]>
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; | |
68435eb2 RB |
100 | |
101 | stmt_info_for_cost si = { count, kind, where, | |
102 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
103 | misalign }; | |
104 | body_cost_vec->safe_push (si); | |
105 | ||
106 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; | |
107 | return (unsigned) | |
108 | (builtin_vectorization_cost (kind, vectype, misalign) * count); | |
c3e7ee41 BS |
109 | } |
110 | ||
272c6793 RS |
111 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
112 | ||
113 | static tree | |
114 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
115 | { | |
116 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
117 | "vect_array"); | |
118 | } | |
119 | ||
120 | /* ARRAY is an array of vectors created by create_vector_array. | |
121 | Return an SSA_NAME for the vector in index N. The reference | |
122 | is part of the vectorization of STMT and the vector is associated | |
123 | with scalar destination SCALAR_DEST. */ | |
124 | ||
125 | static tree | |
355fe088 | 126 | read_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree scalar_dest, |
272c6793 RS |
127 | tree array, unsigned HOST_WIDE_INT n) |
128 | { | |
129 | tree vect_type, vect, vect_name, array_ref; | |
355fe088 | 130 | gimple *new_stmt; |
272c6793 RS |
131 | |
132 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
133 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
134 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
135 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
136 | build_int_cst (size_type_node, n), | |
137 | NULL_TREE, NULL_TREE); | |
138 | ||
139 | new_stmt = gimple_build_assign (vect, array_ref); | |
140 | vect_name = make_ssa_name (vect, new_stmt); | |
141 | gimple_assign_set_lhs (new_stmt, vect_name); | |
142 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
143 | |
144 | return vect_name; | |
145 | } | |
146 | ||
147 | /* ARRAY is an array of vectors created by create_vector_array. | |
148 | Emit code to store SSA_NAME VECT in index N of the array. | |
149 | The store is part of the vectorization of STMT. */ | |
150 | ||
151 | static void | |
355fe088 | 152 | write_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree vect, |
272c6793 RS |
153 | tree array, unsigned HOST_WIDE_INT n) |
154 | { | |
155 | tree array_ref; | |
355fe088 | 156 | gimple *new_stmt; |
272c6793 RS |
157 | |
158 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
159 | build_int_cst (size_type_node, n), | |
160 | NULL_TREE, NULL_TREE); | |
161 | ||
162 | new_stmt = gimple_build_assign (array_ref, vect); | |
163 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
164 | } |
165 | ||
166 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
167 | of *PTR. The memory reference replaces those in FIRST_DR | |
168 | (and its group). */ | |
169 | ||
170 | static tree | |
44fc7854 | 171 | create_array_ref (tree type, tree ptr, tree alias_ptr_type) |
272c6793 | 172 | { |
44fc7854 | 173 | tree mem_ref; |
272c6793 | 174 | |
272c6793 RS |
175 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); |
176 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 177 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
178 | return mem_ref; |
179 | } | |
180 | ||
3ba4ff41 RS |
181 | /* Add a clobber of variable VAR to the vectorization of STMT. |
182 | Emit the clobber before *GSI. */ | |
183 | ||
184 | static void | |
185 | vect_clobber_variable (gimple *stmt, gimple_stmt_iterator *gsi, tree var) | |
186 | { | |
187 | tree clobber = build_clobber (TREE_TYPE (var)); | |
188 | gimple *new_stmt = gimple_build_assign (var, clobber); | |
189 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
190 | } | |
191 | ||
ebfd146a IR |
192 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
193 | ||
194 | /* Function vect_mark_relevant. | |
195 | ||
196 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
197 | ||
198 | static void | |
355fe088 | 199 | vect_mark_relevant (vec<gimple *> *worklist, gimple *stmt, |
97ecdb46 | 200 | enum vect_relevant relevant, bool live_p) |
ebfd146a IR |
201 | { |
202 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
203 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
204 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
355fe088 | 205 | gimple *pattern_stmt; |
ebfd146a | 206 | |
73fbfcad | 207 | if (dump_enabled_p ()) |
66c16fd9 RB |
208 | { |
209 | dump_printf_loc (MSG_NOTE, vect_location, | |
210 | "mark relevant %d, live %d: ", relevant, live_p); | |
211 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
212 | } | |
ebfd146a | 213 | |
83197f37 IR |
214 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
215 | related pattern stmt instead of the original stmt. However, such stmts | |
216 | may have their own uses that are not in any pattern, in such cases the | |
217 | stmt itself should be marked. */ | |
ebfd146a IR |
218 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
219 | { | |
97ecdb46 JJ |
220 | /* This is the last stmt in a sequence that was detected as a |
221 | pattern that can potentially be vectorized. Don't mark the stmt | |
222 | as relevant/live because it's not going to be vectorized. | |
223 | Instead mark the pattern-stmt that replaces it. */ | |
83197f37 | 224 | |
97ecdb46 JJ |
225 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); |
226 | ||
227 | if (dump_enabled_p ()) | |
228 | dump_printf_loc (MSG_NOTE, vect_location, | |
229 | "last stmt in pattern. don't mark" | |
230 | " relevant/live.\n"); | |
231 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
232 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
233 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
234 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
235 | stmt = pattern_stmt; | |
ebfd146a IR |
236 | } |
237 | ||
238 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
239 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
240 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
241 | ||
242 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
243 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
244 | { | |
73fbfcad | 245 | if (dump_enabled_p ()) |
78c60e3d | 246 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 247 | "already marked relevant/live.\n"); |
ebfd146a IR |
248 | return; |
249 | } | |
250 | ||
9771b263 | 251 | worklist->safe_push (stmt); |
ebfd146a IR |
252 | } |
253 | ||
254 | ||
b28ead45 AH |
255 | /* Function is_simple_and_all_uses_invariant |
256 | ||
257 | Return true if STMT is simple and all uses of it are invariant. */ | |
258 | ||
259 | bool | |
260 | is_simple_and_all_uses_invariant (gimple *stmt, loop_vec_info loop_vinfo) | |
261 | { | |
262 | tree op; | |
b28ead45 AH |
263 | ssa_op_iter iter; |
264 | ||
265 | if (!is_gimple_assign (stmt)) | |
266 | return false; | |
267 | ||
268 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
269 | { | |
270 | enum vect_def_type dt = vect_uninitialized_def; | |
271 | ||
894dd753 | 272 | if (!vect_is_simple_use (op, loop_vinfo, &dt)) |
b28ead45 AH |
273 | { |
274 | if (dump_enabled_p ()) | |
275 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
276 | "use not simple.\n"); | |
277 | return false; | |
278 | } | |
279 | ||
280 | if (dt != vect_external_def && dt != vect_constant_def) | |
281 | return false; | |
282 | } | |
283 | return true; | |
284 | } | |
285 | ||
ebfd146a IR |
286 | /* Function vect_stmt_relevant_p. |
287 | ||
288 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
289 | "relevant for vectorization". | |
290 | ||
291 | A stmt is considered "relevant for vectorization" if: | |
292 | - it has uses outside the loop. | |
293 | - it has vdefs (it alters memory). | |
294 | - control stmts in the loop (except for the exit condition). | |
295 | ||
296 | CHECKME: what other side effects would the vectorizer allow? */ | |
297 | ||
298 | static bool | |
355fe088 | 299 | vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo, |
ebfd146a IR |
300 | enum vect_relevant *relevant, bool *live_p) |
301 | { | |
302 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
303 | ssa_op_iter op_iter; | |
304 | imm_use_iterator imm_iter; | |
305 | use_operand_p use_p; | |
306 | def_operand_p def_p; | |
307 | ||
8644a673 | 308 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
309 | *live_p = false; |
310 | ||
311 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
312 | if (is_ctrl_stmt (stmt) |
313 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
314 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 315 | *relevant = vect_used_in_scope; |
ebfd146a IR |
316 | |
317 | /* changing memory. */ | |
318 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
319 | if (gimple_vdef (stmt) |
320 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 321 | { |
73fbfcad | 322 | if (dump_enabled_p ()) |
78c60e3d | 323 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 324 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 325 | *relevant = vect_used_in_scope; |
ebfd146a IR |
326 | } |
327 | ||
328 | /* uses outside the loop. */ | |
329 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
330 | { | |
331 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
332 | { | |
333 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
334 | if (!flow_bb_inside_loop_p (loop, bb)) | |
335 | { | |
73fbfcad | 336 | if (dump_enabled_p ()) |
78c60e3d | 337 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 338 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 339 | |
3157b0c2 AO |
340 | if (is_gimple_debug (USE_STMT (use_p))) |
341 | continue; | |
342 | ||
ebfd146a IR |
343 | /* We expect all such uses to be in the loop exit phis |
344 | (because of loop closed form) */ | |
345 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
346 | gcc_assert (bb == single_exit (loop)->dest); | |
347 | ||
348 | *live_p = true; | |
349 | } | |
350 | } | |
351 | } | |
352 | ||
3a2edf4c AH |
353 | if (*live_p && *relevant == vect_unused_in_scope |
354 | && !is_simple_and_all_uses_invariant (stmt, loop_vinfo)) | |
b28ead45 AH |
355 | { |
356 | if (dump_enabled_p ()) | |
357 | dump_printf_loc (MSG_NOTE, vect_location, | |
358 | "vec_stmt_relevant_p: stmt live but not relevant.\n"); | |
359 | *relevant = vect_used_only_live; | |
360 | } | |
361 | ||
ebfd146a IR |
362 | return (*live_p || *relevant); |
363 | } | |
364 | ||
365 | ||
b8698a0f | 366 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 367 | |
ff802fa1 | 368 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
369 | used in STMT for anything other than indexing an array. */ |
370 | ||
371 | static bool | |
355fe088 | 372 | exist_non_indexing_operands_for_use_p (tree use, gimple *stmt) |
ebfd146a IR |
373 | { |
374 | tree operand; | |
375 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 376 | |
ff802fa1 | 377 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
378 | reference in STMT, then any operand that corresponds to USE |
379 | is not indexing an array. */ | |
380 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
381 | return true; | |
59a05b0c | 382 | |
ebfd146a IR |
383 | /* STMT has a data_ref. FORNOW this means that its of one of |
384 | the following forms: | |
385 | -1- ARRAY_REF = var | |
386 | -2- var = ARRAY_REF | |
387 | (This should have been verified in analyze_data_refs). | |
388 | ||
389 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 390 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
391 | for array indexing. |
392 | ||
393 | Therefore, all we need to check is if STMT falls into the | |
394 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
395 | |
396 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
397 | { |
398 | if (is_gimple_call (stmt) | |
399 | && gimple_call_internal_p (stmt)) | |
bfaa08b7 RS |
400 | { |
401 | internal_fn ifn = gimple_call_internal_fn (stmt); | |
402 | int mask_index = internal_fn_mask_index (ifn); | |
403 | if (mask_index >= 0 | |
404 | && use == gimple_call_arg (stmt, mask_index)) | |
405 | return true; | |
f307441a RS |
406 | int stored_value_index = internal_fn_stored_value_index (ifn); |
407 | if (stored_value_index >= 0 | |
408 | && use == gimple_call_arg (stmt, stored_value_index)) | |
409 | return true; | |
bfaa08b7 RS |
410 | if (internal_gather_scatter_fn_p (ifn) |
411 | && use == gimple_call_arg (stmt, 1)) | |
412 | return true; | |
bfaa08b7 | 413 | } |
5ce9450f JJ |
414 | return false; |
415 | } | |
416 | ||
59a05b0c EB |
417 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
418 | return false; | |
ebfd146a | 419 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
420 | if (TREE_CODE (operand) != SSA_NAME) |
421 | return false; | |
422 | ||
423 | if (operand == use) | |
424 | return true; | |
425 | ||
426 | return false; | |
427 | } | |
428 | ||
429 | ||
b8698a0f | 430 | /* |
ebfd146a IR |
431 | Function process_use. |
432 | ||
433 | Inputs: | |
434 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b28ead45 | 435 | - RELEVANT - enum value to be set in the STMT_VINFO of the stmt |
ff802fa1 | 436 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 437 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
438 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
439 | be performed. | |
ebfd146a IR |
440 | |
441 | Outputs: | |
442 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
443 | relevance info of the DEF_STMT of this USE: | |
444 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
445 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
446 | Exceptions: | |
447 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 448 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 449 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
450 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
451 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
452 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
453 | be modified accordingly. | |
454 | ||
455 | Return true if everything is as expected. Return false otherwise. */ | |
456 | ||
457 | static bool | |
b28ead45 | 458 | process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo, |
355fe088 | 459 | enum vect_relevant relevant, vec<gimple *> *worklist, |
aec7ae7d | 460 | bool force) |
ebfd146a IR |
461 | { |
462 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
463 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
464 | stmt_vec_info dstmt_vinfo; | |
465 | basic_block bb, def_bb; | |
355fe088 | 466 | gimple *def_stmt; |
ebfd146a IR |
467 | enum vect_def_type dt; |
468 | ||
b8698a0f | 469 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 470 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 471 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
472 | return true; |
473 | ||
894dd753 | 474 | if (!vect_is_simple_use (use, loop_vinfo, &dt, &def_stmt)) |
b8698a0f | 475 | { |
73fbfcad | 476 | if (dump_enabled_p ()) |
78c60e3d | 477 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 478 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
479 | return false; |
480 | } | |
481 | ||
482 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
483 | return true; | |
484 | ||
485 | def_bb = gimple_bb (def_stmt); | |
486 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
487 | { | |
73fbfcad | 488 | if (dump_enabled_p ()) |
e645e942 | 489 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
490 | return true; |
491 | } | |
492 | ||
b8698a0f L |
493 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
494 | DEF_STMT must have already been processed, because this should be the | |
495 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
496 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
497 | check that everything is as expected, and we are done. */ |
498 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
499 | bb = gimple_bb (stmt); | |
500 | if (gimple_code (stmt) == GIMPLE_PHI | |
501 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
502 | && gimple_code (def_stmt) != GIMPLE_PHI | |
503 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
504 | && bb->loop_father == def_bb->loop_father) | |
505 | { | |
73fbfcad | 506 | if (dump_enabled_p ()) |
78c60e3d | 507 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 508 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a | 509 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); |
b8698a0f | 510 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 511 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
512 | return true; |
513 | } | |
514 | ||
515 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
516 | outer-loop-header-bb: | |
517 | d = def_stmt | |
518 | inner-loop: | |
519 | stmt # use (d) | |
520 | outer-loop-tail-bb: | |
521 | ... */ | |
522 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
523 | { | |
73fbfcad | 524 | if (dump_enabled_p ()) |
78c60e3d | 525 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 526 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 527 | |
ebfd146a IR |
528 | switch (relevant) |
529 | { | |
8644a673 | 530 | case vect_unused_in_scope: |
7c5222ff IR |
531 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
532 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 533 | break; |
7c5222ff | 534 | |
ebfd146a | 535 | case vect_used_in_outer_by_reduction: |
7c5222ff | 536 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
537 | relevant = vect_used_by_reduction; |
538 | break; | |
7c5222ff | 539 | |
ebfd146a | 540 | case vect_used_in_outer: |
7c5222ff | 541 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 542 | relevant = vect_used_in_scope; |
ebfd146a | 543 | break; |
7c5222ff | 544 | |
8644a673 | 545 | case vect_used_in_scope: |
ebfd146a IR |
546 | break; |
547 | ||
548 | default: | |
549 | gcc_unreachable (); | |
b8698a0f | 550 | } |
ebfd146a IR |
551 | } |
552 | ||
553 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
554 | outer-loop-header-bb: | |
555 | ... | |
556 | inner-loop: | |
557 | d = def_stmt | |
06066f92 | 558 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
559 | stmt # use (d) */ |
560 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
561 | { | |
73fbfcad | 562 | if (dump_enabled_p ()) |
78c60e3d | 563 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 564 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 565 | |
ebfd146a IR |
566 | switch (relevant) |
567 | { | |
8644a673 | 568 | case vect_unused_in_scope: |
b8698a0f | 569 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 570 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 571 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
572 | break; |
573 | ||
ebfd146a | 574 | case vect_used_by_reduction: |
b28ead45 | 575 | case vect_used_only_live: |
ebfd146a IR |
576 | relevant = vect_used_in_outer_by_reduction; |
577 | break; | |
578 | ||
8644a673 | 579 | case vect_used_in_scope: |
ebfd146a IR |
580 | relevant = vect_used_in_outer; |
581 | break; | |
582 | ||
583 | default: | |
584 | gcc_unreachable (); | |
585 | } | |
586 | } | |
643a9684 RB |
587 | /* We are also not interested in uses on loop PHI backedges that are |
588 | inductions. Otherwise we'll needlessly vectorize the IV increment | |
e294f495 RB |
589 | and cause hybrid SLP for SLP inductions. Unless the PHI is live |
590 | of course. */ | |
643a9684 RB |
591 | else if (gimple_code (stmt) == GIMPLE_PHI |
592 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_induction_def | |
e294f495 | 593 | && ! STMT_VINFO_LIVE_P (stmt_vinfo) |
643a9684 RB |
594 | && (PHI_ARG_DEF_FROM_EDGE (stmt, loop_latch_edge (bb->loop_father)) |
595 | == use)) | |
596 | { | |
597 | if (dump_enabled_p ()) | |
598 | dump_printf_loc (MSG_NOTE, vect_location, | |
599 | "induction value on backedge.\n"); | |
600 | return true; | |
601 | } | |
602 | ||
ebfd146a | 603 | |
b28ead45 | 604 | vect_mark_relevant (worklist, def_stmt, relevant, false); |
ebfd146a IR |
605 | return true; |
606 | } | |
607 | ||
608 | ||
609 | /* Function vect_mark_stmts_to_be_vectorized. | |
610 | ||
611 | Not all stmts in the loop need to be vectorized. For example: | |
612 | ||
613 | for i... | |
614 | for j... | |
615 | 1. T0 = i + j | |
616 | 2. T1 = a[T0] | |
617 | ||
618 | 3. j = j + 1 | |
619 | ||
620 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
621 | addressing of vectorized data-refs are handled differently. | |
622 | ||
623 | This pass detects such stmts. */ | |
624 | ||
625 | bool | |
626 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
627 | { | |
ebfd146a IR |
628 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
629 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
630 | unsigned int nbbs = loop->num_nodes; | |
631 | gimple_stmt_iterator si; | |
355fe088 | 632 | gimple *stmt; |
ebfd146a IR |
633 | unsigned int i; |
634 | stmt_vec_info stmt_vinfo; | |
635 | basic_block bb; | |
355fe088 | 636 | gimple *phi; |
ebfd146a | 637 | bool live_p; |
b28ead45 | 638 | enum vect_relevant relevant; |
ebfd146a | 639 | |
adac3a68 | 640 | DUMP_VECT_SCOPE ("vect_mark_stmts_to_be_vectorized"); |
ebfd146a | 641 | |
355fe088 | 642 | auto_vec<gimple *, 64> worklist; |
ebfd146a IR |
643 | |
644 | /* 1. Init worklist. */ | |
645 | for (i = 0; i < nbbs; i++) | |
646 | { | |
647 | bb = bbs[i]; | |
648 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 649 | { |
ebfd146a | 650 | phi = gsi_stmt (si); |
73fbfcad | 651 | if (dump_enabled_p ()) |
ebfd146a | 652 | { |
78c60e3d SS |
653 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
654 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
655 | } |
656 | ||
657 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 658 | vect_mark_relevant (&worklist, phi, relevant, live_p); |
ebfd146a IR |
659 | } |
660 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
661 | { | |
662 | stmt = gsi_stmt (si); | |
73fbfcad | 663 | if (dump_enabled_p ()) |
ebfd146a | 664 | { |
78c60e3d SS |
665 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
666 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 667 | } |
ebfd146a IR |
668 | |
669 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 670 | vect_mark_relevant (&worklist, stmt, relevant, live_p); |
ebfd146a IR |
671 | } |
672 | } | |
673 | ||
674 | /* 2. Process_worklist */ | |
9771b263 | 675 | while (worklist.length () > 0) |
ebfd146a IR |
676 | { |
677 | use_operand_p use_p; | |
678 | ssa_op_iter iter; | |
679 | ||
9771b263 | 680 | stmt = worklist.pop (); |
73fbfcad | 681 | if (dump_enabled_p ()) |
ebfd146a | 682 | { |
78c60e3d SS |
683 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
684 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
685 | } |
686 | ||
b8698a0f | 687 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
b28ead45 AH |
688 | (DEF_STMT) as relevant/irrelevant according to the relevance property |
689 | of STMT. */ | |
ebfd146a IR |
690 | stmt_vinfo = vinfo_for_stmt (stmt); |
691 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
ebfd146a | 692 | |
b28ead45 AH |
693 | /* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is |
694 | propagated as is to the DEF_STMTs of its USEs. | |
ebfd146a IR |
695 | |
696 | One exception is when STMT has been identified as defining a reduction | |
b28ead45 | 697 | variable; in this case we set the relevance to vect_used_by_reduction. |
ebfd146a | 698 | This is because we distinguish between two kinds of relevant stmts - |
b8698a0f | 699 | those that are used by a reduction computation, and those that are |
ff802fa1 | 700 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 701 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 702 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 703 | |
b28ead45 | 704 | switch (STMT_VINFO_DEF_TYPE (stmt_vinfo)) |
ebfd146a | 705 | { |
06066f92 | 706 | case vect_reduction_def: |
b28ead45 AH |
707 | gcc_assert (relevant != vect_unused_in_scope); |
708 | if (relevant != vect_unused_in_scope | |
709 | && relevant != vect_used_in_scope | |
710 | && relevant != vect_used_by_reduction | |
711 | && relevant != vect_used_only_live) | |
06066f92 | 712 | { |
b28ead45 AH |
713 | if (dump_enabled_p ()) |
714 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
715 | "unsupported use of reduction.\n"); | |
716 | return false; | |
06066f92 | 717 | } |
06066f92 | 718 | break; |
b8698a0f | 719 | |
06066f92 | 720 | case vect_nested_cycle: |
b28ead45 AH |
721 | if (relevant != vect_unused_in_scope |
722 | && relevant != vect_used_in_outer_by_reduction | |
723 | && relevant != vect_used_in_outer) | |
06066f92 | 724 | { |
73fbfcad | 725 | if (dump_enabled_p ()) |
78c60e3d | 726 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 727 | "unsupported use of nested cycle.\n"); |
7c5222ff | 728 | |
06066f92 IR |
729 | return false; |
730 | } | |
b8698a0f L |
731 | break; |
732 | ||
06066f92 | 733 | case vect_double_reduction_def: |
b28ead45 AH |
734 | if (relevant != vect_unused_in_scope |
735 | && relevant != vect_used_by_reduction | |
736 | && relevant != vect_used_only_live) | |
06066f92 | 737 | { |
73fbfcad | 738 | if (dump_enabled_p ()) |
78c60e3d | 739 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 740 | "unsupported use of double reduction.\n"); |
7c5222ff | 741 | |
7c5222ff | 742 | return false; |
06066f92 | 743 | } |
b8698a0f | 744 | break; |
7c5222ff | 745 | |
06066f92 IR |
746 | default: |
747 | break; | |
7c5222ff | 748 | } |
b8698a0f | 749 | |
aec7ae7d | 750 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
751 | { |
752 | /* Pattern statements are not inserted into the code, so | |
753 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
754 | have to scan the RHS or function arguments instead. */ | |
755 | if (is_gimple_assign (stmt)) | |
756 | { | |
69d2aade JJ |
757 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
758 | tree op = gimple_assign_rhs1 (stmt); | |
759 | ||
760 | i = 1; | |
761 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
762 | { | |
763 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
b28ead45 | 764 | relevant, &worklist, false) |
69d2aade | 765 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
b28ead45 | 766 | relevant, &worklist, false)) |
566d377a | 767 | return false; |
69d2aade JJ |
768 | i = 2; |
769 | } | |
770 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 771 | { |
69d2aade | 772 | op = gimple_op (stmt, i); |
afbe6325 | 773 | if (TREE_CODE (op) == SSA_NAME |
b28ead45 | 774 | && !process_use (stmt, op, loop_vinfo, relevant, |
afbe6325 | 775 | &worklist, false)) |
07687835 | 776 | return false; |
9d5e7640 IR |
777 | } |
778 | } | |
779 | else if (is_gimple_call (stmt)) | |
780 | { | |
781 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
782 | { | |
783 | tree arg = gimple_call_arg (stmt, i); | |
b28ead45 | 784 | if (!process_use (stmt, arg, loop_vinfo, relevant, |
aec7ae7d | 785 | &worklist, false)) |
07687835 | 786 | return false; |
9d5e7640 IR |
787 | } |
788 | } | |
789 | } | |
790 | else | |
791 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
792 | { | |
793 | tree op = USE_FROM_PTR (use_p); | |
b28ead45 | 794 | if (!process_use (stmt, op, loop_vinfo, relevant, |
aec7ae7d | 795 | &worklist, false)) |
07687835 | 796 | return false; |
9d5e7640 | 797 | } |
aec7ae7d | 798 | |
3bab6342 | 799 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d | 800 | { |
134c85ca RS |
801 | gather_scatter_info gs_info; |
802 | if (!vect_check_gather_scatter (stmt, loop_vinfo, &gs_info)) | |
803 | gcc_unreachable (); | |
804 | if (!process_use (stmt, gs_info.offset, loop_vinfo, relevant, | |
805 | &worklist, true)) | |
566d377a | 806 | return false; |
aec7ae7d | 807 | } |
ebfd146a IR |
808 | } /* while worklist */ |
809 | ||
ebfd146a IR |
810 | return true; |
811 | } | |
812 | ||
68435eb2 RB |
813 | /* Compute the prologue cost for invariant or constant operands. */ |
814 | ||
815 | static unsigned | |
816 | vect_prologue_cost_for_slp_op (slp_tree node, stmt_vec_info stmt_info, | |
817 | unsigned opno, enum vect_def_type dt, | |
818 | stmt_vector_for_cost *cost_vec) | |
819 | { | |
820 | gimple *stmt = SLP_TREE_SCALAR_STMTS (node)[0]; | |
821 | tree op = gimple_op (stmt, opno); | |
822 | unsigned prologue_cost = 0; | |
823 | ||
824 | /* Without looking at the actual initializer a vector of | |
825 | constants can be implemented as load from the constant pool. | |
826 | When all elements are the same we can use a splat. */ | |
827 | tree vectype = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
828 | unsigned group_size = SLP_TREE_SCALAR_STMTS (node).length (); | |
829 | unsigned num_vects_to_check; | |
830 | unsigned HOST_WIDE_INT const_nunits; | |
831 | unsigned nelt_limit; | |
832 | if (TYPE_VECTOR_SUBPARTS (vectype).is_constant (&const_nunits) | |
833 | && ! multiple_p (const_nunits, group_size)) | |
834 | { | |
835 | num_vects_to_check = SLP_TREE_NUMBER_OF_VEC_STMTS (node); | |
836 | nelt_limit = const_nunits; | |
837 | } | |
838 | else | |
839 | { | |
840 | /* If either the vector has variable length or the vectors | |
841 | are composed of repeated whole groups we only need to | |
842 | cost construction once. All vectors will be the same. */ | |
843 | num_vects_to_check = 1; | |
844 | nelt_limit = group_size; | |
845 | } | |
846 | tree elt = NULL_TREE; | |
847 | unsigned nelt = 0; | |
848 | for (unsigned j = 0; j < num_vects_to_check * nelt_limit; ++j) | |
849 | { | |
850 | unsigned si = j % group_size; | |
851 | if (nelt == 0) | |
852 | elt = gimple_op (SLP_TREE_SCALAR_STMTS (node)[si], opno); | |
853 | /* ??? We're just tracking whether all operands of a single | |
854 | vector initializer are the same, ideally we'd check if | |
855 | we emitted the same one already. */ | |
856 | else if (elt != gimple_op (SLP_TREE_SCALAR_STMTS (node)[si], | |
857 | opno)) | |
858 | elt = NULL_TREE; | |
859 | nelt++; | |
860 | if (nelt == nelt_limit) | |
861 | { | |
862 | /* ??? We need to pass down stmt_info for a vector type | |
863 | even if it points to the wrong stmt. */ | |
864 | prologue_cost += record_stmt_cost | |
865 | (cost_vec, 1, | |
866 | dt == vect_external_def | |
867 | ? (elt ? scalar_to_vec : vec_construct) | |
868 | : vector_load, | |
869 | stmt_info, 0, vect_prologue); | |
870 | nelt = 0; | |
871 | } | |
872 | } | |
873 | ||
874 | return prologue_cost; | |
875 | } | |
ebfd146a | 876 | |
b8698a0f | 877 | /* Function vect_model_simple_cost. |
ebfd146a | 878 | |
b8698a0f | 879 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
880 | single op. Right now, this does not account for multiple insns that could |
881 | be generated for the single vector op. We will handle that shortly. */ | |
882 | ||
68435eb2 | 883 | static void |
b8698a0f | 884 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 | 885 | enum vect_def_type *dt, |
4fc5ebf1 | 886 | int ndts, |
68435eb2 RB |
887 | slp_tree node, |
888 | stmt_vector_for_cost *cost_vec) | |
ebfd146a | 889 | { |
92345349 | 890 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a | 891 | |
68435eb2 | 892 | gcc_assert (cost_vec != NULL); |
ebfd146a | 893 | |
68435eb2 RB |
894 | /* ??? Somehow we need to fix this at the callers. */ |
895 | if (node) | |
896 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (node); | |
897 | ||
898 | if (node) | |
899 | { | |
900 | /* Scan operands and account for prologue cost of constants/externals. | |
901 | ??? This over-estimates cost for multiple uses and should be | |
902 | re-engineered. */ | |
903 | gimple *stmt = SLP_TREE_SCALAR_STMTS (node)[0]; | |
904 | tree lhs = gimple_get_lhs (stmt); | |
905 | for (unsigned i = 0; i < gimple_num_ops (stmt); ++i) | |
906 | { | |
907 | tree op = gimple_op (stmt, i); | |
68435eb2 RB |
908 | enum vect_def_type dt; |
909 | if (!op || op == lhs) | |
910 | continue; | |
894dd753 | 911 | if (vect_is_simple_use (op, stmt_info->vinfo, &dt) |
68435eb2 RB |
912 | && (dt == vect_constant_def || dt == vect_external_def)) |
913 | prologue_cost += vect_prologue_cost_for_slp_op (node, stmt_info, | |
914 | i, dt, cost_vec); | |
915 | } | |
916 | } | |
917 | else | |
918 | /* Cost the "broadcast" of a scalar operand in to a vector operand. | |
919 | Use scalar_to_vec to cost the broadcast, as elsewhere in the vector | |
920 | cost model. */ | |
921 | for (int i = 0; i < ndts; i++) | |
922 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) | |
923 | prologue_cost += record_stmt_cost (cost_vec, 1, scalar_to_vec, | |
924 | stmt_info, 0, vect_prologue); | |
925 | ||
926 | /* Adjust for two-operator SLP nodes. */ | |
927 | if (node && SLP_TREE_TWO_OPERATORS (node)) | |
928 | { | |
929 | ncopies *= 2; | |
930 | inside_cost += record_stmt_cost (cost_vec, ncopies, vec_perm, | |
931 | stmt_info, 0, vect_body); | |
932 | } | |
c3e7ee41 BS |
933 | |
934 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
68435eb2 RB |
935 | inside_cost += record_stmt_cost (cost_vec, ncopies, vector_stmt, |
936 | stmt_info, 0, vect_body); | |
c3e7ee41 | 937 | |
73fbfcad | 938 | if (dump_enabled_p ()) |
78c60e3d SS |
939 | dump_printf_loc (MSG_NOTE, vect_location, |
940 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 941 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
942 | } |
943 | ||
944 | ||
8bd37302 BS |
945 | /* Model cost for type demotion and promotion operations. PWR is normally |
946 | zero for single-step promotions and demotions. It will be one if | |
947 | two-step promotion/demotion is required, and so on. Each additional | |
948 | step doubles the number of instructions required. */ | |
949 | ||
950 | static void | |
951 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
68435eb2 RB |
952 | enum vect_def_type *dt, int pwr, |
953 | stmt_vector_for_cost *cost_vec) | |
8bd37302 BS |
954 | { |
955 | int i, tmp; | |
92345349 | 956 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 | 957 | |
8bd37302 BS |
958 | for (i = 0; i < pwr + 1; i++) |
959 | { | |
960 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
961 | (i + 1) : i; | |
68435eb2 RB |
962 | inside_cost += record_stmt_cost (cost_vec, vect_pow2 (tmp), |
963 | vec_promote_demote, stmt_info, 0, | |
964 | vect_body); | |
8bd37302 BS |
965 | } |
966 | ||
967 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
968 | for (i = 0; i < 2; i++) | |
92345349 | 969 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
68435eb2 RB |
970 | prologue_cost += record_stmt_cost (cost_vec, 1, vector_stmt, |
971 | stmt_info, 0, vect_prologue); | |
8bd37302 | 972 | |
73fbfcad | 973 | if (dump_enabled_p ()) |
78c60e3d SS |
974 | dump_printf_loc (MSG_NOTE, vect_location, |
975 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 976 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
977 | } |
978 | ||
ebfd146a IR |
979 | /* Function vect_model_store_cost |
980 | ||
0d0293ac MM |
981 | Models cost for stores. In the case of grouped accesses, one access |
982 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a | 983 | |
68435eb2 | 984 | static void |
b8698a0f | 985 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
68435eb2 | 986 | enum vect_def_type dt, |
2de001ee | 987 | vect_memory_access_type memory_access_type, |
9ce4345a | 988 | vec_load_store_type vls_type, slp_tree slp_node, |
68435eb2 | 989 | stmt_vector_for_cost *cost_vec) |
ebfd146a | 990 | { |
92345349 | 991 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f RS |
992 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); |
993 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); | |
ebfd146a | 994 | |
68435eb2 RB |
995 | /* ??? Somehow we need to fix this at the callers. */ |
996 | if (slp_node) | |
997 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
998 | ||
9ce4345a | 999 | if (vls_type == VLS_STORE_INVARIANT) |
68435eb2 RB |
1000 | { |
1001 | if (slp_node) | |
1002 | prologue_cost += vect_prologue_cost_for_slp_op (slp_node, stmt_info, | |
1003 | 1, dt, cost_vec); | |
1004 | else | |
1005 | prologue_cost += record_stmt_cost (cost_vec, 1, scalar_to_vec, | |
1006 | stmt_info, 0, vect_prologue); | |
1007 | } | |
ebfd146a | 1008 | |
892a981f RS |
1009 | /* Grouped stores update all elements in the group at once, |
1010 | so we want the DR for the first statement. */ | |
1011 | if (!slp_node && grouped_access_p) | |
57c454d2 | 1012 | first_stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a | 1013 | |
892a981f RS |
1014 | /* True if we should include any once-per-group costs as well as |
1015 | the cost of the statement itself. For SLP we only get called | |
1016 | once per group anyhow. */ | |
1017 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
1018 | ||
272c6793 | 1019 | /* We assume that the cost of a single store-lanes instruction is |
2c53b149 | 1020 | equivalent to the cost of DR_GROUP_SIZE separate stores. If a grouped |
272c6793 | 1021 | access is instead being provided by a permute-and-store operation, |
2de001ee RS |
1022 | include the cost of the permutes. */ |
1023 | if (first_stmt_p | |
1024 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 1025 | { |
e1377713 ES |
1026 | /* Uses a high and low interleave or shuffle operations for each |
1027 | needed permute. */ | |
2c53b149 | 1028 | int group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
e1377713 | 1029 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
68435eb2 | 1030 | inside_cost = record_stmt_cost (cost_vec, nstmts, vec_perm, |
92345349 | 1031 | stmt_info, 0, vect_body); |
ebfd146a | 1032 | |
73fbfcad | 1033 | if (dump_enabled_p ()) |
78c60e3d | 1034 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1035 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 1036 | group_size); |
ebfd146a IR |
1037 | } |
1038 | ||
cee62fee | 1039 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 1040 | /* Costs of the stores. */ |
067bc855 RB |
1041 | if (memory_access_type == VMAT_ELEMENTWISE |
1042 | || memory_access_type == VMAT_GATHER_SCATTER) | |
c5126ce8 RS |
1043 | { |
1044 | /* N scalar stores plus extracting the elements. */ | |
1045 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
68435eb2 | 1046 | inside_cost += record_stmt_cost (cost_vec, |
c5126ce8 RS |
1047 | ncopies * assumed_nunits, |
1048 | scalar_store, stmt_info, 0, vect_body); | |
1049 | } | |
f2e2a985 | 1050 | else |
57c454d2 | 1051 | vect_get_store_cost (stmt_info, ncopies, &inside_cost, cost_vec); |
ebfd146a | 1052 | |
2de001ee RS |
1053 | if (memory_access_type == VMAT_ELEMENTWISE |
1054 | || memory_access_type == VMAT_STRIDED_SLP) | |
c5126ce8 RS |
1055 | { |
1056 | /* N scalar stores plus extracting the elements. */ | |
1057 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
68435eb2 | 1058 | inside_cost += record_stmt_cost (cost_vec, |
c5126ce8 RS |
1059 | ncopies * assumed_nunits, |
1060 | vec_to_scalar, stmt_info, 0, vect_body); | |
1061 | } | |
cee62fee | 1062 | |
73fbfcad | 1063 | if (dump_enabled_p ()) |
78c60e3d SS |
1064 | dump_printf_loc (MSG_NOTE, vect_location, |
1065 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 1066 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
1067 | } |
1068 | ||
1069 | ||
720f5239 IR |
1070 | /* Calculate cost of DR's memory access. */ |
1071 | void | |
57c454d2 | 1072 | vect_get_store_cost (stmt_vec_info stmt_info, int ncopies, |
c3e7ee41 | 1073 | unsigned int *inside_cost, |
92345349 | 1074 | stmt_vector_for_cost *body_cost_vec) |
720f5239 | 1075 | { |
57c454d2 | 1076 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
720f5239 IR |
1077 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
1078 | ||
1079 | switch (alignment_support_scheme) | |
1080 | { | |
1081 | case dr_aligned: | |
1082 | { | |
92345349 BS |
1083 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
1084 | vector_store, stmt_info, 0, | |
1085 | vect_body); | |
720f5239 | 1086 | |
73fbfcad | 1087 | if (dump_enabled_p ()) |
78c60e3d | 1088 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1089 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
1090 | break; |
1091 | } | |
1092 | ||
1093 | case dr_unaligned_supported: | |
1094 | { | |
720f5239 | 1095 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1096 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1097 | unaligned_store, stmt_info, |
92345349 | 1098 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1099 | if (dump_enabled_p ()) |
78c60e3d SS |
1100 | dump_printf_loc (MSG_NOTE, vect_location, |
1101 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1102 | "hardware.\n"); |
720f5239 IR |
1103 | break; |
1104 | } | |
1105 | ||
38eec4c6 UW |
1106 | case dr_unaligned_unsupported: |
1107 | { | |
1108 | *inside_cost = VECT_MAX_COST; | |
1109 | ||
73fbfcad | 1110 | if (dump_enabled_p ()) |
78c60e3d | 1111 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1112 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1113 | break; |
1114 | } | |
1115 | ||
720f5239 IR |
1116 | default: |
1117 | gcc_unreachable (); | |
1118 | } | |
1119 | } | |
1120 | ||
1121 | ||
ebfd146a IR |
1122 | /* Function vect_model_load_cost |
1123 | ||
892a981f RS |
1124 | Models cost for loads. In the case of grouped accesses, one access has |
1125 | the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1126 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1127 | access scheme chosen. */ |
1128 | ||
68435eb2 RB |
1129 | static void |
1130 | vect_model_load_cost (stmt_vec_info stmt_info, unsigned ncopies, | |
2de001ee | 1131 | vect_memory_access_type memory_access_type, |
68435eb2 | 1132 | slp_instance instance, |
2de001ee | 1133 | slp_tree slp_node, |
68435eb2 | 1134 | stmt_vector_for_cost *cost_vec) |
ebfd146a | 1135 | { |
892a981f | 1136 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); |
92345349 | 1137 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f | 1138 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
ebfd146a | 1139 | |
68435eb2 RB |
1140 | gcc_assert (cost_vec); |
1141 | ||
1142 | /* ??? Somehow we need to fix this at the callers. */ | |
1143 | if (slp_node) | |
1144 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
1145 | ||
1146 | if (slp_node && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
1147 | { | |
1148 | /* If the load is permuted then the alignment is determined by | |
1149 | the first group element not by the first scalar stmt DR. */ | |
2c53b149 | 1150 | gimple *stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
68435eb2 RB |
1151 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
1152 | /* Record the cost for the permutation. */ | |
1153 | unsigned n_perms; | |
1154 | unsigned assumed_nunits | |
1155 | = vect_nunits_for_cost (STMT_VINFO_VECTYPE (stmt_info)); | |
1156 | unsigned slp_vf = (ncopies * assumed_nunits) / instance->group_size; | |
1157 | vect_transform_slp_perm_load (slp_node, vNULL, NULL, | |
1158 | slp_vf, instance, true, | |
1159 | &n_perms); | |
1160 | inside_cost += record_stmt_cost (cost_vec, n_perms, vec_perm, | |
1161 | stmt_info, 0, vect_body); | |
1162 | /* And adjust the number of loads performed. This handles | |
1163 | redundancies as well as loads that are later dead. */ | |
2c53b149 | 1164 | auto_sbitmap perm (DR_GROUP_SIZE (stmt_info)); |
68435eb2 RB |
1165 | bitmap_clear (perm); |
1166 | for (unsigned i = 0; | |
1167 | i < SLP_TREE_LOAD_PERMUTATION (slp_node).length (); ++i) | |
1168 | bitmap_set_bit (perm, SLP_TREE_LOAD_PERMUTATION (slp_node)[i]); | |
1169 | ncopies = 0; | |
1170 | bool load_seen = false; | |
2c53b149 | 1171 | for (unsigned i = 0; i < DR_GROUP_SIZE (stmt_info); ++i) |
68435eb2 RB |
1172 | { |
1173 | if (i % assumed_nunits == 0) | |
1174 | { | |
1175 | if (load_seen) | |
1176 | ncopies++; | |
1177 | load_seen = false; | |
1178 | } | |
1179 | if (bitmap_bit_p (perm, i)) | |
1180 | load_seen = true; | |
1181 | } | |
1182 | if (load_seen) | |
1183 | ncopies++; | |
1184 | gcc_assert (ncopies | |
2c53b149 | 1185 | <= (DR_GROUP_SIZE (stmt_info) - DR_GROUP_GAP (stmt_info) |
68435eb2 RB |
1186 | + assumed_nunits - 1) / assumed_nunits); |
1187 | } | |
1188 | ||
892a981f RS |
1189 | /* Grouped loads read all elements in the group at once, |
1190 | so we want the DR for the first statement. */ | |
1191 | if (!slp_node && grouped_access_p) | |
57c454d2 | 1192 | first_stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a | 1193 | |
892a981f RS |
1194 | /* True if we should include any once-per-group costs as well as |
1195 | the cost of the statement itself. For SLP we only get called | |
1196 | once per group anyhow. */ | |
1197 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
1198 | ||
272c6793 | 1199 | /* We assume that the cost of a single load-lanes instruction is |
2c53b149 | 1200 | equivalent to the cost of DR_GROUP_SIZE separate loads. If a grouped |
272c6793 | 1201 | access is instead being provided by a load-and-permute operation, |
2de001ee RS |
1202 | include the cost of the permutes. */ |
1203 | if (first_stmt_p | |
1204 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 1205 | { |
2c23db6d ES |
1206 | /* Uses an even and odd extract operations or shuffle operations |
1207 | for each needed permute. */ | |
2c53b149 | 1208 | int group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
2c23db6d | 1209 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
68435eb2 RB |
1210 | inside_cost += record_stmt_cost (cost_vec, nstmts, vec_perm, |
1211 | stmt_info, 0, vect_body); | |
ebfd146a | 1212 | |
73fbfcad | 1213 | if (dump_enabled_p ()) |
e645e942 TJ |
1214 | dump_printf_loc (MSG_NOTE, vect_location, |
1215 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1216 | group_size); |
ebfd146a IR |
1217 | } |
1218 | ||
1219 | /* The loads themselves. */ | |
067bc855 RB |
1220 | if (memory_access_type == VMAT_ELEMENTWISE |
1221 | || memory_access_type == VMAT_GATHER_SCATTER) | |
a82960aa | 1222 | { |
a21892ad BS |
1223 | /* N scalar loads plus gathering them into a vector. */ |
1224 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
c5126ce8 | 1225 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); |
68435eb2 | 1226 | inside_cost += record_stmt_cost (cost_vec, |
c5126ce8 | 1227 | ncopies * assumed_nunits, |
92345349 | 1228 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1229 | } |
1230 | else | |
57c454d2 | 1231 | vect_get_load_cost (stmt_info, ncopies, first_stmt_p, |
92345349 | 1232 | &inside_cost, &prologue_cost, |
68435eb2 | 1233 | cost_vec, cost_vec, true); |
2de001ee RS |
1234 | if (memory_access_type == VMAT_ELEMENTWISE |
1235 | || memory_access_type == VMAT_STRIDED_SLP) | |
68435eb2 | 1236 | inside_cost += record_stmt_cost (cost_vec, ncopies, vec_construct, |
892a981f | 1237 | stmt_info, 0, vect_body); |
720f5239 | 1238 | |
73fbfcad | 1239 | if (dump_enabled_p ()) |
78c60e3d SS |
1240 | dump_printf_loc (MSG_NOTE, vect_location, |
1241 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1242 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1243 | } |
1244 | ||
1245 | ||
1246 | /* Calculate cost of DR's memory access. */ | |
1247 | void | |
57c454d2 | 1248 | vect_get_load_cost (stmt_vec_info stmt_info, int ncopies, |
c3e7ee41 | 1249 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1250 | unsigned int *prologue_cost, |
1251 | stmt_vector_for_cost *prologue_cost_vec, | |
1252 | stmt_vector_for_cost *body_cost_vec, | |
1253 | bool record_prologue_costs) | |
720f5239 | 1254 | { |
57c454d2 | 1255 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
720f5239 IR |
1256 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
1257 | ||
1258 | switch (alignment_support_scheme) | |
ebfd146a IR |
1259 | { |
1260 | case dr_aligned: | |
1261 | { | |
92345349 BS |
1262 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1263 | stmt_info, 0, vect_body); | |
ebfd146a | 1264 | |
73fbfcad | 1265 | if (dump_enabled_p ()) |
78c60e3d | 1266 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1267 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1268 | |
1269 | break; | |
1270 | } | |
1271 | case dr_unaligned_supported: | |
1272 | { | |
720f5239 | 1273 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1274 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1275 | unaligned_load, stmt_info, |
92345349 | 1276 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1277 | |
73fbfcad | 1278 | if (dump_enabled_p ()) |
78c60e3d SS |
1279 | dump_printf_loc (MSG_NOTE, vect_location, |
1280 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1281 | "hardware.\n"); |
ebfd146a IR |
1282 | |
1283 | break; | |
1284 | } | |
1285 | case dr_explicit_realign: | |
1286 | { | |
92345349 BS |
1287 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1288 | vector_load, stmt_info, 0, vect_body); | |
1289 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1290 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1291 | |
1292 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1293 | the containing loop, the following cost should be added to the | |
92345349 | 1294 | prologue costs. */ |
ebfd146a | 1295 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1296 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1297 | stmt_info, 0, vect_body); | |
ebfd146a | 1298 | |
73fbfcad | 1299 | if (dump_enabled_p ()) |
e645e942 TJ |
1300 | dump_printf_loc (MSG_NOTE, vect_location, |
1301 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1302 | |
ebfd146a IR |
1303 | break; |
1304 | } | |
1305 | case dr_explicit_realign_optimized: | |
1306 | { | |
73fbfcad | 1307 | if (dump_enabled_p ()) |
e645e942 | 1308 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1309 | "vect_model_load_cost: unaligned software " |
e645e942 | 1310 | "pipelined.\n"); |
ebfd146a IR |
1311 | |
1312 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1313 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1314 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1315 | access, then the above cost should only be considered for one |
ff802fa1 | 1316 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1317 | and a realignment op. */ |
1318 | ||
92345349 | 1319 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1320 | { |
92345349 BS |
1321 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1322 | vector_stmt, stmt_info, | |
1323 | 0, vect_prologue); | |
ebfd146a | 1324 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1325 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1326 | vector_stmt, stmt_info, | |
1327 | 0, vect_prologue); | |
ebfd146a IR |
1328 | } |
1329 | ||
92345349 BS |
1330 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1331 | stmt_info, 0, vect_body); | |
1332 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1333 | stmt_info, 0, vect_body); | |
8bd37302 | 1334 | |
73fbfcad | 1335 | if (dump_enabled_p ()) |
78c60e3d | 1336 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1337 | "vect_model_load_cost: explicit realign optimized" |
1338 | "\n"); | |
8bd37302 | 1339 | |
ebfd146a IR |
1340 | break; |
1341 | } | |
1342 | ||
38eec4c6 UW |
1343 | case dr_unaligned_unsupported: |
1344 | { | |
1345 | *inside_cost = VECT_MAX_COST; | |
1346 | ||
73fbfcad | 1347 | if (dump_enabled_p ()) |
78c60e3d | 1348 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1349 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1350 | break; |
1351 | } | |
1352 | ||
ebfd146a IR |
1353 | default: |
1354 | gcc_unreachable (); | |
1355 | } | |
ebfd146a IR |
1356 | } |
1357 | ||
418b7df3 RG |
1358 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1359 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1360 | |
418b7df3 | 1361 | static void |
355fe088 | 1362 | vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi) |
ebfd146a | 1363 | { |
ebfd146a | 1364 | if (gsi) |
418b7df3 | 1365 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1366 | else |
1367 | { | |
418b7df3 | 1368 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1369 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1370 | |
a70d6342 IR |
1371 | if (loop_vinfo) |
1372 | { | |
1373 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1374 | basic_block new_bb; |
1375 | edge pe; | |
a70d6342 IR |
1376 | |
1377 | if (nested_in_vect_loop_p (loop, stmt)) | |
1378 | loop = loop->inner; | |
b8698a0f | 1379 | |
a70d6342 | 1380 | pe = loop_preheader_edge (loop); |
418b7df3 | 1381 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1382 | gcc_assert (!new_bb); |
1383 | } | |
1384 | else | |
1385 | { | |
1386 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1387 | basic_block bb; | |
1388 | gimple_stmt_iterator gsi_bb_start; | |
1389 | ||
1390 | gcc_assert (bb_vinfo); | |
1391 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1392 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1393 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1394 | } |
ebfd146a IR |
1395 | } |
1396 | ||
73fbfcad | 1397 | if (dump_enabled_p ()) |
ebfd146a | 1398 | { |
78c60e3d SS |
1399 | dump_printf_loc (MSG_NOTE, vect_location, |
1400 | "created new init_stmt: "); | |
1401 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1402 | } |
418b7df3 RG |
1403 | } |
1404 | ||
1405 | /* Function vect_init_vector. | |
ebfd146a | 1406 | |
5467ee52 RG |
1407 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1408 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1409 | vector type a vector with all elements equal to VAL is created first. | |
1410 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1411 | initialization at the loop preheader. | |
418b7df3 RG |
1412 | Return the DEF of INIT_STMT. |
1413 | It will be used in the vectorization of STMT. */ | |
1414 | ||
1415 | tree | |
355fe088 | 1416 | vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 | 1417 | { |
355fe088 | 1418 | gimple *init_stmt; |
418b7df3 RG |
1419 | tree new_temp; |
1420 | ||
e412ece4 RB |
1421 | /* We abuse this function to push sth to a SSA name with initial 'val'. */ |
1422 | if (! useless_type_conversion_p (type, TREE_TYPE (val))) | |
418b7df3 | 1423 | { |
e412ece4 RB |
1424 | gcc_assert (TREE_CODE (type) == VECTOR_TYPE); |
1425 | if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) | |
418b7df3 | 1426 | { |
5a308cf1 IE |
1427 | /* Scalar boolean value should be transformed into |
1428 | all zeros or all ones value before building a vector. */ | |
1429 | if (VECTOR_BOOLEAN_TYPE_P (type)) | |
1430 | { | |
b3d51f23 IE |
1431 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1432 | tree false_val = build_zero_cst (TREE_TYPE (type)); | |
5a308cf1 IE |
1433 | |
1434 | if (CONSTANT_CLASS_P (val)) | |
1435 | val = integer_zerop (val) ? false_val : true_val; | |
1436 | else | |
1437 | { | |
1438 | new_temp = make_ssa_name (TREE_TYPE (type)); | |
1439 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, | |
1440 | val, true_val, false_val); | |
1441 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
1442 | val = new_temp; | |
1443 | } | |
1444 | } | |
1445 | else if (CONSTANT_CLASS_P (val)) | |
42fd8198 | 1446 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1447 | else |
1448 | { | |
b731b390 | 1449 | new_temp = make_ssa_name (TREE_TYPE (type)); |
e412ece4 RB |
1450 | if (! INTEGRAL_TYPE_P (TREE_TYPE (val))) |
1451 | init_stmt = gimple_build_assign (new_temp, | |
1452 | fold_build1 (VIEW_CONVERT_EXPR, | |
1453 | TREE_TYPE (type), | |
1454 | val)); | |
1455 | else | |
1456 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); | |
418b7df3 | 1457 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1458 | val = new_temp; |
418b7df3 RG |
1459 | } |
1460 | } | |
5467ee52 | 1461 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1462 | } |
1463 | ||
0e22bb5a RB |
1464 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1465 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1466 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1467 | return new_temp; |
ebfd146a IR |
1468 | } |
1469 | ||
c83a894c | 1470 | /* Function vect_get_vec_def_for_operand_1. |
a70d6342 | 1471 | |
c83a894c AH |
1472 | For a defining stmt DEF_STMT of a scalar stmt, return a vector def with type |
1473 | DT that will be used in the vectorized stmt. */ | |
ebfd146a IR |
1474 | |
1475 | tree | |
c83a894c | 1476 | vect_get_vec_def_for_operand_1 (gimple *def_stmt, enum vect_def_type dt) |
ebfd146a IR |
1477 | { |
1478 | tree vec_oprnd; | |
355fe088 | 1479 | gimple *vec_stmt; |
ebfd146a | 1480 | stmt_vec_info def_stmt_info = NULL; |
ebfd146a IR |
1481 | |
1482 | switch (dt) | |
1483 | { | |
81c40241 | 1484 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1485 | case vect_constant_def: |
81c40241 | 1486 | case vect_external_def: |
c83a894c AH |
1487 | /* Code should use vect_get_vec_def_for_operand. */ |
1488 | gcc_unreachable (); | |
ebfd146a | 1489 | |
81c40241 | 1490 | /* operand is defined inside the loop. */ |
8644a673 | 1491 | case vect_internal_def: |
ebfd146a | 1492 | { |
ebfd146a IR |
1493 | /* Get the def from the vectorized stmt. */ |
1494 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1495 | |
ebfd146a | 1496 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1497 | /* Get vectorized pattern statement. */ |
1498 | if (!vec_stmt | |
1499 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1500 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1501 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1502 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1503 | gcc_assert (vec_stmt); |
1504 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1505 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1506 | else if (is_gimple_call (vec_stmt)) | |
1507 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1508 | else | |
1509 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1510 | return vec_oprnd; | |
1511 | } | |
1512 | ||
c78e3652 | 1513 | /* operand is defined by a loop header phi. */ |
ebfd146a | 1514 | case vect_reduction_def: |
06066f92 | 1515 | case vect_double_reduction_def: |
7c5222ff | 1516 | case vect_nested_cycle: |
ebfd146a IR |
1517 | case vect_induction_def: |
1518 | { | |
1519 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1520 | ||
1521 | /* Get the def from the vectorized stmt. */ | |
1522 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1523 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1524 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1525 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1526 | else | |
1527 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1528 | return vec_oprnd; |
1529 | } | |
1530 | ||
1531 | default: | |
1532 | gcc_unreachable (); | |
1533 | } | |
1534 | } | |
1535 | ||
1536 | ||
c83a894c AH |
1537 | /* Function vect_get_vec_def_for_operand. |
1538 | ||
1539 | OP is an operand in STMT. This function returns a (vector) def that will be | |
1540 | used in the vectorized stmt for STMT. | |
1541 | ||
1542 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1543 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1544 | ||
1545 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1546 | needs to be introduced. VECTYPE may be used to specify a required type for | |
1547 | vector invariant. */ | |
1548 | ||
1549 | tree | |
1550 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) | |
1551 | { | |
1552 | gimple *def_stmt; | |
1553 | enum vect_def_type dt; | |
1554 | bool is_simple_use; | |
1555 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
1556 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
1557 | ||
1558 | if (dump_enabled_p ()) | |
1559 | { | |
1560 | dump_printf_loc (MSG_NOTE, vect_location, | |
1561 | "vect_get_vec_def_for_operand: "); | |
1562 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
1563 | dump_printf (MSG_NOTE, "\n"); | |
1564 | } | |
1565 | ||
894dd753 | 1566 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &dt, &def_stmt); |
c83a894c AH |
1567 | gcc_assert (is_simple_use); |
1568 | if (def_stmt && dump_enabled_p ()) | |
1569 | { | |
1570 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1571 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
1572 | } | |
1573 | ||
1574 | if (dt == vect_constant_def || dt == vect_external_def) | |
1575 | { | |
1576 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
1577 | tree vector_type; | |
1578 | ||
1579 | if (vectype) | |
1580 | vector_type = vectype; | |
2568d8a1 | 1581 | else if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op)) |
c83a894c AH |
1582 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) |
1583 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1584 | else | |
1585 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1586 | ||
1587 | gcc_assert (vector_type); | |
1588 | return vect_init_vector (stmt, op, vector_type, NULL); | |
1589 | } | |
1590 | else | |
1591 | return vect_get_vec_def_for_operand_1 (def_stmt, dt); | |
1592 | } | |
1593 | ||
1594 | ||
ebfd146a IR |
1595 | /* Function vect_get_vec_def_for_stmt_copy |
1596 | ||
ff802fa1 | 1597 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1598 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1599 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1600 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1601 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1602 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1603 | DT is the type of the vector def VEC_OPRND. |
1604 | ||
1605 | Context: | |
1606 | In case the vectorization factor (VF) is bigger than the number | |
1607 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1608 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1609 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1610 | smallest data-type determines the VF, and as a result, when vectorizing |
1611 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1612 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1613 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1614 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1615 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1616 | ||
1617 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1618 | |
ebfd146a IR |
1619 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1620 | VS1.1: vx.1 = memref1 VS1.2 | |
1621 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1622 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1623 | |
1624 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1625 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1626 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1627 | VSnew.3: vz3 = vx.3 + ... | |
1628 | ||
1629 | The vectorization of S1 is explained in vectorizable_load. | |
1630 | The vectorization of S2: | |
b8698a0f L |
1631 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1632 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1633 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1634 | returns the vector-def 'vx.0'. |
1635 | ||
b8698a0f L |
1636 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1637 | function is called to get the relevant vector-def for each operand. It is | |
1638 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1639 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1640 | ||
b8698a0f L |
1641 | For example, to obtain the vector-def 'vx.1' in order to create the |
1642 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1643 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1644 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1645 | and return its def ('vx.1'). | |
1646 | Overall, to create the above sequence this function will be called 3 times: | |
1647 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1648 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1649 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1650 | ||
1651 | tree | |
1652 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1653 | { | |
355fe088 | 1654 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1655 | stmt_vec_info def_stmt_info; |
1656 | ||
1657 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1658 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1659 | return vec_oprnd; |
1660 | ||
1661 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1662 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1663 | gcc_assert (def_stmt_info); | |
1664 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1665 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1666 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1667 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1668 | else | |
1669 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1670 | return vec_oprnd; | |
1671 | } | |
1672 | ||
1673 | ||
1674 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1675 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a | 1676 | |
c78e3652 | 1677 | void |
b8698a0f | 1678 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1679 | vec<tree> *vec_oprnds0, |
1680 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1681 | { |
9771b263 | 1682 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1683 | |
1684 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1685 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1686 | |
9771b263 | 1687 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1688 | { |
9771b263 | 1689 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1690 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1691 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1692 | } |
1693 | } | |
1694 | ||
1695 | ||
c78e3652 | 1696 | /* Get vectorized definitions for OP0 and OP1. */ |
ebfd146a | 1697 | |
c78e3652 | 1698 | void |
355fe088 | 1699 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1700 | vec<tree> *vec_oprnds0, |
1701 | vec<tree> *vec_oprnds1, | |
306b0c92 | 1702 | slp_tree slp_node) |
ebfd146a IR |
1703 | { |
1704 | if (slp_node) | |
d092494c IR |
1705 | { |
1706 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1707 | auto_vec<tree> ops (nops); |
1708 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1709 | |
9771b263 | 1710 | ops.quick_push (op0); |
d092494c | 1711 | if (op1) |
9771b263 | 1712 | ops.quick_push (op1); |
d092494c | 1713 | |
306b0c92 | 1714 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
d092494c | 1715 | |
37b5ec8f | 1716 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1717 | if (op1) |
37b5ec8f | 1718 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1719 | } |
ebfd146a IR |
1720 | else |
1721 | { | |
1722 | tree vec_oprnd; | |
1723 | ||
9771b263 | 1724 | vec_oprnds0->create (1); |
81c40241 | 1725 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1726 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1727 | |
1728 | if (op1) | |
1729 | { | |
9771b263 | 1730 | vec_oprnds1->create (1); |
81c40241 | 1731 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1732 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1733 | } |
1734 | } | |
1735 | } | |
1736 | ||
bb6c2b68 RS |
1737 | /* Helper function called by vect_finish_replace_stmt and |
1738 | vect_finish_stmt_generation. Set the location of the new | |
1739 | statement and create a stmt_vec_info for it. */ | |
1740 | ||
1741 | static void | |
1742 | vect_finish_stmt_generation_1 (gimple *stmt, gimple *vec_stmt) | |
1743 | { | |
1744 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1745 | vec_info *vinfo = stmt_info->vinfo; | |
1746 | ||
4fbeb363 | 1747 | vinfo->add_stmt (vec_stmt); |
bb6c2b68 RS |
1748 | |
1749 | if (dump_enabled_p ()) | |
1750 | { | |
1751 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); | |
1752 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
1753 | } | |
1754 | ||
1755 | gimple_set_location (vec_stmt, gimple_location (stmt)); | |
1756 | ||
1757 | /* While EH edges will generally prevent vectorization, stmt might | |
1758 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1759 | that could throw are part of the same region. */ | |
1760 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1761 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1762 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
1763 | } | |
1764 | ||
1765 | /* Replace the scalar statement STMT with a new vector statement VEC_STMT, | |
1766 | which sets the same scalar result as STMT did. */ | |
1767 | ||
1768 | void | |
1769 | vect_finish_replace_stmt (gimple *stmt, gimple *vec_stmt) | |
1770 | { | |
1771 | gcc_assert (gimple_get_lhs (stmt) == gimple_get_lhs (vec_stmt)); | |
1772 | ||
1773 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); | |
1774 | gsi_replace (&gsi, vec_stmt, false); | |
1775 | ||
1776 | vect_finish_stmt_generation_1 (stmt, vec_stmt); | |
1777 | } | |
ebfd146a IR |
1778 | |
1779 | /* Function vect_finish_stmt_generation. | |
1780 | ||
1781 | Insert a new stmt. */ | |
1782 | ||
1783 | void | |
355fe088 | 1784 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1785 | gimple_stmt_iterator *gsi) |
1786 | { | |
ebfd146a IR |
1787 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); |
1788 | ||
54e8e2c3 RG |
1789 | if (!gsi_end_p (*gsi) |
1790 | && gimple_has_mem_ops (vec_stmt)) | |
1791 | { | |
355fe088 | 1792 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1793 | tree vuse = gimple_vuse (at_stmt); |
1794 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1795 | { | |
1796 | tree vdef = gimple_vdef (at_stmt); | |
1797 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1798 | /* If we have an SSA vuse and insert a store, update virtual | |
1799 | SSA form to avoid triggering the renamer. Do so only | |
1800 | if we can easily see all uses - which is what almost always | |
1801 | happens with the way vectorized stmts are inserted. */ | |
1802 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1803 | && ((is_gimple_assign (vec_stmt) | |
1804 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1805 | || (is_gimple_call (vec_stmt) | |
1806 | && !(gimple_call_flags (vec_stmt) | |
1807 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1808 | { | |
1809 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1810 | gimple_set_vdef (vec_stmt, new_vdef); | |
1811 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1812 | } | |
1813 | } | |
1814 | } | |
ebfd146a | 1815 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
bb6c2b68 | 1816 | vect_finish_stmt_generation_1 (stmt, vec_stmt); |
ebfd146a IR |
1817 | } |
1818 | ||
70439f0d RS |
1819 | /* We want to vectorize a call to combined function CFN with function |
1820 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN | |
1821 | as the types of all inputs. Check whether this is possible using | |
1822 | an internal function, returning its code if so or IFN_LAST if not. */ | |
ebfd146a | 1823 | |
70439f0d RS |
1824 | static internal_fn |
1825 | vectorizable_internal_function (combined_fn cfn, tree fndecl, | |
1826 | tree vectype_out, tree vectype_in) | |
ebfd146a | 1827 | { |
70439f0d RS |
1828 | internal_fn ifn; |
1829 | if (internal_fn_p (cfn)) | |
1830 | ifn = as_internal_fn (cfn); | |
1831 | else | |
1832 | ifn = associated_internal_fn (fndecl); | |
1833 | if (ifn != IFN_LAST && direct_internal_fn_p (ifn)) | |
1834 | { | |
1835 | const direct_internal_fn_info &info = direct_internal_fn (ifn); | |
1836 | if (info.vectorizable) | |
1837 | { | |
1838 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); | |
1839 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); | |
d95ab70a RS |
1840 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1841 | OPTIMIZE_FOR_SPEED)) | |
70439f0d RS |
1842 | return ifn; |
1843 | } | |
1844 | } | |
1845 | return IFN_LAST; | |
ebfd146a IR |
1846 | } |
1847 | ||
5ce9450f | 1848 | |
355fe088 | 1849 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1850 | gimple_stmt_iterator *); |
1851 | ||
7cfb4d93 RS |
1852 | /* Check whether a load or store statement in the loop described by |
1853 | LOOP_VINFO is possible in a fully-masked loop. This is testing | |
1854 | whether the vectorizer pass has the appropriate support, as well as | |
1855 | whether the target does. | |
1856 | ||
1857 | VLS_TYPE says whether the statement is a load or store and VECTYPE | |
1858 | is the type of the vector being loaded or stored. MEMORY_ACCESS_TYPE | |
1859 | says how the load or store is going to be implemented and GROUP_SIZE | |
1860 | is the number of load or store statements in the containing group. | |
bfaa08b7 RS |
1861 | If the access is a gather load or scatter store, GS_INFO describes |
1862 | its arguments. | |
7cfb4d93 RS |
1863 | |
1864 | Clear LOOP_VINFO_CAN_FULLY_MASK_P if a fully-masked loop is not | |
1865 | supported, otherwise record the required mask types. */ | |
1866 | ||
1867 | static void | |
1868 | check_load_store_masking (loop_vec_info loop_vinfo, tree vectype, | |
1869 | vec_load_store_type vls_type, int group_size, | |
bfaa08b7 RS |
1870 | vect_memory_access_type memory_access_type, |
1871 | gather_scatter_info *gs_info) | |
7cfb4d93 RS |
1872 | { |
1873 | /* Invariant loads need no special support. */ | |
1874 | if (memory_access_type == VMAT_INVARIANT) | |
1875 | return; | |
1876 | ||
1877 | vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); | |
1878 | machine_mode vecmode = TYPE_MODE (vectype); | |
1879 | bool is_load = (vls_type == VLS_LOAD); | |
1880 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
1881 | { | |
1882 | if (is_load | |
1883 | ? !vect_load_lanes_supported (vectype, group_size, true) | |
1884 | : !vect_store_lanes_supported (vectype, group_size, true)) | |
1885 | { | |
1886 | if (dump_enabled_p ()) | |
1887 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1888 | "can't use a fully-masked loop because the" | |
1889 | " target doesn't have an appropriate masked" | |
1890 | " load/store-lanes instruction.\n"); | |
1891 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1892 | return; | |
1893 | } | |
1894 | unsigned int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
1895 | vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype); | |
1896 | return; | |
1897 | } | |
1898 | ||
bfaa08b7 RS |
1899 | if (memory_access_type == VMAT_GATHER_SCATTER) |
1900 | { | |
f307441a RS |
1901 | internal_fn ifn = (is_load |
1902 | ? IFN_MASK_GATHER_LOAD | |
1903 | : IFN_MASK_SCATTER_STORE); | |
bfaa08b7 | 1904 | tree offset_type = TREE_TYPE (gs_info->offset); |
f307441a | 1905 | if (!internal_gather_scatter_fn_supported_p (ifn, vectype, |
bfaa08b7 RS |
1906 | gs_info->memory_type, |
1907 | TYPE_SIGN (offset_type), | |
1908 | gs_info->scale)) | |
1909 | { | |
1910 | if (dump_enabled_p ()) | |
1911 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1912 | "can't use a fully-masked loop because the" | |
1913 | " target doesn't have an appropriate masked" | |
f307441a | 1914 | " gather load or scatter store instruction.\n"); |
bfaa08b7 RS |
1915 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; |
1916 | return; | |
1917 | } | |
1918 | unsigned int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
1919 | vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype); | |
1920 | return; | |
1921 | } | |
1922 | ||
7cfb4d93 RS |
1923 | if (memory_access_type != VMAT_CONTIGUOUS |
1924 | && memory_access_type != VMAT_CONTIGUOUS_PERMUTE) | |
1925 | { | |
1926 | /* Element X of the data must come from iteration i * VF + X of the | |
1927 | scalar loop. We need more work to support other mappings. */ | |
1928 | if (dump_enabled_p ()) | |
1929 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1930 | "can't use a fully-masked loop because an access" | |
1931 | " isn't contiguous.\n"); | |
1932 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1933 | return; | |
1934 | } | |
1935 | ||
1936 | machine_mode mask_mode; | |
1937 | if (!(targetm.vectorize.get_mask_mode | |
1938 | (GET_MODE_NUNITS (vecmode), | |
1939 | GET_MODE_SIZE (vecmode)).exists (&mask_mode)) | |
1940 | || !can_vec_mask_load_store_p (vecmode, mask_mode, is_load)) | |
1941 | { | |
1942 | if (dump_enabled_p ()) | |
1943 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1944 | "can't use a fully-masked loop because the target" | |
1945 | " doesn't have the appropriate masked load or" | |
1946 | " store.\n"); | |
1947 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1948 | return; | |
1949 | } | |
1950 | /* We might load more scalars than we need for permuting SLP loads. | |
1951 | We checked in get_group_load_store_type that the extra elements | |
1952 | don't leak into a new vector. */ | |
1953 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
1954 | poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
1955 | unsigned int nvectors; | |
1956 | if (can_div_away_from_zero_p (group_size * vf, nunits, &nvectors)) | |
1957 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype); | |
1958 | else | |
1959 | gcc_unreachable (); | |
1960 | } | |
1961 | ||
1962 | /* Return the mask input to a masked load or store. VEC_MASK is the vectorized | |
1963 | form of the scalar mask condition and LOOP_MASK, if nonnull, is the mask | |
1964 | that needs to be applied to all loads and stores in a vectorized loop. | |
1965 | Return VEC_MASK if LOOP_MASK is null, otherwise return VEC_MASK & LOOP_MASK. | |
1966 | ||
1967 | MASK_TYPE is the type of both masks. If new statements are needed, | |
1968 | insert them before GSI. */ | |
1969 | ||
1970 | static tree | |
1971 | prepare_load_store_mask (tree mask_type, tree loop_mask, tree vec_mask, | |
1972 | gimple_stmt_iterator *gsi) | |
1973 | { | |
1974 | gcc_assert (useless_type_conversion_p (mask_type, TREE_TYPE (vec_mask))); | |
1975 | if (!loop_mask) | |
1976 | return vec_mask; | |
1977 | ||
1978 | gcc_assert (TREE_TYPE (loop_mask) == mask_type); | |
1979 | tree and_res = make_temp_ssa_name (mask_type, NULL, "vec_mask_and"); | |
1980 | gimple *and_stmt = gimple_build_assign (and_res, BIT_AND_EXPR, | |
1981 | vec_mask, loop_mask); | |
1982 | gsi_insert_before (gsi, and_stmt, GSI_SAME_STMT); | |
1983 | return and_res; | |
1984 | } | |
1985 | ||
429ef523 RS |
1986 | /* Determine whether we can use a gather load or scatter store to vectorize |
1987 | strided load or store STMT by truncating the current offset to a smaller | |
1988 | width. We need to be able to construct an offset vector: | |
1989 | ||
1990 | { 0, X, X*2, X*3, ... } | |
1991 | ||
1992 | without loss of precision, where X is STMT's DR_STEP. | |
1993 | ||
1994 | Return true if this is possible, describing the gather load or scatter | |
1995 | store in GS_INFO. MASKED_P is true if the load or store is conditional. */ | |
1996 | ||
1997 | static bool | |
1998 | vect_truncate_gather_scatter_offset (gimple *stmt, loop_vec_info loop_vinfo, | |
1999 | bool masked_p, | |
2000 | gather_scatter_info *gs_info) | |
2001 | { | |
2002 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2003 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2004 | tree step = DR_STEP (dr); | |
2005 | if (TREE_CODE (step) != INTEGER_CST) | |
2006 | { | |
2007 | /* ??? Perhaps we could use range information here? */ | |
2008 | if (dump_enabled_p ()) | |
2009 | dump_printf_loc (MSG_NOTE, vect_location, | |
2010 | "cannot truncate variable step.\n"); | |
2011 | return false; | |
2012 | } | |
2013 | ||
2014 | /* Get the number of bits in an element. */ | |
2015 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2016 | scalar_mode element_mode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); | |
2017 | unsigned int element_bits = GET_MODE_BITSIZE (element_mode); | |
2018 | ||
2019 | /* Set COUNT to the upper limit on the number of elements - 1. | |
2020 | Start with the maximum vectorization factor. */ | |
2021 | unsigned HOST_WIDE_INT count = vect_max_vf (loop_vinfo) - 1; | |
2022 | ||
2023 | /* Try lowering COUNT to the number of scalar latch iterations. */ | |
2024 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2025 | widest_int max_iters; | |
2026 | if (max_loop_iterations (loop, &max_iters) | |
2027 | && max_iters < count) | |
2028 | count = max_iters.to_shwi (); | |
2029 | ||
2030 | /* Try scales of 1 and the element size. */ | |
2031 | int scales[] = { 1, vect_get_scalar_dr_size (dr) }; | |
4a669ac3 | 2032 | wi::overflow_type overflow = wi::OVF_NONE; |
429ef523 RS |
2033 | for (int i = 0; i < 2; ++i) |
2034 | { | |
2035 | int scale = scales[i]; | |
2036 | widest_int factor; | |
2037 | if (!wi::multiple_of_p (wi::to_widest (step), scale, SIGNED, &factor)) | |
2038 | continue; | |
2039 | ||
2040 | /* See whether we can calculate (COUNT - 1) * STEP / SCALE | |
2041 | in OFFSET_BITS bits. */ | |
4a669ac3 AH |
2042 | widest_int range = wi::mul (count, factor, SIGNED, &overflow); |
2043 | if (overflow) | |
429ef523 RS |
2044 | continue; |
2045 | signop sign = range >= 0 ? UNSIGNED : SIGNED; | |
2046 | if (wi::min_precision (range, sign) > element_bits) | |
2047 | { | |
4a669ac3 | 2048 | overflow = wi::OVF_UNKNOWN; |
429ef523 RS |
2049 | continue; |
2050 | } | |
2051 | ||
2052 | /* See whether the target supports the operation. */ | |
2053 | tree memory_type = TREE_TYPE (DR_REF (dr)); | |
2054 | if (!vect_gather_scatter_fn_p (DR_IS_READ (dr), masked_p, vectype, | |
2055 | memory_type, element_bits, sign, scale, | |
2056 | &gs_info->ifn, &gs_info->element_type)) | |
2057 | continue; | |
2058 | ||
2059 | tree offset_type = build_nonstandard_integer_type (element_bits, | |
2060 | sign == UNSIGNED); | |
2061 | ||
2062 | gs_info->decl = NULL_TREE; | |
2063 | /* Logically the sum of DR_BASE_ADDRESS, DR_INIT and DR_OFFSET, | |
2064 | but we don't need to store that here. */ | |
2065 | gs_info->base = NULL_TREE; | |
2066 | gs_info->offset = fold_convert (offset_type, step); | |
929b4411 | 2067 | gs_info->offset_dt = vect_constant_def; |
429ef523 RS |
2068 | gs_info->offset_vectype = NULL_TREE; |
2069 | gs_info->scale = scale; | |
2070 | gs_info->memory_type = memory_type; | |
2071 | return true; | |
2072 | } | |
2073 | ||
4a669ac3 | 2074 | if (overflow && dump_enabled_p ()) |
429ef523 RS |
2075 | dump_printf_loc (MSG_NOTE, vect_location, |
2076 | "truncating gather/scatter offset to %d bits" | |
2077 | " might change its value.\n", element_bits); | |
2078 | ||
2079 | return false; | |
2080 | } | |
2081 | ||
ab2fc782 RS |
2082 | /* Return true if we can use gather/scatter internal functions to |
2083 | vectorize STMT, which is a grouped or strided load or store. | |
429ef523 RS |
2084 | MASKED_P is true if load or store is conditional. When returning |
2085 | true, fill in GS_INFO with the information required to perform the | |
2086 | operation. */ | |
ab2fc782 RS |
2087 | |
2088 | static bool | |
2089 | vect_use_strided_gather_scatters_p (gimple *stmt, loop_vec_info loop_vinfo, | |
429ef523 | 2090 | bool masked_p, |
ab2fc782 RS |
2091 | gather_scatter_info *gs_info) |
2092 | { | |
2093 | if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info) | |
2094 | || gs_info->decl) | |
429ef523 RS |
2095 | return vect_truncate_gather_scatter_offset (stmt, loop_vinfo, |
2096 | masked_p, gs_info); | |
ab2fc782 RS |
2097 | |
2098 | scalar_mode element_mode = SCALAR_TYPE_MODE (gs_info->element_type); | |
2099 | unsigned int element_bits = GET_MODE_BITSIZE (element_mode); | |
2100 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2101 | unsigned int offset_bits = TYPE_PRECISION (offset_type); | |
2102 | ||
2103 | /* Enforced by vect_check_gather_scatter. */ | |
2104 | gcc_assert (element_bits >= offset_bits); | |
2105 | ||
2106 | /* If the elements are wider than the offset, convert the offset to the | |
2107 | same width, without changing its sign. */ | |
2108 | if (element_bits > offset_bits) | |
2109 | { | |
2110 | bool unsigned_p = TYPE_UNSIGNED (offset_type); | |
2111 | offset_type = build_nonstandard_integer_type (element_bits, unsigned_p); | |
2112 | gs_info->offset = fold_convert (offset_type, gs_info->offset); | |
2113 | } | |
2114 | ||
2115 | if (dump_enabled_p ()) | |
2116 | dump_printf_loc (MSG_NOTE, vect_location, | |
2117 | "using gather/scatter for strided/grouped access," | |
2118 | " scale = %d\n", gs_info->scale); | |
2119 | ||
2120 | return true; | |
2121 | } | |
2122 | ||
62da9e14 RS |
2123 | /* STMT is a non-strided load or store, meaning that it accesses |
2124 | elements with a known constant step. Return -1 if that step | |
2125 | is negative, 0 if it is zero, and 1 if it is greater than zero. */ | |
2126 | ||
2127 | static int | |
2128 | compare_step_with_zero (gimple *stmt) | |
2129 | { | |
2130 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3f5e8a76 RS |
2131 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
2132 | return tree_int_cst_compare (vect_dr_behavior (dr)->step, | |
2133 | size_zero_node); | |
62da9e14 RS |
2134 | } |
2135 | ||
2136 | /* If the target supports a permute mask that reverses the elements in | |
2137 | a vector of type VECTYPE, return that mask, otherwise return null. */ | |
2138 | ||
2139 | static tree | |
2140 | perm_mask_for_reverse (tree vectype) | |
2141 | { | |
928686b1 | 2142 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
62da9e14 | 2143 | |
d980067b RS |
2144 | /* The encoding has a single stepped pattern. */ |
2145 | vec_perm_builder sel (nunits, 1, 3); | |
928686b1 | 2146 | for (int i = 0; i < 3; ++i) |
908a1a16 | 2147 | sel.quick_push (nunits - 1 - i); |
62da9e14 | 2148 | |
e3342de4 RS |
2149 | vec_perm_indices indices (sel, 1, nunits); |
2150 | if (!can_vec_perm_const_p (TYPE_MODE (vectype), indices)) | |
62da9e14 | 2151 | return NULL_TREE; |
e3342de4 | 2152 | return vect_gen_perm_mask_checked (vectype, indices); |
62da9e14 | 2153 | } |
5ce9450f | 2154 | |
c3a8f964 RS |
2155 | /* STMT is either a masked or unconditional store. Return the value |
2156 | being stored. */ | |
2157 | ||
f307441a | 2158 | tree |
c3a8f964 RS |
2159 | vect_get_store_rhs (gimple *stmt) |
2160 | { | |
2161 | if (gassign *assign = dyn_cast <gassign *> (stmt)) | |
2162 | { | |
2163 | gcc_assert (gimple_assign_single_p (assign)); | |
2164 | return gimple_assign_rhs1 (assign); | |
2165 | } | |
2166 | if (gcall *call = dyn_cast <gcall *> (stmt)) | |
2167 | { | |
2168 | internal_fn ifn = gimple_call_internal_fn (call); | |
f307441a RS |
2169 | int index = internal_fn_stored_value_index (ifn); |
2170 | gcc_assert (index >= 0); | |
2171 | return gimple_call_arg (stmt, index); | |
c3a8f964 RS |
2172 | } |
2173 | gcc_unreachable (); | |
2174 | } | |
2175 | ||
2de001ee RS |
2176 | /* A subroutine of get_load_store_type, with a subset of the same |
2177 | arguments. Handle the case where STMT is part of a grouped load | |
2178 | or store. | |
2179 | ||
2180 | For stores, the statements in the group are all consecutive | |
2181 | and there is no gap at the end. For loads, the statements in the | |
2182 | group might not be consecutive; there can be gaps between statements | |
2183 | as well as at the end. */ | |
2184 | ||
2185 | static bool | |
2186 | get_group_load_store_type (gimple *stmt, tree vectype, bool slp, | |
7e11fc7f | 2187 | bool masked_p, vec_load_store_type vls_type, |
429ef523 RS |
2188 | vect_memory_access_type *memory_access_type, |
2189 | gather_scatter_info *gs_info) | |
2de001ee RS |
2190 | { |
2191 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2192 | vec_info *vinfo = stmt_info->vinfo; | |
2193 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2194 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; | |
2c53b149 | 2195 | gimple *first_stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
f702e7d4 | 2196 | data_reference *first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
2c53b149 | 2197 | unsigned int group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
2de001ee | 2198 | bool single_element_p = (stmt == first_stmt |
2c53b149 RB |
2199 | && !DR_GROUP_NEXT_ELEMENT (stmt_info)); |
2200 | unsigned HOST_WIDE_INT gap = DR_GROUP_GAP (vinfo_for_stmt (first_stmt)); | |
928686b1 | 2201 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
2202 | |
2203 | /* True if the vectorized statements would access beyond the last | |
2204 | statement in the group. */ | |
2205 | bool overrun_p = false; | |
2206 | ||
2207 | /* True if we can cope with such overrun by peeling for gaps, so that | |
2208 | there is at least one final scalar iteration after the vector loop. */ | |
7e11fc7f RS |
2209 | bool can_overrun_p = (!masked_p |
2210 | && vls_type == VLS_LOAD | |
2211 | && loop_vinfo | |
2212 | && !loop->inner); | |
2de001ee RS |
2213 | |
2214 | /* There can only be a gap at the end of the group if the stride is | |
2215 | known at compile time. */ | |
2216 | gcc_assert (!STMT_VINFO_STRIDED_P (stmt_info) || gap == 0); | |
2217 | ||
2218 | /* Stores can't yet have gaps. */ | |
2219 | gcc_assert (slp || vls_type == VLS_LOAD || gap == 0); | |
2220 | ||
2221 | if (slp) | |
2222 | { | |
2223 | if (STMT_VINFO_STRIDED_P (stmt_info)) | |
2224 | { | |
2c53b149 | 2225 | /* Try to use consecutive accesses of DR_GROUP_SIZE elements, |
2de001ee RS |
2226 | separated by the stride, until we have a complete vector. |
2227 | Fall back to scalar accesses if that isn't possible. */ | |
928686b1 | 2228 | if (multiple_p (nunits, group_size)) |
2de001ee RS |
2229 | *memory_access_type = VMAT_STRIDED_SLP; |
2230 | else | |
2231 | *memory_access_type = VMAT_ELEMENTWISE; | |
2232 | } | |
2233 | else | |
2234 | { | |
2235 | overrun_p = loop_vinfo && gap != 0; | |
2236 | if (overrun_p && vls_type != VLS_LOAD) | |
2237 | { | |
2238 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2239 | "Grouped store with gaps requires" | |
2240 | " non-consecutive accesses\n"); | |
2241 | return false; | |
2242 | } | |
f702e7d4 RS |
2243 | /* An overrun is fine if the trailing elements are smaller |
2244 | than the alignment boundary B. Every vector access will | |
2245 | be a multiple of B and so we are guaranteed to access a | |
2246 | non-gap element in the same B-sized block. */ | |
f9ef2c76 | 2247 | if (overrun_p |
f702e7d4 RS |
2248 | && gap < (vect_known_alignment_in_bytes (first_dr) |
2249 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 2250 | overrun_p = false; |
2de001ee RS |
2251 | if (overrun_p && !can_overrun_p) |
2252 | { | |
2253 | if (dump_enabled_p ()) | |
2254 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2255 | "Peeling for outer loop is not supported\n"); | |
2256 | return false; | |
2257 | } | |
2258 | *memory_access_type = VMAT_CONTIGUOUS; | |
2259 | } | |
2260 | } | |
2261 | else | |
2262 | { | |
2263 | /* We can always handle this case using elementwise accesses, | |
2264 | but see if something more efficient is available. */ | |
2265 | *memory_access_type = VMAT_ELEMENTWISE; | |
2266 | ||
2267 | /* If there is a gap at the end of the group then these optimizations | |
2268 | would access excess elements in the last iteration. */ | |
2269 | bool would_overrun_p = (gap != 0); | |
f702e7d4 RS |
2270 | /* An overrun is fine if the trailing elements are smaller than the |
2271 | alignment boundary B. Every vector access will be a multiple of B | |
2272 | and so we are guaranteed to access a non-gap element in the | |
2273 | same B-sized block. */ | |
f9ef2c76 | 2274 | if (would_overrun_p |
7e11fc7f | 2275 | && !masked_p |
f702e7d4 RS |
2276 | && gap < (vect_known_alignment_in_bytes (first_dr) |
2277 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 2278 | would_overrun_p = false; |
f702e7d4 | 2279 | |
2de001ee | 2280 | if (!STMT_VINFO_STRIDED_P (stmt_info) |
62da9e14 RS |
2281 | && (can_overrun_p || !would_overrun_p) |
2282 | && compare_step_with_zero (stmt) > 0) | |
2de001ee | 2283 | { |
6737facb RS |
2284 | /* First cope with the degenerate case of a single-element |
2285 | vector. */ | |
2286 | if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U)) | |
2287 | *memory_access_type = VMAT_CONTIGUOUS; | |
2288 | ||
2289 | /* Otherwise try using LOAD/STORE_LANES. */ | |
2290 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2291 | && (vls_type == VLS_LOAD | |
7e11fc7f RS |
2292 | ? vect_load_lanes_supported (vectype, group_size, masked_p) |
2293 | : vect_store_lanes_supported (vectype, group_size, | |
2294 | masked_p))) | |
2de001ee RS |
2295 | { |
2296 | *memory_access_type = VMAT_LOAD_STORE_LANES; | |
2297 | overrun_p = would_overrun_p; | |
2298 | } | |
2299 | ||
2300 | /* If that fails, try using permuting loads. */ | |
2301 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2302 | && (vls_type == VLS_LOAD | |
2303 | ? vect_grouped_load_supported (vectype, single_element_p, | |
2304 | group_size) | |
2305 | : vect_grouped_store_supported (vectype, group_size))) | |
2306 | { | |
2307 | *memory_access_type = VMAT_CONTIGUOUS_PERMUTE; | |
2308 | overrun_p = would_overrun_p; | |
2309 | } | |
2310 | } | |
429ef523 RS |
2311 | |
2312 | /* As a last resort, trying using a gather load or scatter store. | |
2313 | ||
2314 | ??? Although the code can handle all group sizes correctly, | |
2315 | it probably isn't a win to use separate strided accesses based | |
2316 | on nearby locations. Or, even if it's a win over scalar code, | |
2317 | it might not be a win over vectorizing at a lower VF, if that | |
2318 | allows us to use contiguous accesses. */ | |
2319 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2320 | && single_element_p | |
2321 | && loop_vinfo | |
2322 | && vect_use_strided_gather_scatters_p (stmt, loop_vinfo, | |
2323 | masked_p, gs_info)) | |
2324 | *memory_access_type = VMAT_GATHER_SCATTER; | |
2de001ee RS |
2325 | } |
2326 | ||
2327 | if (vls_type != VLS_LOAD && first_stmt == stmt) | |
2328 | { | |
2329 | /* STMT is the leader of the group. Check the operands of all the | |
2330 | stmts of the group. */ | |
2c53b149 | 2331 | gimple *next_stmt = DR_GROUP_NEXT_ELEMENT (stmt_info); |
2de001ee RS |
2332 | while (next_stmt) |
2333 | { | |
7e11fc7f | 2334 | tree op = vect_get_store_rhs (next_stmt); |
2de001ee | 2335 | enum vect_def_type dt; |
894dd753 | 2336 | if (!vect_is_simple_use (op, vinfo, &dt)) |
2de001ee RS |
2337 | { |
2338 | if (dump_enabled_p ()) | |
2339 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2340 | "use not simple.\n"); | |
2341 | return false; | |
2342 | } | |
2c53b149 | 2343 | next_stmt = DR_GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
2de001ee RS |
2344 | } |
2345 | } | |
2346 | ||
2347 | if (overrun_p) | |
2348 | { | |
2349 | gcc_assert (can_overrun_p); | |
2350 | if (dump_enabled_p ()) | |
2351 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2352 | "Data access with gaps requires scalar " | |
2353 | "epilogue loop\n"); | |
2354 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; | |
2355 | } | |
2356 | ||
2357 | return true; | |
2358 | } | |
2359 | ||
62da9e14 RS |
2360 | /* A subroutine of get_load_store_type, with a subset of the same |
2361 | arguments. Handle the case where STMT is a load or store that | |
2362 | accesses consecutive elements with a negative step. */ | |
2363 | ||
2364 | static vect_memory_access_type | |
2365 | get_negative_load_store_type (gimple *stmt, tree vectype, | |
2366 | vec_load_store_type vls_type, | |
2367 | unsigned int ncopies) | |
2368 | { | |
2369 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2370 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2371 | dr_alignment_support alignment_support_scheme; | |
2372 | ||
2373 | if (ncopies > 1) | |
2374 | { | |
2375 | if (dump_enabled_p ()) | |
2376 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2377 | "multiple types with negative step.\n"); | |
2378 | return VMAT_ELEMENTWISE; | |
2379 | } | |
2380 | ||
2381 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
2382 | if (alignment_support_scheme != dr_aligned | |
2383 | && alignment_support_scheme != dr_unaligned_supported) | |
2384 | { | |
2385 | if (dump_enabled_p ()) | |
2386 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2387 | "negative step but alignment required.\n"); | |
2388 | return VMAT_ELEMENTWISE; | |
2389 | } | |
2390 | ||
2391 | if (vls_type == VLS_STORE_INVARIANT) | |
2392 | { | |
2393 | if (dump_enabled_p ()) | |
2394 | dump_printf_loc (MSG_NOTE, vect_location, | |
2395 | "negative step with invariant source;" | |
2396 | " no permute needed.\n"); | |
2397 | return VMAT_CONTIGUOUS_DOWN; | |
2398 | } | |
2399 | ||
2400 | if (!perm_mask_for_reverse (vectype)) | |
2401 | { | |
2402 | if (dump_enabled_p ()) | |
2403 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2404 | "negative step and reversing not supported.\n"); | |
2405 | return VMAT_ELEMENTWISE; | |
2406 | } | |
2407 | ||
2408 | return VMAT_CONTIGUOUS_REVERSE; | |
2409 | } | |
2410 | ||
2de001ee RS |
2411 | /* Analyze load or store statement STMT of type VLS_TYPE. Return true |
2412 | if there is a memory access type that the vectorized form can use, | |
2413 | storing it in *MEMORY_ACCESS_TYPE if so. If we decide to use gathers | |
2414 | or scatters, fill in GS_INFO accordingly. | |
2415 | ||
2416 | SLP says whether we're performing SLP rather than loop vectorization. | |
7e11fc7f | 2417 | MASKED_P is true if the statement is conditional on a vectorized mask. |
62da9e14 RS |
2418 | VECTYPE is the vector type that the vectorized statements will use. |
2419 | NCOPIES is the number of vector statements that will be needed. */ | |
2de001ee RS |
2420 | |
2421 | static bool | |
7e11fc7f | 2422 | get_load_store_type (gimple *stmt, tree vectype, bool slp, bool masked_p, |
62da9e14 | 2423 | vec_load_store_type vls_type, unsigned int ncopies, |
2de001ee RS |
2424 | vect_memory_access_type *memory_access_type, |
2425 | gather_scatter_info *gs_info) | |
2426 | { | |
2427 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2428 | vec_info *vinfo = stmt_info->vinfo; | |
2429 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4d694b27 | 2430 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
2431 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
2432 | { | |
2433 | *memory_access_type = VMAT_GATHER_SCATTER; | |
2de001ee RS |
2434 | if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info)) |
2435 | gcc_unreachable (); | |
894dd753 | 2436 | else if (!vect_is_simple_use (gs_info->offset, vinfo, |
2de001ee RS |
2437 | &gs_info->offset_dt, |
2438 | &gs_info->offset_vectype)) | |
2439 | { | |
2440 | if (dump_enabled_p ()) | |
2441 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2442 | "%s index use not simple.\n", | |
2443 | vls_type == VLS_LOAD ? "gather" : "scatter"); | |
2444 | return false; | |
2445 | } | |
2446 | } | |
2447 | else if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
2448 | { | |
7e11fc7f | 2449 | if (!get_group_load_store_type (stmt, vectype, slp, masked_p, vls_type, |
429ef523 | 2450 | memory_access_type, gs_info)) |
2de001ee RS |
2451 | return false; |
2452 | } | |
2453 | else if (STMT_VINFO_STRIDED_P (stmt_info)) | |
2454 | { | |
2455 | gcc_assert (!slp); | |
ab2fc782 | 2456 | if (loop_vinfo |
429ef523 RS |
2457 | && vect_use_strided_gather_scatters_p (stmt, loop_vinfo, |
2458 | masked_p, gs_info)) | |
ab2fc782 RS |
2459 | *memory_access_type = VMAT_GATHER_SCATTER; |
2460 | else | |
2461 | *memory_access_type = VMAT_ELEMENTWISE; | |
2de001ee RS |
2462 | } |
2463 | else | |
62da9e14 RS |
2464 | { |
2465 | int cmp = compare_step_with_zero (stmt); | |
2466 | if (cmp < 0) | |
2467 | *memory_access_type = get_negative_load_store_type | |
2468 | (stmt, vectype, vls_type, ncopies); | |
2469 | else if (cmp == 0) | |
2470 | { | |
2471 | gcc_assert (vls_type == VLS_LOAD); | |
2472 | *memory_access_type = VMAT_INVARIANT; | |
2473 | } | |
2474 | else | |
2475 | *memory_access_type = VMAT_CONTIGUOUS; | |
2476 | } | |
2de001ee | 2477 | |
4d694b27 RS |
2478 | if ((*memory_access_type == VMAT_ELEMENTWISE |
2479 | || *memory_access_type == VMAT_STRIDED_SLP) | |
2480 | && !nunits.is_constant ()) | |
2481 | { | |
2482 | if (dump_enabled_p ()) | |
2483 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2484 | "Not using elementwise accesses due to variable " | |
2485 | "vectorization factor.\n"); | |
2486 | return false; | |
2487 | } | |
2488 | ||
2de001ee RS |
2489 | /* FIXME: At the moment the cost model seems to underestimate the |
2490 | cost of using elementwise accesses. This check preserves the | |
2491 | traditional behavior until that can be fixed. */ | |
2492 | if (*memory_access_type == VMAT_ELEMENTWISE | |
4aa157e8 | 2493 | && !STMT_VINFO_STRIDED_P (stmt_info) |
2c53b149 RB |
2494 | && !(stmt == DR_GROUP_FIRST_ELEMENT (stmt_info) |
2495 | && !DR_GROUP_NEXT_ELEMENT (stmt_info) | |
2496 | && !pow2p_hwi (DR_GROUP_SIZE (stmt_info)))) | |
2de001ee RS |
2497 | { |
2498 | if (dump_enabled_p ()) | |
2499 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2500 | "not falling back to elementwise accesses\n"); | |
2501 | return false; | |
2502 | } | |
2503 | return true; | |
2504 | } | |
2505 | ||
aaeefd88 | 2506 | /* Return true if boolean argument MASK is suitable for vectorizing |
929b4411 RS |
2507 | conditional load or store STMT. When returning true, store the type |
2508 | of the definition in *MASK_DT_OUT and the type of the vectorized mask | |
2509 | in *MASK_VECTYPE_OUT. */ | |
aaeefd88 RS |
2510 | |
2511 | static bool | |
929b4411 RS |
2512 | vect_check_load_store_mask (gimple *stmt, tree mask, |
2513 | vect_def_type *mask_dt_out, | |
2514 | tree *mask_vectype_out) | |
aaeefd88 RS |
2515 | { |
2516 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (mask))) | |
2517 | { | |
2518 | if (dump_enabled_p ()) | |
2519 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2520 | "mask argument is not a boolean.\n"); | |
2521 | return false; | |
2522 | } | |
2523 | ||
2524 | if (TREE_CODE (mask) != SSA_NAME) | |
2525 | { | |
2526 | if (dump_enabled_p ()) | |
2527 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2528 | "mask argument is not an SSA name.\n"); | |
2529 | return false; | |
2530 | } | |
2531 | ||
2532 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
929b4411 | 2533 | enum vect_def_type mask_dt; |
aaeefd88 | 2534 | tree mask_vectype; |
894dd753 | 2535 | if (!vect_is_simple_use (mask, stmt_info->vinfo, &mask_dt, &mask_vectype)) |
aaeefd88 RS |
2536 | { |
2537 | if (dump_enabled_p ()) | |
2538 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2539 | "mask use not simple.\n"); | |
2540 | return false; | |
2541 | } | |
2542 | ||
2543 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2544 | if (!mask_vectype) | |
2545 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
2546 | ||
2547 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype)) | |
2548 | { | |
2549 | if (dump_enabled_p ()) | |
2550 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2551 | "could not find an appropriate vector mask type.\n"); | |
2552 | return false; | |
2553 | } | |
2554 | ||
2555 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (mask_vectype), | |
2556 | TYPE_VECTOR_SUBPARTS (vectype))) | |
2557 | { | |
2558 | if (dump_enabled_p ()) | |
2559 | { | |
2560 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2561 | "vector mask type "); | |
2562 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, mask_vectype); | |
2563 | dump_printf (MSG_MISSED_OPTIMIZATION, | |
2564 | " does not match vector data type "); | |
2565 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, vectype); | |
2566 | dump_printf (MSG_MISSED_OPTIMIZATION, ".\n"); | |
2567 | } | |
2568 | return false; | |
2569 | } | |
2570 | ||
929b4411 | 2571 | *mask_dt_out = mask_dt; |
aaeefd88 RS |
2572 | *mask_vectype_out = mask_vectype; |
2573 | return true; | |
2574 | } | |
2575 | ||
3133c3b6 RS |
2576 | /* Return true if stored value RHS is suitable for vectorizing store |
2577 | statement STMT. When returning true, store the type of the | |
929b4411 RS |
2578 | definition in *RHS_DT_OUT, the type of the vectorized store value in |
2579 | *RHS_VECTYPE_OUT and the type of the store in *VLS_TYPE_OUT. */ | |
3133c3b6 RS |
2580 | |
2581 | static bool | |
929b4411 RS |
2582 | vect_check_store_rhs (gimple *stmt, tree rhs, vect_def_type *rhs_dt_out, |
2583 | tree *rhs_vectype_out, vec_load_store_type *vls_type_out) | |
3133c3b6 RS |
2584 | { |
2585 | /* In the case this is a store from a constant make sure | |
2586 | native_encode_expr can handle it. */ | |
2587 | if (CONSTANT_CLASS_P (rhs) && native_encode_expr (rhs, NULL, 64) == 0) | |
2588 | { | |
2589 | if (dump_enabled_p ()) | |
2590 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2591 | "cannot encode constant as a byte sequence.\n"); | |
2592 | return false; | |
2593 | } | |
2594 | ||
2595 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
929b4411 | 2596 | enum vect_def_type rhs_dt; |
3133c3b6 | 2597 | tree rhs_vectype; |
894dd753 | 2598 | if (!vect_is_simple_use (rhs, stmt_info->vinfo, &rhs_dt, &rhs_vectype)) |
3133c3b6 RS |
2599 | { |
2600 | if (dump_enabled_p ()) | |
2601 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2602 | "use not simple.\n"); | |
2603 | return false; | |
2604 | } | |
2605 | ||
2606 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2607 | if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype)) | |
2608 | { | |
2609 | if (dump_enabled_p ()) | |
2610 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2611 | "incompatible vector types.\n"); | |
2612 | return false; | |
2613 | } | |
2614 | ||
929b4411 | 2615 | *rhs_dt_out = rhs_dt; |
3133c3b6 | 2616 | *rhs_vectype_out = rhs_vectype; |
929b4411 | 2617 | if (rhs_dt == vect_constant_def || rhs_dt == vect_external_def) |
3133c3b6 RS |
2618 | *vls_type_out = VLS_STORE_INVARIANT; |
2619 | else | |
2620 | *vls_type_out = VLS_STORE; | |
2621 | return true; | |
2622 | } | |
2623 | ||
bc9587eb RS |
2624 | /* Build an all-ones vector mask of type MASKTYPE while vectorizing STMT. |
2625 | Note that we support masks with floating-point type, in which case the | |
2626 | floats are interpreted as a bitmask. */ | |
2627 | ||
2628 | static tree | |
2629 | vect_build_all_ones_mask (gimple *stmt, tree masktype) | |
2630 | { | |
2631 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
2632 | return build_int_cst (masktype, -1); | |
2633 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
2634 | { | |
2635 | tree mask = build_int_cst (TREE_TYPE (masktype), -1); | |
2636 | mask = build_vector_from_val (masktype, mask); | |
2637 | return vect_init_vector (stmt, mask, masktype, NULL); | |
2638 | } | |
2639 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) | |
2640 | { | |
2641 | REAL_VALUE_TYPE r; | |
2642 | long tmp[6]; | |
2643 | for (int j = 0; j < 6; ++j) | |
2644 | tmp[j] = -1; | |
2645 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
2646 | tree mask = build_real (TREE_TYPE (masktype), r); | |
2647 | mask = build_vector_from_val (masktype, mask); | |
2648 | return vect_init_vector (stmt, mask, masktype, NULL); | |
2649 | } | |
2650 | gcc_unreachable (); | |
2651 | } | |
2652 | ||
2653 | /* Build an all-zero merge value of type VECTYPE while vectorizing | |
2654 | STMT as a gather load. */ | |
2655 | ||
2656 | static tree | |
2657 | vect_build_zero_merge_argument (gimple *stmt, tree vectype) | |
2658 | { | |
2659 | tree merge; | |
2660 | if (TREE_CODE (TREE_TYPE (vectype)) == INTEGER_TYPE) | |
2661 | merge = build_int_cst (TREE_TYPE (vectype), 0); | |
2662 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (vectype))) | |
2663 | { | |
2664 | REAL_VALUE_TYPE r; | |
2665 | long tmp[6]; | |
2666 | for (int j = 0; j < 6; ++j) | |
2667 | tmp[j] = 0; | |
2668 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (vectype))); | |
2669 | merge = build_real (TREE_TYPE (vectype), r); | |
2670 | } | |
2671 | else | |
2672 | gcc_unreachable (); | |
2673 | merge = build_vector_from_val (vectype, merge); | |
2674 | return vect_init_vector (stmt, merge, vectype, NULL); | |
2675 | } | |
2676 | ||
c48d2d35 RS |
2677 | /* Build a gather load call while vectorizing STMT. Insert new instructions |
2678 | before GSI and add them to VEC_STMT. GS_INFO describes the gather load | |
2679 | operation. If the load is conditional, MASK is the unvectorized | |
929b4411 | 2680 | condition and MASK_DT is its definition type, otherwise MASK is null. */ |
c48d2d35 RS |
2681 | |
2682 | static void | |
2683 | vect_build_gather_load_calls (gimple *stmt, gimple_stmt_iterator *gsi, | |
2684 | gimple **vec_stmt, gather_scatter_info *gs_info, | |
929b4411 | 2685 | tree mask, vect_def_type mask_dt) |
c48d2d35 RS |
2686 | { |
2687 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2688 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2689 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2690 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2691 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2692 | int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
2693 | edge pe = loop_preheader_edge (loop); | |
2694 | enum { NARROW, NONE, WIDEN } modifier; | |
2695 | poly_uint64 gather_off_nunits | |
2696 | = TYPE_VECTOR_SUBPARTS (gs_info->offset_vectype); | |
2697 | ||
2698 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info->decl)); | |
2699 | tree rettype = TREE_TYPE (TREE_TYPE (gs_info->decl)); | |
2700 | tree srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2701 | tree ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2702 | tree idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2703 | tree masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2704 | tree scaletype = TREE_VALUE (arglist); | |
2705 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
2706 | && (!mask || types_compatible_p (srctype, masktype))); | |
2707 | ||
2708 | tree perm_mask = NULL_TREE; | |
2709 | tree mask_perm_mask = NULL_TREE; | |
2710 | if (known_eq (nunits, gather_off_nunits)) | |
2711 | modifier = NONE; | |
2712 | else if (known_eq (nunits * 2, gather_off_nunits)) | |
2713 | { | |
2714 | modifier = WIDEN; | |
2715 | ||
2716 | /* Currently widening gathers and scatters are only supported for | |
2717 | fixed-length vectors. */ | |
2718 | int count = gather_off_nunits.to_constant (); | |
2719 | vec_perm_builder sel (count, count, 1); | |
2720 | for (int i = 0; i < count; ++i) | |
2721 | sel.quick_push (i | (count / 2)); | |
2722 | ||
2723 | vec_perm_indices indices (sel, 1, count); | |
2724 | perm_mask = vect_gen_perm_mask_checked (gs_info->offset_vectype, | |
2725 | indices); | |
2726 | } | |
2727 | else if (known_eq (nunits, gather_off_nunits * 2)) | |
2728 | { | |
2729 | modifier = NARROW; | |
2730 | ||
2731 | /* Currently narrowing gathers and scatters are only supported for | |
2732 | fixed-length vectors. */ | |
2733 | int count = nunits.to_constant (); | |
2734 | vec_perm_builder sel (count, count, 1); | |
2735 | sel.quick_grow (count); | |
2736 | for (int i = 0; i < count; ++i) | |
2737 | sel[i] = i < count / 2 ? i : i + count / 2; | |
2738 | vec_perm_indices indices (sel, 2, count); | |
2739 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); | |
2740 | ||
2741 | ncopies *= 2; | |
2742 | ||
2743 | if (mask) | |
2744 | { | |
2745 | for (int i = 0; i < count; ++i) | |
2746 | sel[i] = i | (count / 2); | |
2747 | indices.new_vector (sel, 2, count); | |
2748 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, indices); | |
2749 | } | |
2750 | } | |
2751 | else | |
2752 | gcc_unreachable (); | |
2753 | ||
2754 | tree vec_dest = vect_create_destination_var (gimple_get_lhs (stmt), | |
2755 | vectype); | |
2756 | ||
2757 | tree ptr = fold_convert (ptrtype, gs_info->base); | |
2758 | if (!is_gimple_min_invariant (ptr)) | |
2759 | { | |
2760 | gimple_seq seq; | |
2761 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
2762 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
2763 | gcc_assert (!new_bb); | |
2764 | } | |
2765 | ||
2766 | tree scale = build_int_cst (scaletype, gs_info->scale); | |
2767 | ||
2768 | tree vec_oprnd0 = NULL_TREE; | |
2769 | tree vec_mask = NULL_TREE; | |
2770 | tree src_op = NULL_TREE; | |
2771 | tree mask_op = NULL_TREE; | |
2772 | tree prev_res = NULL_TREE; | |
2773 | stmt_vec_info prev_stmt_info = NULL; | |
2774 | ||
2775 | if (!mask) | |
2776 | { | |
2777 | src_op = vect_build_zero_merge_argument (stmt, rettype); | |
2778 | mask_op = vect_build_all_ones_mask (stmt, masktype); | |
2779 | } | |
2780 | ||
2781 | for (int j = 0; j < ncopies; ++j) | |
2782 | { | |
2783 | tree op, var; | |
2784 | gimple *new_stmt; | |
2785 | if (modifier == WIDEN && (j & 1)) | |
2786 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
2787 | perm_mask, stmt, gsi); | |
2788 | else if (j == 0) | |
2789 | op = vec_oprnd0 | |
2790 | = vect_get_vec_def_for_operand (gs_info->offset, stmt); | |
2791 | else | |
2792 | op = vec_oprnd0 | |
2793 | = vect_get_vec_def_for_stmt_copy (gs_info->offset_dt, vec_oprnd0); | |
2794 | ||
2795 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
2796 | { | |
2797 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), | |
2798 | TYPE_VECTOR_SUBPARTS (idxtype))); | |
2799 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); | |
2800 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); | |
2801 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
2802 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2803 | op = var; | |
2804 | } | |
2805 | ||
2806 | if (mask) | |
2807 | { | |
2808 | if (mask_perm_mask && (j & 1)) | |
2809 | mask_op = permute_vec_elements (mask_op, mask_op, | |
2810 | mask_perm_mask, stmt, gsi); | |
2811 | else | |
2812 | { | |
2813 | if (j == 0) | |
2814 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); | |
2815 | else | |
929b4411 | 2816 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
c48d2d35 RS |
2817 | |
2818 | mask_op = vec_mask; | |
2819 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
2820 | { | |
2821 | gcc_assert | |
2822 | (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)), | |
2823 | TYPE_VECTOR_SUBPARTS (masktype))); | |
2824 | var = vect_get_new_ssa_name (masktype, vect_simple_var); | |
2825 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); | |
2826 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, | |
2827 | mask_op); | |
2828 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2829 | mask_op = var; | |
2830 | } | |
2831 | } | |
2832 | src_op = mask_op; | |
2833 | } | |
2834 | ||
2835 | new_stmt = gimple_build_call (gs_info->decl, 5, src_op, ptr, op, | |
2836 | mask_op, scale); | |
2837 | ||
2838 | if (!useless_type_conversion_p (vectype, rettype)) | |
2839 | { | |
2840 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (vectype), | |
2841 | TYPE_VECTOR_SUBPARTS (rettype))); | |
2842 | op = vect_get_new_ssa_name (rettype, vect_simple_var); | |
2843 | gimple_call_set_lhs (new_stmt, op); | |
2844 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2845 | var = make_ssa_name (vec_dest); | |
2846 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); | |
2847 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
2848 | } | |
2849 | else | |
2850 | { | |
2851 | var = make_ssa_name (vec_dest, new_stmt); | |
2852 | gimple_call_set_lhs (new_stmt, var); | |
2853 | } | |
2854 | ||
2855 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2856 | ||
2857 | if (modifier == NARROW) | |
2858 | { | |
2859 | if ((j & 1) == 0) | |
2860 | { | |
2861 | prev_res = var; | |
2862 | continue; | |
2863 | } | |
2864 | var = permute_vec_elements (prev_res, var, perm_mask, stmt, gsi); | |
2865 | new_stmt = SSA_NAME_DEF_STMT (var); | |
2866 | } | |
2867 | ||
2868 | if (prev_stmt_info == NULL) | |
2869 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2870 | else | |
2871 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2872 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2873 | } | |
2874 | } | |
2875 | ||
bfaa08b7 RS |
2876 | /* Prepare the base and offset in GS_INFO for vectorization. |
2877 | Set *DATAREF_PTR to the loop-invariant base address and *VEC_OFFSET | |
2878 | to the vectorized offset argument for the first copy of STMT. STMT | |
2879 | is the statement described by GS_INFO and LOOP is the containing loop. */ | |
2880 | ||
2881 | static void | |
2882 | vect_get_gather_scatter_ops (struct loop *loop, gimple *stmt, | |
2883 | gather_scatter_info *gs_info, | |
2884 | tree *dataref_ptr, tree *vec_offset) | |
2885 | { | |
2886 | gimple_seq stmts = NULL; | |
2887 | *dataref_ptr = force_gimple_operand (gs_info->base, &stmts, true, NULL_TREE); | |
2888 | if (stmts != NULL) | |
2889 | { | |
2890 | basic_block new_bb; | |
2891 | edge pe = loop_preheader_edge (loop); | |
2892 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
2893 | gcc_assert (!new_bb); | |
2894 | } | |
2895 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2896 | tree offset_vectype = get_vectype_for_scalar_type (offset_type); | |
2897 | *vec_offset = vect_get_vec_def_for_operand (gs_info->offset, stmt, | |
2898 | offset_vectype); | |
2899 | } | |
2900 | ||
ab2fc782 RS |
2901 | /* Prepare to implement a grouped or strided load or store using |
2902 | the gather load or scatter store operation described by GS_INFO. | |
2903 | STMT is the load or store statement. | |
2904 | ||
2905 | Set *DATAREF_BUMP to the amount that should be added to the base | |
2906 | address after each copy of the vectorized statement. Set *VEC_OFFSET | |
2907 | to an invariant offset vector in which element I has the value | |
2908 | I * DR_STEP / SCALE. */ | |
2909 | ||
2910 | static void | |
2911 | vect_get_strided_load_store_ops (gimple *stmt, loop_vec_info loop_vinfo, | |
2912 | gather_scatter_info *gs_info, | |
2913 | tree *dataref_bump, tree *vec_offset) | |
2914 | { | |
2915 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2916 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2917 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2918 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2919 | gimple_seq stmts; | |
2920 | ||
2921 | tree bump = size_binop (MULT_EXPR, | |
2922 | fold_convert (sizetype, DR_STEP (dr)), | |
2923 | size_int (TYPE_VECTOR_SUBPARTS (vectype))); | |
2924 | *dataref_bump = force_gimple_operand (bump, &stmts, true, NULL_TREE); | |
2925 | if (stmts) | |
2926 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
2927 | ||
2928 | /* The offset given in GS_INFO can have pointer type, so use the element | |
2929 | type of the vector instead. */ | |
2930 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2931 | tree offset_vectype = get_vectype_for_scalar_type (offset_type); | |
2932 | offset_type = TREE_TYPE (offset_vectype); | |
2933 | ||
2934 | /* Calculate X = DR_STEP / SCALE and convert it to the appropriate type. */ | |
2935 | tree step = size_binop (EXACT_DIV_EXPR, DR_STEP (dr), | |
2936 | ssize_int (gs_info->scale)); | |
2937 | step = fold_convert (offset_type, step); | |
2938 | step = force_gimple_operand (step, &stmts, true, NULL_TREE); | |
2939 | ||
2940 | /* Create {0, X, X*2, X*3, ...}. */ | |
2941 | *vec_offset = gimple_build (&stmts, VEC_SERIES_EXPR, offset_vectype, | |
2942 | build_zero_cst (offset_type), step); | |
2943 | if (stmts) | |
2944 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
2945 | } | |
2946 | ||
2947 | /* Return the amount that should be added to a vector pointer to move | |
2948 | to the next or previous copy of AGGR_TYPE. DR is the data reference | |
2949 | being vectorized and MEMORY_ACCESS_TYPE describes the type of | |
2950 | vectorization. */ | |
2951 | ||
2952 | static tree | |
2953 | vect_get_data_ptr_increment (data_reference *dr, tree aggr_type, | |
2954 | vect_memory_access_type memory_access_type) | |
2955 | { | |
2956 | if (memory_access_type == VMAT_INVARIANT) | |
2957 | return size_zero_node; | |
2958 | ||
2959 | tree iv_step = TYPE_SIZE_UNIT (aggr_type); | |
2960 | tree step = vect_dr_behavior (dr)->step; | |
2961 | if (tree_int_cst_sgn (step) == -1) | |
2962 | iv_step = fold_build1 (NEGATE_EXPR, TREE_TYPE (iv_step), iv_step); | |
2963 | return iv_step; | |
2964 | } | |
2965 | ||
37b14185 RB |
2966 | /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64}. */ |
2967 | ||
2968 | static bool | |
2969 | vectorizable_bswap (gimple *stmt, gimple_stmt_iterator *gsi, | |
2970 | gimple **vec_stmt, slp_tree slp_node, | |
68435eb2 RB |
2971 | tree vectype_in, enum vect_def_type *dt, |
2972 | stmt_vector_for_cost *cost_vec) | |
37b14185 RB |
2973 | { |
2974 | tree op, vectype; | |
2975 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2976 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
928686b1 RS |
2977 | unsigned ncopies; |
2978 | unsigned HOST_WIDE_INT nunits, num_bytes; | |
37b14185 RB |
2979 | |
2980 | op = gimple_call_arg (stmt, 0); | |
2981 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
928686b1 RS |
2982 | |
2983 | if (!TYPE_VECTOR_SUBPARTS (vectype).is_constant (&nunits)) | |
2984 | return false; | |
37b14185 RB |
2985 | |
2986 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2987 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2988 | case of SLP. */ | |
2989 | if (slp_node) | |
2990 | ncopies = 1; | |
2991 | else | |
e8f142e2 | 2992 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
37b14185 RB |
2993 | |
2994 | gcc_assert (ncopies >= 1); | |
2995 | ||
2996 | tree char_vectype = get_same_sized_vectype (char_type_node, vectype_in); | |
2997 | if (! char_vectype) | |
2998 | return false; | |
2999 | ||
928686b1 RS |
3000 | if (!TYPE_VECTOR_SUBPARTS (char_vectype).is_constant (&num_bytes)) |
3001 | return false; | |
3002 | ||
794e3180 | 3003 | unsigned word_bytes = num_bytes / nunits; |
908a1a16 | 3004 | |
d980067b RS |
3005 | /* The encoding uses one stepped pattern for each byte in the word. */ |
3006 | vec_perm_builder elts (num_bytes, word_bytes, 3); | |
3007 | for (unsigned i = 0; i < 3; ++i) | |
37b14185 | 3008 | for (unsigned j = 0; j < word_bytes; ++j) |
908a1a16 | 3009 | elts.quick_push ((i + 1) * word_bytes - j - 1); |
37b14185 | 3010 | |
e3342de4 RS |
3011 | vec_perm_indices indices (elts, 1, num_bytes); |
3012 | if (!can_vec_perm_const_p (TYPE_MODE (char_vectype), indices)) | |
37b14185 RB |
3013 | return false; |
3014 | ||
3015 | if (! vec_stmt) | |
3016 | { | |
3017 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
adac3a68 | 3018 | DUMP_VECT_SCOPE ("vectorizable_bswap"); |
78604de0 | 3019 | if (! slp_node) |
37b14185 | 3020 | { |
68435eb2 RB |
3021 | record_stmt_cost (cost_vec, |
3022 | 1, vector_stmt, stmt_info, 0, vect_prologue); | |
3023 | record_stmt_cost (cost_vec, | |
3024 | ncopies, vec_perm, stmt_info, 0, vect_body); | |
37b14185 RB |
3025 | } |
3026 | return true; | |
3027 | } | |
3028 | ||
736d0f28 | 3029 | tree bswap_vconst = vec_perm_indices_to_tree (char_vectype, indices); |
37b14185 RB |
3030 | |
3031 | /* Transform. */ | |
3032 | vec<tree> vec_oprnds = vNULL; | |
3033 | gimple *new_stmt = NULL; | |
3034 | stmt_vec_info prev_stmt_info = NULL; | |
3035 | for (unsigned j = 0; j < ncopies; j++) | |
3036 | { | |
3037 | /* Handle uses. */ | |
3038 | if (j == 0) | |
306b0c92 | 3039 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
37b14185 RB |
3040 | else |
3041 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
3042 | ||
3043 | /* Arguments are ready. create the new vector stmt. */ | |
3044 | unsigned i; | |
3045 | tree vop; | |
3046 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) | |
3047 | { | |
3048 | tree tem = make_ssa_name (char_vectype); | |
3049 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
3050 | char_vectype, vop)); | |
3051 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3052 | tree tem2 = make_ssa_name (char_vectype); | |
3053 | new_stmt = gimple_build_assign (tem2, VEC_PERM_EXPR, | |
3054 | tem, tem, bswap_vconst); | |
3055 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3056 | tem = make_ssa_name (vectype); | |
3057 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
3058 | vectype, tem2)); | |
3059 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3060 | if (slp_node) | |
3061 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
3062 | } | |
3063 | ||
3064 | if (slp_node) | |
3065 | continue; | |
3066 | ||
3067 | if (j == 0) | |
3068 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3069 | else | |
3070 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3071 | ||
3072 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3073 | } | |
3074 | ||
3075 | vec_oprnds.release (); | |
3076 | return true; | |
3077 | } | |
3078 | ||
b1b6836e RS |
3079 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
3080 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT | |
3081 | in a single step. On success, store the binary pack code in | |
3082 | *CONVERT_CODE. */ | |
3083 | ||
3084 | static bool | |
3085 | simple_integer_narrowing (tree vectype_out, tree vectype_in, | |
3086 | tree_code *convert_code) | |
3087 | { | |
3088 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) | |
3089 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) | |
3090 | return false; | |
3091 | ||
3092 | tree_code code; | |
3093 | int multi_step_cvt = 0; | |
3094 | auto_vec <tree, 8> interm_types; | |
3095 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, | |
3096 | &code, &multi_step_cvt, | |
3097 | &interm_types) | |
3098 | || multi_step_cvt) | |
3099 | return false; | |
3100 | ||
3101 | *convert_code = code; | |
3102 | return true; | |
3103 | } | |
5ce9450f | 3104 | |
ebfd146a IR |
3105 | /* Function vectorizable_call. |
3106 | ||
538dd0b7 | 3107 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 3108 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
3109 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3110 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3111 | ||
3112 | static bool | |
355fe088 | 3113 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
68435eb2 | 3114 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
ebfd146a | 3115 | { |
538dd0b7 | 3116 | gcall *stmt; |
ebfd146a IR |
3117 | tree vec_dest; |
3118 | tree scalar_dest; | |
0267732b | 3119 | tree op; |
ebfd146a | 3120 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
538dd0b7 | 3121 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a | 3122 | tree vectype_out, vectype_in; |
c7bda0f4 RS |
3123 | poly_uint64 nunits_in; |
3124 | poly_uint64 nunits_out; | |
ebfd146a | 3125 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
190c2236 | 3126 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 3127 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 3128 | tree fndecl, new_temp, rhs_type; |
2c58d42c RS |
3129 | enum vect_def_type dt[4] |
3130 | = { vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type, | |
3131 | vect_unknown_def_type }; | |
3132 | int ndts = ARRAY_SIZE (dt); | |
355fe088 | 3133 | gimple *new_stmt = NULL; |
ebfd146a | 3134 | int ncopies, j; |
2c58d42c RS |
3135 | auto_vec<tree, 8> vargs; |
3136 | auto_vec<tree, 8> orig_vargs; | |
ebfd146a IR |
3137 | enum { NARROW, NONE, WIDEN } modifier; |
3138 | size_t i, nargs; | |
9d5e7640 | 3139 | tree lhs; |
ebfd146a | 3140 | |
190c2236 | 3141 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3142 | return false; |
3143 | ||
66c16fd9 RB |
3144 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3145 | && ! vec_stmt) | |
ebfd146a IR |
3146 | return false; |
3147 | ||
538dd0b7 DM |
3148 | /* Is GS a vectorizable call? */ |
3149 | stmt = dyn_cast <gcall *> (gs); | |
3150 | if (!stmt) | |
ebfd146a IR |
3151 | return false; |
3152 | ||
5ce9450f | 3153 | if (gimple_call_internal_p (stmt) |
bfaa08b7 | 3154 | && (internal_load_fn_p (gimple_call_internal_fn (stmt)) |
f307441a | 3155 | || internal_store_fn_p (gimple_call_internal_fn (stmt)))) |
c3a8f964 RS |
3156 | /* Handled by vectorizable_load and vectorizable_store. */ |
3157 | return false; | |
5ce9450f | 3158 | |
0136f8f0 AH |
3159 | if (gimple_call_lhs (stmt) == NULL_TREE |
3160 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
3161 | return false; |
3162 | ||
0136f8f0 | 3163 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 3164 | |
b690cc0f RG |
3165 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3166 | ||
ebfd146a IR |
3167 | /* Process function arguments. */ |
3168 | rhs_type = NULL_TREE; | |
b690cc0f | 3169 | vectype_in = NULL_TREE; |
ebfd146a IR |
3170 | nargs = gimple_call_num_args (stmt); |
3171 | ||
1b1562a5 MM |
3172 | /* Bail out if the function has more than three arguments, we do not have |
3173 | interesting builtin functions to vectorize with more than two arguments | |
3174 | except for fma. No arguments is also not good. */ | |
2c58d42c | 3175 | if (nargs == 0 || nargs > 4) |
ebfd146a IR |
3176 | return false; |
3177 | ||
74bf76ed | 3178 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2c58d42c RS |
3179 | combined_fn cfn = gimple_call_combined_fn (stmt); |
3180 | if (cfn == CFN_GOMP_SIMD_LANE) | |
74bf76ed JJ |
3181 | { |
3182 | nargs = 0; | |
3183 | rhs_type = unsigned_type_node; | |
3184 | } | |
3185 | ||
2c58d42c RS |
3186 | int mask_opno = -1; |
3187 | if (internal_fn_p (cfn)) | |
3188 | mask_opno = internal_fn_mask_index (as_internal_fn (cfn)); | |
3189 | ||
ebfd146a IR |
3190 | for (i = 0; i < nargs; i++) |
3191 | { | |
b690cc0f RG |
3192 | tree opvectype; |
3193 | ||
ebfd146a | 3194 | op = gimple_call_arg (stmt, i); |
2c58d42c RS |
3195 | if (!vect_is_simple_use (op, vinfo, &dt[i], &opvectype)) |
3196 | { | |
3197 | if (dump_enabled_p ()) | |
3198 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3199 | "use not simple.\n"); | |
3200 | return false; | |
3201 | } | |
3202 | ||
3203 | /* Skip the mask argument to an internal function. This operand | |
3204 | has been converted via a pattern if necessary. */ | |
3205 | if ((int) i == mask_opno) | |
3206 | continue; | |
ebfd146a IR |
3207 | |
3208 | /* We can only handle calls with arguments of the same type. */ | |
3209 | if (rhs_type | |
8533c9d8 | 3210 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 3211 | { |
73fbfcad | 3212 | if (dump_enabled_p ()) |
78c60e3d | 3213 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3214 | "argument types differ.\n"); |
ebfd146a IR |
3215 | return false; |
3216 | } | |
b690cc0f RG |
3217 | if (!rhs_type) |
3218 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 3219 | |
b690cc0f RG |
3220 | if (!vectype_in) |
3221 | vectype_in = opvectype; | |
3222 | else if (opvectype | |
3223 | && opvectype != vectype_in) | |
3224 | { | |
73fbfcad | 3225 | if (dump_enabled_p ()) |
78c60e3d | 3226 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3227 | "argument vector types differ.\n"); |
b690cc0f RG |
3228 | return false; |
3229 | } | |
3230 | } | |
3231 | /* If all arguments are external or constant defs use a vector type with | |
3232 | the same size as the output vector type. */ | |
ebfd146a | 3233 | if (!vectype_in) |
b690cc0f | 3234 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
3235 | if (vec_stmt) |
3236 | gcc_assert (vectype_in); | |
3237 | if (!vectype_in) | |
3238 | { | |
73fbfcad | 3239 | if (dump_enabled_p ()) |
7d8930a0 | 3240 | { |
78c60e3d SS |
3241 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3242 | "no vectype for scalar type "); | |
3243 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 3244 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
3245 | } |
3246 | ||
3247 | return false; | |
3248 | } | |
ebfd146a IR |
3249 | |
3250 | /* FORNOW */ | |
b690cc0f RG |
3251 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
3252 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
c7bda0f4 | 3253 | if (known_eq (nunits_in * 2, nunits_out)) |
ebfd146a | 3254 | modifier = NARROW; |
c7bda0f4 | 3255 | else if (known_eq (nunits_out, nunits_in)) |
ebfd146a | 3256 | modifier = NONE; |
c7bda0f4 | 3257 | else if (known_eq (nunits_out * 2, nunits_in)) |
ebfd146a IR |
3258 | modifier = WIDEN; |
3259 | else | |
3260 | return false; | |
3261 | ||
70439f0d RS |
3262 | /* We only handle functions that do not read or clobber memory. */ |
3263 | if (gimple_vuse (stmt)) | |
3264 | { | |
3265 | if (dump_enabled_p ()) | |
3266 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3267 | "function reads from or writes to memory.\n"); | |
3268 | return false; | |
3269 | } | |
3270 | ||
ebfd146a IR |
3271 | /* For now, we only vectorize functions if a target specific builtin |
3272 | is available. TODO -- in some cases, it might be profitable to | |
3273 | insert the calls for pieces of the vector, in order to be able | |
3274 | to vectorize other operations in the loop. */ | |
70439f0d RS |
3275 | fndecl = NULL_TREE; |
3276 | internal_fn ifn = IFN_LAST; | |
70439f0d RS |
3277 | tree callee = gimple_call_fndecl (stmt); |
3278 | ||
3279 | /* First try using an internal function. */ | |
b1b6836e RS |
3280 | tree_code convert_code = ERROR_MARK; |
3281 | if (cfn != CFN_LAST | |
3282 | && (modifier == NONE | |
3283 | || (modifier == NARROW | |
3284 | && simple_integer_narrowing (vectype_out, vectype_in, | |
3285 | &convert_code)))) | |
70439f0d RS |
3286 | ifn = vectorizable_internal_function (cfn, callee, vectype_out, |
3287 | vectype_in); | |
3288 | ||
3289 | /* If that fails, try asking for a target-specific built-in function. */ | |
3290 | if (ifn == IFN_LAST) | |
3291 | { | |
3292 | if (cfn != CFN_LAST) | |
3293 | fndecl = targetm.vectorize.builtin_vectorized_function | |
3294 | (cfn, vectype_out, vectype_in); | |
7672aa9b | 3295 | else if (callee) |
70439f0d RS |
3296 | fndecl = targetm.vectorize.builtin_md_vectorized_function |
3297 | (callee, vectype_out, vectype_in); | |
3298 | } | |
3299 | ||
3300 | if (ifn == IFN_LAST && !fndecl) | |
ebfd146a | 3301 | { |
70439f0d | 3302 | if (cfn == CFN_GOMP_SIMD_LANE |
74bf76ed JJ |
3303 | && !slp_node |
3304 | && loop_vinfo | |
3305 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
3306 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
3307 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
3308 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
3309 | { | |
3310 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
3311 | { 0, 1, 2, ... vf - 1 } vector. */ | |
3312 | gcc_assert (nargs == 0); | |
3313 | } | |
37b14185 RB |
3314 | else if (modifier == NONE |
3315 | && (gimple_call_builtin_p (stmt, BUILT_IN_BSWAP16) | |
3316 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP32) | |
3317 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP64))) | |
3318 | return vectorizable_bswap (stmt, gsi, vec_stmt, slp_node, | |
68435eb2 | 3319 | vectype_in, dt, cost_vec); |
74bf76ed JJ |
3320 | else |
3321 | { | |
3322 | if (dump_enabled_p ()) | |
3323 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 3324 | "function is not vectorizable.\n"); |
74bf76ed JJ |
3325 | return false; |
3326 | } | |
ebfd146a IR |
3327 | } |
3328 | ||
fce57248 | 3329 | if (slp_node) |
190c2236 | 3330 | ncopies = 1; |
b1b6836e | 3331 | else if (modifier == NARROW && ifn == IFN_LAST) |
e8f142e2 | 3332 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
ebfd146a | 3333 | else |
e8f142e2 | 3334 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
ebfd146a IR |
3335 | |
3336 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3337 | needs to be generated. */ | |
3338 | gcc_assert (ncopies >= 1); | |
3339 | ||
ed623edb | 3340 | vec_loop_masks *masks = (loop_vinfo ? &LOOP_VINFO_MASKS (loop_vinfo) : NULL); |
ebfd146a IR |
3341 | if (!vec_stmt) /* transformation not required. */ |
3342 | { | |
3343 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
adac3a68 | 3344 | DUMP_VECT_SCOPE ("vectorizable_call"); |
68435eb2 RB |
3345 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
3346 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) | |
3347 | record_stmt_cost (cost_vec, ncopies / 2, | |
3348 | vec_promote_demote, stmt_info, 0, vect_body); | |
b1b6836e | 3349 | |
2c58d42c RS |
3350 | if (loop_vinfo && mask_opno >= 0) |
3351 | { | |
3352 | unsigned int nvectors = (slp_node | |
3353 | ? SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node) | |
3354 | : ncopies); | |
3355 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype_out); | |
3356 | } | |
ebfd146a IR |
3357 | return true; |
3358 | } | |
3359 | ||
67b8dbac | 3360 | /* Transform. */ |
ebfd146a | 3361 | |
73fbfcad | 3362 | if (dump_enabled_p ()) |
e645e942 | 3363 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
3364 | |
3365 | /* Handle def. */ | |
3366 | scalar_dest = gimple_call_lhs (stmt); | |
3367 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
3368 | ||
2c58d42c RS |
3369 | bool masked_loop_p = loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo); |
3370 | ||
ebfd146a | 3371 | prev_stmt_info = NULL; |
b1b6836e | 3372 | if (modifier == NONE || ifn != IFN_LAST) |
ebfd146a | 3373 | { |
b1b6836e | 3374 | tree prev_res = NULL_TREE; |
2c58d42c RS |
3375 | vargs.safe_grow (nargs); |
3376 | orig_vargs.safe_grow (nargs); | |
ebfd146a IR |
3377 | for (j = 0; j < ncopies; ++j) |
3378 | { | |
3379 | /* Build argument list for the vectorized call. */ | |
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++) | |
2c58d42c | 3386 | vargs[i] = 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_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
3392 | { |
3393 | size_t k; | |
3394 | for (k = 0; k < nargs; k++) | |
3395 | { | |
37b5ec8f | 3396 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 3397 | vargs[k] = vec_oprndsk[i]; |
190c2236 | 3398 | } |
b1b6836e RS |
3399 | if (modifier == NARROW) |
3400 | { | |
2c58d42c RS |
3401 | /* We don't define any narrowing conditional functions |
3402 | at present. */ | |
3403 | gcc_assert (mask_opno < 0); | |
b1b6836e | 3404 | tree half_res = make_ssa_name (vectype_in); |
a844293d RS |
3405 | gcall *call |
3406 | = gimple_build_call_internal_vec (ifn, vargs); | |
3407 | gimple_call_set_lhs (call, half_res); | |
3408 | gimple_call_set_nothrow (call, true); | |
3409 | new_stmt = call; | |
b1b6836e RS |
3410 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3411 | if ((i & 1) == 0) | |
3412 | { | |
3413 | prev_res = half_res; | |
3414 | continue; | |
3415 | } | |
3416 | new_temp = make_ssa_name (vec_dest); | |
3417 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
3418 | prev_res, half_res); | |
3419 | } | |
70439f0d | 3420 | else |
b1b6836e | 3421 | { |
2c58d42c RS |
3422 | if (mask_opno >= 0 && masked_loop_p) |
3423 | { | |
3424 | unsigned int vec_num = vec_oprnds0.length (); | |
3425 | /* Always true for SLP. */ | |
3426 | gcc_assert (ncopies == 1); | |
3427 | tree mask = vect_get_loop_mask (gsi, masks, vec_num, | |
3428 | vectype_out, i); | |
3429 | vargs[mask_opno] = prepare_load_store_mask | |
3430 | (TREE_TYPE (mask), mask, vargs[mask_opno], gsi); | |
3431 | } | |
3432 | ||
a844293d | 3433 | gcall *call; |
b1b6836e | 3434 | if (ifn != IFN_LAST) |
a844293d | 3435 | call = gimple_build_call_internal_vec (ifn, vargs); |
b1b6836e | 3436 | else |
a844293d RS |
3437 | call = gimple_build_call_vec (fndecl, vargs); |
3438 | new_temp = make_ssa_name (vec_dest, call); | |
3439 | gimple_call_set_lhs (call, new_temp); | |
3440 | gimple_call_set_nothrow (call, true); | |
3441 | new_stmt = call; | |
b1b6836e | 3442 | } |
190c2236 | 3443 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 3444 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
3445 | } |
3446 | ||
3447 | for (i = 0; i < nargs; i++) | |
3448 | { | |
37b5ec8f | 3449 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3450 | vec_oprndsi.release (); |
190c2236 | 3451 | } |
190c2236 JJ |
3452 | continue; |
3453 | } | |
3454 | ||
ebfd146a IR |
3455 | for (i = 0; i < nargs; i++) |
3456 | { | |
3457 | op = gimple_call_arg (stmt, i); | |
3458 | if (j == 0) | |
3459 | vec_oprnd0 | |
81c40241 | 3460 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 3461 | else |
2c58d42c RS |
3462 | vec_oprnd0 |
3463 | = vect_get_vec_def_for_stmt_copy (dt[i], orig_vargs[i]); | |
3464 | ||
3465 | orig_vargs[i] = vargs[i] = vec_oprnd0; | |
3466 | } | |
ebfd146a | 3467 | |
2c58d42c RS |
3468 | if (mask_opno >= 0 && masked_loop_p) |
3469 | { | |
3470 | tree mask = vect_get_loop_mask (gsi, masks, ncopies, | |
3471 | vectype_out, j); | |
3472 | vargs[mask_opno] | |
3473 | = prepare_load_store_mask (TREE_TYPE (mask), mask, | |
3474 | vargs[mask_opno], gsi); | |
ebfd146a IR |
3475 | } |
3476 | ||
2c58d42c | 3477 | if (cfn == CFN_GOMP_SIMD_LANE) |
74bf76ed | 3478 | { |
c7bda0f4 | 3479 | tree cst = build_index_vector (vectype_out, j * nunits_out, 1); |
74bf76ed | 3480 | tree new_var |
0e22bb5a | 3481 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 3482 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 3483 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 3484 | new_temp = make_ssa_name (vec_dest); |
0e22bb5a | 3485 | new_stmt = gimple_build_assign (new_temp, new_var); |
74bf76ed | 3486 | } |
b1b6836e RS |
3487 | else if (modifier == NARROW) |
3488 | { | |
2c58d42c RS |
3489 | /* We don't define any narrowing conditional functions at |
3490 | present. */ | |
3491 | gcc_assert (mask_opno < 0); | |
b1b6836e | 3492 | tree half_res = make_ssa_name (vectype_in); |
a844293d RS |
3493 | gcall *call = gimple_build_call_internal_vec (ifn, vargs); |
3494 | gimple_call_set_lhs (call, half_res); | |
3495 | gimple_call_set_nothrow (call, true); | |
3496 | new_stmt = call; | |
b1b6836e RS |
3497 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3498 | if ((j & 1) == 0) | |
3499 | { | |
3500 | prev_res = half_res; | |
3501 | continue; | |
3502 | } | |
3503 | new_temp = make_ssa_name (vec_dest); | |
3504 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
3505 | prev_res, half_res); | |
3506 | } | |
74bf76ed JJ |
3507 | else |
3508 | { | |
a844293d | 3509 | gcall *call; |
70439f0d | 3510 | if (ifn != IFN_LAST) |
a844293d | 3511 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3512 | else |
a844293d | 3513 | call = gimple_build_call_vec (fndecl, vargs); |
74bf76ed | 3514 | new_temp = make_ssa_name (vec_dest, new_stmt); |
a844293d RS |
3515 | gimple_call_set_lhs (call, new_temp); |
3516 | gimple_call_set_nothrow (call, true); | |
3517 | new_stmt = call; | |
74bf76ed | 3518 | } |
ebfd146a IR |
3519 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3520 | ||
b1b6836e | 3521 | if (j == (modifier == NARROW ? 1 : 0)) |
ebfd146a IR |
3522 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; |
3523 | else | |
3524 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3525 | ||
3526 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3527 | } | |
b1b6836e RS |
3528 | } |
3529 | else if (modifier == NARROW) | |
3530 | { | |
2c58d42c RS |
3531 | /* We don't define any narrowing conditional functions at present. */ |
3532 | gcc_assert (mask_opno < 0); | |
ebfd146a IR |
3533 | for (j = 0; j < ncopies; ++j) |
3534 | { | |
3535 | /* Build argument list for the vectorized call. */ | |
3536 | if (j == 0) | |
9771b263 | 3537 | vargs.create (nargs * 2); |
ebfd146a | 3538 | else |
9771b263 | 3539 | vargs.truncate (0); |
ebfd146a | 3540 | |
190c2236 JJ |
3541 | if (slp_node) |
3542 | { | |
ef062b13 | 3543 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 3544 | vec<tree> vec_oprnds0; |
190c2236 JJ |
3545 | |
3546 | for (i = 0; i < nargs; i++) | |
9771b263 | 3547 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 3548 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 3549 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
3550 | |
3551 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3552 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
3553 | { |
3554 | size_t k; | |
9771b263 | 3555 | vargs.truncate (0); |
190c2236 JJ |
3556 | for (k = 0; k < nargs; k++) |
3557 | { | |
37b5ec8f | 3558 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
3559 | vargs.quick_push (vec_oprndsk[i]); |
3560 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 | 3561 | } |
a844293d | 3562 | gcall *call; |
70439f0d | 3563 | if (ifn != IFN_LAST) |
a844293d | 3564 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3565 | else |
a844293d RS |
3566 | call = gimple_build_call_vec (fndecl, vargs); |
3567 | new_temp = make_ssa_name (vec_dest, call); | |
3568 | gimple_call_set_lhs (call, new_temp); | |
3569 | gimple_call_set_nothrow (call, true); | |
3570 | new_stmt = call; | |
190c2236 | 3571 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 3572 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
3573 | } |
3574 | ||
3575 | for (i = 0; i < nargs; i++) | |
3576 | { | |
37b5ec8f | 3577 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3578 | vec_oprndsi.release (); |
190c2236 | 3579 | } |
190c2236 JJ |
3580 | continue; |
3581 | } | |
3582 | ||
ebfd146a IR |
3583 | for (i = 0; i < nargs; i++) |
3584 | { | |
3585 | op = gimple_call_arg (stmt, i); | |
3586 | if (j == 0) | |
3587 | { | |
3588 | vec_oprnd0 | |
81c40241 | 3589 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 3590 | vec_oprnd1 |
63827fb8 | 3591 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3592 | } |
3593 | else | |
3594 | { | |
336ecb65 | 3595 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 3596 | vec_oprnd0 |
63827fb8 | 3597 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 3598 | vec_oprnd1 |
63827fb8 | 3599 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3600 | } |
3601 | ||
9771b263 DN |
3602 | vargs.quick_push (vec_oprnd0); |
3603 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
3604 | } |
3605 | ||
b1b6836e | 3606 | new_stmt = gimple_build_call_vec (fndecl, vargs); |
ebfd146a IR |
3607 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3608 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
3609 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3610 | ||
3611 | if (j == 0) | |
3612 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3613 | else | |
3614 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3615 | ||
3616 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3617 | } | |
3618 | ||
3619 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a | 3620 | } |
b1b6836e RS |
3621 | else |
3622 | /* No current target implements this case. */ | |
3623 | return false; | |
ebfd146a | 3624 | |
9771b263 | 3625 | vargs.release (); |
ebfd146a | 3626 | |
ebfd146a IR |
3627 | /* The call in STMT might prevent it from being removed in dce. |
3628 | We however cannot remove it here, due to the way the ssa name | |
3629 | it defines is mapped to the new definition. So just replace | |
3630 | rhs of the statement with something harmless. */ | |
3631 | ||
dd34c087 JJ |
3632 | if (slp_node) |
3633 | return true; | |
3634 | ||
9d5e7640 | 3635 | if (is_pattern_stmt_p (stmt_info)) |
ed7b8123 RS |
3636 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
3637 | lhs = gimple_get_lhs (stmt_info->stmt); | |
3cc2fa2a | 3638 | |
0267732b | 3639 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); |
ebfd146a | 3640 | set_vinfo_for_stmt (new_stmt, stmt_info); |
ed7b8123 | 3641 | set_vinfo_for_stmt (stmt_info->stmt, NULL); |
ebfd146a IR |
3642 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
3643 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
3644 | |
3645 | return true; | |
3646 | } | |
3647 | ||
3648 | ||
0136f8f0 AH |
3649 | struct simd_call_arg_info |
3650 | { | |
3651 | tree vectype; | |
3652 | tree op; | |
0136f8f0 | 3653 | HOST_WIDE_INT linear_step; |
34e82342 | 3654 | enum vect_def_type dt; |
0136f8f0 | 3655 | unsigned int align; |
17b658af | 3656 | bool simd_lane_linear; |
0136f8f0 AH |
3657 | }; |
3658 | ||
17b658af JJ |
3659 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
3660 | is linear within simd lane (but not within whole loop), note it in | |
3661 | *ARGINFO. */ | |
3662 | ||
3663 | static void | |
3664 | vect_simd_lane_linear (tree op, struct loop *loop, | |
3665 | struct simd_call_arg_info *arginfo) | |
3666 | { | |
355fe088 | 3667 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
3668 | |
3669 | if (!is_gimple_assign (def_stmt) | |
3670 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
3671 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
3672 | return; | |
3673 | ||
3674 | tree base = gimple_assign_rhs1 (def_stmt); | |
3675 | HOST_WIDE_INT linear_step = 0; | |
3676 | tree v = gimple_assign_rhs2 (def_stmt); | |
3677 | while (TREE_CODE (v) == SSA_NAME) | |
3678 | { | |
3679 | tree t; | |
3680 | def_stmt = SSA_NAME_DEF_STMT (v); | |
3681 | if (is_gimple_assign (def_stmt)) | |
3682 | switch (gimple_assign_rhs_code (def_stmt)) | |
3683 | { | |
3684 | case PLUS_EXPR: | |
3685 | t = gimple_assign_rhs2 (def_stmt); | |
3686 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
3687 | return; | |
3688 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
3689 | v = gimple_assign_rhs1 (def_stmt); | |
3690 | continue; | |
3691 | case MULT_EXPR: | |
3692 | t = gimple_assign_rhs2 (def_stmt); | |
3693 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
3694 | return; | |
3695 | linear_step = tree_to_shwi (t); | |
3696 | v = gimple_assign_rhs1 (def_stmt); | |
3697 | continue; | |
3698 | CASE_CONVERT: | |
3699 | t = gimple_assign_rhs1 (def_stmt); | |
3700 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
3701 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
3702 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
3703 | return; | |
3704 | if (!linear_step) | |
3705 | linear_step = 1; | |
3706 | v = t; | |
3707 | continue; | |
3708 | default: | |
3709 | return; | |
3710 | } | |
8e4284d0 | 3711 | else if (gimple_call_internal_p (def_stmt, IFN_GOMP_SIMD_LANE) |
17b658af JJ |
3712 | && loop->simduid |
3713 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
3714 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
3715 | == loop->simduid)) | |
3716 | { | |
3717 | if (!linear_step) | |
3718 | linear_step = 1; | |
3719 | arginfo->linear_step = linear_step; | |
3720 | arginfo->op = base; | |
3721 | arginfo->simd_lane_linear = true; | |
3722 | return; | |
3723 | } | |
3724 | } | |
3725 | } | |
3726 | ||
cf1b2ba4 RS |
3727 | /* Return the number of elements in vector type VECTYPE, which is associated |
3728 | with a SIMD clone. At present these vectors always have a constant | |
3729 | length. */ | |
3730 | ||
3731 | static unsigned HOST_WIDE_INT | |
3732 | simd_clone_subparts (tree vectype) | |
3733 | { | |
928686b1 | 3734 | return TYPE_VECTOR_SUBPARTS (vectype).to_constant (); |
cf1b2ba4 RS |
3735 | } |
3736 | ||
0136f8f0 AH |
3737 | /* Function vectorizable_simd_clone_call. |
3738 | ||
3739 | Check if STMT performs a function call that can be vectorized | |
3740 | by calling a simd clone of the function. | |
3741 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3742 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3743 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3744 | ||
3745 | static bool | |
355fe088 | 3746 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
68435eb2 RB |
3747 | gimple **vec_stmt, slp_tree slp_node, |
3748 | stmt_vector_for_cost *) | |
0136f8f0 AH |
3749 | { |
3750 | tree vec_dest; | |
3751 | tree scalar_dest; | |
3752 | tree op, type; | |
3753 | tree vec_oprnd0 = NULL_TREE; | |
3754 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
3755 | tree vectype; | |
3756 | unsigned int nunits; | |
3757 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3758 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 3759 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 3760 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 3761 | tree fndecl, new_temp; |
355fe088 | 3762 | gimple *new_stmt = NULL; |
0136f8f0 | 3763 | int ncopies, j; |
00426f9a | 3764 | auto_vec<simd_call_arg_info> arginfo; |
0136f8f0 AH |
3765 | vec<tree> vargs = vNULL; |
3766 | size_t i, nargs; | |
3767 | tree lhs, rtype, ratype; | |
e7a74006 | 3768 | vec<constructor_elt, va_gc> *ret_ctor_elts = NULL; |
0136f8f0 AH |
3769 | |
3770 | /* Is STMT a vectorizable call? */ | |
3771 | if (!is_gimple_call (stmt)) | |
3772 | return false; | |
3773 | ||
3774 | fndecl = gimple_call_fndecl (stmt); | |
3775 | if (fndecl == NULL_TREE) | |
3776 | return false; | |
3777 | ||
d52f5295 | 3778 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
3779 | if (node == NULL || node->simd_clones == NULL) |
3780 | return false; | |
3781 | ||
3782 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
3783 | return false; | |
3784 | ||
66c16fd9 RB |
3785 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3786 | && ! vec_stmt) | |
0136f8f0 AH |
3787 | return false; |
3788 | ||
3789 | if (gimple_call_lhs (stmt) | |
3790 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
3791 | return false; | |
3792 | ||
3793 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
3794 | ||
3795 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3796 | ||
3797 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
3798 | return false; | |
3799 | ||
3800 | /* FORNOW */ | |
fce57248 | 3801 | if (slp_node) |
0136f8f0 AH |
3802 | return false; |
3803 | ||
3804 | /* Process function arguments. */ | |
3805 | nargs = gimple_call_num_args (stmt); | |
3806 | ||
3807 | /* Bail out if the function has zero arguments. */ | |
3808 | if (nargs == 0) | |
3809 | return false; | |
3810 | ||
00426f9a | 3811 | arginfo.reserve (nargs, true); |
0136f8f0 AH |
3812 | |
3813 | for (i = 0; i < nargs; i++) | |
3814 | { | |
3815 | simd_call_arg_info thisarginfo; | |
3816 | affine_iv iv; | |
3817 | ||
3818 | thisarginfo.linear_step = 0; | |
3819 | thisarginfo.align = 0; | |
3820 | thisarginfo.op = NULL_TREE; | |
17b658af | 3821 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
3822 | |
3823 | op = gimple_call_arg (stmt, i); | |
894dd753 | 3824 | if (!vect_is_simple_use (op, vinfo, &thisarginfo.dt, |
81c40241 | 3825 | &thisarginfo.vectype) |
0136f8f0 AH |
3826 | || thisarginfo.dt == vect_uninitialized_def) |
3827 | { | |
3828 | if (dump_enabled_p ()) | |
3829 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3830 | "use not simple.\n"); | |
0136f8f0 AH |
3831 | return false; |
3832 | } | |
3833 | ||
3834 | if (thisarginfo.dt == vect_constant_def | |
3835 | || thisarginfo.dt == vect_external_def) | |
3836 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
3837 | else | |
3838 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
3839 | ||
6c9e85fb JJ |
3840 | /* For linear arguments, the analyze phase should have saved |
3841 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
3842 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
3843 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
3844 | { |
3845 | gcc_assert (vec_stmt); | |
3846 | thisarginfo.linear_step | |
17b658af | 3847 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 3848 | thisarginfo.op |
17b658af JJ |
3849 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
3850 | thisarginfo.simd_lane_linear | |
3851 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
3852 | == boolean_true_node); | |
6c9e85fb JJ |
3853 | /* If loop has been peeled for alignment, we need to adjust it. */ |
3854 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
3855 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 3856 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
3857 | { |
3858 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 3859 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
3860 | tree opt = TREE_TYPE (thisarginfo.op); |
3861 | bias = fold_convert (TREE_TYPE (step), bias); | |
3862 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
3863 | thisarginfo.op | |
3864 | = fold_build2 (POINTER_TYPE_P (opt) | |
3865 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
3866 | thisarginfo.op, bias); | |
3867 | } | |
3868 | } | |
3869 | else if (!vec_stmt | |
3870 | && thisarginfo.dt != vect_constant_def | |
3871 | && thisarginfo.dt != vect_external_def | |
3872 | && loop_vinfo | |
3873 | && TREE_CODE (op) == SSA_NAME | |
3874 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
3875 | &iv, false) | |
3876 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
3877 | { |
3878 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
3879 | thisarginfo.op = iv.base; | |
3880 | } | |
3881 | else if ((thisarginfo.dt == vect_constant_def | |
3882 | || thisarginfo.dt == vect_external_def) | |
3883 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
3884 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
3885 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
3886 | linear too. */ | |
3887 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
3888 | && !thisarginfo.linear_step | |
3889 | && !vec_stmt | |
3890 | && thisarginfo.dt != vect_constant_def | |
3891 | && thisarginfo.dt != vect_external_def | |
3892 | && loop_vinfo | |
3893 | && !slp_node | |
3894 | && TREE_CODE (op) == SSA_NAME) | |
3895 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
3896 | |
3897 | arginfo.quick_push (thisarginfo); | |
3898 | } | |
3899 | ||
d9f21f6a RS |
3900 | unsigned HOST_WIDE_INT vf; |
3901 | if (!LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&vf)) | |
3902 | { | |
3903 | if (dump_enabled_p ()) | |
3904 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3905 | "not considering SIMD clones; not yet supported" | |
3906 | " for variable-width vectors.\n"); | |
3907 | return NULL; | |
3908 | } | |
3909 | ||
0136f8f0 AH |
3910 | unsigned int badness = 0; |
3911 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
3912 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
3913 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
3914 | else |
3915 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
3916 | n = n->simdclone->next_clone) | |
3917 | { | |
3918 | unsigned int this_badness = 0; | |
d9f21f6a | 3919 | if (n->simdclone->simdlen > vf |
0136f8f0 AH |
3920 | || n->simdclone->nargs != nargs) |
3921 | continue; | |
d9f21f6a RS |
3922 | if (n->simdclone->simdlen < vf) |
3923 | this_badness += (exact_log2 (vf) | |
0136f8f0 AH |
3924 | - exact_log2 (n->simdclone->simdlen)) * 1024; |
3925 | if (n->simdclone->inbranch) | |
3926 | this_badness += 2048; | |
3927 | int target_badness = targetm.simd_clone.usable (n); | |
3928 | if (target_badness < 0) | |
3929 | continue; | |
3930 | this_badness += target_badness * 512; | |
3931 | /* FORNOW: Have to add code to add the mask argument. */ | |
3932 | if (n->simdclone->inbranch) | |
3933 | continue; | |
3934 | for (i = 0; i < nargs; i++) | |
3935 | { | |
3936 | switch (n->simdclone->args[i].arg_type) | |
3937 | { | |
3938 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3939 | if (!useless_type_conversion_p | |
3940 | (n->simdclone->args[i].orig_type, | |
3941 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
3942 | i = -1; | |
3943 | else if (arginfo[i].dt == vect_constant_def | |
3944 | || arginfo[i].dt == vect_external_def | |
3945 | || arginfo[i].linear_step) | |
3946 | this_badness += 64; | |
3947 | break; | |
3948 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3949 | if (arginfo[i].dt != vect_constant_def | |
3950 | && arginfo[i].dt != vect_external_def) | |
3951 | i = -1; | |
3952 | break; | |
3953 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 3954 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
3955 | if (arginfo[i].dt == vect_constant_def |
3956 | || arginfo[i].dt == vect_external_def | |
3957 | || (arginfo[i].linear_step | |
3958 | != n->simdclone->args[i].linear_step)) | |
3959 | i = -1; | |
3960 | break; | |
3961 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
3962 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
3963 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
3964 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3965 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3966 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3967 | /* FORNOW */ |
3968 | i = -1; | |
3969 | break; | |
3970 | case SIMD_CLONE_ARG_TYPE_MASK: | |
3971 | gcc_unreachable (); | |
3972 | } | |
3973 | if (i == (size_t) -1) | |
3974 | break; | |
3975 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
3976 | { | |
3977 | i = -1; | |
3978 | break; | |
3979 | } | |
3980 | if (arginfo[i].align) | |
3981 | this_badness += (exact_log2 (arginfo[i].align) | |
3982 | - exact_log2 (n->simdclone->args[i].alignment)); | |
3983 | } | |
3984 | if (i == (size_t) -1) | |
3985 | continue; | |
3986 | if (bestn == NULL || this_badness < badness) | |
3987 | { | |
3988 | bestn = n; | |
3989 | badness = this_badness; | |
3990 | } | |
3991 | } | |
3992 | ||
3993 | if (bestn == NULL) | |
00426f9a | 3994 | return false; |
0136f8f0 AH |
3995 | |
3996 | for (i = 0; i < nargs; i++) | |
3997 | if ((arginfo[i].dt == vect_constant_def | |
3998 | || arginfo[i].dt == vect_external_def) | |
3999 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
4000 | { | |
4001 | arginfo[i].vectype | |
4002 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
4003 | i))); | |
4004 | if (arginfo[i].vectype == NULL | |
cf1b2ba4 | 4005 | || (simd_clone_subparts (arginfo[i].vectype) |
0136f8f0 | 4006 | > bestn->simdclone->simdlen)) |
00426f9a | 4007 | return false; |
0136f8f0 AH |
4008 | } |
4009 | ||
4010 | fndecl = bestn->decl; | |
4011 | nunits = bestn->simdclone->simdlen; | |
d9f21f6a | 4012 | ncopies = vf / nunits; |
0136f8f0 AH |
4013 | |
4014 | /* If the function isn't const, only allow it in simd loops where user | |
4015 | has asserted that at least nunits consecutive iterations can be | |
4016 | performed using SIMD instructions. */ | |
4017 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
4018 | && gimple_vuse (stmt)) | |
00426f9a | 4019 | return false; |
0136f8f0 AH |
4020 | |
4021 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
4022 | needs to be generated. */ | |
4023 | gcc_assert (ncopies >= 1); | |
4024 | ||
4025 | if (!vec_stmt) /* transformation not required. */ | |
4026 | { | |
6c9e85fb JJ |
4027 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
4028 | for (i = 0; i < nargs; i++) | |
7adb26f2 JJ |
4029 | if ((bestn->simdclone->args[i].arg_type |
4030 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
4031 | || (bestn->simdclone->args[i].arg_type | |
4032 | == SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP)) | |
6c9e85fb | 4033 | { |
17b658af | 4034 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
4035 | + 1); |
4036 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
4037 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
4038 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
4039 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
4040 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
4041 | tree sll = arginfo[i].simd_lane_linear |
4042 | ? boolean_true_node : boolean_false_node; | |
4043 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 4044 | } |
0136f8f0 | 4045 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
adac3a68 | 4046 | DUMP_VECT_SCOPE ("vectorizable_simd_clone_call"); |
68435eb2 | 4047 | /* vect_model_simple_cost (stmt_info, ncopies, dt, slp_node, cost_vec); */ |
0136f8f0 AH |
4048 | return true; |
4049 | } | |
4050 | ||
67b8dbac | 4051 | /* Transform. */ |
0136f8f0 AH |
4052 | |
4053 | if (dump_enabled_p ()) | |
4054 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
4055 | ||
4056 | /* Handle def. */ | |
4057 | scalar_dest = gimple_call_lhs (stmt); | |
4058 | vec_dest = NULL_TREE; | |
4059 | rtype = NULL_TREE; | |
4060 | ratype = NULL_TREE; | |
4061 | if (scalar_dest) | |
4062 | { | |
4063 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4064 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
4065 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
4066 | { | |
4067 | ratype = rtype; | |
4068 | rtype = TREE_TYPE (ratype); | |
4069 | } | |
4070 | } | |
4071 | ||
4072 | prev_stmt_info = NULL; | |
4073 | for (j = 0; j < ncopies; ++j) | |
4074 | { | |
4075 | /* Build argument list for the vectorized call. */ | |
4076 | if (j == 0) | |
4077 | vargs.create (nargs); | |
4078 | else | |
4079 | vargs.truncate (0); | |
4080 | ||
4081 | for (i = 0; i < nargs; i++) | |
4082 | { | |
4083 | unsigned int k, l, m, o; | |
4084 | tree atype; | |
4085 | op = gimple_call_arg (stmt, i); | |
4086 | switch (bestn->simdclone->args[i].arg_type) | |
4087 | { | |
4088 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
4089 | atype = bestn->simdclone->args[i].vector_type; | |
cf1b2ba4 | 4090 | o = nunits / simd_clone_subparts (atype); |
0136f8f0 AH |
4091 | for (m = j * o; m < (j + 1) * o; m++) |
4092 | { | |
cf1b2ba4 RS |
4093 | if (simd_clone_subparts (atype) |
4094 | < simd_clone_subparts (arginfo[i].vectype)) | |
0136f8f0 | 4095 | { |
73a699ae | 4096 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); |
cf1b2ba4 RS |
4097 | k = (simd_clone_subparts (arginfo[i].vectype) |
4098 | / simd_clone_subparts (atype)); | |
0136f8f0 AH |
4099 | gcc_assert ((k & (k - 1)) == 0); |
4100 | if (m == 0) | |
4101 | vec_oprnd0 | |
81c40241 | 4102 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
4103 | else |
4104 | { | |
4105 | vec_oprnd0 = arginfo[i].op; | |
4106 | if ((m & (k - 1)) == 0) | |
4107 | vec_oprnd0 | |
4108 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
4109 | vec_oprnd0); | |
4110 | } | |
4111 | arginfo[i].op = vec_oprnd0; | |
4112 | vec_oprnd0 | |
4113 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
92e29a5e | 4114 | bitsize_int (prec), |
0136f8f0 AH |
4115 | bitsize_int ((m & (k - 1)) * prec)); |
4116 | new_stmt | |
b731b390 | 4117 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
4118 | vec_oprnd0); |
4119 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4120 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
4121 | } | |
4122 | else | |
4123 | { | |
cf1b2ba4 RS |
4124 | k = (simd_clone_subparts (atype) |
4125 | / simd_clone_subparts (arginfo[i].vectype)); | |
0136f8f0 AH |
4126 | gcc_assert ((k & (k - 1)) == 0); |
4127 | vec<constructor_elt, va_gc> *ctor_elts; | |
4128 | if (k != 1) | |
4129 | vec_alloc (ctor_elts, k); | |
4130 | else | |
4131 | ctor_elts = NULL; | |
4132 | for (l = 0; l < k; l++) | |
4133 | { | |
4134 | if (m == 0 && l == 0) | |
4135 | vec_oprnd0 | |
81c40241 | 4136 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
4137 | else |
4138 | vec_oprnd0 | |
4139 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
4140 | arginfo[i].op); | |
4141 | arginfo[i].op = vec_oprnd0; | |
4142 | if (k == 1) | |
4143 | break; | |
4144 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
4145 | vec_oprnd0); | |
4146 | } | |
4147 | if (k == 1) | |
4148 | vargs.safe_push (vec_oprnd0); | |
4149 | else | |
4150 | { | |
4151 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
4152 | new_stmt | |
b731b390 | 4153 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
4154 | vec_oprnd0); |
4155 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4156 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
4157 | } | |
4158 | } | |
4159 | } | |
4160 | break; | |
4161 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
4162 | vargs.safe_push (op); | |
4163 | break; | |
4164 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
7adb26f2 | 4165 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
4166 | if (j == 0) |
4167 | { | |
4168 | gimple_seq stmts; | |
4169 | arginfo[i].op | |
4170 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
4171 | NULL_TREE); | |
4172 | if (stmts != NULL) | |
4173 | { | |
4174 | basic_block new_bb; | |
4175 | edge pe = loop_preheader_edge (loop); | |
4176 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
4177 | gcc_assert (!new_bb); | |
4178 | } | |
17b658af JJ |
4179 | if (arginfo[i].simd_lane_linear) |
4180 | { | |
4181 | vargs.safe_push (arginfo[i].op); | |
4182 | break; | |
4183 | } | |
b731b390 | 4184 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 4185 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
4fbeb363 | 4186 | loop_vinfo->add_stmt (new_phi); |
0136f8f0 AH |
4187 | add_phi_arg (new_phi, arginfo[i].op, |
4188 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
4189 | enum tree_code code | |
4190 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
4191 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
4192 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
4193 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
4194 | widest_int cst |
4195 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
4196 | ncopies * nunits); | |
4197 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 4198 | tree phi_arg = copy_ssa_name (op); |
0d0e4a03 JJ |
4199 | new_stmt |
4200 | = gimple_build_assign (phi_arg, code, phi_res, tcst); | |
0136f8f0 AH |
4201 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
4202 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
4fbeb363 | 4203 | loop_vinfo->add_stmt (new_stmt); |
0136f8f0 AH |
4204 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
4205 | UNKNOWN_LOCATION); | |
4206 | arginfo[i].op = phi_res; | |
4207 | vargs.safe_push (phi_res); | |
4208 | } | |
4209 | else | |
4210 | { | |
4211 | enum tree_code code | |
4212 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
4213 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
4214 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
4215 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
4216 | widest_int cst |
4217 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
4218 | j * nunits); | |
4219 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 4220 | new_temp = make_ssa_name (TREE_TYPE (op)); |
0d0e4a03 JJ |
4221 | new_stmt = gimple_build_assign (new_temp, code, |
4222 | arginfo[i].op, tcst); | |
0136f8f0 AH |
4223 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4224 | vargs.safe_push (new_temp); | |
4225 | } | |
4226 | break; | |
7adb26f2 JJ |
4227 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
4228 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
0136f8f0 | 4229 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
e01d41e5 JJ |
4230 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
4231 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
4232 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
4233 | default: |
4234 | gcc_unreachable (); | |
4235 | } | |
4236 | } | |
4237 | ||
4238 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
4239 | if (vec_dest) | |
4240 | { | |
cf1b2ba4 | 4241 | gcc_assert (ratype || simd_clone_subparts (rtype) == nunits); |
0136f8f0 | 4242 | if (ratype) |
b731b390 | 4243 | new_temp = create_tmp_var (ratype); |
cf1b2ba4 RS |
4244 | else if (simd_clone_subparts (vectype) |
4245 | == simd_clone_subparts (rtype)) | |
0136f8f0 AH |
4246 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4247 | else | |
4248 | new_temp = make_ssa_name (rtype, new_stmt); | |
4249 | gimple_call_set_lhs (new_stmt, new_temp); | |
4250 | } | |
4251 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4252 | ||
4253 | if (vec_dest) | |
4254 | { | |
cf1b2ba4 | 4255 | if (simd_clone_subparts (vectype) < nunits) |
0136f8f0 AH |
4256 | { |
4257 | unsigned int k, l; | |
73a699ae RS |
4258 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); |
4259 | poly_uint64 bytes = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
cf1b2ba4 | 4260 | k = nunits / simd_clone_subparts (vectype); |
0136f8f0 AH |
4261 | gcc_assert ((k & (k - 1)) == 0); |
4262 | for (l = 0; l < k; l++) | |
4263 | { | |
4264 | tree t; | |
4265 | if (ratype) | |
4266 | { | |
4267 | t = build_fold_addr_expr (new_temp); | |
4268 | t = build2 (MEM_REF, vectype, t, | |
73a699ae | 4269 | build_int_cst (TREE_TYPE (t), l * bytes)); |
0136f8f0 AH |
4270 | } |
4271 | else | |
4272 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
92e29a5e | 4273 | bitsize_int (prec), bitsize_int (l * prec)); |
0136f8f0 | 4274 | new_stmt |
b731b390 | 4275 | = gimple_build_assign (make_ssa_name (vectype), t); |
0136f8f0 AH |
4276 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4277 | if (j == 0 && l == 0) | |
4278 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4279 | else | |
4280 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4281 | ||
4282 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4283 | } | |
4284 | ||
4285 | if (ratype) | |
3ba4ff41 | 4286 | vect_clobber_variable (stmt, gsi, new_temp); |
0136f8f0 AH |
4287 | continue; |
4288 | } | |
cf1b2ba4 | 4289 | else if (simd_clone_subparts (vectype) > nunits) |
0136f8f0 | 4290 | { |
cf1b2ba4 RS |
4291 | unsigned int k = (simd_clone_subparts (vectype) |
4292 | / simd_clone_subparts (rtype)); | |
0136f8f0 AH |
4293 | gcc_assert ((k & (k - 1)) == 0); |
4294 | if ((j & (k - 1)) == 0) | |
4295 | vec_alloc (ret_ctor_elts, k); | |
4296 | if (ratype) | |
4297 | { | |
cf1b2ba4 | 4298 | unsigned int m, o = nunits / simd_clone_subparts (rtype); |
0136f8f0 AH |
4299 | for (m = 0; m < o; m++) |
4300 | { | |
4301 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
4302 | size_int (m), NULL_TREE, NULL_TREE); | |
4303 | new_stmt | |
b731b390 | 4304 | = gimple_build_assign (make_ssa_name (rtype), tem); |
0136f8f0 AH |
4305 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4306 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
4307 | gimple_assign_lhs (new_stmt)); | |
4308 | } | |
3ba4ff41 | 4309 | vect_clobber_variable (stmt, gsi, new_temp); |
0136f8f0 AH |
4310 | } |
4311 | else | |
4312 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
4313 | if ((j & (k - 1)) != k - 1) | |
4314 | continue; | |
4315 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
4316 | new_stmt | |
b731b390 | 4317 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
0136f8f0 AH |
4318 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4319 | ||
4320 | if ((unsigned) j == k - 1) | |
4321 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4322 | else | |
4323 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4324 | ||
4325 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4326 | continue; | |
4327 | } | |
4328 | else if (ratype) | |
4329 | { | |
4330 | tree t = build_fold_addr_expr (new_temp); | |
4331 | t = build2 (MEM_REF, vectype, t, | |
4332 | build_int_cst (TREE_TYPE (t), 0)); | |
4333 | new_stmt | |
b731b390 | 4334 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
0136f8f0 | 4335 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3ba4ff41 | 4336 | vect_clobber_variable (stmt, gsi, new_temp); |
0136f8f0 AH |
4337 | } |
4338 | } | |
4339 | ||
4340 | if (j == 0) | |
4341 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4342 | else | |
4343 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4344 | ||
4345 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4346 | } | |
4347 | ||
4348 | vargs.release (); | |
4349 | ||
4350 | /* The call in STMT might prevent it from being removed in dce. | |
4351 | We however cannot remove it here, due to the way the ssa name | |
4352 | it defines is mapped to the new definition. So just replace | |
4353 | rhs of the statement with something harmless. */ | |
4354 | ||
4355 | if (slp_node) | |
4356 | return true; | |
4357 | ||
4358 | if (scalar_dest) | |
4359 | { | |
4360 | type = TREE_TYPE (scalar_dest); | |
4361 | if (is_pattern_stmt_p (stmt_info)) | |
4362 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
4363 | else | |
4364 | lhs = gimple_call_lhs (stmt); | |
4365 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
4366 | } | |
4367 | else | |
4368 | new_stmt = gimple_build_nop (); | |
4369 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
4370 | set_vinfo_for_stmt (stmt, NULL); | |
4371 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 4372 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
4373 | unlink_stmt_vdef (stmt); |
4374 | ||
4375 | return true; | |
4376 | } | |
4377 | ||
4378 | ||
ebfd146a IR |
4379 | /* Function vect_gen_widened_results_half |
4380 | ||
4381 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 4382 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 4383 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
4384 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
4385 | needs to be created (DECL is a function-decl of a target-builtin). | |
4386 | STMT is the original scalar stmt that we are vectorizing. */ | |
4387 | ||
355fe088 | 4388 | static gimple * |
ebfd146a IR |
4389 | vect_gen_widened_results_half (enum tree_code code, |
4390 | tree decl, | |
4391 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
4392 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 4393 | gimple *stmt) |
b8698a0f | 4394 | { |
355fe088 | 4395 | gimple *new_stmt; |
b8698a0f L |
4396 | tree new_temp; |
4397 | ||
4398 | /* Generate half of the widened result: */ | |
4399 | if (code == CALL_EXPR) | |
4400 | { | |
4401 | /* Target specific support */ | |
ebfd146a IR |
4402 | if (op_type == binary_op) |
4403 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
4404 | else | |
4405 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
4406 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4407 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
4408 | } |
4409 | else | |
ebfd146a | 4410 | { |
b8698a0f L |
4411 | /* Generic support */ |
4412 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
4413 | if (op_type != binary_op) |
4414 | vec_oprnd1 = NULL; | |
0d0e4a03 | 4415 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
4416 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4417 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 4418 | } |
ebfd146a IR |
4419 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4420 | ||
ebfd146a IR |
4421 | return new_stmt; |
4422 | } | |
4423 | ||
4a00c761 JJ |
4424 | |
4425 | /* Get vectorized definitions for loop-based vectorization. For the first | |
4426 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
4427 | scalar operand), and for the rest we get a copy with | |
4428 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
4429 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
4430 | The vectors are collected into VEC_OPRNDS. */ | |
4431 | ||
4432 | static void | |
355fe088 | 4433 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 4434 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
4435 | { |
4436 | tree vec_oprnd; | |
4437 | ||
4438 | /* Get first vector operand. */ | |
4439 | /* All the vector operands except the very first one (that is scalar oprnd) | |
4440 | are stmt copies. */ | |
4441 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 4442 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
4443 | else |
4444 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
4445 | ||
9771b263 | 4446 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
4447 | |
4448 | /* Get second vector operand. */ | |
4449 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 4450 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
4451 | |
4452 | *oprnd = vec_oprnd; | |
4453 | ||
4454 | /* For conversion in multiple steps, continue to get operands | |
4455 | recursively. */ | |
4456 | if (multi_step_cvt) | |
4457 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
4458 | } | |
4459 | ||
4460 | ||
4461 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
4462 | For multi-step conversions store the resulting vectors and call the function | |
4463 | recursively. */ | |
4464 | ||
4465 | static void | |
9771b263 | 4466 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 4467 | int multi_step_cvt, gimple *stmt, |
9771b263 | 4468 | vec<tree> vec_dsts, |
4a00c761 JJ |
4469 | gimple_stmt_iterator *gsi, |
4470 | slp_tree slp_node, enum tree_code code, | |
4471 | stmt_vec_info *prev_stmt_info) | |
4472 | { | |
4473 | unsigned int i; | |
4474 | tree vop0, vop1, new_tmp, vec_dest; | |
355fe088 | 4475 | gimple *new_stmt; |
4a00c761 JJ |
4476 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
4477 | ||
9771b263 | 4478 | vec_dest = vec_dsts.pop (); |
4a00c761 | 4479 | |
9771b263 | 4480 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
4481 | { |
4482 | /* Create demotion operation. */ | |
9771b263 DN |
4483 | vop0 = (*vec_oprnds)[i]; |
4484 | vop1 = (*vec_oprnds)[i + 1]; | |
0d0e4a03 | 4485 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
4486 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
4487 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
4488 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4489 | ||
4490 | if (multi_step_cvt) | |
4491 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 4492 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
4493 | else |
4494 | { | |
4495 | /* This is the last step of the conversion sequence. Store the | |
4496 | vectors in SLP_NODE or in vector info of the scalar statement | |
4497 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
4498 | if (slp_node) | |
9771b263 | 4499 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4500 | else |
c689ce1e RB |
4501 | { |
4502 | if (!*prev_stmt_info) | |
4503 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4504 | else | |
4505 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
4a00c761 | 4506 | |
c689ce1e RB |
4507 | *prev_stmt_info = vinfo_for_stmt (new_stmt); |
4508 | } | |
4a00c761 JJ |
4509 | } |
4510 | } | |
4511 | ||
4512 | /* For multi-step demotion operations we first generate demotion operations | |
4513 | from the source type to the intermediate types, and then combine the | |
4514 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
4515 | type. */ | |
4516 | if (multi_step_cvt) | |
4517 | { | |
4518 | /* At each level of recursion we have half of the operands we had at the | |
4519 | previous level. */ | |
9771b263 | 4520 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
4521 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
4522 | stmt, vec_dsts, gsi, slp_node, | |
4523 | VEC_PACK_TRUNC_EXPR, | |
4524 | prev_stmt_info); | |
4525 | } | |
4526 | ||
9771b263 | 4527 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4528 | } |
4529 | ||
4530 | ||
4531 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
4532 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
4533 | the resulting vectors and call the function recursively. */ | |
4534 | ||
4535 | static void | |
9771b263 DN |
4536 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
4537 | vec<tree> *vec_oprnds1, | |
355fe088 | 4538 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
4539 | gimple_stmt_iterator *gsi, |
4540 | enum tree_code code1, | |
4541 | enum tree_code code2, tree decl1, | |
4542 | tree decl2, int op_type) | |
4543 | { | |
4544 | int i; | |
4545 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 4546 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 4547 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 4548 | |
9771b263 DN |
4549 | vec_tmp.create (vec_oprnds0->length () * 2); |
4550 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
4551 | { |
4552 | if (op_type == binary_op) | |
9771b263 | 4553 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
4554 | else |
4555 | vop1 = NULL_TREE; | |
4556 | ||
4557 | /* Generate the two halves of promotion operation. */ | |
4558 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
4559 | op_type, vec_dest, gsi, stmt); | |
4560 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
4561 | op_type, vec_dest, gsi, stmt); | |
4562 | if (is_gimple_call (new_stmt1)) | |
4563 | { | |
4564 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
4565 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
4566 | } | |
4567 | else | |
4568 | { | |
4569 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
4570 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
4571 | } | |
4572 | ||
4573 | /* Store the results for the next step. */ | |
9771b263 DN |
4574 | vec_tmp.quick_push (new_tmp1); |
4575 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
4576 | } |
4577 | ||
689eaba3 | 4578 | vec_oprnds0->release (); |
4a00c761 JJ |
4579 | *vec_oprnds0 = vec_tmp; |
4580 | } | |
4581 | ||
4582 | ||
b8698a0f L |
4583 | /* Check if STMT performs a conversion operation, that can be vectorized. |
4584 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 4585 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
4586 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
4587 | ||
4588 | static bool | |
355fe088 | 4589 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
68435eb2 RB |
4590 | gimple **vec_stmt, slp_tree slp_node, |
4591 | stmt_vector_for_cost *cost_vec) | |
ebfd146a IR |
4592 | { |
4593 | tree vec_dest; | |
4594 | tree scalar_dest; | |
4a00c761 | 4595 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
4596 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
4597 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4598 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4599 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 4600 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
4601 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
4602 | tree new_temp; | |
ebfd146a | 4603 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 4604 | int ndts = 2; |
355fe088 | 4605 | gimple *new_stmt = NULL; |
ebfd146a | 4606 | stmt_vec_info prev_stmt_info; |
062d5ccc RS |
4607 | poly_uint64 nunits_in; |
4608 | poly_uint64 nunits_out; | |
ebfd146a | 4609 | tree vectype_out, vectype_in; |
4a00c761 JJ |
4610 | int ncopies, i, j; |
4611 | tree lhs_type, rhs_type; | |
ebfd146a | 4612 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
4613 | vec<tree> vec_oprnds0 = vNULL; |
4614 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 4615 | tree vop0; |
4a00c761 | 4616 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4617 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 4618 | int multi_step_cvt = 0; |
6e1aa848 | 4619 | vec<tree> interm_types = vNULL; |
4a00c761 JJ |
4620 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
4621 | int op_type; | |
4a00c761 | 4622 | unsigned short fltsz; |
ebfd146a IR |
4623 | |
4624 | /* Is STMT a vectorizable conversion? */ | |
4625 | ||
4a00c761 | 4626 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4627 | return false; |
4628 | ||
66c16fd9 RB |
4629 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4630 | && ! vec_stmt) | |
ebfd146a IR |
4631 | return false; |
4632 | ||
4633 | if (!is_gimple_assign (stmt)) | |
4634 | return false; | |
4635 | ||
4636 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4637 | return false; | |
4638 | ||
4639 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
4640 | if (!CONVERT_EXPR_CODE_P (code) |
4641 | && code != FIX_TRUNC_EXPR | |
4642 | && code != FLOAT_EXPR | |
4643 | && code != WIDEN_MULT_EXPR | |
4644 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
4645 | return false; |
4646 | ||
4a00c761 JJ |
4647 | op_type = TREE_CODE_LENGTH (code); |
4648 | ||
ebfd146a | 4649 | /* Check types of lhs and rhs. */ |
b690cc0f | 4650 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 4651 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
4652 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
4653 | ||
ebfd146a IR |
4654 | op0 = gimple_assign_rhs1 (stmt); |
4655 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
4656 | |
4657 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4658 | && !((INTEGRAL_TYPE_P (lhs_type) | |
4659 | && INTEGRAL_TYPE_P (rhs_type)) | |
4660 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
4661 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
4662 | return false; | |
4663 | ||
e6f5c25d IE |
4664 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4665 | && ((INTEGRAL_TYPE_P (lhs_type) | |
2be65d9e | 4666 | && !type_has_mode_precision_p (lhs_type)) |
e6f5c25d | 4667 | || (INTEGRAL_TYPE_P (rhs_type) |
2be65d9e | 4668 | && !type_has_mode_precision_p (rhs_type)))) |
4a00c761 | 4669 | { |
73fbfcad | 4670 | if (dump_enabled_p ()) |
78c60e3d | 4671 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
4672 | "type conversion to/from bit-precision unsupported." |
4673 | "\n"); | |
4a00c761 JJ |
4674 | return false; |
4675 | } | |
4676 | ||
b690cc0f | 4677 | /* Check the operands of the operation. */ |
894dd753 | 4678 | if (!vect_is_simple_use (op0, vinfo, &dt[0], &vectype_in)) |
b690cc0f | 4679 | { |
73fbfcad | 4680 | if (dump_enabled_p ()) |
78c60e3d | 4681 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4682 | "use not simple.\n"); |
b690cc0f RG |
4683 | return false; |
4684 | } | |
4a00c761 JJ |
4685 | if (op_type == binary_op) |
4686 | { | |
4687 | bool ok; | |
4688 | ||
4689 | op1 = gimple_assign_rhs2 (stmt); | |
4690 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
4691 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
4692 | OP1. */ | |
4693 | if (CONSTANT_CLASS_P (op0)) | |
894dd753 | 4694 | ok = vect_is_simple_use (op1, vinfo, &dt[1], &vectype_in); |
4a00c761 | 4695 | else |
894dd753 | 4696 | ok = vect_is_simple_use (op1, vinfo, &dt[1]); |
4a00c761 JJ |
4697 | |
4698 | if (!ok) | |
4699 | { | |
73fbfcad | 4700 | if (dump_enabled_p ()) |
78c60e3d | 4701 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4702 | "use not simple.\n"); |
4a00c761 JJ |
4703 | return false; |
4704 | } | |
4705 | } | |
4706 | ||
b690cc0f RG |
4707 | /* If op0 is an external or constant defs use a vector type of |
4708 | the same size as the output vector type. */ | |
ebfd146a | 4709 | if (!vectype_in) |
b690cc0f | 4710 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
4711 | if (vec_stmt) |
4712 | gcc_assert (vectype_in); | |
4713 | if (!vectype_in) | |
4714 | { | |
73fbfcad | 4715 | if (dump_enabled_p ()) |
4a00c761 | 4716 | { |
78c60e3d SS |
4717 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4718 | "no vectype for scalar type "); | |
4719 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 4720 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 4721 | } |
7d8930a0 IR |
4722 | |
4723 | return false; | |
4724 | } | |
ebfd146a | 4725 | |
e6f5c25d IE |
4726 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4727 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
4728 | { | |
4729 | if (dump_enabled_p ()) | |
4730 | { | |
4731 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4732 | "can't convert between boolean and non " | |
4733 | "boolean vectors"); | |
4734 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
4735 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
4736 | } | |
4737 | ||
4738 | return false; | |
4739 | } | |
4740 | ||
b690cc0f RG |
4741 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
4742 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
062d5ccc | 4743 | if (known_eq (nunits_out, nunits_in)) |
ebfd146a | 4744 | modifier = NONE; |
062d5ccc RS |
4745 | else if (multiple_p (nunits_out, nunits_in)) |
4746 | modifier = NARROW; | |
ebfd146a | 4747 | else |
062d5ccc RS |
4748 | { |
4749 | gcc_checking_assert (multiple_p (nunits_in, nunits_out)); | |
4750 | modifier = WIDEN; | |
4751 | } | |
ebfd146a | 4752 | |
ff802fa1 IR |
4753 | /* Multiple types in SLP are handled by creating the appropriate number of |
4754 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4755 | case of SLP. */ | |
fce57248 | 4756 | if (slp_node) |
ebfd146a | 4757 | ncopies = 1; |
4a00c761 | 4758 | else if (modifier == NARROW) |
e8f142e2 | 4759 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
4a00c761 | 4760 | else |
e8f142e2 | 4761 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
b8698a0f | 4762 | |
ebfd146a IR |
4763 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
4764 | needs to be generated. */ | |
4765 | gcc_assert (ncopies >= 1); | |
4766 | ||
16d22000 RS |
4767 | bool found_mode = false; |
4768 | scalar_mode lhs_mode = SCALAR_TYPE_MODE (lhs_type); | |
4769 | scalar_mode rhs_mode = SCALAR_TYPE_MODE (rhs_type); | |
4770 | opt_scalar_mode rhs_mode_iter; | |
b397965c | 4771 | |
ebfd146a | 4772 | /* Supportable by target? */ |
4a00c761 | 4773 | switch (modifier) |
ebfd146a | 4774 | { |
4a00c761 JJ |
4775 | case NONE: |
4776 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4777 | return false; | |
4778 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
4779 | &decl1, &code1)) | |
4780 | break; | |
4781 | /* FALLTHRU */ | |
4782 | unsupported: | |
73fbfcad | 4783 | if (dump_enabled_p ()) |
78c60e3d | 4784 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4785 | "conversion not supported by target.\n"); |
ebfd146a | 4786 | return false; |
ebfd146a | 4787 | |
4a00c761 JJ |
4788 | case WIDEN: |
4789 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
4790 | &code1, &code2, &multi_step_cvt, |
4791 | &interm_types)) | |
4a00c761 JJ |
4792 | { |
4793 | /* Binary widening operation can only be supported directly by the | |
4794 | architecture. */ | |
4795 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
4796 | break; | |
4797 | } | |
4798 | ||
4799 | if (code != FLOAT_EXPR | |
b397965c | 4800 | || GET_MODE_SIZE (lhs_mode) <= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4801 | goto unsupported; |
4802 | ||
b397965c | 4803 | fltsz = GET_MODE_SIZE (lhs_mode); |
16d22000 | 4804 | FOR_EACH_2XWIDER_MODE (rhs_mode_iter, rhs_mode) |
4a00c761 | 4805 | { |
16d22000 | 4806 | rhs_mode = rhs_mode_iter.require (); |
c94843d2 RS |
4807 | if (GET_MODE_SIZE (rhs_mode) > fltsz) |
4808 | break; | |
4809 | ||
4a00c761 JJ |
4810 | cvt_type |
4811 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4812 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4813 | if (cvt_type == NULL_TREE) | |
4814 | goto unsupported; | |
4815 | ||
4816 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4817 | { | |
4818 | if (!supportable_convert_operation (code, vectype_out, | |
4819 | cvt_type, &decl1, &codecvt1)) | |
4820 | goto unsupported; | |
4821 | } | |
4822 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
4823 | cvt_type, &codecvt1, |
4824 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
4825 | &interm_types)) |
4826 | continue; | |
4827 | else | |
4828 | gcc_assert (multi_step_cvt == 0); | |
4829 | ||
4830 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
4831 | vectype_in, &code1, &code2, |
4832 | &multi_step_cvt, &interm_types)) | |
16d22000 RS |
4833 | { |
4834 | found_mode = true; | |
4835 | break; | |
4836 | } | |
4a00c761 JJ |
4837 | } |
4838 | ||
16d22000 | 4839 | if (!found_mode) |
4a00c761 JJ |
4840 | goto unsupported; |
4841 | ||
4842 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4843 | codecvt2 = ERROR_MARK; | |
4844 | else | |
4845 | { | |
4846 | multi_step_cvt++; | |
9771b263 | 4847 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
4848 | cvt_type = NULL_TREE; |
4849 | } | |
4850 | break; | |
4851 | ||
4852 | case NARROW: | |
4853 | gcc_assert (op_type == unary_op); | |
4854 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
4855 | &code1, &multi_step_cvt, | |
4856 | &interm_types)) | |
4857 | break; | |
4858 | ||
4859 | if (code != FIX_TRUNC_EXPR | |
b397965c | 4860 | || GET_MODE_SIZE (lhs_mode) >= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4861 | goto unsupported; |
4862 | ||
4a00c761 JJ |
4863 | cvt_type |
4864 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4865 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4866 | if (cvt_type == NULL_TREE) | |
4867 | goto unsupported; | |
4868 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
4869 | &decl1, &codecvt1)) | |
4870 | goto unsupported; | |
4871 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
4872 | &code1, &multi_step_cvt, | |
4873 | &interm_types)) | |
4874 | break; | |
4875 | goto unsupported; | |
4876 | ||
4877 | default: | |
4878 | gcc_unreachable (); | |
ebfd146a IR |
4879 | } |
4880 | ||
4881 | if (!vec_stmt) /* transformation not required. */ | |
4882 | { | |
adac3a68 | 4883 | DUMP_VECT_SCOPE ("vectorizable_conversion"); |
4a00c761 | 4884 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
4885 | { |
4886 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
68435eb2 RB |
4887 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, |
4888 | cost_vec); | |
8bd37302 | 4889 | } |
4a00c761 JJ |
4890 | else if (modifier == NARROW) |
4891 | { | |
4892 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
68435eb2 RB |
4893 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt, |
4894 | cost_vec); | |
4a00c761 JJ |
4895 | } |
4896 | else | |
4897 | { | |
4898 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
68435eb2 RB |
4899 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt, |
4900 | cost_vec); | |
4a00c761 | 4901 | } |
9771b263 | 4902 | interm_types.release (); |
ebfd146a IR |
4903 | return true; |
4904 | } | |
4905 | ||
67b8dbac | 4906 | /* Transform. */ |
73fbfcad | 4907 | if (dump_enabled_p ()) |
78c60e3d | 4908 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4909 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 4910 | |
4a00c761 JJ |
4911 | if (op_type == binary_op) |
4912 | { | |
4913 | if (CONSTANT_CLASS_P (op0)) | |
4914 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
4915 | else if (CONSTANT_CLASS_P (op1)) | |
4916 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
4917 | } | |
4918 | ||
4919 | /* In case of multi-step conversion, we first generate conversion operations | |
4920 | to the intermediate types, and then from that types to the final one. | |
4921 | We create vector destinations for the intermediate type (TYPES) received | |
4922 | from supportable_*_operation, and store them in the correct order | |
4923 | for future use in vect_create_vectorized_*_stmts (). */ | |
8c681247 | 4924 | auto_vec<tree> vec_dsts (multi_step_cvt + 1); |
82294ec1 JJ |
4925 | vec_dest = vect_create_destination_var (scalar_dest, |
4926 | (cvt_type && modifier == WIDEN) | |
4927 | ? cvt_type : vectype_out); | |
9771b263 | 4928 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4929 | |
4930 | if (multi_step_cvt) | |
4931 | { | |
9771b263 DN |
4932 | for (i = interm_types.length () - 1; |
4933 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
4934 | { |
4935 | vec_dest = vect_create_destination_var (scalar_dest, | |
4936 | intermediate_type); | |
9771b263 | 4937 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4938 | } |
4939 | } | |
ebfd146a | 4940 | |
4a00c761 | 4941 | if (cvt_type) |
82294ec1 JJ |
4942 | vec_dest = vect_create_destination_var (scalar_dest, |
4943 | modifier == WIDEN | |
4944 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
4945 | |
4946 | if (!slp_node) | |
4947 | { | |
30862efc | 4948 | if (modifier == WIDEN) |
4a00c761 | 4949 | { |
c3284718 | 4950 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 4951 | if (op_type == binary_op) |
9771b263 | 4952 | vec_oprnds1.create (1); |
4a00c761 | 4953 | } |
30862efc | 4954 | else if (modifier == NARROW) |
9771b263 DN |
4955 | vec_oprnds0.create ( |
4956 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
4957 | } |
4958 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 4959 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 4960 | |
4a00c761 | 4961 | last_oprnd = op0; |
ebfd146a IR |
4962 | prev_stmt_info = NULL; |
4963 | switch (modifier) | |
4964 | { | |
4965 | case NONE: | |
4966 | for (j = 0; j < ncopies; j++) | |
4967 | { | |
ebfd146a | 4968 | if (j == 0) |
306b0c92 | 4969 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node); |
ebfd146a IR |
4970 | else |
4971 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
4972 | ||
9771b263 | 4973 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4974 | { |
4975 | /* Arguments are ready, create the new vector stmt. */ | |
4976 | if (code1 == CALL_EXPR) | |
4977 | { | |
4978 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4979 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4980 | gimple_call_set_lhs (new_stmt, new_temp); | |
4981 | } | |
4982 | else | |
4983 | { | |
4984 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
0d0e4a03 | 4985 | new_stmt = gimple_build_assign (vec_dest, code1, vop0); |
4a00c761 JJ |
4986 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4987 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4988 | } | |
4989 | ||
4990 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4991 | if (slp_node) | |
9771b263 | 4992 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
225ce44b RB |
4993 | else |
4994 | { | |
4995 | if (!prev_stmt_info) | |
4996 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4997 | else | |
4998 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4999 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5000 | } | |
4a00c761 | 5001 | } |
ebfd146a IR |
5002 | } |
5003 | break; | |
5004 | ||
5005 | case WIDEN: | |
5006 | /* In case the vectorization factor (VF) is bigger than the number | |
5007 | of elements that we can fit in a vectype (nunits), we have to | |
5008 | generate more than one vector stmt - i.e - we need to "unroll" | |
5009 | the vector stmt by a factor VF/nunits. */ | |
5010 | for (j = 0; j < ncopies; j++) | |
5011 | { | |
4a00c761 | 5012 | /* Handle uses. */ |
ebfd146a | 5013 | if (j == 0) |
4a00c761 JJ |
5014 | { |
5015 | if (slp_node) | |
5016 | { | |
5017 | if (code == WIDEN_LSHIFT_EXPR) | |
5018 | { | |
5019 | unsigned int k; | |
ebfd146a | 5020 | |
4a00c761 JJ |
5021 | vec_oprnd1 = op1; |
5022 | /* Store vec_oprnd1 for every vector stmt to be created | |
5023 | for SLP_NODE. We check during the analysis that all | |
5024 | the shift arguments are the same. */ | |
5025 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 5026 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
5027 | |
5028 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5029 | slp_node); |
4a00c761 JJ |
5030 | } |
5031 | else | |
5032 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
306b0c92 | 5033 | &vec_oprnds1, slp_node); |
4a00c761 JJ |
5034 | } |
5035 | else | |
5036 | { | |
81c40241 | 5037 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 5038 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
5039 | if (op_type == binary_op) |
5040 | { | |
5041 | if (code == WIDEN_LSHIFT_EXPR) | |
5042 | vec_oprnd1 = op1; | |
5043 | else | |
81c40241 | 5044 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 5045 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
5046 | } |
5047 | } | |
5048 | } | |
ebfd146a | 5049 | else |
4a00c761 JJ |
5050 | { |
5051 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
5052 | vec_oprnds0.truncate (0); |
5053 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
5054 | if (op_type == binary_op) |
5055 | { | |
5056 | if (code == WIDEN_LSHIFT_EXPR) | |
5057 | vec_oprnd1 = op1; | |
5058 | else | |
5059 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
5060 | vec_oprnd1); | |
9771b263 DN |
5061 | vec_oprnds1.truncate (0); |
5062 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
5063 | } |
5064 | } | |
ebfd146a | 5065 | |
4a00c761 JJ |
5066 | /* Arguments are ready. Create the new vector stmts. */ |
5067 | for (i = multi_step_cvt; i >= 0; i--) | |
5068 | { | |
9771b263 | 5069 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
5070 | enum tree_code c1 = code1, c2 = code2; |
5071 | if (i == 0 && codecvt2 != ERROR_MARK) | |
5072 | { | |
5073 | c1 = codecvt1; | |
5074 | c2 = codecvt2; | |
5075 | } | |
5076 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
5077 | &vec_oprnds1, | |
5078 | stmt, this_dest, gsi, | |
5079 | c1, c2, decl1, decl2, | |
5080 | op_type); | |
5081 | } | |
5082 | ||
9771b263 | 5083 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
5084 | { |
5085 | if (cvt_type) | |
5086 | { | |
5087 | if (codecvt1 == CALL_EXPR) | |
5088 | { | |
5089 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
5090 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5091 | gimple_call_set_lhs (new_stmt, new_temp); | |
5092 | } | |
5093 | else | |
5094 | { | |
5095 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 5096 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
5097 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
5098 | vop0); | |
4a00c761 JJ |
5099 | } |
5100 | ||
5101 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5102 | } | |
5103 | else | |
5104 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
5105 | ||
5106 | if (slp_node) | |
9771b263 | 5107 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 5108 | else |
c689ce1e RB |
5109 | { |
5110 | if (!prev_stmt_info) | |
5111 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
5112 | else | |
5113 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5114 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5115 | } | |
4a00c761 | 5116 | } |
ebfd146a | 5117 | } |
4a00c761 JJ |
5118 | |
5119 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
5120 | break; |
5121 | ||
5122 | case NARROW: | |
5123 | /* In case the vectorization factor (VF) is bigger than the number | |
5124 | of elements that we can fit in a vectype (nunits), we have to | |
5125 | generate more than one vector stmt - i.e - we need to "unroll" | |
5126 | the vector stmt by a factor VF/nunits. */ | |
5127 | for (j = 0; j < ncopies; j++) | |
5128 | { | |
5129 | /* Handle uses. */ | |
4a00c761 JJ |
5130 | if (slp_node) |
5131 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5132 | slp_node); |
ebfd146a IR |
5133 | else |
5134 | { | |
9771b263 | 5135 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
5136 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
5137 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
5138 | } |
5139 | ||
4a00c761 JJ |
5140 | /* Arguments are ready. Create the new vector stmts. */ |
5141 | if (cvt_type) | |
9771b263 | 5142 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
5143 | { |
5144 | if (codecvt1 == CALL_EXPR) | |
5145 | { | |
5146 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
5147 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5148 | gimple_call_set_lhs (new_stmt, new_temp); | |
5149 | } | |
5150 | else | |
5151 | { | |
5152 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 5153 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
5154 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
5155 | vop0); | |
4a00c761 | 5156 | } |
ebfd146a | 5157 | |
4a00c761 | 5158 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 5159 | vec_oprnds0[i] = new_temp; |
4a00c761 | 5160 | } |
ebfd146a | 5161 | |
4a00c761 JJ |
5162 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
5163 | stmt, vec_dsts, gsi, | |
5164 | slp_node, code1, | |
5165 | &prev_stmt_info); | |
ebfd146a IR |
5166 | } |
5167 | ||
5168 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 5169 | break; |
ebfd146a IR |
5170 | } |
5171 | ||
9771b263 DN |
5172 | vec_oprnds0.release (); |
5173 | vec_oprnds1.release (); | |
9771b263 | 5174 | interm_types.release (); |
ebfd146a IR |
5175 | |
5176 | return true; | |
5177 | } | |
ff802fa1 IR |
5178 | |
5179 | ||
ebfd146a IR |
5180 | /* Function vectorizable_assignment. |
5181 | ||
b8698a0f L |
5182 | Check if STMT performs an assignment (copy) that can be vectorized. |
5183 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5184 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5185 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5186 | ||
5187 | static bool | |
355fe088 | 5188 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
68435eb2 RB |
5189 | gimple **vec_stmt, slp_tree slp_node, |
5190 | stmt_vector_for_cost *cost_vec) | |
ebfd146a IR |
5191 | { |
5192 | tree vec_dest; | |
5193 | tree scalar_dest; | |
5194 | tree op; | |
5195 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
5196 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
5197 | tree new_temp; | |
4fc5ebf1 JG |
5198 | enum vect_def_type dt[1] = {vect_unknown_def_type}; |
5199 | int ndts = 1; | |
ebfd146a | 5200 | int ncopies; |
f18b55bd | 5201 | int i, j; |
6e1aa848 | 5202 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5203 | tree vop; |
a70d6342 | 5204 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5205 | vec_info *vinfo = stmt_info->vinfo; |
355fe088 | 5206 | gimple *new_stmt = NULL; |
f18b55bd | 5207 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
5208 | enum tree_code code; |
5209 | tree vectype_in; | |
ebfd146a | 5210 | |
a70d6342 | 5211 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5212 | return false; |
5213 | ||
66c16fd9 RB |
5214 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5215 | && ! vec_stmt) | |
ebfd146a IR |
5216 | return false; |
5217 | ||
5218 | /* Is vectorizable assignment? */ | |
5219 | if (!is_gimple_assign (stmt)) | |
5220 | return false; | |
5221 | ||
5222 | scalar_dest = gimple_assign_lhs (stmt); | |
5223 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
5224 | return false; | |
5225 | ||
fde9c428 | 5226 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 5227 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
5228 | || code == PAREN_EXPR |
5229 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
5230 | op = gimple_assign_rhs1 (stmt); |
5231 | else | |
5232 | return false; | |
5233 | ||
7b7ec6c5 RG |
5234 | if (code == VIEW_CONVERT_EXPR) |
5235 | op = TREE_OPERAND (op, 0); | |
5236 | ||
465c8c19 | 5237 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
928686b1 | 5238 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 JJ |
5239 | |
5240 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5241 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5242 | case of SLP. */ | |
fce57248 | 5243 | if (slp_node) |
465c8c19 JJ |
5244 | ncopies = 1; |
5245 | else | |
e8f142e2 | 5246 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
5247 | |
5248 | gcc_assert (ncopies >= 1); | |
5249 | ||
894dd753 | 5250 | if (!vect_is_simple_use (op, vinfo, &dt[0], &vectype_in)) |
ebfd146a | 5251 | { |
73fbfcad | 5252 | if (dump_enabled_p ()) |
78c60e3d | 5253 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5254 | "use not simple.\n"); |
ebfd146a IR |
5255 | return false; |
5256 | } | |
5257 | ||
fde9c428 RG |
5258 | /* We can handle NOP_EXPR conversions that do not change the number |
5259 | of elements or the vector size. */ | |
7b7ec6c5 RG |
5260 | if ((CONVERT_EXPR_CODE_P (code) |
5261 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 | 5262 | && (!vectype_in |
928686b1 | 5263 | || maybe_ne (TYPE_VECTOR_SUBPARTS (vectype_in), nunits) |
cf098191 RS |
5264 | || maybe_ne (GET_MODE_SIZE (TYPE_MODE (vectype)), |
5265 | GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
fde9c428 RG |
5266 | return false; |
5267 | ||
7b7b1813 RG |
5268 | /* We do not handle bit-precision changes. */ |
5269 | if ((CONVERT_EXPR_CODE_P (code) | |
5270 | || code == VIEW_CONVERT_EXPR) | |
5271 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2be65d9e RS |
5272 | && (!type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
5273 | || !type_has_mode_precision_p (TREE_TYPE (op))) | |
7b7b1813 RG |
5274 | /* But a conversion that does not change the bit-pattern is ok. */ |
5275 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
5276 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2dab46d5 IE |
5277 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
5278 | /* Conversion between boolean types of different sizes is | |
5279 | a simple assignment in case their vectypes are same | |
5280 | boolean vectors. */ | |
5281 | && (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
5282 | || !VECTOR_BOOLEAN_TYPE_P (vectype_in))) | |
7b7b1813 | 5283 | { |
73fbfcad | 5284 | if (dump_enabled_p ()) |
78c60e3d SS |
5285 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5286 | "type conversion to/from bit-precision " | |
e645e942 | 5287 | "unsupported.\n"); |
7b7b1813 RG |
5288 | return false; |
5289 | } | |
5290 | ||
ebfd146a IR |
5291 | if (!vec_stmt) /* transformation not required. */ |
5292 | { | |
5293 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
adac3a68 | 5294 | DUMP_VECT_SCOPE ("vectorizable_assignment"); |
68435eb2 | 5295 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
ebfd146a IR |
5296 | return true; |
5297 | } | |
5298 | ||
67b8dbac | 5299 | /* Transform. */ |
73fbfcad | 5300 | if (dump_enabled_p ()) |
e645e942 | 5301 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
5302 | |
5303 | /* Handle def. */ | |
5304 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5305 | ||
5306 | /* Handle use. */ | |
f18b55bd | 5307 | for (j = 0; j < ncopies; j++) |
ebfd146a | 5308 | { |
f18b55bd IR |
5309 | /* Handle uses. */ |
5310 | if (j == 0) | |
306b0c92 | 5311 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
f18b55bd IR |
5312 | else |
5313 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
5314 | ||
5315 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 5316 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 5317 | { |
7b7ec6c5 RG |
5318 | if (CONVERT_EXPR_CODE_P (code) |
5319 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 5320 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
5321 | new_stmt = gimple_build_assign (vec_dest, vop); |
5322 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5323 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5324 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5325 | if (slp_node) | |
9771b263 | 5326 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 5327 | } |
ebfd146a IR |
5328 | |
5329 | if (slp_node) | |
f18b55bd IR |
5330 | continue; |
5331 | ||
5332 | if (j == 0) | |
5333 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5334 | else | |
5335 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5336 | ||
5337 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5338 | } | |
b8698a0f | 5339 | |
9771b263 | 5340 | vec_oprnds.release (); |
ebfd146a IR |
5341 | return true; |
5342 | } | |
5343 | ||
9dc3f7de | 5344 | |
1107f3ae IR |
5345 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
5346 | either as shift by a scalar or by a vector. */ | |
5347 | ||
5348 | bool | |
5349 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
5350 | { | |
5351 | ||
ef4bddc2 | 5352 | machine_mode vec_mode; |
1107f3ae IR |
5353 | optab optab; |
5354 | int icode; | |
5355 | tree vectype; | |
5356 | ||
5357 | vectype = get_vectype_for_scalar_type (scalar_type); | |
5358 | if (!vectype) | |
5359 | return false; | |
5360 | ||
5361 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5362 | if (!optab | |
5363 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
5364 | { | |
5365 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5366 | if (!optab | |
5367 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
5368 | == CODE_FOR_nothing)) | |
5369 | return false; | |
5370 | } | |
5371 | ||
5372 | vec_mode = TYPE_MODE (vectype); | |
5373 | icode = (int) optab_handler (optab, vec_mode); | |
5374 | if (icode == CODE_FOR_nothing) | |
5375 | return false; | |
5376 | ||
5377 | return true; | |
5378 | } | |
5379 | ||
5380 | ||
9dc3f7de IR |
5381 | /* Function vectorizable_shift. |
5382 | ||
5383 | Check if STMT performs a shift operation that can be vectorized. | |
5384 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
5385 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
5386 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5387 | ||
5388 | static bool | |
355fe088 | 5389 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
68435eb2 RB |
5390 | gimple **vec_stmt, slp_tree slp_node, |
5391 | stmt_vector_for_cost *cost_vec) | |
9dc3f7de IR |
5392 | { |
5393 | tree vec_dest; | |
5394 | tree scalar_dest; | |
5395 | tree op0, op1 = NULL; | |
5396 | tree vec_oprnd1 = NULL_TREE; | |
5397 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5398 | tree vectype; | |
5399 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
5400 | enum tree_code code; | |
ef4bddc2 | 5401 | machine_mode vec_mode; |
9dc3f7de IR |
5402 | tree new_temp; |
5403 | optab optab; | |
5404 | int icode; | |
ef4bddc2 | 5405 | machine_mode optab_op2_mode; |
9dc3f7de | 5406 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 5407 | int ndts = 2; |
355fe088 | 5408 | gimple *new_stmt = NULL; |
9dc3f7de | 5409 | stmt_vec_info prev_stmt_info; |
928686b1 RS |
5410 | poly_uint64 nunits_in; |
5411 | poly_uint64 nunits_out; | |
9dc3f7de | 5412 | tree vectype_out; |
cede2577 | 5413 | tree op1_vectype; |
9dc3f7de IR |
5414 | int ncopies; |
5415 | int j, i; | |
6e1aa848 DN |
5416 | vec<tree> vec_oprnds0 = vNULL; |
5417 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
5418 | tree vop0, vop1; |
5419 | unsigned int k; | |
49eab32e | 5420 | bool scalar_shift_arg = true; |
9dc3f7de | 5421 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5422 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
5423 | |
5424 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
5425 | return false; | |
5426 | ||
66c16fd9 RB |
5427 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5428 | && ! vec_stmt) | |
9dc3f7de IR |
5429 | return false; |
5430 | ||
5431 | /* Is STMT a vectorizable binary/unary operation? */ | |
5432 | if (!is_gimple_assign (stmt)) | |
5433 | return false; | |
5434 | ||
5435 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5436 | return false; | |
5437 | ||
5438 | code = gimple_assign_rhs_code (stmt); | |
5439 | ||
5440 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
5441 | || code == RROTATE_EXPR)) | |
5442 | return false; | |
5443 | ||
5444 | scalar_dest = gimple_assign_lhs (stmt); | |
5445 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2be65d9e | 5446 | if (!type_has_mode_precision_p (TREE_TYPE (scalar_dest))) |
7b7b1813 | 5447 | { |
73fbfcad | 5448 | if (dump_enabled_p ()) |
78c60e3d | 5449 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5450 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
5451 | return false; |
5452 | } | |
9dc3f7de IR |
5453 | |
5454 | op0 = gimple_assign_rhs1 (stmt); | |
894dd753 | 5455 | if (!vect_is_simple_use (op0, vinfo, &dt[0], &vectype)) |
9dc3f7de | 5456 | { |
73fbfcad | 5457 | if (dump_enabled_p ()) |
78c60e3d | 5458 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5459 | "use not simple.\n"); |
9dc3f7de IR |
5460 | return false; |
5461 | } | |
5462 | /* If op0 is an external or constant def use a vector type with | |
5463 | the same size as the output vector type. */ | |
5464 | if (!vectype) | |
5465 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5466 | if (vec_stmt) | |
5467 | gcc_assert (vectype); | |
5468 | if (!vectype) | |
5469 | { | |
73fbfcad | 5470 | if (dump_enabled_p ()) |
78c60e3d | 5471 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5472 | "no vectype for scalar type\n"); |
9dc3f7de IR |
5473 | return false; |
5474 | } | |
5475 | ||
5476 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5477 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
928686b1 | 5478 | if (maybe_ne (nunits_out, nunits_in)) |
9dc3f7de IR |
5479 | return false; |
5480 | ||
5481 | op1 = gimple_assign_rhs2 (stmt); | |
894dd753 | 5482 | if (!vect_is_simple_use (op1, vinfo, &dt[1], &op1_vectype)) |
9dc3f7de | 5483 | { |
73fbfcad | 5484 | if (dump_enabled_p ()) |
78c60e3d | 5485 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5486 | "use not simple.\n"); |
9dc3f7de IR |
5487 | return false; |
5488 | } | |
5489 | ||
9dc3f7de IR |
5490 | /* Multiple types in SLP are handled by creating the appropriate number of |
5491 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5492 | case of SLP. */ | |
fce57248 | 5493 | if (slp_node) |
9dc3f7de IR |
5494 | ncopies = 1; |
5495 | else | |
e8f142e2 | 5496 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
9dc3f7de IR |
5497 | |
5498 | gcc_assert (ncopies >= 1); | |
5499 | ||
5500 | /* Determine whether the shift amount is a vector, or scalar. If the | |
5501 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
5502 | ||
dbfa87aa YR |
5503 | if ((dt[1] == vect_internal_def |
5504 | || dt[1] == vect_induction_def) | |
5505 | && !slp_node) | |
49eab32e JJ |
5506 | scalar_shift_arg = false; |
5507 | else if (dt[1] == vect_constant_def | |
5508 | || dt[1] == vect_external_def | |
5509 | || dt[1] == vect_internal_def) | |
5510 | { | |
5511 | /* In SLP, need to check whether the shift count is the same, | |
5512 | in loops if it is a constant or invariant, it is always | |
5513 | a scalar shift. */ | |
5514 | if (slp_node) | |
5515 | { | |
355fe088 TS |
5516 | vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
5517 | gimple *slpstmt; | |
49eab32e | 5518 | |
9771b263 | 5519 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
5520 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
5521 | scalar_shift_arg = false; | |
5522 | } | |
60d393e8 RB |
5523 | |
5524 | /* If the shift amount is computed by a pattern stmt we cannot | |
5525 | use the scalar amount directly thus give up and use a vector | |
5526 | shift. */ | |
5527 | if (dt[1] == vect_internal_def) | |
5528 | { | |
5529 | gimple *def = SSA_NAME_DEF_STMT (op1); | |
5530 | if (is_pattern_stmt_p (vinfo_for_stmt (def))) | |
5531 | scalar_shift_arg = false; | |
5532 | } | |
49eab32e JJ |
5533 | } |
5534 | else | |
5535 | { | |
73fbfcad | 5536 | if (dump_enabled_p ()) |
78c60e3d | 5537 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5538 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
5539 | return false; |
5540 | } | |
5541 | ||
9dc3f7de | 5542 | /* Vector shifted by vector. */ |
49eab32e | 5543 | if (!scalar_shift_arg) |
9dc3f7de IR |
5544 | { |
5545 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 5546 | if (dump_enabled_p ()) |
78c60e3d | 5547 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5548 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 5549 | |
aa948027 JJ |
5550 | if (!op1_vectype) |
5551 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
5552 | if (op1_vectype == NULL_TREE | |
5553 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 5554 | { |
73fbfcad | 5555 | if (dump_enabled_p ()) |
78c60e3d SS |
5556 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5557 | "unusable type for last operand in" | |
e645e942 | 5558 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
5559 | return false; |
5560 | } | |
9dc3f7de IR |
5561 | } |
5562 | /* See if the machine has a vector shifted by scalar insn and if not | |
5563 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 5564 | else |
9dc3f7de IR |
5565 | { |
5566 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5567 | if (optab | |
5568 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
5569 | { | |
73fbfcad | 5570 | if (dump_enabled_p ()) |
78c60e3d | 5571 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5572 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
5573 | } |
5574 | else | |
5575 | { | |
5576 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5577 | if (optab | |
5578 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
5579 | != CODE_FOR_nothing)) | |
5580 | { | |
49eab32e JJ |
5581 | scalar_shift_arg = false; |
5582 | ||
73fbfcad | 5583 | if (dump_enabled_p ()) |
78c60e3d | 5584 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5585 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
5586 | |
5587 | /* Unlike the other binary operators, shifts/rotates have | |
5588 | the rhs being int, instead of the same type as the lhs, | |
5589 | so make sure the scalar is the right type if we are | |
aa948027 | 5590 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
5591 | if (dt[1] == vect_constant_def) |
5592 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
5593 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
5594 | TREE_TYPE (op1))) | |
5595 | { | |
5596 | if (slp_node | |
5597 | && TYPE_MODE (TREE_TYPE (vectype)) | |
5598 | != TYPE_MODE (TREE_TYPE (op1))) | |
5599 | { | |
73fbfcad | 5600 | if (dump_enabled_p ()) |
78c60e3d SS |
5601 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5602 | "unusable type for last operand in" | |
e645e942 | 5603 | " vector/vector shift/rotate.\n"); |
21c0a521 | 5604 | return false; |
aa948027 JJ |
5605 | } |
5606 | if (vec_stmt && !slp_node) | |
5607 | { | |
5608 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
5609 | op1 = vect_init_vector (stmt, op1, | |
5610 | TREE_TYPE (vectype), NULL); | |
5611 | } | |
5612 | } | |
9dc3f7de IR |
5613 | } |
5614 | } | |
5615 | } | |
9dc3f7de IR |
5616 | |
5617 | /* Supportable by target? */ | |
5618 | if (!optab) | |
5619 | { | |
73fbfcad | 5620 | if (dump_enabled_p ()) |
78c60e3d | 5621 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5622 | "no optab.\n"); |
9dc3f7de IR |
5623 | return false; |
5624 | } | |
5625 | vec_mode = TYPE_MODE (vectype); | |
5626 | icode = (int) optab_handler (optab, vec_mode); | |
5627 | if (icode == CODE_FOR_nothing) | |
5628 | { | |
73fbfcad | 5629 | if (dump_enabled_p ()) |
78c60e3d | 5630 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5631 | "op not supported by target.\n"); |
9dc3f7de | 5632 | /* Check only during analysis. */ |
cf098191 | 5633 | if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) |
ca09abcb RS |
5634 | || (!vec_stmt |
5635 | && !vect_worthwhile_without_simd_p (vinfo, code))) | |
9dc3f7de | 5636 | return false; |
73fbfcad | 5637 | if (dump_enabled_p ()) |
e645e942 TJ |
5638 | dump_printf_loc (MSG_NOTE, vect_location, |
5639 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
5640 | } |
5641 | ||
5642 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
ca09abcb RS |
5643 | if (!vec_stmt |
5644 | && !VECTOR_MODE_P (TYPE_MODE (vectype)) | |
5645 | && !vect_worthwhile_without_simd_p (vinfo, code)) | |
9dc3f7de | 5646 | { |
73fbfcad | 5647 | if (dump_enabled_p ()) |
78c60e3d | 5648 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5649 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
5650 | return false; |
5651 | } | |
5652 | ||
5653 | if (!vec_stmt) /* transformation not required. */ | |
5654 | { | |
5655 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
adac3a68 | 5656 | DUMP_VECT_SCOPE ("vectorizable_shift"); |
68435eb2 | 5657 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
9dc3f7de IR |
5658 | return true; |
5659 | } | |
5660 | ||
67b8dbac | 5661 | /* Transform. */ |
9dc3f7de | 5662 | |
73fbfcad | 5663 | if (dump_enabled_p ()) |
78c60e3d | 5664 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5665 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
5666 | |
5667 | /* Handle def. */ | |
5668 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5669 | ||
9dc3f7de IR |
5670 | prev_stmt_info = NULL; |
5671 | for (j = 0; j < ncopies; j++) | |
5672 | { | |
5673 | /* Handle uses. */ | |
5674 | if (j == 0) | |
5675 | { | |
5676 | if (scalar_shift_arg) | |
5677 | { | |
5678 | /* Vector shl and shr insn patterns can be defined with scalar | |
5679 | operand 2 (shift operand). In this case, use constant or loop | |
5680 | invariant op1 directly, without extending it to vector mode | |
5681 | first. */ | |
5682 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
5683 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
5684 | { | |
73fbfcad | 5685 | if (dump_enabled_p ()) |
78c60e3d | 5686 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5687 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 5688 | vec_oprnd1 = op1; |
8930f723 | 5689 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 5690 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5691 | if (slp_node) |
5692 | { | |
5693 | /* Store vec_oprnd1 for every vector stmt to be created | |
5694 | for SLP_NODE. We check during the analysis that all | |
5695 | the shift arguments are the same. | |
5696 | TODO: Allow different constants for different vector | |
5697 | stmts generated for an SLP instance. */ | |
5698 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 5699 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5700 | } |
5701 | } | |
5702 | } | |
5703 | ||
5704 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
5705 | (a special case for certain kind of vector shifts); otherwise, | |
5706 | operand 1 should be of a vector type (the usual case). */ | |
5707 | if (vec_oprnd1) | |
5708 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5709 | slp_node); |
9dc3f7de IR |
5710 | else |
5711 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
306b0c92 | 5712 | slp_node); |
9dc3f7de IR |
5713 | } |
5714 | else | |
5715 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5716 | ||
5717 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5718 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 5719 | { |
9771b263 | 5720 | vop1 = vec_oprnds1[i]; |
0d0e4a03 | 5721 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
5722 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5723 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5724 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5725 | if (slp_node) | |
9771b263 | 5726 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
5727 | } |
5728 | ||
5729 | if (slp_node) | |
5730 | continue; | |
5731 | ||
5732 | if (j == 0) | |
5733 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5734 | else | |
5735 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5736 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5737 | } | |
5738 | ||
9771b263 DN |
5739 | vec_oprnds0.release (); |
5740 | vec_oprnds1.release (); | |
9dc3f7de IR |
5741 | |
5742 | return true; | |
5743 | } | |
5744 | ||
5745 | ||
ebfd146a IR |
5746 | /* Function vectorizable_operation. |
5747 | ||
16949072 RG |
5748 | Check if STMT performs a binary, unary or ternary operation that can |
5749 | be vectorized. | |
b8698a0f | 5750 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
5751 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5752 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5753 | ||
5754 | static bool | |
355fe088 | 5755 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
68435eb2 RB |
5756 | gimple **vec_stmt, slp_tree slp_node, |
5757 | stmt_vector_for_cost *cost_vec) | |
ebfd146a | 5758 | { |
00f07b86 | 5759 | tree vec_dest; |
ebfd146a | 5760 | tree scalar_dest; |
16949072 | 5761 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 5762 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 5763 | tree vectype; |
ebfd146a | 5764 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
0eb952ea | 5765 | enum tree_code code, orig_code; |
ef4bddc2 | 5766 | machine_mode vec_mode; |
ebfd146a IR |
5767 | tree new_temp; |
5768 | int op_type; | |
00f07b86 | 5769 | optab optab; |
523ba738 | 5770 | bool target_support_p; |
16949072 RG |
5771 | enum vect_def_type dt[3] |
5772 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 5773 | int ndts = 3; |
355fe088 | 5774 | gimple *new_stmt = NULL; |
ebfd146a | 5775 | stmt_vec_info prev_stmt_info; |
928686b1 RS |
5776 | poly_uint64 nunits_in; |
5777 | poly_uint64 nunits_out; | |
ebfd146a IR |
5778 | tree vectype_out; |
5779 | int ncopies; | |
5780 | int j, i; | |
6e1aa848 DN |
5781 | vec<tree> vec_oprnds0 = vNULL; |
5782 | vec<tree> vec_oprnds1 = vNULL; | |
5783 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 5784 | tree vop0, vop1, vop2; |
a70d6342 | 5785 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5786 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 5787 | |
a70d6342 | 5788 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5789 | return false; |
5790 | ||
66c16fd9 RB |
5791 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5792 | && ! vec_stmt) | |
ebfd146a IR |
5793 | return false; |
5794 | ||
5795 | /* Is STMT a vectorizable binary/unary operation? */ | |
5796 | if (!is_gimple_assign (stmt)) | |
5797 | return false; | |
5798 | ||
5799 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5800 | return false; | |
5801 | ||
0eb952ea | 5802 | orig_code = code = gimple_assign_rhs_code (stmt); |
ebfd146a | 5803 | |
1af4ebf5 MG |
5804 | /* For pointer addition and subtraction, we should use the normal |
5805 | plus and minus for the vector operation. */ | |
ebfd146a IR |
5806 | if (code == POINTER_PLUS_EXPR) |
5807 | code = PLUS_EXPR; | |
1af4ebf5 MG |
5808 | if (code == POINTER_DIFF_EXPR) |
5809 | code = MINUS_EXPR; | |
ebfd146a IR |
5810 | |
5811 | /* Support only unary or binary operations. */ | |
5812 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 5813 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 5814 | { |
73fbfcad | 5815 | if (dump_enabled_p ()) |
78c60e3d | 5816 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5817 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 5818 | op_type); |
ebfd146a IR |
5819 | return false; |
5820 | } | |
5821 | ||
b690cc0f RG |
5822 | scalar_dest = gimple_assign_lhs (stmt); |
5823 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
5824 | ||
7b7b1813 RG |
5825 | /* Most operations cannot handle bit-precision types without extra |
5826 | truncations. */ | |
045c1278 | 5827 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
2be65d9e | 5828 | && !type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
7b7b1813 RG |
5829 | /* Exception are bitwise binary operations. */ |
5830 | && code != BIT_IOR_EXPR | |
5831 | && code != BIT_XOR_EXPR | |
5832 | && code != BIT_AND_EXPR) | |
5833 | { | |
73fbfcad | 5834 | if (dump_enabled_p ()) |
78c60e3d | 5835 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5836 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
5837 | return false; |
5838 | } | |
5839 | ||
ebfd146a | 5840 | op0 = gimple_assign_rhs1 (stmt); |
894dd753 | 5841 | if (!vect_is_simple_use (op0, vinfo, &dt[0], &vectype)) |
ebfd146a | 5842 | { |
73fbfcad | 5843 | if (dump_enabled_p ()) |
78c60e3d | 5844 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5845 | "use not simple.\n"); |
ebfd146a IR |
5846 | return false; |
5847 | } | |
b690cc0f RG |
5848 | /* If op0 is an external or constant def use a vector type with |
5849 | the same size as the output vector type. */ | |
5850 | if (!vectype) | |
b036c6c5 IE |
5851 | { |
5852 | /* For boolean type we cannot determine vectype by | |
5853 | invariant value (don't know whether it is a vector | |
5854 | of booleans or vector of integers). We use output | |
5855 | vectype because operations on boolean don't change | |
5856 | type. */ | |
2568d8a1 | 5857 | if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op0))) |
b036c6c5 | 5858 | { |
2568d8a1 | 5859 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (scalar_dest))) |
b036c6c5 IE |
5860 | { |
5861 | if (dump_enabled_p ()) | |
5862 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5863 | "not supported operation on bool value.\n"); | |
5864 | return false; | |
5865 | } | |
5866 | vectype = vectype_out; | |
5867 | } | |
5868 | else | |
5869 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5870 | } | |
7d8930a0 IR |
5871 | if (vec_stmt) |
5872 | gcc_assert (vectype); | |
5873 | if (!vectype) | |
5874 | { | |
73fbfcad | 5875 | if (dump_enabled_p ()) |
7d8930a0 | 5876 | { |
78c60e3d SS |
5877 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5878 | "no vectype for scalar type "); | |
5879 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
5880 | TREE_TYPE (op0)); | |
e645e942 | 5881 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
5882 | } |
5883 | ||
5884 | return false; | |
5885 | } | |
b690cc0f RG |
5886 | |
5887 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5888 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
928686b1 | 5889 | if (maybe_ne (nunits_out, nunits_in)) |
b690cc0f | 5890 | return false; |
ebfd146a | 5891 | |
16949072 | 5892 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
5893 | { |
5894 | op1 = gimple_assign_rhs2 (stmt); | |
894dd753 | 5895 | if (!vect_is_simple_use (op1, vinfo, &dt[1])) |
ebfd146a | 5896 | { |
73fbfcad | 5897 | if (dump_enabled_p ()) |
78c60e3d | 5898 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5899 | "use not simple.\n"); |
ebfd146a IR |
5900 | return false; |
5901 | } | |
5902 | } | |
16949072 RG |
5903 | if (op_type == ternary_op) |
5904 | { | |
5905 | op2 = gimple_assign_rhs3 (stmt); | |
894dd753 | 5906 | if (!vect_is_simple_use (op2, vinfo, &dt[2])) |
16949072 | 5907 | { |
73fbfcad | 5908 | if (dump_enabled_p ()) |
78c60e3d | 5909 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5910 | "use not simple.\n"); |
16949072 RG |
5911 | return false; |
5912 | } | |
5913 | } | |
ebfd146a | 5914 | |
b690cc0f | 5915 | /* Multiple types in SLP are handled by creating the appropriate number of |
ff802fa1 | 5916 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 5917 | case of SLP. */ |
fce57248 | 5918 | if (slp_node) |
b690cc0f RG |
5919 | ncopies = 1; |
5920 | else | |
e8f142e2 | 5921 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
b690cc0f RG |
5922 | |
5923 | gcc_assert (ncopies >= 1); | |
5924 | ||
9dc3f7de | 5925 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
5926 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
5927 | || code == RROTATE_EXPR) | |
9dc3f7de | 5928 | return false; |
ebfd146a | 5929 | |
ebfd146a | 5930 | /* Supportable by target? */ |
00f07b86 RH |
5931 | |
5932 | vec_mode = TYPE_MODE (vectype); | |
5933 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 5934 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
5935 | else |
5936 | { | |
5937 | optab = optab_for_tree_code (code, vectype, optab_default); | |
5938 | if (!optab) | |
5deb57cb | 5939 | { |
73fbfcad | 5940 | if (dump_enabled_p ()) |
78c60e3d | 5941 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5942 | "no optab.\n"); |
00f07b86 | 5943 | return false; |
5deb57cb | 5944 | } |
523ba738 RS |
5945 | target_support_p = (optab_handler (optab, vec_mode) |
5946 | != CODE_FOR_nothing); | |
5deb57cb JJ |
5947 | } |
5948 | ||
523ba738 | 5949 | if (!target_support_p) |
ebfd146a | 5950 | { |
73fbfcad | 5951 | if (dump_enabled_p ()) |
78c60e3d | 5952 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5953 | "op not supported by target.\n"); |
ebfd146a | 5954 | /* Check only during analysis. */ |
cf098191 | 5955 | if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) |
ca09abcb | 5956 | || (!vec_stmt && !vect_worthwhile_without_simd_p (vinfo, code))) |
ebfd146a | 5957 | return false; |
73fbfcad | 5958 | if (dump_enabled_p ()) |
e645e942 TJ |
5959 | dump_printf_loc (MSG_NOTE, vect_location, |
5960 | "proceeding using word mode.\n"); | |
383d9c83 IR |
5961 | } |
5962 | ||
4a00c761 | 5963 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
5964 | if (!VECTOR_MODE_P (vec_mode) |
5965 | && !vec_stmt | |
ca09abcb | 5966 | && !vect_worthwhile_without_simd_p (vinfo, code)) |
7d8930a0 | 5967 | { |
73fbfcad | 5968 | if (dump_enabled_p ()) |
78c60e3d | 5969 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5970 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 5971 | return false; |
7d8930a0 | 5972 | } |
ebfd146a | 5973 | |
ebfd146a IR |
5974 | if (!vec_stmt) /* transformation not required. */ |
5975 | { | |
4a00c761 | 5976 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
adac3a68 | 5977 | DUMP_VECT_SCOPE ("vectorizable_operation"); |
68435eb2 | 5978 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
ebfd146a IR |
5979 | return true; |
5980 | } | |
5981 | ||
67b8dbac | 5982 | /* Transform. */ |
ebfd146a | 5983 | |
73fbfcad | 5984 | if (dump_enabled_p ()) |
78c60e3d | 5985 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5986 | "transform binary/unary operation.\n"); |
383d9c83 | 5987 | |
0eb952ea JJ |
5988 | /* POINTER_DIFF_EXPR has pointer arguments which are vectorized as |
5989 | vectors with unsigned elements, but the result is signed. So, we | |
5990 | need to compute the MINUS_EXPR into vectype temporary and | |
5991 | VIEW_CONVERT_EXPR it into the final vectype_out result. */ | |
5992 | tree vec_cvt_dest = NULL_TREE; | |
5993 | if (orig_code == POINTER_DIFF_EXPR) | |
7b76867b RB |
5994 | { |
5995 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5996 | vec_cvt_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
5997 | } | |
5998 | /* Handle def. */ | |
5999 | else | |
6000 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
0eb952ea | 6001 | |
ebfd146a IR |
6002 | /* In case the vectorization factor (VF) is bigger than the number |
6003 | of elements that we can fit in a vectype (nunits), we have to generate | |
6004 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
6005 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
6006 | from one copy of the vector stmt to the next, in the field | |
6007 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
6008 | stages to find the correct vector defs to be used when vectorizing | |
6009 | stmts that use the defs of the current stmt. The example below | |
6010 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
6011 | we need to create 4 vectorized stmts): | |
6012 | ||
6013 | before vectorization: | |
6014 | RELATED_STMT VEC_STMT | |
6015 | S1: x = memref - - | |
6016 | S2: z = x + 1 - - | |
6017 | ||
6018 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
6019 | there): | |
6020 | RELATED_STMT VEC_STMT | |
6021 | VS1_0: vx0 = memref0 VS1_1 - | |
6022 | VS1_1: vx1 = memref1 VS1_2 - | |
6023 | VS1_2: vx2 = memref2 VS1_3 - | |
6024 | VS1_3: vx3 = memref3 - - | |
6025 | S1: x = load - VS1_0 | |
6026 | S2: z = x + 1 - - | |
6027 | ||
6028 | step2: vectorize stmt S2 (done here): | |
6029 | To vectorize stmt S2 we first need to find the relevant vector | |
6030 | def for the first operand 'x'. This is, as usual, obtained from | |
6031 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
6032 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
6033 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
6034 | the vector stmt VS2_0, and as usual, record it in the | |
6035 | STMT_VINFO_VEC_STMT of stmt S2. | |
6036 | When creating the second copy (VS2_1), we obtain the relevant vector | |
6037 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
6038 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
6039 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
6040 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
6041 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
6042 | chain of stmts and pointers: | |
6043 | RELATED_STMT VEC_STMT | |
6044 | VS1_0: vx0 = memref0 VS1_1 - | |
6045 | VS1_1: vx1 = memref1 VS1_2 - | |
6046 | VS1_2: vx2 = memref2 VS1_3 - | |
6047 | VS1_3: vx3 = memref3 - - | |
6048 | S1: x = load - VS1_0 | |
6049 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
6050 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
6051 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
6052 | VS2_3: vz3 = vx3 + v1 - - | |
6053 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
6054 | |
6055 | prev_stmt_info = NULL; | |
6056 | for (j = 0; j < ncopies; j++) | |
6057 | { | |
6058 | /* Handle uses. */ | |
6059 | if (j == 0) | |
4a00c761 | 6060 | { |
d6476f90 | 6061 | if (op_type == binary_op) |
4a00c761 | 6062 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, |
306b0c92 | 6063 | slp_node); |
d6476f90 RB |
6064 | else if (op_type == ternary_op) |
6065 | { | |
6066 | if (slp_node) | |
6067 | { | |
6068 | auto_vec<tree> ops(3); | |
6069 | ops.quick_push (op0); | |
6070 | ops.quick_push (op1); | |
6071 | ops.quick_push (op2); | |
6072 | auto_vec<vec<tree> > vec_defs(3); | |
6073 | vect_get_slp_defs (ops, slp_node, &vec_defs); | |
6074 | vec_oprnds0 = vec_defs[0]; | |
6075 | vec_oprnds1 = vec_defs[1]; | |
6076 | vec_oprnds2 = vec_defs[2]; | |
6077 | } | |
6078 | else | |
6079 | { | |
6080 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
6081 | NULL); | |
6082 | vect_get_vec_defs (op2, NULL_TREE, stmt, &vec_oprnds2, NULL, | |
6083 | NULL); | |
6084 | } | |
6085 | } | |
4a00c761 JJ |
6086 | else |
6087 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 6088 | slp_node); |
4a00c761 | 6089 | } |
ebfd146a | 6090 | else |
4a00c761 JJ |
6091 | { |
6092 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
6093 | if (op_type == ternary_op) | |
6094 | { | |
9771b263 DN |
6095 | tree vec_oprnd = vec_oprnds2.pop (); |
6096 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
6097 | vec_oprnd)); | |
4a00c761 JJ |
6098 | } |
6099 | } | |
6100 | ||
6101 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 6102 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 6103 | { |
4a00c761 | 6104 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 6105 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 6106 | vop2 = ((op_type == ternary_op) |
9771b263 | 6107 | ? vec_oprnds2[i] : NULL_TREE); |
0d0e4a03 | 6108 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
4a00c761 JJ |
6109 | new_temp = make_ssa_name (vec_dest, new_stmt); |
6110 | gimple_assign_set_lhs (new_stmt, new_temp); | |
6111 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
0eb952ea JJ |
6112 | if (vec_cvt_dest) |
6113 | { | |
6114 | new_temp = build1 (VIEW_CONVERT_EXPR, vectype_out, new_temp); | |
6115 | new_stmt = gimple_build_assign (vec_cvt_dest, VIEW_CONVERT_EXPR, | |
6116 | new_temp); | |
6117 | new_temp = make_ssa_name (vec_cvt_dest, new_stmt); | |
6118 | gimple_assign_set_lhs (new_stmt, new_temp); | |
6119 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6120 | } | |
4a00c761 | 6121 | if (slp_node) |
9771b263 | 6122 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
6123 | } |
6124 | ||
4a00c761 JJ |
6125 | if (slp_node) |
6126 | continue; | |
6127 | ||
6128 | if (j == 0) | |
6129 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6130 | else | |
6131 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6132 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
6133 | } |
6134 | ||
9771b263 DN |
6135 | vec_oprnds0.release (); |
6136 | vec_oprnds1.release (); | |
6137 | vec_oprnds2.release (); | |
ebfd146a | 6138 | |
ebfd146a IR |
6139 | return true; |
6140 | } | |
6141 | ||
f702e7d4 | 6142 | /* A helper function to ensure data reference DR's base alignment. */ |
c716e67f XDL |
6143 | |
6144 | static void | |
f702e7d4 | 6145 | ensure_base_align (struct data_reference *dr) |
c716e67f | 6146 | { |
ca823c85 | 6147 | if (DR_VECT_AUX (dr)->misalignment == DR_MISALIGNMENT_UNINITIALIZED) |
c716e67f XDL |
6148 | return; |
6149 | ||
52639a61 | 6150 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f | 6151 | { |
52639a61 | 6152 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 6153 | |
f702e7d4 RS |
6154 | unsigned int align_base_to = DR_TARGET_ALIGNMENT (dr) * BITS_PER_UNIT; |
6155 | ||
428f0c67 | 6156 | if (decl_in_symtab_p (base_decl)) |
f702e7d4 | 6157 | symtab_node::get (base_decl)->increase_alignment (align_base_to); |
428f0c67 JH |
6158 | else |
6159 | { | |
f702e7d4 | 6160 | SET_DECL_ALIGN (base_decl, align_base_to); |
428f0c67 JH |
6161 | DECL_USER_ALIGN (base_decl) = 1; |
6162 | } | |
52639a61 | 6163 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
6164 | } |
6165 | } | |
6166 | ||
ebfd146a | 6167 | |
44fc7854 BE |
6168 | /* Function get_group_alias_ptr_type. |
6169 | ||
6170 | Return the alias type for the group starting at FIRST_STMT. */ | |
6171 | ||
6172 | static tree | |
6173 | get_group_alias_ptr_type (gimple *first_stmt) | |
6174 | { | |
6175 | struct data_reference *first_dr, *next_dr; | |
6176 | gimple *next_stmt; | |
6177 | ||
6178 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
2c53b149 | 6179 | next_stmt = DR_GROUP_NEXT_ELEMENT (vinfo_for_stmt (first_stmt)); |
44fc7854 BE |
6180 | while (next_stmt) |
6181 | { | |
6182 | next_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (next_stmt)); | |
6183 | if (get_alias_set (DR_REF (first_dr)) | |
6184 | != get_alias_set (DR_REF (next_dr))) | |
6185 | { | |
6186 | if (dump_enabled_p ()) | |
6187 | dump_printf_loc (MSG_NOTE, vect_location, | |
6188 | "conflicting alias set types.\n"); | |
6189 | return ptr_type_node; | |
6190 | } | |
2c53b149 | 6191 | next_stmt = DR_GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
44fc7854 BE |
6192 | } |
6193 | return reference_alias_ptr_type (DR_REF (first_dr)); | |
6194 | } | |
6195 | ||
6196 | ||
ebfd146a IR |
6197 | /* Function vectorizable_store. |
6198 | ||
b8698a0f L |
6199 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
6200 | can be vectorized. | |
6201 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6202 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6203 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6204 | ||
6205 | static bool | |
355fe088 | 6206 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
68435eb2 | 6207 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
ebfd146a | 6208 | { |
ebfd146a IR |
6209 | tree data_ref; |
6210 | tree op; | |
6211 | tree vec_oprnd = NULL_TREE; | |
6212 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6213 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 6214 | tree elem_type; |
ebfd146a | 6215 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6216 | struct loop *loop = NULL; |
ef4bddc2 | 6217 | machine_mode vec_mode; |
ebfd146a IR |
6218 | tree dummy; |
6219 | enum dr_alignment_support alignment_support_scheme; | |
929b4411 RS |
6220 | enum vect_def_type rhs_dt = vect_unknown_def_type; |
6221 | enum vect_def_type mask_dt = vect_unknown_def_type; | |
ebfd146a IR |
6222 | stmt_vec_info prev_stmt_info = NULL; |
6223 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6224 | tree dataref_offset = NULL_TREE; |
355fe088 | 6225 | gimple *ptr_incr = NULL; |
ebfd146a IR |
6226 | int ncopies; |
6227 | int j; | |
2de001ee RS |
6228 | gimple *next_stmt, *first_stmt; |
6229 | bool grouped_store; | |
ebfd146a | 6230 | unsigned int group_size, i; |
6e1aa848 DN |
6231 | vec<tree> oprnds = vNULL; |
6232 | vec<tree> result_chain = vNULL; | |
ebfd146a | 6233 | bool inv_p; |
09dfa495 | 6234 | tree offset = NULL_TREE; |
6e1aa848 | 6235 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 6236 | bool slp = (slp_node != NULL); |
ebfd146a | 6237 | unsigned int vec_num; |
a70d6342 | 6238 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 6239 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 6240 | tree aggr_type; |
134c85ca | 6241 | gather_scatter_info gs_info; |
355fe088 | 6242 | gimple *new_stmt; |
d9f21f6a | 6243 | poly_uint64 vf; |
2de001ee | 6244 | vec_load_store_type vls_type; |
44fc7854 | 6245 | tree ref_type; |
a70d6342 | 6246 | |
a70d6342 | 6247 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
6248 | return false; |
6249 | ||
66c16fd9 RB |
6250 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6251 | && ! vec_stmt) | |
ebfd146a IR |
6252 | return false; |
6253 | ||
6254 | /* Is vectorizable store? */ | |
6255 | ||
c3a8f964 RS |
6256 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
6257 | if (is_gimple_assign (stmt)) | |
6258 | { | |
6259 | tree scalar_dest = gimple_assign_lhs (stmt); | |
6260 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR | |
6261 | && is_pattern_stmt_p (stmt_info)) | |
6262 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
6263 | if (TREE_CODE (scalar_dest) != ARRAY_REF | |
6264 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF | |
6265 | && TREE_CODE (scalar_dest) != INDIRECT_REF | |
6266 | && TREE_CODE (scalar_dest) != COMPONENT_REF | |
6267 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
6268 | && TREE_CODE (scalar_dest) != REALPART_EXPR | |
6269 | && TREE_CODE (scalar_dest) != MEM_REF) | |
6270 | return false; | |
6271 | } | |
6272 | else | |
6273 | { | |
6274 | gcall *call = dyn_cast <gcall *> (stmt); | |
f307441a RS |
6275 | if (!call || !gimple_call_internal_p (call)) |
6276 | return false; | |
6277 | ||
6278 | internal_fn ifn = gimple_call_internal_fn (call); | |
6279 | if (!internal_store_fn_p (ifn)) | |
c3a8f964 | 6280 | return false; |
ebfd146a | 6281 | |
c3a8f964 RS |
6282 | if (slp_node != NULL) |
6283 | { | |
6284 | if (dump_enabled_p ()) | |
6285 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6286 | "SLP of masked stores not supported.\n"); | |
6287 | return false; | |
6288 | } | |
6289 | ||
f307441a RS |
6290 | int mask_index = internal_fn_mask_index (ifn); |
6291 | if (mask_index >= 0) | |
6292 | { | |
6293 | mask = gimple_call_arg (call, mask_index); | |
929b4411 RS |
6294 | if (!vect_check_load_store_mask (stmt, mask, &mask_dt, |
6295 | &mask_vectype)) | |
f307441a RS |
6296 | return false; |
6297 | } | |
c3a8f964 RS |
6298 | } |
6299 | ||
6300 | op = vect_get_store_rhs (stmt); | |
ebfd146a | 6301 | |
fce57248 RS |
6302 | /* Cannot have hybrid store SLP -- that would mean storing to the |
6303 | same location twice. */ | |
6304 | gcc_assert (slp == PURE_SLP_STMT (stmt_info)); | |
6305 | ||
f4d09712 | 6306 | tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE; |
4d694b27 | 6307 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 JJ |
6308 | |
6309 | if (loop_vinfo) | |
b17dc4d4 RB |
6310 | { |
6311 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6312 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6313 | } | |
6314 | else | |
6315 | vf = 1; | |
465c8c19 JJ |
6316 | |
6317 | /* Multiple types in SLP are handled by creating the appropriate number of | |
6318 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
6319 | case of SLP. */ | |
fce57248 | 6320 | if (slp) |
465c8c19 JJ |
6321 | ncopies = 1; |
6322 | else | |
e8f142e2 | 6323 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
6324 | |
6325 | gcc_assert (ncopies >= 1); | |
6326 | ||
6327 | /* FORNOW. This restriction should be relaxed. */ | |
6328 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
6329 | { | |
6330 | if (dump_enabled_p ()) | |
6331 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6332 | "multiple types in nested loop.\n"); | |
6333 | return false; | |
6334 | } | |
6335 | ||
929b4411 | 6336 | if (!vect_check_store_rhs (stmt, op, &rhs_dt, &rhs_vectype, &vls_type)) |
f4d09712 KY |
6337 | return false; |
6338 | ||
272c6793 | 6339 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 6340 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 6341 | |
ebfd146a IR |
6342 | if (!STMT_VINFO_DATA_REF (stmt_info)) |
6343 | return false; | |
6344 | ||
2de001ee | 6345 | vect_memory_access_type memory_access_type; |
7e11fc7f | 6346 | if (!get_load_store_type (stmt, vectype, slp, mask, vls_type, ncopies, |
2de001ee RS |
6347 | &memory_access_type, &gs_info)) |
6348 | return false; | |
3bab6342 | 6349 | |
c3a8f964 RS |
6350 | if (mask) |
6351 | { | |
7e11fc7f RS |
6352 | if (memory_access_type == VMAT_CONTIGUOUS) |
6353 | { | |
6354 | if (!VECTOR_MODE_P (vec_mode) | |
6355 | || !can_vec_mask_load_store_p (vec_mode, | |
6356 | TYPE_MODE (mask_vectype), false)) | |
6357 | return false; | |
6358 | } | |
f307441a RS |
6359 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
6360 | && (memory_access_type != VMAT_GATHER_SCATTER || gs_info.decl)) | |
c3a8f964 RS |
6361 | { |
6362 | if (dump_enabled_p ()) | |
6363 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6364 | "unsupported access type for masked store.\n"); | |
6365 | return false; | |
6366 | } | |
c3a8f964 RS |
6367 | } |
6368 | else | |
6369 | { | |
6370 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6371 | (e.g. - array initialization with 0). */ | |
6372 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) | |
6373 | return false; | |
6374 | } | |
6375 | ||
f307441a | 6376 | grouped_store = (STMT_VINFO_GROUPED_ACCESS (stmt_info) |
b5ec4de7 RS |
6377 | && memory_access_type != VMAT_GATHER_SCATTER |
6378 | && (slp || memory_access_type != VMAT_CONTIGUOUS)); | |
7cfb4d93 RS |
6379 | if (grouped_store) |
6380 | { | |
2c53b149 | 6381 | first_stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
7cfb4d93 | 6382 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
2c53b149 | 6383 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
7cfb4d93 RS |
6384 | } |
6385 | else | |
6386 | { | |
6387 | first_stmt = stmt; | |
6388 | first_dr = dr; | |
6389 | group_size = vec_num = 1; | |
6390 | } | |
6391 | ||
ebfd146a IR |
6392 | if (!vec_stmt) /* transformation not required. */ |
6393 | { | |
2de001ee | 6394 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
7cfb4d93 RS |
6395 | |
6396 | if (loop_vinfo | |
6397 | && LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) | |
6398 | check_load_store_masking (loop_vinfo, vectype, vls_type, group_size, | |
bfaa08b7 | 6399 | memory_access_type, &gs_info); |
7cfb4d93 | 6400 | |
ebfd146a | 6401 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; |
68435eb2 RB |
6402 | vect_model_store_cost (stmt_info, ncopies, rhs_dt, memory_access_type, |
6403 | vls_type, slp_node, cost_vec); | |
ebfd146a IR |
6404 | return true; |
6405 | } | |
2de001ee | 6406 | gcc_assert (memory_access_type == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
ebfd146a | 6407 | |
67b8dbac | 6408 | /* Transform. */ |
ebfd146a | 6409 | |
f702e7d4 | 6410 | ensure_base_align (dr); |
c716e67f | 6411 | |
f307441a | 6412 | if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
3bab6342 | 6413 | { |
c3a8f964 | 6414 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, src; |
134c85ca | 6415 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
6416 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; |
6417 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
6418 | edge pe = loop_preheader_edge (loop); | |
6419 | gimple_seq seq; | |
6420 | basic_block new_bb; | |
6421 | enum { NARROW, NONE, WIDEN } modifier; | |
4d694b27 RS |
6422 | poly_uint64 scatter_off_nunits |
6423 | = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype); | |
3bab6342 | 6424 | |
4d694b27 | 6425 | if (known_eq (nunits, scatter_off_nunits)) |
3bab6342 | 6426 | modifier = NONE; |
4d694b27 | 6427 | else if (known_eq (nunits * 2, scatter_off_nunits)) |
3bab6342 | 6428 | { |
3bab6342 AT |
6429 | modifier = WIDEN; |
6430 | ||
4d694b27 RS |
6431 | /* Currently gathers and scatters are only supported for |
6432 | fixed-length vectors. */ | |
6433 | unsigned int count = scatter_off_nunits.to_constant (); | |
6434 | vec_perm_builder sel (count, count, 1); | |
6435 | for (i = 0; i < (unsigned int) count; ++i) | |
6436 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 6437 | |
4d694b27 | 6438 | vec_perm_indices indices (sel, 1, count); |
e3342de4 RS |
6439 | perm_mask = vect_gen_perm_mask_checked (gs_info.offset_vectype, |
6440 | indices); | |
3bab6342 AT |
6441 | gcc_assert (perm_mask != NULL_TREE); |
6442 | } | |
4d694b27 | 6443 | else if (known_eq (nunits, scatter_off_nunits * 2)) |
3bab6342 | 6444 | { |
3bab6342 AT |
6445 | modifier = NARROW; |
6446 | ||
4d694b27 RS |
6447 | /* Currently gathers and scatters are only supported for |
6448 | fixed-length vectors. */ | |
6449 | unsigned int count = nunits.to_constant (); | |
6450 | vec_perm_builder sel (count, count, 1); | |
6451 | for (i = 0; i < (unsigned int) count; ++i) | |
6452 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 6453 | |
4d694b27 | 6454 | vec_perm_indices indices (sel, 2, count); |
e3342de4 | 6455 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); |
3bab6342 AT |
6456 | gcc_assert (perm_mask != NULL_TREE); |
6457 | ncopies *= 2; | |
6458 | } | |
6459 | else | |
6460 | gcc_unreachable (); | |
6461 | ||
134c85ca | 6462 | rettype = TREE_TYPE (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
6463 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
6464 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6465 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6466 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6467 | scaletype = TREE_VALUE (arglist); | |
6468 | ||
6469 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
6470 | && TREE_CODE (rettype) == VOID_TYPE); | |
6471 | ||
134c85ca | 6472 | ptr = fold_convert (ptrtype, gs_info.base); |
3bab6342 AT |
6473 | if (!is_gimple_min_invariant (ptr)) |
6474 | { | |
6475 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6476 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6477 | gcc_assert (!new_bb); | |
6478 | } | |
6479 | ||
6480 | /* Currently we support only unconditional scatter stores, | |
6481 | so mask should be all ones. */ | |
6482 | mask = build_int_cst (masktype, -1); | |
6483 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
6484 | ||
134c85ca | 6485 | scale = build_int_cst (scaletype, gs_info.scale); |
3bab6342 AT |
6486 | |
6487 | prev_stmt_info = NULL; | |
6488 | for (j = 0; j < ncopies; ++j) | |
6489 | { | |
6490 | if (j == 0) | |
6491 | { | |
6492 | src = vec_oprnd1 | |
c3a8f964 | 6493 | = vect_get_vec_def_for_operand (op, stmt); |
3bab6342 | 6494 | op = vec_oprnd0 |
134c85ca | 6495 | = vect_get_vec_def_for_operand (gs_info.offset, stmt); |
3bab6342 AT |
6496 | } |
6497 | else if (modifier != NONE && (j & 1)) | |
6498 | { | |
6499 | if (modifier == WIDEN) | |
6500 | { | |
6501 | src = vec_oprnd1 | |
929b4411 | 6502 | = vect_get_vec_def_for_stmt_copy (rhs_dt, vec_oprnd1); |
3bab6342 AT |
6503 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, |
6504 | stmt, gsi); | |
6505 | } | |
6506 | else if (modifier == NARROW) | |
6507 | { | |
6508 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
6509 | stmt, gsi); | |
6510 | op = vec_oprnd0 | |
134c85ca RS |
6511 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
6512 | vec_oprnd0); | |
3bab6342 AT |
6513 | } |
6514 | else | |
6515 | gcc_unreachable (); | |
6516 | } | |
6517 | else | |
6518 | { | |
6519 | src = vec_oprnd1 | |
929b4411 | 6520 | = vect_get_vec_def_for_stmt_copy (rhs_dt, vec_oprnd1); |
3bab6342 | 6521 | op = vec_oprnd0 |
134c85ca RS |
6522 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
6523 | vec_oprnd0); | |
3bab6342 AT |
6524 | } |
6525 | ||
6526 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
6527 | { | |
928686b1 RS |
6528 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)), |
6529 | TYPE_VECTOR_SUBPARTS (srctype))); | |
0e22bb5a | 6530 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 AT |
6531 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
6532 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
6533 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6534 | src = var; | |
6535 | } | |
6536 | ||
6537 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6538 | { | |
928686b1 RS |
6539 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), |
6540 | TYPE_VECTOR_SUBPARTS (idxtype))); | |
0e22bb5a | 6541 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 AT |
6542 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
6543 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
6544 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6545 | op = var; | |
6546 | } | |
6547 | ||
6548 | new_stmt | |
134c85ca | 6549 | = gimple_build_call (gs_info.decl, 5, ptr, mask, op, src, scale); |
3bab6342 AT |
6550 | |
6551 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6552 | ||
6553 | if (prev_stmt_info == NULL) | |
6554 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6555 | else | |
6556 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6557 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6558 | } | |
6559 | return true; | |
6560 | } | |
6561 | ||
f307441a | 6562 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 6563 | { |
2c53b149 RB |
6564 | gimple *group_stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
6565 | DR_GROUP_STORE_COUNT (vinfo_for_stmt (group_stmt))++; | |
f307441a | 6566 | } |
ebfd146a | 6567 | |
f307441a RS |
6568 | if (grouped_store) |
6569 | { | |
ebfd146a | 6570 | /* FORNOW */ |
a70d6342 | 6571 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
6572 | |
6573 | /* We vectorize all the stmts of the interleaving group when we | |
6574 | reach the last stmt in the group. */ | |
2c53b149 RB |
6575 | if (DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
6576 | < DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
6577 | && !slp) |
6578 | { | |
6579 | *vec_stmt = NULL; | |
6580 | return true; | |
6581 | } | |
6582 | ||
6583 | if (slp) | |
4b5caab7 | 6584 | { |
0d0293ac | 6585 | grouped_store = false; |
4b5caab7 IR |
6586 | /* VEC_NUM is the number of vect stmts to be created for this |
6587 | group. */ | |
6588 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 6589 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
2c53b149 | 6590 | gcc_assert (DR_GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt)) == first_stmt); |
4b5caab7 | 6591 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
c3a8f964 | 6592 | op = vect_get_store_rhs (first_stmt); |
4b5caab7 | 6593 | } |
ebfd146a | 6594 | else |
4b5caab7 IR |
6595 | /* VEC_NUM is the number of vect stmts to be created for this |
6596 | group. */ | |
ebfd146a | 6597 | vec_num = group_size; |
44fc7854 BE |
6598 | |
6599 | ref_type = get_group_alias_ptr_type (first_stmt); | |
ebfd146a | 6600 | } |
b8698a0f | 6601 | else |
7cfb4d93 | 6602 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
b8698a0f | 6603 | |
73fbfcad | 6604 | if (dump_enabled_p ()) |
78c60e3d | 6605 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6606 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 6607 | |
2de001ee RS |
6608 | if (memory_access_type == VMAT_ELEMENTWISE |
6609 | || memory_access_type == VMAT_STRIDED_SLP) | |
f2e2a985 MM |
6610 | { |
6611 | gimple_stmt_iterator incr_gsi; | |
6612 | bool insert_after; | |
355fe088 | 6613 | gimple *incr; |
f2e2a985 MM |
6614 | tree offvar; |
6615 | tree ivstep; | |
6616 | tree running_off; | |
f2e2a985 MM |
6617 | tree stride_base, stride_step, alias_off; |
6618 | tree vec_oprnd; | |
f502d50e | 6619 | unsigned int g; |
4d694b27 RS |
6620 | /* Checked by get_load_store_type. */ |
6621 | unsigned int const_nunits = nunits.to_constant (); | |
f2e2a985 | 6622 | |
7cfb4d93 | 6623 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
f2e2a985 MM |
6624 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); |
6625 | ||
6626 | stride_base | |
6627 | = fold_build_pointer_plus | |
b210f45f | 6628 | (DR_BASE_ADDRESS (first_dr), |
f2e2a985 | 6629 | size_binop (PLUS_EXPR, |
b210f45f | 6630 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
44fc7854 | 6631 | convert_to_ptrofftype (DR_INIT (first_dr)))); |
b210f45f | 6632 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); |
f2e2a985 MM |
6633 | |
6634 | /* For a store with loop-invariant (but other than power-of-2) | |
6635 | stride (i.e. not a grouped access) like so: | |
6636 | ||
6637 | for (i = 0; i < n; i += stride) | |
6638 | array[i] = ...; | |
6639 | ||
6640 | we generate a new induction variable and new stores from | |
6641 | the components of the (vectorized) rhs: | |
6642 | ||
6643 | for (j = 0; ; j += VF*stride) | |
6644 | vectemp = ...; | |
6645 | tmp1 = vectemp[0]; | |
6646 | array[j] = tmp1; | |
6647 | tmp2 = vectemp[1]; | |
6648 | array[j + stride] = tmp2; | |
6649 | ... | |
6650 | */ | |
6651 | ||
4d694b27 | 6652 | unsigned nstores = const_nunits; |
b17dc4d4 | 6653 | unsigned lnel = 1; |
cee62fee | 6654 | tree ltype = elem_type; |
04199738 | 6655 | tree lvectype = vectype; |
cee62fee MM |
6656 | if (slp) |
6657 | { | |
4d694b27 RS |
6658 | if (group_size < const_nunits |
6659 | && const_nunits % group_size == 0) | |
b17dc4d4 | 6660 | { |
4d694b27 | 6661 | nstores = const_nunits / group_size; |
b17dc4d4 RB |
6662 | lnel = group_size; |
6663 | ltype = build_vector_type (elem_type, group_size); | |
04199738 RB |
6664 | lvectype = vectype; |
6665 | ||
6666 | /* First check if vec_extract optab doesn't support extraction | |
6667 | of vector elts directly. */ | |
b397965c | 6668 | scalar_mode elmode = SCALAR_TYPE_MODE (elem_type); |
9da15d40 RS |
6669 | machine_mode vmode; |
6670 | if (!mode_for_vector (elmode, group_size).exists (&vmode) | |
6671 | || !VECTOR_MODE_P (vmode) | |
414fef4e | 6672 | || !targetm.vector_mode_supported_p (vmode) |
04199738 RB |
6673 | || (convert_optab_handler (vec_extract_optab, |
6674 | TYPE_MODE (vectype), vmode) | |
6675 | == CODE_FOR_nothing)) | |
6676 | { | |
6677 | /* Try to avoid emitting an extract of vector elements | |
6678 | by performing the extracts using an integer type of the | |
6679 | same size, extracting from a vector of those and then | |
6680 | re-interpreting it as the original vector type if | |
6681 | supported. */ | |
6682 | unsigned lsize | |
6683 | = group_size * GET_MODE_BITSIZE (elmode); | |
fffbab82 | 6684 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 6685 | unsigned int lnunits = const_nunits / group_size; |
04199738 RB |
6686 | /* If we can't construct such a vector fall back to |
6687 | element extracts from the original vector type and | |
6688 | element size stores. */ | |
4d694b27 | 6689 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 6690 | && VECTOR_MODE_P (vmode) |
414fef4e | 6691 | && targetm.vector_mode_supported_p (vmode) |
04199738 RB |
6692 | && (convert_optab_handler (vec_extract_optab, |
6693 | vmode, elmode) | |
6694 | != CODE_FOR_nothing)) | |
6695 | { | |
4d694b27 | 6696 | nstores = lnunits; |
04199738 RB |
6697 | lnel = group_size; |
6698 | ltype = build_nonstandard_integer_type (lsize, 1); | |
6699 | lvectype = build_vector_type (ltype, nstores); | |
6700 | } | |
6701 | /* Else fall back to vector extraction anyway. | |
6702 | Fewer stores are more important than avoiding spilling | |
6703 | of the vector we extract from. Compared to the | |
6704 | construction case in vectorizable_load no store-forwarding | |
6705 | issue exists here for reasonable archs. */ | |
6706 | } | |
b17dc4d4 | 6707 | } |
4d694b27 RS |
6708 | else if (group_size >= const_nunits |
6709 | && group_size % const_nunits == 0) | |
b17dc4d4 RB |
6710 | { |
6711 | nstores = 1; | |
4d694b27 | 6712 | lnel = const_nunits; |
b17dc4d4 | 6713 | ltype = vectype; |
04199738 | 6714 | lvectype = vectype; |
b17dc4d4 | 6715 | } |
cee62fee MM |
6716 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); |
6717 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
6718 | } | |
6719 | ||
f2e2a985 MM |
6720 | ivstep = stride_step; |
6721 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
b17dc4d4 | 6722 | build_int_cst (TREE_TYPE (ivstep), vf)); |
f2e2a985 MM |
6723 | |
6724 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6725 | ||
b210f45f RB |
6726 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
6727 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); | |
f2e2a985 MM |
6728 | create_iv (stride_base, ivstep, NULL, |
6729 | loop, &incr_gsi, insert_after, | |
6730 | &offvar, NULL); | |
6731 | incr = gsi_stmt (incr_gsi); | |
4fbeb363 | 6732 | loop_vinfo->add_stmt (incr); |
f2e2a985 | 6733 | |
b210f45f | 6734 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
f2e2a985 MM |
6735 | |
6736 | prev_stmt_info = NULL; | |
44fc7854 | 6737 | alias_off = build_int_cst (ref_type, 0); |
f502d50e MM |
6738 | next_stmt = first_stmt; |
6739 | for (g = 0; g < group_size; g++) | |
f2e2a985 | 6740 | { |
f502d50e MM |
6741 | running_off = offvar; |
6742 | if (g) | |
f2e2a985 | 6743 | { |
f502d50e MM |
6744 | tree size = TYPE_SIZE_UNIT (ltype); |
6745 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 6746 | size); |
f502d50e | 6747 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 6748 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 6749 | running_off, pos); |
f2e2a985 | 6750 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 6751 | running_off = newoff; |
f502d50e | 6752 | } |
b17dc4d4 RB |
6753 | unsigned int group_el = 0; |
6754 | unsigned HOST_WIDE_INT | |
6755 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
f502d50e MM |
6756 | for (j = 0; j < ncopies; j++) |
6757 | { | |
c3a8f964 | 6758 | /* We've set op and dt above, from vect_get_store_rhs, |
f502d50e MM |
6759 | and first_stmt == stmt. */ |
6760 | if (j == 0) | |
6761 | { | |
6762 | if (slp) | |
6763 | { | |
6764 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
306b0c92 | 6765 | slp_node); |
f502d50e MM |
6766 | vec_oprnd = vec_oprnds[0]; |
6767 | } | |
6768 | else | |
6769 | { | |
c3a8f964 | 6770 | op = vect_get_store_rhs (next_stmt); |
81c40241 | 6771 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
f502d50e MM |
6772 | } |
6773 | } | |
f2e2a985 | 6774 | else |
f502d50e MM |
6775 | { |
6776 | if (slp) | |
6777 | vec_oprnd = vec_oprnds[j]; | |
6778 | else | |
c079cbac | 6779 | { |
894dd753 | 6780 | vect_is_simple_use (op, vinfo, &rhs_dt); |
929b4411 RS |
6781 | vec_oprnd = vect_get_vec_def_for_stmt_copy (rhs_dt, |
6782 | vec_oprnd); | |
c079cbac | 6783 | } |
f502d50e | 6784 | } |
04199738 RB |
6785 | /* Pun the vector to extract from if necessary. */ |
6786 | if (lvectype != vectype) | |
6787 | { | |
6788 | tree tem = make_ssa_name (lvectype); | |
6789 | gimple *pun | |
6790 | = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
6791 | lvectype, vec_oprnd)); | |
6792 | vect_finish_stmt_generation (stmt, pun, gsi); | |
6793 | vec_oprnd = tem; | |
6794 | } | |
f502d50e MM |
6795 | for (i = 0; i < nstores; i++) |
6796 | { | |
6797 | tree newref, newoff; | |
355fe088 | 6798 | gimple *incr, *assign; |
f502d50e MM |
6799 | tree size = TYPE_SIZE (ltype); |
6800 | /* Extract the i'th component. */ | |
6801 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
6802 | bitsize_int (i), size); | |
6803 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
6804 | size, pos); | |
6805 | ||
6806 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
6807 | NULL_TREE, true, | |
6808 | GSI_SAME_STMT); | |
6809 | ||
b17dc4d4 RB |
6810 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
6811 | group_el * elsz); | |
f502d50e | 6812 | newref = build2 (MEM_REF, ltype, |
b17dc4d4 | 6813 | running_off, this_off); |
19986382 | 6814 | vect_copy_ref_info (newref, DR_REF (first_dr)); |
f502d50e MM |
6815 | |
6816 | /* And store it to *running_off. */ | |
6817 | assign = gimple_build_assign (newref, elem); | |
6818 | vect_finish_stmt_generation (stmt, assign, gsi); | |
6819 | ||
b17dc4d4 RB |
6820 | group_el += lnel; |
6821 | if (! slp | |
6822 | || group_el == group_size) | |
6823 | { | |
6824 | newoff = copy_ssa_name (running_off, NULL); | |
6825 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6826 | running_off, stride_step); | |
6827 | vect_finish_stmt_generation (stmt, incr, gsi); | |
f502d50e | 6828 | |
b17dc4d4 RB |
6829 | running_off = newoff; |
6830 | group_el = 0; | |
6831 | } | |
225ce44b RB |
6832 | if (g == group_size - 1 |
6833 | && !slp) | |
f502d50e MM |
6834 | { |
6835 | if (j == 0 && i == 0) | |
225ce44b RB |
6836 | STMT_VINFO_VEC_STMT (stmt_info) |
6837 | = *vec_stmt = assign; | |
f502d50e MM |
6838 | else |
6839 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign; | |
6840 | prev_stmt_info = vinfo_for_stmt (assign); | |
6841 | } | |
6842 | } | |
f2e2a985 | 6843 | } |
2c53b149 | 6844 | next_stmt = DR_GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
b17dc4d4 RB |
6845 | if (slp) |
6846 | break; | |
f2e2a985 | 6847 | } |
778dd3b6 RB |
6848 | |
6849 | vec_oprnds.release (); | |
f2e2a985 MM |
6850 | return true; |
6851 | } | |
6852 | ||
8c681247 | 6853 | auto_vec<tree> dr_chain (group_size); |
9771b263 | 6854 | oprnds.create (group_size); |
ebfd146a | 6855 | |
720f5239 | 6856 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6857 | gcc_assert (alignment_support_scheme); |
70088b95 RS |
6858 | vec_loop_masks *loop_masks |
6859 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) | |
6860 | ? &LOOP_VINFO_MASKS (loop_vinfo) | |
6861 | : NULL); | |
272c6793 | 6862 | /* Targets with store-lane instructions must not require explicit |
c3a8f964 RS |
6863 | realignment. vect_supportable_dr_alignment always returns either |
6864 | dr_aligned or dr_unaligned_supported for masked operations. */ | |
7cfb4d93 RS |
6865 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES |
6866 | && !mask | |
70088b95 | 6867 | && !loop_masks) |
272c6793 RS |
6868 | || alignment_support_scheme == dr_aligned |
6869 | || alignment_support_scheme == dr_unaligned_supported); | |
6870 | ||
62da9e14 RS |
6871 | if (memory_access_type == VMAT_CONTIGUOUS_DOWN |
6872 | || memory_access_type == VMAT_CONTIGUOUS_REVERSE) | |
09dfa495 BM |
6873 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
6874 | ||
f307441a RS |
6875 | tree bump; |
6876 | tree vec_offset = NULL_TREE; | |
6877 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
6878 | { | |
6879 | aggr_type = NULL_TREE; | |
6880 | bump = NULL_TREE; | |
6881 | } | |
6882 | else if (memory_access_type == VMAT_GATHER_SCATTER) | |
6883 | { | |
6884 | aggr_type = elem_type; | |
6885 | vect_get_strided_load_store_ops (stmt, loop_vinfo, &gs_info, | |
6886 | &bump, &vec_offset); | |
6887 | } | |
272c6793 | 6888 | else |
f307441a RS |
6889 | { |
6890 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
6891 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
6892 | else | |
6893 | aggr_type = vectype; | |
6894 | bump = vect_get_data_ptr_increment (dr, aggr_type, memory_access_type); | |
6895 | } | |
ebfd146a | 6896 | |
c3a8f964 RS |
6897 | if (mask) |
6898 | LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true; | |
6899 | ||
ebfd146a IR |
6900 | /* In case the vectorization factor (VF) is bigger than the number |
6901 | of elements that we can fit in a vectype (nunits), we have to generate | |
6902 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 6903 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
6904 | vect_get_vec_def_for_copy_stmt. */ |
6905 | ||
0d0293ac | 6906 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6907 | |
6908 | S1: &base + 2 = x2 | |
6909 | S2: &base = x0 | |
6910 | S3: &base + 1 = x1 | |
6911 | S4: &base + 3 = x3 | |
6912 | ||
6913 | We create vectorized stores starting from base address (the access of the | |
6914 | first stmt in the chain (S2 in the above example), when the last store stmt | |
6915 | of the chain (S4) is reached: | |
6916 | ||
6917 | VS1: &base = vx2 | |
6918 | VS2: &base + vec_size*1 = vx0 | |
6919 | VS3: &base + vec_size*2 = vx1 | |
6920 | VS4: &base + vec_size*3 = vx3 | |
6921 | ||
6922 | Then permutation statements are generated: | |
6923 | ||
3fcc1b55 JJ |
6924 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
6925 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 6926 | ... |
b8698a0f | 6927 | |
ebfd146a IR |
6928 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
6929 | (the order of the data-refs in the output of vect_permute_store_chain | |
6930 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6931 | the documentation of vect_permute_store_chain()). | |
6932 | ||
6933 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 6934 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 6935 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 6936 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
6937 | */ |
6938 | ||
6939 | prev_stmt_info = NULL; | |
c3a8f964 | 6940 | tree vec_mask = NULL_TREE; |
ebfd146a IR |
6941 | for (j = 0; j < ncopies; j++) |
6942 | { | |
ebfd146a IR |
6943 | |
6944 | if (j == 0) | |
6945 | { | |
6946 | if (slp) | |
6947 | { | |
6948 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c | 6949 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
306b0c92 | 6950 | NULL, slp_node); |
ebfd146a | 6951 | |
9771b263 | 6952 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
6953 | } |
6954 | else | |
6955 | { | |
b8698a0f L |
6956 | /* For interleaved stores we collect vectorized defs for all the |
6957 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
6958 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
6959 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
6960 | ||
2c53b149 | 6961 | If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 6962 | OPRNDS are of size 1. */ |
b8698a0f | 6963 | next_stmt = first_stmt; |
ebfd146a IR |
6964 | for (i = 0; i < group_size; i++) |
6965 | { | |
b8698a0f | 6966 | /* Since gaps are not supported for interleaved stores, |
2c53b149 | 6967 | DR_GROUP_SIZE is the exact number of stmts in the chain. |
b8698a0f | 6968 | Therefore, NEXT_STMT can't be NULL_TREE. In case that |
2c53b149 | 6969 | there is no interleaving, DR_GROUP_SIZE is 1, and only one |
ebfd146a | 6970 | iteration of the loop will be executed. */ |
c3a8f964 | 6971 | op = vect_get_store_rhs (next_stmt); |
81c40241 | 6972 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
9771b263 DN |
6973 | dr_chain.quick_push (vec_oprnd); |
6974 | oprnds.quick_push (vec_oprnd); | |
2c53b149 | 6975 | next_stmt = DR_GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a | 6976 | } |
c3a8f964 RS |
6977 | if (mask) |
6978 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, | |
6979 | mask_vectype); | |
ebfd146a IR |
6980 | } |
6981 | ||
6982 | /* We should have catched mismatched types earlier. */ | |
6983 | gcc_assert (useless_type_conversion_p (vectype, | |
6984 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
6985 | bool simd_lane_access_p |
6986 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
6987 | if (simd_lane_access_p | |
6988 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
6989 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
6990 | && integer_zerop (DR_OFFSET (first_dr)) | |
6991 | && integer_zerop (DR_INIT (first_dr)) | |
6992 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 6993 | get_alias_set (TREE_TYPE (ref_type)))) |
74bf76ed JJ |
6994 | { |
6995 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 6996 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 6997 | inv_p = false; |
74bf76ed | 6998 | } |
f307441a RS |
6999 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
7000 | { | |
7001 | vect_get_gather_scatter_ops (loop, stmt, &gs_info, | |
7002 | &dataref_ptr, &vec_offset); | |
7003 | inv_p = false; | |
7004 | } | |
74bf76ed JJ |
7005 | else |
7006 | dataref_ptr | |
7007 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
7008 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 7009 | offset, &dummy, gsi, &ptr_incr, |
f307441a RS |
7010 | simd_lane_access_p, &inv_p, |
7011 | NULL_TREE, bump); | |
a70d6342 | 7012 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 7013 | } |
b8698a0f | 7014 | else |
ebfd146a | 7015 | { |
b8698a0f L |
7016 | /* For interleaved stores we created vectorized defs for all the |
7017 | defs stored in OPRNDS in the previous iteration (previous copy). | |
7018 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
7019 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
7020 | next copy. | |
2c53b149 | 7021 | If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
7022 | OPRNDS are of size 1. */ |
7023 | for (i = 0; i < group_size; i++) | |
7024 | { | |
9771b263 | 7025 | op = oprnds[i]; |
894dd753 | 7026 | vect_is_simple_use (op, vinfo, &rhs_dt); |
929b4411 | 7027 | vec_oprnd = vect_get_vec_def_for_stmt_copy (rhs_dt, op); |
9771b263 DN |
7028 | dr_chain[i] = vec_oprnd; |
7029 | oprnds[i] = vec_oprnd; | |
ebfd146a | 7030 | } |
c3a8f964 | 7031 | if (mask) |
929b4411 | 7032 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
74bf76ed JJ |
7033 | if (dataref_offset) |
7034 | dataref_offset | |
f307441a RS |
7035 | = int_const_binop (PLUS_EXPR, dataref_offset, bump); |
7036 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
929b4411 RS |
7037 | vec_offset = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
7038 | vec_offset); | |
74bf76ed JJ |
7039 | else |
7040 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
f307441a | 7041 | bump); |
ebfd146a IR |
7042 | } |
7043 | ||
2de001ee | 7044 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 7045 | { |
272c6793 | 7046 | tree vec_array; |
267d3070 | 7047 | |
3ba4ff41 | 7048 | /* Get an array into which we can store the individual vectors. */ |
272c6793 | 7049 | vec_array = create_vector_array (vectype, vec_num); |
3ba4ff41 RS |
7050 | |
7051 | /* Invalidate the current contents of VEC_ARRAY. This should | |
7052 | become an RTL clobber too, which prevents the vector registers | |
7053 | from being upward-exposed. */ | |
7054 | vect_clobber_variable (stmt, gsi, vec_array); | |
7055 | ||
7056 | /* Store the individual vectors into the array. */ | |
272c6793 | 7057 | for (i = 0; i < vec_num; i++) |
c2d7ab2a | 7058 | { |
9771b263 | 7059 | vec_oprnd = dr_chain[i]; |
272c6793 | 7060 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 7061 | } |
b8698a0f | 7062 | |
7cfb4d93 | 7063 | tree final_mask = NULL; |
70088b95 RS |
7064 | if (loop_masks) |
7065 | final_mask = vect_get_loop_mask (gsi, loop_masks, ncopies, | |
7066 | vectype, j); | |
7cfb4d93 RS |
7067 | if (vec_mask) |
7068 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
7069 | vec_mask, gsi); | |
7070 | ||
7e11fc7f | 7071 | gcall *call; |
7cfb4d93 | 7072 | if (final_mask) |
7e11fc7f RS |
7073 | { |
7074 | /* Emit: | |
7075 | MASK_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK, | |
7076 | VEC_ARRAY). */ | |
7077 | unsigned int align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype)); | |
7078 | tree alias_ptr = build_int_cst (ref_type, align); | |
7079 | call = gimple_build_call_internal (IFN_MASK_STORE_LANES, 4, | |
7080 | dataref_ptr, alias_ptr, | |
7cfb4d93 | 7081 | final_mask, vec_array); |
7e11fc7f RS |
7082 | } |
7083 | else | |
7084 | { | |
7085 | /* Emit: | |
7086 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
7087 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); | |
7088 | call = gimple_build_call_internal (IFN_STORE_LANES, 1, | |
7089 | vec_array); | |
7090 | gimple_call_set_lhs (call, data_ref); | |
7091 | } | |
a844293d RS |
7092 | gimple_call_set_nothrow (call, true); |
7093 | new_stmt = call; | |
267d3070 | 7094 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3ba4ff41 RS |
7095 | |
7096 | /* Record that VEC_ARRAY is now dead. */ | |
7097 | vect_clobber_variable (stmt, gsi, vec_array); | |
272c6793 RS |
7098 | } |
7099 | else | |
7100 | { | |
7101 | new_stmt = NULL; | |
0d0293ac | 7102 | if (grouped_store) |
272c6793 | 7103 | { |
b6b9227d JJ |
7104 | if (j == 0) |
7105 | result_chain.create (group_size); | |
272c6793 RS |
7106 | /* Permute. */ |
7107 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
7108 | &result_chain); | |
7109 | } | |
c2d7ab2a | 7110 | |
272c6793 RS |
7111 | next_stmt = first_stmt; |
7112 | for (i = 0; i < vec_num; i++) | |
7113 | { | |
644ffefd | 7114 | unsigned align, misalign; |
272c6793 | 7115 | |
7cfb4d93 | 7116 | tree final_mask = NULL_TREE; |
70088b95 RS |
7117 | if (loop_masks) |
7118 | final_mask = vect_get_loop_mask (gsi, loop_masks, | |
7119 | vec_num * ncopies, | |
7cfb4d93 RS |
7120 | vectype, vec_num * j + i); |
7121 | if (vec_mask) | |
7122 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
7123 | vec_mask, gsi); | |
7124 | ||
f307441a RS |
7125 | if (memory_access_type == VMAT_GATHER_SCATTER) |
7126 | { | |
7127 | tree scale = size_int (gs_info.scale); | |
7128 | gcall *call; | |
70088b95 | 7129 | if (loop_masks) |
f307441a RS |
7130 | call = gimple_build_call_internal |
7131 | (IFN_MASK_SCATTER_STORE, 5, dataref_ptr, vec_offset, | |
7132 | scale, vec_oprnd, final_mask); | |
7133 | else | |
7134 | call = gimple_build_call_internal | |
7135 | (IFN_SCATTER_STORE, 4, dataref_ptr, vec_offset, | |
7136 | scale, vec_oprnd); | |
7137 | gimple_call_set_nothrow (call, true); | |
7138 | new_stmt = call; | |
7139 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7140 | break; | |
7141 | } | |
7142 | ||
272c6793 RS |
7143 | if (i > 0) |
7144 | /* Bump the vector pointer. */ | |
7145 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
f307441a | 7146 | stmt, bump); |
272c6793 RS |
7147 | |
7148 | if (slp) | |
9771b263 | 7149 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
7150 | else if (grouped_store) |
7151 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 7152 | vect_permute_store_chain(). */ |
9771b263 | 7153 | vec_oprnd = result_chain[i]; |
272c6793 | 7154 | |
f702e7d4 | 7155 | align = DR_TARGET_ALIGNMENT (first_dr); |
272c6793 | 7156 | if (aligned_access_p (first_dr)) |
644ffefd | 7157 | misalign = 0; |
272c6793 RS |
7158 | else if (DR_MISALIGNMENT (first_dr) == -1) |
7159 | { | |
25f68d90 | 7160 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 7161 | misalign = 0; |
272c6793 RS |
7162 | } |
7163 | else | |
c3a8f964 | 7164 | misalign = DR_MISALIGNMENT (first_dr); |
aed93b23 RB |
7165 | if (dataref_offset == NULL_TREE |
7166 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
7167 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
7168 | misalign); | |
c2d7ab2a | 7169 | |
62da9e14 | 7170 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
09dfa495 BM |
7171 | { |
7172 | tree perm_mask = perm_mask_for_reverse (vectype); | |
7173 | tree perm_dest | |
c3a8f964 | 7174 | = vect_create_destination_var (vect_get_store_rhs (stmt), |
09dfa495 | 7175 | vectype); |
b731b390 | 7176 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
7177 | |
7178 | /* Generate the permute statement. */ | |
355fe088 | 7179 | gimple *perm_stmt |
0d0e4a03 JJ |
7180 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
7181 | vec_oprnd, perm_mask); | |
09dfa495 BM |
7182 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
7183 | ||
7184 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
7185 | vec_oprnd = new_temp; | |
7186 | } | |
7187 | ||
272c6793 | 7188 | /* Arguments are ready. Create the new vector stmt. */ |
7cfb4d93 | 7189 | if (final_mask) |
c3a8f964 RS |
7190 | { |
7191 | align = least_bit_hwi (misalign | align); | |
7192 | tree ptr = build_int_cst (ref_type, align); | |
7193 | gcall *call | |
7194 | = gimple_build_call_internal (IFN_MASK_STORE, 4, | |
7195 | dataref_ptr, ptr, | |
7cfb4d93 | 7196 | final_mask, vec_oprnd); |
c3a8f964 RS |
7197 | gimple_call_set_nothrow (call, true); |
7198 | new_stmt = call; | |
7199 | } | |
7200 | else | |
7201 | { | |
7202 | data_ref = fold_build2 (MEM_REF, vectype, | |
7203 | dataref_ptr, | |
7204 | dataref_offset | |
7205 | ? dataref_offset | |
7206 | : build_int_cst (ref_type, 0)); | |
7207 | if (aligned_access_p (first_dr)) | |
7208 | ; | |
7209 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
7210 | TREE_TYPE (data_ref) | |
7211 | = build_aligned_type (TREE_TYPE (data_ref), | |
7212 | align * BITS_PER_UNIT); | |
7213 | else | |
7214 | TREE_TYPE (data_ref) | |
7215 | = build_aligned_type (TREE_TYPE (data_ref), | |
7216 | TYPE_ALIGN (elem_type)); | |
19986382 | 7217 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); |
c3a8f964 RS |
7218 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); |
7219 | } | |
272c6793 | 7220 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
7221 | |
7222 | if (slp) | |
7223 | continue; | |
7224 | ||
2c53b149 | 7225 | next_stmt = DR_GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
7226 | if (!next_stmt) |
7227 | break; | |
7228 | } | |
ebfd146a | 7229 | } |
1da0876c RS |
7230 | if (!slp) |
7231 | { | |
7232 | if (j == 0) | |
7233 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7234 | else | |
7235 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7236 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7237 | } | |
ebfd146a IR |
7238 | } |
7239 | ||
9771b263 DN |
7240 | oprnds.release (); |
7241 | result_chain.release (); | |
7242 | vec_oprnds.release (); | |
ebfd146a IR |
7243 | |
7244 | return true; | |
7245 | } | |
7246 | ||
557be5a8 AL |
7247 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
7248 | VECTOR_CST mask. No checks are made that the target platform supports the | |
7ac7e286 | 7249 | mask, so callers may wish to test can_vec_perm_const_p separately, or use |
557be5a8 | 7250 | vect_gen_perm_mask_checked. */ |
a1e53f3f | 7251 | |
3fcc1b55 | 7252 | tree |
4aae3cb3 | 7253 | vect_gen_perm_mask_any (tree vectype, const vec_perm_indices &sel) |
a1e53f3f | 7254 | { |
b00cb3bf | 7255 | tree mask_type; |
a1e53f3f | 7256 | |
0ecc2b7d RS |
7257 | poly_uint64 nunits = sel.length (); |
7258 | gcc_assert (known_eq (nunits, TYPE_VECTOR_SUBPARTS (vectype))); | |
b00cb3bf RS |
7259 | |
7260 | mask_type = build_vector_type (ssizetype, nunits); | |
736d0f28 | 7261 | return vec_perm_indices_to_tree (mask_type, sel); |
a1e53f3f L |
7262 | } |
7263 | ||
7ac7e286 | 7264 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_const_p, |
cf7aa6a3 | 7265 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ |
557be5a8 AL |
7266 | |
7267 | tree | |
4aae3cb3 | 7268 | vect_gen_perm_mask_checked (tree vectype, const vec_perm_indices &sel) |
557be5a8 | 7269 | { |
7ac7e286 | 7270 | gcc_assert (can_vec_perm_const_p (TYPE_MODE (vectype), sel)); |
557be5a8 AL |
7271 | return vect_gen_perm_mask_any (vectype, sel); |
7272 | } | |
7273 | ||
aec7ae7d JJ |
7274 | /* Given a vector variable X and Y, that was generated for the scalar |
7275 | STMT, generate instructions to permute the vector elements of X and Y | |
7276 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
7277 | permuted vector variable. */ | |
a1e53f3f L |
7278 | |
7279 | static tree | |
355fe088 | 7280 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 7281 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
7282 | { |
7283 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 7284 | tree perm_dest, data_ref; |
355fe088 | 7285 | gimple *perm_stmt; |
a1e53f3f | 7286 | |
7ad429a4 RS |
7287 | tree scalar_dest = gimple_get_lhs (stmt); |
7288 | if (TREE_CODE (scalar_dest) == SSA_NAME) | |
7289 | perm_dest = vect_create_destination_var (scalar_dest, vectype); | |
7290 | else | |
7291 | perm_dest = vect_get_new_vect_var (vectype, vect_simple_var, NULL); | |
b731b390 | 7292 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
7293 | |
7294 | /* Generate the permute statement. */ | |
0d0e4a03 | 7295 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
7296 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
7297 | ||
7298 | return data_ref; | |
7299 | } | |
7300 | ||
6b916b36 RB |
7301 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
7302 | inserting them on the loops preheader edge. Returns true if we | |
7303 | were successful in doing so (and thus STMT can be moved then), | |
7304 | otherwise returns false. */ | |
7305 | ||
7306 | static bool | |
355fe088 | 7307 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
7308 | { |
7309 | ssa_op_iter i; | |
7310 | tree op; | |
7311 | bool any = false; | |
7312 | ||
7313 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
7314 | { | |
355fe088 | 7315 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
7316 | if (!gimple_nop_p (def_stmt) |
7317 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
7318 | { | |
7319 | /* Make sure we don't need to recurse. While we could do | |
7320 | so in simple cases when there are more complex use webs | |
7321 | we don't have an easy way to preserve stmt order to fulfil | |
7322 | dependencies within them. */ | |
7323 | tree op2; | |
7324 | ssa_op_iter i2; | |
d1417442 JJ |
7325 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
7326 | return false; | |
6b916b36 RB |
7327 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
7328 | { | |
355fe088 | 7329 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
7330 | if (!gimple_nop_p (def_stmt2) |
7331 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
7332 | return false; | |
7333 | } | |
7334 | any = true; | |
7335 | } | |
7336 | } | |
7337 | ||
7338 | if (!any) | |
7339 | return true; | |
7340 | ||
7341 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
7342 | { | |
355fe088 | 7343 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
7344 | if (!gimple_nop_p (def_stmt) |
7345 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
7346 | { | |
7347 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
7348 | gsi_remove (&gsi, false); | |
7349 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
7350 | } | |
7351 | } | |
7352 | ||
7353 | return true; | |
7354 | } | |
7355 | ||
ebfd146a IR |
7356 | /* vectorizable_load. |
7357 | ||
b8698a0f L |
7358 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
7359 | can be vectorized. | |
7360 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
7361 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
7362 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7363 | ||
7364 | static bool | |
355fe088 | 7365 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
68435eb2 RB |
7366 | slp_tree slp_node, slp_instance slp_node_instance, |
7367 | stmt_vector_for_cost *cost_vec) | |
ebfd146a IR |
7368 | { |
7369 | tree scalar_dest; | |
7370 | tree vec_dest = NULL; | |
7371 | tree data_ref = NULL; | |
7372 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 7373 | stmt_vec_info prev_stmt_info; |
ebfd146a | 7374 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 7375 | struct loop *loop = NULL; |
ebfd146a | 7376 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 7377 | bool nested_in_vect_loop = false; |
c716e67f | 7378 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 7379 | tree elem_type; |
ebfd146a | 7380 | tree new_temp; |
ef4bddc2 | 7381 | machine_mode mode; |
355fe088 | 7382 | gimple *new_stmt = NULL; |
ebfd146a IR |
7383 | tree dummy; |
7384 | enum dr_alignment_support alignment_support_scheme; | |
7385 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 7386 | tree dataref_offset = NULL_TREE; |
355fe088 | 7387 | gimple *ptr_incr = NULL; |
ebfd146a | 7388 | int ncopies; |
4d694b27 RS |
7389 | int i, j; |
7390 | unsigned int group_size; | |
7391 | poly_uint64 group_gap_adj; | |
ebfd146a IR |
7392 | tree msq = NULL_TREE, lsq; |
7393 | tree offset = NULL_TREE; | |
356bbc4c | 7394 | tree byte_offset = NULL_TREE; |
ebfd146a | 7395 | tree realignment_token = NULL_TREE; |
538dd0b7 | 7396 | gphi *phi = NULL; |
6e1aa848 | 7397 | vec<tree> dr_chain = vNULL; |
0d0293ac | 7398 | bool grouped_load = false; |
355fe088 | 7399 | gimple *first_stmt; |
4f0a0218 | 7400 | gimple *first_stmt_for_drptr = NULL; |
ebfd146a IR |
7401 | bool inv_p; |
7402 | bool compute_in_loop = false; | |
7403 | struct loop *at_loop; | |
7404 | int vec_num; | |
7405 | bool slp = (slp_node != NULL); | |
7406 | bool slp_perm = false; | |
a70d6342 | 7407 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
d9f21f6a | 7408 | poly_uint64 vf; |
272c6793 | 7409 | tree aggr_type; |
134c85ca | 7410 | gather_scatter_info gs_info; |
310213d4 | 7411 | vec_info *vinfo = stmt_info->vinfo; |
44fc7854 | 7412 | tree ref_type; |
929b4411 | 7413 | enum vect_def_type mask_dt = vect_unknown_def_type; |
a70d6342 | 7414 | |
465c8c19 JJ |
7415 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
7416 | return false; | |
7417 | ||
66c16fd9 RB |
7418 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7419 | && ! vec_stmt) | |
465c8c19 JJ |
7420 | return false; |
7421 | ||
c3a8f964 RS |
7422 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
7423 | if (is_gimple_assign (stmt)) | |
7424 | { | |
7425 | scalar_dest = gimple_assign_lhs (stmt); | |
7426 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
7427 | return false; | |
465c8c19 | 7428 | |
c3a8f964 RS |
7429 | tree_code code = gimple_assign_rhs_code (stmt); |
7430 | if (code != ARRAY_REF | |
7431 | && code != BIT_FIELD_REF | |
7432 | && code != INDIRECT_REF | |
7433 | && code != COMPONENT_REF | |
7434 | && code != IMAGPART_EXPR | |
7435 | && code != REALPART_EXPR | |
7436 | && code != MEM_REF | |
7437 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
7438 | return false; | |
7439 | } | |
7440 | else | |
7441 | { | |
7442 | gcall *call = dyn_cast <gcall *> (stmt); | |
bfaa08b7 RS |
7443 | if (!call || !gimple_call_internal_p (call)) |
7444 | return false; | |
7445 | ||
7446 | internal_fn ifn = gimple_call_internal_fn (call); | |
7447 | if (!internal_load_fn_p (ifn)) | |
c3a8f964 | 7448 | return false; |
465c8c19 | 7449 | |
c3a8f964 RS |
7450 | scalar_dest = gimple_call_lhs (call); |
7451 | if (!scalar_dest) | |
7452 | return false; | |
7453 | ||
7454 | if (slp_node != NULL) | |
7455 | { | |
7456 | if (dump_enabled_p ()) | |
7457 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7458 | "SLP of masked loads not supported.\n"); | |
7459 | return false; | |
7460 | } | |
7461 | ||
bfaa08b7 RS |
7462 | int mask_index = internal_fn_mask_index (ifn); |
7463 | if (mask_index >= 0) | |
7464 | { | |
7465 | mask = gimple_call_arg (call, mask_index); | |
929b4411 RS |
7466 | if (!vect_check_load_store_mask (stmt, mask, &mask_dt, |
7467 | &mask_vectype)) | |
bfaa08b7 RS |
7468 | return false; |
7469 | } | |
c3a8f964 | 7470 | } |
465c8c19 JJ |
7471 | |
7472 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
7473 | return false; | |
7474 | ||
7475 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
4d694b27 | 7476 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 | 7477 | |
a70d6342 IR |
7478 | if (loop_vinfo) |
7479 | { | |
7480 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
7481 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
7482 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
7483 | } | |
7484 | else | |
3533e503 | 7485 | vf = 1; |
ebfd146a IR |
7486 | |
7487 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 7488 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 7489 | case of SLP. */ |
fce57248 | 7490 | if (slp) |
ebfd146a IR |
7491 | ncopies = 1; |
7492 | else | |
e8f142e2 | 7493 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
ebfd146a IR |
7494 | |
7495 | gcc_assert (ncopies >= 1); | |
7496 | ||
7497 | /* FORNOW. This restriction should be relaxed. */ | |
7498 | if (nested_in_vect_loop && ncopies > 1) | |
7499 | { | |
73fbfcad | 7500 | if (dump_enabled_p ()) |
78c60e3d | 7501 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7502 | "multiple types in nested loop.\n"); |
ebfd146a IR |
7503 | return false; |
7504 | } | |
7505 | ||
f2556b68 RB |
7506 | /* Invalidate assumptions made by dependence analysis when vectorization |
7507 | on the unrolled body effectively re-orders stmts. */ | |
7508 | if (ncopies > 1 | |
7509 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
7510 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
7511 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
7512 | { |
7513 | if (dump_enabled_p ()) | |
7514 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7515 | "cannot perform implicit CSE when unrolling " | |
7516 | "with negative dependence distance\n"); | |
7517 | return false; | |
7518 | } | |
7519 | ||
7b7b1813 | 7520 | elem_type = TREE_TYPE (vectype); |
947131ba | 7521 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
7522 | |
7523 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
7524 | (e.g. - data copies). */ | |
947131ba | 7525 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 7526 | { |
73fbfcad | 7527 | if (dump_enabled_p ()) |
78c60e3d | 7528 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7529 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
7530 | return false; |
7531 | } | |
7532 | ||
ebfd146a | 7533 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 7534 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 7535 | { |
0d0293ac | 7536 | grouped_load = true; |
ebfd146a | 7537 | /* FORNOW */ |
2de001ee RS |
7538 | gcc_assert (!nested_in_vect_loop); |
7539 | gcc_assert (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)); | |
ebfd146a | 7540 | |
2c53b149 RB |
7541 | first_stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
7542 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
d5f035ea | 7543 | |
b1af7da6 RB |
7544 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
7545 | slp_perm = true; | |
7546 | ||
f2556b68 RB |
7547 | /* Invalidate assumptions made by dependence analysis when vectorization |
7548 | on the unrolled body effectively re-orders stmts. */ | |
7549 | if (!PURE_SLP_STMT (stmt_info) | |
7550 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
7551 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
7552 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
7553 | { |
7554 | if (dump_enabled_p ()) | |
7555 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7556 | "cannot perform implicit CSE when performing " | |
7557 | "group loads with negative dependence distance\n"); | |
7558 | return false; | |
7559 | } | |
96bb56b2 RB |
7560 | |
7561 | /* Similarly when the stmt is a load that is both part of a SLP | |
7562 | instance and a loop vectorized stmt via the same-dr mechanism | |
7563 | we have to give up. */ | |
2c53b149 | 7564 | if (DR_GROUP_SAME_DR_STMT (stmt_info) |
96bb56b2 RB |
7565 | && (STMT_SLP_TYPE (stmt_info) |
7566 | != STMT_SLP_TYPE (vinfo_for_stmt | |
2c53b149 | 7567 | (DR_GROUP_SAME_DR_STMT (stmt_info))))) |
96bb56b2 RB |
7568 | { |
7569 | if (dump_enabled_p ()) | |
7570 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7571 | "conflicting SLP types for CSEd load\n"); | |
7572 | return false; | |
7573 | } | |
ebfd146a | 7574 | } |
7cfb4d93 RS |
7575 | else |
7576 | group_size = 1; | |
ebfd146a | 7577 | |
2de001ee | 7578 | vect_memory_access_type memory_access_type; |
7e11fc7f | 7579 | if (!get_load_store_type (stmt, vectype, slp, mask, VLS_LOAD, ncopies, |
2de001ee RS |
7580 | &memory_access_type, &gs_info)) |
7581 | return false; | |
a1e53f3f | 7582 | |
c3a8f964 RS |
7583 | if (mask) |
7584 | { | |
7585 | if (memory_access_type == VMAT_CONTIGUOUS) | |
7586 | { | |
7e11fc7f RS |
7587 | machine_mode vec_mode = TYPE_MODE (vectype); |
7588 | if (!VECTOR_MODE_P (vec_mode) | |
7589 | || !can_vec_mask_load_store_p (vec_mode, | |
c3a8f964 RS |
7590 | TYPE_MODE (mask_vectype), true)) |
7591 | return false; | |
7592 | } | |
bfaa08b7 | 7593 | else if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
c3a8f964 RS |
7594 | { |
7595 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); | |
7596 | tree masktype | |
7597 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
7598 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
7599 | { | |
7600 | if (dump_enabled_p ()) | |
7601 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7602 | "masked gather with integer mask not" | |
7603 | " supported."); | |
7604 | return false; | |
7605 | } | |
7606 | } | |
bfaa08b7 RS |
7607 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
7608 | && memory_access_type != VMAT_GATHER_SCATTER) | |
c3a8f964 RS |
7609 | { |
7610 | if (dump_enabled_p ()) | |
7611 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7612 | "unsupported access type for masked load.\n"); | |
7613 | return false; | |
7614 | } | |
7615 | } | |
7616 | ||
ebfd146a IR |
7617 | if (!vec_stmt) /* transformation not required. */ |
7618 | { | |
2de001ee RS |
7619 | if (!slp) |
7620 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; | |
7cfb4d93 RS |
7621 | |
7622 | if (loop_vinfo | |
7623 | && LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) | |
7624 | check_load_store_masking (loop_vinfo, vectype, VLS_LOAD, group_size, | |
bfaa08b7 | 7625 | memory_access_type, &gs_info); |
7cfb4d93 | 7626 | |
ebfd146a | 7627 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; |
68435eb2 RB |
7628 | vect_model_load_cost (stmt_info, ncopies, memory_access_type, |
7629 | slp_node_instance, slp_node, cost_vec); | |
ebfd146a IR |
7630 | return true; |
7631 | } | |
7632 | ||
2de001ee RS |
7633 | if (!slp) |
7634 | gcc_assert (memory_access_type | |
7635 | == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); | |
7636 | ||
73fbfcad | 7637 | if (dump_enabled_p ()) |
78c60e3d | 7638 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 7639 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a | 7640 | |
67b8dbac | 7641 | /* Transform. */ |
ebfd146a | 7642 | |
f702e7d4 | 7643 | ensure_base_align (dr); |
c716e67f | 7644 | |
bfaa08b7 | 7645 | if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
aec7ae7d | 7646 | { |
929b4411 RS |
7647 | vect_build_gather_load_calls (stmt, gsi, vec_stmt, &gs_info, mask, |
7648 | mask_dt); | |
aec7ae7d JJ |
7649 | return true; |
7650 | } | |
2de001ee RS |
7651 | |
7652 | if (memory_access_type == VMAT_ELEMENTWISE | |
7653 | || memory_access_type == VMAT_STRIDED_SLP) | |
7d75abc8 MM |
7654 | { |
7655 | gimple_stmt_iterator incr_gsi; | |
7656 | bool insert_after; | |
355fe088 | 7657 | gimple *incr; |
7d75abc8 | 7658 | tree offvar; |
7d75abc8 MM |
7659 | tree ivstep; |
7660 | tree running_off; | |
9771b263 | 7661 | vec<constructor_elt, va_gc> *v = NULL; |
14ac6aa2 | 7662 | tree stride_base, stride_step, alias_off; |
4d694b27 RS |
7663 | /* Checked by get_load_store_type. */ |
7664 | unsigned int const_nunits = nunits.to_constant (); | |
b210f45f | 7665 | unsigned HOST_WIDE_INT cst_offset = 0; |
14ac6aa2 | 7666 | |
7cfb4d93 | 7667 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
14ac6aa2 | 7668 | gcc_assert (!nested_in_vect_loop); |
7d75abc8 | 7669 | |
b210f45f | 7670 | if (grouped_load) |
44fc7854 | 7671 | { |
2c53b149 | 7672 | first_stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
44fc7854 | 7673 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
44fc7854 | 7674 | } |
ab313a8c | 7675 | else |
44fc7854 BE |
7676 | { |
7677 | first_stmt = stmt; | |
7678 | first_dr = dr; | |
b210f45f RB |
7679 | } |
7680 | if (slp && grouped_load) | |
7681 | { | |
2c53b149 | 7682 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
b210f45f RB |
7683 | ref_type = get_group_alias_ptr_type (first_stmt); |
7684 | } | |
7685 | else | |
7686 | { | |
7687 | if (grouped_load) | |
7688 | cst_offset | |
7689 | = (tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))) | |
7690 | * vect_get_place_in_interleaving_chain (stmt, first_stmt)); | |
44fc7854 | 7691 | group_size = 1; |
b210f45f | 7692 | ref_type = reference_alias_ptr_type (DR_REF (dr)); |
44fc7854 | 7693 | } |
ab313a8c | 7694 | |
14ac6aa2 RB |
7695 | stride_base |
7696 | = fold_build_pointer_plus | |
ab313a8c | 7697 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 7698 | size_binop (PLUS_EXPR, |
ab313a8c RB |
7699 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
7700 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
7701 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
7702 | |
7703 | /* For a load with loop-invariant (but other than power-of-2) | |
7704 | stride (i.e. not a grouped access) like so: | |
7705 | ||
7706 | for (i = 0; i < n; i += stride) | |
7707 | ... = array[i]; | |
7708 | ||
7709 | we generate a new induction variable and new accesses to | |
7710 | form a new vector (or vectors, depending on ncopies): | |
7711 | ||
7712 | for (j = 0; ; j += VF*stride) | |
7713 | tmp1 = array[j]; | |
7714 | tmp2 = array[j + stride]; | |
7715 | ... | |
7716 | vectemp = {tmp1, tmp2, ...} | |
7717 | */ | |
7718 | ||
ab313a8c RB |
7719 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
7720 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
7721 | |
7722 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
7723 | ||
b210f45f RB |
7724 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
7725 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); | |
7726 | create_iv (stride_base, ivstep, NULL, | |
7d75abc8 MM |
7727 | loop, &incr_gsi, insert_after, |
7728 | &offvar, NULL); | |
7729 | incr = gsi_stmt (incr_gsi); | |
4fbeb363 | 7730 | loop_vinfo->add_stmt (incr); |
7d75abc8 | 7731 | |
b210f45f | 7732 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
7d75abc8 MM |
7733 | |
7734 | prev_stmt_info = NULL; | |
7735 | running_off = offvar; | |
44fc7854 | 7736 | alias_off = build_int_cst (ref_type, 0); |
4d694b27 | 7737 | int nloads = const_nunits; |
e09b4c37 | 7738 | int lnel = 1; |
7b5fc413 | 7739 | tree ltype = TREE_TYPE (vectype); |
ea60dd34 | 7740 | tree lvectype = vectype; |
b266b968 | 7741 | auto_vec<tree> dr_chain; |
2de001ee | 7742 | if (memory_access_type == VMAT_STRIDED_SLP) |
7b5fc413 | 7743 | { |
4d694b27 | 7744 | if (group_size < const_nunits) |
e09b4c37 | 7745 | { |
ff03930a JJ |
7746 | /* First check if vec_init optab supports construction from |
7747 | vector elts directly. */ | |
b397965c | 7748 | scalar_mode elmode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); |
9da15d40 RS |
7749 | machine_mode vmode; |
7750 | if (mode_for_vector (elmode, group_size).exists (&vmode) | |
7751 | && VECTOR_MODE_P (vmode) | |
414fef4e | 7752 | && targetm.vector_mode_supported_p (vmode) |
ff03930a JJ |
7753 | && (convert_optab_handler (vec_init_optab, |
7754 | TYPE_MODE (vectype), vmode) | |
7755 | != CODE_FOR_nothing)) | |
ea60dd34 | 7756 | { |
4d694b27 | 7757 | nloads = const_nunits / group_size; |
ea60dd34 | 7758 | lnel = group_size; |
ff03930a JJ |
7759 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); |
7760 | } | |
7761 | else | |
7762 | { | |
7763 | /* Otherwise avoid emitting a constructor of vector elements | |
7764 | by performing the loads using an integer type of the same | |
7765 | size, constructing a vector of those and then | |
7766 | re-interpreting it as the original vector type. | |
7767 | This avoids a huge runtime penalty due to the general | |
7768 | inability to perform store forwarding from smaller stores | |
7769 | to a larger load. */ | |
7770 | unsigned lsize | |
7771 | = group_size * TYPE_PRECISION (TREE_TYPE (vectype)); | |
fffbab82 | 7772 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 7773 | unsigned int lnunits = const_nunits / group_size; |
ff03930a JJ |
7774 | /* If we can't construct such a vector fall back to |
7775 | element loads of the original vector type. */ | |
4d694b27 | 7776 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 7777 | && VECTOR_MODE_P (vmode) |
414fef4e | 7778 | && targetm.vector_mode_supported_p (vmode) |
ff03930a JJ |
7779 | && (convert_optab_handler (vec_init_optab, vmode, elmode) |
7780 | != CODE_FOR_nothing)) | |
7781 | { | |
4d694b27 | 7782 | nloads = lnunits; |
ff03930a JJ |
7783 | lnel = group_size; |
7784 | ltype = build_nonstandard_integer_type (lsize, 1); | |
7785 | lvectype = build_vector_type (ltype, nloads); | |
7786 | } | |
ea60dd34 | 7787 | } |
e09b4c37 | 7788 | } |
2de001ee | 7789 | else |
e09b4c37 | 7790 | { |
ea60dd34 | 7791 | nloads = 1; |
4d694b27 | 7792 | lnel = const_nunits; |
e09b4c37 | 7793 | ltype = vectype; |
e09b4c37 | 7794 | } |
2de001ee RS |
7795 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); |
7796 | } | |
bb4e4747 BC |
7797 | /* Load vector(1) scalar_type if it's 1 element-wise vectype. */ |
7798 | else if (nloads == 1) | |
7799 | ltype = vectype; | |
7800 | ||
2de001ee RS |
7801 | if (slp) |
7802 | { | |
66c16fd9 RB |
7803 | /* For SLP permutation support we need to load the whole group, |
7804 | not only the number of vector stmts the permutation result | |
7805 | fits in. */ | |
b266b968 | 7806 | if (slp_perm) |
66c16fd9 | 7807 | { |
d9f21f6a RS |
7808 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7809 | variable VF. */ | |
7810 | unsigned int const_vf = vf.to_constant (); | |
4d694b27 | 7811 | ncopies = CEIL (group_size * const_vf, const_nunits); |
66c16fd9 RB |
7812 | dr_chain.create (ncopies); |
7813 | } | |
7814 | else | |
7815 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
7b5fc413 | 7816 | } |
4d694b27 | 7817 | unsigned int group_el = 0; |
e09b4c37 RB |
7818 | unsigned HOST_WIDE_INT |
7819 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
7d75abc8 MM |
7820 | for (j = 0; j < ncopies; j++) |
7821 | { | |
7b5fc413 | 7822 | if (nloads > 1) |
e09b4c37 RB |
7823 | vec_alloc (v, nloads); |
7824 | for (i = 0; i < nloads; i++) | |
7b5fc413 | 7825 | { |
e09b4c37 | 7826 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
b210f45f | 7827 | group_el * elsz + cst_offset); |
19986382 RB |
7828 | tree data_ref = build2 (MEM_REF, ltype, running_off, this_off); |
7829 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); | |
7830 | new_stmt = gimple_build_assign (make_ssa_name (ltype), data_ref); | |
e09b4c37 RB |
7831 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7832 | if (nloads > 1) | |
7833 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, | |
7834 | gimple_assign_lhs (new_stmt)); | |
7835 | ||
7836 | group_el += lnel; | |
7837 | if (! slp | |
7838 | || group_el == group_size) | |
7b5fc413 | 7839 | { |
e09b4c37 RB |
7840 | tree newoff = copy_ssa_name (running_off); |
7841 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
7842 | running_off, stride_step); | |
7b5fc413 RB |
7843 | vect_finish_stmt_generation (stmt, incr, gsi); |
7844 | ||
7845 | running_off = newoff; | |
e09b4c37 | 7846 | group_el = 0; |
7b5fc413 | 7847 | } |
7b5fc413 | 7848 | } |
e09b4c37 | 7849 | if (nloads > 1) |
7d75abc8 | 7850 | { |
ea60dd34 RB |
7851 | tree vec_inv = build_constructor (lvectype, v); |
7852 | new_temp = vect_init_vector (stmt, vec_inv, lvectype, gsi); | |
e09b4c37 | 7853 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
ea60dd34 RB |
7854 | if (lvectype != vectype) |
7855 | { | |
7856 | new_stmt = gimple_build_assign (make_ssa_name (vectype), | |
7857 | VIEW_CONVERT_EXPR, | |
7858 | build1 (VIEW_CONVERT_EXPR, | |
7859 | vectype, new_temp)); | |
7860 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7861 | } | |
7d75abc8 MM |
7862 | } |
7863 | ||
7b5fc413 | 7864 | if (slp) |
b266b968 | 7865 | { |
b266b968 RB |
7866 | if (slp_perm) |
7867 | dr_chain.quick_push (gimple_assign_lhs (new_stmt)); | |
66c16fd9 RB |
7868 | else |
7869 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
b266b968 | 7870 | } |
7d75abc8 | 7871 | else |
225ce44b RB |
7872 | { |
7873 | if (j == 0) | |
7874 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7875 | else | |
7876 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7877 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7878 | } | |
7d75abc8 | 7879 | } |
b266b968 | 7880 | if (slp_perm) |
29afecdf RB |
7881 | { |
7882 | unsigned n_perms; | |
7883 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
7884 | slp_node_instance, false, &n_perms); | |
7885 | } | |
7d75abc8 MM |
7886 | return true; |
7887 | } | |
aec7ae7d | 7888 | |
b5ec4de7 RS |
7889 | if (memory_access_type == VMAT_GATHER_SCATTER |
7890 | || (!slp && memory_access_type == VMAT_CONTIGUOUS)) | |
ab2fc782 RS |
7891 | grouped_load = false; |
7892 | ||
0d0293ac | 7893 | if (grouped_load) |
ebfd146a | 7894 | { |
2c53b149 RB |
7895 | first_stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
7896 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
4f0a0218 | 7897 | /* For SLP vectorization we directly vectorize a subchain |
52eab378 RB |
7898 | without permutation. */ |
7899 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
4f0a0218 RB |
7900 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
7901 | /* For BB vectorization always use the first stmt to base | |
7902 | the data ref pointer on. */ | |
7903 | if (bb_vinfo) | |
7904 | first_stmt_for_drptr = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 7905 | |
ebfd146a | 7906 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
7907 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
7908 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
7909 | ??? But we can only do so if there is exactly one | |
7910 | as we have no way to get at the rest. Leave the CSE | |
7911 | opportunity alone. | |
7912 | ??? With the group load eventually participating | |
7913 | in multiple different permutations (having multiple | |
7914 | slp nodes which refer to the same group) the CSE | |
7915 | is even wrong code. See PR56270. */ | |
7916 | && !slp) | |
ebfd146a IR |
7917 | { |
7918 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
7919 | return true; | |
7920 | } | |
7921 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
9b999e8c | 7922 | group_gap_adj = 0; |
ebfd146a IR |
7923 | |
7924 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
7925 | if (slp) | |
7926 | { | |
0d0293ac | 7927 | grouped_load = false; |
91ff1504 RB |
7928 | /* For SLP permutation support we need to load the whole group, |
7929 | not only the number of vector stmts the permutation result | |
7930 | fits in. */ | |
7931 | if (slp_perm) | |
b267968e | 7932 | { |
d9f21f6a RS |
7933 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7934 | variable VF. */ | |
7935 | unsigned int const_vf = vf.to_constant (); | |
4d694b27 RS |
7936 | unsigned int const_nunits = nunits.to_constant (); |
7937 | vec_num = CEIL (group_size * const_vf, const_nunits); | |
b267968e RB |
7938 | group_gap_adj = vf * group_size - nunits * vec_num; |
7939 | } | |
91ff1504 | 7940 | else |
b267968e RB |
7941 | { |
7942 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
796bd467 RB |
7943 | group_gap_adj |
7944 | = group_size - SLP_INSTANCE_GROUP_SIZE (slp_node_instance); | |
b267968e | 7945 | } |
a70d6342 | 7946 | } |
ebfd146a | 7947 | else |
9b999e8c | 7948 | vec_num = group_size; |
44fc7854 BE |
7949 | |
7950 | ref_type = get_group_alias_ptr_type (first_stmt); | |
ebfd146a IR |
7951 | } |
7952 | else | |
7953 | { | |
7954 | first_stmt = stmt; | |
7955 | first_dr = dr; | |
7956 | group_size = vec_num = 1; | |
9b999e8c | 7957 | group_gap_adj = 0; |
44fc7854 | 7958 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
ebfd146a IR |
7959 | } |
7960 | ||
720f5239 | 7961 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 7962 | gcc_assert (alignment_support_scheme); |
70088b95 RS |
7963 | vec_loop_masks *loop_masks |
7964 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) | |
7965 | ? &LOOP_VINFO_MASKS (loop_vinfo) | |
7966 | : NULL); | |
7cfb4d93 RS |
7967 | /* Targets with store-lane instructions must not require explicit |
7968 | realignment. vect_supportable_dr_alignment always returns either | |
7969 | dr_aligned or dr_unaligned_supported for masked operations. */ | |
7970 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES | |
7971 | && !mask | |
70088b95 | 7972 | && !loop_masks) |
272c6793 RS |
7973 | || alignment_support_scheme == dr_aligned |
7974 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
7975 | |
7976 | /* In case the vectorization factor (VF) is bigger than the number | |
7977 | of elements that we can fit in a vectype (nunits), we have to generate | |
7978 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 7979 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 7980 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 7981 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 7982 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
7983 | stmts that use the defs of the current stmt. The example below |
7984 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
7985 | need to create 4 vectorized stmts): | |
ebfd146a IR |
7986 | |
7987 | before vectorization: | |
7988 | RELATED_STMT VEC_STMT | |
7989 | S1: x = memref - - | |
7990 | S2: z = x + 1 - - | |
7991 | ||
7992 | step 1: vectorize stmt S1: | |
7993 | We first create the vector stmt VS1_0, and, as usual, record a | |
7994 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
7995 | Next, we create the vector stmt VS1_1, and record a pointer to | |
7996 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 7997 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
7998 | stmts and pointers: |
7999 | RELATED_STMT VEC_STMT | |
8000 | VS1_0: vx0 = memref0 VS1_1 - | |
8001 | VS1_1: vx1 = memref1 VS1_2 - | |
8002 | VS1_2: vx2 = memref2 VS1_3 - | |
8003 | VS1_3: vx3 = memref3 - - | |
8004 | S1: x = load - VS1_0 | |
8005 | S2: z = x + 1 - - | |
8006 | ||
b8698a0f L |
8007 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
8008 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
8009 | stmt S2. */ |
8010 | ||
0d0293ac | 8011 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
8012 | |
8013 | S1: x2 = &base + 2 | |
8014 | S2: x0 = &base | |
8015 | S3: x1 = &base + 1 | |
8016 | S4: x3 = &base + 3 | |
8017 | ||
b8698a0f | 8018 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
8019 | starting from the access of the first stmt of the chain: |
8020 | ||
8021 | VS1: vx0 = &base | |
8022 | VS2: vx1 = &base + vec_size*1 | |
8023 | VS3: vx3 = &base + vec_size*2 | |
8024 | VS4: vx4 = &base + vec_size*3 | |
8025 | ||
8026 | Then permutation statements are generated: | |
8027 | ||
e2c83630 RH |
8028 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
8029 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
8030 | ... |
8031 | ||
8032 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
8033 | (the order of the data-refs in the output of vect_permute_load_chain | |
8034 | corresponds to the order of scalar stmts in the interleaving chain - see | |
8035 | the documentation of vect_permute_load_chain()). | |
8036 | The generation of permutation stmts and recording them in | |
0d0293ac | 8037 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 8038 | |
b8698a0f | 8039 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
8040 | permutation stmts above are created for every copy. The result vector |
8041 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
8042 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
8043 | |
8044 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
8045 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
8046 | we generate the following code: | |
8047 | p = initial_addr; | |
8048 | indx = 0; | |
8049 | loop { | |
8050 | p = p + indx * vectype_size; | |
8051 | vec_dest = *(p); | |
8052 | indx = indx + 1; | |
8053 | } | |
8054 | ||
8055 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 8056 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
8057 | then generate the following code, in which the data in each iteration is |
8058 | obtained by two vector loads, one from the previous iteration, and one | |
8059 | from the current iteration: | |
8060 | p1 = initial_addr; | |
8061 | msq_init = *(floor(p1)) | |
8062 | p2 = initial_addr + VS - 1; | |
8063 | realignment_token = call target_builtin; | |
8064 | indx = 0; | |
8065 | loop { | |
8066 | p2 = p2 + indx * vectype_size | |
8067 | lsq = *(floor(p2)) | |
8068 | vec_dest = realign_load (msq, lsq, realignment_token) | |
8069 | indx = indx + 1; | |
8070 | msq = lsq; | |
8071 | } */ | |
8072 | ||
8073 | /* If the misalignment remains the same throughout the execution of the | |
8074 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 8075 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
8076 | This can only occur when vectorizing memory accesses in the inner-loop |
8077 | nested within an outer-loop that is being vectorized. */ | |
8078 | ||
d1e4b493 | 8079 | if (nested_in_vect_loop |
cf098191 RS |
8080 | && !multiple_p (DR_STEP_ALIGNMENT (dr), |
8081 | GET_MODE_SIZE (TYPE_MODE (vectype)))) | |
ebfd146a IR |
8082 | { |
8083 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
8084 | compute_in_loop = true; | |
8085 | } | |
8086 | ||
8087 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
8088 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 8089 | && !compute_in_loop) |
ebfd146a IR |
8090 | { |
8091 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
8092 | alignment_support_scheme, NULL_TREE, | |
8093 | &at_loop); | |
8094 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
8095 | { | |
538dd0b7 | 8096 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
8097 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
8098 | size_one_node); | |
ebfd146a IR |
8099 | } |
8100 | } | |
8101 | else | |
8102 | at_loop = loop; | |
8103 | ||
62da9e14 | 8104 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
a1e53f3f L |
8105 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
8106 | ||
ab2fc782 RS |
8107 | tree bump; |
8108 | tree vec_offset = NULL_TREE; | |
8109 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
8110 | { | |
8111 | aggr_type = NULL_TREE; | |
8112 | bump = NULL_TREE; | |
8113 | } | |
8114 | else if (memory_access_type == VMAT_GATHER_SCATTER) | |
8115 | { | |
8116 | aggr_type = elem_type; | |
8117 | vect_get_strided_load_store_ops (stmt, loop_vinfo, &gs_info, | |
8118 | &bump, &vec_offset); | |
8119 | } | |
272c6793 | 8120 | else |
ab2fc782 RS |
8121 | { |
8122 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
8123 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
8124 | else | |
8125 | aggr_type = vectype; | |
8126 | bump = vect_get_data_ptr_increment (dr, aggr_type, memory_access_type); | |
8127 | } | |
272c6793 | 8128 | |
c3a8f964 | 8129 | tree vec_mask = NULL_TREE; |
ebfd146a | 8130 | prev_stmt_info = NULL; |
4d694b27 | 8131 | poly_uint64 group_elt = 0; |
ebfd146a | 8132 | for (j = 0; j < ncopies; j++) |
b8698a0f | 8133 | { |
272c6793 | 8134 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 8135 | if (j == 0) |
74bf76ed JJ |
8136 | { |
8137 | bool simd_lane_access_p | |
8138 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
8139 | if (simd_lane_access_p | |
8140 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
8141 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
8142 | && integer_zerop (DR_OFFSET (first_dr)) | |
8143 | && integer_zerop (DR_INIT (first_dr)) | |
8144 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 8145 | get_alias_set (TREE_TYPE (ref_type))) |
74bf76ed JJ |
8146 | && (alignment_support_scheme == dr_aligned |
8147 | || alignment_support_scheme == dr_unaligned_supported)) | |
8148 | { | |
8149 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 8150 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 8151 | inv_p = false; |
74bf76ed | 8152 | } |
4f0a0218 RB |
8153 | else if (first_stmt_for_drptr |
8154 | && first_stmt != first_stmt_for_drptr) | |
8155 | { | |
8156 | dataref_ptr | |
8157 | = vect_create_data_ref_ptr (first_stmt_for_drptr, aggr_type, | |
8158 | at_loop, offset, &dummy, gsi, | |
8159 | &ptr_incr, simd_lane_access_p, | |
ab2fc782 | 8160 | &inv_p, byte_offset, bump); |
4f0a0218 RB |
8161 | /* Adjust the pointer by the difference to first_stmt. */ |
8162 | data_reference_p ptrdr | |
8163 | = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt_for_drptr)); | |
8164 | tree diff = fold_convert (sizetype, | |
8165 | size_binop (MINUS_EXPR, | |
8166 | DR_INIT (first_dr), | |
8167 | DR_INIT (ptrdr))); | |
8168 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
8169 | stmt, diff); | |
8170 | } | |
bfaa08b7 RS |
8171 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
8172 | { | |
8173 | vect_get_gather_scatter_ops (loop, stmt, &gs_info, | |
8174 | &dataref_ptr, &vec_offset); | |
8175 | inv_p = false; | |
8176 | } | |
74bf76ed JJ |
8177 | else |
8178 | dataref_ptr | |
8179 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
8180 | offset, &dummy, gsi, &ptr_incr, | |
356bbc4c | 8181 | simd_lane_access_p, &inv_p, |
ab2fc782 | 8182 | byte_offset, bump); |
c3a8f964 RS |
8183 | if (mask) |
8184 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, | |
8185 | mask_vectype); | |
74bf76ed | 8186 | } |
ebfd146a | 8187 | else |
c3a8f964 RS |
8188 | { |
8189 | if (dataref_offset) | |
8190 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
ab2fc782 | 8191 | bump); |
bfaa08b7 | 8192 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
929b4411 RS |
8193 | vec_offset = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
8194 | vec_offset); | |
c3a8f964 | 8195 | else |
ab2fc782 RS |
8196 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8197 | stmt, bump); | |
c3a8f964 | 8198 | if (mask) |
929b4411 | 8199 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
c3a8f964 | 8200 | } |
ebfd146a | 8201 | |
0d0293ac | 8202 | if (grouped_load || slp_perm) |
9771b263 | 8203 | dr_chain.create (vec_num); |
5ce1ee7f | 8204 | |
2de001ee | 8205 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 8206 | { |
272c6793 RS |
8207 | tree vec_array; |
8208 | ||
8209 | vec_array = create_vector_array (vectype, vec_num); | |
8210 | ||
7cfb4d93 | 8211 | tree final_mask = NULL_TREE; |
70088b95 RS |
8212 | if (loop_masks) |
8213 | final_mask = vect_get_loop_mask (gsi, loop_masks, ncopies, | |
8214 | vectype, j); | |
7cfb4d93 RS |
8215 | if (vec_mask) |
8216 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
8217 | vec_mask, gsi); | |
8218 | ||
7e11fc7f | 8219 | gcall *call; |
7cfb4d93 | 8220 | if (final_mask) |
7e11fc7f RS |
8221 | { |
8222 | /* Emit: | |
8223 | VEC_ARRAY = MASK_LOAD_LANES (DATAREF_PTR, ALIAS_PTR, | |
8224 | VEC_MASK). */ | |
8225 | unsigned int align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype)); | |
8226 | tree alias_ptr = build_int_cst (ref_type, align); | |
8227 | call = gimple_build_call_internal (IFN_MASK_LOAD_LANES, 3, | |
8228 | dataref_ptr, alias_ptr, | |
7cfb4d93 | 8229 | final_mask); |
7e11fc7f RS |
8230 | } |
8231 | else | |
8232 | { | |
8233 | /* Emit: | |
8234 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
8235 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); | |
8236 | call = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
8237 | } | |
a844293d RS |
8238 | gimple_call_set_lhs (call, vec_array); |
8239 | gimple_call_set_nothrow (call, true); | |
8240 | new_stmt = call; | |
272c6793 | 8241 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a | 8242 | |
272c6793 RS |
8243 | /* Extract each vector into an SSA_NAME. */ |
8244 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 8245 | { |
272c6793 RS |
8246 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
8247 | vec_array, i); | |
9771b263 | 8248 | dr_chain.quick_push (new_temp); |
272c6793 RS |
8249 | } |
8250 | ||
8251 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 8252 | vect_record_grouped_load_vectors (stmt, dr_chain); |
3ba4ff41 RS |
8253 | |
8254 | /* Record that VEC_ARRAY is now dead. */ | |
8255 | vect_clobber_variable (stmt, gsi, vec_array); | |
272c6793 RS |
8256 | } |
8257 | else | |
8258 | { | |
8259 | for (i = 0; i < vec_num; i++) | |
8260 | { | |
7cfb4d93 | 8261 | tree final_mask = NULL_TREE; |
70088b95 | 8262 | if (loop_masks |
7cfb4d93 | 8263 | && memory_access_type != VMAT_INVARIANT) |
70088b95 RS |
8264 | final_mask = vect_get_loop_mask (gsi, loop_masks, |
8265 | vec_num * ncopies, | |
7cfb4d93 RS |
8266 | vectype, vec_num * j + i); |
8267 | if (vec_mask) | |
8268 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
8269 | vec_mask, gsi); | |
8270 | ||
272c6793 RS |
8271 | if (i > 0) |
8272 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
ab2fc782 | 8273 | stmt, bump); |
272c6793 RS |
8274 | |
8275 | /* 2. Create the vector-load in the loop. */ | |
8276 | switch (alignment_support_scheme) | |
8277 | { | |
8278 | case dr_aligned: | |
8279 | case dr_unaligned_supported: | |
be1ac4ec | 8280 | { |
644ffefd MJ |
8281 | unsigned int align, misalign; |
8282 | ||
bfaa08b7 RS |
8283 | if (memory_access_type == VMAT_GATHER_SCATTER) |
8284 | { | |
8285 | tree scale = size_int (gs_info.scale); | |
8286 | gcall *call; | |
70088b95 | 8287 | if (loop_masks) |
bfaa08b7 RS |
8288 | call = gimple_build_call_internal |
8289 | (IFN_MASK_GATHER_LOAD, 4, dataref_ptr, | |
8290 | vec_offset, scale, final_mask); | |
8291 | else | |
8292 | call = gimple_build_call_internal | |
8293 | (IFN_GATHER_LOAD, 3, dataref_ptr, | |
8294 | vec_offset, scale); | |
8295 | gimple_call_set_nothrow (call, true); | |
8296 | new_stmt = call; | |
8297 | data_ref = NULL_TREE; | |
8298 | break; | |
8299 | } | |
8300 | ||
f702e7d4 | 8301 | align = DR_TARGET_ALIGNMENT (dr); |
272c6793 RS |
8302 | if (alignment_support_scheme == dr_aligned) |
8303 | { | |
8304 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 8305 | misalign = 0; |
272c6793 RS |
8306 | } |
8307 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
8308 | { | |
25f68d90 | 8309 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 8310 | misalign = 0; |
272c6793 RS |
8311 | } |
8312 | else | |
c3a8f964 | 8313 | misalign = DR_MISALIGNMENT (first_dr); |
aed93b23 RB |
8314 | if (dataref_offset == NULL_TREE |
8315 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
8316 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
8317 | align, misalign); | |
c3a8f964 | 8318 | |
7cfb4d93 | 8319 | if (final_mask) |
c3a8f964 RS |
8320 | { |
8321 | align = least_bit_hwi (misalign | align); | |
8322 | tree ptr = build_int_cst (ref_type, align); | |
8323 | gcall *call | |
8324 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, | |
8325 | dataref_ptr, ptr, | |
7cfb4d93 | 8326 | final_mask); |
c3a8f964 RS |
8327 | gimple_call_set_nothrow (call, true); |
8328 | new_stmt = call; | |
8329 | data_ref = NULL_TREE; | |
8330 | } | |
8331 | else | |
8332 | { | |
8333 | data_ref | |
8334 | = fold_build2 (MEM_REF, vectype, dataref_ptr, | |
8335 | dataref_offset | |
8336 | ? dataref_offset | |
8337 | : build_int_cst (ref_type, 0)); | |
8338 | if (alignment_support_scheme == dr_aligned) | |
8339 | ; | |
8340 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
8341 | TREE_TYPE (data_ref) | |
8342 | = build_aligned_type (TREE_TYPE (data_ref), | |
8343 | align * BITS_PER_UNIT); | |
8344 | else | |
8345 | TREE_TYPE (data_ref) | |
8346 | = build_aligned_type (TREE_TYPE (data_ref), | |
8347 | TYPE_ALIGN (elem_type)); | |
8348 | } | |
272c6793 | 8349 | break; |
be1ac4ec | 8350 | } |
272c6793 | 8351 | case dr_explicit_realign: |
267d3070 | 8352 | { |
272c6793 | 8353 | tree ptr, bump; |
272c6793 | 8354 | |
d88981fc | 8355 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
8356 | |
8357 | if (compute_in_loop) | |
8358 | msq = vect_setup_realignment (first_stmt, gsi, | |
8359 | &realignment_token, | |
8360 | dr_explicit_realign, | |
8361 | dataref_ptr, NULL); | |
8362 | ||
aed93b23 RB |
8363 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
8364 | ptr = copy_ssa_name (dataref_ptr); | |
8365 | else | |
8366 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
f702e7d4 | 8367 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); |
0d0e4a03 JJ |
8368 | new_stmt = gimple_build_assign |
8369 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
8370 | build_int_cst |
8371 | (TREE_TYPE (dataref_ptr), | |
f702e7d4 | 8372 | -(HOST_WIDE_INT) align)); |
272c6793 RS |
8373 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
8374 | data_ref | |
8375 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 8376 | build_int_cst (ref_type, 0)); |
19986382 | 8377 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); |
272c6793 RS |
8378 | vec_dest = vect_create_destination_var (scalar_dest, |
8379 | vectype); | |
8380 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
8381 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
8382 | gimple_assign_set_lhs (new_stmt, new_temp); | |
8383 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
8384 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
8385 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8386 | msq = new_temp; | |
8387 | ||
d88981fc | 8388 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 8389 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 8390 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 8391 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
8392 | new_stmt = gimple_build_assign |
8393 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 | 8394 | build_int_cst |
f702e7d4 | 8395 | (TREE_TYPE (ptr), -(HOST_WIDE_INT) align)); |
aed93b23 | 8396 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
8397 | gimple_assign_set_lhs (new_stmt, ptr); |
8398 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8399 | data_ref | |
8400 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 8401 | build_int_cst (ref_type, 0)); |
272c6793 | 8402 | break; |
267d3070 | 8403 | } |
272c6793 | 8404 | case dr_explicit_realign_optimized: |
f702e7d4 RS |
8405 | { |
8406 | if (TREE_CODE (dataref_ptr) == SSA_NAME) | |
8407 | new_temp = copy_ssa_name (dataref_ptr); | |
8408 | else | |
8409 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
8410 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); | |
8411 | new_stmt = gimple_build_assign | |
8412 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
8413 | build_int_cst (TREE_TYPE (dataref_ptr), | |
8414 | -(HOST_WIDE_INT) align)); | |
8415 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8416 | data_ref | |
8417 | = build2 (MEM_REF, vectype, new_temp, | |
8418 | build_int_cst (ref_type, 0)); | |
8419 | break; | |
8420 | } | |
272c6793 RS |
8421 | default: |
8422 | gcc_unreachable (); | |
8423 | } | |
ebfd146a | 8424 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
c3a8f964 RS |
8425 | /* DATA_REF is null if we've already built the statement. */ |
8426 | if (data_ref) | |
19986382 RB |
8427 | { |
8428 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); | |
8429 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
8430 | } | |
ebfd146a | 8431 | new_temp = make_ssa_name (vec_dest, new_stmt); |
c3a8f964 | 8432 | gimple_set_lhs (new_stmt, new_temp); |
ebfd146a IR |
8433 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
8434 | ||
272c6793 RS |
8435 | /* 3. Handle explicit realignment if necessary/supported. |
8436 | Create in loop: | |
8437 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
8438 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
8439 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 8440 | { |
272c6793 RS |
8441 | lsq = gimple_assign_lhs (new_stmt); |
8442 | if (!realignment_token) | |
8443 | realignment_token = dataref_ptr; | |
8444 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
8445 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
8446 | msq, lsq, realignment_token); | |
272c6793 RS |
8447 | new_temp = make_ssa_name (vec_dest, new_stmt); |
8448 | gimple_assign_set_lhs (new_stmt, new_temp); | |
8449 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8450 | ||
8451 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
8452 | { | |
8453 | gcc_assert (phi); | |
8454 | if (i == vec_num - 1 && j == ncopies - 1) | |
8455 | add_phi_arg (phi, lsq, | |
8456 | loop_latch_edge (containing_loop), | |
9e227d60 | 8457 | UNKNOWN_LOCATION); |
272c6793 RS |
8458 | msq = lsq; |
8459 | } | |
ebfd146a | 8460 | } |
ebfd146a | 8461 | |
59fd17e3 RB |
8462 | /* 4. Handle invariant-load. */ |
8463 | if (inv_p && !bb_vinfo) | |
8464 | { | |
59fd17e3 | 8465 | gcc_assert (!grouped_load); |
d1417442 JJ |
8466 | /* If we have versioned for aliasing or the loop doesn't |
8467 | have any data dependencies that would preclude this, | |
8468 | then we are sure this is a loop invariant load and | |
8469 | thus we can insert it on the preheader edge. */ | |
8470 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
8471 | && !nested_in_vect_loop | |
6b916b36 | 8472 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
8473 | { |
8474 | if (dump_enabled_p ()) | |
8475 | { | |
8476 | dump_printf_loc (MSG_NOTE, vect_location, | |
8477 | "hoisting out of the vectorized " | |
8478 | "loop: "); | |
8479 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 8480 | } |
b731b390 | 8481 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
8482 | gsi_insert_on_edge_immediate |
8483 | (loop_preheader_edge (loop), | |
8484 | gimple_build_assign (tem, | |
8485 | unshare_expr | |
8486 | (gimple_assign_rhs1 (stmt)))); | |
8487 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
34cd48e5 | 8488 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
4fbeb363 | 8489 | vinfo->add_stmt (new_stmt); |
a0e35eb0 RB |
8490 | } |
8491 | else | |
8492 | { | |
8493 | gimple_stmt_iterator gsi2 = *gsi; | |
8494 | gsi_next (&gsi2); | |
8495 | new_temp = vect_init_vector (stmt, scalar_dest, | |
8496 | vectype, &gsi2); | |
34cd48e5 | 8497 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
a0e35eb0 | 8498 | } |
59fd17e3 RB |
8499 | } |
8500 | ||
62da9e14 | 8501 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
272c6793 | 8502 | { |
aec7ae7d JJ |
8503 | tree perm_mask = perm_mask_for_reverse (vectype); |
8504 | new_temp = permute_vec_elements (new_temp, new_temp, | |
8505 | perm_mask, stmt, gsi); | |
ebfd146a IR |
8506 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
8507 | } | |
267d3070 | 8508 | |
272c6793 | 8509 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
8510 | vect_transform_grouped_load (). */ |
8511 | if (grouped_load || slp_perm) | |
9771b263 | 8512 | dr_chain.quick_push (new_temp); |
267d3070 | 8513 | |
272c6793 RS |
8514 | /* Store vector loads in the corresponding SLP_NODE. */ |
8515 | if (slp && !slp_perm) | |
9771b263 | 8516 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
b267968e RB |
8517 | |
8518 | /* With SLP permutation we load the gaps as well, without | |
8519 | we need to skip the gaps after we manage to fully load | |
2c53b149 | 8520 | all elements. group_gap_adj is DR_GROUP_SIZE here. */ |
b267968e | 8521 | group_elt += nunits; |
d9f21f6a RS |
8522 | if (maybe_ne (group_gap_adj, 0U) |
8523 | && !slp_perm | |
8524 | && known_eq (group_elt, group_size - group_gap_adj)) | |
b267968e | 8525 | { |
d9f21f6a RS |
8526 | poly_wide_int bump_val |
8527 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
8528 | * group_gap_adj); | |
8e6cdc90 | 8529 | tree bump = wide_int_to_tree (sizetype, bump_val); |
b267968e RB |
8530 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8531 | stmt, bump); | |
8532 | group_elt = 0; | |
8533 | } | |
272c6793 | 8534 | } |
9b999e8c RB |
8535 | /* Bump the vector pointer to account for a gap or for excess |
8536 | elements loaded for a permuted SLP load. */ | |
d9f21f6a | 8537 | if (maybe_ne (group_gap_adj, 0U) && slp_perm) |
a64b9c26 | 8538 | { |
d9f21f6a RS |
8539 | poly_wide_int bump_val |
8540 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
8541 | * group_gap_adj); | |
8e6cdc90 | 8542 | tree bump = wide_int_to_tree (sizetype, bump_val); |
a64b9c26 RB |
8543 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8544 | stmt, bump); | |
8545 | } | |
ebfd146a IR |
8546 | } |
8547 | ||
8548 | if (slp && !slp_perm) | |
8549 | continue; | |
8550 | ||
8551 | if (slp_perm) | |
8552 | { | |
29afecdf | 8553 | unsigned n_perms; |
01d8bf07 | 8554 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
29afecdf RB |
8555 | slp_node_instance, false, |
8556 | &n_perms)) | |
ebfd146a | 8557 | { |
9771b263 | 8558 | dr_chain.release (); |
ebfd146a IR |
8559 | return false; |
8560 | } | |
8561 | } | |
8562 | else | |
8563 | { | |
0d0293ac | 8564 | if (grouped_load) |
ebfd146a | 8565 | { |
2de001ee | 8566 | if (memory_access_type != VMAT_LOAD_STORE_LANES) |
0d0293ac | 8567 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 8568 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
8569 | } |
8570 | else | |
8571 | { | |
8572 | if (j == 0) | |
8573 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
8574 | else | |
8575 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
8576 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
8577 | } | |
8578 | } | |
9771b263 | 8579 | dr_chain.release (); |
ebfd146a IR |
8580 | } |
8581 | ||
ebfd146a IR |
8582 | return true; |
8583 | } | |
8584 | ||
8585 | /* Function vect_is_simple_cond. | |
b8698a0f | 8586 | |
ebfd146a IR |
8587 | Input: |
8588 | LOOP - the loop that is being vectorized. | |
8589 | COND - Condition that is checked for simple use. | |
8590 | ||
e9e1d143 RG |
8591 | Output: |
8592 | *COMP_VECTYPE - the vector type for the comparison. | |
4fc5ebf1 | 8593 | *DTS - The def types for the arguments of the comparison |
e9e1d143 | 8594 | |
ebfd146a IR |
8595 | Returns whether a COND can be vectorized. Checks whether |
8596 | condition operands are supportable using vec_is_simple_use. */ | |
8597 | ||
87aab9b2 | 8598 | static bool |
4fc5ebf1 | 8599 | vect_is_simple_cond (tree cond, vec_info *vinfo, |
8da4c8d8 RB |
8600 | tree *comp_vectype, enum vect_def_type *dts, |
8601 | tree vectype) | |
ebfd146a IR |
8602 | { |
8603 | tree lhs, rhs; | |
e9e1d143 | 8604 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 8605 | |
a414c77f IE |
8606 | /* Mask case. */ |
8607 | if (TREE_CODE (cond) == SSA_NAME | |
2568d8a1 | 8608 | && VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (cond))) |
a414c77f | 8609 | { |
894dd753 | 8610 | if (!vect_is_simple_use (cond, vinfo, &dts[0], comp_vectype) |
a414c77f IE |
8611 | || !*comp_vectype |
8612 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
8613 | return false; | |
8614 | return true; | |
8615 | } | |
8616 | ||
ebfd146a IR |
8617 | if (!COMPARISON_CLASS_P (cond)) |
8618 | return false; | |
8619 | ||
8620 | lhs = TREE_OPERAND (cond, 0); | |
8621 | rhs = TREE_OPERAND (cond, 1); | |
8622 | ||
8623 | if (TREE_CODE (lhs) == SSA_NAME) | |
8624 | { | |
894dd753 | 8625 | if (!vect_is_simple_use (lhs, vinfo, &dts[0], &vectype1)) |
ebfd146a IR |
8626 | return false; |
8627 | } | |
4fc5ebf1 JG |
8628 | else if (TREE_CODE (lhs) == INTEGER_CST || TREE_CODE (lhs) == REAL_CST |
8629 | || TREE_CODE (lhs) == FIXED_CST) | |
8630 | dts[0] = vect_constant_def; | |
8631 | else | |
ebfd146a IR |
8632 | return false; |
8633 | ||
8634 | if (TREE_CODE (rhs) == SSA_NAME) | |
8635 | { | |
894dd753 | 8636 | if (!vect_is_simple_use (rhs, vinfo, &dts[1], &vectype2)) |
ebfd146a IR |
8637 | return false; |
8638 | } | |
4fc5ebf1 JG |
8639 | else if (TREE_CODE (rhs) == INTEGER_CST || TREE_CODE (rhs) == REAL_CST |
8640 | || TREE_CODE (rhs) == FIXED_CST) | |
8641 | dts[1] = vect_constant_def; | |
8642 | else | |
ebfd146a IR |
8643 | return false; |
8644 | ||
28b33016 | 8645 | if (vectype1 && vectype2 |
928686b1 RS |
8646 | && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1), |
8647 | TYPE_VECTOR_SUBPARTS (vectype2))) | |
28b33016 IE |
8648 | return false; |
8649 | ||
e9e1d143 | 8650 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
8da4c8d8 | 8651 | /* Invariant comparison. */ |
4515e413 | 8652 | if (! *comp_vectype && vectype) |
8da4c8d8 RB |
8653 | { |
8654 | tree scalar_type = TREE_TYPE (lhs); | |
8655 | /* If we can widen the comparison to match vectype do so. */ | |
8656 | if (INTEGRAL_TYPE_P (scalar_type) | |
8657 | && tree_int_cst_lt (TYPE_SIZE (scalar_type), | |
8658 | TYPE_SIZE (TREE_TYPE (vectype)))) | |
8659 | scalar_type = build_nonstandard_integer_type | |
8660 | (tree_to_uhwi (TYPE_SIZE (TREE_TYPE (vectype))), | |
8661 | TYPE_UNSIGNED (scalar_type)); | |
8662 | *comp_vectype = get_vectype_for_scalar_type (scalar_type); | |
8663 | } | |
8664 | ||
ebfd146a IR |
8665 | return true; |
8666 | } | |
8667 | ||
8668 | /* vectorizable_condition. | |
8669 | ||
b8698a0f L |
8670 | Check if STMT is conditional modify expression that can be vectorized. |
8671 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
8672 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
8673 | at GSI. |
8674 | ||
8675 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
8676 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 8677 | else clause if it is 2). |
ebfd146a IR |
8678 | |
8679 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
8680 | ||
4bbe8262 | 8681 | bool |
355fe088 TS |
8682 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
8683 | gimple **vec_stmt, tree reduc_def, int reduc_index, | |
68435eb2 | 8684 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
ebfd146a IR |
8685 | { |
8686 | tree scalar_dest = NULL_TREE; | |
8687 | tree vec_dest = NULL_TREE; | |
01216d27 JJ |
8688 | tree cond_expr, cond_expr0 = NULL_TREE, cond_expr1 = NULL_TREE; |
8689 | tree then_clause, else_clause; | |
ebfd146a | 8690 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
df11cc78 | 8691 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
8692 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
8693 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
5958f9e2 | 8694 | tree vec_compare; |
ebfd146a IR |
8695 | tree new_temp; |
8696 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4fc5ebf1 JG |
8697 | enum vect_def_type dts[4] |
8698 | = {vect_unknown_def_type, vect_unknown_def_type, | |
8699 | vect_unknown_def_type, vect_unknown_def_type}; | |
8700 | int ndts = 4; | |
f7e531cf | 8701 | int ncopies; |
01216d27 | 8702 | enum tree_code code, cond_code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
a855b1b1 | 8703 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
8704 | int i, j; |
8705 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
8706 | vec<tree> vec_oprnds0 = vNULL; |
8707 | vec<tree> vec_oprnds1 = vNULL; | |
8708 | vec<tree> vec_oprnds2 = vNULL; | |
8709 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 8710 | tree vec_cmp_type; |
a414c77f | 8711 | bool masked = false; |
b8698a0f | 8712 | |
f7e531cf IR |
8713 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
8714 | return false; | |
8715 | ||
bb6c2b68 RS |
8716 | vect_reduction_type reduction_type |
8717 | = STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info); | |
8718 | if (reduction_type == TREE_CODE_REDUCTION) | |
af29617a AH |
8719 | { |
8720 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
8721 | return false; | |
ebfd146a | 8722 | |
af29617a AH |
8723 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
8724 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
8725 | && reduc_def)) | |
8726 | return false; | |
ebfd146a | 8727 | |
af29617a AH |
8728 | /* FORNOW: not yet supported. */ |
8729 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
8730 | { | |
8731 | if (dump_enabled_p ()) | |
8732 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
8733 | "value used after loop.\n"); | |
8734 | return false; | |
8735 | } | |
ebfd146a IR |
8736 | } |
8737 | ||
8738 | /* Is vectorizable conditional operation? */ | |
8739 | if (!is_gimple_assign (stmt)) | |
8740 | return false; | |
8741 | ||
8742 | code = gimple_assign_rhs_code (stmt); | |
8743 | ||
8744 | if (code != COND_EXPR) | |
8745 | return false; | |
8746 | ||
465c8c19 | 8747 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2947d3b2 | 8748 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
465c8c19 | 8749 | |
fce57248 | 8750 | if (slp_node) |
465c8c19 JJ |
8751 | ncopies = 1; |
8752 | else | |
e8f142e2 | 8753 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
8754 | |
8755 | gcc_assert (ncopies >= 1); | |
8756 | if (reduc_index && ncopies > 1) | |
8757 | return false; /* FORNOW */ | |
8758 | ||
4e71066d RG |
8759 | cond_expr = gimple_assign_rhs1 (stmt); |
8760 | then_clause = gimple_assign_rhs2 (stmt); | |
8761 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 8762 | |
4fc5ebf1 | 8763 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, |
4515e413 | 8764 | &comp_vectype, &dts[0], slp_node ? NULL : vectype) |
e9e1d143 | 8765 | || !comp_vectype) |
ebfd146a IR |
8766 | return false; |
8767 | ||
894dd753 | 8768 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &dts[2], &vectype1)) |
2947d3b2 | 8769 | return false; |
894dd753 | 8770 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &dts[3], &vectype2)) |
ebfd146a | 8771 | return false; |
2947d3b2 IE |
8772 | |
8773 | if (vectype1 && !useless_type_conversion_p (vectype, vectype1)) | |
8774 | return false; | |
8775 | ||
8776 | if (vectype2 && !useless_type_conversion_p (vectype, vectype2)) | |
ebfd146a IR |
8777 | return false; |
8778 | ||
28b33016 IE |
8779 | masked = !COMPARISON_CLASS_P (cond_expr); |
8780 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
8781 | ||
74946978 MP |
8782 | if (vec_cmp_type == NULL_TREE) |
8783 | return false; | |
784fb9b3 | 8784 | |
01216d27 JJ |
8785 | cond_code = TREE_CODE (cond_expr); |
8786 | if (!masked) | |
8787 | { | |
8788 | cond_expr0 = TREE_OPERAND (cond_expr, 0); | |
8789 | cond_expr1 = TREE_OPERAND (cond_expr, 1); | |
8790 | } | |
8791 | ||
8792 | if (!masked && VECTOR_BOOLEAN_TYPE_P (comp_vectype)) | |
8793 | { | |
8794 | /* Boolean values may have another representation in vectors | |
8795 | and therefore we prefer bit operations over comparison for | |
8796 | them (which also works for scalar masks). We store opcodes | |
8797 | to use in bitop1 and bitop2. Statement is vectorized as | |
8798 | BITOP2 (rhs1 BITOP1 rhs2) or rhs1 BITOP2 (BITOP1 rhs2) | |
8799 | depending on bitop1 and bitop2 arity. */ | |
8800 | switch (cond_code) | |
8801 | { | |
8802 | case GT_EXPR: | |
8803 | bitop1 = BIT_NOT_EXPR; | |
8804 | bitop2 = BIT_AND_EXPR; | |
8805 | break; | |
8806 | case GE_EXPR: | |
8807 | bitop1 = BIT_NOT_EXPR; | |
8808 | bitop2 = BIT_IOR_EXPR; | |
8809 | break; | |
8810 | case LT_EXPR: | |
8811 | bitop1 = BIT_NOT_EXPR; | |
8812 | bitop2 = BIT_AND_EXPR; | |
8813 | std::swap (cond_expr0, cond_expr1); | |
8814 | break; | |
8815 | case LE_EXPR: | |
8816 | bitop1 = BIT_NOT_EXPR; | |
8817 | bitop2 = BIT_IOR_EXPR; | |
8818 | std::swap (cond_expr0, cond_expr1); | |
8819 | break; | |
8820 | case NE_EXPR: | |
8821 | bitop1 = BIT_XOR_EXPR; | |
8822 | break; | |
8823 | case EQ_EXPR: | |
8824 | bitop1 = BIT_XOR_EXPR; | |
8825 | bitop2 = BIT_NOT_EXPR; | |
8826 | break; | |
8827 | default: | |
8828 | return false; | |
8829 | } | |
8830 | cond_code = SSA_NAME; | |
8831 | } | |
8832 | ||
b8698a0f | 8833 | if (!vec_stmt) |
ebfd146a | 8834 | { |
01216d27 JJ |
8835 | if (bitop1 != NOP_EXPR) |
8836 | { | |
8837 | machine_mode mode = TYPE_MODE (comp_vectype); | |
8838 | optab optab; | |
8839 | ||
8840 | optab = optab_for_tree_code (bitop1, comp_vectype, optab_default); | |
8841 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8842 | return false; | |
8843 | ||
8844 | if (bitop2 != NOP_EXPR) | |
8845 | { | |
8846 | optab = optab_for_tree_code (bitop2, comp_vectype, | |
8847 | optab_default); | |
8848 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8849 | return false; | |
8850 | } | |
8851 | } | |
4fc5ebf1 JG |
8852 | if (expand_vec_cond_expr_p (vectype, comp_vectype, |
8853 | cond_code)) | |
8854 | { | |
68435eb2 RB |
8855 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; |
8856 | vect_model_simple_cost (stmt_info, ncopies, dts, ndts, slp_node, | |
8857 | cost_vec); | |
4fc5ebf1 JG |
8858 | return true; |
8859 | } | |
8860 | return false; | |
ebfd146a IR |
8861 | } |
8862 | ||
f7e531cf IR |
8863 | /* Transform. */ |
8864 | ||
8865 | if (!slp_node) | |
8866 | { | |
9771b263 DN |
8867 | vec_oprnds0.create (1); |
8868 | vec_oprnds1.create (1); | |
8869 | vec_oprnds2.create (1); | |
8870 | vec_oprnds3.create (1); | |
f7e531cf | 8871 | } |
ebfd146a IR |
8872 | |
8873 | /* Handle def. */ | |
8874 | scalar_dest = gimple_assign_lhs (stmt); | |
bb6c2b68 RS |
8875 | if (reduction_type != EXTRACT_LAST_REDUCTION) |
8876 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
ebfd146a IR |
8877 | |
8878 | /* Handle cond expr. */ | |
a855b1b1 MM |
8879 | for (j = 0; j < ncopies; j++) |
8880 | { | |
bb6c2b68 | 8881 | gimple *new_stmt = NULL; |
a855b1b1 MM |
8882 | if (j == 0) |
8883 | { | |
f7e531cf IR |
8884 | if (slp_node) |
8885 | { | |
00f96dc9 TS |
8886 | auto_vec<tree, 4> ops; |
8887 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 8888 | |
a414c77f | 8889 | if (masked) |
01216d27 | 8890 | ops.safe_push (cond_expr); |
a414c77f IE |
8891 | else |
8892 | { | |
01216d27 JJ |
8893 | ops.safe_push (cond_expr0); |
8894 | ops.safe_push (cond_expr1); | |
a414c77f | 8895 | } |
9771b263 DN |
8896 | ops.safe_push (then_clause); |
8897 | ops.safe_push (else_clause); | |
306b0c92 | 8898 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
37b5ec8f JJ |
8899 | vec_oprnds3 = vec_defs.pop (); |
8900 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
8901 | if (!masked) |
8902 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 8903 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf IR |
8904 | } |
8905 | else | |
8906 | { | |
a414c77f IE |
8907 | if (masked) |
8908 | { | |
8909 | vec_cond_lhs | |
8910 | = vect_get_vec_def_for_operand (cond_expr, stmt, | |
8911 | comp_vectype); | |
894dd753 | 8912 | vect_is_simple_use (cond_expr, stmt_info->vinfo, &dts[0]); |
a414c77f IE |
8913 | } |
8914 | else | |
8915 | { | |
01216d27 JJ |
8916 | vec_cond_lhs |
8917 | = vect_get_vec_def_for_operand (cond_expr0, | |
8918 | stmt, comp_vectype); | |
894dd753 | 8919 | vect_is_simple_use (cond_expr0, loop_vinfo, &dts[0]); |
01216d27 JJ |
8920 | |
8921 | vec_cond_rhs | |
8922 | = vect_get_vec_def_for_operand (cond_expr1, | |
8923 | stmt, comp_vectype); | |
894dd753 | 8924 | vect_is_simple_use (cond_expr1, loop_vinfo, &dts[1]); |
a414c77f | 8925 | } |
f7e531cf IR |
8926 | if (reduc_index == 1) |
8927 | vec_then_clause = reduc_def; | |
8928 | else | |
8929 | { | |
8930 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 | 8931 | stmt); |
894dd753 | 8932 | vect_is_simple_use (then_clause, loop_vinfo, &dts[2]); |
f7e531cf IR |
8933 | } |
8934 | if (reduc_index == 2) | |
8935 | vec_else_clause = reduc_def; | |
8936 | else | |
8937 | { | |
8938 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 | 8939 | stmt); |
894dd753 | 8940 | vect_is_simple_use (else_clause, loop_vinfo, &dts[3]); |
f7e531cf | 8941 | } |
a855b1b1 MM |
8942 | } |
8943 | } | |
8944 | else | |
8945 | { | |
a414c77f IE |
8946 | vec_cond_lhs |
8947 | = vect_get_vec_def_for_stmt_copy (dts[0], | |
8948 | vec_oprnds0.pop ()); | |
8949 | if (!masked) | |
8950 | vec_cond_rhs | |
8951 | = vect_get_vec_def_for_stmt_copy (dts[1], | |
8952 | vec_oprnds1.pop ()); | |
8953 | ||
a855b1b1 | 8954 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 8955 | vec_oprnds2.pop ()); |
a855b1b1 | 8956 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 8957 | vec_oprnds3.pop ()); |
f7e531cf IR |
8958 | } |
8959 | ||
8960 | if (!slp_node) | |
8961 | { | |
9771b263 | 8962 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
8963 | if (!masked) |
8964 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
8965 | vec_oprnds2.quick_push (vec_then_clause); |
8966 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
8967 | } |
8968 | ||
9dc3f7de | 8969 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 8970 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 8971 | { |
9771b263 DN |
8972 | vec_then_clause = vec_oprnds2[i]; |
8973 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 8974 | |
a414c77f IE |
8975 | if (masked) |
8976 | vec_compare = vec_cond_lhs; | |
8977 | else | |
8978 | { | |
8979 | vec_cond_rhs = vec_oprnds1[i]; | |
01216d27 JJ |
8980 | if (bitop1 == NOP_EXPR) |
8981 | vec_compare = build2 (cond_code, vec_cmp_type, | |
8982 | vec_cond_lhs, vec_cond_rhs); | |
8983 | else | |
8984 | { | |
8985 | new_temp = make_ssa_name (vec_cmp_type); | |
8986 | if (bitop1 == BIT_NOT_EXPR) | |
8987 | new_stmt = gimple_build_assign (new_temp, bitop1, | |
8988 | vec_cond_rhs); | |
8989 | else | |
8990 | new_stmt | |
8991 | = gimple_build_assign (new_temp, bitop1, vec_cond_lhs, | |
8992 | vec_cond_rhs); | |
8993 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8994 | if (bitop2 == NOP_EXPR) | |
8995 | vec_compare = new_temp; | |
8996 | else if (bitop2 == BIT_NOT_EXPR) | |
8997 | { | |
8998 | /* Instead of doing ~x ? y : z do x ? z : y. */ | |
8999 | vec_compare = new_temp; | |
9000 | std::swap (vec_then_clause, vec_else_clause); | |
9001 | } | |
9002 | else | |
9003 | { | |
9004 | vec_compare = make_ssa_name (vec_cmp_type); | |
9005 | new_stmt | |
9006 | = gimple_build_assign (vec_compare, bitop2, | |
9007 | vec_cond_lhs, new_temp); | |
9008 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9009 | } | |
9010 | } | |
a414c77f | 9011 | } |
bb6c2b68 RS |
9012 | if (reduction_type == EXTRACT_LAST_REDUCTION) |
9013 | { | |
9014 | if (!is_gimple_val (vec_compare)) | |
9015 | { | |
9016 | tree vec_compare_name = make_ssa_name (vec_cmp_type); | |
9017 | new_stmt = gimple_build_assign (vec_compare_name, | |
9018 | vec_compare); | |
9019 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9020 | vec_compare = vec_compare_name; | |
9021 | } | |
9022 | gcc_assert (reduc_index == 2); | |
9023 | new_stmt = gimple_build_call_internal | |
9024 | (IFN_FOLD_EXTRACT_LAST, 3, else_clause, vec_compare, | |
9025 | vec_then_clause); | |
9026 | gimple_call_set_lhs (new_stmt, scalar_dest); | |
9027 | SSA_NAME_DEF_STMT (scalar_dest) = new_stmt; | |
9028 | if (stmt == gsi_stmt (*gsi)) | |
9029 | vect_finish_replace_stmt (stmt, new_stmt); | |
9030 | else | |
9031 | { | |
9032 | /* In this case we're moving the definition to later in the | |
9033 | block. That doesn't matter because the only uses of the | |
9034 | lhs are in phi statements. */ | |
9035 | gimple_stmt_iterator old_gsi = gsi_for_stmt (stmt); | |
9036 | gsi_remove (&old_gsi, true); | |
9037 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9038 | } | |
9039 | } | |
9040 | else | |
9041 | { | |
9042 | new_temp = make_ssa_name (vec_dest); | |
9043 | new_stmt = gimple_build_assign (new_temp, VEC_COND_EXPR, | |
9044 | vec_compare, vec_then_clause, | |
9045 | vec_else_clause); | |
9046 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9047 | } | |
f7e531cf | 9048 | if (slp_node) |
9771b263 | 9049 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
9050 | } |
9051 | ||
9052 | if (slp_node) | |
9053 | continue; | |
9054 | ||
9055 | if (j == 0) | |
9056 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
9057 | else | |
9058 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
9059 | ||
9060 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 9061 | } |
b8698a0f | 9062 | |
9771b263 DN |
9063 | vec_oprnds0.release (); |
9064 | vec_oprnds1.release (); | |
9065 | vec_oprnds2.release (); | |
9066 | vec_oprnds3.release (); | |
f7e531cf | 9067 | |
ebfd146a IR |
9068 | return true; |
9069 | } | |
9070 | ||
42fd8198 IE |
9071 | /* vectorizable_comparison. |
9072 | ||
9073 | Check if STMT is comparison expression that can be vectorized. | |
9074 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
9075 | comparison, put it in VEC_STMT, and insert it at GSI. | |
9076 | ||
9077 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
9078 | ||
fce57248 | 9079 | static bool |
42fd8198 IE |
9080 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, |
9081 | gimple **vec_stmt, tree reduc_def, | |
68435eb2 | 9082 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
42fd8198 IE |
9083 | { |
9084 | tree lhs, rhs1, rhs2; | |
9085 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9086 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
9087 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
9088 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
9089 | tree new_temp; | |
9090 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
9091 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 9092 | int ndts = 2; |
928686b1 | 9093 | poly_uint64 nunits; |
42fd8198 | 9094 | int ncopies; |
49e76ff1 | 9095 | enum tree_code code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
42fd8198 IE |
9096 | stmt_vec_info prev_stmt_info = NULL; |
9097 | int i, j; | |
9098 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
9099 | vec<tree> vec_oprnds0 = vNULL; | |
9100 | vec<tree> vec_oprnds1 = vNULL; | |
42fd8198 IE |
9101 | tree mask_type; |
9102 | tree mask; | |
9103 | ||
c245362b IE |
9104 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
9105 | return false; | |
9106 | ||
30480bcd | 9107 | if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)) |
42fd8198 IE |
9108 | return false; |
9109 | ||
9110 | mask_type = vectype; | |
9111 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
9112 | ||
fce57248 | 9113 | if (slp_node) |
42fd8198 IE |
9114 | ncopies = 1; |
9115 | else | |
e8f142e2 | 9116 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
42fd8198 IE |
9117 | |
9118 | gcc_assert (ncopies >= 1); | |
42fd8198 IE |
9119 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
9120 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
9121 | && reduc_def)) | |
9122 | return false; | |
9123 | ||
9124 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
9125 | { | |
9126 | if (dump_enabled_p ()) | |
9127 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
9128 | "value used after loop.\n"); | |
9129 | return false; | |
9130 | } | |
9131 | ||
9132 | if (!is_gimple_assign (stmt)) | |
9133 | return false; | |
9134 | ||
9135 | code = gimple_assign_rhs_code (stmt); | |
9136 | ||
9137 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
9138 | return false; | |
9139 | ||
9140 | rhs1 = gimple_assign_rhs1 (stmt); | |
9141 | rhs2 = gimple_assign_rhs2 (stmt); | |
9142 | ||
894dd753 | 9143 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &dts[0], &vectype1)) |
42fd8198 IE |
9144 | return false; |
9145 | ||
894dd753 | 9146 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &dts[1], &vectype2)) |
42fd8198 IE |
9147 | return false; |
9148 | ||
9149 | if (vectype1 && vectype2 | |
928686b1 RS |
9150 | && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1), |
9151 | TYPE_VECTOR_SUBPARTS (vectype2))) | |
42fd8198 IE |
9152 | return false; |
9153 | ||
9154 | vectype = vectype1 ? vectype1 : vectype2; | |
9155 | ||
9156 | /* Invariant comparison. */ | |
9157 | if (!vectype) | |
9158 | { | |
69a9a66f | 9159 | vectype = get_vectype_for_scalar_type (TREE_TYPE (rhs1)); |
928686b1 | 9160 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (vectype), nunits)) |
42fd8198 IE |
9161 | return false; |
9162 | } | |
928686b1 | 9163 | else if (maybe_ne (nunits, TYPE_VECTOR_SUBPARTS (vectype))) |
42fd8198 IE |
9164 | return false; |
9165 | ||
49e76ff1 IE |
9166 | /* Can't compare mask and non-mask types. */ |
9167 | if (vectype1 && vectype2 | |
9168 | && (VECTOR_BOOLEAN_TYPE_P (vectype1) ^ VECTOR_BOOLEAN_TYPE_P (vectype2))) | |
9169 | return false; | |
9170 | ||
9171 | /* Boolean values may have another representation in vectors | |
9172 | and therefore we prefer bit operations over comparison for | |
9173 | them (which also works for scalar masks). We store opcodes | |
9174 | to use in bitop1 and bitop2. Statement is vectorized as | |
9175 | BITOP2 (rhs1 BITOP1 rhs2) or | |
9176 | rhs1 BITOP2 (BITOP1 rhs2) | |
9177 | depending on bitop1 and bitop2 arity. */ | |
9178 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) | |
9179 | { | |
9180 | if (code == GT_EXPR) | |
9181 | { | |
9182 | bitop1 = BIT_NOT_EXPR; | |
9183 | bitop2 = BIT_AND_EXPR; | |
9184 | } | |
9185 | else if (code == GE_EXPR) | |
9186 | { | |
9187 | bitop1 = BIT_NOT_EXPR; | |
9188 | bitop2 = BIT_IOR_EXPR; | |
9189 | } | |
9190 | else if (code == LT_EXPR) | |
9191 | { | |
9192 | bitop1 = BIT_NOT_EXPR; | |
9193 | bitop2 = BIT_AND_EXPR; | |
9194 | std::swap (rhs1, rhs2); | |
264d951a | 9195 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
9196 | } |
9197 | else if (code == LE_EXPR) | |
9198 | { | |
9199 | bitop1 = BIT_NOT_EXPR; | |
9200 | bitop2 = BIT_IOR_EXPR; | |
9201 | std::swap (rhs1, rhs2); | |
264d951a | 9202 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
9203 | } |
9204 | else | |
9205 | { | |
9206 | bitop1 = BIT_XOR_EXPR; | |
9207 | if (code == EQ_EXPR) | |
9208 | bitop2 = BIT_NOT_EXPR; | |
9209 | } | |
9210 | } | |
9211 | ||
42fd8198 IE |
9212 | if (!vec_stmt) |
9213 | { | |
49e76ff1 | 9214 | if (bitop1 == NOP_EXPR) |
68435eb2 RB |
9215 | { |
9216 | if (!expand_vec_cmp_expr_p (vectype, mask_type, code)) | |
9217 | return false; | |
9218 | } | |
49e76ff1 IE |
9219 | else |
9220 | { | |
9221 | machine_mode mode = TYPE_MODE (vectype); | |
9222 | optab optab; | |
9223 | ||
9224 | optab = optab_for_tree_code (bitop1, vectype, optab_default); | |
9225 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
9226 | return false; | |
9227 | ||
9228 | if (bitop2 != NOP_EXPR) | |
9229 | { | |
9230 | optab = optab_for_tree_code (bitop2, vectype, optab_default); | |
9231 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
9232 | return false; | |
9233 | } | |
49e76ff1 | 9234 | } |
68435eb2 RB |
9235 | |
9236 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
9237 | vect_model_simple_cost (stmt_info, ncopies * (1 + (bitop2 != NOP_EXPR)), | |
9238 | dts, ndts, slp_node, cost_vec); | |
9239 | return true; | |
42fd8198 IE |
9240 | } |
9241 | ||
9242 | /* Transform. */ | |
9243 | if (!slp_node) | |
9244 | { | |
9245 | vec_oprnds0.create (1); | |
9246 | vec_oprnds1.create (1); | |
9247 | } | |
9248 | ||
9249 | /* Handle def. */ | |
9250 | lhs = gimple_assign_lhs (stmt); | |
9251 | mask = vect_create_destination_var (lhs, mask_type); | |
9252 | ||
9253 | /* Handle cmp expr. */ | |
9254 | for (j = 0; j < ncopies; j++) | |
9255 | { | |
9256 | gassign *new_stmt = NULL; | |
9257 | if (j == 0) | |
9258 | { | |
9259 | if (slp_node) | |
9260 | { | |
9261 | auto_vec<tree, 2> ops; | |
9262 | auto_vec<vec<tree>, 2> vec_defs; | |
9263 | ||
9264 | ops.safe_push (rhs1); | |
9265 | ops.safe_push (rhs2); | |
306b0c92 | 9266 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
42fd8198 IE |
9267 | vec_oprnds1 = vec_defs.pop (); |
9268 | vec_oprnds0 = vec_defs.pop (); | |
9269 | } | |
9270 | else | |
9271 | { | |
e4af0bc4 IE |
9272 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, vectype); |
9273 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, vectype); | |
42fd8198 IE |
9274 | } |
9275 | } | |
9276 | else | |
9277 | { | |
9278 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
9279 | vec_oprnds0.pop ()); | |
9280 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
9281 | vec_oprnds1.pop ()); | |
9282 | } | |
9283 | ||
9284 | if (!slp_node) | |
9285 | { | |
9286 | vec_oprnds0.quick_push (vec_rhs1); | |
9287 | vec_oprnds1.quick_push (vec_rhs2); | |
9288 | } | |
9289 | ||
9290 | /* Arguments are ready. Create the new vector stmt. */ | |
9291 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
9292 | { | |
9293 | vec_rhs2 = vec_oprnds1[i]; | |
9294 | ||
9295 | new_temp = make_ssa_name (mask); | |
49e76ff1 IE |
9296 | if (bitop1 == NOP_EXPR) |
9297 | { | |
9298 | new_stmt = gimple_build_assign (new_temp, code, | |
9299 | vec_rhs1, vec_rhs2); | |
9300 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9301 | } | |
9302 | else | |
9303 | { | |
9304 | if (bitop1 == BIT_NOT_EXPR) | |
9305 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs2); | |
9306 | else | |
9307 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs1, | |
9308 | vec_rhs2); | |
9309 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9310 | if (bitop2 != NOP_EXPR) | |
9311 | { | |
9312 | tree res = make_ssa_name (mask); | |
9313 | if (bitop2 == BIT_NOT_EXPR) | |
9314 | new_stmt = gimple_build_assign (res, bitop2, new_temp); | |
9315 | else | |
9316 | new_stmt = gimple_build_assign (res, bitop2, vec_rhs1, | |
9317 | new_temp); | |
9318 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9319 | } | |
9320 | } | |
42fd8198 IE |
9321 | if (slp_node) |
9322 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
9323 | } | |
9324 | ||
9325 | if (slp_node) | |
9326 | continue; | |
9327 | ||
9328 | if (j == 0) | |
9329 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
9330 | else | |
9331 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
9332 | ||
9333 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
9334 | } | |
9335 | ||
9336 | vec_oprnds0.release (); | |
9337 | vec_oprnds1.release (); | |
9338 | ||
9339 | return true; | |
9340 | } | |
ebfd146a | 9341 | |
68a0f2ff RS |
9342 | /* If SLP_NODE is nonnull, return true if vectorizable_live_operation |
9343 | can handle all live statements in the node. Otherwise return true | |
9344 | if STMT is not live or if vectorizable_live_operation can handle it. | |
9345 | GSI and VEC_STMT are as for vectorizable_live_operation. */ | |
9346 | ||
9347 | static bool | |
9348 | can_vectorize_live_stmts (gimple *stmt, gimple_stmt_iterator *gsi, | |
68435eb2 RB |
9349 | slp_tree slp_node, gimple **vec_stmt, |
9350 | stmt_vector_for_cost *cost_vec) | |
68a0f2ff RS |
9351 | { |
9352 | if (slp_node) | |
9353 | { | |
9354 | gimple *slp_stmt; | |
9355 | unsigned int i; | |
9356 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, slp_stmt) | |
9357 | { | |
9358 | stmt_vec_info slp_stmt_info = vinfo_for_stmt (slp_stmt); | |
9359 | if (STMT_VINFO_LIVE_P (slp_stmt_info) | |
9360 | && !vectorizable_live_operation (slp_stmt, gsi, slp_node, i, | |
68435eb2 | 9361 | vec_stmt, cost_vec)) |
68a0f2ff RS |
9362 | return false; |
9363 | } | |
9364 | } | |
9365 | else if (STMT_VINFO_LIVE_P (vinfo_for_stmt (stmt)) | |
68435eb2 RB |
9366 | && !vectorizable_live_operation (stmt, gsi, slp_node, -1, vec_stmt, |
9367 | cost_vec)) | |
68a0f2ff RS |
9368 | return false; |
9369 | ||
9370 | return true; | |
9371 | } | |
9372 | ||
8644a673 | 9373 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
9374 | |
9375 | bool | |
891ad31c | 9376 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node, |
68435eb2 | 9377 | slp_instance node_instance, stmt_vector_for_cost *cost_vec) |
ebfd146a | 9378 | { |
8644a673 | 9379 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
6585ff8f | 9380 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 9381 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 9382 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 9383 | bool ok; |
355fe088 | 9384 | gimple *pattern_stmt; |
363477c0 | 9385 | gimple_seq pattern_def_seq; |
ebfd146a | 9386 | |
73fbfcad | 9387 | if (dump_enabled_p ()) |
ebfd146a | 9388 | { |
78c60e3d SS |
9389 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
9390 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 9391 | } |
ebfd146a | 9392 | |
1825a1f3 | 9393 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 9394 | { |
73fbfcad | 9395 | if (dump_enabled_p ()) |
78c60e3d | 9396 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9397 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
9398 | |
9399 | return false; | |
9400 | } | |
b8698a0f | 9401 | |
d54a098e RS |
9402 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
9403 | && node == NULL | |
9404 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) | |
9405 | { | |
9406 | gimple_stmt_iterator si; | |
9407 | ||
9408 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) | |
9409 | { | |
9410 | gimple *pattern_def_stmt = gsi_stmt (si); | |
6585ff8f RS |
9411 | stmt_vec_info pattern_def_stmt_info |
9412 | = vinfo->lookup_stmt (gsi_stmt (si)); | |
9413 | if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info) | |
9414 | || STMT_VINFO_LIVE_P (pattern_def_stmt_info)) | |
d54a098e RS |
9415 | { |
9416 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
9417 | if (dump_enabled_p ()) | |
9418 | { | |
9419 | dump_printf_loc (MSG_NOTE, vect_location, | |
9420 | "==> examining pattern def statement: "); | |
9421 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
9422 | } | |
9423 | ||
9424 | if (!vect_analyze_stmt (pattern_def_stmt, | |
9425 | need_to_vectorize, node, node_instance, | |
9426 | cost_vec)) | |
9427 | return false; | |
9428 | } | |
9429 | } | |
9430 | } | |
9431 | ||
b8698a0f | 9432 | /* Skip stmts that do not need to be vectorized. In loops this is expected |
8644a673 IR |
9433 | to include: |
9434 | - the COND_EXPR which is the loop exit condition | |
9435 | - any LABEL_EXPRs in the loop | |
b8698a0f | 9436 | - computations that are used only for array indexing or loop control. |
8644a673 | 9437 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 9438 | instance, therefore, all the statements are relevant. |
ebfd146a | 9439 | |
d092494c | 9440 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 9441 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
9442 | statements. In basic blocks we are called from some SLP instance |
9443 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
9444 | already will be part of SLP instance. */ | |
83197f37 IR |
9445 | |
9446 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 9447 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 9448 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 9449 | { |
9d5e7640 | 9450 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 9451 | && pattern_stmt |
9d5e7640 IR |
9452 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
9453 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
9454 | { | |
83197f37 | 9455 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
9456 | stmt = pattern_stmt; |
9457 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 9458 | if (dump_enabled_p ()) |
9d5e7640 | 9459 | { |
78c60e3d SS |
9460 | dump_printf_loc (MSG_NOTE, vect_location, |
9461 | "==> examining pattern statement: "); | |
9462 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
9463 | } |
9464 | } | |
9465 | else | |
9466 | { | |
73fbfcad | 9467 | if (dump_enabled_p ()) |
e645e942 | 9468 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 9469 | |
9d5e7640 IR |
9470 | return true; |
9471 | } | |
8644a673 | 9472 | } |
83197f37 | 9473 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 9474 | && node == NULL |
83197f37 IR |
9475 | && pattern_stmt |
9476 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
9477 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
9478 | { | |
9479 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 9480 | if (dump_enabled_p ()) |
83197f37 | 9481 | { |
78c60e3d SS |
9482 | dump_printf_loc (MSG_NOTE, vect_location, |
9483 | "==> examining pattern statement: "); | |
9484 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
9485 | } |
9486 | ||
891ad31c | 9487 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node, |
68435eb2 | 9488 | node_instance, cost_vec)) |
83197f37 IR |
9489 | return false; |
9490 | } | |
ebfd146a | 9491 | |
8644a673 IR |
9492 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
9493 | { | |
9494 | case vect_internal_def: | |
9495 | break; | |
ebfd146a | 9496 | |
8644a673 | 9497 | case vect_reduction_def: |
7c5222ff | 9498 | case vect_nested_cycle: |
14a61437 RB |
9499 | gcc_assert (!bb_vinfo |
9500 | && (relevance == vect_used_in_outer | |
9501 | || relevance == vect_used_in_outer_by_reduction | |
9502 | || relevance == vect_used_by_reduction | |
b28ead45 AH |
9503 | || relevance == vect_unused_in_scope |
9504 | || relevance == vect_used_only_live)); | |
8644a673 IR |
9505 | break; |
9506 | ||
9507 | case vect_induction_def: | |
e7baeb39 RB |
9508 | gcc_assert (!bb_vinfo); |
9509 | break; | |
9510 | ||
8644a673 IR |
9511 | case vect_constant_def: |
9512 | case vect_external_def: | |
9513 | case vect_unknown_def_type: | |
9514 | default: | |
9515 | gcc_unreachable (); | |
9516 | } | |
ebfd146a | 9517 | |
8644a673 | 9518 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 9519 | { |
8644a673 | 9520 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
9521 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
9522 | || (is_gimple_call (stmt) | |
9523 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 9524 | *need_to_vectorize = true; |
ebfd146a IR |
9525 | } |
9526 | ||
b1af7da6 RB |
9527 | if (PURE_SLP_STMT (stmt_info) && !node) |
9528 | { | |
9529 | dump_printf_loc (MSG_NOTE, vect_location, | |
9530 | "handled only by SLP analysis\n"); | |
9531 | return true; | |
9532 | } | |
9533 | ||
9534 | ok = true; | |
9535 | if (!bb_vinfo | |
9536 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
9537 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
68435eb2 RB |
9538 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node, cost_vec) |
9539 | || vectorizable_conversion (stmt, NULL, NULL, node, cost_vec) | |
9540 | || vectorizable_shift (stmt, NULL, NULL, node, cost_vec) | |
9541 | || vectorizable_operation (stmt, NULL, NULL, node, cost_vec) | |
9542 | || vectorizable_assignment (stmt, NULL, NULL, node, cost_vec) | |
9543 | || vectorizable_load (stmt, NULL, NULL, node, node_instance, cost_vec) | |
9544 | || vectorizable_call (stmt, NULL, NULL, node, cost_vec) | |
9545 | || vectorizable_store (stmt, NULL, NULL, node, cost_vec) | |
9546 | || vectorizable_reduction (stmt, NULL, NULL, node, node_instance, | |
9547 | cost_vec) | |
9548 | || vectorizable_induction (stmt, NULL, NULL, node, cost_vec) | |
9549 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node, cost_vec) | |
9550 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node, cost_vec)); | |
b1af7da6 RB |
9551 | else |
9552 | { | |
9553 | if (bb_vinfo) | |
68435eb2 RB |
9554 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node, cost_vec) |
9555 | || vectorizable_conversion (stmt, NULL, NULL, node, cost_vec) | |
9556 | || vectorizable_shift (stmt, NULL, NULL, node, cost_vec) | |
9557 | || vectorizable_operation (stmt, NULL, NULL, node, cost_vec) | |
9558 | || vectorizable_assignment (stmt, NULL, NULL, node, cost_vec) | |
9559 | || vectorizable_load (stmt, NULL, NULL, node, node_instance, | |
9560 | cost_vec) | |
9561 | || vectorizable_call (stmt, NULL, NULL, node, cost_vec) | |
9562 | || vectorizable_store (stmt, NULL, NULL, node, cost_vec) | |
9563 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node, | |
9564 | cost_vec) | |
9565 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node, | |
9566 | cost_vec)); | |
b1af7da6 | 9567 | } |
8644a673 IR |
9568 | |
9569 | if (!ok) | |
ebfd146a | 9570 | { |
73fbfcad | 9571 | if (dump_enabled_p ()) |
8644a673 | 9572 | { |
78c60e3d SS |
9573 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9574 | "not vectorized: relevant stmt not "); | |
9575 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
9576 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 9577 | } |
b8698a0f | 9578 | |
ebfd146a IR |
9579 | return false; |
9580 | } | |
9581 | ||
8644a673 IR |
9582 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
9583 | need extra handling, except for vectorizable reductions. */ | |
68435eb2 RB |
9584 | if (!bb_vinfo |
9585 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type | |
9586 | && !can_vectorize_live_stmts (stmt, NULL, node, NULL, cost_vec)) | |
ebfd146a | 9587 | { |
73fbfcad | 9588 | if (dump_enabled_p ()) |
8644a673 | 9589 | { |
78c60e3d | 9590 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
68a0f2ff | 9591 | "not vectorized: live stmt not supported: "); |
78c60e3d | 9592 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); |
8644a673 | 9593 | } |
b8698a0f | 9594 | |
8644a673 | 9595 | return false; |
ebfd146a IR |
9596 | } |
9597 | ||
ebfd146a IR |
9598 | return true; |
9599 | } | |
9600 | ||
9601 | ||
9602 | /* Function vect_transform_stmt. | |
9603 | ||
9604 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
9605 | ||
9606 | bool | |
355fe088 | 9607 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 9608 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
9609 | slp_instance slp_node_instance) |
9610 | { | |
6585ff8f RS |
9611 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
9612 | vec_info *vinfo = stmt_info->vinfo; | |
ebfd146a | 9613 | bool is_store = false; |
355fe088 | 9614 | gimple *vec_stmt = NULL; |
ebfd146a | 9615 | bool done; |
ebfd146a | 9616 | |
fce57248 | 9617 | gcc_assert (slp_node || !PURE_SLP_STMT (stmt_info)); |
355fe088 | 9618 | gimple *old_vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 9619 | |
e57d9a82 RB |
9620 | bool nested_p = (STMT_VINFO_LOOP_VINFO (stmt_info) |
9621 | && nested_in_vect_loop_p | |
9622 | (LOOP_VINFO_LOOP (STMT_VINFO_LOOP_VINFO (stmt_info)), | |
9623 | stmt)); | |
9624 | ||
ebfd146a IR |
9625 | switch (STMT_VINFO_TYPE (stmt_info)) |
9626 | { | |
9627 | case type_demotion_vec_info_type: | |
ebfd146a | 9628 | case type_promotion_vec_info_type: |
ebfd146a | 9629 | case type_conversion_vec_info_type: |
68435eb2 | 9630 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a IR |
9631 | gcc_assert (done); |
9632 | break; | |
9633 | ||
9634 | case induc_vec_info_type: | |
68435eb2 | 9635 | done = vectorizable_induction (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a IR |
9636 | gcc_assert (done); |
9637 | break; | |
9638 | ||
9dc3f7de | 9639 | case shift_vec_info_type: |
68435eb2 | 9640 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node, NULL); |
9dc3f7de IR |
9641 | gcc_assert (done); |
9642 | break; | |
9643 | ||
ebfd146a | 9644 | case op_vec_info_type: |
68435eb2 | 9645 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a IR |
9646 | gcc_assert (done); |
9647 | break; | |
9648 | ||
9649 | case assignment_vec_info_type: | |
68435eb2 | 9650 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a IR |
9651 | gcc_assert (done); |
9652 | break; | |
9653 | ||
9654 | case load_vec_info_type: | |
b8698a0f | 9655 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
68435eb2 | 9656 | slp_node_instance, NULL); |
ebfd146a IR |
9657 | gcc_assert (done); |
9658 | break; | |
9659 | ||
9660 | case store_vec_info_type: | |
68435eb2 | 9661 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a | 9662 | gcc_assert (done); |
0d0293ac | 9663 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
9664 | { |
9665 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 9666 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
9667 | one are skipped, and there vec_stmt_info shouldn't be freed |
9668 | meanwhile. */ | |
0d0293ac | 9669 | *grouped_store = true; |
f307441a | 9670 | stmt_vec_info group_info |
2c53b149 RB |
9671 | = vinfo_for_stmt (DR_GROUP_FIRST_ELEMENT (stmt_info)); |
9672 | if (DR_GROUP_STORE_COUNT (group_info) == DR_GROUP_SIZE (group_info)) | |
ebfd146a | 9673 | is_store = true; |
f307441a | 9674 | } |
ebfd146a IR |
9675 | else |
9676 | is_store = true; | |
9677 | break; | |
9678 | ||
9679 | case condition_vec_info_type: | |
68435eb2 | 9680 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node, NULL); |
ebfd146a IR |
9681 | gcc_assert (done); |
9682 | break; | |
9683 | ||
42fd8198 | 9684 | case comparison_vec_info_type: |
68435eb2 | 9685 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node, NULL); |
42fd8198 IE |
9686 | gcc_assert (done); |
9687 | break; | |
9688 | ||
ebfd146a | 9689 | case call_vec_info_type: |
68435eb2 | 9690 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node, NULL); |
039d9ea1 | 9691 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
9692 | break; |
9693 | ||
0136f8f0 | 9694 | case call_simd_clone_vec_info_type: |
68435eb2 | 9695 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node, NULL); |
0136f8f0 AH |
9696 | stmt = gsi_stmt (*gsi); |
9697 | break; | |
9698 | ||
ebfd146a | 9699 | case reduc_vec_info_type: |
891ad31c | 9700 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node, |
68435eb2 | 9701 | slp_node_instance, NULL); |
ebfd146a IR |
9702 | gcc_assert (done); |
9703 | break; | |
9704 | ||
9705 | default: | |
9706 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
9707 | { | |
73fbfcad | 9708 | if (dump_enabled_p ()) |
78c60e3d | 9709 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9710 | "stmt not supported.\n"); |
ebfd146a IR |
9711 | gcc_unreachable (); |
9712 | } | |
9713 | } | |
9714 | ||
225ce44b RB |
9715 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
9716 | This would break hybrid SLP vectorization. */ | |
9717 | if (slp_node) | |
d90f8440 RB |
9718 | gcc_assert (!vec_stmt |
9719 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt); | |
225ce44b | 9720 | |
ebfd146a IR |
9721 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
9722 | is being vectorized, but outside the immediately enclosing loop. */ | |
9723 | if (vec_stmt | |
e57d9a82 | 9724 | && nested_p |
ebfd146a IR |
9725 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
9726 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 9727 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 9728 | vect_used_in_outer_by_reduction)) |
ebfd146a | 9729 | { |
a70d6342 IR |
9730 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
9731 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
9732 | imm_use_iterator imm_iter; |
9733 | use_operand_p use_p; | |
9734 | tree scalar_dest; | |
ebfd146a | 9735 | |
73fbfcad | 9736 | if (dump_enabled_p ()) |
78c60e3d | 9737 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 9738 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
9739 | |
9740 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
9741 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
9742 | STMT). */ | |
9743 | if (gimple_code (stmt) == GIMPLE_PHI) | |
9744 | scalar_dest = PHI_RESULT (stmt); | |
9745 | else | |
9746 | scalar_dest = gimple_assign_lhs (stmt); | |
9747 | ||
9748 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
6585ff8f RS |
9749 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) |
9750 | { | |
9751 | stmt_vec_info exit_phi_info | |
9752 | = vinfo->lookup_stmt (USE_STMT (use_p)); | |
9753 | STMT_VINFO_VEC_STMT (exit_phi_info) = vec_stmt; | |
9754 | } | |
ebfd146a IR |
9755 | } |
9756 | ||
9757 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
9758 | being vectorized. */ | |
68a0f2ff | 9759 | if (STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) |
ebfd146a | 9760 | { |
68435eb2 | 9761 | done = can_vectorize_live_stmts (stmt, gsi, slp_node, &vec_stmt, NULL); |
ebfd146a IR |
9762 | gcc_assert (done); |
9763 | } | |
9764 | ||
9765 | if (vec_stmt) | |
83197f37 | 9766 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 9767 | |
b8698a0f | 9768 | return is_store; |
ebfd146a IR |
9769 | } |
9770 | ||
9771 | ||
b8698a0f | 9772 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
9773 | stmt_vec_info. */ |
9774 | ||
9775 | void | |
355fe088 | 9776 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 9777 | { |
355fe088 TS |
9778 | gimple *next = first_stmt; |
9779 | gimple *tmp; | |
ebfd146a IR |
9780 | gimple_stmt_iterator next_si; |
9781 | ||
9782 | while (next) | |
9783 | { | |
78048b1c JJ |
9784 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
9785 | ||
2c53b149 | 9786 | tmp = DR_GROUP_NEXT_ELEMENT (stmt_info); |
78048b1c JJ |
9787 | if (is_pattern_stmt_p (stmt_info)) |
9788 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
9789 | /* Free the attached stmt_vec_info and remove the stmt. */ |
9790 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 9791 | unlink_stmt_vdef (next); |
ebfd146a | 9792 | gsi_remove (&next_si, true); |
3d3f2249 | 9793 | release_defs (next); |
ebfd146a IR |
9794 | free_stmt_vec_info (next); |
9795 | next = tmp; | |
9796 | } | |
9797 | } | |
9798 | ||
9799 | ||
9800 | /* Function new_stmt_vec_info. | |
9801 | ||
9802 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
9803 | ||
9804 | stmt_vec_info | |
310213d4 | 9805 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
9806 | { |
9807 | stmt_vec_info res; | |
9808 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
9809 | ||
9810 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
9811 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 9812 | res->vinfo = vinfo; |
8644a673 | 9813 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
9814 | STMT_VINFO_LIVE_P (res) = false; |
9815 | STMT_VINFO_VECTYPE (res) = NULL; | |
9816 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 9817 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
9818 | STMT_VINFO_IN_PATTERN_P (res) = false; |
9819 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 9820 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 9821 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 9822 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
7e16ce79 | 9823 | STMT_VINFO_VEC_CONST_COND_REDUC_CODE (res) = ERROR_MARK; |
ebfd146a | 9824 | |
ebfd146a IR |
9825 | if (gimple_code (stmt) == GIMPLE_PHI |
9826 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
9827 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
9828 | else | |
8644a673 IR |
9829 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
9830 | ||
9771b263 | 9831 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 9832 | STMT_SLP_TYPE (res) = loop_vect; |
78810bd3 RB |
9833 | STMT_VINFO_NUM_SLP_USES (res) = 0; |
9834 | ||
2c53b149 RB |
9835 | res->first_element = NULL; /* GROUP_FIRST_ELEMENT */ |
9836 | res->next_element = NULL; /* GROUP_NEXT_ELEMENT */ | |
9837 | res->size = 0; /* GROUP_SIZE */ | |
9838 | res->store_count = 0; /* GROUP_STORE_COUNT */ | |
9839 | res->gap = 0; /* GROUP_GAP */ | |
9840 | res->same_dr_stmt = NULL; /* GROUP_SAME_DR_STMT */ | |
ebfd146a | 9841 | |
ca823c85 RB |
9842 | /* This is really "uninitialized" until vect_compute_data_ref_alignment. */ |
9843 | res->dr_aux.misalignment = DR_MISALIGNMENT_UNINITIALIZED; | |
9844 | ||
ebfd146a IR |
9845 | return res; |
9846 | } | |
9847 | ||
9848 | ||
f8c0baaf | 9849 | /* Set the current stmt_vec_info vector to V. */ |
ebfd146a IR |
9850 | |
9851 | void | |
f8c0baaf | 9852 | set_stmt_vec_info_vec (vec<stmt_vec_info> *v) |
ebfd146a | 9853 | { |
f8c0baaf | 9854 | stmt_vec_info_vec = v; |
ebfd146a IR |
9855 | } |
9856 | ||
f8c0baaf | 9857 | /* Free the stmt_vec_info entries in V and release V. */ |
ebfd146a IR |
9858 | |
9859 | void | |
f8c0baaf | 9860 | free_stmt_vec_infos (vec<stmt_vec_info> *v) |
ebfd146a | 9861 | { |
93675444 | 9862 | unsigned int i; |
3161455c | 9863 | stmt_vec_info info; |
f8c0baaf | 9864 | FOR_EACH_VEC_ELT (*v, i, info) |
93675444 | 9865 | if (info != NULL) |
3161455c | 9866 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
f8c0baaf RB |
9867 | if (v == stmt_vec_info_vec) |
9868 | stmt_vec_info_vec = NULL; | |
9869 | v->release (); | |
ebfd146a IR |
9870 | } |
9871 | ||
9872 | ||
9873 | /* Free stmt vectorization related info. */ | |
9874 | ||
9875 | void | |
355fe088 | 9876 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
9877 | { |
9878 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9879 | ||
9880 | if (!stmt_info) | |
9881 | return; | |
9882 | ||
78048b1c JJ |
9883 | /* Check if this statement has a related "pattern stmt" |
9884 | (introduced by the vectorizer during the pattern recognition | |
9885 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
9886 | too. */ | |
9887 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
9888 | { | |
e3947d80 RS |
9889 | if (gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info)) |
9890 | for (gimple_stmt_iterator si = gsi_start (seq); | |
9891 | !gsi_end_p (si); gsi_next (&si)) | |
9892 | { | |
9893 | gimple *seq_stmt = gsi_stmt (si); | |
9894 | gimple_set_bb (seq_stmt, NULL); | |
9895 | tree lhs = gimple_get_lhs (seq_stmt); | |
9896 | if (lhs && TREE_CODE (lhs) == SSA_NAME) | |
9897 | release_ssa_name (lhs); | |
9898 | free_stmt_vec_info (seq_stmt); | |
9899 | } | |
78048b1c JJ |
9900 | stmt_vec_info patt_info |
9901 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
9902 | if (patt_info) | |
9903 | { | |
355fe088 | 9904 | gimple *patt_stmt = STMT_VINFO_STMT (patt_info); |
f0281fde RB |
9905 | gimple_set_bb (patt_stmt, NULL); |
9906 | tree lhs = gimple_get_lhs (patt_stmt); | |
e6f5c25d | 9907 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde | 9908 | release_ssa_name (lhs); |
f0281fde | 9909 | free_stmt_vec_info (patt_stmt); |
78048b1c JJ |
9910 | } |
9911 | } | |
9912 | ||
9771b263 | 9913 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 9914 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
9915 | set_vinfo_for_stmt (stmt, NULL); |
9916 | free (stmt_info); | |
9917 | } | |
9918 | ||
9919 | ||
bb67d9c7 | 9920 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 9921 | |
bb67d9c7 | 9922 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
9923 | by the target. */ |
9924 | ||
c803b2a9 | 9925 | tree |
86e36728 | 9926 | get_vectype_for_scalar_type_and_size (tree scalar_type, poly_uint64 size) |
ebfd146a | 9927 | { |
c7d97b28 | 9928 | tree orig_scalar_type = scalar_type; |
3bd8f481 | 9929 | scalar_mode inner_mode; |
ef4bddc2 | 9930 | machine_mode simd_mode; |
86e36728 | 9931 | poly_uint64 nunits; |
ebfd146a IR |
9932 | tree vectype; |
9933 | ||
3bd8f481 RS |
9934 | if (!is_int_mode (TYPE_MODE (scalar_type), &inner_mode) |
9935 | && !is_float_mode (TYPE_MODE (scalar_type), &inner_mode)) | |
ebfd146a IR |
9936 | return NULL_TREE; |
9937 | ||
3bd8f481 | 9938 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
48f2e373 | 9939 | |
7b7b1813 RG |
9940 | /* For vector types of elements whose mode precision doesn't |
9941 | match their types precision we use a element type of mode | |
9942 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
9943 | they support the proper result truncation/extension. |
9944 | We also make sure to build vector types with INTEGER_TYPE | |
9945 | component type only. */ | |
6d7971b8 | 9946 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
9947 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
9948 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
9949 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
9950 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 9951 | |
ccbf5bb4 RG |
9952 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
9953 | When the component mode passes the above test simply use a type | |
9954 | corresponding to that mode. The theory is that any use that | |
9955 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 9956 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 9957 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
9958 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
9959 | ||
9960 | /* We can't build a vector type of elements with alignment bigger than | |
9961 | their size. */ | |
dfc2e2ac | 9962 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
9963 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
9964 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 9965 | |
dfc2e2ac RB |
9966 | /* If we felt back to using the mode fail if there was |
9967 | no scalar type for it. */ | |
9968 | if (scalar_type == NULL_TREE) | |
9969 | return NULL_TREE; | |
9970 | ||
bb67d9c7 RG |
9971 | /* If no size was supplied use the mode the target prefers. Otherwise |
9972 | lookup a vector mode of the specified size. */ | |
86e36728 | 9973 | if (known_eq (size, 0U)) |
bb67d9c7 | 9974 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); |
86e36728 RS |
9975 | else if (!multiple_p (size, nbytes, &nunits) |
9976 | || !mode_for_vector (inner_mode, nunits).exists (&simd_mode)) | |
9da15d40 | 9977 | return NULL_TREE; |
4c8fd8ac | 9978 | /* NOTE: nunits == 1 is allowed to support single element vector types. */ |
86e36728 | 9979 | if (!multiple_p (GET_MODE_SIZE (simd_mode), nbytes, &nunits)) |
cc4b5170 | 9980 | return NULL_TREE; |
ebfd146a IR |
9981 | |
9982 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
9983 | |
9984 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
9985 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 9986 | return NULL_TREE; |
ebfd146a | 9987 | |
c7d97b28 RB |
9988 | /* Re-attach the address-space qualifier if we canonicalized the scalar |
9989 | type. */ | |
9990 | if (TYPE_ADDR_SPACE (orig_scalar_type) != TYPE_ADDR_SPACE (vectype)) | |
9991 | return build_qualified_type | |
9992 | (vectype, KEEP_QUAL_ADDR_SPACE (TYPE_QUALS (orig_scalar_type))); | |
9993 | ||
ebfd146a IR |
9994 | return vectype; |
9995 | } | |
9996 | ||
86e36728 | 9997 | poly_uint64 current_vector_size; |
bb67d9c7 RG |
9998 | |
9999 | /* Function get_vectype_for_scalar_type. | |
10000 | ||
10001 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
10002 | by the target. */ | |
10003 | ||
10004 | tree | |
10005 | get_vectype_for_scalar_type (tree scalar_type) | |
10006 | { | |
10007 | tree vectype; | |
10008 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
10009 | current_vector_size); | |
10010 | if (vectype | |
86e36728 | 10011 | && known_eq (current_vector_size, 0U)) |
bb67d9c7 RG |
10012 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); |
10013 | return vectype; | |
10014 | } | |
10015 | ||
42fd8198 IE |
10016 | /* Function get_mask_type_for_scalar_type. |
10017 | ||
10018 | Returns the mask type corresponding to a result of comparison | |
10019 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
10020 | ||
10021 | tree | |
10022 | get_mask_type_for_scalar_type (tree scalar_type) | |
10023 | { | |
10024 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
10025 | ||
10026 | if (!vectype) | |
10027 | return NULL; | |
10028 | ||
10029 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
10030 | current_vector_size); | |
10031 | } | |
10032 | ||
b690cc0f RG |
10033 | /* Function get_same_sized_vectype |
10034 | ||
10035 | Returns a vector type corresponding to SCALAR_TYPE of size | |
10036 | VECTOR_TYPE if supported by the target. */ | |
10037 | ||
10038 | tree | |
bb67d9c7 | 10039 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 10040 | { |
2568d8a1 | 10041 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type)) |
9f47c7e5 IE |
10042 | return build_same_sized_truth_vector_type (vector_type); |
10043 | ||
bb67d9c7 RG |
10044 | return get_vectype_for_scalar_type_and_size |
10045 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
10046 | } |
10047 | ||
ebfd146a IR |
10048 | /* Function vect_is_simple_use. |
10049 | ||
10050 | Input: | |
81c40241 RB |
10051 | VINFO - the vect info of the loop or basic block that is being vectorized. |
10052 | OPERAND - operand in the loop or bb. | |
10053 | Output: | |
894dd753 | 10054 | DEF_STMT_OUT (optional) - the defining stmt in case OPERAND is an SSA_NAME. |
81c40241 | 10055 | DT - the type of definition |
ebfd146a IR |
10056 | |
10057 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 10058 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 10059 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 10060 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
10061 | is the case in reduction/induction computations). |
10062 | For basic blocks, supportable operands are constants and bb invariants. | |
10063 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
10064 | |
10065 | bool | |
894dd753 RS |
10066 | vect_is_simple_use (tree operand, vec_info *vinfo, enum vect_def_type *dt, |
10067 | gimple **def_stmt_out) | |
b8698a0f | 10068 | { |
894dd753 RS |
10069 | if (def_stmt_out) |
10070 | *def_stmt_out = NULL; | |
3fc356dc | 10071 | *dt = vect_unknown_def_type; |
b8698a0f | 10072 | |
73fbfcad | 10073 | if (dump_enabled_p ()) |
ebfd146a | 10074 | { |
78c60e3d SS |
10075 | dump_printf_loc (MSG_NOTE, vect_location, |
10076 | "vect_is_simple_use: operand "); | |
30f502ed RB |
10077 | if (TREE_CODE (operand) == SSA_NAME |
10078 | && !SSA_NAME_IS_DEFAULT_DEF (operand)) | |
10079 | dump_gimple_expr (MSG_NOTE, TDF_SLIM, SSA_NAME_DEF_STMT (operand), 0); | |
10080 | else | |
10081 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
ebfd146a | 10082 | } |
b8698a0f | 10083 | |
b758f602 | 10084 | if (CONSTANT_CLASS_P (operand)) |
30f502ed RB |
10085 | *dt = vect_constant_def; |
10086 | else if (is_gimple_min_invariant (operand)) | |
10087 | *dt = vect_external_def; | |
10088 | else if (TREE_CODE (operand) != SSA_NAME) | |
10089 | *dt = vect_unknown_def_type; | |
10090 | else if (SSA_NAME_IS_DEFAULT_DEF (operand)) | |
8644a673 | 10091 | *dt = vect_external_def; |
ebfd146a IR |
10092 | else |
10093 | { | |
30f502ed | 10094 | gimple *def_stmt = SSA_NAME_DEF_STMT (operand); |
c98d0595 RS |
10095 | stmt_vec_info stmt_vinfo = vinfo->lookup_def (operand); |
10096 | if (!stmt_vinfo) | |
30f502ed RB |
10097 | *dt = vect_external_def; |
10098 | else | |
0f8c840c | 10099 | { |
30f502ed RB |
10100 | if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo)) |
10101 | { | |
10102 | def_stmt = STMT_VINFO_RELATED_STMT (stmt_vinfo); | |
10103 | stmt_vinfo = vinfo_for_stmt (def_stmt); | |
10104 | } | |
10105 | switch (gimple_code (def_stmt)) | |
10106 | { | |
10107 | case GIMPLE_PHI: | |
10108 | case GIMPLE_ASSIGN: | |
10109 | case GIMPLE_CALL: | |
10110 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
10111 | break; | |
10112 | default: | |
10113 | *dt = vect_unknown_def_type; | |
10114 | break; | |
10115 | } | |
0f8c840c | 10116 | } |
30f502ed RB |
10117 | if (def_stmt_out) |
10118 | *def_stmt_out = def_stmt; | |
ebfd146a IR |
10119 | } |
10120 | ||
2e8ab70c RB |
10121 | if (dump_enabled_p ()) |
10122 | { | |
30f502ed | 10123 | dump_printf (MSG_NOTE, ", type of def: "); |
2e8ab70c RB |
10124 | switch (*dt) |
10125 | { | |
10126 | case vect_uninitialized_def: | |
10127 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
10128 | break; | |
10129 | case vect_constant_def: | |
10130 | dump_printf (MSG_NOTE, "constant\n"); | |
10131 | break; | |
10132 | case vect_external_def: | |
10133 | dump_printf (MSG_NOTE, "external\n"); | |
10134 | break; | |
10135 | case vect_internal_def: | |
10136 | dump_printf (MSG_NOTE, "internal\n"); | |
10137 | break; | |
10138 | case vect_induction_def: | |
10139 | dump_printf (MSG_NOTE, "induction\n"); | |
10140 | break; | |
10141 | case vect_reduction_def: | |
10142 | dump_printf (MSG_NOTE, "reduction\n"); | |
10143 | break; | |
10144 | case vect_double_reduction_def: | |
10145 | dump_printf (MSG_NOTE, "double reduction\n"); | |
10146 | break; | |
10147 | case vect_nested_cycle: | |
10148 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
10149 | break; | |
10150 | case vect_unknown_def_type: | |
10151 | dump_printf (MSG_NOTE, "unknown\n"); | |
10152 | break; | |
10153 | } | |
10154 | } | |
10155 | ||
81c40241 | 10156 | if (*dt == vect_unknown_def_type) |
ebfd146a | 10157 | { |
73fbfcad | 10158 | if (dump_enabled_p ()) |
78c60e3d | 10159 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 10160 | "Unsupported pattern.\n"); |
ebfd146a IR |
10161 | return false; |
10162 | } | |
10163 | ||
ebfd146a IR |
10164 | return true; |
10165 | } | |
10166 | ||
81c40241 | 10167 | /* Function vect_is_simple_use. |
b690cc0f | 10168 | |
81c40241 | 10169 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
10170 | type of OPERAND and stores it to *VECTYPE. If the definition of |
10171 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
10172 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
10173 | is responsible to compute the best suited vector type for the | |
10174 | scalar operand. */ | |
10175 | ||
10176 | bool | |
894dd753 RS |
10177 | vect_is_simple_use (tree operand, vec_info *vinfo, enum vect_def_type *dt, |
10178 | tree *vectype, gimple **def_stmt_out) | |
b690cc0f | 10179 | { |
894dd753 RS |
10180 | gimple *def_stmt; |
10181 | if (!vect_is_simple_use (operand, vinfo, dt, &def_stmt)) | |
b690cc0f RG |
10182 | return false; |
10183 | ||
894dd753 RS |
10184 | if (def_stmt_out) |
10185 | *def_stmt_out = def_stmt; | |
10186 | ||
b690cc0f RG |
10187 | /* Now get a vector type if the def is internal, otherwise supply |
10188 | NULL_TREE and leave it up to the caller to figure out a proper | |
10189 | type for the use stmt. */ | |
10190 | if (*dt == vect_internal_def | |
10191 | || *dt == vect_induction_def | |
10192 | || *dt == vect_reduction_def | |
10193 | || *dt == vect_double_reduction_def | |
10194 | || *dt == vect_nested_cycle) | |
10195 | { | |
894dd753 | 10196 | stmt_vec_info stmt_info = vinfo_for_stmt (def_stmt); |
b690cc0f RG |
10197 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
10198 | gcc_assert (*vectype != NULL_TREE); | |
30f502ed RB |
10199 | if (dump_enabled_p ()) |
10200 | { | |
10201 | dump_printf_loc (MSG_NOTE, vect_location, | |
10202 | "vect_is_simple_use: vectype "); | |
10203 | dump_generic_expr (MSG_NOTE, TDF_SLIM, *vectype); | |
10204 | dump_printf (MSG_NOTE, "\n"); | |
10205 | } | |
b690cc0f RG |
10206 | } |
10207 | else if (*dt == vect_uninitialized_def | |
10208 | || *dt == vect_constant_def | |
10209 | || *dt == vect_external_def) | |
10210 | *vectype = NULL_TREE; | |
10211 | else | |
10212 | gcc_unreachable (); | |
10213 | ||
10214 | return true; | |
10215 | } | |
10216 | ||
ebfd146a IR |
10217 | |
10218 | /* Function supportable_widening_operation | |
10219 | ||
b8698a0f L |
10220 | Check whether an operation represented by the code CODE is a |
10221 | widening operation that is supported by the target platform in | |
b690cc0f RG |
10222 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
10223 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 10224 | |
1bda738b JJ |
10225 | Widening operations we currently support are NOP (CONVERT), FLOAT, |
10226 | FIX_TRUNC and WIDEN_MULT. This function checks if these operations | |
10227 | are supported by the target platform either directly (via vector | |
10228 | tree-codes), or via target builtins. | |
ebfd146a IR |
10229 | |
10230 | Output: | |
b8698a0f L |
10231 | - CODE1 and CODE2 are codes of vector operations to be used when |
10232 | vectorizing the operation, if available. | |
ebfd146a IR |
10233 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
10234 | case of multi-step conversion (like char->short->int - in that case | |
10235 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
10236 | - INTERM_TYPES contains the intermediate type required to perform the |
10237 | widening operation (short in the above example). */ | |
ebfd146a IR |
10238 | |
10239 | bool | |
355fe088 | 10240 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 10241 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
10242 | enum tree_code *code1, enum tree_code *code2, |
10243 | int *multi_step_cvt, | |
9771b263 | 10244 | vec<tree> *interm_types) |
ebfd146a IR |
10245 | { |
10246 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
10247 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 10248 | struct loop *vect_loop = NULL; |
ef4bddc2 | 10249 | machine_mode vec_mode; |
81f40b79 | 10250 | enum insn_code icode1, icode2; |
ebfd146a | 10251 | optab optab1, optab2; |
b690cc0f RG |
10252 | tree vectype = vectype_in; |
10253 | tree wide_vectype = vectype_out; | |
ebfd146a | 10254 | enum tree_code c1, c2; |
4a00c761 JJ |
10255 | int i; |
10256 | tree prev_type, intermediate_type; | |
ef4bddc2 | 10257 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 10258 | optab optab3, optab4; |
ebfd146a | 10259 | |
4a00c761 | 10260 | *multi_step_cvt = 0; |
4ef69dfc IR |
10261 | if (loop_info) |
10262 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
10263 | ||
ebfd146a IR |
10264 | switch (code) |
10265 | { | |
10266 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
10267 | /* The result of a vectorized widening operation usually requires |
10268 | two vectors (because the widened results do not fit into one vector). | |
10269 | The generated vector results would normally be expected to be | |
10270 | generated in the same order as in the original scalar computation, | |
10271 | i.e. if 8 results are generated in each vector iteration, they are | |
10272 | to be organized as follows: | |
10273 | vect1: [res1,res2,res3,res4], | |
10274 | vect2: [res5,res6,res7,res8]. | |
10275 | ||
10276 | However, in the special case that the result of the widening | |
10277 | operation is used in a reduction computation only, the order doesn't | |
10278 | matter (because when vectorizing a reduction we change the order of | |
10279 | the computation). Some targets can take advantage of this and | |
10280 | generate more efficient code. For example, targets like Altivec, | |
10281 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
10282 | generate the following vectors: | |
10283 | vect1: [res1,res3,res5,res7], | |
10284 | vect2: [res2,res4,res6,res8]. | |
10285 | ||
10286 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
10287 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
10288 | iterations in parallel). We therefore don't allow to change the | |
10289 | order of the computation in the inner-loop during outer-loop | |
10290 | vectorization. */ | |
10291 | /* TODO: Another case in which order doesn't *really* matter is when we | |
10292 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
10293 | Normally, pack_trunc performs an even/odd permute, whereas the | |
10294 | repack from an even/odd expansion would be an interleave, which | |
10295 | would be significantly simpler for e.g. AVX2. */ | |
10296 | /* In any case, in order to avoid duplicating the code below, recurse | |
10297 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
10298 | are properly set up for the caller. If we fail, we'll continue with | |
10299 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
10300 | if (vect_loop | |
10301 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
10302 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
10303 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
10304 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
10305 | code1, code2, multi_step_cvt, |
10306 | interm_types)) | |
ebc047a2 CH |
10307 | { |
10308 | /* Elements in a vector with vect_used_by_reduction property cannot | |
10309 | be reordered if the use chain with this property does not have the | |
10310 | same operation. One such an example is s += a * b, where elements | |
10311 | in a and b cannot be reordered. Here we check if the vector defined | |
10312 | by STMT is only directly used in the reduction statement. */ | |
10313 | tree lhs = gimple_assign_lhs (stmt); | |
10314 | use_operand_p dummy; | |
355fe088 | 10315 | gimple *use_stmt; |
ebc047a2 CH |
10316 | stmt_vec_info use_stmt_info = NULL; |
10317 | if (single_imm_use (lhs, &dummy, &use_stmt) | |
10318 | && (use_stmt_info = vinfo_for_stmt (use_stmt)) | |
10319 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
10320 | return true; | |
10321 | } | |
4a00c761 JJ |
10322 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
10323 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
10324 | break; |
10325 | ||
81c40241 RB |
10326 | case DOT_PROD_EXPR: |
10327 | c1 = DOT_PROD_EXPR; | |
10328 | c2 = DOT_PROD_EXPR; | |
10329 | break; | |
10330 | ||
10331 | case SAD_EXPR: | |
10332 | c1 = SAD_EXPR; | |
10333 | c2 = SAD_EXPR; | |
10334 | break; | |
10335 | ||
6ae6116f RH |
10336 | case VEC_WIDEN_MULT_EVEN_EXPR: |
10337 | /* Support the recursion induced just above. */ | |
10338 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
10339 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
10340 | break; | |
10341 | ||
36ba4aae | 10342 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
10343 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
10344 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
10345 | break; |
10346 | ||
ebfd146a | 10347 | CASE_CONVERT: |
4a00c761 JJ |
10348 | c1 = VEC_UNPACK_LO_EXPR; |
10349 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
10350 | break; |
10351 | ||
10352 | case FLOAT_EXPR: | |
4a00c761 JJ |
10353 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
10354 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
10355 | break; |
10356 | ||
10357 | case FIX_TRUNC_EXPR: | |
1bda738b JJ |
10358 | c1 = VEC_UNPACK_FIX_TRUNC_LO_EXPR; |
10359 | c2 = VEC_UNPACK_FIX_TRUNC_HI_EXPR; | |
10360 | break; | |
ebfd146a IR |
10361 | |
10362 | default: | |
10363 | gcc_unreachable (); | |
10364 | } | |
10365 | ||
6ae6116f | 10366 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 10367 | std::swap (c1, c2); |
4a00c761 | 10368 | |
ebfd146a IR |
10369 | if (code == FIX_TRUNC_EXPR) |
10370 | { | |
10371 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
10372 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
10373 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
10374 | } |
10375 | else | |
10376 | { | |
10377 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
10378 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
10379 | } | |
10380 | ||
10381 | if (!optab1 || !optab2) | |
10382 | return false; | |
10383 | ||
10384 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
10385 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
10386 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
10387 | return false; |
10388 | ||
4a00c761 JJ |
10389 | *code1 = c1; |
10390 | *code2 = c2; | |
10391 | ||
10392 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
10393 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
10394 | /* For scalar masks we may have different boolean |
10395 | vector types having the same QImode. Thus we | |
10396 | add additional check for elements number. */ | |
10397 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
928686b1 RS |
10398 | || known_eq (TYPE_VECTOR_SUBPARTS (vectype), |
10399 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)); | |
4a00c761 | 10400 | |
b8698a0f | 10401 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 10402 | types. */ |
ebfd146a | 10403 | |
4a00c761 JJ |
10404 | prev_type = vectype; |
10405 | prev_mode = vec_mode; | |
b8698a0f | 10406 | |
4a00c761 JJ |
10407 | if (!CONVERT_EXPR_CODE_P (code)) |
10408 | return false; | |
b8698a0f | 10409 | |
4a00c761 JJ |
10410 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
10411 | intermediate steps in promotion sequence. We try | |
10412 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
10413 | not. */ | |
9771b263 | 10414 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
10415 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
10416 | { | |
10417 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
10418 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
10419 | { | |
7cfb4d93 | 10420 | intermediate_type = vect_halve_mask_nunits (prev_type); |
3ae0661a IE |
10421 | if (intermediate_mode != TYPE_MODE (intermediate_type)) |
10422 | return false; | |
10423 | } | |
10424 | else | |
10425 | intermediate_type | |
10426 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
10427 | TYPE_UNSIGNED (prev_type)); | |
10428 | ||
4a00c761 JJ |
10429 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); |
10430 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
10431 | ||
10432 | if (!optab3 || !optab4 | |
10433 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
10434 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
10435 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
10436 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
10437 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
10438 | == CODE_FOR_nothing) | |
10439 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
10440 | == CODE_FOR_nothing)) | |
10441 | break; | |
ebfd146a | 10442 | |
9771b263 | 10443 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
10444 | (*multi_step_cvt)++; |
10445 | ||
10446 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
10447 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff | 10448 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
928686b1 RS |
10449 | || known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type), |
10450 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)); | |
4a00c761 JJ |
10451 | |
10452 | prev_type = intermediate_type; | |
10453 | prev_mode = intermediate_mode; | |
ebfd146a IR |
10454 | } |
10455 | ||
9771b263 | 10456 | interm_types->release (); |
4a00c761 | 10457 | return false; |
ebfd146a IR |
10458 | } |
10459 | ||
10460 | ||
10461 | /* Function supportable_narrowing_operation | |
10462 | ||
b8698a0f L |
10463 | Check whether an operation represented by the code CODE is a |
10464 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
10465 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
10466 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 10467 | |
1bda738b JJ |
10468 | Narrowing operations we currently support are NOP (CONVERT), FIX_TRUNC |
10469 | and FLOAT. This function checks if these operations are supported by | |
ebfd146a IR |
10470 | the target platform directly via vector tree-codes. |
10471 | ||
10472 | Output: | |
b8698a0f L |
10473 | - CODE1 is the code of a vector operation to be used when |
10474 | vectorizing the operation, if available. | |
ebfd146a IR |
10475 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
10476 | case of multi-step conversion (like int->short->char - in that case | |
10477 | MULTI_STEP_CVT will be 1). | |
10478 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 10479 | narrowing operation (short in the above example). */ |
ebfd146a IR |
10480 | |
10481 | bool | |
10482 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 10483 | tree vectype_out, tree vectype_in, |
ebfd146a | 10484 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 10485 | vec<tree> *interm_types) |
ebfd146a | 10486 | { |
ef4bddc2 | 10487 | machine_mode vec_mode; |
ebfd146a IR |
10488 | enum insn_code icode1; |
10489 | optab optab1, interm_optab; | |
b690cc0f RG |
10490 | tree vectype = vectype_in; |
10491 | tree narrow_vectype = vectype_out; | |
ebfd146a | 10492 | enum tree_code c1; |
3ae0661a | 10493 | tree intermediate_type, prev_type; |
ef4bddc2 | 10494 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 10495 | int i; |
4a00c761 | 10496 | bool uns; |
ebfd146a | 10497 | |
4a00c761 | 10498 | *multi_step_cvt = 0; |
ebfd146a IR |
10499 | switch (code) |
10500 | { | |
10501 | CASE_CONVERT: | |
10502 | c1 = VEC_PACK_TRUNC_EXPR; | |
10503 | break; | |
10504 | ||
10505 | case FIX_TRUNC_EXPR: | |
10506 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
10507 | break; | |
10508 | ||
10509 | case FLOAT_EXPR: | |
1bda738b JJ |
10510 | c1 = VEC_PACK_FLOAT_EXPR; |
10511 | break; | |
ebfd146a IR |
10512 | |
10513 | default: | |
10514 | gcc_unreachable (); | |
10515 | } | |
10516 | ||
10517 | if (code == FIX_TRUNC_EXPR) | |
10518 | /* The signedness is determined from output operand. */ | |
b690cc0f | 10519 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
10520 | else |
10521 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
10522 | ||
10523 | if (!optab1) | |
10524 | return false; | |
10525 | ||
10526 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 10527 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
10528 | return false; |
10529 | ||
4a00c761 JJ |
10530 | *code1 = c1; |
10531 | ||
10532 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
10533 | /* For scalar masks we may have different boolean |
10534 | vector types having the same QImode. Thus we | |
10535 | add additional check for elements number. */ | |
10536 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
928686b1 RS |
10537 | || known_eq (TYPE_VECTOR_SUBPARTS (vectype) * 2, |
10538 | TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 | 10539 | |
1bda738b JJ |
10540 | if (code == FLOAT_EXPR) |
10541 | return false; | |
10542 | ||
ebfd146a IR |
10543 | /* Check if it's a multi-step conversion that can be done using intermediate |
10544 | types. */ | |
4a00c761 | 10545 | prev_mode = vec_mode; |
3ae0661a | 10546 | prev_type = vectype; |
4a00c761 JJ |
10547 | if (code == FIX_TRUNC_EXPR) |
10548 | uns = TYPE_UNSIGNED (vectype_out); | |
10549 | else | |
10550 | uns = TYPE_UNSIGNED (vectype); | |
10551 | ||
10552 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
10553 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
10554 | costly than signed. */ | |
10555 | if (code == FIX_TRUNC_EXPR && uns) | |
10556 | { | |
10557 | enum insn_code icode2; | |
10558 | ||
10559 | intermediate_type | |
10560 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
10561 | interm_optab | |
10562 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 10563 | if (interm_optab != unknown_optab |
4a00c761 JJ |
10564 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
10565 | && insn_data[icode1].operand[0].mode | |
10566 | == insn_data[icode2].operand[0].mode) | |
10567 | { | |
10568 | uns = false; | |
10569 | optab1 = interm_optab; | |
10570 | icode1 = icode2; | |
10571 | } | |
10572 | } | |
ebfd146a | 10573 | |
4a00c761 JJ |
10574 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
10575 | intermediate steps in promotion sequence. We try | |
10576 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 10577 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
10578 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
10579 | { | |
10580 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
10581 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
10582 | { | |
7cfb4d93 | 10583 | intermediate_type = vect_double_mask_nunits (prev_type); |
3ae0661a | 10584 | if (intermediate_mode != TYPE_MODE (intermediate_type)) |
7cfb4d93 | 10585 | return false; |
3ae0661a IE |
10586 | } |
10587 | else | |
10588 | intermediate_type | |
10589 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
4a00c761 JJ |
10590 | interm_optab |
10591 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
10592 | optab_default); | |
10593 | if (!interm_optab | |
10594 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
10595 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
10596 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
10597 | == CODE_FOR_nothing)) | |
10598 | break; | |
10599 | ||
9771b263 | 10600 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
10601 | (*multi_step_cvt)++; |
10602 | ||
10603 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff | 10604 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
928686b1 RS |
10605 | || known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type) * 2, |
10606 | TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 JJ |
10607 | |
10608 | prev_mode = intermediate_mode; | |
3ae0661a | 10609 | prev_type = intermediate_type; |
4a00c761 | 10610 | optab1 = interm_optab; |
ebfd146a IR |
10611 | } |
10612 | ||
9771b263 | 10613 | interm_types->release (); |
4a00c761 | 10614 | return false; |
ebfd146a | 10615 | } |
7cfb4d93 RS |
10616 | |
10617 | /* Generate and return a statement that sets vector mask MASK such that | |
10618 | MASK[I] is true iff J + START_INDEX < END_INDEX for all J <= I. */ | |
10619 | ||
10620 | gcall * | |
10621 | vect_gen_while (tree mask, tree start_index, tree end_index) | |
10622 | { | |
10623 | tree cmp_type = TREE_TYPE (start_index); | |
10624 | tree mask_type = TREE_TYPE (mask); | |
10625 | gcc_checking_assert (direct_internal_fn_supported_p (IFN_WHILE_ULT, | |
10626 | cmp_type, mask_type, | |
10627 | OPTIMIZE_FOR_SPEED)); | |
10628 | gcall *call = gimple_build_call_internal (IFN_WHILE_ULT, 3, | |
10629 | start_index, end_index, | |
10630 | build_zero_cst (mask_type)); | |
10631 | gimple_call_set_lhs (call, mask); | |
10632 | return call; | |
10633 | } | |
535e7c11 RS |
10634 | |
10635 | /* Generate a vector mask of type MASK_TYPE for which index I is false iff | |
10636 | J + START_INDEX < END_INDEX for all J <= I. Add the statements to SEQ. */ | |
10637 | ||
10638 | tree | |
10639 | vect_gen_while_not (gimple_seq *seq, tree mask_type, tree start_index, | |
10640 | tree end_index) | |
10641 | { | |
10642 | tree tmp = make_ssa_name (mask_type); | |
10643 | gcall *call = vect_gen_while (tmp, start_index, end_index); | |
10644 | gimple_seq_add_stmt (seq, call); | |
10645 | return gimple_build (seq, BIT_NOT_EXPR, mask_type, tmp); | |
10646 | } | |
1f3cb663 RS |
10647 | |
10648 | /* Try to compute the vector types required to vectorize STMT_INFO, | |
10649 | returning true on success and false if vectorization isn't possible. | |
10650 | ||
10651 | On success: | |
10652 | ||
10653 | - Set *STMT_VECTYPE_OUT to: | |
10654 | - NULL_TREE if the statement doesn't need to be vectorized; | |
10655 | - boolean_type_node if the statement is a boolean operation whose | |
10656 | vector type can only be determined once all the other vector types | |
10657 | are known; and | |
10658 | - the equivalent of STMT_VINFO_VECTYPE otherwise. | |
10659 | ||
10660 | - Set *NUNITS_VECTYPE_OUT to the vector type that contains the maximum | |
10661 | number of units needed to vectorize STMT_INFO, or NULL_TREE if the | |
10662 | statement does not help to determine the overall number of units. */ | |
10663 | ||
10664 | bool | |
10665 | vect_get_vector_types_for_stmt (stmt_vec_info stmt_info, | |
10666 | tree *stmt_vectype_out, | |
10667 | tree *nunits_vectype_out) | |
10668 | { | |
10669 | gimple *stmt = stmt_info->stmt; | |
10670 | ||
10671 | *stmt_vectype_out = NULL_TREE; | |
10672 | *nunits_vectype_out = NULL_TREE; | |
10673 | ||
10674 | if (gimple_get_lhs (stmt) == NULL_TREE | |
10675 | /* MASK_STORE has no lhs, but is ok. */ | |
10676 | && !gimple_call_internal_p (stmt, IFN_MASK_STORE)) | |
10677 | { | |
10678 | if (is_a <gcall *> (stmt)) | |
10679 | { | |
10680 | /* Ignore calls with no lhs. These must be calls to | |
10681 | #pragma omp simd functions, and what vectorization factor | |
10682 | it really needs can't be determined until | |
10683 | vectorizable_simd_clone_call. */ | |
10684 | if (dump_enabled_p ()) | |
10685 | dump_printf_loc (MSG_NOTE, vect_location, | |
10686 | "defer to SIMD clone analysis.\n"); | |
10687 | return true; | |
10688 | } | |
10689 | ||
10690 | if (dump_enabled_p ()) | |
10691 | { | |
10692 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10693 | "not vectorized: irregular stmt."); | |
10694 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
10695 | } | |
10696 | return false; | |
10697 | } | |
10698 | ||
10699 | if (VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))) | |
10700 | { | |
10701 | if (dump_enabled_p ()) | |
10702 | { | |
10703 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10704 | "not vectorized: vector stmt in loop:"); | |
10705 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
10706 | } | |
10707 | return false; | |
10708 | } | |
10709 | ||
10710 | tree vectype; | |
10711 | tree scalar_type = NULL_TREE; | |
10712 | if (STMT_VINFO_VECTYPE (stmt_info)) | |
10713 | *stmt_vectype_out = vectype = STMT_VINFO_VECTYPE (stmt_info); | |
10714 | else | |
10715 | { | |
10716 | gcc_assert (!STMT_VINFO_DATA_REF (stmt_info)); | |
10717 | if (gimple_call_internal_p (stmt, IFN_MASK_STORE)) | |
10718 | scalar_type = TREE_TYPE (gimple_call_arg (stmt, 3)); | |
10719 | else | |
10720 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); | |
10721 | ||
10722 | /* Pure bool ops don't participate in number-of-units computation. | |
10723 | For comparisons use the types being compared. */ | |
10724 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type) | |
10725 | && is_gimple_assign (stmt) | |
10726 | && gimple_assign_rhs_code (stmt) != COND_EXPR) | |
10727 | { | |
10728 | *stmt_vectype_out = boolean_type_node; | |
10729 | ||
10730 | tree rhs1 = gimple_assign_rhs1 (stmt); | |
10731 | if (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_comparison | |
10732 | && !VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (rhs1))) | |
10733 | scalar_type = TREE_TYPE (rhs1); | |
10734 | else | |
10735 | { | |
10736 | if (dump_enabled_p ()) | |
10737 | dump_printf_loc (MSG_NOTE, vect_location, | |
10738 | "pure bool operation.\n"); | |
10739 | return true; | |
10740 | } | |
10741 | } | |
10742 | ||
10743 | if (dump_enabled_p ()) | |
10744 | { | |
10745 | dump_printf_loc (MSG_NOTE, vect_location, | |
10746 | "get vectype for scalar type: "); | |
10747 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
10748 | dump_printf (MSG_NOTE, "\n"); | |
10749 | } | |
10750 | vectype = get_vectype_for_scalar_type (scalar_type); | |
10751 | if (!vectype) | |
10752 | { | |
10753 | if (dump_enabled_p ()) | |
10754 | { | |
10755 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10756 | "not vectorized: unsupported data-type "); | |
10757 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10758 | scalar_type); | |
10759 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10760 | } | |
10761 | return false; | |
10762 | } | |
10763 | ||
10764 | if (!*stmt_vectype_out) | |
10765 | *stmt_vectype_out = vectype; | |
10766 | ||
10767 | if (dump_enabled_p ()) | |
10768 | { | |
10769 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); | |
10770 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
10771 | dump_printf (MSG_NOTE, "\n"); | |
10772 | } | |
10773 | } | |
10774 | ||
10775 | /* Don't try to compute scalar types if the stmt produces a boolean | |
10776 | vector; use the existing vector type instead. */ | |
10777 | tree nunits_vectype; | |
10778 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) | |
10779 | nunits_vectype = vectype; | |
10780 | else | |
10781 | { | |
10782 | /* The number of units is set according to the smallest scalar | |
10783 | type (or the largest vector size, but we only support one | |
10784 | vector size per vectorization). */ | |
10785 | if (*stmt_vectype_out != boolean_type_node) | |
10786 | { | |
10787 | HOST_WIDE_INT dummy; | |
10788 | scalar_type = vect_get_smallest_scalar_type (stmt, &dummy, &dummy); | |
10789 | } | |
10790 | if (dump_enabled_p ()) | |
10791 | { | |
10792 | dump_printf_loc (MSG_NOTE, vect_location, | |
10793 | "get vectype for scalar type: "); | |
10794 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
10795 | dump_printf (MSG_NOTE, "\n"); | |
10796 | } | |
10797 | nunits_vectype = get_vectype_for_scalar_type (scalar_type); | |
10798 | } | |
10799 | if (!nunits_vectype) | |
10800 | { | |
10801 | if (dump_enabled_p ()) | |
10802 | { | |
10803 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10804 | "not vectorized: unsupported data-type "); | |
10805 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, scalar_type); | |
10806 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10807 | } | |
10808 | return false; | |
10809 | } | |
10810 | ||
10811 | if (maybe_ne (GET_MODE_SIZE (TYPE_MODE (vectype)), | |
10812 | GET_MODE_SIZE (TYPE_MODE (nunits_vectype)))) | |
10813 | { | |
10814 | if (dump_enabled_p ()) | |
10815 | { | |
10816 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10817 | "not vectorized: different sized vector " | |
10818 | "types in statement, "); | |
10819 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, vectype); | |
10820 | dump_printf (MSG_MISSED_OPTIMIZATION, " and "); | |
10821 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, nunits_vectype); | |
10822 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10823 | } | |
10824 | return false; | |
10825 | } | |
10826 | ||
10827 | if (dump_enabled_p ()) | |
10828 | { | |
10829 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); | |
10830 | dump_generic_expr (MSG_NOTE, TDF_SLIM, nunits_vectype); | |
10831 | dump_printf (MSG_NOTE, "\n"); | |
10832 | ||
10833 | dump_printf_loc (MSG_NOTE, vect_location, "nunits = "); | |
10834 | dump_dec (MSG_NOTE, TYPE_VECTOR_SUBPARTS (nunits_vectype)); | |
10835 | dump_printf (MSG_NOTE, "\n"); | |
10836 | } | |
10837 | ||
10838 | *nunits_vectype_out = nunits_vectype; | |
10839 | return true; | |
10840 | } | |
10841 | ||
10842 | /* Try to determine the correct vector type for STMT_INFO, which is a | |
10843 | statement that produces a scalar boolean result. Return the vector | |
10844 | type on success, otherwise return NULL_TREE. */ | |
10845 | ||
10846 | tree | |
10847 | vect_get_mask_type_for_stmt (stmt_vec_info stmt_info) | |
10848 | { | |
10849 | gimple *stmt = stmt_info->stmt; | |
10850 | tree mask_type = NULL; | |
10851 | tree vectype, scalar_type; | |
10852 | ||
10853 | if (is_gimple_assign (stmt) | |
10854 | && TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_comparison | |
10855 | && !VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (gimple_assign_rhs1 (stmt)))) | |
10856 | { | |
10857 | scalar_type = TREE_TYPE (gimple_assign_rhs1 (stmt)); | |
10858 | mask_type = get_mask_type_for_scalar_type (scalar_type); | |
10859 | ||
10860 | if (!mask_type) | |
10861 | { | |
10862 | if (dump_enabled_p ()) | |
10863 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10864 | "not vectorized: unsupported mask\n"); | |
10865 | return NULL_TREE; | |
10866 | } | |
10867 | } | |
10868 | else | |
10869 | { | |
10870 | tree rhs; | |
10871 | ssa_op_iter iter; | |
1f3cb663 RS |
10872 | enum vect_def_type dt; |
10873 | ||
10874 | FOR_EACH_SSA_TREE_OPERAND (rhs, stmt, iter, SSA_OP_USE) | |
10875 | { | |
894dd753 | 10876 | if (!vect_is_simple_use (rhs, stmt_info->vinfo, &dt, &vectype)) |
1f3cb663 RS |
10877 | { |
10878 | if (dump_enabled_p ()) | |
10879 | { | |
10880 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10881 | "not vectorized: can't compute mask type " | |
10882 | "for statement, "); | |
10883 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, | |
10884 | 0); | |
10885 | } | |
10886 | return NULL_TREE; | |
10887 | } | |
10888 | ||
10889 | /* No vectype probably means external definition. | |
10890 | Allow it in case there is another operand which | |
10891 | allows to determine mask type. */ | |
10892 | if (!vectype) | |
10893 | continue; | |
10894 | ||
10895 | if (!mask_type) | |
10896 | mask_type = vectype; | |
10897 | else if (maybe_ne (TYPE_VECTOR_SUBPARTS (mask_type), | |
10898 | TYPE_VECTOR_SUBPARTS (vectype))) | |
10899 | { | |
10900 | if (dump_enabled_p ()) | |
10901 | { | |
10902 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10903 | "not vectorized: different sized masks " | |
10904 | "types in statement, "); | |
10905 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10906 | mask_type); | |
10907 | dump_printf (MSG_MISSED_OPTIMIZATION, " and "); | |
10908 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10909 | vectype); | |
10910 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10911 | } | |
10912 | return NULL_TREE; | |
10913 | } | |
10914 | else if (VECTOR_BOOLEAN_TYPE_P (mask_type) | |
10915 | != VECTOR_BOOLEAN_TYPE_P (vectype)) | |
10916 | { | |
10917 | if (dump_enabled_p ()) | |
10918 | { | |
10919 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10920 | "not vectorized: mixed mask and " | |
10921 | "nonmask vector types in statement, "); | |
10922 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10923 | mask_type); | |
10924 | dump_printf (MSG_MISSED_OPTIMIZATION, " and "); | |
10925 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10926 | vectype); | |
10927 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10928 | } | |
10929 | return NULL_TREE; | |
10930 | } | |
10931 | } | |
10932 | ||
10933 | /* We may compare boolean value loaded as vector of integers. | |
10934 | Fix mask_type in such case. */ | |
10935 | if (mask_type | |
10936 | && !VECTOR_BOOLEAN_TYPE_P (mask_type) | |
10937 | && gimple_code (stmt) == GIMPLE_ASSIGN | |
10938 | && TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_comparison) | |
10939 | mask_type = build_same_sized_truth_vector_type (mask_type); | |
10940 | } | |
10941 | ||
10942 | /* No mask_type should mean loop invariant predicate. | |
10943 | This is probably a subject for optimization in if-conversion. */ | |
10944 | if (!mask_type && dump_enabled_p ()) | |
10945 | { | |
10946 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10947 | "not vectorized: can't compute mask type " | |
10948 | "for statement, "); | |
10949 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
10950 | } | |
10951 | return mask_type; | |
10952 | } |