]>
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); | |
205 | ||
73fbfcad | 206 | if (dump_enabled_p ()) |
66c16fd9 RB |
207 | { |
208 | dump_printf_loc (MSG_NOTE, vect_location, | |
209 | "mark relevant %d, live %d: ", relevant, live_p); | |
210 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
211 | } | |
ebfd146a | 212 | |
83197f37 IR |
213 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
214 | related pattern stmt instead of the original stmt. However, such stmts | |
215 | may have their own uses that are not in any pattern, in such cases the | |
216 | stmt itself should be marked. */ | |
ebfd146a IR |
217 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
218 | { | |
97ecdb46 JJ |
219 | /* This is the last stmt in a sequence that was detected as a |
220 | pattern that can potentially be vectorized. Don't mark the stmt | |
221 | as relevant/live because it's not going to be vectorized. | |
222 | Instead mark the pattern-stmt that replaces it. */ | |
83197f37 | 223 | |
97ecdb46 JJ |
224 | if (dump_enabled_p ()) |
225 | dump_printf_loc (MSG_NOTE, vect_location, | |
226 | "last stmt in pattern. don't mark" | |
227 | " relevant/live.\n"); | |
10681ce8 RS |
228 | stmt_vec_info old_stmt_info = stmt_info; |
229 | stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); | |
230 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == old_stmt_info); | |
97ecdb46 JJ |
231 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); |
232 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
10681ce8 | 233 | stmt = stmt_info->stmt; |
ebfd146a IR |
234 | } |
235 | ||
236 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
237 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
238 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
239 | ||
240 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
241 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
242 | { | |
73fbfcad | 243 | if (dump_enabled_p ()) |
78c60e3d | 244 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 245 | "already marked relevant/live.\n"); |
ebfd146a IR |
246 | return; |
247 | } | |
248 | ||
9771b263 | 249 | worklist->safe_push (stmt); |
ebfd146a IR |
250 | } |
251 | ||
252 | ||
b28ead45 AH |
253 | /* Function is_simple_and_all_uses_invariant |
254 | ||
255 | Return true if STMT is simple and all uses of it are invariant. */ | |
256 | ||
257 | bool | |
258 | is_simple_and_all_uses_invariant (gimple *stmt, loop_vec_info loop_vinfo) | |
259 | { | |
260 | tree op; | |
b28ead45 AH |
261 | ssa_op_iter iter; |
262 | ||
263 | if (!is_gimple_assign (stmt)) | |
264 | return false; | |
265 | ||
266 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
267 | { | |
268 | enum vect_def_type dt = vect_uninitialized_def; | |
269 | ||
894dd753 | 270 | if (!vect_is_simple_use (op, loop_vinfo, &dt)) |
b28ead45 AH |
271 | { |
272 | if (dump_enabled_p ()) | |
273 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
274 | "use not simple.\n"); | |
275 | return false; | |
276 | } | |
277 | ||
278 | if (dt != vect_external_def && dt != vect_constant_def) | |
279 | return false; | |
280 | } | |
281 | return true; | |
282 | } | |
283 | ||
ebfd146a IR |
284 | /* Function vect_stmt_relevant_p. |
285 | ||
286 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
287 | "relevant for vectorization". | |
288 | ||
289 | A stmt is considered "relevant for vectorization" if: | |
290 | - it has uses outside the loop. | |
291 | - it has vdefs (it alters memory). | |
292 | - control stmts in the loop (except for the exit condition). | |
293 | ||
294 | CHECKME: what other side effects would the vectorizer allow? */ | |
295 | ||
296 | static bool | |
355fe088 | 297 | vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo, |
ebfd146a IR |
298 | enum vect_relevant *relevant, bool *live_p) |
299 | { | |
300 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
301 | ssa_op_iter op_iter; | |
302 | imm_use_iterator imm_iter; | |
303 | use_operand_p use_p; | |
304 | def_operand_p def_p; | |
305 | ||
8644a673 | 306 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
307 | *live_p = false; |
308 | ||
309 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
310 | if (is_ctrl_stmt (stmt) |
311 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
312 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 313 | *relevant = vect_used_in_scope; |
ebfd146a IR |
314 | |
315 | /* changing memory. */ | |
316 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
317 | if (gimple_vdef (stmt) |
318 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 319 | { |
73fbfcad | 320 | if (dump_enabled_p ()) |
78c60e3d | 321 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 322 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 323 | *relevant = vect_used_in_scope; |
ebfd146a IR |
324 | } |
325 | ||
326 | /* uses outside the loop. */ | |
327 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
328 | { | |
329 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
330 | { | |
331 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
332 | if (!flow_bb_inside_loop_p (loop, bb)) | |
333 | { | |
73fbfcad | 334 | if (dump_enabled_p ()) |
78c60e3d | 335 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 336 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 337 | |
3157b0c2 AO |
338 | if (is_gimple_debug (USE_STMT (use_p))) |
339 | continue; | |
340 | ||
ebfd146a IR |
341 | /* We expect all such uses to be in the loop exit phis |
342 | (because of loop closed form) */ | |
343 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
344 | gcc_assert (bb == single_exit (loop)->dest); | |
345 | ||
346 | *live_p = true; | |
347 | } | |
348 | } | |
349 | } | |
350 | ||
3a2edf4c AH |
351 | if (*live_p && *relevant == vect_unused_in_scope |
352 | && !is_simple_and_all_uses_invariant (stmt, loop_vinfo)) | |
b28ead45 AH |
353 | { |
354 | if (dump_enabled_p ()) | |
355 | dump_printf_loc (MSG_NOTE, vect_location, | |
356 | "vec_stmt_relevant_p: stmt live but not relevant.\n"); | |
357 | *relevant = vect_used_only_live; | |
358 | } | |
359 | ||
ebfd146a IR |
360 | return (*live_p || *relevant); |
361 | } | |
362 | ||
363 | ||
b8698a0f | 364 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 365 | |
ff802fa1 | 366 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
367 | used in STMT for anything other than indexing an array. */ |
368 | ||
369 | static bool | |
355fe088 | 370 | exist_non_indexing_operands_for_use_p (tree use, gimple *stmt) |
ebfd146a IR |
371 | { |
372 | tree operand; | |
373 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 374 | |
ff802fa1 | 375 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
376 | reference in STMT, then any operand that corresponds to USE |
377 | is not indexing an array. */ | |
378 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
379 | return true; | |
59a05b0c | 380 | |
ebfd146a IR |
381 | /* STMT has a data_ref. FORNOW this means that its of one of |
382 | the following forms: | |
383 | -1- ARRAY_REF = var | |
384 | -2- var = ARRAY_REF | |
385 | (This should have been verified in analyze_data_refs). | |
386 | ||
387 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 388 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
389 | for array indexing. |
390 | ||
391 | Therefore, all we need to check is if STMT falls into the | |
392 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
393 | |
394 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
395 | { |
396 | if (is_gimple_call (stmt) | |
397 | && gimple_call_internal_p (stmt)) | |
bfaa08b7 RS |
398 | { |
399 | internal_fn ifn = gimple_call_internal_fn (stmt); | |
400 | int mask_index = internal_fn_mask_index (ifn); | |
401 | if (mask_index >= 0 | |
402 | && use == gimple_call_arg (stmt, mask_index)) | |
403 | return true; | |
f307441a RS |
404 | int stored_value_index = internal_fn_stored_value_index (ifn); |
405 | if (stored_value_index >= 0 | |
406 | && use == gimple_call_arg (stmt, stored_value_index)) | |
407 | return true; | |
bfaa08b7 RS |
408 | if (internal_gather_scatter_fn_p (ifn) |
409 | && use == gimple_call_arg (stmt, 1)) | |
410 | return true; | |
bfaa08b7 | 411 | } |
5ce9450f JJ |
412 | return false; |
413 | } | |
414 | ||
59a05b0c EB |
415 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
416 | return false; | |
ebfd146a | 417 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
418 | if (TREE_CODE (operand) != SSA_NAME) |
419 | return false; | |
420 | ||
421 | if (operand == use) | |
422 | return true; | |
423 | ||
424 | return false; | |
425 | } | |
426 | ||
427 | ||
b8698a0f | 428 | /* |
ebfd146a IR |
429 | Function process_use. |
430 | ||
431 | Inputs: | |
432 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b28ead45 | 433 | - RELEVANT - enum value to be set in the STMT_VINFO of the stmt |
ff802fa1 | 434 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 435 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
436 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
437 | be performed. | |
ebfd146a IR |
438 | |
439 | Outputs: | |
440 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
441 | relevance info of the DEF_STMT of this USE: | |
442 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
443 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
444 | Exceptions: | |
445 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 446 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 447 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
448 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
449 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
450 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
451 | be modified accordingly. | |
452 | ||
453 | Return true if everything is as expected. Return false otherwise. */ | |
454 | ||
455 | static bool | |
b28ead45 | 456 | process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo, |
355fe088 | 457 | enum vect_relevant relevant, vec<gimple *> *worklist, |
aec7ae7d | 458 | bool force) |
ebfd146a | 459 | { |
ebfd146a IR |
460 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
461 | stmt_vec_info dstmt_vinfo; | |
462 | basic_block bb, def_bb; | |
ebfd146a IR |
463 | enum vect_def_type dt; |
464 | ||
b8698a0f | 465 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 466 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 467 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
468 | return true; |
469 | ||
fef96d8e | 470 | if (!vect_is_simple_use (use, loop_vinfo, &dt, &dstmt_vinfo)) |
b8698a0f | 471 | { |
73fbfcad | 472 | if (dump_enabled_p ()) |
78c60e3d | 473 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 474 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
475 | return false; |
476 | } | |
477 | ||
fef96d8e | 478 | if (!dstmt_vinfo) |
ebfd146a IR |
479 | return true; |
480 | ||
fef96d8e | 481 | def_bb = gimple_bb (dstmt_vinfo->stmt); |
ebfd146a | 482 | |
fef96d8e RS |
483 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DSTMT_VINFO). |
484 | DSTMT_VINFO must have already been processed, because this should be the | |
b8698a0f | 485 | only way that STMT, which is a reduction-phi, was put in the worklist, |
fef96d8e | 486 | as there should be no other uses for DSTMT_VINFO in the loop. So we just |
ebfd146a | 487 | check that everything is as expected, and we are done. */ |
ebfd146a IR |
488 | bb = gimple_bb (stmt); |
489 | if (gimple_code (stmt) == GIMPLE_PHI | |
490 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
fef96d8e | 491 | && gimple_code (dstmt_vinfo->stmt) != GIMPLE_PHI |
ebfd146a IR |
492 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def |
493 | && bb->loop_father == def_bb->loop_father) | |
494 | { | |
73fbfcad | 495 | if (dump_enabled_p ()) |
78c60e3d | 496 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 497 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a | 498 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); |
b8698a0f | 499 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 500 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
501 | return true; |
502 | } | |
503 | ||
504 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
505 | outer-loop-header-bb: | |
fef96d8e | 506 | d = dstmt_vinfo |
ebfd146a IR |
507 | inner-loop: |
508 | stmt # use (d) | |
509 | outer-loop-tail-bb: | |
510 | ... */ | |
511 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
512 | { | |
73fbfcad | 513 | if (dump_enabled_p ()) |
78c60e3d | 514 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 515 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 516 | |
ebfd146a IR |
517 | switch (relevant) |
518 | { | |
8644a673 | 519 | case vect_unused_in_scope: |
7c5222ff IR |
520 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
521 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 522 | break; |
7c5222ff | 523 | |
ebfd146a | 524 | case vect_used_in_outer_by_reduction: |
7c5222ff | 525 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
526 | relevant = vect_used_by_reduction; |
527 | break; | |
7c5222ff | 528 | |
ebfd146a | 529 | case vect_used_in_outer: |
7c5222ff | 530 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 531 | relevant = vect_used_in_scope; |
ebfd146a | 532 | break; |
7c5222ff | 533 | |
8644a673 | 534 | case vect_used_in_scope: |
ebfd146a IR |
535 | break; |
536 | ||
537 | default: | |
538 | gcc_unreachable (); | |
b8698a0f | 539 | } |
ebfd146a IR |
540 | } |
541 | ||
542 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
543 | outer-loop-header-bb: | |
544 | ... | |
545 | inner-loop: | |
fef96d8e | 546 | d = dstmt_vinfo |
06066f92 | 547 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
548 | stmt # use (d) */ |
549 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
550 | { | |
73fbfcad | 551 | if (dump_enabled_p ()) |
78c60e3d | 552 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 553 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 554 | |
ebfd146a IR |
555 | switch (relevant) |
556 | { | |
8644a673 | 557 | case vect_unused_in_scope: |
b8698a0f | 558 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 559 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 560 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
561 | break; |
562 | ||
ebfd146a | 563 | case vect_used_by_reduction: |
b28ead45 | 564 | case vect_used_only_live: |
ebfd146a IR |
565 | relevant = vect_used_in_outer_by_reduction; |
566 | break; | |
567 | ||
8644a673 | 568 | case vect_used_in_scope: |
ebfd146a IR |
569 | relevant = vect_used_in_outer; |
570 | break; | |
571 | ||
572 | default: | |
573 | gcc_unreachable (); | |
574 | } | |
575 | } | |
643a9684 RB |
576 | /* We are also not interested in uses on loop PHI backedges that are |
577 | inductions. Otherwise we'll needlessly vectorize the IV increment | |
e294f495 RB |
578 | and cause hybrid SLP for SLP inductions. Unless the PHI is live |
579 | of course. */ | |
643a9684 RB |
580 | else if (gimple_code (stmt) == GIMPLE_PHI |
581 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_induction_def | |
e294f495 | 582 | && ! STMT_VINFO_LIVE_P (stmt_vinfo) |
643a9684 RB |
583 | && (PHI_ARG_DEF_FROM_EDGE (stmt, loop_latch_edge (bb->loop_father)) |
584 | == use)) | |
585 | { | |
586 | if (dump_enabled_p ()) | |
587 | dump_printf_loc (MSG_NOTE, vect_location, | |
588 | "induction value on backedge.\n"); | |
589 | return true; | |
590 | } | |
591 | ||
ebfd146a | 592 | |
fef96d8e | 593 | vect_mark_relevant (worklist, dstmt_vinfo, relevant, false); |
ebfd146a IR |
594 | return true; |
595 | } | |
596 | ||
597 | ||
598 | /* Function vect_mark_stmts_to_be_vectorized. | |
599 | ||
600 | Not all stmts in the loop need to be vectorized. For example: | |
601 | ||
602 | for i... | |
603 | for j... | |
604 | 1. T0 = i + j | |
605 | 2. T1 = a[T0] | |
606 | ||
607 | 3. j = j + 1 | |
608 | ||
609 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
610 | addressing of vectorized data-refs are handled differently. | |
611 | ||
612 | This pass detects such stmts. */ | |
613 | ||
614 | bool | |
615 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
616 | { | |
ebfd146a IR |
617 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
618 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
619 | unsigned int nbbs = loop->num_nodes; | |
620 | gimple_stmt_iterator si; | |
355fe088 | 621 | gimple *stmt; |
ebfd146a IR |
622 | unsigned int i; |
623 | stmt_vec_info stmt_vinfo; | |
624 | basic_block bb; | |
355fe088 | 625 | gimple *phi; |
ebfd146a | 626 | bool live_p; |
b28ead45 | 627 | enum vect_relevant relevant; |
ebfd146a | 628 | |
adac3a68 | 629 | DUMP_VECT_SCOPE ("vect_mark_stmts_to_be_vectorized"); |
ebfd146a | 630 | |
355fe088 | 631 | auto_vec<gimple *, 64> worklist; |
ebfd146a IR |
632 | |
633 | /* 1. Init worklist. */ | |
634 | for (i = 0; i < nbbs; i++) | |
635 | { | |
636 | bb = bbs[i]; | |
637 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 638 | { |
ebfd146a | 639 | phi = gsi_stmt (si); |
73fbfcad | 640 | if (dump_enabled_p ()) |
ebfd146a | 641 | { |
78c60e3d SS |
642 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
643 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
644 | } |
645 | ||
646 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 647 | vect_mark_relevant (&worklist, phi, relevant, live_p); |
ebfd146a IR |
648 | } |
649 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
650 | { | |
651 | stmt = gsi_stmt (si); | |
73fbfcad | 652 | if (dump_enabled_p ()) |
ebfd146a | 653 | { |
78c60e3d SS |
654 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
655 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 656 | } |
ebfd146a IR |
657 | |
658 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 659 | vect_mark_relevant (&worklist, stmt, relevant, live_p); |
ebfd146a IR |
660 | } |
661 | } | |
662 | ||
663 | /* 2. Process_worklist */ | |
9771b263 | 664 | while (worklist.length () > 0) |
ebfd146a IR |
665 | { |
666 | use_operand_p use_p; | |
667 | ssa_op_iter iter; | |
668 | ||
9771b263 | 669 | stmt = worklist.pop (); |
73fbfcad | 670 | if (dump_enabled_p ()) |
ebfd146a | 671 | { |
78c60e3d SS |
672 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
673 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
674 | } |
675 | ||
b8698a0f | 676 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
b28ead45 AH |
677 | (DEF_STMT) as relevant/irrelevant according to the relevance property |
678 | of STMT. */ | |
ebfd146a IR |
679 | stmt_vinfo = vinfo_for_stmt (stmt); |
680 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
ebfd146a | 681 | |
b28ead45 AH |
682 | /* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is |
683 | propagated as is to the DEF_STMTs of its USEs. | |
ebfd146a IR |
684 | |
685 | One exception is when STMT has been identified as defining a reduction | |
b28ead45 | 686 | variable; in this case we set the relevance to vect_used_by_reduction. |
ebfd146a | 687 | This is because we distinguish between two kinds of relevant stmts - |
b8698a0f | 688 | those that are used by a reduction computation, and those that are |
ff802fa1 | 689 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 690 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 691 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 692 | |
b28ead45 | 693 | switch (STMT_VINFO_DEF_TYPE (stmt_vinfo)) |
ebfd146a | 694 | { |
06066f92 | 695 | case vect_reduction_def: |
b28ead45 AH |
696 | gcc_assert (relevant != vect_unused_in_scope); |
697 | if (relevant != vect_unused_in_scope | |
698 | && relevant != vect_used_in_scope | |
699 | && relevant != vect_used_by_reduction | |
700 | && relevant != vect_used_only_live) | |
06066f92 | 701 | { |
b28ead45 AH |
702 | if (dump_enabled_p ()) |
703 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
704 | "unsupported use of reduction.\n"); | |
705 | return false; | |
06066f92 | 706 | } |
06066f92 | 707 | break; |
b8698a0f | 708 | |
06066f92 | 709 | case vect_nested_cycle: |
b28ead45 AH |
710 | if (relevant != vect_unused_in_scope |
711 | && relevant != vect_used_in_outer_by_reduction | |
712 | && relevant != vect_used_in_outer) | |
06066f92 | 713 | { |
73fbfcad | 714 | if (dump_enabled_p ()) |
78c60e3d | 715 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 716 | "unsupported use of nested cycle.\n"); |
7c5222ff | 717 | |
06066f92 IR |
718 | return false; |
719 | } | |
b8698a0f L |
720 | break; |
721 | ||
06066f92 | 722 | case vect_double_reduction_def: |
b28ead45 AH |
723 | if (relevant != vect_unused_in_scope |
724 | && relevant != vect_used_by_reduction | |
725 | && relevant != vect_used_only_live) | |
06066f92 | 726 | { |
73fbfcad | 727 | if (dump_enabled_p ()) |
78c60e3d | 728 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 729 | "unsupported use of double reduction.\n"); |
7c5222ff | 730 | |
7c5222ff | 731 | return false; |
06066f92 | 732 | } |
b8698a0f | 733 | break; |
7c5222ff | 734 | |
06066f92 IR |
735 | default: |
736 | break; | |
7c5222ff | 737 | } |
b8698a0f | 738 | |
aec7ae7d | 739 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
740 | { |
741 | /* Pattern statements are not inserted into the code, so | |
742 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
743 | have to scan the RHS or function arguments instead. */ | |
744 | if (is_gimple_assign (stmt)) | |
745 | { | |
69d2aade JJ |
746 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
747 | tree op = gimple_assign_rhs1 (stmt); | |
748 | ||
749 | i = 1; | |
750 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
751 | { | |
752 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
b28ead45 | 753 | relevant, &worklist, false) |
69d2aade | 754 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
b28ead45 | 755 | relevant, &worklist, false)) |
566d377a | 756 | return false; |
69d2aade JJ |
757 | i = 2; |
758 | } | |
759 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 760 | { |
69d2aade | 761 | op = gimple_op (stmt, i); |
afbe6325 | 762 | if (TREE_CODE (op) == SSA_NAME |
b28ead45 | 763 | && !process_use (stmt, op, loop_vinfo, relevant, |
afbe6325 | 764 | &worklist, false)) |
07687835 | 765 | return false; |
9d5e7640 IR |
766 | } |
767 | } | |
768 | else if (is_gimple_call (stmt)) | |
769 | { | |
770 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
771 | { | |
772 | tree arg = gimple_call_arg (stmt, i); | |
b28ead45 | 773 | if (!process_use (stmt, arg, loop_vinfo, relevant, |
aec7ae7d | 774 | &worklist, false)) |
07687835 | 775 | return false; |
9d5e7640 IR |
776 | } |
777 | } | |
778 | } | |
779 | else | |
780 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
781 | { | |
782 | tree op = USE_FROM_PTR (use_p); | |
b28ead45 | 783 | if (!process_use (stmt, op, loop_vinfo, relevant, |
aec7ae7d | 784 | &worklist, false)) |
07687835 | 785 | return false; |
9d5e7640 | 786 | } |
aec7ae7d | 787 | |
3bab6342 | 788 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d | 789 | { |
134c85ca RS |
790 | gather_scatter_info gs_info; |
791 | if (!vect_check_gather_scatter (stmt, loop_vinfo, &gs_info)) | |
792 | gcc_unreachable (); | |
793 | if (!process_use (stmt, gs_info.offset, loop_vinfo, relevant, | |
794 | &worklist, true)) | |
566d377a | 795 | return false; |
aec7ae7d | 796 | } |
ebfd146a IR |
797 | } /* while worklist */ |
798 | ||
ebfd146a IR |
799 | return true; |
800 | } | |
801 | ||
68435eb2 RB |
802 | /* Compute the prologue cost for invariant or constant operands. */ |
803 | ||
804 | static unsigned | |
805 | vect_prologue_cost_for_slp_op (slp_tree node, stmt_vec_info stmt_info, | |
806 | unsigned opno, enum vect_def_type dt, | |
807 | stmt_vector_for_cost *cost_vec) | |
808 | { | |
b9787581 | 809 | gimple *stmt = SLP_TREE_SCALAR_STMTS (node)[0]->stmt; |
68435eb2 RB |
810 | tree op = gimple_op (stmt, opno); |
811 | unsigned prologue_cost = 0; | |
812 | ||
813 | /* Without looking at the actual initializer a vector of | |
814 | constants can be implemented as load from the constant pool. | |
815 | When all elements are the same we can use a splat. */ | |
816 | tree vectype = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
817 | unsigned group_size = SLP_TREE_SCALAR_STMTS (node).length (); | |
818 | unsigned num_vects_to_check; | |
819 | unsigned HOST_WIDE_INT const_nunits; | |
820 | unsigned nelt_limit; | |
821 | if (TYPE_VECTOR_SUBPARTS (vectype).is_constant (&const_nunits) | |
822 | && ! multiple_p (const_nunits, group_size)) | |
823 | { | |
824 | num_vects_to_check = SLP_TREE_NUMBER_OF_VEC_STMTS (node); | |
825 | nelt_limit = const_nunits; | |
826 | } | |
827 | else | |
828 | { | |
829 | /* If either the vector has variable length or the vectors | |
830 | are composed of repeated whole groups we only need to | |
831 | cost construction once. All vectors will be the same. */ | |
832 | num_vects_to_check = 1; | |
833 | nelt_limit = group_size; | |
834 | } | |
835 | tree elt = NULL_TREE; | |
836 | unsigned nelt = 0; | |
837 | for (unsigned j = 0; j < num_vects_to_check * nelt_limit; ++j) | |
838 | { | |
839 | unsigned si = j % group_size; | |
840 | if (nelt == 0) | |
b9787581 | 841 | elt = gimple_op (SLP_TREE_SCALAR_STMTS (node)[si]->stmt, opno); |
68435eb2 RB |
842 | /* ??? We're just tracking whether all operands of a single |
843 | vector initializer are the same, ideally we'd check if | |
844 | we emitted the same one already. */ | |
b9787581 | 845 | else if (elt != gimple_op (SLP_TREE_SCALAR_STMTS (node)[si]->stmt, |
68435eb2 RB |
846 | opno)) |
847 | elt = NULL_TREE; | |
848 | nelt++; | |
849 | if (nelt == nelt_limit) | |
850 | { | |
851 | /* ??? We need to pass down stmt_info for a vector type | |
852 | even if it points to the wrong stmt. */ | |
853 | prologue_cost += record_stmt_cost | |
854 | (cost_vec, 1, | |
855 | dt == vect_external_def | |
856 | ? (elt ? scalar_to_vec : vec_construct) | |
857 | : vector_load, | |
858 | stmt_info, 0, vect_prologue); | |
859 | nelt = 0; | |
860 | } | |
861 | } | |
862 | ||
863 | return prologue_cost; | |
864 | } | |
ebfd146a | 865 | |
b8698a0f | 866 | /* Function vect_model_simple_cost. |
ebfd146a | 867 | |
b8698a0f | 868 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
869 | single op. Right now, this does not account for multiple insns that could |
870 | be generated for the single vector op. We will handle that shortly. */ | |
871 | ||
68435eb2 | 872 | static void |
b8698a0f | 873 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 | 874 | enum vect_def_type *dt, |
4fc5ebf1 | 875 | int ndts, |
68435eb2 RB |
876 | slp_tree node, |
877 | stmt_vector_for_cost *cost_vec) | |
ebfd146a | 878 | { |
92345349 | 879 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a | 880 | |
68435eb2 | 881 | gcc_assert (cost_vec != NULL); |
ebfd146a | 882 | |
68435eb2 RB |
883 | /* ??? Somehow we need to fix this at the callers. */ |
884 | if (node) | |
885 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (node); | |
886 | ||
887 | if (node) | |
888 | { | |
889 | /* Scan operands and account for prologue cost of constants/externals. | |
890 | ??? This over-estimates cost for multiple uses and should be | |
891 | re-engineered. */ | |
b9787581 | 892 | gimple *stmt = SLP_TREE_SCALAR_STMTS (node)[0]->stmt; |
68435eb2 RB |
893 | tree lhs = gimple_get_lhs (stmt); |
894 | for (unsigned i = 0; i < gimple_num_ops (stmt); ++i) | |
895 | { | |
896 | tree op = gimple_op (stmt, i); | |
68435eb2 RB |
897 | enum vect_def_type dt; |
898 | if (!op || op == lhs) | |
899 | continue; | |
894dd753 | 900 | if (vect_is_simple_use (op, stmt_info->vinfo, &dt) |
68435eb2 RB |
901 | && (dt == vect_constant_def || dt == vect_external_def)) |
902 | prologue_cost += vect_prologue_cost_for_slp_op (node, stmt_info, | |
903 | i, dt, cost_vec); | |
904 | } | |
905 | } | |
906 | else | |
907 | /* Cost the "broadcast" of a scalar operand in to a vector operand. | |
908 | Use scalar_to_vec to cost the broadcast, as elsewhere in the vector | |
909 | cost model. */ | |
910 | for (int i = 0; i < ndts; i++) | |
911 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) | |
912 | prologue_cost += record_stmt_cost (cost_vec, 1, scalar_to_vec, | |
913 | stmt_info, 0, vect_prologue); | |
914 | ||
915 | /* Adjust for two-operator SLP nodes. */ | |
916 | if (node && SLP_TREE_TWO_OPERATORS (node)) | |
917 | { | |
918 | ncopies *= 2; | |
919 | inside_cost += record_stmt_cost (cost_vec, ncopies, vec_perm, | |
920 | stmt_info, 0, vect_body); | |
921 | } | |
c3e7ee41 BS |
922 | |
923 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
68435eb2 RB |
924 | inside_cost += record_stmt_cost (cost_vec, ncopies, vector_stmt, |
925 | stmt_info, 0, vect_body); | |
c3e7ee41 | 926 | |
73fbfcad | 927 | if (dump_enabled_p ()) |
78c60e3d SS |
928 | dump_printf_loc (MSG_NOTE, vect_location, |
929 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 930 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
931 | } |
932 | ||
933 | ||
8bd37302 BS |
934 | /* Model cost for type demotion and promotion operations. PWR is normally |
935 | zero for single-step promotions and demotions. It will be one if | |
936 | two-step promotion/demotion is required, and so on. Each additional | |
937 | step doubles the number of instructions required. */ | |
938 | ||
939 | static void | |
940 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
68435eb2 RB |
941 | enum vect_def_type *dt, int pwr, |
942 | stmt_vector_for_cost *cost_vec) | |
8bd37302 BS |
943 | { |
944 | int i, tmp; | |
92345349 | 945 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 | 946 | |
8bd37302 BS |
947 | for (i = 0; i < pwr + 1; i++) |
948 | { | |
949 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
950 | (i + 1) : i; | |
68435eb2 RB |
951 | inside_cost += record_stmt_cost (cost_vec, vect_pow2 (tmp), |
952 | vec_promote_demote, stmt_info, 0, | |
953 | vect_body); | |
8bd37302 BS |
954 | } |
955 | ||
956 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
957 | for (i = 0; i < 2; i++) | |
92345349 | 958 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
68435eb2 RB |
959 | prologue_cost += record_stmt_cost (cost_vec, 1, vector_stmt, |
960 | stmt_info, 0, vect_prologue); | |
8bd37302 | 961 | |
73fbfcad | 962 | if (dump_enabled_p ()) |
78c60e3d SS |
963 | dump_printf_loc (MSG_NOTE, vect_location, |
964 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 965 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
966 | } |
967 | ||
ebfd146a IR |
968 | /* Function vect_model_store_cost |
969 | ||
0d0293ac MM |
970 | Models cost for stores. In the case of grouped accesses, one access |
971 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a | 972 | |
68435eb2 | 973 | static void |
b8698a0f | 974 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
68435eb2 | 975 | enum vect_def_type dt, |
2de001ee | 976 | vect_memory_access_type memory_access_type, |
9ce4345a | 977 | vec_load_store_type vls_type, slp_tree slp_node, |
68435eb2 | 978 | stmt_vector_for_cost *cost_vec) |
ebfd146a | 979 | { |
92345349 | 980 | unsigned int inside_cost = 0, prologue_cost = 0; |
bffb8014 | 981 | stmt_vec_info first_stmt_info = stmt_info; |
892a981f | 982 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
ebfd146a | 983 | |
68435eb2 RB |
984 | /* ??? Somehow we need to fix this at the callers. */ |
985 | if (slp_node) | |
986 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
987 | ||
9ce4345a | 988 | if (vls_type == VLS_STORE_INVARIANT) |
68435eb2 RB |
989 | { |
990 | if (slp_node) | |
991 | prologue_cost += vect_prologue_cost_for_slp_op (slp_node, stmt_info, | |
992 | 1, dt, cost_vec); | |
993 | else | |
994 | prologue_cost += record_stmt_cost (cost_vec, 1, scalar_to_vec, | |
995 | stmt_info, 0, vect_prologue); | |
996 | } | |
ebfd146a | 997 | |
892a981f RS |
998 | /* Grouped stores update all elements in the group at once, |
999 | so we want the DR for the first statement. */ | |
1000 | if (!slp_node && grouped_access_p) | |
bffb8014 | 1001 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a | 1002 | |
892a981f RS |
1003 | /* True if we should include any once-per-group costs as well as |
1004 | the cost of the statement itself. For SLP we only get called | |
1005 | once per group anyhow. */ | |
bffb8014 | 1006 | bool first_stmt_p = (first_stmt_info == stmt_info); |
892a981f | 1007 | |
272c6793 | 1008 | /* We assume that the cost of a single store-lanes instruction is |
2c53b149 | 1009 | equivalent to the cost of DR_GROUP_SIZE separate stores. If a grouped |
272c6793 | 1010 | access is instead being provided by a permute-and-store operation, |
2de001ee RS |
1011 | include the cost of the permutes. */ |
1012 | if (first_stmt_p | |
1013 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 1014 | { |
e1377713 ES |
1015 | /* Uses a high and low interleave or shuffle operations for each |
1016 | needed permute. */ | |
bffb8014 | 1017 | int group_size = DR_GROUP_SIZE (first_stmt_info); |
e1377713 | 1018 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
68435eb2 | 1019 | inside_cost = record_stmt_cost (cost_vec, nstmts, vec_perm, |
92345349 | 1020 | stmt_info, 0, vect_body); |
ebfd146a | 1021 | |
73fbfcad | 1022 | if (dump_enabled_p ()) |
78c60e3d | 1023 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1024 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 1025 | group_size); |
ebfd146a IR |
1026 | } |
1027 | ||
cee62fee | 1028 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 1029 | /* Costs of the stores. */ |
067bc855 RB |
1030 | if (memory_access_type == VMAT_ELEMENTWISE |
1031 | || memory_access_type == VMAT_GATHER_SCATTER) | |
c5126ce8 RS |
1032 | { |
1033 | /* N scalar stores plus extracting the elements. */ | |
1034 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
68435eb2 | 1035 | inside_cost += record_stmt_cost (cost_vec, |
c5126ce8 RS |
1036 | ncopies * assumed_nunits, |
1037 | scalar_store, stmt_info, 0, vect_body); | |
1038 | } | |
f2e2a985 | 1039 | else |
57c454d2 | 1040 | vect_get_store_cost (stmt_info, ncopies, &inside_cost, cost_vec); |
ebfd146a | 1041 | |
2de001ee RS |
1042 | if (memory_access_type == VMAT_ELEMENTWISE |
1043 | || memory_access_type == VMAT_STRIDED_SLP) | |
c5126ce8 RS |
1044 | { |
1045 | /* N scalar stores plus extracting the elements. */ | |
1046 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
68435eb2 | 1047 | inside_cost += record_stmt_cost (cost_vec, |
c5126ce8 RS |
1048 | ncopies * assumed_nunits, |
1049 | vec_to_scalar, stmt_info, 0, vect_body); | |
1050 | } | |
cee62fee | 1051 | |
73fbfcad | 1052 | if (dump_enabled_p ()) |
78c60e3d SS |
1053 | dump_printf_loc (MSG_NOTE, vect_location, |
1054 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 1055 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
1056 | } |
1057 | ||
1058 | ||
720f5239 IR |
1059 | /* Calculate cost of DR's memory access. */ |
1060 | void | |
57c454d2 | 1061 | vect_get_store_cost (stmt_vec_info stmt_info, int ncopies, |
c3e7ee41 | 1062 | unsigned int *inside_cost, |
92345349 | 1063 | stmt_vector_for_cost *body_cost_vec) |
720f5239 | 1064 | { |
57c454d2 | 1065 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
720f5239 IR |
1066 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
1067 | ||
1068 | switch (alignment_support_scheme) | |
1069 | { | |
1070 | case dr_aligned: | |
1071 | { | |
92345349 BS |
1072 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
1073 | vector_store, stmt_info, 0, | |
1074 | vect_body); | |
720f5239 | 1075 | |
73fbfcad | 1076 | if (dump_enabled_p ()) |
78c60e3d | 1077 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1078 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
1079 | break; |
1080 | } | |
1081 | ||
1082 | case dr_unaligned_supported: | |
1083 | { | |
720f5239 | 1084 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1085 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1086 | unaligned_store, stmt_info, |
92345349 | 1087 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1088 | if (dump_enabled_p ()) |
78c60e3d SS |
1089 | dump_printf_loc (MSG_NOTE, vect_location, |
1090 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1091 | "hardware.\n"); |
720f5239 IR |
1092 | break; |
1093 | } | |
1094 | ||
38eec4c6 UW |
1095 | case dr_unaligned_unsupported: |
1096 | { | |
1097 | *inside_cost = VECT_MAX_COST; | |
1098 | ||
73fbfcad | 1099 | if (dump_enabled_p ()) |
78c60e3d | 1100 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1101 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1102 | break; |
1103 | } | |
1104 | ||
720f5239 IR |
1105 | default: |
1106 | gcc_unreachable (); | |
1107 | } | |
1108 | } | |
1109 | ||
1110 | ||
ebfd146a IR |
1111 | /* Function vect_model_load_cost |
1112 | ||
892a981f RS |
1113 | Models cost for loads. In the case of grouped accesses, one access has |
1114 | the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1115 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1116 | access scheme chosen. */ |
1117 | ||
68435eb2 RB |
1118 | static void |
1119 | vect_model_load_cost (stmt_vec_info stmt_info, unsigned ncopies, | |
2de001ee | 1120 | vect_memory_access_type memory_access_type, |
68435eb2 | 1121 | slp_instance instance, |
2de001ee | 1122 | slp_tree slp_node, |
68435eb2 | 1123 | stmt_vector_for_cost *cost_vec) |
ebfd146a | 1124 | { |
92345349 | 1125 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f | 1126 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
ebfd146a | 1127 | |
68435eb2 RB |
1128 | gcc_assert (cost_vec); |
1129 | ||
1130 | /* ??? Somehow we need to fix this at the callers. */ | |
1131 | if (slp_node) | |
1132 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
1133 | ||
1134 | if (slp_node && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
1135 | { | |
1136 | /* If the load is permuted then the alignment is determined by | |
1137 | the first group element not by the first scalar stmt DR. */ | |
bffb8014 | 1138 | stmt_vec_info first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
68435eb2 RB |
1139 | /* Record the cost for the permutation. */ |
1140 | unsigned n_perms; | |
1141 | unsigned assumed_nunits | |
bffb8014 | 1142 | = vect_nunits_for_cost (STMT_VINFO_VECTYPE (first_stmt_info)); |
68435eb2 RB |
1143 | unsigned slp_vf = (ncopies * assumed_nunits) / instance->group_size; |
1144 | vect_transform_slp_perm_load (slp_node, vNULL, NULL, | |
1145 | slp_vf, instance, true, | |
1146 | &n_perms); | |
1147 | inside_cost += record_stmt_cost (cost_vec, n_perms, vec_perm, | |
bffb8014 | 1148 | first_stmt_info, 0, vect_body); |
68435eb2 RB |
1149 | /* And adjust the number of loads performed. This handles |
1150 | redundancies as well as loads that are later dead. */ | |
bffb8014 | 1151 | auto_sbitmap perm (DR_GROUP_SIZE (first_stmt_info)); |
68435eb2 RB |
1152 | bitmap_clear (perm); |
1153 | for (unsigned i = 0; | |
1154 | i < SLP_TREE_LOAD_PERMUTATION (slp_node).length (); ++i) | |
1155 | bitmap_set_bit (perm, SLP_TREE_LOAD_PERMUTATION (slp_node)[i]); | |
1156 | ncopies = 0; | |
1157 | bool load_seen = false; | |
bffb8014 | 1158 | for (unsigned i = 0; i < DR_GROUP_SIZE (first_stmt_info); ++i) |
68435eb2 RB |
1159 | { |
1160 | if (i % assumed_nunits == 0) | |
1161 | { | |
1162 | if (load_seen) | |
1163 | ncopies++; | |
1164 | load_seen = false; | |
1165 | } | |
1166 | if (bitmap_bit_p (perm, i)) | |
1167 | load_seen = true; | |
1168 | } | |
1169 | if (load_seen) | |
1170 | ncopies++; | |
1171 | gcc_assert (ncopies | |
bffb8014 RS |
1172 | <= (DR_GROUP_SIZE (first_stmt_info) |
1173 | - DR_GROUP_GAP (first_stmt_info) | |
68435eb2 RB |
1174 | + assumed_nunits - 1) / assumed_nunits); |
1175 | } | |
1176 | ||
892a981f RS |
1177 | /* Grouped loads read all elements in the group at once, |
1178 | so we want the DR for the first statement. */ | |
bffb8014 | 1179 | stmt_vec_info first_stmt_info = stmt_info; |
892a981f | 1180 | if (!slp_node && grouped_access_p) |
bffb8014 | 1181 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a | 1182 | |
892a981f RS |
1183 | /* True if we should include any once-per-group costs as well as |
1184 | the cost of the statement itself. For SLP we only get called | |
1185 | once per group anyhow. */ | |
bffb8014 | 1186 | bool first_stmt_p = (first_stmt_info == stmt_info); |
892a981f | 1187 | |
272c6793 | 1188 | /* We assume that the cost of a single load-lanes instruction is |
2c53b149 | 1189 | equivalent to the cost of DR_GROUP_SIZE separate loads. If a grouped |
272c6793 | 1190 | access is instead being provided by a load-and-permute operation, |
2de001ee RS |
1191 | include the cost of the permutes. */ |
1192 | if (first_stmt_p | |
1193 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 1194 | { |
2c23db6d ES |
1195 | /* Uses an even and odd extract operations or shuffle operations |
1196 | for each needed permute. */ | |
bffb8014 | 1197 | int group_size = DR_GROUP_SIZE (first_stmt_info); |
2c23db6d | 1198 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
68435eb2 RB |
1199 | inside_cost += record_stmt_cost (cost_vec, nstmts, vec_perm, |
1200 | stmt_info, 0, vect_body); | |
ebfd146a | 1201 | |
73fbfcad | 1202 | if (dump_enabled_p ()) |
e645e942 TJ |
1203 | dump_printf_loc (MSG_NOTE, vect_location, |
1204 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1205 | group_size); |
ebfd146a IR |
1206 | } |
1207 | ||
1208 | /* The loads themselves. */ | |
067bc855 RB |
1209 | if (memory_access_type == VMAT_ELEMENTWISE |
1210 | || memory_access_type == VMAT_GATHER_SCATTER) | |
a82960aa | 1211 | { |
a21892ad BS |
1212 | /* N scalar loads plus gathering them into a vector. */ |
1213 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
c5126ce8 | 1214 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); |
68435eb2 | 1215 | inside_cost += record_stmt_cost (cost_vec, |
c5126ce8 | 1216 | ncopies * assumed_nunits, |
92345349 | 1217 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1218 | } |
1219 | else | |
57c454d2 | 1220 | vect_get_load_cost (stmt_info, ncopies, first_stmt_p, |
92345349 | 1221 | &inside_cost, &prologue_cost, |
68435eb2 | 1222 | cost_vec, cost_vec, true); |
2de001ee RS |
1223 | if (memory_access_type == VMAT_ELEMENTWISE |
1224 | || memory_access_type == VMAT_STRIDED_SLP) | |
68435eb2 | 1225 | inside_cost += record_stmt_cost (cost_vec, ncopies, vec_construct, |
892a981f | 1226 | stmt_info, 0, vect_body); |
720f5239 | 1227 | |
73fbfcad | 1228 | if (dump_enabled_p ()) |
78c60e3d SS |
1229 | dump_printf_loc (MSG_NOTE, vect_location, |
1230 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1231 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1232 | } |
1233 | ||
1234 | ||
1235 | /* Calculate cost of DR's memory access. */ | |
1236 | void | |
57c454d2 | 1237 | vect_get_load_cost (stmt_vec_info stmt_info, int ncopies, |
c3e7ee41 | 1238 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1239 | unsigned int *prologue_cost, |
1240 | stmt_vector_for_cost *prologue_cost_vec, | |
1241 | stmt_vector_for_cost *body_cost_vec, | |
1242 | bool record_prologue_costs) | |
720f5239 | 1243 | { |
57c454d2 | 1244 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
720f5239 IR |
1245 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
1246 | ||
1247 | switch (alignment_support_scheme) | |
ebfd146a IR |
1248 | { |
1249 | case dr_aligned: | |
1250 | { | |
92345349 BS |
1251 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1252 | stmt_info, 0, vect_body); | |
ebfd146a | 1253 | |
73fbfcad | 1254 | if (dump_enabled_p ()) |
78c60e3d | 1255 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1256 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1257 | |
1258 | break; | |
1259 | } | |
1260 | case dr_unaligned_supported: | |
1261 | { | |
720f5239 | 1262 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1263 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1264 | unaligned_load, stmt_info, |
92345349 | 1265 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1266 | |
73fbfcad | 1267 | if (dump_enabled_p ()) |
78c60e3d SS |
1268 | dump_printf_loc (MSG_NOTE, vect_location, |
1269 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1270 | "hardware.\n"); |
ebfd146a IR |
1271 | |
1272 | break; | |
1273 | } | |
1274 | case dr_explicit_realign: | |
1275 | { | |
92345349 BS |
1276 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1277 | vector_load, stmt_info, 0, vect_body); | |
1278 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1279 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1280 | |
1281 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1282 | the containing loop, the following cost should be added to the | |
92345349 | 1283 | prologue costs. */ |
ebfd146a | 1284 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1285 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1286 | stmt_info, 0, vect_body); | |
ebfd146a | 1287 | |
73fbfcad | 1288 | if (dump_enabled_p ()) |
e645e942 TJ |
1289 | dump_printf_loc (MSG_NOTE, vect_location, |
1290 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1291 | |
ebfd146a IR |
1292 | break; |
1293 | } | |
1294 | case dr_explicit_realign_optimized: | |
1295 | { | |
73fbfcad | 1296 | if (dump_enabled_p ()) |
e645e942 | 1297 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1298 | "vect_model_load_cost: unaligned software " |
e645e942 | 1299 | "pipelined.\n"); |
ebfd146a IR |
1300 | |
1301 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1302 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1303 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1304 | access, then the above cost should only be considered for one |
ff802fa1 | 1305 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1306 | and a realignment op. */ |
1307 | ||
92345349 | 1308 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1309 | { |
92345349 BS |
1310 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1311 | vector_stmt, stmt_info, | |
1312 | 0, vect_prologue); | |
ebfd146a | 1313 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1314 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1315 | vector_stmt, stmt_info, | |
1316 | 0, vect_prologue); | |
ebfd146a IR |
1317 | } |
1318 | ||
92345349 BS |
1319 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1320 | stmt_info, 0, vect_body); | |
1321 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1322 | stmt_info, 0, vect_body); | |
8bd37302 | 1323 | |
73fbfcad | 1324 | if (dump_enabled_p ()) |
78c60e3d | 1325 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1326 | "vect_model_load_cost: explicit realign optimized" |
1327 | "\n"); | |
8bd37302 | 1328 | |
ebfd146a IR |
1329 | break; |
1330 | } | |
1331 | ||
38eec4c6 UW |
1332 | case dr_unaligned_unsupported: |
1333 | { | |
1334 | *inside_cost = VECT_MAX_COST; | |
1335 | ||
73fbfcad | 1336 | if (dump_enabled_p ()) |
78c60e3d | 1337 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1338 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1339 | break; |
1340 | } | |
1341 | ||
ebfd146a IR |
1342 | default: |
1343 | gcc_unreachable (); | |
1344 | } | |
ebfd146a IR |
1345 | } |
1346 | ||
418b7df3 RG |
1347 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1348 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1349 | |
418b7df3 | 1350 | static void |
355fe088 | 1351 | vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi) |
ebfd146a | 1352 | { |
ebfd146a | 1353 | if (gsi) |
418b7df3 | 1354 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1355 | else |
1356 | { | |
418b7df3 | 1357 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1358 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1359 | |
a70d6342 IR |
1360 | if (loop_vinfo) |
1361 | { | |
1362 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1363 | basic_block new_bb; |
1364 | edge pe; | |
a70d6342 IR |
1365 | |
1366 | if (nested_in_vect_loop_p (loop, stmt)) | |
1367 | loop = loop->inner; | |
b8698a0f | 1368 | |
a70d6342 | 1369 | pe = loop_preheader_edge (loop); |
418b7df3 | 1370 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1371 | gcc_assert (!new_bb); |
1372 | } | |
1373 | else | |
1374 | { | |
1375 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1376 | basic_block bb; | |
1377 | gimple_stmt_iterator gsi_bb_start; | |
1378 | ||
1379 | gcc_assert (bb_vinfo); | |
1380 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1381 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1382 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1383 | } |
ebfd146a IR |
1384 | } |
1385 | ||
73fbfcad | 1386 | if (dump_enabled_p ()) |
ebfd146a | 1387 | { |
78c60e3d SS |
1388 | dump_printf_loc (MSG_NOTE, vect_location, |
1389 | "created new init_stmt: "); | |
1390 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1391 | } |
418b7df3 RG |
1392 | } |
1393 | ||
1394 | /* Function vect_init_vector. | |
ebfd146a | 1395 | |
5467ee52 RG |
1396 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1397 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1398 | vector type a vector with all elements equal to VAL is created first. | |
1399 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1400 | initialization at the loop preheader. | |
418b7df3 RG |
1401 | Return the DEF of INIT_STMT. |
1402 | It will be used in the vectorization of STMT. */ | |
1403 | ||
1404 | tree | |
355fe088 | 1405 | vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 | 1406 | { |
355fe088 | 1407 | gimple *init_stmt; |
418b7df3 RG |
1408 | tree new_temp; |
1409 | ||
e412ece4 RB |
1410 | /* We abuse this function to push sth to a SSA name with initial 'val'. */ |
1411 | if (! useless_type_conversion_p (type, TREE_TYPE (val))) | |
418b7df3 | 1412 | { |
e412ece4 RB |
1413 | gcc_assert (TREE_CODE (type) == VECTOR_TYPE); |
1414 | if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) | |
418b7df3 | 1415 | { |
5a308cf1 IE |
1416 | /* Scalar boolean value should be transformed into |
1417 | all zeros or all ones value before building a vector. */ | |
1418 | if (VECTOR_BOOLEAN_TYPE_P (type)) | |
1419 | { | |
b3d51f23 IE |
1420 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1421 | tree false_val = build_zero_cst (TREE_TYPE (type)); | |
5a308cf1 IE |
1422 | |
1423 | if (CONSTANT_CLASS_P (val)) | |
1424 | val = integer_zerop (val) ? false_val : true_val; | |
1425 | else | |
1426 | { | |
1427 | new_temp = make_ssa_name (TREE_TYPE (type)); | |
1428 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, | |
1429 | val, true_val, false_val); | |
1430 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
1431 | val = new_temp; | |
1432 | } | |
1433 | } | |
1434 | else if (CONSTANT_CLASS_P (val)) | |
42fd8198 | 1435 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1436 | else |
1437 | { | |
b731b390 | 1438 | new_temp = make_ssa_name (TREE_TYPE (type)); |
e412ece4 RB |
1439 | if (! INTEGRAL_TYPE_P (TREE_TYPE (val))) |
1440 | init_stmt = gimple_build_assign (new_temp, | |
1441 | fold_build1 (VIEW_CONVERT_EXPR, | |
1442 | TREE_TYPE (type), | |
1443 | val)); | |
1444 | else | |
1445 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); | |
418b7df3 | 1446 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1447 | val = new_temp; |
418b7df3 RG |
1448 | } |
1449 | } | |
5467ee52 | 1450 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1451 | } |
1452 | ||
0e22bb5a RB |
1453 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1454 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1455 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1456 | return new_temp; |
ebfd146a IR |
1457 | } |
1458 | ||
c83a894c | 1459 | /* Function vect_get_vec_def_for_operand_1. |
a70d6342 | 1460 | |
c83a894c AH |
1461 | For a defining stmt DEF_STMT of a scalar stmt, return a vector def with type |
1462 | DT that will be used in the vectorized stmt. */ | |
ebfd146a IR |
1463 | |
1464 | tree | |
c83a894c | 1465 | vect_get_vec_def_for_operand_1 (gimple *def_stmt, enum vect_def_type dt) |
ebfd146a IR |
1466 | { |
1467 | tree vec_oprnd; | |
1eede195 | 1468 | stmt_vec_info vec_stmt_info; |
ebfd146a | 1469 | stmt_vec_info def_stmt_info = NULL; |
ebfd146a IR |
1470 | |
1471 | switch (dt) | |
1472 | { | |
81c40241 | 1473 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1474 | case vect_constant_def: |
81c40241 | 1475 | case vect_external_def: |
c83a894c AH |
1476 | /* Code should use vect_get_vec_def_for_operand. */ |
1477 | gcc_unreachable (); | |
ebfd146a | 1478 | |
81c40241 | 1479 | /* operand is defined inside the loop. */ |
8644a673 | 1480 | case vect_internal_def: |
ebfd146a | 1481 | { |
ebfd146a IR |
1482 | /* Get the def from the vectorized stmt. */ |
1483 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1484 | |
1eede195 RS |
1485 | vec_stmt_info = STMT_VINFO_VEC_STMT (def_stmt_info); |
1486 | /* Get vectorized pattern statement. */ | |
1487 | if (!vec_stmt_info | |
1488 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1489 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1490 | vec_stmt_info = (STMT_VINFO_VEC_STMT | |
1491 | (STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
1492 | gcc_assert (vec_stmt_info); | |
1493 | if (gphi *phi = dyn_cast <gphi *> (vec_stmt_info->stmt)) | |
1494 | vec_oprnd = PHI_RESULT (phi); | |
ebfd146a | 1495 | else |
1eede195 RS |
1496 | vec_oprnd = gimple_get_lhs (vec_stmt_info->stmt); |
1497 | return vec_oprnd; | |
ebfd146a IR |
1498 | } |
1499 | ||
c78e3652 | 1500 | /* operand is defined by a loop header phi. */ |
ebfd146a | 1501 | case vect_reduction_def: |
06066f92 | 1502 | case vect_double_reduction_def: |
7c5222ff | 1503 | case vect_nested_cycle: |
ebfd146a IR |
1504 | case vect_induction_def: |
1505 | { | |
1506 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1507 | ||
1eede195 RS |
1508 | /* Get the def from the vectorized stmt. */ |
1509 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1510 | vec_stmt_info = STMT_VINFO_VEC_STMT (def_stmt_info); | |
1511 | if (gphi *phi = dyn_cast <gphi *> (vec_stmt_info->stmt)) | |
1512 | vec_oprnd = PHI_RESULT (phi); | |
6dbbece6 | 1513 | else |
1eede195 RS |
1514 | vec_oprnd = gimple_get_lhs (vec_stmt_info->stmt); |
1515 | return vec_oprnd; | |
ebfd146a IR |
1516 | } |
1517 | ||
1518 | default: | |
1519 | gcc_unreachable (); | |
1520 | } | |
1521 | } | |
1522 | ||
1523 | ||
c83a894c AH |
1524 | /* Function vect_get_vec_def_for_operand. |
1525 | ||
1526 | OP is an operand in STMT. This function returns a (vector) def that will be | |
1527 | used in the vectorized stmt for STMT. | |
1528 | ||
1529 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1530 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1531 | ||
1532 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1533 | needs to be introduced. VECTYPE may be used to specify a required type for | |
1534 | vector invariant. */ | |
1535 | ||
1536 | tree | |
1537 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) | |
1538 | { | |
1539 | gimple *def_stmt; | |
1540 | enum vect_def_type dt; | |
1541 | bool is_simple_use; | |
1542 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
1543 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
1544 | ||
1545 | if (dump_enabled_p ()) | |
1546 | { | |
1547 | dump_printf_loc (MSG_NOTE, vect_location, | |
1548 | "vect_get_vec_def_for_operand: "); | |
1549 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
1550 | dump_printf (MSG_NOTE, "\n"); | |
1551 | } | |
1552 | ||
fef96d8e RS |
1553 | stmt_vec_info def_stmt_info; |
1554 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &dt, | |
1555 | &def_stmt_info, &def_stmt); | |
c83a894c AH |
1556 | gcc_assert (is_simple_use); |
1557 | if (def_stmt && dump_enabled_p ()) | |
1558 | { | |
1559 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1560 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
1561 | } | |
1562 | ||
1563 | if (dt == vect_constant_def || dt == vect_external_def) | |
1564 | { | |
1565 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
1566 | tree vector_type; | |
1567 | ||
1568 | if (vectype) | |
1569 | vector_type = vectype; | |
2568d8a1 | 1570 | else if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op)) |
c83a894c AH |
1571 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) |
1572 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1573 | else | |
1574 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1575 | ||
1576 | gcc_assert (vector_type); | |
1577 | return vect_init_vector (stmt, op, vector_type, NULL); | |
1578 | } | |
1579 | else | |
fef96d8e | 1580 | return vect_get_vec_def_for_operand_1 (def_stmt_info, dt); |
c83a894c AH |
1581 | } |
1582 | ||
1583 | ||
ebfd146a IR |
1584 | /* Function vect_get_vec_def_for_stmt_copy |
1585 | ||
ff802fa1 | 1586 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1587 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1588 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1589 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1590 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1591 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1592 | DT is the type of the vector def VEC_OPRND. |
1593 | ||
1594 | Context: | |
1595 | In case the vectorization factor (VF) is bigger than the number | |
1596 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1597 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1598 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1599 | smallest data-type determines the VF, and as a result, when vectorizing |
1600 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1601 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1602 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1603 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1604 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1605 | ||
1606 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1607 | |
ebfd146a IR |
1608 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1609 | VS1.1: vx.1 = memref1 VS1.2 | |
1610 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1611 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1612 | |
1613 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1614 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1615 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1616 | VSnew.3: vz3 = vx.3 + ... | |
1617 | ||
1618 | The vectorization of S1 is explained in vectorizable_load. | |
1619 | The vectorization of S2: | |
b8698a0f L |
1620 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1621 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1622 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1623 | returns the vector-def 'vx.0'. |
1624 | ||
b8698a0f L |
1625 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1626 | function is called to get the relevant vector-def for each operand. It is | |
1627 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1628 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1629 | ||
b8698a0f L |
1630 | For example, to obtain the vector-def 'vx.1' in order to create the |
1631 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1632 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1633 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1634 | and return its def ('vx.1'). | |
1635 | Overall, to create the above sequence this function will be called 3 times: | |
1636 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1637 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1638 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1639 | ||
1640 | tree | |
1641 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1642 | { | |
355fe088 | 1643 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1644 | stmt_vec_info def_stmt_info; |
1645 | ||
1646 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1647 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1648 | return vec_oprnd; |
1649 | ||
1650 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1651 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1652 | gcc_assert (def_stmt_info); | |
1653 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1654 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1655 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1656 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1657 | else | |
1658 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1659 | return vec_oprnd; | |
1660 | } | |
1661 | ||
1662 | ||
1663 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1664 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a | 1665 | |
c78e3652 | 1666 | void |
b8698a0f | 1667 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1668 | vec<tree> *vec_oprnds0, |
1669 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1670 | { |
9771b263 | 1671 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1672 | |
1673 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1674 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1675 | |
9771b263 | 1676 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1677 | { |
9771b263 | 1678 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1679 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1680 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1681 | } |
1682 | } | |
1683 | ||
1684 | ||
c78e3652 | 1685 | /* Get vectorized definitions for OP0 and OP1. */ |
ebfd146a | 1686 | |
c78e3652 | 1687 | void |
355fe088 | 1688 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1689 | vec<tree> *vec_oprnds0, |
1690 | vec<tree> *vec_oprnds1, | |
306b0c92 | 1691 | slp_tree slp_node) |
ebfd146a IR |
1692 | { |
1693 | if (slp_node) | |
d092494c IR |
1694 | { |
1695 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1696 | auto_vec<tree> ops (nops); |
1697 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1698 | |
9771b263 | 1699 | ops.quick_push (op0); |
d092494c | 1700 | if (op1) |
9771b263 | 1701 | ops.quick_push (op1); |
d092494c | 1702 | |
306b0c92 | 1703 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
d092494c | 1704 | |
37b5ec8f | 1705 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1706 | if (op1) |
37b5ec8f | 1707 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1708 | } |
ebfd146a IR |
1709 | else |
1710 | { | |
1711 | tree vec_oprnd; | |
1712 | ||
9771b263 | 1713 | vec_oprnds0->create (1); |
81c40241 | 1714 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1715 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1716 | |
1717 | if (op1) | |
1718 | { | |
9771b263 | 1719 | vec_oprnds1->create (1); |
81c40241 | 1720 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1721 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1722 | } |
1723 | } | |
1724 | } | |
1725 | ||
bb6c2b68 RS |
1726 | /* Helper function called by vect_finish_replace_stmt and |
1727 | vect_finish_stmt_generation. Set the location of the new | |
e1bd7296 | 1728 | statement and create and return a stmt_vec_info for it. */ |
bb6c2b68 | 1729 | |
e1bd7296 | 1730 | static stmt_vec_info |
bb6c2b68 RS |
1731 | vect_finish_stmt_generation_1 (gimple *stmt, gimple *vec_stmt) |
1732 | { | |
1733 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1734 | vec_info *vinfo = stmt_info->vinfo; | |
1735 | ||
e1bd7296 | 1736 | stmt_vec_info vec_stmt_info = vinfo->add_stmt (vec_stmt); |
bb6c2b68 RS |
1737 | |
1738 | if (dump_enabled_p ()) | |
1739 | { | |
1740 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); | |
1741 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
1742 | } | |
1743 | ||
1744 | gimple_set_location (vec_stmt, gimple_location (stmt)); | |
1745 | ||
1746 | /* While EH edges will generally prevent vectorization, stmt might | |
1747 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1748 | that could throw are part of the same region. */ | |
1749 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1750 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1751 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
e1bd7296 RS |
1752 | |
1753 | return vec_stmt_info; | |
bb6c2b68 RS |
1754 | } |
1755 | ||
1756 | /* Replace the scalar statement STMT with a new vector statement VEC_STMT, | |
e1bd7296 RS |
1757 | which sets the same scalar result as STMT did. Create and return a |
1758 | stmt_vec_info for VEC_STMT. */ | |
bb6c2b68 | 1759 | |
e1bd7296 | 1760 | stmt_vec_info |
bb6c2b68 RS |
1761 | vect_finish_replace_stmt (gimple *stmt, gimple *vec_stmt) |
1762 | { | |
1763 | gcc_assert (gimple_get_lhs (stmt) == gimple_get_lhs (vec_stmt)); | |
1764 | ||
1765 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); | |
1766 | gsi_replace (&gsi, vec_stmt, false); | |
1767 | ||
e1bd7296 | 1768 | return vect_finish_stmt_generation_1 (stmt, vec_stmt); |
bb6c2b68 | 1769 | } |
ebfd146a | 1770 | |
e1bd7296 RS |
1771 | /* Add VEC_STMT to the vectorized implementation of STMT and insert it |
1772 | before *GSI. Create and return a stmt_vec_info for VEC_STMT. */ | |
ebfd146a | 1773 | |
e1bd7296 | 1774 | stmt_vec_info |
355fe088 | 1775 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1776 | gimple_stmt_iterator *gsi) |
1777 | { | |
ebfd146a IR |
1778 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); |
1779 | ||
54e8e2c3 RG |
1780 | if (!gsi_end_p (*gsi) |
1781 | && gimple_has_mem_ops (vec_stmt)) | |
1782 | { | |
355fe088 | 1783 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1784 | tree vuse = gimple_vuse (at_stmt); |
1785 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1786 | { | |
1787 | tree vdef = gimple_vdef (at_stmt); | |
1788 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1789 | /* If we have an SSA vuse and insert a store, update virtual | |
1790 | SSA form to avoid triggering the renamer. Do so only | |
1791 | if we can easily see all uses - which is what almost always | |
1792 | happens with the way vectorized stmts are inserted. */ | |
1793 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1794 | && ((is_gimple_assign (vec_stmt) | |
1795 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1796 | || (is_gimple_call (vec_stmt) | |
1797 | && !(gimple_call_flags (vec_stmt) | |
1798 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1799 | { | |
1800 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1801 | gimple_set_vdef (vec_stmt, new_vdef); | |
1802 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1803 | } | |
1804 | } | |
1805 | } | |
ebfd146a | 1806 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
e1bd7296 | 1807 | return vect_finish_stmt_generation_1 (stmt, vec_stmt); |
ebfd146a IR |
1808 | } |
1809 | ||
70439f0d RS |
1810 | /* We want to vectorize a call to combined function CFN with function |
1811 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN | |
1812 | as the types of all inputs. Check whether this is possible using | |
1813 | an internal function, returning its code if so or IFN_LAST if not. */ | |
ebfd146a | 1814 | |
70439f0d RS |
1815 | static internal_fn |
1816 | vectorizable_internal_function (combined_fn cfn, tree fndecl, | |
1817 | tree vectype_out, tree vectype_in) | |
ebfd146a | 1818 | { |
70439f0d RS |
1819 | internal_fn ifn; |
1820 | if (internal_fn_p (cfn)) | |
1821 | ifn = as_internal_fn (cfn); | |
1822 | else | |
1823 | ifn = associated_internal_fn (fndecl); | |
1824 | if (ifn != IFN_LAST && direct_internal_fn_p (ifn)) | |
1825 | { | |
1826 | const direct_internal_fn_info &info = direct_internal_fn (ifn); | |
1827 | if (info.vectorizable) | |
1828 | { | |
1829 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); | |
1830 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); | |
d95ab70a RS |
1831 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1832 | OPTIMIZE_FOR_SPEED)) | |
70439f0d RS |
1833 | return ifn; |
1834 | } | |
1835 | } | |
1836 | return IFN_LAST; | |
ebfd146a IR |
1837 | } |
1838 | ||
5ce9450f | 1839 | |
355fe088 | 1840 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1841 | gimple_stmt_iterator *); |
1842 | ||
7cfb4d93 RS |
1843 | /* Check whether a load or store statement in the loop described by |
1844 | LOOP_VINFO is possible in a fully-masked loop. This is testing | |
1845 | whether the vectorizer pass has the appropriate support, as well as | |
1846 | whether the target does. | |
1847 | ||
1848 | VLS_TYPE says whether the statement is a load or store and VECTYPE | |
1849 | is the type of the vector being loaded or stored. MEMORY_ACCESS_TYPE | |
1850 | says how the load or store is going to be implemented and GROUP_SIZE | |
1851 | is the number of load or store statements in the containing group. | |
bfaa08b7 RS |
1852 | If the access is a gather load or scatter store, GS_INFO describes |
1853 | its arguments. | |
7cfb4d93 RS |
1854 | |
1855 | Clear LOOP_VINFO_CAN_FULLY_MASK_P if a fully-masked loop is not | |
1856 | supported, otherwise record the required mask types. */ | |
1857 | ||
1858 | static void | |
1859 | check_load_store_masking (loop_vec_info loop_vinfo, tree vectype, | |
1860 | vec_load_store_type vls_type, int group_size, | |
bfaa08b7 RS |
1861 | vect_memory_access_type memory_access_type, |
1862 | gather_scatter_info *gs_info) | |
7cfb4d93 RS |
1863 | { |
1864 | /* Invariant loads need no special support. */ | |
1865 | if (memory_access_type == VMAT_INVARIANT) | |
1866 | return; | |
1867 | ||
1868 | vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); | |
1869 | machine_mode vecmode = TYPE_MODE (vectype); | |
1870 | bool is_load = (vls_type == VLS_LOAD); | |
1871 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
1872 | { | |
1873 | if (is_load | |
1874 | ? !vect_load_lanes_supported (vectype, group_size, true) | |
1875 | : !vect_store_lanes_supported (vectype, group_size, true)) | |
1876 | { | |
1877 | if (dump_enabled_p ()) | |
1878 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1879 | "can't use a fully-masked loop because the" | |
1880 | " target doesn't have an appropriate masked" | |
1881 | " load/store-lanes instruction.\n"); | |
1882 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1883 | return; | |
1884 | } | |
1885 | unsigned int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
1886 | vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype); | |
1887 | return; | |
1888 | } | |
1889 | ||
bfaa08b7 RS |
1890 | if (memory_access_type == VMAT_GATHER_SCATTER) |
1891 | { | |
f307441a RS |
1892 | internal_fn ifn = (is_load |
1893 | ? IFN_MASK_GATHER_LOAD | |
1894 | : IFN_MASK_SCATTER_STORE); | |
bfaa08b7 | 1895 | tree offset_type = TREE_TYPE (gs_info->offset); |
f307441a | 1896 | if (!internal_gather_scatter_fn_supported_p (ifn, vectype, |
bfaa08b7 RS |
1897 | gs_info->memory_type, |
1898 | TYPE_SIGN (offset_type), | |
1899 | gs_info->scale)) | |
1900 | { | |
1901 | if (dump_enabled_p ()) | |
1902 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1903 | "can't use a fully-masked loop because the" | |
1904 | " target doesn't have an appropriate masked" | |
f307441a | 1905 | " gather load or scatter store instruction.\n"); |
bfaa08b7 RS |
1906 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; |
1907 | return; | |
1908 | } | |
1909 | unsigned int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
1910 | vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype); | |
1911 | return; | |
1912 | } | |
1913 | ||
7cfb4d93 RS |
1914 | if (memory_access_type != VMAT_CONTIGUOUS |
1915 | && memory_access_type != VMAT_CONTIGUOUS_PERMUTE) | |
1916 | { | |
1917 | /* Element X of the data must come from iteration i * VF + X of the | |
1918 | scalar loop. We need more work to support other mappings. */ | |
1919 | if (dump_enabled_p ()) | |
1920 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1921 | "can't use a fully-masked loop because an access" | |
1922 | " isn't contiguous.\n"); | |
1923 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1924 | return; | |
1925 | } | |
1926 | ||
1927 | machine_mode mask_mode; | |
1928 | if (!(targetm.vectorize.get_mask_mode | |
1929 | (GET_MODE_NUNITS (vecmode), | |
1930 | GET_MODE_SIZE (vecmode)).exists (&mask_mode)) | |
1931 | || !can_vec_mask_load_store_p (vecmode, mask_mode, is_load)) | |
1932 | { | |
1933 | if (dump_enabled_p ()) | |
1934 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1935 | "can't use a fully-masked loop because the target" | |
1936 | " doesn't have the appropriate masked load or" | |
1937 | " store.\n"); | |
1938 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1939 | return; | |
1940 | } | |
1941 | /* We might load more scalars than we need for permuting SLP loads. | |
1942 | We checked in get_group_load_store_type that the extra elements | |
1943 | don't leak into a new vector. */ | |
1944 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
1945 | poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
1946 | unsigned int nvectors; | |
1947 | if (can_div_away_from_zero_p (group_size * vf, nunits, &nvectors)) | |
1948 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype); | |
1949 | else | |
1950 | gcc_unreachable (); | |
1951 | } | |
1952 | ||
1953 | /* Return the mask input to a masked load or store. VEC_MASK is the vectorized | |
1954 | form of the scalar mask condition and LOOP_MASK, if nonnull, is the mask | |
1955 | that needs to be applied to all loads and stores in a vectorized loop. | |
1956 | Return VEC_MASK if LOOP_MASK is null, otherwise return VEC_MASK & LOOP_MASK. | |
1957 | ||
1958 | MASK_TYPE is the type of both masks. If new statements are needed, | |
1959 | insert them before GSI. */ | |
1960 | ||
1961 | static tree | |
1962 | prepare_load_store_mask (tree mask_type, tree loop_mask, tree vec_mask, | |
1963 | gimple_stmt_iterator *gsi) | |
1964 | { | |
1965 | gcc_assert (useless_type_conversion_p (mask_type, TREE_TYPE (vec_mask))); | |
1966 | if (!loop_mask) | |
1967 | return vec_mask; | |
1968 | ||
1969 | gcc_assert (TREE_TYPE (loop_mask) == mask_type); | |
1970 | tree and_res = make_temp_ssa_name (mask_type, NULL, "vec_mask_and"); | |
1971 | gimple *and_stmt = gimple_build_assign (and_res, BIT_AND_EXPR, | |
1972 | vec_mask, loop_mask); | |
1973 | gsi_insert_before (gsi, and_stmt, GSI_SAME_STMT); | |
1974 | return and_res; | |
1975 | } | |
1976 | ||
429ef523 RS |
1977 | /* Determine whether we can use a gather load or scatter store to vectorize |
1978 | strided load or store STMT by truncating the current offset to a smaller | |
1979 | width. We need to be able to construct an offset vector: | |
1980 | ||
1981 | { 0, X, X*2, X*3, ... } | |
1982 | ||
1983 | without loss of precision, where X is STMT's DR_STEP. | |
1984 | ||
1985 | Return true if this is possible, describing the gather load or scatter | |
1986 | store in GS_INFO. MASKED_P is true if the load or store is conditional. */ | |
1987 | ||
1988 | static bool | |
1989 | vect_truncate_gather_scatter_offset (gimple *stmt, loop_vec_info loop_vinfo, | |
1990 | bool masked_p, | |
1991 | gather_scatter_info *gs_info) | |
1992 | { | |
1993 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1994 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
1995 | tree step = DR_STEP (dr); | |
1996 | if (TREE_CODE (step) != INTEGER_CST) | |
1997 | { | |
1998 | /* ??? Perhaps we could use range information here? */ | |
1999 | if (dump_enabled_p ()) | |
2000 | dump_printf_loc (MSG_NOTE, vect_location, | |
2001 | "cannot truncate variable step.\n"); | |
2002 | return false; | |
2003 | } | |
2004 | ||
2005 | /* Get the number of bits in an element. */ | |
2006 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2007 | scalar_mode element_mode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); | |
2008 | unsigned int element_bits = GET_MODE_BITSIZE (element_mode); | |
2009 | ||
2010 | /* Set COUNT to the upper limit on the number of elements - 1. | |
2011 | Start with the maximum vectorization factor. */ | |
2012 | unsigned HOST_WIDE_INT count = vect_max_vf (loop_vinfo) - 1; | |
2013 | ||
2014 | /* Try lowering COUNT to the number of scalar latch iterations. */ | |
2015 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2016 | widest_int max_iters; | |
2017 | if (max_loop_iterations (loop, &max_iters) | |
2018 | && max_iters < count) | |
2019 | count = max_iters.to_shwi (); | |
2020 | ||
2021 | /* Try scales of 1 and the element size. */ | |
2022 | int scales[] = { 1, vect_get_scalar_dr_size (dr) }; | |
4a669ac3 | 2023 | wi::overflow_type overflow = wi::OVF_NONE; |
429ef523 RS |
2024 | for (int i = 0; i < 2; ++i) |
2025 | { | |
2026 | int scale = scales[i]; | |
2027 | widest_int factor; | |
2028 | if (!wi::multiple_of_p (wi::to_widest (step), scale, SIGNED, &factor)) | |
2029 | continue; | |
2030 | ||
2031 | /* See whether we can calculate (COUNT - 1) * STEP / SCALE | |
2032 | in OFFSET_BITS bits. */ | |
4a669ac3 AH |
2033 | widest_int range = wi::mul (count, factor, SIGNED, &overflow); |
2034 | if (overflow) | |
429ef523 RS |
2035 | continue; |
2036 | signop sign = range >= 0 ? UNSIGNED : SIGNED; | |
2037 | if (wi::min_precision (range, sign) > element_bits) | |
2038 | { | |
4a669ac3 | 2039 | overflow = wi::OVF_UNKNOWN; |
429ef523 RS |
2040 | continue; |
2041 | } | |
2042 | ||
2043 | /* See whether the target supports the operation. */ | |
2044 | tree memory_type = TREE_TYPE (DR_REF (dr)); | |
2045 | if (!vect_gather_scatter_fn_p (DR_IS_READ (dr), masked_p, vectype, | |
2046 | memory_type, element_bits, sign, scale, | |
2047 | &gs_info->ifn, &gs_info->element_type)) | |
2048 | continue; | |
2049 | ||
2050 | tree offset_type = build_nonstandard_integer_type (element_bits, | |
2051 | sign == UNSIGNED); | |
2052 | ||
2053 | gs_info->decl = NULL_TREE; | |
2054 | /* Logically the sum of DR_BASE_ADDRESS, DR_INIT and DR_OFFSET, | |
2055 | but we don't need to store that here. */ | |
2056 | gs_info->base = NULL_TREE; | |
2057 | gs_info->offset = fold_convert (offset_type, step); | |
929b4411 | 2058 | gs_info->offset_dt = vect_constant_def; |
429ef523 RS |
2059 | gs_info->offset_vectype = NULL_TREE; |
2060 | gs_info->scale = scale; | |
2061 | gs_info->memory_type = memory_type; | |
2062 | return true; | |
2063 | } | |
2064 | ||
4a669ac3 | 2065 | if (overflow && dump_enabled_p ()) |
429ef523 RS |
2066 | dump_printf_loc (MSG_NOTE, vect_location, |
2067 | "truncating gather/scatter offset to %d bits" | |
2068 | " might change its value.\n", element_bits); | |
2069 | ||
2070 | return false; | |
2071 | } | |
2072 | ||
ab2fc782 RS |
2073 | /* Return true if we can use gather/scatter internal functions to |
2074 | vectorize STMT, which is a grouped or strided load or store. | |
429ef523 RS |
2075 | MASKED_P is true if load or store is conditional. When returning |
2076 | true, fill in GS_INFO with the information required to perform the | |
2077 | operation. */ | |
ab2fc782 RS |
2078 | |
2079 | static bool | |
2080 | vect_use_strided_gather_scatters_p (gimple *stmt, loop_vec_info loop_vinfo, | |
429ef523 | 2081 | bool masked_p, |
ab2fc782 RS |
2082 | gather_scatter_info *gs_info) |
2083 | { | |
2084 | if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info) | |
2085 | || gs_info->decl) | |
429ef523 RS |
2086 | return vect_truncate_gather_scatter_offset (stmt, loop_vinfo, |
2087 | masked_p, gs_info); | |
ab2fc782 RS |
2088 | |
2089 | scalar_mode element_mode = SCALAR_TYPE_MODE (gs_info->element_type); | |
2090 | unsigned int element_bits = GET_MODE_BITSIZE (element_mode); | |
2091 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2092 | unsigned int offset_bits = TYPE_PRECISION (offset_type); | |
2093 | ||
2094 | /* Enforced by vect_check_gather_scatter. */ | |
2095 | gcc_assert (element_bits >= offset_bits); | |
2096 | ||
2097 | /* If the elements are wider than the offset, convert the offset to the | |
2098 | same width, without changing its sign. */ | |
2099 | if (element_bits > offset_bits) | |
2100 | { | |
2101 | bool unsigned_p = TYPE_UNSIGNED (offset_type); | |
2102 | offset_type = build_nonstandard_integer_type (element_bits, unsigned_p); | |
2103 | gs_info->offset = fold_convert (offset_type, gs_info->offset); | |
2104 | } | |
2105 | ||
2106 | if (dump_enabled_p ()) | |
2107 | dump_printf_loc (MSG_NOTE, vect_location, | |
2108 | "using gather/scatter for strided/grouped access," | |
2109 | " scale = %d\n", gs_info->scale); | |
2110 | ||
2111 | return true; | |
2112 | } | |
2113 | ||
62da9e14 RS |
2114 | /* STMT is a non-strided load or store, meaning that it accesses |
2115 | elements with a known constant step. Return -1 if that step | |
2116 | is negative, 0 if it is zero, and 1 if it is greater than zero. */ | |
2117 | ||
2118 | static int | |
2119 | compare_step_with_zero (gimple *stmt) | |
2120 | { | |
2121 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3f5e8a76 RS |
2122 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
2123 | return tree_int_cst_compare (vect_dr_behavior (dr)->step, | |
2124 | size_zero_node); | |
62da9e14 RS |
2125 | } |
2126 | ||
2127 | /* If the target supports a permute mask that reverses the elements in | |
2128 | a vector of type VECTYPE, return that mask, otherwise return null. */ | |
2129 | ||
2130 | static tree | |
2131 | perm_mask_for_reverse (tree vectype) | |
2132 | { | |
928686b1 | 2133 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
62da9e14 | 2134 | |
d980067b RS |
2135 | /* The encoding has a single stepped pattern. */ |
2136 | vec_perm_builder sel (nunits, 1, 3); | |
928686b1 | 2137 | for (int i = 0; i < 3; ++i) |
908a1a16 | 2138 | sel.quick_push (nunits - 1 - i); |
62da9e14 | 2139 | |
e3342de4 RS |
2140 | vec_perm_indices indices (sel, 1, nunits); |
2141 | if (!can_vec_perm_const_p (TYPE_MODE (vectype), indices)) | |
62da9e14 | 2142 | return NULL_TREE; |
e3342de4 | 2143 | return vect_gen_perm_mask_checked (vectype, indices); |
62da9e14 | 2144 | } |
5ce9450f | 2145 | |
c3a8f964 RS |
2146 | /* STMT is either a masked or unconditional store. Return the value |
2147 | being stored. */ | |
2148 | ||
f307441a | 2149 | tree |
c3a8f964 RS |
2150 | vect_get_store_rhs (gimple *stmt) |
2151 | { | |
2152 | if (gassign *assign = dyn_cast <gassign *> (stmt)) | |
2153 | { | |
2154 | gcc_assert (gimple_assign_single_p (assign)); | |
2155 | return gimple_assign_rhs1 (assign); | |
2156 | } | |
2157 | if (gcall *call = dyn_cast <gcall *> (stmt)) | |
2158 | { | |
2159 | internal_fn ifn = gimple_call_internal_fn (call); | |
f307441a RS |
2160 | int index = internal_fn_stored_value_index (ifn); |
2161 | gcc_assert (index >= 0); | |
2162 | return gimple_call_arg (stmt, index); | |
c3a8f964 RS |
2163 | } |
2164 | gcc_unreachable (); | |
2165 | } | |
2166 | ||
2de001ee RS |
2167 | /* A subroutine of get_load_store_type, with a subset of the same |
2168 | arguments. Handle the case where STMT is part of a grouped load | |
2169 | or store. | |
2170 | ||
2171 | For stores, the statements in the group are all consecutive | |
2172 | and there is no gap at the end. For loads, the statements in the | |
2173 | group might not be consecutive; there can be gaps between statements | |
2174 | as well as at the end. */ | |
2175 | ||
2176 | static bool | |
2177 | get_group_load_store_type (gimple *stmt, tree vectype, bool slp, | |
7e11fc7f | 2178 | bool masked_p, vec_load_store_type vls_type, |
429ef523 RS |
2179 | vect_memory_access_type *memory_access_type, |
2180 | gather_scatter_info *gs_info) | |
2de001ee RS |
2181 | { |
2182 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2183 | vec_info *vinfo = stmt_info->vinfo; | |
2184 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2185 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; | |
bffb8014 RS |
2186 | stmt_vec_info first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
2187 | data_reference *first_dr = STMT_VINFO_DATA_REF (first_stmt_info); | |
2188 | unsigned int group_size = DR_GROUP_SIZE (first_stmt_info); | |
2189 | bool single_element_p = (stmt_info == first_stmt_info | |
2c53b149 | 2190 | && !DR_GROUP_NEXT_ELEMENT (stmt_info)); |
bffb8014 | 2191 | unsigned HOST_WIDE_INT gap = DR_GROUP_GAP (first_stmt_info); |
928686b1 | 2192 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
2193 | |
2194 | /* True if the vectorized statements would access beyond the last | |
2195 | statement in the group. */ | |
2196 | bool overrun_p = false; | |
2197 | ||
2198 | /* True if we can cope with such overrun by peeling for gaps, so that | |
2199 | there is at least one final scalar iteration after the vector loop. */ | |
7e11fc7f RS |
2200 | bool can_overrun_p = (!masked_p |
2201 | && vls_type == VLS_LOAD | |
2202 | && loop_vinfo | |
2203 | && !loop->inner); | |
2de001ee RS |
2204 | |
2205 | /* There can only be a gap at the end of the group if the stride is | |
2206 | known at compile time. */ | |
2207 | gcc_assert (!STMT_VINFO_STRIDED_P (stmt_info) || gap == 0); | |
2208 | ||
2209 | /* Stores can't yet have gaps. */ | |
2210 | gcc_assert (slp || vls_type == VLS_LOAD || gap == 0); | |
2211 | ||
2212 | if (slp) | |
2213 | { | |
2214 | if (STMT_VINFO_STRIDED_P (stmt_info)) | |
2215 | { | |
2c53b149 | 2216 | /* Try to use consecutive accesses of DR_GROUP_SIZE elements, |
2de001ee RS |
2217 | separated by the stride, until we have a complete vector. |
2218 | Fall back to scalar accesses if that isn't possible. */ | |
928686b1 | 2219 | if (multiple_p (nunits, group_size)) |
2de001ee RS |
2220 | *memory_access_type = VMAT_STRIDED_SLP; |
2221 | else | |
2222 | *memory_access_type = VMAT_ELEMENTWISE; | |
2223 | } | |
2224 | else | |
2225 | { | |
2226 | overrun_p = loop_vinfo && gap != 0; | |
2227 | if (overrun_p && vls_type != VLS_LOAD) | |
2228 | { | |
2229 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2230 | "Grouped store with gaps requires" | |
2231 | " non-consecutive accesses\n"); | |
2232 | return false; | |
2233 | } | |
f702e7d4 RS |
2234 | /* An overrun is fine if the trailing elements are smaller |
2235 | than the alignment boundary B. Every vector access will | |
2236 | be a multiple of B and so we are guaranteed to access a | |
2237 | non-gap element in the same B-sized block. */ | |
f9ef2c76 | 2238 | if (overrun_p |
f702e7d4 RS |
2239 | && gap < (vect_known_alignment_in_bytes (first_dr) |
2240 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 2241 | overrun_p = false; |
2de001ee RS |
2242 | if (overrun_p && !can_overrun_p) |
2243 | { | |
2244 | if (dump_enabled_p ()) | |
2245 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2246 | "Peeling for outer loop is not supported\n"); | |
2247 | return false; | |
2248 | } | |
2249 | *memory_access_type = VMAT_CONTIGUOUS; | |
2250 | } | |
2251 | } | |
2252 | else | |
2253 | { | |
2254 | /* We can always handle this case using elementwise accesses, | |
2255 | but see if something more efficient is available. */ | |
2256 | *memory_access_type = VMAT_ELEMENTWISE; | |
2257 | ||
2258 | /* If there is a gap at the end of the group then these optimizations | |
2259 | would access excess elements in the last iteration. */ | |
2260 | bool would_overrun_p = (gap != 0); | |
f702e7d4 RS |
2261 | /* An overrun is fine if the trailing elements are smaller than the |
2262 | alignment boundary B. Every vector access will be a multiple of B | |
2263 | and so we are guaranteed to access a non-gap element in the | |
2264 | same B-sized block. */ | |
f9ef2c76 | 2265 | if (would_overrun_p |
7e11fc7f | 2266 | && !masked_p |
f702e7d4 RS |
2267 | && gap < (vect_known_alignment_in_bytes (first_dr) |
2268 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 2269 | would_overrun_p = false; |
f702e7d4 | 2270 | |
2de001ee | 2271 | if (!STMT_VINFO_STRIDED_P (stmt_info) |
62da9e14 RS |
2272 | && (can_overrun_p || !would_overrun_p) |
2273 | && compare_step_with_zero (stmt) > 0) | |
2de001ee | 2274 | { |
6737facb RS |
2275 | /* First cope with the degenerate case of a single-element |
2276 | vector. */ | |
2277 | if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U)) | |
2278 | *memory_access_type = VMAT_CONTIGUOUS; | |
2279 | ||
2280 | /* Otherwise try using LOAD/STORE_LANES. */ | |
2281 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2282 | && (vls_type == VLS_LOAD | |
7e11fc7f RS |
2283 | ? vect_load_lanes_supported (vectype, group_size, masked_p) |
2284 | : vect_store_lanes_supported (vectype, group_size, | |
2285 | masked_p))) | |
2de001ee RS |
2286 | { |
2287 | *memory_access_type = VMAT_LOAD_STORE_LANES; | |
2288 | overrun_p = would_overrun_p; | |
2289 | } | |
2290 | ||
2291 | /* If that fails, try using permuting loads. */ | |
2292 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2293 | && (vls_type == VLS_LOAD | |
2294 | ? vect_grouped_load_supported (vectype, single_element_p, | |
2295 | group_size) | |
2296 | : vect_grouped_store_supported (vectype, group_size))) | |
2297 | { | |
2298 | *memory_access_type = VMAT_CONTIGUOUS_PERMUTE; | |
2299 | overrun_p = would_overrun_p; | |
2300 | } | |
2301 | } | |
429ef523 RS |
2302 | |
2303 | /* As a last resort, trying using a gather load or scatter store. | |
2304 | ||
2305 | ??? Although the code can handle all group sizes correctly, | |
2306 | it probably isn't a win to use separate strided accesses based | |
2307 | on nearby locations. Or, even if it's a win over scalar code, | |
2308 | it might not be a win over vectorizing at a lower VF, if that | |
2309 | allows us to use contiguous accesses. */ | |
2310 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2311 | && single_element_p | |
2312 | && loop_vinfo | |
2313 | && vect_use_strided_gather_scatters_p (stmt, loop_vinfo, | |
2314 | masked_p, gs_info)) | |
2315 | *memory_access_type = VMAT_GATHER_SCATTER; | |
2de001ee RS |
2316 | } |
2317 | ||
bffb8014 | 2318 | if (vls_type != VLS_LOAD && first_stmt_info == stmt_info) |
2de001ee RS |
2319 | { |
2320 | /* STMT is the leader of the group. Check the operands of all the | |
2321 | stmts of the group. */ | |
bffb8014 RS |
2322 | stmt_vec_info next_stmt_info = DR_GROUP_NEXT_ELEMENT (stmt_info); |
2323 | while (next_stmt_info) | |
2de001ee | 2324 | { |
bffb8014 | 2325 | tree op = vect_get_store_rhs (next_stmt_info); |
2de001ee | 2326 | enum vect_def_type dt; |
894dd753 | 2327 | if (!vect_is_simple_use (op, vinfo, &dt)) |
2de001ee RS |
2328 | { |
2329 | if (dump_enabled_p ()) | |
2330 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2331 | "use not simple.\n"); | |
2332 | return false; | |
2333 | } | |
bffb8014 | 2334 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
2de001ee RS |
2335 | } |
2336 | } | |
2337 | ||
2338 | if (overrun_p) | |
2339 | { | |
2340 | gcc_assert (can_overrun_p); | |
2341 | if (dump_enabled_p ()) | |
2342 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2343 | "Data access with gaps requires scalar " | |
2344 | "epilogue loop\n"); | |
2345 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; | |
2346 | } | |
2347 | ||
2348 | return true; | |
2349 | } | |
2350 | ||
62da9e14 RS |
2351 | /* A subroutine of get_load_store_type, with a subset of the same |
2352 | arguments. Handle the case where STMT is a load or store that | |
2353 | accesses consecutive elements with a negative step. */ | |
2354 | ||
2355 | static vect_memory_access_type | |
2356 | get_negative_load_store_type (gimple *stmt, tree vectype, | |
2357 | vec_load_store_type vls_type, | |
2358 | unsigned int ncopies) | |
2359 | { | |
2360 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2361 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2362 | dr_alignment_support alignment_support_scheme; | |
2363 | ||
2364 | if (ncopies > 1) | |
2365 | { | |
2366 | if (dump_enabled_p ()) | |
2367 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2368 | "multiple types with negative step.\n"); | |
2369 | return VMAT_ELEMENTWISE; | |
2370 | } | |
2371 | ||
2372 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
2373 | if (alignment_support_scheme != dr_aligned | |
2374 | && alignment_support_scheme != dr_unaligned_supported) | |
2375 | { | |
2376 | if (dump_enabled_p ()) | |
2377 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2378 | "negative step but alignment required.\n"); | |
2379 | return VMAT_ELEMENTWISE; | |
2380 | } | |
2381 | ||
2382 | if (vls_type == VLS_STORE_INVARIANT) | |
2383 | { | |
2384 | if (dump_enabled_p ()) | |
2385 | dump_printf_loc (MSG_NOTE, vect_location, | |
2386 | "negative step with invariant source;" | |
2387 | " no permute needed.\n"); | |
2388 | return VMAT_CONTIGUOUS_DOWN; | |
2389 | } | |
2390 | ||
2391 | if (!perm_mask_for_reverse (vectype)) | |
2392 | { | |
2393 | if (dump_enabled_p ()) | |
2394 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2395 | "negative step and reversing not supported.\n"); | |
2396 | return VMAT_ELEMENTWISE; | |
2397 | } | |
2398 | ||
2399 | return VMAT_CONTIGUOUS_REVERSE; | |
2400 | } | |
2401 | ||
2de001ee RS |
2402 | /* Analyze load or store statement STMT of type VLS_TYPE. Return true |
2403 | if there is a memory access type that the vectorized form can use, | |
2404 | storing it in *MEMORY_ACCESS_TYPE if so. If we decide to use gathers | |
2405 | or scatters, fill in GS_INFO accordingly. | |
2406 | ||
2407 | SLP says whether we're performing SLP rather than loop vectorization. | |
7e11fc7f | 2408 | MASKED_P is true if the statement is conditional on a vectorized mask. |
62da9e14 RS |
2409 | VECTYPE is the vector type that the vectorized statements will use. |
2410 | NCOPIES is the number of vector statements that will be needed. */ | |
2de001ee RS |
2411 | |
2412 | static bool | |
7e11fc7f | 2413 | get_load_store_type (gimple *stmt, tree vectype, bool slp, bool masked_p, |
62da9e14 | 2414 | vec_load_store_type vls_type, unsigned int ncopies, |
2de001ee RS |
2415 | vect_memory_access_type *memory_access_type, |
2416 | gather_scatter_info *gs_info) | |
2417 | { | |
2418 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2419 | vec_info *vinfo = stmt_info->vinfo; | |
2420 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4d694b27 | 2421 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
2422 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
2423 | { | |
2424 | *memory_access_type = VMAT_GATHER_SCATTER; | |
2de001ee RS |
2425 | if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info)) |
2426 | gcc_unreachable (); | |
894dd753 | 2427 | else if (!vect_is_simple_use (gs_info->offset, vinfo, |
2de001ee RS |
2428 | &gs_info->offset_dt, |
2429 | &gs_info->offset_vectype)) | |
2430 | { | |
2431 | if (dump_enabled_p ()) | |
2432 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2433 | "%s index use not simple.\n", | |
2434 | vls_type == VLS_LOAD ? "gather" : "scatter"); | |
2435 | return false; | |
2436 | } | |
2437 | } | |
2438 | else if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
2439 | { | |
7e11fc7f | 2440 | if (!get_group_load_store_type (stmt, vectype, slp, masked_p, vls_type, |
429ef523 | 2441 | memory_access_type, gs_info)) |
2de001ee RS |
2442 | return false; |
2443 | } | |
2444 | else if (STMT_VINFO_STRIDED_P (stmt_info)) | |
2445 | { | |
2446 | gcc_assert (!slp); | |
ab2fc782 | 2447 | if (loop_vinfo |
429ef523 RS |
2448 | && vect_use_strided_gather_scatters_p (stmt, loop_vinfo, |
2449 | masked_p, gs_info)) | |
ab2fc782 RS |
2450 | *memory_access_type = VMAT_GATHER_SCATTER; |
2451 | else | |
2452 | *memory_access_type = VMAT_ELEMENTWISE; | |
2de001ee RS |
2453 | } |
2454 | else | |
62da9e14 RS |
2455 | { |
2456 | int cmp = compare_step_with_zero (stmt); | |
2457 | if (cmp < 0) | |
2458 | *memory_access_type = get_negative_load_store_type | |
2459 | (stmt, vectype, vls_type, ncopies); | |
2460 | else if (cmp == 0) | |
2461 | { | |
2462 | gcc_assert (vls_type == VLS_LOAD); | |
2463 | *memory_access_type = VMAT_INVARIANT; | |
2464 | } | |
2465 | else | |
2466 | *memory_access_type = VMAT_CONTIGUOUS; | |
2467 | } | |
2de001ee | 2468 | |
4d694b27 RS |
2469 | if ((*memory_access_type == VMAT_ELEMENTWISE |
2470 | || *memory_access_type == VMAT_STRIDED_SLP) | |
2471 | && !nunits.is_constant ()) | |
2472 | { | |
2473 | if (dump_enabled_p ()) | |
2474 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2475 | "Not using elementwise accesses due to variable " | |
2476 | "vectorization factor.\n"); | |
2477 | return false; | |
2478 | } | |
2479 | ||
2de001ee RS |
2480 | /* FIXME: At the moment the cost model seems to underestimate the |
2481 | cost of using elementwise accesses. This check preserves the | |
2482 | traditional behavior until that can be fixed. */ | |
2483 | if (*memory_access_type == VMAT_ELEMENTWISE | |
4aa157e8 | 2484 | && !STMT_VINFO_STRIDED_P (stmt_info) |
bffb8014 | 2485 | && !(stmt_info == DR_GROUP_FIRST_ELEMENT (stmt_info) |
2c53b149 RB |
2486 | && !DR_GROUP_NEXT_ELEMENT (stmt_info) |
2487 | && !pow2p_hwi (DR_GROUP_SIZE (stmt_info)))) | |
2de001ee RS |
2488 | { |
2489 | if (dump_enabled_p ()) | |
2490 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2491 | "not falling back to elementwise accesses\n"); | |
2492 | return false; | |
2493 | } | |
2494 | return true; | |
2495 | } | |
2496 | ||
aaeefd88 | 2497 | /* Return true if boolean argument MASK is suitable for vectorizing |
929b4411 RS |
2498 | conditional load or store STMT. When returning true, store the type |
2499 | of the definition in *MASK_DT_OUT and the type of the vectorized mask | |
2500 | in *MASK_VECTYPE_OUT. */ | |
aaeefd88 RS |
2501 | |
2502 | static bool | |
929b4411 RS |
2503 | vect_check_load_store_mask (gimple *stmt, tree mask, |
2504 | vect_def_type *mask_dt_out, | |
2505 | tree *mask_vectype_out) | |
aaeefd88 RS |
2506 | { |
2507 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (mask))) | |
2508 | { | |
2509 | if (dump_enabled_p ()) | |
2510 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2511 | "mask argument is not a boolean.\n"); | |
2512 | return false; | |
2513 | } | |
2514 | ||
2515 | if (TREE_CODE (mask) != SSA_NAME) | |
2516 | { | |
2517 | if (dump_enabled_p ()) | |
2518 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2519 | "mask argument is not an SSA name.\n"); | |
2520 | return false; | |
2521 | } | |
2522 | ||
2523 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
929b4411 | 2524 | enum vect_def_type mask_dt; |
aaeefd88 | 2525 | tree mask_vectype; |
894dd753 | 2526 | if (!vect_is_simple_use (mask, stmt_info->vinfo, &mask_dt, &mask_vectype)) |
aaeefd88 RS |
2527 | { |
2528 | if (dump_enabled_p ()) | |
2529 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2530 | "mask use not simple.\n"); | |
2531 | return false; | |
2532 | } | |
2533 | ||
2534 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2535 | if (!mask_vectype) | |
2536 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
2537 | ||
2538 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype)) | |
2539 | { | |
2540 | if (dump_enabled_p ()) | |
2541 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2542 | "could not find an appropriate vector mask type.\n"); | |
2543 | return false; | |
2544 | } | |
2545 | ||
2546 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (mask_vectype), | |
2547 | TYPE_VECTOR_SUBPARTS (vectype))) | |
2548 | { | |
2549 | if (dump_enabled_p ()) | |
2550 | { | |
2551 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2552 | "vector mask type "); | |
2553 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, mask_vectype); | |
2554 | dump_printf (MSG_MISSED_OPTIMIZATION, | |
2555 | " does not match vector data type "); | |
2556 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, vectype); | |
2557 | dump_printf (MSG_MISSED_OPTIMIZATION, ".\n"); | |
2558 | } | |
2559 | return false; | |
2560 | } | |
2561 | ||
929b4411 | 2562 | *mask_dt_out = mask_dt; |
aaeefd88 RS |
2563 | *mask_vectype_out = mask_vectype; |
2564 | return true; | |
2565 | } | |
2566 | ||
3133c3b6 RS |
2567 | /* Return true if stored value RHS is suitable for vectorizing store |
2568 | statement STMT. When returning true, store the type of the | |
929b4411 RS |
2569 | definition in *RHS_DT_OUT, the type of the vectorized store value in |
2570 | *RHS_VECTYPE_OUT and the type of the store in *VLS_TYPE_OUT. */ | |
3133c3b6 RS |
2571 | |
2572 | static bool | |
929b4411 RS |
2573 | vect_check_store_rhs (gimple *stmt, tree rhs, vect_def_type *rhs_dt_out, |
2574 | tree *rhs_vectype_out, vec_load_store_type *vls_type_out) | |
3133c3b6 RS |
2575 | { |
2576 | /* In the case this is a store from a constant make sure | |
2577 | native_encode_expr can handle it. */ | |
2578 | if (CONSTANT_CLASS_P (rhs) && native_encode_expr (rhs, NULL, 64) == 0) | |
2579 | { | |
2580 | if (dump_enabled_p ()) | |
2581 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2582 | "cannot encode constant as a byte sequence.\n"); | |
2583 | return false; | |
2584 | } | |
2585 | ||
2586 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
929b4411 | 2587 | enum vect_def_type rhs_dt; |
3133c3b6 | 2588 | tree rhs_vectype; |
894dd753 | 2589 | if (!vect_is_simple_use (rhs, stmt_info->vinfo, &rhs_dt, &rhs_vectype)) |
3133c3b6 RS |
2590 | { |
2591 | if (dump_enabled_p ()) | |
2592 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2593 | "use not simple.\n"); | |
2594 | return false; | |
2595 | } | |
2596 | ||
2597 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2598 | if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype)) | |
2599 | { | |
2600 | if (dump_enabled_p ()) | |
2601 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2602 | "incompatible vector types.\n"); | |
2603 | return false; | |
2604 | } | |
2605 | ||
929b4411 | 2606 | *rhs_dt_out = rhs_dt; |
3133c3b6 | 2607 | *rhs_vectype_out = rhs_vectype; |
929b4411 | 2608 | if (rhs_dt == vect_constant_def || rhs_dt == vect_external_def) |
3133c3b6 RS |
2609 | *vls_type_out = VLS_STORE_INVARIANT; |
2610 | else | |
2611 | *vls_type_out = VLS_STORE; | |
2612 | return true; | |
2613 | } | |
2614 | ||
bc9587eb RS |
2615 | /* Build an all-ones vector mask of type MASKTYPE while vectorizing STMT. |
2616 | Note that we support masks with floating-point type, in which case the | |
2617 | floats are interpreted as a bitmask. */ | |
2618 | ||
2619 | static tree | |
2620 | vect_build_all_ones_mask (gimple *stmt, tree masktype) | |
2621 | { | |
2622 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
2623 | return build_int_cst (masktype, -1); | |
2624 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
2625 | { | |
2626 | tree mask = build_int_cst (TREE_TYPE (masktype), -1); | |
2627 | mask = build_vector_from_val (masktype, mask); | |
2628 | return vect_init_vector (stmt, mask, masktype, NULL); | |
2629 | } | |
2630 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) | |
2631 | { | |
2632 | REAL_VALUE_TYPE r; | |
2633 | long tmp[6]; | |
2634 | for (int j = 0; j < 6; ++j) | |
2635 | tmp[j] = -1; | |
2636 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
2637 | tree mask = build_real (TREE_TYPE (masktype), r); | |
2638 | mask = build_vector_from_val (masktype, mask); | |
2639 | return vect_init_vector (stmt, mask, masktype, NULL); | |
2640 | } | |
2641 | gcc_unreachable (); | |
2642 | } | |
2643 | ||
2644 | /* Build an all-zero merge value of type VECTYPE while vectorizing | |
2645 | STMT as a gather load. */ | |
2646 | ||
2647 | static tree | |
2648 | vect_build_zero_merge_argument (gimple *stmt, tree vectype) | |
2649 | { | |
2650 | tree merge; | |
2651 | if (TREE_CODE (TREE_TYPE (vectype)) == INTEGER_TYPE) | |
2652 | merge = build_int_cst (TREE_TYPE (vectype), 0); | |
2653 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (vectype))) | |
2654 | { | |
2655 | REAL_VALUE_TYPE r; | |
2656 | long tmp[6]; | |
2657 | for (int j = 0; j < 6; ++j) | |
2658 | tmp[j] = 0; | |
2659 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (vectype))); | |
2660 | merge = build_real (TREE_TYPE (vectype), r); | |
2661 | } | |
2662 | else | |
2663 | gcc_unreachable (); | |
2664 | merge = build_vector_from_val (vectype, merge); | |
2665 | return vect_init_vector (stmt, merge, vectype, NULL); | |
2666 | } | |
2667 | ||
c48d2d35 RS |
2668 | /* Build a gather load call while vectorizing STMT. Insert new instructions |
2669 | before GSI and add them to VEC_STMT. GS_INFO describes the gather load | |
2670 | operation. If the load is conditional, MASK is the unvectorized | |
929b4411 | 2671 | condition and MASK_DT is its definition type, otherwise MASK is null. */ |
c48d2d35 RS |
2672 | |
2673 | static void | |
2674 | vect_build_gather_load_calls (gimple *stmt, gimple_stmt_iterator *gsi, | |
1eede195 RS |
2675 | stmt_vec_info *vec_stmt, |
2676 | gather_scatter_info *gs_info, tree mask, | |
2677 | vect_def_type mask_dt) | |
c48d2d35 RS |
2678 | { |
2679 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2680 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2681 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2682 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2683 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2684 | int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
2685 | edge pe = loop_preheader_edge (loop); | |
2686 | enum { NARROW, NONE, WIDEN } modifier; | |
2687 | poly_uint64 gather_off_nunits | |
2688 | = TYPE_VECTOR_SUBPARTS (gs_info->offset_vectype); | |
2689 | ||
2690 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info->decl)); | |
2691 | tree rettype = TREE_TYPE (TREE_TYPE (gs_info->decl)); | |
2692 | tree srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2693 | tree ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2694 | tree idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2695 | tree masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2696 | tree scaletype = TREE_VALUE (arglist); | |
2697 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
2698 | && (!mask || types_compatible_p (srctype, masktype))); | |
2699 | ||
2700 | tree perm_mask = NULL_TREE; | |
2701 | tree mask_perm_mask = NULL_TREE; | |
2702 | if (known_eq (nunits, gather_off_nunits)) | |
2703 | modifier = NONE; | |
2704 | else if (known_eq (nunits * 2, gather_off_nunits)) | |
2705 | { | |
2706 | modifier = WIDEN; | |
2707 | ||
2708 | /* Currently widening gathers and scatters are only supported for | |
2709 | fixed-length vectors. */ | |
2710 | int count = gather_off_nunits.to_constant (); | |
2711 | vec_perm_builder sel (count, count, 1); | |
2712 | for (int i = 0; i < count; ++i) | |
2713 | sel.quick_push (i | (count / 2)); | |
2714 | ||
2715 | vec_perm_indices indices (sel, 1, count); | |
2716 | perm_mask = vect_gen_perm_mask_checked (gs_info->offset_vectype, | |
2717 | indices); | |
2718 | } | |
2719 | else if (known_eq (nunits, gather_off_nunits * 2)) | |
2720 | { | |
2721 | modifier = NARROW; | |
2722 | ||
2723 | /* Currently narrowing gathers and scatters are only supported for | |
2724 | fixed-length vectors. */ | |
2725 | int count = nunits.to_constant (); | |
2726 | vec_perm_builder sel (count, count, 1); | |
2727 | sel.quick_grow (count); | |
2728 | for (int i = 0; i < count; ++i) | |
2729 | sel[i] = i < count / 2 ? i : i + count / 2; | |
2730 | vec_perm_indices indices (sel, 2, count); | |
2731 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); | |
2732 | ||
2733 | ncopies *= 2; | |
2734 | ||
2735 | if (mask) | |
2736 | { | |
2737 | for (int i = 0; i < count; ++i) | |
2738 | sel[i] = i | (count / 2); | |
2739 | indices.new_vector (sel, 2, count); | |
2740 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, indices); | |
2741 | } | |
2742 | } | |
2743 | else | |
2744 | gcc_unreachable (); | |
2745 | ||
2746 | tree vec_dest = vect_create_destination_var (gimple_get_lhs (stmt), | |
2747 | vectype); | |
2748 | ||
2749 | tree ptr = fold_convert (ptrtype, gs_info->base); | |
2750 | if (!is_gimple_min_invariant (ptr)) | |
2751 | { | |
2752 | gimple_seq seq; | |
2753 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
2754 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
2755 | gcc_assert (!new_bb); | |
2756 | } | |
2757 | ||
2758 | tree scale = build_int_cst (scaletype, gs_info->scale); | |
2759 | ||
2760 | tree vec_oprnd0 = NULL_TREE; | |
2761 | tree vec_mask = NULL_TREE; | |
2762 | tree src_op = NULL_TREE; | |
2763 | tree mask_op = NULL_TREE; | |
2764 | tree prev_res = NULL_TREE; | |
2765 | stmt_vec_info prev_stmt_info = NULL; | |
2766 | ||
2767 | if (!mask) | |
2768 | { | |
2769 | src_op = vect_build_zero_merge_argument (stmt, rettype); | |
2770 | mask_op = vect_build_all_ones_mask (stmt, masktype); | |
2771 | } | |
2772 | ||
2773 | for (int j = 0; j < ncopies; ++j) | |
2774 | { | |
2775 | tree op, var; | |
c48d2d35 RS |
2776 | if (modifier == WIDEN && (j & 1)) |
2777 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
2778 | perm_mask, stmt, gsi); | |
2779 | else if (j == 0) | |
2780 | op = vec_oprnd0 | |
2781 | = vect_get_vec_def_for_operand (gs_info->offset, stmt); | |
2782 | else | |
2783 | op = vec_oprnd0 | |
2784 | = vect_get_vec_def_for_stmt_copy (gs_info->offset_dt, vec_oprnd0); | |
2785 | ||
2786 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
2787 | { | |
2788 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), | |
2789 | TYPE_VECTOR_SUBPARTS (idxtype))); | |
2790 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); | |
2791 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); | |
e1bd7296 | 2792 | gassign *new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
c48d2d35 RS |
2793 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2794 | op = var; | |
2795 | } | |
2796 | ||
2797 | if (mask) | |
2798 | { | |
2799 | if (mask_perm_mask && (j & 1)) | |
2800 | mask_op = permute_vec_elements (mask_op, mask_op, | |
2801 | mask_perm_mask, stmt, gsi); | |
2802 | else | |
2803 | { | |
2804 | if (j == 0) | |
2805 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); | |
2806 | else | |
929b4411 | 2807 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
c48d2d35 RS |
2808 | |
2809 | mask_op = vec_mask; | |
2810 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
2811 | { | |
2812 | gcc_assert | |
2813 | (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)), | |
2814 | TYPE_VECTOR_SUBPARTS (masktype))); | |
2815 | var = vect_get_new_ssa_name (masktype, vect_simple_var); | |
2816 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); | |
e1bd7296 RS |
2817 | gassign *new_stmt |
2818 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_op); | |
c48d2d35 RS |
2819 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2820 | mask_op = var; | |
2821 | } | |
2822 | } | |
2823 | src_op = mask_op; | |
2824 | } | |
2825 | ||
e1bd7296 RS |
2826 | gcall *new_call = gimple_build_call (gs_info->decl, 5, src_op, ptr, op, |
2827 | mask_op, scale); | |
c48d2d35 | 2828 | |
e1bd7296 | 2829 | stmt_vec_info new_stmt_info; |
c48d2d35 RS |
2830 | if (!useless_type_conversion_p (vectype, rettype)) |
2831 | { | |
2832 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (vectype), | |
2833 | TYPE_VECTOR_SUBPARTS (rettype))); | |
2834 | op = vect_get_new_ssa_name (rettype, vect_simple_var); | |
e1bd7296 RS |
2835 | gimple_call_set_lhs (new_call, op); |
2836 | vect_finish_stmt_generation (stmt, new_call, gsi); | |
c48d2d35 RS |
2837 | var = make_ssa_name (vec_dest); |
2838 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); | |
e1bd7296 RS |
2839 | gassign *new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
2840 | new_stmt_info = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
c48d2d35 RS |
2841 | } |
2842 | else | |
2843 | { | |
e1bd7296 RS |
2844 | var = make_ssa_name (vec_dest, new_call); |
2845 | gimple_call_set_lhs (new_call, var); | |
2846 | new_stmt_info = vect_finish_stmt_generation (stmt, new_call, gsi); | |
c48d2d35 RS |
2847 | } |
2848 | ||
c48d2d35 RS |
2849 | if (modifier == NARROW) |
2850 | { | |
2851 | if ((j & 1) == 0) | |
2852 | { | |
2853 | prev_res = var; | |
2854 | continue; | |
2855 | } | |
2856 | var = permute_vec_elements (prev_res, var, perm_mask, stmt, gsi); | |
e1bd7296 | 2857 | new_stmt_info = loop_vinfo->lookup_def (var); |
c48d2d35 RS |
2858 | } |
2859 | ||
dbe1b846 | 2860 | if (prev_stmt_info == NULL_STMT_VEC_INFO) |
e1bd7296 | 2861 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
c48d2d35 | 2862 | else |
e1bd7296 RS |
2863 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
2864 | prev_stmt_info = new_stmt_info; | |
c48d2d35 RS |
2865 | } |
2866 | } | |
2867 | ||
bfaa08b7 RS |
2868 | /* Prepare the base and offset in GS_INFO for vectorization. |
2869 | Set *DATAREF_PTR to the loop-invariant base address and *VEC_OFFSET | |
2870 | to the vectorized offset argument for the first copy of STMT. STMT | |
2871 | is the statement described by GS_INFO and LOOP is the containing loop. */ | |
2872 | ||
2873 | static void | |
2874 | vect_get_gather_scatter_ops (struct loop *loop, gimple *stmt, | |
2875 | gather_scatter_info *gs_info, | |
2876 | tree *dataref_ptr, tree *vec_offset) | |
2877 | { | |
2878 | gimple_seq stmts = NULL; | |
2879 | *dataref_ptr = force_gimple_operand (gs_info->base, &stmts, true, NULL_TREE); | |
2880 | if (stmts != NULL) | |
2881 | { | |
2882 | basic_block new_bb; | |
2883 | edge pe = loop_preheader_edge (loop); | |
2884 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
2885 | gcc_assert (!new_bb); | |
2886 | } | |
2887 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2888 | tree offset_vectype = get_vectype_for_scalar_type (offset_type); | |
2889 | *vec_offset = vect_get_vec_def_for_operand (gs_info->offset, stmt, | |
2890 | offset_vectype); | |
2891 | } | |
2892 | ||
ab2fc782 RS |
2893 | /* Prepare to implement a grouped or strided load or store using |
2894 | the gather load or scatter store operation described by GS_INFO. | |
2895 | STMT is the load or store statement. | |
2896 | ||
2897 | Set *DATAREF_BUMP to the amount that should be added to the base | |
2898 | address after each copy of the vectorized statement. Set *VEC_OFFSET | |
2899 | to an invariant offset vector in which element I has the value | |
2900 | I * DR_STEP / SCALE. */ | |
2901 | ||
2902 | static void | |
2903 | vect_get_strided_load_store_ops (gimple *stmt, loop_vec_info loop_vinfo, | |
2904 | gather_scatter_info *gs_info, | |
2905 | tree *dataref_bump, tree *vec_offset) | |
2906 | { | |
2907 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2908 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2909 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2910 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2911 | gimple_seq stmts; | |
2912 | ||
2913 | tree bump = size_binop (MULT_EXPR, | |
2914 | fold_convert (sizetype, DR_STEP (dr)), | |
2915 | size_int (TYPE_VECTOR_SUBPARTS (vectype))); | |
2916 | *dataref_bump = force_gimple_operand (bump, &stmts, true, NULL_TREE); | |
2917 | if (stmts) | |
2918 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
2919 | ||
2920 | /* The offset given in GS_INFO can have pointer type, so use the element | |
2921 | type of the vector instead. */ | |
2922 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2923 | tree offset_vectype = get_vectype_for_scalar_type (offset_type); | |
2924 | offset_type = TREE_TYPE (offset_vectype); | |
2925 | ||
2926 | /* Calculate X = DR_STEP / SCALE and convert it to the appropriate type. */ | |
2927 | tree step = size_binop (EXACT_DIV_EXPR, DR_STEP (dr), | |
2928 | ssize_int (gs_info->scale)); | |
2929 | step = fold_convert (offset_type, step); | |
2930 | step = force_gimple_operand (step, &stmts, true, NULL_TREE); | |
2931 | ||
2932 | /* Create {0, X, X*2, X*3, ...}. */ | |
2933 | *vec_offset = gimple_build (&stmts, VEC_SERIES_EXPR, offset_vectype, | |
2934 | build_zero_cst (offset_type), step); | |
2935 | if (stmts) | |
2936 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
2937 | } | |
2938 | ||
2939 | /* Return the amount that should be added to a vector pointer to move | |
2940 | to the next or previous copy of AGGR_TYPE. DR is the data reference | |
2941 | being vectorized and MEMORY_ACCESS_TYPE describes the type of | |
2942 | vectorization. */ | |
2943 | ||
2944 | static tree | |
2945 | vect_get_data_ptr_increment (data_reference *dr, tree aggr_type, | |
2946 | vect_memory_access_type memory_access_type) | |
2947 | { | |
2948 | if (memory_access_type == VMAT_INVARIANT) | |
2949 | return size_zero_node; | |
2950 | ||
2951 | tree iv_step = TYPE_SIZE_UNIT (aggr_type); | |
2952 | tree step = vect_dr_behavior (dr)->step; | |
2953 | if (tree_int_cst_sgn (step) == -1) | |
2954 | iv_step = fold_build1 (NEGATE_EXPR, TREE_TYPE (iv_step), iv_step); | |
2955 | return iv_step; | |
2956 | } | |
2957 | ||
37b14185 RB |
2958 | /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64}. */ |
2959 | ||
2960 | static bool | |
2961 | vectorizable_bswap (gimple *stmt, gimple_stmt_iterator *gsi, | |
1eede195 | 2962 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
68435eb2 RB |
2963 | tree vectype_in, enum vect_def_type *dt, |
2964 | stmt_vector_for_cost *cost_vec) | |
37b14185 RB |
2965 | { |
2966 | tree op, vectype; | |
2967 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2968 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
928686b1 RS |
2969 | unsigned ncopies; |
2970 | unsigned HOST_WIDE_INT nunits, num_bytes; | |
37b14185 RB |
2971 | |
2972 | op = gimple_call_arg (stmt, 0); | |
2973 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
928686b1 RS |
2974 | |
2975 | if (!TYPE_VECTOR_SUBPARTS (vectype).is_constant (&nunits)) | |
2976 | return false; | |
37b14185 RB |
2977 | |
2978 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2979 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2980 | case of SLP. */ | |
2981 | if (slp_node) | |
2982 | ncopies = 1; | |
2983 | else | |
e8f142e2 | 2984 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
37b14185 RB |
2985 | |
2986 | gcc_assert (ncopies >= 1); | |
2987 | ||
2988 | tree char_vectype = get_same_sized_vectype (char_type_node, vectype_in); | |
2989 | if (! char_vectype) | |
2990 | return false; | |
2991 | ||
928686b1 RS |
2992 | if (!TYPE_VECTOR_SUBPARTS (char_vectype).is_constant (&num_bytes)) |
2993 | return false; | |
2994 | ||
794e3180 | 2995 | unsigned word_bytes = num_bytes / nunits; |
908a1a16 | 2996 | |
d980067b RS |
2997 | /* The encoding uses one stepped pattern for each byte in the word. */ |
2998 | vec_perm_builder elts (num_bytes, word_bytes, 3); | |
2999 | for (unsigned i = 0; i < 3; ++i) | |
37b14185 | 3000 | for (unsigned j = 0; j < word_bytes; ++j) |
908a1a16 | 3001 | elts.quick_push ((i + 1) * word_bytes - j - 1); |
37b14185 | 3002 | |
e3342de4 RS |
3003 | vec_perm_indices indices (elts, 1, num_bytes); |
3004 | if (!can_vec_perm_const_p (TYPE_MODE (char_vectype), indices)) | |
37b14185 RB |
3005 | return false; |
3006 | ||
3007 | if (! vec_stmt) | |
3008 | { | |
3009 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
adac3a68 | 3010 | DUMP_VECT_SCOPE ("vectorizable_bswap"); |
78604de0 | 3011 | if (! slp_node) |
37b14185 | 3012 | { |
68435eb2 RB |
3013 | record_stmt_cost (cost_vec, |
3014 | 1, vector_stmt, stmt_info, 0, vect_prologue); | |
3015 | record_stmt_cost (cost_vec, | |
3016 | ncopies, vec_perm, stmt_info, 0, vect_body); | |
37b14185 RB |
3017 | } |
3018 | return true; | |
3019 | } | |
3020 | ||
736d0f28 | 3021 | tree bswap_vconst = vec_perm_indices_to_tree (char_vectype, indices); |
37b14185 RB |
3022 | |
3023 | /* Transform. */ | |
3024 | vec<tree> vec_oprnds = vNULL; | |
e1bd7296 | 3025 | stmt_vec_info new_stmt_info = NULL; |
37b14185 RB |
3026 | stmt_vec_info prev_stmt_info = NULL; |
3027 | for (unsigned j = 0; j < ncopies; j++) | |
3028 | { | |
3029 | /* Handle uses. */ | |
3030 | if (j == 0) | |
306b0c92 | 3031 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
37b14185 RB |
3032 | else |
3033 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
3034 | ||
3035 | /* Arguments are ready. create the new vector stmt. */ | |
3036 | unsigned i; | |
3037 | tree vop; | |
3038 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) | |
3039 | { | |
e1bd7296 | 3040 | gimple *new_stmt; |
37b14185 RB |
3041 | tree tem = make_ssa_name (char_vectype); |
3042 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
3043 | char_vectype, vop)); | |
3044 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3045 | tree tem2 = make_ssa_name (char_vectype); | |
3046 | new_stmt = gimple_build_assign (tem2, VEC_PERM_EXPR, | |
3047 | tem, tem, bswap_vconst); | |
3048 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3049 | tem = make_ssa_name (vectype); | |
3050 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
3051 | vectype, tem2)); | |
e1bd7296 | 3052 | new_stmt_info = vect_finish_stmt_generation (stmt, new_stmt, gsi); |
37b14185 | 3053 | if (slp_node) |
e1bd7296 | 3054 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
37b14185 RB |
3055 | } |
3056 | ||
3057 | if (slp_node) | |
3058 | continue; | |
3059 | ||
3060 | if (j == 0) | |
e1bd7296 | 3061 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
37b14185 | 3062 | else |
e1bd7296 | 3063 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
37b14185 | 3064 | |
e1bd7296 | 3065 | prev_stmt_info = new_stmt_info; |
37b14185 RB |
3066 | } |
3067 | ||
3068 | vec_oprnds.release (); | |
3069 | return true; | |
3070 | } | |
3071 | ||
b1b6836e RS |
3072 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
3073 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT | |
3074 | in a single step. On success, store the binary pack code in | |
3075 | *CONVERT_CODE. */ | |
3076 | ||
3077 | static bool | |
3078 | simple_integer_narrowing (tree vectype_out, tree vectype_in, | |
3079 | tree_code *convert_code) | |
3080 | { | |
3081 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) | |
3082 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) | |
3083 | return false; | |
3084 | ||
3085 | tree_code code; | |
3086 | int multi_step_cvt = 0; | |
3087 | auto_vec <tree, 8> interm_types; | |
3088 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, | |
3089 | &code, &multi_step_cvt, | |
3090 | &interm_types) | |
3091 | || multi_step_cvt) | |
3092 | return false; | |
3093 | ||
3094 | *convert_code = code; | |
3095 | return true; | |
3096 | } | |
5ce9450f | 3097 | |
ebfd146a IR |
3098 | /* Function vectorizable_call. |
3099 | ||
538dd0b7 | 3100 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 3101 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
3102 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3103 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3104 | ||
3105 | static bool | |
1eede195 RS |
3106 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, |
3107 | stmt_vec_info *vec_stmt, slp_tree slp_node, | |
3108 | stmt_vector_for_cost *cost_vec) | |
ebfd146a | 3109 | { |
538dd0b7 | 3110 | gcall *stmt; |
ebfd146a IR |
3111 | tree vec_dest; |
3112 | tree scalar_dest; | |
0267732b | 3113 | tree op; |
ebfd146a | 3114 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
538dd0b7 | 3115 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a | 3116 | tree vectype_out, vectype_in; |
c7bda0f4 RS |
3117 | poly_uint64 nunits_in; |
3118 | poly_uint64 nunits_out; | |
ebfd146a | 3119 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
190c2236 | 3120 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 3121 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 3122 | tree fndecl, new_temp, rhs_type; |
2c58d42c RS |
3123 | enum vect_def_type dt[4] |
3124 | = { vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type, | |
3125 | vect_unknown_def_type }; | |
3126 | int ndts = ARRAY_SIZE (dt); | |
ebfd146a | 3127 | int ncopies, j; |
2c58d42c RS |
3128 | auto_vec<tree, 8> vargs; |
3129 | auto_vec<tree, 8> orig_vargs; | |
ebfd146a IR |
3130 | enum { NARROW, NONE, WIDEN } modifier; |
3131 | size_t i, nargs; | |
9d5e7640 | 3132 | tree lhs; |
ebfd146a | 3133 | |
190c2236 | 3134 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3135 | return false; |
3136 | ||
66c16fd9 RB |
3137 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3138 | && ! vec_stmt) | |
ebfd146a IR |
3139 | return false; |
3140 | ||
538dd0b7 DM |
3141 | /* Is GS a vectorizable call? */ |
3142 | stmt = dyn_cast <gcall *> (gs); | |
3143 | if (!stmt) | |
ebfd146a IR |
3144 | return false; |
3145 | ||
5ce9450f | 3146 | if (gimple_call_internal_p (stmt) |
bfaa08b7 | 3147 | && (internal_load_fn_p (gimple_call_internal_fn (stmt)) |
f307441a | 3148 | || internal_store_fn_p (gimple_call_internal_fn (stmt)))) |
c3a8f964 RS |
3149 | /* Handled by vectorizable_load and vectorizable_store. */ |
3150 | return false; | |
5ce9450f | 3151 | |
0136f8f0 AH |
3152 | if (gimple_call_lhs (stmt) == NULL_TREE |
3153 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
3154 | return false; |
3155 | ||
0136f8f0 | 3156 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 3157 | |
b690cc0f RG |
3158 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3159 | ||
ebfd146a IR |
3160 | /* Process function arguments. */ |
3161 | rhs_type = NULL_TREE; | |
b690cc0f | 3162 | vectype_in = NULL_TREE; |
ebfd146a IR |
3163 | nargs = gimple_call_num_args (stmt); |
3164 | ||
1b1562a5 MM |
3165 | /* Bail out if the function has more than three arguments, we do not have |
3166 | interesting builtin functions to vectorize with more than two arguments | |
3167 | except for fma. No arguments is also not good. */ | |
2c58d42c | 3168 | if (nargs == 0 || nargs > 4) |
ebfd146a IR |
3169 | return false; |
3170 | ||
74bf76ed | 3171 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2c58d42c RS |
3172 | combined_fn cfn = gimple_call_combined_fn (stmt); |
3173 | if (cfn == CFN_GOMP_SIMD_LANE) | |
74bf76ed JJ |
3174 | { |
3175 | nargs = 0; | |
3176 | rhs_type = unsigned_type_node; | |
3177 | } | |
3178 | ||
2c58d42c RS |
3179 | int mask_opno = -1; |
3180 | if (internal_fn_p (cfn)) | |
3181 | mask_opno = internal_fn_mask_index (as_internal_fn (cfn)); | |
3182 | ||
ebfd146a IR |
3183 | for (i = 0; i < nargs; i++) |
3184 | { | |
b690cc0f RG |
3185 | tree opvectype; |
3186 | ||
ebfd146a | 3187 | op = gimple_call_arg (stmt, i); |
2c58d42c RS |
3188 | if (!vect_is_simple_use (op, vinfo, &dt[i], &opvectype)) |
3189 | { | |
3190 | if (dump_enabled_p ()) | |
3191 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3192 | "use not simple.\n"); | |
3193 | return false; | |
3194 | } | |
3195 | ||
3196 | /* Skip the mask argument to an internal function. This operand | |
3197 | has been converted via a pattern if necessary. */ | |
3198 | if ((int) i == mask_opno) | |
3199 | continue; | |
ebfd146a IR |
3200 | |
3201 | /* We can only handle calls with arguments of the same type. */ | |
3202 | if (rhs_type | |
8533c9d8 | 3203 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 3204 | { |
73fbfcad | 3205 | if (dump_enabled_p ()) |
78c60e3d | 3206 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3207 | "argument types differ.\n"); |
ebfd146a IR |
3208 | return false; |
3209 | } | |
b690cc0f RG |
3210 | if (!rhs_type) |
3211 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 3212 | |
b690cc0f RG |
3213 | if (!vectype_in) |
3214 | vectype_in = opvectype; | |
3215 | else if (opvectype | |
3216 | && opvectype != vectype_in) | |
3217 | { | |
73fbfcad | 3218 | if (dump_enabled_p ()) |
78c60e3d | 3219 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3220 | "argument vector types differ.\n"); |
b690cc0f RG |
3221 | return false; |
3222 | } | |
3223 | } | |
3224 | /* If all arguments are external or constant defs use a vector type with | |
3225 | the same size as the output vector type. */ | |
ebfd146a | 3226 | if (!vectype_in) |
b690cc0f | 3227 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
3228 | if (vec_stmt) |
3229 | gcc_assert (vectype_in); | |
3230 | if (!vectype_in) | |
3231 | { | |
73fbfcad | 3232 | if (dump_enabled_p ()) |
7d8930a0 | 3233 | { |
78c60e3d SS |
3234 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3235 | "no vectype for scalar type "); | |
3236 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 3237 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
3238 | } |
3239 | ||
3240 | return false; | |
3241 | } | |
ebfd146a IR |
3242 | |
3243 | /* FORNOW */ | |
b690cc0f RG |
3244 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
3245 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
c7bda0f4 | 3246 | if (known_eq (nunits_in * 2, nunits_out)) |
ebfd146a | 3247 | modifier = NARROW; |
c7bda0f4 | 3248 | else if (known_eq (nunits_out, nunits_in)) |
ebfd146a | 3249 | modifier = NONE; |
c7bda0f4 | 3250 | else if (known_eq (nunits_out * 2, nunits_in)) |
ebfd146a IR |
3251 | modifier = WIDEN; |
3252 | else | |
3253 | return false; | |
3254 | ||
70439f0d RS |
3255 | /* We only handle functions that do not read or clobber memory. */ |
3256 | if (gimple_vuse (stmt)) | |
3257 | { | |
3258 | if (dump_enabled_p ()) | |
3259 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3260 | "function reads from or writes to memory.\n"); | |
3261 | return false; | |
3262 | } | |
3263 | ||
ebfd146a IR |
3264 | /* For now, we only vectorize functions if a target specific builtin |
3265 | is available. TODO -- in some cases, it might be profitable to | |
3266 | insert the calls for pieces of the vector, in order to be able | |
3267 | to vectorize other operations in the loop. */ | |
70439f0d RS |
3268 | fndecl = NULL_TREE; |
3269 | internal_fn ifn = IFN_LAST; | |
70439f0d RS |
3270 | tree callee = gimple_call_fndecl (stmt); |
3271 | ||
3272 | /* First try using an internal function. */ | |
b1b6836e RS |
3273 | tree_code convert_code = ERROR_MARK; |
3274 | if (cfn != CFN_LAST | |
3275 | && (modifier == NONE | |
3276 | || (modifier == NARROW | |
3277 | && simple_integer_narrowing (vectype_out, vectype_in, | |
3278 | &convert_code)))) | |
70439f0d RS |
3279 | ifn = vectorizable_internal_function (cfn, callee, vectype_out, |
3280 | vectype_in); | |
3281 | ||
3282 | /* If that fails, try asking for a target-specific built-in function. */ | |
3283 | if (ifn == IFN_LAST) | |
3284 | { | |
3285 | if (cfn != CFN_LAST) | |
3286 | fndecl = targetm.vectorize.builtin_vectorized_function | |
3287 | (cfn, vectype_out, vectype_in); | |
7672aa9b | 3288 | else if (callee) |
70439f0d RS |
3289 | fndecl = targetm.vectorize.builtin_md_vectorized_function |
3290 | (callee, vectype_out, vectype_in); | |
3291 | } | |
3292 | ||
3293 | if (ifn == IFN_LAST && !fndecl) | |
ebfd146a | 3294 | { |
70439f0d | 3295 | if (cfn == CFN_GOMP_SIMD_LANE |
74bf76ed JJ |
3296 | && !slp_node |
3297 | && loop_vinfo | |
3298 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
3299 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
3300 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
3301 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
3302 | { | |
3303 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
3304 | { 0, 1, 2, ... vf - 1 } vector. */ | |
3305 | gcc_assert (nargs == 0); | |
3306 | } | |
37b14185 RB |
3307 | else if (modifier == NONE |
3308 | && (gimple_call_builtin_p (stmt, BUILT_IN_BSWAP16) | |
3309 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP32) | |
3310 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP64))) | |
3311 | return vectorizable_bswap (stmt, gsi, vec_stmt, slp_node, | |
68435eb2 | 3312 | vectype_in, dt, cost_vec); |
74bf76ed JJ |
3313 | else |
3314 | { | |
3315 | if (dump_enabled_p ()) | |
3316 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 3317 | "function is not vectorizable.\n"); |
74bf76ed JJ |
3318 | return false; |
3319 | } | |
ebfd146a IR |
3320 | } |
3321 | ||
fce57248 | 3322 | if (slp_node) |
190c2236 | 3323 | ncopies = 1; |
b1b6836e | 3324 | else if (modifier == NARROW && ifn == IFN_LAST) |
e8f142e2 | 3325 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
ebfd146a | 3326 | else |
e8f142e2 | 3327 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
ebfd146a IR |
3328 | |
3329 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3330 | needs to be generated. */ | |
3331 | gcc_assert (ncopies >= 1); | |
3332 | ||
ed623edb | 3333 | vec_loop_masks *masks = (loop_vinfo ? &LOOP_VINFO_MASKS (loop_vinfo) : NULL); |
ebfd146a IR |
3334 | if (!vec_stmt) /* transformation not required. */ |
3335 | { | |
3336 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
adac3a68 | 3337 | DUMP_VECT_SCOPE ("vectorizable_call"); |
68435eb2 RB |
3338 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
3339 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) | |
3340 | record_stmt_cost (cost_vec, ncopies / 2, | |
3341 | vec_promote_demote, stmt_info, 0, vect_body); | |
b1b6836e | 3342 | |
2c58d42c RS |
3343 | if (loop_vinfo && mask_opno >= 0) |
3344 | { | |
3345 | unsigned int nvectors = (slp_node | |
3346 | ? SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node) | |
3347 | : ncopies); | |
3348 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype_out); | |
3349 | } | |
ebfd146a IR |
3350 | return true; |
3351 | } | |
3352 | ||
67b8dbac | 3353 | /* Transform. */ |
ebfd146a | 3354 | |
73fbfcad | 3355 | if (dump_enabled_p ()) |
e645e942 | 3356 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
3357 | |
3358 | /* Handle def. */ | |
3359 | scalar_dest = gimple_call_lhs (stmt); | |
3360 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
3361 | ||
2c58d42c RS |
3362 | bool masked_loop_p = loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo); |
3363 | ||
e1bd7296 | 3364 | stmt_vec_info new_stmt_info = NULL; |
ebfd146a | 3365 | prev_stmt_info = NULL; |
b1b6836e | 3366 | if (modifier == NONE || ifn != IFN_LAST) |
ebfd146a | 3367 | { |
b1b6836e | 3368 | tree prev_res = NULL_TREE; |
2c58d42c RS |
3369 | vargs.safe_grow (nargs); |
3370 | orig_vargs.safe_grow (nargs); | |
ebfd146a IR |
3371 | for (j = 0; j < ncopies; ++j) |
3372 | { | |
3373 | /* Build argument list for the vectorized call. */ | |
190c2236 JJ |
3374 | if (slp_node) |
3375 | { | |
ef062b13 | 3376 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 3377 | vec<tree> vec_oprnds0; |
190c2236 JJ |
3378 | |
3379 | for (i = 0; i < nargs; i++) | |
2c58d42c | 3380 | vargs[i] = gimple_call_arg (stmt, i); |
306b0c92 | 3381 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 3382 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
3383 | |
3384 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3385 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
3386 | { |
3387 | size_t k; | |
3388 | for (k = 0; k < nargs; k++) | |
3389 | { | |
37b5ec8f | 3390 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 3391 | vargs[k] = vec_oprndsk[i]; |
190c2236 | 3392 | } |
b1b6836e RS |
3393 | if (modifier == NARROW) |
3394 | { | |
2c58d42c RS |
3395 | /* We don't define any narrowing conditional functions |
3396 | at present. */ | |
3397 | gcc_assert (mask_opno < 0); | |
b1b6836e | 3398 | tree half_res = make_ssa_name (vectype_in); |
a844293d RS |
3399 | gcall *call |
3400 | = gimple_build_call_internal_vec (ifn, vargs); | |
3401 | gimple_call_set_lhs (call, half_res); | |
3402 | gimple_call_set_nothrow (call, true); | |
e1bd7296 RS |
3403 | new_stmt_info |
3404 | = vect_finish_stmt_generation (stmt, call, gsi); | |
b1b6836e RS |
3405 | if ((i & 1) == 0) |
3406 | { | |
3407 | prev_res = half_res; | |
3408 | continue; | |
3409 | } | |
3410 | new_temp = make_ssa_name (vec_dest); | |
e1bd7296 RS |
3411 | gimple *new_stmt |
3412 | = gimple_build_assign (new_temp, convert_code, | |
3413 | prev_res, half_res); | |
3414 | new_stmt_info | |
3415 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b1b6836e | 3416 | } |
70439f0d | 3417 | else |
b1b6836e | 3418 | { |
2c58d42c RS |
3419 | if (mask_opno >= 0 && masked_loop_p) |
3420 | { | |
3421 | unsigned int vec_num = vec_oprnds0.length (); | |
3422 | /* Always true for SLP. */ | |
3423 | gcc_assert (ncopies == 1); | |
3424 | tree mask = vect_get_loop_mask (gsi, masks, vec_num, | |
3425 | vectype_out, i); | |
3426 | vargs[mask_opno] = prepare_load_store_mask | |
3427 | (TREE_TYPE (mask), mask, vargs[mask_opno], gsi); | |
3428 | } | |
3429 | ||
a844293d | 3430 | gcall *call; |
b1b6836e | 3431 | if (ifn != IFN_LAST) |
a844293d | 3432 | call = gimple_build_call_internal_vec (ifn, vargs); |
b1b6836e | 3433 | else |
a844293d RS |
3434 | call = gimple_build_call_vec (fndecl, vargs); |
3435 | new_temp = make_ssa_name (vec_dest, call); | |
3436 | gimple_call_set_lhs (call, new_temp); | |
3437 | gimple_call_set_nothrow (call, true); | |
e1bd7296 RS |
3438 | new_stmt_info |
3439 | = vect_finish_stmt_generation (stmt, call, gsi); | |
b1b6836e | 3440 | } |
e1bd7296 | 3441 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
190c2236 JJ |
3442 | } |
3443 | ||
3444 | for (i = 0; i < nargs; i++) | |
3445 | { | |
37b5ec8f | 3446 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3447 | vec_oprndsi.release (); |
190c2236 | 3448 | } |
190c2236 JJ |
3449 | continue; |
3450 | } | |
3451 | ||
ebfd146a IR |
3452 | for (i = 0; i < nargs; i++) |
3453 | { | |
3454 | op = gimple_call_arg (stmt, i); | |
3455 | if (j == 0) | |
3456 | vec_oprnd0 | |
81c40241 | 3457 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 3458 | else |
2c58d42c RS |
3459 | vec_oprnd0 |
3460 | = vect_get_vec_def_for_stmt_copy (dt[i], orig_vargs[i]); | |
3461 | ||
3462 | orig_vargs[i] = vargs[i] = vec_oprnd0; | |
3463 | } | |
ebfd146a | 3464 | |
2c58d42c RS |
3465 | if (mask_opno >= 0 && masked_loop_p) |
3466 | { | |
3467 | tree mask = vect_get_loop_mask (gsi, masks, ncopies, | |
3468 | vectype_out, j); | |
3469 | vargs[mask_opno] | |
3470 | = prepare_load_store_mask (TREE_TYPE (mask), mask, | |
3471 | vargs[mask_opno], gsi); | |
ebfd146a IR |
3472 | } |
3473 | ||
2c58d42c | 3474 | if (cfn == CFN_GOMP_SIMD_LANE) |
74bf76ed | 3475 | { |
c7bda0f4 | 3476 | tree cst = build_index_vector (vectype_out, j * nunits_out, 1); |
74bf76ed | 3477 | tree new_var |
0e22bb5a | 3478 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 3479 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 3480 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 3481 | new_temp = make_ssa_name (vec_dest); |
e1bd7296 RS |
3482 | gimple *new_stmt = gimple_build_assign (new_temp, new_var); |
3483 | new_stmt_info | |
3484 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
74bf76ed | 3485 | } |
b1b6836e RS |
3486 | else if (modifier == NARROW) |
3487 | { | |
2c58d42c RS |
3488 | /* We don't define any narrowing conditional functions at |
3489 | present. */ | |
3490 | gcc_assert (mask_opno < 0); | |
b1b6836e | 3491 | tree half_res = make_ssa_name (vectype_in); |
a844293d RS |
3492 | gcall *call = gimple_build_call_internal_vec (ifn, vargs); |
3493 | gimple_call_set_lhs (call, half_res); | |
3494 | gimple_call_set_nothrow (call, true); | |
e1bd7296 | 3495 | new_stmt_info = vect_finish_stmt_generation (stmt, call, gsi); |
b1b6836e RS |
3496 | if ((j & 1) == 0) |
3497 | { | |
3498 | prev_res = half_res; | |
3499 | continue; | |
3500 | } | |
3501 | new_temp = make_ssa_name (vec_dest); | |
e1bd7296 RS |
3502 | gassign *new_stmt = gimple_build_assign (new_temp, convert_code, |
3503 | prev_res, half_res); | |
3504 | new_stmt_info | |
3505 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b1b6836e | 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); |
e1bd7296 | 3514 | new_temp = make_ssa_name (vec_dest, call); |
a844293d RS |
3515 | gimple_call_set_lhs (call, new_temp); |
3516 | gimple_call_set_nothrow (call, true); | |
e1bd7296 | 3517 | new_stmt_info = vect_finish_stmt_generation (stmt, call, gsi); |
74bf76ed | 3518 | } |
ebfd146a | 3519 | |
b1b6836e | 3520 | if (j == (modifier == NARROW ? 1 : 0)) |
e1bd7296 | 3521 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
ebfd146a | 3522 | else |
e1bd7296 | 3523 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
ebfd146a | 3524 | |
e1bd7296 | 3525 | prev_stmt_info = new_stmt_info; |
ebfd146a | 3526 | } |
b1b6836e RS |
3527 | } |
3528 | else if (modifier == NARROW) | |
3529 | { | |
2c58d42c RS |
3530 | /* We don't define any narrowing conditional functions at present. */ |
3531 | gcc_assert (mask_opno < 0); | |
ebfd146a IR |
3532 | for (j = 0; j < ncopies; ++j) |
3533 | { | |
3534 | /* Build argument list for the vectorized call. */ | |
3535 | if (j == 0) | |
9771b263 | 3536 | vargs.create (nargs * 2); |
ebfd146a | 3537 | else |
9771b263 | 3538 | vargs.truncate (0); |
ebfd146a | 3539 | |
190c2236 JJ |
3540 | if (slp_node) |
3541 | { | |
ef062b13 | 3542 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 3543 | vec<tree> vec_oprnds0; |
190c2236 JJ |
3544 | |
3545 | for (i = 0; i < nargs; i++) | |
9771b263 | 3546 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 3547 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 3548 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
3549 | |
3550 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3551 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
3552 | { |
3553 | size_t k; | |
9771b263 | 3554 | vargs.truncate (0); |
190c2236 JJ |
3555 | for (k = 0; k < nargs; k++) |
3556 | { | |
37b5ec8f | 3557 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
3558 | vargs.quick_push (vec_oprndsk[i]); |
3559 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 | 3560 | } |
a844293d | 3561 | gcall *call; |
70439f0d | 3562 | if (ifn != IFN_LAST) |
a844293d | 3563 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3564 | else |
a844293d RS |
3565 | call = gimple_build_call_vec (fndecl, vargs); |
3566 | new_temp = make_ssa_name (vec_dest, call); | |
3567 | gimple_call_set_lhs (call, new_temp); | |
3568 | gimple_call_set_nothrow (call, true); | |
e1bd7296 RS |
3569 | new_stmt_info |
3570 | = vect_finish_stmt_generation (stmt, call, gsi); | |
3571 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); | |
190c2236 JJ |
3572 | } |
3573 | ||
3574 | for (i = 0; i < nargs; i++) | |
3575 | { | |
37b5ec8f | 3576 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3577 | vec_oprndsi.release (); |
190c2236 | 3578 | } |
190c2236 JJ |
3579 | continue; |
3580 | } | |
3581 | ||
ebfd146a IR |
3582 | for (i = 0; i < nargs; i++) |
3583 | { | |
3584 | op = gimple_call_arg (stmt, i); | |
3585 | if (j == 0) | |
3586 | { | |
3587 | vec_oprnd0 | |
81c40241 | 3588 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 3589 | vec_oprnd1 |
63827fb8 | 3590 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3591 | } |
3592 | else | |
3593 | { | |
e1bd7296 RS |
3594 | vec_oprnd1 = gimple_call_arg (new_stmt_info->stmt, |
3595 | 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 | ||
e1bd7296 | 3606 | gcall *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); | |
e1bd7296 | 3609 | new_stmt_info = vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
3610 | |
3611 | if (j == 0) | |
e1bd7296 | 3612 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt_info; |
ebfd146a | 3613 | else |
e1bd7296 | 3614 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
ebfd146a | 3615 | |
e1bd7296 | 3616 | prev_stmt_info = new_stmt_info; |
ebfd146a IR |
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)) |
10681ce8 | 3636 | stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); |
ed7b8123 | 3637 | lhs = gimple_get_lhs (stmt_info->stmt); |
3cc2fa2a | 3638 | |
e1bd7296 RS |
3639 | gassign *new_stmt |
3640 | = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
ebfd146a | 3641 | set_vinfo_for_stmt (new_stmt, stmt_info); |
ed7b8123 | 3642 | set_vinfo_for_stmt (stmt_info->stmt, NULL); |
ebfd146a IR |
3643 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
3644 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
3645 | |
3646 | return true; | |
3647 | } | |
3648 | ||
3649 | ||
0136f8f0 AH |
3650 | struct simd_call_arg_info |
3651 | { | |
3652 | tree vectype; | |
3653 | tree op; | |
0136f8f0 | 3654 | HOST_WIDE_INT linear_step; |
34e82342 | 3655 | enum vect_def_type dt; |
0136f8f0 | 3656 | unsigned int align; |
17b658af | 3657 | bool simd_lane_linear; |
0136f8f0 AH |
3658 | }; |
3659 | ||
17b658af JJ |
3660 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
3661 | is linear within simd lane (but not within whole loop), note it in | |
3662 | *ARGINFO. */ | |
3663 | ||
3664 | static void | |
3665 | vect_simd_lane_linear (tree op, struct loop *loop, | |
3666 | struct simd_call_arg_info *arginfo) | |
3667 | { | |
355fe088 | 3668 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
3669 | |
3670 | if (!is_gimple_assign (def_stmt) | |
3671 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
3672 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
3673 | return; | |
3674 | ||
3675 | tree base = gimple_assign_rhs1 (def_stmt); | |
3676 | HOST_WIDE_INT linear_step = 0; | |
3677 | tree v = gimple_assign_rhs2 (def_stmt); | |
3678 | while (TREE_CODE (v) == SSA_NAME) | |
3679 | { | |
3680 | tree t; | |
3681 | def_stmt = SSA_NAME_DEF_STMT (v); | |
3682 | if (is_gimple_assign (def_stmt)) | |
3683 | switch (gimple_assign_rhs_code (def_stmt)) | |
3684 | { | |
3685 | case PLUS_EXPR: | |
3686 | t = gimple_assign_rhs2 (def_stmt); | |
3687 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
3688 | return; | |
3689 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
3690 | v = gimple_assign_rhs1 (def_stmt); | |
3691 | continue; | |
3692 | case MULT_EXPR: | |
3693 | t = gimple_assign_rhs2 (def_stmt); | |
3694 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
3695 | return; | |
3696 | linear_step = tree_to_shwi (t); | |
3697 | v = gimple_assign_rhs1 (def_stmt); | |
3698 | continue; | |
3699 | CASE_CONVERT: | |
3700 | t = gimple_assign_rhs1 (def_stmt); | |
3701 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
3702 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
3703 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
3704 | return; | |
3705 | if (!linear_step) | |
3706 | linear_step = 1; | |
3707 | v = t; | |
3708 | continue; | |
3709 | default: | |
3710 | return; | |
3711 | } | |
8e4284d0 | 3712 | else if (gimple_call_internal_p (def_stmt, IFN_GOMP_SIMD_LANE) |
17b658af JJ |
3713 | && loop->simduid |
3714 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
3715 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
3716 | == loop->simduid)) | |
3717 | { | |
3718 | if (!linear_step) | |
3719 | linear_step = 1; | |
3720 | arginfo->linear_step = linear_step; | |
3721 | arginfo->op = base; | |
3722 | arginfo->simd_lane_linear = true; | |
3723 | return; | |
3724 | } | |
3725 | } | |
3726 | } | |
3727 | ||
cf1b2ba4 RS |
3728 | /* Return the number of elements in vector type VECTYPE, which is associated |
3729 | with a SIMD clone. At present these vectors always have a constant | |
3730 | length. */ | |
3731 | ||
3732 | static unsigned HOST_WIDE_INT | |
3733 | simd_clone_subparts (tree vectype) | |
3734 | { | |
928686b1 | 3735 | return TYPE_VECTOR_SUBPARTS (vectype).to_constant (); |
cf1b2ba4 RS |
3736 | } |
3737 | ||
0136f8f0 AH |
3738 | /* Function vectorizable_simd_clone_call. |
3739 | ||
3740 | Check if STMT performs a function call that can be vectorized | |
3741 | by calling a simd clone of the function. | |
3742 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3743 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3744 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3745 | ||
3746 | static bool | |
355fe088 | 3747 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
1eede195 | 3748 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
68435eb2 | 3749 | stmt_vector_for_cost *) |
0136f8f0 AH |
3750 | { |
3751 | tree vec_dest; | |
3752 | tree scalar_dest; | |
3753 | tree op, type; | |
3754 | tree vec_oprnd0 = NULL_TREE; | |
3755 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
3756 | tree vectype; | |
3757 | unsigned int nunits; | |
3758 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3759 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 3760 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 3761 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 3762 | tree fndecl, new_temp; |
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 | 4115 | bitsize_int ((m & (k - 1)) * prec)); |
e1bd7296 | 4116 | gassign *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); | |
e1bd7296 | 4152 | gassign *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); |
e1bd7296 | 4199 | gassign *new_stmt |
0d0e4a03 | 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)); |
e1bd7296 RS |
4221 | gassign *new_stmt |
4222 | = gimple_build_assign (new_temp, code, | |
4223 | arginfo[i].op, tcst); | |
0136f8f0 AH |
4224 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4225 | vargs.safe_push (new_temp); | |
4226 | } | |
4227 | break; | |
7adb26f2 JJ |
4228 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
4229 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
0136f8f0 | 4230 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
e01d41e5 JJ |
4231 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
4232 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
4233 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
4234 | default: |
4235 | gcc_unreachable (); | |
4236 | } | |
4237 | } | |
4238 | ||
e1bd7296 | 4239 | gcall *new_call = gimple_build_call_vec (fndecl, vargs); |
0136f8f0 AH |
4240 | if (vec_dest) |
4241 | { | |
cf1b2ba4 | 4242 | gcc_assert (ratype || simd_clone_subparts (rtype) == nunits); |
0136f8f0 | 4243 | if (ratype) |
b731b390 | 4244 | new_temp = create_tmp_var (ratype); |
cf1b2ba4 RS |
4245 | else if (simd_clone_subparts (vectype) |
4246 | == simd_clone_subparts (rtype)) | |
e1bd7296 | 4247 | new_temp = make_ssa_name (vec_dest, new_call); |
0136f8f0 | 4248 | else |
e1bd7296 RS |
4249 | new_temp = make_ssa_name (rtype, new_call); |
4250 | gimple_call_set_lhs (new_call, new_temp); | |
0136f8f0 | 4251 | } |
e1bd7296 RS |
4252 | stmt_vec_info new_stmt_info |
4253 | = vect_finish_stmt_generation (stmt, new_call, gsi); | |
0136f8f0 AH |
4254 | |
4255 | if (vec_dest) | |
4256 | { | |
cf1b2ba4 | 4257 | if (simd_clone_subparts (vectype) < nunits) |
0136f8f0 AH |
4258 | { |
4259 | unsigned int k, l; | |
73a699ae RS |
4260 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); |
4261 | poly_uint64 bytes = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
cf1b2ba4 | 4262 | k = nunits / simd_clone_subparts (vectype); |
0136f8f0 AH |
4263 | gcc_assert ((k & (k - 1)) == 0); |
4264 | for (l = 0; l < k; l++) | |
4265 | { | |
4266 | tree t; | |
4267 | if (ratype) | |
4268 | { | |
4269 | t = build_fold_addr_expr (new_temp); | |
4270 | t = build2 (MEM_REF, vectype, t, | |
73a699ae | 4271 | build_int_cst (TREE_TYPE (t), l * bytes)); |
0136f8f0 AH |
4272 | } |
4273 | else | |
4274 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
92e29a5e | 4275 | bitsize_int (prec), bitsize_int (l * prec)); |
e1bd7296 | 4276 | gimple *new_stmt |
b731b390 | 4277 | = gimple_build_assign (make_ssa_name (vectype), t); |
e1bd7296 RS |
4278 | new_stmt_info |
4279 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4280 | ||
0136f8f0 | 4281 | if (j == 0 && l == 0) |
e1bd7296 RS |
4282 | STMT_VINFO_VEC_STMT (stmt_info) |
4283 | = *vec_stmt = new_stmt_info; | |
0136f8f0 | 4284 | else |
e1bd7296 | 4285 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
0136f8f0 | 4286 | |
e1bd7296 | 4287 | prev_stmt_info = new_stmt_info; |
0136f8f0 AH |
4288 | } |
4289 | ||
4290 | if (ratype) | |
3ba4ff41 | 4291 | vect_clobber_variable (stmt, gsi, new_temp); |
0136f8f0 AH |
4292 | continue; |
4293 | } | |
cf1b2ba4 | 4294 | else if (simd_clone_subparts (vectype) > nunits) |
0136f8f0 | 4295 | { |
cf1b2ba4 RS |
4296 | unsigned int k = (simd_clone_subparts (vectype) |
4297 | / simd_clone_subparts (rtype)); | |
0136f8f0 AH |
4298 | gcc_assert ((k & (k - 1)) == 0); |
4299 | if ((j & (k - 1)) == 0) | |
4300 | vec_alloc (ret_ctor_elts, k); | |
4301 | if (ratype) | |
4302 | { | |
cf1b2ba4 | 4303 | unsigned int m, o = nunits / simd_clone_subparts (rtype); |
0136f8f0 AH |
4304 | for (m = 0; m < o; m++) |
4305 | { | |
4306 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
4307 | size_int (m), NULL_TREE, NULL_TREE); | |
e1bd7296 | 4308 | gimple *new_stmt |
b731b390 | 4309 | = gimple_build_assign (make_ssa_name (rtype), tem); |
e1bd7296 RS |
4310 | new_stmt_info |
4311 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
0136f8f0 AH |
4312 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, |
4313 | gimple_assign_lhs (new_stmt)); | |
4314 | } | |
3ba4ff41 | 4315 | vect_clobber_variable (stmt, gsi, new_temp); |
0136f8f0 AH |
4316 | } |
4317 | else | |
4318 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
4319 | if ((j & (k - 1)) != k - 1) | |
4320 | continue; | |
4321 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
e1bd7296 | 4322 | gimple *new_stmt |
b731b390 | 4323 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
e1bd7296 RS |
4324 | new_stmt_info |
4325 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
0136f8f0 AH |
4326 | |
4327 | if ((unsigned) j == k - 1) | |
e1bd7296 | 4328 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
0136f8f0 | 4329 | else |
e1bd7296 | 4330 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
0136f8f0 | 4331 | |
e1bd7296 | 4332 | prev_stmt_info = new_stmt_info; |
0136f8f0 AH |
4333 | continue; |
4334 | } | |
4335 | else if (ratype) | |
4336 | { | |
4337 | tree t = build_fold_addr_expr (new_temp); | |
4338 | t = build2 (MEM_REF, vectype, t, | |
4339 | build_int_cst (TREE_TYPE (t), 0)); | |
e1bd7296 | 4340 | gimple *new_stmt |
b731b390 | 4341 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
e1bd7296 RS |
4342 | new_stmt_info |
4343 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3ba4ff41 | 4344 | vect_clobber_variable (stmt, gsi, new_temp); |
0136f8f0 AH |
4345 | } |
4346 | } | |
4347 | ||
4348 | if (j == 0) | |
e1bd7296 | 4349 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
0136f8f0 | 4350 | else |
e1bd7296 | 4351 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
0136f8f0 | 4352 | |
e1bd7296 | 4353 | prev_stmt_info = new_stmt_info; |
0136f8f0 AH |
4354 | } |
4355 | ||
4356 | vargs.release (); | |
4357 | ||
4358 | /* The call in STMT might prevent it from being removed in dce. | |
4359 | We however cannot remove it here, due to the way the ssa name | |
4360 | it defines is mapped to the new definition. So just replace | |
4361 | rhs of the statement with something harmless. */ | |
4362 | ||
4363 | if (slp_node) | |
4364 | return true; | |
4365 | ||
e1bd7296 | 4366 | gimple *new_stmt; |
0136f8f0 AH |
4367 | if (scalar_dest) |
4368 | { | |
4369 | type = TREE_TYPE (scalar_dest); | |
4370 | if (is_pattern_stmt_p (stmt_info)) | |
10681ce8 | 4371 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)->stmt); |
0136f8f0 AH |
4372 | else |
4373 | lhs = gimple_call_lhs (stmt); | |
4374 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
4375 | } | |
4376 | else | |
4377 | new_stmt = gimple_build_nop (); | |
4378 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
4379 | set_vinfo_for_stmt (stmt, NULL); | |
4380 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 4381 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
4382 | unlink_stmt_vdef (stmt); |
4383 | ||
4384 | return true; | |
4385 | } | |
4386 | ||
4387 | ||
ebfd146a IR |
4388 | /* Function vect_gen_widened_results_half |
4389 | ||
4390 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 4391 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 4392 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
4393 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
4394 | needs to be created (DECL is a function-decl of a target-builtin). | |
4395 | STMT is the original scalar stmt that we are vectorizing. */ | |
4396 | ||
355fe088 | 4397 | static gimple * |
ebfd146a IR |
4398 | vect_gen_widened_results_half (enum tree_code code, |
4399 | tree decl, | |
4400 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
4401 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 4402 | gimple *stmt) |
b8698a0f | 4403 | { |
355fe088 | 4404 | gimple *new_stmt; |
b8698a0f L |
4405 | tree new_temp; |
4406 | ||
4407 | /* Generate half of the widened result: */ | |
4408 | if (code == CALL_EXPR) | |
4409 | { | |
4410 | /* Target specific support */ | |
ebfd146a IR |
4411 | if (op_type == binary_op) |
4412 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
4413 | else | |
4414 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
4415 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4416 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
4417 | } |
4418 | else | |
ebfd146a | 4419 | { |
b8698a0f L |
4420 | /* Generic support */ |
4421 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
4422 | if (op_type != binary_op) |
4423 | vec_oprnd1 = NULL; | |
0d0e4a03 | 4424 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
4425 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4426 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 4427 | } |
ebfd146a IR |
4428 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4429 | ||
ebfd146a IR |
4430 | return new_stmt; |
4431 | } | |
4432 | ||
4a00c761 JJ |
4433 | |
4434 | /* Get vectorized definitions for loop-based vectorization. For the first | |
4435 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
4436 | scalar operand), and for the rest we get a copy with | |
4437 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
4438 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
4439 | The vectors are collected into VEC_OPRNDS. */ | |
4440 | ||
4441 | static void | |
355fe088 | 4442 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 4443 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
4444 | { |
4445 | tree vec_oprnd; | |
4446 | ||
4447 | /* Get first vector operand. */ | |
4448 | /* All the vector operands except the very first one (that is scalar oprnd) | |
4449 | are stmt copies. */ | |
4450 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 4451 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
4452 | else |
4453 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
4454 | ||
9771b263 | 4455 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
4456 | |
4457 | /* Get second vector operand. */ | |
4458 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 4459 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
4460 | |
4461 | *oprnd = vec_oprnd; | |
4462 | ||
4463 | /* For conversion in multiple steps, continue to get operands | |
4464 | recursively. */ | |
4465 | if (multi_step_cvt) | |
4466 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
4467 | } | |
4468 | ||
4469 | ||
4470 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
4471 | For multi-step conversions store the resulting vectors and call the function | |
4472 | recursively. */ | |
4473 | ||
4474 | static void | |
9771b263 | 4475 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 4476 | int multi_step_cvt, gimple *stmt, |
9771b263 | 4477 | vec<tree> vec_dsts, |
4a00c761 JJ |
4478 | gimple_stmt_iterator *gsi, |
4479 | slp_tree slp_node, enum tree_code code, | |
4480 | stmt_vec_info *prev_stmt_info) | |
4481 | { | |
4482 | unsigned int i; | |
4483 | tree vop0, vop1, new_tmp, vec_dest; | |
4a00c761 JJ |
4484 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
4485 | ||
9771b263 | 4486 | vec_dest = vec_dsts.pop (); |
4a00c761 | 4487 | |
9771b263 | 4488 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
4489 | { |
4490 | /* Create demotion operation. */ | |
9771b263 DN |
4491 | vop0 = (*vec_oprnds)[i]; |
4492 | vop1 = (*vec_oprnds)[i + 1]; | |
e1bd7296 | 4493 | gassign *new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
4494 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
4495 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
e1bd7296 RS |
4496 | stmt_vec_info new_stmt_info |
4497 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4a00c761 JJ |
4498 | |
4499 | if (multi_step_cvt) | |
4500 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 4501 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
4502 | else |
4503 | { | |
4504 | /* This is the last step of the conversion sequence. Store the | |
4505 | vectors in SLP_NODE or in vector info of the scalar statement | |
4506 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
4507 | if (slp_node) | |
e1bd7296 | 4508 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
4a00c761 | 4509 | else |
c689ce1e RB |
4510 | { |
4511 | if (!*prev_stmt_info) | |
e1bd7296 | 4512 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt_info; |
c689ce1e | 4513 | else |
e1bd7296 | 4514 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt_info; |
4a00c761 | 4515 | |
e1bd7296 | 4516 | *prev_stmt_info = new_stmt_info; |
c689ce1e | 4517 | } |
4a00c761 JJ |
4518 | } |
4519 | } | |
4520 | ||
4521 | /* For multi-step demotion operations we first generate demotion operations | |
4522 | from the source type to the intermediate types, and then combine the | |
4523 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
4524 | type. */ | |
4525 | if (multi_step_cvt) | |
4526 | { | |
4527 | /* At each level of recursion we have half of the operands we had at the | |
4528 | previous level. */ | |
9771b263 | 4529 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
4530 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
4531 | stmt, vec_dsts, gsi, slp_node, | |
4532 | VEC_PACK_TRUNC_EXPR, | |
4533 | prev_stmt_info); | |
4534 | } | |
4535 | ||
9771b263 | 4536 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4537 | } |
4538 | ||
4539 | ||
4540 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
4541 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
4542 | the resulting vectors and call the function recursively. */ | |
4543 | ||
4544 | static void | |
9771b263 DN |
4545 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
4546 | vec<tree> *vec_oprnds1, | |
355fe088 | 4547 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
4548 | gimple_stmt_iterator *gsi, |
4549 | enum tree_code code1, | |
4550 | enum tree_code code2, tree decl1, | |
4551 | tree decl2, int op_type) | |
4552 | { | |
4553 | int i; | |
4554 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 4555 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 4556 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 4557 | |
9771b263 DN |
4558 | vec_tmp.create (vec_oprnds0->length () * 2); |
4559 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
4560 | { |
4561 | if (op_type == binary_op) | |
9771b263 | 4562 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
4563 | else |
4564 | vop1 = NULL_TREE; | |
4565 | ||
4566 | /* Generate the two halves of promotion operation. */ | |
4567 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
4568 | op_type, vec_dest, gsi, stmt); | |
4569 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
4570 | op_type, vec_dest, gsi, stmt); | |
4571 | if (is_gimple_call (new_stmt1)) | |
4572 | { | |
4573 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
4574 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
4575 | } | |
4576 | else | |
4577 | { | |
4578 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
4579 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
4580 | } | |
4581 | ||
4582 | /* Store the results for the next step. */ | |
9771b263 DN |
4583 | vec_tmp.quick_push (new_tmp1); |
4584 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
4585 | } |
4586 | ||
689eaba3 | 4587 | vec_oprnds0->release (); |
4a00c761 JJ |
4588 | *vec_oprnds0 = vec_tmp; |
4589 | } | |
4590 | ||
4591 | ||
b8698a0f L |
4592 | /* Check if STMT performs a conversion operation, that can be vectorized. |
4593 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 4594 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
4595 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
4596 | ||
4597 | static bool | |
355fe088 | 4598 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
1eede195 | 4599 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
68435eb2 | 4600 | stmt_vector_for_cost *cost_vec) |
ebfd146a IR |
4601 | { |
4602 | tree vec_dest; | |
4603 | tree scalar_dest; | |
4a00c761 | 4604 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
4605 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
4606 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4607 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4608 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 4609 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
4610 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
4611 | tree new_temp; | |
ebfd146a | 4612 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 4613 | int ndts = 2; |
ebfd146a | 4614 | stmt_vec_info prev_stmt_info; |
062d5ccc RS |
4615 | poly_uint64 nunits_in; |
4616 | poly_uint64 nunits_out; | |
ebfd146a | 4617 | tree vectype_out, vectype_in; |
4a00c761 JJ |
4618 | int ncopies, i, j; |
4619 | tree lhs_type, rhs_type; | |
ebfd146a | 4620 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
4621 | vec<tree> vec_oprnds0 = vNULL; |
4622 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 4623 | tree vop0; |
4a00c761 | 4624 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4625 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 4626 | int multi_step_cvt = 0; |
6e1aa848 | 4627 | vec<tree> interm_types = vNULL; |
4a00c761 JJ |
4628 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
4629 | int op_type; | |
4a00c761 | 4630 | unsigned short fltsz; |
ebfd146a IR |
4631 | |
4632 | /* Is STMT a vectorizable conversion? */ | |
4633 | ||
4a00c761 | 4634 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4635 | return false; |
4636 | ||
66c16fd9 RB |
4637 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4638 | && ! vec_stmt) | |
ebfd146a IR |
4639 | return false; |
4640 | ||
4641 | if (!is_gimple_assign (stmt)) | |
4642 | return false; | |
4643 | ||
4644 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4645 | return false; | |
4646 | ||
4647 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
4648 | if (!CONVERT_EXPR_CODE_P (code) |
4649 | && code != FIX_TRUNC_EXPR | |
4650 | && code != FLOAT_EXPR | |
4651 | && code != WIDEN_MULT_EXPR | |
4652 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
4653 | return false; |
4654 | ||
4a00c761 JJ |
4655 | op_type = TREE_CODE_LENGTH (code); |
4656 | ||
ebfd146a | 4657 | /* Check types of lhs and rhs. */ |
b690cc0f | 4658 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 4659 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
4660 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
4661 | ||
ebfd146a IR |
4662 | op0 = gimple_assign_rhs1 (stmt); |
4663 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
4664 | |
4665 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4666 | && !((INTEGRAL_TYPE_P (lhs_type) | |
4667 | && INTEGRAL_TYPE_P (rhs_type)) | |
4668 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
4669 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
4670 | return false; | |
4671 | ||
e6f5c25d IE |
4672 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4673 | && ((INTEGRAL_TYPE_P (lhs_type) | |
2be65d9e | 4674 | && !type_has_mode_precision_p (lhs_type)) |
e6f5c25d | 4675 | || (INTEGRAL_TYPE_P (rhs_type) |
2be65d9e | 4676 | && !type_has_mode_precision_p (rhs_type)))) |
4a00c761 | 4677 | { |
73fbfcad | 4678 | if (dump_enabled_p ()) |
78c60e3d | 4679 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
4680 | "type conversion to/from bit-precision unsupported." |
4681 | "\n"); | |
4a00c761 JJ |
4682 | return false; |
4683 | } | |
4684 | ||
b690cc0f | 4685 | /* Check the operands of the operation. */ |
894dd753 | 4686 | if (!vect_is_simple_use (op0, vinfo, &dt[0], &vectype_in)) |
b690cc0f | 4687 | { |
73fbfcad | 4688 | if (dump_enabled_p ()) |
78c60e3d | 4689 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4690 | "use not simple.\n"); |
b690cc0f RG |
4691 | return false; |
4692 | } | |
4a00c761 JJ |
4693 | if (op_type == binary_op) |
4694 | { | |
4695 | bool ok; | |
4696 | ||
4697 | op1 = gimple_assign_rhs2 (stmt); | |
4698 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
4699 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
4700 | OP1. */ | |
4701 | if (CONSTANT_CLASS_P (op0)) | |
894dd753 | 4702 | ok = vect_is_simple_use (op1, vinfo, &dt[1], &vectype_in); |
4a00c761 | 4703 | else |
894dd753 | 4704 | ok = vect_is_simple_use (op1, vinfo, &dt[1]); |
4a00c761 JJ |
4705 | |
4706 | if (!ok) | |
4707 | { | |
73fbfcad | 4708 | if (dump_enabled_p ()) |
78c60e3d | 4709 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4710 | "use not simple.\n"); |
4a00c761 JJ |
4711 | return false; |
4712 | } | |
4713 | } | |
4714 | ||
b690cc0f RG |
4715 | /* If op0 is an external or constant defs use a vector type of |
4716 | the same size as the output vector type. */ | |
ebfd146a | 4717 | if (!vectype_in) |
b690cc0f | 4718 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
4719 | if (vec_stmt) |
4720 | gcc_assert (vectype_in); | |
4721 | if (!vectype_in) | |
4722 | { | |
73fbfcad | 4723 | if (dump_enabled_p ()) |
4a00c761 | 4724 | { |
78c60e3d SS |
4725 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4726 | "no vectype for scalar type "); | |
4727 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 4728 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 4729 | } |
7d8930a0 IR |
4730 | |
4731 | return false; | |
4732 | } | |
ebfd146a | 4733 | |
e6f5c25d IE |
4734 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4735 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
4736 | { | |
4737 | if (dump_enabled_p ()) | |
4738 | { | |
4739 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4740 | "can't convert between boolean and non " | |
4741 | "boolean vectors"); | |
4742 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
4743 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
4744 | } | |
4745 | ||
4746 | return false; | |
4747 | } | |
4748 | ||
b690cc0f RG |
4749 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
4750 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
062d5ccc | 4751 | if (known_eq (nunits_out, nunits_in)) |
ebfd146a | 4752 | modifier = NONE; |
062d5ccc RS |
4753 | else if (multiple_p (nunits_out, nunits_in)) |
4754 | modifier = NARROW; | |
ebfd146a | 4755 | else |
062d5ccc RS |
4756 | { |
4757 | gcc_checking_assert (multiple_p (nunits_in, nunits_out)); | |
4758 | modifier = WIDEN; | |
4759 | } | |
ebfd146a | 4760 | |
ff802fa1 IR |
4761 | /* Multiple types in SLP are handled by creating the appropriate number of |
4762 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4763 | case of SLP. */ | |
fce57248 | 4764 | if (slp_node) |
ebfd146a | 4765 | ncopies = 1; |
4a00c761 | 4766 | else if (modifier == NARROW) |
e8f142e2 | 4767 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
4a00c761 | 4768 | else |
e8f142e2 | 4769 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
b8698a0f | 4770 | |
ebfd146a IR |
4771 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
4772 | needs to be generated. */ | |
4773 | gcc_assert (ncopies >= 1); | |
4774 | ||
16d22000 RS |
4775 | bool found_mode = false; |
4776 | scalar_mode lhs_mode = SCALAR_TYPE_MODE (lhs_type); | |
4777 | scalar_mode rhs_mode = SCALAR_TYPE_MODE (rhs_type); | |
4778 | opt_scalar_mode rhs_mode_iter; | |
b397965c | 4779 | |
ebfd146a | 4780 | /* Supportable by target? */ |
4a00c761 | 4781 | switch (modifier) |
ebfd146a | 4782 | { |
4a00c761 JJ |
4783 | case NONE: |
4784 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4785 | return false; | |
4786 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
4787 | &decl1, &code1)) | |
4788 | break; | |
4789 | /* FALLTHRU */ | |
4790 | unsupported: | |
73fbfcad | 4791 | if (dump_enabled_p ()) |
78c60e3d | 4792 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4793 | "conversion not supported by target.\n"); |
ebfd146a | 4794 | return false; |
ebfd146a | 4795 | |
4a00c761 JJ |
4796 | case WIDEN: |
4797 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
4798 | &code1, &code2, &multi_step_cvt, |
4799 | &interm_types)) | |
4a00c761 JJ |
4800 | { |
4801 | /* Binary widening operation can only be supported directly by the | |
4802 | architecture. */ | |
4803 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
4804 | break; | |
4805 | } | |
4806 | ||
4807 | if (code != FLOAT_EXPR | |
b397965c | 4808 | || GET_MODE_SIZE (lhs_mode) <= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4809 | goto unsupported; |
4810 | ||
b397965c | 4811 | fltsz = GET_MODE_SIZE (lhs_mode); |
16d22000 | 4812 | FOR_EACH_2XWIDER_MODE (rhs_mode_iter, rhs_mode) |
4a00c761 | 4813 | { |
16d22000 | 4814 | rhs_mode = rhs_mode_iter.require (); |
c94843d2 RS |
4815 | if (GET_MODE_SIZE (rhs_mode) > fltsz) |
4816 | break; | |
4817 | ||
4a00c761 JJ |
4818 | cvt_type |
4819 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4820 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4821 | if (cvt_type == NULL_TREE) | |
4822 | goto unsupported; | |
4823 | ||
4824 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4825 | { | |
4826 | if (!supportable_convert_operation (code, vectype_out, | |
4827 | cvt_type, &decl1, &codecvt1)) | |
4828 | goto unsupported; | |
4829 | } | |
4830 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
4831 | cvt_type, &codecvt1, |
4832 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
4833 | &interm_types)) |
4834 | continue; | |
4835 | else | |
4836 | gcc_assert (multi_step_cvt == 0); | |
4837 | ||
4838 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
4839 | vectype_in, &code1, &code2, |
4840 | &multi_step_cvt, &interm_types)) | |
16d22000 RS |
4841 | { |
4842 | found_mode = true; | |
4843 | break; | |
4844 | } | |
4a00c761 JJ |
4845 | } |
4846 | ||
16d22000 | 4847 | if (!found_mode) |
4a00c761 JJ |
4848 | goto unsupported; |
4849 | ||
4850 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4851 | codecvt2 = ERROR_MARK; | |
4852 | else | |
4853 | { | |
4854 | multi_step_cvt++; | |
9771b263 | 4855 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
4856 | cvt_type = NULL_TREE; |
4857 | } | |
4858 | break; | |
4859 | ||
4860 | case NARROW: | |
4861 | gcc_assert (op_type == unary_op); | |
4862 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
4863 | &code1, &multi_step_cvt, | |
4864 | &interm_types)) | |
4865 | break; | |
4866 | ||
4867 | if (code != FIX_TRUNC_EXPR | |
b397965c | 4868 | || GET_MODE_SIZE (lhs_mode) >= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4869 | goto unsupported; |
4870 | ||
4a00c761 JJ |
4871 | cvt_type |
4872 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4873 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4874 | if (cvt_type == NULL_TREE) | |
4875 | goto unsupported; | |
4876 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
4877 | &decl1, &codecvt1)) | |
4878 | goto unsupported; | |
4879 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
4880 | &code1, &multi_step_cvt, | |
4881 | &interm_types)) | |
4882 | break; | |
4883 | goto unsupported; | |
4884 | ||
4885 | default: | |
4886 | gcc_unreachable (); | |
ebfd146a IR |
4887 | } |
4888 | ||
4889 | if (!vec_stmt) /* transformation not required. */ | |
4890 | { | |
adac3a68 | 4891 | DUMP_VECT_SCOPE ("vectorizable_conversion"); |
4a00c761 | 4892 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
4893 | { |
4894 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
68435eb2 RB |
4895 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, |
4896 | cost_vec); | |
8bd37302 | 4897 | } |
4a00c761 JJ |
4898 | else if (modifier == NARROW) |
4899 | { | |
4900 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
68435eb2 RB |
4901 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt, |
4902 | cost_vec); | |
4a00c761 JJ |
4903 | } |
4904 | else | |
4905 | { | |
4906 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
68435eb2 RB |
4907 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt, |
4908 | cost_vec); | |
4a00c761 | 4909 | } |
9771b263 | 4910 | interm_types.release (); |
ebfd146a IR |
4911 | return true; |
4912 | } | |
4913 | ||
67b8dbac | 4914 | /* Transform. */ |
73fbfcad | 4915 | if (dump_enabled_p ()) |
78c60e3d | 4916 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4917 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 4918 | |
4a00c761 JJ |
4919 | if (op_type == binary_op) |
4920 | { | |
4921 | if (CONSTANT_CLASS_P (op0)) | |
4922 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
4923 | else if (CONSTANT_CLASS_P (op1)) | |
4924 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
4925 | } | |
4926 | ||
4927 | /* In case of multi-step conversion, we first generate conversion operations | |
4928 | to the intermediate types, and then from that types to the final one. | |
4929 | We create vector destinations for the intermediate type (TYPES) received | |
4930 | from supportable_*_operation, and store them in the correct order | |
4931 | for future use in vect_create_vectorized_*_stmts (). */ | |
8c681247 | 4932 | auto_vec<tree> vec_dsts (multi_step_cvt + 1); |
82294ec1 JJ |
4933 | vec_dest = vect_create_destination_var (scalar_dest, |
4934 | (cvt_type && modifier == WIDEN) | |
4935 | ? cvt_type : vectype_out); | |
9771b263 | 4936 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4937 | |
4938 | if (multi_step_cvt) | |
4939 | { | |
9771b263 DN |
4940 | for (i = interm_types.length () - 1; |
4941 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
4942 | { |
4943 | vec_dest = vect_create_destination_var (scalar_dest, | |
4944 | intermediate_type); | |
9771b263 | 4945 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4946 | } |
4947 | } | |
ebfd146a | 4948 | |
4a00c761 | 4949 | if (cvt_type) |
82294ec1 JJ |
4950 | vec_dest = vect_create_destination_var (scalar_dest, |
4951 | modifier == WIDEN | |
4952 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
4953 | |
4954 | if (!slp_node) | |
4955 | { | |
30862efc | 4956 | if (modifier == WIDEN) |
4a00c761 | 4957 | { |
c3284718 | 4958 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 4959 | if (op_type == binary_op) |
9771b263 | 4960 | vec_oprnds1.create (1); |
4a00c761 | 4961 | } |
30862efc | 4962 | else if (modifier == NARROW) |
9771b263 DN |
4963 | vec_oprnds0.create ( |
4964 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
4965 | } |
4966 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 4967 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 4968 | |
4a00c761 | 4969 | last_oprnd = op0; |
ebfd146a IR |
4970 | prev_stmt_info = NULL; |
4971 | switch (modifier) | |
4972 | { | |
4973 | case NONE: | |
4974 | for (j = 0; j < ncopies; j++) | |
4975 | { | |
ebfd146a | 4976 | if (j == 0) |
306b0c92 | 4977 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node); |
ebfd146a IR |
4978 | else |
4979 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
4980 | ||
9771b263 | 4981 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 | 4982 | { |
e1bd7296 | 4983 | stmt_vec_info new_stmt_info; |
4a00c761 JJ |
4984 | /* Arguments are ready, create the new vector stmt. */ |
4985 | if (code1 == CALL_EXPR) | |
4986 | { | |
e1bd7296 | 4987 | gcall *new_stmt = gimple_build_call (decl1, 1, vop0); |
4a00c761 JJ |
4988 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4989 | gimple_call_set_lhs (new_stmt, new_temp); | |
e1bd7296 RS |
4990 | new_stmt_info |
4991 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4a00c761 JJ |
4992 | } |
4993 | else | |
4994 | { | |
4995 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
e1bd7296 RS |
4996 | gassign *new_stmt |
4997 | = gimple_build_assign (vec_dest, code1, vop0); | |
4a00c761 JJ |
4998 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4999 | gimple_assign_set_lhs (new_stmt, new_temp); | |
e1bd7296 RS |
5000 | new_stmt_info |
5001 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4a00c761 JJ |
5002 | } |
5003 | ||
4a00c761 | 5004 | if (slp_node) |
e1bd7296 | 5005 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
225ce44b RB |
5006 | else |
5007 | { | |
5008 | if (!prev_stmt_info) | |
e1bd7296 RS |
5009 | STMT_VINFO_VEC_STMT (stmt_info) |
5010 | = *vec_stmt = new_stmt_info; | |
225ce44b | 5011 | else |
e1bd7296 RS |
5012 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
5013 | prev_stmt_info = new_stmt_info; | |
225ce44b | 5014 | } |
4a00c761 | 5015 | } |
ebfd146a IR |
5016 | } |
5017 | break; | |
5018 | ||
5019 | case WIDEN: | |
5020 | /* In case the vectorization factor (VF) is bigger than the number | |
5021 | of elements that we can fit in a vectype (nunits), we have to | |
5022 | generate more than one vector stmt - i.e - we need to "unroll" | |
5023 | the vector stmt by a factor VF/nunits. */ | |
5024 | for (j = 0; j < ncopies; j++) | |
5025 | { | |
4a00c761 | 5026 | /* Handle uses. */ |
ebfd146a | 5027 | if (j == 0) |
4a00c761 JJ |
5028 | { |
5029 | if (slp_node) | |
5030 | { | |
5031 | if (code == WIDEN_LSHIFT_EXPR) | |
5032 | { | |
5033 | unsigned int k; | |
ebfd146a | 5034 | |
4a00c761 JJ |
5035 | vec_oprnd1 = op1; |
5036 | /* Store vec_oprnd1 for every vector stmt to be created | |
5037 | for SLP_NODE. We check during the analysis that all | |
5038 | the shift arguments are the same. */ | |
5039 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 5040 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
5041 | |
5042 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5043 | slp_node); |
4a00c761 JJ |
5044 | } |
5045 | else | |
5046 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
306b0c92 | 5047 | &vec_oprnds1, slp_node); |
4a00c761 JJ |
5048 | } |
5049 | else | |
5050 | { | |
81c40241 | 5051 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 5052 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
5053 | if (op_type == binary_op) |
5054 | { | |
5055 | if (code == WIDEN_LSHIFT_EXPR) | |
5056 | vec_oprnd1 = op1; | |
5057 | else | |
81c40241 | 5058 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 5059 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
5060 | } |
5061 | } | |
5062 | } | |
ebfd146a | 5063 | else |
4a00c761 JJ |
5064 | { |
5065 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
5066 | vec_oprnds0.truncate (0); |
5067 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
5068 | if (op_type == binary_op) |
5069 | { | |
5070 | if (code == WIDEN_LSHIFT_EXPR) | |
5071 | vec_oprnd1 = op1; | |
5072 | else | |
5073 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
5074 | vec_oprnd1); | |
9771b263 DN |
5075 | vec_oprnds1.truncate (0); |
5076 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
5077 | } |
5078 | } | |
ebfd146a | 5079 | |
4a00c761 JJ |
5080 | /* Arguments are ready. Create the new vector stmts. */ |
5081 | for (i = multi_step_cvt; i >= 0; i--) | |
5082 | { | |
9771b263 | 5083 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
5084 | enum tree_code c1 = code1, c2 = code2; |
5085 | if (i == 0 && codecvt2 != ERROR_MARK) | |
5086 | { | |
5087 | c1 = codecvt1; | |
5088 | c2 = codecvt2; | |
5089 | } | |
5090 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
5091 | &vec_oprnds1, | |
5092 | stmt, this_dest, gsi, | |
5093 | c1, c2, decl1, decl2, | |
5094 | op_type); | |
5095 | } | |
5096 | ||
9771b263 | 5097 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 | 5098 | { |
e1bd7296 | 5099 | stmt_vec_info new_stmt_info; |
4a00c761 JJ |
5100 | if (cvt_type) |
5101 | { | |
5102 | if (codecvt1 == CALL_EXPR) | |
5103 | { | |
e1bd7296 | 5104 | gcall *new_stmt = gimple_build_call (decl1, 1, vop0); |
4a00c761 JJ |
5105 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5106 | gimple_call_set_lhs (new_stmt, new_temp); | |
e1bd7296 RS |
5107 | new_stmt_info |
5108 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4a00c761 JJ |
5109 | } |
5110 | else | |
5111 | { | |
5112 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 5113 | new_temp = make_ssa_name (vec_dest); |
e1bd7296 RS |
5114 | gassign *new_stmt |
5115 | = gimple_build_assign (new_temp, codecvt1, vop0); | |
5116 | new_stmt_info | |
5117 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4a00c761 | 5118 | } |
4a00c761 JJ |
5119 | } |
5120 | else | |
e1bd7296 | 5121 | new_stmt_info = vinfo->lookup_def (vop0); |
4a00c761 JJ |
5122 | |
5123 | if (slp_node) | |
e1bd7296 | 5124 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
4a00c761 | 5125 | else |
c689ce1e RB |
5126 | { |
5127 | if (!prev_stmt_info) | |
e1bd7296 | 5128 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt_info; |
c689ce1e | 5129 | else |
e1bd7296 RS |
5130 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
5131 | prev_stmt_info = new_stmt_info; | |
c689ce1e | 5132 | } |
4a00c761 | 5133 | } |
ebfd146a | 5134 | } |
4a00c761 JJ |
5135 | |
5136 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
5137 | break; |
5138 | ||
5139 | case NARROW: | |
5140 | /* In case the vectorization factor (VF) is bigger than the number | |
5141 | of elements that we can fit in a vectype (nunits), we have to | |
5142 | generate more than one vector stmt - i.e - we need to "unroll" | |
5143 | the vector stmt by a factor VF/nunits. */ | |
5144 | for (j = 0; j < ncopies; j++) | |
5145 | { | |
5146 | /* Handle uses. */ | |
4a00c761 JJ |
5147 | if (slp_node) |
5148 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5149 | slp_node); |
ebfd146a IR |
5150 | else |
5151 | { | |
9771b263 | 5152 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
5153 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
5154 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
5155 | } |
5156 | ||
4a00c761 JJ |
5157 | /* Arguments are ready. Create the new vector stmts. */ |
5158 | if (cvt_type) | |
9771b263 | 5159 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
5160 | { |
5161 | if (codecvt1 == CALL_EXPR) | |
5162 | { | |
e1bd7296 | 5163 | gcall *new_stmt = gimple_build_call (decl1, 1, vop0); |
4a00c761 JJ |
5164 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5165 | gimple_call_set_lhs (new_stmt, new_temp); | |
e1bd7296 | 5166 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4a00c761 JJ |
5167 | } |
5168 | else | |
5169 | { | |
5170 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 5171 | new_temp = make_ssa_name (vec_dest); |
e1bd7296 RS |
5172 | gassign *new_stmt |
5173 | = gimple_build_assign (new_temp, codecvt1, vop0); | |
5174 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4a00c761 | 5175 | } |
ebfd146a | 5176 | |
9771b263 | 5177 | vec_oprnds0[i] = new_temp; |
4a00c761 | 5178 | } |
ebfd146a | 5179 | |
4a00c761 JJ |
5180 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
5181 | stmt, vec_dsts, gsi, | |
5182 | slp_node, code1, | |
5183 | &prev_stmt_info); | |
ebfd146a IR |
5184 | } |
5185 | ||
5186 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 5187 | break; |
ebfd146a IR |
5188 | } |
5189 | ||
9771b263 DN |
5190 | vec_oprnds0.release (); |
5191 | vec_oprnds1.release (); | |
9771b263 | 5192 | interm_types.release (); |
ebfd146a IR |
5193 | |
5194 | return true; | |
5195 | } | |
ff802fa1 IR |
5196 | |
5197 | ||
ebfd146a IR |
5198 | /* Function vectorizable_assignment. |
5199 | ||
b8698a0f L |
5200 | Check if STMT performs an assignment (copy) that can be vectorized. |
5201 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5202 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5203 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5204 | ||
5205 | static bool | |
355fe088 | 5206 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
1eede195 | 5207 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
68435eb2 | 5208 | stmt_vector_for_cost *cost_vec) |
ebfd146a IR |
5209 | { |
5210 | tree vec_dest; | |
5211 | tree scalar_dest; | |
5212 | tree op; | |
5213 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
5214 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
5215 | tree new_temp; | |
4fc5ebf1 JG |
5216 | enum vect_def_type dt[1] = {vect_unknown_def_type}; |
5217 | int ndts = 1; | |
ebfd146a | 5218 | int ncopies; |
f18b55bd | 5219 | int i, j; |
6e1aa848 | 5220 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5221 | tree vop; |
a70d6342 | 5222 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5223 | vec_info *vinfo = stmt_info->vinfo; |
f18b55bd | 5224 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
5225 | enum tree_code code; |
5226 | tree vectype_in; | |
ebfd146a | 5227 | |
a70d6342 | 5228 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5229 | return false; |
5230 | ||
66c16fd9 RB |
5231 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5232 | && ! vec_stmt) | |
ebfd146a IR |
5233 | return false; |
5234 | ||
5235 | /* Is vectorizable assignment? */ | |
5236 | if (!is_gimple_assign (stmt)) | |
5237 | return false; | |
5238 | ||
5239 | scalar_dest = gimple_assign_lhs (stmt); | |
5240 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
5241 | return false; | |
5242 | ||
fde9c428 | 5243 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 5244 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
5245 | || code == PAREN_EXPR |
5246 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
5247 | op = gimple_assign_rhs1 (stmt); |
5248 | else | |
5249 | return false; | |
5250 | ||
7b7ec6c5 RG |
5251 | if (code == VIEW_CONVERT_EXPR) |
5252 | op = TREE_OPERAND (op, 0); | |
5253 | ||
465c8c19 | 5254 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
928686b1 | 5255 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 JJ |
5256 | |
5257 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5258 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5259 | case of SLP. */ | |
fce57248 | 5260 | if (slp_node) |
465c8c19 JJ |
5261 | ncopies = 1; |
5262 | else | |
e8f142e2 | 5263 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
5264 | |
5265 | gcc_assert (ncopies >= 1); | |
5266 | ||
894dd753 | 5267 | if (!vect_is_simple_use (op, vinfo, &dt[0], &vectype_in)) |
ebfd146a | 5268 | { |
73fbfcad | 5269 | if (dump_enabled_p ()) |
78c60e3d | 5270 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5271 | "use not simple.\n"); |
ebfd146a IR |
5272 | return false; |
5273 | } | |
5274 | ||
fde9c428 RG |
5275 | /* We can handle NOP_EXPR conversions that do not change the number |
5276 | of elements or the vector size. */ | |
7b7ec6c5 RG |
5277 | if ((CONVERT_EXPR_CODE_P (code) |
5278 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 | 5279 | && (!vectype_in |
928686b1 | 5280 | || maybe_ne (TYPE_VECTOR_SUBPARTS (vectype_in), nunits) |
cf098191 RS |
5281 | || maybe_ne (GET_MODE_SIZE (TYPE_MODE (vectype)), |
5282 | GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
fde9c428 RG |
5283 | return false; |
5284 | ||
7b7b1813 RG |
5285 | /* We do not handle bit-precision changes. */ |
5286 | if ((CONVERT_EXPR_CODE_P (code) | |
5287 | || code == VIEW_CONVERT_EXPR) | |
5288 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2be65d9e RS |
5289 | && (!type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
5290 | || !type_has_mode_precision_p (TREE_TYPE (op))) | |
7b7b1813 RG |
5291 | /* But a conversion that does not change the bit-pattern is ok. */ |
5292 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
5293 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2dab46d5 IE |
5294 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
5295 | /* Conversion between boolean types of different sizes is | |
5296 | a simple assignment in case their vectypes are same | |
5297 | boolean vectors. */ | |
5298 | && (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
5299 | || !VECTOR_BOOLEAN_TYPE_P (vectype_in))) | |
7b7b1813 | 5300 | { |
73fbfcad | 5301 | if (dump_enabled_p ()) |
78c60e3d SS |
5302 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5303 | "type conversion to/from bit-precision " | |
e645e942 | 5304 | "unsupported.\n"); |
7b7b1813 RG |
5305 | return false; |
5306 | } | |
5307 | ||
ebfd146a IR |
5308 | if (!vec_stmt) /* transformation not required. */ |
5309 | { | |
5310 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
adac3a68 | 5311 | DUMP_VECT_SCOPE ("vectorizable_assignment"); |
68435eb2 | 5312 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
ebfd146a IR |
5313 | return true; |
5314 | } | |
5315 | ||
67b8dbac | 5316 | /* Transform. */ |
73fbfcad | 5317 | if (dump_enabled_p ()) |
e645e942 | 5318 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
5319 | |
5320 | /* Handle def. */ | |
5321 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5322 | ||
5323 | /* Handle use. */ | |
f18b55bd | 5324 | for (j = 0; j < ncopies; j++) |
ebfd146a | 5325 | { |
f18b55bd IR |
5326 | /* Handle uses. */ |
5327 | if (j == 0) | |
306b0c92 | 5328 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
f18b55bd IR |
5329 | else |
5330 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
5331 | ||
5332 | /* Arguments are ready. create the new vector stmt. */ | |
e1bd7296 | 5333 | stmt_vec_info new_stmt_info = NULL; |
9771b263 | 5334 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 5335 | { |
7b7ec6c5 RG |
5336 | if (CONVERT_EXPR_CODE_P (code) |
5337 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 5338 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
e1bd7296 | 5339 | gassign *new_stmt = gimple_build_assign (vec_dest, vop); |
f18b55bd IR |
5340 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5341 | gimple_assign_set_lhs (new_stmt, new_temp); | |
e1bd7296 | 5342 | new_stmt_info = vect_finish_stmt_generation (stmt, new_stmt, gsi); |
f18b55bd | 5343 | if (slp_node) |
e1bd7296 | 5344 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
f18b55bd | 5345 | } |
ebfd146a IR |
5346 | |
5347 | if (slp_node) | |
f18b55bd IR |
5348 | continue; |
5349 | ||
5350 | if (j == 0) | |
e1bd7296 | 5351 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
f18b55bd | 5352 | else |
e1bd7296 | 5353 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
f18b55bd | 5354 | |
e1bd7296 | 5355 | prev_stmt_info = new_stmt_info; |
f18b55bd | 5356 | } |
b8698a0f | 5357 | |
9771b263 | 5358 | vec_oprnds.release (); |
ebfd146a IR |
5359 | return true; |
5360 | } | |
5361 | ||
9dc3f7de | 5362 | |
1107f3ae IR |
5363 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
5364 | either as shift by a scalar or by a vector. */ | |
5365 | ||
5366 | bool | |
5367 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
5368 | { | |
5369 | ||
ef4bddc2 | 5370 | machine_mode vec_mode; |
1107f3ae IR |
5371 | optab optab; |
5372 | int icode; | |
5373 | tree vectype; | |
5374 | ||
5375 | vectype = get_vectype_for_scalar_type (scalar_type); | |
5376 | if (!vectype) | |
5377 | return false; | |
5378 | ||
5379 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5380 | if (!optab | |
5381 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
5382 | { | |
5383 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5384 | if (!optab | |
5385 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
5386 | == CODE_FOR_nothing)) | |
5387 | return false; | |
5388 | } | |
5389 | ||
5390 | vec_mode = TYPE_MODE (vectype); | |
5391 | icode = (int) optab_handler (optab, vec_mode); | |
5392 | if (icode == CODE_FOR_nothing) | |
5393 | return false; | |
5394 | ||
5395 | return true; | |
5396 | } | |
5397 | ||
5398 | ||
9dc3f7de IR |
5399 | /* Function vectorizable_shift. |
5400 | ||
5401 | Check if STMT performs a shift operation that can be vectorized. | |
5402 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
5403 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
5404 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5405 | ||
5406 | static bool | |
355fe088 | 5407 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
1eede195 | 5408 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
68435eb2 | 5409 | stmt_vector_for_cost *cost_vec) |
9dc3f7de IR |
5410 | { |
5411 | tree vec_dest; | |
5412 | tree scalar_dest; | |
5413 | tree op0, op1 = NULL; | |
5414 | tree vec_oprnd1 = NULL_TREE; | |
5415 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5416 | tree vectype; | |
5417 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
5418 | enum tree_code code; | |
ef4bddc2 | 5419 | machine_mode vec_mode; |
9dc3f7de IR |
5420 | tree new_temp; |
5421 | optab optab; | |
5422 | int icode; | |
ef4bddc2 | 5423 | machine_mode optab_op2_mode; |
9dc3f7de | 5424 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 5425 | int ndts = 2; |
9dc3f7de | 5426 | stmt_vec_info prev_stmt_info; |
928686b1 RS |
5427 | poly_uint64 nunits_in; |
5428 | poly_uint64 nunits_out; | |
9dc3f7de | 5429 | tree vectype_out; |
cede2577 | 5430 | tree op1_vectype; |
9dc3f7de IR |
5431 | int ncopies; |
5432 | int j, i; | |
6e1aa848 DN |
5433 | vec<tree> vec_oprnds0 = vNULL; |
5434 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
5435 | tree vop0, vop1; |
5436 | unsigned int k; | |
49eab32e | 5437 | bool scalar_shift_arg = true; |
9dc3f7de | 5438 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5439 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
5440 | |
5441 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
5442 | return false; | |
5443 | ||
66c16fd9 RB |
5444 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5445 | && ! vec_stmt) | |
9dc3f7de IR |
5446 | return false; |
5447 | ||
5448 | /* Is STMT a vectorizable binary/unary operation? */ | |
5449 | if (!is_gimple_assign (stmt)) | |
5450 | return false; | |
5451 | ||
5452 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5453 | return false; | |
5454 | ||
5455 | code = gimple_assign_rhs_code (stmt); | |
5456 | ||
5457 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
5458 | || code == RROTATE_EXPR)) | |
5459 | return false; | |
5460 | ||
5461 | scalar_dest = gimple_assign_lhs (stmt); | |
5462 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2be65d9e | 5463 | if (!type_has_mode_precision_p (TREE_TYPE (scalar_dest))) |
7b7b1813 | 5464 | { |
73fbfcad | 5465 | if (dump_enabled_p ()) |
78c60e3d | 5466 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5467 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
5468 | return false; |
5469 | } | |
9dc3f7de IR |
5470 | |
5471 | op0 = gimple_assign_rhs1 (stmt); | |
894dd753 | 5472 | if (!vect_is_simple_use (op0, vinfo, &dt[0], &vectype)) |
9dc3f7de | 5473 | { |
73fbfcad | 5474 | if (dump_enabled_p ()) |
78c60e3d | 5475 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5476 | "use not simple.\n"); |
9dc3f7de IR |
5477 | return false; |
5478 | } | |
5479 | /* If op0 is an external or constant def use a vector type with | |
5480 | the same size as the output vector type. */ | |
5481 | if (!vectype) | |
5482 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5483 | if (vec_stmt) | |
5484 | gcc_assert (vectype); | |
5485 | if (!vectype) | |
5486 | { | |
73fbfcad | 5487 | if (dump_enabled_p ()) |
78c60e3d | 5488 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5489 | "no vectype for scalar type\n"); |
9dc3f7de IR |
5490 | return false; |
5491 | } | |
5492 | ||
5493 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5494 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
928686b1 | 5495 | if (maybe_ne (nunits_out, nunits_in)) |
9dc3f7de IR |
5496 | return false; |
5497 | ||
5498 | op1 = gimple_assign_rhs2 (stmt); | |
fef96d8e RS |
5499 | stmt_vec_info op1_def_stmt_info; |
5500 | if (!vect_is_simple_use (op1, vinfo, &dt[1], &op1_vectype, | |
5501 | &op1_def_stmt_info)) | |
9dc3f7de | 5502 | { |
73fbfcad | 5503 | if (dump_enabled_p ()) |
78c60e3d | 5504 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5505 | "use not simple.\n"); |
9dc3f7de IR |
5506 | return false; |
5507 | } | |
5508 | ||
9dc3f7de IR |
5509 | /* Multiple types in SLP are handled by creating the appropriate number of |
5510 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5511 | case of SLP. */ | |
fce57248 | 5512 | if (slp_node) |
9dc3f7de IR |
5513 | ncopies = 1; |
5514 | else | |
e8f142e2 | 5515 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
9dc3f7de IR |
5516 | |
5517 | gcc_assert (ncopies >= 1); | |
5518 | ||
5519 | /* Determine whether the shift amount is a vector, or scalar. If the | |
5520 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
5521 | ||
dbfa87aa YR |
5522 | if ((dt[1] == vect_internal_def |
5523 | || dt[1] == vect_induction_def) | |
5524 | && !slp_node) | |
49eab32e JJ |
5525 | scalar_shift_arg = false; |
5526 | else if (dt[1] == vect_constant_def | |
5527 | || dt[1] == vect_external_def | |
5528 | || dt[1] == vect_internal_def) | |
5529 | { | |
5530 | /* In SLP, need to check whether the shift count is the same, | |
5531 | in loops if it is a constant or invariant, it is always | |
5532 | a scalar shift. */ | |
5533 | if (slp_node) | |
5534 | { | |
b9787581 RS |
5535 | vec<stmt_vec_info> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
5536 | stmt_vec_info slpstmt_info; | |
49eab32e | 5537 | |
b9787581 RS |
5538 | FOR_EACH_VEC_ELT (stmts, k, slpstmt_info) |
5539 | { | |
5540 | gassign *slpstmt = as_a <gassign *> (slpstmt_info->stmt); | |
5541 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) | |
5542 | scalar_shift_arg = false; | |
5543 | } | |
49eab32e | 5544 | } |
60d393e8 RB |
5545 | |
5546 | /* If the shift amount is computed by a pattern stmt we cannot | |
5547 | use the scalar amount directly thus give up and use a vector | |
5548 | shift. */ | |
fef96d8e RS |
5549 | if (op1_def_stmt_info && is_pattern_stmt_p (op1_def_stmt_info)) |
5550 | scalar_shift_arg = false; | |
49eab32e JJ |
5551 | } |
5552 | else | |
5553 | { | |
73fbfcad | 5554 | if (dump_enabled_p ()) |
78c60e3d | 5555 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5556 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
5557 | return false; |
5558 | } | |
5559 | ||
9dc3f7de | 5560 | /* Vector shifted by vector. */ |
49eab32e | 5561 | if (!scalar_shift_arg) |
9dc3f7de IR |
5562 | { |
5563 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 5564 | if (dump_enabled_p ()) |
78c60e3d | 5565 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5566 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 5567 | |
aa948027 JJ |
5568 | if (!op1_vectype) |
5569 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
5570 | if (op1_vectype == NULL_TREE | |
5571 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 5572 | { |
73fbfcad | 5573 | if (dump_enabled_p ()) |
78c60e3d SS |
5574 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5575 | "unusable type for last operand in" | |
e645e942 | 5576 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
5577 | return false; |
5578 | } | |
9dc3f7de IR |
5579 | } |
5580 | /* See if the machine has a vector shifted by scalar insn and if not | |
5581 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 5582 | else |
9dc3f7de IR |
5583 | { |
5584 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5585 | if (optab | |
5586 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
5587 | { | |
73fbfcad | 5588 | if (dump_enabled_p ()) |
78c60e3d | 5589 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5590 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
5591 | } |
5592 | else | |
5593 | { | |
5594 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5595 | if (optab | |
5596 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
5597 | != CODE_FOR_nothing)) | |
5598 | { | |
49eab32e JJ |
5599 | scalar_shift_arg = false; |
5600 | ||
73fbfcad | 5601 | if (dump_enabled_p ()) |
78c60e3d | 5602 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5603 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
5604 | |
5605 | /* Unlike the other binary operators, shifts/rotates have | |
5606 | the rhs being int, instead of the same type as the lhs, | |
5607 | so make sure the scalar is the right type if we are | |
aa948027 | 5608 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
5609 | if (dt[1] == vect_constant_def) |
5610 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
5611 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
5612 | TREE_TYPE (op1))) | |
5613 | { | |
5614 | if (slp_node | |
5615 | && TYPE_MODE (TREE_TYPE (vectype)) | |
5616 | != TYPE_MODE (TREE_TYPE (op1))) | |
5617 | { | |
73fbfcad | 5618 | if (dump_enabled_p ()) |
78c60e3d SS |
5619 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5620 | "unusable type for last operand in" | |
e645e942 | 5621 | " vector/vector shift/rotate.\n"); |
21c0a521 | 5622 | return false; |
aa948027 JJ |
5623 | } |
5624 | if (vec_stmt && !slp_node) | |
5625 | { | |
5626 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
5627 | op1 = vect_init_vector (stmt, op1, | |
5628 | TREE_TYPE (vectype), NULL); | |
5629 | } | |
5630 | } | |
9dc3f7de IR |
5631 | } |
5632 | } | |
5633 | } | |
9dc3f7de IR |
5634 | |
5635 | /* Supportable by target? */ | |
5636 | if (!optab) | |
5637 | { | |
73fbfcad | 5638 | if (dump_enabled_p ()) |
78c60e3d | 5639 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5640 | "no optab.\n"); |
9dc3f7de IR |
5641 | return false; |
5642 | } | |
5643 | vec_mode = TYPE_MODE (vectype); | |
5644 | icode = (int) optab_handler (optab, vec_mode); | |
5645 | if (icode == CODE_FOR_nothing) | |
5646 | { | |
73fbfcad | 5647 | if (dump_enabled_p ()) |
78c60e3d | 5648 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5649 | "op not supported by target.\n"); |
9dc3f7de | 5650 | /* Check only during analysis. */ |
cf098191 | 5651 | if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) |
ca09abcb RS |
5652 | || (!vec_stmt |
5653 | && !vect_worthwhile_without_simd_p (vinfo, code))) | |
9dc3f7de | 5654 | return false; |
73fbfcad | 5655 | if (dump_enabled_p ()) |
e645e942 TJ |
5656 | dump_printf_loc (MSG_NOTE, vect_location, |
5657 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
5658 | } |
5659 | ||
5660 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
ca09abcb RS |
5661 | if (!vec_stmt |
5662 | && !VECTOR_MODE_P (TYPE_MODE (vectype)) | |
5663 | && !vect_worthwhile_without_simd_p (vinfo, code)) | |
9dc3f7de | 5664 | { |
73fbfcad | 5665 | if (dump_enabled_p ()) |
78c60e3d | 5666 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5667 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
5668 | return false; |
5669 | } | |
5670 | ||
5671 | if (!vec_stmt) /* transformation not required. */ | |
5672 | { | |
5673 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
adac3a68 | 5674 | DUMP_VECT_SCOPE ("vectorizable_shift"); |
68435eb2 | 5675 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
9dc3f7de IR |
5676 | return true; |
5677 | } | |
5678 | ||
67b8dbac | 5679 | /* Transform. */ |
9dc3f7de | 5680 | |
73fbfcad | 5681 | if (dump_enabled_p ()) |
78c60e3d | 5682 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5683 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
5684 | |
5685 | /* Handle def. */ | |
5686 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5687 | ||
9dc3f7de IR |
5688 | prev_stmt_info = NULL; |
5689 | for (j = 0; j < ncopies; j++) | |
5690 | { | |
5691 | /* Handle uses. */ | |
5692 | if (j == 0) | |
5693 | { | |
5694 | if (scalar_shift_arg) | |
5695 | { | |
5696 | /* Vector shl and shr insn patterns can be defined with scalar | |
5697 | operand 2 (shift operand). In this case, use constant or loop | |
5698 | invariant op1 directly, without extending it to vector mode | |
5699 | first. */ | |
5700 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
5701 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
5702 | { | |
73fbfcad | 5703 | if (dump_enabled_p ()) |
78c60e3d | 5704 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5705 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 5706 | vec_oprnd1 = op1; |
8930f723 | 5707 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 5708 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5709 | if (slp_node) |
5710 | { | |
5711 | /* Store vec_oprnd1 for every vector stmt to be created | |
5712 | for SLP_NODE. We check during the analysis that all | |
5713 | the shift arguments are the same. | |
5714 | TODO: Allow different constants for different vector | |
5715 | stmts generated for an SLP instance. */ | |
5716 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 5717 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5718 | } |
5719 | } | |
5720 | } | |
5721 | ||
5722 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
5723 | (a special case for certain kind of vector shifts); otherwise, | |
5724 | operand 1 should be of a vector type (the usual case). */ | |
5725 | if (vec_oprnd1) | |
5726 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5727 | slp_node); |
9dc3f7de IR |
5728 | else |
5729 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
306b0c92 | 5730 | slp_node); |
9dc3f7de IR |
5731 | } |
5732 | else | |
5733 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5734 | ||
5735 | /* Arguments are ready. Create the new vector stmt. */ | |
e1bd7296 | 5736 | stmt_vec_info new_stmt_info = NULL; |
9771b263 | 5737 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 5738 | { |
9771b263 | 5739 | vop1 = vec_oprnds1[i]; |
e1bd7296 | 5740 | gassign *new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
5741 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5742 | gimple_assign_set_lhs (new_stmt, new_temp); | |
e1bd7296 | 5743 | new_stmt_info = vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9dc3f7de | 5744 | if (slp_node) |
e1bd7296 | 5745 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
9dc3f7de IR |
5746 | } |
5747 | ||
5748 | if (slp_node) | |
5749 | continue; | |
5750 | ||
5751 | if (j == 0) | |
e1bd7296 | 5752 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
9dc3f7de | 5753 | else |
e1bd7296 RS |
5754 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
5755 | prev_stmt_info = new_stmt_info; | |
9dc3f7de IR |
5756 | } |
5757 | ||
9771b263 DN |
5758 | vec_oprnds0.release (); |
5759 | vec_oprnds1.release (); | |
9dc3f7de IR |
5760 | |
5761 | return true; | |
5762 | } | |
5763 | ||
5764 | ||
ebfd146a IR |
5765 | /* Function vectorizable_operation. |
5766 | ||
16949072 RG |
5767 | Check if STMT performs a binary, unary or ternary operation that can |
5768 | be vectorized. | |
b8698a0f | 5769 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
5770 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5771 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5772 | ||
5773 | static bool | |
355fe088 | 5774 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
1eede195 | 5775 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
68435eb2 | 5776 | stmt_vector_for_cost *cost_vec) |
ebfd146a | 5777 | { |
00f07b86 | 5778 | tree vec_dest; |
ebfd146a | 5779 | tree scalar_dest; |
16949072 | 5780 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 5781 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 5782 | tree vectype; |
ebfd146a | 5783 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
0eb952ea | 5784 | enum tree_code code, orig_code; |
ef4bddc2 | 5785 | machine_mode vec_mode; |
ebfd146a IR |
5786 | tree new_temp; |
5787 | int op_type; | |
00f07b86 | 5788 | optab optab; |
523ba738 | 5789 | bool target_support_p; |
16949072 RG |
5790 | enum vect_def_type dt[3] |
5791 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 5792 | int ndts = 3; |
ebfd146a | 5793 | stmt_vec_info prev_stmt_info; |
928686b1 RS |
5794 | poly_uint64 nunits_in; |
5795 | poly_uint64 nunits_out; | |
ebfd146a IR |
5796 | tree vectype_out; |
5797 | int ncopies; | |
5798 | int j, i; | |
6e1aa848 DN |
5799 | vec<tree> vec_oprnds0 = vNULL; |
5800 | vec<tree> vec_oprnds1 = vNULL; | |
5801 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 5802 | tree vop0, vop1, vop2; |
a70d6342 | 5803 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5804 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 5805 | |
a70d6342 | 5806 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5807 | return false; |
5808 | ||
66c16fd9 RB |
5809 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5810 | && ! vec_stmt) | |
ebfd146a IR |
5811 | return false; |
5812 | ||
5813 | /* Is STMT a vectorizable binary/unary operation? */ | |
5814 | if (!is_gimple_assign (stmt)) | |
5815 | return false; | |
5816 | ||
5817 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5818 | return false; | |
5819 | ||
0eb952ea | 5820 | orig_code = code = gimple_assign_rhs_code (stmt); |
ebfd146a | 5821 | |
1af4ebf5 MG |
5822 | /* For pointer addition and subtraction, we should use the normal |
5823 | plus and minus for the vector operation. */ | |
ebfd146a IR |
5824 | if (code == POINTER_PLUS_EXPR) |
5825 | code = PLUS_EXPR; | |
1af4ebf5 MG |
5826 | if (code == POINTER_DIFF_EXPR) |
5827 | code = MINUS_EXPR; | |
ebfd146a IR |
5828 | |
5829 | /* Support only unary or binary operations. */ | |
5830 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 5831 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 5832 | { |
73fbfcad | 5833 | if (dump_enabled_p ()) |
78c60e3d | 5834 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5835 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 5836 | op_type); |
ebfd146a IR |
5837 | return false; |
5838 | } | |
5839 | ||
b690cc0f RG |
5840 | scalar_dest = gimple_assign_lhs (stmt); |
5841 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
5842 | ||
7b7b1813 RG |
5843 | /* Most operations cannot handle bit-precision types without extra |
5844 | truncations. */ | |
045c1278 | 5845 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
2be65d9e | 5846 | && !type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
7b7b1813 RG |
5847 | /* Exception are bitwise binary operations. */ |
5848 | && code != BIT_IOR_EXPR | |
5849 | && code != BIT_XOR_EXPR | |
5850 | && code != BIT_AND_EXPR) | |
5851 | { | |
73fbfcad | 5852 | if (dump_enabled_p ()) |
78c60e3d | 5853 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5854 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
5855 | return false; |
5856 | } | |
5857 | ||
ebfd146a | 5858 | op0 = gimple_assign_rhs1 (stmt); |
894dd753 | 5859 | if (!vect_is_simple_use (op0, vinfo, &dt[0], &vectype)) |
ebfd146a | 5860 | { |
73fbfcad | 5861 | if (dump_enabled_p ()) |
78c60e3d | 5862 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5863 | "use not simple.\n"); |
ebfd146a IR |
5864 | return false; |
5865 | } | |
b690cc0f RG |
5866 | /* If op0 is an external or constant def use a vector type with |
5867 | the same size as the output vector type. */ | |
5868 | if (!vectype) | |
b036c6c5 IE |
5869 | { |
5870 | /* For boolean type we cannot determine vectype by | |
5871 | invariant value (don't know whether it is a vector | |
5872 | of booleans or vector of integers). We use output | |
5873 | vectype because operations on boolean don't change | |
5874 | type. */ | |
2568d8a1 | 5875 | if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op0))) |
b036c6c5 | 5876 | { |
2568d8a1 | 5877 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (scalar_dest))) |
b036c6c5 IE |
5878 | { |
5879 | if (dump_enabled_p ()) | |
5880 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5881 | "not supported operation on bool value.\n"); | |
5882 | return false; | |
5883 | } | |
5884 | vectype = vectype_out; | |
5885 | } | |
5886 | else | |
5887 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5888 | } | |
7d8930a0 IR |
5889 | if (vec_stmt) |
5890 | gcc_assert (vectype); | |
5891 | if (!vectype) | |
5892 | { | |
73fbfcad | 5893 | if (dump_enabled_p ()) |
7d8930a0 | 5894 | { |
78c60e3d SS |
5895 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5896 | "no vectype for scalar type "); | |
5897 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
5898 | TREE_TYPE (op0)); | |
e645e942 | 5899 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
5900 | } |
5901 | ||
5902 | return false; | |
5903 | } | |
b690cc0f RG |
5904 | |
5905 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5906 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
928686b1 | 5907 | if (maybe_ne (nunits_out, nunits_in)) |
b690cc0f | 5908 | return false; |
ebfd146a | 5909 | |
16949072 | 5910 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
5911 | { |
5912 | op1 = gimple_assign_rhs2 (stmt); | |
894dd753 | 5913 | if (!vect_is_simple_use (op1, vinfo, &dt[1])) |
ebfd146a | 5914 | { |
73fbfcad | 5915 | if (dump_enabled_p ()) |
78c60e3d | 5916 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5917 | "use not simple.\n"); |
ebfd146a IR |
5918 | return false; |
5919 | } | |
5920 | } | |
16949072 RG |
5921 | if (op_type == ternary_op) |
5922 | { | |
5923 | op2 = gimple_assign_rhs3 (stmt); | |
894dd753 | 5924 | if (!vect_is_simple_use (op2, vinfo, &dt[2])) |
16949072 | 5925 | { |
73fbfcad | 5926 | if (dump_enabled_p ()) |
78c60e3d | 5927 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5928 | "use not simple.\n"); |
16949072 RG |
5929 | return false; |
5930 | } | |
5931 | } | |
ebfd146a | 5932 | |
b690cc0f | 5933 | /* Multiple types in SLP are handled by creating the appropriate number of |
ff802fa1 | 5934 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 5935 | case of SLP. */ |
fce57248 | 5936 | if (slp_node) |
b690cc0f RG |
5937 | ncopies = 1; |
5938 | else | |
e8f142e2 | 5939 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
b690cc0f RG |
5940 | |
5941 | gcc_assert (ncopies >= 1); | |
5942 | ||
9dc3f7de | 5943 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
5944 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
5945 | || code == RROTATE_EXPR) | |
9dc3f7de | 5946 | return false; |
ebfd146a | 5947 | |
ebfd146a | 5948 | /* Supportable by target? */ |
00f07b86 RH |
5949 | |
5950 | vec_mode = TYPE_MODE (vectype); | |
5951 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 5952 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
5953 | else |
5954 | { | |
5955 | optab = optab_for_tree_code (code, vectype, optab_default); | |
5956 | if (!optab) | |
5deb57cb | 5957 | { |
73fbfcad | 5958 | if (dump_enabled_p ()) |
78c60e3d | 5959 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5960 | "no optab.\n"); |
00f07b86 | 5961 | return false; |
5deb57cb | 5962 | } |
523ba738 RS |
5963 | target_support_p = (optab_handler (optab, vec_mode) |
5964 | != CODE_FOR_nothing); | |
5deb57cb JJ |
5965 | } |
5966 | ||
523ba738 | 5967 | if (!target_support_p) |
ebfd146a | 5968 | { |
73fbfcad | 5969 | if (dump_enabled_p ()) |
78c60e3d | 5970 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5971 | "op not supported by target.\n"); |
ebfd146a | 5972 | /* Check only during analysis. */ |
cf098191 | 5973 | if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) |
ca09abcb | 5974 | || (!vec_stmt && !vect_worthwhile_without_simd_p (vinfo, code))) |
ebfd146a | 5975 | return false; |
73fbfcad | 5976 | if (dump_enabled_p ()) |
e645e942 TJ |
5977 | dump_printf_loc (MSG_NOTE, vect_location, |
5978 | "proceeding using word mode.\n"); | |
383d9c83 IR |
5979 | } |
5980 | ||
4a00c761 | 5981 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
5982 | if (!VECTOR_MODE_P (vec_mode) |
5983 | && !vec_stmt | |
ca09abcb | 5984 | && !vect_worthwhile_without_simd_p (vinfo, code)) |
7d8930a0 | 5985 | { |
73fbfcad | 5986 | if (dump_enabled_p ()) |
78c60e3d | 5987 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5988 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 5989 | return false; |
7d8930a0 | 5990 | } |
ebfd146a | 5991 | |
ebfd146a IR |
5992 | if (!vec_stmt) /* transformation not required. */ |
5993 | { | |
4a00c761 | 5994 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
adac3a68 | 5995 | DUMP_VECT_SCOPE ("vectorizable_operation"); |
68435eb2 | 5996 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
ebfd146a IR |
5997 | return true; |
5998 | } | |
5999 | ||
67b8dbac | 6000 | /* Transform. */ |
ebfd146a | 6001 | |
73fbfcad | 6002 | if (dump_enabled_p ()) |
78c60e3d | 6003 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6004 | "transform binary/unary operation.\n"); |
383d9c83 | 6005 | |
0eb952ea JJ |
6006 | /* POINTER_DIFF_EXPR has pointer arguments which are vectorized as |
6007 | vectors with unsigned elements, but the result is signed. So, we | |
6008 | need to compute the MINUS_EXPR into vectype temporary and | |
6009 | VIEW_CONVERT_EXPR it into the final vectype_out result. */ | |
6010 | tree vec_cvt_dest = NULL_TREE; | |
6011 | if (orig_code == POINTER_DIFF_EXPR) | |
7b76867b RB |
6012 | { |
6013 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
6014 | vec_cvt_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
6015 | } | |
6016 | /* Handle def. */ | |
6017 | else | |
6018 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
0eb952ea | 6019 | |
ebfd146a IR |
6020 | /* In case the vectorization factor (VF) is bigger than the number |
6021 | of elements that we can fit in a vectype (nunits), we have to generate | |
6022 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
6023 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
6024 | from one copy of the vector stmt to the next, in the field | |
6025 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
6026 | stages to find the correct vector defs to be used when vectorizing | |
6027 | stmts that use the defs of the current stmt. The example below | |
6028 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
6029 | we need to create 4 vectorized stmts): | |
6030 | ||
6031 | before vectorization: | |
6032 | RELATED_STMT VEC_STMT | |
6033 | S1: x = memref - - | |
6034 | S2: z = x + 1 - - | |
6035 | ||
6036 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
6037 | there): | |
6038 | RELATED_STMT VEC_STMT | |
6039 | VS1_0: vx0 = memref0 VS1_1 - | |
6040 | VS1_1: vx1 = memref1 VS1_2 - | |
6041 | VS1_2: vx2 = memref2 VS1_3 - | |
6042 | VS1_3: vx3 = memref3 - - | |
6043 | S1: x = load - VS1_0 | |
6044 | S2: z = x + 1 - - | |
6045 | ||
6046 | step2: vectorize stmt S2 (done here): | |
6047 | To vectorize stmt S2 we first need to find the relevant vector | |
6048 | def for the first operand 'x'. This is, as usual, obtained from | |
6049 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
6050 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
6051 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
6052 | the vector stmt VS2_0, and as usual, record it in the | |
6053 | STMT_VINFO_VEC_STMT of stmt S2. | |
6054 | When creating the second copy (VS2_1), we obtain the relevant vector | |
6055 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
6056 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
6057 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
6058 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
6059 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
6060 | chain of stmts and pointers: | |
6061 | RELATED_STMT VEC_STMT | |
6062 | VS1_0: vx0 = memref0 VS1_1 - | |
6063 | VS1_1: vx1 = memref1 VS1_2 - | |
6064 | VS1_2: vx2 = memref2 VS1_3 - | |
6065 | VS1_3: vx3 = memref3 - - | |
6066 | S1: x = load - VS1_0 | |
6067 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
6068 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
6069 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
6070 | VS2_3: vz3 = vx3 + v1 - - | |
6071 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
6072 | |
6073 | prev_stmt_info = NULL; | |
6074 | for (j = 0; j < ncopies; j++) | |
6075 | { | |
6076 | /* Handle uses. */ | |
6077 | if (j == 0) | |
4a00c761 | 6078 | { |
d6476f90 | 6079 | if (op_type == binary_op) |
4a00c761 | 6080 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, |
306b0c92 | 6081 | slp_node); |
d6476f90 RB |
6082 | else if (op_type == ternary_op) |
6083 | { | |
6084 | if (slp_node) | |
6085 | { | |
6086 | auto_vec<tree> ops(3); | |
6087 | ops.quick_push (op0); | |
6088 | ops.quick_push (op1); | |
6089 | ops.quick_push (op2); | |
6090 | auto_vec<vec<tree> > vec_defs(3); | |
6091 | vect_get_slp_defs (ops, slp_node, &vec_defs); | |
6092 | vec_oprnds0 = vec_defs[0]; | |
6093 | vec_oprnds1 = vec_defs[1]; | |
6094 | vec_oprnds2 = vec_defs[2]; | |
6095 | } | |
6096 | else | |
6097 | { | |
6098 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
6099 | NULL); | |
6100 | vect_get_vec_defs (op2, NULL_TREE, stmt, &vec_oprnds2, NULL, | |
6101 | NULL); | |
6102 | } | |
6103 | } | |
4a00c761 JJ |
6104 | else |
6105 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 6106 | slp_node); |
4a00c761 | 6107 | } |
ebfd146a | 6108 | else |
4a00c761 JJ |
6109 | { |
6110 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
6111 | if (op_type == ternary_op) | |
6112 | { | |
9771b263 DN |
6113 | tree vec_oprnd = vec_oprnds2.pop (); |
6114 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
6115 | vec_oprnd)); | |
4a00c761 JJ |
6116 | } |
6117 | } | |
6118 | ||
6119 | /* Arguments are ready. Create the new vector stmt. */ | |
e1bd7296 | 6120 | stmt_vec_info new_stmt_info = NULL; |
9771b263 | 6121 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 6122 | { |
4a00c761 | 6123 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 6124 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 6125 | vop2 = ((op_type == ternary_op) |
9771b263 | 6126 | ? vec_oprnds2[i] : NULL_TREE); |
e1bd7296 RS |
6127 | gassign *new_stmt = gimple_build_assign (vec_dest, code, |
6128 | vop0, vop1, vop2); | |
4a00c761 JJ |
6129 | new_temp = make_ssa_name (vec_dest, new_stmt); |
6130 | gimple_assign_set_lhs (new_stmt, new_temp); | |
e1bd7296 | 6131 | new_stmt_info = vect_finish_stmt_generation (stmt, new_stmt, gsi); |
0eb952ea JJ |
6132 | if (vec_cvt_dest) |
6133 | { | |
6134 | new_temp = build1 (VIEW_CONVERT_EXPR, vectype_out, new_temp); | |
e1bd7296 RS |
6135 | gassign *new_stmt |
6136 | = gimple_build_assign (vec_cvt_dest, VIEW_CONVERT_EXPR, | |
6137 | new_temp); | |
0eb952ea JJ |
6138 | new_temp = make_ssa_name (vec_cvt_dest, new_stmt); |
6139 | gimple_assign_set_lhs (new_stmt, new_temp); | |
e1bd7296 RS |
6140 | new_stmt_info |
6141 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
0eb952ea | 6142 | } |
4a00c761 | 6143 | if (slp_node) |
e1bd7296 | 6144 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
ebfd146a IR |
6145 | } |
6146 | ||
4a00c761 JJ |
6147 | if (slp_node) |
6148 | continue; | |
6149 | ||
6150 | if (j == 0) | |
e1bd7296 | 6151 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
4a00c761 | 6152 | else |
e1bd7296 RS |
6153 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
6154 | prev_stmt_info = new_stmt_info; | |
ebfd146a IR |
6155 | } |
6156 | ||
9771b263 DN |
6157 | vec_oprnds0.release (); |
6158 | vec_oprnds1.release (); | |
6159 | vec_oprnds2.release (); | |
ebfd146a | 6160 | |
ebfd146a IR |
6161 | return true; |
6162 | } | |
6163 | ||
f702e7d4 | 6164 | /* A helper function to ensure data reference DR's base alignment. */ |
c716e67f XDL |
6165 | |
6166 | static void | |
f702e7d4 | 6167 | ensure_base_align (struct data_reference *dr) |
c716e67f | 6168 | { |
ca823c85 | 6169 | if (DR_VECT_AUX (dr)->misalignment == DR_MISALIGNMENT_UNINITIALIZED) |
c716e67f XDL |
6170 | return; |
6171 | ||
52639a61 | 6172 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f | 6173 | { |
52639a61 | 6174 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 6175 | |
f702e7d4 RS |
6176 | unsigned int align_base_to = DR_TARGET_ALIGNMENT (dr) * BITS_PER_UNIT; |
6177 | ||
428f0c67 | 6178 | if (decl_in_symtab_p (base_decl)) |
f702e7d4 | 6179 | symtab_node::get (base_decl)->increase_alignment (align_base_to); |
428f0c67 JH |
6180 | else |
6181 | { | |
f702e7d4 | 6182 | SET_DECL_ALIGN (base_decl, align_base_to); |
428f0c67 JH |
6183 | DECL_USER_ALIGN (base_decl) = 1; |
6184 | } | |
52639a61 | 6185 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
6186 | } |
6187 | } | |
6188 | ||
ebfd146a | 6189 | |
44fc7854 BE |
6190 | /* Function get_group_alias_ptr_type. |
6191 | ||
6192 | Return the alias type for the group starting at FIRST_STMT. */ | |
6193 | ||
6194 | static tree | |
6195 | get_group_alias_ptr_type (gimple *first_stmt) | |
6196 | { | |
6197 | struct data_reference *first_dr, *next_dr; | |
44fc7854 BE |
6198 | |
6199 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
bffb8014 RS |
6200 | stmt_vec_info next_stmt_info |
6201 | = DR_GROUP_NEXT_ELEMENT (vinfo_for_stmt (first_stmt)); | |
6202 | while (next_stmt_info) | |
44fc7854 | 6203 | { |
bffb8014 | 6204 | next_dr = STMT_VINFO_DATA_REF (next_stmt_info); |
44fc7854 BE |
6205 | if (get_alias_set (DR_REF (first_dr)) |
6206 | != get_alias_set (DR_REF (next_dr))) | |
6207 | { | |
6208 | if (dump_enabled_p ()) | |
6209 | dump_printf_loc (MSG_NOTE, vect_location, | |
6210 | "conflicting alias set types.\n"); | |
6211 | return ptr_type_node; | |
6212 | } | |
bffb8014 | 6213 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
44fc7854 BE |
6214 | } |
6215 | return reference_alias_ptr_type (DR_REF (first_dr)); | |
6216 | } | |
6217 | ||
6218 | ||
ebfd146a IR |
6219 | /* Function vectorizable_store. |
6220 | ||
b8698a0f L |
6221 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
6222 | can be vectorized. | |
6223 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6224 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6225 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6226 | ||
6227 | static bool | |
1eede195 RS |
6228 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, |
6229 | stmt_vec_info *vec_stmt, slp_tree slp_node, | |
6230 | stmt_vector_for_cost *cost_vec) | |
ebfd146a | 6231 | { |
ebfd146a IR |
6232 | tree data_ref; |
6233 | tree op; | |
6234 | tree vec_oprnd = NULL_TREE; | |
6235 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6236 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 6237 | tree elem_type; |
ebfd146a | 6238 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6239 | struct loop *loop = NULL; |
ef4bddc2 | 6240 | machine_mode vec_mode; |
ebfd146a IR |
6241 | tree dummy; |
6242 | enum dr_alignment_support alignment_support_scheme; | |
929b4411 RS |
6243 | enum vect_def_type rhs_dt = vect_unknown_def_type; |
6244 | enum vect_def_type mask_dt = vect_unknown_def_type; | |
ebfd146a IR |
6245 | stmt_vec_info prev_stmt_info = NULL; |
6246 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6247 | tree dataref_offset = NULL_TREE; |
355fe088 | 6248 | gimple *ptr_incr = NULL; |
ebfd146a IR |
6249 | int ncopies; |
6250 | int j; | |
bffb8014 | 6251 | stmt_vec_info first_stmt_info; |
2de001ee | 6252 | bool grouped_store; |
ebfd146a | 6253 | unsigned int group_size, i; |
6e1aa848 DN |
6254 | vec<tree> oprnds = vNULL; |
6255 | vec<tree> result_chain = vNULL; | |
ebfd146a | 6256 | bool inv_p; |
09dfa495 | 6257 | tree offset = NULL_TREE; |
6e1aa848 | 6258 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 6259 | bool slp = (slp_node != NULL); |
ebfd146a | 6260 | unsigned int vec_num; |
a70d6342 | 6261 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 6262 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 6263 | tree aggr_type; |
134c85ca | 6264 | gather_scatter_info gs_info; |
d9f21f6a | 6265 | poly_uint64 vf; |
2de001ee | 6266 | vec_load_store_type vls_type; |
44fc7854 | 6267 | tree ref_type; |
a70d6342 | 6268 | |
a70d6342 | 6269 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
6270 | return false; |
6271 | ||
66c16fd9 RB |
6272 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6273 | && ! vec_stmt) | |
ebfd146a IR |
6274 | return false; |
6275 | ||
6276 | /* Is vectorizable store? */ | |
6277 | ||
c3a8f964 RS |
6278 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
6279 | if (is_gimple_assign (stmt)) | |
6280 | { | |
6281 | tree scalar_dest = gimple_assign_lhs (stmt); | |
6282 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR | |
6283 | && is_pattern_stmt_p (stmt_info)) | |
6284 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
6285 | if (TREE_CODE (scalar_dest) != ARRAY_REF | |
6286 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF | |
6287 | && TREE_CODE (scalar_dest) != INDIRECT_REF | |
6288 | && TREE_CODE (scalar_dest) != COMPONENT_REF | |
6289 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
6290 | && TREE_CODE (scalar_dest) != REALPART_EXPR | |
6291 | && TREE_CODE (scalar_dest) != MEM_REF) | |
6292 | return false; | |
6293 | } | |
6294 | else | |
6295 | { | |
6296 | gcall *call = dyn_cast <gcall *> (stmt); | |
f307441a RS |
6297 | if (!call || !gimple_call_internal_p (call)) |
6298 | return false; | |
6299 | ||
6300 | internal_fn ifn = gimple_call_internal_fn (call); | |
6301 | if (!internal_store_fn_p (ifn)) | |
c3a8f964 | 6302 | return false; |
ebfd146a | 6303 | |
c3a8f964 RS |
6304 | if (slp_node != NULL) |
6305 | { | |
6306 | if (dump_enabled_p ()) | |
6307 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6308 | "SLP of masked stores not supported.\n"); | |
6309 | return false; | |
6310 | } | |
6311 | ||
f307441a RS |
6312 | int mask_index = internal_fn_mask_index (ifn); |
6313 | if (mask_index >= 0) | |
6314 | { | |
6315 | mask = gimple_call_arg (call, mask_index); | |
929b4411 RS |
6316 | if (!vect_check_load_store_mask (stmt, mask, &mask_dt, |
6317 | &mask_vectype)) | |
f307441a RS |
6318 | return false; |
6319 | } | |
c3a8f964 RS |
6320 | } |
6321 | ||
6322 | op = vect_get_store_rhs (stmt); | |
ebfd146a | 6323 | |
fce57248 RS |
6324 | /* Cannot have hybrid store SLP -- that would mean storing to the |
6325 | same location twice. */ | |
6326 | gcc_assert (slp == PURE_SLP_STMT (stmt_info)); | |
6327 | ||
f4d09712 | 6328 | tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE; |
4d694b27 | 6329 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 JJ |
6330 | |
6331 | if (loop_vinfo) | |
b17dc4d4 RB |
6332 | { |
6333 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6334 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6335 | } | |
6336 | else | |
6337 | vf = 1; | |
465c8c19 JJ |
6338 | |
6339 | /* Multiple types in SLP are handled by creating the appropriate number of | |
6340 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
6341 | case of SLP. */ | |
fce57248 | 6342 | if (slp) |
465c8c19 JJ |
6343 | ncopies = 1; |
6344 | else | |
e8f142e2 | 6345 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
6346 | |
6347 | gcc_assert (ncopies >= 1); | |
6348 | ||
6349 | /* FORNOW. This restriction should be relaxed. */ | |
6350 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
6351 | { | |
6352 | if (dump_enabled_p ()) | |
6353 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6354 | "multiple types in nested loop.\n"); | |
6355 | return false; | |
6356 | } | |
6357 | ||
929b4411 | 6358 | if (!vect_check_store_rhs (stmt, op, &rhs_dt, &rhs_vectype, &vls_type)) |
f4d09712 KY |
6359 | return false; |
6360 | ||
272c6793 | 6361 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 6362 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 6363 | |
ebfd146a IR |
6364 | if (!STMT_VINFO_DATA_REF (stmt_info)) |
6365 | return false; | |
6366 | ||
2de001ee | 6367 | vect_memory_access_type memory_access_type; |
7e11fc7f | 6368 | if (!get_load_store_type (stmt, vectype, slp, mask, vls_type, ncopies, |
2de001ee RS |
6369 | &memory_access_type, &gs_info)) |
6370 | return false; | |
3bab6342 | 6371 | |
c3a8f964 RS |
6372 | if (mask) |
6373 | { | |
7e11fc7f RS |
6374 | if (memory_access_type == VMAT_CONTIGUOUS) |
6375 | { | |
6376 | if (!VECTOR_MODE_P (vec_mode) | |
6377 | || !can_vec_mask_load_store_p (vec_mode, | |
6378 | TYPE_MODE (mask_vectype), false)) | |
6379 | return false; | |
6380 | } | |
f307441a RS |
6381 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
6382 | && (memory_access_type != VMAT_GATHER_SCATTER || gs_info.decl)) | |
c3a8f964 RS |
6383 | { |
6384 | if (dump_enabled_p ()) | |
6385 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6386 | "unsupported access type for masked store.\n"); | |
6387 | return false; | |
6388 | } | |
c3a8f964 RS |
6389 | } |
6390 | else | |
6391 | { | |
6392 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6393 | (e.g. - array initialization with 0). */ | |
6394 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) | |
6395 | return false; | |
6396 | } | |
6397 | ||
f307441a | 6398 | grouped_store = (STMT_VINFO_GROUPED_ACCESS (stmt_info) |
b5ec4de7 RS |
6399 | && memory_access_type != VMAT_GATHER_SCATTER |
6400 | && (slp || memory_access_type != VMAT_CONTIGUOUS)); | |
7cfb4d93 RS |
6401 | if (grouped_store) |
6402 | { | |
bffb8014 RS |
6403 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
6404 | first_dr = STMT_VINFO_DATA_REF (first_stmt_info); | |
6405 | group_size = DR_GROUP_SIZE (first_stmt_info); | |
7cfb4d93 RS |
6406 | } |
6407 | else | |
6408 | { | |
bffb8014 | 6409 | first_stmt_info = stmt_info; |
7cfb4d93 RS |
6410 | first_dr = dr; |
6411 | group_size = vec_num = 1; | |
6412 | } | |
6413 | ||
ebfd146a IR |
6414 | if (!vec_stmt) /* transformation not required. */ |
6415 | { | |
2de001ee | 6416 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
7cfb4d93 RS |
6417 | |
6418 | if (loop_vinfo | |
6419 | && LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) | |
6420 | check_load_store_masking (loop_vinfo, vectype, vls_type, group_size, | |
bfaa08b7 | 6421 | memory_access_type, &gs_info); |
7cfb4d93 | 6422 | |
ebfd146a | 6423 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; |
68435eb2 RB |
6424 | vect_model_store_cost (stmt_info, ncopies, rhs_dt, memory_access_type, |
6425 | vls_type, slp_node, cost_vec); | |
ebfd146a IR |
6426 | return true; |
6427 | } | |
2de001ee | 6428 | gcc_assert (memory_access_type == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
ebfd146a | 6429 | |
67b8dbac | 6430 | /* Transform. */ |
ebfd146a | 6431 | |
f702e7d4 | 6432 | ensure_base_align (dr); |
c716e67f | 6433 | |
f307441a | 6434 | if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
3bab6342 | 6435 | { |
c3a8f964 | 6436 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, src; |
134c85ca | 6437 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
6438 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; |
6439 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
6440 | edge pe = loop_preheader_edge (loop); | |
6441 | gimple_seq seq; | |
6442 | basic_block new_bb; | |
6443 | enum { NARROW, NONE, WIDEN } modifier; | |
4d694b27 RS |
6444 | poly_uint64 scatter_off_nunits |
6445 | = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype); | |
3bab6342 | 6446 | |
4d694b27 | 6447 | if (known_eq (nunits, scatter_off_nunits)) |
3bab6342 | 6448 | modifier = NONE; |
4d694b27 | 6449 | else if (known_eq (nunits * 2, scatter_off_nunits)) |
3bab6342 | 6450 | { |
3bab6342 AT |
6451 | modifier = WIDEN; |
6452 | ||
4d694b27 RS |
6453 | /* Currently gathers and scatters are only supported for |
6454 | fixed-length vectors. */ | |
6455 | unsigned int count = scatter_off_nunits.to_constant (); | |
6456 | vec_perm_builder sel (count, count, 1); | |
6457 | for (i = 0; i < (unsigned int) count; ++i) | |
6458 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 6459 | |
4d694b27 | 6460 | vec_perm_indices indices (sel, 1, count); |
e3342de4 RS |
6461 | perm_mask = vect_gen_perm_mask_checked (gs_info.offset_vectype, |
6462 | indices); | |
3bab6342 AT |
6463 | gcc_assert (perm_mask != NULL_TREE); |
6464 | } | |
4d694b27 | 6465 | else if (known_eq (nunits, scatter_off_nunits * 2)) |
3bab6342 | 6466 | { |
3bab6342 AT |
6467 | modifier = NARROW; |
6468 | ||
4d694b27 RS |
6469 | /* Currently gathers and scatters are only supported for |
6470 | fixed-length vectors. */ | |
6471 | unsigned int count = nunits.to_constant (); | |
6472 | vec_perm_builder sel (count, count, 1); | |
6473 | for (i = 0; i < (unsigned int) count; ++i) | |
6474 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 6475 | |
4d694b27 | 6476 | vec_perm_indices indices (sel, 2, count); |
e3342de4 | 6477 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); |
3bab6342 AT |
6478 | gcc_assert (perm_mask != NULL_TREE); |
6479 | ncopies *= 2; | |
6480 | } | |
6481 | else | |
6482 | gcc_unreachable (); | |
6483 | ||
134c85ca | 6484 | rettype = TREE_TYPE (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
6485 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
6486 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6487 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6488 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6489 | scaletype = TREE_VALUE (arglist); | |
6490 | ||
6491 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
6492 | && TREE_CODE (rettype) == VOID_TYPE); | |
6493 | ||
134c85ca | 6494 | ptr = fold_convert (ptrtype, gs_info.base); |
3bab6342 AT |
6495 | if (!is_gimple_min_invariant (ptr)) |
6496 | { | |
6497 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6498 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6499 | gcc_assert (!new_bb); | |
6500 | } | |
6501 | ||
6502 | /* Currently we support only unconditional scatter stores, | |
6503 | so mask should be all ones. */ | |
6504 | mask = build_int_cst (masktype, -1); | |
6505 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
6506 | ||
134c85ca | 6507 | scale = build_int_cst (scaletype, gs_info.scale); |
3bab6342 AT |
6508 | |
6509 | prev_stmt_info = NULL; | |
6510 | for (j = 0; j < ncopies; ++j) | |
6511 | { | |
6512 | if (j == 0) | |
6513 | { | |
6514 | src = vec_oprnd1 | |
c3a8f964 | 6515 | = vect_get_vec_def_for_operand (op, stmt); |
3bab6342 | 6516 | op = vec_oprnd0 |
134c85ca | 6517 | = vect_get_vec_def_for_operand (gs_info.offset, stmt); |
3bab6342 AT |
6518 | } |
6519 | else if (modifier != NONE && (j & 1)) | |
6520 | { | |
6521 | if (modifier == WIDEN) | |
6522 | { | |
6523 | src = vec_oprnd1 | |
929b4411 | 6524 | = vect_get_vec_def_for_stmt_copy (rhs_dt, vec_oprnd1); |
3bab6342 AT |
6525 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, |
6526 | stmt, gsi); | |
6527 | } | |
6528 | else if (modifier == NARROW) | |
6529 | { | |
6530 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
6531 | stmt, gsi); | |
6532 | op = vec_oprnd0 | |
134c85ca RS |
6533 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
6534 | vec_oprnd0); | |
3bab6342 AT |
6535 | } |
6536 | else | |
6537 | gcc_unreachable (); | |
6538 | } | |
6539 | else | |
6540 | { | |
6541 | src = vec_oprnd1 | |
929b4411 | 6542 | = vect_get_vec_def_for_stmt_copy (rhs_dt, vec_oprnd1); |
3bab6342 | 6543 | op = vec_oprnd0 |
134c85ca RS |
6544 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
6545 | vec_oprnd0); | |
3bab6342 AT |
6546 | } |
6547 | ||
6548 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
6549 | { | |
928686b1 RS |
6550 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)), |
6551 | TYPE_VECTOR_SUBPARTS (srctype))); | |
0e22bb5a | 6552 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 | 6553 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
e1bd7296 RS |
6554 | gassign *new_stmt |
6555 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
3bab6342 AT |
6556 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6557 | src = var; | |
6558 | } | |
6559 | ||
6560 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6561 | { | |
928686b1 RS |
6562 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), |
6563 | TYPE_VECTOR_SUBPARTS (idxtype))); | |
0e22bb5a | 6564 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 | 6565 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
e1bd7296 RS |
6566 | gassign *new_stmt |
6567 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
3bab6342 AT |
6568 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6569 | op = var; | |
6570 | } | |
6571 | ||
e1bd7296 | 6572 | gcall *new_stmt |
134c85ca | 6573 | = gimple_build_call (gs_info.decl, 5, ptr, mask, op, src, scale); |
e1bd7296 RS |
6574 | stmt_vec_info new_stmt_info |
6575 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3bab6342 | 6576 | |
dbe1b846 | 6577 | if (prev_stmt_info == NULL_STMT_VEC_INFO) |
e1bd7296 | 6578 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
3bab6342 | 6579 | else |
e1bd7296 RS |
6580 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
6581 | prev_stmt_info = new_stmt_info; | |
3bab6342 AT |
6582 | } |
6583 | return true; | |
6584 | } | |
6585 | ||
f307441a | 6586 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
bffb8014 | 6587 | DR_GROUP_STORE_COUNT (DR_GROUP_FIRST_ELEMENT (stmt_info))++; |
ebfd146a | 6588 | |
f307441a RS |
6589 | if (grouped_store) |
6590 | { | |
ebfd146a | 6591 | /* FORNOW */ |
a70d6342 | 6592 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
6593 | |
6594 | /* We vectorize all the stmts of the interleaving group when we | |
6595 | reach the last stmt in the group. */ | |
bffb8014 RS |
6596 | if (DR_GROUP_STORE_COUNT (first_stmt_info) |
6597 | < DR_GROUP_SIZE (first_stmt_info) | |
ebfd146a IR |
6598 | && !slp) |
6599 | { | |
6600 | *vec_stmt = NULL; | |
6601 | return true; | |
6602 | } | |
6603 | ||
6604 | if (slp) | |
4b5caab7 | 6605 | { |
0d0293ac | 6606 | grouped_store = false; |
4b5caab7 IR |
6607 | /* VEC_NUM is the number of vect stmts to be created for this |
6608 | group. */ | |
6609 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
bffb8014 RS |
6610 | first_stmt_info = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
6611 | gcc_assert (DR_GROUP_FIRST_ELEMENT (first_stmt_info) | |
6612 | == first_stmt_info); | |
6613 | first_dr = STMT_VINFO_DATA_REF (first_stmt_info); | |
6614 | op = vect_get_store_rhs (first_stmt_info); | |
4b5caab7 | 6615 | } |
ebfd146a | 6616 | else |
4b5caab7 IR |
6617 | /* VEC_NUM is the number of vect stmts to be created for this |
6618 | group. */ | |
ebfd146a | 6619 | vec_num = group_size; |
44fc7854 | 6620 | |
bffb8014 | 6621 | ref_type = get_group_alias_ptr_type (first_stmt_info); |
ebfd146a | 6622 | } |
b8698a0f | 6623 | else |
7cfb4d93 | 6624 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
b8698a0f | 6625 | |
73fbfcad | 6626 | if (dump_enabled_p ()) |
78c60e3d | 6627 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6628 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 6629 | |
2de001ee RS |
6630 | if (memory_access_type == VMAT_ELEMENTWISE |
6631 | || memory_access_type == VMAT_STRIDED_SLP) | |
f2e2a985 MM |
6632 | { |
6633 | gimple_stmt_iterator incr_gsi; | |
6634 | bool insert_after; | |
355fe088 | 6635 | gimple *incr; |
f2e2a985 MM |
6636 | tree offvar; |
6637 | tree ivstep; | |
6638 | tree running_off; | |
f2e2a985 MM |
6639 | tree stride_base, stride_step, alias_off; |
6640 | tree vec_oprnd; | |
f502d50e | 6641 | unsigned int g; |
4d694b27 RS |
6642 | /* Checked by get_load_store_type. */ |
6643 | unsigned int const_nunits = nunits.to_constant (); | |
f2e2a985 | 6644 | |
7cfb4d93 | 6645 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
f2e2a985 MM |
6646 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); |
6647 | ||
6648 | stride_base | |
6649 | = fold_build_pointer_plus | |
b210f45f | 6650 | (DR_BASE_ADDRESS (first_dr), |
f2e2a985 | 6651 | size_binop (PLUS_EXPR, |
b210f45f | 6652 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
44fc7854 | 6653 | convert_to_ptrofftype (DR_INIT (first_dr)))); |
b210f45f | 6654 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); |
f2e2a985 MM |
6655 | |
6656 | /* For a store with loop-invariant (but other than power-of-2) | |
6657 | stride (i.e. not a grouped access) like so: | |
6658 | ||
6659 | for (i = 0; i < n; i += stride) | |
6660 | array[i] = ...; | |
6661 | ||
6662 | we generate a new induction variable and new stores from | |
6663 | the components of the (vectorized) rhs: | |
6664 | ||
6665 | for (j = 0; ; j += VF*stride) | |
6666 | vectemp = ...; | |
6667 | tmp1 = vectemp[0]; | |
6668 | array[j] = tmp1; | |
6669 | tmp2 = vectemp[1]; | |
6670 | array[j + stride] = tmp2; | |
6671 | ... | |
6672 | */ | |
6673 | ||
4d694b27 | 6674 | unsigned nstores = const_nunits; |
b17dc4d4 | 6675 | unsigned lnel = 1; |
cee62fee | 6676 | tree ltype = elem_type; |
04199738 | 6677 | tree lvectype = vectype; |
cee62fee MM |
6678 | if (slp) |
6679 | { | |
4d694b27 RS |
6680 | if (group_size < const_nunits |
6681 | && const_nunits % group_size == 0) | |
b17dc4d4 | 6682 | { |
4d694b27 | 6683 | nstores = const_nunits / group_size; |
b17dc4d4 RB |
6684 | lnel = group_size; |
6685 | ltype = build_vector_type (elem_type, group_size); | |
04199738 RB |
6686 | lvectype = vectype; |
6687 | ||
6688 | /* First check if vec_extract optab doesn't support extraction | |
6689 | of vector elts directly. */ | |
b397965c | 6690 | scalar_mode elmode = SCALAR_TYPE_MODE (elem_type); |
9da15d40 RS |
6691 | machine_mode vmode; |
6692 | if (!mode_for_vector (elmode, group_size).exists (&vmode) | |
6693 | || !VECTOR_MODE_P (vmode) | |
414fef4e | 6694 | || !targetm.vector_mode_supported_p (vmode) |
04199738 RB |
6695 | || (convert_optab_handler (vec_extract_optab, |
6696 | TYPE_MODE (vectype), vmode) | |
6697 | == CODE_FOR_nothing)) | |
6698 | { | |
6699 | /* Try to avoid emitting an extract of vector elements | |
6700 | by performing the extracts using an integer type of the | |
6701 | same size, extracting from a vector of those and then | |
6702 | re-interpreting it as the original vector type if | |
6703 | supported. */ | |
6704 | unsigned lsize | |
6705 | = group_size * GET_MODE_BITSIZE (elmode); | |
fffbab82 | 6706 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 6707 | unsigned int lnunits = const_nunits / group_size; |
04199738 RB |
6708 | /* If we can't construct such a vector fall back to |
6709 | element extracts from the original vector type and | |
6710 | element size stores. */ | |
4d694b27 | 6711 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 6712 | && VECTOR_MODE_P (vmode) |
414fef4e | 6713 | && targetm.vector_mode_supported_p (vmode) |
04199738 RB |
6714 | && (convert_optab_handler (vec_extract_optab, |
6715 | vmode, elmode) | |
6716 | != CODE_FOR_nothing)) | |
6717 | { | |
4d694b27 | 6718 | nstores = lnunits; |
04199738 RB |
6719 | lnel = group_size; |
6720 | ltype = build_nonstandard_integer_type (lsize, 1); | |
6721 | lvectype = build_vector_type (ltype, nstores); | |
6722 | } | |
6723 | /* Else fall back to vector extraction anyway. | |
6724 | Fewer stores are more important than avoiding spilling | |
6725 | of the vector we extract from. Compared to the | |
6726 | construction case in vectorizable_load no store-forwarding | |
6727 | issue exists here for reasonable archs. */ | |
6728 | } | |
b17dc4d4 | 6729 | } |
4d694b27 RS |
6730 | else if (group_size >= const_nunits |
6731 | && group_size % const_nunits == 0) | |
b17dc4d4 RB |
6732 | { |
6733 | nstores = 1; | |
4d694b27 | 6734 | lnel = const_nunits; |
b17dc4d4 | 6735 | ltype = vectype; |
04199738 | 6736 | lvectype = vectype; |
b17dc4d4 | 6737 | } |
cee62fee MM |
6738 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); |
6739 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
6740 | } | |
6741 | ||
f2e2a985 MM |
6742 | ivstep = stride_step; |
6743 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
b17dc4d4 | 6744 | build_int_cst (TREE_TYPE (ivstep), vf)); |
f2e2a985 MM |
6745 | |
6746 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6747 | ||
b210f45f RB |
6748 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
6749 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); | |
f2e2a985 MM |
6750 | create_iv (stride_base, ivstep, NULL, |
6751 | loop, &incr_gsi, insert_after, | |
6752 | &offvar, NULL); | |
6753 | incr = gsi_stmt (incr_gsi); | |
4fbeb363 | 6754 | loop_vinfo->add_stmt (incr); |
f2e2a985 | 6755 | |
b210f45f | 6756 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
f2e2a985 MM |
6757 | |
6758 | prev_stmt_info = NULL; | |
44fc7854 | 6759 | alias_off = build_int_cst (ref_type, 0); |
bffb8014 | 6760 | stmt_vec_info next_stmt_info = first_stmt_info; |
f502d50e | 6761 | for (g = 0; g < group_size; g++) |
f2e2a985 | 6762 | { |
f502d50e MM |
6763 | running_off = offvar; |
6764 | if (g) | |
f2e2a985 | 6765 | { |
f502d50e MM |
6766 | tree size = TYPE_SIZE_UNIT (ltype); |
6767 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 6768 | size); |
f502d50e | 6769 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 6770 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 6771 | running_off, pos); |
f2e2a985 | 6772 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 6773 | running_off = newoff; |
f502d50e | 6774 | } |
b17dc4d4 RB |
6775 | unsigned int group_el = 0; |
6776 | unsigned HOST_WIDE_INT | |
6777 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
f502d50e MM |
6778 | for (j = 0; j < ncopies; j++) |
6779 | { | |
c3a8f964 | 6780 | /* We've set op and dt above, from vect_get_store_rhs, |
bffb8014 | 6781 | and first_stmt_info == stmt_info. */ |
f502d50e MM |
6782 | if (j == 0) |
6783 | { | |
6784 | if (slp) | |
6785 | { | |
6786 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
306b0c92 | 6787 | slp_node); |
f502d50e MM |
6788 | vec_oprnd = vec_oprnds[0]; |
6789 | } | |
6790 | else | |
6791 | { | |
bffb8014 RS |
6792 | op = vect_get_store_rhs (next_stmt_info); |
6793 | vec_oprnd = vect_get_vec_def_for_operand | |
6794 | (op, next_stmt_info); | |
f502d50e MM |
6795 | } |
6796 | } | |
f2e2a985 | 6797 | else |
f502d50e MM |
6798 | { |
6799 | if (slp) | |
6800 | vec_oprnd = vec_oprnds[j]; | |
6801 | else | |
c079cbac | 6802 | { |
894dd753 | 6803 | vect_is_simple_use (op, vinfo, &rhs_dt); |
929b4411 RS |
6804 | vec_oprnd = vect_get_vec_def_for_stmt_copy (rhs_dt, |
6805 | vec_oprnd); | |
c079cbac | 6806 | } |
f502d50e | 6807 | } |
04199738 RB |
6808 | /* Pun the vector to extract from if necessary. */ |
6809 | if (lvectype != vectype) | |
6810 | { | |
6811 | tree tem = make_ssa_name (lvectype); | |
6812 | gimple *pun | |
6813 | = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
6814 | lvectype, vec_oprnd)); | |
6815 | vect_finish_stmt_generation (stmt, pun, gsi); | |
6816 | vec_oprnd = tem; | |
6817 | } | |
f502d50e MM |
6818 | for (i = 0; i < nstores; i++) |
6819 | { | |
6820 | tree newref, newoff; | |
355fe088 | 6821 | gimple *incr, *assign; |
f502d50e MM |
6822 | tree size = TYPE_SIZE (ltype); |
6823 | /* Extract the i'th component. */ | |
6824 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
6825 | bitsize_int (i), size); | |
6826 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
6827 | size, pos); | |
6828 | ||
6829 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
6830 | NULL_TREE, true, | |
6831 | GSI_SAME_STMT); | |
6832 | ||
b17dc4d4 RB |
6833 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
6834 | group_el * elsz); | |
f502d50e | 6835 | newref = build2 (MEM_REF, ltype, |
b17dc4d4 | 6836 | running_off, this_off); |
19986382 | 6837 | vect_copy_ref_info (newref, DR_REF (first_dr)); |
f502d50e MM |
6838 | |
6839 | /* And store it to *running_off. */ | |
6840 | assign = gimple_build_assign (newref, elem); | |
e1bd7296 RS |
6841 | stmt_vec_info assign_info |
6842 | = vect_finish_stmt_generation (stmt, assign, gsi); | |
f502d50e | 6843 | |
b17dc4d4 RB |
6844 | group_el += lnel; |
6845 | if (! slp | |
6846 | || group_el == group_size) | |
6847 | { | |
6848 | newoff = copy_ssa_name (running_off, NULL); | |
6849 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6850 | running_off, stride_step); | |
6851 | vect_finish_stmt_generation (stmt, incr, gsi); | |
f502d50e | 6852 | |
b17dc4d4 RB |
6853 | running_off = newoff; |
6854 | group_el = 0; | |
6855 | } | |
225ce44b RB |
6856 | if (g == group_size - 1 |
6857 | && !slp) | |
f502d50e MM |
6858 | { |
6859 | if (j == 0 && i == 0) | |
225ce44b | 6860 | STMT_VINFO_VEC_STMT (stmt_info) |
e1bd7296 | 6861 | = *vec_stmt = assign_info; |
f502d50e | 6862 | else |
e1bd7296 RS |
6863 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign_info; |
6864 | prev_stmt_info = assign_info; | |
f502d50e MM |
6865 | } |
6866 | } | |
f2e2a985 | 6867 | } |
bffb8014 | 6868 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
b17dc4d4 RB |
6869 | if (slp) |
6870 | break; | |
f2e2a985 | 6871 | } |
778dd3b6 RB |
6872 | |
6873 | vec_oprnds.release (); | |
f2e2a985 MM |
6874 | return true; |
6875 | } | |
6876 | ||
8c681247 | 6877 | auto_vec<tree> dr_chain (group_size); |
9771b263 | 6878 | oprnds.create (group_size); |
ebfd146a | 6879 | |
720f5239 | 6880 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6881 | gcc_assert (alignment_support_scheme); |
70088b95 RS |
6882 | vec_loop_masks *loop_masks |
6883 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) | |
6884 | ? &LOOP_VINFO_MASKS (loop_vinfo) | |
6885 | : NULL); | |
272c6793 | 6886 | /* Targets with store-lane instructions must not require explicit |
c3a8f964 RS |
6887 | realignment. vect_supportable_dr_alignment always returns either |
6888 | dr_aligned or dr_unaligned_supported for masked operations. */ | |
7cfb4d93 RS |
6889 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES |
6890 | && !mask | |
70088b95 | 6891 | && !loop_masks) |
272c6793 RS |
6892 | || alignment_support_scheme == dr_aligned |
6893 | || alignment_support_scheme == dr_unaligned_supported); | |
6894 | ||
62da9e14 RS |
6895 | if (memory_access_type == VMAT_CONTIGUOUS_DOWN |
6896 | || memory_access_type == VMAT_CONTIGUOUS_REVERSE) | |
09dfa495 BM |
6897 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
6898 | ||
f307441a RS |
6899 | tree bump; |
6900 | tree vec_offset = NULL_TREE; | |
6901 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
6902 | { | |
6903 | aggr_type = NULL_TREE; | |
6904 | bump = NULL_TREE; | |
6905 | } | |
6906 | else if (memory_access_type == VMAT_GATHER_SCATTER) | |
6907 | { | |
6908 | aggr_type = elem_type; | |
6909 | vect_get_strided_load_store_ops (stmt, loop_vinfo, &gs_info, | |
6910 | &bump, &vec_offset); | |
6911 | } | |
272c6793 | 6912 | else |
f307441a RS |
6913 | { |
6914 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
6915 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
6916 | else | |
6917 | aggr_type = vectype; | |
6918 | bump = vect_get_data_ptr_increment (dr, aggr_type, memory_access_type); | |
6919 | } | |
ebfd146a | 6920 | |
c3a8f964 RS |
6921 | if (mask) |
6922 | LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true; | |
6923 | ||
ebfd146a IR |
6924 | /* In case the vectorization factor (VF) is bigger than the number |
6925 | of elements that we can fit in a vectype (nunits), we have to generate | |
6926 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 6927 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
6928 | vect_get_vec_def_for_copy_stmt. */ |
6929 | ||
0d0293ac | 6930 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6931 | |
6932 | S1: &base + 2 = x2 | |
6933 | S2: &base = x0 | |
6934 | S3: &base + 1 = x1 | |
6935 | S4: &base + 3 = x3 | |
6936 | ||
6937 | We create vectorized stores starting from base address (the access of the | |
6938 | first stmt in the chain (S2 in the above example), when the last store stmt | |
6939 | of the chain (S4) is reached: | |
6940 | ||
6941 | VS1: &base = vx2 | |
6942 | VS2: &base + vec_size*1 = vx0 | |
6943 | VS3: &base + vec_size*2 = vx1 | |
6944 | VS4: &base + vec_size*3 = vx3 | |
6945 | ||
6946 | Then permutation statements are generated: | |
6947 | ||
3fcc1b55 JJ |
6948 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
6949 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 6950 | ... |
b8698a0f | 6951 | |
ebfd146a IR |
6952 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
6953 | (the order of the data-refs in the output of vect_permute_store_chain | |
6954 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6955 | the documentation of vect_permute_store_chain()). | |
6956 | ||
6957 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 6958 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 6959 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 6960 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
6961 | */ |
6962 | ||
6963 | prev_stmt_info = NULL; | |
c3a8f964 | 6964 | tree vec_mask = NULL_TREE; |
ebfd146a IR |
6965 | for (j = 0; j < ncopies; j++) |
6966 | { | |
e1bd7296 | 6967 | stmt_vec_info new_stmt_info; |
ebfd146a IR |
6968 | if (j == 0) |
6969 | { | |
6970 | if (slp) | |
6971 | { | |
6972 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c | 6973 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
306b0c92 | 6974 | NULL, slp_node); |
ebfd146a | 6975 | |
9771b263 | 6976 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
6977 | } |
6978 | else | |
6979 | { | |
b8698a0f L |
6980 | /* For interleaved stores we collect vectorized defs for all the |
6981 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
6982 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
6983 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
6984 | ||
2c53b149 | 6985 | If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 6986 | OPRNDS are of size 1. */ |
bffb8014 | 6987 | stmt_vec_info next_stmt_info = first_stmt_info; |
ebfd146a IR |
6988 | for (i = 0; i < group_size; i++) |
6989 | { | |
b8698a0f | 6990 | /* Since gaps are not supported for interleaved stores, |
2c53b149 | 6991 | DR_GROUP_SIZE is the exact number of stmts in the chain. |
bffb8014 RS |
6992 | Therefore, NEXT_STMT_INFO can't be NULL_TREE. In case |
6993 | that there is no interleaving, DR_GROUP_SIZE is 1, | |
6994 | and only one iteration of the loop will be executed. */ | |
6995 | op = vect_get_store_rhs (next_stmt_info); | |
6996 | vec_oprnd = vect_get_vec_def_for_operand | |
6997 | (op, next_stmt_info); | |
9771b263 DN |
6998 | dr_chain.quick_push (vec_oprnd); |
6999 | oprnds.quick_push (vec_oprnd); | |
bffb8014 | 7000 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
ebfd146a | 7001 | } |
c3a8f964 RS |
7002 | if (mask) |
7003 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, | |
7004 | mask_vectype); | |
ebfd146a IR |
7005 | } |
7006 | ||
7007 | /* We should have catched mismatched types earlier. */ | |
7008 | gcc_assert (useless_type_conversion_p (vectype, | |
7009 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
7010 | bool simd_lane_access_p |
7011 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
7012 | if (simd_lane_access_p | |
7013 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
7014 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
7015 | && integer_zerop (DR_OFFSET (first_dr)) | |
7016 | && integer_zerop (DR_INIT (first_dr)) | |
7017 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 7018 | get_alias_set (TREE_TYPE (ref_type)))) |
74bf76ed JJ |
7019 | { |
7020 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 7021 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 7022 | inv_p = false; |
74bf76ed | 7023 | } |
f307441a RS |
7024 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
7025 | { | |
7026 | vect_get_gather_scatter_ops (loop, stmt, &gs_info, | |
7027 | &dataref_ptr, &vec_offset); | |
7028 | inv_p = false; | |
7029 | } | |
74bf76ed JJ |
7030 | else |
7031 | dataref_ptr | |
bffb8014 | 7032 | = vect_create_data_ref_ptr (first_stmt_info, aggr_type, |
74bf76ed | 7033 | simd_lane_access_p ? loop : NULL, |
09dfa495 | 7034 | offset, &dummy, gsi, &ptr_incr, |
f307441a RS |
7035 | simd_lane_access_p, &inv_p, |
7036 | NULL_TREE, bump); | |
a70d6342 | 7037 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 7038 | } |
b8698a0f | 7039 | else |
ebfd146a | 7040 | { |
b8698a0f L |
7041 | /* For interleaved stores we created vectorized defs for all the |
7042 | defs stored in OPRNDS in the previous iteration (previous copy). | |
7043 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
7044 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
7045 | next copy. | |
2c53b149 | 7046 | If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
7047 | OPRNDS are of size 1. */ |
7048 | for (i = 0; i < group_size; i++) | |
7049 | { | |
9771b263 | 7050 | op = oprnds[i]; |
894dd753 | 7051 | vect_is_simple_use (op, vinfo, &rhs_dt); |
929b4411 | 7052 | vec_oprnd = vect_get_vec_def_for_stmt_copy (rhs_dt, op); |
9771b263 DN |
7053 | dr_chain[i] = vec_oprnd; |
7054 | oprnds[i] = vec_oprnd; | |
ebfd146a | 7055 | } |
c3a8f964 | 7056 | if (mask) |
929b4411 | 7057 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
74bf76ed JJ |
7058 | if (dataref_offset) |
7059 | dataref_offset | |
f307441a RS |
7060 | = int_const_binop (PLUS_EXPR, dataref_offset, bump); |
7061 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
929b4411 RS |
7062 | vec_offset = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
7063 | vec_offset); | |
74bf76ed JJ |
7064 | else |
7065 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
f307441a | 7066 | bump); |
ebfd146a IR |
7067 | } |
7068 | ||
2de001ee | 7069 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 7070 | { |
272c6793 | 7071 | tree vec_array; |
267d3070 | 7072 | |
3ba4ff41 | 7073 | /* Get an array into which we can store the individual vectors. */ |
272c6793 | 7074 | vec_array = create_vector_array (vectype, vec_num); |
3ba4ff41 RS |
7075 | |
7076 | /* Invalidate the current contents of VEC_ARRAY. This should | |
7077 | become an RTL clobber too, which prevents the vector registers | |
7078 | from being upward-exposed. */ | |
7079 | vect_clobber_variable (stmt, gsi, vec_array); | |
7080 | ||
7081 | /* Store the individual vectors into the array. */ | |
272c6793 | 7082 | for (i = 0; i < vec_num; i++) |
c2d7ab2a | 7083 | { |
9771b263 | 7084 | vec_oprnd = dr_chain[i]; |
272c6793 | 7085 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 7086 | } |
b8698a0f | 7087 | |
7cfb4d93 | 7088 | tree final_mask = NULL; |
70088b95 RS |
7089 | if (loop_masks) |
7090 | final_mask = vect_get_loop_mask (gsi, loop_masks, ncopies, | |
7091 | vectype, j); | |
7cfb4d93 RS |
7092 | if (vec_mask) |
7093 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
7094 | vec_mask, gsi); | |
7095 | ||
7e11fc7f | 7096 | gcall *call; |
7cfb4d93 | 7097 | if (final_mask) |
7e11fc7f RS |
7098 | { |
7099 | /* Emit: | |
7100 | MASK_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK, | |
7101 | VEC_ARRAY). */ | |
7102 | unsigned int align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype)); | |
7103 | tree alias_ptr = build_int_cst (ref_type, align); | |
7104 | call = gimple_build_call_internal (IFN_MASK_STORE_LANES, 4, | |
7105 | dataref_ptr, alias_ptr, | |
7cfb4d93 | 7106 | final_mask, vec_array); |
7e11fc7f RS |
7107 | } |
7108 | else | |
7109 | { | |
7110 | /* Emit: | |
7111 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
7112 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); | |
7113 | call = gimple_build_call_internal (IFN_STORE_LANES, 1, | |
7114 | vec_array); | |
7115 | gimple_call_set_lhs (call, data_ref); | |
7116 | } | |
a844293d | 7117 | gimple_call_set_nothrow (call, true); |
e1bd7296 | 7118 | new_stmt_info = vect_finish_stmt_generation (stmt, call, gsi); |
3ba4ff41 RS |
7119 | |
7120 | /* Record that VEC_ARRAY is now dead. */ | |
7121 | vect_clobber_variable (stmt, gsi, vec_array); | |
272c6793 RS |
7122 | } |
7123 | else | |
7124 | { | |
e1bd7296 | 7125 | new_stmt_info = NULL; |
0d0293ac | 7126 | if (grouped_store) |
272c6793 | 7127 | { |
b6b9227d JJ |
7128 | if (j == 0) |
7129 | result_chain.create (group_size); | |
272c6793 RS |
7130 | /* Permute. */ |
7131 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
7132 | &result_chain); | |
7133 | } | |
c2d7ab2a | 7134 | |
bffb8014 | 7135 | stmt_vec_info next_stmt_info = first_stmt_info; |
272c6793 RS |
7136 | for (i = 0; i < vec_num; i++) |
7137 | { | |
644ffefd | 7138 | unsigned align, misalign; |
272c6793 | 7139 | |
7cfb4d93 | 7140 | tree final_mask = NULL_TREE; |
70088b95 RS |
7141 | if (loop_masks) |
7142 | final_mask = vect_get_loop_mask (gsi, loop_masks, | |
7143 | vec_num * ncopies, | |
7cfb4d93 RS |
7144 | vectype, vec_num * j + i); |
7145 | if (vec_mask) | |
7146 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
7147 | vec_mask, gsi); | |
7148 | ||
f307441a RS |
7149 | if (memory_access_type == VMAT_GATHER_SCATTER) |
7150 | { | |
7151 | tree scale = size_int (gs_info.scale); | |
7152 | gcall *call; | |
70088b95 | 7153 | if (loop_masks) |
f307441a RS |
7154 | call = gimple_build_call_internal |
7155 | (IFN_MASK_SCATTER_STORE, 5, dataref_ptr, vec_offset, | |
7156 | scale, vec_oprnd, final_mask); | |
7157 | else | |
7158 | call = gimple_build_call_internal | |
7159 | (IFN_SCATTER_STORE, 4, dataref_ptr, vec_offset, | |
7160 | scale, vec_oprnd); | |
7161 | gimple_call_set_nothrow (call, true); | |
e1bd7296 RS |
7162 | new_stmt_info |
7163 | = vect_finish_stmt_generation (stmt, call, gsi); | |
f307441a RS |
7164 | break; |
7165 | } | |
7166 | ||
272c6793 RS |
7167 | if (i > 0) |
7168 | /* Bump the vector pointer. */ | |
7169 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
f307441a | 7170 | stmt, bump); |
272c6793 RS |
7171 | |
7172 | if (slp) | |
9771b263 | 7173 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
7174 | else if (grouped_store) |
7175 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 7176 | vect_permute_store_chain(). */ |
9771b263 | 7177 | vec_oprnd = result_chain[i]; |
272c6793 | 7178 | |
f702e7d4 | 7179 | align = DR_TARGET_ALIGNMENT (first_dr); |
272c6793 | 7180 | if (aligned_access_p (first_dr)) |
644ffefd | 7181 | misalign = 0; |
272c6793 RS |
7182 | else if (DR_MISALIGNMENT (first_dr) == -1) |
7183 | { | |
25f68d90 | 7184 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 7185 | misalign = 0; |
272c6793 RS |
7186 | } |
7187 | else | |
c3a8f964 | 7188 | misalign = DR_MISALIGNMENT (first_dr); |
aed93b23 RB |
7189 | if (dataref_offset == NULL_TREE |
7190 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
7191 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
7192 | misalign); | |
c2d7ab2a | 7193 | |
62da9e14 | 7194 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
09dfa495 BM |
7195 | { |
7196 | tree perm_mask = perm_mask_for_reverse (vectype); | |
7197 | tree perm_dest | |
c3a8f964 | 7198 | = vect_create_destination_var (vect_get_store_rhs (stmt), |
09dfa495 | 7199 | vectype); |
b731b390 | 7200 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
7201 | |
7202 | /* Generate the permute statement. */ | |
355fe088 | 7203 | gimple *perm_stmt |
0d0e4a03 JJ |
7204 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
7205 | vec_oprnd, perm_mask); | |
09dfa495 BM |
7206 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
7207 | ||
7208 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
7209 | vec_oprnd = new_temp; | |
7210 | } | |
7211 | ||
272c6793 | 7212 | /* Arguments are ready. Create the new vector stmt. */ |
7cfb4d93 | 7213 | if (final_mask) |
c3a8f964 RS |
7214 | { |
7215 | align = least_bit_hwi (misalign | align); | |
7216 | tree ptr = build_int_cst (ref_type, align); | |
7217 | gcall *call | |
7218 | = gimple_build_call_internal (IFN_MASK_STORE, 4, | |
7219 | dataref_ptr, ptr, | |
7cfb4d93 | 7220 | final_mask, vec_oprnd); |
c3a8f964 | 7221 | gimple_call_set_nothrow (call, true); |
e1bd7296 RS |
7222 | new_stmt_info |
7223 | = vect_finish_stmt_generation (stmt, call, gsi); | |
c3a8f964 RS |
7224 | } |
7225 | else | |
7226 | { | |
7227 | data_ref = fold_build2 (MEM_REF, vectype, | |
7228 | dataref_ptr, | |
7229 | dataref_offset | |
7230 | ? dataref_offset | |
7231 | : build_int_cst (ref_type, 0)); | |
7232 | if (aligned_access_p (first_dr)) | |
7233 | ; | |
7234 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
7235 | TREE_TYPE (data_ref) | |
7236 | = build_aligned_type (TREE_TYPE (data_ref), | |
7237 | align * BITS_PER_UNIT); | |
7238 | else | |
7239 | TREE_TYPE (data_ref) | |
7240 | = build_aligned_type (TREE_TYPE (data_ref), | |
7241 | TYPE_ALIGN (elem_type)); | |
19986382 | 7242 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); |
e1bd7296 RS |
7243 | gassign *new_stmt |
7244 | = gimple_build_assign (data_ref, vec_oprnd); | |
7245 | new_stmt_info | |
7246 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
c3a8f964 | 7247 | } |
272c6793 RS |
7248 | |
7249 | if (slp) | |
7250 | continue; | |
7251 | ||
bffb8014 RS |
7252 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
7253 | if (!next_stmt_info) | |
272c6793 RS |
7254 | break; |
7255 | } | |
ebfd146a | 7256 | } |
1da0876c RS |
7257 | if (!slp) |
7258 | { | |
7259 | if (j == 0) | |
e1bd7296 | 7260 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
1da0876c | 7261 | else |
e1bd7296 RS |
7262 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
7263 | prev_stmt_info = new_stmt_info; | |
1da0876c | 7264 | } |
ebfd146a IR |
7265 | } |
7266 | ||
9771b263 DN |
7267 | oprnds.release (); |
7268 | result_chain.release (); | |
7269 | vec_oprnds.release (); | |
ebfd146a IR |
7270 | |
7271 | return true; | |
7272 | } | |
7273 | ||
557be5a8 AL |
7274 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
7275 | VECTOR_CST mask. No checks are made that the target platform supports the | |
7ac7e286 | 7276 | mask, so callers may wish to test can_vec_perm_const_p separately, or use |
557be5a8 | 7277 | vect_gen_perm_mask_checked. */ |
a1e53f3f | 7278 | |
3fcc1b55 | 7279 | tree |
4aae3cb3 | 7280 | vect_gen_perm_mask_any (tree vectype, const vec_perm_indices &sel) |
a1e53f3f | 7281 | { |
b00cb3bf | 7282 | tree mask_type; |
a1e53f3f | 7283 | |
0ecc2b7d RS |
7284 | poly_uint64 nunits = sel.length (); |
7285 | gcc_assert (known_eq (nunits, TYPE_VECTOR_SUBPARTS (vectype))); | |
b00cb3bf RS |
7286 | |
7287 | mask_type = build_vector_type (ssizetype, nunits); | |
736d0f28 | 7288 | return vec_perm_indices_to_tree (mask_type, sel); |
a1e53f3f L |
7289 | } |
7290 | ||
7ac7e286 | 7291 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_const_p, |
cf7aa6a3 | 7292 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ |
557be5a8 AL |
7293 | |
7294 | tree | |
4aae3cb3 | 7295 | vect_gen_perm_mask_checked (tree vectype, const vec_perm_indices &sel) |
557be5a8 | 7296 | { |
7ac7e286 | 7297 | gcc_assert (can_vec_perm_const_p (TYPE_MODE (vectype), sel)); |
557be5a8 AL |
7298 | return vect_gen_perm_mask_any (vectype, sel); |
7299 | } | |
7300 | ||
aec7ae7d JJ |
7301 | /* Given a vector variable X and Y, that was generated for the scalar |
7302 | STMT, generate instructions to permute the vector elements of X and Y | |
7303 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
7304 | permuted vector variable. */ | |
a1e53f3f L |
7305 | |
7306 | static tree | |
355fe088 | 7307 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 7308 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
7309 | { |
7310 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 7311 | tree perm_dest, data_ref; |
355fe088 | 7312 | gimple *perm_stmt; |
a1e53f3f | 7313 | |
7ad429a4 RS |
7314 | tree scalar_dest = gimple_get_lhs (stmt); |
7315 | if (TREE_CODE (scalar_dest) == SSA_NAME) | |
7316 | perm_dest = vect_create_destination_var (scalar_dest, vectype); | |
7317 | else | |
7318 | perm_dest = vect_get_new_vect_var (vectype, vect_simple_var, NULL); | |
b731b390 | 7319 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
7320 | |
7321 | /* Generate the permute statement. */ | |
0d0e4a03 | 7322 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
7323 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
7324 | ||
7325 | return data_ref; | |
7326 | } | |
7327 | ||
6b916b36 RB |
7328 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
7329 | inserting them on the loops preheader edge. Returns true if we | |
7330 | were successful in doing so (and thus STMT can be moved then), | |
7331 | otherwise returns false. */ | |
7332 | ||
7333 | static bool | |
355fe088 | 7334 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
7335 | { |
7336 | ssa_op_iter i; | |
7337 | tree op; | |
7338 | bool any = false; | |
7339 | ||
7340 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
7341 | { | |
355fe088 | 7342 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
7343 | if (!gimple_nop_p (def_stmt) |
7344 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
7345 | { | |
7346 | /* Make sure we don't need to recurse. While we could do | |
7347 | so in simple cases when there are more complex use webs | |
7348 | we don't have an easy way to preserve stmt order to fulfil | |
7349 | dependencies within them. */ | |
7350 | tree op2; | |
7351 | ssa_op_iter i2; | |
d1417442 JJ |
7352 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
7353 | return false; | |
6b916b36 RB |
7354 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
7355 | { | |
355fe088 | 7356 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
7357 | if (!gimple_nop_p (def_stmt2) |
7358 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
7359 | return false; | |
7360 | } | |
7361 | any = true; | |
7362 | } | |
7363 | } | |
7364 | ||
7365 | if (!any) | |
7366 | return true; | |
7367 | ||
7368 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
7369 | { | |
355fe088 | 7370 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
7371 | if (!gimple_nop_p (def_stmt) |
7372 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
7373 | { | |
7374 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
7375 | gsi_remove (&gsi, false); | |
7376 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
7377 | } | |
7378 | } | |
7379 | ||
7380 | return true; | |
7381 | } | |
7382 | ||
ebfd146a IR |
7383 | /* vectorizable_load. |
7384 | ||
b8698a0f L |
7385 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
7386 | can be vectorized. | |
7387 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
7388 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
7389 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7390 | ||
7391 | static bool | |
1eede195 RS |
7392 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, |
7393 | stmt_vec_info *vec_stmt, slp_tree slp_node, | |
7394 | slp_instance slp_node_instance, | |
68435eb2 | 7395 | stmt_vector_for_cost *cost_vec) |
ebfd146a IR |
7396 | { |
7397 | tree scalar_dest; | |
7398 | tree vec_dest = NULL; | |
7399 | tree data_ref = NULL; | |
7400 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 7401 | stmt_vec_info prev_stmt_info; |
ebfd146a | 7402 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 7403 | struct loop *loop = NULL; |
ebfd146a | 7404 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 7405 | bool nested_in_vect_loop = false; |
c716e67f | 7406 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 7407 | tree elem_type; |
ebfd146a | 7408 | tree new_temp; |
ef4bddc2 | 7409 | machine_mode mode; |
ebfd146a IR |
7410 | tree dummy; |
7411 | enum dr_alignment_support alignment_support_scheme; | |
7412 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 7413 | tree dataref_offset = NULL_TREE; |
355fe088 | 7414 | gimple *ptr_incr = NULL; |
ebfd146a | 7415 | int ncopies; |
4d694b27 RS |
7416 | int i, j; |
7417 | unsigned int group_size; | |
7418 | poly_uint64 group_gap_adj; | |
ebfd146a IR |
7419 | tree msq = NULL_TREE, lsq; |
7420 | tree offset = NULL_TREE; | |
356bbc4c | 7421 | tree byte_offset = NULL_TREE; |
ebfd146a | 7422 | tree realignment_token = NULL_TREE; |
538dd0b7 | 7423 | gphi *phi = NULL; |
6e1aa848 | 7424 | vec<tree> dr_chain = vNULL; |
0d0293ac | 7425 | bool grouped_load = false; |
bffb8014 | 7426 | stmt_vec_info first_stmt_info; |
b9787581 | 7427 | stmt_vec_info first_stmt_info_for_drptr = NULL; |
ebfd146a IR |
7428 | bool inv_p; |
7429 | bool compute_in_loop = false; | |
7430 | struct loop *at_loop; | |
7431 | int vec_num; | |
7432 | bool slp = (slp_node != NULL); | |
7433 | bool slp_perm = false; | |
a70d6342 | 7434 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
d9f21f6a | 7435 | poly_uint64 vf; |
272c6793 | 7436 | tree aggr_type; |
134c85ca | 7437 | gather_scatter_info gs_info; |
310213d4 | 7438 | vec_info *vinfo = stmt_info->vinfo; |
44fc7854 | 7439 | tree ref_type; |
929b4411 | 7440 | enum vect_def_type mask_dt = vect_unknown_def_type; |
a70d6342 | 7441 | |
465c8c19 JJ |
7442 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
7443 | return false; | |
7444 | ||
66c16fd9 RB |
7445 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7446 | && ! vec_stmt) | |
465c8c19 JJ |
7447 | return false; |
7448 | ||
c3a8f964 RS |
7449 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
7450 | if (is_gimple_assign (stmt)) | |
7451 | { | |
7452 | scalar_dest = gimple_assign_lhs (stmt); | |
7453 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
7454 | return false; | |
465c8c19 | 7455 | |
c3a8f964 RS |
7456 | tree_code code = gimple_assign_rhs_code (stmt); |
7457 | if (code != ARRAY_REF | |
7458 | && code != BIT_FIELD_REF | |
7459 | && code != INDIRECT_REF | |
7460 | && code != COMPONENT_REF | |
7461 | && code != IMAGPART_EXPR | |
7462 | && code != REALPART_EXPR | |
7463 | && code != MEM_REF | |
7464 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
7465 | return false; | |
7466 | } | |
7467 | else | |
7468 | { | |
7469 | gcall *call = dyn_cast <gcall *> (stmt); | |
bfaa08b7 RS |
7470 | if (!call || !gimple_call_internal_p (call)) |
7471 | return false; | |
7472 | ||
7473 | internal_fn ifn = gimple_call_internal_fn (call); | |
7474 | if (!internal_load_fn_p (ifn)) | |
c3a8f964 | 7475 | return false; |
465c8c19 | 7476 | |
c3a8f964 RS |
7477 | scalar_dest = gimple_call_lhs (call); |
7478 | if (!scalar_dest) | |
7479 | return false; | |
7480 | ||
7481 | if (slp_node != NULL) | |
7482 | { | |
7483 | if (dump_enabled_p ()) | |
7484 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7485 | "SLP of masked loads not supported.\n"); | |
7486 | return false; | |
7487 | } | |
7488 | ||
bfaa08b7 RS |
7489 | int mask_index = internal_fn_mask_index (ifn); |
7490 | if (mask_index >= 0) | |
7491 | { | |
7492 | mask = gimple_call_arg (call, mask_index); | |
929b4411 RS |
7493 | if (!vect_check_load_store_mask (stmt, mask, &mask_dt, |
7494 | &mask_vectype)) | |
bfaa08b7 RS |
7495 | return false; |
7496 | } | |
c3a8f964 | 7497 | } |
465c8c19 JJ |
7498 | |
7499 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
7500 | return false; | |
7501 | ||
7502 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
4d694b27 | 7503 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 | 7504 | |
a70d6342 IR |
7505 | if (loop_vinfo) |
7506 | { | |
7507 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
7508 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
7509 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
7510 | } | |
7511 | else | |
3533e503 | 7512 | vf = 1; |
ebfd146a IR |
7513 | |
7514 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 7515 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 7516 | case of SLP. */ |
fce57248 | 7517 | if (slp) |
ebfd146a IR |
7518 | ncopies = 1; |
7519 | else | |
e8f142e2 | 7520 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
ebfd146a IR |
7521 | |
7522 | gcc_assert (ncopies >= 1); | |
7523 | ||
7524 | /* FORNOW. This restriction should be relaxed. */ | |
7525 | if (nested_in_vect_loop && ncopies > 1) | |
7526 | { | |
73fbfcad | 7527 | if (dump_enabled_p ()) |
78c60e3d | 7528 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7529 | "multiple types in nested loop.\n"); |
ebfd146a IR |
7530 | return false; |
7531 | } | |
7532 | ||
f2556b68 RB |
7533 | /* Invalidate assumptions made by dependence analysis when vectorization |
7534 | on the unrolled body effectively re-orders stmts. */ | |
7535 | if (ncopies > 1 | |
7536 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
7537 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
7538 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
7539 | { |
7540 | if (dump_enabled_p ()) | |
7541 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7542 | "cannot perform implicit CSE when unrolling " | |
7543 | "with negative dependence distance\n"); | |
7544 | return false; | |
7545 | } | |
7546 | ||
7b7b1813 | 7547 | elem_type = TREE_TYPE (vectype); |
947131ba | 7548 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
7549 | |
7550 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
7551 | (e.g. - data copies). */ | |
947131ba | 7552 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 7553 | { |
73fbfcad | 7554 | if (dump_enabled_p ()) |
78c60e3d | 7555 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7556 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
7557 | return false; |
7558 | } | |
7559 | ||
ebfd146a | 7560 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 7561 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 7562 | { |
0d0293ac | 7563 | grouped_load = true; |
ebfd146a | 7564 | /* FORNOW */ |
2de001ee RS |
7565 | gcc_assert (!nested_in_vect_loop); |
7566 | gcc_assert (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)); | |
ebfd146a | 7567 | |
bffb8014 RS |
7568 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
7569 | group_size = DR_GROUP_SIZE (first_stmt_info); | |
d5f035ea | 7570 | |
b1af7da6 RB |
7571 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
7572 | slp_perm = true; | |
7573 | ||
f2556b68 RB |
7574 | /* Invalidate assumptions made by dependence analysis when vectorization |
7575 | on the unrolled body effectively re-orders stmts. */ | |
7576 | if (!PURE_SLP_STMT (stmt_info) | |
7577 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
7578 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
7579 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
7580 | { |
7581 | if (dump_enabled_p ()) | |
7582 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7583 | "cannot perform implicit CSE when performing " | |
7584 | "group loads with negative dependence distance\n"); | |
7585 | return false; | |
7586 | } | |
96bb56b2 RB |
7587 | |
7588 | /* Similarly when the stmt is a load that is both part of a SLP | |
7589 | instance and a loop vectorized stmt via the same-dr mechanism | |
7590 | we have to give up. */ | |
2c53b149 | 7591 | if (DR_GROUP_SAME_DR_STMT (stmt_info) |
96bb56b2 | 7592 | && (STMT_SLP_TYPE (stmt_info) |
c26228d4 | 7593 | != STMT_SLP_TYPE (DR_GROUP_SAME_DR_STMT (stmt_info)))) |
96bb56b2 RB |
7594 | { |
7595 | if (dump_enabled_p ()) | |
7596 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7597 | "conflicting SLP types for CSEd load\n"); | |
7598 | return false; | |
7599 | } | |
ebfd146a | 7600 | } |
7cfb4d93 RS |
7601 | else |
7602 | group_size = 1; | |
ebfd146a | 7603 | |
2de001ee | 7604 | vect_memory_access_type memory_access_type; |
7e11fc7f | 7605 | if (!get_load_store_type (stmt, vectype, slp, mask, VLS_LOAD, ncopies, |
2de001ee RS |
7606 | &memory_access_type, &gs_info)) |
7607 | return false; | |
a1e53f3f | 7608 | |
c3a8f964 RS |
7609 | if (mask) |
7610 | { | |
7611 | if (memory_access_type == VMAT_CONTIGUOUS) | |
7612 | { | |
7e11fc7f RS |
7613 | machine_mode vec_mode = TYPE_MODE (vectype); |
7614 | if (!VECTOR_MODE_P (vec_mode) | |
7615 | || !can_vec_mask_load_store_p (vec_mode, | |
c3a8f964 RS |
7616 | TYPE_MODE (mask_vectype), true)) |
7617 | return false; | |
7618 | } | |
bfaa08b7 | 7619 | else if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
c3a8f964 RS |
7620 | { |
7621 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); | |
7622 | tree masktype | |
7623 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
7624 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
7625 | { | |
7626 | if (dump_enabled_p ()) | |
7627 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7628 | "masked gather with integer mask not" | |
7629 | " supported."); | |
7630 | return false; | |
7631 | } | |
7632 | } | |
bfaa08b7 RS |
7633 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
7634 | && memory_access_type != VMAT_GATHER_SCATTER) | |
c3a8f964 RS |
7635 | { |
7636 | if (dump_enabled_p ()) | |
7637 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7638 | "unsupported access type for masked load.\n"); | |
7639 | return false; | |
7640 | } | |
7641 | } | |
7642 | ||
ebfd146a IR |
7643 | if (!vec_stmt) /* transformation not required. */ |
7644 | { | |
2de001ee RS |
7645 | if (!slp) |
7646 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; | |
7cfb4d93 RS |
7647 | |
7648 | if (loop_vinfo | |
7649 | && LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) | |
7650 | check_load_store_masking (loop_vinfo, vectype, VLS_LOAD, group_size, | |
bfaa08b7 | 7651 | memory_access_type, &gs_info); |
7cfb4d93 | 7652 | |
ebfd146a | 7653 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; |
68435eb2 RB |
7654 | vect_model_load_cost (stmt_info, ncopies, memory_access_type, |
7655 | slp_node_instance, slp_node, cost_vec); | |
ebfd146a IR |
7656 | return true; |
7657 | } | |
7658 | ||
2de001ee RS |
7659 | if (!slp) |
7660 | gcc_assert (memory_access_type | |
7661 | == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); | |
7662 | ||
73fbfcad | 7663 | if (dump_enabled_p ()) |
78c60e3d | 7664 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 7665 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a | 7666 | |
67b8dbac | 7667 | /* Transform. */ |
ebfd146a | 7668 | |
f702e7d4 | 7669 | ensure_base_align (dr); |
c716e67f | 7670 | |
bfaa08b7 | 7671 | if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
aec7ae7d | 7672 | { |
929b4411 RS |
7673 | vect_build_gather_load_calls (stmt, gsi, vec_stmt, &gs_info, mask, |
7674 | mask_dt); | |
aec7ae7d JJ |
7675 | return true; |
7676 | } | |
2de001ee RS |
7677 | |
7678 | if (memory_access_type == VMAT_ELEMENTWISE | |
7679 | || memory_access_type == VMAT_STRIDED_SLP) | |
7d75abc8 MM |
7680 | { |
7681 | gimple_stmt_iterator incr_gsi; | |
7682 | bool insert_after; | |
355fe088 | 7683 | gimple *incr; |
7d75abc8 | 7684 | tree offvar; |
7d75abc8 MM |
7685 | tree ivstep; |
7686 | tree running_off; | |
9771b263 | 7687 | vec<constructor_elt, va_gc> *v = NULL; |
14ac6aa2 | 7688 | tree stride_base, stride_step, alias_off; |
4d694b27 RS |
7689 | /* Checked by get_load_store_type. */ |
7690 | unsigned int const_nunits = nunits.to_constant (); | |
b210f45f | 7691 | unsigned HOST_WIDE_INT cst_offset = 0; |
14ac6aa2 | 7692 | |
7cfb4d93 | 7693 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
14ac6aa2 | 7694 | gcc_assert (!nested_in_vect_loop); |
7d75abc8 | 7695 | |
b210f45f | 7696 | if (grouped_load) |
44fc7854 | 7697 | { |
bffb8014 RS |
7698 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
7699 | first_dr = STMT_VINFO_DATA_REF (first_stmt_info); | |
44fc7854 | 7700 | } |
ab313a8c | 7701 | else |
44fc7854 | 7702 | { |
bffb8014 | 7703 | first_stmt_info = stmt_info; |
44fc7854 | 7704 | first_dr = dr; |
b210f45f RB |
7705 | } |
7706 | if (slp && grouped_load) | |
7707 | { | |
bffb8014 RS |
7708 | group_size = DR_GROUP_SIZE (first_stmt_info); |
7709 | ref_type = get_group_alias_ptr_type (first_stmt_info); | |
b210f45f RB |
7710 | } |
7711 | else | |
7712 | { | |
7713 | if (grouped_load) | |
7714 | cst_offset | |
7715 | = (tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))) | |
bffb8014 RS |
7716 | * vect_get_place_in_interleaving_chain (stmt, |
7717 | first_stmt_info)); | |
44fc7854 | 7718 | group_size = 1; |
b210f45f | 7719 | ref_type = reference_alias_ptr_type (DR_REF (dr)); |
44fc7854 | 7720 | } |
ab313a8c | 7721 | |
14ac6aa2 RB |
7722 | stride_base |
7723 | = fold_build_pointer_plus | |
ab313a8c | 7724 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 7725 | size_binop (PLUS_EXPR, |
ab313a8c RB |
7726 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
7727 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
7728 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
7729 | |
7730 | /* For a load with loop-invariant (but other than power-of-2) | |
7731 | stride (i.e. not a grouped access) like so: | |
7732 | ||
7733 | for (i = 0; i < n; i += stride) | |
7734 | ... = array[i]; | |
7735 | ||
7736 | we generate a new induction variable and new accesses to | |
7737 | form a new vector (or vectors, depending on ncopies): | |
7738 | ||
7739 | for (j = 0; ; j += VF*stride) | |
7740 | tmp1 = array[j]; | |
7741 | tmp2 = array[j + stride]; | |
7742 | ... | |
7743 | vectemp = {tmp1, tmp2, ...} | |
7744 | */ | |
7745 | ||
ab313a8c RB |
7746 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
7747 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
7748 | |
7749 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
7750 | ||
b210f45f RB |
7751 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
7752 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); | |
7753 | create_iv (stride_base, ivstep, NULL, | |
7d75abc8 MM |
7754 | loop, &incr_gsi, insert_after, |
7755 | &offvar, NULL); | |
7756 | incr = gsi_stmt (incr_gsi); | |
4fbeb363 | 7757 | loop_vinfo->add_stmt (incr); |
7d75abc8 | 7758 | |
b210f45f | 7759 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
7d75abc8 MM |
7760 | |
7761 | prev_stmt_info = NULL; | |
7762 | running_off = offvar; | |
44fc7854 | 7763 | alias_off = build_int_cst (ref_type, 0); |
4d694b27 | 7764 | int nloads = const_nunits; |
e09b4c37 | 7765 | int lnel = 1; |
7b5fc413 | 7766 | tree ltype = TREE_TYPE (vectype); |
ea60dd34 | 7767 | tree lvectype = vectype; |
b266b968 | 7768 | auto_vec<tree> dr_chain; |
2de001ee | 7769 | if (memory_access_type == VMAT_STRIDED_SLP) |
7b5fc413 | 7770 | { |
4d694b27 | 7771 | if (group_size < const_nunits) |
e09b4c37 | 7772 | { |
ff03930a JJ |
7773 | /* First check if vec_init optab supports construction from |
7774 | vector elts directly. */ | |
b397965c | 7775 | scalar_mode elmode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); |
9da15d40 RS |
7776 | machine_mode vmode; |
7777 | if (mode_for_vector (elmode, group_size).exists (&vmode) | |
7778 | && VECTOR_MODE_P (vmode) | |
414fef4e | 7779 | && targetm.vector_mode_supported_p (vmode) |
ff03930a JJ |
7780 | && (convert_optab_handler (vec_init_optab, |
7781 | TYPE_MODE (vectype), vmode) | |
7782 | != CODE_FOR_nothing)) | |
ea60dd34 | 7783 | { |
4d694b27 | 7784 | nloads = const_nunits / group_size; |
ea60dd34 | 7785 | lnel = group_size; |
ff03930a JJ |
7786 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); |
7787 | } | |
7788 | else | |
7789 | { | |
7790 | /* Otherwise avoid emitting a constructor of vector elements | |
7791 | by performing the loads using an integer type of the same | |
7792 | size, constructing a vector of those and then | |
7793 | re-interpreting it as the original vector type. | |
7794 | This avoids a huge runtime penalty due to the general | |
7795 | inability to perform store forwarding from smaller stores | |
7796 | to a larger load. */ | |
7797 | unsigned lsize | |
7798 | = group_size * TYPE_PRECISION (TREE_TYPE (vectype)); | |
fffbab82 | 7799 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 7800 | unsigned int lnunits = const_nunits / group_size; |
ff03930a JJ |
7801 | /* If we can't construct such a vector fall back to |
7802 | element loads of the original vector type. */ | |
4d694b27 | 7803 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 7804 | && VECTOR_MODE_P (vmode) |
414fef4e | 7805 | && targetm.vector_mode_supported_p (vmode) |
ff03930a JJ |
7806 | && (convert_optab_handler (vec_init_optab, vmode, elmode) |
7807 | != CODE_FOR_nothing)) | |
7808 | { | |
4d694b27 | 7809 | nloads = lnunits; |
ff03930a JJ |
7810 | lnel = group_size; |
7811 | ltype = build_nonstandard_integer_type (lsize, 1); | |
7812 | lvectype = build_vector_type (ltype, nloads); | |
7813 | } | |
ea60dd34 | 7814 | } |
e09b4c37 | 7815 | } |
2de001ee | 7816 | else |
e09b4c37 | 7817 | { |
ea60dd34 | 7818 | nloads = 1; |
4d694b27 | 7819 | lnel = const_nunits; |
e09b4c37 | 7820 | ltype = vectype; |
e09b4c37 | 7821 | } |
2de001ee RS |
7822 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); |
7823 | } | |
bb4e4747 BC |
7824 | /* Load vector(1) scalar_type if it's 1 element-wise vectype. */ |
7825 | else if (nloads == 1) | |
7826 | ltype = vectype; | |
7827 | ||
2de001ee RS |
7828 | if (slp) |
7829 | { | |
66c16fd9 RB |
7830 | /* For SLP permutation support we need to load the whole group, |
7831 | not only the number of vector stmts the permutation result | |
7832 | fits in. */ | |
b266b968 | 7833 | if (slp_perm) |
66c16fd9 | 7834 | { |
d9f21f6a RS |
7835 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7836 | variable VF. */ | |
7837 | unsigned int const_vf = vf.to_constant (); | |
4d694b27 | 7838 | ncopies = CEIL (group_size * const_vf, const_nunits); |
66c16fd9 RB |
7839 | dr_chain.create (ncopies); |
7840 | } | |
7841 | else | |
7842 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
7b5fc413 | 7843 | } |
4d694b27 | 7844 | unsigned int group_el = 0; |
e09b4c37 RB |
7845 | unsigned HOST_WIDE_INT |
7846 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
7d75abc8 MM |
7847 | for (j = 0; j < ncopies; j++) |
7848 | { | |
7b5fc413 | 7849 | if (nloads > 1) |
e09b4c37 | 7850 | vec_alloc (v, nloads); |
e1bd7296 | 7851 | stmt_vec_info new_stmt_info = NULL; |
e09b4c37 | 7852 | for (i = 0; i < nloads; i++) |
7b5fc413 | 7853 | { |
e09b4c37 | 7854 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
b210f45f | 7855 | group_el * elsz + cst_offset); |
19986382 RB |
7856 | tree data_ref = build2 (MEM_REF, ltype, running_off, this_off); |
7857 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); | |
e1bd7296 RS |
7858 | gassign *new_stmt |
7859 | = gimple_build_assign (make_ssa_name (ltype), data_ref); | |
7860 | new_stmt_info | |
7861 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
e09b4c37 RB |
7862 | if (nloads > 1) |
7863 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, | |
7864 | gimple_assign_lhs (new_stmt)); | |
7865 | ||
7866 | group_el += lnel; | |
7867 | if (! slp | |
7868 | || group_el == group_size) | |
7b5fc413 | 7869 | { |
e09b4c37 RB |
7870 | tree newoff = copy_ssa_name (running_off); |
7871 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
7872 | running_off, stride_step); | |
7b5fc413 RB |
7873 | vect_finish_stmt_generation (stmt, incr, gsi); |
7874 | ||
7875 | running_off = newoff; | |
e09b4c37 | 7876 | group_el = 0; |
7b5fc413 | 7877 | } |
7b5fc413 | 7878 | } |
e09b4c37 | 7879 | if (nloads > 1) |
7d75abc8 | 7880 | { |
ea60dd34 RB |
7881 | tree vec_inv = build_constructor (lvectype, v); |
7882 | new_temp = vect_init_vector (stmt, vec_inv, lvectype, gsi); | |
e1bd7296 | 7883 | new_stmt_info = vinfo->lookup_def (new_temp); |
ea60dd34 RB |
7884 | if (lvectype != vectype) |
7885 | { | |
e1bd7296 RS |
7886 | gassign *new_stmt |
7887 | = gimple_build_assign (make_ssa_name (vectype), | |
7888 | VIEW_CONVERT_EXPR, | |
7889 | build1 (VIEW_CONVERT_EXPR, | |
7890 | vectype, new_temp)); | |
7891 | new_stmt_info | |
7892 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ea60dd34 | 7893 | } |
7d75abc8 MM |
7894 | } |
7895 | ||
7b5fc413 | 7896 | if (slp) |
b266b968 | 7897 | { |
b266b968 | 7898 | if (slp_perm) |
e1bd7296 | 7899 | dr_chain.quick_push (gimple_assign_lhs (new_stmt_info->stmt)); |
66c16fd9 | 7900 | else |
e1bd7296 | 7901 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
b266b968 | 7902 | } |
7d75abc8 | 7903 | else |
225ce44b RB |
7904 | { |
7905 | if (j == 0) | |
e1bd7296 | 7906 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
225ce44b | 7907 | else |
e1bd7296 RS |
7908 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
7909 | prev_stmt_info = new_stmt_info; | |
225ce44b | 7910 | } |
7d75abc8 | 7911 | } |
b266b968 | 7912 | if (slp_perm) |
29afecdf RB |
7913 | { |
7914 | unsigned n_perms; | |
7915 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
7916 | slp_node_instance, false, &n_perms); | |
7917 | } | |
7d75abc8 MM |
7918 | return true; |
7919 | } | |
aec7ae7d | 7920 | |
b5ec4de7 RS |
7921 | if (memory_access_type == VMAT_GATHER_SCATTER |
7922 | || (!slp && memory_access_type == VMAT_CONTIGUOUS)) | |
ab2fc782 RS |
7923 | grouped_load = false; |
7924 | ||
0d0293ac | 7925 | if (grouped_load) |
ebfd146a | 7926 | { |
bffb8014 RS |
7927 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
7928 | group_size = DR_GROUP_SIZE (first_stmt_info); | |
4f0a0218 | 7929 | /* For SLP vectorization we directly vectorize a subchain |
52eab378 RB |
7930 | without permutation. */ |
7931 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
bffb8014 | 7932 | first_stmt_info = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
4f0a0218 RB |
7933 | /* For BB vectorization always use the first stmt to base |
7934 | the data ref pointer on. */ | |
7935 | if (bb_vinfo) | |
b9787581 | 7936 | first_stmt_info_for_drptr = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
6aa904c4 | 7937 | |
ebfd146a | 7938 | /* Check if the chain of loads is already vectorized. */ |
bffb8014 | 7939 | if (STMT_VINFO_VEC_STMT (first_stmt_info) |
01d8bf07 RB |
7940 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. |
7941 | ??? But we can only do so if there is exactly one | |
7942 | as we have no way to get at the rest. Leave the CSE | |
7943 | opportunity alone. | |
7944 | ??? With the group load eventually participating | |
7945 | in multiple different permutations (having multiple | |
7946 | slp nodes which refer to the same group) the CSE | |
7947 | is even wrong code. See PR56270. */ | |
7948 | && !slp) | |
ebfd146a IR |
7949 | { |
7950 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
7951 | return true; | |
7952 | } | |
bffb8014 | 7953 | first_dr = STMT_VINFO_DATA_REF (first_stmt_info); |
9b999e8c | 7954 | group_gap_adj = 0; |
ebfd146a IR |
7955 | |
7956 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
7957 | if (slp) | |
7958 | { | |
0d0293ac | 7959 | grouped_load = false; |
91ff1504 RB |
7960 | /* For SLP permutation support we need to load the whole group, |
7961 | not only the number of vector stmts the permutation result | |
7962 | fits in. */ | |
7963 | if (slp_perm) | |
b267968e | 7964 | { |
d9f21f6a RS |
7965 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7966 | variable VF. */ | |
7967 | unsigned int const_vf = vf.to_constant (); | |
4d694b27 RS |
7968 | unsigned int const_nunits = nunits.to_constant (); |
7969 | vec_num = CEIL (group_size * const_vf, const_nunits); | |
b267968e RB |
7970 | group_gap_adj = vf * group_size - nunits * vec_num; |
7971 | } | |
91ff1504 | 7972 | else |
b267968e RB |
7973 | { |
7974 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
796bd467 RB |
7975 | group_gap_adj |
7976 | = group_size - SLP_INSTANCE_GROUP_SIZE (slp_node_instance); | |
b267968e | 7977 | } |
a70d6342 | 7978 | } |
ebfd146a | 7979 | else |
9b999e8c | 7980 | vec_num = group_size; |
44fc7854 | 7981 | |
bffb8014 | 7982 | ref_type = get_group_alias_ptr_type (first_stmt_info); |
ebfd146a IR |
7983 | } |
7984 | else | |
7985 | { | |
bffb8014 | 7986 | first_stmt_info = stmt_info; |
ebfd146a IR |
7987 | first_dr = dr; |
7988 | group_size = vec_num = 1; | |
9b999e8c | 7989 | group_gap_adj = 0; |
44fc7854 | 7990 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
ebfd146a IR |
7991 | } |
7992 | ||
720f5239 | 7993 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 7994 | gcc_assert (alignment_support_scheme); |
70088b95 RS |
7995 | vec_loop_masks *loop_masks |
7996 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) | |
7997 | ? &LOOP_VINFO_MASKS (loop_vinfo) | |
7998 | : NULL); | |
7cfb4d93 RS |
7999 | /* Targets with store-lane instructions must not require explicit |
8000 | realignment. vect_supportable_dr_alignment always returns either | |
8001 | dr_aligned or dr_unaligned_supported for masked operations. */ | |
8002 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES | |
8003 | && !mask | |
70088b95 | 8004 | && !loop_masks) |
272c6793 RS |
8005 | || alignment_support_scheme == dr_aligned |
8006 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
8007 | |
8008 | /* In case the vectorization factor (VF) is bigger than the number | |
8009 | of elements that we can fit in a vectype (nunits), we have to generate | |
8010 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 8011 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 8012 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 8013 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 8014 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
8015 | stmts that use the defs of the current stmt. The example below |
8016 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
8017 | need to create 4 vectorized stmts): | |
ebfd146a IR |
8018 | |
8019 | before vectorization: | |
8020 | RELATED_STMT VEC_STMT | |
8021 | S1: x = memref - - | |
8022 | S2: z = x + 1 - - | |
8023 | ||
8024 | step 1: vectorize stmt S1: | |
8025 | We first create the vector stmt VS1_0, and, as usual, record a | |
8026 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
8027 | Next, we create the vector stmt VS1_1, and record a pointer to | |
8028 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 8029 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
8030 | stmts and pointers: |
8031 | RELATED_STMT VEC_STMT | |
8032 | VS1_0: vx0 = memref0 VS1_1 - | |
8033 | VS1_1: vx1 = memref1 VS1_2 - | |
8034 | VS1_2: vx2 = memref2 VS1_3 - | |
8035 | VS1_3: vx3 = memref3 - - | |
8036 | S1: x = load - VS1_0 | |
8037 | S2: z = x + 1 - - | |
8038 | ||
b8698a0f L |
8039 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
8040 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
8041 | stmt S2. */ |
8042 | ||
0d0293ac | 8043 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
8044 | |
8045 | S1: x2 = &base + 2 | |
8046 | S2: x0 = &base | |
8047 | S3: x1 = &base + 1 | |
8048 | S4: x3 = &base + 3 | |
8049 | ||
b8698a0f | 8050 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
8051 | starting from the access of the first stmt of the chain: |
8052 | ||
8053 | VS1: vx0 = &base | |
8054 | VS2: vx1 = &base + vec_size*1 | |
8055 | VS3: vx3 = &base + vec_size*2 | |
8056 | VS4: vx4 = &base + vec_size*3 | |
8057 | ||
8058 | Then permutation statements are generated: | |
8059 | ||
e2c83630 RH |
8060 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
8061 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
8062 | ... |
8063 | ||
8064 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
8065 | (the order of the data-refs in the output of vect_permute_load_chain | |
8066 | corresponds to the order of scalar stmts in the interleaving chain - see | |
8067 | the documentation of vect_permute_load_chain()). | |
8068 | The generation of permutation stmts and recording them in | |
0d0293ac | 8069 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 8070 | |
b8698a0f | 8071 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
8072 | permutation stmts above are created for every copy. The result vector |
8073 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
8074 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
8075 | |
8076 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
8077 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
8078 | we generate the following code: | |
8079 | p = initial_addr; | |
8080 | indx = 0; | |
8081 | loop { | |
8082 | p = p + indx * vectype_size; | |
8083 | vec_dest = *(p); | |
8084 | indx = indx + 1; | |
8085 | } | |
8086 | ||
8087 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 8088 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
8089 | then generate the following code, in which the data in each iteration is |
8090 | obtained by two vector loads, one from the previous iteration, and one | |
8091 | from the current iteration: | |
8092 | p1 = initial_addr; | |
8093 | msq_init = *(floor(p1)) | |
8094 | p2 = initial_addr + VS - 1; | |
8095 | realignment_token = call target_builtin; | |
8096 | indx = 0; | |
8097 | loop { | |
8098 | p2 = p2 + indx * vectype_size | |
8099 | lsq = *(floor(p2)) | |
8100 | vec_dest = realign_load (msq, lsq, realignment_token) | |
8101 | indx = indx + 1; | |
8102 | msq = lsq; | |
8103 | } */ | |
8104 | ||
8105 | /* If the misalignment remains the same throughout the execution of the | |
8106 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 8107 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
8108 | This can only occur when vectorizing memory accesses in the inner-loop |
8109 | nested within an outer-loop that is being vectorized. */ | |
8110 | ||
d1e4b493 | 8111 | if (nested_in_vect_loop |
cf098191 RS |
8112 | && !multiple_p (DR_STEP_ALIGNMENT (dr), |
8113 | GET_MODE_SIZE (TYPE_MODE (vectype)))) | |
ebfd146a IR |
8114 | { |
8115 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
8116 | compute_in_loop = true; | |
8117 | } | |
8118 | ||
8119 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
8120 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 8121 | && !compute_in_loop) |
ebfd146a | 8122 | { |
bffb8014 | 8123 | msq = vect_setup_realignment (first_stmt_info, gsi, &realignment_token, |
ebfd146a IR |
8124 | alignment_support_scheme, NULL_TREE, |
8125 | &at_loop); | |
8126 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
8127 | { | |
538dd0b7 | 8128 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
8129 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
8130 | size_one_node); | |
ebfd146a IR |
8131 | } |
8132 | } | |
8133 | else | |
8134 | at_loop = loop; | |
8135 | ||
62da9e14 | 8136 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
a1e53f3f L |
8137 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
8138 | ||
ab2fc782 RS |
8139 | tree bump; |
8140 | tree vec_offset = NULL_TREE; | |
8141 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
8142 | { | |
8143 | aggr_type = NULL_TREE; | |
8144 | bump = NULL_TREE; | |
8145 | } | |
8146 | else if (memory_access_type == VMAT_GATHER_SCATTER) | |
8147 | { | |
8148 | aggr_type = elem_type; | |
8149 | vect_get_strided_load_store_ops (stmt, loop_vinfo, &gs_info, | |
8150 | &bump, &vec_offset); | |
8151 | } | |
272c6793 | 8152 | else |
ab2fc782 RS |
8153 | { |
8154 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
8155 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
8156 | else | |
8157 | aggr_type = vectype; | |
8158 | bump = vect_get_data_ptr_increment (dr, aggr_type, memory_access_type); | |
8159 | } | |
272c6793 | 8160 | |
c3a8f964 | 8161 | tree vec_mask = NULL_TREE; |
ebfd146a | 8162 | prev_stmt_info = NULL; |
4d694b27 | 8163 | poly_uint64 group_elt = 0; |
ebfd146a | 8164 | for (j = 0; j < ncopies; j++) |
b8698a0f | 8165 | { |
e1bd7296 | 8166 | stmt_vec_info new_stmt_info = NULL; |
272c6793 | 8167 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 8168 | if (j == 0) |
74bf76ed JJ |
8169 | { |
8170 | bool simd_lane_access_p | |
8171 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
8172 | if (simd_lane_access_p | |
8173 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
8174 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
8175 | && integer_zerop (DR_OFFSET (first_dr)) | |
8176 | && integer_zerop (DR_INIT (first_dr)) | |
8177 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 8178 | get_alias_set (TREE_TYPE (ref_type))) |
74bf76ed JJ |
8179 | && (alignment_support_scheme == dr_aligned |
8180 | || alignment_support_scheme == dr_unaligned_supported)) | |
8181 | { | |
8182 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 8183 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 8184 | inv_p = false; |
74bf76ed | 8185 | } |
b9787581 | 8186 | else if (first_stmt_info_for_drptr |
bffb8014 | 8187 | && first_stmt_info != first_stmt_info_for_drptr) |
4f0a0218 RB |
8188 | { |
8189 | dataref_ptr | |
b9787581 RS |
8190 | = vect_create_data_ref_ptr (first_stmt_info_for_drptr, |
8191 | aggr_type, at_loop, offset, &dummy, | |
8192 | gsi, &ptr_incr, simd_lane_access_p, | |
ab2fc782 | 8193 | &inv_p, byte_offset, bump); |
4f0a0218 RB |
8194 | /* Adjust the pointer by the difference to first_stmt. */ |
8195 | data_reference_p ptrdr | |
b9787581 | 8196 | = STMT_VINFO_DATA_REF (first_stmt_info_for_drptr); |
4f0a0218 RB |
8197 | tree diff = fold_convert (sizetype, |
8198 | size_binop (MINUS_EXPR, | |
8199 | DR_INIT (first_dr), | |
8200 | DR_INIT (ptrdr))); | |
8201 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
8202 | stmt, diff); | |
8203 | } | |
bfaa08b7 RS |
8204 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
8205 | { | |
8206 | vect_get_gather_scatter_ops (loop, stmt, &gs_info, | |
8207 | &dataref_ptr, &vec_offset); | |
8208 | inv_p = false; | |
8209 | } | |
74bf76ed JJ |
8210 | else |
8211 | dataref_ptr | |
bffb8014 | 8212 | = vect_create_data_ref_ptr (first_stmt_info, aggr_type, at_loop, |
74bf76ed | 8213 | offset, &dummy, gsi, &ptr_incr, |
356bbc4c | 8214 | simd_lane_access_p, &inv_p, |
ab2fc782 | 8215 | byte_offset, bump); |
c3a8f964 RS |
8216 | if (mask) |
8217 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, | |
8218 | mask_vectype); | |
74bf76ed | 8219 | } |
ebfd146a | 8220 | else |
c3a8f964 RS |
8221 | { |
8222 | if (dataref_offset) | |
8223 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
ab2fc782 | 8224 | bump); |
bfaa08b7 | 8225 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
929b4411 RS |
8226 | vec_offset = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
8227 | vec_offset); | |
c3a8f964 | 8228 | else |
ab2fc782 RS |
8229 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8230 | stmt, bump); | |
c3a8f964 | 8231 | if (mask) |
929b4411 | 8232 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
c3a8f964 | 8233 | } |
ebfd146a | 8234 | |
0d0293ac | 8235 | if (grouped_load || slp_perm) |
9771b263 | 8236 | dr_chain.create (vec_num); |
5ce1ee7f | 8237 | |
2de001ee | 8238 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 8239 | { |
272c6793 RS |
8240 | tree vec_array; |
8241 | ||
8242 | vec_array = create_vector_array (vectype, vec_num); | |
8243 | ||
7cfb4d93 | 8244 | tree final_mask = NULL_TREE; |
70088b95 RS |
8245 | if (loop_masks) |
8246 | final_mask = vect_get_loop_mask (gsi, loop_masks, ncopies, | |
8247 | vectype, j); | |
7cfb4d93 RS |
8248 | if (vec_mask) |
8249 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
8250 | vec_mask, gsi); | |
8251 | ||
7e11fc7f | 8252 | gcall *call; |
7cfb4d93 | 8253 | if (final_mask) |
7e11fc7f RS |
8254 | { |
8255 | /* Emit: | |
8256 | VEC_ARRAY = MASK_LOAD_LANES (DATAREF_PTR, ALIAS_PTR, | |
8257 | VEC_MASK). */ | |
8258 | unsigned int align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype)); | |
8259 | tree alias_ptr = build_int_cst (ref_type, align); | |
8260 | call = gimple_build_call_internal (IFN_MASK_LOAD_LANES, 3, | |
8261 | dataref_ptr, alias_ptr, | |
7cfb4d93 | 8262 | final_mask); |
7e11fc7f RS |
8263 | } |
8264 | else | |
8265 | { | |
8266 | /* Emit: | |
8267 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
8268 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); | |
8269 | call = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
8270 | } | |
a844293d RS |
8271 | gimple_call_set_lhs (call, vec_array); |
8272 | gimple_call_set_nothrow (call, true); | |
e1bd7296 | 8273 | new_stmt_info = vect_finish_stmt_generation (stmt, call, gsi); |
ebfd146a | 8274 | |
272c6793 RS |
8275 | /* Extract each vector into an SSA_NAME. */ |
8276 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 8277 | { |
272c6793 RS |
8278 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
8279 | vec_array, i); | |
9771b263 | 8280 | dr_chain.quick_push (new_temp); |
272c6793 RS |
8281 | } |
8282 | ||
8283 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 8284 | vect_record_grouped_load_vectors (stmt, dr_chain); |
3ba4ff41 RS |
8285 | |
8286 | /* Record that VEC_ARRAY is now dead. */ | |
8287 | vect_clobber_variable (stmt, gsi, vec_array); | |
272c6793 RS |
8288 | } |
8289 | else | |
8290 | { | |
8291 | for (i = 0; i < vec_num; i++) | |
8292 | { | |
7cfb4d93 | 8293 | tree final_mask = NULL_TREE; |
70088b95 | 8294 | if (loop_masks |
7cfb4d93 | 8295 | && memory_access_type != VMAT_INVARIANT) |
70088b95 RS |
8296 | final_mask = vect_get_loop_mask (gsi, loop_masks, |
8297 | vec_num * ncopies, | |
7cfb4d93 RS |
8298 | vectype, vec_num * j + i); |
8299 | if (vec_mask) | |
8300 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
8301 | vec_mask, gsi); | |
8302 | ||
272c6793 RS |
8303 | if (i > 0) |
8304 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
ab2fc782 | 8305 | stmt, bump); |
272c6793 RS |
8306 | |
8307 | /* 2. Create the vector-load in the loop. */ | |
e1bd7296 | 8308 | gimple *new_stmt = NULL; |
272c6793 RS |
8309 | switch (alignment_support_scheme) |
8310 | { | |
8311 | case dr_aligned: | |
8312 | case dr_unaligned_supported: | |
be1ac4ec | 8313 | { |
644ffefd MJ |
8314 | unsigned int align, misalign; |
8315 | ||
bfaa08b7 RS |
8316 | if (memory_access_type == VMAT_GATHER_SCATTER) |
8317 | { | |
8318 | tree scale = size_int (gs_info.scale); | |
8319 | gcall *call; | |
70088b95 | 8320 | if (loop_masks) |
bfaa08b7 RS |
8321 | call = gimple_build_call_internal |
8322 | (IFN_MASK_GATHER_LOAD, 4, dataref_ptr, | |
8323 | vec_offset, scale, final_mask); | |
8324 | else | |
8325 | call = gimple_build_call_internal | |
8326 | (IFN_GATHER_LOAD, 3, dataref_ptr, | |
8327 | vec_offset, scale); | |
8328 | gimple_call_set_nothrow (call, true); | |
8329 | new_stmt = call; | |
8330 | data_ref = NULL_TREE; | |
8331 | break; | |
8332 | } | |
8333 | ||
f702e7d4 | 8334 | align = DR_TARGET_ALIGNMENT (dr); |
272c6793 RS |
8335 | if (alignment_support_scheme == dr_aligned) |
8336 | { | |
8337 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 8338 | misalign = 0; |
272c6793 RS |
8339 | } |
8340 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
8341 | { | |
25f68d90 | 8342 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 8343 | misalign = 0; |
272c6793 RS |
8344 | } |
8345 | else | |
c3a8f964 | 8346 | misalign = DR_MISALIGNMENT (first_dr); |
aed93b23 RB |
8347 | if (dataref_offset == NULL_TREE |
8348 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
8349 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
8350 | align, misalign); | |
c3a8f964 | 8351 | |
7cfb4d93 | 8352 | if (final_mask) |
c3a8f964 RS |
8353 | { |
8354 | align = least_bit_hwi (misalign | align); | |
8355 | tree ptr = build_int_cst (ref_type, align); | |
8356 | gcall *call | |
8357 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, | |
8358 | dataref_ptr, ptr, | |
7cfb4d93 | 8359 | final_mask); |
c3a8f964 RS |
8360 | gimple_call_set_nothrow (call, true); |
8361 | new_stmt = call; | |
8362 | data_ref = NULL_TREE; | |
8363 | } | |
8364 | else | |
8365 | { | |
8366 | data_ref | |
8367 | = fold_build2 (MEM_REF, vectype, dataref_ptr, | |
8368 | dataref_offset | |
8369 | ? dataref_offset | |
8370 | : build_int_cst (ref_type, 0)); | |
8371 | if (alignment_support_scheme == dr_aligned) | |
8372 | ; | |
8373 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
8374 | TREE_TYPE (data_ref) | |
8375 | = build_aligned_type (TREE_TYPE (data_ref), | |
8376 | align * BITS_PER_UNIT); | |
8377 | else | |
8378 | TREE_TYPE (data_ref) | |
8379 | = build_aligned_type (TREE_TYPE (data_ref), | |
8380 | TYPE_ALIGN (elem_type)); | |
8381 | } | |
272c6793 | 8382 | break; |
be1ac4ec | 8383 | } |
272c6793 | 8384 | case dr_explicit_realign: |
267d3070 | 8385 | { |
272c6793 | 8386 | tree ptr, bump; |
272c6793 | 8387 | |
d88981fc | 8388 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
8389 | |
8390 | if (compute_in_loop) | |
bffb8014 | 8391 | msq = vect_setup_realignment (first_stmt_info, gsi, |
272c6793 RS |
8392 | &realignment_token, |
8393 | dr_explicit_realign, | |
8394 | dataref_ptr, NULL); | |
8395 | ||
aed93b23 RB |
8396 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
8397 | ptr = copy_ssa_name (dataref_ptr); | |
8398 | else | |
8399 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
f702e7d4 | 8400 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); |
0d0e4a03 JJ |
8401 | new_stmt = gimple_build_assign |
8402 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
8403 | build_int_cst |
8404 | (TREE_TYPE (dataref_ptr), | |
f702e7d4 | 8405 | -(HOST_WIDE_INT) align)); |
272c6793 RS |
8406 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
8407 | data_ref | |
8408 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 8409 | build_int_cst (ref_type, 0)); |
19986382 | 8410 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); |
272c6793 RS |
8411 | vec_dest = vect_create_destination_var (scalar_dest, |
8412 | vectype); | |
8413 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
8414 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
8415 | gimple_assign_set_lhs (new_stmt, new_temp); | |
8416 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
8417 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
8418 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8419 | msq = new_temp; | |
8420 | ||
d88981fc | 8421 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 8422 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 8423 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 8424 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
8425 | new_stmt = gimple_build_assign |
8426 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 | 8427 | build_int_cst |
f702e7d4 | 8428 | (TREE_TYPE (ptr), -(HOST_WIDE_INT) align)); |
aed93b23 | 8429 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
8430 | gimple_assign_set_lhs (new_stmt, ptr); |
8431 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8432 | data_ref | |
8433 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 8434 | build_int_cst (ref_type, 0)); |
272c6793 | 8435 | break; |
267d3070 | 8436 | } |
272c6793 | 8437 | case dr_explicit_realign_optimized: |
f702e7d4 RS |
8438 | { |
8439 | if (TREE_CODE (dataref_ptr) == SSA_NAME) | |
8440 | new_temp = copy_ssa_name (dataref_ptr); | |
8441 | else | |
8442 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
8443 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); | |
8444 | new_stmt = gimple_build_assign | |
8445 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
8446 | build_int_cst (TREE_TYPE (dataref_ptr), | |
8447 | -(HOST_WIDE_INT) align)); | |
8448 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8449 | data_ref | |
8450 | = build2 (MEM_REF, vectype, new_temp, | |
8451 | build_int_cst (ref_type, 0)); | |
8452 | break; | |
8453 | } | |
272c6793 RS |
8454 | default: |
8455 | gcc_unreachable (); | |
8456 | } | |
ebfd146a | 8457 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
c3a8f964 RS |
8458 | /* DATA_REF is null if we've already built the statement. */ |
8459 | if (data_ref) | |
19986382 RB |
8460 | { |
8461 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); | |
8462 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
8463 | } | |
ebfd146a | 8464 | new_temp = make_ssa_name (vec_dest, new_stmt); |
c3a8f964 | 8465 | gimple_set_lhs (new_stmt, new_temp); |
e1bd7296 RS |
8466 | new_stmt_info |
8467 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 8468 | |
272c6793 RS |
8469 | /* 3. Handle explicit realignment if necessary/supported. |
8470 | Create in loop: | |
8471 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
8472 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
8473 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 8474 | { |
272c6793 RS |
8475 | lsq = gimple_assign_lhs (new_stmt); |
8476 | if (!realignment_token) | |
8477 | realignment_token = dataref_ptr; | |
8478 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
8479 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
8480 | msq, lsq, realignment_token); | |
272c6793 RS |
8481 | new_temp = make_ssa_name (vec_dest, new_stmt); |
8482 | gimple_assign_set_lhs (new_stmt, new_temp); | |
e1bd7296 RS |
8483 | new_stmt_info |
8484 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
8485 | |
8486 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
8487 | { | |
8488 | gcc_assert (phi); | |
8489 | if (i == vec_num - 1 && j == ncopies - 1) | |
8490 | add_phi_arg (phi, lsq, | |
8491 | loop_latch_edge (containing_loop), | |
9e227d60 | 8492 | UNKNOWN_LOCATION); |
272c6793 RS |
8493 | msq = lsq; |
8494 | } | |
ebfd146a | 8495 | } |
ebfd146a | 8496 | |
59fd17e3 RB |
8497 | /* 4. Handle invariant-load. */ |
8498 | if (inv_p && !bb_vinfo) | |
8499 | { | |
59fd17e3 | 8500 | gcc_assert (!grouped_load); |
d1417442 JJ |
8501 | /* If we have versioned for aliasing or the loop doesn't |
8502 | have any data dependencies that would preclude this, | |
8503 | then we are sure this is a loop invariant load and | |
8504 | thus we can insert it on the preheader edge. */ | |
8505 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
8506 | && !nested_in_vect_loop | |
6b916b36 | 8507 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
8508 | { |
8509 | if (dump_enabled_p ()) | |
8510 | { | |
8511 | dump_printf_loc (MSG_NOTE, vect_location, | |
8512 | "hoisting out of the vectorized " | |
8513 | "loop: "); | |
8514 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 8515 | } |
b731b390 | 8516 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
8517 | gsi_insert_on_edge_immediate |
8518 | (loop_preheader_edge (loop), | |
8519 | gimple_build_assign (tem, | |
8520 | unshare_expr | |
8521 | (gimple_assign_rhs1 (stmt)))); | |
8522 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
34cd48e5 | 8523 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
e1bd7296 | 8524 | new_stmt_info = vinfo->add_stmt (new_stmt); |
a0e35eb0 RB |
8525 | } |
8526 | else | |
8527 | { | |
8528 | gimple_stmt_iterator gsi2 = *gsi; | |
8529 | gsi_next (&gsi2); | |
8530 | new_temp = vect_init_vector (stmt, scalar_dest, | |
8531 | vectype, &gsi2); | |
e1bd7296 | 8532 | new_stmt_info = vinfo->lookup_def (new_temp); |
a0e35eb0 | 8533 | } |
59fd17e3 RB |
8534 | } |
8535 | ||
62da9e14 | 8536 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
272c6793 | 8537 | { |
aec7ae7d JJ |
8538 | tree perm_mask = perm_mask_for_reverse (vectype); |
8539 | new_temp = permute_vec_elements (new_temp, new_temp, | |
8540 | perm_mask, stmt, gsi); | |
e1bd7296 | 8541 | new_stmt_info = vinfo->lookup_def (new_temp); |
ebfd146a | 8542 | } |
267d3070 | 8543 | |
272c6793 | 8544 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
8545 | vect_transform_grouped_load (). */ |
8546 | if (grouped_load || slp_perm) | |
9771b263 | 8547 | dr_chain.quick_push (new_temp); |
267d3070 | 8548 | |
272c6793 RS |
8549 | /* Store vector loads in the corresponding SLP_NODE. */ |
8550 | if (slp && !slp_perm) | |
e1bd7296 | 8551 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
b267968e RB |
8552 | |
8553 | /* With SLP permutation we load the gaps as well, without | |
8554 | we need to skip the gaps after we manage to fully load | |
2c53b149 | 8555 | all elements. group_gap_adj is DR_GROUP_SIZE here. */ |
b267968e | 8556 | group_elt += nunits; |
d9f21f6a RS |
8557 | if (maybe_ne (group_gap_adj, 0U) |
8558 | && !slp_perm | |
8559 | && known_eq (group_elt, group_size - group_gap_adj)) | |
b267968e | 8560 | { |
d9f21f6a RS |
8561 | poly_wide_int bump_val |
8562 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
8563 | * group_gap_adj); | |
8e6cdc90 | 8564 | tree bump = wide_int_to_tree (sizetype, bump_val); |
b267968e RB |
8565 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8566 | stmt, bump); | |
8567 | group_elt = 0; | |
8568 | } | |
272c6793 | 8569 | } |
9b999e8c RB |
8570 | /* Bump the vector pointer to account for a gap or for excess |
8571 | elements loaded for a permuted SLP load. */ | |
d9f21f6a | 8572 | if (maybe_ne (group_gap_adj, 0U) && slp_perm) |
a64b9c26 | 8573 | { |
d9f21f6a RS |
8574 | poly_wide_int bump_val |
8575 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
8576 | * group_gap_adj); | |
8e6cdc90 | 8577 | tree bump = wide_int_to_tree (sizetype, bump_val); |
a64b9c26 RB |
8578 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8579 | stmt, bump); | |
8580 | } | |
ebfd146a IR |
8581 | } |
8582 | ||
8583 | if (slp && !slp_perm) | |
8584 | continue; | |
8585 | ||
8586 | if (slp_perm) | |
8587 | { | |
29afecdf | 8588 | unsigned n_perms; |
01d8bf07 | 8589 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
29afecdf RB |
8590 | slp_node_instance, false, |
8591 | &n_perms)) | |
ebfd146a | 8592 | { |
9771b263 | 8593 | dr_chain.release (); |
ebfd146a IR |
8594 | return false; |
8595 | } | |
8596 | } | |
8597 | else | |
8598 | { | |
0d0293ac | 8599 | if (grouped_load) |
ebfd146a | 8600 | { |
2de001ee | 8601 | if (memory_access_type != VMAT_LOAD_STORE_LANES) |
0d0293ac | 8602 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 8603 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
8604 | } |
8605 | else | |
8606 | { | |
8607 | if (j == 0) | |
e1bd7296 | 8608 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
ebfd146a | 8609 | else |
e1bd7296 RS |
8610 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
8611 | prev_stmt_info = new_stmt_info; | |
ebfd146a IR |
8612 | } |
8613 | } | |
9771b263 | 8614 | dr_chain.release (); |
ebfd146a IR |
8615 | } |
8616 | ||
ebfd146a IR |
8617 | return true; |
8618 | } | |
8619 | ||
8620 | /* Function vect_is_simple_cond. | |
b8698a0f | 8621 | |
ebfd146a IR |
8622 | Input: |
8623 | LOOP - the loop that is being vectorized. | |
8624 | COND - Condition that is checked for simple use. | |
8625 | ||
e9e1d143 RG |
8626 | Output: |
8627 | *COMP_VECTYPE - the vector type for the comparison. | |
4fc5ebf1 | 8628 | *DTS - The def types for the arguments of the comparison |
e9e1d143 | 8629 | |
ebfd146a IR |
8630 | Returns whether a COND can be vectorized. Checks whether |
8631 | condition operands are supportable using vec_is_simple_use. */ | |
8632 | ||
87aab9b2 | 8633 | static bool |
4fc5ebf1 | 8634 | vect_is_simple_cond (tree cond, vec_info *vinfo, |
8da4c8d8 RB |
8635 | tree *comp_vectype, enum vect_def_type *dts, |
8636 | tree vectype) | |
ebfd146a IR |
8637 | { |
8638 | tree lhs, rhs; | |
e9e1d143 | 8639 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 8640 | |
a414c77f IE |
8641 | /* Mask case. */ |
8642 | if (TREE_CODE (cond) == SSA_NAME | |
2568d8a1 | 8643 | && VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (cond))) |
a414c77f | 8644 | { |
894dd753 | 8645 | if (!vect_is_simple_use (cond, vinfo, &dts[0], comp_vectype) |
a414c77f IE |
8646 | || !*comp_vectype |
8647 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
8648 | return false; | |
8649 | return true; | |
8650 | } | |
8651 | ||
ebfd146a IR |
8652 | if (!COMPARISON_CLASS_P (cond)) |
8653 | return false; | |
8654 | ||
8655 | lhs = TREE_OPERAND (cond, 0); | |
8656 | rhs = TREE_OPERAND (cond, 1); | |
8657 | ||
8658 | if (TREE_CODE (lhs) == SSA_NAME) | |
8659 | { | |
894dd753 | 8660 | if (!vect_is_simple_use (lhs, vinfo, &dts[0], &vectype1)) |
ebfd146a IR |
8661 | return false; |
8662 | } | |
4fc5ebf1 JG |
8663 | else if (TREE_CODE (lhs) == INTEGER_CST || TREE_CODE (lhs) == REAL_CST |
8664 | || TREE_CODE (lhs) == FIXED_CST) | |
8665 | dts[0] = vect_constant_def; | |
8666 | else | |
ebfd146a IR |
8667 | return false; |
8668 | ||
8669 | if (TREE_CODE (rhs) == SSA_NAME) | |
8670 | { | |
894dd753 | 8671 | if (!vect_is_simple_use (rhs, vinfo, &dts[1], &vectype2)) |
ebfd146a IR |
8672 | return false; |
8673 | } | |
4fc5ebf1 JG |
8674 | else if (TREE_CODE (rhs) == INTEGER_CST || TREE_CODE (rhs) == REAL_CST |
8675 | || TREE_CODE (rhs) == FIXED_CST) | |
8676 | dts[1] = vect_constant_def; | |
8677 | else | |
ebfd146a IR |
8678 | return false; |
8679 | ||
28b33016 | 8680 | if (vectype1 && vectype2 |
928686b1 RS |
8681 | && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1), |
8682 | TYPE_VECTOR_SUBPARTS (vectype2))) | |
28b33016 IE |
8683 | return false; |
8684 | ||
e9e1d143 | 8685 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
8da4c8d8 | 8686 | /* Invariant comparison. */ |
4515e413 | 8687 | if (! *comp_vectype && vectype) |
8da4c8d8 RB |
8688 | { |
8689 | tree scalar_type = TREE_TYPE (lhs); | |
8690 | /* If we can widen the comparison to match vectype do so. */ | |
8691 | if (INTEGRAL_TYPE_P (scalar_type) | |
8692 | && tree_int_cst_lt (TYPE_SIZE (scalar_type), | |
8693 | TYPE_SIZE (TREE_TYPE (vectype)))) | |
8694 | scalar_type = build_nonstandard_integer_type | |
8695 | (tree_to_uhwi (TYPE_SIZE (TREE_TYPE (vectype))), | |
8696 | TYPE_UNSIGNED (scalar_type)); | |
8697 | *comp_vectype = get_vectype_for_scalar_type (scalar_type); | |
8698 | } | |
8699 | ||
ebfd146a IR |
8700 | return true; |
8701 | } | |
8702 | ||
8703 | /* vectorizable_condition. | |
8704 | ||
b8698a0f L |
8705 | Check if STMT is conditional modify expression that can be vectorized. |
8706 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
8707 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
8708 | at GSI. |
8709 | ||
8710 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
8711 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 8712 | else clause if it is 2). |
ebfd146a IR |
8713 | |
8714 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
8715 | ||
4bbe8262 | 8716 | bool |
355fe088 | 8717 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
1eede195 RS |
8718 | stmt_vec_info *vec_stmt, tree reduc_def, |
8719 | int reduc_index, slp_tree slp_node, | |
8720 | stmt_vector_for_cost *cost_vec) | |
ebfd146a IR |
8721 | { |
8722 | tree scalar_dest = NULL_TREE; | |
8723 | tree vec_dest = NULL_TREE; | |
01216d27 JJ |
8724 | tree cond_expr, cond_expr0 = NULL_TREE, cond_expr1 = NULL_TREE; |
8725 | tree then_clause, else_clause; | |
ebfd146a | 8726 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
df11cc78 | 8727 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
8728 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
8729 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
5958f9e2 | 8730 | tree vec_compare; |
ebfd146a IR |
8731 | tree new_temp; |
8732 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4fc5ebf1 JG |
8733 | enum vect_def_type dts[4] |
8734 | = {vect_unknown_def_type, vect_unknown_def_type, | |
8735 | vect_unknown_def_type, vect_unknown_def_type}; | |
8736 | int ndts = 4; | |
f7e531cf | 8737 | int ncopies; |
01216d27 | 8738 | enum tree_code code, cond_code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
a855b1b1 | 8739 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
8740 | int i, j; |
8741 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
8742 | vec<tree> vec_oprnds0 = vNULL; |
8743 | vec<tree> vec_oprnds1 = vNULL; | |
8744 | vec<tree> vec_oprnds2 = vNULL; | |
8745 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 8746 | tree vec_cmp_type; |
a414c77f | 8747 | bool masked = false; |
b8698a0f | 8748 | |
f7e531cf IR |
8749 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
8750 | return false; | |
8751 | ||
bb6c2b68 RS |
8752 | vect_reduction_type reduction_type |
8753 | = STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info); | |
8754 | if (reduction_type == TREE_CODE_REDUCTION) | |
af29617a AH |
8755 | { |
8756 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
8757 | return false; | |
ebfd146a | 8758 | |
af29617a AH |
8759 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
8760 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
8761 | && reduc_def)) | |
8762 | return false; | |
ebfd146a | 8763 | |
af29617a AH |
8764 | /* FORNOW: not yet supported. */ |
8765 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
8766 | { | |
8767 | if (dump_enabled_p ()) | |
8768 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
8769 | "value used after loop.\n"); | |
8770 | return false; | |
8771 | } | |
ebfd146a IR |
8772 | } |
8773 | ||
8774 | /* Is vectorizable conditional operation? */ | |
8775 | if (!is_gimple_assign (stmt)) | |
8776 | return false; | |
8777 | ||
8778 | code = gimple_assign_rhs_code (stmt); | |
8779 | ||
8780 | if (code != COND_EXPR) | |
8781 | return false; | |
8782 | ||
465c8c19 | 8783 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2947d3b2 | 8784 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
465c8c19 | 8785 | |
fce57248 | 8786 | if (slp_node) |
465c8c19 JJ |
8787 | ncopies = 1; |
8788 | else | |
e8f142e2 | 8789 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
8790 | |
8791 | gcc_assert (ncopies >= 1); | |
8792 | if (reduc_index && ncopies > 1) | |
8793 | return false; /* FORNOW */ | |
8794 | ||
4e71066d RG |
8795 | cond_expr = gimple_assign_rhs1 (stmt); |
8796 | then_clause = gimple_assign_rhs2 (stmt); | |
8797 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 8798 | |
4fc5ebf1 | 8799 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, |
4515e413 | 8800 | &comp_vectype, &dts[0], slp_node ? NULL : vectype) |
e9e1d143 | 8801 | || !comp_vectype) |
ebfd146a IR |
8802 | return false; |
8803 | ||
894dd753 | 8804 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &dts[2], &vectype1)) |
2947d3b2 | 8805 | return false; |
894dd753 | 8806 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &dts[3], &vectype2)) |
ebfd146a | 8807 | return false; |
2947d3b2 IE |
8808 | |
8809 | if (vectype1 && !useless_type_conversion_p (vectype, vectype1)) | |
8810 | return false; | |
8811 | ||
8812 | if (vectype2 && !useless_type_conversion_p (vectype, vectype2)) | |
ebfd146a IR |
8813 | return false; |
8814 | ||
28b33016 IE |
8815 | masked = !COMPARISON_CLASS_P (cond_expr); |
8816 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
8817 | ||
74946978 MP |
8818 | if (vec_cmp_type == NULL_TREE) |
8819 | return false; | |
784fb9b3 | 8820 | |
01216d27 JJ |
8821 | cond_code = TREE_CODE (cond_expr); |
8822 | if (!masked) | |
8823 | { | |
8824 | cond_expr0 = TREE_OPERAND (cond_expr, 0); | |
8825 | cond_expr1 = TREE_OPERAND (cond_expr, 1); | |
8826 | } | |
8827 | ||
8828 | if (!masked && VECTOR_BOOLEAN_TYPE_P (comp_vectype)) | |
8829 | { | |
8830 | /* Boolean values may have another representation in vectors | |
8831 | and therefore we prefer bit operations over comparison for | |
8832 | them (which also works for scalar masks). We store opcodes | |
8833 | to use in bitop1 and bitop2. Statement is vectorized as | |
8834 | BITOP2 (rhs1 BITOP1 rhs2) or rhs1 BITOP2 (BITOP1 rhs2) | |
8835 | depending on bitop1 and bitop2 arity. */ | |
8836 | switch (cond_code) | |
8837 | { | |
8838 | case GT_EXPR: | |
8839 | bitop1 = BIT_NOT_EXPR; | |
8840 | bitop2 = BIT_AND_EXPR; | |
8841 | break; | |
8842 | case GE_EXPR: | |
8843 | bitop1 = BIT_NOT_EXPR; | |
8844 | bitop2 = BIT_IOR_EXPR; | |
8845 | break; | |
8846 | case LT_EXPR: | |
8847 | bitop1 = BIT_NOT_EXPR; | |
8848 | bitop2 = BIT_AND_EXPR; | |
8849 | std::swap (cond_expr0, cond_expr1); | |
8850 | break; | |
8851 | case LE_EXPR: | |
8852 | bitop1 = BIT_NOT_EXPR; | |
8853 | bitop2 = BIT_IOR_EXPR; | |
8854 | std::swap (cond_expr0, cond_expr1); | |
8855 | break; | |
8856 | case NE_EXPR: | |
8857 | bitop1 = BIT_XOR_EXPR; | |
8858 | break; | |
8859 | case EQ_EXPR: | |
8860 | bitop1 = BIT_XOR_EXPR; | |
8861 | bitop2 = BIT_NOT_EXPR; | |
8862 | break; | |
8863 | default: | |
8864 | return false; | |
8865 | } | |
8866 | cond_code = SSA_NAME; | |
8867 | } | |
8868 | ||
b8698a0f | 8869 | if (!vec_stmt) |
ebfd146a | 8870 | { |
01216d27 JJ |
8871 | if (bitop1 != NOP_EXPR) |
8872 | { | |
8873 | machine_mode mode = TYPE_MODE (comp_vectype); | |
8874 | optab optab; | |
8875 | ||
8876 | optab = optab_for_tree_code (bitop1, comp_vectype, optab_default); | |
8877 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8878 | return false; | |
8879 | ||
8880 | if (bitop2 != NOP_EXPR) | |
8881 | { | |
8882 | optab = optab_for_tree_code (bitop2, comp_vectype, | |
8883 | optab_default); | |
8884 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8885 | return false; | |
8886 | } | |
8887 | } | |
4fc5ebf1 JG |
8888 | if (expand_vec_cond_expr_p (vectype, comp_vectype, |
8889 | cond_code)) | |
8890 | { | |
68435eb2 RB |
8891 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; |
8892 | vect_model_simple_cost (stmt_info, ncopies, dts, ndts, slp_node, | |
8893 | cost_vec); | |
4fc5ebf1 JG |
8894 | return true; |
8895 | } | |
8896 | return false; | |
ebfd146a IR |
8897 | } |
8898 | ||
f7e531cf IR |
8899 | /* Transform. */ |
8900 | ||
8901 | if (!slp_node) | |
8902 | { | |
9771b263 DN |
8903 | vec_oprnds0.create (1); |
8904 | vec_oprnds1.create (1); | |
8905 | vec_oprnds2.create (1); | |
8906 | vec_oprnds3.create (1); | |
f7e531cf | 8907 | } |
ebfd146a IR |
8908 | |
8909 | /* Handle def. */ | |
8910 | scalar_dest = gimple_assign_lhs (stmt); | |
bb6c2b68 RS |
8911 | if (reduction_type != EXTRACT_LAST_REDUCTION) |
8912 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
ebfd146a IR |
8913 | |
8914 | /* Handle cond expr. */ | |
a855b1b1 MM |
8915 | for (j = 0; j < ncopies; j++) |
8916 | { | |
e1bd7296 | 8917 | stmt_vec_info new_stmt_info = NULL; |
a855b1b1 MM |
8918 | if (j == 0) |
8919 | { | |
f7e531cf IR |
8920 | if (slp_node) |
8921 | { | |
00f96dc9 TS |
8922 | auto_vec<tree, 4> ops; |
8923 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 8924 | |
a414c77f | 8925 | if (masked) |
01216d27 | 8926 | ops.safe_push (cond_expr); |
a414c77f IE |
8927 | else |
8928 | { | |
01216d27 JJ |
8929 | ops.safe_push (cond_expr0); |
8930 | ops.safe_push (cond_expr1); | |
a414c77f | 8931 | } |
9771b263 DN |
8932 | ops.safe_push (then_clause); |
8933 | ops.safe_push (else_clause); | |
306b0c92 | 8934 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
37b5ec8f JJ |
8935 | vec_oprnds3 = vec_defs.pop (); |
8936 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
8937 | if (!masked) |
8938 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 8939 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf IR |
8940 | } |
8941 | else | |
8942 | { | |
a414c77f IE |
8943 | if (masked) |
8944 | { | |
8945 | vec_cond_lhs | |
8946 | = vect_get_vec_def_for_operand (cond_expr, stmt, | |
8947 | comp_vectype); | |
894dd753 | 8948 | vect_is_simple_use (cond_expr, stmt_info->vinfo, &dts[0]); |
a414c77f IE |
8949 | } |
8950 | else | |
8951 | { | |
01216d27 JJ |
8952 | vec_cond_lhs |
8953 | = vect_get_vec_def_for_operand (cond_expr0, | |
8954 | stmt, comp_vectype); | |
894dd753 | 8955 | vect_is_simple_use (cond_expr0, loop_vinfo, &dts[0]); |
01216d27 JJ |
8956 | |
8957 | vec_cond_rhs | |
8958 | = vect_get_vec_def_for_operand (cond_expr1, | |
8959 | stmt, comp_vectype); | |
894dd753 | 8960 | vect_is_simple_use (cond_expr1, loop_vinfo, &dts[1]); |
a414c77f | 8961 | } |
f7e531cf IR |
8962 | if (reduc_index == 1) |
8963 | vec_then_clause = reduc_def; | |
8964 | else | |
8965 | { | |
8966 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 | 8967 | stmt); |
894dd753 | 8968 | vect_is_simple_use (then_clause, loop_vinfo, &dts[2]); |
f7e531cf IR |
8969 | } |
8970 | if (reduc_index == 2) | |
8971 | vec_else_clause = reduc_def; | |
8972 | else | |
8973 | { | |
8974 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 | 8975 | stmt); |
894dd753 | 8976 | vect_is_simple_use (else_clause, loop_vinfo, &dts[3]); |
f7e531cf | 8977 | } |
a855b1b1 MM |
8978 | } |
8979 | } | |
8980 | else | |
8981 | { | |
a414c77f IE |
8982 | vec_cond_lhs |
8983 | = vect_get_vec_def_for_stmt_copy (dts[0], | |
8984 | vec_oprnds0.pop ()); | |
8985 | if (!masked) | |
8986 | vec_cond_rhs | |
8987 | = vect_get_vec_def_for_stmt_copy (dts[1], | |
8988 | vec_oprnds1.pop ()); | |
8989 | ||
a855b1b1 | 8990 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 8991 | vec_oprnds2.pop ()); |
a855b1b1 | 8992 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 8993 | vec_oprnds3.pop ()); |
f7e531cf IR |
8994 | } |
8995 | ||
8996 | if (!slp_node) | |
8997 | { | |
9771b263 | 8998 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
8999 | if (!masked) |
9000 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
9001 | vec_oprnds2.quick_push (vec_then_clause); |
9002 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
9003 | } |
9004 | ||
9dc3f7de | 9005 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 9006 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 9007 | { |
9771b263 DN |
9008 | vec_then_clause = vec_oprnds2[i]; |
9009 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 9010 | |
a414c77f IE |
9011 | if (masked) |
9012 | vec_compare = vec_cond_lhs; | |
9013 | else | |
9014 | { | |
9015 | vec_cond_rhs = vec_oprnds1[i]; | |
01216d27 JJ |
9016 | if (bitop1 == NOP_EXPR) |
9017 | vec_compare = build2 (cond_code, vec_cmp_type, | |
9018 | vec_cond_lhs, vec_cond_rhs); | |
9019 | else | |
9020 | { | |
9021 | new_temp = make_ssa_name (vec_cmp_type); | |
e1bd7296 | 9022 | gassign *new_stmt; |
01216d27 JJ |
9023 | if (bitop1 == BIT_NOT_EXPR) |
9024 | new_stmt = gimple_build_assign (new_temp, bitop1, | |
9025 | vec_cond_rhs); | |
9026 | else | |
9027 | new_stmt | |
9028 | = gimple_build_assign (new_temp, bitop1, vec_cond_lhs, | |
9029 | vec_cond_rhs); | |
9030 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9031 | if (bitop2 == NOP_EXPR) | |
9032 | vec_compare = new_temp; | |
9033 | else if (bitop2 == BIT_NOT_EXPR) | |
9034 | { | |
9035 | /* Instead of doing ~x ? y : z do x ? z : y. */ | |
9036 | vec_compare = new_temp; | |
9037 | std::swap (vec_then_clause, vec_else_clause); | |
9038 | } | |
9039 | else | |
9040 | { | |
9041 | vec_compare = make_ssa_name (vec_cmp_type); | |
9042 | new_stmt | |
9043 | = gimple_build_assign (vec_compare, bitop2, | |
9044 | vec_cond_lhs, new_temp); | |
9045 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9046 | } | |
9047 | } | |
a414c77f | 9048 | } |
bb6c2b68 RS |
9049 | if (reduction_type == EXTRACT_LAST_REDUCTION) |
9050 | { | |
9051 | if (!is_gimple_val (vec_compare)) | |
9052 | { | |
9053 | tree vec_compare_name = make_ssa_name (vec_cmp_type); | |
e1bd7296 RS |
9054 | gassign *new_stmt = gimple_build_assign (vec_compare_name, |
9055 | vec_compare); | |
bb6c2b68 RS |
9056 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9057 | vec_compare = vec_compare_name; | |
9058 | } | |
9059 | gcc_assert (reduc_index == 2); | |
e1bd7296 | 9060 | gcall *new_stmt = gimple_build_call_internal |
bb6c2b68 RS |
9061 | (IFN_FOLD_EXTRACT_LAST, 3, else_clause, vec_compare, |
9062 | vec_then_clause); | |
9063 | gimple_call_set_lhs (new_stmt, scalar_dest); | |
9064 | SSA_NAME_DEF_STMT (scalar_dest) = new_stmt; | |
9065 | if (stmt == gsi_stmt (*gsi)) | |
e1bd7296 | 9066 | new_stmt_info = vect_finish_replace_stmt (stmt, new_stmt); |
bb6c2b68 RS |
9067 | else |
9068 | { | |
9069 | /* In this case we're moving the definition to later in the | |
9070 | block. That doesn't matter because the only uses of the | |
9071 | lhs are in phi statements. */ | |
9072 | gimple_stmt_iterator old_gsi = gsi_for_stmt (stmt); | |
9073 | gsi_remove (&old_gsi, true); | |
e1bd7296 RS |
9074 | new_stmt_info |
9075 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
bb6c2b68 RS |
9076 | } |
9077 | } | |
9078 | else | |
9079 | { | |
9080 | new_temp = make_ssa_name (vec_dest); | |
e1bd7296 RS |
9081 | gassign *new_stmt |
9082 | = gimple_build_assign (new_temp, VEC_COND_EXPR, vec_compare, | |
9083 | vec_then_clause, vec_else_clause); | |
9084 | new_stmt_info | |
9085 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
bb6c2b68 | 9086 | } |
f7e531cf | 9087 | if (slp_node) |
e1bd7296 | 9088 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
f7e531cf IR |
9089 | } |
9090 | ||
9091 | if (slp_node) | |
9092 | continue; | |
9093 | ||
e1bd7296 RS |
9094 | if (j == 0) |
9095 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; | |
9096 | else | |
9097 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; | |
f7e531cf | 9098 | |
e1bd7296 | 9099 | prev_stmt_info = new_stmt_info; |
a855b1b1 | 9100 | } |
b8698a0f | 9101 | |
9771b263 DN |
9102 | vec_oprnds0.release (); |
9103 | vec_oprnds1.release (); | |
9104 | vec_oprnds2.release (); | |
9105 | vec_oprnds3.release (); | |
f7e531cf | 9106 | |
ebfd146a IR |
9107 | return true; |
9108 | } | |
9109 | ||
42fd8198 IE |
9110 | /* vectorizable_comparison. |
9111 | ||
9112 | Check if STMT is comparison expression that can be vectorized. | |
9113 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
9114 | comparison, put it in VEC_STMT, and insert it at GSI. | |
9115 | ||
9116 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
9117 | ||
fce57248 | 9118 | static bool |
42fd8198 | 9119 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, |
1eede195 | 9120 | stmt_vec_info *vec_stmt, tree reduc_def, |
68435eb2 | 9121 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
42fd8198 IE |
9122 | { |
9123 | tree lhs, rhs1, rhs2; | |
9124 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9125 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
9126 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
9127 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
9128 | tree new_temp; | |
9129 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
9130 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 9131 | int ndts = 2; |
928686b1 | 9132 | poly_uint64 nunits; |
42fd8198 | 9133 | int ncopies; |
49e76ff1 | 9134 | enum tree_code code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
42fd8198 IE |
9135 | stmt_vec_info prev_stmt_info = NULL; |
9136 | int i, j; | |
9137 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
9138 | vec<tree> vec_oprnds0 = vNULL; | |
9139 | vec<tree> vec_oprnds1 = vNULL; | |
42fd8198 IE |
9140 | tree mask_type; |
9141 | tree mask; | |
9142 | ||
c245362b IE |
9143 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
9144 | return false; | |
9145 | ||
30480bcd | 9146 | if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)) |
42fd8198 IE |
9147 | return false; |
9148 | ||
9149 | mask_type = vectype; | |
9150 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
9151 | ||
fce57248 | 9152 | if (slp_node) |
42fd8198 IE |
9153 | ncopies = 1; |
9154 | else | |
e8f142e2 | 9155 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
42fd8198 IE |
9156 | |
9157 | gcc_assert (ncopies >= 1); | |
42fd8198 IE |
9158 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
9159 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
9160 | && reduc_def)) | |
9161 | return false; | |
9162 | ||
9163 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
9164 | { | |
9165 | if (dump_enabled_p ()) | |
9166 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
9167 | "value used after loop.\n"); | |
9168 | return false; | |
9169 | } | |
9170 | ||
9171 | if (!is_gimple_assign (stmt)) | |
9172 | return false; | |
9173 | ||
9174 | code = gimple_assign_rhs_code (stmt); | |
9175 | ||
9176 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
9177 | return false; | |
9178 | ||
9179 | rhs1 = gimple_assign_rhs1 (stmt); | |
9180 | rhs2 = gimple_assign_rhs2 (stmt); | |
9181 | ||
894dd753 | 9182 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &dts[0], &vectype1)) |
42fd8198 IE |
9183 | return false; |
9184 | ||
894dd753 | 9185 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &dts[1], &vectype2)) |
42fd8198 IE |
9186 | return false; |
9187 | ||
9188 | if (vectype1 && vectype2 | |
928686b1 RS |
9189 | && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1), |
9190 | TYPE_VECTOR_SUBPARTS (vectype2))) | |
42fd8198 IE |
9191 | return false; |
9192 | ||
9193 | vectype = vectype1 ? vectype1 : vectype2; | |
9194 | ||
9195 | /* Invariant comparison. */ | |
9196 | if (!vectype) | |
9197 | { | |
69a9a66f | 9198 | vectype = get_vectype_for_scalar_type (TREE_TYPE (rhs1)); |
928686b1 | 9199 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (vectype), nunits)) |
42fd8198 IE |
9200 | return false; |
9201 | } | |
928686b1 | 9202 | else if (maybe_ne (nunits, TYPE_VECTOR_SUBPARTS (vectype))) |
42fd8198 IE |
9203 | return false; |
9204 | ||
49e76ff1 IE |
9205 | /* Can't compare mask and non-mask types. */ |
9206 | if (vectype1 && vectype2 | |
9207 | && (VECTOR_BOOLEAN_TYPE_P (vectype1) ^ VECTOR_BOOLEAN_TYPE_P (vectype2))) | |
9208 | return false; | |
9209 | ||
9210 | /* Boolean values may have another representation in vectors | |
9211 | and therefore we prefer bit operations over comparison for | |
9212 | them (which also works for scalar masks). We store opcodes | |
9213 | to use in bitop1 and bitop2. Statement is vectorized as | |
9214 | BITOP2 (rhs1 BITOP1 rhs2) or | |
9215 | rhs1 BITOP2 (BITOP1 rhs2) | |
9216 | depending on bitop1 and bitop2 arity. */ | |
9217 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) | |
9218 | { | |
9219 | if (code == GT_EXPR) | |
9220 | { | |
9221 | bitop1 = BIT_NOT_EXPR; | |
9222 | bitop2 = BIT_AND_EXPR; | |
9223 | } | |
9224 | else if (code == GE_EXPR) | |
9225 | { | |
9226 | bitop1 = BIT_NOT_EXPR; | |
9227 | bitop2 = BIT_IOR_EXPR; | |
9228 | } | |
9229 | else if (code == LT_EXPR) | |
9230 | { | |
9231 | bitop1 = BIT_NOT_EXPR; | |
9232 | bitop2 = BIT_AND_EXPR; | |
9233 | std::swap (rhs1, rhs2); | |
264d951a | 9234 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
9235 | } |
9236 | else if (code == LE_EXPR) | |
9237 | { | |
9238 | bitop1 = BIT_NOT_EXPR; | |
9239 | bitop2 = BIT_IOR_EXPR; | |
9240 | std::swap (rhs1, rhs2); | |
264d951a | 9241 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
9242 | } |
9243 | else | |
9244 | { | |
9245 | bitop1 = BIT_XOR_EXPR; | |
9246 | if (code == EQ_EXPR) | |
9247 | bitop2 = BIT_NOT_EXPR; | |
9248 | } | |
9249 | } | |
9250 | ||
42fd8198 IE |
9251 | if (!vec_stmt) |
9252 | { | |
49e76ff1 | 9253 | if (bitop1 == NOP_EXPR) |
68435eb2 RB |
9254 | { |
9255 | if (!expand_vec_cmp_expr_p (vectype, mask_type, code)) | |
9256 | return false; | |
9257 | } | |
49e76ff1 IE |
9258 | else |
9259 | { | |
9260 | machine_mode mode = TYPE_MODE (vectype); | |
9261 | optab optab; | |
9262 | ||
9263 | optab = optab_for_tree_code (bitop1, vectype, optab_default); | |
9264 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
9265 | return false; | |
9266 | ||
9267 | if (bitop2 != NOP_EXPR) | |
9268 | { | |
9269 | optab = optab_for_tree_code (bitop2, vectype, optab_default); | |
9270 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
9271 | return false; | |
9272 | } | |
49e76ff1 | 9273 | } |
68435eb2 RB |
9274 | |
9275 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
9276 | vect_model_simple_cost (stmt_info, ncopies * (1 + (bitop2 != NOP_EXPR)), | |
9277 | dts, ndts, slp_node, cost_vec); | |
9278 | return true; | |
42fd8198 IE |
9279 | } |
9280 | ||
9281 | /* Transform. */ | |
9282 | if (!slp_node) | |
9283 | { | |
9284 | vec_oprnds0.create (1); | |
9285 | vec_oprnds1.create (1); | |
9286 | } | |
9287 | ||
9288 | /* Handle def. */ | |
9289 | lhs = gimple_assign_lhs (stmt); | |
9290 | mask = vect_create_destination_var (lhs, mask_type); | |
9291 | ||
9292 | /* Handle cmp expr. */ | |
9293 | for (j = 0; j < ncopies; j++) | |
9294 | { | |
e1bd7296 | 9295 | stmt_vec_info new_stmt_info = NULL; |
42fd8198 IE |
9296 | if (j == 0) |
9297 | { | |
9298 | if (slp_node) | |
9299 | { | |
9300 | auto_vec<tree, 2> ops; | |
9301 | auto_vec<vec<tree>, 2> vec_defs; | |
9302 | ||
9303 | ops.safe_push (rhs1); | |
9304 | ops.safe_push (rhs2); | |
306b0c92 | 9305 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
42fd8198 IE |
9306 | vec_oprnds1 = vec_defs.pop (); |
9307 | vec_oprnds0 = vec_defs.pop (); | |
9308 | } | |
9309 | else | |
9310 | { | |
e4af0bc4 IE |
9311 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, vectype); |
9312 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, vectype); | |
42fd8198 IE |
9313 | } |
9314 | } | |
9315 | else | |
9316 | { | |
9317 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
9318 | vec_oprnds0.pop ()); | |
9319 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
9320 | vec_oprnds1.pop ()); | |
9321 | } | |
9322 | ||
9323 | if (!slp_node) | |
9324 | { | |
9325 | vec_oprnds0.quick_push (vec_rhs1); | |
9326 | vec_oprnds1.quick_push (vec_rhs2); | |
9327 | } | |
9328 | ||
9329 | /* Arguments are ready. Create the new vector stmt. */ | |
9330 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
9331 | { | |
9332 | vec_rhs2 = vec_oprnds1[i]; | |
9333 | ||
9334 | new_temp = make_ssa_name (mask); | |
49e76ff1 IE |
9335 | if (bitop1 == NOP_EXPR) |
9336 | { | |
e1bd7296 RS |
9337 | gassign *new_stmt = gimple_build_assign (new_temp, code, |
9338 | vec_rhs1, vec_rhs2); | |
9339 | new_stmt_info | |
9340 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
49e76ff1 IE |
9341 | } |
9342 | else | |
9343 | { | |
e1bd7296 | 9344 | gassign *new_stmt; |
49e76ff1 IE |
9345 | if (bitop1 == BIT_NOT_EXPR) |
9346 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs2); | |
9347 | else | |
9348 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs1, | |
9349 | vec_rhs2); | |
e1bd7296 RS |
9350 | new_stmt_info |
9351 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
49e76ff1 IE |
9352 | if (bitop2 != NOP_EXPR) |
9353 | { | |
9354 | tree res = make_ssa_name (mask); | |
9355 | if (bitop2 == BIT_NOT_EXPR) | |
9356 | new_stmt = gimple_build_assign (res, bitop2, new_temp); | |
9357 | else | |
9358 | new_stmt = gimple_build_assign (res, bitop2, vec_rhs1, | |
9359 | new_temp); | |
e1bd7296 RS |
9360 | new_stmt_info |
9361 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
49e76ff1 IE |
9362 | } |
9363 | } | |
42fd8198 | 9364 | if (slp_node) |
e1bd7296 | 9365 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
42fd8198 IE |
9366 | } |
9367 | ||
9368 | if (slp_node) | |
9369 | continue; | |
9370 | ||
9371 | if (j == 0) | |
e1bd7296 | 9372 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
42fd8198 | 9373 | else |
e1bd7296 | 9374 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
42fd8198 | 9375 | |
e1bd7296 | 9376 | prev_stmt_info = new_stmt_info; |
42fd8198 IE |
9377 | } |
9378 | ||
9379 | vec_oprnds0.release (); | |
9380 | vec_oprnds1.release (); | |
9381 | ||
9382 | return true; | |
9383 | } | |
ebfd146a | 9384 | |
68a0f2ff RS |
9385 | /* If SLP_NODE is nonnull, return true if vectorizable_live_operation |
9386 | can handle all live statements in the node. Otherwise return true | |
9387 | if STMT is not live or if vectorizable_live_operation can handle it. | |
9388 | GSI and VEC_STMT are as for vectorizable_live_operation. */ | |
9389 | ||
9390 | static bool | |
9391 | can_vectorize_live_stmts (gimple *stmt, gimple_stmt_iterator *gsi, | |
1eede195 | 9392 | slp_tree slp_node, stmt_vec_info *vec_stmt, |
68435eb2 | 9393 | stmt_vector_for_cost *cost_vec) |
68a0f2ff RS |
9394 | { |
9395 | if (slp_node) | |
9396 | { | |
b9787581 | 9397 | stmt_vec_info slp_stmt_info; |
68a0f2ff | 9398 | unsigned int i; |
b9787581 | 9399 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, slp_stmt_info) |
68a0f2ff | 9400 | { |
68a0f2ff | 9401 | if (STMT_VINFO_LIVE_P (slp_stmt_info) |
b9787581 | 9402 | && !vectorizable_live_operation (slp_stmt_info, gsi, slp_node, i, |
68435eb2 | 9403 | vec_stmt, cost_vec)) |
68a0f2ff RS |
9404 | return false; |
9405 | } | |
9406 | } | |
9407 | else if (STMT_VINFO_LIVE_P (vinfo_for_stmt (stmt)) | |
68435eb2 RB |
9408 | && !vectorizable_live_operation (stmt, gsi, slp_node, -1, vec_stmt, |
9409 | cost_vec)) | |
68a0f2ff RS |
9410 | return false; |
9411 | ||
9412 | return true; | |
9413 | } | |
9414 | ||
8644a673 | 9415 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
9416 | |
9417 | bool | |
891ad31c | 9418 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node, |
68435eb2 | 9419 | slp_instance node_instance, stmt_vector_for_cost *cost_vec) |
ebfd146a | 9420 | { |
8644a673 | 9421 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
6585ff8f | 9422 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 9423 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 9424 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 9425 | bool ok; |
363477c0 | 9426 | gimple_seq pattern_def_seq; |
ebfd146a | 9427 | |
73fbfcad | 9428 | if (dump_enabled_p ()) |
ebfd146a | 9429 | { |
78c60e3d SS |
9430 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
9431 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 9432 | } |
ebfd146a | 9433 | |
1825a1f3 | 9434 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 9435 | { |
73fbfcad | 9436 | if (dump_enabled_p ()) |
78c60e3d | 9437 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9438 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
9439 | |
9440 | return false; | |
9441 | } | |
b8698a0f | 9442 | |
d54a098e RS |
9443 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
9444 | && node == NULL | |
9445 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) | |
9446 | { | |
9447 | gimple_stmt_iterator si; | |
9448 | ||
9449 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) | |
9450 | { | |
9451 | gimple *pattern_def_stmt = gsi_stmt (si); | |
6585ff8f RS |
9452 | stmt_vec_info pattern_def_stmt_info |
9453 | = vinfo->lookup_stmt (gsi_stmt (si)); | |
9454 | if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info) | |
9455 | || STMT_VINFO_LIVE_P (pattern_def_stmt_info)) | |
d54a098e RS |
9456 | { |
9457 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
9458 | if (dump_enabled_p ()) | |
9459 | { | |
9460 | dump_printf_loc (MSG_NOTE, vect_location, | |
9461 | "==> examining pattern def statement: "); | |
9462 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
9463 | } | |
9464 | ||
9465 | if (!vect_analyze_stmt (pattern_def_stmt, | |
9466 | need_to_vectorize, node, node_instance, | |
9467 | cost_vec)) | |
9468 | return false; | |
9469 | } | |
9470 | } | |
9471 | } | |
9472 | ||
b8698a0f | 9473 | /* Skip stmts that do not need to be vectorized. In loops this is expected |
8644a673 IR |
9474 | to include: |
9475 | - the COND_EXPR which is the loop exit condition | |
9476 | - any LABEL_EXPRs in the loop | |
b8698a0f | 9477 | - computations that are used only for array indexing or loop control. |
8644a673 | 9478 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 9479 | instance, therefore, all the statements are relevant. |
ebfd146a | 9480 | |
d092494c | 9481 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 9482 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
9483 | statements. In basic blocks we are called from some SLP instance |
9484 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
9485 | already will be part of SLP instance. */ | |
83197f37 | 9486 | |
10681ce8 | 9487 | stmt_vec_info pattern_stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); |
b8698a0f | 9488 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 9489 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 9490 | { |
9d5e7640 | 9491 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
10681ce8 RS |
9492 | && pattern_stmt_info |
9493 | && (STMT_VINFO_RELEVANT_P (pattern_stmt_info) | |
9494 | || STMT_VINFO_LIVE_P (pattern_stmt_info))) | |
9d5e7640 | 9495 | { |
83197f37 | 9496 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
10681ce8 RS |
9497 | stmt = pattern_stmt_info->stmt; |
9498 | stmt_info = pattern_stmt_info; | |
73fbfcad | 9499 | if (dump_enabled_p ()) |
9d5e7640 | 9500 | { |
78c60e3d SS |
9501 | dump_printf_loc (MSG_NOTE, vect_location, |
9502 | "==> examining pattern statement: "); | |
9503 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
9504 | } |
9505 | } | |
9506 | else | |
9507 | { | |
73fbfcad | 9508 | if (dump_enabled_p ()) |
e645e942 | 9509 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 9510 | |
9d5e7640 IR |
9511 | return true; |
9512 | } | |
8644a673 | 9513 | } |
83197f37 | 9514 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 9515 | && node == NULL |
10681ce8 RS |
9516 | && pattern_stmt_info |
9517 | && (STMT_VINFO_RELEVANT_P (pattern_stmt_info) | |
9518 | || STMT_VINFO_LIVE_P (pattern_stmt_info))) | |
83197f37 IR |
9519 | { |
9520 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 9521 | if (dump_enabled_p ()) |
83197f37 | 9522 | { |
78c60e3d SS |
9523 | dump_printf_loc (MSG_NOTE, vect_location, |
9524 | "==> examining pattern statement: "); | |
9525 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
9526 | } |
9527 | ||
10681ce8 | 9528 | if (!vect_analyze_stmt (pattern_stmt_info, need_to_vectorize, node, |
68435eb2 | 9529 | node_instance, cost_vec)) |
83197f37 IR |
9530 | return false; |
9531 | } | |
ebfd146a | 9532 | |
8644a673 IR |
9533 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
9534 | { | |
9535 | case vect_internal_def: | |
9536 | break; | |
ebfd146a | 9537 | |
8644a673 | 9538 | case vect_reduction_def: |
7c5222ff | 9539 | case vect_nested_cycle: |
14a61437 RB |
9540 | gcc_assert (!bb_vinfo |
9541 | && (relevance == vect_used_in_outer | |
9542 | || relevance == vect_used_in_outer_by_reduction | |
9543 | || relevance == vect_used_by_reduction | |
b28ead45 AH |
9544 | || relevance == vect_unused_in_scope |
9545 | || relevance == vect_used_only_live)); | |
8644a673 IR |
9546 | break; |
9547 | ||
9548 | case vect_induction_def: | |
e7baeb39 RB |
9549 | gcc_assert (!bb_vinfo); |
9550 | break; | |
9551 | ||
8644a673 IR |
9552 | case vect_constant_def: |
9553 | case vect_external_def: | |
9554 | case vect_unknown_def_type: | |
9555 | default: | |
9556 | gcc_unreachable (); | |
9557 | } | |
ebfd146a | 9558 | |
8644a673 | 9559 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 9560 | { |
8644a673 | 9561 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
9562 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
9563 | || (is_gimple_call (stmt) | |
9564 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 9565 | *need_to_vectorize = true; |
ebfd146a IR |
9566 | } |
9567 | ||
b1af7da6 RB |
9568 | if (PURE_SLP_STMT (stmt_info) && !node) |
9569 | { | |
9570 | dump_printf_loc (MSG_NOTE, vect_location, | |
9571 | "handled only by SLP analysis\n"); | |
9572 | return true; | |
9573 | } | |
9574 | ||
9575 | ok = true; | |
9576 | if (!bb_vinfo | |
9577 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
9578 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
68435eb2 RB |
9579 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node, cost_vec) |
9580 | || vectorizable_conversion (stmt, NULL, NULL, node, cost_vec) | |
9581 | || vectorizable_shift (stmt, NULL, NULL, node, cost_vec) | |
9582 | || vectorizable_operation (stmt, NULL, NULL, node, cost_vec) | |
9583 | || vectorizable_assignment (stmt, NULL, NULL, node, cost_vec) | |
9584 | || vectorizable_load (stmt, NULL, NULL, node, node_instance, cost_vec) | |
9585 | || vectorizable_call (stmt, NULL, NULL, node, cost_vec) | |
9586 | || vectorizable_store (stmt, NULL, NULL, node, cost_vec) | |
9587 | || vectorizable_reduction (stmt, NULL, NULL, node, node_instance, | |
9588 | cost_vec) | |
9589 | || vectorizable_induction (stmt, NULL, NULL, node, cost_vec) | |
9590 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node, cost_vec) | |
9591 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node, cost_vec)); | |
b1af7da6 RB |
9592 | else |
9593 | { | |
9594 | if (bb_vinfo) | |
68435eb2 RB |
9595 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node, cost_vec) |
9596 | || vectorizable_conversion (stmt, NULL, NULL, node, cost_vec) | |
9597 | || vectorizable_shift (stmt, NULL, NULL, node, cost_vec) | |
9598 | || vectorizable_operation (stmt, NULL, NULL, node, cost_vec) | |
9599 | || vectorizable_assignment (stmt, NULL, NULL, node, cost_vec) | |
9600 | || vectorizable_load (stmt, NULL, NULL, node, node_instance, | |
9601 | cost_vec) | |
9602 | || vectorizable_call (stmt, NULL, NULL, node, cost_vec) | |
9603 | || vectorizable_store (stmt, NULL, NULL, node, cost_vec) | |
9604 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node, | |
9605 | cost_vec) | |
9606 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node, | |
9607 | cost_vec)); | |
b1af7da6 | 9608 | } |
8644a673 IR |
9609 | |
9610 | if (!ok) | |
ebfd146a | 9611 | { |
73fbfcad | 9612 | if (dump_enabled_p ()) |
8644a673 | 9613 | { |
78c60e3d SS |
9614 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9615 | "not vectorized: relevant stmt not "); | |
9616 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
9617 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 9618 | } |
b8698a0f | 9619 | |
ebfd146a IR |
9620 | return false; |
9621 | } | |
9622 | ||
8644a673 IR |
9623 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
9624 | need extra handling, except for vectorizable reductions. */ | |
68435eb2 RB |
9625 | if (!bb_vinfo |
9626 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type | |
9627 | && !can_vectorize_live_stmts (stmt, NULL, node, NULL, cost_vec)) | |
ebfd146a | 9628 | { |
73fbfcad | 9629 | if (dump_enabled_p ()) |
8644a673 | 9630 | { |
78c60e3d | 9631 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
68a0f2ff | 9632 | "not vectorized: live stmt not supported: "); |
78c60e3d | 9633 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); |
8644a673 | 9634 | } |
b8698a0f | 9635 | |
8644a673 | 9636 | return false; |
ebfd146a IR |
9637 | } |
9638 | ||
ebfd146a IR |
9639 | return true; |
9640 | } | |
9641 | ||
9642 | ||
9643 | /* Function vect_transform_stmt. | |
9644 | ||
9645 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
9646 | ||
9647 | bool | |
355fe088 | 9648 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 9649 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
9650 | slp_instance slp_node_instance) |
9651 | { | |
6585ff8f RS |
9652 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
9653 | vec_info *vinfo = stmt_info->vinfo; | |
ebfd146a | 9654 | bool is_store = false; |
1eede195 | 9655 | stmt_vec_info vec_stmt = NULL; |
ebfd146a | 9656 | bool done; |
ebfd146a | 9657 | |
fce57248 | 9658 | gcc_assert (slp_node || !PURE_SLP_STMT (stmt_info)); |
1eede195 | 9659 | stmt_vec_info old_vec_stmt_info = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 9660 | |
e57d9a82 RB |
9661 | bool nested_p = (STMT_VINFO_LOOP_VINFO (stmt_info) |
9662 | && nested_in_vect_loop_p | |
9663 | (LOOP_VINFO_LOOP (STMT_VINFO_LOOP_VINFO (stmt_info)), | |
9664 | stmt)); | |
9665 | ||
ebfd146a IR |
9666 | switch (STMT_VINFO_TYPE (stmt_info)) |
9667 | { | |
9668 | case type_demotion_vec_info_type: | |
ebfd146a | 9669 | case type_promotion_vec_info_type: |
ebfd146a | 9670 | case type_conversion_vec_info_type: |
68435eb2 | 9671 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a IR |
9672 | gcc_assert (done); |
9673 | break; | |
9674 | ||
9675 | case induc_vec_info_type: | |
68435eb2 | 9676 | done = vectorizable_induction (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a IR |
9677 | gcc_assert (done); |
9678 | break; | |
9679 | ||
9dc3f7de | 9680 | case shift_vec_info_type: |
68435eb2 | 9681 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node, NULL); |
9dc3f7de IR |
9682 | gcc_assert (done); |
9683 | break; | |
9684 | ||
ebfd146a | 9685 | case op_vec_info_type: |
68435eb2 | 9686 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a IR |
9687 | gcc_assert (done); |
9688 | break; | |
9689 | ||
9690 | case assignment_vec_info_type: | |
68435eb2 | 9691 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a IR |
9692 | gcc_assert (done); |
9693 | break; | |
9694 | ||
9695 | case load_vec_info_type: | |
b8698a0f | 9696 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
68435eb2 | 9697 | slp_node_instance, NULL); |
ebfd146a IR |
9698 | gcc_assert (done); |
9699 | break; | |
9700 | ||
9701 | case store_vec_info_type: | |
68435eb2 | 9702 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a | 9703 | gcc_assert (done); |
0d0293ac | 9704 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
9705 | { |
9706 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 9707 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
9708 | one are skipped, and there vec_stmt_info shouldn't be freed |
9709 | meanwhile. */ | |
0d0293ac | 9710 | *grouped_store = true; |
bffb8014 | 9711 | stmt_vec_info group_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
2c53b149 | 9712 | if (DR_GROUP_STORE_COUNT (group_info) == DR_GROUP_SIZE (group_info)) |
ebfd146a | 9713 | is_store = true; |
f307441a | 9714 | } |
ebfd146a IR |
9715 | else |
9716 | is_store = true; | |
9717 | break; | |
9718 | ||
9719 | case condition_vec_info_type: | |
68435eb2 | 9720 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node, NULL); |
ebfd146a IR |
9721 | gcc_assert (done); |
9722 | break; | |
9723 | ||
42fd8198 | 9724 | case comparison_vec_info_type: |
68435eb2 | 9725 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node, NULL); |
42fd8198 IE |
9726 | gcc_assert (done); |
9727 | break; | |
9728 | ||
ebfd146a | 9729 | case call_vec_info_type: |
68435eb2 | 9730 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node, NULL); |
039d9ea1 | 9731 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
9732 | break; |
9733 | ||
0136f8f0 | 9734 | case call_simd_clone_vec_info_type: |
68435eb2 | 9735 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node, NULL); |
0136f8f0 AH |
9736 | stmt = gsi_stmt (*gsi); |
9737 | break; | |
9738 | ||
ebfd146a | 9739 | case reduc_vec_info_type: |
891ad31c | 9740 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node, |
68435eb2 | 9741 | slp_node_instance, NULL); |
ebfd146a IR |
9742 | gcc_assert (done); |
9743 | break; | |
9744 | ||
9745 | default: | |
9746 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
9747 | { | |
73fbfcad | 9748 | if (dump_enabled_p ()) |
78c60e3d | 9749 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9750 | "stmt not supported.\n"); |
ebfd146a IR |
9751 | gcc_unreachable (); |
9752 | } | |
9753 | } | |
9754 | ||
225ce44b RB |
9755 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
9756 | This would break hybrid SLP vectorization. */ | |
9757 | if (slp_node) | |
d90f8440 | 9758 | gcc_assert (!vec_stmt |
1eede195 | 9759 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt_info); |
225ce44b | 9760 | |
ebfd146a IR |
9761 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
9762 | is being vectorized, but outside the immediately enclosing loop. */ | |
9763 | if (vec_stmt | |
e57d9a82 | 9764 | && nested_p |
ebfd146a IR |
9765 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
9766 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 9767 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 9768 | vect_used_in_outer_by_reduction)) |
ebfd146a | 9769 | { |
a70d6342 IR |
9770 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
9771 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
9772 | imm_use_iterator imm_iter; |
9773 | use_operand_p use_p; | |
9774 | tree scalar_dest; | |
ebfd146a | 9775 | |
73fbfcad | 9776 | if (dump_enabled_p ()) |
78c60e3d | 9777 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 9778 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
9779 | |
9780 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
9781 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
9782 | STMT). */ | |
9783 | if (gimple_code (stmt) == GIMPLE_PHI) | |
9784 | scalar_dest = PHI_RESULT (stmt); | |
9785 | else | |
9786 | scalar_dest = gimple_assign_lhs (stmt); | |
9787 | ||
9788 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
6585ff8f RS |
9789 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) |
9790 | { | |
9791 | stmt_vec_info exit_phi_info | |
9792 | = vinfo->lookup_stmt (USE_STMT (use_p)); | |
9793 | STMT_VINFO_VEC_STMT (exit_phi_info) = vec_stmt; | |
9794 | } | |
ebfd146a IR |
9795 | } |
9796 | ||
9797 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
9798 | being vectorized. */ | |
68a0f2ff | 9799 | if (STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) |
ebfd146a | 9800 | { |
68435eb2 | 9801 | done = can_vectorize_live_stmts (stmt, gsi, slp_node, &vec_stmt, NULL); |
ebfd146a IR |
9802 | gcc_assert (done); |
9803 | } | |
9804 | ||
9805 | if (vec_stmt) | |
83197f37 | 9806 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 9807 | |
b8698a0f | 9808 | return is_store; |
ebfd146a IR |
9809 | } |
9810 | ||
9811 | ||
b8698a0f | 9812 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
9813 | stmt_vec_info. */ |
9814 | ||
9815 | void | |
355fe088 | 9816 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 9817 | { |
355fe088 | 9818 | gimple *next = first_stmt; |
ebfd146a IR |
9819 | gimple_stmt_iterator next_si; |
9820 | ||
9821 | while (next) | |
9822 | { | |
78048b1c JJ |
9823 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
9824 | ||
bffb8014 | 9825 | stmt_vec_info tmp = DR_GROUP_NEXT_ELEMENT (stmt_info); |
78048b1c JJ |
9826 | if (is_pattern_stmt_p (stmt_info)) |
9827 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
9828 | /* Free the attached stmt_vec_info and remove the stmt. */ |
9829 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 9830 | unlink_stmt_vdef (next); |
ebfd146a | 9831 | gsi_remove (&next_si, true); |
3d3f2249 | 9832 | release_defs (next); |
ebfd146a IR |
9833 | free_stmt_vec_info (next); |
9834 | next = tmp; | |
9835 | } | |
9836 | } | |
9837 | ||
9838 | ||
9839 | /* Function new_stmt_vec_info. | |
9840 | ||
9841 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
9842 | ||
9843 | stmt_vec_info | |
310213d4 | 9844 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
9845 | { |
9846 | stmt_vec_info res; | |
dbe1b846 | 9847 | res = (_stmt_vec_info *) xcalloc (1, sizeof (struct _stmt_vec_info)); |
ebfd146a IR |
9848 | |
9849 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
9850 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 9851 | res->vinfo = vinfo; |
8644a673 | 9852 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
9853 | STMT_VINFO_LIVE_P (res) = false; |
9854 | STMT_VINFO_VECTYPE (res) = NULL; | |
9855 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 9856 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a | 9857 | STMT_VINFO_IN_PATTERN_P (res) = false; |
363477c0 | 9858 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 9859 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 9860 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
7e16ce79 | 9861 | STMT_VINFO_VEC_CONST_COND_REDUC_CODE (res) = ERROR_MARK; |
ebfd146a | 9862 | |
ebfd146a IR |
9863 | if (gimple_code (stmt) == GIMPLE_PHI |
9864 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
9865 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
9866 | else | |
8644a673 IR |
9867 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
9868 | ||
9771b263 | 9869 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 9870 | STMT_SLP_TYPE (res) = loop_vect; |
78810bd3 RB |
9871 | STMT_VINFO_NUM_SLP_USES (res) = 0; |
9872 | ||
2c53b149 RB |
9873 | res->first_element = NULL; /* GROUP_FIRST_ELEMENT */ |
9874 | res->next_element = NULL; /* GROUP_NEXT_ELEMENT */ | |
9875 | res->size = 0; /* GROUP_SIZE */ | |
9876 | res->store_count = 0; /* GROUP_STORE_COUNT */ | |
9877 | res->gap = 0; /* GROUP_GAP */ | |
9878 | res->same_dr_stmt = NULL; /* GROUP_SAME_DR_STMT */ | |
ebfd146a | 9879 | |
ca823c85 RB |
9880 | /* This is really "uninitialized" until vect_compute_data_ref_alignment. */ |
9881 | res->dr_aux.misalignment = DR_MISALIGNMENT_UNINITIALIZED; | |
9882 | ||
ebfd146a IR |
9883 | return res; |
9884 | } | |
9885 | ||
9886 | ||
f8c0baaf | 9887 | /* Set the current stmt_vec_info vector to V. */ |
ebfd146a IR |
9888 | |
9889 | void | |
f8c0baaf | 9890 | set_stmt_vec_info_vec (vec<stmt_vec_info> *v) |
ebfd146a | 9891 | { |
f8c0baaf | 9892 | stmt_vec_info_vec = v; |
ebfd146a IR |
9893 | } |
9894 | ||
f8c0baaf | 9895 | /* Free the stmt_vec_info entries in V and release V. */ |
ebfd146a IR |
9896 | |
9897 | void | |
f8c0baaf | 9898 | free_stmt_vec_infos (vec<stmt_vec_info> *v) |
ebfd146a | 9899 | { |
93675444 | 9900 | unsigned int i; |
3161455c | 9901 | stmt_vec_info info; |
f8c0baaf | 9902 | FOR_EACH_VEC_ELT (*v, i, info) |
dbe1b846 | 9903 | if (info != NULL_STMT_VEC_INFO) |
3161455c | 9904 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
f8c0baaf RB |
9905 | if (v == stmt_vec_info_vec) |
9906 | stmt_vec_info_vec = NULL; | |
9907 | v->release (); | |
ebfd146a IR |
9908 | } |
9909 | ||
9910 | ||
9911 | /* Free stmt vectorization related info. */ | |
9912 | ||
9913 | void | |
355fe088 | 9914 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
9915 | { |
9916 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9917 | ||
9918 | if (!stmt_info) | |
9919 | return; | |
9920 | ||
78048b1c JJ |
9921 | /* Check if this statement has a related "pattern stmt" |
9922 | (introduced by the vectorizer during the pattern recognition | |
9923 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
9924 | too. */ | |
9925 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
9926 | { | |
e3947d80 RS |
9927 | if (gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info)) |
9928 | for (gimple_stmt_iterator si = gsi_start (seq); | |
9929 | !gsi_end_p (si); gsi_next (&si)) | |
9930 | { | |
9931 | gimple *seq_stmt = gsi_stmt (si); | |
9932 | gimple_set_bb (seq_stmt, NULL); | |
9933 | tree lhs = gimple_get_lhs (seq_stmt); | |
9934 | if (lhs && TREE_CODE (lhs) == SSA_NAME) | |
9935 | release_ssa_name (lhs); | |
9936 | free_stmt_vec_info (seq_stmt); | |
9937 | } | |
10681ce8 RS |
9938 | stmt_vec_info patt_stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); |
9939 | if (patt_stmt_info) | |
78048b1c | 9940 | { |
10681ce8 RS |
9941 | gimple_set_bb (patt_stmt_info->stmt, NULL); |
9942 | tree lhs = gimple_get_lhs (patt_stmt_info->stmt); | |
e6f5c25d | 9943 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde | 9944 | release_ssa_name (lhs); |
10681ce8 | 9945 | free_stmt_vec_info (patt_stmt_info); |
78048b1c JJ |
9946 | } |
9947 | } | |
9948 | ||
9771b263 | 9949 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 9950 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
9951 | set_vinfo_for_stmt (stmt, NULL); |
9952 | free (stmt_info); | |
9953 | } | |
9954 | ||
9955 | ||
bb67d9c7 | 9956 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 9957 | |
bb67d9c7 | 9958 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
9959 | by the target. */ |
9960 | ||
c803b2a9 | 9961 | tree |
86e36728 | 9962 | get_vectype_for_scalar_type_and_size (tree scalar_type, poly_uint64 size) |
ebfd146a | 9963 | { |
c7d97b28 | 9964 | tree orig_scalar_type = scalar_type; |
3bd8f481 | 9965 | scalar_mode inner_mode; |
ef4bddc2 | 9966 | machine_mode simd_mode; |
86e36728 | 9967 | poly_uint64 nunits; |
ebfd146a IR |
9968 | tree vectype; |
9969 | ||
3bd8f481 RS |
9970 | if (!is_int_mode (TYPE_MODE (scalar_type), &inner_mode) |
9971 | && !is_float_mode (TYPE_MODE (scalar_type), &inner_mode)) | |
ebfd146a IR |
9972 | return NULL_TREE; |
9973 | ||
3bd8f481 | 9974 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
48f2e373 | 9975 | |
7b7b1813 RG |
9976 | /* For vector types of elements whose mode precision doesn't |
9977 | match their types precision we use a element type of mode | |
9978 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
9979 | they support the proper result truncation/extension. |
9980 | We also make sure to build vector types with INTEGER_TYPE | |
9981 | component type only. */ | |
6d7971b8 | 9982 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
9983 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
9984 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
9985 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
9986 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 9987 | |
ccbf5bb4 RG |
9988 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
9989 | When the component mode passes the above test simply use a type | |
9990 | corresponding to that mode. The theory is that any use that | |
9991 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 9992 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 9993 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
9994 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
9995 | ||
9996 | /* We can't build a vector type of elements with alignment bigger than | |
9997 | their size. */ | |
dfc2e2ac | 9998 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
9999 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
10000 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 10001 | |
dfc2e2ac RB |
10002 | /* If we felt back to using the mode fail if there was |
10003 | no scalar type for it. */ | |
10004 | if (scalar_type == NULL_TREE) | |
10005 | return NULL_TREE; | |
10006 | ||
bb67d9c7 RG |
10007 | /* If no size was supplied use the mode the target prefers. Otherwise |
10008 | lookup a vector mode of the specified size. */ | |
86e36728 | 10009 | if (known_eq (size, 0U)) |
bb67d9c7 | 10010 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); |
86e36728 RS |
10011 | else if (!multiple_p (size, nbytes, &nunits) |
10012 | || !mode_for_vector (inner_mode, nunits).exists (&simd_mode)) | |
9da15d40 | 10013 | return NULL_TREE; |
4c8fd8ac | 10014 | /* NOTE: nunits == 1 is allowed to support single element vector types. */ |
86e36728 | 10015 | if (!multiple_p (GET_MODE_SIZE (simd_mode), nbytes, &nunits)) |
cc4b5170 | 10016 | return NULL_TREE; |
ebfd146a IR |
10017 | |
10018 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
10019 | |
10020 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
10021 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 10022 | return NULL_TREE; |
ebfd146a | 10023 | |
c7d97b28 RB |
10024 | /* Re-attach the address-space qualifier if we canonicalized the scalar |
10025 | type. */ | |
10026 | if (TYPE_ADDR_SPACE (orig_scalar_type) != TYPE_ADDR_SPACE (vectype)) | |
10027 | return build_qualified_type | |
10028 | (vectype, KEEP_QUAL_ADDR_SPACE (TYPE_QUALS (orig_scalar_type))); | |
10029 | ||
ebfd146a IR |
10030 | return vectype; |
10031 | } | |
10032 | ||
86e36728 | 10033 | poly_uint64 current_vector_size; |
bb67d9c7 RG |
10034 | |
10035 | /* Function get_vectype_for_scalar_type. | |
10036 | ||
10037 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
10038 | by the target. */ | |
10039 | ||
10040 | tree | |
10041 | get_vectype_for_scalar_type (tree scalar_type) | |
10042 | { | |
10043 | tree vectype; | |
10044 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
10045 | current_vector_size); | |
10046 | if (vectype | |
86e36728 | 10047 | && known_eq (current_vector_size, 0U)) |
bb67d9c7 RG |
10048 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); |
10049 | return vectype; | |
10050 | } | |
10051 | ||
42fd8198 IE |
10052 | /* Function get_mask_type_for_scalar_type. |
10053 | ||
10054 | Returns the mask type corresponding to a result of comparison | |
10055 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
10056 | ||
10057 | tree | |
10058 | get_mask_type_for_scalar_type (tree scalar_type) | |
10059 | { | |
10060 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
10061 | ||
10062 | if (!vectype) | |
10063 | return NULL; | |
10064 | ||
10065 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
10066 | current_vector_size); | |
10067 | } | |
10068 | ||
b690cc0f RG |
10069 | /* Function get_same_sized_vectype |
10070 | ||
10071 | Returns a vector type corresponding to SCALAR_TYPE of size | |
10072 | VECTOR_TYPE if supported by the target. */ | |
10073 | ||
10074 | tree | |
bb67d9c7 | 10075 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 10076 | { |
2568d8a1 | 10077 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type)) |
9f47c7e5 IE |
10078 | return build_same_sized_truth_vector_type (vector_type); |
10079 | ||
bb67d9c7 RG |
10080 | return get_vectype_for_scalar_type_and_size |
10081 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
10082 | } |
10083 | ||
ebfd146a IR |
10084 | /* Function vect_is_simple_use. |
10085 | ||
10086 | Input: | |
81c40241 RB |
10087 | VINFO - the vect info of the loop or basic block that is being vectorized. |
10088 | OPERAND - operand in the loop or bb. | |
10089 | Output: | |
fef96d8e RS |
10090 | DEF_STMT_INFO_OUT (optional) - information about the defining stmt in |
10091 | case OPERAND is an SSA_NAME that is defined in the vectorizable region | |
10092 | DEF_STMT_OUT (optional) - the defining stmt in case OPERAND is an SSA_NAME; | |
10093 | the definition could be anywhere in the function | |
81c40241 | 10094 | DT - the type of definition |
ebfd146a IR |
10095 | |
10096 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 10097 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 10098 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 10099 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
10100 | is the case in reduction/induction computations). |
10101 | For basic blocks, supportable operands are constants and bb invariants. | |
10102 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
10103 | |
10104 | bool | |
894dd753 | 10105 | vect_is_simple_use (tree operand, vec_info *vinfo, enum vect_def_type *dt, |
fef96d8e | 10106 | stmt_vec_info *def_stmt_info_out, gimple **def_stmt_out) |
b8698a0f | 10107 | { |
fef96d8e RS |
10108 | if (def_stmt_info_out) |
10109 | *def_stmt_info_out = NULL; | |
894dd753 RS |
10110 | if (def_stmt_out) |
10111 | *def_stmt_out = NULL; | |
3fc356dc | 10112 | *dt = vect_unknown_def_type; |
b8698a0f | 10113 | |
73fbfcad | 10114 | if (dump_enabled_p ()) |
ebfd146a | 10115 | { |
78c60e3d SS |
10116 | dump_printf_loc (MSG_NOTE, vect_location, |
10117 | "vect_is_simple_use: operand "); | |
30f502ed RB |
10118 | if (TREE_CODE (operand) == SSA_NAME |
10119 | && !SSA_NAME_IS_DEFAULT_DEF (operand)) | |
10120 | dump_gimple_expr (MSG_NOTE, TDF_SLIM, SSA_NAME_DEF_STMT (operand), 0); | |
10121 | else | |
10122 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
ebfd146a | 10123 | } |
b8698a0f | 10124 | |
b758f602 | 10125 | if (CONSTANT_CLASS_P (operand)) |
30f502ed RB |
10126 | *dt = vect_constant_def; |
10127 | else if (is_gimple_min_invariant (operand)) | |
10128 | *dt = vect_external_def; | |
10129 | else if (TREE_CODE (operand) != SSA_NAME) | |
10130 | *dt = vect_unknown_def_type; | |
10131 | else if (SSA_NAME_IS_DEFAULT_DEF (operand)) | |
8644a673 | 10132 | *dt = vect_external_def; |
ebfd146a IR |
10133 | else |
10134 | { | |
30f502ed | 10135 | gimple *def_stmt = SSA_NAME_DEF_STMT (operand); |
c98d0595 RS |
10136 | stmt_vec_info stmt_vinfo = vinfo->lookup_def (operand); |
10137 | if (!stmt_vinfo) | |
30f502ed RB |
10138 | *dt = vect_external_def; |
10139 | else | |
0f8c840c | 10140 | { |
30f502ed RB |
10141 | if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo)) |
10142 | { | |
10681ce8 RS |
10143 | stmt_vinfo = STMT_VINFO_RELATED_STMT (stmt_vinfo); |
10144 | def_stmt = stmt_vinfo->stmt; | |
30f502ed RB |
10145 | } |
10146 | switch (gimple_code (def_stmt)) | |
10147 | { | |
10148 | case GIMPLE_PHI: | |
10149 | case GIMPLE_ASSIGN: | |
10150 | case GIMPLE_CALL: | |
10151 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
10152 | break; | |
10153 | default: | |
10154 | *dt = vect_unknown_def_type; | |
10155 | break; | |
10156 | } | |
fef96d8e RS |
10157 | if (def_stmt_info_out) |
10158 | *def_stmt_info_out = stmt_vinfo; | |
0f8c840c | 10159 | } |
30f502ed RB |
10160 | if (def_stmt_out) |
10161 | *def_stmt_out = def_stmt; | |
ebfd146a IR |
10162 | } |
10163 | ||
2e8ab70c RB |
10164 | if (dump_enabled_p ()) |
10165 | { | |
30f502ed | 10166 | dump_printf (MSG_NOTE, ", type of def: "); |
2e8ab70c RB |
10167 | switch (*dt) |
10168 | { | |
10169 | case vect_uninitialized_def: | |
10170 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
10171 | break; | |
10172 | case vect_constant_def: | |
10173 | dump_printf (MSG_NOTE, "constant\n"); | |
10174 | break; | |
10175 | case vect_external_def: | |
10176 | dump_printf (MSG_NOTE, "external\n"); | |
10177 | break; | |
10178 | case vect_internal_def: | |
10179 | dump_printf (MSG_NOTE, "internal\n"); | |
10180 | break; | |
10181 | case vect_induction_def: | |
10182 | dump_printf (MSG_NOTE, "induction\n"); | |
10183 | break; | |
10184 | case vect_reduction_def: | |
10185 | dump_printf (MSG_NOTE, "reduction\n"); | |
10186 | break; | |
10187 | case vect_double_reduction_def: | |
10188 | dump_printf (MSG_NOTE, "double reduction\n"); | |
10189 | break; | |
10190 | case vect_nested_cycle: | |
10191 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
10192 | break; | |
10193 | case vect_unknown_def_type: | |
10194 | dump_printf (MSG_NOTE, "unknown\n"); | |
10195 | break; | |
10196 | } | |
10197 | } | |
10198 | ||
81c40241 | 10199 | if (*dt == vect_unknown_def_type) |
ebfd146a | 10200 | { |
73fbfcad | 10201 | if (dump_enabled_p ()) |
78c60e3d | 10202 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 10203 | "Unsupported pattern.\n"); |
ebfd146a IR |
10204 | return false; |
10205 | } | |
10206 | ||
ebfd146a IR |
10207 | return true; |
10208 | } | |
10209 | ||
81c40241 | 10210 | /* Function vect_is_simple_use. |
b690cc0f | 10211 | |
81c40241 | 10212 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
10213 | type of OPERAND and stores it to *VECTYPE. If the definition of |
10214 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
10215 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
10216 | is responsible to compute the best suited vector type for the | |
10217 | scalar operand. */ | |
10218 | ||
10219 | bool | |
894dd753 | 10220 | vect_is_simple_use (tree operand, vec_info *vinfo, enum vect_def_type *dt, |
fef96d8e RS |
10221 | tree *vectype, stmt_vec_info *def_stmt_info_out, |
10222 | gimple **def_stmt_out) | |
b690cc0f | 10223 | { |
fef96d8e | 10224 | stmt_vec_info def_stmt_info; |
894dd753 | 10225 | gimple *def_stmt; |
fef96d8e | 10226 | if (!vect_is_simple_use (operand, vinfo, dt, &def_stmt_info, &def_stmt)) |
b690cc0f RG |
10227 | return false; |
10228 | ||
894dd753 RS |
10229 | if (def_stmt_out) |
10230 | *def_stmt_out = def_stmt; | |
fef96d8e RS |
10231 | if (def_stmt_info_out) |
10232 | *def_stmt_info_out = def_stmt_info; | |
894dd753 | 10233 | |
b690cc0f RG |
10234 | /* Now get a vector type if the def is internal, otherwise supply |
10235 | NULL_TREE and leave it up to the caller to figure out a proper | |
10236 | type for the use stmt. */ | |
10237 | if (*dt == vect_internal_def | |
10238 | || *dt == vect_induction_def | |
10239 | || *dt == vect_reduction_def | |
10240 | || *dt == vect_double_reduction_def | |
10241 | || *dt == vect_nested_cycle) | |
10242 | { | |
fef96d8e | 10243 | *vectype = STMT_VINFO_VECTYPE (def_stmt_info); |
b690cc0f | 10244 | gcc_assert (*vectype != NULL_TREE); |
30f502ed RB |
10245 | if (dump_enabled_p ()) |
10246 | { | |
10247 | dump_printf_loc (MSG_NOTE, vect_location, | |
10248 | "vect_is_simple_use: vectype "); | |
10249 | dump_generic_expr (MSG_NOTE, TDF_SLIM, *vectype); | |
10250 | dump_printf (MSG_NOTE, "\n"); | |
10251 | } | |
b690cc0f RG |
10252 | } |
10253 | else if (*dt == vect_uninitialized_def | |
10254 | || *dt == vect_constant_def | |
10255 | || *dt == vect_external_def) | |
10256 | *vectype = NULL_TREE; | |
10257 | else | |
10258 | gcc_unreachable (); | |
10259 | ||
10260 | return true; | |
10261 | } | |
10262 | ||
ebfd146a IR |
10263 | |
10264 | /* Function supportable_widening_operation | |
10265 | ||
b8698a0f L |
10266 | Check whether an operation represented by the code CODE is a |
10267 | widening operation that is supported by the target platform in | |
b690cc0f RG |
10268 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
10269 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 10270 | |
1bda738b JJ |
10271 | Widening operations we currently support are NOP (CONVERT), FLOAT, |
10272 | FIX_TRUNC and WIDEN_MULT. This function checks if these operations | |
10273 | are supported by the target platform either directly (via vector | |
10274 | tree-codes), or via target builtins. | |
ebfd146a IR |
10275 | |
10276 | Output: | |
b8698a0f L |
10277 | - CODE1 and CODE2 are codes of vector operations to be used when |
10278 | vectorizing the operation, if available. | |
ebfd146a IR |
10279 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
10280 | case of multi-step conversion (like char->short->int - in that case | |
10281 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
10282 | - INTERM_TYPES contains the intermediate type required to perform the |
10283 | widening operation (short in the above example). */ | |
ebfd146a IR |
10284 | |
10285 | bool | |
355fe088 | 10286 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 10287 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
10288 | enum tree_code *code1, enum tree_code *code2, |
10289 | int *multi_step_cvt, | |
9771b263 | 10290 | vec<tree> *interm_types) |
ebfd146a IR |
10291 | { |
10292 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
10293 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 10294 | struct loop *vect_loop = NULL; |
ef4bddc2 | 10295 | machine_mode vec_mode; |
81f40b79 | 10296 | enum insn_code icode1, icode2; |
ebfd146a | 10297 | optab optab1, optab2; |
b690cc0f RG |
10298 | tree vectype = vectype_in; |
10299 | tree wide_vectype = vectype_out; | |
ebfd146a | 10300 | enum tree_code c1, c2; |
4a00c761 JJ |
10301 | int i; |
10302 | tree prev_type, intermediate_type; | |
ef4bddc2 | 10303 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 10304 | optab optab3, optab4; |
ebfd146a | 10305 | |
4a00c761 | 10306 | *multi_step_cvt = 0; |
4ef69dfc IR |
10307 | if (loop_info) |
10308 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
10309 | ||
ebfd146a IR |
10310 | switch (code) |
10311 | { | |
10312 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
10313 | /* The result of a vectorized widening operation usually requires |
10314 | two vectors (because the widened results do not fit into one vector). | |
10315 | The generated vector results would normally be expected to be | |
10316 | generated in the same order as in the original scalar computation, | |
10317 | i.e. if 8 results are generated in each vector iteration, they are | |
10318 | to be organized as follows: | |
10319 | vect1: [res1,res2,res3,res4], | |
10320 | vect2: [res5,res6,res7,res8]. | |
10321 | ||
10322 | However, in the special case that the result of the widening | |
10323 | operation is used in a reduction computation only, the order doesn't | |
10324 | matter (because when vectorizing a reduction we change the order of | |
10325 | the computation). Some targets can take advantage of this and | |
10326 | generate more efficient code. For example, targets like Altivec, | |
10327 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
10328 | generate the following vectors: | |
10329 | vect1: [res1,res3,res5,res7], | |
10330 | vect2: [res2,res4,res6,res8]. | |
10331 | ||
10332 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
10333 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
10334 | iterations in parallel). We therefore don't allow to change the | |
10335 | order of the computation in the inner-loop during outer-loop | |
10336 | vectorization. */ | |
10337 | /* TODO: Another case in which order doesn't *really* matter is when we | |
10338 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
10339 | Normally, pack_trunc performs an even/odd permute, whereas the | |
10340 | repack from an even/odd expansion would be an interleave, which | |
10341 | would be significantly simpler for e.g. AVX2. */ | |
10342 | /* In any case, in order to avoid duplicating the code below, recurse | |
10343 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
10344 | are properly set up for the caller. If we fail, we'll continue with | |
10345 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
10346 | if (vect_loop | |
10347 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
10348 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
10349 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
10350 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
10351 | code1, code2, multi_step_cvt, |
10352 | interm_types)) | |
ebc047a2 CH |
10353 | { |
10354 | /* Elements in a vector with vect_used_by_reduction property cannot | |
10355 | be reordered if the use chain with this property does not have the | |
10356 | same operation. One such an example is s += a * b, where elements | |
10357 | in a and b cannot be reordered. Here we check if the vector defined | |
10358 | by STMT is only directly used in the reduction statement. */ | |
0d0a4e20 RS |
10359 | tree lhs = gimple_assign_lhs (stmt); |
10360 | stmt_vec_info use_stmt_info = loop_info->lookup_single_use (lhs); | |
10361 | if (use_stmt_info | |
10362 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
10363 | return true; | |
ebc047a2 | 10364 | } |
4a00c761 JJ |
10365 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
10366 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
10367 | break; |
10368 | ||
81c40241 RB |
10369 | case DOT_PROD_EXPR: |
10370 | c1 = DOT_PROD_EXPR; | |
10371 | c2 = DOT_PROD_EXPR; | |
10372 | break; | |
10373 | ||
10374 | case SAD_EXPR: | |
10375 | c1 = SAD_EXPR; | |
10376 | c2 = SAD_EXPR; | |
10377 | break; | |
10378 | ||
6ae6116f RH |
10379 | case VEC_WIDEN_MULT_EVEN_EXPR: |
10380 | /* Support the recursion induced just above. */ | |
10381 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
10382 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
10383 | break; | |
10384 | ||
36ba4aae | 10385 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
10386 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
10387 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
10388 | break; |
10389 | ||
ebfd146a | 10390 | CASE_CONVERT: |
4a00c761 JJ |
10391 | c1 = VEC_UNPACK_LO_EXPR; |
10392 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
10393 | break; |
10394 | ||
10395 | case FLOAT_EXPR: | |
4a00c761 JJ |
10396 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
10397 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
10398 | break; |
10399 | ||
10400 | case FIX_TRUNC_EXPR: | |
1bda738b JJ |
10401 | c1 = VEC_UNPACK_FIX_TRUNC_LO_EXPR; |
10402 | c2 = VEC_UNPACK_FIX_TRUNC_HI_EXPR; | |
10403 | break; | |
ebfd146a IR |
10404 | |
10405 | default: | |
10406 | gcc_unreachable (); | |
10407 | } | |
10408 | ||
6ae6116f | 10409 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 10410 | std::swap (c1, c2); |
4a00c761 | 10411 | |
ebfd146a IR |
10412 | if (code == FIX_TRUNC_EXPR) |
10413 | { | |
10414 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
10415 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
10416 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
10417 | } |
10418 | else | |
10419 | { | |
10420 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
10421 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
10422 | } | |
10423 | ||
10424 | if (!optab1 || !optab2) | |
10425 | return false; | |
10426 | ||
10427 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
10428 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
10429 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
10430 | return false; |
10431 | ||
4a00c761 JJ |
10432 | *code1 = c1; |
10433 | *code2 = c2; | |
10434 | ||
10435 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
10436 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
10437 | /* For scalar masks we may have different boolean |
10438 | vector types having the same QImode. Thus we | |
10439 | add additional check for elements number. */ | |
10440 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
928686b1 RS |
10441 | || known_eq (TYPE_VECTOR_SUBPARTS (vectype), |
10442 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)); | |
4a00c761 | 10443 | |
b8698a0f | 10444 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 10445 | types. */ |
ebfd146a | 10446 | |
4a00c761 JJ |
10447 | prev_type = vectype; |
10448 | prev_mode = vec_mode; | |
b8698a0f | 10449 | |
4a00c761 JJ |
10450 | if (!CONVERT_EXPR_CODE_P (code)) |
10451 | return false; | |
b8698a0f | 10452 | |
4a00c761 JJ |
10453 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
10454 | intermediate steps in promotion sequence. We try | |
10455 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
10456 | not. */ | |
9771b263 | 10457 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
10458 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
10459 | { | |
10460 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
10461 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
10462 | { | |
7cfb4d93 | 10463 | intermediate_type = vect_halve_mask_nunits (prev_type); |
3ae0661a IE |
10464 | if (intermediate_mode != TYPE_MODE (intermediate_type)) |
10465 | return false; | |
10466 | } | |
10467 | else | |
10468 | intermediate_type | |
10469 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
10470 | TYPE_UNSIGNED (prev_type)); | |
10471 | ||
4a00c761 JJ |
10472 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); |
10473 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
10474 | ||
10475 | if (!optab3 || !optab4 | |
10476 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
10477 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
10478 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
10479 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
10480 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
10481 | == CODE_FOR_nothing) | |
10482 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
10483 | == CODE_FOR_nothing)) | |
10484 | break; | |
ebfd146a | 10485 | |
9771b263 | 10486 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
10487 | (*multi_step_cvt)++; |
10488 | ||
10489 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
10490 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff | 10491 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
928686b1 RS |
10492 | || known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type), |
10493 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)); | |
4a00c761 JJ |
10494 | |
10495 | prev_type = intermediate_type; | |
10496 | prev_mode = intermediate_mode; | |
ebfd146a IR |
10497 | } |
10498 | ||
9771b263 | 10499 | interm_types->release (); |
4a00c761 | 10500 | return false; |
ebfd146a IR |
10501 | } |
10502 | ||
10503 | ||
10504 | /* Function supportable_narrowing_operation | |
10505 | ||
b8698a0f L |
10506 | Check whether an operation represented by the code CODE is a |
10507 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
10508 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
10509 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 10510 | |
1bda738b JJ |
10511 | Narrowing operations we currently support are NOP (CONVERT), FIX_TRUNC |
10512 | and FLOAT. This function checks if these operations are supported by | |
ebfd146a IR |
10513 | the target platform directly via vector tree-codes. |
10514 | ||
10515 | Output: | |
b8698a0f L |
10516 | - CODE1 is the code of a vector operation to be used when |
10517 | vectorizing the operation, if available. | |
ebfd146a IR |
10518 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
10519 | case of multi-step conversion (like int->short->char - in that case | |
10520 | MULTI_STEP_CVT will be 1). | |
10521 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 10522 | narrowing operation (short in the above example). */ |
ebfd146a IR |
10523 | |
10524 | bool | |
10525 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 10526 | tree vectype_out, tree vectype_in, |
ebfd146a | 10527 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 10528 | vec<tree> *interm_types) |
ebfd146a | 10529 | { |
ef4bddc2 | 10530 | machine_mode vec_mode; |
ebfd146a IR |
10531 | enum insn_code icode1; |
10532 | optab optab1, interm_optab; | |
b690cc0f RG |
10533 | tree vectype = vectype_in; |
10534 | tree narrow_vectype = vectype_out; | |
ebfd146a | 10535 | enum tree_code c1; |
3ae0661a | 10536 | tree intermediate_type, prev_type; |
ef4bddc2 | 10537 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 10538 | int i; |
4a00c761 | 10539 | bool uns; |
ebfd146a | 10540 | |
4a00c761 | 10541 | *multi_step_cvt = 0; |
ebfd146a IR |
10542 | switch (code) |
10543 | { | |
10544 | CASE_CONVERT: | |
10545 | c1 = VEC_PACK_TRUNC_EXPR; | |
10546 | break; | |
10547 | ||
10548 | case FIX_TRUNC_EXPR: | |
10549 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
10550 | break; | |
10551 | ||
10552 | case FLOAT_EXPR: | |
1bda738b JJ |
10553 | c1 = VEC_PACK_FLOAT_EXPR; |
10554 | break; | |
ebfd146a IR |
10555 | |
10556 | default: | |
10557 | gcc_unreachable (); | |
10558 | } | |
10559 | ||
10560 | if (code == FIX_TRUNC_EXPR) | |
10561 | /* The signedness is determined from output operand. */ | |
b690cc0f | 10562 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
10563 | else |
10564 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
10565 | ||
10566 | if (!optab1) | |
10567 | return false; | |
10568 | ||
10569 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 10570 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
10571 | return false; |
10572 | ||
4a00c761 JJ |
10573 | *code1 = c1; |
10574 | ||
10575 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
10576 | /* For scalar masks we may have different boolean |
10577 | vector types having the same QImode. Thus we | |
10578 | add additional check for elements number. */ | |
10579 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
928686b1 RS |
10580 | || known_eq (TYPE_VECTOR_SUBPARTS (vectype) * 2, |
10581 | TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 | 10582 | |
1bda738b JJ |
10583 | if (code == FLOAT_EXPR) |
10584 | return false; | |
10585 | ||
ebfd146a IR |
10586 | /* Check if it's a multi-step conversion that can be done using intermediate |
10587 | types. */ | |
4a00c761 | 10588 | prev_mode = vec_mode; |
3ae0661a | 10589 | prev_type = vectype; |
4a00c761 JJ |
10590 | if (code == FIX_TRUNC_EXPR) |
10591 | uns = TYPE_UNSIGNED (vectype_out); | |
10592 | else | |
10593 | uns = TYPE_UNSIGNED (vectype); | |
10594 | ||
10595 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
10596 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
10597 | costly than signed. */ | |
10598 | if (code == FIX_TRUNC_EXPR && uns) | |
10599 | { | |
10600 | enum insn_code icode2; | |
10601 | ||
10602 | intermediate_type | |
10603 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
10604 | interm_optab | |
10605 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 10606 | if (interm_optab != unknown_optab |
4a00c761 JJ |
10607 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
10608 | && insn_data[icode1].operand[0].mode | |
10609 | == insn_data[icode2].operand[0].mode) | |
10610 | { | |
10611 | uns = false; | |
10612 | optab1 = interm_optab; | |
10613 | icode1 = icode2; | |
10614 | } | |
10615 | } | |
ebfd146a | 10616 | |
4a00c761 JJ |
10617 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
10618 | intermediate steps in promotion sequence. We try | |
10619 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 10620 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
10621 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
10622 | { | |
10623 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
10624 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
10625 | { | |
7cfb4d93 | 10626 | intermediate_type = vect_double_mask_nunits (prev_type); |
3ae0661a | 10627 | if (intermediate_mode != TYPE_MODE (intermediate_type)) |
7cfb4d93 | 10628 | return false; |
3ae0661a IE |
10629 | } |
10630 | else | |
10631 | intermediate_type | |
10632 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
4a00c761 JJ |
10633 | interm_optab |
10634 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
10635 | optab_default); | |
10636 | if (!interm_optab | |
10637 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
10638 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
10639 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
10640 | == CODE_FOR_nothing)) | |
10641 | break; | |
10642 | ||
9771b263 | 10643 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
10644 | (*multi_step_cvt)++; |
10645 | ||
10646 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff | 10647 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
928686b1 RS |
10648 | || known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type) * 2, |
10649 | TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 JJ |
10650 | |
10651 | prev_mode = intermediate_mode; | |
3ae0661a | 10652 | prev_type = intermediate_type; |
4a00c761 | 10653 | optab1 = interm_optab; |
ebfd146a IR |
10654 | } |
10655 | ||
9771b263 | 10656 | interm_types->release (); |
4a00c761 | 10657 | return false; |
ebfd146a | 10658 | } |
7cfb4d93 RS |
10659 | |
10660 | /* Generate and return a statement that sets vector mask MASK such that | |
10661 | MASK[I] is true iff J + START_INDEX < END_INDEX for all J <= I. */ | |
10662 | ||
10663 | gcall * | |
10664 | vect_gen_while (tree mask, tree start_index, tree end_index) | |
10665 | { | |
10666 | tree cmp_type = TREE_TYPE (start_index); | |
10667 | tree mask_type = TREE_TYPE (mask); | |
10668 | gcc_checking_assert (direct_internal_fn_supported_p (IFN_WHILE_ULT, | |
10669 | cmp_type, mask_type, | |
10670 | OPTIMIZE_FOR_SPEED)); | |
10671 | gcall *call = gimple_build_call_internal (IFN_WHILE_ULT, 3, | |
10672 | start_index, end_index, | |
10673 | build_zero_cst (mask_type)); | |
10674 | gimple_call_set_lhs (call, mask); | |
10675 | return call; | |
10676 | } | |
535e7c11 RS |
10677 | |
10678 | /* Generate a vector mask of type MASK_TYPE for which index I is false iff | |
10679 | J + START_INDEX < END_INDEX for all J <= I. Add the statements to SEQ. */ | |
10680 | ||
10681 | tree | |
10682 | vect_gen_while_not (gimple_seq *seq, tree mask_type, tree start_index, | |
10683 | tree end_index) | |
10684 | { | |
10685 | tree tmp = make_ssa_name (mask_type); | |
10686 | gcall *call = vect_gen_while (tmp, start_index, end_index); | |
10687 | gimple_seq_add_stmt (seq, call); | |
10688 | return gimple_build (seq, BIT_NOT_EXPR, mask_type, tmp); | |
10689 | } | |
1f3cb663 RS |
10690 | |
10691 | /* Try to compute the vector types required to vectorize STMT_INFO, | |
10692 | returning true on success and false if vectorization isn't possible. | |
10693 | ||
10694 | On success: | |
10695 | ||
10696 | - Set *STMT_VECTYPE_OUT to: | |
10697 | - NULL_TREE if the statement doesn't need to be vectorized; | |
10698 | - boolean_type_node if the statement is a boolean operation whose | |
10699 | vector type can only be determined once all the other vector types | |
10700 | are known; and | |
10701 | - the equivalent of STMT_VINFO_VECTYPE otherwise. | |
10702 | ||
10703 | - Set *NUNITS_VECTYPE_OUT to the vector type that contains the maximum | |
10704 | number of units needed to vectorize STMT_INFO, or NULL_TREE if the | |
10705 | statement does not help to determine the overall number of units. */ | |
10706 | ||
10707 | bool | |
10708 | vect_get_vector_types_for_stmt (stmt_vec_info stmt_info, | |
10709 | tree *stmt_vectype_out, | |
10710 | tree *nunits_vectype_out) | |
10711 | { | |
10712 | gimple *stmt = stmt_info->stmt; | |
10713 | ||
10714 | *stmt_vectype_out = NULL_TREE; | |
10715 | *nunits_vectype_out = NULL_TREE; | |
10716 | ||
10717 | if (gimple_get_lhs (stmt) == NULL_TREE | |
10718 | /* MASK_STORE has no lhs, but is ok. */ | |
10719 | && !gimple_call_internal_p (stmt, IFN_MASK_STORE)) | |
10720 | { | |
10721 | if (is_a <gcall *> (stmt)) | |
10722 | { | |
10723 | /* Ignore calls with no lhs. These must be calls to | |
10724 | #pragma omp simd functions, and what vectorization factor | |
10725 | it really needs can't be determined until | |
10726 | vectorizable_simd_clone_call. */ | |
10727 | if (dump_enabled_p ()) | |
10728 | dump_printf_loc (MSG_NOTE, vect_location, | |
10729 | "defer to SIMD clone analysis.\n"); | |
10730 | return true; | |
10731 | } | |
10732 | ||
10733 | if (dump_enabled_p ()) | |
10734 | { | |
10735 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10736 | "not vectorized: irregular stmt."); | |
10737 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
10738 | } | |
10739 | return false; | |
10740 | } | |
10741 | ||
10742 | if (VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))) | |
10743 | { | |
10744 | if (dump_enabled_p ()) | |
10745 | { | |
10746 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10747 | "not vectorized: vector stmt in loop:"); | |
10748 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
10749 | } | |
10750 | return false; | |
10751 | } | |
10752 | ||
10753 | tree vectype; | |
10754 | tree scalar_type = NULL_TREE; | |
10755 | if (STMT_VINFO_VECTYPE (stmt_info)) | |
10756 | *stmt_vectype_out = vectype = STMT_VINFO_VECTYPE (stmt_info); | |
10757 | else | |
10758 | { | |
10759 | gcc_assert (!STMT_VINFO_DATA_REF (stmt_info)); | |
10760 | if (gimple_call_internal_p (stmt, IFN_MASK_STORE)) | |
10761 | scalar_type = TREE_TYPE (gimple_call_arg (stmt, 3)); | |
10762 | else | |
10763 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); | |
10764 | ||
10765 | /* Pure bool ops don't participate in number-of-units computation. | |
10766 | For comparisons use the types being compared. */ | |
10767 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type) | |
10768 | && is_gimple_assign (stmt) | |
10769 | && gimple_assign_rhs_code (stmt) != COND_EXPR) | |
10770 | { | |
10771 | *stmt_vectype_out = boolean_type_node; | |
10772 | ||
10773 | tree rhs1 = gimple_assign_rhs1 (stmt); | |
10774 | if (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_comparison | |
10775 | && !VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (rhs1))) | |
10776 | scalar_type = TREE_TYPE (rhs1); | |
10777 | else | |
10778 | { | |
10779 | if (dump_enabled_p ()) | |
10780 | dump_printf_loc (MSG_NOTE, vect_location, | |
10781 | "pure bool operation.\n"); | |
10782 | return true; | |
10783 | } | |
10784 | } | |
10785 | ||
10786 | if (dump_enabled_p ()) | |
10787 | { | |
10788 | dump_printf_loc (MSG_NOTE, vect_location, | |
10789 | "get vectype for scalar type: "); | |
10790 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
10791 | dump_printf (MSG_NOTE, "\n"); | |
10792 | } | |
10793 | vectype = get_vectype_for_scalar_type (scalar_type); | |
10794 | if (!vectype) | |
10795 | { | |
10796 | if (dump_enabled_p ()) | |
10797 | { | |
10798 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10799 | "not vectorized: unsupported data-type "); | |
10800 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10801 | scalar_type); | |
10802 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10803 | } | |
10804 | return false; | |
10805 | } | |
10806 | ||
10807 | if (!*stmt_vectype_out) | |
10808 | *stmt_vectype_out = vectype; | |
10809 | ||
10810 | if (dump_enabled_p ()) | |
10811 | { | |
10812 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); | |
10813 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
10814 | dump_printf (MSG_NOTE, "\n"); | |
10815 | } | |
10816 | } | |
10817 | ||
10818 | /* Don't try to compute scalar types if the stmt produces a boolean | |
10819 | vector; use the existing vector type instead. */ | |
10820 | tree nunits_vectype; | |
10821 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) | |
10822 | nunits_vectype = vectype; | |
10823 | else | |
10824 | { | |
10825 | /* The number of units is set according to the smallest scalar | |
10826 | type (or the largest vector size, but we only support one | |
10827 | vector size per vectorization). */ | |
10828 | if (*stmt_vectype_out != boolean_type_node) | |
10829 | { | |
10830 | HOST_WIDE_INT dummy; | |
10831 | scalar_type = vect_get_smallest_scalar_type (stmt, &dummy, &dummy); | |
10832 | } | |
10833 | if (dump_enabled_p ()) | |
10834 | { | |
10835 | dump_printf_loc (MSG_NOTE, vect_location, | |
10836 | "get vectype for scalar type: "); | |
10837 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
10838 | dump_printf (MSG_NOTE, "\n"); | |
10839 | } | |
10840 | nunits_vectype = get_vectype_for_scalar_type (scalar_type); | |
10841 | } | |
10842 | if (!nunits_vectype) | |
10843 | { | |
10844 | if (dump_enabled_p ()) | |
10845 | { | |
10846 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10847 | "not vectorized: unsupported data-type "); | |
10848 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, scalar_type); | |
10849 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10850 | } | |
10851 | return false; | |
10852 | } | |
10853 | ||
10854 | if (maybe_ne (GET_MODE_SIZE (TYPE_MODE (vectype)), | |
10855 | GET_MODE_SIZE (TYPE_MODE (nunits_vectype)))) | |
10856 | { | |
10857 | if (dump_enabled_p ()) | |
10858 | { | |
10859 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10860 | "not vectorized: different sized vector " | |
10861 | "types in statement, "); | |
10862 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, vectype); | |
10863 | dump_printf (MSG_MISSED_OPTIMIZATION, " and "); | |
10864 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, nunits_vectype); | |
10865 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10866 | } | |
10867 | return false; | |
10868 | } | |
10869 | ||
10870 | if (dump_enabled_p ()) | |
10871 | { | |
10872 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); | |
10873 | dump_generic_expr (MSG_NOTE, TDF_SLIM, nunits_vectype); | |
10874 | dump_printf (MSG_NOTE, "\n"); | |
10875 | ||
10876 | dump_printf_loc (MSG_NOTE, vect_location, "nunits = "); | |
10877 | dump_dec (MSG_NOTE, TYPE_VECTOR_SUBPARTS (nunits_vectype)); | |
10878 | dump_printf (MSG_NOTE, "\n"); | |
10879 | } | |
10880 | ||
10881 | *nunits_vectype_out = nunits_vectype; | |
10882 | return true; | |
10883 | } | |
10884 | ||
10885 | /* Try to determine the correct vector type for STMT_INFO, which is a | |
10886 | statement that produces a scalar boolean result. Return the vector | |
10887 | type on success, otherwise return NULL_TREE. */ | |
10888 | ||
10889 | tree | |
10890 | vect_get_mask_type_for_stmt (stmt_vec_info stmt_info) | |
10891 | { | |
10892 | gimple *stmt = stmt_info->stmt; | |
10893 | tree mask_type = NULL; | |
10894 | tree vectype, scalar_type; | |
10895 | ||
10896 | if (is_gimple_assign (stmt) | |
10897 | && TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_comparison | |
10898 | && !VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (gimple_assign_rhs1 (stmt)))) | |
10899 | { | |
10900 | scalar_type = TREE_TYPE (gimple_assign_rhs1 (stmt)); | |
10901 | mask_type = get_mask_type_for_scalar_type (scalar_type); | |
10902 | ||
10903 | if (!mask_type) | |
10904 | { | |
10905 | if (dump_enabled_p ()) | |
10906 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10907 | "not vectorized: unsupported mask\n"); | |
10908 | return NULL_TREE; | |
10909 | } | |
10910 | } | |
10911 | else | |
10912 | { | |
10913 | tree rhs; | |
10914 | ssa_op_iter iter; | |
1f3cb663 RS |
10915 | enum vect_def_type dt; |
10916 | ||
10917 | FOR_EACH_SSA_TREE_OPERAND (rhs, stmt, iter, SSA_OP_USE) | |
10918 | { | |
894dd753 | 10919 | if (!vect_is_simple_use (rhs, stmt_info->vinfo, &dt, &vectype)) |
1f3cb663 RS |
10920 | { |
10921 | if (dump_enabled_p ()) | |
10922 | { | |
10923 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10924 | "not vectorized: can't compute mask type " | |
10925 | "for statement, "); | |
10926 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, | |
10927 | 0); | |
10928 | } | |
10929 | return NULL_TREE; | |
10930 | } | |
10931 | ||
10932 | /* No vectype probably means external definition. | |
10933 | Allow it in case there is another operand which | |
10934 | allows to determine mask type. */ | |
10935 | if (!vectype) | |
10936 | continue; | |
10937 | ||
10938 | if (!mask_type) | |
10939 | mask_type = vectype; | |
10940 | else if (maybe_ne (TYPE_VECTOR_SUBPARTS (mask_type), | |
10941 | TYPE_VECTOR_SUBPARTS (vectype))) | |
10942 | { | |
10943 | if (dump_enabled_p ()) | |
10944 | { | |
10945 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10946 | "not vectorized: different sized masks " | |
10947 | "types in statement, "); | |
10948 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10949 | mask_type); | |
10950 | dump_printf (MSG_MISSED_OPTIMIZATION, " and "); | |
10951 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10952 | vectype); | |
10953 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10954 | } | |
10955 | return NULL_TREE; | |
10956 | } | |
10957 | else if (VECTOR_BOOLEAN_TYPE_P (mask_type) | |
10958 | != VECTOR_BOOLEAN_TYPE_P (vectype)) | |
10959 | { | |
10960 | if (dump_enabled_p ()) | |
10961 | { | |
10962 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10963 | "not vectorized: mixed mask and " | |
10964 | "nonmask vector types in statement, "); | |
10965 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10966 | mask_type); | |
10967 | dump_printf (MSG_MISSED_OPTIMIZATION, " and "); | |
10968 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10969 | vectype); | |
10970 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10971 | } | |
10972 | return NULL_TREE; | |
10973 | } | |
10974 | } | |
10975 | ||
10976 | /* We may compare boolean value loaded as vector of integers. | |
10977 | Fix mask_type in such case. */ | |
10978 | if (mask_type | |
10979 | && !VECTOR_BOOLEAN_TYPE_P (mask_type) | |
10980 | && gimple_code (stmt) == GIMPLE_ASSIGN | |
10981 | && TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_comparison) | |
10982 | mask_type = build_same_sized_truth_vector_type (mask_type); | |
10983 | } | |
10984 | ||
10985 | /* No mask_type should mean loop invariant predicate. | |
10986 | This is probably a subject for optimization in if-conversion. */ | |
10987 | if (!mask_type && dump_enabled_p ()) | |
10988 | { | |
10989 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10990 | "not vectorized: can't compute mask type " | |
10991 | "for statement, "); | |
10992 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
10993 | } | |
10994 | return mask_type; | |
10995 | } |