]>
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 | 100 | |
211ee39b | 101 | stmt_info_for_cost si = { count, kind, where, stmt_info, misalign }; |
68435eb2 RB |
102 | body_cost_vec->safe_push (si); |
103 | ||
104 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; | |
105 | return (unsigned) | |
106 | (builtin_vectorization_cost (kind, vectype, misalign) * count); | |
c3e7ee41 BS |
107 | } |
108 | ||
272c6793 RS |
109 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
110 | ||
111 | static tree | |
112 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
113 | { | |
114 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
115 | "vect_array"); | |
116 | } | |
117 | ||
118 | /* ARRAY is an array of vectors created by create_vector_array. | |
119 | Return an SSA_NAME for the vector in index N. The reference | |
120 | is part of the vectorization of STMT and the vector is associated | |
121 | with scalar destination SCALAR_DEST. */ | |
122 | ||
123 | static tree | |
355fe088 | 124 | read_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree scalar_dest, |
272c6793 RS |
125 | tree array, unsigned HOST_WIDE_INT n) |
126 | { | |
127 | tree vect_type, vect, vect_name, array_ref; | |
355fe088 | 128 | gimple *new_stmt; |
272c6793 RS |
129 | |
130 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
131 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
132 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
133 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
134 | build_int_cst (size_type_node, n), | |
135 | NULL_TREE, NULL_TREE); | |
136 | ||
137 | new_stmt = gimple_build_assign (vect, array_ref); | |
138 | vect_name = make_ssa_name (vect, new_stmt); | |
139 | gimple_assign_set_lhs (new_stmt, vect_name); | |
140 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
141 | |
142 | return vect_name; | |
143 | } | |
144 | ||
145 | /* ARRAY is an array of vectors created by create_vector_array. | |
146 | Emit code to store SSA_NAME VECT in index N of the array. | |
147 | The store is part of the vectorization of STMT. */ | |
148 | ||
149 | static void | |
355fe088 | 150 | write_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree vect, |
272c6793 RS |
151 | tree array, unsigned HOST_WIDE_INT n) |
152 | { | |
153 | tree array_ref; | |
355fe088 | 154 | gimple *new_stmt; |
272c6793 RS |
155 | |
156 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
157 | build_int_cst (size_type_node, n), | |
158 | NULL_TREE, NULL_TREE); | |
159 | ||
160 | new_stmt = gimple_build_assign (array_ref, vect); | |
161 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
162 | } |
163 | ||
164 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
165 | of *PTR. The memory reference replaces those in FIRST_DR | |
166 | (and its group). */ | |
167 | ||
168 | static tree | |
44fc7854 | 169 | create_array_ref (tree type, tree ptr, tree alias_ptr_type) |
272c6793 | 170 | { |
44fc7854 | 171 | tree mem_ref; |
272c6793 | 172 | |
272c6793 RS |
173 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); |
174 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 175 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
176 | return mem_ref; |
177 | } | |
178 | ||
3ba4ff41 RS |
179 | /* Add a clobber of variable VAR to the vectorization of STMT. |
180 | Emit the clobber before *GSI. */ | |
181 | ||
182 | static void | |
183 | vect_clobber_variable (gimple *stmt, gimple_stmt_iterator *gsi, tree var) | |
184 | { | |
185 | tree clobber = build_clobber (TREE_TYPE (var)); | |
186 | gimple *new_stmt = gimple_build_assign (var, clobber); | |
187 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
188 | } | |
189 | ||
ebfd146a IR |
190 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
191 | ||
192 | /* Function vect_mark_relevant. | |
193 | ||
194 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
195 | ||
196 | static void | |
355fe088 | 197 | vect_mark_relevant (vec<gimple *> *worklist, gimple *stmt, |
97ecdb46 | 198 | enum vect_relevant relevant, bool live_p) |
ebfd146a IR |
199 | { |
200 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
201 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
202 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
203 | ||
73fbfcad | 204 | if (dump_enabled_p ()) |
66c16fd9 RB |
205 | { |
206 | dump_printf_loc (MSG_NOTE, vect_location, | |
207 | "mark relevant %d, live %d: ", relevant, live_p); | |
208 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
209 | } | |
ebfd146a | 210 | |
83197f37 IR |
211 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
212 | related pattern stmt instead of the original stmt. However, such stmts | |
213 | may have their own uses that are not in any pattern, in such cases the | |
214 | stmt itself should be marked. */ | |
ebfd146a IR |
215 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
216 | { | |
97ecdb46 JJ |
217 | /* This is the last stmt in a sequence that was detected as a |
218 | pattern that can potentially be vectorized. Don't mark the stmt | |
219 | as relevant/live because it's not going to be vectorized. | |
220 | Instead mark the pattern-stmt that replaces it. */ | |
83197f37 | 221 | |
97ecdb46 JJ |
222 | if (dump_enabled_p ()) |
223 | dump_printf_loc (MSG_NOTE, vect_location, | |
224 | "last stmt in pattern. don't mark" | |
225 | " relevant/live.\n"); | |
10681ce8 RS |
226 | stmt_vec_info old_stmt_info = stmt_info; |
227 | stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); | |
228 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == old_stmt_info); | |
97ecdb46 JJ |
229 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); |
230 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
10681ce8 | 231 | stmt = stmt_info->stmt; |
ebfd146a IR |
232 | } |
233 | ||
234 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
235 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
236 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
237 | ||
238 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
239 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
240 | { | |
73fbfcad | 241 | if (dump_enabled_p ()) |
78c60e3d | 242 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 243 | "already marked relevant/live.\n"); |
ebfd146a IR |
244 | return; |
245 | } | |
246 | ||
9771b263 | 247 | worklist->safe_push (stmt); |
ebfd146a IR |
248 | } |
249 | ||
250 | ||
b28ead45 AH |
251 | /* Function is_simple_and_all_uses_invariant |
252 | ||
253 | Return true if STMT is simple and all uses of it are invariant. */ | |
254 | ||
255 | bool | |
256 | is_simple_and_all_uses_invariant (gimple *stmt, loop_vec_info loop_vinfo) | |
257 | { | |
258 | tree op; | |
b28ead45 AH |
259 | ssa_op_iter iter; |
260 | ||
261 | if (!is_gimple_assign (stmt)) | |
262 | return false; | |
263 | ||
264 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
265 | { | |
266 | enum vect_def_type dt = vect_uninitialized_def; | |
267 | ||
894dd753 | 268 | if (!vect_is_simple_use (op, loop_vinfo, &dt)) |
b28ead45 AH |
269 | { |
270 | if (dump_enabled_p ()) | |
271 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
272 | "use not simple.\n"); | |
273 | return false; | |
274 | } | |
275 | ||
276 | if (dt != vect_external_def && dt != vect_constant_def) | |
277 | return false; | |
278 | } | |
279 | return true; | |
280 | } | |
281 | ||
ebfd146a IR |
282 | /* Function vect_stmt_relevant_p. |
283 | ||
284 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
285 | "relevant for vectorization". | |
286 | ||
287 | A stmt is considered "relevant for vectorization" if: | |
288 | - it has uses outside the loop. | |
289 | - it has vdefs (it alters memory). | |
290 | - control stmts in the loop (except for the exit condition). | |
291 | ||
292 | CHECKME: what other side effects would the vectorizer allow? */ | |
293 | ||
294 | static bool | |
355fe088 | 295 | vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo, |
ebfd146a IR |
296 | enum vect_relevant *relevant, bool *live_p) |
297 | { | |
298 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
299 | ssa_op_iter op_iter; | |
300 | imm_use_iterator imm_iter; | |
301 | use_operand_p use_p; | |
302 | def_operand_p def_p; | |
303 | ||
8644a673 | 304 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
305 | *live_p = false; |
306 | ||
307 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
308 | if (is_ctrl_stmt (stmt) |
309 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
310 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 311 | *relevant = vect_used_in_scope; |
ebfd146a IR |
312 | |
313 | /* changing memory. */ | |
314 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
315 | if (gimple_vdef (stmt) |
316 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 317 | { |
73fbfcad | 318 | if (dump_enabled_p ()) |
78c60e3d | 319 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 320 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 321 | *relevant = vect_used_in_scope; |
ebfd146a IR |
322 | } |
323 | ||
324 | /* uses outside the loop. */ | |
325 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
326 | { | |
327 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
328 | { | |
329 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
330 | if (!flow_bb_inside_loop_p (loop, bb)) | |
331 | { | |
73fbfcad | 332 | if (dump_enabled_p ()) |
78c60e3d | 333 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 334 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 335 | |
3157b0c2 AO |
336 | if (is_gimple_debug (USE_STMT (use_p))) |
337 | continue; | |
338 | ||
ebfd146a IR |
339 | /* We expect all such uses to be in the loop exit phis |
340 | (because of loop closed form) */ | |
341 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
342 | gcc_assert (bb == single_exit (loop)->dest); | |
343 | ||
344 | *live_p = true; | |
345 | } | |
346 | } | |
347 | } | |
348 | ||
3a2edf4c AH |
349 | if (*live_p && *relevant == vect_unused_in_scope |
350 | && !is_simple_and_all_uses_invariant (stmt, loop_vinfo)) | |
b28ead45 AH |
351 | { |
352 | if (dump_enabled_p ()) | |
353 | dump_printf_loc (MSG_NOTE, vect_location, | |
354 | "vec_stmt_relevant_p: stmt live but not relevant.\n"); | |
355 | *relevant = vect_used_only_live; | |
356 | } | |
357 | ||
ebfd146a IR |
358 | return (*live_p || *relevant); |
359 | } | |
360 | ||
361 | ||
b8698a0f | 362 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 363 | |
ff802fa1 | 364 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
365 | used in STMT for anything other than indexing an array. */ |
366 | ||
367 | static bool | |
355fe088 | 368 | exist_non_indexing_operands_for_use_p (tree use, gimple *stmt) |
ebfd146a IR |
369 | { |
370 | tree operand; | |
371 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 372 | |
ff802fa1 | 373 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
374 | reference in STMT, then any operand that corresponds to USE |
375 | is not indexing an array. */ | |
376 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
377 | return true; | |
59a05b0c | 378 | |
ebfd146a IR |
379 | /* STMT has a data_ref. FORNOW this means that its of one of |
380 | the following forms: | |
381 | -1- ARRAY_REF = var | |
382 | -2- var = ARRAY_REF | |
383 | (This should have been verified in analyze_data_refs). | |
384 | ||
385 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 386 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
387 | for array indexing. |
388 | ||
389 | Therefore, all we need to check is if STMT falls into the | |
390 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
391 | |
392 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
393 | { |
394 | if (is_gimple_call (stmt) | |
395 | && gimple_call_internal_p (stmt)) | |
bfaa08b7 RS |
396 | { |
397 | internal_fn ifn = gimple_call_internal_fn (stmt); | |
398 | int mask_index = internal_fn_mask_index (ifn); | |
399 | if (mask_index >= 0 | |
400 | && use == gimple_call_arg (stmt, mask_index)) | |
401 | return true; | |
f307441a RS |
402 | int stored_value_index = internal_fn_stored_value_index (ifn); |
403 | if (stored_value_index >= 0 | |
404 | && use == gimple_call_arg (stmt, stored_value_index)) | |
405 | return true; | |
bfaa08b7 RS |
406 | if (internal_gather_scatter_fn_p (ifn) |
407 | && use == gimple_call_arg (stmt, 1)) | |
408 | return true; | |
bfaa08b7 | 409 | } |
5ce9450f JJ |
410 | return false; |
411 | } | |
412 | ||
59a05b0c EB |
413 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
414 | return false; | |
ebfd146a | 415 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
416 | if (TREE_CODE (operand) != SSA_NAME) |
417 | return false; | |
418 | ||
419 | if (operand == use) | |
420 | return true; | |
421 | ||
422 | return false; | |
423 | } | |
424 | ||
425 | ||
b8698a0f | 426 | /* |
ebfd146a IR |
427 | Function process_use. |
428 | ||
429 | Inputs: | |
430 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b28ead45 | 431 | - RELEVANT - enum value to be set in the STMT_VINFO of the stmt |
ff802fa1 | 432 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 433 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
434 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
435 | be performed. | |
ebfd146a IR |
436 | |
437 | Outputs: | |
438 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
439 | relevance info of the DEF_STMT of this USE: | |
440 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
441 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
442 | Exceptions: | |
443 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 444 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 445 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
446 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
447 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
448 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
449 | be modified accordingly. | |
450 | ||
451 | Return true if everything is as expected. Return false otherwise. */ | |
452 | ||
453 | static bool | |
b28ead45 | 454 | process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo, |
355fe088 | 455 | enum vect_relevant relevant, vec<gimple *> *worklist, |
aec7ae7d | 456 | bool force) |
ebfd146a | 457 | { |
ebfd146a IR |
458 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
459 | stmt_vec_info dstmt_vinfo; | |
460 | basic_block bb, def_bb; | |
ebfd146a IR |
461 | enum vect_def_type dt; |
462 | ||
b8698a0f | 463 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 464 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 465 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
466 | return true; |
467 | ||
fef96d8e | 468 | if (!vect_is_simple_use (use, loop_vinfo, &dt, &dstmt_vinfo)) |
b8698a0f | 469 | { |
73fbfcad | 470 | if (dump_enabled_p ()) |
78c60e3d | 471 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 472 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
473 | return false; |
474 | } | |
475 | ||
fef96d8e | 476 | if (!dstmt_vinfo) |
ebfd146a IR |
477 | return true; |
478 | ||
fef96d8e | 479 | def_bb = gimple_bb (dstmt_vinfo->stmt); |
ebfd146a | 480 | |
fef96d8e RS |
481 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DSTMT_VINFO). |
482 | DSTMT_VINFO must have already been processed, because this should be the | |
b8698a0f | 483 | only way that STMT, which is a reduction-phi, was put in the worklist, |
fef96d8e | 484 | as there should be no other uses for DSTMT_VINFO in the loop. So we just |
ebfd146a | 485 | check that everything is as expected, and we are done. */ |
ebfd146a IR |
486 | bb = gimple_bb (stmt); |
487 | if (gimple_code (stmt) == GIMPLE_PHI | |
488 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
fef96d8e | 489 | && gimple_code (dstmt_vinfo->stmt) != GIMPLE_PHI |
ebfd146a IR |
490 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def |
491 | && bb->loop_father == def_bb->loop_father) | |
492 | { | |
73fbfcad | 493 | if (dump_enabled_p ()) |
78c60e3d | 494 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 495 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a | 496 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); |
b8698a0f | 497 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 498 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
499 | return true; |
500 | } | |
501 | ||
502 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
503 | outer-loop-header-bb: | |
fef96d8e | 504 | d = dstmt_vinfo |
ebfd146a IR |
505 | inner-loop: |
506 | stmt # use (d) | |
507 | outer-loop-tail-bb: | |
508 | ... */ | |
509 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
510 | { | |
73fbfcad | 511 | if (dump_enabled_p ()) |
78c60e3d | 512 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 513 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 514 | |
ebfd146a IR |
515 | switch (relevant) |
516 | { | |
8644a673 | 517 | case vect_unused_in_scope: |
7c5222ff IR |
518 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
519 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 520 | break; |
7c5222ff | 521 | |
ebfd146a | 522 | case vect_used_in_outer_by_reduction: |
7c5222ff | 523 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
524 | relevant = vect_used_by_reduction; |
525 | break; | |
7c5222ff | 526 | |
ebfd146a | 527 | case vect_used_in_outer: |
7c5222ff | 528 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 529 | relevant = vect_used_in_scope; |
ebfd146a | 530 | break; |
7c5222ff | 531 | |
8644a673 | 532 | case vect_used_in_scope: |
ebfd146a IR |
533 | break; |
534 | ||
535 | default: | |
536 | gcc_unreachable (); | |
b8698a0f | 537 | } |
ebfd146a IR |
538 | } |
539 | ||
540 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
541 | outer-loop-header-bb: | |
542 | ... | |
543 | inner-loop: | |
fef96d8e | 544 | d = dstmt_vinfo |
06066f92 | 545 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
546 | stmt # use (d) */ |
547 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
548 | { | |
73fbfcad | 549 | if (dump_enabled_p ()) |
78c60e3d | 550 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 551 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 552 | |
ebfd146a IR |
553 | switch (relevant) |
554 | { | |
8644a673 | 555 | case vect_unused_in_scope: |
b8698a0f | 556 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 557 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 558 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
559 | break; |
560 | ||
ebfd146a | 561 | case vect_used_by_reduction: |
b28ead45 | 562 | case vect_used_only_live: |
ebfd146a IR |
563 | relevant = vect_used_in_outer_by_reduction; |
564 | break; | |
565 | ||
8644a673 | 566 | case vect_used_in_scope: |
ebfd146a IR |
567 | relevant = vect_used_in_outer; |
568 | break; | |
569 | ||
570 | default: | |
571 | gcc_unreachable (); | |
572 | } | |
573 | } | |
643a9684 RB |
574 | /* We are also not interested in uses on loop PHI backedges that are |
575 | inductions. Otherwise we'll needlessly vectorize the IV increment | |
e294f495 RB |
576 | and cause hybrid SLP for SLP inductions. Unless the PHI is live |
577 | of course. */ | |
643a9684 RB |
578 | else if (gimple_code (stmt) == GIMPLE_PHI |
579 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_induction_def | |
e294f495 | 580 | && ! STMT_VINFO_LIVE_P (stmt_vinfo) |
643a9684 RB |
581 | && (PHI_ARG_DEF_FROM_EDGE (stmt, loop_latch_edge (bb->loop_father)) |
582 | == use)) | |
583 | { | |
584 | if (dump_enabled_p ()) | |
585 | dump_printf_loc (MSG_NOTE, vect_location, | |
586 | "induction value on backedge.\n"); | |
587 | return true; | |
588 | } | |
589 | ||
ebfd146a | 590 | |
fef96d8e | 591 | vect_mark_relevant (worklist, dstmt_vinfo, relevant, false); |
ebfd146a IR |
592 | return true; |
593 | } | |
594 | ||
595 | ||
596 | /* Function vect_mark_stmts_to_be_vectorized. | |
597 | ||
598 | Not all stmts in the loop need to be vectorized. For example: | |
599 | ||
600 | for i... | |
601 | for j... | |
602 | 1. T0 = i + j | |
603 | 2. T1 = a[T0] | |
604 | ||
605 | 3. j = j + 1 | |
606 | ||
607 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
608 | addressing of vectorized data-refs are handled differently. | |
609 | ||
610 | This pass detects such stmts. */ | |
611 | ||
612 | bool | |
613 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
614 | { | |
ebfd146a IR |
615 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
616 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
617 | unsigned int nbbs = loop->num_nodes; | |
618 | gimple_stmt_iterator si; | |
355fe088 | 619 | gimple *stmt; |
ebfd146a IR |
620 | unsigned int i; |
621 | stmt_vec_info stmt_vinfo; | |
622 | basic_block bb; | |
355fe088 | 623 | gimple *phi; |
ebfd146a | 624 | bool live_p; |
b28ead45 | 625 | enum vect_relevant relevant; |
ebfd146a | 626 | |
adac3a68 | 627 | DUMP_VECT_SCOPE ("vect_mark_stmts_to_be_vectorized"); |
ebfd146a | 628 | |
355fe088 | 629 | auto_vec<gimple *, 64> worklist; |
ebfd146a IR |
630 | |
631 | /* 1. Init worklist. */ | |
632 | for (i = 0; i < nbbs; i++) | |
633 | { | |
634 | bb = bbs[i]; | |
635 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 636 | { |
ebfd146a | 637 | phi = gsi_stmt (si); |
73fbfcad | 638 | if (dump_enabled_p ()) |
ebfd146a | 639 | { |
78c60e3d SS |
640 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
641 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
642 | } |
643 | ||
644 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 645 | vect_mark_relevant (&worklist, phi, relevant, live_p); |
ebfd146a IR |
646 | } |
647 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
648 | { | |
649 | stmt = gsi_stmt (si); | |
73fbfcad | 650 | if (dump_enabled_p ()) |
ebfd146a | 651 | { |
78c60e3d SS |
652 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
653 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 654 | } |
ebfd146a IR |
655 | |
656 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 657 | vect_mark_relevant (&worklist, stmt, relevant, live_p); |
ebfd146a IR |
658 | } |
659 | } | |
660 | ||
661 | /* 2. Process_worklist */ | |
9771b263 | 662 | while (worklist.length () > 0) |
ebfd146a IR |
663 | { |
664 | use_operand_p use_p; | |
665 | ssa_op_iter iter; | |
666 | ||
9771b263 | 667 | stmt = worklist.pop (); |
73fbfcad | 668 | if (dump_enabled_p ()) |
ebfd146a | 669 | { |
78c60e3d SS |
670 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
671 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
672 | } |
673 | ||
b8698a0f | 674 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
b28ead45 AH |
675 | (DEF_STMT) as relevant/irrelevant according to the relevance property |
676 | of STMT. */ | |
ebfd146a IR |
677 | stmt_vinfo = vinfo_for_stmt (stmt); |
678 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
ebfd146a | 679 | |
b28ead45 AH |
680 | /* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is |
681 | propagated as is to the DEF_STMTs of its USEs. | |
ebfd146a IR |
682 | |
683 | One exception is when STMT has been identified as defining a reduction | |
b28ead45 | 684 | variable; in this case we set the relevance to vect_used_by_reduction. |
ebfd146a | 685 | This is because we distinguish between two kinds of relevant stmts - |
b8698a0f | 686 | those that are used by a reduction computation, and those that are |
ff802fa1 | 687 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 688 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 689 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 690 | |
b28ead45 | 691 | switch (STMT_VINFO_DEF_TYPE (stmt_vinfo)) |
ebfd146a | 692 | { |
06066f92 | 693 | case vect_reduction_def: |
b28ead45 AH |
694 | gcc_assert (relevant != vect_unused_in_scope); |
695 | if (relevant != vect_unused_in_scope | |
696 | && relevant != vect_used_in_scope | |
697 | && relevant != vect_used_by_reduction | |
698 | && relevant != vect_used_only_live) | |
06066f92 | 699 | { |
b28ead45 AH |
700 | if (dump_enabled_p ()) |
701 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
702 | "unsupported use of reduction.\n"); | |
703 | return false; | |
06066f92 | 704 | } |
06066f92 | 705 | break; |
b8698a0f | 706 | |
06066f92 | 707 | case vect_nested_cycle: |
b28ead45 AH |
708 | if (relevant != vect_unused_in_scope |
709 | && relevant != vect_used_in_outer_by_reduction | |
710 | && relevant != vect_used_in_outer) | |
06066f92 | 711 | { |
73fbfcad | 712 | if (dump_enabled_p ()) |
78c60e3d | 713 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 714 | "unsupported use of nested cycle.\n"); |
7c5222ff | 715 | |
06066f92 IR |
716 | return false; |
717 | } | |
b8698a0f L |
718 | break; |
719 | ||
06066f92 | 720 | case vect_double_reduction_def: |
b28ead45 AH |
721 | if (relevant != vect_unused_in_scope |
722 | && relevant != vect_used_by_reduction | |
723 | && relevant != vect_used_only_live) | |
06066f92 | 724 | { |
73fbfcad | 725 | if (dump_enabled_p ()) |
78c60e3d | 726 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 727 | "unsupported use of double reduction.\n"); |
7c5222ff | 728 | |
7c5222ff | 729 | return false; |
06066f92 | 730 | } |
b8698a0f | 731 | break; |
7c5222ff | 732 | |
06066f92 IR |
733 | default: |
734 | break; | |
7c5222ff | 735 | } |
b8698a0f | 736 | |
aec7ae7d | 737 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
738 | { |
739 | /* Pattern statements are not inserted into the code, so | |
740 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
741 | have to scan the RHS or function arguments instead. */ | |
742 | if (is_gimple_assign (stmt)) | |
743 | { | |
69d2aade JJ |
744 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
745 | tree op = gimple_assign_rhs1 (stmt); | |
746 | ||
747 | i = 1; | |
748 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
749 | { | |
750 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
b28ead45 | 751 | relevant, &worklist, false) |
69d2aade | 752 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
b28ead45 | 753 | relevant, &worklist, false)) |
566d377a | 754 | return false; |
69d2aade JJ |
755 | i = 2; |
756 | } | |
757 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 758 | { |
69d2aade | 759 | op = gimple_op (stmt, i); |
afbe6325 | 760 | if (TREE_CODE (op) == SSA_NAME |
b28ead45 | 761 | && !process_use (stmt, op, loop_vinfo, relevant, |
afbe6325 | 762 | &worklist, false)) |
07687835 | 763 | return false; |
9d5e7640 IR |
764 | } |
765 | } | |
766 | else if (is_gimple_call (stmt)) | |
767 | { | |
768 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
769 | { | |
770 | tree arg = gimple_call_arg (stmt, i); | |
b28ead45 | 771 | if (!process_use (stmt, arg, loop_vinfo, relevant, |
aec7ae7d | 772 | &worklist, false)) |
07687835 | 773 | return false; |
9d5e7640 IR |
774 | } |
775 | } | |
776 | } | |
777 | else | |
778 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
779 | { | |
780 | tree op = USE_FROM_PTR (use_p); | |
b28ead45 | 781 | if (!process_use (stmt, op, loop_vinfo, relevant, |
aec7ae7d | 782 | &worklist, false)) |
07687835 | 783 | return false; |
9d5e7640 | 784 | } |
aec7ae7d | 785 | |
3bab6342 | 786 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d | 787 | { |
134c85ca RS |
788 | gather_scatter_info gs_info; |
789 | if (!vect_check_gather_scatter (stmt, loop_vinfo, &gs_info)) | |
790 | gcc_unreachable (); | |
791 | if (!process_use (stmt, gs_info.offset, loop_vinfo, relevant, | |
792 | &worklist, true)) | |
566d377a | 793 | return false; |
aec7ae7d | 794 | } |
ebfd146a IR |
795 | } /* while worklist */ |
796 | ||
ebfd146a IR |
797 | return true; |
798 | } | |
799 | ||
68435eb2 RB |
800 | /* Compute the prologue cost for invariant or constant operands. */ |
801 | ||
802 | static unsigned | |
803 | vect_prologue_cost_for_slp_op (slp_tree node, stmt_vec_info stmt_info, | |
804 | unsigned opno, enum vect_def_type dt, | |
805 | stmt_vector_for_cost *cost_vec) | |
806 | { | |
b9787581 | 807 | gimple *stmt = SLP_TREE_SCALAR_STMTS (node)[0]->stmt; |
68435eb2 RB |
808 | tree op = gimple_op (stmt, opno); |
809 | unsigned prologue_cost = 0; | |
810 | ||
811 | /* Without looking at the actual initializer a vector of | |
812 | constants can be implemented as load from the constant pool. | |
813 | When all elements are the same we can use a splat. */ | |
814 | tree vectype = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
815 | unsigned group_size = SLP_TREE_SCALAR_STMTS (node).length (); | |
816 | unsigned num_vects_to_check; | |
817 | unsigned HOST_WIDE_INT const_nunits; | |
818 | unsigned nelt_limit; | |
819 | if (TYPE_VECTOR_SUBPARTS (vectype).is_constant (&const_nunits) | |
820 | && ! multiple_p (const_nunits, group_size)) | |
821 | { | |
822 | num_vects_to_check = SLP_TREE_NUMBER_OF_VEC_STMTS (node); | |
823 | nelt_limit = const_nunits; | |
824 | } | |
825 | else | |
826 | { | |
827 | /* If either the vector has variable length or the vectors | |
828 | are composed of repeated whole groups we only need to | |
829 | cost construction once. All vectors will be the same. */ | |
830 | num_vects_to_check = 1; | |
831 | nelt_limit = group_size; | |
832 | } | |
833 | tree elt = NULL_TREE; | |
834 | unsigned nelt = 0; | |
835 | for (unsigned j = 0; j < num_vects_to_check * nelt_limit; ++j) | |
836 | { | |
837 | unsigned si = j % group_size; | |
838 | if (nelt == 0) | |
b9787581 | 839 | elt = gimple_op (SLP_TREE_SCALAR_STMTS (node)[si]->stmt, opno); |
68435eb2 RB |
840 | /* ??? We're just tracking whether all operands of a single |
841 | vector initializer are the same, ideally we'd check if | |
842 | we emitted the same one already. */ | |
b9787581 | 843 | else if (elt != gimple_op (SLP_TREE_SCALAR_STMTS (node)[si]->stmt, |
68435eb2 RB |
844 | opno)) |
845 | elt = NULL_TREE; | |
846 | nelt++; | |
847 | if (nelt == nelt_limit) | |
848 | { | |
849 | /* ??? We need to pass down stmt_info for a vector type | |
850 | even if it points to the wrong stmt. */ | |
851 | prologue_cost += record_stmt_cost | |
852 | (cost_vec, 1, | |
853 | dt == vect_external_def | |
854 | ? (elt ? scalar_to_vec : vec_construct) | |
855 | : vector_load, | |
856 | stmt_info, 0, vect_prologue); | |
857 | nelt = 0; | |
858 | } | |
859 | } | |
860 | ||
861 | return prologue_cost; | |
862 | } | |
ebfd146a | 863 | |
b8698a0f | 864 | /* Function vect_model_simple_cost. |
ebfd146a | 865 | |
b8698a0f | 866 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
867 | single op. Right now, this does not account for multiple insns that could |
868 | be generated for the single vector op. We will handle that shortly. */ | |
869 | ||
68435eb2 | 870 | static void |
b8698a0f | 871 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 | 872 | enum vect_def_type *dt, |
4fc5ebf1 | 873 | int ndts, |
68435eb2 RB |
874 | slp_tree node, |
875 | stmt_vector_for_cost *cost_vec) | |
ebfd146a | 876 | { |
92345349 | 877 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a | 878 | |
68435eb2 | 879 | gcc_assert (cost_vec != NULL); |
ebfd146a | 880 | |
68435eb2 RB |
881 | /* ??? Somehow we need to fix this at the callers. */ |
882 | if (node) | |
883 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (node); | |
884 | ||
885 | if (node) | |
886 | { | |
887 | /* Scan operands and account for prologue cost of constants/externals. | |
888 | ??? This over-estimates cost for multiple uses and should be | |
889 | re-engineered. */ | |
b9787581 | 890 | gimple *stmt = SLP_TREE_SCALAR_STMTS (node)[0]->stmt; |
68435eb2 RB |
891 | tree lhs = gimple_get_lhs (stmt); |
892 | for (unsigned i = 0; i < gimple_num_ops (stmt); ++i) | |
893 | { | |
894 | tree op = gimple_op (stmt, i); | |
68435eb2 RB |
895 | enum vect_def_type dt; |
896 | if (!op || op == lhs) | |
897 | continue; | |
894dd753 | 898 | if (vect_is_simple_use (op, stmt_info->vinfo, &dt) |
68435eb2 RB |
899 | && (dt == vect_constant_def || dt == vect_external_def)) |
900 | prologue_cost += vect_prologue_cost_for_slp_op (node, stmt_info, | |
901 | i, dt, cost_vec); | |
902 | } | |
903 | } | |
904 | else | |
905 | /* Cost the "broadcast" of a scalar operand in to a vector operand. | |
906 | Use scalar_to_vec to cost the broadcast, as elsewhere in the vector | |
907 | cost model. */ | |
908 | for (int i = 0; i < ndts; i++) | |
909 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) | |
910 | prologue_cost += record_stmt_cost (cost_vec, 1, scalar_to_vec, | |
911 | stmt_info, 0, vect_prologue); | |
912 | ||
913 | /* Adjust for two-operator SLP nodes. */ | |
914 | if (node && SLP_TREE_TWO_OPERATORS (node)) | |
915 | { | |
916 | ncopies *= 2; | |
917 | inside_cost += record_stmt_cost (cost_vec, ncopies, vec_perm, | |
918 | stmt_info, 0, vect_body); | |
919 | } | |
c3e7ee41 BS |
920 | |
921 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
68435eb2 RB |
922 | inside_cost += record_stmt_cost (cost_vec, ncopies, vector_stmt, |
923 | stmt_info, 0, vect_body); | |
c3e7ee41 | 924 | |
73fbfcad | 925 | if (dump_enabled_p ()) |
78c60e3d SS |
926 | dump_printf_loc (MSG_NOTE, vect_location, |
927 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 928 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
929 | } |
930 | ||
931 | ||
8bd37302 BS |
932 | /* Model cost for type demotion and promotion operations. PWR is normally |
933 | zero for single-step promotions and demotions. It will be one if | |
934 | two-step promotion/demotion is required, and so on. Each additional | |
935 | step doubles the number of instructions required. */ | |
936 | ||
937 | static void | |
938 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
68435eb2 RB |
939 | enum vect_def_type *dt, int pwr, |
940 | stmt_vector_for_cost *cost_vec) | |
8bd37302 BS |
941 | { |
942 | int i, tmp; | |
92345349 | 943 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 | 944 | |
8bd37302 BS |
945 | for (i = 0; i < pwr + 1; i++) |
946 | { | |
947 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
948 | (i + 1) : i; | |
68435eb2 RB |
949 | inside_cost += record_stmt_cost (cost_vec, vect_pow2 (tmp), |
950 | vec_promote_demote, stmt_info, 0, | |
951 | vect_body); | |
8bd37302 BS |
952 | } |
953 | ||
954 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
955 | for (i = 0; i < 2; i++) | |
92345349 | 956 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
68435eb2 RB |
957 | prologue_cost += record_stmt_cost (cost_vec, 1, vector_stmt, |
958 | stmt_info, 0, vect_prologue); | |
8bd37302 | 959 | |
73fbfcad | 960 | if (dump_enabled_p ()) |
78c60e3d SS |
961 | dump_printf_loc (MSG_NOTE, vect_location, |
962 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 963 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
964 | } |
965 | ||
ebfd146a IR |
966 | /* Function vect_model_store_cost |
967 | ||
0d0293ac MM |
968 | Models cost for stores. In the case of grouped accesses, one access |
969 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a | 970 | |
68435eb2 | 971 | static void |
b8698a0f | 972 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
68435eb2 | 973 | enum vect_def_type dt, |
2de001ee | 974 | vect_memory_access_type memory_access_type, |
9ce4345a | 975 | vec_load_store_type vls_type, slp_tree slp_node, |
68435eb2 | 976 | stmt_vector_for_cost *cost_vec) |
ebfd146a | 977 | { |
92345349 | 978 | unsigned int inside_cost = 0, prologue_cost = 0; |
bffb8014 | 979 | stmt_vec_info first_stmt_info = stmt_info; |
892a981f | 980 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
ebfd146a | 981 | |
68435eb2 RB |
982 | /* ??? Somehow we need to fix this at the callers. */ |
983 | if (slp_node) | |
984 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
985 | ||
9ce4345a | 986 | if (vls_type == VLS_STORE_INVARIANT) |
68435eb2 RB |
987 | { |
988 | if (slp_node) | |
989 | prologue_cost += vect_prologue_cost_for_slp_op (slp_node, stmt_info, | |
990 | 1, dt, cost_vec); | |
991 | else | |
992 | prologue_cost += record_stmt_cost (cost_vec, 1, scalar_to_vec, | |
993 | stmt_info, 0, vect_prologue); | |
994 | } | |
ebfd146a | 995 | |
892a981f RS |
996 | /* Grouped stores update all elements in the group at once, |
997 | so we want the DR for the first statement. */ | |
998 | if (!slp_node && grouped_access_p) | |
bffb8014 | 999 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a | 1000 | |
892a981f RS |
1001 | /* True if we should include any once-per-group costs as well as |
1002 | the cost of the statement itself. For SLP we only get called | |
1003 | once per group anyhow. */ | |
bffb8014 | 1004 | bool first_stmt_p = (first_stmt_info == stmt_info); |
892a981f | 1005 | |
272c6793 | 1006 | /* We assume that the cost of a single store-lanes instruction is |
2c53b149 | 1007 | equivalent to the cost of DR_GROUP_SIZE separate stores. If a grouped |
272c6793 | 1008 | access is instead being provided by a permute-and-store operation, |
2de001ee RS |
1009 | include the cost of the permutes. */ |
1010 | if (first_stmt_p | |
1011 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 1012 | { |
e1377713 ES |
1013 | /* Uses a high and low interleave or shuffle operations for each |
1014 | needed permute. */ | |
bffb8014 | 1015 | int group_size = DR_GROUP_SIZE (first_stmt_info); |
e1377713 | 1016 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
68435eb2 | 1017 | inside_cost = record_stmt_cost (cost_vec, nstmts, vec_perm, |
92345349 | 1018 | stmt_info, 0, vect_body); |
ebfd146a | 1019 | |
73fbfcad | 1020 | if (dump_enabled_p ()) |
78c60e3d | 1021 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1022 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 1023 | group_size); |
ebfd146a IR |
1024 | } |
1025 | ||
cee62fee | 1026 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 1027 | /* Costs of the stores. */ |
067bc855 RB |
1028 | if (memory_access_type == VMAT_ELEMENTWISE |
1029 | || memory_access_type == VMAT_GATHER_SCATTER) | |
c5126ce8 RS |
1030 | { |
1031 | /* N scalar stores plus extracting the elements. */ | |
1032 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
68435eb2 | 1033 | inside_cost += record_stmt_cost (cost_vec, |
c5126ce8 RS |
1034 | ncopies * assumed_nunits, |
1035 | scalar_store, stmt_info, 0, vect_body); | |
1036 | } | |
f2e2a985 | 1037 | else |
57c454d2 | 1038 | vect_get_store_cost (stmt_info, ncopies, &inside_cost, cost_vec); |
ebfd146a | 1039 | |
2de001ee RS |
1040 | if (memory_access_type == VMAT_ELEMENTWISE |
1041 | || memory_access_type == VMAT_STRIDED_SLP) | |
c5126ce8 RS |
1042 | { |
1043 | /* N scalar stores plus extracting the elements. */ | |
1044 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
68435eb2 | 1045 | inside_cost += record_stmt_cost (cost_vec, |
c5126ce8 RS |
1046 | ncopies * assumed_nunits, |
1047 | vec_to_scalar, stmt_info, 0, vect_body); | |
1048 | } | |
cee62fee | 1049 | |
73fbfcad | 1050 | if (dump_enabled_p ()) |
78c60e3d SS |
1051 | dump_printf_loc (MSG_NOTE, vect_location, |
1052 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 1053 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
1054 | } |
1055 | ||
1056 | ||
720f5239 IR |
1057 | /* Calculate cost of DR's memory access. */ |
1058 | void | |
57c454d2 | 1059 | vect_get_store_cost (stmt_vec_info stmt_info, int ncopies, |
c3e7ee41 | 1060 | unsigned int *inside_cost, |
92345349 | 1061 | stmt_vector_for_cost *body_cost_vec) |
720f5239 | 1062 | { |
57c454d2 | 1063 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
720f5239 IR |
1064 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
1065 | ||
1066 | switch (alignment_support_scheme) | |
1067 | { | |
1068 | case dr_aligned: | |
1069 | { | |
92345349 BS |
1070 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
1071 | vector_store, stmt_info, 0, | |
1072 | vect_body); | |
720f5239 | 1073 | |
73fbfcad | 1074 | if (dump_enabled_p ()) |
78c60e3d | 1075 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1076 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
1077 | break; |
1078 | } | |
1079 | ||
1080 | case dr_unaligned_supported: | |
1081 | { | |
720f5239 | 1082 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1083 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1084 | unaligned_store, stmt_info, |
92345349 | 1085 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1086 | if (dump_enabled_p ()) |
78c60e3d SS |
1087 | dump_printf_loc (MSG_NOTE, vect_location, |
1088 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1089 | "hardware.\n"); |
720f5239 IR |
1090 | break; |
1091 | } | |
1092 | ||
38eec4c6 UW |
1093 | case dr_unaligned_unsupported: |
1094 | { | |
1095 | *inside_cost = VECT_MAX_COST; | |
1096 | ||
73fbfcad | 1097 | if (dump_enabled_p ()) |
78c60e3d | 1098 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1099 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1100 | break; |
1101 | } | |
1102 | ||
720f5239 IR |
1103 | default: |
1104 | gcc_unreachable (); | |
1105 | } | |
1106 | } | |
1107 | ||
1108 | ||
ebfd146a IR |
1109 | /* Function vect_model_load_cost |
1110 | ||
892a981f RS |
1111 | Models cost for loads. In the case of grouped accesses, one access has |
1112 | the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1113 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1114 | access scheme chosen. */ |
1115 | ||
68435eb2 RB |
1116 | static void |
1117 | vect_model_load_cost (stmt_vec_info stmt_info, unsigned ncopies, | |
2de001ee | 1118 | vect_memory_access_type memory_access_type, |
68435eb2 | 1119 | slp_instance instance, |
2de001ee | 1120 | slp_tree slp_node, |
68435eb2 | 1121 | stmt_vector_for_cost *cost_vec) |
ebfd146a | 1122 | { |
92345349 | 1123 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f | 1124 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
ebfd146a | 1125 | |
68435eb2 RB |
1126 | gcc_assert (cost_vec); |
1127 | ||
1128 | /* ??? Somehow we need to fix this at the callers. */ | |
1129 | if (slp_node) | |
1130 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
1131 | ||
1132 | if (slp_node && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
1133 | { | |
1134 | /* If the load is permuted then the alignment is determined by | |
1135 | the first group element not by the first scalar stmt DR. */ | |
bffb8014 | 1136 | stmt_vec_info first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
68435eb2 RB |
1137 | /* Record the cost for the permutation. */ |
1138 | unsigned n_perms; | |
1139 | unsigned assumed_nunits | |
bffb8014 | 1140 | = vect_nunits_for_cost (STMT_VINFO_VECTYPE (first_stmt_info)); |
68435eb2 RB |
1141 | unsigned slp_vf = (ncopies * assumed_nunits) / instance->group_size; |
1142 | vect_transform_slp_perm_load (slp_node, vNULL, NULL, | |
1143 | slp_vf, instance, true, | |
1144 | &n_perms); | |
1145 | inside_cost += record_stmt_cost (cost_vec, n_perms, vec_perm, | |
bffb8014 | 1146 | first_stmt_info, 0, vect_body); |
68435eb2 RB |
1147 | /* And adjust the number of loads performed. This handles |
1148 | redundancies as well as loads that are later dead. */ | |
bffb8014 | 1149 | auto_sbitmap perm (DR_GROUP_SIZE (first_stmt_info)); |
68435eb2 RB |
1150 | bitmap_clear (perm); |
1151 | for (unsigned i = 0; | |
1152 | i < SLP_TREE_LOAD_PERMUTATION (slp_node).length (); ++i) | |
1153 | bitmap_set_bit (perm, SLP_TREE_LOAD_PERMUTATION (slp_node)[i]); | |
1154 | ncopies = 0; | |
1155 | bool load_seen = false; | |
bffb8014 | 1156 | for (unsigned i = 0; i < DR_GROUP_SIZE (first_stmt_info); ++i) |
68435eb2 RB |
1157 | { |
1158 | if (i % assumed_nunits == 0) | |
1159 | { | |
1160 | if (load_seen) | |
1161 | ncopies++; | |
1162 | load_seen = false; | |
1163 | } | |
1164 | if (bitmap_bit_p (perm, i)) | |
1165 | load_seen = true; | |
1166 | } | |
1167 | if (load_seen) | |
1168 | ncopies++; | |
1169 | gcc_assert (ncopies | |
bffb8014 RS |
1170 | <= (DR_GROUP_SIZE (first_stmt_info) |
1171 | - DR_GROUP_GAP (first_stmt_info) | |
68435eb2 RB |
1172 | + assumed_nunits - 1) / assumed_nunits); |
1173 | } | |
1174 | ||
892a981f RS |
1175 | /* Grouped loads read all elements in the group at once, |
1176 | so we want the DR for the first statement. */ | |
bffb8014 | 1177 | stmt_vec_info first_stmt_info = stmt_info; |
892a981f | 1178 | if (!slp_node && grouped_access_p) |
bffb8014 | 1179 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a | 1180 | |
892a981f RS |
1181 | /* True if we should include any once-per-group costs as well as |
1182 | the cost of the statement itself. For SLP we only get called | |
1183 | once per group anyhow. */ | |
bffb8014 | 1184 | bool first_stmt_p = (first_stmt_info == stmt_info); |
892a981f | 1185 | |
272c6793 | 1186 | /* We assume that the cost of a single load-lanes instruction is |
2c53b149 | 1187 | equivalent to the cost of DR_GROUP_SIZE separate loads. If a grouped |
272c6793 | 1188 | access is instead being provided by a load-and-permute operation, |
2de001ee RS |
1189 | include the cost of the permutes. */ |
1190 | if (first_stmt_p | |
1191 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 1192 | { |
2c23db6d ES |
1193 | /* Uses an even and odd extract operations or shuffle operations |
1194 | for each needed permute. */ | |
bffb8014 | 1195 | int group_size = DR_GROUP_SIZE (first_stmt_info); |
2c23db6d | 1196 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
68435eb2 RB |
1197 | inside_cost += record_stmt_cost (cost_vec, nstmts, vec_perm, |
1198 | stmt_info, 0, vect_body); | |
ebfd146a | 1199 | |
73fbfcad | 1200 | if (dump_enabled_p ()) |
e645e942 TJ |
1201 | dump_printf_loc (MSG_NOTE, vect_location, |
1202 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1203 | group_size); |
ebfd146a IR |
1204 | } |
1205 | ||
1206 | /* The loads themselves. */ | |
067bc855 RB |
1207 | if (memory_access_type == VMAT_ELEMENTWISE |
1208 | || memory_access_type == VMAT_GATHER_SCATTER) | |
a82960aa | 1209 | { |
a21892ad BS |
1210 | /* N scalar loads plus gathering them into a vector. */ |
1211 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
c5126ce8 | 1212 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); |
68435eb2 | 1213 | inside_cost += record_stmt_cost (cost_vec, |
c5126ce8 | 1214 | ncopies * assumed_nunits, |
92345349 | 1215 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1216 | } |
1217 | else | |
57c454d2 | 1218 | vect_get_load_cost (stmt_info, ncopies, first_stmt_p, |
92345349 | 1219 | &inside_cost, &prologue_cost, |
68435eb2 | 1220 | cost_vec, cost_vec, true); |
2de001ee RS |
1221 | if (memory_access_type == VMAT_ELEMENTWISE |
1222 | || memory_access_type == VMAT_STRIDED_SLP) | |
68435eb2 | 1223 | inside_cost += record_stmt_cost (cost_vec, ncopies, vec_construct, |
892a981f | 1224 | stmt_info, 0, vect_body); |
720f5239 | 1225 | |
73fbfcad | 1226 | if (dump_enabled_p ()) |
78c60e3d SS |
1227 | dump_printf_loc (MSG_NOTE, vect_location, |
1228 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1229 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1230 | } |
1231 | ||
1232 | ||
1233 | /* Calculate cost of DR's memory access. */ | |
1234 | void | |
57c454d2 | 1235 | vect_get_load_cost (stmt_vec_info stmt_info, int ncopies, |
c3e7ee41 | 1236 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1237 | unsigned int *prologue_cost, |
1238 | stmt_vector_for_cost *prologue_cost_vec, | |
1239 | stmt_vector_for_cost *body_cost_vec, | |
1240 | bool record_prologue_costs) | |
720f5239 | 1241 | { |
57c454d2 | 1242 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
720f5239 IR |
1243 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
1244 | ||
1245 | switch (alignment_support_scheme) | |
ebfd146a IR |
1246 | { |
1247 | case dr_aligned: | |
1248 | { | |
92345349 BS |
1249 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1250 | stmt_info, 0, vect_body); | |
ebfd146a | 1251 | |
73fbfcad | 1252 | if (dump_enabled_p ()) |
78c60e3d | 1253 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1254 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1255 | |
1256 | break; | |
1257 | } | |
1258 | case dr_unaligned_supported: | |
1259 | { | |
720f5239 | 1260 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1261 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1262 | unaligned_load, stmt_info, |
92345349 | 1263 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1264 | |
73fbfcad | 1265 | if (dump_enabled_p ()) |
78c60e3d SS |
1266 | dump_printf_loc (MSG_NOTE, vect_location, |
1267 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1268 | "hardware.\n"); |
ebfd146a IR |
1269 | |
1270 | break; | |
1271 | } | |
1272 | case dr_explicit_realign: | |
1273 | { | |
92345349 BS |
1274 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1275 | vector_load, stmt_info, 0, vect_body); | |
1276 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1277 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1278 | |
1279 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1280 | the containing loop, the following cost should be added to the | |
92345349 | 1281 | prologue costs. */ |
ebfd146a | 1282 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1283 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1284 | stmt_info, 0, vect_body); | |
ebfd146a | 1285 | |
73fbfcad | 1286 | if (dump_enabled_p ()) |
e645e942 TJ |
1287 | dump_printf_loc (MSG_NOTE, vect_location, |
1288 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1289 | |
ebfd146a IR |
1290 | break; |
1291 | } | |
1292 | case dr_explicit_realign_optimized: | |
1293 | { | |
73fbfcad | 1294 | if (dump_enabled_p ()) |
e645e942 | 1295 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1296 | "vect_model_load_cost: unaligned software " |
e645e942 | 1297 | "pipelined.\n"); |
ebfd146a IR |
1298 | |
1299 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1300 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1301 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1302 | access, then the above cost should only be considered for one |
ff802fa1 | 1303 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1304 | and a realignment op. */ |
1305 | ||
92345349 | 1306 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1307 | { |
92345349 BS |
1308 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1309 | vector_stmt, stmt_info, | |
1310 | 0, vect_prologue); | |
ebfd146a | 1311 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1312 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1313 | vector_stmt, stmt_info, | |
1314 | 0, vect_prologue); | |
ebfd146a IR |
1315 | } |
1316 | ||
92345349 BS |
1317 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1318 | stmt_info, 0, vect_body); | |
1319 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1320 | stmt_info, 0, vect_body); | |
8bd37302 | 1321 | |
73fbfcad | 1322 | if (dump_enabled_p ()) |
78c60e3d | 1323 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1324 | "vect_model_load_cost: explicit realign optimized" |
1325 | "\n"); | |
8bd37302 | 1326 | |
ebfd146a IR |
1327 | break; |
1328 | } | |
1329 | ||
38eec4c6 UW |
1330 | case dr_unaligned_unsupported: |
1331 | { | |
1332 | *inside_cost = VECT_MAX_COST; | |
1333 | ||
73fbfcad | 1334 | if (dump_enabled_p ()) |
78c60e3d | 1335 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1336 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1337 | break; |
1338 | } | |
1339 | ||
ebfd146a IR |
1340 | default: |
1341 | gcc_unreachable (); | |
1342 | } | |
ebfd146a IR |
1343 | } |
1344 | ||
418b7df3 RG |
1345 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1346 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1347 | |
418b7df3 | 1348 | static void |
355fe088 | 1349 | vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi) |
ebfd146a | 1350 | { |
ebfd146a | 1351 | if (gsi) |
418b7df3 | 1352 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1353 | else |
1354 | { | |
418b7df3 | 1355 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1356 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1357 | |
a70d6342 IR |
1358 | if (loop_vinfo) |
1359 | { | |
1360 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1361 | basic_block new_bb; |
1362 | edge pe; | |
a70d6342 IR |
1363 | |
1364 | if (nested_in_vect_loop_p (loop, stmt)) | |
1365 | loop = loop->inner; | |
b8698a0f | 1366 | |
a70d6342 | 1367 | pe = loop_preheader_edge (loop); |
418b7df3 | 1368 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1369 | gcc_assert (!new_bb); |
1370 | } | |
1371 | else | |
1372 | { | |
1373 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1374 | basic_block bb; | |
1375 | gimple_stmt_iterator gsi_bb_start; | |
1376 | ||
1377 | gcc_assert (bb_vinfo); | |
1378 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1379 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1380 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1381 | } |
ebfd146a IR |
1382 | } |
1383 | ||
73fbfcad | 1384 | if (dump_enabled_p ()) |
ebfd146a | 1385 | { |
78c60e3d SS |
1386 | dump_printf_loc (MSG_NOTE, vect_location, |
1387 | "created new init_stmt: "); | |
1388 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1389 | } |
418b7df3 RG |
1390 | } |
1391 | ||
1392 | /* Function vect_init_vector. | |
ebfd146a | 1393 | |
5467ee52 RG |
1394 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1395 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1396 | vector type a vector with all elements equal to VAL is created first. | |
1397 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1398 | initialization at the loop preheader. | |
418b7df3 RG |
1399 | Return the DEF of INIT_STMT. |
1400 | It will be used in the vectorization of STMT. */ | |
1401 | ||
1402 | tree | |
355fe088 | 1403 | vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 | 1404 | { |
355fe088 | 1405 | gimple *init_stmt; |
418b7df3 RG |
1406 | tree new_temp; |
1407 | ||
e412ece4 RB |
1408 | /* We abuse this function to push sth to a SSA name with initial 'val'. */ |
1409 | if (! useless_type_conversion_p (type, TREE_TYPE (val))) | |
418b7df3 | 1410 | { |
e412ece4 RB |
1411 | gcc_assert (TREE_CODE (type) == VECTOR_TYPE); |
1412 | if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) | |
418b7df3 | 1413 | { |
5a308cf1 IE |
1414 | /* Scalar boolean value should be transformed into |
1415 | all zeros or all ones value before building a vector. */ | |
1416 | if (VECTOR_BOOLEAN_TYPE_P (type)) | |
1417 | { | |
b3d51f23 IE |
1418 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1419 | tree false_val = build_zero_cst (TREE_TYPE (type)); | |
5a308cf1 IE |
1420 | |
1421 | if (CONSTANT_CLASS_P (val)) | |
1422 | val = integer_zerop (val) ? false_val : true_val; | |
1423 | else | |
1424 | { | |
1425 | new_temp = make_ssa_name (TREE_TYPE (type)); | |
1426 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, | |
1427 | val, true_val, false_val); | |
1428 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
1429 | val = new_temp; | |
1430 | } | |
1431 | } | |
1432 | else if (CONSTANT_CLASS_P (val)) | |
42fd8198 | 1433 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1434 | else |
1435 | { | |
b731b390 | 1436 | new_temp = make_ssa_name (TREE_TYPE (type)); |
e412ece4 RB |
1437 | if (! INTEGRAL_TYPE_P (TREE_TYPE (val))) |
1438 | init_stmt = gimple_build_assign (new_temp, | |
1439 | fold_build1 (VIEW_CONVERT_EXPR, | |
1440 | TREE_TYPE (type), | |
1441 | val)); | |
1442 | else | |
1443 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); | |
418b7df3 | 1444 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1445 | val = new_temp; |
418b7df3 RG |
1446 | } |
1447 | } | |
5467ee52 | 1448 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1449 | } |
1450 | ||
0e22bb5a RB |
1451 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1452 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1453 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1454 | return new_temp; |
ebfd146a IR |
1455 | } |
1456 | ||
c83a894c | 1457 | /* Function vect_get_vec_def_for_operand_1. |
a70d6342 | 1458 | |
c83a894c AH |
1459 | For a defining stmt DEF_STMT of a scalar stmt, return a vector def with type |
1460 | DT that will be used in the vectorized stmt. */ | |
ebfd146a IR |
1461 | |
1462 | tree | |
c83a894c | 1463 | vect_get_vec_def_for_operand_1 (gimple *def_stmt, enum vect_def_type dt) |
ebfd146a IR |
1464 | { |
1465 | tree vec_oprnd; | |
1eede195 | 1466 | stmt_vec_info vec_stmt_info; |
ebfd146a | 1467 | stmt_vec_info def_stmt_info = NULL; |
ebfd146a IR |
1468 | |
1469 | switch (dt) | |
1470 | { | |
81c40241 | 1471 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1472 | case vect_constant_def: |
81c40241 | 1473 | case vect_external_def: |
c83a894c AH |
1474 | /* Code should use vect_get_vec_def_for_operand. */ |
1475 | gcc_unreachable (); | |
ebfd146a | 1476 | |
81c40241 | 1477 | /* operand is defined inside the loop. */ |
8644a673 | 1478 | case vect_internal_def: |
ebfd146a | 1479 | { |
ebfd146a IR |
1480 | /* Get the def from the vectorized stmt. */ |
1481 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1482 | |
1eede195 RS |
1483 | vec_stmt_info = STMT_VINFO_VEC_STMT (def_stmt_info); |
1484 | /* Get vectorized pattern statement. */ | |
1485 | if (!vec_stmt_info | |
1486 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1487 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1488 | vec_stmt_info = (STMT_VINFO_VEC_STMT | |
1489 | (STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
1490 | gcc_assert (vec_stmt_info); | |
1491 | if (gphi *phi = dyn_cast <gphi *> (vec_stmt_info->stmt)) | |
1492 | vec_oprnd = PHI_RESULT (phi); | |
ebfd146a | 1493 | else |
1eede195 RS |
1494 | vec_oprnd = gimple_get_lhs (vec_stmt_info->stmt); |
1495 | return vec_oprnd; | |
ebfd146a IR |
1496 | } |
1497 | ||
c78e3652 | 1498 | /* operand is defined by a loop header phi. */ |
ebfd146a | 1499 | case vect_reduction_def: |
06066f92 | 1500 | case vect_double_reduction_def: |
7c5222ff | 1501 | case vect_nested_cycle: |
ebfd146a IR |
1502 | case vect_induction_def: |
1503 | { | |
1504 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1505 | ||
1eede195 RS |
1506 | /* Get the def from the vectorized stmt. */ |
1507 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1508 | vec_stmt_info = STMT_VINFO_VEC_STMT (def_stmt_info); | |
1509 | if (gphi *phi = dyn_cast <gphi *> (vec_stmt_info->stmt)) | |
1510 | vec_oprnd = PHI_RESULT (phi); | |
6dbbece6 | 1511 | else |
1eede195 RS |
1512 | vec_oprnd = gimple_get_lhs (vec_stmt_info->stmt); |
1513 | return vec_oprnd; | |
ebfd146a IR |
1514 | } |
1515 | ||
1516 | default: | |
1517 | gcc_unreachable (); | |
1518 | } | |
1519 | } | |
1520 | ||
1521 | ||
c83a894c AH |
1522 | /* Function vect_get_vec_def_for_operand. |
1523 | ||
1524 | OP is an operand in STMT. This function returns a (vector) def that will be | |
1525 | used in the vectorized stmt for STMT. | |
1526 | ||
1527 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1528 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1529 | ||
1530 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1531 | needs to be introduced. VECTYPE may be used to specify a required type for | |
1532 | vector invariant. */ | |
1533 | ||
1534 | tree | |
1535 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) | |
1536 | { | |
1537 | gimple *def_stmt; | |
1538 | enum vect_def_type dt; | |
1539 | bool is_simple_use; | |
1540 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
1541 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
1542 | ||
1543 | if (dump_enabled_p ()) | |
1544 | { | |
1545 | dump_printf_loc (MSG_NOTE, vect_location, | |
1546 | "vect_get_vec_def_for_operand: "); | |
1547 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
1548 | dump_printf (MSG_NOTE, "\n"); | |
1549 | } | |
1550 | ||
fef96d8e RS |
1551 | stmt_vec_info def_stmt_info; |
1552 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &dt, | |
1553 | &def_stmt_info, &def_stmt); | |
c83a894c AH |
1554 | gcc_assert (is_simple_use); |
1555 | if (def_stmt && dump_enabled_p ()) | |
1556 | { | |
1557 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1558 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
1559 | } | |
1560 | ||
1561 | if (dt == vect_constant_def || dt == vect_external_def) | |
1562 | { | |
1563 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
1564 | tree vector_type; | |
1565 | ||
1566 | if (vectype) | |
1567 | vector_type = vectype; | |
2568d8a1 | 1568 | else if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op)) |
c83a894c AH |
1569 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) |
1570 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1571 | else | |
1572 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1573 | ||
1574 | gcc_assert (vector_type); | |
1575 | return vect_init_vector (stmt, op, vector_type, NULL); | |
1576 | } | |
1577 | else | |
fef96d8e | 1578 | return vect_get_vec_def_for_operand_1 (def_stmt_info, dt); |
c83a894c AH |
1579 | } |
1580 | ||
1581 | ||
ebfd146a IR |
1582 | /* Function vect_get_vec_def_for_stmt_copy |
1583 | ||
ff802fa1 | 1584 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1585 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1586 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1587 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1588 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1589 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1590 | DT is the type of the vector def VEC_OPRND. |
1591 | ||
1592 | Context: | |
1593 | In case the vectorization factor (VF) is bigger than the number | |
1594 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1595 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1596 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1597 | smallest data-type determines the VF, and as a result, when vectorizing |
1598 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1599 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1600 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1601 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1602 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1603 | ||
1604 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1605 | |
ebfd146a IR |
1606 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1607 | VS1.1: vx.1 = memref1 VS1.2 | |
1608 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1609 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1610 | |
1611 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1612 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1613 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1614 | VSnew.3: vz3 = vx.3 + ... | |
1615 | ||
1616 | The vectorization of S1 is explained in vectorizable_load. | |
1617 | The vectorization of S2: | |
b8698a0f L |
1618 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1619 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1620 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1621 | returns the vector-def 'vx.0'. |
1622 | ||
b8698a0f L |
1623 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1624 | function is called to get the relevant vector-def for each operand. It is | |
1625 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1626 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1627 | ||
b8698a0f L |
1628 | For example, to obtain the vector-def 'vx.1' in order to create the |
1629 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1630 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1631 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1632 | and return its def ('vx.1'). | |
1633 | Overall, to create the above sequence this function will be called 3 times: | |
1634 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1635 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1636 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1637 | ||
1638 | tree | |
1639 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1640 | { | |
355fe088 | 1641 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1642 | stmt_vec_info def_stmt_info; |
1643 | ||
1644 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1645 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1646 | return vec_oprnd; |
1647 | ||
1648 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1649 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1650 | gcc_assert (def_stmt_info); | |
1651 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1652 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1653 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1654 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1655 | else | |
1656 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1657 | return vec_oprnd; | |
1658 | } | |
1659 | ||
1660 | ||
1661 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1662 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a | 1663 | |
c78e3652 | 1664 | void |
b8698a0f | 1665 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1666 | vec<tree> *vec_oprnds0, |
1667 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1668 | { |
9771b263 | 1669 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1670 | |
1671 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1672 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1673 | |
9771b263 | 1674 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1675 | { |
9771b263 | 1676 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1677 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1678 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1679 | } |
1680 | } | |
1681 | ||
1682 | ||
c78e3652 | 1683 | /* Get vectorized definitions for OP0 and OP1. */ |
ebfd146a | 1684 | |
c78e3652 | 1685 | void |
355fe088 | 1686 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1687 | vec<tree> *vec_oprnds0, |
1688 | vec<tree> *vec_oprnds1, | |
306b0c92 | 1689 | slp_tree slp_node) |
ebfd146a IR |
1690 | { |
1691 | if (slp_node) | |
d092494c IR |
1692 | { |
1693 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1694 | auto_vec<tree> ops (nops); |
1695 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1696 | |
9771b263 | 1697 | ops.quick_push (op0); |
d092494c | 1698 | if (op1) |
9771b263 | 1699 | ops.quick_push (op1); |
d092494c | 1700 | |
306b0c92 | 1701 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
d092494c | 1702 | |
37b5ec8f | 1703 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1704 | if (op1) |
37b5ec8f | 1705 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1706 | } |
ebfd146a IR |
1707 | else |
1708 | { | |
1709 | tree vec_oprnd; | |
1710 | ||
9771b263 | 1711 | vec_oprnds0->create (1); |
81c40241 | 1712 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1713 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1714 | |
1715 | if (op1) | |
1716 | { | |
9771b263 | 1717 | vec_oprnds1->create (1); |
81c40241 | 1718 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1719 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1720 | } |
1721 | } | |
1722 | } | |
1723 | ||
bb6c2b68 RS |
1724 | /* Helper function called by vect_finish_replace_stmt and |
1725 | vect_finish_stmt_generation. Set the location of the new | |
e1bd7296 | 1726 | statement and create and return a stmt_vec_info for it. */ |
bb6c2b68 | 1727 | |
e1bd7296 | 1728 | static stmt_vec_info |
bb6c2b68 RS |
1729 | vect_finish_stmt_generation_1 (gimple *stmt, gimple *vec_stmt) |
1730 | { | |
1731 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1732 | vec_info *vinfo = stmt_info->vinfo; | |
1733 | ||
e1bd7296 | 1734 | stmt_vec_info vec_stmt_info = vinfo->add_stmt (vec_stmt); |
bb6c2b68 RS |
1735 | |
1736 | if (dump_enabled_p ()) | |
1737 | { | |
1738 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); | |
1739 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
1740 | } | |
1741 | ||
1742 | gimple_set_location (vec_stmt, gimple_location (stmt)); | |
1743 | ||
1744 | /* While EH edges will generally prevent vectorization, stmt might | |
1745 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1746 | that could throw are part of the same region. */ | |
1747 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1748 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1749 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
e1bd7296 RS |
1750 | |
1751 | return vec_stmt_info; | |
bb6c2b68 RS |
1752 | } |
1753 | ||
1754 | /* Replace the scalar statement STMT with a new vector statement VEC_STMT, | |
e1bd7296 RS |
1755 | which sets the same scalar result as STMT did. Create and return a |
1756 | stmt_vec_info for VEC_STMT. */ | |
bb6c2b68 | 1757 | |
e1bd7296 | 1758 | stmt_vec_info |
bb6c2b68 RS |
1759 | vect_finish_replace_stmt (gimple *stmt, gimple *vec_stmt) |
1760 | { | |
1761 | gcc_assert (gimple_get_lhs (stmt) == gimple_get_lhs (vec_stmt)); | |
1762 | ||
1763 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); | |
1764 | gsi_replace (&gsi, vec_stmt, false); | |
1765 | ||
e1bd7296 | 1766 | return vect_finish_stmt_generation_1 (stmt, vec_stmt); |
bb6c2b68 | 1767 | } |
ebfd146a | 1768 | |
e1bd7296 RS |
1769 | /* Add VEC_STMT to the vectorized implementation of STMT and insert it |
1770 | before *GSI. Create and return a stmt_vec_info for VEC_STMT. */ | |
ebfd146a | 1771 | |
e1bd7296 | 1772 | stmt_vec_info |
355fe088 | 1773 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1774 | gimple_stmt_iterator *gsi) |
1775 | { | |
ebfd146a IR |
1776 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); |
1777 | ||
54e8e2c3 RG |
1778 | if (!gsi_end_p (*gsi) |
1779 | && gimple_has_mem_ops (vec_stmt)) | |
1780 | { | |
355fe088 | 1781 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1782 | tree vuse = gimple_vuse (at_stmt); |
1783 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1784 | { | |
1785 | tree vdef = gimple_vdef (at_stmt); | |
1786 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1787 | /* If we have an SSA vuse and insert a store, update virtual | |
1788 | SSA form to avoid triggering the renamer. Do so only | |
1789 | if we can easily see all uses - which is what almost always | |
1790 | happens with the way vectorized stmts are inserted. */ | |
1791 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1792 | && ((is_gimple_assign (vec_stmt) | |
1793 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1794 | || (is_gimple_call (vec_stmt) | |
1795 | && !(gimple_call_flags (vec_stmt) | |
1796 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1797 | { | |
1798 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1799 | gimple_set_vdef (vec_stmt, new_vdef); | |
1800 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1801 | } | |
1802 | } | |
1803 | } | |
ebfd146a | 1804 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
e1bd7296 | 1805 | return vect_finish_stmt_generation_1 (stmt, vec_stmt); |
ebfd146a IR |
1806 | } |
1807 | ||
70439f0d RS |
1808 | /* We want to vectorize a call to combined function CFN with function |
1809 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN | |
1810 | as the types of all inputs. Check whether this is possible using | |
1811 | an internal function, returning its code if so or IFN_LAST if not. */ | |
ebfd146a | 1812 | |
70439f0d RS |
1813 | static internal_fn |
1814 | vectorizable_internal_function (combined_fn cfn, tree fndecl, | |
1815 | tree vectype_out, tree vectype_in) | |
ebfd146a | 1816 | { |
70439f0d RS |
1817 | internal_fn ifn; |
1818 | if (internal_fn_p (cfn)) | |
1819 | ifn = as_internal_fn (cfn); | |
1820 | else | |
1821 | ifn = associated_internal_fn (fndecl); | |
1822 | if (ifn != IFN_LAST && direct_internal_fn_p (ifn)) | |
1823 | { | |
1824 | const direct_internal_fn_info &info = direct_internal_fn (ifn); | |
1825 | if (info.vectorizable) | |
1826 | { | |
1827 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); | |
1828 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); | |
d95ab70a RS |
1829 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1830 | OPTIMIZE_FOR_SPEED)) | |
70439f0d RS |
1831 | return ifn; |
1832 | } | |
1833 | } | |
1834 | return IFN_LAST; | |
ebfd146a IR |
1835 | } |
1836 | ||
5ce9450f | 1837 | |
355fe088 | 1838 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1839 | gimple_stmt_iterator *); |
1840 | ||
7cfb4d93 RS |
1841 | /* Check whether a load or store statement in the loop described by |
1842 | LOOP_VINFO is possible in a fully-masked loop. This is testing | |
1843 | whether the vectorizer pass has the appropriate support, as well as | |
1844 | whether the target does. | |
1845 | ||
1846 | VLS_TYPE says whether the statement is a load or store and VECTYPE | |
1847 | is the type of the vector being loaded or stored. MEMORY_ACCESS_TYPE | |
1848 | says how the load or store is going to be implemented and GROUP_SIZE | |
1849 | is the number of load or store statements in the containing group. | |
bfaa08b7 RS |
1850 | If the access is a gather load or scatter store, GS_INFO describes |
1851 | its arguments. | |
7cfb4d93 RS |
1852 | |
1853 | Clear LOOP_VINFO_CAN_FULLY_MASK_P if a fully-masked loop is not | |
1854 | supported, otherwise record the required mask types. */ | |
1855 | ||
1856 | static void | |
1857 | check_load_store_masking (loop_vec_info loop_vinfo, tree vectype, | |
1858 | vec_load_store_type vls_type, int group_size, | |
bfaa08b7 RS |
1859 | vect_memory_access_type memory_access_type, |
1860 | gather_scatter_info *gs_info) | |
7cfb4d93 RS |
1861 | { |
1862 | /* Invariant loads need no special support. */ | |
1863 | if (memory_access_type == VMAT_INVARIANT) | |
1864 | return; | |
1865 | ||
1866 | vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); | |
1867 | machine_mode vecmode = TYPE_MODE (vectype); | |
1868 | bool is_load = (vls_type == VLS_LOAD); | |
1869 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
1870 | { | |
1871 | if (is_load | |
1872 | ? !vect_load_lanes_supported (vectype, group_size, true) | |
1873 | : !vect_store_lanes_supported (vectype, group_size, true)) | |
1874 | { | |
1875 | if (dump_enabled_p ()) | |
1876 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1877 | "can't use a fully-masked loop because the" | |
1878 | " target doesn't have an appropriate masked" | |
1879 | " load/store-lanes instruction.\n"); | |
1880 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1881 | return; | |
1882 | } | |
1883 | unsigned int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
1884 | vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype); | |
1885 | return; | |
1886 | } | |
1887 | ||
bfaa08b7 RS |
1888 | if (memory_access_type == VMAT_GATHER_SCATTER) |
1889 | { | |
f307441a RS |
1890 | internal_fn ifn = (is_load |
1891 | ? IFN_MASK_GATHER_LOAD | |
1892 | : IFN_MASK_SCATTER_STORE); | |
bfaa08b7 | 1893 | tree offset_type = TREE_TYPE (gs_info->offset); |
f307441a | 1894 | if (!internal_gather_scatter_fn_supported_p (ifn, vectype, |
bfaa08b7 RS |
1895 | gs_info->memory_type, |
1896 | TYPE_SIGN (offset_type), | |
1897 | gs_info->scale)) | |
1898 | { | |
1899 | if (dump_enabled_p ()) | |
1900 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1901 | "can't use a fully-masked loop because the" | |
1902 | " target doesn't have an appropriate masked" | |
f307441a | 1903 | " gather load or scatter store instruction.\n"); |
bfaa08b7 RS |
1904 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; |
1905 | return; | |
1906 | } | |
1907 | unsigned int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
1908 | vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype); | |
1909 | return; | |
1910 | } | |
1911 | ||
7cfb4d93 RS |
1912 | if (memory_access_type != VMAT_CONTIGUOUS |
1913 | && memory_access_type != VMAT_CONTIGUOUS_PERMUTE) | |
1914 | { | |
1915 | /* Element X of the data must come from iteration i * VF + X of the | |
1916 | scalar loop. We need more work to support other mappings. */ | |
1917 | if (dump_enabled_p ()) | |
1918 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1919 | "can't use a fully-masked loop because an access" | |
1920 | " isn't contiguous.\n"); | |
1921 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1922 | return; | |
1923 | } | |
1924 | ||
1925 | machine_mode mask_mode; | |
1926 | if (!(targetm.vectorize.get_mask_mode | |
1927 | (GET_MODE_NUNITS (vecmode), | |
1928 | GET_MODE_SIZE (vecmode)).exists (&mask_mode)) | |
1929 | || !can_vec_mask_load_store_p (vecmode, mask_mode, is_load)) | |
1930 | { | |
1931 | if (dump_enabled_p ()) | |
1932 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1933 | "can't use a fully-masked loop because the target" | |
1934 | " doesn't have the appropriate masked load or" | |
1935 | " store.\n"); | |
1936 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1937 | return; | |
1938 | } | |
1939 | /* We might load more scalars than we need for permuting SLP loads. | |
1940 | We checked in get_group_load_store_type that the extra elements | |
1941 | don't leak into a new vector. */ | |
1942 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
1943 | poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
1944 | unsigned int nvectors; | |
1945 | if (can_div_away_from_zero_p (group_size * vf, nunits, &nvectors)) | |
1946 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype); | |
1947 | else | |
1948 | gcc_unreachable (); | |
1949 | } | |
1950 | ||
1951 | /* Return the mask input to a masked load or store. VEC_MASK is the vectorized | |
1952 | form of the scalar mask condition and LOOP_MASK, if nonnull, is the mask | |
1953 | that needs to be applied to all loads and stores in a vectorized loop. | |
1954 | Return VEC_MASK if LOOP_MASK is null, otherwise return VEC_MASK & LOOP_MASK. | |
1955 | ||
1956 | MASK_TYPE is the type of both masks. If new statements are needed, | |
1957 | insert them before GSI. */ | |
1958 | ||
1959 | static tree | |
1960 | prepare_load_store_mask (tree mask_type, tree loop_mask, tree vec_mask, | |
1961 | gimple_stmt_iterator *gsi) | |
1962 | { | |
1963 | gcc_assert (useless_type_conversion_p (mask_type, TREE_TYPE (vec_mask))); | |
1964 | if (!loop_mask) | |
1965 | return vec_mask; | |
1966 | ||
1967 | gcc_assert (TREE_TYPE (loop_mask) == mask_type); | |
1968 | tree and_res = make_temp_ssa_name (mask_type, NULL, "vec_mask_and"); | |
1969 | gimple *and_stmt = gimple_build_assign (and_res, BIT_AND_EXPR, | |
1970 | vec_mask, loop_mask); | |
1971 | gsi_insert_before (gsi, and_stmt, GSI_SAME_STMT); | |
1972 | return and_res; | |
1973 | } | |
1974 | ||
429ef523 RS |
1975 | /* Determine whether we can use a gather load or scatter store to vectorize |
1976 | strided load or store STMT by truncating the current offset to a smaller | |
1977 | width. We need to be able to construct an offset vector: | |
1978 | ||
1979 | { 0, X, X*2, X*3, ... } | |
1980 | ||
1981 | without loss of precision, where X is STMT's DR_STEP. | |
1982 | ||
1983 | Return true if this is possible, describing the gather load or scatter | |
1984 | store in GS_INFO. MASKED_P is true if the load or store is conditional. */ | |
1985 | ||
1986 | static bool | |
1987 | vect_truncate_gather_scatter_offset (gimple *stmt, loop_vec_info loop_vinfo, | |
1988 | bool masked_p, | |
1989 | gather_scatter_info *gs_info) | |
1990 | { | |
1991 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1992 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
1993 | tree step = DR_STEP (dr); | |
1994 | if (TREE_CODE (step) != INTEGER_CST) | |
1995 | { | |
1996 | /* ??? Perhaps we could use range information here? */ | |
1997 | if (dump_enabled_p ()) | |
1998 | dump_printf_loc (MSG_NOTE, vect_location, | |
1999 | "cannot truncate variable step.\n"); | |
2000 | return false; | |
2001 | } | |
2002 | ||
2003 | /* Get the number of bits in an element. */ | |
2004 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2005 | scalar_mode element_mode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); | |
2006 | unsigned int element_bits = GET_MODE_BITSIZE (element_mode); | |
2007 | ||
2008 | /* Set COUNT to the upper limit on the number of elements - 1. | |
2009 | Start with the maximum vectorization factor. */ | |
2010 | unsigned HOST_WIDE_INT count = vect_max_vf (loop_vinfo) - 1; | |
2011 | ||
2012 | /* Try lowering COUNT to the number of scalar latch iterations. */ | |
2013 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2014 | widest_int max_iters; | |
2015 | if (max_loop_iterations (loop, &max_iters) | |
2016 | && max_iters < count) | |
2017 | count = max_iters.to_shwi (); | |
2018 | ||
2019 | /* Try scales of 1 and the element size. */ | |
2020 | int scales[] = { 1, vect_get_scalar_dr_size (dr) }; | |
4a669ac3 | 2021 | wi::overflow_type overflow = wi::OVF_NONE; |
429ef523 RS |
2022 | for (int i = 0; i < 2; ++i) |
2023 | { | |
2024 | int scale = scales[i]; | |
2025 | widest_int factor; | |
2026 | if (!wi::multiple_of_p (wi::to_widest (step), scale, SIGNED, &factor)) | |
2027 | continue; | |
2028 | ||
2029 | /* See whether we can calculate (COUNT - 1) * STEP / SCALE | |
2030 | in OFFSET_BITS bits. */ | |
4a669ac3 AH |
2031 | widest_int range = wi::mul (count, factor, SIGNED, &overflow); |
2032 | if (overflow) | |
429ef523 RS |
2033 | continue; |
2034 | signop sign = range >= 0 ? UNSIGNED : SIGNED; | |
2035 | if (wi::min_precision (range, sign) > element_bits) | |
2036 | { | |
4a669ac3 | 2037 | overflow = wi::OVF_UNKNOWN; |
429ef523 RS |
2038 | continue; |
2039 | } | |
2040 | ||
2041 | /* See whether the target supports the operation. */ | |
2042 | tree memory_type = TREE_TYPE (DR_REF (dr)); | |
2043 | if (!vect_gather_scatter_fn_p (DR_IS_READ (dr), masked_p, vectype, | |
2044 | memory_type, element_bits, sign, scale, | |
2045 | &gs_info->ifn, &gs_info->element_type)) | |
2046 | continue; | |
2047 | ||
2048 | tree offset_type = build_nonstandard_integer_type (element_bits, | |
2049 | sign == UNSIGNED); | |
2050 | ||
2051 | gs_info->decl = NULL_TREE; | |
2052 | /* Logically the sum of DR_BASE_ADDRESS, DR_INIT and DR_OFFSET, | |
2053 | but we don't need to store that here. */ | |
2054 | gs_info->base = NULL_TREE; | |
2055 | gs_info->offset = fold_convert (offset_type, step); | |
929b4411 | 2056 | gs_info->offset_dt = vect_constant_def; |
429ef523 RS |
2057 | gs_info->offset_vectype = NULL_TREE; |
2058 | gs_info->scale = scale; | |
2059 | gs_info->memory_type = memory_type; | |
2060 | return true; | |
2061 | } | |
2062 | ||
4a669ac3 | 2063 | if (overflow && dump_enabled_p ()) |
429ef523 RS |
2064 | dump_printf_loc (MSG_NOTE, vect_location, |
2065 | "truncating gather/scatter offset to %d bits" | |
2066 | " might change its value.\n", element_bits); | |
2067 | ||
2068 | return false; | |
2069 | } | |
2070 | ||
ab2fc782 RS |
2071 | /* Return true if we can use gather/scatter internal functions to |
2072 | vectorize STMT, which is a grouped or strided load or store. | |
429ef523 RS |
2073 | MASKED_P is true if load or store is conditional. When returning |
2074 | true, fill in GS_INFO with the information required to perform the | |
2075 | operation. */ | |
ab2fc782 RS |
2076 | |
2077 | static bool | |
2078 | vect_use_strided_gather_scatters_p (gimple *stmt, loop_vec_info loop_vinfo, | |
429ef523 | 2079 | bool masked_p, |
ab2fc782 RS |
2080 | gather_scatter_info *gs_info) |
2081 | { | |
2082 | if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info) | |
2083 | || gs_info->decl) | |
429ef523 RS |
2084 | return vect_truncate_gather_scatter_offset (stmt, loop_vinfo, |
2085 | masked_p, gs_info); | |
ab2fc782 RS |
2086 | |
2087 | scalar_mode element_mode = SCALAR_TYPE_MODE (gs_info->element_type); | |
2088 | unsigned int element_bits = GET_MODE_BITSIZE (element_mode); | |
2089 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2090 | unsigned int offset_bits = TYPE_PRECISION (offset_type); | |
2091 | ||
2092 | /* Enforced by vect_check_gather_scatter. */ | |
2093 | gcc_assert (element_bits >= offset_bits); | |
2094 | ||
2095 | /* If the elements are wider than the offset, convert the offset to the | |
2096 | same width, without changing its sign. */ | |
2097 | if (element_bits > offset_bits) | |
2098 | { | |
2099 | bool unsigned_p = TYPE_UNSIGNED (offset_type); | |
2100 | offset_type = build_nonstandard_integer_type (element_bits, unsigned_p); | |
2101 | gs_info->offset = fold_convert (offset_type, gs_info->offset); | |
2102 | } | |
2103 | ||
2104 | if (dump_enabled_p ()) | |
2105 | dump_printf_loc (MSG_NOTE, vect_location, | |
2106 | "using gather/scatter for strided/grouped access," | |
2107 | " scale = %d\n", gs_info->scale); | |
2108 | ||
2109 | return true; | |
2110 | } | |
2111 | ||
62da9e14 RS |
2112 | /* STMT is a non-strided load or store, meaning that it accesses |
2113 | elements with a known constant step. Return -1 if that step | |
2114 | is negative, 0 if it is zero, and 1 if it is greater than zero. */ | |
2115 | ||
2116 | static int | |
2117 | compare_step_with_zero (gimple *stmt) | |
2118 | { | |
2119 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3f5e8a76 RS |
2120 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
2121 | return tree_int_cst_compare (vect_dr_behavior (dr)->step, | |
2122 | size_zero_node); | |
62da9e14 RS |
2123 | } |
2124 | ||
2125 | /* If the target supports a permute mask that reverses the elements in | |
2126 | a vector of type VECTYPE, return that mask, otherwise return null. */ | |
2127 | ||
2128 | static tree | |
2129 | perm_mask_for_reverse (tree vectype) | |
2130 | { | |
928686b1 | 2131 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
62da9e14 | 2132 | |
d980067b RS |
2133 | /* The encoding has a single stepped pattern. */ |
2134 | vec_perm_builder sel (nunits, 1, 3); | |
928686b1 | 2135 | for (int i = 0; i < 3; ++i) |
908a1a16 | 2136 | sel.quick_push (nunits - 1 - i); |
62da9e14 | 2137 | |
e3342de4 RS |
2138 | vec_perm_indices indices (sel, 1, nunits); |
2139 | if (!can_vec_perm_const_p (TYPE_MODE (vectype), indices)) | |
62da9e14 | 2140 | return NULL_TREE; |
e3342de4 | 2141 | return vect_gen_perm_mask_checked (vectype, indices); |
62da9e14 | 2142 | } |
5ce9450f | 2143 | |
c3a8f964 RS |
2144 | /* STMT is either a masked or unconditional store. Return the value |
2145 | being stored. */ | |
2146 | ||
f307441a | 2147 | tree |
c3a8f964 RS |
2148 | vect_get_store_rhs (gimple *stmt) |
2149 | { | |
2150 | if (gassign *assign = dyn_cast <gassign *> (stmt)) | |
2151 | { | |
2152 | gcc_assert (gimple_assign_single_p (assign)); | |
2153 | return gimple_assign_rhs1 (assign); | |
2154 | } | |
2155 | if (gcall *call = dyn_cast <gcall *> (stmt)) | |
2156 | { | |
2157 | internal_fn ifn = gimple_call_internal_fn (call); | |
f307441a RS |
2158 | int index = internal_fn_stored_value_index (ifn); |
2159 | gcc_assert (index >= 0); | |
2160 | return gimple_call_arg (stmt, index); | |
c3a8f964 RS |
2161 | } |
2162 | gcc_unreachable (); | |
2163 | } | |
2164 | ||
2de001ee RS |
2165 | /* A subroutine of get_load_store_type, with a subset of the same |
2166 | arguments. Handle the case where STMT is part of a grouped load | |
2167 | or store. | |
2168 | ||
2169 | For stores, the statements in the group are all consecutive | |
2170 | and there is no gap at the end. For loads, the statements in the | |
2171 | group might not be consecutive; there can be gaps between statements | |
2172 | as well as at the end. */ | |
2173 | ||
2174 | static bool | |
2175 | get_group_load_store_type (gimple *stmt, tree vectype, bool slp, | |
7e11fc7f | 2176 | bool masked_p, vec_load_store_type vls_type, |
429ef523 RS |
2177 | vect_memory_access_type *memory_access_type, |
2178 | gather_scatter_info *gs_info) | |
2de001ee RS |
2179 | { |
2180 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2181 | vec_info *vinfo = stmt_info->vinfo; | |
2182 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2183 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; | |
bffb8014 RS |
2184 | stmt_vec_info first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
2185 | data_reference *first_dr = STMT_VINFO_DATA_REF (first_stmt_info); | |
2186 | unsigned int group_size = DR_GROUP_SIZE (first_stmt_info); | |
2187 | bool single_element_p = (stmt_info == first_stmt_info | |
2c53b149 | 2188 | && !DR_GROUP_NEXT_ELEMENT (stmt_info)); |
bffb8014 | 2189 | unsigned HOST_WIDE_INT gap = DR_GROUP_GAP (first_stmt_info); |
928686b1 | 2190 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
2191 | |
2192 | /* True if the vectorized statements would access beyond the last | |
2193 | statement in the group. */ | |
2194 | bool overrun_p = false; | |
2195 | ||
2196 | /* True if we can cope with such overrun by peeling for gaps, so that | |
2197 | there is at least one final scalar iteration after the vector loop. */ | |
7e11fc7f RS |
2198 | bool can_overrun_p = (!masked_p |
2199 | && vls_type == VLS_LOAD | |
2200 | && loop_vinfo | |
2201 | && !loop->inner); | |
2de001ee RS |
2202 | |
2203 | /* There can only be a gap at the end of the group if the stride is | |
2204 | known at compile time. */ | |
2205 | gcc_assert (!STMT_VINFO_STRIDED_P (stmt_info) || gap == 0); | |
2206 | ||
2207 | /* Stores can't yet have gaps. */ | |
2208 | gcc_assert (slp || vls_type == VLS_LOAD || gap == 0); | |
2209 | ||
2210 | if (slp) | |
2211 | { | |
2212 | if (STMT_VINFO_STRIDED_P (stmt_info)) | |
2213 | { | |
2c53b149 | 2214 | /* Try to use consecutive accesses of DR_GROUP_SIZE elements, |
2de001ee RS |
2215 | separated by the stride, until we have a complete vector. |
2216 | Fall back to scalar accesses if that isn't possible. */ | |
928686b1 | 2217 | if (multiple_p (nunits, group_size)) |
2de001ee RS |
2218 | *memory_access_type = VMAT_STRIDED_SLP; |
2219 | else | |
2220 | *memory_access_type = VMAT_ELEMENTWISE; | |
2221 | } | |
2222 | else | |
2223 | { | |
2224 | overrun_p = loop_vinfo && gap != 0; | |
2225 | if (overrun_p && vls_type != VLS_LOAD) | |
2226 | { | |
2227 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2228 | "Grouped store with gaps requires" | |
2229 | " non-consecutive accesses\n"); | |
2230 | return false; | |
2231 | } | |
f702e7d4 RS |
2232 | /* An overrun is fine if the trailing elements are smaller |
2233 | than the alignment boundary B. Every vector access will | |
2234 | be a multiple of B and so we are guaranteed to access a | |
2235 | non-gap element in the same B-sized block. */ | |
f9ef2c76 | 2236 | if (overrun_p |
f702e7d4 RS |
2237 | && gap < (vect_known_alignment_in_bytes (first_dr) |
2238 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 2239 | overrun_p = false; |
2de001ee RS |
2240 | if (overrun_p && !can_overrun_p) |
2241 | { | |
2242 | if (dump_enabled_p ()) | |
2243 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2244 | "Peeling for outer loop is not supported\n"); | |
2245 | return false; | |
2246 | } | |
2247 | *memory_access_type = VMAT_CONTIGUOUS; | |
2248 | } | |
2249 | } | |
2250 | else | |
2251 | { | |
2252 | /* We can always handle this case using elementwise accesses, | |
2253 | but see if something more efficient is available. */ | |
2254 | *memory_access_type = VMAT_ELEMENTWISE; | |
2255 | ||
2256 | /* If there is a gap at the end of the group then these optimizations | |
2257 | would access excess elements in the last iteration. */ | |
2258 | bool would_overrun_p = (gap != 0); | |
f702e7d4 RS |
2259 | /* An overrun is fine if the trailing elements are smaller than the |
2260 | alignment boundary B. Every vector access will be a multiple of B | |
2261 | and so we are guaranteed to access a non-gap element in the | |
2262 | same B-sized block. */ | |
f9ef2c76 | 2263 | if (would_overrun_p |
7e11fc7f | 2264 | && !masked_p |
f702e7d4 RS |
2265 | && gap < (vect_known_alignment_in_bytes (first_dr) |
2266 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 2267 | would_overrun_p = false; |
f702e7d4 | 2268 | |
2de001ee | 2269 | if (!STMT_VINFO_STRIDED_P (stmt_info) |
62da9e14 RS |
2270 | && (can_overrun_p || !would_overrun_p) |
2271 | && compare_step_with_zero (stmt) > 0) | |
2de001ee | 2272 | { |
6737facb RS |
2273 | /* First cope with the degenerate case of a single-element |
2274 | vector. */ | |
2275 | if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U)) | |
2276 | *memory_access_type = VMAT_CONTIGUOUS; | |
2277 | ||
2278 | /* Otherwise try using LOAD/STORE_LANES. */ | |
2279 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2280 | && (vls_type == VLS_LOAD | |
7e11fc7f RS |
2281 | ? vect_load_lanes_supported (vectype, group_size, masked_p) |
2282 | : vect_store_lanes_supported (vectype, group_size, | |
2283 | masked_p))) | |
2de001ee RS |
2284 | { |
2285 | *memory_access_type = VMAT_LOAD_STORE_LANES; | |
2286 | overrun_p = would_overrun_p; | |
2287 | } | |
2288 | ||
2289 | /* If that fails, try using permuting loads. */ | |
2290 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2291 | && (vls_type == VLS_LOAD | |
2292 | ? vect_grouped_load_supported (vectype, single_element_p, | |
2293 | group_size) | |
2294 | : vect_grouped_store_supported (vectype, group_size))) | |
2295 | { | |
2296 | *memory_access_type = VMAT_CONTIGUOUS_PERMUTE; | |
2297 | overrun_p = would_overrun_p; | |
2298 | } | |
2299 | } | |
429ef523 RS |
2300 | |
2301 | /* As a last resort, trying using a gather load or scatter store. | |
2302 | ||
2303 | ??? Although the code can handle all group sizes correctly, | |
2304 | it probably isn't a win to use separate strided accesses based | |
2305 | on nearby locations. Or, even if it's a win over scalar code, | |
2306 | it might not be a win over vectorizing at a lower VF, if that | |
2307 | allows us to use contiguous accesses. */ | |
2308 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2309 | && single_element_p | |
2310 | && loop_vinfo | |
2311 | && vect_use_strided_gather_scatters_p (stmt, loop_vinfo, | |
2312 | masked_p, gs_info)) | |
2313 | *memory_access_type = VMAT_GATHER_SCATTER; | |
2de001ee RS |
2314 | } |
2315 | ||
bffb8014 | 2316 | if (vls_type != VLS_LOAD && first_stmt_info == stmt_info) |
2de001ee RS |
2317 | { |
2318 | /* STMT is the leader of the group. Check the operands of all the | |
2319 | stmts of the group. */ | |
bffb8014 RS |
2320 | stmt_vec_info next_stmt_info = DR_GROUP_NEXT_ELEMENT (stmt_info); |
2321 | while (next_stmt_info) | |
2de001ee | 2322 | { |
bffb8014 | 2323 | tree op = vect_get_store_rhs (next_stmt_info); |
2de001ee | 2324 | enum vect_def_type dt; |
894dd753 | 2325 | if (!vect_is_simple_use (op, vinfo, &dt)) |
2de001ee RS |
2326 | { |
2327 | if (dump_enabled_p ()) | |
2328 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2329 | "use not simple.\n"); | |
2330 | return false; | |
2331 | } | |
bffb8014 | 2332 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
2de001ee RS |
2333 | } |
2334 | } | |
2335 | ||
2336 | if (overrun_p) | |
2337 | { | |
2338 | gcc_assert (can_overrun_p); | |
2339 | if (dump_enabled_p ()) | |
2340 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2341 | "Data access with gaps requires scalar " | |
2342 | "epilogue loop\n"); | |
2343 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; | |
2344 | } | |
2345 | ||
2346 | return true; | |
2347 | } | |
2348 | ||
62da9e14 RS |
2349 | /* A subroutine of get_load_store_type, with a subset of the same |
2350 | arguments. Handle the case where STMT is a load or store that | |
2351 | accesses consecutive elements with a negative step. */ | |
2352 | ||
2353 | static vect_memory_access_type | |
2354 | get_negative_load_store_type (gimple *stmt, tree vectype, | |
2355 | vec_load_store_type vls_type, | |
2356 | unsigned int ncopies) | |
2357 | { | |
2358 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2359 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2360 | dr_alignment_support alignment_support_scheme; | |
2361 | ||
2362 | if (ncopies > 1) | |
2363 | { | |
2364 | if (dump_enabled_p ()) | |
2365 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2366 | "multiple types with negative step.\n"); | |
2367 | return VMAT_ELEMENTWISE; | |
2368 | } | |
2369 | ||
2370 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
2371 | if (alignment_support_scheme != dr_aligned | |
2372 | && alignment_support_scheme != dr_unaligned_supported) | |
2373 | { | |
2374 | if (dump_enabled_p ()) | |
2375 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2376 | "negative step but alignment required.\n"); | |
2377 | return VMAT_ELEMENTWISE; | |
2378 | } | |
2379 | ||
2380 | if (vls_type == VLS_STORE_INVARIANT) | |
2381 | { | |
2382 | if (dump_enabled_p ()) | |
2383 | dump_printf_loc (MSG_NOTE, vect_location, | |
2384 | "negative step with invariant source;" | |
2385 | " no permute needed.\n"); | |
2386 | return VMAT_CONTIGUOUS_DOWN; | |
2387 | } | |
2388 | ||
2389 | if (!perm_mask_for_reverse (vectype)) | |
2390 | { | |
2391 | if (dump_enabled_p ()) | |
2392 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2393 | "negative step and reversing not supported.\n"); | |
2394 | return VMAT_ELEMENTWISE; | |
2395 | } | |
2396 | ||
2397 | return VMAT_CONTIGUOUS_REVERSE; | |
2398 | } | |
2399 | ||
2de001ee RS |
2400 | /* Analyze load or store statement STMT of type VLS_TYPE. Return true |
2401 | if there is a memory access type that the vectorized form can use, | |
2402 | storing it in *MEMORY_ACCESS_TYPE if so. If we decide to use gathers | |
2403 | or scatters, fill in GS_INFO accordingly. | |
2404 | ||
2405 | SLP says whether we're performing SLP rather than loop vectorization. | |
7e11fc7f | 2406 | MASKED_P is true if the statement is conditional on a vectorized mask. |
62da9e14 RS |
2407 | VECTYPE is the vector type that the vectorized statements will use. |
2408 | NCOPIES is the number of vector statements that will be needed. */ | |
2de001ee RS |
2409 | |
2410 | static bool | |
7e11fc7f | 2411 | get_load_store_type (gimple *stmt, tree vectype, bool slp, bool masked_p, |
62da9e14 | 2412 | vec_load_store_type vls_type, unsigned int ncopies, |
2de001ee RS |
2413 | vect_memory_access_type *memory_access_type, |
2414 | gather_scatter_info *gs_info) | |
2415 | { | |
2416 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2417 | vec_info *vinfo = stmt_info->vinfo; | |
2418 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4d694b27 | 2419 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
2420 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
2421 | { | |
2422 | *memory_access_type = VMAT_GATHER_SCATTER; | |
2de001ee RS |
2423 | if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info)) |
2424 | gcc_unreachable (); | |
894dd753 | 2425 | else if (!vect_is_simple_use (gs_info->offset, vinfo, |
2de001ee RS |
2426 | &gs_info->offset_dt, |
2427 | &gs_info->offset_vectype)) | |
2428 | { | |
2429 | if (dump_enabled_p ()) | |
2430 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2431 | "%s index use not simple.\n", | |
2432 | vls_type == VLS_LOAD ? "gather" : "scatter"); | |
2433 | return false; | |
2434 | } | |
2435 | } | |
2436 | else if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
2437 | { | |
7e11fc7f | 2438 | if (!get_group_load_store_type (stmt, vectype, slp, masked_p, vls_type, |
429ef523 | 2439 | memory_access_type, gs_info)) |
2de001ee RS |
2440 | return false; |
2441 | } | |
2442 | else if (STMT_VINFO_STRIDED_P (stmt_info)) | |
2443 | { | |
2444 | gcc_assert (!slp); | |
ab2fc782 | 2445 | if (loop_vinfo |
429ef523 RS |
2446 | && vect_use_strided_gather_scatters_p (stmt, loop_vinfo, |
2447 | masked_p, gs_info)) | |
ab2fc782 RS |
2448 | *memory_access_type = VMAT_GATHER_SCATTER; |
2449 | else | |
2450 | *memory_access_type = VMAT_ELEMENTWISE; | |
2de001ee RS |
2451 | } |
2452 | else | |
62da9e14 RS |
2453 | { |
2454 | int cmp = compare_step_with_zero (stmt); | |
2455 | if (cmp < 0) | |
2456 | *memory_access_type = get_negative_load_store_type | |
2457 | (stmt, vectype, vls_type, ncopies); | |
2458 | else if (cmp == 0) | |
2459 | { | |
2460 | gcc_assert (vls_type == VLS_LOAD); | |
2461 | *memory_access_type = VMAT_INVARIANT; | |
2462 | } | |
2463 | else | |
2464 | *memory_access_type = VMAT_CONTIGUOUS; | |
2465 | } | |
2de001ee | 2466 | |
4d694b27 RS |
2467 | if ((*memory_access_type == VMAT_ELEMENTWISE |
2468 | || *memory_access_type == VMAT_STRIDED_SLP) | |
2469 | && !nunits.is_constant ()) | |
2470 | { | |
2471 | if (dump_enabled_p ()) | |
2472 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2473 | "Not using elementwise accesses due to variable " | |
2474 | "vectorization factor.\n"); | |
2475 | return false; | |
2476 | } | |
2477 | ||
2de001ee RS |
2478 | /* FIXME: At the moment the cost model seems to underestimate the |
2479 | cost of using elementwise accesses. This check preserves the | |
2480 | traditional behavior until that can be fixed. */ | |
2481 | if (*memory_access_type == VMAT_ELEMENTWISE | |
4aa157e8 | 2482 | && !STMT_VINFO_STRIDED_P (stmt_info) |
bffb8014 | 2483 | && !(stmt_info == DR_GROUP_FIRST_ELEMENT (stmt_info) |
2c53b149 RB |
2484 | && !DR_GROUP_NEXT_ELEMENT (stmt_info) |
2485 | && !pow2p_hwi (DR_GROUP_SIZE (stmt_info)))) | |
2de001ee RS |
2486 | { |
2487 | if (dump_enabled_p ()) | |
2488 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2489 | "not falling back to elementwise accesses\n"); | |
2490 | return false; | |
2491 | } | |
2492 | return true; | |
2493 | } | |
2494 | ||
aaeefd88 | 2495 | /* Return true if boolean argument MASK is suitable for vectorizing |
929b4411 RS |
2496 | conditional load or store STMT. When returning true, store the type |
2497 | of the definition in *MASK_DT_OUT and the type of the vectorized mask | |
2498 | in *MASK_VECTYPE_OUT. */ | |
aaeefd88 RS |
2499 | |
2500 | static bool | |
929b4411 RS |
2501 | vect_check_load_store_mask (gimple *stmt, tree mask, |
2502 | vect_def_type *mask_dt_out, | |
2503 | tree *mask_vectype_out) | |
aaeefd88 RS |
2504 | { |
2505 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (mask))) | |
2506 | { | |
2507 | if (dump_enabled_p ()) | |
2508 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2509 | "mask argument is not a boolean.\n"); | |
2510 | return false; | |
2511 | } | |
2512 | ||
2513 | if (TREE_CODE (mask) != SSA_NAME) | |
2514 | { | |
2515 | if (dump_enabled_p ()) | |
2516 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2517 | "mask argument is not an SSA name.\n"); | |
2518 | return false; | |
2519 | } | |
2520 | ||
2521 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
929b4411 | 2522 | enum vect_def_type mask_dt; |
aaeefd88 | 2523 | tree mask_vectype; |
894dd753 | 2524 | if (!vect_is_simple_use (mask, stmt_info->vinfo, &mask_dt, &mask_vectype)) |
aaeefd88 RS |
2525 | { |
2526 | if (dump_enabled_p ()) | |
2527 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2528 | "mask use not simple.\n"); | |
2529 | return false; | |
2530 | } | |
2531 | ||
2532 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2533 | if (!mask_vectype) | |
2534 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
2535 | ||
2536 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype)) | |
2537 | { | |
2538 | if (dump_enabled_p ()) | |
2539 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2540 | "could not find an appropriate vector mask type.\n"); | |
2541 | return false; | |
2542 | } | |
2543 | ||
2544 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (mask_vectype), | |
2545 | TYPE_VECTOR_SUBPARTS (vectype))) | |
2546 | { | |
2547 | if (dump_enabled_p ()) | |
2548 | { | |
2549 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2550 | "vector mask type "); | |
2551 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, mask_vectype); | |
2552 | dump_printf (MSG_MISSED_OPTIMIZATION, | |
2553 | " does not match vector data type "); | |
2554 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, vectype); | |
2555 | dump_printf (MSG_MISSED_OPTIMIZATION, ".\n"); | |
2556 | } | |
2557 | return false; | |
2558 | } | |
2559 | ||
929b4411 | 2560 | *mask_dt_out = mask_dt; |
aaeefd88 RS |
2561 | *mask_vectype_out = mask_vectype; |
2562 | return true; | |
2563 | } | |
2564 | ||
3133c3b6 RS |
2565 | /* Return true if stored value RHS is suitable for vectorizing store |
2566 | statement STMT. When returning true, store the type of the | |
929b4411 RS |
2567 | definition in *RHS_DT_OUT, the type of the vectorized store value in |
2568 | *RHS_VECTYPE_OUT and the type of the store in *VLS_TYPE_OUT. */ | |
3133c3b6 RS |
2569 | |
2570 | static bool | |
929b4411 RS |
2571 | vect_check_store_rhs (gimple *stmt, tree rhs, vect_def_type *rhs_dt_out, |
2572 | tree *rhs_vectype_out, vec_load_store_type *vls_type_out) | |
3133c3b6 RS |
2573 | { |
2574 | /* In the case this is a store from a constant make sure | |
2575 | native_encode_expr can handle it. */ | |
2576 | if (CONSTANT_CLASS_P (rhs) && native_encode_expr (rhs, NULL, 64) == 0) | |
2577 | { | |
2578 | if (dump_enabled_p ()) | |
2579 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2580 | "cannot encode constant as a byte sequence.\n"); | |
2581 | return false; | |
2582 | } | |
2583 | ||
2584 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
929b4411 | 2585 | enum vect_def_type rhs_dt; |
3133c3b6 | 2586 | tree rhs_vectype; |
894dd753 | 2587 | if (!vect_is_simple_use (rhs, stmt_info->vinfo, &rhs_dt, &rhs_vectype)) |
3133c3b6 RS |
2588 | { |
2589 | if (dump_enabled_p ()) | |
2590 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2591 | "use not simple.\n"); | |
2592 | return false; | |
2593 | } | |
2594 | ||
2595 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2596 | if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype)) | |
2597 | { | |
2598 | if (dump_enabled_p ()) | |
2599 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2600 | "incompatible vector types.\n"); | |
2601 | return false; | |
2602 | } | |
2603 | ||
929b4411 | 2604 | *rhs_dt_out = rhs_dt; |
3133c3b6 | 2605 | *rhs_vectype_out = rhs_vectype; |
929b4411 | 2606 | if (rhs_dt == vect_constant_def || rhs_dt == vect_external_def) |
3133c3b6 RS |
2607 | *vls_type_out = VLS_STORE_INVARIANT; |
2608 | else | |
2609 | *vls_type_out = VLS_STORE; | |
2610 | return true; | |
2611 | } | |
2612 | ||
bc9587eb RS |
2613 | /* Build an all-ones vector mask of type MASKTYPE while vectorizing STMT. |
2614 | Note that we support masks with floating-point type, in which case the | |
2615 | floats are interpreted as a bitmask. */ | |
2616 | ||
2617 | static tree | |
2618 | vect_build_all_ones_mask (gimple *stmt, tree masktype) | |
2619 | { | |
2620 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
2621 | return build_int_cst (masktype, -1); | |
2622 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
2623 | { | |
2624 | tree mask = build_int_cst (TREE_TYPE (masktype), -1); | |
2625 | mask = build_vector_from_val (masktype, mask); | |
2626 | return vect_init_vector (stmt, mask, masktype, NULL); | |
2627 | } | |
2628 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) | |
2629 | { | |
2630 | REAL_VALUE_TYPE r; | |
2631 | long tmp[6]; | |
2632 | for (int j = 0; j < 6; ++j) | |
2633 | tmp[j] = -1; | |
2634 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
2635 | tree mask = build_real (TREE_TYPE (masktype), r); | |
2636 | mask = build_vector_from_val (masktype, mask); | |
2637 | return vect_init_vector (stmt, mask, masktype, NULL); | |
2638 | } | |
2639 | gcc_unreachable (); | |
2640 | } | |
2641 | ||
2642 | /* Build an all-zero merge value of type VECTYPE while vectorizing | |
2643 | STMT as a gather load. */ | |
2644 | ||
2645 | static tree | |
2646 | vect_build_zero_merge_argument (gimple *stmt, tree vectype) | |
2647 | { | |
2648 | tree merge; | |
2649 | if (TREE_CODE (TREE_TYPE (vectype)) == INTEGER_TYPE) | |
2650 | merge = build_int_cst (TREE_TYPE (vectype), 0); | |
2651 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (vectype))) | |
2652 | { | |
2653 | REAL_VALUE_TYPE r; | |
2654 | long tmp[6]; | |
2655 | for (int j = 0; j < 6; ++j) | |
2656 | tmp[j] = 0; | |
2657 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (vectype))); | |
2658 | merge = build_real (TREE_TYPE (vectype), r); | |
2659 | } | |
2660 | else | |
2661 | gcc_unreachable (); | |
2662 | merge = build_vector_from_val (vectype, merge); | |
2663 | return vect_init_vector (stmt, merge, vectype, NULL); | |
2664 | } | |
2665 | ||
c48d2d35 RS |
2666 | /* Build a gather load call while vectorizing STMT. Insert new instructions |
2667 | before GSI and add them to VEC_STMT. GS_INFO describes the gather load | |
2668 | operation. If the load is conditional, MASK is the unvectorized | |
929b4411 | 2669 | condition and MASK_DT is its definition type, otherwise MASK is null. */ |
c48d2d35 RS |
2670 | |
2671 | static void | |
2672 | vect_build_gather_load_calls (gimple *stmt, gimple_stmt_iterator *gsi, | |
1eede195 RS |
2673 | stmt_vec_info *vec_stmt, |
2674 | gather_scatter_info *gs_info, tree mask, | |
2675 | vect_def_type mask_dt) | |
c48d2d35 RS |
2676 | { |
2677 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2678 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2679 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2680 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2681 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2682 | int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
2683 | edge pe = loop_preheader_edge (loop); | |
2684 | enum { NARROW, NONE, WIDEN } modifier; | |
2685 | poly_uint64 gather_off_nunits | |
2686 | = TYPE_VECTOR_SUBPARTS (gs_info->offset_vectype); | |
2687 | ||
2688 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info->decl)); | |
2689 | tree rettype = TREE_TYPE (TREE_TYPE (gs_info->decl)); | |
2690 | tree srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2691 | tree ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2692 | tree idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2693 | tree masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2694 | tree scaletype = TREE_VALUE (arglist); | |
2695 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
2696 | && (!mask || types_compatible_p (srctype, masktype))); | |
2697 | ||
2698 | tree perm_mask = NULL_TREE; | |
2699 | tree mask_perm_mask = NULL_TREE; | |
2700 | if (known_eq (nunits, gather_off_nunits)) | |
2701 | modifier = NONE; | |
2702 | else if (known_eq (nunits * 2, gather_off_nunits)) | |
2703 | { | |
2704 | modifier = WIDEN; | |
2705 | ||
2706 | /* Currently widening gathers and scatters are only supported for | |
2707 | fixed-length vectors. */ | |
2708 | int count = gather_off_nunits.to_constant (); | |
2709 | vec_perm_builder sel (count, count, 1); | |
2710 | for (int i = 0; i < count; ++i) | |
2711 | sel.quick_push (i | (count / 2)); | |
2712 | ||
2713 | vec_perm_indices indices (sel, 1, count); | |
2714 | perm_mask = vect_gen_perm_mask_checked (gs_info->offset_vectype, | |
2715 | indices); | |
2716 | } | |
2717 | else if (known_eq (nunits, gather_off_nunits * 2)) | |
2718 | { | |
2719 | modifier = NARROW; | |
2720 | ||
2721 | /* Currently narrowing gathers and scatters are only supported for | |
2722 | fixed-length vectors. */ | |
2723 | int count = nunits.to_constant (); | |
2724 | vec_perm_builder sel (count, count, 1); | |
2725 | sel.quick_grow (count); | |
2726 | for (int i = 0; i < count; ++i) | |
2727 | sel[i] = i < count / 2 ? i : i + count / 2; | |
2728 | vec_perm_indices indices (sel, 2, count); | |
2729 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); | |
2730 | ||
2731 | ncopies *= 2; | |
2732 | ||
2733 | if (mask) | |
2734 | { | |
2735 | for (int i = 0; i < count; ++i) | |
2736 | sel[i] = i | (count / 2); | |
2737 | indices.new_vector (sel, 2, count); | |
2738 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, indices); | |
2739 | } | |
2740 | } | |
2741 | else | |
2742 | gcc_unreachable (); | |
2743 | ||
2744 | tree vec_dest = vect_create_destination_var (gimple_get_lhs (stmt), | |
2745 | vectype); | |
2746 | ||
2747 | tree ptr = fold_convert (ptrtype, gs_info->base); | |
2748 | if (!is_gimple_min_invariant (ptr)) | |
2749 | { | |
2750 | gimple_seq seq; | |
2751 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
2752 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
2753 | gcc_assert (!new_bb); | |
2754 | } | |
2755 | ||
2756 | tree scale = build_int_cst (scaletype, gs_info->scale); | |
2757 | ||
2758 | tree vec_oprnd0 = NULL_TREE; | |
2759 | tree vec_mask = NULL_TREE; | |
2760 | tree src_op = NULL_TREE; | |
2761 | tree mask_op = NULL_TREE; | |
2762 | tree prev_res = NULL_TREE; | |
2763 | stmt_vec_info prev_stmt_info = NULL; | |
2764 | ||
2765 | if (!mask) | |
2766 | { | |
2767 | src_op = vect_build_zero_merge_argument (stmt, rettype); | |
2768 | mask_op = vect_build_all_ones_mask (stmt, masktype); | |
2769 | } | |
2770 | ||
2771 | for (int j = 0; j < ncopies; ++j) | |
2772 | { | |
2773 | tree op, var; | |
c48d2d35 RS |
2774 | if (modifier == WIDEN && (j & 1)) |
2775 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
2776 | perm_mask, stmt, gsi); | |
2777 | else if (j == 0) | |
2778 | op = vec_oprnd0 | |
2779 | = vect_get_vec_def_for_operand (gs_info->offset, stmt); | |
2780 | else | |
2781 | op = vec_oprnd0 | |
2782 | = vect_get_vec_def_for_stmt_copy (gs_info->offset_dt, vec_oprnd0); | |
2783 | ||
2784 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
2785 | { | |
2786 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), | |
2787 | TYPE_VECTOR_SUBPARTS (idxtype))); | |
2788 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); | |
2789 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); | |
e1bd7296 | 2790 | gassign *new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
c48d2d35 RS |
2791 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2792 | op = var; | |
2793 | } | |
2794 | ||
2795 | if (mask) | |
2796 | { | |
2797 | if (mask_perm_mask && (j & 1)) | |
2798 | mask_op = permute_vec_elements (mask_op, mask_op, | |
2799 | mask_perm_mask, stmt, gsi); | |
2800 | else | |
2801 | { | |
2802 | if (j == 0) | |
2803 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); | |
2804 | else | |
929b4411 | 2805 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
c48d2d35 RS |
2806 | |
2807 | mask_op = vec_mask; | |
2808 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
2809 | { | |
2810 | gcc_assert | |
2811 | (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)), | |
2812 | TYPE_VECTOR_SUBPARTS (masktype))); | |
2813 | var = vect_get_new_ssa_name (masktype, vect_simple_var); | |
2814 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); | |
e1bd7296 RS |
2815 | gassign *new_stmt |
2816 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_op); | |
c48d2d35 RS |
2817 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2818 | mask_op = var; | |
2819 | } | |
2820 | } | |
2821 | src_op = mask_op; | |
2822 | } | |
2823 | ||
e1bd7296 RS |
2824 | gcall *new_call = gimple_build_call (gs_info->decl, 5, src_op, ptr, op, |
2825 | mask_op, scale); | |
c48d2d35 | 2826 | |
e1bd7296 | 2827 | stmt_vec_info new_stmt_info; |
c48d2d35 RS |
2828 | if (!useless_type_conversion_p (vectype, rettype)) |
2829 | { | |
2830 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (vectype), | |
2831 | TYPE_VECTOR_SUBPARTS (rettype))); | |
2832 | op = vect_get_new_ssa_name (rettype, vect_simple_var); | |
e1bd7296 RS |
2833 | gimple_call_set_lhs (new_call, op); |
2834 | vect_finish_stmt_generation (stmt, new_call, gsi); | |
c48d2d35 RS |
2835 | var = make_ssa_name (vec_dest); |
2836 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); | |
e1bd7296 RS |
2837 | gassign *new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
2838 | new_stmt_info = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
c48d2d35 RS |
2839 | } |
2840 | else | |
2841 | { | |
e1bd7296 RS |
2842 | var = make_ssa_name (vec_dest, new_call); |
2843 | gimple_call_set_lhs (new_call, var); | |
2844 | new_stmt_info = vect_finish_stmt_generation (stmt, new_call, gsi); | |
c48d2d35 RS |
2845 | } |
2846 | ||
c48d2d35 RS |
2847 | if (modifier == NARROW) |
2848 | { | |
2849 | if ((j & 1) == 0) | |
2850 | { | |
2851 | prev_res = var; | |
2852 | continue; | |
2853 | } | |
2854 | var = permute_vec_elements (prev_res, var, perm_mask, stmt, gsi); | |
e1bd7296 | 2855 | new_stmt_info = loop_vinfo->lookup_def (var); |
c48d2d35 RS |
2856 | } |
2857 | ||
dbe1b846 | 2858 | if (prev_stmt_info == NULL_STMT_VEC_INFO) |
e1bd7296 | 2859 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
c48d2d35 | 2860 | else |
e1bd7296 RS |
2861 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
2862 | prev_stmt_info = new_stmt_info; | |
c48d2d35 RS |
2863 | } |
2864 | } | |
2865 | ||
bfaa08b7 RS |
2866 | /* Prepare the base and offset in GS_INFO for vectorization. |
2867 | Set *DATAREF_PTR to the loop-invariant base address and *VEC_OFFSET | |
2868 | to the vectorized offset argument for the first copy of STMT. STMT | |
2869 | is the statement described by GS_INFO and LOOP is the containing loop. */ | |
2870 | ||
2871 | static void | |
2872 | vect_get_gather_scatter_ops (struct loop *loop, gimple *stmt, | |
2873 | gather_scatter_info *gs_info, | |
2874 | tree *dataref_ptr, tree *vec_offset) | |
2875 | { | |
2876 | gimple_seq stmts = NULL; | |
2877 | *dataref_ptr = force_gimple_operand (gs_info->base, &stmts, true, NULL_TREE); | |
2878 | if (stmts != NULL) | |
2879 | { | |
2880 | basic_block new_bb; | |
2881 | edge pe = loop_preheader_edge (loop); | |
2882 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
2883 | gcc_assert (!new_bb); | |
2884 | } | |
2885 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2886 | tree offset_vectype = get_vectype_for_scalar_type (offset_type); | |
2887 | *vec_offset = vect_get_vec_def_for_operand (gs_info->offset, stmt, | |
2888 | offset_vectype); | |
2889 | } | |
2890 | ||
ab2fc782 RS |
2891 | /* Prepare to implement a grouped or strided load or store using |
2892 | the gather load or scatter store operation described by GS_INFO. | |
2893 | STMT is the load or store statement. | |
2894 | ||
2895 | Set *DATAREF_BUMP to the amount that should be added to the base | |
2896 | address after each copy of the vectorized statement. Set *VEC_OFFSET | |
2897 | to an invariant offset vector in which element I has the value | |
2898 | I * DR_STEP / SCALE. */ | |
2899 | ||
2900 | static void | |
2901 | vect_get_strided_load_store_ops (gimple *stmt, loop_vec_info loop_vinfo, | |
2902 | gather_scatter_info *gs_info, | |
2903 | tree *dataref_bump, tree *vec_offset) | |
2904 | { | |
2905 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2906 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2907 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2908 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2909 | gimple_seq stmts; | |
2910 | ||
2911 | tree bump = size_binop (MULT_EXPR, | |
2912 | fold_convert (sizetype, DR_STEP (dr)), | |
2913 | size_int (TYPE_VECTOR_SUBPARTS (vectype))); | |
2914 | *dataref_bump = force_gimple_operand (bump, &stmts, true, NULL_TREE); | |
2915 | if (stmts) | |
2916 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
2917 | ||
2918 | /* The offset given in GS_INFO can have pointer type, so use the element | |
2919 | type of the vector instead. */ | |
2920 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2921 | tree offset_vectype = get_vectype_for_scalar_type (offset_type); | |
2922 | offset_type = TREE_TYPE (offset_vectype); | |
2923 | ||
2924 | /* Calculate X = DR_STEP / SCALE and convert it to the appropriate type. */ | |
2925 | tree step = size_binop (EXACT_DIV_EXPR, DR_STEP (dr), | |
2926 | ssize_int (gs_info->scale)); | |
2927 | step = fold_convert (offset_type, step); | |
2928 | step = force_gimple_operand (step, &stmts, true, NULL_TREE); | |
2929 | ||
2930 | /* Create {0, X, X*2, X*3, ...}. */ | |
2931 | *vec_offset = gimple_build (&stmts, VEC_SERIES_EXPR, offset_vectype, | |
2932 | build_zero_cst (offset_type), step); | |
2933 | if (stmts) | |
2934 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
2935 | } | |
2936 | ||
2937 | /* Return the amount that should be added to a vector pointer to move | |
2938 | to the next or previous copy of AGGR_TYPE. DR is the data reference | |
2939 | being vectorized and MEMORY_ACCESS_TYPE describes the type of | |
2940 | vectorization. */ | |
2941 | ||
2942 | static tree | |
2943 | vect_get_data_ptr_increment (data_reference *dr, tree aggr_type, | |
2944 | vect_memory_access_type memory_access_type) | |
2945 | { | |
2946 | if (memory_access_type == VMAT_INVARIANT) | |
2947 | return size_zero_node; | |
2948 | ||
2949 | tree iv_step = TYPE_SIZE_UNIT (aggr_type); | |
2950 | tree step = vect_dr_behavior (dr)->step; | |
2951 | if (tree_int_cst_sgn (step) == -1) | |
2952 | iv_step = fold_build1 (NEGATE_EXPR, TREE_TYPE (iv_step), iv_step); | |
2953 | return iv_step; | |
2954 | } | |
2955 | ||
37b14185 RB |
2956 | /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64}. */ |
2957 | ||
2958 | static bool | |
2959 | vectorizable_bswap (gimple *stmt, gimple_stmt_iterator *gsi, | |
1eede195 | 2960 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
68435eb2 RB |
2961 | tree vectype_in, enum vect_def_type *dt, |
2962 | stmt_vector_for_cost *cost_vec) | |
37b14185 RB |
2963 | { |
2964 | tree op, vectype; | |
2965 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2966 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
928686b1 RS |
2967 | unsigned ncopies; |
2968 | unsigned HOST_WIDE_INT nunits, num_bytes; | |
37b14185 RB |
2969 | |
2970 | op = gimple_call_arg (stmt, 0); | |
2971 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
928686b1 RS |
2972 | |
2973 | if (!TYPE_VECTOR_SUBPARTS (vectype).is_constant (&nunits)) | |
2974 | return false; | |
37b14185 RB |
2975 | |
2976 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2977 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2978 | case of SLP. */ | |
2979 | if (slp_node) | |
2980 | ncopies = 1; | |
2981 | else | |
e8f142e2 | 2982 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
37b14185 RB |
2983 | |
2984 | gcc_assert (ncopies >= 1); | |
2985 | ||
2986 | tree char_vectype = get_same_sized_vectype (char_type_node, vectype_in); | |
2987 | if (! char_vectype) | |
2988 | return false; | |
2989 | ||
928686b1 RS |
2990 | if (!TYPE_VECTOR_SUBPARTS (char_vectype).is_constant (&num_bytes)) |
2991 | return false; | |
2992 | ||
794e3180 | 2993 | unsigned word_bytes = num_bytes / nunits; |
908a1a16 | 2994 | |
d980067b RS |
2995 | /* The encoding uses one stepped pattern for each byte in the word. */ |
2996 | vec_perm_builder elts (num_bytes, word_bytes, 3); | |
2997 | for (unsigned i = 0; i < 3; ++i) | |
37b14185 | 2998 | for (unsigned j = 0; j < word_bytes; ++j) |
908a1a16 | 2999 | elts.quick_push ((i + 1) * word_bytes - j - 1); |
37b14185 | 3000 | |
e3342de4 RS |
3001 | vec_perm_indices indices (elts, 1, num_bytes); |
3002 | if (!can_vec_perm_const_p (TYPE_MODE (char_vectype), indices)) | |
37b14185 RB |
3003 | return false; |
3004 | ||
3005 | if (! vec_stmt) | |
3006 | { | |
3007 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
adac3a68 | 3008 | DUMP_VECT_SCOPE ("vectorizable_bswap"); |
78604de0 | 3009 | if (! slp_node) |
37b14185 | 3010 | { |
68435eb2 RB |
3011 | record_stmt_cost (cost_vec, |
3012 | 1, vector_stmt, stmt_info, 0, vect_prologue); | |
3013 | record_stmt_cost (cost_vec, | |
3014 | ncopies, vec_perm, stmt_info, 0, vect_body); | |
37b14185 RB |
3015 | } |
3016 | return true; | |
3017 | } | |
3018 | ||
736d0f28 | 3019 | tree bswap_vconst = vec_perm_indices_to_tree (char_vectype, indices); |
37b14185 RB |
3020 | |
3021 | /* Transform. */ | |
3022 | vec<tree> vec_oprnds = vNULL; | |
e1bd7296 | 3023 | stmt_vec_info new_stmt_info = NULL; |
37b14185 RB |
3024 | stmt_vec_info prev_stmt_info = NULL; |
3025 | for (unsigned j = 0; j < ncopies; j++) | |
3026 | { | |
3027 | /* Handle uses. */ | |
3028 | if (j == 0) | |
306b0c92 | 3029 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
37b14185 RB |
3030 | else |
3031 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
3032 | ||
3033 | /* Arguments are ready. create the new vector stmt. */ | |
3034 | unsigned i; | |
3035 | tree vop; | |
3036 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) | |
3037 | { | |
e1bd7296 | 3038 | gimple *new_stmt; |
37b14185 RB |
3039 | tree tem = make_ssa_name (char_vectype); |
3040 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
3041 | char_vectype, vop)); | |
3042 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3043 | tree tem2 = make_ssa_name (char_vectype); | |
3044 | new_stmt = gimple_build_assign (tem2, VEC_PERM_EXPR, | |
3045 | tem, tem, bswap_vconst); | |
3046 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3047 | tem = make_ssa_name (vectype); | |
3048 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
3049 | vectype, tem2)); | |
e1bd7296 | 3050 | new_stmt_info = vect_finish_stmt_generation (stmt, new_stmt, gsi); |
37b14185 | 3051 | if (slp_node) |
e1bd7296 | 3052 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
37b14185 RB |
3053 | } |
3054 | ||
3055 | if (slp_node) | |
3056 | continue; | |
3057 | ||
3058 | if (j == 0) | |
e1bd7296 | 3059 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
37b14185 | 3060 | else |
e1bd7296 | 3061 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
37b14185 | 3062 | |
e1bd7296 | 3063 | prev_stmt_info = new_stmt_info; |
37b14185 RB |
3064 | } |
3065 | ||
3066 | vec_oprnds.release (); | |
3067 | return true; | |
3068 | } | |
3069 | ||
b1b6836e RS |
3070 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
3071 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT | |
3072 | in a single step. On success, store the binary pack code in | |
3073 | *CONVERT_CODE. */ | |
3074 | ||
3075 | static bool | |
3076 | simple_integer_narrowing (tree vectype_out, tree vectype_in, | |
3077 | tree_code *convert_code) | |
3078 | { | |
3079 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) | |
3080 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) | |
3081 | return false; | |
3082 | ||
3083 | tree_code code; | |
3084 | int multi_step_cvt = 0; | |
3085 | auto_vec <tree, 8> interm_types; | |
3086 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, | |
3087 | &code, &multi_step_cvt, | |
3088 | &interm_types) | |
3089 | || multi_step_cvt) | |
3090 | return false; | |
3091 | ||
3092 | *convert_code = code; | |
3093 | return true; | |
3094 | } | |
5ce9450f | 3095 | |
ebfd146a IR |
3096 | /* Function vectorizable_call. |
3097 | ||
538dd0b7 | 3098 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 3099 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
3100 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3101 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3102 | ||
3103 | static bool | |
1eede195 RS |
3104 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, |
3105 | stmt_vec_info *vec_stmt, slp_tree slp_node, | |
3106 | stmt_vector_for_cost *cost_vec) | |
ebfd146a | 3107 | { |
538dd0b7 | 3108 | gcall *stmt; |
ebfd146a IR |
3109 | tree vec_dest; |
3110 | tree scalar_dest; | |
0267732b | 3111 | tree op; |
ebfd146a | 3112 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
538dd0b7 | 3113 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a | 3114 | tree vectype_out, vectype_in; |
c7bda0f4 RS |
3115 | poly_uint64 nunits_in; |
3116 | poly_uint64 nunits_out; | |
ebfd146a | 3117 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
190c2236 | 3118 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 3119 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 3120 | tree fndecl, new_temp, rhs_type; |
2c58d42c RS |
3121 | enum vect_def_type dt[4] |
3122 | = { vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type, | |
3123 | vect_unknown_def_type }; | |
3124 | int ndts = ARRAY_SIZE (dt); | |
ebfd146a | 3125 | int ncopies, j; |
2c58d42c RS |
3126 | auto_vec<tree, 8> vargs; |
3127 | auto_vec<tree, 8> orig_vargs; | |
ebfd146a IR |
3128 | enum { NARROW, NONE, WIDEN } modifier; |
3129 | size_t i, nargs; | |
9d5e7640 | 3130 | tree lhs; |
ebfd146a | 3131 | |
190c2236 | 3132 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3133 | return false; |
3134 | ||
66c16fd9 RB |
3135 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3136 | && ! vec_stmt) | |
ebfd146a IR |
3137 | return false; |
3138 | ||
538dd0b7 DM |
3139 | /* Is GS a vectorizable call? */ |
3140 | stmt = dyn_cast <gcall *> (gs); | |
3141 | if (!stmt) | |
ebfd146a IR |
3142 | return false; |
3143 | ||
5ce9450f | 3144 | if (gimple_call_internal_p (stmt) |
bfaa08b7 | 3145 | && (internal_load_fn_p (gimple_call_internal_fn (stmt)) |
f307441a | 3146 | || internal_store_fn_p (gimple_call_internal_fn (stmt)))) |
c3a8f964 RS |
3147 | /* Handled by vectorizable_load and vectorizable_store. */ |
3148 | return false; | |
5ce9450f | 3149 | |
0136f8f0 AH |
3150 | if (gimple_call_lhs (stmt) == NULL_TREE |
3151 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
3152 | return false; |
3153 | ||
0136f8f0 | 3154 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 3155 | |
b690cc0f RG |
3156 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3157 | ||
ebfd146a IR |
3158 | /* Process function arguments. */ |
3159 | rhs_type = NULL_TREE; | |
b690cc0f | 3160 | vectype_in = NULL_TREE; |
ebfd146a IR |
3161 | nargs = gimple_call_num_args (stmt); |
3162 | ||
1b1562a5 MM |
3163 | /* Bail out if the function has more than three arguments, we do not have |
3164 | interesting builtin functions to vectorize with more than two arguments | |
3165 | except for fma. No arguments is also not good. */ | |
2c58d42c | 3166 | if (nargs == 0 || nargs > 4) |
ebfd146a IR |
3167 | return false; |
3168 | ||
74bf76ed | 3169 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2c58d42c RS |
3170 | combined_fn cfn = gimple_call_combined_fn (stmt); |
3171 | if (cfn == CFN_GOMP_SIMD_LANE) | |
74bf76ed JJ |
3172 | { |
3173 | nargs = 0; | |
3174 | rhs_type = unsigned_type_node; | |
3175 | } | |
3176 | ||
2c58d42c RS |
3177 | int mask_opno = -1; |
3178 | if (internal_fn_p (cfn)) | |
3179 | mask_opno = internal_fn_mask_index (as_internal_fn (cfn)); | |
3180 | ||
ebfd146a IR |
3181 | for (i = 0; i < nargs; i++) |
3182 | { | |
b690cc0f RG |
3183 | tree opvectype; |
3184 | ||
ebfd146a | 3185 | op = gimple_call_arg (stmt, i); |
2c58d42c RS |
3186 | if (!vect_is_simple_use (op, vinfo, &dt[i], &opvectype)) |
3187 | { | |
3188 | if (dump_enabled_p ()) | |
3189 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3190 | "use not simple.\n"); | |
3191 | return false; | |
3192 | } | |
3193 | ||
3194 | /* Skip the mask argument to an internal function. This operand | |
3195 | has been converted via a pattern if necessary. */ | |
3196 | if ((int) i == mask_opno) | |
3197 | continue; | |
ebfd146a IR |
3198 | |
3199 | /* We can only handle calls with arguments of the same type. */ | |
3200 | if (rhs_type | |
8533c9d8 | 3201 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 3202 | { |
73fbfcad | 3203 | if (dump_enabled_p ()) |
78c60e3d | 3204 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3205 | "argument types differ.\n"); |
ebfd146a IR |
3206 | return false; |
3207 | } | |
b690cc0f RG |
3208 | if (!rhs_type) |
3209 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 3210 | |
b690cc0f RG |
3211 | if (!vectype_in) |
3212 | vectype_in = opvectype; | |
3213 | else if (opvectype | |
3214 | && opvectype != vectype_in) | |
3215 | { | |
73fbfcad | 3216 | if (dump_enabled_p ()) |
78c60e3d | 3217 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3218 | "argument vector types differ.\n"); |
b690cc0f RG |
3219 | return false; |
3220 | } | |
3221 | } | |
3222 | /* If all arguments are external or constant defs use a vector type with | |
3223 | the same size as the output vector type. */ | |
ebfd146a | 3224 | if (!vectype_in) |
b690cc0f | 3225 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
3226 | if (vec_stmt) |
3227 | gcc_assert (vectype_in); | |
3228 | if (!vectype_in) | |
3229 | { | |
73fbfcad | 3230 | if (dump_enabled_p ()) |
7d8930a0 | 3231 | { |
78c60e3d SS |
3232 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3233 | "no vectype for scalar type "); | |
3234 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 3235 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
3236 | } |
3237 | ||
3238 | return false; | |
3239 | } | |
ebfd146a IR |
3240 | |
3241 | /* FORNOW */ | |
b690cc0f RG |
3242 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
3243 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
c7bda0f4 | 3244 | if (known_eq (nunits_in * 2, nunits_out)) |
ebfd146a | 3245 | modifier = NARROW; |
c7bda0f4 | 3246 | else if (known_eq (nunits_out, nunits_in)) |
ebfd146a | 3247 | modifier = NONE; |
c7bda0f4 | 3248 | else if (known_eq (nunits_out * 2, nunits_in)) |
ebfd146a IR |
3249 | modifier = WIDEN; |
3250 | else | |
3251 | return false; | |
3252 | ||
70439f0d RS |
3253 | /* We only handle functions that do not read or clobber memory. */ |
3254 | if (gimple_vuse (stmt)) | |
3255 | { | |
3256 | if (dump_enabled_p ()) | |
3257 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3258 | "function reads from or writes to memory.\n"); | |
3259 | return false; | |
3260 | } | |
3261 | ||
ebfd146a IR |
3262 | /* For now, we only vectorize functions if a target specific builtin |
3263 | is available. TODO -- in some cases, it might be profitable to | |
3264 | insert the calls for pieces of the vector, in order to be able | |
3265 | to vectorize other operations in the loop. */ | |
70439f0d RS |
3266 | fndecl = NULL_TREE; |
3267 | internal_fn ifn = IFN_LAST; | |
70439f0d RS |
3268 | tree callee = gimple_call_fndecl (stmt); |
3269 | ||
3270 | /* First try using an internal function. */ | |
b1b6836e RS |
3271 | tree_code convert_code = ERROR_MARK; |
3272 | if (cfn != CFN_LAST | |
3273 | && (modifier == NONE | |
3274 | || (modifier == NARROW | |
3275 | && simple_integer_narrowing (vectype_out, vectype_in, | |
3276 | &convert_code)))) | |
70439f0d RS |
3277 | ifn = vectorizable_internal_function (cfn, callee, vectype_out, |
3278 | vectype_in); | |
3279 | ||
3280 | /* If that fails, try asking for a target-specific built-in function. */ | |
3281 | if (ifn == IFN_LAST) | |
3282 | { | |
3283 | if (cfn != CFN_LAST) | |
3284 | fndecl = targetm.vectorize.builtin_vectorized_function | |
3285 | (cfn, vectype_out, vectype_in); | |
7672aa9b | 3286 | else if (callee) |
70439f0d RS |
3287 | fndecl = targetm.vectorize.builtin_md_vectorized_function |
3288 | (callee, vectype_out, vectype_in); | |
3289 | } | |
3290 | ||
3291 | if (ifn == IFN_LAST && !fndecl) | |
ebfd146a | 3292 | { |
70439f0d | 3293 | if (cfn == CFN_GOMP_SIMD_LANE |
74bf76ed JJ |
3294 | && !slp_node |
3295 | && loop_vinfo | |
3296 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
3297 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
3298 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
3299 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
3300 | { | |
3301 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
3302 | { 0, 1, 2, ... vf - 1 } vector. */ | |
3303 | gcc_assert (nargs == 0); | |
3304 | } | |
37b14185 RB |
3305 | else if (modifier == NONE |
3306 | && (gimple_call_builtin_p (stmt, BUILT_IN_BSWAP16) | |
3307 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP32) | |
3308 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP64))) | |
3309 | return vectorizable_bswap (stmt, gsi, vec_stmt, slp_node, | |
68435eb2 | 3310 | vectype_in, dt, cost_vec); |
74bf76ed JJ |
3311 | else |
3312 | { | |
3313 | if (dump_enabled_p ()) | |
3314 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 3315 | "function is not vectorizable.\n"); |
74bf76ed JJ |
3316 | return false; |
3317 | } | |
ebfd146a IR |
3318 | } |
3319 | ||
fce57248 | 3320 | if (slp_node) |
190c2236 | 3321 | ncopies = 1; |
b1b6836e | 3322 | else if (modifier == NARROW && ifn == IFN_LAST) |
e8f142e2 | 3323 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
ebfd146a | 3324 | else |
e8f142e2 | 3325 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
ebfd146a IR |
3326 | |
3327 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3328 | needs to be generated. */ | |
3329 | gcc_assert (ncopies >= 1); | |
3330 | ||
ed623edb | 3331 | vec_loop_masks *masks = (loop_vinfo ? &LOOP_VINFO_MASKS (loop_vinfo) : NULL); |
ebfd146a IR |
3332 | if (!vec_stmt) /* transformation not required. */ |
3333 | { | |
3334 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
adac3a68 | 3335 | DUMP_VECT_SCOPE ("vectorizable_call"); |
68435eb2 RB |
3336 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
3337 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) | |
3338 | record_stmt_cost (cost_vec, ncopies / 2, | |
3339 | vec_promote_demote, stmt_info, 0, vect_body); | |
b1b6836e | 3340 | |
2c58d42c RS |
3341 | if (loop_vinfo && mask_opno >= 0) |
3342 | { | |
3343 | unsigned int nvectors = (slp_node | |
3344 | ? SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node) | |
3345 | : ncopies); | |
3346 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype_out); | |
3347 | } | |
ebfd146a IR |
3348 | return true; |
3349 | } | |
3350 | ||
67b8dbac | 3351 | /* Transform. */ |
ebfd146a | 3352 | |
73fbfcad | 3353 | if (dump_enabled_p ()) |
e645e942 | 3354 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
3355 | |
3356 | /* Handle def. */ | |
3357 | scalar_dest = gimple_call_lhs (stmt); | |
3358 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
3359 | ||
2c58d42c RS |
3360 | bool masked_loop_p = loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo); |
3361 | ||
e1bd7296 | 3362 | stmt_vec_info new_stmt_info = NULL; |
ebfd146a | 3363 | prev_stmt_info = NULL; |
b1b6836e | 3364 | if (modifier == NONE || ifn != IFN_LAST) |
ebfd146a | 3365 | { |
b1b6836e | 3366 | tree prev_res = NULL_TREE; |
2c58d42c RS |
3367 | vargs.safe_grow (nargs); |
3368 | orig_vargs.safe_grow (nargs); | |
ebfd146a IR |
3369 | for (j = 0; j < ncopies; ++j) |
3370 | { | |
3371 | /* Build argument list for the vectorized call. */ | |
190c2236 JJ |
3372 | if (slp_node) |
3373 | { | |
ef062b13 | 3374 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 3375 | vec<tree> vec_oprnds0; |
190c2236 JJ |
3376 | |
3377 | for (i = 0; i < nargs; i++) | |
2c58d42c | 3378 | vargs[i] = gimple_call_arg (stmt, i); |
306b0c92 | 3379 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 3380 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
3381 | |
3382 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3383 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
3384 | { |
3385 | size_t k; | |
3386 | for (k = 0; k < nargs; k++) | |
3387 | { | |
37b5ec8f | 3388 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 3389 | vargs[k] = vec_oprndsk[i]; |
190c2236 | 3390 | } |
b1b6836e RS |
3391 | if (modifier == NARROW) |
3392 | { | |
2c58d42c RS |
3393 | /* We don't define any narrowing conditional functions |
3394 | at present. */ | |
3395 | gcc_assert (mask_opno < 0); | |
b1b6836e | 3396 | tree half_res = make_ssa_name (vectype_in); |
a844293d RS |
3397 | gcall *call |
3398 | = gimple_build_call_internal_vec (ifn, vargs); | |
3399 | gimple_call_set_lhs (call, half_res); | |
3400 | gimple_call_set_nothrow (call, true); | |
e1bd7296 RS |
3401 | new_stmt_info |
3402 | = vect_finish_stmt_generation (stmt, call, gsi); | |
b1b6836e RS |
3403 | if ((i & 1) == 0) |
3404 | { | |
3405 | prev_res = half_res; | |
3406 | continue; | |
3407 | } | |
3408 | new_temp = make_ssa_name (vec_dest); | |
e1bd7296 RS |
3409 | gimple *new_stmt |
3410 | = gimple_build_assign (new_temp, convert_code, | |
3411 | prev_res, half_res); | |
3412 | new_stmt_info | |
3413 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b1b6836e | 3414 | } |
70439f0d | 3415 | else |
b1b6836e | 3416 | { |
2c58d42c RS |
3417 | if (mask_opno >= 0 && masked_loop_p) |
3418 | { | |
3419 | unsigned int vec_num = vec_oprnds0.length (); | |
3420 | /* Always true for SLP. */ | |
3421 | gcc_assert (ncopies == 1); | |
3422 | tree mask = vect_get_loop_mask (gsi, masks, vec_num, | |
3423 | vectype_out, i); | |
3424 | vargs[mask_opno] = prepare_load_store_mask | |
3425 | (TREE_TYPE (mask), mask, vargs[mask_opno], gsi); | |
3426 | } | |
3427 | ||
a844293d | 3428 | gcall *call; |
b1b6836e | 3429 | if (ifn != IFN_LAST) |
a844293d | 3430 | call = gimple_build_call_internal_vec (ifn, vargs); |
b1b6836e | 3431 | else |
a844293d RS |
3432 | call = gimple_build_call_vec (fndecl, vargs); |
3433 | new_temp = make_ssa_name (vec_dest, call); | |
3434 | gimple_call_set_lhs (call, new_temp); | |
3435 | gimple_call_set_nothrow (call, true); | |
e1bd7296 RS |
3436 | new_stmt_info |
3437 | = vect_finish_stmt_generation (stmt, call, gsi); | |
b1b6836e | 3438 | } |
e1bd7296 | 3439 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
190c2236 JJ |
3440 | } |
3441 | ||
3442 | for (i = 0; i < nargs; i++) | |
3443 | { | |
37b5ec8f | 3444 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3445 | vec_oprndsi.release (); |
190c2236 | 3446 | } |
190c2236 JJ |
3447 | continue; |
3448 | } | |
3449 | ||
ebfd146a IR |
3450 | for (i = 0; i < nargs; i++) |
3451 | { | |
3452 | op = gimple_call_arg (stmt, i); | |
3453 | if (j == 0) | |
3454 | vec_oprnd0 | |
81c40241 | 3455 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 3456 | else |
2c58d42c RS |
3457 | vec_oprnd0 |
3458 | = vect_get_vec_def_for_stmt_copy (dt[i], orig_vargs[i]); | |
3459 | ||
3460 | orig_vargs[i] = vargs[i] = vec_oprnd0; | |
3461 | } | |
ebfd146a | 3462 | |
2c58d42c RS |
3463 | if (mask_opno >= 0 && masked_loop_p) |
3464 | { | |
3465 | tree mask = vect_get_loop_mask (gsi, masks, ncopies, | |
3466 | vectype_out, j); | |
3467 | vargs[mask_opno] | |
3468 | = prepare_load_store_mask (TREE_TYPE (mask), mask, | |
3469 | vargs[mask_opno], gsi); | |
ebfd146a IR |
3470 | } |
3471 | ||
2c58d42c | 3472 | if (cfn == CFN_GOMP_SIMD_LANE) |
74bf76ed | 3473 | { |
c7bda0f4 | 3474 | tree cst = build_index_vector (vectype_out, j * nunits_out, 1); |
74bf76ed | 3475 | tree new_var |
0e22bb5a | 3476 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 3477 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 3478 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 3479 | new_temp = make_ssa_name (vec_dest); |
e1bd7296 RS |
3480 | gimple *new_stmt = gimple_build_assign (new_temp, new_var); |
3481 | new_stmt_info | |
3482 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
74bf76ed | 3483 | } |
b1b6836e RS |
3484 | else if (modifier == NARROW) |
3485 | { | |
2c58d42c RS |
3486 | /* We don't define any narrowing conditional functions at |
3487 | present. */ | |
3488 | gcc_assert (mask_opno < 0); | |
b1b6836e | 3489 | tree half_res = make_ssa_name (vectype_in); |
a844293d RS |
3490 | gcall *call = gimple_build_call_internal_vec (ifn, vargs); |
3491 | gimple_call_set_lhs (call, half_res); | |
3492 | gimple_call_set_nothrow (call, true); | |
e1bd7296 | 3493 | new_stmt_info = vect_finish_stmt_generation (stmt, call, gsi); |
b1b6836e RS |
3494 | if ((j & 1) == 0) |
3495 | { | |
3496 | prev_res = half_res; | |
3497 | continue; | |
3498 | } | |
3499 | new_temp = make_ssa_name (vec_dest); | |
e1bd7296 RS |
3500 | gassign *new_stmt = gimple_build_assign (new_temp, convert_code, |
3501 | prev_res, half_res); | |
3502 | new_stmt_info | |
3503 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b1b6836e | 3504 | } |
74bf76ed JJ |
3505 | else |
3506 | { | |
a844293d | 3507 | gcall *call; |
70439f0d | 3508 | if (ifn != IFN_LAST) |
a844293d | 3509 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3510 | else |
a844293d | 3511 | call = gimple_build_call_vec (fndecl, vargs); |
e1bd7296 | 3512 | new_temp = make_ssa_name (vec_dest, call); |
a844293d RS |
3513 | gimple_call_set_lhs (call, new_temp); |
3514 | gimple_call_set_nothrow (call, true); | |
e1bd7296 | 3515 | new_stmt_info = vect_finish_stmt_generation (stmt, call, gsi); |
74bf76ed | 3516 | } |
ebfd146a | 3517 | |
b1b6836e | 3518 | if (j == (modifier == NARROW ? 1 : 0)) |
e1bd7296 | 3519 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
ebfd146a | 3520 | else |
e1bd7296 | 3521 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
ebfd146a | 3522 | |
e1bd7296 | 3523 | prev_stmt_info = new_stmt_info; |
ebfd146a | 3524 | } |
b1b6836e RS |
3525 | } |
3526 | else if (modifier == NARROW) | |
3527 | { | |
2c58d42c RS |
3528 | /* We don't define any narrowing conditional functions at present. */ |
3529 | gcc_assert (mask_opno < 0); | |
ebfd146a IR |
3530 | for (j = 0; j < ncopies; ++j) |
3531 | { | |
3532 | /* Build argument list for the vectorized call. */ | |
3533 | if (j == 0) | |
9771b263 | 3534 | vargs.create (nargs * 2); |
ebfd146a | 3535 | else |
9771b263 | 3536 | vargs.truncate (0); |
ebfd146a | 3537 | |
190c2236 JJ |
3538 | if (slp_node) |
3539 | { | |
ef062b13 | 3540 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 3541 | vec<tree> vec_oprnds0; |
190c2236 JJ |
3542 | |
3543 | for (i = 0; i < nargs; i++) | |
9771b263 | 3544 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 3545 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 3546 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
3547 | |
3548 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3549 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
3550 | { |
3551 | size_t k; | |
9771b263 | 3552 | vargs.truncate (0); |
190c2236 JJ |
3553 | for (k = 0; k < nargs; k++) |
3554 | { | |
37b5ec8f | 3555 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
3556 | vargs.quick_push (vec_oprndsk[i]); |
3557 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 | 3558 | } |
a844293d | 3559 | gcall *call; |
70439f0d | 3560 | if (ifn != IFN_LAST) |
a844293d | 3561 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3562 | else |
a844293d RS |
3563 | call = gimple_build_call_vec (fndecl, vargs); |
3564 | new_temp = make_ssa_name (vec_dest, call); | |
3565 | gimple_call_set_lhs (call, new_temp); | |
3566 | gimple_call_set_nothrow (call, true); | |
e1bd7296 RS |
3567 | new_stmt_info |
3568 | = vect_finish_stmt_generation (stmt, call, gsi); | |
3569 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); | |
190c2236 JJ |
3570 | } |
3571 | ||
3572 | for (i = 0; i < nargs; i++) | |
3573 | { | |
37b5ec8f | 3574 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3575 | vec_oprndsi.release (); |
190c2236 | 3576 | } |
190c2236 JJ |
3577 | continue; |
3578 | } | |
3579 | ||
ebfd146a IR |
3580 | for (i = 0; i < nargs; i++) |
3581 | { | |
3582 | op = gimple_call_arg (stmt, i); | |
3583 | if (j == 0) | |
3584 | { | |
3585 | vec_oprnd0 | |
81c40241 | 3586 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 3587 | vec_oprnd1 |
63827fb8 | 3588 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3589 | } |
3590 | else | |
3591 | { | |
e1bd7296 RS |
3592 | vec_oprnd1 = gimple_call_arg (new_stmt_info->stmt, |
3593 | 2 * i + 1); | |
ebfd146a | 3594 | vec_oprnd0 |
63827fb8 | 3595 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 3596 | vec_oprnd1 |
63827fb8 | 3597 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3598 | } |
3599 | ||
9771b263 DN |
3600 | vargs.quick_push (vec_oprnd0); |
3601 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
3602 | } |
3603 | ||
e1bd7296 | 3604 | gcall *new_stmt = gimple_build_call_vec (fndecl, vargs); |
ebfd146a IR |
3605 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3606 | gimple_call_set_lhs (new_stmt, new_temp); | |
e1bd7296 | 3607 | new_stmt_info = vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
3608 | |
3609 | if (j == 0) | |
e1bd7296 | 3610 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt_info; |
ebfd146a | 3611 | else |
e1bd7296 | 3612 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
ebfd146a | 3613 | |
e1bd7296 | 3614 | prev_stmt_info = new_stmt_info; |
ebfd146a IR |
3615 | } |
3616 | ||
3617 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a | 3618 | } |
b1b6836e RS |
3619 | else |
3620 | /* No current target implements this case. */ | |
3621 | return false; | |
ebfd146a | 3622 | |
9771b263 | 3623 | vargs.release (); |
ebfd146a | 3624 | |
ebfd146a IR |
3625 | /* The call in STMT might prevent it from being removed in dce. |
3626 | We however cannot remove it here, due to the way the ssa name | |
3627 | it defines is mapped to the new definition. So just replace | |
3628 | rhs of the statement with something harmless. */ | |
3629 | ||
dd34c087 JJ |
3630 | if (slp_node) |
3631 | return true; | |
3632 | ||
9d5e7640 | 3633 | if (is_pattern_stmt_p (stmt_info)) |
10681ce8 | 3634 | stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); |
ed7b8123 | 3635 | lhs = gimple_get_lhs (stmt_info->stmt); |
3cc2fa2a | 3636 | |
e1bd7296 RS |
3637 | gassign *new_stmt |
3638 | = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
ebfd146a | 3639 | set_vinfo_for_stmt (new_stmt, stmt_info); |
ed7b8123 | 3640 | set_vinfo_for_stmt (stmt_info->stmt, NULL); |
ebfd146a IR |
3641 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
3642 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
3643 | |
3644 | return true; | |
3645 | } | |
3646 | ||
3647 | ||
0136f8f0 AH |
3648 | struct simd_call_arg_info |
3649 | { | |
3650 | tree vectype; | |
3651 | tree op; | |
0136f8f0 | 3652 | HOST_WIDE_INT linear_step; |
34e82342 | 3653 | enum vect_def_type dt; |
0136f8f0 | 3654 | unsigned int align; |
17b658af | 3655 | bool simd_lane_linear; |
0136f8f0 AH |
3656 | }; |
3657 | ||
17b658af JJ |
3658 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
3659 | is linear within simd lane (but not within whole loop), note it in | |
3660 | *ARGINFO. */ | |
3661 | ||
3662 | static void | |
3663 | vect_simd_lane_linear (tree op, struct loop *loop, | |
3664 | struct simd_call_arg_info *arginfo) | |
3665 | { | |
355fe088 | 3666 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
3667 | |
3668 | if (!is_gimple_assign (def_stmt) | |
3669 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
3670 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
3671 | return; | |
3672 | ||
3673 | tree base = gimple_assign_rhs1 (def_stmt); | |
3674 | HOST_WIDE_INT linear_step = 0; | |
3675 | tree v = gimple_assign_rhs2 (def_stmt); | |
3676 | while (TREE_CODE (v) == SSA_NAME) | |
3677 | { | |
3678 | tree t; | |
3679 | def_stmt = SSA_NAME_DEF_STMT (v); | |
3680 | if (is_gimple_assign (def_stmt)) | |
3681 | switch (gimple_assign_rhs_code (def_stmt)) | |
3682 | { | |
3683 | case PLUS_EXPR: | |
3684 | t = gimple_assign_rhs2 (def_stmt); | |
3685 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
3686 | return; | |
3687 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
3688 | v = gimple_assign_rhs1 (def_stmt); | |
3689 | continue; | |
3690 | case MULT_EXPR: | |
3691 | t = gimple_assign_rhs2 (def_stmt); | |
3692 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
3693 | return; | |
3694 | linear_step = tree_to_shwi (t); | |
3695 | v = gimple_assign_rhs1 (def_stmt); | |
3696 | continue; | |
3697 | CASE_CONVERT: | |
3698 | t = gimple_assign_rhs1 (def_stmt); | |
3699 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
3700 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
3701 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
3702 | return; | |
3703 | if (!linear_step) | |
3704 | linear_step = 1; | |
3705 | v = t; | |
3706 | continue; | |
3707 | default: | |
3708 | return; | |
3709 | } | |
8e4284d0 | 3710 | else if (gimple_call_internal_p (def_stmt, IFN_GOMP_SIMD_LANE) |
17b658af JJ |
3711 | && loop->simduid |
3712 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
3713 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
3714 | == loop->simduid)) | |
3715 | { | |
3716 | if (!linear_step) | |
3717 | linear_step = 1; | |
3718 | arginfo->linear_step = linear_step; | |
3719 | arginfo->op = base; | |
3720 | arginfo->simd_lane_linear = true; | |
3721 | return; | |
3722 | } | |
3723 | } | |
3724 | } | |
3725 | ||
cf1b2ba4 RS |
3726 | /* Return the number of elements in vector type VECTYPE, which is associated |
3727 | with a SIMD clone. At present these vectors always have a constant | |
3728 | length. */ | |
3729 | ||
3730 | static unsigned HOST_WIDE_INT | |
3731 | simd_clone_subparts (tree vectype) | |
3732 | { | |
928686b1 | 3733 | return TYPE_VECTOR_SUBPARTS (vectype).to_constant (); |
cf1b2ba4 RS |
3734 | } |
3735 | ||
0136f8f0 AH |
3736 | /* Function vectorizable_simd_clone_call. |
3737 | ||
3738 | Check if STMT performs a function call that can be vectorized | |
3739 | by calling a simd clone of the function. | |
3740 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3741 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3742 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3743 | ||
3744 | static bool | |
355fe088 | 3745 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
1eede195 | 3746 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
68435eb2 | 3747 | stmt_vector_for_cost *) |
0136f8f0 AH |
3748 | { |
3749 | tree vec_dest; | |
3750 | tree scalar_dest; | |
3751 | tree op, type; | |
3752 | tree vec_oprnd0 = NULL_TREE; | |
3753 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
3754 | tree vectype; | |
3755 | unsigned int nunits; | |
3756 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3757 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 3758 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 3759 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 3760 | tree fndecl, new_temp; |
0136f8f0 | 3761 | int ncopies, j; |
00426f9a | 3762 | auto_vec<simd_call_arg_info> arginfo; |
0136f8f0 AH |
3763 | vec<tree> vargs = vNULL; |
3764 | size_t i, nargs; | |
3765 | tree lhs, rtype, ratype; | |
e7a74006 | 3766 | vec<constructor_elt, va_gc> *ret_ctor_elts = NULL; |
0136f8f0 AH |
3767 | |
3768 | /* Is STMT a vectorizable call? */ | |
3769 | if (!is_gimple_call (stmt)) | |
3770 | return false; | |
3771 | ||
3772 | fndecl = gimple_call_fndecl (stmt); | |
3773 | if (fndecl == NULL_TREE) | |
3774 | return false; | |
3775 | ||
d52f5295 | 3776 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
3777 | if (node == NULL || node->simd_clones == NULL) |
3778 | return false; | |
3779 | ||
3780 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
3781 | return false; | |
3782 | ||
66c16fd9 RB |
3783 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3784 | && ! vec_stmt) | |
0136f8f0 AH |
3785 | return false; |
3786 | ||
3787 | if (gimple_call_lhs (stmt) | |
3788 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
3789 | return false; | |
3790 | ||
3791 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
3792 | ||
3793 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3794 | ||
3795 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
3796 | return false; | |
3797 | ||
3798 | /* FORNOW */ | |
fce57248 | 3799 | if (slp_node) |
0136f8f0 AH |
3800 | return false; |
3801 | ||
3802 | /* Process function arguments. */ | |
3803 | nargs = gimple_call_num_args (stmt); | |
3804 | ||
3805 | /* Bail out if the function has zero arguments. */ | |
3806 | if (nargs == 0) | |
3807 | return false; | |
3808 | ||
00426f9a | 3809 | arginfo.reserve (nargs, true); |
0136f8f0 AH |
3810 | |
3811 | for (i = 0; i < nargs; i++) | |
3812 | { | |
3813 | simd_call_arg_info thisarginfo; | |
3814 | affine_iv iv; | |
3815 | ||
3816 | thisarginfo.linear_step = 0; | |
3817 | thisarginfo.align = 0; | |
3818 | thisarginfo.op = NULL_TREE; | |
17b658af | 3819 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
3820 | |
3821 | op = gimple_call_arg (stmt, i); | |
894dd753 | 3822 | if (!vect_is_simple_use (op, vinfo, &thisarginfo.dt, |
81c40241 | 3823 | &thisarginfo.vectype) |
0136f8f0 AH |
3824 | || thisarginfo.dt == vect_uninitialized_def) |
3825 | { | |
3826 | if (dump_enabled_p ()) | |
3827 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3828 | "use not simple.\n"); | |
0136f8f0 AH |
3829 | return false; |
3830 | } | |
3831 | ||
3832 | if (thisarginfo.dt == vect_constant_def | |
3833 | || thisarginfo.dt == vect_external_def) | |
3834 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
3835 | else | |
3836 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
3837 | ||
6c9e85fb JJ |
3838 | /* For linear arguments, the analyze phase should have saved |
3839 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
3840 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
3841 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
3842 | { |
3843 | gcc_assert (vec_stmt); | |
3844 | thisarginfo.linear_step | |
17b658af | 3845 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 3846 | thisarginfo.op |
17b658af JJ |
3847 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
3848 | thisarginfo.simd_lane_linear | |
3849 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
3850 | == boolean_true_node); | |
6c9e85fb JJ |
3851 | /* If loop has been peeled for alignment, we need to adjust it. */ |
3852 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
3853 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 3854 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
3855 | { |
3856 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 3857 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
3858 | tree opt = TREE_TYPE (thisarginfo.op); |
3859 | bias = fold_convert (TREE_TYPE (step), bias); | |
3860 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
3861 | thisarginfo.op | |
3862 | = fold_build2 (POINTER_TYPE_P (opt) | |
3863 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
3864 | thisarginfo.op, bias); | |
3865 | } | |
3866 | } | |
3867 | else if (!vec_stmt | |
3868 | && thisarginfo.dt != vect_constant_def | |
3869 | && thisarginfo.dt != vect_external_def | |
3870 | && loop_vinfo | |
3871 | && TREE_CODE (op) == SSA_NAME | |
3872 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
3873 | &iv, false) | |
3874 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
3875 | { |
3876 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
3877 | thisarginfo.op = iv.base; | |
3878 | } | |
3879 | else if ((thisarginfo.dt == vect_constant_def | |
3880 | || thisarginfo.dt == vect_external_def) | |
3881 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
3882 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
3883 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
3884 | linear too. */ | |
3885 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
3886 | && !thisarginfo.linear_step | |
3887 | && !vec_stmt | |
3888 | && thisarginfo.dt != vect_constant_def | |
3889 | && thisarginfo.dt != vect_external_def | |
3890 | && loop_vinfo | |
3891 | && !slp_node | |
3892 | && TREE_CODE (op) == SSA_NAME) | |
3893 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
3894 | |
3895 | arginfo.quick_push (thisarginfo); | |
3896 | } | |
3897 | ||
d9f21f6a RS |
3898 | unsigned HOST_WIDE_INT vf; |
3899 | if (!LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&vf)) | |
3900 | { | |
3901 | if (dump_enabled_p ()) | |
3902 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3903 | "not considering SIMD clones; not yet supported" | |
3904 | " for variable-width vectors.\n"); | |
3905 | return NULL; | |
3906 | } | |
3907 | ||
0136f8f0 AH |
3908 | unsigned int badness = 0; |
3909 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
3910 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
3911 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
3912 | else |
3913 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
3914 | n = n->simdclone->next_clone) | |
3915 | { | |
3916 | unsigned int this_badness = 0; | |
d9f21f6a | 3917 | if (n->simdclone->simdlen > vf |
0136f8f0 AH |
3918 | || n->simdclone->nargs != nargs) |
3919 | continue; | |
d9f21f6a RS |
3920 | if (n->simdclone->simdlen < vf) |
3921 | this_badness += (exact_log2 (vf) | |
0136f8f0 AH |
3922 | - exact_log2 (n->simdclone->simdlen)) * 1024; |
3923 | if (n->simdclone->inbranch) | |
3924 | this_badness += 2048; | |
3925 | int target_badness = targetm.simd_clone.usable (n); | |
3926 | if (target_badness < 0) | |
3927 | continue; | |
3928 | this_badness += target_badness * 512; | |
3929 | /* FORNOW: Have to add code to add the mask argument. */ | |
3930 | if (n->simdclone->inbranch) | |
3931 | continue; | |
3932 | for (i = 0; i < nargs; i++) | |
3933 | { | |
3934 | switch (n->simdclone->args[i].arg_type) | |
3935 | { | |
3936 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3937 | if (!useless_type_conversion_p | |
3938 | (n->simdclone->args[i].orig_type, | |
3939 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
3940 | i = -1; | |
3941 | else if (arginfo[i].dt == vect_constant_def | |
3942 | || arginfo[i].dt == vect_external_def | |
3943 | || arginfo[i].linear_step) | |
3944 | this_badness += 64; | |
3945 | break; | |
3946 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3947 | if (arginfo[i].dt != vect_constant_def | |
3948 | && arginfo[i].dt != vect_external_def) | |
3949 | i = -1; | |
3950 | break; | |
3951 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 3952 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
3953 | if (arginfo[i].dt == vect_constant_def |
3954 | || arginfo[i].dt == vect_external_def | |
3955 | || (arginfo[i].linear_step | |
3956 | != n->simdclone->args[i].linear_step)) | |
3957 | i = -1; | |
3958 | break; | |
3959 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
3960 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
3961 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
3962 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3963 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3964 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3965 | /* FORNOW */ |
3966 | i = -1; | |
3967 | break; | |
3968 | case SIMD_CLONE_ARG_TYPE_MASK: | |
3969 | gcc_unreachable (); | |
3970 | } | |
3971 | if (i == (size_t) -1) | |
3972 | break; | |
3973 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
3974 | { | |
3975 | i = -1; | |
3976 | break; | |
3977 | } | |
3978 | if (arginfo[i].align) | |
3979 | this_badness += (exact_log2 (arginfo[i].align) | |
3980 | - exact_log2 (n->simdclone->args[i].alignment)); | |
3981 | } | |
3982 | if (i == (size_t) -1) | |
3983 | continue; | |
3984 | if (bestn == NULL || this_badness < badness) | |
3985 | { | |
3986 | bestn = n; | |
3987 | badness = this_badness; | |
3988 | } | |
3989 | } | |
3990 | ||
3991 | if (bestn == NULL) | |
00426f9a | 3992 | return false; |
0136f8f0 AH |
3993 | |
3994 | for (i = 0; i < nargs; i++) | |
3995 | if ((arginfo[i].dt == vect_constant_def | |
3996 | || arginfo[i].dt == vect_external_def) | |
3997 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
3998 | { | |
3999 | arginfo[i].vectype | |
4000 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
4001 | i))); | |
4002 | if (arginfo[i].vectype == NULL | |
cf1b2ba4 | 4003 | || (simd_clone_subparts (arginfo[i].vectype) |
0136f8f0 | 4004 | > bestn->simdclone->simdlen)) |
00426f9a | 4005 | return false; |
0136f8f0 AH |
4006 | } |
4007 | ||
4008 | fndecl = bestn->decl; | |
4009 | nunits = bestn->simdclone->simdlen; | |
d9f21f6a | 4010 | ncopies = vf / nunits; |
0136f8f0 AH |
4011 | |
4012 | /* If the function isn't const, only allow it in simd loops where user | |
4013 | has asserted that at least nunits consecutive iterations can be | |
4014 | performed using SIMD instructions. */ | |
4015 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
4016 | && gimple_vuse (stmt)) | |
00426f9a | 4017 | return false; |
0136f8f0 AH |
4018 | |
4019 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
4020 | needs to be generated. */ | |
4021 | gcc_assert (ncopies >= 1); | |
4022 | ||
4023 | if (!vec_stmt) /* transformation not required. */ | |
4024 | { | |
6c9e85fb JJ |
4025 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
4026 | for (i = 0; i < nargs; i++) | |
7adb26f2 JJ |
4027 | if ((bestn->simdclone->args[i].arg_type |
4028 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
4029 | || (bestn->simdclone->args[i].arg_type | |
4030 | == SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP)) | |
6c9e85fb | 4031 | { |
17b658af | 4032 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
4033 | + 1); |
4034 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
4035 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
4036 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
4037 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
4038 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
4039 | tree sll = arginfo[i].simd_lane_linear |
4040 | ? boolean_true_node : boolean_false_node; | |
4041 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 4042 | } |
0136f8f0 | 4043 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
adac3a68 | 4044 | DUMP_VECT_SCOPE ("vectorizable_simd_clone_call"); |
68435eb2 | 4045 | /* vect_model_simple_cost (stmt_info, ncopies, dt, slp_node, cost_vec); */ |
0136f8f0 AH |
4046 | return true; |
4047 | } | |
4048 | ||
67b8dbac | 4049 | /* Transform. */ |
0136f8f0 AH |
4050 | |
4051 | if (dump_enabled_p ()) | |
4052 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
4053 | ||
4054 | /* Handle def. */ | |
4055 | scalar_dest = gimple_call_lhs (stmt); | |
4056 | vec_dest = NULL_TREE; | |
4057 | rtype = NULL_TREE; | |
4058 | ratype = NULL_TREE; | |
4059 | if (scalar_dest) | |
4060 | { | |
4061 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4062 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
4063 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
4064 | { | |
4065 | ratype = rtype; | |
4066 | rtype = TREE_TYPE (ratype); | |
4067 | } | |
4068 | } | |
4069 | ||
4070 | prev_stmt_info = NULL; | |
4071 | for (j = 0; j < ncopies; ++j) | |
4072 | { | |
4073 | /* Build argument list for the vectorized call. */ | |
4074 | if (j == 0) | |
4075 | vargs.create (nargs); | |
4076 | else | |
4077 | vargs.truncate (0); | |
4078 | ||
4079 | for (i = 0; i < nargs; i++) | |
4080 | { | |
4081 | unsigned int k, l, m, o; | |
4082 | tree atype; | |
4083 | op = gimple_call_arg (stmt, i); | |
4084 | switch (bestn->simdclone->args[i].arg_type) | |
4085 | { | |
4086 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
4087 | atype = bestn->simdclone->args[i].vector_type; | |
cf1b2ba4 | 4088 | o = nunits / simd_clone_subparts (atype); |
0136f8f0 AH |
4089 | for (m = j * o; m < (j + 1) * o; m++) |
4090 | { | |
cf1b2ba4 RS |
4091 | if (simd_clone_subparts (atype) |
4092 | < simd_clone_subparts (arginfo[i].vectype)) | |
0136f8f0 | 4093 | { |
73a699ae | 4094 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); |
cf1b2ba4 RS |
4095 | k = (simd_clone_subparts (arginfo[i].vectype) |
4096 | / simd_clone_subparts (atype)); | |
0136f8f0 AH |
4097 | gcc_assert ((k & (k - 1)) == 0); |
4098 | if (m == 0) | |
4099 | vec_oprnd0 | |
81c40241 | 4100 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
4101 | else |
4102 | { | |
4103 | vec_oprnd0 = arginfo[i].op; | |
4104 | if ((m & (k - 1)) == 0) | |
4105 | vec_oprnd0 | |
4106 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
4107 | vec_oprnd0); | |
4108 | } | |
4109 | arginfo[i].op = vec_oprnd0; | |
4110 | vec_oprnd0 | |
4111 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
92e29a5e | 4112 | bitsize_int (prec), |
0136f8f0 | 4113 | bitsize_int ((m & (k - 1)) * prec)); |
e1bd7296 | 4114 | gassign *new_stmt |
b731b390 | 4115 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
4116 | vec_oprnd0); |
4117 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4118 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
4119 | } | |
4120 | else | |
4121 | { | |
cf1b2ba4 RS |
4122 | k = (simd_clone_subparts (atype) |
4123 | / simd_clone_subparts (arginfo[i].vectype)); | |
0136f8f0 AH |
4124 | gcc_assert ((k & (k - 1)) == 0); |
4125 | vec<constructor_elt, va_gc> *ctor_elts; | |
4126 | if (k != 1) | |
4127 | vec_alloc (ctor_elts, k); | |
4128 | else | |
4129 | ctor_elts = NULL; | |
4130 | for (l = 0; l < k; l++) | |
4131 | { | |
4132 | if (m == 0 && l == 0) | |
4133 | vec_oprnd0 | |
81c40241 | 4134 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
4135 | else |
4136 | vec_oprnd0 | |
4137 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
4138 | arginfo[i].op); | |
4139 | arginfo[i].op = vec_oprnd0; | |
4140 | if (k == 1) | |
4141 | break; | |
4142 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
4143 | vec_oprnd0); | |
4144 | } | |
4145 | if (k == 1) | |
4146 | vargs.safe_push (vec_oprnd0); | |
4147 | else | |
4148 | { | |
4149 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
e1bd7296 | 4150 | gassign *new_stmt |
b731b390 | 4151 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
4152 | vec_oprnd0); |
4153 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4154 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
4155 | } | |
4156 | } | |
4157 | } | |
4158 | break; | |
4159 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
4160 | vargs.safe_push (op); | |
4161 | break; | |
4162 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
7adb26f2 | 4163 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
4164 | if (j == 0) |
4165 | { | |
4166 | gimple_seq stmts; | |
4167 | arginfo[i].op | |
4168 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
4169 | NULL_TREE); | |
4170 | if (stmts != NULL) | |
4171 | { | |
4172 | basic_block new_bb; | |
4173 | edge pe = loop_preheader_edge (loop); | |
4174 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
4175 | gcc_assert (!new_bb); | |
4176 | } | |
17b658af JJ |
4177 | if (arginfo[i].simd_lane_linear) |
4178 | { | |
4179 | vargs.safe_push (arginfo[i].op); | |
4180 | break; | |
4181 | } | |
b731b390 | 4182 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 4183 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
4fbeb363 | 4184 | loop_vinfo->add_stmt (new_phi); |
0136f8f0 AH |
4185 | add_phi_arg (new_phi, arginfo[i].op, |
4186 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
4187 | enum tree_code code | |
4188 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
4189 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
4190 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
4191 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
4192 | widest_int cst |
4193 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
4194 | ncopies * nunits); | |
4195 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 4196 | tree phi_arg = copy_ssa_name (op); |
e1bd7296 | 4197 | gassign *new_stmt |
0d0e4a03 | 4198 | = gimple_build_assign (phi_arg, code, phi_res, tcst); |
0136f8f0 AH |
4199 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
4200 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
4fbeb363 | 4201 | loop_vinfo->add_stmt (new_stmt); |
0136f8f0 AH |
4202 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
4203 | UNKNOWN_LOCATION); | |
4204 | arginfo[i].op = phi_res; | |
4205 | vargs.safe_push (phi_res); | |
4206 | } | |
4207 | else | |
4208 | { | |
4209 | enum tree_code code | |
4210 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
4211 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
4212 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
4213 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
4214 | widest_int cst |
4215 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
4216 | j * nunits); | |
4217 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 4218 | new_temp = make_ssa_name (TREE_TYPE (op)); |
e1bd7296 RS |
4219 | gassign *new_stmt |
4220 | = gimple_build_assign (new_temp, code, | |
4221 | arginfo[i].op, tcst); | |
0136f8f0 AH |
4222 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4223 | vargs.safe_push (new_temp); | |
4224 | } | |
4225 | break; | |
7adb26f2 JJ |
4226 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
4227 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
0136f8f0 | 4228 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
e01d41e5 JJ |
4229 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
4230 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
4231 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
4232 | default: |
4233 | gcc_unreachable (); | |
4234 | } | |
4235 | } | |
4236 | ||
e1bd7296 | 4237 | gcall *new_call = gimple_build_call_vec (fndecl, vargs); |
0136f8f0 AH |
4238 | if (vec_dest) |
4239 | { | |
cf1b2ba4 | 4240 | gcc_assert (ratype || simd_clone_subparts (rtype) == nunits); |
0136f8f0 | 4241 | if (ratype) |
b731b390 | 4242 | new_temp = create_tmp_var (ratype); |
cf1b2ba4 RS |
4243 | else if (simd_clone_subparts (vectype) |
4244 | == simd_clone_subparts (rtype)) | |
e1bd7296 | 4245 | new_temp = make_ssa_name (vec_dest, new_call); |
0136f8f0 | 4246 | else |
e1bd7296 RS |
4247 | new_temp = make_ssa_name (rtype, new_call); |
4248 | gimple_call_set_lhs (new_call, new_temp); | |
0136f8f0 | 4249 | } |
e1bd7296 RS |
4250 | stmt_vec_info new_stmt_info |
4251 | = vect_finish_stmt_generation (stmt, new_call, gsi); | |
0136f8f0 AH |
4252 | |
4253 | if (vec_dest) | |
4254 | { | |
cf1b2ba4 | 4255 | if (simd_clone_subparts (vectype) < nunits) |
0136f8f0 AH |
4256 | { |
4257 | unsigned int k, l; | |
73a699ae RS |
4258 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); |
4259 | poly_uint64 bytes = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
cf1b2ba4 | 4260 | k = nunits / simd_clone_subparts (vectype); |
0136f8f0 AH |
4261 | gcc_assert ((k & (k - 1)) == 0); |
4262 | for (l = 0; l < k; l++) | |
4263 | { | |
4264 | tree t; | |
4265 | if (ratype) | |
4266 | { | |
4267 | t = build_fold_addr_expr (new_temp); | |
4268 | t = build2 (MEM_REF, vectype, t, | |
73a699ae | 4269 | build_int_cst (TREE_TYPE (t), l * bytes)); |
0136f8f0 AH |
4270 | } |
4271 | else | |
4272 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
92e29a5e | 4273 | bitsize_int (prec), bitsize_int (l * prec)); |
e1bd7296 | 4274 | gimple *new_stmt |
b731b390 | 4275 | = gimple_build_assign (make_ssa_name (vectype), t); |
e1bd7296 RS |
4276 | new_stmt_info |
4277 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4278 | ||
0136f8f0 | 4279 | if (j == 0 && l == 0) |
e1bd7296 RS |
4280 | STMT_VINFO_VEC_STMT (stmt_info) |
4281 | = *vec_stmt = new_stmt_info; | |
0136f8f0 | 4282 | else |
e1bd7296 | 4283 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
0136f8f0 | 4284 | |
e1bd7296 | 4285 | prev_stmt_info = new_stmt_info; |
0136f8f0 AH |
4286 | } |
4287 | ||
4288 | if (ratype) | |
3ba4ff41 | 4289 | vect_clobber_variable (stmt, gsi, new_temp); |
0136f8f0 AH |
4290 | continue; |
4291 | } | |
cf1b2ba4 | 4292 | else if (simd_clone_subparts (vectype) > nunits) |
0136f8f0 | 4293 | { |
cf1b2ba4 RS |
4294 | unsigned int k = (simd_clone_subparts (vectype) |
4295 | / simd_clone_subparts (rtype)); | |
0136f8f0 AH |
4296 | gcc_assert ((k & (k - 1)) == 0); |
4297 | if ((j & (k - 1)) == 0) | |
4298 | vec_alloc (ret_ctor_elts, k); | |
4299 | if (ratype) | |
4300 | { | |
cf1b2ba4 | 4301 | unsigned int m, o = nunits / simd_clone_subparts (rtype); |
0136f8f0 AH |
4302 | for (m = 0; m < o; m++) |
4303 | { | |
4304 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
4305 | size_int (m), NULL_TREE, NULL_TREE); | |
e1bd7296 | 4306 | gimple *new_stmt |
b731b390 | 4307 | = gimple_build_assign (make_ssa_name (rtype), tem); |
e1bd7296 RS |
4308 | new_stmt_info |
4309 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
0136f8f0 AH |
4310 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, |
4311 | gimple_assign_lhs (new_stmt)); | |
4312 | } | |
3ba4ff41 | 4313 | vect_clobber_variable (stmt, gsi, new_temp); |
0136f8f0 AH |
4314 | } |
4315 | else | |
4316 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
4317 | if ((j & (k - 1)) != k - 1) | |
4318 | continue; | |
4319 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
e1bd7296 | 4320 | gimple *new_stmt |
b731b390 | 4321 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
e1bd7296 RS |
4322 | new_stmt_info |
4323 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
0136f8f0 AH |
4324 | |
4325 | if ((unsigned) j == k - 1) | |
e1bd7296 | 4326 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
0136f8f0 | 4327 | else |
e1bd7296 | 4328 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
0136f8f0 | 4329 | |
e1bd7296 | 4330 | prev_stmt_info = new_stmt_info; |
0136f8f0 AH |
4331 | continue; |
4332 | } | |
4333 | else if (ratype) | |
4334 | { | |
4335 | tree t = build_fold_addr_expr (new_temp); | |
4336 | t = build2 (MEM_REF, vectype, t, | |
4337 | build_int_cst (TREE_TYPE (t), 0)); | |
e1bd7296 | 4338 | gimple *new_stmt |
b731b390 | 4339 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
e1bd7296 RS |
4340 | new_stmt_info |
4341 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3ba4ff41 | 4342 | vect_clobber_variable (stmt, gsi, new_temp); |
0136f8f0 AH |
4343 | } |
4344 | } | |
4345 | ||
4346 | if (j == 0) | |
e1bd7296 | 4347 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
0136f8f0 | 4348 | else |
e1bd7296 | 4349 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
0136f8f0 | 4350 | |
e1bd7296 | 4351 | prev_stmt_info = new_stmt_info; |
0136f8f0 AH |
4352 | } |
4353 | ||
4354 | vargs.release (); | |
4355 | ||
4356 | /* The call in STMT might prevent it from being removed in dce. | |
4357 | We however cannot remove it here, due to the way the ssa name | |
4358 | it defines is mapped to the new definition. So just replace | |
4359 | rhs of the statement with something harmless. */ | |
4360 | ||
4361 | if (slp_node) | |
4362 | return true; | |
4363 | ||
e1bd7296 | 4364 | gimple *new_stmt; |
0136f8f0 AH |
4365 | if (scalar_dest) |
4366 | { | |
4367 | type = TREE_TYPE (scalar_dest); | |
4368 | if (is_pattern_stmt_p (stmt_info)) | |
10681ce8 | 4369 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)->stmt); |
0136f8f0 AH |
4370 | else |
4371 | lhs = gimple_call_lhs (stmt); | |
4372 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
4373 | } | |
4374 | else | |
4375 | new_stmt = gimple_build_nop (); | |
4376 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
4377 | set_vinfo_for_stmt (stmt, NULL); | |
4378 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 4379 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
4380 | unlink_stmt_vdef (stmt); |
4381 | ||
4382 | return true; | |
4383 | } | |
4384 | ||
4385 | ||
ebfd146a IR |
4386 | /* Function vect_gen_widened_results_half |
4387 | ||
4388 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 4389 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 4390 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
4391 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
4392 | needs to be created (DECL is a function-decl of a target-builtin). | |
4393 | STMT is the original scalar stmt that we are vectorizing. */ | |
4394 | ||
355fe088 | 4395 | static gimple * |
ebfd146a IR |
4396 | vect_gen_widened_results_half (enum tree_code code, |
4397 | tree decl, | |
4398 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
4399 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 4400 | gimple *stmt) |
b8698a0f | 4401 | { |
355fe088 | 4402 | gimple *new_stmt; |
b8698a0f L |
4403 | tree new_temp; |
4404 | ||
4405 | /* Generate half of the widened result: */ | |
4406 | if (code == CALL_EXPR) | |
4407 | { | |
4408 | /* Target specific support */ | |
ebfd146a IR |
4409 | if (op_type == binary_op) |
4410 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
4411 | else | |
4412 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
4413 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4414 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
4415 | } |
4416 | else | |
ebfd146a | 4417 | { |
b8698a0f L |
4418 | /* Generic support */ |
4419 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
4420 | if (op_type != binary_op) |
4421 | vec_oprnd1 = NULL; | |
0d0e4a03 | 4422 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
4423 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4424 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 4425 | } |
ebfd146a IR |
4426 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4427 | ||
ebfd146a IR |
4428 | return new_stmt; |
4429 | } | |
4430 | ||
4a00c761 JJ |
4431 | |
4432 | /* Get vectorized definitions for loop-based vectorization. For the first | |
4433 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
4434 | scalar operand), and for the rest we get a copy with | |
4435 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
4436 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
4437 | The vectors are collected into VEC_OPRNDS. */ | |
4438 | ||
4439 | static void | |
355fe088 | 4440 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 4441 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
4442 | { |
4443 | tree vec_oprnd; | |
4444 | ||
4445 | /* Get first vector operand. */ | |
4446 | /* All the vector operands except the very first one (that is scalar oprnd) | |
4447 | are stmt copies. */ | |
4448 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 4449 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
4450 | else |
4451 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
4452 | ||
9771b263 | 4453 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
4454 | |
4455 | /* Get second vector operand. */ | |
4456 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 4457 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
4458 | |
4459 | *oprnd = vec_oprnd; | |
4460 | ||
4461 | /* For conversion in multiple steps, continue to get operands | |
4462 | recursively. */ | |
4463 | if (multi_step_cvt) | |
4464 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
4465 | } | |
4466 | ||
4467 | ||
4468 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
4469 | For multi-step conversions store the resulting vectors and call the function | |
4470 | recursively. */ | |
4471 | ||
4472 | static void | |
9771b263 | 4473 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 4474 | int multi_step_cvt, gimple *stmt, |
9771b263 | 4475 | vec<tree> vec_dsts, |
4a00c761 JJ |
4476 | gimple_stmt_iterator *gsi, |
4477 | slp_tree slp_node, enum tree_code code, | |
4478 | stmt_vec_info *prev_stmt_info) | |
4479 | { | |
4480 | unsigned int i; | |
4481 | tree vop0, vop1, new_tmp, vec_dest; | |
4a00c761 JJ |
4482 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
4483 | ||
9771b263 | 4484 | vec_dest = vec_dsts.pop (); |
4a00c761 | 4485 | |
9771b263 | 4486 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
4487 | { |
4488 | /* Create demotion operation. */ | |
9771b263 DN |
4489 | vop0 = (*vec_oprnds)[i]; |
4490 | vop1 = (*vec_oprnds)[i + 1]; | |
e1bd7296 | 4491 | gassign *new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
4492 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
4493 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
e1bd7296 RS |
4494 | stmt_vec_info new_stmt_info |
4495 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4a00c761 JJ |
4496 | |
4497 | if (multi_step_cvt) | |
4498 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 4499 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
4500 | else |
4501 | { | |
4502 | /* This is the last step of the conversion sequence. Store the | |
4503 | vectors in SLP_NODE or in vector info of the scalar statement | |
4504 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
4505 | if (slp_node) | |
e1bd7296 | 4506 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
4a00c761 | 4507 | else |
c689ce1e RB |
4508 | { |
4509 | if (!*prev_stmt_info) | |
e1bd7296 | 4510 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt_info; |
c689ce1e | 4511 | else |
e1bd7296 | 4512 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt_info; |
4a00c761 | 4513 | |
e1bd7296 | 4514 | *prev_stmt_info = new_stmt_info; |
c689ce1e | 4515 | } |
4a00c761 JJ |
4516 | } |
4517 | } | |
4518 | ||
4519 | /* For multi-step demotion operations we first generate demotion operations | |
4520 | from the source type to the intermediate types, and then combine the | |
4521 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
4522 | type. */ | |
4523 | if (multi_step_cvt) | |
4524 | { | |
4525 | /* At each level of recursion we have half of the operands we had at the | |
4526 | previous level. */ | |
9771b263 | 4527 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
4528 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
4529 | stmt, vec_dsts, gsi, slp_node, | |
4530 | VEC_PACK_TRUNC_EXPR, | |
4531 | prev_stmt_info); | |
4532 | } | |
4533 | ||
9771b263 | 4534 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4535 | } |
4536 | ||
4537 | ||
4538 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
4539 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
4540 | the resulting vectors and call the function recursively. */ | |
4541 | ||
4542 | static void | |
9771b263 DN |
4543 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
4544 | vec<tree> *vec_oprnds1, | |
355fe088 | 4545 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
4546 | gimple_stmt_iterator *gsi, |
4547 | enum tree_code code1, | |
4548 | enum tree_code code2, tree decl1, | |
4549 | tree decl2, int op_type) | |
4550 | { | |
4551 | int i; | |
4552 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 4553 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 4554 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 4555 | |
9771b263 DN |
4556 | vec_tmp.create (vec_oprnds0->length () * 2); |
4557 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
4558 | { |
4559 | if (op_type == binary_op) | |
9771b263 | 4560 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
4561 | else |
4562 | vop1 = NULL_TREE; | |
4563 | ||
4564 | /* Generate the two halves of promotion operation. */ | |
4565 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
4566 | op_type, vec_dest, gsi, stmt); | |
4567 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
4568 | op_type, vec_dest, gsi, stmt); | |
4569 | if (is_gimple_call (new_stmt1)) | |
4570 | { | |
4571 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
4572 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
4573 | } | |
4574 | else | |
4575 | { | |
4576 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
4577 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
4578 | } | |
4579 | ||
4580 | /* Store the results for the next step. */ | |
9771b263 DN |
4581 | vec_tmp.quick_push (new_tmp1); |
4582 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
4583 | } |
4584 | ||
689eaba3 | 4585 | vec_oprnds0->release (); |
4a00c761 JJ |
4586 | *vec_oprnds0 = vec_tmp; |
4587 | } | |
4588 | ||
4589 | ||
b8698a0f L |
4590 | /* Check if STMT performs a conversion operation, that can be vectorized. |
4591 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 4592 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
4593 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
4594 | ||
4595 | static bool | |
355fe088 | 4596 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
1eede195 | 4597 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
68435eb2 | 4598 | stmt_vector_for_cost *cost_vec) |
ebfd146a IR |
4599 | { |
4600 | tree vec_dest; | |
4601 | tree scalar_dest; | |
4a00c761 | 4602 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
4603 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
4604 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4605 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4606 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 4607 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
4608 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
4609 | tree new_temp; | |
ebfd146a | 4610 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 4611 | int ndts = 2; |
ebfd146a | 4612 | stmt_vec_info prev_stmt_info; |
062d5ccc RS |
4613 | poly_uint64 nunits_in; |
4614 | poly_uint64 nunits_out; | |
ebfd146a | 4615 | tree vectype_out, vectype_in; |
4a00c761 JJ |
4616 | int ncopies, i, j; |
4617 | tree lhs_type, rhs_type; | |
ebfd146a | 4618 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
4619 | vec<tree> vec_oprnds0 = vNULL; |
4620 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 4621 | tree vop0; |
4a00c761 | 4622 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4623 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 4624 | int multi_step_cvt = 0; |
6e1aa848 | 4625 | vec<tree> interm_types = vNULL; |
4a00c761 JJ |
4626 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
4627 | int op_type; | |
4a00c761 | 4628 | unsigned short fltsz; |
ebfd146a IR |
4629 | |
4630 | /* Is STMT a vectorizable conversion? */ | |
4631 | ||
4a00c761 | 4632 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4633 | return false; |
4634 | ||
66c16fd9 RB |
4635 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4636 | && ! vec_stmt) | |
ebfd146a IR |
4637 | return false; |
4638 | ||
4639 | if (!is_gimple_assign (stmt)) | |
4640 | return false; | |
4641 | ||
4642 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4643 | return false; | |
4644 | ||
4645 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
4646 | if (!CONVERT_EXPR_CODE_P (code) |
4647 | && code != FIX_TRUNC_EXPR | |
4648 | && code != FLOAT_EXPR | |
4649 | && code != WIDEN_MULT_EXPR | |
4650 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
4651 | return false; |
4652 | ||
4a00c761 JJ |
4653 | op_type = TREE_CODE_LENGTH (code); |
4654 | ||
ebfd146a | 4655 | /* Check types of lhs and rhs. */ |
b690cc0f | 4656 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 4657 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
4658 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
4659 | ||
ebfd146a IR |
4660 | op0 = gimple_assign_rhs1 (stmt); |
4661 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
4662 | |
4663 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4664 | && !((INTEGRAL_TYPE_P (lhs_type) | |
4665 | && INTEGRAL_TYPE_P (rhs_type)) | |
4666 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
4667 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
4668 | return false; | |
4669 | ||
e6f5c25d IE |
4670 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4671 | && ((INTEGRAL_TYPE_P (lhs_type) | |
2be65d9e | 4672 | && !type_has_mode_precision_p (lhs_type)) |
e6f5c25d | 4673 | || (INTEGRAL_TYPE_P (rhs_type) |
2be65d9e | 4674 | && !type_has_mode_precision_p (rhs_type)))) |
4a00c761 | 4675 | { |
73fbfcad | 4676 | if (dump_enabled_p ()) |
78c60e3d | 4677 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
4678 | "type conversion to/from bit-precision unsupported." |
4679 | "\n"); | |
4a00c761 JJ |
4680 | return false; |
4681 | } | |
4682 | ||
b690cc0f | 4683 | /* Check the operands of the operation. */ |
894dd753 | 4684 | if (!vect_is_simple_use (op0, vinfo, &dt[0], &vectype_in)) |
b690cc0f | 4685 | { |
73fbfcad | 4686 | if (dump_enabled_p ()) |
78c60e3d | 4687 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4688 | "use not simple.\n"); |
b690cc0f RG |
4689 | return false; |
4690 | } | |
4a00c761 JJ |
4691 | if (op_type == binary_op) |
4692 | { | |
4693 | bool ok; | |
4694 | ||
4695 | op1 = gimple_assign_rhs2 (stmt); | |
4696 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
4697 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
4698 | OP1. */ | |
4699 | if (CONSTANT_CLASS_P (op0)) | |
894dd753 | 4700 | ok = vect_is_simple_use (op1, vinfo, &dt[1], &vectype_in); |
4a00c761 | 4701 | else |
894dd753 | 4702 | ok = vect_is_simple_use (op1, vinfo, &dt[1]); |
4a00c761 JJ |
4703 | |
4704 | if (!ok) | |
4705 | { | |
73fbfcad | 4706 | if (dump_enabled_p ()) |
78c60e3d | 4707 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4708 | "use not simple.\n"); |
4a00c761 JJ |
4709 | return false; |
4710 | } | |
4711 | } | |
4712 | ||
b690cc0f RG |
4713 | /* If op0 is an external or constant defs use a vector type of |
4714 | the same size as the output vector type. */ | |
ebfd146a | 4715 | if (!vectype_in) |
b690cc0f | 4716 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
4717 | if (vec_stmt) |
4718 | gcc_assert (vectype_in); | |
4719 | if (!vectype_in) | |
4720 | { | |
73fbfcad | 4721 | if (dump_enabled_p ()) |
4a00c761 | 4722 | { |
78c60e3d SS |
4723 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4724 | "no vectype for scalar type "); | |
4725 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 4726 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 4727 | } |
7d8930a0 IR |
4728 | |
4729 | return false; | |
4730 | } | |
ebfd146a | 4731 | |
e6f5c25d IE |
4732 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4733 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
4734 | { | |
4735 | if (dump_enabled_p ()) | |
4736 | { | |
4737 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4738 | "can't convert between boolean and non " | |
4739 | "boolean vectors"); | |
4740 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
4741 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
4742 | } | |
4743 | ||
4744 | return false; | |
4745 | } | |
4746 | ||
b690cc0f RG |
4747 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
4748 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
062d5ccc | 4749 | if (known_eq (nunits_out, nunits_in)) |
ebfd146a | 4750 | modifier = NONE; |
062d5ccc RS |
4751 | else if (multiple_p (nunits_out, nunits_in)) |
4752 | modifier = NARROW; | |
ebfd146a | 4753 | else |
062d5ccc RS |
4754 | { |
4755 | gcc_checking_assert (multiple_p (nunits_in, nunits_out)); | |
4756 | modifier = WIDEN; | |
4757 | } | |
ebfd146a | 4758 | |
ff802fa1 IR |
4759 | /* Multiple types in SLP are handled by creating the appropriate number of |
4760 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4761 | case of SLP. */ | |
fce57248 | 4762 | if (slp_node) |
ebfd146a | 4763 | ncopies = 1; |
4a00c761 | 4764 | else if (modifier == NARROW) |
e8f142e2 | 4765 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
4a00c761 | 4766 | else |
e8f142e2 | 4767 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
b8698a0f | 4768 | |
ebfd146a IR |
4769 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
4770 | needs to be generated. */ | |
4771 | gcc_assert (ncopies >= 1); | |
4772 | ||
16d22000 RS |
4773 | bool found_mode = false; |
4774 | scalar_mode lhs_mode = SCALAR_TYPE_MODE (lhs_type); | |
4775 | scalar_mode rhs_mode = SCALAR_TYPE_MODE (rhs_type); | |
4776 | opt_scalar_mode rhs_mode_iter; | |
b397965c | 4777 | |
ebfd146a | 4778 | /* Supportable by target? */ |
4a00c761 | 4779 | switch (modifier) |
ebfd146a | 4780 | { |
4a00c761 JJ |
4781 | case NONE: |
4782 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4783 | return false; | |
4784 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
4785 | &decl1, &code1)) | |
4786 | break; | |
4787 | /* FALLTHRU */ | |
4788 | unsupported: | |
73fbfcad | 4789 | if (dump_enabled_p ()) |
78c60e3d | 4790 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4791 | "conversion not supported by target.\n"); |
ebfd146a | 4792 | return false; |
ebfd146a | 4793 | |
4a00c761 JJ |
4794 | case WIDEN: |
4795 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
4796 | &code1, &code2, &multi_step_cvt, |
4797 | &interm_types)) | |
4a00c761 JJ |
4798 | { |
4799 | /* Binary widening operation can only be supported directly by the | |
4800 | architecture. */ | |
4801 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
4802 | break; | |
4803 | } | |
4804 | ||
4805 | if (code != FLOAT_EXPR | |
b397965c | 4806 | || GET_MODE_SIZE (lhs_mode) <= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4807 | goto unsupported; |
4808 | ||
b397965c | 4809 | fltsz = GET_MODE_SIZE (lhs_mode); |
16d22000 | 4810 | FOR_EACH_2XWIDER_MODE (rhs_mode_iter, rhs_mode) |
4a00c761 | 4811 | { |
16d22000 | 4812 | rhs_mode = rhs_mode_iter.require (); |
c94843d2 RS |
4813 | if (GET_MODE_SIZE (rhs_mode) > fltsz) |
4814 | break; | |
4815 | ||
4a00c761 JJ |
4816 | cvt_type |
4817 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4818 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4819 | if (cvt_type == NULL_TREE) | |
4820 | goto unsupported; | |
4821 | ||
4822 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4823 | { | |
4824 | if (!supportable_convert_operation (code, vectype_out, | |
4825 | cvt_type, &decl1, &codecvt1)) | |
4826 | goto unsupported; | |
4827 | } | |
4828 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
4829 | cvt_type, &codecvt1, |
4830 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
4831 | &interm_types)) |
4832 | continue; | |
4833 | else | |
4834 | gcc_assert (multi_step_cvt == 0); | |
4835 | ||
4836 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
4837 | vectype_in, &code1, &code2, |
4838 | &multi_step_cvt, &interm_types)) | |
16d22000 RS |
4839 | { |
4840 | found_mode = true; | |
4841 | break; | |
4842 | } | |
4a00c761 JJ |
4843 | } |
4844 | ||
16d22000 | 4845 | if (!found_mode) |
4a00c761 JJ |
4846 | goto unsupported; |
4847 | ||
4848 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4849 | codecvt2 = ERROR_MARK; | |
4850 | else | |
4851 | { | |
4852 | multi_step_cvt++; | |
9771b263 | 4853 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
4854 | cvt_type = NULL_TREE; |
4855 | } | |
4856 | break; | |
4857 | ||
4858 | case NARROW: | |
4859 | gcc_assert (op_type == unary_op); | |
4860 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
4861 | &code1, &multi_step_cvt, | |
4862 | &interm_types)) | |
4863 | break; | |
4864 | ||
4865 | if (code != FIX_TRUNC_EXPR | |
b397965c | 4866 | || GET_MODE_SIZE (lhs_mode) >= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4867 | goto unsupported; |
4868 | ||
4a00c761 JJ |
4869 | cvt_type |
4870 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4871 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4872 | if (cvt_type == NULL_TREE) | |
4873 | goto unsupported; | |
4874 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
4875 | &decl1, &codecvt1)) | |
4876 | goto unsupported; | |
4877 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
4878 | &code1, &multi_step_cvt, | |
4879 | &interm_types)) | |
4880 | break; | |
4881 | goto unsupported; | |
4882 | ||
4883 | default: | |
4884 | gcc_unreachable (); | |
ebfd146a IR |
4885 | } |
4886 | ||
4887 | if (!vec_stmt) /* transformation not required. */ | |
4888 | { | |
adac3a68 | 4889 | DUMP_VECT_SCOPE ("vectorizable_conversion"); |
4a00c761 | 4890 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
4891 | { |
4892 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
68435eb2 RB |
4893 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, |
4894 | cost_vec); | |
8bd37302 | 4895 | } |
4a00c761 JJ |
4896 | else if (modifier == NARROW) |
4897 | { | |
4898 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
68435eb2 RB |
4899 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt, |
4900 | cost_vec); | |
4a00c761 JJ |
4901 | } |
4902 | else | |
4903 | { | |
4904 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
68435eb2 RB |
4905 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt, |
4906 | cost_vec); | |
4a00c761 | 4907 | } |
9771b263 | 4908 | interm_types.release (); |
ebfd146a IR |
4909 | return true; |
4910 | } | |
4911 | ||
67b8dbac | 4912 | /* Transform. */ |
73fbfcad | 4913 | if (dump_enabled_p ()) |
78c60e3d | 4914 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4915 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 4916 | |
4a00c761 JJ |
4917 | if (op_type == binary_op) |
4918 | { | |
4919 | if (CONSTANT_CLASS_P (op0)) | |
4920 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
4921 | else if (CONSTANT_CLASS_P (op1)) | |
4922 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
4923 | } | |
4924 | ||
4925 | /* In case of multi-step conversion, we first generate conversion operations | |
4926 | to the intermediate types, and then from that types to the final one. | |
4927 | We create vector destinations for the intermediate type (TYPES) received | |
4928 | from supportable_*_operation, and store them in the correct order | |
4929 | for future use in vect_create_vectorized_*_stmts (). */ | |
8c681247 | 4930 | auto_vec<tree> vec_dsts (multi_step_cvt + 1); |
82294ec1 JJ |
4931 | vec_dest = vect_create_destination_var (scalar_dest, |
4932 | (cvt_type && modifier == WIDEN) | |
4933 | ? cvt_type : vectype_out); | |
9771b263 | 4934 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4935 | |
4936 | if (multi_step_cvt) | |
4937 | { | |
9771b263 DN |
4938 | for (i = interm_types.length () - 1; |
4939 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
4940 | { |
4941 | vec_dest = vect_create_destination_var (scalar_dest, | |
4942 | intermediate_type); | |
9771b263 | 4943 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4944 | } |
4945 | } | |
ebfd146a | 4946 | |
4a00c761 | 4947 | if (cvt_type) |
82294ec1 JJ |
4948 | vec_dest = vect_create_destination_var (scalar_dest, |
4949 | modifier == WIDEN | |
4950 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
4951 | |
4952 | if (!slp_node) | |
4953 | { | |
30862efc | 4954 | if (modifier == WIDEN) |
4a00c761 | 4955 | { |
c3284718 | 4956 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 4957 | if (op_type == binary_op) |
9771b263 | 4958 | vec_oprnds1.create (1); |
4a00c761 | 4959 | } |
30862efc | 4960 | else if (modifier == NARROW) |
9771b263 DN |
4961 | vec_oprnds0.create ( |
4962 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
4963 | } |
4964 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 4965 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 4966 | |
4a00c761 | 4967 | last_oprnd = op0; |
ebfd146a IR |
4968 | prev_stmt_info = NULL; |
4969 | switch (modifier) | |
4970 | { | |
4971 | case NONE: | |
4972 | for (j = 0; j < ncopies; j++) | |
4973 | { | |
ebfd146a | 4974 | if (j == 0) |
306b0c92 | 4975 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node); |
ebfd146a IR |
4976 | else |
4977 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
4978 | ||
9771b263 | 4979 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 | 4980 | { |
e1bd7296 | 4981 | stmt_vec_info new_stmt_info; |
4a00c761 JJ |
4982 | /* Arguments are ready, create the new vector stmt. */ |
4983 | if (code1 == CALL_EXPR) | |
4984 | { | |
e1bd7296 | 4985 | gcall *new_stmt = gimple_build_call (decl1, 1, vop0); |
4a00c761 JJ |
4986 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4987 | gimple_call_set_lhs (new_stmt, new_temp); | |
e1bd7296 RS |
4988 | new_stmt_info |
4989 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4a00c761 JJ |
4990 | } |
4991 | else | |
4992 | { | |
4993 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
e1bd7296 RS |
4994 | gassign *new_stmt |
4995 | = gimple_build_assign (vec_dest, code1, vop0); | |
4a00c761 JJ |
4996 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4997 | gimple_assign_set_lhs (new_stmt, new_temp); | |
e1bd7296 RS |
4998 | new_stmt_info |
4999 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4a00c761 JJ |
5000 | } |
5001 | ||
4a00c761 | 5002 | if (slp_node) |
e1bd7296 | 5003 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
225ce44b RB |
5004 | else |
5005 | { | |
5006 | if (!prev_stmt_info) | |
e1bd7296 RS |
5007 | STMT_VINFO_VEC_STMT (stmt_info) |
5008 | = *vec_stmt = new_stmt_info; | |
225ce44b | 5009 | else |
e1bd7296 RS |
5010 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
5011 | prev_stmt_info = new_stmt_info; | |
225ce44b | 5012 | } |
4a00c761 | 5013 | } |
ebfd146a IR |
5014 | } |
5015 | break; | |
5016 | ||
5017 | case WIDEN: | |
5018 | /* In case the vectorization factor (VF) is bigger than the number | |
5019 | of elements that we can fit in a vectype (nunits), we have to | |
5020 | generate more than one vector stmt - i.e - we need to "unroll" | |
5021 | the vector stmt by a factor VF/nunits. */ | |
5022 | for (j = 0; j < ncopies; j++) | |
5023 | { | |
4a00c761 | 5024 | /* Handle uses. */ |
ebfd146a | 5025 | if (j == 0) |
4a00c761 JJ |
5026 | { |
5027 | if (slp_node) | |
5028 | { | |
5029 | if (code == WIDEN_LSHIFT_EXPR) | |
5030 | { | |
5031 | unsigned int k; | |
ebfd146a | 5032 | |
4a00c761 JJ |
5033 | vec_oprnd1 = op1; |
5034 | /* Store vec_oprnd1 for every vector stmt to be created | |
5035 | for SLP_NODE. We check during the analysis that all | |
5036 | the shift arguments are the same. */ | |
5037 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 5038 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
5039 | |
5040 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5041 | slp_node); |
4a00c761 JJ |
5042 | } |
5043 | else | |
5044 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
306b0c92 | 5045 | &vec_oprnds1, slp_node); |
4a00c761 JJ |
5046 | } |
5047 | else | |
5048 | { | |
81c40241 | 5049 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 5050 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
5051 | if (op_type == binary_op) |
5052 | { | |
5053 | if (code == WIDEN_LSHIFT_EXPR) | |
5054 | vec_oprnd1 = op1; | |
5055 | else | |
81c40241 | 5056 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 5057 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
5058 | } |
5059 | } | |
5060 | } | |
ebfd146a | 5061 | else |
4a00c761 JJ |
5062 | { |
5063 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
5064 | vec_oprnds0.truncate (0); |
5065 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
5066 | if (op_type == binary_op) |
5067 | { | |
5068 | if (code == WIDEN_LSHIFT_EXPR) | |
5069 | vec_oprnd1 = op1; | |
5070 | else | |
5071 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
5072 | vec_oprnd1); | |
9771b263 DN |
5073 | vec_oprnds1.truncate (0); |
5074 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
5075 | } |
5076 | } | |
ebfd146a | 5077 | |
4a00c761 JJ |
5078 | /* Arguments are ready. Create the new vector stmts. */ |
5079 | for (i = multi_step_cvt; i >= 0; i--) | |
5080 | { | |
9771b263 | 5081 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
5082 | enum tree_code c1 = code1, c2 = code2; |
5083 | if (i == 0 && codecvt2 != ERROR_MARK) | |
5084 | { | |
5085 | c1 = codecvt1; | |
5086 | c2 = codecvt2; | |
5087 | } | |
5088 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
5089 | &vec_oprnds1, | |
5090 | stmt, this_dest, gsi, | |
5091 | c1, c2, decl1, decl2, | |
5092 | op_type); | |
5093 | } | |
5094 | ||
9771b263 | 5095 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 | 5096 | { |
e1bd7296 | 5097 | stmt_vec_info new_stmt_info; |
4a00c761 JJ |
5098 | if (cvt_type) |
5099 | { | |
5100 | if (codecvt1 == CALL_EXPR) | |
5101 | { | |
e1bd7296 | 5102 | gcall *new_stmt = gimple_build_call (decl1, 1, vop0); |
4a00c761 JJ |
5103 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5104 | gimple_call_set_lhs (new_stmt, new_temp); | |
e1bd7296 RS |
5105 | new_stmt_info |
5106 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4a00c761 JJ |
5107 | } |
5108 | else | |
5109 | { | |
5110 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 5111 | new_temp = make_ssa_name (vec_dest); |
e1bd7296 RS |
5112 | gassign *new_stmt |
5113 | = gimple_build_assign (new_temp, codecvt1, vop0); | |
5114 | new_stmt_info | |
5115 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4a00c761 | 5116 | } |
4a00c761 JJ |
5117 | } |
5118 | else | |
e1bd7296 | 5119 | new_stmt_info = vinfo->lookup_def (vop0); |
4a00c761 JJ |
5120 | |
5121 | if (slp_node) | |
e1bd7296 | 5122 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
4a00c761 | 5123 | else |
c689ce1e RB |
5124 | { |
5125 | if (!prev_stmt_info) | |
e1bd7296 | 5126 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt_info; |
c689ce1e | 5127 | else |
e1bd7296 RS |
5128 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
5129 | prev_stmt_info = new_stmt_info; | |
c689ce1e | 5130 | } |
4a00c761 | 5131 | } |
ebfd146a | 5132 | } |
4a00c761 JJ |
5133 | |
5134 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
5135 | break; |
5136 | ||
5137 | case NARROW: | |
5138 | /* In case the vectorization factor (VF) is bigger than the number | |
5139 | of elements that we can fit in a vectype (nunits), we have to | |
5140 | generate more than one vector stmt - i.e - we need to "unroll" | |
5141 | the vector stmt by a factor VF/nunits. */ | |
5142 | for (j = 0; j < ncopies; j++) | |
5143 | { | |
5144 | /* Handle uses. */ | |
4a00c761 JJ |
5145 | if (slp_node) |
5146 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5147 | slp_node); |
ebfd146a IR |
5148 | else |
5149 | { | |
9771b263 | 5150 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
5151 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
5152 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
5153 | } |
5154 | ||
4a00c761 JJ |
5155 | /* Arguments are ready. Create the new vector stmts. */ |
5156 | if (cvt_type) | |
9771b263 | 5157 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
5158 | { |
5159 | if (codecvt1 == CALL_EXPR) | |
5160 | { | |
e1bd7296 | 5161 | gcall *new_stmt = gimple_build_call (decl1, 1, vop0); |
4a00c761 JJ |
5162 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5163 | gimple_call_set_lhs (new_stmt, new_temp); | |
e1bd7296 | 5164 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4a00c761 JJ |
5165 | } |
5166 | else | |
5167 | { | |
5168 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 5169 | new_temp = make_ssa_name (vec_dest); |
e1bd7296 RS |
5170 | gassign *new_stmt |
5171 | = gimple_build_assign (new_temp, codecvt1, vop0); | |
5172 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4a00c761 | 5173 | } |
ebfd146a | 5174 | |
9771b263 | 5175 | vec_oprnds0[i] = new_temp; |
4a00c761 | 5176 | } |
ebfd146a | 5177 | |
4a00c761 JJ |
5178 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
5179 | stmt, vec_dsts, gsi, | |
5180 | slp_node, code1, | |
5181 | &prev_stmt_info); | |
ebfd146a IR |
5182 | } |
5183 | ||
5184 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 5185 | break; |
ebfd146a IR |
5186 | } |
5187 | ||
9771b263 DN |
5188 | vec_oprnds0.release (); |
5189 | vec_oprnds1.release (); | |
9771b263 | 5190 | interm_types.release (); |
ebfd146a IR |
5191 | |
5192 | return true; | |
5193 | } | |
ff802fa1 IR |
5194 | |
5195 | ||
ebfd146a IR |
5196 | /* Function vectorizable_assignment. |
5197 | ||
b8698a0f L |
5198 | Check if STMT performs an assignment (copy) that can be vectorized. |
5199 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5200 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5201 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5202 | ||
5203 | static bool | |
355fe088 | 5204 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
1eede195 | 5205 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
68435eb2 | 5206 | stmt_vector_for_cost *cost_vec) |
ebfd146a IR |
5207 | { |
5208 | tree vec_dest; | |
5209 | tree scalar_dest; | |
5210 | tree op; | |
5211 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
5212 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
5213 | tree new_temp; | |
4fc5ebf1 JG |
5214 | enum vect_def_type dt[1] = {vect_unknown_def_type}; |
5215 | int ndts = 1; | |
ebfd146a | 5216 | int ncopies; |
f18b55bd | 5217 | int i, j; |
6e1aa848 | 5218 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5219 | tree vop; |
a70d6342 | 5220 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5221 | vec_info *vinfo = stmt_info->vinfo; |
f18b55bd | 5222 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
5223 | enum tree_code code; |
5224 | tree vectype_in; | |
ebfd146a | 5225 | |
a70d6342 | 5226 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5227 | return false; |
5228 | ||
66c16fd9 RB |
5229 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5230 | && ! vec_stmt) | |
ebfd146a IR |
5231 | return false; |
5232 | ||
5233 | /* Is vectorizable assignment? */ | |
5234 | if (!is_gimple_assign (stmt)) | |
5235 | return false; | |
5236 | ||
5237 | scalar_dest = gimple_assign_lhs (stmt); | |
5238 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
5239 | return false; | |
5240 | ||
fde9c428 | 5241 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 5242 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
5243 | || code == PAREN_EXPR |
5244 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
5245 | op = gimple_assign_rhs1 (stmt); |
5246 | else | |
5247 | return false; | |
5248 | ||
7b7ec6c5 RG |
5249 | if (code == VIEW_CONVERT_EXPR) |
5250 | op = TREE_OPERAND (op, 0); | |
5251 | ||
465c8c19 | 5252 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
928686b1 | 5253 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 JJ |
5254 | |
5255 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5256 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5257 | case of SLP. */ | |
fce57248 | 5258 | if (slp_node) |
465c8c19 JJ |
5259 | ncopies = 1; |
5260 | else | |
e8f142e2 | 5261 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
5262 | |
5263 | gcc_assert (ncopies >= 1); | |
5264 | ||
894dd753 | 5265 | if (!vect_is_simple_use (op, vinfo, &dt[0], &vectype_in)) |
ebfd146a | 5266 | { |
73fbfcad | 5267 | if (dump_enabled_p ()) |
78c60e3d | 5268 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5269 | "use not simple.\n"); |
ebfd146a IR |
5270 | return false; |
5271 | } | |
5272 | ||
fde9c428 RG |
5273 | /* We can handle NOP_EXPR conversions that do not change the number |
5274 | of elements or the vector size. */ | |
7b7ec6c5 RG |
5275 | if ((CONVERT_EXPR_CODE_P (code) |
5276 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 | 5277 | && (!vectype_in |
928686b1 | 5278 | || maybe_ne (TYPE_VECTOR_SUBPARTS (vectype_in), nunits) |
cf098191 RS |
5279 | || maybe_ne (GET_MODE_SIZE (TYPE_MODE (vectype)), |
5280 | GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
fde9c428 RG |
5281 | return false; |
5282 | ||
7b7b1813 RG |
5283 | /* We do not handle bit-precision changes. */ |
5284 | if ((CONVERT_EXPR_CODE_P (code) | |
5285 | || code == VIEW_CONVERT_EXPR) | |
5286 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2be65d9e RS |
5287 | && (!type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
5288 | || !type_has_mode_precision_p (TREE_TYPE (op))) | |
7b7b1813 RG |
5289 | /* But a conversion that does not change the bit-pattern is ok. */ |
5290 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
5291 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2dab46d5 IE |
5292 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
5293 | /* Conversion between boolean types of different sizes is | |
5294 | a simple assignment in case their vectypes are same | |
5295 | boolean vectors. */ | |
5296 | && (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
5297 | || !VECTOR_BOOLEAN_TYPE_P (vectype_in))) | |
7b7b1813 | 5298 | { |
73fbfcad | 5299 | if (dump_enabled_p ()) |
78c60e3d SS |
5300 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5301 | "type conversion to/from bit-precision " | |
e645e942 | 5302 | "unsupported.\n"); |
7b7b1813 RG |
5303 | return false; |
5304 | } | |
5305 | ||
ebfd146a IR |
5306 | if (!vec_stmt) /* transformation not required. */ |
5307 | { | |
5308 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
adac3a68 | 5309 | DUMP_VECT_SCOPE ("vectorizable_assignment"); |
68435eb2 | 5310 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
ebfd146a IR |
5311 | return true; |
5312 | } | |
5313 | ||
67b8dbac | 5314 | /* Transform. */ |
73fbfcad | 5315 | if (dump_enabled_p ()) |
e645e942 | 5316 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
5317 | |
5318 | /* Handle def. */ | |
5319 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5320 | ||
5321 | /* Handle use. */ | |
f18b55bd | 5322 | for (j = 0; j < ncopies; j++) |
ebfd146a | 5323 | { |
f18b55bd IR |
5324 | /* Handle uses. */ |
5325 | if (j == 0) | |
306b0c92 | 5326 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
f18b55bd IR |
5327 | else |
5328 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
5329 | ||
5330 | /* Arguments are ready. create the new vector stmt. */ | |
e1bd7296 | 5331 | stmt_vec_info new_stmt_info = NULL; |
9771b263 | 5332 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 5333 | { |
7b7ec6c5 RG |
5334 | if (CONVERT_EXPR_CODE_P (code) |
5335 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 5336 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
e1bd7296 | 5337 | gassign *new_stmt = gimple_build_assign (vec_dest, vop); |
f18b55bd IR |
5338 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5339 | gimple_assign_set_lhs (new_stmt, new_temp); | |
e1bd7296 | 5340 | new_stmt_info = vect_finish_stmt_generation (stmt, new_stmt, gsi); |
f18b55bd | 5341 | if (slp_node) |
e1bd7296 | 5342 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
f18b55bd | 5343 | } |
ebfd146a IR |
5344 | |
5345 | if (slp_node) | |
f18b55bd IR |
5346 | continue; |
5347 | ||
5348 | if (j == 0) | |
e1bd7296 | 5349 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
f18b55bd | 5350 | else |
e1bd7296 | 5351 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
f18b55bd | 5352 | |
e1bd7296 | 5353 | prev_stmt_info = new_stmt_info; |
f18b55bd | 5354 | } |
b8698a0f | 5355 | |
9771b263 | 5356 | vec_oprnds.release (); |
ebfd146a IR |
5357 | return true; |
5358 | } | |
5359 | ||
9dc3f7de | 5360 | |
1107f3ae IR |
5361 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
5362 | either as shift by a scalar or by a vector. */ | |
5363 | ||
5364 | bool | |
5365 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
5366 | { | |
5367 | ||
ef4bddc2 | 5368 | machine_mode vec_mode; |
1107f3ae IR |
5369 | optab optab; |
5370 | int icode; | |
5371 | tree vectype; | |
5372 | ||
5373 | vectype = get_vectype_for_scalar_type (scalar_type); | |
5374 | if (!vectype) | |
5375 | return false; | |
5376 | ||
5377 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5378 | if (!optab | |
5379 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
5380 | { | |
5381 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5382 | if (!optab | |
5383 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
5384 | == CODE_FOR_nothing)) | |
5385 | return false; | |
5386 | } | |
5387 | ||
5388 | vec_mode = TYPE_MODE (vectype); | |
5389 | icode = (int) optab_handler (optab, vec_mode); | |
5390 | if (icode == CODE_FOR_nothing) | |
5391 | return false; | |
5392 | ||
5393 | return true; | |
5394 | } | |
5395 | ||
5396 | ||
9dc3f7de IR |
5397 | /* Function vectorizable_shift. |
5398 | ||
5399 | Check if STMT performs a shift operation that can be vectorized. | |
5400 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
5401 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
5402 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5403 | ||
5404 | static bool | |
355fe088 | 5405 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
1eede195 | 5406 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
68435eb2 | 5407 | stmt_vector_for_cost *cost_vec) |
9dc3f7de IR |
5408 | { |
5409 | tree vec_dest; | |
5410 | tree scalar_dest; | |
5411 | tree op0, op1 = NULL; | |
5412 | tree vec_oprnd1 = NULL_TREE; | |
5413 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5414 | tree vectype; | |
5415 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
5416 | enum tree_code code; | |
ef4bddc2 | 5417 | machine_mode vec_mode; |
9dc3f7de IR |
5418 | tree new_temp; |
5419 | optab optab; | |
5420 | int icode; | |
ef4bddc2 | 5421 | machine_mode optab_op2_mode; |
9dc3f7de | 5422 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 5423 | int ndts = 2; |
9dc3f7de | 5424 | stmt_vec_info prev_stmt_info; |
928686b1 RS |
5425 | poly_uint64 nunits_in; |
5426 | poly_uint64 nunits_out; | |
9dc3f7de | 5427 | tree vectype_out; |
cede2577 | 5428 | tree op1_vectype; |
9dc3f7de IR |
5429 | int ncopies; |
5430 | int j, i; | |
6e1aa848 DN |
5431 | vec<tree> vec_oprnds0 = vNULL; |
5432 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
5433 | tree vop0, vop1; |
5434 | unsigned int k; | |
49eab32e | 5435 | bool scalar_shift_arg = true; |
9dc3f7de | 5436 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5437 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
5438 | |
5439 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
5440 | return false; | |
5441 | ||
66c16fd9 RB |
5442 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5443 | && ! vec_stmt) | |
9dc3f7de IR |
5444 | return false; |
5445 | ||
5446 | /* Is STMT a vectorizable binary/unary operation? */ | |
5447 | if (!is_gimple_assign (stmt)) | |
5448 | return false; | |
5449 | ||
5450 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5451 | return false; | |
5452 | ||
5453 | code = gimple_assign_rhs_code (stmt); | |
5454 | ||
5455 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
5456 | || code == RROTATE_EXPR)) | |
5457 | return false; | |
5458 | ||
5459 | scalar_dest = gimple_assign_lhs (stmt); | |
5460 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2be65d9e | 5461 | if (!type_has_mode_precision_p (TREE_TYPE (scalar_dest))) |
7b7b1813 | 5462 | { |
73fbfcad | 5463 | if (dump_enabled_p ()) |
78c60e3d | 5464 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5465 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
5466 | return false; |
5467 | } | |
9dc3f7de IR |
5468 | |
5469 | op0 = gimple_assign_rhs1 (stmt); | |
894dd753 | 5470 | if (!vect_is_simple_use (op0, vinfo, &dt[0], &vectype)) |
9dc3f7de | 5471 | { |
73fbfcad | 5472 | if (dump_enabled_p ()) |
78c60e3d | 5473 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5474 | "use not simple.\n"); |
9dc3f7de IR |
5475 | return false; |
5476 | } | |
5477 | /* If op0 is an external or constant def use a vector type with | |
5478 | the same size as the output vector type. */ | |
5479 | if (!vectype) | |
5480 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5481 | if (vec_stmt) | |
5482 | gcc_assert (vectype); | |
5483 | if (!vectype) | |
5484 | { | |
73fbfcad | 5485 | if (dump_enabled_p ()) |
78c60e3d | 5486 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5487 | "no vectype for scalar type\n"); |
9dc3f7de IR |
5488 | return false; |
5489 | } | |
5490 | ||
5491 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5492 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
928686b1 | 5493 | if (maybe_ne (nunits_out, nunits_in)) |
9dc3f7de IR |
5494 | return false; |
5495 | ||
5496 | op1 = gimple_assign_rhs2 (stmt); | |
fef96d8e RS |
5497 | stmt_vec_info op1_def_stmt_info; |
5498 | if (!vect_is_simple_use (op1, vinfo, &dt[1], &op1_vectype, | |
5499 | &op1_def_stmt_info)) | |
9dc3f7de | 5500 | { |
73fbfcad | 5501 | if (dump_enabled_p ()) |
78c60e3d | 5502 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5503 | "use not simple.\n"); |
9dc3f7de IR |
5504 | return false; |
5505 | } | |
5506 | ||
9dc3f7de IR |
5507 | /* Multiple types in SLP are handled by creating the appropriate number of |
5508 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5509 | case of SLP. */ | |
fce57248 | 5510 | if (slp_node) |
9dc3f7de IR |
5511 | ncopies = 1; |
5512 | else | |
e8f142e2 | 5513 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
9dc3f7de IR |
5514 | |
5515 | gcc_assert (ncopies >= 1); | |
5516 | ||
5517 | /* Determine whether the shift amount is a vector, or scalar. If the | |
5518 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
5519 | ||
dbfa87aa YR |
5520 | if ((dt[1] == vect_internal_def |
5521 | || dt[1] == vect_induction_def) | |
5522 | && !slp_node) | |
49eab32e JJ |
5523 | scalar_shift_arg = false; |
5524 | else if (dt[1] == vect_constant_def | |
5525 | || dt[1] == vect_external_def | |
5526 | || dt[1] == vect_internal_def) | |
5527 | { | |
5528 | /* In SLP, need to check whether the shift count is the same, | |
5529 | in loops if it is a constant or invariant, it is always | |
5530 | a scalar shift. */ | |
5531 | if (slp_node) | |
5532 | { | |
b9787581 RS |
5533 | vec<stmt_vec_info> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
5534 | stmt_vec_info slpstmt_info; | |
49eab32e | 5535 | |
b9787581 RS |
5536 | FOR_EACH_VEC_ELT (stmts, k, slpstmt_info) |
5537 | { | |
5538 | gassign *slpstmt = as_a <gassign *> (slpstmt_info->stmt); | |
5539 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) | |
5540 | scalar_shift_arg = false; | |
5541 | } | |
49eab32e | 5542 | } |
60d393e8 RB |
5543 | |
5544 | /* If the shift amount is computed by a pattern stmt we cannot | |
5545 | use the scalar amount directly thus give up and use a vector | |
5546 | shift. */ | |
fef96d8e RS |
5547 | if (op1_def_stmt_info && is_pattern_stmt_p (op1_def_stmt_info)) |
5548 | scalar_shift_arg = false; | |
49eab32e JJ |
5549 | } |
5550 | else | |
5551 | { | |
73fbfcad | 5552 | if (dump_enabled_p ()) |
78c60e3d | 5553 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5554 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
5555 | return false; |
5556 | } | |
5557 | ||
9dc3f7de | 5558 | /* Vector shifted by vector. */ |
49eab32e | 5559 | if (!scalar_shift_arg) |
9dc3f7de IR |
5560 | { |
5561 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 5562 | if (dump_enabled_p ()) |
78c60e3d | 5563 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5564 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 5565 | |
aa948027 JJ |
5566 | if (!op1_vectype) |
5567 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
5568 | if (op1_vectype == NULL_TREE | |
5569 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 5570 | { |
73fbfcad | 5571 | if (dump_enabled_p ()) |
78c60e3d SS |
5572 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5573 | "unusable type for last operand in" | |
e645e942 | 5574 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
5575 | return false; |
5576 | } | |
9dc3f7de IR |
5577 | } |
5578 | /* See if the machine has a vector shifted by scalar insn and if not | |
5579 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 5580 | else |
9dc3f7de IR |
5581 | { |
5582 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5583 | if (optab | |
5584 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
5585 | { | |
73fbfcad | 5586 | if (dump_enabled_p ()) |
78c60e3d | 5587 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5588 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
5589 | } |
5590 | else | |
5591 | { | |
5592 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5593 | if (optab | |
5594 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
5595 | != CODE_FOR_nothing)) | |
5596 | { | |
49eab32e JJ |
5597 | scalar_shift_arg = false; |
5598 | ||
73fbfcad | 5599 | if (dump_enabled_p ()) |
78c60e3d | 5600 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5601 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
5602 | |
5603 | /* Unlike the other binary operators, shifts/rotates have | |
5604 | the rhs being int, instead of the same type as the lhs, | |
5605 | so make sure the scalar is the right type if we are | |
aa948027 | 5606 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
5607 | if (dt[1] == vect_constant_def) |
5608 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
5609 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
5610 | TREE_TYPE (op1))) | |
5611 | { | |
5612 | if (slp_node | |
5613 | && TYPE_MODE (TREE_TYPE (vectype)) | |
5614 | != TYPE_MODE (TREE_TYPE (op1))) | |
5615 | { | |
73fbfcad | 5616 | if (dump_enabled_p ()) |
78c60e3d SS |
5617 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5618 | "unusable type for last operand in" | |
e645e942 | 5619 | " vector/vector shift/rotate.\n"); |
21c0a521 | 5620 | return false; |
aa948027 JJ |
5621 | } |
5622 | if (vec_stmt && !slp_node) | |
5623 | { | |
5624 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
5625 | op1 = vect_init_vector (stmt, op1, | |
5626 | TREE_TYPE (vectype), NULL); | |
5627 | } | |
5628 | } | |
9dc3f7de IR |
5629 | } |
5630 | } | |
5631 | } | |
9dc3f7de IR |
5632 | |
5633 | /* Supportable by target? */ | |
5634 | if (!optab) | |
5635 | { | |
73fbfcad | 5636 | if (dump_enabled_p ()) |
78c60e3d | 5637 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5638 | "no optab.\n"); |
9dc3f7de IR |
5639 | return false; |
5640 | } | |
5641 | vec_mode = TYPE_MODE (vectype); | |
5642 | icode = (int) optab_handler (optab, vec_mode); | |
5643 | if (icode == CODE_FOR_nothing) | |
5644 | { | |
73fbfcad | 5645 | if (dump_enabled_p ()) |
78c60e3d | 5646 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5647 | "op not supported by target.\n"); |
9dc3f7de | 5648 | /* Check only during analysis. */ |
cf098191 | 5649 | if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) |
ca09abcb RS |
5650 | || (!vec_stmt |
5651 | && !vect_worthwhile_without_simd_p (vinfo, code))) | |
9dc3f7de | 5652 | return false; |
73fbfcad | 5653 | if (dump_enabled_p ()) |
e645e942 TJ |
5654 | dump_printf_loc (MSG_NOTE, vect_location, |
5655 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
5656 | } |
5657 | ||
5658 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
ca09abcb RS |
5659 | if (!vec_stmt |
5660 | && !VECTOR_MODE_P (TYPE_MODE (vectype)) | |
5661 | && !vect_worthwhile_without_simd_p (vinfo, code)) | |
9dc3f7de | 5662 | { |
73fbfcad | 5663 | if (dump_enabled_p ()) |
78c60e3d | 5664 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5665 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
5666 | return false; |
5667 | } | |
5668 | ||
5669 | if (!vec_stmt) /* transformation not required. */ | |
5670 | { | |
5671 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
adac3a68 | 5672 | DUMP_VECT_SCOPE ("vectorizable_shift"); |
68435eb2 | 5673 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
9dc3f7de IR |
5674 | return true; |
5675 | } | |
5676 | ||
67b8dbac | 5677 | /* Transform. */ |
9dc3f7de | 5678 | |
73fbfcad | 5679 | if (dump_enabled_p ()) |
78c60e3d | 5680 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5681 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
5682 | |
5683 | /* Handle def. */ | |
5684 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5685 | ||
9dc3f7de IR |
5686 | prev_stmt_info = NULL; |
5687 | for (j = 0; j < ncopies; j++) | |
5688 | { | |
5689 | /* Handle uses. */ | |
5690 | if (j == 0) | |
5691 | { | |
5692 | if (scalar_shift_arg) | |
5693 | { | |
5694 | /* Vector shl and shr insn patterns can be defined with scalar | |
5695 | operand 2 (shift operand). In this case, use constant or loop | |
5696 | invariant op1 directly, without extending it to vector mode | |
5697 | first. */ | |
5698 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
5699 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
5700 | { | |
73fbfcad | 5701 | if (dump_enabled_p ()) |
78c60e3d | 5702 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5703 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 5704 | vec_oprnd1 = op1; |
8930f723 | 5705 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 5706 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5707 | if (slp_node) |
5708 | { | |
5709 | /* Store vec_oprnd1 for every vector stmt to be created | |
5710 | for SLP_NODE. We check during the analysis that all | |
5711 | the shift arguments are the same. | |
5712 | TODO: Allow different constants for different vector | |
5713 | stmts generated for an SLP instance. */ | |
5714 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 5715 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5716 | } |
5717 | } | |
5718 | } | |
5719 | ||
5720 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
5721 | (a special case for certain kind of vector shifts); otherwise, | |
5722 | operand 1 should be of a vector type (the usual case). */ | |
5723 | if (vec_oprnd1) | |
5724 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5725 | slp_node); |
9dc3f7de IR |
5726 | else |
5727 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
306b0c92 | 5728 | slp_node); |
9dc3f7de IR |
5729 | } |
5730 | else | |
5731 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5732 | ||
5733 | /* Arguments are ready. Create the new vector stmt. */ | |
e1bd7296 | 5734 | stmt_vec_info new_stmt_info = NULL; |
9771b263 | 5735 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 5736 | { |
9771b263 | 5737 | vop1 = vec_oprnds1[i]; |
e1bd7296 | 5738 | gassign *new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
5739 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5740 | gimple_assign_set_lhs (new_stmt, new_temp); | |
e1bd7296 | 5741 | new_stmt_info = vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9dc3f7de | 5742 | if (slp_node) |
e1bd7296 | 5743 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
9dc3f7de IR |
5744 | } |
5745 | ||
5746 | if (slp_node) | |
5747 | continue; | |
5748 | ||
5749 | if (j == 0) | |
e1bd7296 | 5750 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
9dc3f7de | 5751 | else |
e1bd7296 RS |
5752 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
5753 | prev_stmt_info = new_stmt_info; | |
9dc3f7de IR |
5754 | } |
5755 | ||
9771b263 DN |
5756 | vec_oprnds0.release (); |
5757 | vec_oprnds1.release (); | |
9dc3f7de IR |
5758 | |
5759 | return true; | |
5760 | } | |
5761 | ||
5762 | ||
ebfd146a IR |
5763 | /* Function vectorizable_operation. |
5764 | ||
16949072 RG |
5765 | Check if STMT performs a binary, unary or ternary operation that can |
5766 | be vectorized. | |
b8698a0f | 5767 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
5768 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5769 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5770 | ||
5771 | static bool | |
355fe088 | 5772 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
1eede195 | 5773 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
68435eb2 | 5774 | stmt_vector_for_cost *cost_vec) |
ebfd146a | 5775 | { |
00f07b86 | 5776 | tree vec_dest; |
ebfd146a | 5777 | tree scalar_dest; |
16949072 | 5778 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 5779 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 5780 | tree vectype; |
ebfd146a | 5781 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
0eb952ea | 5782 | enum tree_code code, orig_code; |
ef4bddc2 | 5783 | machine_mode vec_mode; |
ebfd146a IR |
5784 | tree new_temp; |
5785 | int op_type; | |
00f07b86 | 5786 | optab optab; |
523ba738 | 5787 | bool target_support_p; |
16949072 RG |
5788 | enum vect_def_type dt[3] |
5789 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 5790 | int ndts = 3; |
ebfd146a | 5791 | stmt_vec_info prev_stmt_info; |
928686b1 RS |
5792 | poly_uint64 nunits_in; |
5793 | poly_uint64 nunits_out; | |
ebfd146a IR |
5794 | tree vectype_out; |
5795 | int ncopies; | |
5796 | int j, i; | |
6e1aa848 DN |
5797 | vec<tree> vec_oprnds0 = vNULL; |
5798 | vec<tree> vec_oprnds1 = vNULL; | |
5799 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 5800 | tree vop0, vop1, vop2; |
a70d6342 | 5801 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5802 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 5803 | |
a70d6342 | 5804 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5805 | return false; |
5806 | ||
66c16fd9 RB |
5807 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5808 | && ! vec_stmt) | |
ebfd146a IR |
5809 | return false; |
5810 | ||
5811 | /* Is STMT a vectorizable binary/unary operation? */ | |
5812 | if (!is_gimple_assign (stmt)) | |
5813 | return false; | |
5814 | ||
5815 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5816 | return false; | |
5817 | ||
0eb952ea | 5818 | orig_code = code = gimple_assign_rhs_code (stmt); |
ebfd146a | 5819 | |
1af4ebf5 MG |
5820 | /* For pointer addition and subtraction, we should use the normal |
5821 | plus and minus for the vector operation. */ | |
ebfd146a IR |
5822 | if (code == POINTER_PLUS_EXPR) |
5823 | code = PLUS_EXPR; | |
1af4ebf5 MG |
5824 | if (code == POINTER_DIFF_EXPR) |
5825 | code = MINUS_EXPR; | |
ebfd146a IR |
5826 | |
5827 | /* Support only unary or binary operations. */ | |
5828 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 5829 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 5830 | { |
73fbfcad | 5831 | if (dump_enabled_p ()) |
78c60e3d | 5832 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5833 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 5834 | op_type); |
ebfd146a IR |
5835 | return false; |
5836 | } | |
5837 | ||
b690cc0f RG |
5838 | scalar_dest = gimple_assign_lhs (stmt); |
5839 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
5840 | ||
7b7b1813 RG |
5841 | /* Most operations cannot handle bit-precision types without extra |
5842 | truncations. */ | |
045c1278 | 5843 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
2be65d9e | 5844 | && !type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
7b7b1813 RG |
5845 | /* Exception are bitwise binary operations. */ |
5846 | && code != BIT_IOR_EXPR | |
5847 | && code != BIT_XOR_EXPR | |
5848 | && code != BIT_AND_EXPR) | |
5849 | { | |
73fbfcad | 5850 | if (dump_enabled_p ()) |
78c60e3d | 5851 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5852 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
5853 | return false; |
5854 | } | |
5855 | ||
ebfd146a | 5856 | op0 = gimple_assign_rhs1 (stmt); |
894dd753 | 5857 | if (!vect_is_simple_use (op0, vinfo, &dt[0], &vectype)) |
ebfd146a | 5858 | { |
73fbfcad | 5859 | if (dump_enabled_p ()) |
78c60e3d | 5860 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5861 | "use not simple.\n"); |
ebfd146a IR |
5862 | return false; |
5863 | } | |
b690cc0f RG |
5864 | /* If op0 is an external or constant def use a vector type with |
5865 | the same size as the output vector type. */ | |
5866 | if (!vectype) | |
b036c6c5 IE |
5867 | { |
5868 | /* For boolean type we cannot determine vectype by | |
5869 | invariant value (don't know whether it is a vector | |
5870 | of booleans or vector of integers). We use output | |
5871 | vectype because operations on boolean don't change | |
5872 | type. */ | |
2568d8a1 | 5873 | if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op0))) |
b036c6c5 | 5874 | { |
2568d8a1 | 5875 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (scalar_dest))) |
b036c6c5 IE |
5876 | { |
5877 | if (dump_enabled_p ()) | |
5878 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5879 | "not supported operation on bool value.\n"); | |
5880 | return false; | |
5881 | } | |
5882 | vectype = vectype_out; | |
5883 | } | |
5884 | else | |
5885 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5886 | } | |
7d8930a0 IR |
5887 | if (vec_stmt) |
5888 | gcc_assert (vectype); | |
5889 | if (!vectype) | |
5890 | { | |
73fbfcad | 5891 | if (dump_enabled_p ()) |
7d8930a0 | 5892 | { |
78c60e3d SS |
5893 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5894 | "no vectype for scalar type "); | |
5895 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
5896 | TREE_TYPE (op0)); | |
e645e942 | 5897 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
5898 | } |
5899 | ||
5900 | return false; | |
5901 | } | |
b690cc0f RG |
5902 | |
5903 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5904 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
928686b1 | 5905 | if (maybe_ne (nunits_out, nunits_in)) |
b690cc0f | 5906 | return false; |
ebfd146a | 5907 | |
16949072 | 5908 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
5909 | { |
5910 | op1 = gimple_assign_rhs2 (stmt); | |
894dd753 | 5911 | if (!vect_is_simple_use (op1, vinfo, &dt[1])) |
ebfd146a | 5912 | { |
73fbfcad | 5913 | if (dump_enabled_p ()) |
78c60e3d | 5914 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5915 | "use not simple.\n"); |
ebfd146a IR |
5916 | return false; |
5917 | } | |
5918 | } | |
16949072 RG |
5919 | if (op_type == ternary_op) |
5920 | { | |
5921 | op2 = gimple_assign_rhs3 (stmt); | |
894dd753 | 5922 | if (!vect_is_simple_use (op2, vinfo, &dt[2])) |
16949072 | 5923 | { |
73fbfcad | 5924 | if (dump_enabled_p ()) |
78c60e3d | 5925 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5926 | "use not simple.\n"); |
16949072 RG |
5927 | return false; |
5928 | } | |
5929 | } | |
ebfd146a | 5930 | |
b690cc0f | 5931 | /* Multiple types in SLP are handled by creating the appropriate number of |
ff802fa1 | 5932 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 5933 | case of SLP. */ |
fce57248 | 5934 | if (slp_node) |
b690cc0f RG |
5935 | ncopies = 1; |
5936 | else | |
e8f142e2 | 5937 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
b690cc0f RG |
5938 | |
5939 | gcc_assert (ncopies >= 1); | |
5940 | ||
9dc3f7de | 5941 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
5942 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
5943 | || code == RROTATE_EXPR) | |
9dc3f7de | 5944 | return false; |
ebfd146a | 5945 | |
ebfd146a | 5946 | /* Supportable by target? */ |
00f07b86 RH |
5947 | |
5948 | vec_mode = TYPE_MODE (vectype); | |
5949 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 5950 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
5951 | else |
5952 | { | |
5953 | optab = optab_for_tree_code (code, vectype, optab_default); | |
5954 | if (!optab) | |
5deb57cb | 5955 | { |
73fbfcad | 5956 | if (dump_enabled_p ()) |
78c60e3d | 5957 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5958 | "no optab.\n"); |
00f07b86 | 5959 | return false; |
5deb57cb | 5960 | } |
523ba738 RS |
5961 | target_support_p = (optab_handler (optab, vec_mode) |
5962 | != CODE_FOR_nothing); | |
5deb57cb JJ |
5963 | } |
5964 | ||
523ba738 | 5965 | if (!target_support_p) |
ebfd146a | 5966 | { |
73fbfcad | 5967 | if (dump_enabled_p ()) |
78c60e3d | 5968 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5969 | "op not supported by target.\n"); |
ebfd146a | 5970 | /* Check only during analysis. */ |
cf098191 | 5971 | if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) |
ca09abcb | 5972 | || (!vec_stmt && !vect_worthwhile_without_simd_p (vinfo, code))) |
ebfd146a | 5973 | return false; |
73fbfcad | 5974 | if (dump_enabled_p ()) |
e645e942 TJ |
5975 | dump_printf_loc (MSG_NOTE, vect_location, |
5976 | "proceeding using word mode.\n"); | |
383d9c83 IR |
5977 | } |
5978 | ||
4a00c761 | 5979 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
5980 | if (!VECTOR_MODE_P (vec_mode) |
5981 | && !vec_stmt | |
ca09abcb | 5982 | && !vect_worthwhile_without_simd_p (vinfo, code)) |
7d8930a0 | 5983 | { |
73fbfcad | 5984 | if (dump_enabled_p ()) |
78c60e3d | 5985 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5986 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 5987 | return false; |
7d8930a0 | 5988 | } |
ebfd146a | 5989 | |
ebfd146a IR |
5990 | if (!vec_stmt) /* transformation not required. */ |
5991 | { | |
4a00c761 | 5992 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
adac3a68 | 5993 | DUMP_VECT_SCOPE ("vectorizable_operation"); |
68435eb2 | 5994 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
ebfd146a IR |
5995 | return true; |
5996 | } | |
5997 | ||
67b8dbac | 5998 | /* Transform. */ |
ebfd146a | 5999 | |
73fbfcad | 6000 | if (dump_enabled_p ()) |
78c60e3d | 6001 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6002 | "transform binary/unary operation.\n"); |
383d9c83 | 6003 | |
0eb952ea JJ |
6004 | /* POINTER_DIFF_EXPR has pointer arguments which are vectorized as |
6005 | vectors with unsigned elements, but the result is signed. So, we | |
6006 | need to compute the MINUS_EXPR into vectype temporary and | |
6007 | VIEW_CONVERT_EXPR it into the final vectype_out result. */ | |
6008 | tree vec_cvt_dest = NULL_TREE; | |
6009 | if (orig_code == POINTER_DIFF_EXPR) | |
7b76867b RB |
6010 | { |
6011 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
6012 | vec_cvt_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
6013 | } | |
6014 | /* Handle def. */ | |
6015 | else | |
6016 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
0eb952ea | 6017 | |
ebfd146a IR |
6018 | /* In case the vectorization factor (VF) is bigger than the number |
6019 | of elements that we can fit in a vectype (nunits), we have to generate | |
6020 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
6021 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
6022 | from one copy of the vector stmt to the next, in the field | |
6023 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
6024 | stages to find the correct vector defs to be used when vectorizing | |
6025 | stmts that use the defs of the current stmt. The example below | |
6026 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
6027 | we need to create 4 vectorized stmts): | |
6028 | ||
6029 | before vectorization: | |
6030 | RELATED_STMT VEC_STMT | |
6031 | S1: x = memref - - | |
6032 | S2: z = x + 1 - - | |
6033 | ||
6034 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
6035 | there): | |
6036 | RELATED_STMT VEC_STMT | |
6037 | VS1_0: vx0 = memref0 VS1_1 - | |
6038 | VS1_1: vx1 = memref1 VS1_2 - | |
6039 | VS1_2: vx2 = memref2 VS1_3 - | |
6040 | VS1_3: vx3 = memref3 - - | |
6041 | S1: x = load - VS1_0 | |
6042 | S2: z = x + 1 - - | |
6043 | ||
6044 | step2: vectorize stmt S2 (done here): | |
6045 | To vectorize stmt S2 we first need to find the relevant vector | |
6046 | def for the first operand 'x'. This is, as usual, obtained from | |
6047 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
6048 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
6049 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
6050 | the vector stmt VS2_0, and as usual, record it in the | |
6051 | STMT_VINFO_VEC_STMT of stmt S2. | |
6052 | When creating the second copy (VS2_1), we obtain the relevant vector | |
6053 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
6054 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
6055 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
6056 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
6057 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
6058 | chain of stmts and pointers: | |
6059 | RELATED_STMT VEC_STMT | |
6060 | VS1_0: vx0 = memref0 VS1_1 - | |
6061 | VS1_1: vx1 = memref1 VS1_2 - | |
6062 | VS1_2: vx2 = memref2 VS1_3 - | |
6063 | VS1_3: vx3 = memref3 - - | |
6064 | S1: x = load - VS1_0 | |
6065 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
6066 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
6067 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
6068 | VS2_3: vz3 = vx3 + v1 - - | |
6069 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
6070 | |
6071 | prev_stmt_info = NULL; | |
6072 | for (j = 0; j < ncopies; j++) | |
6073 | { | |
6074 | /* Handle uses. */ | |
6075 | if (j == 0) | |
4a00c761 | 6076 | { |
d6476f90 | 6077 | if (op_type == binary_op) |
4a00c761 | 6078 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, |
306b0c92 | 6079 | slp_node); |
d6476f90 RB |
6080 | else if (op_type == ternary_op) |
6081 | { | |
6082 | if (slp_node) | |
6083 | { | |
6084 | auto_vec<tree> ops(3); | |
6085 | ops.quick_push (op0); | |
6086 | ops.quick_push (op1); | |
6087 | ops.quick_push (op2); | |
6088 | auto_vec<vec<tree> > vec_defs(3); | |
6089 | vect_get_slp_defs (ops, slp_node, &vec_defs); | |
6090 | vec_oprnds0 = vec_defs[0]; | |
6091 | vec_oprnds1 = vec_defs[1]; | |
6092 | vec_oprnds2 = vec_defs[2]; | |
6093 | } | |
6094 | else | |
6095 | { | |
6096 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
6097 | NULL); | |
6098 | vect_get_vec_defs (op2, NULL_TREE, stmt, &vec_oprnds2, NULL, | |
6099 | NULL); | |
6100 | } | |
6101 | } | |
4a00c761 JJ |
6102 | else |
6103 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 6104 | slp_node); |
4a00c761 | 6105 | } |
ebfd146a | 6106 | else |
4a00c761 JJ |
6107 | { |
6108 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
6109 | if (op_type == ternary_op) | |
6110 | { | |
9771b263 DN |
6111 | tree vec_oprnd = vec_oprnds2.pop (); |
6112 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
6113 | vec_oprnd)); | |
4a00c761 JJ |
6114 | } |
6115 | } | |
6116 | ||
6117 | /* Arguments are ready. Create the new vector stmt. */ | |
e1bd7296 | 6118 | stmt_vec_info new_stmt_info = NULL; |
9771b263 | 6119 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 6120 | { |
4a00c761 | 6121 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 6122 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 6123 | vop2 = ((op_type == ternary_op) |
9771b263 | 6124 | ? vec_oprnds2[i] : NULL_TREE); |
e1bd7296 RS |
6125 | gassign *new_stmt = gimple_build_assign (vec_dest, code, |
6126 | vop0, vop1, vop2); | |
4a00c761 JJ |
6127 | new_temp = make_ssa_name (vec_dest, new_stmt); |
6128 | gimple_assign_set_lhs (new_stmt, new_temp); | |
e1bd7296 | 6129 | new_stmt_info = vect_finish_stmt_generation (stmt, new_stmt, gsi); |
0eb952ea JJ |
6130 | if (vec_cvt_dest) |
6131 | { | |
6132 | new_temp = build1 (VIEW_CONVERT_EXPR, vectype_out, new_temp); | |
e1bd7296 RS |
6133 | gassign *new_stmt |
6134 | = gimple_build_assign (vec_cvt_dest, VIEW_CONVERT_EXPR, | |
6135 | new_temp); | |
0eb952ea JJ |
6136 | new_temp = make_ssa_name (vec_cvt_dest, new_stmt); |
6137 | gimple_assign_set_lhs (new_stmt, new_temp); | |
e1bd7296 RS |
6138 | new_stmt_info |
6139 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
0eb952ea | 6140 | } |
4a00c761 | 6141 | if (slp_node) |
e1bd7296 | 6142 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
ebfd146a IR |
6143 | } |
6144 | ||
4a00c761 JJ |
6145 | if (slp_node) |
6146 | continue; | |
6147 | ||
6148 | if (j == 0) | |
e1bd7296 | 6149 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
4a00c761 | 6150 | else |
e1bd7296 RS |
6151 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
6152 | prev_stmt_info = new_stmt_info; | |
ebfd146a IR |
6153 | } |
6154 | ||
9771b263 DN |
6155 | vec_oprnds0.release (); |
6156 | vec_oprnds1.release (); | |
6157 | vec_oprnds2.release (); | |
ebfd146a | 6158 | |
ebfd146a IR |
6159 | return true; |
6160 | } | |
6161 | ||
f702e7d4 | 6162 | /* A helper function to ensure data reference DR's base alignment. */ |
c716e67f XDL |
6163 | |
6164 | static void | |
f702e7d4 | 6165 | ensure_base_align (struct data_reference *dr) |
c716e67f | 6166 | { |
ca823c85 | 6167 | if (DR_VECT_AUX (dr)->misalignment == DR_MISALIGNMENT_UNINITIALIZED) |
c716e67f XDL |
6168 | return; |
6169 | ||
52639a61 | 6170 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f | 6171 | { |
52639a61 | 6172 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 6173 | |
f702e7d4 RS |
6174 | unsigned int align_base_to = DR_TARGET_ALIGNMENT (dr) * BITS_PER_UNIT; |
6175 | ||
428f0c67 | 6176 | if (decl_in_symtab_p (base_decl)) |
f702e7d4 | 6177 | symtab_node::get (base_decl)->increase_alignment (align_base_to); |
428f0c67 JH |
6178 | else |
6179 | { | |
f702e7d4 | 6180 | SET_DECL_ALIGN (base_decl, align_base_to); |
428f0c67 JH |
6181 | DECL_USER_ALIGN (base_decl) = 1; |
6182 | } | |
52639a61 | 6183 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
6184 | } |
6185 | } | |
6186 | ||
ebfd146a | 6187 | |
44fc7854 BE |
6188 | /* Function get_group_alias_ptr_type. |
6189 | ||
6190 | Return the alias type for the group starting at FIRST_STMT. */ | |
6191 | ||
6192 | static tree | |
6193 | get_group_alias_ptr_type (gimple *first_stmt) | |
6194 | { | |
6195 | struct data_reference *first_dr, *next_dr; | |
44fc7854 BE |
6196 | |
6197 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
bffb8014 RS |
6198 | stmt_vec_info next_stmt_info |
6199 | = DR_GROUP_NEXT_ELEMENT (vinfo_for_stmt (first_stmt)); | |
6200 | while (next_stmt_info) | |
44fc7854 | 6201 | { |
bffb8014 | 6202 | next_dr = STMT_VINFO_DATA_REF (next_stmt_info); |
44fc7854 BE |
6203 | if (get_alias_set (DR_REF (first_dr)) |
6204 | != get_alias_set (DR_REF (next_dr))) | |
6205 | { | |
6206 | if (dump_enabled_p ()) | |
6207 | dump_printf_loc (MSG_NOTE, vect_location, | |
6208 | "conflicting alias set types.\n"); | |
6209 | return ptr_type_node; | |
6210 | } | |
bffb8014 | 6211 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
44fc7854 BE |
6212 | } |
6213 | return reference_alias_ptr_type (DR_REF (first_dr)); | |
6214 | } | |
6215 | ||
6216 | ||
ebfd146a IR |
6217 | /* Function vectorizable_store. |
6218 | ||
b8698a0f L |
6219 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
6220 | can be vectorized. | |
6221 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6222 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6223 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6224 | ||
6225 | static bool | |
1eede195 RS |
6226 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, |
6227 | stmt_vec_info *vec_stmt, slp_tree slp_node, | |
6228 | stmt_vector_for_cost *cost_vec) | |
ebfd146a | 6229 | { |
ebfd146a IR |
6230 | tree data_ref; |
6231 | tree op; | |
6232 | tree vec_oprnd = NULL_TREE; | |
6233 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6234 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 6235 | tree elem_type; |
ebfd146a | 6236 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6237 | struct loop *loop = NULL; |
ef4bddc2 | 6238 | machine_mode vec_mode; |
ebfd146a IR |
6239 | tree dummy; |
6240 | enum dr_alignment_support alignment_support_scheme; | |
929b4411 RS |
6241 | enum vect_def_type rhs_dt = vect_unknown_def_type; |
6242 | enum vect_def_type mask_dt = vect_unknown_def_type; | |
ebfd146a IR |
6243 | stmt_vec_info prev_stmt_info = NULL; |
6244 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6245 | tree dataref_offset = NULL_TREE; |
355fe088 | 6246 | gimple *ptr_incr = NULL; |
ebfd146a IR |
6247 | int ncopies; |
6248 | int j; | |
bffb8014 | 6249 | stmt_vec_info first_stmt_info; |
2de001ee | 6250 | bool grouped_store; |
ebfd146a | 6251 | unsigned int group_size, i; |
6e1aa848 DN |
6252 | vec<tree> oprnds = vNULL; |
6253 | vec<tree> result_chain = vNULL; | |
ebfd146a | 6254 | bool inv_p; |
09dfa495 | 6255 | tree offset = NULL_TREE; |
6e1aa848 | 6256 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 6257 | bool slp = (slp_node != NULL); |
ebfd146a | 6258 | unsigned int vec_num; |
a70d6342 | 6259 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 6260 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 6261 | tree aggr_type; |
134c85ca | 6262 | gather_scatter_info gs_info; |
d9f21f6a | 6263 | poly_uint64 vf; |
2de001ee | 6264 | vec_load_store_type vls_type; |
44fc7854 | 6265 | tree ref_type; |
a70d6342 | 6266 | |
a70d6342 | 6267 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
6268 | return false; |
6269 | ||
66c16fd9 RB |
6270 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6271 | && ! vec_stmt) | |
ebfd146a IR |
6272 | return false; |
6273 | ||
6274 | /* Is vectorizable store? */ | |
6275 | ||
c3a8f964 RS |
6276 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
6277 | if (is_gimple_assign (stmt)) | |
6278 | { | |
6279 | tree scalar_dest = gimple_assign_lhs (stmt); | |
6280 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR | |
6281 | && is_pattern_stmt_p (stmt_info)) | |
6282 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
6283 | if (TREE_CODE (scalar_dest) != ARRAY_REF | |
6284 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF | |
6285 | && TREE_CODE (scalar_dest) != INDIRECT_REF | |
6286 | && TREE_CODE (scalar_dest) != COMPONENT_REF | |
6287 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
6288 | && TREE_CODE (scalar_dest) != REALPART_EXPR | |
6289 | && TREE_CODE (scalar_dest) != MEM_REF) | |
6290 | return false; | |
6291 | } | |
6292 | else | |
6293 | { | |
6294 | gcall *call = dyn_cast <gcall *> (stmt); | |
f307441a RS |
6295 | if (!call || !gimple_call_internal_p (call)) |
6296 | return false; | |
6297 | ||
6298 | internal_fn ifn = gimple_call_internal_fn (call); | |
6299 | if (!internal_store_fn_p (ifn)) | |
c3a8f964 | 6300 | return false; |
ebfd146a | 6301 | |
c3a8f964 RS |
6302 | if (slp_node != NULL) |
6303 | { | |
6304 | if (dump_enabled_p ()) | |
6305 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6306 | "SLP of masked stores not supported.\n"); | |
6307 | return false; | |
6308 | } | |
6309 | ||
f307441a RS |
6310 | int mask_index = internal_fn_mask_index (ifn); |
6311 | if (mask_index >= 0) | |
6312 | { | |
6313 | mask = gimple_call_arg (call, mask_index); | |
929b4411 RS |
6314 | if (!vect_check_load_store_mask (stmt, mask, &mask_dt, |
6315 | &mask_vectype)) | |
f307441a RS |
6316 | return false; |
6317 | } | |
c3a8f964 RS |
6318 | } |
6319 | ||
6320 | op = vect_get_store_rhs (stmt); | |
ebfd146a | 6321 | |
fce57248 RS |
6322 | /* Cannot have hybrid store SLP -- that would mean storing to the |
6323 | same location twice. */ | |
6324 | gcc_assert (slp == PURE_SLP_STMT (stmt_info)); | |
6325 | ||
f4d09712 | 6326 | tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE; |
4d694b27 | 6327 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 JJ |
6328 | |
6329 | if (loop_vinfo) | |
b17dc4d4 RB |
6330 | { |
6331 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6332 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6333 | } | |
6334 | else | |
6335 | vf = 1; | |
465c8c19 JJ |
6336 | |
6337 | /* Multiple types in SLP are handled by creating the appropriate number of | |
6338 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
6339 | case of SLP. */ | |
fce57248 | 6340 | if (slp) |
465c8c19 JJ |
6341 | ncopies = 1; |
6342 | else | |
e8f142e2 | 6343 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
6344 | |
6345 | gcc_assert (ncopies >= 1); | |
6346 | ||
6347 | /* FORNOW. This restriction should be relaxed. */ | |
6348 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
6349 | { | |
6350 | if (dump_enabled_p ()) | |
6351 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6352 | "multiple types in nested loop.\n"); | |
6353 | return false; | |
6354 | } | |
6355 | ||
929b4411 | 6356 | if (!vect_check_store_rhs (stmt, op, &rhs_dt, &rhs_vectype, &vls_type)) |
f4d09712 KY |
6357 | return false; |
6358 | ||
272c6793 | 6359 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 6360 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 6361 | |
ebfd146a IR |
6362 | if (!STMT_VINFO_DATA_REF (stmt_info)) |
6363 | return false; | |
6364 | ||
2de001ee | 6365 | vect_memory_access_type memory_access_type; |
7e11fc7f | 6366 | if (!get_load_store_type (stmt, vectype, slp, mask, vls_type, ncopies, |
2de001ee RS |
6367 | &memory_access_type, &gs_info)) |
6368 | return false; | |
3bab6342 | 6369 | |
c3a8f964 RS |
6370 | if (mask) |
6371 | { | |
7e11fc7f RS |
6372 | if (memory_access_type == VMAT_CONTIGUOUS) |
6373 | { | |
6374 | if (!VECTOR_MODE_P (vec_mode) | |
6375 | || !can_vec_mask_load_store_p (vec_mode, | |
6376 | TYPE_MODE (mask_vectype), false)) | |
6377 | return false; | |
6378 | } | |
f307441a RS |
6379 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
6380 | && (memory_access_type != VMAT_GATHER_SCATTER || gs_info.decl)) | |
c3a8f964 RS |
6381 | { |
6382 | if (dump_enabled_p ()) | |
6383 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6384 | "unsupported access type for masked store.\n"); | |
6385 | return false; | |
6386 | } | |
c3a8f964 RS |
6387 | } |
6388 | else | |
6389 | { | |
6390 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6391 | (e.g. - array initialization with 0). */ | |
6392 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) | |
6393 | return false; | |
6394 | } | |
6395 | ||
f307441a | 6396 | grouped_store = (STMT_VINFO_GROUPED_ACCESS (stmt_info) |
b5ec4de7 RS |
6397 | && memory_access_type != VMAT_GATHER_SCATTER |
6398 | && (slp || memory_access_type != VMAT_CONTIGUOUS)); | |
7cfb4d93 RS |
6399 | if (grouped_store) |
6400 | { | |
bffb8014 RS |
6401 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
6402 | first_dr = STMT_VINFO_DATA_REF (first_stmt_info); | |
6403 | group_size = DR_GROUP_SIZE (first_stmt_info); | |
7cfb4d93 RS |
6404 | } |
6405 | else | |
6406 | { | |
bffb8014 | 6407 | first_stmt_info = stmt_info; |
7cfb4d93 RS |
6408 | first_dr = dr; |
6409 | group_size = vec_num = 1; | |
6410 | } | |
6411 | ||
ebfd146a IR |
6412 | if (!vec_stmt) /* transformation not required. */ |
6413 | { | |
2de001ee | 6414 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
7cfb4d93 RS |
6415 | |
6416 | if (loop_vinfo | |
6417 | && LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) | |
6418 | check_load_store_masking (loop_vinfo, vectype, vls_type, group_size, | |
bfaa08b7 | 6419 | memory_access_type, &gs_info); |
7cfb4d93 | 6420 | |
ebfd146a | 6421 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; |
68435eb2 RB |
6422 | vect_model_store_cost (stmt_info, ncopies, rhs_dt, memory_access_type, |
6423 | vls_type, slp_node, cost_vec); | |
ebfd146a IR |
6424 | return true; |
6425 | } | |
2de001ee | 6426 | gcc_assert (memory_access_type == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
ebfd146a | 6427 | |
67b8dbac | 6428 | /* Transform. */ |
ebfd146a | 6429 | |
f702e7d4 | 6430 | ensure_base_align (dr); |
c716e67f | 6431 | |
f307441a | 6432 | if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
3bab6342 | 6433 | { |
c3a8f964 | 6434 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, src; |
134c85ca | 6435 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
6436 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; |
6437 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
6438 | edge pe = loop_preheader_edge (loop); | |
6439 | gimple_seq seq; | |
6440 | basic_block new_bb; | |
6441 | enum { NARROW, NONE, WIDEN } modifier; | |
4d694b27 RS |
6442 | poly_uint64 scatter_off_nunits |
6443 | = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype); | |
3bab6342 | 6444 | |
4d694b27 | 6445 | if (known_eq (nunits, scatter_off_nunits)) |
3bab6342 | 6446 | modifier = NONE; |
4d694b27 | 6447 | else if (known_eq (nunits * 2, scatter_off_nunits)) |
3bab6342 | 6448 | { |
3bab6342 AT |
6449 | modifier = WIDEN; |
6450 | ||
4d694b27 RS |
6451 | /* Currently gathers and scatters are only supported for |
6452 | fixed-length vectors. */ | |
6453 | unsigned int count = scatter_off_nunits.to_constant (); | |
6454 | vec_perm_builder sel (count, count, 1); | |
6455 | for (i = 0; i < (unsigned int) count; ++i) | |
6456 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 6457 | |
4d694b27 | 6458 | vec_perm_indices indices (sel, 1, count); |
e3342de4 RS |
6459 | perm_mask = vect_gen_perm_mask_checked (gs_info.offset_vectype, |
6460 | indices); | |
3bab6342 AT |
6461 | gcc_assert (perm_mask != NULL_TREE); |
6462 | } | |
4d694b27 | 6463 | else if (known_eq (nunits, scatter_off_nunits * 2)) |
3bab6342 | 6464 | { |
3bab6342 AT |
6465 | modifier = NARROW; |
6466 | ||
4d694b27 RS |
6467 | /* Currently gathers and scatters are only supported for |
6468 | fixed-length vectors. */ | |
6469 | unsigned int count = nunits.to_constant (); | |
6470 | vec_perm_builder sel (count, count, 1); | |
6471 | for (i = 0; i < (unsigned int) count; ++i) | |
6472 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 6473 | |
4d694b27 | 6474 | vec_perm_indices indices (sel, 2, count); |
e3342de4 | 6475 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); |
3bab6342 AT |
6476 | gcc_assert (perm_mask != NULL_TREE); |
6477 | ncopies *= 2; | |
6478 | } | |
6479 | else | |
6480 | gcc_unreachable (); | |
6481 | ||
134c85ca | 6482 | rettype = TREE_TYPE (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
6483 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
6484 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6485 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6486 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6487 | scaletype = TREE_VALUE (arglist); | |
6488 | ||
6489 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
6490 | && TREE_CODE (rettype) == VOID_TYPE); | |
6491 | ||
134c85ca | 6492 | ptr = fold_convert (ptrtype, gs_info.base); |
3bab6342 AT |
6493 | if (!is_gimple_min_invariant (ptr)) |
6494 | { | |
6495 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6496 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6497 | gcc_assert (!new_bb); | |
6498 | } | |
6499 | ||
6500 | /* Currently we support only unconditional scatter stores, | |
6501 | so mask should be all ones. */ | |
6502 | mask = build_int_cst (masktype, -1); | |
6503 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
6504 | ||
134c85ca | 6505 | scale = build_int_cst (scaletype, gs_info.scale); |
3bab6342 AT |
6506 | |
6507 | prev_stmt_info = NULL; | |
6508 | for (j = 0; j < ncopies; ++j) | |
6509 | { | |
6510 | if (j == 0) | |
6511 | { | |
6512 | src = vec_oprnd1 | |
c3a8f964 | 6513 | = vect_get_vec_def_for_operand (op, stmt); |
3bab6342 | 6514 | op = vec_oprnd0 |
134c85ca | 6515 | = vect_get_vec_def_for_operand (gs_info.offset, stmt); |
3bab6342 AT |
6516 | } |
6517 | else if (modifier != NONE && (j & 1)) | |
6518 | { | |
6519 | if (modifier == WIDEN) | |
6520 | { | |
6521 | src = vec_oprnd1 | |
929b4411 | 6522 | = vect_get_vec_def_for_stmt_copy (rhs_dt, vec_oprnd1); |
3bab6342 AT |
6523 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, |
6524 | stmt, gsi); | |
6525 | } | |
6526 | else if (modifier == NARROW) | |
6527 | { | |
6528 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
6529 | stmt, gsi); | |
6530 | op = vec_oprnd0 | |
134c85ca RS |
6531 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
6532 | vec_oprnd0); | |
3bab6342 AT |
6533 | } |
6534 | else | |
6535 | gcc_unreachable (); | |
6536 | } | |
6537 | else | |
6538 | { | |
6539 | src = vec_oprnd1 | |
929b4411 | 6540 | = vect_get_vec_def_for_stmt_copy (rhs_dt, vec_oprnd1); |
3bab6342 | 6541 | op = vec_oprnd0 |
134c85ca RS |
6542 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
6543 | vec_oprnd0); | |
3bab6342 AT |
6544 | } |
6545 | ||
6546 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
6547 | { | |
928686b1 RS |
6548 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)), |
6549 | TYPE_VECTOR_SUBPARTS (srctype))); | |
0e22bb5a | 6550 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 | 6551 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
e1bd7296 RS |
6552 | gassign *new_stmt |
6553 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
3bab6342 AT |
6554 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6555 | src = var; | |
6556 | } | |
6557 | ||
6558 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6559 | { | |
928686b1 RS |
6560 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), |
6561 | TYPE_VECTOR_SUBPARTS (idxtype))); | |
0e22bb5a | 6562 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 | 6563 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
e1bd7296 RS |
6564 | gassign *new_stmt |
6565 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
3bab6342 AT |
6566 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6567 | op = var; | |
6568 | } | |
6569 | ||
e1bd7296 | 6570 | gcall *new_stmt |
134c85ca | 6571 | = gimple_build_call (gs_info.decl, 5, ptr, mask, op, src, scale); |
e1bd7296 RS |
6572 | stmt_vec_info new_stmt_info |
6573 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3bab6342 | 6574 | |
dbe1b846 | 6575 | if (prev_stmt_info == NULL_STMT_VEC_INFO) |
e1bd7296 | 6576 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
3bab6342 | 6577 | else |
e1bd7296 RS |
6578 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
6579 | prev_stmt_info = new_stmt_info; | |
3bab6342 AT |
6580 | } |
6581 | return true; | |
6582 | } | |
6583 | ||
f307441a | 6584 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
bffb8014 | 6585 | DR_GROUP_STORE_COUNT (DR_GROUP_FIRST_ELEMENT (stmt_info))++; |
ebfd146a | 6586 | |
f307441a RS |
6587 | if (grouped_store) |
6588 | { | |
ebfd146a | 6589 | /* FORNOW */ |
a70d6342 | 6590 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
6591 | |
6592 | /* We vectorize all the stmts of the interleaving group when we | |
6593 | reach the last stmt in the group. */ | |
bffb8014 RS |
6594 | if (DR_GROUP_STORE_COUNT (first_stmt_info) |
6595 | < DR_GROUP_SIZE (first_stmt_info) | |
ebfd146a IR |
6596 | && !slp) |
6597 | { | |
6598 | *vec_stmt = NULL; | |
6599 | return true; | |
6600 | } | |
6601 | ||
6602 | if (slp) | |
4b5caab7 | 6603 | { |
0d0293ac | 6604 | grouped_store = false; |
4b5caab7 IR |
6605 | /* VEC_NUM is the number of vect stmts to be created for this |
6606 | group. */ | |
6607 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
bffb8014 RS |
6608 | first_stmt_info = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
6609 | gcc_assert (DR_GROUP_FIRST_ELEMENT (first_stmt_info) | |
6610 | == first_stmt_info); | |
6611 | first_dr = STMT_VINFO_DATA_REF (first_stmt_info); | |
6612 | op = vect_get_store_rhs (first_stmt_info); | |
4b5caab7 | 6613 | } |
ebfd146a | 6614 | else |
4b5caab7 IR |
6615 | /* VEC_NUM is the number of vect stmts to be created for this |
6616 | group. */ | |
ebfd146a | 6617 | vec_num = group_size; |
44fc7854 | 6618 | |
bffb8014 | 6619 | ref_type = get_group_alias_ptr_type (first_stmt_info); |
ebfd146a | 6620 | } |
b8698a0f | 6621 | else |
7cfb4d93 | 6622 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
b8698a0f | 6623 | |
73fbfcad | 6624 | if (dump_enabled_p ()) |
78c60e3d | 6625 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6626 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 6627 | |
2de001ee RS |
6628 | if (memory_access_type == VMAT_ELEMENTWISE |
6629 | || memory_access_type == VMAT_STRIDED_SLP) | |
f2e2a985 MM |
6630 | { |
6631 | gimple_stmt_iterator incr_gsi; | |
6632 | bool insert_after; | |
355fe088 | 6633 | gimple *incr; |
f2e2a985 MM |
6634 | tree offvar; |
6635 | tree ivstep; | |
6636 | tree running_off; | |
f2e2a985 MM |
6637 | tree stride_base, stride_step, alias_off; |
6638 | tree vec_oprnd; | |
f502d50e | 6639 | unsigned int g; |
4d694b27 RS |
6640 | /* Checked by get_load_store_type. */ |
6641 | unsigned int const_nunits = nunits.to_constant (); | |
f2e2a985 | 6642 | |
7cfb4d93 | 6643 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
f2e2a985 MM |
6644 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); |
6645 | ||
6646 | stride_base | |
6647 | = fold_build_pointer_plus | |
b210f45f | 6648 | (DR_BASE_ADDRESS (first_dr), |
f2e2a985 | 6649 | size_binop (PLUS_EXPR, |
b210f45f | 6650 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
44fc7854 | 6651 | convert_to_ptrofftype (DR_INIT (first_dr)))); |
b210f45f | 6652 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); |
f2e2a985 MM |
6653 | |
6654 | /* For a store with loop-invariant (but other than power-of-2) | |
6655 | stride (i.e. not a grouped access) like so: | |
6656 | ||
6657 | for (i = 0; i < n; i += stride) | |
6658 | array[i] = ...; | |
6659 | ||
6660 | we generate a new induction variable and new stores from | |
6661 | the components of the (vectorized) rhs: | |
6662 | ||
6663 | for (j = 0; ; j += VF*stride) | |
6664 | vectemp = ...; | |
6665 | tmp1 = vectemp[0]; | |
6666 | array[j] = tmp1; | |
6667 | tmp2 = vectemp[1]; | |
6668 | array[j + stride] = tmp2; | |
6669 | ... | |
6670 | */ | |
6671 | ||
4d694b27 | 6672 | unsigned nstores = const_nunits; |
b17dc4d4 | 6673 | unsigned lnel = 1; |
cee62fee | 6674 | tree ltype = elem_type; |
04199738 | 6675 | tree lvectype = vectype; |
cee62fee MM |
6676 | if (slp) |
6677 | { | |
4d694b27 RS |
6678 | if (group_size < const_nunits |
6679 | && const_nunits % group_size == 0) | |
b17dc4d4 | 6680 | { |
4d694b27 | 6681 | nstores = const_nunits / group_size; |
b17dc4d4 RB |
6682 | lnel = group_size; |
6683 | ltype = build_vector_type (elem_type, group_size); | |
04199738 RB |
6684 | lvectype = vectype; |
6685 | ||
6686 | /* First check if vec_extract optab doesn't support extraction | |
6687 | of vector elts directly. */ | |
b397965c | 6688 | scalar_mode elmode = SCALAR_TYPE_MODE (elem_type); |
9da15d40 RS |
6689 | machine_mode vmode; |
6690 | if (!mode_for_vector (elmode, group_size).exists (&vmode) | |
6691 | || !VECTOR_MODE_P (vmode) | |
414fef4e | 6692 | || !targetm.vector_mode_supported_p (vmode) |
04199738 RB |
6693 | || (convert_optab_handler (vec_extract_optab, |
6694 | TYPE_MODE (vectype), vmode) | |
6695 | == CODE_FOR_nothing)) | |
6696 | { | |
6697 | /* Try to avoid emitting an extract of vector elements | |
6698 | by performing the extracts using an integer type of the | |
6699 | same size, extracting from a vector of those and then | |
6700 | re-interpreting it as the original vector type if | |
6701 | supported. */ | |
6702 | unsigned lsize | |
6703 | = group_size * GET_MODE_BITSIZE (elmode); | |
fffbab82 | 6704 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 6705 | unsigned int lnunits = const_nunits / group_size; |
04199738 RB |
6706 | /* If we can't construct such a vector fall back to |
6707 | element extracts from the original vector type and | |
6708 | element size stores. */ | |
4d694b27 | 6709 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 6710 | && VECTOR_MODE_P (vmode) |
414fef4e | 6711 | && targetm.vector_mode_supported_p (vmode) |
04199738 RB |
6712 | && (convert_optab_handler (vec_extract_optab, |
6713 | vmode, elmode) | |
6714 | != CODE_FOR_nothing)) | |
6715 | { | |
4d694b27 | 6716 | nstores = lnunits; |
04199738 RB |
6717 | lnel = group_size; |
6718 | ltype = build_nonstandard_integer_type (lsize, 1); | |
6719 | lvectype = build_vector_type (ltype, nstores); | |
6720 | } | |
6721 | /* Else fall back to vector extraction anyway. | |
6722 | Fewer stores are more important than avoiding spilling | |
6723 | of the vector we extract from. Compared to the | |
6724 | construction case in vectorizable_load no store-forwarding | |
6725 | issue exists here for reasonable archs. */ | |
6726 | } | |
b17dc4d4 | 6727 | } |
4d694b27 RS |
6728 | else if (group_size >= const_nunits |
6729 | && group_size % const_nunits == 0) | |
b17dc4d4 RB |
6730 | { |
6731 | nstores = 1; | |
4d694b27 | 6732 | lnel = const_nunits; |
b17dc4d4 | 6733 | ltype = vectype; |
04199738 | 6734 | lvectype = vectype; |
b17dc4d4 | 6735 | } |
cee62fee MM |
6736 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); |
6737 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
6738 | } | |
6739 | ||
f2e2a985 MM |
6740 | ivstep = stride_step; |
6741 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
b17dc4d4 | 6742 | build_int_cst (TREE_TYPE (ivstep), vf)); |
f2e2a985 MM |
6743 | |
6744 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6745 | ||
b210f45f RB |
6746 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
6747 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); | |
f2e2a985 MM |
6748 | create_iv (stride_base, ivstep, NULL, |
6749 | loop, &incr_gsi, insert_after, | |
6750 | &offvar, NULL); | |
6751 | incr = gsi_stmt (incr_gsi); | |
4fbeb363 | 6752 | loop_vinfo->add_stmt (incr); |
f2e2a985 | 6753 | |
b210f45f | 6754 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
f2e2a985 MM |
6755 | |
6756 | prev_stmt_info = NULL; | |
44fc7854 | 6757 | alias_off = build_int_cst (ref_type, 0); |
bffb8014 | 6758 | stmt_vec_info next_stmt_info = first_stmt_info; |
f502d50e | 6759 | for (g = 0; g < group_size; g++) |
f2e2a985 | 6760 | { |
f502d50e MM |
6761 | running_off = offvar; |
6762 | if (g) | |
f2e2a985 | 6763 | { |
f502d50e MM |
6764 | tree size = TYPE_SIZE_UNIT (ltype); |
6765 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 6766 | size); |
f502d50e | 6767 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 6768 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 6769 | running_off, pos); |
f2e2a985 | 6770 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 6771 | running_off = newoff; |
f502d50e | 6772 | } |
b17dc4d4 RB |
6773 | unsigned int group_el = 0; |
6774 | unsigned HOST_WIDE_INT | |
6775 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
f502d50e MM |
6776 | for (j = 0; j < ncopies; j++) |
6777 | { | |
c3a8f964 | 6778 | /* We've set op and dt above, from vect_get_store_rhs, |
bffb8014 | 6779 | and first_stmt_info == stmt_info. */ |
f502d50e MM |
6780 | if (j == 0) |
6781 | { | |
6782 | if (slp) | |
6783 | { | |
6784 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
306b0c92 | 6785 | slp_node); |
f502d50e MM |
6786 | vec_oprnd = vec_oprnds[0]; |
6787 | } | |
6788 | else | |
6789 | { | |
bffb8014 RS |
6790 | op = vect_get_store_rhs (next_stmt_info); |
6791 | vec_oprnd = vect_get_vec_def_for_operand | |
6792 | (op, next_stmt_info); | |
f502d50e MM |
6793 | } |
6794 | } | |
f2e2a985 | 6795 | else |
f502d50e MM |
6796 | { |
6797 | if (slp) | |
6798 | vec_oprnd = vec_oprnds[j]; | |
6799 | else | |
c079cbac | 6800 | { |
894dd753 | 6801 | vect_is_simple_use (op, vinfo, &rhs_dt); |
929b4411 RS |
6802 | vec_oprnd = vect_get_vec_def_for_stmt_copy (rhs_dt, |
6803 | vec_oprnd); | |
c079cbac | 6804 | } |
f502d50e | 6805 | } |
04199738 RB |
6806 | /* Pun the vector to extract from if necessary. */ |
6807 | if (lvectype != vectype) | |
6808 | { | |
6809 | tree tem = make_ssa_name (lvectype); | |
6810 | gimple *pun | |
6811 | = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
6812 | lvectype, vec_oprnd)); | |
6813 | vect_finish_stmt_generation (stmt, pun, gsi); | |
6814 | vec_oprnd = tem; | |
6815 | } | |
f502d50e MM |
6816 | for (i = 0; i < nstores; i++) |
6817 | { | |
6818 | tree newref, newoff; | |
355fe088 | 6819 | gimple *incr, *assign; |
f502d50e MM |
6820 | tree size = TYPE_SIZE (ltype); |
6821 | /* Extract the i'th component. */ | |
6822 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
6823 | bitsize_int (i), size); | |
6824 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
6825 | size, pos); | |
6826 | ||
6827 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
6828 | NULL_TREE, true, | |
6829 | GSI_SAME_STMT); | |
6830 | ||
b17dc4d4 RB |
6831 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
6832 | group_el * elsz); | |
f502d50e | 6833 | newref = build2 (MEM_REF, ltype, |
b17dc4d4 | 6834 | running_off, this_off); |
19986382 | 6835 | vect_copy_ref_info (newref, DR_REF (first_dr)); |
f502d50e MM |
6836 | |
6837 | /* And store it to *running_off. */ | |
6838 | assign = gimple_build_assign (newref, elem); | |
e1bd7296 RS |
6839 | stmt_vec_info assign_info |
6840 | = vect_finish_stmt_generation (stmt, assign, gsi); | |
f502d50e | 6841 | |
b17dc4d4 RB |
6842 | group_el += lnel; |
6843 | if (! slp | |
6844 | || group_el == group_size) | |
6845 | { | |
6846 | newoff = copy_ssa_name (running_off, NULL); | |
6847 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6848 | running_off, stride_step); | |
6849 | vect_finish_stmt_generation (stmt, incr, gsi); | |
f502d50e | 6850 | |
b17dc4d4 RB |
6851 | running_off = newoff; |
6852 | group_el = 0; | |
6853 | } | |
225ce44b RB |
6854 | if (g == group_size - 1 |
6855 | && !slp) | |
f502d50e MM |
6856 | { |
6857 | if (j == 0 && i == 0) | |
225ce44b | 6858 | STMT_VINFO_VEC_STMT (stmt_info) |
e1bd7296 | 6859 | = *vec_stmt = assign_info; |
f502d50e | 6860 | else |
e1bd7296 RS |
6861 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign_info; |
6862 | prev_stmt_info = assign_info; | |
f502d50e MM |
6863 | } |
6864 | } | |
f2e2a985 | 6865 | } |
bffb8014 | 6866 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
b17dc4d4 RB |
6867 | if (slp) |
6868 | break; | |
f2e2a985 | 6869 | } |
778dd3b6 RB |
6870 | |
6871 | vec_oprnds.release (); | |
f2e2a985 MM |
6872 | return true; |
6873 | } | |
6874 | ||
8c681247 | 6875 | auto_vec<tree> dr_chain (group_size); |
9771b263 | 6876 | oprnds.create (group_size); |
ebfd146a | 6877 | |
720f5239 | 6878 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6879 | gcc_assert (alignment_support_scheme); |
70088b95 RS |
6880 | vec_loop_masks *loop_masks |
6881 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) | |
6882 | ? &LOOP_VINFO_MASKS (loop_vinfo) | |
6883 | : NULL); | |
272c6793 | 6884 | /* Targets with store-lane instructions must not require explicit |
c3a8f964 RS |
6885 | realignment. vect_supportable_dr_alignment always returns either |
6886 | dr_aligned or dr_unaligned_supported for masked operations. */ | |
7cfb4d93 RS |
6887 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES |
6888 | && !mask | |
70088b95 | 6889 | && !loop_masks) |
272c6793 RS |
6890 | || alignment_support_scheme == dr_aligned |
6891 | || alignment_support_scheme == dr_unaligned_supported); | |
6892 | ||
62da9e14 RS |
6893 | if (memory_access_type == VMAT_CONTIGUOUS_DOWN |
6894 | || memory_access_type == VMAT_CONTIGUOUS_REVERSE) | |
09dfa495 BM |
6895 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
6896 | ||
f307441a RS |
6897 | tree bump; |
6898 | tree vec_offset = NULL_TREE; | |
6899 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
6900 | { | |
6901 | aggr_type = NULL_TREE; | |
6902 | bump = NULL_TREE; | |
6903 | } | |
6904 | else if (memory_access_type == VMAT_GATHER_SCATTER) | |
6905 | { | |
6906 | aggr_type = elem_type; | |
6907 | vect_get_strided_load_store_ops (stmt, loop_vinfo, &gs_info, | |
6908 | &bump, &vec_offset); | |
6909 | } | |
272c6793 | 6910 | else |
f307441a RS |
6911 | { |
6912 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
6913 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
6914 | else | |
6915 | aggr_type = vectype; | |
6916 | bump = vect_get_data_ptr_increment (dr, aggr_type, memory_access_type); | |
6917 | } | |
ebfd146a | 6918 | |
c3a8f964 RS |
6919 | if (mask) |
6920 | LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true; | |
6921 | ||
ebfd146a IR |
6922 | /* In case the vectorization factor (VF) is bigger than the number |
6923 | of elements that we can fit in a vectype (nunits), we have to generate | |
6924 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 6925 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
6926 | vect_get_vec_def_for_copy_stmt. */ |
6927 | ||
0d0293ac | 6928 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6929 | |
6930 | S1: &base + 2 = x2 | |
6931 | S2: &base = x0 | |
6932 | S3: &base + 1 = x1 | |
6933 | S4: &base + 3 = x3 | |
6934 | ||
6935 | We create vectorized stores starting from base address (the access of the | |
6936 | first stmt in the chain (S2 in the above example), when the last store stmt | |
6937 | of the chain (S4) is reached: | |
6938 | ||
6939 | VS1: &base = vx2 | |
6940 | VS2: &base + vec_size*1 = vx0 | |
6941 | VS3: &base + vec_size*2 = vx1 | |
6942 | VS4: &base + vec_size*3 = vx3 | |
6943 | ||
6944 | Then permutation statements are generated: | |
6945 | ||
3fcc1b55 JJ |
6946 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
6947 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 6948 | ... |
b8698a0f | 6949 | |
ebfd146a IR |
6950 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
6951 | (the order of the data-refs in the output of vect_permute_store_chain | |
6952 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6953 | the documentation of vect_permute_store_chain()). | |
6954 | ||
6955 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 6956 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 6957 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 6958 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
6959 | */ |
6960 | ||
6961 | prev_stmt_info = NULL; | |
c3a8f964 | 6962 | tree vec_mask = NULL_TREE; |
ebfd146a IR |
6963 | for (j = 0; j < ncopies; j++) |
6964 | { | |
e1bd7296 | 6965 | stmt_vec_info new_stmt_info; |
ebfd146a IR |
6966 | if (j == 0) |
6967 | { | |
6968 | if (slp) | |
6969 | { | |
6970 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c | 6971 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
306b0c92 | 6972 | NULL, slp_node); |
ebfd146a | 6973 | |
9771b263 | 6974 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
6975 | } |
6976 | else | |
6977 | { | |
b8698a0f L |
6978 | /* For interleaved stores we collect vectorized defs for all the |
6979 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
6980 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
6981 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
6982 | ||
2c53b149 | 6983 | If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 6984 | OPRNDS are of size 1. */ |
bffb8014 | 6985 | stmt_vec_info next_stmt_info = first_stmt_info; |
ebfd146a IR |
6986 | for (i = 0; i < group_size; i++) |
6987 | { | |
b8698a0f | 6988 | /* Since gaps are not supported for interleaved stores, |
2c53b149 | 6989 | DR_GROUP_SIZE is the exact number of stmts in the chain. |
bffb8014 RS |
6990 | Therefore, NEXT_STMT_INFO can't be NULL_TREE. In case |
6991 | that there is no interleaving, DR_GROUP_SIZE is 1, | |
6992 | and only one iteration of the loop will be executed. */ | |
6993 | op = vect_get_store_rhs (next_stmt_info); | |
6994 | vec_oprnd = vect_get_vec_def_for_operand | |
6995 | (op, next_stmt_info); | |
9771b263 DN |
6996 | dr_chain.quick_push (vec_oprnd); |
6997 | oprnds.quick_push (vec_oprnd); | |
bffb8014 | 6998 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
ebfd146a | 6999 | } |
c3a8f964 RS |
7000 | if (mask) |
7001 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, | |
7002 | mask_vectype); | |
ebfd146a IR |
7003 | } |
7004 | ||
7005 | /* We should have catched mismatched types earlier. */ | |
7006 | gcc_assert (useless_type_conversion_p (vectype, | |
7007 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
7008 | bool simd_lane_access_p |
7009 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
7010 | if (simd_lane_access_p | |
7011 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
7012 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
7013 | && integer_zerop (DR_OFFSET (first_dr)) | |
7014 | && integer_zerop (DR_INIT (first_dr)) | |
7015 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 7016 | get_alias_set (TREE_TYPE (ref_type)))) |
74bf76ed JJ |
7017 | { |
7018 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 7019 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 7020 | inv_p = false; |
74bf76ed | 7021 | } |
f307441a RS |
7022 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
7023 | { | |
7024 | vect_get_gather_scatter_ops (loop, stmt, &gs_info, | |
7025 | &dataref_ptr, &vec_offset); | |
7026 | inv_p = false; | |
7027 | } | |
74bf76ed JJ |
7028 | else |
7029 | dataref_ptr | |
bffb8014 | 7030 | = vect_create_data_ref_ptr (first_stmt_info, aggr_type, |
74bf76ed | 7031 | simd_lane_access_p ? loop : NULL, |
09dfa495 | 7032 | offset, &dummy, gsi, &ptr_incr, |
f307441a RS |
7033 | simd_lane_access_p, &inv_p, |
7034 | NULL_TREE, bump); | |
a70d6342 | 7035 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 7036 | } |
b8698a0f | 7037 | else |
ebfd146a | 7038 | { |
b8698a0f L |
7039 | /* For interleaved stores we created vectorized defs for all the |
7040 | defs stored in OPRNDS in the previous iteration (previous copy). | |
7041 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
7042 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
7043 | next copy. | |
2c53b149 | 7044 | If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
7045 | OPRNDS are of size 1. */ |
7046 | for (i = 0; i < group_size; i++) | |
7047 | { | |
9771b263 | 7048 | op = oprnds[i]; |
894dd753 | 7049 | vect_is_simple_use (op, vinfo, &rhs_dt); |
929b4411 | 7050 | vec_oprnd = vect_get_vec_def_for_stmt_copy (rhs_dt, op); |
9771b263 DN |
7051 | dr_chain[i] = vec_oprnd; |
7052 | oprnds[i] = vec_oprnd; | |
ebfd146a | 7053 | } |
c3a8f964 | 7054 | if (mask) |
929b4411 | 7055 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
74bf76ed JJ |
7056 | if (dataref_offset) |
7057 | dataref_offset | |
f307441a RS |
7058 | = int_const_binop (PLUS_EXPR, dataref_offset, bump); |
7059 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
929b4411 RS |
7060 | vec_offset = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
7061 | vec_offset); | |
74bf76ed JJ |
7062 | else |
7063 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
f307441a | 7064 | bump); |
ebfd146a IR |
7065 | } |
7066 | ||
2de001ee | 7067 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 7068 | { |
272c6793 | 7069 | tree vec_array; |
267d3070 | 7070 | |
3ba4ff41 | 7071 | /* Get an array into which we can store the individual vectors. */ |
272c6793 | 7072 | vec_array = create_vector_array (vectype, vec_num); |
3ba4ff41 RS |
7073 | |
7074 | /* Invalidate the current contents of VEC_ARRAY. This should | |
7075 | become an RTL clobber too, which prevents the vector registers | |
7076 | from being upward-exposed. */ | |
7077 | vect_clobber_variable (stmt, gsi, vec_array); | |
7078 | ||
7079 | /* Store the individual vectors into the array. */ | |
272c6793 | 7080 | for (i = 0; i < vec_num; i++) |
c2d7ab2a | 7081 | { |
9771b263 | 7082 | vec_oprnd = dr_chain[i]; |
272c6793 | 7083 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 7084 | } |
b8698a0f | 7085 | |
7cfb4d93 | 7086 | tree final_mask = NULL; |
70088b95 RS |
7087 | if (loop_masks) |
7088 | final_mask = vect_get_loop_mask (gsi, loop_masks, ncopies, | |
7089 | vectype, j); | |
7cfb4d93 RS |
7090 | if (vec_mask) |
7091 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
7092 | vec_mask, gsi); | |
7093 | ||
7e11fc7f | 7094 | gcall *call; |
7cfb4d93 | 7095 | if (final_mask) |
7e11fc7f RS |
7096 | { |
7097 | /* Emit: | |
7098 | MASK_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK, | |
7099 | VEC_ARRAY). */ | |
7100 | unsigned int align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype)); | |
7101 | tree alias_ptr = build_int_cst (ref_type, align); | |
7102 | call = gimple_build_call_internal (IFN_MASK_STORE_LANES, 4, | |
7103 | dataref_ptr, alias_ptr, | |
7cfb4d93 | 7104 | final_mask, vec_array); |
7e11fc7f RS |
7105 | } |
7106 | else | |
7107 | { | |
7108 | /* Emit: | |
7109 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
7110 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); | |
7111 | call = gimple_build_call_internal (IFN_STORE_LANES, 1, | |
7112 | vec_array); | |
7113 | gimple_call_set_lhs (call, data_ref); | |
7114 | } | |
a844293d | 7115 | gimple_call_set_nothrow (call, true); |
e1bd7296 | 7116 | new_stmt_info = vect_finish_stmt_generation (stmt, call, gsi); |
3ba4ff41 RS |
7117 | |
7118 | /* Record that VEC_ARRAY is now dead. */ | |
7119 | vect_clobber_variable (stmt, gsi, vec_array); | |
272c6793 RS |
7120 | } |
7121 | else | |
7122 | { | |
e1bd7296 | 7123 | new_stmt_info = NULL; |
0d0293ac | 7124 | if (grouped_store) |
272c6793 | 7125 | { |
b6b9227d JJ |
7126 | if (j == 0) |
7127 | result_chain.create (group_size); | |
272c6793 RS |
7128 | /* Permute. */ |
7129 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
7130 | &result_chain); | |
7131 | } | |
c2d7ab2a | 7132 | |
bffb8014 | 7133 | stmt_vec_info next_stmt_info = first_stmt_info; |
272c6793 RS |
7134 | for (i = 0; i < vec_num; i++) |
7135 | { | |
644ffefd | 7136 | unsigned align, misalign; |
272c6793 | 7137 | |
7cfb4d93 | 7138 | tree final_mask = NULL_TREE; |
70088b95 RS |
7139 | if (loop_masks) |
7140 | final_mask = vect_get_loop_mask (gsi, loop_masks, | |
7141 | vec_num * ncopies, | |
7cfb4d93 RS |
7142 | vectype, vec_num * j + i); |
7143 | if (vec_mask) | |
7144 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
7145 | vec_mask, gsi); | |
7146 | ||
f307441a RS |
7147 | if (memory_access_type == VMAT_GATHER_SCATTER) |
7148 | { | |
7149 | tree scale = size_int (gs_info.scale); | |
7150 | gcall *call; | |
70088b95 | 7151 | if (loop_masks) |
f307441a RS |
7152 | call = gimple_build_call_internal |
7153 | (IFN_MASK_SCATTER_STORE, 5, dataref_ptr, vec_offset, | |
7154 | scale, vec_oprnd, final_mask); | |
7155 | else | |
7156 | call = gimple_build_call_internal | |
7157 | (IFN_SCATTER_STORE, 4, dataref_ptr, vec_offset, | |
7158 | scale, vec_oprnd); | |
7159 | gimple_call_set_nothrow (call, true); | |
e1bd7296 RS |
7160 | new_stmt_info |
7161 | = vect_finish_stmt_generation (stmt, call, gsi); | |
f307441a RS |
7162 | break; |
7163 | } | |
7164 | ||
272c6793 RS |
7165 | if (i > 0) |
7166 | /* Bump the vector pointer. */ | |
7167 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
f307441a | 7168 | stmt, bump); |
272c6793 RS |
7169 | |
7170 | if (slp) | |
9771b263 | 7171 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
7172 | else if (grouped_store) |
7173 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 7174 | vect_permute_store_chain(). */ |
9771b263 | 7175 | vec_oprnd = result_chain[i]; |
272c6793 | 7176 | |
f702e7d4 | 7177 | align = DR_TARGET_ALIGNMENT (first_dr); |
272c6793 | 7178 | if (aligned_access_p (first_dr)) |
644ffefd | 7179 | misalign = 0; |
272c6793 RS |
7180 | else if (DR_MISALIGNMENT (first_dr) == -1) |
7181 | { | |
25f68d90 | 7182 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 7183 | misalign = 0; |
272c6793 RS |
7184 | } |
7185 | else | |
c3a8f964 | 7186 | misalign = DR_MISALIGNMENT (first_dr); |
aed93b23 RB |
7187 | if (dataref_offset == NULL_TREE |
7188 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
7189 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
7190 | misalign); | |
c2d7ab2a | 7191 | |
62da9e14 | 7192 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
09dfa495 BM |
7193 | { |
7194 | tree perm_mask = perm_mask_for_reverse (vectype); | |
7195 | tree perm_dest | |
c3a8f964 | 7196 | = vect_create_destination_var (vect_get_store_rhs (stmt), |
09dfa495 | 7197 | vectype); |
b731b390 | 7198 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
7199 | |
7200 | /* Generate the permute statement. */ | |
355fe088 | 7201 | gimple *perm_stmt |
0d0e4a03 JJ |
7202 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
7203 | vec_oprnd, perm_mask); | |
09dfa495 BM |
7204 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
7205 | ||
7206 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
7207 | vec_oprnd = new_temp; | |
7208 | } | |
7209 | ||
272c6793 | 7210 | /* Arguments are ready. Create the new vector stmt. */ |
7cfb4d93 | 7211 | if (final_mask) |
c3a8f964 RS |
7212 | { |
7213 | align = least_bit_hwi (misalign | align); | |
7214 | tree ptr = build_int_cst (ref_type, align); | |
7215 | gcall *call | |
7216 | = gimple_build_call_internal (IFN_MASK_STORE, 4, | |
7217 | dataref_ptr, ptr, | |
7cfb4d93 | 7218 | final_mask, vec_oprnd); |
c3a8f964 | 7219 | gimple_call_set_nothrow (call, true); |
e1bd7296 RS |
7220 | new_stmt_info |
7221 | = vect_finish_stmt_generation (stmt, call, gsi); | |
c3a8f964 RS |
7222 | } |
7223 | else | |
7224 | { | |
7225 | data_ref = fold_build2 (MEM_REF, vectype, | |
7226 | dataref_ptr, | |
7227 | dataref_offset | |
7228 | ? dataref_offset | |
7229 | : build_int_cst (ref_type, 0)); | |
7230 | if (aligned_access_p (first_dr)) | |
7231 | ; | |
7232 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
7233 | TREE_TYPE (data_ref) | |
7234 | = build_aligned_type (TREE_TYPE (data_ref), | |
7235 | align * BITS_PER_UNIT); | |
7236 | else | |
7237 | TREE_TYPE (data_ref) | |
7238 | = build_aligned_type (TREE_TYPE (data_ref), | |
7239 | TYPE_ALIGN (elem_type)); | |
19986382 | 7240 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); |
e1bd7296 RS |
7241 | gassign *new_stmt |
7242 | = gimple_build_assign (data_ref, vec_oprnd); | |
7243 | new_stmt_info | |
7244 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
c3a8f964 | 7245 | } |
272c6793 RS |
7246 | |
7247 | if (slp) | |
7248 | continue; | |
7249 | ||
bffb8014 RS |
7250 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
7251 | if (!next_stmt_info) | |
272c6793 RS |
7252 | break; |
7253 | } | |
ebfd146a | 7254 | } |
1da0876c RS |
7255 | if (!slp) |
7256 | { | |
7257 | if (j == 0) | |
e1bd7296 | 7258 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
1da0876c | 7259 | else |
e1bd7296 RS |
7260 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
7261 | prev_stmt_info = new_stmt_info; | |
1da0876c | 7262 | } |
ebfd146a IR |
7263 | } |
7264 | ||
9771b263 DN |
7265 | oprnds.release (); |
7266 | result_chain.release (); | |
7267 | vec_oprnds.release (); | |
ebfd146a IR |
7268 | |
7269 | return true; | |
7270 | } | |
7271 | ||
557be5a8 AL |
7272 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
7273 | VECTOR_CST mask. No checks are made that the target platform supports the | |
7ac7e286 | 7274 | mask, so callers may wish to test can_vec_perm_const_p separately, or use |
557be5a8 | 7275 | vect_gen_perm_mask_checked. */ |
a1e53f3f | 7276 | |
3fcc1b55 | 7277 | tree |
4aae3cb3 | 7278 | vect_gen_perm_mask_any (tree vectype, const vec_perm_indices &sel) |
a1e53f3f | 7279 | { |
b00cb3bf | 7280 | tree mask_type; |
a1e53f3f | 7281 | |
0ecc2b7d RS |
7282 | poly_uint64 nunits = sel.length (); |
7283 | gcc_assert (known_eq (nunits, TYPE_VECTOR_SUBPARTS (vectype))); | |
b00cb3bf RS |
7284 | |
7285 | mask_type = build_vector_type (ssizetype, nunits); | |
736d0f28 | 7286 | return vec_perm_indices_to_tree (mask_type, sel); |
a1e53f3f L |
7287 | } |
7288 | ||
7ac7e286 | 7289 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_const_p, |
cf7aa6a3 | 7290 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ |
557be5a8 AL |
7291 | |
7292 | tree | |
4aae3cb3 | 7293 | vect_gen_perm_mask_checked (tree vectype, const vec_perm_indices &sel) |
557be5a8 | 7294 | { |
7ac7e286 | 7295 | gcc_assert (can_vec_perm_const_p (TYPE_MODE (vectype), sel)); |
557be5a8 AL |
7296 | return vect_gen_perm_mask_any (vectype, sel); |
7297 | } | |
7298 | ||
aec7ae7d JJ |
7299 | /* Given a vector variable X and Y, that was generated for the scalar |
7300 | STMT, generate instructions to permute the vector elements of X and Y | |
7301 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
7302 | permuted vector variable. */ | |
a1e53f3f L |
7303 | |
7304 | static tree | |
355fe088 | 7305 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 7306 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
7307 | { |
7308 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 7309 | tree perm_dest, data_ref; |
355fe088 | 7310 | gimple *perm_stmt; |
a1e53f3f | 7311 | |
7ad429a4 RS |
7312 | tree scalar_dest = gimple_get_lhs (stmt); |
7313 | if (TREE_CODE (scalar_dest) == SSA_NAME) | |
7314 | perm_dest = vect_create_destination_var (scalar_dest, vectype); | |
7315 | else | |
7316 | perm_dest = vect_get_new_vect_var (vectype, vect_simple_var, NULL); | |
b731b390 | 7317 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
7318 | |
7319 | /* Generate the permute statement. */ | |
0d0e4a03 | 7320 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
7321 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
7322 | ||
7323 | return data_ref; | |
7324 | } | |
7325 | ||
6b916b36 RB |
7326 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
7327 | inserting them on the loops preheader edge. Returns true if we | |
7328 | were successful in doing so (and thus STMT can be moved then), | |
7329 | otherwise returns false. */ | |
7330 | ||
7331 | static bool | |
355fe088 | 7332 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
7333 | { |
7334 | ssa_op_iter i; | |
7335 | tree op; | |
7336 | bool any = false; | |
7337 | ||
7338 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
7339 | { | |
355fe088 | 7340 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
7341 | if (!gimple_nop_p (def_stmt) |
7342 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
7343 | { | |
7344 | /* Make sure we don't need to recurse. While we could do | |
7345 | so in simple cases when there are more complex use webs | |
7346 | we don't have an easy way to preserve stmt order to fulfil | |
7347 | dependencies within them. */ | |
7348 | tree op2; | |
7349 | ssa_op_iter i2; | |
d1417442 JJ |
7350 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
7351 | return false; | |
6b916b36 RB |
7352 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
7353 | { | |
355fe088 | 7354 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
7355 | if (!gimple_nop_p (def_stmt2) |
7356 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
7357 | return false; | |
7358 | } | |
7359 | any = true; | |
7360 | } | |
7361 | } | |
7362 | ||
7363 | if (!any) | |
7364 | return true; | |
7365 | ||
7366 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
7367 | { | |
355fe088 | 7368 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
7369 | if (!gimple_nop_p (def_stmt) |
7370 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
7371 | { | |
7372 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
7373 | gsi_remove (&gsi, false); | |
7374 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
7375 | } | |
7376 | } | |
7377 | ||
7378 | return true; | |
7379 | } | |
7380 | ||
ebfd146a IR |
7381 | /* vectorizable_load. |
7382 | ||
b8698a0f L |
7383 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
7384 | can be vectorized. | |
7385 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
7386 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
7387 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7388 | ||
7389 | static bool | |
1eede195 RS |
7390 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, |
7391 | stmt_vec_info *vec_stmt, slp_tree slp_node, | |
7392 | slp_instance slp_node_instance, | |
68435eb2 | 7393 | stmt_vector_for_cost *cost_vec) |
ebfd146a IR |
7394 | { |
7395 | tree scalar_dest; | |
7396 | tree vec_dest = NULL; | |
7397 | tree data_ref = NULL; | |
7398 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 7399 | stmt_vec_info prev_stmt_info; |
ebfd146a | 7400 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 7401 | struct loop *loop = NULL; |
ebfd146a | 7402 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 7403 | bool nested_in_vect_loop = false; |
c716e67f | 7404 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 7405 | tree elem_type; |
ebfd146a | 7406 | tree new_temp; |
ef4bddc2 | 7407 | machine_mode mode; |
ebfd146a IR |
7408 | tree dummy; |
7409 | enum dr_alignment_support alignment_support_scheme; | |
7410 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 7411 | tree dataref_offset = NULL_TREE; |
355fe088 | 7412 | gimple *ptr_incr = NULL; |
ebfd146a | 7413 | int ncopies; |
4d694b27 RS |
7414 | int i, j; |
7415 | unsigned int group_size; | |
7416 | poly_uint64 group_gap_adj; | |
ebfd146a IR |
7417 | tree msq = NULL_TREE, lsq; |
7418 | tree offset = NULL_TREE; | |
356bbc4c | 7419 | tree byte_offset = NULL_TREE; |
ebfd146a | 7420 | tree realignment_token = NULL_TREE; |
538dd0b7 | 7421 | gphi *phi = NULL; |
6e1aa848 | 7422 | vec<tree> dr_chain = vNULL; |
0d0293ac | 7423 | bool grouped_load = false; |
bffb8014 | 7424 | stmt_vec_info first_stmt_info; |
b9787581 | 7425 | stmt_vec_info first_stmt_info_for_drptr = NULL; |
ebfd146a IR |
7426 | bool inv_p; |
7427 | bool compute_in_loop = false; | |
7428 | struct loop *at_loop; | |
7429 | int vec_num; | |
7430 | bool slp = (slp_node != NULL); | |
7431 | bool slp_perm = false; | |
a70d6342 | 7432 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
d9f21f6a | 7433 | poly_uint64 vf; |
272c6793 | 7434 | tree aggr_type; |
134c85ca | 7435 | gather_scatter_info gs_info; |
310213d4 | 7436 | vec_info *vinfo = stmt_info->vinfo; |
44fc7854 | 7437 | tree ref_type; |
929b4411 | 7438 | enum vect_def_type mask_dt = vect_unknown_def_type; |
a70d6342 | 7439 | |
465c8c19 JJ |
7440 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
7441 | return false; | |
7442 | ||
66c16fd9 RB |
7443 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7444 | && ! vec_stmt) | |
465c8c19 JJ |
7445 | return false; |
7446 | ||
c3a8f964 RS |
7447 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
7448 | if (is_gimple_assign (stmt)) | |
7449 | { | |
7450 | scalar_dest = gimple_assign_lhs (stmt); | |
7451 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
7452 | return false; | |
465c8c19 | 7453 | |
c3a8f964 RS |
7454 | tree_code code = gimple_assign_rhs_code (stmt); |
7455 | if (code != ARRAY_REF | |
7456 | && code != BIT_FIELD_REF | |
7457 | && code != INDIRECT_REF | |
7458 | && code != COMPONENT_REF | |
7459 | && code != IMAGPART_EXPR | |
7460 | && code != REALPART_EXPR | |
7461 | && code != MEM_REF | |
7462 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
7463 | return false; | |
7464 | } | |
7465 | else | |
7466 | { | |
7467 | gcall *call = dyn_cast <gcall *> (stmt); | |
bfaa08b7 RS |
7468 | if (!call || !gimple_call_internal_p (call)) |
7469 | return false; | |
7470 | ||
7471 | internal_fn ifn = gimple_call_internal_fn (call); | |
7472 | if (!internal_load_fn_p (ifn)) | |
c3a8f964 | 7473 | return false; |
465c8c19 | 7474 | |
c3a8f964 RS |
7475 | scalar_dest = gimple_call_lhs (call); |
7476 | if (!scalar_dest) | |
7477 | return false; | |
7478 | ||
7479 | if (slp_node != NULL) | |
7480 | { | |
7481 | if (dump_enabled_p ()) | |
7482 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7483 | "SLP of masked loads not supported.\n"); | |
7484 | return false; | |
7485 | } | |
7486 | ||
bfaa08b7 RS |
7487 | int mask_index = internal_fn_mask_index (ifn); |
7488 | if (mask_index >= 0) | |
7489 | { | |
7490 | mask = gimple_call_arg (call, mask_index); | |
929b4411 RS |
7491 | if (!vect_check_load_store_mask (stmt, mask, &mask_dt, |
7492 | &mask_vectype)) | |
bfaa08b7 RS |
7493 | return false; |
7494 | } | |
c3a8f964 | 7495 | } |
465c8c19 JJ |
7496 | |
7497 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
7498 | return false; | |
7499 | ||
7500 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
4d694b27 | 7501 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 | 7502 | |
a70d6342 IR |
7503 | if (loop_vinfo) |
7504 | { | |
7505 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
7506 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
7507 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
7508 | } | |
7509 | else | |
3533e503 | 7510 | vf = 1; |
ebfd146a IR |
7511 | |
7512 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 7513 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 7514 | case of SLP. */ |
fce57248 | 7515 | if (slp) |
ebfd146a IR |
7516 | ncopies = 1; |
7517 | else | |
e8f142e2 | 7518 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
ebfd146a IR |
7519 | |
7520 | gcc_assert (ncopies >= 1); | |
7521 | ||
7522 | /* FORNOW. This restriction should be relaxed. */ | |
7523 | if (nested_in_vect_loop && ncopies > 1) | |
7524 | { | |
73fbfcad | 7525 | if (dump_enabled_p ()) |
78c60e3d | 7526 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7527 | "multiple types in nested loop.\n"); |
ebfd146a IR |
7528 | return false; |
7529 | } | |
7530 | ||
f2556b68 RB |
7531 | /* Invalidate assumptions made by dependence analysis when vectorization |
7532 | on the unrolled body effectively re-orders stmts. */ | |
7533 | if (ncopies > 1 | |
7534 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
7535 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
7536 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
7537 | { |
7538 | if (dump_enabled_p ()) | |
7539 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7540 | "cannot perform implicit CSE when unrolling " | |
7541 | "with negative dependence distance\n"); | |
7542 | return false; | |
7543 | } | |
7544 | ||
7b7b1813 | 7545 | elem_type = TREE_TYPE (vectype); |
947131ba | 7546 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
7547 | |
7548 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
7549 | (e.g. - data copies). */ | |
947131ba | 7550 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 7551 | { |
73fbfcad | 7552 | if (dump_enabled_p ()) |
78c60e3d | 7553 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7554 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
7555 | return false; |
7556 | } | |
7557 | ||
ebfd146a | 7558 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 7559 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 7560 | { |
0d0293ac | 7561 | grouped_load = true; |
ebfd146a | 7562 | /* FORNOW */ |
2de001ee RS |
7563 | gcc_assert (!nested_in_vect_loop); |
7564 | gcc_assert (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)); | |
ebfd146a | 7565 | |
bffb8014 RS |
7566 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
7567 | group_size = DR_GROUP_SIZE (first_stmt_info); | |
d5f035ea | 7568 | |
b1af7da6 RB |
7569 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
7570 | slp_perm = true; | |
7571 | ||
f2556b68 RB |
7572 | /* Invalidate assumptions made by dependence analysis when vectorization |
7573 | on the unrolled body effectively re-orders stmts. */ | |
7574 | if (!PURE_SLP_STMT (stmt_info) | |
7575 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
7576 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
7577 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
7578 | { |
7579 | if (dump_enabled_p ()) | |
7580 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7581 | "cannot perform implicit CSE when performing " | |
7582 | "group loads with negative dependence distance\n"); | |
7583 | return false; | |
7584 | } | |
96bb56b2 RB |
7585 | |
7586 | /* Similarly when the stmt is a load that is both part of a SLP | |
7587 | instance and a loop vectorized stmt via the same-dr mechanism | |
7588 | we have to give up. */ | |
2c53b149 | 7589 | if (DR_GROUP_SAME_DR_STMT (stmt_info) |
96bb56b2 | 7590 | && (STMT_SLP_TYPE (stmt_info) |
c26228d4 | 7591 | != STMT_SLP_TYPE (DR_GROUP_SAME_DR_STMT (stmt_info)))) |
96bb56b2 RB |
7592 | { |
7593 | if (dump_enabled_p ()) | |
7594 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7595 | "conflicting SLP types for CSEd load\n"); | |
7596 | return false; | |
7597 | } | |
ebfd146a | 7598 | } |
7cfb4d93 RS |
7599 | else |
7600 | group_size = 1; | |
ebfd146a | 7601 | |
2de001ee | 7602 | vect_memory_access_type memory_access_type; |
7e11fc7f | 7603 | if (!get_load_store_type (stmt, vectype, slp, mask, VLS_LOAD, ncopies, |
2de001ee RS |
7604 | &memory_access_type, &gs_info)) |
7605 | return false; | |
a1e53f3f | 7606 | |
c3a8f964 RS |
7607 | if (mask) |
7608 | { | |
7609 | if (memory_access_type == VMAT_CONTIGUOUS) | |
7610 | { | |
7e11fc7f RS |
7611 | machine_mode vec_mode = TYPE_MODE (vectype); |
7612 | if (!VECTOR_MODE_P (vec_mode) | |
7613 | || !can_vec_mask_load_store_p (vec_mode, | |
c3a8f964 RS |
7614 | TYPE_MODE (mask_vectype), true)) |
7615 | return false; | |
7616 | } | |
bfaa08b7 | 7617 | else if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
c3a8f964 RS |
7618 | { |
7619 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); | |
7620 | tree masktype | |
7621 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
7622 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
7623 | { | |
7624 | if (dump_enabled_p ()) | |
7625 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7626 | "masked gather with integer mask not" | |
7627 | " supported."); | |
7628 | return false; | |
7629 | } | |
7630 | } | |
bfaa08b7 RS |
7631 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
7632 | && memory_access_type != VMAT_GATHER_SCATTER) | |
c3a8f964 RS |
7633 | { |
7634 | if (dump_enabled_p ()) | |
7635 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7636 | "unsupported access type for masked load.\n"); | |
7637 | return false; | |
7638 | } | |
7639 | } | |
7640 | ||
ebfd146a IR |
7641 | if (!vec_stmt) /* transformation not required. */ |
7642 | { | |
2de001ee RS |
7643 | if (!slp) |
7644 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; | |
7cfb4d93 RS |
7645 | |
7646 | if (loop_vinfo | |
7647 | && LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) | |
7648 | check_load_store_masking (loop_vinfo, vectype, VLS_LOAD, group_size, | |
bfaa08b7 | 7649 | memory_access_type, &gs_info); |
7cfb4d93 | 7650 | |
ebfd146a | 7651 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; |
68435eb2 RB |
7652 | vect_model_load_cost (stmt_info, ncopies, memory_access_type, |
7653 | slp_node_instance, slp_node, cost_vec); | |
ebfd146a IR |
7654 | return true; |
7655 | } | |
7656 | ||
2de001ee RS |
7657 | if (!slp) |
7658 | gcc_assert (memory_access_type | |
7659 | == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); | |
7660 | ||
73fbfcad | 7661 | if (dump_enabled_p ()) |
78c60e3d | 7662 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 7663 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a | 7664 | |
67b8dbac | 7665 | /* Transform. */ |
ebfd146a | 7666 | |
f702e7d4 | 7667 | ensure_base_align (dr); |
c716e67f | 7668 | |
bfaa08b7 | 7669 | if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
aec7ae7d | 7670 | { |
929b4411 RS |
7671 | vect_build_gather_load_calls (stmt, gsi, vec_stmt, &gs_info, mask, |
7672 | mask_dt); | |
aec7ae7d JJ |
7673 | return true; |
7674 | } | |
2de001ee RS |
7675 | |
7676 | if (memory_access_type == VMAT_ELEMENTWISE | |
7677 | || memory_access_type == VMAT_STRIDED_SLP) | |
7d75abc8 MM |
7678 | { |
7679 | gimple_stmt_iterator incr_gsi; | |
7680 | bool insert_after; | |
355fe088 | 7681 | gimple *incr; |
7d75abc8 | 7682 | tree offvar; |
7d75abc8 MM |
7683 | tree ivstep; |
7684 | tree running_off; | |
9771b263 | 7685 | vec<constructor_elt, va_gc> *v = NULL; |
14ac6aa2 | 7686 | tree stride_base, stride_step, alias_off; |
4d694b27 RS |
7687 | /* Checked by get_load_store_type. */ |
7688 | unsigned int const_nunits = nunits.to_constant (); | |
b210f45f | 7689 | unsigned HOST_WIDE_INT cst_offset = 0; |
14ac6aa2 | 7690 | |
7cfb4d93 | 7691 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
14ac6aa2 | 7692 | gcc_assert (!nested_in_vect_loop); |
7d75abc8 | 7693 | |
b210f45f | 7694 | if (grouped_load) |
44fc7854 | 7695 | { |
bffb8014 RS |
7696 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
7697 | first_dr = STMT_VINFO_DATA_REF (first_stmt_info); | |
44fc7854 | 7698 | } |
ab313a8c | 7699 | else |
44fc7854 | 7700 | { |
bffb8014 | 7701 | first_stmt_info = stmt_info; |
44fc7854 | 7702 | first_dr = dr; |
b210f45f RB |
7703 | } |
7704 | if (slp && grouped_load) | |
7705 | { | |
bffb8014 RS |
7706 | group_size = DR_GROUP_SIZE (first_stmt_info); |
7707 | ref_type = get_group_alias_ptr_type (first_stmt_info); | |
b210f45f RB |
7708 | } |
7709 | else | |
7710 | { | |
7711 | if (grouped_load) | |
7712 | cst_offset | |
7713 | = (tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))) | |
bffb8014 RS |
7714 | * vect_get_place_in_interleaving_chain (stmt, |
7715 | first_stmt_info)); | |
44fc7854 | 7716 | group_size = 1; |
b210f45f | 7717 | ref_type = reference_alias_ptr_type (DR_REF (dr)); |
44fc7854 | 7718 | } |
ab313a8c | 7719 | |
14ac6aa2 RB |
7720 | stride_base |
7721 | = fold_build_pointer_plus | |
ab313a8c | 7722 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 7723 | size_binop (PLUS_EXPR, |
ab313a8c RB |
7724 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
7725 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
7726 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
7727 | |
7728 | /* For a load with loop-invariant (but other than power-of-2) | |
7729 | stride (i.e. not a grouped access) like so: | |
7730 | ||
7731 | for (i = 0; i < n; i += stride) | |
7732 | ... = array[i]; | |
7733 | ||
7734 | we generate a new induction variable and new accesses to | |
7735 | form a new vector (or vectors, depending on ncopies): | |
7736 | ||
7737 | for (j = 0; ; j += VF*stride) | |
7738 | tmp1 = array[j]; | |
7739 | tmp2 = array[j + stride]; | |
7740 | ... | |
7741 | vectemp = {tmp1, tmp2, ...} | |
7742 | */ | |
7743 | ||
ab313a8c RB |
7744 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
7745 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
7746 | |
7747 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
7748 | ||
b210f45f RB |
7749 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
7750 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); | |
7751 | create_iv (stride_base, ivstep, NULL, | |
7d75abc8 MM |
7752 | loop, &incr_gsi, insert_after, |
7753 | &offvar, NULL); | |
7754 | incr = gsi_stmt (incr_gsi); | |
4fbeb363 | 7755 | loop_vinfo->add_stmt (incr); |
7d75abc8 | 7756 | |
b210f45f | 7757 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
7d75abc8 MM |
7758 | |
7759 | prev_stmt_info = NULL; | |
7760 | running_off = offvar; | |
44fc7854 | 7761 | alias_off = build_int_cst (ref_type, 0); |
4d694b27 | 7762 | int nloads = const_nunits; |
e09b4c37 | 7763 | int lnel = 1; |
7b5fc413 | 7764 | tree ltype = TREE_TYPE (vectype); |
ea60dd34 | 7765 | tree lvectype = vectype; |
b266b968 | 7766 | auto_vec<tree> dr_chain; |
2de001ee | 7767 | if (memory_access_type == VMAT_STRIDED_SLP) |
7b5fc413 | 7768 | { |
4d694b27 | 7769 | if (group_size < const_nunits) |
e09b4c37 | 7770 | { |
ff03930a JJ |
7771 | /* First check if vec_init optab supports construction from |
7772 | vector elts directly. */ | |
b397965c | 7773 | scalar_mode elmode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); |
9da15d40 RS |
7774 | machine_mode vmode; |
7775 | if (mode_for_vector (elmode, group_size).exists (&vmode) | |
7776 | && VECTOR_MODE_P (vmode) | |
414fef4e | 7777 | && targetm.vector_mode_supported_p (vmode) |
ff03930a JJ |
7778 | && (convert_optab_handler (vec_init_optab, |
7779 | TYPE_MODE (vectype), vmode) | |
7780 | != CODE_FOR_nothing)) | |
ea60dd34 | 7781 | { |
4d694b27 | 7782 | nloads = const_nunits / group_size; |
ea60dd34 | 7783 | lnel = group_size; |
ff03930a JJ |
7784 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); |
7785 | } | |
7786 | else | |
7787 | { | |
7788 | /* Otherwise avoid emitting a constructor of vector elements | |
7789 | by performing the loads using an integer type of the same | |
7790 | size, constructing a vector of those and then | |
7791 | re-interpreting it as the original vector type. | |
7792 | This avoids a huge runtime penalty due to the general | |
7793 | inability to perform store forwarding from smaller stores | |
7794 | to a larger load. */ | |
7795 | unsigned lsize | |
7796 | = group_size * TYPE_PRECISION (TREE_TYPE (vectype)); | |
fffbab82 | 7797 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 7798 | unsigned int lnunits = const_nunits / group_size; |
ff03930a JJ |
7799 | /* If we can't construct such a vector fall back to |
7800 | element loads of the original vector type. */ | |
4d694b27 | 7801 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 7802 | && VECTOR_MODE_P (vmode) |
414fef4e | 7803 | && targetm.vector_mode_supported_p (vmode) |
ff03930a JJ |
7804 | && (convert_optab_handler (vec_init_optab, vmode, elmode) |
7805 | != CODE_FOR_nothing)) | |
7806 | { | |
4d694b27 | 7807 | nloads = lnunits; |
ff03930a JJ |
7808 | lnel = group_size; |
7809 | ltype = build_nonstandard_integer_type (lsize, 1); | |
7810 | lvectype = build_vector_type (ltype, nloads); | |
7811 | } | |
ea60dd34 | 7812 | } |
e09b4c37 | 7813 | } |
2de001ee | 7814 | else |
e09b4c37 | 7815 | { |
ea60dd34 | 7816 | nloads = 1; |
4d694b27 | 7817 | lnel = const_nunits; |
e09b4c37 | 7818 | ltype = vectype; |
e09b4c37 | 7819 | } |
2de001ee RS |
7820 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); |
7821 | } | |
bb4e4747 BC |
7822 | /* Load vector(1) scalar_type if it's 1 element-wise vectype. */ |
7823 | else if (nloads == 1) | |
7824 | ltype = vectype; | |
7825 | ||
2de001ee RS |
7826 | if (slp) |
7827 | { | |
66c16fd9 RB |
7828 | /* For SLP permutation support we need to load the whole group, |
7829 | not only the number of vector stmts the permutation result | |
7830 | fits in. */ | |
b266b968 | 7831 | if (slp_perm) |
66c16fd9 | 7832 | { |
d9f21f6a RS |
7833 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7834 | variable VF. */ | |
7835 | unsigned int const_vf = vf.to_constant (); | |
4d694b27 | 7836 | ncopies = CEIL (group_size * const_vf, const_nunits); |
66c16fd9 RB |
7837 | dr_chain.create (ncopies); |
7838 | } | |
7839 | else | |
7840 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
7b5fc413 | 7841 | } |
4d694b27 | 7842 | unsigned int group_el = 0; |
e09b4c37 RB |
7843 | unsigned HOST_WIDE_INT |
7844 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
7d75abc8 MM |
7845 | for (j = 0; j < ncopies; j++) |
7846 | { | |
7b5fc413 | 7847 | if (nloads > 1) |
e09b4c37 | 7848 | vec_alloc (v, nloads); |
e1bd7296 | 7849 | stmt_vec_info new_stmt_info = NULL; |
e09b4c37 | 7850 | for (i = 0; i < nloads; i++) |
7b5fc413 | 7851 | { |
e09b4c37 | 7852 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
b210f45f | 7853 | group_el * elsz + cst_offset); |
19986382 RB |
7854 | tree data_ref = build2 (MEM_REF, ltype, running_off, this_off); |
7855 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); | |
e1bd7296 RS |
7856 | gassign *new_stmt |
7857 | = gimple_build_assign (make_ssa_name (ltype), data_ref); | |
7858 | new_stmt_info | |
7859 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
e09b4c37 RB |
7860 | if (nloads > 1) |
7861 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, | |
7862 | gimple_assign_lhs (new_stmt)); | |
7863 | ||
7864 | group_el += lnel; | |
7865 | if (! slp | |
7866 | || group_el == group_size) | |
7b5fc413 | 7867 | { |
e09b4c37 RB |
7868 | tree newoff = copy_ssa_name (running_off); |
7869 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
7870 | running_off, stride_step); | |
7b5fc413 RB |
7871 | vect_finish_stmt_generation (stmt, incr, gsi); |
7872 | ||
7873 | running_off = newoff; | |
e09b4c37 | 7874 | group_el = 0; |
7b5fc413 | 7875 | } |
7b5fc413 | 7876 | } |
e09b4c37 | 7877 | if (nloads > 1) |
7d75abc8 | 7878 | { |
ea60dd34 RB |
7879 | tree vec_inv = build_constructor (lvectype, v); |
7880 | new_temp = vect_init_vector (stmt, vec_inv, lvectype, gsi); | |
e1bd7296 | 7881 | new_stmt_info = vinfo->lookup_def (new_temp); |
ea60dd34 RB |
7882 | if (lvectype != vectype) |
7883 | { | |
e1bd7296 RS |
7884 | gassign *new_stmt |
7885 | = gimple_build_assign (make_ssa_name (vectype), | |
7886 | VIEW_CONVERT_EXPR, | |
7887 | build1 (VIEW_CONVERT_EXPR, | |
7888 | vectype, new_temp)); | |
7889 | new_stmt_info | |
7890 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ea60dd34 | 7891 | } |
7d75abc8 MM |
7892 | } |
7893 | ||
7b5fc413 | 7894 | if (slp) |
b266b968 | 7895 | { |
b266b968 | 7896 | if (slp_perm) |
e1bd7296 | 7897 | dr_chain.quick_push (gimple_assign_lhs (new_stmt_info->stmt)); |
66c16fd9 | 7898 | else |
e1bd7296 | 7899 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
b266b968 | 7900 | } |
7d75abc8 | 7901 | else |
225ce44b RB |
7902 | { |
7903 | if (j == 0) | |
e1bd7296 | 7904 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
225ce44b | 7905 | else |
e1bd7296 RS |
7906 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
7907 | prev_stmt_info = new_stmt_info; | |
225ce44b | 7908 | } |
7d75abc8 | 7909 | } |
b266b968 | 7910 | if (slp_perm) |
29afecdf RB |
7911 | { |
7912 | unsigned n_perms; | |
7913 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
7914 | slp_node_instance, false, &n_perms); | |
7915 | } | |
7d75abc8 MM |
7916 | return true; |
7917 | } | |
aec7ae7d | 7918 | |
b5ec4de7 RS |
7919 | if (memory_access_type == VMAT_GATHER_SCATTER |
7920 | || (!slp && memory_access_type == VMAT_CONTIGUOUS)) | |
ab2fc782 RS |
7921 | grouped_load = false; |
7922 | ||
0d0293ac | 7923 | if (grouped_load) |
ebfd146a | 7924 | { |
bffb8014 RS |
7925 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
7926 | group_size = DR_GROUP_SIZE (first_stmt_info); | |
4f0a0218 | 7927 | /* For SLP vectorization we directly vectorize a subchain |
52eab378 RB |
7928 | without permutation. */ |
7929 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
bffb8014 | 7930 | first_stmt_info = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
4f0a0218 RB |
7931 | /* For BB vectorization always use the first stmt to base |
7932 | the data ref pointer on. */ | |
7933 | if (bb_vinfo) | |
b9787581 | 7934 | first_stmt_info_for_drptr = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
6aa904c4 | 7935 | |
ebfd146a | 7936 | /* Check if the chain of loads is already vectorized. */ |
bffb8014 | 7937 | if (STMT_VINFO_VEC_STMT (first_stmt_info) |
01d8bf07 RB |
7938 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. |
7939 | ??? But we can only do so if there is exactly one | |
7940 | as we have no way to get at the rest. Leave the CSE | |
7941 | opportunity alone. | |
7942 | ??? With the group load eventually participating | |
7943 | in multiple different permutations (having multiple | |
7944 | slp nodes which refer to the same group) the CSE | |
7945 | is even wrong code. See PR56270. */ | |
7946 | && !slp) | |
ebfd146a IR |
7947 | { |
7948 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
7949 | return true; | |
7950 | } | |
bffb8014 | 7951 | first_dr = STMT_VINFO_DATA_REF (first_stmt_info); |
9b999e8c | 7952 | group_gap_adj = 0; |
ebfd146a IR |
7953 | |
7954 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
7955 | if (slp) | |
7956 | { | |
0d0293ac | 7957 | grouped_load = false; |
91ff1504 RB |
7958 | /* For SLP permutation support we need to load the whole group, |
7959 | not only the number of vector stmts the permutation result | |
7960 | fits in. */ | |
7961 | if (slp_perm) | |
b267968e | 7962 | { |
d9f21f6a RS |
7963 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7964 | variable VF. */ | |
7965 | unsigned int const_vf = vf.to_constant (); | |
4d694b27 RS |
7966 | unsigned int const_nunits = nunits.to_constant (); |
7967 | vec_num = CEIL (group_size * const_vf, const_nunits); | |
b267968e RB |
7968 | group_gap_adj = vf * group_size - nunits * vec_num; |
7969 | } | |
91ff1504 | 7970 | else |
b267968e RB |
7971 | { |
7972 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
796bd467 RB |
7973 | group_gap_adj |
7974 | = group_size - SLP_INSTANCE_GROUP_SIZE (slp_node_instance); | |
b267968e | 7975 | } |
a70d6342 | 7976 | } |
ebfd146a | 7977 | else |
9b999e8c | 7978 | vec_num = group_size; |
44fc7854 | 7979 | |
bffb8014 | 7980 | ref_type = get_group_alias_ptr_type (first_stmt_info); |
ebfd146a IR |
7981 | } |
7982 | else | |
7983 | { | |
bffb8014 | 7984 | first_stmt_info = stmt_info; |
ebfd146a IR |
7985 | first_dr = dr; |
7986 | group_size = vec_num = 1; | |
9b999e8c | 7987 | group_gap_adj = 0; |
44fc7854 | 7988 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
ebfd146a IR |
7989 | } |
7990 | ||
720f5239 | 7991 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 7992 | gcc_assert (alignment_support_scheme); |
70088b95 RS |
7993 | vec_loop_masks *loop_masks |
7994 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) | |
7995 | ? &LOOP_VINFO_MASKS (loop_vinfo) | |
7996 | : NULL); | |
7cfb4d93 RS |
7997 | /* Targets with store-lane instructions must not require explicit |
7998 | realignment. vect_supportable_dr_alignment always returns either | |
7999 | dr_aligned or dr_unaligned_supported for masked operations. */ | |
8000 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES | |
8001 | && !mask | |
70088b95 | 8002 | && !loop_masks) |
272c6793 RS |
8003 | || alignment_support_scheme == dr_aligned |
8004 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
8005 | |
8006 | /* In case the vectorization factor (VF) is bigger than the number | |
8007 | of elements that we can fit in a vectype (nunits), we have to generate | |
8008 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 8009 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 8010 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 8011 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 8012 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
8013 | stmts that use the defs of the current stmt. The example below |
8014 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
8015 | need to create 4 vectorized stmts): | |
ebfd146a IR |
8016 | |
8017 | before vectorization: | |
8018 | RELATED_STMT VEC_STMT | |
8019 | S1: x = memref - - | |
8020 | S2: z = x + 1 - - | |
8021 | ||
8022 | step 1: vectorize stmt S1: | |
8023 | We first create the vector stmt VS1_0, and, as usual, record a | |
8024 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
8025 | Next, we create the vector stmt VS1_1, and record a pointer to | |
8026 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 8027 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
8028 | stmts and pointers: |
8029 | RELATED_STMT VEC_STMT | |
8030 | VS1_0: vx0 = memref0 VS1_1 - | |
8031 | VS1_1: vx1 = memref1 VS1_2 - | |
8032 | VS1_2: vx2 = memref2 VS1_3 - | |
8033 | VS1_3: vx3 = memref3 - - | |
8034 | S1: x = load - VS1_0 | |
8035 | S2: z = x + 1 - - | |
8036 | ||
b8698a0f L |
8037 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
8038 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
8039 | stmt S2. */ |
8040 | ||
0d0293ac | 8041 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
8042 | |
8043 | S1: x2 = &base + 2 | |
8044 | S2: x0 = &base | |
8045 | S3: x1 = &base + 1 | |
8046 | S4: x3 = &base + 3 | |
8047 | ||
b8698a0f | 8048 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
8049 | starting from the access of the first stmt of the chain: |
8050 | ||
8051 | VS1: vx0 = &base | |
8052 | VS2: vx1 = &base + vec_size*1 | |
8053 | VS3: vx3 = &base + vec_size*2 | |
8054 | VS4: vx4 = &base + vec_size*3 | |
8055 | ||
8056 | Then permutation statements are generated: | |
8057 | ||
e2c83630 RH |
8058 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
8059 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
8060 | ... |
8061 | ||
8062 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
8063 | (the order of the data-refs in the output of vect_permute_load_chain | |
8064 | corresponds to the order of scalar stmts in the interleaving chain - see | |
8065 | the documentation of vect_permute_load_chain()). | |
8066 | The generation of permutation stmts and recording them in | |
0d0293ac | 8067 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 8068 | |
b8698a0f | 8069 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
8070 | permutation stmts above are created for every copy. The result vector |
8071 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
8072 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
8073 | |
8074 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
8075 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
8076 | we generate the following code: | |
8077 | p = initial_addr; | |
8078 | indx = 0; | |
8079 | loop { | |
8080 | p = p + indx * vectype_size; | |
8081 | vec_dest = *(p); | |
8082 | indx = indx + 1; | |
8083 | } | |
8084 | ||
8085 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 8086 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
8087 | then generate the following code, in which the data in each iteration is |
8088 | obtained by two vector loads, one from the previous iteration, and one | |
8089 | from the current iteration: | |
8090 | p1 = initial_addr; | |
8091 | msq_init = *(floor(p1)) | |
8092 | p2 = initial_addr + VS - 1; | |
8093 | realignment_token = call target_builtin; | |
8094 | indx = 0; | |
8095 | loop { | |
8096 | p2 = p2 + indx * vectype_size | |
8097 | lsq = *(floor(p2)) | |
8098 | vec_dest = realign_load (msq, lsq, realignment_token) | |
8099 | indx = indx + 1; | |
8100 | msq = lsq; | |
8101 | } */ | |
8102 | ||
8103 | /* If the misalignment remains the same throughout the execution of the | |
8104 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 8105 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
8106 | This can only occur when vectorizing memory accesses in the inner-loop |
8107 | nested within an outer-loop that is being vectorized. */ | |
8108 | ||
d1e4b493 | 8109 | if (nested_in_vect_loop |
cf098191 RS |
8110 | && !multiple_p (DR_STEP_ALIGNMENT (dr), |
8111 | GET_MODE_SIZE (TYPE_MODE (vectype)))) | |
ebfd146a IR |
8112 | { |
8113 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
8114 | compute_in_loop = true; | |
8115 | } | |
8116 | ||
8117 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
8118 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 8119 | && !compute_in_loop) |
ebfd146a | 8120 | { |
bffb8014 | 8121 | msq = vect_setup_realignment (first_stmt_info, gsi, &realignment_token, |
ebfd146a IR |
8122 | alignment_support_scheme, NULL_TREE, |
8123 | &at_loop); | |
8124 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
8125 | { | |
538dd0b7 | 8126 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
8127 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
8128 | size_one_node); | |
ebfd146a IR |
8129 | } |
8130 | } | |
8131 | else | |
8132 | at_loop = loop; | |
8133 | ||
62da9e14 | 8134 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
a1e53f3f L |
8135 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
8136 | ||
ab2fc782 RS |
8137 | tree bump; |
8138 | tree vec_offset = NULL_TREE; | |
8139 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
8140 | { | |
8141 | aggr_type = NULL_TREE; | |
8142 | bump = NULL_TREE; | |
8143 | } | |
8144 | else if (memory_access_type == VMAT_GATHER_SCATTER) | |
8145 | { | |
8146 | aggr_type = elem_type; | |
8147 | vect_get_strided_load_store_ops (stmt, loop_vinfo, &gs_info, | |
8148 | &bump, &vec_offset); | |
8149 | } | |
272c6793 | 8150 | else |
ab2fc782 RS |
8151 | { |
8152 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
8153 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
8154 | else | |
8155 | aggr_type = vectype; | |
8156 | bump = vect_get_data_ptr_increment (dr, aggr_type, memory_access_type); | |
8157 | } | |
272c6793 | 8158 | |
c3a8f964 | 8159 | tree vec_mask = NULL_TREE; |
ebfd146a | 8160 | prev_stmt_info = NULL; |
4d694b27 | 8161 | poly_uint64 group_elt = 0; |
ebfd146a | 8162 | for (j = 0; j < ncopies; j++) |
b8698a0f | 8163 | { |
e1bd7296 | 8164 | stmt_vec_info new_stmt_info = NULL; |
272c6793 | 8165 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 8166 | if (j == 0) |
74bf76ed JJ |
8167 | { |
8168 | bool simd_lane_access_p | |
8169 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
8170 | if (simd_lane_access_p | |
8171 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
8172 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
8173 | && integer_zerop (DR_OFFSET (first_dr)) | |
8174 | && integer_zerop (DR_INIT (first_dr)) | |
8175 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 8176 | get_alias_set (TREE_TYPE (ref_type))) |
74bf76ed JJ |
8177 | && (alignment_support_scheme == dr_aligned |
8178 | || alignment_support_scheme == dr_unaligned_supported)) | |
8179 | { | |
8180 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 8181 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 8182 | inv_p = false; |
74bf76ed | 8183 | } |
b9787581 | 8184 | else if (first_stmt_info_for_drptr |
bffb8014 | 8185 | && first_stmt_info != first_stmt_info_for_drptr) |
4f0a0218 RB |
8186 | { |
8187 | dataref_ptr | |
b9787581 RS |
8188 | = vect_create_data_ref_ptr (first_stmt_info_for_drptr, |
8189 | aggr_type, at_loop, offset, &dummy, | |
8190 | gsi, &ptr_incr, simd_lane_access_p, | |
ab2fc782 | 8191 | &inv_p, byte_offset, bump); |
4f0a0218 RB |
8192 | /* Adjust the pointer by the difference to first_stmt. */ |
8193 | data_reference_p ptrdr | |
b9787581 | 8194 | = STMT_VINFO_DATA_REF (first_stmt_info_for_drptr); |
4f0a0218 RB |
8195 | tree diff = fold_convert (sizetype, |
8196 | size_binop (MINUS_EXPR, | |
8197 | DR_INIT (first_dr), | |
8198 | DR_INIT (ptrdr))); | |
8199 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
8200 | stmt, diff); | |
8201 | } | |
bfaa08b7 RS |
8202 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
8203 | { | |
8204 | vect_get_gather_scatter_ops (loop, stmt, &gs_info, | |
8205 | &dataref_ptr, &vec_offset); | |
8206 | inv_p = false; | |
8207 | } | |
74bf76ed JJ |
8208 | else |
8209 | dataref_ptr | |
bffb8014 | 8210 | = vect_create_data_ref_ptr (first_stmt_info, aggr_type, at_loop, |
74bf76ed | 8211 | offset, &dummy, gsi, &ptr_incr, |
356bbc4c | 8212 | simd_lane_access_p, &inv_p, |
ab2fc782 | 8213 | byte_offset, bump); |
c3a8f964 RS |
8214 | if (mask) |
8215 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, | |
8216 | mask_vectype); | |
74bf76ed | 8217 | } |
ebfd146a | 8218 | else |
c3a8f964 RS |
8219 | { |
8220 | if (dataref_offset) | |
8221 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
ab2fc782 | 8222 | bump); |
bfaa08b7 | 8223 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
929b4411 RS |
8224 | vec_offset = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
8225 | vec_offset); | |
c3a8f964 | 8226 | else |
ab2fc782 RS |
8227 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8228 | stmt, bump); | |
c3a8f964 | 8229 | if (mask) |
929b4411 | 8230 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
c3a8f964 | 8231 | } |
ebfd146a | 8232 | |
0d0293ac | 8233 | if (grouped_load || slp_perm) |
9771b263 | 8234 | dr_chain.create (vec_num); |
5ce1ee7f | 8235 | |
2de001ee | 8236 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 8237 | { |
272c6793 RS |
8238 | tree vec_array; |
8239 | ||
8240 | vec_array = create_vector_array (vectype, vec_num); | |
8241 | ||
7cfb4d93 | 8242 | tree final_mask = NULL_TREE; |
70088b95 RS |
8243 | if (loop_masks) |
8244 | final_mask = vect_get_loop_mask (gsi, loop_masks, ncopies, | |
8245 | vectype, j); | |
7cfb4d93 RS |
8246 | if (vec_mask) |
8247 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
8248 | vec_mask, gsi); | |
8249 | ||
7e11fc7f | 8250 | gcall *call; |
7cfb4d93 | 8251 | if (final_mask) |
7e11fc7f RS |
8252 | { |
8253 | /* Emit: | |
8254 | VEC_ARRAY = MASK_LOAD_LANES (DATAREF_PTR, ALIAS_PTR, | |
8255 | VEC_MASK). */ | |
8256 | unsigned int align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype)); | |
8257 | tree alias_ptr = build_int_cst (ref_type, align); | |
8258 | call = gimple_build_call_internal (IFN_MASK_LOAD_LANES, 3, | |
8259 | dataref_ptr, alias_ptr, | |
7cfb4d93 | 8260 | final_mask); |
7e11fc7f RS |
8261 | } |
8262 | else | |
8263 | { | |
8264 | /* Emit: | |
8265 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
8266 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); | |
8267 | call = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
8268 | } | |
a844293d RS |
8269 | gimple_call_set_lhs (call, vec_array); |
8270 | gimple_call_set_nothrow (call, true); | |
e1bd7296 | 8271 | new_stmt_info = vect_finish_stmt_generation (stmt, call, gsi); |
ebfd146a | 8272 | |
272c6793 RS |
8273 | /* Extract each vector into an SSA_NAME. */ |
8274 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 8275 | { |
272c6793 RS |
8276 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
8277 | vec_array, i); | |
9771b263 | 8278 | dr_chain.quick_push (new_temp); |
272c6793 RS |
8279 | } |
8280 | ||
8281 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 8282 | vect_record_grouped_load_vectors (stmt, dr_chain); |
3ba4ff41 RS |
8283 | |
8284 | /* Record that VEC_ARRAY is now dead. */ | |
8285 | vect_clobber_variable (stmt, gsi, vec_array); | |
272c6793 RS |
8286 | } |
8287 | else | |
8288 | { | |
8289 | for (i = 0; i < vec_num; i++) | |
8290 | { | |
7cfb4d93 | 8291 | tree final_mask = NULL_TREE; |
70088b95 | 8292 | if (loop_masks |
7cfb4d93 | 8293 | && memory_access_type != VMAT_INVARIANT) |
70088b95 RS |
8294 | final_mask = vect_get_loop_mask (gsi, loop_masks, |
8295 | vec_num * ncopies, | |
7cfb4d93 RS |
8296 | vectype, vec_num * j + i); |
8297 | if (vec_mask) | |
8298 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
8299 | vec_mask, gsi); | |
8300 | ||
272c6793 RS |
8301 | if (i > 0) |
8302 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
ab2fc782 | 8303 | stmt, bump); |
272c6793 RS |
8304 | |
8305 | /* 2. Create the vector-load in the loop. */ | |
e1bd7296 | 8306 | gimple *new_stmt = NULL; |
272c6793 RS |
8307 | switch (alignment_support_scheme) |
8308 | { | |
8309 | case dr_aligned: | |
8310 | case dr_unaligned_supported: | |
be1ac4ec | 8311 | { |
644ffefd MJ |
8312 | unsigned int align, misalign; |
8313 | ||
bfaa08b7 RS |
8314 | if (memory_access_type == VMAT_GATHER_SCATTER) |
8315 | { | |
8316 | tree scale = size_int (gs_info.scale); | |
8317 | gcall *call; | |
70088b95 | 8318 | if (loop_masks) |
bfaa08b7 RS |
8319 | call = gimple_build_call_internal |
8320 | (IFN_MASK_GATHER_LOAD, 4, dataref_ptr, | |
8321 | vec_offset, scale, final_mask); | |
8322 | else | |
8323 | call = gimple_build_call_internal | |
8324 | (IFN_GATHER_LOAD, 3, dataref_ptr, | |
8325 | vec_offset, scale); | |
8326 | gimple_call_set_nothrow (call, true); | |
8327 | new_stmt = call; | |
8328 | data_ref = NULL_TREE; | |
8329 | break; | |
8330 | } | |
8331 | ||
f702e7d4 | 8332 | align = DR_TARGET_ALIGNMENT (dr); |
272c6793 RS |
8333 | if (alignment_support_scheme == dr_aligned) |
8334 | { | |
8335 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 8336 | misalign = 0; |
272c6793 RS |
8337 | } |
8338 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
8339 | { | |
25f68d90 | 8340 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 8341 | misalign = 0; |
272c6793 RS |
8342 | } |
8343 | else | |
c3a8f964 | 8344 | misalign = DR_MISALIGNMENT (first_dr); |
aed93b23 RB |
8345 | if (dataref_offset == NULL_TREE |
8346 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
8347 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
8348 | align, misalign); | |
c3a8f964 | 8349 | |
7cfb4d93 | 8350 | if (final_mask) |
c3a8f964 RS |
8351 | { |
8352 | align = least_bit_hwi (misalign | align); | |
8353 | tree ptr = build_int_cst (ref_type, align); | |
8354 | gcall *call | |
8355 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, | |
8356 | dataref_ptr, ptr, | |
7cfb4d93 | 8357 | final_mask); |
c3a8f964 RS |
8358 | gimple_call_set_nothrow (call, true); |
8359 | new_stmt = call; | |
8360 | data_ref = NULL_TREE; | |
8361 | } | |
8362 | else | |
8363 | { | |
8364 | data_ref | |
8365 | = fold_build2 (MEM_REF, vectype, dataref_ptr, | |
8366 | dataref_offset | |
8367 | ? dataref_offset | |
8368 | : build_int_cst (ref_type, 0)); | |
8369 | if (alignment_support_scheme == dr_aligned) | |
8370 | ; | |
8371 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
8372 | TREE_TYPE (data_ref) | |
8373 | = build_aligned_type (TREE_TYPE (data_ref), | |
8374 | align * BITS_PER_UNIT); | |
8375 | else | |
8376 | TREE_TYPE (data_ref) | |
8377 | = build_aligned_type (TREE_TYPE (data_ref), | |
8378 | TYPE_ALIGN (elem_type)); | |
8379 | } | |
272c6793 | 8380 | break; |
be1ac4ec | 8381 | } |
272c6793 | 8382 | case dr_explicit_realign: |
267d3070 | 8383 | { |
272c6793 | 8384 | tree ptr, bump; |
272c6793 | 8385 | |
d88981fc | 8386 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
8387 | |
8388 | if (compute_in_loop) | |
bffb8014 | 8389 | msq = vect_setup_realignment (first_stmt_info, gsi, |
272c6793 RS |
8390 | &realignment_token, |
8391 | dr_explicit_realign, | |
8392 | dataref_ptr, NULL); | |
8393 | ||
aed93b23 RB |
8394 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
8395 | ptr = copy_ssa_name (dataref_ptr); | |
8396 | else | |
8397 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
f702e7d4 | 8398 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); |
0d0e4a03 JJ |
8399 | new_stmt = gimple_build_assign |
8400 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
8401 | build_int_cst |
8402 | (TREE_TYPE (dataref_ptr), | |
f702e7d4 | 8403 | -(HOST_WIDE_INT) align)); |
272c6793 RS |
8404 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
8405 | data_ref | |
8406 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 8407 | build_int_cst (ref_type, 0)); |
19986382 | 8408 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); |
272c6793 RS |
8409 | vec_dest = vect_create_destination_var (scalar_dest, |
8410 | vectype); | |
8411 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
8412 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
8413 | gimple_assign_set_lhs (new_stmt, new_temp); | |
8414 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
8415 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
8416 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8417 | msq = new_temp; | |
8418 | ||
d88981fc | 8419 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 8420 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 8421 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 8422 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
8423 | new_stmt = gimple_build_assign |
8424 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 | 8425 | build_int_cst |
f702e7d4 | 8426 | (TREE_TYPE (ptr), -(HOST_WIDE_INT) align)); |
aed93b23 | 8427 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
8428 | gimple_assign_set_lhs (new_stmt, ptr); |
8429 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8430 | data_ref | |
8431 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 8432 | build_int_cst (ref_type, 0)); |
272c6793 | 8433 | break; |
267d3070 | 8434 | } |
272c6793 | 8435 | case dr_explicit_realign_optimized: |
f702e7d4 RS |
8436 | { |
8437 | if (TREE_CODE (dataref_ptr) == SSA_NAME) | |
8438 | new_temp = copy_ssa_name (dataref_ptr); | |
8439 | else | |
8440 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
8441 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); | |
8442 | new_stmt = gimple_build_assign | |
8443 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
8444 | build_int_cst (TREE_TYPE (dataref_ptr), | |
8445 | -(HOST_WIDE_INT) align)); | |
8446 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8447 | data_ref | |
8448 | = build2 (MEM_REF, vectype, new_temp, | |
8449 | build_int_cst (ref_type, 0)); | |
8450 | break; | |
8451 | } | |
272c6793 RS |
8452 | default: |
8453 | gcc_unreachable (); | |
8454 | } | |
ebfd146a | 8455 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
c3a8f964 RS |
8456 | /* DATA_REF is null if we've already built the statement. */ |
8457 | if (data_ref) | |
19986382 RB |
8458 | { |
8459 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); | |
8460 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
8461 | } | |
ebfd146a | 8462 | new_temp = make_ssa_name (vec_dest, new_stmt); |
c3a8f964 | 8463 | gimple_set_lhs (new_stmt, new_temp); |
e1bd7296 RS |
8464 | new_stmt_info |
8465 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 8466 | |
272c6793 RS |
8467 | /* 3. Handle explicit realignment if necessary/supported. |
8468 | Create in loop: | |
8469 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
8470 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
8471 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 8472 | { |
272c6793 RS |
8473 | lsq = gimple_assign_lhs (new_stmt); |
8474 | if (!realignment_token) | |
8475 | realignment_token = dataref_ptr; | |
8476 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
8477 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
8478 | msq, lsq, realignment_token); | |
272c6793 RS |
8479 | new_temp = make_ssa_name (vec_dest, new_stmt); |
8480 | gimple_assign_set_lhs (new_stmt, new_temp); | |
e1bd7296 RS |
8481 | new_stmt_info |
8482 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
8483 | |
8484 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
8485 | { | |
8486 | gcc_assert (phi); | |
8487 | if (i == vec_num - 1 && j == ncopies - 1) | |
8488 | add_phi_arg (phi, lsq, | |
8489 | loop_latch_edge (containing_loop), | |
9e227d60 | 8490 | UNKNOWN_LOCATION); |
272c6793 RS |
8491 | msq = lsq; |
8492 | } | |
ebfd146a | 8493 | } |
ebfd146a | 8494 | |
59fd17e3 RB |
8495 | /* 4. Handle invariant-load. */ |
8496 | if (inv_p && !bb_vinfo) | |
8497 | { | |
59fd17e3 | 8498 | gcc_assert (!grouped_load); |
d1417442 JJ |
8499 | /* If we have versioned for aliasing or the loop doesn't |
8500 | have any data dependencies that would preclude this, | |
8501 | then we are sure this is a loop invariant load and | |
8502 | thus we can insert it on the preheader edge. */ | |
8503 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
8504 | && !nested_in_vect_loop | |
6b916b36 | 8505 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
8506 | { |
8507 | if (dump_enabled_p ()) | |
8508 | { | |
8509 | dump_printf_loc (MSG_NOTE, vect_location, | |
8510 | "hoisting out of the vectorized " | |
8511 | "loop: "); | |
8512 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 8513 | } |
b731b390 | 8514 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
8515 | gsi_insert_on_edge_immediate |
8516 | (loop_preheader_edge (loop), | |
8517 | gimple_build_assign (tem, | |
8518 | unshare_expr | |
8519 | (gimple_assign_rhs1 (stmt)))); | |
8520 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
34cd48e5 | 8521 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
e1bd7296 | 8522 | new_stmt_info = vinfo->add_stmt (new_stmt); |
a0e35eb0 RB |
8523 | } |
8524 | else | |
8525 | { | |
8526 | gimple_stmt_iterator gsi2 = *gsi; | |
8527 | gsi_next (&gsi2); | |
8528 | new_temp = vect_init_vector (stmt, scalar_dest, | |
8529 | vectype, &gsi2); | |
e1bd7296 | 8530 | new_stmt_info = vinfo->lookup_def (new_temp); |
a0e35eb0 | 8531 | } |
59fd17e3 RB |
8532 | } |
8533 | ||
62da9e14 | 8534 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
272c6793 | 8535 | { |
aec7ae7d JJ |
8536 | tree perm_mask = perm_mask_for_reverse (vectype); |
8537 | new_temp = permute_vec_elements (new_temp, new_temp, | |
8538 | perm_mask, stmt, gsi); | |
e1bd7296 | 8539 | new_stmt_info = vinfo->lookup_def (new_temp); |
ebfd146a | 8540 | } |
267d3070 | 8541 | |
272c6793 | 8542 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
8543 | vect_transform_grouped_load (). */ |
8544 | if (grouped_load || slp_perm) | |
9771b263 | 8545 | dr_chain.quick_push (new_temp); |
267d3070 | 8546 | |
272c6793 RS |
8547 | /* Store vector loads in the corresponding SLP_NODE. */ |
8548 | if (slp && !slp_perm) | |
e1bd7296 | 8549 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
b267968e RB |
8550 | |
8551 | /* With SLP permutation we load the gaps as well, without | |
8552 | we need to skip the gaps after we manage to fully load | |
2c53b149 | 8553 | all elements. group_gap_adj is DR_GROUP_SIZE here. */ |
b267968e | 8554 | group_elt += nunits; |
d9f21f6a RS |
8555 | if (maybe_ne (group_gap_adj, 0U) |
8556 | && !slp_perm | |
8557 | && known_eq (group_elt, group_size - group_gap_adj)) | |
b267968e | 8558 | { |
d9f21f6a RS |
8559 | poly_wide_int bump_val |
8560 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
8561 | * group_gap_adj); | |
8e6cdc90 | 8562 | tree bump = wide_int_to_tree (sizetype, bump_val); |
b267968e RB |
8563 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8564 | stmt, bump); | |
8565 | group_elt = 0; | |
8566 | } | |
272c6793 | 8567 | } |
9b999e8c RB |
8568 | /* Bump the vector pointer to account for a gap or for excess |
8569 | elements loaded for a permuted SLP load. */ | |
d9f21f6a | 8570 | if (maybe_ne (group_gap_adj, 0U) && slp_perm) |
a64b9c26 | 8571 | { |
d9f21f6a RS |
8572 | poly_wide_int bump_val |
8573 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
8574 | * group_gap_adj); | |
8e6cdc90 | 8575 | tree bump = wide_int_to_tree (sizetype, bump_val); |
a64b9c26 RB |
8576 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8577 | stmt, bump); | |
8578 | } | |
ebfd146a IR |
8579 | } |
8580 | ||
8581 | if (slp && !slp_perm) | |
8582 | continue; | |
8583 | ||
8584 | if (slp_perm) | |
8585 | { | |
29afecdf | 8586 | unsigned n_perms; |
01d8bf07 | 8587 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
29afecdf RB |
8588 | slp_node_instance, false, |
8589 | &n_perms)) | |
ebfd146a | 8590 | { |
9771b263 | 8591 | dr_chain.release (); |
ebfd146a IR |
8592 | return false; |
8593 | } | |
8594 | } | |
8595 | else | |
8596 | { | |
0d0293ac | 8597 | if (grouped_load) |
ebfd146a | 8598 | { |
2de001ee | 8599 | if (memory_access_type != VMAT_LOAD_STORE_LANES) |
0d0293ac | 8600 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 8601 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
8602 | } |
8603 | else | |
8604 | { | |
8605 | if (j == 0) | |
e1bd7296 | 8606 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
ebfd146a | 8607 | else |
e1bd7296 RS |
8608 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
8609 | prev_stmt_info = new_stmt_info; | |
ebfd146a IR |
8610 | } |
8611 | } | |
9771b263 | 8612 | dr_chain.release (); |
ebfd146a IR |
8613 | } |
8614 | ||
ebfd146a IR |
8615 | return true; |
8616 | } | |
8617 | ||
8618 | /* Function vect_is_simple_cond. | |
b8698a0f | 8619 | |
ebfd146a IR |
8620 | Input: |
8621 | LOOP - the loop that is being vectorized. | |
8622 | COND - Condition that is checked for simple use. | |
8623 | ||
e9e1d143 RG |
8624 | Output: |
8625 | *COMP_VECTYPE - the vector type for the comparison. | |
4fc5ebf1 | 8626 | *DTS - The def types for the arguments of the comparison |
e9e1d143 | 8627 | |
ebfd146a IR |
8628 | Returns whether a COND can be vectorized. Checks whether |
8629 | condition operands are supportable using vec_is_simple_use. */ | |
8630 | ||
87aab9b2 | 8631 | static bool |
4fc5ebf1 | 8632 | vect_is_simple_cond (tree cond, vec_info *vinfo, |
8da4c8d8 RB |
8633 | tree *comp_vectype, enum vect_def_type *dts, |
8634 | tree vectype) | |
ebfd146a IR |
8635 | { |
8636 | tree lhs, rhs; | |
e9e1d143 | 8637 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 8638 | |
a414c77f IE |
8639 | /* Mask case. */ |
8640 | if (TREE_CODE (cond) == SSA_NAME | |
2568d8a1 | 8641 | && VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (cond))) |
a414c77f | 8642 | { |
894dd753 | 8643 | if (!vect_is_simple_use (cond, vinfo, &dts[0], comp_vectype) |
a414c77f IE |
8644 | || !*comp_vectype |
8645 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
8646 | return false; | |
8647 | return true; | |
8648 | } | |
8649 | ||
ebfd146a IR |
8650 | if (!COMPARISON_CLASS_P (cond)) |
8651 | return false; | |
8652 | ||
8653 | lhs = TREE_OPERAND (cond, 0); | |
8654 | rhs = TREE_OPERAND (cond, 1); | |
8655 | ||
8656 | if (TREE_CODE (lhs) == SSA_NAME) | |
8657 | { | |
894dd753 | 8658 | if (!vect_is_simple_use (lhs, vinfo, &dts[0], &vectype1)) |
ebfd146a IR |
8659 | return false; |
8660 | } | |
4fc5ebf1 JG |
8661 | else if (TREE_CODE (lhs) == INTEGER_CST || TREE_CODE (lhs) == REAL_CST |
8662 | || TREE_CODE (lhs) == FIXED_CST) | |
8663 | dts[0] = vect_constant_def; | |
8664 | else | |
ebfd146a IR |
8665 | return false; |
8666 | ||
8667 | if (TREE_CODE (rhs) == SSA_NAME) | |
8668 | { | |
894dd753 | 8669 | if (!vect_is_simple_use (rhs, vinfo, &dts[1], &vectype2)) |
ebfd146a IR |
8670 | return false; |
8671 | } | |
4fc5ebf1 JG |
8672 | else if (TREE_CODE (rhs) == INTEGER_CST || TREE_CODE (rhs) == REAL_CST |
8673 | || TREE_CODE (rhs) == FIXED_CST) | |
8674 | dts[1] = vect_constant_def; | |
8675 | else | |
ebfd146a IR |
8676 | return false; |
8677 | ||
28b33016 | 8678 | if (vectype1 && vectype2 |
928686b1 RS |
8679 | && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1), |
8680 | TYPE_VECTOR_SUBPARTS (vectype2))) | |
28b33016 IE |
8681 | return false; |
8682 | ||
e9e1d143 | 8683 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
8da4c8d8 | 8684 | /* Invariant comparison. */ |
4515e413 | 8685 | if (! *comp_vectype && vectype) |
8da4c8d8 RB |
8686 | { |
8687 | tree scalar_type = TREE_TYPE (lhs); | |
8688 | /* If we can widen the comparison to match vectype do so. */ | |
8689 | if (INTEGRAL_TYPE_P (scalar_type) | |
8690 | && tree_int_cst_lt (TYPE_SIZE (scalar_type), | |
8691 | TYPE_SIZE (TREE_TYPE (vectype)))) | |
8692 | scalar_type = build_nonstandard_integer_type | |
8693 | (tree_to_uhwi (TYPE_SIZE (TREE_TYPE (vectype))), | |
8694 | TYPE_UNSIGNED (scalar_type)); | |
8695 | *comp_vectype = get_vectype_for_scalar_type (scalar_type); | |
8696 | } | |
8697 | ||
ebfd146a IR |
8698 | return true; |
8699 | } | |
8700 | ||
8701 | /* vectorizable_condition. | |
8702 | ||
b8698a0f L |
8703 | Check if STMT is conditional modify expression that can be vectorized. |
8704 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
8705 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
8706 | at GSI. |
8707 | ||
8708 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
8709 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 8710 | else clause if it is 2). |
ebfd146a IR |
8711 | |
8712 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
8713 | ||
4bbe8262 | 8714 | bool |
355fe088 | 8715 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
1eede195 RS |
8716 | stmt_vec_info *vec_stmt, tree reduc_def, |
8717 | int reduc_index, slp_tree slp_node, | |
8718 | stmt_vector_for_cost *cost_vec) | |
ebfd146a IR |
8719 | { |
8720 | tree scalar_dest = NULL_TREE; | |
8721 | tree vec_dest = NULL_TREE; | |
01216d27 JJ |
8722 | tree cond_expr, cond_expr0 = NULL_TREE, cond_expr1 = NULL_TREE; |
8723 | tree then_clause, else_clause; | |
ebfd146a | 8724 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
df11cc78 | 8725 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
8726 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
8727 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
5958f9e2 | 8728 | tree vec_compare; |
ebfd146a IR |
8729 | tree new_temp; |
8730 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4fc5ebf1 JG |
8731 | enum vect_def_type dts[4] |
8732 | = {vect_unknown_def_type, vect_unknown_def_type, | |
8733 | vect_unknown_def_type, vect_unknown_def_type}; | |
8734 | int ndts = 4; | |
f7e531cf | 8735 | int ncopies; |
01216d27 | 8736 | enum tree_code code, cond_code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
a855b1b1 | 8737 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
8738 | int i, j; |
8739 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
8740 | vec<tree> vec_oprnds0 = vNULL; |
8741 | vec<tree> vec_oprnds1 = vNULL; | |
8742 | vec<tree> vec_oprnds2 = vNULL; | |
8743 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 8744 | tree vec_cmp_type; |
a414c77f | 8745 | bool masked = false; |
b8698a0f | 8746 | |
f7e531cf IR |
8747 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
8748 | return false; | |
8749 | ||
bb6c2b68 RS |
8750 | vect_reduction_type reduction_type |
8751 | = STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info); | |
8752 | if (reduction_type == TREE_CODE_REDUCTION) | |
af29617a AH |
8753 | { |
8754 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
8755 | return false; | |
ebfd146a | 8756 | |
af29617a AH |
8757 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
8758 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
8759 | && reduc_def)) | |
8760 | return false; | |
ebfd146a | 8761 | |
af29617a AH |
8762 | /* FORNOW: not yet supported. */ |
8763 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
8764 | { | |
8765 | if (dump_enabled_p ()) | |
8766 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
8767 | "value used after loop.\n"); | |
8768 | return false; | |
8769 | } | |
ebfd146a IR |
8770 | } |
8771 | ||
8772 | /* Is vectorizable conditional operation? */ | |
8773 | if (!is_gimple_assign (stmt)) | |
8774 | return false; | |
8775 | ||
8776 | code = gimple_assign_rhs_code (stmt); | |
8777 | ||
8778 | if (code != COND_EXPR) | |
8779 | return false; | |
8780 | ||
465c8c19 | 8781 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2947d3b2 | 8782 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
465c8c19 | 8783 | |
fce57248 | 8784 | if (slp_node) |
465c8c19 JJ |
8785 | ncopies = 1; |
8786 | else | |
e8f142e2 | 8787 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
8788 | |
8789 | gcc_assert (ncopies >= 1); | |
8790 | if (reduc_index && ncopies > 1) | |
8791 | return false; /* FORNOW */ | |
8792 | ||
4e71066d RG |
8793 | cond_expr = gimple_assign_rhs1 (stmt); |
8794 | then_clause = gimple_assign_rhs2 (stmt); | |
8795 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 8796 | |
4fc5ebf1 | 8797 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, |
4515e413 | 8798 | &comp_vectype, &dts[0], slp_node ? NULL : vectype) |
e9e1d143 | 8799 | || !comp_vectype) |
ebfd146a IR |
8800 | return false; |
8801 | ||
894dd753 | 8802 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &dts[2], &vectype1)) |
2947d3b2 | 8803 | return false; |
894dd753 | 8804 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &dts[3], &vectype2)) |
ebfd146a | 8805 | return false; |
2947d3b2 IE |
8806 | |
8807 | if (vectype1 && !useless_type_conversion_p (vectype, vectype1)) | |
8808 | return false; | |
8809 | ||
8810 | if (vectype2 && !useless_type_conversion_p (vectype, vectype2)) | |
ebfd146a IR |
8811 | return false; |
8812 | ||
28b33016 IE |
8813 | masked = !COMPARISON_CLASS_P (cond_expr); |
8814 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
8815 | ||
74946978 MP |
8816 | if (vec_cmp_type == NULL_TREE) |
8817 | return false; | |
784fb9b3 | 8818 | |
01216d27 JJ |
8819 | cond_code = TREE_CODE (cond_expr); |
8820 | if (!masked) | |
8821 | { | |
8822 | cond_expr0 = TREE_OPERAND (cond_expr, 0); | |
8823 | cond_expr1 = TREE_OPERAND (cond_expr, 1); | |
8824 | } | |
8825 | ||
8826 | if (!masked && VECTOR_BOOLEAN_TYPE_P (comp_vectype)) | |
8827 | { | |
8828 | /* Boolean values may have another representation in vectors | |
8829 | and therefore we prefer bit operations over comparison for | |
8830 | them (which also works for scalar masks). We store opcodes | |
8831 | to use in bitop1 and bitop2. Statement is vectorized as | |
8832 | BITOP2 (rhs1 BITOP1 rhs2) or rhs1 BITOP2 (BITOP1 rhs2) | |
8833 | depending on bitop1 and bitop2 arity. */ | |
8834 | switch (cond_code) | |
8835 | { | |
8836 | case GT_EXPR: | |
8837 | bitop1 = BIT_NOT_EXPR; | |
8838 | bitop2 = BIT_AND_EXPR; | |
8839 | break; | |
8840 | case GE_EXPR: | |
8841 | bitop1 = BIT_NOT_EXPR; | |
8842 | bitop2 = BIT_IOR_EXPR; | |
8843 | break; | |
8844 | case LT_EXPR: | |
8845 | bitop1 = BIT_NOT_EXPR; | |
8846 | bitop2 = BIT_AND_EXPR; | |
8847 | std::swap (cond_expr0, cond_expr1); | |
8848 | break; | |
8849 | case LE_EXPR: | |
8850 | bitop1 = BIT_NOT_EXPR; | |
8851 | bitop2 = BIT_IOR_EXPR; | |
8852 | std::swap (cond_expr0, cond_expr1); | |
8853 | break; | |
8854 | case NE_EXPR: | |
8855 | bitop1 = BIT_XOR_EXPR; | |
8856 | break; | |
8857 | case EQ_EXPR: | |
8858 | bitop1 = BIT_XOR_EXPR; | |
8859 | bitop2 = BIT_NOT_EXPR; | |
8860 | break; | |
8861 | default: | |
8862 | return false; | |
8863 | } | |
8864 | cond_code = SSA_NAME; | |
8865 | } | |
8866 | ||
b8698a0f | 8867 | if (!vec_stmt) |
ebfd146a | 8868 | { |
01216d27 JJ |
8869 | if (bitop1 != NOP_EXPR) |
8870 | { | |
8871 | machine_mode mode = TYPE_MODE (comp_vectype); | |
8872 | optab optab; | |
8873 | ||
8874 | optab = optab_for_tree_code (bitop1, comp_vectype, optab_default); | |
8875 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8876 | return false; | |
8877 | ||
8878 | if (bitop2 != NOP_EXPR) | |
8879 | { | |
8880 | optab = optab_for_tree_code (bitop2, comp_vectype, | |
8881 | optab_default); | |
8882 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8883 | return false; | |
8884 | } | |
8885 | } | |
4fc5ebf1 JG |
8886 | if (expand_vec_cond_expr_p (vectype, comp_vectype, |
8887 | cond_code)) | |
8888 | { | |
68435eb2 RB |
8889 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; |
8890 | vect_model_simple_cost (stmt_info, ncopies, dts, ndts, slp_node, | |
8891 | cost_vec); | |
4fc5ebf1 JG |
8892 | return true; |
8893 | } | |
8894 | return false; | |
ebfd146a IR |
8895 | } |
8896 | ||
f7e531cf IR |
8897 | /* Transform. */ |
8898 | ||
8899 | if (!slp_node) | |
8900 | { | |
9771b263 DN |
8901 | vec_oprnds0.create (1); |
8902 | vec_oprnds1.create (1); | |
8903 | vec_oprnds2.create (1); | |
8904 | vec_oprnds3.create (1); | |
f7e531cf | 8905 | } |
ebfd146a IR |
8906 | |
8907 | /* Handle def. */ | |
8908 | scalar_dest = gimple_assign_lhs (stmt); | |
bb6c2b68 RS |
8909 | if (reduction_type != EXTRACT_LAST_REDUCTION) |
8910 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
ebfd146a IR |
8911 | |
8912 | /* Handle cond expr. */ | |
a855b1b1 MM |
8913 | for (j = 0; j < ncopies; j++) |
8914 | { | |
e1bd7296 | 8915 | stmt_vec_info new_stmt_info = NULL; |
a855b1b1 MM |
8916 | if (j == 0) |
8917 | { | |
f7e531cf IR |
8918 | if (slp_node) |
8919 | { | |
00f96dc9 TS |
8920 | auto_vec<tree, 4> ops; |
8921 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 8922 | |
a414c77f | 8923 | if (masked) |
01216d27 | 8924 | ops.safe_push (cond_expr); |
a414c77f IE |
8925 | else |
8926 | { | |
01216d27 JJ |
8927 | ops.safe_push (cond_expr0); |
8928 | ops.safe_push (cond_expr1); | |
a414c77f | 8929 | } |
9771b263 DN |
8930 | ops.safe_push (then_clause); |
8931 | ops.safe_push (else_clause); | |
306b0c92 | 8932 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
37b5ec8f JJ |
8933 | vec_oprnds3 = vec_defs.pop (); |
8934 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
8935 | if (!masked) |
8936 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 8937 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf IR |
8938 | } |
8939 | else | |
8940 | { | |
a414c77f IE |
8941 | if (masked) |
8942 | { | |
8943 | vec_cond_lhs | |
8944 | = vect_get_vec_def_for_operand (cond_expr, stmt, | |
8945 | comp_vectype); | |
894dd753 | 8946 | vect_is_simple_use (cond_expr, stmt_info->vinfo, &dts[0]); |
a414c77f IE |
8947 | } |
8948 | else | |
8949 | { | |
01216d27 JJ |
8950 | vec_cond_lhs |
8951 | = vect_get_vec_def_for_operand (cond_expr0, | |
8952 | stmt, comp_vectype); | |
894dd753 | 8953 | vect_is_simple_use (cond_expr0, loop_vinfo, &dts[0]); |
01216d27 JJ |
8954 | |
8955 | vec_cond_rhs | |
8956 | = vect_get_vec_def_for_operand (cond_expr1, | |
8957 | stmt, comp_vectype); | |
894dd753 | 8958 | vect_is_simple_use (cond_expr1, loop_vinfo, &dts[1]); |
a414c77f | 8959 | } |
f7e531cf IR |
8960 | if (reduc_index == 1) |
8961 | vec_then_clause = reduc_def; | |
8962 | else | |
8963 | { | |
8964 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 | 8965 | stmt); |
894dd753 | 8966 | vect_is_simple_use (then_clause, loop_vinfo, &dts[2]); |
f7e531cf IR |
8967 | } |
8968 | if (reduc_index == 2) | |
8969 | vec_else_clause = reduc_def; | |
8970 | else | |
8971 | { | |
8972 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 | 8973 | stmt); |
894dd753 | 8974 | vect_is_simple_use (else_clause, loop_vinfo, &dts[3]); |
f7e531cf | 8975 | } |
a855b1b1 MM |
8976 | } |
8977 | } | |
8978 | else | |
8979 | { | |
a414c77f IE |
8980 | vec_cond_lhs |
8981 | = vect_get_vec_def_for_stmt_copy (dts[0], | |
8982 | vec_oprnds0.pop ()); | |
8983 | if (!masked) | |
8984 | vec_cond_rhs | |
8985 | = vect_get_vec_def_for_stmt_copy (dts[1], | |
8986 | vec_oprnds1.pop ()); | |
8987 | ||
a855b1b1 | 8988 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 8989 | vec_oprnds2.pop ()); |
a855b1b1 | 8990 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 8991 | vec_oprnds3.pop ()); |
f7e531cf IR |
8992 | } |
8993 | ||
8994 | if (!slp_node) | |
8995 | { | |
9771b263 | 8996 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
8997 | if (!masked) |
8998 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
8999 | vec_oprnds2.quick_push (vec_then_clause); |
9000 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
9001 | } |
9002 | ||
9dc3f7de | 9003 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 9004 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 9005 | { |
9771b263 DN |
9006 | vec_then_clause = vec_oprnds2[i]; |
9007 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 9008 | |
a414c77f IE |
9009 | if (masked) |
9010 | vec_compare = vec_cond_lhs; | |
9011 | else | |
9012 | { | |
9013 | vec_cond_rhs = vec_oprnds1[i]; | |
01216d27 JJ |
9014 | if (bitop1 == NOP_EXPR) |
9015 | vec_compare = build2 (cond_code, vec_cmp_type, | |
9016 | vec_cond_lhs, vec_cond_rhs); | |
9017 | else | |
9018 | { | |
9019 | new_temp = make_ssa_name (vec_cmp_type); | |
e1bd7296 | 9020 | gassign *new_stmt; |
01216d27 JJ |
9021 | if (bitop1 == BIT_NOT_EXPR) |
9022 | new_stmt = gimple_build_assign (new_temp, bitop1, | |
9023 | vec_cond_rhs); | |
9024 | else | |
9025 | new_stmt | |
9026 | = gimple_build_assign (new_temp, bitop1, vec_cond_lhs, | |
9027 | vec_cond_rhs); | |
9028 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9029 | if (bitop2 == NOP_EXPR) | |
9030 | vec_compare = new_temp; | |
9031 | else if (bitop2 == BIT_NOT_EXPR) | |
9032 | { | |
9033 | /* Instead of doing ~x ? y : z do x ? z : y. */ | |
9034 | vec_compare = new_temp; | |
9035 | std::swap (vec_then_clause, vec_else_clause); | |
9036 | } | |
9037 | else | |
9038 | { | |
9039 | vec_compare = make_ssa_name (vec_cmp_type); | |
9040 | new_stmt | |
9041 | = gimple_build_assign (vec_compare, bitop2, | |
9042 | vec_cond_lhs, new_temp); | |
9043 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9044 | } | |
9045 | } | |
a414c77f | 9046 | } |
bb6c2b68 RS |
9047 | if (reduction_type == EXTRACT_LAST_REDUCTION) |
9048 | { | |
9049 | if (!is_gimple_val (vec_compare)) | |
9050 | { | |
9051 | tree vec_compare_name = make_ssa_name (vec_cmp_type); | |
e1bd7296 RS |
9052 | gassign *new_stmt = gimple_build_assign (vec_compare_name, |
9053 | vec_compare); | |
bb6c2b68 RS |
9054 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9055 | vec_compare = vec_compare_name; | |
9056 | } | |
9057 | gcc_assert (reduc_index == 2); | |
e1bd7296 | 9058 | gcall *new_stmt = gimple_build_call_internal |
bb6c2b68 RS |
9059 | (IFN_FOLD_EXTRACT_LAST, 3, else_clause, vec_compare, |
9060 | vec_then_clause); | |
9061 | gimple_call_set_lhs (new_stmt, scalar_dest); | |
9062 | SSA_NAME_DEF_STMT (scalar_dest) = new_stmt; | |
9063 | if (stmt == gsi_stmt (*gsi)) | |
e1bd7296 | 9064 | new_stmt_info = vect_finish_replace_stmt (stmt, new_stmt); |
bb6c2b68 RS |
9065 | else |
9066 | { | |
9067 | /* In this case we're moving the definition to later in the | |
9068 | block. That doesn't matter because the only uses of the | |
9069 | lhs are in phi statements. */ | |
9070 | gimple_stmt_iterator old_gsi = gsi_for_stmt (stmt); | |
9071 | gsi_remove (&old_gsi, true); | |
e1bd7296 RS |
9072 | new_stmt_info |
9073 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
bb6c2b68 RS |
9074 | } |
9075 | } | |
9076 | else | |
9077 | { | |
9078 | new_temp = make_ssa_name (vec_dest); | |
e1bd7296 RS |
9079 | gassign *new_stmt |
9080 | = gimple_build_assign (new_temp, VEC_COND_EXPR, vec_compare, | |
9081 | vec_then_clause, vec_else_clause); | |
9082 | new_stmt_info | |
9083 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
bb6c2b68 | 9084 | } |
f7e531cf | 9085 | if (slp_node) |
e1bd7296 | 9086 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
f7e531cf IR |
9087 | } |
9088 | ||
9089 | if (slp_node) | |
9090 | continue; | |
9091 | ||
e1bd7296 RS |
9092 | if (j == 0) |
9093 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; | |
9094 | else | |
9095 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; | |
f7e531cf | 9096 | |
e1bd7296 | 9097 | prev_stmt_info = new_stmt_info; |
a855b1b1 | 9098 | } |
b8698a0f | 9099 | |
9771b263 DN |
9100 | vec_oprnds0.release (); |
9101 | vec_oprnds1.release (); | |
9102 | vec_oprnds2.release (); | |
9103 | vec_oprnds3.release (); | |
f7e531cf | 9104 | |
ebfd146a IR |
9105 | return true; |
9106 | } | |
9107 | ||
42fd8198 IE |
9108 | /* vectorizable_comparison. |
9109 | ||
9110 | Check if STMT is comparison expression that can be vectorized. | |
9111 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
9112 | comparison, put it in VEC_STMT, and insert it at GSI. | |
9113 | ||
9114 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
9115 | ||
fce57248 | 9116 | static bool |
42fd8198 | 9117 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, |
1eede195 | 9118 | stmt_vec_info *vec_stmt, tree reduc_def, |
68435eb2 | 9119 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
42fd8198 IE |
9120 | { |
9121 | tree lhs, rhs1, rhs2; | |
9122 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9123 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
9124 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
9125 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
9126 | tree new_temp; | |
9127 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
9128 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 9129 | int ndts = 2; |
928686b1 | 9130 | poly_uint64 nunits; |
42fd8198 | 9131 | int ncopies; |
49e76ff1 | 9132 | enum tree_code code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
42fd8198 IE |
9133 | stmt_vec_info prev_stmt_info = NULL; |
9134 | int i, j; | |
9135 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
9136 | vec<tree> vec_oprnds0 = vNULL; | |
9137 | vec<tree> vec_oprnds1 = vNULL; | |
42fd8198 IE |
9138 | tree mask_type; |
9139 | tree mask; | |
9140 | ||
c245362b IE |
9141 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
9142 | return false; | |
9143 | ||
30480bcd | 9144 | if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)) |
42fd8198 IE |
9145 | return false; |
9146 | ||
9147 | mask_type = vectype; | |
9148 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
9149 | ||
fce57248 | 9150 | if (slp_node) |
42fd8198 IE |
9151 | ncopies = 1; |
9152 | else | |
e8f142e2 | 9153 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
42fd8198 IE |
9154 | |
9155 | gcc_assert (ncopies >= 1); | |
42fd8198 IE |
9156 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
9157 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
9158 | && reduc_def)) | |
9159 | return false; | |
9160 | ||
9161 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
9162 | { | |
9163 | if (dump_enabled_p ()) | |
9164 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
9165 | "value used after loop.\n"); | |
9166 | return false; | |
9167 | } | |
9168 | ||
9169 | if (!is_gimple_assign (stmt)) | |
9170 | return false; | |
9171 | ||
9172 | code = gimple_assign_rhs_code (stmt); | |
9173 | ||
9174 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
9175 | return false; | |
9176 | ||
9177 | rhs1 = gimple_assign_rhs1 (stmt); | |
9178 | rhs2 = gimple_assign_rhs2 (stmt); | |
9179 | ||
894dd753 | 9180 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &dts[0], &vectype1)) |
42fd8198 IE |
9181 | return false; |
9182 | ||
894dd753 | 9183 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &dts[1], &vectype2)) |
42fd8198 IE |
9184 | return false; |
9185 | ||
9186 | if (vectype1 && vectype2 | |
928686b1 RS |
9187 | && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1), |
9188 | TYPE_VECTOR_SUBPARTS (vectype2))) | |
42fd8198 IE |
9189 | return false; |
9190 | ||
9191 | vectype = vectype1 ? vectype1 : vectype2; | |
9192 | ||
9193 | /* Invariant comparison. */ | |
9194 | if (!vectype) | |
9195 | { | |
69a9a66f | 9196 | vectype = get_vectype_for_scalar_type (TREE_TYPE (rhs1)); |
928686b1 | 9197 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (vectype), nunits)) |
42fd8198 IE |
9198 | return false; |
9199 | } | |
928686b1 | 9200 | else if (maybe_ne (nunits, TYPE_VECTOR_SUBPARTS (vectype))) |
42fd8198 IE |
9201 | return false; |
9202 | ||
49e76ff1 IE |
9203 | /* Can't compare mask and non-mask types. */ |
9204 | if (vectype1 && vectype2 | |
9205 | && (VECTOR_BOOLEAN_TYPE_P (vectype1) ^ VECTOR_BOOLEAN_TYPE_P (vectype2))) | |
9206 | return false; | |
9207 | ||
9208 | /* Boolean values may have another representation in vectors | |
9209 | and therefore we prefer bit operations over comparison for | |
9210 | them (which also works for scalar masks). We store opcodes | |
9211 | to use in bitop1 and bitop2. Statement is vectorized as | |
9212 | BITOP2 (rhs1 BITOP1 rhs2) or | |
9213 | rhs1 BITOP2 (BITOP1 rhs2) | |
9214 | depending on bitop1 and bitop2 arity. */ | |
9215 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) | |
9216 | { | |
9217 | if (code == GT_EXPR) | |
9218 | { | |
9219 | bitop1 = BIT_NOT_EXPR; | |
9220 | bitop2 = BIT_AND_EXPR; | |
9221 | } | |
9222 | else if (code == GE_EXPR) | |
9223 | { | |
9224 | bitop1 = BIT_NOT_EXPR; | |
9225 | bitop2 = BIT_IOR_EXPR; | |
9226 | } | |
9227 | else if (code == LT_EXPR) | |
9228 | { | |
9229 | bitop1 = BIT_NOT_EXPR; | |
9230 | bitop2 = BIT_AND_EXPR; | |
9231 | std::swap (rhs1, rhs2); | |
264d951a | 9232 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
9233 | } |
9234 | else if (code == LE_EXPR) | |
9235 | { | |
9236 | bitop1 = BIT_NOT_EXPR; | |
9237 | bitop2 = BIT_IOR_EXPR; | |
9238 | std::swap (rhs1, rhs2); | |
264d951a | 9239 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
9240 | } |
9241 | else | |
9242 | { | |
9243 | bitop1 = BIT_XOR_EXPR; | |
9244 | if (code == EQ_EXPR) | |
9245 | bitop2 = BIT_NOT_EXPR; | |
9246 | } | |
9247 | } | |
9248 | ||
42fd8198 IE |
9249 | if (!vec_stmt) |
9250 | { | |
49e76ff1 | 9251 | if (bitop1 == NOP_EXPR) |
68435eb2 RB |
9252 | { |
9253 | if (!expand_vec_cmp_expr_p (vectype, mask_type, code)) | |
9254 | return false; | |
9255 | } | |
49e76ff1 IE |
9256 | else |
9257 | { | |
9258 | machine_mode mode = TYPE_MODE (vectype); | |
9259 | optab optab; | |
9260 | ||
9261 | optab = optab_for_tree_code (bitop1, vectype, optab_default); | |
9262 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
9263 | return false; | |
9264 | ||
9265 | if (bitop2 != NOP_EXPR) | |
9266 | { | |
9267 | optab = optab_for_tree_code (bitop2, vectype, optab_default); | |
9268 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
9269 | return false; | |
9270 | } | |
49e76ff1 | 9271 | } |
68435eb2 RB |
9272 | |
9273 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
9274 | vect_model_simple_cost (stmt_info, ncopies * (1 + (bitop2 != NOP_EXPR)), | |
9275 | dts, ndts, slp_node, cost_vec); | |
9276 | return true; | |
42fd8198 IE |
9277 | } |
9278 | ||
9279 | /* Transform. */ | |
9280 | if (!slp_node) | |
9281 | { | |
9282 | vec_oprnds0.create (1); | |
9283 | vec_oprnds1.create (1); | |
9284 | } | |
9285 | ||
9286 | /* Handle def. */ | |
9287 | lhs = gimple_assign_lhs (stmt); | |
9288 | mask = vect_create_destination_var (lhs, mask_type); | |
9289 | ||
9290 | /* Handle cmp expr. */ | |
9291 | for (j = 0; j < ncopies; j++) | |
9292 | { | |
e1bd7296 | 9293 | stmt_vec_info new_stmt_info = NULL; |
42fd8198 IE |
9294 | if (j == 0) |
9295 | { | |
9296 | if (slp_node) | |
9297 | { | |
9298 | auto_vec<tree, 2> ops; | |
9299 | auto_vec<vec<tree>, 2> vec_defs; | |
9300 | ||
9301 | ops.safe_push (rhs1); | |
9302 | ops.safe_push (rhs2); | |
306b0c92 | 9303 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
42fd8198 IE |
9304 | vec_oprnds1 = vec_defs.pop (); |
9305 | vec_oprnds0 = vec_defs.pop (); | |
9306 | } | |
9307 | else | |
9308 | { | |
e4af0bc4 IE |
9309 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, vectype); |
9310 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, vectype); | |
42fd8198 IE |
9311 | } |
9312 | } | |
9313 | else | |
9314 | { | |
9315 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
9316 | vec_oprnds0.pop ()); | |
9317 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
9318 | vec_oprnds1.pop ()); | |
9319 | } | |
9320 | ||
9321 | if (!slp_node) | |
9322 | { | |
9323 | vec_oprnds0.quick_push (vec_rhs1); | |
9324 | vec_oprnds1.quick_push (vec_rhs2); | |
9325 | } | |
9326 | ||
9327 | /* Arguments are ready. Create the new vector stmt. */ | |
9328 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
9329 | { | |
9330 | vec_rhs2 = vec_oprnds1[i]; | |
9331 | ||
9332 | new_temp = make_ssa_name (mask); | |
49e76ff1 IE |
9333 | if (bitop1 == NOP_EXPR) |
9334 | { | |
e1bd7296 RS |
9335 | gassign *new_stmt = gimple_build_assign (new_temp, code, |
9336 | vec_rhs1, vec_rhs2); | |
9337 | new_stmt_info | |
9338 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
49e76ff1 IE |
9339 | } |
9340 | else | |
9341 | { | |
e1bd7296 | 9342 | gassign *new_stmt; |
49e76ff1 IE |
9343 | if (bitop1 == BIT_NOT_EXPR) |
9344 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs2); | |
9345 | else | |
9346 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs1, | |
9347 | vec_rhs2); | |
e1bd7296 RS |
9348 | new_stmt_info |
9349 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
49e76ff1 IE |
9350 | if (bitop2 != NOP_EXPR) |
9351 | { | |
9352 | tree res = make_ssa_name (mask); | |
9353 | if (bitop2 == BIT_NOT_EXPR) | |
9354 | new_stmt = gimple_build_assign (res, bitop2, new_temp); | |
9355 | else | |
9356 | new_stmt = gimple_build_assign (res, bitop2, vec_rhs1, | |
9357 | new_temp); | |
e1bd7296 RS |
9358 | new_stmt_info |
9359 | = vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
49e76ff1 IE |
9360 | } |
9361 | } | |
42fd8198 | 9362 | if (slp_node) |
e1bd7296 | 9363 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
42fd8198 IE |
9364 | } |
9365 | ||
9366 | if (slp_node) | |
9367 | continue; | |
9368 | ||
9369 | if (j == 0) | |
e1bd7296 | 9370 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
42fd8198 | 9371 | else |
e1bd7296 | 9372 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
42fd8198 | 9373 | |
e1bd7296 | 9374 | prev_stmt_info = new_stmt_info; |
42fd8198 IE |
9375 | } |
9376 | ||
9377 | vec_oprnds0.release (); | |
9378 | vec_oprnds1.release (); | |
9379 | ||
9380 | return true; | |
9381 | } | |
ebfd146a | 9382 | |
68a0f2ff RS |
9383 | /* If SLP_NODE is nonnull, return true if vectorizable_live_operation |
9384 | can handle all live statements in the node. Otherwise return true | |
9385 | if STMT is not live or if vectorizable_live_operation can handle it. | |
9386 | GSI and VEC_STMT are as for vectorizable_live_operation. */ | |
9387 | ||
9388 | static bool | |
9389 | can_vectorize_live_stmts (gimple *stmt, gimple_stmt_iterator *gsi, | |
1eede195 | 9390 | slp_tree slp_node, stmt_vec_info *vec_stmt, |
68435eb2 | 9391 | stmt_vector_for_cost *cost_vec) |
68a0f2ff RS |
9392 | { |
9393 | if (slp_node) | |
9394 | { | |
b9787581 | 9395 | stmt_vec_info slp_stmt_info; |
68a0f2ff | 9396 | unsigned int i; |
b9787581 | 9397 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, slp_stmt_info) |
68a0f2ff | 9398 | { |
68a0f2ff | 9399 | if (STMT_VINFO_LIVE_P (slp_stmt_info) |
b9787581 | 9400 | && !vectorizable_live_operation (slp_stmt_info, gsi, slp_node, i, |
68435eb2 | 9401 | vec_stmt, cost_vec)) |
68a0f2ff RS |
9402 | return false; |
9403 | } | |
9404 | } | |
9405 | else if (STMT_VINFO_LIVE_P (vinfo_for_stmt (stmt)) | |
68435eb2 RB |
9406 | && !vectorizable_live_operation (stmt, gsi, slp_node, -1, vec_stmt, |
9407 | cost_vec)) | |
68a0f2ff RS |
9408 | return false; |
9409 | ||
9410 | return true; | |
9411 | } | |
9412 | ||
8644a673 | 9413 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
9414 | |
9415 | bool | |
891ad31c | 9416 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node, |
68435eb2 | 9417 | slp_instance node_instance, stmt_vector_for_cost *cost_vec) |
ebfd146a | 9418 | { |
8644a673 | 9419 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
6585ff8f | 9420 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 9421 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 9422 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 9423 | bool ok; |
363477c0 | 9424 | gimple_seq pattern_def_seq; |
ebfd146a | 9425 | |
73fbfcad | 9426 | if (dump_enabled_p ()) |
ebfd146a | 9427 | { |
78c60e3d SS |
9428 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
9429 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 9430 | } |
ebfd146a | 9431 | |
1825a1f3 | 9432 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 9433 | { |
73fbfcad | 9434 | if (dump_enabled_p ()) |
78c60e3d | 9435 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9436 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
9437 | |
9438 | return false; | |
9439 | } | |
b8698a0f | 9440 | |
d54a098e RS |
9441 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
9442 | && node == NULL | |
9443 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) | |
9444 | { | |
9445 | gimple_stmt_iterator si; | |
9446 | ||
9447 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) | |
9448 | { | |
9449 | gimple *pattern_def_stmt = gsi_stmt (si); | |
6585ff8f RS |
9450 | stmt_vec_info pattern_def_stmt_info |
9451 | = vinfo->lookup_stmt (gsi_stmt (si)); | |
9452 | if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info) | |
9453 | || STMT_VINFO_LIVE_P (pattern_def_stmt_info)) | |
d54a098e RS |
9454 | { |
9455 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
9456 | if (dump_enabled_p ()) | |
9457 | { | |
9458 | dump_printf_loc (MSG_NOTE, vect_location, | |
9459 | "==> examining pattern def statement: "); | |
9460 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
9461 | } | |
9462 | ||
9463 | if (!vect_analyze_stmt (pattern_def_stmt, | |
9464 | need_to_vectorize, node, node_instance, | |
9465 | cost_vec)) | |
9466 | return false; | |
9467 | } | |
9468 | } | |
9469 | } | |
9470 | ||
b8698a0f | 9471 | /* Skip stmts that do not need to be vectorized. In loops this is expected |
8644a673 IR |
9472 | to include: |
9473 | - the COND_EXPR which is the loop exit condition | |
9474 | - any LABEL_EXPRs in the loop | |
b8698a0f | 9475 | - computations that are used only for array indexing or loop control. |
8644a673 | 9476 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 9477 | instance, therefore, all the statements are relevant. |
ebfd146a | 9478 | |
d092494c | 9479 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 9480 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
9481 | statements. In basic blocks we are called from some SLP instance |
9482 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
9483 | already will be part of SLP instance. */ | |
83197f37 | 9484 | |
10681ce8 | 9485 | stmt_vec_info pattern_stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); |
b8698a0f | 9486 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 9487 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 9488 | { |
9d5e7640 | 9489 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
10681ce8 RS |
9490 | && pattern_stmt_info |
9491 | && (STMT_VINFO_RELEVANT_P (pattern_stmt_info) | |
9492 | || STMT_VINFO_LIVE_P (pattern_stmt_info))) | |
9d5e7640 | 9493 | { |
83197f37 | 9494 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
10681ce8 RS |
9495 | stmt = pattern_stmt_info->stmt; |
9496 | stmt_info = pattern_stmt_info; | |
73fbfcad | 9497 | if (dump_enabled_p ()) |
9d5e7640 | 9498 | { |
78c60e3d SS |
9499 | dump_printf_loc (MSG_NOTE, vect_location, |
9500 | "==> examining pattern statement: "); | |
9501 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
9502 | } |
9503 | } | |
9504 | else | |
9505 | { | |
73fbfcad | 9506 | if (dump_enabled_p ()) |
e645e942 | 9507 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 9508 | |
9d5e7640 IR |
9509 | return true; |
9510 | } | |
8644a673 | 9511 | } |
83197f37 | 9512 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 9513 | && node == NULL |
10681ce8 RS |
9514 | && pattern_stmt_info |
9515 | && (STMT_VINFO_RELEVANT_P (pattern_stmt_info) | |
9516 | || STMT_VINFO_LIVE_P (pattern_stmt_info))) | |
83197f37 IR |
9517 | { |
9518 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 9519 | if (dump_enabled_p ()) |
83197f37 | 9520 | { |
78c60e3d SS |
9521 | dump_printf_loc (MSG_NOTE, vect_location, |
9522 | "==> examining pattern statement: "); | |
9523 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
9524 | } |
9525 | ||
10681ce8 | 9526 | if (!vect_analyze_stmt (pattern_stmt_info, need_to_vectorize, node, |
68435eb2 | 9527 | node_instance, cost_vec)) |
83197f37 IR |
9528 | return false; |
9529 | } | |
ebfd146a | 9530 | |
8644a673 IR |
9531 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
9532 | { | |
9533 | case vect_internal_def: | |
9534 | break; | |
ebfd146a | 9535 | |
8644a673 | 9536 | case vect_reduction_def: |
7c5222ff | 9537 | case vect_nested_cycle: |
14a61437 RB |
9538 | gcc_assert (!bb_vinfo |
9539 | && (relevance == vect_used_in_outer | |
9540 | || relevance == vect_used_in_outer_by_reduction | |
9541 | || relevance == vect_used_by_reduction | |
b28ead45 AH |
9542 | || relevance == vect_unused_in_scope |
9543 | || relevance == vect_used_only_live)); | |
8644a673 IR |
9544 | break; |
9545 | ||
9546 | case vect_induction_def: | |
e7baeb39 RB |
9547 | gcc_assert (!bb_vinfo); |
9548 | break; | |
9549 | ||
8644a673 IR |
9550 | case vect_constant_def: |
9551 | case vect_external_def: | |
9552 | case vect_unknown_def_type: | |
9553 | default: | |
9554 | gcc_unreachable (); | |
9555 | } | |
ebfd146a | 9556 | |
8644a673 | 9557 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 9558 | { |
8644a673 | 9559 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
9560 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
9561 | || (is_gimple_call (stmt) | |
9562 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 9563 | *need_to_vectorize = true; |
ebfd146a IR |
9564 | } |
9565 | ||
b1af7da6 RB |
9566 | if (PURE_SLP_STMT (stmt_info) && !node) |
9567 | { | |
9568 | dump_printf_loc (MSG_NOTE, vect_location, | |
9569 | "handled only by SLP analysis\n"); | |
9570 | return true; | |
9571 | } | |
9572 | ||
9573 | ok = true; | |
9574 | if (!bb_vinfo | |
9575 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
9576 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
68435eb2 RB |
9577 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node, cost_vec) |
9578 | || vectorizable_conversion (stmt, NULL, NULL, node, cost_vec) | |
9579 | || vectorizable_shift (stmt, NULL, NULL, node, cost_vec) | |
9580 | || vectorizable_operation (stmt, NULL, NULL, node, cost_vec) | |
9581 | || vectorizable_assignment (stmt, NULL, NULL, node, cost_vec) | |
9582 | || vectorizable_load (stmt, NULL, NULL, node, node_instance, cost_vec) | |
9583 | || vectorizable_call (stmt, NULL, NULL, node, cost_vec) | |
9584 | || vectorizable_store (stmt, NULL, NULL, node, cost_vec) | |
9585 | || vectorizable_reduction (stmt, NULL, NULL, node, node_instance, | |
9586 | cost_vec) | |
9587 | || vectorizable_induction (stmt, NULL, NULL, node, cost_vec) | |
9588 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node, cost_vec) | |
9589 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node, cost_vec)); | |
b1af7da6 RB |
9590 | else |
9591 | { | |
9592 | if (bb_vinfo) | |
68435eb2 RB |
9593 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node, cost_vec) |
9594 | || vectorizable_conversion (stmt, NULL, NULL, node, cost_vec) | |
9595 | || vectorizable_shift (stmt, NULL, NULL, node, cost_vec) | |
9596 | || vectorizable_operation (stmt, NULL, NULL, node, cost_vec) | |
9597 | || vectorizable_assignment (stmt, NULL, NULL, node, cost_vec) | |
9598 | || vectorizable_load (stmt, NULL, NULL, node, node_instance, | |
9599 | cost_vec) | |
9600 | || vectorizable_call (stmt, NULL, NULL, node, cost_vec) | |
9601 | || vectorizable_store (stmt, NULL, NULL, node, cost_vec) | |
9602 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node, | |
9603 | cost_vec) | |
9604 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node, | |
9605 | cost_vec)); | |
b1af7da6 | 9606 | } |
8644a673 IR |
9607 | |
9608 | if (!ok) | |
ebfd146a | 9609 | { |
73fbfcad | 9610 | if (dump_enabled_p ()) |
8644a673 | 9611 | { |
78c60e3d SS |
9612 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9613 | "not vectorized: relevant stmt not "); | |
9614 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
9615 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 9616 | } |
b8698a0f | 9617 | |
ebfd146a IR |
9618 | return false; |
9619 | } | |
9620 | ||
8644a673 IR |
9621 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
9622 | need extra handling, except for vectorizable reductions. */ | |
68435eb2 RB |
9623 | if (!bb_vinfo |
9624 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type | |
9625 | && !can_vectorize_live_stmts (stmt, NULL, node, NULL, cost_vec)) | |
ebfd146a | 9626 | { |
73fbfcad | 9627 | if (dump_enabled_p ()) |
8644a673 | 9628 | { |
78c60e3d | 9629 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
68a0f2ff | 9630 | "not vectorized: live stmt not supported: "); |
78c60e3d | 9631 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); |
8644a673 | 9632 | } |
b8698a0f | 9633 | |
8644a673 | 9634 | return false; |
ebfd146a IR |
9635 | } |
9636 | ||
ebfd146a IR |
9637 | return true; |
9638 | } | |
9639 | ||
9640 | ||
9641 | /* Function vect_transform_stmt. | |
9642 | ||
9643 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
9644 | ||
9645 | bool | |
355fe088 | 9646 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 9647 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
9648 | slp_instance slp_node_instance) |
9649 | { | |
6585ff8f RS |
9650 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
9651 | vec_info *vinfo = stmt_info->vinfo; | |
ebfd146a | 9652 | bool is_store = false; |
1eede195 | 9653 | stmt_vec_info vec_stmt = NULL; |
ebfd146a | 9654 | bool done; |
ebfd146a | 9655 | |
fce57248 | 9656 | gcc_assert (slp_node || !PURE_SLP_STMT (stmt_info)); |
1eede195 | 9657 | stmt_vec_info old_vec_stmt_info = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 9658 | |
e57d9a82 RB |
9659 | bool nested_p = (STMT_VINFO_LOOP_VINFO (stmt_info) |
9660 | && nested_in_vect_loop_p | |
9661 | (LOOP_VINFO_LOOP (STMT_VINFO_LOOP_VINFO (stmt_info)), | |
9662 | stmt)); | |
9663 | ||
ebfd146a IR |
9664 | switch (STMT_VINFO_TYPE (stmt_info)) |
9665 | { | |
9666 | case type_demotion_vec_info_type: | |
ebfd146a | 9667 | case type_promotion_vec_info_type: |
ebfd146a | 9668 | case type_conversion_vec_info_type: |
68435eb2 | 9669 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a IR |
9670 | gcc_assert (done); |
9671 | break; | |
9672 | ||
9673 | case induc_vec_info_type: | |
68435eb2 | 9674 | done = vectorizable_induction (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a IR |
9675 | gcc_assert (done); |
9676 | break; | |
9677 | ||
9dc3f7de | 9678 | case shift_vec_info_type: |
68435eb2 | 9679 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node, NULL); |
9dc3f7de IR |
9680 | gcc_assert (done); |
9681 | break; | |
9682 | ||
ebfd146a | 9683 | case op_vec_info_type: |
68435eb2 | 9684 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a IR |
9685 | gcc_assert (done); |
9686 | break; | |
9687 | ||
9688 | case assignment_vec_info_type: | |
68435eb2 | 9689 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a IR |
9690 | gcc_assert (done); |
9691 | break; | |
9692 | ||
9693 | case load_vec_info_type: | |
b8698a0f | 9694 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
68435eb2 | 9695 | slp_node_instance, NULL); |
ebfd146a IR |
9696 | gcc_assert (done); |
9697 | break; | |
9698 | ||
9699 | case store_vec_info_type: | |
68435eb2 | 9700 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a | 9701 | gcc_assert (done); |
0d0293ac | 9702 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
9703 | { |
9704 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 9705 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
9706 | one are skipped, and there vec_stmt_info shouldn't be freed |
9707 | meanwhile. */ | |
0d0293ac | 9708 | *grouped_store = true; |
bffb8014 | 9709 | stmt_vec_info group_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
2c53b149 | 9710 | if (DR_GROUP_STORE_COUNT (group_info) == DR_GROUP_SIZE (group_info)) |
ebfd146a | 9711 | is_store = true; |
f307441a | 9712 | } |
ebfd146a IR |
9713 | else |
9714 | is_store = true; | |
9715 | break; | |
9716 | ||
9717 | case condition_vec_info_type: | |
68435eb2 | 9718 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node, NULL); |
ebfd146a IR |
9719 | gcc_assert (done); |
9720 | break; | |
9721 | ||
42fd8198 | 9722 | case comparison_vec_info_type: |
68435eb2 | 9723 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node, NULL); |
42fd8198 IE |
9724 | gcc_assert (done); |
9725 | break; | |
9726 | ||
ebfd146a | 9727 | case call_vec_info_type: |
68435eb2 | 9728 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node, NULL); |
039d9ea1 | 9729 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
9730 | break; |
9731 | ||
0136f8f0 | 9732 | case call_simd_clone_vec_info_type: |
68435eb2 | 9733 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node, NULL); |
0136f8f0 AH |
9734 | stmt = gsi_stmt (*gsi); |
9735 | break; | |
9736 | ||
ebfd146a | 9737 | case reduc_vec_info_type: |
891ad31c | 9738 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node, |
68435eb2 | 9739 | slp_node_instance, NULL); |
ebfd146a IR |
9740 | gcc_assert (done); |
9741 | break; | |
9742 | ||
9743 | default: | |
9744 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
9745 | { | |
73fbfcad | 9746 | if (dump_enabled_p ()) |
78c60e3d | 9747 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9748 | "stmt not supported.\n"); |
ebfd146a IR |
9749 | gcc_unreachable (); |
9750 | } | |
9751 | } | |
9752 | ||
225ce44b RB |
9753 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
9754 | This would break hybrid SLP vectorization. */ | |
9755 | if (slp_node) | |
d90f8440 | 9756 | gcc_assert (!vec_stmt |
1eede195 | 9757 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt_info); |
225ce44b | 9758 | |
ebfd146a IR |
9759 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
9760 | is being vectorized, but outside the immediately enclosing loop. */ | |
9761 | if (vec_stmt | |
e57d9a82 | 9762 | && nested_p |
ebfd146a IR |
9763 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
9764 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 9765 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 9766 | vect_used_in_outer_by_reduction)) |
ebfd146a | 9767 | { |
a70d6342 IR |
9768 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
9769 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
9770 | imm_use_iterator imm_iter; |
9771 | use_operand_p use_p; | |
9772 | tree scalar_dest; | |
ebfd146a | 9773 | |
73fbfcad | 9774 | if (dump_enabled_p ()) |
78c60e3d | 9775 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 9776 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
9777 | |
9778 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
9779 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
9780 | STMT). */ | |
9781 | if (gimple_code (stmt) == GIMPLE_PHI) | |
9782 | scalar_dest = PHI_RESULT (stmt); | |
9783 | else | |
9784 | scalar_dest = gimple_assign_lhs (stmt); | |
9785 | ||
9786 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
6585ff8f RS |
9787 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) |
9788 | { | |
9789 | stmt_vec_info exit_phi_info | |
9790 | = vinfo->lookup_stmt (USE_STMT (use_p)); | |
9791 | STMT_VINFO_VEC_STMT (exit_phi_info) = vec_stmt; | |
9792 | } | |
ebfd146a IR |
9793 | } |
9794 | ||
9795 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
9796 | being vectorized. */ | |
68a0f2ff | 9797 | if (STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) |
ebfd146a | 9798 | { |
68435eb2 | 9799 | done = can_vectorize_live_stmts (stmt, gsi, slp_node, &vec_stmt, NULL); |
ebfd146a IR |
9800 | gcc_assert (done); |
9801 | } | |
9802 | ||
9803 | if (vec_stmt) | |
83197f37 | 9804 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 9805 | |
b8698a0f | 9806 | return is_store; |
ebfd146a IR |
9807 | } |
9808 | ||
9809 | ||
b8698a0f | 9810 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
9811 | stmt_vec_info. */ |
9812 | ||
9813 | void | |
355fe088 | 9814 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 9815 | { |
355fe088 | 9816 | gimple *next = first_stmt; |
ebfd146a IR |
9817 | gimple_stmt_iterator next_si; |
9818 | ||
9819 | while (next) | |
9820 | { | |
78048b1c JJ |
9821 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
9822 | ||
bffb8014 | 9823 | stmt_vec_info tmp = DR_GROUP_NEXT_ELEMENT (stmt_info); |
78048b1c JJ |
9824 | if (is_pattern_stmt_p (stmt_info)) |
9825 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
9826 | /* Free the attached stmt_vec_info and remove the stmt. */ |
9827 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 9828 | unlink_stmt_vdef (next); |
ebfd146a | 9829 | gsi_remove (&next_si, true); |
3d3f2249 | 9830 | release_defs (next); |
ebfd146a IR |
9831 | free_stmt_vec_info (next); |
9832 | next = tmp; | |
9833 | } | |
9834 | } | |
9835 | ||
9836 | ||
9837 | /* Function new_stmt_vec_info. | |
9838 | ||
9839 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
9840 | ||
9841 | stmt_vec_info | |
310213d4 | 9842 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
9843 | { |
9844 | stmt_vec_info res; | |
dbe1b846 | 9845 | res = (_stmt_vec_info *) xcalloc (1, sizeof (struct _stmt_vec_info)); |
ebfd146a IR |
9846 | |
9847 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
9848 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 9849 | res->vinfo = vinfo; |
8644a673 | 9850 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
9851 | STMT_VINFO_LIVE_P (res) = false; |
9852 | STMT_VINFO_VECTYPE (res) = NULL; | |
9853 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 9854 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a | 9855 | STMT_VINFO_IN_PATTERN_P (res) = false; |
363477c0 | 9856 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 9857 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 9858 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
7e16ce79 | 9859 | STMT_VINFO_VEC_CONST_COND_REDUC_CODE (res) = ERROR_MARK; |
ebfd146a | 9860 | |
ebfd146a IR |
9861 | if (gimple_code (stmt) == GIMPLE_PHI |
9862 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
9863 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
9864 | else | |
8644a673 IR |
9865 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
9866 | ||
9771b263 | 9867 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 9868 | STMT_SLP_TYPE (res) = loop_vect; |
78810bd3 RB |
9869 | STMT_VINFO_NUM_SLP_USES (res) = 0; |
9870 | ||
2c53b149 RB |
9871 | res->first_element = NULL; /* GROUP_FIRST_ELEMENT */ |
9872 | res->next_element = NULL; /* GROUP_NEXT_ELEMENT */ | |
9873 | res->size = 0; /* GROUP_SIZE */ | |
9874 | res->store_count = 0; /* GROUP_STORE_COUNT */ | |
9875 | res->gap = 0; /* GROUP_GAP */ | |
9876 | res->same_dr_stmt = NULL; /* GROUP_SAME_DR_STMT */ | |
ebfd146a | 9877 | |
ca823c85 RB |
9878 | /* This is really "uninitialized" until vect_compute_data_ref_alignment. */ |
9879 | res->dr_aux.misalignment = DR_MISALIGNMENT_UNINITIALIZED; | |
9880 | ||
ebfd146a IR |
9881 | return res; |
9882 | } | |
9883 | ||
9884 | ||
f8c0baaf | 9885 | /* Set the current stmt_vec_info vector to V. */ |
ebfd146a IR |
9886 | |
9887 | void | |
f8c0baaf | 9888 | set_stmt_vec_info_vec (vec<stmt_vec_info> *v) |
ebfd146a | 9889 | { |
f8c0baaf | 9890 | stmt_vec_info_vec = v; |
ebfd146a IR |
9891 | } |
9892 | ||
f8c0baaf | 9893 | /* Free the stmt_vec_info entries in V and release V. */ |
ebfd146a IR |
9894 | |
9895 | void | |
f8c0baaf | 9896 | free_stmt_vec_infos (vec<stmt_vec_info> *v) |
ebfd146a | 9897 | { |
93675444 | 9898 | unsigned int i; |
3161455c | 9899 | stmt_vec_info info; |
f8c0baaf | 9900 | FOR_EACH_VEC_ELT (*v, i, info) |
dbe1b846 | 9901 | if (info != NULL_STMT_VEC_INFO) |
3161455c | 9902 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
f8c0baaf RB |
9903 | if (v == stmt_vec_info_vec) |
9904 | stmt_vec_info_vec = NULL; | |
9905 | v->release (); | |
ebfd146a IR |
9906 | } |
9907 | ||
9908 | ||
9909 | /* Free stmt vectorization related info. */ | |
9910 | ||
9911 | void | |
355fe088 | 9912 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
9913 | { |
9914 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9915 | ||
9916 | if (!stmt_info) | |
9917 | return; | |
9918 | ||
78048b1c JJ |
9919 | /* Check if this statement has a related "pattern stmt" |
9920 | (introduced by the vectorizer during the pattern recognition | |
9921 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
9922 | too. */ | |
9923 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
9924 | { | |
e3947d80 RS |
9925 | if (gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info)) |
9926 | for (gimple_stmt_iterator si = gsi_start (seq); | |
9927 | !gsi_end_p (si); gsi_next (&si)) | |
9928 | { | |
9929 | gimple *seq_stmt = gsi_stmt (si); | |
9930 | gimple_set_bb (seq_stmt, NULL); | |
9931 | tree lhs = gimple_get_lhs (seq_stmt); | |
9932 | if (lhs && TREE_CODE (lhs) == SSA_NAME) | |
9933 | release_ssa_name (lhs); | |
9934 | free_stmt_vec_info (seq_stmt); | |
9935 | } | |
10681ce8 RS |
9936 | stmt_vec_info patt_stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); |
9937 | if (patt_stmt_info) | |
78048b1c | 9938 | { |
10681ce8 RS |
9939 | gimple_set_bb (patt_stmt_info->stmt, NULL); |
9940 | tree lhs = gimple_get_lhs (patt_stmt_info->stmt); | |
e6f5c25d | 9941 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde | 9942 | release_ssa_name (lhs); |
10681ce8 | 9943 | free_stmt_vec_info (patt_stmt_info); |
78048b1c JJ |
9944 | } |
9945 | } | |
9946 | ||
9771b263 | 9947 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 9948 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
9949 | set_vinfo_for_stmt (stmt, NULL); |
9950 | free (stmt_info); | |
9951 | } | |
9952 | ||
9953 | ||
bb67d9c7 | 9954 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 9955 | |
bb67d9c7 | 9956 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
9957 | by the target. */ |
9958 | ||
c803b2a9 | 9959 | tree |
86e36728 | 9960 | get_vectype_for_scalar_type_and_size (tree scalar_type, poly_uint64 size) |
ebfd146a | 9961 | { |
c7d97b28 | 9962 | tree orig_scalar_type = scalar_type; |
3bd8f481 | 9963 | scalar_mode inner_mode; |
ef4bddc2 | 9964 | machine_mode simd_mode; |
86e36728 | 9965 | poly_uint64 nunits; |
ebfd146a IR |
9966 | tree vectype; |
9967 | ||
3bd8f481 RS |
9968 | if (!is_int_mode (TYPE_MODE (scalar_type), &inner_mode) |
9969 | && !is_float_mode (TYPE_MODE (scalar_type), &inner_mode)) | |
ebfd146a IR |
9970 | return NULL_TREE; |
9971 | ||
3bd8f481 | 9972 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
48f2e373 | 9973 | |
7b7b1813 RG |
9974 | /* For vector types of elements whose mode precision doesn't |
9975 | match their types precision we use a element type of mode | |
9976 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
9977 | they support the proper result truncation/extension. |
9978 | We also make sure to build vector types with INTEGER_TYPE | |
9979 | component type only. */ | |
6d7971b8 | 9980 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
9981 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
9982 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
9983 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
9984 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 9985 | |
ccbf5bb4 RG |
9986 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
9987 | When the component mode passes the above test simply use a type | |
9988 | corresponding to that mode. The theory is that any use that | |
9989 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 9990 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 9991 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
9992 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
9993 | ||
9994 | /* We can't build a vector type of elements with alignment bigger than | |
9995 | their size. */ | |
dfc2e2ac | 9996 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
9997 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
9998 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 9999 | |
dfc2e2ac RB |
10000 | /* If we felt back to using the mode fail if there was |
10001 | no scalar type for it. */ | |
10002 | if (scalar_type == NULL_TREE) | |
10003 | return NULL_TREE; | |
10004 | ||
bb67d9c7 RG |
10005 | /* If no size was supplied use the mode the target prefers. Otherwise |
10006 | lookup a vector mode of the specified size. */ | |
86e36728 | 10007 | if (known_eq (size, 0U)) |
bb67d9c7 | 10008 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); |
86e36728 RS |
10009 | else if (!multiple_p (size, nbytes, &nunits) |
10010 | || !mode_for_vector (inner_mode, nunits).exists (&simd_mode)) | |
9da15d40 | 10011 | return NULL_TREE; |
4c8fd8ac | 10012 | /* NOTE: nunits == 1 is allowed to support single element vector types. */ |
86e36728 | 10013 | if (!multiple_p (GET_MODE_SIZE (simd_mode), nbytes, &nunits)) |
cc4b5170 | 10014 | return NULL_TREE; |
ebfd146a IR |
10015 | |
10016 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
10017 | |
10018 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
10019 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 10020 | return NULL_TREE; |
ebfd146a | 10021 | |
c7d97b28 RB |
10022 | /* Re-attach the address-space qualifier if we canonicalized the scalar |
10023 | type. */ | |
10024 | if (TYPE_ADDR_SPACE (orig_scalar_type) != TYPE_ADDR_SPACE (vectype)) | |
10025 | return build_qualified_type | |
10026 | (vectype, KEEP_QUAL_ADDR_SPACE (TYPE_QUALS (orig_scalar_type))); | |
10027 | ||
ebfd146a IR |
10028 | return vectype; |
10029 | } | |
10030 | ||
86e36728 | 10031 | poly_uint64 current_vector_size; |
bb67d9c7 RG |
10032 | |
10033 | /* Function get_vectype_for_scalar_type. | |
10034 | ||
10035 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
10036 | by the target. */ | |
10037 | ||
10038 | tree | |
10039 | get_vectype_for_scalar_type (tree scalar_type) | |
10040 | { | |
10041 | tree vectype; | |
10042 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
10043 | current_vector_size); | |
10044 | if (vectype | |
86e36728 | 10045 | && known_eq (current_vector_size, 0U)) |
bb67d9c7 RG |
10046 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); |
10047 | return vectype; | |
10048 | } | |
10049 | ||
42fd8198 IE |
10050 | /* Function get_mask_type_for_scalar_type. |
10051 | ||
10052 | Returns the mask type corresponding to a result of comparison | |
10053 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
10054 | ||
10055 | tree | |
10056 | get_mask_type_for_scalar_type (tree scalar_type) | |
10057 | { | |
10058 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
10059 | ||
10060 | if (!vectype) | |
10061 | return NULL; | |
10062 | ||
10063 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
10064 | current_vector_size); | |
10065 | } | |
10066 | ||
b690cc0f RG |
10067 | /* Function get_same_sized_vectype |
10068 | ||
10069 | Returns a vector type corresponding to SCALAR_TYPE of size | |
10070 | VECTOR_TYPE if supported by the target. */ | |
10071 | ||
10072 | tree | |
bb67d9c7 | 10073 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 10074 | { |
2568d8a1 | 10075 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type)) |
9f47c7e5 IE |
10076 | return build_same_sized_truth_vector_type (vector_type); |
10077 | ||
bb67d9c7 RG |
10078 | return get_vectype_for_scalar_type_and_size |
10079 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
10080 | } |
10081 | ||
ebfd146a IR |
10082 | /* Function vect_is_simple_use. |
10083 | ||
10084 | Input: | |
81c40241 RB |
10085 | VINFO - the vect info of the loop or basic block that is being vectorized. |
10086 | OPERAND - operand in the loop or bb. | |
10087 | Output: | |
fef96d8e RS |
10088 | DEF_STMT_INFO_OUT (optional) - information about the defining stmt in |
10089 | case OPERAND is an SSA_NAME that is defined in the vectorizable region | |
10090 | DEF_STMT_OUT (optional) - the defining stmt in case OPERAND is an SSA_NAME; | |
10091 | the definition could be anywhere in the function | |
81c40241 | 10092 | DT - the type of definition |
ebfd146a IR |
10093 | |
10094 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 10095 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 10096 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 10097 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
10098 | is the case in reduction/induction computations). |
10099 | For basic blocks, supportable operands are constants and bb invariants. | |
10100 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
10101 | |
10102 | bool | |
894dd753 | 10103 | vect_is_simple_use (tree operand, vec_info *vinfo, enum vect_def_type *dt, |
fef96d8e | 10104 | stmt_vec_info *def_stmt_info_out, gimple **def_stmt_out) |
b8698a0f | 10105 | { |
fef96d8e RS |
10106 | if (def_stmt_info_out) |
10107 | *def_stmt_info_out = NULL; | |
894dd753 RS |
10108 | if (def_stmt_out) |
10109 | *def_stmt_out = NULL; | |
3fc356dc | 10110 | *dt = vect_unknown_def_type; |
b8698a0f | 10111 | |
73fbfcad | 10112 | if (dump_enabled_p ()) |
ebfd146a | 10113 | { |
78c60e3d SS |
10114 | dump_printf_loc (MSG_NOTE, vect_location, |
10115 | "vect_is_simple_use: operand "); | |
30f502ed RB |
10116 | if (TREE_CODE (operand) == SSA_NAME |
10117 | && !SSA_NAME_IS_DEFAULT_DEF (operand)) | |
10118 | dump_gimple_expr (MSG_NOTE, TDF_SLIM, SSA_NAME_DEF_STMT (operand), 0); | |
10119 | else | |
10120 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
ebfd146a | 10121 | } |
b8698a0f | 10122 | |
b758f602 | 10123 | if (CONSTANT_CLASS_P (operand)) |
30f502ed RB |
10124 | *dt = vect_constant_def; |
10125 | else if (is_gimple_min_invariant (operand)) | |
10126 | *dt = vect_external_def; | |
10127 | else if (TREE_CODE (operand) != SSA_NAME) | |
10128 | *dt = vect_unknown_def_type; | |
10129 | else if (SSA_NAME_IS_DEFAULT_DEF (operand)) | |
8644a673 | 10130 | *dt = vect_external_def; |
ebfd146a IR |
10131 | else |
10132 | { | |
30f502ed | 10133 | gimple *def_stmt = SSA_NAME_DEF_STMT (operand); |
c98d0595 RS |
10134 | stmt_vec_info stmt_vinfo = vinfo->lookup_def (operand); |
10135 | if (!stmt_vinfo) | |
30f502ed RB |
10136 | *dt = vect_external_def; |
10137 | else | |
0f8c840c | 10138 | { |
30f502ed RB |
10139 | if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo)) |
10140 | { | |
10681ce8 RS |
10141 | stmt_vinfo = STMT_VINFO_RELATED_STMT (stmt_vinfo); |
10142 | def_stmt = stmt_vinfo->stmt; | |
30f502ed RB |
10143 | } |
10144 | switch (gimple_code (def_stmt)) | |
10145 | { | |
10146 | case GIMPLE_PHI: | |
10147 | case GIMPLE_ASSIGN: | |
10148 | case GIMPLE_CALL: | |
10149 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
10150 | break; | |
10151 | default: | |
10152 | *dt = vect_unknown_def_type; | |
10153 | break; | |
10154 | } | |
fef96d8e RS |
10155 | if (def_stmt_info_out) |
10156 | *def_stmt_info_out = stmt_vinfo; | |
0f8c840c | 10157 | } |
30f502ed RB |
10158 | if (def_stmt_out) |
10159 | *def_stmt_out = def_stmt; | |
ebfd146a IR |
10160 | } |
10161 | ||
2e8ab70c RB |
10162 | if (dump_enabled_p ()) |
10163 | { | |
30f502ed | 10164 | dump_printf (MSG_NOTE, ", type of def: "); |
2e8ab70c RB |
10165 | switch (*dt) |
10166 | { | |
10167 | case vect_uninitialized_def: | |
10168 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
10169 | break; | |
10170 | case vect_constant_def: | |
10171 | dump_printf (MSG_NOTE, "constant\n"); | |
10172 | break; | |
10173 | case vect_external_def: | |
10174 | dump_printf (MSG_NOTE, "external\n"); | |
10175 | break; | |
10176 | case vect_internal_def: | |
10177 | dump_printf (MSG_NOTE, "internal\n"); | |
10178 | break; | |
10179 | case vect_induction_def: | |
10180 | dump_printf (MSG_NOTE, "induction\n"); | |
10181 | break; | |
10182 | case vect_reduction_def: | |
10183 | dump_printf (MSG_NOTE, "reduction\n"); | |
10184 | break; | |
10185 | case vect_double_reduction_def: | |
10186 | dump_printf (MSG_NOTE, "double reduction\n"); | |
10187 | break; | |
10188 | case vect_nested_cycle: | |
10189 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
10190 | break; | |
10191 | case vect_unknown_def_type: | |
10192 | dump_printf (MSG_NOTE, "unknown\n"); | |
10193 | break; | |
10194 | } | |
10195 | } | |
10196 | ||
81c40241 | 10197 | if (*dt == vect_unknown_def_type) |
ebfd146a | 10198 | { |
73fbfcad | 10199 | if (dump_enabled_p ()) |
78c60e3d | 10200 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 10201 | "Unsupported pattern.\n"); |
ebfd146a IR |
10202 | return false; |
10203 | } | |
10204 | ||
ebfd146a IR |
10205 | return true; |
10206 | } | |
10207 | ||
81c40241 | 10208 | /* Function vect_is_simple_use. |
b690cc0f | 10209 | |
81c40241 | 10210 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
10211 | type of OPERAND and stores it to *VECTYPE. If the definition of |
10212 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
10213 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
10214 | is responsible to compute the best suited vector type for the | |
10215 | scalar operand. */ | |
10216 | ||
10217 | bool | |
894dd753 | 10218 | vect_is_simple_use (tree operand, vec_info *vinfo, enum vect_def_type *dt, |
fef96d8e RS |
10219 | tree *vectype, stmt_vec_info *def_stmt_info_out, |
10220 | gimple **def_stmt_out) | |
b690cc0f | 10221 | { |
fef96d8e | 10222 | stmt_vec_info def_stmt_info; |
894dd753 | 10223 | gimple *def_stmt; |
fef96d8e | 10224 | if (!vect_is_simple_use (operand, vinfo, dt, &def_stmt_info, &def_stmt)) |
b690cc0f RG |
10225 | return false; |
10226 | ||
894dd753 RS |
10227 | if (def_stmt_out) |
10228 | *def_stmt_out = def_stmt; | |
fef96d8e RS |
10229 | if (def_stmt_info_out) |
10230 | *def_stmt_info_out = def_stmt_info; | |
894dd753 | 10231 | |
b690cc0f RG |
10232 | /* Now get a vector type if the def is internal, otherwise supply |
10233 | NULL_TREE and leave it up to the caller to figure out a proper | |
10234 | type for the use stmt. */ | |
10235 | if (*dt == vect_internal_def | |
10236 | || *dt == vect_induction_def | |
10237 | || *dt == vect_reduction_def | |
10238 | || *dt == vect_double_reduction_def | |
10239 | || *dt == vect_nested_cycle) | |
10240 | { | |
fef96d8e | 10241 | *vectype = STMT_VINFO_VECTYPE (def_stmt_info); |
b690cc0f | 10242 | gcc_assert (*vectype != NULL_TREE); |
30f502ed RB |
10243 | if (dump_enabled_p ()) |
10244 | { | |
10245 | dump_printf_loc (MSG_NOTE, vect_location, | |
10246 | "vect_is_simple_use: vectype "); | |
10247 | dump_generic_expr (MSG_NOTE, TDF_SLIM, *vectype); | |
10248 | dump_printf (MSG_NOTE, "\n"); | |
10249 | } | |
b690cc0f RG |
10250 | } |
10251 | else if (*dt == vect_uninitialized_def | |
10252 | || *dt == vect_constant_def | |
10253 | || *dt == vect_external_def) | |
10254 | *vectype = NULL_TREE; | |
10255 | else | |
10256 | gcc_unreachable (); | |
10257 | ||
10258 | return true; | |
10259 | } | |
10260 | ||
ebfd146a IR |
10261 | |
10262 | /* Function supportable_widening_operation | |
10263 | ||
b8698a0f L |
10264 | Check whether an operation represented by the code CODE is a |
10265 | widening operation that is supported by the target platform in | |
b690cc0f RG |
10266 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
10267 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 10268 | |
1bda738b JJ |
10269 | Widening operations we currently support are NOP (CONVERT), FLOAT, |
10270 | FIX_TRUNC and WIDEN_MULT. This function checks if these operations | |
10271 | are supported by the target platform either directly (via vector | |
10272 | tree-codes), or via target builtins. | |
ebfd146a IR |
10273 | |
10274 | Output: | |
b8698a0f L |
10275 | - CODE1 and CODE2 are codes of vector operations to be used when |
10276 | vectorizing the operation, if available. | |
ebfd146a IR |
10277 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
10278 | case of multi-step conversion (like char->short->int - in that case | |
10279 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
10280 | - INTERM_TYPES contains the intermediate type required to perform the |
10281 | widening operation (short in the above example). */ | |
ebfd146a IR |
10282 | |
10283 | bool | |
355fe088 | 10284 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 10285 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
10286 | enum tree_code *code1, enum tree_code *code2, |
10287 | int *multi_step_cvt, | |
9771b263 | 10288 | vec<tree> *interm_types) |
ebfd146a IR |
10289 | { |
10290 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
10291 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 10292 | struct loop *vect_loop = NULL; |
ef4bddc2 | 10293 | machine_mode vec_mode; |
81f40b79 | 10294 | enum insn_code icode1, icode2; |
ebfd146a | 10295 | optab optab1, optab2; |
b690cc0f RG |
10296 | tree vectype = vectype_in; |
10297 | tree wide_vectype = vectype_out; | |
ebfd146a | 10298 | enum tree_code c1, c2; |
4a00c761 JJ |
10299 | int i; |
10300 | tree prev_type, intermediate_type; | |
ef4bddc2 | 10301 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 10302 | optab optab3, optab4; |
ebfd146a | 10303 | |
4a00c761 | 10304 | *multi_step_cvt = 0; |
4ef69dfc IR |
10305 | if (loop_info) |
10306 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
10307 | ||
ebfd146a IR |
10308 | switch (code) |
10309 | { | |
10310 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
10311 | /* The result of a vectorized widening operation usually requires |
10312 | two vectors (because the widened results do not fit into one vector). | |
10313 | The generated vector results would normally be expected to be | |
10314 | generated in the same order as in the original scalar computation, | |
10315 | i.e. if 8 results are generated in each vector iteration, they are | |
10316 | to be organized as follows: | |
10317 | vect1: [res1,res2,res3,res4], | |
10318 | vect2: [res5,res6,res7,res8]. | |
10319 | ||
10320 | However, in the special case that the result of the widening | |
10321 | operation is used in a reduction computation only, the order doesn't | |
10322 | matter (because when vectorizing a reduction we change the order of | |
10323 | the computation). Some targets can take advantage of this and | |
10324 | generate more efficient code. For example, targets like Altivec, | |
10325 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
10326 | generate the following vectors: | |
10327 | vect1: [res1,res3,res5,res7], | |
10328 | vect2: [res2,res4,res6,res8]. | |
10329 | ||
10330 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
10331 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
10332 | iterations in parallel). We therefore don't allow to change the | |
10333 | order of the computation in the inner-loop during outer-loop | |
10334 | vectorization. */ | |
10335 | /* TODO: Another case in which order doesn't *really* matter is when we | |
10336 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
10337 | Normally, pack_trunc performs an even/odd permute, whereas the | |
10338 | repack from an even/odd expansion would be an interleave, which | |
10339 | would be significantly simpler for e.g. AVX2. */ | |
10340 | /* In any case, in order to avoid duplicating the code below, recurse | |
10341 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
10342 | are properly set up for the caller. If we fail, we'll continue with | |
10343 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
10344 | if (vect_loop | |
10345 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
10346 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
10347 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
10348 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
10349 | code1, code2, multi_step_cvt, |
10350 | interm_types)) | |
ebc047a2 CH |
10351 | { |
10352 | /* Elements in a vector with vect_used_by_reduction property cannot | |
10353 | be reordered if the use chain with this property does not have the | |
10354 | same operation. One such an example is s += a * b, where elements | |
10355 | in a and b cannot be reordered. Here we check if the vector defined | |
10356 | by STMT is only directly used in the reduction statement. */ | |
0d0a4e20 RS |
10357 | tree lhs = gimple_assign_lhs (stmt); |
10358 | stmt_vec_info use_stmt_info = loop_info->lookup_single_use (lhs); | |
10359 | if (use_stmt_info | |
10360 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
10361 | return true; | |
ebc047a2 | 10362 | } |
4a00c761 JJ |
10363 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
10364 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
10365 | break; |
10366 | ||
81c40241 RB |
10367 | case DOT_PROD_EXPR: |
10368 | c1 = DOT_PROD_EXPR; | |
10369 | c2 = DOT_PROD_EXPR; | |
10370 | break; | |
10371 | ||
10372 | case SAD_EXPR: | |
10373 | c1 = SAD_EXPR; | |
10374 | c2 = SAD_EXPR; | |
10375 | break; | |
10376 | ||
6ae6116f RH |
10377 | case VEC_WIDEN_MULT_EVEN_EXPR: |
10378 | /* Support the recursion induced just above. */ | |
10379 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
10380 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
10381 | break; | |
10382 | ||
36ba4aae | 10383 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
10384 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
10385 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
10386 | break; |
10387 | ||
ebfd146a | 10388 | CASE_CONVERT: |
4a00c761 JJ |
10389 | c1 = VEC_UNPACK_LO_EXPR; |
10390 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
10391 | break; |
10392 | ||
10393 | case FLOAT_EXPR: | |
4a00c761 JJ |
10394 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
10395 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
10396 | break; |
10397 | ||
10398 | case FIX_TRUNC_EXPR: | |
1bda738b JJ |
10399 | c1 = VEC_UNPACK_FIX_TRUNC_LO_EXPR; |
10400 | c2 = VEC_UNPACK_FIX_TRUNC_HI_EXPR; | |
10401 | break; | |
ebfd146a IR |
10402 | |
10403 | default: | |
10404 | gcc_unreachable (); | |
10405 | } | |
10406 | ||
6ae6116f | 10407 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 10408 | std::swap (c1, c2); |
4a00c761 | 10409 | |
ebfd146a IR |
10410 | if (code == FIX_TRUNC_EXPR) |
10411 | { | |
10412 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
10413 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
10414 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
10415 | } |
10416 | else | |
10417 | { | |
10418 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
10419 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
10420 | } | |
10421 | ||
10422 | if (!optab1 || !optab2) | |
10423 | return false; | |
10424 | ||
10425 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
10426 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
10427 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
10428 | return false; |
10429 | ||
4a00c761 JJ |
10430 | *code1 = c1; |
10431 | *code2 = c2; | |
10432 | ||
10433 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
10434 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
10435 | /* For scalar masks we may have different boolean |
10436 | vector types having the same QImode. Thus we | |
10437 | add additional check for elements number. */ | |
10438 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
928686b1 RS |
10439 | || known_eq (TYPE_VECTOR_SUBPARTS (vectype), |
10440 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)); | |
4a00c761 | 10441 | |
b8698a0f | 10442 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 10443 | types. */ |
ebfd146a | 10444 | |
4a00c761 JJ |
10445 | prev_type = vectype; |
10446 | prev_mode = vec_mode; | |
b8698a0f | 10447 | |
4a00c761 JJ |
10448 | if (!CONVERT_EXPR_CODE_P (code)) |
10449 | return false; | |
b8698a0f | 10450 | |
4a00c761 JJ |
10451 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
10452 | intermediate steps in promotion sequence. We try | |
10453 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
10454 | not. */ | |
9771b263 | 10455 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
10456 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
10457 | { | |
10458 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
10459 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
10460 | { | |
7cfb4d93 | 10461 | intermediate_type = vect_halve_mask_nunits (prev_type); |
3ae0661a IE |
10462 | if (intermediate_mode != TYPE_MODE (intermediate_type)) |
10463 | return false; | |
10464 | } | |
10465 | else | |
10466 | intermediate_type | |
10467 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
10468 | TYPE_UNSIGNED (prev_type)); | |
10469 | ||
4a00c761 JJ |
10470 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); |
10471 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
10472 | ||
10473 | if (!optab3 || !optab4 | |
10474 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
10475 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
10476 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
10477 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
10478 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
10479 | == CODE_FOR_nothing) | |
10480 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
10481 | == CODE_FOR_nothing)) | |
10482 | break; | |
ebfd146a | 10483 | |
9771b263 | 10484 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
10485 | (*multi_step_cvt)++; |
10486 | ||
10487 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
10488 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff | 10489 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
928686b1 RS |
10490 | || known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type), |
10491 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)); | |
4a00c761 JJ |
10492 | |
10493 | prev_type = intermediate_type; | |
10494 | prev_mode = intermediate_mode; | |
ebfd146a IR |
10495 | } |
10496 | ||
9771b263 | 10497 | interm_types->release (); |
4a00c761 | 10498 | return false; |
ebfd146a IR |
10499 | } |
10500 | ||
10501 | ||
10502 | /* Function supportable_narrowing_operation | |
10503 | ||
b8698a0f L |
10504 | Check whether an operation represented by the code CODE is a |
10505 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
10506 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
10507 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 10508 | |
1bda738b JJ |
10509 | Narrowing operations we currently support are NOP (CONVERT), FIX_TRUNC |
10510 | and FLOAT. This function checks if these operations are supported by | |
ebfd146a IR |
10511 | the target platform directly via vector tree-codes. |
10512 | ||
10513 | Output: | |
b8698a0f L |
10514 | - CODE1 is the code of a vector operation to be used when |
10515 | vectorizing the operation, if available. | |
ebfd146a IR |
10516 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
10517 | case of multi-step conversion (like int->short->char - in that case | |
10518 | MULTI_STEP_CVT will be 1). | |
10519 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 10520 | narrowing operation (short in the above example). */ |
ebfd146a IR |
10521 | |
10522 | bool | |
10523 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 10524 | tree vectype_out, tree vectype_in, |
ebfd146a | 10525 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 10526 | vec<tree> *interm_types) |
ebfd146a | 10527 | { |
ef4bddc2 | 10528 | machine_mode vec_mode; |
ebfd146a IR |
10529 | enum insn_code icode1; |
10530 | optab optab1, interm_optab; | |
b690cc0f RG |
10531 | tree vectype = vectype_in; |
10532 | tree narrow_vectype = vectype_out; | |
ebfd146a | 10533 | enum tree_code c1; |
3ae0661a | 10534 | tree intermediate_type, prev_type; |
ef4bddc2 | 10535 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 10536 | int i; |
4a00c761 | 10537 | bool uns; |
ebfd146a | 10538 | |
4a00c761 | 10539 | *multi_step_cvt = 0; |
ebfd146a IR |
10540 | switch (code) |
10541 | { | |
10542 | CASE_CONVERT: | |
10543 | c1 = VEC_PACK_TRUNC_EXPR; | |
10544 | break; | |
10545 | ||
10546 | case FIX_TRUNC_EXPR: | |
10547 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
10548 | break; | |
10549 | ||
10550 | case FLOAT_EXPR: | |
1bda738b JJ |
10551 | c1 = VEC_PACK_FLOAT_EXPR; |
10552 | break; | |
ebfd146a IR |
10553 | |
10554 | default: | |
10555 | gcc_unreachable (); | |
10556 | } | |
10557 | ||
10558 | if (code == FIX_TRUNC_EXPR) | |
10559 | /* The signedness is determined from output operand. */ | |
b690cc0f | 10560 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
10561 | else |
10562 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
10563 | ||
10564 | if (!optab1) | |
10565 | return false; | |
10566 | ||
10567 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 10568 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
10569 | return false; |
10570 | ||
4a00c761 JJ |
10571 | *code1 = c1; |
10572 | ||
10573 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
10574 | /* For scalar masks we may have different boolean |
10575 | vector types having the same QImode. Thus we | |
10576 | add additional check for elements number. */ | |
10577 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
928686b1 RS |
10578 | || known_eq (TYPE_VECTOR_SUBPARTS (vectype) * 2, |
10579 | TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 | 10580 | |
1bda738b JJ |
10581 | if (code == FLOAT_EXPR) |
10582 | return false; | |
10583 | ||
ebfd146a IR |
10584 | /* Check if it's a multi-step conversion that can be done using intermediate |
10585 | types. */ | |
4a00c761 | 10586 | prev_mode = vec_mode; |
3ae0661a | 10587 | prev_type = vectype; |
4a00c761 JJ |
10588 | if (code == FIX_TRUNC_EXPR) |
10589 | uns = TYPE_UNSIGNED (vectype_out); | |
10590 | else | |
10591 | uns = TYPE_UNSIGNED (vectype); | |
10592 | ||
10593 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
10594 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
10595 | costly than signed. */ | |
10596 | if (code == FIX_TRUNC_EXPR && uns) | |
10597 | { | |
10598 | enum insn_code icode2; | |
10599 | ||
10600 | intermediate_type | |
10601 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
10602 | interm_optab | |
10603 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 10604 | if (interm_optab != unknown_optab |
4a00c761 JJ |
10605 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
10606 | && insn_data[icode1].operand[0].mode | |
10607 | == insn_data[icode2].operand[0].mode) | |
10608 | { | |
10609 | uns = false; | |
10610 | optab1 = interm_optab; | |
10611 | icode1 = icode2; | |
10612 | } | |
10613 | } | |
ebfd146a | 10614 | |
4a00c761 JJ |
10615 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
10616 | intermediate steps in promotion sequence. We try | |
10617 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 10618 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
10619 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
10620 | { | |
10621 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
10622 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
10623 | { | |
7cfb4d93 | 10624 | intermediate_type = vect_double_mask_nunits (prev_type); |
3ae0661a | 10625 | if (intermediate_mode != TYPE_MODE (intermediate_type)) |
7cfb4d93 | 10626 | return false; |
3ae0661a IE |
10627 | } |
10628 | else | |
10629 | intermediate_type | |
10630 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
4a00c761 JJ |
10631 | interm_optab |
10632 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
10633 | optab_default); | |
10634 | if (!interm_optab | |
10635 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
10636 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
10637 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
10638 | == CODE_FOR_nothing)) | |
10639 | break; | |
10640 | ||
9771b263 | 10641 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
10642 | (*multi_step_cvt)++; |
10643 | ||
10644 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff | 10645 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
928686b1 RS |
10646 | || known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type) * 2, |
10647 | TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 JJ |
10648 | |
10649 | prev_mode = intermediate_mode; | |
3ae0661a | 10650 | prev_type = intermediate_type; |
4a00c761 | 10651 | optab1 = interm_optab; |
ebfd146a IR |
10652 | } |
10653 | ||
9771b263 | 10654 | interm_types->release (); |
4a00c761 | 10655 | return false; |
ebfd146a | 10656 | } |
7cfb4d93 RS |
10657 | |
10658 | /* Generate and return a statement that sets vector mask MASK such that | |
10659 | MASK[I] is true iff J + START_INDEX < END_INDEX for all J <= I. */ | |
10660 | ||
10661 | gcall * | |
10662 | vect_gen_while (tree mask, tree start_index, tree end_index) | |
10663 | { | |
10664 | tree cmp_type = TREE_TYPE (start_index); | |
10665 | tree mask_type = TREE_TYPE (mask); | |
10666 | gcc_checking_assert (direct_internal_fn_supported_p (IFN_WHILE_ULT, | |
10667 | cmp_type, mask_type, | |
10668 | OPTIMIZE_FOR_SPEED)); | |
10669 | gcall *call = gimple_build_call_internal (IFN_WHILE_ULT, 3, | |
10670 | start_index, end_index, | |
10671 | build_zero_cst (mask_type)); | |
10672 | gimple_call_set_lhs (call, mask); | |
10673 | return call; | |
10674 | } | |
535e7c11 RS |
10675 | |
10676 | /* Generate a vector mask of type MASK_TYPE for which index I is false iff | |
10677 | J + START_INDEX < END_INDEX for all J <= I. Add the statements to SEQ. */ | |
10678 | ||
10679 | tree | |
10680 | vect_gen_while_not (gimple_seq *seq, tree mask_type, tree start_index, | |
10681 | tree end_index) | |
10682 | { | |
10683 | tree tmp = make_ssa_name (mask_type); | |
10684 | gcall *call = vect_gen_while (tmp, start_index, end_index); | |
10685 | gimple_seq_add_stmt (seq, call); | |
10686 | return gimple_build (seq, BIT_NOT_EXPR, mask_type, tmp); | |
10687 | } | |
1f3cb663 RS |
10688 | |
10689 | /* Try to compute the vector types required to vectorize STMT_INFO, | |
10690 | returning true on success and false if vectorization isn't possible. | |
10691 | ||
10692 | On success: | |
10693 | ||
10694 | - Set *STMT_VECTYPE_OUT to: | |
10695 | - NULL_TREE if the statement doesn't need to be vectorized; | |
10696 | - boolean_type_node if the statement is a boolean operation whose | |
10697 | vector type can only be determined once all the other vector types | |
10698 | are known; and | |
10699 | - the equivalent of STMT_VINFO_VECTYPE otherwise. | |
10700 | ||
10701 | - Set *NUNITS_VECTYPE_OUT to the vector type that contains the maximum | |
10702 | number of units needed to vectorize STMT_INFO, or NULL_TREE if the | |
10703 | statement does not help to determine the overall number of units. */ | |
10704 | ||
10705 | bool | |
10706 | vect_get_vector_types_for_stmt (stmt_vec_info stmt_info, | |
10707 | tree *stmt_vectype_out, | |
10708 | tree *nunits_vectype_out) | |
10709 | { | |
10710 | gimple *stmt = stmt_info->stmt; | |
10711 | ||
10712 | *stmt_vectype_out = NULL_TREE; | |
10713 | *nunits_vectype_out = NULL_TREE; | |
10714 | ||
10715 | if (gimple_get_lhs (stmt) == NULL_TREE | |
10716 | /* MASK_STORE has no lhs, but is ok. */ | |
10717 | && !gimple_call_internal_p (stmt, IFN_MASK_STORE)) | |
10718 | { | |
10719 | if (is_a <gcall *> (stmt)) | |
10720 | { | |
10721 | /* Ignore calls with no lhs. These must be calls to | |
10722 | #pragma omp simd functions, and what vectorization factor | |
10723 | it really needs can't be determined until | |
10724 | vectorizable_simd_clone_call. */ | |
10725 | if (dump_enabled_p ()) | |
10726 | dump_printf_loc (MSG_NOTE, vect_location, | |
10727 | "defer to SIMD clone analysis.\n"); | |
10728 | return true; | |
10729 | } | |
10730 | ||
10731 | if (dump_enabled_p ()) | |
10732 | { | |
10733 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10734 | "not vectorized: irregular stmt."); | |
10735 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
10736 | } | |
10737 | return false; | |
10738 | } | |
10739 | ||
10740 | if (VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))) | |
10741 | { | |
10742 | if (dump_enabled_p ()) | |
10743 | { | |
10744 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10745 | "not vectorized: vector stmt in loop:"); | |
10746 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
10747 | } | |
10748 | return false; | |
10749 | } | |
10750 | ||
10751 | tree vectype; | |
10752 | tree scalar_type = NULL_TREE; | |
10753 | if (STMT_VINFO_VECTYPE (stmt_info)) | |
10754 | *stmt_vectype_out = vectype = STMT_VINFO_VECTYPE (stmt_info); | |
10755 | else | |
10756 | { | |
10757 | gcc_assert (!STMT_VINFO_DATA_REF (stmt_info)); | |
10758 | if (gimple_call_internal_p (stmt, IFN_MASK_STORE)) | |
10759 | scalar_type = TREE_TYPE (gimple_call_arg (stmt, 3)); | |
10760 | else | |
10761 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); | |
10762 | ||
10763 | /* Pure bool ops don't participate in number-of-units computation. | |
10764 | For comparisons use the types being compared. */ | |
10765 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type) | |
10766 | && is_gimple_assign (stmt) | |
10767 | && gimple_assign_rhs_code (stmt) != COND_EXPR) | |
10768 | { | |
10769 | *stmt_vectype_out = boolean_type_node; | |
10770 | ||
10771 | tree rhs1 = gimple_assign_rhs1 (stmt); | |
10772 | if (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_comparison | |
10773 | && !VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (rhs1))) | |
10774 | scalar_type = TREE_TYPE (rhs1); | |
10775 | else | |
10776 | { | |
10777 | if (dump_enabled_p ()) | |
10778 | dump_printf_loc (MSG_NOTE, vect_location, | |
10779 | "pure bool operation.\n"); | |
10780 | return true; | |
10781 | } | |
10782 | } | |
10783 | ||
10784 | if (dump_enabled_p ()) | |
10785 | { | |
10786 | dump_printf_loc (MSG_NOTE, vect_location, | |
10787 | "get vectype for scalar type: "); | |
10788 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
10789 | dump_printf (MSG_NOTE, "\n"); | |
10790 | } | |
10791 | vectype = get_vectype_for_scalar_type (scalar_type); | |
10792 | if (!vectype) | |
10793 | { | |
10794 | if (dump_enabled_p ()) | |
10795 | { | |
10796 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10797 | "not vectorized: unsupported data-type "); | |
10798 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10799 | scalar_type); | |
10800 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10801 | } | |
10802 | return false; | |
10803 | } | |
10804 | ||
10805 | if (!*stmt_vectype_out) | |
10806 | *stmt_vectype_out = vectype; | |
10807 | ||
10808 | if (dump_enabled_p ()) | |
10809 | { | |
10810 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); | |
10811 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
10812 | dump_printf (MSG_NOTE, "\n"); | |
10813 | } | |
10814 | } | |
10815 | ||
10816 | /* Don't try to compute scalar types if the stmt produces a boolean | |
10817 | vector; use the existing vector type instead. */ | |
10818 | tree nunits_vectype; | |
10819 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) | |
10820 | nunits_vectype = vectype; | |
10821 | else | |
10822 | { | |
10823 | /* The number of units is set according to the smallest scalar | |
10824 | type (or the largest vector size, but we only support one | |
10825 | vector size per vectorization). */ | |
10826 | if (*stmt_vectype_out != boolean_type_node) | |
10827 | { | |
10828 | HOST_WIDE_INT dummy; | |
10829 | scalar_type = vect_get_smallest_scalar_type (stmt, &dummy, &dummy); | |
10830 | } | |
10831 | if (dump_enabled_p ()) | |
10832 | { | |
10833 | dump_printf_loc (MSG_NOTE, vect_location, | |
10834 | "get vectype for scalar type: "); | |
10835 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
10836 | dump_printf (MSG_NOTE, "\n"); | |
10837 | } | |
10838 | nunits_vectype = get_vectype_for_scalar_type (scalar_type); | |
10839 | } | |
10840 | if (!nunits_vectype) | |
10841 | { | |
10842 | if (dump_enabled_p ()) | |
10843 | { | |
10844 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10845 | "not vectorized: unsupported data-type "); | |
10846 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, scalar_type); | |
10847 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10848 | } | |
10849 | return false; | |
10850 | } | |
10851 | ||
10852 | if (maybe_ne (GET_MODE_SIZE (TYPE_MODE (vectype)), | |
10853 | GET_MODE_SIZE (TYPE_MODE (nunits_vectype)))) | |
10854 | { | |
10855 | if (dump_enabled_p ()) | |
10856 | { | |
10857 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10858 | "not vectorized: different sized vector " | |
10859 | "types in statement, "); | |
10860 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, vectype); | |
10861 | dump_printf (MSG_MISSED_OPTIMIZATION, " and "); | |
10862 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, nunits_vectype); | |
10863 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10864 | } | |
10865 | return false; | |
10866 | } | |
10867 | ||
10868 | if (dump_enabled_p ()) | |
10869 | { | |
10870 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); | |
10871 | dump_generic_expr (MSG_NOTE, TDF_SLIM, nunits_vectype); | |
10872 | dump_printf (MSG_NOTE, "\n"); | |
10873 | ||
10874 | dump_printf_loc (MSG_NOTE, vect_location, "nunits = "); | |
10875 | dump_dec (MSG_NOTE, TYPE_VECTOR_SUBPARTS (nunits_vectype)); | |
10876 | dump_printf (MSG_NOTE, "\n"); | |
10877 | } | |
10878 | ||
10879 | *nunits_vectype_out = nunits_vectype; | |
10880 | return true; | |
10881 | } | |
10882 | ||
10883 | /* Try to determine the correct vector type for STMT_INFO, which is a | |
10884 | statement that produces a scalar boolean result. Return the vector | |
10885 | type on success, otherwise return NULL_TREE. */ | |
10886 | ||
10887 | tree | |
10888 | vect_get_mask_type_for_stmt (stmt_vec_info stmt_info) | |
10889 | { | |
10890 | gimple *stmt = stmt_info->stmt; | |
10891 | tree mask_type = NULL; | |
10892 | tree vectype, scalar_type; | |
10893 | ||
10894 | if (is_gimple_assign (stmt) | |
10895 | && TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_comparison | |
10896 | && !VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (gimple_assign_rhs1 (stmt)))) | |
10897 | { | |
10898 | scalar_type = TREE_TYPE (gimple_assign_rhs1 (stmt)); | |
10899 | mask_type = get_mask_type_for_scalar_type (scalar_type); | |
10900 | ||
10901 | if (!mask_type) | |
10902 | { | |
10903 | if (dump_enabled_p ()) | |
10904 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10905 | "not vectorized: unsupported mask\n"); | |
10906 | return NULL_TREE; | |
10907 | } | |
10908 | } | |
10909 | else | |
10910 | { | |
10911 | tree rhs; | |
10912 | ssa_op_iter iter; | |
1f3cb663 RS |
10913 | enum vect_def_type dt; |
10914 | ||
10915 | FOR_EACH_SSA_TREE_OPERAND (rhs, stmt, iter, SSA_OP_USE) | |
10916 | { | |
894dd753 | 10917 | if (!vect_is_simple_use (rhs, stmt_info->vinfo, &dt, &vectype)) |
1f3cb663 RS |
10918 | { |
10919 | if (dump_enabled_p ()) | |
10920 | { | |
10921 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10922 | "not vectorized: can't compute mask type " | |
10923 | "for statement, "); | |
10924 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, | |
10925 | 0); | |
10926 | } | |
10927 | return NULL_TREE; | |
10928 | } | |
10929 | ||
10930 | /* No vectype probably means external definition. | |
10931 | Allow it in case there is another operand which | |
10932 | allows to determine mask type. */ | |
10933 | if (!vectype) | |
10934 | continue; | |
10935 | ||
10936 | if (!mask_type) | |
10937 | mask_type = vectype; | |
10938 | else if (maybe_ne (TYPE_VECTOR_SUBPARTS (mask_type), | |
10939 | TYPE_VECTOR_SUBPARTS (vectype))) | |
10940 | { | |
10941 | if (dump_enabled_p ()) | |
10942 | { | |
10943 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10944 | "not vectorized: different sized masks " | |
10945 | "types in statement, "); | |
10946 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10947 | mask_type); | |
10948 | dump_printf (MSG_MISSED_OPTIMIZATION, " and "); | |
10949 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10950 | vectype); | |
10951 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10952 | } | |
10953 | return NULL_TREE; | |
10954 | } | |
10955 | else if (VECTOR_BOOLEAN_TYPE_P (mask_type) | |
10956 | != VECTOR_BOOLEAN_TYPE_P (vectype)) | |
10957 | { | |
10958 | if (dump_enabled_p ()) | |
10959 | { | |
10960 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10961 | "not vectorized: mixed mask and " | |
10962 | "nonmask vector types in statement, "); | |
10963 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10964 | mask_type); | |
10965 | dump_printf (MSG_MISSED_OPTIMIZATION, " and "); | |
10966 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10967 | vectype); | |
10968 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10969 | } | |
10970 | return NULL_TREE; | |
10971 | } | |
10972 | } | |
10973 | ||
10974 | /* We may compare boolean value loaded as vector of integers. | |
10975 | Fix mask_type in such case. */ | |
10976 | if (mask_type | |
10977 | && !VECTOR_BOOLEAN_TYPE_P (mask_type) | |
10978 | && gimple_code (stmt) == GIMPLE_ASSIGN | |
10979 | && TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_comparison) | |
10980 | mask_type = build_same_sized_truth_vector_type (mask_type); | |
10981 | } | |
10982 | ||
10983 | /* No mask_type should mean loop invariant predicate. | |
10984 | This is probably a subject for optimization in if-conversion. */ | |
10985 | if (!mask_type && dump_enabled_p ()) | |
10986 | { | |
10987 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10988 | "not vectorized: can't compute mask type " | |
10989 | "for statement, "); | |
10990 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
10991 | } | |
10992 | return mask_type; | |
10993 | } |