||
4373 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4374 | if (!vectype) | |
4375 | return false; | |
4376 | ||
4377 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4378 | if (!optab | |
4379 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
4380 | { | |
4381 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4382 | if (!optab | |
4383 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
4384 | == CODE_FOR_nothing)) | |
4385 | return false; | |
4386 | } | |
4387 | ||
4388 | vec_mode = TYPE_MODE (vectype); | |
4389 | icode = (int) optab_handler (optab, vec_mode); | |
4390 | if (icode == CODE_FOR_nothing) | |
4391 | return false; | |
4392 | ||
4393 | return true; | |
4394 | } | |
4395 | ||
4396 | ||
9dc3f7de IR |
4397 | /* Function vectorizable_shift. |
4398 | ||
4399 | Check if STMT performs a shift operation that can be vectorized. | |
4400 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4401 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
4402 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4403 | ||
4404 | static bool | |
355fe088 TS |
4405 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
4406 | gimple **vec_stmt, slp_tree slp_node) | |
9dc3f7de IR |
4407 | { |
4408 | tree vec_dest; | |
4409 | tree scalar_dest; | |
4410 | tree op0, op1 = NULL; | |
4411 | tree vec_oprnd1 = NULL_TREE; | |
4412 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4413 | tree vectype; | |
4414 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4415 | enum tree_code code; | |
ef4bddc2 | 4416 | machine_mode vec_mode; |
9dc3f7de IR |
4417 | tree new_temp; |
4418 | optab optab; | |
4419 | int icode; | |
ef4bddc2 | 4420 | machine_mode optab_op2_mode; |
355fe088 | 4421 | gimple *def_stmt; |
9dc3f7de | 4422 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
355fe088 | 4423 | gimple *new_stmt = NULL; |
9dc3f7de IR |
4424 | stmt_vec_info prev_stmt_info; |
4425 | int nunits_in; | |
4426 | int nunits_out; | |
4427 | tree vectype_out; | |
cede2577 | 4428 | tree op1_vectype; |
9dc3f7de IR |
4429 | int ncopies; |
4430 | int j, i; | |
6e1aa848 DN |
4431 | vec<tree> vec_oprnds0 = vNULL; |
4432 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
4433 | tree vop0, vop1; |
4434 | unsigned int k; | |
49eab32e | 4435 | bool scalar_shift_arg = true; |
9dc3f7de | 4436 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4437 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
4438 | int vf; |
4439 | ||
4440 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
4441 | return false; | |
4442 | ||
66c16fd9 RB |
4443 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4444 | && ! vec_stmt) | |
9dc3f7de IR |
4445 | return false; |
4446 | ||
4447 | /* Is STMT a vectorizable binary/unary operation? */ | |
4448 | if (!is_gimple_assign (stmt)) | |
4449 | return false; | |
4450 | ||
4451 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4452 | return false; | |
4453 | ||
4454 | code = gimple_assign_rhs_code (stmt); | |
4455 | ||
4456 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
4457 | || code == RROTATE_EXPR)) | |
4458 | return false; | |
4459 | ||
4460 | scalar_dest = gimple_assign_lhs (stmt); | |
4461 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
7b7b1813 RG |
4462 | if (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
4463 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4464 | { | |
73fbfcad | 4465 | if (dump_enabled_p ()) |
78c60e3d | 4466 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4467 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
4468 | return false; |
4469 | } | |
9dc3f7de IR |
4470 | |
4471 | op0 = gimple_assign_rhs1 (stmt); | |
81c40241 | 4472 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
9dc3f7de | 4473 | { |
73fbfcad | 4474 | if (dump_enabled_p ()) |
78c60e3d | 4475 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4476 | "use not simple.\n"); |
9dc3f7de IR |
4477 | return false; |
4478 | } | |
4479 | /* If op0 is an external or constant def use a vector type with | |
4480 | the same size as the output vector type. */ | |
4481 | if (!vectype) | |
4482 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4483 | if (vec_stmt) | |
4484 | gcc_assert (vectype); | |
4485 | if (!vectype) | |
4486 | { | |
73fbfcad | 4487 | if (dump_enabled_p ()) |
78c60e3d | 4488 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4489 | "no vectype for scalar type\n"); |
9dc3f7de IR |
4490 | return false; |
4491 | } | |
4492 | ||
4493 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4494 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4495 | if (nunits_out != nunits_in) | |
4496 | return false; | |
4497 | ||
4498 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4499 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &op1_vectype)) |
9dc3f7de | 4500 | { |
73fbfcad | 4501 | if (dump_enabled_p ()) |
78c60e3d | 4502 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4503 | "use not simple.\n"); |
9dc3f7de IR |
4504 | return false; |
4505 | } | |
4506 | ||
4507 | if (loop_vinfo) | |
4508 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4509 | else | |
4510 | vf = 1; | |
4511 | ||
4512 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4513 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4514 | case of SLP. */ | |
fce57248 | 4515 | if (slp_node) |
9dc3f7de IR |
4516 | ncopies = 1; |
4517 | else | |
4518 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4519 | ||
4520 | gcc_assert (ncopies >= 1); | |
4521 | ||
4522 | /* Determine whether the shift amount is a vector, or scalar. If the | |
4523 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
4524 | ||
dbfa87aa YR |
4525 | if ((dt[1] == vect_internal_def |
4526 | || dt[1] == vect_induction_def) | |
4527 | && !slp_node) | |
49eab32e JJ |
4528 | scalar_shift_arg = false; |
4529 | else if (dt[1] == vect_constant_def | |
4530 | || dt[1] == vect_external_def | |
4531 | || dt[1] == vect_internal_def) | |
4532 | { | |
4533 | /* In SLP, need to check whether the shift count is the same, | |
4534 | in loops if it is a constant or invariant, it is always | |
4535 | a scalar shift. */ | |
4536 | if (slp_node) | |
4537 | { | |
355fe088 TS |
4538 | vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
4539 | gimple *slpstmt; | |
49eab32e | 4540 | |
9771b263 | 4541 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
4542 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
4543 | scalar_shift_arg = false; | |
4544 | } | |
60d393e8 RB |
4545 | |
4546 | /* If the shift amount is computed by a pattern stmt we cannot | |
4547 | use the scalar amount directly thus give up and use a vector | |
4548 | shift. */ | |
4549 | if (dt[1] == vect_internal_def) | |
4550 | { | |
4551 | gimple *def = SSA_NAME_DEF_STMT (op1); | |
4552 | if (is_pattern_stmt_p (vinfo_for_stmt (def))) | |
4553 | scalar_shift_arg = false; | |
4554 | } | |
49eab32e JJ |
4555 | } |
4556 | else | |
4557 | { | |
73fbfcad | 4558 | if (dump_enabled_p ()) |
78c60e3d | 4559 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4560 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
4561 | return false; |
4562 | } | |
4563 | ||
9dc3f7de | 4564 | /* Vector shifted by vector. */ |
49eab32e | 4565 | if (!scalar_shift_arg) |
9dc3f7de IR |
4566 | { |
4567 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 4568 | if (dump_enabled_p ()) |
78c60e3d | 4569 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4570 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 4571 | |
aa948027 JJ |
4572 | if (!op1_vectype) |
4573 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
4574 | if (op1_vectype == NULL_TREE | |
4575 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 4576 | { |
73fbfcad | 4577 | if (dump_enabled_p ()) |
78c60e3d SS |
4578 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4579 | "unusable type for last operand in" | |
e645e942 | 4580 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
4581 | return false; |
4582 | } | |
9dc3f7de IR |
4583 | } |
4584 | /* See if the machine has a vector shifted by scalar insn and if not | |
4585 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 4586 | else |
9dc3f7de IR |
4587 | { |
4588 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4589 | if (optab | |
4590 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
4591 | { | |
73fbfcad | 4592 | if (dump_enabled_p ()) |
78c60e3d | 4593 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4594 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
4595 | } |
4596 | else | |
4597 | { | |
4598 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4599 | if (optab | |
4600 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
4601 | != CODE_FOR_nothing)) | |
4602 | { | |
49eab32e JJ |
4603 | scalar_shift_arg = false; |
4604 | ||
73fbfcad | 4605 | if (dump_enabled_p ()) |
78c60e3d | 4606 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4607 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
4608 | |
4609 | /* Unlike the other binary operators, shifts/rotates have | |
4610 | the rhs being int, instead of the same type as the lhs, | |
4611 | so make sure the scalar is the right type if we are | |
aa948027 | 4612 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
4613 | if (dt[1] == vect_constant_def) |
4614 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
4615 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
4616 | TREE_TYPE (op1))) | |
4617 | { | |
4618 | if (slp_node | |
4619 | && TYPE_MODE (TREE_TYPE (vectype)) | |
4620 | != TYPE_MODE (TREE_TYPE (op1))) | |
4621 | { | |
73fbfcad | 4622 | if (dump_enabled_p ()) |
78c60e3d SS |
4623 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4624 | "unusable type for last operand in" | |
e645e942 | 4625 | " vector/vector shift/rotate.\n"); |
21c0a521 | 4626 | return false; |
aa948027 JJ |
4627 | } |
4628 | if (vec_stmt && !slp_node) | |
4629 | { | |
4630 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
4631 | op1 = vect_init_vector (stmt, op1, | |
4632 | TREE_TYPE (vectype), NULL); | |
4633 | } | |
4634 | } | |
9dc3f7de IR |
4635 | } |
4636 | } | |
4637 | } | |
9dc3f7de IR |
4638 | |
4639 | /* Supportable by target? */ | |
4640 | if (!optab) | |
4641 | { | |
73fbfcad | 4642 | if (dump_enabled_p ()) |
78c60e3d | 4643 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4644 | "no optab.\n"); |
9dc3f7de IR |
4645 | return false; |
4646 | } | |
4647 | vec_mode = TYPE_MODE (vectype); | |
4648 | icode = (int) optab_handler (optab, vec_mode); | |
4649 | if (icode == CODE_FOR_nothing) | |
4650 | { | |
73fbfcad | 4651 | if (dump_enabled_p ()) |
78c60e3d | 4652 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4653 | "op not supported by target.\n"); |
9dc3f7de IR |
4654 | /* Check only during analysis. */ |
4655 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
4656 | || (vf < vect_min_worthwhile_factor (code) | |
4657 | && !vec_stmt)) | |
4658 | return false; | |
73fbfcad | 4659 | if (dump_enabled_p ()) |
e645e942 TJ |
4660 | dump_printf_loc (MSG_NOTE, vect_location, |
4661 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
4662 | } |
4663 | ||
4664 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
4665 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
4666 | && vf < vect_min_worthwhile_factor (code) | |
4667 | && !vec_stmt) | |
4668 | { | |
73fbfcad | 4669 | if (dump_enabled_p ()) |
78c60e3d | 4670 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4671 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
4672 | return false; |
4673 | } | |
4674 | ||
4675 | if (!vec_stmt) /* transformation not required. */ | |
4676 | { | |
4677 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 4678 | if (dump_enabled_p ()) |
e645e942 TJ |
4679 | dump_printf_loc (MSG_NOTE, vect_location, |
4680 | "=== vectorizable_shift ===\n"); | |
c3e7ee41 | 4681 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
9dc3f7de IR |
4682 | return true; |
4683 | } | |
4684 | ||
4685 | /** Transform. **/ | |
4686 | ||
73fbfcad | 4687 | if (dump_enabled_p ()) |
78c60e3d | 4688 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4689 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
4690 | |
4691 | /* Handle def. */ | |
4692 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4693 | ||
9dc3f7de IR |
4694 | prev_stmt_info = NULL; |
4695 | for (j = 0; j < ncopies; j++) | |
4696 | { | |
4697 | /* Handle uses. */ | |
4698 | if (j == 0) | |
4699 | { | |
4700 | if (scalar_shift_arg) | |
4701 | { | |
4702 | /* Vector shl and shr insn patterns can be defined with scalar | |
4703 | operand 2 (shift operand). In this case, use constant or loop | |
4704 | invariant op1 directly, without extending it to vector mode | |
4705 | first. */ | |
4706 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
4707 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
4708 | { | |
73fbfcad | 4709 | if (dump_enabled_p ()) |
78c60e3d | 4710 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4711 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 4712 | vec_oprnd1 = op1; |
8930f723 | 4713 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 4714 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4715 | if (slp_node) |
4716 | { | |
4717 | /* Store vec_oprnd1 for every vector stmt to be created | |
4718 | for SLP_NODE. We check during the analysis that all | |
4719 | the shift arguments are the same. | |
4720 | TODO: Allow different constants for different vector | |
4721 | stmts generated for an SLP instance. */ | |
4722 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4723 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4724 | } |
4725 | } | |
4726 | } | |
4727 | ||
4728 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
4729 | (a special case for certain kind of vector shifts); otherwise, | |
4730 | operand 1 should be of a vector type (the usual case). */ | |
4731 | if (vec_oprnd1) | |
4732 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
d092494c | 4733 | slp_node, -1); |
9dc3f7de IR |
4734 | else |
4735 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
d092494c | 4736 | slp_node, -1); |
9dc3f7de IR |
4737 | } |
4738 | else | |
4739 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
4740 | ||
4741 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 4742 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 4743 | { |
9771b263 | 4744 | vop1 = vec_oprnds1[i]; |
0d0e4a03 | 4745 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
4746 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4747 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4748 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4749 | if (slp_node) | |
9771b263 | 4750 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
4751 | } |
4752 | ||
4753 | if (slp_node) | |
4754 | continue; | |
4755 | ||
4756 | if (j == 0) | |
4757 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4758 | else | |
4759 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4760 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4761 | } | |
4762 | ||
9771b263 DN |
4763 | vec_oprnds0.release (); |
4764 | vec_oprnds1.release (); | |
9dc3f7de IR |
4765 | |
4766 | return true; | |
4767 | } | |
4768 | ||
4769 | ||
ebfd146a IR |
4770 | /* Function vectorizable_operation. |
4771 | ||
16949072 RG |
4772 | Check if STMT performs a binary, unary or ternary operation that can |
4773 | be vectorized. | |
b8698a0f | 4774 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
4775 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4776 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4777 | ||
4778 | static bool | |
355fe088 TS |
4779 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
4780 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a | 4781 | { |
00f07b86 | 4782 | tree vec_dest; |
ebfd146a | 4783 | tree scalar_dest; |
16949072 | 4784 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 4785 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 4786 | tree vectype; |
ebfd146a IR |
4787 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4788 | enum tree_code code; | |
ef4bddc2 | 4789 | machine_mode vec_mode; |
ebfd146a IR |
4790 | tree new_temp; |
4791 | int op_type; | |
00f07b86 | 4792 | optab optab; |
523ba738 | 4793 | bool target_support_p; |
355fe088 | 4794 | gimple *def_stmt; |
16949072 RG |
4795 | enum vect_def_type dt[3] |
4796 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
355fe088 | 4797 | gimple *new_stmt = NULL; |
ebfd146a | 4798 | stmt_vec_info prev_stmt_info; |
b690cc0f | 4799 | int nunits_in; |
ebfd146a IR |
4800 | int nunits_out; |
4801 | tree vectype_out; | |
4802 | int ncopies; | |
4803 | int j, i; | |
6e1aa848 DN |
4804 | vec<tree> vec_oprnds0 = vNULL; |
4805 | vec<tree> vec_oprnds1 = vNULL; | |
4806 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 4807 | tree vop0, vop1, vop2; |
a70d6342 | 4808 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4809 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 IR |
4810 | int vf; |
4811 | ||
a70d6342 | 4812 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4813 | return false; |
4814 | ||
66c16fd9 RB |
4815 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4816 | && ! vec_stmt) | |
ebfd146a IR |
4817 | return false; |
4818 | ||
4819 | /* Is STMT a vectorizable binary/unary operation? */ | |
4820 | if (!is_gimple_assign (stmt)) | |
4821 | return false; | |
4822 | ||
4823 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4824 | return false; | |
4825 | ||
ebfd146a IR |
4826 | code = gimple_assign_rhs_code (stmt); |
4827 | ||
4828 | /* For pointer addition, we should use the normal plus for | |
4829 | the vector addition. */ | |
4830 | if (code == POINTER_PLUS_EXPR) | |
4831 | code = PLUS_EXPR; | |
4832 | ||
4833 | /* Support only unary or binary operations. */ | |
4834 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 4835 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 4836 | { |
73fbfcad | 4837 | if (dump_enabled_p ()) |
78c60e3d | 4838 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4839 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 4840 | op_type); |
ebfd146a IR |
4841 | return false; |
4842 | } | |
4843 | ||
b690cc0f RG |
4844 | scalar_dest = gimple_assign_lhs (stmt); |
4845 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
4846 | ||
7b7b1813 RG |
4847 | /* Most operations cannot handle bit-precision types without extra |
4848 | truncations. */ | |
045c1278 IE |
4849 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4850 | && (TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4851 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
7b7b1813 RG |
4852 | /* Exception are bitwise binary operations. */ |
4853 | && code != BIT_IOR_EXPR | |
4854 | && code != BIT_XOR_EXPR | |
4855 | && code != BIT_AND_EXPR) | |
4856 | { | |
73fbfcad | 4857 | if (dump_enabled_p ()) |
78c60e3d | 4858 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4859 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
4860 | return false; |
4861 | } | |
4862 | ||
ebfd146a | 4863 | op0 = gimple_assign_rhs1 (stmt); |
81c40241 | 4864 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
ebfd146a | 4865 | { |
73fbfcad | 4866 | if (dump_enabled_p ()) |
78c60e3d | 4867 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4868 | "use not simple.\n"); |
ebfd146a IR |
4869 | return false; |
4870 | } | |
b690cc0f RG |
4871 | /* If op0 is an external or constant def use a vector type with |
4872 | the same size as the output vector type. */ | |
4873 | if (!vectype) | |
b036c6c5 IE |
4874 | { |
4875 | /* For boolean type we cannot determine vectype by | |
4876 | invariant value (don't know whether it is a vector | |
4877 | of booleans or vector of integers). We use output | |
4878 | vectype because operations on boolean don't change | |
4879 | type. */ | |
4880 | if (TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE) | |
4881 | { | |
4882 | if (TREE_CODE (TREE_TYPE (scalar_dest)) != BOOLEAN_TYPE) | |
4883 | { | |
4884 | if (dump_enabled_p ()) | |
4885 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4886 | "not supported operation on bool value.\n"); | |
4887 | return false; | |
4888 | } | |
4889 | vectype = vectype_out; | |
4890 | } | |
4891 | else | |
4892 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4893 | } | |
7d8930a0 IR |
4894 | if (vec_stmt) |
4895 | gcc_assert (vectype); | |
4896 | if (!vectype) | |
4897 | { | |
73fbfcad | 4898 | if (dump_enabled_p ()) |
7d8930a0 | 4899 | { |
78c60e3d SS |
4900 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4901 | "no vectype for scalar type "); | |
4902 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
4903 | TREE_TYPE (op0)); | |
e645e942 | 4904 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
4905 | } |
4906 | ||
4907 | return false; | |
4908 | } | |
b690cc0f RG |
4909 | |
4910 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4911 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4912 | if (nunits_out != nunits_in) | |
4913 | return false; | |
ebfd146a | 4914 | |
16949072 | 4915 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
4916 | { |
4917 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4918 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1])) |
ebfd146a | 4919 | { |
73fbfcad | 4920 | if (dump_enabled_p ()) |
78c60e3d | 4921 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4922 | "use not simple.\n"); |
ebfd146a IR |
4923 | return false; |
4924 | } | |
4925 | } | |
16949072 RG |
4926 | if (op_type == ternary_op) |
4927 | { | |
4928 | op2 = gimple_assign_rhs3 (stmt); | |
81c40241 | 4929 | if (!vect_is_simple_use (op2, vinfo, &def_stmt, &dt[2])) |
16949072 | 4930 | { |
73fbfcad | 4931 | if (dump_enabled_p ()) |
78c60e3d | 4932 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4933 | "use not simple.\n"); |
16949072 RG |
4934 | return false; |
4935 | } | |
4936 | } | |
ebfd146a | 4937 | |
b690cc0f RG |
4938 | if (loop_vinfo) |
4939 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4940 | else | |
4941 | vf = 1; | |
4942 | ||
4943 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4944 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 4945 | case of SLP. */ |
fce57248 | 4946 | if (slp_node) |
b690cc0f RG |
4947 | ncopies = 1; |
4948 | else | |
4949 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4950 | ||
4951 | gcc_assert (ncopies >= 1); | |
4952 | ||
9dc3f7de | 4953 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
4954 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
4955 | || code == RROTATE_EXPR) | |
9dc3f7de | 4956 | return false; |
ebfd146a | 4957 | |
ebfd146a | 4958 | /* Supportable by target? */ |
00f07b86 RH |
4959 | |
4960 | vec_mode = TYPE_MODE (vectype); | |
4961 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 4962 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
4963 | else |
4964 | { | |
4965 | optab = optab_for_tree_code (code, vectype, optab_default); | |
4966 | if (!optab) | |
5deb57cb | 4967 | { |
73fbfcad | 4968 | if (dump_enabled_p ()) |
78c60e3d | 4969 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4970 | "no optab.\n"); |
00f07b86 | 4971 | return false; |
5deb57cb | 4972 | } |
523ba738 RS |
4973 | target_support_p = (optab_handler (optab, vec_mode) |
4974 | != CODE_FOR_nothing); | |
5deb57cb JJ |
4975 | } |
4976 | ||
523ba738 | 4977 | if (!target_support_p) |
ebfd146a | 4978 | { |
73fbfcad | 4979 | if (dump_enabled_p ()) |
78c60e3d | 4980 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4981 | "op not supported by target.\n"); |
ebfd146a IR |
4982 | /* Check only during analysis. */ |
4983 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5deb57cb | 4984 | || (!vec_stmt && vf < vect_min_worthwhile_factor (code))) |
ebfd146a | 4985 | return false; |
73fbfcad | 4986 | if (dump_enabled_p ()) |
e645e942 TJ |
4987 | dump_printf_loc (MSG_NOTE, vect_location, |
4988 | "proceeding using word mode.\n"); | |
383d9c83 IR |
4989 | } |
4990 | ||
4a00c761 | 4991 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
4992 | if (!VECTOR_MODE_P (vec_mode) |
4993 | && !vec_stmt | |
4994 | && vf < vect_min_worthwhile_factor (code)) | |
7d8930a0 | 4995 | { |
73fbfcad | 4996 | if (dump_enabled_p ()) |
78c60e3d | 4997 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4998 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 4999 | return false; |
7d8930a0 | 5000 | } |
ebfd146a | 5001 | |
ebfd146a IR |
5002 | if (!vec_stmt) /* transformation not required. */ |
5003 | { | |
4a00c761 | 5004 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 5005 | if (dump_enabled_p ()) |
78c60e3d | 5006 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5007 | "=== vectorizable_operation ===\n"); |
c3e7ee41 | 5008 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
5009 | return true; |
5010 | } | |
5011 | ||
5012 | /** Transform. **/ | |
5013 | ||
73fbfcad | 5014 | if (dump_enabled_p ()) |
78c60e3d | 5015 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5016 | "transform binary/unary operation.\n"); |
383d9c83 | 5017 | |
ebfd146a | 5018 | /* Handle def. */ |
00f07b86 | 5019 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 5020 | |
ebfd146a IR |
5021 | /* In case the vectorization factor (VF) is bigger than the number |
5022 | of elements that we can fit in a vectype (nunits), we have to generate | |
5023 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
5024 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
5025 | from one copy of the vector stmt to the next, in the field | |
5026 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
5027 | stages to find the correct vector defs to be used when vectorizing | |
5028 | stmts that use the defs of the current stmt. The example below | |
5029 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
5030 | we need to create 4 vectorized stmts): | |
5031 | ||
5032 | before vectorization: | |
5033 | RELATED_STMT VEC_STMT | |
5034 | S1: x = memref - - | |
5035 | S2: z = x + 1 - - | |
5036 | ||
5037 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
5038 | there): | |
5039 | RELATED_STMT VEC_STMT | |
5040 | VS1_0: vx0 = memref0 VS1_1 - | |
5041 | VS1_1: vx1 = memref1 VS1_2 - | |
5042 | VS1_2: vx2 = memref2 VS1_3 - | |
5043 | VS1_3: vx3 = memref3 - - | |
5044 | S1: x = load - VS1_0 | |
5045 | S2: z = x + 1 - - | |
5046 | ||
5047 | step2: vectorize stmt S2 (done here): | |
5048 | To vectorize stmt S2 we first need to find the relevant vector | |
5049 | def for the first operand 'x'. This is, as usual, obtained from | |
5050 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
5051 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
5052 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
5053 | the vector stmt VS2_0, and as usual, record it in the | |
5054 | STMT_VINFO_VEC_STMT of stmt S2. | |
5055 | When creating the second copy (VS2_1), we obtain the relevant vector | |
5056 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
5057 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
5058 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
5059 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
5060 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
5061 | chain of stmts and pointers: | |
5062 | RELATED_STMT VEC_STMT | |
5063 | VS1_0: vx0 = memref0 VS1_1 - | |
5064 | VS1_1: vx1 = memref1 VS1_2 - | |
5065 | VS1_2: vx2 = memref2 VS1_3 - | |
5066 | VS1_3: vx3 = memref3 - - | |
5067 | S1: x = load - VS1_0 | |
5068 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
5069 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
5070 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
5071 | VS2_3: vz3 = vx3 + v1 - - | |
5072 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
5073 | |
5074 | prev_stmt_info = NULL; | |
5075 | for (j = 0; j < ncopies; j++) | |
5076 | { | |
5077 | /* Handle uses. */ | |
5078 | if (j == 0) | |
4a00c761 JJ |
5079 | { |
5080 | if (op_type == binary_op || op_type == ternary_op) | |
5081 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
5082 | slp_node, -1); | |
5083 | else | |
5084 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
5085 | slp_node, -1); | |
5086 | if (op_type == ternary_op) | |
36ba4aae | 5087 | { |
9771b263 DN |
5088 | vec_oprnds2.create (1); |
5089 | vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2, | |
81c40241 | 5090 | stmt)); |
36ba4aae | 5091 | } |
4a00c761 | 5092 | } |
ebfd146a | 5093 | else |
4a00c761 JJ |
5094 | { |
5095 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5096 | if (op_type == ternary_op) | |
5097 | { | |
9771b263 DN |
5098 | tree vec_oprnd = vec_oprnds2.pop (); |
5099 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
5100 | vec_oprnd)); | |
4a00c761 JJ |
5101 | } |
5102 | } | |
5103 | ||
5104 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5105 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 5106 | { |
4a00c761 | 5107 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 5108 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 5109 | vop2 = ((op_type == ternary_op) |
9771b263 | 5110 | ? vec_oprnds2[i] : NULL_TREE); |
0d0e4a03 | 5111 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
4a00c761 JJ |
5112 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5113 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5114 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5115 | if (slp_node) | |
9771b263 | 5116 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
5117 | } |
5118 | ||
4a00c761 JJ |
5119 | if (slp_node) |
5120 | continue; | |
5121 | ||
5122 | if (j == 0) | |
5123 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5124 | else | |
5125 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5126 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
5127 | } |
5128 | ||
9771b263 DN |
5129 | vec_oprnds0.release (); |
5130 | vec_oprnds1.release (); | |
5131 | vec_oprnds2.release (); | |
ebfd146a | 5132 | |
ebfd146a IR |
5133 | return true; |
5134 | } | |
5135 | ||
c716e67f XDL |
5136 | /* A helper function to ensure data reference DR's base alignment |
5137 | for STMT_INFO. */ | |
5138 | ||
5139 | static void | |
5140 | ensure_base_align (stmt_vec_info stmt_info, struct data_reference *dr) | |
5141 | { | |
5142 | if (!dr->aux) | |
5143 | return; | |
5144 | ||
52639a61 | 5145 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f XDL |
5146 | { |
5147 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
52639a61 | 5148 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 5149 | |
428f0c67 JH |
5150 | if (decl_in_symtab_p (base_decl)) |
5151 | symtab_node::get (base_decl)->increase_alignment (TYPE_ALIGN (vectype)); | |
5152 | else | |
5153 | { | |
fe37c7af | 5154 | SET_DECL_ALIGN (base_decl, TYPE_ALIGN (vectype)); |
428f0c67 JH |
5155 | DECL_USER_ALIGN (base_decl) = 1; |
5156 | } | |
52639a61 | 5157 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
5158 | } |
5159 | } | |
5160 | ||
ebfd146a | 5161 | |
09dfa495 BM |
5162 | /* Given a vector type VECTYPE returns the VECTOR_CST mask that implements |
5163 | reversal of the vector elements. If that is impossible to do, | |
5164 | returns NULL. */ | |
5165 | ||
5166 | static tree | |
5167 | perm_mask_for_reverse (tree vectype) | |
5168 | { | |
5169 | int i, nunits; | |
5170 | unsigned char *sel; | |
5171 | ||
5172 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
5173 | sel = XALLOCAVEC (unsigned char, nunits); | |
5174 | ||
5175 | for (i = 0; i < nunits; ++i) | |
5176 | sel[i] = nunits - 1 - i; | |
5177 | ||
557be5a8 AL |
5178 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) |
5179 | return NULL_TREE; | |
5180 | return vect_gen_perm_mask_checked (vectype, sel); | |
09dfa495 BM |
5181 | } |
5182 | ||
ebfd146a IR |
5183 | /* Function vectorizable_store. |
5184 | ||
b8698a0f L |
5185 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
5186 | can be vectorized. | |
5187 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5188 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5189 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5190 | ||
5191 | static bool | |
355fe088 | 5192 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 5193 | slp_tree slp_node) |
ebfd146a IR |
5194 | { |
5195 | tree scalar_dest; | |
5196 | tree data_ref; | |
5197 | tree op; | |
5198 | tree vec_oprnd = NULL_TREE; | |
5199 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5200 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 5201 | tree elem_type; |
ebfd146a | 5202 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 5203 | struct loop *loop = NULL; |
ef4bddc2 | 5204 | machine_mode vec_mode; |
ebfd146a IR |
5205 | tree dummy; |
5206 | enum dr_alignment_support alignment_support_scheme; | |
355fe088 | 5207 | gimple *def_stmt; |
ebfd146a IR |
5208 | enum vect_def_type dt; |
5209 | stmt_vec_info prev_stmt_info = NULL; | |
5210 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 5211 | tree dataref_offset = NULL_TREE; |
355fe088 | 5212 | gimple *ptr_incr = NULL; |
ebfd146a IR |
5213 | int ncopies; |
5214 | int j; | |
355fe088 | 5215 | gimple *next_stmt, *first_stmt = NULL; |
0d0293ac | 5216 | bool grouped_store = false; |
272c6793 | 5217 | bool store_lanes_p = false; |
ebfd146a | 5218 | unsigned int group_size, i; |
6e1aa848 DN |
5219 | vec<tree> dr_chain = vNULL; |
5220 | vec<tree> oprnds = vNULL; | |
5221 | vec<tree> result_chain = vNULL; | |
ebfd146a | 5222 | bool inv_p; |
09dfa495 BM |
5223 | bool negative = false; |
5224 | tree offset = NULL_TREE; | |
6e1aa848 | 5225 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5226 | bool slp = (slp_node != NULL); |
ebfd146a | 5227 | unsigned int vec_num; |
a70d6342 | 5228 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5229 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 5230 | tree aggr_type; |
3bab6342 AT |
5231 | tree scatter_base = NULL_TREE, scatter_off = NULL_TREE; |
5232 | tree scatter_off_vectype = NULL_TREE, scatter_decl = NULL_TREE; | |
5233 | int scatter_scale = 1; | |
5234 | enum vect_def_type scatter_idx_dt = vect_unknown_def_type; | |
5235 | enum vect_def_type scatter_src_dt = vect_unknown_def_type; | |
355fe088 | 5236 | gimple *new_stmt; |
a70d6342 | 5237 | |
a70d6342 | 5238 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5239 | return false; |
5240 | ||
66c16fd9 RB |
5241 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5242 | && ! vec_stmt) | |
ebfd146a IR |
5243 | return false; |
5244 | ||
5245 | /* Is vectorizable store? */ | |
5246 | ||
5247 | if (!is_gimple_assign (stmt)) | |
5248 | return false; | |
5249 | ||
5250 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
5251 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
5252 | && is_pattern_stmt_p (stmt_info)) | |
5253 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 5254 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 5255 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 5256 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
5257 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
5258 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
5259 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
5260 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
5261 | return false; |
5262 | ||
fce57248 RS |
5263 | /* Cannot have hybrid store SLP -- that would mean storing to the |
5264 | same location twice. */ | |
5265 | gcc_assert (slp == PURE_SLP_STMT (stmt_info)); | |
5266 | ||
ebfd146a | 5267 | gcc_assert (gimple_assign_single_p (stmt)); |
465c8c19 | 5268 | |
f4d09712 | 5269 | tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE; |
465c8c19 JJ |
5270 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
5271 | ||
5272 | if (loop_vinfo) | |
5273 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
5274 | ||
5275 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5276 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5277 | case of SLP. */ | |
fce57248 | 5278 | if (slp) |
465c8c19 JJ |
5279 | ncopies = 1; |
5280 | else | |
5281 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
5282 | ||
5283 | gcc_assert (ncopies >= 1); | |
5284 | ||
5285 | /* FORNOW. This restriction should be relaxed. */ | |
5286 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
5287 | { | |
5288 | if (dump_enabled_p ()) | |
5289 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5290 | "multiple types in nested loop.\n"); | |
5291 | return false; | |
5292 | } | |
5293 | ||
ebfd146a | 5294 | op = gimple_assign_rhs1 (stmt); |
f4d09712 KY |
5295 | |
5296 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt, &rhs_vectype)) | |
ebfd146a | 5297 | { |
73fbfcad | 5298 | if (dump_enabled_p ()) |
78c60e3d | 5299 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5300 | "use not simple.\n"); |
ebfd146a IR |
5301 | return false; |
5302 | } | |
5303 | ||
f4d09712 KY |
5304 | if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype)) |
5305 | return false; | |
5306 | ||
272c6793 | 5307 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 5308 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 5309 | |
ebfd146a IR |
5310 | /* FORNOW. In some cases can vectorize even if data-type not supported |
5311 | (e.g. - array initialization with 0). */ | |
947131ba | 5312 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
5313 | return false; |
5314 | ||
5315 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
5316 | return false; | |
5317 | ||
f2e2a985 | 5318 | if (!STMT_VINFO_STRIDED_P (stmt_info)) |
09dfa495 | 5319 | { |
f2e2a985 MM |
5320 | negative = |
5321 | tree_int_cst_compare (loop && nested_in_vect_loop_p (loop, stmt) | |
5322 | ? STMT_VINFO_DR_STEP (stmt_info) : DR_STEP (dr), | |
5323 | size_zero_node) < 0; | |
5324 | if (negative && ncopies > 1) | |
09dfa495 BM |
5325 | { |
5326 | if (dump_enabled_p ()) | |
5327 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
f2e2a985 | 5328 | "multiple types with negative step.\n"); |
09dfa495 BM |
5329 | return false; |
5330 | } | |
f2e2a985 | 5331 | if (negative) |
09dfa495 | 5332 | { |
f2e2a985 MM |
5333 | gcc_assert (!grouped_store); |
5334 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
5335 | if (alignment_support_scheme != dr_aligned | |
5336 | && alignment_support_scheme != dr_unaligned_supported) | |
5337 | { | |
5338 | if (dump_enabled_p ()) | |
5339 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5340 | "negative step but alignment required.\n"); | |
5341 | return false; | |
5342 | } | |
5343 | if (dt != vect_constant_def | |
5344 | && dt != vect_external_def | |
5345 | && !perm_mask_for_reverse (vectype)) | |
5346 | { | |
5347 | if (dump_enabled_p ()) | |
5348 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5349 | "negative step and reversing not supported.\n"); | |
5350 | return false; | |
5351 | } | |
09dfa495 BM |
5352 | } |
5353 | } | |
5354 | ||
0d0293ac | 5355 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 5356 | { |
0d0293ac | 5357 | grouped_store = true; |
e14c1050 | 5358 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
cee62fee | 5359 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
fce57248 | 5360 | if (!slp && !STMT_VINFO_STRIDED_P (stmt_info)) |
b602d918 | 5361 | { |
272c6793 RS |
5362 | if (vect_store_lanes_supported (vectype, group_size)) |
5363 | store_lanes_p = true; | |
0d0293ac | 5364 | else if (!vect_grouped_store_supported (vectype, group_size)) |
b602d918 RS |
5365 | return false; |
5366 | } | |
b8698a0f | 5367 | |
cee62fee | 5368 | if (STMT_VINFO_STRIDED_P (stmt_info) |
fce57248 | 5369 | && slp |
cee62fee MM |
5370 | && (group_size > nunits |
5371 | || nunits % group_size != 0)) | |
5372 | { | |
5373 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5374 | "unhandled strided group store\n"); | |
5375 | return false; | |
5376 | } | |
5377 | ||
ebfd146a IR |
5378 | if (first_stmt == stmt) |
5379 | { | |
5380 | /* STMT is the leader of the group. Check the operands of all the | |
5381 | stmts of the group. */ | |
e14c1050 | 5382 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
5383 | while (next_stmt) |
5384 | { | |
5385 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5386 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 5387 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) |
ebfd146a | 5388 | { |
73fbfcad | 5389 | if (dump_enabled_p ()) |
78c60e3d | 5390 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5391 | "use not simple.\n"); |
ebfd146a IR |
5392 | return false; |
5393 | } | |
e14c1050 | 5394 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5395 | } |
5396 | } | |
5397 | } | |
5398 | ||
3bab6342 AT |
5399 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5400 | { | |
355fe088 | 5401 | gimple *def_stmt; |
3bab6342 AT |
5402 | scatter_decl = vect_check_gather_scatter (stmt, loop_vinfo, &scatter_base, |
5403 | &scatter_off, &scatter_scale); | |
5404 | gcc_assert (scatter_decl); | |
81c40241 RB |
5405 | if (!vect_is_simple_use (scatter_off, vinfo, &def_stmt, &scatter_idx_dt, |
5406 | &scatter_off_vectype)) | |
3bab6342 AT |
5407 | { |
5408 | if (dump_enabled_p ()) | |
5409 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5410 | "scatter index use not simple."); | |
5411 | return false; | |
5412 | } | |
5413 | } | |
5414 | ||
ebfd146a IR |
5415 | if (!vec_stmt) /* transformation not required. */ |
5416 | { | |
5417 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
2e8ab70c RB |
5418 | /* The SLP costs are calculated during SLP analysis. */ |
5419 | if (!PURE_SLP_STMT (stmt_info)) | |
5420 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, | |
5421 | NULL, NULL, NULL); | |
ebfd146a IR |
5422 | return true; |
5423 | } | |
5424 | ||
5425 | /** Transform. **/ | |
5426 | ||
c716e67f XDL |
5427 | ensure_base_align (stmt_info, dr); |
5428 | ||
3bab6342 AT |
5429 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5430 | { | |
5431 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, op, src; | |
5432 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (scatter_decl)); | |
5433 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
5434 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
5435 | edge pe = loop_preheader_edge (loop); | |
5436 | gimple_seq seq; | |
5437 | basic_block new_bb; | |
5438 | enum { NARROW, NONE, WIDEN } modifier; | |
5439 | int scatter_off_nunits = TYPE_VECTOR_SUBPARTS (scatter_off_vectype); | |
5440 | ||
5441 | if (nunits == (unsigned int) scatter_off_nunits) | |
5442 | modifier = NONE; | |
5443 | else if (nunits == (unsigned int) scatter_off_nunits / 2) | |
5444 | { | |
5445 | unsigned char *sel = XALLOCAVEC (unsigned char, scatter_off_nunits); | |
5446 | modifier = WIDEN; | |
5447 | ||
5448 | for (i = 0; i < (unsigned int) scatter_off_nunits; ++i) | |
5449 | sel[i] = i | nunits; | |
5450 | ||
5451 | perm_mask = vect_gen_perm_mask_checked (scatter_off_vectype, sel); | |
5452 | gcc_assert (perm_mask != NULL_TREE); | |
5453 | } | |
5454 | else if (nunits == (unsigned int) scatter_off_nunits * 2) | |
5455 | { | |
5456 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
5457 | modifier = NARROW; | |
5458 | ||
5459 | for (i = 0; i < (unsigned int) nunits; ++i) | |
5460 | sel[i] = i | scatter_off_nunits; | |
5461 | ||
5462 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); | |
5463 | gcc_assert (perm_mask != NULL_TREE); | |
5464 | ncopies *= 2; | |
5465 | } | |
5466 | else | |
5467 | gcc_unreachable (); | |
5468 | ||
5469 | rettype = TREE_TYPE (TREE_TYPE (scatter_decl)); | |
5470 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5471 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5472 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5473 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5474 | scaletype = TREE_VALUE (arglist); | |
5475 | ||
5476 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
5477 | && TREE_CODE (rettype) == VOID_TYPE); | |
5478 | ||
5479 | ptr = fold_convert (ptrtype, scatter_base); | |
5480 | if (!is_gimple_min_invariant (ptr)) | |
5481 | { | |
5482 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
5483 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
5484 | gcc_assert (!new_bb); | |
5485 | } | |
5486 | ||
5487 | /* Currently we support only unconditional scatter stores, | |
5488 | so mask should be all ones. */ | |
5489 | mask = build_int_cst (masktype, -1); | |
5490 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
5491 | ||
5492 | scale = build_int_cst (scaletype, scatter_scale); | |
5493 | ||
5494 | prev_stmt_info = NULL; | |
5495 | for (j = 0; j < ncopies; ++j) | |
5496 | { | |
5497 | if (j == 0) | |
5498 | { | |
5499 | src = vec_oprnd1 | |
81c40241 | 5500 | = vect_get_vec_def_for_operand (gimple_assign_rhs1 (stmt), stmt); |
3bab6342 | 5501 | op = vec_oprnd0 |
81c40241 | 5502 | = vect_get_vec_def_for_operand (scatter_off, stmt); |
3bab6342 AT |
5503 | } |
5504 | else if (modifier != NONE && (j & 1)) | |
5505 | { | |
5506 | if (modifier == WIDEN) | |
5507 | { | |
5508 | src = vec_oprnd1 | |
5509 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5510 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, | |
5511 | stmt, gsi); | |
5512 | } | |
5513 | else if (modifier == NARROW) | |
5514 | { | |
5515 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
5516 | stmt, gsi); | |
5517 | op = vec_oprnd0 | |
5518 | = vect_get_vec_def_for_stmt_copy (scatter_idx_dt, vec_oprnd0); | |
5519 | } | |
5520 | else | |
5521 | gcc_unreachable (); | |
5522 | } | |
5523 | else | |
5524 | { | |
5525 | src = vec_oprnd1 | |
5526 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5527 | op = vec_oprnd0 | |
5528 | = vect_get_vec_def_for_stmt_copy (scatter_idx_dt, vec_oprnd0); | |
5529 | } | |
5530 | ||
5531 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
5532 | { | |
5533 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)) | |
5534 | == TYPE_VECTOR_SUBPARTS (srctype)); | |
0e22bb5a | 5535 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 AT |
5536 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
5537 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
5538 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5539 | src = var; | |
5540 | } | |
5541 | ||
5542 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
5543 | { | |
5544 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
5545 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 5546 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 AT |
5547 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
5548 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
5549 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5550 | op = var; | |
5551 | } | |
5552 | ||
5553 | new_stmt | |
5554 | = gimple_build_call (scatter_decl, 5, ptr, mask, op, src, scale); | |
5555 | ||
5556 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5557 | ||
5558 | if (prev_stmt_info == NULL) | |
5559 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5560 | else | |
5561 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5562 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5563 | } | |
5564 | return true; | |
5565 | } | |
5566 | ||
0d0293ac | 5567 | if (grouped_store) |
ebfd146a IR |
5568 | { |
5569 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 5570 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 5571 | |
e14c1050 | 5572 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
5573 | |
5574 | /* FORNOW */ | |
a70d6342 | 5575 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
5576 | |
5577 | /* We vectorize all the stmts of the interleaving group when we | |
5578 | reach the last stmt in the group. */ | |
e14c1050 IR |
5579 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
5580 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
5581 | && !slp) |
5582 | { | |
5583 | *vec_stmt = NULL; | |
5584 | return true; | |
5585 | } | |
5586 | ||
5587 | if (slp) | |
4b5caab7 | 5588 | { |
0d0293ac | 5589 | grouped_store = false; |
4b5caab7 IR |
5590 | /* VEC_NUM is the number of vect stmts to be created for this |
5591 | group. */ | |
5592 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 5593 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
52eab378 | 5594 | gcc_assert (GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt)) == first_stmt); |
4b5caab7 | 5595 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 5596 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 5597 | } |
ebfd146a | 5598 | else |
4b5caab7 IR |
5599 | /* VEC_NUM is the number of vect stmts to be created for this |
5600 | group. */ | |
ebfd146a IR |
5601 | vec_num = group_size; |
5602 | } | |
b8698a0f | 5603 | else |
ebfd146a IR |
5604 | { |
5605 | first_stmt = stmt; | |
5606 | first_dr = dr; | |
5607 | group_size = vec_num = 1; | |
ebfd146a | 5608 | } |
b8698a0f | 5609 | |
73fbfcad | 5610 | if (dump_enabled_p ()) |
78c60e3d | 5611 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5612 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 5613 | |
f2e2a985 MM |
5614 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
5615 | { | |
5616 | gimple_stmt_iterator incr_gsi; | |
5617 | bool insert_after; | |
355fe088 | 5618 | gimple *incr; |
f2e2a985 MM |
5619 | tree offvar; |
5620 | tree ivstep; | |
5621 | tree running_off; | |
5622 | gimple_seq stmts = NULL; | |
5623 | tree stride_base, stride_step, alias_off; | |
5624 | tree vec_oprnd; | |
f502d50e | 5625 | unsigned int g; |
f2e2a985 MM |
5626 | |
5627 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); | |
5628 | ||
5629 | stride_base | |
5630 | = fold_build_pointer_plus | |
f502d50e | 5631 | (unshare_expr (DR_BASE_ADDRESS (first_dr)), |
f2e2a985 | 5632 | size_binop (PLUS_EXPR, |
f502d50e MM |
5633 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (first_dr))), |
5634 | convert_to_ptrofftype (DR_INIT(first_dr)))); | |
5635 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (first_dr))); | |
f2e2a985 MM |
5636 | |
5637 | /* For a store with loop-invariant (but other than power-of-2) | |
5638 | stride (i.e. not a grouped access) like so: | |
5639 | ||
5640 | for (i = 0; i < n; i += stride) | |
5641 | array[i] = ...; | |
5642 | ||
5643 | we generate a new induction variable and new stores from | |
5644 | the components of the (vectorized) rhs: | |
5645 | ||
5646 | for (j = 0; ; j += VF*stride) | |
5647 | vectemp = ...; | |
5648 | tmp1 = vectemp[0]; | |
5649 | array[j] = tmp1; | |
5650 | tmp2 = vectemp[1]; | |
5651 | array[j + stride] = tmp2; | |
5652 | ... | |
5653 | */ | |
5654 | ||
cee62fee MM |
5655 | unsigned nstores = nunits; |
5656 | tree ltype = elem_type; | |
5657 | if (slp) | |
5658 | { | |
5659 | nstores = nunits / group_size; | |
5660 | if (group_size < nunits) | |
5661 | ltype = build_vector_type (elem_type, group_size); | |
5662 | else | |
5663 | ltype = vectype; | |
5664 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); | |
5665 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
f502d50e | 5666 | group_size = 1; |
cee62fee MM |
5667 | } |
5668 | ||
f2e2a985 MM |
5669 | ivstep = stride_step; |
5670 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
5671 | build_int_cst (TREE_TYPE (ivstep), | |
cee62fee | 5672 | ncopies * nstores)); |
f2e2a985 MM |
5673 | |
5674 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
5675 | ||
5676 | create_iv (stride_base, ivstep, NULL, | |
5677 | loop, &incr_gsi, insert_after, | |
5678 | &offvar, NULL); | |
5679 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 5680 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
f2e2a985 MM |
5681 | |
5682 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
5683 | if (stmts) | |
5684 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
5685 | ||
5686 | prev_stmt_info = NULL; | |
f502d50e MM |
5687 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (first_dr)), 0); |
5688 | next_stmt = first_stmt; | |
5689 | for (g = 0; g < group_size; g++) | |
f2e2a985 | 5690 | { |
f502d50e MM |
5691 | running_off = offvar; |
5692 | if (g) | |
f2e2a985 | 5693 | { |
f502d50e MM |
5694 | tree size = TYPE_SIZE_UNIT (ltype); |
5695 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 5696 | size); |
f502d50e | 5697 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 5698 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 5699 | running_off, pos); |
f2e2a985 | 5700 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 5701 | running_off = newoff; |
f502d50e MM |
5702 | } |
5703 | for (j = 0; j < ncopies; j++) | |
5704 | { | |
5705 | /* We've set op and dt above, from gimple_assign_rhs1(stmt), | |
5706 | and first_stmt == stmt. */ | |
5707 | if (j == 0) | |
5708 | { | |
5709 | if (slp) | |
5710 | { | |
5711 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
5712 | slp_node, -1); | |
5713 | vec_oprnd = vec_oprnds[0]; | |
5714 | } | |
5715 | else | |
5716 | { | |
5717 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5718 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 5719 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
f502d50e MM |
5720 | } |
5721 | } | |
f2e2a985 | 5722 | else |
f502d50e MM |
5723 | { |
5724 | if (slp) | |
5725 | vec_oprnd = vec_oprnds[j]; | |
5726 | else | |
c079cbac | 5727 | { |
81c40241 | 5728 | vect_is_simple_use (vec_oprnd, vinfo, &def_stmt, &dt); |
c079cbac RB |
5729 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); |
5730 | } | |
f502d50e MM |
5731 | } |
5732 | ||
5733 | for (i = 0; i < nstores; i++) | |
5734 | { | |
5735 | tree newref, newoff; | |
355fe088 | 5736 | gimple *incr, *assign; |
f502d50e MM |
5737 | tree size = TYPE_SIZE (ltype); |
5738 | /* Extract the i'th component. */ | |
5739 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
5740 | bitsize_int (i), size); | |
5741 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
5742 | size, pos); | |
5743 | ||
5744 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
5745 | NULL_TREE, true, | |
5746 | GSI_SAME_STMT); | |
5747 | ||
5748 | newref = build2 (MEM_REF, ltype, | |
5749 | running_off, alias_off); | |
5750 | ||
5751 | /* And store it to *running_off. */ | |
5752 | assign = gimple_build_assign (newref, elem); | |
5753 | vect_finish_stmt_generation (stmt, assign, gsi); | |
5754 | ||
5755 | newoff = copy_ssa_name (running_off, NULL); | |
5756 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
5757 | running_off, stride_step); | |
5758 | vect_finish_stmt_generation (stmt, incr, gsi); | |
5759 | ||
5760 | running_off = newoff; | |
225ce44b RB |
5761 | if (g == group_size - 1 |
5762 | && !slp) | |
f502d50e MM |
5763 | { |
5764 | if (j == 0 && i == 0) | |
225ce44b RB |
5765 | STMT_VINFO_VEC_STMT (stmt_info) |
5766 | = *vec_stmt = assign; | |
f502d50e MM |
5767 | else |
5768 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign; | |
5769 | prev_stmt_info = vinfo_for_stmt (assign); | |
5770 | } | |
5771 | } | |
f2e2a985 | 5772 | } |
f502d50e | 5773 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
f2e2a985 MM |
5774 | } |
5775 | return true; | |
5776 | } | |
5777 | ||
9771b263 DN |
5778 | dr_chain.create (group_size); |
5779 | oprnds.create (group_size); | |
ebfd146a | 5780 | |
720f5239 | 5781 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 5782 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
5783 | /* Targets with store-lane instructions must not require explicit |
5784 | realignment. */ | |
5785 | gcc_assert (!store_lanes_p | |
5786 | || alignment_support_scheme == dr_aligned | |
5787 | || alignment_support_scheme == dr_unaligned_supported); | |
5788 | ||
09dfa495 BM |
5789 | if (negative) |
5790 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
5791 | ||
272c6793 RS |
5792 | if (store_lanes_p) |
5793 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
5794 | else | |
5795 | aggr_type = vectype; | |
ebfd146a IR |
5796 | |
5797 | /* In case the vectorization factor (VF) is bigger than the number | |
5798 | of elements that we can fit in a vectype (nunits), we have to generate | |
5799 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 5800 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
5801 | vect_get_vec_def_for_copy_stmt. */ |
5802 | ||
0d0293ac | 5803 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
5804 | |
5805 | S1: &base + 2 = x2 | |
5806 | S2: &base = x0 | |
5807 | S3: &base + 1 = x1 | |
5808 | S4: &base + 3 = x3 | |
5809 | ||
5810 | We create vectorized stores starting from base address (the access of the | |
5811 | first stmt in the chain (S2 in the above example), when the last store stmt | |
5812 | of the chain (S4) is reached: | |
5813 | ||
5814 | VS1: &base = vx2 | |
5815 | VS2: &base + vec_size*1 = vx0 | |
5816 | VS3: &base + vec_size*2 = vx1 | |
5817 | VS4: &base + vec_size*3 = vx3 | |
5818 | ||
5819 | Then permutation statements are generated: | |
5820 | ||
3fcc1b55 JJ |
5821 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
5822 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 5823 | ... |
b8698a0f | 5824 | |
ebfd146a IR |
5825 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
5826 | (the order of the data-refs in the output of vect_permute_store_chain | |
5827 | corresponds to the order of scalar stmts in the interleaving chain - see | |
5828 | the documentation of vect_permute_store_chain()). | |
5829 | ||
5830 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 5831 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 5832 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 5833 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
5834 | */ |
5835 | ||
5836 | prev_stmt_info = NULL; | |
5837 | for (j = 0; j < ncopies; j++) | |
5838 | { | |
ebfd146a IR |
5839 | |
5840 | if (j == 0) | |
5841 | { | |
5842 | if (slp) | |
5843 | { | |
5844 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c IR |
5845 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
5846 | NULL, slp_node, -1); | |
ebfd146a | 5847 | |
9771b263 | 5848 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
5849 | } |
5850 | else | |
5851 | { | |
b8698a0f L |
5852 | /* For interleaved stores we collect vectorized defs for all the |
5853 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
5854 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
5855 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
5856 | ||
0d0293ac | 5857 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 5858 | OPRNDS are of size 1. */ |
b8698a0f | 5859 | next_stmt = first_stmt; |
ebfd146a IR |
5860 | for (i = 0; i < group_size; i++) |
5861 | { | |
b8698a0f L |
5862 | /* Since gaps are not supported for interleaved stores, |
5863 | GROUP_SIZE is the exact number of stmts in the chain. | |
5864 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
5865 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
5866 | iteration of the loop will be executed. */ |
5867 | gcc_assert (next_stmt | |
5868 | && gimple_assign_single_p (next_stmt)); | |
5869 | op = gimple_assign_rhs1 (next_stmt); | |
5870 | ||
81c40241 | 5871 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
9771b263 DN |
5872 | dr_chain.quick_push (vec_oprnd); |
5873 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 5874 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5875 | } |
5876 | } | |
5877 | ||
5878 | /* We should have catched mismatched types earlier. */ | |
5879 | gcc_assert (useless_type_conversion_p (vectype, | |
5880 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
5881 | bool simd_lane_access_p |
5882 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
5883 | if (simd_lane_access_p | |
5884 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
5885 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
5886 | && integer_zerop (DR_OFFSET (first_dr)) | |
5887 | && integer_zerop (DR_INIT (first_dr)) | |
5888 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
5889 | get_alias_set (DR_REF (first_dr)))) | |
5890 | { | |
5891 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
5892 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
5893 | (DR_REF (first_dr)), 0); | |
8928eff3 | 5894 | inv_p = false; |
74bf76ed JJ |
5895 | } |
5896 | else | |
5897 | dataref_ptr | |
5898 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
5899 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 5900 | offset, &dummy, gsi, &ptr_incr, |
74bf76ed | 5901 | simd_lane_access_p, &inv_p); |
a70d6342 | 5902 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 5903 | } |
b8698a0f | 5904 | else |
ebfd146a | 5905 | { |
b8698a0f L |
5906 | /* For interleaved stores we created vectorized defs for all the |
5907 | defs stored in OPRNDS in the previous iteration (previous copy). | |
5908 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
5909 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
5910 | next copy. | |
0d0293ac | 5911 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
5912 | OPRNDS are of size 1. */ |
5913 | for (i = 0; i < group_size; i++) | |
5914 | { | |
9771b263 | 5915 | op = oprnds[i]; |
81c40241 | 5916 | vect_is_simple_use (op, vinfo, &def_stmt, &dt); |
b8698a0f | 5917 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
5918 | dr_chain[i] = vec_oprnd; |
5919 | oprnds[i] = vec_oprnd; | |
ebfd146a | 5920 | } |
74bf76ed JJ |
5921 | if (dataref_offset) |
5922 | dataref_offset | |
5923 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
5924 | TYPE_SIZE_UNIT (aggr_type)); | |
5925 | else | |
5926 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
5927 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
5928 | } |
5929 | ||
272c6793 | 5930 | if (store_lanes_p) |
ebfd146a | 5931 | { |
272c6793 | 5932 | tree vec_array; |
267d3070 | 5933 | |
272c6793 RS |
5934 | /* Combine all the vectors into an array. */ |
5935 | vec_array = create_vector_array (vectype, vec_num); | |
5936 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 5937 | { |
9771b263 | 5938 | vec_oprnd = dr_chain[i]; |
272c6793 | 5939 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 5940 | } |
b8698a0f | 5941 | |
272c6793 RS |
5942 | /* Emit: |
5943 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
5944 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
5945 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
5946 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 | 5947 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
5948 | } |
5949 | else | |
5950 | { | |
5951 | new_stmt = NULL; | |
0d0293ac | 5952 | if (grouped_store) |
272c6793 | 5953 | { |
b6b9227d JJ |
5954 | if (j == 0) |
5955 | result_chain.create (group_size); | |
272c6793 RS |
5956 | /* Permute. */ |
5957 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
5958 | &result_chain); | |
5959 | } | |
c2d7ab2a | 5960 | |
272c6793 RS |
5961 | next_stmt = first_stmt; |
5962 | for (i = 0; i < vec_num; i++) | |
5963 | { | |
644ffefd | 5964 | unsigned align, misalign; |
272c6793 RS |
5965 | |
5966 | if (i > 0) | |
5967 | /* Bump the vector pointer. */ | |
5968 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5969 | stmt, NULL_TREE); | |
5970 | ||
5971 | if (slp) | |
9771b263 | 5972 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
5973 | else if (grouped_store) |
5974 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 5975 | vect_permute_store_chain(). */ |
9771b263 | 5976 | vec_oprnd = result_chain[i]; |
272c6793 | 5977 | |
aed93b23 RB |
5978 | data_ref = fold_build2 (MEM_REF, TREE_TYPE (vec_oprnd), |
5979 | dataref_ptr, | |
5980 | dataref_offset | |
5981 | ? dataref_offset | |
5982 | : build_int_cst (reference_alias_ptr_type | |
5983 | (DR_REF (first_dr)), 0)); | |
644ffefd | 5984 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 | 5985 | if (aligned_access_p (first_dr)) |
644ffefd | 5986 | misalign = 0; |
272c6793 RS |
5987 | else if (DR_MISALIGNMENT (first_dr) == -1) |
5988 | { | |
52639a61 RB |
5989 | if (DR_VECT_AUX (first_dr)->base_element_aligned) |
5990 | align = TYPE_ALIGN_UNIT (elem_type); | |
5991 | else | |
5992 | align = get_object_alignment (DR_REF (first_dr)) | |
5993 | / BITS_PER_UNIT; | |
5994 | misalign = 0; | |
272c6793 RS |
5995 | TREE_TYPE (data_ref) |
5996 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 5997 | align * BITS_PER_UNIT); |
272c6793 RS |
5998 | } |
5999 | else | |
6000 | { | |
6001 | TREE_TYPE (data_ref) | |
6002 | = build_aligned_type (TREE_TYPE (data_ref), | |
6003 | TYPE_ALIGN (elem_type)); | |
644ffefd | 6004 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 6005 | } |
aed93b23 RB |
6006 | if (dataref_offset == NULL_TREE |
6007 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
6008 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
6009 | misalign); | |
c2d7ab2a | 6010 | |
f234d260 BM |
6011 | if (negative |
6012 | && dt != vect_constant_def | |
6013 | && dt != vect_external_def) | |
09dfa495 BM |
6014 | { |
6015 | tree perm_mask = perm_mask_for_reverse (vectype); | |
6016 | tree perm_dest | |
6017 | = vect_create_destination_var (gimple_assign_rhs1 (stmt), | |
6018 | vectype); | |
b731b390 | 6019 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
6020 | |
6021 | /* Generate the permute statement. */ | |
355fe088 | 6022 | gimple *perm_stmt |
0d0e4a03 JJ |
6023 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
6024 | vec_oprnd, perm_mask); | |
09dfa495 BM |
6025 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6026 | ||
6027 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
6028 | vec_oprnd = new_temp; | |
6029 | } | |
6030 | ||
272c6793 RS |
6031 | /* Arguments are ready. Create the new vector stmt. */ |
6032 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
6033 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
6034 | |
6035 | if (slp) | |
6036 | continue; | |
6037 | ||
e14c1050 | 6038 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
6039 | if (!next_stmt) |
6040 | break; | |
6041 | } | |
ebfd146a | 6042 | } |
1da0876c RS |
6043 | if (!slp) |
6044 | { | |
6045 | if (j == 0) | |
6046 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6047 | else | |
6048 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6049 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6050 | } | |
ebfd146a IR |
6051 | } |
6052 | ||
9771b263 DN |
6053 | dr_chain.release (); |
6054 | oprnds.release (); | |
6055 | result_chain.release (); | |
6056 | vec_oprnds.release (); | |
ebfd146a IR |
6057 | |
6058 | return true; | |
6059 | } | |
6060 | ||
557be5a8 AL |
6061 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
6062 | VECTOR_CST mask. No checks are made that the target platform supports the | |
6063 | mask, so callers may wish to test can_vec_perm_p separately, or use | |
6064 | vect_gen_perm_mask_checked. */ | |
a1e53f3f | 6065 | |
3fcc1b55 | 6066 | tree |
557be5a8 | 6067 | vect_gen_perm_mask_any (tree vectype, const unsigned char *sel) |
a1e53f3f | 6068 | { |
d2a12ae7 | 6069 | tree mask_elt_type, mask_type, mask_vec, *mask_elts; |
2635892a | 6070 | int i, nunits; |
a1e53f3f | 6071 | |
22e4dee7 | 6072 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 | 6073 | |
96f9265a RG |
6074 | mask_elt_type = lang_hooks.types.type_for_mode |
6075 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
22e4dee7 | 6076 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 6077 | |
d2a12ae7 | 6078 | mask_elts = XALLOCAVEC (tree, nunits); |
aec7ae7d | 6079 | for (i = nunits - 1; i >= 0; i--) |
d2a12ae7 RG |
6080 | mask_elts[i] = build_int_cst (mask_elt_type, sel[i]); |
6081 | mask_vec = build_vector (mask_type, mask_elts); | |
a1e53f3f | 6082 | |
2635892a | 6083 | return mask_vec; |
a1e53f3f L |
6084 | } |
6085 | ||
cf7aa6a3 AL |
6086 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_p, |
6087 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ | |
557be5a8 AL |
6088 | |
6089 | tree | |
6090 | vect_gen_perm_mask_checked (tree vectype, const unsigned char *sel) | |
6091 | { | |
6092 | gcc_assert (can_vec_perm_p (TYPE_MODE (vectype), false, sel)); | |
6093 | return vect_gen_perm_mask_any (vectype, sel); | |
6094 | } | |
6095 | ||
aec7ae7d JJ |
6096 | /* Given a vector variable X and Y, that was generated for the scalar |
6097 | STMT, generate instructions to permute the vector elements of X and Y | |
6098 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
6099 | permuted vector variable. */ | |
a1e53f3f L |
6100 | |
6101 | static tree | |
355fe088 | 6102 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 6103 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
6104 | { |
6105 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 6106 | tree perm_dest, data_ref; |
355fe088 | 6107 | gimple *perm_stmt; |
a1e53f3f | 6108 | |
acdcd61b | 6109 | perm_dest = vect_create_destination_var (gimple_get_lhs (stmt), vectype); |
b731b390 | 6110 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
6111 | |
6112 | /* Generate the permute statement. */ | |
0d0e4a03 | 6113 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
6114 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6115 | ||
6116 | return data_ref; | |
6117 | } | |
6118 | ||
6b916b36 RB |
6119 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
6120 | inserting them on the loops preheader edge. Returns true if we | |
6121 | were successful in doing so (and thus STMT can be moved then), | |
6122 | otherwise returns false. */ | |
6123 | ||
6124 | static bool | |
355fe088 | 6125 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
6126 | { |
6127 | ssa_op_iter i; | |
6128 | tree op; | |
6129 | bool any = false; | |
6130 | ||
6131 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6132 | { | |
355fe088 | 6133 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6134 | if (!gimple_nop_p (def_stmt) |
6135 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6136 | { | |
6137 | /* Make sure we don't need to recurse. While we could do | |
6138 | so in simple cases when there are more complex use webs | |
6139 | we don't have an easy way to preserve stmt order to fulfil | |
6140 | dependencies within them. */ | |
6141 | tree op2; | |
6142 | ssa_op_iter i2; | |
d1417442 JJ |
6143 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
6144 | return false; | |
6b916b36 RB |
6145 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
6146 | { | |
355fe088 | 6147 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
6148 | if (!gimple_nop_p (def_stmt2) |
6149 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
6150 | return false; | |
6151 | } | |
6152 | any = true; | |
6153 | } | |
6154 | } | |
6155 | ||
6156 | if (!any) | |
6157 | return true; | |
6158 | ||
6159 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6160 | { | |
355fe088 | 6161 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6162 | if (!gimple_nop_p (def_stmt) |
6163 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6164 | { | |
6165 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
6166 | gsi_remove (&gsi, false); | |
6167 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
6168 | } | |
6169 | } | |
6170 | ||
6171 | return true; | |
6172 | } | |
6173 | ||
ebfd146a IR |
6174 | /* vectorizable_load. |
6175 | ||
b8698a0f L |
6176 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
6177 | can be vectorized. | |
6178 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6179 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6180 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6181 | ||
6182 | static bool | |
355fe088 | 6183 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 6184 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
6185 | { |
6186 | tree scalar_dest; | |
6187 | tree vec_dest = NULL; | |
6188 | tree data_ref = NULL; | |
6189 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 6190 | stmt_vec_info prev_stmt_info; |
ebfd146a | 6191 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6192 | struct loop *loop = NULL; |
ebfd146a | 6193 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 6194 | bool nested_in_vect_loop = false; |
c716e67f | 6195 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 6196 | tree elem_type; |
ebfd146a | 6197 | tree new_temp; |
ef4bddc2 | 6198 | machine_mode mode; |
355fe088 | 6199 | gimple *new_stmt = NULL; |
ebfd146a IR |
6200 | tree dummy; |
6201 | enum dr_alignment_support alignment_support_scheme; | |
6202 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6203 | tree dataref_offset = NULL_TREE; |
355fe088 | 6204 | gimple *ptr_incr = NULL; |
ebfd146a | 6205 | int ncopies; |
9b999e8c | 6206 | int i, j, group_size = -1, group_gap_adj; |
ebfd146a IR |
6207 | tree msq = NULL_TREE, lsq; |
6208 | tree offset = NULL_TREE; | |
356bbc4c | 6209 | tree byte_offset = NULL_TREE; |
ebfd146a | 6210 | tree realignment_token = NULL_TREE; |
538dd0b7 | 6211 | gphi *phi = NULL; |
6e1aa848 | 6212 | vec<tree> dr_chain = vNULL; |
0d0293ac | 6213 | bool grouped_load = false; |
272c6793 | 6214 | bool load_lanes_p = false; |
355fe088 | 6215 | gimple *first_stmt; |
4f0a0218 | 6216 | gimple *first_stmt_for_drptr = NULL; |
ebfd146a | 6217 | bool inv_p; |
319e6439 | 6218 | bool negative = false; |
ebfd146a IR |
6219 | bool compute_in_loop = false; |
6220 | struct loop *at_loop; | |
6221 | int vec_num; | |
6222 | bool slp = (slp_node != NULL); | |
6223 | bool slp_perm = false; | |
6224 | enum tree_code code; | |
a70d6342 IR |
6225 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
6226 | int vf; | |
272c6793 | 6227 | tree aggr_type; |
aec7ae7d JJ |
6228 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; |
6229 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
6230 | int gather_scale = 1; | |
6231 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
310213d4 | 6232 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 6233 | |
465c8c19 JJ |
6234 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
6235 | return false; | |
6236 | ||
66c16fd9 RB |
6237 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6238 | && ! vec_stmt) | |
465c8c19 JJ |
6239 | return false; |
6240 | ||
6241 | /* Is vectorizable load? */ | |
6242 | if (!is_gimple_assign (stmt)) | |
6243 | return false; | |
6244 | ||
6245 | scalar_dest = gimple_assign_lhs (stmt); | |
6246 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
6247 | return false; | |
6248 | ||
6249 | code = gimple_assign_rhs_code (stmt); | |
6250 | if (code != ARRAY_REF | |
6251 | && code != BIT_FIELD_REF | |
6252 | && code != INDIRECT_REF | |
6253 | && code != COMPONENT_REF | |
6254 | && code != IMAGPART_EXPR | |
6255 | && code != REALPART_EXPR | |
6256 | && code != MEM_REF | |
6257 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
6258 | return false; | |
6259 | ||
6260 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
6261 | return false; | |
6262 | ||
6263 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
6264 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
6265 | ||
a70d6342 IR |
6266 | if (loop_vinfo) |
6267 | { | |
6268 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6269 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
6270 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6271 | } | |
6272 | else | |
3533e503 | 6273 | vf = 1; |
ebfd146a IR |
6274 | |
6275 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 6276 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 6277 | case of SLP. */ |
fce57248 | 6278 | if (slp) |
ebfd146a IR |
6279 | ncopies = 1; |
6280 | else | |
6281 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
6282 | ||
6283 | gcc_assert (ncopies >= 1); | |
6284 | ||
6285 | /* FORNOW. This restriction should be relaxed. */ | |
6286 | if (nested_in_vect_loop && ncopies > 1) | |
6287 | { | |
73fbfcad | 6288 | if (dump_enabled_p ()) |
78c60e3d | 6289 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6290 | "multiple types in nested loop.\n"); |
ebfd146a IR |
6291 | return false; |
6292 | } | |
6293 | ||
f2556b68 RB |
6294 | /* Invalidate assumptions made by dependence analysis when vectorization |
6295 | on the unrolled body effectively re-orders stmts. */ | |
6296 | if (ncopies > 1 | |
6297 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6298 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6299 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6300 | { | |
6301 | if (dump_enabled_p ()) | |
6302 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6303 | "cannot perform implicit CSE when unrolling " | |
6304 | "with negative dependence distance\n"); | |
6305 | return false; | |
6306 | } | |
6307 | ||
7b7b1813 | 6308 | elem_type = TREE_TYPE (vectype); |
947131ba | 6309 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
6310 | |
6311 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6312 | (e.g. - data copies). */ | |
947131ba | 6313 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 6314 | { |
73fbfcad | 6315 | if (dump_enabled_p ()) |
78c60e3d | 6316 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6317 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
6318 | return false; |
6319 | } | |
6320 | ||
ebfd146a | 6321 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 6322 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 6323 | { |
0d0293ac | 6324 | grouped_load = true; |
ebfd146a | 6325 | /* FORNOW */ |
3bab6342 | 6326 | gcc_assert (!nested_in_vect_loop && !STMT_VINFO_GATHER_SCATTER_P (stmt_info)); |
ebfd146a | 6327 | |
e14c1050 | 6328 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
d3465d72 RS |
6329 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
6330 | ||
fce57248 | 6331 | if (!slp && !STMT_VINFO_STRIDED_P (stmt_info)) |
d3465d72 RS |
6332 | { |
6333 | if (vect_load_lanes_supported (vectype, group_size)) | |
6334 | load_lanes_p = true; | |
6335 | else if (!vect_grouped_load_supported (vectype, group_size)) | |
6336 | return false; | |
6337 | } | |
d5f035ea RB |
6338 | |
6339 | /* If this is single-element interleaving with an element distance | |
6340 | that leaves unused vector loads around punt - we at least create | |
6341 | very sub-optimal code in that case (and blow up memory, | |
6342 | see PR65518). */ | |
6343 | if (first_stmt == stmt | |
72c0f643 RB |
6344 | && !GROUP_NEXT_ELEMENT (stmt_info)) |
6345 | { | |
6346 | if (GROUP_SIZE (stmt_info) > TYPE_VECTOR_SUBPARTS (vectype)) | |
6347 | { | |
6348 | if (dump_enabled_p ()) | |
6349 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6350 | "single-element interleaving not supported " | |
6351 | "for not adjacent vector loads\n"); | |
6352 | return false; | |
6353 | } | |
6354 | ||
6355 | /* Single-element interleaving requires peeling for gaps. */ | |
836dbb1a | 6356 | gcc_assert (GROUP_GAP (stmt_info)); |
72c0f643 RB |
6357 | } |
6358 | ||
6359 | /* If there is a gap in the end of the group or the group size cannot | |
6360 | be made a multiple of the vector element count then we access excess | |
6361 | elements in the last iteration and thus need to peel that off. */ | |
6362 | if (loop_vinfo | |
6363 | && ! STMT_VINFO_STRIDED_P (stmt_info) | |
836dbb1a | 6364 | && (GROUP_GAP (vinfo_for_stmt (first_stmt)) != 0 |
d3465d72 | 6365 | || (!slp && !load_lanes_p && vf % group_size != 0))) |
d5f035ea RB |
6366 | { |
6367 | if (dump_enabled_p ()) | |
6368 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
72c0f643 RB |
6369 | "Data access with gaps requires scalar " |
6370 | "epilogue loop\n"); | |
6371 | if (loop->inner) | |
6372 | { | |
6373 | if (dump_enabled_p ()) | |
6374 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6375 | "Peeling for outer loop is not supported\n"); | |
6376 | return false; | |
6377 | } | |
6378 | ||
6379 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; | |
d5f035ea RB |
6380 | } |
6381 | ||
b1af7da6 RB |
6382 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
6383 | slp_perm = true; | |
6384 | ||
47d3fdb2 RB |
6385 | /* ??? The following is overly pessimistic (as well as the loop |
6386 | case above) in the case we can statically determine the excess | |
6387 | elements loaded are within the bounds of a decl that is accessed. | |
6388 | Likewise for BB vectorizations using masked loads is a possibility. */ | |
6389 | if (bb_vinfo && slp_perm && group_size % nunits != 0) | |
6390 | { | |
6391 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6392 | "BB vectorization with gaps at the end of a load " | |
6393 | "is not supported\n"); | |
6394 | return false; | |
6395 | } | |
6396 | ||
f2556b68 RB |
6397 | /* Invalidate assumptions made by dependence analysis when vectorization |
6398 | on the unrolled body effectively re-orders stmts. */ | |
6399 | if (!PURE_SLP_STMT (stmt_info) | |
6400 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6401 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6402 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6403 | { | |
6404 | if (dump_enabled_p ()) | |
6405 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6406 | "cannot perform implicit CSE when performing " | |
6407 | "group loads with negative dependence distance\n"); | |
6408 | return false; | |
6409 | } | |
96bb56b2 RB |
6410 | |
6411 | /* Similarly when the stmt is a load that is both part of a SLP | |
6412 | instance and a loop vectorized stmt via the same-dr mechanism | |
6413 | we have to give up. */ | |
6414 | if (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info) | |
6415 | && (STMT_SLP_TYPE (stmt_info) | |
6416 | != STMT_SLP_TYPE (vinfo_for_stmt | |
6417 | (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info))))) | |
6418 | { | |
6419 | if (dump_enabled_p ()) | |
6420 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6421 | "conflicting SLP types for CSEd load\n"); | |
6422 | return false; | |
6423 | } | |
ebfd146a IR |
6424 | } |
6425 | ||
a1e53f3f | 6426 | |
3bab6342 | 6427 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
aec7ae7d | 6428 | { |
355fe088 | 6429 | gimple *def_stmt; |
3bab6342 AT |
6430 | gather_decl = vect_check_gather_scatter (stmt, loop_vinfo, &gather_base, |
6431 | &gather_off, &gather_scale); | |
aec7ae7d | 6432 | gcc_assert (gather_decl); |
81c40241 RB |
6433 | if (!vect_is_simple_use (gather_off, vinfo, &def_stmt, &gather_dt, |
6434 | &gather_off_vectype)) | |
aec7ae7d | 6435 | { |
73fbfcad | 6436 | if (dump_enabled_p ()) |
78c60e3d | 6437 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6438 | "gather index use not simple.\n"); |
aec7ae7d JJ |
6439 | return false; |
6440 | } | |
6441 | } | |
f2e2a985 | 6442 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
7b5fc413 | 6443 | { |
fce57248 RS |
6444 | if (grouped_load |
6445 | && slp | |
7b5fc413 | 6446 | && (group_size > nunits |
b266b968 | 6447 | || nunits % group_size != 0)) |
7b5fc413 RB |
6448 | { |
6449 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6450 | "unhandled strided group load\n"); | |
6451 | return false; | |
6452 | } | |
6453 | } | |
319e6439 RG |
6454 | else |
6455 | { | |
6456 | negative = tree_int_cst_compare (nested_in_vect_loop | |
6457 | ? STMT_VINFO_DR_STEP (stmt_info) | |
6458 | : DR_STEP (dr), | |
6459 | size_zero_node) < 0; | |
6460 | if (negative && ncopies > 1) | |
6461 | { | |
73fbfcad | 6462 | if (dump_enabled_p ()) |
78c60e3d | 6463 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6464 | "multiple types with negative step.\n"); |
319e6439 RG |
6465 | return false; |
6466 | } | |
6467 | ||
6468 | if (negative) | |
6469 | { | |
08940f33 RB |
6470 | if (grouped_load) |
6471 | { | |
6472 | if (dump_enabled_p ()) | |
6473 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 TJ |
6474 | "negative step for group load not supported" |
6475 | "\n"); | |
08940f33 RB |
6476 | return false; |
6477 | } | |
319e6439 RG |
6478 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
6479 | if (alignment_support_scheme != dr_aligned | |
6480 | && alignment_support_scheme != dr_unaligned_supported) | |
6481 | { | |
73fbfcad | 6482 | if (dump_enabled_p ()) |
78c60e3d | 6483 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6484 | "negative step but alignment required.\n"); |
319e6439 RG |
6485 | return false; |
6486 | } | |
6487 | if (!perm_mask_for_reverse (vectype)) | |
6488 | { | |
73fbfcad | 6489 | if (dump_enabled_p ()) |
78c60e3d | 6490 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
6491 | "negative step and reversing not supported." |
6492 | "\n"); | |
319e6439 RG |
6493 | return false; |
6494 | } | |
6495 | } | |
7d75abc8 | 6496 | } |
aec7ae7d | 6497 | |
ebfd146a IR |
6498 | if (!vec_stmt) /* transformation not required. */ |
6499 | { | |
6500 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
2e8ab70c RB |
6501 | /* The SLP costs are calculated during SLP analysis. */ |
6502 | if (!PURE_SLP_STMT (stmt_info)) | |
6503 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, | |
6504 | NULL, NULL, NULL); | |
ebfd146a IR |
6505 | return true; |
6506 | } | |
6507 | ||
73fbfcad | 6508 | if (dump_enabled_p ()) |
78c60e3d | 6509 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6510 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a IR |
6511 | |
6512 | /** Transform. **/ | |
6513 | ||
c716e67f XDL |
6514 | ensure_base_align (stmt_info, dr); |
6515 | ||
3bab6342 | 6516 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
aec7ae7d JJ |
6517 | { |
6518 | tree vec_oprnd0 = NULL_TREE, op; | |
6519 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
6520 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
d3c2fee0 | 6521 | tree ptr, mask, var, scale, merge, perm_mask = NULL_TREE, prev_res = NULL_TREE; |
aec7ae7d JJ |
6522 | edge pe = loop_preheader_edge (loop); |
6523 | gimple_seq seq; | |
6524 | basic_block new_bb; | |
6525 | enum { NARROW, NONE, WIDEN } modifier; | |
6526 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
6527 | ||
6528 | if (nunits == gather_off_nunits) | |
6529 | modifier = NONE; | |
6530 | else if (nunits == gather_off_nunits / 2) | |
6531 | { | |
6532 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
6533 | modifier = WIDEN; | |
6534 | ||
6535 | for (i = 0; i < gather_off_nunits; ++i) | |
6536 | sel[i] = i | nunits; | |
6537 | ||
557be5a8 | 6538 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
aec7ae7d JJ |
6539 | } |
6540 | else if (nunits == gather_off_nunits * 2) | |
6541 | { | |
6542 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
6543 | modifier = NARROW; | |
6544 | ||
6545 | for (i = 0; i < nunits; ++i) | |
6546 | sel[i] = i < gather_off_nunits | |
6547 | ? i : i + nunits - gather_off_nunits; | |
6548 | ||
557be5a8 | 6549 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
aec7ae7d JJ |
6550 | ncopies *= 2; |
6551 | } | |
6552 | else | |
6553 | gcc_unreachable (); | |
6554 | ||
6555 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
6556 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6557 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6558 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6559 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6560 | scaletype = TREE_VALUE (arglist); | |
d3c2fee0 | 6561 | gcc_checking_assert (types_compatible_p (srctype, rettype)); |
aec7ae7d JJ |
6562 | |
6563 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
6564 | ||
6565 | ptr = fold_convert (ptrtype, gather_base); | |
6566 | if (!is_gimple_min_invariant (ptr)) | |
6567 | { | |
6568 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6569 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6570 | gcc_assert (!new_bb); | |
6571 | } | |
6572 | ||
6573 | /* Currently we support only unconditional gather loads, | |
6574 | so mask should be all ones. */ | |
d3c2fee0 AI |
6575 | if (TREE_CODE (masktype) == INTEGER_TYPE) |
6576 | mask = build_int_cst (masktype, -1); | |
6577 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
6578 | { | |
6579 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
6580 | mask = build_vector_from_val (masktype, mask); | |
03b9e8e4 | 6581 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
d3c2fee0 | 6582 | } |
aec7ae7d JJ |
6583 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) |
6584 | { | |
6585 | REAL_VALUE_TYPE r; | |
6586 | long tmp[6]; | |
6587 | for (j = 0; j < 6; ++j) | |
6588 | tmp[j] = -1; | |
6589 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
6590 | mask = build_real (TREE_TYPE (masktype), r); | |
d3c2fee0 | 6591 | mask = build_vector_from_val (masktype, mask); |
03b9e8e4 | 6592 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
aec7ae7d JJ |
6593 | } |
6594 | else | |
6595 | gcc_unreachable (); | |
aec7ae7d JJ |
6596 | |
6597 | scale = build_int_cst (scaletype, gather_scale); | |
6598 | ||
d3c2fee0 AI |
6599 | if (TREE_CODE (TREE_TYPE (rettype)) == INTEGER_TYPE) |
6600 | merge = build_int_cst (TREE_TYPE (rettype), 0); | |
6601 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (rettype))) | |
6602 | { | |
6603 | REAL_VALUE_TYPE r; | |
6604 | long tmp[6]; | |
6605 | for (j = 0; j < 6; ++j) | |
6606 | tmp[j] = 0; | |
6607 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (rettype))); | |
6608 | merge = build_real (TREE_TYPE (rettype), r); | |
6609 | } | |
6610 | else | |
6611 | gcc_unreachable (); | |
6612 | merge = build_vector_from_val (rettype, merge); | |
6613 | merge = vect_init_vector (stmt, merge, rettype, NULL); | |
6614 | ||
aec7ae7d JJ |
6615 | prev_stmt_info = NULL; |
6616 | for (j = 0; j < ncopies; ++j) | |
6617 | { | |
6618 | if (modifier == WIDEN && (j & 1)) | |
6619 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
6620 | perm_mask, stmt, gsi); | |
6621 | else if (j == 0) | |
6622 | op = vec_oprnd0 | |
81c40241 | 6623 | = vect_get_vec_def_for_operand (gather_off, stmt); |
aec7ae7d JJ |
6624 | else |
6625 | op = vec_oprnd0 | |
6626 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
6627 | ||
6628 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6629 | { | |
6630 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
6631 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 6632 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
aec7ae7d JJ |
6633 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
6634 | new_stmt | |
0d0e4a03 | 6635 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6636 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6637 | op = var; | |
6638 | } | |
6639 | ||
6640 | new_stmt | |
d3c2fee0 | 6641 | = gimple_build_call (gather_decl, 5, merge, ptr, op, mask, scale); |
aec7ae7d JJ |
6642 | |
6643 | if (!useless_type_conversion_p (vectype, rettype)) | |
6644 | { | |
6645 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
6646 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 6647 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
aec7ae7d JJ |
6648 | gimple_call_set_lhs (new_stmt, op); |
6649 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 6650 | var = make_ssa_name (vec_dest); |
aec7ae7d JJ |
6651 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
6652 | new_stmt | |
0d0e4a03 | 6653 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6654 | } |
6655 | else | |
6656 | { | |
6657 | var = make_ssa_name (vec_dest, new_stmt); | |
6658 | gimple_call_set_lhs (new_stmt, var); | |
6659 | } | |
6660 | ||
6661 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6662 | ||
6663 | if (modifier == NARROW) | |
6664 | { | |
6665 | if ((j & 1) == 0) | |
6666 | { | |
6667 | prev_res = var; | |
6668 | continue; | |
6669 | } | |
6670 | var = permute_vec_elements (prev_res, var, | |
6671 | perm_mask, stmt, gsi); | |
6672 | new_stmt = SSA_NAME_DEF_STMT (var); | |
6673 | } | |
6674 | ||
6675 | if (prev_stmt_info == NULL) | |
6676 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6677 | else | |
6678 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6679 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6680 | } | |
6681 | return true; | |
6682 | } | |
f2e2a985 | 6683 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
7d75abc8 MM |
6684 | { |
6685 | gimple_stmt_iterator incr_gsi; | |
6686 | bool insert_after; | |
355fe088 | 6687 | gimple *incr; |
7d75abc8 | 6688 | tree offvar; |
7d75abc8 MM |
6689 | tree ivstep; |
6690 | tree running_off; | |
9771b263 | 6691 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 6692 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
6693 | tree stride_base, stride_step, alias_off; |
6694 | ||
6695 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 6696 | |
f502d50e | 6697 | if (slp && grouped_load) |
ab313a8c RB |
6698 | first_dr = STMT_VINFO_DATA_REF |
6699 | (vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info))); | |
6700 | else | |
6701 | first_dr = dr; | |
6702 | ||
14ac6aa2 RB |
6703 | stride_base |
6704 | = fold_build_pointer_plus | |
ab313a8c | 6705 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 6706 | size_binop (PLUS_EXPR, |
ab313a8c RB |
6707 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
6708 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
6709 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
6710 | |
6711 | /* For a load with loop-invariant (but other than power-of-2) | |
6712 | stride (i.e. not a grouped access) like so: | |
6713 | ||
6714 | for (i = 0; i < n; i += stride) | |
6715 | ... = array[i]; | |
6716 | ||
6717 | we generate a new induction variable and new accesses to | |
6718 | form a new vector (or vectors, depending on ncopies): | |
6719 | ||
6720 | for (j = 0; ; j += VF*stride) | |
6721 | tmp1 = array[j]; | |
6722 | tmp2 = array[j + stride]; | |
6723 | ... | |
6724 | vectemp = {tmp1, tmp2, ...} | |
6725 | */ | |
6726 | ||
ab313a8c RB |
6727 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
6728 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
6729 | |
6730 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6731 | ||
ab313a8c | 6732 | create_iv (unshare_expr (stride_base), unshare_expr (ivstep), NULL, |
7d75abc8 MM |
6733 | loop, &incr_gsi, insert_after, |
6734 | &offvar, NULL); | |
6735 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 6736 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
7d75abc8 | 6737 | |
ab313a8c RB |
6738 | stride_step = force_gimple_operand (unshare_expr (stride_step), |
6739 | &stmts, true, NULL_TREE); | |
7d75abc8 MM |
6740 | if (stmts) |
6741 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
6742 | ||
6743 | prev_stmt_info = NULL; | |
6744 | running_off = offvar; | |
ab313a8c | 6745 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (first_dr)), 0); |
7b5fc413 RB |
6746 | int nloads = nunits; |
6747 | tree ltype = TREE_TYPE (vectype); | |
b266b968 | 6748 | auto_vec<tree> dr_chain; |
7b5fc413 RB |
6749 | if (slp) |
6750 | { | |
6751 | nloads = nunits / group_size; | |
6752 | if (group_size < nunits) | |
6753 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); | |
6754 | else | |
6755 | ltype = vectype; | |
6756 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); | |
66c16fd9 RB |
6757 | /* For SLP permutation support we need to load the whole group, |
6758 | not only the number of vector stmts the permutation result | |
6759 | fits in. */ | |
b266b968 | 6760 | if (slp_perm) |
66c16fd9 RB |
6761 | { |
6762 | ncopies = (group_size * vf + nunits - 1) / nunits; | |
6763 | dr_chain.create (ncopies); | |
6764 | } | |
6765 | else | |
6766 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
7b5fc413 | 6767 | } |
7d75abc8 MM |
6768 | for (j = 0; j < ncopies; j++) |
6769 | { | |
6770 | tree vec_inv; | |
6771 | ||
7b5fc413 RB |
6772 | if (nloads > 1) |
6773 | { | |
6774 | vec_alloc (v, nloads); | |
6775 | for (i = 0; i < nloads; i++) | |
6776 | { | |
6777 | tree newref, newoff; | |
355fe088 | 6778 | gimple *incr; |
7b5fc413 RB |
6779 | newref = build2 (MEM_REF, ltype, running_off, alias_off); |
6780 | ||
6781 | newref = force_gimple_operand_gsi (gsi, newref, true, | |
6782 | NULL_TREE, true, | |
6783 | GSI_SAME_STMT); | |
6784 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, newref); | |
6785 | newoff = copy_ssa_name (running_off); | |
6786 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6787 | running_off, stride_step); | |
6788 | vect_finish_stmt_generation (stmt, incr, gsi); | |
6789 | ||
6790 | running_off = newoff; | |
6791 | } | |
6792 | ||
6793 | vec_inv = build_constructor (vectype, v); | |
6794 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
6795 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
6796 | } | |
6797 | else | |
7d75abc8 | 6798 | { |
7b5fc413 RB |
6799 | new_stmt = gimple_build_assign (make_ssa_name (ltype), |
6800 | build2 (MEM_REF, ltype, | |
6801 | running_off, alias_off)); | |
6802 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6803 | ||
6804 | tree newoff = copy_ssa_name (running_off); | |
355fe088 | 6805 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
0d0e4a03 | 6806 | running_off, stride_step); |
7d75abc8 MM |
6807 | vect_finish_stmt_generation (stmt, incr, gsi); |
6808 | ||
6809 | running_off = newoff; | |
6810 | } | |
6811 | ||
7b5fc413 | 6812 | if (slp) |
b266b968 | 6813 | { |
b266b968 RB |
6814 | if (slp_perm) |
6815 | dr_chain.quick_push (gimple_assign_lhs (new_stmt)); | |
66c16fd9 RB |
6816 | else |
6817 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
b266b968 | 6818 | } |
7d75abc8 | 6819 | else |
225ce44b RB |
6820 | { |
6821 | if (j == 0) | |
6822 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6823 | else | |
6824 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6825 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6826 | } | |
7d75abc8 | 6827 | } |
b266b968 RB |
6828 | if (slp_perm) |
6829 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
6830 | slp_node_instance, false); | |
7d75abc8 MM |
6831 | return true; |
6832 | } | |
aec7ae7d | 6833 | |
0d0293ac | 6834 | if (grouped_load) |
ebfd146a | 6835 | { |
e14c1050 | 6836 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
4f0a0218 | 6837 | /* For SLP vectorization we directly vectorize a subchain |
52eab378 RB |
6838 | without permutation. */ |
6839 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
4f0a0218 RB |
6840 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
6841 | /* For BB vectorization always use the first stmt to base | |
6842 | the data ref pointer on. */ | |
6843 | if (bb_vinfo) | |
6844 | first_stmt_for_drptr = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 6845 | |
ebfd146a | 6846 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
6847 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
6848 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
6849 | ??? But we can only do so if there is exactly one | |
6850 | as we have no way to get at the rest. Leave the CSE | |
6851 | opportunity alone. | |
6852 | ??? With the group load eventually participating | |
6853 | in multiple different permutations (having multiple | |
6854 | slp nodes which refer to the same group) the CSE | |
6855 | is even wrong code. See PR56270. */ | |
6856 | && !slp) | |
ebfd146a IR |
6857 | { |
6858 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
6859 | return true; | |
6860 | } | |
6861 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 6862 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
9b999e8c | 6863 | group_gap_adj = 0; |
ebfd146a IR |
6864 | |
6865 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
6866 | if (slp) | |
6867 | { | |
0d0293ac | 6868 | grouped_load = false; |
91ff1504 RB |
6869 | /* For SLP permutation support we need to load the whole group, |
6870 | not only the number of vector stmts the permutation result | |
6871 | fits in. */ | |
6872 | if (slp_perm) | |
6873 | vec_num = (group_size * vf + nunits - 1) / nunits; | |
6874 | else | |
6875 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9b999e8c | 6876 | group_gap_adj = vf * group_size - nunits * vec_num; |
a70d6342 | 6877 | } |
ebfd146a | 6878 | else |
9b999e8c | 6879 | vec_num = group_size; |
ebfd146a IR |
6880 | } |
6881 | else | |
6882 | { | |
6883 | first_stmt = stmt; | |
6884 | first_dr = dr; | |
6885 | group_size = vec_num = 1; | |
9b999e8c | 6886 | group_gap_adj = 0; |
ebfd146a IR |
6887 | } |
6888 | ||
720f5239 | 6889 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6890 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
6891 | /* Targets with load-lane instructions must not require explicit |
6892 | realignment. */ | |
6893 | gcc_assert (!load_lanes_p | |
6894 | || alignment_support_scheme == dr_aligned | |
6895 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
6896 | |
6897 | /* In case the vectorization factor (VF) is bigger than the number | |
6898 | of elements that we can fit in a vectype (nunits), we have to generate | |
6899 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 6900 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 6901 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 6902 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 6903 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
6904 | stmts that use the defs of the current stmt. The example below |
6905 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
6906 | need to create 4 vectorized stmts): | |
ebfd146a IR |
6907 | |
6908 | before vectorization: | |
6909 | RELATED_STMT VEC_STMT | |
6910 | S1: x = memref - - | |
6911 | S2: z = x + 1 - - | |
6912 | ||
6913 | step 1: vectorize stmt S1: | |
6914 | We first create the vector stmt VS1_0, and, as usual, record a | |
6915 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
6916 | Next, we create the vector stmt VS1_1, and record a pointer to | |
6917 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 6918 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
6919 | stmts and pointers: |
6920 | RELATED_STMT VEC_STMT | |
6921 | VS1_0: vx0 = memref0 VS1_1 - | |
6922 | VS1_1: vx1 = memref1 VS1_2 - | |
6923 | VS1_2: vx2 = memref2 VS1_3 - | |
6924 | VS1_3: vx3 = memref3 - - | |
6925 | S1: x = load - VS1_0 | |
6926 | S2: z = x + 1 - - | |
6927 | ||
b8698a0f L |
6928 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
6929 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
6930 | stmt S2. */ |
6931 | ||
0d0293ac | 6932 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6933 | |
6934 | S1: x2 = &base + 2 | |
6935 | S2: x0 = &base | |
6936 | S3: x1 = &base + 1 | |
6937 | S4: x3 = &base + 3 | |
6938 | ||
b8698a0f | 6939 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
6940 | starting from the access of the first stmt of the chain: |
6941 | ||
6942 | VS1: vx0 = &base | |
6943 | VS2: vx1 = &base + vec_size*1 | |
6944 | VS3: vx3 = &base + vec_size*2 | |
6945 | VS4: vx4 = &base + vec_size*3 | |
6946 | ||
6947 | Then permutation statements are generated: | |
6948 | ||
e2c83630 RH |
6949 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
6950 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
6951 | ... |
6952 | ||
6953 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
6954 | (the order of the data-refs in the output of vect_permute_load_chain | |
6955 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6956 | the documentation of vect_permute_load_chain()). | |
6957 | The generation of permutation stmts and recording them in | |
0d0293ac | 6958 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 6959 | |
b8698a0f | 6960 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
6961 | permutation stmts above are created for every copy. The result vector |
6962 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
6963 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
6964 | |
6965 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
6966 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
6967 | we generate the following code: | |
6968 | p = initial_addr; | |
6969 | indx = 0; | |
6970 | loop { | |
6971 | p = p + indx * vectype_size; | |
6972 | vec_dest = *(p); | |
6973 | indx = indx + 1; | |
6974 | } | |
6975 | ||
6976 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 6977 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
6978 | then generate the following code, in which the data in each iteration is |
6979 | obtained by two vector loads, one from the previous iteration, and one | |
6980 | from the current iteration: | |
6981 | p1 = initial_addr; | |
6982 | msq_init = *(floor(p1)) | |
6983 | p2 = initial_addr + VS - 1; | |
6984 | realignment_token = call target_builtin; | |
6985 | indx = 0; | |
6986 | loop { | |
6987 | p2 = p2 + indx * vectype_size | |
6988 | lsq = *(floor(p2)) | |
6989 | vec_dest = realign_load (msq, lsq, realignment_token) | |
6990 | indx = indx + 1; | |
6991 | msq = lsq; | |
6992 | } */ | |
6993 | ||
6994 | /* If the misalignment remains the same throughout the execution of the | |
6995 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 6996 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
6997 | This can only occur when vectorizing memory accesses in the inner-loop |
6998 | nested within an outer-loop that is being vectorized. */ | |
6999 | ||
d1e4b493 | 7000 | if (nested_in_vect_loop |
211bea38 | 7001 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
ebfd146a IR |
7002 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) |
7003 | { | |
7004 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
7005 | compute_in_loop = true; | |
7006 | } | |
7007 | ||
7008 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
7009 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 7010 | && !compute_in_loop) |
ebfd146a IR |
7011 | { |
7012 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
7013 | alignment_support_scheme, NULL_TREE, | |
7014 | &at_loop); | |
7015 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7016 | { | |
538dd0b7 | 7017 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
7018 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
7019 | size_one_node); | |
ebfd146a IR |
7020 | } |
7021 | } | |
7022 | else | |
7023 | at_loop = loop; | |
7024 | ||
a1e53f3f L |
7025 | if (negative) |
7026 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
7027 | ||
272c6793 RS |
7028 | if (load_lanes_p) |
7029 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
7030 | else | |
7031 | aggr_type = vectype; | |
7032 | ||
ebfd146a IR |
7033 | prev_stmt_info = NULL; |
7034 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 7035 | { |
272c6793 | 7036 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 7037 | if (j == 0) |
74bf76ed JJ |
7038 | { |
7039 | bool simd_lane_access_p | |
7040 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
7041 | if (simd_lane_access_p | |
7042 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
7043 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
7044 | && integer_zerop (DR_OFFSET (first_dr)) | |
7045 | && integer_zerop (DR_INIT (first_dr)) | |
7046 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
7047 | get_alias_set (DR_REF (first_dr))) | |
7048 | && (alignment_support_scheme == dr_aligned | |
7049 | || alignment_support_scheme == dr_unaligned_supported)) | |
7050 | { | |
7051 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
7052 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
7053 | (DR_REF (first_dr)), 0); | |
8928eff3 | 7054 | inv_p = false; |
74bf76ed | 7055 | } |
4f0a0218 RB |
7056 | else if (first_stmt_for_drptr |
7057 | && first_stmt != first_stmt_for_drptr) | |
7058 | { | |
7059 | dataref_ptr | |
7060 | = vect_create_data_ref_ptr (first_stmt_for_drptr, aggr_type, | |
7061 | at_loop, offset, &dummy, gsi, | |
7062 | &ptr_incr, simd_lane_access_p, | |
7063 | &inv_p, byte_offset); | |
7064 | /* Adjust the pointer by the difference to first_stmt. */ | |
7065 | data_reference_p ptrdr | |
7066 | = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt_for_drptr)); | |
7067 | tree diff = fold_convert (sizetype, | |
7068 | size_binop (MINUS_EXPR, | |
7069 | DR_INIT (first_dr), | |
7070 | DR_INIT (ptrdr))); | |
7071 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7072 | stmt, diff); | |
7073 | } | |
74bf76ed JJ |
7074 | else |
7075 | dataref_ptr | |
7076 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
7077 | offset, &dummy, gsi, &ptr_incr, | |
356bbc4c JJ |
7078 | simd_lane_access_p, &inv_p, |
7079 | byte_offset); | |
74bf76ed JJ |
7080 | } |
7081 | else if (dataref_offset) | |
7082 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
7083 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 7084 | else |
272c6793 RS |
7085 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
7086 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 7087 | |
0d0293ac | 7088 | if (grouped_load || slp_perm) |
9771b263 | 7089 | dr_chain.create (vec_num); |
5ce1ee7f | 7090 | |
272c6793 | 7091 | if (load_lanes_p) |
ebfd146a | 7092 | { |
272c6793 RS |
7093 | tree vec_array; |
7094 | ||
7095 | vec_array = create_vector_array (vectype, vec_num); | |
7096 | ||
7097 | /* Emit: | |
7098 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
7099 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
7100 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
7101 | gimple_call_set_lhs (new_stmt, vec_array); | |
7102 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 7103 | |
272c6793 RS |
7104 | /* Extract each vector into an SSA_NAME. */ |
7105 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 7106 | { |
272c6793 RS |
7107 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
7108 | vec_array, i); | |
9771b263 | 7109 | dr_chain.quick_push (new_temp); |
272c6793 RS |
7110 | } |
7111 | ||
7112 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 7113 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
7114 | } |
7115 | else | |
7116 | { | |
7117 | for (i = 0; i < vec_num; i++) | |
7118 | { | |
7119 | if (i > 0) | |
7120 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7121 | stmt, NULL_TREE); | |
7122 | ||
7123 | /* 2. Create the vector-load in the loop. */ | |
7124 | switch (alignment_support_scheme) | |
7125 | { | |
7126 | case dr_aligned: | |
7127 | case dr_unaligned_supported: | |
be1ac4ec | 7128 | { |
644ffefd MJ |
7129 | unsigned int align, misalign; |
7130 | ||
272c6793 | 7131 | data_ref |
aed93b23 RB |
7132 | = fold_build2 (MEM_REF, vectype, dataref_ptr, |
7133 | dataref_offset | |
7134 | ? dataref_offset | |
7135 | : build_int_cst (reference_alias_ptr_type | |
7136 | (DR_REF (first_dr)), 0)); | |
644ffefd | 7137 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 RS |
7138 | if (alignment_support_scheme == dr_aligned) |
7139 | { | |
7140 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 7141 | misalign = 0; |
272c6793 RS |
7142 | } |
7143 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
7144 | { | |
52639a61 RB |
7145 | if (DR_VECT_AUX (first_dr)->base_element_aligned) |
7146 | align = TYPE_ALIGN_UNIT (elem_type); | |
7147 | else | |
7148 | align = (get_object_alignment (DR_REF (first_dr)) | |
7149 | / BITS_PER_UNIT); | |
7150 | misalign = 0; | |
272c6793 RS |
7151 | TREE_TYPE (data_ref) |
7152 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 7153 | align * BITS_PER_UNIT); |
272c6793 RS |
7154 | } |
7155 | else | |
7156 | { | |
7157 | TREE_TYPE (data_ref) | |
7158 | = build_aligned_type (TREE_TYPE (data_ref), | |
7159 | TYPE_ALIGN (elem_type)); | |
644ffefd | 7160 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 7161 | } |
aed93b23 RB |
7162 | if (dataref_offset == NULL_TREE |
7163 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
7164 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
7165 | align, misalign); | |
272c6793 | 7166 | break; |
be1ac4ec | 7167 | } |
272c6793 | 7168 | case dr_explicit_realign: |
267d3070 | 7169 | { |
272c6793 | 7170 | tree ptr, bump; |
272c6793 | 7171 | |
d88981fc | 7172 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
7173 | |
7174 | if (compute_in_loop) | |
7175 | msq = vect_setup_realignment (first_stmt, gsi, | |
7176 | &realignment_token, | |
7177 | dr_explicit_realign, | |
7178 | dataref_ptr, NULL); | |
7179 | ||
aed93b23 RB |
7180 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7181 | ptr = copy_ssa_name (dataref_ptr); | |
7182 | else | |
7183 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
7184 | new_stmt = gimple_build_assign |
7185 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7186 | build_int_cst |
7187 | (TREE_TYPE (dataref_ptr), | |
7188 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
7189 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7190 | data_ref | |
7191 | = build2 (MEM_REF, vectype, ptr, | |
7192 | build_int_cst (reference_alias_ptr_type | |
7193 | (DR_REF (first_dr)), 0)); | |
7194 | vec_dest = vect_create_destination_var (scalar_dest, | |
7195 | vectype); | |
7196 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
7197 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
7198 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7199 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
7200 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
7201 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7202 | msq = new_temp; | |
7203 | ||
d88981fc | 7204 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 7205 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 7206 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 7207 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
7208 | new_stmt = gimple_build_assign |
7209 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 RS |
7210 | build_int_cst |
7211 | (TREE_TYPE (ptr), | |
7212 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
aed93b23 | 7213 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
7214 | gimple_assign_set_lhs (new_stmt, ptr); |
7215 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7216 | data_ref | |
7217 | = build2 (MEM_REF, vectype, ptr, | |
7218 | build_int_cst (reference_alias_ptr_type | |
7219 | (DR_REF (first_dr)), 0)); | |
7220 | break; | |
267d3070 | 7221 | } |
272c6793 | 7222 | case dr_explicit_realign_optimized: |
aed93b23 RB |
7223 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7224 | new_temp = copy_ssa_name (dataref_ptr); | |
7225 | else | |
7226 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
7227 | new_stmt = gimple_build_assign |
7228 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7229 | build_int_cst |
7230 | (TREE_TYPE (dataref_ptr), | |
7231 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
7232 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7233 | data_ref | |
7234 | = build2 (MEM_REF, vectype, new_temp, | |
7235 | build_int_cst (reference_alias_ptr_type | |
7236 | (DR_REF (first_dr)), 0)); | |
7237 | break; | |
7238 | default: | |
7239 | gcc_unreachable (); | |
7240 | } | |
ebfd146a | 7241 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 7242 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
7243 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7244 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7245 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7246 | ||
272c6793 RS |
7247 | /* 3. Handle explicit realignment if necessary/supported. |
7248 | Create in loop: | |
7249 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
7250 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
7251 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 7252 | { |
272c6793 RS |
7253 | lsq = gimple_assign_lhs (new_stmt); |
7254 | if (!realignment_token) | |
7255 | realignment_token = dataref_ptr; | |
7256 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
7257 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
7258 | msq, lsq, realignment_token); | |
272c6793 RS |
7259 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7260 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7261 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7262 | ||
7263 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7264 | { | |
7265 | gcc_assert (phi); | |
7266 | if (i == vec_num - 1 && j == ncopies - 1) | |
7267 | add_phi_arg (phi, lsq, | |
7268 | loop_latch_edge (containing_loop), | |
9e227d60 | 7269 | UNKNOWN_LOCATION); |
272c6793 RS |
7270 | msq = lsq; |
7271 | } | |
ebfd146a | 7272 | } |
ebfd146a | 7273 | |
59fd17e3 RB |
7274 | /* 4. Handle invariant-load. */ |
7275 | if (inv_p && !bb_vinfo) | |
7276 | { | |
59fd17e3 | 7277 | gcc_assert (!grouped_load); |
d1417442 JJ |
7278 | /* If we have versioned for aliasing or the loop doesn't |
7279 | have any data dependencies that would preclude this, | |
7280 | then we are sure this is a loop invariant load and | |
7281 | thus we can insert it on the preheader edge. */ | |
7282 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
7283 | && !nested_in_vect_loop | |
6b916b36 | 7284 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
7285 | { |
7286 | if (dump_enabled_p ()) | |
7287 | { | |
7288 | dump_printf_loc (MSG_NOTE, vect_location, | |
7289 | "hoisting out of the vectorized " | |
7290 | "loop: "); | |
7291 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 7292 | } |
b731b390 | 7293 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
7294 | gsi_insert_on_edge_immediate |
7295 | (loop_preheader_edge (loop), | |
7296 | gimple_build_assign (tem, | |
7297 | unshare_expr | |
7298 | (gimple_assign_rhs1 (stmt)))); | |
7299 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
34cd48e5 RB |
7300 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7301 | set_vinfo_for_stmt (new_stmt, | |
7302 | new_stmt_vec_info (new_stmt, vinfo)); | |
a0e35eb0 RB |
7303 | } |
7304 | else | |
7305 | { | |
7306 | gimple_stmt_iterator gsi2 = *gsi; | |
7307 | gsi_next (&gsi2); | |
7308 | new_temp = vect_init_vector (stmt, scalar_dest, | |
7309 | vectype, &gsi2); | |
34cd48e5 | 7310 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
a0e35eb0 | 7311 | } |
59fd17e3 RB |
7312 | } |
7313 | ||
272c6793 RS |
7314 | if (negative) |
7315 | { | |
aec7ae7d JJ |
7316 | tree perm_mask = perm_mask_for_reverse (vectype); |
7317 | new_temp = permute_vec_elements (new_temp, new_temp, | |
7318 | perm_mask, stmt, gsi); | |
ebfd146a IR |
7319 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7320 | } | |
267d3070 | 7321 | |
272c6793 | 7322 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
7323 | vect_transform_grouped_load (). */ |
7324 | if (grouped_load || slp_perm) | |
9771b263 | 7325 | dr_chain.quick_push (new_temp); |
267d3070 | 7326 | |
272c6793 RS |
7327 | /* Store vector loads in the corresponding SLP_NODE. */ |
7328 | if (slp && !slp_perm) | |
9771b263 | 7329 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
272c6793 | 7330 | } |
9b999e8c RB |
7331 | /* Bump the vector pointer to account for a gap or for excess |
7332 | elements loaded for a permuted SLP load. */ | |
7333 | if (group_gap_adj != 0) | |
a64b9c26 | 7334 | { |
9b999e8c RB |
7335 | bool ovf; |
7336 | tree bump | |
7337 | = wide_int_to_tree (sizetype, | |
7338 | wi::smul (TYPE_SIZE_UNIT (elem_type), | |
7339 | group_gap_adj, &ovf)); | |
a64b9c26 RB |
7340 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
7341 | stmt, bump); | |
7342 | } | |
ebfd146a IR |
7343 | } |
7344 | ||
7345 | if (slp && !slp_perm) | |
7346 | continue; | |
7347 | ||
7348 | if (slp_perm) | |
7349 | { | |
01d8bf07 | 7350 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
ebfd146a IR |
7351 | slp_node_instance, false)) |
7352 | { | |
9771b263 | 7353 | dr_chain.release (); |
ebfd146a IR |
7354 | return false; |
7355 | } | |
7356 | } | |
7357 | else | |
7358 | { | |
0d0293ac | 7359 | if (grouped_load) |
ebfd146a | 7360 | { |
272c6793 | 7361 | if (!load_lanes_p) |
0d0293ac | 7362 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 7363 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
7364 | } |
7365 | else | |
7366 | { | |
7367 | if (j == 0) | |
7368 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7369 | else | |
7370 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7371 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7372 | } | |
7373 | } | |
9771b263 | 7374 | dr_chain.release (); |
ebfd146a IR |
7375 | } |
7376 | ||
ebfd146a IR |
7377 | return true; |
7378 | } | |
7379 | ||
7380 | /* Function vect_is_simple_cond. | |
b8698a0f | 7381 | |
ebfd146a IR |
7382 | Input: |
7383 | LOOP - the loop that is being vectorized. | |
7384 | COND - Condition that is checked for simple use. | |
7385 | ||
e9e1d143 RG |
7386 | Output: |
7387 | *COMP_VECTYPE - the vector type for the comparison. | |
7388 | ||
ebfd146a IR |
7389 | Returns whether a COND can be vectorized. Checks whether |
7390 | condition operands are supportable using vec_is_simple_use. */ | |
7391 | ||
87aab9b2 | 7392 | static bool |
81c40241 | 7393 | vect_is_simple_cond (tree cond, vec_info *vinfo, tree *comp_vectype) |
ebfd146a IR |
7394 | { |
7395 | tree lhs, rhs; | |
ebfd146a | 7396 | enum vect_def_type dt; |
e9e1d143 | 7397 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 7398 | |
a414c77f IE |
7399 | /* Mask case. */ |
7400 | if (TREE_CODE (cond) == SSA_NAME | |
7401 | && TREE_CODE (TREE_TYPE (cond)) == BOOLEAN_TYPE) | |
7402 | { | |
7403 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (cond); | |
7404 | if (!vect_is_simple_use (cond, vinfo, &lhs_def_stmt, | |
7405 | &dt, comp_vectype) | |
7406 | || !*comp_vectype | |
7407 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
7408 | return false; | |
7409 | return true; | |
7410 | } | |
7411 | ||
ebfd146a IR |
7412 | if (!COMPARISON_CLASS_P (cond)) |
7413 | return false; | |
7414 | ||
7415 | lhs = TREE_OPERAND (cond, 0); | |
7416 | rhs = TREE_OPERAND (cond, 1); | |
7417 | ||
7418 | if (TREE_CODE (lhs) == SSA_NAME) | |
7419 | { | |
355fe088 | 7420 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); |
81c40241 | 7421 | if (!vect_is_simple_use (lhs, vinfo, &lhs_def_stmt, &dt, &vectype1)) |
ebfd146a IR |
7422 | return false; |
7423 | } | |
7424 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
7425 | && TREE_CODE (lhs) != FIXED_CST) | |
7426 | return false; | |
7427 | ||
7428 | if (TREE_CODE (rhs) == SSA_NAME) | |
7429 | { | |
355fe088 | 7430 | gimple *rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); |
81c40241 | 7431 | if (!vect_is_simple_use (rhs, vinfo, &rhs_def_stmt, &dt, &vectype2)) |
ebfd146a IR |
7432 | return false; |
7433 | } | |
f7e531cf | 7434 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST |
ebfd146a IR |
7435 | && TREE_CODE (rhs) != FIXED_CST) |
7436 | return false; | |
7437 | ||
28b33016 IE |
7438 | if (vectype1 && vectype2 |
7439 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
7440 | return false; | |
7441 | ||
e9e1d143 | 7442 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
7443 | return true; |
7444 | } | |
7445 | ||
7446 | /* vectorizable_condition. | |
7447 | ||
b8698a0f L |
7448 | Check if STMT is conditional modify expression that can be vectorized. |
7449 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7450 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
7451 | at GSI. |
7452 | ||
7453 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
7454 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 7455 | else clause if it is 2). |
ebfd146a IR |
7456 | |
7457 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7458 | ||
4bbe8262 | 7459 | bool |
355fe088 TS |
7460 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
7461 | gimple **vec_stmt, tree reduc_def, int reduc_index, | |
f7e531cf | 7462 | slp_tree slp_node) |
ebfd146a IR |
7463 | { |
7464 | tree scalar_dest = NULL_TREE; | |
7465 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
7466 | tree cond_expr, then_clause, else_clause; |
7467 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
df11cc78 | 7468 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
7469 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
7470 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
5958f9e2 | 7471 | tree vec_compare; |
ebfd146a IR |
7472 | tree new_temp; |
7473 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a855b1b1 | 7474 | enum vect_def_type dt, dts[4]; |
f7e531cf | 7475 | int ncopies; |
ebfd146a | 7476 | enum tree_code code; |
a855b1b1 | 7477 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
7478 | int i, j; |
7479 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
7480 | vec<tree> vec_oprnds0 = vNULL; |
7481 | vec<tree> vec_oprnds1 = vNULL; | |
7482 | vec<tree> vec_oprnds2 = vNULL; | |
7483 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 7484 | tree vec_cmp_type; |
a414c77f | 7485 | bool masked = false; |
b8698a0f | 7486 | |
f7e531cf IR |
7487 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
7488 | return false; | |
7489 | ||
af29617a AH |
7490 | if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == TREE_CODE_REDUCTION) |
7491 | { | |
7492 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
7493 | return false; | |
ebfd146a | 7494 | |
af29617a AH |
7495 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7496 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7497 | && reduc_def)) | |
7498 | return false; | |
ebfd146a | 7499 | |
af29617a AH |
7500 | /* FORNOW: not yet supported. */ |
7501 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7502 | { | |
7503 | if (dump_enabled_p ()) | |
7504 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7505 | "value used after loop.\n"); | |
7506 | return false; | |
7507 | } | |
ebfd146a IR |
7508 | } |
7509 | ||
7510 | /* Is vectorizable conditional operation? */ | |
7511 | if (!is_gimple_assign (stmt)) | |
7512 | return false; | |
7513 | ||
7514 | code = gimple_assign_rhs_code (stmt); | |
7515 | ||
7516 | if (code != COND_EXPR) | |
7517 | return false; | |
7518 | ||
465c8c19 JJ |
7519 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
7520 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2947d3b2 | 7521 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
465c8c19 | 7522 | |
fce57248 | 7523 | if (slp_node) |
465c8c19 JJ |
7524 | ncopies = 1; |
7525 | else | |
7526 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
7527 | ||
7528 | gcc_assert (ncopies >= 1); | |
7529 | if (reduc_index && ncopies > 1) | |
7530 | return false; /* FORNOW */ | |
7531 | ||
4e71066d RG |
7532 | cond_expr = gimple_assign_rhs1 (stmt); |
7533 | then_clause = gimple_assign_rhs2 (stmt); | |
7534 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 7535 | |
81c40241 | 7536 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, &comp_vectype) |
e9e1d143 | 7537 | || !comp_vectype) |
ebfd146a IR |
7538 | return false; |
7539 | ||
81c40241 | 7540 | gimple *def_stmt; |
2947d3b2 IE |
7541 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &def_stmt, &dt, |
7542 | &vectype1)) | |
7543 | return false; | |
7544 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &def_stmt, &dt, | |
7545 | &vectype2)) | |
ebfd146a | 7546 | return false; |
2947d3b2 IE |
7547 | |
7548 | if (vectype1 && !useless_type_conversion_p (vectype, vectype1)) | |
7549 | return false; | |
7550 | ||
7551 | if (vectype2 && !useless_type_conversion_p (vectype, vectype2)) | |
ebfd146a IR |
7552 | return false; |
7553 | ||
28b33016 IE |
7554 | masked = !COMPARISON_CLASS_P (cond_expr); |
7555 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
7556 | ||
74946978 MP |
7557 | if (vec_cmp_type == NULL_TREE) |
7558 | return false; | |
784fb9b3 | 7559 | |
b8698a0f | 7560 | if (!vec_stmt) |
ebfd146a IR |
7561 | { |
7562 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
e9e1d143 | 7563 | return expand_vec_cond_expr_p (vectype, comp_vectype); |
ebfd146a IR |
7564 | } |
7565 | ||
f7e531cf IR |
7566 | /* Transform. */ |
7567 | ||
7568 | if (!slp_node) | |
7569 | { | |
9771b263 DN |
7570 | vec_oprnds0.create (1); |
7571 | vec_oprnds1.create (1); | |
7572 | vec_oprnds2.create (1); | |
7573 | vec_oprnds3.create (1); | |
f7e531cf | 7574 | } |
ebfd146a IR |
7575 | |
7576 | /* Handle def. */ | |
7577 | scalar_dest = gimple_assign_lhs (stmt); | |
7578 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
7579 | ||
7580 | /* Handle cond expr. */ | |
a855b1b1 MM |
7581 | for (j = 0; j < ncopies; j++) |
7582 | { | |
538dd0b7 | 7583 | gassign *new_stmt = NULL; |
a855b1b1 MM |
7584 | if (j == 0) |
7585 | { | |
f7e531cf IR |
7586 | if (slp_node) |
7587 | { | |
00f96dc9 TS |
7588 | auto_vec<tree, 4> ops; |
7589 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 7590 | |
a414c77f IE |
7591 | if (masked) |
7592 | ops.safe_push (cond_expr); | |
7593 | else | |
7594 | { | |
7595 | ops.safe_push (TREE_OPERAND (cond_expr, 0)); | |
7596 | ops.safe_push (TREE_OPERAND (cond_expr, 1)); | |
7597 | } | |
9771b263 DN |
7598 | ops.safe_push (then_clause); |
7599 | ops.safe_push (else_clause); | |
f7e531cf | 7600 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); |
37b5ec8f JJ |
7601 | vec_oprnds3 = vec_defs.pop (); |
7602 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
7603 | if (!masked) |
7604 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 7605 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf | 7606 | |
9771b263 DN |
7607 | ops.release (); |
7608 | vec_defs.release (); | |
f7e531cf IR |
7609 | } |
7610 | else | |
7611 | { | |
355fe088 | 7612 | gimple *gtemp; |
a414c77f IE |
7613 | if (masked) |
7614 | { | |
7615 | vec_cond_lhs | |
7616 | = vect_get_vec_def_for_operand (cond_expr, stmt, | |
7617 | comp_vectype); | |
7618 | vect_is_simple_use (cond_expr, stmt_info->vinfo, | |
7619 | >emp, &dts[0]); | |
7620 | } | |
7621 | else | |
7622 | { | |
7623 | vec_cond_lhs = | |
7624 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), | |
7625 | stmt, comp_vectype); | |
7626 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), | |
7627 | loop_vinfo, >emp, &dts[0]); | |
7628 | ||
7629 | vec_cond_rhs = | |
7630 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
7631 | stmt, comp_vectype); | |
7632 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), | |
7633 | loop_vinfo, >emp, &dts[1]); | |
7634 | } | |
f7e531cf IR |
7635 | if (reduc_index == 1) |
7636 | vec_then_clause = reduc_def; | |
7637 | else | |
7638 | { | |
7639 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 RB |
7640 | stmt); |
7641 | vect_is_simple_use (then_clause, loop_vinfo, | |
7642 | >emp, &dts[2]); | |
f7e531cf IR |
7643 | } |
7644 | if (reduc_index == 2) | |
7645 | vec_else_clause = reduc_def; | |
7646 | else | |
7647 | { | |
7648 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 RB |
7649 | stmt); |
7650 | vect_is_simple_use (else_clause, loop_vinfo, >emp, &dts[3]); | |
f7e531cf | 7651 | } |
a855b1b1 MM |
7652 | } |
7653 | } | |
7654 | else | |
7655 | { | |
a414c77f IE |
7656 | vec_cond_lhs |
7657 | = vect_get_vec_def_for_stmt_copy (dts[0], | |
7658 | vec_oprnds0.pop ()); | |
7659 | if (!masked) | |
7660 | vec_cond_rhs | |
7661 | = vect_get_vec_def_for_stmt_copy (dts[1], | |
7662 | vec_oprnds1.pop ()); | |
7663 | ||
a855b1b1 | 7664 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 7665 | vec_oprnds2.pop ()); |
a855b1b1 | 7666 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 7667 | vec_oprnds3.pop ()); |
f7e531cf IR |
7668 | } |
7669 | ||
7670 | if (!slp_node) | |
7671 | { | |
9771b263 | 7672 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
7673 | if (!masked) |
7674 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
7675 | vec_oprnds2.quick_push (vec_then_clause); |
7676 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
7677 | } |
7678 | ||
9dc3f7de | 7679 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 7680 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 7681 | { |
9771b263 DN |
7682 | vec_then_clause = vec_oprnds2[i]; |
7683 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 7684 | |
a414c77f IE |
7685 | if (masked) |
7686 | vec_compare = vec_cond_lhs; | |
7687 | else | |
7688 | { | |
7689 | vec_cond_rhs = vec_oprnds1[i]; | |
7690 | vec_compare = build2 (TREE_CODE (cond_expr), vec_cmp_type, | |
7691 | vec_cond_lhs, vec_cond_rhs); | |
7692 | } | |
5958f9e2 JJ |
7693 | new_temp = make_ssa_name (vec_dest); |
7694 | new_stmt = gimple_build_assign (new_temp, VEC_COND_EXPR, | |
7695 | vec_compare, vec_then_clause, | |
7696 | vec_else_clause); | |
f7e531cf IR |
7697 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7698 | if (slp_node) | |
9771b263 | 7699 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
7700 | } |
7701 | ||
7702 | if (slp_node) | |
7703 | continue; | |
7704 | ||
7705 | if (j == 0) | |
7706 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7707 | else | |
7708 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7709 | ||
7710 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 7711 | } |
b8698a0f | 7712 | |
9771b263 DN |
7713 | vec_oprnds0.release (); |
7714 | vec_oprnds1.release (); | |
7715 | vec_oprnds2.release (); | |
7716 | vec_oprnds3.release (); | |
f7e531cf | 7717 | |
ebfd146a IR |
7718 | return true; |
7719 | } | |
7720 | ||
42fd8198 IE |
7721 | /* vectorizable_comparison. |
7722 | ||
7723 | Check if STMT is comparison expression that can be vectorized. | |
7724 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7725 | comparison, put it in VEC_STMT, and insert it at GSI. | |
7726 | ||
7727 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7728 | ||
fce57248 | 7729 | static bool |
42fd8198 IE |
7730 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, |
7731 | gimple **vec_stmt, tree reduc_def, | |
7732 | slp_tree slp_node) | |
7733 | { | |
7734 | tree lhs, rhs1, rhs2; | |
7735 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
7736 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
7737 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
7738 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
7739 | tree new_temp; | |
7740 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
7741 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
7742 | unsigned nunits; | |
7743 | int ncopies; | |
7744 | enum tree_code code; | |
7745 | stmt_vec_info prev_stmt_info = NULL; | |
7746 | int i, j; | |
7747 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
7748 | vec<tree> vec_oprnds0 = vNULL; | |
7749 | vec<tree> vec_oprnds1 = vNULL; | |
7750 | gimple *def_stmt; | |
7751 | tree mask_type; | |
7752 | tree mask; | |
7753 | ||
c245362b IE |
7754 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
7755 | return false; | |
7756 | ||
30480bcd | 7757 | if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)) |
42fd8198 IE |
7758 | return false; |
7759 | ||
7760 | mask_type = vectype; | |
7761 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
7762 | ||
fce57248 | 7763 | if (slp_node) |
42fd8198 IE |
7764 | ncopies = 1; |
7765 | else | |
7766 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
7767 | ||
7768 | gcc_assert (ncopies >= 1); | |
42fd8198 IE |
7769 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7770 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7771 | && reduc_def)) | |
7772 | return false; | |
7773 | ||
7774 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7775 | { | |
7776 | if (dump_enabled_p ()) | |
7777 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7778 | "value used after loop.\n"); | |
7779 | return false; | |
7780 | } | |
7781 | ||
7782 | if (!is_gimple_assign (stmt)) | |
7783 | return false; | |
7784 | ||
7785 | code = gimple_assign_rhs_code (stmt); | |
7786 | ||
7787 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
7788 | return false; | |
7789 | ||
7790 | rhs1 = gimple_assign_rhs1 (stmt); | |
7791 | rhs2 = gimple_assign_rhs2 (stmt); | |
7792 | ||
7793 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &def_stmt, | |
7794 | &dts[0], &vectype1)) | |
7795 | return false; | |
7796 | ||
7797 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &def_stmt, | |
7798 | &dts[1], &vectype2)) | |
7799 | return false; | |
7800 | ||
7801 | if (vectype1 && vectype2 | |
7802 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
7803 | return false; | |
7804 | ||
7805 | vectype = vectype1 ? vectype1 : vectype2; | |
7806 | ||
7807 | /* Invariant comparison. */ | |
7808 | if (!vectype) | |
7809 | { | |
69a9a66f RB |
7810 | vectype = get_vectype_for_scalar_type (TREE_TYPE (rhs1)); |
7811 | if (TYPE_VECTOR_SUBPARTS (vectype) != nunits) | |
42fd8198 IE |
7812 | return false; |
7813 | } | |
7814 | else if (nunits != TYPE_VECTOR_SUBPARTS (vectype)) | |
7815 | return false; | |
7816 | ||
7817 | if (!vec_stmt) | |
7818 | { | |
7819 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
7820 | vect_model_simple_cost (stmt_info, ncopies, dts, NULL, NULL); | |
7821 | return expand_vec_cmp_expr_p (vectype, mask_type); | |
7822 | } | |
7823 | ||
7824 | /* Transform. */ | |
7825 | if (!slp_node) | |
7826 | { | |
7827 | vec_oprnds0.create (1); | |
7828 | vec_oprnds1.create (1); | |
7829 | } | |
7830 | ||
7831 | /* Handle def. */ | |
7832 | lhs = gimple_assign_lhs (stmt); | |
7833 | mask = vect_create_destination_var (lhs, mask_type); | |
7834 | ||
7835 | /* Handle cmp expr. */ | |
7836 | for (j = 0; j < ncopies; j++) | |
7837 | { | |
7838 | gassign *new_stmt = NULL; | |
7839 | if (j == 0) | |
7840 | { | |
7841 | if (slp_node) | |
7842 | { | |
7843 | auto_vec<tree, 2> ops; | |
7844 | auto_vec<vec<tree>, 2> vec_defs; | |
7845 | ||
7846 | ops.safe_push (rhs1); | |
7847 | ops.safe_push (rhs2); | |
7848 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); | |
7849 | vec_oprnds1 = vec_defs.pop (); | |
7850 | vec_oprnds0 = vec_defs.pop (); | |
7851 | } | |
7852 | else | |
7853 | { | |
e4af0bc4 IE |
7854 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, vectype); |
7855 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, vectype); | |
42fd8198 IE |
7856 | } |
7857 | } | |
7858 | else | |
7859 | { | |
7860 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
7861 | vec_oprnds0.pop ()); | |
7862 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
7863 | vec_oprnds1.pop ()); | |
7864 | } | |
7865 | ||
7866 | if (!slp_node) | |
7867 | { | |
7868 | vec_oprnds0.quick_push (vec_rhs1); | |
7869 | vec_oprnds1.quick_push (vec_rhs2); | |
7870 | } | |
7871 | ||
7872 | /* Arguments are ready. Create the new vector stmt. */ | |
7873 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
7874 | { | |
7875 | vec_rhs2 = vec_oprnds1[i]; | |
7876 | ||
7877 | new_temp = make_ssa_name (mask); | |
7878 | new_stmt = gimple_build_assign (new_temp, code, vec_rhs1, vec_rhs2); | |
7879 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7880 | if (slp_node) | |
7881 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
7882 | } | |
7883 | ||
7884 | if (slp_node) | |
7885 | continue; | |
7886 | ||
7887 | if (j == 0) | |
7888 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7889 | else | |
7890 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7891 | ||
7892 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7893 | } | |
7894 | ||
7895 | vec_oprnds0.release (); | |
7896 | vec_oprnds1.release (); | |
7897 | ||
7898 | return true; | |
7899 | } | |
ebfd146a | 7900 | |
8644a673 | 7901 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
7902 | |
7903 | bool | |
355fe088 | 7904 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 7905 | { |
8644a673 | 7906 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 7907 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 7908 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 7909 | bool ok; |
a70d6342 | 7910 | tree scalar_type, vectype; |
355fe088 | 7911 | gimple *pattern_stmt; |
363477c0 | 7912 | gimple_seq pattern_def_seq; |
ebfd146a | 7913 | |
73fbfcad | 7914 | if (dump_enabled_p ()) |
ebfd146a | 7915 | { |
78c60e3d SS |
7916 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
7917 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 7918 | } |
ebfd146a | 7919 | |
1825a1f3 | 7920 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 7921 | { |
73fbfcad | 7922 | if (dump_enabled_p ()) |
78c60e3d | 7923 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7924 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
7925 | |
7926 | return false; | |
7927 | } | |
b8698a0f L |
7928 | |
7929 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
7930 | to include: |
7931 | - the COND_EXPR which is the loop exit condition | |
7932 | - any LABEL_EXPRs in the loop | |
b8698a0f | 7933 | - computations that are used only for array indexing or loop control. |
8644a673 | 7934 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 7935 | instance, therefore, all the statements are relevant. |
ebfd146a | 7936 | |
d092494c | 7937 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 7938 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
7939 | statements. In basic blocks we are called from some SLP instance |
7940 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
7941 | already will be part of SLP instance. */ | |
83197f37 IR |
7942 | |
7943 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 7944 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 7945 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 7946 | { |
9d5e7640 | 7947 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 7948 | && pattern_stmt |
9d5e7640 IR |
7949 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
7950 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
7951 | { | |
83197f37 | 7952 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
7953 | stmt = pattern_stmt; |
7954 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 7955 | if (dump_enabled_p ()) |
9d5e7640 | 7956 | { |
78c60e3d SS |
7957 | dump_printf_loc (MSG_NOTE, vect_location, |
7958 | "==> examining pattern statement: "); | |
7959 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
7960 | } |
7961 | } | |
7962 | else | |
7963 | { | |
73fbfcad | 7964 | if (dump_enabled_p ()) |
e645e942 | 7965 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 7966 | |
9d5e7640 IR |
7967 | return true; |
7968 | } | |
8644a673 | 7969 | } |
83197f37 | 7970 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 7971 | && node == NULL |
83197f37 IR |
7972 | && pattern_stmt |
7973 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
7974 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
7975 | { | |
7976 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 7977 | if (dump_enabled_p ()) |
83197f37 | 7978 | { |
78c60e3d SS |
7979 | dump_printf_loc (MSG_NOTE, vect_location, |
7980 | "==> examining pattern statement: "); | |
7981 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
7982 | } |
7983 | ||
7984 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
7985 | return false; | |
7986 | } | |
ebfd146a | 7987 | |
1107f3ae | 7988 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 7989 | && node == NULL |
363477c0 | 7990 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 7991 | { |
363477c0 | 7992 | gimple_stmt_iterator si; |
1107f3ae | 7993 | |
363477c0 JJ |
7994 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
7995 | { | |
355fe088 | 7996 | gimple *pattern_def_stmt = gsi_stmt (si); |
363477c0 JJ |
7997 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) |
7998 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
7999 | { | |
8000 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 8001 | if (dump_enabled_p ()) |
363477c0 | 8002 | { |
78c60e3d SS |
8003 | dump_printf_loc (MSG_NOTE, vect_location, |
8004 | "==> examining pattern def statement: "); | |
8005 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 8006 | } |
1107f3ae | 8007 | |
363477c0 JJ |
8008 | if (!vect_analyze_stmt (pattern_def_stmt, |
8009 | need_to_vectorize, node)) | |
8010 | return false; | |
8011 | } | |
8012 | } | |
8013 | } | |
1107f3ae | 8014 | |
8644a673 IR |
8015 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
8016 | { | |
8017 | case vect_internal_def: | |
8018 | break; | |
ebfd146a | 8019 | |
8644a673 | 8020 | case vect_reduction_def: |
7c5222ff | 8021 | case vect_nested_cycle: |
14a61437 RB |
8022 | gcc_assert (!bb_vinfo |
8023 | && (relevance == vect_used_in_outer | |
8024 | || relevance == vect_used_in_outer_by_reduction | |
8025 | || relevance == vect_used_by_reduction | |
b28ead45 AH |
8026 | || relevance == vect_unused_in_scope |
8027 | || relevance == vect_used_only_live)); | |
8644a673 IR |
8028 | break; |
8029 | ||
8030 | case vect_induction_def: | |
8031 | case vect_constant_def: | |
8032 | case vect_external_def: | |
8033 | case vect_unknown_def_type: | |
8034 | default: | |
8035 | gcc_unreachable (); | |
8036 | } | |
ebfd146a | 8037 | |
a70d6342 IR |
8038 | if (bb_vinfo) |
8039 | { | |
8040 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
8041 | ||
b690cc0f | 8042 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
73fbfcad | 8043 | if (dump_enabled_p ()) |
a70d6342 | 8044 | { |
78c60e3d SS |
8045 | dump_printf_loc (MSG_NOTE, vect_location, |
8046 | "get vectype for scalar type: "); | |
8047 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
e645e942 | 8048 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
8049 | } |
8050 | ||
8051 | vectype = get_vectype_for_scalar_type (scalar_type); | |
8052 | if (!vectype) | |
8053 | { | |
73fbfcad | 8054 | if (dump_enabled_p ()) |
a70d6342 | 8055 | { |
78c60e3d SS |
8056 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8057 | "not SLPed: unsupported data-type "); | |
8058 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
8059 | scalar_type); | |
e645e942 | 8060 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
a70d6342 IR |
8061 | } |
8062 | return false; | |
8063 | } | |
8064 | ||
73fbfcad | 8065 | if (dump_enabled_p ()) |
a70d6342 | 8066 | { |
78c60e3d SS |
8067 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); |
8068 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
e645e942 | 8069 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
8070 | } |
8071 | ||
8072 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
8073 | } | |
8074 | ||
8644a673 | 8075 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 8076 | { |
8644a673 | 8077 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
8078 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
8079 | || (is_gimple_call (stmt) | |
8080 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 8081 | *need_to_vectorize = true; |
ebfd146a IR |
8082 | } |
8083 | ||
b1af7da6 RB |
8084 | if (PURE_SLP_STMT (stmt_info) && !node) |
8085 | { | |
8086 | dump_printf_loc (MSG_NOTE, vect_location, | |
8087 | "handled only by SLP analysis\n"); | |
8088 | return true; | |
8089 | } | |
8090 | ||
8091 | ok = true; | |
8092 | if (!bb_vinfo | |
8093 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
8094 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
8095 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8096 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8097 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8098 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8099 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8100 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8101 | || vectorizable_call (stmt, NULL, NULL, node) | |
8102 | || vectorizable_store (stmt, NULL, NULL, node) | |
8103 | || vectorizable_reduction (stmt, NULL, NULL, node) | |
42fd8198 IE |
8104 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8105 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 RB |
8106 | else |
8107 | { | |
8108 | if (bb_vinfo) | |
8109 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8110 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8111 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8112 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8113 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8114 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8115 | || vectorizable_call (stmt, NULL, NULL, node) | |
8116 | || vectorizable_store (stmt, NULL, NULL, node) | |
42fd8198 IE |
8117 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8118 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 | 8119 | } |
8644a673 IR |
8120 | |
8121 | if (!ok) | |
ebfd146a | 8122 | { |
73fbfcad | 8123 | if (dump_enabled_p ()) |
8644a673 | 8124 | { |
78c60e3d SS |
8125 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8126 | "not vectorized: relevant stmt not "); | |
8127 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
8128 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 8129 | } |
b8698a0f | 8130 | |
ebfd146a IR |
8131 | return false; |
8132 | } | |
8133 | ||
a70d6342 IR |
8134 | if (bb_vinfo) |
8135 | return true; | |
8136 | ||
8644a673 IR |
8137 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
8138 | need extra handling, except for vectorizable reductions. */ | |
8139 | if (STMT_VINFO_LIVE_P (stmt_info) | |
8140 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
b28ead45 | 8141 | ok = vectorizable_live_operation (stmt, NULL, NULL, -1, NULL); |
ebfd146a | 8142 | |
8644a673 | 8143 | if (!ok) |
ebfd146a | 8144 | { |
73fbfcad | 8145 | if (dump_enabled_p ()) |
8644a673 | 8146 | { |
78c60e3d SS |
8147 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8148 | "not vectorized: live stmt not "); | |
8149 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
8150 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 8151 | } |
b8698a0f | 8152 | |
8644a673 | 8153 | return false; |
ebfd146a IR |
8154 | } |
8155 | ||
ebfd146a IR |
8156 | return true; |
8157 | } | |
8158 | ||
8159 | ||
8160 | /* Function vect_transform_stmt. | |
8161 | ||
8162 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
8163 | ||
8164 | bool | |
355fe088 | 8165 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 8166 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
8167 | slp_instance slp_node_instance) |
8168 | { | |
8169 | bool is_store = false; | |
355fe088 | 8170 | gimple *vec_stmt = NULL; |
ebfd146a | 8171 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
ebfd146a | 8172 | bool done; |
ebfd146a | 8173 | |
fce57248 | 8174 | gcc_assert (slp_node || !PURE_SLP_STMT (stmt_info)); |
355fe088 | 8175 | gimple *old_vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 8176 | |
ebfd146a IR |
8177 | switch (STMT_VINFO_TYPE (stmt_info)) |
8178 | { | |
8179 | case type_demotion_vec_info_type: | |
ebfd146a | 8180 | case type_promotion_vec_info_type: |
ebfd146a IR |
8181 | case type_conversion_vec_info_type: |
8182 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
8183 | gcc_assert (done); | |
8184 | break; | |
8185 | ||
8186 | case induc_vec_info_type: | |
8187 | gcc_assert (!slp_node); | |
8188 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
8189 | gcc_assert (done); | |
8190 | break; | |
8191 | ||
9dc3f7de IR |
8192 | case shift_vec_info_type: |
8193 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
8194 | gcc_assert (done); | |
8195 | break; | |
8196 | ||
ebfd146a IR |
8197 | case op_vec_info_type: |
8198 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
8199 | gcc_assert (done); | |
8200 | break; | |
8201 | ||
8202 | case assignment_vec_info_type: | |
8203 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
8204 | gcc_assert (done); | |
8205 | break; | |
8206 | ||
8207 | case load_vec_info_type: | |
b8698a0f | 8208 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
8209 | slp_node_instance); |
8210 | gcc_assert (done); | |
8211 | break; | |
8212 | ||
8213 | case store_vec_info_type: | |
8214 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
8215 | gcc_assert (done); | |
0d0293ac | 8216 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
8217 | { |
8218 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 8219 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
8220 | one are skipped, and there vec_stmt_info shouldn't be freed |
8221 | meanwhile. */ | |
0d0293ac | 8222 | *grouped_store = true; |
ebfd146a IR |
8223 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
8224 | is_store = true; | |
8225 | } | |
8226 | else | |
8227 | is_store = true; | |
8228 | break; | |
8229 | ||
8230 | case condition_vec_info_type: | |
f7e531cf | 8231 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
8232 | gcc_assert (done); |
8233 | break; | |
8234 | ||
42fd8198 IE |
8235 | case comparison_vec_info_type: |
8236 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node); | |
8237 | gcc_assert (done); | |
8238 | break; | |
8239 | ||
ebfd146a | 8240 | case call_vec_info_type: |
190c2236 | 8241 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 8242 | stmt = gsi_stmt (*gsi); |
5ce9450f JJ |
8243 | if (is_gimple_call (stmt) |
8244 | && gimple_call_internal_p (stmt) | |
8245 | && gimple_call_internal_fn (stmt) == IFN_MASK_STORE) | |
8246 | is_store = true; | |
ebfd146a IR |
8247 | break; |
8248 | ||
0136f8f0 AH |
8249 | case call_simd_clone_vec_info_type: |
8250 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node); | |
8251 | stmt = gsi_stmt (*gsi); | |
8252 | break; | |
8253 | ||
ebfd146a | 8254 | case reduc_vec_info_type: |
b5aeb3bb | 8255 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
8256 | gcc_assert (done); |
8257 | break; | |
8258 | ||
8259 | default: | |
8260 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
8261 | { | |
73fbfcad | 8262 | if (dump_enabled_p ()) |
78c60e3d | 8263 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8264 | "stmt not supported.\n"); |
ebfd146a IR |
8265 | gcc_unreachable (); |
8266 | } | |
8267 | } | |
8268 | ||
225ce44b RB |
8269 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
8270 | This would break hybrid SLP vectorization. */ | |
8271 | if (slp_node) | |
d90f8440 RB |
8272 | gcc_assert (!vec_stmt |
8273 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt); | |
225ce44b | 8274 | |
ebfd146a IR |
8275 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
8276 | is being vectorized, but outside the immediately enclosing loop. */ | |
8277 | if (vec_stmt | |
a70d6342 IR |
8278 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
8279 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
8280 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
8281 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
8282 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 8283 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 8284 | vect_used_in_outer_by_reduction)) |
ebfd146a | 8285 | { |
a70d6342 IR |
8286 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
8287 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
8288 | imm_use_iterator imm_iter; |
8289 | use_operand_p use_p; | |
8290 | tree scalar_dest; | |
355fe088 | 8291 | gimple *exit_phi; |
ebfd146a | 8292 | |
73fbfcad | 8293 | if (dump_enabled_p ()) |
78c60e3d | 8294 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 8295 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
8296 | |
8297 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
8298 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
8299 | STMT). */ | |
8300 | if (gimple_code (stmt) == GIMPLE_PHI) | |
8301 | scalar_dest = PHI_RESULT (stmt); | |
8302 | else | |
8303 | scalar_dest = gimple_assign_lhs (stmt); | |
8304 | ||
8305 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
8306 | { | |
8307 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
8308 | { | |
8309 | exit_phi = USE_STMT (use_p); | |
8310 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
8311 | } | |
8312 | } | |
8313 | } | |
8314 | ||
8315 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
8316 | being vectorized. */ | |
b28ead45 AH |
8317 | if (slp_node) |
8318 | { | |
8319 | gimple *slp_stmt; | |
8320 | int i; | |
8321 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, slp_stmt) | |
8322 | { | |
8323 | stmt_vec_info slp_stmt_info = vinfo_for_stmt (slp_stmt); | |
8324 | if (STMT_VINFO_LIVE_P (slp_stmt_info) | |
8325 | && STMT_VINFO_TYPE (slp_stmt_info) != reduc_vec_info_type) | |
8326 | { | |
8327 | done = vectorizable_live_operation (slp_stmt, gsi, slp_node, i, | |
8328 | &vec_stmt); | |
8329 | gcc_assert (done); | |
8330 | } | |
8331 | } | |
8332 | } | |
8333 | else if (STMT_VINFO_LIVE_P (stmt_info) | |
ebfd146a IR |
8334 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) |
8335 | { | |
b28ead45 | 8336 | done = vectorizable_live_operation (stmt, gsi, slp_node, -1, &vec_stmt); |
ebfd146a IR |
8337 | gcc_assert (done); |
8338 | } | |
8339 | ||
8340 | if (vec_stmt) | |
83197f37 | 8341 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 8342 | |
b8698a0f | 8343 | return is_store; |
ebfd146a IR |
8344 | } |
8345 | ||
8346 | ||
b8698a0f | 8347 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
8348 | stmt_vec_info. */ |
8349 | ||
8350 | void | |
355fe088 | 8351 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 8352 | { |
355fe088 TS |
8353 | gimple *next = first_stmt; |
8354 | gimple *tmp; | |
ebfd146a IR |
8355 | gimple_stmt_iterator next_si; |
8356 | ||
8357 | while (next) | |
8358 | { | |
78048b1c JJ |
8359 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
8360 | ||
8361 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
8362 | if (is_pattern_stmt_p (stmt_info)) | |
8363 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
8364 | /* Free the attached stmt_vec_info and remove the stmt. */ |
8365 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 8366 | unlink_stmt_vdef (next); |
ebfd146a | 8367 | gsi_remove (&next_si, true); |
3d3f2249 | 8368 | release_defs (next); |
ebfd146a IR |
8369 | free_stmt_vec_info (next); |
8370 | next = tmp; | |
8371 | } | |
8372 | } | |
8373 | ||
8374 | ||
8375 | /* Function new_stmt_vec_info. | |
8376 | ||
8377 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
8378 | ||
8379 | stmt_vec_info | |
310213d4 | 8380 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
8381 | { |
8382 | stmt_vec_info res; | |
8383 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
8384 | ||
8385 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
8386 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 8387 | res->vinfo = vinfo; |
8644a673 | 8388 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
8389 | STMT_VINFO_LIVE_P (res) = false; |
8390 | STMT_VINFO_VECTYPE (res) = NULL; | |
8391 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 8392 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
8393 | STMT_VINFO_IN_PATTERN_P (res) = false; |
8394 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 8395 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 8396 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 8397 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
ebfd146a IR |
8398 | |
8399 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
8400 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
8401 | STMT_VINFO_DR_INIT (res) = NULL; | |
8402 | STMT_VINFO_DR_STEP (res) = NULL; | |
8403 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
8404 | ||
8405 | if (gimple_code (stmt) == GIMPLE_PHI | |
8406 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
8407 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
8408 | else | |
8644a673 IR |
8409 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
8410 | ||
9771b263 | 8411 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 8412 | STMT_SLP_TYPE (res) = loop_vect; |
78810bd3 RB |
8413 | STMT_VINFO_NUM_SLP_USES (res) = 0; |
8414 | ||
e14c1050 IR |
8415 | GROUP_FIRST_ELEMENT (res) = NULL; |
8416 | GROUP_NEXT_ELEMENT (res) = NULL; | |
8417 | GROUP_SIZE (res) = 0; | |
8418 | GROUP_STORE_COUNT (res) = 0; | |
8419 | GROUP_GAP (res) = 0; | |
8420 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
8421 | |
8422 | return res; | |
8423 | } | |
8424 | ||
8425 | ||
8426 | /* Create a hash table for stmt_vec_info. */ | |
8427 | ||
8428 | void | |
8429 | init_stmt_vec_info_vec (void) | |
8430 | { | |
9771b263 DN |
8431 | gcc_assert (!stmt_vec_info_vec.exists ()); |
8432 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
8433 | } |
8434 | ||
8435 | ||
8436 | /* Free hash table for stmt_vec_info. */ | |
8437 | ||
8438 | void | |
8439 | free_stmt_vec_info_vec (void) | |
8440 | { | |
93675444 | 8441 | unsigned int i; |
3161455c | 8442 | stmt_vec_info info; |
93675444 JJ |
8443 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) |
8444 | if (info != NULL) | |
3161455c | 8445 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
9771b263 DN |
8446 | gcc_assert (stmt_vec_info_vec.exists ()); |
8447 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
8448 | } |
8449 | ||
8450 | ||
8451 | /* Free stmt vectorization related info. */ | |
8452 | ||
8453 | void | |
355fe088 | 8454 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
8455 | { |
8456 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8457 | ||
8458 | if (!stmt_info) | |
8459 | return; | |
8460 | ||
78048b1c JJ |
8461 | /* Check if this statement has a related "pattern stmt" |
8462 | (introduced by the vectorizer during the pattern recognition | |
8463 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
8464 | too. */ | |
8465 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
8466 | { | |
8467 | stmt_vec_info patt_info | |
8468 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
8469 | if (patt_info) | |
8470 | { | |
363477c0 | 8471 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
355fe088 | 8472 | gimple *patt_stmt = STMT_VINFO_STMT (patt_info); |
f0281fde RB |
8473 | gimple_set_bb (patt_stmt, NULL); |
8474 | tree lhs = gimple_get_lhs (patt_stmt); | |
e6f5c25d | 8475 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde | 8476 | release_ssa_name (lhs); |
363477c0 JJ |
8477 | if (seq) |
8478 | { | |
8479 | gimple_stmt_iterator si; | |
8480 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
f0281fde | 8481 | { |
355fe088 | 8482 | gimple *seq_stmt = gsi_stmt (si); |
f0281fde | 8483 | gimple_set_bb (seq_stmt, NULL); |
7532abf2 | 8484 | lhs = gimple_get_lhs (seq_stmt); |
e6f5c25d | 8485 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde RB |
8486 | release_ssa_name (lhs); |
8487 | free_stmt_vec_info (seq_stmt); | |
8488 | } | |
363477c0 | 8489 | } |
f0281fde | 8490 | free_stmt_vec_info (patt_stmt); |
78048b1c JJ |
8491 | } |
8492 | } | |
8493 | ||
9771b263 | 8494 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 8495 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
8496 | set_vinfo_for_stmt (stmt, NULL); |
8497 | free (stmt_info); | |
8498 | } | |
8499 | ||
8500 | ||
bb67d9c7 | 8501 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 8502 | |
bb67d9c7 | 8503 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
8504 | by the target. */ |
8505 | ||
bb67d9c7 RG |
8506 | static tree |
8507 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a | 8508 | { |
ef4bddc2 RS |
8509 | machine_mode inner_mode = TYPE_MODE (scalar_type); |
8510 | machine_mode simd_mode; | |
2f816591 | 8511 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
8512 | int nunits; |
8513 | tree vectype; | |
8514 | ||
cc4b5170 | 8515 | if (nbytes == 0) |
ebfd146a IR |
8516 | return NULL_TREE; |
8517 | ||
48f2e373 RB |
8518 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
8519 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
8520 | return NULL_TREE; | |
8521 | ||
7b7b1813 RG |
8522 | /* For vector types of elements whose mode precision doesn't |
8523 | match their types precision we use a element type of mode | |
8524 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
8525 | they support the proper result truncation/extension. |
8526 | We also make sure to build vector types with INTEGER_TYPE | |
8527 | component type only. */ | |
6d7971b8 | 8528 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
8529 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
8530 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
8531 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
8532 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 8533 | |
ccbf5bb4 RG |
8534 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
8535 | When the component mode passes the above test simply use a type | |
8536 | corresponding to that mode. The theory is that any use that | |
8537 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 8538 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 8539 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
8540 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
8541 | ||
8542 | /* We can't build a vector type of elements with alignment bigger than | |
8543 | their size. */ | |
dfc2e2ac | 8544 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
8545 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
8546 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 8547 | |
dfc2e2ac RB |
8548 | /* If we felt back to using the mode fail if there was |
8549 | no scalar type for it. */ | |
8550 | if (scalar_type == NULL_TREE) | |
8551 | return NULL_TREE; | |
8552 | ||
bb67d9c7 RG |
8553 | /* If no size was supplied use the mode the target prefers. Otherwise |
8554 | lookup a vector mode of the specified size. */ | |
8555 | if (size == 0) | |
8556 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
8557 | else | |
8558 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
8559 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
8560 | if (nunits <= 1) | |
8561 | return NULL_TREE; | |
ebfd146a IR |
8562 | |
8563 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
8564 | |
8565 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
8566 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 8567 | return NULL_TREE; |
ebfd146a IR |
8568 | |
8569 | return vectype; | |
8570 | } | |
8571 | ||
bb67d9c7 RG |
8572 | unsigned int current_vector_size; |
8573 | ||
8574 | /* Function get_vectype_for_scalar_type. | |
8575 | ||
8576 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
8577 | by the target. */ | |
8578 | ||
8579 | tree | |
8580 | get_vectype_for_scalar_type (tree scalar_type) | |
8581 | { | |
8582 | tree vectype; | |
8583 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
8584 | current_vector_size); | |
8585 | if (vectype | |
8586 | && current_vector_size == 0) | |
8587 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
8588 | return vectype; | |
8589 | } | |
8590 | ||
42fd8198 IE |
8591 | /* Function get_mask_type_for_scalar_type. |
8592 | ||
8593 | Returns the mask type corresponding to a result of comparison | |
8594 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
8595 | ||
8596 | tree | |
8597 | get_mask_type_for_scalar_type (tree scalar_type) | |
8598 | { | |
8599 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
8600 | ||
8601 | if (!vectype) | |
8602 | return NULL; | |
8603 | ||
8604 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
8605 | current_vector_size); | |
8606 | } | |
8607 | ||
b690cc0f RG |
8608 | /* Function get_same_sized_vectype |
8609 | ||
8610 | Returns a vector type corresponding to SCALAR_TYPE of size | |
8611 | VECTOR_TYPE if supported by the target. */ | |
8612 | ||
8613 | tree | |
bb67d9c7 | 8614 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 8615 | { |
9f47c7e5 IE |
8616 | if (TREE_CODE (scalar_type) == BOOLEAN_TYPE) |
8617 | return build_same_sized_truth_vector_type (vector_type); | |
8618 | ||
bb67d9c7 RG |
8619 | return get_vectype_for_scalar_type_and_size |
8620 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
8621 | } |
8622 | ||
ebfd146a IR |
8623 | /* Function vect_is_simple_use. |
8624 | ||
8625 | Input: | |
81c40241 RB |
8626 | VINFO - the vect info of the loop or basic block that is being vectorized. |
8627 | OPERAND - operand in the loop or bb. | |
8628 | Output: | |
8629 | DEF_STMT - the defining stmt in case OPERAND is an SSA_NAME. | |
8630 | DT - the type of definition | |
ebfd146a IR |
8631 | |
8632 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 8633 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 8634 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 8635 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
8636 | is the case in reduction/induction computations). |
8637 | For basic blocks, supportable operands are constants and bb invariants. | |
8638 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
8639 | |
8640 | bool | |
81c40241 RB |
8641 | vect_is_simple_use (tree operand, vec_info *vinfo, |
8642 | gimple **def_stmt, enum vect_def_type *dt) | |
b8698a0f | 8643 | { |
ebfd146a | 8644 | *def_stmt = NULL; |
3fc356dc | 8645 | *dt = vect_unknown_def_type; |
b8698a0f | 8646 | |
73fbfcad | 8647 | if (dump_enabled_p ()) |
ebfd146a | 8648 | { |
78c60e3d SS |
8649 | dump_printf_loc (MSG_NOTE, vect_location, |
8650 | "vect_is_simple_use: operand "); | |
8651 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 8652 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 8653 | } |
b8698a0f | 8654 | |
b758f602 | 8655 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
8656 | { |
8657 | *dt = vect_constant_def; | |
8658 | return true; | |
8659 | } | |
b8698a0f | 8660 | |
ebfd146a IR |
8661 | if (is_gimple_min_invariant (operand)) |
8662 | { | |
8644a673 | 8663 | *dt = vect_external_def; |
ebfd146a IR |
8664 | return true; |
8665 | } | |
8666 | ||
ebfd146a IR |
8667 | if (TREE_CODE (operand) != SSA_NAME) |
8668 | { | |
73fbfcad | 8669 | if (dump_enabled_p ()) |
af29617a AH |
8670 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8671 | "not ssa-name.\n"); | |
ebfd146a IR |
8672 | return false; |
8673 | } | |
b8698a0f | 8674 | |
3fc356dc | 8675 | if (SSA_NAME_IS_DEFAULT_DEF (operand)) |
ebfd146a | 8676 | { |
3fc356dc RB |
8677 | *dt = vect_external_def; |
8678 | return true; | |
ebfd146a IR |
8679 | } |
8680 | ||
3fc356dc | 8681 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
73fbfcad | 8682 | if (dump_enabled_p ()) |
ebfd146a | 8683 | { |
78c60e3d SS |
8684 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
8685 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
8686 | } |
8687 | ||
61d371eb | 8688 | if (! vect_stmt_in_region_p (vinfo, *def_stmt)) |
8644a673 | 8689 | *dt = vect_external_def; |
ebfd146a IR |
8690 | else |
8691 | { | |
3fc356dc | 8692 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (*def_stmt); |
603cca93 | 8693 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
ebfd146a IR |
8694 | } |
8695 | ||
2e8ab70c RB |
8696 | if (dump_enabled_p ()) |
8697 | { | |
8698 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: "); | |
8699 | switch (*dt) | |
8700 | { | |
8701 | case vect_uninitialized_def: | |
8702 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
8703 | break; | |
8704 | case vect_constant_def: | |
8705 | dump_printf (MSG_NOTE, "constant\n"); | |
8706 | break; | |
8707 | case vect_external_def: | |
8708 | dump_printf (MSG_NOTE, "external\n"); | |
8709 | break; | |
8710 | case vect_internal_def: | |
8711 | dump_printf (MSG_NOTE, "internal\n"); | |
8712 | break; | |
8713 | case vect_induction_def: | |
8714 | dump_printf (MSG_NOTE, "induction\n"); | |
8715 | break; | |
8716 | case vect_reduction_def: | |
8717 | dump_printf (MSG_NOTE, "reduction\n"); | |
8718 | break; | |
8719 | case vect_double_reduction_def: | |
8720 | dump_printf (MSG_NOTE, "double reduction\n"); | |
8721 | break; | |
8722 | case vect_nested_cycle: | |
8723 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
8724 | break; | |
8725 | case vect_unknown_def_type: | |
8726 | dump_printf (MSG_NOTE, "unknown\n"); | |
8727 | break; | |
8728 | } | |
8729 | } | |
8730 | ||
81c40241 | 8731 | if (*dt == vect_unknown_def_type) |
ebfd146a | 8732 | { |
73fbfcad | 8733 | if (dump_enabled_p ()) |
78c60e3d | 8734 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8735 | "Unsupported pattern.\n"); |
ebfd146a IR |
8736 | return false; |
8737 | } | |
8738 | ||
ebfd146a IR |
8739 | switch (gimple_code (*def_stmt)) |
8740 | { | |
8741 | case GIMPLE_PHI: | |
ebfd146a | 8742 | case GIMPLE_ASSIGN: |
ebfd146a | 8743 | case GIMPLE_CALL: |
81c40241 | 8744 | break; |
ebfd146a | 8745 | default: |
73fbfcad | 8746 | if (dump_enabled_p ()) |
78c60e3d | 8747 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8748 | "unsupported defining stmt:\n"); |
ebfd146a IR |
8749 | return false; |
8750 | } | |
8751 | ||
8752 | return true; | |
8753 | } | |
8754 | ||
81c40241 | 8755 | /* Function vect_is_simple_use. |
b690cc0f | 8756 | |
81c40241 | 8757 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
8758 | type of OPERAND and stores it to *VECTYPE. If the definition of |
8759 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
8760 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
8761 | is responsible to compute the best suited vector type for the | |
8762 | scalar operand. */ | |
8763 | ||
8764 | bool | |
81c40241 RB |
8765 | vect_is_simple_use (tree operand, vec_info *vinfo, |
8766 | gimple **def_stmt, enum vect_def_type *dt, tree *vectype) | |
b690cc0f | 8767 | { |
81c40241 | 8768 | if (!vect_is_simple_use (operand, vinfo, def_stmt, dt)) |
b690cc0f RG |
8769 | return false; |
8770 | ||
8771 | /* Now get a vector type if the def is internal, otherwise supply | |
8772 | NULL_TREE and leave it up to the caller to figure out a proper | |
8773 | type for the use stmt. */ | |
8774 | if (*dt == vect_internal_def | |
8775 | || *dt == vect_induction_def | |
8776 | || *dt == vect_reduction_def | |
8777 | || *dt == vect_double_reduction_def | |
8778 | || *dt == vect_nested_cycle) | |
8779 | { | |
8780 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
8781 | |
8782 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
8783 | && !STMT_VINFO_RELEVANT (stmt_info) | |
8784 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 8785 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 8786 | |
b690cc0f RG |
8787 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
8788 | gcc_assert (*vectype != NULL_TREE); | |
8789 | } | |
8790 | else if (*dt == vect_uninitialized_def | |
8791 | || *dt == vect_constant_def | |
8792 | || *dt == vect_external_def) | |
8793 | *vectype = NULL_TREE; | |
8794 | else | |
8795 | gcc_unreachable (); | |
8796 | ||
8797 | return true; | |
8798 | } | |
8799 | ||
ebfd146a IR |
8800 | |
8801 | /* Function supportable_widening_operation | |
8802 | ||
b8698a0f L |
8803 | Check whether an operation represented by the code CODE is a |
8804 | widening operation that is supported by the target platform in | |
b690cc0f RG |
8805 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
8806 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 8807 | |
ebfd146a IR |
8808 | Widening operations we currently support are NOP (CONVERT), FLOAT |
8809 | and WIDEN_MULT. This function checks if these operations are supported | |
8810 | by the target platform either directly (via vector tree-codes), or via | |
8811 | target builtins. | |
8812 | ||
8813 | Output: | |
b8698a0f L |
8814 | - CODE1 and CODE2 are codes of vector operations to be used when |
8815 | vectorizing the operation, if available. | |
ebfd146a IR |
8816 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
8817 | case of multi-step conversion (like char->short->int - in that case | |
8818 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
8819 | - INTERM_TYPES contains the intermediate type required to perform the |
8820 | widening operation (short in the above example). */ | |
ebfd146a IR |
8821 | |
8822 | bool | |
355fe088 | 8823 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 8824 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
8825 | enum tree_code *code1, enum tree_code *code2, |
8826 | int *multi_step_cvt, | |
9771b263 | 8827 | vec<tree> *interm_types) |
ebfd146a IR |
8828 | { |
8829 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8830 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 8831 | struct loop *vect_loop = NULL; |
ef4bddc2 | 8832 | machine_mode vec_mode; |
81f40b79 | 8833 | enum insn_code icode1, icode2; |
ebfd146a | 8834 | optab optab1, optab2; |
b690cc0f RG |
8835 | tree vectype = vectype_in; |
8836 | tree wide_vectype = vectype_out; | |
ebfd146a | 8837 | enum tree_code c1, c2; |
4a00c761 JJ |
8838 | int i; |
8839 | tree prev_type, intermediate_type; | |
ef4bddc2 | 8840 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 8841 | optab optab3, optab4; |
ebfd146a | 8842 | |
4a00c761 | 8843 | *multi_step_cvt = 0; |
4ef69dfc IR |
8844 | if (loop_info) |
8845 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
8846 | ||
ebfd146a IR |
8847 | switch (code) |
8848 | { | |
8849 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
8850 | /* The result of a vectorized widening operation usually requires |
8851 | two vectors (because the widened results do not fit into one vector). | |
8852 | The generated vector results would normally be expected to be | |
8853 | generated in the same order as in the original scalar computation, | |
8854 | i.e. if 8 results are generated in each vector iteration, they are | |
8855 | to be organized as follows: | |
8856 | vect1: [res1,res2,res3,res4], | |
8857 | vect2: [res5,res6,res7,res8]. | |
8858 | ||
8859 | However, in the special case that the result of the widening | |
8860 | operation is used in a reduction computation only, the order doesn't | |
8861 | matter (because when vectorizing a reduction we change the order of | |
8862 | the computation). Some targets can take advantage of this and | |
8863 | generate more efficient code. For example, targets like Altivec, | |
8864 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
8865 | generate the following vectors: | |
8866 | vect1: [res1,res3,res5,res7], | |
8867 | vect2: [res2,res4,res6,res8]. | |
8868 | ||
8869 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
8870 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
8871 | iterations in parallel). We therefore don't allow to change the | |
8872 | order of the computation in the inner-loop during outer-loop | |
8873 | vectorization. */ | |
8874 | /* TODO: Another case in which order doesn't *really* matter is when we | |
8875 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
8876 | Normally, pack_trunc performs an even/odd permute, whereas the | |
8877 | repack from an even/odd expansion would be an interleave, which | |
8878 | would be significantly simpler for e.g. AVX2. */ | |
8879 | /* In any case, in order to avoid duplicating the code below, recurse | |
8880 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
8881 | are properly set up for the caller. If we fail, we'll continue with | |
8882 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
8883 | if (vect_loop | |
8884 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
8885 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
8886 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
8887 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
8888 | code1, code2, multi_step_cvt, |
8889 | interm_types)) | |
ebc047a2 CH |
8890 | { |
8891 | /* Elements in a vector with vect_used_by_reduction property cannot | |
8892 | be reordered if the use chain with this property does not have the | |
8893 | same operation. One such an example is s += a * b, where elements | |
8894 | in a and b cannot be reordered. Here we check if the vector defined | |
8895 | by STMT is only directly used in the reduction statement. */ | |
8896 | tree lhs = gimple_assign_lhs (stmt); | |
8897 | use_operand_p dummy; | |
355fe088 | 8898 | gimple *use_stmt; |
ebc047a2 CH |
8899 | stmt_vec_info use_stmt_info = NULL; |
8900 | if (single_imm_use (lhs, &dummy, &use_stmt) | |
8901 | && (use_stmt_info = vinfo_for_stmt (use_stmt)) | |
8902 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
8903 | return true; | |
8904 | } | |
4a00c761 JJ |
8905 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
8906 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
8907 | break; |
8908 | ||
81c40241 RB |
8909 | case DOT_PROD_EXPR: |
8910 | c1 = DOT_PROD_EXPR; | |
8911 | c2 = DOT_PROD_EXPR; | |
8912 | break; | |
8913 | ||
8914 | case SAD_EXPR: | |
8915 | c1 = SAD_EXPR; | |
8916 | c2 = SAD_EXPR; | |
8917 | break; | |
8918 | ||
6ae6116f RH |
8919 | case VEC_WIDEN_MULT_EVEN_EXPR: |
8920 | /* Support the recursion induced just above. */ | |
8921 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
8922 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
8923 | break; | |
8924 | ||
36ba4aae | 8925 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
8926 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
8927 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
8928 | break; |
8929 | ||
ebfd146a | 8930 | CASE_CONVERT: |
4a00c761 JJ |
8931 | c1 = VEC_UNPACK_LO_EXPR; |
8932 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
8933 | break; |
8934 | ||
8935 | case FLOAT_EXPR: | |
4a00c761 JJ |
8936 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
8937 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
8938 | break; |
8939 | ||
8940 | case FIX_TRUNC_EXPR: | |
8941 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
8942 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
8943 | computing the operation. */ | |
8944 | return false; | |
8945 | ||
8946 | default: | |
8947 | gcc_unreachable (); | |
8948 | } | |
8949 | ||
6ae6116f | 8950 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 8951 | std::swap (c1, c2); |
4a00c761 | 8952 | |
ebfd146a IR |
8953 | if (code == FIX_TRUNC_EXPR) |
8954 | { | |
8955 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
8956 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
8957 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
8958 | } |
8959 | else | |
8960 | { | |
8961 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
8962 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
8963 | } | |
8964 | ||
8965 | if (!optab1 || !optab2) | |
8966 | return false; | |
8967 | ||
8968 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
8969 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
8970 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
8971 | return false; |
8972 | ||
4a00c761 JJ |
8973 | *code1 = c1; |
8974 | *code2 = c2; | |
8975 | ||
8976 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
8977 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
8978 | /* For scalar masks we may have different boolean |
8979 | vector types having the same QImode. Thus we | |
8980 | add additional check for elements number. */ | |
8981 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
8982 | || (TYPE_VECTOR_SUBPARTS (vectype) / 2 | |
8983 | == TYPE_VECTOR_SUBPARTS (wide_vectype))); | |
4a00c761 | 8984 | |
b8698a0f | 8985 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 8986 | types. */ |
ebfd146a | 8987 | |
4a00c761 JJ |
8988 | prev_type = vectype; |
8989 | prev_mode = vec_mode; | |
b8698a0f | 8990 | |
4a00c761 JJ |
8991 | if (!CONVERT_EXPR_CODE_P (code)) |
8992 | return false; | |
b8698a0f | 8993 | |
4a00c761 JJ |
8994 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
8995 | intermediate steps in promotion sequence. We try | |
8996 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
8997 | not. */ | |
9771b263 | 8998 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
8999 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
9000 | { | |
9001 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
9002 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
9003 | { | |
9004 | intermediate_type | |
9005 | = build_truth_vector_type (TYPE_VECTOR_SUBPARTS (prev_type) / 2, | |
9006 | current_vector_size); | |
9007 | if (intermediate_mode != TYPE_MODE (intermediate_type)) | |
9008 | return false; | |
9009 | } | |
9010 | else | |
9011 | intermediate_type | |
9012 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
9013 | TYPE_UNSIGNED (prev_type)); | |
9014 | ||
4a00c761 JJ |
9015 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); |
9016 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
9017 | ||
9018 | if (!optab3 || !optab4 | |
9019 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
9020 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
9021 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
9022 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
9023 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
9024 | == CODE_FOR_nothing) | |
9025 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
9026 | == CODE_FOR_nothing)) | |
9027 | break; | |
ebfd146a | 9028 | |
9771b263 | 9029 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
9030 | (*multi_step_cvt)++; |
9031 | ||
9032 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
9033 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
9034 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
9035 | || (TYPE_VECTOR_SUBPARTS (intermediate_type) / 2 | |
9036 | == TYPE_VECTOR_SUBPARTS (wide_vectype))); | |
4a00c761 JJ |
9037 | |
9038 | prev_type = intermediate_type; | |
9039 | prev_mode = intermediate_mode; | |
ebfd146a IR |
9040 | } |
9041 | ||
9771b263 | 9042 | interm_types->release (); |
4a00c761 | 9043 | return false; |
ebfd146a IR |
9044 | } |
9045 | ||
9046 | ||
9047 | /* Function supportable_narrowing_operation | |
9048 | ||
b8698a0f L |
9049 | Check whether an operation represented by the code CODE is a |
9050 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
9051 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
9052 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 9053 | |
ebfd146a | 9054 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 9055 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
9056 | the target platform directly via vector tree-codes. |
9057 | ||
9058 | Output: | |
b8698a0f L |
9059 | - CODE1 is the code of a vector operation to be used when |
9060 | vectorizing the operation, if available. | |
ebfd146a IR |
9061 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
9062 | case of multi-step conversion (like int->short->char - in that case | |
9063 | MULTI_STEP_CVT will be 1). | |
9064 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 9065 | narrowing operation (short in the above example). */ |
ebfd146a IR |
9066 | |
9067 | bool | |
9068 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 9069 | tree vectype_out, tree vectype_in, |
ebfd146a | 9070 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 9071 | vec<tree> *interm_types) |
ebfd146a | 9072 | { |
ef4bddc2 | 9073 | machine_mode vec_mode; |
ebfd146a IR |
9074 | enum insn_code icode1; |
9075 | optab optab1, interm_optab; | |
b690cc0f RG |
9076 | tree vectype = vectype_in; |
9077 | tree narrow_vectype = vectype_out; | |
ebfd146a | 9078 | enum tree_code c1; |
3ae0661a | 9079 | tree intermediate_type, prev_type; |
ef4bddc2 | 9080 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 9081 | int i; |
4a00c761 | 9082 | bool uns; |
ebfd146a | 9083 | |
4a00c761 | 9084 | *multi_step_cvt = 0; |
ebfd146a IR |
9085 | switch (code) |
9086 | { | |
9087 | CASE_CONVERT: | |
9088 | c1 = VEC_PACK_TRUNC_EXPR; | |
9089 | break; | |
9090 | ||
9091 | case FIX_TRUNC_EXPR: | |
9092 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
9093 | break; | |
9094 | ||
9095 | case FLOAT_EXPR: | |
9096 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
9097 | tree code and optabs used for computing the operation. */ | |
9098 | return false; | |
9099 | ||
9100 | default: | |
9101 | gcc_unreachable (); | |
9102 | } | |
9103 | ||
9104 | if (code == FIX_TRUNC_EXPR) | |
9105 | /* The signedness is determined from output operand. */ | |
b690cc0f | 9106 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
9107 | else |
9108 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
9109 | ||
9110 | if (!optab1) | |
9111 | return false; | |
9112 | ||
9113 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 9114 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
9115 | return false; |
9116 | ||
4a00c761 JJ |
9117 | *code1 = c1; |
9118 | ||
9119 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
9120 | /* For scalar masks we may have different boolean |
9121 | vector types having the same QImode. Thus we | |
9122 | add additional check for elements number. */ | |
9123 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
9124 | || (TYPE_VECTOR_SUBPARTS (vectype) * 2 | |
9125 | == TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 | 9126 | |
ebfd146a IR |
9127 | /* Check if it's a multi-step conversion that can be done using intermediate |
9128 | types. */ | |
4a00c761 | 9129 | prev_mode = vec_mode; |
3ae0661a | 9130 | prev_type = vectype; |
4a00c761 JJ |
9131 | if (code == FIX_TRUNC_EXPR) |
9132 | uns = TYPE_UNSIGNED (vectype_out); | |
9133 | else | |
9134 | uns = TYPE_UNSIGNED (vectype); | |
9135 | ||
9136 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
9137 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
9138 | costly than signed. */ | |
9139 | if (code == FIX_TRUNC_EXPR && uns) | |
9140 | { | |
9141 | enum insn_code icode2; | |
9142 | ||
9143 | intermediate_type | |
9144 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
9145 | interm_optab | |
9146 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 9147 | if (interm_optab != unknown_optab |
4a00c761 JJ |
9148 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
9149 | && insn_data[icode1].operand[0].mode | |
9150 | == insn_data[icode2].operand[0].mode) | |
9151 | { | |
9152 | uns = false; | |
9153 | optab1 = interm_optab; | |
9154 | icode1 = icode2; | |
9155 | } | |
9156 | } | |
ebfd146a | 9157 | |
4a00c761 JJ |
9158 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
9159 | intermediate steps in promotion sequence. We try | |
9160 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 9161 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
9162 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
9163 | { | |
9164 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
9165 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
9166 | { | |
9167 | intermediate_type | |
9168 | = build_truth_vector_type (TYPE_VECTOR_SUBPARTS (prev_type) * 2, | |
9169 | current_vector_size); | |
9170 | if (intermediate_mode != TYPE_MODE (intermediate_type)) | |
9171 | return false; | |
9172 | } | |
9173 | else | |
9174 | intermediate_type | |
9175 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
4a00c761 JJ |
9176 | interm_optab |
9177 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
9178 | optab_default); | |
9179 | if (!interm_optab | |
9180 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
9181 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
9182 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
9183 | == CODE_FOR_nothing)) | |
9184 | break; | |
9185 | ||
9771b263 | 9186 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
9187 | (*multi_step_cvt)++; |
9188 | ||
9189 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
9190 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
9191 | || (TYPE_VECTOR_SUBPARTS (intermediate_type) * 2 | |
9192 | == TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 JJ |
9193 | |
9194 | prev_mode = intermediate_mode; | |
3ae0661a | 9195 | prev_type = intermediate_type; |
4a00c761 | 9196 | optab1 = interm_optab; |
ebfd146a IR |
9197 | } |
9198 | ||
9771b263 | 9199 | interm_types->release (); |
4a00c761 | 9200 | return false; |
ebfd146a | 9201 | } |