]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
cbe34bb5 | 2 | Copyright (C) 2003-2017 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" |
5ebaa477 | 51 | #include "tree-vector-builder.h" |
f151c9e1 | 52 | #include "vec-perm-indices.h" |
ebfd146a | 53 | |
7ee2468b SB |
54 | /* For lang_hooks.types.type_for_mode. */ |
55 | #include "langhooks.h" | |
ebfd146a | 56 | |
2de001ee RS |
57 | /* Says whether a statement is a load, a store of a vectorized statement |
58 | result, or a store of an invariant value. */ | |
59 | enum vec_load_store_type { | |
60 | VLS_LOAD, | |
61 | VLS_STORE, | |
62 | VLS_STORE_INVARIANT | |
63 | }; | |
64 | ||
c3e7ee41 BS |
65 | /* Return the vectorized type for the given statement. */ |
66 | ||
67 | tree | |
68 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
69 | { | |
70 | return STMT_VINFO_VECTYPE (stmt_info); | |
71 | } | |
72 | ||
73 | /* Return TRUE iff the given statement is in an inner loop relative to | |
74 | the loop being vectorized. */ | |
75 | bool | |
76 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
77 | { | |
355fe088 | 78 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
c3e7ee41 BS |
79 | basic_block bb = gimple_bb (stmt); |
80 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
81 | struct loop* loop; | |
82 | ||
83 | if (!loop_vinfo) | |
84 | return false; | |
85 | ||
86 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
87 | ||
88 | return (bb->loop_father == loop->inner); | |
89 | } | |
90 | ||
91 | /* Record the cost of a statement, either by directly informing the | |
92 | target model or by saving it in a vector for later processing. | |
93 | Return a preliminary estimate of the statement's cost. */ | |
94 | ||
95 | unsigned | |
92345349 | 96 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 97 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 98 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 99 | { |
cc9fe6bb JH |
100 | if ((kind == vector_load || kind == unaligned_load) |
101 | && STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
102 | kind = vector_gather_load; | |
103 | if ((kind == vector_store || kind == unaligned_store) | |
104 | && STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
105 | kind = vector_scatter_store; | |
92345349 | 106 | if (body_cost_vec) |
c3e7ee41 | 107 | { |
92345349 | 108 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
ddf56386 RB |
109 | stmt_info_for_cost si = { count, kind, |
110 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
111 | misalign }; | |
112 | body_cost_vec->safe_push (si); | |
c3e7ee41 | 113 | return (unsigned) |
92345349 | 114 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
115 | } |
116 | else | |
310213d4 RB |
117 | return add_stmt_cost (stmt_info->vinfo->target_cost_data, |
118 | count, kind, stmt_info, misalign, where); | |
c3e7ee41 BS |
119 | } |
120 | ||
272c6793 RS |
121 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
122 | ||
123 | static tree | |
124 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
125 | { | |
126 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
127 | "vect_array"); | |
128 | } | |
129 | ||
130 | /* ARRAY is an array of vectors created by create_vector_array. | |
131 | Return an SSA_NAME for the vector in index N. The reference | |
132 | is part of the vectorization of STMT and the vector is associated | |
133 | with scalar destination SCALAR_DEST. */ | |
134 | ||
135 | static tree | |
355fe088 | 136 | read_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree scalar_dest, |
272c6793 RS |
137 | tree array, unsigned HOST_WIDE_INT n) |
138 | { | |
139 | tree vect_type, vect, vect_name, array_ref; | |
355fe088 | 140 | gimple *new_stmt; |
272c6793 RS |
141 | |
142 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
143 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
144 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
145 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
146 | build_int_cst (size_type_node, n), | |
147 | NULL_TREE, NULL_TREE); | |
148 | ||
149 | new_stmt = gimple_build_assign (vect, array_ref); | |
150 | vect_name = make_ssa_name (vect, new_stmt); | |
151 | gimple_assign_set_lhs (new_stmt, vect_name); | |
152 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
153 | |
154 | return vect_name; | |
155 | } | |
156 | ||
157 | /* ARRAY is an array of vectors created by create_vector_array. | |
158 | Emit code to store SSA_NAME VECT in index N of the array. | |
159 | The store is part of the vectorization of STMT. */ | |
160 | ||
161 | static void | |
355fe088 | 162 | write_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree vect, |
272c6793 RS |
163 | tree array, unsigned HOST_WIDE_INT n) |
164 | { | |
165 | tree array_ref; | |
355fe088 | 166 | gimple *new_stmt; |
272c6793 RS |
167 | |
168 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
169 | build_int_cst (size_type_node, n), | |
170 | NULL_TREE, NULL_TREE); | |
171 | ||
172 | new_stmt = gimple_build_assign (array_ref, vect); | |
173 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
174 | } |
175 | ||
176 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
177 | of *PTR. The memory reference replaces those in FIRST_DR | |
178 | (and its group). */ | |
179 | ||
180 | static tree | |
44fc7854 | 181 | create_array_ref (tree type, tree ptr, tree alias_ptr_type) |
272c6793 | 182 | { |
44fc7854 | 183 | tree mem_ref; |
272c6793 | 184 | |
272c6793 RS |
185 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); |
186 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 187 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
188 | return mem_ref; |
189 | } | |
190 | ||
ebfd146a IR |
191 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
192 | ||
193 | /* Function vect_mark_relevant. | |
194 | ||
195 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
196 | ||
197 | static void | |
355fe088 | 198 | vect_mark_relevant (vec<gimple *> *worklist, gimple *stmt, |
97ecdb46 | 199 | enum vect_relevant relevant, bool live_p) |
ebfd146a IR |
200 | { |
201 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
202 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
203 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
355fe088 | 204 | gimple *pattern_stmt; |
ebfd146a | 205 | |
73fbfcad | 206 | if (dump_enabled_p ()) |
66c16fd9 RB |
207 | { |
208 | dump_printf_loc (MSG_NOTE, vect_location, | |
209 | "mark relevant %d, live %d: ", relevant, live_p); | |
210 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
211 | } | |
ebfd146a | 212 | |
83197f37 IR |
213 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
214 | related pattern stmt instead of the original stmt. However, such stmts | |
215 | may have their own uses that are not in any pattern, in such cases the | |
216 | stmt itself should be marked. */ | |
ebfd146a IR |
217 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
218 | { | |
97ecdb46 JJ |
219 | /* This is the last stmt in a sequence that was detected as a |
220 | pattern that can potentially be vectorized. Don't mark the stmt | |
221 | as relevant/live because it's not going to be vectorized. | |
222 | Instead mark the pattern-stmt that replaces it. */ | |
83197f37 | 223 | |
97ecdb46 JJ |
224 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); |
225 | ||
226 | if (dump_enabled_p ()) | |
227 | dump_printf_loc (MSG_NOTE, vect_location, | |
228 | "last stmt in pattern. don't mark" | |
229 | " relevant/live.\n"); | |
230 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
231 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
232 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
233 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
234 | stmt = pattern_stmt; | |
ebfd146a IR |
235 | } |
236 | ||
237 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
238 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
239 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
240 | ||
241 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
242 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
243 | { | |
73fbfcad | 244 | if (dump_enabled_p ()) |
78c60e3d | 245 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 246 | "already marked relevant/live.\n"); |
ebfd146a IR |
247 | return; |
248 | } | |
249 | ||
9771b263 | 250 | worklist->safe_push (stmt); |
ebfd146a IR |
251 | } |
252 | ||
253 | ||
b28ead45 AH |
254 | /* Function is_simple_and_all_uses_invariant |
255 | ||
256 | Return true if STMT is simple and all uses of it are invariant. */ | |
257 | ||
258 | bool | |
259 | is_simple_and_all_uses_invariant (gimple *stmt, loop_vec_info loop_vinfo) | |
260 | { | |
261 | tree op; | |
262 | gimple *def_stmt; | |
263 | ssa_op_iter iter; | |
264 | ||
265 | if (!is_gimple_assign (stmt)) | |
266 | return false; | |
267 | ||
268 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
269 | { | |
270 | enum vect_def_type dt = vect_uninitialized_def; | |
271 | ||
272 | if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt)) | |
273 | { | |
274 | if (dump_enabled_p ()) | |
275 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
276 | "use not simple.\n"); | |
277 | return false; | |
278 | } | |
279 | ||
280 | if (dt != vect_external_def && dt != vect_constant_def) | |
281 | return false; | |
282 | } | |
283 | return true; | |
284 | } | |
285 | ||
ebfd146a IR |
286 | /* Function vect_stmt_relevant_p. |
287 | ||
288 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
289 | "relevant for vectorization". | |
290 | ||
291 | A stmt is considered "relevant for vectorization" if: | |
292 | - it has uses outside the loop. | |
293 | - it has vdefs (it alters memory). | |
294 | - control stmts in the loop (except for the exit condition). | |
295 | ||
296 | CHECKME: what other side effects would the vectorizer allow? */ | |
297 | ||
298 | static bool | |
355fe088 | 299 | vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo, |
ebfd146a IR |
300 | enum vect_relevant *relevant, bool *live_p) |
301 | { | |
302 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
303 | ssa_op_iter op_iter; | |
304 | imm_use_iterator imm_iter; | |
305 | use_operand_p use_p; | |
306 | def_operand_p def_p; | |
307 | ||
8644a673 | 308 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
309 | *live_p = false; |
310 | ||
311 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
312 | if (is_ctrl_stmt (stmt) |
313 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
314 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 315 | *relevant = vect_used_in_scope; |
ebfd146a IR |
316 | |
317 | /* changing memory. */ | |
318 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
319 | if (gimple_vdef (stmt) |
320 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 321 | { |
73fbfcad | 322 | if (dump_enabled_p ()) |
78c60e3d | 323 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 324 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 325 | *relevant = vect_used_in_scope; |
ebfd146a IR |
326 | } |
327 | ||
328 | /* uses outside the loop. */ | |
329 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
330 | { | |
331 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
332 | { | |
333 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
334 | if (!flow_bb_inside_loop_p (loop, bb)) | |
335 | { | |
73fbfcad | 336 | if (dump_enabled_p ()) |
78c60e3d | 337 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 338 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 339 | |
3157b0c2 AO |
340 | if (is_gimple_debug (USE_STMT (use_p))) |
341 | continue; | |
342 | ||
ebfd146a IR |
343 | /* We expect all such uses to be in the loop exit phis |
344 | (because of loop closed form) */ | |
345 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
346 | gcc_assert (bb == single_exit (loop)->dest); | |
347 | ||
348 | *live_p = true; | |
349 | } | |
350 | } | |
351 | } | |
352 | ||
3a2edf4c AH |
353 | if (*live_p && *relevant == vect_unused_in_scope |
354 | && !is_simple_and_all_uses_invariant (stmt, loop_vinfo)) | |
b28ead45 AH |
355 | { |
356 | if (dump_enabled_p ()) | |
357 | dump_printf_loc (MSG_NOTE, vect_location, | |
358 | "vec_stmt_relevant_p: stmt live but not relevant.\n"); | |
359 | *relevant = vect_used_only_live; | |
360 | } | |
361 | ||
ebfd146a IR |
362 | return (*live_p || *relevant); |
363 | } | |
364 | ||
365 | ||
b8698a0f | 366 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 367 | |
ff802fa1 | 368 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
369 | used in STMT for anything other than indexing an array. */ |
370 | ||
371 | static bool | |
355fe088 | 372 | exist_non_indexing_operands_for_use_p (tree use, gimple *stmt) |
ebfd146a IR |
373 | { |
374 | tree operand; | |
375 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 376 | |
ff802fa1 | 377 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
378 | reference in STMT, then any operand that corresponds to USE |
379 | is not indexing an array. */ | |
380 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
381 | return true; | |
59a05b0c | 382 | |
ebfd146a IR |
383 | /* STMT has a data_ref. FORNOW this means that its of one of |
384 | the following forms: | |
385 | -1- ARRAY_REF = var | |
386 | -2- var = ARRAY_REF | |
387 | (This should have been verified in analyze_data_refs). | |
388 | ||
389 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 390 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
391 | for array indexing. |
392 | ||
393 | Therefore, all we need to check is if STMT falls into the | |
394 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
395 | |
396 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
397 | { |
398 | if (is_gimple_call (stmt) | |
399 | && gimple_call_internal_p (stmt)) | |
400 | switch (gimple_call_internal_fn (stmt)) | |
401 | { | |
402 | case IFN_MASK_STORE: | |
403 | operand = gimple_call_arg (stmt, 3); | |
404 | if (operand == use) | |
405 | return true; | |
406 | /* FALLTHRU */ | |
407 | case IFN_MASK_LOAD: | |
408 | operand = gimple_call_arg (stmt, 2); | |
409 | if (operand == use) | |
410 | return true; | |
411 | break; | |
412 | default: | |
413 | break; | |
414 | } | |
415 | return false; | |
416 | } | |
417 | ||
59a05b0c EB |
418 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
419 | return false; | |
ebfd146a | 420 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
421 | if (TREE_CODE (operand) != SSA_NAME) |
422 | return false; | |
423 | ||
424 | if (operand == use) | |
425 | return true; | |
426 | ||
427 | return false; | |
428 | } | |
429 | ||
430 | ||
b8698a0f | 431 | /* |
ebfd146a IR |
432 | Function process_use. |
433 | ||
434 | Inputs: | |
435 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b28ead45 | 436 | - RELEVANT - enum value to be set in the STMT_VINFO of the stmt |
ff802fa1 | 437 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 438 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
439 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
440 | be performed. | |
ebfd146a IR |
441 | |
442 | Outputs: | |
443 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
444 | relevance info of the DEF_STMT of this USE: | |
445 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
446 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
447 | Exceptions: | |
448 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 449 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 450 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
451 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
452 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
453 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
454 | be modified accordingly. | |
455 | ||
456 | Return true if everything is as expected. Return false otherwise. */ | |
457 | ||
458 | static bool | |
b28ead45 | 459 | process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo, |
355fe088 | 460 | enum vect_relevant relevant, vec<gimple *> *worklist, |
aec7ae7d | 461 | bool force) |
ebfd146a IR |
462 | { |
463 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
464 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
465 | stmt_vec_info dstmt_vinfo; | |
466 | basic_block bb, def_bb; | |
355fe088 | 467 | gimple *def_stmt; |
ebfd146a IR |
468 | enum vect_def_type dt; |
469 | ||
b8698a0f | 470 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 471 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 472 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
473 | return true; |
474 | ||
81c40241 | 475 | if (!vect_is_simple_use (use, loop_vinfo, &def_stmt, &dt)) |
b8698a0f | 476 | { |
73fbfcad | 477 | if (dump_enabled_p ()) |
78c60e3d | 478 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 479 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
480 | return false; |
481 | } | |
482 | ||
483 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
484 | return true; | |
485 | ||
486 | def_bb = gimple_bb (def_stmt); | |
487 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
488 | { | |
73fbfcad | 489 | if (dump_enabled_p ()) |
e645e942 | 490 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
491 | return true; |
492 | } | |
493 | ||
b8698a0f L |
494 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
495 | DEF_STMT must have already been processed, because this should be the | |
496 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
497 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
498 | check that everything is as expected, and we are done. */ |
499 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
500 | bb = gimple_bb (stmt); | |
501 | if (gimple_code (stmt) == GIMPLE_PHI | |
502 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
503 | && gimple_code (def_stmt) != GIMPLE_PHI | |
504 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
505 | && bb->loop_father == def_bb->loop_father) | |
506 | { | |
73fbfcad | 507 | if (dump_enabled_p ()) |
78c60e3d | 508 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 509 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
510 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
511 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
512 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 513 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 514 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
515 | return true; |
516 | } | |
517 | ||
518 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
519 | outer-loop-header-bb: | |
520 | d = def_stmt | |
521 | inner-loop: | |
522 | stmt # use (d) | |
523 | outer-loop-tail-bb: | |
524 | ... */ | |
525 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
526 | { | |
73fbfcad | 527 | if (dump_enabled_p ()) |
78c60e3d | 528 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 529 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 530 | |
ebfd146a IR |
531 | switch (relevant) |
532 | { | |
8644a673 | 533 | case vect_unused_in_scope: |
7c5222ff IR |
534 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
535 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 536 | break; |
7c5222ff | 537 | |
ebfd146a | 538 | case vect_used_in_outer_by_reduction: |
7c5222ff | 539 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
540 | relevant = vect_used_by_reduction; |
541 | break; | |
7c5222ff | 542 | |
ebfd146a | 543 | case vect_used_in_outer: |
7c5222ff | 544 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 545 | relevant = vect_used_in_scope; |
ebfd146a | 546 | break; |
7c5222ff | 547 | |
8644a673 | 548 | case vect_used_in_scope: |
ebfd146a IR |
549 | break; |
550 | ||
551 | default: | |
552 | gcc_unreachable (); | |
b8698a0f | 553 | } |
ebfd146a IR |
554 | } |
555 | ||
556 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
557 | outer-loop-header-bb: | |
558 | ... | |
559 | inner-loop: | |
560 | d = def_stmt | |
06066f92 | 561 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
562 | stmt # use (d) */ |
563 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
564 | { | |
73fbfcad | 565 | if (dump_enabled_p ()) |
78c60e3d | 566 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 567 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 568 | |
ebfd146a IR |
569 | switch (relevant) |
570 | { | |
8644a673 | 571 | case vect_unused_in_scope: |
b8698a0f | 572 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 573 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 574 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
575 | break; |
576 | ||
ebfd146a | 577 | case vect_used_by_reduction: |
b28ead45 | 578 | case vect_used_only_live: |
ebfd146a IR |
579 | relevant = vect_used_in_outer_by_reduction; |
580 | break; | |
581 | ||
8644a673 | 582 | case vect_used_in_scope: |
ebfd146a IR |
583 | relevant = vect_used_in_outer; |
584 | break; | |
585 | ||
586 | default: | |
587 | gcc_unreachable (); | |
588 | } | |
589 | } | |
643a9684 RB |
590 | /* We are also not interested in uses on loop PHI backedges that are |
591 | inductions. Otherwise we'll needlessly vectorize the IV increment | |
e294f495 RB |
592 | and cause hybrid SLP for SLP inductions. Unless the PHI is live |
593 | of course. */ | |
643a9684 RB |
594 | else if (gimple_code (stmt) == GIMPLE_PHI |
595 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_induction_def | |
e294f495 | 596 | && ! STMT_VINFO_LIVE_P (stmt_vinfo) |
643a9684 RB |
597 | && (PHI_ARG_DEF_FROM_EDGE (stmt, loop_latch_edge (bb->loop_father)) |
598 | == use)) | |
599 | { | |
600 | if (dump_enabled_p ()) | |
601 | dump_printf_loc (MSG_NOTE, vect_location, | |
602 | "induction value on backedge.\n"); | |
603 | return true; | |
604 | } | |
605 | ||
ebfd146a | 606 | |
b28ead45 | 607 | vect_mark_relevant (worklist, def_stmt, relevant, false); |
ebfd146a IR |
608 | return true; |
609 | } | |
610 | ||
611 | ||
612 | /* Function vect_mark_stmts_to_be_vectorized. | |
613 | ||
614 | Not all stmts in the loop need to be vectorized. For example: | |
615 | ||
616 | for i... | |
617 | for j... | |
618 | 1. T0 = i + j | |
619 | 2. T1 = a[T0] | |
620 | ||
621 | 3. j = j + 1 | |
622 | ||
623 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
624 | addressing of vectorized data-refs are handled differently. | |
625 | ||
626 | This pass detects such stmts. */ | |
627 | ||
628 | bool | |
629 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
630 | { | |
ebfd146a IR |
631 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
632 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
633 | unsigned int nbbs = loop->num_nodes; | |
634 | gimple_stmt_iterator si; | |
355fe088 | 635 | gimple *stmt; |
ebfd146a IR |
636 | unsigned int i; |
637 | stmt_vec_info stmt_vinfo; | |
638 | basic_block bb; | |
355fe088 | 639 | gimple *phi; |
ebfd146a | 640 | bool live_p; |
b28ead45 | 641 | enum vect_relevant relevant; |
ebfd146a | 642 | |
73fbfcad | 643 | if (dump_enabled_p ()) |
78c60e3d | 644 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 645 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 646 | |
355fe088 | 647 | auto_vec<gimple *, 64> worklist; |
ebfd146a IR |
648 | |
649 | /* 1. Init worklist. */ | |
650 | for (i = 0; i < nbbs; i++) | |
651 | { | |
652 | bb = bbs[i]; | |
653 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 654 | { |
ebfd146a | 655 | phi = gsi_stmt (si); |
73fbfcad | 656 | if (dump_enabled_p ()) |
ebfd146a | 657 | { |
78c60e3d SS |
658 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
659 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
660 | } |
661 | ||
662 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 663 | vect_mark_relevant (&worklist, phi, relevant, live_p); |
ebfd146a IR |
664 | } |
665 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
666 | { | |
667 | stmt = gsi_stmt (si); | |
73fbfcad | 668 | if (dump_enabled_p ()) |
ebfd146a | 669 | { |
78c60e3d SS |
670 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
671 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 672 | } |
ebfd146a IR |
673 | |
674 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 675 | vect_mark_relevant (&worklist, stmt, relevant, live_p); |
ebfd146a IR |
676 | } |
677 | } | |
678 | ||
679 | /* 2. Process_worklist */ | |
9771b263 | 680 | while (worklist.length () > 0) |
ebfd146a IR |
681 | { |
682 | use_operand_p use_p; | |
683 | ssa_op_iter iter; | |
684 | ||
9771b263 | 685 | stmt = worklist.pop (); |
73fbfcad | 686 | if (dump_enabled_p ()) |
ebfd146a | 687 | { |
78c60e3d SS |
688 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
689 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
690 | } |
691 | ||
b8698a0f | 692 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
b28ead45 AH |
693 | (DEF_STMT) as relevant/irrelevant according to the relevance property |
694 | of STMT. */ | |
ebfd146a IR |
695 | stmt_vinfo = vinfo_for_stmt (stmt); |
696 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
ebfd146a | 697 | |
b28ead45 AH |
698 | /* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is |
699 | propagated as is to the DEF_STMTs of its USEs. | |
ebfd146a IR |
700 | |
701 | One exception is when STMT has been identified as defining a reduction | |
b28ead45 | 702 | variable; in this case we set the relevance to vect_used_by_reduction. |
ebfd146a | 703 | This is because we distinguish between two kinds of relevant stmts - |
b8698a0f | 704 | those that are used by a reduction computation, and those that are |
ff802fa1 | 705 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 706 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 707 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 708 | |
b28ead45 | 709 | switch (STMT_VINFO_DEF_TYPE (stmt_vinfo)) |
ebfd146a | 710 | { |
06066f92 | 711 | case vect_reduction_def: |
b28ead45 AH |
712 | gcc_assert (relevant != vect_unused_in_scope); |
713 | if (relevant != vect_unused_in_scope | |
714 | && relevant != vect_used_in_scope | |
715 | && relevant != vect_used_by_reduction | |
716 | && relevant != vect_used_only_live) | |
06066f92 | 717 | { |
b28ead45 AH |
718 | if (dump_enabled_p ()) |
719 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
720 | "unsupported use of reduction.\n"); | |
721 | return false; | |
06066f92 | 722 | } |
06066f92 | 723 | break; |
b8698a0f | 724 | |
06066f92 | 725 | case vect_nested_cycle: |
b28ead45 AH |
726 | if (relevant != vect_unused_in_scope |
727 | && relevant != vect_used_in_outer_by_reduction | |
728 | && relevant != vect_used_in_outer) | |
06066f92 | 729 | { |
73fbfcad | 730 | if (dump_enabled_p ()) |
78c60e3d | 731 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 732 | "unsupported use of nested cycle.\n"); |
7c5222ff | 733 | |
06066f92 IR |
734 | return false; |
735 | } | |
b8698a0f L |
736 | break; |
737 | ||
06066f92 | 738 | case vect_double_reduction_def: |
b28ead45 AH |
739 | if (relevant != vect_unused_in_scope |
740 | && relevant != vect_used_by_reduction | |
741 | && relevant != vect_used_only_live) | |
06066f92 | 742 | { |
73fbfcad | 743 | if (dump_enabled_p ()) |
78c60e3d | 744 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 745 | "unsupported use of double reduction.\n"); |
7c5222ff | 746 | |
7c5222ff | 747 | return false; |
06066f92 | 748 | } |
b8698a0f | 749 | break; |
7c5222ff | 750 | |
06066f92 IR |
751 | default: |
752 | break; | |
7c5222ff | 753 | } |
b8698a0f | 754 | |
aec7ae7d | 755 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
756 | { |
757 | /* Pattern statements are not inserted into the code, so | |
758 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
759 | have to scan the RHS or function arguments instead. */ | |
760 | if (is_gimple_assign (stmt)) | |
761 | { | |
69d2aade JJ |
762 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
763 | tree op = gimple_assign_rhs1 (stmt); | |
764 | ||
765 | i = 1; | |
766 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
767 | { | |
768 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
b28ead45 | 769 | relevant, &worklist, false) |
69d2aade | 770 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
b28ead45 | 771 | relevant, &worklist, false)) |
566d377a | 772 | return false; |
69d2aade JJ |
773 | i = 2; |
774 | } | |
775 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 776 | { |
69d2aade | 777 | op = gimple_op (stmt, i); |
afbe6325 | 778 | if (TREE_CODE (op) == SSA_NAME |
b28ead45 | 779 | && !process_use (stmt, op, loop_vinfo, relevant, |
afbe6325 | 780 | &worklist, false)) |
07687835 | 781 | return false; |
9d5e7640 IR |
782 | } |
783 | } | |
784 | else if (is_gimple_call (stmt)) | |
785 | { | |
786 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
787 | { | |
788 | tree arg = gimple_call_arg (stmt, i); | |
b28ead45 | 789 | if (!process_use (stmt, arg, loop_vinfo, relevant, |
aec7ae7d | 790 | &worklist, false)) |
07687835 | 791 | return false; |
9d5e7640 IR |
792 | } |
793 | } | |
794 | } | |
795 | else | |
796 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
797 | { | |
798 | tree op = USE_FROM_PTR (use_p); | |
b28ead45 | 799 | if (!process_use (stmt, op, loop_vinfo, relevant, |
aec7ae7d | 800 | &worklist, false)) |
07687835 | 801 | return false; |
9d5e7640 | 802 | } |
aec7ae7d | 803 | |
3bab6342 | 804 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d | 805 | { |
134c85ca RS |
806 | gather_scatter_info gs_info; |
807 | if (!vect_check_gather_scatter (stmt, loop_vinfo, &gs_info)) | |
808 | gcc_unreachable (); | |
809 | if (!process_use (stmt, gs_info.offset, loop_vinfo, relevant, | |
810 | &worklist, true)) | |
566d377a | 811 | return false; |
aec7ae7d | 812 | } |
ebfd146a IR |
813 | } /* while worklist */ |
814 | ||
ebfd146a IR |
815 | return true; |
816 | } | |
817 | ||
818 | ||
b8698a0f | 819 | /* Function vect_model_simple_cost. |
ebfd146a | 820 | |
b8698a0f | 821 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
822 | single op. Right now, this does not account for multiple insns that could |
823 | be generated for the single vector op. We will handle that shortly. */ | |
824 | ||
825 | void | |
b8698a0f | 826 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 | 827 | enum vect_def_type *dt, |
4fc5ebf1 | 828 | int ndts, |
92345349 BS |
829 | stmt_vector_for_cost *prologue_cost_vec, |
830 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
831 | { |
832 | int i; | |
92345349 | 833 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
834 | |
835 | /* The SLP costs were already calculated during SLP tree build. */ | |
836 | if (PURE_SLP_STMT (stmt_info)) | |
837 | return; | |
838 | ||
4fc5ebf1 JG |
839 | /* Cost the "broadcast" of a scalar operand in to a vector operand. |
840 | Use scalar_to_vec to cost the broadcast, as elsewhere in the vector | |
841 | cost model. */ | |
842 | for (i = 0; i < ndts; i++) | |
92345349 | 843 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
4fc5ebf1 | 844 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
92345349 | 845 | stmt_info, 0, vect_prologue); |
c3e7ee41 BS |
846 | |
847 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
848 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
849 | stmt_info, 0, vect_body); | |
c3e7ee41 | 850 | |
73fbfcad | 851 | if (dump_enabled_p ()) |
78c60e3d SS |
852 | dump_printf_loc (MSG_NOTE, vect_location, |
853 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 854 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
855 | } |
856 | ||
857 | ||
8bd37302 BS |
858 | /* Model cost for type demotion and promotion operations. PWR is normally |
859 | zero for single-step promotions and demotions. It will be one if | |
860 | two-step promotion/demotion is required, and so on. Each additional | |
861 | step doubles the number of instructions required. */ | |
862 | ||
863 | static void | |
864 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
865 | enum vect_def_type *dt, int pwr) | |
866 | { | |
867 | int i, tmp; | |
92345349 | 868 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
869 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
870 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
871 | void *target_cost_data; | |
8bd37302 BS |
872 | |
873 | /* The SLP costs were already calculated during SLP tree build. */ | |
874 | if (PURE_SLP_STMT (stmt_info)) | |
875 | return; | |
876 | ||
c3e7ee41 BS |
877 | if (loop_vinfo) |
878 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
879 | else | |
880 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
881 | ||
8bd37302 BS |
882 | for (i = 0; i < pwr + 1; i++) |
883 | { | |
884 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
885 | (i + 1) : i; | |
c3e7ee41 | 886 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
887 | vec_promote_demote, stmt_info, 0, |
888 | vect_body); | |
8bd37302 BS |
889 | } |
890 | ||
891 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
892 | for (i = 0; i < 2; i++) | |
92345349 BS |
893 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
894 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
895 | stmt_info, 0, vect_prologue); | |
8bd37302 | 896 | |
73fbfcad | 897 | if (dump_enabled_p ()) |
78c60e3d SS |
898 | dump_printf_loc (MSG_NOTE, vect_location, |
899 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 900 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
901 | } |
902 | ||
ebfd146a IR |
903 | /* Function vect_model_store_cost |
904 | ||
0d0293ac MM |
905 | Models cost for stores. In the case of grouped accesses, one access |
906 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
907 | |
908 | void | |
b8698a0f | 909 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
2de001ee RS |
910 | vect_memory_access_type memory_access_type, |
911 | enum vect_def_type dt, slp_tree slp_node, | |
92345349 BS |
912 | stmt_vector_for_cost *prologue_cost_vec, |
913 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a | 914 | { |
92345349 | 915 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f RS |
916 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
917 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); | |
918 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); | |
ebfd146a | 919 | |
8644a673 | 920 | if (dt == vect_constant_def || dt == vect_external_def) |
92345349 BS |
921 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
922 | stmt_info, 0, vect_prologue); | |
ebfd146a | 923 | |
892a981f RS |
924 | /* Grouped stores update all elements in the group at once, |
925 | so we want the DR for the first statement. */ | |
926 | if (!slp_node && grouped_access_p) | |
720f5239 | 927 | { |
892a981f RS |
928 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
929 | dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
720f5239 | 930 | } |
ebfd146a | 931 | |
892a981f RS |
932 | /* True if we should include any once-per-group costs as well as |
933 | the cost of the statement itself. For SLP we only get called | |
934 | once per group anyhow. */ | |
935 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
936 | ||
272c6793 | 937 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 938 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 | 939 | access is instead being provided by a permute-and-store operation, |
2de001ee RS |
940 | include the cost of the permutes. */ |
941 | if (first_stmt_p | |
942 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 943 | { |
e1377713 ES |
944 | /* Uses a high and low interleave or shuffle operations for each |
945 | needed permute. */ | |
892a981f | 946 | int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
e1377713 | 947 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
92345349 BS |
948 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
949 | stmt_info, 0, vect_body); | |
ebfd146a | 950 | |
73fbfcad | 951 | if (dump_enabled_p ()) |
78c60e3d | 952 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 953 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 954 | group_size); |
ebfd146a IR |
955 | } |
956 | ||
cee62fee | 957 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 958 | /* Costs of the stores. */ |
067bc855 RB |
959 | if (memory_access_type == VMAT_ELEMENTWISE |
960 | || memory_access_type == VMAT_GATHER_SCATTER) | |
2de001ee RS |
961 | /* N scalar stores plus extracting the elements. */ |
962 | inside_cost += record_stmt_cost (body_cost_vec, | |
963 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
964 | scalar_store, stmt_info, 0, vect_body); | |
f2e2a985 | 965 | else |
892a981f | 966 | vect_get_store_cost (dr, ncopies, &inside_cost, body_cost_vec); |
ebfd146a | 967 | |
2de001ee RS |
968 | if (memory_access_type == VMAT_ELEMENTWISE |
969 | || memory_access_type == VMAT_STRIDED_SLP) | |
cee62fee MM |
970 | inside_cost += record_stmt_cost (body_cost_vec, |
971 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
972 | vec_to_scalar, stmt_info, 0, vect_body); | |
973 | ||
73fbfcad | 974 | if (dump_enabled_p ()) |
78c60e3d SS |
975 | dump_printf_loc (MSG_NOTE, vect_location, |
976 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 977 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
978 | } |
979 | ||
980 | ||
720f5239 IR |
981 | /* Calculate cost of DR's memory access. */ |
982 | void | |
983 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 984 | unsigned int *inside_cost, |
92345349 | 985 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
986 | { |
987 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 988 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 989 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
990 | |
991 | switch (alignment_support_scheme) | |
992 | { | |
993 | case dr_aligned: | |
994 | { | |
92345349 BS |
995 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
996 | vector_store, stmt_info, 0, | |
997 | vect_body); | |
720f5239 | 998 | |
73fbfcad | 999 | if (dump_enabled_p ()) |
78c60e3d | 1000 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1001 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
1002 | break; |
1003 | } | |
1004 | ||
1005 | case dr_unaligned_supported: | |
1006 | { | |
720f5239 | 1007 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1008 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1009 | unaligned_store, stmt_info, |
92345349 | 1010 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1011 | if (dump_enabled_p ()) |
78c60e3d SS |
1012 | dump_printf_loc (MSG_NOTE, vect_location, |
1013 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1014 | "hardware.\n"); |
720f5239 IR |
1015 | break; |
1016 | } | |
1017 | ||
38eec4c6 UW |
1018 | case dr_unaligned_unsupported: |
1019 | { | |
1020 | *inside_cost = VECT_MAX_COST; | |
1021 | ||
73fbfcad | 1022 | if (dump_enabled_p ()) |
78c60e3d | 1023 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1024 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1025 | break; |
1026 | } | |
1027 | ||
720f5239 IR |
1028 | default: |
1029 | gcc_unreachable (); | |
1030 | } | |
1031 | } | |
1032 | ||
1033 | ||
ebfd146a IR |
1034 | /* Function vect_model_load_cost |
1035 | ||
892a981f RS |
1036 | Models cost for loads. In the case of grouped accesses, one access has |
1037 | the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1038 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1039 | access scheme chosen. */ |
1040 | ||
1041 | void | |
92345349 | 1042 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
2de001ee RS |
1043 | vect_memory_access_type memory_access_type, |
1044 | slp_tree slp_node, | |
92345349 BS |
1045 | stmt_vector_for_cost *prologue_cost_vec, |
1046 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a | 1047 | { |
892a981f RS |
1048 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); |
1049 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
92345349 | 1050 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f | 1051 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
ebfd146a | 1052 | |
892a981f RS |
1053 | /* Grouped loads read all elements in the group at once, |
1054 | so we want the DR for the first statement. */ | |
1055 | if (!slp_node && grouped_access_p) | |
ebfd146a | 1056 | { |
892a981f RS |
1057 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
1058 | dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
ebfd146a IR |
1059 | } |
1060 | ||
892a981f RS |
1061 | /* True if we should include any once-per-group costs as well as |
1062 | the cost of the statement itself. For SLP we only get called | |
1063 | once per group anyhow. */ | |
1064 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
1065 | ||
272c6793 | 1066 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1067 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 | 1068 | access is instead being provided by a load-and-permute operation, |
2de001ee RS |
1069 | include the cost of the permutes. */ |
1070 | if (first_stmt_p | |
1071 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 1072 | { |
2c23db6d ES |
1073 | /* Uses an even and odd extract operations or shuffle operations |
1074 | for each needed permute. */ | |
892a981f | 1075 | int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
2c23db6d ES |
1076 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
1077 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, | |
1078 | stmt_info, 0, vect_body); | |
ebfd146a | 1079 | |
73fbfcad | 1080 | if (dump_enabled_p ()) |
e645e942 TJ |
1081 | dump_printf_loc (MSG_NOTE, vect_location, |
1082 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1083 | group_size); |
ebfd146a IR |
1084 | } |
1085 | ||
1086 | /* The loads themselves. */ | |
067bc855 RB |
1087 | if (memory_access_type == VMAT_ELEMENTWISE |
1088 | || memory_access_type == VMAT_GATHER_SCATTER) | |
a82960aa | 1089 | { |
a21892ad BS |
1090 | /* N scalar loads plus gathering them into a vector. */ |
1091 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
92345349 | 1092 | inside_cost += record_stmt_cost (body_cost_vec, |
c3e7ee41 | 1093 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), |
92345349 | 1094 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1095 | } |
1096 | else | |
892a981f | 1097 | vect_get_load_cost (dr, ncopies, first_stmt_p, |
92345349 BS |
1098 | &inside_cost, &prologue_cost, |
1099 | prologue_cost_vec, body_cost_vec, true); | |
2de001ee RS |
1100 | if (memory_access_type == VMAT_ELEMENTWISE |
1101 | || memory_access_type == VMAT_STRIDED_SLP) | |
892a981f RS |
1102 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, |
1103 | stmt_info, 0, vect_body); | |
720f5239 | 1104 | |
73fbfcad | 1105 | if (dump_enabled_p ()) |
78c60e3d SS |
1106 | dump_printf_loc (MSG_NOTE, vect_location, |
1107 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1108 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1109 | } |
1110 | ||
1111 | ||
1112 | /* Calculate cost of DR's memory access. */ | |
1113 | void | |
1114 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1115 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1116 | unsigned int *prologue_cost, |
1117 | stmt_vector_for_cost *prologue_cost_vec, | |
1118 | stmt_vector_for_cost *body_cost_vec, | |
1119 | bool record_prologue_costs) | |
720f5239 IR |
1120 | { |
1121 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 1122 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 1123 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
1124 | |
1125 | switch (alignment_support_scheme) | |
ebfd146a IR |
1126 | { |
1127 | case dr_aligned: | |
1128 | { | |
92345349 BS |
1129 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1130 | stmt_info, 0, vect_body); | |
ebfd146a | 1131 | |
73fbfcad | 1132 | if (dump_enabled_p ()) |
78c60e3d | 1133 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1134 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1135 | |
1136 | break; | |
1137 | } | |
1138 | case dr_unaligned_supported: | |
1139 | { | |
720f5239 | 1140 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1141 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1142 | unaligned_load, stmt_info, |
92345349 | 1143 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1144 | |
73fbfcad | 1145 | if (dump_enabled_p ()) |
78c60e3d SS |
1146 | dump_printf_loc (MSG_NOTE, vect_location, |
1147 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1148 | "hardware.\n"); |
ebfd146a IR |
1149 | |
1150 | break; | |
1151 | } | |
1152 | case dr_explicit_realign: | |
1153 | { | |
92345349 BS |
1154 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1155 | vector_load, stmt_info, 0, vect_body); | |
1156 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1157 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1158 | |
1159 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1160 | the containing loop, the following cost should be added to the | |
92345349 | 1161 | prologue costs. */ |
ebfd146a | 1162 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1163 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1164 | stmt_info, 0, vect_body); | |
ebfd146a | 1165 | |
73fbfcad | 1166 | if (dump_enabled_p ()) |
e645e942 TJ |
1167 | dump_printf_loc (MSG_NOTE, vect_location, |
1168 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1169 | |
ebfd146a IR |
1170 | break; |
1171 | } | |
1172 | case dr_explicit_realign_optimized: | |
1173 | { | |
73fbfcad | 1174 | if (dump_enabled_p ()) |
e645e942 | 1175 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1176 | "vect_model_load_cost: unaligned software " |
e645e942 | 1177 | "pipelined.\n"); |
ebfd146a IR |
1178 | |
1179 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1180 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1181 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1182 | access, then the above cost should only be considered for one |
ff802fa1 | 1183 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1184 | and a realignment op. */ |
1185 | ||
92345349 | 1186 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1187 | { |
92345349 BS |
1188 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1189 | vector_stmt, stmt_info, | |
1190 | 0, vect_prologue); | |
ebfd146a | 1191 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1192 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1193 | vector_stmt, stmt_info, | |
1194 | 0, vect_prologue); | |
ebfd146a IR |
1195 | } |
1196 | ||
92345349 BS |
1197 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1198 | stmt_info, 0, vect_body); | |
1199 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1200 | stmt_info, 0, vect_body); | |
8bd37302 | 1201 | |
73fbfcad | 1202 | if (dump_enabled_p ()) |
78c60e3d | 1203 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1204 | "vect_model_load_cost: explicit realign optimized" |
1205 | "\n"); | |
8bd37302 | 1206 | |
ebfd146a IR |
1207 | break; |
1208 | } | |
1209 | ||
38eec4c6 UW |
1210 | case dr_unaligned_unsupported: |
1211 | { | |
1212 | *inside_cost = VECT_MAX_COST; | |
1213 | ||
73fbfcad | 1214 | if (dump_enabled_p ()) |
78c60e3d | 1215 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1216 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1217 | break; |
1218 | } | |
1219 | ||
ebfd146a IR |
1220 | default: |
1221 | gcc_unreachable (); | |
1222 | } | |
ebfd146a IR |
1223 | } |
1224 | ||
418b7df3 RG |
1225 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1226 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1227 | |
418b7df3 | 1228 | static void |
355fe088 | 1229 | vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi) |
ebfd146a | 1230 | { |
ebfd146a | 1231 | if (gsi) |
418b7df3 | 1232 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1233 | else |
1234 | { | |
418b7df3 | 1235 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1236 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1237 | |
a70d6342 IR |
1238 | if (loop_vinfo) |
1239 | { | |
1240 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1241 | basic_block new_bb; |
1242 | edge pe; | |
a70d6342 IR |
1243 | |
1244 | if (nested_in_vect_loop_p (loop, stmt)) | |
1245 | loop = loop->inner; | |
b8698a0f | 1246 | |
a70d6342 | 1247 | pe = loop_preheader_edge (loop); |
418b7df3 | 1248 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1249 | gcc_assert (!new_bb); |
1250 | } | |
1251 | else | |
1252 | { | |
1253 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1254 | basic_block bb; | |
1255 | gimple_stmt_iterator gsi_bb_start; | |
1256 | ||
1257 | gcc_assert (bb_vinfo); | |
1258 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1259 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1260 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1261 | } |
ebfd146a IR |
1262 | } |
1263 | ||
73fbfcad | 1264 | if (dump_enabled_p ()) |
ebfd146a | 1265 | { |
78c60e3d SS |
1266 | dump_printf_loc (MSG_NOTE, vect_location, |
1267 | "created new init_stmt: "); | |
1268 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1269 | } |
418b7df3 RG |
1270 | } |
1271 | ||
1272 | /* Function vect_init_vector. | |
ebfd146a | 1273 | |
5467ee52 RG |
1274 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1275 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1276 | vector type a vector with all elements equal to VAL is created first. | |
1277 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1278 | initialization at the loop preheader. | |
418b7df3 RG |
1279 | Return the DEF of INIT_STMT. |
1280 | It will be used in the vectorization of STMT. */ | |
1281 | ||
1282 | tree | |
355fe088 | 1283 | vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 | 1284 | { |
355fe088 | 1285 | gimple *init_stmt; |
418b7df3 RG |
1286 | tree new_temp; |
1287 | ||
e412ece4 RB |
1288 | /* We abuse this function to push sth to a SSA name with initial 'val'. */ |
1289 | if (! useless_type_conversion_p (type, TREE_TYPE (val))) | |
418b7df3 | 1290 | { |
e412ece4 RB |
1291 | gcc_assert (TREE_CODE (type) == VECTOR_TYPE); |
1292 | if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) | |
418b7df3 | 1293 | { |
5a308cf1 IE |
1294 | /* Scalar boolean value should be transformed into |
1295 | all zeros or all ones value before building a vector. */ | |
1296 | if (VECTOR_BOOLEAN_TYPE_P (type)) | |
1297 | { | |
b3d51f23 IE |
1298 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1299 | tree false_val = build_zero_cst (TREE_TYPE (type)); | |
5a308cf1 IE |
1300 | |
1301 | if (CONSTANT_CLASS_P (val)) | |
1302 | val = integer_zerop (val) ? false_val : true_val; | |
1303 | else | |
1304 | { | |
1305 | new_temp = make_ssa_name (TREE_TYPE (type)); | |
1306 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, | |
1307 | val, true_val, false_val); | |
1308 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
1309 | val = new_temp; | |
1310 | } | |
1311 | } | |
1312 | else if (CONSTANT_CLASS_P (val)) | |
42fd8198 | 1313 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1314 | else |
1315 | { | |
b731b390 | 1316 | new_temp = make_ssa_name (TREE_TYPE (type)); |
e412ece4 RB |
1317 | if (! INTEGRAL_TYPE_P (TREE_TYPE (val))) |
1318 | init_stmt = gimple_build_assign (new_temp, | |
1319 | fold_build1 (VIEW_CONVERT_EXPR, | |
1320 | TREE_TYPE (type), | |
1321 | val)); | |
1322 | else | |
1323 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); | |
418b7df3 | 1324 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1325 | val = new_temp; |
418b7df3 RG |
1326 | } |
1327 | } | |
5467ee52 | 1328 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1329 | } |
1330 | ||
0e22bb5a RB |
1331 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1332 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1333 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1334 | return new_temp; |
ebfd146a IR |
1335 | } |
1336 | ||
c83a894c | 1337 | /* Function vect_get_vec_def_for_operand_1. |
a70d6342 | 1338 | |
c83a894c AH |
1339 | For a defining stmt DEF_STMT of a scalar stmt, return a vector def with type |
1340 | DT that will be used in the vectorized stmt. */ | |
ebfd146a IR |
1341 | |
1342 | tree | |
c83a894c | 1343 | vect_get_vec_def_for_operand_1 (gimple *def_stmt, enum vect_def_type dt) |
ebfd146a IR |
1344 | { |
1345 | tree vec_oprnd; | |
355fe088 | 1346 | gimple *vec_stmt; |
ebfd146a | 1347 | stmt_vec_info def_stmt_info = NULL; |
ebfd146a IR |
1348 | |
1349 | switch (dt) | |
1350 | { | |
81c40241 | 1351 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1352 | case vect_constant_def: |
81c40241 | 1353 | case vect_external_def: |
c83a894c AH |
1354 | /* Code should use vect_get_vec_def_for_operand. */ |
1355 | gcc_unreachable (); | |
ebfd146a | 1356 | |
81c40241 | 1357 | /* operand is defined inside the loop. */ |
8644a673 | 1358 | case vect_internal_def: |
ebfd146a | 1359 | { |
ebfd146a IR |
1360 | /* Get the def from the vectorized stmt. */ |
1361 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1362 | |
ebfd146a | 1363 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1364 | /* Get vectorized pattern statement. */ |
1365 | if (!vec_stmt | |
1366 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1367 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1368 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1369 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1370 | gcc_assert (vec_stmt); |
1371 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1372 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1373 | else if (is_gimple_call (vec_stmt)) | |
1374 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1375 | else | |
1376 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1377 | return vec_oprnd; | |
1378 | } | |
1379 | ||
c78e3652 | 1380 | /* operand is defined by a loop header phi. */ |
ebfd146a | 1381 | case vect_reduction_def: |
06066f92 | 1382 | case vect_double_reduction_def: |
7c5222ff | 1383 | case vect_nested_cycle: |
ebfd146a IR |
1384 | case vect_induction_def: |
1385 | { | |
1386 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1387 | ||
1388 | /* Get the def from the vectorized stmt. */ | |
1389 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1390 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1391 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1392 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1393 | else | |
1394 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1395 | return vec_oprnd; |
1396 | } | |
1397 | ||
1398 | default: | |
1399 | gcc_unreachable (); | |
1400 | } | |
1401 | } | |
1402 | ||
1403 | ||
c83a894c AH |
1404 | /* Function vect_get_vec_def_for_operand. |
1405 | ||
1406 | OP is an operand in STMT. This function returns a (vector) def that will be | |
1407 | used in the vectorized stmt for STMT. | |
1408 | ||
1409 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1410 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1411 | ||
1412 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1413 | needs to be introduced. VECTYPE may be used to specify a required type for | |
1414 | vector invariant. */ | |
1415 | ||
1416 | tree | |
1417 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) | |
1418 | { | |
1419 | gimple *def_stmt; | |
1420 | enum vect_def_type dt; | |
1421 | bool is_simple_use; | |
1422 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
1423 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
1424 | ||
1425 | if (dump_enabled_p ()) | |
1426 | { | |
1427 | dump_printf_loc (MSG_NOTE, vect_location, | |
1428 | "vect_get_vec_def_for_operand: "); | |
1429 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
1430 | dump_printf (MSG_NOTE, "\n"); | |
1431 | } | |
1432 | ||
1433 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt); | |
1434 | gcc_assert (is_simple_use); | |
1435 | if (def_stmt && dump_enabled_p ()) | |
1436 | { | |
1437 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1438 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
1439 | } | |
1440 | ||
1441 | if (dt == vect_constant_def || dt == vect_external_def) | |
1442 | { | |
1443 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
1444 | tree vector_type; | |
1445 | ||
1446 | if (vectype) | |
1447 | vector_type = vectype; | |
2568d8a1 | 1448 | else if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op)) |
c83a894c AH |
1449 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) |
1450 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1451 | else | |
1452 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1453 | ||
1454 | gcc_assert (vector_type); | |
1455 | return vect_init_vector (stmt, op, vector_type, NULL); | |
1456 | } | |
1457 | else | |
1458 | return vect_get_vec_def_for_operand_1 (def_stmt, dt); | |
1459 | } | |
1460 | ||
1461 | ||
ebfd146a IR |
1462 | /* Function vect_get_vec_def_for_stmt_copy |
1463 | ||
ff802fa1 | 1464 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1465 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1466 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1467 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1468 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1469 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1470 | DT is the type of the vector def VEC_OPRND. |
1471 | ||
1472 | Context: | |
1473 | In case the vectorization factor (VF) is bigger than the number | |
1474 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1475 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1476 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1477 | smallest data-type determines the VF, and as a result, when vectorizing |
1478 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1479 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1480 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1481 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1482 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1483 | ||
1484 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1485 | |
ebfd146a IR |
1486 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1487 | VS1.1: vx.1 = memref1 VS1.2 | |
1488 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1489 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1490 | |
1491 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1492 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1493 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1494 | VSnew.3: vz3 = vx.3 + ... | |
1495 | ||
1496 | The vectorization of S1 is explained in vectorizable_load. | |
1497 | The vectorization of S2: | |
b8698a0f L |
1498 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1499 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1500 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1501 | returns the vector-def 'vx.0'. |
1502 | ||
b8698a0f L |
1503 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1504 | function is called to get the relevant vector-def for each operand. It is | |
1505 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1506 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1507 | ||
b8698a0f L |
1508 | For example, to obtain the vector-def 'vx.1' in order to create the |
1509 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1510 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1511 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1512 | and return its def ('vx.1'). | |
1513 | Overall, to create the above sequence this function will be called 3 times: | |
1514 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1515 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1516 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1517 | ||
1518 | tree | |
1519 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1520 | { | |
355fe088 | 1521 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1522 | stmt_vec_info def_stmt_info; |
1523 | ||
1524 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1525 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1526 | return vec_oprnd; |
1527 | ||
1528 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1529 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1530 | gcc_assert (def_stmt_info); | |
1531 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1532 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1533 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1534 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1535 | else | |
1536 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1537 | return vec_oprnd; | |
1538 | } | |
1539 | ||
1540 | ||
1541 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1542 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a | 1543 | |
c78e3652 | 1544 | void |
b8698a0f | 1545 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1546 | vec<tree> *vec_oprnds0, |
1547 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1548 | { |
9771b263 | 1549 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1550 | |
1551 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1552 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1553 | |
9771b263 | 1554 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1555 | { |
9771b263 | 1556 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1557 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1558 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1559 | } |
1560 | } | |
1561 | ||
1562 | ||
c78e3652 | 1563 | /* Get vectorized definitions for OP0 and OP1. */ |
ebfd146a | 1564 | |
c78e3652 | 1565 | void |
355fe088 | 1566 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1567 | vec<tree> *vec_oprnds0, |
1568 | vec<tree> *vec_oprnds1, | |
306b0c92 | 1569 | slp_tree slp_node) |
ebfd146a IR |
1570 | { |
1571 | if (slp_node) | |
d092494c IR |
1572 | { |
1573 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1574 | auto_vec<tree> ops (nops); |
1575 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1576 | |
9771b263 | 1577 | ops.quick_push (op0); |
d092494c | 1578 | if (op1) |
9771b263 | 1579 | ops.quick_push (op1); |
d092494c | 1580 | |
306b0c92 | 1581 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
d092494c | 1582 | |
37b5ec8f | 1583 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1584 | if (op1) |
37b5ec8f | 1585 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1586 | } |
ebfd146a IR |
1587 | else |
1588 | { | |
1589 | tree vec_oprnd; | |
1590 | ||
9771b263 | 1591 | vec_oprnds0->create (1); |
81c40241 | 1592 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1593 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1594 | |
1595 | if (op1) | |
1596 | { | |
9771b263 | 1597 | vec_oprnds1->create (1); |
81c40241 | 1598 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1599 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1600 | } |
1601 | } | |
1602 | } | |
1603 | ||
1604 | ||
1605 | /* Function vect_finish_stmt_generation. | |
1606 | ||
1607 | Insert a new stmt. */ | |
1608 | ||
1609 | void | |
355fe088 | 1610 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1611 | gimple_stmt_iterator *gsi) |
1612 | { | |
1613 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
310213d4 | 1614 | vec_info *vinfo = stmt_info->vinfo; |
ebfd146a IR |
1615 | |
1616 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1617 | ||
54e8e2c3 RG |
1618 | if (!gsi_end_p (*gsi) |
1619 | && gimple_has_mem_ops (vec_stmt)) | |
1620 | { | |
355fe088 | 1621 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1622 | tree vuse = gimple_vuse (at_stmt); |
1623 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1624 | { | |
1625 | tree vdef = gimple_vdef (at_stmt); | |
1626 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1627 | /* If we have an SSA vuse and insert a store, update virtual | |
1628 | SSA form to avoid triggering the renamer. Do so only | |
1629 | if we can easily see all uses - which is what almost always | |
1630 | happens with the way vectorized stmts are inserted. */ | |
1631 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1632 | && ((is_gimple_assign (vec_stmt) | |
1633 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1634 | || (is_gimple_call (vec_stmt) | |
1635 | && !(gimple_call_flags (vec_stmt) | |
1636 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1637 | { | |
1638 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1639 | gimple_set_vdef (vec_stmt, new_vdef); | |
1640 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1641 | } | |
1642 | } | |
1643 | } | |
ebfd146a IR |
1644 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1645 | ||
310213d4 | 1646 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, vinfo)); |
ebfd146a | 1647 | |
73fbfcad | 1648 | if (dump_enabled_p ()) |
ebfd146a | 1649 | { |
78c60e3d SS |
1650 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1651 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
ebfd146a IR |
1652 | } |
1653 | ||
ad885386 | 1654 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
8e91d222 JJ |
1655 | |
1656 | /* While EH edges will generally prevent vectorization, stmt might | |
1657 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1658 | that could throw are part of the same region. */ | |
1659 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1660 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1661 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
ebfd146a IR |
1662 | } |
1663 | ||
70439f0d RS |
1664 | /* We want to vectorize a call to combined function CFN with function |
1665 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN | |
1666 | as the types of all inputs. Check whether this is possible using | |
1667 | an internal function, returning its code if so or IFN_LAST if not. */ | |
ebfd146a | 1668 | |
70439f0d RS |
1669 | static internal_fn |
1670 | vectorizable_internal_function (combined_fn cfn, tree fndecl, | |
1671 | tree vectype_out, tree vectype_in) | |
ebfd146a | 1672 | { |
70439f0d RS |
1673 | internal_fn ifn; |
1674 | if (internal_fn_p (cfn)) | |
1675 | ifn = as_internal_fn (cfn); | |
1676 | else | |
1677 | ifn = associated_internal_fn (fndecl); | |
1678 | if (ifn != IFN_LAST && direct_internal_fn_p (ifn)) | |
1679 | { | |
1680 | const direct_internal_fn_info &info = direct_internal_fn (ifn); | |
1681 | if (info.vectorizable) | |
1682 | { | |
1683 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); | |
1684 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); | |
d95ab70a RS |
1685 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1686 | OPTIMIZE_FOR_SPEED)) | |
70439f0d RS |
1687 | return ifn; |
1688 | } | |
1689 | } | |
1690 | return IFN_LAST; | |
ebfd146a IR |
1691 | } |
1692 | ||
5ce9450f | 1693 | |
355fe088 | 1694 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1695 | gimple_stmt_iterator *); |
1696 | ||
62da9e14 RS |
1697 | /* STMT is a non-strided load or store, meaning that it accesses |
1698 | elements with a known constant step. Return -1 if that step | |
1699 | is negative, 0 if it is zero, and 1 if it is greater than zero. */ | |
1700 | ||
1701 | static int | |
1702 | compare_step_with_zero (gimple *stmt) | |
1703 | { | |
1704 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3f5e8a76 RS |
1705 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
1706 | return tree_int_cst_compare (vect_dr_behavior (dr)->step, | |
1707 | size_zero_node); | |
62da9e14 RS |
1708 | } |
1709 | ||
1710 | /* If the target supports a permute mask that reverses the elements in | |
1711 | a vector of type VECTYPE, return that mask, otherwise return null. */ | |
1712 | ||
1713 | static tree | |
1714 | perm_mask_for_reverse (tree vectype) | |
1715 | { | |
1716 | int i, nunits; | |
62da9e14 RS |
1717 | |
1718 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
62da9e14 | 1719 | |
d980067b RS |
1720 | /* The encoding has a single stepped pattern. */ |
1721 | vec_perm_builder sel (nunits, 1, 3); | |
1722 | for (i = 0; i < 3; ++i) | |
908a1a16 | 1723 | sel.quick_push (nunits - 1 - i); |
62da9e14 | 1724 | |
e3342de4 RS |
1725 | vec_perm_indices indices (sel, 1, nunits); |
1726 | if (!can_vec_perm_const_p (TYPE_MODE (vectype), indices)) | |
62da9e14 | 1727 | return NULL_TREE; |
e3342de4 | 1728 | return vect_gen_perm_mask_checked (vectype, indices); |
62da9e14 | 1729 | } |
5ce9450f | 1730 | |
2de001ee RS |
1731 | /* A subroutine of get_load_store_type, with a subset of the same |
1732 | arguments. Handle the case where STMT is part of a grouped load | |
1733 | or store. | |
1734 | ||
1735 | For stores, the statements in the group are all consecutive | |
1736 | and there is no gap at the end. For loads, the statements in the | |
1737 | group might not be consecutive; there can be gaps between statements | |
1738 | as well as at the end. */ | |
1739 | ||
1740 | static bool | |
1741 | get_group_load_store_type (gimple *stmt, tree vectype, bool slp, | |
1742 | vec_load_store_type vls_type, | |
1743 | vect_memory_access_type *memory_access_type) | |
1744 | { | |
1745 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1746 | vec_info *vinfo = stmt_info->vinfo; | |
1747 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1748 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; | |
1749 | gimple *first_stmt = GROUP_FIRST_ELEMENT (stmt_info); | |
f702e7d4 | 1750 | data_reference *first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
2de001ee RS |
1751 | unsigned int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
1752 | bool single_element_p = (stmt == first_stmt | |
1753 | && !GROUP_NEXT_ELEMENT (stmt_info)); | |
1754 | unsigned HOST_WIDE_INT gap = GROUP_GAP (vinfo_for_stmt (first_stmt)); | |
522fcdd7 | 1755 | unsigned nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
1756 | |
1757 | /* True if the vectorized statements would access beyond the last | |
1758 | statement in the group. */ | |
1759 | bool overrun_p = false; | |
1760 | ||
1761 | /* True if we can cope with such overrun by peeling for gaps, so that | |
1762 | there is at least one final scalar iteration after the vector loop. */ | |
1763 | bool can_overrun_p = (vls_type == VLS_LOAD && loop_vinfo && !loop->inner); | |
1764 | ||
1765 | /* There can only be a gap at the end of the group if the stride is | |
1766 | known at compile time. */ | |
1767 | gcc_assert (!STMT_VINFO_STRIDED_P (stmt_info) || gap == 0); | |
1768 | ||
1769 | /* Stores can't yet have gaps. */ | |
1770 | gcc_assert (slp || vls_type == VLS_LOAD || gap == 0); | |
1771 | ||
1772 | if (slp) | |
1773 | { | |
1774 | if (STMT_VINFO_STRIDED_P (stmt_info)) | |
1775 | { | |
1776 | /* Try to use consecutive accesses of GROUP_SIZE elements, | |
1777 | separated by the stride, until we have a complete vector. | |
1778 | Fall back to scalar accesses if that isn't possible. */ | |
1779 | if (nunits % group_size == 0) | |
1780 | *memory_access_type = VMAT_STRIDED_SLP; | |
1781 | else | |
1782 | *memory_access_type = VMAT_ELEMENTWISE; | |
1783 | } | |
1784 | else | |
1785 | { | |
1786 | overrun_p = loop_vinfo && gap != 0; | |
1787 | if (overrun_p && vls_type != VLS_LOAD) | |
1788 | { | |
1789 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1790 | "Grouped store with gaps requires" | |
1791 | " non-consecutive accesses\n"); | |
1792 | return false; | |
1793 | } | |
f702e7d4 RS |
1794 | /* An overrun is fine if the trailing elements are smaller |
1795 | than the alignment boundary B. Every vector access will | |
1796 | be a multiple of B and so we are guaranteed to access a | |
1797 | non-gap element in the same B-sized block. */ | |
f9ef2c76 | 1798 | if (overrun_p |
f702e7d4 RS |
1799 | && gap < (vect_known_alignment_in_bytes (first_dr) |
1800 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 1801 | overrun_p = false; |
2de001ee RS |
1802 | if (overrun_p && !can_overrun_p) |
1803 | { | |
1804 | if (dump_enabled_p ()) | |
1805 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1806 | "Peeling for outer loop is not supported\n"); | |
1807 | return false; | |
1808 | } | |
1809 | *memory_access_type = VMAT_CONTIGUOUS; | |
1810 | } | |
1811 | } | |
1812 | else | |
1813 | { | |
1814 | /* We can always handle this case using elementwise accesses, | |
1815 | but see if something more efficient is available. */ | |
1816 | *memory_access_type = VMAT_ELEMENTWISE; | |
1817 | ||
1818 | /* If there is a gap at the end of the group then these optimizations | |
1819 | would access excess elements in the last iteration. */ | |
1820 | bool would_overrun_p = (gap != 0); | |
f702e7d4 RS |
1821 | /* An overrun is fine if the trailing elements are smaller than the |
1822 | alignment boundary B. Every vector access will be a multiple of B | |
1823 | and so we are guaranteed to access a non-gap element in the | |
1824 | same B-sized block. */ | |
f9ef2c76 | 1825 | if (would_overrun_p |
f702e7d4 RS |
1826 | && gap < (vect_known_alignment_in_bytes (first_dr) |
1827 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 1828 | would_overrun_p = false; |
f702e7d4 | 1829 | |
2de001ee | 1830 | if (!STMT_VINFO_STRIDED_P (stmt_info) |
62da9e14 RS |
1831 | && (can_overrun_p || !would_overrun_p) |
1832 | && compare_step_with_zero (stmt) > 0) | |
2de001ee RS |
1833 | { |
1834 | /* First try using LOAD/STORE_LANES. */ | |
1835 | if (vls_type == VLS_LOAD | |
1836 | ? vect_load_lanes_supported (vectype, group_size) | |
1837 | : vect_store_lanes_supported (vectype, group_size)) | |
1838 | { | |
1839 | *memory_access_type = VMAT_LOAD_STORE_LANES; | |
1840 | overrun_p = would_overrun_p; | |
1841 | } | |
1842 | ||
1843 | /* If that fails, try using permuting loads. */ | |
1844 | if (*memory_access_type == VMAT_ELEMENTWISE | |
1845 | && (vls_type == VLS_LOAD | |
1846 | ? vect_grouped_load_supported (vectype, single_element_p, | |
1847 | group_size) | |
1848 | : vect_grouped_store_supported (vectype, group_size))) | |
1849 | { | |
1850 | *memory_access_type = VMAT_CONTIGUOUS_PERMUTE; | |
1851 | overrun_p = would_overrun_p; | |
1852 | } | |
1853 | } | |
1854 | } | |
1855 | ||
1856 | if (vls_type != VLS_LOAD && first_stmt == stmt) | |
1857 | { | |
1858 | /* STMT is the leader of the group. Check the operands of all the | |
1859 | stmts of the group. */ | |
1860 | gimple *next_stmt = GROUP_NEXT_ELEMENT (stmt_info); | |
1861 | while (next_stmt) | |
1862 | { | |
1863 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
1864 | tree op = gimple_assign_rhs1 (next_stmt); | |
1865 | gimple *def_stmt; | |
1866 | enum vect_def_type dt; | |
1867 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) | |
1868 | { | |
1869 | if (dump_enabled_p ()) | |
1870 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1871 | "use not simple.\n"); | |
1872 | return false; | |
1873 | } | |
1874 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); | |
1875 | } | |
1876 | } | |
1877 | ||
1878 | if (overrun_p) | |
1879 | { | |
1880 | gcc_assert (can_overrun_p); | |
1881 | if (dump_enabled_p ()) | |
1882 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1883 | "Data access with gaps requires scalar " | |
1884 | "epilogue loop\n"); | |
1885 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; | |
1886 | } | |
1887 | ||
1888 | return true; | |
1889 | } | |
1890 | ||
62da9e14 RS |
1891 | /* A subroutine of get_load_store_type, with a subset of the same |
1892 | arguments. Handle the case where STMT is a load or store that | |
1893 | accesses consecutive elements with a negative step. */ | |
1894 | ||
1895 | static vect_memory_access_type | |
1896 | get_negative_load_store_type (gimple *stmt, tree vectype, | |
1897 | vec_load_store_type vls_type, | |
1898 | unsigned int ncopies) | |
1899 | { | |
1900 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1901 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
1902 | dr_alignment_support alignment_support_scheme; | |
1903 | ||
1904 | if (ncopies > 1) | |
1905 | { | |
1906 | if (dump_enabled_p ()) | |
1907 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1908 | "multiple types with negative step.\n"); | |
1909 | return VMAT_ELEMENTWISE; | |
1910 | } | |
1911 | ||
1912 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
1913 | if (alignment_support_scheme != dr_aligned | |
1914 | && alignment_support_scheme != dr_unaligned_supported) | |
1915 | { | |
1916 | if (dump_enabled_p ()) | |
1917 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1918 | "negative step but alignment required.\n"); | |
1919 | return VMAT_ELEMENTWISE; | |
1920 | } | |
1921 | ||
1922 | if (vls_type == VLS_STORE_INVARIANT) | |
1923 | { | |
1924 | if (dump_enabled_p ()) | |
1925 | dump_printf_loc (MSG_NOTE, vect_location, | |
1926 | "negative step with invariant source;" | |
1927 | " no permute needed.\n"); | |
1928 | return VMAT_CONTIGUOUS_DOWN; | |
1929 | } | |
1930 | ||
1931 | if (!perm_mask_for_reverse (vectype)) | |
1932 | { | |
1933 | if (dump_enabled_p ()) | |
1934 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1935 | "negative step and reversing not supported.\n"); | |
1936 | return VMAT_ELEMENTWISE; | |
1937 | } | |
1938 | ||
1939 | return VMAT_CONTIGUOUS_REVERSE; | |
1940 | } | |
1941 | ||
2de001ee RS |
1942 | /* Analyze load or store statement STMT of type VLS_TYPE. Return true |
1943 | if there is a memory access type that the vectorized form can use, | |
1944 | storing it in *MEMORY_ACCESS_TYPE if so. If we decide to use gathers | |
1945 | or scatters, fill in GS_INFO accordingly. | |
1946 | ||
1947 | SLP says whether we're performing SLP rather than loop vectorization. | |
62da9e14 RS |
1948 | VECTYPE is the vector type that the vectorized statements will use. |
1949 | NCOPIES is the number of vector statements that will be needed. */ | |
2de001ee RS |
1950 | |
1951 | static bool | |
1952 | get_load_store_type (gimple *stmt, tree vectype, bool slp, | |
62da9e14 | 1953 | vec_load_store_type vls_type, unsigned int ncopies, |
2de001ee RS |
1954 | vect_memory_access_type *memory_access_type, |
1955 | gather_scatter_info *gs_info) | |
1956 | { | |
1957 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1958 | vec_info *vinfo = stmt_info->vinfo; | |
1959 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1960 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
1961 | { | |
1962 | *memory_access_type = VMAT_GATHER_SCATTER; | |
1963 | gimple *def_stmt; | |
1964 | if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info)) | |
1965 | gcc_unreachable (); | |
1966 | else if (!vect_is_simple_use (gs_info->offset, vinfo, &def_stmt, | |
1967 | &gs_info->offset_dt, | |
1968 | &gs_info->offset_vectype)) | |
1969 | { | |
1970 | if (dump_enabled_p ()) | |
1971 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1972 | "%s index use not simple.\n", | |
1973 | vls_type == VLS_LOAD ? "gather" : "scatter"); | |
1974 | return false; | |
1975 | } | |
1976 | } | |
1977 | else if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1978 | { | |
1979 | if (!get_group_load_store_type (stmt, vectype, slp, vls_type, | |
1980 | memory_access_type)) | |
1981 | return false; | |
1982 | } | |
1983 | else if (STMT_VINFO_STRIDED_P (stmt_info)) | |
1984 | { | |
1985 | gcc_assert (!slp); | |
1986 | *memory_access_type = VMAT_ELEMENTWISE; | |
1987 | } | |
1988 | else | |
62da9e14 RS |
1989 | { |
1990 | int cmp = compare_step_with_zero (stmt); | |
1991 | if (cmp < 0) | |
1992 | *memory_access_type = get_negative_load_store_type | |
1993 | (stmt, vectype, vls_type, ncopies); | |
1994 | else if (cmp == 0) | |
1995 | { | |
1996 | gcc_assert (vls_type == VLS_LOAD); | |
1997 | *memory_access_type = VMAT_INVARIANT; | |
1998 | } | |
1999 | else | |
2000 | *memory_access_type = VMAT_CONTIGUOUS; | |
2001 | } | |
2de001ee RS |
2002 | |
2003 | /* FIXME: At the moment the cost model seems to underestimate the | |
2004 | cost of using elementwise accesses. This check preserves the | |
2005 | traditional behavior until that can be fixed. */ | |
2006 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2007 | && !STMT_VINFO_STRIDED_P (stmt_info)) | |
2008 | { | |
2009 | if (dump_enabled_p ()) | |
2010 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2011 | "not falling back to elementwise accesses\n"); | |
2012 | return false; | |
2013 | } | |
2014 | return true; | |
2015 | } | |
2016 | ||
5ce9450f JJ |
2017 | /* Function vectorizable_mask_load_store. |
2018 | ||
2019 | Check if STMT performs a conditional load or store that can be vectorized. | |
2020 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2021 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
2022 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2023 | ||
2024 | static bool | |
355fe088 TS |
2025 | vectorizable_mask_load_store (gimple *stmt, gimple_stmt_iterator *gsi, |
2026 | gimple **vec_stmt, slp_tree slp_node) | |
5ce9450f JJ |
2027 | { |
2028 | tree vec_dest = NULL; | |
2029 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2030 | stmt_vec_info prev_stmt_info; | |
2031 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2032 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2033 | bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
2034 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2035 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
57e2f6ad | 2036 | tree rhs_vectype = NULL_TREE; |
045c1278 | 2037 | tree mask_vectype; |
5ce9450f | 2038 | tree elem_type; |
355fe088 | 2039 | gimple *new_stmt; |
5ce9450f JJ |
2040 | tree dummy; |
2041 | tree dataref_ptr = NULL_TREE; | |
355fe088 | 2042 | gimple *ptr_incr; |
5ce9450f JJ |
2043 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2044 | int ncopies; | |
2045 | int i, j; | |
2046 | bool inv_p; | |
134c85ca | 2047 | gather_scatter_info gs_info; |
2de001ee | 2048 | vec_load_store_type vls_type; |
5ce9450f | 2049 | tree mask; |
355fe088 | 2050 | gimple *def_stmt; |
5ce9450f JJ |
2051 | enum vect_def_type dt; |
2052 | ||
2053 | if (slp_node != NULL) | |
2054 | return false; | |
2055 | ||
e8f142e2 | 2056 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
5ce9450f JJ |
2057 | gcc_assert (ncopies >= 1); |
2058 | ||
5ce9450f | 2059 | mask = gimple_call_arg (stmt, 2); |
045c1278 | 2060 | |
2568d8a1 | 2061 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (mask))) |
5ce9450f JJ |
2062 | return false; |
2063 | ||
2064 | /* FORNOW. This restriction should be relaxed. */ | |
2065 | if (nested_in_vect_loop && ncopies > 1) | |
2066 | { | |
2067 | if (dump_enabled_p ()) | |
2068 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2069 | "multiple types in nested loop."); | |
2070 | return false; | |
2071 | } | |
2072 | ||
2073 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
2074 | return false; | |
2075 | ||
66c16fd9 RB |
2076 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
2077 | && ! vec_stmt) | |
5ce9450f JJ |
2078 | return false; |
2079 | ||
2080 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
2081 | return false; | |
2082 | ||
2083 | elem_type = TREE_TYPE (vectype); | |
2084 | ||
045c1278 IE |
2085 | if (TREE_CODE (mask) != SSA_NAME) |
2086 | return false; | |
2087 | ||
2088 | if (!vect_is_simple_use (mask, loop_vinfo, &def_stmt, &dt, &mask_vectype)) | |
2089 | return false; | |
2090 | ||
2091 | if (!mask_vectype) | |
2092 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
2093 | ||
dc6a3147 IE |
2094 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype) |
2095 | || TYPE_VECTOR_SUBPARTS (mask_vectype) != TYPE_VECTOR_SUBPARTS (vectype)) | |
045c1278 IE |
2096 | return false; |
2097 | ||
2de001ee | 2098 | if (gimple_call_internal_fn (stmt) == IFN_MASK_STORE) |
57e2f6ad IE |
2099 | { |
2100 | tree rhs = gimple_call_arg (stmt, 3); | |
2101 | if (!vect_is_simple_use (rhs, loop_vinfo, &def_stmt, &dt, &rhs_vectype)) | |
2102 | return false; | |
2de001ee RS |
2103 | if (dt == vect_constant_def || dt == vect_external_def) |
2104 | vls_type = VLS_STORE_INVARIANT; | |
2105 | else | |
2106 | vls_type = VLS_STORE; | |
57e2f6ad | 2107 | } |
2de001ee RS |
2108 | else |
2109 | vls_type = VLS_LOAD; | |
57e2f6ad | 2110 | |
2de001ee | 2111 | vect_memory_access_type memory_access_type; |
62da9e14 | 2112 | if (!get_load_store_type (stmt, vectype, false, vls_type, ncopies, |
2de001ee RS |
2113 | &memory_access_type, &gs_info)) |
2114 | return false; | |
03b9e8e4 | 2115 | |
2de001ee RS |
2116 | if (memory_access_type == VMAT_GATHER_SCATTER) |
2117 | { | |
134c85ca | 2118 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
03b9e8e4 JJ |
2119 | tree masktype |
2120 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
2121 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
2122 | { | |
2123 | if (dump_enabled_p ()) | |
2124 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2125 | "masked gather with integer mask not supported."); | |
2126 | return false; | |
2127 | } | |
5ce9450f | 2128 | } |
2de001ee RS |
2129 | else if (memory_access_type != VMAT_CONTIGUOUS) |
2130 | { | |
2131 | if (dump_enabled_p ()) | |
2132 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2133 | "unsupported access type for masked %s.\n", | |
2134 | vls_type == VLS_LOAD ? "load" : "store"); | |
2135 | return false; | |
2136 | } | |
5ce9450f | 2137 | else if (!VECTOR_MODE_P (TYPE_MODE (vectype)) |
045c1278 IE |
2138 | || !can_vec_mask_load_store_p (TYPE_MODE (vectype), |
2139 | TYPE_MODE (mask_vectype), | |
2de001ee | 2140 | vls_type == VLS_LOAD) |
57e2f6ad IE |
2141 | || (rhs_vectype |
2142 | && !useless_type_conversion_p (vectype, rhs_vectype))) | |
5ce9450f JJ |
2143 | return false; |
2144 | ||
5ce9450f JJ |
2145 | if (!vec_stmt) /* transformation not required. */ |
2146 | { | |
2de001ee | 2147 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
5ce9450f | 2148 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; |
2de001ee RS |
2149 | if (vls_type == VLS_LOAD) |
2150 | vect_model_load_cost (stmt_info, ncopies, memory_access_type, | |
2151 | NULL, NULL, NULL); | |
5ce9450f | 2152 | else |
2de001ee RS |
2153 | vect_model_store_cost (stmt_info, ncopies, memory_access_type, |
2154 | dt, NULL, NULL, NULL); | |
5ce9450f JJ |
2155 | return true; |
2156 | } | |
2de001ee | 2157 | gcc_assert (memory_access_type == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
5ce9450f | 2158 | |
67b8dbac | 2159 | /* Transform. */ |
5ce9450f | 2160 | |
2de001ee | 2161 | if (memory_access_type == VMAT_GATHER_SCATTER) |
5ce9450f JJ |
2162 | { |
2163 | tree vec_oprnd0 = NULL_TREE, op; | |
134c85ca | 2164 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
5ce9450f | 2165 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; |
acdcd61b | 2166 | tree ptr, vec_mask = NULL_TREE, mask_op = NULL_TREE, var, scale; |
5ce9450f | 2167 | tree perm_mask = NULL_TREE, prev_res = NULL_TREE; |
acdcd61b | 2168 | tree mask_perm_mask = NULL_TREE; |
5ce9450f JJ |
2169 | edge pe = loop_preheader_edge (loop); |
2170 | gimple_seq seq; | |
2171 | basic_block new_bb; | |
2172 | enum { NARROW, NONE, WIDEN } modifier; | |
134c85ca | 2173 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype); |
5ce9450f | 2174 | |
134c85ca | 2175 | rettype = TREE_TYPE (TREE_TYPE (gs_info.decl)); |
acdcd61b JJ |
2176 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
2177 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2178 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2179 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2180 | scaletype = TREE_VALUE (arglist); | |
2181 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
2182 | && types_compatible_p (srctype, masktype)); | |
2183 | ||
5ce9450f JJ |
2184 | if (nunits == gather_off_nunits) |
2185 | modifier = NONE; | |
2186 | else if (nunits == gather_off_nunits / 2) | |
2187 | { | |
5ce9450f JJ |
2188 | modifier = WIDEN; |
2189 | ||
e3342de4 | 2190 | vec_perm_builder sel (gather_off_nunits, gather_off_nunits, 1); |
5ce9450f | 2191 | for (i = 0; i < gather_off_nunits; ++i) |
908a1a16 | 2192 | sel.quick_push (i | nunits); |
5ce9450f | 2193 | |
e3342de4 RS |
2194 | vec_perm_indices indices (sel, 1, gather_off_nunits); |
2195 | perm_mask = vect_gen_perm_mask_checked (gs_info.offset_vectype, | |
2196 | indices); | |
5ce9450f JJ |
2197 | } |
2198 | else if (nunits == gather_off_nunits * 2) | |
2199 | { | |
5ce9450f JJ |
2200 | modifier = NARROW; |
2201 | ||
e3342de4 | 2202 | vec_perm_builder sel (nunits, nunits, 1); |
908a1a16 | 2203 | sel.quick_grow (nunits); |
5ce9450f JJ |
2204 | for (i = 0; i < nunits; ++i) |
2205 | sel[i] = i < gather_off_nunits | |
2206 | ? i : i + nunits - gather_off_nunits; | |
e3342de4 RS |
2207 | vec_perm_indices indices (sel, 2, nunits); |
2208 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); | |
5ce9450f | 2209 | |
5ce9450f | 2210 | ncopies *= 2; |
e3342de4 | 2211 | |
acdcd61b JJ |
2212 | for (i = 0; i < nunits; ++i) |
2213 | sel[i] = i | gather_off_nunits; | |
e3342de4 RS |
2214 | indices.new_vector (sel, 2, gather_off_nunits); |
2215 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, indices); | |
5ce9450f JJ |
2216 | } |
2217 | else | |
2218 | gcc_unreachable (); | |
2219 | ||
5ce9450f JJ |
2220 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); |
2221 | ||
134c85ca | 2222 | ptr = fold_convert (ptrtype, gs_info.base); |
5ce9450f JJ |
2223 | if (!is_gimple_min_invariant (ptr)) |
2224 | { | |
2225 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
2226 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
2227 | gcc_assert (!new_bb); | |
2228 | } | |
2229 | ||
134c85ca | 2230 | scale = build_int_cst (scaletype, gs_info.scale); |
5ce9450f JJ |
2231 | |
2232 | prev_stmt_info = NULL; | |
2233 | for (j = 0; j < ncopies; ++j) | |
2234 | { | |
2235 | if (modifier == WIDEN && (j & 1)) | |
2236 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
2237 | perm_mask, stmt, gsi); | |
2238 | else if (j == 0) | |
2239 | op = vec_oprnd0 | |
134c85ca | 2240 | = vect_get_vec_def_for_operand (gs_info.offset, stmt); |
5ce9450f JJ |
2241 | else |
2242 | op = vec_oprnd0 | |
134c85ca | 2243 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, vec_oprnd0); |
5ce9450f JJ |
2244 | |
2245 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
2246 | { | |
2247 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
2248 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 2249 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
5ce9450f JJ |
2250 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
2251 | new_stmt | |
0d0e4a03 | 2252 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
2253 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2254 | op = var; | |
2255 | } | |
2256 | ||
acdcd61b JJ |
2257 | if (mask_perm_mask && (j & 1)) |
2258 | mask_op = permute_vec_elements (mask_op, mask_op, | |
2259 | mask_perm_mask, stmt, gsi); | |
5ce9450f JJ |
2260 | else |
2261 | { | |
acdcd61b | 2262 | if (j == 0) |
81c40241 | 2263 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
acdcd61b JJ |
2264 | else |
2265 | { | |
81c40241 | 2266 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
acdcd61b JJ |
2267 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2268 | } | |
5ce9450f | 2269 | |
acdcd61b JJ |
2270 | mask_op = vec_mask; |
2271 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
2272 | { | |
2273 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)) | |
2274 | == TYPE_VECTOR_SUBPARTS (masktype)); | |
0e22bb5a | 2275 | var = vect_get_new_ssa_name (masktype, vect_simple_var); |
acdcd61b JJ |
2276 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); |
2277 | new_stmt | |
0d0e4a03 | 2278 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_op); |
acdcd61b JJ |
2279 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2280 | mask_op = var; | |
2281 | } | |
5ce9450f JJ |
2282 | } |
2283 | ||
2284 | new_stmt | |
134c85ca | 2285 | = gimple_build_call (gs_info.decl, 5, mask_op, ptr, op, mask_op, |
5ce9450f JJ |
2286 | scale); |
2287 | ||
2288 | if (!useless_type_conversion_p (vectype, rettype)) | |
2289 | { | |
2290 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
2291 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 2292 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
5ce9450f JJ |
2293 | gimple_call_set_lhs (new_stmt, op); |
2294 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 2295 | var = make_ssa_name (vec_dest); |
5ce9450f | 2296 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
0d0e4a03 | 2297 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
2298 | } |
2299 | else | |
2300 | { | |
2301 | var = make_ssa_name (vec_dest, new_stmt); | |
2302 | gimple_call_set_lhs (new_stmt, var); | |
2303 | } | |
2304 | ||
2305 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2306 | ||
2307 | if (modifier == NARROW) | |
2308 | { | |
2309 | if ((j & 1) == 0) | |
2310 | { | |
2311 | prev_res = var; | |
2312 | continue; | |
2313 | } | |
2314 | var = permute_vec_elements (prev_res, var, | |
2315 | perm_mask, stmt, gsi); | |
2316 | new_stmt = SSA_NAME_DEF_STMT (var); | |
2317 | } | |
2318 | ||
2319 | if (prev_stmt_info == NULL) | |
2320 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2321 | else | |
2322 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2323 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2324 | } | |
3efe2e2c JJ |
2325 | |
2326 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2327 | from the IL. */ | |
e6f5c25d IE |
2328 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
2329 | { | |
2330 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
2331 | stmt_info = vinfo_for_stmt (stmt); | |
2332 | } | |
3efe2e2c JJ |
2333 | tree lhs = gimple_call_lhs (stmt); |
2334 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2335 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2336 | set_vinfo_for_stmt (stmt, NULL); | |
2337 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2338 | gsi_replace (gsi, new_stmt, true); | |
5ce9450f JJ |
2339 | return true; |
2340 | } | |
2de001ee | 2341 | else if (vls_type != VLS_LOAD) |
5ce9450f JJ |
2342 | { |
2343 | tree vec_rhs = NULL_TREE, vec_mask = NULL_TREE; | |
2344 | prev_stmt_info = NULL; | |
2d4dc223 | 2345 | LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true; |
5ce9450f JJ |
2346 | for (i = 0; i < ncopies; i++) |
2347 | { | |
2348 | unsigned align, misalign; | |
2349 | ||
2350 | if (i == 0) | |
2351 | { | |
2352 | tree rhs = gimple_call_arg (stmt, 3); | |
81c40241 | 2353 | vec_rhs = vect_get_vec_def_for_operand (rhs, stmt); |
7251b0bf RS |
2354 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, |
2355 | mask_vectype); | |
5ce9450f JJ |
2356 | /* We should have catched mismatched types earlier. */ |
2357 | gcc_assert (useless_type_conversion_p (vectype, | |
2358 | TREE_TYPE (vec_rhs))); | |
2359 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, | |
2360 | NULL_TREE, &dummy, gsi, | |
2361 | &ptr_incr, false, &inv_p); | |
2362 | gcc_assert (!inv_p); | |
2363 | } | |
2364 | else | |
2365 | { | |
81c40241 | 2366 | vect_is_simple_use (vec_rhs, loop_vinfo, &def_stmt, &dt); |
5ce9450f | 2367 | vec_rhs = vect_get_vec_def_for_stmt_copy (dt, vec_rhs); |
81c40241 | 2368 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2369 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2370 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2371 | TYPE_SIZE_UNIT (vectype)); | |
2372 | } | |
2373 | ||
f702e7d4 | 2374 | align = DR_TARGET_ALIGNMENT (dr); |
5ce9450f JJ |
2375 | if (aligned_access_p (dr)) |
2376 | misalign = 0; | |
2377 | else if (DR_MISALIGNMENT (dr) == -1) | |
2378 | { | |
2379 | align = TYPE_ALIGN_UNIT (elem_type); | |
2380 | misalign = 0; | |
2381 | } | |
2382 | else | |
2383 | misalign = DR_MISALIGNMENT (dr); | |
2384 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2385 | misalign); | |
08554c26 | 2386 | tree ptr = build_int_cst (TREE_TYPE (gimple_call_arg (stmt, 1)), |
146ec50f | 2387 | misalign ? least_bit_hwi (misalign) : align); |
a844293d | 2388 | gcall *call |
5ce9450f | 2389 | = gimple_build_call_internal (IFN_MASK_STORE, 4, dataref_ptr, |
08554c26 | 2390 | ptr, vec_mask, vec_rhs); |
a844293d RS |
2391 | gimple_call_set_nothrow (call, true); |
2392 | new_stmt = call; | |
5ce9450f JJ |
2393 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2394 | if (i == 0) | |
2395 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2396 | else | |
2397 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2398 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2399 | } | |
2400 | } | |
2401 | else | |
2402 | { | |
2403 | tree vec_mask = NULL_TREE; | |
2404 | prev_stmt_info = NULL; | |
2405 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); | |
2406 | for (i = 0; i < ncopies; i++) | |
2407 | { | |
2408 | unsigned align, misalign; | |
2409 | ||
2410 | if (i == 0) | |
2411 | { | |
7251b0bf RS |
2412 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, |
2413 | mask_vectype); | |
5ce9450f JJ |
2414 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, |
2415 | NULL_TREE, &dummy, gsi, | |
2416 | &ptr_incr, false, &inv_p); | |
2417 | gcc_assert (!inv_p); | |
2418 | } | |
2419 | else | |
2420 | { | |
81c40241 | 2421 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2422 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2423 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2424 | TYPE_SIZE_UNIT (vectype)); | |
2425 | } | |
2426 | ||
f702e7d4 | 2427 | align = DR_TARGET_ALIGNMENT (dr); |
5ce9450f JJ |
2428 | if (aligned_access_p (dr)) |
2429 | misalign = 0; | |
2430 | else if (DR_MISALIGNMENT (dr) == -1) | |
2431 | { | |
2432 | align = TYPE_ALIGN_UNIT (elem_type); | |
2433 | misalign = 0; | |
2434 | } | |
2435 | else | |
2436 | misalign = DR_MISALIGNMENT (dr); | |
2437 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2438 | misalign); | |
08554c26 | 2439 | tree ptr = build_int_cst (TREE_TYPE (gimple_call_arg (stmt, 1)), |
146ec50f | 2440 | misalign ? least_bit_hwi (misalign) : align); |
a844293d | 2441 | gcall *call |
5ce9450f | 2442 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, dataref_ptr, |
08554c26 | 2443 | ptr, vec_mask); |
a844293d RS |
2444 | gimple_call_set_lhs (call, make_ssa_name (vec_dest)); |
2445 | gimple_call_set_nothrow (call, true); | |
2446 | vect_finish_stmt_generation (stmt, call, gsi); | |
5ce9450f | 2447 | if (i == 0) |
a844293d | 2448 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = call; |
5ce9450f | 2449 | else |
a844293d RS |
2450 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = call; |
2451 | prev_stmt_info = vinfo_for_stmt (call); | |
5ce9450f JJ |
2452 | } |
2453 | } | |
2454 | ||
2de001ee | 2455 | if (vls_type == VLS_LOAD) |
3efe2e2c JJ |
2456 | { |
2457 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2458 | from the IL. */ | |
e6f5c25d IE |
2459 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
2460 | { | |
2461 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
2462 | stmt_info = vinfo_for_stmt (stmt); | |
2463 | } | |
3efe2e2c JJ |
2464 | tree lhs = gimple_call_lhs (stmt); |
2465 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2466 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2467 | set_vinfo_for_stmt (stmt, NULL); | |
2468 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2469 | gsi_replace (gsi, new_stmt, true); | |
2470 | } | |
2471 | ||
5ce9450f JJ |
2472 | return true; |
2473 | } | |
2474 | ||
37b14185 RB |
2475 | /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64}. */ |
2476 | ||
2477 | static bool | |
2478 | vectorizable_bswap (gimple *stmt, gimple_stmt_iterator *gsi, | |
2479 | gimple **vec_stmt, slp_tree slp_node, | |
2480 | tree vectype_in, enum vect_def_type *dt) | |
2481 | { | |
2482 | tree op, vectype; | |
2483 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2484 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2485 | unsigned ncopies, nunits; | |
2486 | ||
2487 | op = gimple_call_arg (stmt, 0); | |
2488 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2489 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2490 | ||
2491 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2492 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2493 | case of SLP. */ | |
2494 | if (slp_node) | |
2495 | ncopies = 1; | |
2496 | else | |
e8f142e2 | 2497 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
37b14185 RB |
2498 | |
2499 | gcc_assert (ncopies >= 1); | |
2500 | ||
2501 | tree char_vectype = get_same_sized_vectype (char_type_node, vectype_in); | |
2502 | if (! char_vectype) | |
2503 | return false; | |
2504 | ||
794e3180 | 2505 | unsigned int num_bytes = TYPE_VECTOR_SUBPARTS (char_vectype); |
794e3180 | 2506 | unsigned word_bytes = num_bytes / nunits; |
908a1a16 | 2507 | |
d980067b RS |
2508 | /* The encoding uses one stepped pattern for each byte in the word. */ |
2509 | vec_perm_builder elts (num_bytes, word_bytes, 3); | |
2510 | for (unsigned i = 0; i < 3; ++i) | |
37b14185 | 2511 | for (unsigned j = 0; j < word_bytes; ++j) |
908a1a16 | 2512 | elts.quick_push ((i + 1) * word_bytes - j - 1); |
37b14185 | 2513 | |
e3342de4 RS |
2514 | vec_perm_indices indices (elts, 1, num_bytes); |
2515 | if (!can_vec_perm_const_p (TYPE_MODE (char_vectype), indices)) | |
37b14185 RB |
2516 | return false; |
2517 | ||
2518 | if (! vec_stmt) | |
2519 | { | |
2520 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
2521 | if (dump_enabled_p ()) | |
2522 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_bswap ===" | |
2523 | "\n"); | |
2524 | if (! PURE_SLP_STMT (stmt_info)) | |
2525 | { | |
2526 | add_stmt_cost (stmt_info->vinfo->target_cost_data, | |
2527 | 1, vector_stmt, stmt_info, 0, vect_prologue); | |
2528 | add_stmt_cost (stmt_info->vinfo->target_cost_data, | |
2529 | ncopies, vec_perm, stmt_info, 0, vect_body); | |
2530 | } | |
2531 | return true; | |
2532 | } | |
2533 | ||
736d0f28 | 2534 | tree bswap_vconst = vec_perm_indices_to_tree (char_vectype, indices); |
37b14185 RB |
2535 | |
2536 | /* Transform. */ | |
2537 | vec<tree> vec_oprnds = vNULL; | |
2538 | gimple *new_stmt = NULL; | |
2539 | stmt_vec_info prev_stmt_info = NULL; | |
2540 | for (unsigned j = 0; j < ncopies; j++) | |
2541 | { | |
2542 | /* Handle uses. */ | |
2543 | if (j == 0) | |
306b0c92 | 2544 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
37b14185 RB |
2545 | else |
2546 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
2547 | ||
2548 | /* Arguments are ready. create the new vector stmt. */ | |
2549 | unsigned i; | |
2550 | tree vop; | |
2551 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) | |
2552 | { | |
2553 | tree tem = make_ssa_name (char_vectype); | |
2554 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
2555 | char_vectype, vop)); | |
2556 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2557 | tree tem2 = make_ssa_name (char_vectype); | |
2558 | new_stmt = gimple_build_assign (tem2, VEC_PERM_EXPR, | |
2559 | tem, tem, bswap_vconst); | |
2560 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2561 | tem = make_ssa_name (vectype); | |
2562 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
2563 | vectype, tem2)); | |
2564 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2565 | if (slp_node) | |
2566 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
2567 | } | |
2568 | ||
2569 | if (slp_node) | |
2570 | continue; | |
2571 | ||
2572 | if (j == 0) | |
2573 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2574 | else | |
2575 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2576 | ||
2577 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2578 | } | |
2579 | ||
2580 | vec_oprnds.release (); | |
2581 | return true; | |
2582 | } | |
2583 | ||
b1b6836e RS |
2584 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
2585 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT | |
2586 | in a single step. On success, store the binary pack code in | |
2587 | *CONVERT_CODE. */ | |
2588 | ||
2589 | static bool | |
2590 | simple_integer_narrowing (tree vectype_out, tree vectype_in, | |
2591 | tree_code *convert_code) | |
2592 | { | |
2593 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) | |
2594 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) | |
2595 | return false; | |
2596 | ||
2597 | tree_code code; | |
2598 | int multi_step_cvt = 0; | |
2599 | auto_vec <tree, 8> interm_types; | |
2600 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, | |
2601 | &code, &multi_step_cvt, | |
2602 | &interm_types) | |
2603 | || multi_step_cvt) | |
2604 | return false; | |
2605 | ||
2606 | *convert_code = code; | |
2607 | return true; | |
2608 | } | |
5ce9450f | 2609 | |
ebfd146a IR |
2610 | /* Function vectorizable_call. |
2611 | ||
538dd0b7 | 2612 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 2613 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2614 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2615 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2616 | ||
2617 | static bool | |
355fe088 | 2618 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
190c2236 | 2619 | slp_tree slp_node) |
ebfd146a | 2620 | { |
538dd0b7 | 2621 | gcall *stmt; |
ebfd146a IR |
2622 | tree vec_dest; |
2623 | tree scalar_dest; | |
2624 | tree op, type; | |
2625 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
538dd0b7 | 2626 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a IR |
2627 | tree vectype_out, vectype_in; |
2628 | int nunits_in; | |
2629 | int nunits_out; | |
2630 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 2631 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 2632 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 2633 | tree fndecl, new_temp, rhs_type; |
355fe088 | 2634 | gimple *def_stmt; |
0502fb85 UB |
2635 | enum vect_def_type dt[3] |
2636 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 2637 | int ndts = 3; |
355fe088 | 2638 | gimple *new_stmt = NULL; |
ebfd146a | 2639 | int ncopies, j; |
6e1aa848 | 2640 | vec<tree> vargs = vNULL; |
ebfd146a IR |
2641 | enum { NARROW, NONE, WIDEN } modifier; |
2642 | size_t i, nargs; | |
9d5e7640 | 2643 | tree lhs; |
ebfd146a | 2644 | |
190c2236 | 2645 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2646 | return false; |
2647 | ||
66c16fd9 RB |
2648 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
2649 | && ! vec_stmt) | |
ebfd146a IR |
2650 | return false; |
2651 | ||
538dd0b7 DM |
2652 | /* Is GS a vectorizable call? */ |
2653 | stmt = dyn_cast <gcall *> (gs); | |
2654 | if (!stmt) | |
ebfd146a IR |
2655 | return false; |
2656 | ||
5ce9450f JJ |
2657 | if (gimple_call_internal_p (stmt) |
2658 | && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD | |
2659 | || gimple_call_internal_fn (stmt) == IFN_MASK_STORE)) | |
2660 | return vectorizable_mask_load_store (stmt, gsi, vec_stmt, | |
2661 | slp_node); | |
2662 | ||
0136f8f0 AH |
2663 | if (gimple_call_lhs (stmt) == NULL_TREE |
2664 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
2665 | return false; |
2666 | ||
0136f8f0 | 2667 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 2668 | |
b690cc0f RG |
2669 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2670 | ||
ebfd146a IR |
2671 | /* Process function arguments. */ |
2672 | rhs_type = NULL_TREE; | |
b690cc0f | 2673 | vectype_in = NULL_TREE; |
ebfd146a IR |
2674 | nargs = gimple_call_num_args (stmt); |
2675 | ||
1b1562a5 MM |
2676 | /* Bail out if the function has more than three arguments, we do not have |
2677 | interesting builtin functions to vectorize with more than two arguments | |
2678 | except for fma. No arguments is also not good. */ | |
2679 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
2680 | return false; |
2681 | ||
74bf76ed JJ |
2682 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2683 | if (gimple_call_internal_p (stmt) | |
2684 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2685 | { | |
2686 | nargs = 0; | |
2687 | rhs_type = unsigned_type_node; | |
2688 | } | |
2689 | ||
ebfd146a IR |
2690 | for (i = 0; i < nargs; i++) |
2691 | { | |
b690cc0f RG |
2692 | tree opvectype; |
2693 | ||
ebfd146a IR |
2694 | op = gimple_call_arg (stmt, i); |
2695 | ||
2696 | /* We can only handle calls with arguments of the same type. */ | |
2697 | if (rhs_type | |
8533c9d8 | 2698 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 2699 | { |
73fbfcad | 2700 | if (dump_enabled_p ()) |
78c60e3d | 2701 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2702 | "argument types differ.\n"); |
ebfd146a IR |
2703 | return false; |
2704 | } | |
b690cc0f RG |
2705 | if (!rhs_type) |
2706 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 2707 | |
81c40241 | 2708 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[i], &opvectype)) |
ebfd146a | 2709 | { |
73fbfcad | 2710 | if (dump_enabled_p ()) |
78c60e3d | 2711 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2712 | "use not simple.\n"); |
ebfd146a IR |
2713 | return false; |
2714 | } | |
ebfd146a | 2715 | |
b690cc0f RG |
2716 | if (!vectype_in) |
2717 | vectype_in = opvectype; | |
2718 | else if (opvectype | |
2719 | && opvectype != vectype_in) | |
2720 | { | |
73fbfcad | 2721 | if (dump_enabled_p ()) |
78c60e3d | 2722 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2723 | "argument vector types differ.\n"); |
b690cc0f RG |
2724 | return false; |
2725 | } | |
2726 | } | |
2727 | /* If all arguments are external or constant defs use a vector type with | |
2728 | the same size as the output vector type. */ | |
ebfd146a | 2729 | if (!vectype_in) |
b690cc0f | 2730 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2731 | if (vec_stmt) |
2732 | gcc_assert (vectype_in); | |
2733 | if (!vectype_in) | |
2734 | { | |
73fbfcad | 2735 | if (dump_enabled_p ()) |
7d8930a0 | 2736 | { |
78c60e3d SS |
2737 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2738 | "no vectype for scalar type "); | |
2739 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2740 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
2741 | } |
2742 | ||
2743 | return false; | |
2744 | } | |
ebfd146a IR |
2745 | |
2746 | /* FORNOW */ | |
b690cc0f RG |
2747 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2748 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
2749 | if (nunits_in == nunits_out / 2) |
2750 | modifier = NARROW; | |
2751 | else if (nunits_out == nunits_in) | |
2752 | modifier = NONE; | |
2753 | else if (nunits_out == nunits_in / 2) | |
2754 | modifier = WIDEN; | |
2755 | else | |
2756 | return false; | |
2757 | ||
70439f0d RS |
2758 | /* We only handle functions that do not read or clobber memory. */ |
2759 | if (gimple_vuse (stmt)) | |
2760 | { | |
2761 | if (dump_enabled_p ()) | |
2762 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2763 | "function reads from or writes to memory.\n"); | |
2764 | return false; | |
2765 | } | |
2766 | ||
ebfd146a IR |
2767 | /* For now, we only vectorize functions if a target specific builtin |
2768 | is available. TODO -- in some cases, it might be profitable to | |
2769 | insert the calls for pieces of the vector, in order to be able | |
2770 | to vectorize other operations in the loop. */ | |
70439f0d RS |
2771 | fndecl = NULL_TREE; |
2772 | internal_fn ifn = IFN_LAST; | |
2773 | combined_fn cfn = gimple_call_combined_fn (stmt); | |
2774 | tree callee = gimple_call_fndecl (stmt); | |
2775 | ||
2776 | /* First try using an internal function. */ | |
b1b6836e RS |
2777 | tree_code convert_code = ERROR_MARK; |
2778 | if (cfn != CFN_LAST | |
2779 | && (modifier == NONE | |
2780 | || (modifier == NARROW | |
2781 | && simple_integer_narrowing (vectype_out, vectype_in, | |
2782 | &convert_code)))) | |
70439f0d RS |
2783 | ifn = vectorizable_internal_function (cfn, callee, vectype_out, |
2784 | vectype_in); | |
2785 | ||
2786 | /* If that fails, try asking for a target-specific built-in function. */ | |
2787 | if (ifn == IFN_LAST) | |
2788 | { | |
2789 | if (cfn != CFN_LAST) | |
2790 | fndecl = targetm.vectorize.builtin_vectorized_function | |
2791 | (cfn, vectype_out, vectype_in); | |
2792 | else | |
2793 | fndecl = targetm.vectorize.builtin_md_vectorized_function | |
2794 | (callee, vectype_out, vectype_in); | |
2795 | } | |
2796 | ||
2797 | if (ifn == IFN_LAST && !fndecl) | |
ebfd146a | 2798 | { |
70439f0d | 2799 | if (cfn == CFN_GOMP_SIMD_LANE |
74bf76ed JJ |
2800 | && !slp_node |
2801 | && loop_vinfo | |
2802 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2803 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
2804 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2805 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
2806 | { | |
2807 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
2808 | { 0, 1, 2, ... vf - 1 } vector. */ | |
2809 | gcc_assert (nargs == 0); | |
2810 | } | |
37b14185 RB |
2811 | else if (modifier == NONE |
2812 | && (gimple_call_builtin_p (stmt, BUILT_IN_BSWAP16) | |
2813 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP32) | |
2814 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP64))) | |
2815 | return vectorizable_bswap (stmt, gsi, vec_stmt, slp_node, | |
2816 | vectype_in, dt); | |
74bf76ed JJ |
2817 | else |
2818 | { | |
2819 | if (dump_enabled_p ()) | |
2820 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 2821 | "function is not vectorizable.\n"); |
74bf76ed JJ |
2822 | return false; |
2823 | } | |
ebfd146a IR |
2824 | } |
2825 | ||
fce57248 | 2826 | if (slp_node) |
190c2236 | 2827 | ncopies = 1; |
b1b6836e | 2828 | else if (modifier == NARROW && ifn == IFN_LAST) |
e8f142e2 | 2829 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
ebfd146a | 2830 | else |
e8f142e2 | 2831 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
ebfd146a IR |
2832 | |
2833 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2834 | needs to be generated. */ | |
2835 | gcc_assert (ncopies >= 1); | |
2836 | ||
2837 | if (!vec_stmt) /* transformation not required. */ | |
2838 | { | |
2839 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 2840 | if (dump_enabled_p ()) |
e645e942 TJ |
2841 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
2842 | "\n"); | |
4fc5ebf1 | 2843 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
b1b6836e RS |
2844 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) |
2845 | add_stmt_cost (stmt_info->vinfo->target_cost_data, ncopies / 2, | |
2846 | vec_promote_demote, stmt_info, 0, vect_body); | |
2847 | ||
ebfd146a IR |
2848 | return true; |
2849 | } | |
2850 | ||
67b8dbac | 2851 | /* Transform. */ |
ebfd146a | 2852 | |
73fbfcad | 2853 | if (dump_enabled_p ()) |
e645e942 | 2854 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
2855 | |
2856 | /* Handle def. */ | |
2857 | scalar_dest = gimple_call_lhs (stmt); | |
2858 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2859 | ||
2860 | prev_stmt_info = NULL; | |
b1b6836e | 2861 | if (modifier == NONE || ifn != IFN_LAST) |
ebfd146a | 2862 | { |
b1b6836e | 2863 | tree prev_res = NULL_TREE; |
ebfd146a IR |
2864 | for (j = 0; j < ncopies; ++j) |
2865 | { | |
2866 | /* Build argument list for the vectorized call. */ | |
2867 | if (j == 0) | |
9771b263 | 2868 | vargs.create (nargs); |
ebfd146a | 2869 | else |
9771b263 | 2870 | vargs.truncate (0); |
ebfd146a | 2871 | |
190c2236 JJ |
2872 | if (slp_node) |
2873 | { | |
ef062b13 | 2874 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2875 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2876 | |
2877 | for (i = 0; i < nargs; i++) | |
9771b263 | 2878 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 2879 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 2880 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2881 | |
2882 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2883 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
2884 | { |
2885 | size_t k; | |
2886 | for (k = 0; k < nargs; k++) | |
2887 | { | |
37b5ec8f | 2888 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 2889 | vargs[k] = vec_oprndsk[i]; |
190c2236 | 2890 | } |
b1b6836e RS |
2891 | if (modifier == NARROW) |
2892 | { | |
2893 | tree half_res = make_ssa_name (vectype_in); | |
a844293d RS |
2894 | gcall *call |
2895 | = gimple_build_call_internal_vec (ifn, vargs); | |
2896 | gimple_call_set_lhs (call, half_res); | |
2897 | gimple_call_set_nothrow (call, true); | |
2898 | new_stmt = call; | |
b1b6836e RS |
2899 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2900 | if ((i & 1) == 0) | |
2901 | { | |
2902 | prev_res = half_res; | |
2903 | continue; | |
2904 | } | |
2905 | new_temp = make_ssa_name (vec_dest); | |
2906 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
2907 | prev_res, half_res); | |
2908 | } | |
70439f0d | 2909 | else |
b1b6836e | 2910 | { |
a844293d | 2911 | gcall *call; |
b1b6836e | 2912 | if (ifn != IFN_LAST) |
a844293d | 2913 | call = gimple_build_call_internal_vec (ifn, vargs); |
b1b6836e | 2914 | else |
a844293d RS |
2915 | call = gimple_build_call_vec (fndecl, vargs); |
2916 | new_temp = make_ssa_name (vec_dest, call); | |
2917 | gimple_call_set_lhs (call, new_temp); | |
2918 | gimple_call_set_nothrow (call, true); | |
2919 | new_stmt = call; | |
b1b6836e | 2920 | } |
190c2236 | 2921 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 2922 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2923 | } |
2924 | ||
2925 | for (i = 0; i < nargs; i++) | |
2926 | { | |
37b5ec8f | 2927 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2928 | vec_oprndsi.release (); |
190c2236 | 2929 | } |
190c2236 JJ |
2930 | continue; |
2931 | } | |
2932 | ||
ebfd146a IR |
2933 | for (i = 0; i < nargs; i++) |
2934 | { | |
2935 | op = gimple_call_arg (stmt, i); | |
2936 | if (j == 0) | |
2937 | vec_oprnd0 | |
81c40241 | 2938 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2939 | else |
63827fb8 IR |
2940 | { |
2941 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
2942 | vec_oprnd0 | |
2943 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
2944 | } | |
ebfd146a | 2945 | |
9771b263 | 2946 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
2947 | } |
2948 | ||
74bf76ed JJ |
2949 | if (gimple_call_internal_p (stmt) |
2950 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2951 | { | |
5ebaa477 RS |
2952 | tree_vector_builder v (vectype_out, 1, 3); |
2953 | for (int k = 0; k < 3; ++k) | |
794e3180 RS |
2954 | v.quick_push (build_int_cst (unsigned_type_node, |
2955 | j * nunits_out + k)); | |
5ebaa477 | 2956 | tree cst = v.build (); |
74bf76ed | 2957 | tree new_var |
0e22bb5a | 2958 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 2959 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 2960 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 2961 | new_temp = make_ssa_name (vec_dest); |
0e22bb5a | 2962 | new_stmt = gimple_build_assign (new_temp, new_var); |
74bf76ed | 2963 | } |
b1b6836e RS |
2964 | else if (modifier == NARROW) |
2965 | { | |
2966 | tree half_res = make_ssa_name (vectype_in); | |
a844293d RS |
2967 | gcall *call = gimple_build_call_internal_vec (ifn, vargs); |
2968 | gimple_call_set_lhs (call, half_res); | |
2969 | gimple_call_set_nothrow (call, true); | |
2970 | new_stmt = call; | |
b1b6836e RS |
2971 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2972 | if ((j & 1) == 0) | |
2973 | { | |
2974 | prev_res = half_res; | |
2975 | continue; | |
2976 | } | |
2977 | new_temp = make_ssa_name (vec_dest); | |
2978 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
2979 | prev_res, half_res); | |
2980 | } | |
74bf76ed JJ |
2981 | else |
2982 | { | |
a844293d | 2983 | gcall *call; |
70439f0d | 2984 | if (ifn != IFN_LAST) |
a844293d | 2985 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 2986 | else |
a844293d | 2987 | call = gimple_build_call_vec (fndecl, vargs); |
74bf76ed | 2988 | new_temp = make_ssa_name (vec_dest, new_stmt); |
a844293d RS |
2989 | gimple_call_set_lhs (call, new_temp); |
2990 | gimple_call_set_nothrow (call, true); | |
2991 | new_stmt = call; | |
74bf76ed | 2992 | } |
ebfd146a IR |
2993 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2994 | ||
b1b6836e | 2995 | if (j == (modifier == NARROW ? 1 : 0)) |
ebfd146a IR |
2996 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; |
2997 | else | |
2998 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2999 | ||
3000 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3001 | } | |
b1b6836e RS |
3002 | } |
3003 | else if (modifier == NARROW) | |
3004 | { | |
ebfd146a IR |
3005 | for (j = 0; j < ncopies; ++j) |
3006 | { | |
3007 | /* Build argument list for the vectorized call. */ | |
3008 | if (j == 0) | |
9771b263 | 3009 | vargs.create (nargs * 2); |
ebfd146a | 3010 | else |
9771b263 | 3011 | vargs.truncate (0); |
ebfd146a | 3012 | |
190c2236 JJ |
3013 | if (slp_node) |
3014 | { | |
ef062b13 | 3015 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 3016 | vec<tree> vec_oprnds0; |
190c2236 JJ |
3017 | |
3018 | for (i = 0; i < nargs; i++) | |
9771b263 | 3019 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 3020 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 3021 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
3022 | |
3023 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3024 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
3025 | { |
3026 | size_t k; | |
9771b263 | 3027 | vargs.truncate (0); |
190c2236 JJ |
3028 | for (k = 0; k < nargs; k++) |
3029 | { | |
37b5ec8f | 3030 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
3031 | vargs.quick_push (vec_oprndsk[i]); |
3032 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 | 3033 | } |
a844293d | 3034 | gcall *call; |
70439f0d | 3035 | if (ifn != IFN_LAST) |
a844293d | 3036 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3037 | else |
a844293d RS |
3038 | call = gimple_build_call_vec (fndecl, vargs); |
3039 | new_temp = make_ssa_name (vec_dest, call); | |
3040 | gimple_call_set_lhs (call, new_temp); | |
3041 | gimple_call_set_nothrow (call, true); | |
3042 | new_stmt = call; | |
190c2236 | 3043 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 3044 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
3045 | } |
3046 | ||
3047 | for (i = 0; i < nargs; i++) | |
3048 | { | |
37b5ec8f | 3049 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3050 | vec_oprndsi.release (); |
190c2236 | 3051 | } |
190c2236 JJ |
3052 | continue; |
3053 | } | |
3054 | ||
ebfd146a IR |
3055 | for (i = 0; i < nargs; i++) |
3056 | { | |
3057 | op = gimple_call_arg (stmt, i); | |
3058 | if (j == 0) | |
3059 | { | |
3060 | vec_oprnd0 | |
81c40241 | 3061 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 3062 | vec_oprnd1 |
63827fb8 | 3063 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3064 | } |
3065 | else | |
3066 | { | |
336ecb65 | 3067 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 3068 | vec_oprnd0 |
63827fb8 | 3069 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 3070 | vec_oprnd1 |
63827fb8 | 3071 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3072 | } |
3073 | ||
9771b263 DN |
3074 | vargs.quick_push (vec_oprnd0); |
3075 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
3076 | } |
3077 | ||
b1b6836e | 3078 | new_stmt = gimple_build_call_vec (fndecl, vargs); |
ebfd146a IR |
3079 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3080 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
3081 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3082 | ||
3083 | if (j == 0) | |
3084 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3085 | else | |
3086 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3087 | ||
3088 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3089 | } | |
3090 | ||
3091 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a | 3092 | } |
b1b6836e RS |
3093 | else |
3094 | /* No current target implements this case. */ | |
3095 | return false; | |
ebfd146a | 3096 | |
9771b263 | 3097 | vargs.release (); |
ebfd146a | 3098 | |
ebfd146a IR |
3099 | /* The call in STMT might prevent it from being removed in dce. |
3100 | We however cannot remove it here, due to the way the ssa name | |
3101 | it defines is mapped to the new definition. So just replace | |
3102 | rhs of the statement with something harmless. */ | |
3103 | ||
dd34c087 JJ |
3104 | if (slp_node) |
3105 | return true; | |
3106 | ||
ebfd146a | 3107 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
3108 | if (is_pattern_stmt_p (stmt_info)) |
3109 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3110 | else | |
3111 | lhs = gimple_call_lhs (stmt); | |
3cc2fa2a | 3112 | |
9d5e7640 | 3113 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
ebfd146a | 3114 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 3115 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
3116 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
3117 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
3118 | |
3119 | return true; | |
3120 | } | |
3121 | ||
3122 | ||
0136f8f0 AH |
3123 | struct simd_call_arg_info |
3124 | { | |
3125 | tree vectype; | |
3126 | tree op; | |
0136f8f0 | 3127 | HOST_WIDE_INT linear_step; |
34e82342 | 3128 | enum vect_def_type dt; |
0136f8f0 | 3129 | unsigned int align; |
17b658af | 3130 | bool simd_lane_linear; |
0136f8f0 AH |
3131 | }; |
3132 | ||
17b658af JJ |
3133 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
3134 | is linear within simd lane (but not within whole loop), note it in | |
3135 | *ARGINFO. */ | |
3136 | ||
3137 | static void | |
3138 | vect_simd_lane_linear (tree op, struct loop *loop, | |
3139 | struct simd_call_arg_info *arginfo) | |
3140 | { | |
355fe088 | 3141 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
3142 | |
3143 | if (!is_gimple_assign (def_stmt) | |
3144 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
3145 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
3146 | return; | |
3147 | ||
3148 | tree base = gimple_assign_rhs1 (def_stmt); | |
3149 | HOST_WIDE_INT linear_step = 0; | |
3150 | tree v = gimple_assign_rhs2 (def_stmt); | |
3151 | while (TREE_CODE (v) == SSA_NAME) | |
3152 | { | |
3153 | tree t; | |
3154 | def_stmt = SSA_NAME_DEF_STMT (v); | |
3155 | if (is_gimple_assign (def_stmt)) | |
3156 | switch (gimple_assign_rhs_code (def_stmt)) | |
3157 | { | |
3158 | case PLUS_EXPR: | |
3159 | t = gimple_assign_rhs2 (def_stmt); | |
3160 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
3161 | return; | |
3162 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
3163 | v = gimple_assign_rhs1 (def_stmt); | |
3164 | continue; | |
3165 | case MULT_EXPR: | |
3166 | t = gimple_assign_rhs2 (def_stmt); | |
3167 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
3168 | return; | |
3169 | linear_step = tree_to_shwi (t); | |
3170 | v = gimple_assign_rhs1 (def_stmt); | |
3171 | continue; | |
3172 | CASE_CONVERT: | |
3173 | t = gimple_assign_rhs1 (def_stmt); | |
3174 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
3175 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
3176 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
3177 | return; | |
3178 | if (!linear_step) | |
3179 | linear_step = 1; | |
3180 | v = t; | |
3181 | continue; | |
3182 | default: | |
3183 | return; | |
3184 | } | |
8e4284d0 | 3185 | else if (gimple_call_internal_p (def_stmt, IFN_GOMP_SIMD_LANE) |
17b658af JJ |
3186 | && loop->simduid |
3187 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
3188 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
3189 | == loop->simduid)) | |
3190 | { | |
3191 | if (!linear_step) | |
3192 | linear_step = 1; | |
3193 | arginfo->linear_step = linear_step; | |
3194 | arginfo->op = base; | |
3195 | arginfo->simd_lane_linear = true; | |
3196 | return; | |
3197 | } | |
3198 | } | |
3199 | } | |
3200 | ||
0136f8f0 AH |
3201 | /* Function vectorizable_simd_clone_call. |
3202 | ||
3203 | Check if STMT performs a function call that can be vectorized | |
3204 | by calling a simd clone of the function. | |
3205 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3206 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3207 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3208 | ||
3209 | static bool | |
355fe088 TS |
3210 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
3211 | gimple **vec_stmt, slp_tree slp_node) | |
0136f8f0 AH |
3212 | { |
3213 | tree vec_dest; | |
3214 | tree scalar_dest; | |
3215 | tree op, type; | |
3216 | tree vec_oprnd0 = NULL_TREE; | |
3217 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
3218 | tree vectype; | |
3219 | unsigned int nunits; | |
3220 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3221 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 3222 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 3223 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 3224 | tree fndecl, new_temp; |
355fe088 TS |
3225 | gimple *def_stmt; |
3226 | gimple *new_stmt = NULL; | |
0136f8f0 | 3227 | int ncopies, j; |
00426f9a | 3228 | auto_vec<simd_call_arg_info> arginfo; |
0136f8f0 AH |
3229 | vec<tree> vargs = vNULL; |
3230 | size_t i, nargs; | |
3231 | tree lhs, rtype, ratype; | |
e7a74006 | 3232 | vec<constructor_elt, va_gc> *ret_ctor_elts = NULL; |
0136f8f0 AH |
3233 | |
3234 | /* Is STMT a vectorizable call? */ | |
3235 | if (!is_gimple_call (stmt)) | |
3236 | return false; | |
3237 | ||
3238 | fndecl = gimple_call_fndecl (stmt); | |
3239 | if (fndecl == NULL_TREE) | |
3240 | return false; | |
3241 | ||
d52f5295 | 3242 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
3243 | if (node == NULL || node->simd_clones == NULL) |
3244 | return false; | |
3245 | ||
3246 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
3247 | return false; | |
3248 | ||
66c16fd9 RB |
3249 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3250 | && ! vec_stmt) | |
0136f8f0 AH |
3251 | return false; |
3252 | ||
3253 | if (gimple_call_lhs (stmt) | |
3254 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
3255 | return false; | |
3256 | ||
3257 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
3258 | ||
3259 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3260 | ||
3261 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
3262 | return false; | |
3263 | ||
3264 | /* FORNOW */ | |
fce57248 | 3265 | if (slp_node) |
0136f8f0 AH |
3266 | return false; |
3267 | ||
3268 | /* Process function arguments. */ | |
3269 | nargs = gimple_call_num_args (stmt); | |
3270 | ||
3271 | /* Bail out if the function has zero arguments. */ | |
3272 | if (nargs == 0) | |
3273 | return false; | |
3274 | ||
00426f9a | 3275 | arginfo.reserve (nargs, true); |
0136f8f0 AH |
3276 | |
3277 | for (i = 0; i < nargs; i++) | |
3278 | { | |
3279 | simd_call_arg_info thisarginfo; | |
3280 | affine_iv iv; | |
3281 | ||
3282 | thisarginfo.linear_step = 0; | |
3283 | thisarginfo.align = 0; | |
3284 | thisarginfo.op = NULL_TREE; | |
17b658af | 3285 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
3286 | |
3287 | op = gimple_call_arg (stmt, i); | |
81c40241 RB |
3288 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &thisarginfo.dt, |
3289 | &thisarginfo.vectype) | |
0136f8f0 AH |
3290 | || thisarginfo.dt == vect_uninitialized_def) |
3291 | { | |
3292 | if (dump_enabled_p ()) | |
3293 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3294 | "use not simple.\n"); | |
0136f8f0 AH |
3295 | return false; |
3296 | } | |
3297 | ||
3298 | if (thisarginfo.dt == vect_constant_def | |
3299 | || thisarginfo.dt == vect_external_def) | |
3300 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
3301 | else | |
3302 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
3303 | ||
6c9e85fb JJ |
3304 | /* For linear arguments, the analyze phase should have saved |
3305 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
3306 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
3307 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
3308 | { |
3309 | gcc_assert (vec_stmt); | |
3310 | thisarginfo.linear_step | |
17b658af | 3311 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 3312 | thisarginfo.op |
17b658af JJ |
3313 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
3314 | thisarginfo.simd_lane_linear | |
3315 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
3316 | == boolean_true_node); | |
6c9e85fb JJ |
3317 | /* If loop has been peeled for alignment, we need to adjust it. */ |
3318 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
3319 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 3320 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
3321 | { |
3322 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 3323 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
3324 | tree opt = TREE_TYPE (thisarginfo.op); |
3325 | bias = fold_convert (TREE_TYPE (step), bias); | |
3326 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
3327 | thisarginfo.op | |
3328 | = fold_build2 (POINTER_TYPE_P (opt) | |
3329 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
3330 | thisarginfo.op, bias); | |
3331 | } | |
3332 | } | |
3333 | else if (!vec_stmt | |
3334 | && thisarginfo.dt != vect_constant_def | |
3335 | && thisarginfo.dt != vect_external_def | |
3336 | && loop_vinfo | |
3337 | && TREE_CODE (op) == SSA_NAME | |
3338 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
3339 | &iv, false) | |
3340 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
3341 | { |
3342 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
3343 | thisarginfo.op = iv.base; | |
3344 | } | |
3345 | else if ((thisarginfo.dt == vect_constant_def | |
3346 | || thisarginfo.dt == vect_external_def) | |
3347 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
3348 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
3349 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
3350 | linear too. */ | |
3351 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
3352 | && !thisarginfo.linear_step | |
3353 | && !vec_stmt | |
3354 | && thisarginfo.dt != vect_constant_def | |
3355 | && thisarginfo.dt != vect_external_def | |
3356 | && loop_vinfo | |
3357 | && !slp_node | |
3358 | && TREE_CODE (op) == SSA_NAME) | |
3359 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
3360 | |
3361 | arginfo.quick_push (thisarginfo); | |
3362 | } | |
3363 | ||
d9f21f6a RS |
3364 | unsigned HOST_WIDE_INT vf; |
3365 | if (!LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&vf)) | |
3366 | { | |
3367 | if (dump_enabled_p ()) | |
3368 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3369 | "not considering SIMD clones; not yet supported" | |
3370 | " for variable-width vectors.\n"); | |
3371 | return NULL; | |
3372 | } | |
3373 | ||
0136f8f0 AH |
3374 | unsigned int badness = 0; |
3375 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
3376 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
3377 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
3378 | else |
3379 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
3380 | n = n->simdclone->next_clone) | |
3381 | { | |
3382 | unsigned int this_badness = 0; | |
d9f21f6a | 3383 | if (n->simdclone->simdlen > vf |
0136f8f0 AH |
3384 | || n->simdclone->nargs != nargs) |
3385 | continue; | |
d9f21f6a RS |
3386 | if (n->simdclone->simdlen < vf) |
3387 | this_badness += (exact_log2 (vf) | |
0136f8f0 AH |
3388 | - exact_log2 (n->simdclone->simdlen)) * 1024; |
3389 | if (n->simdclone->inbranch) | |
3390 | this_badness += 2048; | |
3391 | int target_badness = targetm.simd_clone.usable (n); | |
3392 | if (target_badness < 0) | |
3393 | continue; | |
3394 | this_badness += target_badness * 512; | |
3395 | /* FORNOW: Have to add code to add the mask argument. */ | |
3396 | if (n->simdclone->inbranch) | |
3397 | continue; | |
3398 | for (i = 0; i < nargs; i++) | |
3399 | { | |
3400 | switch (n->simdclone->args[i].arg_type) | |
3401 | { | |
3402 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3403 | if (!useless_type_conversion_p | |
3404 | (n->simdclone->args[i].orig_type, | |
3405 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
3406 | i = -1; | |
3407 | else if (arginfo[i].dt == vect_constant_def | |
3408 | || arginfo[i].dt == vect_external_def | |
3409 | || arginfo[i].linear_step) | |
3410 | this_badness += 64; | |
3411 | break; | |
3412 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3413 | if (arginfo[i].dt != vect_constant_def | |
3414 | && arginfo[i].dt != vect_external_def) | |
3415 | i = -1; | |
3416 | break; | |
3417 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 3418 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
3419 | if (arginfo[i].dt == vect_constant_def |
3420 | || arginfo[i].dt == vect_external_def | |
3421 | || (arginfo[i].linear_step | |
3422 | != n->simdclone->args[i].linear_step)) | |
3423 | i = -1; | |
3424 | break; | |
3425 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
3426 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
3427 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
3428 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3429 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3430 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3431 | /* FORNOW */ |
3432 | i = -1; | |
3433 | break; | |
3434 | case SIMD_CLONE_ARG_TYPE_MASK: | |
3435 | gcc_unreachable (); | |
3436 | } | |
3437 | if (i == (size_t) -1) | |
3438 | break; | |
3439 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
3440 | { | |
3441 | i = -1; | |
3442 | break; | |
3443 | } | |
3444 | if (arginfo[i].align) | |
3445 | this_badness += (exact_log2 (arginfo[i].align) | |
3446 | - exact_log2 (n->simdclone->args[i].alignment)); | |
3447 | } | |
3448 | if (i == (size_t) -1) | |
3449 | continue; | |
3450 | if (bestn == NULL || this_badness < badness) | |
3451 | { | |
3452 | bestn = n; | |
3453 | badness = this_badness; | |
3454 | } | |
3455 | } | |
3456 | ||
3457 | if (bestn == NULL) | |
00426f9a | 3458 | return false; |
0136f8f0 AH |
3459 | |
3460 | for (i = 0; i < nargs; i++) | |
3461 | if ((arginfo[i].dt == vect_constant_def | |
3462 | || arginfo[i].dt == vect_external_def) | |
3463 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
3464 | { | |
3465 | arginfo[i].vectype | |
3466 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
3467 | i))); | |
3468 | if (arginfo[i].vectype == NULL | |
3469 | || (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
3470 | > bestn->simdclone->simdlen)) | |
00426f9a | 3471 | return false; |
0136f8f0 AH |
3472 | } |
3473 | ||
3474 | fndecl = bestn->decl; | |
3475 | nunits = bestn->simdclone->simdlen; | |
d9f21f6a | 3476 | ncopies = vf / nunits; |
0136f8f0 AH |
3477 | |
3478 | /* If the function isn't const, only allow it in simd loops where user | |
3479 | has asserted that at least nunits consecutive iterations can be | |
3480 | performed using SIMD instructions. */ | |
3481 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
3482 | && gimple_vuse (stmt)) | |
00426f9a | 3483 | return false; |
0136f8f0 AH |
3484 | |
3485 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3486 | needs to be generated. */ | |
3487 | gcc_assert (ncopies >= 1); | |
3488 | ||
3489 | if (!vec_stmt) /* transformation not required. */ | |
3490 | { | |
6c9e85fb JJ |
3491 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
3492 | for (i = 0; i < nargs; i++) | |
7adb26f2 JJ |
3493 | if ((bestn->simdclone->args[i].arg_type |
3494 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
3495 | || (bestn->simdclone->args[i].arg_type | |
3496 | == SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP)) | |
6c9e85fb | 3497 | { |
17b658af | 3498 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
3499 | + 1); |
3500 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
3501 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
3502 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
3503 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
3504 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
3505 | tree sll = arginfo[i].simd_lane_linear |
3506 | ? boolean_true_node : boolean_false_node; | |
3507 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 3508 | } |
0136f8f0 AH |
3509 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
3510 | if (dump_enabled_p ()) | |
3511 | dump_printf_loc (MSG_NOTE, vect_location, | |
3512 | "=== vectorizable_simd_clone_call ===\n"); | |
3513 | /* vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); */ | |
0136f8f0 AH |
3514 | return true; |
3515 | } | |
3516 | ||
67b8dbac | 3517 | /* Transform. */ |
0136f8f0 AH |
3518 | |
3519 | if (dump_enabled_p ()) | |
3520 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
3521 | ||
3522 | /* Handle def. */ | |
3523 | scalar_dest = gimple_call_lhs (stmt); | |
3524 | vec_dest = NULL_TREE; | |
3525 | rtype = NULL_TREE; | |
3526 | ratype = NULL_TREE; | |
3527 | if (scalar_dest) | |
3528 | { | |
3529 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3530 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
3531 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
3532 | { | |
3533 | ratype = rtype; | |
3534 | rtype = TREE_TYPE (ratype); | |
3535 | } | |
3536 | } | |
3537 | ||
3538 | prev_stmt_info = NULL; | |
3539 | for (j = 0; j < ncopies; ++j) | |
3540 | { | |
3541 | /* Build argument list for the vectorized call. */ | |
3542 | if (j == 0) | |
3543 | vargs.create (nargs); | |
3544 | else | |
3545 | vargs.truncate (0); | |
3546 | ||
3547 | for (i = 0; i < nargs; i++) | |
3548 | { | |
3549 | unsigned int k, l, m, o; | |
3550 | tree atype; | |
3551 | op = gimple_call_arg (stmt, i); | |
3552 | switch (bestn->simdclone->args[i].arg_type) | |
3553 | { | |
3554 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3555 | atype = bestn->simdclone->args[i].vector_type; | |
3556 | o = nunits / TYPE_VECTOR_SUBPARTS (atype); | |
3557 | for (m = j * o; m < (j + 1) * o; m++) | |
3558 | { | |
3559 | if (TYPE_VECTOR_SUBPARTS (atype) | |
3560 | < TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)) | |
3561 | { | |
3562 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); | |
3563 | k = (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
3564 | / TYPE_VECTOR_SUBPARTS (atype)); | |
3565 | gcc_assert ((k & (k - 1)) == 0); | |
3566 | if (m == 0) | |
3567 | vec_oprnd0 | |
81c40241 | 3568 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3569 | else |
3570 | { | |
3571 | vec_oprnd0 = arginfo[i].op; | |
3572 | if ((m & (k - 1)) == 0) | |
3573 | vec_oprnd0 | |
3574 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3575 | vec_oprnd0); | |
3576 | } | |
3577 | arginfo[i].op = vec_oprnd0; | |
3578 | vec_oprnd0 | |
3579 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
92e29a5e | 3580 | bitsize_int (prec), |
0136f8f0 AH |
3581 | bitsize_int ((m & (k - 1)) * prec)); |
3582 | new_stmt | |
b731b390 | 3583 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3584 | vec_oprnd0); |
3585 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3586 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3587 | } | |
3588 | else | |
3589 | { | |
3590 | k = (TYPE_VECTOR_SUBPARTS (atype) | |
3591 | / TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)); | |
3592 | gcc_assert ((k & (k - 1)) == 0); | |
3593 | vec<constructor_elt, va_gc> *ctor_elts; | |
3594 | if (k != 1) | |
3595 | vec_alloc (ctor_elts, k); | |
3596 | else | |
3597 | ctor_elts = NULL; | |
3598 | for (l = 0; l < k; l++) | |
3599 | { | |
3600 | if (m == 0 && l == 0) | |
3601 | vec_oprnd0 | |
81c40241 | 3602 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3603 | else |
3604 | vec_oprnd0 | |
3605 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3606 | arginfo[i].op); | |
3607 | arginfo[i].op = vec_oprnd0; | |
3608 | if (k == 1) | |
3609 | break; | |
3610 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
3611 | vec_oprnd0); | |
3612 | } | |
3613 | if (k == 1) | |
3614 | vargs.safe_push (vec_oprnd0); | |
3615 | else | |
3616 | { | |
3617 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
3618 | new_stmt | |
b731b390 | 3619 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3620 | vec_oprnd0); |
3621 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3622 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3623 | } | |
3624 | } | |
3625 | } | |
3626 | break; | |
3627 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3628 | vargs.safe_push (op); | |
3629 | break; | |
3630 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
7adb26f2 | 3631 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
3632 | if (j == 0) |
3633 | { | |
3634 | gimple_seq stmts; | |
3635 | arginfo[i].op | |
3636 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
3637 | NULL_TREE); | |
3638 | if (stmts != NULL) | |
3639 | { | |
3640 | basic_block new_bb; | |
3641 | edge pe = loop_preheader_edge (loop); | |
3642 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
3643 | gcc_assert (!new_bb); | |
3644 | } | |
17b658af JJ |
3645 | if (arginfo[i].simd_lane_linear) |
3646 | { | |
3647 | vargs.safe_push (arginfo[i].op); | |
3648 | break; | |
3649 | } | |
b731b390 | 3650 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 3651 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
0136f8f0 | 3652 | set_vinfo_for_stmt (new_phi, |
310213d4 | 3653 | new_stmt_vec_info (new_phi, loop_vinfo)); |
0136f8f0 AH |
3654 | add_phi_arg (new_phi, arginfo[i].op, |
3655 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
3656 | enum tree_code code | |
3657 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3658 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3659 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3660 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3661 | widest_int cst |
3662 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3663 | ncopies * nunits); | |
3664 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3665 | tree phi_arg = copy_ssa_name (op); |
0d0e4a03 JJ |
3666 | new_stmt |
3667 | = gimple_build_assign (phi_arg, code, phi_res, tcst); | |
0136f8f0 AH |
3668 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
3669 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
3670 | set_vinfo_for_stmt (new_stmt, | |
310213d4 | 3671 | new_stmt_vec_info (new_stmt, loop_vinfo)); |
0136f8f0 AH |
3672 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
3673 | UNKNOWN_LOCATION); | |
3674 | arginfo[i].op = phi_res; | |
3675 | vargs.safe_push (phi_res); | |
3676 | } | |
3677 | else | |
3678 | { | |
3679 | enum tree_code code | |
3680 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3681 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3682 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3683 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3684 | widest_int cst |
3685 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3686 | j * nunits); | |
3687 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3688 | new_temp = make_ssa_name (TREE_TYPE (op)); |
0d0e4a03 JJ |
3689 | new_stmt = gimple_build_assign (new_temp, code, |
3690 | arginfo[i].op, tcst); | |
0136f8f0 AH |
3691 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3692 | vargs.safe_push (new_temp); | |
3693 | } | |
3694 | break; | |
7adb26f2 JJ |
3695 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
3696 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
0136f8f0 | 3697 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
e01d41e5 JJ |
3698 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3699 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3700 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3701 | default: |
3702 | gcc_unreachable (); | |
3703 | } | |
3704 | } | |
3705 | ||
3706 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
3707 | if (vec_dest) | |
3708 | { | |
3709 | gcc_assert (ratype || TYPE_VECTOR_SUBPARTS (rtype) == nunits); | |
3710 | if (ratype) | |
b731b390 | 3711 | new_temp = create_tmp_var (ratype); |
0136f8f0 AH |
3712 | else if (TYPE_VECTOR_SUBPARTS (vectype) |
3713 | == TYPE_VECTOR_SUBPARTS (rtype)) | |
3714 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3715 | else | |
3716 | new_temp = make_ssa_name (rtype, new_stmt); | |
3717 | gimple_call_set_lhs (new_stmt, new_temp); | |
3718 | } | |
3719 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3720 | ||
3721 | if (vec_dest) | |
3722 | { | |
3723 | if (TYPE_VECTOR_SUBPARTS (vectype) < nunits) | |
3724 | { | |
3725 | unsigned int k, l; | |
3726 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); | |
3727 | k = nunits / TYPE_VECTOR_SUBPARTS (vectype); | |
3728 | gcc_assert ((k & (k - 1)) == 0); | |
3729 | for (l = 0; l < k; l++) | |
3730 | { | |
3731 | tree t; | |
3732 | if (ratype) | |
3733 | { | |
3734 | t = build_fold_addr_expr (new_temp); | |
3735 | t = build2 (MEM_REF, vectype, t, | |
3736 | build_int_cst (TREE_TYPE (t), | |
3737 | l * prec / BITS_PER_UNIT)); | |
3738 | } | |
3739 | else | |
3740 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
92e29a5e | 3741 | bitsize_int (prec), bitsize_int (l * prec)); |
0136f8f0 | 3742 | new_stmt |
b731b390 | 3743 | = gimple_build_assign (make_ssa_name (vectype), t); |
0136f8f0 AH |
3744 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3745 | if (j == 0 && l == 0) | |
3746 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3747 | else | |
3748 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3749 | ||
3750 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3751 | } | |
3752 | ||
3753 | if (ratype) | |
3754 | { | |
3755 | tree clobber = build_constructor (ratype, NULL); | |
3756 | TREE_THIS_VOLATILE (clobber) = 1; | |
3757 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3758 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3759 | } | |
3760 | continue; | |
3761 | } | |
3762 | else if (TYPE_VECTOR_SUBPARTS (vectype) > nunits) | |
3763 | { | |
3764 | unsigned int k = (TYPE_VECTOR_SUBPARTS (vectype) | |
3765 | / TYPE_VECTOR_SUBPARTS (rtype)); | |
3766 | gcc_assert ((k & (k - 1)) == 0); | |
3767 | if ((j & (k - 1)) == 0) | |
3768 | vec_alloc (ret_ctor_elts, k); | |
3769 | if (ratype) | |
3770 | { | |
3771 | unsigned int m, o = nunits / TYPE_VECTOR_SUBPARTS (rtype); | |
3772 | for (m = 0; m < o; m++) | |
3773 | { | |
3774 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
3775 | size_int (m), NULL_TREE, NULL_TREE); | |
3776 | new_stmt | |
b731b390 | 3777 | = gimple_build_assign (make_ssa_name (rtype), tem); |
0136f8f0 AH |
3778 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3779 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
3780 | gimple_assign_lhs (new_stmt)); | |
3781 | } | |
3782 | tree clobber = build_constructor (ratype, NULL); | |
3783 | TREE_THIS_VOLATILE (clobber) = 1; | |
3784 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3785 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3786 | } | |
3787 | else | |
3788 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
3789 | if ((j & (k - 1)) != k - 1) | |
3790 | continue; | |
3791 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
3792 | new_stmt | |
b731b390 | 3793 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
0136f8f0 AH |
3794 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3795 | ||
3796 | if ((unsigned) j == k - 1) | |
3797 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3798 | else | |
3799 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3800 | ||
3801 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3802 | continue; | |
3803 | } | |
3804 | else if (ratype) | |
3805 | { | |
3806 | tree t = build_fold_addr_expr (new_temp); | |
3807 | t = build2 (MEM_REF, vectype, t, | |
3808 | build_int_cst (TREE_TYPE (t), 0)); | |
3809 | new_stmt | |
b731b390 | 3810 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
0136f8f0 AH |
3811 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3812 | tree clobber = build_constructor (ratype, NULL); | |
3813 | TREE_THIS_VOLATILE (clobber) = 1; | |
3814 | vect_finish_stmt_generation (stmt, | |
3815 | gimple_build_assign (new_temp, | |
3816 | clobber), gsi); | |
3817 | } | |
3818 | } | |
3819 | ||
3820 | if (j == 0) | |
3821 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3822 | else | |
3823 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3824 | ||
3825 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3826 | } | |
3827 | ||
3828 | vargs.release (); | |
3829 | ||
3830 | /* The call in STMT might prevent it from being removed in dce. | |
3831 | We however cannot remove it here, due to the way the ssa name | |
3832 | it defines is mapped to the new definition. So just replace | |
3833 | rhs of the statement with something harmless. */ | |
3834 | ||
3835 | if (slp_node) | |
3836 | return true; | |
3837 | ||
3838 | if (scalar_dest) | |
3839 | { | |
3840 | type = TREE_TYPE (scalar_dest); | |
3841 | if (is_pattern_stmt_p (stmt_info)) | |
3842 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3843 | else | |
3844 | lhs = gimple_call_lhs (stmt); | |
3845 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
3846 | } | |
3847 | else | |
3848 | new_stmt = gimple_build_nop (); | |
3849 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
3850 | set_vinfo_for_stmt (stmt, NULL); | |
3851 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 3852 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
3853 | unlink_stmt_vdef (stmt); |
3854 | ||
3855 | return true; | |
3856 | } | |
3857 | ||
3858 | ||
ebfd146a IR |
3859 | /* Function vect_gen_widened_results_half |
3860 | ||
3861 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 3862 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 3863 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
3864 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
3865 | needs to be created (DECL is a function-decl of a target-builtin). | |
3866 | STMT is the original scalar stmt that we are vectorizing. */ | |
3867 | ||
355fe088 | 3868 | static gimple * |
ebfd146a IR |
3869 | vect_gen_widened_results_half (enum tree_code code, |
3870 | tree decl, | |
3871 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
3872 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 3873 | gimple *stmt) |
b8698a0f | 3874 | { |
355fe088 | 3875 | gimple *new_stmt; |
b8698a0f L |
3876 | tree new_temp; |
3877 | ||
3878 | /* Generate half of the widened result: */ | |
3879 | if (code == CALL_EXPR) | |
3880 | { | |
3881 | /* Target specific support */ | |
ebfd146a IR |
3882 | if (op_type == binary_op) |
3883 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
3884 | else | |
3885 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
3886 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3887 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
3888 | } |
3889 | else | |
ebfd146a | 3890 | { |
b8698a0f L |
3891 | /* Generic support */ |
3892 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
3893 | if (op_type != binary_op) |
3894 | vec_oprnd1 = NULL; | |
0d0e4a03 | 3895 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
3896 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3897 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 3898 | } |
ebfd146a IR |
3899 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3900 | ||
ebfd146a IR |
3901 | return new_stmt; |
3902 | } | |
3903 | ||
4a00c761 JJ |
3904 | |
3905 | /* Get vectorized definitions for loop-based vectorization. For the first | |
3906 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
3907 | scalar operand), and for the rest we get a copy with | |
3908 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
3909 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
3910 | The vectors are collected into VEC_OPRNDS. */ | |
3911 | ||
3912 | static void | |
355fe088 | 3913 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 3914 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
3915 | { |
3916 | tree vec_oprnd; | |
3917 | ||
3918 | /* Get first vector operand. */ | |
3919 | /* All the vector operands except the very first one (that is scalar oprnd) | |
3920 | are stmt copies. */ | |
3921 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 3922 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
3923 | else |
3924 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
3925 | ||
9771b263 | 3926 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3927 | |
3928 | /* Get second vector operand. */ | |
3929 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 3930 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3931 | |
3932 | *oprnd = vec_oprnd; | |
3933 | ||
3934 | /* For conversion in multiple steps, continue to get operands | |
3935 | recursively. */ | |
3936 | if (multi_step_cvt) | |
3937 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
3938 | } | |
3939 | ||
3940 | ||
3941 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
3942 | For multi-step conversions store the resulting vectors and call the function | |
3943 | recursively. */ | |
3944 | ||
3945 | static void | |
9771b263 | 3946 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 3947 | int multi_step_cvt, gimple *stmt, |
9771b263 | 3948 | vec<tree> vec_dsts, |
4a00c761 JJ |
3949 | gimple_stmt_iterator *gsi, |
3950 | slp_tree slp_node, enum tree_code code, | |
3951 | stmt_vec_info *prev_stmt_info) | |
3952 | { | |
3953 | unsigned int i; | |
3954 | tree vop0, vop1, new_tmp, vec_dest; | |
355fe088 | 3955 | gimple *new_stmt; |
4a00c761 JJ |
3956 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
3957 | ||
9771b263 | 3958 | vec_dest = vec_dsts.pop (); |
4a00c761 | 3959 | |
9771b263 | 3960 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
3961 | { |
3962 | /* Create demotion operation. */ | |
9771b263 DN |
3963 | vop0 = (*vec_oprnds)[i]; |
3964 | vop1 = (*vec_oprnds)[i + 1]; | |
0d0e4a03 | 3965 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
3966 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
3967 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
3968 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3969 | ||
3970 | if (multi_step_cvt) | |
3971 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 3972 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
3973 | else |
3974 | { | |
3975 | /* This is the last step of the conversion sequence. Store the | |
3976 | vectors in SLP_NODE or in vector info of the scalar statement | |
3977 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
3978 | if (slp_node) | |
9771b263 | 3979 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 3980 | else |
c689ce1e RB |
3981 | { |
3982 | if (!*prev_stmt_info) | |
3983 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3984 | else | |
3985 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
4a00c761 | 3986 | |
c689ce1e RB |
3987 | *prev_stmt_info = vinfo_for_stmt (new_stmt); |
3988 | } | |
4a00c761 JJ |
3989 | } |
3990 | } | |
3991 | ||
3992 | /* For multi-step demotion operations we first generate demotion operations | |
3993 | from the source type to the intermediate types, and then combine the | |
3994 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
3995 | type. */ | |
3996 | if (multi_step_cvt) | |
3997 | { | |
3998 | /* At each level of recursion we have half of the operands we had at the | |
3999 | previous level. */ | |
9771b263 | 4000 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
4001 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
4002 | stmt, vec_dsts, gsi, slp_node, | |
4003 | VEC_PACK_TRUNC_EXPR, | |
4004 | prev_stmt_info); | |
4005 | } | |
4006 | ||
9771b263 | 4007 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4008 | } |
4009 | ||
4010 | ||
4011 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
4012 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
4013 | the resulting vectors and call the function recursively. */ | |
4014 | ||
4015 | static void | |
9771b263 DN |
4016 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
4017 | vec<tree> *vec_oprnds1, | |
355fe088 | 4018 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
4019 | gimple_stmt_iterator *gsi, |
4020 | enum tree_code code1, | |
4021 | enum tree_code code2, tree decl1, | |
4022 | tree decl2, int op_type) | |
4023 | { | |
4024 | int i; | |
4025 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 4026 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 4027 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 4028 | |
9771b263 DN |
4029 | vec_tmp.create (vec_oprnds0->length () * 2); |
4030 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
4031 | { |
4032 | if (op_type == binary_op) | |
9771b263 | 4033 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
4034 | else |
4035 | vop1 = NULL_TREE; | |
4036 | ||
4037 | /* Generate the two halves of promotion operation. */ | |
4038 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
4039 | op_type, vec_dest, gsi, stmt); | |
4040 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
4041 | op_type, vec_dest, gsi, stmt); | |
4042 | if (is_gimple_call (new_stmt1)) | |
4043 | { | |
4044 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
4045 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
4046 | } | |
4047 | else | |
4048 | { | |
4049 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
4050 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
4051 | } | |
4052 | ||
4053 | /* Store the results for the next step. */ | |
9771b263 DN |
4054 | vec_tmp.quick_push (new_tmp1); |
4055 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
4056 | } |
4057 | ||
689eaba3 | 4058 | vec_oprnds0->release (); |
4a00c761 JJ |
4059 | *vec_oprnds0 = vec_tmp; |
4060 | } | |
4061 | ||
4062 | ||
b8698a0f L |
4063 | /* Check if STMT performs a conversion operation, that can be vectorized. |
4064 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 4065 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
4066 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
4067 | ||
4068 | static bool | |
355fe088 TS |
4069 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
4070 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4071 | { |
4072 | tree vec_dest; | |
4073 | tree scalar_dest; | |
4a00c761 | 4074 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
4075 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
4076 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4077 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4078 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 4079 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
4080 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
4081 | tree new_temp; | |
355fe088 | 4082 | gimple *def_stmt; |
ebfd146a | 4083 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 4084 | int ndts = 2; |
355fe088 | 4085 | gimple *new_stmt = NULL; |
ebfd146a IR |
4086 | stmt_vec_info prev_stmt_info; |
4087 | int nunits_in; | |
4088 | int nunits_out; | |
4089 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
4090 | int ncopies, i, j; |
4091 | tree lhs_type, rhs_type; | |
ebfd146a | 4092 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
4093 | vec<tree> vec_oprnds0 = vNULL; |
4094 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 4095 | tree vop0; |
4a00c761 | 4096 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4097 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 4098 | int multi_step_cvt = 0; |
6e1aa848 | 4099 | vec<tree> interm_types = vNULL; |
4a00c761 JJ |
4100 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
4101 | int op_type; | |
4a00c761 | 4102 | unsigned short fltsz; |
ebfd146a IR |
4103 | |
4104 | /* Is STMT a vectorizable conversion? */ | |
4105 | ||
4a00c761 | 4106 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4107 | return false; |
4108 | ||
66c16fd9 RB |
4109 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4110 | && ! vec_stmt) | |
ebfd146a IR |
4111 | return false; |
4112 | ||
4113 | if (!is_gimple_assign (stmt)) | |
4114 | return false; | |
4115 | ||
4116 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4117 | return false; | |
4118 | ||
4119 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
4120 | if (!CONVERT_EXPR_CODE_P (code) |
4121 | && code != FIX_TRUNC_EXPR | |
4122 | && code != FLOAT_EXPR | |
4123 | && code != WIDEN_MULT_EXPR | |
4124 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
4125 | return false; |
4126 | ||
4a00c761 JJ |
4127 | op_type = TREE_CODE_LENGTH (code); |
4128 | ||
ebfd146a | 4129 | /* Check types of lhs and rhs. */ |
b690cc0f | 4130 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 4131 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
4132 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
4133 | ||
ebfd146a IR |
4134 | op0 = gimple_assign_rhs1 (stmt); |
4135 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
4136 | |
4137 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4138 | && !((INTEGRAL_TYPE_P (lhs_type) | |
4139 | && INTEGRAL_TYPE_P (rhs_type)) | |
4140 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
4141 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
4142 | return false; | |
4143 | ||
e6f5c25d IE |
4144 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4145 | && ((INTEGRAL_TYPE_P (lhs_type) | |
2be65d9e | 4146 | && !type_has_mode_precision_p (lhs_type)) |
e6f5c25d | 4147 | || (INTEGRAL_TYPE_P (rhs_type) |
2be65d9e | 4148 | && !type_has_mode_precision_p (rhs_type)))) |
4a00c761 | 4149 | { |
73fbfcad | 4150 | if (dump_enabled_p ()) |
78c60e3d | 4151 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
4152 | "type conversion to/from bit-precision unsupported." |
4153 | "\n"); | |
4a00c761 JJ |
4154 | return false; |
4155 | } | |
4156 | ||
b690cc0f | 4157 | /* Check the operands of the operation. */ |
81c40241 | 4158 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype_in)) |
b690cc0f | 4159 | { |
73fbfcad | 4160 | if (dump_enabled_p ()) |
78c60e3d | 4161 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4162 | "use not simple.\n"); |
b690cc0f RG |
4163 | return false; |
4164 | } | |
4a00c761 JJ |
4165 | if (op_type == binary_op) |
4166 | { | |
4167 | bool ok; | |
4168 | ||
4169 | op1 = gimple_assign_rhs2 (stmt); | |
4170 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
4171 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
4172 | OP1. */ | |
4173 | if (CONSTANT_CLASS_P (op0)) | |
81c40241 | 4174 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &vectype_in); |
4a00c761 | 4175 | else |
81c40241 | 4176 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1]); |
4a00c761 JJ |
4177 | |
4178 | if (!ok) | |
4179 | { | |
73fbfcad | 4180 | if (dump_enabled_p ()) |
78c60e3d | 4181 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4182 | "use not simple.\n"); |
4a00c761 JJ |
4183 | return false; |
4184 | } | |
4185 | } | |
4186 | ||
b690cc0f RG |
4187 | /* If op0 is an external or constant defs use a vector type of |
4188 | the same size as the output vector type. */ | |
ebfd146a | 4189 | if (!vectype_in) |
b690cc0f | 4190 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
4191 | if (vec_stmt) |
4192 | gcc_assert (vectype_in); | |
4193 | if (!vectype_in) | |
4194 | { | |
73fbfcad | 4195 | if (dump_enabled_p ()) |
4a00c761 | 4196 | { |
78c60e3d SS |
4197 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4198 | "no vectype for scalar type "); | |
4199 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 4200 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 4201 | } |
7d8930a0 IR |
4202 | |
4203 | return false; | |
4204 | } | |
ebfd146a | 4205 | |
e6f5c25d IE |
4206 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4207 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
4208 | { | |
4209 | if (dump_enabled_p ()) | |
4210 | { | |
4211 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4212 | "can't convert between boolean and non " | |
4213 | "boolean vectors"); | |
4214 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
4215 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
4216 | } | |
4217 | ||
4218 | return false; | |
4219 | } | |
4220 | ||
b690cc0f RG |
4221 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
4222 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 4223 | if (nunits_in < nunits_out) |
ebfd146a IR |
4224 | modifier = NARROW; |
4225 | else if (nunits_out == nunits_in) | |
4226 | modifier = NONE; | |
ebfd146a | 4227 | else |
4a00c761 | 4228 | modifier = WIDEN; |
ebfd146a | 4229 | |
ff802fa1 IR |
4230 | /* Multiple types in SLP are handled by creating the appropriate number of |
4231 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4232 | case of SLP. */ | |
fce57248 | 4233 | if (slp_node) |
ebfd146a | 4234 | ncopies = 1; |
4a00c761 | 4235 | else if (modifier == NARROW) |
e8f142e2 | 4236 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
4a00c761 | 4237 | else |
e8f142e2 | 4238 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
b8698a0f | 4239 | |
ebfd146a IR |
4240 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
4241 | needs to be generated. */ | |
4242 | gcc_assert (ncopies >= 1); | |
4243 | ||
16d22000 RS |
4244 | bool found_mode = false; |
4245 | scalar_mode lhs_mode = SCALAR_TYPE_MODE (lhs_type); | |
4246 | scalar_mode rhs_mode = SCALAR_TYPE_MODE (rhs_type); | |
4247 | opt_scalar_mode rhs_mode_iter; | |
b397965c | 4248 | |
ebfd146a | 4249 | /* Supportable by target? */ |
4a00c761 | 4250 | switch (modifier) |
ebfd146a | 4251 | { |
4a00c761 JJ |
4252 | case NONE: |
4253 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4254 | return false; | |
4255 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
4256 | &decl1, &code1)) | |
4257 | break; | |
4258 | /* FALLTHRU */ | |
4259 | unsupported: | |
73fbfcad | 4260 | if (dump_enabled_p ()) |
78c60e3d | 4261 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4262 | "conversion not supported by target.\n"); |
ebfd146a | 4263 | return false; |
ebfd146a | 4264 | |
4a00c761 JJ |
4265 | case WIDEN: |
4266 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
4267 | &code1, &code2, &multi_step_cvt, |
4268 | &interm_types)) | |
4a00c761 JJ |
4269 | { |
4270 | /* Binary widening operation can only be supported directly by the | |
4271 | architecture. */ | |
4272 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
4273 | break; | |
4274 | } | |
4275 | ||
4276 | if (code != FLOAT_EXPR | |
b397965c | 4277 | || GET_MODE_SIZE (lhs_mode) <= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4278 | goto unsupported; |
4279 | ||
b397965c | 4280 | fltsz = GET_MODE_SIZE (lhs_mode); |
16d22000 | 4281 | FOR_EACH_2XWIDER_MODE (rhs_mode_iter, rhs_mode) |
4a00c761 | 4282 | { |
16d22000 | 4283 | rhs_mode = rhs_mode_iter.require (); |
c94843d2 RS |
4284 | if (GET_MODE_SIZE (rhs_mode) > fltsz) |
4285 | break; | |
4286 | ||
4a00c761 JJ |
4287 | cvt_type |
4288 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4289 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4290 | if (cvt_type == NULL_TREE) | |
4291 | goto unsupported; | |
4292 | ||
4293 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4294 | { | |
4295 | if (!supportable_convert_operation (code, vectype_out, | |
4296 | cvt_type, &decl1, &codecvt1)) | |
4297 | goto unsupported; | |
4298 | } | |
4299 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
4300 | cvt_type, &codecvt1, |
4301 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
4302 | &interm_types)) |
4303 | continue; | |
4304 | else | |
4305 | gcc_assert (multi_step_cvt == 0); | |
4306 | ||
4307 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
4308 | vectype_in, &code1, &code2, |
4309 | &multi_step_cvt, &interm_types)) | |
16d22000 RS |
4310 | { |
4311 | found_mode = true; | |
4312 | break; | |
4313 | } | |
4a00c761 JJ |
4314 | } |
4315 | ||
16d22000 | 4316 | if (!found_mode) |
4a00c761 JJ |
4317 | goto unsupported; |
4318 | ||
4319 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4320 | codecvt2 = ERROR_MARK; | |
4321 | else | |
4322 | { | |
4323 | multi_step_cvt++; | |
9771b263 | 4324 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
4325 | cvt_type = NULL_TREE; |
4326 | } | |
4327 | break; | |
4328 | ||
4329 | case NARROW: | |
4330 | gcc_assert (op_type == unary_op); | |
4331 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
4332 | &code1, &multi_step_cvt, | |
4333 | &interm_types)) | |
4334 | break; | |
4335 | ||
4336 | if (code != FIX_TRUNC_EXPR | |
b397965c | 4337 | || GET_MODE_SIZE (lhs_mode) >= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4338 | goto unsupported; |
4339 | ||
4a00c761 JJ |
4340 | cvt_type |
4341 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4342 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4343 | if (cvt_type == NULL_TREE) | |
4344 | goto unsupported; | |
4345 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
4346 | &decl1, &codecvt1)) | |
4347 | goto unsupported; | |
4348 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
4349 | &code1, &multi_step_cvt, | |
4350 | &interm_types)) | |
4351 | break; | |
4352 | goto unsupported; | |
4353 | ||
4354 | default: | |
4355 | gcc_unreachable (); | |
ebfd146a IR |
4356 | } |
4357 | ||
4358 | if (!vec_stmt) /* transformation not required. */ | |
4359 | { | |
73fbfcad | 4360 | if (dump_enabled_p ()) |
78c60e3d | 4361 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4362 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 4363 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
4364 | { |
4365 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
4fc5ebf1 | 4366 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
8bd37302 | 4367 | } |
4a00c761 JJ |
4368 | else if (modifier == NARROW) |
4369 | { | |
4370 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 4371 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
4372 | } |
4373 | else | |
4374 | { | |
4375 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 4376 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 4377 | } |
9771b263 | 4378 | interm_types.release (); |
ebfd146a IR |
4379 | return true; |
4380 | } | |
4381 | ||
67b8dbac | 4382 | /* Transform. */ |
73fbfcad | 4383 | if (dump_enabled_p ()) |
78c60e3d | 4384 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4385 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 4386 | |
4a00c761 JJ |
4387 | if (op_type == binary_op) |
4388 | { | |
4389 | if (CONSTANT_CLASS_P (op0)) | |
4390 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
4391 | else if (CONSTANT_CLASS_P (op1)) | |
4392 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
4393 | } | |
4394 | ||
4395 | /* In case of multi-step conversion, we first generate conversion operations | |
4396 | to the intermediate types, and then from that types to the final one. | |
4397 | We create vector destinations for the intermediate type (TYPES) received | |
4398 | from supportable_*_operation, and store them in the correct order | |
4399 | for future use in vect_create_vectorized_*_stmts (). */ | |
8c681247 | 4400 | auto_vec<tree> vec_dsts (multi_step_cvt + 1); |
82294ec1 JJ |
4401 | vec_dest = vect_create_destination_var (scalar_dest, |
4402 | (cvt_type && modifier == WIDEN) | |
4403 | ? cvt_type : vectype_out); | |
9771b263 | 4404 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4405 | |
4406 | if (multi_step_cvt) | |
4407 | { | |
9771b263 DN |
4408 | for (i = interm_types.length () - 1; |
4409 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
4410 | { |
4411 | vec_dest = vect_create_destination_var (scalar_dest, | |
4412 | intermediate_type); | |
9771b263 | 4413 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4414 | } |
4415 | } | |
ebfd146a | 4416 | |
4a00c761 | 4417 | if (cvt_type) |
82294ec1 JJ |
4418 | vec_dest = vect_create_destination_var (scalar_dest, |
4419 | modifier == WIDEN | |
4420 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
4421 | |
4422 | if (!slp_node) | |
4423 | { | |
30862efc | 4424 | if (modifier == WIDEN) |
4a00c761 | 4425 | { |
c3284718 | 4426 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 4427 | if (op_type == binary_op) |
9771b263 | 4428 | vec_oprnds1.create (1); |
4a00c761 | 4429 | } |
30862efc | 4430 | else if (modifier == NARROW) |
9771b263 DN |
4431 | vec_oprnds0.create ( |
4432 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
4433 | } |
4434 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 4435 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 4436 | |
4a00c761 | 4437 | last_oprnd = op0; |
ebfd146a IR |
4438 | prev_stmt_info = NULL; |
4439 | switch (modifier) | |
4440 | { | |
4441 | case NONE: | |
4442 | for (j = 0; j < ncopies; j++) | |
4443 | { | |
ebfd146a | 4444 | if (j == 0) |
306b0c92 | 4445 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node); |
ebfd146a IR |
4446 | else |
4447 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
4448 | ||
9771b263 | 4449 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4450 | { |
4451 | /* Arguments are ready, create the new vector stmt. */ | |
4452 | if (code1 == CALL_EXPR) | |
4453 | { | |
4454 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4455 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4456 | gimple_call_set_lhs (new_stmt, new_temp); | |
4457 | } | |
4458 | else | |
4459 | { | |
4460 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
0d0e4a03 | 4461 | new_stmt = gimple_build_assign (vec_dest, code1, vop0); |
4a00c761 JJ |
4462 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4463 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4464 | } | |
4465 | ||
4466 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4467 | if (slp_node) | |
9771b263 | 4468 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
225ce44b RB |
4469 | else |
4470 | { | |
4471 | if (!prev_stmt_info) | |
4472 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4473 | else | |
4474 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4475 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4476 | } | |
4a00c761 | 4477 | } |
ebfd146a IR |
4478 | } |
4479 | break; | |
4480 | ||
4481 | case WIDEN: | |
4482 | /* In case the vectorization factor (VF) is bigger than the number | |
4483 | of elements that we can fit in a vectype (nunits), we have to | |
4484 | generate more than one vector stmt - i.e - we need to "unroll" | |
4485 | the vector stmt by a factor VF/nunits. */ | |
4486 | for (j = 0; j < ncopies; j++) | |
4487 | { | |
4a00c761 | 4488 | /* Handle uses. */ |
ebfd146a | 4489 | if (j == 0) |
4a00c761 JJ |
4490 | { |
4491 | if (slp_node) | |
4492 | { | |
4493 | if (code == WIDEN_LSHIFT_EXPR) | |
4494 | { | |
4495 | unsigned int k; | |
ebfd146a | 4496 | |
4a00c761 JJ |
4497 | vec_oprnd1 = op1; |
4498 | /* Store vec_oprnd1 for every vector stmt to be created | |
4499 | for SLP_NODE. We check during the analysis that all | |
4500 | the shift arguments are the same. */ | |
4501 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4502 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4503 | |
4504 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 4505 | slp_node); |
4a00c761 JJ |
4506 | } |
4507 | else | |
4508 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
306b0c92 | 4509 | &vec_oprnds1, slp_node); |
4a00c761 JJ |
4510 | } |
4511 | else | |
4512 | { | |
81c40241 | 4513 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 4514 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
4515 | if (op_type == binary_op) |
4516 | { | |
4517 | if (code == WIDEN_LSHIFT_EXPR) | |
4518 | vec_oprnd1 = op1; | |
4519 | else | |
81c40241 | 4520 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 4521 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4522 | } |
4523 | } | |
4524 | } | |
ebfd146a | 4525 | else |
4a00c761 JJ |
4526 | { |
4527 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
4528 | vec_oprnds0.truncate (0); |
4529 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
4530 | if (op_type == binary_op) |
4531 | { | |
4532 | if (code == WIDEN_LSHIFT_EXPR) | |
4533 | vec_oprnd1 = op1; | |
4534 | else | |
4535 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
4536 | vec_oprnd1); | |
9771b263 DN |
4537 | vec_oprnds1.truncate (0); |
4538 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
4539 | } |
4540 | } | |
ebfd146a | 4541 | |
4a00c761 JJ |
4542 | /* Arguments are ready. Create the new vector stmts. */ |
4543 | for (i = multi_step_cvt; i >= 0; i--) | |
4544 | { | |
9771b263 | 4545 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
4546 | enum tree_code c1 = code1, c2 = code2; |
4547 | if (i == 0 && codecvt2 != ERROR_MARK) | |
4548 | { | |
4549 | c1 = codecvt1; | |
4550 | c2 = codecvt2; | |
4551 | } | |
4552 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
4553 | &vec_oprnds1, | |
4554 | stmt, this_dest, gsi, | |
4555 | c1, c2, decl1, decl2, | |
4556 | op_type); | |
4557 | } | |
4558 | ||
9771b263 | 4559 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4560 | { |
4561 | if (cvt_type) | |
4562 | { | |
4563 | if (codecvt1 == CALL_EXPR) | |
4564 | { | |
4565 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4566 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4567 | gimple_call_set_lhs (new_stmt, new_temp); | |
4568 | } | |
4569 | else | |
4570 | { | |
4571 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4572 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4573 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4574 | vop0); | |
4a00c761 JJ |
4575 | } |
4576 | ||
4577 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4578 | } | |
4579 | else | |
4580 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
4581 | ||
4582 | if (slp_node) | |
9771b263 | 4583 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4584 | else |
c689ce1e RB |
4585 | { |
4586 | if (!prev_stmt_info) | |
4587 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4588 | else | |
4589 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4590 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4591 | } | |
4a00c761 | 4592 | } |
ebfd146a | 4593 | } |
4a00c761 JJ |
4594 | |
4595 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
4596 | break; |
4597 | ||
4598 | case NARROW: | |
4599 | /* In case the vectorization factor (VF) is bigger than the number | |
4600 | of elements that we can fit in a vectype (nunits), we have to | |
4601 | generate more than one vector stmt - i.e - we need to "unroll" | |
4602 | the vector stmt by a factor VF/nunits. */ | |
4603 | for (j = 0; j < ncopies; j++) | |
4604 | { | |
4605 | /* Handle uses. */ | |
4a00c761 JJ |
4606 | if (slp_node) |
4607 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 4608 | slp_node); |
ebfd146a IR |
4609 | else |
4610 | { | |
9771b263 | 4611 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
4612 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
4613 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
4614 | } |
4615 | ||
4a00c761 JJ |
4616 | /* Arguments are ready. Create the new vector stmts. */ |
4617 | if (cvt_type) | |
9771b263 | 4618 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4619 | { |
4620 | if (codecvt1 == CALL_EXPR) | |
4621 | { | |
4622 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4623 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4624 | gimple_call_set_lhs (new_stmt, new_temp); | |
4625 | } | |
4626 | else | |
4627 | { | |
4628 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4629 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4630 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4631 | vop0); | |
4a00c761 | 4632 | } |
ebfd146a | 4633 | |
4a00c761 | 4634 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 4635 | vec_oprnds0[i] = new_temp; |
4a00c761 | 4636 | } |
ebfd146a | 4637 | |
4a00c761 JJ |
4638 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
4639 | stmt, vec_dsts, gsi, | |
4640 | slp_node, code1, | |
4641 | &prev_stmt_info); | |
ebfd146a IR |
4642 | } |
4643 | ||
4644 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 4645 | break; |
ebfd146a IR |
4646 | } |
4647 | ||
9771b263 DN |
4648 | vec_oprnds0.release (); |
4649 | vec_oprnds1.release (); | |
9771b263 | 4650 | interm_types.release (); |
ebfd146a IR |
4651 | |
4652 | return true; | |
4653 | } | |
ff802fa1 IR |
4654 | |
4655 | ||
ebfd146a IR |
4656 | /* Function vectorizable_assignment. |
4657 | ||
b8698a0f L |
4658 | Check if STMT performs an assignment (copy) that can be vectorized. |
4659 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4660 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4661 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4662 | ||
4663 | static bool | |
355fe088 TS |
4664 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
4665 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4666 | { |
4667 | tree vec_dest; | |
4668 | tree scalar_dest; | |
4669 | tree op; | |
4670 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
4671 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4672 | tree new_temp; | |
355fe088 | 4673 | gimple *def_stmt; |
4fc5ebf1 JG |
4674 | enum vect_def_type dt[1] = {vect_unknown_def_type}; |
4675 | int ndts = 1; | |
ebfd146a | 4676 | int ncopies; |
f18b55bd | 4677 | int i, j; |
6e1aa848 | 4678 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 4679 | tree vop; |
a70d6342 | 4680 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4681 | vec_info *vinfo = stmt_info->vinfo; |
355fe088 | 4682 | gimple *new_stmt = NULL; |
f18b55bd | 4683 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
4684 | enum tree_code code; |
4685 | tree vectype_in; | |
ebfd146a | 4686 | |
a70d6342 | 4687 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4688 | return false; |
4689 | ||
66c16fd9 RB |
4690 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4691 | && ! vec_stmt) | |
ebfd146a IR |
4692 | return false; |
4693 | ||
4694 | /* Is vectorizable assignment? */ | |
4695 | if (!is_gimple_assign (stmt)) | |
4696 | return false; | |
4697 | ||
4698 | scalar_dest = gimple_assign_lhs (stmt); | |
4699 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4700 | return false; | |
4701 | ||
fde9c428 | 4702 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 4703 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
4704 | || code == PAREN_EXPR |
4705 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
4706 | op = gimple_assign_rhs1 (stmt); |
4707 | else | |
4708 | return false; | |
4709 | ||
7b7ec6c5 RG |
4710 | if (code == VIEW_CONVERT_EXPR) |
4711 | op = TREE_OPERAND (op, 0); | |
4712 | ||
465c8c19 JJ |
4713 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
4714 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4715 | ||
4716 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4717 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4718 | case of SLP. */ | |
fce57248 | 4719 | if (slp_node) |
465c8c19 JJ |
4720 | ncopies = 1; |
4721 | else | |
e8f142e2 | 4722 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
4723 | |
4724 | gcc_assert (ncopies >= 1); | |
4725 | ||
81c40241 | 4726 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[0], &vectype_in)) |
ebfd146a | 4727 | { |
73fbfcad | 4728 | if (dump_enabled_p ()) |
78c60e3d | 4729 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4730 | "use not simple.\n"); |
ebfd146a IR |
4731 | return false; |
4732 | } | |
4733 | ||
fde9c428 RG |
4734 | /* We can handle NOP_EXPR conversions that do not change the number |
4735 | of elements or the vector size. */ | |
7b7ec6c5 RG |
4736 | if ((CONVERT_EXPR_CODE_P (code) |
4737 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
4738 | && (!vectype_in |
4739 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
4740 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
4741 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
4742 | return false; | |
4743 | ||
7b7b1813 RG |
4744 | /* We do not handle bit-precision changes. */ |
4745 | if ((CONVERT_EXPR_CODE_P (code) | |
4746 | || code == VIEW_CONVERT_EXPR) | |
4747 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2be65d9e RS |
4748 | && (!type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
4749 | || !type_has_mode_precision_p (TREE_TYPE (op))) | |
7b7b1813 RG |
4750 | /* But a conversion that does not change the bit-pattern is ok. */ |
4751 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4752 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2dab46d5 IE |
4753 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
4754 | /* Conversion between boolean types of different sizes is | |
4755 | a simple assignment in case their vectypes are same | |
4756 | boolean vectors. */ | |
4757 | && (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
4758 | || !VECTOR_BOOLEAN_TYPE_P (vectype_in))) | |
7b7b1813 | 4759 | { |
73fbfcad | 4760 | if (dump_enabled_p ()) |
78c60e3d SS |
4761 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4762 | "type conversion to/from bit-precision " | |
e645e942 | 4763 | "unsupported.\n"); |
7b7b1813 RG |
4764 | return false; |
4765 | } | |
4766 | ||
ebfd146a IR |
4767 | if (!vec_stmt) /* transformation not required. */ |
4768 | { | |
4769 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 4770 | if (dump_enabled_p ()) |
78c60e3d | 4771 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4772 | "=== vectorizable_assignment ===\n"); |
4fc5ebf1 | 4773 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
ebfd146a IR |
4774 | return true; |
4775 | } | |
4776 | ||
67b8dbac | 4777 | /* Transform. */ |
73fbfcad | 4778 | if (dump_enabled_p ()) |
e645e942 | 4779 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
4780 | |
4781 | /* Handle def. */ | |
4782 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4783 | ||
4784 | /* Handle use. */ | |
f18b55bd | 4785 | for (j = 0; j < ncopies; j++) |
ebfd146a | 4786 | { |
f18b55bd IR |
4787 | /* Handle uses. */ |
4788 | if (j == 0) | |
306b0c92 | 4789 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
f18b55bd IR |
4790 | else |
4791 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
4792 | ||
4793 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 4794 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 4795 | { |
7b7ec6c5 RG |
4796 | if (CONVERT_EXPR_CODE_P (code) |
4797 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 4798 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
4799 | new_stmt = gimple_build_assign (vec_dest, vop); |
4800 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4801 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4802 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4803 | if (slp_node) | |
9771b263 | 4804 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 4805 | } |
ebfd146a IR |
4806 | |
4807 | if (slp_node) | |
f18b55bd IR |
4808 | continue; |
4809 | ||
4810 | if (j == 0) | |
4811 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4812 | else | |
4813 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4814 | ||
4815 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4816 | } | |
b8698a0f | 4817 | |
9771b263 | 4818 | vec_oprnds.release (); |
ebfd146a IR |
4819 | return true; |
4820 | } | |
4821 | ||
9dc3f7de | 4822 | |
1107f3ae IR |
4823 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
4824 | either as shift by a scalar or by a vector. */ | |
4825 | ||
4826 | bool | |
4827 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
4828 | { | |
4829 | ||
ef4bddc2 | 4830 | machine_mode vec_mode; |
1107f3ae IR |
4831 | optab optab; |
4832 | int icode; | |
4833 | tree vectype; | |
4834 | ||
4835 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4836 | if (!vectype) | |
4837 | return false; | |
4838 | ||
4839 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4840 | if (!optab | |
4841 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
4842 | { | |
4843 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4844 | if (!optab | |
4845 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
4846 | == CODE_FOR_nothing)) | |
4847 | return false; | |
4848 | } | |
4849 | ||
4850 | vec_mode = TYPE_MODE (vectype); | |
4851 | icode = (int) optab_handler (optab, vec_mode); | |
4852 | if (icode == CODE_FOR_nothing) | |
4853 | return false; | |
4854 | ||
4855 | return true; | |
4856 | } | |
4857 | ||
4858 | ||
9dc3f7de IR |
4859 | /* Function vectorizable_shift. |
4860 | ||
4861 | Check if STMT performs a shift operation that can be vectorized. | |
4862 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4863 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
4864 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4865 | ||
4866 | static bool | |
355fe088 TS |
4867 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
4868 | gimple **vec_stmt, slp_tree slp_node) | |
9dc3f7de IR |
4869 | { |
4870 | tree vec_dest; | |
4871 | tree scalar_dest; | |
4872 | tree op0, op1 = NULL; | |
4873 | tree vec_oprnd1 = NULL_TREE; | |
4874 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4875 | tree vectype; | |
4876 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4877 | enum tree_code code; | |
ef4bddc2 | 4878 | machine_mode vec_mode; |
9dc3f7de IR |
4879 | tree new_temp; |
4880 | optab optab; | |
4881 | int icode; | |
ef4bddc2 | 4882 | machine_mode optab_op2_mode; |
355fe088 | 4883 | gimple *def_stmt; |
9dc3f7de | 4884 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 4885 | int ndts = 2; |
355fe088 | 4886 | gimple *new_stmt = NULL; |
9dc3f7de IR |
4887 | stmt_vec_info prev_stmt_info; |
4888 | int nunits_in; | |
4889 | int nunits_out; | |
4890 | tree vectype_out; | |
cede2577 | 4891 | tree op1_vectype; |
9dc3f7de IR |
4892 | int ncopies; |
4893 | int j, i; | |
6e1aa848 DN |
4894 | vec<tree> vec_oprnds0 = vNULL; |
4895 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
4896 | tree vop0, vop1; |
4897 | unsigned int k; | |
49eab32e | 4898 | bool scalar_shift_arg = true; |
9dc3f7de | 4899 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4900 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
4901 | |
4902 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
4903 | return false; | |
4904 | ||
66c16fd9 RB |
4905 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4906 | && ! vec_stmt) | |
9dc3f7de IR |
4907 | return false; |
4908 | ||
4909 | /* Is STMT a vectorizable binary/unary operation? */ | |
4910 | if (!is_gimple_assign (stmt)) | |
4911 | return false; | |
4912 | ||
4913 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4914 | return false; | |
4915 | ||
4916 | code = gimple_assign_rhs_code (stmt); | |
4917 | ||
4918 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
4919 | || code == RROTATE_EXPR)) | |
4920 | return false; | |
4921 | ||
4922 | scalar_dest = gimple_assign_lhs (stmt); | |
4923 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2be65d9e | 4924 | if (!type_has_mode_precision_p (TREE_TYPE (scalar_dest))) |
7b7b1813 | 4925 | { |
73fbfcad | 4926 | if (dump_enabled_p ()) |
78c60e3d | 4927 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4928 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
4929 | return false; |
4930 | } | |
9dc3f7de IR |
4931 | |
4932 | op0 = gimple_assign_rhs1 (stmt); | |
81c40241 | 4933 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
9dc3f7de | 4934 | { |
73fbfcad | 4935 | if (dump_enabled_p ()) |
78c60e3d | 4936 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4937 | "use not simple.\n"); |
9dc3f7de IR |
4938 | return false; |
4939 | } | |
4940 | /* If op0 is an external or constant def use a vector type with | |
4941 | the same size as the output vector type. */ | |
4942 | if (!vectype) | |
4943 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4944 | if (vec_stmt) | |
4945 | gcc_assert (vectype); | |
4946 | if (!vectype) | |
4947 | { | |
73fbfcad | 4948 | if (dump_enabled_p ()) |
78c60e3d | 4949 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4950 | "no vectype for scalar type\n"); |
9dc3f7de IR |
4951 | return false; |
4952 | } | |
4953 | ||
4954 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4955 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4956 | if (nunits_out != nunits_in) | |
4957 | return false; | |
4958 | ||
4959 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4960 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &op1_vectype)) |
9dc3f7de | 4961 | { |
73fbfcad | 4962 | if (dump_enabled_p ()) |
78c60e3d | 4963 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4964 | "use not simple.\n"); |
9dc3f7de IR |
4965 | return false; |
4966 | } | |
4967 | ||
9dc3f7de IR |
4968 | /* Multiple types in SLP are handled by creating the appropriate number of |
4969 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4970 | case of SLP. */ | |
fce57248 | 4971 | if (slp_node) |
9dc3f7de IR |
4972 | ncopies = 1; |
4973 | else | |
e8f142e2 | 4974 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
9dc3f7de IR |
4975 | |
4976 | gcc_assert (ncopies >= 1); | |
4977 | ||
4978 | /* Determine whether the shift amount is a vector, or scalar. If the | |
4979 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
4980 | ||
dbfa87aa YR |
4981 | if ((dt[1] == vect_internal_def |
4982 | || dt[1] == vect_induction_def) | |
4983 | && !slp_node) | |
49eab32e JJ |
4984 | scalar_shift_arg = false; |
4985 | else if (dt[1] == vect_constant_def | |
4986 | || dt[1] == vect_external_def | |
4987 | || dt[1] == vect_internal_def) | |
4988 | { | |
4989 | /* In SLP, need to check whether the shift count is the same, | |
4990 | in loops if it is a constant or invariant, it is always | |
4991 | a scalar shift. */ | |
4992 | if (slp_node) | |
4993 | { | |
355fe088 TS |
4994 | vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
4995 | gimple *slpstmt; | |
49eab32e | 4996 | |
9771b263 | 4997 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
4998 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
4999 | scalar_shift_arg = false; | |
5000 | } | |
60d393e8 RB |
5001 | |
5002 | /* If the shift amount is computed by a pattern stmt we cannot | |
5003 | use the scalar amount directly thus give up and use a vector | |
5004 | shift. */ | |
5005 | if (dt[1] == vect_internal_def) | |
5006 | { | |
5007 | gimple *def = SSA_NAME_DEF_STMT (op1); | |
5008 | if (is_pattern_stmt_p (vinfo_for_stmt (def))) | |
5009 | scalar_shift_arg = false; | |
5010 | } | |
49eab32e JJ |
5011 | } |
5012 | else | |
5013 | { | |
73fbfcad | 5014 | if (dump_enabled_p ()) |
78c60e3d | 5015 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5016 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
5017 | return false; |
5018 | } | |
5019 | ||
9dc3f7de | 5020 | /* Vector shifted by vector. */ |
49eab32e | 5021 | if (!scalar_shift_arg) |
9dc3f7de IR |
5022 | { |
5023 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 5024 | if (dump_enabled_p ()) |
78c60e3d | 5025 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5026 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 5027 | |
aa948027 JJ |
5028 | if (!op1_vectype) |
5029 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
5030 | if (op1_vectype == NULL_TREE | |
5031 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 5032 | { |
73fbfcad | 5033 | if (dump_enabled_p ()) |
78c60e3d SS |
5034 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5035 | "unusable type for last operand in" | |
e645e942 | 5036 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
5037 | return false; |
5038 | } | |
9dc3f7de IR |
5039 | } |
5040 | /* See if the machine has a vector shifted by scalar insn and if not | |
5041 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 5042 | else |
9dc3f7de IR |
5043 | { |
5044 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5045 | if (optab | |
5046 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
5047 | { | |
73fbfcad | 5048 | if (dump_enabled_p ()) |
78c60e3d | 5049 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5050 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
5051 | } |
5052 | else | |
5053 | { | |
5054 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5055 | if (optab | |
5056 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
5057 | != CODE_FOR_nothing)) | |
5058 | { | |
49eab32e JJ |
5059 | scalar_shift_arg = false; |
5060 | ||
73fbfcad | 5061 | if (dump_enabled_p ()) |
78c60e3d | 5062 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5063 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
5064 | |
5065 | /* Unlike the other binary operators, shifts/rotates have | |
5066 | the rhs being int, instead of the same type as the lhs, | |
5067 | so make sure the scalar is the right type if we are | |
aa948027 | 5068 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
5069 | if (dt[1] == vect_constant_def) |
5070 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
5071 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
5072 | TREE_TYPE (op1))) | |
5073 | { | |
5074 | if (slp_node | |
5075 | && TYPE_MODE (TREE_TYPE (vectype)) | |
5076 | != TYPE_MODE (TREE_TYPE (op1))) | |
5077 | { | |
73fbfcad | 5078 | if (dump_enabled_p ()) |
78c60e3d SS |
5079 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5080 | "unusable type for last operand in" | |
e645e942 | 5081 | " vector/vector shift/rotate.\n"); |
21c0a521 | 5082 | return false; |
aa948027 JJ |
5083 | } |
5084 | if (vec_stmt && !slp_node) | |
5085 | { | |
5086 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
5087 | op1 = vect_init_vector (stmt, op1, | |
5088 | TREE_TYPE (vectype), NULL); | |
5089 | } | |
5090 | } | |
9dc3f7de IR |
5091 | } |
5092 | } | |
5093 | } | |
9dc3f7de IR |
5094 | |
5095 | /* Supportable by target? */ | |
5096 | if (!optab) | |
5097 | { | |
73fbfcad | 5098 | if (dump_enabled_p ()) |
78c60e3d | 5099 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5100 | "no optab.\n"); |
9dc3f7de IR |
5101 | return false; |
5102 | } | |
5103 | vec_mode = TYPE_MODE (vectype); | |
5104 | icode = (int) optab_handler (optab, vec_mode); | |
5105 | if (icode == CODE_FOR_nothing) | |
5106 | { | |
73fbfcad | 5107 | if (dump_enabled_p ()) |
78c60e3d | 5108 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5109 | "op not supported by target.\n"); |
9dc3f7de IR |
5110 | /* Check only during analysis. */ |
5111 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
ca09abcb RS |
5112 | || (!vec_stmt |
5113 | && !vect_worthwhile_without_simd_p (vinfo, code))) | |
9dc3f7de | 5114 | return false; |
73fbfcad | 5115 | if (dump_enabled_p ()) |
e645e942 TJ |
5116 | dump_printf_loc (MSG_NOTE, vect_location, |
5117 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
5118 | } |
5119 | ||
5120 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
ca09abcb RS |
5121 | if (!vec_stmt |
5122 | && !VECTOR_MODE_P (TYPE_MODE (vectype)) | |
5123 | && !vect_worthwhile_without_simd_p (vinfo, code)) | |
9dc3f7de | 5124 | { |
73fbfcad | 5125 | if (dump_enabled_p ()) |
78c60e3d | 5126 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5127 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
5128 | return false; |
5129 | } | |
5130 | ||
5131 | if (!vec_stmt) /* transformation not required. */ | |
5132 | { | |
5133 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 5134 | if (dump_enabled_p ()) |
e645e942 TJ |
5135 | dump_printf_loc (MSG_NOTE, vect_location, |
5136 | "=== vectorizable_shift ===\n"); | |
4fc5ebf1 | 5137 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
9dc3f7de IR |
5138 | return true; |
5139 | } | |
5140 | ||
67b8dbac | 5141 | /* Transform. */ |
9dc3f7de | 5142 | |
73fbfcad | 5143 | if (dump_enabled_p ()) |
78c60e3d | 5144 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5145 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
5146 | |
5147 | /* Handle def. */ | |
5148 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5149 | ||
9dc3f7de IR |
5150 | prev_stmt_info = NULL; |
5151 | for (j = 0; j < ncopies; j++) | |
5152 | { | |
5153 | /* Handle uses. */ | |
5154 | if (j == 0) | |
5155 | { | |
5156 | if (scalar_shift_arg) | |
5157 | { | |
5158 | /* Vector shl and shr insn patterns can be defined with scalar | |
5159 | operand 2 (shift operand). In this case, use constant or loop | |
5160 | invariant op1 directly, without extending it to vector mode | |
5161 | first. */ | |
5162 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
5163 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
5164 | { | |
73fbfcad | 5165 | if (dump_enabled_p ()) |
78c60e3d | 5166 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5167 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 5168 | vec_oprnd1 = op1; |
8930f723 | 5169 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 5170 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5171 | if (slp_node) |
5172 | { | |
5173 | /* Store vec_oprnd1 for every vector stmt to be created | |
5174 | for SLP_NODE. We check during the analysis that all | |
5175 | the shift arguments are the same. | |
5176 | TODO: Allow different constants for different vector | |
5177 | stmts generated for an SLP instance. */ | |
5178 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 5179 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5180 | } |
5181 | } | |
5182 | } | |
5183 | ||
5184 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
5185 | (a special case for certain kind of vector shifts); otherwise, | |
5186 | operand 1 should be of a vector type (the usual case). */ | |
5187 | if (vec_oprnd1) | |
5188 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5189 | slp_node); |
9dc3f7de IR |
5190 | else |
5191 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
306b0c92 | 5192 | slp_node); |
9dc3f7de IR |
5193 | } |
5194 | else | |
5195 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5196 | ||
5197 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5198 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 5199 | { |
9771b263 | 5200 | vop1 = vec_oprnds1[i]; |
0d0e4a03 | 5201 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
5202 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5203 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5204 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5205 | if (slp_node) | |
9771b263 | 5206 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
5207 | } |
5208 | ||
5209 | if (slp_node) | |
5210 | continue; | |
5211 | ||
5212 | if (j == 0) | |
5213 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5214 | else | |
5215 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5216 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5217 | } | |
5218 | ||
9771b263 DN |
5219 | vec_oprnds0.release (); |
5220 | vec_oprnds1.release (); | |
9dc3f7de IR |
5221 | |
5222 | return true; | |
5223 | } | |
5224 | ||
5225 | ||
ebfd146a IR |
5226 | /* Function vectorizable_operation. |
5227 | ||
16949072 RG |
5228 | Check if STMT performs a binary, unary or ternary operation that can |
5229 | be vectorized. | |
b8698a0f | 5230 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
5231 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5232 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5233 | ||
5234 | static bool | |
355fe088 TS |
5235 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
5236 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a | 5237 | { |
00f07b86 | 5238 | tree vec_dest; |
ebfd146a | 5239 | tree scalar_dest; |
16949072 | 5240 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 5241 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 5242 | tree vectype; |
ebfd146a | 5243 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
0eb952ea | 5244 | enum tree_code code, orig_code; |
ef4bddc2 | 5245 | machine_mode vec_mode; |
ebfd146a IR |
5246 | tree new_temp; |
5247 | int op_type; | |
00f07b86 | 5248 | optab optab; |
523ba738 | 5249 | bool target_support_p; |
355fe088 | 5250 | gimple *def_stmt; |
16949072 RG |
5251 | enum vect_def_type dt[3] |
5252 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 5253 | int ndts = 3; |
355fe088 | 5254 | gimple *new_stmt = NULL; |
ebfd146a | 5255 | stmt_vec_info prev_stmt_info; |
b690cc0f | 5256 | int nunits_in; |
ebfd146a IR |
5257 | int nunits_out; |
5258 | tree vectype_out; | |
5259 | int ncopies; | |
5260 | int j, i; | |
6e1aa848 DN |
5261 | vec<tree> vec_oprnds0 = vNULL; |
5262 | vec<tree> vec_oprnds1 = vNULL; | |
5263 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 5264 | tree vop0, vop1, vop2; |
a70d6342 | 5265 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5266 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 5267 | |
a70d6342 | 5268 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5269 | return false; |
5270 | ||
66c16fd9 RB |
5271 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5272 | && ! vec_stmt) | |
ebfd146a IR |
5273 | return false; |
5274 | ||
5275 | /* Is STMT a vectorizable binary/unary operation? */ | |
5276 | if (!is_gimple_assign (stmt)) | |
5277 | return false; | |
5278 | ||
5279 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5280 | return false; | |
5281 | ||
0eb952ea | 5282 | orig_code = code = gimple_assign_rhs_code (stmt); |
ebfd146a | 5283 | |
1af4ebf5 MG |
5284 | /* For pointer addition and subtraction, we should use the normal |
5285 | plus and minus for the vector operation. */ | |
ebfd146a IR |
5286 | if (code == POINTER_PLUS_EXPR) |
5287 | code = PLUS_EXPR; | |
1af4ebf5 MG |
5288 | if (code == POINTER_DIFF_EXPR) |
5289 | code = MINUS_EXPR; | |
ebfd146a IR |
5290 | |
5291 | /* Support only unary or binary operations. */ | |
5292 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 5293 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 5294 | { |
73fbfcad | 5295 | if (dump_enabled_p ()) |
78c60e3d | 5296 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5297 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 5298 | op_type); |
ebfd146a IR |
5299 | return false; |
5300 | } | |
5301 | ||
b690cc0f RG |
5302 | scalar_dest = gimple_assign_lhs (stmt); |
5303 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
5304 | ||
7b7b1813 RG |
5305 | /* Most operations cannot handle bit-precision types without extra |
5306 | truncations. */ | |
045c1278 | 5307 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
2be65d9e | 5308 | && !type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
7b7b1813 RG |
5309 | /* Exception are bitwise binary operations. */ |
5310 | && code != BIT_IOR_EXPR | |
5311 | && code != BIT_XOR_EXPR | |
5312 | && code != BIT_AND_EXPR) | |
5313 | { | |
73fbfcad | 5314 | if (dump_enabled_p ()) |
78c60e3d | 5315 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5316 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
5317 | return false; |
5318 | } | |
5319 | ||
ebfd146a | 5320 | op0 = gimple_assign_rhs1 (stmt); |
81c40241 | 5321 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
ebfd146a | 5322 | { |
73fbfcad | 5323 | if (dump_enabled_p ()) |
78c60e3d | 5324 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5325 | "use not simple.\n"); |
ebfd146a IR |
5326 | return false; |
5327 | } | |
b690cc0f RG |
5328 | /* If op0 is an external or constant def use a vector type with |
5329 | the same size as the output vector type. */ | |
5330 | if (!vectype) | |
b036c6c5 IE |
5331 | { |
5332 | /* For boolean type we cannot determine vectype by | |
5333 | invariant value (don't know whether it is a vector | |
5334 | of booleans or vector of integers). We use output | |
5335 | vectype because operations on boolean don't change | |
5336 | type. */ | |
2568d8a1 | 5337 | if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op0))) |
b036c6c5 | 5338 | { |
2568d8a1 | 5339 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (scalar_dest))) |
b036c6c5 IE |
5340 | { |
5341 | if (dump_enabled_p ()) | |
5342 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5343 | "not supported operation on bool value.\n"); | |
5344 | return false; | |
5345 | } | |
5346 | vectype = vectype_out; | |
5347 | } | |
5348 | else | |
5349 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5350 | } | |
7d8930a0 IR |
5351 | if (vec_stmt) |
5352 | gcc_assert (vectype); | |
5353 | if (!vectype) | |
5354 | { | |
73fbfcad | 5355 | if (dump_enabled_p ()) |
7d8930a0 | 5356 | { |
78c60e3d SS |
5357 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5358 | "no vectype for scalar type "); | |
5359 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
5360 | TREE_TYPE (op0)); | |
e645e942 | 5361 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
5362 | } |
5363 | ||
5364 | return false; | |
5365 | } | |
b690cc0f RG |
5366 | |
5367 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5368 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
5369 | if (nunits_out != nunits_in) | |
5370 | return false; | |
ebfd146a | 5371 | |
16949072 | 5372 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
5373 | { |
5374 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 5375 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1])) |
ebfd146a | 5376 | { |
73fbfcad | 5377 | if (dump_enabled_p ()) |
78c60e3d | 5378 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5379 | "use not simple.\n"); |
ebfd146a IR |
5380 | return false; |
5381 | } | |
5382 | } | |
16949072 RG |
5383 | if (op_type == ternary_op) |
5384 | { | |
5385 | op2 = gimple_assign_rhs3 (stmt); | |
81c40241 | 5386 | if (!vect_is_simple_use (op2, vinfo, &def_stmt, &dt[2])) |
16949072 | 5387 | { |
73fbfcad | 5388 | if (dump_enabled_p ()) |
78c60e3d | 5389 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5390 | "use not simple.\n"); |
16949072 RG |
5391 | return false; |
5392 | } | |
5393 | } | |
ebfd146a | 5394 | |
b690cc0f | 5395 | /* Multiple types in SLP are handled by creating the appropriate number of |
ff802fa1 | 5396 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 5397 | case of SLP. */ |
fce57248 | 5398 | if (slp_node) |
b690cc0f RG |
5399 | ncopies = 1; |
5400 | else | |
e8f142e2 | 5401 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
b690cc0f RG |
5402 | |
5403 | gcc_assert (ncopies >= 1); | |
5404 | ||
9dc3f7de | 5405 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
5406 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
5407 | || code == RROTATE_EXPR) | |
9dc3f7de | 5408 | return false; |
ebfd146a | 5409 | |
ebfd146a | 5410 | /* Supportable by target? */ |
00f07b86 RH |
5411 | |
5412 | vec_mode = TYPE_MODE (vectype); | |
5413 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 5414 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
5415 | else |
5416 | { | |
5417 | optab = optab_for_tree_code (code, vectype, optab_default); | |
5418 | if (!optab) | |
5deb57cb | 5419 | { |
73fbfcad | 5420 | if (dump_enabled_p ()) |
78c60e3d | 5421 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5422 | "no optab.\n"); |
00f07b86 | 5423 | return false; |
5deb57cb | 5424 | } |
523ba738 RS |
5425 | target_support_p = (optab_handler (optab, vec_mode) |
5426 | != CODE_FOR_nothing); | |
5deb57cb JJ |
5427 | } |
5428 | ||
523ba738 | 5429 | if (!target_support_p) |
ebfd146a | 5430 | { |
73fbfcad | 5431 | if (dump_enabled_p ()) |
78c60e3d | 5432 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5433 | "op not supported by target.\n"); |
ebfd146a IR |
5434 | /* Check only during analysis. */ |
5435 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
ca09abcb | 5436 | || (!vec_stmt && !vect_worthwhile_without_simd_p (vinfo, code))) |
ebfd146a | 5437 | return false; |
73fbfcad | 5438 | if (dump_enabled_p ()) |
e645e942 TJ |
5439 | dump_printf_loc (MSG_NOTE, vect_location, |
5440 | "proceeding using word mode.\n"); | |
383d9c83 IR |
5441 | } |
5442 | ||
4a00c761 | 5443 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
5444 | if (!VECTOR_MODE_P (vec_mode) |
5445 | && !vec_stmt | |
ca09abcb | 5446 | && !vect_worthwhile_without_simd_p (vinfo, code)) |
7d8930a0 | 5447 | { |
73fbfcad | 5448 | if (dump_enabled_p ()) |
78c60e3d | 5449 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5450 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 5451 | return false; |
7d8930a0 | 5452 | } |
ebfd146a | 5453 | |
ebfd146a IR |
5454 | if (!vec_stmt) /* transformation not required. */ |
5455 | { | |
4a00c761 | 5456 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 5457 | if (dump_enabled_p ()) |
78c60e3d | 5458 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5459 | "=== vectorizable_operation ===\n"); |
4fc5ebf1 | 5460 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
ebfd146a IR |
5461 | return true; |
5462 | } | |
5463 | ||
67b8dbac | 5464 | /* Transform. */ |
ebfd146a | 5465 | |
73fbfcad | 5466 | if (dump_enabled_p ()) |
78c60e3d | 5467 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5468 | "transform binary/unary operation.\n"); |
383d9c83 | 5469 | |
ebfd146a | 5470 | /* Handle def. */ |
00f07b86 | 5471 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 5472 | |
0eb952ea JJ |
5473 | /* POINTER_DIFF_EXPR has pointer arguments which are vectorized as |
5474 | vectors with unsigned elements, but the result is signed. So, we | |
5475 | need to compute the MINUS_EXPR into vectype temporary and | |
5476 | VIEW_CONVERT_EXPR it into the final vectype_out result. */ | |
5477 | tree vec_cvt_dest = NULL_TREE; | |
5478 | if (orig_code == POINTER_DIFF_EXPR) | |
5479 | vec_cvt_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
5480 | ||
ebfd146a IR |
5481 | /* In case the vectorization factor (VF) is bigger than the number |
5482 | of elements that we can fit in a vectype (nunits), we have to generate | |
5483 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
5484 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
5485 | from one copy of the vector stmt to the next, in the field | |
5486 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
5487 | stages to find the correct vector defs to be used when vectorizing | |
5488 | stmts that use the defs of the current stmt. The example below | |
5489 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
5490 | we need to create 4 vectorized stmts): | |
5491 | ||
5492 | before vectorization: | |
5493 | RELATED_STMT VEC_STMT | |
5494 | S1: x = memref - - | |
5495 | S2: z = x + 1 - - | |
5496 | ||
5497 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
5498 | there): | |
5499 | RELATED_STMT VEC_STMT | |
5500 | VS1_0: vx0 = memref0 VS1_1 - | |
5501 | VS1_1: vx1 = memref1 VS1_2 - | |
5502 | VS1_2: vx2 = memref2 VS1_3 - | |
5503 | VS1_3: vx3 = memref3 - - | |
5504 | S1: x = load - VS1_0 | |
5505 | S2: z = x + 1 - - | |
5506 | ||
5507 | step2: vectorize stmt S2 (done here): | |
5508 | To vectorize stmt S2 we first need to find the relevant vector | |
5509 | def for the first operand 'x'. This is, as usual, obtained from | |
5510 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
5511 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
5512 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
5513 | the vector stmt VS2_0, and as usual, record it in the | |
5514 | STMT_VINFO_VEC_STMT of stmt S2. | |
5515 | When creating the second copy (VS2_1), we obtain the relevant vector | |
5516 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
5517 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
5518 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
5519 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
5520 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
5521 | chain of stmts and pointers: | |
5522 | RELATED_STMT VEC_STMT | |
5523 | VS1_0: vx0 = memref0 VS1_1 - | |
5524 | VS1_1: vx1 = memref1 VS1_2 - | |
5525 | VS1_2: vx2 = memref2 VS1_3 - | |
5526 | VS1_3: vx3 = memref3 - - | |
5527 | S1: x = load - VS1_0 | |
5528 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
5529 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
5530 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
5531 | VS2_3: vz3 = vx3 + v1 - - | |
5532 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
5533 | |
5534 | prev_stmt_info = NULL; | |
5535 | for (j = 0; j < ncopies; j++) | |
5536 | { | |
5537 | /* Handle uses. */ | |
5538 | if (j == 0) | |
4a00c761 JJ |
5539 | { |
5540 | if (op_type == binary_op || op_type == ternary_op) | |
5541 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
306b0c92 | 5542 | slp_node); |
4a00c761 JJ |
5543 | else |
5544 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5545 | slp_node); |
4a00c761 | 5546 | if (op_type == ternary_op) |
c392943c | 5547 | vect_get_vec_defs (op2, NULL_TREE, stmt, &vec_oprnds2, NULL, |
306b0c92 | 5548 | slp_node); |
4a00c761 | 5549 | } |
ebfd146a | 5550 | else |
4a00c761 JJ |
5551 | { |
5552 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5553 | if (op_type == ternary_op) | |
5554 | { | |
9771b263 DN |
5555 | tree vec_oprnd = vec_oprnds2.pop (); |
5556 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
5557 | vec_oprnd)); | |
4a00c761 JJ |
5558 | } |
5559 | } | |
5560 | ||
5561 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5562 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 5563 | { |
4a00c761 | 5564 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 5565 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 5566 | vop2 = ((op_type == ternary_op) |
9771b263 | 5567 | ? vec_oprnds2[i] : NULL_TREE); |
0d0e4a03 | 5568 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
4a00c761 JJ |
5569 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5570 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5571 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
0eb952ea JJ |
5572 | if (vec_cvt_dest) |
5573 | { | |
5574 | new_temp = build1 (VIEW_CONVERT_EXPR, vectype_out, new_temp); | |
5575 | new_stmt = gimple_build_assign (vec_cvt_dest, VIEW_CONVERT_EXPR, | |
5576 | new_temp); | |
5577 | new_temp = make_ssa_name (vec_cvt_dest, new_stmt); | |
5578 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5579 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5580 | } | |
4a00c761 | 5581 | if (slp_node) |
9771b263 | 5582 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
5583 | } |
5584 | ||
4a00c761 JJ |
5585 | if (slp_node) |
5586 | continue; | |
5587 | ||
5588 | if (j == 0) | |
5589 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5590 | else | |
5591 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5592 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
5593 | } |
5594 | ||
9771b263 DN |
5595 | vec_oprnds0.release (); |
5596 | vec_oprnds1.release (); | |
5597 | vec_oprnds2.release (); | |
ebfd146a | 5598 | |
ebfd146a IR |
5599 | return true; |
5600 | } | |
5601 | ||
f702e7d4 | 5602 | /* A helper function to ensure data reference DR's base alignment. */ |
c716e67f XDL |
5603 | |
5604 | static void | |
f702e7d4 | 5605 | ensure_base_align (struct data_reference *dr) |
c716e67f XDL |
5606 | { |
5607 | if (!dr->aux) | |
5608 | return; | |
5609 | ||
52639a61 | 5610 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f | 5611 | { |
52639a61 | 5612 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 5613 | |
f702e7d4 RS |
5614 | unsigned int align_base_to = DR_TARGET_ALIGNMENT (dr) * BITS_PER_UNIT; |
5615 | ||
428f0c67 | 5616 | if (decl_in_symtab_p (base_decl)) |
f702e7d4 | 5617 | symtab_node::get (base_decl)->increase_alignment (align_base_to); |
428f0c67 JH |
5618 | else |
5619 | { | |
f702e7d4 | 5620 | SET_DECL_ALIGN (base_decl, align_base_to); |
428f0c67 JH |
5621 | DECL_USER_ALIGN (base_decl) = 1; |
5622 | } | |
52639a61 | 5623 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
5624 | } |
5625 | } | |
5626 | ||
ebfd146a | 5627 | |
44fc7854 BE |
5628 | /* Function get_group_alias_ptr_type. |
5629 | ||
5630 | Return the alias type for the group starting at FIRST_STMT. */ | |
5631 | ||
5632 | static tree | |
5633 | get_group_alias_ptr_type (gimple *first_stmt) | |
5634 | { | |
5635 | struct data_reference *first_dr, *next_dr; | |
5636 | gimple *next_stmt; | |
5637 | ||
5638 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
5639 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (first_stmt)); | |
5640 | while (next_stmt) | |
5641 | { | |
5642 | next_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (next_stmt)); | |
5643 | if (get_alias_set (DR_REF (first_dr)) | |
5644 | != get_alias_set (DR_REF (next_dr))) | |
5645 | { | |
5646 | if (dump_enabled_p ()) | |
5647 | dump_printf_loc (MSG_NOTE, vect_location, | |
5648 | "conflicting alias set types.\n"); | |
5649 | return ptr_type_node; | |
5650 | } | |
5651 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); | |
5652 | } | |
5653 | return reference_alias_ptr_type (DR_REF (first_dr)); | |
5654 | } | |
5655 | ||
5656 | ||
ebfd146a IR |
5657 | /* Function vectorizable_store. |
5658 | ||
b8698a0f L |
5659 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
5660 | can be vectorized. | |
5661 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5662 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5663 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5664 | ||
5665 | static bool | |
355fe088 | 5666 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 5667 | slp_tree slp_node) |
ebfd146a IR |
5668 | { |
5669 | tree scalar_dest; | |
5670 | tree data_ref; | |
5671 | tree op; | |
5672 | tree vec_oprnd = NULL_TREE; | |
5673 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5674 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 5675 | tree elem_type; |
ebfd146a | 5676 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 5677 | struct loop *loop = NULL; |
ef4bddc2 | 5678 | machine_mode vec_mode; |
ebfd146a IR |
5679 | tree dummy; |
5680 | enum dr_alignment_support alignment_support_scheme; | |
355fe088 | 5681 | gimple *def_stmt; |
ebfd146a IR |
5682 | enum vect_def_type dt; |
5683 | stmt_vec_info prev_stmt_info = NULL; | |
5684 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 5685 | tree dataref_offset = NULL_TREE; |
355fe088 | 5686 | gimple *ptr_incr = NULL; |
ebfd146a IR |
5687 | int ncopies; |
5688 | int j; | |
2de001ee RS |
5689 | gimple *next_stmt, *first_stmt; |
5690 | bool grouped_store; | |
ebfd146a | 5691 | unsigned int group_size, i; |
6e1aa848 DN |
5692 | vec<tree> oprnds = vNULL; |
5693 | vec<tree> result_chain = vNULL; | |
ebfd146a | 5694 | bool inv_p; |
09dfa495 | 5695 | tree offset = NULL_TREE; |
6e1aa848 | 5696 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5697 | bool slp = (slp_node != NULL); |
ebfd146a | 5698 | unsigned int vec_num; |
a70d6342 | 5699 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5700 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 5701 | tree aggr_type; |
134c85ca | 5702 | gather_scatter_info gs_info; |
3bab6342 | 5703 | enum vect_def_type scatter_src_dt = vect_unknown_def_type; |
355fe088 | 5704 | gimple *new_stmt; |
d9f21f6a | 5705 | poly_uint64 vf; |
2de001ee | 5706 | vec_load_store_type vls_type; |
44fc7854 | 5707 | tree ref_type; |
a70d6342 | 5708 | |
a70d6342 | 5709 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5710 | return false; |
5711 | ||
66c16fd9 RB |
5712 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5713 | && ! vec_stmt) | |
ebfd146a IR |
5714 | return false; |
5715 | ||
5716 | /* Is vectorizable store? */ | |
5717 | ||
5718 | if (!is_gimple_assign (stmt)) | |
5719 | return false; | |
5720 | ||
5721 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
5722 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
5723 | && is_pattern_stmt_p (stmt_info)) | |
5724 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 5725 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 5726 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 5727 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
5728 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
5729 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
5730 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
5731 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
5732 | return false; |
5733 | ||
fce57248 RS |
5734 | /* Cannot have hybrid store SLP -- that would mean storing to the |
5735 | same location twice. */ | |
5736 | gcc_assert (slp == PURE_SLP_STMT (stmt_info)); | |
5737 | ||
ebfd146a | 5738 | gcc_assert (gimple_assign_single_p (stmt)); |
465c8c19 | 5739 | |
f4d09712 | 5740 | tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE; |
465c8c19 JJ |
5741 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
5742 | ||
5743 | if (loop_vinfo) | |
b17dc4d4 RB |
5744 | { |
5745 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
5746 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
5747 | } | |
5748 | else | |
5749 | vf = 1; | |
465c8c19 JJ |
5750 | |
5751 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5752 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5753 | case of SLP. */ | |
fce57248 | 5754 | if (slp) |
465c8c19 JJ |
5755 | ncopies = 1; |
5756 | else | |
e8f142e2 | 5757 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
5758 | |
5759 | gcc_assert (ncopies >= 1); | |
5760 | ||
5761 | /* FORNOW. This restriction should be relaxed. */ | |
5762 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
5763 | { | |
5764 | if (dump_enabled_p ()) | |
5765 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5766 | "multiple types in nested loop.\n"); | |
5767 | return false; | |
5768 | } | |
5769 | ||
ebfd146a | 5770 | op = gimple_assign_rhs1 (stmt); |
f4d09712 | 5771 | |
2f391428 | 5772 | /* In the case this is a store from a constant make sure |
11a82e25 | 5773 | native_encode_expr can handle it. */ |
2f391428 | 5774 | if (CONSTANT_CLASS_P (op) && native_encode_expr (op, NULL, 64) == 0) |
11a82e25 RB |
5775 | return false; |
5776 | ||
f4d09712 | 5777 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt, &rhs_vectype)) |
ebfd146a | 5778 | { |
73fbfcad | 5779 | if (dump_enabled_p ()) |
78c60e3d | 5780 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5781 | "use not simple.\n"); |
ebfd146a IR |
5782 | return false; |
5783 | } | |
5784 | ||
2de001ee RS |
5785 | if (dt == vect_constant_def || dt == vect_external_def) |
5786 | vls_type = VLS_STORE_INVARIANT; | |
5787 | else | |
5788 | vls_type = VLS_STORE; | |
5789 | ||
f4d09712 KY |
5790 | if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype)) |
5791 | return false; | |
5792 | ||
272c6793 | 5793 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 5794 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 5795 | |
ebfd146a IR |
5796 | /* FORNOW. In some cases can vectorize even if data-type not supported |
5797 | (e.g. - array initialization with 0). */ | |
947131ba | 5798 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
5799 | return false; |
5800 | ||
5801 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
5802 | return false; | |
5803 | ||
2de001ee | 5804 | vect_memory_access_type memory_access_type; |
62da9e14 | 5805 | if (!get_load_store_type (stmt, vectype, slp, vls_type, ncopies, |
2de001ee RS |
5806 | &memory_access_type, &gs_info)) |
5807 | return false; | |
3bab6342 | 5808 | |
ebfd146a IR |
5809 | if (!vec_stmt) /* transformation not required. */ |
5810 | { | |
2de001ee | 5811 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
ebfd146a | 5812 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; |
2e8ab70c RB |
5813 | /* The SLP costs are calculated during SLP analysis. */ |
5814 | if (!PURE_SLP_STMT (stmt_info)) | |
2de001ee | 5815 | vect_model_store_cost (stmt_info, ncopies, memory_access_type, dt, |
2e8ab70c | 5816 | NULL, NULL, NULL); |
ebfd146a IR |
5817 | return true; |
5818 | } | |
2de001ee | 5819 | gcc_assert (memory_access_type == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
ebfd146a | 5820 | |
67b8dbac | 5821 | /* Transform. */ |
ebfd146a | 5822 | |
f702e7d4 | 5823 | ensure_base_align (dr); |
c716e67f | 5824 | |
2de001ee | 5825 | if (memory_access_type == VMAT_GATHER_SCATTER) |
3bab6342 AT |
5826 | { |
5827 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, op, src; | |
134c85ca | 5828 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
5829 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; |
5830 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
5831 | edge pe = loop_preheader_edge (loop); | |
5832 | gimple_seq seq; | |
5833 | basic_block new_bb; | |
5834 | enum { NARROW, NONE, WIDEN } modifier; | |
134c85ca | 5835 | int scatter_off_nunits = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype); |
3bab6342 AT |
5836 | |
5837 | if (nunits == (unsigned int) scatter_off_nunits) | |
5838 | modifier = NONE; | |
5839 | else if (nunits == (unsigned int) scatter_off_nunits / 2) | |
5840 | { | |
3bab6342 AT |
5841 | modifier = WIDEN; |
5842 | ||
e3342de4 | 5843 | vec_perm_builder sel (scatter_off_nunits, scatter_off_nunits, 1); |
3bab6342 | 5844 | for (i = 0; i < (unsigned int) scatter_off_nunits; ++i) |
908a1a16 | 5845 | sel.quick_push (i | nunits); |
3bab6342 | 5846 | |
e3342de4 RS |
5847 | vec_perm_indices indices (sel, 1, scatter_off_nunits); |
5848 | perm_mask = vect_gen_perm_mask_checked (gs_info.offset_vectype, | |
5849 | indices); | |
3bab6342 AT |
5850 | gcc_assert (perm_mask != NULL_TREE); |
5851 | } | |
5852 | else if (nunits == (unsigned int) scatter_off_nunits * 2) | |
5853 | { | |
3bab6342 AT |
5854 | modifier = NARROW; |
5855 | ||
e3342de4 | 5856 | vec_perm_builder sel (nunits, nunits, 1); |
3bab6342 | 5857 | for (i = 0; i < (unsigned int) nunits; ++i) |
908a1a16 | 5858 | sel.quick_push (i | scatter_off_nunits); |
3bab6342 | 5859 | |
e3342de4 RS |
5860 | vec_perm_indices indices (sel, 2, nunits); |
5861 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); | |
3bab6342 AT |
5862 | gcc_assert (perm_mask != NULL_TREE); |
5863 | ncopies *= 2; | |
5864 | } | |
5865 | else | |
5866 | gcc_unreachable (); | |
5867 | ||
134c85ca | 5868 | rettype = TREE_TYPE (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
5869 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
5870 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5871 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5872 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5873 | scaletype = TREE_VALUE (arglist); | |
5874 | ||
5875 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
5876 | && TREE_CODE (rettype) == VOID_TYPE); | |
5877 | ||
134c85ca | 5878 | ptr = fold_convert (ptrtype, gs_info.base); |
3bab6342 AT |
5879 | if (!is_gimple_min_invariant (ptr)) |
5880 | { | |
5881 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
5882 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
5883 | gcc_assert (!new_bb); | |
5884 | } | |
5885 | ||
5886 | /* Currently we support only unconditional scatter stores, | |
5887 | so mask should be all ones. */ | |
5888 | mask = build_int_cst (masktype, -1); | |
5889 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
5890 | ||
134c85ca | 5891 | scale = build_int_cst (scaletype, gs_info.scale); |
3bab6342 AT |
5892 | |
5893 | prev_stmt_info = NULL; | |
5894 | for (j = 0; j < ncopies; ++j) | |
5895 | { | |
5896 | if (j == 0) | |
5897 | { | |
5898 | src = vec_oprnd1 | |
81c40241 | 5899 | = vect_get_vec_def_for_operand (gimple_assign_rhs1 (stmt), stmt); |
3bab6342 | 5900 | op = vec_oprnd0 |
134c85ca | 5901 | = vect_get_vec_def_for_operand (gs_info.offset, stmt); |
3bab6342 AT |
5902 | } |
5903 | else if (modifier != NONE && (j & 1)) | |
5904 | { | |
5905 | if (modifier == WIDEN) | |
5906 | { | |
5907 | src = vec_oprnd1 | |
5908 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5909 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, | |
5910 | stmt, gsi); | |
5911 | } | |
5912 | else if (modifier == NARROW) | |
5913 | { | |
5914 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
5915 | stmt, gsi); | |
5916 | op = vec_oprnd0 | |
134c85ca RS |
5917 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
5918 | vec_oprnd0); | |
3bab6342 AT |
5919 | } |
5920 | else | |
5921 | gcc_unreachable (); | |
5922 | } | |
5923 | else | |
5924 | { | |
5925 | src = vec_oprnd1 | |
5926 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5927 | op = vec_oprnd0 | |
134c85ca RS |
5928 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
5929 | vec_oprnd0); | |
3bab6342 AT |
5930 | } |
5931 | ||
5932 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
5933 | { | |
5934 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)) | |
5935 | == TYPE_VECTOR_SUBPARTS (srctype)); | |
0e22bb5a | 5936 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 AT |
5937 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
5938 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
5939 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5940 | src = var; | |
5941 | } | |
5942 | ||
5943 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
5944 | { | |
5945 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
5946 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 5947 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 AT |
5948 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
5949 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
5950 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5951 | op = var; | |
5952 | } | |
5953 | ||
5954 | new_stmt | |
134c85ca | 5955 | = gimple_build_call (gs_info.decl, 5, ptr, mask, op, src, scale); |
3bab6342 AT |
5956 | |
5957 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5958 | ||
5959 | if (prev_stmt_info == NULL) | |
5960 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5961 | else | |
5962 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5963 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5964 | } | |
5965 | return true; | |
5966 | } | |
5967 | ||
2de001ee | 5968 | grouped_store = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
0d0293ac | 5969 | if (grouped_store) |
ebfd146a | 5970 | { |
2de001ee | 5971 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a | 5972 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
e14c1050 | 5973 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 5974 | |
e14c1050 | 5975 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
5976 | |
5977 | /* FORNOW */ | |
a70d6342 | 5978 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
5979 | |
5980 | /* We vectorize all the stmts of the interleaving group when we | |
5981 | reach the last stmt in the group. */ | |
e14c1050 IR |
5982 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
5983 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
5984 | && !slp) |
5985 | { | |
5986 | *vec_stmt = NULL; | |
5987 | return true; | |
5988 | } | |
5989 | ||
5990 | if (slp) | |
4b5caab7 | 5991 | { |
0d0293ac | 5992 | grouped_store = false; |
4b5caab7 IR |
5993 | /* VEC_NUM is the number of vect stmts to be created for this |
5994 | group. */ | |
5995 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 5996 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
52eab378 | 5997 | gcc_assert (GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt)) == first_stmt); |
4b5caab7 | 5998 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 5999 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 6000 | } |
ebfd146a | 6001 | else |
4b5caab7 IR |
6002 | /* VEC_NUM is the number of vect stmts to be created for this |
6003 | group. */ | |
ebfd146a | 6004 | vec_num = group_size; |
44fc7854 BE |
6005 | |
6006 | ref_type = get_group_alias_ptr_type (first_stmt); | |
ebfd146a | 6007 | } |
b8698a0f | 6008 | else |
ebfd146a IR |
6009 | { |
6010 | first_stmt = stmt; | |
6011 | first_dr = dr; | |
6012 | group_size = vec_num = 1; | |
44fc7854 | 6013 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
ebfd146a | 6014 | } |
b8698a0f | 6015 | |
73fbfcad | 6016 | if (dump_enabled_p ()) |
78c60e3d | 6017 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6018 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 6019 | |
2de001ee RS |
6020 | if (memory_access_type == VMAT_ELEMENTWISE |
6021 | || memory_access_type == VMAT_STRIDED_SLP) | |
f2e2a985 MM |
6022 | { |
6023 | gimple_stmt_iterator incr_gsi; | |
6024 | bool insert_after; | |
355fe088 | 6025 | gimple *incr; |
f2e2a985 MM |
6026 | tree offvar; |
6027 | tree ivstep; | |
6028 | tree running_off; | |
6029 | gimple_seq stmts = NULL; | |
6030 | tree stride_base, stride_step, alias_off; | |
6031 | tree vec_oprnd; | |
f502d50e | 6032 | unsigned int g; |
f2e2a985 MM |
6033 | |
6034 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); | |
6035 | ||
6036 | stride_base | |
6037 | = fold_build_pointer_plus | |
f502d50e | 6038 | (unshare_expr (DR_BASE_ADDRESS (first_dr)), |
f2e2a985 | 6039 | size_binop (PLUS_EXPR, |
f502d50e | 6040 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (first_dr))), |
44fc7854 | 6041 | convert_to_ptrofftype (DR_INIT (first_dr)))); |
f502d50e | 6042 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (first_dr))); |
f2e2a985 MM |
6043 | |
6044 | /* For a store with loop-invariant (but other than power-of-2) | |
6045 | stride (i.e. not a grouped access) like so: | |
6046 | ||
6047 | for (i = 0; i < n; i += stride) | |
6048 | array[i] = ...; | |
6049 | ||
6050 | we generate a new induction variable and new stores from | |
6051 | the components of the (vectorized) rhs: | |
6052 | ||
6053 | for (j = 0; ; j += VF*stride) | |
6054 | vectemp = ...; | |
6055 | tmp1 = vectemp[0]; | |
6056 | array[j] = tmp1; | |
6057 | tmp2 = vectemp[1]; | |
6058 | array[j + stride] = tmp2; | |
6059 | ... | |
6060 | */ | |
6061 | ||
cee62fee | 6062 | unsigned nstores = nunits; |
b17dc4d4 | 6063 | unsigned lnel = 1; |
cee62fee | 6064 | tree ltype = elem_type; |
04199738 | 6065 | tree lvectype = vectype; |
cee62fee MM |
6066 | if (slp) |
6067 | { | |
b17dc4d4 RB |
6068 | if (group_size < nunits |
6069 | && nunits % group_size == 0) | |
6070 | { | |
6071 | nstores = nunits / group_size; | |
6072 | lnel = group_size; | |
6073 | ltype = build_vector_type (elem_type, group_size); | |
04199738 RB |
6074 | lvectype = vectype; |
6075 | ||
6076 | /* First check if vec_extract optab doesn't support extraction | |
6077 | of vector elts directly. */ | |
b397965c | 6078 | scalar_mode elmode = SCALAR_TYPE_MODE (elem_type); |
9da15d40 RS |
6079 | machine_mode vmode; |
6080 | if (!mode_for_vector (elmode, group_size).exists (&vmode) | |
6081 | || !VECTOR_MODE_P (vmode) | |
04199738 RB |
6082 | || (convert_optab_handler (vec_extract_optab, |
6083 | TYPE_MODE (vectype), vmode) | |
6084 | == CODE_FOR_nothing)) | |
6085 | { | |
6086 | /* Try to avoid emitting an extract of vector elements | |
6087 | by performing the extracts using an integer type of the | |
6088 | same size, extracting from a vector of those and then | |
6089 | re-interpreting it as the original vector type if | |
6090 | supported. */ | |
6091 | unsigned lsize | |
6092 | = group_size * GET_MODE_BITSIZE (elmode); | |
fffbab82 | 6093 | elmode = int_mode_for_size (lsize, 0).require (); |
04199738 RB |
6094 | /* If we can't construct such a vector fall back to |
6095 | element extracts from the original vector type and | |
6096 | element size stores. */ | |
9da15d40 RS |
6097 | if (mode_for_vector (elmode, |
6098 | nunits / group_size).exists (&vmode) | |
6099 | && VECTOR_MODE_P (vmode) | |
04199738 RB |
6100 | && (convert_optab_handler (vec_extract_optab, |
6101 | vmode, elmode) | |
6102 | != CODE_FOR_nothing)) | |
6103 | { | |
6104 | nstores = nunits / group_size; | |
6105 | lnel = group_size; | |
6106 | ltype = build_nonstandard_integer_type (lsize, 1); | |
6107 | lvectype = build_vector_type (ltype, nstores); | |
6108 | } | |
6109 | /* Else fall back to vector extraction anyway. | |
6110 | Fewer stores are more important than avoiding spilling | |
6111 | of the vector we extract from. Compared to the | |
6112 | construction case in vectorizable_load no store-forwarding | |
6113 | issue exists here for reasonable archs. */ | |
6114 | } | |
b17dc4d4 RB |
6115 | } |
6116 | else if (group_size >= nunits | |
6117 | && group_size % nunits == 0) | |
6118 | { | |
6119 | nstores = 1; | |
6120 | lnel = nunits; | |
6121 | ltype = vectype; | |
04199738 | 6122 | lvectype = vectype; |
b17dc4d4 | 6123 | } |
cee62fee MM |
6124 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); |
6125 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
6126 | } | |
6127 | ||
f2e2a985 MM |
6128 | ivstep = stride_step; |
6129 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
b17dc4d4 | 6130 | build_int_cst (TREE_TYPE (ivstep), vf)); |
f2e2a985 MM |
6131 | |
6132 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6133 | ||
6134 | create_iv (stride_base, ivstep, NULL, | |
6135 | loop, &incr_gsi, insert_after, | |
6136 | &offvar, NULL); | |
6137 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 6138 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
f2e2a985 MM |
6139 | |
6140 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
6141 | if (stmts) | |
6142 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
6143 | ||
6144 | prev_stmt_info = NULL; | |
44fc7854 | 6145 | alias_off = build_int_cst (ref_type, 0); |
f502d50e MM |
6146 | next_stmt = first_stmt; |
6147 | for (g = 0; g < group_size; g++) | |
f2e2a985 | 6148 | { |
f502d50e MM |
6149 | running_off = offvar; |
6150 | if (g) | |
f2e2a985 | 6151 | { |
f502d50e MM |
6152 | tree size = TYPE_SIZE_UNIT (ltype); |
6153 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 6154 | size); |
f502d50e | 6155 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 6156 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 6157 | running_off, pos); |
f2e2a985 | 6158 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 6159 | running_off = newoff; |
f502d50e | 6160 | } |
b17dc4d4 RB |
6161 | unsigned int group_el = 0; |
6162 | unsigned HOST_WIDE_INT | |
6163 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
f502d50e MM |
6164 | for (j = 0; j < ncopies; j++) |
6165 | { | |
6166 | /* We've set op and dt above, from gimple_assign_rhs1(stmt), | |
6167 | and first_stmt == stmt. */ | |
6168 | if (j == 0) | |
6169 | { | |
6170 | if (slp) | |
6171 | { | |
6172 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
306b0c92 | 6173 | slp_node); |
f502d50e MM |
6174 | vec_oprnd = vec_oprnds[0]; |
6175 | } | |
6176 | else | |
6177 | { | |
6178 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
6179 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 6180 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
f502d50e MM |
6181 | } |
6182 | } | |
f2e2a985 | 6183 | else |
f502d50e MM |
6184 | { |
6185 | if (slp) | |
6186 | vec_oprnd = vec_oprnds[j]; | |
6187 | else | |
c079cbac | 6188 | { |
81c40241 | 6189 | vect_is_simple_use (vec_oprnd, vinfo, &def_stmt, &dt); |
c079cbac RB |
6190 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); |
6191 | } | |
f502d50e | 6192 | } |
04199738 RB |
6193 | /* Pun the vector to extract from if necessary. */ |
6194 | if (lvectype != vectype) | |
6195 | { | |
6196 | tree tem = make_ssa_name (lvectype); | |
6197 | gimple *pun | |
6198 | = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
6199 | lvectype, vec_oprnd)); | |
6200 | vect_finish_stmt_generation (stmt, pun, gsi); | |
6201 | vec_oprnd = tem; | |
6202 | } | |
f502d50e MM |
6203 | for (i = 0; i < nstores; i++) |
6204 | { | |
6205 | tree newref, newoff; | |
355fe088 | 6206 | gimple *incr, *assign; |
f502d50e MM |
6207 | tree size = TYPE_SIZE (ltype); |
6208 | /* Extract the i'th component. */ | |
6209 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
6210 | bitsize_int (i), size); | |
6211 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
6212 | size, pos); | |
6213 | ||
6214 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
6215 | NULL_TREE, true, | |
6216 | GSI_SAME_STMT); | |
6217 | ||
b17dc4d4 RB |
6218 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
6219 | group_el * elsz); | |
f502d50e | 6220 | newref = build2 (MEM_REF, ltype, |
b17dc4d4 | 6221 | running_off, this_off); |
f502d50e MM |
6222 | |
6223 | /* And store it to *running_off. */ | |
6224 | assign = gimple_build_assign (newref, elem); | |
6225 | vect_finish_stmt_generation (stmt, assign, gsi); | |
6226 | ||
b17dc4d4 RB |
6227 | group_el += lnel; |
6228 | if (! slp | |
6229 | || group_el == group_size) | |
6230 | { | |
6231 | newoff = copy_ssa_name (running_off, NULL); | |
6232 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6233 | running_off, stride_step); | |
6234 | vect_finish_stmt_generation (stmt, incr, gsi); | |
f502d50e | 6235 | |
b17dc4d4 RB |
6236 | running_off = newoff; |
6237 | group_el = 0; | |
6238 | } | |
225ce44b RB |
6239 | if (g == group_size - 1 |
6240 | && !slp) | |
f502d50e MM |
6241 | { |
6242 | if (j == 0 && i == 0) | |
225ce44b RB |
6243 | STMT_VINFO_VEC_STMT (stmt_info) |
6244 | = *vec_stmt = assign; | |
f502d50e MM |
6245 | else |
6246 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign; | |
6247 | prev_stmt_info = vinfo_for_stmt (assign); | |
6248 | } | |
6249 | } | |
f2e2a985 | 6250 | } |
f502d50e | 6251 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
b17dc4d4 RB |
6252 | if (slp) |
6253 | break; | |
f2e2a985 | 6254 | } |
778dd3b6 RB |
6255 | |
6256 | vec_oprnds.release (); | |
f2e2a985 MM |
6257 | return true; |
6258 | } | |
6259 | ||
8c681247 | 6260 | auto_vec<tree> dr_chain (group_size); |
9771b263 | 6261 | oprnds.create (group_size); |
ebfd146a | 6262 | |
720f5239 | 6263 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6264 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
6265 | /* Targets with store-lane instructions must not require explicit |
6266 | realignment. */ | |
2de001ee | 6267 | gcc_assert (memory_access_type != VMAT_LOAD_STORE_LANES |
272c6793 RS |
6268 | || alignment_support_scheme == dr_aligned |
6269 | || alignment_support_scheme == dr_unaligned_supported); | |
6270 | ||
62da9e14 RS |
6271 | if (memory_access_type == VMAT_CONTIGUOUS_DOWN |
6272 | || memory_access_type == VMAT_CONTIGUOUS_REVERSE) | |
09dfa495 BM |
6273 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
6274 | ||
2de001ee | 6275 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
272c6793 RS |
6276 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); |
6277 | else | |
6278 | aggr_type = vectype; | |
ebfd146a IR |
6279 | |
6280 | /* In case the vectorization factor (VF) is bigger than the number | |
6281 | of elements that we can fit in a vectype (nunits), we have to generate | |
6282 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 6283 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
6284 | vect_get_vec_def_for_copy_stmt. */ |
6285 | ||
0d0293ac | 6286 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6287 | |
6288 | S1: &base + 2 = x2 | |
6289 | S2: &base = x0 | |
6290 | S3: &base + 1 = x1 | |
6291 | S4: &base + 3 = x3 | |
6292 | ||
6293 | We create vectorized stores starting from base address (the access of the | |
6294 | first stmt in the chain (S2 in the above example), when the last store stmt | |
6295 | of the chain (S4) is reached: | |
6296 | ||
6297 | VS1: &base = vx2 | |
6298 | VS2: &base + vec_size*1 = vx0 | |
6299 | VS3: &base + vec_size*2 = vx1 | |
6300 | VS4: &base + vec_size*3 = vx3 | |
6301 | ||
6302 | Then permutation statements are generated: | |
6303 | ||
3fcc1b55 JJ |
6304 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
6305 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 6306 | ... |
b8698a0f | 6307 | |
ebfd146a IR |
6308 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
6309 | (the order of the data-refs in the output of vect_permute_store_chain | |
6310 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6311 | the documentation of vect_permute_store_chain()). | |
6312 | ||
6313 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 6314 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 6315 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 6316 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
6317 | */ |
6318 | ||
6319 | prev_stmt_info = NULL; | |
6320 | for (j = 0; j < ncopies; j++) | |
6321 | { | |
ebfd146a IR |
6322 | |
6323 | if (j == 0) | |
6324 | { | |
6325 | if (slp) | |
6326 | { | |
6327 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c | 6328 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
306b0c92 | 6329 | NULL, slp_node); |
ebfd146a | 6330 | |
9771b263 | 6331 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
6332 | } |
6333 | else | |
6334 | { | |
b8698a0f L |
6335 | /* For interleaved stores we collect vectorized defs for all the |
6336 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
6337 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
6338 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
6339 | ||
0d0293ac | 6340 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 6341 | OPRNDS are of size 1. */ |
b8698a0f | 6342 | next_stmt = first_stmt; |
ebfd146a IR |
6343 | for (i = 0; i < group_size; i++) |
6344 | { | |
b8698a0f L |
6345 | /* Since gaps are not supported for interleaved stores, |
6346 | GROUP_SIZE is the exact number of stmts in the chain. | |
6347 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
6348 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
6349 | iteration of the loop will be executed. */ |
6350 | gcc_assert (next_stmt | |
6351 | && gimple_assign_single_p (next_stmt)); | |
6352 | op = gimple_assign_rhs1 (next_stmt); | |
6353 | ||
81c40241 | 6354 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
9771b263 DN |
6355 | dr_chain.quick_push (vec_oprnd); |
6356 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 6357 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
6358 | } |
6359 | } | |
6360 | ||
6361 | /* We should have catched mismatched types earlier. */ | |
6362 | gcc_assert (useless_type_conversion_p (vectype, | |
6363 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
6364 | bool simd_lane_access_p |
6365 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
6366 | if (simd_lane_access_p | |
6367 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
6368 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
6369 | && integer_zerop (DR_OFFSET (first_dr)) | |
6370 | && integer_zerop (DR_INIT (first_dr)) | |
6371 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 6372 | get_alias_set (TREE_TYPE (ref_type)))) |
74bf76ed JJ |
6373 | { |
6374 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 6375 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 6376 | inv_p = false; |
74bf76ed JJ |
6377 | } |
6378 | else | |
6379 | dataref_ptr | |
6380 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
6381 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 6382 | offset, &dummy, gsi, &ptr_incr, |
74bf76ed | 6383 | simd_lane_access_p, &inv_p); |
a70d6342 | 6384 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 6385 | } |
b8698a0f | 6386 | else |
ebfd146a | 6387 | { |
b8698a0f L |
6388 | /* For interleaved stores we created vectorized defs for all the |
6389 | defs stored in OPRNDS in the previous iteration (previous copy). | |
6390 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
6391 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
6392 | next copy. | |
0d0293ac | 6393 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
6394 | OPRNDS are of size 1. */ |
6395 | for (i = 0; i < group_size; i++) | |
6396 | { | |
9771b263 | 6397 | op = oprnds[i]; |
81c40241 | 6398 | vect_is_simple_use (op, vinfo, &def_stmt, &dt); |
b8698a0f | 6399 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
6400 | dr_chain[i] = vec_oprnd; |
6401 | oprnds[i] = vec_oprnd; | |
ebfd146a | 6402 | } |
74bf76ed JJ |
6403 | if (dataref_offset) |
6404 | dataref_offset | |
6405 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
6406 | TYPE_SIZE_UNIT (aggr_type)); | |
6407 | else | |
6408 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
6409 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
6410 | } |
6411 | ||
2de001ee | 6412 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 6413 | { |
272c6793 | 6414 | tree vec_array; |
267d3070 | 6415 | |
272c6793 RS |
6416 | /* Combine all the vectors into an array. */ |
6417 | vec_array = create_vector_array (vectype, vec_num); | |
6418 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 6419 | { |
9771b263 | 6420 | vec_oprnd = dr_chain[i]; |
272c6793 | 6421 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 6422 | } |
b8698a0f | 6423 | |
272c6793 RS |
6424 | /* Emit: |
6425 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
44fc7854 | 6426 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); |
a844293d RS |
6427 | gcall *call = gimple_build_call_internal (IFN_STORE_LANES, 1, |
6428 | vec_array); | |
6429 | gimple_call_set_lhs (call, data_ref); | |
6430 | gimple_call_set_nothrow (call, true); | |
6431 | new_stmt = call; | |
267d3070 | 6432 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
6433 | } |
6434 | else | |
6435 | { | |
6436 | new_stmt = NULL; | |
0d0293ac | 6437 | if (grouped_store) |
272c6793 | 6438 | { |
b6b9227d JJ |
6439 | if (j == 0) |
6440 | result_chain.create (group_size); | |
272c6793 RS |
6441 | /* Permute. */ |
6442 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
6443 | &result_chain); | |
6444 | } | |
c2d7ab2a | 6445 | |
272c6793 RS |
6446 | next_stmt = first_stmt; |
6447 | for (i = 0; i < vec_num; i++) | |
6448 | { | |
644ffefd | 6449 | unsigned align, misalign; |
272c6793 RS |
6450 | |
6451 | if (i > 0) | |
6452 | /* Bump the vector pointer. */ | |
6453 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
6454 | stmt, NULL_TREE); | |
6455 | ||
6456 | if (slp) | |
9771b263 | 6457 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
6458 | else if (grouped_store) |
6459 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 6460 | vect_permute_store_chain(). */ |
9771b263 | 6461 | vec_oprnd = result_chain[i]; |
272c6793 | 6462 | |
69a2e8a1 | 6463 | data_ref = fold_build2 (MEM_REF, vectype, |
aed93b23 RB |
6464 | dataref_ptr, |
6465 | dataref_offset | |
6466 | ? dataref_offset | |
44fc7854 | 6467 | : build_int_cst (ref_type, 0)); |
f702e7d4 | 6468 | align = DR_TARGET_ALIGNMENT (first_dr); |
272c6793 | 6469 | if (aligned_access_p (first_dr)) |
644ffefd | 6470 | misalign = 0; |
272c6793 RS |
6471 | else if (DR_MISALIGNMENT (first_dr) == -1) |
6472 | { | |
25f68d90 | 6473 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 6474 | misalign = 0; |
272c6793 RS |
6475 | TREE_TYPE (data_ref) |
6476 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 6477 | align * BITS_PER_UNIT); |
272c6793 RS |
6478 | } |
6479 | else | |
6480 | { | |
6481 | TREE_TYPE (data_ref) | |
6482 | = build_aligned_type (TREE_TYPE (data_ref), | |
6483 | TYPE_ALIGN (elem_type)); | |
644ffefd | 6484 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 6485 | } |
aed93b23 RB |
6486 | if (dataref_offset == NULL_TREE |
6487 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
6488 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
6489 | misalign); | |
c2d7ab2a | 6490 | |
62da9e14 | 6491 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
09dfa495 BM |
6492 | { |
6493 | tree perm_mask = perm_mask_for_reverse (vectype); | |
6494 | tree perm_dest | |
6495 | = vect_create_destination_var (gimple_assign_rhs1 (stmt), | |
6496 | vectype); | |
b731b390 | 6497 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
6498 | |
6499 | /* Generate the permute statement. */ | |
355fe088 | 6500 | gimple *perm_stmt |
0d0e4a03 JJ |
6501 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
6502 | vec_oprnd, perm_mask); | |
09dfa495 BM |
6503 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6504 | ||
6505 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
6506 | vec_oprnd = new_temp; | |
6507 | } | |
6508 | ||
272c6793 RS |
6509 | /* Arguments are ready. Create the new vector stmt. */ |
6510 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
6511 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
6512 | |
6513 | if (slp) | |
6514 | continue; | |
6515 | ||
e14c1050 | 6516 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
6517 | if (!next_stmt) |
6518 | break; | |
6519 | } | |
ebfd146a | 6520 | } |
1da0876c RS |
6521 | if (!slp) |
6522 | { | |
6523 | if (j == 0) | |
6524 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6525 | else | |
6526 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6527 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6528 | } | |
ebfd146a IR |
6529 | } |
6530 | ||
9771b263 DN |
6531 | oprnds.release (); |
6532 | result_chain.release (); | |
6533 | vec_oprnds.release (); | |
ebfd146a IR |
6534 | |
6535 | return true; | |
6536 | } | |
6537 | ||
557be5a8 AL |
6538 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
6539 | VECTOR_CST mask. No checks are made that the target platform supports the | |
7ac7e286 | 6540 | mask, so callers may wish to test can_vec_perm_const_p separately, or use |
557be5a8 | 6541 | vect_gen_perm_mask_checked. */ |
a1e53f3f | 6542 | |
3fcc1b55 | 6543 | tree |
4aae3cb3 | 6544 | vect_gen_perm_mask_any (tree vectype, const vec_perm_indices &sel) |
a1e53f3f | 6545 | { |
b00cb3bf | 6546 | tree mask_type; |
a1e53f3f | 6547 | |
b00cb3bf RS |
6548 | unsigned int nunits = sel.length (); |
6549 | gcc_assert (nunits == TYPE_VECTOR_SUBPARTS (vectype)); | |
6550 | ||
6551 | mask_type = build_vector_type (ssizetype, nunits); | |
736d0f28 | 6552 | return vec_perm_indices_to_tree (mask_type, sel); |
a1e53f3f L |
6553 | } |
6554 | ||
7ac7e286 | 6555 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_const_p, |
cf7aa6a3 | 6556 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ |
557be5a8 AL |
6557 | |
6558 | tree | |
4aae3cb3 | 6559 | vect_gen_perm_mask_checked (tree vectype, const vec_perm_indices &sel) |
557be5a8 | 6560 | { |
7ac7e286 | 6561 | gcc_assert (can_vec_perm_const_p (TYPE_MODE (vectype), sel)); |
557be5a8 AL |
6562 | return vect_gen_perm_mask_any (vectype, sel); |
6563 | } | |
6564 | ||
aec7ae7d JJ |
6565 | /* Given a vector variable X and Y, that was generated for the scalar |
6566 | STMT, generate instructions to permute the vector elements of X and Y | |
6567 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
6568 | permuted vector variable. */ | |
a1e53f3f L |
6569 | |
6570 | static tree | |
355fe088 | 6571 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 6572 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
6573 | { |
6574 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 6575 | tree perm_dest, data_ref; |
355fe088 | 6576 | gimple *perm_stmt; |
a1e53f3f | 6577 | |
acdcd61b | 6578 | perm_dest = vect_create_destination_var (gimple_get_lhs (stmt), vectype); |
b731b390 | 6579 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
6580 | |
6581 | /* Generate the permute statement. */ | |
0d0e4a03 | 6582 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
6583 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6584 | ||
6585 | return data_ref; | |
6586 | } | |
6587 | ||
6b916b36 RB |
6588 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
6589 | inserting them on the loops preheader edge. Returns true if we | |
6590 | were successful in doing so (and thus STMT can be moved then), | |
6591 | otherwise returns false. */ | |
6592 | ||
6593 | static bool | |
355fe088 | 6594 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
6595 | { |
6596 | ssa_op_iter i; | |
6597 | tree op; | |
6598 | bool any = false; | |
6599 | ||
6600 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6601 | { | |
355fe088 | 6602 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6603 | if (!gimple_nop_p (def_stmt) |
6604 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6605 | { | |
6606 | /* Make sure we don't need to recurse. While we could do | |
6607 | so in simple cases when there are more complex use webs | |
6608 | we don't have an easy way to preserve stmt order to fulfil | |
6609 | dependencies within them. */ | |
6610 | tree op2; | |
6611 | ssa_op_iter i2; | |
d1417442 JJ |
6612 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
6613 | return false; | |
6b916b36 RB |
6614 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
6615 | { | |
355fe088 | 6616 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
6617 | if (!gimple_nop_p (def_stmt2) |
6618 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
6619 | return false; | |
6620 | } | |
6621 | any = true; | |
6622 | } | |
6623 | } | |
6624 | ||
6625 | if (!any) | |
6626 | return true; | |
6627 | ||
6628 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6629 | { | |
355fe088 | 6630 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6631 | if (!gimple_nop_p (def_stmt) |
6632 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6633 | { | |
6634 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
6635 | gsi_remove (&gsi, false); | |
6636 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
6637 | } | |
6638 | } | |
6639 | ||
6640 | return true; | |
6641 | } | |
6642 | ||
ebfd146a IR |
6643 | /* vectorizable_load. |
6644 | ||
b8698a0f L |
6645 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
6646 | can be vectorized. | |
6647 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6648 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6649 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6650 | ||
6651 | static bool | |
355fe088 | 6652 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 6653 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
6654 | { |
6655 | tree scalar_dest; | |
6656 | tree vec_dest = NULL; | |
6657 | tree data_ref = NULL; | |
6658 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 6659 | stmt_vec_info prev_stmt_info; |
ebfd146a | 6660 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6661 | struct loop *loop = NULL; |
ebfd146a | 6662 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 6663 | bool nested_in_vect_loop = false; |
c716e67f | 6664 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 6665 | tree elem_type; |
ebfd146a | 6666 | tree new_temp; |
ef4bddc2 | 6667 | machine_mode mode; |
355fe088 | 6668 | gimple *new_stmt = NULL; |
ebfd146a IR |
6669 | tree dummy; |
6670 | enum dr_alignment_support alignment_support_scheme; | |
6671 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6672 | tree dataref_offset = NULL_TREE; |
355fe088 | 6673 | gimple *ptr_incr = NULL; |
ebfd146a | 6674 | int ncopies; |
d9f21f6a RS |
6675 | int i, j, group_size; |
6676 | poly_int64 group_gap_adj; | |
ebfd146a IR |
6677 | tree msq = NULL_TREE, lsq; |
6678 | tree offset = NULL_TREE; | |
356bbc4c | 6679 | tree byte_offset = NULL_TREE; |
ebfd146a | 6680 | tree realignment_token = NULL_TREE; |
538dd0b7 | 6681 | gphi *phi = NULL; |
6e1aa848 | 6682 | vec<tree> dr_chain = vNULL; |
0d0293ac | 6683 | bool grouped_load = false; |
355fe088 | 6684 | gimple *first_stmt; |
4f0a0218 | 6685 | gimple *first_stmt_for_drptr = NULL; |
ebfd146a IR |
6686 | bool inv_p; |
6687 | bool compute_in_loop = false; | |
6688 | struct loop *at_loop; | |
6689 | int vec_num; | |
6690 | bool slp = (slp_node != NULL); | |
6691 | bool slp_perm = false; | |
6692 | enum tree_code code; | |
a70d6342 | 6693 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
d9f21f6a | 6694 | poly_uint64 vf; |
272c6793 | 6695 | tree aggr_type; |
134c85ca | 6696 | gather_scatter_info gs_info; |
310213d4 | 6697 | vec_info *vinfo = stmt_info->vinfo; |
44fc7854 | 6698 | tree ref_type; |
a70d6342 | 6699 | |
465c8c19 JJ |
6700 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
6701 | return false; | |
6702 | ||
66c16fd9 RB |
6703 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6704 | && ! vec_stmt) | |
465c8c19 JJ |
6705 | return false; |
6706 | ||
6707 | /* Is vectorizable load? */ | |
6708 | if (!is_gimple_assign (stmt)) | |
6709 | return false; | |
6710 | ||
6711 | scalar_dest = gimple_assign_lhs (stmt); | |
6712 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
6713 | return false; | |
6714 | ||
6715 | code = gimple_assign_rhs_code (stmt); | |
6716 | if (code != ARRAY_REF | |
6717 | && code != BIT_FIELD_REF | |
6718 | && code != INDIRECT_REF | |
6719 | && code != COMPONENT_REF | |
6720 | && code != IMAGPART_EXPR | |
6721 | && code != REALPART_EXPR | |
6722 | && code != MEM_REF | |
6723 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
6724 | return false; | |
6725 | ||
6726 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
6727 | return false; | |
6728 | ||
6729 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
6730 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
6731 | ||
a70d6342 IR |
6732 | if (loop_vinfo) |
6733 | { | |
6734 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6735 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
6736 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6737 | } | |
6738 | else | |
3533e503 | 6739 | vf = 1; |
ebfd146a IR |
6740 | |
6741 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 6742 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 6743 | case of SLP. */ |
fce57248 | 6744 | if (slp) |
ebfd146a IR |
6745 | ncopies = 1; |
6746 | else | |
e8f142e2 | 6747 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
ebfd146a IR |
6748 | |
6749 | gcc_assert (ncopies >= 1); | |
6750 | ||
6751 | /* FORNOW. This restriction should be relaxed. */ | |
6752 | if (nested_in_vect_loop && ncopies > 1) | |
6753 | { | |
73fbfcad | 6754 | if (dump_enabled_p ()) |
78c60e3d | 6755 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6756 | "multiple types in nested loop.\n"); |
ebfd146a IR |
6757 | return false; |
6758 | } | |
6759 | ||
f2556b68 RB |
6760 | /* Invalidate assumptions made by dependence analysis when vectorization |
6761 | on the unrolled body effectively re-orders stmts. */ | |
6762 | if (ncopies > 1 | |
6763 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
6764 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
6765 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
6766 | { |
6767 | if (dump_enabled_p ()) | |
6768 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6769 | "cannot perform implicit CSE when unrolling " | |
6770 | "with negative dependence distance\n"); | |
6771 | return false; | |
6772 | } | |
6773 | ||
7b7b1813 | 6774 | elem_type = TREE_TYPE (vectype); |
947131ba | 6775 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
6776 | |
6777 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6778 | (e.g. - data copies). */ | |
947131ba | 6779 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 6780 | { |
73fbfcad | 6781 | if (dump_enabled_p ()) |
78c60e3d | 6782 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6783 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
6784 | return false; |
6785 | } | |
6786 | ||
ebfd146a | 6787 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 6788 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 6789 | { |
0d0293ac | 6790 | grouped_load = true; |
ebfd146a | 6791 | /* FORNOW */ |
2de001ee RS |
6792 | gcc_assert (!nested_in_vect_loop); |
6793 | gcc_assert (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)); | |
ebfd146a | 6794 | |
e14c1050 | 6795 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
d3465d72 | 6796 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
d5f035ea | 6797 | |
b1af7da6 RB |
6798 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
6799 | slp_perm = true; | |
6800 | ||
f2556b68 RB |
6801 | /* Invalidate assumptions made by dependence analysis when vectorization |
6802 | on the unrolled body effectively re-orders stmts. */ | |
6803 | if (!PURE_SLP_STMT (stmt_info) | |
6804 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
6805 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
6806 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
6807 | { |
6808 | if (dump_enabled_p ()) | |
6809 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6810 | "cannot perform implicit CSE when performing " | |
6811 | "group loads with negative dependence distance\n"); | |
6812 | return false; | |
6813 | } | |
96bb56b2 RB |
6814 | |
6815 | /* Similarly when the stmt is a load that is both part of a SLP | |
6816 | instance and a loop vectorized stmt via the same-dr mechanism | |
6817 | we have to give up. */ | |
6818 | if (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info) | |
6819 | && (STMT_SLP_TYPE (stmt_info) | |
6820 | != STMT_SLP_TYPE (vinfo_for_stmt | |
6821 | (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info))))) | |
6822 | { | |
6823 | if (dump_enabled_p ()) | |
6824 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6825 | "conflicting SLP types for CSEd load\n"); | |
6826 | return false; | |
6827 | } | |
ebfd146a IR |
6828 | } |
6829 | ||
2de001ee | 6830 | vect_memory_access_type memory_access_type; |
62da9e14 | 6831 | if (!get_load_store_type (stmt, vectype, slp, VLS_LOAD, ncopies, |
2de001ee RS |
6832 | &memory_access_type, &gs_info)) |
6833 | return false; | |
a1e53f3f | 6834 | |
ebfd146a IR |
6835 | if (!vec_stmt) /* transformation not required. */ |
6836 | { | |
2de001ee RS |
6837 | if (!slp) |
6838 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; | |
ebfd146a | 6839 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; |
2e8ab70c RB |
6840 | /* The SLP costs are calculated during SLP analysis. */ |
6841 | if (!PURE_SLP_STMT (stmt_info)) | |
2de001ee | 6842 | vect_model_load_cost (stmt_info, ncopies, memory_access_type, |
2e8ab70c | 6843 | NULL, NULL, NULL); |
ebfd146a IR |
6844 | return true; |
6845 | } | |
6846 | ||
2de001ee RS |
6847 | if (!slp) |
6848 | gcc_assert (memory_access_type | |
6849 | == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); | |
6850 | ||
73fbfcad | 6851 | if (dump_enabled_p ()) |
78c60e3d | 6852 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6853 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a | 6854 | |
67b8dbac | 6855 | /* Transform. */ |
ebfd146a | 6856 | |
f702e7d4 | 6857 | ensure_base_align (dr); |
c716e67f | 6858 | |
2de001ee | 6859 | if (memory_access_type == VMAT_GATHER_SCATTER) |
aec7ae7d JJ |
6860 | { |
6861 | tree vec_oprnd0 = NULL_TREE, op; | |
134c85ca | 6862 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
aec7ae7d | 6863 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; |
d3c2fee0 | 6864 | tree ptr, mask, var, scale, merge, perm_mask = NULL_TREE, prev_res = NULL_TREE; |
aec7ae7d JJ |
6865 | edge pe = loop_preheader_edge (loop); |
6866 | gimple_seq seq; | |
6867 | basic_block new_bb; | |
6868 | enum { NARROW, NONE, WIDEN } modifier; | |
134c85ca | 6869 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype); |
aec7ae7d JJ |
6870 | |
6871 | if (nunits == gather_off_nunits) | |
6872 | modifier = NONE; | |
6873 | else if (nunits == gather_off_nunits / 2) | |
6874 | { | |
aec7ae7d JJ |
6875 | modifier = WIDEN; |
6876 | ||
e3342de4 | 6877 | vec_perm_builder sel (gather_off_nunits, gather_off_nunits, 1); |
aec7ae7d | 6878 | for (i = 0; i < gather_off_nunits; ++i) |
908a1a16 | 6879 | sel.quick_push (i | nunits); |
aec7ae7d | 6880 | |
e3342de4 RS |
6881 | vec_perm_indices indices (sel, 1, gather_off_nunits); |
6882 | perm_mask = vect_gen_perm_mask_checked (gs_info.offset_vectype, | |
6883 | indices); | |
aec7ae7d JJ |
6884 | } |
6885 | else if (nunits == gather_off_nunits * 2) | |
6886 | { | |
aec7ae7d JJ |
6887 | modifier = NARROW; |
6888 | ||
e3342de4 | 6889 | vec_perm_builder sel (nunits, nunits, 1); |
aec7ae7d | 6890 | for (i = 0; i < nunits; ++i) |
908a1a16 RS |
6891 | sel.quick_push (i < gather_off_nunits |
6892 | ? i : i + nunits - gather_off_nunits); | |
aec7ae7d | 6893 | |
e3342de4 RS |
6894 | vec_perm_indices indices (sel, 2, nunits); |
6895 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); | |
aec7ae7d JJ |
6896 | ncopies *= 2; |
6897 | } | |
6898 | else | |
6899 | gcc_unreachable (); | |
6900 | ||
134c85ca | 6901 | rettype = TREE_TYPE (TREE_TYPE (gs_info.decl)); |
aec7ae7d JJ |
6902 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
6903 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6904 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6905 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6906 | scaletype = TREE_VALUE (arglist); | |
d3c2fee0 | 6907 | gcc_checking_assert (types_compatible_p (srctype, rettype)); |
aec7ae7d JJ |
6908 | |
6909 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
6910 | ||
134c85ca | 6911 | ptr = fold_convert (ptrtype, gs_info.base); |
aec7ae7d JJ |
6912 | if (!is_gimple_min_invariant (ptr)) |
6913 | { | |
6914 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6915 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6916 | gcc_assert (!new_bb); | |
6917 | } | |
6918 | ||
6919 | /* Currently we support only unconditional gather loads, | |
6920 | so mask should be all ones. */ | |
d3c2fee0 AI |
6921 | if (TREE_CODE (masktype) == INTEGER_TYPE) |
6922 | mask = build_int_cst (masktype, -1); | |
6923 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
6924 | { | |
6925 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
6926 | mask = build_vector_from_val (masktype, mask); | |
03b9e8e4 | 6927 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
d3c2fee0 | 6928 | } |
aec7ae7d JJ |
6929 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) |
6930 | { | |
6931 | REAL_VALUE_TYPE r; | |
6932 | long tmp[6]; | |
6933 | for (j = 0; j < 6; ++j) | |
6934 | tmp[j] = -1; | |
6935 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
6936 | mask = build_real (TREE_TYPE (masktype), r); | |
d3c2fee0 | 6937 | mask = build_vector_from_val (masktype, mask); |
03b9e8e4 | 6938 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
aec7ae7d JJ |
6939 | } |
6940 | else | |
6941 | gcc_unreachable (); | |
aec7ae7d | 6942 | |
134c85ca | 6943 | scale = build_int_cst (scaletype, gs_info.scale); |
aec7ae7d | 6944 | |
d3c2fee0 AI |
6945 | if (TREE_CODE (TREE_TYPE (rettype)) == INTEGER_TYPE) |
6946 | merge = build_int_cst (TREE_TYPE (rettype), 0); | |
6947 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (rettype))) | |
6948 | { | |
6949 | REAL_VALUE_TYPE r; | |
6950 | long tmp[6]; | |
6951 | for (j = 0; j < 6; ++j) | |
6952 | tmp[j] = 0; | |
6953 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (rettype))); | |
6954 | merge = build_real (TREE_TYPE (rettype), r); | |
6955 | } | |
6956 | else | |
6957 | gcc_unreachable (); | |
6958 | merge = build_vector_from_val (rettype, merge); | |
6959 | merge = vect_init_vector (stmt, merge, rettype, NULL); | |
6960 | ||
aec7ae7d JJ |
6961 | prev_stmt_info = NULL; |
6962 | for (j = 0; j < ncopies; ++j) | |
6963 | { | |
6964 | if (modifier == WIDEN && (j & 1)) | |
6965 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
6966 | perm_mask, stmt, gsi); | |
6967 | else if (j == 0) | |
6968 | op = vec_oprnd0 | |
134c85ca | 6969 | = vect_get_vec_def_for_operand (gs_info.offset, stmt); |
aec7ae7d JJ |
6970 | else |
6971 | op = vec_oprnd0 | |
134c85ca | 6972 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, vec_oprnd0); |
aec7ae7d JJ |
6973 | |
6974 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6975 | { | |
6976 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
6977 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 6978 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
aec7ae7d JJ |
6979 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
6980 | new_stmt | |
0d0e4a03 | 6981 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6982 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6983 | op = var; | |
6984 | } | |
6985 | ||
6986 | new_stmt | |
134c85ca | 6987 | = gimple_build_call (gs_info.decl, 5, merge, ptr, op, mask, scale); |
aec7ae7d JJ |
6988 | |
6989 | if (!useless_type_conversion_p (vectype, rettype)) | |
6990 | { | |
6991 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
6992 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 6993 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
aec7ae7d JJ |
6994 | gimple_call_set_lhs (new_stmt, op); |
6995 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 6996 | var = make_ssa_name (vec_dest); |
aec7ae7d JJ |
6997 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
6998 | new_stmt | |
0d0e4a03 | 6999 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
7000 | } |
7001 | else | |
7002 | { | |
7003 | var = make_ssa_name (vec_dest, new_stmt); | |
7004 | gimple_call_set_lhs (new_stmt, var); | |
7005 | } | |
7006 | ||
7007 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7008 | ||
7009 | if (modifier == NARROW) | |
7010 | { | |
7011 | if ((j & 1) == 0) | |
7012 | { | |
7013 | prev_res = var; | |
7014 | continue; | |
7015 | } | |
7016 | var = permute_vec_elements (prev_res, var, | |
7017 | perm_mask, stmt, gsi); | |
7018 | new_stmt = SSA_NAME_DEF_STMT (var); | |
7019 | } | |
7020 | ||
7021 | if (prev_stmt_info == NULL) | |
7022 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7023 | else | |
7024 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7025 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7026 | } | |
7027 | return true; | |
7028 | } | |
2de001ee RS |
7029 | |
7030 | if (memory_access_type == VMAT_ELEMENTWISE | |
7031 | || memory_access_type == VMAT_STRIDED_SLP) | |
7d75abc8 MM |
7032 | { |
7033 | gimple_stmt_iterator incr_gsi; | |
7034 | bool insert_after; | |
355fe088 | 7035 | gimple *incr; |
7d75abc8 | 7036 | tree offvar; |
7d75abc8 MM |
7037 | tree ivstep; |
7038 | tree running_off; | |
9771b263 | 7039 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 7040 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
7041 | tree stride_base, stride_step, alias_off; |
7042 | ||
7043 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 7044 | |
f502d50e | 7045 | if (slp && grouped_load) |
44fc7854 BE |
7046 | { |
7047 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); | |
7048 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
7049 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
7050 | ref_type = get_group_alias_ptr_type (first_stmt); | |
7051 | } | |
ab313a8c | 7052 | else |
44fc7854 BE |
7053 | { |
7054 | first_stmt = stmt; | |
7055 | first_dr = dr; | |
7056 | group_size = 1; | |
7057 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
7058 | } | |
ab313a8c | 7059 | |
14ac6aa2 RB |
7060 | stride_base |
7061 | = fold_build_pointer_plus | |
ab313a8c | 7062 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 7063 | size_binop (PLUS_EXPR, |
ab313a8c RB |
7064 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
7065 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
7066 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
7067 | |
7068 | /* For a load with loop-invariant (but other than power-of-2) | |
7069 | stride (i.e. not a grouped access) like so: | |
7070 | ||
7071 | for (i = 0; i < n; i += stride) | |
7072 | ... = array[i]; | |
7073 | ||
7074 | we generate a new induction variable and new accesses to | |
7075 | form a new vector (or vectors, depending on ncopies): | |
7076 | ||
7077 | for (j = 0; ; j += VF*stride) | |
7078 | tmp1 = array[j]; | |
7079 | tmp2 = array[j + stride]; | |
7080 | ... | |
7081 | vectemp = {tmp1, tmp2, ...} | |
7082 | */ | |
7083 | ||
ab313a8c RB |
7084 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
7085 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
7086 | |
7087 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
7088 | ||
ab313a8c | 7089 | create_iv (unshare_expr (stride_base), unshare_expr (ivstep), NULL, |
7d75abc8 MM |
7090 | loop, &incr_gsi, insert_after, |
7091 | &offvar, NULL); | |
7092 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 7093 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
7d75abc8 | 7094 | |
ab313a8c RB |
7095 | stride_step = force_gimple_operand (unshare_expr (stride_step), |
7096 | &stmts, true, NULL_TREE); | |
7d75abc8 MM |
7097 | if (stmts) |
7098 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
7099 | ||
7100 | prev_stmt_info = NULL; | |
7101 | running_off = offvar; | |
44fc7854 | 7102 | alias_off = build_int_cst (ref_type, 0); |
7b5fc413 | 7103 | int nloads = nunits; |
e09b4c37 | 7104 | int lnel = 1; |
7b5fc413 | 7105 | tree ltype = TREE_TYPE (vectype); |
ea60dd34 | 7106 | tree lvectype = vectype; |
b266b968 | 7107 | auto_vec<tree> dr_chain; |
2de001ee | 7108 | if (memory_access_type == VMAT_STRIDED_SLP) |
7b5fc413 | 7109 | { |
2de001ee | 7110 | if (group_size < nunits) |
e09b4c37 | 7111 | { |
ff03930a JJ |
7112 | /* First check if vec_init optab supports construction from |
7113 | vector elts directly. */ | |
b397965c | 7114 | scalar_mode elmode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); |
9da15d40 RS |
7115 | machine_mode vmode; |
7116 | if (mode_for_vector (elmode, group_size).exists (&vmode) | |
7117 | && VECTOR_MODE_P (vmode) | |
ff03930a JJ |
7118 | && (convert_optab_handler (vec_init_optab, |
7119 | TYPE_MODE (vectype), vmode) | |
7120 | != CODE_FOR_nothing)) | |
ea60dd34 RB |
7121 | { |
7122 | nloads = nunits / group_size; | |
7123 | lnel = group_size; | |
ff03930a JJ |
7124 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); |
7125 | } | |
7126 | else | |
7127 | { | |
7128 | /* Otherwise avoid emitting a constructor of vector elements | |
7129 | by performing the loads using an integer type of the same | |
7130 | size, constructing a vector of those and then | |
7131 | re-interpreting it as the original vector type. | |
7132 | This avoids a huge runtime penalty due to the general | |
7133 | inability to perform store forwarding from smaller stores | |
7134 | to a larger load. */ | |
7135 | unsigned lsize | |
7136 | = group_size * TYPE_PRECISION (TREE_TYPE (vectype)); | |
fffbab82 | 7137 | elmode = int_mode_for_size (lsize, 0).require (); |
ff03930a JJ |
7138 | /* If we can't construct such a vector fall back to |
7139 | element loads of the original vector type. */ | |
9da15d40 RS |
7140 | if (mode_for_vector (elmode, |
7141 | nunits / group_size).exists (&vmode) | |
7142 | && VECTOR_MODE_P (vmode) | |
ff03930a JJ |
7143 | && (convert_optab_handler (vec_init_optab, vmode, elmode) |
7144 | != CODE_FOR_nothing)) | |
7145 | { | |
7146 | nloads = nunits / group_size; | |
7147 | lnel = group_size; | |
7148 | ltype = build_nonstandard_integer_type (lsize, 1); | |
7149 | lvectype = build_vector_type (ltype, nloads); | |
7150 | } | |
ea60dd34 | 7151 | } |
e09b4c37 | 7152 | } |
2de001ee | 7153 | else |
e09b4c37 | 7154 | { |
ea60dd34 | 7155 | nloads = 1; |
e09b4c37 RB |
7156 | lnel = nunits; |
7157 | ltype = vectype; | |
e09b4c37 | 7158 | } |
2de001ee RS |
7159 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); |
7160 | } | |
7161 | if (slp) | |
7162 | { | |
66c16fd9 RB |
7163 | /* For SLP permutation support we need to load the whole group, |
7164 | not only the number of vector stmts the permutation result | |
7165 | fits in. */ | |
b266b968 | 7166 | if (slp_perm) |
66c16fd9 | 7167 | { |
d9f21f6a RS |
7168 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7169 | variable VF. */ | |
7170 | unsigned int const_vf = vf.to_constant (); | |
7171 | ncopies = (group_size * const_vf + nunits - 1) / nunits; | |
66c16fd9 RB |
7172 | dr_chain.create (ncopies); |
7173 | } | |
7174 | else | |
7175 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
7b5fc413 | 7176 | } |
e09b4c37 RB |
7177 | int group_el = 0; |
7178 | unsigned HOST_WIDE_INT | |
7179 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
7d75abc8 MM |
7180 | for (j = 0; j < ncopies; j++) |
7181 | { | |
7b5fc413 | 7182 | if (nloads > 1) |
e09b4c37 RB |
7183 | vec_alloc (v, nloads); |
7184 | for (i = 0; i < nloads; i++) | |
7b5fc413 | 7185 | { |
e09b4c37 RB |
7186 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
7187 | group_el * elsz); | |
7188 | new_stmt = gimple_build_assign (make_ssa_name (ltype), | |
7189 | build2 (MEM_REF, ltype, | |
7190 | running_off, this_off)); | |
7191 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7192 | if (nloads > 1) | |
7193 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, | |
7194 | gimple_assign_lhs (new_stmt)); | |
7195 | ||
7196 | group_el += lnel; | |
7197 | if (! slp | |
7198 | || group_el == group_size) | |
7b5fc413 | 7199 | { |
e09b4c37 RB |
7200 | tree newoff = copy_ssa_name (running_off); |
7201 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
7202 | running_off, stride_step); | |
7b5fc413 RB |
7203 | vect_finish_stmt_generation (stmt, incr, gsi); |
7204 | ||
7205 | running_off = newoff; | |
e09b4c37 | 7206 | group_el = 0; |
7b5fc413 | 7207 | } |
7b5fc413 | 7208 | } |
e09b4c37 | 7209 | if (nloads > 1) |
7d75abc8 | 7210 | { |
ea60dd34 RB |
7211 | tree vec_inv = build_constructor (lvectype, v); |
7212 | new_temp = vect_init_vector (stmt, vec_inv, lvectype, gsi); | |
e09b4c37 | 7213 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
ea60dd34 RB |
7214 | if (lvectype != vectype) |
7215 | { | |
7216 | new_stmt = gimple_build_assign (make_ssa_name (vectype), | |
7217 | VIEW_CONVERT_EXPR, | |
7218 | build1 (VIEW_CONVERT_EXPR, | |
7219 | vectype, new_temp)); | |
7220 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7221 | } | |
7d75abc8 MM |
7222 | } |
7223 | ||
7b5fc413 | 7224 | if (slp) |
b266b968 | 7225 | { |
b266b968 RB |
7226 | if (slp_perm) |
7227 | dr_chain.quick_push (gimple_assign_lhs (new_stmt)); | |
66c16fd9 RB |
7228 | else |
7229 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
b266b968 | 7230 | } |
7d75abc8 | 7231 | else |
225ce44b RB |
7232 | { |
7233 | if (j == 0) | |
7234 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7235 | else | |
7236 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7237 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7238 | } | |
7d75abc8 | 7239 | } |
b266b968 | 7240 | if (slp_perm) |
29afecdf RB |
7241 | { |
7242 | unsigned n_perms; | |
7243 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
7244 | slp_node_instance, false, &n_perms); | |
7245 | } | |
7d75abc8 MM |
7246 | return true; |
7247 | } | |
aec7ae7d | 7248 | |
0d0293ac | 7249 | if (grouped_load) |
ebfd146a | 7250 | { |
e14c1050 | 7251 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
44fc7854 | 7252 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
4f0a0218 | 7253 | /* For SLP vectorization we directly vectorize a subchain |
52eab378 RB |
7254 | without permutation. */ |
7255 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
4f0a0218 RB |
7256 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
7257 | /* For BB vectorization always use the first stmt to base | |
7258 | the data ref pointer on. */ | |
7259 | if (bb_vinfo) | |
7260 | first_stmt_for_drptr = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 7261 | |
ebfd146a | 7262 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
7263 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
7264 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
7265 | ??? But we can only do so if there is exactly one | |
7266 | as we have no way to get at the rest. Leave the CSE | |
7267 | opportunity alone. | |
7268 | ??? With the group load eventually participating | |
7269 | in multiple different permutations (having multiple | |
7270 | slp nodes which refer to the same group) the CSE | |
7271 | is even wrong code. See PR56270. */ | |
7272 | && !slp) | |
ebfd146a IR |
7273 | { |
7274 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
7275 | return true; | |
7276 | } | |
7277 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
9b999e8c | 7278 | group_gap_adj = 0; |
ebfd146a IR |
7279 | |
7280 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
7281 | if (slp) | |
7282 | { | |
0d0293ac | 7283 | grouped_load = false; |
91ff1504 RB |
7284 | /* For SLP permutation support we need to load the whole group, |
7285 | not only the number of vector stmts the permutation result | |
7286 | fits in. */ | |
7287 | if (slp_perm) | |
b267968e | 7288 | { |
d9f21f6a RS |
7289 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7290 | variable VF. */ | |
7291 | unsigned int const_vf = vf.to_constant (); | |
7292 | vec_num = (group_size * const_vf + nunits - 1) / nunits; | |
b267968e RB |
7293 | group_gap_adj = vf * group_size - nunits * vec_num; |
7294 | } | |
91ff1504 | 7295 | else |
b267968e RB |
7296 | { |
7297 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
796bd467 RB |
7298 | group_gap_adj |
7299 | = group_size - SLP_INSTANCE_GROUP_SIZE (slp_node_instance); | |
b267968e | 7300 | } |
a70d6342 | 7301 | } |
ebfd146a | 7302 | else |
9b999e8c | 7303 | vec_num = group_size; |
44fc7854 BE |
7304 | |
7305 | ref_type = get_group_alias_ptr_type (first_stmt); | |
ebfd146a IR |
7306 | } |
7307 | else | |
7308 | { | |
7309 | first_stmt = stmt; | |
7310 | first_dr = dr; | |
7311 | group_size = vec_num = 1; | |
9b999e8c | 7312 | group_gap_adj = 0; |
44fc7854 | 7313 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
ebfd146a IR |
7314 | } |
7315 | ||
720f5239 | 7316 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 7317 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
7318 | /* Targets with load-lane instructions must not require explicit |
7319 | realignment. */ | |
2de001ee | 7320 | gcc_assert (memory_access_type != VMAT_LOAD_STORE_LANES |
272c6793 RS |
7321 | || alignment_support_scheme == dr_aligned |
7322 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
7323 | |
7324 | /* In case the vectorization factor (VF) is bigger than the number | |
7325 | of elements that we can fit in a vectype (nunits), we have to generate | |
7326 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 7327 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 7328 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 7329 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 7330 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
7331 | stmts that use the defs of the current stmt. The example below |
7332 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
7333 | need to create 4 vectorized stmts): | |
ebfd146a IR |
7334 | |
7335 | before vectorization: | |
7336 | RELATED_STMT VEC_STMT | |
7337 | S1: x = memref - - | |
7338 | S2: z = x + 1 - - | |
7339 | ||
7340 | step 1: vectorize stmt S1: | |
7341 | We first create the vector stmt VS1_0, and, as usual, record a | |
7342 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
7343 | Next, we create the vector stmt VS1_1, and record a pointer to | |
7344 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 7345 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
7346 | stmts and pointers: |
7347 | RELATED_STMT VEC_STMT | |
7348 | VS1_0: vx0 = memref0 VS1_1 - | |
7349 | VS1_1: vx1 = memref1 VS1_2 - | |
7350 | VS1_2: vx2 = memref2 VS1_3 - | |
7351 | VS1_3: vx3 = memref3 - - | |
7352 | S1: x = load - VS1_0 | |
7353 | S2: z = x + 1 - - | |
7354 | ||
b8698a0f L |
7355 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
7356 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
7357 | stmt S2. */ |
7358 | ||
0d0293ac | 7359 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
7360 | |
7361 | S1: x2 = &base + 2 | |
7362 | S2: x0 = &base | |
7363 | S3: x1 = &base + 1 | |
7364 | S4: x3 = &base + 3 | |
7365 | ||
b8698a0f | 7366 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
7367 | starting from the access of the first stmt of the chain: |
7368 | ||
7369 | VS1: vx0 = &base | |
7370 | VS2: vx1 = &base + vec_size*1 | |
7371 | VS3: vx3 = &base + vec_size*2 | |
7372 | VS4: vx4 = &base + vec_size*3 | |
7373 | ||
7374 | Then permutation statements are generated: | |
7375 | ||
e2c83630 RH |
7376 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
7377 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
7378 | ... |
7379 | ||
7380 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
7381 | (the order of the data-refs in the output of vect_permute_load_chain | |
7382 | corresponds to the order of scalar stmts in the interleaving chain - see | |
7383 | the documentation of vect_permute_load_chain()). | |
7384 | The generation of permutation stmts and recording them in | |
0d0293ac | 7385 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 7386 | |
b8698a0f | 7387 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
7388 | permutation stmts above are created for every copy. The result vector |
7389 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
7390 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
7391 | |
7392 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
7393 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
7394 | we generate the following code: | |
7395 | p = initial_addr; | |
7396 | indx = 0; | |
7397 | loop { | |
7398 | p = p + indx * vectype_size; | |
7399 | vec_dest = *(p); | |
7400 | indx = indx + 1; | |
7401 | } | |
7402 | ||
7403 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 7404 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
7405 | then generate the following code, in which the data in each iteration is |
7406 | obtained by two vector loads, one from the previous iteration, and one | |
7407 | from the current iteration: | |
7408 | p1 = initial_addr; | |
7409 | msq_init = *(floor(p1)) | |
7410 | p2 = initial_addr + VS - 1; | |
7411 | realignment_token = call target_builtin; | |
7412 | indx = 0; | |
7413 | loop { | |
7414 | p2 = p2 + indx * vectype_size | |
7415 | lsq = *(floor(p2)) | |
7416 | vec_dest = realign_load (msq, lsq, realignment_token) | |
7417 | indx = indx + 1; | |
7418 | msq = lsq; | |
7419 | } */ | |
7420 | ||
7421 | /* If the misalignment remains the same throughout the execution of the | |
7422 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 7423 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
7424 | This can only occur when vectorizing memory accesses in the inner-loop |
7425 | nested within an outer-loop that is being vectorized. */ | |
7426 | ||
d1e4b493 | 7427 | if (nested_in_vect_loop |
832b4117 | 7428 | && (DR_STEP_ALIGNMENT (dr) % GET_MODE_SIZE (TYPE_MODE (vectype))) != 0) |
ebfd146a IR |
7429 | { |
7430 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
7431 | compute_in_loop = true; | |
7432 | } | |
7433 | ||
7434 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
7435 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 7436 | && !compute_in_loop) |
ebfd146a IR |
7437 | { |
7438 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
7439 | alignment_support_scheme, NULL_TREE, | |
7440 | &at_loop); | |
7441 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7442 | { | |
538dd0b7 | 7443 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
7444 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
7445 | size_one_node); | |
ebfd146a IR |
7446 | } |
7447 | } | |
7448 | else | |
7449 | at_loop = loop; | |
7450 | ||
62da9e14 | 7451 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
a1e53f3f L |
7452 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
7453 | ||
2de001ee | 7454 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
272c6793 RS |
7455 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); |
7456 | else | |
7457 | aggr_type = vectype; | |
7458 | ||
ebfd146a | 7459 | prev_stmt_info = NULL; |
b267968e | 7460 | int group_elt = 0; |
ebfd146a | 7461 | for (j = 0; j < ncopies; j++) |
b8698a0f | 7462 | { |
272c6793 | 7463 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 7464 | if (j == 0) |
74bf76ed JJ |
7465 | { |
7466 | bool simd_lane_access_p | |
7467 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
7468 | if (simd_lane_access_p | |
7469 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
7470 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
7471 | && integer_zerop (DR_OFFSET (first_dr)) | |
7472 | && integer_zerop (DR_INIT (first_dr)) | |
7473 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 7474 | get_alias_set (TREE_TYPE (ref_type))) |
74bf76ed JJ |
7475 | && (alignment_support_scheme == dr_aligned |
7476 | || alignment_support_scheme == dr_unaligned_supported)) | |
7477 | { | |
7478 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 7479 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 7480 | inv_p = false; |
74bf76ed | 7481 | } |
4f0a0218 RB |
7482 | else if (first_stmt_for_drptr |
7483 | && first_stmt != first_stmt_for_drptr) | |
7484 | { | |
7485 | dataref_ptr | |
7486 | = vect_create_data_ref_ptr (first_stmt_for_drptr, aggr_type, | |
7487 | at_loop, offset, &dummy, gsi, | |
7488 | &ptr_incr, simd_lane_access_p, | |
7489 | &inv_p, byte_offset); | |
7490 | /* Adjust the pointer by the difference to first_stmt. */ | |
7491 | data_reference_p ptrdr | |
7492 | = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt_for_drptr)); | |
7493 | tree diff = fold_convert (sizetype, | |
7494 | size_binop (MINUS_EXPR, | |
7495 | DR_INIT (first_dr), | |
7496 | DR_INIT (ptrdr))); | |
7497 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7498 | stmt, diff); | |
7499 | } | |
74bf76ed JJ |
7500 | else |
7501 | dataref_ptr | |
7502 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
7503 | offset, &dummy, gsi, &ptr_incr, | |
356bbc4c JJ |
7504 | simd_lane_access_p, &inv_p, |
7505 | byte_offset); | |
74bf76ed JJ |
7506 | } |
7507 | else if (dataref_offset) | |
7508 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
7509 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 7510 | else |
272c6793 RS |
7511 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
7512 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 7513 | |
0d0293ac | 7514 | if (grouped_load || slp_perm) |
9771b263 | 7515 | dr_chain.create (vec_num); |
5ce1ee7f | 7516 | |
2de001ee | 7517 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 7518 | { |
272c6793 RS |
7519 | tree vec_array; |
7520 | ||
7521 | vec_array = create_vector_array (vectype, vec_num); | |
7522 | ||
7523 | /* Emit: | |
7524 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
44fc7854 | 7525 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); |
a844293d RS |
7526 | gcall *call = gimple_build_call_internal (IFN_LOAD_LANES, 1, |
7527 | data_ref); | |
7528 | gimple_call_set_lhs (call, vec_array); | |
7529 | gimple_call_set_nothrow (call, true); | |
7530 | new_stmt = call; | |
272c6793 | 7531 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a | 7532 | |
272c6793 RS |
7533 | /* Extract each vector into an SSA_NAME. */ |
7534 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 7535 | { |
272c6793 RS |
7536 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
7537 | vec_array, i); | |
9771b263 | 7538 | dr_chain.quick_push (new_temp); |
272c6793 RS |
7539 | } |
7540 | ||
7541 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 7542 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
7543 | } |
7544 | else | |
7545 | { | |
7546 | for (i = 0; i < vec_num; i++) | |
7547 | { | |
7548 | if (i > 0) | |
7549 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7550 | stmt, NULL_TREE); | |
7551 | ||
7552 | /* 2. Create the vector-load in the loop. */ | |
7553 | switch (alignment_support_scheme) | |
7554 | { | |
7555 | case dr_aligned: | |
7556 | case dr_unaligned_supported: | |
be1ac4ec | 7557 | { |
644ffefd MJ |
7558 | unsigned int align, misalign; |
7559 | ||
272c6793 | 7560 | data_ref |
aed93b23 RB |
7561 | = fold_build2 (MEM_REF, vectype, dataref_ptr, |
7562 | dataref_offset | |
7563 | ? dataref_offset | |
44fc7854 | 7564 | : build_int_cst (ref_type, 0)); |
f702e7d4 | 7565 | align = DR_TARGET_ALIGNMENT (dr); |
272c6793 RS |
7566 | if (alignment_support_scheme == dr_aligned) |
7567 | { | |
7568 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 7569 | misalign = 0; |
272c6793 RS |
7570 | } |
7571 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
7572 | { | |
25f68d90 | 7573 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 7574 | misalign = 0; |
272c6793 RS |
7575 | TREE_TYPE (data_ref) |
7576 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 7577 | align * BITS_PER_UNIT); |
272c6793 RS |
7578 | } |
7579 | else | |
7580 | { | |
7581 | TREE_TYPE (data_ref) | |
7582 | = build_aligned_type (TREE_TYPE (data_ref), | |
7583 | TYPE_ALIGN (elem_type)); | |
644ffefd | 7584 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 7585 | } |
aed93b23 RB |
7586 | if (dataref_offset == NULL_TREE |
7587 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
7588 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
7589 | align, misalign); | |
272c6793 | 7590 | break; |
be1ac4ec | 7591 | } |
272c6793 | 7592 | case dr_explicit_realign: |
267d3070 | 7593 | { |
272c6793 | 7594 | tree ptr, bump; |
272c6793 | 7595 | |
d88981fc | 7596 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
7597 | |
7598 | if (compute_in_loop) | |
7599 | msq = vect_setup_realignment (first_stmt, gsi, | |
7600 | &realignment_token, | |
7601 | dr_explicit_realign, | |
7602 | dataref_ptr, NULL); | |
7603 | ||
aed93b23 RB |
7604 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7605 | ptr = copy_ssa_name (dataref_ptr); | |
7606 | else | |
7607 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
f702e7d4 | 7608 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); |
0d0e4a03 JJ |
7609 | new_stmt = gimple_build_assign |
7610 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7611 | build_int_cst |
7612 | (TREE_TYPE (dataref_ptr), | |
f702e7d4 | 7613 | -(HOST_WIDE_INT) align)); |
272c6793 RS |
7614 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7615 | data_ref | |
7616 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 7617 | build_int_cst (ref_type, 0)); |
272c6793 RS |
7618 | vec_dest = vect_create_destination_var (scalar_dest, |
7619 | vectype); | |
7620 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
7621 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
7622 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7623 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
7624 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
7625 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7626 | msq = new_temp; | |
7627 | ||
d88981fc | 7628 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 7629 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 7630 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 7631 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
7632 | new_stmt = gimple_build_assign |
7633 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 | 7634 | build_int_cst |
f702e7d4 | 7635 | (TREE_TYPE (ptr), -(HOST_WIDE_INT) align)); |
aed93b23 | 7636 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
7637 | gimple_assign_set_lhs (new_stmt, ptr); |
7638 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7639 | data_ref | |
7640 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 7641 | build_int_cst (ref_type, 0)); |
272c6793 | 7642 | break; |
267d3070 | 7643 | } |
272c6793 | 7644 | case dr_explicit_realign_optimized: |
f702e7d4 RS |
7645 | { |
7646 | if (TREE_CODE (dataref_ptr) == SSA_NAME) | |
7647 | new_temp = copy_ssa_name (dataref_ptr); | |
7648 | else | |
7649 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
7650 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); | |
7651 | new_stmt = gimple_build_assign | |
7652 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
7653 | build_int_cst (TREE_TYPE (dataref_ptr), | |
7654 | -(HOST_WIDE_INT) align)); | |
7655 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7656 | data_ref | |
7657 | = build2 (MEM_REF, vectype, new_temp, | |
7658 | build_int_cst (ref_type, 0)); | |
7659 | break; | |
7660 | } | |
272c6793 RS |
7661 | default: |
7662 | gcc_unreachable (); | |
7663 | } | |
ebfd146a | 7664 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 7665 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
7666 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7667 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7668 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7669 | ||
272c6793 RS |
7670 | /* 3. Handle explicit realignment if necessary/supported. |
7671 | Create in loop: | |
7672 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
7673 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
7674 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 7675 | { |
272c6793 RS |
7676 | lsq = gimple_assign_lhs (new_stmt); |
7677 | if (!realignment_token) | |
7678 | realignment_token = dataref_ptr; | |
7679 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
7680 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
7681 | msq, lsq, realignment_token); | |
272c6793 RS |
7682 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7683 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7684 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7685 | ||
7686 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7687 | { | |
7688 | gcc_assert (phi); | |
7689 | if (i == vec_num - 1 && j == ncopies - 1) | |
7690 | add_phi_arg (phi, lsq, | |
7691 | loop_latch_edge (containing_loop), | |
9e227d60 | 7692 | UNKNOWN_LOCATION); |
272c6793 RS |
7693 | msq = lsq; |
7694 | } | |
ebfd146a | 7695 | } |
ebfd146a | 7696 | |
59fd17e3 RB |
7697 | /* 4. Handle invariant-load. */ |
7698 | if (inv_p && !bb_vinfo) | |
7699 | { | |
59fd17e3 | 7700 | gcc_assert (!grouped_load); |
d1417442 JJ |
7701 | /* If we have versioned for aliasing or the loop doesn't |
7702 | have any data dependencies that would preclude this, | |
7703 | then we are sure this is a loop invariant load and | |
7704 | thus we can insert it on the preheader edge. */ | |
7705 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
7706 | && !nested_in_vect_loop | |
6b916b36 | 7707 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
7708 | { |
7709 | if (dump_enabled_p ()) | |
7710 | { | |
7711 | dump_printf_loc (MSG_NOTE, vect_location, | |
7712 | "hoisting out of the vectorized " | |
7713 | "loop: "); | |
7714 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 7715 | } |
b731b390 | 7716 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
7717 | gsi_insert_on_edge_immediate |
7718 | (loop_preheader_edge (loop), | |
7719 | gimple_build_assign (tem, | |
7720 | unshare_expr | |
7721 | (gimple_assign_rhs1 (stmt)))); | |
7722 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
34cd48e5 RB |
7723 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7724 | set_vinfo_for_stmt (new_stmt, | |
7725 | new_stmt_vec_info (new_stmt, vinfo)); | |
a0e35eb0 RB |
7726 | } |
7727 | else | |
7728 | { | |
7729 | gimple_stmt_iterator gsi2 = *gsi; | |
7730 | gsi_next (&gsi2); | |
7731 | new_temp = vect_init_vector (stmt, scalar_dest, | |
7732 | vectype, &gsi2); | |
34cd48e5 | 7733 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
a0e35eb0 | 7734 | } |
59fd17e3 RB |
7735 | } |
7736 | ||
62da9e14 | 7737 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
272c6793 | 7738 | { |
aec7ae7d JJ |
7739 | tree perm_mask = perm_mask_for_reverse (vectype); |
7740 | new_temp = permute_vec_elements (new_temp, new_temp, | |
7741 | perm_mask, stmt, gsi); | |
ebfd146a IR |
7742 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7743 | } | |
267d3070 | 7744 | |
272c6793 | 7745 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
7746 | vect_transform_grouped_load (). */ |
7747 | if (grouped_load || slp_perm) | |
9771b263 | 7748 | dr_chain.quick_push (new_temp); |
267d3070 | 7749 | |
272c6793 RS |
7750 | /* Store vector loads in the corresponding SLP_NODE. */ |
7751 | if (slp && !slp_perm) | |
9771b263 | 7752 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
b267968e RB |
7753 | |
7754 | /* With SLP permutation we load the gaps as well, without | |
7755 | we need to skip the gaps after we manage to fully load | |
7756 | all elements. group_gap_adj is GROUP_SIZE here. */ | |
7757 | group_elt += nunits; | |
d9f21f6a RS |
7758 | if (maybe_ne (group_gap_adj, 0U) |
7759 | && !slp_perm | |
7760 | && known_eq (group_elt, group_size - group_gap_adj)) | |
b267968e | 7761 | { |
d9f21f6a RS |
7762 | poly_wide_int bump_val |
7763 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
7764 | * group_gap_adj); | |
8e6cdc90 | 7765 | tree bump = wide_int_to_tree (sizetype, bump_val); |
b267968e RB |
7766 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
7767 | stmt, bump); | |
7768 | group_elt = 0; | |
7769 | } | |
272c6793 | 7770 | } |
9b999e8c RB |
7771 | /* Bump the vector pointer to account for a gap or for excess |
7772 | elements loaded for a permuted SLP load. */ | |
d9f21f6a | 7773 | if (maybe_ne (group_gap_adj, 0U) && slp_perm) |
a64b9c26 | 7774 | { |
d9f21f6a RS |
7775 | poly_wide_int bump_val |
7776 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
7777 | * group_gap_adj); | |
8e6cdc90 | 7778 | tree bump = wide_int_to_tree (sizetype, bump_val); |
a64b9c26 RB |
7779 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
7780 | stmt, bump); | |
7781 | } | |
ebfd146a IR |
7782 | } |
7783 | ||
7784 | if (slp && !slp_perm) | |
7785 | continue; | |
7786 | ||
7787 | if (slp_perm) | |
7788 | { | |
29afecdf | 7789 | unsigned n_perms; |
01d8bf07 | 7790 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
29afecdf RB |
7791 | slp_node_instance, false, |
7792 | &n_perms)) | |
ebfd146a | 7793 | { |
9771b263 | 7794 | dr_chain.release (); |
ebfd146a IR |
7795 | return false; |
7796 | } | |
7797 | } | |
7798 | else | |
7799 | { | |
0d0293ac | 7800 | if (grouped_load) |
ebfd146a | 7801 | { |
2de001ee | 7802 | if (memory_access_type != VMAT_LOAD_STORE_LANES) |
0d0293ac | 7803 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 7804 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
7805 | } |
7806 | else | |
7807 | { | |
7808 | if (j == 0) | |
7809 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7810 | else | |
7811 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7812 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7813 | } | |
7814 | } | |
9771b263 | 7815 | dr_chain.release (); |
ebfd146a IR |
7816 | } |
7817 | ||
ebfd146a IR |
7818 | return true; |
7819 | } | |
7820 | ||
7821 | /* Function vect_is_simple_cond. | |
b8698a0f | 7822 | |
ebfd146a IR |
7823 | Input: |
7824 | LOOP - the loop that is being vectorized. | |
7825 | COND - Condition that is checked for simple use. | |
7826 | ||
e9e1d143 RG |
7827 | Output: |
7828 | *COMP_VECTYPE - the vector type for the comparison. | |
4fc5ebf1 | 7829 | *DTS - The def types for the arguments of the comparison |
e9e1d143 | 7830 | |
ebfd146a IR |
7831 | Returns whether a COND can be vectorized. Checks whether |
7832 | condition operands are supportable using vec_is_simple_use. */ | |
7833 | ||
87aab9b2 | 7834 | static bool |
4fc5ebf1 | 7835 | vect_is_simple_cond (tree cond, vec_info *vinfo, |
8da4c8d8 RB |
7836 | tree *comp_vectype, enum vect_def_type *dts, |
7837 | tree vectype) | |
ebfd146a IR |
7838 | { |
7839 | tree lhs, rhs; | |
e9e1d143 | 7840 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 7841 | |
a414c77f IE |
7842 | /* Mask case. */ |
7843 | if (TREE_CODE (cond) == SSA_NAME | |
2568d8a1 | 7844 | && VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (cond))) |
a414c77f IE |
7845 | { |
7846 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (cond); | |
7847 | if (!vect_is_simple_use (cond, vinfo, &lhs_def_stmt, | |
4fc5ebf1 | 7848 | &dts[0], comp_vectype) |
a414c77f IE |
7849 | || !*comp_vectype |
7850 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
7851 | return false; | |
7852 | return true; | |
7853 | } | |
7854 | ||
ebfd146a IR |
7855 | if (!COMPARISON_CLASS_P (cond)) |
7856 | return false; | |
7857 | ||
7858 | lhs = TREE_OPERAND (cond, 0); | |
7859 | rhs = TREE_OPERAND (cond, 1); | |
7860 | ||
7861 | if (TREE_CODE (lhs) == SSA_NAME) | |
7862 | { | |
355fe088 | 7863 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); |
4fc5ebf1 | 7864 | if (!vect_is_simple_use (lhs, vinfo, &lhs_def_stmt, &dts[0], &vectype1)) |
ebfd146a IR |
7865 | return false; |
7866 | } | |
4fc5ebf1 JG |
7867 | else if (TREE_CODE (lhs) == INTEGER_CST || TREE_CODE (lhs) == REAL_CST |
7868 | || TREE_CODE (lhs) == FIXED_CST) | |
7869 | dts[0] = vect_constant_def; | |
7870 | else | |
ebfd146a IR |
7871 | return false; |
7872 | ||
7873 | if (TREE_CODE (rhs) == SSA_NAME) | |
7874 | { | |
355fe088 | 7875 | gimple *rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); |
4fc5ebf1 | 7876 | if (!vect_is_simple_use (rhs, vinfo, &rhs_def_stmt, &dts[1], &vectype2)) |
ebfd146a IR |
7877 | return false; |
7878 | } | |
4fc5ebf1 JG |
7879 | else if (TREE_CODE (rhs) == INTEGER_CST || TREE_CODE (rhs) == REAL_CST |
7880 | || TREE_CODE (rhs) == FIXED_CST) | |
7881 | dts[1] = vect_constant_def; | |
7882 | else | |
ebfd146a IR |
7883 | return false; |
7884 | ||
28b33016 IE |
7885 | if (vectype1 && vectype2 |
7886 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
7887 | return false; | |
7888 | ||
e9e1d143 | 7889 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
8da4c8d8 RB |
7890 | /* Invariant comparison. */ |
7891 | if (! *comp_vectype) | |
7892 | { | |
7893 | tree scalar_type = TREE_TYPE (lhs); | |
7894 | /* If we can widen the comparison to match vectype do so. */ | |
7895 | if (INTEGRAL_TYPE_P (scalar_type) | |
7896 | && tree_int_cst_lt (TYPE_SIZE (scalar_type), | |
7897 | TYPE_SIZE (TREE_TYPE (vectype)))) | |
7898 | scalar_type = build_nonstandard_integer_type | |
7899 | (tree_to_uhwi (TYPE_SIZE (TREE_TYPE (vectype))), | |
7900 | TYPE_UNSIGNED (scalar_type)); | |
7901 | *comp_vectype = get_vectype_for_scalar_type (scalar_type); | |
7902 | } | |
7903 | ||
ebfd146a IR |
7904 | return true; |
7905 | } | |
7906 | ||
7907 | /* vectorizable_condition. | |
7908 | ||
b8698a0f L |
7909 | Check if STMT is conditional modify expression that can be vectorized. |
7910 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7911 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
7912 | at GSI. |
7913 | ||
7914 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
7915 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 7916 | else clause if it is 2). |
ebfd146a IR |
7917 | |
7918 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7919 | ||
4bbe8262 | 7920 | bool |
355fe088 TS |
7921 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
7922 | gimple **vec_stmt, tree reduc_def, int reduc_index, | |
f7e531cf | 7923 | slp_tree slp_node) |
ebfd146a IR |
7924 | { |
7925 | tree scalar_dest = NULL_TREE; | |
7926 | tree vec_dest = NULL_TREE; | |
01216d27 JJ |
7927 | tree cond_expr, cond_expr0 = NULL_TREE, cond_expr1 = NULL_TREE; |
7928 | tree then_clause, else_clause; | |
ebfd146a | 7929 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
df11cc78 | 7930 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
7931 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
7932 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
5958f9e2 | 7933 | tree vec_compare; |
ebfd146a IR |
7934 | tree new_temp; |
7935 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4fc5ebf1 JG |
7936 | enum vect_def_type dts[4] |
7937 | = {vect_unknown_def_type, vect_unknown_def_type, | |
7938 | vect_unknown_def_type, vect_unknown_def_type}; | |
7939 | int ndts = 4; | |
f7e531cf | 7940 | int ncopies; |
01216d27 | 7941 | enum tree_code code, cond_code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
a855b1b1 | 7942 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
7943 | int i, j; |
7944 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
7945 | vec<tree> vec_oprnds0 = vNULL; |
7946 | vec<tree> vec_oprnds1 = vNULL; | |
7947 | vec<tree> vec_oprnds2 = vNULL; | |
7948 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 7949 | tree vec_cmp_type; |
a414c77f | 7950 | bool masked = false; |
b8698a0f | 7951 | |
f7e531cf IR |
7952 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
7953 | return false; | |
7954 | ||
af29617a AH |
7955 | if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == TREE_CODE_REDUCTION) |
7956 | { | |
7957 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
7958 | return false; | |
ebfd146a | 7959 | |
af29617a AH |
7960 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7961 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7962 | && reduc_def)) | |
7963 | return false; | |
ebfd146a | 7964 | |
af29617a AH |
7965 | /* FORNOW: not yet supported. */ |
7966 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7967 | { | |
7968 | if (dump_enabled_p ()) | |
7969 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7970 | "value used after loop.\n"); | |
7971 | return false; | |
7972 | } | |
ebfd146a IR |
7973 | } |
7974 | ||
7975 | /* Is vectorizable conditional operation? */ | |
7976 | if (!is_gimple_assign (stmt)) | |
7977 | return false; | |
7978 | ||
7979 | code = gimple_assign_rhs_code (stmt); | |
7980 | ||
7981 | if (code != COND_EXPR) | |
7982 | return false; | |
7983 | ||
465c8c19 | 7984 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2947d3b2 | 7985 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
465c8c19 | 7986 | |
fce57248 | 7987 | if (slp_node) |
465c8c19 JJ |
7988 | ncopies = 1; |
7989 | else | |
e8f142e2 | 7990 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
7991 | |
7992 | gcc_assert (ncopies >= 1); | |
7993 | if (reduc_index && ncopies > 1) | |
7994 | return false; /* FORNOW */ | |
7995 | ||
4e71066d RG |
7996 | cond_expr = gimple_assign_rhs1 (stmt); |
7997 | then_clause = gimple_assign_rhs2 (stmt); | |
7998 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 7999 | |
4fc5ebf1 | 8000 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, |
8da4c8d8 | 8001 | &comp_vectype, &dts[0], vectype) |
e9e1d143 | 8002 | || !comp_vectype) |
ebfd146a IR |
8003 | return false; |
8004 | ||
81c40241 | 8005 | gimple *def_stmt; |
4fc5ebf1 | 8006 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &def_stmt, &dts[2], |
2947d3b2 IE |
8007 | &vectype1)) |
8008 | return false; | |
4fc5ebf1 | 8009 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &def_stmt, &dts[3], |
2947d3b2 | 8010 | &vectype2)) |
ebfd146a | 8011 | return false; |
2947d3b2 IE |
8012 | |
8013 | if (vectype1 && !useless_type_conversion_p (vectype, vectype1)) | |
8014 | return false; | |
8015 | ||
8016 | if (vectype2 && !useless_type_conversion_p (vectype, vectype2)) | |
ebfd146a IR |
8017 | return false; |
8018 | ||
28b33016 IE |
8019 | masked = !COMPARISON_CLASS_P (cond_expr); |
8020 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
8021 | ||
74946978 MP |
8022 | if (vec_cmp_type == NULL_TREE) |
8023 | return false; | |
784fb9b3 | 8024 | |
01216d27 JJ |
8025 | cond_code = TREE_CODE (cond_expr); |
8026 | if (!masked) | |
8027 | { | |
8028 | cond_expr0 = TREE_OPERAND (cond_expr, 0); | |
8029 | cond_expr1 = TREE_OPERAND (cond_expr, 1); | |
8030 | } | |
8031 | ||
8032 | if (!masked && VECTOR_BOOLEAN_TYPE_P (comp_vectype)) | |
8033 | { | |
8034 | /* Boolean values may have another representation in vectors | |
8035 | and therefore we prefer bit operations over comparison for | |
8036 | them (which also works for scalar masks). We store opcodes | |
8037 | to use in bitop1 and bitop2. Statement is vectorized as | |
8038 | BITOP2 (rhs1 BITOP1 rhs2) or rhs1 BITOP2 (BITOP1 rhs2) | |
8039 | depending on bitop1 and bitop2 arity. */ | |
8040 | switch (cond_code) | |
8041 | { | |
8042 | case GT_EXPR: | |
8043 | bitop1 = BIT_NOT_EXPR; | |
8044 | bitop2 = BIT_AND_EXPR; | |
8045 | break; | |
8046 | case GE_EXPR: | |
8047 | bitop1 = BIT_NOT_EXPR; | |
8048 | bitop2 = BIT_IOR_EXPR; | |
8049 | break; | |
8050 | case LT_EXPR: | |
8051 | bitop1 = BIT_NOT_EXPR; | |
8052 | bitop2 = BIT_AND_EXPR; | |
8053 | std::swap (cond_expr0, cond_expr1); | |
8054 | break; | |
8055 | case LE_EXPR: | |
8056 | bitop1 = BIT_NOT_EXPR; | |
8057 | bitop2 = BIT_IOR_EXPR; | |
8058 | std::swap (cond_expr0, cond_expr1); | |
8059 | break; | |
8060 | case NE_EXPR: | |
8061 | bitop1 = BIT_XOR_EXPR; | |
8062 | break; | |
8063 | case EQ_EXPR: | |
8064 | bitop1 = BIT_XOR_EXPR; | |
8065 | bitop2 = BIT_NOT_EXPR; | |
8066 | break; | |
8067 | default: | |
8068 | return false; | |
8069 | } | |
8070 | cond_code = SSA_NAME; | |
8071 | } | |
8072 | ||
b8698a0f | 8073 | if (!vec_stmt) |
ebfd146a IR |
8074 | { |
8075 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
01216d27 JJ |
8076 | if (bitop1 != NOP_EXPR) |
8077 | { | |
8078 | machine_mode mode = TYPE_MODE (comp_vectype); | |
8079 | optab optab; | |
8080 | ||
8081 | optab = optab_for_tree_code (bitop1, comp_vectype, optab_default); | |
8082 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8083 | return false; | |
8084 | ||
8085 | if (bitop2 != NOP_EXPR) | |
8086 | { | |
8087 | optab = optab_for_tree_code (bitop2, comp_vectype, | |
8088 | optab_default); | |
8089 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8090 | return false; | |
8091 | } | |
8092 | } | |
4fc5ebf1 JG |
8093 | if (expand_vec_cond_expr_p (vectype, comp_vectype, |
8094 | cond_code)) | |
8095 | { | |
8096 | vect_model_simple_cost (stmt_info, ncopies, dts, ndts, NULL, NULL); | |
8097 | return true; | |
8098 | } | |
8099 | return false; | |
ebfd146a IR |
8100 | } |
8101 | ||
f7e531cf IR |
8102 | /* Transform. */ |
8103 | ||
8104 | if (!slp_node) | |
8105 | { | |
9771b263 DN |
8106 | vec_oprnds0.create (1); |
8107 | vec_oprnds1.create (1); | |
8108 | vec_oprnds2.create (1); | |
8109 | vec_oprnds3.create (1); | |
f7e531cf | 8110 | } |
ebfd146a IR |
8111 | |
8112 | /* Handle def. */ | |
8113 | scalar_dest = gimple_assign_lhs (stmt); | |
8114 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
8115 | ||
8116 | /* Handle cond expr. */ | |
a855b1b1 MM |
8117 | for (j = 0; j < ncopies; j++) |
8118 | { | |
538dd0b7 | 8119 | gassign *new_stmt = NULL; |
a855b1b1 MM |
8120 | if (j == 0) |
8121 | { | |
f7e531cf IR |
8122 | if (slp_node) |
8123 | { | |
00f96dc9 TS |
8124 | auto_vec<tree, 4> ops; |
8125 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 8126 | |
a414c77f | 8127 | if (masked) |
01216d27 | 8128 | ops.safe_push (cond_expr); |
a414c77f IE |
8129 | else |
8130 | { | |
01216d27 JJ |
8131 | ops.safe_push (cond_expr0); |
8132 | ops.safe_push (cond_expr1); | |
a414c77f | 8133 | } |
9771b263 DN |
8134 | ops.safe_push (then_clause); |
8135 | ops.safe_push (else_clause); | |
306b0c92 | 8136 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
37b5ec8f JJ |
8137 | vec_oprnds3 = vec_defs.pop (); |
8138 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
8139 | if (!masked) |
8140 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 8141 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf IR |
8142 | } |
8143 | else | |
8144 | { | |
355fe088 | 8145 | gimple *gtemp; |
a414c77f IE |
8146 | if (masked) |
8147 | { | |
8148 | vec_cond_lhs | |
8149 | = vect_get_vec_def_for_operand (cond_expr, stmt, | |
8150 | comp_vectype); | |
8151 | vect_is_simple_use (cond_expr, stmt_info->vinfo, | |
8152 | >emp, &dts[0]); | |
8153 | } | |
8154 | else | |
8155 | { | |
01216d27 JJ |
8156 | vec_cond_lhs |
8157 | = vect_get_vec_def_for_operand (cond_expr0, | |
8158 | stmt, comp_vectype); | |
8159 | vect_is_simple_use (cond_expr0, loop_vinfo, >emp, &dts[0]); | |
8160 | ||
8161 | vec_cond_rhs | |
8162 | = vect_get_vec_def_for_operand (cond_expr1, | |
8163 | stmt, comp_vectype); | |
8164 | vect_is_simple_use (cond_expr1, loop_vinfo, >emp, &dts[1]); | |
a414c77f | 8165 | } |
f7e531cf IR |
8166 | if (reduc_index == 1) |
8167 | vec_then_clause = reduc_def; | |
8168 | else | |
8169 | { | |
8170 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 RB |
8171 | stmt); |
8172 | vect_is_simple_use (then_clause, loop_vinfo, | |
8173 | >emp, &dts[2]); | |
f7e531cf IR |
8174 | } |
8175 | if (reduc_index == 2) | |
8176 | vec_else_clause = reduc_def; | |
8177 | else | |
8178 | { | |
8179 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 RB |
8180 | stmt); |
8181 | vect_is_simple_use (else_clause, loop_vinfo, >emp, &dts[3]); | |
f7e531cf | 8182 | } |
a855b1b1 MM |
8183 | } |
8184 | } | |
8185 | else | |
8186 | { | |
a414c77f IE |
8187 | vec_cond_lhs |
8188 | = vect_get_vec_def_for_stmt_copy (dts[0], | |
8189 | vec_oprnds0.pop ()); | |
8190 | if (!masked) | |
8191 | vec_cond_rhs | |
8192 | = vect_get_vec_def_for_stmt_copy (dts[1], | |
8193 | vec_oprnds1.pop ()); | |
8194 | ||
a855b1b1 | 8195 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 8196 | vec_oprnds2.pop ()); |
a855b1b1 | 8197 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 8198 | vec_oprnds3.pop ()); |
f7e531cf IR |
8199 | } |
8200 | ||
8201 | if (!slp_node) | |
8202 | { | |
9771b263 | 8203 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
8204 | if (!masked) |
8205 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
8206 | vec_oprnds2.quick_push (vec_then_clause); |
8207 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
8208 | } |
8209 | ||
9dc3f7de | 8210 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 8211 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 8212 | { |
9771b263 DN |
8213 | vec_then_clause = vec_oprnds2[i]; |
8214 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 8215 | |
a414c77f IE |
8216 | if (masked) |
8217 | vec_compare = vec_cond_lhs; | |
8218 | else | |
8219 | { | |
8220 | vec_cond_rhs = vec_oprnds1[i]; | |
01216d27 JJ |
8221 | if (bitop1 == NOP_EXPR) |
8222 | vec_compare = build2 (cond_code, vec_cmp_type, | |
8223 | vec_cond_lhs, vec_cond_rhs); | |
8224 | else | |
8225 | { | |
8226 | new_temp = make_ssa_name (vec_cmp_type); | |
8227 | if (bitop1 == BIT_NOT_EXPR) | |
8228 | new_stmt = gimple_build_assign (new_temp, bitop1, | |
8229 | vec_cond_rhs); | |
8230 | else | |
8231 | new_stmt | |
8232 | = gimple_build_assign (new_temp, bitop1, vec_cond_lhs, | |
8233 | vec_cond_rhs); | |
8234 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8235 | if (bitop2 == NOP_EXPR) | |
8236 | vec_compare = new_temp; | |
8237 | else if (bitop2 == BIT_NOT_EXPR) | |
8238 | { | |
8239 | /* Instead of doing ~x ? y : z do x ? z : y. */ | |
8240 | vec_compare = new_temp; | |
8241 | std::swap (vec_then_clause, vec_else_clause); | |
8242 | } | |
8243 | else | |
8244 | { | |
8245 | vec_compare = make_ssa_name (vec_cmp_type); | |
8246 | new_stmt | |
8247 | = gimple_build_assign (vec_compare, bitop2, | |
8248 | vec_cond_lhs, new_temp); | |
8249 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8250 | } | |
8251 | } | |
a414c77f | 8252 | } |
5958f9e2 JJ |
8253 | new_temp = make_ssa_name (vec_dest); |
8254 | new_stmt = gimple_build_assign (new_temp, VEC_COND_EXPR, | |
8255 | vec_compare, vec_then_clause, | |
8256 | vec_else_clause); | |
f7e531cf IR |
8257 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
8258 | if (slp_node) | |
9771b263 | 8259 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
8260 | } |
8261 | ||
8262 | if (slp_node) | |
8263 | continue; | |
8264 | ||
8265 | if (j == 0) | |
8266 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
8267 | else | |
8268 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
8269 | ||
8270 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 8271 | } |
b8698a0f | 8272 | |
9771b263 DN |
8273 | vec_oprnds0.release (); |
8274 | vec_oprnds1.release (); | |
8275 | vec_oprnds2.release (); | |
8276 | vec_oprnds3.release (); | |
f7e531cf | 8277 | |
ebfd146a IR |
8278 | return true; |
8279 | } | |
8280 | ||
42fd8198 IE |
8281 | /* vectorizable_comparison. |
8282 | ||
8283 | Check if STMT is comparison expression that can be vectorized. | |
8284 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
8285 | comparison, put it in VEC_STMT, and insert it at GSI. | |
8286 | ||
8287 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
8288 | ||
fce57248 | 8289 | static bool |
42fd8198 IE |
8290 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, |
8291 | gimple **vec_stmt, tree reduc_def, | |
8292 | slp_tree slp_node) | |
8293 | { | |
8294 | tree lhs, rhs1, rhs2; | |
8295 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8296 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
8297 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
8298 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
8299 | tree new_temp; | |
8300 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
8301 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 8302 | int ndts = 2; |
42fd8198 IE |
8303 | unsigned nunits; |
8304 | int ncopies; | |
49e76ff1 | 8305 | enum tree_code code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
42fd8198 IE |
8306 | stmt_vec_info prev_stmt_info = NULL; |
8307 | int i, j; | |
8308 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
8309 | vec<tree> vec_oprnds0 = vNULL; | |
8310 | vec<tree> vec_oprnds1 = vNULL; | |
8311 | gimple *def_stmt; | |
8312 | tree mask_type; | |
8313 | tree mask; | |
8314 | ||
c245362b IE |
8315 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
8316 | return false; | |
8317 | ||
30480bcd | 8318 | if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)) |
42fd8198 IE |
8319 | return false; |
8320 | ||
8321 | mask_type = vectype; | |
8322 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
8323 | ||
fce57248 | 8324 | if (slp_node) |
42fd8198 IE |
8325 | ncopies = 1; |
8326 | else | |
e8f142e2 | 8327 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
42fd8198 IE |
8328 | |
8329 | gcc_assert (ncopies >= 1); | |
42fd8198 IE |
8330 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
8331 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
8332 | && reduc_def)) | |
8333 | return false; | |
8334 | ||
8335 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
8336 | { | |
8337 | if (dump_enabled_p ()) | |
8338 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
8339 | "value used after loop.\n"); | |
8340 | return false; | |
8341 | } | |
8342 | ||
8343 | if (!is_gimple_assign (stmt)) | |
8344 | return false; | |
8345 | ||
8346 | code = gimple_assign_rhs_code (stmt); | |
8347 | ||
8348 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
8349 | return false; | |
8350 | ||
8351 | rhs1 = gimple_assign_rhs1 (stmt); | |
8352 | rhs2 = gimple_assign_rhs2 (stmt); | |
8353 | ||
8354 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &def_stmt, | |
8355 | &dts[0], &vectype1)) | |
8356 | return false; | |
8357 | ||
8358 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &def_stmt, | |
8359 | &dts[1], &vectype2)) | |
8360 | return false; | |
8361 | ||
8362 | if (vectype1 && vectype2 | |
8363 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
8364 | return false; | |
8365 | ||
8366 | vectype = vectype1 ? vectype1 : vectype2; | |
8367 | ||
8368 | /* Invariant comparison. */ | |
8369 | if (!vectype) | |
8370 | { | |
69a9a66f RB |
8371 | vectype = get_vectype_for_scalar_type (TREE_TYPE (rhs1)); |
8372 | if (TYPE_VECTOR_SUBPARTS (vectype) != nunits) | |
42fd8198 IE |
8373 | return false; |
8374 | } | |
8375 | else if (nunits != TYPE_VECTOR_SUBPARTS (vectype)) | |
8376 | return false; | |
8377 | ||
49e76ff1 IE |
8378 | /* Can't compare mask and non-mask types. */ |
8379 | if (vectype1 && vectype2 | |
8380 | && (VECTOR_BOOLEAN_TYPE_P (vectype1) ^ VECTOR_BOOLEAN_TYPE_P (vectype2))) | |
8381 | return false; | |
8382 | ||
8383 | /* Boolean values may have another representation in vectors | |
8384 | and therefore we prefer bit operations over comparison for | |
8385 | them (which also works for scalar masks). We store opcodes | |
8386 | to use in bitop1 and bitop2. Statement is vectorized as | |
8387 | BITOP2 (rhs1 BITOP1 rhs2) or | |
8388 | rhs1 BITOP2 (BITOP1 rhs2) | |
8389 | depending on bitop1 and bitop2 arity. */ | |
8390 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) | |
8391 | { | |
8392 | if (code == GT_EXPR) | |
8393 | { | |
8394 | bitop1 = BIT_NOT_EXPR; | |
8395 | bitop2 = BIT_AND_EXPR; | |
8396 | } | |
8397 | else if (code == GE_EXPR) | |
8398 | { | |
8399 | bitop1 = BIT_NOT_EXPR; | |
8400 | bitop2 = BIT_IOR_EXPR; | |
8401 | } | |
8402 | else if (code == LT_EXPR) | |
8403 | { | |
8404 | bitop1 = BIT_NOT_EXPR; | |
8405 | bitop2 = BIT_AND_EXPR; | |
8406 | std::swap (rhs1, rhs2); | |
264d951a | 8407 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
8408 | } |
8409 | else if (code == LE_EXPR) | |
8410 | { | |
8411 | bitop1 = BIT_NOT_EXPR; | |
8412 | bitop2 = BIT_IOR_EXPR; | |
8413 | std::swap (rhs1, rhs2); | |
264d951a | 8414 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
8415 | } |
8416 | else | |
8417 | { | |
8418 | bitop1 = BIT_XOR_EXPR; | |
8419 | if (code == EQ_EXPR) | |
8420 | bitop2 = BIT_NOT_EXPR; | |
8421 | } | |
8422 | } | |
8423 | ||
42fd8198 IE |
8424 | if (!vec_stmt) |
8425 | { | |
8426 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
49e76ff1 | 8427 | vect_model_simple_cost (stmt_info, ncopies * (1 + (bitop2 != NOP_EXPR)), |
4fc5ebf1 | 8428 | dts, ndts, NULL, NULL); |
49e76ff1 | 8429 | if (bitop1 == NOP_EXPR) |
96592eed | 8430 | return expand_vec_cmp_expr_p (vectype, mask_type, code); |
49e76ff1 IE |
8431 | else |
8432 | { | |
8433 | machine_mode mode = TYPE_MODE (vectype); | |
8434 | optab optab; | |
8435 | ||
8436 | optab = optab_for_tree_code (bitop1, vectype, optab_default); | |
8437 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8438 | return false; | |
8439 | ||
8440 | if (bitop2 != NOP_EXPR) | |
8441 | { | |
8442 | optab = optab_for_tree_code (bitop2, vectype, optab_default); | |
8443 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8444 | return false; | |
8445 | } | |
8446 | return true; | |
8447 | } | |
42fd8198 IE |
8448 | } |
8449 | ||
8450 | /* Transform. */ | |
8451 | if (!slp_node) | |
8452 | { | |
8453 | vec_oprnds0.create (1); | |
8454 | vec_oprnds1.create (1); | |
8455 | } | |
8456 | ||
8457 | /* Handle def. */ | |
8458 | lhs = gimple_assign_lhs (stmt); | |
8459 | mask = vect_create_destination_var (lhs, mask_type); | |
8460 | ||
8461 | /* Handle cmp expr. */ | |
8462 | for (j = 0; j < ncopies; j++) | |
8463 | { | |
8464 | gassign *new_stmt = NULL; | |
8465 | if (j == 0) | |
8466 | { | |
8467 | if (slp_node) | |
8468 | { | |
8469 | auto_vec<tree, 2> ops; | |
8470 | auto_vec<vec<tree>, 2> vec_defs; | |
8471 | ||
8472 | ops.safe_push (rhs1); | |
8473 | ops.safe_push (rhs2); | |
306b0c92 | 8474 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
42fd8198 IE |
8475 | vec_oprnds1 = vec_defs.pop (); |
8476 | vec_oprnds0 = vec_defs.pop (); | |
8477 | } | |
8478 | else | |
8479 | { | |
e4af0bc4 IE |
8480 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, vectype); |
8481 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, vectype); | |
42fd8198 IE |
8482 | } |
8483 | } | |
8484 | else | |
8485 | { | |
8486 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
8487 | vec_oprnds0.pop ()); | |
8488 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
8489 | vec_oprnds1.pop ()); | |
8490 | } | |
8491 | ||
8492 | if (!slp_node) | |
8493 | { | |
8494 | vec_oprnds0.quick_push (vec_rhs1); | |
8495 | vec_oprnds1.quick_push (vec_rhs2); | |
8496 | } | |
8497 | ||
8498 | /* Arguments are ready. Create the new vector stmt. */ | |
8499 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
8500 | { | |
8501 | vec_rhs2 = vec_oprnds1[i]; | |
8502 | ||
8503 | new_temp = make_ssa_name (mask); | |
49e76ff1 IE |
8504 | if (bitop1 == NOP_EXPR) |
8505 | { | |
8506 | new_stmt = gimple_build_assign (new_temp, code, | |
8507 | vec_rhs1, vec_rhs2); | |
8508 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8509 | } | |
8510 | else | |
8511 | { | |
8512 | if (bitop1 == BIT_NOT_EXPR) | |
8513 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs2); | |
8514 | else | |
8515 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs1, | |
8516 | vec_rhs2); | |
8517 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8518 | if (bitop2 != NOP_EXPR) | |
8519 | { | |
8520 | tree res = make_ssa_name (mask); | |
8521 | if (bitop2 == BIT_NOT_EXPR) | |
8522 | new_stmt = gimple_build_assign (res, bitop2, new_temp); | |
8523 | else | |
8524 | new_stmt = gimple_build_assign (res, bitop2, vec_rhs1, | |
8525 | new_temp); | |
8526 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8527 | } | |
8528 | } | |
42fd8198 IE |
8529 | if (slp_node) |
8530 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
8531 | } | |
8532 | ||
8533 | if (slp_node) | |
8534 | continue; | |
8535 | ||
8536 | if (j == 0) | |
8537 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
8538 | else | |
8539 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
8540 | ||
8541 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
8542 | } | |
8543 | ||
8544 | vec_oprnds0.release (); | |
8545 | vec_oprnds1.release (); | |
8546 | ||
8547 | return true; | |
8548 | } | |
ebfd146a | 8549 | |
68a0f2ff RS |
8550 | /* If SLP_NODE is nonnull, return true if vectorizable_live_operation |
8551 | can handle all live statements in the node. Otherwise return true | |
8552 | if STMT is not live or if vectorizable_live_operation can handle it. | |
8553 | GSI and VEC_STMT are as for vectorizable_live_operation. */ | |
8554 | ||
8555 | static bool | |
8556 | can_vectorize_live_stmts (gimple *stmt, gimple_stmt_iterator *gsi, | |
8557 | slp_tree slp_node, gimple **vec_stmt) | |
8558 | { | |
8559 | if (slp_node) | |
8560 | { | |
8561 | gimple *slp_stmt; | |
8562 | unsigned int i; | |
8563 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, slp_stmt) | |
8564 | { | |
8565 | stmt_vec_info slp_stmt_info = vinfo_for_stmt (slp_stmt); | |
8566 | if (STMT_VINFO_LIVE_P (slp_stmt_info) | |
8567 | && !vectorizable_live_operation (slp_stmt, gsi, slp_node, i, | |
8568 | vec_stmt)) | |
8569 | return false; | |
8570 | } | |
8571 | } | |
8572 | else if (STMT_VINFO_LIVE_P (vinfo_for_stmt (stmt)) | |
8573 | && !vectorizable_live_operation (stmt, gsi, slp_node, -1, vec_stmt)) | |
8574 | return false; | |
8575 | ||
8576 | return true; | |
8577 | } | |
8578 | ||
8644a673 | 8579 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
8580 | |
8581 | bool | |
891ad31c RB |
8582 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node, |
8583 | slp_instance node_instance) | |
ebfd146a | 8584 | { |
8644a673 | 8585 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 8586 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 8587 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 8588 | bool ok; |
355fe088 | 8589 | gimple *pattern_stmt; |
363477c0 | 8590 | gimple_seq pattern_def_seq; |
ebfd146a | 8591 | |
73fbfcad | 8592 | if (dump_enabled_p ()) |
ebfd146a | 8593 | { |
78c60e3d SS |
8594 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
8595 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 8596 | } |
ebfd146a | 8597 | |
1825a1f3 | 8598 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 8599 | { |
73fbfcad | 8600 | if (dump_enabled_p ()) |
78c60e3d | 8601 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8602 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
8603 | |
8604 | return false; | |
8605 | } | |
b8698a0f L |
8606 | |
8607 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
8608 | to include: |
8609 | - the COND_EXPR which is the loop exit condition | |
8610 | - any LABEL_EXPRs in the loop | |
b8698a0f | 8611 | - computations that are used only for array indexing or loop control. |
8644a673 | 8612 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 8613 | instance, therefore, all the statements are relevant. |
ebfd146a | 8614 | |
d092494c | 8615 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 8616 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
8617 | statements. In basic blocks we are called from some SLP instance |
8618 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
8619 | already will be part of SLP instance. */ | |
83197f37 IR |
8620 | |
8621 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 8622 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 8623 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 8624 | { |
9d5e7640 | 8625 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 8626 | && pattern_stmt |
9d5e7640 IR |
8627 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
8628 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
8629 | { | |
83197f37 | 8630 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
8631 | stmt = pattern_stmt; |
8632 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 8633 | if (dump_enabled_p ()) |
9d5e7640 | 8634 | { |
78c60e3d SS |
8635 | dump_printf_loc (MSG_NOTE, vect_location, |
8636 | "==> examining pattern statement: "); | |
8637 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
8638 | } |
8639 | } | |
8640 | else | |
8641 | { | |
73fbfcad | 8642 | if (dump_enabled_p ()) |
e645e942 | 8643 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 8644 | |
9d5e7640 IR |
8645 | return true; |
8646 | } | |
8644a673 | 8647 | } |
83197f37 | 8648 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 8649 | && node == NULL |
83197f37 IR |
8650 | && pattern_stmt |
8651 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
8652 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
8653 | { | |
8654 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 8655 | if (dump_enabled_p ()) |
83197f37 | 8656 | { |
78c60e3d SS |
8657 | dump_printf_loc (MSG_NOTE, vect_location, |
8658 | "==> examining pattern statement: "); | |
8659 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
8660 | } |
8661 | ||
891ad31c RB |
8662 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node, |
8663 | node_instance)) | |
83197f37 IR |
8664 | return false; |
8665 | } | |
ebfd146a | 8666 | |
1107f3ae | 8667 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 8668 | && node == NULL |
363477c0 | 8669 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 8670 | { |
363477c0 | 8671 | gimple_stmt_iterator si; |
1107f3ae | 8672 | |
363477c0 JJ |
8673 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
8674 | { | |
355fe088 | 8675 | gimple *pattern_def_stmt = gsi_stmt (si); |
363477c0 JJ |
8676 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) |
8677 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
8678 | { | |
8679 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 8680 | if (dump_enabled_p ()) |
363477c0 | 8681 | { |
78c60e3d SS |
8682 | dump_printf_loc (MSG_NOTE, vect_location, |
8683 | "==> examining pattern def statement: "); | |
8684 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 8685 | } |
1107f3ae | 8686 | |
363477c0 | 8687 | if (!vect_analyze_stmt (pattern_def_stmt, |
891ad31c | 8688 | need_to_vectorize, node, node_instance)) |
363477c0 JJ |
8689 | return false; |
8690 | } | |
8691 | } | |
8692 | } | |
1107f3ae | 8693 | |
8644a673 IR |
8694 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
8695 | { | |
8696 | case vect_internal_def: | |
8697 | break; | |
ebfd146a | 8698 | |
8644a673 | 8699 | case vect_reduction_def: |
7c5222ff | 8700 | case vect_nested_cycle: |
14a61437 RB |
8701 | gcc_assert (!bb_vinfo |
8702 | && (relevance == vect_used_in_outer | |
8703 | || relevance == vect_used_in_outer_by_reduction | |
8704 | || relevance == vect_used_by_reduction | |
b28ead45 AH |
8705 | || relevance == vect_unused_in_scope |
8706 | || relevance == vect_used_only_live)); | |
8644a673 IR |
8707 | break; |
8708 | ||
8709 | case vect_induction_def: | |
e7baeb39 RB |
8710 | gcc_assert (!bb_vinfo); |
8711 | break; | |
8712 | ||
8644a673 IR |
8713 | case vect_constant_def: |
8714 | case vect_external_def: | |
8715 | case vect_unknown_def_type: | |
8716 | default: | |
8717 | gcc_unreachable (); | |
8718 | } | |
ebfd146a | 8719 | |
8644a673 | 8720 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 8721 | { |
8644a673 | 8722 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
8723 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
8724 | || (is_gimple_call (stmt) | |
8725 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 8726 | *need_to_vectorize = true; |
ebfd146a IR |
8727 | } |
8728 | ||
b1af7da6 RB |
8729 | if (PURE_SLP_STMT (stmt_info) && !node) |
8730 | { | |
8731 | dump_printf_loc (MSG_NOTE, vect_location, | |
8732 | "handled only by SLP analysis\n"); | |
8733 | return true; | |
8734 | } | |
8735 | ||
8736 | ok = true; | |
8737 | if (!bb_vinfo | |
8738 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
8739 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
8740 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8741 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8742 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8743 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8744 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8745 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8746 | || vectorizable_call (stmt, NULL, NULL, node) | |
8747 | || vectorizable_store (stmt, NULL, NULL, node) | |
891ad31c | 8748 | || vectorizable_reduction (stmt, NULL, NULL, node, node_instance) |
e7baeb39 | 8749 | || vectorizable_induction (stmt, NULL, NULL, node) |
42fd8198 IE |
8750 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8751 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 RB |
8752 | else |
8753 | { | |
8754 | if (bb_vinfo) | |
8755 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8756 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8757 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8758 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8759 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8760 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8761 | || vectorizable_call (stmt, NULL, NULL, node) | |
8762 | || vectorizable_store (stmt, NULL, NULL, node) | |
42fd8198 IE |
8763 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8764 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 | 8765 | } |
8644a673 IR |
8766 | |
8767 | if (!ok) | |
ebfd146a | 8768 | { |
73fbfcad | 8769 | if (dump_enabled_p ()) |
8644a673 | 8770 | { |
78c60e3d SS |
8771 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8772 | "not vectorized: relevant stmt not "); | |
8773 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
8774 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 8775 | } |
b8698a0f | 8776 | |
ebfd146a IR |
8777 | return false; |
8778 | } | |
8779 | ||
a70d6342 IR |
8780 | if (bb_vinfo) |
8781 | return true; | |
8782 | ||
8644a673 IR |
8783 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
8784 | need extra handling, except for vectorizable reductions. */ | |
68a0f2ff RS |
8785 | if (STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
8786 | && !can_vectorize_live_stmts (stmt, NULL, node, NULL)) | |
ebfd146a | 8787 | { |
73fbfcad | 8788 | if (dump_enabled_p ()) |
8644a673 | 8789 | { |
78c60e3d | 8790 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
68a0f2ff | 8791 | "not vectorized: live stmt not supported: "); |
78c60e3d | 8792 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); |
8644a673 | 8793 | } |
b8698a0f | 8794 | |
8644a673 | 8795 | return false; |
ebfd146a IR |
8796 | } |
8797 | ||
ebfd146a IR |
8798 | return true; |
8799 | } | |
8800 | ||
8801 | ||
8802 | /* Function vect_transform_stmt. | |
8803 | ||
8804 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
8805 | ||
8806 | bool | |
355fe088 | 8807 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 8808 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
8809 | slp_instance slp_node_instance) |
8810 | { | |
8811 | bool is_store = false; | |
355fe088 | 8812 | gimple *vec_stmt = NULL; |
ebfd146a | 8813 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
ebfd146a | 8814 | bool done; |
ebfd146a | 8815 | |
fce57248 | 8816 | gcc_assert (slp_node || !PURE_SLP_STMT (stmt_info)); |
355fe088 | 8817 | gimple *old_vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 8818 | |
ebfd146a IR |
8819 | switch (STMT_VINFO_TYPE (stmt_info)) |
8820 | { | |
8821 | case type_demotion_vec_info_type: | |
ebfd146a | 8822 | case type_promotion_vec_info_type: |
ebfd146a IR |
8823 | case type_conversion_vec_info_type: |
8824 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
8825 | gcc_assert (done); | |
8826 | break; | |
8827 | ||
8828 | case induc_vec_info_type: | |
e7baeb39 | 8829 | done = vectorizable_induction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
8830 | gcc_assert (done); |
8831 | break; | |
8832 | ||
9dc3f7de IR |
8833 | case shift_vec_info_type: |
8834 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
8835 | gcc_assert (done); | |
8836 | break; | |
8837 | ||
ebfd146a IR |
8838 | case op_vec_info_type: |
8839 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
8840 | gcc_assert (done); | |
8841 | break; | |
8842 | ||
8843 | case assignment_vec_info_type: | |
8844 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
8845 | gcc_assert (done); | |
8846 | break; | |
8847 | ||
8848 | case load_vec_info_type: | |
b8698a0f | 8849 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
8850 | slp_node_instance); |
8851 | gcc_assert (done); | |
8852 | break; | |
8853 | ||
8854 | case store_vec_info_type: | |
8855 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
8856 | gcc_assert (done); | |
0d0293ac | 8857 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
8858 | { |
8859 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 8860 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
8861 | one are skipped, and there vec_stmt_info shouldn't be freed |
8862 | meanwhile. */ | |
0d0293ac | 8863 | *grouped_store = true; |
ebfd146a IR |
8864 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
8865 | is_store = true; | |
8866 | } | |
8867 | else | |
8868 | is_store = true; | |
8869 | break; | |
8870 | ||
8871 | case condition_vec_info_type: | |
f7e531cf | 8872 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
8873 | gcc_assert (done); |
8874 | break; | |
8875 | ||
42fd8198 IE |
8876 | case comparison_vec_info_type: |
8877 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node); | |
8878 | gcc_assert (done); | |
8879 | break; | |
8880 | ||
ebfd146a | 8881 | case call_vec_info_type: |
190c2236 | 8882 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 8883 | stmt = gsi_stmt (*gsi); |
8e4284d0 | 8884 | if (gimple_call_internal_p (stmt, IFN_MASK_STORE)) |
5ce9450f | 8885 | is_store = true; |
ebfd146a IR |
8886 | break; |
8887 | ||
0136f8f0 AH |
8888 | case call_simd_clone_vec_info_type: |
8889 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node); | |
8890 | stmt = gsi_stmt (*gsi); | |
8891 | break; | |
8892 | ||
ebfd146a | 8893 | case reduc_vec_info_type: |
891ad31c RB |
8894 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node, |
8895 | slp_node_instance); | |
ebfd146a IR |
8896 | gcc_assert (done); |
8897 | break; | |
8898 | ||
8899 | default: | |
8900 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
8901 | { | |
73fbfcad | 8902 | if (dump_enabled_p ()) |
78c60e3d | 8903 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8904 | "stmt not supported.\n"); |
ebfd146a IR |
8905 | gcc_unreachable (); |
8906 | } | |
8907 | } | |
8908 | ||
225ce44b RB |
8909 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
8910 | This would break hybrid SLP vectorization. */ | |
8911 | if (slp_node) | |
d90f8440 RB |
8912 | gcc_assert (!vec_stmt |
8913 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt); | |
225ce44b | 8914 | |
ebfd146a IR |
8915 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
8916 | is being vectorized, but outside the immediately enclosing loop. */ | |
8917 | if (vec_stmt | |
a70d6342 IR |
8918 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
8919 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
8920 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
8921 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
8922 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 8923 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 8924 | vect_used_in_outer_by_reduction)) |
ebfd146a | 8925 | { |
a70d6342 IR |
8926 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
8927 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
8928 | imm_use_iterator imm_iter; |
8929 | use_operand_p use_p; | |
8930 | tree scalar_dest; | |
355fe088 | 8931 | gimple *exit_phi; |
ebfd146a | 8932 | |
73fbfcad | 8933 | if (dump_enabled_p ()) |
78c60e3d | 8934 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 8935 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
8936 | |
8937 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
8938 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
8939 | STMT). */ | |
8940 | if (gimple_code (stmt) == GIMPLE_PHI) | |
8941 | scalar_dest = PHI_RESULT (stmt); | |
8942 | else | |
8943 | scalar_dest = gimple_assign_lhs (stmt); | |
8944 | ||
8945 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
8946 | { | |
8947 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
8948 | { | |
8949 | exit_phi = USE_STMT (use_p); | |
8950 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
8951 | } | |
8952 | } | |
8953 | } | |
8954 | ||
8955 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
8956 | being vectorized. */ | |
68a0f2ff | 8957 | if (STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) |
ebfd146a | 8958 | { |
68a0f2ff | 8959 | done = can_vectorize_live_stmts (stmt, gsi, slp_node, &vec_stmt); |
ebfd146a IR |
8960 | gcc_assert (done); |
8961 | } | |
8962 | ||
8963 | if (vec_stmt) | |
83197f37 | 8964 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 8965 | |
b8698a0f | 8966 | return is_store; |
ebfd146a IR |
8967 | } |
8968 | ||
8969 | ||
b8698a0f | 8970 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
8971 | stmt_vec_info. */ |
8972 | ||
8973 | void | |
355fe088 | 8974 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 8975 | { |
355fe088 TS |
8976 | gimple *next = first_stmt; |
8977 | gimple *tmp; | |
ebfd146a IR |
8978 | gimple_stmt_iterator next_si; |
8979 | ||
8980 | while (next) | |
8981 | { | |
78048b1c JJ |
8982 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
8983 | ||
8984 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
8985 | if (is_pattern_stmt_p (stmt_info)) | |
8986 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
8987 | /* Free the attached stmt_vec_info and remove the stmt. */ |
8988 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 8989 | unlink_stmt_vdef (next); |
ebfd146a | 8990 | gsi_remove (&next_si, true); |
3d3f2249 | 8991 | release_defs (next); |
ebfd146a IR |
8992 | free_stmt_vec_info (next); |
8993 | next = tmp; | |
8994 | } | |
8995 | } | |
8996 | ||
8997 | ||
8998 | /* Function new_stmt_vec_info. | |
8999 | ||
9000 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
9001 | ||
9002 | stmt_vec_info | |
310213d4 | 9003 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
9004 | { |
9005 | stmt_vec_info res; | |
9006 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
9007 | ||
9008 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
9009 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 9010 | res->vinfo = vinfo; |
8644a673 | 9011 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
9012 | STMT_VINFO_LIVE_P (res) = false; |
9013 | STMT_VINFO_VECTYPE (res) = NULL; | |
9014 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 9015 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
9016 | STMT_VINFO_IN_PATTERN_P (res) = false; |
9017 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 9018 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 9019 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 9020 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
7e16ce79 | 9021 | STMT_VINFO_VEC_CONST_COND_REDUC_CODE (res) = ERROR_MARK; |
ebfd146a | 9022 | |
ebfd146a IR |
9023 | if (gimple_code (stmt) == GIMPLE_PHI |
9024 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
9025 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
9026 | else | |
8644a673 IR |
9027 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
9028 | ||
9771b263 | 9029 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 9030 | STMT_SLP_TYPE (res) = loop_vect; |
78810bd3 RB |
9031 | STMT_VINFO_NUM_SLP_USES (res) = 0; |
9032 | ||
e14c1050 IR |
9033 | GROUP_FIRST_ELEMENT (res) = NULL; |
9034 | GROUP_NEXT_ELEMENT (res) = NULL; | |
9035 | GROUP_SIZE (res) = 0; | |
9036 | GROUP_STORE_COUNT (res) = 0; | |
9037 | GROUP_GAP (res) = 0; | |
9038 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
9039 | |
9040 | return res; | |
9041 | } | |
9042 | ||
9043 | ||
9044 | /* Create a hash table for stmt_vec_info. */ | |
9045 | ||
9046 | void | |
9047 | init_stmt_vec_info_vec (void) | |
9048 | { | |
9771b263 DN |
9049 | gcc_assert (!stmt_vec_info_vec.exists ()); |
9050 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
9051 | } |
9052 | ||
9053 | ||
9054 | /* Free hash table for stmt_vec_info. */ | |
9055 | ||
9056 | void | |
9057 | free_stmt_vec_info_vec (void) | |
9058 | { | |
93675444 | 9059 | unsigned int i; |
3161455c | 9060 | stmt_vec_info info; |
93675444 JJ |
9061 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) |
9062 | if (info != NULL) | |
3161455c | 9063 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
9771b263 DN |
9064 | gcc_assert (stmt_vec_info_vec.exists ()); |
9065 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
9066 | } |
9067 | ||
9068 | ||
9069 | /* Free stmt vectorization related info. */ | |
9070 | ||
9071 | void | |
355fe088 | 9072 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
9073 | { |
9074 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9075 | ||
9076 | if (!stmt_info) | |
9077 | return; | |
9078 | ||
78048b1c JJ |
9079 | /* Check if this statement has a related "pattern stmt" |
9080 | (introduced by the vectorizer during the pattern recognition | |
9081 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
9082 | too. */ | |
9083 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
9084 | { | |
9085 | stmt_vec_info patt_info | |
9086 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
9087 | if (patt_info) | |
9088 | { | |
363477c0 | 9089 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
355fe088 | 9090 | gimple *patt_stmt = STMT_VINFO_STMT (patt_info); |
f0281fde RB |
9091 | gimple_set_bb (patt_stmt, NULL); |
9092 | tree lhs = gimple_get_lhs (patt_stmt); | |
e6f5c25d | 9093 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde | 9094 | release_ssa_name (lhs); |
363477c0 JJ |
9095 | if (seq) |
9096 | { | |
9097 | gimple_stmt_iterator si; | |
9098 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
f0281fde | 9099 | { |
355fe088 | 9100 | gimple *seq_stmt = gsi_stmt (si); |
f0281fde | 9101 | gimple_set_bb (seq_stmt, NULL); |
7532abf2 | 9102 | lhs = gimple_get_lhs (seq_stmt); |
e6f5c25d | 9103 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde RB |
9104 | release_ssa_name (lhs); |
9105 | free_stmt_vec_info (seq_stmt); | |
9106 | } | |
363477c0 | 9107 | } |
f0281fde | 9108 | free_stmt_vec_info (patt_stmt); |
78048b1c JJ |
9109 | } |
9110 | } | |
9111 | ||
9771b263 | 9112 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 9113 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
9114 | set_vinfo_for_stmt (stmt, NULL); |
9115 | free (stmt_info); | |
9116 | } | |
9117 | ||
9118 | ||
bb67d9c7 | 9119 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 9120 | |
bb67d9c7 | 9121 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
9122 | by the target. */ |
9123 | ||
bb67d9c7 RG |
9124 | static tree |
9125 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a | 9126 | { |
c7d97b28 | 9127 | tree orig_scalar_type = scalar_type; |
3bd8f481 | 9128 | scalar_mode inner_mode; |
ef4bddc2 | 9129 | machine_mode simd_mode; |
ebfd146a IR |
9130 | int nunits; |
9131 | tree vectype; | |
9132 | ||
3bd8f481 RS |
9133 | if (!is_int_mode (TYPE_MODE (scalar_type), &inner_mode) |
9134 | && !is_float_mode (TYPE_MODE (scalar_type), &inner_mode)) | |
ebfd146a IR |
9135 | return NULL_TREE; |
9136 | ||
3bd8f481 | 9137 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
48f2e373 | 9138 | |
7b7b1813 RG |
9139 | /* For vector types of elements whose mode precision doesn't |
9140 | match their types precision we use a element type of mode | |
9141 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
9142 | they support the proper result truncation/extension. |
9143 | We also make sure to build vector types with INTEGER_TYPE | |
9144 | component type only. */ | |
6d7971b8 | 9145 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
9146 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
9147 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
9148 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
9149 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 9150 | |
ccbf5bb4 RG |
9151 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
9152 | When the component mode passes the above test simply use a type | |
9153 | corresponding to that mode. The theory is that any use that | |
9154 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 9155 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 9156 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
9157 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
9158 | ||
9159 | /* We can't build a vector type of elements with alignment bigger than | |
9160 | their size. */ | |
dfc2e2ac | 9161 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
9162 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
9163 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 9164 | |
dfc2e2ac RB |
9165 | /* If we felt back to using the mode fail if there was |
9166 | no scalar type for it. */ | |
9167 | if (scalar_type == NULL_TREE) | |
9168 | return NULL_TREE; | |
9169 | ||
bb67d9c7 RG |
9170 | /* If no size was supplied use the mode the target prefers. Otherwise |
9171 | lookup a vector mode of the specified size. */ | |
9172 | if (size == 0) | |
9173 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
9da15d40 RS |
9174 | else if (!mode_for_vector (inner_mode, size / nbytes).exists (&simd_mode)) |
9175 | return NULL_TREE; | |
cc4b5170 | 9176 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
4c8fd8ac JB |
9177 | /* NOTE: nunits == 1 is allowed to support single element vector types. */ |
9178 | if (nunits < 1) | |
cc4b5170 | 9179 | return NULL_TREE; |
ebfd146a IR |
9180 | |
9181 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
9182 | |
9183 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
9184 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 9185 | return NULL_TREE; |
ebfd146a | 9186 | |
c7d97b28 RB |
9187 | /* Re-attach the address-space qualifier if we canonicalized the scalar |
9188 | type. */ | |
9189 | if (TYPE_ADDR_SPACE (orig_scalar_type) != TYPE_ADDR_SPACE (vectype)) | |
9190 | return build_qualified_type | |
9191 | (vectype, KEEP_QUAL_ADDR_SPACE (TYPE_QUALS (orig_scalar_type))); | |
9192 | ||
ebfd146a IR |
9193 | return vectype; |
9194 | } | |
9195 | ||
bb67d9c7 RG |
9196 | unsigned int current_vector_size; |
9197 | ||
9198 | /* Function get_vectype_for_scalar_type. | |
9199 | ||
9200 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
9201 | by the target. */ | |
9202 | ||
9203 | tree | |
9204 | get_vectype_for_scalar_type (tree scalar_type) | |
9205 | { | |
9206 | tree vectype; | |
9207 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
9208 | current_vector_size); | |
9209 | if (vectype | |
9210 | && current_vector_size == 0) | |
9211 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
9212 | return vectype; | |
9213 | } | |
9214 | ||
42fd8198 IE |
9215 | /* Function get_mask_type_for_scalar_type. |
9216 | ||
9217 | Returns the mask type corresponding to a result of comparison | |
9218 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
9219 | ||
9220 | tree | |
9221 | get_mask_type_for_scalar_type (tree scalar_type) | |
9222 | { | |
9223 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
9224 | ||
9225 | if (!vectype) | |
9226 | return NULL; | |
9227 | ||
9228 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
9229 | current_vector_size); | |
9230 | } | |
9231 | ||
b690cc0f RG |
9232 | /* Function get_same_sized_vectype |
9233 | ||
9234 | Returns a vector type corresponding to SCALAR_TYPE of size | |
9235 | VECTOR_TYPE if supported by the target. */ | |
9236 | ||
9237 | tree | |
bb67d9c7 | 9238 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 9239 | { |
2568d8a1 | 9240 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type)) |
9f47c7e5 IE |
9241 | return build_same_sized_truth_vector_type (vector_type); |
9242 | ||
bb67d9c7 RG |
9243 | return get_vectype_for_scalar_type_and_size |
9244 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
9245 | } |
9246 | ||
ebfd146a IR |
9247 | /* Function vect_is_simple_use. |
9248 | ||
9249 | Input: | |
81c40241 RB |
9250 | VINFO - the vect info of the loop or basic block that is being vectorized. |
9251 | OPERAND - operand in the loop or bb. | |
9252 | Output: | |
9253 | DEF_STMT - the defining stmt in case OPERAND is an SSA_NAME. | |
9254 | DT - the type of definition | |
ebfd146a IR |
9255 | |
9256 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 9257 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 9258 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 9259 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
9260 | is the case in reduction/induction computations). |
9261 | For basic blocks, supportable operands are constants and bb invariants. | |
9262 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
9263 | |
9264 | bool | |
81c40241 RB |
9265 | vect_is_simple_use (tree operand, vec_info *vinfo, |
9266 | gimple **def_stmt, enum vect_def_type *dt) | |
b8698a0f | 9267 | { |
ebfd146a | 9268 | *def_stmt = NULL; |
3fc356dc | 9269 | *dt = vect_unknown_def_type; |
b8698a0f | 9270 | |
73fbfcad | 9271 | if (dump_enabled_p ()) |
ebfd146a | 9272 | { |
78c60e3d SS |
9273 | dump_printf_loc (MSG_NOTE, vect_location, |
9274 | "vect_is_simple_use: operand "); | |
9275 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 9276 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 9277 | } |
b8698a0f | 9278 | |
b758f602 | 9279 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
9280 | { |
9281 | *dt = vect_constant_def; | |
9282 | return true; | |
9283 | } | |
b8698a0f | 9284 | |
ebfd146a IR |
9285 | if (is_gimple_min_invariant (operand)) |
9286 | { | |
8644a673 | 9287 | *dt = vect_external_def; |
ebfd146a IR |
9288 | return true; |
9289 | } | |
9290 | ||
ebfd146a IR |
9291 | if (TREE_CODE (operand) != SSA_NAME) |
9292 | { | |
73fbfcad | 9293 | if (dump_enabled_p ()) |
af29617a AH |
9294 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9295 | "not ssa-name.\n"); | |
ebfd146a IR |
9296 | return false; |
9297 | } | |
b8698a0f | 9298 | |
3fc356dc | 9299 | if (SSA_NAME_IS_DEFAULT_DEF (operand)) |
ebfd146a | 9300 | { |
3fc356dc RB |
9301 | *dt = vect_external_def; |
9302 | return true; | |
ebfd146a IR |
9303 | } |
9304 | ||
3fc356dc | 9305 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
73fbfcad | 9306 | if (dump_enabled_p ()) |
ebfd146a | 9307 | { |
78c60e3d SS |
9308 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
9309 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
9310 | } |
9311 | ||
61d371eb | 9312 | if (! vect_stmt_in_region_p (vinfo, *def_stmt)) |
8644a673 | 9313 | *dt = vect_external_def; |
ebfd146a IR |
9314 | else |
9315 | { | |
3fc356dc | 9316 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (*def_stmt); |
603cca93 | 9317 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
ebfd146a IR |
9318 | } |
9319 | ||
2e8ab70c RB |
9320 | if (dump_enabled_p ()) |
9321 | { | |
9322 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: "); | |
9323 | switch (*dt) | |
9324 | { | |
9325 | case vect_uninitialized_def: | |
9326 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
9327 | break; | |
9328 | case vect_constant_def: | |
9329 | dump_printf (MSG_NOTE, "constant\n"); | |
9330 | break; | |
9331 | case vect_external_def: | |
9332 | dump_printf (MSG_NOTE, "external\n"); | |
9333 | break; | |
9334 | case vect_internal_def: | |
9335 | dump_printf (MSG_NOTE, "internal\n"); | |
9336 | break; | |
9337 | case vect_induction_def: | |
9338 | dump_printf (MSG_NOTE, "induction\n"); | |
9339 | break; | |
9340 | case vect_reduction_def: | |
9341 | dump_printf (MSG_NOTE, "reduction\n"); | |
9342 | break; | |
9343 | case vect_double_reduction_def: | |
9344 | dump_printf (MSG_NOTE, "double reduction\n"); | |
9345 | break; | |
9346 | case vect_nested_cycle: | |
9347 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
9348 | break; | |
9349 | case vect_unknown_def_type: | |
9350 | dump_printf (MSG_NOTE, "unknown\n"); | |
9351 | break; | |
9352 | } | |
9353 | } | |
9354 | ||
81c40241 | 9355 | if (*dt == vect_unknown_def_type) |
ebfd146a | 9356 | { |
73fbfcad | 9357 | if (dump_enabled_p ()) |
78c60e3d | 9358 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9359 | "Unsupported pattern.\n"); |
ebfd146a IR |
9360 | return false; |
9361 | } | |
9362 | ||
ebfd146a IR |
9363 | switch (gimple_code (*def_stmt)) |
9364 | { | |
9365 | case GIMPLE_PHI: | |
ebfd146a | 9366 | case GIMPLE_ASSIGN: |
ebfd146a | 9367 | case GIMPLE_CALL: |
81c40241 | 9368 | break; |
ebfd146a | 9369 | default: |
73fbfcad | 9370 | if (dump_enabled_p ()) |
78c60e3d | 9371 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9372 | "unsupported defining stmt:\n"); |
ebfd146a IR |
9373 | return false; |
9374 | } | |
9375 | ||
9376 | return true; | |
9377 | } | |
9378 | ||
81c40241 | 9379 | /* Function vect_is_simple_use. |
b690cc0f | 9380 | |
81c40241 | 9381 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
9382 | type of OPERAND and stores it to *VECTYPE. If the definition of |
9383 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
9384 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
9385 | is responsible to compute the best suited vector type for the | |
9386 | scalar operand. */ | |
9387 | ||
9388 | bool | |
81c40241 RB |
9389 | vect_is_simple_use (tree operand, vec_info *vinfo, |
9390 | gimple **def_stmt, enum vect_def_type *dt, tree *vectype) | |
b690cc0f | 9391 | { |
81c40241 | 9392 | if (!vect_is_simple_use (operand, vinfo, def_stmt, dt)) |
b690cc0f RG |
9393 | return false; |
9394 | ||
9395 | /* Now get a vector type if the def is internal, otherwise supply | |
9396 | NULL_TREE and leave it up to the caller to figure out a proper | |
9397 | type for the use stmt. */ | |
9398 | if (*dt == vect_internal_def | |
9399 | || *dt == vect_induction_def | |
9400 | || *dt == vect_reduction_def | |
9401 | || *dt == vect_double_reduction_def | |
9402 | || *dt == vect_nested_cycle) | |
9403 | { | |
9404 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
9405 | |
9406 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
9407 | && !STMT_VINFO_RELEVANT (stmt_info) | |
9408 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 9409 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 9410 | |
b690cc0f RG |
9411 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
9412 | gcc_assert (*vectype != NULL_TREE); | |
9413 | } | |
9414 | else if (*dt == vect_uninitialized_def | |
9415 | || *dt == vect_constant_def | |
9416 | || *dt == vect_external_def) | |
9417 | *vectype = NULL_TREE; | |
9418 | else | |
9419 | gcc_unreachable (); | |
9420 | ||
9421 | return true; | |
9422 | } | |
9423 | ||
ebfd146a IR |
9424 | |
9425 | /* Function supportable_widening_operation | |
9426 | ||
b8698a0f L |
9427 | Check whether an operation represented by the code CODE is a |
9428 | widening operation that is supported by the target platform in | |
b690cc0f RG |
9429 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
9430 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 9431 | |
ebfd146a IR |
9432 | Widening operations we currently support are NOP (CONVERT), FLOAT |
9433 | and WIDEN_MULT. This function checks if these operations are supported | |
9434 | by the target platform either directly (via vector tree-codes), or via | |
9435 | target builtins. | |
9436 | ||
9437 | Output: | |
b8698a0f L |
9438 | - CODE1 and CODE2 are codes of vector operations to be used when |
9439 | vectorizing the operation, if available. | |
ebfd146a IR |
9440 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
9441 | case of multi-step conversion (like char->short->int - in that case | |
9442 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
9443 | - INTERM_TYPES contains the intermediate type required to perform the |
9444 | widening operation (short in the above example). */ | |
ebfd146a IR |
9445 | |
9446 | bool | |
355fe088 | 9447 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 9448 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
9449 | enum tree_code *code1, enum tree_code *code2, |
9450 | int *multi_step_cvt, | |
9771b263 | 9451 | vec<tree> *interm_types) |
ebfd146a IR |
9452 | { |
9453 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9454 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 9455 | struct loop *vect_loop = NULL; |
ef4bddc2 | 9456 | machine_mode vec_mode; |
81f40b79 | 9457 | enum insn_code icode1, icode2; |
ebfd146a | 9458 | optab optab1, optab2; |
b690cc0f RG |
9459 | tree vectype = vectype_in; |
9460 | tree wide_vectype = vectype_out; | |
ebfd146a | 9461 | enum tree_code c1, c2; |
4a00c761 JJ |
9462 | int i; |
9463 | tree prev_type, intermediate_type; | |
ef4bddc2 | 9464 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 9465 | optab optab3, optab4; |
ebfd146a | 9466 | |
4a00c761 | 9467 | *multi_step_cvt = 0; |
4ef69dfc IR |
9468 | if (loop_info) |
9469 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
9470 | ||
ebfd146a IR |
9471 | switch (code) |
9472 | { | |
9473 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
9474 | /* The result of a vectorized widening operation usually requires |
9475 | two vectors (because the widened results do not fit into one vector). | |
9476 | The generated vector results would normally be expected to be | |
9477 | generated in the same order as in the original scalar computation, | |
9478 | i.e. if 8 results are generated in each vector iteration, they are | |
9479 | to be organized as follows: | |
9480 | vect1: [res1,res2,res3,res4], | |
9481 | vect2: [res5,res6,res7,res8]. | |
9482 | ||
9483 | However, in the special case that the result of the widening | |
9484 | operation is used in a reduction computation only, the order doesn't | |
9485 | matter (because when vectorizing a reduction we change the order of | |
9486 | the computation). Some targets can take advantage of this and | |
9487 | generate more efficient code. For example, targets like Altivec, | |
9488 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
9489 | generate the following vectors: | |
9490 | vect1: [res1,res3,res5,res7], | |
9491 | vect2: [res2,res4,res6,res8]. | |
9492 | ||
9493 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
9494 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
9495 | iterations in parallel). We therefore don't allow to change the | |
9496 | order of the computation in the inner-loop during outer-loop | |
9497 | vectorization. */ | |
9498 | /* TODO: Another case in which order doesn't *really* matter is when we | |
9499 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
9500 | Normally, pack_trunc performs an even/odd permute, whereas the | |
9501 | repack from an even/odd expansion would be an interleave, which | |
9502 | would be significantly simpler for e.g. AVX2. */ | |
9503 | /* In any case, in order to avoid duplicating the code below, recurse | |
9504 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
9505 | are properly set up for the caller. If we fail, we'll continue with | |
9506 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
9507 | if (vect_loop | |
9508 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
9509 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
9510 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
9511 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
9512 | code1, code2, multi_step_cvt, |
9513 | interm_types)) | |
ebc047a2 CH |
9514 | { |
9515 | /* Elements in a vector with vect_used_by_reduction property cannot | |
9516 | be reordered if the use chain with this property does not have the | |
9517 | same operation. One such an example is s += a * b, where elements | |
9518 | in a and b cannot be reordered. Here we check if the vector defined | |
9519 | by STMT is only directly used in the reduction statement. */ | |
9520 | tree lhs = gimple_assign_lhs (stmt); | |
9521 | use_operand_p dummy; | |
355fe088 | 9522 | gimple *use_stmt; |
ebc047a2 CH |
9523 | stmt_vec_info use_stmt_info = NULL; |
9524 | if (single_imm_use (lhs, &dummy, &use_stmt) | |
9525 | && (use_stmt_info = vinfo_for_stmt (use_stmt)) | |
9526 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
9527 | return true; | |
9528 | } | |
4a00c761 JJ |
9529 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
9530 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
9531 | break; |
9532 | ||
81c40241 RB |
9533 | case DOT_PROD_EXPR: |
9534 | c1 = DOT_PROD_EXPR; | |
9535 | c2 = DOT_PROD_EXPR; | |
9536 | break; | |
9537 | ||
9538 | case SAD_EXPR: | |
9539 | c1 = SAD_EXPR; | |
9540 | c2 = SAD_EXPR; | |
9541 | break; | |
9542 | ||
6ae6116f RH |
9543 | case VEC_WIDEN_MULT_EVEN_EXPR: |
9544 | /* Support the recursion induced just above. */ | |
9545 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
9546 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
9547 | break; | |
9548 | ||
36ba4aae | 9549 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
9550 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
9551 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
9552 | break; |
9553 | ||
ebfd146a | 9554 | CASE_CONVERT: |
4a00c761 JJ |
9555 | c1 = VEC_UNPACK_LO_EXPR; |
9556 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
9557 | break; |
9558 | ||
9559 | case FLOAT_EXPR: | |
4a00c761 JJ |
9560 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
9561 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
9562 | break; |
9563 | ||
9564 | case FIX_TRUNC_EXPR: | |
9565 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
9566 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
9567 | computing the operation. */ | |
9568 | return false; | |
9569 | ||
9570 | default: | |
9571 | gcc_unreachable (); | |
9572 | } | |
9573 | ||
6ae6116f | 9574 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 9575 | std::swap (c1, c2); |
4a00c761 | 9576 | |
ebfd146a IR |
9577 | if (code == FIX_TRUNC_EXPR) |
9578 | { | |
9579 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
9580 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
9581 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
9582 | } |
9583 | else | |
9584 | { | |
9585 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
9586 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
9587 | } | |
9588 | ||
9589 | if (!optab1 || !optab2) | |
9590 | return false; | |
9591 | ||
9592 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
9593 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
9594 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
9595 | return false; |
9596 | ||
4a00c761 JJ |
9597 | *code1 = c1; |
9598 | *code2 = c2; | |
9599 | ||
9600 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
9601 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
9602 | /* For scalar masks we may have different boolean |
9603 | vector types having the same QImode. Thus we | |
9604 | add additional check for elements number. */ | |
9605 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
9606 | || (TYPE_VECTOR_SUBPARTS (vectype) / 2 | |
9607 | == TYPE_VECTOR_SUBPARTS (wide_vectype))); | |
4a00c761 | 9608 | |
b8698a0f | 9609 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 9610 | types. */ |
ebfd146a | 9611 | |
4a00c761 JJ |
9612 | prev_type = vectype; |
9613 | prev_mode = vec_mode; | |
b8698a0f | 9614 | |
4a00c761 JJ |
9615 | if (!CONVERT_EXPR_CODE_P (code)) |
9616 | return false; | |
b8698a0f | 9617 | |
4a00c761 JJ |
9618 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
9619 | intermediate steps in promotion sequence. We try | |
9620 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
9621 | not. */ | |
9771b263 | 9622 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
9623 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
9624 | { | |
9625 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
9626 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
9627 | { | |
9628 | intermediate_type | |
9629 | = build_truth_vector_type (TYPE_VECTOR_SUBPARTS (prev_type) / 2, | |
9630 | current_vector_size); | |
9631 | if (intermediate_mode != TYPE_MODE (intermediate_type)) | |
9632 | return false; | |
9633 | } | |
9634 | else | |
9635 | intermediate_type | |
9636 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
9637 | TYPE_UNSIGNED (prev_type)); | |
9638 | ||
4a00c761 JJ |
9639 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); |
9640 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
9641 | ||
9642 | if (!optab3 || !optab4 | |
9643 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
9644 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
9645 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
9646 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
9647 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
9648 | == CODE_FOR_nothing) | |
9649 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
9650 | == CODE_FOR_nothing)) | |
9651 | break; | |
ebfd146a | 9652 | |
9771b263 | 9653 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
9654 | (*multi_step_cvt)++; |
9655 | ||
9656 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
9657 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
9658 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
9659 | || (TYPE_VECTOR_SUBPARTS (intermediate_type) / 2 | |
9660 | == TYPE_VECTOR_SUBPARTS (wide_vectype))); | |
4a00c761 JJ |
9661 | |
9662 | prev_type = intermediate_type; | |
9663 | prev_mode = intermediate_mode; | |
ebfd146a IR |
9664 | } |
9665 | ||
9771b263 | 9666 | interm_types->release (); |
4a00c761 | 9667 | return false; |
ebfd146a IR |
9668 | } |
9669 | ||
9670 | ||
9671 | /* Function supportable_narrowing_operation | |
9672 | ||
b8698a0f L |
9673 | Check whether an operation represented by the code CODE is a |
9674 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
9675 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
9676 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 9677 | |
ebfd146a | 9678 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 9679 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
9680 | the target platform directly via vector tree-codes. |
9681 | ||
9682 | Output: | |
b8698a0f L |
9683 | - CODE1 is the code of a vector operation to be used when |
9684 | vectorizing the operation, if available. | |
ebfd146a IR |
9685 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
9686 | case of multi-step conversion (like int->short->char - in that case | |
9687 | MULTI_STEP_CVT will be 1). | |
9688 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 9689 | narrowing operation (short in the above example). */ |
ebfd146a IR |
9690 | |
9691 | bool | |
9692 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 9693 | tree vectype_out, tree vectype_in, |
ebfd146a | 9694 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 9695 | vec<tree> *interm_types) |
ebfd146a | 9696 | { |
ef4bddc2 | 9697 | machine_mode vec_mode; |
ebfd146a IR |
9698 | enum insn_code icode1; |
9699 | optab optab1, interm_optab; | |
b690cc0f RG |
9700 | tree vectype = vectype_in; |
9701 | tree narrow_vectype = vectype_out; | |
ebfd146a | 9702 | enum tree_code c1; |
3ae0661a | 9703 | tree intermediate_type, prev_type; |
ef4bddc2 | 9704 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 9705 | int i; |
4a00c761 | 9706 | bool uns; |
ebfd146a | 9707 | |
4a00c761 | 9708 | *multi_step_cvt = 0; |
ebfd146a IR |
9709 | switch (code) |
9710 | { | |
9711 | CASE_CONVERT: | |
9712 | c1 = VEC_PACK_TRUNC_EXPR; | |
9713 | break; | |
9714 | ||
9715 | case FIX_TRUNC_EXPR: | |
9716 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
9717 | break; | |
9718 | ||
9719 | case FLOAT_EXPR: | |
9720 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
9721 | tree code and optabs used for computing the operation. */ | |
9722 | return false; | |
9723 | ||
9724 | default: | |
9725 | gcc_unreachable (); | |
9726 | } | |
9727 | ||
9728 | if (code == FIX_TRUNC_EXPR) | |
9729 | /* The signedness is determined from output operand. */ | |
b690cc0f | 9730 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
9731 | else |
9732 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
9733 | ||
9734 | if (!optab1) | |
9735 | return false; | |
9736 | ||
9737 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 9738 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
9739 | return false; |
9740 | ||
4a00c761 JJ |
9741 | *code1 = c1; |
9742 | ||
9743 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
9744 | /* For scalar masks we may have different boolean |
9745 | vector types having the same QImode. Thus we | |
9746 | add additional check for elements number. */ | |
9747 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
9748 | || (TYPE_VECTOR_SUBPARTS (vectype) * 2 | |
9749 | == TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 | 9750 | |
ebfd146a IR |
9751 | /* Check if it's a multi-step conversion that can be done using intermediate |
9752 | types. */ | |
4a00c761 | 9753 | prev_mode = vec_mode; |
3ae0661a | 9754 | prev_type = vectype; |
4a00c761 JJ |
9755 | if (code == FIX_TRUNC_EXPR) |
9756 | uns = TYPE_UNSIGNED (vectype_out); | |
9757 | else | |
9758 | uns = TYPE_UNSIGNED (vectype); | |
9759 | ||
9760 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
9761 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
9762 | costly than signed. */ | |
9763 | if (code == FIX_TRUNC_EXPR && uns) | |
9764 | { | |
9765 | enum insn_code icode2; | |
9766 | ||
9767 | intermediate_type | |
9768 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
9769 | interm_optab | |
9770 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 9771 | if (interm_optab != unknown_optab |
4a00c761 JJ |
9772 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
9773 | && insn_data[icode1].operand[0].mode | |
9774 | == insn_data[icode2].operand[0].mode) | |
9775 | { | |
9776 | uns = false; | |
9777 | optab1 = interm_optab; | |
9778 | icode1 = icode2; | |
9779 | } | |
9780 | } | |
ebfd146a | 9781 | |
4a00c761 JJ |
9782 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
9783 | intermediate steps in promotion sequence. We try | |
9784 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 9785 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
9786 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
9787 | { | |
9788 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
9789 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
9790 | { | |
9791 | intermediate_type | |
9792 | = build_truth_vector_type (TYPE_VECTOR_SUBPARTS (prev_type) * 2, | |
9793 | current_vector_size); | |
9794 | if (intermediate_mode != TYPE_MODE (intermediate_type)) | |
9795 | return false; | |
9796 | } | |
9797 | else | |
9798 | intermediate_type | |
9799 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
4a00c761 JJ |
9800 | interm_optab |
9801 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
9802 | optab_default); | |
9803 | if (!interm_optab | |
9804 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
9805 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
9806 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
9807 | == CODE_FOR_nothing)) | |
9808 | break; | |
9809 | ||
9771b263 | 9810 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
9811 | (*multi_step_cvt)++; |
9812 | ||
9813 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
9814 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
9815 | || (TYPE_VECTOR_SUBPARTS (intermediate_type) * 2 | |
9816 | == TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 JJ |
9817 | |
9818 | prev_mode = intermediate_mode; | |
3ae0661a | 9819 | prev_type = intermediate_type; |
4a00c761 | 9820 | optab1 = interm_optab; |
ebfd146a IR |
9821 | } |
9822 | ||
9771b263 | 9823 | interm_types->release (); |
4a00c761 | 9824 | return false; |
ebfd146a | 9825 | } |