]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
818ab71a | 2 | Copyright (C) 2003-2016 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
c7131fb2 | 25 | #include "backend.h" |
957060b5 AM |
26 | #include "target.h" |
27 | #include "rtl.h" | |
ebfd146a | 28 | #include "tree.h" |
c7131fb2 | 29 | #include "gimple.h" |
c7131fb2 | 30 | #include "ssa.h" |
957060b5 AM |
31 | #include "optabs-tree.h" |
32 | #include "insn-config.h" | |
33 | #include "recog.h" /* FIXME: for insn_data */ | |
34 | #include "cgraph.h" | |
957060b5 | 35 | #include "dumpfile.h" |
c7131fb2 | 36 | #include "alias.h" |
40e23961 | 37 | #include "fold-const.h" |
d8a2d370 | 38 | #include "stor-layout.h" |
2fb9a547 | 39 | #include "tree-eh.h" |
45b0be94 | 40 | #include "gimplify.h" |
5be5c238 | 41 | #include "gimple-iterator.h" |
18f429e2 | 42 | #include "gimplify-me.h" |
442b4905 | 43 | #include "tree-cfg.h" |
e28030cf | 44 | #include "tree-ssa-loop-manip.h" |
ebfd146a | 45 | #include "cfgloop.h" |
0136f8f0 AH |
46 | #include "tree-ssa-loop.h" |
47 | #include "tree-scalar-evolution.h" | |
ebfd146a | 48 | #include "tree-vectorizer.h" |
9b2b7279 | 49 | #include "builtins.h" |
70439f0d | 50 | #include "internal-fn.h" |
ebfd146a | 51 | |
7ee2468b SB |
52 | /* For lang_hooks.types.type_for_mode. */ |
53 | #include "langhooks.h" | |
ebfd146a | 54 | |
c3e7ee41 BS |
55 | /* Return the vectorized type for the given statement. */ |
56 | ||
57 | tree | |
58 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
59 | { | |
60 | return STMT_VINFO_VECTYPE (stmt_info); | |
61 | } | |
62 | ||
63 | /* Return TRUE iff the given statement is in an inner loop relative to | |
64 | the loop being vectorized. */ | |
65 | bool | |
66 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
67 | { | |
355fe088 | 68 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
c3e7ee41 BS |
69 | basic_block bb = gimple_bb (stmt); |
70 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
71 | struct loop* loop; | |
72 | ||
73 | if (!loop_vinfo) | |
74 | return false; | |
75 | ||
76 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
77 | ||
78 | return (bb->loop_father == loop->inner); | |
79 | } | |
80 | ||
81 | /* Record the cost of a statement, either by directly informing the | |
82 | target model or by saving it in a vector for later processing. | |
83 | Return a preliminary estimate of the statement's cost. */ | |
84 | ||
85 | unsigned | |
92345349 | 86 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 87 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 88 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 89 | { |
92345349 | 90 | if (body_cost_vec) |
c3e7ee41 | 91 | { |
92345349 | 92 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
ddf56386 RB |
93 | stmt_info_for_cost si = { count, kind, |
94 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
95 | misalign }; | |
96 | body_cost_vec->safe_push (si); | |
c3e7ee41 | 97 | return (unsigned) |
92345349 | 98 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
99 | } |
100 | else | |
310213d4 RB |
101 | return add_stmt_cost (stmt_info->vinfo->target_cost_data, |
102 | count, kind, stmt_info, misalign, where); | |
c3e7ee41 BS |
103 | } |
104 | ||
272c6793 RS |
105 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
106 | ||
107 | static tree | |
108 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
109 | { | |
110 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
111 | "vect_array"); | |
112 | } | |
113 | ||
114 | /* ARRAY is an array of vectors created by create_vector_array. | |
115 | Return an SSA_NAME for the vector in index N. The reference | |
116 | is part of the vectorization of STMT and the vector is associated | |
117 | with scalar destination SCALAR_DEST. */ | |
118 | ||
119 | static tree | |
355fe088 | 120 | read_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree scalar_dest, |
272c6793 RS |
121 | tree array, unsigned HOST_WIDE_INT n) |
122 | { | |
123 | tree vect_type, vect, vect_name, array_ref; | |
355fe088 | 124 | gimple *new_stmt; |
272c6793 RS |
125 | |
126 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
127 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
128 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
129 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
130 | build_int_cst (size_type_node, n), | |
131 | NULL_TREE, NULL_TREE); | |
132 | ||
133 | new_stmt = gimple_build_assign (vect, array_ref); | |
134 | vect_name = make_ssa_name (vect, new_stmt); | |
135 | gimple_assign_set_lhs (new_stmt, vect_name); | |
136 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
137 | |
138 | return vect_name; | |
139 | } | |
140 | ||
141 | /* ARRAY is an array of vectors created by create_vector_array. | |
142 | Emit code to store SSA_NAME VECT in index N of the array. | |
143 | The store is part of the vectorization of STMT. */ | |
144 | ||
145 | static void | |
355fe088 | 146 | write_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree vect, |
272c6793 RS |
147 | tree array, unsigned HOST_WIDE_INT n) |
148 | { | |
149 | tree array_ref; | |
355fe088 | 150 | gimple *new_stmt; |
272c6793 RS |
151 | |
152 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
153 | build_int_cst (size_type_node, n), | |
154 | NULL_TREE, NULL_TREE); | |
155 | ||
156 | new_stmt = gimple_build_assign (array_ref, vect); | |
157 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
158 | } |
159 | ||
160 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
161 | of *PTR. The memory reference replaces those in FIRST_DR | |
162 | (and its group). */ | |
163 | ||
164 | static tree | |
165 | create_array_ref (tree type, tree ptr, struct data_reference *first_dr) | |
166 | { | |
272c6793 RS |
167 | tree mem_ref, alias_ptr_type; |
168 | ||
169 | alias_ptr_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
170 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); | |
171 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 172 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
173 | return mem_ref; |
174 | } | |
175 | ||
ebfd146a IR |
176 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
177 | ||
178 | /* Function vect_mark_relevant. | |
179 | ||
180 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
181 | ||
182 | static void | |
355fe088 | 183 | vect_mark_relevant (vec<gimple *> *worklist, gimple *stmt, |
97ecdb46 | 184 | enum vect_relevant relevant, bool live_p) |
ebfd146a IR |
185 | { |
186 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
187 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
188 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
355fe088 | 189 | gimple *pattern_stmt; |
ebfd146a | 190 | |
73fbfcad | 191 | if (dump_enabled_p ()) |
66c16fd9 RB |
192 | { |
193 | dump_printf_loc (MSG_NOTE, vect_location, | |
194 | "mark relevant %d, live %d: ", relevant, live_p); | |
195 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
196 | } | |
ebfd146a | 197 | |
83197f37 IR |
198 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
199 | related pattern stmt instead of the original stmt. However, such stmts | |
200 | may have their own uses that are not in any pattern, in such cases the | |
201 | stmt itself should be marked. */ | |
ebfd146a IR |
202 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
203 | { | |
97ecdb46 JJ |
204 | /* This is the last stmt in a sequence that was detected as a |
205 | pattern that can potentially be vectorized. Don't mark the stmt | |
206 | as relevant/live because it's not going to be vectorized. | |
207 | Instead mark the pattern-stmt that replaces it. */ | |
83197f37 | 208 | |
97ecdb46 JJ |
209 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); |
210 | ||
211 | if (dump_enabled_p ()) | |
212 | dump_printf_loc (MSG_NOTE, vect_location, | |
213 | "last stmt in pattern. don't mark" | |
214 | " relevant/live.\n"); | |
215 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
216 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
217 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
218 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
219 | stmt = pattern_stmt; | |
ebfd146a IR |
220 | } |
221 | ||
222 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
223 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
224 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
225 | ||
226 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
227 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
228 | { | |
73fbfcad | 229 | if (dump_enabled_p ()) |
78c60e3d | 230 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 231 | "already marked relevant/live.\n"); |
ebfd146a IR |
232 | return; |
233 | } | |
234 | ||
9771b263 | 235 | worklist->safe_push (stmt); |
ebfd146a IR |
236 | } |
237 | ||
238 | ||
239 | /* Function vect_stmt_relevant_p. | |
240 | ||
241 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
242 | "relevant for vectorization". | |
243 | ||
244 | A stmt is considered "relevant for vectorization" if: | |
245 | - it has uses outside the loop. | |
246 | - it has vdefs (it alters memory). | |
247 | - control stmts in the loop (except for the exit condition). | |
248 | ||
249 | CHECKME: what other side effects would the vectorizer allow? */ | |
250 | ||
251 | static bool | |
355fe088 | 252 | vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo, |
ebfd146a IR |
253 | enum vect_relevant *relevant, bool *live_p) |
254 | { | |
255 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
256 | ssa_op_iter op_iter; | |
257 | imm_use_iterator imm_iter; | |
258 | use_operand_p use_p; | |
259 | def_operand_p def_p; | |
260 | ||
8644a673 | 261 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
262 | *live_p = false; |
263 | ||
264 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
265 | if (is_ctrl_stmt (stmt) |
266 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
267 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 268 | *relevant = vect_used_in_scope; |
ebfd146a IR |
269 | |
270 | /* changing memory. */ | |
271 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
272 | if (gimple_vdef (stmt) |
273 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 274 | { |
73fbfcad | 275 | if (dump_enabled_p ()) |
78c60e3d | 276 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 277 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 278 | *relevant = vect_used_in_scope; |
ebfd146a IR |
279 | } |
280 | ||
281 | /* uses outside the loop. */ | |
282 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
283 | { | |
284 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
285 | { | |
286 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
287 | if (!flow_bb_inside_loop_p (loop, bb)) | |
288 | { | |
73fbfcad | 289 | if (dump_enabled_p ()) |
78c60e3d | 290 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 291 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 292 | |
3157b0c2 AO |
293 | if (is_gimple_debug (USE_STMT (use_p))) |
294 | continue; | |
295 | ||
ebfd146a IR |
296 | /* We expect all such uses to be in the loop exit phis |
297 | (because of loop closed form) */ | |
298 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
299 | gcc_assert (bb == single_exit (loop)->dest); | |
300 | ||
301 | *live_p = true; | |
302 | } | |
303 | } | |
304 | } | |
305 | ||
306 | return (*live_p || *relevant); | |
307 | } | |
308 | ||
309 | ||
b8698a0f | 310 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 311 | |
ff802fa1 | 312 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
313 | used in STMT for anything other than indexing an array. */ |
314 | ||
315 | static bool | |
355fe088 | 316 | exist_non_indexing_operands_for_use_p (tree use, gimple *stmt) |
ebfd146a IR |
317 | { |
318 | tree operand; | |
319 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 320 | |
ff802fa1 | 321 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
322 | reference in STMT, then any operand that corresponds to USE |
323 | is not indexing an array. */ | |
324 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
325 | return true; | |
59a05b0c | 326 | |
ebfd146a IR |
327 | /* STMT has a data_ref. FORNOW this means that its of one of |
328 | the following forms: | |
329 | -1- ARRAY_REF = var | |
330 | -2- var = ARRAY_REF | |
331 | (This should have been verified in analyze_data_refs). | |
332 | ||
333 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 334 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
335 | for array indexing. |
336 | ||
337 | Therefore, all we need to check is if STMT falls into the | |
338 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
339 | |
340 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
341 | { |
342 | if (is_gimple_call (stmt) | |
343 | && gimple_call_internal_p (stmt)) | |
344 | switch (gimple_call_internal_fn (stmt)) | |
345 | { | |
346 | case IFN_MASK_STORE: | |
347 | operand = gimple_call_arg (stmt, 3); | |
348 | if (operand == use) | |
349 | return true; | |
350 | /* FALLTHRU */ | |
351 | case IFN_MASK_LOAD: | |
352 | operand = gimple_call_arg (stmt, 2); | |
353 | if (operand == use) | |
354 | return true; | |
355 | break; | |
356 | default: | |
357 | break; | |
358 | } | |
359 | return false; | |
360 | } | |
361 | ||
59a05b0c EB |
362 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
363 | return false; | |
ebfd146a | 364 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
365 | if (TREE_CODE (operand) != SSA_NAME) |
366 | return false; | |
367 | ||
368 | if (operand == use) | |
369 | return true; | |
370 | ||
371 | return false; | |
372 | } | |
373 | ||
374 | ||
b8698a0f | 375 | /* |
ebfd146a IR |
376 | Function process_use. |
377 | ||
378 | Inputs: | |
379 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 380 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ff802fa1 | 381 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 382 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
383 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
384 | be performed. | |
ebfd146a IR |
385 | |
386 | Outputs: | |
387 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
388 | relevance info of the DEF_STMT of this USE: | |
389 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
390 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
391 | Exceptions: | |
392 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 393 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 394 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
395 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
396 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
397 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
398 | be modified accordingly. | |
399 | ||
400 | Return true if everything is as expected. Return false otherwise. */ | |
401 | ||
402 | static bool | |
355fe088 TS |
403 | process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
404 | enum vect_relevant relevant, vec<gimple *> *worklist, | |
aec7ae7d | 405 | bool force) |
ebfd146a IR |
406 | { |
407 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
408 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
409 | stmt_vec_info dstmt_vinfo; | |
410 | basic_block bb, def_bb; | |
355fe088 | 411 | gimple *def_stmt; |
ebfd146a IR |
412 | enum vect_def_type dt; |
413 | ||
b8698a0f | 414 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 415 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 416 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
417 | return true; |
418 | ||
81c40241 | 419 | if (!vect_is_simple_use (use, loop_vinfo, &def_stmt, &dt)) |
b8698a0f | 420 | { |
73fbfcad | 421 | if (dump_enabled_p ()) |
78c60e3d | 422 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 423 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
424 | return false; |
425 | } | |
426 | ||
427 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
428 | return true; | |
429 | ||
430 | def_bb = gimple_bb (def_stmt); | |
431 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
432 | { | |
73fbfcad | 433 | if (dump_enabled_p ()) |
e645e942 | 434 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
435 | return true; |
436 | } | |
437 | ||
b8698a0f L |
438 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
439 | DEF_STMT must have already been processed, because this should be the | |
440 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
441 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
442 | check that everything is as expected, and we are done. */ |
443 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
444 | bb = gimple_bb (stmt); | |
445 | if (gimple_code (stmt) == GIMPLE_PHI | |
446 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
447 | && gimple_code (def_stmt) != GIMPLE_PHI | |
448 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
449 | && bb->loop_father == def_bb->loop_father) | |
450 | { | |
73fbfcad | 451 | if (dump_enabled_p ()) |
78c60e3d | 452 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 453 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
454 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
455 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
456 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 457 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 458 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
459 | return true; |
460 | } | |
461 | ||
462 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
463 | outer-loop-header-bb: | |
464 | d = def_stmt | |
465 | inner-loop: | |
466 | stmt # use (d) | |
467 | outer-loop-tail-bb: | |
468 | ... */ | |
469 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
470 | { | |
73fbfcad | 471 | if (dump_enabled_p ()) |
78c60e3d | 472 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 473 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 474 | |
ebfd146a IR |
475 | switch (relevant) |
476 | { | |
8644a673 | 477 | case vect_unused_in_scope: |
7c5222ff IR |
478 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
479 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 480 | break; |
7c5222ff | 481 | |
ebfd146a | 482 | case vect_used_in_outer_by_reduction: |
7c5222ff | 483 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
484 | relevant = vect_used_by_reduction; |
485 | break; | |
7c5222ff | 486 | |
ebfd146a | 487 | case vect_used_in_outer: |
7c5222ff | 488 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 489 | relevant = vect_used_in_scope; |
ebfd146a | 490 | break; |
7c5222ff | 491 | |
8644a673 | 492 | case vect_used_in_scope: |
ebfd146a IR |
493 | break; |
494 | ||
495 | default: | |
496 | gcc_unreachable (); | |
b8698a0f | 497 | } |
ebfd146a IR |
498 | } |
499 | ||
500 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
501 | outer-loop-header-bb: | |
502 | ... | |
503 | inner-loop: | |
504 | d = def_stmt | |
06066f92 | 505 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
506 | stmt # use (d) */ |
507 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
508 | { | |
73fbfcad | 509 | if (dump_enabled_p ()) |
78c60e3d | 510 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 511 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 512 | |
ebfd146a IR |
513 | switch (relevant) |
514 | { | |
8644a673 | 515 | case vect_unused_in_scope: |
b8698a0f | 516 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 517 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 518 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
519 | break; |
520 | ||
ebfd146a IR |
521 | case vect_used_by_reduction: |
522 | relevant = vect_used_in_outer_by_reduction; | |
523 | break; | |
524 | ||
8644a673 | 525 | case vect_used_in_scope: |
ebfd146a IR |
526 | relevant = vect_used_in_outer; |
527 | break; | |
528 | ||
529 | default: | |
530 | gcc_unreachable (); | |
531 | } | |
532 | } | |
533 | ||
97ecdb46 | 534 | vect_mark_relevant (worklist, def_stmt, relevant, live_p); |
ebfd146a IR |
535 | return true; |
536 | } | |
537 | ||
538 | ||
539 | /* Function vect_mark_stmts_to_be_vectorized. | |
540 | ||
541 | Not all stmts in the loop need to be vectorized. For example: | |
542 | ||
543 | for i... | |
544 | for j... | |
545 | 1. T0 = i + j | |
546 | 2. T1 = a[T0] | |
547 | ||
548 | 3. j = j + 1 | |
549 | ||
550 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
551 | addressing of vectorized data-refs are handled differently. | |
552 | ||
553 | This pass detects such stmts. */ | |
554 | ||
555 | bool | |
556 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
557 | { | |
ebfd146a IR |
558 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
559 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
560 | unsigned int nbbs = loop->num_nodes; | |
561 | gimple_stmt_iterator si; | |
355fe088 | 562 | gimple *stmt; |
ebfd146a IR |
563 | unsigned int i; |
564 | stmt_vec_info stmt_vinfo; | |
565 | basic_block bb; | |
355fe088 | 566 | gimple *phi; |
ebfd146a | 567 | bool live_p; |
06066f92 IR |
568 | enum vect_relevant relevant, tmp_relevant; |
569 | enum vect_def_type def_type; | |
ebfd146a | 570 | |
73fbfcad | 571 | if (dump_enabled_p ()) |
78c60e3d | 572 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 573 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 574 | |
355fe088 | 575 | auto_vec<gimple *, 64> worklist; |
ebfd146a IR |
576 | |
577 | /* 1. Init worklist. */ | |
578 | for (i = 0; i < nbbs; i++) | |
579 | { | |
580 | bb = bbs[i]; | |
581 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 582 | { |
ebfd146a | 583 | phi = gsi_stmt (si); |
73fbfcad | 584 | if (dump_enabled_p ()) |
ebfd146a | 585 | { |
78c60e3d SS |
586 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
587 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
588 | } |
589 | ||
590 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 591 | vect_mark_relevant (&worklist, phi, relevant, live_p); |
ebfd146a IR |
592 | } |
593 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
594 | { | |
595 | stmt = gsi_stmt (si); | |
73fbfcad | 596 | if (dump_enabled_p ()) |
ebfd146a | 597 | { |
78c60e3d SS |
598 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
599 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 600 | } |
ebfd146a IR |
601 | |
602 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 603 | vect_mark_relevant (&worklist, stmt, relevant, live_p); |
ebfd146a IR |
604 | } |
605 | } | |
606 | ||
607 | /* 2. Process_worklist */ | |
9771b263 | 608 | while (worklist.length () > 0) |
ebfd146a IR |
609 | { |
610 | use_operand_p use_p; | |
611 | ssa_op_iter iter; | |
612 | ||
9771b263 | 613 | stmt = worklist.pop (); |
73fbfcad | 614 | if (dump_enabled_p ()) |
ebfd146a | 615 | { |
78c60e3d SS |
616 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
617 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
618 | } |
619 | ||
b8698a0f L |
620 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
621 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
622 | liveness and relevance properties of STMT. */ |
623 | stmt_vinfo = vinfo_for_stmt (stmt); | |
624 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
625 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
626 | ||
627 | /* Generally, the liveness and relevance properties of STMT are | |
628 | propagated as is to the DEF_STMTs of its USEs: | |
629 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
630 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
631 | ||
632 | One exception is when STMT has been identified as defining a reduction | |
633 | variable; in this case we set the liveness/relevance as follows: | |
634 | live_p = false | |
635 | relevant = vect_used_by_reduction | |
636 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f | 637 | those that are used by a reduction computation, and those that are |
ff802fa1 | 638 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 639 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 640 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 641 | |
06066f92 IR |
642 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
643 | tmp_relevant = relevant; | |
644 | switch (def_type) | |
ebfd146a | 645 | { |
06066f92 IR |
646 | case vect_reduction_def: |
647 | switch (tmp_relevant) | |
648 | { | |
649 | case vect_unused_in_scope: | |
650 | relevant = vect_used_by_reduction; | |
651 | break; | |
652 | ||
653 | case vect_used_by_reduction: | |
654 | if (gimple_code (stmt) == GIMPLE_PHI) | |
655 | break; | |
656 | /* fall through */ | |
657 | ||
658 | default: | |
73fbfcad | 659 | if (dump_enabled_p ()) |
78c60e3d | 660 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 661 | "unsupported use of reduction.\n"); |
06066f92 IR |
662 | return false; |
663 | } | |
664 | ||
b8698a0f | 665 | live_p = false; |
06066f92 | 666 | break; |
b8698a0f | 667 | |
06066f92 IR |
668 | case vect_nested_cycle: |
669 | if (tmp_relevant != vect_unused_in_scope | |
670 | && tmp_relevant != vect_used_in_outer_by_reduction | |
671 | && tmp_relevant != vect_used_in_outer) | |
672 | { | |
73fbfcad | 673 | if (dump_enabled_p ()) |
78c60e3d | 674 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 675 | "unsupported use of nested cycle.\n"); |
7c5222ff | 676 | |
06066f92 IR |
677 | return false; |
678 | } | |
7c5222ff | 679 | |
b8698a0f L |
680 | live_p = false; |
681 | break; | |
682 | ||
06066f92 IR |
683 | case vect_double_reduction_def: |
684 | if (tmp_relevant != vect_unused_in_scope | |
685 | && tmp_relevant != vect_used_by_reduction) | |
686 | { | |
73fbfcad | 687 | if (dump_enabled_p ()) |
78c60e3d | 688 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 689 | "unsupported use of double reduction.\n"); |
7c5222ff | 690 | |
7c5222ff | 691 | return false; |
06066f92 IR |
692 | } |
693 | ||
694 | live_p = false; | |
b8698a0f | 695 | break; |
7c5222ff | 696 | |
06066f92 IR |
697 | default: |
698 | break; | |
7c5222ff | 699 | } |
b8698a0f | 700 | |
aec7ae7d | 701 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
702 | { |
703 | /* Pattern statements are not inserted into the code, so | |
704 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
705 | have to scan the RHS or function arguments instead. */ | |
706 | if (is_gimple_assign (stmt)) | |
707 | { | |
69d2aade JJ |
708 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
709 | tree op = gimple_assign_rhs1 (stmt); | |
710 | ||
711 | i = 1; | |
712 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
713 | { | |
714 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
aec7ae7d | 715 | live_p, relevant, &worklist, false) |
69d2aade | 716 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
aec7ae7d | 717 | live_p, relevant, &worklist, false)) |
566d377a | 718 | return false; |
69d2aade JJ |
719 | i = 2; |
720 | } | |
721 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 722 | { |
69d2aade | 723 | op = gimple_op (stmt, i); |
afbe6325 RB |
724 | if (TREE_CODE (op) == SSA_NAME |
725 | && !process_use (stmt, op, loop_vinfo, live_p, relevant, | |
726 | &worklist, false)) | |
07687835 | 727 | return false; |
9d5e7640 IR |
728 | } |
729 | } | |
730 | else if (is_gimple_call (stmt)) | |
731 | { | |
732 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
733 | { | |
734 | tree arg = gimple_call_arg (stmt, i); | |
735 | if (!process_use (stmt, arg, loop_vinfo, live_p, relevant, | |
aec7ae7d | 736 | &worklist, false)) |
07687835 | 737 | return false; |
9d5e7640 IR |
738 | } |
739 | } | |
740 | } | |
741 | else | |
742 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
743 | { | |
744 | tree op = USE_FROM_PTR (use_p); | |
745 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, | |
aec7ae7d | 746 | &worklist, false)) |
07687835 | 747 | return false; |
9d5e7640 | 748 | } |
aec7ae7d | 749 | |
3bab6342 | 750 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d JJ |
751 | { |
752 | tree off; | |
3bab6342 | 753 | tree decl = vect_check_gather_scatter (stmt, loop_vinfo, NULL, &off, NULL); |
aec7ae7d JJ |
754 | gcc_assert (decl); |
755 | if (!process_use (stmt, off, loop_vinfo, live_p, relevant, | |
756 | &worklist, true)) | |
566d377a | 757 | return false; |
aec7ae7d | 758 | } |
ebfd146a IR |
759 | } /* while worklist */ |
760 | ||
ebfd146a IR |
761 | return true; |
762 | } | |
763 | ||
764 | ||
b8698a0f | 765 | /* Function vect_model_simple_cost. |
ebfd146a | 766 | |
b8698a0f | 767 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
768 | single op. Right now, this does not account for multiple insns that could |
769 | be generated for the single vector op. We will handle that shortly. */ | |
770 | ||
771 | void | |
b8698a0f | 772 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 BS |
773 | enum vect_def_type *dt, |
774 | stmt_vector_for_cost *prologue_cost_vec, | |
775 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
776 | { |
777 | int i; | |
92345349 | 778 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
779 | |
780 | /* The SLP costs were already calculated during SLP tree build. */ | |
781 | if (PURE_SLP_STMT (stmt_info)) | |
782 | return; | |
783 | ||
ebfd146a IR |
784 | /* FORNOW: Assuming maximum 2 args per stmts. */ |
785 | for (i = 0; i < 2; i++) | |
92345349 BS |
786 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
787 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, vector_stmt, | |
788 | stmt_info, 0, vect_prologue); | |
c3e7ee41 BS |
789 | |
790 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
791 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
792 | stmt_info, 0, vect_body); | |
c3e7ee41 | 793 | |
73fbfcad | 794 | if (dump_enabled_p ()) |
78c60e3d SS |
795 | dump_printf_loc (MSG_NOTE, vect_location, |
796 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 797 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
798 | } |
799 | ||
800 | ||
8bd37302 BS |
801 | /* Model cost for type demotion and promotion operations. PWR is normally |
802 | zero for single-step promotions and demotions. It will be one if | |
803 | two-step promotion/demotion is required, and so on. Each additional | |
804 | step doubles the number of instructions required. */ | |
805 | ||
806 | static void | |
807 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
808 | enum vect_def_type *dt, int pwr) | |
809 | { | |
810 | int i, tmp; | |
92345349 | 811 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
812 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
813 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
814 | void *target_cost_data; | |
8bd37302 BS |
815 | |
816 | /* The SLP costs were already calculated during SLP tree build. */ | |
817 | if (PURE_SLP_STMT (stmt_info)) | |
818 | return; | |
819 | ||
c3e7ee41 BS |
820 | if (loop_vinfo) |
821 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
822 | else | |
823 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
824 | ||
8bd37302 BS |
825 | for (i = 0; i < pwr + 1; i++) |
826 | { | |
827 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
828 | (i + 1) : i; | |
c3e7ee41 | 829 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
830 | vec_promote_demote, stmt_info, 0, |
831 | vect_body); | |
8bd37302 BS |
832 | } |
833 | ||
834 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
835 | for (i = 0; i < 2; i++) | |
92345349 BS |
836 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
837 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
838 | stmt_info, 0, vect_prologue); | |
8bd37302 | 839 | |
73fbfcad | 840 | if (dump_enabled_p ()) |
78c60e3d SS |
841 | dump_printf_loc (MSG_NOTE, vect_location, |
842 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 843 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
844 | } |
845 | ||
0d0293ac | 846 | /* Function vect_cost_group_size |
b8698a0f | 847 | |
0d0293ac | 848 | For grouped load or store, return the group_size only if it is the first |
ebfd146a IR |
849 | load or store of a group, else return 1. This ensures that group size is |
850 | only returned once per group. */ | |
851 | ||
852 | static int | |
0d0293ac | 853 | vect_cost_group_size (stmt_vec_info stmt_info) |
ebfd146a | 854 | { |
355fe088 | 855 | gimple *first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
856 | |
857 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
e14c1050 | 858 | return GROUP_SIZE (stmt_info); |
ebfd146a IR |
859 | |
860 | return 1; | |
861 | } | |
862 | ||
863 | ||
864 | /* Function vect_model_store_cost | |
865 | ||
0d0293ac MM |
866 | Models cost for stores. In the case of grouped accesses, one access |
867 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
868 | |
869 | void | |
b8698a0f | 870 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
272c6793 | 871 | bool store_lanes_p, enum vect_def_type dt, |
92345349 BS |
872 | slp_tree slp_node, |
873 | stmt_vector_for_cost *prologue_cost_vec, | |
874 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
875 | { |
876 | int group_size; | |
92345349 | 877 | unsigned int inside_cost = 0, prologue_cost = 0; |
720f5239 | 878 | struct data_reference *first_dr; |
355fe088 | 879 | gimple *first_stmt; |
ebfd146a | 880 | |
8644a673 | 881 | if (dt == vect_constant_def || dt == vect_external_def) |
92345349 BS |
882 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
883 | stmt_info, 0, vect_prologue); | |
ebfd146a | 884 | |
0d0293ac MM |
885 | /* Grouped access? */ |
886 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
720f5239 IR |
887 | { |
888 | if (slp_node) | |
889 | { | |
9771b263 | 890 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
720f5239 IR |
891 | group_size = 1; |
892 | } | |
893 | else | |
894 | { | |
e14c1050 | 895 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 896 | group_size = vect_cost_group_size (stmt_info); |
720f5239 IR |
897 | } |
898 | ||
899 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
900 | } | |
0d0293ac | 901 | /* Not a grouped access. */ |
ebfd146a | 902 | else |
720f5239 IR |
903 | { |
904 | group_size = 1; | |
905 | first_dr = STMT_VINFO_DATA_REF (stmt_info); | |
906 | } | |
ebfd146a | 907 | |
272c6793 | 908 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 909 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 RS |
910 | access is instead being provided by a permute-and-store operation, |
911 | include the cost of the permutes. */ | |
cee62fee MM |
912 | if (!store_lanes_p && group_size > 1 |
913 | && !STMT_VINFO_STRIDED_P (stmt_info)) | |
ebfd146a | 914 | { |
e1377713 ES |
915 | /* Uses a high and low interleave or shuffle operations for each |
916 | needed permute. */ | |
917 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; | |
92345349 BS |
918 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
919 | stmt_info, 0, vect_body); | |
ebfd146a | 920 | |
73fbfcad | 921 | if (dump_enabled_p ()) |
78c60e3d | 922 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 923 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 924 | group_size); |
ebfd146a IR |
925 | } |
926 | ||
cee62fee | 927 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 928 | /* Costs of the stores. */ |
cee62fee MM |
929 | if (STMT_VINFO_STRIDED_P (stmt_info) |
930 | && !STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
f2e2a985 MM |
931 | { |
932 | /* N scalar stores plus extracting the elements. */ | |
f2e2a985 MM |
933 | inside_cost += record_stmt_cost (body_cost_vec, |
934 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
935 | scalar_store, stmt_info, 0, vect_body); | |
f2e2a985 MM |
936 | } |
937 | else | |
938 | vect_get_store_cost (first_dr, ncopies, &inside_cost, body_cost_vec); | |
ebfd146a | 939 | |
cee62fee MM |
940 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
941 | inside_cost += record_stmt_cost (body_cost_vec, | |
942 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
943 | vec_to_scalar, stmt_info, 0, vect_body); | |
944 | ||
73fbfcad | 945 | if (dump_enabled_p ()) |
78c60e3d SS |
946 | dump_printf_loc (MSG_NOTE, vect_location, |
947 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 948 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
949 | } |
950 | ||
951 | ||
720f5239 IR |
952 | /* Calculate cost of DR's memory access. */ |
953 | void | |
954 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 955 | unsigned int *inside_cost, |
92345349 | 956 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
957 | { |
958 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 959 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 960 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
961 | |
962 | switch (alignment_support_scheme) | |
963 | { | |
964 | case dr_aligned: | |
965 | { | |
92345349 BS |
966 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
967 | vector_store, stmt_info, 0, | |
968 | vect_body); | |
720f5239 | 969 | |
73fbfcad | 970 | if (dump_enabled_p ()) |
78c60e3d | 971 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 972 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
973 | break; |
974 | } | |
975 | ||
976 | case dr_unaligned_supported: | |
977 | { | |
720f5239 | 978 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 979 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 980 | unaligned_store, stmt_info, |
92345349 | 981 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 982 | if (dump_enabled_p ()) |
78c60e3d SS |
983 | dump_printf_loc (MSG_NOTE, vect_location, |
984 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 985 | "hardware.\n"); |
720f5239 IR |
986 | break; |
987 | } | |
988 | ||
38eec4c6 UW |
989 | case dr_unaligned_unsupported: |
990 | { | |
991 | *inside_cost = VECT_MAX_COST; | |
992 | ||
73fbfcad | 993 | if (dump_enabled_p ()) |
78c60e3d | 994 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 995 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
996 | break; |
997 | } | |
998 | ||
720f5239 IR |
999 | default: |
1000 | gcc_unreachable (); | |
1001 | } | |
1002 | } | |
1003 | ||
1004 | ||
ebfd146a IR |
1005 | /* Function vect_model_load_cost |
1006 | ||
0d0293ac MM |
1007 | Models cost for loads. In the case of grouped accesses, the last access |
1008 | has the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1009 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1010 | access scheme chosen. */ |
1011 | ||
1012 | void | |
92345349 BS |
1013 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
1014 | bool load_lanes_p, slp_tree slp_node, | |
1015 | stmt_vector_for_cost *prologue_cost_vec, | |
1016 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
1017 | { |
1018 | int group_size; | |
355fe088 | 1019 | gimple *first_stmt; |
ebfd146a | 1020 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; |
92345349 | 1021 | unsigned int inside_cost = 0, prologue_cost = 0; |
ebfd146a | 1022 | |
0d0293ac | 1023 | /* Grouped accesses? */ |
e14c1050 | 1024 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 1025 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && first_stmt && !slp_node) |
ebfd146a | 1026 | { |
0d0293ac | 1027 | group_size = vect_cost_group_size (stmt_info); |
ebfd146a IR |
1028 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
1029 | } | |
0d0293ac | 1030 | /* Not a grouped access. */ |
ebfd146a IR |
1031 | else |
1032 | { | |
1033 | group_size = 1; | |
1034 | first_dr = dr; | |
1035 | } | |
1036 | ||
272c6793 | 1037 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1038 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 RS |
1039 | access is instead being provided by a load-and-permute operation, |
1040 | include the cost of the permutes. */ | |
7b5fc413 | 1041 | if (!load_lanes_p && group_size > 1 |
f2e2a985 | 1042 | && !STMT_VINFO_STRIDED_P (stmt_info)) |
ebfd146a | 1043 | { |
2c23db6d ES |
1044 | /* Uses an even and odd extract operations or shuffle operations |
1045 | for each needed permute. */ | |
1046 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; | |
1047 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, | |
1048 | stmt_info, 0, vect_body); | |
ebfd146a | 1049 | |
73fbfcad | 1050 | if (dump_enabled_p ()) |
e645e942 TJ |
1051 | dump_printf_loc (MSG_NOTE, vect_location, |
1052 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1053 | group_size); |
ebfd146a IR |
1054 | } |
1055 | ||
1056 | /* The loads themselves. */ | |
f2e2a985 | 1057 | if (STMT_VINFO_STRIDED_P (stmt_info) |
7b5fc413 | 1058 | && !STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
a82960aa | 1059 | { |
a21892ad BS |
1060 | /* N scalar loads plus gathering them into a vector. */ |
1061 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
92345349 | 1062 | inside_cost += record_stmt_cost (body_cost_vec, |
c3e7ee41 | 1063 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), |
92345349 | 1064 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1065 | } |
1066 | else | |
1067 | vect_get_load_cost (first_dr, ncopies, | |
1068 | ((!STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1069 | || group_size > 1 || slp_node), | |
92345349 BS |
1070 | &inside_cost, &prologue_cost, |
1071 | prologue_cost_vec, body_cost_vec, true); | |
f2e2a985 | 1072 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
7b5fc413 RB |
1073 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, |
1074 | stmt_info, 0, vect_body); | |
720f5239 | 1075 | |
73fbfcad | 1076 | if (dump_enabled_p ()) |
78c60e3d SS |
1077 | dump_printf_loc (MSG_NOTE, vect_location, |
1078 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1079 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1080 | } |
1081 | ||
1082 | ||
1083 | /* Calculate cost of DR's memory access. */ | |
1084 | void | |
1085 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1086 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1087 | unsigned int *prologue_cost, |
1088 | stmt_vector_for_cost *prologue_cost_vec, | |
1089 | stmt_vector_for_cost *body_cost_vec, | |
1090 | bool record_prologue_costs) | |
720f5239 IR |
1091 | { |
1092 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 1093 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 1094 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
1095 | |
1096 | switch (alignment_support_scheme) | |
ebfd146a IR |
1097 | { |
1098 | case dr_aligned: | |
1099 | { | |
92345349 BS |
1100 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1101 | stmt_info, 0, vect_body); | |
ebfd146a | 1102 | |
73fbfcad | 1103 | if (dump_enabled_p ()) |
78c60e3d | 1104 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1105 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1106 | |
1107 | break; | |
1108 | } | |
1109 | case dr_unaligned_supported: | |
1110 | { | |
720f5239 | 1111 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1112 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1113 | unaligned_load, stmt_info, |
92345349 | 1114 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1115 | |
73fbfcad | 1116 | if (dump_enabled_p ()) |
78c60e3d SS |
1117 | dump_printf_loc (MSG_NOTE, vect_location, |
1118 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1119 | "hardware.\n"); |
ebfd146a IR |
1120 | |
1121 | break; | |
1122 | } | |
1123 | case dr_explicit_realign: | |
1124 | { | |
92345349 BS |
1125 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1126 | vector_load, stmt_info, 0, vect_body); | |
1127 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1128 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1129 | |
1130 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1131 | the containing loop, the following cost should be added to the | |
92345349 | 1132 | prologue costs. */ |
ebfd146a | 1133 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1134 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1135 | stmt_info, 0, vect_body); | |
ebfd146a | 1136 | |
73fbfcad | 1137 | if (dump_enabled_p ()) |
e645e942 TJ |
1138 | dump_printf_loc (MSG_NOTE, vect_location, |
1139 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1140 | |
ebfd146a IR |
1141 | break; |
1142 | } | |
1143 | case dr_explicit_realign_optimized: | |
1144 | { | |
73fbfcad | 1145 | if (dump_enabled_p ()) |
e645e942 | 1146 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1147 | "vect_model_load_cost: unaligned software " |
e645e942 | 1148 | "pipelined.\n"); |
ebfd146a IR |
1149 | |
1150 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1151 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1152 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1153 | access, then the above cost should only be considered for one |
ff802fa1 | 1154 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1155 | and a realignment op. */ |
1156 | ||
92345349 | 1157 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1158 | { |
92345349 BS |
1159 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1160 | vector_stmt, stmt_info, | |
1161 | 0, vect_prologue); | |
ebfd146a | 1162 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1163 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1164 | vector_stmt, stmt_info, | |
1165 | 0, vect_prologue); | |
ebfd146a IR |
1166 | } |
1167 | ||
92345349 BS |
1168 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1169 | stmt_info, 0, vect_body); | |
1170 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1171 | stmt_info, 0, vect_body); | |
8bd37302 | 1172 | |
73fbfcad | 1173 | if (dump_enabled_p ()) |
78c60e3d | 1174 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1175 | "vect_model_load_cost: explicit realign optimized" |
1176 | "\n"); | |
8bd37302 | 1177 | |
ebfd146a IR |
1178 | break; |
1179 | } | |
1180 | ||
38eec4c6 UW |
1181 | case dr_unaligned_unsupported: |
1182 | { | |
1183 | *inside_cost = VECT_MAX_COST; | |
1184 | ||
73fbfcad | 1185 | if (dump_enabled_p ()) |
78c60e3d | 1186 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1187 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1188 | break; |
1189 | } | |
1190 | ||
ebfd146a IR |
1191 | default: |
1192 | gcc_unreachable (); | |
1193 | } | |
ebfd146a IR |
1194 | } |
1195 | ||
418b7df3 RG |
1196 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1197 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1198 | |
418b7df3 | 1199 | static void |
355fe088 | 1200 | vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi) |
ebfd146a | 1201 | { |
ebfd146a | 1202 | if (gsi) |
418b7df3 | 1203 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1204 | else |
1205 | { | |
418b7df3 | 1206 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1207 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1208 | |
a70d6342 IR |
1209 | if (loop_vinfo) |
1210 | { | |
1211 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1212 | basic_block new_bb; |
1213 | edge pe; | |
a70d6342 IR |
1214 | |
1215 | if (nested_in_vect_loop_p (loop, stmt)) | |
1216 | loop = loop->inner; | |
b8698a0f | 1217 | |
a70d6342 | 1218 | pe = loop_preheader_edge (loop); |
418b7df3 | 1219 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1220 | gcc_assert (!new_bb); |
1221 | } | |
1222 | else | |
1223 | { | |
1224 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1225 | basic_block bb; | |
1226 | gimple_stmt_iterator gsi_bb_start; | |
1227 | ||
1228 | gcc_assert (bb_vinfo); | |
1229 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1230 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1231 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1232 | } |
ebfd146a IR |
1233 | } |
1234 | ||
73fbfcad | 1235 | if (dump_enabled_p ()) |
ebfd146a | 1236 | { |
78c60e3d SS |
1237 | dump_printf_loc (MSG_NOTE, vect_location, |
1238 | "created new init_stmt: "); | |
1239 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1240 | } |
418b7df3 RG |
1241 | } |
1242 | ||
1243 | /* Function vect_init_vector. | |
ebfd146a | 1244 | |
5467ee52 RG |
1245 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1246 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1247 | vector type a vector with all elements equal to VAL is created first. | |
1248 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1249 | initialization at the loop preheader. | |
418b7df3 RG |
1250 | Return the DEF of INIT_STMT. |
1251 | It will be used in the vectorization of STMT. */ | |
1252 | ||
1253 | tree | |
355fe088 | 1254 | vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 | 1255 | { |
355fe088 | 1256 | gimple *init_stmt; |
418b7df3 RG |
1257 | tree new_temp; |
1258 | ||
5467ee52 RG |
1259 | if (TREE_CODE (type) == VECTOR_TYPE |
1260 | && TREE_CODE (TREE_TYPE (val)) != VECTOR_TYPE) | |
418b7df3 | 1261 | { |
5467ee52 | 1262 | if (!types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) |
418b7df3 | 1263 | { |
5a308cf1 IE |
1264 | /* Scalar boolean value should be transformed into |
1265 | all zeros or all ones value before building a vector. */ | |
1266 | if (VECTOR_BOOLEAN_TYPE_P (type)) | |
1267 | { | |
b3d51f23 IE |
1268 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1269 | tree false_val = build_zero_cst (TREE_TYPE (type)); | |
5a308cf1 IE |
1270 | |
1271 | if (CONSTANT_CLASS_P (val)) | |
1272 | val = integer_zerop (val) ? false_val : true_val; | |
1273 | else | |
1274 | { | |
1275 | new_temp = make_ssa_name (TREE_TYPE (type)); | |
1276 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, | |
1277 | val, true_val, false_val); | |
1278 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
1279 | val = new_temp; | |
1280 | } | |
1281 | } | |
1282 | else if (CONSTANT_CLASS_P (val)) | |
42fd8198 | 1283 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1284 | else |
1285 | { | |
b731b390 | 1286 | new_temp = make_ssa_name (TREE_TYPE (type)); |
0d0e4a03 | 1287 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); |
418b7df3 | 1288 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1289 | val = new_temp; |
418b7df3 RG |
1290 | } |
1291 | } | |
5467ee52 | 1292 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1293 | } |
1294 | ||
0e22bb5a RB |
1295 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1296 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1297 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1298 | return new_temp; |
ebfd146a IR |
1299 | } |
1300 | ||
a70d6342 | 1301 | |
ebfd146a IR |
1302 | /* Function vect_get_vec_def_for_operand. |
1303 | ||
ff802fa1 | 1304 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
1305 | used in the vectorized stmt for STMT. |
1306 | ||
1307 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1308 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1309 | ||
1310 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
42fd8198 IE |
1311 | needs to be introduced. VECTYPE may be used to specify a required type for |
1312 | vector invariant. */ | |
ebfd146a IR |
1313 | |
1314 | tree | |
42fd8198 | 1315 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) |
ebfd146a IR |
1316 | { |
1317 | tree vec_oprnd; | |
355fe088 TS |
1318 | gimple *vec_stmt; |
1319 | gimple *def_stmt; | |
ebfd146a IR |
1320 | stmt_vec_info def_stmt_info = NULL; |
1321 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
42fd8198 | 1322 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); |
ebfd146a | 1323 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
ebfd146a IR |
1324 | enum vect_def_type dt; |
1325 | bool is_simple_use; | |
1326 | tree vector_type; | |
1327 | ||
73fbfcad | 1328 | if (dump_enabled_p ()) |
ebfd146a | 1329 | { |
78c60e3d SS |
1330 | dump_printf_loc (MSG_NOTE, vect_location, |
1331 | "vect_get_vec_def_for_operand: "); | |
1332 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
e645e942 | 1333 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1334 | } |
1335 | ||
81c40241 | 1336 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt); |
ebfd146a | 1337 | gcc_assert (is_simple_use); |
73fbfcad | 1338 | if (dump_enabled_p ()) |
ebfd146a | 1339 | { |
78c60e3d | 1340 | int loc_printed = 0; |
ebfd146a IR |
1341 | if (def_stmt) |
1342 | { | |
78c60e3d SS |
1343 | if (loc_printed) |
1344 | dump_printf (MSG_NOTE, " def_stmt = "); | |
1345 | else | |
1346 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1347 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
ebfd146a IR |
1348 | } |
1349 | } | |
1350 | ||
1351 | switch (dt) | |
1352 | { | |
81c40241 | 1353 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1354 | case vect_constant_def: |
81c40241 | 1355 | case vect_external_def: |
ebfd146a | 1356 | { |
42fd8198 IE |
1357 | if (vectype) |
1358 | vector_type = vectype; | |
1359 | else if (TREE_CODE (TREE_TYPE (op)) == BOOLEAN_TYPE | |
1360 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) | |
1361 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1362 | else | |
1363 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1364 | ||
7569a6cc | 1365 | gcc_assert (vector_type); |
418b7df3 | 1366 | return vect_init_vector (stmt, op, vector_type, NULL); |
ebfd146a IR |
1367 | } |
1368 | ||
81c40241 | 1369 | /* operand is defined inside the loop. */ |
8644a673 | 1370 | case vect_internal_def: |
ebfd146a | 1371 | { |
ebfd146a IR |
1372 | /* Get the def from the vectorized stmt. */ |
1373 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1374 | |
ebfd146a | 1375 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1376 | /* Get vectorized pattern statement. */ |
1377 | if (!vec_stmt | |
1378 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1379 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1380 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1381 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1382 | gcc_assert (vec_stmt); |
1383 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1384 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1385 | else if (is_gimple_call (vec_stmt)) | |
1386 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1387 | else | |
1388 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1389 | return vec_oprnd; | |
1390 | } | |
1391 | ||
81c40241 | 1392 | /* operand is defined by a loop header phi - reduction */ |
ebfd146a | 1393 | case vect_reduction_def: |
06066f92 | 1394 | case vect_double_reduction_def: |
7c5222ff | 1395 | case vect_nested_cycle: |
81c40241 RB |
1396 | /* Code should use get_initial_def_for_reduction. */ |
1397 | gcc_unreachable (); | |
ebfd146a | 1398 | |
81c40241 | 1399 | /* operand is defined by loop-header phi - induction. */ |
ebfd146a IR |
1400 | case vect_induction_def: |
1401 | { | |
1402 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1403 | ||
1404 | /* Get the def from the vectorized stmt. */ | |
1405 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1406 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1407 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1408 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1409 | else | |
1410 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1411 | return vec_oprnd; |
1412 | } | |
1413 | ||
1414 | default: | |
1415 | gcc_unreachable (); | |
1416 | } | |
1417 | } | |
1418 | ||
1419 | ||
1420 | /* Function vect_get_vec_def_for_stmt_copy | |
1421 | ||
ff802fa1 | 1422 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1423 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1424 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1425 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1426 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1427 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1428 | DT is the type of the vector def VEC_OPRND. |
1429 | ||
1430 | Context: | |
1431 | In case the vectorization factor (VF) is bigger than the number | |
1432 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1433 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1434 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1435 | smallest data-type determines the VF, and as a result, when vectorizing |
1436 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1437 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1438 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1439 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1440 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1441 | ||
1442 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1443 | |
ebfd146a IR |
1444 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1445 | VS1.1: vx.1 = memref1 VS1.2 | |
1446 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1447 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1448 | |
1449 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1450 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1451 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1452 | VSnew.3: vz3 = vx.3 + ... | |
1453 | ||
1454 | The vectorization of S1 is explained in vectorizable_load. | |
1455 | The vectorization of S2: | |
b8698a0f L |
1456 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1457 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1458 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1459 | returns the vector-def 'vx.0'. |
1460 | ||
b8698a0f L |
1461 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1462 | function is called to get the relevant vector-def for each operand. It is | |
1463 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1464 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1465 | ||
b8698a0f L |
1466 | For example, to obtain the vector-def 'vx.1' in order to create the |
1467 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1468 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1469 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1470 | and return its def ('vx.1'). | |
1471 | Overall, to create the above sequence this function will be called 3 times: | |
1472 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1473 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1474 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1475 | ||
1476 | tree | |
1477 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1478 | { | |
355fe088 | 1479 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1480 | stmt_vec_info def_stmt_info; |
1481 | ||
1482 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1483 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1484 | return vec_oprnd; |
1485 | ||
1486 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1487 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1488 | gcc_assert (def_stmt_info); | |
1489 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1490 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1491 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1492 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1493 | else | |
1494 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1495 | return vec_oprnd; | |
1496 | } | |
1497 | ||
1498 | ||
1499 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1500 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1501 | |
1502 | static void | |
b8698a0f | 1503 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1504 | vec<tree> *vec_oprnds0, |
1505 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1506 | { |
9771b263 | 1507 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1508 | |
1509 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1510 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1511 | |
9771b263 | 1512 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1513 | { |
9771b263 | 1514 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1515 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1516 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1517 | } |
1518 | } | |
1519 | ||
1520 | ||
d092494c IR |
1521 | /* Get vectorized definitions for OP0 and OP1. |
1522 | REDUC_INDEX is the index of reduction operand in case of reduction, | |
1523 | and -1 otherwise. */ | |
ebfd146a | 1524 | |
d092494c | 1525 | void |
355fe088 | 1526 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1527 | vec<tree> *vec_oprnds0, |
1528 | vec<tree> *vec_oprnds1, | |
d092494c | 1529 | slp_tree slp_node, int reduc_index) |
ebfd146a IR |
1530 | { |
1531 | if (slp_node) | |
d092494c IR |
1532 | { |
1533 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1534 | auto_vec<tree> ops (nops); |
1535 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1536 | |
9771b263 | 1537 | ops.quick_push (op0); |
d092494c | 1538 | if (op1) |
9771b263 | 1539 | ops.quick_push (op1); |
d092494c IR |
1540 | |
1541 | vect_get_slp_defs (ops, slp_node, &vec_defs, reduc_index); | |
1542 | ||
37b5ec8f | 1543 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1544 | if (op1) |
37b5ec8f | 1545 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1546 | } |
ebfd146a IR |
1547 | else |
1548 | { | |
1549 | tree vec_oprnd; | |
1550 | ||
9771b263 | 1551 | vec_oprnds0->create (1); |
81c40241 | 1552 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1553 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1554 | |
1555 | if (op1) | |
1556 | { | |
9771b263 | 1557 | vec_oprnds1->create (1); |
81c40241 | 1558 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1559 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1560 | } |
1561 | } | |
1562 | } | |
1563 | ||
1564 | ||
1565 | /* Function vect_finish_stmt_generation. | |
1566 | ||
1567 | Insert a new stmt. */ | |
1568 | ||
1569 | void | |
355fe088 | 1570 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1571 | gimple_stmt_iterator *gsi) |
1572 | { | |
1573 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
310213d4 | 1574 | vec_info *vinfo = stmt_info->vinfo; |
ebfd146a IR |
1575 | |
1576 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1577 | ||
54e8e2c3 RG |
1578 | if (!gsi_end_p (*gsi) |
1579 | && gimple_has_mem_ops (vec_stmt)) | |
1580 | { | |
355fe088 | 1581 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1582 | tree vuse = gimple_vuse (at_stmt); |
1583 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1584 | { | |
1585 | tree vdef = gimple_vdef (at_stmt); | |
1586 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1587 | /* If we have an SSA vuse and insert a store, update virtual | |
1588 | SSA form to avoid triggering the renamer. Do so only | |
1589 | if we can easily see all uses - which is what almost always | |
1590 | happens with the way vectorized stmts are inserted. */ | |
1591 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1592 | && ((is_gimple_assign (vec_stmt) | |
1593 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1594 | || (is_gimple_call (vec_stmt) | |
1595 | && !(gimple_call_flags (vec_stmt) | |
1596 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1597 | { | |
1598 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1599 | gimple_set_vdef (vec_stmt, new_vdef); | |
1600 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1601 | } | |
1602 | } | |
1603 | } | |
ebfd146a IR |
1604 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1605 | ||
310213d4 | 1606 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, vinfo)); |
ebfd146a | 1607 | |
73fbfcad | 1608 | if (dump_enabled_p ()) |
ebfd146a | 1609 | { |
78c60e3d SS |
1610 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1611 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
ebfd146a IR |
1612 | } |
1613 | ||
ad885386 | 1614 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
8e91d222 JJ |
1615 | |
1616 | /* While EH edges will generally prevent vectorization, stmt might | |
1617 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1618 | that could throw are part of the same region. */ | |
1619 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1620 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1621 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
ebfd146a IR |
1622 | } |
1623 | ||
70439f0d RS |
1624 | /* We want to vectorize a call to combined function CFN with function |
1625 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN | |
1626 | as the types of all inputs. Check whether this is possible using | |
1627 | an internal function, returning its code if so or IFN_LAST if not. */ | |
ebfd146a | 1628 | |
70439f0d RS |
1629 | static internal_fn |
1630 | vectorizable_internal_function (combined_fn cfn, tree fndecl, | |
1631 | tree vectype_out, tree vectype_in) | |
ebfd146a | 1632 | { |
70439f0d RS |
1633 | internal_fn ifn; |
1634 | if (internal_fn_p (cfn)) | |
1635 | ifn = as_internal_fn (cfn); | |
1636 | else | |
1637 | ifn = associated_internal_fn (fndecl); | |
1638 | if (ifn != IFN_LAST && direct_internal_fn_p (ifn)) | |
1639 | { | |
1640 | const direct_internal_fn_info &info = direct_internal_fn (ifn); | |
1641 | if (info.vectorizable) | |
1642 | { | |
1643 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); | |
1644 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); | |
d95ab70a RS |
1645 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1646 | OPTIMIZE_FOR_SPEED)) | |
70439f0d RS |
1647 | return ifn; |
1648 | } | |
1649 | } | |
1650 | return IFN_LAST; | |
ebfd146a IR |
1651 | } |
1652 | ||
5ce9450f | 1653 | |
355fe088 | 1654 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1655 | gimple_stmt_iterator *); |
1656 | ||
1657 | ||
1658 | /* Function vectorizable_mask_load_store. | |
1659 | ||
1660 | Check if STMT performs a conditional load or store that can be vectorized. | |
1661 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1662 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
1663 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1664 | ||
1665 | static bool | |
355fe088 TS |
1666 | vectorizable_mask_load_store (gimple *stmt, gimple_stmt_iterator *gsi, |
1667 | gimple **vec_stmt, slp_tree slp_node) | |
5ce9450f JJ |
1668 | { |
1669 | tree vec_dest = NULL; | |
1670 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1671 | stmt_vec_info prev_stmt_info; | |
1672 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1673 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1674 | bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
1675 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
1676 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
57e2f6ad | 1677 | tree rhs_vectype = NULL_TREE; |
045c1278 | 1678 | tree mask_vectype; |
5ce9450f | 1679 | tree elem_type; |
355fe088 | 1680 | gimple *new_stmt; |
5ce9450f JJ |
1681 | tree dummy; |
1682 | tree dataref_ptr = NULL_TREE; | |
355fe088 | 1683 | gimple *ptr_incr; |
5ce9450f JJ |
1684 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
1685 | int ncopies; | |
1686 | int i, j; | |
1687 | bool inv_p; | |
1688 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; | |
1689 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
1690 | int gather_scale = 1; | |
1691 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
1692 | bool is_store; | |
1693 | tree mask; | |
355fe088 | 1694 | gimple *def_stmt; |
5ce9450f JJ |
1695 | enum vect_def_type dt; |
1696 | ||
1697 | if (slp_node != NULL) | |
1698 | return false; | |
1699 | ||
1700 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
1701 | gcc_assert (ncopies >= 1); | |
1702 | ||
1703 | is_store = gimple_call_internal_fn (stmt) == IFN_MASK_STORE; | |
1704 | mask = gimple_call_arg (stmt, 2); | |
045c1278 IE |
1705 | |
1706 | if (TREE_CODE (TREE_TYPE (mask)) != BOOLEAN_TYPE) | |
5ce9450f JJ |
1707 | return false; |
1708 | ||
1709 | /* FORNOW. This restriction should be relaxed. */ | |
1710 | if (nested_in_vect_loop && ncopies > 1) | |
1711 | { | |
1712 | if (dump_enabled_p ()) | |
1713 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1714 | "multiple types in nested loop."); | |
1715 | return false; | |
1716 | } | |
1717 | ||
1718 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
1719 | return false; | |
1720 | ||
66c16fd9 RB |
1721 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
1722 | && ! vec_stmt) | |
5ce9450f JJ |
1723 | return false; |
1724 | ||
1725 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
1726 | return false; | |
1727 | ||
1728 | elem_type = TREE_TYPE (vectype); | |
1729 | ||
1730 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1731 | return false; | |
1732 | ||
f2e2a985 | 1733 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
5ce9450f JJ |
1734 | return false; |
1735 | ||
045c1278 IE |
1736 | if (TREE_CODE (mask) != SSA_NAME) |
1737 | return false; | |
1738 | ||
1739 | if (!vect_is_simple_use (mask, loop_vinfo, &def_stmt, &dt, &mask_vectype)) | |
1740 | return false; | |
1741 | ||
1742 | if (!mask_vectype) | |
1743 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
1744 | ||
dc6a3147 IE |
1745 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype) |
1746 | || TYPE_VECTOR_SUBPARTS (mask_vectype) != TYPE_VECTOR_SUBPARTS (vectype)) | |
045c1278 IE |
1747 | return false; |
1748 | ||
57e2f6ad IE |
1749 | if (is_store) |
1750 | { | |
1751 | tree rhs = gimple_call_arg (stmt, 3); | |
1752 | if (!vect_is_simple_use (rhs, loop_vinfo, &def_stmt, &dt, &rhs_vectype)) | |
1753 | return false; | |
1754 | } | |
1755 | ||
3bab6342 | 1756 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5ce9450f | 1757 | { |
355fe088 | 1758 | gimple *def_stmt; |
3bab6342 | 1759 | gather_decl = vect_check_gather_scatter (stmt, loop_vinfo, &gather_base, |
5ce9450f JJ |
1760 | &gather_off, &gather_scale); |
1761 | gcc_assert (gather_decl); | |
81c40241 RB |
1762 | if (!vect_is_simple_use (gather_off, loop_vinfo, &def_stmt, &gather_dt, |
1763 | &gather_off_vectype)) | |
5ce9450f JJ |
1764 | { |
1765 | if (dump_enabled_p ()) | |
1766 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1767 | "gather index use not simple."); | |
1768 | return false; | |
1769 | } | |
03b9e8e4 JJ |
1770 | |
1771 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1772 | tree masktype | |
1773 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
1774 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
1775 | { | |
1776 | if (dump_enabled_p ()) | |
1777 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1778 | "masked gather with integer mask not supported."); | |
1779 | return false; | |
1780 | } | |
5ce9450f JJ |
1781 | } |
1782 | else if (tree_int_cst_compare (nested_in_vect_loop | |
1783 | ? STMT_VINFO_DR_STEP (stmt_info) | |
1784 | : DR_STEP (dr), size_zero_node) <= 0) | |
1785 | return false; | |
1786 | else if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
045c1278 IE |
1787 | || !can_vec_mask_load_store_p (TYPE_MODE (vectype), |
1788 | TYPE_MODE (mask_vectype), | |
57e2f6ad IE |
1789 | !is_store) |
1790 | || (rhs_vectype | |
1791 | && !useless_type_conversion_p (vectype, rhs_vectype))) | |
5ce9450f JJ |
1792 | return false; |
1793 | ||
5ce9450f JJ |
1794 | if (!vec_stmt) /* transformation not required. */ |
1795 | { | |
1796 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1797 | if (is_store) | |
1798 | vect_model_store_cost (stmt_info, ncopies, false, dt, | |
1799 | NULL, NULL, NULL); | |
1800 | else | |
1801 | vect_model_load_cost (stmt_info, ncopies, false, NULL, NULL, NULL); | |
1802 | return true; | |
1803 | } | |
1804 | ||
1805 | /** Transform. **/ | |
1806 | ||
3bab6342 | 1807 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5ce9450f JJ |
1808 | { |
1809 | tree vec_oprnd0 = NULL_TREE, op; | |
1810 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1811 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
acdcd61b | 1812 | tree ptr, vec_mask = NULL_TREE, mask_op = NULL_TREE, var, scale; |
5ce9450f | 1813 | tree perm_mask = NULL_TREE, prev_res = NULL_TREE; |
acdcd61b | 1814 | tree mask_perm_mask = NULL_TREE; |
5ce9450f JJ |
1815 | edge pe = loop_preheader_edge (loop); |
1816 | gimple_seq seq; | |
1817 | basic_block new_bb; | |
1818 | enum { NARROW, NONE, WIDEN } modifier; | |
1819 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
1820 | ||
acdcd61b JJ |
1821 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); |
1822 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1823 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1824 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1825 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1826 | scaletype = TREE_VALUE (arglist); | |
1827 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
1828 | && types_compatible_p (srctype, masktype)); | |
1829 | ||
5ce9450f JJ |
1830 | if (nunits == gather_off_nunits) |
1831 | modifier = NONE; | |
1832 | else if (nunits == gather_off_nunits / 2) | |
1833 | { | |
1834 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
1835 | modifier = WIDEN; | |
1836 | ||
1837 | for (i = 0; i < gather_off_nunits; ++i) | |
1838 | sel[i] = i | nunits; | |
1839 | ||
557be5a8 | 1840 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
5ce9450f JJ |
1841 | } |
1842 | else if (nunits == gather_off_nunits * 2) | |
1843 | { | |
1844 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
1845 | modifier = NARROW; | |
1846 | ||
1847 | for (i = 0; i < nunits; ++i) | |
1848 | sel[i] = i < gather_off_nunits | |
1849 | ? i : i + nunits - gather_off_nunits; | |
1850 | ||
557be5a8 | 1851 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
5ce9450f | 1852 | ncopies *= 2; |
acdcd61b JJ |
1853 | for (i = 0; i < nunits; ++i) |
1854 | sel[i] = i | gather_off_nunits; | |
557be5a8 | 1855 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, sel); |
5ce9450f JJ |
1856 | } |
1857 | else | |
1858 | gcc_unreachable (); | |
1859 | ||
5ce9450f JJ |
1860 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); |
1861 | ||
1862 | ptr = fold_convert (ptrtype, gather_base); | |
1863 | if (!is_gimple_min_invariant (ptr)) | |
1864 | { | |
1865 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
1866 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
1867 | gcc_assert (!new_bb); | |
1868 | } | |
1869 | ||
1870 | scale = build_int_cst (scaletype, gather_scale); | |
1871 | ||
1872 | prev_stmt_info = NULL; | |
1873 | for (j = 0; j < ncopies; ++j) | |
1874 | { | |
1875 | if (modifier == WIDEN && (j & 1)) | |
1876 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
1877 | perm_mask, stmt, gsi); | |
1878 | else if (j == 0) | |
1879 | op = vec_oprnd0 | |
81c40241 | 1880 | = vect_get_vec_def_for_operand (gather_off, stmt); |
5ce9450f JJ |
1881 | else |
1882 | op = vec_oprnd0 | |
1883 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
1884 | ||
1885 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
1886 | { | |
1887 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
1888 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 1889 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
5ce9450f JJ |
1890 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
1891 | new_stmt | |
0d0e4a03 | 1892 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
1893 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1894 | op = var; | |
1895 | } | |
1896 | ||
acdcd61b JJ |
1897 | if (mask_perm_mask && (j & 1)) |
1898 | mask_op = permute_vec_elements (mask_op, mask_op, | |
1899 | mask_perm_mask, stmt, gsi); | |
5ce9450f JJ |
1900 | else |
1901 | { | |
acdcd61b | 1902 | if (j == 0) |
81c40241 | 1903 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
acdcd61b JJ |
1904 | else |
1905 | { | |
81c40241 | 1906 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
acdcd61b JJ |
1907 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
1908 | } | |
5ce9450f | 1909 | |
acdcd61b JJ |
1910 | mask_op = vec_mask; |
1911 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
1912 | { | |
1913 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)) | |
1914 | == TYPE_VECTOR_SUBPARTS (masktype)); | |
0e22bb5a | 1915 | var = vect_get_new_ssa_name (masktype, vect_simple_var); |
acdcd61b JJ |
1916 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); |
1917 | new_stmt | |
0d0e4a03 | 1918 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_op); |
acdcd61b JJ |
1919 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1920 | mask_op = var; | |
1921 | } | |
5ce9450f JJ |
1922 | } |
1923 | ||
1924 | new_stmt | |
1925 | = gimple_build_call (gather_decl, 5, mask_op, ptr, op, mask_op, | |
1926 | scale); | |
1927 | ||
1928 | if (!useless_type_conversion_p (vectype, rettype)) | |
1929 | { | |
1930 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
1931 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 1932 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
5ce9450f JJ |
1933 | gimple_call_set_lhs (new_stmt, op); |
1934 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 1935 | var = make_ssa_name (vec_dest); |
5ce9450f | 1936 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
0d0e4a03 | 1937 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
1938 | } |
1939 | else | |
1940 | { | |
1941 | var = make_ssa_name (vec_dest, new_stmt); | |
1942 | gimple_call_set_lhs (new_stmt, var); | |
1943 | } | |
1944 | ||
1945 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1946 | ||
1947 | if (modifier == NARROW) | |
1948 | { | |
1949 | if ((j & 1) == 0) | |
1950 | { | |
1951 | prev_res = var; | |
1952 | continue; | |
1953 | } | |
1954 | var = permute_vec_elements (prev_res, var, | |
1955 | perm_mask, stmt, gsi); | |
1956 | new_stmt = SSA_NAME_DEF_STMT (var); | |
1957 | } | |
1958 | ||
1959 | if (prev_stmt_info == NULL) | |
1960 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1961 | else | |
1962 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1963 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1964 | } | |
3efe2e2c JJ |
1965 | |
1966 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
1967 | from the IL. */ | |
e6f5c25d IE |
1968 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
1969 | { | |
1970 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
1971 | stmt_info = vinfo_for_stmt (stmt); | |
1972 | } | |
3efe2e2c JJ |
1973 | tree lhs = gimple_call_lhs (stmt); |
1974 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
1975 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
1976 | set_vinfo_for_stmt (stmt, NULL); | |
1977 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
1978 | gsi_replace (gsi, new_stmt, true); | |
5ce9450f JJ |
1979 | return true; |
1980 | } | |
1981 | else if (is_store) | |
1982 | { | |
1983 | tree vec_rhs = NULL_TREE, vec_mask = NULL_TREE; | |
1984 | prev_stmt_info = NULL; | |
2d4dc223 | 1985 | LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true; |
5ce9450f JJ |
1986 | for (i = 0; i < ncopies; i++) |
1987 | { | |
1988 | unsigned align, misalign; | |
1989 | ||
1990 | if (i == 0) | |
1991 | { | |
1992 | tree rhs = gimple_call_arg (stmt, 3); | |
81c40241 RB |
1993 | vec_rhs = vect_get_vec_def_for_operand (rhs, stmt); |
1994 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); | |
5ce9450f JJ |
1995 | /* We should have catched mismatched types earlier. */ |
1996 | gcc_assert (useless_type_conversion_p (vectype, | |
1997 | TREE_TYPE (vec_rhs))); | |
1998 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, | |
1999 | NULL_TREE, &dummy, gsi, | |
2000 | &ptr_incr, false, &inv_p); | |
2001 | gcc_assert (!inv_p); | |
2002 | } | |
2003 | else | |
2004 | { | |
81c40241 | 2005 | vect_is_simple_use (vec_rhs, loop_vinfo, &def_stmt, &dt); |
5ce9450f | 2006 | vec_rhs = vect_get_vec_def_for_stmt_copy (dt, vec_rhs); |
81c40241 | 2007 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2008 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2009 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2010 | TYPE_SIZE_UNIT (vectype)); | |
2011 | } | |
2012 | ||
2013 | align = TYPE_ALIGN_UNIT (vectype); | |
2014 | if (aligned_access_p (dr)) | |
2015 | misalign = 0; | |
2016 | else if (DR_MISALIGNMENT (dr) == -1) | |
2017 | { | |
2018 | align = TYPE_ALIGN_UNIT (elem_type); | |
2019 | misalign = 0; | |
2020 | } | |
2021 | else | |
2022 | misalign = DR_MISALIGNMENT (dr); | |
2023 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2024 | misalign); | |
08554c26 JJ |
2025 | tree ptr = build_int_cst (TREE_TYPE (gimple_call_arg (stmt, 1)), |
2026 | misalign ? misalign & -misalign : align); | |
5ce9450f JJ |
2027 | new_stmt |
2028 | = gimple_build_call_internal (IFN_MASK_STORE, 4, dataref_ptr, | |
08554c26 | 2029 | ptr, vec_mask, vec_rhs); |
5ce9450f JJ |
2030 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2031 | if (i == 0) | |
2032 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2033 | else | |
2034 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2035 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2036 | } | |
2037 | } | |
2038 | else | |
2039 | { | |
2040 | tree vec_mask = NULL_TREE; | |
2041 | prev_stmt_info = NULL; | |
2042 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); | |
2043 | for (i = 0; i < ncopies; i++) | |
2044 | { | |
2045 | unsigned align, misalign; | |
2046 | ||
2047 | if (i == 0) | |
2048 | { | |
81c40241 | 2049 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
5ce9450f JJ |
2050 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, |
2051 | NULL_TREE, &dummy, gsi, | |
2052 | &ptr_incr, false, &inv_p); | |
2053 | gcc_assert (!inv_p); | |
2054 | } | |
2055 | else | |
2056 | { | |
81c40241 | 2057 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2058 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2059 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2060 | TYPE_SIZE_UNIT (vectype)); | |
2061 | } | |
2062 | ||
2063 | align = TYPE_ALIGN_UNIT (vectype); | |
2064 | if (aligned_access_p (dr)) | |
2065 | misalign = 0; | |
2066 | else if (DR_MISALIGNMENT (dr) == -1) | |
2067 | { | |
2068 | align = TYPE_ALIGN_UNIT (elem_type); | |
2069 | misalign = 0; | |
2070 | } | |
2071 | else | |
2072 | misalign = DR_MISALIGNMENT (dr); | |
2073 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2074 | misalign); | |
08554c26 JJ |
2075 | tree ptr = build_int_cst (TREE_TYPE (gimple_call_arg (stmt, 1)), |
2076 | misalign ? misalign & -misalign : align); | |
5ce9450f JJ |
2077 | new_stmt |
2078 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, dataref_ptr, | |
08554c26 | 2079 | ptr, vec_mask); |
b731b390 | 2080 | gimple_call_set_lhs (new_stmt, make_ssa_name (vec_dest)); |
5ce9450f JJ |
2081 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2082 | if (i == 0) | |
2083 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2084 | else | |
2085 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2086 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2087 | } | |
2088 | } | |
2089 | ||
3efe2e2c JJ |
2090 | if (!is_store) |
2091 | { | |
2092 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2093 | from the IL. */ | |
e6f5c25d IE |
2094 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
2095 | { | |
2096 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
2097 | stmt_info = vinfo_for_stmt (stmt); | |
2098 | } | |
3efe2e2c JJ |
2099 | tree lhs = gimple_call_lhs (stmt); |
2100 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2101 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2102 | set_vinfo_for_stmt (stmt, NULL); | |
2103 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2104 | gsi_replace (gsi, new_stmt, true); | |
2105 | } | |
2106 | ||
5ce9450f JJ |
2107 | return true; |
2108 | } | |
2109 | ||
b1b6836e RS |
2110 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
2111 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT | |
2112 | in a single step. On success, store the binary pack code in | |
2113 | *CONVERT_CODE. */ | |
2114 | ||
2115 | static bool | |
2116 | simple_integer_narrowing (tree vectype_out, tree vectype_in, | |
2117 | tree_code *convert_code) | |
2118 | { | |
2119 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) | |
2120 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) | |
2121 | return false; | |
2122 | ||
2123 | tree_code code; | |
2124 | int multi_step_cvt = 0; | |
2125 | auto_vec <tree, 8> interm_types; | |
2126 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, | |
2127 | &code, &multi_step_cvt, | |
2128 | &interm_types) | |
2129 | || multi_step_cvt) | |
2130 | return false; | |
2131 | ||
2132 | *convert_code = code; | |
2133 | return true; | |
2134 | } | |
5ce9450f | 2135 | |
ebfd146a IR |
2136 | /* Function vectorizable_call. |
2137 | ||
538dd0b7 | 2138 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 2139 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2140 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2141 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2142 | ||
2143 | static bool | |
355fe088 | 2144 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
190c2236 | 2145 | slp_tree slp_node) |
ebfd146a | 2146 | { |
538dd0b7 | 2147 | gcall *stmt; |
ebfd146a IR |
2148 | tree vec_dest; |
2149 | tree scalar_dest; | |
2150 | tree op, type; | |
2151 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
538dd0b7 | 2152 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a IR |
2153 | tree vectype_out, vectype_in; |
2154 | int nunits_in; | |
2155 | int nunits_out; | |
2156 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 2157 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 2158 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 2159 | tree fndecl, new_temp, rhs_type; |
355fe088 | 2160 | gimple *def_stmt; |
0502fb85 UB |
2161 | enum vect_def_type dt[3] |
2162 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
355fe088 | 2163 | gimple *new_stmt = NULL; |
ebfd146a | 2164 | int ncopies, j; |
6e1aa848 | 2165 | vec<tree> vargs = vNULL; |
ebfd146a IR |
2166 | enum { NARROW, NONE, WIDEN } modifier; |
2167 | size_t i, nargs; | |
9d5e7640 | 2168 | tree lhs; |
ebfd146a | 2169 | |
190c2236 | 2170 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2171 | return false; |
2172 | ||
66c16fd9 RB |
2173 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
2174 | && ! vec_stmt) | |
ebfd146a IR |
2175 | return false; |
2176 | ||
538dd0b7 DM |
2177 | /* Is GS a vectorizable call? */ |
2178 | stmt = dyn_cast <gcall *> (gs); | |
2179 | if (!stmt) | |
ebfd146a IR |
2180 | return false; |
2181 | ||
5ce9450f JJ |
2182 | if (gimple_call_internal_p (stmt) |
2183 | && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD | |
2184 | || gimple_call_internal_fn (stmt) == IFN_MASK_STORE)) | |
2185 | return vectorizable_mask_load_store (stmt, gsi, vec_stmt, | |
2186 | slp_node); | |
2187 | ||
0136f8f0 AH |
2188 | if (gimple_call_lhs (stmt) == NULL_TREE |
2189 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
2190 | return false; |
2191 | ||
0136f8f0 | 2192 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 2193 | |
b690cc0f RG |
2194 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2195 | ||
ebfd146a IR |
2196 | /* Process function arguments. */ |
2197 | rhs_type = NULL_TREE; | |
b690cc0f | 2198 | vectype_in = NULL_TREE; |
ebfd146a IR |
2199 | nargs = gimple_call_num_args (stmt); |
2200 | ||
1b1562a5 MM |
2201 | /* Bail out if the function has more than three arguments, we do not have |
2202 | interesting builtin functions to vectorize with more than two arguments | |
2203 | except for fma. No arguments is also not good. */ | |
2204 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
2205 | return false; |
2206 | ||
74bf76ed JJ |
2207 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2208 | if (gimple_call_internal_p (stmt) | |
2209 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2210 | { | |
2211 | nargs = 0; | |
2212 | rhs_type = unsigned_type_node; | |
2213 | } | |
2214 | ||
ebfd146a IR |
2215 | for (i = 0; i < nargs; i++) |
2216 | { | |
b690cc0f RG |
2217 | tree opvectype; |
2218 | ||
ebfd146a IR |
2219 | op = gimple_call_arg (stmt, i); |
2220 | ||
2221 | /* We can only handle calls with arguments of the same type. */ | |
2222 | if (rhs_type | |
8533c9d8 | 2223 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 2224 | { |
73fbfcad | 2225 | if (dump_enabled_p ()) |
78c60e3d | 2226 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2227 | "argument types differ.\n"); |
ebfd146a IR |
2228 | return false; |
2229 | } | |
b690cc0f RG |
2230 | if (!rhs_type) |
2231 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 2232 | |
81c40241 | 2233 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[i], &opvectype)) |
ebfd146a | 2234 | { |
73fbfcad | 2235 | if (dump_enabled_p ()) |
78c60e3d | 2236 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2237 | "use not simple.\n"); |
ebfd146a IR |
2238 | return false; |
2239 | } | |
ebfd146a | 2240 | |
b690cc0f RG |
2241 | if (!vectype_in) |
2242 | vectype_in = opvectype; | |
2243 | else if (opvectype | |
2244 | && opvectype != vectype_in) | |
2245 | { | |
73fbfcad | 2246 | if (dump_enabled_p ()) |
78c60e3d | 2247 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2248 | "argument vector types differ.\n"); |
b690cc0f RG |
2249 | return false; |
2250 | } | |
2251 | } | |
2252 | /* If all arguments are external or constant defs use a vector type with | |
2253 | the same size as the output vector type. */ | |
ebfd146a | 2254 | if (!vectype_in) |
b690cc0f | 2255 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2256 | if (vec_stmt) |
2257 | gcc_assert (vectype_in); | |
2258 | if (!vectype_in) | |
2259 | { | |
73fbfcad | 2260 | if (dump_enabled_p ()) |
7d8930a0 | 2261 | { |
78c60e3d SS |
2262 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2263 | "no vectype for scalar type "); | |
2264 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2265 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
2266 | } |
2267 | ||
2268 | return false; | |
2269 | } | |
ebfd146a IR |
2270 | |
2271 | /* FORNOW */ | |
b690cc0f RG |
2272 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2273 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
2274 | if (nunits_in == nunits_out / 2) |
2275 | modifier = NARROW; | |
2276 | else if (nunits_out == nunits_in) | |
2277 | modifier = NONE; | |
2278 | else if (nunits_out == nunits_in / 2) | |
2279 | modifier = WIDEN; | |
2280 | else | |
2281 | return false; | |
2282 | ||
70439f0d RS |
2283 | /* We only handle functions that do not read or clobber memory. */ |
2284 | if (gimple_vuse (stmt)) | |
2285 | { | |
2286 | if (dump_enabled_p ()) | |
2287 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2288 | "function reads from or writes to memory.\n"); | |
2289 | return false; | |
2290 | } | |
2291 | ||
ebfd146a IR |
2292 | /* For now, we only vectorize functions if a target specific builtin |
2293 | is available. TODO -- in some cases, it might be profitable to | |
2294 | insert the calls for pieces of the vector, in order to be able | |
2295 | to vectorize other operations in the loop. */ | |
70439f0d RS |
2296 | fndecl = NULL_TREE; |
2297 | internal_fn ifn = IFN_LAST; | |
2298 | combined_fn cfn = gimple_call_combined_fn (stmt); | |
2299 | tree callee = gimple_call_fndecl (stmt); | |
2300 | ||
2301 | /* First try using an internal function. */ | |
b1b6836e RS |
2302 | tree_code convert_code = ERROR_MARK; |
2303 | if (cfn != CFN_LAST | |
2304 | && (modifier == NONE | |
2305 | || (modifier == NARROW | |
2306 | && simple_integer_narrowing (vectype_out, vectype_in, | |
2307 | &convert_code)))) | |
70439f0d RS |
2308 | ifn = vectorizable_internal_function (cfn, callee, vectype_out, |
2309 | vectype_in); | |
2310 | ||
2311 | /* If that fails, try asking for a target-specific built-in function. */ | |
2312 | if (ifn == IFN_LAST) | |
2313 | { | |
2314 | if (cfn != CFN_LAST) | |
2315 | fndecl = targetm.vectorize.builtin_vectorized_function | |
2316 | (cfn, vectype_out, vectype_in); | |
2317 | else | |
2318 | fndecl = targetm.vectorize.builtin_md_vectorized_function | |
2319 | (callee, vectype_out, vectype_in); | |
2320 | } | |
2321 | ||
2322 | if (ifn == IFN_LAST && !fndecl) | |
ebfd146a | 2323 | { |
70439f0d | 2324 | if (cfn == CFN_GOMP_SIMD_LANE |
74bf76ed JJ |
2325 | && !slp_node |
2326 | && loop_vinfo | |
2327 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2328 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
2329 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2330 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
2331 | { | |
2332 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
2333 | { 0, 1, 2, ... vf - 1 } vector. */ | |
2334 | gcc_assert (nargs == 0); | |
2335 | } | |
2336 | else | |
2337 | { | |
2338 | if (dump_enabled_p ()) | |
2339 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 2340 | "function is not vectorizable.\n"); |
74bf76ed JJ |
2341 | return false; |
2342 | } | |
ebfd146a IR |
2343 | } |
2344 | ||
fce57248 | 2345 | if (slp_node) |
190c2236 | 2346 | ncopies = 1; |
b1b6836e | 2347 | else if (modifier == NARROW && ifn == IFN_LAST) |
ebfd146a IR |
2348 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
2349 | else | |
2350 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2351 | ||
2352 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2353 | needs to be generated. */ | |
2354 | gcc_assert (ncopies >= 1); | |
2355 | ||
2356 | if (!vec_stmt) /* transformation not required. */ | |
2357 | { | |
2358 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 2359 | if (dump_enabled_p ()) |
e645e942 TJ |
2360 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
2361 | "\n"); | |
c3e7ee41 | 2362 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
b1b6836e RS |
2363 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) |
2364 | add_stmt_cost (stmt_info->vinfo->target_cost_data, ncopies / 2, | |
2365 | vec_promote_demote, stmt_info, 0, vect_body); | |
2366 | ||
ebfd146a IR |
2367 | return true; |
2368 | } | |
2369 | ||
2370 | /** Transform. **/ | |
2371 | ||
73fbfcad | 2372 | if (dump_enabled_p ()) |
e645e942 | 2373 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
2374 | |
2375 | /* Handle def. */ | |
2376 | scalar_dest = gimple_call_lhs (stmt); | |
2377 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2378 | ||
2379 | prev_stmt_info = NULL; | |
b1b6836e | 2380 | if (modifier == NONE || ifn != IFN_LAST) |
ebfd146a | 2381 | { |
b1b6836e | 2382 | tree prev_res = NULL_TREE; |
ebfd146a IR |
2383 | for (j = 0; j < ncopies; ++j) |
2384 | { | |
2385 | /* Build argument list for the vectorized call. */ | |
2386 | if (j == 0) | |
9771b263 | 2387 | vargs.create (nargs); |
ebfd146a | 2388 | else |
9771b263 | 2389 | vargs.truncate (0); |
ebfd146a | 2390 | |
190c2236 JJ |
2391 | if (slp_node) |
2392 | { | |
ef062b13 | 2393 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2394 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2395 | |
2396 | for (i = 0; i < nargs; i++) | |
9771b263 | 2397 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2398 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2399 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2400 | |
2401 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2402 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
2403 | { |
2404 | size_t k; | |
2405 | for (k = 0; k < nargs; k++) | |
2406 | { | |
37b5ec8f | 2407 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 2408 | vargs[k] = vec_oprndsk[i]; |
190c2236 | 2409 | } |
b1b6836e RS |
2410 | if (modifier == NARROW) |
2411 | { | |
2412 | tree half_res = make_ssa_name (vectype_in); | |
2413 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2414 | gimple_call_set_lhs (new_stmt, half_res); | |
2415 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2416 | if ((i & 1) == 0) | |
2417 | { | |
2418 | prev_res = half_res; | |
2419 | continue; | |
2420 | } | |
2421 | new_temp = make_ssa_name (vec_dest); | |
2422 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
2423 | prev_res, half_res); | |
2424 | } | |
70439f0d | 2425 | else |
b1b6836e RS |
2426 | { |
2427 | if (ifn != IFN_LAST) | |
2428 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2429 | else | |
2430 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2431 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2432 | gimple_call_set_lhs (new_stmt, new_temp); | |
2433 | } | |
190c2236 | 2434 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 2435 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2436 | } |
2437 | ||
2438 | for (i = 0; i < nargs; i++) | |
2439 | { | |
37b5ec8f | 2440 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2441 | vec_oprndsi.release (); |
190c2236 | 2442 | } |
190c2236 JJ |
2443 | continue; |
2444 | } | |
2445 | ||
ebfd146a IR |
2446 | for (i = 0; i < nargs; i++) |
2447 | { | |
2448 | op = gimple_call_arg (stmt, i); | |
2449 | if (j == 0) | |
2450 | vec_oprnd0 | |
81c40241 | 2451 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2452 | else |
63827fb8 IR |
2453 | { |
2454 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
2455 | vec_oprnd0 | |
2456 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
2457 | } | |
ebfd146a | 2458 | |
9771b263 | 2459 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
2460 | } |
2461 | ||
74bf76ed JJ |
2462 | if (gimple_call_internal_p (stmt) |
2463 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2464 | { | |
2465 | tree *v = XALLOCAVEC (tree, nunits_out); | |
2466 | int k; | |
2467 | for (k = 0; k < nunits_out; ++k) | |
2468 | v[k] = build_int_cst (unsigned_type_node, j * nunits_out + k); | |
2469 | tree cst = build_vector (vectype_out, v); | |
2470 | tree new_var | |
0e22bb5a | 2471 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 2472 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 2473 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 2474 | new_temp = make_ssa_name (vec_dest); |
0e22bb5a | 2475 | new_stmt = gimple_build_assign (new_temp, new_var); |
74bf76ed | 2476 | } |
b1b6836e RS |
2477 | else if (modifier == NARROW) |
2478 | { | |
2479 | tree half_res = make_ssa_name (vectype_in); | |
2480 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2481 | gimple_call_set_lhs (new_stmt, half_res); | |
2482 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2483 | if ((j & 1) == 0) | |
2484 | { | |
2485 | prev_res = half_res; | |
2486 | continue; | |
2487 | } | |
2488 | new_temp = make_ssa_name (vec_dest); | |
2489 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
2490 | prev_res, half_res); | |
2491 | } | |
74bf76ed JJ |
2492 | else |
2493 | { | |
70439f0d RS |
2494 | if (ifn != IFN_LAST) |
2495 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2496 | else | |
2497 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
74bf76ed JJ |
2498 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2499 | gimple_call_set_lhs (new_stmt, new_temp); | |
2500 | } | |
ebfd146a IR |
2501 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2502 | ||
b1b6836e | 2503 | if (j == (modifier == NARROW ? 1 : 0)) |
ebfd146a IR |
2504 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; |
2505 | else | |
2506 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2507 | ||
2508 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2509 | } | |
b1b6836e RS |
2510 | } |
2511 | else if (modifier == NARROW) | |
2512 | { | |
ebfd146a IR |
2513 | for (j = 0; j < ncopies; ++j) |
2514 | { | |
2515 | /* Build argument list for the vectorized call. */ | |
2516 | if (j == 0) | |
9771b263 | 2517 | vargs.create (nargs * 2); |
ebfd146a | 2518 | else |
9771b263 | 2519 | vargs.truncate (0); |
ebfd146a | 2520 | |
190c2236 JJ |
2521 | if (slp_node) |
2522 | { | |
ef062b13 | 2523 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2524 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2525 | |
2526 | for (i = 0; i < nargs; i++) | |
9771b263 | 2527 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2528 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2529 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2530 | |
2531 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2532 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
2533 | { |
2534 | size_t k; | |
9771b263 | 2535 | vargs.truncate (0); |
190c2236 JJ |
2536 | for (k = 0; k < nargs; k++) |
2537 | { | |
37b5ec8f | 2538 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
2539 | vargs.quick_push (vec_oprndsk[i]); |
2540 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 | 2541 | } |
70439f0d RS |
2542 | if (ifn != IFN_LAST) |
2543 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2544 | else | |
2545 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
190c2236 JJ |
2546 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2547 | gimple_call_set_lhs (new_stmt, new_temp); | |
2548 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2549 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2550 | } |
2551 | ||
2552 | for (i = 0; i < nargs; i++) | |
2553 | { | |
37b5ec8f | 2554 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2555 | vec_oprndsi.release (); |
190c2236 | 2556 | } |
190c2236 JJ |
2557 | continue; |
2558 | } | |
2559 | ||
ebfd146a IR |
2560 | for (i = 0; i < nargs; i++) |
2561 | { | |
2562 | op = gimple_call_arg (stmt, i); | |
2563 | if (j == 0) | |
2564 | { | |
2565 | vec_oprnd0 | |
81c40241 | 2566 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2567 | vec_oprnd1 |
63827fb8 | 2568 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2569 | } |
2570 | else | |
2571 | { | |
336ecb65 | 2572 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 2573 | vec_oprnd0 |
63827fb8 | 2574 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 2575 | vec_oprnd1 |
63827fb8 | 2576 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2577 | } |
2578 | ||
9771b263 DN |
2579 | vargs.quick_push (vec_oprnd0); |
2580 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
2581 | } |
2582 | ||
b1b6836e | 2583 | new_stmt = gimple_build_call_vec (fndecl, vargs); |
ebfd146a IR |
2584 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2585 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
2586 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2587 | ||
2588 | if (j == 0) | |
2589 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2590 | else | |
2591 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2592 | ||
2593 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2594 | } | |
2595 | ||
2596 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a | 2597 | } |
b1b6836e RS |
2598 | else |
2599 | /* No current target implements this case. */ | |
2600 | return false; | |
ebfd146a | 2601 | |
9771b263 | 2602 | vargs.release (); |
ebfd146a | 2603 | |
ebfd146a IR |
2604 | /* The call in STMT might prevent it from being removed in dce. |
2605 | We however cannot remove it here, due to the way the ssa name | |
2606 | it defines is mapped to the new definition. So just replace | |
2607 | rhs of the statement with something harmless. */ | |
2608 | ||
dd34c087 JJ |
2609 | if (slp_node) |
2610 | return true; | |
2611 | ||
ebfd146a | 2612 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
2613 | if (is_pattern_stmt_p (stmt_info)) |
2614 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
2615 | else | |
2616 | lhs = gimple_call_lhs (stmt); | |
3cc2fa2a JJ |
2617 | |
2618 | if (gimple_call_internal_p (stmt) | |
2619 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2620 | { | |
2621 | /* Replace uses of the lhs of GOMP_SIMD_LANE call outside the loop | |
2622 | with vf - 1 rather than 0, that is the last iteration of the | |
2623 | vectorized loop. */ | |
2624 | imm_use_iterator iter; | |
2625 | use_operand_p use_p; | |
355fe088 | 2626 | gimple *use_stmt; |
3cc2fa2a JJ |
2627 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs) |
2628 | { | |
2629 | basic_block use_bb = gimple_bb (use_stmt); | |
2630 | if (use_bb | |
2631 | && !flow_bb_inside_loop_p (LOOP_VINFO_LOOP (loop_vinfo), use_bb)) | |
2632 | { | |
2633 | FOR_EACH_IMM_USE_ON_STMT (use_p, iter) | |
2634 | SET_USE (use_p, build_int_cst (TREE_TYPE (lhs), | |
2635 | ncopies * nunits_out - 1)); | |
2636 | update_stmt (use_stmt); | |
2637 | } | |
2638 | } | |
2639 | } | |
2640 | ||
9d5e7640 | 2641 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
ebfd146a | 2642 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 2643 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
2644 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
2645 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
2646 | |
2647 | return true; | |
2648 | } | |
2649 | ||
2650 | ||
0136f8f0 AH |
2651 | struct simd_call_arg_info |
2652 | { | |
2653 | tree vectype; | |
2654 | tree op; | |
2655 | enum vect_def_type dt; | |
2656 | HOST_WIDE_INT linear_step; | |
2657 | unsigned int align; | |
17b658af | 2658 | bool simd_lane_linear; |
0136f8f0 AH |
2659 | }; |
2660 | ||
17b658af JJ |
2661 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
2662 | is linear within simd lane (but not within whole loop), note it in | |
2663 | *ARGINFO. */ | |
2664 | ||
2665 | static void | |
2666 | vect_simd_lane_linear (tree op, struct loop *loop, | |
2667 | struct simd_call_arg_info *arginfo) | |
2668 | { | |
355fe088 | 2669 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
2670 | |
2671 | if (!is_gimple_assign (def_stmt) | |
2672 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
2673 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
2674 | return; | |
2675 | ||
2676 | tree base = gimple_assign_rhs1 (def_stmt); | |
2677 | HOST_WIDE_INT linear_step = 0; | |
2678 | tree v = gimple_assign_rhs2 (def_stmt); | |
2679 | while (TREE_CODE (v) == SSA_NAME) | |
2680 | { | |
2681 | tree t; | |
2682 | def_stmt = SSA_NAME_DEF_STMT (v); | |
2683 | if (is_gimple_assign (def_stmt)) | |
2684 | switch (gimple_assign_rhs_code (def_stmt)) | |
2685 | { | |
2686 | case PLUS_EXPR: | |
2687 | t = gimple_assign_rhs2 (def_stmt); | |
2688 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
2689 | return; | |
2690 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
2691 | v = gimple_assign_rhs1 (def_stmt); | |
2692 | continue; | |
2693 | case MULT_EXPR: | |
2694 | t = gimple_assign_rhs2 (def_stmt); | |
2695 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
2696 | return; | |
2697 | linear_step = tree_to_shwi (t); | |
2698 | v = gimple_assign_rhs1 (def_stmt); | |
2699 | continue; | |
2700 | CASE_CONVERT: | |
2701 | t = gimple_assign_rhs1 (def_stmt); | |
2702 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
2703 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
2704 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
2705 | return; | |
2706 | if (!linear_step) | |
2707 | linear_step = 1; | |
2708 | v = t; | |
2709 | continue; | |
2710 | default: | |
2711 | return; | |
2712 | } | |
2713 | else if (is_gimple_call (def_stmt) | |
2714 | && gimple_call_internal_p (def_stmt) | |
2715 | && gimple_call_internal_fn (def_stmt) == IFN_GOMP_SIMD_LANE | |
2716 | && loop->simduid | |
2717 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
2718 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
2719 | == loop->simduid)) | |
2720 | { | |
2721 | if (!linear_step) | |
2722 | linear_step = 1; | |
2723 | arginfo->linear_step = linear_step; | |
2724 | arginfo->op = base; | |
2725 | arginfo->simd_lane_linear = true; | |
2726 | return; | |
2727 | } | |
2728 | } | |
2729 | } | |
2730 | ||
0136f8f0 AH |
2731 | /* Function vectorizable_simd_clone_call. |
2732 | ||
2733 | Check if STMT performs a function call that can be vectorized | |
2734 | by calling a simd clone of the function. | |
2735 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2736 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2737 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2738 | ||
2739 | static bool | |
355fe088 TS |
2740 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
2741 | gimple **vec_stmt, slp_tree slp_node) | |
0136f8f0 AH |
2742 | { |
2743 | tree vec_dest; | |
2744 | tree scalar_dest; | |
2745 | tree op, type; | |
2746 | tree vec_oprnd0 = NULL_TREE; | |
2747 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
2748 | tree vectype; | |
2749 | unsigned int nunits; | |
2750 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2751 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 2752 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 2753 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 2754 | tree fndecl, new_temp; |
355fe088 TS |
2755 | gimple *def_stmt; |
2756 | gimple *new_stmt = NULL; | |
0136f8f0 | 2757 | int ncopies, j; |
00426f9a | 2758 | auto_vec<simd_call_arg_info> arginfo; |
0136f8f0 AH |
2759 | vec<tree> vargs = vNULL; |
2760 | size_t i, nargs; | |
2761 | tree lhs, rtype, ratype; | |
2762 | vec<constructor_elt, va_gc> *ret_ctor_elts; | |
2763 | ||
2764 | /* Is STMT a vectorizable call? */ | |
2765 | if (!is_gimple_call (stmt)) | |
2766 | return false; | |
2767 | ||
2768 | fndecl = gimple_call_fndecl (stmt); | |
2769 | if (fndecl == NULL_TREE) | |
2770 | return false; | |
2771 | ||
d52f5295 | 2772 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
2773 | if (node == NULL || node->simd_clones == NULL) |
2774 | return false; | |
2775 | ||
2776 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
2777 | return false; | |
2778 | ||
66c16fd9 RB |
2779 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
2780 | && ! vec_stmt) | |
0136f8f0 AH |
2781 | return false; |
2782 | ||
2783 | if (gimple_call_lhs (stmt) | |
2784 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
2785 | return false; | |
2786 | ||
2787 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
2788 | ||
2789 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2790 | ||
2791 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
2792 | return false; | |
2793 | ||
2794 | /* FORNOW */ | |
fce57248 | 2795 | if (slp_node) |
0136f8f0 AH |
2796 | return false; |
2797 | ||
2798 | /* Process function arguments. */ | |
2799 | nargs = gimple_call_num_args (stmt); | |
2800 | ||
2801 | /* Bail out if the function has zero arguments. */ | |
2802 | if (nargs == 0) | |
2803 | return false; | |
2804 | ||
00426f9a | 2805 | arginfo.reserve (nargs, true); |
0136f8f0 AH |
2806 | |
2807 | for (i = 0; i < nargs; i++) | |
2808 | { | |
2809 | simd_call_arg_info thisarginfo; | |
2810 | affine_iv iv; | |
2811 | ||
2812 | thisarginfo.linear_step = 0; | |
2813 | thisarginfo.align = 0; | |
2814 | thisarginfo.op = NULL_TREE; | |
17b658af | 2815 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
2816 | |
2817 | op = gimple_call_arg (stmt, i); | |
81c40241 RB |
2818 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &thisarginfo.dt, |
2819 | &thisarginfo.vectype) | |
0136f8f0 AH |
2820 | || thisarginfo.dt == vect_uninitialized_def) |
2821 | { | |
2822 | if (dump_enabled_p ()) | |
2823 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2824 | "use not simple.\n"); | |
0136f8f0 AH |
2825 | return false; |
2826 | } | |
2827 | ||
2828 | if (thisarginfo.dt == vect_constant_def | |
2829 | || thisarginfo.dt == vect_external_def) | |
2830 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
2831 | else | |
2832 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
2833 | ||
6c9e85fb JJ |
2834 | /* For linear arguments, the analyze phase should have saved |
2835 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
2836 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
2837 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
2838 | { |
2839 | gcc_assert (vec_stmt); | |
2840 | thisarginfo.linear_step | |
17b658af | 2841 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 2842 | thisarginfo.op |
17b658af JJ |
2843 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
2844 | thisarginfo.simd_lane_linear | |
2845 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
2846 | == boolean_true_node); | |
6c9e85fb JJ |
2847 | /* If loop has been peeled for alignment, we need to adjust it. */ |
2848 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
2849 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 2850 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
2851 | { |
2852 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 2853 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
2854 | tree opt = TREE_TYPE (thisarginfo.op); |
2855 | bias = fold_convert (TREE_TYPE (step), bias); | |
2856 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
2857 | thisarginfo.op | |
2858 | = fold_build2 (POINTER_TYPE_P (opt) | |
2859 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
2860 | thisarginfo.op, bias); | |
2861 | } | |
2862 | } | |
2863 | else if (!vec_stmt | |
2864 | && thisarginfo.dt != vect_constant_def | |
2865 | && thisarginfo.dt != vect_external_def | |
2866 | && loop_vinfo | |
2867 | && TREE_CODE (op) == SSA_NAME | |
2868 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
2869 | &iv, false) | |
2870 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
2871 | { |
2872 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
2873 | thisarginfo.op = iv.base; | |
2874 | } | |
2875 | else if ((thisarginfo.dt == vect_constant_def | |
2876 | || thisarginfo.dt == vect_external_def) | |
2877 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
2878 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
2879 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
2880 | linear too. */ | |
2881 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
2882 | && !thisarginfo.linear_step | |
2883 | && !vec_stmt | |
2884 | && thisarginfo.dt != vect_constant_def | |
2885 | && thisarginfo.dt != vect_external_def | |
2886 | && loop_vinfo | |
2887 | && !slp_node | |
2888 | && TREE_CODE (op) == SSA_NAME) | |
2889 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
2890 | |
2891 | arginfo.quick_push (thisarginfo); | |
2892 | } | |
2893 | ||
2894 | unsigned int badness = 0; | |
2895 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
2896 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
2897 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
2898 | else |
2899 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
2900 | n = n->simdclone->next_clone) | |
2901 | { | |
2902 | unsigned int this_badness = 0; | |
2903 | if (n->simdclone->simdlen | |
2904 | > (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
2905 | || n->simdclone->nargs != nargs) | |
2906 | continue; | |
2907 | if (n->simdclone->simdlen | |
2908 | < (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2909 | this_badness += (exact_log2 (LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2910 | - exact_log2 (n->simdclone->simdlen)) * 1024; | |
2911 | if (n->simdclone->inbranch) | |
2912 | this_badness += 2048; | |
2913 | int target_badness = targetm.simd_clone.usable (n); | |
2914 | if (target_badness < 0) | |
2915 | continue; | |
2916 | this_badness += target_badness * 512; | |
2917 | /* FORNOW: Have to add code to add the mask argument. */ | |
2918 | if (n->simdclone->inbranch) | |
2919 | continue; | |
2920 | for (i = 0; i < nargs; i++) | |
2921 | { | |
2922 | switch (n->simdclone->args[i].arg_type) | |
2923 | { | |
2924 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
2925 | if (!useless_type_conversion_p | |
2926 | (n->simdclone->args[i].orig_type, | |
2927 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
2928 | i = -1; | |
2929 | else if (arginfo[i].dt == vect_constant_def | |
2930 | || arginfo[i].dt == vect_external_def | |
2931 | || arginfo[i].linear_step) | |
2932 | this_badness += 64; | |
2933 | break; | |
2934 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
2935 | if (arginfo[i].dt != vect_constant_def | |
2936 | && arginfo[i].dt != vect_external_def) | |
2937 | i = -1; | |
2938 | break; | |
2939 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 2940 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
2941 | if (arginfo[i].dt == vect_constant_def |
2942 | || arginfo[i].dt == vect_external_def | |
2943 | || (arginfo[i].linear_step | |
2944 | != n->simdclone->args[i].linear_step)) | |
2945 | i = -1; | |
2946 | break; | |
2947 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
2948 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
2949 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
2950 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
2951 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
2952 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
2953 | /* FORNOW */ |
2954 | i = -1; | |
2955 | break; | |
2956 | case SIMD_CLONE_ARG_TYPE_MASK: | |
2957 | gcc_unreachable (); | |
2958 | } | |
2959 | if (i == (size_t) -1) | |
2960 | break; | |
2961 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
2962 | { | |
2963 | i = -1; | |
2964 | break; | |
2965 | } | |
2966 | if (arginfo[i].align) | |
2967 | this_badness += (exact_log2 (arginfo[i].align) | |
2968 | - exact_log2 (n->simdclone->args[i].alignment)); | |
2969 | } | |
2970 | if (i == (size_t) -1) | |
2971 | continue; | |
2972 | if (bestn == NULL || this_badness < badness) | |
2973 | { | |
2974 | bestn = n; | |
2975 | badness = this_badness; | |
2976 | } | |
2977 | } | |
2978 | ||
2979 | if (bestn == NULL) | |
00426f9a | 2980 | return false; |
0136f8f0 AH |
2981 | |
2982 | for (i = 0; i < nargs; i++) | |
2983 | if ((arginfo[i].dt == vect_constant_def | |
2984 | || arginfo[i].dt == vect_external_def) | |
2985 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
2986 | { | |
2987 | arginfo[i].vectype | |
2988 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
2989 | i))); | |
2990 | if (arginfo[i].vectype == NULL | |
2991 | || (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
2992 | > bestn->simdclone->simdlen)) | |
00426f9a | 2993 | return false; |
0136f8f0 AH |
2994 | } |
2995 | ||
2996 | fndecl = bestn->decl; | |
2997 | nunits = bestn->simdclone->simdlen; | |
2998 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2999 | ||
3000 | /* If the function isn't const, only allow it in simd loops where user | |
3001 | has asserted that at least nunits consecutive iterations can be | |
3002 | performed using SIMD instructions. */ | |
3003 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
3004 | && gimple_vuse (stmt)) | |
00426f9a | 3005 | return false; |
0136f8f0 AH |
3006 | |
3007 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3008 | needs to be generated. */ | |
3009 | gcc_assert (ncopies >= 1); | |
3010 | ||
3011 | if (!vec_stmt) /* transformation not required. */ | |
3012 | { | |
6c9e85fb JJ |
3013 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
3014 | for (i = 0; i < nargs; i++) | |
7adb26f2 JJ |
3015 | if ((bestn->simdclone->args[i].arg_type |
3016 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
3017 | || (bestn->simdclone->args[i].arg_type | |
3018 | == SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP)) | |
6c9e85fb | 3019 | { |
17b658af | 3020 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
3021 | + 1); |
3022 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
3023 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
3024 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
3025 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
3026 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
3027 | tree sll = arginfo[i].simd_lane_linear |
3028 | ? boolean_true_node : boolean_false_node; | |
3029 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 3030 | } |
0136f8f0 AH |
3031 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
3032 | if (dump_enabled_p ()) | |
3033 | dump_printf_loc (MSG_NOTE, vect_location, | |
3034 | "=== vectorizable_simd_clone_call ===\n"); | |
3035 | /* vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); */ | |
0136f8f0 AH |
3036 | return true; |
3037 | } | |
3038 | ||
3039 | /** Transform. **/ | |
3040 | ||
3041 | if (dump_enabled_p ()) | |
3042 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
3043 | ||
3044 | /* Handle def. */ | |
3045 | scalar_dest = gimple_call_lhs (stmt); | |
3046 | vec_dest = NULL_TREE; | |
3047 | rtype = NULL_TREE; | |
3048 | ratype = NULL_TREE; | |
3049 | if (scalar_dest) | |
3050 | { | |
3051 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3052 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
3053 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
3054 | { | |
3055 | ratype = rtype; | |
3056 | rtype = TREE_TYPE (ratype); | |
3057 | } | |
3058 | } | |
3059 | ||
3060 | prev_stmt_info = NULL; | |
3061 | for (j = 0; j < ncopies; ++j) | |
3062 | { | |
3063 | /* Build argument list for the vectorized call. */ | |
3064 | if (j == 0) | |
3065 | vargs.create (nargs); | |
3066 | else | |
3067 | vargs.truncate (0); | |
3068 | ||
3069 | for (i = 0; i < nargs; i++) | |
3070 | { | |
3071 | unsigned int k, l, m, o; | |
3072 | tree atype; | |
3073 | op = gimple_call_arg (stmt, i); | |
3074 | switch (bestn->simdclone->args[i].arg_type) | |
3075 | { | |
3076 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3077 | atype = bestn->simdclone->args[i].vector_type; | |
3078 | o = nunits / TYPE_VECTOR_SUBPARTS (atype); | |
3079 | for (m = j * o; m < (j + 1) * o; m++) | |
3080 | { | |
3081 | if (TYPE_VECTOR_SUBPARTS (atype) | |
3082 | < TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)) | |
3083 | { | |
3084 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); | |
3085 | k = (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
3086 | / TYPE_VECTOR_SUBPARTS (atype)); | |
3087 | gcc_assert ((k & (k - 1)) == 0); | |
3088 | if (m == 0) | |
3089 | vec_oprnd0 | |
81c40241 | 3090 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3091 | else |
3092 | { | |
3093 | vec_oprnd0 = arginfo[i].op; | |
3094 | if ((m & (k - 1)) == 0) | |
3095 | vec_oprnd0 | |
3096 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3097 | vec_oprnd0); | |
3098 | } | |
3099 | arginfo[i].op = vec_oprnd0; | |
3100 | vec_oprnd0 | |
3101 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
3102 | size_int (prec), | |
3103 | bitsize_int ((m & (k - 1)) * prec)); | |
3104 | new_stmt | |
b731b390 | 3105 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3106 | vec_oprnd0); |
3107 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3108 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3109 | } | |
3110 | else | |
3111 | { | |
3112 | k = (TYPE_VECTOR_SUBPARTS (atype) | |
3113 | / TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)); | |
3114 | gcc_assert ((k & (k - 1)) == 0); | |
3115 | vec<constructor_elt, va_gc> *ctor_elts; | |
3116 | if (k != 1) | |
3117 | vec_alloc (ctor_elts, k); | |
3118 | else | |
3119 | ctor_elts = NULL; | |
3120 | for (l = 0; l < k; l++) | |
3121 | { | |
3122 | if (m == 0 && l == 0) | |
3123 | vec_oprnd0 | |
81c40241 | 3124 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3125 | else |
3126 | vec_oprnd0 | |
3127 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3128 | arginfo[i].op); | |
3129 | arginfo[i].op = vec_oprnd0; | |
3130 | if (k == 1) | |
3131 | break; | |
3132 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
3133 | vec_oprnd0); | |
3134 | } | |
3135 | if (k == 1) | |
3136 | vargs.safe_push (vec_oprnd0); | |
3137 | else | |
3138 | { | |
3139 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
3140 | new_stmt | |
b731b390 | 3141 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3142 | vec_oprnd0); |
3143 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3144 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3145 | } | |
3146 | } | |
3147 | } | |
3148 | break; | |
3149 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3150 | vargs.safe_push (op); | |
3151 | break; | |
3152 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
7adb26f2 | 3153 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
3154 | if (j == 0) |
3155 | { | |
3156 | gimple_seq stmts; | |
3157 | arginfo[i].op | |
3158 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
3159 | NULL_TREE); | |
3160 | if (stmts != NULL) | |
3161 | { | |
3162 | basic_block new_bb; | |
3163 | edge pe = loop_preheader_edge (loop); | |
3164 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
3165 | gcc_assert (!new_bb); | |
3166 | } | |
17b658af JJ |
3167 | if (arginfo[i].simd_lane_linear) |
3168 | { | |
3169 | vargs.safe_push (arginfo[i].op); | |
3170 | break; | |
3171 | } | |
b731b390 | 3172 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 3173 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
0136f8f0 | 3174 | set_vinfo_for_stmt (new_phi, |
310213d4 | 3175 | new_stmt_vec_info (new_phi, loop_vinfo)); |
0136f8f0 AH |
3176 | add_phi_arg (new_phi, arginfo[i].op, |
3177 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
3178 | enum tree_code code | |
3179 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3180 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3181 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3182 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3183 | widest_int cst |
3184 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3185 | ncopies * nunits); | |
3186 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3187 | tree phi_arg = copy_ssa_name (op); |
0d0e4a03 JJ |
3188 | new_stmt |
3189 | = gimple_build_assign (phi_arg, code, phi_res, tcst); | |
0136f8f0 AH |
3190 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
3191 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
3192 | set_vinfo_for_stmt (new_stmt, | |
310213d4 | 3193 | new_stmt_vec_info (new_stmt, loop_vinfo)); |
0136f8f0 AH |
3194 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
3195 | UNKNOWN_LOCATION); | |
3196 | arginfo[i].op = phi_res; | |
3197 | vargs.safe_push (phi_res); | |
3198 | } | |
3199 | else | |
3200 | { | |
3201 | enum tree_code code | |
3202 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3203 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3204 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3205 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3206 | widest_int cst |
3207 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3208 | j * nunits); | |
3209 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3210 | new_temp = make_ssa_name (TREE_TYPE (op)); |
0d0e4a03 JJ |
3211 | new_stmt = gimple_build_assign (new_temp, code, |
3212 | arginfo[i].op, tcst); | |
0136f8f0 AH |
3213 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3214 | vargs.safe_push (new_temp); | |
3215 | } | |
3216 | break; | |
7adb26f2 JJ |
3217 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
3218 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
0136f8f0 | 3219 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
e01d41e5 JJ |
3220 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3221 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3222 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3223 | default: |
3224 | gcc_unreachable (); | |
3225 | } | |
3226 | } | |
3227 | ||
3228 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
3229 | if (vec_dest) | |
3230 | { | |
3231 | gcc_assert (ratype || TYPE_VECTOR_SUBPARTS (rtype) == nunits); | |
3232 | if (ratype) | |
b731b390 | 3233 | new_temp = create_tmp_var (ratype); |
0136f8f0 AH |
3234 | else if (TYPE_VECTOR_SUBPARTS (vectype) |
3235 | == TYPE_VECTOR_SUBPARTS (rtype)) | |
3236 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3237 | else | |
3238 | new_temp = make_ssa_name (rtype, new_stmt); | |
3239 | gimple_call_set_lhs (new_stmt, new_temp); | |
3240 | } | |
3241 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3242 | ||
3243 | if (vec_dest) | |
3244 | { | |
3245 | if (TYPE_VECTOR_SUBPARTS (vectype) < nunits) | |
3246 | { | |
3247 | unsigned int k, l; | |
3248 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); | |
3249 | k = nunits / TYPE_VECTOR_SUBPARTS (vectype); | |
3250 | gcc_assert ((k & (k - 1)) == 0); | |
3251 | for (l = 0; l < k; l++) | |
3252 | { | |
3253 | tree t; | |
3254 | if (ratype) | |
3255 | { | |
3256 | t = build_fold_addr_expr (new_temp); | |
3257 | t = build2 (MEM_REF, vectype, t, | |
3258 | build_int_cst (TREE_TYPE (t), | |
3259 | l * prec / BITS_PER_UNIT)); | |
3260 | } | |
3261 | else | |
3262 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
3263 | size_int (prec), bitsize_int (l * prec)); | |
3264 | new_stmt | |
b731b390 | 3265 | = gimple_build_assign (make_ssa_name (vectype), t); |
0136f8f0 AH |
3266 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3267 | if (j == 0 && l == 0) | |
3268 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3269 | else | |
3270 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3271 | ||
3272 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3273 | } | |
3274 | ||
3275 | if (ratype) | |
3276 | { | |
3277 | tree clobber = build_constructor (ratype, NULL); | |
3278 | TREE_THIS_VOLATILE (clobber) = 1; | |
3279 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3280 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3281 | } | |
3282 | continue; | |
3283 | } | |
3284 | else if (TYPE_VECTOR_SUBPARTS (vectype) > nunits) | |
3285 | { | |
3286 | unsigned int k = (TYPE_VECTOR_SUBPARTS (vectype) | |
3287 | / TYPE_VECTOR_SUBPARTS (rtype)); | |
3288 | gcc_assert ((k & (k - 1)) == 0); | |
3289 | if ((j & (k - 1)) == 0) | |
3290 | vec_alloc (ret_ctor_elts, k); | |
3291 | if (ratype) | |
3292 | { | |
3293 | unsigned int m, o = nunits / TYPE_VECTOR_SUBPARTS (rtype); | |
3294 | for (m = 0; m < o; m++) | |
3295 | { | |
3296 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
3297 | size_int (m), NULL_TREE, NULL_TREE); | |
3298 | new_stmt | |
b731b390 | 3299 | = gimple_build_assign (make_ssa_name (rtype), tem); |
0136f8f0 AH |
3300 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3301 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
3302 | gimple_assign_lhs (new_stmt)); | |
3303 | } | |
3304 | tree clobber = build_constructor (ratype, NULL); | |
3305 | TREE_THIS_VOLATILE (clobber) = 1; | |
3306 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3307 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3308 | } | |
3309 | else | |
3310 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
3311 | if ((j & (k - 1)) != k - 1) | |
3312 | continue; | |
3313 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
3314 | new_stmt | |
b731b390 | 3315 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
0136f8f0 AH |
3316 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3317 | ||
3318 | if ((unsigned) j == k - 1) | |
3319 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3320 | else | |
3321 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3322 | ||
3323 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3324 | continue; | |
3325 | } | |
3326 | else if (ratype) | |
3327 | { | |
3328 | tree t = build_fold_addr_expr (new_temp); | |
3329 | t = build2 (MEM_REF, vectype, t, | |
3330 | build_int_cst (TREE_TYPE (t), 0)); | |
3331 | new_stmt | |
b731b390 | 3332 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
0136f8f0 AH |
3333 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3334 | tree clobber = build_constructor (ratype, NULL); | |
3335 | TREE_THIS_VOLATILE (clobber) = 1; | |
3336 | vect_finish_stmt_generation (stmt, | |
3337 | gimple_build_assign (new_temp, | |
3338 | clobber), gsi); | |
3339 | } | |
3340 | } | |
3341 | ||
3342 | if (j == 0) | |
3343 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3344 | else | |
3345 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3346 | ||
3347 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3348 | } | |
3349 | ||
3350 | vargs.release (); | |
3351 | ||
3352 | /* The call in STMT might prevent it from being removed in dce. | |
3353 | We however cannot remove it here, due to the way the ssa name | |
3354 | it defines is mapped to the new definition. So just replace | |
3355 | rhs of the statement with something harmless. */ | |
3356 | ||
3357 | if (slp_node) | |
3358 | return true; | |
3359 | ||
3360 | if (scalar_dest) | |
3361 | { | |
3362 | type = TREE_TYPE (scalar_dest); | |
3363 | if (is_pattern_stmt_p (stmt_info)) | |
3364 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3365 | else | |
3366 | lhs = gimple_call_lhs (stmt); | |
3367 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
3368 | } | |
3369 | else | |
3370 | new_stmt = gimple_build_nop (); | |
3371 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
3372 | set_vinfo_for_stmt (stmt, NULL); | |
3373 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 3374 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
3375 | unlink_stmt_vdef (stmt); |
3376 | ||
3377 | return true; | |
3378 | } | |
3379 | ||
3380 | ||
ebfd146a IR |
3381 | /* Function vect_gen_widened_results_half |
3382 | ||
3383 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 3384 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 3385 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
3386 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
3387 | needs to be created (DECL is a function-decl of a target-builtin). | |
3388 | STMT is the original scalar stmt that we are vectorizing. */ | |
3389 | ||
355fe088 | 3390 | static gimple * |
ebfd146a IR |
3391 | vect_gen_widened_results_half (enum tree_code code, |
3392 | tree decl, | |
3393 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
3394 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 3395 | gimple *stmt) |
b8698a0f | 3396 | { |
355fe088 | 3397 | gimple *new_stmt; |
b8698a0f L |
3398 | tree new_temp; |
3399 | ||
3400 | /* Generate half of the widened result: */ | |
3401 | if (code == CALL_EXPR) | |
3402 | { | |
3403 | /* Target specific support */ | |
ebfd146a IR |
3404 | if (op_type == binary_op) |
3405 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
3406 | else | |
3407 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
3408 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3409 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
3410 | } |
3411 | else | |
ebfd146a | 3412 | { |
b8698a0f L |
3413 | /* Generic support */ |
3414 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
3415 | if (op_type != binary_op) |
3416 | vec_oprnd1 = NULL; | |
0d0e4a03 | 3417 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
3418 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3419 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 3420 | } |
ebfd146a IR |
3421 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3422 | ||
ebfd146a IR |
3423 | return new_stmt; |
3424 | } | |
3425 | ||
4a00c761 JJ |
3426 | |
3427 | /* Get vectorized definitions for loop-based vectorization. For the first | |
3428 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
3429 | scalar operand), and for the rest we get a copy with | |
3430 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
3431 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
3432 | The vectors are collected into VEC_OPRNDS. */ | |
3433 | ||
3434 | static void | |
355fe088 | 3435 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 3436 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
3437 | { |
3438 | tree vec_oprnd; | |
3439 | ||
3440 | /* Get first vector operand. */ | |
3441 | /* All the vector operands except the very first one (that is scalar oprnd) | |
3442 | are stmt copies. */ | |
3443 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 3444 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
3445 | else |
3446 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
3447 | ||
9771b263 | 3448 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3449 | |
3450 | /* Get second vector operand. */ | |
3451 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 3452 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3453 | |
3454 | *oprnd = vec_oprnd; | |
3455 | ||
3456 | /* For conversion in multiple steps, continue to get operands | |
3457 | recursively. */ | |
3458 | if (multi_step_cvt) | |
3459 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
3460 | } | |
3461 | ||
3462 | ||
3463 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
3464 | For multi-step conversions store the resulting vectors and call the function | |
3465 | recursively. */ | |
3466 | ||
3467 | static void | |
9771b263 | 3468 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 3469 | int multi_step_cvt, gimple *stmt, |
9771b263 | 3470 | vec<tree> vec_dsts, |
4a00c761 JJ |
3471 | gimple_stmt_iterator *gsi, |
3472 | slp_tree slp_node, enum tree_code code, | |
3473 | stmt_vec_info *prev_stmt_info) | |
3474 | { | |
3475 | unsigned int i; | |
3476 | tree vop0, vop1, new_tmp, vec_dest; | |
355fe088 | 3477 | gimple *new_stmt; |
4a00c761 JJ |
3478 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
3479 | ||
9771b263 | 3480 | vec_dest = vec_dsts.pop (); |
4a00c761 | 3481 | |
9771b263 | 3482 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
3483 | { |
3484 | /* Create demotion operation. */ | |
9771b263 DN |
3485 | vop0 = (*vec_oprnds)[i]; |
3486 | vop1 = (*vec_oprnds)[i + 1]; | |
0d0e4a03 | 3487 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
3488 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
3489 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
3490 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3491 | ||
3492 | if (multi_step_cvt) | |
3493 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 3494 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
3495 | else |
3496 | { | |
3497 | /* This is the last step of the conversion sequence. Store the | |
3498 | vectors in SLP_NODE or in vector info of the scalar statement | |
3499 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
3500 | if (slp_node) | |
9771b263 | 3501 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 3502 | else |
c689ce1e RB |
3503 | { |
3504 | if (!*prev_stmt_info) | |
3505 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3506 | else | |
3507 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
4a00c761 | 3508 | |
c689ce1e RB |
3509 | *prev_stmt_info = vinfo_for_stmt (new_stmt); |
3510 | } | |
4a00c761 JJ |
3511 | } |
3512 | } | |
3513 | ||
3514 | /* For multi-step demotion operations we first generate demotion operations | |
3515 | from the source type to the intermediate types, and then combine the | |
3516 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
3517 | type. */ | |
3518 | if (multi_step_cvt) | |
3519 | { | |
3520 | /* At each level of recursion we have half of the operands we had at the | |
3521 | previous level. */ | |
9771b263 | 3522 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
3523 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
3524 | stmt, vec_dsts, gsi, slp_node, | |
3525 | VEC_PACK_TRUNC_EXPR, | |
3526 | prev_stmt_info); | |
3527 | } | |
3528 | ||
9771b263 | 3529 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3530 | } |
3531 | ||
3532 | ||
3533 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
3534 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
3535 | the resulting vectors and call the function recursively. */ | |
3536 | ||
3537 | static void | |
9771b263 DN |
3538 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
3539 | vec<tree> *vec_oprnds1, | |
355fe088 | 3540 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
3541 | gimple_stmt_iterator *gsi, |
3542 | enum tree_code code1, | |
3543 | enum tree_code code2, tree decl1, | |
3544 | tree decl2, int op_type) | |
3545 | { | |
3546 | int i; | |
3547 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 3548 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 3549 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 3550 | |
9771b263 DN |
3551 | vec_tmp.create (vec_oprnds0->length () * 2); |
3552 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
3553 | { |
3554 | if (op_type == binary_op) | |
9771b263 | 3555 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
3556 | else |
3557 | vop1 = NULL_TREE; | |
3558 | ||
3559 | /* Generate the two halves of promotion operation. */ | |
3560 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
3561 | op_type, vec_dest, gsi, stmt); | |
3562 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
3563 | op_type, vec_dest, gsi, stmt); | |
3564 | if (is_gimple_call (new_stmt1)) | |
3565 | { | |
3566 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
3567 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
3568 | } | |
3569 | else | |
3570 | { | |
3571 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
3572 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
3573 | } | |
3574 | ||
3575 | /* Store the results for the next step. */ | |
9771b263 DN |
3576 | vec_tmp.quick_push (new_tmp1); |
3577 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
3578 | } |
3579 | ||
689eaba3 | 3580 | vec_oprnds0->release (); |
4a00c761 JJ |
3581 | *vec_oprnds0 = vec_tmp; |
3582 | } | |
3583 | ||
3584 | ||
b8698a0f L |
3585 | /* Check if STMT performs a conversion operation, that can be vectorized. |
3586 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 3587 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
3588 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
3589 | ||
3590 | static bool | |
355fe088 TS |
3591 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
3592 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
3593 | { |
3594 | tree vec_dest; | |
3595 | tree scalar_dest; | |
4a00c761 | 3596 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
3597 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
3598 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3599 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3600 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 3601 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
3602 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
3603 | tree new_temp; | |
355fe088 | 3604 | gimple *def_stmt; |
ebfd146a | 3605 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
355fe088 | 3606 | gimple *new_stmt = NULL; |
ebfd146a IR |
3607 | stmt_vec_info prev_stmt_info; |
3608 | int nunits_in; | |
3609 | int nunits_out; | |
3610 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
3611 | int ncopies, i, j; |
3612 | tree lhs_type, rhs_type; | |
ebfd146a | 3613 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
3614 | vec<tree> vec_oprnds0 = vNULL; |
3615 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 3616 | tree vop0; |
4a00c761 | 3617 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 3618 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 3619 | int multi_step_cvt = 0; |
6e1aa848 DN |
3620 | vec<tree> vec_dsts = vNULL; |
3621 | vec<tree> interm_types = vNULL; | |
4a00c761 JJ |
3622 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
3623 | int op_type; | |
ef4bddc2 | 3624 | machine_mode rhs_mode; |
4a00c761 | 3625 | unsigned short fltsz; |
ebfd146a IR |
3626 | |
3627 | /* Is STMT a vectorizable conversion? */ | |
3628 | ||
4a00c761 | 3629 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3630 | return false; |
3631 | ||
66c16fd9 RB |
3632 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3633 | && ! vec_stmt) | |
ebfd146a IR |
3634 | return false; |
3635 | ||
3636 | if (!is_gimple_assign (stmt)) | |
3637 | return false; | |
3638 | ||
3639 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3640 | return false; | |
3641 | ||
3642 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
3643 | if (!CONVERT_EXPR_CODE_P (code) |
3644 | && code != FIX_TRUNC_EXPR | |
3645 | && code != FLOAT_EXPR | |
3646 | && code != WIDEN_MULT_EXPR | |
3647 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
3648 | return false; |
3649 | ||
4a00c761 JJ |
3650 | op_type = TREE_CODE_LENGTH (code); |
3651 | ||
ebfd146a | 3652 | /* Check types of lhs and rhs. */ |
b690cc0f | 3653 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 3654 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
3655 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3656 | ||
ebfd146a IR |
3657 | op0 = gimple_assign_rhs1 (stmt); |
3658 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
3659 | |
3660 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3661 | && !((INTEGRAL_TYPE_P (lhs_type) | |
3662 | && INTEGRAL_TYPE_P (rhs_type)) | |
3663 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
3664 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
3665 | return false; | |
3666 | ||
e6f5c25d IE |
3667 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
3668 | && ((INTEGRAL_TYPE_P (lhs_type) | |
3669 | && (TYPE_PRECISION (lhs_type) | |
3670 | != GET_MODE_PRECISION (TYPE_MODE (lhs_type)))) | |
3671 | || (INTEGRAL_TYPE_P (rhs_type) | |
3672 | && (TYPE_PRECISION (rhs_type) | |
3673 | != GET_MODE_PRECISION (TYPE_MODE (rhs_type)))))) | |
4a00c761 | 3674 | { |
73fbfcad | 3675 | if (dump_enabled_p ()) |
78c60e3d | 3676 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
3677 | "type conversion to/from bit-precision unsupported." |
3678 | "\n"); | |
4a00c761 JJ |
3679 | return false; |
3680 | } | |
3681 | ||
b690cc0f | 3682 | /* Check the operands of the operation. */ |
81c40241 | 3683 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype_in)) |
b690cc0f | 3684 | { |
73fbfcad | 3685 | if (dump_enabled_p ()) |
78c60e3d | 3686 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3687 | "use not simple.\n"); |
b690cc0f RG |
3688 | return false; |
3689 | } | |
4a00c761 JJ |
3690 | if (op_type == binary_op) |
3691 | { | |
3692 | bool ok; | |
3693 | ||
3694 | op1 = gimple_assign_rhs2 (stmt); | |
3695 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
3696 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
3697 | OP1. */ | |
3698 | if (CONSTANT_CLASS_P (op0)) | |
81c40241 | 3699 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &vectype_in); |
4a00c761 | 3700 | else |
81c40241 | 3701 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1]); |
4a00c761 JJ |
3702 | |
3703 | if (!ok) | |
3704 | { | |
73fbfcad | 3705 | if (dump_enabled_p ()) |
78c60e3d | 3706 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3707 | "use not simple.\n"); |
4a00c761 JJ |
3708 | return false; |
3709 | } | |
3710 | } | |
3711 | ||
b690cc0f RG |
3712 | /* If op0 is an external or constant defs use a vector type of |
3713 | the same size as the output vector type. */ | |
ebfd146a | 3714 | if (!vectype_in) |
b690cc0f | 3715 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
3716 | if (vec_stmt) |
3717 | gcc_assert (vectype_in); | |
3718 | if (!vectype_in) | |
3719 | { | |
73fbfcad | 3720 | if (dump_enabled_p ()) |
4a00c761 | 3721 | { |
78c60e3d SS |
3722 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3723 | "no vectype for scalar type "); | |
3724 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 3725 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 3726 | } |
7d8930a0 IR |
3727 | |
3728 | return false; | |
3729 | } | |
ebfd146a | 3730 | |
e6f5c25d IE |
3731 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
3732 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
3733 | { | |
3734 | if (dump_enabled_p ()) | |
3735 | { | |
3736 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3737 | "can't convert between boolean and non " | |
3738 | "boolean vectors"); | |
3739 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
3740 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
3741 | } | |
3742 | ||
3743 | return false; | |
3744 | } | |
3745 | ||
b690cc0f RG |
3746 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
3747 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 3748 | if (nunits_in < nunits_out) |
ebfd146a IR |
3749 | modifier = NARROW; |
3750 | else if (nunits_out == nunits_in) | |
3751 | modifier = NONE; | |
ebfd146a | 3752 | else |
4a00c761 | 3753 | modifier = WIDEN; |
ebfd146a | 3754 | |
ff802fa1 IR |
3755 | /* Multiple types in SLP are handled by creating the appropriate number of |
3756 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3757 | case of SLP. */ | |
fce57248 | 3758 | if (slp_node) |
ebfd146a | 3759 | ncopies = 1; |
4a00c761 JJ |
3760 | else if (modifier == NARROW) |
3761 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
3762 | else | |
3763 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 3764 | |
ebfd146a IR |
3765 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
3766 | needs to be generated. */ | |
3767 | gcc_assert (ncopies >= 1); | |
3768 | ||
ebfd146a | 3769 | /* Supportable by target? */ |
4a00c761 | 3770 | switch (modifier) |
ebfd146a | 3771 | { |
4a00c761 JJ |
3772 | case NONE: |
3773 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3774 | return false; | |
3775 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
3776 | &decl1, &code1)) | |
3777 | break; | |
3778 | /* FALLTHRU */ | |
3779 | unsupported: | |
73fbfcad | 3780 | if (dump_enabled_p ()) |
78c60e3d | 3781 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3782 | "conversion not supported by target.\n"); |
ebfd146a | 3783 | return false; |
ebfd146a | 3784 | |
4a00c761 JJ |
3785 | case WIDEN: |
3786 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
3787 | &code1, &code2, &multi_step_cvt, |
3788 | &interm_types)) | |
4a00c761 JJ |
3789 | { |
3790 | /* Binary widening operation can only be supported directly by the | |
3791 | architecture. */ | |
3792 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
3793 | break; | |
3794 | } | |
3795 | ||
3796 | if (code != FLOAT_EXPR | |
3797 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3798 | <= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3799 | goto unsupported; | |
3800 | ||
3801 | rhs_mode = TYPE_MODE (rhs_type); | |
3802 | fltsz = GET_MODE_SIZE (TYPE_MODE (lhs_type)); | |
3803 | for (rhs_mode = GET_MODE_2XWIDER_MODE (TYPE_MODE (rhs_type)); | |
3804 | rhs_mode != VOIDmode && GET_MODE_SIZE (rhs_mode) <= fltsz; | |
3805 | rhs_mode = GET_MODE_2XWIDER_MODE (rhs_mode)) | |
3806 | { | |
3807 | cvt_type | |
3808 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3809 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3810 | if (cvt_type == NULL_TREE) | |
3811 | goto unsupported; | |
3812 | ||
3813 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3814 | { | |
3815 | if (!supportable_convert_operation (code, vectype_out, | |
3816 | cvt_type, &decl1, &codecvt1)) | |
3817 | goto unsupported; | |
3818 | } | |
3819 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
3820 | cvt_type, &codecvt1, |
3821 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
3822 | &interm_types)) |
3823 | continue; | |
3824 | else | |
3825 | gcc_assert (multi_step_cvt == 0); | |
3826 | ||
3827 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
3828 | vectype_in, &code1, &code2, |
3829 | &multi_step_cvt, &interm_types)) | |
4a00c761 JJ |
3830 | break; |
3831 | } | |
3832 | ||
3833 | if (rhs_mode == VOIDmode || GET_MODE_SIZE (rhs_mode) > fltsz) | |
3834 | goto unsupported; | |
3835 | ||
3836 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3837 | codecvt2 = ERROR_MARK; | |
3838 | else | |
3839 | { | |
3840 | multi_step_cvt++; | |
9771b263 | 3841 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
3842 | cvt_type = NULL_TREE; |
3843 | } | |
3844 | break; | |
3845 | ||
3846 | case NARROW: | |
3847 | gcc_assert (op_type == unary_op); | |
3848 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
3849 | &code1, &multi_step_cvt, | |
3850 | &interm_types)) | |
3851 | break; | |
3852 | ||
3853 | if (code != FIX_TRUNC_EXPR | |
3854 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3855 | >= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3856 | goto unsupported; | |
3857 | ||
3858 | rhs_mode = TYPE_MODE (rhs_type); | |
3859 | cvt_type | |
3860 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3861 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3862 | if (cvt_type == NULL_TREE) | |
3863 | goto unsupported; | |
3864 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
3865 | &decl1, &codecvt1)) | |
3866 | goto unsupported; | |
3867 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
3868 | &code1, &multi_step_cvt, | |
3869 | &interm_types)) | |
3870 | break; | |
3871 | goto unsupported; | |
3872 | ||
3873 | default: | |
3874 | gcc_unreachable (); | |
ebfd146a IR |
3875 | } |
3876 | ||
3877 | if (!vec_stmt) /* transformation not required. */ | |
3878 | { | |
73fbfcad | 3879 | if (dump_enabled_p ()) |
78c60e3d | 3880 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3881 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 3882 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
3883 | { |
3884 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
c3e7ee41 | 3885 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
8bd37302 | 3886 | } |
4a00c761 JJ |
3887 | else if (modifier == NARROW) |
3888 | { | |
3889 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 3890 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
3891 | } |
3892 | else | |
3893 | { | |
3894 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 3895 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 3896 | } |
9771b263 | 3897 | interm_types.release (); |
ebfd146a IR |
3898 | return true; |
3899 | } | |
3900 | ||
3901 | /** Transform. **/ | |
73fbfcad | 3902 | if (dump_enabled_p ()) |
78c60e3d | 3903 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3904 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 3905 | |
4a00c761 JJ |
3906 | if (op_type == binary_op) |
3907 | { | |
3908 | if (CONSTANT_CLASS_P (op0)) | |
3909 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
3910 | else if (CONSTANT_CLASS_P (op1)) | |
3911 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
3912 | } | |
3913 | ||
3914 | /* In case of multi-step conversion, we first generate conversion operations | |
3915 | to the intermediate types, and then from that types to the final one. | |
3916 | We create vector destinations for the intermediate type (TYPES) received | |
3917 | from supportable_*_operation, and store them in the correct order | |
3918 | for future use in vect_create_vectorized_*_stmts (). */ | |
9771b263 | 3919 | vec_dsts.create (multi_step_cvt + 1); |
82294ec1 JJ |
3920 | vec_dest = vect_create_destination_var (scalar_dest, |
3921 | (cvt_type && modifier == WIDEN) | |
3922 | ? cvt_type : vectype_out); | |
9771b263 | 3923 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3924 | |
3925 | if (multi_step_cvt) | |
3926 | { | |
9771b263 DN |
3927 | for (i = interm_types.length () - 1; |
3928 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
3929 | { |
3930 | vec_dest = vect_create_destination_var (scalar_dest, | |
3931 | intermediate_type); | |
9771b263 | 3932 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3933 | } |
3934 | } | |
ebfd146a | 3935 | |
4a00c761 | 3936 | if (cvt_type) |
82294ec1 JJ |
3937 | vec_dest = vect_create_destination_var (scalar_dest, |
3938 | modifier == WIDEN | |
3939 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
3940 | |
3941 | if (!slp_node) | |
3942 | { | |
30862efc | 3943 | if (modifier == WIDEN) |
4a00c761 | 3944 | { |
c3284718 | 3945 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 3946 | if (op_type == binary_op) |
9771b263 | 3947 | vec_oprnds1.create (1); |
4a00c761 | 3948 | } |
30862efc | 3949 | else if (modifier == NARROW) |
9771b263 DN |
3950 | vec_oprnds0.create ( |
3951 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
3952 | } |
3953 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 3954 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 3955 | |
4a00c761 | 3956 | last_oprnd = op0; |
ebfd146a IR |
3957 | prev_stmt_info = NULL; |
3958 | switch (modifier) | |
3959 | { | |
3960 | case NONE: | |
3961 | for (j = 0; j < ncopies; j++) | |
3962 | { | |
ebfd146a | 3963 | if (j == 0) |
d092494c IR |
3964 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node, |
3965 | -1); | |
ebfd146a IR |
3966 | else |
3967 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
3968 | ||
9771b263 | 3969 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
3970 | { |
3971 | /* Arguments are ready, create the new vector stmt. */ | |
3972 | if (code1 == CALL_EXPR) | |
3973 | { | |
3974 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
3975 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3976 | gimple_call_set_lhs (new_stmt, new_temp); | |
3977 | } | |
3978 | else | |
3979 | { | |
3980 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
0d0e4a03 | 3981 | new_stmt = gimple_build_assign (vec_dest, code1, vop0); |
4a00c761 JJ |
3982 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3983 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3984 | } | |
3985 | ||
3986 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3987 | if (slp_node) | |
9771b263 | 3988 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
225ce44b RB |
3989 | else |
3990 | { | |
3991 | if (!prev_stmt_info) | |
3992 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3993 | else | |
3994 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3995 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3996 | } | |
4a00c761 | 3997 | } |
ebfd146a IR |
3998 | } |
3999 | break; | |
4000 | ||
4001 | case WIDEN: | |
4002 | /* In case the vectorization factor (VF) is bigger than the number | |
4003 | of elements that we can fit in a vectype (nunits), we have to | |
4004 | generate more than one vector stmt - i.e - we need to "unroll" | |
4005 | the vector stmt by a factor VF/nunits. */ | |
4006 | for (j = 0; j < ncopies; j++) | |
4007 | { | |
4a00c761 | 4008 | /* Handle uses. */ |
ebfd146a | 4009 | if (j == 0) |
4a00c761 JJ |
4010 | { |
4011 | if (slp_node) | |
4012 | { | |
4013 | if (code == WIDEN_LSHIFT_EXPR) | |
4014 | { | |
4015 | unsigned int k; | |
ebfd146a | 4016 | |
4a00c761 JJ |
4017 | vec_oprnd1 = op1; |
4018 | /* Store vec_oprnd1 for every vector stmt to be created | |
4019 | for SLP_NODE. We check during the analysis that all | |
4020 | the shift arguments are the same. */ | |
4021 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4022 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4023 | |
4024 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
4025 | slp_node, -1); | |
4026 | } | |
4027 | else | |
4028 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
4029 | &vec_oprnds1, slp_node, -1); | |
4030 | } | |
4031 | else | |
4032 | { | |
81c40241 | 4033 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 4034 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
4035 | if (op_type == binary_op) |
4036 | { | |
4037 | if (code == WIDEN_LSHIFT_EXPR) | |
4038 | vec_oprnd1 = op1; | |
4039 | else | |
81c40241 | 4040 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 4041 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4042 | } |
4043 | } | |
4044 | } | |
ebfd146a | 4045 | else |
4a00c761 JJ |
4046 | { |
4047 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
4048 | vec_oprnds0.truncate (0); |
4049 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
4050 | if (op_type == binary_op) |
4051 | { | |
4052 | if (code == WIDEN_LSHIFT_EXPR) | |
4053 | vec_oprnd1 = op1; | |
4054 | else | |
4055 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
4056 | vec_oprnd1); | |
9771b263 DN |
4057 | vec_oprnds1.truncate (0); |
4058 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
4059 | } |
4060 | } | |
ebfd146a | 4061 | |
4a00c761 JJ |
4062 | /* Arguments are ready. Create the new vector stmts. */ |
4063 | for (i = multi_step_cvt; i >= 0; i--) | |
4064 | { | |
9771b263 | 4065 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
4066 | enum tree_code c1 = code1, c2 = code2; |
4067 | if (i == 0 && codecvt2 != ERROR_MARK) | |
4068 | { | |
4069 | c1 = codecvt1; | |
4070 | c2 = codecvt2; | |
4071 | } | |
4072 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
4073 | &vec_oprnds1, | |
4074 | stmt, this_dest, gsi, | |
4075 | c1, c2, decl1, decl2, | |
4076 | op_type); | |
4077 | } | |
4078 | ||
9771b263 | 4079 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4080 | { |
4081 | if (cvt_type) | |
4082 | { | |
4083 | if (codecvt1 == CALL_EXPR) | |
4084 | { | |
4085 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4086 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4087 | gimple_call_set_lhs (new_stmt, new_temp); | |
4088 | } | |
4089 | else | |
4090 | { | |
4091 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4092 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4093 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4094 | vop0); | |
4a00c761 JJ |
4095 | } |
4096 | ||
4097 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4098 | } | |
4099 | else | |
4100 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
4101 | ||
4102 | if (slp_node) | |
9771b263 | 4103 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4104 | else |
c689ce1e RB |
4105 | { |
4106 | if (!prev_stmt_info) | |
4107 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4108 | else | |
4109 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4110 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4111 | } | |
4a00c761 | 4112 | } |
ebfd146a | 4113 | } |
4a00c761 JJ |
4114 | |
4115 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
4116 | break; |
4117 | ||
4118 | case NARROW: | |
4119 | /* In case the vectorization factor (VF) is bigger than the number | |
4120 | of elements that we can fit in a vectype (nunits), we have to | |
4121 | generate more than one vector stmt - i.e - we need to "unroll" | |
4122 | the vector stmt by a factor VF/nunits. */ | |
4123 | for (j = 0; j < ncopies; j++) | |
4124 | { | |
4125 | /* Handle uses. */ | |
4a00c761 JJ |
4126 | if (slp_node) |
4127 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
4128 | slp_node, -1); | |
ebfd146a IR |
4129 | else |
4130 | { | |
9771b263 | 4131 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
4132 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
4133 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
4134 | } |
4135 | ||
4a00c761 JJ |
4136 | /* Arguments are ready. Create the new vector stmts. */ |
4137 | if (cvt_type) | |
9771b263 | 4138 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4139 | { |
4140 | if (codecvt1 == CALL_EXPR) | |
4141 | { | |
4142 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4143 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4144 | gimple_call_set_lhs (new_stmt, new_temp); | |
4145 | } | |
4146 | else | |
4147 | { | |
4148 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4149 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4150 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4151 | vop0); | |
4a00c761 | 4152 | } |
ebfd146a | 4153 | |
4a00c761 | 4154 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 4155 | vec_oprnds0[i] = new_temp; |
4a00c761 | 4156 | } |
ebfd146a | 4157 | |
4a00c761 JJ |
4158 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
4159 | stmt, vec_dsts, gsi, | |
4160 | slp_node, code1, | |
4161 | &prev_stmt_info); | |
ebfd146a IR |
4162 | } |
4163 | ||
4164 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 4165 | break; |
ebfd146a IR |
4166 | } |
4167 | ||
9771b263 DN |
4168 | vec_oprnds0.release (); |
4169 | vec_oprnds1.release (); | |
4170 | vec_dsts.release (); | |
4171 | interm_types.release (); | |
ebfd146a IR |
4172 | |
4173 | return true; | |
4174 | } | |
ff802fa1 IR |
4175 | |
4176 | ||
ebfd146a IR |
4177 | /* Function vectorizable_assignment. |
4178 | ||
b8698a0f L |
4179 | Check if STMT performs an assignment (copy) that can be vectorized. |
4180 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4181 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4182 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4183 | ||
4184 | static bool | |
355fe088 TS |
4185 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
4186 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4187 | { |
4188 | tree vec_dest; | |
4189 | tree scalar_dest; | |
4190 | tree op; | |
4191 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
4192 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4193 | tree new_temp; | |
355fe088 | 4194 | gimple *def_stmt; |
ebfd146a | 4195 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
ebfd146a | 4196 | int ncopies; |
f18b55bd | 4197 | int i, j; |
6e1aa848 | 4198 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 4199 | tree vop; |
a70d6342 | 4200 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4201 | vec_info *vinfo = stmt_info->vinfo; |
355fe088 | 4202 | gimple *new_stmt = NULL; |
f18b55bd | 4203 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
4204 | enum tree_code code; |
4205 | tree vectype_in; | |
ebfd146a | 4206 | |
a70d6342 | 4207 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4208 | return false; |
4209 | ||
66c16fd9 RB |
4210 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4211 | && ! vec_stmt) | |
ebfd146a IR |
4212 | return false; |
4213 | ||
4214 | /* Is vectorizable assignment? */ | |
4215 | if (!is_gimple_assign (stmt)) | |
4216 | return false; | |
4217 | ||
4218 | scalar_dest = gimple_assign_lhs (stmt); | |
4219 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4220 | return false; | |
4221 | ||
fde9c428 | 4222 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 4223 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
4224 | || code == PAREN_EXPR |
4225 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
4226 | op = gimple_assign_rhs1 (stmt); |
4227 | else | |
4228 | return false; | |
4229 | ||
7b7ec6c5 RG |
4230 | if (code == VIEW_CONVERT_EXPR) |
4231 | op = TREE_OPERAND (op, 0); | |
4232 | ||
465c8c19 JJ |
4233 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
4234 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4235 | ||
4236 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4237 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4238 | case of SLP. */ | |
fce57248 | 4239 | if (slp_node) |
465c8c19 JJ |
4240 | ncopies = 1; |
4241 | else | |
4242 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4243 | ||
4244 | gcc_assert (ncopies >= 1); | |
4245 | ||
81c40241 | 4246 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[0], &vectype_in)) |
ebfd146a | 4247 | { |
73fbfcad | 4248 | if (dump_enabled_p ()) |
78c60e3d | 4249 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4250 | "use not simple.\n"); |
ebfd146a IR |
4251 | return false; |
4252 | } | |
4253 | ||
fde9c428 RG |
4254 | /* We can handle NOP_EXPR conversions that do not change the number |
4255 | of elements or the vector size. */ | |
7b7ec6c5 RG |
4256 | if ((CONVERT_EXPR_CODE_P (code) |
4257 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
4258 | && (!vectype_in |
4259 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
4260 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
4261 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
4262 | return false; | |
4263 | ||
7b7b1813 RG |
4264 | /* We do not handle bit-precision changes. */ |
4265 | if ((CONVERT_EXPR_CODE_P (code) | |
4266 | || code == VIEW_CONVERT_EXPR) | |
4267 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
4268 | && ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4269 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4270 | || ((TYPE_PRECISION (TREE_TYPE (op)) | |
4271 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))) | |
4272 | /* But a conversion that does not change the bit-pattern is ok. */ | |
4273 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4274 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2dab46d5 IE |
4275 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
4276 | /* Conversion between boolean types of different sizes is | |
4277 | a simple assignment in case their vectypes are same | |
4278 | boolean vectors. */ | |
4279 | && (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
4280 | || !VECTOR_BOOLEAN_TYPE_P (vectype_in))) | |
7b7b1813 | 4281 | { |
73fbfcad | 4282 | if (dump_enabled_p ()) |
78c60e3d SS |
4283 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4284 | "type conversion to/from bit-precision " | |
e645e942 | 4285 | "unsupported.\n"); |
7b7b1813 RG |
4286 | return false; |
4287 | } | |
4288 | ||
ebfd146a IR |
4289 | if (!vec_stmt) /* transformation not required. */ |
4290 | { | |
4291 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 4292 | if (dump_enabled_p ()) |
78c60e3d | 4293 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4294 | "=== vectorizable_assignment ===\n"); |
c3e7ee41 | 4295 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
4296 | return true; |
4297 | } | |
4298 | ||
4299 | /** Transform. **/ | |
73fbfcad | 4300 | if (dump_enabled_p ()) |
e645e942 | 4301 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
4302 | |
4303 | /* Handle def. */ | |
4304 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4305 | ||
4306 | /* Handle use. */ | |
f18b55bd | 4307 | for (j = 0; j < ncopies; j++) |
ebfd146a | 4308 | { |
f18b55bd IR |
4309 | /* Handle uses. */ |
4310 | if (j == 0) | |
d092494c | 4311 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node, -1); |
f18b55bd IR |
4312 | else |
4313 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
4314 | ||
4315 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 4316 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 4317 | { |
7b7ec6c5 RG |
4318 | if (CONVERT_EXPR_CODE_P (code) |
4319 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 4320 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
4321 | new_stmt = gimple_build_assign (vec_dest, vop); |
4322 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4323 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4324 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4325 | if (slp_node) | |
9771b263 | 4326 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 4327 | } |
ebfd146a IR |
4328 | |
4329 | if (slp_node) | |
f18b55bd IR |
4330 | continue; |
4331 | ||
4332 | if (j == 0) | |
4333 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4334 | else | |
4335 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4336 | ||
4337 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4338 | } | |
b8698a0f | 4339 | |
9771b263 | 4340 | vec_oprnds.release (); |
ebfd146a IR |
4341 | return true; |
4342 | } | |
4343 | ||
9dc3f7de | 4344 | |
1107f3ae IR |
4345 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
4346 | either as shift by a scalar or by a vector. */ | |
4347 | ||
4348 | bool | |
4349 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
4350 | { | |
4351 | ||
ef4bddc2 | 4352 | machine_mode vec_mode; |
1107f3ae IR |
4353 | optab optab; |
4354 | int icode; | |
4355 | tree vectype; | |
4356 | ||
4357 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4358 | if (!vectype) | |
4359 | return false; | |
4360 | ||
4361 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4362 | if (!optab | |
4363 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
4364 | { | |
4365 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4366 | if (!optab | |
4367 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
4368 | == CODE_FOR_nothing)) | |
4369 | return false; | |
4370 | } | |
4371 | ||
4372 | vec_mode = TYPE_MODE (vectype); | |
4373 | icode = (int) optab_handler (optab, vec_mode); | |
4374 | if (icode == CODE_FOR_nothing) | |
4375 | return false; | |
4376 | ||
4377 | return true; | |
4378 | } | |
4379 | ||
4380 | ||
9dc3f7de IR |
4381 | /* Function vectorizable_shift. |
4382 | ||
4383 | Check if STMT performs a shift operation that can be vectorized. | |
4384 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4385 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
4386 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4387 | ||
4388 | static bool | |
355fe088 TS |
4389 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
4390 | gimple **vec_stmt, slp_tree slp_node) | |
9dc3f7de IR |
4391 | { |
4392 | tree vec_dest; | |
4393 | tree scalar_dest; | |
4394 | tree op0, op1 = NULL; | |
4395 | tree vec_oprnd1 = NULL_TREE; | |
4396 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4397 | tree vectype; | |
4398 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4399 | enum tree_code code; | |
ef4bddc2 | 4400 | machine_mode vec_mode; |
9dc3f7de IR |
4401 | tree new_temp; |
4402 | optab optab; | |
4403 | int icode; | |
ef4bddc2 | 4404 | machine_mode optab_op2_mode; |
355fe088 | 4405 | gimple *def_stmt; |
9dc3f7de | 4406 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
355fe088 | 4407 | gimple *new_stmt = NULL; |
9dc3f7de IR |
4408 | stmt_vec_info prev_stmt_info; |
4409 | int nunits_in; | |
4410 | int nunits_out; | |
4411 | tree vectype_out; | |
cede2577 | 4412 | tree op1_vectype; |
9dc3f7de IR |
4413 | int ncopies; |
4414 | int j, i; | |
6e1aa848 DN |
4415 | vec<tree> vec_oprnds0 = vNULL; |
4416 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
4417 | tree vop0, vop1; |
4418 | unsigned int k; | |
49eab32e | 4419 | bool scalar_shift_arg = true; |
9dc3f7de | 4420 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4421 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
4422 | int vf; |
4423 | ||
4424 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
4425 | return false; | |
4426 | ||
66c16fd9 RB |
4427 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4428 | && ! vec_stmt) | |
9dc3f7de IR |
4429 | return false; |
4430 | ||
4431 | /* Is STMT a vectorizable binary/unary operation? */ | |
4432 | if (!is_gimple_assign (stmt)) | |
4433 | return false; | |
4434 | ||
4435 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4436 | return false; | |
4437 | ||
4438 | code = gimple_assign_rhs_code (stmt); | |
4439 | ||
4440 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
4441 | || code == RROTATE_EXPR)) | |
4442 | return false; | |
4443 | ||
4444 | scalar_dest = gimple_assign_lhs (stmt); | |
4445 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
7b7b1813 RG |
4446 | if (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
4447 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4448 | { | |
73fbfcad | 4449 | if (dump_enabled_p ()) |
78c60e3d | 4450 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4451 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
4452 | return false; |
4453 | } | |
9dc3f7de IR |
4454 | |
4455 | op0 = gimple_assign_rhs1 (stmt); | |
81c40241 | 4456 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
9dc3f7de | 4457 | { |
73fbfcad | 4458 | if (dump_enabled_p ()) |
78c60e3d | 4459 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4460 | "use not simple.\n"); |
9dc3f7de IR |
4461 | return false; |
4462 | } | |
4463 | /* If op0 is an external or constant def use a vector type with | |
4464 | the same size as the output vector type. */ | |
4465 | if (!vectype) | |
4466 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4467 | if (vec_stmt) | |
4468 | gcc_assert (vectype); | |
4469 | if (!vectype) | |
4470 | { | |
73fbfcad | 4471 | if (dump_enabled_p ()) |
78c60e3d | 4472 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4473 | "no vectype for scalar type\n"); |
9dc3f7de IR |
4474 | return false; |
4475 | } | |
4476 | ||
4477 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4478 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4479 | if (nunits_out != nunits_in) | |
4480 | return false; | |
4481 | ||
4482 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4483 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &op1_vectype)) |
9dc3f7de | 4484 | { |
73fbfcad | 4485 | if (dump_enabled_p ()) |
78c60e3d | 4486 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4487 | "use not simple.\n"); |
9dc3f7de IR |
4488 | return false; |
4489 | } | |
4490 | ||
4491 | if (loop_vinfo) | |
4492 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4493 | else | |
4494 | vf = 1; | |
4495 | ||
4496 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4497 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4498 | case of SLP. */ | |
fce57248 | 4499 | if (slp_node) |
9dc3f7de IR |
4500 | ncopies = 1; |
4501 | else | |
4502 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4503 | ||
4504 | gcc_assert (ncopies >= 1); | |
4505 | ||
4506 | /* Determine whether the shift amount is a vector, or scalar. If the | |
4507 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
4508 | ||
dbfa87aa YR |
4509 | if ((dt[1] == vect_internal_def |
4510 | || dt[1] == vect_induction_def) | |
4511 | && !slp_node) | |
49eab32e JJ |
4512 | scalar_shift_arg = false; |
4513 | else if (dt[1] == vect_constant_def | |
4514 | || dt[1] == vect_external_def | |
4515 | || dt[1] == vect_internal_def) | |
4516 | { | |
4517 | /* In SLP, need to check whether the shift count is the same, | |
4518 | in loops if it is a constant or invariant, it is always | |
4519 | a scalar shift. */ | |
4520 | if (slp_node) | |
4521 | { | |
355fe088 TS |
4522 | vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
4523 | gimple *slpstmt; | |
49eab32e | 4524 | |
9771b263 | 4525 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
4526 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
4527 | scalar_shift_arg = false; | |
4528 | } | |
60d393e8 RB |
4529 | |
4530 | /* If the shift amount is computed by a pattern stmt we cannot | |
4531 | use the scalar amount directly thus give up and use a vector | |
4532 | shift. */ | |
4533 | if (dt[1] == vect_internal_def) | |
4534 | { | |
4535 | gimple *def = SSA_NAME_DEF_STMT (op1); | |
4536 | if (is_pattern_stmt_p (vinfo_for_stmt (def))) | |
4537 | scalar_shift_arg = false; | |
4538 | } | |
49eab32e JJ |
4539 | } |
4540 | else | |
4541 | { | |
73fbfcad | 4542 | if (dump_enabled_p ()) |
78c60e3d | 4543 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4544 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
4545 | return false; |
4546 | } | |
4547 | ||
9dc3f7de | 4548 | /* Vector shifted by vector. */ |
49eab32e | 4549 | if (!scalar_shift_arg) |
9dc3f7de IR |
4550 | { |
4551 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 4552 | if (dump_enabled_p ()) |
78c60e3d | 4553 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4554 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 4555 | |
aa948027 JJ |
4556 | if (!op1_vectype) |
4557 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
4558 | if (op1_vectype == NULL_TREE | |
4559 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 4560 | { |
73fbfcad | 4561 | if (dump_enabled_p ()) |
78c60e3d SS |
4562 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4563 | "unusable type for last operand in" | |
e645e942 | 4564 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
4565 | return false; |
4566 | } | |
9dc3f7de IR |
4567 | } |
4568 | /* See if the machine has a vector shifted by scalar insn and if not | |
4569 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 4570 | else |
9dc3f7de IR |
4571 | { |
4572 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4573 | if (optab | |
4574 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
4575 | { | |
73fbfcad | 4576 | if (dump_enabled_p ()) |
78c60e3d | 4577 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4578 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
4579 | } |
4580 | else | |
4581 | { | |
4582 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4583 | if (optab | |
4584 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
4585 | != CODE_FOR_nothing)) | |
4586 | { | |
49eab32e JJ |
4587 | scalar_shift_arg = false; |
4588 | ||
73fbfcad | 4589 | if (dump_enabled_p ()) |
78c60e3d | 4590 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4591 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
4592 | |
4593 | /* Unlike the other binary operators, shifts/rotates have | |
4594 | the rhs being int, instead of the same type as the lhs, | |
4595 | so make sure the scalar is the right type if we are | |
aa948027 | 4596 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
4597 | if (dt[1] == vect_constant_def) |
4598 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
4599 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
4600 | TREE_TYPE (op1))) | |
4601 | { | |
4602 | if (slp_node | |
4603 | && TYPE_MODE (TREE_TYPE (vectype)) | |
4604 | != TYPE_MODE (TREE_TYPE (op1))) | |
4605 | { | |
73fbfcad | 4606 | if (dump_enabled_p ()) |
78c60e3d SS |
4607 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4608 | "unusable type for last operand in" | |
e645e942 | 4609 | " vector/vector shift/rotate.\n"); |
21c0a521 | 4610 | return false; |
aa948027 JJ |
4611 | } |
4612 | if (vec_stmt && !slp_node) | |
4613 | { | |
4614 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
4615 | op1 = vect_init_vector (stmt, op1, | |
4616 | TREE_TYPE (vectype), NULL); | |
4617 | } | |
4618 | } | |
9dc3f7de IR |
4619 | } |
4620 | } | |
4621 | } | |
9dc3f7de IR |
4622 | |
4623 | /* Supportable by target? */ | |
4624 | if (!optab) | |
4625 | { | |
73fbfcad | 4626 | if (dump_enabled_p ()) |
78c60e3d | 4627 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4628 | "no optab.\n"); |
9dc3f7de IR |
4629 | return false; |
4630 | } | |
4631 | vec_mode = TYPE_MODE (vectype); | |
4632 | icode = (int) optab_handler (optab, vec_mode); | |
4633 | if (icode == CODE_FOR_nothing) | |
4634 | { | |
73fbfcad | 4635 | if (dump_enabled_p ()) |
78c60e3d | 4636 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4637 | "op not supported by target.\n"); |
9dc3f7de IR |
4638 | /* Check only during analysis. */ |
4639 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
4640 | || (vf < vect_min_worthwhile_factor (code) | |
4641 | && !vec_stmt)) | |
4642 | return false; | |
73fbfcad | 4643 | if (dump_enabled_p ()) |
e645e942 TJ |
4644 | dump_printf_loc (MSG_NOTE, vect_location, |
4645 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
4646 | } |
4647 | ||
4648 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
4649 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
4650 | && vf < vect_min_worthwhile_factor (code) | |
4651 | && !vec_stmt) | |
4652 | { | |
73fbfcad | 4653 | if (dump_enabled_p ()) |
78c60e3d | 4654 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4655 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
4656 | return false; |
4657 | } | |
4658 | ||
4659 | if (!vec_stmt) /* transformation not required. */ | |
4660 | { | |
4661 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 4662 | if (dump_enabled_p ()) |
e645e942 TJ |
4663 | dump_printf_loc (MSG_NOTE, vect_location, |
4664 | "=== vectorizable_shift ===\n"); | |
c3e7ee41 | 4665 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
9dc3f7de IR |
4666 | return true; |
4667 | } | |
4668 | ||
4669 | /** Transform. **/ | |
4670 | ||
73fbfcad | 4671 | if (dump_enabled_p ()) |
78c60e3d | 4672 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4673 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
4674 | |
4675 | /* Handle def. */ | |
4676 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4677 | ||
9dc3f7de IR |
4678 | prev_stmt_info = NULL; |
4679 | for (j = 0; j < ncopies; j++) | |
4680 | { | |
4681 | /* Handle uses. */ | |
4682 | if (j == 0) | |
4683 | { | |
4684 | if (scalar_shift_arg) | |
4685 | { | |
4686 | /* Vector shl and shr insn patterns can be defined with scalar | |
4687 | operand 2 (shift operand). In this case, use constant or loop | |
4688 | invariant op1 directly, without extending it to vector mode | |
4689 | first. */ | |
4690 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
4691 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
4692 | { | |
73fbfcad | 4693 | if (dump_enabled_p ()) |
78c60e3d | 4694 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4695 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 4696 | vec_oprnd1 = op1; |
8930f723 | 4697 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 4698 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4699 | if (slp_node) |
4700 | { | |
4701 | /* Store vec_oprnd1 for every vector stmt to be created | |
4702 | for SLP_NODE. We check during the analysis that all | |
4703 | the shift arguments are the same. | |
4704 | TODO: Allow different constants for different vector | |
4705 | stmts generated for an SLP instance. */ | |
4706 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4707 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4708 | } |
4709 | } | |
4710 | } | |
4711 | ||
4712 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
4713 | (a special case for certain kind of vector shifts); otherwise, | |
4714 | operand 1 should be of a vector type (the usual case). */ | |
4715 | if (vec_oprnd1) | |
4716 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
d092494c | 4717 | slp_node, -1); |
9dc3f7de IR |
4718 | else |
4719 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
d092494c | 4720 | slp_node, -1); |
9dc3f7de IR |
4721 | } |
4722 | else | |
4723 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
4724 | ||
4725 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 4726 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 4727 | { |
9771b263 | 4728 | vop1 = vec_oprnds1[i]; |
0d0e4a03 | 4729 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
4730 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4731 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4732 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4733 | if (slp_node) | |
9771b263 | 4734 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
4735 | } |
4736 | ||
4737 | if (slp_node) | |
4738 | continue; | |
4739 | ||
4740 | if (j == 0) | |
4741 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4742 | else | |
4743 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4744 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4745 | } | |
4746 | ||
9771b263 DN |
4747 | vec_oprnds0.release (); |
4748 | vec_oprnds1.release (); | |
9dc3f7de IR |
4749 | |
4750 | return true; | |
4751 | } | |
4752 | ||
4753 | ||
ebfd146a IR |
4754 | /* Function vectorizable_operation. |
4755 | ||
16949072 RG |
4756 | Check if STMT performs a binary, unary or ternary operation that can |
4757 | be vectorized. | |
b8698a0f | 4758 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
4759 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4760 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4761 | ||
4762 | static bool | |
355fe088 TS |
4763 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
4764 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a | 4765 | { |
00f07b86 | 4766 | tree vec_dest; |
ebfd146a | 4767 | tree scalar_dest; |
16949072 | 4768 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 4769 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 4770 | tree vectype; |
ebfd146a IR |
4771 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4772 | enum tree_code code; | |
ef4bddc2 | 4773 | machine_mode vec_mode; |
ebfd146a IR |
4774 | tree new_temp; |
4775 | int op_type; | |
00f07b86 | 4776 | optab optab; |
523ba738 | 4777 | bool target_support_p; |
355fe088 | 4778 | gimple *def_stmt; |
16949072 RG |
4779 | enum vect_def_type dt[3] |
4780 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
355fe088 | 4781 | gimple *new_stmt = NULL; |
ebfd146a | 4782 | stmt_vec_info prev_stmt_info; |
b690cc0f | 4783 | int nunits_in; |
ebfd146a IR |
4784 | int nunits_out; |
4785 | tree vectype_out; | |
4786 | int ncopies; | |
4787 | int j, i; | |
6e1aa848 DN |
4788 | vec<tree> vec_oprnds0 = vNULL; |
4789 | vec<tree> vec_oprnds1 = vNULL; | |
4790 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 4791 | tree vop0, vop1, vop2; |
a70d6342 | 4792 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4793 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 IR |
4794 | int vf; |
4795 | ||
a70d6342 | 4796 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4797 | return false; |
4798 | ||
66c16fd9 RB |
4799 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4800 | && ! vec_stmt) | |
ebfd146a IR |
4801 | return false; |
4802 | ||
4803 | /* Is STMT a vectorizable binary/unary operation? */ | |
4804 | if (!is_gimple_assign (stmt)) | |
4805 | return false; | |
4806 | ||
4807 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4808 | return false; | |
4809 | ||
ebfd146a IR |
4810 | code = gimple_assign_rhs_code (stmt); |
4811 | ||
4812 | /* For pointer addition, we should use the normal plus for | |
4813 | the vector addition. */ | |
4814 | if (code == POINTER_PLUS_EXPR) | |
4815 | code = PLUS_EXPR; | |
4816 | ||
4817 | /* Support only unary or binary operations. */ | |
4818 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 4819 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 4820 | { |
73fbfcad | 4821 | if (dump_enabled_p ()) |
78c60e3d | 4822 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4823 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 4824 | op_type); |
ebfd146a IR |
4825 | return false; |
4826 | } | |
4827 | ||
b690cc0f RG |
4828 | scalar_dest = gimple_assign_lhs (stmt); |
4829 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
4830 | ||
7b7b1813 RG |
4831 | /* Most operations cannot handle bit-precision types without extra |
4832 | truncations. */ | |
045c1278 IE |
4833 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4834 | && (TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4835 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
7b7b1813 RG |
4836 | /* Exception are bitwise binary operations. */ |
4837 | && code != BIT_IOR_EXPR | |
4838 | && code != BIT_XOR_EXPR | |
4839 | && code != BIT_AND_EXPR) | |
4840 | { | |
73fbfcad | 4841 | if (dump_enabled_p ()) |
78c60e3d | 4842 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4843 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
4844 | return false; |
4845 | } | |
4846 | ||
ebfd146a | 4847 | op0 = gimple_assign_rhs1 (stmt); |
81c40241 | 4848 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
ebfd146a | 4849 | { |
73fbfcad | 4850 | if (dump_enabled_p ()) |
78c60e3d | 4851 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4852 | "use not simple.\n"); |
ebfd146a IR |
4853 | return false; |
4854 | } | |
b690cc0f RG |
4855 | /* If op0 is an external or constant def use a vector type with |
4856 | the same size as the output vector type. */ | |
4857 | if (!vectype) | |
b036c6c5 IE |
4858 | { |
4859 | /* For boolean type we cannot determine vectype by | |
4860 | invariant value (don't know whether it is a vector | |
4861 | of booleans or vector of integers). We use output | |
4862 | vectype because operations on boolean don't change | |
4863 | type. */ | |
4864 | if (TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE) | |
4865 | { | |
4866 | if (TREE_CODE (TREE_TYPE (scalar_dest)) != BOOLEAN_TYPE) | |
4867 | { | |
4868 | if (dump_enabled_p ()) | |
4869 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4870 | "not supported operation on bool value.\n"); | |
4871 | return false; | |
4872 | } | |
4873 | vectype = vectype_out; | |
4874 | } | |
4875 | else | |
4876 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4877 | } | |
7d8930a0 IR |
4878 | if (vec_stmt) |
4879 | gcc_assert (vectype); | |
4880 | if (!vectype) | |
4881 | { | |
73fbfcad | 4882 | if (dump_enabled_p ()) |
7d8930a0 | 4883 | { |
78c60e3d SS |
4884 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4885 | "no vectype for scalar type "); | |
4886 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
4887 | TREE_TYPE (op0)); | |
e645e942 | 4888 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
4889 | } |
4890 | ||
4891 | return false; | |
4892 | } | |
b690cc0f RG |
4893 | |
4894 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4895 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4896 | if (nunits_out != nunits_in) | |
4897 | return false; | |
ebfd146a | 4898 | |
16949072 | 4899 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
4900 | { |
4901 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4902 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1])) |
ebfd146a | 4903 | { |
73fbfcad | 4904 | if (dump_enabled_p ()) |
78c60e3d | 4905 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4906 | "use not simple.\n"); |
ebfd146a IR |
4907 | return false; |
4908 | } | |
4909 | } | |
16949072 RG |
4910 | if (op_type == ternary_op) |
4911 | { | |
4912 | op2 = gimple_assign_rhs3 (stmt); | |
81c40241 | 4913 | if (!vect_is_simple_use (op2, vinfo, &def_stmt, &dt[2])) |
16949072 | 4914 | { |
73fbfcad | 4915 | if (dump_enabled_p ()) |
78c60e3d | 4916 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4917 | "use not simple.\n"); |
16949072 RG |
4918 | return false; |
4919 | } | |
4920 | } | |
ebfd146a | 4921 | |
b690cc0f RG |
4922 | if (loop_vinfo) |
4923 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4924 | else | |
4925 | vf = 1; | |
4926 | ||
4927 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4928 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 4929 | case of SLP. */ |
fce57248 | 4930 | if (slp_node) |
b690cc0f RG |
4931 | ncopies = 1; |
4932 | else | |
4933 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4934 | ||
4935 | gcc_assert (ncopies >= 1); | |
4936 | ||
9dc3f7de | 4937 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
4938 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
4939 | || code == RROTATE_EXPR) | |
9dc3f7de | 4940 | return false; |
ebfd146a | 4941 | |
ebfd146a | 4942 | /* Supportable by target? */ |
00f07b86 RH |
4943 | |
4944 | vec_mode = TYPE_MODE (vectype); | |
4945 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 4946 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
4947 | else |
4948 | { | |
4949 | optab = optab_for_tree_code (code, vectype, optab_default); | |
4950 | if (!optab) | |
5deb57cb | 4951 | { |
73fbfcad | 4952 | if (dump_enabled_p ()) |
78c60e3d | 4953 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4954 | "no optab.\n"); |
00f07b86 | 4955 | return false; |
5deb57cb | 4956 | } |
523ba738 RS |
4957 | target_support_p = (optab_handler (optab, vec_mode) |
4958 | != CODE_FOR_nothing); | |
5deb57cb JJ |
4959 | } |
4960 | ||
523ba738 | 4961 | if (!target_support_p) |
ebfd146a | 4962 | { |
73fbfcad | 4963 | if (dump_enabled_p ()) |
78c60e3d | 4964 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4965 | "op not supported by target.\n"); |
ebfd146a IR |
4966 | /* Check only during analysis. */ |
4967 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5deb57cb | 4968 | || (!vec_stmt && vf < vect_min_worthwhile_factor (code))) |
ebfd146a | 4969 | return false; |
73fbfcad | 4970 | if (dump_enabled_p ()) |
e645e942 TJ |
4971 | dump_printf_loc (MSG_NOTE, vect_location, |
4972 | "proceeding using word mode.\n"); | |
383d9c83 IR |
4973 | } |
4974 | ||
4a00c761 | 4975 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
4976 | if (!VECTOR_MODE_P (vec_mode) |
4977 | && !vec_stmt | |
4978 | && vf < vect_min_worthwhile_factor (code)) | |
7d8930a0 | 4979 | { |
73fbfcad | 4980 | if (dump_enabled_p ()) |
78c60e3d | 4981 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4982 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 4983 | return false; |
7d8930a0 | 4984 | } |
ebfd146a | 4985 | |
ebfd146a IR |
4986 | if (!vec_stmt) /* transformation not required. */ |
4987 | { | |
4a00c761 | 4988 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 4989 | if (dump_enabled_p ()) |
78c60e3d | 4990 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4991 | "=== vectorizable_operation ===\n"); |
c3e7ee41 | 4992 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
4993 | return true; |
4994 | } | |
4995 | ||
4996 | /** Transform. **/ | |
4997 | ||
73fbfcad | 4998 | if (dump_enabled_p ()) |
78c60e3d | 4999 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5000 | "transform binary/unary operation.\n"); |
383d9c83 | 5001 | |
ebfd146a | 5002 | /* Handle def. */ |
00f07b86 | 5003 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 5004 | |
ebfd146a IR |
5005 | /* In case the vectorization factor (VF) is bigger than the number |
5006 | of elements that we can fit in a vectype (nunits), we have to generate | |
5007 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
5008 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
5009 | from one copy of the vector stmt to the next, in the field | |
5010 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
5011 | stages to find the correct vector defs to be used when vectorizing | |
5012 | stmts that use the defs of the current stmt. The example below | |
5013 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
5014 | we need to create 4 vectorized stmts): | |
5015 | ||
5016 | before vectorization: | |
5017 | RELATED_STMT VEC_STMT | |
5018 | S1: x = memref - - | |
5019 | S2: z = x + 1 - - | |
5020 | ||
5021 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
5022 | there): | |
5023 | RELATED_STMT VEC_STMT | |
5024 | VS1_0: vx0 = memref0 VS1_1 - | |
5025 | VS1_1: vx1 = memref1 VS1_2 - | |
5026 | VS1_2: vx2 = memref2 VS1_3 - | |
5027 | VS1_3: vx3 = memref3 - - | |
5028 | S1: x = load - VS1_0 | |
5029 | S2: z = x + 1 - - | |
5030 | ||
5031 | step2: vectorize stmt S2 (done here): | |
5032 | To vectorize stmt S2 we first need to find the relevant vector | |
5033 | def for the first operand 'x'. This is, as usual, obtained from | |
5034 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
5035 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
5036 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
5037 | the vector stmt VS2_0, and as usual, record it in the | |
5038 | STMT_VINFO_VEC_STMT of stmt S2. | |
5039 | When creating the second copy (VS2_1), we obtain the relevant vector | |
5040 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
5041 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
5042 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
5043 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
5044 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
5045 | chain of stmts and pointers: | |
5046 | RELATED_STMT VEC_STMT | |
5047 | VS1_0: vx0 = memref0 VS1_1 - | |
5048 | VS1_1: vx1 = memref1 VS1_2 - | |
5049 | VS1_2: vx2 = memref2 VS1_3 - | |
5050 | VS1_3: vx3 = memref3 - - | |
5051 | S1: x = load - VS1_0 | |
5052 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
5053 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
5054 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
5055 | VS2_3: vz3 = vx3 + v1 - - | |
5056 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
5057 | |
5058 | prev_stmt_info = NULL; | |
5059 | for (j = 0; j < ncopies; j++) | |
5060 | { | |
5061 | /* Handle uses. */ | |
5062 | if (j == 0) | |
4a00c761 JJ |
5063 | { |
5064 | if (op_type == binary_op || op_type == ternary_op) | |
5065 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
5066 | slp_node, -1); | |
5067 | else | |
5068 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
5069 | slp_node, -1); | |
5070 | if (op_type == ternary_op) | |
36ba4aae | 5071 | { |
9771b263 DN |
5072 | vec_oprnds2.create (1); |
5073 | vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2, | |
81c40241 | 5074 | stmt)); |
36ba4aae | 5075 | } |
4a00c761 | 5076 | } |
ebfd146a | 5077 | else |
4a00c761 JJ |
5078 | { |
5079 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5080 | if (op_type == ternary_op) | |
5081 | { | |
9771b263 DN |
5082 | tree vec_oprnd = vec_oprnds2.pop (); |
5083 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
5084 | vec_oprnd)); | |
4a00c761 JJ |
5085 | } |
5086 | } | |
5087 | ||
5088 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5089 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 5090 | { |
4a00c761 | 5091 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 5092 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 5093 | vop2 = ((op_type == ternary_op) |
9771b263 | 5094 | ? vec_oprnds2[i] : NULL_TREE); |
0d0e4a03 | 5095 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
4a00c761 JJ |
5096 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5097 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5098 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5099 | if (slp_node) | |
9771b263 | 5100 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
5101 | } |
5102 | ||
4a00c761 JJ |
5103 | if (slp_node) |
5104 | continue; | |
5105 | ||
5106 | if (j == 0) | |
5107 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5108 | else | |
5109 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5110 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
5111 | } |
5112 | ||
9771b263 DN |
5113 | vec_oprnds0.release (); |
5114 | vec_oprnds1.release (); | |
5115 | vec_oprnds2.release (); | |
ebfd146a | 5116 | |
ebfd146a IR |
5117 | return true; |
5118 | } | |
5119 | ||
c716e67f XDL |
5120 | /* A helper function to ensure data reference DR's base alignment |
5121 | for STMT_INFO. */ | |
5122 | ||
5123 | static void | |
5124 | ensure_base_align (stmt_vec_info stmt_info, struct data_reference *dr) | |
5125 | { | |
5126 | if (!dr->aux) | |
5127 | return; | |
5128 | ||
52639a61 | 5129 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f XDL |
5130 | { |
5131 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
52639a61 | 5132 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 5133 | |
428f0c67 JH |
5134 | if (decl_in_symtab_p (base_decl)) |
5135 | symtab_node::get (base_decl)->increase_alignment (TYPE_ALIGN (vectype)); | |
5136 | else | |
5137 | { | |
fe37c7af | 5138 | SET_DECL_ALIGN (base_decl, TYPE_ALIGN (vectype)); |
428f0c67 JH |
5139 | DECL_USER_ALIGN (base_decl) = 1; |
5140 | } | |
52639a61 | 5141 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
5142 | } |
5143 | } | |
5144 | ||
ebfd146a | 5145 | |
09dfa495 BM |
5146 | /* Given a vector type VECTYPE returns the VECTOR_CST mask that implements |
5147 | reversal of the vector elements. If that is impossible to do, | |
5148 | returns NULL. */ | |
5149 | ||
5150 | static tree | |
5151 | perm_mask_for_reverse (tree vectype) | |
5152 | { | |
5153 | int i, nunits; | |
5154 | unsigned char *sel; | |
5155 | ||
5156 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
5157 | sel = XALLOCAVEC (unsigned char, nunits); | |
5158 | ||
5159 | for (i = 0; i < nunits; ++i) | |
5160 | sel[i] = nunits - 1 - i; | |
5161 | ||
557be5a8 AL |
5162 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) |
5163 | return NULL_TREE; | |
5164 | return vect_gen_perm_mask_checked (vectype, sel); | |
09dfa495 BM |
5165 | } |
5166 | ||
ebfd146a IR |
5167 | /* Function vectorizable_store. |
5168 | ||
b8698a0f L |
5169 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
5170 | can be vectorized. | |
5171 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5172 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5173 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5174 | ||
5175 | static bool | |
355fe088 | 5176 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 5177 | slp_tree slp_node) |
ebfd146a IR |
5178 | { |
5179 | tree scalar_dest; | |
5180 | tree data_ref; | |
5181 | tree op; | |
5182 | tree vec_oprnd = NULL_TREE; | |
5183 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5184 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 5185 | tree elem_type; |
ebfd146a | 5186 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 5187 | struct loop *loop = NULL; |
ef4bddc2 | 5188 | machine_mode vec_mode; |
ebfd146a IR |
5189 | tree dummy; |
5190 | enum dr_alignment_support alignment_support_scheme; | |
355fe088 | 5191 | gimple *def_stmt; |
ebfd146a IR |
5192 | enum vect_def_type dt; |
5193 | stmt_vec_info prev_stmt_info = NULL; | |
5194 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 5195 | tree dataref_offset = NULL_TREE; |
355fe088 | 5196 | gimple *ptr_incr = NULL; |
ebfd146a IR |
5197 | int ncopies; |
5198 | int j; | |
355fe088 | 5199 | gimple *next_stmt, *first_stmt = NULL; |
0d0293ac | 5200 | bool grouped_store = false; |
272c6793 | 5201 | bool store_lanes_p = false; |
ebfd146a | 5202 | unsigned int group_size, i; |
6e1aa848 DN |
5203 | vec<tree> dr_chain = vNULL; |
5204 | vec<tree> oprnds = vNULL; | |
5205 | vec<tree> result_chain = vNULL; | |
ebfd146a | 5206 | bool inv_p; |
09dfa495 BM |
5207 | bool negative = false; |
5208 | tree offset = NULL_TREE; | |
6e1aa848 | 5209 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5210 | bool slp = (slp_node != NULL); |
ebfd146a | 5211 | unsigned int vec_num; |
a70d6342 | 5212 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5213 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 5214 | tree aggr_type; |
3bab6342 AT |
5215 | tree scatter_base = NULL_TREE, scatter_off = NULL_TREE; |
5216 | tree scatter_off_vectype = NULL_TREE, scatter_decl = NULL_TREE; | |
5217 | int scatter_scale = 1; | |
5218 | enum vect_def_type scatter_idx_dt = vect_unknown_def_type; | |
5219 | enum vect_def_type scatter_src_dt = vect_unknown_def_type; | |
355fe088 | 5220 | gimple *new_stmt; |
a70d6342 | 5221 | |
a70d6342 | 5222 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5223 | return false; |
5224 | ||
66c16fd9 RB |
5225 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5226 | && ! vec_stmt) | |
ebfd146a IR |
5227 | return false; |
5228 | ||
5229 | /* Is vectorizable store? */ | |
5230 | ||
5231 | if (!is_gimple_assign (stmt)) | |
5232 | return false; | |
5233 | ||
5234 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
5235 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
5236 | && is_pattern_stmt_p (stmt_info)) | |
5237 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 5238 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 5239 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 5240 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
5241 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
5242 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
5243 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
5244 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
5245 | return false; |
5246 | ||
fce57248 RS |
5247 | /* Cannot have hybrid store SLP -- that would mean storing to the |
5248 | same location twice. */ | |
5249 | gcc_assert (slp == PURE_SLP_STMT (stmt_info)); | |
5250 | ||
ebfd146a | 5251 | gcc_assert (gimple_assign_single_p (stmt)); |
465c8c19 | 5252 | |
f4d09712 | 5253 | tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE; |
465c8c19 JJ |
5254 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
5255 | ||
5256 | if (loop_vinfo) | |
5257 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
5258 | ||
5259 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5260 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5261 | case of SLP. */ | |
fce57248 | 5262 | if (slp) |
465c8c19 JJ |
5263 | ncopies = 1; |
5264 | else | |
5265 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
5266 | ||
5267 | gcc_assert (ncopies >= 1); | |
5268 | ||
5269 | /* FORNOW. This restriction should be relaxed. */ | |
5270 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
5271 | { | |
5272 | if (dump_enabled_p ()) | |
5273 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5274 | "multiple types in nested loop.\n"); | |
5275 | return false; | |
5276 | } | |
5277 | ||
ebfd146a | 5278 | op = gimple_assign_rhs1 (stmt); |
f4d09712 KY |
5279 | |
5280 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt, &rhs_vectype)) | |
ebfd146a | 5281 | { |
73fbfcad | 5282 | if (dump_enabled_p ()) |
78c60e3d | 5283 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5284 | "use not simple.\n"); |
ebfd146a IR |
5285 | return false; |
5286 | } | |
5287 | ||
f4d09712 KY |
5288 | if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype)) |
5289 | return false; | |
5290 | ||
272c6793 | 5291 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 5292 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 5293 | |
ebfd146a IR |
5294 | /* FORNOW. In some cases can vectorize even if data-type not supported |
5295 | (e.g. - array initialization with 0). */ | |
947131ba | 5296 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
5297 | return false; |
5298 | ||
5299 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
5300 | return false; | |
5301 | ||
f2e2a985 | 5302 | if (!STMT_VINFO_STRIDED_P (stmt_info)) |
09dfa495 | 5303 | { |
f2e2a985 MM |
5304 | negative = |
5305 | tree_int_cst_compare (loop && nested_in_vect_loop_p (loop, stmt) | |
5306 | ? STMT_VINFO_DR_STEP (stmt_info) : DR_STEP (dr), | |
5307 | size_zero_node) < 0; | |
5308 | if (negative && ncopies > 1) | |
09dfa495 BM |
5309 | { |
5310 | if (dump_enabled_p ()) | |
5311 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
f2e2a985 | 5312 | "multiple types with negative step.\n"); |
09dfa495 BM |
5313 | return false; |
5314 | } | |
f2e2a985 | 5315 | if (negative) |
09dfa495 | 5316 | { |
f2e2a985 MM |
5317 | gcc_assert (!grouped_store); |
5318 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
5319 | if (alignment_support_scheme != dr_aligned | |
5320 | && alignment_support_scheme != dr_unaligned_supported) | |
5321 | { | |
5322 | if (dump_enabled_p ()) | |
5323 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5324 | "negative step but alignment required.\n"); | |
5325 | return false; | |
5326 | } | |
5327 | if (dt != vect_constant_def | |
5328 | && dt != vect_external_def | |
5329 | && !perm_mask_for_reverse (vectype)) | |
5330 | { | |
5331 | if (dump_enabled_p ()) | |
5332 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5333 | "negative step and reversing not supported.\n"); | |
5334 | return false; | |
5335 | } | |
09dfa495 BM |
5336 | } |
5337 | } | |
5338 | ||
0d0293ac | 5339 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 5340 | { |
0d0293ac | 5341 | grouped_store = true; |
e14c1050 | 5342 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
cee62fee | 5343 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
fce57248 | 5344 | if (!slp && !STMT_VINFO_STRIDED_P (stmt_info)) |
b602d918 | 5345 | { |
272c6793 RS |
5346 | if (vect_store_lanes_supported (vectype, group_size)) |
5347 | store_lanes_p = true; | |
0d0293ac | 5348 | else if (!vect_grouped_store_supported (vectype, group_size)) |
b602d918 RS |
5349 | return false; |
5350 | } | |
b8698a0f | 5351 | |
cee62fee | 5352 | if (STMT_VINFO_STRIDED_P (stmt_info) |
fce57248 | 5353 | && slp |
cee62fee MM |
5354 | && (group_size > nunits |
5355 | || nunits % group_size != 0)) | |
5356 | { | |
5357 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5358 | "unhandled strided group store\n"); | |
5359 | return false; | |
5360 | } | |
5361 | ||
ebfd146a IR |
5362 | if (first_stmt == stmt) |
5363 | { | |
5364 | /* STMT is the leader of the group. Check the operands of all the | |
5365 | stmts of the group. */ | |
e14c1050 | 5366 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
5367 | while (next_stmt) |
5368 | { | |
5369 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5370 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 5371 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) |
ebfd146a | 5372 | { |
73fbfcad | 5373 | if (dump_enabled_p ()) |
78c60e3d | 5374 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5375 | "use not simple.\n"); |
ebfd146a IR |
5376 | return false; |
5377 | } | |
e14c1050 | 5378 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5379 | } |
5380 | } | |
5381 | } | |
5382 | ||
3bab6342 AT |
5383 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5384 | { | |
355fe088 | 5385 | gimple *def_stmt; |
3bab6342 AT |
5386 | scatter_decl = vect_check_gather_scatter (stmt, loop_vinfo, &scatter_base, |
5387 | &scatter_off, &scatter_scale); | |
5388 | gcc_assert (scatter_decl); | |
81c40241 RB |
5389 | if (!vect_is_simple_use (scatter_off, vinfo, &def_stmt, &scatter_idx_dt, |
5390 | &scatter_off_vectype)) | |
3bab6342 AT |
5391 | { |
5392 | if (dump_enabled_p ()) | |
5393 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5394 | "scatter index use not simple."); | |
5395 | return false; | |
5396 | } | |
5397 | } | |
5398 | ||
ebfd146a IR |
5399 | if (!vec_stmt) /* transformation not required. */ |
5400 | { | |
5401 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
2e8ab70c RB |
5402 | /* The SLP costs are calculated during SLP analysis. */ |
5403 | if (!PURE_SLP_STMT (stmt_info)) | |
5404 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, | |
5405 | NULL, NULL, NULL); | |
ebfd146a IR |
5406 | return true; |
5407 | } | |
5408 | ||
5409 | /** Transform. **/ | |
5410 | ||
c716e67f XDL |
5411 | ensure_base_align (stmt_info, dr); |
5412 | ||
3bab6342 AT |
5413 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5414 | { | |
5415 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, op, src; | |
5416 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (scatter_decl)); | |
5417 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
5418 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
5419 | edge pe = loop_preheader_edge (loop); | |
5420 | gimple_seq seq; | |
5421 | basic_block new_bb; | |
5422 | enum { NARROW, NONE, WIDEN } modifier; | |
5423 | int scatter_off_nunits = TYPE_VECTOR_SUBPARTS (scatter_off_vectype); | |
5424 | ||
5425 | if (nunits == (unsigned int) scatter_off_nunits) | |
5426 | modifier = NONE; | |
5427 | else if (nunits == (unsigned int) scatter_off_nunits / 2) | |
5428 | { | |
5429 | unsigned char *sel = XALLOCAVEC (unsigned char, scatter_off_nunits); | |
5430 | modifier = WIDEN; | |
5431 | ||
5432 | for (i = 0; i < (unsigned int) scatter_off_nunits; ++i) | |
5433 | sel[i] = i | nunits; | |
5434 | ||
5435 | perm_mask = vect_gen_perm_mask_checked (scatter_off_vectype, sel); | |
5436 | gcc_assert (perm_mask != NULL_TREE); | |
5437 | } | |
5438 | else if (nunits == (unsigned int) scatter_off_nunits * 2) | |
5439 | { | |
5440 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
5441 | modifier = NARROW; | |
5442 | ||
5443 | for (i = 0; i < (unsigned int) nunits; ++i) | |
5444 | sel[i] = i | scatter_off_nunits; | |
5445 | ||
5446 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); | |
5447 | gcc_assert (perm_mask != NULL_TREE); | |
5448 | ncopies *= 2; | |
5449 | } | |
5450 | else | |
5451 | gcc_unreachable (); | |
5452 | ||
5453 | rettype = TREE_TYPE (TREE_TYPE (scatter_decl)); | |
5454 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5455 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5456 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5457 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5458 | scaletype = TREE_VALUE (arglist); | |
5459 | ||
5460 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
5461 | && TREE_CODE (rettype) == VOID_TYPE); | |
5462 | ||
5463 | ptr = fold_convert (ptrtype, scatter_base); | |
5464 | if (!is_gimple_min_invariant (ptr)) | |
5465 | { | |
5466 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
5467 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
5468 | gcc_assert (!new_bb); | |
5469 | } | |
5470 | ||
5471 | /* Currently we support only unconditional scatter stores, | |
5472 | so mask should be all ones. */ | |
5473 | mask = build_int_cst (masktype, -1); | |
5474 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
5475 | ||
5476 | scale = build_int_cst (scaletype, scatter_scale); | |
5477 | ||
5478 | prev_stmt_info = NULL; | |
5479 | for (j = 0; j < ncopies; ++j) | |
5480 | { | |
5481 | if (j == 0) | |
5482 | { | |
5483 | src = vec_oprnd1 | |
81c40241 | 5484 | = vect_get_vec_def_for_operand (gimple_assign_rhs1 (stmt), stmt); |
3bab6342 | 5485 | op = vec_oprnd0 |
81c40241 | 5486 | = vect_get_vec_def_for_operand (scatter_off, stmt); |
3bab6342 AT |
5487 | } |
5488 | else if (modifier != NONE && (j & 1)) | |
5489 | { | |
5490 | if (modifier == WIDEN) | |
5491 | { | |
5492 | src = vec_oprnd1 | |
5493 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5494 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, | |
5495 | stmt, gsi); | |
5496 | } | |
5497 | else if (modifier == NARROW) | |
5498 | { | |
5499 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
5500 | stmt, gsi); | |
5501 | op = vec_oprnd0 | |
5502 | = vect_get_vec_def_for_stmt_copy (scatter_idx_dt, vec_oprnd0); | |
5503 | } | |
5504 | else | |
5505 | gcc_unreachable (); | |
5506 | } | |
5507 | else | |
5508 | { | |
5509 | src = vec_oprnd1 | |
5510 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5511 | op = vec_oprnd0 | |
5512 | = vect_get_vec_def_for_stmt_copy (scatter_idx_dt, vec_oprnd0); | |
5513 | } | |
5514 | ||
5515 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
5516 | { | |
5517 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)) | |
5518 | == TYPE_VECTOR_SUBPARTS (srctype)); | |
0e22bb5a | 5519 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 AT |
5520 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
5521 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
5522 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5523 | src = var; | |
5524 | } | |
5525 | ||
5526 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
5527 | { | |
5528 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
5529 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 5530 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 AT |
5531 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
5532 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
5533 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5534 | op = var; | |
5535 | } | |
5536 | ||
5537 | new_stmt | |
5538 | = gimple_build_call (scatter_decl, 5, ptr, mask, op, src, scale); | |
5539 | ||
5540 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5541 | ||
5542 | if (prev_stmt_info == NULL) | |
5543 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5544 | else | |
5545 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5546 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5547 | } | |
5548 | return true; | |
5549 | } | |
5550 | ||
0d0293ac | 5551 | if (grouped_store) |
ebfd146a IR |
5552 | { |
5553 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 5554 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 5555 | |
e14c1050 | 5556 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
5557 | |
5558 | /* FORNOW */ | |
a70d6342 | 5559 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
5560 | |
5561 | /* We vectorize all the stmts of the interleaving group when we | |
5562 | reach the last stmt in the group. */ | |
e14c1050 IR |
5563 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
5564 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
5565 | && !slp) |
5566 | { | |
5567 | *vec_stmt = NULL; | |
5568 | return true; | |
5569 | } | |
5570 | ||
5571 | if (slp) | |
4b5caab7 | 5572 | { |
0d0293ac | 5573 | grouped_store = false; |
4b5caab7 IR |
5574 | /* VEC_NUM is the number of vect stmts to be created for this |
5575 | group. */ | |
5576 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 5577 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
52eab378 | 5578 | gcc_assert (GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt)) == first_stmt); |
4b5caab7 | 5579 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 5580 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 5581 | } |
ebfd146a | 5582 | else |
4b5caab7 IR |
5583 | /* VEC_NUM is the number of vect stmts to be created for this |
5584 | group. */ | |
ebfd146a IR |
5585 | vec_num = group_size; |
5586 | } | |
b8698a0f | 5587 | else |
ebfd146a IR |
5588 | { |
5589 | first_stmt = stmt; | |
5590 | first_dr = dr; | |
5591 | group_size = vec_num = 1; | |
ebfd146a | 5592 | } |
b8698a0f | 5593 | |
73fbfcad | 5594 | if (dump_enabled_p ()) |
78c60e3d | 5595 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5596 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 5597 | |
f2e2a985 MM |
5598 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
5599 | { | |
5600 | gimple_stmt_iterator incr_gsi; | |
5601 | bool insert_after; | |
355fe088 | 5602 | gimple *incr; |
f2e2a985 MM |
5603 | tree offvar; |
5604 | tree ivstep; | |
5605 | tree running_off; | |
5606 | gimple_seq stmts = NULL; | |
5607 | tree stride_base, stride_step, alias_off; | |
5608 | tree vec_oprnd; | |
f502d50e | 5609 | unsigned int g; |
f2e2a985 MM |
5610 | |
5611 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); | |
5612 | ||
5613 | stride_base | |
5614 | = fold_build_pointer_plus | |
f502d50e | 5615 | (unshare_expr (DR_BASE_ADDRESS (first_dr)), |
f2e2a985 | 5616 | size_binop (PLUS_EXPR, |
f502d50e MM |
5617 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (first_dr))), |
5618 | convert_to_ptrofftype (DR_INIT(first_dr)))); | |
5619 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (first_dr))); | |
f2e2a985 MM |
5620 | |
5621 | /* For a store with loop-invariant (but other than power-of-2) | |
5622 | stride (i.e. not a grouped access) like so: | |
5623 | ||
5624 | for (i = 0; i < n; i += stride) | |
5625 | array[i] = ...; | |
5626 | ||
5627 | we generate a new induction variable and new stores from | |
5628 | the components of the (vectorized) rhs: | |
5629 | ||
5630 | for (j = 0; ; j += VF*stride) | |
5631 | vectemp = ...; | |
5632 | tmp1 = vectemp[0]; | |
5633 | array[j] = tmp1; | |
5634 | tmp2 = vectemp[1]; | |
5635 | array[j + stride] = tmp2; | |
5636 | ... | |
5637 | */ | |
5638 | ||
cee62fee MM |
5639 | unsigned nstores = nunits; |
5640 | tree ltype = elem_type; | |
5641 | if (slp) | |
5642 | { | |
5643 | nstores = nunits / group_size; | |
5644 | if (group_size < nunits) | |
5645 | ltype = build_vector_type (elem_type, group_size); | |
5646 | else | |
5647 | ltype = vectype; | |
5648 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); | |
5649 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
f502d50e | 5650 | group_size = 1; |
cee62fee MM |
5651 | } |
5652 | ||
f2e2a985 MM |
5653 | ivstep = stride_step; |
5654 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
5655 | build_int_cst (TREE_TYPE (ivstep), | |
cee62fee | 5656 | ncopies * nstores)); |
f2e2a985 MM |
5657 | |
5658 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
5659 | ||
5660 | create_iv (stride_base, ivstep, NULL, | |
5661 | loop, &incr_gsi, insert_after, | |
5662 | &offvar, NULL); | |
5663 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 5664 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
f2e2a985 MM |
5665 | |
5666 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
5667 | if (stmts) | |
5668 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
5669 | ||
5670 | prev_stmt_info = NULL; | |
f502d50e MM |
5671 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (first_dr)), 0); |
5672 | next_stmt = first_stmt; | |
5673 | for (g = 0; g < group_size; g++) | |
f2e2a985 | 5674 | { |
f502d50e MM |
5675 | running_off = offvar; |
5676 | if (g) | |
f2e2a985 | 5677 | { |
f502d50e MM |
5678 | tree size = TYPE_SIZE_UNIT (ltype); |
5679 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 5680 | size); |
f502d50e | 5681 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 5682 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 5683 | running_off, pos); |
f2e2a985 | 5684 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 5685 | running_off = newoff; |
f502d50e MM |
5686 | } |
5687 | for (j = 0; j < ncopies; j++) | |
5688 | { | |
5689 | /* We've set op and dt above, from gimple_assign_rhs1(stmt), | |
5690 | and first_stmt == stmt. */ | |
5691 | if (j == 0) | |
5692 | { | |
5693 | if (slp) | |
5694 | { | |
5695 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
5696 | slp_node, -1); | |
5697 | vec_oprnd = vec_oprnds[0]; | |
5698 | } | |
5699 | else | |
5700 | { | |
5701 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5702 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 5703 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
f502d50e MM |
5704 | } |
5705 | } | |
f2e2a985 | 5706 | else |
f502d50e MM |
5707 | { |
5708 | if (slp) | |
5709 | vec_oprnd = vec_oprnds[j]; | |
5710 | else | |
c079cbac | 5711 | { |
81c40241 | 5712 | vect_is_simple_use (vec_oprnd, vinfo, &def_stmt, &dt); |
c079cbac RB |
5713 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); |
5714 | } | |
f502d50e MM |
5715 | } |
5716 | ||
5717 | for (i = 0; i < nstores; i++) | |
5718 | { | |
5719 | tree newref, newoff; | |
355fe088 | 5720 | gimple *incr, *assign; |
f502d50e MM |
5721 | tree size = TYPE_SIZE (ltype); |
5722 | /* Extract the i'th component. */ | |
5723 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
5724 | bitsize_int (i), size); | |
5725 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
5726 | size, pos); | |
5727 | ||
5728 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
5729 | NULL_TREE, true, | |
5730 | GSI_SAME_STMT); | |
5731 | ||
5732 | newref = build2 (MEM_REF, ltype, | |
5733 | running_off, alias_off); | |
5734 | ||
5735 | /* And store it to *running_off. */ | |
5736 | assign = gimple_build_assign (newref, elem); | |
5737 | vect_finish_stmt_generation (stmt, assign, gsi); | |
5738 | ||
5739 | newoff = copy_ssa_name (running_off, NULL); | |
5740 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
5741 | running_off, stride_step); | |
5742 | vect_finish_stmt_generation (stmt, incr, gsi); | |
5743 | ||
5744 | running_off = newoff; | |
225ce44b RB |
5745 | if (g == group_size - 1 |
5746 | && !slp) | |
f502d50e MM |
5747 | { |
5748 | if (j == 0 && i == 0) | |
225ce44b RB |
5749 | STMT_VINFO_VEC_STMT (stmt_info) |
5750 | = *vec_stmt = assign; | |
f502d50e MM |
5751 | else |
5752 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign; | |
5753 | prev_stmt_info = vinfo_for_stmt (assign); | |
5754 | } | |
5755 | } | |
f2e2a985 | 5756 | } |
f502d50e | 5757 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
f2e2a985 MM |
5758 | } |
5759 | return true; | |
5760 | } | |
5761 | ||
9771b263 DN |
5762 | dr_chain.create (group_size); |
5763 | oprnds.create (group_size); | |
ebfd146a | 5764 | |
720f5239 | 5765 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 5766 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
5767 | /* Targets with store-lane instructions must not require explicit |
5768 | realignment. */ | |
5769 | gcc_assert (!store_lanes_p | |
5770 | || alignment_support_scheme == dr_aligned | |
5771 | || alignment_support_scheme == dr_unaligned_supported); | |
5772 | ||
09dfa495 BM |
5773 | if (negative) |
5774 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
5775 | ||
272c6793 RS |
5776 | if (store_lanes_p) |
5777 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
5778 | else | |
5779 | aggr_type = vectype; | |
ebfd146a IR |
5780 | |
5781 | /* In case the vectorization factor (VF) is bigger than the number | |
5782 | of elements that we can fit in a vectype (nunits), we have to generate | |
5783 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 5784 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
5785 | vect_get_vec_def_for_copy_stmt. */ |
5786 | ||
0d0293ac | 5787 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
5788 | |
5789 | S1: &base + 2 = x2 | |
5790 | S2: &base = x0 | |
5791 | S3: &base + 1 = x1 | |
5792 | S4: &base + 3 = x3 | |
5793 | ||
5794 | We create vectorized stores starting from base address (the access of the | |
5795 | first stmt in the chain (S2 in the above example), when the last store stmt | |
5796 | of the chain (S4) is reached: | |
5797 | ||
5798 | VS1: &base = vx2 | |
5799 | VS2: &base + vec_size*1 = vx0 | |
5800 | VS3: &base + vec_size*2 = vx1 | |
5801 | VS4: &base + vec_size*3 = vx3 | |
5802 | ||
5803 | Then permutation statements are generated: | |
5804 | ||
3fcc1b55 JJ |
5805 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
5806 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 5807 | ... |
b8698a0f | 5808 | |
ebfd146a IR |
5809 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
5810 | (the order of the data-refs in the output of vect_permute_store_chain | |
5811 | corresponds to the order of scalar stmts in the interleaving chain - see | |
5812 | the documentation of vect_permute_store_chain()). | |
5813 | ||
5814 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 5815 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 5816 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 5817 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
5818 | */ |
5819 | ||
5820 | prev_stmt_info = NULL; | |
5821 | for (j = 0; j < ncopies; j++) | |
5822 | { | |
ebfd146a IR |
5823 | |
5824 | if (j == 0) | |
5825 | { | |
5826 | if (slp) | |
5827 | { | |
5828 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c IR |
5829 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
5830 | NULL, slp_node, -1); | |
ebfd146a | 5831 | |
9771b263 | 5832 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
5833 | } |
5834 | else | |
5835 | { | |
b8698a0f L |
5836 | /* For interleaved stores we collect vectorized defs for all the |
5837 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
5838 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
5839 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
5840 | ||
0d0293ac | 5841 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 5842 | OPRNDS are of size 1. */ |
b8698a0f | 5843 | next_stmt = first_stmt; |
ebfd146a IR |
5844 | for (i = 0; i < group_size; i++) |
5845 | { | |
b8698a0f L |
5846 | /* Since gaps are not supported for interleaved stores, |
5847 | GROUP_SIZE is the exact number of stmts in the chain. | |
5848 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
5849 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
5850 | iteration of the loop will be executed. */ |
5851 | gcc_assert (next_stmt | |
5852 | && gimple_assign_single_p (next_stmt)); | |
5853 | op = gimple_assign_rhs1 (next_stmt); | |
5854 | ||
81c40241 | 5855 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
9771b263 DN |
5856 | dr_chain.quick_push (vec_oprnd); |
5857 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 5858 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5859 | } |
5860 | } | |
5861 | ||
5862 | /* We should have catched mismatched types earlier. */ | |
5863 | gcc_assert (useless_type_conversion_p (vectype, | |
5864 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
5865 | bool simd_lane_access_p |
5866 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
5867 | if (simd_lane_access_p | |
5868 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
5869 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
5870 | && integer_zerop (DR_OFFSET (first_dr)) | |
5871 | && integer_zerop (DR_INIT (first_dr)) | |
5872 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
5873 | get_alias_set (DR_REF (first_dr)))) | |
5874 | { | |
5875 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
5876 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
5877 | (DR_REF (first_dr)), 0); | |
8928eff3 | 5878 | inv_p = false; |
74bf76ed JJ |
5879 | } |
5880 | else | |
5881 | dataref_ptr | |
5882 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
5883 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 5884 | offset, &dummy, gsi, &ptr_incr, |
74bf76ed | 5885 | simd_lane_access_p, &inv_p); |
a70d6342 | 5886 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 5887 | } |
b8698a0f | 5888 | else |
ebfd146a | 5889 | { |
b8698a0f L |
5890 | /* For interleaved stores we created vectorized defs for all the |
5891 | defs stored in OPRNDS in the previous iteration (previous copy). | |
5892 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
5893 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
5894 | next copy. | |
0d0293ac | 5895 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
5896 | OPRNDS are of size 1. */ |
5897 | for (i = 0; i < group_size; i++) | |
5898 | { | |
9771b263 | 5899 | op = oprnds[i]; |
81c40241 | 5900 | vect_is_simple_use (op, vinfo, &def_stmt, &dt); |
b8698a0f | 5901 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
5902 | dr_chain[i] = vec_oprnd; |
5903 | oprnds[i] = vec_oprnd; | |
ebfd146a | 5904 | } |
74bf76ed JJ |
5905 | if (dataref_offset) |
5906 | dataref_offset | |
5907 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
5908 | TYPE_SIZE_UNIT (aggr_type)); | |
5909 | else | |
5910 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
5911 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
5912 | } |
5913 | ||
272c6793 | 5914 | if (store_lanes_p) |
ebfd146a | 5915 | { |
272c6793 | 5916 | tree vec_array; |
267d3070 | 5917 | |
272c6793 RS |
5918 | /* Combine all the vectors into an array. */ |
5919 | vec_array = create_vector_array (vectype, vec_num); | |
5920 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 5921 | { |
9771b263 | 5922 | vec_oprnd = dr_chain[i]; |
272c6793 | 5923 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 5924 | } |
b8698a0f | 5925 | |
272c6793 RS |
5926 | /* Emit: |
5927 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
5928 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
5929 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
5930 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 | 5931 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
5932 | } |
5933 | else | |
5934 | { | |
5935 | new_stmt = NULL; | |
0d0293ac | 5936 | if (grouped_store) |
272c6793 | 5937 | { |
b6b9227d JJ |
5938 | if (j == 0) |
5939 | result_chain.create (group_size); | |
272c6793 RS |
5940 | /* Permute. */ |
5941 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
5942 | &result_chain); | |
5943 | } | |
c2d7ab2a | 5944 | |
272c6793 RS |
5945 | next_stmt = first_stmt; |
5946 | for (i = 0; i < vec_num; i++) | |
5947 | { | |
644ffefd | 5948 | unsigned align, misalign; |
272c6793 RS |
5949 | |
5950 | if (i > 0) | |
5951 | /* Bump the vector pointer. */ | |
5952 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5953 | stmt, NULL_TREE); | |
5954 | ||
5955 | if (slp) | |
9771b263 | 5956 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
5957 | else if (grouped_store) |
5958 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 5959 | vect_permute_store_chain(). */ |
9771b263 | 5960 | vec_oprnd = result_chain[i]; |
272c6793 | 5961 | |
aed93b23 RB |
5962 | data_ref = fold_build2 (MEM_REF, TREE_TYPE (vec_oprnd), |
5963 | dataref_ptr, | |
5964 | dataref_offset | |
5965 | ? dataref_offset | |
5966 | : build_int_cst (reference_alias_ptr_type | |
5967 | (DR_REF (first_dr)), 0)); | |
644ffefd | 5968 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 | 5969 | if (aligned_access_p (first_dr)) |
644ffefd | 5970 | misalign = 0; |
272c6793 RS |
5971 | else if (DR_MISALIGNMENT (first_dr) == -1) |
5972 | { | |
52639a61 RB |
5973 | if (DR_VECT_AUX (first_dr)->base_element_aligned) |
5974 | align = TYPE_ALIGN_UNIT (elem_type); | |
5975 | else | |
5976 | align = get_object_alignment (DR_REF (first_dr)) | |
5977 | / BITS_PER_UNIT; | |
5978 | misalign = 0; | |
272c6793 RS |
5979 | TREE_TYPE (data_ref) |
5980 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 5981 | align * BITS_PER_UNIT); |
272c6793 RS |
5982 | } |
5983 | else | |
5984 | { | |
5985 | TREE_TYPE (data_ref) | |
5986 | = build_aligned_type (TREE_TYPE (data_ref), | |
5987 | TYPE_ALIGN (elem_type)); | |
644ffefd | 5988 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 5989 | } |
aed93b23 RB |
5990 | if (dataref_offset == NULL_TREE |
5991 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
5992 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
5993 | misalign); | |
c2d7ab2a | 5994 | |
f234d260 BM |
5995 | if (negative |
5996 | && dt != vect_constant_def | |
5997 | && dt != vect_external_def) | |
09dfa495 BM |
5998 | { |
5999 | tree perm_mask = perm_mask_for_reverse (vectype); | |
6000 | tree perm_dest | |
6001 | = vect_create_destination_var (gimple_assign_rhs1 (stmt), | |
6002 | vectype); | |
b731b390 | 6003 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
6004 | |
6005 | /* Generate the permute statement. */ | |
355fe088 | 6006 | gimple *perm_stmt |
0d0e4a03 JJ |
6007 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
6008 | vec_oprnd, perm_mask); | |
09dfa495 BM |
6009 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6010 | ||
6011 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
6012 | vec_oprnd = new_temp; | |
6013 | } | |
6014 | ||
272c6793 RS |
6015 | /* Arguments are ready. Create the new vector stmt. */ |
6016 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
6017 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
6018 | |
6019 | if (slp) | |
6020 | continue; | |
6021 | ||
e14c1050 | 6022 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
6023 | if (!next_stmt) |
6024 | break; | |
6025 | } | |
ebfd146a | 6026 | } |
1da0876c RS |
6027 | if (!slp) |
6028 | { | |
6029 | if (j == 0) | |
6030 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6031 | else | |
6032 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6033 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6034 | } | |
ebfd146a IR |
6035 | } |
6036 | ||
9771b263 DN |
6037 | dr_chain.release (); |
6038 | oprnds.release (); | |
6039 | result_chain.release (); | |
6040 | vec_oprnds.release (); | |
ebfd146a IR |
6041 | |
6042 | return true; | |
6043 | } | |
6044 | ||
557be5a8 AL |
6045 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
6046 | VECTOR_CST mask. No checks are made that the target platform supports the | |
6047 | mask, so callers may wish to test can_vec_perm_p separately, or use | |
6048 | vect_gen_perm_mask_checked. */ | |
a1e53f3f | 6049 | |
3fcc1b55 | 6050 | tree |
557be5a8 | 6051 | vect_gen_perm_mask_any (tree vectype, const unsigned char *sel) |
a1e53f3f | 6052 | { |
d2a12ae7 | 6053 | tree mask_elt_type, mask_type, mask_vec, *mask_elts; |
2635892a | 6054 | int i, nunits; |
a1e53f3f | 6055 | |
22e4dee7 | 6056 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 | 6057 | |
96f9265a RG |
6058 | mask_elt_type = lang_hooks.types.type_for_mode |
6059 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
22e4dee7 | 6060 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 6061 | |
d2a12ae7 | 6062 | mask_elts = XALLOCAVEC (tree, nunits); |
aec7ae7d | 6063 | for (i = nunits - 1; i >= 0; i--) |
d2a12ae7 RG |
6064 | mask_elts[i] = build_int_cst (mask_elt_type, sel[i]); |
6065 | mask_vec = build_vector (mask_type, mask_elts); | |
a1e53f3f | 6066 | |
2635892a | 6067 | return mask_vec; |
a1e53f3f L |
6068 | } |
6069 | ||
cf7aa6a3 AL |
6070 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_p, |
6071 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ | |
557be5a8 AL |
6072 | |
6073 | tree | |
6074 | vect_gen_perm_mask_checked (tree vectype, const unsigned char *sel) | |
6075 | { | |
6076 | gcc_assert (can_vec_perm_p (TYPE_MODE (vectype), false, sel)); | |
6077 | return vect_gen_perm_mask_any (vectype, sel); | |
6078 | } | |
6079 | ||
aec7ae7d JJ |
6080 | /* Given a vector variable X and Y, that was generated for the scalar |
6081 | STMT, generate instructions to permute the vector elements of X and Y | |
6082 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
6083 | permuted vector variable. */ | |
a1e53f3f L |
6084 | |
6085 | static tree | |
355fe088 | 6086 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 6087 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
6088 | { |
6089 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 6090 | tree perm_dest, data_ref; |
355fe088 | 6091 | gimple *perm_stmt; |
a1e53f3f | 6092 | |
acdcd61b | 6093 | perm_dest = vect_create_destination_var (gimple_get_lhs (stmt), vectype); |
b731b390 | 6094 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
6095 | |
6096 | /* Generate the permute statement. */ | |
0d0e4a03 | 6097 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
6098 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6099 | ||
6100 | return data_ref; | |
6101 | } | |
6102 | ||
6b916b36 RB |
6103 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
6104 | inserting them on the loops preheader edge. Returns true if we | |
6105 | were successful in doing so (and thus STMT can be moved then), | |
6106 | otherwise returns false. */ | |
6107 | ||
6108 | static bool | |
355fe088 | 6109 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
6110 | { |
6111 | ssa_op_iter i; | |
6112 | tree op; | |
6113 | bool any = false; | |
6114 | ||
6115 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6116 | { | |
355fe088 | 6117 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6118 | if (!gimple_nop_p (def_stmt) |
6119 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6120 | { | |
6121 | /* Make sure we don't need to recurse. While we could do | |
6122 | so in simple cases when there are more complex use webs | |
6123 | we don't have an easy way to preserve stmt order to fulfil | |
6124 | dependencies within them. */ | |
6125 | tree op2; | |
6126 | ssa_op_iter i2; | |
d1417442 JJ |
6127 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
6128 | return false; | |
6b916b36 RB |
6129 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
6130 | { | |
355fe088 | 6131 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
6132 | if (!gimple_nop_p (def_stmt2) |
6133 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
6134 | return false; | |
6135 | } | |
6136 | any = true; | |
6137 | } | |
6138 | } | |
6139 | ||
6140 | if (!any) | |
6141 | return true; | |
6142 | ||
6143 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6144 | { | |
355fe088 | 6145 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6146 | if (!gimple_nop_p (def_stmt) |
6147 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6148 | { | |
6149 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
6150 | gsi_remove (&gsi, false); | |
6151 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
6152 | } | |
6153 | } | |
6154 | ||
6155 | return true; | |
6156 | } | |
6157 | ||
ebfd146a IR |
6158 | /* vectorizable_load. |
6159 | ||
b8698a0f L |
6160 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
6161 | can be vectorized. | |
6162 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6163 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6164 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6165 | ||
6166 | static bool | |
355fe088 | 6167 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 6168 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
6169 | { |
6170 | tree scalar_dest; | |
6171 | tree vec_dest = NULL; | |
6172 | tree data_ref = NULL; | |
6173 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 6174 | stmt_vec_info prev_stmt_info; |
ebfd146a | 6175 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6176 | struct loop *loop = NULL; |
ebfd146a | 6177 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 6178 | bool nested_in_vect_loop = false; |
c716e67f | 6179 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 6180 | tree elem_type; |
ebfd146a | 6181 | tree new_temp; |
ef4bddc2 | 6182 | machine_mode mode; |
355fe088 | 6183 | gimple *new_stmt = NULL; |
ebfd146a IR |
6184 | tree dummy; |
6185 | enum dr_alignment_support alignment_support_scheme; | |
6186 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6187 | tree dataref_offset = NULL_TREE; |
355fe088 | 6188 | gimple *ptr_incr = NULL; |
ebfd146a | 6189 | int ncopies; |
9b999e8c | 6190 | int i, j, group_size = -1, group_gap_adj; |
ebfd146a IR |
6191 | tree msq = NULL_TREE, lsq; |
6192 | tree offset = NULL_TREE; | |
356bbc4c | 6193 | tree byte_offset = NULL_TREE; |
ebfd146a | 6194 | tree realignment_token = NULL_TREE; |
538dd0b7 | 6195 | gphi *phi = NULL; |
6e1aa848 | 6196 | vec<tree> dr_chain = vNULL; |
0d0293ac | 6197 | bool grouped_load = false; |
272c6793 | 6198 | bool load_lanes_p = false; |
355fe088 | 6199 | gimple *first_stmt; |
4f0a0218 | 6200 | gimple *first_stmt_for_drptr = NULL; |
ebfd146a | 6201 | bool inv_p; |
319e6439 | 6202 | bool negative = false; |
ebfd146a IR |
6203 | bool compute_in_loop = false; |
6204 | struct loop *at_loop; | |
6205 | int vec_num; | |
6206 | bool slp = (slp_node != NULL); | |
6207 | bool slp_perm = false; | |
6208 | enum tree_code code; | |
a70d6342 IR |
6209 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
6210 | int vf; | |
272c6793 | 6211 | tree aggr_type; |
aec7ae7d JJ |
6212 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; |
6213 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
6214 | int gather_scale = 1; | |
6215 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
310213d4 | 6216 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 6217 | |
465c8c19 JJ |
6218 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
6219 | return false; | |
6220 | ||
66c16fd9 RB |
6221 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6222 | && ! vec_stmt) | |
465c8c19 JJ |
6223 | return false; |
6224 | ||
6225 | /* Is vectorizable load? */ | |
6226 | if (!is_gimple_assign (stmt)) | |
6227 | return false; | |
6228 | ||
6229 | scalar_dest = gimple_assign_lhs (stmt); | |
6230 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
6231 | return false; | |
6232 | ||
6233 | code = gimple_assign_rhs_code (stmt); | |
6234 | if (code != ARRAY_REF | |
6235 | && code != BIT_FIELD_REF | |
6236 | && code != INDIRECT_REF | |
6237 | && code != COMPONENT_REF | |
6238 | && code != IMAGPART_EXPR | |
6239 | && code != REALPART_EXPR | |
6240 | && code != MEM_REF | |
6241 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
6242 | return false; | |
6243 | ||
6244 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
6245 | return false; | |
6246 | ||
6247 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
6248 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
6249 | ||
a70d6342 IR |
6250 | if (loop_vinfo) |
6251 | { | |
6252 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6253 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
6254 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6255 | } | |
6256 | else | |
3533e503 | 6257 | vf = 1; |
ebfd146a IR |
6258 | |
6259 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 6260 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 6261 | case of SLP. */ |
fce57248 | 6262 | if (slp) |
ebfd146a IR |
6263 | ncopies = 1; |
6264 | else | |
6265 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
6266 | ||
6267 | gcc_assert (ncopies >= 1); | |
6268 | ||
6269 | /* FORNOW. This restriction should be relaxed. */ | |
6270 | if (nested_in_vect_loop && ncopies > 1) | |
6271 | { | |
73fbfcad | 6272 | if (dump_enabled_p ()) |
78c60e3d | 6273 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6274 | "multiple types in nested loop.\n"); |
ebfd146a IR |
6275 | return false; |
6276 | } | |
6277 | ||
f2556b68 RB |
6278 | /* Invalidate assumptions made by dependence analysis when vectorization |
6279 | on the unrolled body effectively re-orders stmts. */ | |
6280 | if (ncopies > 1 | |
6281 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6282 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6283 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6284 | { | |
6285 | if (dump_enabled_p ()) | |
6286 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6287 | "cannot perform implicit CSE when unrolling " | |
6288 | "with negative dependence distance\n"); | |
6289 | return false; | |
6290 | } | |
6291 | ||
7b7b1813 | 6292 | elem_type = TREE_TYPE (vectype); |
947131ba | 6293 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
6294 | |
6295 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6296 | (e.g. - data copies). */ | |
947131ba | 6297 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 6298 | { |
73fbfcad | 6299 | if (dump_enabled_p ()) |
78c60e3d | 6300 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6301 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
6302 | return false; |
6303 | } | |
6304 | ||
ebfd146a | 6305 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 6306 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 6307 | { |
0d0293ac | 6308 | grouped_load = true; |
ebfd146a | 6309 | /* FORNOW */ |
3bab6342 | 6310 | gcc_assert (!nested_in_vect_loop && !STMT_VINFO_GATHER_SCATTER_P (stmt_info)); |
ebfd146a | 6311 | |
e14c1050 | 6312 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
d3465d72 RS |
6313 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
6314 | ||
fce57248 | 6315 | if (!slp && !STMT_VINFO_STRIDED_P (stmt_info)) |
d3465d72 RS |
6316 | { |
6317 | if (vect_load_lanes_supported (vectype, group_size)) | |
6318 | load_lanes_p = true; | |
6319 | else if (!vect_grouped_load_supported (vectype, group_size)) | |
6320 | return false; | |
6321 | } | |
d5f035ea RB |
6322 | |
6323 | /* If this is single-element interleaving with an element distance | |
6324 | that leaves unused vector loads around punt - we at least create | |
6325 | very sub-optimal code in that case (and blow up memory, | |
6326 | see PR65518). */ | |
6327 | if (first_stmt == stmt | |
72c0f643 RB |
6328 | && !GROUP_NEXT_ELEMENT (stmt_info)) |
6329 | { | |
6330 | if (GROUP_SIZE (stmt_info) > TYPE_VECTOR_SUBPARTS (vectype)) | |
6331 | { | |
6332 | if (dump_enabled_p ()) | |
6333 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6334 | "single-element interleaving not supported " | |
6335 | "for not adjacent vector loads\n"); | |
6336 | return false; | |
6337 | } | |
6338 | ||
6339 | /* Single-element interleaving requires peeling for gaps. */ | |
836dbb1a | 6340 | gcc_assert (GROUP_GAP (stmt_info)); |
72c0f643 RB |
6341 | } |
6342 | ||
6343 | /* If there is a gap in the end of the group or the group size cannot | |
6344 | be made a multiple of the vector element count then we access excess | |
6345 | elements in the last iteration and thus need to peel that off. */ | |
6346 | if (loop_vinfo | |
6347 | && ! STMT_VINFO_STRIDED_P (stmt_info) | |
836dbb1a | 6348 | && (GROUP_GAP (vinfo_for_stmt (first_stmt)) != 0 |
d3465d72 | 6349 | || (!slp && !load_lanes_p && vf % group_size != 0))) |
d5f035ea RB |
6350 | { |
6351 | if (dump_enabled_p ()) | |
6352 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
72c0f643 RB |
6353 | "Data access with gaps requires scalar " |
6354 | "epilogue loop\n"); | |
6355 | if (loop->inner) | |
6356 | { | |
6357 | if (dump_enabled_p ()) | |
6358 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6359 | "Peeling for outer loop is not supported\n"); | |
6360 | return false; | |
6361 | } | |
6362 | ||
6363 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; | |
d5f035ea RB |
6364 | } |
6365 | ||
b1af7da6 RB |
6366 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
6367 | slp_perm = true; | |
6368 | ||
47d3fdb2 RB |
6369 | /* ??? The following is overly pessimistic (as well as the loop |
6370 | case above) in the case we can statically determine the excess | |
6371 | elements loaded are within the bounds of a decl that is accessed. | |
6372 | Likewise for BB vectorizations using masked loads is a possibility. */ | |
6373 | if (bb_vinfo && slp_perm && group_size % nunits != 0) | |
6374 | { | |
6375 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6376 | "BB vectorization with gaps at the end of a load " | |
6377 | "is not supported\n"); | |
6378 | return false; | |
6379 | } | |
6380 | ||
f2556b68 RB |
6381 | /* Invalidate assumptions made by dependence analysis when vectorization |
6382 | on the unrolled body effectively re-orders stmts. */ | |
6383 | if (!PURE_SLP_STMT (stmt_info) | |
6384 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6385 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6386 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6387 | { | |
6388 | if (dump_enabled_p ()) | |
6389 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6390 | "cannot perform implicit CSE when performing " | |
6391 | "group loads with negative dependence distance\n"); | |
6392 | return false; | |
6393 | } | |
96bb56b2 RB |
6394 | |
6395 | /* Similarly when the stmt is a load that is both part of a SLP | |
6396 | instance and a loop vectorized stmt via the same-dr mechanism | |
6397 | we have to give up. */ | |
6398 | if (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info) | |
6399 | && (STMT_SLP_TYPE (stmt_info) | |
6400 | != STMT_SLP_TYPE (vinfo_for_stmt | |
6401 | (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info))))) | |
6402 | { | |
6403 | if (dump_enabled_p ()) | |
6404 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6405 | "conflicting SLP types for CSEd load\n"); | |
6406 | return false; | |
6407 | } | |
ebfd146a IR |
6408 | } |
6409 | ||
a1e53f3f | 6410 | |
3bab6342 | 6411 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
aec7ae7d | 6412 | { |
355fe088 | 6413 | gimple *def_stmt; |
3bab6342 AT |
6414 | gather_decl = vect_check_gather_scatter (stmt, loop_vinfo, &gather_base, |
6415 | &gather_off, &gather_scale); | |
aec7ae7d | 6416 | gcc_assert (gather_decl); |
81c40241 RB |
6417 | if (!vect_is_simple_use (gather_off, vinfo, &def_stmt, &gather_dt, |
6418 | &gather_off_vectype)) | |
aec7ae7d | 6419 | { |
73fbfcad | 6420 | if (dump_enabled_p ()) |
78c60e3d | 6421 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6422 | "gather index use not simple.\n"); |
aec7ae7d JJ |
6423 | return false; |
6424 | } | |
6425 | } | |
f2e2a985 | 6426 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
7b5fc413 | 6427 | { |
fce57248 RS |
6428 | if (grouped_load |
6429 | && slp | |
7b5fc413 | 6430 | && (group_size > nunits |
b266b968 | 6431 | || nunits % group_size != 0)) |
7b5fc413 RB |
6432 | { |
6433 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6434 | "unhandled strided group load\n"); | |
6435 | return false; | |
6436 | } | |
6437 | } | |
319e6439 RG |
6438 | else |
6439 | { | |
6440 | negative = tree_int_cst_compare (nested_in_vect_loop | |
6441 | ? STMT_VINFO_DR_STEP (stmt_info) | |
6442 | : DR_STEP (dr), | |
6443 | size_zero_node) < 0; | |
6444 | if (negative && ncopies > 1) | |
6445 | { | |
73fbfcad | 6446 | if (dump_enabled_p ()) |
78c60e3d | 6447 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6448 | "multiple types with negative step.\n"); |
319e6439 RG |
6449 | return false; |
6450 | } | |
6451 | ||
6452 | if (negative) | |
6453 | { | |
08940f33 RB |
6454 | if (grouped_load) |
6455 | { | |
6456 | if (dump_enabled_p ()) | |
6457 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 TJ |
6458 | "negative step for group load not supported" |
6459 | "\n"); | |
08940f33 RB |
6460 | return false; |
6461 | } | |
319e6439 RG |
6462 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
6463 | if (alignment_support_scheme != dr_aligned | |
6464 | && alignment_support_scheme != dr_unaligned_supported) | |
6465 | { | |
73fbfcad | 6466 | if (dump_enabled_p ()) |
78c60e3d | 6467 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6468 | "negative step but alignment required.\n"); |
319e6439 RG |
6469 | return false; |
6470 | } | |
6471 | if (!perm_mask_for_reverse (vectype)) | |
6472 | { | |
73fbfcad | 6473 | if (dump_enabled_p ()) |
78c60e3d | 6474 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
6475 | "negative step and reversing not supported." |
6476 | "\n"); | |
319e6439 RG |
6477 | return false; |
6478 | } | |
6479 | } | |
7d75abc8 | 6480 | } |
aec7ae7d | 6481 | |
ebfd146a IR |
6482 | if (!vec_stmt) /* transformation not required. */ |
6483 | { | |
6484 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
2e8ab70c RB |
6485 | /* The SLP costs are calculated during SLP analysis. */ |
6486 | if (!PURE_SLP_STMT (stmt_info)) | |
6487 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, | |
6488 | NULL, NULL, NULL); | |
ebfd146a IR |
6489 | return true; |
6490 | } | |
6491 | ||
73fbfcad | 6492 | if (dump_enabled_p ()) |
78c60e3d | 6493 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6494 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a IR |
6495 | |
6496 | /** Transform. **/ | |
6497 | ||
c716e67f XDL |
6498 | ensure_base_align (stmt_info, dr); |
6499 | ||
3bab6342 | 6500 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
aec7ae7d JJ |
6501 | { |
6502 | tree vec_oprnd0 = NULL_TREE, op; | |
6503 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
6504 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
d3c2fee0 | 6505 | tree ptr, mask, var, scale, merge, perm_mask = NULL_TREE, prev_res = NULL_TREE; |
aec7ae7d JJ |
6506 | edge pe = loop_preheader_edge (loop); |
6507 | gimple_seq seq; | |
6508 | basic_block new_bb; | |
6509 | enum { NARROW, NONE, WIDEN } modifier; | |
6510 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
6511 | ||
6512 | if (nunits == gather_off_nunits) | |
6513 | modifier = NONE; | |
6514 | else if (nunits == gather_off_nunits / 2) | |
6515 | { | |
6516 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
6517 | modifier = WIDEN; | |
6518 | ||
6519 | for (i = 0; i < gather_off_nunits; ++i) | |
6520 | sel[i] = i | nunits; | |
6521 | ||
557be5a8 | 6522 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
aec7ae7d JJ |
6523 | } |
6524 | else if (nunits == gather_off_nunits * 2) | |
6525 | { | |
6526 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
6527 | modifier = NARROW; | |
6528 | ||
6529 | for (i = 0; i < nunits; ++i) | |
6530 | sel[i] = i < gather_off_nunits | |
6531 | ? i : i + nunits - gather_off_nunits; | |
6532 | ||
557be5a8 | 6533 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
aec7ae7d JJ |
6534 | ncopies *= 2; |
6535 | } | |
6536 | else | |
6537 | gcc_unreachable (); | |
6538 | ||
6539 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
6540 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6541 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6542 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6543 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6544 | scaletype = TREE_VALUE (arglist); | |
d3c2fee0 | 6545 | gcc_checking_assert (types_compatible_p (srctype, rettype)); |
aec7ae7d JJ |
6546 | |
6547 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
6548 | ||
6549 | ptr = fold_convert (ptrtype, gather_base); | |
6550 | if (!is_gimple_min_invariant (ptr)) | |
6551 | { | |
6552 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6553 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6554 | gcc_assert (!new_bb); | |
6555 | } | |
6556 | ||
6557 | /* Currently we support only unconditional gather loads, | |
6558 | so mask should be all ones. */ | |
d3c2fee0 AI |
6559 | if (TREE_CODE (masktype) == INTEGER_TYPE) |
6560 | mask = build_int_cst (masktype, -1); | |
6561 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
6562 | { | |
6563 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
6564 | mask = build_vector_from_val (masktype, mask); | |
03b9e8e4 | 6565 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
d3c2fee0 | 6566 | } |
aec7ae7d JJ |
6567 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) |
6568 | { | |
6569 | REAL_VALUE_TYPE r; | |
6570 | long tmp[6]; | |
6571 | for (j = 0; j < 6; ++j) | |
6572 | tmp[j] = -1; | |
6573 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
6574 | mask = build_real (TREE_TYPE (masktype), r); | |
d3c2fee0 | 6575 | mask = build_vector_from_val (masktype, mask); |
03b9e8e4 | 6576 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
aec7ae7d JJ |
6577 | } |
6578 | else | |
6579 | gcc_unreachable (); | |
aec7ae7d JJ |
6580 | |
6581 | scale = build_int_cst (scaletype, gather_scale); | |
6582 | ||
d3c2fee0 AI |
6583 | if (TREE_CODE (TREE_TYPE (rettype)) == INTEGER_TYPE) |
6584 | merge = build_int_cst (TREE_TYPE (rettype), 0); | |
6585 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (rettype))) | |
6586 | { | |
6587 | REAL_VALUE_TYPE r; | |
6588 | long tmp[6]; | |
6589 | for (j = 0; j < 6; ++j) | |
6590 | tmp[j] = 0; | |
6591 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (rettype))); | |
6592 | merge = build_real (TREE_TYPE (rettype), r); | |
6593 | } | |
6594 | else | |
6595 | gcc_unreachable (); | |
6596 | merge = build_vector_from_val (rettype, merge); | |
6597 | merge = vect_init_vector (stmt, merge, rettype, NULL); | |
6598 | ||
aec7ae7d JJ |
6599 | prev_stmt_info = NULL; |
6600 | for (j = 0; j < ncopies; ++j) | |
6601 | { | |
6602 | if (modifier == WIDEN && (j & 1)) | |
6603 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
6604 | perm_mask, stmt, gsi); | |
6605 | else if (j == 0) | |
6606 | op = vec_oprnd0 | |
81c40241 | 6607 | = vect_get_vec_def_for_operand (gather_off, stmt); |
aec7ae7d JJ |
6608 | else |
6609 | op = vec_oprnd0 | |
6610 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
6611 | ||
6612 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6613 | { | |
6614 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
6615 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 6616 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
aec7ae7d JJ |
6617 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
6618 | new_stmt | |
0d0e4a03 | 6619 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6620 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6621 | op = var; | |
6622 | } | |
6623 | ||
6624 | new_stmt | |
d3c2fee0 | 6625 | = gimple_build_call (gather_decl, 5, merge, ptr, op, mask, scale); |
aec7ae7d JJ |
6626 | |
6627 | if (!useless_type_conversion_p (vectype, rettype)) | |
6628 | { | |
6629 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
6630 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 6631 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
aec7ae7d JJ |
6632 | gimple_call_set_lhs (new_stmt, op); |
6633 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 6634 | var = make_ssa_name (vec_dest); |
aec7ae7d JJ |
6635 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
6636 | new_stmt | |
0d0e4a03 | 6637 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6638 | } |
6639 | else | |
6640 | { | |
6641 | var = make_ssa_name (vec_dest, new_stmt); | |
6642 | gimple_call_set_lhs (new_stmt, var); | |
6643 | } | |
6644 | ||
6645 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6646 | ||
6647 | if (modifier == NARROW) | |
6648 | { | |
6649 | if ((j & 1) == 0) | |
6650 | { | |
6651 | prev_res = var; | |
6652 | continue; | |
6653 | } | |
6654 | var = permute_vec_elements (prev_res, var, | |
6655 | perm_mask, stmt, gsi); | |
6656 | new_stmt = SSA_NAME_DEF_STMT (var); | |
6657 | } | |
6658 | ||
6659 | if (prev_stmt_info == NULL) | |
6660 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6661 | else | |
6662 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6663 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6664 | } | |
6665 | return true; | |
6666 | } | |
f2e2a985 | 6667 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
7d75abc8 MM |
6668 | { |
6669 | gimple_stmt_iterator incr_gsi; | |
6670 | bool insert_after; | |
355fe088 | 6671 | gimple *incr; |
7d75abc8 | 6672 | tree offvar; |
7d75abc8 MM |
6673 | tree ivstep; |
6674 | tree running_off; | |
9771b263 | 6675 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 6676 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
6677 | tree stride_base, stride_step, alias_off; |
6678 | ||
6679 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 6680 | |
f502d50e | 6681 | if (slp && grouped_load) |
ab313a8c RB |
6682 | first_dr = STMT_VINFO_DATA_REF |
6683 | (vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info))); | |
6684 | else | |
6685 | first_dr = dr; | |
6686 | ||
14ac6aa2 RB |
6687 | stride_base |
6688 | = fold_build_pointer_plus | |
ab313a8c | 6689 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 6690 | size_binop (PLUS_EXPR, |
ab313a8c RB |
6691 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
6692 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
6693 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
6694 | |
6695 | /* For a load with loop-invariant (but other than power-of-2) | |
6696 | stride (i.e. not a grouped access) like so: | |
6697 | ||
6698 | for (i = 0; i < n; i += stride) | |
6699 | ... = array[i]; | |
6700 | ||
6701 | we generate a new induction variable and new accesses to | |
6702 | form a new vector (or vectors, depending on ncopies): | |
6703 | ||
6704 | for (j = 0; ; j += VF*stride) | |
6705 | tmp1 = array[j]; | |
6706 | tmp2 = array[j + stride]; | |
6707 | ... | |
6708 | vectemp = {tmp1, tmp2, ...} | |
6709 | */ | |
6710 | ||
ab313a8c RB |
6711 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
6712 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
6713 | |
6714 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6715 | ||
ab313a8c | 6716 | create_iv (unshare_expr (stride_base), unshare_expr (ivstep), NULL, |
7d75abc8 MM |
6717 | loop, &incr_gsi, insert_after, |
6718 | &offvar, NULL); | |
6719 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 6720 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
7d75abc8 | 6721 | |
ab313a8c RB |
6722 | stride_step = force_gimple_operand (unshare_expr (stride_step), |
6723 | &stmts, true, NULL_TREE); | |
7d75abc8 MM |
6724 | if (stmts) |
6725 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
6726 | ||
6727 | prev_stmt_info = NULL; | |
6728 | running_off = offvar; | |
ab313a8c | 6729 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (first_dr)), 0); |
7b5fc413 RB |
6730 | int nloads = nunits; |
6731 | tree ltype = TREE_TYPE (vectype); | |
b266b968 | 6732 | auto_vec<tree> dr_chain; |
7b5fc413 RB |
6733 | if (slp) |
6734 | { | |
6735 | nloads = nunits / group_size; | |
6736 | if (group_size < nunits) | |
6737 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); | |
6738 | else | |
6739 | ltype = vectype; | |
6740 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); | |
66c16fd9 RB |
6741 | /* For SLP permutation support we need to load the whole group, |
6742 | not only the number of vector stmts the permutation result | |
6743 | fits in. */ | |
b266b968 | 6744 | if (slp_perm) |
66c16fd9 RB |
6745 | { |
6746 | ncopies = (group_size * vf + nunits - 1) / nunits; | |
6747 | dr_chain.create (ncopies); | |
6748 | } | |
6749 | else | |
6750 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
7b5fc413 | 6751 | } |
7d75abc8 MM |
6752 | for (j = 0; j < ncopies; j++) |
6753 | { | |
6754 | tree vec_inv; | |
6755 | ||
7b5fc413 RB |
6756 | if (nloads > 1) |
6757 | { | |
6758 | vec_alloc (v, nloads); | |
6759 | for (i = 0; i < nloads; i++) | |
6760 | { | |
6761 | tree newref, newoff; | |
355fe088 | 6762 | gimple *incr; |
7b5fc413 RB |
6763 | newref = build2 (MEM_REF, ltype, running_off, alias_off); |
6764 | ||
6765 | newref = force_gimple_operand_gsi (gsi, newref, true, | |
6766 | NULL_TREE, true, | |
6767 | GSI_SAME_STMT); | |
6768 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, newref); | |
6769 | newoff = copy_ssa_name (running_off); | |
6770 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6771 | running_off, stride_step); | |
6772 | vect_finish_stmt_generation (stmt, incr, gsi); | |
6773 | ||
6774 | running_off = newoff; | |
6775 | } | |
6776 | ||
6777 | vec_inv = build_constructor (vectype, v); | |
6778 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
6779 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
6780 | } | |
6781 | else | |
7d75abc8 | 6782 | { |
7b5fc413 RB |
6783 | new_stmt = gimple_build_assign (make_ssa_name (ltype), |
6784 | build2 (MEM_REF, ltype, | |
6785 | running_off, alias_off)); | |
6786 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6787 | ||
6788 | tree newoff = copy_ssa_name (running_off); | |
355fe088 | 6789 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
0d0e4a03 | 6790 | running_off, stride_step); |
7d75abc8 MM |
6791 | vect_finish_stmt_generation (stmt, incr, gsi); |
6792 | ||
6793 | running_off = newoff; | |
6794 | } | |
6795 | ||
7b5fc413 | 6796 | if (slp) |
b266b968 | 6797 | { |
b266b968 RB |
6798 | if (slp_perm) |
6799 | dr_chain.quick_push (gimple_assign_lhs (new_stmt)); | |
66c16fd9 RB |
6800 | else |
6801 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
b266b968 | 6802 | } |
7d75abc8 | 6803 | else |
225ce44b RB |
6804 | { |
6805 | if (j == 0) | |
6806 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6807 | else | |
6808 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6809 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6810 | } | |
7d75abc8 | 6811 | } |
b266b968 RB |
6812 | if (slp_perm) |
6813 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
6814 | slp_node_instance, false); | |
7d75abc8 MM |
6815 | return true; |
6816 | } | |
aec7ae7d | 6817 | |
0d0293ac | 6818 | if (grouped_load) |
ebfd146a | 6819 | { |
e14c1050 | 6820 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
4f0a0218 | 6821 | /* For SLP vectorization we directly vectorize a subchain |
52eab378 RB |
6822 | without permutation. */ |
6823 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
4f0a0218 RB |
6824 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
6825 | /* For BB vectorization always use the first stmt to base | |
6826 | the data ref pointer on. */ | |
6827 | if (bb_vinfo) | |
6828 | first_stmt_for_drptr = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 6829 | |
ebfd146a | 6830 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
6831 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
6832 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
6833 | ??? But we can only do so if there is exactly one | |
6834 | as we have no way to get at the rest. Leave the CSE | |
6835 | opportunity alone. | |
6836 | ??? With the group load eventually participating | |
6837 | in multiple different permutations (having multiple | |
6838 | slp nodes which refer to the same group) the CSE | |
6839 | is even wrong code. See PR56270. */ | |
6840 | && !slp) | |
ebfd146a IR |
6841 | { |
6842 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
6843 | return true; | |
6844 | } | |
6845 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 6846 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
9b999e8c | 6847 | group_gap_adj = 0; |
ebfd146a IR |
6848 | |
6849 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
6850 | if (slp) | |
6851 | { | |
0d0293ac | 6852 | grouped_load = false; |
91ff1504 RB |
6853 | /* For SLP permutation support we need to load the whole group, |
6854 | not only the number of vector stmts the permutation result | |
6855 | fits in. */ | |
6856 | if (slp_perm) | |
6857 | vec_num = (group_size * vf + nunits - 1) / nunits; | |
6858 | else | |
6859 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9b999e8c | 6860 | group_gap_adj = vf * group_size - nunits * vec_num; |
a70d6342 | 6861 | } |
ebfd146a | 6862 | else |
9b999e8c | 6863 | vec_num = group_size; |
ebfd146a IR |
6864 | } |
6865 | else | |
6866 | { | |
6867 | first_stmt = stmt; | |
6868 | first_dr = dr; | |
6869 | group_size = vec_num = 1; | |
9b999e8c | 6870 | group_gap_adj = 0; |
ebfd146a IR |
6871 | } |
6872 | ||
720f5239 | 6873 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6874 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
6875 | /* Targets with load-lane instructions must not require explicit |
6876 | realignment. */ | |
6877 | gcc_assert (!load_lanes_p | |
6878 | || alignment_support_scheme == dr_aligned | |
6879 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
6880 | |
6881 | /* In case the vectorization factor (VF) is bigger than the number | |
6882 | of elements that we can fit in a vectype (nunits), we have to generate | |
6883 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 6884 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 6885 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 6886 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 6887 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
6888 | stmts that use the defs of the current stmt. The example below |
6889 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
6890 | need to create 4 vectorized stmts): | |
ebfd146a IR |
6891 | |
6892 | before vectorization: | |
6893 | RELATED_STMT VEC_STMT | |
6894 | S1: x = memref - - | |
6895 | S2: z = x + 1 - - | |
6896 | ||
6897 | step 1: vectorize stmt S1: | |
6898 | We first create the vector stmt VS1_0, and, as usual, record a | |
6899 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
6900 | Next, we create the vector stmt VS1_1, and record a pointer to | |
6901 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 6902 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
6903 | stmts and pointers: |
6904 | RELATED_STMT VEC_STMT | |
6905 | VS1_0: vx0 = memref0 VS1_1 - | |
6906 | VS1_1: vx1 = memref1 VS1_2 - | |
6907 | VS1_2: vx2 = memref2 VS1_3 - | |
6908 | VS1_3: vx3 = memref3 - - | |
6909 | S1: x = load - VS1_0 | |
6910 | S2: z = x + 1 - - | |
6911 | ||
b8698a0f L |
6912 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
6913 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
6914 | stmt S2. */ |
6915 | ||
0d0293ac | 6916 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6917 | |
6918 | S1: x2 = &base + 2 | |
6919 | S2: x0 = &base | |
6920 | S3: x1 = &base + 1 | |
6921 | S4: x3 = &base + 3 | |
6922 | ||
b8698a0f | 6923 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
6924 | starting from the access of the first stmt of the chain: |
6925 | ||
6926 | VS1: vx0 = &base | |
6927 | VS2: vx1 = &base + vec_size*1 | |
6928 | VS3: vx3 = &base + vec_size*2 | |
6929 | VS4: vx4 = &base + vec_size*3 | |
6930 | ||
6931 | Then permutation statements are generated: | |
6932 | ||
e2c83630 RH |
6933 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
6934 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
6935 | ... |
6936 | ||
6937 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
6938 | (the order of the data-refs in the output of vect_permute_load_chain | |
6939 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6940 | the documentation of vect_permute_load_chain()). | |
6941 | The generation of permutation stmts and recording them in | |
0d0293ac | 6942 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 6943 | |
b8698a0f | 6944 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
6945 | permutation stmts above are created for every copy. The result vector |
6946 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
6947 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
6948 | |
6949 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
6950 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
6951 | we generate the following code: | |
6952 | p = initial_addr; | |
6953 | indx = 0; | |
6954 | loop { | |
6955 | p = p + indx * vectype_size; | |
6956 | vec_dest = *(p); | |
6957 | indx = indx + 1; | |
6958 | } | |
6959 | ||
6960 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 6961 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
6962 | then generate the following code, in which the data in each iteration is |
6963 | obtained by two vector loads, one from the previous iteration, and one | |
6964 | from the current iteration: | |
6965 | p1 = initial_addr; | |
6966 | msq_init = *(floor(p1)) | |
6967 | p2 = initial_addr + VS - 1; | |
6968 | realignment_token = call target_builtin; | |
6969 | indx = 0; | |
6970 | loop { | |
6971 | p2 = p2 + indx * vectype_size | |
6972 | lsq = *(floor(p2)) | |
6973 | vec_dest = realign_load (msq, lsq, realignment_token) | |
6974 | indx = indx + 1; | |
6975 | msq = lsq; | |
6976 | } */ | |
6977 | ||
6978 | /* If the misalignment remains the same throughout the execution of the | |
6979 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 6980 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
6981 | This can only occur when vectorizing memory accesses in the inner-loop |
6982 | nested within an outer-loop that is being vectorized. */ | |
6983 | ||
d1e4b493 | 6984 | if (nested_in_vect_loop |
211bea38 | 6985 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
ebfd146a IR |
6986 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) |
6987 | { | |
6988 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
6989 | compute_in_loop = true; | |
6990 | } | |
6991 | ||
6992 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
6993 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 6994 | && !compute_in_loop) |
ebfd146a IR |
6995 | { |
6996 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
6997 | alignment_support_scheme, NULL_TREE, | |
6998 | &at_loop); | |
6999 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7000 | { | |
538dd0b7 | 7001 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
7002 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
7003 | size_one_node); | |
ebfd146a IR |
7004 | } |
7005 | } | |
7006 | else | |
7007 | at_loop = loop; | |
7008 | ||
a1e53f3f L |
7009 | if (negative) |
7010 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
7011 | ||
272c6793 RS |
7012 | if (load_lanes_p) |
7013 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
7014 | else | |
7015 | aggr_type = vectype; | |
7016 | ||
ebfd146a IR |
7017 | prev_stmt_info = NULL; |
7018 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 7019 | { |
272c6793 | 7020 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 7021 | if (j == 0) |
74bf76ed JJ |
7022 | { |
7023 | bool simd_lane_access_p | |
7024 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
7025 | if (simd_lane_access_p | |
7026 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
7027 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
7028 | && integer_zerop (DR_OFFSET (first_dr)) | |
7029 | && integer_zerop (DR_INIT (first_dr)) | |
7030 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
7031 | get_alias_set (DR_REF (first_dr))) | |
7032 | && (alignment_support_scheme == dr_aligned | |
7033 | || alignment_support_scheme == dr_unaligned_supported)) | |
7034 | { | |
7035 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
7036 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
7037 | (DR_REF (first_dr)), 0); | |
8928eff3 | 7038 | inv_p = false; |
74bf76ed | 7039 | } |
4f0a0218 RB |
7040 | else if (first_stmt_for_drptr |
7041 | && first_stmt != first_stmt_for_drptr) | |
7042 | { | |
7043 | dataref_ptr | |
7044 | = vect_create_data_ref_ptr (first_stmt_for_drptr, aggr_type, | |
7045 | at_loop, offset, &dummy, gsi, | |
7046 | &ptr_incr, simd_lane_access_p, | |
7047 | &inv_p, byte_offset); | |
7048 | /* Adjust the pointer by the difference to first_stmt. */ | |
7049 | data_reference_p ptrdr | |
7050 | = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt_for_drptr)); | |
7051 | tree diff = fold_convert (sizetype, | |
7052 | size_binop (MINUS_EXPR, | |
7053 | DR_INIT (first_dr), | |
7054 | DR_INIT (ptrdr))); | |
7055 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7056 | stmt, diff); | |
7057 | } | |
74bf76ed JJ |
7058 | else |
7059 | dataref_ptr | |
7060 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
7061 | offset, &dummy, gsi, &ptr_incr, | |
356bbc4c JJ |
7062 | simd_lane_access_p, &inv_p, |
7063 | byte_offset); | |
74bf76ed JJ |
7064 | } |
7065 | else if (dataref_offset) | |
7066 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
7067 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 7068 | else |
272c6793 RS |
7069 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
7070 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 7071 | |
0d0293ac | 7072 | if (grouped_load || slp_perm) |
9771b263 | 7073 | dr_chain.create (vec_num); |
5ce1ee7f | 7074 | |
272c6793 | 7075 | if (load_lanes_p) |
ebfd146a | 7076 | { |
272c6793 RS |
7077 | tree vec_array; |
7078 | ||
7079 | vec_array = create_vector_array (vectype, vec_num); | |
7080 | ||
7081 | /* Emit: | |
7082 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
7083 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
7084 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
7085 | gimple_call_set_lhs (new_stmt, vec_array); | |
7086 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 7087 | |
272c6793 RS |
7088 | /* Extract each vector into an SSA_NAME. */ |
7089 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 7090 | { |
272c6793 RS |
7091 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
7092 | vec_array, i); | |
9771b263 | 7093 | dr_chain.quick_push (new_temp); |
272c6793 RS |
7094 | } |
7095 | ||
7096 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 7097 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
7098 | } |
7099 | else | |
7100 | { | |
7101 | for (i = 0; i < vec_num; i++) | |
7102 | { | |
7103 | if (i > 0) | |
7104 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7105 | stmt, NULL_TREE); | |
7106 | ||
7107 | /* 2. Create the vector-load in the loop. */ | |
7108 | switch (alignment_support_scheme) | |
7109 | { | |
7110 | case dr_aligned: | |
7111 | case dr_unaligned_supported: | |
be1ac4ec | 7112 | { |
644ffefd MJ |
7113 | unsigned int align, misalign; |
7114 | ||
272c6793 | 7115 | data_ref |
aed93b23 RB |
7116 | = fold_build2 (MEM_REF, vectype, dataref_ptr, |
7117 | dataref_offset | |
7118 | ? dataref_offset | |
7119 | : build_int_cst (reference_alias_ptr_type | |
7120 | (DR_REF (first_dr)), 0)); | |
644ffefd | 7121 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 RS |
7122 | if (alignment_support_scheme == dr_aligned) |
7123 | { | |
7124 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 7125 | misalign = 0; |
272c6793 RS |
7126 | } |
7127 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
7128 | { | |
52639a61 RB |
7129 | if (DR_VECT_AUX (first_dr)->base_element_aligned) |
7130 | align = TYPE_ALIGN_UNIT (elem_type); | |
7131 | else | |
7132 | align = (get_object_alignment (DR_REF (first_dr)) | |
7133 | / BITS_PER_UNIT); | |
7134 | misalign = 0; | |
272c6793 RS |
7135 | TREE_TYPE (data_ref) |
7136 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 7137 | align * BITS_PER_UNIT); |
272c6793 RS |
7138 | } |
7139 | else | |
7140 | { | |
7141 | TREE_TYPE (data_ref) | |
7142 | = build_aligned_type (TREE_TYPE (data_ref), | |
7143 | TYPE_ALIGN (elem_type)); | |
644ffefd | 7144 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 7145 | } |
aed93b23 RB |
7146 | if (dataref_offset == NULL_TREE |
7147 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
7148 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
7149 | align, misalign); | |
272c6793 | 7150 | break; |
be1ac4ec | 7151 | } |
272c6793 | 7152 | case dr_explicit_realign: |
267d3070 | 7153 | { |
272c6793 | 7154 | tree ptr, bump; |
272c6793 | 7155 | |
d88981fc | 7156 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
7157 | |
7158 | if (compute_in_loop) | |
7159 | msq = vect_setup_realignment (first_stmt, gsi, | |
7160 | &realignment_token, | |
7161 | dr_explicit_realign, | |
7162 | dataref_ptr, NULL); | |
7163 | ||
aed93b23 RB |
7164 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7165 | ptr = copy_ssa_name (dataref_ptr); | |
7166 | else | |
7167 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
7168 | new_stmt = gimple_build_assign |
7169 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7170 | build_int_cst |
7171 | (TREE_TYPE (dataref_ptr), | |
7172 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
7173 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7174 | data_ref | |
7175 | = build2 (MEM_REF, vectype, ptr, | |
7176 | build_int_cst (reference_alias_ptr_type | |
7177 | (DR_REF (first_dr)), 0)); | |
7178 | vec_dest = vect_create_destination_var (scalar_dest, | |
7179 | vectype); | |
7180 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
7181 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
7182 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7183 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
7184 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
7185 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7186 | msq = new_temp; | |
7187 | ||
d88981fc | 7188 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 7189 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 7190 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 7191 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
7192 | new_stmt = gimple_build_assign |
7193 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 RS |
7194 | build_int_cst |
7195 | (TREE_TYPE (ptr), | |
7196 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
aed93b23 | 7197 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
7198 | gimple_assign_set_lhs (new_stmt, ptr); |
7199 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7200 | data_ref | |
7201 | = build2 (MEM_REF, vectype, ptr, | |
7202 | build_int_cst (reference_alias_ptr_type | |
7203 | (DR_REF (first_dr)), 0)); | |
7204 | break; | |
267d3070 | 7205 | } |
272c6793 | 7206 | case dr_explicit_realign_optimized: |
aed93b23 RB |
7207 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7208 | new_temp = copy_ssa_name (dataref_ptr); | |
7209 | else | |
7210 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
7211 | new_stmt = gimple_build_assign |
7212 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7213 | build_int_cst |
7214 | (TREE_TYPE (dataref_ptr), | |
7215 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
7216 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7217 | data_ref | |
7218 | = build2 (MEM_REF, vectype, new_temp, | |
7219 | build_int_cst (reference_alias_ptr_type | |
7220 | (DR_REF (first_dr)), 0)); | |
7221 | break; | |
7222 | default: | |
7223 | gcc_unreachable (); | |
7224 | } | |
ebfd146a | 7225 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 7226 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
7227 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7228 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7229 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7230 | ||
272c6793 RS |
7231 | /* 3. Handle explicit realignment if necessary/supported. |
7232 | Create in loop: | |
7233 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
7234 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
7235 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 7236 | { |
272c6793 RS |
7237 | lsq = gimple_assign_lhs (new_stmt); |
7238 | if (!realignment_token) | |
7239 | realignment_token = dataref_ptr; | |
7240 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
7241 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
7242 | msq, lsq, realignment_token); | |
272c6793 RS |
7243 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7244 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7245 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7246 | ||
7247 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7248 | { | |
7249 | gcc_assert (phi); | |
7250 | if (i == vec_num - 1 && j == ncopies - 1) | |
7251 | add_phi_arg (phi, lsq, | |
7252 | loop_latch_edge (containing_loop), | |
9e227d60 | 7253 | UNKNOWN_LOCATION); |
272c6793 RS |
7254 | msq = lsq; |
7255 | } | |
ebfd146a | 7256 | } |
ebfd146a | 7257 | |
59fd17e3 RB |
7258 | /* 4. Handle invariant-load. */ |
7259 | if (inv_p && !bb_vinfo) | |
7260 | { | |
59fd17e3 | 7261 | gcc_assert (!grouped_load); |
d1417442 JJ |
7262 | /* If we have versioned for aliasing or the loop doesn't |
7263 | have any data dependencies that would preclude this, | |
7264 | then we are sure this is a loop invariant load and | |
7265 | thus we can insert it on the preheader edge. */ | |
7266 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
7267 | && !nested_in_vect_loop | |
6b916b36 | 7268 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
7269 | { |
7270 | if (dump_enabled_p ()) | |
7271 | { | |
7272 | dump_printf_loc (MSG_NOTE, vect_location, | |
7273 | "hoisting out of the vectorized " | |
7274 | "loop: "); | |
7275 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 7276 | } |
b731b390 | 7277 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
7278 | gsi_insert_on_edge_immediate |
7279 | (loop_preheader_edge (loop), | |
7280 | gimple_build_assign (tem, | |
7281 | unshare_expr | |
7282 | (gimple_assign_rhs1 (stmt)))); | |
7283 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
34cd48e5 RB |
7284 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7285 | set_vinfo_for_stmt (new_stmt, | |
7286 | new_stmt_vec_info (new_stmt, vinfo)); | |
a0e35eb0 RB |
7287 | } |
7288 | else | |
7289 | { | |
7290 | gimple_stmt_iterator gsi2 = *gsi; | |
7291 | gsi_next (&gsi2); | |
7292 | new_temp = vect_init_vector (stmt, scalar_dest, | |
7293 | vectype, &gsi2); | |
34cd48e5 | 7294 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
a0e35eb0 | 7295 | } |
59fd17e3 RB |
7296 | } |
7297 | ||
272c6793 RS |
7298 | if (negative) |
7299 | { | |
aec7ae7d JJ |
7300 | tree perm_mask = perm_mask_for_reverse (vectype); |
7301 | new_temp = permute_vec_elements (new_temp, new_temp, | |
7302 | perm_mask, stmt, gsi); | |
ebfd146a IR |
7303 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7304 | } | |
267d3070 | 7305 | |
272c6793 | 7306 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
7307 | vect_transform_grouped_load (). */ |
7308 | if (grouped_load || slp_perm) | |
9771b263 | 7309 | dr_chain.quick_push (new_temp); |
267d3070 | 7310 | |
272c6793 RS |
7311 | /* Store vector loads in the corresponding SLP_NODE. */ |
7312 | if (slp && !slp_perm) | |
9771b263 | 7313 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
272c6793 | 7314 | } |
9b999e8c RB |
7315 | /* Bump the vector pointer to account for a gap or for excess |
7316 | elements loaded for a permuted SLP load. */ | |
7317 | if (group_gap_adj != 0) | |
a64b9c26 | 7318 | { |
9b999e8c RB |
7319 | bool ovf; |
7320 | tree bump | |
7321 | = wide_int_to_tree (sizetype, | |
7322 | wi::smul (TYPE_SIZE_UNIT (elem_type), | |
7323 | group_gap_adj, &ovf)); | |
a64b9c26 RB |
7324 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
7325 | stmt, bump); | |
7326 | } | |
ebfd146a IR |
7327 | } |
7328 | ||
7329 | if (slp && !slp_perm) | |
7330 | continue; | |
7331 | ||
7332 | if (slp_perm) | |
7333 | { | |
01d8bf07 | 7334 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
ebfd146a IR |
7335 | slp_node_instance, false)) |
7336 | { | |
9771b263 | 7337 | dr_chain.release (); |
ebfd146a IR |
7338 | return false; |
7339 | } | |
7340 | } | |
7341 | else | |
7342 | { | |
0d0293ac | 7343 | if (grouped_load) |
ebfd146a | 7344 | { |
272c6793 | 7345 | if (!load_lanes_p) |
0d0293ac | 7346 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 7347 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
7348 | } |
7349 | else | |
7350 | { | |
7351 | if (j == 0) | |
7352 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7353 | else | |
7354 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7355 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7356 | } | |
7357 | } | |
9771b263 | 7358 | dr_chain.release (); |
ebfd146a IR |
7359 | } |
7360 | ||
ebfd146a IR |
7361 | return true; |
7362 | } | |
7363 | ||
7364 | /* Function vect_is_simple_cond. | |
b8698a0f | 7365 | |
ebfd146a IR |
7366 | Input: |
7367 | LOOP - the loop that is being vectorized. | |
7368 | COND - Condition that is checked for simple use. | |
7369 | ||
e9e1d143 RG |
7370 | Output: |
7371 | *COMP_VECTYPE - the vector type for the comparison. | |
7372 | ||
ebfd146a IR |
7373 | Returns whether a COND can be vectorized. Checks whether |
7374 | condition operands are supportable using vec_is_simple_use. */ | |
7375 | ||
87aab9b2 | 7376 | static bool |
81c40241 | 7377 | vect_is_simple_cond (tree cond, vec_info *vinfo, tree *comp_vectype) |
ebfd146a IR |
7378 | { |
7379 | tree lhs, rhs; | |
ebfd146a | 7380 | enum vect_def_type dt; |
e9e1d143 | 7381 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 7382 | |
a414c77f IE |
7383 | /* Mask case. */ |
7384 | if (TREE_CODE (cond) == SSA_NAME | |
7385 | && TREE_CODE (TREE_TYPE (cond)) == BOOLEAN_TYPE) | |
7386 | { | |
7387 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (cond); | |
7388 | if (!vect_is_simple_use (cond, vinfo, &lhs_def_stmt, | |
7389 | &dt, comp_vectype) | |
7390 | || !*comp_vectype | |
7391 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
7392 | return false; | |
7393 | return true; | |
7394 | } | |
7395 | ||
ebfd146a IR |
7396 | if (!COMPARISON_CLASS_P (cond)) |
7397 | return false; | |
7398 | ||
7399 | lhs = TREE_OPERAND (cond, 0); | |
7400 | rhs = TREE_OPERAND (cond, 1); | |
7401 | ||
7402 | if (TREE_CODE (lhs) == SSA_NAME) | |
7403 | { | |
355fe088 | 7404 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); |
81c40241 | 7405 | if (!vect_is_simple_use (lhs, vinfo, &lhs_def_stmt, &dt, &vectype1)) |
ebfd146a IR |
7406 | return false; |
7407 | } | |
7408 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
7409 | && TREE_CODE (lhs) != FIXED_CST) | |
7410 | return false; | |
7411 | ||
7412 | if (TREE_CODE (rhs) == SSA_NAME) | |
7413 | { | |
355fe088 | 7414 | gimple *rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); |
81c40241 | 7415 | if (!vect_is_simple_use (rhs, vinfo, &rhs_def_stmt, &dt, &vectype2)) |
ebfd146a IR |
7416 | return false; |
7417 | } | |
f7e531cf | 7418 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST |
ebfd146a IR |
7419 | && TREE_CODE (rhs) != FIXED_CST) |
7420 | return false; | |
7421 | ||
28b33016 IE |
7422 | if (vectype1 && vectype2 |
7423 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
7424 | return false; | |
7425 | ||
e9e1d143 | 7426 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
7427 | return true; |
7428 | } | |
7429 | ||
7430 | /* vectorizable_condition. | |
7431 | ||
b8698a0f L |
7432 | Check if STMT is conditional modify expression that can be vectorized. |
7433 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7434 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
7435 | at GSI. |
7436 | ||
7437 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
7438 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 7439 | else clause if it is 2). |
ebfd146a IR |
7440 | |
7441 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7442 | ||
4bbe8262 | 7443 | bool |
355fe088 TS |
7444 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
7445 | gimple **vec_stmt, tree reduc_def, int reduc_index, | |
f7e531cf | 7446 | slp_tree slp_node) |
ebfd146a IR |
7447 | { |
7448 | tree scalar_dest = NULL_TREE; | |
7449 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
7450 | tree cond_expr, then_clause, else_clause; |
7451 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
df11cc78 | 7452 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
7453 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
7454 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
5958f9e2 | 7455 | tree vec_compare; |
ebfd146a IR |
7456 | tree new_temp; |
7457 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a855b1b1 | 7458 | enum vect_def_type dt, dts[4]; |
f7e531cf | 7459 | int ncopies; |
ebfd146a | 7460 | enum tree_code code; |
a855b1b1 | 7461 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
7462 | int i, j; |
7463 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
7464 | vec<tree> vec_oprnds0 = vNULL; |
7465 | vec<tree> vec_oprnds1 = vNULL; | |
7466 | vec<tree> vec_oprnds2 = vNULL; | |
7467 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 7468 | tree vec_cmp_type; |
a414c77f | 7469 | bool masked = false; |
b8698a0f | 7470 | |
f7e531cf IR |
7471 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
7472 | return false; | |
7473 | ||
af29617a AH |
7474 | if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == TREE_CODE_REDUCTION) |
7475 | { | |
7476 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
7477 | return false; | |
ebfd146a | 7478 | |
af29617a AH |
7479 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7480 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7481 | && reduc_def)) | |
7482 | return false; | |
ebfd146a | 7483 | |
af29617a AH |
7484 | /* FORNOW: not yet supported. */ |
7485 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7486 | { | |
7487 | if (dump_enabled_p ()) | |
7488 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7489 | "value used after loop.\n"); | |
7490 | return false; | |
7491 | } | |
ebfd146a IR |
7492 | } |
7493 | ||
7494 | /* Is vectorizable conditional operation? */ | |
7495 | if (!is_gimple_assign (stmt)) | |
7496 | return false; | |
7497 | ||
7498 | code = gimple_assign_rhs_code (stmt); | |
7499 | ||
7500 | if (code != COND_EXPR) | |
7501 | return false; | |
7502 | ||
465c8c19 JJ |
7503 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
7504 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2947d3b2 | 7505 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
465c8c19 | 7506 | |
fce57248 | 7507 | if (slp_node) |
465c8c19 JJ |
7508 | ncopies = 1; |
7509 | else | |
7510 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
7511 | ||
7512 | gcc_assert (ncopies >= 1); | |
7513 | if (reduc_index && ncopies > 1) | |
7514 | return false; /* FORNOW */ | |
7515 | ||
4e71066d RG |
7516 | cond_expr = gimple_assign_rhs1 (stmt); |
7517 | then_clause = gimple_assign_rhs2 (stmt); | |
7518 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 7519 | |
81c40241 | 7520 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, &comp_vectype) |
e9e1d143 | 7521 | || !comp_vectype) |
ebfd146a IR |
7522 | return false; |
7523 | ||
81c40241 | 7524 | gimple *def_stmt; |
2947d3b2 IE |
7525 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &def_stmt, &dt, |
7526 | &vectype1)) | |
7527 | return false; | |
7528 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &def_stmt, &dt, | |
7529 | &vectype2)) | |
ebfd146a | 7530 | return false; |
2947d3b2 IE |
7531 | |
7532 | if (vectype1 && !useless_type_conversion_p (vectype, vectype1)) | |
7533 | return false; | |
7534 | ||
7535 | if (vectype2 && !useless_type_conversion_p (vectype, vectype2)) | |
ebfd146a IR |
7536 | return false; |
7537 | ||
28b33016 IE |
7538 | masked = !COMPARISON_CLASS_P (cond_expr); |
7539 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
7540 | ||
74946978 MP |
7541 | if (vec_cmp_type == NULL_TREE) |
7542 | return false; | |
784fb9b3 | 7543 | |
b8698a0f | 7544 | if (!vec_stmt) |
ebfd146a IR |
7545 | { |
7546 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
e9e1d143 | 7547 | return expand_vec_cond_expr_p (vectype, comp_vectype); |
ebfd146a IR |
7548 | } |
7549 | ||
f7e531cf IR |
7550 | /* Transform. */ |
7551 | ||
7552 | if (!slp_node) | |
7553 | { | |
9771b263 DN |
7554 | vec_oprnds0.create (1); |
7555 | vec_oprnds1.create (1); | |
7556 | vec_oprnds2.create (1); | |
7557 | vec_oprnds3.create (1); | |
f7e531cf | 7558 | } |
ebfd146a IR |
7559 | |
7560 | /* Handle def. */ | |
7561 | scalar_dest = gimple_assign_lhs (stmt); | |
7562 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
7563 | ||
7564 | /* Handle cond expr. */ | |
a855b1b1 MM |
7565 | for (j = 0; j < ncopies; j++) |
7566 | { | |
538dd0b7 | 7567 | gassign *new_stmt = NULL; |
a855b1b1 MM |
7568 | if (j == 0) |
7569 | { | |
f7e531cf IR |
7570 | if (slp_node) |
7571 | { | |
00f96dc9 TS |
7572 | auto_vec<tree, 4> ops; |
7573 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 7574 | |
a414c77f IE |
7575 | if (masked) |
7576 | ops.safe_push (cond_expr); | |
7577 | else | |
7578 | { | |
7579 | ops.safe_push (TREE_OPERAND (cond_expr, 0)); | |
7580 | ops.safe_push (TREE_OPERAND (cond_expr, 1)); | |
7581 | } | |
9771b263 DN |
7582 | ops.safe_push (then_clause); |
7583 | ops.safe_push (else_clause); | |
f7e531cf | 7584 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); |
37b5ec8f JJ |
7585 | vec_oprnds3 = vec_defs.pop (); |
7586 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
7587 | if (!masked) |
7588 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 7589 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf | 7590 | |
9771b263 DN |
7591 | ops.release (); |
7592 | vec_defs.release (); | |
f7e531cf IR |
7593 | } |
7594 | else | |
7595 | { | |
355fe088 | 7596 | gimple *gtemp; |
a414c77f IE |
7597 | if (masked) |
7598 | { | |
7599 | vec_cond_lhs | |
7600 | = vect_get_vec_def_for_operand (cond_expr, stmt, | |
7601 | comp_vectype); | |
7602 | vect_is_simple_use (cond_expr, stmt_info->vinfo, | |
7603 | >emp, &dts[0]); | |
7604 | } | |
7605 | else | |
7606 | { | |
7607 | vec_cond_lhs = | |
7608 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), | |
7609 | stmt, comp_vectype); | |
7610 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), | |
7611 | loop_vinfo, >emp, &dts[0]); | |
7612 | ||
7613 | vec_cond_rhs = | |
7614 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
7615 | stmt, comp_vectype); | |
7616 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), | |
7617 | loop_vinfo, >emp, &dts[1]); | |
7618 | } | |
f7e531cf IR |
7619 | if (reduc_index == 1) |
7620 | vec_then_clause = reduc_def; | |
7621 | else | |
7622 | { | |
7623 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 RB |
7624 | stmt); |
7625 | vect_is_simple_use (then_clause, loop_vinfo, | |
7626 | >emp, &dts[2]); | |
f7e531cf IR |
7627 | } |
7628 | if (reduc_index == 2) | |
7629 | vec_else_clause = reduc_def; | |
7630 | else | |
7631 | { | |
7632 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 RB |
7633 | stmt); |
7634 | vect_is_simple_use (else_clause, loop_vinfo, >emp, &dts[3]); | |
f7e531cf | 7635 | } |
a855b1b1 MM |
7636 | } |
7637 | } | |
7638 | else | |
7639 | { | |
a414c77f IE |
7640 | vec_cond_lhs |
7641 | = vect_get_vec_def_for_stmt_copy (dts[0], | |
7642 | vec_oprnds0.pop ()); | |
7643 | if (!masked) | |
7644 | vec_cond_rhs | |
7645 | = vect_get_vec_def_for_stmt_copy (dts[1], | |
7646 | vec_oprnds1.pop ()); | |
7647 | ||
a855b1b1 | 7648 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 7649 | vec_oprnds2.pop ()); |
a855b1b1 | 7650 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 7651 | vec_oprnds3.pop ()); |
f7e531cf IR |
7652 | } |
7653 | ||
7654 | if (!slp_node) | |
7655 | { | |
9771b263 | 7656 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
7657 | if (!masked) |
7658 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
7659 | vec_oprnds2.quick_push (vec_then_clause); |
7660 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
7661 | } |
7662 | ||
9dc3f7de | 7663 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 7664 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 7665 | { |
9771b263 DN |
7666 | vec_then_clause = vec_oprnds2[i]; |
7667 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 7668 | |
a414c77f IE |
7669 | if (masked) |
7670 | vec_compare = vec_cond_lhs; | |
7671 | else | |
7672 | { | |
7673 | vec_cond_rhs = vec_oprnds1[i]; | |
7674 | vec_compare = build2 (TREE_CODE (cond_expr), vec_cmp_type, | |
7675 | vec_cond_lhs, vec_cond_rhs); | |
7676 | } | |
5958f9e2 JJ |
7677 | new_temp = make_ssa_name (vec_dest); |
7678 | new_stmt = gimple_build_assign (new_temp, VEC_COND_EXPR, | |
7679 | vec_compare, vec_then_clause, | |
7680 | vec_else_clause); | |
f7e531cf IR |
7681 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7682 | if (slp_node) | |
9771b263 | 7683 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
7684 | } |
7685 | ||
7686 | if (slp_node) | |
7687 | continue; | |
7688 | ||
7689 | if (j == 0) | |
7690 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7691 | else | |
7692 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7693 | ||
7694 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 7695 | } |
b8698a0f | 7696 | |
9771b263 DN |
7697 | vec_oprnds0.release (); |
7698 | vec_oprnds1.release (); | |
7699 | vec_oprnds2.release (); | |
7700 | vec_oprnds3.release (); | |
f7e531cf | 7701 | |
ebfd146a IR |
7702 | return true; |
7703 | } | |
7704 | ||
42fd8198 IE |
7705 | /* vectorizable_comparison. |
7706 | ||
7707 | Check if STMT is comparison expression that can be vectorized. | |
7708 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7709 | comparison, put it in VEC_STMT, and insert it at GSI. | |
7710 | ||
7711 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7712 | ||
fce57248 | 7713 | static bool |
42fd8198 IE |
7714 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, |
7715 | gimple **vec_stmt, tree reduc_def, | |
7716 | slp_tree slp_node) | |
7717 | { | |
7718 | tree lhs, rhs1, rhs2; | |
7719 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
7720 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
7721 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
7722 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
7723 | tree new_temp; | |
7724 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
7725 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
7726 | unsigned nunits; | |
7727 | int ncopies; | |
7728 | enum tree_code code; | |
7729 | stmt_vec_info prev_stmt_info = NULL; | |
7730 | int i, j; | |
7731 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
7732 | vec<tree> vec_oprnds0 = vNULL; | |
7733 | vec<tree> vec_oprnds1 = vNULL; | |
7734 | gimple *def_stmt; | |
7735 | tree mask_type; | |
7736 | tree mask; | |
7737 | ||
c245362b IE |
7738 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
7739 | return false; | |
7740 | ||
30480bcd | 7741 | if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)) |
42fd8198 IE |
7742 | return false; |
7743 | ||
7744 | mask_type = vectype; | |
7745 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
7746 | ||
fce57248 | 7747 | if (slp_node) |
42fd8198 IE |
7748 | ncopies = 1; |
7749 | else | |
7750 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
7751 | ||
7752 | gcc_assert (ncopies >= 1); | |
42fd8198 IE |
7753 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7754 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7755 | && reduc_def)) | |
7756 | return false; | |
7757 | ||
7758 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7759 | { | |
7760 | if (dump_enabled_p ()) | |
7761 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7762 | "value used after loop.\n"); | |
7763 | return false; | |
7764 | } | |
7765 | ||
7766 | if (!is_gimple_assign (stmt)) | |
7767 | return false; | |
7768 | ||
7769 | code = gimple_assign_rhs_code (stmt); | |
7770 | ||
7771 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
7772 | return false; | |
7773 | ||
7774 | rhs1 = gimple_assign_rhs1 (stmt); | |
7775 | rhs2 = gimple_assign_rhs2 (stmt); | |
7776 | ||
7777 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &def_stmt, | |
7778 | &dts[0], &vectype1)) | |
7779 | return false; | |
7780 | ||
7781 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &def_stmt, | |
7782 | &dts[1], &vectype2)) | |
7783 | return false; | |
7784 | ||
7785 | if (vectype1 && vectype2 | |
7786 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
7787 | return false; | |
7788 | ||
7789 | vectype = vectype1 ? vectype1 : vectype2; | |
7790 | ||
7791 | /* Invariant comparison. */ | |
7792 | if (!vectype) | |
7793 | { | |
69a9a66f RB |
7794 | vectype = get_vectype_for_scalar_type (TREE_TYPE (rhs1)); |
7795 | if (TYPE_VECTOR_SUBPARTS (vectype) != nunits) | |
42fd8198 IE |
7796 | return false; |
7797 | } | |
7798 | else if (nunits != TYPE_VECTOR_SUBPARTS (vectype)) | |
7799 | return false; | |
7800 | ||
7801 | if (!vec_stmt) | |
7802 | { | |
7803 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
7804 | vect_model_simple_cost (stmt_info, ncopies, dts, NULL, NULL); | |
7805 | return expand_vec_cmp_expr_p (vectype, mask_type); | |
7806 | } | |
7807 | ||
7808 | /* Transform. */ | |
7809 | if (!slp_node) | |
7810 | { | |
7811 | vec_oprnds0.create (1); | |
7812 | vec_oprnds1.create (1); | |
7813 | } | |
7814 | ||
7815 | /* Handle def. */ | |
7816 | lhs = gimple_assign_lhs (stmt); | |
7817 | mask = vect_create_destination_var (lhs, mask_type); | |
7818 | ||
7819 | /* Handle cmp expr. */ | |
7820 | for (j = 0; j < ncopies; j++) | |
7821 | { | |
7822 | gassign *new_stmt = NULL; | |
7823 | if (j == 0) | |
7824 | { | |
7825 | if (slp_node) | |
7826 | { | |
7827 | auto_vec<tree, 2> ops; | |
7828 | auto_vec<vec<tree>, 2> vec_defs; | |
7829 | ||
7830 | ops.safe_push (rhs1); | |
7831 | ops.safe_push (rhs2); | |
7832 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); | |
7833 | vec_oprnds1 = vec_defs.pop (); | |
7834 | vec_oprnds0 = vec_defs.pop (); | |
7835 | } | |
7836 | else | |
7837 | { | |
e4af0bc4 IE |
7838 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, vectype); |
7839 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, vectype); | |
42fd8198 IE |
7840 | } |
7841 | } | |
7842 | else | |
7843 | { | |
7844 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
7845 | vec_oprnds0.pop ()); | |
7846 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
7847 | vec_oprnds1.pop ()); | |
7848 | } | |
7849 | ||
7850 | if (!slp_node) | |
7851 | { | |
7852 | vec_oprnds0.quick_push (vec_rhs1); | |
7853 | vec_oprnds1.quick_push (vec_rhs2); | |
7854 | } | |
7855 | ||
7856 | /* Arguments are ready. Create the new vector stmt. */ | |
7857 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
7858 | { | |
7859 | vec_rhs2 = vec_oprnds1[i]; | |
7860 | ||
7861 | new_temp = make_ssa_name (mask); | |
7862 | new_stmt = gimple_build_assign (new_temp, code, vec_rhs1, vec_rhs2); | |
7863 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7864 | if (slp_node) | |
7865 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
7866 | } | |
7867 | ||
7868 | if (slp_node) | |
7869 | continue; | |
7870 | ||
7871 | if (j == 0) | |
7872 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7873 | else | |
7874 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7875 | ||
7876 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7877 | } | |
7878 | ||
7879 | vec_oprnds0.release (); | |
7880 | vec_oprnds1.release (); | |
7881 | ||
7882 | return true; | |
7883 | } | |
ebfd146a | 7884 | |
8644a673 | 7885 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
7886 | |
7887 | bool | |
355fe088 | 7888 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 7889 | { |
8644a673 | 7890 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 7891 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 7892 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 7893 | bool ok; |
a70d6342 | 7894 | tree scalar_type, vectype; |
355fe088 | 7895 | gimple *pattern_stmt; |
363477c0 | 7896 | gimple_seq pattern_def_seq; |
ebfd146a | 7897 | |
73fbfcad | 7898 | if (dump_enabled_p ()) |
ebfd146a | 7899 | { |
78c60e3d SS |
7900 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
7901 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 7902 | } |
ebfd146a | 7903 | |
1825a1f3 | 7904 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 7905 | { |
73fbfcad | 7906 | if (dump_enabled_p ()) |
78c60e3d | 7907 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7908 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
7909 | |
7910 | return false; | |
7911 | } | |
b8698a0f L |
7912 | |
7913 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
7914 | to include: |
7915 | - the COND_EXPR which is the loop exit condition | |
7916 | - any LABEL_EXPRs in the loop | |
b8698a0f | 7917 | - computations that are used only for array indexing or loop control. |
8644a673 | 7918 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 7919 | instance, therefore, all the statements are relevant. |
ebfd146a | 7920 | |
d092494c | 7921 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 7922 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
7923 | statements. In basic blocks we are called from some SLP instance |
7924 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
7925 | already will be part of SLP instance. */ | |
83197f37 IR |
7926 | |
7927 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 7928 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 7929 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 7930 | { |
9d5e7640 | 7931 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 7932 | && pattern_stmt |
9d5e7640 IR |
7933 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
7934 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
7935 | { | |
83197f37 | 7936 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
7937 | stmt = pattern_stmt; |
7938 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 7939 | if (dump_enabled_p ()) |
9d5e7640 | 7940 | { |
78c60e3d SS |
7941 | dump_printf_loc (MSG_NOTE, vect_location, |
7942 | "==> examining pattern statement: "); | |
7943 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
7944 | } |
7945 | } | |
7946 | else | |
7947 | { | |
73fbfcad | 7948 | if (dump_enabled_p ()) |
e645e942 | 7949 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 7950 | |
9d5e7640 IR |
7951 | return true; |
7952 | } | |
8644a673 | 7953 | } |
83197f37 | 7954 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 7955 | && node == NULL |
83197f37 IR |
7956 | && pattern_stmt |
7957 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
7958 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
7959 | { | |
7960 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 7961 | if (dump_enabled_p ()) |
83197f37 | 7962 | { |
78c60e3d SS |
7963 | dump_printf_loc (MSG_NOTE, vect_location, |
7964 | "==> examining pattern statement: "); | |
7965 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
7966 | } |
7967 | ||
7968 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
7969 | return false; | |
7970 | } | |
ebfd146a | 7971 | |
1107f3ae | 7972 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 7973 | && node == NULL |
363477c0 | 7974 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 7975 | { |
363477c0 | 7976 | gimple_stmt_iterator si; |
1107f3ae | 7977 | |
363477c0 JJ |
7978 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
7979 | { | |
355fe088 | 7980 | gimple *pattern_def_stmt = gsi_stmt (si); |
363477c0 JJ |
7981 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) |
7982 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
7983 | { | |
7984 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 7985 | if (dump_enabled_p ()) |
363477c0 | 7986 | { |
78c60e3d SS |
7987 | dump_printf_loc (MSG_NOTE, vect_location, |
7988 | "==> examining pattern def statement: "); | |
7989 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 7990 | } |
1107f3ae | 7991 | |
363477c0 JJ |
7992 | if (!vect_analyze_stmt (pattern_def_stmt, |
7993 | need_to_vectorize, node)) | |
7994 | return false; | |
7995 | } | |
7996 | } | |
7997 | } | |
1107f3ae | 7998 | |
8644a673 IR |
7999 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
8000 | { | |
8001 | case vect_internal_def: | |
8002 | break; | |
ebfd146a | 8003 | |
8644a673 | 8004 | case vect_reduction_def: |
7c5222ff | 8005 | case vect_nested_cycle: |
14a61437 RB |
8006 | gcc_assert (!bb_vinfo |
8007 | && (relevance == vect_used_in_outer | |
8008 | || relevance == vect_used_in_outer_by_reduction | |
8009 | || relevance == vect_used_by_reduction | |
8010 | || relevance == vect_unused_in_scope)); | |
8644a673 IR |
8011 | break; |
8012 | ||
8013 | case vect_induction_def: | |
8014 | case vect_constant_def: | |
8015 | case vect_external_def: | |
8016 | case vect_unknown_def_type: | |
8017 | default: | |
8018 | gcc_unreachable (); | |
8019 | } | |
ebfd146a | 8020 | |
a70d6342 IR |
8021 | if (bb_vinfo) |
8022 | { | |
8023 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
8024 | ||
b690cc0f | 8025 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
73fbfcad | 8026 | if (dump_enabled_p ()) |
a70d6342 | 8027 | { |
78c60e3d SS |
8028 | dump_printf_loc (MSG_NOTE, vect_location, |
8029 | "get vectype for scalar type: "); | |
8030 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
e645e942 | 8031 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
8032 | } |
8033 | ||
8034 | vectype = get_vectype_for_scalar_type (scalar_type); | |
8035 | if (!vectype) | |
8036 | { | |
73fbfcad | 8037 | if (dump_enabled_p ()) |
a70d6342 | 8038 | { |
78c60e3d SS |
8039 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8040 | "not SLPed: unsupported data-type "); | |
8041 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
8042 | scalar_type); | |
e645e942 | 8043 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
a70d6342 IR |
8044 | } |
8045 | return false; | |
8046 | } | |
8047 | ||
73fbfcad | 8048 | if (dump_enabled_p ()) |
a70d6342 | 8049 | { |
78c60e3d SS |
8050 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); |
8051 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
e645e942 | 8052 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
8053 | } |
8054 | ||
8055 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
8056 | } | |
8057 | ||
8644a673 | 8058 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 8059 | { |
8644a673 | 8060 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
8061 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
8062 | || (is_gimple_call (stmt) | |
8063 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 8064 | *need_to_vectorize = true; |
ebfd146a IR |
8065 | } |
8066 | ||
b1af7da6 RB |
8067 | if (PURE_SLP_STMT (stmt_info) && !node) |
8068 | { | |
8069 | dump_printf_loc (MSG_NOTE, vect_location, | |
8070 | "handled only by SLP analysis\n"); | |
8071 | return true; | |
8072 | } | |
8073 | ||
8074 | ok = true; | |
8075 | if (!bb_vinfo | |
8076 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
8077 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
8078 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8079 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8080 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8081 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8082 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8083 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8084 | || vectorizable_call (stmt, NULL, NULL, node) | |
8085 | || vectorizable_store (stmt, NULL, NULL, node) | |
8086 | || vectorizable_reduction (stmt, NULL, NULL, node) | |
42fd8198 IE |
8087 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8088 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 RB |
8089 | else |
8090 | { | |
8091 | if (bb_vinfo) | |
8092 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8093 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8094 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8095 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8096 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8097 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8098 | || vectorizable_call (stmt, NULL, NULL, node) | |
8099 | || vectorizable_store (stmt, NULL, NULL, node) | |
42fd8198 IE |
8100 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8101 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 | 8102 | } |
8644a673 IR |
8103 | |
8104 | if (!ok) | |
ebfd146a | 8105 | { |
73fbfcad | 8106 | if (dump_enabled_p ()) |
8644a673 | 8107 | { |
78c60e3d SS |
8108 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8109 | "not vectorized: relevant stmt not "); | |
8110 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
8111 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 8112 | } |
b8698a0f | 8113 | |
ebfd146a IR |
8114 | return false; |
8115 | } | |
8116 | ||
a70d6342 IR |
8117 | if (bb_vinfo) |
8118 | return true; | |
8119 | ||
8644a673 IR |
8120 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
8121 | need extra handling, except for vectorizable reductions. */ | |
8122 | if (STMT_VINFO_LIVE_P (stmt_info) | |
8123 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
8124 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 8125 | |
8644a673 | 8126 | if (!ok) |
ebfd146a | 8127 | { |
73fbfcad | 8128 | if (dump_enabled_p ()) |
8644a673 | 8129 | { |
78c60e3d SS |
8130 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8131 | "not vectorized: live stmt not "); | |
8132 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
8133 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 8134 | } |
b8698a0f | 8135 | |
8644a673 | 8136 | return false; |
ebfd146a IR |
8137 | } |
8138 | ||
ebfd146a IR |
8139 | return true; |
8140 | } | |
8141 | ||
8142 | ||
8143 | /* Function vect_transform_stmt. | |
8144 | ||
8145 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
8146 | ||
8147 | bool | |
355fe088 | 8148 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 8149 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
8150 | slp_instance slp_node_instance) |
8151 | { | |
8152 | bool is_store = false; | |
355fe088 | 8153 | gimple *vec_stmt = NULL; |
ebfd146a | 8154 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
ebfd146a | 8155 | bool done; |
ebfd146a | 8156 | |
fce57248 | 8157 | gcc_assert (slp_node || !PURE_SLP_STMT (stmt_info)); |
355fe088 | 8158 | gimple *old_vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 8159 | |
ebfd146a IR |
8160 | switch (STMT_VINFO_TYPE (stmt_info)) |
8161 | { | |
8162 | case type_demotion_vec_info_type: | |
ebfd146a | 8163 | case type_promotion_vec_info_type: |
ebfd146a IR |
8164 | case type_conversion_vec_info_type: |
8165 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
8166 | gcc_assert (done); | |
8167 | break; | |
8168 | ||
8169 | case induc_vec_info_type: | |
8170 | gcc_assert (!slp_node); | |
8171 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
8172 | gcc_assert (done); | |
8173 | break; | |
8174 | ||
9dc3f7de IR |
8175 | case shift_vec_info_type: |
8176 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
8177 | gcc_assert (done); | |
8178 | break; | |
8179 | ||
ebfd146a IR |
8180 | case op_vec_info_type: |
8181 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
8182 | gcc_assert (done); | |
8183 | break; | |
8184 | ||
8185 | case assignment_vec_info_type: | |
8186 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
8187 | gcc_assert (done); | |
8188 | break; | |
8189 | ||
8190 | case load_vec_info_type: | |
b8698a0f | 8191 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
8192 | slp_node_instance); |
8193 | gcc_assert (done); | |
8194 | break; | |
8195 | ||
8196 | case store_vec_info_type: | |
8197 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
8198 | gcc_assert (done); | |
0d0293ac | 8199 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
8200 | { |
8201 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 8202 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
8203 | one are skipped, and there vec_stmt_info shouldn't be freed |
8204 | meanwhile. */ | |
0d0293ac | 8205 | *grouped_store = true; |
ebfd146a IR |
8206 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
8207 | is_store = true; | |
8208 | } | |
8209 | else | |
8210 | is_store = true; | |
8211 | break; | |
8212 | ||
8213 | case condition_vec_info_type: | |
f7e531cf | 8214 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
8215 | gcc_assert (done); |
8216 | break; | |
8217 | ||
42fd8198 IE |
8218 | case comparison_vec_info_type: |
8219 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node); | |
8220 | gcc_assert (done); | |
8221 | break; | |
8222 | ||
ebfd146a | 8223 | case call_vec_info_type: |
190c2236 | 8224 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 8225 | stmt = gsi_stmt (*gsi); |
5ce9450f JJ |
8226 | if (is_gimple_call (stmt) |
8227 | && gimple_call_internal_p (stmt) | |
8228 | && gimple_call_internal_fn (stmt) == IFN_MASK_STORE) | |
8229 | is_store = true; | |
ebfd146a IR |
8230 | break; |
8231 | ||
0136f8f0 AH |
8232 | case call_simd_clone_vec_info_type: |
8233 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node); | |
8234 | stmt = gsi_stmt (*gsi); | |
8235 | break; | |
8236 | ||
ebfd146a | 8237 | case reduc_vec_info_type: |
b5aeb3bb | 8238 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
8239 | gcc_assert (done); |
8240 | break; | |
8241 | ||
8242 | default: | |
8243 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
8244 | { | |
73fbfcad | 8245 | if (dump_enabled_p ()) |
78c60e3d | 8246 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8247 | "stmt not supported.\n"); |
ebfd146a IR |
8248 | gcc_unreachable (); |
8249 | } | |
8250 | } | |
8251 | ||
225ce44b RB |
8252 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
8253 | This would break hybrid SLP vectorization. */ | |
8254 | if (slp_node) | |
d90f8440 RB |
8255 | gcc_assert (!vec_stmt |
8256 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt); | |
225ce44b | 8257 | |
ebfd146a IR |
8258 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
8259 | is being vectorized, but outside the immediately enclosing loop. */ | |
8260 | if (vec_stmt | |
a70d6342 IR |
8261 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
8262 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
8263 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
8264 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
8265 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 8266 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 8267 | vect_used_in_outer_by_reduction)) |
ebfd146a | 8268 | { |
a70d6342 IR |
8269 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
8270 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
8271 | imm_use_iterator imm_iter; |
8272 | use_operand_p use_p; | |
8273 | tree scalar_dest; | |
355fe088 | 8274 | gimple *exit_phi; |
ebfd146a | 8275 | |
73fbfcad | 8276 | if (dump_enabled_p ()) |
78c60e3d | 8277 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 8278 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
8279 | |
8280 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
8281 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
8282 | STMT). */ | |
8283 | if (gimple_code (stmt) == GIMPLE_PHI) | |
8284 | scalar_dest = PHI_RESULT (stmt); | |
8285 | else | |
8286 | scalar_dest = gimple_assign_lhs (stmt); | |
8287 | ||
8288 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
8289 | { | |
8290 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
8291 | { | |
8292 | exit_phi = USE_STMT (use_p); | |
8293 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
8294 | } | |
8295 | } | |
8296 | } | |
8297 | ||
8298 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
8299 | being vectorized. */ | |
8300 | if (STMT_VINFO_LIVE_P (stmt_info) | |
8301 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
8302 | { | |
8303 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
8304 | gcc_assert (done); | |
8305 | } | |
8306 | ||
8307 | if (vec_stmt) | |
83197f37 | 8308 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 8309 | |
b8698a0f | 8310 | return is_store; |
ebfd146a IR |
8311 | } |
8312 | ||
8313 | ||
b8698a0f | 8314 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
8315 | stmt_vec_info. */ |
8316 | ||
8317 | void | |
355fe088 | 8318 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 8319 | { |
355fe088 TS |
8320 | gimple *next = first_stmt; |
8321 | gimple *tmp; | |
ebfd146a IR |
8322 | gimple_stmt_iterator next_si; |
8323 | ||
8324 | while (next) | |
8325 | { | |
78048b1c JJ |
8326 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
8327 | ||
8328 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
8329 | if (is_pattern_stmt_p (stmt_info)) | |
8330 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
8331 | /* Free the attached stmt_vec_info and remove the stmt. */ |
8332 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 8333 | unlink_stmt_vdef (next); |
ebfd146a | 8334 | gsi_remove (&next_si, true); |
3d3f2249 | 8335 | release_defs (next); |
ebfd146a IR |
8336 | free_stmt_vec_info (next); |
8337 | next = tmp; | |
8338 | } | |
8339 | } | |
8340 | ||
8341 | ||
8342 | /* Function new_stmt_vec_info. | |
8343 | ||
8344 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
8345 | ||
8346 | stmt_vec_info | |
310213d4 | 8347 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
8348 | { |
8349 | stmt_vec_info res; | |
8350 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
8351 | ||
8352 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
8353 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 8354 | res->vinfo = vinfo; |
8644a673 | 8355 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
8356 | STMT_VINFO_LIVE_P (res) = false; |
8357 | STMT_VINFO_VECTYPE (res) = NULL; | |
8358 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 8359 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
8360 | STMT_VINFO_IN_PATTERN_P (res) = false; |
8361 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 8362 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 8363 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 8364 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
ebfd146a IR |
8365 | |
8366 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
8367 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
8368 | STMT_VINFO_DR_INIT (res) = NULL; | |
8369 | STMT_VINFO_DR_STEP (res) = NULL; | |
8370 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
8371 | ||
8372 | if (gimple_code (stmt) == GIMPLE_PHI | |
8373 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
8374 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
8375 | else | |
8644a673 IR |
8376 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
8377 | ||
9771b263 | 8378 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 8379 | STMT_SLP_TYPE (res) = loop_vect; |
78810bd3 RB |
8380 | STMT_VINFO_NUM_SLP_USES (res) = 0; |
8381 | ||
e14c1050 IR |
8382 | GROUP_FIRST_ELEMENT (res) = NULL; |
8383 | GROUP_NEXT_ELEMENT (res) = NULL; | |
8384 | GROUP_SIZE (res) = 0; | |
8385 | GROUP_STORE_COUNT (res) = 0; | |
8386 | GROUP_GAP (res) = 0; | |
8387 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
8388 | |
8389 | return res; | |
8390 | } | |
8391 | ||
8392 | ||
8393 | /* Create a hash table for stmt_vec_info. */ | |
8394 | ||
8395 | void | |
8396 | init_stmt_vec_info_vec (void) | |
8397 | { | |
9771b263 DN |
8398 | gcc_assert (!stmt_vec_info_vec.exists ()); |
8399 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
8400 | } |
8401 | ||
8402 | ||
8403 | /* Free hash table for stmt_vec_info. */ | |
8404 | ||
8405 | void | |
8406 | free_stmt_vec_info_vec (void) | |
8407 | { | |
93675444 | 8408 | unsigned int i; |
3161455c | 8409 | stmt_vec_info info; |
93675444 JJ |
8410 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) |
8411 | if (info != NULL) | |
3161455c | 8412 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
9771b263 DN |
8413 | gcc_assert (stmt_vec_info_vec.exists ()); |
8414 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
8415 | } |
8416 | ||
8417 | ||
8418 | /* Free stmt vectorization related info. */ | |
8419 | ||
8420 | void | |
355fe088 | 8421 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
8422 | { |
8423 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8424 | ||
8425 | if (!stmt_info) | |
8426 | return; | |
8427 | ||
78048b1c JJ |
8428 | /* Check if this statement has a related "pattern stmt" |
8429 | (introduced by the vectorizer during the pattern recognition | |
8430 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
8431 | too. */ | |
8432 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
8433 | { | |
8434 | stmt_vec_info patt_info | |
8435 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
8436 | if (patt_info) | |
8437 | { | |
363477c0 | 8438 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
355fe088 | 8439 | gimple *patt_stmt = STMT_VINFO_STMT (patt_info); |
f0281fde RB |
8440 | gimple_set_bb (patt_stmt, NULL); |
8441 | tree lhs = gimple_get_lhs (patt_stmt); | |
e6f5c25d | 8442 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde | 8443 | release_ssa_name (lhs); |
363477c0 JJ |
8444 | if (seq) |
8445 | { | |
8446 | gimple_stmt_iterator si; | |
8447 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
f0281fde | 8448 | { |
355fe088 | 8449 | gimple *seq_stmt = gsi_stmt (si); |
f0281fde | 8450 | gimple_set_bb (seq_stmt, NULL); |
7532abf2 | 8451 | lhs = gimple_get_lhs (seq_stmt); |
e6f5c25d | 8452 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde RB |
8453 | release_ssa_name (lhs); |
8454 | free_stmt_vec_info (seq_stmt); | |
8455 | } | |
363477c0 | 8456 | } |
f0281fde | 8457 | free_stmt_vec_info (patt_stmt); |
78048b1c JJ |
8458 | } |
8459 | } | |
8460 | ||
9771b263 | 8461 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 8462 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
8463 | set_vinfo_for_stmt (stmt, NULL); |
8464 | free (stmt_info); | |
8465 | } | |
8466 | ||
8467 | ||
bb67d9c7 | 8468 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 8469 | |
bb67d9c7 | 8470 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
8471 | by the target. */ |
8472 | ||
bb67d9c7 RG |
8473 | static tree |
8474 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a | 8475 | { |
ef4bddc2 RS |
8476 | machine_mode inner_mode = TYPE_MODE (scalar_type); |
8477 | machine_mode simd_mode; | |
2f816591 | 8478 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
8479 | int nunits; |
8480 | tree vectype; | |
8481 | ||
cc4b5170 | 8482 | if (nbytes == 0) |
ebfd146a IR |
8483 | return NULL_TREE; |
8484 | ||
48f2e373 RB |
8485 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
8486 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
8487 | return NULL_TREE; | |
8488 | ||
7b7b1813 RG |
8489 | /* For vector types of elements whose mode precision doesn't |
8490 | match their types precision we use a element type of mode | |
8491 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
8492 | they support the proper result truncation/extension. |
8493 | We also make sure to build vector types with INTEGER_TYPE | |
8494 | component type only. */ | |
6d7971b8 | 8495 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
8496 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
8497 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
8498 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
8499 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 8500 | |
ccbf5bb4 RG |
8501 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
8502 | When the component mode passes the above test simply use a type | |
8503 | corresponding to that mode. The theory is that any use that | |
8504 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 8505 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 8506 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
8507 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
8508 | ||
8509 | /* We can't build a vector type of elements with alignment bigger than | |
8510 | their size. */ | |
dfc2e2ac | 8511 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
8512 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
8513 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 8514 | |
dfc2e2ac RB |
8515 | /* If we felt back to using the mode fail if there was |
8516 | no scalar type for it. */ | |
8517 | if (scalar_type == NULL_TREE) | |
8518 | return NULL_TREE; | |
8519 | ||
bb67d9c7 RG |
8520 | /* If no size was supplied use the mode the target prefers. Otherwise |
8521 | lookup a vector mode of the specified size. */ | |
8522 | if (size == 0) | |
8523 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
8524 | else | |
8525 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
8526 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
8527 | if (nunits <= 1) | |
8528 | return NULL_TREE; | |
ebfd146a IR |
8529 | |
8530 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
8531 | |
8532 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
8533 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 8534 | return NULL_TREE; |
ebfd146a IR |
8535 | |
8536 | return vectype; | |
8537 | } | |
8538 | ||
bb67d9c7 RG |
8539 | unsigned int current_vector_size; |
8540 | ||
8541 | /* Function get_vectype_for_scalar_type. | |
8542 | ||
8543 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
8544 | by the target. */ | |
8545 | ||
8546 | tree | |
8547 | get_vectype_for_scalar_type (tree scalar_type) | |
8548 | { | |
8549 | tree vectype; | |
8550 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
8551 | current_vector_size); | |
8552 | if (vectype | |
8553 | && current_vector_size == 0) | |
8554 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
8555 | return vectype; | |
8556 | } | |
8557 | ||
42fd8198 IE |
8558 | /* Function get_mask_type_for_scalar_type. |
8559 | ||
8560 | Returns the mask type corresponding to a result of comparison | |
8561 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
8562 | ||
8563 | tree | |
8564 | get_mask_type_for_scalar_type (tree scalar_type) | |
8565 | { | |
8566 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
8567 | ||
8568 | if (!vectype) | |
8569 | return NULL; | |
8570 | ||
8571 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
8572 | current_vector_size); | |
8573 | } | |
8574 | ||
b690cc0f RG |
8575 | /* Function get_same_sized_vectype |
8576 | ||
8577 | Returns a vector type corresponding to SCALAR_TYPE of size | |
8578 | VECTOR_TYPE if supported by the target. */ | |
8579 | ||
8580 | tree | |
bb67d9c7 | 8581 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 8582 | { |
9f47c7e5 IE |
8583 | if (TREE_CODE (scalar_type) == BOOLEAN_TYPE) |
8584 | return build_same_sized_truth_vector_type (vector_type); | |
8585 | ||
bb67d9c7 RG |
8586 | return get_vectype_for_scalar_type_and_size |
8587 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
8588 | } |
8589 | ||
ebfd146a IR |
8590 | /* Function vect_is_simple_use. |
8591 | ||
8592 | Input: | |
81c40241 RB |
8593 | VINFO - the vect info of the loop or basic block that is being vectorized. |
8594 | OPERAND - operand in the loop or bb. | |
8595 | Output: | |
8596 | DEF_STMT - the defining stmt in case OPERAND is an SSA_NAME. | |
8597 | DT - the type of definition | |
ebfd146a IR |
8598 | |
8599 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 8600 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 8601 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 8602 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
8603 | is the case in reduction/induction computations). |
8604 | For basic blocks, supportable operands are constants and bb invariants. | |
8605 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
8606 | |
8607 | bool | |
81c40241 RB |
8608 | vect_is_simple_use (tree operand, vec_info *vinfo, |
8609 | gimple **def_stmt, enum vect_def_type *dt) | |
b8698a0f | 8610 | { |
ebfd146a | 8611 | *def_stmt = NULL; |
3fc356dc | 8612 | *dt = vect_unknown_def_type; |
b8698a0f | 8613 | |
73fbfcad | 8614 | if (dump_enabled_p ()) |
ebfd146a | 8615 | { |
78c60e3d SS |
8616 | dump_printf_loc (MSG_NOTE, vect_location, |
8617 | "vect_is_simple_use: operand "); | |
8618 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 8619 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 8620 | } |
b8698a0f | 8621 | |
b758f602 | 8622 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
8623 | { |
8624 | *dt = vect_constant_def; | |
8625 | return true; | |
8626 | } | |
b8698a0f | 8627 | |
ebfd146a IR |
8628 | if (is_gimple_min_invariant (operand)) |
8629 | { | |
8644a673 | 8630 | *dt = vect_external_def; |
ebfd146a IR |
8631 | return true; |
8632 | } | |
8633 | ||
ebfd146a IR |
8634 | if (TREE_CODE (operand) != SSA_NAME) |
8635 | { | |
73fbfcad | 8636 | if (dump_enabled_p ()) |
af29617a AH |
8637 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8638 | "not ssa-name.\n"); | |
ebfd146a IR |
8639 | return false; |
8640 | } | |
b8698a0f | 8641 | |
3fc356dc | 8642 | if (SSA_NAME_IS_DEFAULT_DEF (operand)) |
ebfd146a | 8643 | { |
3fc356dc RB |
8644 | *dt = vect_external_def; |
8645 | return true; | |
ebfd146a IR |
8646 | } |
8647 | ||
3fc356dc | 8648 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
73fbfcad | 8649 | if (dump_enabled_p ()) |
ebfd146a | 8650 | { |
78c60e3d SS |
8651 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
8652 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
8653 | } |
8654 | ||
61d371eb | 8655 | if (! vect_stmt_in_region_p (vinfo, *def_stmt)) |
8644a673 | 8656 | *dt = vect_external_def; |
ebfd146a IR |
8657 | else |
8658 | { | |
3fc356dc | 8659 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (*def_stmt); |
603cca93 | 8660 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
ebfd146a IR |
8661 | } |
8662 | ||
2e8ab70c RB |
8663 | if (dump_enabled_p ()) |
8664 | { | |
8665 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: "); | |
8666 | switch (*dt) | |
8667 | { | |
8668 | case vect_uninitialized_def: | |
8669 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
8670 | break; | |
8671 | case vect_constant_def: | |
8672 | dump_printf (MSG_NOTE, "constant\n"); | |
8673 | break; | |
8674 | case vect_external_def: | |
8675 | dump_printf (MSG_NOTE, "external\n"); | |
8676 | break; | |
8677 | case vect_internal_def: | |
8678 | dump_printf (MSG_NOTE, "internal\n"); | |
8679 | break; | |
8680 | case vect_induction_def: | |
8681 | dump_printf (MSG_NOTE, "induction\n"); | |
8682 | break; | |
8683 | case vect_reduction_def: | |
8684 | dump_printf (MSG_NOTE, "reduction\n"); | |
8685 | break; | |
8686 | case vect_double_reduction_def: | |
8687 | dump_printf (MSG_NOTE, "double reduction\n"); | |
8688 | break; | |
8689 | case vect_nested_cycle: | |
8690 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
8691 | break; | |
8692 | case vect_unknown_def_type: | |
8693 | dump_printf (MSG_NOTE, "unknown\n"); | |
8694 | break; | |
8695 | } | |
8696 | } | |
8697 | ||
81c40241 | 8698 | if (*dt == vect_unknown_def_type) |
ebfd146a | 8699 | { |
73fbfcad | 8700 | if (dump_enabled_p ()) |
78c60e3d | 8701 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8702 | "Unsupported pattern.\n"); |
ebfd146a IR |
8703 | return false; |
8704 | } | |
8705 | ||
ebfd146a IR |
8706 | switch (gimple_code (*def_stmt)) |
8707 | { | |
8708 | case GIMPLE_PHI: | |
ebfd146a | 8709 | case GIMPLE_ASSIGN: |
ebfd146a | 8710 | case GIMPLE_CALL: |
81c40241 | 8711 | break; |
ebfd146a | 8712 | default: |
73fbfcad | 8713 | if (dump_enabled_p ()) |
78c60e3d | 8714 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8715 | "unsupported defining stmt:\n"); |
ebfd146a IR |
8716 | return false; |
8717 | } | |
8718 | ||
8719 | return true; | |
8720 | } | |
8721 | ||
81c40241 | 8722 | /* Function vect_is_simple_use. |
b690cc0f | 8723 | |
81c40241 | 8724 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
8725 | type of OPERAND and stores it to *VECTYPE. If the definition of |
8726 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
8727 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
8728 | is responsible to compute the best suited vector type for the | |
8729 | scalar operand. */ | |
8730 | ||
8731 | bool | |
81c40241 RB |
8732 | vect_is_simple_use (tree operand, vec_info *vinfo, |
8733 | gimple **def_stmt, enum vect_def_type *dt, tree *vectype) | |
b690cc0f | 8734 | { |
81c40241 | 8735 | if (!vect_is_simple_use (operand, vinfo, def_stmt, dt)) |
b690cc0f RG |
8736 | return false; |
8737 | ||
8738 | /* Now get a vector type if the def is internal, otherwise supply | |
8739 | NULL_TREE and leave it up to the caller to figure out a proper | |
8740 | type for the use stmt. */ | |
8741 | if (*dt == vect_internal_def | |
8742 | || *dt == vect_induction_def | |
8743 | || *dt == vect_reduction_def | |
8744 | || *dt == vect_double_reduction_def | |
8745 | || *dt == vect_nested_cycle) | |
8746 | { | |
8747 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
8748 | |
8749 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
8750 | && !STMT_VINFO_RELEVANT (stmt_info) | |
8751 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 8752 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 8753 | |
b690cc0f RG |
8754 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
8755 | gcc_assert (*vectype != NULL_TREE); | |
8756 | } | |
8757 | else if (*dt == vect_uninitialized_def | |
8758 | || *dt == vect_constant_def | |
8759 | || *dt == vect_external_def) | |
8760 | *vectype = NULL_TREE; | |
8761 | else | |
8762 | gcc_unreachable (); | |
8763 | ||
8764 | return true; | |
8765 | } | |
8766 | ||
ebfd146a IR |
8767 | |
8768 | /* Function supportable_widening_operation | |
8769 | ||
b8698a0f L |
8770 | Check whether an operation represented by the code CODE is a |
8771 | widening operation that is supported by the target platform in | |
b690cc0f RG |
8772 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
8773 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 8774 | |
ebfd146a IR |
8775 | Widening operations we currently support are NOP (CONVERT), FLOAT |
8776 | and WIDEN_MULT. This function checks if these operations are supported | |
8777 | by the target platform either directly (via vector tree-codes), or via | |
8778 | target builtins. | |
8779 | ||
8780 | Output: | |
b8698a0f L |
8781 | - CODE1 and CODE2 are codes of vector operations to be used when |
8782 | vectorizing the operation, if available. | |
ebfd146a IR |
8783 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
8784 | case of multi-step conversion (like char->short->int - in that case | |
8785 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
8786 | - INTERM_TYPES contains the intermediate type required to perform the |
8787 | widening operation (short in the above example). */ | |
ebfd146a IR |
8788 | |
8789 | bool | |
355fe088 | 8790 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 8791 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
8792 | enum tree_code *code1, enum tree_code *code2, |
8793 | int *multi_step_cvt, | |
9771b263 | 8794 | vec<tree> *interm_types) |
ebfd146a IR |
8795 | { |
8796 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8797 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 8798 | struct loop *vect_loop = NULL; |
ef4bddc2 | 8799 | machine_mode vec_mode; |
81f40b79 | 8800 | enum insn_code icode1, icode2; |
ebfd146a | 8801 | optab optab1, optab2; |
b690cc0f RG |
8802 | tree vectype = vectype_in; |
8803 | tree wide_vectype = vectype_out; | |
ebfd146a | 8804 | enum tree_code c1, c2; |
4a00c761 JJ |
8805 | int i; |
8806 | tree prev_type, intermediate_type; | |
ef4bddc2 | 8807 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 8808 | optab optab3, optab4; |
ebfd146a | 8809 | |
4a00c761 | 8810 | *multi_step_cvt = 0; |
4ef69dfc IR |
8811 | if (loop_info) |
8812 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
8813 | ||
ebfd146a IR |
8814 | switch (code) |
8815 | { | |
8816 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
8817 | /* The result of a vectorized widening operation usually requires |
8818 | two vectors (because the widened results do not fit into one vector). | |
8819 | The generated vector results would normally be expected to be | |
8820 | generated in the same order as in the original scalar computation, | |
8821 | i.e. if 8 results are generated in each vector iteration, they are | |
8822 | to be organized as follows: | |
8823 | vect1: [res1,res2,res3,res4], | |
8824 | vect2: [res5,res6,res7,res8]. | |
8825 | ||
8826 | However, in the special case that the result of the widening | |
8827 | operation is used in a reduction computation only, the order doesn't | |
8828 | matter (because when vectorizing a reduction we change the order of | |
8829 | the computation). Some targets can take advantage of this and | |
8830 | generate more efficient code. For example, targets like Altivec, | |
8831 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
8832 | generate the following vectors: | |
8833 | vect1: [res1,res3,res5,res7], | |
8834 | vect2: [res2,res4,res6,res8]. | |
8835 | ||
8836 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
8837 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
8838 | iterations in parallel). We therefore don't allow to change the | |
8839 | order of the computation in the inner-loop during outer-loop | |
8840 | vectorization. */ | |
8841 | /* TODO: Another case in which order doesn't *really* matter is when we | |
8842 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
8843 | Normally, pack_trunc performs an even/odd permute, whereas the | |
8844 | repack from an even/odd expansion would be an interleave, which | |
8845 | would be significantly simpler for e.g. AVX2. */ | |
8846 | /* In any case, in order to avoid duplicating the code below, recurse | |
8847 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
8848 | are properly set up for the caller. If we fail, we'll continue with | |
8849 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
8850 | if (vect_loop | |
8851 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
8852 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
8853 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
8854 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
8855 | code1, code2, multi_step_cvt, |
8856 | interm_types)) | |
ebc047a2 CH |
8857 | { |
8858 | /* Elements in a vector with vect_used_by_reduction property cannot | |
8859 | be reordered if the use chain with this property does not have the | |
8860 | same operation. One such an example is s += a * b, where elements | |
8861 | in a and b cannot be reordered. Here we check if the vector defined | |
8862 | by STMT is only directly used in the reduction statement. */ | |
8863 | tree lhs = gimple_assign_lhs (stmt); | |
8864 | use_operand_p dummy; | |
355fe088 | 8865 | gimple *use_stmt; |
ebc047a2 CH |
8866 | stmt_vec_info use_stmt_info = NULL; |
8867 | if (single_imm_use (lhs, &dummy, &use_stmt) | |
8868 | && (use_stmt_info = vinfo_for_stmt (use_stmt)) | |
8869 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
8870 | return true; | |
8871 | } | |
4a00c761 JJ |
8872 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
8873 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
8874 | break; |
8875 | ||
81c40241 RB |
8876 | case DOT_PROD_EXPR: |
8877 | c1 = DOT_PROD_EXPR; | |
8878 | c2 = DOT_PROD_EXPR; | |
8879 | break; | |
8880 | ||
8881 | case SAD_EXPR: | |
8882 | c1 = SAD_EXPR; | |
8883 | c2 = SAD_EXPR; | |
8884 | break; | |
8885 | ||
6ae6116f RH |
8886 | case VEC_WIDEN_MULT_EVEN_EXPR: |
8887 | /* Support the recursion induced just above. */ | |
8888 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
8889 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
8890 | break; | |
8891 | ||
36ba4aae | 8892 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
8893 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
8894 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
8895 | break; |
8896 | ||
ebfd146a | 8897 | CASE_CONVERT: |
4a00c761 JJ |
8898 | c1 = VEC_UNPACK_LO_EXPR; |
8899 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
8900 | break; |
8901 | ||
8902 | case FLOAT_EXPR: | |
4a00c761 JJ |
8903 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
8904 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
8905 | break; |
8906 | ||
8907 | case FIX_TRUNC_EXPR: | |
8908 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
8909 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
8910 | computing the operation. */ | |
8911 | return false; | |
8912 | ||
8913 | default: | |
8914 | gcc_unreachable (); | |
8915 | } | |
8916 | ||
6ae6116f | 8917 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 8918 | std::swap (c1, c2); |
4a00c761 | 8919 | |
ebfd146a IR |
8920 | if (code == FIX_TRUNC_EXPR) |
8921 | { | |
8922 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
8923 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
8924 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
8925 | } |
8926 | else | |
8927 | { | |
8928 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
8929 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
8930 | } | |
8931 | ||
8932 | if (!optab1 || !optab2) | |
8933 | return false; | |
8934 | ||
8935 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
8936 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
8937 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
8938 | return false; |
8939 | ||
4a00c761 JJ |
8940 | *code1 = c1; |
8941 | *code2 = c2; | |
8942 | ||
8943 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
8944 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
8945 | /* For scalar masks we may have different boolean |
8946 | vector types having the same QImode. Thus we | |
8947 | add additional check for elements number. */ | |
8948 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
8949 | || (TYPE_VECTOR_SUBPARTS (vectype) / 2 | |
8950 | == TYPE_VECTOR_SUBPARTS (wide_vectype))); | |
4a00c761 | 8951 | |
b8698a0f | 8952 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 8953 | types. */ |
ebfd146a | 8954 | |
4a00c761 JJ |
8955 | prev_type = vectype; |
8956 | prev_mode = vec_mode; | |
b8698a0f | 8957 | |
4a00c761 JJ |
8958 | if (!CONVERT_EXPR_CODE_P (code)) |
8959 | return false; | |
b8698a0f | 8960 | |
4a00c761 JJ |
8961 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
8962 | intermediate steps in promotion sequence. We try | |
8963 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
8964 | not. */ | |
9771b263 | 8965 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
8966 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
8967 | { | |
8968 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
8969 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
8970 | { | |
8971 | intermediate_type | |
8972 | = build_truth_vector_type (TYPE_VECTOR_SUBPARTS (prev_type) / 2, | |
8973 | current_vector_size); | |
8974 | if (intermediate_mode != TYPE_MODE (intermediate_type)) | |
8975 | return false; | |
8976 | } | |
8977 | else | |
8978 | intermediate_type | |
8979 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
8980 | TYPE_UNSIGNED (prev_type)); | |
8981 | ||
4a00c761 JJ |
8982 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); |
8983 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
8984 | ||
8985 | if (!optab3 || !optab4 | |
8986 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
8987 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
8988 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
8989 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
8990 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
8991 | == CODE_FOR_nothing) | |
8992 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
8993 | == CODE_FOR_nothing)) | |
8994 | break; | |
ebfd146a | 8995 | |
9771b263 | 8996 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
8997 | (*multi_step_cvt)++; |
8998 | ||
8999 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
9000 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
9001 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
9002 | || (TYPE_VECTOR_SUBPARTS (intermediate_type) / 2 | |
9003 | == TYPE_VECTOR_SUBPARTS (wide_vectype))); | |
4a00c761 JJ |
9004 | |
9005 | prev_type = intermediate_type; | |
9006 | prev_mode = intermediate_mode; | |
ebfd146a IR |
9007 | } |
9008 | ||
9771b263 | 9009 | interm_types->release (); |
4a00c761 | 9010 | return false; |
ebfd146a IR |
9011 | } |
9012 | ||
9013 | ||
9014 | /* Function supportable_narrowing_operation | |
9015 | ||
b8698a0f L |
9016 | Check whether an operation represented by the code CODE is a |
9017 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
9018 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
9019 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 9020 | |
ebfd146a | 9021 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 9022 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
9023 | the target platform directly via vector tree-codes. |
9024 | ||
9025 | Output: | |
b8698a0f L |
9026 | - CODE1 is the code of a vector operation to be used when |
9027 | vectorizing the operation, if available. | |
ebfd146a IR |
9028 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
9029 | case of multi-step conversion (like int->short->char - in that case | |
9030 | MULTI_STEP_CVT will be 1). | |
9031 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 9032 | narrowing operation (short in the above example). */ |
ebfd146a IR |
9033 | |
9034 | bool | |
9035 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 9036 | tree vectype_out, tree vectype_in, |
ebfd146a | 9037 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 9038 | vec<tree> *interm_types) |
ebfd146a | 9039 | { |
ef4bddc2 | 9040 | machine_mode vec_mode; |
ebfd146a IR |
9041 | enum insn_code icode1; |
9042 | optab optab1, interm_optab; | |
b690cc0f RG |
9043 | tree vectype = vectype_in; |
9044 | tree narrow_vectype = vectype_out; | |
ebfd146a | 9045 | enum tree_code c1; |
3ae0661a | 9046 | tree intermediate_type, prev_type; |
ef4bddc2 | 9047 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 9048 | int i; |
4a00c761 | 9049 | bool uns; |
ebfd146a | 9050 | |
4a00c761 | 9051 | *multi_step_cvt = 0; |
ebfd146a IR |
9052 | switch (code) |
9053 | { | |
9054 | CASE_CONVERT: | |
9055 | c1 = VEC_PACK_TRUNC_EXPR; | |
9056 | break; | |
9057 | ||
9058 | case FIX_TRUNC_EXPR: | |
9059 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
9060 | break; | |
9061 | ||
9062 | case FLOAT_EXPR: | |
9063 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
9064 | tree code and optabs used for computing the operation. */ | |
9065 | return false; | |
9066 | ||
9067 | default: | |
9068 | gcc_unreachable (); | |
9069 | } | |
9070 | ||
9071 | if (code == FIX_TRUNC_EXPR) | |
9072 | /* The signedness is determined from output operand. */ | |
b690cc0f | 9073 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
9074 | else |
9075 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
9076 | ||
9077 | if (!optab1) | |
9078 | return false; | |
9079 | ||
9080 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 9081 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
9082 | return false; |
9083 | ||
4a00c761 JJ |
9084 | *code1 = c1; |
9085 | ||
9086 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
9087 | /* For scalar masks we may have different boolean |
9088 | vector types having the same QImode. Thus we | |
9089 | add additional check for elements number. */ | |
9090 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
9091 | || (TYPE_VECTOR_SUBPARTS (vectype) * 2 | |
9092 | == TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 | 9093 | |
ebfd146a IR |
9094 | /* Check if it's a multi-step conversion that can be done using intermediate |
9095 | types. */ | |
4a00c761 | 9096 | prev_mode = vec_mode; |
3ae0661a | 9097 | prev_type = vectype; |
4a00c761 JJ |
9098 | if (code == FIX_TRUNC_EXPR) |
9099 | uns = TYPE_UNSIGNED (vectype_out); | |
9100 | else | |
9101 | uns = TYPE_UNSIGNED (vectype); | |
9102 | ||
9103 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
9104 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
9105 | costly than signed. */ | |
9106 | if (code == FIX_TRUNC_EXPR && uns) | |
9107 | { | |
9108 | enum insn_code icode2; | |
9109 | ||
9110 | intermediate_type | |
9111 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
9112 | interm_optab | |
9113 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 9114 | if (interm_optab != unknown_optab |
4a00c761 JJ |
9115 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
9116 | && insn_data[icode1].operand[0].mode | |
9117 | == insn_data[icode2].operand[0].mode) | |
9118 | { | |
9119 | uns = false; | |
9120 | optab1 = interm_optab; | |
9121 | icode1 = icode2; | |
9122 | } | |
9123 | } | |
ebfd146a | 9124 | |
4a00c761 JJ |
9125 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
9126 | intermediate steps in promotion sequence. We try | |
9127 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 9128 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
9129 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
9130 | { | |
9131 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
9132 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
9133 | { | |
9134 | intermediate_type | |
9135 | = build_truth_vector_type (TYPE_VECTOR_SUBPARTS (prev_type) * 2, | |
9136 | current_vector_size); | |
9137 | if (intermediate_mode != TYPE_MODE (intermediate_type)) | |
9138 | return false; | |
9139 | } | |
9140 | else | |
9141 | intermediate_type | |
9142 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
4a00c761 JJ |
9143 | interm_optab |
9144 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
9145 | optab_default); | |
9146 | if (!interm_optab | |
9147 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
9148 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
9149 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
9150 | == CODE_FOR_nothing)) | |
9151 | break; | |
9152 | ||
9771b263 | 9153 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
9154 | (*multi_step_cvt)++; |
9155 | ||
9156 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
9157 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
9158 | || (TYPE_VECTOR_SUBPARTS (intermediate_type) * 2 | |
9159 | == TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 JJ |
9160 | |
9161 | prev_mode = intermediate_mode; | |
3ae0661a | 9162 | prev_type = intermediate_type; |
4a00c761 | 9163 | optab1 = interm_optab; |
ebfd146a IR |
9164 | } |
9165 | ||
9771b263 | 9166 | interm_types->release (); |
4a00c761 | 9167 | return false; |
ebfd146a | 9168 | } |