]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
85ec4feb | 2 | Copyright (C) 2003-2018 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
c7131fb2 | 25 | #include "backend.h" |
957060b5 AM |
26 | #include "target.h" |
27 | #include "rtl.h" | |
ebfd146a | 28 | #include "tree.h" |
c7131fb2 | 29 | #include "gimple.h" |
c7131fb2 | 30 | #include "ssa.h" |
957060b5 AM |
31 | #include "optabs-tree.h" |
32 | #include "insn-config.h" | |
33 | #include "recog.h" /* FIXME: for insn_data */ | |
34 | #include "cgraph.h" | |
957060b5 | 35 | #include "dumpfile.h" |
c7131fb2 | 36 | #include "alias.h" |
40e23961 | 37 | #include "fold-const.h" |
d8a2d370 | 38 | #include "stor-layout.h" |
2fb9a547 | 39 | #include "tree-eh.h" |
45b0be94 | 40 | #include "gimplify.h" |
5be5c238 | 41 | #include "gimple-iterator.h" |
18f429e2 | 42 | #include "gimplify-me.h" |
442b4905 | 43 | #include "tree-cfg.h" |
e28030cf | 44 | #include "tree-ssa-loop-manip.h" |
ebfd146a | 45 | #include "cfgloop.h" |
0136f8f0 AH |
46 | #include "tree-ssa-loop.h" |
47 | #include "tree-scalar-evolution.h" | |
ebfd146a | 48 | #include "tree-vectorizer.h" |
9b2b7279 | 49 | #include "builtins.h" |
70439f0d | 50 | #include "internal-fn.h" |
5ebaa477 | 51 | #include "tree-vector-builder.h" |
f151c9e1 | 52 | #include "vec-perm-indices.h" |
7cfb4d93 RS |
53 | #include "tree-ssa-loop-niter.h" |
54 | #include "gimple-fold.h" | |
ebfd146a | 55 | |
7ee2468b SB |
56 | /* For lang_hooks.types.type_for_mode. */ |
57 | #include "langhooks.h" | |
ebfd146a | 58 | |
c3e7ee41 BS |
59 | /* Return the vectorized type for the given statement. */ |
60 | ||
61 | tree | |
62 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
63 | { | |
64 | return STMT_VINFO_VECTYPE (stmt_info); | |
65 | } | |
66 | ||
67 | /* Return TRUE iff the given statement is in an inner loop relative to | |
68 | the loop being vectorized. */ | |
69 | bool | |
70 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
71 | { | |
355fe088 | 72 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
c3e7ee41 BS |
73 | basic_block bb = gimple_bb (stmt); |
74 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
75 | struct loop* loop; | |
76 | ||
77 | if (!loop_vinfo) | |
78 | return false; | |
79 | ||
80 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
81 | ||
82 | return (bb->loop_father == loop->inner); | |
83 | } | |
84 | ||
85 | /* Record the cost of a statement, either by directly informing the | |
86 | target model or by saving it in a vector for later processing. | |
87 | Return a preliminary estimate of the statement's cost. */ | |
88 | ||
89 | unsigned | |
92345349 | 90 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 91 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 92 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 93 | { |
cc9fe6bb JH |
94 | if ((kind == vector_load || kind == unaligned_load) |
95 | && STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
96 | kind = vector_gather_load; | |
97 | if ((kind == vector_store || kind == unaligned_store) | |
98 | && STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
99 | kind = vector_scatter_store; | |
92345349 | 100 | if (body_cost_vec) |
c3e7ee41 | 101 | { |
92345349 | 102 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
ddf56386 RB |
103 | stmt_info_for_cost si = { count, kind, |
104 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
105 | misalign }; | |
106 | body_cost_vec->safe_push (si); | |
c3e7ee41 | 107 | return (unsigned) |
92345349 | 108 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
109 | } |
110 | else | |
310213d4 RB |
111 | return add_stmt_cost (stmt_info->vinfo->target_cost_data, |
112 | count, kind, stmt_info, misalign, where); | |
c3e7ee41 BS |
113 | } |
114 | ||
272c6793 RS |
115 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
116 | ||
117 | static tree | |
118 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
119 | { | |
120 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
121 | "vect_array"); | |
122 | } | |
123 | ||
124 | /* ARRAY is an array of vectors created by create_vector_array. | |
125 | Return an SSA_NAME for the vector in index N. The reference | |
126 | is part of the vectorization of STMT and the vector is associated | |
127 | with scalar destination SCALAR_DEST. */ | |
128 | ||
129 | static tree | |
355fe088 | 130 | read_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree scalar_dest, |
272c6793 RS |
131 | tree array, unsigned HOST_WIDE_INT n) |
132 | { | |
133 | tree vect_type, vect, vect_name, array_ref; | |
355fe088 | 134 | gimple *new_stmt; |
272c6793 RS |
135 | |
136 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
137 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
138 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
139 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
140 | build_int_cst (size_type_node, n), | |
141 | NULL_TREE, NULL_TREE); | |
142 | ||
143 | new_stmt = gimple_build_assign (vect, array_ref); | |
144 | vect_name = make_ssa_name (vect, new_stmt); | |
145 | gimple_assign_set_lhs (new_stmt, vect_name); | |
146 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
147 | |
148 | return vect_name; | |
149 | } | |
150 | ||
151 | /* ARRAY is an array of vectors created by create_vector_array. | |
152 | Emit code to store SSA_NAME VECT in index N of the array. | |
153 | The store is part of the vectorization of STMT. */ | |
154 | ||
155 | static void | |
355fe088 | 156 | write_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree vect, |
272c6793 RS |
157 | tree array, unsigned HOST_WIDE_INT n) |
158 | { | |
159 | tree array_ref; | |
355fe088 | 160 | gimple *new_stmt; |
272c6793 RS |
161 | |
162 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
163 | build_int_cst (size_type_node, n), | |
164 | NULL_TREE, NULL_TREE); | |
165 | ||
166 | new_stmt = gimple_build_assign (array_ref, vect); | |
167 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
168 | } |
169 | ||
170 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
171 | of *PTR. The memory reference replaces those in FIRST_DR | |
172 | (and its group). */ | |
173 | ||
174 | static tree | |
44fc7854 | 175 | create_array_ref (tree type, tree ptr, tree alias_ptr_type) |
272c6793 | 176 | { |
44fc7854 | 177 | tree mem_ref; |
272c6793 | 178 | |
272c6793 RS |
179 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); |
180 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 181 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
182 | return mem_ref; |
183 | } | |
184 | ||
ebfd146a IR |
185 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
186 | ||
187 | /* Function vect_mark_relevant. | |
188 | ||
189 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
190 | ||
191 | static void | |
355fe088 | 192 | vect_mark_relevant (vec<gimple *> *worklist, gimple *stmt, |
97ecdb46 | 193 | enum vect_relevant relevant, bool live_p) |
ebfd146a IR |
194 | { |
195 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
196 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
197 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
355fe088 | 198 | gimple *pattern_stmt; |
ebfd146a | 199 | |
73fbfcad | 200 | if (dump_enabled_p ()) |
66c16fd9 RB |
201 | { |
202 | dump_printf_loc (MSG_NOTE, vect_location, | |
203 | "mark relevant %d, live %d: ", relevant, live_p); | |
204 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
205 | } | |
ebfd146a | 206 | |
83197f37 IR |
207 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
208 | related pattern stmt instead of the original stmt. However, such stmts | |
209 | may have their own uses that are not in any pattern, in such cases the | |
210 | stmt itself should be marked. */ | |
ebfd146a IR |
211 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
212 | { | |
97ecdb46 JJ |
213 | /* This is the last stmt in a sequence that was detected as a |
214 | pattern that can potentially be vectorized. Don't mark the stmt | |
215 | as relevant/live because it's not going to be vectorized. | |
216 | Instead mark the pattern-stmt that replaces it. */ | |
83197f37 | 217 | |
97ecdb46 JJ |
218 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); |
219 | ||
220 | if (dump_enabled_p ()) | |
221 | dump_printf_loc (MSG_NOTE, vect_location, | |
222 | "last stmt in pattern. don't mark" | |
223 | " relevant/live.\n"); | |
224 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
225 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
226 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
227 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
228 | stmt = pattern_stmt; | |
ebfd146a IR |
229 | } |
230 | ||
231 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
232 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
233 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
234 | ||
235 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
236 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
237 | { | |
73fbfcad | 238 | if (dump_enabled_p ()) |
78c60e3d | 239 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 240 | "already marked relevant/live.\n"); |
ebfd146a IR |
241 | return; |
242 | } | |
243 | ||
9771b263 | 244 | worklist->safe_push (stmt); |
ebfd146a IR |
245 | } |
246 | ||
247 | ||
b28ead45 AH |
248 | /* Function is_simple_and_all_uses_invariant |
249 | ||
250 | Return true if STMT is simple and all uses of it are invariant. */ | |
251 | ||
252 | bool | |
253 | is_simple_and_all_uses_invariant (gimple *stmt, loop_vec_info loop_vinfo) | |
254 | { | |
255 | tree op; | |
256 | gimple *def_stmt; | |
257 | ssa_op_iter iter; | |
258 | ||
259 | if (!is_gimple_assign (stmt)) | |
260 | return false; | |
261 | ||
262 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
263 | { | |
264 | enum vect_def_type dt = vect_uninitialized_def; | |
265 | ||
266 | if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt)) | |
267 | { | |
268 | if (dump_enabled_p ()) | |
269 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
270 | "use not simple.\n"); | |
271 | return false; | |
272 | } | |
273 | ||
274 | if (dt != vect_external_def && dt != vect_constant_def) | |
275 | return false; | |
276 | } | |
277 | return true; | |
278 | } | |
279 | ||
ebfd146a IR |
280 | /* Function vect_stmt_relevant_p. |
281 | ||
282 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
283 | "relevant for vectorization". | |
284 | ||
285 | A stmt is considered "relevant for vectorization" if: | |
286 | - it has uses outside the loop. | |
287 | - it has vdefs (it alters memory). | |
288 | - control stmts in the loop (except for the exit condition). | |
289 | ||
290 | CHECKME: what other side effects would the vectorizer allow? */ | |
291 | ||
292 | static bool | |
355fe088 | 293 | vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo, |
ebfd146a IR |
294 | enum vect_relevant *relevant, bool *live_p) |
295 | { | |
296 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
297 | ssa_op_iter op_iter; | |
298 | imm_use_iterator imm_iter; | |
299 | use_operand_p use_p; | |
300 | def_operand_p def_p; | |
301 | ||
8644a673 | 302 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
303 | *live_p = false; |
304 | ||
305 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
306 | if (is_ctrl_stmt (stmt) |
307 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
308 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 309 | *relevant = vect_used_in_scope; |
ebfd146a IR |
310 | |
311 | /* changing memory. */ | |
312 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
313 | if (gimple_vdef (stmt) |
314 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 315 | { |
73fbfcad | 316 | if (dump_enabled_p ()) |
78c60e3d | 317 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 318 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 319 | *relevant = vect_used_in_scope; |
ebfd146a IR |
320 | } |
321 | ||
322 | /* uses outside the loop. */ | |
323 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
324 | { | |
325 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
326 | { | |
327 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
328 | if (!flow_bb_inside_loop_p (loop, bb)) | |
329 | { | |
73fbfcad | 330 | if (dump_enabled_p ()) |
78c60e3d | 331 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 332 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 333 | |
3157b0c2 AO |
334 | if (is_gimple_debug (USE_STMT (use_p))) |
335 | continue; | |
336 | ||
ebfd146a IR |
337 | /* We expect all such uses to be in the loop exit phis |
338 | (because of loop closed form) */ | |
339 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
340 | gcc_assert (bb == single_exit (loop)->dest); | |
341 | ||
342 | *live_p = true; | |
343 | } | |
344 | } | |
345 | } | |
346 | ||
3a2edf4c AH |
347 | if (*live_p && *relevant == vect_unused_in_scope |
348 | && !is_simple_and_all_uses_invariant (stmt, loop_vinfo)) | |
b28ead45 AH |
349 | { |
350 | if (dump_enabled_p ()) | |
351 | dump_printf_loc (MSG_NOTE, vect_location, | |
352 | "vec_stmt_relevant_p: stmt live but not relevant.\n"); | |
353 | *relevant = vect_used_only_live; | |
354 | } | |
355 | ||
ebfd146a IR |
356 | return (*live_p || *relevant); |
357 | } | |
358 | ||
359 | ||
b8698a0f | 360 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 361 | |
ff802fa1 | 362 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
363 | used in STMT for anything other than indexing an array. */ |
364 | ||
365 | static bool | |
355fe088 | 366 | exist_non_indexing_operands_for_use_p (tree use, gimple *stmt) |
ebfd146a IR |
367 | { |
368 | tree operand; | |
369 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 370 | |
ff802fa1 | 371 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
372 | reference in STMT, then any operand that corresponds to USE |
373 | is not indexing an array. */ | |
374 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
375 | return true; | |
59a05b0c | 376 | |
ebfd146a IR |
377 | /* STMT has a data_ref. FORNOW this means that its of one of |
378 | the following forms: | |
379 | -1- ARRAY_REF = var | |
380 | -2- var = ARRAY_REF | |
381 | (This should have been verified in analyze_data_refs). | |
382 | ||
383 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 384 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
385 | for array indexing. |
386 | ||
387 | Therefore, all we need to check is if STMT falls into the | |
388 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
389 | |
390 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
391 | { |
392 | if (is_gimple_call (stmt) | |
393 | && gimple_call_internal_p (stmt)) | |
bfaa08b7 RS |
394 | { |
395 | internal_fn ifn = gimple_call_internal_fn (stmt); | |
396 | int mask_index = internal_fn_mask_index (ifn); | |
397 | if (mask_index >= 0 | |
398 | && use == gimple_call_arg (stmt, mask_index)) | |
399 | return true; | |
f307441a RS |
400 | int stored_value_index = internal_fn_stored_value_index (ifn); |
401 | if (stored_value_index >= 0 | |
402 | && use == gimple_call_arg (stmt, stored_value_index)) | |
403 | return true; | |
bfaa08b7 RS |
404 | if (internal_gather_scatter_fn_p (ifn) |
405 | && use == gimple_call_arg (stmt, 1)) | |
406 | return true; | |
bfaa08b7 | 407 | } |
5ce9450f JJ |
408 | return false; |
409 | } | |
410 | ||
59a05b0c EB |
411 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
412 | return false; | |
ebfd146a | 413 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
414 | if (TREE_CODE (operand) != SSA_NAME) |
415 | return false; | |
416 | ||
417 | if (operand == use) | |
418 | return true; | |
419 | ||
420 | return false; | |
421 | } | |
422 | ||
423 | ||
b8698a0f | 424 | /* |
ebfd146a IR |
425 | Function process_use. |
426 | ||
427 | Inputs: | |
428 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b28ead45 | 429 | - RELEVANT - enum value to be set in the STMT_VINFO of the stmt |
ff802fa1 | 430 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 431 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
432 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
433 | be performed. | |
ebfd146a IR |
434 | |
435 | Outputs: | |
436 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
437 | relevance info of the DEF_STMT of this USE: | |
438 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
439 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
440 | Exceptions: | |
441 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 442 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 443 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
444 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
445 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
446 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
447 | be modified accordingly. | |
448 | ||
449 | Return true if everything is as expected. Return false otherwise. */ | |
450 | ||
451 | static bool | |
b28ead45 | 452 | process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo, |
355fe088 | 453 | enum vect_relevant relevant, vec<gimple *> *worklist, |
aec7ae7d | 454 | bool force) |
ebfd146a IR |
455 | { |
456 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
457 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
458 | stmt_vec_info dstmt_vinfo; | |
459 | basic_block bb, def_bb; | |
355fe088 | 460 | gimple *def_stmt; |
ebfd146a IR |
461 | enum vect_def_type dt; |
462 | ||
b8698a0f | 463 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 464 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 465 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
466 | return true; |
467 | ||
81c40241 | 468 | if (!vect_is_simple_use (use, loop_vinfo, &def_stmt, &dt)) |
b8698a0f | 469 | { |
73fbfcad | 470 | if (dump_enabled_p ()) |
78c60e3d | 471 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 472 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
473 | return false; |
474 | } | |
475 | ||
476 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
477 | return true; | |
478 | ||
479 | def_bb = gimple_bb (def_stmt); | |
480 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
481 | { | |
73fbfcad | 482 | if (dump_enabled_p ()) |
e645e942 | 483 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
484 | return true; |
485 | } | |
486 | ||
b8698a0f L |
487 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
488 | DEF_STMT must have already been processed, because this should be the | |
489 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
490 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
491 | check that everything is as expected, and we are done. */ |
492 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
493 | bb = gimple_bb (stmt); | |
494 | if (gimple_code (stmt) == GIMPLE_PHI | |
495 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
496 | && gimple_code (def_stmt) != GIMPLE_PHI | |
497 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
498 | && bb->loop_father == def_bb->loop_father) | |
499 | { | |
73fbfcad | 500 | if (dump_enabled_p ()) |
78c60e3d | 501 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 502 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
503 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
504 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
505 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 506 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 507 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
508 | return true; |
509 | } | |
510 | ||
511 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
512 | outer-loop-header-bb: | |
513 | d = def_stmt | |
514 | inner-loop: | |
515 | stmt # use (d) | |
516 | outer-loop-tail-bb: | |
517 | ... */ | |
518 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
519 | { | |
73fbfcad | 520 | if (dump_enabled_p ()) |
78c60e3d | 521 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 522 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 523 | |
ebfd146a IR |
524 | switch (relevant) |
525 | { | |
8644a673 | 526 | case vect_unused_in_scope: |
7c5222ff IR |
527 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
528 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 529 | break; |
7c5222ff | 530 | |
ebfd146a | 531 | case vect_used_in_outer_by_reduction: |
7c5222ff | 532 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
533 | relevant = vect_used_by_reduction; |
534 | break; | |
7c5222ff | 535 | |
ebfd146a | 536 | case vect_used_in_outer: |
7c5222ff | 537 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 538 | relevant = vect_used_in_scope; |
ebfd146a | 539 | break; |
7c5222ff | 540 | |
8644a673 | 541 | case vect_used_in_scope: |
ebfd146a IR |
542 | break; |
543 | ||
544 | default: | |
545 | gcc_unreachable (); | |
b8698a0f | 546 | } |
ebfd146a IR |
547 | } |
548 | ||
549 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
550 | outer-loop-header-bb: | |
551 | ... | |
552 | inner-loop: | |
553 | d = def_stmt | |
06066f92 | 554 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
555 | stmt # use (d) */ |
556 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
557 | { | |
73fbfcad | 558 | if (dump_enabled_p ()) |
78c60e3d | 559 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 560 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 561 | |
ebfd146a IR |
562 | switch (relevant) |
563 | { | |
8644a673 | 564 | case vect_unused_in_scope: |
b8698a0f | 565 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 566 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 567 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
568 | break; |
569 | ||
ebfd146a | 570 | case vect_used_by_reduction: |
b28ead45 | 571 | case vect_used_only_live: |
ebfd146a IR |
572 | relevant = vect_used_in_outer_by_reduction; |
573 | break; | |
574 | ||
8644a673 | 575 | case vect_used_in_scope: |
ebfd146a IR |
576 | relevant = vect_used_in_outer; |
577 | break; | |
578 | ||
579 | default: | |
580 | gcc_unreachable (); | |
581 | } | |
582 | } | |
643a9684 RB |
583 | /* We are also not interested in uses on loop PHI backedges that are |
584 | inductions. Otherwise we'll needlessly vectorize the IV increment | |
e294f495 RB |
585 | and cause hybrid SLP for SLP inductions. Unless the PHI is live |
586 | of course. */ | |
643a9684 RB |
587 | else if (gimple_code (stmt) == GIMPLE_PHI |
588 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_induction_def | |
e294f495 | 589 | && ! STMT_VINFO_LIVE_P (stmt_vinfo) |
643a9684 RB |
590 | && (PHI_ARG_DEF_FROM_EDGE (stmt, loop_latch_edge (bb->loop_father)) |
591 | == use)) | |
592 | { | |
593 | if (dump_enabled_p ()) | |
594 | dump_printf_loc (MSG_NOTE, vect_location, | |
595 | "induction value on backedge.\n"); | |
596 | return true; | |
597 | } | |
598 | ||
ebfd146a | 599 | |
b28ead45 | 600 | vect_mark_relevant (worklist, def_stmt, relevant, false); |
ebfd146a IR |
601 | return true; |
602 | } | |
603 | ||
604 | ||
605 | /* Function vect_mark_stmts_to_be_vectorized. | |
606 | ||
607 | Not all stmts in the loop need to be vectorized. For example: | |
608 | ||
609 | for i... | |
610 | for j... | |
611 | 1. T0 = i + j | |
612 | 2. T1 = a[T0] | |
613 | ||
614 | 3. j = j + 1 | |
615 | ||
616 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
617 | addressing of vectorized data-refs are handled differently. | |
618 | ||
619 | This pass detects such stmts. */ | |
620 | ||
621 | bool | |
622 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
623 | { | |
ebfd146a IR |
624 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
625 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
626 | unsigned int nbbs = loop->num_nodes; | |
627 | gimple_stmt_iterator si; | |
355fe088 | 628 | gimple *stmt; |
ebfd146a IR |
629 | unsigned int i; |
630 | stmt_vec_info stmt_vinfo; | |
631 | basic_block bb; | |
355fe088 | 632 | gimple *phi; |
ebfd146a | 633 | bool live_p; |
b28ead45 | 634 | enum vect_relevant relevant; |
ebfd146a | 635 | |
73fbfcad | 636 | if (dump_enabled_p ()) |
78c60e3d | 637 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 638 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 639 | |
355fe088 | 640 | auto_vec<gimple *, 64> worklist; |
ebfd146a IR |
641 | |
642 | /* 1. Init worklist. */ | |
643 | for (i = 0; i < nbbs; i++) | |
644 | { | |
645 | bb = bbs[i]; | |
646 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 647 | { |
ebfd146a | 648 | phi = gsi_stmt (si); |
73fbfcad | 649 | if (dump_enabled_p ()) |
ebfd146a | 650 | { |
78c60e3d SS |
651 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
652 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
653 | } |
654 | ||
655 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 656 | vect_mark_relevant (&worklist, phi, relevant, live_p); |
ebfd146a IR |
657 | } |
658 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
659 | { | |
660 | stmt = gsi_stmt (si); | |
73fbfcad | 661 | if (dump_enabled_p ()) |
ebfd146a | 662 | { |
78c60e3d SS |
663 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
664 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 665 | } |
ebfd146a IR |
666 | |
667 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 668 | vect_mark_relevant (&worklist, stmt, relevant, live_p); |
ebfd146a IR |
669 | } |
670 | } | |
671 | ||
672 | /* 2. Process_worklist */ | |
9771b263 | 673 | while (worklist.length () > 0) |
ebfd146a IR |
674 | { |
675 | use_operand_p use_p; | |
676 | ssa_op_iter iter; | |
677 | ||
9771b263 | 678 | stmt = worklist.pop (); |
73fbfcad | 679 | if (dump_enabled_p ()) |
ebfd146a | 680 | { |
78c60e3d SS |
681 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
682 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
683 | } |
684 | ||
b8698a0f | 685 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
b28ead45 AH |
686 | (DEF_STMT) as relevant/irrelevant according to the relevance property |
687 | of STMT. */ | |
ebfd146a IR |
688 | stmt_vinfo = vinfo_for_stmt (stmt); |
689 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
ebfd146a | 690 | |
b28ead45 AH |
691 | /* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is |
692 | propagated as is to the DEF_STMTs of its USEs. | |
ebfd146a IR |
693 | |
694 | One exception is when STMT has been identified as defining a reduction | |
b28ead45 | 695 | variable; in this case we set the relevance to vect_used_by_reduction. |
ebfd146a | 696 | This is because we distinguish between two kinds of relevant stmts - |
b8698a0f | 697 | those that are used by a reduction computation, and those that are |
ff802fa1 | 698 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 699 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 700 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 701 | |
b28ead45 | 702 | switch (STMT_VINFO_DEF_TYPE (stmt_vinfo)) |
ebfd146a | 703 | { |
06066f92 | 704 | case vect_reduction_def: |
b28ead45 AH |
705 | gcc_assert (relevant != vect_unused_in_scope); |
706 | if (relevant != vect_unused_in_scope | |
707 | && relevant != vect_used_in_scope | |
708 | && relevant != vect_used_by_reduction | |
709 | && relevant != vect_used_only_live) | |
06066f92 | 710 | { |
b28ead45 AH |
711 | if (dump_enabled_p ()) |
712 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
713 | "unsupported use of reduction.\n"); | |
714 | return false; | |
06066f92 | 715 | } |
06066f92 | 716 | break; |
b8698a0f | 717 | |
06066f92 | 718 | case vect_nested_cycle: |
b28ead45 AH |
719 | if (relevant != vect_unused_in_scope |
720 | && relevant != vect_used_in_outer_by_reduction | |
721 | && relevant != vect_used_in_outer) | |
06066f92 | 722 | { |
73fbfcad | 723 | if (dump_enabled_p ()) |
78c60e3d | 724 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 725 | "unsupported use of nested cycle.\n"); |
7c5222ff | 726 | |
06066f92 IR |
727 | return false; |
728 | } | |
b8698a0f L |
729 | break; |
730 | ||
06066f92 | 731 | case vect_double_reduction_def: |
b28ead45 AH |
732 | if (relevant != vect_unused_in_scope |
733 | && relevant != vect_used_by_reduction | |
734 | && relevant != vect_used_only_live) | |
06066f92 | 735 | { |
73fbfcad | 736 | if (dump_enabled_p ()) |
78c60e3d | 737 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 738 | "unsupported use of double reduction.\n"); |
7c5222ff | 739 | |
7c5222ff | 740 | return false; |
06066f92 | 741 | } |
b8698a0f | 742 | break; |
7c5222ff | 743 | |
06066f92 IR |
744 | default: |
745 | break; | |
7c5222ff | 746 | } |
b8698a0f | 747 | |
aec7ae7d | 748 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
749 | { |
750 | /* Pattern statements are not inserted into the code, so | |
751 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
752 | have to scan the RHS or function arguments instead. */ | |
753 | if (is_gimple_assign (stmt)) | |
754 | { | |
69d2aade JJ |
755 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
756 | tree op = gimple_assign_rhs1 (stmt); | |
757 | ||
758 | i = 1; | |
759 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
760 | { | |
761 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
b28ead45 | 762 | relevant, &worklist, false) |
69d2aade | 763 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
b28ead45 | 764 | relevant, &worklist, false)) |
566d377a | 765 | return false; |
69d2aade JJ |
766 | i = 2; |
767 | } | |
768 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 769 | { |
69d2aade | 770 | op = gimple_op (stmt, i); |
afbe6325 | 771 | if (TREE_CODE (op) == SSA_NAME |
b28ead45 | 772 | && !process_use (stmt, op, loop_vinfo, relevant, |
afbe6325 | 773 | &worklist, false)) |
07687835 | 774 | return false; |
9d5e7640 IR |
775 | } |
776 | } | |
777 | else if (is_gimple_call (stmt)) | |
778 | { | |
779 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
780 | { | |
781 | tree arg = gimple_call_arg (stmt, i); | |
b28ead45 | 782 | if (!process_use (stmt, arg, loop_vinfo, relevant, |
aec7ae7d | 783 | &worklist, false)) |
07687835 | 784 | return false; |
9d5e7640 IR |
785 | } |
786 | } | |
787 | } | |
788 | else | |
789 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
790 | { | |
791 | tree op = USE_FROM_PTR (use_p); | |
b28ead45 | 792 | if (!process_use (stmt, op, loop_vinfo, relevant, |
aec7ae7d | 793 | &worklist, false)) |
07687835 | 794 | return false; |
9d5e7640 | 795 | } |
aec7ae7d | 796 | |
3bab6342 | 797 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d | 798 | { |
134c85ca RS |
799 | gather_scatter_info gs_info; |
800 | if (!vect_check_gather_scatter (stmt, loop_vinfo, &gs_info)) | |
801 | gcc_unreachable (); | |
802 | if (!process_use (stmt, gs_info.offset, loop_vinfo, relevant, | |
803 | &worklist, true)) | |
566d377a | 804 | return false; |
aec7ae7d | 805 | } |
ebfd146a IR |
806 | } /* while worklist */ |
807 | ||
ebfd146a IR |
808 | return true; |
809 | } | |
810 | ||
811 | ||
b8698a0f | 812 | /* Function vect_model_simple_cost. |
ebfd146a | 813 | |
b8698a0f | 814 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
815 | single op. Right now, this does not account for multiple insns that could |
816 | be generated for the single vector op. We will handle that shortly. */ | |
817 | ||
818 | void | |
b8698a0f | 819 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 | 820 | enum vect_def_type *dt, |
4fc5ebf1 | 821 | int ndts, |
92345349 BS |
822 | stmt_vector_for_cost *prologue_cost_vec, |
823 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
824 | { |
825 | int i; | |
92345349 | 826 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
827 | |
828 | /* The SLP costs were already calculated during SLP tree build. */ | |
78604de0 | 829 | gcc_assert (!PURE_SLP_STMT (stmt_info)); |
ebfd146a | 830 | |
4fc5ebf1 JG |
831 | /* Cost the "broadcast" of a scalar operand in to a vector operand. |
832 | Use scalar_to_vec to cost the broadcast, as elsewhere in the vector | |
833 | cost model. */ | |
834 | for (i = 0; i < ndts; i++) | |
92345349 | 835 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
4fc5ebf1 | 836 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
92345349 | 837 | stmt_info, 0, vect_prologue); |
c3e7ee41 BS |
838 | |
839 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
840 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
841 | stmt_info, 0, vect_body); | |
c3e7ee41 | 842 | |
73fbfcad | 843 | if (dump_enabled_p ()) |
78c60e3d SS |
844 | dump_printf_loc (MSG_NOTE, vect_location, |
845 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 846 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
847 | } |
848 | ||
849 | ||
8bd37302 BS |
850 | /* Model cost for type demotion and promotion operations. PWR is normally |
851 | zero for single-step promotions and demotions. It will be one if | |
852 | two-step promotion/demotion is required, and so on. Each additional | |
853 | step doubles the number of instructions required. */ | |
854 | ||
855 | static void | |
856 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
857 | enum vect_def_type *dt, int pwr) | |
858 | { | |
859 | int i, tmp; | |
92345349 | 860 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
861 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
862 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
863 | void *target_cost_data; | |
8bd37302 BS |
864 | |
865 | /* The SLP costs were already calculated during SLP tree build. */ | |
78604de0 | 866 | gcc_assert (!PURE_SLP_STMT (stmt_info)); |
8bd37302 | 867 | |
c3e7ee41 BS |
868 | if (loop_vinfo) |
869 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
870 | else | |
871 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
872 | ||
8bd37302 BS |
873 | for (i = 0; i < pwr + 1; i++) |
874 | { | |
875 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
876 | (i + 1) : i; | |
c3e7ee41 | 877 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
878 | vec_promote_demote, stmt_info, 0, |
879 | vect_body); | |
8bd37302 BS |
880 | } |
881 | ||
882 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
883 | for (i = 0; i < 2; i++) | |
92345349 BS |
884 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
885 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
886 | stmt_info, 0, vect_prologue); | |
8bd37302 | 887 | |
73fbfcad | 888 | if (dump_enabled_p ()) |
78c60e3d SS |
889 | dump_printf_loc (MSG_NOTE, vect_location, |
890 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 891 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
892 | } |
893 | ||
ebfd146a IR |
894 | /* Function vect_model_store_cost |
895 | ||
0d0293ac MM |
896 | Models cost for stores. In the case of grouped accesses, one access |
897 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
898 | |
899 | void | |
b8698a0f | 900 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
2de001ee | 901 | vect_memory_access_type memory_access_type, |
9ce4345a | 902 | vec_load_store_type vls_type, slp_tree slp_node, |
92345349 BS |
903 | stmt_vector_for_cost *prologue_cost_vec, |
904 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a | 905 | { |
92345349 | 906 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f RS |
907 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
908 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); | |
909 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); | |
ebfd146a | 910 | |
9ce4345a | 911 | if (vls_type == VLS_STORE_INVARIANT) |
92345349 BS |
912 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
913 | stmt_info, 0, vect_prologue); | |
ebfd146a | 914 | |
892a981f RS |
915 | /* Grouped stores update all elements in the group at once, |
916 | so we want the DR for the first statement. */ | |
917 | if (!slp_node && grouped_access_p) | |
720f5239 | 918 | { |
892a981f RS |
919 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
920 | dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
720f5239 | 921 | } |
ebfd146a | 922 | |
892a981f RS |
923 | /* True if we should include any once-per-group costs as well as |
924 | the cost of the statement itself. For SLP we only get called | |
925 | once per group anyhow. */ | |
926 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
927 | ||
272c6793 | 928 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 929 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 | 930 | access is instead being provided by a permute-and-store operation, |
2de001ee RS |
931 | include the cost of the permutes. */ |
932 | if (first_stmt_p | |
933 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 934 | { |
e1377713 ES |
935 | /* Uses a high and low interleave or shuffle operations for each |
936 | needed permute. */ | |
892a981f | 937 | int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
e1377713 | 938 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
92345349 BS |
939 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
940 | stmt_info, 0, vect_body); | |
ebfd146a | 941 | |
73fbfcad | 942 | if (dump_enabled_p ()) |
78c60e3d | 943 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 944 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 945 | group_size); |
ebfd146a IR |
946 | } |
947 | ||
cee62fee | 948 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 949 | /* Costs of the stores. */ |
067bc855 RB |
950 | if (memory_access_type == VMAT_ELEMENTWISE |
951 | || memory_access_type == VMAT_GATHER_SCATTER) | |
c5126ce8 RS |
952 | { |
953 | /* N scalar stores plus extracting the elements. */ | |
954 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
955 | inside_cost += record_stmt_cost (body_cost_vec, | |
956 | ncopies * assumed_nunits, | |
957 | scalar_store, stmt_info, 0, vect_body); | |
958 | } | |
f2e2a985 | 959 | else |
892a981f | 960 | vect_get_store_cost (dr, ncopies, &inside_cost, body_cost_vec); |
ebfd146a | 961 | |
2de001ee RS |
962 | if (memory_access_type == VMAT_ELEMENTWISE |
963 | || memory_access_type == VMAT_STRIDED_SLP) | |
c5126ce8 RS |
964 | { |
965 | /* N scalar stores plus extracting the elements. */ | |
966 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
967 | inside_cost += record_stmt_cost (body_cost_vec, | |
968 | ncopies * assumed_nunits, | |
969 | vec_to_scalar, stmt_info, 0, vect_body); | |
970 | } | |
cee62fee | 971 | |
73fbfcad | 972 | if (dump_enabled_p ()) |
78c60e3d SS |
973 | dump_printf_loc (MSG_NOTE, vect_location, |
974 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 975 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
976 | } |
977 | ||
978 | ||
720f5239 IR |
979 | /* Calculate cost of DR's memory access. */ |
980 | void | |
981 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 982 | unsigned int *inside_cost, |
92345349 | 983 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
984 | { |
985 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 986 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 987 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
988 | |
989 | switch (alignment_support_scheme) | |
990 | { | |
991 | case dr_aligned: | |
992 | { | |
92345349 BS |
993 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
994 | vector_store, stmt_info, 0, | |
995 | vect_body); | |
720f5239 | 996 | |
73fbfcad | 997 | if (dump_enabled_p ()) |
78c60e3d | 998 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 999 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
1000 | break; |
1001 | } | |
1002 | ||
1003 | case dr_unaligned_supported: | |
1004 | { | |
720f5239 | 1005 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1006 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1007 | unaligned_store, stmt_info, |
92345349 | 1008 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1009 | if (dump_enabled_p ()) |
78c60e3d SS |
1010 | dump_printf_loc (MSG_NOTE, vect_location, |
1011 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1012 | "hardware.\n"); |
720f5239 IR |
1013 | break; |
1014 | } | |
1015 | ||
38eec4c6 UW |
1016 | case dr_unaligned_unsupported: |
1017 | { | |
1018 | *inside_cost = VECT_MAX_COST; | |
1019 | ||
73fbfcad | 1020 | if (dump_enabled_p ()) |
78c60e3d | 1021 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1022 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1023 | break; |
1024 | } | |
1025 | ||
720f5239 IR |
1026 | default: |
1027 | gcc_unreachable (); | |
1028 | } | |
1029 | } | |
1030 | ||
1031 | ||
ebfd146a IR |
1032 | /* Function vect_model_load_cost |
1033 | ||
892a981f RS |
1034 | Models cost for loads. In the case of grouped accesses, one access has |
1035 | the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1036 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1037 | access scheme chosen. */ |
1038 | ||
1039 | void | |
92345349 | 1040 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
2de001ee RS |
1041 | vect_memory_access_type memory_access_type, |
1042 | slp_tree slp_node, | |
92345349 BS |
1043 | stmt_vector_for_cost *prologue_cost_vec, |
1044 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a | 1045 | { |
892a981f RS |
1046 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); |
1047 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
92345349 | 1048 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f | 1049 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
ebfd146a | 1050 | |
892a981f RS |
1051 | /* Grouped loads read all elements in the group at once, |
1052 | so we want the DR for the first statement. */ | |
1053 | if (!slp_node && grouped_access_p) | |
ebfd146a | 1054 | { |
892a981f RS |
1055 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
1056 | dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
ebfd146a IR |
1057 | } |
1058 | ||
892a981f RS |
1059 | /* True if we should include any once-per-group costs as well as |
1060 | the cost of the statement itself. For SLP we only get called | |
1061 | once per group anyhow. */ | |
1062 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
1063 | ||
272c6793 | 1064 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1065 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 | 1066 | access is instead being provided by a load-and-permute operation, |
2de001ee RS |
1067 | include the cost of the permutes. */ |
1068 | if (first_stmt_p | |
1069 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 1070 | { |
2c23db6d ES |
1071 | /* Uses an even and odd extract operations or shuffle operations |
1072 | for each needed permute. */ | |
892a981f | 1073 | int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
2c23db6d ES |
1074 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
1075 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, | |
1076 | stmt_info, 0, vect_body); | |
ebfd146a | 1077 | |
73fbfcad | 1078 | if (dump_enabled_p ()) |
e645e942 TJ |
1079 | dump_printf_loc (MSG_NOTE, vect_location, |
1080 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1081 | group_size); |
ebfd146a IR |
1082 | } |
1083 | ||
1084 | /* The loads themselves. */ | |
067bc855 RB |
1085 | if (memory_access_type == VMAT_ELEMENTWISE |
1086 | || memory_access_type == VMAT_GATHER_SCATTER) | |
a82960aa | 1087 | { |
a21892ad BS |
1088 | /* N scalar loads plus gathering them into a vector. */ |
1089 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
c5126ce8 | 1090 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); |
92345349 | 1091 | inside_cost += record_stmt_cost (body_cost_vec, |
c5126ce8 | 1092 | ncopies * assumed_nunits, |
92345349 | 1093 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1094 | } |
1095 | else | |
892a981f | 1096 | vect_get_load_cost (dr, ncopies, first_stmt_p, |
92345349 BS |
1097 | &inside_cost, &prologue_cost, |
1098 | prologue_cost_vec, body_cost_vec, true); | |
2de001ee RS |
1099 | if (memory_access_type == VMAT_ELEMENTWISE |
1100 | || memory_access_type == VMAT_STRIDED_SLP) | |
892a981f RS |
1101 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, |
1102 | stmt_info, 0, vect_body); | |
720f5239 | 1103 | |
73fbfcad | 1104 | if (dump_enabled_p ()) |
78c60e3d SS |
1105 | dump_printf_loc (MSG_NOTE, vect_location, |
1106 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1107 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1108 | } |
1109 | ||
1110 | ||
1111 | /* Calculate cost of DR's memory access. */ | |
1112 | void | |
1113 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1114 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1115 | unsigned int *prologue_cost, |
1116 | stmt_vector_for_cost *prologue_cost_vec, | |
1117 | stmt_vector_for_cost *body_cost_vec, | |
1118 | bool record_prologue_costs) | |
720f5239 IR |
1119 | { |
1120 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 1121 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 1122 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
1123 | |
1124 | switch (alignment_support_scheme) | |
ebfd146a IR |
1125 | { |
1126 | case dr_aligned: | |
1127 | { | |
92345349 BS |
1128 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1129 | stmt_info, 0, vect_body); | |
ebfd146a | 1130 | |
73fbfcad | 1131 | if (dump_enabled_p ()) |
78c60e3d | 1132 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1133 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1134 | |
1135 | break; | |
1136 | } | |
1137 | case dr_unaligned_supported: | |
1138 | { | |
720f5239 | 1139 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1140 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1141 | unaligned_load, stmt_info, |
92345349 | 1142 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1143 | |
73fbfcad | 1144 | if (dump_enabled_p ()) |
78c60e3d SS |
1145 | dump_printf_loc (MSG_NOTE, vect_location, |
1146 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1147 | "hardware.\n"); |
ebfd146a IR |
1148 | |
1149 | break; | |
1150 | } | |
1151 | case dr_explicit_realign: | |
1152 | { | |
92345349 BS |
1153 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1154 | vector_load, stmt_info, 0, vect_body); | |
1155 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1156 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1157 | |
1158 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1159 | the containing loop, the following cost should be added to the | |
92345349 | 1160 | prologue costs. */ |
ebfd146a | 1161 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1162 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1163 | stmt_info, 0, vect_body); | |
ebfd146a | 1164 | |
73fbfcad | 1165 | if (dump_enabled_p ()) |
e645e942 TJ |
1166 | dump_printf_loc (MSG_NOTE, vect_location, |
1167 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1168 | |
ebfd146a IR |
1169 | break; |
1170 | } | |
1171 | case dr_explicit_realign_optimized: | |
1172 | { | |
73fbfcad | 1173 | if (dump_enabled_p ()) |
e645e942 | 1174 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1175 | "vect_model_load_cost: unaligned software " |
e645e942 | 1176 | "pipelined.\n"); |
ebfd146a IR |
1177 | |
1178 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1179 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1180 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1181 | access, then the above cost should only be considered for one |
ff802fa1 | 1182 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1183 | and a realignment op. */ |
1184 | ||
92345349 | 1185 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1186 | { |
92345349 BS |
1187 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1188 | vector_stmt, stmt_info, | |
1189 | 0, vect_prologue); | |
ebfd146a | 1190 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1191 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1192 | vector_stmt, stmt_info, | |
1193 | 0, vect_prologue); | |
ebfd146a IR |
1194 | } |
1195 | ||
92345349 BS |
1196 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1197 | stmt_info, 0, vect_body); | |
1198 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1199 | stmt_info, 0, vect_body); | |
8bd37302 | 1200 | |
73fbfcad | 1201 | if (dump_enabled_p ()) |
78c60e3d | 1202 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1203 | "vect_model_load_cost: explicit realign optimized" |
1204 | "\n"); | |
8bd37302 | 1205 | |
ebfd146a IR |
1206 | break; |
1207 | } | |
1208 | ||
38eec4c6 UW |
1209 | case dr_unaligned_unsupported: |
1210 | { | |
1211 | *inside_cost = VECT_MAX_COST; | |
1212 | ||
73fbfcad | 1213 | if (dump_enabled_p ()) |
78c60e3d | 1214 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1215 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1216 | break; |
1217 | } | |
1218 | ||
ebfd146a IR |
1219 | default: |
1220 | gcc_unreachable (); | |
1221 | } | |
ebfd146a IR |
1222 | } |
1223 | ||
418b7df3 RG |
1224 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1225 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1226 | |
418b7df3 | 1227 | static void |
355fe088 | 1228 | vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi) |
ebfd146a | 1229 | { |
ebfd146a | 1230 | if (gsi) |
418b7df3 | 1231 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1232 | else |
1233 | { | |
418b7df3 | 1234 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1235 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1236 | |
a70d6342 IR |
1237 | if (loop_vinfo) |
1238 | { | |
1239 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1240 | basic_block new_bb; |
1241 | edge pe; | |
a70d6342 IR |
1242 | |
1243 | if (nested_in_vect_loop_p (loop, stmt)) | |
1244 | loop = loop->inner; | |
b8698a0f | 1245 | |
a70d6342 | 1246 | pe = loop_preheader_edge (loop); |
418b7df3 | 1247 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1248 | gcc_assert (!new_bb); |
1249 | } | |
1250 | else | |
1251 | { | |
1252 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1253 | basic_block bb; | |
1254 | gimple_stmt_iterator gsi_bb_start; | |
1255 | ||
1256 | gcc_assert (bb_vinfo); | |
1257 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1258 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1259 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1260 | } |
ebfd146a IR |
1261 | } |
1262 | ||
73fbfcad | 1263 | if (dump_enabled_p ()) |
ebfd146a | 1264 | { |
78c60e3d SS |
1265 | dump_printf_loc (MSG_NOTE, vect_location, |
1266 | "created new init_stmt: "); | |
1267 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1268 | } |
418b7df3 RG |
1269 | } |
1270 | ||
1271 | /* Function vect_init_vector. | |
ebfd146a | 1272 | |
5467ee52 RG |
1273 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1274 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1275 | vector type a vector with all elements equal to VAL is created first. | |
1276 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1277 | initialization at the loop preheader. | |
418b7df3 RG |
1278 | Return the DEF of INIT_STMT. |
1279 | It will be used in the vectorization of STMT. */ | |
1280 | ||
1281 | tree | |
355fe088 | 1282 | vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 | 1283 | { |
355fe088 | 1284 | gimple *init_stmt; |
418b7df3 RG |
1285 | tree new_temp; |
1286 | ||
e412ece4 RB |
1287 | /* We abuse this function to push sth to a SSA name with initial 'val'. */ |
1288 | if (! useless_type_conversion_p (type, TREE_TYPE (val))) | |
418b7df3 | 1289 | { |
e412ece4 RB |
1290 | gcc_assert (TREE_CODE (type) == VECTOR_TYPE); |
1291 | if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) | |
418b7df3 | 1292 | { |
5a308cf1 IE |
1293 | /* Scalar boolean value should be transformed into |
1294 | all zeros or all ones value before building a vector. */ | |
1295 | if (VECTOR_BOOLEAN_TYPE_P (type)) | |
1296 | { | |
b3d51f23 IE |
1297 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1298 | tree false_val = build_zero_cst (TREE_TYPE (type)); | |
5a308cf1 IE |
1299 | |
1300 | if (CONSTANT_CLASS_P (val)) | |
1301 | val = integer_zerop (val) ? false_val : true_val; | |
1302 | else | |
1303 | { | |
1304 | new_temp = make_ssa_name (TREE_TYPE (type)); | |
1305 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, | |
1306 | val, true_val, false_val); | |
1307 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
1308 | val = new_temp; | |
1309 | } | |
1310 | } | |
1311 | else if (CONSTANT_CLASS_P (val)) | |
42fd8198 | 1312 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1313 | else |
1314 | { | |
b731b390 | 1315 | new_temp = make_ssa_name (TREE_TYPE (type)); |
e412ece4 RB |
1316 | if (! INTEGRAL_TYPE_P (TREE_TYPE (val))) |
1317 | init_stmt = gimple_build_assign (new_temp, | |
1318 | fold_build1 (VIEW_CONVERT_EXPR, | |
1319 | TREE_TYPE (type), | |
1320 | val)); | |
1321 | else | |
1322 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); | |
418b7df3 | 1323 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1324 | val = new_temp; |
418b7df3 RG |
1325 | } |
1326 | } | |
5467ee52 | 1327 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1328 | } |
1329 | ||
0e22bb5a RB |
1330 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1331 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1332 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1333 | return new_temp; |
ebfd146a IR |
1334 | } |
1335 | ||
c83a894c | 1336 | /* Function vect_get_vec_def_for_operand_1. |
a70d6342 | 1337 | |
c83a894c AH |
1338 | For a defining stmt DEF_STMT of a scalar stmt, return a vector def with type |
1339 | DT that will be used in the vectorized stmt. */ | |
ebfd146a IR |
1340 | |
1341 | tree | |
c83a894c | 1342 | vect_get_vec_def_for_operand_1 (gimple *def_stmt, enum vect_def_type dt) |
ebfd146a IR |
1343 | { |
1344 | tree vec_oprnd; | |
355fe088 | 1345 | gimple *vec_stmt; |
ebfd146a | 1346 | stmt_vec_info def_stmt_info = NULL; |
ebfd146a IR |
1347 | |
1348 | switch (dt) | |
1349 | { | |
81c40241 | 1350 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1351 | case vect_constant_def: |
81c40241 | 1352 | case vect_external_def: |
c83a894c AH |
1353 | /* Code should use vect_get_vec_def_for_operand. */ |
1354 | gcc_unreachable (); | |
ebfd146a | 1355 | |
81c40241 | 1356 | /* operand is defined inside the loop. */ |
8644a673 | 1357 | case vect_internal_def: |
ebfd146a | 1358 | { |
ebfd146a IR |
1359 | /* Get the def from the vectorized stmt. */ |
1360 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1361 | |
ebfd146a | 1362 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1363 | /* Get vectorized pattern statement. */ |
1364 | if (!vec_stmt | |
1365 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1366 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1367 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1368 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1369 | gcc_assert (vec_stmt); |
1370 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1371 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1372 | else if (is_gimple_call (vec_stmt)) | |
1373 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1374 | else | |
1375 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1376 | return vec_oprnd; | |
1377 | } | |
1378 | ||
c78e3652 | 1379 | /* operand is defined by a loop header phi. */ |
ebfd146a | 1380 | case vect_reduction_def: |
06066f92 | 1381 | case vect_double_reduction_def: |
7c5222ff | 1382 | case vect_nested_cycle: |
ebfd146a IR |
1383 | case vect_induction_def: |
1384 | { | |
1385 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1386 | ||
1387 | /* Get the def from the vectorized stmt. */ | |
1388 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1389 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1390 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1391 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1392 | else | |
1393 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1394 | return vec_oprnd; |
1395 | } | |
1396 | ||
1397 | default: | |
1398 | gcc_unreachable (); | |
1399 | } | |
1400 | } | |
1401 | ||
1402 | ||
c83a894c AH |
1403 | /* Function vect_get_vec_def_for_operand. |
1404 | ||
1405 | OP is an operand in STMT. This function returns a (vector) def that will be | |
1406 | used in the vectorized stmt for STMT. | |
1407 | ||
1408 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1409 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1410 | ||
1411 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1412 | needs to be introduced. VECTYPE may be used to specify a required type for | |
1413 | vector invariant. */ | |
1414 | ||
1415 | tree | |
1416 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) | |
1417 | { | |
1418 | gimple *def_stmt; | |
1419 | enum vect_def_type dt; | |
1420 | bool is_simple_use; | |
1421 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
1422 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
1423 | ||
1424 | if (dump_enabled_p ()) | |
1425 | { | |
1426 | dump_printf_loc (MSG_NOTE, vect_location, | |
1427 | "vect_get_vec_def_for_operand: "); | |
1428 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
1429 | dump_printf (MSG_NOTE, "\n"); | |
1430 | } | |
1431 | ||
1432 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt); | |
1433 | gcc_assert (is_simple_use); | |
1434 | if (def_stmt && dump_enabled_p ()) | |
1435 | { | |
1436 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1437 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
1438 | } | |
1439 | ||
1440 | if (dt == vect_constant_def || dt == vect_external_def) | |
1441 | { | |
1442 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
1443 | tree vector_type; | |
1444 | ||
1445 | if (vectype) | |
1446 | vector_type = vectype; | |
2568d8a1 | 1447 | else if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op)) |
c83a894c AH |
1448 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) |
1449 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1450 | else | |
1451 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1452 | ||
1453 | gcc_assert (vector_type); | |
1454 | return vect_init_vector (stmt, op, vector_type, NULL); | |
1455 | } | |
1456 | else | |
1457 | return vect_get_vec_def_for_operand_1 (def_stmt, dt); | |
1458 | } | |
1459 | ||
1460 | ||
ebfd146a IR |
1461 | /* Function vect_get_vec_def_for_stmt_copy |
1462 | ||
ff802fa1 | 1463 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1464 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1465 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1466 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1467 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1468 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1469 | DT is the type of the vector def VEC_OPRND. |
1470 | ||
1471 | Context: | |
1472 | In case the vectorization factor (VF) is bigger than the number | |
1473 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1474 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1475 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1476 | smallest data-type determines the VF, and as a result, when vectorizing |
1477 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1478 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1479 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1480 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1481 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1482 | ||
1483 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1484 | |
ebfd146a IR |
1485 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1486 | VS1.1: vx.1 = memref1 VS1.2 | |
1487 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1488 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1489 | |
1490 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1491 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1492 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1493 | VSnew.3: vz3 = vx.3 + ... | |
1494 | ||
1495 | The vectorization of S1 is explained in vectorizable_load. | |
1496 | The vectorization of S2: | |
b8698a0f L |
1497 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1498 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1499 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1500 | returns the vector-def 'vx.0'. |
1501 | ||
b8698a0f L |
1502 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1503 | function is called to get the relevant vector-def for each operand. It is | |
1504 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1505 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1506 | ||
b8698a0f L |
1507 | For example, to obtain the vector-def 'vx.1' in order to create the |
1508 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1509 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1510 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1511 | and return its def ('vx.1'). | |
1512 | Overall, to create the above sequence this function will be called 3 times: | |
1513 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1514 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1515 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1516 | ||
1517 | tree | |
1518 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1519 | { | |
355fe088 | 1520 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1521 | stmt_vec_info def_stmt_info; |
1522 | ||
1523 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1524 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1525 | return vec_oprnd; |
1526 | ||
1527 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1528 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1529 | gcc_assert (def_stmt_info); | |
1530 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1531 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1532 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1533 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1534 | else | |
1535 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1536 | return vec_oprnd; | |
1537 | } | |
1538 | ||
1539 | ||
1540 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1541 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a | 1542 | |
c78e3652 | 1543 | void |
b8698a0f | 1544 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1545 | vec<tree> *vec_oprnds0, |
1546 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1547 | { |
9771b263 | 1548 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1549 | |
1550 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1551 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1552 | |
9771b263 | 1553 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1554 | { |
9771b263 | 1555 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1556 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1557 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1558 | } |
1559 | } | |
1560 | ||
1561 | ||
c78e3652 | 1562 | /* Get vectorized definitions for OP0 and OP1. */ |
ebfd146a | 1563 | |
c78e3652 | 1564 | void |
355fe088 | 1565 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1566 | vec<tree> *vec_oprnds0, |
1567 | vec<tree> *vec_oprnds1, | |
306b0c92 | 1568 | slp_tree slp_node) |
ebfd146a IR |
1569 | { |
1570 | if (slp_node) | |
d092494c IR |
1571 | { |
1572 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1573 | auto_vec<tree> ops (nops); |
1574 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1575 | |
9771b263 | 1576 | ops.quick_push (op0); |
d092494c | 1577 | if (op1) |
9771b263 | 1578 | ops.quick_push (op1); |
d092494c | 1579 | |
306b0c92 | 1580 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
d092494c | 1581 | |
37b5ec8f | 1582 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1583 | if (op1) |
37b5ec8f | 1584 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1585 | } |
ebfd146a IR |
1586 | else |
1587 | { | |
1588 | tree vec_oprnd; | |
1589 | ||
9771b263 | 1590 | vec_oprnds0->create (1); |
81c40241 | 1591 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1592 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1593 | |
1594 | if (op1) | |
1595 | { | |
9771b263 | 1596 | vec_oprnds1->create (1); |
81c40241 | 1597 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1598 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1599 | } |
1600 | } | |
1601 | } | |
1602 | ||
bb6c2b68 RS |
1603 | /* Helper function called by vect_finish_replace_stmt and |
1604 | vect_finish_stmt_generation. Set the location of the new | |
1605 | statement and create a stmt_vec_info for it. */ | |
1606 | ||
1607 | static void | |
1608 | vect_finish_stmt_generation_1 (gimple *stmt, gimple *vec_stmt) | |
1609 | { | |
1610 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1611 | vec_info *vinfo = stmt_info->vinfo; | |
1612 | ||
1613 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, vinfo)); | |
1614 | ||
1615 | if (dump_enabled_p ()) | |
1616 | { | |
1617 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); | |
1618 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
1619 | } | |
1620 | ||
1621 | gimple_set_location (vec_stmt, gimple_location (stmt)); | |
1622 | ||
1623 | /* While EH edges will generally prevent vectorization, stmt might | |
1624 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1625 | that could throw are part of the same region. */ | |
1626 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1627 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1628 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
1629 | } | |
1630 | ||
1631 | /* Replace the scalar statement STMT with a new vector statement VEC_STMT, | |
1632 | which sets the same scalar result as STMT did. */ | |
1633 | ||
1634 | void | |
1635 | vect_finish_replace_stmt (gimple *stmt, gimple *vec_stmt) | |
1636 | { | |
1637 | gcc_assert (gimple_get_lhs (stmt) == gimple_get_lhs (vec_stmt)); | |
1638 | ||
1639 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); | |
1640 | gsi_replace (&gsi, vec_stmt, false); | |
1641 | ||
1642 | vect_finish_stmt_generation_1 (stmt, vec_stmt); | |
1643 | } | |
ebfd146a IR |
1644 | |
1645 | /* Function vect_finish_stmt_generation. | |
1646 | ||
1647 | Insert a new stmt. */ | |
1648 | ||
1649 | void | |
355fe088 | 1650 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1651 | gimple_stmt_iterator *gsi) |
1652 | { | |
ebfd146a IR |
1653 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); |
1654 | ||
54e8e2c3 RG |
1655 | if (!gsi_end_p (*gsi) |
1656 | && gimple_has_mem_ops (vec_stmt)) | |
1657 | { | |
355fe088 | 1658 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1659 | tree vuse = gimple_vuse (at_stmt); |
1660 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1661 | { | |
1662 | tree vdef = gimple_vdef (at_stmt); | |
1663 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1664 | /* If we have an SSA vuse and insert a store, update virtual | |
1665 | SSA form to avoid triggering the renamer. Do so only | |
1666 | if we can easily see all uses - which is what almost always | |
1667 | happens with the way vectorized stmts are inserted. */ | |
1668 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1669 | && ((is_gimple_assign (vec_stmt) | |
1670 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1671 | || (is_gimple_call (vec_stmt) | |
1672 | && !(gimple_call_flags (vec_stmt) | |
1673 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1674 | { | |
1675 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1676 | gimple_set_vdef (vec_stmt, new_vdef); | |
1677 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1678 | } | |
1679 | } | |
1680 | } | |
ebfd146a | 1681 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
bb6c2b68 | 1682 | vect_finish_stmt_generation_1 (stmt, vec_stmt); |
ebfd146a IR |
1683 | } |
1684 | ||
70439f0d RS |
1685 | /* We want to vectorize a call to combined function CFN with function |
1686 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN | |
1687 | as the types of all inputs. Check whether this is possible using | |
1688 | an internal function, returning its code if so or IFN_LAST if not. */ | |
ebfd146a | 1689 | |
70439f0d RS |
1690 | static internal_fn |
1691 | vectorizable_internal_function (combined_fn cfn, tree fndecl, | |
1692 | tree vectype_out, tree vectype_in) | |
ebfd146a | 1693 | { |
70439f0d RS |
1694 | internal_fn ifn; |
1695 | if (internal_fn_p (cfn)) | |
1696 | ifn = as_internal_fn (cfn); | |
1697 | else | |
1698 | ifn = associated_internal_fn (fndecl); | |
1699 | if (ifn != IFN_LAST && direct_internal_fn_p (ifn)) | |
1700 | { | |
1701 | const direct_internal_fn_info &info = direct_internal_fn (ifn); | |
1702 | if (info.vectorizable) | |
1703 | { | |
1704 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); | |
1705 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); | |
d95ab70a RS |
1706 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1707 | OPTIMIZE_FOR_SPEED)) | |
70439f0d RS |
1708 | return ifn; |
1709 | } | |
1710 | } | |
1711 | return IFN_LAST; | |
ebfd146a IR |
1712 | } |
1713 | ||
5ce9450f | 1714 | |
355fe088 | 1715 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1716 | gimple_stmt_iterator *); |
1717 | ||
7cfb4d93 RS |
1718 | /* Check whether a load or store statement in the loop described by |
1719 | LOOP_VINFO is possible in a fully-masked loop. This is testing | |
1720 | whether the vectorizer pass has the appropriate support, as well as | |
1721 | whether the target does. | |
1722 | ||
1723 | VLS_TYPE says whether the statement is a load or store and VECTYPE | |
1724 | is the type of the vector being loaded or stored. MEMORY_ACCESS_TYPE | |
1725 | says how the load or store is going to be implemented and GROUP_SIZE | |
1726 | is the number of load or store statements in the containing group. | |
bfaa08b7 RS |
1727 | If the access is a gather load or scatter store, GS_INFO describes |
1728 | its arguments. | |
7cfb4d93 RS |
1729 | |
1730 | Clear LOOP_VINFO_CAN_FULLY_MASK_P if a fully-masked loop is not | |
1731 | supported, otherwise record the required mask types. */ | |
1732 | ||
1733 | static void | |
1734 | check_load_store_masking (loop_vec_info loop_vinfo, tree vectype, | |
1735 | vec_load_store_type vls_type, int group_size, | |
bfaa08b7 RS |
1736 | vect_memory_access_type memory_access_type, |
1737 | gather_scatter_info *gs_info) | |
7cfb4d93 RS |
1738 | { |
1739 | /* Invariant loads need no special support. */ | |
1740 | if (memory_access_type == VMAT_INVARIANT) | |
1741 | return; | |
1742 | ||
1743 | vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); | |
1744 | machine_mode vecmode = TYPE_MODE (vectype); | |
1745 | bool is_load = (vls_type == VLS_LOAD); | |
1746 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
1747 | { | |
1748 | if (is_load | |
1749 | ? !vect_load_lanes_supported (vectype, group_size, true) | |
1750 | : !vect_store_lanes_supported (vectype, group_size, true)) | |
1751 | { | |
1752 | if (dump_enabled_p ()) | |
1753 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1754 | "can't use a fully-masked loop because the" | |
1755 | " target doesn't have an appropriate masked" | |
1756 | " load/store-lanes instruction.\n"); | |
1757 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1758 | return; | |
1759 | } | |
1760 | unsigned int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
1761 | vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype); | |
1762 | return; | |
1763 | } | |
1764 | ||
bfaa08b7 RS |
1765 | if (memory_access_type == VMAT_GATHER_SCATTER) |
1766 | { | |
f307441a RS |
1767 | internal_fn ifn = (is_load |
1768 | ? IFN_MASK_GATHER_LOAD | |
1769 | : IFN_MASK_SCATTER_STORE); | |
bfaa08b7 | 1770 | tree offset_type = TREE_TYPE (gs_info->offset); |
f307441a | 1771 | if (!internal_gather_scatter_fn_supported_p (ifn, vectype, |
bfaa08b7 RS |
1772 | gs_info->memory_type, |
1773 | TYPE_SIGN (offset_type), | |
1774 | gs_info->scale)) | |
1775 | { | |
1776 | if (dump_enabled_p ()) | |
1777 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1778 | "can't use a fully-masked loop because the" | |
1779 | " target doesn't have an appropriate masked" | |
f307441a | 1780 | " gather load or scatter store instruction.\n"); |
bfaa08b7 RS |
1781 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; |
1782 | return; | |
1783 | } | |
1784 | unsigned int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
1785 | vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype); | |
1786 | return; | |
1787 | } | |
1788 | ||
7cfb4d93 RS |
1789 | if (memory_access_type != VMAT_CONTIGUOUS |
1790 | && memory_access_type != VMAT_CONTIGUOUS_PERMUTE) | |
1791 | { | |
1792 | /* Element X of the data must come from iteration i * VF + X of the | |
1793 | scalar loop. We need more work to support other mappings. */ | |
1794 | if (dump_enabled_p ()) | |
1795 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1796 | "can't use a fully-masked loop because an access" | |
1797 | " isn't contiguous.\n"); | |
1798 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1799 | return; | |
1800 | } | |
1801 | ||
1802 | machine_mode mask_mode; | |
1803 | if (!(targetm.vectorize.get_mask_mode | |
1804 | (GET_MODE_NUNITS (vecmode), | |
1805 | GET_MODE_SIZE (vecmode)).exists (&mask_mode)) | |
1806 | || !can_vec_mask_load_store_p (vecmode, mask_mode, is_load)) | |
1807 | { | |
1808 | if (dump_enabled_p ()) | |
1809 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1810 | "can't use a fully-masked loop because the target" | |
1811 | " doesn't have the appropriate masked load or" | |
1812 | " store.\n"); | |
1813 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1814 | return; | |
1815 | } | |
1816 | /* We might load more scalars than we need for permuting SLP loads. | |
1817 | We checked in get_group_load_store_type that the extra elements | |
1818 | don't leak into a new vector. */ | |
1819 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
1820 | poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
1821 | unsigned int nvectors; | |
1822 | if (can_div_away_from_zero_p (group_size * vf, nunits, &nvectors)) | |
1823 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype); | |
1824 | else | |
1825 | gcc_unreachable (); | |
1826 | } | |
1827 | ||
1828 | /* Return the mask input to a masked load or store. VEC_MASK is the vectorized | |
1829 | form of the scalar mask condition and LOOP_MASK, if nonnull, is the mask | |
1830 | that needs to be applied to all loads and stores in a vectorized loop. | |
1831 | Return VEC_MASK if LOOP_MASK is null, otherwise return VEC_MASK & LOOP_MASK. | |
1832 | ||
1833 | MASK_TYPE is the type of both masks. If new statements are needed, | |
1834 | insert them before GSI. */ | |
1835 | ||
1836 | static tree | |
1837 | prepare_load_store_mask (tree mask_type, tree loop_mask, tree vec_mask, | |
1838 | gimple_stmt_iterator *gsi) | |
1839 | { | |
1840 | gcc_assert (useless_type_conversion_p (mask_type, TREE_TYPE (vec_mask))); | |
1841 | if (!loop_mask) | |
1842 | return vec_mask; | |
1843 | ||
1844 | gcc_assert (TREE_TYPE (loop_mask) == mask_type); | |
1845 | tree and_res = make_temp_ssa_name (mask_type, NULL, "vec_mask_and"); | |
1846 | gimple *and_stmt = gimple_build_assign (and_res, BIT_AND_EXPR, | |
1847 | vec_mask, loop_mask); | |
1848 | gsi_insert_before (gsi, and_stmt, GSI_SAME_STMT); | |
1849 | return and_res; | |
1850 | } | |
1851 | ||
429ef523 RS |
1852 | /* Determine whether we can use a gather load or scatter store to vectorize |
1853 | strided load or store STMT by truncating the current offset to a smaller | |
1854 | width. We need to be able to construct an offset vector: | |
1855 | ||
1856 | { 0, X, X*2, X*3, ... } | |
1857 | ||
1858 | without loss of precision, where X is STMT's DR_STEP. | |
1859 | ||
1860 | Return true if this is possible, describing the gather load or scatter | |
1861 | store in GS_INFO. MASKED_P is true if the load or store is conditional. */ | |
1862 | ||
1863 | static bool | |
1864 | vect_truncate_gather_scatter_offset (gimple *stmt, loop_vec_info loop_vinfo, | |
1865 | bool masked_p, | |
1866 | gather_scatter_info *gs_info) | |
1867 | { | |
1868 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1869 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
1870 | tree step = DR_STEP (dr); | |
1871 | if (TREE_CODE (step) != INTEGER_CST) | |
1872 | { | |
1873 | /* ??? Perhaps we could use range information here? */ | |
1874 | if (dump_enabled_p ()) | |
1875 | dump_printf_loc (MSG_NOTE, vect_location, | |
1876 | "cannot truncate variable step.\n"); | |
1877 | return false; | |
1878 | } | |
1879 | ||
1880 | /* Get the number of bits in an element. */ | |
1881 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
1882 | scalar_mode element_mode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); | |
1883 | unsigned int element_bits = GET_MODE_BITSIZE (element_mode); | |
1884 | ||
1885 | /* Set COUNT to the upper limit on the number of elements - 1. | |
1886 | Start with the maximum vectorization factor. */ | |
1887 | unsigned HOST_WIDE_INT count = vect_max_vf (loop_vinfo) - 1; | |
1888 | ||
1889 | /* Try lowering COUNT to the number of scalar latch iterations. */ | |
1890 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1891 | widest_int max_iters; | |
1892 | if (max_loop_iterations (loop, &max_iters) | |
1893 | && max_iters < count) | |
1894 | count = max_iters.to_shwi (); | |
1895 | ||
1896 | /* Try scales of 1 and the element size. */ | |
1897 | int scales[] = { 1, vect_get_scalar_dr_size (dr) }; | |
1898 | bool overflow_p = false; | |
1899 | for (int i = 0; i < 2; ++i) | |
1900 | { | |
1901 | int scale = scales[i]; | |
1902 | widest_int factor; | |
1903 | if (!wi::multiple_of_p (wi::to_widest (step), scale, SIGNED, &factor)) | |
1904 | continue; | |
1905 | ||
1906 | /* See whether we can calculate (COUNT - 1) * STEP / SCALE | |
1907 | in OFFSET_BITS bits. */ | |
1908 | widest_int range = wi::mul (count, factor, SIGNED, &overflow_p); | |
1909 | if (overflow_p) | |
1910 | continue; | |
1911 | signop sign = range >= 0 ? UNSIGNED : SIGNED; | |
1912 | if (wi::min_precision (range, sign) > element_bits) | |
1913 | { | |
1914 | overflow_p = true; | |
1915 | continue; | |
1916 | } | |
1917 | ||
1918 | /* See whether the target supports the operation. */ | |
1919 | tree memory_type = TREE_TYPE (DR_REF (dr)); | |
1920 | if (!vect_gather_scatter_fn_p (DR_IS_READ (dr), masked_p, vectype, | |
1921 | memory_type, element_bits, sign, scale, | |
1922 | &gs_info->ifn, &gs_info->element_type)) | |
1923 | continue; | |
1924 | ||
1925 | tree offset_type = build_nonstandard_integer_type (element_bits, | |
1926 | sign == UNSIGNED); | |
1927 | ||
1928 | gs_info->decl = NULL_TREE; | |
1929 | /* Logically the sum of DR_BASE_ADDRESS, DR_INIT and DR_OFFSET, | |
1930 | but we don't need to store that here. */ | |
1931 | gs_info->base = NULL_TREE; | |
1932 | gs_info->offset = fold_convert (offset_type, step); | |
929b4411 | 1933 | gs_info->offset_dt = vect_constant_def; |
429ef523 RS |
1934 | gs_info->offset_vectype = NULL_TREE; |
1935 | gs_info->scale = scale; | |
1936 | gs_info->memory_type = memory_type; | |
1937 | return true; | |
1938 | } | |
1939 | ||
1940 | if (overflow_p && dump_enabled_p ()) | |
1941 | dump_printf_loc (MSG_NOTE, vect_location, | |
1942 | "truncating gather/scatter offset to %d bits" | |
1943 | " might change its value.\n", element_bits); | |
1944 | ||
1945 | return false; | |
1946 | } | |
1947 | ||
ab2fc782 RS |
1948 | /* Return true if we can use gather/scatter internal functions to |
1949 | vectorize STMT, which is a grouped or strided load or store. | |
429ef523 RS |
1950 | MASKED_P is true if load or store is conditional. When returning |
1951 | true, fill in GS_INFO with the information required to perform the | |
1952 | operation. */ | |
ab2fc782 RS |
1953 | |
1954 | static bool | |
1955 | vect_use_strided_gather_scatters_p (gimple *stmt, loop_vec_info loop_vinfo, | |
429ef523 | 1956 | bool masked_p, |
ab2fc782 RS |
1957 | gather_scatter_info *gs_info) |
1958 | { | |
1959 | if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info) | |
1960 | || gs_info->decl) | |
429ef523 RS |
1961 | return vect_truncate_gather_scatter_offset (stmt, loop_vinfo, |
1962 | masked_p, gs_info); | |
ab2fc782 RS |
1963 | |
1964 | scalar_mode element_mode = SCALAR_TYPE_MODE (gs_info->element_type); | |
1965 | unsigned int element_bits = GET_MODE_BITSIZE (element_mode); | |
1966 | tree offset_type = TREE_TYPE (gs_info->offset); | |
1967 | unsigned int offset_bits = TYPE_PRECISION (offset_type); | |
1968 | ||
1969 | /* Enforced by vect_check_gather_scatter. */ | |
1970 | gcc_assert (element_bits >= offset_bits); | |
1971 | ||
1972 | /* If the elements are wider than the offset, convert the offset to the | |
1973 | same width, without changing its sign. */ | |
1974 | if (element_bits > offset_bits) | |
1975 | { | |
1976 | bool unsigned_p = TYPE_UNSIGNED (offset_type); | |
1977 | offset_type = build_nonstandard_integer_type (element_bits, unsigned_p); | |
1978 | gs_info->offset = fold_convert (offset_type, gs_info->offset); | |
1979 | } | |
1980 | ||
1981 | if (dump_enabled_p ()) | |
1982 | dump_printf_loc (MSG_NOTE, vect_location, | |
1983 | "using gather/scatter for strided/grouped access," | |
1984 | " scale = %d\n", gs_info->scale); | |
1985 | ||
1986 | return true; | |
1987 | } | |
1988 | ||
62da9e14 RS |
1989 | /* STMT is a non-strided load or store, meaning that it accesses |
1990 | elements with a known constant step. Return -1 if that step | |
1991 | is negative, 0 if it is zero, and 1 if it is greater than zero. */ | |
1992 | ||
1993 | static int | |
1994 | compare_step_with_zero (gimple *stmt) | |
1995 | { | |
1996 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3f5e8a76 RS |
1997 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
1998 | return tree_int_cst_compare (vect_dr_behavior (dr)->step, | |
1999 | size_zero_node); | |
62da9e14 RS |
2000 | } |
2001 | ||
2002 | /* If the target supports a permute mask that reverses the elements in | |
2003 | a vector of type VECTYPE, return that mask, otherwise return null. */ | |
2004 | ||
2005 | static tree | |
2006 | perm_mask_for_reverse (tree vectype) | |
2007 | { | |
928686b1 | 2008 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
62da9e14 | 2009 | |
d980067b RS |
2010 | /* The encoding has a single stepped pattern. */ |
2011 | vec_perm_builder sel (nunits, 1, 3); | |
928686b1 | 2012 | for (int i = 0; i < 3; ++i) |
908a1a16 | 2013 | sel.quick_push (nunits - 1 - i); |
62da9e14 | 2014 | |
e3342de4 RS |
2015 | vec_perm_indices indices (sel, 1, nunits); |
2016 | if (!can_vec_perm_const_p (TYPE_MODE (vectype), indices)) | |
62da9e14 | 2017 | return NULL_TREE; |
e3342de4 | 2018 | return vect_gen_perm_mask_checked (vectype, indices); |
62da9e14 | 2019 | } |
5ce9450f | 2020 | |
c3a8f964 RS |
2021 | /* STMT is either a masked or unconditional store. Return the value |
2022 | being stored. */ | |
2023 | ||
f307441a | 2024 | tree |
c3a8f964 RS |
2025 | vect_get_store_rhs (gimple *stmt) |
2026 | { | |
2027 | if (gassign *assign = dyn_cast <gassign *> (stmt)) | |
2028 | { | |
2029 | gcc_assert (gimple_assign_single_p (assign)); | |
2030 | return gimple_assign_rhs1 (assign); | |
2031 | } | |
2032 | if (gcall *call = dyn_cast <gcall *> (stmt)) | |
2033 | { | |
2034 | internal_fn ifn = gimple_call_internal_fn (call); | |
f307441a RS |
2035 | int index = internal_fn_stored_value_index (ifn); |
2036 | gcc_assert (index >= 0); | |
2037 | return gimple_call_arg (stmt, index); | |
c3a8f964 RS |
2038 | } |
2039 | gcc_unreachable (); | |
2040 | } | |
2041 | ||
2de001ee RS |
2042 | /* A subroutine of get_load_store_type, with a subset of the same |
2043 | arguments. Handle the case where STMT is part of a grouped load | |
2044 | or store. | |
2045 | ||
2046 | For stores, the statements in the group are all consecutive | |
2047 | and there is no gap at the end. For loads, the statements in the | |
2048 | group might not be consecutive; there can be gaps between statements | |
2049 | as well as at the end. */ | |
2050 | ||
2051 | static bool | |
2052 | get_group_load_store_type (gimple *stmt, tree vectype, bool slp, | |
7e11fc7f | 2053 | bool masked_p, vec_load_store_type vls_type, |
429ef523 RS |
2054 | vect_memory_access_type *memory_access_type, |
2055 | gather_scatter_info *gs_info) | |
2de001ee RS |
2056 | { |
2057 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2058 | vec_info *vinfo = stmt_info->vinfo; | |
2059 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2060 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; | |
2061 | gimple *first_stmt = GROUP_FIRST_ELEMENT (stmt_info); | |
f702e7d4 | 2062 | data_reference *first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
2de001ee RS |
2063 | unsigned int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
2064 | bool single_element_p = (stmt == first_stmt | |
2065 | && !GROUP_NEXT_ELEMENT (stmt_info)); | |
2066 | unsigned HOST_WIDE_INT gap = GROUP_GAP (vinfo_for_stmt (first_stmt)); | |
928686b1 | 2067 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
2068 | |
2069 | /* True if the vectorized statements would access beyond the last | |
2070 | statement in the group. */ | |
2071 | bool overrun_p = false; | |
2072 | ||
2073 | /* True if we can cope with such overrun by peeling for gaps, so that | |
2074 | there is at least one final scalar iteration after the vector loop. */ | |
7e11fc7f RS |
2075 | bool can_overrun_p = (!masked_p |
2076 | && vls_type == VLS_LOAD | |
2077 | && loop_vinfo | |
2078 | && !loop->inner); | |
2de001ee RS |
2079 | |
2080 | /* There can only be a gap at the end of the group if the stride is | |
2081 | known at compile time. */ | |
2082 | gcc_assert (!STMT_VINFO_STRIDED_P (stmt_info) || gap == 0); | |
2083 | ||
2084 | /* Stores can't yet have gaps. */ | |
2085 | gcc_assert (slp || vls_type == VLS_LOAD || gap == 0); | |
2086 | ||
2087 | if (slp) | |
2088 | { | |
2089 | if (STMT_VINFO_STRIDED_P (stmt_info)) | |
2090 | { | |
2091 | /* Try to use consecutive accesses of GROUP_SIZE elements, | |
2092 | separated by the stride, until we have a complete vector. | |
2093 | Fall back to scalar accesses if that isn't possible. */ | |
928686b1 | 2094 | if (multiple_p (nunits, group_size)) |
2de001ee RS |
2095 | *memory_access_type = VMAT_STRIDED_SLP; |
2096 | else | |
2097 | *memory_access_type = VMAT_ELEMENTWISE; | |
2098 | } | |
2099 | else | |
2100 | { | |
2101 | overrun_p = loop_vinfo && gap != 0; | |
2102 | if (overrun_p && vls_type != VLS_LOAD) | |
2103 | { | |
2104 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2105 | "Grouped store with gaps requires" | |
2106 | " non-consecutive accesses\n"); | |
2107 | return false; | |
2108 | } | |
f702e7d4 RS |
2109 | /* An overrun is fine if the trailing elements are smaller |
2110 | than the alignment boundary B. Every vector access will | |
2111 | be a multiple of B and so we are guaranteed to access a | |
2112 | non-gap element in the same B-sized block. */ | |
f9ef2c76 | 2113 | if (overrun_p |
f702e7d4 RS |
2114 | && gap < (vect_known_alignment_in_bytes (first_dr) |
2115 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 2116 | overrun_p = false; |
2de001ee RS |
2117 | if (overrun_p && !can_overrun_p) |
2118 | { | |
2119 | if (dump_enabled_p ()) | |
2120 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2121 | "Peeling for outer loop is not supported\n"); | |
2122 | return false; | |
2123 | } | |
2124 | *memory_access_type = VMAT_CONTIGUOUS; | |
2125 | } | |
2126 | } | |
2127 | else | |
2128 | { | |
2129 | /* We can always handle this case using elementwise accesses, | |
2130 | but see if something more efficient is available. */ | |
2131 | *memory_access_type = VMAT_ELEMENTWISE; | |
2132 | ||
2133 | /* If there is a gap at the end of the group then these optimizations | |
2134 | would access excess elements in the last iteration. */ | |
2135 | bool would_overrun_p = (gap != 0); | |
f702e7d4 RS |
2136 | /* An overrun is fine if the trailing elements are smaller than the |
2137 | alignment boundary B. Every vector access will be a multiple of B | |
2138 | and so we are guaranteed to access a non-gap element in the | |
2139 | same B-sized block. */ | |
f9ef2c76 | 2140 | if (would_overrun_p |
7e11fc7f | 2141 | && !masked_p |
f702e7d4 RS |
2142 | && gap < (vect_known_alignment_in_bytes (first_dr) |
2143 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 2144 | would_overrun_p = false; |
f702e7d4 | 2145 | |
2de001ee | 2146 | if (!STMT_VINFO_STRIDED_P (stmt_info) |
62da9e14 RS |
2147 | && (can_overrun_p || !would_overrun_p) |
2148 | && compare_step_with_zero (stmt) > 0) | |
2de001ee | 2149 | { |
6737facb RS |
2150 | /* First cope with the degenerate case of a single-element |
2151 | vector. */ | |
2152 | if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U)) | |
2153 | *memory_access_type = VMAT_CONTIGUOUS; | |
2154 | ||
2155 | /* Otherwise try using LOAD/STORE_LANES. */ | |
2156 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2157 | && (vls_type == VLS_LOAD | |
7e11fc7f RS |
2158 | ? vect_load_lanes_supported (vectype, group_size, masked_p) |
2159 | : vect_store_lanes_supported (vectype, group_size, | |
2160 | masked_p))) | |
2de001ee RS |
2161 | { |
2162 | *memory_access_type = VMAT_LOAD_STORE_LANES; | |
2163 | overrun_p = would_overrun_p; | |
2164 | } | |
2165 | ||
2166 | /* If that fails, try using permuting loads. */ | |
2167 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2168 | && (vls_type == VLS_LOAD | |
2169 | ? vect_grouped_load_supported (vectype, single_element_p, | |
2170 | group_size) | |
2171 | : vect_grouped_store_supported (vectype, group_size))) | |
2172 | { | |
2173 | *memory_access_type = VMAT_CONTIGUOUS_PERMUTE; | |
2174 | overrun_p = would_overrun_p; | |
2175 | } | |
2176 | } | |
429ef523 RS |
2177 | |
2178 | /* As a last resort, trying using a gather load or scatter store. | |
2179 | ||
2180 | ??? Although the code can handle all group sizes correctly, | |
2181 | it probably isn't a win to use separate strided accesses based | |
2182 | on nearby locations. Or, even if it's a win over scalar code, | |
2183 | it might not be a win over vectorizing at a lower VF, if that | |
2184 | allows us to use contiguous accesses. */ | |
2185 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2186 | && single_element_p | |
2187 | && loop_vinfo | |
2188 | && vect_use_strided_gather_scatters_p (stmt, loop_vinfo, | |
2189 | masked_p, gs_info)) | |
2190 | *memory_access_type = VMAT_GATHER_SCATTER; | |
2de001ee RS |
2191 | } |
2192 | ||
2193 | if (vls_type != VLS_LOAD && first_stmt == stmt) | |
2194 | { | |
2195 | /* STMT is the leader of the group. Check the operands of all the | |
2196 | stmts of the group. */ | |
2197 | gimple *next_stmt = GROUP_NEXT_ELEMENT (stmt_info); | |
2198 | while (next_stmt) | |
2199 | { | |
7e11fc7f | 2200 | tree op = vect_get_store_rhs (next_stmt); |
2de001ee RS |
2201 | gimple *def_stmt; |
2202 | enum vect_def_type dt; | |
2203 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) | |
2204 | { | |
2205 | if (dump_enabled_p ()) | |
2206 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2207 | "use not simple.\n"); | |
2208 | return false; | |
2209 | } | |
2210 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); | |
2211 | } | |
2212 | } | |
2213 | ||
2214 | if (overrun_p) | |
2215 | { | |
2216 | gcc_assert (can_overrun_p); | |
2217 | if (dump_enabled_p ()) | |
2218 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2219 | "Data access with gaps requires scalar " | |
2220 | "epilogue loop\n"); | |
2221 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; | |
2222 | } | |
2223 | ||
2224 | return true; | |
2225 | } | |
2226 | ||
62da9e14 RS |
2227 | /* A subroutine of get_load_store_type, with a subset of the same |
2228 | arguments. Handle the case where STMT is a load or store that | |
2229 | accesses consecutive elements with a negative step. */ | |
2230 | ||
2231 | static vect_memory_access_type | |
2232 | get_negative_load_store_type (gimple *stmt, tree vectype, | |
2233 | vec_load_store_type vls_type, | |
2234 | unsigned int ncopies) | |
2235 | { | |
2236 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2237 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2238 | dr_alignment_support alignment_support_scheme; | |
2239 | ||
2240 | if (ncopies > 1) | |
2241 | { | |
2242 | if (dump_enabled_p ()) | |
2243 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2244 | "multiple types with negative step.\n"); | |
2245 | return VMAT_ELEMENTWISE; | |
2246 | } | |
2247 | ||
2248 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
2249 | if (alignment_support_scheme != dr_aligned | |
2250 | && alignment_support_scheme != dr_unaligned_supported) | |
2251 | { | |
2252 | if (dump_enabled_p ()) | |
2253 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2254 | "negative step but alignment required.\n"); | |
2255 | return VMAT_ELEMENTWISE; | |
2256 | } | |
2257 | ||
2258 | if (vls_type == VLS_STORE_INVARIANT) | |
2259 | { | |
2260 | if (dump_enabled_p ()) | |
2261 | dump_printf_loc (MSG_NOTE, vect_location, | |
2262 | "negative step with invariant source;" | |
2263 | " no permute needed.\n"); | |
2264 | return VMAT_CONTIGUOUS_DOWN; | |
2265 | } | |
2266 | ||
2267 | if (!perm_mask_for_reverse (vectype)) | |
2268 | { | |
2269 | if (dump_enabled_p ()) | |
2270 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2271 | "negative step and reversing not supported.\n"); | |
2272 | return VMAT_ELEMENTWISE; | |
2273 | } | |
2274 | ||
2275 | return VMAT_CONTIGUOUS_REVERSE; | |
2276 | } | |
2277 | ||
2de001ee RS |
2278 | /* Analyze load or store statement STMT of type VLS_TYPE. Return true |
2279 | if there is a memory access type that the vectorized form can use, | |
2280 | storing it in *MEMORY_ACCESS_TYPE if so. If we decide to use gathers | |
2281 | or scatters, fill in GS_INFO accordingly. | |
2282 | ||
2283 | SLP says whether we're performing SLP rather than loop vectorization. | |
7e11fc7f | 2284 | MASKED_P is true if the statement is conditional on a vectorized mask. |
62da9e14 RS |
2285 | VECTYPE is the vector type that the vectorized statements will use. |
2286 | NCOPIES is the number of vector statements that will be needed. */ | |
2de001ee RS |
2287 | |
2288 | static bool | |
7e11fc7f | 2289 | get_load_store_type (gimple *stmt, tree vectype, bool slp, bool masked_p, |
62da9e14 | 2290 | vec_load_store_type vls_type, unsigned int ncopies, |
2de001ee RS |
2291 | vect_memory_access_type *memory_access_type, |
2292 | gather_scatter_info *gs_info) | |
2293 | { | |
2294 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2295 | vec_info *vinfo = stmt_info->vinfo; | |
2296 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4d694b27 | 2297 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
2298 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
2299 | { | |
2300 | *memory_access_type = VMAT_GATHER_SCATTER; | |
2301 | gimple *def_stmt; | |
2302 | if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info)) | |
2303 | gcc_unreachable (); | |
2304 | else if (!vect_is_simple_use (gs_info->offset, vinfo, &def_stmt, | |
2305 | &gs_info->offset_dt, | |
2306 | &gs_info->offset_vectype)) | |
2307 | { | |
2308 | if (dump_enabled_p ()) | |
2309 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2310 | "%s index use not simple.\n", | |
2311 | vls_type == VLS_LOAD ? "gather" : "scatter"); | |
2312 | return false; | |
2313 | } | |
2314 | } | |
2315 | else if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
2316 | { | |
7e11fc7f | 2317 | if (!get_group_load_store_type (stmt, vectype, slp, masked_p, vls_type, |
429ef523 | 2318 | memory_access_type, gs_info)) |
2de001ee RS |
2319 | return false; |
2320 | } | |
2321 | else if (STMT_VINFO_STRIDED_P (stmt_info)) | |
2322 | { | |
2323 | gcc_assert (!slp); | |
ab2fc782 | 2324 | if (loop_vinfo |
429ef523 RS |
2325 | && vect_use_strided_gather_scatters_p (stmt, loop_vinfo, |
2326 | masked_p, gs_info)) | |
ab2fc782 RS |
2327 | *memory_access_type = VMAT_GATHER_SCATTER; |
2328 | else | |
2329 | *memory_access_type = VMAT_ELEMENTWISE; | |
2de001ee RS |
2330 | } |
2331 | else | |
62da9e14 RS |
2332 | { |
2333 | int cmp = compare_step_with_zero (stmt); | |
2334 | if (cmp < 0) | |
2335 | *memory_access_type = get_negative_load_store_type | |
2336 | (stmt, vectype, vls_type, ncopies); | |
2337 | else if (cmp == 0) | |
2338 | { | |
2339 | gcc_assert (vls_type == VLS_LOAD); | |
2340 | *memory_access_type = VMAT_INVARIANT; | |
2341 | } | |
2342 | else | |
2343 | *memory_access_type = VMAT_CONTIGUOUS; | |
2344 | } | |
2de001ee | 2345 | |
4d694b27 RS |
2346 | if ((*memory_access_type == VMAT_ELEMENTWISE |
2347 | || *memory_access_type == VMAT_STRIDED_SLP) | |
2348 | && !nunits.is_constant ()) | |
2349 | { | |
2350 | if (dump_enabled_p ()) | |
2351 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2352 | "Not using elementwise accesses due to variable " | |
2353 | "vectorization factor.\n"); | |
2354 | return false; | |
2355 | } | |
2356 | ||
2de001ee RS |
2357 | /* FIXME: At the moment the cost model seems to underestimate the |
2358 | cost of using elementwise accesses. This check preserves the | |
2359 | traditional behavior until that can be fixed. */ | |
2360 | if (*memory_access_type == VMAT_ELEMENTWISE | |
4aa157e8 RS |
2361 | && !STMT_VINFO_STRIDED_P (stmt_info) |
2362 | && !(stmt == GROUP_FIRST_ELEMENT (stmt_info) | |
2363 | && !GROUP_NEXT_ELEMENT (stmt_info) | |
2364 | && !pow2p_hwi (GROUP_SIZE (stmt_info)))) | |
2de001ee RS |
2365 | { |
2366 | if (dump_enabled_p ()) | |
2367 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2368 | "not falling back to elementwise accesses\n"); | |
2369 | return false; | |
2370 | } | |
2371 | return true; | |
2372 | } | |
2373 | ||
aaeefd88 | 2374 | /* Return true if boolean argument MASK is suitable for vectorizing |
929b4411 RS |
2375 | conditional load or store STMT. When returning true, store the type |
2376 | of the definition in *MASK_DT_OUT and the type of the vectorized mask | |
2377 | in *MASK_VECTYPE_OUT. */ | |
aaeefd88 RS |
2378 | |
2379 | static bool | |
929b4411 RS |
2380 | vect_check_load_store_mask (gimple *stmt, tree mask, |
2381 | vect_def_type *mask_dt_out, | |
2382 | tree *mask_vectype_out) | |
aaeefd88 RS |
2383 | { |
2384 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (mask))) | |
2385 | { | |
2386 | if (dump_enabled_p ()) | |
2387 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2388 | "mask argument is not a boolean.\n"); | |
2389 | return false; | |
2390 | } | |
2391 | ||
2392 | if (TREE_CODE (mask) != SSA_NAME) | |
2393 | { | |
2394 | if (dump_enabled_p ()) | |
2395 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2396 | "mask argument is not an SSA name.\n"); | |
2397 | return false; | |
2398 | } | |
2399 | ||
2400 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2401 | gimple *def_stmt; | |
929b4411 | 2402 | enum vect_def_type mask_dt; |
aaeefd88 | 2403 | tree mask_vectype; |
929b4411 | 2404 | if (!vect_is_simple_use (mask, stmt_info->vinfo, &def_stmt, &mask_dt, |
aaeefd88 RS |
2405 | &mask_vectype)) |
2406 | { | |
2407 | if (dump_enabled_p ()) | |
2408 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2409 | "mask use not simple.\n"); | |
2410 | return false; | |
2411 | } | |
2412 | ||
2413 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2414 | if (!mask_vectype) | |
2415 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
2416 | ||
2417 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype)) | |
2418 | { | |
2419 | if (dump_enabled_p ()) | |
2420 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2421 | "could not find an appropriate vector mask type.\n"); | |
2422 | return false; | |
2423 | } | |
2424 | ||
2425 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (mask_vectype), | |
2426 | TYPE_VECTOR_SUBPARTS (vectype))) | |
2427 | { | |
2428 | if (dump_enabled_p ()) | |
2429 | { | |
2430 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2431 | "vector mask type "); | |
2432 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, mask_vectype); | |
2433 | dump_printf (MSG_MISSED_OPTIMIZATION, | |
2434 | " does not match vector data type "); | |
2435 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, vectype); | |
2436 | dump_printf (MSG_MISSED_OPTIMIZATION, ".\n"); | |
2437 | } | |
2438 | return false; | |
2439 | } | |
2440 | ||
929b4411 | 2441 | *mask_dt_out = mask_dt; |
aaeefd88 RS |
2442 | *mask_vectype_out = mask_vectype; |
2443 | return true; | |
2444 | } | |
2445 | ||
3133c3b6 RS |
2446 | /* Return true if stored value RHS is suitable for vectorizing store |
2447 | statement STMT. When returning true, store the type of the | |
929b4411 RS |
2448 | definition in *RHS_DT_OUT, the type of the vectorized store value in |
2449 | *RHS_VECTYPE_OUT and the type of the store in *VLS_TYPE_OUT. */ | |
3133c3b6 RS |
2450 | |
2451 | static bool | |
929b4411 RS |
2452 | vect_check_store_rhs (gimple *stmt, tree rhs, vect_def_type *rhs_dt_out, |
2453 | tree *rhs_vectype_out, vec_load_store_type *vls_type_out) | |
3133c3b6 RS |
2454 | { |
2455 | /* In the case this is a store from a constant make sure | |
2456 | native_encode_expr can handle it. */ | |
2457 | if (CONSTANT_CLASS_P (rhs) && native_encode_expr (rhs, NULL, 64) == 0) | |
2458 | { | |
2459 | if (dump_enabled_p ()) | |
2460 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2461 | "cannot encode constant as a byte sequence.\n"); | |
2462 | return false; | |
2463 | } | |
2464 | ||
2465 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2466 | gimple *def_stmt; | |
929b4411 | 2467 | enum vect_def_type rhs_dt; |
3133c3b6 | 2468 | tree rhs_vectype; |
929b4411 | 2469 | if (!vect_is_simple_use (rhs, stmt_info->vinfo, &def_stmt, &rhs_dt, |
3133c3b6 RS |
2470 | &rhs_vectype)) |
2471 | { | |
2472 | if (dump_enabled_p ()) | |
2473 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2474 | "use not simple.\n"); | |
2475 | return false; | |
2476 | } | |
2477 | ||
2478 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2479 | if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype)) | |
2480 | { | |
2481 | if (dump_enabled_p ()) | |
2482 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2483 | "incompatible vector types.\n"); | |
2484 | return false; | |
2485 | } | |
2486 | ||
929b4411 | 2487 | *rhs_dt_out = rhs_dt; |
3133c3b6 | 2488 | *rhs_vectype_out = rhs_vectype; |
929b4411 | 2489 | if (rhs_dt == vect_constant_def || rhs_dt == vect_external_def) |
3133c3b6 RS |
2490 | *vls_type_out = VLS_STORE_INVARIANT; |
2491 | else | |
2492 | *vls_type_out = VLS_STORE; | |
2493 | return true; | |
2494 | } | |
2495 | ||
bc9587eb RS |
2496 | /* Build an all-ones vector mask of type MASKTYPE while vectorizing STMT. |
2497 | Note that we support masks with floating-point type, in which case the | |
2498 | floats are interpreted as a bitmask. */ | |
2499 | ||
2500 | static tree | |
2501 | vect_build_all_ones_mask (gimple *stmt, tree masktype) | |
2502 | { | |
2503 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
2504 | return build_int_cst (masktype, -1); | |
2505 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
2506 | { | |
2507 | tree mask = build_int_cst (TREE_TYPE (masktype), -1); | |
2508 | mask = build_vector_from_val (masktype, mask); | |
2509 | return vect_init_vector (stmt, mask, masktype, NULL); | |
2510 | } | |
2511 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) | |
2512 | { | |
2513 | REAL_VALUE_TYPE r; | |
2514 | long tmp[6]; | |
2515 | for (int j = 0; j < 6; ++j) | |
2516 | tmp[j] = -1; | |
2517 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
2518 | tree mask = build_real (TREE_TYPE (masktype), r); | |
2519 | mask = build_vector_from_val (masktype, mask); | |
2520 | return vect_init_vector (stmt, mask, masktype, NULL); | |
2521 | } | |
2522 | gcc_unreachable (); | |
2523 | } | |
2524 | ||
2525 | /* Build an all-zero merge value of type VECTYPE while vectorizing | |
2526 | STMT as a gather load. */ | |
2527 | ||
2528 | static tree | |
2529 | vect_build_zero_merge_argument (gimple *stmt, tree vectype) | |
2530 | { | |
2531 | tree merge; | |
2532 | if (TREE_CODE (TREE_TYPE (vectype)) == INTEGER_TYPE) | |
2533 | merge = build_int_cst (TREE_TYPE (vectype), 0); | |
2534 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (vectype))) | |
2535 | { | |
2536 | REAL_VALUE_TYPE r; | |
2537 | long tmp[6]; | |
2538 | for (int j = 0; j < 6; ++j) | |
2539 | tmp[j] = 0; | |
2540 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (vectype))); | |
2541 | merge = build_real (TREE_TYPE (vectype), r); | |
2542 | } | |
2543 | else | |
2544 | gcc_unreachable (); | |
2545 | merge = build_vector_from_val (vectype, merge); | |
2546 | return vect_init_vector (stmt, merge, vectype, NULL); | |
2547 | } | |
2548 | ||
c48d2d35 RS |
2549 | /* Build a gather load call while vectorizing STMT. Insert new instructions |
2550 | before GSI and add them to VEC_STMT. GS_INFO describes the gather load | |
2551 | operation. If the load is conditional, MASK is the unvectorized | |
929b4411 | 2552 | condition and MASK_DT is its definition type, otherwise MASK is null. */ |
c48d2d35 RS |
2553 | |
2554 | static void | |
2555 | vect_build_gather_load_calls (gimple *stmt, gimple_stmt_iterator *gsi, | |
2556 | gimple **vec_stmt, gather_scatter_info *gs_info, | |
929b4411 | 2557 | tree mask, vect_def_type mask_dt) |
c48d2d35 RS |
2558 | { |
2559 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2560 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2561 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2562 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2563 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2564 | int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
2565 | edge pe = loop_preheader_edge (loop); | |
2566 | enum { NARROW, NONE, WIDEN } modifier; | |
2567 | poly_uint64 gather_off_nunits | |
2568 | = TYPE_VECTOR_SUBPARTS (gs_info->offset_vectype); | |
2569 | ||
2570 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info->decl)); | |
2571 | tree rettype = TREE_TYPE (TREE_TYPE (gs_info->decl)); | |
2572 | tree srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2573 | tree ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2574 | tree idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2575 | tree masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2576 | tree scaletype = TREE_VALUE (arglist); | |
2577 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
2578 | && (!mask || types_compatible_p (srctype, masktype))); | |
2579 | ||
2580 | tree perm_mask = NULL_TREE; | |
2581 | tree mask_perm_mask = NULL_TREE; | |
2582 | if (known_eq (nunits, gather_off_nunits)) | |
2583 | modifier = NONE; | |
2584 | else if (known_eq (nunits * 2, gather_off_nunits)) | |
2585 | { | |
2586 | modifier = WIDEN; | |
2587 | ||
2588 | /* Currently widening gathers and scatters are only supported for | |
2589 | fixed-length vectors. */ | |
2590 | int count = gather_off_nunits.to_constant (); | |
2591 | vec_perm_builder sel (count, count, 1); | |
2592 | for (int i = 0; i < count; ++i) | |
2593 | sel.quick_push (i | (count / 2)); | |
2594 | ||
2595 | vec_perm_indices indices (sel, 1, count); | |
2596 | perm_mask = vect_gen_perm_mask_checked (gs_info->offset_vectype, | |
2597 | indices); | |
2598 | } | |
2599 | else if (known_eq (nunits, gather_off_nunits * 2)) | |
2600 | { | |
2601 | modifier = NARROW; | |
2602 | ||
2603 | /* Currently narrowing gathers and scatters are only supported for | |
2604 | fixed-length vectors. */ | |
2605 | int count = nunits.to_constant (); | |
2606 | vec_perm_builder sel (count, count, 1); | |
2607 | sel.quick_grow (count); | |
2608 | for (int i = 0; i < count; ++i) | |
2609 | sel[i] = i < count / 2 ? i : i + count / 2; | |
2610 | vec_perm_indices indices (sel, 2, count); | |
2611 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); | |
2612 | ||
2613 | ncopies *= 2; | |
2614 | ||
2615 | if (mask) | |
2616 | { | |
2617 | for (int i = 0; i < count; ++i) | |
2618 | sel[i] = i | (count / 2); | |
2619 | indices.new_vector (sel, 2, count); | |
2620 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, indices); | |
2621 | } | |
2622 | } | |
2623 | else | |
2624 | gcc_unreachable (); | |
2625 | ||
2626 | tree vec_dest = vect_create_destination_var (gimple_get_lhs (stmt), | |
2627 | vectype); | |
2628 | ||
2629 | tree ptr = fold_convert (ptrtype, gs_info->base); | |
2630 | if (!is_gimple_min_invariant (ptr)) | |
2631 | { | |
2632 | gimple_seq seq; | |
2633 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
2634 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
2635 | gcc_assert (!new_bb); | |
2636 | } | |
2637 | ||
2638 | tree scale = build_int_cst (scaletype, gs_info->scale); | |
2639 | ||
2640 | tree vec_oprnd0 = NULL_TREE; | |
2641 | tree vec_mask = NULL_TREE; | |
2642 | tree src_op = NULL_TREE; | |
2643 | tree mask_op = NULL_TREE; | |
2644 | tree prev_res = NULL_TREE; | |
2645 | stmt_vec_info prev_stmt_info = NULL; | |
2646 | ||
2647 | if (!mask) | |
2648 | { | |
2649 | src_op = vect_build_zero_merge_argument (stmt, rettype); | |
2650 | mask_op = vect_build_all_ones_mask (stmt, masktype); | |
2651 | } | |
2652 | ||
2653 | for (int j = 0; j < ncopies; ++j) | |
2654 | { | |
2655 | tree op, var; | |
2656 | gimple *new_stmt; | |
2657 | if (modifier == WIDEN && (j & 1)) | |
2658 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
2659 | perm_mask, stmt, gsi); | |
2660 | else if (j == 0) | |
2661 | op = vec_oprnd0 | |
2662 | = vect_get_vec_def_for_operand (gs_info->offset, stmt); | |
2663 | else | |
2664 | op = vec_oprnd0 | |
2665 | = vect_get_vec_def_for_stmt_copy (gs_info->offset_dt, vec_oprnd0); | |
2666 | ||
2667 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
2668 | { | |
2669 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), | |
2670 | TYPE_VECTOR_SUBPARTS (idxtype))); | |
2671 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); | |
2672 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); | |
2673 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
2674 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2675 | op = var; | |
2676 | } | |
2677 | ||
2678 | if (mask) | |
2679 | { | |
2680 | if (mask_perm_mask && (j & 1)) | |
2681 | mask_op = permute_vec_elements (mask_op, mask_op, | |
2682 | mask_perm_mask, stmt, gsi); | |
2683 | else | |
2684 | { | |
2685 | if (j == 0) | |
2686 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); | |
2687 | else | |
929b4411 | 2688 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
c48d2d35 RS |
2689 | |
2690 | mask_op = vec_mask; | |
2691 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
2692 | { | |
2693 | gcc_assert | |
2694 | (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)), | |
2695 | TYPE_VECTOR_SUBPARTS (masktype))); | |
2696 | var = vect_get_new_ssa_name (masktype, vect_simple_var); | |
2697 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); | |
2698 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, | |
2699 | mask_op); | |
2700 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2701 | mask_op = var; | |
2702 | } | |
2703 | } | |
2704 | src_op = mask_op; | |
2705 | } | |
2706 | ||
2707 | new_stmt = gimple_build_call (gs_info->decl, 5, src_op, ptr, op, | |
2708 | mask_op, scale); | |
2709 | ||
2710 | if (!useless_type_conversion_p (vectype, rettype)) | |
2711 | { | |
2712 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (vectype), | |
2713 | TYPE_VECTOR_SUBPARTS (rettype))); | |
2714 | op = vect_get_new_ssa_name (rettype, vect_simple_var); | |
2715 | gimple_call_set_lhs (new_stmt, op); | |
2716 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2717 | var = make_ssa_name (vec_dest); | |
2718 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); | |
2719 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
2720 | } | |
2721 | else | |
2722 | { | |
2723 | var = make_ssa_name (vec_dest, new_stmt); | |
2724 | gimple_call_set_lhs (new_stmt, var); | |
2725 | } | |
2726 | ||
2727 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2728 | ||
2729 | if (modifier == NARROW) | |
2730 | { | |
2731 | if ((j & 1) == 0) | |
2732 | { | |
2733 | prev_res = var; | |
2734 | continue; | |
2735 | } | |
2736 | var = permute_vec_elements (prev_res, var, perm_mask, stmt, gsi); | |
2737 | new_stmt = SSA_NAME_DEF_STMT (var); | |
2738 | } | |
2739 | ||
2740 | if (prev_stmt_info == NULL) | |
2741 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2742 | else | |
2743 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2744 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2745 | } | |
2746 | } | |
2747 | ||
bfaa08b7 RS |
2748 | /* Prepare the base and offset in GS_INFO for vectorization. |
2749 | Set *DATAREF_PTR to the loop-invariant base address and *VEC_OFFSET | |
2750 | to the vectorized offset argument for the first copy of STMT. STMT | |
2751 | is the statement described by GS_INFO and LOOP is the containing loop. */ | |
2752 | ||
2753 | static void | |
2754 | vect_get_gather_scatter_ops (struct loop *loop, gimple *stmt, | |
2755 | gather_scatter_info *gs_info, | |
2756 | tree *dataref_ptr, tree *vec_offset) | |
2757 | { | |
2758 | gimple_seq stmts = NULL; | |
2759 | *dataref_ptr = force_gimple_operand (gs_info->base, &stmts, true, NULL_TREE); | |
2760 | if (stmts != NULL) | |
2761 | { | |
2762 | basic_block new_bb; | |
2763 | edge pe = loop_preheader_edge (loop); | |
2764 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
2765 | gcc_assert (!new_bb); | |
2766 | } | |
2767 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2768 | tree offset_vectype = get_vectype_for_scalar_type (offset_type); | |
2769 | *vec_offset = vect_get_vec_def_for_operand (gs_info->offset, stmt, | |
2770 | offset_vectype); | |
2771 | } | |
2772 | ||
ab2fc782 RS |
2773 | /* Prepare to implement a grouped or strided load or store using |
2774 | the gather load or scatter store operation described by GS_INFO. | |
2775 | STMT is the load or store statement. | |
2776 | ||
2777 | Set *DATAREF_BUMP to the amount that should be added to the base | |
2778 | address after each copy of the vectorized statement. Set *VEC_OFFSET | |
2779 | to an invariant offset vector in which element I has the value | |
2780 | I * DR_STEP / SCALE. */ | |
2781 | ||
2782 | static void | |
2783 | vect_get_strided_load_store_ops (gimple *stmt, loop_vec_info loop_vinfo, | |
2784 | gather_scatter_info *gs_info, | |
2785 | tree *dataref_bump, tree *vec_offset) | |
2786 | { | |
2787 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2788 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2789 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2790 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2791 | gimple_seq stmts; | |
2792 | ||
2793 | tree bump = size_binop (MULT_EXPR, | |
2794 | fold_convert (sizetype, DR_STEP (dr)), | |
2795 | size_int (TYPE_VECTOR_SUBPARTS (vectype))); | |
2796 | *dataref_bump = force_gimple_operand (bump, &stmts, true, NULL_TREE); | |
2797 | if (stmts) | |
2798 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
2799 | ||
2800 | /* The offset given in GS_INFO can have pointer type, so use the element | |
2801 | type of the vector instead. */ | |
2802 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2803 | tree offset_vectype = get_vectype_for_scalar_type (offset_type); | |
2804 | offset_type = TREE_TYPE (offset_vectype); | |
2805 | ||
2806 | /* Calculate X = DR_STEP / SCALE and convert it to the appropriate type. */ | |
2807 | tree step = size_binop (EXACT_DIV_EXPR, DR_STEP (dr), | |
2808 | ssize_int (gs_info->scale)); | |
2809 | step = fold_convert (offset_type, step); | |
2810 | step = force_gimple_operand (step, &stmts, true, NULL_TREE); | |
2811 | ||
2812 | /* Create {0, X, X*2, X*3, ...}. */ | |
2813 | *vec_offset = gimple_build (&stmts, VEC_SERIES_EXPR, offset_vectype, | |
2814 | build_zero_cst (offset_type), step); | |
2815 | if (stmts) | |
2816 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
2817 | } | |
2818 | ||
2819 | /* Return the amount that should be added to a vector pointer to move | |
2820 | to the next or previous copy of AGGR_TYPE. DR is the data reference | |
2821 | being vectorized and MEMORY_ACCESS_TYPE describes the type of | |
2822 | vectorization. */ | |
2823 | ||
2824 | static tree | |
2825 | vect_get_data_ptr_increment (data_reference *dr, tree aggr_type, | |
2826 | vect_memory_access_type memory_access_type) | |
2827 | { | |
2828 | if (memory_access_type == VMAT_INVARIANT) | |
2829 | return size_zero_node; | |
2830 | ||
2831 | tree iv_step = TYPE_SIZE_UNIT (aggr_type); | |
2832 | tree step = vect_dr_behavior (dr)->step; | |
2833 | if (tree_int_cst_sgn (step) == -1) | |
2834 | iv_step = fold_build1 (NEGATE_EXPR, TREE_TYPE (iv_step), iv_step); | |
2835 | return iv_step; | |
2836 | } | |
2837 | ||
37b14185 RB |
2838 | /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64}. */ |
2839 | ||
2840 | static bool | |
2841 | vectorizable_bswap (gimple *stmt, gimple_stmt_iterator *gsi, | |
2842 | gimple **vec_stmt, slp_tree slp_node, | |
2843 | tree vectype_in, enum vect_def_type *dt) | |
2844 | { | |
2845 | tree op, vectype; | |
2846 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2847 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
928686b1 RS |
2848 | unsigned ncopies; |
2849 | unsigned HOST_WIDE_INT nunits, num_bytes; | |
37b14185 RB |
2850 | |
2851 | op = gimple_call_arg (stmt, 0); | |
2852 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
928686b1 RS |
2853 | |
2854 | if (!TYPE_VECTOR_SUBPARTS (vectype).is_constant (&nunits)) | |
2855 | return false; | |
37b14185 RB |
2856 | |
2857 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2858 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2859 | case of SLP. */ | |
2860 | if (slp_node) | |
2861 | ncopies = 1; | |
2862 | else | |
e8f142e2 | 2863 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
37b14185 RB |
2864 | |
2865 | gcc_assert (ncopies >= 1); | |
2866 | ||
2867 | tree char_vectype = get_same_sized_vectype (char_type_node, vectype_in); | |
2868 | if (! char_vectype) | |
2869 | return false; | |
2870 | ||
928686b1 RS |
2871 | if (!TYPE_VECTOR_SUBPARTS (char_vectype).is_constant (&num_bytes)) |
2872 | return false; | |
2873 | ||
794e3180 | 2874 | unsigned word_bytes = num_bytes / nunits; |
908a1a16 | 2875 | |
d980067b RS |
2876 | /* The encoding uses one stepped pattern for each byte in the word. */ |
2877 | vec_perm_builder elts (num_bytes, word_bytes, 3); | |
2878 | for (unsigned i = 0; i < 3; ++i) | |
37b14185 | 2879 | for (unsigned j = 0; j < word_bytes; ++j) |
908a1a16 | 2880 | elts.quick_push ((i + 1) * word_bytes - j - 1); |
37b14185 | 2881 | |
e3342de4 RS |
2882 | vec_perm_indices indices (elts, 1, num_bytes); |
2883 | if (!can_vec_perm_const_p (TYPE_MODE (char_vectype), indices)) | |
37b14185 RB |
2884 | return false; |
2885 | ||
2886 | if (! vec_stmt) | |
2887 | { | |
2888 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
2889 | if (dump_enabled_p ()) | |
2890 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_bswap ===" | |
2891 | "\n"); | |
78604de0 | 2892 | if (! slp_node) |
37b14185 RB |
2893 | { |
2894 | add_stmt_cost (stmt_info->vinfo->target_cost_data, | |
2895 | 1, vector_stmt, stmt_info, 0, vect_prologue); | |
2896 | add_stmt_cost (stmt_info->vinfo->target_cost_data, | |
2897 | ncopies, vec_perm, stmt_info, 0, vect_body); | |
2898 | } | |
2899 | return true; | |
2900 | } | |
2901 | ||
736d0f28 | 2902 | tree bswap_vconst = vec_perm_indices_to_tree (char_vectype, indices); |
37b14185 RB |
2903 | |
2904 | /* Transform. */ | |
2905 | vec<tree> vec_oprnds = vNULL; | |
2906 | gimple *new_stmt = NULL; | |
2907 | stmt_vec_info prev_stmt_info = NULL; | |
2908 | for (unsigned j = 0; j < ncopies; j++) | |
2909 | { | |
2910 | /* Handle uses. */ | |
2911 | if (j == 0) | |
306b0c92 | 2912 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
37b14185 RB |
2913 | else |
2914 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
2915 | ||
2916 | /* Arguments are ready. create the new vector stmt. */ | |
2917 | unsigned i; | |
2918 | tree vop; | |
2919 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) | |
2920 | { | |
2921 | tree tem = make_ssa_name (char_vectype); | |
2922 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
2923 | char_vectype, vop)); | |
2924 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2925 | tree tem2 = make_ssa_name (char_vectype); | |
2926 | new_stmt = gimple_build_assign (tem2, VEC_PERM_EXPR, | |
2927 | tem, tem, bswap_vconst); | |
2928 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2929 | tem = make_ssa_name (vectype); | |
2930 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
2931 | vectype, tem2)); | |
2932 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2933 | if (slp_node) | |
2934 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
2935 | } | |
2936 | ||
2937 | if (slp_node) | |
2938 | continue; | |
2939 | ||
2940 | if (j == 0) | |
2941 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2942 | else | |
2943 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2944 | ||
2945 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2946 | } | |
2947 | ||
2948 | vec_oprnds.release (); | |
2949 | return true; | |
2950 | } | |
2951 | ||
b1b6836e RS |
2952 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
2953 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT | |
2954 | in a single step. On success, store the binary pack code in | |
2955 | *CONVERT_CODE. */ | |
2956 | ||
2957 | static bool | |
2958 | simple_integer_narrowing (tree vectype_out, tree vectype_in, | |
2959 | tree_code *convert_code) | |
2960 | { | |
2961 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) | |
2962 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) | |
2963 | return false; | |
2964 | ||
2965 | tree_code code; | |
2966 | int multi_step_cvt = 0; | |
2967 | auto_vec <tree, 8> interm_types; | |
2968 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, | |
2969 | &code, &multi_step_cvt, | |
2970 | &interm_types) | |
2971 | || multi_step_cvt) | |
2972 | return false; | |
2973 | ||
2974 | *convert_code = code; | |
2975 | return true; | |
2976 | } | |
5ce9450f | 2977 | |
ebfd146a IR |
2978 | /* Function vectorizable_call. |
2979 | ||
538dd0b7 | 2980 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 2981 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2982 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2983 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2984 | ||
2985 | static bool | |
355fe088 | 2986 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
190c2236 | 2987 | slp_tree slp_node) |
ebfd146a | 2988 | { |
538dd0b7 | 2989 | gcall *stmt; |
ebfd146a IR |
2990 | tree vec_dest; |
2991 | tree scalar_dest; | |
2992 | tree op, type; | |
2993 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
538dd0b7 | 2994 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a | 2995 | tree vectype_out, vectype_in; |
c7bda0f4 RS |
2996 | poly_uint64 nunits_in; |
2997 | poly_uint64 nunits_out; | |
ebfd146a | 2998 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
190c2236 | 2999 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 3000 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 3001 | tree fndecl, new_temp, rhs_type; |
355fe088 | 3002 | gimple *def_stmt; |
0502fb85 UB |
3003 | enum vect_def_type dt[3] |
3004 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 3005 | int ndts = 3; |
355fe088 | 3006 | gimple *new_stmt = NULL; |
ebfd146a | 3007 | int ncopies, j; |
6e1aa848 | 3008 | vec<tree> vargs = vNULL; |
ebfd146a IR |
3009 | enum { NARROW, NONE, WIDEN } modifier; |
3010 | size_t i, nargs; | |
9d5e7640 | 3011 | tree lhs; |
ebfd146a | 3012 | |
190c2236 | 3013 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3014 | return false; |
3015 | ||
66c16fd9 RB |
3016 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3017 | && ! vec_stmt) | |
ebfd146a IR |
3018 | return false; |
3019 | ||
538dd0b7 DM |
3020 | /* Is GS a vectorizable call? */ |
3021 | stmt = dyn_cast <gcall *> (gs); | |
3022 | if (!stmt) | |
ebfd146a IR |
3023 | return false; |
3024 | ||
5ce9450f | 3025 | if (gimple_call_internal_p (stmt) |
bfaa08b7 | 3026 | && (internal_load_fn_p (gimple_call_internal_fn (stmt)) |
f307441a | 3027 | || internal_store_fn_p (gimple_call_internal_fn (stmt)))) |
c3a8f964 RS |
3028 | /* Handled by vectorizable_load and vectorizable_store. */ |
3029 | return false; | |
5ce9450f | 3030 | |
0136f8f0 AH |
3031 | if (gimple_call_lhs (stmt) == NULL_TREE |
3032 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
3033 | return false; |
3034 | ||
0136f8f0 | 3035 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 3036 | |
b690cc0f RG |
3037 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3038 | ||
ebfd146a IR |
3039 | /* Process function arguments. */ |
3040 | rhs_type = NULL_TREE; | |
b690cc0f | 3041 | vectype_in = NULL_TREE; |
ebfd146a IR |
3042 | nargs = gimple_call_num_args (stmt); |
3043 | ||
1b1562a5 MM |
3044 | /* Bail out if the function has more than three arguments, we do not have |
3045 | interesting builtin functions to vectorize with more than two arguments | |
3046 | except for fma. No arguments is also not good. */ | |
3047 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
3048 | return false; |
3049 | ||
74bf76ed JJ |
3050 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
3051 | if (gimple_call_internal_p (stmt) | |
3052 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
3053 | { | |
3054 | nargs = 0; | |
3055 | rhs_type = unsigned_type_node; | |
3056 | } | |
3057 | ||
ebfd146a IR |
3058 | for (i = 0; i < nargs; i++) |
3059 | { | |
b690cc0f RG |
3060 | tree opvectype; |
3061 | ||
ebfd146a IR |
3062 | op = gimple_call_arg (stmt, i); |
3063 | ||
3064 | /* We can only handle calls with arguments of the same type. */ | |
3065 | if (rhs_type | |
8533c9d8 | 3066 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 3067 | { |
73fbfcad | 3068 | if (dump_enabled_p ()) |
78c60e3d | 3069 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3070 | "argument types differ.\n"); |
ebfd146a IR |
3071 | return false; |
3072 | } | |
b690cc0f RG |
3073 | if (!rhs_type) |
3074 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 3075 | |
81c40241 | 3076 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[i], &opvectype)) |
ebfd146a | 3077 | { |
73fbfcad | 3078 | if (dump_enabled_p ()) |
78c60e3d | 3079 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3080 | "use not simple.\n"); |
ebfd146a IR |
3081 | return false; |
3082 | } | |
ebfd146a | 3083 | |
b690cc0f RG |
3084 | if (!vectype_in) |
3085 | vectype_in = opvectype; | |
3086 | else if (opvectype | |
3087 | && opvectype != vectype_in) | |
3088 | { | |
73fbfcad | 3089 | if (dump_enabled_p ()) |
78c60e3d | 3090 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3091 | "argument vector types differ.\n"); |
b690cc0f RG |
3092 | return false; |
3093 | } | |
3094 | } | |
3095 | /* If all arguments are external or constant defs use a vector type with | |
3096 | the same size as the output vector type. */ | |
ebfd146a | 3097 | if (!vectype_in) |
b690cc0f | 3098 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
3099 | if (vec_stmt) |
3100 | gcc_assert (vectype_in); | |
3101 | if (!vectype_in) | |
3102 | { | |
73fbfcad | 3103 | if (dump_enabled_p ()) |
7d8930a0 | 3104 | { |
78c60e3d SS |
3105 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3106 | "no vectype for scalar type "); | |
3107 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 3108 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
3109 | } |
3110 | ||
3111 | return false; | |
3112 | } | |
ebfd146a IR |
3113 | |
3114 | /* FORNOW */ | |
b690cc0f RG |
3115 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
3116 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
c7bda0f4 | 3117 | if (known_eq (nunits_in * 2, nunits_out)) |
ebfd146a | 3118 | modifier = NARROW; |
c7bda0f4 | 3119 | else if (known_eq (nunits_out, nunits_in)) |
ebfd146a | 3120 | modifier = NONE; |
c7bda0f4 | 3121 | else if (known_eq (nunits_out * 2, nunits_in)) |
ebfd146a IR |
3122 | modifier = WIDEN; |
3123 | else | |
3124 | return false; | |
3125 | ||
70439f0d RS |
3126 | /* We only handle functions that do not read or clobber memory. */ |
3127 | if (gimple_vuse (stmt)) | |
3128 | { | |
3129 | if (dump_enabled_p ()) | |
3130 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3131 | "function reads from or writes to memory.\n"); | |
3132 | return false; | |
3133 | } | |
3134 | ||
ebfd146a IR |
3135 | /* For now, we only vectorize functions if a target specific builtin |
3136 | is available. TODO -- in some cases, it might be profitable to | |
3137 | insert the calls for pieces of the vector, in order to be able | |
3138 | to vectorize other operations in the loop. */ | |
70439f0d RS |
3139 | fndecl = NULL_TREE; |
3140 | internal_fn ifn = IFN_LAST; | |
3141 | combined_fn cfn = gimple_call_combined_fn (stmt); | |
3142 | tree callee = gimple_call_fndecl (stmt); | |
3143 | ||
3144 | /* First try using an internal function. */ | |
b1b6836e RS |
3145 | tree_code convert_code = ERROR_MARK; |
3146 | if (cfn != CFN_LAST | |
3147 | && (modifier == NONE | |
3148 | || (modifier == NARROW | |
3149 | && simple_integer_narrowing (vectype_out, vectype_in, | |
3150 | &convert_code)))) | |
70439f0d RS |
3151 | ifn = vectorizable_internal_function (cfn, callee, vectype_out, |
3152 | vectype_in); | |
3153 | ||
3154 | /* If that fails, try asking for a target-specific built-in function. */ | |
3155 | if (ifn == IFN_LAST) | |
3156 | { | |
3157 | if (cfn != CFN_LAST) | |
3158 | fndecl = targetm.vectorize.builtin_vectorized_function | |
3159 | (cfn, vectype_out, vectype_in); | |
7672aa9b | 3160 | else if (callee) |
70439f0d RS |
3161 | fndecl = targetm.vectorize.builtin_md_vectorized_function |
3162 | (callee, vectype_out, vectype_in); | |
3163 | } | |
3164 | ||
3165 | if (ifn == IFN_LAST && !fndecl) | |
ebfd146a | 3166 | { |
70439f0d | 3167 | if (cfn == CFN_GOMP_SIMD_LANE |
74bf76ed JJ |
3168 | && !slp_node |
3169 | && loop_vinfo | |
3170 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
3171 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
3172 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
3173 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
3174 | { | |
3175 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
3176 | { 0, 1, 2, ... vf - 1 } vector. */ | |
3177 | gcc_assert (nargs == 0); | |
3178 | } | |
37b14185 RB |
3179 | else if (modifier == NONE |
3180 | && (gimple_call_builtin_p (stmt, BUILT_IN_BSWAP16) | |
3181 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP32) | |
3182 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP64))) | |
3183 | return vectorizable_bswap (stmt, gsi, vec_stmt, slp_node, | |
3184 | vectype_in, dt); | |
74bf76ed JJ |
3185 | else |
3186 | { | |
3187 | if (dump_enabled_p ()) | |
3188 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 3189 | "function is not vectorizable.\n"); |
74bf76ed JJ |
3190 | return false; |
3191 | } | |
ebfd146a IR |
3192 | } |
3193 | ||
fce57248 | 3194 | if (slp_node) |
190c2236 | 3195 | ncopies = 1; |
b1b6836e | 3196 | else if (modifier == NARROW && ifn == IFN_LAST) |
e8f142e2 | 3197 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
ebfd146a | 3198 | else |
e8f142e2 | 3199 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
ebfd146a IR |
3200 | |
3201 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3202 | needs to be generated. */ | |
3203 | gcc_assert (ncopies >= 1); | |
3204 | ||
3205 | if (!vec_stmt) /* transformation not required. */ | |
3206 | { | |
3207 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 3208 | if (dump_enabled_p ()) |
e645e942 TJ |
3209 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
3210 | "\n"); | |
78604de0 RB |
3211 | if (!slp_node) |
3212 | { | |
3213 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); | |
3214 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) | |
3215 | add_stmt_cost (stmt_info->vinfo->target_cost_data, ncopies / 2, | |
3216 | vec_promote_demote, stmt_info, 0, vect_body); | |
3217 | } | |
b1b6836e | 3218 | |
ebfd146a IR |
3219 | return true; |
3220 | } | |
3221 | ||
67b8dbac | 3222 | /* Transform. */ |
ebfd146a | 3223 | |
73fbfcad | 3224 | if (dump_enabled_p ()) |
e645e942 | 3225 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
3226 | |
3227 | /* Handle def. */ | |
3228 | scalar_dest = gimple_call_lhs (stmt); | |
3229 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
3230 | ||
3231 | prev_stmt_info = NULL; | |
b1b6836e | 3232 | if (modifier == NONE || ifn != IFN_LAST) |
ebfd146a | 3233 | { |
b1b6836e | 3234 | tree prev_res = NULL_TREE; |
ebfd146a IR |
3235 | for (j = 0; j < ncopies; ++j) |
3236 | { | |
3237 | /* Build argument list for the vectorized call. */ | |
3238 | if (j == 0) | |
9771b263 | 3239 | vargs.create (nargs); |
ebfd146a | 3240 | else |
9771b263 | 3241 | vargs.truncate (0); |
ebfd146a | 3242 | |
190c2236 JJ |
3243 | if (slp_node) |
3244 | { | |
ef062b13 | 3245 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 3246 | vec<tree> vec_oprnds0; |
190c2236 JJ |
3247 | |
3248 | for (i = 0; i < nargs; i++) | |
9771b263 | 3249 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 3250 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 3251 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
3252 | |
3253 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3254 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
3255 | { |
3256 | size_t k; | |
3257 | for (k = 0; k < nargs; k++) | |
3258 | { | |
37b5ec8f | 3259 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 3260 | vargs[k] = vec_oprndsk[i]; |
190c2236 | 3261 | } |
b1b6836e RS |
3262 | if (modifier == NARROW) |
3263 | { | |
3264 | tree half_res = make_ssa_name (vectype_in); | |
a844293d RS |
3265 | gcall *call |
3266 | = gimple_build_call_internal_vec (ifn, vargs); | |
3267 | gimple_call_set_lhs (call, half_res); | |
3268 | gimple_call_set_nothrow (call, true); | |
3269 | new_stmt = call; | |
b1b6836e RS |
3270 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3271 | if ((i & 1) == 0) | |
3272 | { | |
3273 | prev_res = half_res; | |
3274 | continue; | |
3275 | } | |
3276 | new_temp = make_ssa_name (vec_dest); | |
3277 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
3278 | prev_res, half_res); | |
3279 | } | |
70439f0d | 3280 | else |
b1b6836e | 3281 | { |
a844293d | 3282 | gcall *call; |
b1b6836e | 3283 | if (ifn != IFN_LAST) |
a844293d | 3284 | call = gimple_build_call_internal_vec (ifn, vargs); |
b1b6836e | 3285 | else |
a844293d RS |
3286 | call = gimple_build_call_vec (fndecl, vargs); |
3287 | new_temp = make_ssa_name (vec_dest, call); | |
3288 | gimple_call_set_lhs (call, new_temp); | |
3289 | gimple_call_set_nothrow (call, true); | |
3290 | new_stmt = call; | |
b1b6836e | 3291 | } |
190c2236 | 3292 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 3293 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
3294 | } |
3295 | ||
3296 | for (i = 0; i < nargs; i++) | |
3297 | { | |
37b5ec8f | 3298 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3299 | vec_oprndsi.release (); |
190c2236 | 3300 | } |
190c2236 JJ |
3301 | continue; |
3302 | } | |
3303 | ||
ebfd146a IR |
3304 | for (i = 0; i < nargs; i++) |
3305 | { | |
3306 | op = gimple_call_arg (stmt, i); | |
3307 | if (j == 0) | |
3308 | vec_oprnd0 | |
81c40241 | 3309 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 3310 | else |
63827fb8 IR |
3311 | { |
3312 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
3313 | vec_oprnd0 | |
3314 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
3315 | } | |
ebfd146a | 3316 | |
9771b263 | 3317 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
3318 | } |
3319 | ||
74bf76ed JJ |
3320 | if (gimple_call_internal_p (stmt) |
3321 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
3322 | { | |
c7bda0f4 | 3323 | tree cst = build_index_vector (vectype_out, j * nunits_out, 1); |
74bf76ed | 3324 | tree new_var |
0e22bb5a | 3325 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 3326 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 3327 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 3328 | new_temp = make_ssa_name (vec_dest); |
0e22bb5a | 3329 | new_stmt = gimple_build_assign (new_temp, new_var); |
74bf76ed | 3330 | } |
b1b6836e RS |
3331 | else if (modifier == NARROW) |
3332 | { | |
3333 | tree half_res = make_ssa_name (vectype_in); | |
a844293d RS |
3334 | gcall *call = gimple_build_call_internal_vec (ifn, vargs); |
3335 | gimple_call_set_lhs (call, half_res); | |
3336 | gimple_call_set_nothrow (call, true); | |
3337 | new_stmt = call; | |
b1b6836e RS |
3338 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3339 | if ((j & 1) == 0) | |
3340 | { | |
3341 | prev_res = half_res; | |
3342 | continue; | |
3343 | } | |
3344 | new_temp = make_ssa_name (vec_dest); | |
3345 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
3346 | prev_res, half_res); | |
3347 | } | |
74bf76ed JJ |
3348 | else |
3349 | { | |
a844293d | 3350 | gcall *call; |
70439f0d | 3351 | if (ifn != IFN_LAST) |
a844293d | 3352 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3353 | else |
a844293d | 3354 | call = gimple_build_call_vec (fndecl, vargs); |
74bf76ed | 3355 | new_temp = make_ssa_name (vec_dest, new_stmt); |
a844293d RS |
3356 | gimple_call_set_lhs (call, new_temp); |
3357 | gimple_call_set_nothrow (call, true); | |
3358 | new_stmt = call; | |
74bf76ed | 3359 | } |
ebfd146a IR |
3360 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3361 | ||
b1b6836e | 3362 | if (j == (modifier == NARROW ? 1 : 0)) |
ebfd146a IR |
3363 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; |
3364 | else | |
3365 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3366 | ||
3367 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3368 | } | |
b1b6836e RS |
3369 | } |
3370 | else if (modifier == NARROW) | |
3371 | { | |
ebfd146a IR |
3372 | for (j = 0; j < ncopies; ++j) |
3373 | { | |
3374 | /* Build argument list for the vectorized call. */ | |
3375 | if (j == 0) | |
9771b263 | 3376 | vargs.create (nargs * 2); |
ebfd146a | 3377 | else |
9771b263 | 3378 | vargs.truncate (0); |
ebfd146a | 3379 | |
190c2236 JJ |
3380 | if (slp_node) |
3381 | { | |
ef062b13 | 3382 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 3383 | vec<tree> vec_oprnds0; |
190c2236 JJ |
3384 | |
3385 | for (i = 0; i < nargs; i++) | |
9771b263 | 3386 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 3387 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 3388 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
3389 | |
3390 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3391 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
3392 | { |
3393 | size_t k; | |
9771b263 | 3394 | vargs.truncate (0); |
190c2236 JJ |
3395 | for (k = 0; k < nargs; k++) |
3396 | { | |
37b5ec8f | 3397 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
3398 | vargs.quick_push (vec_oprndsk[i]); |
3399 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 | 3400 | } |
a844293d | 3401 | gcall *call; |
70439f0d | 3402 | if (ifn != IFN_LAST) |
a844293d | 3403 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3404 | else |
a844293d RS |
3405 | call = gimple_build_call_vec (fndecl, vargs); |
3406 | new_temp = make_ssa_name (vec_dest, call); | |
3407 | gimple_call_set_lhs (call, new_temp); | |
3408 | gimple_call_set_nothrow (call, true); | |
3409 | new_stmt = call; | |
190c2236 | 3410 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 3411 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
3412 | } |
3413 | ||
3414 | for (i = 0; i < nargs; i++) | |
3415 | { | |
37b5ec8f | 3416 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3417 | vec_oprndsi.release (); |
190c2236 | 3418 | } |
190c2236 JJ |
3419 | continue; |
3420 | } | |
3421 | ||
ebfd146a IR |
3422 | for (i = 0; i < nargs; i++) |
3423 | { | |
3424 | op = gimple_call_arg (stmt, i); | |
3425 | if (j == 0) | |
3426 | { | |
3427 | vec_oprnd0 | |
81c40241 | 3428 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 3429 | vec_oprnd1 |
63827fb8 | 3430 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3431 | } |
3432 | else | |
3433 | { | |
336ecb65 | 3434 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 3435 | vec_oprnd0 |
63827fb8 | 3436 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 3437 | vec_oprnd1 |
63827fb8 | 3438 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3439 | } |
3440 | ||
9771b263 DN |
3441 | vargs.quick_push (vec_oprnd0); |
3442 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
3443 | } |
3444 | ||
b1b6836e | 3445 | new_stmt = gimple_build_call_vec (fndecl, vargs); |
ebfd146a IR |
3446 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3447 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
3448 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3449 | ||
3450 | if (j == 0) | |
3451 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3452 | else | |
3453 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3454 | ||
3455 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3456 | } | |
3457 | ||
3458 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a | 3459 | } |
b1b6836e RS |
3460 | else |
3461 | /* No current target implements this case. */ | |
3462 | return false; | |
ebfd146a | 3463 | |
9771b263 | 3464 | vargs.release (); |
ebfd146a | 3465 | |
ebfd146a IR |
3466 | /* The call in STMT might prevent it from being removed in dce. |
3467 | We however cannot remove it here, due to the way the ssa name | |
3468 | it defines is mapped to the new definition. So just replace | |
3469 | rhs of the statement with something harmless. */ | |
3470 | ||
dd34c087 JJ |
3471 | if (slp_node) |
3472 | return true; | |
3473 | ||
ebfd146a | 3474 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
3475 | if (is_pattern_stmt_p (stmt_info)) |
3476 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3477 | else | |
3478 | lhs = gimple_call_lhs (stmt); | |
3cc2fa2a | 3479 | |
9d5e7640 | 3480 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
ebfd146a | 3481 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 3482 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
3483 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
3484 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
3485 | |
3486 | return true; | |
3487 | } | |
3488 | ||
3489 | ||
0136f8f0 AH |
3490 | struct simd_call_arg_info |
3491 | { | |
3492 | tree vectype; | |
3493 | tree op; | |
0136f8f0 | 3494 | HOST_WIDE_INT linear_step; |
34e82342 | 3495 | enum vect_def_type dt; |
0136f8f0 | 3496 | unsigned int align; |
17b658af | 3497 | bool simd_lane_linear; |
0136f8f0 AH |
3498 | }; |
3499 | ||
17b658af JJ |
3500 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
3501 | is linear within simd lane (but not within whole loop), note it in | |
3502 | *ARGINFO. */ | |
3503 | ||
3504 | static void | |
3505 | vect_simd_lane_linear (tree op, struct loop *loop, | |
3506 | struct simd_call_arg_info *arginfo) | |
3507 | { | |
355fe088 | 3508 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
3509 | |
3510 | if (!is_gimple_assign (def_stmt) | |
3511 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
3512 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
3513 | return; | |
3514 | ||
3515 | tree base = gimple_assign_rhs1 (def_stmt); | |
3516 | HOST_WIDE_INT linear_step = 0; | |
3517 | tree v = gimple_assign_rhs2 (def_stmt); | |
3518 | while (TREE_CODE (v) == SSA_NAME) | |
3519 | { | |
3520 | tree t; | |
3521 | def_stmt = SSA_NAME_DEF_STMT (v); | |
3522 | if (is_gimple_assign (def_stmt)) | |
3523 | switch (gimple_assign_rhs_code (def_stmt)) | |
3524 | { | |
3525 | case PLUS_EXPR: | |
3526 | t = gimple_assign_rhs2 (def_stmt); | |
3527 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
3528 | return; | |
3529 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
3530 | v = gimple_assign_rhs1 (def_stmt); | |
3531 | continue; | |
3532 | case MULT_EXPR: | |
3533 | t = gimple_assign_rhs2 (def_stmt); | |
3534 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
3535 | return; | |
3536 | linear_step = tree_to_shwi (t); | |
3537 | v = gimple_assign_rhs1 (def_stmt); | |
3538 | continue; | |
3539 | CASE_CONVERT: | |
3540 | t = gimple_assign_rhs1 (def_stmt); | |
3541 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
3542 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
3543 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
3544 | return; | |
3545 | if (!linear_step) | |
3546 | linear_step = 1; | |
3547 | v = t; | |
3548 | continue; | |
3549 | default: | |
3550 | return; | |
3551 | } | |
8e4284d0 | 3552 | else if (gimple_call_internal_p (def_stmt, IFN_GOMP_SIMD_LANE) |
17b658af JJ |
3553 | && loop->simduid |
3554 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
3555 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
3556 | == loop->simduid)) | |
3557 | { | |
3558 | if (!linear_step) | |
3559 | linear_step = 1; | |
3560 | arginfo->linear_step = linear_step; | |
3561 | arginfo->op = base; | |
3562 | arginfo->simd_lane_linear = true; | |
3563 | return; | |
3564 | } | |
3565 | } | |
3566 | } | |
3567 | ||
cf1b2ba4 RS |
3568 | /* Return the number of elements in vector type VECTYPE, which is associated |
3569 | with a SIMD clone. At present these vectors always have a constant | |
3570 | length. */ | |
3571 | ||
3572 | static unsigned HOST_WIDE_INT | |
3573 | simd_clone_subparts (tree vectype) | |
3574 | { | |
928686b1 | 3575 | return TYPE_VECTOR_SUBPARTS (vectype).to_constant (); |
cf1b2ba4 RS |
3576 | } |
3577 | ||
0136f8f0 AH |
3578 | /* Function vectorizable_simd_clone_call. |
3579 | ||
3580 | Check if STMT performs a function call that can be vectorized | |
3581 | by calling a simd clone of the function. | |
3582 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3583 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3584 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3585 | ||
3586 | static bool | |
355fe088 TS |
3587 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
3588 | gimple **vec_stmt, slp_tree slp_node) | |
0136f8f0 AH |
3589 | { |
3590 | tree vec_dest; | |
3591 | tree scalar_dest; | |
3592 | tree op, type; | |
3593 | tree vec_oprnd0 = NULL_TREE; | |
3594 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
3595 | tree vectype; | |
3596 | unsigned int nunits; | |
3597 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3598 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 3599 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 3600 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 3601 | tree fndecl, new_temp; |
355fe088 TS |
3602 | gimple *def_stmt; |
3603 | gimple *new_stmt = NULL; | |
0136f8f0 | 3604 | int ncopies, j; |
00426f9a | 3605 | auto_vec<simd_call_arg_info> arginfo; |
0136f8f0 AH |
3606 | vec<tree> vargs = vNULL; |
3607 | size_t i, nargs; | |
3608 | tree lhs, rtype, ratype; | |
e7a74006 | 3609 | vec<constructor_elt, va_gc> *ret_ctor_elts = NULL; |
0136f8f0 AH |
3610 | |
3611 | /* Is STMT a vectorizable call? */ | |
3612 | if (!is_gimple_call (stmt)) | |
3613 | return false; | |
3614 | ||
3615 | fndecl = gimple_call_fndecl (stmt); | |
3616 | if (fndecl == NULL_TREE) | |
3617 | return false; | |
3618 | ||
d52f5295 | 3619 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
3620 | if (node == NULL || node->simd_clones == NULL) |
3621 | return false; | |
3622 | ||
3623 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
3624 | return false; | |
3625 | ||
66c16fd9 RB |
3626 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3627 | && ! vec_stmt) | |
0136f8f0 AH |
3628 | return false; |
3629 | ||
3630 | if (gimple_call_lhs (stmt) | |
3631 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
3632 | return false; | |
3633 | ||
3634 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
3635 | ||
3636 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3637 | ||
3638 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
3639 | return false; | |
3640 | ||
3641 | /* FORNOW */ | |
fce57248 | 3642 | if (slp_node) |
0136f8f0 AH |
3643 | return false; |
3644 | ||
3645 | /* Process function arguments. */ | |
3646 | nargs = gimple_call_num_args (stmt); | |
3647 | ||
3648 | /* Bail out if the function has zero arguments. */ | |
3649 | if (nargs == 0) | |
3650 | return false; | |
3651 | ||
00426f9a | 3652 | arginfo.reserve (nargs, true); |
0136f8f0 AH |
3653 | |
3654 | for (i = 0; i < nargs; i++) | |
3655 | { | |
3656 | simd_call_arg_info thisarginfo; | |
3657 | affine_iv iv; | |
3658 | ||
3659 | thisarginfo.linear_step = 0; | |
3660 | thisarginfo.align = 0; | |
3661 | thisarginfo.op = NULL_TREE; | |
17b658af | 3662 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
3663 | |
3664 | op = gimple_call_arg (stmt, i); | |
81c40241 RB |
3665 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &thisarginfo.dt, |
3666 | &thisarginfo.vectype) | |
0136f8f0 AH |
3667 | || thisarginfo.dt == vect_uninitialized_def) |
3668 | { | |
3669 | if (dump_enabled_p ()) | |
3670 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3671 | "use not simple.\n"); | |
0136f8f0 AH |
3672 | return false; |
3673 | } | |
3674 | ||
3675 | if (thisarginfo.dt == vect_constant_def | |
3676 | || thisarginfo.dt == vect_external_def) | |
3677 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
3678 | else | |
3679 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
3680 | ||
6c9e85fb JJ |
3681 | /* For linear arguments, the analyze phase should have saved |
3682 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
3683 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
3684 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
3685 | { |
3686 | gcc_assert (vec_stmt); | |
3687 | thisarginfo.linear_step | |
17b658af | 3688 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 3689 | thisarginfo.op |
17b658af JJ |
3690 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
3691 | thisarginfo.simd_lane_linear | |
3692 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
3693 | == boolean_true_node); | |
6c9e85fb JJ |
3694 | /* If loop has been peeled for alignment, we need to adjust it. */ |
3695 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
3696 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 3697 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
3698 | { |
3699 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 3700 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
3701 | tree opt = TREE_TYPE (thisarginfo.op); |
3702 | bias = fold_convert (TREE_TYPE (step), bias); | |
3703 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
3704 | thisarginfo.op | |
3705 | = fold_build2 (POINTER_TYPE_P (opt) | |
3706 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
3707 | thisarginfo.op, bias); | |
3708 | } | |
3709 | } | |
3710 | else if (!vec_stmt | |
3711 | && thisarginfo.dt != vect_constant_def | |
3712 | && thisarginfo.dt != vect_external_def | |
3713 | && loop_vinfo | |
3714 | && TREE_CODE (op) == SSA_NAME | |
3715 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
3716 | &iv, false) | |
3717 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
3718 | { |
3719 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
3720 | thisarginfo.op = iv.base; | |
3721 | } | |
3722 | else if ((thisarginfo.dt == vect_constant_def | |
3723 | || thisarginfo.dt == vect_external_def) | |
3724 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
3725 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
3726 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
3727 | linear too. */ | |
3728 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
3729 | && !thisarginfo.linear_step | |
3730 | && !vec_stmt | |
3731 | && thisarginfo.dt != vect_constant_def | |
3732 | && thisarginfo.dt != vect_external_def | |
3733 | && loop_vinfo | |
3734 | && !slp_node | |
3735 | && TREE_CODE (op) == SSA_NAME) | |
3736 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
3737 | |
3738 | arginfo.quick_push (thisarginfo); | |
3739 | } | |
3740 | ||
d9f21f6a RS |
3741 | unsigned HOST_WIDE_INT vf; |
3742 | if (!LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&vf)) | |
3743 | { | |
3744 | if (dump_enabled_p ()) | |
3745 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3746 | "not considering SIMD clones; not yet supported" | |
3747 | " for variable-width vectors.\n"); | |
3748 | return NULL; | |
3749 | } | |
3750 | ||
0136f8f0 AH |
3751 | unsigned int badness = 0; |
3752 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
3753 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
3754 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
3755 | else |
3756 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
3757 | n = n->simdclone->next_clone) | |
3758 | { | |
3759 | unsigned int this_badness = 0; | |
d9f21f6a | 3760 | if (n->simdclone->simdlen > vf |
0136f8f0 AH |
3761 | || n->simdclone->nargs != nargs) |
3762 | continue; | |
d9f21f6a RS |
3763 | if (n->simdclone->simdlen < vf) |
3764 | this_badness += (exact_log2 (vf) | |
0136f8f0 AH |
3765 | - exact_log2 (n->simdclone->simdlen)) * 1024; |
3766 | if (n->simdclone->inbranch) | |
3767 | this_badness += 2048; | |
3768 | int target_badness = targetm.simd_clone.usable (n); | |
3769 | if (target_badness < 0) | |
3770 | continue; | |
3771 | this_badness += target_badness * 512; | |
3772 | /* FORNOW: Have to add code to add the mask argument. */ | |
3773 | if (n->simdclone->inbranch) | |
3774 | continue; | |
3775 | for (i = 0; i < nargs; i++) | |
3776 | { | |
3777 | switch (n->simdclone->args[i].arg_type) | |
3778 | { | |
3779 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3780 | if (!useless_type_conversion_p | |
3781 | (n->simdclone->args[i].orig_type, | |
3782 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
3783 | i = -1; | |
3784 | else if (arginfo[i].dt == vect_constant_def | |
3785 | || arginfo[i].dt == vect_external_def | |
3786 | || arginfo[i].linear_step) | |
3787 | this_badness += 64; | |
3788 | break; | |
3789 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3790 | if (arginfo[i].dt != vect_constant_def | |
3791 | && arginfo[i].dt != vect_external_def) | |
3792 | i = -1; | |
3793 | break; | |
3794 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 3795 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
3796 | if (arginfo[i].dt == vect_constant_def |
3797 | || arginfo[i].dt == vect_external_def | |
3798 | || (arginfo[i].linear_step | |
3799 | != n->simdclone->args[i].linear_step)) | |
3800 | i = -1; | |
3801 | break; | |
3802 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
3803 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
3804 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
3805 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3806 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3807 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3808 | /* FORNOW */ |
3809 | i = -1; | |
3810 | break; | |
3811 | case SIMD_CLONE_ARG_TYPE_MASK: | |
3812 | gcc_unreachable (); | |
3813 | } | |
3814 | if (i == (size_t) -1) | |
3815 | break; | |
3816 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
3817 | { | |
3818 | i = -1; | |
3819 | break; | |
3820 | } | |
3821 | if (arginfo[i].align) | |
3822 | this_badness += (exact_log2 (arginfo[i].align) | |
3823 | - exact_log2 (n->simdclone->args[i].alignment)); | |
3824 | } | |
3825 | if (i == (size_t) -1) | |
3826 | continue; | |
3827 | if (bestn == NULL || this_badness < badness) | |
3828 | { | |
3829 | bestn = n; | |
3830 | badness = this_badness; | |
3831 | } | |
3832 | } | |
3833 | ||
3834 | if (bestn == NULL) | |
00426f9a | 3835 | return false; |
0136f8f0 AH |
3836 | |
3837 | for (i = 0; i < nargs; i++) | |
3838 | if ((arginfo[i].dt == vect_constant_def | |
3839 | || arginfo[i].dt == vect_external_def) | |
3840 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
3841 | { | |
3842 | arginfo[i].vectype | |
3843 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
3844 | i))); | |
3845 | if (arginfo[i].vectype == NULL | |
cf1b2ba4 | 3846 | || (simd_clone_subparts (arginfo[i].vectype) |
0136f8f0 | 3847 | > bestn->simdclone->simdlen)) |
00426f9a | 3848 | return false; |
0136f8f0 AH |
3849 | } |
3850 | ||
3851 | fndecl = bestn->decl; | |
3852 | nunits = bestn->simdclone->simdlen; | |
d9f21f6a | 3853 | ncopies = vf / nunits; |
0136f8f0 AH |
3854 | |
3855 | /* If the function isn't const, only allow it in simd loops where user | |
3856 | has asserted that at least nunits consecutive iterations can be | |
3857 | performed using SIMD instructions. */ | |
3858 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
3859 | && gimple_vuse (stmt)) | |
00426f9a | 3860 | return false; |
0136f8f0 AH |
3861 | |
3862 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3863 | needs to be generated. */ | |
3864 | gcc_assert (ncopies >= 1); | |
3865 | ||
3866 | if (!vec_stmt) /* transformation not required. */ | |
3867 | { | |
6c9e85fb JJ |
3868 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
3869 | for (i = 0; i < nargs; i++) | |
7adb26f2 JJ |
3870 | if ((bestn->simdclone->args[i].arg_type |
3871 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
3872 | || (bestn->simdclone->args[i].arg_type | |
3873 | == SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP)) | |
6c9e85fb | 3874 | { |
17b658af | 3875 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
3876 | + 1); |
3877 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
3878 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
3879 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
3880 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
3881 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
3882 | tree sll = arginfo[i].simd_lane_linear |
3883 | ? boolean_true_node : boolean_false_node; | |
3884 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 3885 | } |
0136f8f0 AH |
3886 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
3887 | if (dump_enabled_p ()) | |
3888 | dump_printf_loc (MSG_NOTE, vect_location, | |
3889 | "=== vectorizable_simd_clone_call ===\n"); | |
3890 | /* vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); */ | |
0136f8f0 AH |
3891 | return true; |
3892 | } | |
3893 | ||
67b8dbac | 3894 | /* Transform. */ |
0136f8f0 AH |
3895 | |
3896 | if (dump_enabled_p ()) | |
3897 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
3898 | ||
3899 | /* Handle def. */ | |
3900 | scalar_dest = gimple_call_lhs (stmt); | |
3901 | vec_dest = NULL_TREE; | |
3902 | rtype = NULL_TREE; | |
3903 | ratype = NULL_TREE; | |
3904 | if (scalar_dest) | |
3905 | { | |
3906 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3907 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
3908 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
3909 | { | |
3910 | ratype = rtype; | |
3911 | rtype = TREE_TYPE (ratype); | |
3912 | } | |
3913 | } | |
3914 | ||
3915 | prev_stmt_info = NULL; | |
3916 | for (j = 0; j < ncopies; ++j) | |
3917 | { | |
3918 | /* Build argument list for the vectorized call. */ | |
3919 | if (j == 0) | |
3920 | vargs.create (nargs); | |
3921 | else | |
3922 | vargs.truncate (0); | |
3923 | ||
3924 | for (i = 0; i < nargs; i++) | |
3925 | { | |
3926 | unsigned int k, l, m, o; | |
3927 | tree atype; | |
3928 | op = gimple_call_arg (stmt, i); | |
3929 | switch (bestn->simdclone->args[i].arg_type) | |
3930 | { | |
3931 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3932 | atype = bestn->simdclone->args[i].vector_type; | |
cf1b2ba4 | 3933 | o = nunits / simd_clone_subparts (atype); |
0136f8f0 AH |
3934 | for (m = j * o; m < (j + 1) * o; m++) |
3935 | { | |
cf1b2ba4 RS |
3936 | if (simd_clone_subparts (atype) |
3937 | < simd_clone_subparts (arginfo[i].vectype)) | |
0136f8f0 | 3938 | { |
73a699ae | 3939 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); |
cf1b2ba4 RS |
3940 | k = (simd_clone_subparts (arginfo[i].vectype) |
3941 | / simd_clone_subparts (atype)); | |
0136f8f0 AH |
3942 | gcc_assert ((k & (k - 1)) == 0); |
3943 | if (m == 0) | |
3944 | vec_oprnd0 | |
81c40241 | 3945 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3946 | else |
3947 | { | |
3948 | vec_oprnd0 = arginfo[i].op; | |
3949 | if ((m & (k - 1)) == 0) | |
3950 | vec_oprnd0 | |
3951 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3952 | vec_oprnd0); | |
3953 | } | |
3954 | arginfo[i].op = vec_oprnd0; | |
3955 | vec_oprnd0 | |
3956 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
92e29a5e | 3957 | bitsize_int (prec), |
0136f8f0 AH |
3958 | bitsize_int ((m & (k - 1)) * prec)); |
3959 | new_stmt | |
b731b390 | 3960 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3961 | vec_oprnd0); |
3962 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3963 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3964 | } | |
3965 | else | |
3966 | { | |
cf1b2ba4 RS |
3967 | k = (simd_clone_subparts (atype) |
3968 | / simd_clone_subparts (arginfo[i].vectype)); | |
0136f8f0 AH |
3969 | gcc_assert ((k & (k - 1)) == 0); |
3970 | vec<constructor_elt, va_gc> *ctor_elts; | |
3971 | if (k != 1) | |
3972 | vec_alloc (ctor_elts, k); | |
3973 | else | |
3974 | ctor_elts = NULL; | |
3975 | for (l = 0; l < k; l++) | |
3976 | { | |
3977 | if (m == 0 && l == 0) | |
3978 | vec_oprnd0 | |
81c40241 | 3979 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3980 | else |
3981 | vec_oprnd0 | |
3982 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3983 | arginfo[i].op); | |
3984 | arginfo[i].op = vec_oprnd0; | |
3985 | if (k == 1) | |
3986 | break; | |
3987 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
3988 | vec_oprnd0); | |
3989 | } | |
3990 | if (k == 1) | |
3991 | vargs.safe_push (vec_oprnd0); | |
3992 | else | |
3993 | { | |
3994 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
3995 | new_stmt | |
b731b390 | 3996 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3997 | vec_oprnd0); |
3998 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3999 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
4000 | } | |
4001 | } | |
4002 | } | |
4003 | break; | |
4004 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
4005 | vargs.safe_push (op); | |
4006 | break; | |
4007 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
7adb26f2 | 4008 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
4009 | if (j == 0) |
4010 | { | |
4011 | gimple_seq stmts; | |
4012 | arginfo[i].op | |
4013 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
4014 | NULL_TREE); | |
4015 | if (stmts != NULL) | |
4016 | { | |
4017 | basic_block new_bb; | |
4018 | edge pe = loop_preheader_edge (loop); | |
4019 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
4020 | gcc_assert (!new_bb); | |
4021 | } | |
17b658af JJ |
4022 | if (arginfo[i].simd_lane_linear) |
4023 | { | |
4024 | vargs.safe_push (arginfo[i].op); | |
4025 | break; | |
4026 | } | |
b731b390 | 4027 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 4028 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
0136f8f0 | 4029 | set_vinfo_for_stmt (new_phi, |
310213d4 | 4030 | new_stmt_vec_info (new_phi, loop_vinfo)); |
0136f8f0 AH |
4031 | add_phi_arg (new_phi, arginfo[i].op, |
4032 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
4033 | enum tree_code code | |
4034 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
4035 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
4036 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
4037 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
4038 | widest_int cst |
4039 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
4040 | ncopies * nunits); | |
4041 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 4042 | tree phi_arg = copy_ssa_name (op); |
0d0e4a03 JJ |
4043 | new_stmt |
4044 | = gimple_build_assign (phi_arg, code, phi_res, tcst); | |
0136f8f0 AH |
4045 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
4046 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
4047 | set_vinfo_for_stmt (new_stmt, | |
310213d4 | 4048 | new_stmt_vec_info (new_stmt, loop_vinfo)); |
0136f8f0 AH |
4049 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
4050 | UNKNOWN_LOCATION); | |
4051 | arginfo[i].op = phi_res; | |
4052 | vargs.safe_push (phi_res); | |
4053 | } | |
4054 | else | |
4055 | { | |
4056 | enum tree_code code | |
4057 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
4058 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
4059 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
4060 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
4061 | widest_int cst |
4062 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
4063 | j * nunits); | |
4064 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 4065 | new_temp = make_ssa_name (TREE_TYPE (op)); |
0d0e4a03 JJ |
4066 | new_stmt = gimple_build_assign (new_temp, code, |
4067 | arginfo[i].op, tcst); | |
0136f8f0 AH |
4068 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4069 | vargs.safe_push (new_temp); | |
4070 | } | |
4071 | break; | |
7adb26f2 JJ |
4072 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
4073 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
0136f8f0 | 4074 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
e01d41e5 JJ |
4075 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
4076 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
4077 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
4078 | default: |
4079 | gcc_unreachable (); | |
4080 | } | |
4081 | } | |
4082 | ||
4083 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
4084 | if (vec_dest) | |
4085 | { | |
cf1b2ba4 | 4086 | gcc_assert (ratype || simd_clone_subparts (rtype) == nunits); |
0136f8f0 | 4087 | if (ratype) |
b731b390 | 4088 | new_temp = create_tmp_var (ratype); |
cf1b2ba4 RS |
4089 | else if (simd_clone_subparts (vectype) |
4090 | == simd_clone_subparts (rtype)) | |
0136f8f0 AH |
4091 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4092 | else | |
4093 | new_temp = make_ssa_name (rtype, new_stmt); | |
4094 | gimple_call_set_lhs (new_stmt, new_temp); | |
4095 | } | |
4096 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4097 | ||
4098 | if (vec_dest) | |
4099 | { | |
cf1b2ba4 | 4100 | if (simd_clone_subparts (vectype) < nunits) |
0136f8f0 AH |
4101 | { |
4102 | unsigned int k, l; | |
73a699ae RS |
4103 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); |
4104 | poly_uint64 bytes = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
cf1b2ba4 | 4105 | k = nunits / simd_clone_subparts (vectype); |
0136f8f0 AH |
4106 | gcc_assert ((k & (k - 1)) == 0); |
4107 | for (l = 0; l < k; l++) | |
4108 | { | |
4109 | tree t; | |
4110 | if (ratype) | |
4111 | { | |
4112 | t = build_fold_addr_expr (new_temp); | |
4113 | t = build2 (MEM_REF, vectype, t, | |
73a699ae | 4114 | build_int_cst (TREE_TYPE (t), l * bytes)); |
0136f8f0 AH |
4115 | } |
4116 | else | |
4117 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
92e29a5e | 4118 | bitsize_int (prec), bitsize_int (l * prec)); |
0136f8f0 | 4119 | new_stmt |
b731b390 | 4120 | = gimple_build_assign (make_ssa_name (vectype), t); |
0136f8f0 AH |
4121 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4122 | if (j == 0 && l == 0) | |
4123 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4124 | else | |
4125 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4126 | ||
4127 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4128 | } | |
4129 | ||
4130 | if (ratype) | |
4131 | { | |
4132 | tree clobber = build_constructor (ratype, NULL); | |
4133 | TREE_THIS_VOLATILE (clobber) = 1; | |
4134 | new_stmt = gimple_build_assign (new_temp, clobber); | |
4135 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4136 | } | |
4137 | continue; | |
4138 | } | |
cf1b2ba4 | 4139 | else if (simd_clone_subparts (vectype) > nunits) |
0136f8f0 | 4140 | { |
cf1b2ba4 RS |
4141 | unsigned int k = (simd_clone_subparts (vectype) |
4142 | / simd_clone_subparts (rtype)); | |
0136f8f0 AH |
4143 | gcc_assert ((k & (k - 1)) == 0); |
4144 | if ((j & (k - 1)) == 0) | |
4145 | vec_alloc (ret_ctor_elts, k); | |
4146 | if (ratype) | |
4147 | { | |
cf1b2ba4 | 4148 | unsigned int m, o = nunits / simd_clone_subparts (rtype); |
0136f8f0 AH |
4149 | for (m = 0; m < o; m++) |
4150 | { | |
4151 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
4152 | size_int (m), NULL_TREE, NULL_TREE); | |
4153 | new_stmt | |
b731b390 | 4154 | = gimple_build_assign (make_ssa_name (rtype), tem); |
0136f8f0 AH |
4155 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4156 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
4157 | gimple_assign_lhs (new_stmt)); | |
4158 | } | |
4159 | tree clobber = build_constructor (ratype, NULL); | |
4160 | TREE_THIS_VOLATILE (clobber) = 1; | |
4161 | new_stmt = gimple_build_assign (new_temp, clobber); | |
4162 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4163 | } | |
4164 | else | |
4165 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
4166 | if ((j & (k - 1)) != k - 1) | |
4167 | continue; | |
4168 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
4169 | new_stmt | |
b731b390 | 4170 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
0136f8f0 AH |
4171 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4172 | ||
4173 | if ((unsigned) j == k - 1) | |
4174 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4175 | else | |
4176 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4177 | ||
4178 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4179 | continue; | |
4180 | } | |
4181 | else if (ratype) | |
4182 | { | |
4183 | tree t = build_fold_addr_expr (new_temp); | |
4184 | t = build2 (MEM_REF, vectype, t, | |
4185 | build_int_cst (TREE_TYPE (t), 0)); | |
4186 | new_stmt | |
b731b390 | 4187 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
0136f8f0 AH |
4188 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4189 | tree clobber = build_constructor (ratype, NULL); | |
4190 | TREE_THIS_VOLATILE (clobber) = 1; | |
4191 | vect_finish_stmt_generation (stmt, | |
4192 | gimple_build_assign (new_temp, | |
4193 | clobber), gsi); | |
4194 | } | |
4195 | } | |
4196 | ||
4197 | if (j == 0) | |
4198 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4199 | else | |
4200 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4201 | ||
4202 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4203 | } | |
4204 | ||
4205 | vargs.release (); | |
4206 | ||
4207 | /* The call in STMT might prevent it from being removed in dce. | |
4208 | We however cannot remove it here, due to the way the ssa name | |
4209 | it defines is mapped to the new definition. So just replace | |
4210 | rhs of the statement with something harmless. */ | |
4211 | ||
4212 | if (slp_node) | |
4213 | return true; | |
4214 | ||
4215 | if (scalar_dest) | |
4216 | { | |
4217 | type = TREE_TYPE (scalar_dest); | |
4218 | if (is_pattern_stmt_p (stmt_info)) | |
4219 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
4220 | else | |
4221 | lhs = gimple_call_lhs (stmt); | |
4222 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
4223 | } | |
4224 | else | |
4225 | new_stmt = gimple_build_nop (); | |
4226 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
4227 | set_vinfo_for_stmt (stmt, NULL); | |
4228 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 4229 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
4230 | unlink_stmt_vdef (stmt); |
4231 | ||
4232 | return true; | |
4233 | } | |
4234 | ||
4235 | ||
ebfd146a IR |
4236 | /* Function vect_gen_widened_results_half |
4237 | ||
4238 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 4239 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 4240 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
4241 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
4242 | needs to be created (DECL is a function-decl of a target-builtin). | |
4243 | STMT is the original scalar stmt that we are vectorizing. */ | |
4244 | ||
355fe088 | 4245 | static gimple * |
ebfd146a IR |
4246 | vect_gen_widened_results_half (enum tree_code code, |
4247 | tree decl, | |
4248 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
4249 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 4250 | gimple *stmt) |
b8698a0f | 4251 | { |
355fe088 | 4252 | gimple *new_stmt; |
b8698a0f L |
4253 | tree new_temp; |
4254 | ||
4255 | /* Generate half of the widened result: */ | |
4256 | if (code == CALL_EXPR) | |
4257 | { | |
4258 | /* Target specific support */ | |
ebfd146a IR |
4259 | if (op_type == binary_op) |
4260 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
4261 | else | |
4262 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
4263 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4264 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
4265 | } |
4266 | else | |
ebfd146a | 4267 | { |
b8698a0f L |
4268 | /* Generic support */ |
4269 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
4270 | if (op_type != binary_op) |
4271 | vec_oprnd1 = NULL; | |
0d0e4a03 | 4272 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
4273 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4274 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 4275 | } |
ebfd146a IR |
4276 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4277 | ||
ebfd146a IR |
4278 | return new_stmt; |
4279 | } | |
4280 | ||
4a00c761 JJ |
4281 | |
4282 | /* Get vectorized definitions for loop-based vectorization. For the first | |
4283 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
4284 | scalar operand), and for the rest we get a copy with | |
4285 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
4286 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
4287 | The vectors are collected into VEC_OPRNDS. */ | |
4288 | ||
4289 | static void | |
355fe088 | 4290 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 4291 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
4292 | { |
4293 | tree vec_oprnd; | |
4294 | ||
4295 | /* Get first vector operand. */ | |
4296 | /* All the vector operands except the very first one (that is scalar oprnd) | |
4297 | are stmt copies. */ | |
4298 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 4299 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
4300 | else |
4301 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
4302 | ||
9771b263 | 4303 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
4304 | |
4305 | /* Get second vector operand. */ | |
4306 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 4307 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
4308 | |
4309 | *oprnd = vec_oprnd; | |
4310 | ||
4311 | /* For conversion in multiple steps, continue to get operands | |
4312 | recursively. */ | |
4313 | if (multi_step_cvt) | |
4314 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
4315 | } | |
4316 | ||
4317 | ||
4318 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
4319 | For multi-step conversions store the resulting vectors and call the function | |
4320 | recursively. */ | |
4321 | ||
4322 | static void | |
9771b263 | 4323 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 4324 | int multi_step_cvt, gimple *stmt, |
9771b263 | 4325 | vec<tree> vec_dsts, |
4a00c761 JJ |
4326 | gimple_stmt_iterator *gsi, |
4327 | slp_tree slp_node, enum tree_code code, | |
4328 | stmt_vec_info *prev_stmt_info) | |
4329 | { | |
4330 | unsigned int i; | |
4331 | tree vop0, vop1, new_tmp, vec_dest; | |
355fe088 | 4332 | gimple *new_stmt; |
4a00c761 JJ |
4333 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
4334 | ||
9771b263 | 4335 | vec_dest = vec_dsts.pop (); |
4a00c761 | 4336 | |
9771b263 | 4337 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
4338 | { |
4339 | /* Create demotion operation. */ | |
9771b263 DN |
4340 | vop0 = (*vec_oprnds)[i]; |
4341 | vop1 = (*vec_oprnds)[i + 1]; | |
0d0e4a03 | 4342 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
4343 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
4344 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
4345 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4346 | ||
4347 | if (multi_step_cvt) | |
4348 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 4349 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
4350 | else |
4351 | { | |
4352 | /* This is the last step of the conversion sequence. Store the | |
4353 | vectors in SLP_NODE or in vector info of the scalar statement | |
4354 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
4355 | if (slp_node) | |
9771b263 | 4356 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4357 | else |
c689ce1e RB |
4358 | { |
4359 | if (!*prev_stmt_info) | |
4360 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4361 | else | |
4362 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
4a00c761 | 4363 | |
c689ce1e RB |
4364 | *prev_stmt_info = vinfo_for_stmt (new_stmt); |
4365 | } | |
4a00c761 JJ |
4366 | } |
4367 | } | |
4368 | ||
4369 | /* For multi-step demotion operations we first generate demotion operations | |
4370 | from the source type to the intermediate types, and then combine the | |
4371 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
4372 | type. */ | |
4373 | if (multi_step_cvt) | |
4374 | { | |
4375 | /* At each level of recursion we have half of the operands we had at the | |
4376 | previous level. */ | |
9771b263 | 4377 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
4378 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
4379 | stmt, vec_dsts, gsi, slp_node, | |
4380 | VEC_PACK_TRUNC_EXPR, | |
4381 | prev_stmt_info); | |
4382 | } | |
4383 | ||
9771b263 | 4384 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4385 | } |
4386 | ||
4387 | ||
4388 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
4389 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
4390 | the resulting vectors and call the function recursively. */ | |
4391 | ||
4392 | static void | |
9771b263 DN |
4393 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
4394 | vec<tree> *vec_oprnds1, | |
355fe088 | 4395 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
4396 | gimple_stmt_iterator *gsi, |
4397 | enum tree_code code1, | |
4398 | enum tree_code code2, tree decl1, | |
4399 | tree decl2, int op_type) | |
4400 | { | |
4401 | int i; | |
4402 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 4403 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 4404 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 4405 | |
9771b263 DN |
4406 | vec_tmp.create (vec_oprnds0->length () * 2); |
4407 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
4408 | { |
4409 | if (op_type == binary_op) | |
9771b263 | 4410 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
4411 | else |
4412 | vop1 = NULL_TREE; | |
4413 | ||
4414 | /* Generate the two halves of promotion operation. */ | |
4415 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
4416 | op_type, vec_dest, gsi, stmt); | |
4417 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
4418 | op_type, vec_dest, gsi, stmt); | |
4419 | if (is_gimple_call (new_stmt1)) | |
4420 | { | |
4421 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
4422 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
4423 | } | |
4424 | else | |
4425 | { | |
4426 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
4427 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
4428 | } | |
4429 | ||
4430 | /* Store the results for the next step. */ | |
9771b263 DN |
4431 | vec_tmp.quick_push (new_tmp1); |
4432 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
4433 | } |
4434 | ||
689eaba3 | 4435 | vec_oprnds0->release (); |
4a00c761 JJ |
4436 | *vec_oprnds0 = vec_tmp; |
4437 | } | |
4438 | ||
4439 | ||
b8698a0f L |
4440 | /* Check if STMT performs a conversion operation, that can be vectorized. |
4441 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 4442 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
4443 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
4444 | ||
4445 | static bool | |
355fe088 TS |
4446 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
4447 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4448 | { |
4449 | tree vec_dest; | |
4450 | tree scalar_dest; | |
4a00c761 | 4451 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
4452 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
4453 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4454 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4455 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 4456 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
4457 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
4458 | tree new_temp; | |
355fe088 | 4459 | gimple *def_stmt; |
ebfd146a | 4460 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 4461 | int ndts = 2; |
355fe088 | 4462 | gimple *new_stmt = NULL; |
ebfd146a | 4463 | stmt_vec_info prev_stmt_info; |
062d5ccc RS |
4464 | poly_uint64 nunits_in; |
4465 | poly_uint64 nunits_out; | |
ebfd146a | 4466 | tree vectype_out, vectype_in; |
4a00c761 JJ |
4467 | int ncopies, i, j; |
4468 | tree lhs_type, rhs_type; | |
ebfd146a | 4469 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
4470 | vec<tree> vec_oprnds0 = vNULL; |
4471 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 4472 | tree vop0; |
4a00c761 | 4473 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4474 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 4475 | int multi_step_cvt = 0; |
6e1aa848 | 4476 | vec<tree> interm_types = vNULL; |
4a00c761 JJ |
4477 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
4478 | int op_type; | |
4a00c761 | 4479 | unsigned short fltsz; |
ebfd146a IR |
4480 | |
4481 | /* Is STMT a vectorizable conversion? */ | |
4482 | ||
4a00c761 | 4483 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4484 | return false; |
4485 | ||
66c16fd9 RB |
4486 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4487 | && ! vec_stmt) | |
ebfd146a IR |
4488 | return false; |
4489 | ||
4490 | if (!is_gimple_assign (stmt)) | |
4491 | return false; | |
4492 | ||
4493 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4494 | return false; | |
4495 | ||
4496 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
4497 | if (!CONVERT_EXPR_CODE_P (code) |
4498 | && code != FIX_TRUNC_EXPR | |
4499 | && code != FLOAT_EXPR | |
4500 | && code != WIDEN_MULT_EXPR | |
4501 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
4502 | return false; |
4503 | ||
4a00c761 JJ |
4504 | op_type = TREE_CODE_LENGTH (code); |
4505 | ||
ebfd146a | 4506 | /* Check types of lhs and rhs. */ |
b690cc0f | 4507 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 4508 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
4509 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
4510 | ||
ebfd146a IR |
4511 | op0 = gimple_assign_rhs1 (stmt); |
4512 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
4513 | |
4514 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4515 | && !((INTEGRAL_TYPE_P (lhs_type) | |
4516 | && INTEGRAL_TYPE_P (rhs_type)) | |
4517 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
4518 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
4519 | return false; | |
4520 | ||
e6f5c25d IE |
4521 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4522 | && ((INTEGRAL_TYPE_P (lhs_type) | |
2be65d9e | 4523 | && !type_has_mode_precision_p (lhs_type)) |
e6f5c25d | 4524 | || (INTEGRAL_TYPE_P (rhs_type) |
2be65d9e | 4525 | && !type_has_mode_precision_p (rhs_type)))) |
4a00c761 | 4526 | { |
73fbfcad | 4527 | if (dump_enabled_p ()) |
78c60e3d | 4528 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
4529 | "type conversion to/from bit-precision unsupported." |
4530 | "\n"); | |
4a00c761 JJ |
4531 | return false; |
4532 | } | |
4533 | ||
b690cc0f | 4534 | /* Check the operands of the operation. */ |
81c40241 | 4535 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype_in)) |
b690cc0f | 4536 | { |
73fbfcad | 4537 | if (dump_enabled_p ()) |
78c60e3d | 4538 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4539 | "use not simple.\n"); |
b690cc0f RG |
4540 | return false; |
4541 | } | |
4a00c761 JJ |
4542 | if (op_type == binary_op) |
4543 | { | |
4544 | bool ok; | |
4545 | ||
4546 | op1 = gimple_assign_rhs2 (stmt); | |
4547 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
4548 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
4549 | OP1. */ | |
4550 | if (CONSTANT_CLASS_P (op0)) | |
81c40241 | 4551 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &vectype_in); |
4a00c761 | 4552 | else |
81c40241 | 4553 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1]); |
4a00c761 JJ |
4554 | |
4555 | if (!ok) | |
4556 | { | |
73fbfcad | 4557 | if (dump_enabled_p ()) |
78c60e3d | 4558 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4559 | "use not simple.\n"); |
4a00c761 JJ |
4560 | return false; |
4561 | } | |
4562 | } | |
4563 | ||
b690cc0f RG |
4564 | /* If op0 is an external or constant defs use a vector type of |
4565 | the same size as the output vector type. */ | |
ebfd146a | 4566 | if (!vectype_in) |
b690cc0f | 4567 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
4568 | if (vec_stmt) |
4569 | gcc_assert (vectype_in); | |
4570 | if (!vectype_in) | |
4571 | { | |
73fbfcad | 4572 | if (dump_enabled_p ()) |
4a00c761 | 4573 | { |
78c60e3d SS |
4574 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4575 | "no vectype for scalar type "); | |
4576 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 4577 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 4578 | } |
7d8930a0 IR |
4579 | |
4580 | return false; | |
4581 | } | |
ebfd146a | 4582 | |
e6f5c25d IE |
4583 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4584 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
4585 | { | |
4586 | if (dump_enabled_p ()) | |
4587 | { | |
4588 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4589 | "can't convert between boolean and non " | |
4590 | "boolean vectors"); | |
4591 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
4592 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
4593 | } | |
4594 | ||
4595 | return false; | |
4596 | } | |
4597 | ||
b690cc0f RG |
4598 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
4599 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
062d5ccc | 4600 | if (known_eq (nunits_out, nunits_in)) |
ebfd146a | 4601 | modifier = NONE; |
062d5ccc RS |
4602 | else if (multiple_p (nunits_out, nunits_in)) |
4603 | modifier = NARROW; | |
ebfd146a | 4604 | else |
062d5ccc RS |
4605 | { |
4606 | gcc_checking_assert (multiple_p (nunits_in, nunits_out)); | |
4607 | modifier = WIDEN; | |
4608 | } | |
ebfd146a | 4609 | |
ff802fa1 IR |
4610 | /* Multiple types in SLP are handled by creating the appropriate number of |
4611 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4612 | case of SLP. */ | |
fce57248 | 4613 | if (slp_node) |
ebfd146a | 4614 | ncopies = 1; |
4a00c761 | 4615 | else if (modifier == NARROW) |
e8f142e2 | 4616 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
4a00c761 | 4617 | else |
e8f142e2 | 4618 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
b8698a0f | 4619 | |
ebfd146a IR |
4620 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
4621 | needs to be generated. */ | |
4622 | gcc_assert (ncopies >= 1); | |
4623 | ||
16d22000 RS |
4624 | bool found_mode = false; |
4625 | scalar_mode lhs_mode = SCALAR_TYPE_MODE (lhs_type); | |
4626 | scalar_mode rhs_mode = SCALAR_TYPE_MODE (rhs_type); | |
4627 | opt_scalar_mode rhs_mode_iter; | |
b397965c | 4628 | |
ebfd146a | 4629 | /* Supportable by target? */ |
4a00c761 | 4630 | switch (modifier) |
ebfd146a | 4631 | { |
4a00c761 JJ |
4632 | case NONE: |
4633 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4634 | return false; | |
4635 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
4636 | &decl1, &code1)) | |
4637 | break; | |
4638 | /* FALLTHRU */ | |
4639 | unsupported: | |
73fbfcad | 4640 | if (dump_enabled_p ()) |
78c60e3d | 4641 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4642 | "conversion not supported by target.\n"); |
ebfd146a | 4643 | return false; |
ebfd146a | 4644 | |
4a00c761 JJ |
4645 | case WIDEN: |
4646 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
4647 | &code1, &code2, &multi_step_cvt, |
4648 | &interm_types)) | |
4a00c761 JJ |
4649 | { |
4650 | /* Binary widening operation can only be supported directly by the | |
4651 | architecture. */ | |
4652 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
4653 | break; | |
4654 | } | |
4655 | ||
4656 | if (code != FLOAT_EXPR | |
b397965c | 4657 | || GET_MODE_SIZE (lhs_mode) <= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4658 | goto unsupported; |
4659 | ||
b397965c | 4660 | fltsz = GET_MODE_SIZE (lhs_mode); |
16d22000 | 4661 | FOR_EACH_2XWIDER_MODE (rhs_mode_iter, rhs_mode) |
4a00c761 | 4662 | { |
16d22000 | 4663 | rhs_mode = rhs_mode_iter.require (); |
c94843d2 RS |
4664 | if (GET_MODE_SIZE (rhs_mode) > fltsz) |
4665 | break; | |
4666 | ||
4a00c761 JJ |
4667 | cvt_type |
4668 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4669 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4670 | if (cvt_type == NULL_TREE) | |
4671 | goto unsupported; | |
4672 | ||
4673 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4674 | { | |
4675 | if (!supportable_convert_operation (code, vectype_out, | |
4676 | cvt_type, &decl1, &codecvt1)) | |
4677 | goto unsupported; | |
4678 | } | |
4679 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
4680 | cvt_type, &codecvt1, |
4681 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
4682 | &interm_types)) |
4683 | continue; | |
4684 | else | |
4685 | gcc_assert (multi_step_cvt == 0); | |
4686 | ||
4687 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
4688 | vectype_in, &code1, &code2, |
4689 | &multi_step_cvt, &interm_types)) | |
16d22000 RS |
4690 | { |
4691 | found_mode = true; | |
4692 | break; | |
4693 | } | |
4a00c761 JJ |
4694 | } |
4695 | ||
16d22000 | 4696 | if (!found_mode) |
4a00c761 JJ |
4697 | goto unsupported; |
4698 | ||
4699 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4700 | codecvt2 = ERROR_MARK; | |
4701 | else | |
4702 | { | |
4703 | multi_step_cvt++; | |
9771b263 | 4704 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
4705 | cvt_type = NULL_TREE; |
4706 | } | |
4707 | break; | |
4708 | ||
4709 | case NARROW: | |
4710 | gcc_assert (op_type == unary_op); | |
4711 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
4712 | &code1, &multi_step_cvt, | |
4713 | &interm_types)) | |
4714 | break; | |
4715 | ||
4716 | if (code != FIX_TRUNC_EXPR | |
b397965c | 4717 | || GET_MODE_SIZE (lhs_mode) >= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4718 | goto unsupported; |
4719 | ||
4a00c761 JJ |
4720 | cvt_type |
4721 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4722 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4723 | if (cvt_type == NULL_TREE) | |
4724 | goto unsupported; | |
4725 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
4726 | &decl1, &codecvt1)) | |
4727 | goto unsupported; | |
4728 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
4729 | &code1, &multi_step_cvt, | |
4730 | &interm_types)) | |
4731 | break; | |
4732 | goto unsupported; | |
4733 | ||
4734 | default: | |
4735 | gcc_unreachable (); | |
ebfd146a IR |
4736 | } |
4737 | ||
4738 | if (!vec_stmt) /* transformation not required. */ | |
4739 | { | |
73fbfcad | 4740 | if (dump_enabled_p ()) |
78c60e3d | 4741 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4742 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 4743 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
4744 | { |
4745 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
78604de0 RB |
4746 | if (!slp_node) |
4747 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); | |
8bd37302 | 4748 | } |
4a00c761 JJ |
4749 | else if (modifier == NARROW) |
4750 | { | |
4751 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
78604de0 RB |
4752 | if (!slp_node) |
4753 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); | |
4a00c761 JJ |
4754 | } |
4755 | else | |
4756 | { | |
4757 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
78604de0 RB |
4758 | if (!slp_node) |
4759 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); | |
4a00c761 | 4760 | } |
9771b263 | 4761 | interm_types.release (); |
ebfd146a IR |
4762 | return true; |
4763 | } | |
4764 | ||
67b8dbac | 4765 | /* Transform. */ |
73fbfcad | 4766 | if (dump_enabled_p ()) |
78c60e3d | 4767 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4768 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 4769 | |
4a00c761 JJ |
4770 | if (op_type == binary_op) |
4771 | { | |
4772 | if (CONSTANT_CLASS_P (op0)) | |
4773 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
4774 | else if (CONSTANT_CLASS_P (op1)) | |
4775 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
4776 | } | |
4777 | ||
4778 | /* In case of multi-step conversion, we first generate conversion operations | |
4779 | to the intermediate types, and then from that types to the final one. | |
4780 | We create vector destinations for the intermediate type (TYPES) received | |
4781 | from supportable_*_operation, and store them in the correct order | |
4782 | for future use in vect_create_vectorized_*_stmts (). */ | |
8c681247 | 4783 | auto_vec<tree> vec_dsts (multi_step_cvt + 1); |
82294ec1 JJ |
4784 | vec_dest = vect_create_destination_var (scalar_dest, |
4785 | (cvt_type && modifier == WIDEN) | |
4786 | ? cvt_type : vectype_out); | |
9771b263 | 4787 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4788 | |
4789 | if (multi_step_cvt) | |
4790 | { | |
9771b263 DN |
4791 | for (i = interm_types.length () - 1; |
4792 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
4793 | { |
4794 | vec_dest = vect_create_destination_var (scalar_dest, | |
4795 | intermediate_type); | |
9771b263 | 4796 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4797 | } |
4798 | } | |
ebfd146a | 4799 | |
4a00c761 | 4800 | if (cvt_type) |
82294ec1 JJ |
4801 | vec_dest = vect_create_destination_var (scalar_dest, |
4802 | modifier == WIDEN | |
4803 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
4804 | |
4805 | if (!slp_node) | |
4806 | { | |
30862efc | 4807 | if (modifier == WIDEN) |
4a00c761 | 4808 | { |
c3284718 | 4809 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 4810 | if (op_type == binary_op) |
9771b263 | 4811 | vec_oprnds1.create (1); |
4a00c761 | 4812 | } |
30862efc | 4813 | else if (modifier == NARROW) |
9771b263 DN |
4814 | vec_oprnds0.create ( |
4815 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
4816 | } |
4817 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 4818 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 4819 | |
4a00c761 | 4820 | last_oprnd = op0; |
ebfd146a IR |
4821 | prev_stmt_info = NULL; |
4822 | switch (modifier) | |
4823 | { | |
4824 | case NONE: | |
4825 | for (j = 0; j < ncopies; j++) | |
4826 | { | |
ebfd146a | 4827 | if (j == 0) |
306b0c92 | 4828 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node); |
ebfd146a IR |
4829 | else |
4830 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
4831 | ||
9771b263 | 4832 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4833 | { |
4834 | /* Arguments are ready, create the new vector stmt. */ | |
4835 | if (code1 == CALL_EXPR) | |
4836 | { | |
4837 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4838 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4839 | gimple_call_set_lhs (new_stmt, new_temp); | |
4840 | } | |
4841 | else | |
4842 | { | |
4843 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
0d0e4a03 | 4844 | new_stmt = gimple_build_assign (vec_dest, code1, vop0); |
4a00c761 JJ |
4845 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4846 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4847 | } | |
4848 | ||
4849 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4850 | if (slp_node) | |
9771b263 | 4851 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
225ce44b RB |
4852 | else |
4853 | { | |
4854 | if (!prev_stmt_info) | |
4855 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4856 | else | |
4857 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4858 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4859 | } | |
4a00c761 | 4860 | } |
ebfd146a IR |
4861 | } |
4862 | break; | |
4863 | ||
4864 | case WIDEN: | |
4865 | /* In case the vectorization factor (VF) is bigger than the number | |
4866 | of elements that we can fit in a vectype (nunits), we have to | |
4867 | generate more than one vector stmt - i.e - we need to "unroll" | |
4868 | the vector stmt by a factor VF/nunits. */ | |
4869 | for (j = 0; j < ncopies; j++) | |
4870 | { | |
4a00c761 | 4871 | /* Handle uses. */ |
ebfd146a | 4872 | if (j == 0) |
4a00c761 JJ |
4873 | { |
4874 | if (slp_node) | |
4875 | { | |
4876 | if (code == WIDEN_LSHIFT_EXPR) | |
4877 | { | |
4878 | unsigned int k; | |
ebfd146a | 4879 | |
4a00c761 JJ |
4880 | vec_oprnd1 = op1; |
4881 | /* Store vec_oprnd1 for every vector stmt to be created | |
4882 | for SLP_NODE. We check during the analysis that all | |
4883 | the shift arguments are the same. */ | |
4884 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4885 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4886 | |
4887 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 4888 | slp_node); |
4a00c761 JJ |
4889 | } |
4890 | else | |
4891 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
306b0c92 | 4892 | &vec_oprnds1, slp_node); |
4a00c761 JJ |
4893 | } |
4894 | else | |
4895 | { | |
81c40241 | 4896 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 4897 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
4898 | if (op_type == binary_op) |
4899 | { | |
4900 | if (code == WIDEN_LSHIFT_EXPR) | |
4901 | vec_oprnd1 = op1; | |
4902 | else | |
81c40241 | 4903 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 4904 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4905 | } |
4906 | } | |
4907 | } | |
ebfd146a | 4908 | else |
4a00c761 JJ |
4909 | { |
4910 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
4911 | vec_oprnds0.truncate (0); |
4912 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
4913 | if (op_type == binary_op) |
4914 | { | |
4915 | if (code == WIDEN_LSHIFT_EXPR) | |
4916 | vec_oprnd1 = op1; | |
4917 | else | |
4918 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
4919 | vec_oprnd1); | |
9771b263 DN |
4920 | vec_oprnds1.truncate (0); |
4921 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
4922 | } |
4923 | } | |
ebfd146a | 4924 | |
4a00c761 JJ |
4925 | /* Arguments are ready. Create the new vector stmts. */ |
4926 | for (i = multi_step_cvt; i >= 0; i--) | |
4927 | { | |
9771b263 | 4928 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
4929 | enum tree_code c1 = code1, c2 = code2; |
4930 | if (i == 0 && codecvt2 != ERROR_MARK) | |
4931 | { | |
4932 | c1 = codecvt1; | |
4933 | c2 = codecvt2; | |
4934 | } | |
4935 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
4936 | &vec_oprnds1, | |
4937 | stmt, this_dest, gsi, | |
4938 | c1, c2, decl1, decl2, | |
4939 | op_type); | |
4940 | } | |
4941 | ||
9771b263 | 4942 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4943 | { |
4944 | if (cvt_type) | |
4945 | { | |
4946 | if (codecvt1 == CALL_EXPR) | |
4947 | { | |
4948 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4949 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4950 | gimple_call_set_lhs (new_stmt, new_temp); | |
4951 | } | |
4952 | else | |
4953 | { | |
4954 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4955 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4956 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4957 | vop0); | |
4a00c761 JJ |
4958 | } |
4959 | ||
4960 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4961 | } | |
4962 | else | |
4963 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
4964 | ||
4965 | if (slp_node) | |
9771b263 | 4966 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4967 | else |
c689ce1e RB |
4968 | { |
4969 | if (!prev_stmt_info) | |
4970 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4971 | else | |
4972 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4973 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4974 | } | |
4a00c761 | 4975 | } |
ebfd146a | 4976 | } |
4a00c761 JJ |
4977 | |
4978 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
4979 | break; |
4980 | ||
4981 | case NARROW: | |
4982 | /* In case the vectorization factor (VF) is bigger than the number | |
4983 | of elements that we can fit in a vectype (nunits), we have to | |
4984 | generate more than one vector stmt - i.e - we need to "unroll" | |
4985 | the vector stmt by a factor VF/nunits. */ | |
4986 | for (j = 0; j < ncopies; j++) | |
4987 | { | |
4988 | /* Handle uses. */ | |
4a00c761 JJ |
4989 | if (slp_node) |
4990 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 4991 | slp_node); |
ebfd146a IR |
4992 | else |
4993 | { | |
9771b263 | 4994 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
4995 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
4996 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
4997 | } |
4998 | ||
4a00c761 JJ |
4999 | /* Arguments are ready. Create the new vector stmts. */ |
5000 | if (cvt_type) | |
9771b263 | 5001 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
5002 | { |
5003 | if (codecvt1 == CALL_EXPR) | |
5004 | { | |
5005 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
5006 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5007 | gimple_call_set_lhs (new_stmt, new_temp); | |
5008 | } | |
5009 | else | |
5010 | { | |
5011 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 5012 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
5013 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
5014 | vop0); | |
4a00c761 | 5015 | } |
ebfd146a | 5016 | |
4a00c761 | 5017 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 5018 | vec_oprnds0[i] = new_temp; |
4a00c761 | 5019 | } |
ebfd146a | 5020 | |
4a00c761 JJ |
5021 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
5022 | stmt, vec_dsts, gsi, | |
5023 | slp_node, code1, | |
5024 | &prev_stmt_info); | |
ebfd146a IR |
5025 | } |
5026 | ||
5027 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 5028 | break; |
ebfd146a IR |
5029 | } |
5030 | ||
9771b263 DN |
5031 | vec_oprnds0.release (); |
5032 | vec_oprnds1.release (); | |
9771b263 | 5033 | interm_types.release (); |
ebfd146a IR |
5034 | |
5035 | return true; | |
5036 | } | |
ff802fa1 IR |
5037 | |
5038 | ||
ebfd146a IR |
5039 | /* Function vectorizable_assignment. |
5040 | ||
b8698a0f L |
5041 | Check if STMT performs an assignment (copy) that can be vectorized. |
5042 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5043 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5044 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5045 | ||
5046 | static bool | |
355fe088 TS |
5047 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
5048 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
5049 | { |
5050 | tree vec_dest; | |
5051 | tree scalar_dest; | |
5052 | tree op; | |
5053 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
5054 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
5055 | tree new_temp; | |
355fe088 | 5056 | gimple *def_stmt; |
4fc5ebf1 JG |
5057 | enum vect_def_type dt[1] = {vect_unknown_def_type}; |
5058 | int ndts = 1; | |
ebfd146a | 5059 | int ncopies; |
f18b55bd | 5060 | int i, j; |
6e1aa848 | 5061 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5062 | tree vop; |
a70d6342 | 5063 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5064 | vec_info *vinfo = stmt_info->vinfo; |
355fe088 | 5065 | gimple *new_stmt = NULL; |
f18b55bd | 5066 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
5067 | enum tree_code code; |
5068 | tree vectype_in; | |
ebfd146a | 5069 | |
a70d6342 | 5070 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5071 | return false; |
5072 | ||
66c16fd9 RB |
5073 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5074 | && ! vec_stmt) | |
ebfd146a IR |
5075 | return false; |
5076 | ||
5077 | /* Is vectorizable assignment? */ | |
5078 | if (!is_gimple_assign (stmt)) | |
5079 | return false; | |
5080 | ||
5081 | scalar_dest = gimple_assign_lhs (stmt); | |
5082 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
5083 | return false; | |
5084 | ||
fde9c428 | 5085 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 5086 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
5087 | || code == PAREN_EXPR |
5088 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
5089 | op = gimple_assign_rhs1 (stmt); |
5090 | else | |
5091 | return false; | |
5092 | ||
7b7ec6c5 RG |
5093 | if (code == VIEW_CONVERT_EXPR) |
5094 | op = TREE_OPERAND (op, 0); | |
5095 | ||
465c8c19 | 5096 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
928686b1 | 5097 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 JJ |
5098 | |
5099 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5100 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5101 | case of SLP. */ | |
fce57248 | 5102 | if (slp_node) |
465c8c19 JJ |
5103 | ncopies = 1; |
5104 | else | |
e8f142e2 | 5105 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
5106 | |
5107 | gcc_assert (ncopies >= 1); | |
5108 | ||
81c40241 | 5109 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[0], &vectype_in)) |
ebfd146a | 5110 | { |
73fbfcad | 5111 | if (dump_enabled_p ()) |
78c60e3d | 5112 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5113 | "use not simple.\n"); |
ebfd146a IR |
5114 | return false; |
5115 | } | |
5116 | ||
fde9c428 RG |
5117 | /* We can handle NOP_EXPR conversions that do not change the number |
5118 | of elements or the vector size. */ | |
7b7ec6c5 RG |
5119 | if ((CONVERT_EXPR_CODE_P (code) |
5120 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 | 5121 | && (!vectype_in |
928686b1 | 5122 | || maybe_ne (TYPE_VECTOR_SUBPARTS (vectype_in), nunits) |
cf098191 RS |
5123 | || maybe_ne (GET_MODE_SIZE (TYPE_MODE (vectype)), |
5124 | GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
fde9c428 RG |
5125 | return false; |
5126 | ||
7b7b1813 RG |
5127 | /* We do not handle bit-precision changes. */ |
5128 | if ((CONVERT_EXPR_CODE_P (code) | |
5129 | || code == VIEW_CONVERT_EXPR) | |
5130 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2be65d9e RS |
5131 | && (!type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
5132 | || !type_has_mode_precision_p (TREE_TYPE (op))) | |
7b7b1813 RG |
5133 | /* But a conversion that does not change the bit-pattern is ok. */ |
5134 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
5135 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2dab46d5 IE |
5136 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
5137 | /* Conversion between boolean types of different sizes is | |
5138 | a simple assignment in case their vectypes are same | |
5139 | boolean vectors. */ | |
5140 | && (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
5141 | || !VECTOR_BOOLEAN_TYPE_P (vectype_in))) | |
7b7b1813 | 5142 | { |
73fbfcad | 5143 | if (dump_enabled_p ()) |
78c60e3d SS |
5144 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5145 | "type conversion to/from bit-precision " | |
e645e942 | 5146 | "unsupported.\n"); |
7b7b1813 RG |
5147 | return false; |
5148 | } | |
5149 | ||
ebfd146a IR |
5150 | if (!vec_stmt) /* transformation not required. */ |
5151 | { | |
5152 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 5153 | if (dump_enabled_p ()) |
78c60e3d | 5154 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5155 | "=== vectorizable_assignment ===\n"); |
78604de0 RB |
5156 | if (!slp_node) |
5157 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); | |
ebfd146a IR |
5158 | return true; |
5159 | } | |
5160 | ||
67b8dbac | 5161 | /* Transform. */ |
73fbfcad | 5162 | if (dump_enabled_p ()) |
e645e942 | 5163 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
5164 | |
5165 | /* Handle def. */ | |
5166 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5167 | ||
5168 | /* Handle use. */ | |
f18b55bd | 5169 | for (j = 0; j < ncopies; j++) |
ebfd146a | 5170 | { |
f18b55bd IR |
5171 | /* Handle uses. */ |
5172 | if (j == 0) | |
306b0c92 | 5173 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
f18b55bd IR |
5174 | else |
5175 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
5176 | ||
5177 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 5178 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 5179 | { |
7b7ec6c5 RG |
5180 | if (CONVERT_EXPR_CODE_P (code) |
5181 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 5182 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
5183 | new_stmt = gimple_build_assign (vec_dest, vop); |
5184 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5185 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5186 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5187 | if (slp_node) | |
9771b263 | 5188 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 5189 | } |
ebfd146a IR |
5190 | |
5191 | if (slp_node) | |
f18b55bd IR |
5192 | continue; |
5193 | ||
5194 | if (j == 0) | |
5195 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5196 | else | |
5197 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5198 | ||
5199 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5200 | } | |
b8698a0f | 5201 | |
9771b263 | 5202 | vec_oprnds.release (); |
ebfd146a IR |
5203 | return true; |
5204 | } | |
5205 | ||
9dc3f7de | 5206 | |
1107f3ae IR |
5207 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
5208 | either as shift by a scalar or by a vector. */ | |
5209 | ||
5210 | bool | |
5211 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
5212 | { | |
5213 | ||
ef4bddc2 | 5214 | machine_mode vec_mode; |
1107f3ae IR |
5215 | optab optab; |
5216 | int icode; | |
5217 | tree vectype; | |
5218 | ||
5219 | vectype = get_vectype_for_scalar_type (scalar_type); | |
5220 | if (!vectype) | |
5221 | return false; | |
5222 | ||
5223 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5224 | if (!optab | |
5225 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
5226 | { | |
5227 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5228 | if (!optab | |
5229 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
5230 | == CODE_FOR_nothing)) | |
5231 | return false; | |
5232 | } | |
5233 | ||
5234 | vec_mode = TYPE_MODE (vectype); | |
5235 | icode = (int) optab_handler (optab, vec_mode); | |
5236 | if (icode == CODE_FOR_nothing) | |
5237 | return false; | |
5238 | ||
5239 | return true; | |
5240 | } | |
5241 | ||
5242 | ||
9dc3f7de IR |
5243 | /* Function vectorizable_shift. |
5244 | ||
5245 | Check if STMT performs a shift operation that can be vectorized. | |
5246 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
5247 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
5248 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5249 | ||
5250 | static bool | |
355fe088 TS |
5251 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
5252 | gimple **vec_stmt, slp_tree slp_node) | |
9dc3f7de IR |
5253 | { |
5254 | tree vec_dest; | |
5255 | tree scalar_dest; | |
5256 | tree op0, op1 = NULL; | |
5257 | tree vec_oprnd1 = NULL_TREE; | |
5258 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5259 | tree vectype; | |
5260 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
5261 | enum tree_code code; | |
ef4bddc2 | 5262 | machine_mode vec_mode; |
9dc3f7de IR |
5263 | tree new_temp; |
5264 | optab optab; | |
5265 | int icode; | |
ef4bddc2 | 5266 | machine_mode optab_op2_mode; |
355fe088 | 5267 | gimple *def_stmt; |
9dc3f7de | 5268 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 5269 | int ndts = 2; |
355fe088 | 5270 | gimple *new_stmt = NULL; |
9dc3f7de | 5271 | stmt_vec_info prev_stmt_info; |
928686b1 RS |
5272 | poly_uint64 nunits_in; |
5273 | poly_uint64 nunits_out; | |
9dc3f7de | 5274 | tree vectype_out; |
cede2577 | 5275 | tree op1_vectype; |
9dc3f7de IR |
5276 | int ncopies; |
5277 | int j, i; | |
6e1aa848 DN |
5278 | vec<tree> vec_oprnds0 = vNULL; |
5279 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
5280 | tree vop0, vop1; |
5281 | unsigned int k; | |
49eab32e | 5282 | bool scalar_shift_arg = true; |
9dc3f7de | 5283 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5284 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
5285 | |
5286 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
5287 | return false; | |
5288 | ||
66c16fd9 RB |
5289 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5290 | && ! vec_stmt) | |
9dc3f7de IR |
5291 | return false; |
5292 | ||
5293 | /* Is STMT a vectorizable binary/unary operation? */ | |
5294 | if (!is_gimple_assign (stmt)) | |
5295 | return false; | |
5296 | ||
5297 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5298 | return false; | |
5299 | ||
5300 | code = gimple_assign_rhs_code (stmt); | |
5301 | ||
5302 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
5303 | || code == RROTATE_EXPR)) | |
5304 | return false; | |
5305 | ||
5306 | scalar_dest = gimple_assign_lhs (stmt); | |
5307 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2be65d9e | 5308 | if (!type_has_mode_precision_p (TREE_TYPE (scalar_dest))) |
7b7b1813 | 5309 | { |
73fbfcad | 5310 | if (dump_enabled_p ()) |
78c60e3d | 5311 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5312 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
5313 | return false; |
5314 | } | |
9dc3f7de IR |
5315 | |
5316 | op0 = gimple_assign_rhs1 (stmt); | |
81c40241 | 5317 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
9dc3f7de | 5318 | { |
73fbfcad | 5319 | if (dump_enabled_p ()) |
78c60e3d | 5320 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5321 | "use not simple.\n"); |
9dc3f7de IR |
5322 | return false; |
5323 | } | |
5324 | /* If op0 is an external or constant def use a vector type with | |
5325 | the same size as the output vector type. */ | |
5326 | if (!vectype) | |
5327 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5328 | if (vec_stmt) | |
5329 | gcc_assert (vectype); | |
5330 | if (!vectype) | |
5331 | { | |
73fbfcad | 5332 | if (dump_enabled_p ()) |
78c60e3d | 5333 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5334 | "no vectype for scalar type\n"); |
9dc3f7de IR |
5335 | return false; |
5336 | } | |
5337 | ||
5338 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5339 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
928686b1 | 5340 | if (maybe_ne (nunits_out, nunits_in)) |
9dc3f7de IR |
5341 | return false; |
5342 | ||
5343 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 5344 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &op1_vectype)) |
9dc3f7de | 5345 | { |
73fbfcad | 5346 | if (dump_enabled_p ()) |
78c60e3d | 5347 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5348 | "use not simple.\n"); |
9dc3f7de IR |
5349 | return false; |
5350 | } | |
5351 | ||
9dc3f7de IR |
5352 | /* Multiple types in SLP are handled by creating the appropriate number of |
5353 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5354 | case of SLP. */ | |
fce57248 | 5355 | if (slp_node) |
9dc3f7de IR |
5356 | ncopies = 1; |
5357 | else | |
e8f142e2 | 5358 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
9dc3f7de IR |
5359 | |
5360 | gcc_assert (ncopies >= 1); | |
5361 | ||
5362 | /* Determine whether the shift amount is a vector, or scalar. If the | |
5363 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
5364 | ||
dbfa87aa YR |
5365 | if ((dt[1] == vect_internal_def |
5366 | || dt[1] == vect_induction_def) | |
5367 | && !slp_node) | |
49eab32e JJ |
5368 | scalar_shift_arg = false; |
5369 | else if (dt[1] == vect_constant_def | |
5370 | || dt[1] == vect_external_def | |
5371 | || dt[1] == vect_internal_def) | |
5372 | { | |
5373 | /* In SLP, need to check whether the shift count is the same, | |
5374 | in loops if it is a constant or invariant, it is always | |
5375 | a scalar shift. */ | |
5376 | if (slp_node) | |
5377 | { | |
355fe088 TS |
5378 | vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
5379 | gimple *slpstmt; | |
49eab32e | 5380 | |
9771b263 | 5381 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
5382 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
5383 | scalar_shift_arg = false; | |
5384 | } | |
60d393e8 RB |
5385 | |
5386 | /* If the shift amount is computed by a pattern stmt we cannot | |
5387 | use the scalar amount directly thus give up and use a vector | |
5388 | shift. */ | |
5389 | if (dt[1] == vect_internal_def) | |
5390 | { | |
5391 | gimple *def = SSA_NAME_DEF_STMT (op1); | |
5392 | if (is_pattern_stmt_p (vinfo_for_stmt (def))) | |
5393 | scalar_shift_arg = false; | |
5394 | } | |
49eab32e JJ |
5395 | } |
5396 | else | |
5397 | { | |
73fbfcad | 5398 | if (dump_enabled_p ()) |
78c60e3d | 5399 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5400 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
5401 | return false; |
5402 | } | |
5403 | ||
9dc3f7de | 5404 | /* Vector shifted by vector. */ |
49eab32e | 5405 | if (!scalar_shift_arg) |
9dc3f7de IR |
5406 | { |
5407 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 5408 | if (dump_enabled_p ()) |
78c60e3d | 5409 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5410 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 5411 | |
aa948027 JJ |
5412 | if (!op1_vectype) |
5413 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
5414 | if (op1_vectype == NULL_TREE | |
5415 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 5416 | { |
73fbfcad | 5417 | if (dump_enabled_p ()) |
78c60e3d SS |
5418 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5419 | "unusable type for last operand in" | |
e645e942 | 5420 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
5421 | return false; |
5422 | } | |
9dc3f7de IR |
5423 | } |
5424 | /* See if the machine has a vector shifted by scalar insn and if not | |
5425 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 5426 | else |
9dc3f7de IR |
5427 | { |
5428 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5429 | if (optab | |
5430 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
5431 | { | |
73fbfcad | 5432 | if (dump_enabled_p ()) |
78c60e3d | 5433 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5434 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
5435 | } |
5436 | else | |
5437 | { | |
5438 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5439 | if (optab | |
5440 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
5441 | != CODE_FOR_nothing)) | |
5442 | { | |
49eab32e JJ |
5443 | scalar_shift_arg = false; |
5444 | ||
73fbfcad | 5445 | if (dump_enabled_p ()) |
78c60e3d | 5446 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5447 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
5448 | |
5449 | /* Unlike the other binary operators, shifts/rotates have | |
5450 | the rhs being int, instead of the same type as the lhs, | |
5451 | so make sure the scalar is the right type if we are | |
aa948027 | 5452 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
5453 | if (dt[1] == vect_constant_def) |
5454 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
5455 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
5456 | TREE_TYPE (op1))) | |
5457 | { | |
5458 | if (slp_node | |
5459 | && TYPE_MODE (TREE_TYPE (vectype)) | |
5460 | != TYPE_MODE (TREE_TYPE (op1))) | |
5461 | { | |
73fbfcad | 5462 | if (dump_enabled_p ()) |
78c60e3d SS |
5463 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5464 | "unusable type for last operand in" | |
e645e942 | 5465 | " vector/vector shift/rotate.\n"); |
21c0a521 | 5466 | return false; |
aa948027 JJ |
5467 | } |
5468 | if (vec_stmt && !slp_node) | |
5469 | { | |
5470 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
5471 | op1 = vect_init_vector (stmt, op1, | |
5472 | TREE_TYPE (vectype), NULL); | |
5473 | } | |
5474 | } | |
9dc3f7de IR |
5475 | } |
5476 | } | |
5477 | } | |
9dc3f7de IR |
5478 | |
5479 | /* Supportable by target? */ | |
5480 | if (!optab) | |
5481 | { | |
73fbfcad | 5482 | if (dump_enabled_p ()) |
78c60e3d | 5483 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5484 | "no optab.\n"); |
9dc3f7de IR |
5485 | return false; |
5486 | } | |
5487 | vec_mode = TYPE_MODE (vectype); | |
5488 | icode = (int) optab_handler (optab, vec_mode); | |
5489 | if (icode == CODE_FOR_nothing) | |
5490 | { | |
73fbfcad | 5491 | if (dump_enabled_p ()) |
78c60e3d | 5492 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5493 | "op not supported by target.\n"); |
9dc3f7de | 5494 | /* Check only during analysis. */ |
cf098191 | 5495 | if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) |
ca09abcb RS |
5496 | || (!vec_stmt |
5497 | && !vect_worthwhile_without_simd_p (vinfo, code))) | |
9dc3f7de | 5498 | return false; |
73fbfcad | 5499 | if (dump_enabled_p ()) |
e645e942 TJ |
5500 | dump_printf_loc (MSG_NOTE, vect_location, |
5501 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
5502 | } |
5503 | ||
5504 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
ca09abcb RS |
5505 | if (!vec_stmt |
5506 | && !VECTOR_MODE_P (TYPE_MODE (vectype)) | |
5507 | && !vect_worthwhile_without_simd_p (vinfo, code)) | |
9dc3f7de | 5508 | { |
73fbfcad | 5509 | if (dump_enabled_p ()) |
78c60e3d | 5510 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5511 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
5512 | return false; |
5513 | } | |
5514 | ||
5515 | if (!vec_stmt) /* transformation not required. */ | |
5516 | { | |
5517 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 5518 | if (dump_enabled_p ()) |
e645e942 TJ |
5519 | dump_printf_loc (MSG_NOTE, vect_location, |
5520 | "=== vectorizable_shift ===\n"); | |
78604de0 RB |
5521 | if (!slp_node) |
5522 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); | |
9dc3f7de IR |
5523 | return true; |
5524 | } | |
5525 | ||
67b8dbac | 5526 | /* Transform. */ |
9dc3f7de | 5527 | |
73fbfcad | 5528 | if (dump_enabled_p ()) |
78c60e3d | 5529 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5530 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
5531 | |
5532 | /* Handle def. */ | |
5533 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5534 | ||
9dc3f7de IR |
5535 | prev_stmt_info = NULL; |
5536 | for (j = 0; j < ncopies; j++) | |
5537 | { | |
5538 | /* Handle uses. */ | |
5539 | if (j == 0) | |
5540 | { | |
5541 | if (scalar_shift_arg) | |
5542 | { | |
5543 | /* Vector shl and shr insn patterns can be defined with scalar | |
5544 | operand 2 (shift operand). In this case, use constant or loop | |
5545 | invariant op1 directly, without extending it to vector mode | |
5546 | first. */ | |
5547 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
5548 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
5549 | { | |
73fbfcad | 5550 | if (dump_enabled_p ()) |
78c60e3d | 5551 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5552 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 5553 | vec_oprnd1 = op1; |
8930f723 | 5554 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 5555 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5556 | if (slp_node) |
5557 | { | |
5558 | /* Store vec_oprnd1 for every vector stmt to be created | |
5559 | for SLP_NODE. We check during the analysis that all | |
5560 | the shift arguments are the same. | |
5561 | TODO: Allow different constants for different vector | |
5562 | stmts generated for an SLP instance. */ | |
5563 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 5564 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5565 | } |
5566 | } | |
5567 | } | |
5568 | ||
5569 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
5570 | (a special case for certain kind of vector shifts); otherwise, | |
5571 | operand 1 should be of a vector type (the usual case). */ | |
5572 | if (vec_oprnd1) | |
5573 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5574 | slp_node); |
9dc3f7de IR |
5575 | else |
5576 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
306b0c92 | 5577 | slp_node); |
9dc3f7de IR |
5578 | } |
5579 | else | |
5580 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5581 | ||
5582 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5583 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 5584 | { |
9771b263 | 5585 | vop1 = vec_oprnds1[i]; |
0d0e4a03 | 5586 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
5587 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5588 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5589 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5590 | if (slp_node) | |
9771b263 | 5591 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
5592 | } |
5593 | ||
5594 | if (slp_node) | |
5595 | continue; | |
5596 | ||
5597 | if (j == 0) | |
5598 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5599 | else | |
5600 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5601 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5602 | } | |
5603 | ||
9771b263 DN |
5604 | vec_oprnds0.release (); |
5605 | vec_oprnds1.release (); | |
9dc3f7de IR |
5606 | |
5607 | return true; | |
5608 | } | |
5609 | ||
5610 | ||
ebfd146a IR |
5611 | /* Function vectorizable_operation. |
5612 | ||
16949072 RG |
5613 | Check if STMT performs a binary, unary or ternary operation that can |
5614 | be vectorized. | |
b8698a0f | 5615 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
5616 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5617 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5618 | ||
5619 | static bool | |
355fe088 TS |
5620 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
5621 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a | 5622 | { |
00f07b86 | 5623 | tree vec_dest; |
ebfd146a | 5624 | tree scalar_dest; |
16949072 | 5625 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 5626 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 5627 | tree vectype; |
ebfd146a | 5628 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
0eb952ea | 5629 | enum tree_code code, orig_code; |
ef4bddc2 | 5630 | machine_mode vec_mode; |
ebfd146a IR |
5631 | tree new_temp; |
5632 | int op_type; | |
00f07b86 | 5633 | optab optab; |
523ba738 | 5634 | bool target_support_p; |
355fe088 | 5635 | gimple *def_stmt; |
16949072 RG |
5636 | enum vect_def_type dt[3] |
5637 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 5638 | int ndts = 3; |
355fe088 | 5639 | gimple *new_stmt = NULL; |
ebfd146a | 5640 | stmt_vec_info prev_stmt_info; |
928686b1 RS |
5641 | poly_uint64 nunits_in; |
5642 | poly_uint64 nunits_out; | |
ebfd146a IR |
5643 | tree vectype_out; |
5644 | int ncopies; | |
5645 | int j, i; | |
6e1aa848 DN |
5646 | vec<tree> vec_oprnds0 = vNULL; |
5647 | vec<tree> vec_oprnds1 = vNULL; | |
5648 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 5649 | tree vop0, vop1, vop2; |
a70d6342 | 5650 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5651 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 5652 | |
a70d6342 | 5653 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5654 | return false; |
5655 | ||
66c16fd9 RB |
5656 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5657 | && ! vec_stmt) | |
ebfd146a IR |
5658 | return false; |
5659 | ||
5660 | /* Is STMT a vectorizable binary/unary operation? */ | |
5661 | if (!is_gimple_assign (stmt)) | |
5662 | return false; | |
5663 | ||
5664 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5665 | return false; | |
5666 | ||
0eb952ea | 5667 | orig_code = code = gimple_assign_rhs_code (stmt); |
ebfd146a | 5668 | |
1af4ebf5 MG |
5669 | /* For pointer addition and subtraction, we should use the normal |
5670 | plus and minus for the vector operation. */ | |
ebfd146a IR |
5671 | if (code == POINTER_PLUS_EXPR) |
5672 | code = PLUS_EXPR; | |
1af4ebf5 MG |
5673 | if (code == POINTER_DIFF_EXPR) |
5674 | code = MINUS_EXPR; | |
ebfd146a IR |
5675 | |
5676 | /* Support only unary or binary operations. */ | |
5677 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 5678 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 5679 | { |
73fbfcad | 5680 | if (dump_enabled_p ()) |
78c60e3d | 5681 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5682 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 5683 | op_type); |
ebfd146a IR |
5684 | return false; |
5685 | } | |
5686 | ||
b690cc0f RG |
5687 | scalar_dest = gimple_assign_lhs (stmt); |
5688 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
5689 | ||
7b7b1813 RG |
5690 | /* Most operations cannot handle bit-precision types without extra |
5691 | truncations. */ | |
045c1278 | 5692 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
2be65d9e | 5693 | && !type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
7b7b1813 RG |
5694 | /* Exception are bitwise binary operations. */ |
5695 | && code != BIT_IOR_EXPR | |
5696 | && code != BIT_XOR_EXPR | |
5697 | && code != BIT_AND_EXPR) | |
5698 | { | |
73fbfcad | 5699 | if (dump_enabled_p ()) |
78c60e3d | 5700 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5701 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
5702 | return false; |
5703 | } | |
5704 | ||
ebfd146a | 5705 | op0 = gimple_assign_rhs1 (stmt); |
81c40241 | 5706 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
ebfd146a | 5707 | { |
73fbfcad | 5708 | if (dump_enabled_p ()) |
78c60e3d | 5709 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5710 | "use not simple.\n"); |
ebfd146a IR |
5711 | return false; |
5712 | } | |
b690cc0f RG |
5713 | /* If op0 is an external or constant def use a vector type with |
5714 | the same size as the output vector type. */ | |
5715 | if (!vectype) | |
b036c6c5 IE |
5716 | { |
5717 | /* For boolean type we cannot determine vectype by | |
5718 | invariant value (don't know whether it is a vector | |
5719 | of booleans or vector of integers). We use output | |
5720 | vectype because operations on boolean don't change | |
5721 | type. */ | |
2568d8a1 | 5722 | if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op0))) |
b036c6c5 | 5723 | { |
2568d8a1 | 5724 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (scalar_dest))) |
b036c6c5 IE |
5725 | { |
5726 | if (dump_enabled_p ()) | |
5727 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5728 | "not supported operation on bool value.\n"); | |
5729 | return false; | |
5730 | } | |
5731 | vectype = vectype_out; | |
5732 | } | |
5733 | else | |
5734 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5735 | } | |
7d8930a0 IR |
5736 | if (vec_stmt) |
5737 | gcc_assert (vectype); | |
5738 | if (!vectype) | |
5739 | { | |
73fbfcad | 5740 | if (dump_enabled_p ()) |
7d8930a0 | 5741 | { |
78c60e3d SS |
5742 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5743 | "no vectype for scalar type "); | |
5744 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
5745 | TREE_TYPE (op0)); | |
e645e942 | 5746 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
5747 | } |
5748 | ||
5749 | return false; | |
5750 | } | |
b690cc0f RG |
5751 | |
5752 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5753 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
928686b1 | 5754 | if (maybe_ne (nunits_out, nunits_in)) |
b690cc0f | 5755 | return false; |
ebfd146a | 5756 | |
16949072 | 5757 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
5758 | { |
5759 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 5760 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1])) |
ebfd146a | 5761 | { |
73fbfcad | 5762 | if (dump_enabled_p ()) |
78c60e3d | 5763 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5764 | "use not simple.\n"); |
ebfd146a IR |
5765 | return false; |
5766 | } | |
5767 | } | |
16949072 RG |
5768 | if (op_type == ternary_op) |
5769 | { | |
5770 | op2 = gimple_assign_rhs3 (stmt); | |
81c40241 | 5771 | if (!vect_is_simple_use (op2, vinfo, &def_stmt, &dt[2])) |
16949072 | 5772 | { |
73fbfcad | 5773 | if (dump_enabled_p ()) |
78c60e3d | 5774 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5775 | "use not simple.\n"); |
16949072 RG |
5776 | return false; |
5777 | } | |
5778 | } | |
ebfd146a | 5779 | |
b690cc0f | 5780 | /* Multiple types in SLP are handled by creating the appropriate number of |
ff802fa1 | 5781 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 5782 | case of SLP. */ |
fce57248 | 5783 | if (slp_node) |
b690cc0f RG |
5784 | ncopies = 1; |
5785 | else | |
e8f142e2 | 5786 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
b690cc0f RG |
5787 | |
5788 | gcc_assert (ncopies >= 1); | |
5789 | ||
9dc3f7de | 5790 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
5791 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
5792 | || code == RROTATE_EXPR) | |
9dc3f7de | 5793 | return false; |
ebfd146a | 5794 | |
ebfd146a | 5795 | /* Supportable by target? */ |
00f07b86 RH |
5796 | |
5797 | vec_mode = TYPE_MODE (vectype); | |
5798 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 5799 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
5800 | else |
5801 | { | |
5802 | optab = optab_for_tree_code (code, vectype, optab_default); | |
5803 | if (!optab) | |
5deb57cb | 5804 | { |
73fbfcad | 5805 | if (dump_enabled_p ()) |
78c60e3d | 5806 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5807 | "no optab.\n"); |
00f07b86 | 5808 | return false; |
5deb57cb | 5809 | } |
523ba738 RS |
5810 | target_support_p = (optab_handler (optab, vec_mode) |
5811 | != CODE_FOR_nothing); | |
5deb57cb JJ |
5812 | } |
5813 | ||
523ba738 | 5814 | if (!target_support_p) |
ebfd146a | 5815 | { |
73fbfcad | 5816 | if (dump_enabled_p ()) |
78c60e3d | 5817 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5818 | "op not supported by target.\n"); |
ebfd146a | 5819 | /* Check only during analysis. */ |
cf098191 | 5820 | if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) |
ca09abcb | 5821 | || (!vec_stmt && !vect_worthwhile_without_simd_p (vinfo, code))) |
ebfd146a | 5822 | return false; |
73fbfcad | 5823 | if (dump_enabled_p ()) |
e645e942 TJ |
5824 | dump_printf_loc (MSG_NOTE, vect_location, |
5825 | "proceeding using word mode.\n"); | |
383d9c83 IR |
5826 | } |
5827 | ||
4a00c761 | 5828 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
5829 | if (!VECTOR_MODE_P (vec_mode) |
5830 | && !vec_stmt | |
ca09abcb | 5831 | && !vect_worthwhile_without_simd_p (vinfo, code)) |
7d8930a0 | 5832 | { |
73fbfcad | 5833 | if (dump_enabled_p ()) |
78c60e3d | 5834 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5835 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 5836 | return false; |
7d8930a0 | 5837 | } |
ebfd146a | 5838 | |
ebfd146a IR |
5839 | if (!vec_stmt) /* transformation not required. */ |
5840 | { | |
4a00c761 | 5841 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 5842 | if (dump_enabled_p ()) |
78c60e3d | 5843 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5844 | "=== vectorizable_operation ===\n"); |
78604de0 RB |
5845 | if (!slp_node) |
5846 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); | |
ebfd146a IR |
5847 | return true; |
5848 | } | |
5849 | ||
67b8dbac | 5850 | /* Transform. */ |
ebfd146a | 5851 | |
73fbfcad | 5852 | if (dump_enabled_p ()) |
78c60e3d | 5853 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5854 | "transform binary/unary operation.\n"); |
383d9c83 | 5855 | |
ebfd146a | 5856 | /* Handle def. */ |
00f07b86 | 5857 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 5858 | |
0eb952ea JJ |
5859 | /* POINTER_DIFF_EXPR has pointer arguments which are vectorized as |
5860 | vectors with unsigned elements, but the result is signed. So, we | |
5861 | need to compute the MINUS_EXPR into vectype temporary and | |
5862 | VIEW_CONVERT_EXPR it into the final vectype_out result. */ | |
5863 | tree vec_cvt_dest = NULL_TREE; | |
5864 | if (orig_code == POINTER_DIFF_EXPR) | |
5865 | vec_cvt_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
5866 | ||
ebfd146a IR |
5867 | /* In case the vectorization factor (VF) is bigger than the number |
5868 | of elements that we can fit in a vectype (nunits), we have to generate | |
5869 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
5870 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
5871 | from one copy of the vector stmt to the next, in the field | |
5872 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
5873 | stages to find the correct vector defs to be used when vectorizing | |
5874 | stmts that use the defs of the current stmt. The example below | |
5875 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
5876 | we need to create 4 vectorized stmts): | |
5877 | ||
5878 | before vectorization: | |
5879 | RELATED_STMT VEC_STMT | |
5880 | S1: x = memref - - | |
5881 | S2: z = x + 1 - - | |
5882 | ||
5883 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
5884 | there): | |
5885 | RELATED_STMT VEC_STMT | |
5886 | VS1_0: vx0 = memref0 VS1_1 - | |
5887 | VS1_1: vx1 = memref1 VS1_2 - | |
5888 | VS1_2: vx2 = memref2 VS1_3 - | |
5889 | VS1_3: vx3 = memref3 - - | |
5890 | S1: x = load - VS1_0 | |
5891 | S2: z = x + 1 - - | |
5892 | ||
5893 | step2: vectorize stmt S2 (done here): | |
5894 | To vectorize stmt S2 we first need to find the relevant vector | |
5895 | def for the first operand 'x'. This is, as usual, obtained from | |
5896 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
5897 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
5898 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
5899 | the vector stmt VS2_0, and as usual, record it in the | |
5900 | STMT_VINFO_VEC_STMT of stmt S2. | |
5901 | When creating the second copy (VS2_1), we obtain the relevant vector | |
5902 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
5903 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
5904 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
5905 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
5906 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
5907 | chain of stmts and pointers: | |
5908 | RELATED_STMT VEC_STMT | |
5909 | VS1_0: vx0 = memref0 VS1_1 - | |
5910 | VS1_1: vx1 = memref1 VS1_2 - | |
5911 | VS1_2: vx2 = memref2 VS1_3 - | |
5912 | VS1_3: vx3 = memref3 - - | |
5913 | S1: x = load - VS1_0 | |
5914 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
5915 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
5916 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
5917 | VS2_3: vz3 = vx3 + v1 - - | |
5918 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
5919 | |
5920 | prev_stmt_info = NULL; | |
5921 | for (j = 0; j < ncopies; j++) | |
5922 | { | |
5923 | /* Handle uses. */ | |
5924 | if (j == 0) | |
4a00c761 JJ |
5925 | { |
5926 | if (op_type == binary_op || op_type == ternary_op) | |
5927 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
306b0c92 | 5928 | slp_node); |
4a00c761 JJ |
5929 | else |
5930 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5931 | slp_node); |
4a00c761 | 5932 | if (op_type == ternary_op) |
c392943c | 5933 | vect_get_vec_defs (op2, NULL_TREE, stmt, &vec_oprnds2, NULL, |
306b0c92 | 5934 | slp_node); |
4a00c761 | 5935 | } |
ebfd146a | 5936 | else |
4a00c761 JJ |
5937 | { |
5938 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5939 | if (op_type == ternary_op) | |
5940 | { | |
9771b263 DN |
5941 | tree vec_oprnd = vec_oprnds2.pop (); |
5942 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
5943 | vec_oprnd)); | |
4a00c761 JJ |
5944 | } |
5945 | } | |
5946 | ||
5947 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5948 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 5949 | { |
4a00c761 | 5950 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 5951 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 5952 | vop2 = ((op_type == ternary_op) |
9771b263 | 5953 | ? vec_oprnds2[i] : NULL_TREE); |
0d0e4a03 | 5954 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
4a00c761 JJ |
5955 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5956 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5957 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
0eb952ea JJ |
5958 | if (vec_cvt_dest) |
5959 | { | |
5960 | new_temp = build1 (VIEW_CONVERT_EXPR, vectype_out, new_temp); | |
5961 | new_stmt = gimple_build_assign (vec_cvt_dest, VIEW_CONVERT_EXPR, | |
5962 | new_temp); | |
5963 | new_temp = make_ssa_name (vec_cvt_dest, new_stmt); | |
5964 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5965 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5966 | } | |
4a00c761 | 5967 | if (slp_node) |
9771b263 | 5968 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
5969 | } |
5970 | ||
4a00c761 JJ |
5971 | if (slp_node) |
5972 | continue; | |
5973 | ||
5974 | if (j == 0) | |
5975 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5976 | else | |
5977 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5978 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
5979 | } |
5980 | ||
9771b263 DN |
5981 | vec_oprnds0.release (); |
5982 | vec_oprnds1.release (); | |
5983 | vec_oprnds2.release (); | |
ebfd146a | 5984 | |
ebfd146a IR |
5985 | return true; |
5986 | } | |
5987 | ||
f702e7d4 | 5988 | /* A helper function to ensure data reference DR's base alignment. */ |
c716e67f XDL |
5989 | |
5990 | static void | |
f702e7d4 | 5991 | ensure_base_align (struct data_reference *dr) |
c716e67f XDL |
5992 | { |
5993 | if (!dr->aux) | |
5994 | return; | |
5995 | ||
52639a61 | 5996 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f | 5997 | { |
52639a61 | 5998 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 5999 | |
f702e7d4 RS |
6000 | unsigned int align_base_to = DR_TARGET_ALIGNMENT (dr) * BITS_PER_UNIT; |
6001 | ||
428f0c67 | 6002 | if (decl_in_symtab_p (base_decl)) |
f702e7d4 | 6003 | symtab_node::get (base_decl)->increase_alignment (align_base_to); |
428f0c67 JH |
6004 | else |
6005 | { | |
f702e7d4 | 6006 | SET_DECL_ALIGN (base_decl, align_base_to); |
428f0c67 JH |
6007 | DECL_USER_ALIGN (base_decl) = 1; |
6008 | } | |
52639a61 | 6009 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
6010 | } |
6011 | } | |
6012 | ||
ebfd146a | 6013 | |
44fc7854 BE |
6014 | /* Function get_group_alias_ptr_type. |
6015 | ||
6016 | Return the alias type for the group starting at FIRST_STMT. */ | |
6017 | ||
6018 | static tree | |
6019 | get_group_alias_ptr_type (gimple *first_stmt) | |
6020 | { | |
6021 | struct data_reference *first_dr, *next_dr; | |
6022 | gimple *next_stmt; | |
6023 | ||
6024 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
6025 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (first_stmt)); | |
6026 | while (next_stmt) | |
6027 | { | |
6028 | next_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (next_stmt)); | |
6029 | if (get_alias_set (DR_REF (first_dr)) | |
6030 | != get_alias_set (DR_REF (next_dr))) | |
6031 | { | |
6032 | if (dump_enabled_p ()) | |
6033 | dump_printf_loc (MSG_NOTE, vect_location, | |
6034 | "conflicting alias set types.\n"); | |
6035 | return ptr_type_node; | |
6036 | } | |
6037 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); | |
6038 | } | |
6039 | return reference_alias_ptr_type (DR_REF (first_dr)); | |
6040 | } | |
6041 | ||
6042 | ||
ebfd146a IR |
6043 | /* Function vectorizable_store. |
6044 | ||
b8698a0f L |
6045 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
6046 | can be vectorized. | |
6047 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6048 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6049 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6050 | ||
6051 | static bool | |
355fe088 | 6052 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 6053 | slp_tree slp_node) |
ebfd146a | 6054 | { |
ebfd146a IR |
6055 | tree data_ref; |
6056 | tree op; | |
6057 | tree vec_oprnd = NULL_TREE; | |
6058 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6059 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 6060 | tree elem_type; |
ebfd146a | 6061 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6062 | struct loop *loop = NULL; |
ef4bddc2 | 6063 | machine_mode vec_mode; |
ebfd146a IR |
6064 | tree dummy; |
6065 | enum dr_alignment_support alignment_support_scheme; | |
355fe088 | 6066 | gimple *def_stmt; |
929b4411 RS |
6067 | enum vect_def_type rhs_dt = vect_unknown_def_type; |
6068 | enum vect_def_type mask_dt = vect_unknown_def_type; | |
ebfd146a IR |
6069 | stmt_vec_info prev_stmt_info = NULL; |
6070 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6071 | tree dataref_offset = NULL_TREE; |
355fe088 | 6072 | gimple *ptr_incr = NULL; |
ebfd146a IR |
6073 | int ncopies; |
6074 | int j; | |
2de001ee RS |
6075 | gimple *next_stmt, *first_stmt; |
6076 | bool grouped_store; | |
ebfd146a | 6077 | unsigned int group_size, i; |
6e1aa848 DN |
6078 | vec<tree> oprnds = vNULL; |
6079 | vec<tree> result_chain = vNULL; | |
ebfd146a | 6080 | bool inv_p; |
09dfa495 | 6081 | tree offset = NULL_TREE; |
6e1aa848 | 6082 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 6083 | bool slp = (slp_node != NULL); |
ebfd146a | 6084 | unsigned int vec_num; |
a70d6342 | 6085 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 6086 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 6087 | tree aggr_type; |
134c85ca | 6088 | gather_scatter_info gs_info; |
355fe088 | 6089 | gimple *new_stmt; |
d9f21f6a | 6090 | poly_uint64 vf; |
2de001ee | 6091 | vec_load_store_type vls_type; |
44fc7854 | 6092 | tree ref_type; |
a70d6342 | 6093 | |
a70d6342 | 6094 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
6095 | return false; |
6096 | ||
66c16fd9 RB |
6097 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6098 | && ! vec_stmt) | |
ebfd146a IR |
6099 | return false; |
6100 | ||
6101 | /* Is vectorizable store? */ | |
6102 | ||
c3a8f964 RS |
6103 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
6104 | if (is_gimple_assign (stmt)) | |
6105 | { | |
6106 | tree scalar_dest = gimple_assign_lhs (stmt); | |
6107 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR | |
6108 | && is_pattern_stmt_p (stmt_info)) | |
6109 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
6110 | if (TREE_CODE (scalar_dest) != ARRAY_REF | |
6111 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF | |
6112 | && TREE_CODE (scalar_dest) != INDIRECT_REF | |
6113 | && TREE_CODE (scalar_dest) != COMPONENT_REF | |
6114 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
6115 | && TREE_CODE (scalar_dest) != REALPART_EXPR | |
6116 | && TREE_CODE (scalar_dest) != MEM_REF) | |
6117 | return false; | |
6118 | } | |
6119 | else | |
6120 | { | |
6121 | gcall *call = dyn_cast <gcall *> (stmt); | |
f307441a RS |
6122 | if (!call || !gimple_call_internal_p (call)) |
6123 | return false; | |
6124 | ||
6125 | internal_fn ifn = gimple_call_internal_fn (call); | |
6126 | if (!internal_store_fn_p (ifn)) | |
c3a8f964 | 6127 | return false; |
ebfd146a | 6128 | |
c3a8f964 RS |
6129 | if (slp_node != NULL) |
6130 | { | |
6131 | if (dump_enabled_p ()) | |
6132 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6133 | "SLP of masked stores not supported.\n"); | |
6134 | return false; | |
6135 | } | |
6136 | ||
f307441a RS |
6137 | int mask_index = internal_fn_mask_index (ifn); |
6138 | if (mask_index >= 0) | |
6139 | { | |
6140 | mask = gimple_call_arg (call, mask_index); | |
929b4411 RS |
6141 | if (!vect_check_load_store_mask (stmt, mask, &mask_dt, |
6142 | &mask_vectype)) | |
f307441a RS |
6143 | return false; |
6144 | } | |
c3a8f964 RS |
6145 | } |
6146 | ||
6147 | op = vect_get_store_rhs (stmt); | |
ebfd146a | 6148 | |
fce57248 RS |
6149 | /* Cannot have hybrid store SLP -- that would mean storing to the |
6150 | same location twice. */ | |
6151 | gcc_assert (slp == PURE_SLP_STMT (stmt_info)); | |
6152 | ||
f4d09712 | 6153 | tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE; |
4d694b27 | 6154 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 JJ |
6155 | |
6156 | if (loop_vinfo) | |
b17dc4d4 RB |
6157 | { |
6158 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6159 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6160 | } | |
6161 | else | |
6162 | vf = 1; | |
465c8c19 JJ |
6163 | |
6164 | /* Multiple types in SLP are handled by creating the appropriate number of | |
6165 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
6166 | case of SLP. */ | |
fce57248 | 6167 | if (slp) |
465c8c19 JJ |
6168 | ncopies = 1; |
6169 | else | |
e8f142e2 | 6170 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
6171 | |
6172 | gcc_assert (ncopies >= 1); | |
6173 | ||
6174 | /* FORNOW. This restriction should be relaxed. */ | |
6175 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
6176 | { | |
6177 | if (dump_enabled_p ()) | |
6178 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6179 | "multiple types in nested loop.\n"); | |
6180 | return false; | |
6181 | } | |
6182 | ||
929b4411 | 6183 | if (!vect_check_store_rhs (stmt, op, &rhs_dt, &rhs_vectype, &vls_type)) |
f4d09712 KY |
6184 | return false; |
6185 | ||
272c6793 | 6186 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 6187 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 6188 | |
ebfd146a IR |
6189 | if (!STMT_VINFO_DATA_REF (stmt_info)) |
6190 | return false; | |
6191 | ||
2de001ee | 6192 | vect_memory_access_type memory_access_type; |
7e11fc7f | 6193 | if (!get_load_store_type (stmt, vectype, slp, mask, vls_type, ncopies, |
2de001ee RS |
6194 | &memory_access_type, &gs_info)) |
6195 | return false; | |
3bab6342 | 6196 | |
c3a8f964 RS |
6197 | if (mask) |
6198 | { | |
7e11fc7f RS |
6199 | if (memory_access_type == VMAT_CONTIGUOUS) |
6200 | { | |
6201 | if (!VECTOR_MODE_P (vec_mode) | |
6202 | || !can_vec_mask_load_store_p (vec_mode, | |
6203 | TYPE_MODE (mask_vectype), false)) | |
6204 | return false; | |
6205 | } | |
f307441a RS |
6206 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
6207 | && (memory_access_type != VMAT_GATHER_SCATTER || gs_info.decl)) | |
c3a8f964 RS |
6208 | { |
6209 | if (dump_enabled_p ()) | |
6210 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6211 | "unsupported access type for masked store.\n"); | |
6212 | return false; | |
6213 | } | |
c3a8f964 RS |
6214 | } |
6215 | else | |
6216 | { | |
6217 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6218 | (e.g. - array initialization with 0). */ | |
6219 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) | |
6220 | return false; | |
6221 | } | |
6222 | ||
f307441a | 6223 | grouped_store = (STMT_VINFO_GROUPED_ACCESS (stmt_info) |
b5ec4de7 RS |
6224 | && memory_access_type != VMAT_GATHER_SCATTER |
6225 | && (slp || memory_access_type != VMAT_CONTIGUOUS)); | |
7cfb4d93 RS |
6226 | if (grouped_store) |
6227 | { | |
6228 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); | |
6229 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
6230 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
6231 | } | |
6232 | else | |
6233 | { | |
6234 | first_stmt = stmt; | |
6235 | first_dr = dr; | |
6236 | group_size = vec_num = 1; | |
6237 | } | |
6238 | ||
ebfd146a IR |
6239 | if (!vec_stmt) /* transformation not required. */ |
6240 | { | |
2de001ee | 6241 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
7cfb4d93 RS |
6242 | |
6243 | if (loop_vinfo | |
6244 | && LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) | |
6245 | check_load_store_masking (loop_vinfo, vectype, vls_type, group_size, | |
bfaa08b7 | 6246 | memory_access_type, &gs_info); |
7cfb4d93 | 6247 | |
ebfd146a | 6248 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; |
2e8ab70c | 6249 | /* The SLP costs are calculated during SLP analysis. */ |
78604de0 | 6250 | if (!slp_node) |
9ce4345a RS |
6251 | vect_model_store_cost (stmt_info, ncopies, memory_access_type, |
6252 | vls_type, NULL, NULL, NULL); | |
ebfd146a IR |
6253 | return true; |
6254 | } | |
2de001ee | 6255 | gcc_assert (memory_access_type == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
ebfd146a | 6256 | |
67b8dbac | 6257 | /* Transform. */ |
ebfd146a | 6258 | |
f702e7d4 | 6259 | ensure_base_align (dr); |
c716e67f | 6260 | |
f307441a | 6261 | if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
3bab6342 | 6262 | { |
c3a8f964 | 6263 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, src; |
134c85ca | 6264 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
6265 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; |
6266 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
6267 | edge pe = loop_preheader_edge (loop); | |
6268 | gimple_seq seq; | |
6269 | basic_block new_bb; | |
6270 | enum { NARROW, NONE, WIDEN } modifier; | |
4d694b27 RS |
6271 | poly_uint64 scatter_off_nunits |
6272 | = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype); | |
3bab6342 | 6273 | |
4d694b27 | 6274 | if (known_eq (nunits, scatter_off_nunits)) |
3bab6342 | 6275 | modifier = NONE; |
4d694b27 | 6276 | else if (known_eq (nunits * 2, scatter_off_nunits)) |
3bab6342 | 6277 | { |
3bab6342 AT |
6278 | modifier = WIDEN; |
6279 | ||
4d694b27 RS |
6280 | /* Currently gathers and scatters are only supported for |
6281 | fixed-length vectors. */ | |
6282 | unsigned int count = scatter_off_nunits.to_constant (); | |
6283 | vec_perm_builder sel (count, count, 1); | |
6284 | for (i = 0; i < (unsigned int) count; ++i) | |
6285 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 6286 | |
4d694b27 | 6287 | vec_perm_indices indices (sel, 1, count); |
e3342de4 RS |
6288 | perm_mask = vect_gen_perm_mask_checked (gs_info.offset_vectype, |
6289 | indices); | |
3bab6342 AT |
6290 | gcc_assert (perm_mask != NULL_TREE); |
6291 | } | |
4d694b27 | 6292 | else if (known_eq (nunits, scatter_off_nunits * 2)) |
3bab6342 | 6293 | { |
3bab6342 AT |
6294 | modifier = NARROW; |
6295 | ||
4d694b27 RS |
6296 | /* Currently gathers and scatters are only supported for |
6297 | fixed-length vectors. */ | |
6298 | unsigned int count = nunits.to_constant (); | |
6299 | vec_perm_builder sel (count, count, 1); | |
6300 | for (i = 0; i < (unsigned int) count; ++i) | |
6301 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 6302 | |
4d694b27 | 6303 | vec_perm_indices indices (sel, 2, count); |
e3342de4 | 6304 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); |
3bab6342 AT |
6305 | gcc_assert (perm_mask != NULL_TREE); |
6306 | ncopies *= 2; | |
6307 | } | |
6308 | else | |
6309 | gcc_unreachable (); | |
6310 | ||
134c85ca | 6311 | rettype = TREE_TYPE (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
6312 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
6313 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6314 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6315 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6316 | scaletype = TREE_VALUE (arglist); | |
6317 | ||
6318 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
6319 | && TREE_CODE (rettype) == VOID_TYPE); | |
6320 | ||
134c85ca | 6321 | ptr = fold_convert (ptrtype, gs_info.base); |
3bab6342 AT |
6322 | if (!is_gimple_min_invariant (ptr)) |
6323 | { | |
6324 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6325 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6326 | gcc_assert (!new_bb); | |
6327 | } | |
6328 | ||
6329 | /* Currently we support only unconditional scatter stores, | |
6330 | so mask should be all ones. */ | |
6331 | mask = build_int_cst (masktype, -1); | |
6332 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
6333 | ||
134c85ca | 6334 | scale = build_int_cst (scaletype, gs_info.scale); |
3bab6342 AT |
6335 | |
6336 | prev_stmt_info = NULL; | |
6337 | for (j = 0; j < ncopies; ++j) | |
6338 | { | |
6339 | if (j == 0) | |
6340 | { | |
6341 | src = vec_oprnd1 | |
c3a8f964 | 6342 | = vect_get_vec_def_for_operand (op, stmt); |
3bab6342 | 6343 | op = vec_oprnd0 |
134c85ca | 6344 | = vect_get_vec_def_for_operand (gs_info.offset, stmt); |
3bab6342 AT |
6345 | } |
6346 | else if (modifier != NONE && (j & 1)) | |
6347 | { | |
6348 | if (modifier == WIDEN) | |
6349 | { | |
6350 | src = vec_oprnd1 | |
929b4411 | 6351 | = vect_get_vec_def_for_stmt_copy (rhs_dt, vec_oprnd1); |
3bab6342 AT |
6352 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, |
6353 | stmt, gsi); | |
6354 | } | |
6355 | else if (modifier == NARROW) | |
6356 | { | |
6357 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
6358 | stmt, gsi); | |
6359 | op = vec_oprnd0 | |
134c85ca RS |
6360 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
6361 | vec_oprnd0); | |
3bab6342 AT |
6362 | } |
6363 | else | |
6364 | gcc_unreachable (); | |
6365 | } | |
6366 | else | |
6367 | { | |
6368 | src = vec_oprnd1 | |
929b4411 | 6369 | = vect_get_vec_def_for_stmt_copy (rhs_dt, vec_oprnd1); |
3bab6342 | 6370 | op = vec_oprnd0 |
134c85ca RS |
6371 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
6372 | vec_oprnd0); | |
3bab6342 AT |
6373 | } |
6374 | ||
6375 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
6376 | { | |
928686b1 RS |
6377 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)), |
6378 | TYPE_VECTOR_SUBPARTS (srctype))); | |
0e22bb5a | 6379 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 AT |
6380 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
6381 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
6382 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6383 | src = var; | |
6384 | } | |
6385 | ||
6386 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6387 | { | |
928686b1 RS |
6388 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), |
6389 | TYPE_VECTOR_SUBPARTS (idxtype))); | |
0e22bb5a | 6390 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 AT |
6391 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
6392 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
6393 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6394 | op = var; | |
6395 | } | |
6396 | ||
6397 | new_stmt | |
134c85ca | 6398 | = gimple_build_call (gs_info.decl, 5, ptr, mask, op, src, scale); |
3bab6342 AT |
6399 | |
6400 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6401 | ||
6402 | if (prev_stmt_info == NULL) | |
6403 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6404 | else | |
6405 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6406 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6407 | } | |
6408 | return true; | |
6409 | } | |
6410 | ||
f307441a | 6411 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 6412 | { |
f307441a RS |
6413 | gimple *group_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
6414 | GROUP_STORE_COUNT (vinfo_for_stmt (group_stmt))++; | |
6415 | } | |
ebfd146a | 6416 | |
f307441a RS |
6417 | if (grouped_store) |
6418 | { | |
ebfd146a | 6419 | /* FORNOW */ |
a70d6342 | 6420 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
6421 | |
6422 | /* We vectorize all the stmts of the interleaving group when we | |
6423 | reach the last stmt in the group. */ | |
e14c1050 IR |
6424 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
6425 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
6426 | && !slp) |
6427 | { | |
6428 | *vec_stmt = NULL; | |
6429 | return true; | |
6430 | } | |
6431 | ||
6432 | if (slp) | |
4b5caab7 | 6433 | { |
0d0293ac | 6434 | grouped_store = false; |
4b5caab7 IR |
6435 | /* VEC_NUM is the number of vect stmts to be created for this |
6436 | group. */ | |
6437 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 6438 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
52eab378 | 6439 | gcc_assert (GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt)) == first_stmt); |
4b5caab7 | 6440 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
c3a8f964 | 6441 | op = vect_get_store_rhs (first_stmt); |
4b5caab7 | 6442 | } |
ebfd146a | 6443 | else |
4b5caab7 IR |
6444 | /* VEC_NUM is the number of vect stmts to be created for this |
6445 | group. */ | |
ebfd146a | 6446 | vec_num = group_size; |
44fc7854 BE |
6447 | |
6448 | ref_type = get_group_alias_ptr_type (first_stmt); | |
ebfd146a | 6449 | } |
b8698a0f | 6450 | else |
7cfb4d93 | 6451 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
b8698a0f | 6452 | |
73fbfcad | 6453 | if (dump_enabled_p ()) |
78c60e3d | 6454 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6455 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 6456 | |
2de001ee RS |
6457 | if (memory_access_type == VMAT_ELEMENTWISE |
6458 | || memory_access_type == VMAT_STRIDED_SLP) | |
f2e2a985 MM |
6459 | { |
6460 | gimple_stmt_iterator incr_gsi; | |
6461 | bool insert_after; | |
355fe088 | 6462 | gimple *incr; |
f2e2a985 MM |
6463 | tree offvar; |
6464 | tree ivstep; | |
6465 | tree running_off; | |
f2e2a985 MM |
6466 | tree stride_base, stride_step, alias_off; |
6467 | tree vec_oprnd; | |
f502d50e | 6468 | unsigned int g; |
4d694b27 RS |
6469 | /* Checked by get_load_store_type. */ |
6470 | unsigned int const_nunits = nunits.to_constant (); | |
f2e2a985 | 6471 | |
7cfb4d93 | 6472 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
f2e2a985 MM |
6473 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); |
6474 | ||
6475 | stride_base | |
6476 | = fold_build_pointer_plus | |
b210f45f | 6477 | (DR_BASE_ADDRESS (first_dr), |
f2e2a985 | 6478 | size_binop (PLUS_EXPR, |
b210f45f | 6479 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
44fc7854 | 6480 | convert_to_ptrofftype (DR_INIT (first_dr)))); |
b210f45f | 6481 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); |
f2e2a985 MM |
6482 | |
6483 | /* For a store with loop-invariant (but other than power-of-2) | |
6484 | stride (i.e. not a grouped access) like so: | |
6485 | ||
6486 | for (i = 0; i < n; i += stride) | |
6487 | array[i] = ...; | |
6488 | ||
6489 | we generate a new induction variable and new stores from | |
6490 | the components of the (vectorized) rhs: | |
6491 | ||
6492 | for (j = 0; ; j += VF*stride) | |
6493 | vectemp = ...; | |
6494 | tmp1 = vectemp[0]; | |
6495 | array[j] = tmp1; | |
6496 | tmp2 = vectemp[1]; | |
6497 | array[j + stride] = tmp2; | |
6498 | ... | |
6499 | */ | |
6500 | ||
4d694b27 | 6501 | unsigned nstores = const_nunits; |
b17dc4d4 | 6502 | unsigned lnel = 1; |
cee62fee | 6503 | tree ltype = elem_type; |
04199738 | 6504 | tree lvectype = vectype; |
cee62fee MM |
6505 | if (slp) |
6506 | { | |
4d694b27 RS |
6507 | if (group_size < const_nunits |
6508 | && const_nunits % group_size == 0) | |
b17dc4d4 | 6509 | { |
4d694b27 | 6510 | nstores = const_nunits / group_size; |
b17dc4d4 RB |
6511 | lnel = group_size; |
6512 | ltype = build_vector_type (elem_type, group_size); | |
04199738 RB |
6513 | lvectype = vectype; |
6514 | ||
6515 | /* First check if vec_extract optab doesn't support extraction | |
6516 | of vector elts directly. */ | |
b397965c | 6517 | scalar_mode elmode = SCALAR_TYPE_MODE (elem_type); |
9da15d40 RS |
6518 | machine_mode vmode; |
6519 | if (!mode_for_vector (elmode, group_size).exists (&vmode) | |
6520 | || !VECTOR_MODE_P (vmode) | |
414fef4e | 6521 | || !targetm.vector_mode_supported_p (vmode) |
04199738 RB |
6522 | || (convert_optab_handler (vec_extract_optab, |
6523 | TYPE_MODE (vectype), vmode) | |
6524 | == CODE_FOR_nothing)) | |
6525 | { | |
6526 | /* Try to avoid emitting an extract of vector elements | |
6527 | by performing the extracts using an integer type of the | |
6528 | same size, extracting from a vector of those and then | |
6529 | re-interpreting it as the original vector type if | |
6530 | supported. */ | |
6531 | unsigned lsize | |
6532 | = group_size * GET_MODE_BITSIZE (elmode); | |
fffbab82 | 6533 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 6534 | unsigned int lnunits = const_nunits / group_size; |
04199738 RB |
6535 | /* If we can't construct such a vector fall back to |
6536 | element extracts from the original vector type and | |
6537 | element size stores. */ | |
4d694b27 | 6538 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 6539 | && VECTOR_MODE_P (vmode) |
414fef4e | 6540 | && targetm.vector_mode_supported_p (vmode) |
04199738 RB |
6541 | && (convert_optab_handler (vec_extract_optab, |
6542 | vmode, elmode) | |
6543 | != CODE_FOR_nothing)) | |
6544 | { | |
4d694b27 | 6545 | nstores = lnunits; |
04199738 RB |
6546 | lnel = group_size; |
6547 | ltype = build_nonstandard_integer_type (lsize, 1); | |
6548 | lvectype = build_vector_type (ltype, nstores); | |
6549 | } | |
6550 | /* Else fall back to vector extraction anyway. | |
6551 | Fewer stores are more important than avoiding spilling | |
6552 | of the vector we extract from. Compared to the | |
6553 | construction case in vectorizable_load no store-forwarding | |
6554 | issue exists here for reasonable archs. */ | |
6555 | } | |
b17dc4d4 | 6556 | } |
4d694b27 RS |
6557 | else if (group_size >= const_nunits |
6558 | && group_size % const_nunits == 0) | |
b17dc4d4 RB |
6559 | { |
6560 | nstores = 1; | |
4d694b27 | 6561 | lnel = const_nunits; |
b17dc4d4 | 6562 | ltype = vectype; |
04199738 | 6563 | lvectype = vectype; |
b17dc4d4 | 6564 | } |
cee62fee MM |
6565 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); |
6566 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
6567 | } | |
6568 | ||
f2e2a985 MM |
6569 | ivstep = stride_step; |
6570 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
b17dc4d4 | 6571 | build_int_cst (TREE_TYPE (ivstep), vf)); |
f2e2a985 MM |
6572 | |
6573 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6574 | ||
b210f45f RB |
6575 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
6576 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); | |
f2e2a985 MM |
6577 | create_iv (stride_base, ivstep, NULL, |
6578 | loop, &incr_gsi, insert_after, | |
6579 | &offvar, NULL); | |
6580 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 6581 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
f2e2a985 | 6582 | |
b210f45f | 6583 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
f2e2a985 MM |
6584 | |
6585 | prev_stmt_info = NULL; | |
44fc7854 | 6586 | alias_off = build_int_cst (ref_type, 0); |
f502d50e MM |
6587 | next_stmt = first_stmt; |
6588 | for (g = 0; g < group_size; g++) | |
f2e2a985 | 6589 | { |
f502d50e MM |
6590 | running_off = offvar; |
6591 | if (g) | |
f2e2a985 | 6592 | { |
f502d50e MM |
6593 | tree size = TYPE_SIZE_UNIT (ltype); |
6594 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 6595 | size); |
f502d50e | 6596 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 6597 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 6598 | running_off, pos); |
f2e2a985 | 6599 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 6600 | running_off = newoff; |
f502d50e | 6601 | } |
b17dc4d4 RB |
6602 | unsigned int group_el = 0; |
6603 | unsigned HOST_WIDE_INT | |
6604 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
f502d50e MM |
6605 | for (j = 0; j < ncopies; j++) |
6606 | { | |
c3a8f964 | 6607 | /* We've set op and dt above, from vect_get_store_rhs, |
f502d50e MM |
6608 | and first_stmt == stmt. */ |
6609 | if (j == 0) | |
6610 | { | |
6611 | if (slp) | |
6612 | { | |
6613 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
306b0c92 | 6614 | slp_node); |
f502d50e MM |
6615 | vec_oprnd = vec_oprnds[0]; |
6616 | } | |
6617 | else | |
6618 | { | |
c3a8f964 | 6619 | op = vect_get_store_rhs (next_stmt); |
81c40241 | 6620 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
f502d50e MM |
6621 | } |
6622 | } | |
f2e2a985 | 6623 | else |
f502d50e MM |
6624 | { |
6625 | if (slp) | |
6626 | vec_oprnd = vec_oprnds[j]; | |
6627 | else | |
c079cbac | 6628 | { |
929b4411 RS |
6629 | vect_is_simple_use (op, vinfo, &def_stmt, &rhs_dt); |
6630 | vec_oprnd = vect_get_vec_def_for_stmt_copy (rhs_dt, | |
6631 | vec_oprnd); | |
c079cbac | 6632 | } |
f502d50e | 6633 | } |
04199738 RB |
6634 | /* Pun the vector to extract from if necessary. */ |
6635 | if (lvectype != vectype) | |
6636 | { | |
6637 | tree tem = make_ssa_name (lvectype); | |
6638 | gimple *pun | |
6639 | = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
6640 | lvectype, vec_oprnd)); | |
6641 | vect_finish_stmt_generation (stmt, pun, gsi); | |
6642 | vec_oprnd = tem; | |
6643 | } | |
f502d50e MM |
6644 | for (i = 0; i < nstores; i++) |
6645 | { | |
6646 | tree newref, newoff; | |
355fe088 | 6647 | gimple *incr, *assign; |
f502d50e MM |
6648 | tree size = TYPE_SIZE (ltype); |
6649 | /* Extract the i'th component. */ | |
6650 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
6651 | bitsize_int (i), size); | |
6652 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
6653 | size, pos); | |
6654 | ||
6655 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
6656 | NULL_TREE, true, | |
6657 | GSI_SAME_STMT); | |
6658 | ||
b17dc4d4 RB |
6659 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
6660 | group_el * elsz); | |
f502d50e | 6661 | newref = build2 (MEM_REF, ltype, |
b17dc4d4 | 6662 | running_off, this_off); |
f502d50e MM |
6663 | |
6664 | /* And store it to *running_off. */ | |
6665 | assign = gimple_build_assign (newref, elem); | |
6666 | vect_finish_stmt_generation (stmt, assign, gsi); | |
6667 | ||
b17dc4d4 RB |
6668 | group_el += lnel; |
6669 | if (! slp | |
6670 | || group_el == group_size) | |
6671 | { | |
6672 | newoff = copy_ssa_name (running_off, NULL); | |
6673 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6674 | running_off, stride_step); | |
6675 | vect_finish_stmt_generation (stmt, incr, gsi); | |
f502d50e | 6676 | |
b17dc4d4 RB |
6677 | running_off = newoff; |
6678 | group_el = 0; | |
6679 | } | |
225ce44b RB |
6680 | if (g == group_size - 1 |
6681 | && !slp) | |
f502d50e MM |
6682 | { |
6683 | if (j == 0 && i == 0) | |
225ce44b RB |
6684 | STMT_VINFO_VEC_STMT (stmt_info) |
6685 | = *vec_stmt = assign; | |
f502d50e MM |
6686 | else |
6687 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign; | |
6688 | prev_stmt_info = vinfo_for_stmt (assign); | |
6689 | } | |
6690 | } | |
f2e2a985 | 6691 | } |
f502d50e | 6692 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
b17dc4d4 RB |
6693 | if (slp) |
6694 | break; | |
f2e2a985 | 6695 | } |
778dd3b6 RB |
6696 | |
6697 | vec_oprnds.release (); | |
f2e2a985 MM |
6698 | return true; |
6699 | } | |
6700 | ||
8c681247 | 6701 | auto_vec<tree> dr_chain (group_size); |
9771b263 | 6702 | oprnds.create (group_size); |
ebfd146a | 6703 | |
720f5239 | 6704 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6705 | gcc_assert (alignment_support_scheme); |
70088b95 RS |
6706 | vec_loop_masks *loop_masks |
6707 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) | |
6708 | ? &LOOP_VINFO_MASKS (loop_vinfo) | |
6709 | : NULL); | |
272c6793 | 6710 | /* Targets with store-lane instructions must not require explicit |
c3a8f964 RS |
6711 | realignment. vect_supportable_dr_alignment always returns either |
6712 | dr_aligned or dr_unaligned_supported for masked operations. */ | |
7cfb4d93 RS |
6713 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES |
6714 | && !mask | |
70088b95 | 6715 | && !loop_masks) |
272c6793 RS |
6716 | || alignment_support_scheme == dr_aligned |
6717 | || alignment_support_scheme == dr_unaligned_supported); | |
6718 | ||
62da9e14 RS |
6719 | if (memory_access_type == VMAT_CONTIGUOUS_DOWN |
6720 | || memory_access_type == VMAT_CONTIGUOUS_REVERSE) | |
09dfa495 BM |
6721 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
6722 | ||
f307441a RS |
6723 | tree bump; |
6724 | tree vec_offset = NULL_TREE; | |
6725 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
6726 | { | |
6727 | aggr_type = NULL_TREE; | |
6728 | bump = NULL_TREE; | |
6729 | } | |
6730 | else if (memory_access_type == VMAT_GATHER_SCATTER) | |
6731 | { | |
6732 | aggr_type = elem_type; | |
6733 | vect_get_strided_load_store_ops (stmt, loop_vinfo, &gs_info, | |
6734 | &bump, &vec_offset); | |
6735 | } | |
272c6793 | 6736 | else |
f307441a RS |
6737 | { |
6738 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
6739 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
6740 | else | |
6741 | aggr_type = vectype; | |
6742 | bump = vect_get_data_ptr_increment (dr, aggr_type, memory_access_type); | |
6743 | } | |
ebfd146a | 6744 | |
c3a8f964 RS |
6745 | if (mask) |
6746 | LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true; | |
6747 | ||
ebfd146a IR |
6748 | /* In case the vectorization factor (VF) is bigger than the number |
6749 | of elements that we can fit in a vectype (nunits), we have to generate | |
6750 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 6751 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
6752 | vect_get_vec_def_for_copy_stmt. */ |
6753 | ||
0d0293ac | 6754 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6755 | |
6756 | S1: &base + 2 = x2 | |
6757 | S2: &base = x0 | |
6758 | S3: &base + 1 = x1 | |
6759 | S4: &base + 3 = x3 | |
6760 | ||
6761 | We create vectorized stores starting from base address (the access of the | |
6762 | first stmt in the chain (S2 in the above example), when the last store stmt | |
6763 | of the chain (S4) is reached: | |
6764 | ||
6765 | VS1: &base = vx2 | |
6766 | VS2: &base + vec_size*1 = vx0 | |
6767 | VS3: &base + vec_size*2 = vx1 | |
6768 | VS4: &base + vec_size*3 = vx3 | |
6769 | ||
6770 | Then permutation statements are generated: | |
6771 | ||
3fcc1b55 JJ |
6772 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
6773 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 6774 | ... |
b8698a0f | 6775 | |
ebfd146a IR |
6776 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
6777 | (the order of the data-refs in the output of vect_permute_store_chain | |
6778 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6779 | the documentation of vect_permute_store_chain()). | |
6780 | ||
6781 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 6782 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 6783 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 6784 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
6785 | */ |
6786 | ||
6787 | prev_stmt_info = NULL; | |
c3a8f964 | 6788 | tree vec_mask = NULL_TREE; |
ebfd146a IR |
6789 | for (j = 0; j < ncopies; j++) |
6790 | { | |
ebfd146a IR |
6791 | |
6792 | if (j == 0) | |
6793 | { | |
6794 | if (slp) | |
6795 | { | |
6796 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c | 6797 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
306b0c92 | 6798 | NULL, slp_node); |
ebfd146a | 6799 | |
9771b263 | 6800 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
6801 | } |
6802 | else | |
6803 | { | |
b8698a0f L |
6804 | /* For interleaved stores we collect vectorized defs for all the |
6805 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
6806 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
6807 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
6808 | ||
0d0293ac | 6809 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 6810 | OPRNDS are of size 1. */ |
b8698a0f | 6811 | next_stmt = first_stmt; |
ebfd146a IR |
6812 | for (i = 0; i < group_size; i++) |
6813 | { | |
b8698a0f L |
6814 | /* Since gaps are not supported for interleaved stores, |
6815 | GROUP_SIZE is the exact number of stmts in the chain. | |
6816 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
6817 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a | 6818 | iteration of the loop will be executed. */ |
c3a8f964 | 6819 | op = vect_get_store_rhs (next_stmt); |
81c40241 | 6820 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
9771b263 DN |
6821 | dr_chain.quick_push (vec_oprnd); |
6822 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 6823 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a | 6824 | } |
c3a8f964 RS |
6825 | if (mask) |
6826 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, | |
6827 | mask_vectype); | |
ebfd146a IR |
6828 | } |
6829 | ||
6830 | /* We should have catched mismatched types earlier. */ | |
6831 | gcc_assert (useless_type_conversion_p (vectype, | |
6832 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
6833 | bool simd_lane_access_p |
6834 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
6835 | if (simd_lane_access_p | |
6836 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
6837 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
6838 | && integer_zerop (DR_OFFSET (first_dr)) | |
6839 | && integer_zerop (DR_INIT (first_dr)) | |
6840 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 6841 | get_alias_set (TREE_TYPE (ref_type)))) |
74bf76ed JJ |
6842 | { |
6843 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 6844 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 6845 | inv_p = false; |
74bf76ed | 6846 | } |
f307441a RS |
6847 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
6848 | { | |
6849 | vect_get_gather_scatter_ops (loop, stmt, &gs_info, | |
6850 | &dataref_ptr, &vec_offset); | |
6851 | inv_p = false; | |
6852 | } | |
74bf76ed JJ |
6853 | else |
6854 | dataref_ptr | |
6855 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
6856 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 6857 | offset, &dummy, gsi, &ptr_incr, |
f307441a RS |
6858 | simd_lane_access_p, &inv_p, |
6859 | NULL_TREE, bump); | |
a70d6342 | 6860 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 6861 | } |
b8698a0f | 6862 | else |
ebfd146a | 6863 | { |
b8698a0f L |
6864 | /* For interleaved stores we created vectorized defs for all the |
6865 | defs stored in OPRNDS in the previous iteration (previous copy). | |
6866 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
6867 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
6868 | next copy. | |
0d0293ac | 6869 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
6870 | OPRNDS are of size 1. */ |
6871 | for (i = 0; i < group_size; i++) | |
6872 | { | |
9771b263 | 6873 | op = oprnds[i]; |
929b4411 RS |
6874 | vect_is_simple_use (op, vinfo, &def_stmt, &rhs_dt); |
6875 | vec_oprnd = vect_get_vec_def_for_stmt_copy (rhs_dt, op); | |
9771b263 DN |
6876 | dr_chain[i] = vec_oprnd; |
6877 | oprnds[i] = vec_oprnd; | |
ebfd146a | 6878 | } |
c3a8f964 | 6879 | if (mask) |
929b4411 | 6880 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
74bf76ed JJ |
6881 | if (dataref_offset) |
6882 | dataref_offset | |
f307441a RS |
6883 | = int_const_binop (PLUS_EXPR, dataref_offset, bump); |
6884 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
929b4411 RS |
6885 | vec_offset = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
6886 | vec_offset); | |
74bf76ed JJ |
6887 | else |
6888 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
f307441a | 6889 | bump); |
ebfd146a IR |
6890 | } |
6891 | ||
2de001ee | 6892 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 6893 | { |
272c6793 | 6894 | tree vec_array; |
267d3070 | 6895 | |
272c6793 RS |
6896 | /* Combine all the vectors into an array. */ |
6897 | vec_array = create_vector_array (vectype, vec_num); | |
6898 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 6899 | { |
9771b263 | 6900 | vec_oprnd = dr_chain[i]; |
272c6793 | 6901 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 6902 | } |
b8698a0f | 6903 | |
7cfb4d93 | 6904 | tree final_mask = NULL; |
70088b95 RS |
6905 | if (loop_masks) |
6906 | final_mask = vect_get_loop_mask (gsi, loop_masks, ncopies, | |
6907 | vectype, j); | |
7cfb4d93 RS |
6908 | if (vec_mask) |
6909 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
6910 | vec_mask, gsi); | |
6911 | ||
7e11fc7f | 6912 | gcall *call; |
7cfb4d93 | 6913 | if (final_mask) |
7e11fc7f RS |
6914 | { |
6915 | /* Emit: | |
6916 | MASK_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK, | |
6917 | VEC_ARRAY). */ | |
6918 | unsigned int align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype)); | |
6919 | tree alias_ptr = build_int_cst (ref_type, align); | |
6920 | call = gimple_build_call_internal (IFN_MASK_STORE_LANES, 4, | |
6921 | dataref_ptr, alias_ptr, | |
7cfb4d93 | 6922 | final_mask, vec_array); |
7e11fc7f RS |
6923 | } |
6924 | else | |
6925 | { | |
6926 | /* Emit: | |
6927 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
6928 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); | |
6929 | call = gimple_build_call_internal (IFN_STORE_LANES, 1, | |
6930 | vec_array); | |
6931 | gimple_call_set_lhs (call, data_ref); | |
6932 | } | |
a844293d RS |
6933 | gimple_call_set_nothrow (call, true); |
6934 | new_stmt = call; | |
267d3070 | 6935 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
6936 | } |
6937 | else | |
6938 | { | |
6939 | new_stmt = NULL; | |
0d0293ac | 6940 | if (grouped_store) |
272c6793 | 6941 | { |
b6b9227d JJ |
6942 | if (j == 0) |
6943 | result_chain.create (group_size); | |
272c6793 RS |
6944 | /* Permute. */ |
6945 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
6946 | &result_chain); | |
6947 | } | |
c2d7ab2a | 6948 | |
272c6793 RS |
6949 | next_stmt = first_stmt; |
6950 | for (i = 0; i < vec_num; i++) | |
6951 | { | |
644ffefd | 6952 | unsigned align, misalign; |
272c6793 | 6953 | |
7cfb4d93 | 6954 | tree final_mask = NULL_TREE; |
70088b95 RS |
6955 | if (loop_masks) |
6956 | final_mask = vect_get_loop_mask (gsi, loop_masks, | |
6957 | vec_num * ncopies, | |
7cfb4d93 RS |
6958 | vectype, vec_num * j + i); |
6959 | if (vec_mask) | |
6960 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
6961 | vec_mask, gsi); | |
6962 | ||
f307441a RS |
6963 | if (memory_access_type == VMAT_GATHER_SCATTER) |
6964 | { | |
6965 | tree scale = size_int (gs_info.scale); | |
6966 | gcall *call; | |
70088b95 | 6967 | if (loop_masks) |
f307441a RS |
6968 | call = gimple_build_call_internal |
6969 | (IFN_MASK_SCATTER_STORE, 5, dataref_ptr, vec_offset, | |
6970 | scale, vec_oprnd, final_mask); | |
6971 | else | |
6972 | call = gimple_build_call_internal | |
6973 | (IFN_SCATTER_STORE, 4, dataref_ptr, vec_offset, | |
6974 | scale, vec_oprnd); | |
6975 | gimple_call_set_nothrow (call, true); | |
6976 | new_stmt = call; | |
6977 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6978 | break; | |
6979 | } | |
6980 | ||
272c6793 RS |
6981 | if (i > 0) |
6982 | /* Bump the vector pointer. */ | |
6983 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
f307441a | 6984 | stmt, bump); |
272c6793 RS |
6985 | |
6986 | if (slp) | |
9771b263 | 6987 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
6988 | else if (grouped_store) |
6989 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 6990 | vect_permute_store_chain(). */ |
9771b263 | 6991 | vec_oprnd = result_chain[i]; |
272c6793 | 6992 | |
f702e7d4 | 6993 | align = DR_TARGET_ALIGNMENT (first_dr); |
272c6793 | 6994 | if (aligned_access_p (first_dr)) |
644ffefd | 6995 | misalign = 0; |
272c6793 RS |
6996 | else if (DR_MISALIGNMENT (first_dr) == -1) |
6997 | { | |
25f68d90 | 6998 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 6999 | misalign = 0; |
272c6793 RS |
7000 | } |
7001 | else | |
c3a8f964 | 7002 | misalign = DR_MISALIGNMENT (first_dr); |
aed93b23 RB |
7003 | if (dataref_offset == NULL_TREE |
7004 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
7005 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
7006 | misalign); | |
c2d7ab2a | 7007 | |
62da9e14 | 7008 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
09dfa495 BM |
7009 | { |
7010 | tree perm_mask = perm_mask_for_reverse (vectype); | |
7011 | tree perm_dest | |
c3a8f964 | 7012 | = vect_create_destination_var (vect_get_store_rhs (stmt), |
09dfa495 | 7013 | vectype); |
b731b390 | 7014 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
7015 | |
7016 | /* Generate the permute statement. */ | |
355fe088 | 7017 | gimple *perm_stmt |
0d0e4a03 JJ |
7018 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
7019 | vec_oprnd, perm_mask); | |
09dfa495 BM |
7020 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
7021 | ||
7022 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
7023 | vec_oprnd = new_temp; | |
7024 | } | |
7025 | ||
272c6793 | 7026 | /* Arguments are ready. Create the new vector stmt. */ |
7cfb4d93 | 7027 | if (final_mask) |
c3a8f964 RS |
7028 | { |
7029 | align = least_bit_hwi (misalign | align); | |
7030 | tree ptr = build_int_cst (ref_type, align); | |
7031 | gcall *call | |
7032 | = gimple_build_call_internal (IFN_MASK_STORE, 4, | |
7033 | dataref_ptr, ptr, | |
7cfb4d93 | 7034 | final_mask, vec_oprnd); |
c3a8f964 RS |
7035 | gimple_call_set_nothrow (call, true); |
7036 | new_stmt = call; | |
7037 | } | |
7038 | else | |
7039 | { | |
7040 | data_ref = fold_build2 (MEM_REF, vectype, | |
7041 | dataref_ptr, | |
7042 | dataref_offset | |
7043 | ? dataref_offset | |
7044 | : build_int_cst (ref_type, 0)); | |
7045 | if (aligned_access_p (first_dr)) | |
7046 | ; | |
7047 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
7048 | TREE_TYPE (data_ref) | |
7049 | = build_aligned_type (TREE_TYPE (data_ref), | |
7050 | align * BITS_PER_UNIT); | |
7051 | else | |
7052 | TREE_TYPE (data_ref) | |
7053 | = build_aligned_type (TREE_TYPE (data_ref), | |
7054 | TYPE_ALIGN (elem_type)); | |
7055 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
7056 | } | |
272c6793 | 7057 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
7058 | |
7059 | if (slp) | |
7060 | continue; | |
7061 | ||
e14c1050 | 7062 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
7063 | if (!next_stmt) |
7064 | break; | |
7065 | } | |
ebfd146a | 7066 | } |
1da0876c RS |
7067 | if (!slp) |
7068 | { | |
7069 | if (j == 0) | |
7070 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7071 | else | |
7072 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7073 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7074 | } | |
ebfd146a IR |
7075 | } |
7076 | ||
9771b263 DN |
7077 | oprnds.release (); |
7078 | result_chain.release (); | |
7079 | vec_oprnds.release (); | |
ebfd146a IR |
7080 | |
7081 | return true; | |
7082 | } | |
7083 | ||
557be5a8 AL |
7084 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
7085 | VECTOR_CST mask. No checks are made that the target platform supports the | |
7ac7e286 | 7086 | mask, so callers may wish to test can_vec_perm_const_p separately, or use |
557be5a8 | 7087 | vect_gen_perm_mask_checked. */ |
a1e53f3f | 7088 | |
3fcc1b55 | 7089 | tree |
4aae3cb3 | 7090 | vect_gen_perm_mask_any (tree vectype, const vec_perm_indices &sel) |
a1e53f3f | 7091 | { |
b00cb3bf | 7092 | tree mask_type; |
a1e53f3f | 7093 | |
0ecc2b7d RS |
7094 | poly_uint64 nunits = sel.length (); |
7095 | gcc_assert (known_eq (nunits, TYPE_VECTOR_SUBPARTS (vectype))); | |
b00cb3bf RS |
7096 | |
7097 | mask_type = build_vector_type (ssizetype, nunits); | |
736d0f28 | 7098 | return vec_perm_indices_to_tree (mask_type, sel); |
a1e53f3f L |
7099 | } |
7100 | ||
7ac7e286 | 7101 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_const_p, |
cf7aa6a3 | 7102 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ |
557be5a8 AL |
7103 | |
7104 | tree | |
4aae3cb3 | 7105 | vect_gen_perm_mask_checked (tree vectype, const vec_perm_indices &sel) |
557be5a8 | 7106 | { |
7ac7e286 | 7107 | gcc_assert (can_vec_perm_const_p (TYPE_MODE (vectype), sel)); |
557be5a8 AL |
7108 | return vect_gen_perm_mask_any (vectype, sel); |
7109 | } | |
7110 | ||
aec7ae7d JJ |
7111 | /* Given a vector variable X and Y, that was generated for the scalar |
7112 | STMT, generate instructions to permute the vector elements of X and Y | |
7113 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
7114 | permuted vector variable. */ | |
a1e53f3f L |
7115 | |
7116 | static tree | |
355fe088 | 7117 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 7118 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
7119 | { |
7120 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 7121 | tree perm_dest, data_ref; |
355fe088 | 7122 | gimple *perm_stmt; |
a1e53f3f | 7123 | |
7ad429a4 RS |
7124 | tree scalar_dest = gimple_get_lhs (stmt); |
7125 | if (TREE_CODE (scalar_dest) == SSA_NAME) | |
7126 | perm_dest = vect_create_destination_var (scalar_dest, vectype); | |
7127 | else | |
7128 | perm_dest = vect_get_new_vect_var (vectype, vect_simple_var, NULL); | |
b731b390 | 7129 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
7130 | |
7131 | /* Generate the permute statement. */ | |
0d0e4a03 | 7132 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
7133 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
7134 | ||
7135 | return data_ref; | |
7136 | } | |
7137 | ||
6b916b36 RB |
7138 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
7139 | inserting them on the loops preheader edge. Returns true if we | |
7140 | were successful in doing so (and thus STMT can be moved then), | |
7141 | otherwise returns false. */ | |
7142 | ||
7143 | static bool | |
355fe088 | 7144 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
7145 | { |
7146 | ssa_op_iter i; | |
7147 | tree op; | |
7148 | bool any = false; | |
7149 | ||
7150 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
7151 | { | |
355fe088 | 7152 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
7153 | if (!gimple_nop_p (def_stmt) |
7154 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
7155 | { | |
7156 | /* Make sure we don't need to recurse. While we could do | |
7157 | so in simple cases when there are more complex use webs | |
7158 | we don't have an easy way to preserve stmt order to fulfil | |
7159 | dependencies within them. */ | |
7160 | tree op2; | |
7161 | ssa_op_iter i2; | |
d1417442 JJ |
7162 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
7163 | return false; | |
6b916b36 RB |
7164 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
7165 | { | |
355fe088 | 7166 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
7167 | if (!gimple_nop_p (def_stmt2) |
7168 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
7169 | return false; | |
7170 | } | |
7171 | any = true; | |
7172 | } | |
7173 | } | |
7174 | ||
7175 | if (!any) | |
7176 | return true; | |
7177 | ||
7178 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
7179 | { | |
355fe088 | 7180 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
7181 | if (!gimple_nop_p (def_stmt) |
7182 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
7183 | { | |
7184 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
7185 | gsi_remove (&gsi, false); | |
7186 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
7187 | } | |
7188 | } | |
7189 | ||
7190 | return true; | |
7191 | } | |
7192 | ||
ebfd146a IR |
7193 | /* vectorizable_load. |
7194 | ||
b8698a0f L |
7195 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
7196 | can be vectorized. | |
7197 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
7198 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
7199 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7200 | ||
7201 | static bool | |
355fe088 | 7202 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 7203 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
7204 | { |
7205 | tree scalar_dest; | |
7206 | tree vec_dest = NULL; | |
7207 | tree data_ref = NULL; | |
7208 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 7209 | stmt_vec_info prev_stmt_info; |
ebfd146a | 7210 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 7211 | struct loop *loop = NULL; |
ebfd146a | 7212 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 7213 | bool nested_in_vect_loop = false; |
c716e67f | 7214 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 7215 | tree elem_type; |
ebfd146a | 7216 | tree new_temp; |
ef4bddc2 | 7217 | machine_mode mode; |
355fe088 | 7218 | gimple *new_stmt = NULL; |
ebfd146a IR |
7219 | tree dummy; |
7220 | enum dr_alignment_support alignment_support_scheme; | |
7221 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 7222 | tree dataref_offset = NULL_TREE; |
355fe088 | 7223 | gimple *ptr_incr = NULL; |
ebfd146a | 7224 | int ncopies; |
4d694b27 RS |
7225 | int i, j; |
7226 | unsigned int group_size; | |
7227 | poly_uint64 group_gap_adj; | |
ebfd146a IR |
7228 | tree msq = NULL_TREE, lsq; |
7229 | tree offset = NULL_TREE; | |
356bbc4c | 7230 | tree byte_offset = NULL_TREE; |
ebfd146a | 7231 | tree realignment_token = NULL_TREE; |
538dd0b7 | 7232 | gphi *phi = NULL; |
6e1aa848 | 7233 | vec<tree> dr_chain = vNULL; |
0d0293ac | 7234 | bool grouped_load = false; |
355fe088 | 7235 | gimple *first_stmt; |
4f0a0218 | 7236 | gimple *first_stmt_for_drptr = NULL; |
ebfd146a IR |
7237 | bool inv_p; |
7238 | bool compute_in_loop = false; | |
7239 | struct loop *at_loop; | |
7240 | int vec_num; | |
7241 | bool slp = (slp_node != NULL); | |
7242 | bool slp_perm = false; | |
a70d6342 | 7243 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
d9f21f6a | 7244 | poly_uint64 vf; |
272c6793 | 7245 | tree aggr_type; |
134c85ca | 7246 | gather_scatter_info gs_info; |
310213d4 | 7247 | vec_info *vinfo = stmt_info->vinfo; |
44fc7854 | 7248 | tree ref_type; |
929b4411 | 7249 | enum vect_def_type mask_dt = vect_unknown_def_type; |
a70d6342 | 7250 | |
465c8c19 JJ |
7251 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
7252 | return false; | |
7253 | ||
66c16fd9 RB |
7254 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7255 | && ! vec_stmt) | |
465c8c19 JJ |
7256 | return false; |
7257 | ||
c3a8f964 RS |
7258 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
7259 | if (is_gimple_assign (stmt)) | |
7260 | { | |
7261 | scalar_dest = gimple_assign_lhs (stmt); | |
7262 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
7263 | return false; | |
465c8c19 | 7264 | |
c3a8f964 RS |
7265 | tree_code code = gimple_assign_rhs_code (stmt); |
7266 | if (code != ARRAY_REF | |
7267 | && code != BIT_FIELD_REF | |
7268 | && code != INDIRECT_REF | |
7269 | && code != COMPONENT_REF | |
7270 | && code != IMAGPART_EXPR | |
7271 | && code != REALPART_EXPR | |
7272 | && code != MEM_REF | |
7273 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
7274 | return false; | |
7275 | } | |
7276 | else | |
7277 | { | |
7278 | gcall *call = dyn_cast <gcall *> (stmt); | |
bfaa08b7 RS |
7279 | if (!call || !gimple_call_internal_p (call)) |
7280 | return false; | |
7281 | ||
7282 | internal_fn ifn = gimple_call_internal_fn (call); | |
7283 | if (!internal_load_fn_p (ifn)) | |
c3a8f964 | 7284 | return false; |
465c8c19 | 7285 | |
c3a8f964 RS |
7286 | scalar_dest = gimple_call_lhs (call); |
7287 | if (!scalar_dest) | |
7288 | return false; | |
7289 | ||
7290 | if (slp_node != NULL) | |
7291 | { | |
7292 | if (dump_enabled_p ()) | |
7293 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7294 | "SLP of masked loads not supported.\n"); | |
7295 | return false; | |
7296 | } | |
7297 | ||
bfaa08b7 RS |
7298 | int mask_index = internal_fn_mask_index (ifn); |
7299 | if (mask_index >= 0) | |
7300 | { | |
7301 | mask = gimple_call_arg (call, mask_index); | |
929b4411 RS |
7302 | if (!vect_check_load_store_mask (stmt, mask, &mask_dt, |
7303 | &mask_vectype)) | |
bfaa08b7 RS |
7304 | return false; |
7305 | } | |
c3a8f964 | 7306 | } |
465c8c19 JJ |
7307 | |
7308 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
7309 | return false; | |
7310 | ||
7311 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
4d694b27 | 7312 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 | 7313 | |
a70d6342 IR |
7314 | if (loop_vinfo) |
7315 | { | |
7316 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
7317 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
7318 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
7319 | } | |
7320 | else | |
3533e503 | 7321 | vf = 1; |
ebfd146a IR |
7322 | |
7323 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 7324 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 7325 | case of SLP. */ |
fce57248 | 7326 | if (slp) |
ebfd146a IR |
7327 | ncopies = 1; |
7328 | else | |
e8f142e2 | 7329 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
ebfd146a IR |
7330 | |
7331 | gcc_assert (ncopies >= 1); | |
7332 | ||
7333 | /* FORNOW. This restriction should be relaxed. */ | |
7334 | if (nested_in_vect_loop && ncopies > 1) | |
7335 | { | |
73fbfcad | 7336 | if (dump_enabled_p ()) |
78c60e3d | 7337 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7338 | "multiple types in nested loop.\n"); |
ebfd146a IR |
7339 | return false; |
7340 | } | |
7341 | ||
f2556b68 RB |
7342 | /* Invalidate assumptions made by dependence analysis when vectorization |
7343 | on the unrolled body effectively re-orders stmts. */ | |
7344 | if (ncopies > 1 | |
7345 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
7346 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
7347 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
7348 | { |
7349 | if (dump_enabled_p ()) | |
7350 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7351 | "cannot perform implicit CSE when unrolling " | |
7352 | "with negative dependence distance\n"); | |
7353 | return false; | |
7354 | } | |
7355 | ||
7b7b1813 | 7356 | elem_type = TREE_TYPE (vectype); |
947131ba | 7357 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
7358 | |
7359 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
7360 | (e.g. - data copies). */ | |
947131ba | 7361 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 7362 | { |
73fbfcad | 7363 | if (dump_enabled_p ()) |
78c60e3d | 7364 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7365 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
7366 | return false; |
7367 | } | |
7368 | ||
ebfd146a | 7369 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 7370 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 7371 | { |
0d0293ac | 7372 | grouped_load = true; |
ebfd146a | 7373 | /* FORNOW */ |
2de001ee RS |
7374 | gcc_assert (!nested_in_vect_loop); |
7375 | gcc_assert (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)); | |
ebfd146a | 7376 | |
e14c1050 | 7377 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
d3465d72 | 7378 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
d5f035ea | 7379 | |
b1af7da6 RB |
7380 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
7381 | slp_perm = true; | |
7382 | ||
f2556b68 RB |
7383 | /* Invalidate assumptions made by dependence analysis when vectorization |
7384 | on the unrolled body effectively re-orders stmts. */ | |
7385 | if (!PURE_SLP_STMT (stmt_info) | |
7386 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
7387 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
7388 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
7389 | { |
7390 | if (dump_enabled_p ()) | |
7391 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7392 | "cannot perform implicit CSE when performing " | |
7393 | "group loads with negative dependence distance\n"); | |
7394 | return false; | |
7395 | } | |
96bb56b2 RB |
7396 | |
7397 | /* Similarly when the stmt is a load that is both part of a SLP | |
7398 | instance and a loop vectorized stmt via the same-dr mechanism | |
7399 | we have to give up. */ | |
7400 | if (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info) | |
7401 | && (STMT_SLP_TYPE (stmt_info) | |
7402 | != STMT_SLP_TYPE (vinfo_for_stmt | |
7403 | (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info))))) | |
7404 | { | |
7405 | if (dump_enabled_p ()) | |
7406 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7407 | "conflicting SLP types for CSEd load\n"); | |
7408 | return false; | |
7409 | } | |
ebfd146a | 7410 | } |
7cfb4d93 RS |
7411 | else |
7412 | group_size = 1; | |
ebfd146a | 7413 | |
2de001ee | 7414 | vect_memory_access_type memory_access_type; |
7e11fc7f | 7415 | if (!get_load_store_type (stmt, vectype, slp, mask, VLS_LOAD, ncopies, |
2de001ee RS |
7416 | &memory_access_type, &gs_info)) |
7417 | return false; | |
a1e53f3f | 7418 | |
c3a8f964 RS |
7419 | if (mask) |
7420 | { | |
7421 | if (memory_access_type == VMAT_CONTIGUOUS) | |
7422 | { | |
7e11fc7f RS |
7423 | machine_mode vec_mode = TYPE_MODE (vectype); |
7424 | if (!VECTOR_MODE_P (vec_mode) | |
7425 | || !can_vec_mask_load_store_p (vec_mode, | |
c3a8f964 RS |
7426 | TYPE_MODE (mask_vectype), true)) |
7427 | return false; | |
7428 | } | |
bfaa08b7 | 7429 | else if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
c3a8f964 RS |
7430 | { |
7431 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); | |
7432 | tree masktype | |
7433 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
7434 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
7435 | { | |
7436 | if (dump_enabled_p ()) | |
7437 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7438 | "masked gather with integer mask not" | |
7439 | " supported."); | |
7440 | return false; | |
7441 | } | |
7442 | } | |
bfaa08b7 RS |
7443 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
7444 | && memory_access_type != VMAT_GATHER_SCATTER) | |
c3a8f964 RS |
7445 | { |
7446 | if (dump_enabled_p ()) | |
7447 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7448 | "unsupported access type for masked load.\n"); | |
7449 | return false; | |
7450 | } | |
7451 | } | |
7452 | ||
ebfd146a IR |
7453 | if (!vec_stmt) /* transformation not required. */ |
7454 | { | |
2de001ee RS |
7455 | if (!slp) |
7456 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; | |
7cfb4d93 RS |
7457 | |
7458 | if (loop_vinfo | |
7459 | && LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) | |
7460 | check_load_store_masking (loop_vinfo, vectype, VLS_LOAD, group_size, | |
bfaa08b7 | 7461 | memory_access_type, &gs_info); |
7cfb4d93 | 7462 | |
ebfd146a | 7463 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; |
2e8ab70c | 7464 | /* The SLP costs are calculated during SLP analysis. */ |
78604de0 | 7465 | if (! slp_node) |
2de001ee | 7466 | vect_model_load_cost (stmt_info, ncopies, memory_access_type, |
2e8ab70c | 7467 | NULL, NULL, NULL); |
ebfd146a IR |
7468 | return true; |
7469 | } | |
7470 | ||
2de001ee RS |
7471 | if (!slp) |
7472 | gcc_assert (memory_access_type | |
7473 | == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); | |
7474 | ||
73fbfcad | 7475 | if (dump_enabled_p ()) |
78c60e3d | 7476 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 7477 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a | 7478 | |
67b8dbac | 7479 | /* Transform. */ |
ebfd146a | 7480 | |
f702e7d4 | 7481 | ensure_base_align (dr); |
c716e67f | 7482 | |
bfaa08b7 | 7483 | if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
aec7ae7d | 7484 | { |
929b4411 RS |
7485 | vect_build_gather_load_calls (stmt, gsi, vec_stmt, &gs_info, mask, |
7486 | mask_dt); | |
aec7ae7d JJ |
7487 | return true; |
7488 | } | |
2de001ee RS |
7489 | |
7490 | if (memory_access_type == VMAT_ELEMENTWISE | |
7491 | || memory_access_type == VMAT_STRIDED_SLP) | |
7d75abc8 MM |
7492 | { |
7493 | gimple_stmt_iterator incr_gsi; | |
7494 | bool insert_after; | |
355fe088 | 7495 | gimple *incr; |
7d75abc8 | 7496 | tree offvar; |
7d75abc8 MM |
7497 | tree ivstep; |
7498 | tree running_off; | |
9771b263 | 7499 | vec<constructor_elt, va_gc> *v = NULL; |
14ac6aa2 | 7500 | tree stride_base, stride_step, alias_off; |
4d694b27 RS |
7501 | /* Checked by get_load_store_type. */ |
7502 | unsigned int const_nunits = nunits.to_constant (); | |
b210f45f | 7503 | unsigned HOST_WIDE_INT cst_offset = 0; |
14ac6aa2 | 7504 | |
7cfb4d93 | 7505 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
14ac6aa2 | 7506 | gcc_assert (!nested_in_vect_loop); |
7d75abc8 | 7507 | |
b210f45f | 7508 | if (grouped_load) |
44fc7854 BE |
7509 | { |
7510 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); | |
7511 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
44fc7854 | 7512 | } |
ab313a8c | 7513 | else |
44fc7854 BE |
7514 | { |
7515 | first_stmt = stmt; | |
7516 | first_dr = dr; | |
b210f45f RB |
7517 | } |
7518 | if (slp && grouped_load) | |
7519 | { | |
7520 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
7521 | ref_type = get_group_alias_ptr_type (first_stmt); | |
7522 | } | |
7523 | else | |
7524 | { | |
7525 | if (grouped_load) | |
7526 | cst_offset | |
7527 | = (tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))) | |
7528 | * vect_get_place_in_interleaving_chain (stmt, first_stmt)); | |
44fc7854 | 7529 | group_size = 1; |
b210f45f | 7530 | ref_type = reference_alias_ptr_type (DR_REF (dr)); |
44fc7854 | 7531 | } |
ab313a8c | 7532 | |
14ac6aa2 RB |
7533 | stride_base |
7534 | = fold_build_pointer_plus | |
ab313a8c | 7535 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 7536 | size_binop (PLUS_EXPR, |
ab313a8c RB |
7537 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
7538 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
7539 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
7540 | |
7541 | /* For a load with loop-invariant (but other than power-of-2) | |
7542 | stride (i.e. not a grouped access) like so: | |
7543 | ||
7544 | for (i = 0; i < n; i += stride) | |
7545 | ... = array[i]; | |
7546 | ||
7547 | we generate a new induction variable and new accesses to | |
7548 | form a new vector (or vectors, depending on ncopies): | |
7549 | ||
7550 | for (j = 0; ; j += VF*stride) | |
7551 | tmp1 = array[j]; | |
7552 | tmp2 = array[j + stride]; | |
7553 | ... | |
7554 | vectemp = {tmp1, tmp2, ...} | |
7555 | */ | |
7556 | ||
ab313a8c RB |
7557 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
7558 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
7559 | |
7560 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
7561 | ||
b210f45f RB |
7562 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
7563 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); | |
7564 | create_iv (stride_base, ivstep, NULL, | |
7d75abc8 MM |
7565 | loop, &incr_gsi, insert_after, |
7566 | &offvar, NULL); | |
7567 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 7568 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
7d75abc8 | 7569 | |
b210f45f | 7570 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
7d75abc8 MM |
7571 | |
7572 | prev_stmt_info = NULL; | |
7573 | running_off = offvar; | |
44fc7854 | 7574 | alias_off = build_int_cst (ref_type, 0); |
4d694b27 | 7575 | int nloads = const_nunits; |
e09b4c37 | 7576 | int lnel = 1; |
7b5fc413 | 7577 | tree ltype = TREE_TYPE (vectype); |
ea60dd34 | 7578 | tree lvectype = vectype; |
b266b968 | 7579 | auto_vec<tree> dr_chain; |
2de001ee | 7580 | if (memory_access_type == VMAT_STRIDED_SLP) |
7b5fc413 | 7581 | { |
4d694b27 | 7582 | if (group_size < const_nunits) |
e09b4c37 | 7583 | { |
ff03930a JJ |
7584 | /* First check if vec_init optab supports construction from |
7585 | vector elts directly. */ | |
b397965c | 7586 | scalar_mode elmode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); |
9da15d40 RS |
7587 | machine_mode vmode; |
7588 | if (mode_for_vector (elmode, group_size).exists (&vmode) | |
7589 | && VECTOR_MODE_P (vmode) | |
414fef4e | 7590 | && targetm.vector_mode_supported_p (vmode) |
ff03930a JJ |
7591 | && (convert_optab_handler (vec_init_optab, |
7592 | TYPE_MODE (vectype), vmode) | |
7593 | != CODE_FOR_nothing)) | |
ea60dd34 | 7594 | { |
4d694b27 | 7595 | nloads = const_nunits / group_size; |
ea60dd34 | 7596 | lnel = group_size; |
ff03930a JJ |
7597 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); |
7598 | } | |
7599 | else | |
7600 | { | |
7601 | /* Otherwise avoid emitting a constructor of vector elements | |
7602 | by performing the loads using an integer type of the same | |
7603 | size, constructing a vector of those and then | |
7604 | re-interpreting it as the original vector type. | |
7605 | This avoids a huge runtime penalty due to the general | |
7606 | inability to perform store forwarding from smaller stores | |
7607 | to a larger load. */ | |
7608 | unsigned lsize | |
7609 | = group_size * TYPE_PRECISION (TREE_TYPE (vectype)); | |
fffbab82 | 7610 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 7611 | unsigned int lnunits = const_nunits / group_size; |
ff03930a JJ |
7612 | /* If we can't construct such a vector fall back to |
7613 | element loads of the original vector type. */ | |
4d694b27 | 7614 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 7615 | && VECTOR_MODE_P (vmode) |
414fef4e | 7616 | && targetm.vector_mode_supported_p (vmode) |
ff03930a JJ |
7617 | && (convert_optab_handler (vec_init_optab, vmode, elmode) |
7618 | != CODE_FOR_nothing)) | |
7619 | { | |
4d694b27 | 7620 | nloads = lnunits; |
ff03930a JJ |
7621 | lnel = group_size; |
7622 | ltype = build_nonstandard_integer_type (lsize, 1); | |
7623 | lvectype = build_vector_type (ltype, nloads); | |
7624 | } | |
ea60dd34 | 7625 | } |
e09b4c37 | 7626 | } |
2de001ee | 7627 | else |
e09b4c37 | 7628 | { |
ea60dd34 | 7629 | nloads = 1; |
4d694b27 | 7630 | lnel = const_nunits; |
e09b4c37 | 7631 | ltype = vectype; |
e09b4c37 | 7632 | } |
2de001ee RS |
7633 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); |
7634 | } | |
7635 | if (slp) | |
7636 | { | |
66c16fd9 RB |
7637 | /* For SLP permutation support we need to load the whole group, |
7638 | not only the number of vector stmts the permutation result | |
7639 | fits in. */ | |
b266b968 | 7640 | if (slp_perm) |
66c16fd9 | 7641 | { |
d9f21f6a RS |
7642 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7643 | variable VF. */ | |
7644 | unsigned int const_vf = vf.to_constant (); | |
4d694b27 | 7645 | ncopies = CEIL (group_size * const_vf, const_nunits); |
66c16fd9 RB |
7646 | dr_chain.create (ncopies); |
7647 | } | |
7648 | else | |
7649 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
7b5fc413 | 7650 | } |
4d694b27 | 7651 | unsigned int group_el = 0; |
e09b4c37 RB |
7652 | unsigned HOST_WIDE_INT |
7653 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
7d75abc8 MM |
7654 | for (j = 0; j < ncopies; j++) |
7655 | { | |
7b5fc413 | 7656 | if (nloads > 1) |
e09b4c37 RB |
7657 | vec_alloc (v, nloads); |
7658 | for (i = 0; i < nloads; i++) | |
7b5fc413 | 7659 | { |
e09b4c37 | 7660 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
b210f45f | 7661 | group_el * elsz + cst_offset); |
e09b4c37 RB |
7662 | new_stmt = gimple_build_assign (make_ssa_name (ltype), |
7663 | build2 (MEM_REF, ltype, | |
7664 | running_off, this_off)); | |
7665 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7666 | if (nloads > 1) | |
7667 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, | |
7668 | gimple_assign_lhs (new_stmt)); | |
7669 | ||
7670 | group_el += lnel; | |
7671 | if (! slp | |
7672 | || group_el == group_size) | |
7b5fc413 | 7673 | { |
e09b4c37 RB |
7674 | tree newoff = copy_ssa_name (running_off); |
7675 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
7676 | running_off, stride_step); | |
7b5fc413 RB |
7677 | vect_finish_stmt_generation (stmt, incr, gsi); |
7678 | ||
7679 | running_off = newoff; | |
e09b4c37 | 7680 | group_el = 0; |
7b5fc413 | 7681 | } |
7b5fc413 | 7682 | } |
e09b4c37 | 7683 | if (nloads > 1) |
7d75abc8 | 7684 | { |
ea60dd34 RB |
7685 | tree vec_inv = build_constructor (lvectype, v); |
7686 | new_temp = vect_init_vector (stmt, vec_inv, lvectype, gsi); | |
e09b4c37 | 7687 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
ea60dd34 RB |
7688 | if (lvectype != vectype) |
7689 | { | |
7690 | new_stmt = gimple_build_assign (make_ssa_name (vectype), | |
7691 | VIEW_CONVERT_EXPR, | |
7692 | build1 (VIEW_CONVERT_EXPR, | |
7693 | vectype, new_temp)); | |
7694 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7695 | } | |
7d75abc8 MM |
7696 | } |
7697 | ||
7b5fc413 | 7698 | if (slp) |
b266b968 | 7699 | { |
b266b968 RB |
7700 | if (slp_perm) |
7701 | dr_chain.quick_push (gimple_assign_lhs (new_stmt)); | |
66c16fd9 RB |
7702 | else |
7703 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
b266b968 | 7704 | } |
7d75abc8 | 7705 | else |
225ce44b RB |
7706 | { |
7707 | if (j == 0) | |
7708 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7709 | else | |
7710 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7711 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7712 | } | |
7d75abc8 | 7713 | } |
b266b968 | 7714 | if (slp_perm) |
29afecdf RB |
7715 | { |
7716 | unsigned n_perms; | |
7717 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
7718 | slp_node_instance, false, &n_perms); | |
7719 | } | |
7d75abc8 MM |
7720 | return true; |
7721 | } | |
aec7ae7d | 7722 | |
b5ec4de7 RS |
7723 | if (memory_access_type == VMAT_GATHER_SCATTER |
7724 | || (!slp && memory_access_type == VMAT_CONTIGUOUS)) | |
ab2fc782 RS |
7725 | grouped_load = false; |
7726 | ||
0d0293ac | 7727 | if (grouped_load) |
ebfd146a | 7728 | { |
e14c1050 | 7729 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
44fc7854 | 7730 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
4f0a0218 | 7731 | /* For SLP vectorization we directly vectorize a subchain |
52eab378 RB |
7732 | without permutation. */ |
7733 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
4f0a0218 RB |
7734 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
7735 | /* For BB vectorization always use the first stmt to base | |
7736 | the data ref pointer on. */ | |
7737 | if (bb_vinfo) | |
7738 | first_stmt_for_drptr = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 7739 | |
ebfd146a | 7740 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
7741 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
7742 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
7743 | ??? But we can only do so if there is exactly one | |
7744 | as we have no way to get at the rest. Leave the CSE | |
7745 | opportunity alone. | |
7746 | ??? With the group load eventually participating | |
7747 | in multiple different permutations (having multiple | |
7748 | slp nodes which refer to the same group) the CSE | |
7749 | is even wrong code. See PR56270. */ | |
7750 | && !slp) | |
ebfd146a IR |
7751 | { |
7752 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
7753 | return true; | |
7754 | } | |
7755 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
9b999e8c | 7756 | group_gap_adj = 0; |
ebfd146a IR |
7757 | |
7758 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
7759 | if (slp) | |
7760 | { | |
0d0293ac | 7761 | grouped_load = false; |
91ff1504 RB |
7762 | /* For SLP permutation support we need to load the whole group, |
7763 | not only the number of vector stmts the permutation result | |
7764 | fits in. */ | |
7765 | if (slp_perm) | |
b267968e | 7766 | { |
d9f21f6a RS |
7767 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7768 | variable VF. */ | |
7769 | unsigned int const_vf = vf.to_constant (); | |
4d694b27 RS |
7770 | unsigned int const_nunits = nunits.to_constant (); |
7771 | vec_num = CEIL (group_size * const_vf, const_nunits); | |
b267968e RB |
7772 | group_gap_adj = vf * group_size - nunits * vec_num; |
7773 | } | |
91ff1504 | 7774 | else |
b267968e RB |
7775 | { |
7776 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
796bd467 RB |
7777 | group_gap_adj |
7778 | = group_size - SLP_INSTANCE_GROUP_SIZE (slp_node_instance); | |
b267968e | 7779 | } |
a70d6342 | 7780 | } |
ebfd146a | 7781 | else |
9b999e8c | 7782 | vec_num = group_size; |
44fc7854 BE |
7783 | |
7784 | ref_type = get_group_alias_ptr_type (first_stmt); | |
ebfd146a IR |
7785 | } |
7786 | else | |
7787 | { | |
7788 | first_stmt = stmt; | |
7789 | first_dr = dr; | |
7790 | group_size = vec_num = 1; | |
9b999e8c | 7791 | group_gap_adj = 0; |
44fc7854 | 7792 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
ebfd146a IR |
7793 | } |
7794 | ||
720f5239 | 7795 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 7796 | gcc_assert (alignment_support_scheme); |
70088b95 RS |
7797 | vec_loop_masks *loop_masks |
7798 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) | |
7799 | ? &LOOP_VINFO_MASKS (loop_vinfo) | |
7800 | : NULL); | |
7cfb4d93 RS |
7801 | /* Targets with store-lane instructions must not require explicit |
7802 | realignment. vect_supportable_dr_alignment always returns either | |
7803 | dr_aligned or dr_unaligned_supported for masked operations. */ | |
7804 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES | |
7805 | && !mask | |
70088b95 | 7806 | && !loop_masks) |
272c6793 RS |
7807 | || alignment_support_scheme == dr_aligned |
7808 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
7809 | |
7810 | /* In case the vectorization factor (VF) is bigger than the number | |
7811 | of elements that we can fit in a vectype (nunits), we have to generate | |
7812 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 7813 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 7814 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 7815 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 7816 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
7817 | stmts that use the defs of the current stmt. The example below |
7818 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
7819 | need to create 4 vectorized stmts): | |
ebfd146a IR |
7820 | |
7821 | before vectorization: | |
7822 | RELATED_STMT VEC_STMT | |
7823 | S1: x = memref - - | |
7824 | S2: z = x + 1 - - | |
7825 | ||
7826 | step 1: vectorize stmt S1: | |
7827 | We first create the vector stmt VS1_0, and, as usual, record a | |
7828 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
7829 | Next, we create the vector stmt VS1_1, and record a pointer to | |
7830 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 7831 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
7832 | stmts and pointers: |
7833 | RELATED_STMT VEC_STMT | |
7834 | VS1_0: vx0 = memref0 VS1_1 - | |
7835 | VS1_1: vx1 = memref1 VS1_2 - | |
7836 | VS1_2: vx2 = memref2 VS1_3 - | |
7837 | VS1_3: vx3 = memref3 - - | |
7838 | S1: x = load - VS1_0 | |
7839 | S2: z = x + 1 - - | |
7840 | ||
b8698a0f L |
7841 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
7842 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
7843 | stmt S2. */ |
7844 | ||
0d0293ac | 7845 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
7846 | |
7847 | S1: x2 = &base + 2 | |
7848 | S2: x0 = &base | |
7849 | S3: x1 = &base + 1 | |
7850 | S4: x3 = &base + 3 | |
7851 | ||
b8698a0f | 7852 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
7853 | starting from the access of the first stmt of the chain: |
7854 | ||
7855 | VS1: vx0 = &base | |
7856 | VS2: vx1 = &base + vec_size*1 | |
7857 | VS3: vx3 = &base + vec_size*2 | |
7858 | VS4: vx4 = &base + vec_size*3 | |
7859 | ||
7860 | Then permutation statements are generated: | |
7861 | ||
e2c83630 RH |
7862 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
7863 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
7864 | ... |
7865 | ||
7866 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
7867 | (the order of the data-refs in the output of vect_permute_load_chain | |
7868 | corresponds to the order of scalar stmts in the interleaving chain - see | |
7869 | the documentation of vect_permute_load_chain()). | |
7870 | The generation of permutation stmts and recording them in | |
0d0293ac | 7871 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 7872 | |
b8698a0f | 7873 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
7874 | permutation stmts above are created for every copy. The result vector |
7875 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
7876 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
7877 | |
7878 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
7879 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
7880 | we generate the following code: | |
7881 | p = initial_addr; | |
7882 | indx = 0; | |
7883 | loop { | |
7884 | p = p + indx * vectype_size; | |
7885 | vec_dest = *(p); | |
7886 | indx = indx + 1; | |
7887 | } | |
7888 | ||
7889 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 7890 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
7891 | then generate the following code, in which the data in each iteration is |
7892 | obtained by two vector loads, one from the previous iteration, and one | |
7893 | from the current iteration: | |
7894 | p1 = initial_addr; | |
7895 | msq_init = *(floor(p1)) | |
7896 | p2 = initial_addr + VS - 1; | |
7897 | realignment_token = call target_builtin; | |
7898 | indx = 0; | |
7899 | loop { | |
7900 | p2 = p2 + indx * vectype_size | |
7901 | lsq = *(floor(p2)) | |
7902 | vec_dest = realign_load (msq, lsq, realignment_token) | |
7903 | indx = indx + 1; | |
7904 | msq = lsq; | |
7905 | } */ | |
7906 | ||
7907 | /* If the misalignment remains the same throughout the execution of the | |
7908 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 7909 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
7910 | This can only occur when vectorizing memory accesses in the inner-loop |
7911 | nested within an outer-loop that is being vectorized. */ | |
7912 | ||
d1e4b493 | 7913 | if (nested_in_vect_loop |
cf098191 RS |
7914 | && !multiple_p (DR_STEP_ALIGNMENT (dr), |
7915 | GET_MODE_SIZE (TYPE_MODE (vectype)))) | |
ebfd146a IR |
7916 | { |
7917 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
7918 | compute_in_loop = true; | |
7919 | } | |
7920 | ||
7921 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
7922 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 7923 | && !compute_in_loop) |
ebfd146a IR |
7924 | { |
7925 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
7926 | alignment_support_scheme, NULL_TREE, | |
7927 | &at_loop); | |
7928 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7929 | { | |
538dd0b7 | 7930 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
7931 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
7932 | size_one_node); | |
ebfd146a IR |
7933 | } |
7934 | } | |
7935 | else | |
7936 | at_loop = loop; | |
7937 | ||
62da9e14 | 7938 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
a1e53f3f L |
7939 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
7940 | ||
ab2fc782 RS |
7941 | tree bump; |
7942 | tree vec_offset = NULL_TREE; | |
7943 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
7944 | { | |
7945 | aggr_type = NULL_TREE; | |
7946 | bump = NULL_TREE; | |
7947 | } | |
7948 | else if (memory_access_type == VMAT_GATHER_SCATTER) | |
7949 | { | |
7950 | aggr_type = elem_type; | |
7951 | vect_get_strided_load_store_ops (stmt, loop_vinfo, &gs_info, | |
7952 | &bump, &vec_offset); | |
7953 | } | |
272c6793 | 7954 | else |
ab2fc782 RS |
7955 | { |
7956 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
7957 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
7958 | else | |
7959 | aggr_type = vectype; | |
7960 | bump = vect_get_data_ptr_increment (dr, aggr_type, memory_access_type); | |
7961 | } | |
272c6793 | 7962 | |
c3a8f964 | 7963 | tree vec_mask = NULL_TREE; |
ebfd146a | 7964 | prev_stmt_info = NULL; |
4d694b27 | 7965 | poly_uint64 group_elt = 0; |
ebfd146a | 7966 | for (j = 0; j < ncopies; j++) |
b8698a0f | 7967 | { |
272c6793 | 7968 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 7969 | if (j == 0) |
74bf76ed JJ |
7970 | { |
7971 | bool simd_lane_access_p | |
7972 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
7973 | if (simd_lane_access_p | |
7974 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
7975 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
7976 | && integer_zerop (DR_OFFSET (first_dr)) | |
7977 | && integer_zerop (DR_INIT (first_dr)) | |
7978 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 7979 | get_alias_set (TREE_TYPE (ref_type))) |
74bf76ed JJ |
7980 | && (alignment_support_scheme == dr_aligned |
7981 | || alignment_support_scheme == dr_unaligned_supported)) | |
7982 | { | |
7983 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 7984 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 7985 | inv_p = false; |
74bf76ed | 7986 | } |
4f0a0218 RB |
7987 | else if (first_stmt_for_drptr |
7988 | && first_stmt != first_stmt_for_drptr) | |
7989 | { | |
7990 | dataref_ptr | |
7991 | = vect_create_data_ref_ptr (first_stmt_for_drptr, aggr_type, | |
7992 | at_loop, offset, &dummy, gsi, | |
7993 | &ptr_incr, simd_lane_access_p, | |
ab2fc782 | 7994 | &inv_p, byte_offset, bump); |
4f0a0218 RB |
7995 | /* Adjust the pointer by the difference to first_stmt. */ |
7996 | data_reference_p ptrdr | |
7997 | = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt_for_drptr)); | |
7998 | tree diff = fold_convert (sizetype, | |
7999 | size_binop (MINUS_EXPR, | |
8000 | DR_INIT (first_dr), | |
8001 | DR_INIT (ptrdr))); | |
8002 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
8003 | stmt, diff); | |
8004 | } | |
bfaa08b7 RS |
8005 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
8006 | { | |
8007 | vect_get_gather_scatter_ops (loop, stmt, &gs_info, | |
8008 | &dataref_ptr, &vec_offset); | |
8009 | inv_p = false; | |
8010 | } | |
74bf76ed JJ |
8011 | else |
8012 | dataref_ptr | |
8013 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
8014 | offset, &dummy, gsi, &ptr_incr, | |
356bbc4c | 8015 | simd_lane_access_p, &inv_p, |
ab2fc782 | 8016 | byte_offset, bump); |
c3a8f964 RS |
8017 | if (mask) |
8018 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, | |
8019 | mask_vectype); | |
74bf76ed | 8020 | } |
ebfd146a | 8021 | else |
c3a8f964 RS |
8022 | { |
8023 | if (dataref_offset) | |
8024 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
ab2fc782 | 8025 | bump); |
bfaa08b7 | 8026 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
929b4411 RS |
8027 | vec_offset = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
8028 | vec_offset); | |
c3a8f964 | 8029 | else |
ab2fc782 RS |
8030 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8031 | stmt, bump); | |
c3a8f964 | 8032 | if (mask) |
929b4411 | 8033 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
c3a8f964 | 8034 | } |
ebfd146a | 8035 | |
0d0293ac | 8036 | if (grouped_load || slp_perm) |
9771b263 | 8037 | dr_chain.create (vec_num); |
5ce1ee7f | 8038 | |
2de001ee | 8039 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 8040 | { |
272c6793 RS |
8041 | tree vec_array; |
8042 | ||
8043 | vec_array = create_vector_array (vectype, vec_num); | |
8044 | ||
7cfb4d93 | 8045 | tree final_mask = NULL_TREE; |
70088b95 RS |
8046 | if (loop_masks) |
8047 | final_mask = vect_get_loop_mask (gsi, loop_masks, ncopies, | |
8048 | vectype, j); | |
7cfb4d93 RS |
8049 | if (vec_mask) |
8050 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
8051 | vec_mask, gsi); | |
8052 | ||
7e11fc7f | 8053 | gcall *call; |
7cfb4d93 | 8054 | if (final_mask) |
7e11fc7f RS |
8055 | { |
8056 | /* Emit: | |
8057 | VEC_ARRAY = MASK_LOAD_LANES (DATAREF_PTR, ALIAS_PTR, | |
8058 | VEC_MASK). */ | |
8059 | unsigned int align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype)); | |
8060 | tree alias_ptr = build_int_cst (ref_type, align); | |
8061 | call = gimple_build_call_internal (IFN_MASK_LOAD_LANES, 3, | |
8062 | dataref_ptr, alias_ptr, | |
7cfb4d93 | 8063 | final_mask); |
7e11fc7f RS |
8064 | } |
8065 | else | |
8066 | { | |
8067 | /* Emit: | |
8068 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
8069 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); | |
8070 | call = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
8071 | } | |
a844293d RS |
8072 | gimple_call_set_lhs (call, vec_array); |
8073 | gimple_call_set_nothrow (call, true); | |
8074 | new_stmt = call; | |
272c6793 | 8075 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a | 8076 | |
272c6793 RS |
8077 | /* Extract each vector into an SSA_NAME. */ |
8078 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 8079 | { |
272c6793 RS |
8080 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
8081 | vec_array, i); | |
9771b263 | 8082 | dr_chain.quick_push (new_temp); |
272c6793 RS |
8083 | } |
8084 | ||
8085 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 8086 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
8087 | } |
8088 | else | |
8089 | { | |
8090 | for (i = 0; i < vec_num; i++) | |
8091 | { | |
7cfb4d93 | 8092 | tree final_mask = NULL_TREE; |
70088b95 | 8093 | if (loop_masks |
7cfb4d93 | 8094 | && memory_access_type != VMAT_INVARIANT) |
70088b95 RS |
8095 | final_mask = vect_get_loop_mask (gsi, loop_masks, |
8096 | vec_num * ncopies, | |
7cfb4d93 RS |
8097 | vectype, vec_num * j + i); |
8098 | if (vec_mask) | |
8099 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
8100 | vec_mask, gsi); | |
8101 | ||
272c6793 RS |
8102 | if (i > 0) |
8103 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
ab2fc782 | 8104 | stmt, bump); |
272c6793 RS |
8105 | |
8106 | /* 2. Create the vector-load in the loop. */ | |
8107 | switch (alignment_support_scheme) | |
8108 | { | |
8109 | case dr_aligned: | |
8110 | case dr_unaligned_supported: | |
be1ac4ec | 8111 | { |
644ffefd MJ |
8112 | unsigned int align, misalign; |
8113 | ||
bfaa08b7 RS |
8114 | if (memory_access_type == VMAT_GATHER_SCATTER) |
8115 | { | |
8116 | tree scale = size_int (gs_info.scale); | |
8117 | gcall *call; | |
70088b95 | 8118 | if (loop_masks) |
bfaa08b7 RS |
8119 | call = gimple_build_call_internal |
8120 | (IFN_MASK_GATHER_LOAD, 4, dataref_ptr, | |
8121 | vec_offset, scale, final_mask); | |
8122 | else | |
8123 | call = gimple_build_call_internal | |
8124 | (IFN_GATHER_LOAD, 3, dataref_ptr, | |
8125 | vec_offset, scale); | |
8126 | gimple_call_set_nothrow (call, true); | |
8127 | new_stmt = call; | |
8128 | data_ref = NULL_TREE; | |
8129 | break; | |
8130 | } | |
8131 | ||
f702e7d4 | 8132 | align = DR_TARGET_ALIGNMENT (dr); |
272c6793 RS |
8133 | if (alignment_support_scheme == dr_aligned) |
8134 | { | |
8135 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 8136 | misalign = 0; |
272c6793 RS |
8137 | } |
8138 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
8139 | { | |
25f68d90 | 8140 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 8141 | misalign = 0; |
272c6793 RS |
8142 | } |
8143 | else | |
c3a8f964 | 8144 | misalign = DR_MISALIGNMENT (first_dr); |
aed93b23 RB |
8145 | if (dataref_offset == NULL_TREE |
8146 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
8147 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
8148 | align, misalign); | |
c3a8f964 | 8149 | |
7cfb4d93 | 8150 | if (final_mask) |
c3a8f964 RS |
8151 | { |
8152 | align = least_bit_hwi (misalign | align); | |
8153 | tree ptr = build_int_cst (ref_type, align); | |
8154 | gcall *call | |
8155 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, | |
8156 | dataref_ptr, ptr, | |
7cfb4d93 | 8157 | final_mask); |
c3a8f964 RS |
8158 | gimple_call_set_nothrow (call, true); |
8159 | new_stmt = call; | |
8160 | data_ref = NULL_TREE; | |
8161 | } | |
8162 | else | |
8163 | { | |
8164 | data_ref | |
8165 | = fold_build2 (MEM_REF, vectype, dataref_ptr, | |
8166 | dataref_offset | |
8167 | ? dataref_offset | |
8168 | : build_int_cst (ref_type, 0)); | |
8169 | if (alignment_support_scheme == dr_aligned) | |
8170 | ; | |
8171 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
8172 | TREE_TYPE (data_ref) | |
8173 | = build_aligned_type (TREE_TYPE (data_ref), | |
8174 | align * BITS_PER_UNIT); | |
8175 | else | |
8176 | TREE_TYPE (data_ref) | |
8177 | = build_aligned_type (TREE_TYPE (data_ref), | |
8178 | TYPE_ALIGN (elem_type)); | |
8179 | } | |
272c6793 | 8180 | break; |
be1ac4ec | 8181 | } |
272c6793 | 8182 | case dr_explicit_realign: |
267d3070 | 8183 | { |
272c6793 | 8184 | tree ptr, bump; |
272c6793 | 8185 | |
d88981fc | 8186 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
8187 | |
8188 | if (compute_in_loop) | |
8189 | msq = vect_setup_realignment (first_stmt, gsi, | |
8190 | &realignment_token, | |
8191 | dr_explicit_realign, | |
8192 | dataref_ptr, NULL); | |
8193 | ||
aed93b23 RB |
8194 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
8195 | ptr = copy_ssa_name (dataref_ptr); | |
8196 | else | |
8197 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
f702e7d4 | 8198 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); |
0d0e4a03 JJ |
8199 | new_stmt = gimple_build_assign |
8200 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
8201 | build_int_cst |
8202 | (TREE_TYPE (dataref_ptr), | |
f702e7d4 | 8203 | -(HOST_WIDE_INT) align)); |
272c6793 RS |
8204 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
8205 | data_ref | |
8206 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 8207 | build_int_cst (ref_type, 0)); |
272c6793 RS |
8208 | vec_dest = vect_create_destination_var (scalar_dest, |
8209 | vectype); | |
8210 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
8211 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
8212 | gimple_assign_set_lhs (new_stmt, new_temp); | |
8213 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
8214 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
8215 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8216 | msq = new_temp; | |
8217 | ||
d88981fc | 8218 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 8219 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 8220 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 8221 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
8222 | new_stmt = gimple_build_assign |
8223 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 | 8224 | build_int_cst |
f702e7d4 | 8225 | (TREE_TYPE (ptr), -(HOST_WIDE_INT) align)); |
aed93b23 | 8226 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
8227 | gimple_assign_set_lhs (new_stmt, ptr); |
8228 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8229 | data_ref | |
8230 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 8231 | build_int_cst (ref_type, 0)); |
272c6793 | 8232 | break; |
267d3070 | 8233 | } |
272c6793 | 8234 | case dr_explicit_realign_optimized: |
f702e7d4 RS |
8235 | { |
8236 | if (TREE_CODE (dataref_ptr) == SSA_NAME) | |
8237 | new_temp = copy_ssa_name (dataref_ptr); | |
8238 | else | |
8239 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
8240 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); | |
8241 | new_stmt = gimple_build_assign | |
8242 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
8243 | build_int_cst (TREE_TYPE (dataref_ptr), | |
8244 | -(HOST_WIDE_INT) align)); | |
8245 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8246 | data_ref | |
8247 | = build2 (MEM_REF, vectype, new_temp, | |
8248 | build_int_cst (ref_type, 0)); | |
8249 | break; | |
8250 | } | |
272c6793 RS |
8251 | default: |
8252 | gcc_unreachable (); | |
8253 | } | |
ebfd146a | 8254 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
c3a8f964 RS |
8255 | /* DATA_REF is null if we've already built the statement. */ |
8256 | if (data_ref) | |
8257 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
ebfd146a | 8258 | new_temp = make_ssa_name (vec_dest, new_stmt); |
c3a8f964 | 8259 | gimple_set_lhs (new_stmt, new_temp); |
ebfd146a IR |
8260 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
8261 | ||
272c6793 RS |
8262 | /* 3. Handle explicit realignment if necessary/supported. |
8263 | Create in loop: | |
8264 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
8265 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
8266 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 8267 | { |
272c6793 RS |
8268 | lsq = gimple_assign_lhs (new_stmt); |
8269 | if (!realignment_token) | |
8270 | realignment_token = dataref_ptr; | |
8271 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
8272 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
8273 | msq, lsq, realignment_token); | |
272c6793 RS |
8274 | new_temp = make_ssa_name (vec_dest, new_stmt); |
8275 | gimple_assign_set_lhs (new_stmt, new_temp); | |
8276 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8277 | ||
8278 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
8279 | { | |
8280 | gcc_assert (phi); | |
8281 | if (i == vec_num - 1 && j == ncopies - 1) | |
8282 | add_phi_arg (phi, lsq, | |
8283 | loop_latch_edge (containing_loop), | |
9e227d60 | 8284 | UNKNOWN_LOCATION); |
272c6793 RS |
8285 | msq = lsq; |
8286 | } | |
ebfd146a | 8287 | } |
ebfd146a | 8288 | |
59fd17e3 RB |
8289 | /* 4. Handle invariant-load. */ |
8290 | if (inv_p && !bb_vinfo) | |
8291 | { | |
59fd17e3 | 8292 | gcc_assert (!grouped_load); |
d1417442 JJ |
8293 | /* If we have versioned for aliasing or the loop doesn't |
8294 | have any data dependencies that would preclude this, | |
8295 | then we are sure this is a loop invariant load and | |
8296 | thus we can insert it on the preheader edge. */ | |
8297 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
8298 | && !nested_in_vect_loop | |
6b916b36 | 8299 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
8300 | { |
8301 | if (dump_enabled_p ()) | |
8302 | { | |
8303 | dump_printf_loc (MSG_NOTE, vect_location, | |
8304 | "hoisting out of the vectorized " | |
8305 | "loop: "); | |
8306 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 8307 | } |
b731b390 | 8308 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
8309 | gsi_insert_on_edge_immediate |
8310 | (loop_preheader_edge (loop), | |
8311 | gimple_build_assign (tem, | |
8312 | unshare_expr | |
8313 | (gimple_assign_rhs1 (stmt)))); | |
8314 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
34cd48e5 RB |
8315 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
8316 | set_vinfo_for_stmt (new_stmt, | |
8317 | new_stmt_vec_info (new_stmt, vinfo)); | |
a0e35eb0 RB |
8318 | } |
8319 | else | |
8320 | { | |
8321 | gimple_stmt_iterator gsi2 = *gsi; | |
8322 | gsi_next (&gsi2); | |
8323 | new_temp = vect_init_vector (stmt, scalar_dest, | |
8324 | vectype, &gsi2); | |
34cd48e5 | 8325 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
a0e35eb0 | 8326 | } |
59fd17e3 RB |
8327 | } |
8328 | ||
62da9e14 | 8329 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
272c6793 | 8330 | { |
aec7ae7d JJ |
8331 | tree perm_mask = perm_mask_for_reverse (vectype); |
8332 | new_temp = permute_vec_elements (new_temp, new_temp, | |
8333 | perm_mask, stmt, gsi); | |
ebfd146a IR |
8334 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
8335 | } | |
267d3070 | 8336 | |
272c6793 | 8337 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
8338 | vect_transform_grouped_load (). */ |
8339 | if (grouped_load || slp_perm) | |
9771b263 | 8340 | dr_chain.quick_push (new_temp); |
267d3070 | 8341 | |
272c6793 RS |
8342 | /* Store vector loads in the corresponding SLP_NODE. */ |
8343 | if (slp && !slp_perm) | |
9771b263 | 8344 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
b267968e RB |
8345 | |
8346 | /* With SLP permutation we load the gaps as well, without | |
8347 | we need to skip the gaps after we manage to fully load | |
8348 | all elements. group_gap_adj is GROUP_SIZE here. */ | |
8349 | group_elt += nunits; | |
d9f21f6a RS |
8350 | if (maybe_ne (group_gap_adj, 0U) |
8351 | && !slp_perm | |
8352 | && known_eq (group_elt, group_size - group_gap_adj)) | |
b267968e | 8353 | { |
d9f21f6a RS |
8354 | poly_wide_int bump_val |
8355 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
8356 | * group_gap_adj); | |
8e6cdc90 | 8357 | tree bump = wide_int_to_tree (sizetype, bump_val); |
b267968e RB |
8358 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8359 | stmt, bump); | |
8360 | group_elt = 0; | |
8361 | } | |
272c6793 | 8362 | } |
9b999e8c RB |
8363 | /* Bump the vector pointer to account for a gap or for excess |
8364 | elements loaded for a permuted SLP load. */ | |
d9f21f6a | 8365 | if (maybe_ne (group_gap_adj, 0U) && slp_perm) |
a64b9c26 | 8366 | { |
d9f21f6a RS |
8367 | poly_wide_int bump_val |
8368 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
8369 | * group_gap_adj); | |
8e6cdc90 | 8370 | tree bump = wide_int_to_tree (sizetype, bump_val); |
a64b9c26 RB |
8371 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8372 | stmt, bump); | |
8373 | } | |
ebfd146a IR |
8374 | } |
8375 | ||
8376 | if (slp && !slp_perm) | |
8377 | continue; | |
8378 | ||
8379 | if (slp_perm) | |
8380 | { | |
29afecdf | 8381 | unsigned n_perms; |
01d8bf07 | 8382 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
29afecdf RB |
8383 | slp_node_instance, false, |
8384 | &n_perms)) | |
ebfd146a | 8385 | { |
9771b263 | 8386 | dr_chain.release (); |
ebfd146a IR |
8387 | return false; |
8388 | } | |
8389 | } | |
8390 | else | |
8391 | { | |
0d0293ac | 8392 | if (grouped_load) |
ebfd146a | 8393 | { |
2de001ee | 8394 | if (memory_access_type != VMAT_LOAD_STORE_LANES) |
0d0293ac | 8395 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 8396 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
8397 | } |
8398 | else | |
8399 | { | |
8400 | if (j == 0) | |
8401 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
8402 | else | |
8403 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
8404 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
8405 | } | |
8406 | } | |
9771b263 | 8407 | dr_chain.release (); |
ebfd146a IR |
8408 | } |
8409 | ||
ebfd146a IR |
8410 | return true; |
8411 | } | |
8412 | ||
8413 | /* Function vect_is_simple_cond. | |
b8698a0f | 8414 | |
ebfd146a IR |
8415 | Input: |
8416 | LOOP - the loop that is being vectorized. | |
8417 | COND - Condition that is checked for simple use. | |
8418 | ||
e9e1d143 RG |
8419 | Output: |
8420 | *COMP_VECTYPE - the vector type for the comparison. | |
4fc5ebf1 | 8421 | *DTS - The def types for the arguments of the comparison |
e9e1d143 | 8422 | |
ebfd146a IR |
8423 | Returns whether a COND can be vectorized. Checks whether |
8424 | condition operands are supportable using vec_is_simple_use. */ | |
8425 | ||
87aab9b2 | 8426 | static bool |
4fc5ebf1 | 8427 | vect_is_simple_cond (tree cond, vec_info *vinfo, |
8da4c8d8 RB |
8428 | tree *comp_vectype, enum vect_def_type *dts, |
8429 | tree vectype) | |
ebfd146a IR |
8430 | { |
8431 | tree lhs, rhs; | |
e9e1d143 | 8432 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 8433 | |
a414c77f IE |
8434 | /* Mask case. */ |
8435 | if (TREE_CODE (cond) == SSA_NAME | |
2568d8a1 | 8436 | && VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (cond))) |
a414c77f IE |
8437 | { |
8438 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (cond); | |
8439 | if (!vect_is_simple_use (cond, vinfo, &lhs_def_stmt, | |
4fc5ebf1 | 8440 | &dts[0], comp_vectype) |
a414c77f IE |
8441 | || !*comp_vectype |
8442 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
8443 | return false; | |
8444 | return true; | |
8445 | } | |
8446 | ||
ebfd146a IR |
8447 | if (!COMPARISON_CLASS_P (cond)) |
8448 | return false; | |
8449 | ||
8450 | lhs = TREE_OPERAND (cond, 0); | |
8451 | rhs = TREE_OPERAND (cond, 1); | |
8452 | ||
8453 | if (TREE_CODE (lhs) == SSA_NAME) | |
8454 | { | |
355fe088 | 8455 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); |
4fc5ebf1 | 8456 | if (!vect_is_simple_use (lhs, vinfo, &lhs_def_stmt, &dts[0], &vectype1)) |
ebfd146a IR |
8457 | return false; |
8458 | } | |
4fc5ebf1 JG |
8459 | else if (TREE_CODE (lhs) == INTEGER_CST || TREE_CODE (lhs) == REAL_CST |
8460 | || TREE_CODE (lhs) == FIXED_CST) | |
8461 | dts[0] = vect_constant_def; | |
8462 | else | |
ebfd146a IR |
8463 | return false; |
8464 | ||
8465 | if (TREE_CODE (rhs) == SSA_NAME) | |
8466 | { | |
355fe088 | 8467 | gimple *rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); |
4fc5ebf1 | 8468 | if (!vect_is_simple_use (rhs, vinfo, &rhs_def_stmt, &dts[1], &vectype2)) |
ebfd146a IR |
8469 | return false; |
8470 | } | |
4fc5ebf1 JG |
8471 | else if (TREE_CODE (rhs) == INTEGER_CST || TREE_CODE (rhs) == REAL_CST |
8472 | || TREE_CODE (rhs) == FIXED_CST) | |
8473 | dts[1] = vect_constant_def; | |
8474 | else | |
ebfd146a IR |
8475 | return false; |
8476 | ||
28b33016 | 8477 | if (vectype1 && vectype2 |
928686b1 RS |
8478 | && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1), |
8479 | TYPE_VECTOR_SUBPARTS (vectype2))) | |
28b33016 IE |
8480 | return false; |
8481 | ||
e9e1d143 | 8482 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
8da4c8d8 RB |
8483 | /* Invariant comparison. */ |
8484 | if (! *comp_vectype) | |
8485 | { | |
8486 | tree scalar_type = TREE_TYPE (lhs); | |
8487 | /* If we can widen the comparison to match vectype do so. */ | |
8488 | if (INTEGRAL_TYPE_P (scalar_type) | |
8489 | && tree_int_cst_lt (TYPE_SIZE (scalar_type), | |
8490 | TYPE_SIZE (TREE_TYPE (vectype)))) | |
8491 | scalar_type = build_nonstandard_integer_type | |
8492 | (tree_to_uhwi (TYPE_SIZE (TREE_TYPE (vectype))), | |
8493 | TYPE_UNSIGNED (scalar_type)); | |
8494 | *comp_vectype = get_vectype_for_scalar_type (scalar_type); | |
8495 | } | |
8496 | ||
ebfd146a IR |
8497 | return true; |
8498 | } | |
8499 | ||
8500 | /* vectorizable_condition. | |
8501 | ||
b8698a0f L |
8502 | Check if STMT is conditional modify expression that can be vectorized. |
8503 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
8504 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
8505 | at GSI. |
8506 | ||
8507 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
8508 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 8509 | else clause if it is 2). |
ebfd146a IR |
8510 | |
8511 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
8512 | ||
4bbe8262 | 8513 | bool |
355fe088 TS |
8514 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
8515 | gimple **vec_stmt, tree reduc_def, int reduc_index, | |
f7e531cf | 8516 | slp_tree slp_node) |
ebfd146a IR |
8517 | { |
8518 | tree scalar_dest = NULL_TREE; | |
8519 | tree vec_dest = NULL_TREE; | |
01216d27 JJ |
8520 | tree cond_expr, cond_expr0 = NULL_TREE, cond_expr1 = NULL_TREE; |
8521 | tree then_clause, else_clause; | |
ebfd146a | 8522 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
df11cc78 | 8523 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
8524 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
8525 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
5958f9e2 | 8526 | tree vec_compare; |
ebfd146a IR |
8527 | tree new_temp; |
8528 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4fc5ebf1 JG |
8529 | enum vect_def_type dts[4] |
8530 | = {vect_unknown_def_type, vect_unknown_def_type, | |
8531 | vect_unknown_def_type, vect_unknown_def_type}; | |
8532 | int ndts = 4; | |
f7e531cf | 8533 | int ncopies; |
01216d27 | 8534 | enum tree_code code, cond_code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
a855b1b1 | 8535 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
8536 | int i, j; |
8537 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
8538 | vec<tree> vec_oprnds0 = vNULL; |
8539 | vec<tree> vec_oprnds1 = vNULL; | |
8540 | vec<tree> vec_oprnds2 = vNULL; | |
8541 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 8542 | tree vec_cmp_type; |
a414c77f | 8543 | bool masked = false; |
b8698a0f | 8544 | |
f7e531cf IR |
8545 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
8546 | return false; | |
8547 | ||
bb6c2b68 RS |
8548 | vect_reduction_type reduction_type |
8549 | = STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info); | |
8550 | if (reduction_type == TREE_CODE_REDUCTION) | |
af29617a AH |
8551 | { |
8552 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
8553 | return false; | |
ebfd146a | 8554 | |
af29617a AH |
8555 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
8556 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
8557 | && reduc_def)) | |
8558 | return false; | |
ebfd146a | 8559 | |
af29617a AH |
8560 | /* FORNOW: not yet supported. */ |
8561 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
8562 | { | |
8563 | if (dump_enabled_p ()) | |
8564 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
8565 | "value used after loop.\n"); | |
8566 | return false; | |
8567 | } | |
ebfd146a IR |
8568 | } |
8569 | ||
8570 | /* Is vectorizable conditional operation? */ | |
8571 | if (!is_gimple_assign (stmt)) | |
8572 | return false; | |
8573 | ||
8574 | code = gimple_assign_rhs_code (stmt); | |
8575 | ||
8576 | if (code != COND_EXPR) | |
8577 | return false; | |
8578 | ||
465c8c19 | 8579 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2947d3b2 | 8580 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
465c8c19 | 8581 | |
fce57248 | 8582 | if (slp_node) |
465c8c19 JJ |
8583 | ncopies = 1; |
8584 | else | |
e8f142e2 | 8585 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
8586 | |
8587 | gcc_assert (ncopies >= 1); | |
8588 | if (reduc_index && ncopies > 1) | |
8589 | return false; /* FORNOW */ | |
8590 | ||
4e71066d RG |
8591 | cond_expr = gimple_assign_rhs1 (stmt); |
8592 | then_clause = gimple_assign_rhs2 (stmt); | |
8593 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 8594 | |
4fc5ebf1 | 8595 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, |
8da4c8d8 | 8596 | &comp_vectype, &dts[0], vectype) |
e9e1d143 | 8597 | || !comp_vectype) |
ebfd146a IR |
8598 | return false; |
8599 | ||
81c40241 | 8600 | gimple *def_stmt; |
4fc5ebf1 | 8601 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &def_stmt, &dts[2], |
2947d3b2 IE |
8602 | &vectype1)) |
8603 | return false; | |
4fc5ebf1 | 8604 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &def_stmt, &dts[3], |
2947d3b2 | 8605 | &vectype2)) |
ebfd146a | 8606 | return false; |
2947d3b2 IE |
8607 | |
8608 | if (vectype1 && !useless_type_conversion_p (vectype, vectype1)) | |
8609 | return false; | |
8610 | ||
8611 | if (vectype2 && !useless_type_conversion_p (vectype, vectype2)) | |
ebfd146a IR |
8612 | return false; |
8613 | ||
28b33016 IE |
8614 | masked = !COMPARISON_CLASS_P (cond_expr); |
8615 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
8616 | ||
74946978 MP |
8617 | if (vec_cmp_type == NULL_TREE) |
8618 | return false; | |
784fb9b3 | 8619 | |
01216d27 JJ |
8620 | cond_code = TREE_CODE (cond_expr); |
8621 | if (!masked) | |
8622 | { | |
8623 | cond_expr0 = TREE_OPERAND (cond_expr, 0); | |
8624 | cond_expr1 = TREE_OPERAND (cond_expr, 1); | |
8625 | } | |
8626 | ||
8627 | if (!masked && VECTOR_BOOLEAN_TYPE_P (comp_vectype)) | |
8628 | { | |
8629 | /* Boolean values may have another representation in vectors | |
8630 | and therefore we prefer bit operations over comparison for | |
8631 | them (which also works for scalar masks). We store opcodes | |
8632 | to use in bitop1 and bitop2. Statement is vectorized as | |
8633 | BITOP2 (rhs1 BITOP1 rhs2) or rhs1 BITOP2 (BITOP1 rhs2) | |
8634 | depending on bitop1 and bitop2 arity. */ | |
8635 | switch (cond_code) | |
8636 | { | |
8637 | case GT_EXPR: | |
8638 | bitop1 = BIT_NOT_EXPR; | |
8639 | bitop2 = BIT_AND_EXPR; | |
8640 | break; | |
8641 | case GE_EXPR: | |
8642 | bitop1 = BIT_NOT_EXPR; | |
8643 | bitop2 = BIT_IOR_EXPR; | |
8644 | break; | |
8645 | case LT_EXPR: | |
8646 | bitop1 = BIT_NOT_EXPR; | |
8647 | bitop2 = BIT_AND_EXPR; | |
8648 | std::swap (cond_expr0, cond_expr1); | |
8649 | break; | |
8650 | case LE_EXPR: | |
8651 | bitop1 = BIT_NOT_EXPR; | |
8652 | bitop2 = BIT_IOR_EXPR; | |
8653 | std::swap (cond_expr0, cond_expr1); | |
8654 | break; | |
8655 | case NE_EXPR: | |
8656 | bitop1 = BIT_XOR_EXPR; | |
8657 | break; | |
8658 | case EQ_EXPR: | |
8659 | bitop1 = BIT_XOR_EXPR; | |
8660 | bitop2 = BIT_NOT_EXPR; | |
8661 | break; | |
8662 | default: | |
8663 | return false; | |
8664 | } | |
8665 | cond_code = SSA_NAME; | |
8666 | } | |
8667 | ||
b8698a0f | 8668 | if (!vec_stmt) |
ebfd146a IR |
8669 | { |
8670 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
01216d27 JJ |
8671 | if (bitop1 != NOP_EXPR) |
8672 | { | |
8673 | machine_mode mode = TYPE_MODE (comp_vectype); | |
8674 | optab optab; | |
8675 | ||
8676 | optab = optab_for_tree_code (bitop1, comp_vectype, optab_default); | |
8677 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8678 | return false; | |
8679 | ||
8680 | if (bitop2 != NOP_EXPR) | |
8681 | { | |
8682 | optab = optab_for_tree_code (bitop2, comp_vectype, | |
8683 | optab_default); | |
8684 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8685 | return false; | |
8686 | } | |
8687 | } | |
4fc5ebf1 JG |
8688 | if (expand_vec_cond_expr_p (vectype, comp_vectype, |
8689 | cond_code)) | |
8690 | { | |
78604de0 RB |
8691 | if (!slp_node) |
8692 | vect_model_simple_cost (stmt_info, ncopies, dts, ndts, NULL, NULL); | |
4fc5ebf1 JG |
8693 | return true; |
8694 | } | |
8695 | return false; | |
ebfd146a IR |
8696 | } |
8697 | ||
f7e531cf IR |
8698 | /* Transform. */ |
8699 | ||
8700 | if (!slp_node) | |
8701 | { | |
9771b263 DN |
8702 | vec_oprnds0.create (1); |
8703 | vec_oprnds1.create (1); | |
8704 | vec_oprnds2.create (1); | |
8705 | vec_oprnds3.create (1); | |
f7e531cf | 8706 | } |
ebfd146a IR |
8707 | |
8708 | /* Handle def. */ | |
8709 | scalar_dest = gimple_assign_lhs (stmt); | |
bb6c2b68 RS |
8710 | if (reduction_type != EXTRACT_LAST_REDUCTION) |
8711 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
ebfd146a IR |
8712 | |
8713 | /* Handle cond expr. */ | |
a855b1b1 MM |
8714 | for (j = 0; j < ncopies; j++) |
8715 | { | |
bb6c2b68 | 8716 | gimple *new_stmt = NULL; |
a855b1b1 MM |
8717 | if (j == 0) |
8718 | { | |
f7e531cf IR |
8719 | if (slp_node) |
8720 | { | |
00f96dc9 TS |
8721 | auto_vec<tree, 4> ops; |
8722 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 8723 | |
a414c77f | 8724 | if (masked) |
01216d27 | 8725 | ops.safe_push (cond_expr); |
a414c77f IE |
8726 | else |
8727 | { | |
01216d27 JJ |
8728 | ops.safe_push (cond_expr0); |
8729 | ops.safe_push (cond_expr1); | |
a414c77f | 8730 | } |
9771b263 DN |
8731 | ops.safe_push (then_clause); |
8732 | ops.safe_push (else_clause); | |
306b0c92 | 8733 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
37b5ec8f JJ |
8734 | vec_oprnds3 = vec_defs.pop (); |
8735 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
8736 | if (!masked) |
8737 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 8738 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf IR |
8739 | } |
8740 | else | |
8741 | { | |
355fe088 | 8742 | gimple *gtemp; |
a414c77f IE |
8743 | if (masked) |
8744 | { | |
8745 | vec_cond_lhs | |
8746 | = vect_get_vec_def_for_operand (cond_expr, stmt, | |
8747 | comp_vectype); | |
8748 | vect_is_simple_use (cond_expr, stmt_info->vinfo, | |
8749 | >emp, &dts[0]); | |
8750 | } | |
8751 | else | |
8752 | { | |
01216d27 JJ |
8753 | vec_cond_lhs |
8754 | = vect_get_vec_def_for_operand (cond_expr0, | |
8755 | stmt, comp_vectype); | |
8756 | vect_is_simple_use (cond_expr0, loop_vinfo, >emp, &dts[0]); | |
8757 | ||
8758 | vec_cond_rhs | |
8759 | = vect_get_vec_def_for_operand (cond_expr1, | |
8760 | stmt, comp_vectype); | |
8761 | vect_is_simple_use (cond_expr1, loop_vinfo, >emp, &dts[1]); | |
a414c77f | 8762 | } |
f7e531cf IR |
8763 | if (reduc_index == 1) |
8764 | vec_then_clause = reduc_def; | |
8765 | else | |
8766 | { | |
8767 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 RB |
8768 | stmt); |
8769 | vect_is_simple_use (then_clause, loop_vinfo, | |
8770 | >emp, &dts[2]); | |
f7e531cf IR |
8771 | } |
8772 | if (reduc_index == 2) | |
8773 | vec_else_clause = reduc_def; | |
8774 | else | |
8775 | { | |
8776 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 RB |
8777 | stmt); |
8778 | vect_is_simple_use (else_clause, loop_vinfo, >emp, &dts[3]); | |
f7e531cf | 8779 | } |
a855b1b1 MM |
8780 | } |
8781 | } | |
8782 | else | |
8783 | { | |
a414c77f IE |
8784 | vec_cond_lhs |
8785 | = vect_get_vec_def_for_stmt_copy (dts[0], | |
8786 | vec_oprnds0.pop ()); | |
8787 | if (!masked) | |
8788 | vec_cond_rhs | |
8789 | = vect_get_vec_def_for_stmt_copy (dts[1], | |
8790 | vec_oprnds1.pop ()); | |
8791 | ||
a855b1b1 | 8792 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 8793 | vec_oprnds2.pop ()); |
a855b1b1 | 8794 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 8795 | vec_oprnds3.pop ()); |
f7e531cf IR |
8796 | } |
8797 | ||
8798 | if (!slp_node) | |
8799 | { | |
9771b263 | 8800 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
8801 | if (!masked) |
8802 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
8803 | vec_oprnds2.quick_push (vec_then_clause); |
8804 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
8805 | } |
8806 | ||
9dc3f7de | 8807 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 8808 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 8809 | { |
9771b263 DN |
8810 | vec_then_clause = vec_oprnds2[i]; |
8811 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 8812 | |
a414c77f IE |
8813 | if (masked) |
8814 | vec_compare = vec_cond_lhs; | |
8815 | else | |
8816 | { | |
8817 | vec_cond_rhs = vec_oprnds1[i]; | |
01216d27 JJ |
8818 | if (bitop1 == NOP_EXPR) |
8819 | vec_compare = build2 (cond_code, vec_cmp_type, | |
8820 | vec_cond_lhs, vec_cond_rhs); | |
8821 | else | |
8822 | { | |
8823 | new_temp = make_ssa_name (vec_cmp_type); | |
8824 | if (bitop1 == BIT_NOT_EXPR) | |
8825 | new_stmt = gimple_build_assign (new_temp, bitop1, | |
8826 | vec_cond_rhs); | |
8827 | else | |
8828 | new_stmt | |
8829 | = gimple_build_assign (new_temp, bitop1, vec_cond_lhs, | |
8830 | vec_cond_rhs); | |
8831 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8832 | if (bitop2 == NOP_EXPR) | |
8833 | vec_compare = new_temp; | |
8834 | else if (bitop2 == BIT_NOT_EXPR) | |
8835 | { | |
8836 | /* Instead of doing ~x ? y : z do x ? z : y. */ | |
8837 | vec_compare = new_temp; | |
8838 | std::swap (vec_then_clause, vec_else_clause); | |
8839 | } | |
8840 | else | |
8841 | { | |
8842 | vec_compare = make_ssa_name (vec_cmp_type); | |
8843 | new_stmt | |
8844 | = gimple_build_assign (vec_compare, bitop2, | |
8845 | vec_cond_lhs, new_temp); | |
8846 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8847 | } | |
8848 | } | |
a414c77f | 8849 | } |
bb6c2b68 RS |
8850 | if (reduction_type == EXTRACT_LAST_REDUCTION) |
8851 | { | |
8852 | if (!is_gimple_val (vec_compare)) | |
8853 | { | |
8854 | tree vec_compare_name = make_ssa_name (vec_cmp_type); | |
8855 | new_stmt = gimple_build_assign (vec_compare_name, | |
8856 | vec_compare); | |
8857 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8858 | vec_compare = vec_compare_name; | |
8859 | } | |
8860 | gcc_assert (reduc_index == 2); | |
8861 | new_stmt = gimple_build_call_internal | |
8862 | (IFN_FOLD_EXTRACT_LAST, 3, else_clause, vec_compare, | |
8863 | vec_then_clause); | |
8864 | gimple_call_set_lhs (new_stmt, scalar_dest); | |
8865 | SSA_NAME_DEF_STMT (scalar_dest) = new_stmt; | |
8866 | if (stmt == gsi_stmt (*gsi)) | |
8867 | vect_finish_replace_stmt (stmt, new_stmt); | |
8868 | else | |
8869 | { | |
8870 | /* In this case we're moving the definition to later in the | |
8871 | block. That doesn't matter because the only uses of the | |
8872 | lhs are in phi statements. */ | |
8873 | gimple_stmt_iterator old_gsi = gsi_for_stmt (stmt); | |
8874 | gsi_remove (&old_gsi, true); | |
8875 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8876 | } | |
8877 | } | |
8878 | else | |
8879 | { | |
8880 | new_temp = make_ssa_name (vec_dest); | |
8881 | new_stmt = gimple_build_assign (new_temp, VEC_COND_EXPR, | |
8882 | vec_compare, vec_then_clause, | |
8883 | vec_else_clause); | |
8884 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8885 | } | |
f7e531cf | 8886 | if (slp_node) |
9771b263 | 8887 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
8888 | } |
8889 | ||
8890 | if (slp_node) | |
8891 | continue; | |
8892 | ||
8893 | if (j == 0) | |
8894 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
8895 | else | |
8896 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
8897 | ||
8898 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 8899 | } |
b8698a0f | 8900 | |
9771b263 DN |
8901 | vec_oprnds0.release (); |
8902 | vec_oprnds1.release (); | |
8903 | vec_oprnds2.release (); | |
8904 | vec_oprnds3.release (); | |
f7e531cf | 8905 | |
ebfd146a IR |
8906 | return true; |
8907 | } | |
8908 | ||
42fd8198 IE |
8909 | /* vectorizable_comparison. |
8910 | ||
8911 | Check if STMT is comparison expression that can be vectorized. | |
8912 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
8913 | comparison, put it in VEC_STMT, and insert it at GSI. | |
8914 | ||
8915 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
8916 | ||
fce57248 | 8917 | static bool |
42fd8198 IE |
8918 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, |
8919 | gimple **vec_stmt, tree reduc_def, | |
8920 | slp_tree slp_node) | |
8921 | { | |
8922 | tree lhs, rhs1, rhs2; | |
8923 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8924 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
8925 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
8926 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
8927 | tree new_temp; | |
8928 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
8929 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 8930 | int ndts = 2; |
928686b1 | 8931 | poly_uint64 nunits; |
42fd8198 | 8932 | int ncopies; |
49e76ff1 | 8933 | enum tree_code code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
42fd8198 IE |
8934 | stmt_vec_info prev_stmt_info = NULL; |
8935 | int i, j; | |
8936 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
8937 | vec<tree> vec_oprnds0 = vNULL; | |
8938 | vec<tree> vec_oprnds1 = vNULL; | |
8939 | gimple *def_stmt; | |
8940 | tree mask_type; | |
8941 | tree mask; | |
8942 | ||
c245362b IE |
8943 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
8944 | return false; | |
8945 | ||
30480bcd | 8946 | if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)) |
42fd8198 IE |
8947 | return false; |
8948 | ||
8949 | mask_type = vectype; | |
8950 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
8951 | ||
fce57248 | 8952 | if (slp_node) |
42fd8198 IE |
8953 | ncopies = 1; |
8954 | else | |
e8f142e2 | 8955 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
42fd8198 IE |
8956 | |
8957 | gcc_assert (ncopies >= 1); | |
42fd8198 IE |
8958 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
8959 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
8960 | && reduc_def)) | |
8961 | return false; | |
8962 | ||
8963 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
8964 | { | |
8965 | if (dump_enabled_p ()) | |
8966 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
8967 | "value used after loop.\n"); | |
8968 | return false; | |
8969 | } | |
8970 | ||
8971 | if (!is_gimple_assign (stmt)) | |
8972 | return false; | |
8973 | ||
8974 | code = gimple_assign_rhs_code (stmt); | |
8975 | ||
8976 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
8977 | return false; | |
8978 | ||
8979 | rhs1 = gimple_assign_rhs1 (stmt); | |
8980 | rhs2 = gimple_assign_rhs2 (stmt); | |
8981 | ||
8982 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &def_stmt, | |
8983 | &dts[0], &vectype1)) | |
8984 | return false; | |
8985 | ||
8986 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &def_stmt, | |
8987 | &dts[1], &vectype2)) | |
8988 | return false; | |
8989 | ||
8990 | if (vectype1 && vectype2 | |
928686b1 RS |
8991 | && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1), |
8992 | TYPE_VECTOR_SUBPARTS (vectype2))) | |
42fd8198 IE |
8993 | return false; |
8994 | ||
8995 | vectype = vectype1 ? vectype1 : vectype2; | |
8996 | ||
8997 | /* Invariant comparison. */ | |
8998 | if (!vectype) | |
8999 | { | |
69a9a66f | 9000 | vectype = get_vectype_for_scalar_type (TREE_TYPE (rhs1)); |
928686b1 | 9001 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (vectype), nunits)) |
42fd8198 IE |
9002 | return false; |
9003 | } | |
928686b1 | 9004 | else if (maybe_ne (nunits, TYPE_VECTOR_SUBPARTS (vectype))) |
42fd8198 IE |
9005 | return false; |
9006 | ||
49e76ff1 IE |
9007 | /* Can't compare mask and non-mask types. */ |
9008 | if (vectype1 && vectype2 | |
9009 | && (VECTOR_BOOLEAN_TYPE_P (vectype1) ^ VECTOR_BOOLEAN_TYPE_P (vectype2))) | |
9010 | return false; | |
9011 | ||
9012 | /* Boolean values may have another representation in vectors | |
9013 | and therefore we prefer bit operations over comparison for | |
9014 | them (which also works for scalar masks). We store opcodes | |
9015 | to use in bitop1 and bitop2. Statement is vectorized as | |
9016 | BITOP2 (rhs1 BITOP1 rhs2) or | |
9017 | rhs1 BITOP2 (BITOP1 rhs2) | |
9018 | depending on bitop1 and bitop2 arity. */ | |
9019 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) | |
9020 | { | |
9021 | if (code == GT_EXPR) | |
9022 | { | |
9023 | bitop1 = BIT_NOT_EXPR; | |
9024 | bitop2 = BIT_AND_EXPR; | |
9025 | } | |
9026 | else if (code == GE_EXPR) | |
9027 | { | |
9028 | bitop1 = BIT_NOT_EXPR; | |
9029 | bitop2 = BIT_IOR_EXPR; | |
9030 | } | |
9031 | else if (code == LT_EXPR) | |
9032 | { | |
9033 | bitop1 = BIT_NOT_EXPR; | |
9034 | bitop2 = BIT_AND_EXPR; | |
9035 | std::swap (rhs1, rhs2); | |
264d951a | 9036 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
9037 | } |
9038 | else if (code == LE_EXPR) | |
9039 | { | |
9040 | bitop1 = BIT_NOT_EXPR; | |
9041 | bitop2 = BIT_IOR_EXPR; | |
9042 | std::swap (rhs1, rhs2); | |
264d951a | 9043 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
9044 | } |
9045 | else | |
9046 | { | |
9047 | bitop1 = BIT_XOR_EXPR; | |
9048 | if (code == EQ_EXPR) | |
9049 | bitop2 = BIT_NOT_EXPR; | |
9050 | } | |
9051 | } | |
9052 | ||
42fd8198 IE |
9053 | if (!vec_stmt) |
9054 | { | |
9055 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
78604de0 RB |
9056 | if (!slp_node) |
9057 | vect_model_simple_cost (stmt_info, ncopies * (1 + (bitop2 != NOP_EXPR)), | |
9058 | dts, ndts, NULL, NULL); | |
49e76ff1 | 9059 | if (bitop1 == NOP_EXPR) |
96592eed | 9060 | return expand_vec_cmp_expr_p (vectype, mask_type, code); |
49e76ff1 IE |
9061 | else |
9062 | { | |
9063 | machine_mode mode = TYPE_MODE (vectype); | |
9064 | optab optab; | |
9065 | ||
9066 | optab = optab_for_tree_code (bitop1, vectype, optab_default); | |
9067 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
9068 | return false; | |
9069 | ||
9070 | if (bitop2 != NOP_EXPR) | |
9071 | { | |
9072 | optab = optab_for_tree_code (bitop2, vectype, optab_default); | |
9073 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
9074 | return false; | |
9075 | } | |
9076 | return true; | |
9077 | } | |
42fd8198 IE |
9078 | } |
9079 | ||
9080 | /* Transform. */ | |
9081 | if (!slp_node) | |
9082 | { | |
9083 | vec_oprnds0.create (1); | |
9084 | vec_oprnds1.create (1); | |
9085 | } | |
9086 | ||
9087 | /* Handle def. */ | |
9088 | lhs = gimple_assign_lhs (stmt); | |
9089 | mask = vect_create_destination_var (lhs, mask_type); | |
9090 | ||
9091 | /* Handle cmp expr. */ | |
9092 | for (j = 0; j < ncopies; j++) | |
9093 | { | |
9094 | gassign *new_stmt = NULL; | |
9095 | if (j == 0) | |
9096 | { | |
9097 | if (slp_node) | |
9098 | { | |
9099 | auto_vec<tree, 2> ops; | |
9100 | auto_vec<vec<tree>, 2> vec_defs; | |
9101 | ||
9102 | ops.safe_push (rhs1); | |
9103 | ops.safe_push (rhs2); | |
306b0c92 | 9104 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
42fd8198 IE |
9105 | vec_oprnds1 = vec_defs.pop (); |
9106 | vec_oprnds0 = vec_defs.pop (); | |
9107 | } | |
9108 | else | |
9109 | { | |
e4af0bc4 IE |
9110 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, vectype); |
9111 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, vectype); | |
42fd8198 IE |
9112 | } |
9113 | } | |
9114 | else | |
9115 | { | |
9116 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
9117 | vec_oprnds0.pop ()); | |
9118 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
9119 | vec_oprnds1.pop ()); | |
9120 | } | |
9121 | ||
9122 | if (!slp_node) | |
9123 | { | |
9124 | vec_oprnds0.quick_push (vec_rhs1); | |
9125 | vec_oprnds1.quick_push (vec_rhs2); | |
9126 | } | |
9127 | ||
9128 | /* Arguments are ready. Create the new vector stmt. */ | |
9129 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
9130 | { | |
9131 | vec_rhs2 = vec_oprnds1[i]; | |
9132 | ||
9133 | new_temp = make_ssa_name (mask); | |
49e76ff1 IE |
9134 | if (bitop1 == NOP_EXPR) |
9135 | { | |
9136 | new_stmt = gimple_build_assign (new_temp, code, | |
9137 | vec_rhs1, vec_rhs2); | |
9138 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9139 | } | |
9140 | else | |
9141 | { | |
9142 | if (bitop1 == BIT_NOT_EXPR) | |
9143 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs2); | |
9144 | else | |
9145 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs1, | |
9146 | vec_rhs2); | |
9147 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9148 | if (bitop2 != NOP_EXPR) | |
9149 | { | |
9150 | tree res = make_ssa_name (mask); | |
9151 | if (bitop2 == BIT_NOT_EXPR) | |
9152 | new_stmt = gimple_build_assign (res, bitop2, new_temp); | |
9153 | else | |
9154 | new_stmt = gimple_build_assign (res, bitop2, vec_rhs1, | |
9155 | new_temp); | |
9156 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9157 | } | |
9158 | } | |
42fd8198 IE |
9159 | if (slp_node) |
9160 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
9161 | } | |
9162 | ||
9163 | if (slp_node) | |
9164 | continue; | |
9165 | ||
9166 | if (j == 0) | |
9167 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
9168 | else | |
9169 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
9170 | ||
9171 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
9172 | } | |
9173 | ||
9174 | vec_oprnds0.release (); | |
9175 | vec_oprnds1.release (); | |
9176 | ||
9177 | return true; | |
9178 | } | |
ebfd146a | 9179 | |
68a0f2ff RS |
9180 | /* If SLP_NODE is nonnull, return true if vectorizable_live_operation |
9181 | can handle all live statements in the node. Otherwise return true | |
9182 | if STMT is not live or if vectorizable_live_operation can handle it. | |
9183 | GSI and VEC_STMT are as for vectorizable_live_operation. */ | |
9184 | ||
9185 | static bool | |
9186 | can_vectorize_live_stmts (gimple *stmt, gimple_stmt_iterator *gsi, | |
9187 | slp_tree slp_node, gimple **vec_stmt) | |
9188 | { | |
9189 | if (slp_node) | |
9190 | { | |
9191 | gimple *slp_stmt; | |
9192 | unsigned int i; | |
9193 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, slp_stmt) | |
9194 | { | |
9195 | stmt_vec_info slp_stmt_info = vinfo_for_stmt (slp_stmt); | |
9196 | if (STMT_VINFO_LIVE_P (slp_stmt_info) | |
9197 | && !vectorizable_live_operation (slp_stmt, gsi, slp_node, i, | |
9198 | vec_stmt)) | |
9199 | return false; | |
9200 | } | |
9201 | } | |
9202 | else if (STMT_VINFO_LIVE_P (vinfo_for_stmt (stmt)) | |
9203 | && !vectorizable_live_operation (stmt, gsi, slp_node, -1, vec_stmt)) | |
9204 | return false; | |
9205 | ||
9206 | return true; | |
9207 | } | |
9208 | ||
8644a673 | 9209 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
9210 | |
9211 | bool | |
891ad31c RB |
9212 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node, |
9213 | slp_instance node_instance) | |
ebfd146a | 9214 | { |
8644a673 | 9215 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 9216 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 9217 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 9218 | bool ok; |
355fe088 | 9219 | gimple *pattern_stmt; |
363477c0 | 9220 | gimple_seq pattern_def_seq; |
ebfd146a | 9221 | |
73fbfcad | 9222 | if (dump_enabled_p ()) |
ebfd146a | 9223 | { |
78c60e3d SS |
9224 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
9225 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 9226 | } |
ebfd146a | 9227 | |
1825a1f3 | 9228 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 9229 | { |
73fbfcad | 9230 | if (dump_enabled_p ()) |
78c60e3d | 9231 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9232 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
9233 | |
9234 | return false; | |
9235 | } | |
b8698a0f L |
9236 | |
9237 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
9238 | to include: |
9239 | - the COND_EXPR which is the loop exit condition | |
9240 | - any LABEL_EXPRs in the loop | |
b8698a0f | 9241 | - computations that are used only for array indexing or loop control. |
8644a673 | 9242 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 9243 | instance, therefore, all the statements are relevant. |
ebfd146a | 9244 | |
d092494c | 9245 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 9246 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
9247 | statements. In basic blocks we are called from some SLP instance |
9248 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
9249 | already will be part of SLP instance. */ | |
83197f37 IR |
9250 | |
9251 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 9252 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 9253 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 9254 | { |
9d5e7640 | 9255 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 9256 | && pattern_stmt |
9d5e7640 IR |
9257 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
9258 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
9259 | { | |
83197f37 | 9260 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
9261 | stmt = pattern_stmt; |
9262 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 9263 | if (dump_enabled_p ()) |
9d5e7640 | 9264 | { |
78c60e3d SS |
9265 | dump_printf_loc (MSG_NOTE, vect_location, |
9266 | "==> examining pattern statement: "); | |
9267 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
9268 | } |
9269 | } | |
9270 | else | |
9271 | { | |
73fbfcad | 9272 | if (dump_enabled_p ()) |
e645e942 | 9273 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 9274 | |
9d5e7640 IR |
9275 | return true; |
9276 | } | |
8644a673 | 9277 | } |
83197f37 | 9278 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 9279 | && node == NULL |
83197f37 IR |
9280 | && pattern_stmt |
9281 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
9282 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
9283 | { | |
9284 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 9285 | if (dump_enabled_p ()) |
83197f37 | 9286 | { |
78c60e3d SS |
9287 | dump_printf_loc (MSG_NOTE, vect_location, |
9288 | "==> examining pattern statement: "); | |
9289 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
9290 | } |
9291 | ||
891ad31c RB |
9292 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node, |
9293 | node_instance)) | |
83197f37 IR |
9294 | return false; |
9295 | } | |
ebfd146a | 9296 | |
1107f3ae | 9297 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 9298 | && node == NULL |
363477c0 | 9299 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 9300 | { |
363477c0 | 9301 | gimple_stmt_iterator si; |
1107f3ae | 9302 | |
363477c0 JJ |
9303 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
9304 | { | |
355fe088 | 9305 | gimple *pattern_def_stmt = gsi_stmt (si); |
363477c0 JJ |
9306 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) |
9307 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
9308 | { | |
9309 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 9310 | if (dump_enabled_p ()) |
363477c0 | 9311 | { |
78c60e3d SS |
9312 | dump_printf_loc (MSG_NOTE, vect_location, |
9313 | "==> examining pattern def statement: "); | |
9314 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 9315 | } |
1107f3ae | 9316 | |
363477c0 | 9317 | if (!vect_analyze_stmt (pattern_def_stmt, |
891ad31c | 9318 | need_to_vectorize, node, node_instance)) |
363477c0 JJ |
9319 | return false; |
9320 | } | |
9321 | } | |
9322 | } | |
1107f3ae | 9323 | |
8644a673 IR |
9324 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
9325 | { | |
9326 | case vect_internal_def: | |
9327 | break; | |
ebfd146a | 9328 | |
8644a673 | 9329 | case vect_reduction_def: |
7c5222ff | 9330 | case vect_nested_cycle: |
14a61437 RB |
9331 | gcc_assert (!bb_vinfo |
9332 | && (relevance == vect_used_in_outer | |
9333 | || relevance == vect_used_in_outer_by_reduction | |
9334 | || relevance == vect_used_by_reduction | |
b28ead45 AH |
9335 | || relevance == vect_unused_in_scope |
9336 | || relevance == vect_used_only_live)); | |
8644a673 IR |
9337 | break; |
9338 | ||
9339 | case vect_induction_def: | |
e7baeb39 RB |
9340 | gcc_assert (!bb_vinfo); |
9341 | break; | |
9342 | ||
8644a673 IR |
9343 | case vect_constant_def: |
9344 | case vect_external_def: | |
9345 | case vect_unknown_def_type: | |
9346 | default: | |
9347 | gcc_unreachable (); | |
9348 | } | |
ebfd146a | 9349 | |
8644a673 | 9350 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 9351 | { |
8644a673 | 9352 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
9353 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
9354 | || (is_gimple_call (stmt) | |
9355 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 9356 | *need_to_vectorize = true; |
ebfd146a IR |
9357 | } |
9358 | ||
b1af7da6 RB |
9359 | if (PURE_SLP_STMT (stmt_info) && !node) |
9360 | { | |
9361 | dump_printf_loc (MSG_NOTE, vect_location, | |
9362 | "handled only by SLP analysis\n"); | |
9363 | return true; | |
9364 | } | |
9365 | ||
9366 | ok = true; | |
9367 | if (!bb_vinfo | |
9368 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
9369 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
9370 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
9371 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
9372 | || vectorizable_shift (stmt, NULL, NULL, node) | |
9373 | || vectorizable_operation (stmt, NULL, NULL, node) | |
9374 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
9375 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
9376 | || vectorizable_call (stmt, NULL, NULL, node) | |
9377 | || vectorizable_store (stmt, NULL, NULL, node) | |
891ad31c | 9378 | || vectorizable_reduction (stmt, NULL, NULL, node, node_instance) |
e7baeb39 | 9379 | || vectorizable_induction (stmt, NULL, NULL, node) |
42fd8198 IE |
9380 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
9381 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 RB |
9382 | else |
9383 | { | |
9384 | if (bb_vinfo) | |
9385 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
9386 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
9387 | || vectorizable_shift (stmt, NULL, NULL, node) | |
9388 | || vectorizable_operation (stmt, NULL, NULL, node) | |
9389 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
9390 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
9391 | || vectorizable_call (stmt, NULL, NULL, node) | |
9392 | || vectorizable_store (stmt, NULL, NULL, node) | |
42fd8198 IE |
9393 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
9394 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 | 9395 | } |
8644a673 IR |
9396 | |
9397 | if (!ok) | |
ebfd146a | 9398 | { |
73fbfcad | 9399 | if (dump_enabled_p ()) |
8644a673 | 9400 | { |
78c60e3d SS |
9401 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9402 | "not vectorized: relevant stmt not "); | |
9403 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
9404 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 9405 | } |
b8698a0f | 9406 | |
ebfd146a IR |
9407 | return false; |
9408 | } | |
9409 | ||
a70d6342 IR |
9410 | if (bb_vinfo) |
9411 | return true; | |
9412 | ||
8644a673 IR |
9413 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
9414 | need extra handling, except for vectorizable reductions. */ | |
68a0f2ff RS |
9415 | if (STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
9416 | && !can_vectorize_live_stmts (stmt, NULL, node, NULL)) | |
ebfd146a | 9417 | { |
73fbfcad | 9418 | if (dump_enabled_p ()) |
8644a673 | 9419 | { |
78c60e3d | 9420 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
68a0f2ff | 9421 | "not vectorized: live stmt not supported: "); |
78c60e3d | 9422 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); |
8644a673 | 9423 | } |
b8698a0f | 9424 | |
8644a673 | 9425 | return false; |
ebfd146a IR |
9426 | } |
9427 | ||
ebfd146a IR |
9428 | return true; |
9429 | } | |
9430 | ||
9431 | ||
9432 | /* Function vect_transform_stmt. | |
9433 | ||
9434 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
9435 | ||
9436 | bool | |
355fe088 | 9437 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 9438 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
9439 | slp_instance slp_node_instance) |
9440 | { | |
9441 | bool is_store = false; | |
355fe088 | 9442 | gimple *vec_stmt = NULL; |
ebfd146a | 9443 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
ebfd146a | 9444 | bool done; |
ebfd146a | 9445 | |
fce57248 | 9446 | gcc_assert (slp_node || !PURE_SLP_STMT (stmt_info)); |
355fe088 | 9447 | gimple *old_vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 9448 | |
e57d9a82 RB |
9449 | bool nested_p = (STMT_VINFO_LOOP_VINFO (stmt_info) |
9450 | && nested_in_vect_loop_p | |
9451 | (LOOP_VINFO_LOOP (STMT_VINFO_LOOP_VINFO (stmt_info)), | |
9452 | stmt)); | |
9453 | ||
ebfd146a IR |
9454 | switch (STMT_VINFO_TYPE (stmt_info)) |
9455 | { | |
9456 | case type_demotion_vec_info_type: | |
ebfd146a | 9457 | case type_promotion_vec_info_type: |
ebfd146a IR |
9458 | case type_conversion_vec_info_type: |
9459 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
9460 | gcc_assert (done); | |
9461 | break; | |
9462 | ||
9463 | case induc_vec_info_type: | |
e7baeb39 | 9464 | done = vectorizable_induction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
9465 | gcc_assert (done); |
9466 | break; | |
9467 | ||
9dc3f7de IR |
9468 | case shift_vec_info_type: |
9469 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
9470 | gcc_assert (done); | |
9471 | break; | |
9472 | ||
ebfd146a IR |
9473 | case op_vec_info_type: |
9474 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
9475 | gcc_assert (done); | |
9476 | break; | |
9477 | ||
9478 | case assignment_vec_info_type: | |
9479 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
9480 | gcc_assert (done); | |
9481 | break; | |
9482 | ||
9483 | case load_vec_info_type: | |
b8698a0f | 9484 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
9485 | slp_node_instance); |
9486 | gcc_assert (done); | |
9487 | break; | |
9488 | ||
9489 | case store_vec_info_type: | |
9490 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
9491 | gcc_assert (done); | |
0d0293ac | 9492 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
9493 | { |
9494 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 9495 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
9496 | one are skipped, and there vec_stmt_info shouldn't be freed |
9497 | meanwhile. */ | |
0d0293ac | 9498 | *grouped_store = true; |
f307441a RS |
9499 | stmt_vec_info group_info |
9500 | = vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info)); | |
9501 | if (GROUP_STORE_COUNT (group_info) == GROUP_SIZE (group_info)) | |
ebfd146a | 9502 | is_store = true; |
f307441a | 9503 | } |
ebfd146a IR |
9504 | else |
9505 | is_store = true; | |
9506 | break; | |
9507 | ||
9508 | case condition_vec_info_type: | |
f7e531cf | 9509 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
9510 | gcc_assert (done); |
9511 | break; | |
9512 | ||
42fd8198 IE |
9513 | case comparison_vec_info_type: |
9514 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node); | |
9515 | gcc_assert (done); | |
9516 | break; | |
9517 | ||
ebfd146a | 9518 | case call_vec_info_type: |
190c2236 | 9519 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 9520 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
9521 | break; |
9522 | ||
0136f8f0 AH |
9523 | case call_simd_clone_vec_info_type: |
9524 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node); | |
9525 | stmt = gsi_stmt (*gsi); | |
9526 | break; | |
9527 | ||
ebfd146a | 9528 | case reduc_vec_info_type: |
891ad31c RB |
9529 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node, |
9530 | slp_node_instance); | |
ebfd146a IR |
9531 | gcc_assert (done); |
9532 | break; | |
9533 | ||
9534 | default: | |
9535 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
9536 | { | |
73fbfcad | 9537 | if (dump_enabled_p ()) |
78c60e3d | 9538 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9539 | "stmt not supported.\n"); |
ebfd146a IR |
9540 | gcc_unreachable (); |
9541 | } | |
9542 | } | |
9543 | ||
225ce44b RB |
9544 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
9545 | This would break hybrid SLP vectorization. */ | |
9546 | if (slp_node) | |
d90f8440 RB |
9547 | gcc_assert (!vec_stmt |
9548 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt); | |
225ce44b | 9549 | |
ebfd146a IR |
9550 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
9551 | is being vectorized, but outside the immediately enclosing loop. */ | |
9552 | if (vec_stmt | |
e57d9a82 | 9553 | && nested_p |
ebfd146a IR |
9554 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
9555 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 9556 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 9557 | vect_used_in_outer_by_reduction)) |
ebfd146a | 9558 | { |
a70d6342 IR |
9559 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
9560 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
9561 | imm_use_iterator imm_iter; |
9562 | use_operand_p use_p; | |
9563 | tree scalar_dest; | |
355fe088 | 9564 | gimple *exit_phi; |
ebfd146a | 9565 | |
73fbfcad | 9566 | if (dump_enabled_p ()) |
78c60e3d | 9567 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 9568 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
9569 | |
9570 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
9571 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
9572 | STMT). */ | |
9573 | if (gimple_code (stmt) == GIMPLE_PHI) | |
9574 | scalar_dest = PHI_RESULT (stmt); | |
9575 | else | |
9576 | scalar_dest = gimple_assign_lhs (stmt); | |
9577 | ||
9578 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
9579 | { | |
9580 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
9581 | { | |
9582 | exit_phi = USE_STMT (use_p); | |
9583 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
9584 | } | |
9585 | } | |
9586 | } | |
9587 | ||
9588 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
9589 | being vectorized. */ | |
68a0f2ff | 9590 | if (STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) |
ebfd146a | 9591 | { |
68a0f2ff | 9592 | done = can_vectorize_live_stmts (stmt, gsi, slp_node, &vec_stmt); |
ebfd146a IR |
9593 | gcc_assert (done); |
9594 | } | |
9595 | ||
9596 | if (vec_stmt) | |
83197f37 | 9597 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 9598 | |
b8698a0f | 9599 | return is_store; |
ebfd146a IR |
9600 | } |
9601 | ||
9602 | ||
b8698a0f | 9603 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
9604 | stmt_vec_info. */ |
9605 | ||
9606 | void | |
355fe088 | 9607 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 9608 | { |
355fe088 TS |
9609 | gimple *next = first_stmt; |
9610 | gimple *tmp; | |
ebfd146a IR |
9611 | gimple_stmt_iterator next_si; |
9612 | ||
9613 | while (next) | |
9614 | { | |
78048b1c JJ |
9615 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
9616 | ||
9617 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
9618 | if (is_pattern_stmt_p (stmt_info)) | |
9619 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
9620 | /* Free the attached stmt_vec_info and remove the stmt. */ |
9621 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 9622 | unlink_stmt_vdef (next); |
ebfd146a | 9623 | gsi_remove (&next_si, true); |
3d3f2249 | 9624 | release_defs (next); |
ebfd146a IR |
9625 | free_stmt_vec_info (next); |
9626 | next = tmp; | |
9627 | } | |
9628 | } | |
9629 | ||
9630 | ||
9631 | /* Function new_stmt_vec_info. | |
9632 | ||
9633 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
9634 | ||
9635 | stmt_vec_info | |
310213d4 | 9636 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
9637 | { |
9638 | stmt_vec_info res; | |
9639 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
9640 | ||
9641 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
9642 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 9643 | res->vinfo = vinfo; |
8644a673 | 9644 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
9645 | STMT_VINFO_LIVE_P (res) = false; |
9646 | STMT_VINFO_VECTYPE (res) = NULL; | |
9647 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 9648 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
9649 | STMT_VINFO_IN_PATTERN_P (res) = false; |
9650 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 9651 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 9652 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 9653 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
7e16ce79 | 9654 | STMT_VINFO_VEC_CONST_COND_REDUC_CODE (res) = ERROR_MARK; |
ebfd146a | 9655 | |
ebfd146a IR |
9656 | if (gimple_code (stmt) == GIMPLE_PHI |
9657 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
9658 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
9659 | else | |
8644a673 IR |
9660 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
9661 | ||
9771b263 | 9662 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 9663 | STMT_SLP_TYPE (res) = loop_vect; |
78810bd3 RB |
9664 | STMT_VINFO_NUM_SLP_USES (res) = 0; |
9665 | ||
e14c1050 IR |
9666 | GROUP_FIRST_ELEMENT (res) = NULL; |
9667 | GROUP_NEXT_ELEMENT (res) = NULL; | |
9668 | GROUP_SIZE (res) = 0; | |
9669 | GROUP_STORE_COUNT (res) = 0; | |
9670 | GROUP_GAP (res) = 0; | |
9671 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
9672 | |
9673 | return res; | |
9674 | } | |
9675 | ||
9676 | ||
9677 | /* Create a hash table for stmt_vec_info. */ | |
9678 | ||
9679 | void | |
9680 | init_stmt_vec_info_vec (void) | |
9681 | { | |
9771b263 DN |
9682 | gcc_assert (!stmt_vec_info_vec.exists ()); |
9683 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
9684 | } |
9685 | ||
9686 | ||
9687 | /* Free hash table for stmt_vec_info. */ | |
9688 | ||
9689 | void | |
9690 | free_stmt_vec_info_vec (void) | |
9691 | { | |
93675444 | 9692 | unsigned int i; |
3161455c | 9693 | stmt_vec_info info; |
93675444 JJ |
9694 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) |
9695 | if (info != NULL) | |
3161455c | 9696 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
9771b263 DN |
9697 | gcc_assert (stmt_vec_info_vec.exists ()); |
9698 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
9699 | } |
9700 | ||
9701 | ||
9702 | /* Free stmt vectorization related info. */ | |
9703 | ||
9704 | void | |
355fe088 | 9705 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
9706 | { |
9707 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9708 | ||
9709 | if (!stmt_info) | |
9710 | return; | |
9711 | ||
78048b1c JJ |
9712 | /* Check if this statement has a related "pattern stmt" |
9713 | (introduced by the vectorizer during the pattern recognition | |
9714 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
9715 | too. */ | |
9716 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
9717 | { | |
9718 | stmt_vec_info patt_info | |
9719 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
9720 | if (patt_info) | |
9721 | { | |
363477c0 | 9722 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
355fe088 | 9723 | gimple *patt_stmt = STMT_VINFO_STMT (patt_info); |
f0281fde RB |
9724 | gimple_set_bb (patt_stmt, NULL); |
9725 | tree lhs = gimple_get_lhs (patt_stmt); | |
e6f5c25d | 9726 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde | 9727 | release_ssa_name (lhs); |
363477c0 JJ |
9728 | if (seq) |
9729 | { | |
9730 | gimple_stmt_iterator si; | |
9731 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
f0281fde | 9732 | { |
355fe088 | 9733 | gimple *seq_stmt = gsi_stmt (si); |
f0281fde | 9734 | gimple_set_bb (seq_stmt, NULL); |
7532abf2 | 9735 | lhs = gimple_get_lhs (seq_stmt); |
e6f5c25d | 9736 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde RB |
9737 | release_ssa_name (lhs); |
9738 | free_stmt_vec_info (seq_stmt); | |
9739 | } | |
363477c0 | 9740 | } |
f0281fde | 9741 | free_stmt_vec_info (patt_stmt); |
78048b1c JJ |
9742 | } |
9743 | } | |
9744 | ||
9771b263 | 9745 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 9746 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
9747 | set_vinfo_for_stmt (stmt, NULL); |
9748 | free (stmt_info); | |
9749 | } | |
9750 | ||
9751 | ||
bb67d9c7 | 9752 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 9753 | |
bb67d9c7 | 9754 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
9755 | by the target. */ |
9756 | ||
c803b2a9 | 9757 | tree |
86e36728 | 9758 | get_vectype_for_scalar_type_and_size (tree scalar_type, poly_uint64 size) |
ebfd146a | 9759 | { |
c7d97b28 | 9760 | tree orig_scalar_type = scalar_type; |
3bd8f481 | 9761 | scalar_mode inner_mode; |
ef4bddc2 | 9762 | machine_mode simd_mode; |
86e36728 | 9763 | poly_uint64 nunits; |
ebfd146a IR |
9764 | tree vectype; |
9765 | ||
3bd8f481 RS |
9766 | if (!is_int_mode (TYPE_MODE (scalar_type), &inner_mode) |
9767 | && !is_float_mode (TYPE_MODE (scalar_type), &inner_mode)) | |
ebfd146a IR |
9768 | return NULL_TREE; |
9769 | ||
3bd8f481 | 9770 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
48f2e373 | 9771 | |
7b7b1813 RG |
9772 | /* For vector types of elements whose mode precision doesn't |
9773 | match their types precision we use a element type of mode | |
9774 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
9775 | they support the proper result truncation/extension. |
9776 | We also make sure to build vector types with INTEGER_TYPE | |
9777 | component type only. */ | |
6d7971b8 | 9778 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
9779 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
9780 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
9781 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
9782 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 9783 | |
ccbf5bb4 RG |
9784 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
9785 | When the component mode passes the above test simply use a type | |
9786 | corresponding to that mode. The theory is that any use that | |
9787 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 9788 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 9789 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
9790 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
9791 | ||
9792 | /* We can't build a vector type of elements with alignment bigger than | |
9793 | their size. */ | |
dfc2e2ac | 9794 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
9795 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
9796 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 9797 | |
dfc2e2ac RB |
9798 | /* If we felt back to using the mode fail if there was |
9799 | no scalar type for it. */ | |
9800 | if (scalar_type == NULL_TREE) | |
9801 | return NULL_TREE; | |
9802 | ||
bb67d9c7 RG |
9803 | /* If no size was supplied use the mode the target prefers. Otherwise |
9804 | lookup a vector mode of the specified size. */ | |
86e36728 | 9805 | if (known_eq (size, 0U)) |
bb67d9c7 | 9806 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); |
86e36728 RS |
9807 | else if (!multiple_p (size, nbytes, &nunits) |
9808 | || !mode_for_vector (inner_mode, nunits).exists (&simd_mode)) | |
9da15d40 | 9809 | return NULL_TREE; |
4c8fd8ac | 9810 | /* NOTE: nunits == 1 is allowed to support single element vector types. */ |
86e36728 | 9811 | if (!multiple_p (GET_MODE_SIZE (simd_mode), nbytes, &nunits)) |
cc4b5170 | 9812 | return NULL_TREE; |
ebfd146a IR |
9813 | |
9814 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
9815 | |
9816 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
9817 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 9818 | return NULL_TREE; |
ebfd146a | 9819 | |
c7d97b28 RB |
9820 | /* Re-attach the address-space qualifier if we canonicalized the scalar |
9821 | type. */ | |
9822 | if (TYPE_ADDR_SPACE (orig_scalar_type) != TYPE_ADDR_SPACE (vectype)) | |
9823 | return build_qualified_type | |
9824 | (vectype, KEEP_QUAL_ADDR_SPACE (TYPE_QUALS (orig_scalar_type))); | |
9825 | ||
ebfd146a IR |
9826 | return vectype; |
9827 | } | |
9828 | ||
86e36728 | 9829 | poly_uint64 current_vector_size; |
bb67d9c7 RG |
9830 | |
9831 | /* Function get_vectype_for_scalar_type. | |
9832 | ||
9833 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
9834 | by the target. */ | |
9835 | ||
9836 | tree | |
9837 | get_vectype_for_scalar_type (tree scalar_type) | |
9838 | { | |
9839 | tree vectype; | |
9840 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
9841 | current_vector_size); | |
9842 | if (vectype | |
86e36728 | 9843 | && known_eq (current_vector_size, 0U)) |
bb67d9c7 RG |
9844 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); |
9845 | return vectype; | |
9846 | } | |
9847 | ||
42fd8198 IE |
9848 | /* Function get_mask_type_for_scalar_type. |
9849 | ||
9850 | Returns the mask type corresponding to a result of comparison | |
9851 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
9852 | ||
9853 | tree | |
9854 | get_mask_type_for_scalar_type (tree scalar_type) | |
9855 | { | |
9856 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
9857 | ||
9858 | if (!vectype) | |
9859 | return NULL; | |
9860 | ||
9861 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
9862 | current_vector_size); | |
9863 | } | |
9864 | ||
b690cc0f RG |
9865 | /* Function get_same_sized_vectype |
9866 | ||
9867 | Returns a vector type corresponding to SCALAR_TYPE of size | |
9868 | VECTOR_TYPE if supported by the target. */ | |
9869 | ||
9870 | tree | |
bb67d9c7 | 9871 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 9872 | { |
2568d8a1 | 9873 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type)) |
9f47c7e5 IE |
9874 | return build_same_sized_truth_vector_type (vector_type); |
9875 | ||
bb67d9c7 RG |
9876 | return get_vectype_for_scalar_type_and_size |
9877 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
9878 | } |
9879 | ||
ebfd146a IR |
9880 | /* Function vect_is_simple_use. |
9881 | ||
9882 | Input: | |
81c40241 RB |
9883 | VINFO - the vect info of the loop or basic block that is being vectorized. |
9884 | OPERAND - operand in the loop or bb. | |
9885 | Output: | |
9886 | DEF_STMT - the defining stmt in case OPERAND is an SSA_NAME. | |
9887 | DT - the type of definition | |
ebfd146a IR |
9888 | |
9889 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 9890 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 9891 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 9892 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
9893 | is the case in reduction/induction computations). |
9894 | For basic blocks, supportable operands are constants and bb invariants. | |
9895 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
9896 | |
9897 | bool | |
81c40241 RB |
9898 | vect_is_simple_use (tree operand, vec_info *vinfo, |
9899 | gimple **def_stmt, enum vect_def_type *dt) | |
b8698a0f | 9900 | { |
ebfd146a | 9901 | *def_stmt = NULL; |
3fc356dc | 9902 | *dt = vect_unknown_def_type; |
b8698a0f | 9903 | |
73fbfcad | 9904 | if (dump_enabled_p ()) |
ebfd146a | 9905 | { |
78c60e3d SS |
9906 | dump_printf_loc (MSG_NOTE, vect_location, |
9907 | "vect_is_simple_use: operand "); | |
9908 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 9909 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 9910 | } |
b8698a0f | 9911 | |
b758f602 | 9912 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
9913 | { |
9914 | *dt = vect_constant_def; | |
9915 | return true; | |
9916 | } | |
b8698a0f | 9917 | |
ebfd146a IR |
9918 | if (is_gimple_min_invariant (operand)) |
9919 | { | |
8644a673 | 9920 | *dt = vect_external_def; |
ebfd146a IR |
9921 | return true; |
9922 | } | |
9923 | ||
ebfd146a IR |
9924 | if (TREE_CODE (operand) != SSA_NAME) |
9925 | { | |
73fbfcad | 9926 | if (dump_enabled_p ()) |
af29617a AH |
9927 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9928 | "not ssa-name.\n"); | |
ebfd146a IR |
9929 | return false; |
9930 | } | |
b8698a0f | 9931 | |
3fc356dc | 9932 | if (SSA_NAME_IS_DEFAULT_DEF (operand)) |
ebfd146a | 9933 | { |
3fc356dc RB |
9934 | *dt = vect_external_def; |
9935 | return true; | |
ebfd146a IR |
9936 | } |
9937 | ||
3fc356dc | 9938 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
73fbfcad | 9939 | if (dump_enabled_p ()) |
ebfd146a | 9940 | { |
78c60e3d SS |
9941 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
9942 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
9943 | } |
9944 | ||
61d371eb | 9945 | if (! vect_stmt_in_region_p (vinfo, *def_stmt)) |
8644a673 | 9946 | *dt = vect_external_def; |
ebfd146a IR |
9947 | else |
9948 | { | |
3fc356dc | 9949 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (*def_stmt); |
603cca93 | 9950 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
ebfd146a IR |
9951 | } |
9952 | ||
2e8ab70c RB |
9953 | if (dump_enabled_p ()) |
9954 | { | |
9955 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: "); | |
9956 | switch (*dt) | |
9957 | { | |
9958 | case vect_uninitialized_def: | |
9959 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
9960 | break; | |
9961 | case vect_constant_def: | |
9962 | dump_printf (MSG_NOTE, "constant\n"); | |
9963 | break; | |
9964 | case vect_external_def: | |
9965 | dump_printf (MSG_NOTE, "external\n"); | |
9966 | break; | |
9967 | case vect_internal_def: | |
9968 | dump_printf (MSG_NOTE, "internal\n"); | |
9969 | break; | |
9970 | case vect_induction_def: | |
9971 | dump_printf (MSG_NOTE, "induction\n"); | |
9972 | break; | |
9973 | case vect_reduction_def: | |
9974 | dump_printf (MSG_NOTE, "reduction\n"); | |
9975 | break; | |
9976 | case vect_double_reduction_def: | |
9977 | dump_printf (MSG_NOTE, "double reduction\n"); | |
9978 | break; | |
9979 | case vect_nested_cycle: | |
9980 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
9981 | break; | |
9982 | case vect_unknown_def_type: | |
9983 | dump_printf (MSG_NOTE, "unknown\n"); | |
9984 | break; | |
9985 | } | |
9986 | } | |
9987 | ||
81c40241 | 9988 | if (*dt == vect_unknown_def_type) |
ebfd146a | 9989 | { |
73fbfcad | 9990 | if (dump_enabled_p ()) |
78c60e3d | 9991 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9992 | "Unsupported pattern.\n"); |
ebfd146a IR |
9993 | return false; |
9994 | } | |
9995 | ||
ebfd146a IR |
9996 | switch (gimple_code (*def_stmt)) |
9997 | { | |
9998 | case GIMPLE_PHI: | |
ebfd146a | 9999 | case GIMPLE_ASSIGN: |
ebfd146a | 10000 | case GIMPLE_CALL: |
81c40241 | 10001 | break; |
ebfd146a | 10002 | default: |
73fbfcad | 10003 | if (dump_enabled_p ()) |
78c60e3d | 10004 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 10005 | "unsupported defining stmt:\n"); |
ebfd146a IR |
10006 | return false; |
10007 | } | |
10008 | ||
10009 | return true; | |
10010 | } | |
10011 | ||
81c40241 | 10012 | /* Function vect_is_simple_use. |
b690cc0f | 10013 | |
81c40241 | 10014 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
10015 | type of OPERAND and stores it to *VECTYPE. If the definition of |
10016 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
10017 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
10018 | is responsible to compute the best suited vector type for the | |
10019 | scalar operand. */ | |
10020 | ||
10021 | bool | |
81c40241 RB |
10022 | vect_is_simple_use (tree operand, vec_info *vinfo, |
10023 | gimple **def_stmt, enum vect_def_type *dt, tree *vectype) | |
b690cc0f | 10024 | { |
81c40241 | 10025 | if (!vect_is_simple_use (operand, vinfo, def_stmt, dt)) |
b690cc0f RG |
10026 | return false; |
10027 | ||
10028 | /* Now get a vector type if the def is internal, otherwise supply | |
10029 | NULL_TREE and leave it up to the caller to figure out a proper | |
10030 | type for the use stmt. */ | |
10031 | if (*dt == vect_internal_def | |
10032 | || *dt == vect_induction_def | |
10033 | || *dt == vect_reduction_def | |
10034 | || *dt == vect_double_reduction_def | |
10035 | || *dt == vect_nested_cycle) | |
10036 | { | |
10037 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
10038 | |
10039 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
10040 | && !STMT_VINFO_RELEVANT (stmt_info) | |
10041 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 10042 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 10043 | |
b690cc0f RG |
10044 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
10045 | gcc_assert (*vectype != NULL_TREE); | |
10046 | } | |
10047 | else if (*dt == vect_uninitialized_def | |
10048 | || *dt == vect_constant_def | |
10049 | || *dt == vect_external_def) | |
10050 | *vectype = NULL_TREE; | |
10051 | else | |
10052 | gcc_unreachable (); | |
10053 | ||
10054 | return true; | |
10055 | } | |
10056 | ||
ebfd146a IR |
10057 | |
10058 | /* Function supportable_widening_operation | |
10059 | ||
b8698a0f L |
10060 | Check whether an operation represented by the code CODE is a |
10061 | widening operation that is supported by the target platform in | |
b690cc0f RG |
10062 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
10063 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 10064 | |
ebfd146a IR |
10065 | Widening operations we currently support are NOP (CONVERT), FLOAT |
10066 | and WIDEN_MULT. This function checks if these operations are supported | |
10067 | by the target platform either directly (via vector tree-codes), or via | |
10068 | target builtins. | |
10069 | ||
10070 | Output: | |
b8698a0f L |
10071 | - CODE1 and CODE2 are codes of vector operations to be used when |
10072 | vectorizing the operation, if available. | |
ebfd146a IR |
10073 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
10074 | case of multi-step conversion (like char->short->int - in that case | |
10075 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
10076 | - INTERM_TYPES contains the intermediate type required to perform the |
10077 | widening operation (short in the above example). */ | |
ebfd146a IR |
10078 | |
10079 | bool | |
355fe088 | 10080 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 10081 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
10082 | enum tree_code *code1, enum tree_code *code2, |
10083 | int *multi_step_cvt, | |
9771b263 | 10084 | vec<tree> *interm_types) |
ebfd146a IR |
10085 | { |
10086 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
10087 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 10088 | struct loop *vect_loop = NULL; |
ef4bddc2 | 10089 | machine_mode vec_mode; |
81f40b79 | 10090 | enum insn_code icode1, icode2; |
ebfd146a | 10091 | optab optab1, optab2; |
b690cc0f RG |
10092 | tree vectype = vectype_in; |
10093 | tree wide_vectype = vectype_out; | |
ebfd146a | 10094 | enum tree_code c1, c2; |
4a00c761 JJ |
10095 | int i; |
10096 | tree prev_type, intermediate_type; | |
ef4bddc2 | 10097 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 10098 | optab optab3, optab4; |
ebfd146a | 10099 | |
4a00c761 | 10100 | *multi_step_cvt = 0; |
4ef69dfc IR |
10101 | if (loop_info) |
10102 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
10103 | ||
ebfd146a IR |
10104 | switch (code) |
10105 | { | |
10106 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
10107 | /* The result of a vectorized widening operation usually requires |
10108 | two vectors (because the widened results do not fit into one vector). | |
10109 | The generated vector results would normally be expected to be | |
10110 | generated in the same order as in the original scalar computation, | |
10111 | i.e. if 8 results are generated in each vector iteration, they are | |
10112 | to be organized as follows: | |
10113 | vect1: [res1,res2,res3,res4], | |
10114 | vect2: [res5,res6,res7,res8]. | |
10115 | ||
10116 | However, in the special case that the result of the widening | |
10117 | operation is used in a reduction computation only, the order doesn't | |
10118 | matter (because when vectorizing a reduction we change the order of | |
10119 | the computation). Some targets can take advantage of this and | |
10120 | generate more efficient code. For example, targets like Altivec, | |
10121 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
10122 | generate the following vectors: | |
10123 | vect1: [res1,res3,res5,res7], | |
10124 | vect2: [res2,res4,res6,res8]. | |
10125 | ||
10126 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
10127 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
10128 | iterations in parallel). We therefore don't allow to change the | |
10129 | order of the computation in the inner-loop during outer-loop | |
10130 | vectorization. */ | |
10131 | /* TODO: Another case in which order doesn't *really* matter is when we | |
10132 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
10133 | Normally, pack_trunc performs an even/odd permute, whereas the | |
10134 | repack from an even/odd expansion would be an interleave, which | |
10135 | would be significantly simpler for e.g. AVX2. */ | |
10136 | /* In any case, in order to avoid duplicating the code below, recurse | |
10137 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
10138 | are properly set up for the caller. If we fail, we'll continue with | |
10139 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
10140 | if (vect_loop | |
10141 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
10142 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
10143 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
10144 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
10145 | code1, code2, multi_step_cvt, |
10146 | interm_types)) | |
ebc047a2 CH |
10147 | { |
10148 | /* Elements in a vector with vect_used_by_reduction property cannot | |
10149 | be reordered if the use chain with this property does not have the | |
10150 | same operation. One such an example is s += a * b, where elements | |
10151 | in a and b cannot be reordered. Here we check if the vector defined | |
10152 | by STMT is only directly used in the reduction statement. */ | |
10153 | tree lhs = gimple_assign_lhs (stmt); | |
10154 | use_operand_p dummy; | |
355fe088 | 10155 | gimple *use_stmt; |
ebc047a2 CH |
10156 | stmt_vec_info use_stmt_info = NULL; |
10157 | if (single_imm_use (lhs, &dummy, &use_stmt) | |
10158 | && (use_stmt_info = vinfo_for_stmt (use_stmt)) | |
10159 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
10160 | return true; | |
10161 | } | |
4a00c761 JJ |
10162 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
10163 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
10164 | break; |
10165 | ||
81c40241 RB |
10166 | case DOT_PROD_EXPR: |
10167 | c1 = DOT_PROD_EXPR; | |
10168 | c2 = DOT_PROD_EXPR; | |
10169 | break; | |
10170 | ||
10171 | case SAD_EXPR: | |
10172 | c1 = SAD_EXPR; | |
10173 | c2 = SAD_EXPR; | |
10174 | break; | |
10175 | ||
6ae6116f RH |
10176 | case VEC_WIDEN_MULT_EVEN_EXPR: |
10177 | /* Support the recursion induced just above. */ | |
10178 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
10179 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
10180 | break; | |
10181 | ||
36ba4aae | 10182 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
10183 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
10184 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
10185 | break; |
10186 | ||
ebfd146a | 10187 | CASE_CONVERT: |
4a00c761 JJ |
10188 | c1 = VEC_UNPACK_LO_EXPR; |
10189 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
10190 | break; |
10191 | ||
10192 | case FLOAT_EXPR: | |
4a00c761 JJ |
10193 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
10194 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
10195 | break; |
10196 | ||
10197 | case FIX_TRUNC_EXPR: | |
10198 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
10199 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
10200 | computing the operation. */ | |
10201 | return false; | |
10202 | ||
10203 | default: | |
10204 | gcc_unreachable (); | |
10205 | } | |
10206 | ||
6ae6116f | 10207 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 10208 | std::swap (c1, c2); |
4a00c761 | 10209 | |
ebfd146a IR |
10210 | if (code == FIX_TRUNC_EXPR) |
10211 | { | |
10212 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
10213 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
10214 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
10215 | } |
10216 | else | |
10217 | { | |
10218 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
10219 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
10220 | } | |
10221 | ||
10222 | if (!optab1 || !optab2) | |
10223 | return false; | |
10224 | ||
10225 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
10226 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
10227 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
10228 | return false; |
10229 | ||
4a00c761 JJ |
10230 | *code1 = c1; |
10231 | *code2 = c2; | |
10232 | ||
10233 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
10234 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
10235 | /* For scalar masks we may have different boolean |
10236 | vector types having the same QImode. Thus we | |
10237 | add additional check for elements number. */ | |
10238 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
928686b1 RS |
10239 | || known_eq (TYPE_VECTOR_SUBPARTS (vectype), |
10240 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)); | |
4a00c761 | 10241 | |
b8698a0f | 10242 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 10243 | types. */ |
ebfd146a | 10244 | |
4a00c761 JJ |
10245 | prev_type = vectype; |
10246 | prev_mode = vec_mode; | |
b8698a0f | 10247 | |
4a00c761 JJ |
10248 | if (!CONVERT_EXPR_CODE_P (code)) |
10249 | return false; | |
b8698a0f | 10250 | |
4a00c761 JJ |
10251 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
10252 | intermediate steps in promotion sequence. We try | |
10253 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
10254 | not. */ | |
9771b263 | 10255 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
10256 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
10257 | { | |
10258 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
10259 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
10260 | { | |
7cfb4d93 | 10261 | intermediate_type = vect_halve_mask_nunits (prev_type); |
3ae0661a IE |
10262 | if (intermediate_mode != TYPE_MODE (intermediate_type)) |
10263 | return false; | |
10264 | } | |
10265 | else | |
10266 | intermediate_type | |
10267 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
10268 | TYPE_UNSIGNED (prev_type)); | |
10269 | ||
4a00c761 JJ |
10270 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); |
10271 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
10272 | ||
10273 | if (!optab3 || !optab4 | |
10274 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
10275 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
10276 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
10277 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
10278 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
10279 | == CODE_FOR_nothing) | |
10280 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
10281 | == CODE_FOR_nothing)) | |
10282 | break; | |
ebfd146a | 10283 | |
9771b263 | 10284 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
10285 | (*multi_step_cvt)++; |
10286 | ||
10287 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
10288 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff | 10289 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
928686b1 RS |
10290 | || known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type), |
10291 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)); | |
4a00c761 JJ |
10292 | |
10293 | prev_type = intermediate_type; | |
10294 | prev_mode = intermediate_mode; | |
ebfd146a IR |
10295 | } |
10296 | ||
9771b263 | 10297 | interm_types->release (); |
4a00c761 | 10298 | return false; |
ebfd146a IR |
10299 | } |
10300 | ||
10301 | ||
10302 | /* Function supportable_narrowing_operation | |
10303 | ||
b8698a0f L |
10304 | Check whether an operation represented by the code CODE is a |
10305 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
10306 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
10307 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 10308 | |
ebfd146a | 10309 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 10310 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
10311 | the target platform directly via vector tree-codes. |
10312 | ||
10313 | Output: | |
b8698a0f L |
10314 | - CODE1 is the code of a vector operation to be used when |
10315 | vectorizing the operation, if available. | |
ebfd146a IR |
10316 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
10317 | case of multi-step conversion (like int->short->char - in that case | |
10318 | MULTI_STEP_CVT will be 1). | |
10319 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 10320 | narrowing operation (short in the above example). */ |
ebfd146a IR |
10321 | |
10322 | bool | |
10323 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 10324 | tree vectype_out, tree vectype_in, |
ebfd146a | 10325 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 10326 | vec<tree> *interm_types) |
ebfd146a | 10327 | { |
ef4bddc2 | 10328 | machine_mode vec_mode; |
ebfd146a IR |
10329 | enum insn_code icode1; |
10330 | optab optab1, interm_optab; | |
b690cc0f RG |
10331 | tree vectype = vectype_in; |
10332 | tree narrow_vectype = vectype_out; | |
ebfd146a | 10333 | enum tree_code c1; |
3ae0661a | 10334 | tree intermediate_type, prev_type; |
ef4bddc2 | 10335 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 10336 | int i; |
4a00c761 | 10337 | bool uns; |
ebfd146a | 10338 | |
4a00c761 | 10339 | *multi_step_cvt = 0; |
ebfd146a IR |
10340 | switch (code) |
10341 | { | |
10342 | CASE_CONVERT: | |
10343 | c1 = VEC_PACK_TRUNC_EXPR; | |
10344 | break; | |
10345 | ||
10346 | case FIX_TRUNC_EXPR: | |
10347 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
10348 | break; | |
10349 | ||
10350 | case FLOAT_EXPR: | |
10351 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
10352 | tree code and optabs used for computing the operation. */ | |
10353 | return false; | |
10354 | ||
10355 | default: | |
10356 | gcc_unreachable (); | |
10357 | } | |
10358 | ||
10359 | if (code == FIX_TRUNC_EXPR) | |
10360 | /* The signedness is determined from output operand. */ | |
b690cc0f | 10361 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
10362 | else |
10363 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
10364 | ||
10365 | if (!optab1) | |
10366 | return false; | |
10367 | ||
10368 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 10369 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
10370 | return false; |
10371 | ||
4a00c761 JJ |
10372 | *code1 = c1; |
10373 | ||
10374 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
10375 | /* For scalar masks we may have different boolean |
10376 | vector types having the same QImode. Thus we | |
10377 | add additional check for elements number. */ | |
10378 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
928686b1 RS |
10379 | || known_eq (TYPE_VECTOR_SUBPARTS (vectype) * 2, |
10380 | TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 | 10381 | |
ebfd146a IR |
10382 | /* Check if it's a multi-step conversion that can be done using intermediate |
10383 | types. */ | |
4a00c761 | 10384 | prev_mode = vec_mode; |
3ae0661a | 10385 | prev_type = vectype; |
4a00c761 JJ |
10386 | if (code == FIX_TRUNC_EXPR) |
10387 | uns = TYPE_UNSIGNED (vectype_out); | |
10388 | else | |
10389 | uns = TYPE_UNSIGNED (vectype); | |
10390 | ||
10391 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
10392 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
10393 | costly than signed. */ | |
10394 | if (code == FIX_TRUNC_EXPR && uns) | |
10395 | { | |
10396 | enum insn_code icode2; | |
10397 | ||
10398 | intermediate_type | |
10399 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
10400 | interm_optab | |
10401 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 10402 | if (interm_optab != unknown_optab |
4a00c761 JJ |
10403 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
10404 | && insn_data[icode1].operand[0].mode | |
10405 | == insn_data[icode2].operand[0].mode) | |
10406 | { | |
10407 | uns = false; | |
10408 | optab1 = interm_optab; | |
10409 | icode1 = icode2; | |
10410 | } | |
10411 | } | |
ebfd146a | 10412 | |
4a00c761 JJ |
10413 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
10414 | intermediate steps in promotion sequence. We try | |
10415 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 10416 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
10417 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
10418 | { | |
10419 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
10420 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
10421 | { | |
7cfb4d93 | 10422 | intermediate_type = vect_double_mask_nunits (prev_type); |
3ae0661a | 10423 | if (intermediate_mode != TYPE_MODE (intermediate_type)) |
7cfb4d93 | 10424 | return false; |
3ae0661a IE |
10425 | } |
10426 | else | |
10427 | intermediate_type | |
10428 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
4a00c761 JJ |
10429 | interm_optab |
10430 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
10431 | optab_default); | |
10432 | if (!interm_optab | |
10433 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
10434 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
10435 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
10436 | == CODE_FOR_nothing)) | |
10437 | break; | |
10438 | ||
9771b263 | 10439 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
10440 | (*multi_step_cvt)++; |
10441 | ||
10442 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff | 10443 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
928686b1 RS |
10444 | || known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type) * 2, |
10445 | TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 JJ |
10446 | |
10447 | prev_mode = intermediate_mode; | |
3ae0661a | 10448 | prev_type = intermediate_type; |
4a00c761 | 10449 | optab1 = interm_optab; |
ebfd146a IR |
10450 | } |
10451 | ||
9771b263 | 10452 | interm_types->release (); |
4a00c761 | 10453 | return false; |
ebfd146a | 10454 | } |
7cfb4d93 RS |
10455 | |
10456 | /* Generate and return a statement that sets vector mask MASK such that | |
10457 | MASK[I] is true iff J + START_INDEX < END_INDEX for all J <= I. */ | |
10458 | ||
10459 | gcall * | |
10460 | vect_gen_while (tree mask, tree start_index, tree end_index) | |
10461 | { | |
10462 | tree cmp_type = TREE_TYPE (start_index); | |
10463 | tree mask_type = TREE_TYPE (mask); | |
10464 | gcc_checking_assert (direct_internal_fn_supported_p (IFN_WHILE_ULT, | |
10465 | cmp_type, mask_type, | |
10466 | OPTIMIZE_FOR_SPEED)); | |
10467 | gcall *call = gimple_build_call_internal (IFN_WHILE_ULT, 3, | |
10468 | start_index, end_index, | |
10469 | build_zero_cst (mask_type)); | |
10470 | gimple_call_set_lhs (call, mask); | |
10471 | return call; | |
10472 | } | |
535e7c11 RS |
10473 | |
10474 | /* Generate a vector mask of type MASK_TYPE for which index I is false iff | |
10475 | J + START_INDEX < END_INDEX for all J <= I. Add the statements to SEQ. */ | |
10476 | ||
10477 | tree | |
10478 | vect_gen_while_not (gimple_seq *seq, tree mask_type, tree start_index, | |
10479 | tree end_index) | |
10480 | { | |
10481 | tree tmp = make_ssa_name (mask_type); | |
10482 | gcall *call = vect_gen_while (tmp, start_index, end_index); | |
10483 | gimple_seq_add_stmt (seq, call); | |
10484 | return gimple_build (seq, BIT_NOT_EXPR, mask_type, tmp); | |
10485 | } |