]>
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 | |
82570274 | 120 | is part of the vectorization of STMT_INFO and the vector is associated |
272c6793 RS |
121 | with scalar destination SCALAR_DEST. */ |
122 | ||
123 | static tree | |
82570274 RS |
124 | read_vector_array (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
125 | tree scalar_dest, tree array, unsigned HOST_WIDE_INT n) | |
272c6793 RS |
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); | |
82570274 | 140 | vect_finish_stmt_generation (stmt_info, 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. | |
82570274 | 147 | The store is part of the vectorization of STMT_INFO. */ |
272c6793 RS |
148 | |
149 | static void | |
82570274 RS |
150 | write_vector_array (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
151 | tree vect, tree array, unsigned HOST_WIDE_INT n) | |
272c6793 RS |
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); | |
82570274 | 161 | vect_finish_stmt_generation (stmt_info, 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 | ||
82570274 | 179 | /* Add a clobber of variable VAR to the vectorization of STMT_INFO. |
3ba4ff41 RS |
180 | Emit the clobber before *GSI. */ |
181 | ||
182 | static void | |
82570274 RS |
183 | vect_clobber_variable (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
184 | tree var) | |
3ba4ff41 RS |
185 | { |
186 | tree clobber = build_clobber (TREE_TYPE (var)); | |
187 | gimple *new_stmt = gimple_build_assign (var, clobber); | |
82570274 | 188 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
3ba4ff41 RS |
189 | } |
190 | ||
ebfd146a IR |
191 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
192 | ||
193 | /* Function vect_mark_relevant. | |
194 | ||
32e8e429 | 195 | Mark STMT_INFO as "relevant for vectorization" and add it to WORKLIST. */ |
ebfd146a IR |
196 | |
197 | static void | |
32e8e429 | 198 | vect_mark_relevant (vec<stmt_vec_info> *worklist, stmt_vec_info stmt_info, |
97ecdb46 | 199 | enum vect_relevant relevant, bool live_p) |
ebfd146a | 200 | { |
ebfd146a IR |
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 ()) |
3c2a8ed0 DM |
205 | dump_printf_loc (MSG_NOTE, vect_location, |
206 | "mark relevant %d, live %d: %G", relevant, live_p, | |
207 | stmt_info->stmt); | |
ebfd146a | 208 | |
83197f37 IR |
209 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
210 | related pattern stmt instead of the original stmt. However, such stmts | |
211 | may have their own uses that are not in any pattern, in such cases the | |
212 | stmt itself should be marked. */ | |
ebfd146a IR |
213 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
214 | { | |
97ecdb46 JJ |
215 | /* This is the last stmt in a sequence that was detected as a |
216 | pattern that can potentially be vectorized. Don't mark the stmt | |
217 | as relevant/live because it's not going to be vectorized. | |
218 | Instead mark the pattern-stmt that replaces it. */ | |
83197f37 | 219 | |
97ecdb46 JJ |
220 | if (dump_enabled_p ()) |
221 | dump_printf_loc (MSG_NOTE, vect_location, | |
222 | "last stmt in pattern. don't mark" | |
223 | " relevant/live.\n"); | |
10681ce8 RS |
224 | stmt_vec_info old_stmt_info = stmt_info; |
225 | stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); | |
226 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == old_stmt_info); | |
97ecdb46 JJ |
227 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); |
228 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
ebfd146a IR |
229 | } |
230 | ||
231 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
232 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
233 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
234 | ||
235 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
236 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
237 | { | |
73fbfcad | 238 | if (dump_enabled_p ()) |
78c60e3d | 239 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 240 | "already marked relevant/live.\n"); |
ebfd146a IR |
241 | return; |
242 | } | |
243 | ||
86a91c0a | 244 | worklist->safe_push (stmt_info); |
ebfd146a IR |
245 | } |
246 | ||
247 | ||
b28ead45 AH |
248 | /* Function is_simple_and_all_uses_invariant |
249 | ||
32e8e429 | 250 | Return true if STMT_INFO is simple and all uses of it are invariant. */ |
b28ead45 AH |
251 | |
252 | bool | |
32e8e429 RS |
253 | is_simple_and_all_uses_invariant (stmt_vec_info stmt_info, |
254 | loop_vec_info loop_vinfo) | |
b28ead45 AH |
255 | { |
256 | tree op; | |
b28ead45 AH |
257 | ssa_op_iter iter; |
258 | ||
32e8e429 RS |
259 | gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt); |
260 | if (!stmt) | |
b28ead45 AH |
261 | return false; |
262 | ||
263 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
264 | { | |
265 | enum vect_def_type dt = vect_uninitialized_def; | |
266 | ||
894dd753 | 267 | if (!vect_is_simple_use (op, loop_vinfo, &dt)) |
b28ead45 AH |
268 | { |
269 | if (dump_enabled_p ()) | |
270 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
271 | "use not simple.\n"); | |
272 | return false; | |
273 | } | |
274 | ||
275 | if (dt != vect_external_def && dt != vect_constant_def) | |
276 | return false; | |
277 | } | |
278 | return true; | |
279 | } | |
280 | ||
ebfd146a IR |
281 | /* Function vect_stmt_relevant_p. |
282 | ||
82570274 RS |
283 | Return true if STMT_INFO, in the loop that is represented by LOOP_VINFO, |
284 | is "relevant for vectorization". | |
ebfd146a IR |
285 | |
286 | A stmt is considered "relevant for vectorization" if: | |
287 | - it has uses outside the loop. | |
288 | - it has vdefs (it alters memory). | |
289 | - control stmts in the loop (except for the exit condition). | |
290 | ||
291 | CHECKME: what other side effects would the vectorizer allow? */ | |
292 | ||
293 | static bool | |
82570274 | 294 | vect_stmt_relevant_p (stmt_vec_info stmt_info, loop_vec_info loop_vinfo, |
ebfd146a IR |
295 | enum vect_relevant *relevant, bool *live_p) |
296 | { | |
297 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
298 | ssa_op_iter op_iter; | |
299 | imm_use_iterator imm_iter; | |
300 | use_operand_p use_p; | |
301 | def_operand_p def_p; | |
302 | ||
8644a673 | 303 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
304 | *live_p = false; |
305 | ||
306 | /* cond stmt other than loop exit cond. */ | |
82570274 RS |
307 | if (is_ctrl_stmt (stmt_info->stmt) |
308 | && STMT_VINFO_TYPE (stmt_info) != loop_exit_ctrl_vec_info_type) | |
8644a673 | 309 | *relevant = vect_used_in_scope; |
ebfd146a IR |
310 | |
311 | /* changing memory. */ | |
82570274 RS |
312 | if (gimple_code (stmt_info->stmt) != GIMPLE_PHI) |
313 | if (gimple_vdef (stmt_info->stmt) | |
314 | && !gimple_clobber_p (stmt_info->stmt)) | |
ebfd146a | 315 | { |
73fbfcad | 316 | if (dump_enabled_p ()) |
78c60e3d | 317 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 318 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 319 | *relevant = vect_used_in_scope; |
ebfd146a IR |
320 | } |
321 | ||
322 | /* uses outside the loop. */ | |
82570274 | 323 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt_info->stmt, op_iter, SSA_OP_DEF) |
ebfd146a IR |
324 | { |
325 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
326 | { | |
327 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
328 | if (!flow_bb_inside_loop_p (loop, bb)) | |
329 | { | |
73fbfcad | 330 | if (dump_enabled_p ()) |
78c60e3d | 331 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 332 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 333 | |
3157b0c2 AO |
334 | if (is_gimple_debug (USE_STMT (use_p))) |
335 | continue; | |
336 | ||
ebfd146a IR |
337 | /* We expect all such uses to be in the loop exit phis |
338 | (because of loop closed form) */ | |
339 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
340 | gcc_assert (bb == single_exit (loop)->dest); | |
341 | ||
342 | *live_p = true; | |
343 | } | |
344 | } | |
345 | } | |
346 | ||
3a2edf4c | 347 | if (*live_p && *relevant == vect_unused_in_scope |
82570274 | 348 | && !is_simple_and_all_uses_invariant (stmt_info, loop_vinfo)) |
b28ead45 AH |
349 | { |
350 | if (dump_enabled_p ()) | |
351 | dump_printf_loc (MSG_NOTE, vect_location, | |
352 | "vec_stmt_relevant_p: stmt live but not relevant.\n"); | |
353 | *relevant = vect_used_only_live; | |
354 | } | |
355 | ||
ebfd146a IR |
356 | return (*live_p || *relevant); |
357 | } | |
358 | ||
359 | ||
b8698a0f | 360 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 361 | |
32e8e429 RS |
362 | USE is one of the uses attached to STMT_INFO. Check if USE is |
363 | used in STMT_INFO for anything other than indexing an array. */ | |
ebfd146a IR |
364 | |
365 | static bool | |
32e8e429 | 366 | exist_non_indexing_operands_for_use_p (tree use, stmt_vec_info stmt_info) |
ebfd146a IR |
367 | { |
368 | tree operand; | |
59a05b0c | 369 | |
ff802fa1 | 370 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
371 | reference in STMT, then any operand that corresponds to USE |
372 | is not indexing an array. */ | |
373 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
374 | return true; | |
59a05b0c | 375 | |
ebfd146a IR |
376 | /* STMT has a data_ref. FORNOW this means that its of one of |
377 | the following forms: | |
378 | -1- ARRAY_REF = var | |
379 | -2- var = ARRAY_REF | |
380 | (This should have been verified in analyze_data_refs). | |
381 | ||
382 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 383 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
384 | for array indexing. |
385 | ||
386 | Therefore, all we need to check is if STMT falls into the | |
387 | first case, and whether var corresponds to USE. */ | |
ebfd146a | 388 | |
86a91c0a | 389 | gassign *assign = dyn_cast <gassign *> (stmt_info->stmt); |
beb456c3 | 390 | if (!assign || !gimple_assign_copy_p (assign)) |
5ce9450f | 391 | { |
86a91c0a | 392 | gcall *call = dyn_cast <gcall *> (stmt_info->stmt); |
beb456c3 | 393 | if (call && gimple_call_internal_p (call)) |
bfaa08b7 | 394 | { |
beb456c3 | 395 | internal_fn ifn = gimple_call_internal_fn (call); |
bfaa08b7 RS |
396 | int mask_index = internal_fn_mask_index (ifn); |
397 | if (mask_index >= 0 | |
beb456c3 | 398 | && use == gimple_call_arg (call, mask_index)) |
bfaa08b7 | 399 | return true; |
f307441a RS |
400 | int stored_value_index = internal_fn_stored_value_index (ifn); |
401 | if (stored_value_index >= 0 | |
beb456c3 | 402 | && use == gimple_call_arg (call, stored_value_index)) |
f307441a | 403 | return true; |
bfaa08b7 | 404 | if (internal_gather_scatter_fn_p (ifn) |
beb456c3 | 405 | && use == gimple_call_arg (call, 1)) |
bfaa08b7 | 406 | return true; |
bfaa08b7 | 407 | } |
5ce9450f JJ |
408 | return false; |
409 | } | |
410 | ||
beb456c3 | 411 | if (TREE_CODE (gimple_assign_lhs (assign)) == SSA_NAME) |
59a05b0c | 412 | return false; |
beb456c3 | 413 | operand = gimple_assign_rhs1 (assign); |
ebfd146a IR |
414 | if (TREE_CODE (operand) != SSA_NAME) |
415 | return false; | |
416 | ||
417 | if (operand == use) | |
418 | return true; | |
419 | ||
420 | return false; | |
421 | } | |
422 | ||
423 | ||
b8698a0f | 424 | /* |
ebfd146a IR |
425 | Function process_use. |
426 | ||
427 | Inputs: | |
32e8e429 | 428 | - a USE in STMT_VINFO in a loop represented by LOOP_VINFO |
b28ead45 | 429 | - RELEVANT - enum value to be set in the STMT_VINFO of the stmt |
ff802fa1 | 430 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 431 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
432 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
433 | be performed. | |
ebfd146a IR |
434 | |
435 | Outputs: | |
436 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
437 | relevance info of the DEF_STMT of this USE: | |
32e8e429 RS |
438 | STMT_VINFO_LIVE_P (DEF_stmt_vinfo) <-- live_p |
439 | STMT_VINFO_RELEVANT (DEF_stmt_vinfo) <-- relevant | |
ebfd146a IR |
440 | Exceptions: |
441 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 442 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 443 | of the respective DEF_STMT is left unchanged. |
32e8e429 RS |
444 | - case 2: If STMT_VINFO is a reduction phi and DEF_STMT is a reduction stmt, |
445 | we skip DEF_STMT cause it had already been processed. | |
446 | - case 3: If DEF_STMT and STMT_VINFO are in different nests, then | |
447 | "relevant" will be modified accordingly. | |
ebfd146a IR |
448 | |
449 | Return true if everything is as expected. Return false otherwise. */ | |
450 | ||
f4ebbd24 | 451 | static opt_result |
32e8e429 | 452 | process_use (stmt_vec_info stmt_vinfo, tree use, loop_vec_info loop_vinfo, |
eca52fdd | 453 | enum vect_relevant relevant, vec<stmt_vec_info> *worklist, |
aec7ae7d | 454 | bool force) |
ebfd146a | 455 | { |
ebfd146a IR |
456 | stmt_vec_info dstmt_vinfo; |
457 | basic_block bb, def_bb; | |
ebfd146a IR |
458 | enum vect_def_type dt; |
459 | ||
b8698a0f | 460 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 461 | that are used for address computation are not considered relevant. */ |
86a91c0a | 462 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt_vinfo)) |
f4ebbd24 | 463 | return opt_result::success (); |
ebfd146a | 464 | |
fef96d8e | 465 | if (!vect_is_simple_use (use, loop_vinfo, &dt, &dstmt_vinfo)) |
f4ebbd24 DM |
466 | return opt_result::failure_at (stmt_vinfo->stmt, |
467 | "not vectorized:" | |
468 | " unsupported use in stmt.\n"); | |
ebfd146a | 469 | |
fef96d8e | 470 | if (!dstmt_vinfo) |
f4ebbd24 | 471 | return opt_result::success (); |
ebfd146a | 472 | |
fef96d8e | 473 | def_bb = gimple_bb (dstmt_vinfo->stmt); |
ebfd146a | 474 | |
fef96d8e RS |
475 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DSTMT_VINFO). |
476 | DSTMT_VINFO must have already been processed, because this should be the | |
b8698a0f | 477 | only way that STMT, which is a reduction-phi, was put in the worklist, |
fef96d8e | 478 | as there should be no other uses for DSTMT_VINFO in the loop. So we just |
ebfd146a | 479 | check that everything is as expected, and we are done. */ |
86a91c0a RS |
480 | bb = gimple_bb (stmt_vinfo->stmt); |
481 | if (gimple_code (stmt_vinfo->stmt) == GIMPLE_PHI | |
ebfd146a | 482 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
fef96d8e | 483 | && gimple_code (dstmt_vinfo->stmt) != GIMPLE_PHI |
ebfd146a IR |
484 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def |
485 | && bb->loop_father == def_bb->loop_father) | |
486 | { | |
73fbfcad | 487 | if (dump_enabled_p ()) |
78c60e3d | 488 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 489 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a | 490 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); |
b8698a0f | 491 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 492 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
f4ebbd24 | 493 | return opt_result::success (); |
ebfd146a IR |
494 | } |
495 | ||
496 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
497 | outer-loop-header-bb: | |
fef96d8e | 498 | d = dstmt_vinfo |
ebfd146a IR |
499 | inner-loop: |
500 | stmt # use (d) | |
501 | outer-loop-tail-bb: | |
502 | ... */ | |
503 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
504 | { | |
73fbfcad | 505 | if (dump_enabled_p ()) |
78c60e3d | 506 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 507 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 508 | |
ebfd146a IR |
509 | switch (relevant) |
510 | { | |
8644a673 | 511 | case vect_unused_in_scope: |
7c5222ff IR |
512 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
513 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 514 | break; |
7c5222ff | 515 | |
ebfd146a | 516 | case vect_used_in_outer_by_reduction: |
7c5222ff | 517 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
518 | relevant = vect_used_by_reduction; |
519 | break; | |
7c5222ff | 520 | |
ebfd146a | 521 | case vect_used_in_outer: |
7c5222ff | 522 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 523 | relevant = vect_used_in_scope; |
ebfd146a | 524 | break; |
7c5222ff | 525 | |
8644a673 | 526 | case vect_used_in_scope: |
ebfd146a IR |
527 | break; |
528 | ||
529 | default: | |
530 | gcc_unreachable (); | |
b8698a0f | 531 | } |
ebfd146a IR |
532 | } |
533 | ||
534 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
535 | outer-loop-header-bb: | |
536 | ... | |
537 | inner-loop: | |
fef96d8e | 538 | d = dstmt_vinfo |
06066f92 | 539 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
540 | stmt # use (d) */ |
541 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
542 | { | |
73fbfcad | 543 | if (dump_enabled_p ()) |
78c60e3d | 544 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 545 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 546 | |
ebfd146a IR |
547 | switch (relevant) |
548 | { | |
8644a673 | 549 | case vect_unused_in_scope: |
b8698a0f | 550 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 551 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 552 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
553 | break; |
554 | ||
ebfd146a | 555 | case vect_used_by_reduction: |
b28ead45 | 556 | case vect_used_only_live: |
ebfd146a IR |
557 | relevant = vect_used_in_outer_by_reduction; |
558 | break; | |
559 | ||
8644a673 | 560 | case vect_used_in_scope: |
ebfd146a IR |
561 | relevant = vect_used_in_outer; |
562 | break; | |
563 | ||
564 | default: | |
565 | gcc_unreachable (); | |
566 | } | |
567 | } | |
643a9684 RB |
568 | /* We are also not interested in uses on loop PHI backedges that are |
569 | inductions. Otherwise we'll needlessly vectorize the IV increment | |
e294f495 RB |
570 | and cause hybrid SLP for SLP inductions. Unless the PHI is live |
571 | of course. */ | |
86a91c0a | 572 | else if (gimple_code (stmt_vinfo->stmt) == GIMPLE_PHI |
643a9684 | 573 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_induction_def |
e294f495 | 574 | && ! STMT_VINFO_LIVE_P (stmt_vinfo) |
86a91c0a RS |
575 | && (PHI_ARG_DEF_FROM_EDGE (stmt_vinfo->stmt, |
576 | loop_latch_edge (bb->loop_father)) | |
643a9684 RB |
577 | == use)) |
578 | { | |
579 | if (dump_enabled_p ()) | |
580 | dump_printf_loc (MSG_NOTE, vect_location, | |
581 | "induction value on backedge.\n"); | |
f4ebbd24 | 582 | return opt_result::success (); |
643a9684 RB |
583 | } |
584 | ||
ebfd146a | 585 | |
fef96d8e | 586 | vect_mark_relevant (worklist, dstmt_vinfo, relevant, false); |
f4ebbd24 | 587 | return opt_result::success (); |
ebfd146a IR |
588 | } |
589 | ||
590 | ||
591 | /* Function vect_mark_stmts_to_be_vectorized. | |
592 | ||
593 | Not all stmts in the loop need to be vectorized. For example: | |
594 | ||
595 | for i... | |
596 | for j... | |
597 | 1. T0 = i + j | |
598 | 2. T1 = a[T0] | |
599 | ||
600 | 3. j = j + 1 | |
601 | ||
602 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
603 | addressing of vectorized data-refs are handled differently. | |
604 | ||
605 | This pass detects such stmts. */ | |
606 | ||
f4ebbd24 | 607 | opt_result |
ebfd146a IR |
608 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) |
609 | { | |
ebfd146a IR |
610 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
611 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
612 | unsigned int nbbs = loop->num_nodes; | |
613 | gimple_stmt_iterator si; | |
ebfd146a | 614 | unsigned int i; |
ebfd146a | 615 | basic_block bb; |
ebfd146a | 616 | bool live_p; |
b28ead45 | 617 | enum vect_relevant relevant; |
ebfd146a | 618 | |
adac3a68 | 619 | DUMP_VECT_SCOPE ("vect_mark_stmts_to_be_vectorized"); |
ebfd146a | 620 | |
eca52fdd | 621 | auto_vec<stmt_vec_info, 64> worklist; |
ebfd146a IR |
622 | |
623 | /* 1. Init worklist. */ | |
624 | for (i = 0; i < nbbs; i++) | |
625 | { | |
626 | bb = bbs[i]; | |
627 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 628 | { |
a1824cfd | 629 | stmt_vec_info phi_info = loop_vinfo->lookup_stmt (gsi_stmt (si)); |
73fbfcad | 630 | if (dump_enabled_p ()) |
3c2a8ed0 DM |
631 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? %G", |
632 | phi_info->stmt); | |
ebfd146a | 633 | |
a1824cfd RS |
634 | if (vect_stmt_relevant_p (phi_info, loop_vinfo, &relevant, &live_p)) |
635 | vect_mark_relevant (&worklist, phi_info, relevant, live_p); | |
ebfd146a IR |
636 | } |
637 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
638 | { | |
a1824cfd | 639 | stmt_vec_info stmt_info = loop_vinfo->lookup_stmt (gsi_stmt (si)); |
73fbfcad | 640 | if (dump_enabled_p ()) |
3c2a8ed0 DM |
641 | dump_printf_loc (MSG_NOTE, vect_location, |
642 | "init: stmt relevant? %G", stmt_info->stmt); | |
ebfd146a | 643 | |
a1824cfd RS |
644 | if (vect_stmt_relevant_p (stmt_info, loop_vinfo, &relevant, &live_p)) |
645 | vect_mark_relevant (&worklist, stmt_info, relevant, live_p); | |
ebfd146a IR |
646 | } |
647 | } | |
648 | ||
649 | /* 2. Process_worklist */ | |
9771b263 | 650 | while (worklist.length () > 0) |
ebfd146a IR |
651 | { |
652 | use_operand_p use_p; | |
653 | ssa_op_iter iter; | |
654 | ||
eca52fdd | 655 | stmt_vec_info stmt_vinfo = worklist.pop (); |
73fbfcad | 656 | if (dump_enabled_p ()) |
3c2a8ed0 DM |
657 | dump_printf_loc (MSG_NOTE, vect_location, |
658 | "worklist: examine stmt: %G", stmt_vinfo->stmt); | |
ebfd146a | 659 | |
b8698a0f | 660 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
b28ead45 AH |
661 | (DEF_STMT) as relevant/irrelevant according to the relevance property |
662 | of STMT. */ | |
ebfd146a | 663 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); |
ebfd146a | 664 | |
b28ead45 AH |
665 | /* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is |
666 | propagated as is to the DEF_STMTs of its USEs. | |
ebfd146a IR |
667 | |
668 | One exception is when STMT has been identified as defining a reduction | |
b28ead45 | 669 | variable; in this case we set the relevance to vect_used_by_reduction. |
ebfd146a | 670 | This is because we distinguish between two kinds of relevant stmts - |
b8698a0f | 671 | those that are used by a reduction computation, and those that are |
ff802fa1 | 672 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 673 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 674 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 675 | |
b28ead45 | 676 | switch (STMT_VINFO_DEF_TYPE (stmt_vinfo)) |
ebfd146a | 677 | { |
06066f92 | 678 | case vect_reduction_def: |
b28ead45 AH |
679 | gcc_assert (relevant != vect_unused_in_scope); |
680 | if (relevant != vect_unused_in_scope | |
681 | && relevant != vect_used_in_scope | |
682 | && relevant != vect_used_by_reduction | |
683 | && relevant != vect_used_only_live) | |
f4ebbd24 DM |
684 | return opt_result::failure_at |
685 | (stmt_vinfo->stmt, "unsupported use of reduction.\n"); | |
06066f92 | 686 | break; |
b8698a0f | 687 | |
06066f92 | 688 | case vect_nested_cycle: |
b28ead45 AH |
689 | if (relevant != vect_unused_in_scope |
690 | && relevant != vect_used_in_outer_by_reduction | |
691 | && relevant != vect_used_in_outer) | |
f4ebbd24 DM |
692 | return opt_result::failure_at |
693 | (stmt_vinfo->stmt, "unsupported use of nested cycle.\n"); | |
b8698a0f L |
694 | break; |
695 | ||
06066f92 | 696 | case vect_double_reduction_def: |
b28ead45 AH |
697 | if (relevant != vect_unused_in_scope |
698 | && relevant != vect_used_by_reduction | |
699 | && relevant != vect_used_only_live) | |
f4ebbd24 DM |
700 | return opt_result::failure_at |
701 | (stmt_vinfo->stmt, "unsupported use of double reduction.\n"); | |
b8698a0f | 702 | break; |
7c5222ff | 703 | |
06066f92 IR |
704 | default: |
705 | break; | |
7c5222ff | 706 | } |
b8698a0f | 707 | |
aec7ae7d | 708 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
709 | { |
710 | /* Pattern statements are not inserted into the code, so | |
711 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
712 | have to scan the RHS or function arguments instead. */ | |
86a91c0a | 713 | if (gassign *assign = dyn_cast <gassign *> (stmt_vinfo->stmt)) |
beb456c3 RS |
714 | { |
715 | enum tree_code rhs_code = gimple_assign_rhs_code (assign); | |
716 | tree op = gimple_assign_rhs1 (assign); | |
69d2aade JJ |
717 | |
718 | i = 1; | |
719 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
720 | { | |
f4ebbd24 DM |
721 | opt_result res |
722 | = process_use (stmt_vinfo, TREE_OPERAND (op, 0), | |
723 | loop_vinfo, relevant, &worklist, false); | |
724 | if (!res) | |
725 | return res; | |
726 | res = process_use (stmt_vinfo, TREE_OPERAND (op, 1), | |
727 | loop_vinfo, relevant, &worklist, false); | |
728 | if (!res) | |
729 | return res; | |
69d2aade JJ |
730 | i = 2; |
731 | } | |
beb456c3 RS |
732 | for (; i < gimple_num_ops (assign); i++) |
733 | { | |
734 | op = gimple_op (assign, i); | |
f4ebbd24 DM |
735 | if (TREE_CODE (op) == SSA_NAME) |
736 | { | |
737 | opt_result res | |
738 | = process_use (stmt_vinfo, op, loop_vinfo, relevant, | |
739 | &worklist, false); | |
740 | if (!res) | |
741 | return res; | |
742 | } | |
9d5e7640 IR |
743 | } |
744 | } | |
86a91c0a | 745 | else if (gcall *call = dyn_cast <gcall *> (stmt_vinfo->stmt)) |
beb456c3 RS |
746 | { |
747 | for (i = 0; i < gimple_call_num_args (call); i++) | |
748 | { | |
749 | tree arg = gimple_call_arg (call, i); | |
f4ebbd24 DM |
750 | opt_result res |
751 | = process_use (stmt_vinfo, arg, loop_vinfo, relevant, | |
752 | &worklist, false); | |
753 | if (!res) | |
754 | return res; | |
beb456c3 RS |
755 | } |
756 | } | |
9d5e7640 IR |
757 | } |
758 | else | |
86a91c0a | 759 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt_vinfo->stmt, iter, SSA_OP_USE) |
9d5e7640 IR |
760 | { |
761 | tree op = USE_FROM_PTR (use_p); | |
f4ebbd24 DM |
762 | opt_result res |
763 | = process_use (stmt_vinfo, op, loop_vinfo, relevant, | |
764 | &worklist, false); | |
765 | if (!res) | |
766 | return res; | |
9d5e7640 | 767 | } |
aec7ae7d | 768 | |
3bab6342 | 769 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d | 770 | { |
134c85ca | 771 | gather_scatter_info gs_info; |
86a91c0a | 772 | if (!vect_check_gather_scatter (stmt_vinfo, loop_vinfo, &gs_info)) |
134c85ca | 773 | gcc_unreachable (); |
f4ebbd24 DM |
774 | opt_result res |
775 | = process_use (stmt_vinfo, gs_info.offset, loop_vinfo, relevant, | |
776 | &worklist, true); | |
777 | if (!res) | |
778 | return res; | |
aec7ae7d | 779 | } |
ebfd146a IR |
780 | } /* while worklist */ |
781 | ||
f4ebbd24 | 782 | return opt_result::success (); |
ebfd146a IR |
783 | } |
784 | ||
68435eb2 RB |
785 | /* Compute the prologue cost for invariant or constant operands. */ |
786 | ||
787 | static unsigned | |
788 | vect_prologue_cost_for_slp_op (slp_tree node, stmt_vec_info stmt_info, | |
789 | unsigned opno, enum vect_def_type dt, | |
790 | stmt_vector_for_cost *cost_vec) | |
791 | { | |
b9787581 | 792 | gimple *stmt = SLP_TREE_SCALAR_STMTS (node)[0]->stmt; |
68435eb2 RB |
793 | tree op = gimple_op (stmt, opno); |
794 | unsigned prologue_cost = 0; | |
795 | ||
796 | /* Without looking at the actual initializer a vector of | |
797 | constants can be implemented as load from the constant pool. | |
798 | When all elements are the same we can use a splat. */ | |
799 | tree vectype = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
800 | unsigned group_size = SLP_TREE_SCALAR_STMTS (node).length (); | |
801 | unsigned num_vects_to_check; | |
802 | unsigned HOST_WIDE_INT const_nunits; | |
803 | unsigned nelt_limit; | |
804 | if (TYPE_VECTOR_SUBPARTS (vectype).is_constant (&const_nunits) | |
805 | && ! multiple_p (const_nunits, group_size)) | |
806 | { | |
807 | num_vects_to_check = SLP_TREE_NUMBER_OF_VEC_STMTS (node); | |
808 | nelt_limit = const_nunits; | |
809 | } | |
810 | else | |
811 | { | |
812 | /* If either the vector has variable length or the vectors | |
813 | are composed of repeated whole groups we only need to | |
814 | cost construction once. All vectors will be the same. */ | |
815 | num_vects_to_check = 1; | |
816 | nelt_limit = group_size; | |
817 | } | |
818 | tree elt = NULL_TREE; | |
819 | unsigned nelt = 0; | |
820 | for (unsigned j = 0; j < num_vects_to_check * nelt_limit; ++j) | |
821 | { | |
822 | unsigned si = j % group_size; | |
823 | if (nelt == 0) | |
b9787581 | 824 | elt = gimple_op (SLP_TREE_SCALAR_STMTS (node)[si]->stmt, opno); |
68435eb2 RB |
825 | /* ??? We're just tracking whether all operands of a single |
826 | vector initializer are the same, ideally we'd check if | |
827 | we emitted the same one already. */ | |
b9787581 | 828 | else if (elt != gimple_op (SLP_TREE_SCALAR_STMTS (node)[si]->stmt, |
68435eb2 RB |
829 | opno)) |
830 | elt = NULL_TREE; | |
831 | nelt++; | |
832 | if (nelt == nelt_limit) | |
833 | { | |
834 | /* ??? We need to pass down stmt_info for a vector type | |
835 | even if it points to the wrong stmt. */ | |
836 | prologue_cost += record_stmt_cost | |
837 | (cost_vec, 1, | |
838 | dt == vect_external_def | |
839 | ? (elt ? scalar_to_vec : vec_construct) | |
840 | : vector_load, | |
841 | stmt_info, 0, vect_prologue); | |
842 | nelt = 0; | |
843 | } | |
844 | } | |
845 | ||
846 | return prologue_cost; | |
847 | } | |
ebfd146a | 848 | |
b8698a0f | 849 | /* Function vect_model_simple_cost. |
ebfd146a | 850 | |
b8698a0f | 851 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
852 | single op. Right now, this does not account for multiple insns that could |
853 | be generated for the single vector op. We will handle that shortly. */ | |
854 | ||
68435eb2 | 855 | static void |
b8698a0f | 856 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 | 857 | enum vect_def_type *dt, |
4fc5ebf1 | 858 | int ndts, |
68435eb2 RB |
859 | slp_tree node, |
860 | stmt_vector_for_cost *cost_vec) | |
ebfd146a | 861 | { |
92345349 | 862 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a | 863 | |
68435eb2 | 864 | gcc_assert (cost_vec != NULL); |
ebfd146a | 865 | |
68435eb2 RB |
866 | /* ??? Somehow we need to fix this at the callers. */ |
867 | if (node) | |
868 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (node); | |
869 | ||
870 | if (node) | |
871 | { | |
872 | /* Scan operands and account for prologue cost of constants/externals. | |
873 | ??? This over-estimates cost for multiple uses and should be | |
874 | re-engineered. */ | |
b9787581 | 875 | gimple *stmt = SLP_TREE_SCALAR_STMTS (node)[0]->stmt; |
68435eb2 RB |
876 | tree lhs = gimple_get_lhs (stmt); |
877 | for (unsigned i = 0; i < gimple_num_ops (stmt); ++i) | |
878 | { | |
879 | tree op = gimple_op (stmt, i); | |
68435eb2 RB |
880 | enum vect_def_type dt; |
881 | if (!op || op == lhs) | |
882 | continue; | |
894dd753 | 883 | if (vect_is_simple_use (op, stmt_info->vinfo, &dt) |
68435eb2 RB |
884 | && (dt == vect_constant_def || dt == vect_external_def)) |
885 | prologue_cost += vect_prologue_cost_for_slp_op (node, stmt_info, | |
886 | i, dt, cost_vec); | |
887 | } | |
888 | } | |
889 | else | |
890 | /* Cost the "broadcast" of a scalar operand in to a vector operand. | |
891 | Use scalar_to_vec to cost the broadcast, as elsewhere in the vector | |
892 | cost model. */ | |
893 | for (int i = 0; i < ndts; i++) | |
894 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) | |
895 | prologue_cost += record_stmt_cost (cost_vec, 1, scalar_to_vec, | |
896 | stmt_info, 0, vect_prologue); | |
897 | ||
898 | /* Adjust for two-operator SLP nodes. */ | |
899 | if (node && SLP_TREE_TWO_OPERATORS (node)) | |
900 | { | |
901 | ncopies *= 2; | |
902 | inside_cost += record_stmt_cost (cost_vec, ncopies, vec_perm, | |
903 | stmt_info, 0, vect_body); | |
904 | } | |
c3e7ee41 BS |
905 | |
906 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
68435eb2 RB |
907 | inside_cost += record_stmt_cost (cost_vec, ncopies, vector_stmt, |
908 | stmt_info, 0, vect_body); | |
c3e7ee41 | 909 | |
73fbfcad | 910 | if (dump_enabled_p ()) |
78c60e3d SS |
911 | dump_printf_loc (MSG_NOTE, vect_location, |
912 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 913 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
914 | } |
915 | ||
916 | ||
8bd37302 BS |
917 | /* Model cost for type demotion and promotion operations. PWR is normally |
918 | zero for single-step promotions and demotions. It will be one if | |
919 | two-step promotion/demotion is required, and so on. Each additional | |
920 | step doubles the number of instructions required. */ | |
921 | ||
922 | static void | |
923 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
68435eb2 RB |
924 | enum vect_def_type *dt, int pwr, |
925 | stmt_vector_for_cost *cost_vec) | |
8bd37302 BS |
926 | { |
927 | int i, tmp; | |
92345349 | 928 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 | 929 | |
8bd37302 BS |
930 | for (i = 0; i < pwr + 1; i++) |
931 | { | |
932 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
933 | (i + 1) : i; | |
68435eb2 RB |
934 | inside_cost += record_stmt_cost (cost_vec, vect_pow2 (tmp), |
935 | vec_promote_demote, stmt_info, 0, | |
936 | vect_body); | |
8bd37302 BS |
937 | } |
938 | ||
939 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
940 | for (i = 0; i < 2; i++) | |
92345349 | 941 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
68435eb2 RB |
942 | prologue_cost += record_stmt_cost (cost_vec, 1, vector_stmt, |
943 | stmt_info, 0, vect_prologue); | |
8bd37302 | 944 | |
73fbfcad | 945 | if (dump_enabled_p ()) |
78c60e3d SS |
946 | dump_printf_loc (MSG_NOTE, vect_location, |
947 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 948 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
949 | } |
950 | ||
ebfd146a IR |
951 | /* Function vect_model_store_cost |
952 | ||
0d0293ac MM |
953 | Models cost for stores. In the case of grouped accesses, one access |
954 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a | 955 | |
68435eb2 | 956 | static void |
b8698a0f | 957 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
68435eb2 | 958 | enum vect_def_type dt, |
2de001ee | 959 | vect_memory_access_type memory_access_type, |
9ce4345a | 960 | vec_load_store_type vls_type, slp_tree slp_node, |
68435eb2 | 961 | stmt_vector_for_cost *cost_vec) |
ebfd146a | 962 | { |
92345349 | 963 | unsigned int inside_cost = 0, prologue_cost = 0; |
bffb8014 | 964 | stmt_vec_info first_stmt_info = stmt_info; |
892a981f | 965 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
ebfd146a | 966 | |
68435eb2 RB |
967 | /* ??? Somehow we need to fix this at the callers. */ |
968 | if (slp_node) | |
969 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
970 | ||
9ce4345a | 971 | if (vls_type == VLS_STORE_INVARIANT) |
68435eb2 RB |
972 | { |
973 | if (slp_node) | |
974 | prologue_cost += vect_prologue_cost_for_slp_op (slp_node, stmt_info, | |
975 | 1, dt, cost_vec); | |
976 | else | |
977 | prologue_cost += record_stmt_cost (cost_vec, 1, scalar_to_vec, | |
978 | stmt_info, 0, vect_prologue); | |
979 | } | |
ebfd146a | 980 | |
892a981f RS |
981 | /* Grouped stores update all elements in the group at once, |
982 | so we want the DR for the first statement. */ | |
983 | if (!slp_node && grouped_access_p) | |
bffb8014 | 984 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a | 985 | |
892a981f RS |
986 | /* True if we should include any once-per-group costs as well as |
987 | the cost of the statement itself. For SLP we only get called | |
988 | once per group anyhow. */ | |
bffb8014 | 989 | bool first_stmt_p = (first_stmt_info == stmt_info); |
892a981f | 990 | |
272c6793 | 991 | /* We assume that the cost of a single store-lanes instruction is |
2c53b149 | 992 | equivalent to the cost of DR_GROUP_SIZE separate stores. If a grouped |
272c6793 | 993 | access is instead being provided by a permute-and-store operation, |
2de001ee RS |
994 | include the cost of the permutes. */ |
995 | if (first_stmt_p | |
996 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 997 | { |
e1377713 ES |
998 | /* Uses a high and low interleave or shuffle operations for each |
999 | needed permute. */ | |
bffb8014 | 1000 | int group_size = DR_GROUP_SIZE (first_stmt_info); |
e1377713 | 1001 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
68435eb2 | 1002 | inside_cost = record_stmt_cost (cost_vec, nstmts, vec_perm, |
92345349 | 1003 | stmt_info, 0, vect_body); |
ebfd146a | 1004 | |
73fbfcad | 1005 | if (dump_enabled_p ()) |
78c60e3d | 1006 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1007 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 1008 | group_size); |
ebfd146a IR |
1009 | } |
1010 | ||
cee62fee | 1011 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 1012 | /* Costs of the stores. */ |
067bc855 RB |
1013 | if (memory_access_type == VMAT_ELEMENTWISE |
1014 | || memory_access_type == VMAT_GATHER_SCATTER) | |
c5126ce8 RS |
1015 | { |
1016 | /* N scalar stores plus extracting the elements. */ | |
1017 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
68435eb2 | 1018 | inside_cost += record_stmt_cost (cost_vec, |
c5126ce8 RS |
1019 | ncopies * assumed_nunits, |
1020 | scalar_store, stmt_info, 0, vect_body); | |
1021 | } | |
f2e2a985 | 1022 | else |
57c454d2 | 1023 | vect_get_store_cost (stmt_info, ncopies, &inside_cost, cost_vec); |
ebfd146a | 1024 | |
2de001ee RS |
1025 | if (memory_access_type == VMAT_ELEMENTWISE |
1026 | || memory_access_type == VMAT_STRIDED_SLP) | |
c5126ce8 RS |
1027 | { |
1028 | /* N scalar stores plus extracting the elements. */ | |
1029 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
68435eb2 | 1030 | inside_cost += record_stmt_cost (cost_vec, |
c5126ce8 RS |
1031 | ncopies * assumed_nunits, |
1032 | vec_to_scalar, stmt_info, 0, vect_body); | |
1033 | } | |
cee62fee | 1034 | |
73fbfcad | 1035 | if (dump_enabled_p ()) |
78c60e3d SS |
1036 | dump_printf_loc (MSG_NOTE, vect_location, |
1037 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 1038 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
1039 | } |
1040 | ||
1041 | ||
720f5239 IR |
1042 | /* Calculate cost of DR's memory access. */ |
1043 | void | |
57c454d2 | 1044 | vect_get_store_cost (stmt_vec_info stmt_info, int ncopies, |
c3e7ee41 | 1045 | unsigned int *inside_cost, |
92345349 | 1046 | stmt_vector_for_cost *body_cost_vec) |
720f5239 | 1047 | { |
89fa689a RS |
1048 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info); |
1049 | int alignment_support_scheme | |
1050 | = vect_supportable_dr_alignment (dr_info, false); | |
720f5239 IR |
1051 | |
1052 | switch (alignment_support_scheme) | |
1053 | { | |
1054 | case dr_aligned: | |
1055 | { | |
92345349 BS |
1056 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
1057 | vector_store, stmt_info, 0, | |
1058 | vect_body); | |
720f5239 | 1059 | |
73fbfcad | 1060 | if (dump_enabled_p ()) |
78c60e3d | 1061 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1062 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
1063 | break; |
1064 | } | |
1065 | ||
1066 | case dr_unaligned_supported: | |
1067 | { | |
720f5239 | 1068 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1069 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1070 | unaligned_store, stmt_info, |
89fa689a RS |
1071 | DR_MISALIGNMENT (dr_info), |
1072 | vect_body); | |
73fbfcad | 1073 | if (dump_enabled_p ()) |
78c60e3d SS |
1074 | dump_printf_loc (MSG_NOTE, vect_location, |
1075 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1076 | "hardware.\n"); |
720f5239 IR |
1077 | break; |
1078 | } | |
1079 | ||
38eec4c6 UW |
1080 | case dr_unaligned_unsupported: |
1081 | { | |
1082 | *inside_cost = VECT_MAX_COST; | |
1083 | ||
73fbfcad | 1084 | if (dump_enabled_p ()) |
78c60e3d | 1085 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1086 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1087 | break; |
1088 | } | |
1089 | ||
720f5239 IR |
1090 | default: |
1091 | gcc_unreachable (); | |
1092 | } | |
1093 | } | |
1094 | ||
1095 | ||
ebfd146a IR |
1096 | /* Function vect_model_load_cost |
1097 | ||
892a981f RS |
1098 | Models cost for loads. In the case of grouped accesses, one access has |
1099 | the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1100 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1101 | access scheme chosen. */ |
1102 | ||
68435eb2 RB |
1103 | static void |
1104 | vect_model_load_cost (stmt_vec_info stmt_info, unsigned ncopies, | |
2de001ee | 1105 | vect_memory_access_type memory_access_type, |
68435eb2 | 1106 | slp_instance instance, |
2de001ee | 1107 | slp_tree slp_node, |
68435eb2 | 1108 | stmt_vector_for_cost *cost_vec) |
ebfd146a | 1109 | { |
92345349 | 1110 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f | 1111 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
ebfd146a | 1112 | |
68435eb2 RB |
1113 | gcc_assert (cost_vec); |
1114 | ||
1115 | /* ??? Somehow we need to fix this at the callers. */ | |
1116 | if (slp_node) | |
1117 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
1118 | ||
1119 | if (slp_node && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
1120 | { | |
1121 | /* If the load is permuted then the alignment is determined by | |
1122 | the first group element not by the first scalar stmt DR. */ | |
bffb8014 | 1123 | stmt_vec_info first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
68435eb2 RB |
1124 | /* Record the cost for the permutation. */ |
1125 | unsigned n_perms; | |
1126 | unsigned assumed_nunits | |
bffb8014 | 1127 | = vect_nunits_for_cost (STMT_VINFO_VECTYPE (first_stmt_info)); |
68435eb2 RB |
1128 | unsigned slp_vf = (ncopies * assumed_nunits) / instance->group_size; |
1129 | vect_transform_slp_perm_load (slp_node, vNULL, NULL, | |
1130 | slp_vf, instance, true, | |
1131 | &n_perms); | |
1132 | inside_cost += record_stmt_cost (cost_vec, n_perms, vec_perm, | |
bffb8014 | 1133 | first_stmt_info, 0, vect_body); |
68435eb2 RB |
1134 | /* And adjust the number of loads performed. This handles |
1135 | redundancies as well as loads that are later dead. */ | |
bffb8014 | 1136 | auto_sbitmap perm (DR_GROUP_SIZE (first_stmt_info)); |
68435eb2 RB |
1137 | bitmap_clear (perm); |
1138 | for (unsigned i = 0; | |
1139 | i < SLP_TREE_LOAD_PERMUTATION (slp_node).length (); ++i) | |
1140 | bitmap_set_bit (perm, SLP_TREE_LOAD_PERMUTATION (slp_node)[i]); | |
1141 | ncopies = 0; | |
1142 | bool load_seen = false; | |
bffb8014 | 1143 | for (unsigned i = 0; i < DR_GROUP_SIZE (first_stmt_info); ++i) |
68435eb2 RB |
1144 | { |
1145 | if (i % assumed_nunits == 0) | |
1146 | { | |
1147 | if (load_seen) | |
1148 | ncopies++; | |
1149 | load_seen = false; | |
1150 | } | |
1151 | if (bitmap_bit_p (perm, i)) | |
1152 | load_seen = true; | |
1153 | } | |
1154 | if (load_seen) | |
1155 | ncopies++; | |
1156 | gcc_assert (ncopies | |
bffb8014 RS |
1157 | <= (DR_GROUP_SIZE (first_stmt_info) |
1158 | - DR_GROUP_GAP (first_stmt_info) | |
68435eb2 RB |
1159 | + assumed_nunits - 1) / assumed_nunits); |
1160 | } | |
1161 | ||
892a981f RS |
1162 | /* Grouped loads read all elements in the group at once, |
1163 | so we want the DR for the first statement. */ | |
bffb8014 | 1164 | stmt_vec_info first_stmt_info = stmt_info; |
892a981f | 1165 | if (!slp_node && grouped_access_p) |
bffb8014 | 1166 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a | 1167 | |
892a981f RS |
1168 | /* True if we should include any once-per-group costs as well as |
1169 | the cost of the statement itself. For SLP we only get called | |
1170 | once per group anyhow. */ | |
bffb8014 | 1171 | bool first_stmt_p = (first_stmt_info == stmt_info); |
892a981f | 1172 | |
272c6793 | 1173 | /* We assume that the cost of a single load-lanes instruction is |
2c53b149 | 1174 | equivalent to the cost of DR_GROUP_SIZE separate loads. If a grouped |
272c6793 | 1175 | access is instead being provided by a load-and-permute operation, |
2de001ee RS |
1176 | include the cost of the permutes. */ |
1177 | if (first_stmt_p | |
1178 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 1179 | { |
2c23db6d ES |
1180 | /* Uses an even and odd extract operations or shuffle operations |
1181 | for each needed permute. */ | |
bffb8014 | 1182 | int group_size = DR_GROUP_SIZE (first_stmt_info); |
2c23db6d | 1183 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
68435eb2 RB |
1184 | inside_cost += record_stmt_cost (cost_vec, nstmts, vec_perm, |
1185 | stmt_info, 0, vect_body); | |
ebfd146a | 1186 | |
73fbfcad | 1187 | if (dump_enabled_p ()) |
e645e942 TJ |
1188 | dump_printf_loc (MSG_NOTE, vect_location, |
1189 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1190 | group_size); |
ebfd146a IR |
1191 | } |
1192 | ||
1193 | /* The loads themselves. */ | |
067bc855 RB |
1194 | if (memory_access_type == VMAT_ELEMENTWISE |
1195 | || memory_access_type == VMAT_GATHER_SCATTER) | |
a82960aa | 1196 | { |
a21892ad BS |
1197 | /* N scalar loads plus gathering them into a vector. */ |
1198 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
c5126ce8 | 1199 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); |
68435eb2 | 1200 | inside_cost += record_stmt_cost (cost_vec, |
c5126ce8 | 1201 | ncopies * assumed_nunits, |
92345349 | 1202 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1203 | } |
1204 | else | |
57c454d2 | 1205 | vect_get_load_cost (stmt_info, ncopies, first_stmt_p, |
92345349 | 1206 | &inside_cost, &prologue_cost, |
68435eb2 | 1207 | cost_vec, cost_vec, true); |
2de001ee RS |
1208 | if (memory_access_type == VMAT_ELEMENTWISE |
1209 | || memory_access_type == VMAT_STRIDED_SLP) | |
68435eb2 | 1210 | inside_cost += record_stmt_cost (cost_vec, ncopies, vec_construct, |
892a981f | 1211 | stmt_info, 0, vect_body); |
720f5239 | 1212 | |
73fbfcad | 1213 | if (dump_enabled_p ()) |
78c60e3d SS |
1214 | dump_printf_loc (MSG_NOTE, vect_location, |
1215 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1216 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1217 | } |
1218 | ||
1219 | ||
1220 | /* Calculate cost of DR's memory access. */ | |
1221 | void | |
57c454d2 | 1222 | vect_get_load_cost (stmt_vec_info stmt_info, int ncopies, |
c3e7ee41 | 1223 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1224 | unsigned int *prologue_cost, |
1225 | stmt_vector_for_cost *prologue_cost_vec, | |
1226 | stmt_vector_for_cost *body_cost_vec, | |
1227 | bool record_prologue_costs) | |
720f5239 | 1228 | { |
89fa689a RS |
1229 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info); |
1230 | int alignment_support_scheme | |
1231 | = vect_supportable_dr_alignment (dr_info, false); | |
720f5239 IR |
1232 | |
1233 | switch (alignment_support_scheme) | |
ebfd146a IR |
1234 | { |
1235 | case dr_aligned: | |
1236 | { | |
92345349 BS |
1237 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1238 | stmt_info, 0, vect_body); | |
ebfd146a | 1239 | |
73fbfcad | 1240 | if (dump_enabled_p ()) |
78c60e3d | 1241 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1242 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1243 | |
1244 | break; | |
1245 | } | |
1246 | case dr_unaligned_supported: | |
1247 | { | |
720f5239 | 1248 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1249 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1250 | unaligned_load, stmt_info, |
89fa689a RS |
1251 | DR_MISALIGNMENT (dr_info), |
1252 | vect_body); | |
c3e7ee41 | 1253 | |
73fbfcad | 1254 | if (dump_enabled_p ()) |
78c60e3d SS |
1255 | dump_printf_loc (MSG_NOTE, vect_location, |
1256 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1257 | "hardware.\n"); |
ebfd146a IR |
1258 | |
1259 | break; | |
1260 | } | |
1261 | case dr_explicit_realign: | |
1262 | { | |
92345349 BS |
1263 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1264 | vector_load, stmt_info, 0, vect_body); | |
1265 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1266 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1267 | |
1268 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1269 | the containing loop, the following cost should be added to the | |
92345349 | 1270 | prologue costs. */ |
ebfd146a | 1271 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1272 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1273 | stmt_info, 0, vect_body); | |
ebfd146a | 1274 | |
73fbfcad | 1275 | if (dump_enabled_p ()) |
e645e942 TJ |
1276 | dump_printf_loc (MSG_NOTE, vect_location, |
1277 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1278 | |
ebfd146a IR |
1279 | break; |
1280 | } | |
1281 | case dr_explicit_realign_optimized: | |
1282 | { | |
73fbfcad | 1283 | if (dump_enabled_p ()) |
e645e942 | 1284 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1285 | "vect_model_load_cost: unaligned software " |
e645e942 | 1286 | "pipelined.\n"); |
ebfd146a IR |
1287 | |
1288 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1289 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1290 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1291 | access, then the above cost should only be considered for one |
ff802fa1 | 1292 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1293 | and a realignment op. */ |
1294 | ||
92345349 | 1295 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1296 | { |
92345349 BS |
1297 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1298 | vector_stmt, stmt_info, | |
1299 | 0, vect_prologue); | |
ebfd146a | 1300 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1301 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1302 | vector_stmt, stmt_info, | |
1303 | 0, vect_prologue); | |
ebfd146a IR |
1304 | } |
1305 | ||
92345349 BS |
1306 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1307 | stmt_info, 0, vect_body); | |
1308 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1309 | stmt_info, 0, vect_body); | |
8bd37302 | 1310 | |
73fbfcad | 1311 | if (dump_enabled_p ()) |
78c60e3d | 1312 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1313 | "vect_model_load_cost: explicit realign optimized" |
1314 | "\n"); | |
8bd37302 | 1315 | |
ebfd146a IR |
1316 | break; |
1317 | } | |
1318 | ||
38eec4c6 UW |
1319 | case dr_unaligned_unsupported: |
1320 | { | |
1321 | *inside_cost = VECT_MAX_COST; | |
1322 | ||
73fbfcad | 1323 | if (dump_enabled_p ()) |
78c60e3d | 1324 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1325 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1326 | break; |
1327 | } | |
1328 | ||
ebfd146a IR |
1329 | default: |
1330 | gcc_unreachable (); | |
1331 | } | |
ebfd146a IR |
1332 | } |
1333 | ||
418b7df3 | 1334 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
32e8e429 | 1335 | the loop preheader for the vectorized stmt STMT_VINFO. */ |
ebfd146a | 1336 | |
418b7df3 | 1337 | static void |
32e8e429 RS |
1338 | vect_init_vector_1 (stmt_vec_info stmt_vinfo, gimple *new_stmt, |
1339 | gimple_stmt_iterator *gsi) | |
ebfd146a | 1340 | { |
ebfd146a | 1341 | if (gsi) |
a1824cfd | 1342 | vect_finish_stmt_generation (stmt_vinfo, new_stmt, gsi); |
ebfd146a IR |
1343 | else |
1344 | { | |
1345 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
b8698a0f | 1346 | |
a70d6342 IR |
1347 | if (loop_vinfo) |
1348 | { | |
1349 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1350 | basic_block new_bb; |
1351 | edge pe; | |
a70d6342 | 1352 | |
86a91c0a RS |
1353 | if (nested_in_vect_loop_p (loop, stmt_vinfo)) |
1354 | loop = loop->inner; | |
b8698a0f | 1355 | |
a70d6342 | 1356 | pe = loop_preheader_edge (loop); |
418b7df3 | 1357 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1358 | gcc_assert (!new_bb); |
1359 | } | |
1360 | else | |
1361 | { | |
1362 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1363 | basic_block bb; | |
1364 | gimple_stmt_iterator gsi_bb_start; | |
1365 | ||
1366 | gcc_assert (bb_vinfo); | |
1367 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1368 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1369 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1370 | } |
ebfd146a IR |
1371 | } |
1372 | ||
73fbfcad | 1373 | if (dump_enabled_p ()) |
3c2a8ed0 DM |
1374 | dump_printf_loc (MSG_NOTE, vect_location, |
1375 | "created new init_stmt: %G", new_stmt); | |
418b7df3 RG |
1376 | } |
1377 | ||
1378 | /* Function vect_init_vector. | |
ebfd146a | 1379 | |
5467ee52 RG |
1380 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1381 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1382 | vector type a vector with all elements equal to VAL is created first. | |
1383 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1384 | initialization at the loop preheader. | |
418b7df3 | 1385 | Return the DEF of INIT_STMT. |
32e8e429 | 1386 | It will be used in the vectorization of STMT_INFO. */ |
418b7df3 RG |
1387 | |
1388 | tree | |
32e8e429 RS |
1389 | vect_init_vector (stmt_vec_info stmt_info, tree val, tree type, |
1390 | gimple_stmt_iterator *gsi) | |
418b7df3 | 1391 | { |
355fe088 | 1392 | gimple *init_stmt; |
418b7df3 RG |
1393 | tree new_temp; |
1394 | ||
e412ece4 RB |
1395 | /* We abuse this function to push sth to a SSA name with initial 'val'. */ |
1396 | if (! useless_type_conversion_p (type, TREE_TYPE (val))) | |
418b7df3 | 1397 | { |
e412ece4 RB |
1398 | gcc_assert (TREE_CODE (type) == VECTOR_TYPE); |
1399 | if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) | |
418b7df3 | 1400 | { |
5a308cf1 IE |
1401 | /* Scalar boolean value should be transformed into |
1402 | all zeros or all ones value before building a vector. */ | |
1403 | if (VECTOR_BOOLEAN_TYPE_P (type)) | |
1404 | { | |
b3d51f23 IE |
1405 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1406 | tree false_val = build_zero_cst (TREE_TYPE (type)); | |
5a308cf1 IE |
1407 | |
1408 | if (CONSTANT_CLASS_P (val)) | |
1409 | val = integer_zerop (val) ? false_val : true_val; | |
1410 | else | |
1411 | { | |
1412 | new_temp = make_ssa_name (TREE_TYPE (type)); | |
1413 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, | |
1414 | val, true_val, false_val); | |
a1824cfd | 1415 | vect_init_vector_1 (stmt_info, init_stmt, gsi); |
5a308cf1 IE |
1416 | val = new_temp; |
1417 | } | |
1418 | } | |
1419 | else if (CONSTANT_CLASS_P (val)) | |
42fd8198 | 1420 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1421 | else |
1422 | { | |
b731b390 | 1423 | new_temp = make_ssa_name (TREE_TYPE (type)); |
e412ece4 RB |
1424 | if (! INTEGRAL_TYPE_P (TREE_TYPE (val))) |
1425 | init_stmt = gimple_build_assign (new_temp, | |
1426 | fold_build1 (VIEW_CONVERT_EXPR, | |
1427 | TREE_TYPE (type), | |
1428 | val)); | |
1429 | else | |
1430 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); | |
a1824cfd | 1431 | vect_init_vector_1 (stmt_info, init_stmt, gsi); |
5467ee52 | 1432 | val = new_temp; |
418b7df3 RG |
1433 | } |
1434 | } | |
5467ee52 | 1435 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1436 | } |
1437 | ||
0e22bb5a RB |
1438 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1439 | init_stmt = gimple_build_assign (new_temp, val); | |
a1824cfd | 1440 | vect_init_vector_1 (stmt_info, init_stmt, gsi); |
0e22bb5a | 1441 | return new_temp; |
ebfd146a IR |
1442 | } |
1443 | ||
c83a894c | 1444 | /* Function vect_get_vec_def_for_operand_1. |
a70d6342 | 1445 | |
32e8e429 RS |
1446 | For a defining stmt DEF_STMT_INFO of a scalar stmt, return a vector def |
1447 | with type DT that will be used in the vectorized stmt. */ | |
ebfd146a IR |
1448 | |
1449 | tree | |
32e8e429 RS |
1450 | vect_get_vec_def_for_operand_1 (stmt_vec_info def_stmt_info, |
1451 | enum vect_def_type dt) | |
ebfd146a IR |
1452 | { |
1453 | tree vec_oprnd; | |
1eede195 | 1454 | stmt_vec_info vec_stmt_info; |
ebfd146a IR |
1455 | |
1456 | switch (dt) | |
1457 | { | |
81c40241 | 1458 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1459 | case vect_constant_def: |
81c40241 | 1460 | case vect_external_def: |
c83a894c AH |
1461 | /* Code should use vect_get_vec_def_for_operand. */ |
1462 | gcc_unreachable (); | |
ebfd146a | 1463 | |
81c40241 | 1464 | /* operand is defined inside the loop. */ |
8644a673 | 1465 | case vect_internal_def: |
ebfd146a | 1466 | { |
ebfd146a | 1467 | /* Get the def from the vectorized stmt. */ |
1eede195 RS |
1468 | vec_stmt_info = STMT_VINFO_VEC_STMT (def_stmt_info); |
1469 | /* Get vectorized pattern statement. */ | |
1470 | if (!vec_stmt_info | |
1471 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1472 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1473 | vec_stmt_info = (STMT_VINFO_VEC_STMT | |
1474 | (STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
1475 | gcc_assert (vec_stmt_info); | |
1476 | if (gphi *phi = dyn_cast <gphi *> (vec_stmt_info->stmt)) | |
1477 | vec_oprnd = PHI_RESULT (phi); | |
ebfd146a | 1478 | else |
1eede195 RS |
1479 | vec_oprnd = gimple_get_lhs (vec_stmt_info->stmt); |
1480 | return vec_oprnd; | |
ebfd146a IR |
1481 | } |
1482 | ||
c78e3652 | 1483 | /* operand is defined by a loop header phi. */ |
ebfd146a | 1484 | case vect_reduction_def: |
06066f92 | 1485 | case vect_double_reduction_def: |
7c5222ff | 1486 | case vect_nested_cycle: |
ebfd146a IR |
1487 | case vect_induction_def: |
1488 | { | |
32e8e429 | 1489 | gcc_assert (gimple_code (def_stmt_info->stmt) == GIMPLE_PHI); |
ebfd146a | 1490 | |
1eede195 | 1491 | /* Get the def from the vectorized stmt. */ |
1eede195 RS |
1492 | vec_stmt_info = STMT_VINFO_VEC_STMT (def_stmt_info); |
1493 | if (gphi *phi = dyn_cast <gphi *> (vec_stmt_info->stmt)) | |
1494 | vec_oprnd = PHI_RESULT (phi); | |
6dbbece6 | 1495 | else |
1eede195 RS |
1496 | vec_oprnd = gimple_get_lhs (vec_stmt_info->stmt); |
1497 | return vec_oprnd; | |
ebfd146a IR |
1498 | } |
1499 | ||
1500 | default: | |
1501 | gcc_unreachable (); | |
1502 | } | |
1503 | } | |
1504 | ||
1505 | ||
c83a894c AH |
1506 | /* Function vect_get_vec_def_for_operand. |
1507 | ||
32e8e429 RS |
1508 | OP is an operand in STMT_VINFO. This function returns a (vector) def |
1509 | that will be used in the vectorized stmt for STMT_VINFO. | |
c83a894c AH |
1510 | |
1511 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1512 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1513 | ||
1514 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1515 | needs to be introduced. VECTYPE may be used to specify a required type for | |
1516 | vector invariant. */ | |
1517 | ||
1518 | tree | |
32e8e429 | 1519 | vect_get_vec_def_for_operand (tree op, stmt_vec_info stmt_vinfo, tree vectype) |
c83a894c AH |
1520 | { |
1521 | gimple *def_stmt; | |
1522 | enum vect_def_type dt; | |
1523 | bool is_simple_use; | |
c83a894c AH |
1524 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
1525 | ||
1526 | if (dump_enabled_p ()) | |
3c2a8ed0 DM |
1527 | dump_printf_loc (MSG_NOTE, vect_location, |
1528 | "vect_get_vec_def_for_operand: %T\n", op); | |
c83a894c | 1529 | |
fef96d8e RS |
1530 | stmt_vec_info def_stmt_info; |
1531 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &dt, | |
1532 | &def_stmt_info, &def_stmt); | |
c83a894c AH |
1533 | gcc_assert (is_simple_use); |
1534 | if (def_stmt && dump_enabled_p ()) | |
3c2a8ed0 | 1535 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = %G", def_stmt); |
c83a894c AH |
1536 | |
1537 | if (dt == vect_constant_def || dt == vect_external_def) | |
1538 | { | |
1539 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
1540 | tree vector_type; | |
1541 | ||
1542 | if (vectype) | |
1543 | vector_type = vectype; | |
2568d8a1 | 1544 | else if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op)) |
c83a894c AH |
1545 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) |
1546 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1547 | else | |
1548 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1549 | ||
1550 | gcc_assert (vector_type); | |
86a91c0a | 1551 | return vect_init_vector (stmt_vinfo, op, vector_type, NULL); |
c83a894c AH |
1552 | } |
1553 | else | |
fef96d8e | 1554 | return vect_get_vec_def_for_operand_1 (def_stmt_info, dt); |
c83a894c AH |
1555 | } |
1556 | ||
1557 | ||
ebfd146a IR |
1558 | /* Function vect_get_vec_def_for_stmt_copy |
1559 | ||
ff802fa1 | 1560 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1561 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1562 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1563 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1564 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
e4057a39 | 1565 | of the stmt that defines VEC_OPRND. VINFO describes the vectorization. |
ebfd146a IR |
1566 | |
1567 | Context: | |
1568 | In case the vectorization factor (VF) is bigger than the number | |
1569 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1570 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1571 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1572 | smallest data-type determines the VF, and as a result, when vectorizing |
1573 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1574 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1575 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1576 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1577 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1578 | ||
1579 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1580 | |
ebfd146a IR |
1581 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1582 | VS1.1: vx.1 = memref1 VS1.2 | |
1583 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1584 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1585 | |
1586 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1587 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1588 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1589 | VSnew.3: vz3 = vx.3 + ... | |
1590 | ||
1591 | The vectorization of S1 is explained in vectorizable_load. | |
1592 | The vectorization of S2: | |
b8698a0f L |
1593 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1594 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1595 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1596 | returns the vector-def 'vx.0'. |
1597 | ||
b8698a0f L |
1598 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1599 | function is called to get the relevant vector-def for each operand. It is | |
1600 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1601 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1602 | ||
b8698a0f L |
1603 | For example, to obtain the vector-def 'vx.1' in order to create the |
1604 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1605 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1606 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1607 | and return its def ('vx.1'). | |
1608 | Overall, to create the above sequence this function will be called 3 times: | |
e4057a39 RS |
1609 | vx.1 = vect_get_vec_def_for_stmt_copy (vinfo, vx.0); |
1610 | vx.2 = vect_get_vec_def_for_stmt_copy (vinfo, vx.1); | |
1611 | vx.3 = vect_get_vec_def_for_stmt_copy (vinfo, vx.2); */ | |
ebfd146a IR |
1612 | |
1613 | tree | |
e4057a39 | 1614 | vect_get_vec_def_for_stmt_copy (vec_info *vinfo, tree vec_oprnd) |
ebfd146a | 1615 | { |
e4057a39 RS |
1616 | stmt_vec_info def_stmt_info = vinfo->lookup_def (vec_oprnd); |
1617 | if (!def_stmt_info) | |
1618 | /* Do nothing; can reuse same def. */ | |
ebfd146a IR |
1619 | return vec_oprnd; |
1620 | ||
e4057a39 | 1621 | def_stmt_info = STMT_VINFO_RELATED_STMT (def_stmt_info); |
ebfd146a | 1622 | gcc_assert (def_stmt_info); |
e4057a39 RS |
1623 | if (gphi *phi = dyn_cast <gphi *> (def_stmt_info->stmt)) |
1624 | vec_oprnd = PHI_RESULT (phi); | |
ebfd146a | 1625 | else |
e4057a39 | 1626 | vec_oprnd = gimple_get_lhs (def_stmt_info->stmt); |
ebfd146a IR |
1627 | return vec_oprnd; |
1628 | } | |
1629 | ||
1630 | ||
1631 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1632 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a | 1633 | |
c78e3652 | 1634 | void |
e4057a39 | 1635 | vect_get_vec_defs_for_stmt_copy (vec_info *vinfo, |
9771b263 DN |
1636 | vec<tree> *vec_oprnds0, |
1637 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1638 | { |
9771b263 | 1639 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a | 1640 | |
e4057a39 | 1641 | vec_oprnd = vect_get_vec_def_for_stmt_copy (vinfo, vec_oprnd); |
9771b263 | 1642 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1643 | |
9771b263 | 1644 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1645 | { |
9771b263 | 1646 | vec_oprnd = vec_oprnds1->pop (); |
e4057a39 | 1647 | vec_oprnd = vect_get_vec_def_for_stmt_copy (vinfo, vec_oprnd); |
9771b263 | 1648 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1649 | } |
1650 | } | |
1651 | ||
1652 | ||
c78e3652 | 1653 | /* Get vectorized definitions for OP0 and OP1. */ |
ebfd146a | 1654 | |
c78e3652 | 1655 | void |
32e8e429 | 1656 | vect_get_vec_defs (tree op0, tree op1, stmt_vec_info stmt_info, |
9771b263 DN |
1657 | vec<tree> *vec_oprnds0, |
1658 | vec<tree> *vec_oprnds1, | |
306b0c92 | 1659 | slp_tree slp_node) |
ebfd146a IR |
1660 | { |
1661 | if (slp_node) | |
d092494c IR |
1662 | { |
1663 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1664 | auto_vec<tree> ops (nops); |
1665 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1666 | |
9771b263 | 1667 | ops.quick_push (op0); |
d092494c | 1668 | if (op1) |
9771b263 | 1669 | ops.quick_push (op1); |
d092494c | 1670 | |
306b0c92 | 1671 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
d092494c | 1672 | |
37b5ec8f | 1673 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1674 | if (op1) |
37b5ec8f | 1675 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1676 | } |
ebfd146a IR |
1677 | else |
1678 | { | |
1679 | tree vec_oprnd; | |
1680 | ||
9771b263 | 1681 | vec_oprnds0->create (1); |
a1824cfd | 1682 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt_info); |
9771b263 | 1683 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1684 | |
1685 | if (op1) | |
1686 | { | |
9771b263 | 1687 | vec_oprnds1->create (1); |
a1824cfd | 1688 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt_info); |
9771b263 | 1689 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1690 | } |
1691 | } | |
1692 | } | |
1693 | ||
bb6c2b68 RS |
1694 | /* Helper function called by vect_finish_replace_stmt and |
1695 | vect_finish_stmt_generation. Set the location of the new | |
e1bd7296 | 1696 | statement and create and return a stmt_vec_info for it. */ |
bb6c2b68 | 1697 | |
e1bd7296 | 1698 | static stmt_vec_info |
32e8e429 | 1699 | vect_finish_stmt_generation_1 (stmt_vec_info stmt_info, gimple *vec_stmt) |
bb6c2b68 | 1700 | { |
bb6c2b68 RS |
1701 | vec_info *vinfo = stmt_info->vinfo; |
1702 | ||
e1bd7296 | 1703 | stmt_vec_info vec_stmt_info = vinfo->add_stmt (vec_stmt); |
bb6c2b68 RS |
1704 | |
1705 | if (dump_enabled_p ()) | |
3c2a8ed0 | 1706 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: %G", vec_stmt); |
bb6c2b68 | 1707 | |
86a91c0a | 1708 | gimple_set_location (vec_stmt, gimple_location (stmt_info->stmt)); |
bb6c2b68 RS |
1709 | |
1710 | /* While EH edges will generally prevent vectorization, stmt might | |
1711 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1712 | that could throw are part of the same region. */ | |
86a91c0a | 1713 | int lp_nr = lookup_stmt_eh_lp (stmt_info->stmt); |
36bbc05d | 1714 | if (lp_nr != 0 && stmt_could_throw_p (cfun, vec_stmt)) |
bb6c2b68 | 1715 | add_stmt_to_eh_lp (vec_stmt, lp_nr); |
e1bd7296 RS |
1716 | |
1717 | return vec_stmt_info; | |
bb6c2b68 RS |
1718 | } |
1719 | ||
32e8e429 RS |
1720 | /* Replace the scalar statement STMT_INFO with a new vector statement VEC_STMT, |
1721 | which sets the same scalar result as STMT_INFO did. Create and return a | |
e1bd7296 | 1722 | stmt_vec_info for VEC_STMT. */ |
bb6c2b68 | 1723 | |
e1bd7296 | 1724 | stmt_vec_info |
32e8e429 | 1725 | vect_finish_replace_stmt (stmt_vec_info stmt_info, gimple *vec_stmt) |
bb6c2b68 | 1726 | { |
a1824cfd | 1727 | gcc_assert (gimple_get_lhs (stmt_info->stmt) == gimple_get_lhs (vec_stmt)); |
bb6c2b68 | 1728 | |
a1824cfd | 1729 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt_info->stmt); |
bb6c2b68 RS |
1730 | gsi_replace (&gsi, vec_stmt, false); |
1731 | ||
a1824cfd | 1732 | return vect_finish_stmt_generation_1 (stmt_info, vec_stmt); |
bb6c2b68 | 1733 | } |
ebfd146a | 1734 | |
32e8e429 | 1735 | /* Add VEC_STMT to the vectorized implementation of STMT_INFO and insert it |
e1bd7296 | 1736 | before *GSI. Create and return a stmt_vec_info for VEC_STMT. */ |
ebfd146a | 1737 | |
e1bd7296 | 1738 | stmt_vec_info |
32e8e429 | 1739 | vect_finish_stmt_generation (stmt_vec_info stmt_info, gimple *vec_stmt, |
ebfd146a IR |
1740 | gimple_stmt_iterator *gsi) |
1741 | { | |
a1824cfd | 1742 | gcc_assert (gimple_code (stmt_info->stmt) != GIMPLE_LABEL); |
ebfd146a | 1743 | |
54e8e2c3 RG |
1744 | if (!gsi_end_p (*gsi) |
1745 | && gimple_has_mem_ops (vec_stmt)) | |
1746 | { | |
355fe088 | 1747 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1748 | tree vuse = gimple_vuse (at_stmt); |
1749 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1750 | { | |
1751 | tree vdef = gimple_vdef (at_stmt); | |
1752 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1753 | /* If we have an SSA vuse and insert a store, update virtual | |
1754 | SSA form to avoid triggering the renamer. Do so only | |
1755 | if we can easily see all uses - which is what almost always | |
1756 | happens with the way vectorized stmts are inserted. */ | |
1757 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1758 | && ((is_gimple_assign (vec_stmt) | |
1759 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1760 | || (is_gimple_call (vec_stmt) | |
1761 | && !(gimple_call_flags (vec_stmt) | |
1762 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1763 | { | |
1764 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1765 | gimple_set_vdef (vec_stmt, new_vdef); | |
1766 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1767 | } | |
1768 | } | |
1769 | } | |
ebfd146a | 1770 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
a1824cfd | 1771 | return vect_finish_stmt_generation_1 (stmt_info, vec_stmt); |
ebfd146a IR |
1772 | } |
1773 | ||
70439f0d RS |
1774 | /* We want to vectorize a call to combined function CFN with function |
1775 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN | |
1776 | as the types of all inputs. Check whether this is possible using | |
1777 | an internal function, returning its code if so or IFN_LAST if not. */ | |
ebfd146a | 1778 | |
70439f0d RS |
1779 | static internal_fn |
1780 | vectorizable_internal_function (combined_fn cfn, tree fndecl, | |
1781 | tree vectype_out, tree vectype_in) | |
ebfd146a | 1782 | { |
70439f0d RS |
1783 | internal_fn ifn; |
1784 | if (internal_fn_p (cfn)) | |
1785 | ifn = as_internal_fn (cfn); | |
1786 | else | |
1787 | ifn = associated_internal_fn (fndecl); | |
1788 | if (ifn != IFN_LAST && direct_internal_fn_p (ifn)) | |
1789 | { | |
1790 | const direct_internal_fn_info &info = direct_internal_fn (ifn); | |
1791 | if (info.vectorizable) | |
1792 | { | |
1793 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); | |
1794 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); | |
d95ab70a RS |
1795 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1796 | OPTIMIZE_FOR_SPEED)) | |
70439f0d RS |
1797 | return ifn; |
1798 | } | |
1799 | } | |
1800 | return IFN_LAST; | |
ebfd146a IR |
1801 | } |
1802 | ||
5ce9450f | 1803 | |
82570274 | 1804 | static tree permute_vec_elements (tree, tree, tree, stmt_vec_info, |
5ce9450f JJ |
1805 | gimple_stmt_iterator *); |
1806 | ||
7cfb4d93 RS |
1807 | /* Check whether a load or store statement in the loop described by |
1808 | LOOP_VINFO is possible in a fully-masked loop. This is testing | |
1809 | whether the vectorizer pass has the appropriate support, as well as | |
1810 | whether the target does. | |
1811 | ||
1812 | VLS_TYPE says whether the statement is a load or store and VECTYPE | |
1813 | is the type of the vector being loaded or stored. MEMORY_ACCESS_TYPE | |
1814 | says how the load or store is going to be implemented and GROUP_SIZE | |
1815 | is the number of load or store statements in the containing group. | |
bfaa08b7 RS |
1816 | If the access is a gather load or scatter store, GS_INFO describes |
1817 | its arguments. | |
7cfb4d93 RS |
1818 | |
1819 | Clear LOOP_VINFO_CAN_FULLY_MASK_P if a fully-masked loop is not | |
1820 | supported, otherwise record the required mask types. */ | |
1821 | ||
1822 | static void | |
1823 | check_load_store_masking (loop_vec_info loop_vinfo, tree vectype, | |
1824 | vec_load_store_type vls_type, int group_size, | |
bfaa08b7 RS |
1825 | vect_memory_access_type memory_access_type, |
1826 | gather_scatter_info *gs_info) | |
7cfb4d93 RS |
1827 | { |
1828 | /* Invariant loads need no special support. */ | |
1829 | if (memory_access_type == VMAT_INVARIANT) | |
1830 | return; | |
1831 | ||
1832 | vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); | |
1833 | machine_mode vecmode = TYPE_MODE (vectype); | |
1834 | bool is_load = (vls_type == VLS_LOAD); | |
1835 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
1836 | { | |
1837 | if (is_load | |
1838 | ? !vect_load_lanes_supported (vectype, group_size, true) | |
1839 | : !vect_store_lanes_supported (vectype, group_size, true)) | |
1840 | { | |
1841 | if (dump_enabled_p ()) | |
1842 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1843 | "can't use a fully-masked loop because the" | |
1844 | " target doesn't have an appropriate masked" | |
1845 | " load/store-lanes instruction.\n"); | |
1846 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1847 | return; | |
1848 | } | |
1849 | unsigned int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
1850 | vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype); | |
1851 | return; | |
1852 | } | |
1853 | ||
bfaa08b7 RS |
1854 | if (memory_access_type == VMAT_GATHER_SCATTER) |
1855 | { | |
f307441a RS |
1856 | internal_fn ifn = (is_load |
1857 | ? IFN_MASK_GATHER_LOAD | |
1858 | : IFN_MASK_SCATTER_STORE); | |
bfaa08b7 | 1859 | tree offset_type = TREE_TYPE (gs_info->offset); |
f307441a | 1860 | if (!internal_gather_scatter_fn_supported_p (ifn, vectype, |
bfaa08b7 RS |
1861 | gs_info->memory_type, |
1862 | TYPE_SIGN (offset_type), | |
1863 | gs_info->scale)) | |
1864 | { | |
1865 | if (dump_enabled_p ()) | |
1866 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1867 | "can't use a fully-masked loop because the" | |
1868 | " target doesn't have an appropriate masked" | |
f307441a | 1869 | " gather load or scatter store instruction.\n"); |
bfaa08b7 RS |
1870 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; |
1871 | return; | |
1872 | } | |
1873 | unsigned int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
1874 | vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype); | |
1875 | return; | |
1876 | } | |
1877 | ||
7cfb4d93 RS |
1878 | if (memory_access_type != VMAT_CONTIGUOUS |
1879 | && memory_access_type != VMAT_CONTIGUOUS_PERMUTE) | |
1880 | { | |
1881 | /* Element X of the data must come from iteration i * VF + X of the | |
1882 | scalar loop. We need more work to support other mappings. */ | |
1883 | if (dump_enabled_p ()) | |
1884 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1885 | "can't use a fully-masked loop because an access" | |
1886 | " isn't contiguous.\n"); | |
1887 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1888 | return; | |
1889 | } | |
1890 | ||
1891 | machine_mode mask_mode; | |
1892 | if (!(targetm.vectorize.get_mask_mode | |
1893 | (GET_MODE_NUNITS (vecmode), | |
1894 | GET_MODE_SIZE (vecmode)).exists (&mask_mode)) | |
1895 | || !can_vec_mask_load_store_p (vecmode, mask_mode, is_load)) | |
1896 | { | |
1897 | if (dump_enabled_p ()) | |
1898 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1899 | "can't use a fully-masked loop because the target" | |
1900 | " doesn't have the appropriate masked load or" | |
1901 | " store.\n"); | |
1902 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1903 | return; | |
1904 | } | |
1905 | /* We might load more scalars than we need for permuting SLP loads. | |
1906 | We checked in get_group_load_store_type that the extra elements | |
1907 | don't leak into a new vector. */ | |
1908 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
1909 | poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
1910 | unsigned int nvectors; | |
1911 | if (can_div_away_from_zero_p (group_size * vf, nunits, &nvectors)) | |
1912 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype); | |
1913 | else | |
1914 | gcc_unreachable (); | |
1915 | } | |
1916 | ||
1917 | /* Return the mask input to a masked load or store. VEC_MASK is the vectorized | |
1918 | form of the scalar mask condition and LOOP_MASK, if nonnull, is the mask | |
1919 | that needs to be applied to all loads and stores in a vectorized loop. | |
1920 | Return VEC_MASK if LOOP_MASK is null, otherwise return VEC_MASK & LOOP_MASK. | |
1921 | ||
1922 | MASK_TYPE is the type of both masks. If new statements are needed, | |
1923 | insert them before GSI. */ | |
1924 | ||
1925 | static tree | |
1926 | prepare_load_store_mask (tree mask_type, tree loop_mask, tree vec_mask, | |
1927 | gimple_stmt_iterator *gsi) | |
1928 | { | |
1929 | gcc_assert (useless_type_conversion_p (mask_type, TREE_TYPE (vec_mask))); | |
1930 | if (!loop_mask) | |
1931 | return vec_mask; | |
1932 | ||
1933 | gcc_assert (TREE_TYPE (loop_mask) == mask_type); | |
1934 | tree and_res = make_temp_ssa_name (mask_type, NULL, "vec_mask_and"); | |
1935 | gimple *and_stmt = gimple_build_assign (and_res, BIT_AND_EXPR, | |
1936 | vec_mask, loop_mask); | |
1937 | gsi_insert_before (gsi, and_stmt, GSI_SAME_STMT); | |
1938 | return and_res; | |
1939 | } | |
1940 | ||
429ef523 | 1941 | /* Determine whether we can use a gather load or scatter store to vectorize |
32e8e429 RS |
1942 | strided load or store STMT_INFO by truncating the current offset to a |
1943 | smaller width. We need to be able to construct an offset vector: | |
429ef523 RS |
1944 | |
1945 | { 0, X, X*2, X*3, ... } | |
1946 | ||
32e8e429 | 1947 | without loss of precision, where X is STMT_INFO's DR_STEP. |
429ef523 RS |
1948 | |
1949 | Return true if this is possible, describing the gather load or scatter | |
1950 | store in GS_INFO. MASKED_P is true if the load or store is conditional. */ | |
1951 | ||
1952 | static bool | |
32e8e429 RS |
1953 | vect_truncate_gather_scatter_offset (stmt_vec_info stmt_info, |
1954 | loop_vec_info loop_vinfo, bool masked_p, | |
429ef523 RS |
1955 | gather_scatter_info *gs_info) |
1956 | { | |
89fa689a RS |
1957 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info); |
1958 | data_reference *dr = dr_info->dr; | |
429ef523 RS |
1959 | tree step = DR_STEP (dr); |
1960 | if (TREE_CODE (step) != INTEGER_CST) | |
1961 | { | |
1962 | /* ??? Perhaps we could use range information here? */ | |
1963 | if (dump_enabled_p ()) | |
1964 | dump_printf_loc (MSG_NOTE, vect_location, | |
1965 | "cannot truncate variable step.\n"); | |
1966 | return false; | |
1967 | } | |
1968 | ||
1969 | /* Get the number of bits in an element. */ | |
1970 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
1971 | scalar_mode element_mode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); | |
1972 | unsigned int element_bits = GET_MODE_BITSIZE (element_mode); | |
1973 | ||
1974 | /* Set COUNT to the upper limit on the number of elements - 1. | |
1975 | Start with the maximum vectorization factor. */ | |
1976 | unsigned HOST_WIDE_INT count = vect_max_vf (loop_vinfo) - 1; | |
1977 | ||
1978 | /* Try lowering COUNT to the number of scalar latch iterations. */ | |
1979 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1980 | widest_int max_iters; | |
1981 | if (max_loop_iterations (loop, &max_iters) | |
1982 | && max_iters < count) | |
1983 | count = max_iters.to_shwi (); | |
1984 | ||
1985 | /* Try scales of 1 and the element size. */ | |
89fa689a | 1986 | int scales[] = { 1, vect_get_scalar_dr_size (dr_info) }; |
4a669ac3 | 1987 | wi::overflow_type overflow = wi::OVF_NONE; |
429ef523 RS |
1988 | for (int i = 0; i < 2; ++i) |
1989 | { | |
1990 | int scale = scales[i]; | |
1991 | widest_int factor; | |
1992 | if (!wi::multiple_of_p (wi::to_widest (step), scale, SIGNED, &factor)) | |
1993 | continue; | |
1994 | ||
1995 | /* See whether we can calculate (COUNT - 1) * STEP / SCALE | |
1996 | in OFFSET_BITS bits. */ | |
4a669ac3 AH |
1997 | widest_int range = wi::mul (count, factor, SIGNED, &overflow); |
1998 | if (overflow) | |
429ef523 RS |
1999 | continue; |
2000 | signop sign = range >= 0 ? UNSIGNED : SIGNED; | |
2001 | if (wi::min_precision (range, sign) > element_bits) | |
2002 | { | |
4a669ac3 | 2003 | overflow = wi::OVF_UNKNOWN; |
429ef523 RS |
2004 | continue; |
2005 | } | |
2006 | ||
2007 | /* See whether the target supports the operation. */ | |
2008 | tree memory_type = TREE_TYPE (DR_REF (dr)); | |
2009 | if (!vect_gather_scatter_fn_p (DR_IS_READ (dr), masked_p, vectype, | |
2010 | memory_type, element_bits, sign, scale, | |
2011 | &gs_info->ifn, &gs_info->element_type)) | |
2012 | continue; | |
2013 | ||
2014 | tree offset_type = build_nonstandard_integer_type (element_bits, | |
2015 | sign == UNSIGNED); | |
2016 | ||
2017 | gs_info->decl = NULL_TREE; | |
2018 | /* Logically the sum of DR_BASE_ADDRESS, DR_INIT and DR_OFFSET, | |
2019 | but we don't need to store that here. */ | |
2020 | gs_info->base = NULL_TREE; | |
2021 | gs_info->offset = fold_convert (offset_type, step); | |
929b4411 | 2022 | gs_info->offset_dt = vect_constant_def; |
429ef523 RS |
2023 | gs_info->offset_vectype = NULL_TREE; |
2024 | gs_info->scale = scale; | |
2025 | gs_info->memory_type = memory_type; | |
2026 | return true; | |
2027 | } | |
2028 | ||
4a669ac3 | 2029 | if (overflow && dump_enabled_p ()) |
429ef523 RS |
2030 | dump_printf_loc (MSG_NOTE, vect_location, |
2031 | "truncating gather/scatter offset to %d bits" | |
2032 | " might change its value.\n", element_bits); | |
2033 | ||
2034 | return false; | |
2035 | } | |
2036 | ||
ab2fc782 | 2037 | /* Return true if we can use gather/scatter internal functions to |
82570274 | 2038 | vectorize STMT_INFO, which is a grouped or strided load or store. |
429ef523 RS |
2039 | MASKED_P is true if load or store is conditional. When returning |
2040 | true, fill in GS_INFO with the information required to perform the | |
2041 | operation. */ | |
ab2fc782 RS |
2042 | |
2043 | static bool | |
82570274 RS |
2044 | vect_use_strided_gather_scatters_p (stmt_vec_info stmt_info, |
2045 | loop_vec_info loop_vinfo, bool masked_p, | |
ab2fc782 RS |
2046 | gather_scatter_info *gs_info) |
2047 | { | |
82570274 | 2048 | if (!vect_check_gather_scatter (stmt_info, loop_vinfo, gs_info) |
ab2fc782 | 2049 | || gs_info->decl) |
82570274 | 2050 | return vect_truncate_gather_scatter_offset (stmt_info, loop_vinfo, |
429ef523 | 2051 | masked_p, gs_info); |
ab2fc782 RS |
2052 | |
2053 | scalar_mode element_mode = SCALAR_TYPE_MODE (gs_info->element_type); | |
2054 | unsigned int element_bits = GET_MODE_BITSIZE (element_mode); | |
2055 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2056 | unsigned int offset_bits = TYPE_PRECISION (offset_type); | |
2057 | ||
2058 | /* Enforced by vect_check_gather_scatter. */ | |
2059 | gcc_assert (element_bits >= offset_bits); | |
2060 | ||
2061 | /* If the elements are wider than the offset, convert the offset to the | |
2062 | same width, without changing its sign. */ | |
2063 | if (element_bits > offset_bits) | |
2064 | { | |
2065 | bool unsigned_p = TYPE_UNSIGNED (offset_type); | |
2066 | offset_type = build_nonstandard_integer_type (element_bits, unsigned_p); | |
2067 | gs_info->offset = fold_convert (offset_type, gs_info->offset); | |
2068 | } | |
2069 | ||
2070 | if (dump_enabled_p ()) | |
2071 | dump_printf_loc (MSG_NOTE, vect_location, | |
2072 | "using gather/scatter for strided/grouped access," | |
2073 | " scale = %d\n", gs_info->scale); | |
2074 | ||
2075 | return true; | |
2076 | } | |
2077 | ||
32e8e429 | 2078 | /* STMT_INFO is a non-strided load or store, meaning that it accesses |
62da9e14 RS |
2079 | elements with a known constant step. Return -1 if that step |
2080 | is negative, 0 if it is zero, and 1 if it is greater than zero. */ | |
2081 | ||
2082 | static int | |
32e8e429 | 2083 | compare_step_with_zero (stmt_vec_info stmt_info) |
62da9e14 | 2084 | { |
89fa689a RS |
2085 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info); |
2086 | return tree_int_cst_compare (vect_dr_behavior (dr_info)->step, | |
3f5e8a76 | 2087 | size_zero_node); |
62da9e14 RS |
2088 | } |
2089 | ||
2090 | /* If the target supports a permute mask that reverses the elements in | |
2091 | a vector of type VECTYPE, return that mask, otherwise return null. */ | |
2092 | ||
2093 | static tree | |
2094 | perm_mask_for_reverse (tree vectype) | |
2095 | { | |
928686b1 | 2096 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
62da9e14 | 2097 | |
d980067b RS |
2098 | /* The encoding has a single stepped pattern. */ |
2099 | vec_perm_builder sel (nunits, 1, 3); | |
928686b1 | 2100 | for (int i = 0; i < 3; ++i) |
908a1a16 | 2101 | sel.quick_push (nunits - 1 - i); |
62da9e14 | 2102 | |
e3342de4 RS |
2103 | vec_perm_indices indices (sel, 1, nunits); |
2104 | if (!can_vec_perm_const_p (TYPE_MODE (vectype), indices)) | |
62da9e14 | 2105 | return NULL_TREE; |
e3342de4 | 2106 | return vect_gen_perm_mask_checked (vectype, indices); |
62da9e14 | 2107 | } |
5ce9450f | 2108 | |
32e8e429 | 2109 | /* STMT_INFO is either a masked or unconditional store. Return the value |
c3a8f964 RS |
2110 | being stored. */ |
2111 | ||
f307441a | 2112 | tree |
32e8e429 | 2113 | vect_get_store_rhs (stmt_vec_info stmt_info) |
c3a8f964 | 2114 | { |
32e8e429 | 2115 | if (gassign *assign = dyn_cast <gassign *> (stmt_info->stmt)) |
c3a8f964 RS |
2116 | { |
2117 | gcc_assert (gimple_assign_single_p (assign)); | |
2118 | return gimple_assign_rhs1 (assign); | |
2119 | } | |
32e8e429 | 2120 | if (gcall *call = dyn_cast <gcall *> (stmt_info->stmt)) |
c3a8f964 RS |
2121 | { |
2122 | internal_fn ifn = gimple_call_internal_fn (call); | |
f307441a RS |
2123 | int index = internal_fn_stored_value_index (ifn); |
2124 | gcc_assert (index >= 0); | |
32e8e429 | 2125 | return gimple_call_arg (call, index); |
c3a8f964 RS |
2126 | } |
2127 | gcc_unreachable (); | |
2128 | } | |
2129 | ||
2de001ee | 2130 | /* A subroutine of get_load_store_type, with a subset of the same |
32e8e429 | 2131 | arguments. Handle the case where STMT_INFO is part of a grouped load |
2de001ee RS |
2132 | or store. |
2133 | ||
2134 | For stores, the statements in the group are all consecutive | |
2135 | and there is no gap at the end. For loads, the statements in the | |
2136 | group might not be consecutive; there can be gaps between statements | |
2137 | as well as at the end. */ | |
2138 | ||
2139 | static bool | |
32e8e429 | 2140 | get_group_load_store_type (stmt_vec_info stmt_info, tree vectype, bool slp, |
7e11fc7f | 2141 | bool masked_p, vec_load_store_type vls_type, |
429ef523 RS |
2142 | vect_memory_access_type *memory_access_type, |
2143 | gather_scatter_info *gs_info) | |
2de001ee | 2144 | { |
2de001ee RS |
2145 | vec_info *vinfo = stmt_info->vinfo; |
2146 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2147 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; | |
bffb8014 | 2148 | stmt_vec_info first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
89fa689a | 2149 | dr_vec_info *first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info); |
bffb8014 RS |
2150 | unsigned int group_size = DR_GROUP_SIZE (first_stmt_info); |
2151 | bool single_element_p = (stmt_info == first_stmt_info | |
2c53b149 | 2152 | && !DR_GROUP_NEXT_ELEMENT (stmt_info)); |
bffb8014 | 2153 | unsigned HOST_WIDE_INT gap = DR_GROUP_GAP (first_stmt_info); |
928686b1 | 2154 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
2155 | |
2156 | /* True if the vectorized statements would access beyond the last | |
2157 | statement in the group. */ | |
2158 | bool overrun_p = false; | |
2159 | ||
2160 | /* True if we can cope with such overrun by peeling for gaps, so that | |
2161 | there is at least one final scalar iteration after the vector loop. */ | |
7e11fc7f RS |
2162 | bool can_overrun_p = (!masked_p |
2163 | && vls_type == VLS_LOAD | |
2164 | && loop_vinfo | |
2165 | && !loop->inner); | |
2de001ee RS |
2166 | |
2167 | /* There can only be a gap at the end of the group if the stride is | |
2168 | known at compile time. */ | |
3ad3b3ac | 2169 | gcc_assert (!STMT_VINFO_STRIDED_P (first_stmt_info) || gap == 0); |
2de001ee RS |
2170 | |
2171 | /* Stores can't yet have gaps. */ | |
2172 | gcc_assert (slp || vls_type == VLS_LOAD || gap == 0); | |
2173 | ||
2174 | if (slp) | |
2175 | { | |
3ad3b3ac | 2176 | if (STMT_VINFO_STRIDED_P (first_stmt_info)) |
2de001ee | 2177 | { |
2c53b149 | 2178 | /* Try to use consecutive accesses of DR_GROUP_SIZE elements, |
2de001ee RS |
2179 | separated by the stride, until we have a complete vector. |
2180 | Fall back to scalar accesses if that isn't possible. */ | |
928686b1 | 2181 | if (multiple_p (nunits, group_size)) |
2de001ee RS |
2182 | *memory_access_type = VMAT_STRIDED_SLP; |
2183 | else | |
2184 | *memory_access_type = VMAT_ELEMENTWISE; | |
2185 | } | |
2186 | else | |
2187 | { | |
2188 | overrun_p = loop_vinfo && gap != 0; | |
2189 | if (overrun_p && vls_type != VLS_LOAD) | |
2190 | { | |
2191 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2192 | "Grouped store with gaps requires" | |
2193 | " non-consecutive accesses\n"); | |
2194 | return false; | |
2195 | } | |
f702e7d4 RS |
2196 | /* An overrun is fine if the trailing elements are smaller |
2197 | than the alignment boundary B. Every vector access will | |
2198 | be a multiple of B and so we are guaranteed to access a | |
2199 | non-gap element in the same B-sized block. */ | |
f9ef2c76 | 2200 | if (overrun_p |
89fa689a RS |
2201 | && gap < (vect_known_alignment_in_bytes (first_dr_info) |
2202 | / vect_get_scalar_dr_size (first_dr_info))) | |
f9ef2c76 | 2203 | overrun_p = false; |
2de001ee RS |
2204 | if (overrun_p && !can_overrun_p) |
2205 | { | |
2206 | if (dump_enabled_p ()) | |
2207 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2208 | "Peeling for outer loop is not supported\n"); | |
2209 | return false; | |
2210 | } | |
2211 | *memory_access_type = VMAT_CONTIGUOUS; | |
2212 | } | |
2213 | } | |
2214 | else | |
2215 | { | |
2216 | /* We can always handle this case using elementwise accesses, | |
2217 | but see if something more efficient is available. */ | |
2218 | *memory_access_type = VMAT_ELEMENTWISE; | |
2219 | ||
2220 | /* If there is a gap at the end of the group then these optimizations | |
2221 | would access excess elements in the last iteration. */ | |
2222 | bool would_overrun_p = (gap != 0); | |
f702e7d4 RS |
2223 | /* An overrun is fine if the trailing elements are smaller than the |
2224 | alignment boundary B. Every vector access will be a multiple of B | |
2225 | and so we are guaranteed to access a non-gap element in the | |
2226 | same B-sized block. */ | |
f9ef2c76 | 2227 | if (would_overrun_p |
7e11fc7f | 2228 | && !masked_p |
89fa689a RS |
2229 | && gap < (vect_known_alignment_in_bytes (first_dr_info) |
2230 | / vect_get_scalar_dr_size (first_dr_info))) | |
f9ef2c76 | 2231 | would_overrun_p = false; |
f702e7d4 | 2232 | |
3ad3b3ac | 2233 | if (!STMT_VINFO_STRIDED_P (first_stmt_info) |
62da9e14 | 2234 | && (can_overrun_p || !would_overrun_p) |
86a91c0a | 2235 | && compare_step_with_zero (stmt_info) > 0) |
2de001ee | 2236 | { |
6737facb RS |
2237 | /* First cope with the degenerate case of a single-element |
2238 | vector. */ | |
2239 | if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U)) | |
2240 | *memory_access_type = VMAT_CONTIGUOUS; | |
2241 | ||
2242 | /* Otherwise try using LOAD/STORE_LANES. */ | |
2243 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2244 | && (vls_type == VLS_LOAD | |
7e11fc7f RS |
2245 | ? vect_load_lanes_supported (vectype, group_size, masked_p) |
2246 | : vect_store_lanes_supported (vectype, group_size, | |
2247 | masked_p))) | |
2de001ee RS |
2248 | { |
2249 | *memory_access_type = VMAT_LOAD_STORE_LANES; | |
2250 | overrun_p = would_overrun_p; | |
2251 | } | |
2252 | ||
2253 | /* If that fails, try using permuting loads. */ | |
2254 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2255 | && (vls_type == VLS_LOAD | |
2256 | ? vect_grouped_load_supported (vectype, single_element_p, | |
2257 | group_size) | |
2258 | : vect_grouped_store_supported (vectype, group_size))) | |
2259 | { | |
2260 | *memory_access_type = VMAT_CONTIGUOUS_PERMUTE; | |
2261 | overrun_p = would_overrun_p; | |
2262 | } | |
2263 | } | |
429ef523 RS |
2264 | |
2265 | /* As a last resort, trying using a gather load or scatter store. | |
2266 | ||
2267 | ??? Although the code can handle all group sizes correctly, | |
2268 | it probably isn't a win to use separate strided accesses based | |
2269 | on nearby locations. Or, even if it's a win over scalar code, | |
2270 | it might not be a win over vectorizing at a lower VF, if that | |
2271 | allows us to use contiguous accesses. */ | |
2272 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2273 | && single_element_p | |
2274 | && loop_vinfo | |
86a91c0a | 2275 | && vect_use_strided_gather_scatters_p (stmt_info, loop_vinfo, |
429ef523 RS |
2276 | masked_p, gs_info)) |
2277 | *memory_access_type = VMAT_GATHER_SCATTER; | |
2de001ee RS |
2278 | } |
2279 | ||
bffb8014 | 2280 | if (vls_type != VLS_LOAD && first_stmt_info == stmt_info) |
2de001ee RS |
2281 | { |
2282 | /* STMT is the leader of the group. Check the operands of all the | |
2283 | stmts of the group. */ | |
bffb8014 RS |
2284 | stmt_vec_info next_stmt_info = DR_GROUP_NEXT_ELEMENT (stmt_info); |
2285 | while (next_stmt_info) | |
2de001ee | 2286 | { |
bffb8014 | 2287 | tree op = vect_get_store_rhs (next_stmt_info); |
2de001ee | 2288 | enum vect_def_type dt; |
894dd753 | 2289 | if (!vect_is_simple_use (op, vinfo, &dt)) |
2de001ee RS |
2290 | { |
2291 | if (dump_enabled_p ()) | |
2292 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2293 | "use not simple.\n"); | |
2294 | return false; | |
2295 | } | |
bffb8014 | 2296 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
2de001ee RS |
2297 | } |
2298 | } | |
2299 | ||
2300 | if (overrun_p) | |
2301 | { | |
2302 | gcc_assert (can_overrun_p); | |
2303 | if (dump_enabled_p ()) | |
2304 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2305 | "Data access with gaps requires scalar " | |
2306 | "epilogue loop\n"); | |
2307 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; | |
2308 | } | |
2309 | ||
2310 | return true; | |
2311 | } | |
2312 | ||
62da9e14 | 2313 | /* A subroutine of get_load_store_type, with a subset of the same |
32e8e429 | 2314 | arguments. Handle the case where STMT_INFO is a load or store that |
62da9e14 RS |
2315 | accesses consecutive elements with a negative step. */ |
2316 | ||
2317 | static vect_memory_access_type | |
32e8e429 | 2318 | get_negative_load_store_type (stmt_vec_info stmt_info, tree vectype, |
62da9e14 RS |
2319 | vec_load_store_type vls_type, |
2320 | unsigned int ncopies) | |
2321 | { | |
89fa689a | 2322 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info); |
62da9e14 RS |
2323 | dr_alignment_support alignment_support_scheme; |
2324 | ||
2325 | if (ncopies > 1) | |
2326 | { | |
2327 | if (dump_enabled_p ()) | |
2328 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2329 | "multiple types with negative step.\n"); | |
2330 | return VMAT_ELEMENTWISE; | |
2331 | } | |
2332 | ||
89fa689a | 2333 | alignment_support_scheme = vect_supportable_dr_alignment (dr_info, false); |
62da9e14 RS |
2334 | if (alignment_support_scheme != dr_aligned |
2335 | && alignment_support_scheme != dr_unaligned_supported) | |
2336 | { | |
2337 | if (dump_enabled_p ()) | |
2338 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2339 | "negative step but alignment required.\n"); | |
2340 | return VMAT_ELEMENTWISE; | |
2341 | } | |
2342 | ||
2343 | if (vls_type == VLS_STORE_INVARIANT) | |
2344 | { | |
2345 | if (dump_enabled_p ()) | |
2346 | dump_printf_loc (MSG_NOTE, vect_location, | |
2347 | "negative step with invariant source;" | |
2348 | " no permute needed.\n"); | |
2349 | return VMAT_CONTIGUOUS_DOWN; | |
2350 | } | |
2351 | ||
2352 | if (!perm_mask_for_reverse (vectype)) | |
2353 | { | |
2354 | if (dump_enabled_p ()) | |
2355 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2356 | "negative step and reversing not supported.\n"); | |
2357 | return VMAT_ELEMENTWISE; | |
2358 | } | |
2359 | ||
2360 | return VMAT_CONTIGUOUS_REVERSE; | |
2361 | } | |
2362 | ||
32e8e429 | 2363 | /* Analyze load or store statement STMT_INFO of type VLS_TYPE. Return true |
2de001ee RS |
2364 | if there is a memory access type that the vectorized form can use, |
2365 | storing it in *MEMORY_ACCESS_TYPE if so. If we decide to use gathers | |
2366 | or scatters, fill in GS_INFO accordingly. | |
2367 | ||
2368 | SLP says whether we're performing SLP rather than loop vectorization. | |
7e11fc7f | 2369 | MASKED_P is true if the statement is conditional on a vectorized mask. |
62da9e14 RS |
2370 | VECTYPE is the vector type that the vectorized statements will use. |
2371 | NCOPIES is the number of vector statements that will be needed. */ | |
2de001ee RS |
2372 | |
2373 | static bool | |
32e8e429 RS |
2374 | get_load_store_type (stmt_vec_info stmt_info, tree vectype, bool slp, |
2375 | bool masked_p, vec_load_store_type vls_type, | |
2376 | unsigned int ncopies, | |
2de001ee RS |
2377 | vect_memory_access_type *memory_access_type, |
2378 | gather_scatter_info *gs_info) | |
2379 | { | |
2de001ee RS |
2380 | vec_info *vinfo = stmt_info->vinfo; |
2381 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4d694b27 | 2382 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
2383 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
2384 | { | |
2385 | *memory_access_type = VMAT_GATHER_SCATTER; | |
86a91c0a | 2386 | if (!vect_check_gather_scatter (stmt_info, loop_vinfo, gs_info)) |
2de001ee | 2387 | gcc_unreachable (); |
894dd753 | 2388 | else if (!vect_is_simple_use (gs_info->offset, vinfo, |
2de001ee RS |
2389 | &gs_info->offset_dt, |
2390 | &gs_info->offset_vectype)) | |
2391 | { | |
2392 | if (dump_enabled_p ()) | |
2393 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2394 | "%s index use not simple.\n", | |
2395 | vls_type == VLS_LOAD ? "gather" : "scatter"); | |
2396 | return false; | |
2397 | } | |
2398 | } | |
2399 | else if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
2400 | { | |
86a91c0a RS |
2401 | if (!get_group_load_store_type (stmt_info, vectype, slp, masked_p, |
2402 | vls_type, memory_access_type, gs_info)) | |
2de001ee RS |
2403 | return false; |
2404 | } | |
2405 | else if (STMT_VINFO_STRIDED_P (stmt_info)) | |
2406 | { | |
2407 | gcc_assert (!slp); | |
ab2fc782 | 2408 | if (loop_vinfo |
86a91c0a | 2409 | && vect_use_strided_gather_scatters_p (stmt_info, loop_vinfo, |
429ef523 | 2410 | masked_p, gs_info)) |
ab2fc782 RS |
2411 | *memory_access_type = VMAT_GATHER_SCATTER; |
2412 | else | |
2413 | *memory_access_type = VMAT_ELEMENTWISE; | |
2de001ee RS |
2414 | } |
2415 | else | |
62da9e14 | 2416 | { |
86a91c0a | 2417 | int cmp = compare_step_with_zero (stmt_info); |
62da9e14 RS |
2418 | if (cmp < 0) |
2419 | *memory_access_type = get_negative_load_store_type | |
86a91c0a | 2420 | (stmt_info, vectype, vls_type, ncopies); |
62da9e14 RS |
2421 | else if (cmp == 0) |
2422 | { | |
2423 | gcc_assert (vls_type == VLS_LOAD); | |
2424 | *memory_access_type = VMAT_INVARIANT; | |
2425 | } | |
2426 | else | |
2427 | *memory_access_type = VMAT_CONTIGUOUS; | |
2428 | } | |
2de001ee | 2429 | |
4d694b27 RS |
2430 | if ((*memory_access_type == VMAT_ELEMENTWISE |
2431 | || *memory_access_type == VMAT_STRIDED_SLP) | |
2432 | && !nunits.is_constant ()) | |
2433 | { | |
2434 | if (dump_enabled_p ()) | |
2435 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2436 | "Not using elementwise accesses due to variable " | |
2437 | "vectorization factor.\n"); | |
2438 | return false; | |
2439 | } | |
2440 | ||
2de001ee RS |
2441 | /* FIXME: At the moment the cost model seems to underestimate the |
2442 | cost of using elementwise accesses. This check preserves the | |
2443 | traditional behavior until that can be fixed. */ | |
3ad3b3ac RS |
2444 | stmt_vec_info first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
2445 | if (!first_stmt_info) | |
2446 | first_stmt_info = stmt_info; | |
2de001ee | 2447 | if (*memory_access_type == VMAT_ELEMENTWISE |
3ad3b3ac | 2448 | && !STMT_VINFO_STRIDED_P (first_stmt_info) |
bffb8014 | 2449 | && !(stmt_info == DR_GROUP_FIRST_ELEMENT (stmt_info) |
2c53b149 RB |
2450 | && !DR_GROUP_NEXT_ELEMENT (stmt_info) |
2451 | && !pow2p_hwi (DR_GROUP_SIZE (stmt_info)))) | |
2de001ee RS |
2452 | { |
2453 | if (dump_enabled_p ()) | |
2454 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2455 | "not falling back to elementwise accesses\n"); | |
2456 | return false; | |
2457 | } | |
2458 | return true; | |
2459 | } | |
2460 | ||
aaeefd88 | 2461 | /* Return true if boolean argument MASK is suitable for vectorizing |
32e8e429 | 2462 | conditional load or store STMT_INFO. When returning true, store the type |
929b4411 RS |
2463 | of the definition in *MASK_DT_OUT and the type of the vectorized mask |
2464 | in *MASK_VECTYPE_OUT. */ | |
aaeefd88 RS |
2465 | |
2466 | static bool | |
32e8e429 | 2467 | vect_check_load_store_mask (stmt_vec_info stmt_info, tree mask, |
929b4411 RS |
2468 | vect_def_type *mask_dt_out, |
2469 | tree *mask_vectype_out) | |
aaeefd88 RS |
2470 | { |
2471 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (mask))) | |
2472 | { | |
2473 | if (dump_enabled_p ()) | |
2474 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2475 | "mask argument is not a boolean.\n"); | |
2476 | return false; | |
2477 | } | |
2478 | ||
2479 | if (TREE_CODE (mask) != SSA_NAME) | |
2480 | { | |
2481 | if (dump_enabled_p ()) | |
2482 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2483 | "mask argument is not an SSA name.\n"); | |
2484 | return false; | |
2485 | } | |
2486 | ||
929b4411 | 2487 | enum vect_def_type mask_dt; |
aaeefd88 | 2488 | tree mask_vectype; |
894dd753 | 2489 | if (!vect_is_simple_use (mask, stmt_info->vinfo, &mask_dt, &mask_vectype)) |
aaeefd88 RS |
2490 | { |
2491 | if (dump_enabled_p ()) | |
2492 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2493 | "mask use not simple.\n"); | |
2494 | return false; | |
2495 | } | |
2496 | ||
2497 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2498 | if (!mask_vectype) | |
2499 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
2500 | ||
2501 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype)) | |
2502 | { | |
2503 | if (dump_enabled_p ()) | |
2504 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2505 | "could not find an appropriate vector mask type.\n"); | |
2506 | return false; | |
2507 | } | |
2508 | ||
2509 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (mask_vectype), | |
2510 | TYPE_VECTOR_SUBPARTS (vectype))) | |
2511 | { | |
2512 | if (dump_enabled_p ()) | |
3c2a8ed0 DM |
2513 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2514 | "vector mask type %T", | |
2515 | " does not match vector data type %T.\n", | |
2516 | mask_vectype, vectype); | |
2517 | ||
aaeefd88 RS |
2518 | return false; |
2519 | } | |
2520 | ||
929b4411 | 2521 | *mask_dt_out = mask_dt; |
aaeefd88 RS |
2522 | *mask_vectype_out = mask_vectype; |
2523 | return true; | |
2524 | } | |
2525 | ||
3133c3b6 | 2526 | /* Return true if stored value RHS is suitable for vectorizing store |
32e8e429 | 2527 | statement STMT_INFO. When returning true, store the type of the |
929b4411 RS |
2528 | definition in *RHS_DT_OUT, the type of the vectorized store value in |
2529 | *RHS_VECTYPE_OUT and the type of the store in *VLS_TYPE_OUT. */ | |
3133c3b6 RS |
2530 | |
2531 | static bool | |
32e8e429 RS |
2532 | vect_check_store_rhs (stmt_vec_info stmt_info, tree rhs, |
2533 | vect_def_type *rhs_dt_out, tree *rhs_vectype_out, | |
2534 | vec_load_store_type *vls_type_out) | |
3133c3b6 RS |
2535 | { |
2536 | /* In the case this is a store from a constant make sure | |
2537 | native_encode_expr can handle it. */ | |
2538 | if (CONSTANT_CLASS_P (rhs) && native_encode_expr (rhs, NULL, 64) == 0) | |
2539 | { | |
2540 | if (dump_enabled_p ()) | |
2541 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2542 | "cannot encode constant as a byte sequence.\n"); | |
2543 | return false; | |
2544 | } | |
2545 | ||
929b4411 | 2546 | enum vect_def_type rhs_dt; |
3133c3b6 | 2547 | tree rhs_vectype; |
894dd753 | 2548 | if (!vect_is_simple_use (rhs, stmt_info->vinfo, &rhs_dt, &rhs_vectype)) |
3133c3b6 RS |
2549 | { |
2550 | if (dump_enabled_p ()) | |
2551 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2552 | "use not simple.\n"); | |
2553 | return false; | |
2554 | } | |
2555 | ||
2556 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2557 | if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype)) | |
2558 | { | |
2559 | if (dump_enabled_p ()) | |
2560 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2561 | "incompatible vector types.\n"); | |
2562 | return false; | |
2563 | } | |
2564 | ||
929b4411 | 2565 | *rhs_dt_out = rhs_dt; |
3133c3b6 | 2566 | *rhs_vectype_out = rhs_vectype; |
929b4411 | 2567 | if (rhs_dt == vect_constant_def || rhs_dt == vect_external_def) |
3133c3b6 RS |
2568 | *vls_type_out = VLS_STORE_INVARIANT; |
2569 | else | |
2570 | *vls_type_out = VLS_STORE; | |
2571 | return true; | |
2572 | } | |
2573 | ||
82570274 | 2574 | /* Build an all-ones vector mask of type MASKTYPE while vectorizing STMT_INFO. |
bc9587eb RS |
2575 | Note that we support masks with floating-point type, in which case the |
2576 | floats are interpreted as a bitmask. */ | |
2577 | ||
2578 | static tree | |
82570274 | 2579 | vect_build_all_ones_mask (stmt_vec_info stmt_info, tree masktype) |
bc9587eb RS |
2580 | { |
2581 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
2582 | return build_int_cst (masktype, -1); | |
2583 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
2584 | { | |
2585 | tree mask = build_int_cst (TREE_TYPE (masktype), -1); | |
2586 | mask = build_vector_from_val (masktype, mask); | |
82570274 | 2587 | return vect_init_vector (stmt_info, mask, masktype, NULL); |
bc9587eb RS |
2588 | } |
2589 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) | |
2590 | { | |
2591 | REAL_VALUE_TYPE r; | |
2592 | long tmp[6]; | |
2593 | for (int j = 0; j < 6; ++j) | |
2594 | tmp[j] = -1; | |
2595 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
2596 | tree mask = build_real (TREE_TYPE (masktype), r); | |
2597 | mask = build_vector_from_val (masktype, mask); | |
82570274 | 2598 | return vect_init_vector (stmt_info, mask, masktype, NULL); |
bc9587eb RS |
2599 | } |
2600 | gcc_unreachable (); | |
2601 | } | |
2602 | ||
2603 | /* Build an all-zero merge value of type VECTYPE while vectorizing | |
82570274 | 2604 | STMT_INFO as a gather load. */ |
bc9587eb RS |
2605 | |
2606 | static tree | |
82570274 | 2607 | vect_build_zero_merge_argument (stmt_vec_info stmt_info, tree vectype) |
bc9587eb RS |
2608 | { |
2609 | tree merge; | |
2610 | if (TREE_CODE (TREE_TYPE (vectype)) == INTEGER_TYPE) | |
2611 | merge = build_int_cst (TREE_TYPE (vectype), 0); | |
2612 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (vectype))) | |
2613 | { | |
2614 | REAL_VALUE_TYPE r; | |
2615 | long tmp[6]; | |
2616 | for (int j = 0; j < 6; ++j) | |
2617 | tmp[j] = 0; | |
2618 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (vectype))); | |
2619 | merge = build_real (TREE_TYPE (vectype), r); | |
2620 | } | |
2621 | else | |
2622 | gcc_unreachable (); | |
2623 | merge = build_vector_from_val (vectype, merge); | |
82570274 | 2624 | return vect_init_vector (stmt_info, merge, vectype, NULL); |
bc9587eb RS |
2625 | } |
2626 | ||
32e8e429 RS |
2627 | /* Build a gather load call while vectorizing STMT_INFO. Insert new |
2628 | instructions before GSI and add them to VEC_STMT. GS_INFO describes | |
2629 | the gather load operation. If the load is conditional, MASK is the | |
2630 | unvectorized condition and MASK_DT is its definition type, otherwise | |
2631 | MASK is null. */ | |
c48d2d35 RS |
2632 | |
2633 | static void | |
32e8e429 RS |
2634 | vect_build_gather_load_calls (stmt_vec_info stmt_info, |
2635 | gimple_stmt_iterator *gsi, | |
1eede195 | 2636 | stmt_vec_info *vec_stmt, |
32e8e429 | 2637 | gather_scatter_info *gs_info, |
e4057a39 | 2638 | tree mask) |
c48d2d35 | 2639 | { |
c48d2d35 RS |
2640 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
2641 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2642 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2643 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2644 | int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
2645 | edge pe = loop_preheader_edge (loop); | |
2646 | enum { NARROW, NONE, WIDEN } modifier; | |
2647 | poly_uint64 gather_off_nunits | |
2648 | = TYPE_VECTOR_SUBPARTS (gs_info->offset_vectype); | |
2649 | ||
2650 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info->decl)); | |
2651 | tree rettype = TREE_TYPE (TREE_TYPE (gs_info->decl)); | |
2652 | tree srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2653 | tree ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2654 | tree idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2655 | tree masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2656 | tree scaletype = TREE_VALUE (arglist); | |
2657 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
2658 | && (!mask || types_compatible_p (srctype, masktype))); | |
2659 | ||
2660 | tree perm_mask = NULL_TREE; | |
2661 | tree mask_perm_mask = NULL_TREE; | |
2662 | if (known_eq (nunits, gather_off_nunits)) | |
2663 | modifier = NONE; | |
2664 | else if (known_eq (nunits * 2, gather_off_nunits)) | |
2665 | { | |
2666 | modifier = WIDEN; | |
2667 | ||
2668 | /* Currently widening gathers and scatters are only supported for | |
2669 | fixed-length vectors. */ | |
2670 | int count = gather_off_nunits.to_constant (); | |
2671 | vec_perm_builder sel (count, count, 1); | |
2672 | for (int i = 0; i < count; ++i) | |
2673 | sel.quick_push (i | (count / 2)); | |
2674 | ||
2675 | vec_perm_indices indices (sel, 1, count); | |
2676 | perm_mask = vect_gen_perm_mask_checked (gs_info->offset_vectype, | |
2677 | indices); | |
2678 | } | |
2679 | else if (known_eq (nunits, gather_off_nunits * 2)) | |
2680 | { | |
2681 | modifier = NARROW; | |
2682 | ||
2683 | /* Currently narrowing gathers and scatters are only supported for | |
2684 | fixed-length vectors. */ | |
2685 | int count = nunits.to_constant (); | |
2686 | vec_perm_builder sel (count, count, 1); | |
2687 | sel.quick_grow (count); | |
2688 | for (int i = 0; i < count; ++i) | |
2689 | sel[i] = i < count / 2 ? i : i + count / 2; | |
2690 | vec_perm_indices indices (sel, 2, count); | |
2691 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); | |
2692 | ||
2693 | ncopies *= 2; | |
2694 | ||
2695 | if (mask) | |
2696 | { | |
2697 | for (int i = 0; i < count; ++i) | |
2698 | sel[i] = i | (count / 2); | |
2699 | indices.new_vector (sel, 2, count); | |
2700 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, indices); | |
2701 | } | |
2702 | } | |
2703 | else | |
2704 | gcc_unreachable (); | |
2705 | ||
86a91c0a RS |
2706 | tree scalar_dest = gimple_get_lhs (stmt_info->stmt); |
2707 | tree vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
c48d2d35 RS |
2708 | |
2709 | tree ptr = fold_convert (ptrtype, gs_info->base); | |
2710 | if (!is_gimple_min_invariant (ptr)) | |
2711 | { | |
2712 | gimple_seq seq; | |
2713 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
2714 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
2715 | gcc_assert (!new_bb); | |
2716 | } | |
2717 | ||
2718 | tree scale = build_int_cst (scaletype, gs_info->scale); | |
2719 | ||
2720 | tree vec_oprnd0 = NULL_TREE; | |
2721 | tree vec_mask = NULL_TREE; | |
2722 | tree src_op = NULL_TREE; | |
2723 | tree mask_op = NULL_TREE; | |
2724 | tree prev_res = NULL_TREE; | |
2725 | stmt_vec_info prev_stmt_info = NULL; | |
2726 | ||
2727 | if (!mask) | |
2728 | { | |
86a91c0a RS |
2729 | src_op = vect_build_zero_merge_argument (stmt_info, rettype); |
2730 | mask_op = vect_build_all_ones_mask (stmt_info, masktype); | |
c48d2d35 RS |
2731 | } |
2732 | ||
2733 | for (int j = 0; j < ncopies; ++j) | |
2734 | { | |
2735 | tree op, var; | |
c48d2d35 RS |
2736 | if (modifier == WIDEN && (j & 1)) |
2737 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
86a91c0a | 2738 | perm_mask, stmt_info, gsi); |
c48d2d35 RS |
2739 | else if (j == 0) |
2740 | op = vec_oprnd0 | |
86a91c0a | 2741 | = vect_get_vec_def_for_operand (gs_info->offset, stmt_info); |
c48d2d35 | 2742 | else |
e4057a39 RS |
2743 | op = vec_oprnd0 = vect_get_vec_def_for_stmt_copy (loop_vinfo, |
2744 | vec_oprnd0); | |
c48d2d35 RS |
2745 | |
2746 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
2747 | { | |
2748 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), | |
2749 | TYPE_VECTOR_SUBPARTS (idxtype))); | |
2750 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); | |
2751 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); | |
e1bd7296 | 2752 | gassign *new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
86a91c0a | 2753 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
c48d2d35 RS |
2754 | op = var; |
2755 | } | |
2756 | ||
2757 | if (mask) | |
2758 | { | |
2759 | if (mask_perm_mask && (j & 1)) | |
2760 | mask_op = permute_vec_elements (mask_op, mask_op, | |
86a91c0a | 2761 | mask_perm_mask, stmt_info, gsi); |
c48d2d35 RS |
2762 | else |
2763 | { | |
2764 | if (j == 0) | |
86a91c0a | 2765 | vec_mask = vect_get_vec_def_for_operand (mask, stmt_info); |
c48d2d35 | 2766 | else |
e4057a39 RS |
2767 | vec_mask = vect_get_vec_def_for_stmt_copy (loop_vinfo, |
2768 | vec_mask); | |
c48d2d35 RS |
2769 | |
2770 | mask_op = vec_mask; | |
2771 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
2772 | { | |
2773 | gcc_assert | |
2774 | (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)), | |
2775 | TYPE_VECTOR_SUBPARTS (masktype))); | |
2776 | var = vect_get_new_ssa_name (masktype, vect_simple_var); | |
2777 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); | |
e1bd7296 RS |
2778 | gassign *new_stmt |
2779 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_op); | |
86a91c0a | 2780 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
c48d2d35 RS |
2781 | mask_op = var; |
2782 | } | |
2783 | } | |
2784 | src_op = mask_op; | |
2785 | } | |
2786 | ||
e1bd7296 RS |
2787 | gcall *new_call = gimple_build_call (gs_info->decl, 5, src_op, ptr, op, |
2788 | mask_op, scale); | |
c48d2d35 | 2789 | |
e1bd7296 | 2790 | stmt_vec_info new_stmt_info; |
c48d2d35 RS |
2791 | if (!useless_type_conversion_p (vectype, rettype)) |
2792 | { | |
2793 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (vectype), | |
2794 | TYPE_VECTOR_SUBPARTS (rettype))); | |
2795 | op = vect_get_new_ssa_name (rettype, vect_simple_var); | |
e1bd7296 | 2796 | gimple_call_set_lhs (new_call, op); |
86a91c0a | 2797 | vect_finish_stmt_generation (stmt_info, new_call, gsi); |
c48d2d35 RS |
2798 | var = make_ssa_name (vec_dest); |
2799 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); | |
e1bd7296 | 2800 | gassign *new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
86a91c0a RS |
2801 | new_stmt_info |
2802 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); | |
c48d2d35 RS |
2803 | } |
2804 | else | |
2805 | { | |
e1bd7296 RS |
2806 | var = make_ssa_name (vec_dest, new_call); |
2807 | gimple_call_set_lhs (new_call, var); | |
86a91c0a RS |
2808 | new_stmt_info |
2809 | = vect_finish_stmt_generation (stmt_info, new_call, gsi); | |
c48d2d35 RS |
2810 | } |
2811 | ||
c48d2d35 RS |
2812 | if (modifier == NARROW) |
2813 | { | |
2814 | if ((j & 1) == 0) | |
2815 | { | |
2816 | prev_res = var; | |
2817 | continue; | |
2818 | } | |
86a91c0a RS |
2819 | var = permute_vec_elements (prev_res, var, perm_mask, |
2820 | stmt_info, gsi); | |
e1bd7296 | 2821 | new_stmt_info = loop_vinfo->lookup_def (var); |
c48d2d35 RS |
2822 | } |
2823 | ||
ddf98a96 | 2824 | if (prev_stmt_info == NULL) |
e1bd7296 | 2825 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
c48d2d35 | 2826 | else |
e1bd7296 RS |
2827 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
2828 | prev_stmt_info = new_stmt_info; | |
c48d2d35 RS |
2829 | } |
2830 | } | |
2831 | ||
bfaa08b7 RS |
2832 | /* Prepare the base and offset in GS_INFO for vectorization. |
2833 | Set *DATAREF_PTR to the loop-invariant base address and *VEC_OFFSET | |
82570274 RS |
2834 | to the vectorized offset argument for the first copy of STMT_INFO. |
2835 | STMT_INFO is the statement described by GS_INFO and LOOP is the | |
2836 | containing loop. */ | |
bfaa08b7 RS |
2837 | |
2838 | static void | |
82570274 | 2839 | vect_get_gather_scatter_ops (struct loop *loop, stmt_vec_info stmt_info, |
bfaa08b7 RS |
2840 | gather_scatter_info *gs_info, |
2841 | tree *dataref_ptr, tree *vec_offset) | |
2842 | { | |
2843 | gimple_seq stmts = NULL; | |
2844 | *dataref_ptr = force_gimple_operand (gs_info->base, &stmts, true, NULL_TREE); | |
2845 | if (stmts != NULL) | |
2846 | { | |
2847 | basic_block new_bb; | |
2848 | edge pe = loop_preheader_edge (loop); | |
2849 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
2850 | gcc_assert (!new_bb); | |
2851 | } | |
2852 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2853 | tree offset_vectype = get_vectype_for_scalar_type (offset_type); | |
82570274 | 2854 | *vec_offset = vect_get_vec_def_for_operand (gs_info->offset, stmt_info, |
bfaa08b7 RS |
2855 | offset_vectype); |
2856 | } | |
2857 | ||
ab2fc782 RS |
2858 | /* Prepare to implement a grouped or strided load or store using |
2859 | the gather load or scatter store operation described by GS_INFO. | |
32e8e429 | 2860 | STMT_INFO is the load or store statement. |
ab2fc782 RS |
2861 | |
2862 | Set *DATAREF_BUMP to the amount that should be added to the base | |
2863 | address after each copy of the vectorized statement. Set *VEC_OFFSET | |
2864 | to an invariant offset vector in which element I has the value | |
2865 | I * DR_STEP / SCALE. */ | |
2866 | ||
2867 | static void | |
32e8e429 RS |
2868 | vect_get_strided_load_store_ops (stmt_vec_info stmt_info, |
2869 | loop_vec_info loop_vinfo, | |
ab2fc782 RS |
2870 | gather_scatter_info *gs_info, |
2871 | tree *dataref_bump, tree *vec_offset) | |
2872 | { | |
ab2fc782 RS |
2873 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
2874 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2875 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2876 | gimple_seq stmts; | |
2877 | ||
2878 | tree bump = size_binop (MULT_EXPR, | |
2879 | fold_convert (sizetype, DR_STEP (dr)), | |
2880 | size_int (TYPE_VECTOR_SUBPARTS (vectype))); | |
2881 | *dataref_bump = force_gimple_operand (bump, &stmts, true, NULL_TREE); | |
2882 | if (stmts) | |
2883 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
2884 | ||
2885 | /* The offset given in GS_INFO can have pointer type, so use the element | |
2886 | type of the vector instead. */ | |
2887 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2888 | tree offset_vectype = get_vectype_for_scalar_type (offset_type); | |
2889 | offset_type = TREE_TYPE (offset_vectype); | |
2890 | ||
2891 | /* Calculate X = DR_STEP / SCALE and convert it to the appropriate type. */ | |
2892 | tree step = size_binop (EXACT_DIV_EXPR, DR_STEP (dr), | |
2893 | ssize_int (gs_info->scale)); | |
2894 | step = fold_convert (offset_type, step); | |
2895 | step = force_gimple_operand (step, &stmts, true, NULL_TREE); | |
2896 | ||
2897 | /* Create {0, X, X*2, X*3, ...}. */ | |
2898 | *vec_offset = gimple_build (&stmts, VEC_SERIES_EXPR, offset_vectype, | |
2899 | build_zero_cst (offset_type), step); | |
2900 | if (stmts) | |
2901 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
2902 | } | |
2903 | ||
2904 | /* Return the amount that should be added to a vector pointer to move | |
89fa689a | 2905 | to the next or previous copy of AGGR_TYPE. DR_INFO is the data reference |
ab2fc782 RS |
2906 | being vectorized and MEMORY_ACCESS_TYPE describes the type of |
2907 | vectorization. */ | |
2908 | ||
2909 | static tree | |
89fa689a | 2910 | vect_get_data_ptr_increment (dr_vec_info *dr_info, tree aggr_type, |
ab2fc782 RS |
2911 | vect_memory_access_type memory_access_type) |
2912 | { | |
2913 | if (memory_access_type == VMAT_INVARIANT) | |
2914 | return size_zero_node; | |
2915 | ||
2916 | tree iv_step = TYPE_SIZE_UNIT (aggr_type); | |
89fa689a | 2917 | tree step = vect_dr_behavior (dr_info)->step; |
ab2fc782 RS |
2918 | if (tree_int_cst_sgn (step) == -1) |
2919 | iv_step = fold_build1 (NEGATE_EXPR, TREE_TYPE (iv_step), iv_step); | |
2920 | return iv_step; | |
2921 | } | |
2922 | ||
37b14185 RB |
2923 | /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64}. */ |
2924 | ||
2925 | static bool | |
32e8e429 | 2926 | vectorizable_bswap (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
1eede195 | 2927 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
e4057a39 | 2928 | tree vectype_in, stmt_vector_for_cost *cost_vec) |
37b14185 RB |
2929 | { |
2930 | tree op, vectype; | |
32e8e429 | 2931 | gcall *stmt = as_a <gcall *> (stmt_info->stmt); |
e4057a39 | 2932 | vec_info *vinfo = stmt_info->vinfo; |
37b14185 | 2933 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
928686b1 | 2934 | unsigned ncopies; |
37b14185 RB |
2935 | |
2936 | op = gimple_call_arg (stmt, 0); | |
2937 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
8c2f568c | 2938 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
37b14185 RB |
2939 | |
2940 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2941 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2942 | case of SLP. */ | |
2943 | if (slp_node) | |
2944 | ncopies = 1; | |
2945 | else | |
e8f142e2 | 2946 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
37b14185 RB |
2947 | |
2948 | gcc_assert (ncopies >= 1); | |
2949 | ||
2950 | tree char_vectype = get_same_sized_vectype (char_type_node, vectype_in); | |
2951 | if (! char_vectype) | |
2952 | return false; | |
2953 | ||
8c2f568c RS |
2954 | poly_uint64 num_bytes = TYPE_VECTOR_SUBPARTS (char_vectype); |
2955 | unsigned word_bytes; | |
2956 | if (!constant_multiple_p (num_bytes, nunits, &word_bytes)) | |
928686b1 RS |
2957 | return false; |
2958 | ||
d980067b RS |
2959 | /* The encoding uses one stepped pattern for each byte in the word. */ |
2960 | vec_perm_builder elts (num_bytes, word_bytes, 3); | |
2961 | for (unsigned i = 0; i < 3; ++i) | |
37b14185 | 2962 | for (unsigned j = 0; j < word_bytes; ++j) |
908a1a16 | 2963 | elts.quick_push ((i + 1) * word_bytes - j - 1); |
37b14185 | 2964 | |
e3342de4 RS |
2965 | vec_perm_indices indices (elts, 1, num_bytes); |
2966 | if (!can_vec_perm_const_p (TYPE_MODE (char_vectype), indices)) | |
37b14185 RB |
2967 | return false; |
2968 | ||
2969 | if (! vec_stmt) | |
2970 | { | |
2971 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
adac3a68 | 2972 | DUMP_VECT_SCOPE ("vectorizable_bswap"); |
78604de0 | 2973 | if (! slp_node) |
37b14185 | 2974 | { |
68435eb2 RB |
2975 | record_stmt_cost (cost_vec, |
2976 | 1, vector_stmt, stmt_info, 0, vect_prologue); | |
2977 | record_stmt_cost (cost_vec, | |
2978 | ncopies, vec_perm, stmt_info, 0, vect_body); | |
37b14185 RB |
2979 | } |
2980 | return true; | |
2981 | } | |
2982 | ||
736d0f28 | 2983 | tree bswap_vconst = vec_perm_indices_to_tree (char_vectype, indices); |
37b14185 RB |
2984 | |
2985 | /* Transform. */ | |
2986 | vec<tree> vec_oprnds = vNULL; | |
e1bd7296 | 2987 | stmt_vec_info new_stmt_info = NULL; |
37b14185 RB |
2988 | stmt_vec_info prev_stmt_info = NULL; |
2989 | for (unsigned j = 0; j < ncopies; j++) | |
2990 | { | |
2991 | /* Handle uses. */ | |
2992 | if (j == 0) | |
86a91c0a | 2993 | vect_get_vec_defs (op, NULL, stmt_info, &vec_oprnds, NULL, slp_node); |
37b14185 | 2994 | else |
e4057a39 | 2995 | vect_get_vec_defs_for_stmt_copy (vinfo, &vec_oprnds, NULL); |
37b14185 RB |
2996 | |
2997 | /* Arguments are ready. create the new vector stmt. */ | |
2998 | unsigned i; | |
2999 | tree vop; | |
3000 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) | |
3001 | { | |
e1bd7296 | 3002 | gimple *new_stmt; |
37b14185 RB |
3003 | tree tem = make_ssa_name (char_vectype); |
3004 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
3005 | char_vectype, vop)); | |
86a91c0a | 3006 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
37b14185 RB |
3007 | tree tem2 = make_ssa_name (char_vectype); |
3008 | new_stmt = gimple_build_assign (tem2, VEC_PERM_EXPR, | |
3009 | tem, tem, bswap_vconst); | |
86a91c0a | 3010 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
37b14185 RB |
3011 | tem = make_ssa_name (vectype); |
3012 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
3013 | vectype, tem2)); | |
86a91c0a RS |
3014 | new_stmt_info |
3015 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); | |
37b14185 | 3016 | if (slp_node) |
e1bd7296 | 3017 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
37b14185 RB |
3018 | } |
3019 | ||
3020 | if (slp_node) | |
3021 | continue; | |
3022 | ||
3023 | if (j == 0) | |
e1bd7296 | 3024 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
37b14185 | 3025 | else |
e1bd7296 | 3026 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
37b14185 | 3027 | |
e1bd7296 | 3028 | prev_stmt_info = new_stmt_info; |
37b14185 RB |
3029 | } |
3030 | ||
3031 | vec_oprnds.release (); | |
3032 | return true; | |
3033 | } | |
3034 | ||
b1b6836e RS |
3035 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
3036 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT | |
3037 | in a single step. On success, store the binary pack code in | |
3038 | *CONVERT_CODE. */ | |
3039 | ||
3040 | static bool | |
3041 | simple_integer_narrowing (tree vectype_out, tree vectype_in, | |
3042 | tree_code *convert_code) | |
3043 | { | |
3044 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) | |
3045 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) | |
3046 | return false; | |
3047 | ||
3048 | tree_code code; | |
3049 | int multi_step_cvt = 0; | |
3050 | auto_vec <tree, 8> interm_types; | |
3051 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, | |
3052 | &code, &multi_step_cvt, | |
3053 | &interm_types) | |
3054 | || multi_step_cvt) | |
3055 | return false; | |
3056 | ||
3057 | *convert_code = code; | |
3058 | return true; | |
3059 | } | |
5ce9450f | 3060 | |
ebfd146a IR |
3061 | /* Function vectorizable_call. |
3062 | ||
32e8e429 RS |
3063 | Check if STMT_INFO performs a function call that can be vectorized. |
3064 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized | |
3065 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
3066 | Return true if STMT_INFO is vectorizable in this way. */ | |
ebfd146a IR |
3067 | |
3068 | static bool | |
32e8e429 | 3069 | vectorizable_call (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
1eede195 RS |
3070 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
3071 | stmt_vector_for_cost *cost_vec) | |
ebfd146a | 3072 | { |
538dd0b7 | 3073 | gcall *stmt; |
ebfd146a IR |
3074 | tree vec_dest; |
3075 | tree scalar_dest; | |
0267732b | 3076 | tree op; |
ebfd146a | 3077 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
32e8e429 | 3078 | stmt_vec_info prev_stmt_info; |
ebfd146a | 3079 | tree vectype_out, vectype_in; |
c7bda0f4 RS |
3080 | poly_uint64 nunits_in; |
3081 | poly_uint64 nunits_out; | |
ebfd146a | 3082 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
190c2236 | 3083 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 3084 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 3085 | tree fndecl, new_temp, rhs_type; |
2c58d42c RS |
3086 | enum vect_def_type dt[4] |
3087 | = { vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type, | |
3088 | vect_unknown_def_type }; | |
3089 | int ndts = ARRAY_SIZE (dt); | |
ebfd146a | 3090 | int ncopies, j; |
2c58d42c RS |
3091 | auto_vec<tree, 8> vargs; |
3092 | auto_vec<tree, 8> orig_vargs; | |
ebfd146a IR |
3093 | enum { NARROW, NONE, WIDEN } modifier; |
3094 | size_t i, nargs; | |
9d5e7640 | 3095 | tree lhs; |
ebfd146a | 3096 | |
190c2236 | 3097 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3098 | return false; |
3099 | ||
66c16fd9 RB |
3100 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3101 | && ! vec_stmt) | |
ebfd146a IR |
3102 | return false; |
3103 | ||
86a91c0a RS |
3104 | /* Is STMT_INFO a vectorizable call? */ |
3105 | stmt = dyn_cast <gcall *> (stmt_info->stmt); | |
538dd0b7 | 3106 | if (!stmt) |
ebfd146a IR |
3107 | return false; |
3108 | ||
5ce9450f | 3109 | if (gimple_call_internal_p (stmt) |
bfaa08b7 | 3110 | && (internal_load_fn_p (gimple_call_internal_fn (stmt)) |
f307441a | 3111 | || internal_store_fn_p (gimple_call_internal_fn (stmt)))) |
c3a8f964 RS |
3112 | /* Handled by vectorizable_load and vectorizable_store. */ |
3113 | return false; | |
5ce9450f | 3114 | |
0136f8f0 AH |
3115 | if (gimple_call_lhs (stmt) == NULL_TREE |
3116 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
3117 | return false; |
3118 | ||
36bbc05d | 3119 | gcc_checking_assert (!stmt_can_throw_internal (cfun, stmt)); |
5a2c1986 | 3120 | |
b690cc0f RG |
3121 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3122 | ||
ebfd146a IR |
3123 | /* Process function arguments. */ |
3124 | rhs_type = NULL_TREE; | |
b690cc0f | 3125 | vectype_in = NULL_TREE; |
ebfd146a IR |
3126 | nargs = gimple_call_num_args (stmt); |
3127 | ||
1b1562a5 MM |
3128 | /* Bail out if the function has more than three arguments, we do not have |
3129 | interesting builtin functions to vectorize with more than two arguments | |
3130 | except for fma. No arguments is also not good. */ | |
2c58d42c | 3131 | if (nargs == 0 || nargs > 4) |
ebfd146a IR |
3132 | return false; |
3133 | ||
74bf76ed | 3134 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2c58d42c RS |
3135 | combined_fn cfn = gimple_call_combined_fn (stmt); |
3136 | if (cfn == CFN_GOMP_SIMD_LANE) | |
74bf76ed JJ |
3137 | { |
3138 | nargs = 0; | |
3139 | rhs_type = unsigned_type_node; | |
3140 | } | |
3141 | ||
2c58d42c RS |
3142 | int mask_opno = -1; |
3143 | if (internal_fn_p (cfn)) | |
3144 | mask_opno = internal_fn_mask_index (as_internal_fn (cfn)); | |
3145 | ||
ebfd146a IR |
3146 | for (i = 0; i < nargs; i++) |
3147 | { | |
b690cc0f RG |
3148 | tree opvectype; |
3149 | ||
ebfd146a | 3150 | op = gimple_call_arg (stmt, i); |
2c58d42c RS |
3151 | if (!vect_is_simple_use (op, vinfo, &dt[i], &opvectype)) |
3152 | { | |
3153 | if (dump_enabled_p ()) | |
3154 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3155 | "use not simple.\n"); | |
3156 | return false; | |
3157 | } | |
3158 | ||
3159 | /* Skip the mask argument to an internal function. This operand | |
3160 | has been converted via a pattern if necessary. */ | |
3161 | if ((int) i == mask_opno) | |
3162 | continue; | |
ebfd146a IR |
3163 | |
3164 | /* We can only handle calls with arguments of the same type. */ | |
3165 | if (rhs_type | |
8533c9d8 | 3166 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 3167 | { |
73fbfcad | 3168 | if (dump_enabled_p ()) |
78c60e3d | 3169 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3170 | "argument types differ.\n"); |
ebfd146a IR |
3171 | return false; |
3172 | } | |
b690cc0f RG |
3173 | if (!rhs_type) |
3174 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 3175 | |
b690cc0f RG |
3176 | if (!vectype_in) |
3177 | vectype_in = opvectype; | |
3178 | else if (opvectype | |
3179 | && opvectype != vectype_in) | |
3180 | { | |
73fbfcad | 3181 | if (dump_enabled_p ()) |
78c60e3d | 3182 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3183 | "argument vector types differ.\n"); |
b690cc0f RG |
3184 | return false; |
3185 | } | |
3186 | } | |
3187 | /* If all arguments are external or constant defs use a vector type with | |
3188 | the same size as the output vector type. */ | |
ebfd146a | 3189 | if (!vectype_in) |
b690cc0f | 3190 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
3191 | if (vec_stmt) |
3192 | gcc_assert (vectype_in); | |
3193 | if (!vectype_in) | |
3194 | { | |
73fbfcad | 3195 | if (dump_enabled_p ()) |
3c2a8ed0 DM |
3196 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3197 | "no vectype for scalar type %T\n", rhs_type); | |
7d8930a0 IR |
3198 | |
3199 | return false; | |
3200 | } | |
ebfd146a IR |
3201 | |
3202 | /* FORNOW */ | |
b690cc0f RG |
3203 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
3204 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
c7bda0f4 | 3205 | if (known_eq (nunits_in * 2, nunits_out)) |
ebfd146a | 3206 | modifier = NARROW; |
c7bda0f4 | 3207 | else if (known_eq (nunits_out, nunits_in)) |
ebfd146a | 3208 | modifier = NONE; |
c7bda0f4 | 3209 | else if (known_eq (nunits_out * 2, nunits_in)) |
ebfd146a IR |
3210 | modifier = WIDEN; |
3211 | else | |
3212 | return false; | |
3213 | ||
70439f0d RS |
3214 | /* We only handle functions that do not read or clobber memory. */ |
3215 | if (gimple_vuse (stmt)) | |
3216 | { | |
3217 | if (dump_enabled_p ()) | |
3218 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3219 | "function reads from or writes to memory.\n"); | |
3220 | return false; | |
3221 | } | |
3222 | ||
ebfd146a IR |
3223 | /* For now, we only vectorize functions if a target specific builtin |
3224 | is available. TODO -- in some cases, it might be profitable to | |
3225 | insert the calls for pieces of the vector, in order to be able | |
3226 | to vectorize other operations in the loop. */ | |
70439f0d RS |
3227 | fndecl = NULL_TREE; |
3228 | internal_fn ifn = IFN_LAST; | |
70439f0d RS |
3229 | tree callee = gimple_call_fndecl (stmt); |
3230 | ||
3231 | /* First try using an internal function. */ | |
b1b6836e RS |
3232 | tree_code convert_code = ERROR_MARK; |
3233 | if (cfn != CFN_LAST | |
3234 | && (modifier == NONE | |
3235 | || (modifier == NARROW | |
3236 | && simple_integer_narrowing (vectype_out, vectype_in, | |
3237 | &convert_code)))) | |
70439f0d RS |
3238 | ifn = vectorizable_internal_function (cfn, callee, vectype_out, |
3239 | vectype_in); | |
3240 | ||
3241 | /* If that fails, try asking for a target-specific built-in function. */ | |
3242 | if (ifn == IFN_LAST) | |
3243 | { | |
3244 | if (cfn != CFN_LAST) | |
3245 | fndecl = targetm.vectorize.builtin_vectorized_function | |
3246 | (cfn, vectype_out, vectype_in); | |
7672aa9b | 3247 | else if (callee) |
70439f0d RS |
3248 | fndecl = targetm.vectorize.builtin_md_vectorized_function |
3249 | (callee, vectype_out, vectype_in); | |
3250 | } | |
3251 | ||
3252 | if (ifn == IFN_LAST && !fndecl) | |
ebfd146a | 3253 | { |
70439f0d | 3254 | if (cfn == CFN_GOMP_SIMD_LANE |
74bf76ed JJ |
3255 | && !slp_node |
3256 | && loop_vinfo | |
3257 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
3258 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
3259 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
3260 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
3261 | { | |
3262 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
3263 | { 0, 1, 2, ... vf - 1 } vector. */ | |
3264 | gcc_assert (nargs == 0); | |
3265 | } | |
37b14185 RB |
3266 | else if (modifier == NONE |
3267 | && (gimple_call_builtin_p (stmt, BUILT_IN_BSWAP16) | |
3268 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP32) | |
3269 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP64))) | |
86a91c0a | 3270 | return vectorizable_bswap (stmt_info, gsi, vec_stmt, slp_node, |
e4057a39 | 3271 | vectype_in, cost_vec); |
74bf76ed JJ |
3272 | else |
3273 | { | |
3274 | if (dump_enabled_p ()) | |
3275 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 3276 | "function is not vectorizable.\n"); |
74bf76ed JJ |
3277 | return false; |
3278 | } | |
ebfd146a IR |
3279 | } |
3280 | ||
fce57248 | 3281 | if (slp_node) |
190c2236 | 3282 | ncopies = 1; |
b1b6836e | 3283 | else if (modifier == NARROW && ifn == IFN_LAST) |
e8f142e2 | 3284 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
ebfd146a | 3285 | else |
e8f142e2 | 3286 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
ebfd146a IR |
3287 | |
3288 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3289 | needs to be generated. */ | |
3290 | gcc_assert (ncopies >= 1); | |
3291 | ||
ed623edb | 3292 | vec_loop_masks *masks = (loop_vinfo ? &LOOP_VINFO_MASKS (loop_vinfo) : NULL); |
ebfd146a IR |
3293 | if (!vec_stmt) /* transformation not required. */ |
3294 | { | |
3295 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
adac3a68 | 3296 | DUMP_VECT_SCOPE ("vectorizable_call"); |
68435eb2 RB |
3297 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
3298 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) | |
3299 | record_stmt_cost (cost_vec, ncopies / 2, | |
3300 | vec_promote_demote, stmt_info, 0, vect_body); | |
b1b6836e | 3301 | |
2c58d42c RS |
3302 | if (loop_vinfo && mask_opno >= 0) |
3303 | { | |
3304 | unsigned int nvectors = (slp_node | |
3305 | ? SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node) | |
3306 | : ncopies); | |
3307 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype_out); | |
3308 | } | |
ebfd146a IR |
3309 | return true; |
3310 | } | |
3311 | ||
67b8dbac | 3312 | /* Transform. */ |
ebfd146a | 3313 | |
73fbfcad | 3314 | if (dump_enabled_p ()) |
e645e942 | 3315 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
3316 | |
3317 | /* Handle def. */ | |
3318 | scalar_dest = gimple_call_lhs (stmt); | |
3319 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
3320 | ||
2c58d42c RS |
3321 | bool masked_loop_p = loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo); |
3322 | ||
e1bd7296 | 3323 | stmt_vec_info new_stmt_info = NULL; |
ebfd146a | 3324 | prev_stmt_info = NULL; |
b1b6836e | 3325 | if (modifier == NONE || ifn != IFN_LAST) |
ebfd146a | 3326 | { |
b1b6836e | 3327 | tree prev_res = NULL_TREE; |
2c58d42c RS |
3328 | vargs.safe_grow (nargs); |
3329 | orig_vargs.safe_grow (nargs); | |
ebfd146a IR |
3330 | for (j = 0; j < ncopies; ++j) |
3331 | { | |
3332 | /* Build argument list for the vectorized call. */ | |
190c2236 JJ |
3333 | if (slp_node) |
3334 | { | |
ef062b13 | 3335 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 3336 | vec<tree> vec_oprnds0; |
190c2236 JJ |
3337 | |
3338 | for (i = 0; i < nargs; i++) | |
2c58d42c | 3339 | vargs[i] = gimple_call_arg (stmt, i); |
306b0c92 | 3340 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 3341 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
3342 | |
3343 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3344 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
3345 | { |
3346 | size_t k; | |
3347 | for (k = 0; k < nargs; k++) | |
3348 | { | |
37b5ec8f | 3349 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 3350 | vargs[k] = vec_oprndsk[i]; |
190c2236 | 3351 | } |
b1b6836e RS |
3352 | if (modifier == NARROW) |
3353 | { | |
2c58d42c RS |
3354 | /* We don't define any narrowing conditional functions |
3355 | at present. */ | |
3356 | gcc_assert (mask_opno < 0); | |
b1b6836e | 3357 | tree half_res = make_ssa_name (vectype_in); |
a844293d RS |
3358 | gcall *call |
3359 | = gimple_build_call_internal_vec (ifn, vargs); | |
3360 | gimple_call_set_lhs (call, half_res); | |
3361 | gimple_call_set_nothrow (call, true); | |
e1bd7296 | 3362 | new_stmt_info |
86a91c0a | 3363 | = vect_finish_stmt_generation (stmt_info, call, gsi); |
b1b6836e RS |
3364 | if ((i & 1) == 0) |
3365 | { | |
3366 | prev_res = half_res; | |
3367 | continue; | |
3368 | } | |
3369 | new_temp = make_ssa_name (vec_dest); | |
e1bd7296 RS |
3370 | gimple *new_stmt |
3371 | = gimple_build_assign (new_temp, convert_code, | |
3372 | prev_res, half_res); | |
3373 | new_stmt_info | |
86a91c0a RS |
3374 | = vect_finish_stmt_generation (stmt_info, new_stmt, |
3375 | gsi); | |
b1b6836e | 3376 | } |
70439f0d | 3377 | else |
b1b6836e | 3378 | { |
2c58d42c RS |
3379 | if (mask_opno >= 0 && masked_loop_p) |
3380 | { | |
3381 | unsigned int vec_num = vec_oprnds0.length (); | |
3382 | /* Always true for SLP. */ | |
3383 | gcc_assert (ncopies == 1); | |
3384 | tree mask = vect_get_loop_mask (gsi, masks, vec_num, | |
3385 | vectype_out, i); | |
3386 | vargs[mask_opno] = prepare_load_store_mask | |
3387 | (TREE_TYPE (mask), mask, vargs[mask_opno], gsi); | |
3388 | } | |
3389 | ||
a844293d | 3390 | gcall *call; |
b1b6836e | 3391 | if (ifn != IFN_LAST) |
a844293d | 3392 | call = gimple_build_call_internal_vec (ifn, vargs); |
b1b6836e | 3393 | else |
a844293d RS |
3394 | call = gimple_build_call_vec (fndecl, vargs); |
3395 | new_temp = make_ssa_name (vec_dest, call); | |
3396 | gimple_call_set_lhs (call, new_temp); | |
3397 | gimple_call_set_nothrow (call, true); | |
e1bd7296 | 3398 | new_stmt_info |
86a91c0a | 3399 | = vect_finish_stmt_generation (stmt_info, call, gsi); |
b1b6836e | 3400 | } |
e1bd7296 | 3401 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
190c2236 JJ |
3402 | } |
3403 | ||
3404 | for (i = 0; i < nargs; i++) | |
3405 | { | |
37b5ec8f | 3406 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3407 | vec_oprndsi.release (); |
190c2236 | 3408 | } |
190c2236 JJ |
3409 | continue; |
3410 | } | |
3411 | ||
ebfd146a IR |
3412 | for (i = 0; i < nargs; i++) |
3413 | { | |
3414 | op = gimple_call_arg (stmt, i); | |
3415 | if (j == 0) | |
3416 | vec_oprnd0 | |
86a91c0a | 3417 | = vect_get_vec_def_for_operand (op, stmt_info); |
ebfd146a | 3418 | else |
2c58d42c | 3419 | vec_oprnd0 |
e4057a39 | 3420 | = vect_get_vec_def_for_stmt_copy (vinfo, orig_vargs[i]); |
2c58d42c RS |
3421 | |
3422 | orig_vargs[i] = vargs[i] = vec_oprnd0; | |
3423 | } | |
ebfd146a | 3424 | |
2c58d42c RS |
3425 | if (mask_opno >= 0 && masked_loop_p) |
3426 | { | |
3427 | tree mask = vect_get_loop_mask (gsi, masks, ncopies, | |
3428 | vectype_out, j); | |
3429 | vargs[mask_opno] | |
3430 | = prepare_load_store_mask (TREE_TYPE (mask), mask, | |
3431 | vargs[mask_opno], gsi); | |
ebfd146a IR |
3432 | } |
3433 | ||
2c58d42c | 3434 | if (cfn == CFN_GOMP_SIMD_LANE) |
74bf76ed | 3435 | { |
c7bda0f4 | 3436 | tree cst = build_index_vector (vectype_out, j * nunits_out, 1); |
74bf76ed | 3437 | tree new_var |
0e22bb5a | 3438 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 3439 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
86a91c0a | 3440 | vect_init_vector_1 (stmt_info, init_stmt, NULL); |
b731b390 | 3441 | new_temp = make_ssa_name (vec_dest); |
e1bd7296 RS |
3442 | gimple *new_stmt = gimple_build_assign (new_temp, new_var); |
3443 | new_stmt_info | |
86a91c0a | 3444 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
74bf76ed | 3445 | } |
b1b6836e RS |
3446 | else if (modifier == NARROW) |
3447 | { | |
2c58d42c RS |
3448 | /* We don't define any narrowing conditional functions at |
3449 | present. */ | |
3450 | gcc_assert (mask_opno < 0); | |
b1b6836e | 3451 | tree half_res = make_ssa_name (vectype_in); |
a844293d RS |
3452 | gcall *call = gimple_build_call_internal_vec (ifn, vargs); |
3453 | gimple_call_set_lhs (call, half_res); | |
3454 | gimple_call_set_nothrow (call, true); | |
86a91c0a RS |
3455 | new_stmt_info |
3456 | = vect_finish_stmt_generation (stmt_info, call, gsi); | |
b1b6836e RS |
3457 | if ((j & 1) == 0) |
3458 | { | |
3459 | prev_res = half_res; | |
3460 | continue; | |
3461 | } | |
3462 | new_temp = make_ssa_name (vec_dest); | |
e1bd7296 RS |
3463 | gassign *new_stmt = gimple_build_assign (new_temp, convert_code, |
3464 | prev_res, half_res); | |
3465 | new_stmt_info | |
86a91c0a | 3466 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
b1b6836e | 3467 | } |
74bf76ed JJ |
3468 | else |
3469 | { | |
a844293d | 3470 | gcall *call; |
70439f0d | 3471 | if (ifn != IFN_LAST) |
a844293d | 3472 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3473 | else |
a844293d | 3474 | call = gimple_build_call_vec (fndecl, vargs); |
e1bd7296 | 3475 | new_temp = make_ssa_name (vec_dest, call); |
a844293d RS |
3476 | gimple_call_set_lhs (call, new_temp); |
3477 | gimple_call_set_nothrow (call, true); | |
86a91c0a RS |
3478 | new_stmt_info |
3479 | = vect_finish_stmt_generation (stmt_info, call, gsi); | |
74bf76ed | 3480 | } |
ebfd146a | 3481 | |
b1b6836e | 3482 | if (j == (modifier == NARROW ? 1 : 0)) |
e1bd7296 | 3483 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
ebfd146a | 3484 | else |
e1bd7296 | 3485 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
ebfd146a | 3486 | |
e1bd7296 | 3487 | prev_stmt_info = new_stmt_info; |
ebfd146a | 3488 | } |
b1b6836e RS |
3489 | } |
3490 | else if (modifier == NARROW) | |
3491 | { | |
2c58d42c RS |
3492 | /* We don't define any narrowing conditional functions at present. */ |
3493 | gcc_assert (mask_opno < 0); | |
ebfd146a IR |
3494 | for (j = 0; j < ncopies; ++j) |
3495 | { | |
3496 | /* Build argument list for the vectorized call. */ | |
3497 | if (j == 0) | |
9771b263 | 3498 | vargs.create (nargs * 2); |
ebfd146a | 3499 | else |
9771b263 | 3500 | vargs.truncate (0); |
ebfd146a | 3501 | |
190c2236 JJ |
3502 | if (slp_node) |
3503 | { | |
ef062b13 | 3504 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 3505 | vec<tree> vec_oprnds0; |
190c2236 JJ |
3506 | |
3507 | for (i = 0; i < nargs; i++) | |
9771b263 | 3508 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 3509 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 3510 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
3511 | |
3512 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3513 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
3514 | { |
3515 | size_t k; | |
9771b263 | 3516 | vargs.truncate (0); |
190c2236 JJ |
3517 | for (k = 0; k < nargs; k++) |
3518 | { | |
37b5ec8f | 3519 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
3520 | vargs.quick_push (vec_oprndsk[i]); |
3521 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 | 3522 | } |
a844293d | 3523 | gcall *call; |
70439f0d | 3524 | if (ifn != IFN_LAST) |
a844293d | 3525 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3526 | else |
a844293d RS |
3527 | call = gimple_build_call_vec (fndecl, vargs); |
3528 | new_temp = make_ssa_name (vec_dest, call); | |
3529 | gimple_call_set_lhs (call, new_temp); | |
3530 | gimple_call_set_nothrow (call, true); | |
e1bd7296 | 3531 | new_stmt_info |
86a91c0a | 3532 | = vect_finish_stmt_generation (stmt_info, call, gsi); |
e1bd7296 | 3533 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
190c2236 JJ |
3534 | } |
3535 | ||
3536 | for (i = 0; i < nargs; i++) | |
3537 | { | |
37b5ec8f | 3538 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3539 | vec_oprndsi.release (); |
190c2236 | 3540 | } |
190c2236 JJ |
3541 | continue; |
3542 | } | |
3543 | ||
ebfd146a IR |
3544 | for (i = 0; i < nargs; i++) |
3545 | { | |
3546 | op = gimple_call_arg (stmt, i); | |
3547 | if (j == 0) | |
3548 | { | |
3549 | vec_oprnd0 | |
86a91c0a | 3550 | = vect_get_vec_def_for_operand (op, stmt_info); |
ebfd146a | 3551 | vec_oprnd1 |
e4057a39 | 3552 | = vect_get_vec_def_for_stmt_copy (vinfo, vec_oprnd0); |
ebfd146a IR |
3553 | } |
3554 | else | |
3555 | { | |
e1bd7296 RS |
3556 | vec_oprnd1 = gimple_call_arg (new_stmt_info->stmt, |
3557 | 2 * i + 1); | |
ebfd146a | 3558 | vec_oprnd0 |
e4057a39 | 3559 | = vect_get_vec_def_for_stmt_copy (vinfo, vec_oprnd1); |
ebfd146a | 3560 | vec_oprnd1 |
e4057a39 | 3561 | = vect_get_vec_def_for_stmt_copy (vinfo, vec_oprnd0); |
ebfd146a IR |
3562 | } |
3563 | ||
9771b263 DN |
3564 | vargs.quick_push (vec_oprnd0); |
3565 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
3566 | } |
3567 | ||
e1bd7296 | 3568 | gcall *new_stmt = gimple_build_call_vec (fndecl, vargs); |
ebfd146a IR |
3569 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3570 | gimple_call_set_lhs (new_stmt, new_temp); | |
86a91c0a RS |
3571 | new_stmt_info |
3572 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); | |
ebfd146a IR |
3573 | |
3574 | if (j == 0) | |
e1bd7296 | 3575 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt_info; |
ebfd146a | 3576 | else |
e1bd7296 | 3577 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
ebfd146a | 3578 | |
e1bd7296 | 3579 | prev_stmt_info = new_stmt_info; |
ebfd146a IR |
3580 | } |
3581 | ||
3582 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a | 3583 | } |
b1b6836e RS |
3584 | else |
3585 | /* No current target implements this case. */ | |
3586 | return false; | |
ebfd146a | 3587 | |
9771b263 | 3588 | vargs.release (); |
ebfd146a | 3589 | |
ebfd146a IR |
3590 | /* The call in STMT might prevent it from being removed in dce. |
3591 | We however cannot remove it here, due to the way the ssa name | |
3592 | it defines is mapped to the new definition. So just replace | |
3593 | rhs of the statement with something harmless. */ | |
3594 | ||
dd34c087 JJ |
3595 | if (slp_node) |
3596 | return true; | |
3597 | ||
211cd1e2 | 3598 | stmt_info = vect_orig_stmt (stmt_info); |
ed7b8123 | 3599 | lhs = gimple_get_lhs (stmt_info->stmt); |
3cc2fa2a | 3600 | |
e1bd7296 RS |
3601 | gassign *new_stmt |
3602 | = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
9d97912b | 3603 | vinfo->replace_stmt (gsi, stmt_info, new_stmt); |
ebfd146a IR |
3604 | |
3605 | return true; | |
3606 | } | |
3607 | ||
3608 | ||
0136f8f0 AH |
3609 | struct simd_call_arg_info |
3610 | { | |
3611 | tree vectype; | |
3612 | tree op; | |
0136f8f0 | 3613 | HOST_WIDE_INT linear_step; |
34e82342 | 3614 | enum vect_def_type dt; |
0136f8f0 | 3615 | unsigned int align; |
17b658af | 3616 | bool simd_lane_linear; |
0136f8f0 AH |
3617 | }; |
3618 | ||
17b658af JJ |
3619 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
3620 | is linear within simd lane (but not within whole loop), note it in | |
3621 | *ARGINFO. */ | |
3622 | ||
3623 | static void | |
3624 | vect_simd_lane_linear (tree op, struct loop *loop, | |
3625 | struct simd_call_arg_info *arginfo) | |
3626 | { | |
355fe088 | 3627 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
3628 | |
3629 | if (!is_gimple_assign (def_stmt) | |
3630 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
3631 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
3632 | return; | |
3633 | ||
3634 | tree base = gimple_assign_rhs1 (def_stmt); | |
3635 | HOST_WIDE_INT linear_step = 0; | |
3636 | tree v = gimple_assign_rhs2 (def_stmt); | |
3637 | while (TREE_CODE (v) == SSA_NAME) | |
3638 | { | |
3639 | tree t; | |
3640 | def_stmt = SSA_NAME_DEF_STMT (v); | |
3641 | if (is_gimple_assign (def_stmt)) | |
3642 | switch (gimple_assign_rhs_code (def_stmt)) | |
3643 | { | |
3644 | case PLUS_EXPR: | |
3645 | t = gimple_assign_rhs2 (def_stmt); | |
3646 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
3647 | return; | |
3648 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
3649 | v = gimple_assign_rhs1 (def_stmt); | |
3650 | continue; | |
3651 | case MULT_EXPR: | |
3652 | t = gimple_assign_rhs2 (def_stmt); | |
3653 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
3654 | return; | |
3655 | linear_step = tree_to_shwi (t); | |
3656 | v = gimple_assign_rhs1 (def_stmt); | |
3657 | continue; | |
3658 | CASE_CONVERT: | |
3659 | t = gimple_assign_rhs1 (def_stmt); | |
3660 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
3661 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
3662 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
3663 | return; | |
3664 | if (!linear_step) | |
3665 | linear_step = 1; | |
3666 | v = t; | |
3667 | continue; | |
3668 | default: | |
3669 | return; | |
3670 | } | |
8e4284d0 | 3671 | else if (gimple_call_internal_p (def_stmt, IFN_GOMP_SIMD_LANE) |
17b658af JJ |
3672 | && loop->simduid |
3673 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
3674 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
3675 | == loop->simduid)) | |
3676 | { | |
3677 | if (!linear_step) | |
3678 | linear_step = 1; | |
3679 | arginfo->linear_step = linear_step; | |
3680 | arginfo->op = base; | |
3681 | arginfo->simd_lane_linear = true; | |
3682 | return; | |
3683 | } | |
3684 | } | |
3685 | } | |
3686 | ||
cf1b2ba4 RS |
3687 | /* Return the number of elements in vector type VECTYPE, which is associated |
3688 | with a SIMD clone. At present these vectors always have a constant | |
3689 | length. */ | |
3690 | ||
3691 | static unsigned HOST_WIDE_INT | |
3692 | simd_clone_subparts (tree vectype) | |
3693 | { | |
928686b1 | 3694 | return TYPE_VECTOR_SUBPARTS (vectype).to_constant (); |
cf1b2ba4 RS |
3695 | } |
3696 | ||
0136f8f0 AH |
3697 | /* Function vectorizable_simd_clone_call. |
3698 | ||
32e8e429 | 3699 | Check if STMT_INFO performs a function call that can be vectorized |
0136f8f0 | 3700 | by calling a simd clone of the function. |
32e8e429 RS |
3701 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
3702 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
3703 | Return true if STMT_INFO is vectorizable in this way. */ | |
0136f8f0 AH |
3704 | |
3705 | static bool | |
32e8e429 RS |
3706 | vectorizable_simd_clone_call (stmt_vec_info stmt_info, |
3707 | gimple_stmt_iterator *gsi, | |
1eede195 | 3708 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
68435eb2 | 3709 | stmt_vector_for_cost *) |
0136f8f0 AH |
3710 | { |
3711 | tree vec_dest; | |
3712 | tree scalar_dest; | |
3713 | tree op, type; | |
3714 | tree vec_oprnd0 = NULL_TREE; | |
32e8e429 | 3715 | stmt_vec_info prev_stmt_info; |
0136f8f0 AH |
3716 | tree vectype; |
3717 | unsigned int nunits; | |
3718 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3719 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 3720 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 3721 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 3722 | tree fndecl, new_temp; |
0136f8f0 | 3723 | int ncopies, j; |
00426f9a | 3724 | auto_vec<simd_call_arg_info> arginfo; |
0136f8f0 AH |
3725 | vec<tree> vargs = vNULL; |
3726 | size_t i, nargs; | |
3727 | tree lhs, rtype, ratype; | |
e7a74006 | 3728 | vec<constructor_elt, va_gc> *ret_ctor_elts = NULL; |
0136f8f0 AH |
3729 | |
3730 | /* Is STMT a vectorizable call? */ | |
32e8e429 RS |
3731 | gcall *stmt = dyn_cast <gcall *> (stmt_info->stmt); |
3732 | if (!stmt) | |
0136f8f0 AH |
3733 | return false; |
3734 | ||
3735 | fndecl = gimple_call_fndecl (stmt); | |
3736 | if (fndecl == NULL_TREE) | |
3737 | return false; | |
3738 | ||
d52f5295 | 3739 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
3740 | if (node == NULL || node->simd_clones == NULL) |
3741 | return false; | |
3742 | ||
3743 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
3744 | return false; | |
3745 | ||
66c16fd9 RB |
3746 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3747 | && ! vec_stmt) | |
0136f8f0 AH |
3748 | return false; |
3749 | ||
3750 | if (gimple_call_lhs (stmt) | |
3751 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
3752 | return false; | |
3753 | ||
36bbc05d | 3754 | gcc_checking_assert (!stmt_can_throw_internal (cfun, stmt)); |
0136f8f0 AH |
3755 | |
3756 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3757 | ||
86a91c0a | 3758 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt_info)) |
0136f8f0 AH |
3759 | return false; |
3760 | ||
3761 | /* FORNOW */ | |
fce57248 | 3762 | if (slp_node) |
0136f8f0 AH |
3763 | return false; |
3764 | ||
3765 | /* Process function arguments. */ | |
3766 | nargs = gimple_call_num_args (stmt); | |
3767 | ||
3768 | /* Bail out if the function has zero arguments. */ | |
3769 | if (nargs == 0) | |
3770 | return false; | |
3771 | ||
00426f9a | 3772 | arginfo.reserve (nargs, true); |
0136f8f0 AH |
3773 | |
3774 | for (i = 0; i < nargs; i++) | |
3775 | { | |
3776 | simd_call_arg_info thisarginfo; | |
3777 | affine_iv iv; | |
3778 | ||
3779 | thisarginfo.linear_step = 0; | |
3780 | thisarginfo.align = 0; | |
3781 | thisarginfo.op = NULL_TREE; | |
17b658af | 3782 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
3783 | |
3784 | op = gimple_call_arg (stmt, i); | |
894dd753 | 3785 | if (!vect_is_simple_use (op, vinfo, &thisarginfo.dt, |
81c40241 | 3786 | &thisarginfo.vectype) |
0136f8f0 AH |
3787 | || thisarginfo.dt == vect_uninitialized_def) |
3788 | { | |
3789 | if (dump_enabled_p ()) | |
3790 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3791 | "use not simple.\n"); | |
0136f8f0 AH |
3792 | return false; |
3793 | } | |
3794 | ||
3795 | if (thisarginfo.dt == vect_constant_def | |
3796 | || thisarginfo.dt == vect_external_def) | |
3797 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
3798 | else | |
3799 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
3800 | ||
6c9e85fb JJ |
3801 | /* For linear arguments, the analyze phase should have saved |
3802 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
3803 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
3804 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
3805 | { |
3806 | gcc_assert (vec_stmt); | |
3807 | thisarginfo.linear_step | |
17b658af | 3808 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 3809 | thisarginfo.op |
17b658af JJ |
3810 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
3811 | thisarginfo.simd_lane_linear | |
3812 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
3813 | == boolean_true_node); | |
6c9e85fb JJ |
3814 | /* If loop has been peeled for alignment, we need to adjust it. */ |
3815 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
3816 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 3817 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
3818 | { |
3819 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 3820 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
3821 | tree opt = TREE_TYPE (thisarginfo.op); |
3822 | bias = fold_convert (TREE_TYPE (step), bias); | |
3823 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
3824 | thisarginfo.op | |
3825 | = fold_build2 (POINTER_TYPE_P (opt) | |
3826 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
3827 | thisarginfo.op, bias); | |
3828 | } | |
3829 | } | |
3830 | else if (!vec_stmt | |
3831 | && thisarginfo.dt != vect_constant_def | |
3832 | && thisarginfo.dt != vect_external_def | |
3833 | && loop_vinfo | |
3834 | && TREE_CODE (op) == SSA_NAME | |
3835 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
3836 | &iv, false) | |
3837 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
3838 | { |
3839 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
3840 | thisarginfo.op = iv.base; | |
3841 | } | |
3842 | else if ((thisarginfo.dt == vect_constant_def | |
3843 | || thisarginfo.dt == vect_external_def) | |
3844 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
3845 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
3846 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
3847 | linear too. */ | |
3848 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
3849 | && !thisarginfo.linear_step | |
3850 | && !vec_stmt | |
3851 | && thisarginfo.dt != vect_constant_def | |
3852 | && thisarginfo.dt != vect_external_def | |
3853 | && loop_vinfo | |
3854 | && !slp_node | |
3855 | && TREE_CODE (op) == SSA_NAME) | |
3856 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
3857 | |
3858 | arginfo.quick_push (thisarginfo); | |
3859 | } | |
3860 | ||
d9f21f6a RS |
3861 | unsigned HOST_WIDE_INT vf; |
3862 | if (!LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&vf)) | |
3863 | { | |
3864 | if (dump_enabled_p ()) | |
3865 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3866 | "not considering SIMD clones; not yet supported" | |
3867 | " for variable-width vectors.\n"); | |
888157af | 3868 | return false; |
d9f21f6a RS |
3869 | } |
3870 | ||
0136f8f0 AH |
3871 | unsigned int badness = 0; |
3872 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
3873 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
3874 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
3875 | else |
3876 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
3877 | n = n->simdclone->next_clone) | |
3878 | { | |
3879 | unsigned int this_badness = 0; | |
d9f21f6a | 3880 | if (n->simdclone->simdlen > vf |
0136f8f0 AH |
3881 | || n->simdclone->nargs != nargs) |
3882 | continue; | |
d9f21f6a RS |
3883 | if (n->simdclone->simdlen < vf) |
3884 | this_badness += (exact_log2 (vf) | |
0136f8f0 AH |
3885 | - exact_log2 (n->simdclone->simdlen)) * 1024; |
3886 | if (n->simdclone->inbranch) | |
3887 | this_badness += 2048; | |
3888 | int target_badness = targetm.simd_clone.usable (n); | |
3889 | if (target_badness < 0) | |
3890 | continue; | |
3891 | this_badness += target_badness * 512; | |
3892 | /* FORNOW: Have to add code to add the mask argument. */ | |
3893 | if (n->simdclone->inbranch) | |
3894 | continue; | |
3895 | for (i = 0; i < nargs; i++) | |
3896 | { | |
3897 | switch (n->simdclone->args[i].arg_type) | |
3898 | { | |
3899 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3900 | if (!useless_type_conversion_p | |
3901 | (n->simdclone->args[i].orig_type, | |
3902 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
3903 | i = -1; | |
3904 | else if (arginfo[i].dt == vect_constant_def | |
3905 | || arginfo[i].dt == vect_external_def | |
3906 | || arginfo[i].linear_step) | |
3907 | this_badness += 64; | |
3908 | break; | |
3909 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3910 | if (arginfo[i].dt != vect_constant_def | |
3911 | && arginfo[i].dt != vect_external_def) | |
3912 | i = -1; | |
3913 | break; | |
3914 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 3915 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
3916 | if (arginfo[i].dt == vect_constant_def |
3917 | || arginfo[i].dt == vect_external_def | |
3918 | || (arginfo[i].linear_step | |
3919 | != n->simdclone->args[i].linear_step)) | |
3920 | i = -1; | |
3921 | break; | |
3922 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
3923 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
3924 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
3925 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3926 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3927 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3928 | /* FORNOW */ |
3929 | i = -1; | |
3930 | break; | |
3931 | case SIMD_CLONE_ARG_TYPE_MASK: | |
3932 | gcc_unreachable (); | |
3933 | } | |
3934 | if (i == (size_t) -1) | |
3935 | break; | |
3936 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
3937 | { | |
3938 | i = -1; | |
3939 | break; | |
3940 | } | |
3941 | if (arginfo[i].align) | |
3942 | this_badness += (exact_log2 (arginfo[i].align) | |
3943 | - exact_log2 (n->simdclone->args[i].alignment)); | |
3944 | } | |
3945 | if (i == (size_t) -1) | |
3946 | continue; | |
3947 | if (bestn == NULL || this_badness < badness) | |
3948 | { | |
3949 | bestn = n; | |
3950 | badness = this_badness; | |
3951 | } | |
3952 | } | |
3953 | ||
3954 | if (bestn == NULL) | |
00426f9a | 3955 | return false; |
0136f8f0 AH |
3956 | |
3957 | for (i = 0; i < nargs; i++) | |
3958 | if ((arginfo[i].dt == vect_constant_def | |
3959 | || arginfo[i].dt == vect_external_def) | |
3960 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
3961 | { | |
3962 | arginfo[i].vectype | |
3963 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
3964 | i))); | |
3965 | if (arginfo[i].vectype == NULL | |
cf1b2ba4 | 3966 | || (simd_clone_subparts (arginfo[i].vectype) |
0136f8f0 | 3967 | > bestn->simdclone->simdlen)) |
00426f9a | 3968 | return false; |
0136f8f0 AH |
3969 | } |
3970 | ||
3971 | fndecl = bestn->decl; | |
3972 | nunits = bestn->simdclone->simdlen; | |
d9f21f6a | 3973 | ncopies = vf / nunits; |
0136f8f0 AH |
3974 | |
3975 | /* If the function isn't const, only allow it in simd loops where user | |
3976 | has asserted that at least nunits consecutive iterations can be | |
3977 | performed using SIMD instructions. */ | |
3978 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
3979 | && gimple_vuse (stmt)) | |
00426f9a | 3980 | return false; |
0136f8f0 AH |
3981 | |
3982 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3983 | needs to be generated. */ | |
3984 | gcc_assert (ncopies >= 1); | |
3985 | ||
3986 | if (!vec_stmt) /* transformation not required. */ | |
3987 | { | |
6c9e85fb JJ |
3988 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
3989 | for (i = 0; i < nargs; i++) | |
7adb26f2 JJ |
3990 | if ((bestn->simdclone->args[i].arg_type |
3991 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
3992 | || (bestn->simdclone->args[i].arg_type | |
3993 | == SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP)) | |
6c9e85fb | 3994 | { |
17b658af | 3995 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
3996 | + 1); |
3997 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
3998 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
3999 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
4000 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
4001 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
4002 | tree sll = arginfo[i].simd_lane_linear |
4003 | ? boolean_true_node : boolean_false_node; | |
4004 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 4005 | } |
0136f8f0 | 4006 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
adac3a68 | 4007 | DUMP_VECT_SCOPE ("vectorizable_simd_clone_call"); |
68435eb2 | 4008 | /* vect_model_simple_cost (stmt_info, ncopies, dt, slp_node, cost_vec); */ |
0136f8f0 AH |
4009 | return true; |
4010 | } | |
4011 | ||
67b8dbac | 4012 | /* Transform. */ |
0136f8f0 AH |
4013 | |
4014 | if (dump_enabled_p ()) | |
4015 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
4016 | ||
4017 | /* Handle def. */ | |
4018 | scalar_dest = gimple_call_lhs (stmt); | |
4019 | vec_dest = NULL_TREE; | |
4020 | rtype = NULL_TREE; | |
4021 | ratype = NULL_TREE; | |
4022 | if (scalar_dest) | |
4023 | { | |
4024 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4025 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
4026 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
4027 | { | |
4028 | ratype = rtype; | |
4029 | rtype = TREE_TYPE (ratype); | |
4030 | } | |
4031 | } | |
4032 | ||
4033 | prev_stmt_info = NULL; | |
4034 | for (j = 0; j < ncopies; ++j) | |
4035 | { | |
4036 | /* Build argument list for the vectorized call. */ | |
4037 | if (j == 0) | |
4038 | vargs.create (nargs); | |
4039 | else | |
4040 | vargs.truncate (0); | |
4041 | ||
4042 | for (i = 0; i < nargs; i++) | |
4043 | { | |
4044 | unsigned int k, l, m, o; | |
4045 | tree atype; | |
4046 | op = gimple_call_arg (stmt, i); | |
4047 | switch (bestn->simdclone->args[i].arg_type) | |
4048 | { | |
4049 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
4050 | atype = bestn->simdclone->args[i].vector_type; | |
cf1b2ba4 | 4051 | o = nunits / simd_clone_subparts (atype); |
0136f8f0 AH |
4052 | for (m = j * o; m < (j + 1) * o; m++) |
4053 | { | |
cf1b2ba4 RS |
4054 | if (simd_clone_subparts (atype) |
4055 | < simd_clone_subparts (arginfo[i].vectype)) | |
0136f8f0 | 4056 | { |
73a699ae | 4057 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); |
cf1b2ba4 RS |
4058 | k = (simd_clone_subparts (arginfo[i].vectype) |
4059 | / simd_clone_subparts (atype)); | |
0136f8f0 AH |
4060 | gcc_assert ((k & (k - 1)) == 0); |
4061 | if (m == 0) | |
4062 | vec_oprnd0 | |
86a91c0a | 4063 | = vect_get_vec_def_for_operand (op, stmt_info); |
0136f8f0 AH |
4064 | else |
4065 | { | |
4066 | vec_oprnd0 = arginfo[i].op; | |
4067 | if ((m & (k - 1)) == 0) | |
4068 | vec_oprnd0 | |
e4057a39 | 4069 | = vect_get_vec_def_for_stmt_copy (vinfo, |
0136f8f0 AH |
4070 | vec_oprnd0); |
4071 | } | |
4072 | arginfo[i].op = vec_oprnd0; | |
4073 | vec_oprnd0 | |
4074 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
92e29a5e | 4075 | bitsize_int (prec), |
0136f8f0 | 4076 | bitsize_int ((m & (k - 1)) * prec)); |
e1bd7296 | 4077 | gassign *new_stmt |
b731b390 | 4078 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 | 4079 | vec_oprnd0); |
86a91c0a | 4080 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
0136f8f0 AH |
4081 | vargs.safe_push (gimple_assign_lhs (new_stmt)); |
4082 | } | |
4083 | else | |
4084 | { | |
cf1b2ba4 RS |
4085 | k = (simd_clone_subparts (atype) |
4086 | / simd_clone_subparts (arginfo[i].vectype)); | |
0136f8f0 AH |
4087 | gcc_assert ((k & (k - 1)) == 0); |
4088 | vec<constructor_elt, va_gc> *ctor_elts; | |
4089 | if (k != 1) | |
4090 | vec_alloc (ctor_elts, k); | |
4091 | else | |
4092 | ctor_elts = NULL; | |
4093 | for (l = 0; l < k; l++) | |
4094 | { | |
4095 | if (m == 0 && l == 0) | |
4096 | vec_oprnd0 | |
86a91c0a | 4097 | = vect_get_vec_def_for_operand (op, stmt_info); |
0136f8f0 AH |
4098 | else |
4099 | vec_oprnd0 | |
e4057a39 | 4100 | = vect_get_vec_def_for_stmt_copy (vinfo, |
0136f8f0 AH |
4101 | arginfo[i].op); |
4102 | arginfo[i].op = vec_oprnd0; | |
4103 | if (k == 1) | |
4104 | break; | |
4105 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
4106 | vec_oprnd0); | |
4107 | } | |
4108 | if (k == 1) | |
4109 | vargs.safe_push (vec_oprnd0); | |
4110 | else | |
4111 | { | |
4112 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
e1bd7296 | 4113 | gassign *new_stmt |
b731b390 | 4114 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 | 4115 | vec_oprnd0); |
86a91c0a RS |
4116 | vect_finish_stmt_generation (stmt_info, new_stmt, |
4117 | gsi); | |
0136f8f0 AH |
4118 | vargs.safe_push (gimple_assign_lhs (new_stmt)); |
4119 | } | |
4120 | } | |
4121 | } | |
4122 | break; | |
4123 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
4124 | vargs.safe_push (op); | |
4125 | break; | |
4126 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
7adb26f2 | 4127 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
4128 | if (j == 0) |
4129 | { | |
4130 | gimple_seq stmts; | |
4131 | arginfo[i].op | |
4132 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
4133 | NULL_TREE); | |
4134 | if (stmts != NULL) | |
4135 | { | |
4136 | basic_block new_bb; | |
4137 | edge pe = loop_preheader_edge (loop); | |
4138 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
4139 | gcc_assert (!new_bb); | |
4140 | } | |
17b658af JJ |
4141 | if (arginfo[i].simd_lane_linear) |
4142 | { | |
4143 | vargs.safe_push (arginfo[i].op); | |
4144 | break; | |
4145 | } | |
b731b390 | 4146 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 4147 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
4fbeb363 | 4148 | loop_vinfo->add_stmt (new_phi); |
0136f8f0 AH |
4149 | add_phi_arg (new_phi, arginfo[i].op, |
4150 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
4151 | enum tree_code code | |
4152 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
4153 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
4154 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
4155 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
4156 | widest_int cst |
4157 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
4158 | ncopies * nunits); | |
4159 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 4160 | tree phi_arg = copy_ssa_name (op); |
e1bd7296 | 4161 | gassign *new_stmt |
0d0e4a03 | 4162 | = gimple_build_assign (phi_arg, code, phi_res, tcst); |
0136f8f0 AH |
4163 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
4164 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
4fbeb363 | 4165 | loop_vinfo->add_stmt (new_stmt); |
0136f8f0 AH |
4166 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
4167 | UNKNOWN_LOCATION); | |
4168 | arginfo[i].op = phi_res; | |
4169 | vargs.safe_push (phi_res); | |
4170 | } | |
4171 | else | |
4172 | { | |
4173 | enum tree_code code | |
4174 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
4175 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
4176 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
4177 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
4178 | widest_int cst |
4179 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
4180 | j * nunits); | |
4181 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 4182 | new_temp = make_ssa_name (TREE_TYPE (op)); |
e1bd7296 RS |
4183 | gassign *new_stmt |
4184 | = gimple_build_assign (new_temp, code, | |
4185 | arginfo[i].op, tcst); | |
86a91c0a | 4186 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
0136f8f0 AH |
4187 | vargs.safe_push (new_temp); |
4188 | } | |
4189 | break; | |
7adb26f2 JJ |
4190 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
4191 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
0136f8f0 | 4192 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
e01d41e5 JJ |
4193 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
4194 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
4195 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
4196 | default: |
4197 | gcc_unreachable (); | |
4198 | } | |
4199 | } | |
4200 | ||
e1bd7296 | 4201 | gcall *new_call = gimple_build_call_vec (fndecl, vargs); |
0136f8f0 AH |
4202 | if (vec_dest) |
4203 | { | |
cf1b2ba4 | 4204 | gcc_assert (ratype || simd_clone_subparts (rtype) == nunits); |
0136f8f0 | 4205 | if (ratype) |
b731b390 | 4206 | new_temp = create_tmp_var (ratype); |
cf1b2ba4 RS |
4207 | else if (simd_clone_subparts (vectype) |
4208 | == simd_clone_subparts (rtype)) | |
e1bd7296 | 4209 | new_temp = make_ssa_name (vec_dest, new_call); |
0136f8f0 | 4210 | else |
e1bd7296 RS |
4211 | new_temp = make_ssa_name (rtype, new_call); |
4212 | gimple_call_set_lhs (new_call, new_temp); | |
0136f8f0 | 4213 | } |
e1bd7296 | 4214 | stmt_vec_info new_stmt_info |
86a91c0a | 4215 | = vect_finish_stmt_generation (stmt_info, new_call, gsi); |
0136f8f0 AH |
4216 | |
4217 | if (vec_dest) | |
4218 | { | |
cf1b2ba4 | 4219 | if (simd_clone_subparts (vectype) < nunits) |
0136f8f0 AH |
4220 | { |
4221 | unsigned int k, l; | |
73a699ae RS |
4222 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); |
4223 | poly_uint64 bytes = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
cf1b2ba4 | 4224 | k = nunits / simd_clone_subparts (vectype); |
0136f8f0 AH |
4225 | gcc_assert ((k & (k - 1)) == 0); |
4226 | for (l = 0; l < k; l++) | |
4227 | { | |
4228 | tree t; | |
4229 | if (ratype) | |
4230 | { | |
4231 | t = build_fold_addr_expr (new_temp); | |
4232 | t = build2 (MEM_REF, vectype, t, | |
73a699ae | 4233 | build_int_cst (TREE_TYPE (t), l * bytes)); |
0136f8f0 AH |
4234 | } |
4235 | else | |
4236 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
92e29a5e | 4237 | bitsize_int (prec), bitsize_int (l * prec)); |
e1bd7296 | 4238 | gimple *new_stmt |
b731b390 | 4239 | = gimple_build_assign (make_ssa_name (vectype), t); |
e1bd7296 | 4240 | new_stmt_info |
86a91c0a | 4241 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
e1bd7296 | 4242 | |
0136f8f0 | 4243 | if (j == 0 && l == 0) |
e1bd7296 RS |
4244 | STMT_VINFO_VEC_STMT (stmt_info) |
4245 | = *vec_stmt = new_stmt_info; | |
0136f8f0 | 4246 | else |
e1bd7296 | 4247 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
0136f8f0 | 4248 | |
e1bd7296 | 4249 | prev_stmt_info = new_stmt_info; |
0136f8f0 AH |
4250 | } |
4251 | ||
4252 | if (ratype) | |
86a91c0a | 4253 | vect_clobber_variable (stmt_info, gsi, new_temp); |
0136f8f0 AH |
4254 | continue; |
4255 | } | |
cf1b2ba4 | 4256 | else if (simd_clone_subparts (vectype) > nunits) |
0136f8f0 | 4257 | { |
cf1b2ba4 RS |
4258 | unsigned int k = (simd_clone_subparts (vectype) |
4259 | / simd_clone_subparts (rtype)); | |
0136f8f0 AH |
4260 | gcc_assert ((k & (k - 1)) == 0); |
4261 | if ((j & (k - 1)) == 0) | |
4262 | vec_alloc (ret_ctor_elts, k); | |
4263 | if (ratype) | |
4264 | { | |
cf1b2ba4 | 4265 | unsigned int m, o = nunits / simd_clone_subparts (rtype); |
0136f8f0 AH |
4266 | for (m = 0; m < o; m++) |
4267 | { | |
4268 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
4269 | size_int (m), NULL_TREE, NULL_TREE); | |
e1bd7296 | 4270 | gimple *new_stmt |
b731b390 | 4271 | = gimple_build_assign (make_ssa_name (rtype), tem); |
e1bd7296 | 4272 | new_stmt_info |
86a91c0a RS |
4273 | = vect_finish_stmt_generation (stmt_info, new_stmt, |
4274 | gsi); | |
0136f8f0 AH |
4275 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, |
4276 | gimple_assign_lhs (new_stmt)); | |
4277 | } | |
86a91c0a | 4278 | vect_clobber_variable (stmt_info, gsi, new_temp); |
0136f8f0 AH |
4279 | } |
4280 | else | |
4281 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
4282 | if ((j & (k - 1)) != k - 1) | |
4283 | continue; | |
4284 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
e1bd7296 | 4285 | gimple *new_stmt |
b731b390 | 4286 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
e1bd7296 | 4287 | new_stmt_info |
86a91c0a | 4288 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
0136f8f0 AH |
4289 | |
4290 | if ((unsigned) j == k - 1) | |
e1bd7296 | 4291 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
0136f8f0 | 4292 | else |
e1bd7296 | 4293 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
0136f8f0 | 4294 | |
e1bd7296 | 4295 | prev_stmt_info = new_stmt_info; |
0136f8f0 AH |
4296 | continue; |
4297 | } | |
4298 | else if (ratype) | |
4299 | { | |
4300 | tree t = build_fold_addr_expr (new_temp); | |
4301 | t = build2 (MEM_REF, vectype, t, | |
4302 | build_int_cst (TREE_TYPE (t), 0)); | |
e1bd7296 | 4303 | gimple *new_stmt |
b731b390 | 4304 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
e1bd7296 | 4305 | new_stmt_info |
86a91c0a RS |
4306 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
4307 | vect_clobber_variable (stmt_info, gsi, new_temp); | |
0136f8f0 AH |
4308 | } |
4309 | } | |
4310 | ||
4311 | if (j == 0) | |
e1bd7296 | 4312 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
0136f8f0 | 4313 | else |
e1bd7296 | 4314 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
0136f8f0 | 4315 | |
e1bd7296 | 4316 | prev_stmt_info = new_stmt_info; |
0136f8f0 AH |
4317 | } |
4318 | ||
4319 | vargs.release (); | |
4320 | ||
4321 | /* The call in STMT might prevent it from being removed in dce. | |
4322 | We however cannot remove it here, due to the way the ssa name | |
4323 | it defines is mapped to the new definition. So just replace | |
4324 | rhs of the statement with something harmless. */ | |
4325 | ||
4326 | if (slp_node) | |
4327 | return true; | |
4328 | ||
e1bd7296 | 4329 | gimple *new_stmt; |
0136f8f0 AH |
4330 | if (scalar_dest) |
4331 | { | |
4332 | type = TREE_TYPE (scalar_dest); | |
211cd1e2 | 4333 | lhs = gimple_call_lhs (vect_orig_stmt (stmt_info)->stmt); |
0136f8f0 AH |
4334 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
4335 | } | |
4336 | else | |
4337 | new_stmt = gimple_build_nop (); | |
41b6b80e | 4338 | vinfo->replace_stmt (gsi, vect_orig_stmt (stmt_info), new_stmt); |
0136f8f0 AH |
4339 | unlink_stmt_vdef (stmt); |
4340 | ||
4341 | return true; | |
4342 | } | |
4343 | ||
4344 | ||
ebfd146a IR |
4345 | /* Function vect_gen_widened_results_half |
4346 | ||
4347 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 4348 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 4349 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
4350 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
4351 | needs to be created (DECL is a function-decl of a target-builtin). | |
82570274 | 4352 | STMT_INFO is the original scalar stmt that we are vectorizing. */ |
ebfd146a | 4353 | |
355fe088 | 4354 | static gimple * |
ebfd146a IR |
4355 | vect_gen_widened_results_half (enum tree_code code, |
4356 | tree decl, | |
4357 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
4358 | tree vec_dest, gimple_stmt_iterator *gsi, | |
82570274 | 4359 | stmt_vec_info stmt_info) |
b8698a0f | 4360 | { |
355fe088 | 4361 | gimple *new_stmt; |
b8698a0f L |
4362 | tree new_temp; |
4363 | ||
4364 | /* Generate half of the widened result: */ | |
4365 | if (code == CALL_EXPR) | |
4366 | { | |
4367 | /* Target specific support */ | |
ebfd146a IR |
4368 | if (op_type == binary_op) |
4369 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
4370 | else | |
4371 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
4372 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4373 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
4374 | } |
4375 | else | |
ebfd146a | 4376 | { |
b8698a0f L |
4377 | /* Generic support */ |
4378 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
4379 | if (op_type != binary_op) |
4380 | vec_oprnd1 = NULL; | |
0d0e4a03 | 4381 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
4382 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4383 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 4384 | } |
82570274 | 4385 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
ebfd146a | 4386 | |
ebfd146a IR |
4387 | return new_stmt; |
4388 | } | |
4389 | ||
4a00c761 | 4390 | |
82570274 RS |
4391 | /* Get vectorized definitions for loop-based vectorization of STMT_INFO. |
4392 | For the first operand we call vect_get_vec_def_for_operand (with OPRND | |
4393 | containing scalar operand), and for the rest we get a copy with | |
4a00c761 JJ |
4394 | vect_get_vec_def_for_stmt_copy() using the previous vector definition |
4395 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
4396 | The vectors are collected into VEC_OPRNDS. */ | |
4397 | ||
4398 | static void | |
82570274 | 4399 | vect_get_loop_based_defs (tree *oprnd, stmt_vec_info stmt_info, |
e4057a39 | 4400 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 | 4401 | { |
e4057a39 | 4402 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 JJ |
4403 | tree vec_oprnd; |
4404 | ||
4405 | /* Get first vector operand. */ | |
4406 | /* All the vector operands except the very first one (that is scalar oprnd) | |
4407 | are stmt copies. */ | |
4408 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
82570274 | 4409 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt_info); |
4a00c761 | 4410 | else |
e4057a39 | 4411 | vec_oprnd = vect_get_vec_def_for_stmt_copy (vinfo, *oprnd); |
4a00c761 | 4412 | |
9771b263 | 4413 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
4414 | |
4415 | /* Get second vector operand. */ | |
e4057a39 | 4416 | vec_oprnd = vect_get_vec_def_for_stmt_copy (vinfo, vec_oprnd); |
9771b263 | 4417 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
4418 | |
4419 | *oprnd = vec_oprnd; | |
4420 | ||
4421 | /* For conversion in multiple steps, continue to get operands | |
4422 | recursively. */ | |
4423 | if (multi_step_cvt) | |
e4057a39 | 4424 | vect_get_loop_based_defs (oprnd, stmt_info, vec_oprnds, |
82570274 | 4425 | multi_step_cvt - 1); |
4a00c761 JJ |
4426 | } |
4427 | ||
4428 | ||
4429 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
4430 | For multi-step conversions store the resulting vectors and call the function | |
4431 | recursively. */ | |
4432 | ||
4433 | static void | |
9771b263 | 4434 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
32e8e429 RS |
4435 | int multi_step_cvt, |
4436 | stmt_vec_info stmt_info, | |
9771b263 | 4437 | vec<tree> vec_dsts, |
4a00c761 JJ |
4438 | gimple_stmt_iterator *gsi, |
4439 | slp_tree slp_node, enum tree_code code, | |
4440 | stmt_vec_info *prev_stmt_info) | |
4441 | { | |
4442 | unsigned int i; | |
4443 | tree vop0, vop1, new_tmp, vec_dest; | |
4a00c761 | 4444 | |
9771b263 | 4445 | vec_dest = vec_dsts.pop (); |
4a00c761 | 4446 | |
9771b263 | 4447 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
4448 | { |
4449 | /* Create demotion operation. */ | |
9771b263 DN |
4450 | vop0 = (*vec_oprnds)[i]; |
4451 | vop1 = (*vec_oprnds)[i + 1]; | |
e1bd7296 | 4452 | gassign *new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
4453 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
4454 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
e1bd7296 | 4455 | stmt_vec_info new_stmt_info |
86a91c0a | 4456 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
4a00c761 JJ |
4457 | |
4458 | if (multi_step_cvt) | |
4459 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 4460 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
4461 | else |
4462 | { | |
4463 | /* This is the last step of the conversion sequence. Store the | |
4464 | vectors in SLP_NODE or in vector info of the scalar statement | |
4465 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
4466 | if (slp_node) | |
e1bd7296 | 4467 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
4a00c761 | 4468 | else |
c689ce1e RB |
4469 | { |
4470 | if (!*prev_stmt_info) | |
e1bd7296 | 4471 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt_info; |
c689ce1e | 4472 | else |
e1bd7296 | 4473 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt_info; |
4a00c761 | 4474 | |
e1bd7296 | 4475 | *prev_stmt_info = new_stmt_info; |
c689ce1e | 4476 | } |
4a00c761 JJ |
4477 | } |
4478 | } | |
4479 | ||
4480 | /* For multi-step demotion operations we first generate demotion operations | |
4481 | from the source type to the intermediate types, and then combine the | |
4482 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
4483 | type. */ | |
4484 | if (multi_step_cvt) | |
4485 | { | |
4486 | /* At each level of recursion we have half of the operands we had at the | |
4487 | previous level. */ | |
9771b263 | 4488 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 | 4489 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
86a91c0a RS |
4490 | stmt_info, vec_dsts, gsi, |
4491 | slp_node, VEC_PACK_TRUNC_EXPR, | |
4a00c761 JJ |
4492 | prev_stmt_info); |
4493 | } | |
4494 | ||
9771b263 | 4495 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4496 | } |
4497 | ||
4498 | ||
4499 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
82570274 RS |
4500 | and VEC_OPRNDS1, for a binary operation associated with scalar statement |
4501 | STMT_INFO. For multi-step conversions store the resulting vectors and | |
4502 | call the function recursively. */ | |
4a00c761 JJ |
4503 | |
4504 | static void | |
9771b263 DN |
4505 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
4506 | vec<tree> *vec_oprnds1, | |
82570274 | 4507 | stmt_vec_info stmt_info, tree vec_dest, |
4a00c761 JJ |
4508 | gimple_stmt_iterator *gsi, |
4509 | enum tree_code code1, | |
4510 | enum tree_code code2, tree decl1, | |
4511 | tree decl2, int op_type) | |
4512 | { | |
4513 | int i; | |
4514 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 4515 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 4516 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 4517 | |
9771b263 DN |
4518 | vec_tmp.create (vec_oprnds0->length () * 2); |
4519 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
4520 | { |
4521 | if (op_type == binary_op) | |
9771b263 | 4522 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
4523 | else |
4524 | vop1 = NULL_TREE; | |
4525 | ||
4526 | /* Generate the two halves of promotion operation. */ | |
4527 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
82570274 RS |
4528 | op_type, vec_dest, gsi, |
4529 | stmt_info); | |
4a00c761 | 4530 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, |
82570274 RS |
4531 | op_type, vec_dest, gsi, |
4532 | stmt_info); | |
4a00c761 JJ |
4533 | if (is_gimple_call (new_stmt1)) |
4534 | { | |
4535 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
4536 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
4537 | } | |
4538 | else | |
4539 | { | |
4540 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
4541 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
4542 | } | |
4543 | ||
4544 | /* Store the results for the next step. */ | |
9771b263 DN |
4545 | vec_tmp.quick_push (new_tmp1); |
4546 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
4547 | } |
4548 | ||
689eaba3 | 4549 | vec_oprnds0->release (); |
4a00c761 JJ |
4550 | *vec_oprnds0 = vec_tmp; |
4551 | } | |
4552 | ||
4553 | ||
32e8e429 RS |
4554 | /* Check if STMT_INFO performs a conversion operation that can be vectorized. |
4555 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized | |
4a00c761 | 4556 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
32e8e429 | 4557 | Return true if STMT_INFO is vectorizable in this way. */ |
ebfd146a IR |
4558 | |
4559 | static bool | |
32e8e429 | 4560 | vectorizable_conversion (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
1eede195 | 4561 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
68435eb2 | 4562 | stmt_vector_for_cost *cost_vec) |
ebfd146a IR |
4563 | { |
4564 | tree vec_dest; | |
4565 | tree scalar_dest; | |
4a00c761 | 4566 | tree op0, op1 = NULL_TREE; |
ebfd146a | 4567 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
ebfd146a IR |
4568 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4569 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 4570 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
4571 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
4572 | tree new_temp; | |
ebfd146a | 4573 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 4574 | int ndts = 2; |
ebfd146a | 4575 | stmt_vec_info prev_stmt_info; |
062d5ccc RS |
4576 | poly_uint64 nunits_in; |
4577 | poly_uint64 nunits_out; | |
ebfd146a | 4578 | tree vectype_out, vectype_in; |
4a00c761 JJ |
4579 | int ncopies, i, j; |
4580 | tree lhs_type, rhs_type; | |
ebfd146a | 4581 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
4582 | vec<tree> vec_oprnds0 = vNULL; |
4583 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 4584 | tree vop0; |
4a00c761 | 4585 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4586 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 4587 | int multi_step_cvt = 0; |
6e1aa848 | 4588 | vec<tree> interm_types = vNULL; |
4a00c761 JJ |
4589 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
4590 | int op_type; | |
4a00c761 | 4591 | unsigned short fltsz; |
ebfd146a IR |
4592 | |
4593 | /* Is STMT a vectorizable conversion? */ | |
4594 | ||
4a00c761 | 4595 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4596 | return false; |
4597 | ||
66c16fd9 RB |
4598 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4599 | && ! vec_stmt) | |
ebfd146a IR |
4600 | return false; |
4601 | ||
32e8e429 RS |
4602 | gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt); |
4603 | if (!stmt) | |
ebfd146a IR |
4604 | return false; |
4605 | ||
4606 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4607 | return false; | |
4608 | ||
4609 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
4610 | if (!CONVERT_EXPR_CODE_P (code) |
4611 | && code != FIX_TRUNC_EXPR | |
4612 | && code != FLOAT_EXPR | |
4613 | && code != WIDEN_MULT_EXPR | |
4614 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
4615 | return false; |
4616 | ||
4a00c761 JJ |
4617 | op_type = TREE_CODE_LENGTH (code); |
4618 | ||
ebfd146a | 4619 | /* Check types of lhs and rhs. */ |
b690cc0f | 4620 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 4621 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
4622 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
4623 | ||
ebfd146a IR |
4624 | op0 = gimple_assign_rhs1 (stmt); |
4625 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
4626 | |
4627 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4628 | && !((INTEGRAL_TYPE_P (lhs_type) | |
4629 | && INTEGRAL_TYPE_P (rhs_type)) | |
4630 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
4631 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
4632 | return false; | |
4633 | ||
e6f5c25d IE |
4634 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4635 | && ((INTEGRAL_TYPE_P (lhs_type) | |
2be65d9e | 4636 | && !type_has_mode_precision_p (lhs_type)) |
e6f5c25d | 4637 | || (INTEGRAL_TYPE_P (rhs_type) |
2be65d9e | 4638 | && !type_has_mode_precision_p (rhs_type)))) |
4a00c761 | 4639 | { |
73fbfcad | 4640 | if (dump_enabled_p ()) |
78c60e3d | 4641 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
4642 | "type conversion to/from bit-precision unsupported." |
4643 | "\n"); | |
4a00c761 JJ |
4644 | return false; |
4645 | } | |
4646 | ||
b690cc0f | 4647 | /* Check the operands of the operation. */ |
894dd753 | 4648 | if (!vect_is_simple_use (op0, vinfo, &dt[0], &vectype_in)) |
b690cc0f | 4649 | { |
73fbfcad | 4650 | if (dump_enabled_p ()) |
78c60e3d | 4651 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4652 | "use not simple.\n"); |
b690cc0f RG |
4653 | return false; |
4654 | } | |
4a00c761 JJ |
4655 | if (op_type == binary_op) |
4656 | { | |
4657 | bool ok; | |
4658 | ||
4659 | op1 = gimple_assign_rhs2 (stmt); | |
4660 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
4661 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
4662 | OP1. */ | |
4663 | if (CONSTANT_CLASS_P (op0)) | |
894dd753 | 4664 | ok = vect_is_simple_use (op1, vinfo, &dt[1], &vectype_in); |
4a00c761 | 4665 | else |
894dd753 | 4666 | ok = vect_is_simple_use (op1, vinfo, &dt[1]); |
4a00c761 JJ |
4667 | |
4668 | if (!ok) | |
4669 | { | |
73fbfcad | 4670 | if (dump_enabled_p ()) |
78c60e3d | 4671 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4672 | "use not simple.\n"); |
4a00c761 JJ |
4673 | return false; |
4674 | } | |
4675 | } | |
4676 | ||
b690cc0f RG |
4677 | /* If op0 is an external or constant defs use a vector type of |
4678 | the same size as the output vector type. */ | |
ebfd146a | 4679 | if (!vectype_in) |
b690cc0f | 4680 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
4681 | if (vec_stmt) |
4682 | gcc_assert (vectype_in); | |
4683 | if (!vectype_in) | |
4684 | { | |
73fbfcad | 4685 | if (dump_enabled_p ()) |
3c2a8ed0 DM |
4686 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4687 | "no vectype for scalar type %T\n", rhs_type); | |
7d8930a0 IR |
4688 | |
4689 | return false; | |
4690 | } | |
ebfd146a | 4691 | |
e6f5c25d IE |
4692 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4693 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
4694 | { | |
4695 | if (dump_enabled_p ()) | |
3c2a8ed0 DM |
4696 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4697 | "can't convert between boolean and non " | |
4698 | "boolean vectors %T\n", rhs_type); | |
e6f5c25d IE |
4699 | |
4700 | return false; | |
4701 | } | |
4702 | ||
b690cc0f RG |
4703 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
4704 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
062d5ccc | 4705 | if (known_eq (nunits_out, nunits_in)) |
ebfd146a | 4706 | modifier = NONE; |
062d5ccc RS |
4707 | else if (multiple_p (nunits_out, nunits_in)) |
4708 | modifier = NARROW; | |
ebfd146a | 4709 | else |
062d5ccc RS |
4710 | { |
4711 | gcc_checking_assert (multiple_p (nunits_in, nunits_out)); | |
4712 | modifier = WIDEN; | |
4713 | } | |
ebfd146a | 4714 | |
ff802fa1 IR |
4715 | /* Multiple types in SLP are handled by creating the appropriate number of |
4716 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4717 | case of SLP. */ | |
fce57248 | 4718 | if (slp_node) |
ebfd146a | 4719 | ncopies = 1; |
4a00c761 | 4720 | else if (modifier == NARROW) |
e8f142e2 | 4721 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
4a00c761 | 4722 | else |
e8f142e2 | 4723 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
b8698a0f | 4724 | |
ebfd146a IR |
4725 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
4726 | needs to be generated. */ | |
4727 | gcc_assert (ncopies >= 1); | |
4728 | ||
16d22000 RS |
4729 | bool found_mode = false; |
4730 | scalar_mode lhs_mode = SCALAR_TYPE_MODE (lhs_type); | |
4731 | scalar_mode rhs_mode = SCALAR_TYPE_MODE (rhs_type); | |
4732 | opt_scalar_mode rhs_mode_iter; | |
b397965c | 4733 | |
ebfd146a | 4734 | /* Supportable by target? */ |
4a00c761 | 4735 | switch (modifier) |
ebfd146a | 4736 | { |
4a00c761 JJ |
4737 | case NONE: |
4738 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4739 | return false; | |
4740 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
4741 | &decl1, &code1)) | |
4742 | break; | |
4743 | /* FALLTHRU */ | |
4744 | unsupported: | |
73fbfcad | 4745 | if (dump_enabled_p ()) |
78c60e3d | 4746 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4747 | "conversion not supported by target.\n"); |
ebfd146a | 4748 | return false; |
ebfd146a | 4749 | |
4a00c761 | 4750 | case WIDEN: |
86a91c0a RS |
4751 | if (supportable_widening_operation (code, stmt_info, vectype_out, |
4752 | vectype_in, &code1, &code2, | |
4753 | &multi_step_cvt, &interm_types)) | |
4a00c761 JJ |
4754 | { |
4755 | /* Binary widening operation can only be supported directly by the | |
4756 | architecture. */ | |
4757 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
4758 | break; | |
4759 | } | |
4760 | ||
4761 | if (code != FLOAT_EXPR | |
b397965c | 4762 | || GET_MODE_SIZE (lhs_mode) <= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4763 | goto unsupported; |
4764 | ||
b397965c | 4765 | fltsz = GET_MODE_SIZE (lhs_mode); |
16d22000 | 4766 | FOR_EACH_2XWIDER_MODE (rhs_mode_iter, rhs_mode) |
4a00c761 | 4767 | { |
16d22000 | 4768 | rhs_mode = rhs_mode_iter.require (); |
c94843d2 RS |
4769 | if (GET_MODE_SIZE (rhs_mode) > fltsz) |
4770 | break; | |
4771 | ||
4a00c761 JJ |
4772 | cvt_type |
4773 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4774 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4775 | if (cvt_type == NULL_TREE) | |
4776 | goto unsupported; | |
4777 | ||
4778 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4779 | { | |
4780 | if (!supportable_convert_operation (code, vectype_out, | |
4781 | cvt_type, &decl1, &codecvt1)) | |
4782 | goto unsupported; | |
4783 | } | |
86a91c0a RS |
4784 | else if (!supportable_widening_operation (code, stmt_info, |
4785 | vectype_out, cvt_type, | |
4786 | &codecvt1, &codecvt2, | |
4787 | &multi_step_cvt, | |
4a00c761 JJ |
4788 | &interm_types)) |
4789 | continue; | |
4790 | else | |
4791 | gcc_assert (multi_step_cvt == 0); | |
4792 | ||
86a91c0a | 4793 | if (supportable_widening_operation (NOP_EXPR, stmt_info, cvt_type, |
a86ec597 RH |
4794 | vectype_in, &code1, &code2, |
4795 | &multi_step_cvt, &interm_types)) | |
16d22000 RS |
4796 | { |
4797 | found_mode = true; | |
4798 | break; | |
4799 | } | |
4a00c761 JJ |
4800 | } |
4801 | ||
16d22000 | 4802 | if (!found_mode) |
4a00c761 JJ |
4803 | goto unsupported; |
4804 | ||
4805 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4806 | codecvt2 = ERROR_MARK; | |
4807 | else | |
4808 | { | |
4809 | multi_step_cvt++; | |
9771b263 | 4810 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
4811 | cvt_type = NULL_TREE; |
4812 | } | |
4813 | break; | |
4814 | ||
4815 | case NARROW: | |
4816 | gcc_assert (op_type == unary_op); | |
4817 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
4818 | &code1, &multi_step_cvt, | |
4819 | &interm_types)) | |
4820 | break; | |
4821 | ||
4822 | if (code != FIX_TRUNC_EXPR | |
b397965c | 4823 | || GET_MODE_SIZE (lhs_mode) >= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4824 | goto unsupported; |
4825 | ||
4a00c761 JJ |
4826 | cvt_type |
4827 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4828 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4829 | if (cvt_type == NULL_TREE) | |
4830 | goto unsupported; | |
4831 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
4832 | &decl1, &codecvt1)) | |
4833 | goto unsupported; | |
4834 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
4835 | &code1, &multi_step_cvt, | |
4836 | &interm_types)) | |
4837 | break; | |
4838 | goto unsupported; | |
4839 | ||
4840 | default: | |
4841 | gcc_unreachable (); | |
ebfd146a IR |
4842 | } |
4843 | ||
4844 | if (!vec_stmt) /* transformation not required. */ | |
4845 | { | |
adac3a68 | 4846 | DUMP_VECT_SCOPE ("vectorizable_conversion"); |
4a00c761 | 4847 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
4848 | { |
4849 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
68435eb2 RB |
4850 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, |
4851 | cost_vec); | |
8bd37302 | 4852 | } |
4a00c761 JJ |
4853 | else if (modifier == NARROW) |
4854 | { | |
4855 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
68435eb2 RB |
4856 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt, |
4857 | cost_vec); | |
4a00c761 JJ |
4858 | } |
4859 | else | |
4860 | { | |
4861 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
68435eb2 RB |
4862 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt, |
4863 | cost_vec); | |
4a00c761 | 4864 | } |
9771b263 | 4865 | interm_types.release (); |
ebfd146a IR |
4866 | return true; |
4867 | } | |
4868 | ||
67b8dbac | 4869 | /* Transform. */ |
73fbfcad | 4870 | if (dump_enabled_p ()) |
78c60e3d | 4871 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4872 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 4873 | |
4a00c761 JJ |
4874 | if (op_type == binary_op) |
4875 | { | |
4876 | if (CONSTANT_CLASS_P (op0)) | |
4877 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
4878 | else if (CONSTANT_CLASS_P (op1)) | |
4879 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
4880 | } | |
4881 | ||
4882 | /* In case of multi-step conversion, we first generate conversion operations | |
4883 | to the intermediate types, and then from that types to the final one. | |
4884 | We create vector destinations for the intermediate type (TYPES) received | |
4885 | from supportable_*_operation, and store them in the correct order | |
4886 | for future use in vect_create_vectorized_*_stmts (). */ | |
8c681247 | 4887 | auto_vec<tree> vec_dsts (multi_step_cvt + 1); |
82294ec1 JJ |
4888 | vec_dest = vect_create_destination_var (scalar_dest, |
4889 | (cvt_type && modifier == WIDEN) | |
4890 | ? cvt_type : vectype_out); | |
9771b263 | 4891 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4892 | |
4893 | if (multi_step_cvt) | |
4894 | { | |
9771b263 DN |
4895 | for (i = interm_types.length () - 1; |
4896 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
4897 | { |
4898 | vec_dest = vect_create_destination_var (scalar_dest, | |
4899 | intermediate_type); | |
9771b263 | 4900 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4901 | } |
4902 | } | |
ebfd146a | 4903 | |
4a00c761 | 4904 | if (cvt_type) |
82294ec1 JJ |
4905 | vec_dest = vect_create_destination_var (scalar_dest, |
4906 | modifier == WIDEN | |
4907 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
4908 | |
4909 | if (!slp_node) | |
4910 | { | |
30862efc | 4911 | if (modifier == WIDEN) |
4a00c761 | 4912 | { |
c3284718 | 4913 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 4914 | if (op_type == binary_op) |
9771b263 | 4915 | vec_oprnds1.create (1); |
4a00c761 | 4916 | } |
30862efc | 4917 | else if (modifier == NARROW) |
9771b263 DN |
4918 | vec_oprnds0.create ( |
4919 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
4920 | } |
4921 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 4922 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 4923 | |
4a00c761 | 4924 | last_oprnd = op0; |
ebfd146a IR |
4925 | prev_stmt_info = NULL; |
4926 | switch (modifier) | |
4927 | { | |
4928 | case NONE: | |
4929 | for (j = 0; j < ncopies; j++) | |
4930 | { | |
ebfd146a | 4931 | if (j == 0) |
86a91c0a RS |
4932 | vect_get_vec_defs (op0, NULL, stmt_info, &vec_oprnds0, |
4933 | NULL, slp_node); | |
ebfd146a | 4934 | else |
e4057a39 | 4935 | vect_get_vec_defs_for_stmt_copy (vinfo, &vec_oprnds0, NULL); |
ebfd146a | 4936 | |
9771b263 | 4937 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 | 4938 | { |
e1bd7296 | 4939 | stmt_vec_info new_stmt_info; |
4a00c761 JJ |
4940 | /* Arguments are ready, create the new vector stmt. */ |
4941 | if (code1 == CALL_EXPR) | |
4942 | { | |
e1bd7296 | 4943 | gcall *new_stmt = gimple_build_call (decl1, 1, vop0); |
4a00c761 JJ |
4944 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4945 | gimple_call_set_lhs (new_stmt, new_temp); | |
e1bd7296 | 4946 | new_stmt_info |
86a91c0a | 4947 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
4a00c761 JJ |
4948 | } |
4949 | else | |
4950 | { | |
4951 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
e1bd7296 RS |
4952 | gassign *new_stmt |
4953 | = gimple_build_assign (vec_dest, code1, vop0); | |
4a00c761 JJ |
4954 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4955 | gimple_assign_set_lhs (new_stmt, new_temp); | |
e1bd7296 | 4956 | new_stmt_info |
86a91c0a | 4957 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
4a00c761 JJ |
4958 | } |
4959 | ||
4a00c761 | 4960 | if (slp_node) |
e1bd7296 | 4961 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
225ce44b RB |
4962 | else |
4963 | { | |
4964 | if (!prev_stmt_info) | |
e1bd7296 RS |
4965 | STMT_VINFO_VEC_STMT (stmt_info) |
4966 | = *vec_stmt = new_stmt_info; | |
225ce44b | 4967 | else |
e1bd7296 RS |
4968 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
4969 | prev_stmt_info = new_stmt_info; | |
225ce44b | 4970 | } |
4a00c761 | 4971 | } |
ebfd146a IR |
4972 | } |
4973 | break; | |
4974 | ||
4975 | case WIDEN: | |
4976 | /* In case the vectorization factor (VF) is bigger than the number | |
4977 | of elements that we can fit in a vectype (nunits), we have to | |
4978 | generate more than one vector stmt - i.e - we need to "unroll" | |
4979 | the vector stmt by a factor VF/nunits. */ | |
4980 | for (j = 0; j < ncopies; j++) | |
4981 | { | |
4a00c761 | 4982 | /* Handle uses. */ |
ebfd146a | 4983 | if (j == 0) |
4a00c761 JJ |
4984 | { |
4985 | if (slp_node) | |
4986 | { | |
4987 | if (code == WIDEN_LSHIFT_EXPR) | |
4988 | { | |
4989 | unsigned int k; | |
ebfd146a | 4990 | |
4a00c761 JJ |
4991 | vec_oprnd1 = op1; |
4992 | /* Store vec_oprnd1 for every vector stmt to be created | |
4993 | for SLP_NODE. We check during the analysis that all | |
4994 | the shift arguments are the same. */ | |
4995 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4996 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 | 4997 | |
86a91c0a RS |
4998 | vect_get_vec_defs (op0, NULL_TREE, stmt_info, |
4999 | &vec_oprnds0, NULL, slp_node); | |
4a00c761 JJ |
5000 | } |
5001 | else | |
86a91c0a | 5002 | vect_get_vec_defs (op0, op1, stmt_info, &vec_oprnds0, |
306b0c92 | 5003 | &vec_oprnds1, slp_node); |
4a00c761 JJ |
5004 | } |
5005 | else | |
5006 | { | |
86a91c0a | 5007 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt_info); |
9771b263 | 5008 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
5009 | if (op_type == binary_op) |
5010 | { | |
5011 | if (code == WIDEN_LSHIFT_EXPR) | |
5012 | vec_oprnd1 = op1; | |
5013 | else | |
86a91c0a RS |
5014 | vec_oprnd1 |
5015 | = vect_get_vec_def_for_operand (op1, stmt_info); | |
9771b263 | 5016 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
5017 | } |
5018 | } | |
5019 | } | |
ebfd146a | 5020 | else |
4a00c761 | 5021 | { |
e4057a39 | 5022 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (vinfo, vec_oprnd0); |
9771b263 DN |
5023 | vec_oprnds0.truncate (0); |
5024 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
5025 | if (op_type == binary_op) |
5026 | { | |
5027 | if (code == WIDEN_LSHIFT_EXPR) | |
5028 | vec_oprnd1 = op1; | |
5029 | else | |
e4057a39 | 5030 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (vinfo, |
4a00c761 | 5031 | vec_oprnd1); |
9771b263 DN |
5032 | vec_oprnds1.truncate (0); |
5033 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
5034 | } |
5035 | } | |
ebfd146a | 5036 | |
4a00c761 JJ |
5037 | /* Arguments are ready. Create the new vector stmts. */ |
5038 | for (i = multi_step_cvt; i >= 0; i--) | |
5039 | { | |
9771b263 | 5040 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
5041 | enum tree_code c1 = code1, c2 = code2; |
5042 | if (i == 0 && codecvt2 != ERROR_MARK) | |
5043 | { | |
5044 | c1 = codecvt1; | |
5045 | c2 = codecvt2; | |
5046 | } | |
5047 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
86a91c0a RS |
5048 | &vec_oprnds1, stmt_info, |
5049 | this_dest, gsi, | |
4a00c761 JJ |
5050 | c1, c2, decl1, decl2, |
5051 | op_type); | |
5052 | } | |
5053 | ||
9771b263 | 5054 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 | 5055 | { |
e1bd7296 | 5056 | stmt_vec_info new_stmt_info; |
4a00c761 JJ |
5057 | if (cvt_type) |
5058 | { | |
5059 | if (codecvt1 == CALL_EXPR) | |
5060 | { | |
e1bd7296 | 5061 | gcall *new_stmt = gimple_build_call (decl1, 1, vop0); |
4a00c761 JJ |
5062 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5063 | gimple_call_set_lhs (new_stmt, new_temp); | |
e1bd7296 | 5064 | new_stmt_info |
86a91c0a RS |
5065 | = vect_finish_stmt_generation (stmt_info, new_stmt, |
5066 | gsi); | |
4a00c761 JJ |
5067 | } |
5068 | else | |
5069 | { | |
5070 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 5071 | new_temp = make_ssa_name (vec_dest); |
e1bd7296 RS |
5072 | gassign *new_stmt |
5073 | = gimple_build_assign (new_temp, codecvt1, vop0); | |
5074 | new_stmt_info | |
86a91c0a RS |
5075 | = vect_finish_stmt_generation (stmt_info, new_stmt, |
5076 | gsi); | |
4a00c761 | 5077 | } |
4a00c761 JJ |
5078 | } |
5079 | else | |
e1bd7296 | 5080 | new_stmt_info = vinfo->lookup_def (vop0); |
4a00c761 JJ |
5081 | |
5082 | if (slp_node) | |
e1bd7296 | 5083 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
4a00c761 | 5084 | else |
c689ce1e RB |
5085 | { |
5086 | if (!prev_stmt_info) | |
e1bd7296 | 5087 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt_info; |
c689ce1e | 5088 | else |
e1bd7296 RS |
5089 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
5090 | prev_stmt_info = new_stmt_info; | |
c689ce1e | 5091 | } |
4a00c761 | 5092 | } |
ebfd146a | 5093 | } |
4a00c761 JJ |
5094 | |
5095 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
5096 | break; |
5097 | ||
5098 | case NARROW: | |
5099 | /* In case the vectorization factor (VF) is bigger than the number | |
5100 | of elements that we can fit in a vectype (nunits), we have to | |
5101 | generate more than one vector stmt - i.e - we need to "unroll" | |
5102 | the vector stmt by a factor VF/nunits. */ | |
5103 | for (j = 0; j < ncopies; j++) | |
5104 | { | |
5105 | /* Handle uses. */ | |
4a00c761 | 5106 | if (slp_node) |
86a91c0a | 5107 | vect_get_vec_defs (op0, NULL_TREE, stmt_info, &vec_oprnds0, NULL, |
306b0c92 | 5108 | slp_node); |
ebfd146a IR |
5109 | else |
5110 | { | |
9771b263 | 5111 | vec_oprnds0.truncate (0); |
e4057a39 | 5112 | vect_get_loop_based_defs (&last_oprnd, stmt_info, &vec_oprnds0, |
4a00c761 | 5113 | vect_pow2 (multi_step_cvt) - 1); |
ebfd146a IR |
5114 | } |
5115 | ||
4a00c761 JJ |
5116 | /* Arguments are ready. Create the new vector stmts. */ |
5117 | if (cvt_type) | |
9771b263 | 5118 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
5119 | { |
5120 | if (codecvt1 == CALL_EXPR) | |
5121 | { | |
e1bd7296 | 5122 | gcall *new_stmt = gimple_build_call (decl1, 1, vop0); |
4a00c761 JJ |
5123 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5124 | gimple_call_set_lhs (new_stmt, new_temp); | |
86a91c0a | 5125 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
4a00c761 JJ |
5126 | } |
5127 | else | |
5128 | { | |
5129 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 5130 | new_temp = make_ssa_name (vec_dest); |
e1bd7296 RS |
5131 | gassign *new_stmt |
5132 | = gimple_build_assign (new_temp, codecvt1, vop0); | |
86a91c0a | 5133 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
4a00c761 | 5134 | } |
ebfd146a | 5135 | |
9771b263 | 5136 | vec_oprnds0[i] = new_temp; |
4a00c761 | 5137 | } |
ebfd146a | 5138 | |
4a00c761 | 5139 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
86a91c0a | 5140 | stmt_info, vec_dsts, gsi, |
4a00c761 JJ |
5141 | slp_node, code1, |
5142 | &prev_stmt_info); | |
ebfd146a IR |
5143 | } |
5144 | ||
5145 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 5146 | break; |
ebfd146a IR |
5147 | } |
5148 | ||
9771b263 DN |
5149 | vec_oprnds0.release (); |
5150 | vec_oprnds1.release (); | |
9771b263 | 5151 | interm_types.release (); |
ebfd146a IR |
5152 | |
5153 | return true; | |
5154 | } | |
ff802fa1 IR |
5155 | |
5156 | ||
ebfd146a IR |
5157 | /* Function vectorizable_assignment. |
5158 | ||
32e8e429 RS |
5159 | Check if STMT_INFO performs an assignment (copy) that can be vectorized. |
5160 | If VEC_STMT is also passed, vectorize the STMT_INFO: create a vectorized | |
5161 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
5162 | Return true if STMT_INFO is vectorizable in this way. */ | |
ebfd146a IR |
5163 | |
5164 | static bool | |
32e8e429 | 5165 | vectorizable_assignment (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
1eede195 | 5166 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
68435eb2 | 5167 | stmt_vector_for_cost *cost_vec) |
ebfd146a IR |
5168 | { |
5169 | tree vec_dest; | |
5170 | tree scalar_dest; | |
5171 | tree op; | |
ebfd146a IR |
5172 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
5173 | tree new_temp; | |
4fc5ebf1 JG |
5174 | enum vect_def_type dt[1] = {vect_unknown_def_type}; |
5175 | int ndts = 1; | |
ebfd146a | 5176 | int ncopies; |
f18b55bd | 5177 | int i, j; |
6e1aa848 | 5178 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5179 | tree vop; |
a70d6342 | 5180 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5181 | vec_info *vinfo = stmt_info->vinfo; |
f18b55bd | 5182 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
5183 | enum tree_code code; |
5184 | tree vectype_in; | |
ebfd146a | 5185 | |
a70d6342 | 5186 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5187 | return false; |
5188 | ||
66c16fd9 RB |
5189 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5190 | && ! vec_stmt) | |
ebfd146a IR |
5191 | return false; |
5192 | ||
5193 | /* Is vectorizable assignment? */ | |
32e8e429 RS |
5194 | gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt); |
5195 | if (!stmt) | |
ebfd146a IR |
5196 | return false; |
5197 | ||
5198 | scalar_dest = gimple_assign_lhs (stmt); | |
5199 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
5200 | return false; | |
5201 | ||
fde9c428 | 5202 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 5203 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
5204 | || code == PAREN_EXPR |
5205 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
5206 | op = gimple_assign_rhs1 (stmt); |
5207 | else | |
5208 | return false; | |
5209 | ||
7b7ec6c5 RG |
5210 | if (code == VIEW_CONVERT_EXPR) |
5211 | op = TREE_OPERAND (op, 0); | |
5212 | ||
465c8c19 | 5213 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
928686b1 | 5214 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 JJ |
5215 | |
5216 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5217 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5218 | case of SLP. */ | |
fce57248 | 5219 | if (slp_node) |
465c8c19 JJ |
5220 | ncopies = 1; |
5221 | else | |
e8f142e2 | 5222 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
5223 | |
5224 | gcc_assert (ncopies >= 1); | |
5225 | ||
894dd753 | 5226 | if (!vect_is_simple_use (op, vinfo, &dt[0], &vectype_in)) |
ebfd146a | 5227 | { |
73fbfcad | 5228 | if (dump_enabled_p ()) |
78c60e3d | 5229 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5230 | "use not simple.\n"); |
ebfd146a IR |
5231 | return false; |
5232 | } | |
5233 | ||
fde9c428 RG |
5234 | /* We can handle NOP_EXPR conversions that do not change the number |
5235 | of elements or the vector size. */ | |
7b7ec6c5 RG |
5236 | if ((CONVERT_EXPR_CODE_P (code) |
5237 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 | 5238 | && (!vectype_in |
928686b1 | 5239 | || maybe_ne (TYPE_VECTOR_SUBPARTS (vectype_in), nunits) |
cf098191 RS |
5240 | || maybe_ne (GET_MODE_SIZE (TYPE_MODE (vectype)), |
5241 | GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
fde9c428 RG |
5242 | return false; |
5243 | ||
7b7b1813 RG |
5244 | /* We do not handle bit-precision changes. */ |
5245 | if ((CONVERT_EXPR_CODE_P (code) | |
5246 | || code == VIEW_CONVERT_EXPR) | |
5247 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2be65d9e RS |
5248 | && (!type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
5249 | || !type_has_mode_precision_p (TREE_TYPE (op))) | |
7b7b1813 RG |
5250 | /* But a conversion that does not change the bit-pattern is ok. */ |
5251 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
5252 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2dab46d5 IE |
5253 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
5254 | /* Conversion between boolean types of different sizes is | |
5255 | a simple assignment in case their vectypes are same | |
5256 | boolean vectors. */ | |
5257 | && (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
5258 | || !VECTOR_BOOLEAN_TYPE_P (vectype_in))) | |
7b7b1813 | 5259 | { |
73fbfcad | 5260 | if (dump_enabled_p ()) |
78c60e3d SS |
5261 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5262 | "type conversion to/from bit-precision " | |
e645e942 | 5263 | "unsupported.\n"); |
7b7b1813 RG |
5264 | return false; |
5265 | } | |
5266 | ||
ebfd146a IR |
5267 | if (!vec_stmt) /* transformation not required. */ |
5268 | { | |
5269 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
adac3a68 | 5270 | DUMP_VECT_SCOPE ("vectorizable_assignment"); |
68435eb2 | 5271 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
ebfd146a IR |
5272 | return true; |
5273 | } | |
5274 | ||
67b8dbac | 5275 | /* Transform. */ |
73fbfcad | 5276 | if (dump_enabled_p ()) |
e645e942 | 5277 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
5278 | |
5279 | /* Handle def. */ | |
5280 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5281 | ||
5282 | /* Handle use. */ | |
f18b55bd | 5283 | for (j = 0; j < ncopies; j++) |
ebfd146a | 5284 | { |
f18b55bd IR |
5285 | /* Handle uses. */ |
5286 | if (j == 0) | |
86a91c0a | 5287 | vect_get_vec_defs (op, NULL, stmt_info, &vec_oprnds, NULL, slp_node); |
f18b55bd | 5288 | else |
e4057a39 | 5289 | vect_get_vec_defs_for_stmt_copy (vinfo, &vec_oprnds, NULL); |
f18b55bd IR |
5290 | |
5291 | /* Arguments are ready. create the new vector stmt. */ | |
e1bd7296 | 5292 | stmt_vec_info new_stmt_info = NULL; |
9771b263 | 5293 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 5294 | { |
7b7ec6c5 RG |
5295 | if (CONVERT_EXPR_CODE_P (code) |
5296 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 5297 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
e1bd7296 | 5298 | gassign *new_stmt = gimple_build_assign (vec_dest, vop); |
f18b55bd IR |
5299 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5300 | gimple_assign_set_lhs (new_stmt, new_temp); | |
86a91c0a RS |
5301 | new_stmt_info |
5302 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); | |
f18b55bd | 5303 | if (slp_node) |
e1bd7296 | 5304 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
f18b55bd | 5305 | } |
ebfd146a IR |
5306 | |
5307 | if (slp_node) | |
f18b55bd IR |
5308 | continue; |
5309 | ||
5310 | if (j == 0) | |
e1bd7296 | 5311 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
f18b55bd | 5312 | else |
e1bd7296 | 5313 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
f18b55bd | 5314 | |
e1bd7296 | 5315 | prev_stmt_info = new_stmt_info; |
f18b55bd | 5316 | } |
b8698a0f | 5317 | |
9771b263 | 5318 | vec_oprnds.release (); |
ebfd146a IR |
5319 | return true; |
5320 | } | |
5321 | ||
9dc3f7de | 5322 | |
1107f3ae IR |
5323 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
5324 | either as shift by a scalar or by a vector. */ | |
5325 | ||
5326 | bool | |
5327 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
5328 | { | |
5329 | ||
ef4bddc2 | 5330 | machine_mode vec_mode; |
1107f3ae IR |
5331 | optab optab; |
5332 | int icode; | |
5333 | tree vectype; | |
5334 | ||
5335 | vectype = get_vectype_for_scalar_type (scalar_type); | |
5336 | if (!vectype) | |
5337 | return false; | |
5338 | ||
5339 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5340 | if (!optab | |
5341 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
5342 | { | |
5343 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5344 | if (!optab | |
5345 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
5346 | == CODE_FOR_nothing)) | |
5347 | return false; | |
5348 | } | |
5349 | ||
5350 | vec_mode = TYPE_MODE (vectype); | |
5351 | icode = (int) optab_handler (optab, vec_mode); | |
5352 | if (icode == CODE_FOR_nothing) | |
5353 | return false; | |
5354 | ||
5355 | return true; | |
5356 | } | |
5357 | ||
5358 | ||
9dc3f7de IR |
5359 | /* Function vectorizable_shift. |
5360 | ||
32e8e429 RS |
5361 | Check if STMT_INFO performs a shift operation that can be vectorized. |
5362 | If VEC_STMT is also passed, vectorize the STMT_INFO: create a vectorized | |
5363 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
5364 | Return true if STMT_INFO is vectorizable in this way. */ | |
9dc3f7de IR |
5365 | |
5366 | static bool | |
32e8e429 | 5367 | vectorizable_shift (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
1eede195 | 5368 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
68435eb2 | 5369 | stmt_vector_for_cost *cost_vec) |
9dc3f7de IR |
5370 | { |
5371 | tree vec_dest; | |
5372 | tree scalar_dest; | |
5373 | tree op0, op1 = NULL; | |
5374 | tree vec_oprnd1 = NULL_TREE; | |
9dc3f7de IR |
5375 | tree vectype; |
5376 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
5377 | enum tree_code code; | |
ef4bddc2 | 5378 | machine_mode vec_mode; |
9dc3f7de IR |
5379 | tree new_temp; |
5380 | optab optab; | |
5381 | int icode; | |
ef4bddc2 | 5382 | machine_mode optab_op2_mode; |
9dc3f7de | 5383 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 5384 | int ndts = 2; |
9dc3f7de | 5385 | stmt_vec_info prev_stmt_info; |
928686b1 RS |
5386 | poly_uint64 nunits_in; |
5387 | poly_uint64 nunits_out; | |
9dc3f7de | 5388 | tree vectype_out; |
cede2577 | 5389 | tree op1_vectype; |
9dc3f7de IR |
5390 | int ncopies; |
5391 | int j, i; | |
6e1aa848 DN |
5392 | vec<tree> vec_oprnds0 = vNULL; |
5393 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
5394 | tree vop0, vop1; |
5395 | unsigned int k; | |
49eab32e | 5396 | bool scalar_shift_arg = true; |
9dc3f7de | 5397 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5398 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
5399 | |
5400 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
5401 | return false; | |
5402 | ||
66c16fd9 RB |
5403 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5404 | && ! vec_stmt) | |
9dc3f7de IR |
5405 | return false; |
5406 | ||
5407 | /* Is STMT a vectorizable binary/unary operation? */ | |
32e8e429 RS |
5408 | gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt); |
5409 | if (!stmt) | |
9dc3f7de IR |
5410 | return false; |
5411 | ||
5412 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5413 | return false; | |
5414 | ||
5415 | code = gimple_assign_rhs_code (stmt); | |
5416 | ||
5417 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
5418 | || code == RROTATE_EXPR)) | |
5419 | return false; | |
5420 | ||
5421 | scalar_dest = gimple_assign_lhs (stmt); | |
5422 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2be65d9e | 5423 | if (!type_has_mode_precision_p (TREE_TYPE (scalar_dest))) |
7b7b1813 | 5424 | { |
73fbfcad | 5425 | if (dump_enabled_p ()) |
78c60e3d | 5426 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5427 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
5428 | return false; |
5429 | } | |
9dc3f7de IR |
5430 | |
5431 | op0 = gimple_assign_rhs1 (stmt); | |
894dd753 | 5432 | if (!vect_is_simple_use (op0, vinfo, &dt[0], &vectype)) |
9dc3f7de | 5433 | { |
73fbfcad | 5434 | if (dump_enabled_p ()) |
78c60e3d | 5435 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5436 | "use not simple.\n"); |
9dc3f7de IR |
5437 | return false; |
5438 | } | |
5439 | /* If op0 is an external or constant def use a vector type with | |
5440 | the same size as the output vector type. */ | |
5441 | if (!vectype) | |
5442 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5443 | if (vec_stmt) | |
5444 | gcc_assert (vectype); | |
5445 | if (!vectype) | |
5446 | { | |
73fbfcad | 5447 | if (dump_enabled_p ()) |
78c60e3d | 5448 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5449 | "no vectype for scalar type\n"); |
9dc3f7de IR |
5450 | return false; |
5451 | } | |
5452 | ||
5453 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5454 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
928686b1 | 5455 | if (maybe_ne (nunits_out, nunits_in)) |
9dc3f7de IR |
5456 | return false; |
5457 | ||
5458 | op1 = gimple_assign_rhs2 (stmt); | |
fef96d8e RS |
5459 | stmt_vec_info op1_def_stmt_info; |
5460 | if (!vect_is_simple_use (op1, vinfo, &dt[1], &op1_vectype, | |
5461 | &op1_def_stmt_info)) | |
9dc3f7de | 5462 | { |
73fbfcad | 5463 | if (dump_enabled_p ()) |
78c60e3d | 5464 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5465 | "use not simple.\n"); |
9dc3f7de IR |
5466 | return false; |
5467 | } | |
5468 | ||
9dc3f7de IR |
5469 | /* Multiple types in SLP are handled by creating the appropriate number of |
5470 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5471 | case of SLP. */ | |
fce57248 | 5472 | if (slp_node) |
9dc3f7de IR |
5473 | ncopies = 1; |
5474 | else | |
e8f142e2 | 5475 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
9dc3f7de IR |
5476 | |
5477 | gcc_assert (ncopies >= 1); | |
5478 | ||
5479 | /* Determine whether the shift amount is a vector, or scalar. If the | |
5480 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
5481 | ||
dbfa87aa YR |
5482 | if ((dt[1] == vect_internal_def |
5483 | || dt[1] == vect_induction_def) | |
5484 | && !slp_node) | |
49eab32e JJ |
5485 | scalar_shift_arg = false; |
5486 | else if (dt[1] == vect_constant_def | |
5487 | || dt[1] == vect_external_def | |
5488 | || dt[1] == vect_internal_def) | |
5489 | { | |
5490 | /* In SLP, need to check whether the shift count is the same, | |
5491 | in loops if it is a constant or invariant, it is always | |
5492 | a scalar shift. */ | |
5493 | if (slp_node) | |
5494 | { | |
b9787581 RS |
5495 | vec<stmt_vec_info> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
5496 | stmt_vec_info slpstmt_info; | |
49eab32e | 5497 | |
b9787581 RS |
5498 | FOR_EACH_VEC_ELT (stmts, k, slpstmt_info) |
5499 | { | |
5500 | gassign *slpstmt = as_a <gassign *> (slpstmt_info->stmt); | |
5501 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) | |
5502 | scalar_shift_arg = false; | |
5503 | } | |
49eab32e | 5504 | } |
60d393e8 RB |
5505 | |
5506 | /* If the shift amount is computed by a pattern stmt we cannot | |
5507 | use the scalar amount directly thus give up and use a vector | |
5508 | shift. */ | |
fef96d8e RS |
5509 | if (op1_def_stmt_info && is_pattern_stmt_p (op1_def_stmt_info)) |
5510 | scalar_shift_arg = false; | |
49eab32e JJ |
5511 | } |
5512 | else | |
5513 | { | |
73fbfcad | 5514 | if (dump_enabled_p ()) |
78c60e3d | 5515 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5516 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
5517 | return false; |
5518 | } | |
5519 | ||
9dc3f7de | 5520 | /* Vector shifted by vector. */ |
49eab32e | 5521 | if (!scalar_shift_arg) |
9dc3f7de IR |
5522 | { |
5523 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 5524 | if (dump_enabled_p ()) |
78c60e3d | 5525 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5526 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 5527 | |
aa948027 JJ |
5528 | if (!op1_vectype) |
5529 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
5530 | if (op1_vectype == NULL_TREE | |
5531 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 5532 | { |
73fbfcad | 5533 | if (dump_enabled_p ()) |
78c60e3d SS |
5534 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5535 | "unusable type for last operand in" | |
e645e942 | 5536 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
5537 | return false; |
5538 | } | |
9dc3f7de IR |
5539 | } |
5540 | /* See if the machine has a vector shifted by scalar insn and if not | |
5541 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 5542 | else |
9dc3f7de IR |
5543 | { |
5544 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5545 | if (optab | |
5546 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
5547 | { | |
73fbfcad | 5548 | if (dump_enabled_p ()) |
78c60e3d | 5549 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5550 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
5551 | } |
5552 | else | |
5553 | { | |
5554 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5555 | if (optab | |
5556 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
5557 | != CODE_FOR_nothing)) | |
5558 | { | |
49eab32e JJ |
5559 | scalar_shift_arg = false; |
5560 | ||
73fbfcad | 5561 | if (dump_enabled_p ()) |
78c60e3d | 5562 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5563 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
5564 | |
5565 | /* Unlike the other binary operators, shifts/rotates have | |
5566 | the rhs being int, instead of the same type as the lhs, | |
5567 | so make sure the scalar is the right type if we are | |
aa948027 | 5568 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
5569 | if (dt[1] == vect_constant_def) |
5570 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
5571 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
5572 | TREE_TYPE (op1))) | |
5573 | { | |
5574 | if (slp_node | |
5575 | && TYPE_MODE (TREE_TYPE (vectype)) | |
5576 | != TYPE_MODE (TREE_TYPE (op1))) | |
5577 | { | |
73fbfcad | 5578 | if (dump_enabled_p ()) |
78c60e3d SS |
5579 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5580 | "unusable type for last operand in" | |
e645e942 | 5581 | " vector/vector shift/rotate.\n"); |
21c0a521 | 5582 | return false; |
aa948027 JJ |
5583 | } |
5584 | if (vec_stmt && !slp_node) | |
5585 | { | |
5586 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
86a91c0a | 5587 | op1 = vect_init_vector (stmt_info, op1, |
aa948027 JJ |
5588 | TREE_TYPE (vectype), NULL); |
5589 | } | |
5590 | } | |
9dc3f7de IR |
5591 | } |
5592 | } | |
5593 | } | |
9dc3f7de IR |
5594 | |
5595 | /* Supportable by target? */ | |
5596 | if (!optab) | |
5597 | { | |
73fbfcad | 5598 | if (dump_enabled_p ()) |
78c60e3d | 5599 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5600 | "no optab.\n"); |
9dc3f7de IR |
5601 | return false; |
5602 | } | |
5603 | vec_mode = TYPE_MODE (vectype); | |
5604 | icode = (int) optab_handler (optab, vec_mode); | |
5605 | if (icode == CODE_FOR_nothing) | |
5606 | { | |
73fbfcad | 5607 | if (dump_enabled_p ()) |
78c60e3d | 5608 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5609 | "op not supported by target.\n"); |
9dc3f7de | 5610 | /* Check only during analysis. */ |
cf098191 | 5611 | if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) |
ca09abcb RS |
5612 | || (!vec_stmt |
5613 | && !vect_worthwhile_without_simd_p (vinfo, code))) | |
9dc3f7de | 5614 | return false; |
73fbfcad | 5615 | if (dump_enabled_p ()) |
e645e942 TJ |
5616 | dump_printf_loc (MSG_NOTE, vect_location, |
5617 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
5618 | } |
5619 | ||
5620 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
ca09abcb RS |
5621 | if (!vec_stmt |
5622 | && !VECTOR_MODE_P (TYPE_MODE (vectype)) | |
5623 | && !vect_worthwhile_without_simd_p (vinfo, code)) | |
9dc3f7de | 5624 | { |
73fbfcad | 5625 | if (dump_enabled_p ()) |
78c60e3d | 5626 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5627 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
5628 | return false; |
5629 | } | |
5630 | ||
5631 | if (!vec_stmt) /* transformation not required. */ | |
5632 | { | |
5633 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
adac3a68 | 5634 | DUMP_VECT_SCOPE ("vectorizable_shift"); |
68435eb2 | 5635 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
9dc3f7de IR |
5636 | return true; |
5637 | } | |
5638 | ||
67b8dbac | 5639 | /* Transform. */ |
9dc3f7de | 5640 | |
73fbfcad | 5641 | if (dump_enabled_p ()) |
78c60e3d | 5642 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5643 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
5644 | |
5645 | /* Handle def. */ | |
5646 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5647 | ||
9dc3f7de IR |
5648 | prev_stmt_info = NULL; |
5649 | for (j = 0; j < ncopies; j++) | |
5650 | { | |
5651 | /* Handle uses. */ | |
5652 | if (j == 0) | |
5653 | { | |
5654 | if (scalar_shift_arg) | |
5655 | { | |
5656 | /* Vector shl and shr insn patterns can be defined with scalar | |
5657 | operand 2 (shift operand). In this case, use constant or loop | |
5658 | invariant op1 directly, without extending it to vector mode | |
5659 | first. */ | |
5660 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
5661 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
5662 | { | |
73fbfcad | 5663 | if (dump_enabled_p ()) |
78c60e3d | 5664 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5665 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 5666 | vec_oprnd1 = op1; |
8930f723 | 5667 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 5668 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5669 | if (slp_node) |
5670 | { | |
5671 | /* Store vec_oprnd1 for every vector stmt to be created | |
5672 | for SLP_NODE. We check during the analysis that all | |
5673 | the shift arguments are the same. | |
5674 | TODO: Allow different constants for different vector | |
5675 | stmts generated for an SLP instance. */ | |
5676 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 5677 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5678 | } |
5679 | } | |
5680 | } | |
5681 | ||
5682 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
5683 | (a special case for certain kind of vector shifts); otherwise, | |
5684 | operand 1 should be of a vector type (the usual case). */ | |
5685 | if (vec_oprnd1) | |
86a91c0a RS |
5686 | vect_get_vec_defs (op0, NULL_TREE, stmt_info, &vec_oprnds0, NULL, |
5687 | slp_node); | |
9dc3f7de | 5688 | else |
86a91c0a RS |
5689 | vect_get_vec_defs (op0, op1, stmt_info, &vec_oprnds0, &vec_oprnds1, |
5690 | slp_node); | |
9dc3f7de IR |
5691 | } |
5692 | else | |
e4057a39 | 5693 | vect_get_vec_defs_for_stmt_copy (vinfo, &vec_oprnds0, &vec_oprnds1); |
9dc3f7de IR |
5694 | |
5695 | /* Arguments are ready. Create the new vector stmt. */ | |
e1bd7296 | 5696 | stmt_vec_info new_stmt_info = NULL; |
9771b263 | 5697 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 5698 | { |
9771b263 | 5699 | vop1 = vec_oprnds1[i]; |
e1bd7296 | 5700 | gassign *new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
5701 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5702 | gimple_assign_set_lhs (new_stmt, new_temp); | |
86a91c0a RS |
5703 | new_stmt_info |
5704 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); | |
9dc3f7de | 5705 | if (slp_node) |
e1bd7296 | 5706 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
9dc3f7de IR |
5707 | } |
5708 | ||
5709 | if (slp_node) | |
5710 | continue; | |
5711 | ||
5712 | if (j == 0) | |
e1bd7296 | 5713 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
9dc3f7de | 5714 | else |
e1bd7296 RS |
5715 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
5716 | prev_stmt_info = new_stmt_info; | |
9dc3f7de IR |
5717 | } |
5718 | ||
9771b263 DN |
5719 | vec_oprnds0.release (); |
5720 | vec_oprnds1.release (); | |
9dc3f7de IR |
5721 | |
5722 | return true; | |
5723 | } | |
5724 | ||
5725 | ||
ebfd146a IR |
5726 | /* Function vectorizable_operation. |
5727 | ||
32e8e429 | 5728 | Check if STMT_INFO performs a binary, unary or ternary operation that can |
16949072 | 5729 | be vectorized. |
32e8e429 RS |
5730 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
5731 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
5732 | Return true if STMT_INFO is vectorizable in this way. */ | |
ebfd146a IR |
5733 | |
5734 | static bool | |
32e8e429 | 5735 | vectorizable_operation (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
1eede195 | 5736 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
68435eb2 | 5737 | stmt_vector_for_cost *cost_vec) |
ebfd146a | 5738 | { |
00f07b86 | 5739 | tree vec_dest; |
ebfd146a | 5740 | tree scalar_dest; |
16949072 | 5741 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
00f07b86 | 5742 | tree vectype; |
ebfd146a | 5743 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
0eb952ea | 5744 | enum tree_code code, orig_code; |
ef4bddc2 | 5745 | machine_mode vec_mode; |
ebfd146a IR |
5746 | tree new_temp; |
5747 | int op_type; | |
00f07b86 | 5748 | optab optab; |
523ba738 | 5749 | bool target_support_p; |
16949072 RG |
5750 | enum vect_def_type dt[3] |
5751 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 5752 | int ndts = 3; |
ebfd146a | 5753 | stmt_vec_info prev_stmt_info; |
928686b1 RS |
5754 | poly_uint64 nunits_in; |
5755 | poly_uint64 nunits_out; | |
ebfd146a IR |
5756 | tree vectype_out; |
5757 | int ncopies; | |
5758 | int j, i; | |
6e1aa848 DN |
5759 | vec<tree> vec_oprnds0 = vNULL; |
5760 | vec<tree> vec_oprnds1 = vNULL; | |
5761 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 5762 | tree vop0, vop1, vop2; |
a70d6342 | 5763 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5764 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 5765 | |
a70d6342 | 5766 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5767 | return false; |
5768 | ||
66c16fd9 RB |
5769 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5770 | && ! vec_stmt) | |
ebfd146a IR |
5771 | return false; |
5772 | ||
5773 | /* Is STMT a vectorizable binary/unary operation? */ | |
32e8e429 RS |
5774 | gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt); |
5775 | if (!stmt) | |
ebfd146a IR |
5776 | return false; |
5777 | ||
5778 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5779 | return false; | |
5780 | ||
0eb952ea | 5781 | orig_code = code = gimple_assign_rhs_code (stmt); |
ebfd146a | 5782 | |
1af4ebf5 MG |
5783 | /* For pointer addition and subtraction, we should use the normal |
5784 | plus and minus for the vector operation. */ | |
ebfd146a IR |
5785 | if (code == POINTER_PLUS_EXPR) |
5786 | code = PLUS_EXPR; | |
1af4ebf5 MG |
5787 | if (code == POINTER_DIFF_EXPR) |
5788 | code = MINUS_EXPR; | |
ebfd146a IR |
5789 | |
5790 | /* Support only unary or binary operations. */ | |
5791 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 5792 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 5793 | { |
73fbfcad | 5794 | if (dump_enabled_p ()) |
78c60e3d | 5795 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5796 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 5797 | op_type); |
ebfd146a IR |
5798 | return false; |
5799 | } | |
5800 | ||
b690cc0f RG |
5801 | scalar_dest = gimple_assign_lhs (stmt); |
5802 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
5803 | ||
7b7b1813 RG |
5804 | /* Most operations cannot handle bit-precision types without extra |
5805 | truncations. */ | |
045c1278 | 5806 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
2be65d9e | 5807 | && !type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
7b7b1813 RG |
5808 | /* Exception are bitwise binary operations. */ |
5809 | && code != BIT_IOR_EXPR | |
5810 | && code != BIT_XOR_EXPR | |
5811 | && code != BIT_AND_EXPR) | |
5812 | { | |
73fbfcad | 5813 | if (dump_enabled_p ()) |
78c60e3d | 5814 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5815 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
5816 | return false; |
5817 | } | |
5818 | ||
ebfd146a | 5819 | op0 = gimple_assign_rhs1 (stmt); |
894dd753 | 5820 | if (!vect_is_simple_use (op0, vinfo, &dt[0], &vectype)) |
ebfd146a | 5821 | { |
73fbfcad | 5822 | if (dump_enabled_p ()) |
78c60e3d | 5823 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5824 | "use not simple.\n"); |
ebfd146a IR |
5825 | return false; |
5826 | } | |
b690cc0f RG |
5827 | /* If op0 is an external or constant def use a vector type with |
5828 | the same size as the output vector type. */ | |
5829 | if (!vectype) | |
b036c6c5 IE |
5830 | { |
5831 | /* For boolean type we cannot determine vectype by | |
5832 | invariant value (don't know whether it is a vector | |
5833 | of booleans or vector of integers). We use output | |
5834 | vectype because operations on boolean don't change | |
5835 | type. */ | |
2568d8a1 | 5836 | if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op0))) |
b036c6c5 | 5837 | { |
2568d8a1 | 5838 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (scalar_dest))) |
b036c6c5 IE |
5839 | { |
5840 | if (dump_enabled_p ()) | |
5841 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5842 | "not supported operation on bool value.\n"); | |
5843 | return false; | |
5844 | } | |
5845 | vectype = vectype_out; | |
5846 | } | |
5847 | else | |
5848 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5849 | } | |
7d8930a0 IR |
5850 | if (vec_stmt) |
5851 | gcc_assert (vectype); | |
5852 | if (!vectype) | |
5853 | { | |
73fbfcad | 5854 | if (dump_enabled_p ()) |
3c2a8ed0 DM |
5855 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5856 | "no vectype for scalar type %T\n", | |
5857 | TREE_TYPE (op0)); | |
7d8930a0 IR |
5858 | |
5859 | return false; | |
5860 | } | |
b690cc0f RG |
5861 | |
5862 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5863 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
928686b1 | 5864 | if (maybe_ne (nunits_out, nunits_in)) |
b690cc0f | 5865 | return false; |
ebfd146a | 5866 | |
16949072 | 5867 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
5868 | { |
5869 | op1 = gimple_assign_rhs2 (stmt); | |
894dd753 | 5870 | if (!vect_is_simple_use (op1, vinfo, &dt[1])) |
ebfd146a | 5871 | { |
73fbfcad | 5872 | if (dump_enabled_p ()) |
78c60e3d | 5873 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5874 | "use not simple.\n"); |
ebfd146a IR |
5875 | return false; |
5876 | } | |
5877 | } | |
16949072 RG |
5878 | if (op_type == ternary_op) |
5879 | { | |
5880 | op2 = gimple_assign_rhs3 (stmt); | |
894dd753 | 5881 | if (!vect_is_simple_use (op2, vinfo, &dt[2])) |
16949072 | 5882 | { |
73fbfcad | 5883 | if (dump_enabled_p ()) |
78c60e3d | 5884 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5885 | "use not simple.\n"); |
16949072 RG |
5886 | return false; |
5887 | } | |
5888 | } | |
ebfd146a | 5889 | |
b690cc0f | 5890 | /* Multiple types in SLP are handled by creating the appropriate number of |
ff802fa1 | 5891 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 5892 | case of SLP. */ |
fce57248 | 5893 | if (slp_node) |
b690cc0f RG |
5894 | ncopies = 1; |
5895 | else | |
e8f142e2 | 5896 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
b690cc0f RG |
5897 | |
5898 | gcc_assert (ncopies >= 1); | |
5899 | ||
9dc3f7de | 5900 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
5901 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
5902 | || code == RROTATE_EXPR) | |
9dc3f7de | 5903 | return false; |
ebfd146a | 5904 | |
ebfd146a | 5905 | /* Supportable by target? */ |
00f07b86 RH |
5906 | |
5907 | vec_mode = TYPE_MODE (vectype); | |
5908 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 5909 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
5910 | else |
5911 | { | |
5912 | optab = optab_for_tree_code (code, vectype, optab_default); | |
5913 | if (!optab) | |
5deb57cb | 5914 | { |
73fbfcad | 5915 | if (dump_enabled_p ()) |
78c60e3d | 5916 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5917 | "no optab.\n"); |
00f07b86 | 5918 | return false; |
5deb57cb | 5919 | } |
523ba738 RS |
5920 | target_support_p = (optab_handler (optab, vec_mode) |
5921 | != CODE_FOR_nothing); | |
5deb57cb JJ |
5922 | } |
5923 | ||
523ba738 | 5924 | if (!target_support_p) |
ebfd146a | 5925 | { |
73fbfcad | 5926 | if (dump_enabled_p ()) |
78c60e3d | 5927 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5928 | "op not supported by target.\n"); |
ebfd146a | 5929 | /* Check only during analysis. */ |
cf098191 | 5930 | if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) |
ca09abcb | 5931 | || (!vec_stmt && !vect_worthwhile_without_simd_p (vinfo, code))) |
ebfd146a | 5932 | return false; |
73fbfcad | 5933 | if (dump_enabled_p ()) |
e645e942 TJ |
5934 | dump_printf_loc (MSG_NOTE, vect_location, |
5935 | "proceeding using word mode.\n"); | |
383d9c83 IR |
5936 | } |
5937 | ||
4a00c761 | 5938 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
5939 | if (!VECTOR_MODE_P (vec_mode) |
5940 | && !vec_stmt | |
ca09abcb | 5941 | && !vect_worthwhile_without_simd_p (vinfo, code)) |
7d8930a0 | 5942 | { |
73fbfcad | 5943 | if (dump_enabled_p ()) |
78c60e3d | 5944 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5945 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 5946 | return false; |
7d8930a0 | 5947 | } |
ebfd146a | 5948 | |
ebfd146a IR |
5949 | if (!vec_stmt) /* transformation not required. */ |
5950 | { | |
4a00c761 | 5951 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
adac3a68 | 5952 | DUMP_VECT_SCOPE ("vectorizable_operation"); |
68435eb2 | 5953 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
ebfd146a IR |
5954 | return true; |
5955 | } | |
5956 | ||
67b8dbac | 5957 | /* Transform. */ |
ebfd146a | 5958 | |
73fbfcad | 5959 | if (dump_enabled_p ()) |
78c60e3d | 5960 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5961 | "transform binary/unary operation.\n"); |
383d9c83 | 5962 | |
0eb952ea JJ |
5963 | /* POINTER_DIFF_EXPR has pointer arguments which are vectorized as |
5964 | vectors with unsigned elements, but the result is signed. So, we | |
5965 | need to compute the MINUS_EXPR into vectype temporary and | |
5966 | VIEW_CONVERT_EXPR it into the final vectype_out result. */ | |
5967 | tree vec_cvt_dest = NULL_TREE; | |
5968 | if (orig_code == POINTER_DIFF_EXPR) | |
7b76867b RB |
5969 | { |
5970 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5971 | vec_cvt_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
5972 | } | |
5973 | /* Handle def. */ | |
5974 | else | |
5975 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
0eb952ea | 5976 | |
ebfd146a IR |
5977 | /* In case the vectorization factor (VF) is bigger than the number |
5978 | of elements that we can fit in a vectype (nunits), we have to generate | |
5979 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
5980 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
5981 | from one copy of the vector stmt to the next, in the field | |
5982 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
5983 | stages to find the correct vector defs to be used when vectorizing | |
5984 | stmts that use the defs of the current stmt. The example below | |
5985 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
5986 | we need to create 4 vectorized stmts): | |
5987 | ||
5988 | before vectorization: | |
5989 | RELATED_STMT VEC_STMT | |
5990 | S1: x = memref - - | |
5991 | S2: z = x + 1 - - | |
5992 | ||
5993 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
5994 | there): | |
5995 | RELATED_STMT VEC_STMT | |
5996 | VS1_0: vx0 = memref0 VS1_1 - | |
5997 | VS1_1: vx1 = memref1 VS1_2 - | |
5998 | VS1_2: vx2 = memref2 VS1_3 - | |
5999 | VS1_3: vx3 = memref3 - - | |
6000 | S1: x = load - VS1_0 | |
6001 | S2: z = x + 1 - - | |
6002 | ||
6003 | step2: vectorize stmt S2 (done here): | |
6004 | To vectorize stmt S2 we first need to find the relevant vector | |
6005 | def for the first operand 'x'. This is, as usual, obtained from | |
6006 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
6007 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
6008 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
6009 | the vector stmt VS2_0, and as usual, record it in the | |
6010 | STMT_VINFO_VEC_STMT of stmt S2. | |
6011 | When creating the second copy (VS2_1), we obtain the relevant vector | |
6012 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
6013 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
6014 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
6015 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
6016 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
6017 | chain of stmts and pointers: | |
6018 | RELATED_STMT VEC_STMT | |
6019 | VS1_0: vx0 = memref0 VS1_1 - | |
6020 | VS1_1: vx1 = memref1 VS1_2 - | |
6021 | VS1_2: vx2 = memref2 VS1_3 - | |
6022 | VS1_3: vx3 = memref3 - - | |
6023 | S1: x = load - VS1_0 | |
6024 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
6025 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
6026 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
6027 | VS2_3: vz3 = vx3 + v1 - - | |
6028 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
6029 | |
6030 | prev_stmt_info = NULL; | |
6031 | for (j = 0; j < ncopies; j++) | |
6032 | { | |
6033 | /* Handle uses. */ | |
6034 | if (j == 0) | |
4a00c761 | 6035 | { |
d6476f90 | 6036 | if (op_type == binary_op) |
86a91c0a | 6037 | vect_get_vec_defs (op0, op1, stmt_info, &vec_oprnds0, &vec_oprnds1, |
306b0c92 | 6038 | slp_node); |
d6476f90 RB |
6039 | else if (op_type == ternary_op) |
6040 | { | |
6041 | if (slp_node) | |
6042 | { | |
6043 | auto_vec<tree> ops(3); | |
6044 | ops.quick_push (op0); | |
6045 | ops.quick_push (op1); | |
6046 | ops.quick_push (op2); | |
6047 | auto_vec<vec<tree> > vec_defs(3); | |
6048 | vect_get_slp_defs (ops, slp_node, &vec_defs); | |
6049 | vec_oprnds0 = vec_defs[0]; | |
6050 | vec_oprnds1 = vec_defs[1]; | |
6051 | vec_oprnds2 = vec_defs[2]; | |
6052 | } | |
6053 | else | |
6054 | { | |
86a91c0a RS |
6055 | vect_get_vec_defs (op0, op1, stmt_info, &vec_oprnds0, |
6056 | &vec_oprnds1, NULL); | |
6057 | vect_get_vec_defs (op2, NULL_TREE, stmt_info, &vec_oprnds2, | |
6058 | NULL, NULL); | |
d6476f90 RB |
6059 | } |
6060 | } | |
4a00c761 | 6061 | else |
86a91c0a | 6062 | vect_get_vec_defs (op0, NULL_TREE, stmt_info, &vec_oprnds0, NULL, |
306b0c92 | 6063 | slp_node); |
4a00c761 | 6064 | } |
ebfd146a | 6065 | else |
4a00c761 | 6066 | { |
e4057a39 | 6067 | vect_get_vec_defs_for_stmt_copy (vinfo, &vec_oprnds0, &vec_oprnds1); |
4a00c761 JJ |
6068 | if (op_type == ternary_op) |
6069 | { | |
9771b263 | 6070 | tree vec_oprnd = vec_oprnds2.pop (); |
e4057a39 | 6071 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (vinfo, |
9771b263 | 6072 | vec_oprnd)); |
4a00c761 JJ |
6073 | } |
6074 | } | |
6075 | ||
6076 | /* Arguments are ready. Create the new vector stmt. */ | |
e1bd7296 | 6077 | stmt_vec_info new_stmt_info = NULL; |
9771b263 | 6078 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 6079 | { |
4a00c761 | 6080 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 6081 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 6082 | vop2 = ((op_type == ternary_op) |
9771b263 | 6083 | ? vec_oprnds2[i] : NULL_TREE); |
e1bd7296 RS |
6084 | gassign *new_stmt = gimple_build_assign (vec_dest, code, |
6085 | vop0, vop1, vop2); | |
4a00c761 JJ |
6086 | new_temp = make_ssa_name (vec_dest, new_stmt); |
6087 | gimple_assign_set_lhs (new_stmt, new_temp); | |
86a91c0a RS |
6088 | new_stmt_info |
6089 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); | |
0eb952ea JJ |
6090 | if (vec_cvt_dest) |
6091 | { | |
6092 | new_temp = build1 (VIEW_CONVERT_EXPR, vectype_out, new_temp); | |
e1bd7296 RS |
6093 | gassign *new_stmt |
6094 | = gimple_build_assign (vec_cvt_dest, VIEW_CONVERT_EXPR, | |
6095 | new_temp); | |
0eb952ea JJ |
6096 | new_temp = make_ssa_name (vec_cvt_dest, new_stmt); |
6097 | gimple_assign_set_lhs (new_stmt, new_temp); | |
e1bd7296 | 6098 | new_stmt_info |
86a91c0a | 6099 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
0eb952ea | 6100 | } |
4a00c761 | 6101 | if (slp_node) |
e1bd7296 | 6102 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
ebfd146a IR |
6103 | } |
6104 | ||
4a00c761 JJ |
6105 | if (slp_node) |
6106 | continue; | |
6107 | ||
6108 | if (j == 0) | |
e1bd7296 | 6109 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
4a00c761 | 6110 | else |
e1bd7296 RS |
6111 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
6112 | prev_stmt_info = new_stmt_info; | |
ebfd146a IR |
6113 | } |
6114 | ||
9771b263 DN |
6115 | vec_oprnds0.release (); |
6116 | vec_oprnds1.release (); | |
6117 | vec_oprnds2.release (); | |
ebfd146a | 6118 | |
ebfd146a IR |
6119 | return true; |
6120 | } | |
6121 | ||
89fa689a | 6122 | /* A helper function to ensure data reference DR_INFO's base alignment. */ |
c716e67f XDL |
6123 | |
6124 | static void | |
89fa689a | 6125 | ensure_base_align (dr_vec_info *dr_info) |
c716e67f | 6126 | { |
89fa689a | 6127 | if (dr_info->misalignment == DR_MISALIGNMENT_UNINITIALIZED) |
c716e67f XDL |
6128 | return; |
6129 | ||
89fa689a | 6130 | if (dr_info->base_misaligned) |
c716e67f | 6131 | { |
89fa689a | 6132 | tree base_decl = dr_info->base_decl; |
c716e67f | 6133 | |
89fa689a RS |
6134 | unsigned int align_base_to |
6135 | = DR_TARGET_ALIGNMENT (dr_info) * BITS_PER_UNIT; | |
f702e7d4 | 6136 | |
428f0c67 | 6137 | if (decl_in_symtab_p (base_decl)) |
f702e7d4 | 6138 | symtab_node::get (base_decl)->increase_alignment (align_base_to); |
428f0c67 JH |
6139 | else |
6140 | { | |
f702e7d4 | 6141 | SET_DECL_ALIGN (base_decl, align_base_to); |
428f0c67 JH |
6142 | DECL_USER_ALIGN (base_decl) = 1; |
6143 | } | |
89fa689a | 6144 | dr_info->base_misaligned = false; |
c716e67f XDL |
6145 | } |
6146 | } | |
6147 | ||
ebfd146a | 6148 | |
44fc7854 BE |
6149 | /* Function get_group_alias_ptr_type. |
6150 | ||
32e8e429 | 6151 | Return the alias type for the group starting at FIRST_STMT_INFO. */ |
44fc7854 BE |
6152 | |
6153 | static tree | |
32e8e429 | 6154 | get_group_alias_ptr_type (stmt_vec_info first_stmt_info) |
44fc7854 BE |
6155 | { |
6156 | struct data_reference *first_dr, *next_dr; | |
44fc7854 | 6157 | |
91987857 RS |
6158 | first_dr = STMT_VINFO_DATA_REF (first_stmt_info); |
6159 | stmt_vec_info next_stmt_info = DR_GROUP_NEXT_ELEMENT (first_stmt_info); | |
bffb8014 | 6160 | while (next_stmt_info) |
44fc7854 | 6161 | { |
bffb8014 | 6162 | next_dr = STMT_VINFO_DATA_REF (next_stmt_info); |
44fc7854 BE |
6163 | if (get_alias_set (DR_REF (first_dr)) |
6164 | != get_alias_set (DR_REF (next_dr))) | |
6165 | { | |
6166 | if (dump_enabled_p ()) | |
6167 | dump_printf_loc (MSG_NOTE, vect_location, | |
6168 | "conflicting alias set types.\n"); | |
6169 | return ptr_type_node; | |
6170 | } | |
bffb8014 | 6171 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
44fc7854 BE |
6172 | } |
6173 | return reference_alias_ptr_type (DR_REF (first_dr)); | |
6174 | } | |
6175 | ||
6176 | ||
ebfd146a IR |
6177 | /* Function vectorizable_store. |
6178 | ||
32e8e429 RS |
6179 | Check if STMT_INFO defines a non scalar data-ref (array/pointer/structure) |
6180 | that can be vectorized. | |
6181 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized | |
6182 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
6183 | Return true if STMT_INFO is vectorizable in this way. */ | |
ebfd146a IR |
6184 | |
6185 | static bool | |
32e8e429 | 6186 | vectorizable_store (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
1eede195 RS |
6187 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
6188 | stmt_vector_for_cost *cost_vec) | |
ebfd146a | 6189 | { |
ebfd146a IR |
6190 | tree data_ref; |
6191 | tree op; | |
6192 | tree vec_oprnd = NULL_TREE; | |
272c6793 | 6193 | tree elem_type; |
ebfd146a | 6194 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6195 | struct loop *loop = NULL; |
ef4bddc2 | 6196 | machine_mode vec_mode; |
ebfd146a IR |
6197 | tree dummy; |
6198 | enum dr_alignment_support alignment_support_scheme; | |
929b4411 RS |
6199 | enum vect_def_type rhs_dt = vect_unknown_def_type; |
6200 | enum vect_def_type mask_dt = vect_unknown_def_type; | |
ebfd146a IR |
6201 | stmt_vec_info prev_stmt_info = NULL; |
6202 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6203 | tree dataref_offset = NULL_TREE; |
355fe088 | 6204 | gimple *ptr_incr = NULL; |
ebfd146a IR |
6205 | int ncopies; |
6206 | int j; | |
bffb8014 | 6207 | stmt_vec_info first_stmt_info; |
2de001ee | 6208 | bool grouped_store; |
ebfd146a | 6209 | unsigned int group_size, i; |
6e1aa848 DN |
6210 | vec<tree> oprnds = vNULL; |
6211 | vec<tree> result_chain = vNULL; | |
09dfa495 | 6212 | tree offset = NULL_TREE; |
6e1aa848 | 6213 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 6214 | bool slp = (slp_node != NULL); |
ebfd146a | 6215 | unsigned int vec_num; |
a70d6342 | 6216 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 6217 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 6218 | tree aggr_type; |
134c85ca | 6219 | gather_scatter_info gs_info; |
d9f21f6a | 6220 | poly_uint64 vf; |
2de001ee | 6221 | vec_load_store_type vls_type; |
44fc7854 | 6222 | tree ref_type; |
a70d6342 | 6223 | |
a70d6342 | 6224 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
6225 | return false; |
6226 | ||
66c16fd9 RB |
6227 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6228 | && ! vec_stmt) | |
ebfd146a IR |
6229 | return false; |
6230 | ||
6231 | /* Is vectorizable store? */ | |
6232 | ||
c3a8f964 | 6233 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
86a91c0a | 6234 | if (gassign *assign = dyn_cast <gassign *> (stmt_info->stmt)) |
c3a8f964 | 6235 | { |
beb456c3 | 6236 | tree scalar_dest = gimple_assign_lhs (assign); |
c3a8f964 RS |
6237 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
6238 | && is_pattern_stmt_p (stmt_info)) | |
6239 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
6240 | if (TREE_CODE (scalar_dest) != ARRAY_REF | |
6241 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF | |
6242 | && TREE_CODE (scalar_dest) != INDIRECT_REF | |
6243 | && TREE_CODE (scalar_dest) != COMPONENT_REF | |
6244 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
6245 | && TREE_CODE (scalar_dest) != REALPART_EXPR | |
6246 | && TREE_CODE (scalar_dest) != MEM_REF) | |
6247 | return false; | |
6248 | } | |
6249 | else | |
6250 | { | |
86a91c0a | 6251 | gcall *call = dyn_cast <gcall *> (stmt_info->stmt); |
f307441a RS |
6252 | if (!call || !gimple_call_internal_p (call)) |
6253 | return false; | |
6254 | ||
6255 | internal_fn ifn = gimple_call_internal_fn (call); | |
6256 | if (!internal_store_fn_p (ifn)) | |
c3a8f964 | 6257 | return false; |
ebfd146a | 6258 | |
c3a8f964 RS |
6259 | if (slp_node != NULL) |
6260 | { | |
6261 | if (dump_enabled_p ()) | |
6262 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6263 | "SLP of masked stores not supported.\n"); | |
6264 | return false; | |
6265 | } | |
6266 | ||
f307441a RS |
6267 | int mask_index = internal_fn_mask_index (ifn); |
6268 | if (mask_index >= 0) | |
6269 | { | |
6270 | mask = gimple_call_arg (call, mask_index); | |
86a91c0a | 6271 | if (!vect_check_load_store_mask (stmt_info, mask, &mask_dt, |
929b4411 | 6272 | &mask_vectype)) |
f307441a RS |
6273 | return false; |
6274 | } | |
c3a8f964 RS |
6275 | } |
6276 | ||
86a91c0a | 6277 | op = vect_get_store_rhs (stmt_info); |
ebfd146a | 6278 | |
fce57248 RS |
6279 | /* Cannot have hybrid store SLP -- that would mean storing to the |
6280 | same location twice. */ | |
6281 | gcc_assert (slp == PURE_SLP_STMT (stmt_info)); | |
6282 | ||
f4d09712 | 6283 | tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE; |
4d694b27 | 6284 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 JJ |
6285 | |
6286 | if (loop_vinfo) | |
b17dc4d4 RB |
6287 | { |
6288 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6289 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6290 | } | |
6291 | else | |
6292 | vf = 1; | |
465c8c19 JJ |
6293 | |
6294 | /* Multiple types in SLP are handled by creating the appropriate number of | |
6295 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
6296 | case of SLP. */ | |
fce57248 | 6297 | if (slp) |
465c8c19 JJ |
6298 | ncopies = 1; |
6299 | else | |
e8f142e2 | 6300 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
6301 | |
6302 | gcc_assert (ncopies >= 1); | |
6303 | ||
6304 | /* FORNOW. This restriction should be relaxed. */ | |
86a91c0a | 6305 | if (loop && nested_in_vect_loop_p (loop, stmt_info) && ncopies > 1) |
465c8c19 JJ |
6306 | { |
6307 | if (dump_enabled_p ()) | |
6308 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6309 | "multiple types in nested loop.\n"); | |
6310 | return false; | |
6311 | } | |
6312 | ||
86a91c0a | 6313 | if (!vect_check_store_rhs (stmt_info, op, &rhs_dt, &rhs_vectype, &vls_type)) |
f4d09712 KY |
6314 | return false; |
6315 | ||
272c6793 | 6316 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 6317 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 6318 | |
ebfd146a IR |
6319 | if (!STMT_VINFO_DATA_REF (stmt_info)) |
6320 | return false; | |
6321 | ||
2de001ee | 6322 | vect_memory_access_type memory_access_type; |
86a91c0a | 6323 | if (!get_load_store_type (stmt_info, vectype, slp, mask, vls_type, ncopies, |
2de001ee RS |
6324 | &memory_access_type, &gs_info)) |
6325 | return false; | |
3bab6342 | 6326 | |
c3a8f964 RS |
6327 | if (mask) |
6328 | { | |
7e11fc7f RS |
6329 | if (memory_access_type == VMAT_CONTIGUOUS) |
6330 | { | |
6331 | if (!VECTOR_MODE_P (vec_mode) | |
6332 | || !can_vec_mask_load_store_p (vec_mode, | |
6333 | TYPE_MODE (mask_vectype), false)) | |
6334 | return false; | |
6335 | } | |
f307441a RS |
6336 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
6337 | && (memory_access_type != VMAT_GATHER_SCATTER || gs_info.decl)) | |
c3a8f964 RS |
6338 | { |
6339 | if (dump_enabled_p ()) | |
6340 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6341 | "unsupported access type for masked store.\n"); | |
6342 | return false; | |
6343 | } | |
c3a8f964 RS |
6344 | } |
6345 | else | |
6346 | { | |
6347 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6348 | (e.g. - array initialization with 0). */ | |
6349 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) | |
6350 | return false; | |
6351 | } | |
6352 | ||
89fa689a | 6353 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info), *first_dr_info = NULL; |
f307441a | 6354 | grouped_store = (STMT_VINFO_GROUPED_ACCESS (stmt_info) |
b5ec4de7 RS |
6355 | && memory_access_type != VMAT_GATHER_SCATTER |
6356 | && (slp || memory_access_type != VMAT_CONTIGUOUS)); | |
7cfb4d93 RS |
6357 | if (grouped_store) |
6358 | { | |
bffb8014 | 6359 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
89fa689a | 6360 | first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info); |
bffb8014 | 6361 | group_size = DR_GROUP_SIZE (first_stmt_info); |
7cfb4d93 RS |
6362 | } |
6363 | else | |
6364 | { | |
bffb8014 | 6365 | first_stmt_info = stmt_info; |
89fa689a | 6366 | first_dr_info = dr_info; |
7cfb4d93 RS |
6367 | group_size = vec_num = 1; |
6368 | } | |
6369 | ||
ebfd146a IR |
6370 | if (!vec_stmt) /* transformation not required. */ |
6371 | { | |
2de001ee | 6372 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
7cfb4d93 RS |
6373 | |
6374 | if (loop_vinfo | |
6375 | && LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) | |
6376 | check_load_store_masking (loop_vinfo, vectype, vls_type, group_size, | |
bfaa08b7 | 6377 | memory_access_type, &gs_info); |
7cfb4d93 | 6378 | |
ebfd146a | 6379 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; |
68435eb2 RB |
6380 | vect_model_store_cost (stmt_info, ncopies, rhs_dt, memory_access_type, |
6381 | vls_type, slp_node, cost_vec); | |
ebfd146a IR |
6382 | return true; |
6383 | } | |
2de001ee | 6384 | gcc_assert (memory_access_type == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
ebfd146a | 6385 | |
67b8dbac | 6386 | /* Transform. */ |
ebfd146a | 6387 | |
89fa689a | 6388 | ensure_base_align (dr_info); |
c716e67f | 6389 | |
f307441a | 6390 | if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
3bab6342 | 6391 | { |
c3a8f964 | 6392 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, src; |
134c85ca | 6393 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
6394 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; |
6395 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
6396 | edge pe = loop_preheader_edge (loop); | |
6397 | gimple_seq seq; | |
6398 | basic_block new_bb; | |
6399 | enum { NARROW, NONE, WIDEN } modifier; | |
4d694b27 RS |
6400 | poly_uint64 scatter_off_nunits |
6401 | = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype); | |
3bab6342 | 6402 | |
4d694b27 | 6403 | if (known_eq (nunits, scatter_off_nunits)) |
3bab6342 | 6404 | modifier = NONE; |
4d694b27 | 6405 | else if (known_eq (nunits * 2, scatter_off_nunits)) |
3bab6342 | 6406 | { |
3bab6342 AT |
6407 | modifier = WIDEN; |
6408 | ||
4d694b27 RS |
6409 | /* Currently gathers and scatters are only supported for |
6410 | fixed-length vectors. */ | |
6411 | unsigned int count = scatter_off_nunits.to_constant (); | |
6412 | vec_perm_builder sel (count, count, 1); | |
6413 | for (i = 0; i < (unsigned int) count; ++i) | |
6414 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 6415 | |
4d694b27 | 6416 | vec_perm_indices indices (sel, 1, count); |
e3342de4 RS |
6417 | perm_mask = vect_gen_perm_mask_checked (gs_info.offset_vectype, |
6418 | indices); | |
3bab6342 AT |
6419 | gcc_assert (perm_mask != NULL_TREE); |
6420 | } | |
4d694b27 | 6421 | else if (known_eq (nunits, scatter_off_nunits * 2)) |
3bab6342 | 6422 | { |
3bab6342 AT |
6423 | modifier = NARROW; |
6424 | ||
4d694b27 RS |
6425 | /* Currently gathers and scatters are only supported for |
6426 | fixed-length vectors. */ | |
6427 | unsigned int count = nunits.to_constant (); | |
6428 | vec_perm_builder sel (count, count, 1); | |
6429 | for (i = 0; i < (unsigned int) count; ++i) | |
6430 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 6431 | |
4d694b27 | 6432 | vec_perm_indices indices (sel, 2, count); |
e3342de4 | 6433 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); |
3bab6342 AT |
6434 | gcc_assert (perm_mask != NULL_TREE); |
6435 | ncopies *= 2; | |
6436 | } | |
6437 | else | |
6438 | gcc_unreachable (); | |
6439 | ||
134c85ca | 6440 | rettype = TREE_TYPE (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
6441 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
6442 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6443 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6444 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6445 | scaletype = TREE_VALUE (arglist); | |
6446 | ||
6447 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
6448 | && TREE_CODE (rettype) == VOID_TYPE); | |
6449 | ||
134c85ca | 6450 | ptr = fold_convert (ptrtype, gs_info.base); |
3bab6342 AT |
6451 | if (!is_gimple_min_invariant (ptr)) |
6452 | { | |
6453 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6454 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6455 | gcc_assert (!new_bb); | |
6456 | } | |
6457 | ||
6458 | /* Currently we support only unconditional scatter stores, | |
6459 | so mask should be all ones. */ | |
6460 | mask = build_int_cst (masktype, -1); | |
86a91c0a | 6461 | mask = vect_init_vector (stmt_info, mask, masktype, NULL); |
3bab6342 | 6462 | |
134c85ca | 6463 | scale = build_int_cst (scaletype, gs_info.scale); |
3bab6342 AT |
6464 | |
6465 | prev_stmt_info = NULL; | |
6466 | for (j = 0; j < ncopies; ++j) | |
6467 | { | |
6468 | if (j == 0) | |
6469 | { | |
6470 | src = vec_oprnd1 | |
86a91c0a | 6471 | = vect_get_vec_def_for_operand (op, stmt_info); |
3bab6342 | 6472 | op = vec_oprnd0 |
86a91c0a | 6473 | = vect_get_vec_def_for_operand (gs_info.offset, stmt_info); |
3bab6342 AT |
6474 | } |
6475 | else if (modifier != NONE && (j & 1)) | |
6476 | { | |
6477 | if (modifier == WIDEN) | |
6478 | { | |
6479 | src = vec_oprnd1 | |
e4057a39 | 6480 | = vect_get_vec_def_for_stmt_copy (vinfo, vec_oprnd1); |
3bab6342 | 6481 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, |
86a91c0a | 6482 | stmt_info, gsi); |
3bab6342 AT |
6483 | } |
6484 | else if (modifier == NARROW) | |
6485 | { | |
6486 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
86a91c0a | 6487 | stmt_info, gsi); |
3bab6342 | 6488 | op = vec_oprnd0 |
e4057a39 | 6489 | = vect_get_vec_def_for_stmt_copy (vinfo, vec_oprnd0); |
3bab6342 AT |
6490 | } |
6491 | else | |
6492 | gcc_unreachable (); | |
6493 | } | |
6494 | else | |
6495 | { | |
6496 | src = vec_oprnd1 | |
e4057a39 | 6497 | = vect_get_vec_def_for_stmt_copy (vinfo, vec_oprnd1); |
3bab6342 | 6498 | op = vec_oprnd0 |
e4057a39 | 6499 | = vect_get_vec_def_for_stmt_copy (vinfo, vec_oprnd0); |
3bab6342 AT |
6500 | } |
6501 | ||
6502 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
6503 | { | |
928686b1 RS |
6504 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)), |
6505 | TYPE_VECTOR_SUBPARTS (srctype))); | |
0e22bb5a | 6506 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 | 6507 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
e1bd7296 RS |
6508 | gassign *new_stmt |
6509 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
86a91c0a | 6510 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
3bab6342 AT |
6511 | src = var; |
6512 | } | |
6513 | ||
6514 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6515 | { | |
928686b1 RS |
6516 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), |
6517 | TYPE_VECTOR_SUBPARTS (idxtype))); | |
0e22bb5a | 6518 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 | 6519 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
e1bd7296 RS |
6520 | gassign *new_stmt |
6521 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
86a91c0a | 6522 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
3bab6342 AT |
6523 | op = var; |
6524 | } | |
6525 | ||
e1bd7296 | 6526 | gcall *new_stmt |
134c85ca | 6527 | = gimple_build_call (gs_info.decl, 5, ptr, mask, op, src, scale); |
e1bd7296 | 6528 | stmt_vec_info new_stmt_info |
86a91c0a | 6529 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
3bab6342 | 6530 | |
ddf98a96 | 6531 | if (prev_stmt_info == NULL) |
e1bd7296 | 6532 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
3bab6342 | 6533 | else |
e1bd7296 RS |
6534 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
6535 | prev_stmt_info = new_stmt_info; | |
3bab6342 AT |
6536 | } |
6537 | return true; | |
6538 | } | |
6539 | ||
f307441a | 6540 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
bffb8014 | 6541 | DR_GROUP_STORE_COUNT (DR_GROUP_FIRST_ELEMENT (stmt_info))++; |
ebfd146a | 6542 | |
f307441a RS |
6543 | if (grouped_store) |
6544 | { | |
ebfd146a | 6545 | /* FORNOW */ |
86a91c0a | 6546 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt_info)); |
ebfd146a IR |
6547 | |
6548 | /* We vectorize all the stmts of the interleaving group when we | |
6549 | reach the last stmt in the group. */ | |
bffb8014 RS |
6550 | if (DR_GROUP_STORE_COUNT (first_stmt_info) |
6551 | < DR_GROUP_SIZE (first_stmt_info) | |
ebfd146a IR |
6552 | && !slp) |
6553 | { | |
6554 | *vec_stmt = NULL; | |
6555 | return true; | |
6556 | } | |
6557 | ||
6558 | if (slp) | |
4b5caab7 | 6559 | { |
0d0293ac | 6560 | grouped_store = false; |
4b5caab7 IR |
6561 | /* VEC_NUM is the number of vect stmts to be created for this |
6562 | group. */ | |
6563 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
bffb8014 RS |
6564 | first_stmt_info = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
6565 | gcc_assert (DR_GROUP_FIRST_ELEMENT (first_stmt_info) | |
6566 | == first_stmt_info); | |
89fa689a | 6567 | first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info); |
bffb8014 | 6568 | op = vect_get_store_rhs (first_stmt_info); |
4b5caab7 | 6569 | } |
ebfd146a | 6570 | else |
4b5caab7 IR |
6571 | /* VEC_NUM is the number of vect stmts to be created for this |
6572 | group. */ | |
ebfd146a | 6573 | vec_num = group_size; |
44fc7854 | 6574 | |
bffb8014 | 6575 | ref_type = get_group_alias_ptr_type (first_stmt_info); |
ebfd146a | 6576 | } |
b8698a0f | 6577 | else |
89fa689a | 6578 | ref_type = reference_alias_ptr_type (DR_REF (first_dr_info->dr)); |
b8698a0f | 6579 | |
73fbfcad | 6580 | if (dump_enabled_p ()) |
78c60e3d | 6581 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6582 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 6583 | |
2de001ee RS |
6584 | if (memory_access_type == VMAT_ELEMENTWISE |
6585 | || memory_access_type == VMAT_STRIDED_SLP) | |
f2e2a985 MM |
6586 | { |
6587 | gimple_stmt_iterator incr_gsi; | |
6588 | bool insert_after; | |
355fe088 | 6589 | gimple *incr; |
f2e2a985 MM |
6590 | tree offvar; |
6591 | tree ivstep; | |
6592 | tree running_off; | |
f2e2a985 MM |
6593 | tree stride_base, stride_step, alias_off; |
6594 | tree vec_oprnd; | |
f502d50e | 6595 | unsigned int g; |
4d694b27 RS |
6596 | /* Checked by get_load_store_type. */ |
6597 | unsigned int const_nunits = nunits.to_constant (); | |
f2e2a985 | 6598 | |
7cfb4d93 | 6599 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
86a91c0a | 6600 | gcc_assert (!nested_in_vect_loop_p (loop, stmt_info)); |
f2e2a985 MM |
6601 | |
6602 | stride_base | |
6603 | = fold_build_pointer_plus | |
89fa689a | 6604 | (DR_BASE_ADDRESS (first_dr_info->dr), |
f2e2a985 | 6605 | size_binop (PLUS_EXPR, |
89fa689a RS |
6606 | convert_to_ptrofftype (DR_OFFSET (first_dr_info->dr)), |
6607 | convert_to_ptrofftype (DR_INIT (first_dr_info->dr)))); | |
6608 | stride_step = fold_convert (sizetype, DR_STEP (first_dr_info->dr)); | |
f2e2a985 MM |
6609 | |
6610 | /* For a store with loop-invariant (but other than power-of-2) | |
6611 | stride (i.e. not a grouped access) like so: | |
6612 | ||
6613 | for (i = 0; i < n; i += stride) | |
6614 | array[i] = ...; | |
6615 | ||
6616 | we generate a new induction variable and new stores from | |
6617 | the components of the (vectorized) rhs: | |
6618 | ||
6619 | for (j = 0; ; j += VF*stride) | |
6620 | vectemp = ...; | |
6621 | tmp1 = vectemp[0]; | |
6622 | array[j] = tmp1; | |
6623 | tmp2 = vectemp[1]; | |
6624 | array[j + stride] = tmp2; | |
6625 | ... | |
6626 | */ | |
6627 | ||
4d694b27 | 6628 | unsigned nstores = const_nunits; |
b17dc4d4 | 6629 | unsigned lnel = 1; |
cee62fee | 6630 | tree ltype = elem_type; |
04199738 | 6631 | tree lvectype = vectype; |
cee62fee MM |
6632 | if (slp) |
6633 | { | |
4d694b27 RS |
6634 | if (group_size < const_nunits |
6635 | && const_nunits % group_size == 0) | |
b17dc4d4 | 6636 | { |
4d694b27 | 6637 | nstores = const_nunits / group_size; |
b17dc4d4 RB |
6638 | lnel = group_size; |
6639 | ltype = build_vector_type (elem_type, group_size); | |
04199738 RB |
6640 | lvectype = vectype; |
6641 | ||
6642 | /* First check if vec_extract optab doesn't support extraction | |
6643 | of vector elts directly. */ | |
b397965c | 6644 | scalar_mode elmode = SCALAR_TYPE_MODE (elem_type); |
9da15d40 RS |
6645 | machine_mode vmode; |
6646 | if (!mode_for_vector (elmode, group_size).exists (&vmode) | |
6647 | || !VECTOR_MODE_P (vmode) | |
414fef4e | 6648 | || !targetm.vector_mode_supported_p (vmode) |
04199738 RB |
6649 | || (convert_optab_handler (vec_extract_optab, |
6650 | TYPE_MODE (vectype), vmode) | |
6651 | == CODE_FOR_nothing)) | |
6652 | { | |
6653 | /* Try to avoid emitting an extract of vector elements | |
6654 | by performing the extracts using an integer type of the | |
6655 | same size, extracting from a vector of those and then | |
6656 | re-interpreting it as the original vector type if | |
6657 | supported. */ | |
6658 | unsigned lsize | |
6659 | = group_size * GET_MODE_BITSIZE (elmode); | |
fffbab82 | 6660 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 6661 | unsigned int lnunits = const_nunits / group_size; |
04199738 RB |
6662 | /* If we can't construct such a vector fall back to |
6663 | element extracts from the original vector type and | |
6664 | element size stores. */ | |
4d694b27 | 6665 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 6666 | && VECTOR_MODE_P (vmode) |
414fef4e | 6667 | && targetm.vector_mode_supported_p (vmode) |
04199738 RB |
6668 | && (convert_optab_handler (vec_extract_optab, |
6669 | vmode, elmode) | |
6670 | != CODE_FOR_nothing)) | |
6671 | { | |
4d694b27 | 6672 | nstores = lnunits; |
04199738 RB |
6673 | lnel = group_size; |
6674 | ltype = build_nonstandard_integer_type (lsize, 1); | |
6675 | lvectype = build_vector_type (ltype, nstores); | |
6676 | } | |
6677 | /* Else fall back to vector extraction anyway. | |
6678 | Fewer stores are more important than avoiding spilling | |
6679 | of the vector we extract from. Compared to the | |
6680 | construction case in vectorizable_load no store-forwarding | |
6681 | issue exists here for reasonable archs. */ | |
6682 | } | |
b17dc4d4 | 6683 | } |
4d694b27 RS |
6684 | else if (group_size >= const_nunits |
6685 | && group_size % const_nunits == 0) | |
b17dc4d4 RB |
6686 | { |
6687 | nstores = 1; | |
4d694b27 | 6688 | lnel = const_nunits; |
b17dc4d4 | 6689 | ltype = vectype; |
04199738 | 6690 | lvectype = vectype; |
b17dc4d4 | 6691 | } |
cee62fee MM |
6692 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); |
6693 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
6694 | } | |
6695 | ||
f2e2a985 MM |
6696 | ivstep = stride_step; |
6697 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
b17dc4d4 | 6698 | build_int_cst (TREE_TYPE (ivstep), vf)); |
f2e2a985 MM |
6699 | |
6700 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6701 | ||
b210f45f RB |
6702 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
6703 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); | |
f2e2a985 MM |
6704 | create_iv (stride_base, ivstep, NULL, |
6705 | loop, &incr_gsi, insert_after, | |
6706 | &offvar, NULL); | |
6707 | incr = gsi_stmt (incr_gsi); | |
4fbeb363 | 6708 | loop_vinfo->add_stmt (incr); |
f2e2a985 | 6709 | |
b210f45f | 6710 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
f2e2a985 MM |
6711 | |
6712 | prev_stmt_info = NULL; | |
44fc7854 | 6713 | alias_off = build_int_cst (ref_type, 0); |
bffb8014 | 6714 | stmt_vec_info next_stmt_info = first_stmt_info; |
f502d50e | 6715 | for (g = 0; g < group_size; g++) |
f2e2a985 | 6716 | { |
f502d50e MM |
6717 | running_off = offvar; |
6718 | if (g) | |
f2e2a985 | 6719 | { |
f502d50e MM |
6720 | tree size = TYPE_SIZE_UNIT (ltype); |
6721 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 6722 | size); |
f502d50e | 6723 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 6724 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 6725 | running_off, pos); |
86a91c0a | 6726 | vect_finish_stmt_generation (stmt_info, incr, gsi); |
f2e2a985 | 6727 | running_off = newoff; |
f502d50e | 6728 | } |
b17dc4d4 RB |
6729 | unsigned int group_el = 0; |
6730 | unsigned HOST_WIDE_INT | |
6731 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
f502d50e MM |
6732 | for (j = 0; j < ncopies; j++) |
6733 | { | |
c3a8f964 | 6734 | /* We've set op and dt above, from vect_get_store_rhs, |
bffb8014 | 6735 | and first_stmt_info == stmt_info. */ |
f502d50e MM |
6736 | if (j == 0) |
6737 | { | |
6738 | if (slp) | |
6739 | { | |
86a91c0a RS |
6740 | vect_get_vec_defs (op, NULL_TREE, stmt_info, |
6741 | &vec_oprnds, NULL, slp_node); | |
f502d50e MM |
6742 | vec_oprnd = vec_oprnds[0]; |
6743 | } | |
6744 | else | |
6745 | { | |
bffb8014 RS |
6746 | op = vect_get_store_rhs (next_stmt_info); |
6747 | vec_oprnd = vect_get_vec_def_for_operand | |
6748 | (op, next_stmt_info); | |
f502d50e MM |
6749 | } |
6750 | } | |
f2e2a985 | 6751 | else |
f502d50e MM |
6752 | { |
6753 | if (slp) | |
6754 | vec_oprnd = vec_oprnds[j]; | |
6755 | else | |
e4057a39 RS |
6756 | vec_oprnd = vect_get_vec_def_for_stmt_copy (vinfo, |
6757 | vec_oprnd); | |
f502d50e | 6758 | } |
04199738 RB |
6759 | /* Pun the vector to extract from if necessary. */ |
6760 | if (lvectype != vectype) | |
6761 | { | |
6762 | tree tem = make_ssa_name (lvectype); | |
6763 | gimple *pun | |
6764 | = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
6765 | lvectype, vec_oprnd)); | |
86a91c0a | 6766 | vect_finish_stmt_generation (stmt_info, pun, gsi); |
04199738 RB |
6767 | vec_oprnd = tem; |
6768 | } | |
f502d50e MM |
6769 | for (i = 0; i < nstores; i++) |
6770 | { | |
6771 | tree newref, newoff; | |
355fe088 | 6772 | gimple *incr, *assign; |
f502d50e MM |
6773 | tree size = TYPE_SIZE (ltype); |
6774 | /* Extract the i'th component. */ | |
6775 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
6776 | bitsize_int (i), size); | |
6777 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
6778 | size, pos); | |
6779 | ||
6780 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
6781 | NULL_TREE, true, | |
6782 | GSI_SAME_STMT); | |
6783 | ||
b17dc4d4 RB |
6784 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
6785 | group_el * elsz); | |
f502d50e | 6786 | newref = build2 (MEM_REF, ltype, |
b17dc4d4 | 6787 | running_off, this_off); |
89fa689a | 6788 | vect_copy_ref_info (newref, DR_REF (first_dr_info->dr)); |
f502d50e MM |
6789 | |
6790 | /* And store it to *running_off. */ | |
6791 | assign = gimple_build_assign (newref, elem); | |
e1bd7296 | 6792 | stmt_vec_info assign_info |
86a91c0a | 6793 | = vect_finish_stmt_generation (stmt_info, assign, gsi); |
f502d50e | 6794 | |
b17dc4d4 RB |
6795 | group_el += lnel; |
6796 | if (! slp | |
6797 | || group_el == group_size) | |
6798 | { | |
6799 | newoff = copy_ssa_name (running_off, NULL); | |
6800 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6801 | running_off, stride_step); | |
86a91c0a | 6802 | vect_finish_stmt_generation (stmt_info, incr, gsi); |
f502d50e | 6803 | |
b17dc4d4 RB |
6804 | running_off = newoff; |
6805 | group_el = 0; | |
6806 | } | |
225ce44b RB |
6807 | if (g == group_size - 1 |
6808 | && !slp) | |
f502d50e MM |
6809 | { |
6810 | if (j == 0 && i == 0) | |
225ce44b | 6811 | STMT_VINFO_VEC_STMT (stmt_info) |
e1bd7296 | 6812 | = *vec_stmt = assign_info; |
f502d50e | 6813 | else |
e1bd7296 RS |
6814 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign_info; |
6815 | prev_stmt_info = assign_info; | |
f502d50e MM |
6816 | } |
6817 | } | |
f2e2a985 | 6818 | } |
bffb8014 | 6819 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
b17dc4d4 RB |
6820 | if (slp) |
6821 | break; | |
f2e2a985 | 6822 | } |
778dd3b6 RB |
6823 | |
6824 | vec_oprnds.release (); | |
f2e2a985 MM |
6825 | return true; |
6826 | } | |
6827 | ||
8c681247 | 6828 | auto_vec<tree> dr_chain (group_size); |
9771b263 | 6829 | oprnds.create (group_size); |
ebfd146a | 6830 | |
89fa689a RS |
6831 | alignment_support_scheme |
6832 | = vect_supportable_dr_alignment (first_dr_info, false); | |
ebfd146a | 6833 | gcc_assert (alignment_support_scheme); |
70088b95 RS |
6834 | vec_loop_masks *loop_masks |
6835 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) | |
6836 | ? &LOOP_VINFO_MASKS (loop_vinfo) | |
6837 | : NULL); | |
272c6793 | 6838 | /* Targets with store-lane instructions must not require explicit |
c3a8f964 RS |
6839 | realignment. vect_supportable_dr_alignment always returns either |
6840 | dr_aligned or dr_unaligned_supported for masked operations. */ | |
7cfb4d93 RS |
6841 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES |
6842 | && !mask | |
70088b95 | 6843 | && !loop_masks) |
272c6793 RS |
6844 | || alignment_support_scheme == dr_aligned |
6845 | || alignment_support_scheme == dr_unaligned_supported); | |
6846 | ||
62da9e14 RS |
6847 | if (memory_access_type == VMAT_CONTIGUOUS_DOWN |
6848 | || memory_access_type == VMAT_CONTIGUOUS_REVERSE) | |
09dfa495 BM |
6849 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
6850 | ||
f307441a RS |
6851 | tree bump; |
6852 | tree vec_offset = NULL_TREE; | |
6853 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
6854 | { | |
6855 | aggr_type = NULL_TREE; | |
6856 | bump = NULL_TREE; | |
6857 | } | |
6858 | else if (memory_access_type == VMAT_GATHER_SCATTER) | |
6859 | { | |
6860 | aggr_type = elem_type; | |
86a91c0a | 6861 | vect_get_strided_load_store_ops (stmt_info, loop_vinfo, &gs_info, |
f307441a RS |
6862 | &bump, &vec_offset); |
6863 | } | |
272c6793 | 6864 | else |
f307441a RS |
6865 | { |
6866 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
6867 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
6868 | else | |
6869 | aggr_type = vectype; | |
89fa689a RS |
6870 | bump = vect_get_data_ptr_increment (dr_info, aggr_type, |
6871 | memory_access_type); | |
f307441a | 6872 | } |
ebfd146a | 6873 | |
c3a8f964 RS |
6874 | if (mask) |
6875 | LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true; | |
6876 | ||
ebfd146a IR |
6877 | /* In case the vectorization factor (VF) is bigger than the number |
6878 | of elements that we can fit in a vectype (nunits), we have to generate | |
6879 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 6880 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
6881 | vect_get_vec_def_for_copy_stmt. */ |
6882 | ||
0d0293ac | 6883 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6884 | |
6885 | S1: &base + 2 = x2 | |
6886 | S2: &base = x0 | |
6887 | S3: &base + 1 = x1 | |
6888 | S4: &base + 3 = x3 | |
6889 | ||
6890 | We create vectorized stores starting from base address (the access of the | |
6891 | first stmt in the chain (S2 in the above example), when the last store stmt | |
6892 | of the chain (S4) is reached: | |
6893 | ||
6894 | VS1: &base = vx2 | |
6895 | VS2: &base + vec_size*1 = vx0 | |
6896 | VS3: &base + vec_size*2 = vx1 | |
6897 | VS4: &base + vec_size*3 = vx3 | |
6898 | ||
6899 | Then permutation statements are generated: | |
6900 | ||
3fcc1b55 JJ |
6901 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
6902 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 6903 | ... |
b8698a0f | 6904 | |
ebfd146a IR |
6905 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
6906 | (the order of the data-refs in the output of vect_permute_store_chain | |
6907 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6908 | the documentation of vect_permute_store_chain()). | |
6909 | ||
6910 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 6911 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 6912 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 6913 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
6914 | */ |
6915 | ||
6916 | prev_stmt_info = NULL; | |
c3a8f964 | 6917 | tree vec_mask = NULL_TREE; |
ebfd146a IR |
6918 | for (j = 0; j < ncopies; j++) |
6919 | { | |
e1bd7296 | 6920 | stmt_vec_info new_stmt_info; |
ebfd146a IR |
6921 | if (j == 0) |
6922 | { | |
6923 | if (slp) | |
6924 | { | |
6925 | /* Get vectorized arguments for SLP_NODE. */ | |
86a91c0a RS |
6926 | vect_get_vec_defs (op, NULL_TREE, stmt_info, &vec_oprnds, |
6927 | NULL, slp_node); | |
ebfd146a | 6928 | |
9771b263 | 6929 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
6930 | } |
6931 | else | |
6932 | { | |
b8698a0f L |
6933 | /* For interleaved stores we collect vectorized defs for all the |
6934 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
6935 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
6936 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
6937 | ||
2c53b149 | 6938 | If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 6939 | OPRNDS are of size 1. */ |
bffb8014 | 6940 | stmt_vec_info next_stmt_info = first_stmt_info; |
ebfd146a IR |
6941 | for (i = 0; i < group_size; i++) |
6942 | { | |
b8698a0f | 6943 | /* Since gaps are not supported for interleaved stores, |
2c53b149 | 6944 | DR_GROUP_SIZE is the exact number of stmts in the chain. |
bffb8014 RS |
6945 | Therefore, NEXT_STMT_INFO can't be NULL_TREE. In case |
6946 | that there is no interleaving, DR_GROUP_SIZE is 1, | |
6947 | and only one iteration of the loop will be executed. */ | |
6948 | op = vect_get_store_rhs (next_stmt_info); | |
6949 | vec_oprnd = vect_get_vec_def_for_operand | |
6950 | (op, next_stmt_info); | |
9771b263 DN |
6951 | dr_chain.quick_push (vec_oprnd); |
6952 | oprnds.quick_push (vec_oprnd); | |
bffb8014 | 6953 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
ebfd146a | 6954 | } |
c3a8f964 | 6955 | if (mask) |
86a91c0a | 6956 | vec_mask = vect_get_vec_def_for_operand (mask, stmt_info, |
c3a8f964 | 6957 | mask_vectype); |
ebfd146a IR |
6958 | } |
6959 | ||
6960 | /* We should have catched mismatched types earlier. */ | |
6961 | gcc_assert (useless_type_conversion_p (vectype, | |
6962 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
6963 | bool simd_lane_access_p |
6964 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
6965 | if (simd_lane_access_p | |
89fa689a RS |
6966 | && TREE_CODE (DR_BASE_ADDRESS (first_dr_info->dr)) == ADDR_EXPR |
6967 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr_info->dr), 0)) | |
6968 | && integer_zerop (DR_OFFSET (first_dr_info->dr)) | |
6969 | && integer_zerop (DR_INIT (first_dr_info->dr)) | |
74bf76ed | 6970 | && alias_sets_conflict_p (get_alias_set (aggr_type), |
44fc7854 | 6971 | get_alias_set (TREE_TYPE (ref_type)))) |
74bf76ed | 6972 | { |
89fa689a | 6973 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr_info->dr)); |
44fc7854 | 6974 | dataref_offset = build_int_cst (ref_type, 0); |
74bf76ed | 6975 | } |
f307441a | 6976 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
2d4bca81 RS |
6977 | vect_get_gather_scatter_ops (loop, stmt_info, &gs_info, |
6978 | &dataref_ptr, &vec_offset); | |
74bf76ed JJ |
6979 | else |
6980 | dataref_ptr | |
bffb8014 | 6981 | = vect_create_data_ref_ptr (first_stmt_info, aggr_type, |
74bf76ed | 6982 | simd_lane_access_p ? loop : NULL, |
09dfa495 | 6983 | offset, &dummy, gsi, &ptr_incr, |
2d4bca81 | 6984 | simd_lane_access_p, NULL_TREE, bump); |
ebfd146a | 6985 | } |
b8698a0f | 6986 | else |
ebfd146a | 6987 | { |
b8698a0f L |
6988 | /* For interleaved stores we created vectorized defs for all the |
6989 | defs stored in OPRNDS in the previous iteration (previous copy). | |
6990 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
6991 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
6992 | next copy. | |
2c53b149 | 6993 | If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
6994 | OPRNDS are of size 1. */ |
6995 | for (i = 0; i < group_size; i++) | |
6996 | { | |
9771b263 | 6997 | op = oprnds[i]; |
e4057a39 | 6998 | vec_oprnd = vect_get_vec_def_for_stmt_copy (vinfo, op); |
9771b263 DN |
6999 | dr_chain[i] = vec_oprnd; |
7000 | oprnds[i] = vec_oprnd; | |
ebfd146a | 7001 | } |
c3a8f964 | 7002 | if (mask) |
e4057a39 | 7003 | vec_mask = vect_get_vec_def_for_stmt_copy (vinfo, vec_mask); |
74bf76ed JJ |
7004 | if (dataref_offset) |
7005 | dataref_offset | |
f307441a RS |
7006 | = int_const_binop (PLUS_EXPR, dataref_offset, bump); |
7007 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
e4057a39 | 7008 | vec_offset = vect_get_vec_def_for_stmt_copy (vinfo, vec_offset); |
74bf76ed | 7009 | else |
86a91c0a RS |
7010 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
7011 | stmt_info, bump); | |
ebfd146a IR |
7012 | } |
7013 | ||
2de001ee | 7014 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 7015 | { |
272c6793 | 7016 | tree vec_array; |
267d3070 | 7017 | |
3ba4ff41 | 7018 | /* Get an array into which we can store the individual vectors. */ |
272c6793 | 7019 | vec_array = create_vector_array (vectype, vec_num); |
3ba4ff41 RS |
7020 | |
7021 | /* Invalidate the current contents of VEC_ARRAY. This should | |
7022 | become an RTL clobber too, which prevents the vector registers | |
7023 | from being upward-exposed. */ | |
86a91c0a | 7024 | vect_clobber_variable (stmt_info, gsi, vec_array); |
3ba4ff41 RS |
7025 | |
7026 | /* Store the individual vectors into the array. */ | |
272c6793 | 7027 | for (i = 0; i < vec_num; i++) |
c2d7ab2a | 7028 | { |
9771b263 | 7029 | vec_oprnd = dr_chain[i]; |
86a91c0a | 7030 | write_vector_array (stmt_info, gsi, vec_oprnd, vec_array, i); |
267d3070 | 7031 | } |
b8698a0f | 7032 | |
7cfb4d93 | 7033 | tree final_mask = NULL; |
70088b95 RS |
7034 | if (loop_masks) |
7035 | final_mask = vect_get_loop_mask (gsi, loop_masks, ncopies, | |
7036 | vectype, j); | |
7cfb4d93 RS |
7037 | if (vec_mask) |
7038 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
7039 | vec_mask, gsi); | |
7040 | ||
7e11fc7f | 7041 | gcall *call; |
7cfb4d93 | 7042 | if (final_mask) |
7e11fc7f RS |
7043 | { |
7044 | /* Emit: | |
7045 | MASK_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK, | |
7046 | VEC_ARRAY). */ | |
7047 | unsigned int align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype)); | |
7048 | tree alias_ptr = build_int_cst (ref_type, align); | |
7049 | call = gimple_build_call_internal (IFN_MASK_STORE_LANES, 4, | |
7050 | dataref_ptr, alias_ptr, | |
7cfb4d93 | 7051 | final_mask, vec_array); |
7e11fc7f RS |
7052 | } |
7053 | else | |
7054 | { | |
7055 | /* Emit: | |
7056 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
7057 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); | |
7058 | call = gimple_build_call_internal (IFN_STORE_LANES, 1, | |
7059 | vec_array); | |
7060 | gimple_call_set_lhs (call, data_ref); | |
7061 | } | |
a844293d | 7062 | gimple_call_set_nothrow (call, true); |
86a91c0a | 7063 | new_stmt_info = vect_finish_stmt_generation (stmt_info, call, gsi); |
3ba4ff41 RS |
7064 | |
7065 | /* Record that VEC_ARRAY is now dead. */ | |
86a91c0a | 7066 | vect_clobber_variable (stmt_info, gsi, vec_array); |
272c6793 RS |
7067 | } |
7068 | else | |
7069 | { | |
e1bd7296 | 7070 | new_stmt_info = NULL; |
0d0293ac | 7071 | if (grouped_store) |
272c6793 | 7072 | { |
b6b9227d JJ |
7073 | if (j == 0) |
7074 | result_chain.create (group_size); | |
272c6793 | 7075 | /* Permute. */ |
86a91c0a | 7076 | vect_permute_store_chain (dr_chain, group_size, stmt_info, gsi, |
272c6793 RS |
7077 | &result_chain); |
7078 | } | |
c2d7ab2a | 7079 | |
bffb8014 | 7080 | stmt_vec_info next_stmt_info = first_stmt_info; |
272c6793 RS |
7081 | for (i = 0; i < vec_num; i++) |
7082 | { | |
644ffefd | 7083 | unsigned align, misalign; |
272c6793 | 7084 | |
7cfb4d93 | 7085 | tree final_mask = NULL_TREE; |
70088b95 RS |
7086 | if (loop_masks) |
7087 | final_mask = vect_get_loop_mask (gsi, loop_masks, | |
7088 | vec_num * ncopies, | |
7cfb4d93 RS |
7089 | vectype, vec_num * j + i); |
7090 | if (vec_mask) | |
7091 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
7092 | vec_mask, gsi); | |
7093 | ||
f307441a RS |
7094 | if (memory_access_type == VMAT_GATHER_SCATTER) |
7095 | { | |
7096 | tree scale = size_int (gs_info.scale); | |
7097 | gcall *call; | |
70088b95 | 7098 | if (loop_masks) |
f307441a RS |
7099 | call = gimple_build_call_internal |
7100 | (IFN_MASK_SCATTER_STORE, 5, dataref_ptr, vec_offset, | |
7101 | scale, vec_oprnd, final_mask); | |
7102 | else | |
7103 | call = gimple_build_call_internal | |
7104 | (IFN_SCATTER_STORE, 4, dataref_ptr, vec_offset, | |
7105 | scale, vec_oprnd); | |
7106 | gimple_call_set_nothrow (call, true); | |
e1bd7296 | 7107 | new_stmt_info |
86a91c0a | 7108 | = vect_finish_stmt_generation (stmt_info, call, gsi); |
f307441a RS |
7109 | break; |
7110 | } | |
7111 | ||
272c6793 RS |
7112 | if (i > 0) |
7113 | /* Bump the vector pointer. */ | |
7114 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
86a91c0a | 7115 | stmt_info, bump); |
272c6793 RS |
7116 | |
7117 | if (slp) | |
9771b263 | 7118 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
7119 | else if (grouped_store) |
7120 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 7121 | vect_permute_store_chain(). */ |
9771b263 | 7122 | vec_oprnd = result_chain[i]; |
272c6793 | 7123 | |
89fa689a RS |
7124 | align = DR_TARGET_ALIGNMENT (first_dr_info); |
7125 | if (aligned_access_p (first_dr_info)) | |
644ffefd | 7126 | misalign = 0; |
89fa689a | 7127 | else if (DR_MISALIGNMENT (first_dr_info) == -1) |
272c6793 | 7128 | { |
89fa689a | 7129 | align = dr_alignment (vect_dr_behavior (first_dr_info)); |
52639a61 | 7130 | misalign = 0; |
272c6793 RS |
7131 | } |
7132 | else | |
89fa689a | 7133 | misalign = DR_MISALIGNMENT (first_dr_info); |
aed93b23 RB |
7134 | if (dataref_offset == NULL_TREE |
7135 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
7136 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
7137 | misalign); | |
c2d7ab2a | 7138 | |
62da9e14 | 7139 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
09dfa495 BM |
7140 | { |
7141 | tree perm_mask = perm_mask_for_reverse (vectype); | |
86a91c0a RS |
7142 | tree perm_dest = vect_create_destination_var |
7143 | (vect_get_store_rhs (stmt_info), vectype); | |
b731b390 | 7144 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
7145 | |
7146 | /* Generate the permute statement. */ | |
355fe088 | 7147 | gimple *perm_stmt |
0d0e4a03 JJ |
7148 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
7149 | vec_oprnd, perm_mask); | |
86a91c0a | 7150 | vect_finish_stmt_generation (stmt_info, perm_stmt, gsi); |
09dfa495 BM |
7151 | |
7152 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
7153 | vec_oprnd = new_temp; | |
7154 | } | |
7155 | ||
272c6793 | 7156 | /* Arguments are ready. Create the new vector stmt. */ |
7cfb4d93 | 7157 | if (final_mask) |
c3a8f964 RS |
7158 | { |
7159 | align = least_bit_hwi (misalign | align); | |
7160 | tree ptr = build_int_cst (ref_type, align); | |
7161 | gcall *call | |
7162 | = gimple_build_call_internal (IFN_MASK_STORE, 4, | |
7163 | dataref_ptr, ptr, | |
7cfb4d93 | 7164 | final_mask, vec_oprnd); |
c3a8f964 | 7165 | gimple_call_set_nothrow (call, true); |
e1bd7296 | 7166 | new_stmt_info |
86a91c0a | 7167 | = vect_finish_stmt_generation (stmt_info, call, gsi); |
c3a8f964 RS |
7168 | } |
7169 | else | |
7170 | { | |
7171 | data_ref = fold_build2 (MEM_REF, vectype, | |
7172 | dataref_ptr, | |
7173 | dataref_offset | |
7174 | ? dataref_offset | |
7175 | : build_int_cst (ref_type, 0)); | |
89fa689a | 7176 | if (aligned_access_p (first_dr_info)) |
c3a8f964 | 7177 | ; |
89fa689a | 7178 | else if (DR_MISALIGNMENT (first_dr_info) == -1) |
c3a8f964 RS |
7179 | TREE_TYPE (data_ref) |
7180 | = build_aligned_type (TREE_TYPE (data_ref), | |
7181 | align * BITS_PER_UNIT); | |
7182 | else | |
7183 | TREE_TYPE (data_ref) | |
7184 | = build_aligned_type (TREE_TYPE (data_ref), | |
7185 | TYPE_ALIGN (elem_type)); | |
89fa689a | 7186 | vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr)); |
e1bd7296 RS |
7187 | gassign *new_stmt |
7188 | = gimple_build_assign (data_ref, vec_oprnd); | |
7189 | new_stmt_info | |
86a91c0a | 7190 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
c3a8f964 | 7191 | } |
272c6793 RS |
7192 | |
7193 | if (slp) | |
7194 | continue; | |
7195 | ||
bffb8014 RS |
7196 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
7197 | if (!next_stmt_info) | |
272c6793 RS |
7198 | break; |
7199 | } | |
ebfd146a | 7200 | } |
1da0876c RS |
7201 | if (!slp) |
7202 | { | |
7203 | if (j == 0) | |
e1bd7296 | 7204 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
1da0876c | 7205 | else |
e1bd7296 RS |
7206 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
7207 | prev_stmt_info = new_stmt_info; | |
1da0876c | 7208 | } |
ebfd146a IR |
7209 | } |
7210 | ||
9771b263 DN |
7211 | oprnds.release (); |
7212 | result_chain.release (); | |
7213 | vec_oprnds.release (); | |
ebfd146a IR |
7214 | |
7215 | return true; | |
7216 | } | |
7217 | ||
557be5a8 AL |
7218 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
7219 | VECTOR_CST mask. No checks are made that the target platform supports the | |
7ac7e286 | 7220 | mask, so callers may wish to test can_vec_perm_const_p separately, or use |
557be5a8 | 7221 | vect_gen_perm_mask_checked. */ |
a1e53f3f | 7222 | |
3fcc1b55 | 7223 | tree |
4aae3cb3 | 7224 | vect_gen_perm_mask_any (tree vectype, const vec_perm_indices &sel) |
a1e53f3f | 7225 | { |
b00cb3bf | 7226 | tree mask_type; |
a1e53f3f | 7227 | |
0ecc2b7d RS |
7228 | poly_uint64 nunits = sel.length (); |
7229 | gcc_assert (known_eq (nunits, TYPE_VECTOR_SUBPARTS (vectype))); | |
b00cb3bf RS |
7230 | |
7231 | mask_type = build_vector_type (ssizetype, nunits); | |
736d0f28 | 7232 | return vec_perm_indices_to_tree (mask_type, sel); |
a1e53f3f L |
7233 | } |
7234 | ||
7ac7e286 | 7235 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_const_p, |
cf7aa6a3 | 7236 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ |
557be5a8 AL |
7237 | |
7238 | tree | |
4aae3cb3 | 7239 | vect_gen_perm_mask_checked (tree vectype, const vec_perm_indices &sel) |
557be5a8 | 7240 | { |
7ac7e286 | 7241 | gcc_assert (can_vec_perm_const_p (TYPE_MODE (vectype), sel)); |
557be5a8 AL |
7242 | return vect_gen_perm_mask_any (vectype, sel); |
7243 | } | |
7244 | ||
aec7ae7d | 7245 | /* Given a vector variable X and Y, that was generated for the scalar |
82570274 | 7246 | STMT_INFO, generate instructions to permute the vector elements of X and Y |
aec7ae7d JJ |
7247 | using permutation mask MASK_VEC, insert them at *GSI and return the |
7248 | permuted vector variable. */ | |
a1e53f3f L |
7249 | |
7250 | static tree | |
82570274 | 7251 | permute_vec_elements (tree x, tree y, tree mask_vec, stmt_vec_info stmt_info, |
aec7ae7d | 7252 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
7253 | { |
7254 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 7255 | tree perm_dest, data_ref; |
355fe088 | 7256 | gimple *perm_stmt; |
a1e53f3f | 7257 | |
82570274 | 7258 | tree scalar_dest = gimple_get_lhs (stmt_info->stmt); |
7ad429a4 RS |
7259 | if (TREE_CODE (scalar_dest) == SSA_NAME) |
7260 | perm_dest = vect_create_destination_var (scalar_dest, vectype); | |
7261 | else | |
7262 | perm_dest = vect_get_new_vect_var (vectype, vect_simple_var, NULL); | |
b731b390 | 7263 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
7264 | |
7265 | /* Generate the permute statement. */ | |
0d0e4a03 | 7266 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
82570274 | 7267 | vect_finish_stmt_generation (stmt_info, perm_stmt, gsi); |
a1e53f3f L |
7268 | |
7269 | return data_ref; | |
7270 | } | |
7271 | ||
32e8e429 | 7272 | /* Hoist the definitions of all SSA uses on STMT_INFO out of the loop LOOP, |
6b916b36 | 7273 | inserting them on the loops preheader edge. Returns true if we |
32e8e429 | 7274 | were successful in doing so (and thus STMT_INFO can be moved then), |
6b916b36 RB |
7275 | otherwise returns false. */ |
7276 | ||
7277 | static bool | |
32e8e429 | 7278 | hoist_defs_of_uses (stmt_vec_info stmt_info, struct loop *loop) |
6b916b36 RB |
7279 | { |
7280 | ssa_op_iter i; | |
7281 | tree op; | |
7282 | bool any = false; | |
7283 | ||
32e8e429 | 7284 | FOR_EACH_SSA_TREE_OPERAND (op, stmt_info->stmt, i, SSA_OP_USE) |
6b916b36 | 7285 | { |
355fe088 | 7286 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
7287 | if (!gimple_nop_p (def_stmt) |
7288 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
7289 | { | |
7290 | /* Make sure we don't need to recurse. While we could do | |
7291 | so in simple cases when there are more complex use webs | |
7292 | we don't have an easy way to preserve stmt order to fulfil | |
7293 | dependencies within them. */ | |
7294 | tree op2; | |
7295 | ssa_op_iter i2; | |
d1417442 JJ |
7296 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
7297 | return false; | |
6b916b36 RB |
7298 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
7299 | { | |
355fe088 | 7300 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
7301 | if (!gimple_nop_p (def_stmt2) |
7302 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
7303 | return false; | |
7304 | } | |
7305 | any = true; | |
7306 | } | |
7307 | } | |
7308 | ||
7309 | if (!any) | |
7310 | return true; | |
7311 | ||
32e8e429 | 7312 | FOR_EACH_SSA_TREE_OPERAND (op, stmt_info->stmt, i, SSA_OP_USE) |
6b916b36 | 7313 | { |
355fe088 | 7314 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
7315 | if (!gimple_nop_p (def_stmt) |
7316 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
7317 | { | |
7318 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
7319 | gsi_remove (&gsi, false); | |
7320 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
7321 | } | |
7322 | } | |
7323 | ||
7324 | return true; | |
7325 | } | |
7326 | ||
ebfd146a IR |
7327 | /* vectorizable_load. |
7328 | ||
32e8e429 RS |
7329 | Check if STMT_INFO reads a non scalar data-ref (array/pointer/structure) |
7330 | that can be vectorized. | |
7331 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized | |
7332 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
7333 | Return true if STMT_INFO is vectorizable in this way. */ | |
ebfd146a IR |
7334 | |
7335 | static bool | |
32e8e429 | 7336 | vectorizable_load (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
1eede195 RS |
7337 | stmt_vec_info *vec_stmt, slp_tree slp_node, |
7338 | slp_instance slp_node_instance, | |
68435eb2 | 7339 | stmt_vector_for_cost *cost_vec) |
ebfd146a IR |
7340 | { |
7341 | tree scalar_dest; | |
7342 | tree vec_dest = NULL; | |
7343 | tree data_ref = NULL; | |
b8698a0f | 7344 | stmt_vec_info prev_stmt_info; |
ebfd146a | 7345 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 7346 | struct loop *loop = NULL; |
32e8e429 | 7347 | struct loop *containing_loop = gimple_bb (stmt_info->stmt)->loop_father; |
a70d6342 | 7348 | bool nested_in_vect_loop = false; |
272c6793 | 7349 | tree elem_type; |
ebfd146a | 7350 | tree new_temp; |
ef4bddc2 | 7351 | machine_mode mode; |
ebfd146a IR |
7352 | tree dummy; |
7353 | enum dr_alignment_support alignment_support_scheme; | |
7354 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 7355 | tree dataref_offset = NULL_TREE; |
355fe088 | 7356 | gimple *ptr_incr = NULL; |
ebfd146a | 7357 | int ncopies; |
4d694b27 RS |
7358 | int i, j; |
7359 | unsigned int group_size; | |
7360 | poly_uint64 group_gap_adj; | |
ebfd146a IR |
7361 | tree msq = NULL_TREE, lsq; |
7362 | tree offset = NULL_TREE; | |
356bbc4c | 7363 | tree byte_offset = NULL_TREE; |
ebfd146a | 7364 | tree realignment_token = NULL_TREE; |
538dd0b7 | 7365 | gphi *phi = NULL; |
6e1aa848 | 7366 | vec<tree> dr_chain = vNULL; |
0d0293ac | 7367 | bool grouped_load = false; |
bffb8014 | 7368 | stmt_vec_info first_stmt_info; |
b9787581 | 7369 | stmt_vec_info first_stmt_info_for_drptr = NULL; |
ebfd146a IR |
7370 | bool compute_in_loop = false; |
7371 | struct loop *at_loop; | |
7372 | int vec_num; | |
7373 | bool slp = (slp_node != NULL); | |
7374 | bool slp_perm = false; | |
a70d6342 | 7375 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
d9f21f6a | 7376 | poly_uint64 vf; |
272c6793 | 7377 | tree aggr_type; |
134c85ca | 7378 | gather_scatter_info gs_info; |
310213d4 | 7379 | vec_info *vinfo = stmt_info->vinfo; |
44fc7854 | 7380 | tree ref_type; |
929b4411 | 7381 | enum vect_def_type mask_dt = vect_unknown_def_type; |
a70d6342 | 7382 | |
465c8c19 JJ |
7383 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
7384 | return false; | |
7385 | ||
66c16fd9 RB |
7386 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7387 | && ! vec_stmt) | |
465c8c19 JJ |
7388 | return false; |
7389 | ||
c3a8f964 | 7390 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
86a91c0a | 7391 | if (gassign *assign = dyn_cast <gassign *> (stmt_info->stmt)) |
c3a8f964 | 7392 | { |
beb456c3 | 7393 | scalar_dest = gimple_assign_lhs (assign); |
c3a8f964 RS |
7394 | if (TREE_CODE (scalar_dest) != SSA_NAME) |
7395 | return false; | |
465c8c19 | 7396 | |
beb456c3 | 7397 | tree_code code = gimple_assign_rhs_code (assign); |
c3a8f964 RS |
7398 | if (code != ARRAY_REF |
7399 | && code != BIT_FIELD_REF | |
7400 | && code != INDIRECT_REF | |
7401 | && code != COMPONENT_REF | |
7402 | && code != IMAGPART_EXPR | |
7403 | && code != REALPART_EXPR | |
7404 | && code != MEM_REF | |
7405 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
7406 | return false; | |
7407 | } | |
7408 | else | |
7409 | { | |
86a91c0a | 7410 | gcall *call = dyn_cast <gcall *> (stmt_info->stmt); |
bfaa08b7 RS |
7411 | if (!call || !gimple_call_internal_p (call)) |
7412 | return false; | |
7413 | ||
7414 | internal_fn ifn = gimple_call_internal_fn (call); | |
7415 | if (!internal_load_fn_p (ifn)) | |
c3a8f964 | 7416 | return false; |
465c8c19 | 7417 | |
c3a8f964 RS |
7418 | scalar_dest = gimple_call_lhs (call); |
7419 | if (!scalar_dest) | |
7420 | return false; | |
7421 | ||
7422 | if (slp_node != NULL) | |
7423 | { | |
7424 | if (dump_enabled_p ()) | |
7425 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7426 | "SLP of masked loads not supported.\n"); | |
7427 | return false; | |
7428 | } | |
7429 | ||
bfaa08b7 RS |
7430 | int mask_index = internal_fn_mask_index (ifn); |
7431 | if (mask_index >= 0) | |
7432 | { | |
7433 | mask = gimple_call_arg (call, mask_index); | |
86a91c0a | 7434 | if (!vect_check_load_store_mask (stmt_info, mask, &mask_dt, |
929b4411 | 7435 | &mask_vectype)) |
bfaa08b7 RS |
7436 | return false; |
7437 | } | |
c3a8f964 | 7438 | } |
465c8c19 JJ |
7439 | |
7440 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
7441 | return false; | |
7442 | ||
7443 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
4d694b27 | 7444 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 | 7445 | |
a70d6342 IR |
7446 | if (loop_vinfo) |
7447 | { | |
7448 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
86a91c0a | 7449 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt_info); |
a70d6342 IR |
7450 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); |
7451 | } | |
7452 | else | |
3533e503 | 7453 | vf = 1; |
ebfd146a IR |
7454 | |
7455 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 7456 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 7457 | case of SLP. */ |
fce57248 | 7458 | if (slp) |
ebfd146a IR |
7459 | ncopies = 1; |
7460 | else | |
e8f142e2 | 7461 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
ebfd146a IR |
7462 | |
7463 | gcc_assert (ncopies >= 1); | |
7464 | ||
7465 | /* FORNOW. This restriction should be relaxed. */ | |
7466 | if (nested_in_vect_loop && ncopies > 1) | |
7467 | { | |
73fbfcad | 7468 | if (dump_enabled_p ()) |
78c60e3d | 7469 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7470 | "multiple types in nested loop.\n"); |
ebfd146a IR |
7471 | return false; |
7472 | } | |
7473 | ||
f2556b68 RB |
7474 | /* Invalidate assumptions made by dependence analysis when vectorization |
7475 | on the unrolled body effectively re-orders stmts. */ | |
7476 | if (ncopies > 1 | |
7477 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
7478 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
7479 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
7480 | { |
7481 | if (dump_enabled_p ()) | |
7482 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7483 | "cannot perform implicit CSE when unrolling " | |
7484 | "with negative dependence distance\n"); | |
7485 | return false; | |
7486 | } | |
7487 | ||
7b7b1813 | 7488 | elem_type = TREE_TYPE (vectype); |
947131ba | 7489 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
7490 | |
7491 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
7492 | (e.g. - data copies). */ | |
947131ba | 7493 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 7494 | { |
73fbfcad | 7495 | if (dump_enabled_p ()) |
78c60e3d | 7496 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7497 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
7498 | return false; |
7499 | } | |
7500 | ||
ebfd146a | 7501 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 7502 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 7503 | { |
0d0293ac | 7504 | grouped_load = true; |
ebfd146a | 7505 | /* FORNOW */ |
2de001ee RS |
7506 | gcc_assert (!nested_in_vect_loop); |
7507 | gcc_assert (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)); | |
ebfd146a | 7508 | |
bffb8014 RS |
7509 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
7510 | group_size = DR_GROUP_SIZE (first_stmt_info); | |
d5f035ea | 7511 | |
b1af7da6 RB |
7512 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
7513 | slp_perm = true; | |
7514 | ||
f2556b68 RB |
7515 | /* Invalidate assumptions made by dependence analysis when vectorization |
7516 | on the unrolled body effectively re-orders stmts. */ | |
7517 | if (!PURE_SLP_STMT (stmt_info) | |
7518 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
7519 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
7520 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
7521 | { |
7522 | if (dump_enabled_p ()) | |
7523 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7524 | "cannot perform implicit CSE when performing " | |
7525 | "group loads with negative dependence distance\n"); | |
7526 | return false; | |
7527 | } | |
96bb56b2 RB |
7528 | |
7529 | /* Similarly when the stmt is a load that is both part of a SLP | |
7530 | instance and a loop vectorized stmt via the same-dr mechanism | |
7531 | we have to give up. */ | |
2c53b149 | 7532 | if (DR_GROUP_SAME_DR_STMT (stmt_info) |
96bb56b2 | 7533 | && (STMT_SLP_TYPE (stmt_info) |
c26228d4 | 7534 | != STMT_SLP_TYPE (DR_GROUP_SAME_DR_STMT (stmt_info)))) |
96bb56b2 RB |
7535 | { |
7536 | if (dump_enabled_p ()) | |
7537 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7538 | "conflicting SLP types for CSEd load\n"); | |
7539 | return false; | |
7540 | } | |
ebfd146a | 7541 | } |
7cfb4d93 RS |
7542 | else |
7543 | group_size = 1; | |
ebfd146a | 7544 | |
2de001ee | 7545 | vect_memory_access_type memory_access_type; |
86a91c0a | 7546 | if (!get_load_store_type (stmt_info, vectype, slp, mask, VLS_LOAD, ncopies, |
2de001ee RS |
7547 | &memory_access_type, &gs_info)) |
7548 | return false; | |
a1e53f3f | 7549 | |
c3a8f964 RS |
7550 | if (mask) |
7551 | { | |
7552 | if (memory_access_type == VMAT_CONTIGUOUS) | |
7553 | { | |
7e11fc7f RS |
7554 | machine_mode vec_mode = TYPE_MODE (vectype); |
7555 | if (!VECTOR_MODE_P (vec_mode) | |
7556 | || !can_vec_mask_load_store_p (vec_mode, | |
c3a8f964 RS |
7557 | TYPE_MODE (mask_vectype), true)) |
7558 | return false; | |
7559 | } | |
bfaa08b7 | 7560 | else if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
c3a8f964 RS |
7561 | { |
7562 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); | |
7563 | tree masktype | |
7564 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
7565 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
7566 | { | |
7567 | if (dump_enabled_p ()) | |
7568 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7569 | "masked gather with integer mask not" | |
7570 | " supported."); | |
7571 | return false; | |
7572 | } | |
7573 | } | |
bfaa08b7 RS |
7574 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
7575 | && memory_access_type != VMAT_GATHER_SCATTER) | |
c3a8f964 RS |
7576 | { |
7577 | if (dump_enabled_p ()) | |
7578 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7579 | "unsupported access type for masked load.\n"); | |
7580 | return false; | |
7581 | } | |
7582 | } | |
7583 | ||
ebfd146a IR |
7584 | if (!vec_stmt) /* transformation not required. */ |
7585 | { | |
2de001ee RS |
7586 | if (!slp) |
7587 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; | |
7cfb4d93 RS |
7588 | |
7589 | if (loop_vinfo | |
7590 | && LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) | |
7591 | check_load_store_masking (loop_vinfo, vectype, VLS_LOAD, group_size, | |
bfaa08b7 | 7592 | memory_access_type, &gs_info); |
7cfb4d93 | 7593 | |
ebfd146a | 7594 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; |
68435eb2 RB |
7595 | vect_model_load_cost (stmt_info, ncopies, memory_access_type, |
7596 | slp_node_instance, slp_node, cost_vec); | |
ebfd146a IR |
7597 | return true; |
7598 | } | |
7599 | ||
2de001ee RS |
7600 | if (!slp) |
7601 | gcc_assert (memory_access_type | |
7602 | == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); | |
7603 | ||
73fbfcad | 7604 | if (dump_enabled_p ()) |
78c60e3d | 7605 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 7606 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a | 7607 | |
67b8dbac | 7608 | /* Transform. */ |
ebfd146a | 7609 | |
89fa689a RS |
7610 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info), *first_dr_info = NULL; |
7611 | ensure_base_align (dr_info); | |
c716e67f | 7612 | |
bfaa08b7 | 7613 | if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
aec7ae7d | 7614 | { |
e4057a39 | 7615 | vect_build_gather_load_calls (stmt_info, gsi, vec_stmt, &gs_info, mask); |
aec7ae7d JJ |
7616 | return true; |
7617 | } | |
2de001ee | 7618 | |
2d4bca81 RS |
7619 | if (memory_access_type == VMAT_INVARIANT) |
7620 | { | |
7621 | gcc_assert (!grouped_load && !mask && !bb_vinfo); | |
7622 | /* If we have versioned for aliasing or the loop doesn't | |
7623 | have any data dependencies that would preclude this, | |
7624 | then we are sure this is a loop invariant load and | |
7625 | thus we can insert it on the preheader edge. */ | |
7626 | bool hoist_p = (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
7627 | && !nested_in_vect_loop | |
7628 | && hoist_defs_of_uses (stmt_info, loop)); | |
7629 | if (hoist_p) | |
7630 | { | |
7631 | gassign *stmt = as_a <gassign *> (stmt_info->stmt); | |
7632 | if (dump_enabled_p ()) | |
3c2a8ed0 DM |
7633 | dump_printf_loc (MSG_NOTE, vect_location, |
7634 | "hoisting out of the vectorized loop: %G", stmt); | |
2d4bca81 RS |
7635 | scalar_dest = copy_ssa_name (scalar_dest); |
7636 | tree rhs = unshare_expr (gimple_assign_rhs1 (stmt)); | |
7637 | gsi_insert_on_edge_immediate | |
7638 | (loop_preheader_edge (loop), | |
7639 | gimple_build_assign (scalar_dest, rhs)); | |
7640 | } | |
7641 | /* These copies are all equivalent, but currently the representation | |
7642 | requires a separate STMT_VINFO_VEC_STMT for each one. */ | |
7643 | prev_stmt_info = NULL; | |
7644 | gimple_stmt_iterator gsi2 = *gsi; | |
7645 | gsi_next (&gsi2); | |
7646 | for (j = 0; j < ncopies; j++) | |
7647 | { | |
7648 | stmt_vec_info new_stmt_info; | |
7649 | if (hoist_p) | |
7650 | { | |
7651 | new_temp = vect_init_vector (stmt_info, scalar_dest, | |
7652 | vectype, NULL); | |
7653 | gimple *new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
7654 | new_stmt_info = vinfo->add_stmt (new_stmt); | |
7655 | } | |
7656 | else | |
7657 | { | |
7658 | new_temp = vect_init_vector (stmt_info, scalar_dest, | |
7659 | vectype, &gsi2); | |
7660 | new_stmt_info = vinfo->lookup_def (new_temp); | |
7661 | } | |
7662 | if (slp) | |
7663 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); | |
7664 | else if (j == 0) | |
7665 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; | |
7666 | else | |
7667 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; | |
7668 | prev_stmt_info = new_stmt_info; | |
7669 | } | |
7670 | return true; | |
7671 | } | |
7672 | ||
2de001ee RS |
7673 | if (memory_access_type == VMAT_ELEMENTWISE |
7674 | || memory_access_type == VMAT_STRIDED_SLP) | |
7d75abc8 MM |
7675 | { |
7676 | gimple_stmt_iterator incr_gsi; | |
7677 | bool insert_after; | |
355fe088 | 7678 | gimple *incr; |
7d75abc8 | 7679 | tree offvar; |
7d75abc8 MM |
7680 | tree ivstep; |
7681 | tree running_off; | |
9771b263 | 7682 | vec<constructor_elt, va_gc> *v = NULL; |
14ac6aa2 | 7683 | tree stride_base, stride_step, alias_off; |
4d694b27 RS |
7684 | /* Checked by get_load_store_type. */ |
7685 | unsigned int const_nunits = nunits.to_constant (); | |
b210f45f | 7686 | unsigned HOST_WIDE_INT cst_offset = 0; |
14ac6aa2 | 7687 | |
7cfb4d93 | 7688 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
14ac6aa2 | 7689 | gcc_assert (!nested_in_vect_loop); |
7d75abc8 | 7690 | |
b210f45f | 7691 | if (grouped_load) |
44fc7854 | 7692 | { |
bffb8014 | 7693 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
89fa689a | 7694 | first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info); |
44fc7854 | 7695 | } |
ab313a8c | 7696 | else |
44fc7854 | 7697 | { |
bffb8014 | 7698 | first_stmt_info = stmt_info; |
89fa689a | 7699 | first_dr_info = dr_info; |
b210f45f RB |
7700 | } |
7701 | if (slp && grouped_load) | |
7702 | { | |
bffb8014 RS |
7703 | group_size = DR_GROUP_SIZE (first_stmt_info); |
7704 | ref_type = get_group_alias_ptr_type (first_stmt_info); | |
b210f45f RB |
7705 | } |
7706 | else | |
7707 | { | |
7708 | if (grouped_load) | |
7709 | cst_offset | |
7710 | = (tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))) | |
86a91c0a | 7711 | * vect_get_place_in_interleaving_chain (stmt_info, |
bffb8014 | 7712 | first_stmt_info)); |
44fc7854 | 7713 | group_size = 1; |
89fa689a | 7714 | ref_type = reference_alias_ptr_type (DR_REF (dr_info->dr)); |
44fc7854 | 7715 | } |
ab313a8c | 7716 | |
14ac6aa2 RB |
7717 | stride_base |
7718 | = fold_build_pointer_plus | |
89fa689a | 7719 | (DR_BASE_ADDRESS (first_dr_info->dr), |
14ac6aa2 | 7720 | size_binop (PLUS_EXPR, |
89fa689a RS |
7721 | convert_to_ptrofftype (DR_OFFSET (first_dr_info->dr)), |
7722 | convert_to_ptrofftype (DR_INIT (first_dr_info->dr)))); | |
7723 | stride_step = fold_convert (sizetype, DR_STEP (first_dr_info->dr)); | |
7d75abc8 MM |
7724 | |
7725 | /* For a load with loop-invariant (but other than power-of-2) | |
7726 | stride (i.e. not a grouped access) like so: | |
7727 | ||
7728 | for (i = 0; i < n; i += stride) | |
7729 | ... = array[i]; | |
7730 | ||
7731 | we generate a new induction variable and new accesses to | |
7732 | form a new vector (or vectors, depending on ncopies): | |
7733 | ||
7734 | for (j = 0; ; j += VF*stride) | |
7735 | tmp1 = array[j]; | |
7736 | tmp2 = array[j + stride]; | |
7737 | ... | |
7738 | vectemp = {tmp1, tmp2, ...} | |
7739 | */ | |
7740 | ||
ab313a8c RB |
7741 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
7742 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
7743 | |
7744 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
7745 | ||
b210f45f RB |
7746 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
7747 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); | |
7748 | create_iv (stride_base, ivstep, NULL, | |
7d75abc8 MM |
7749 | loop, &incr_gsi, insert_after, |
7750 | &offvar, NULL); | |
7751 | incr = gsi_stmt (incr_gsi); | |
4fbeb363 | 7752 | loop_vinfo->add_stmt (incr); |
7d75abc8 | 7753 | |
b210f45f | 7754 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
7d75abc8 MM |
7755 | |
7756 | prev_stmt_info = NULL; | |
7757 | running_off = offvar; | |
44fc7854 | 7758 | alias_off = build_int_cst (ref_type, 0); |
4d694b27 | 7759 | int nloads = const_nunits; |
e09b4c37 | 7760 | int lnel = 1; |
7b5fc413 | 7761 | tree ltype = TREE_TYPE (vectype); |
ea60dd34 | 7762 | tree lvectype = vectype; |
b266b968 | 7763 | auto_vec<tree> dr_chain; |
2de001ee | 7764 | if (memory_access_type == VMAT_STRIDED_SLP) |
7b5fc413 | 7765 | { |
4d694b27 | 7766 | if (group_size < const_nunits) |
e09b4c37 | 7767 | { |
ff03930a JJ |
7768 | /* First check if vec_init optab supports construction from |
7769 | vector elts directly. */ | |
b397965c | 7770 | scalar_mode elmode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); |
9da15d40 RS |
7771 | machine_mode vmode; |
7772 | if (mode_for_vector (elmode, group_size).exists (&vmode) | |
7773 | && VECTOR_MODE_P (vmode) | |
414fef4e | 7774 | && targetm.vector_mode_supported_p (vmode) |
ff03930a JJ |
7775 | && (convert_optab_handler (vec_init_optab, |
7776 | TYPE_MODE (vectype), vmode) | |
7777 | != CODE_FOR_nothing)) | |
ea60dd34 | 7778 | { |
4d694b27 | 7779 | nloads = const_nunits / group_size; |
ea60dd34 | 7780 | lnel = group_size; |
ff03930a JJ |
7781 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); |
7782 | } | |
7783 | else | |
7784 | { | |
7785 | /* Otherwise avoid emitting a constructor of vector elements | |
7786 | by performing the loads using an integer type of the same | |
7787 | size, constructing a vector of those and then | |
7788 | re-interpreting it as the original vector type. | |
7789 | This avoids a huge runtime penalty due to the general | |
7790 | inability to perform store forwarding from smaller stores | |
7791 | to a larger load. */ | |
7792 | unsigned lsize | |
7793 | = group_size * TYPE_PRECISION (TREE_TYPE (vectype)); | |
fffbab82 | 7794 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 7795 | unsigned int lnunits = const_nunits / group_size; |
ff03930a JJ |
7796 | /* If we can't construct such a vector fall back to |
7797 | element loads of the original vector type. */ | |
4d694b27 | 7798 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 7799 | && VECTOR_MODE_P (vmode) |
414fef4e | 7800 | && targetm.vector_mode_supported_p (vmode) |
ff03930a JJ |
7801 | && (convert_optab_handler (vec_init_optab, vmode, elmode) |
7802 | != CODE_FOR_nothing)) | |
7803 | { | |
4d694b27 | 7804 | nloads = lnunits; |
ff03930a JJ |
7805 | lnel = group_size; |
7806 | ltype = build_nonstandard_integer_type (lsize, 1); | |
7807 | lvectype = build_vector_type (ltype, nloads); | |
7808 | } | |
ea60dd34 | 7809 | } |
e09b4c37 | 7810 | } |
2de001ee | 7811 | else |
e09b4c37 | 7812 | { |
ea60dd34 | 7813 | nloads = 1; |
4d694b27 | 7814 | lnel = const_nunits; |
e09b4c37 | 7815 | ltype = vectype; |
e09b4c37 | 7816 | } |
2de001ee RS |
7817 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); |
7818 | } | |
bb4e4747 BC |
7819 | /* Load vector(1) scalar_type if it's 1 element-wise vectype. */ |
7820 | else if (nloads == 1) | |
7821 | ltype = vectype; | |
7822 | ||
2de001ee RS |
7823 | if (slp) |
7824 | { | |
66c16fd9 RB |
7825 | /* For SLP permutation support we need to load the whole group, |
7826 | not only the number of vector stmts the permutation result | |
7827 | fits in. */ | |
b266b968 | 7828 | if (slp_perm) |
66c16fd9 | 7829 | { |
d9f21f6a RS |
7830 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7831 | variable VF. */ | |
7832 | unsigned int const_vf = vf.to_constant (); | |
4d694b27 | 7833 | ncopies = CEIL (group_size * const_vf, const_nunits); |
66c16fd9 RB |
7834 | dr_chain.create (ncopies); |
7835 | } | |
7836 | else | |
7837 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
7b5fc413 | 7838 | } |
4d694b27 | 7839 | unsigned int group_el = 0; |
e09b4c37 RB |
7840 | unsigned HOST_WIDE_INT |
7841 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
7d75abc8 MM |
7842 | for (j = 0; j < ncopies; j++) |
7843 | { | |
7b5fc413 | 7844 | if (nloads > 1) |
e09b4c37 | 7845 | vec_alloc (v, nloads); |
e1bd7296 | 7846 | stmt_vec_info new_stmt_info = NULL; |
e09b4c37 | 7847 | for (i = 0; i < nloads; i++) |
7b5fc413 | 7848 | { |
e09b4c37 | 7849 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
b210f45f | 7850 | group_el * elsz + cst_offset); |
19986382 | 7851 | tree data_ref = build2 (MEM_REF, ltype, running_off, this_off); |
89fa689a | 7852 | vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr)); |
e1bd7296 RS |
7853 | gassign *new_stmt |
7854 | = gimple_build_assign (make_ssa_name (ltype), data_ref); | |
7855 | new_stmt_info | |
86a91c0a | 7856 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
e09b4c37 RB |
7857 | if (nloads > 1) |
7858 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, | |
7859 | gimple_assign_lhs (new_stmt)); | |
7860 | ||
7861 | group_el += lnel; | |
7862 | if (! slp | |
7863 | || group_el == group_size) | |
7b5fc413 | 7864 | { |
e09b4c37 RB |
7865 | tree newoff = copy_ssa_name (running_off); |
7866 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
7867 | running_off, stride_step); | |
86a91c0a | 7868 | vect_finish_stmt_generation (stmt_info, incr, gsi); |
7b5fc413 RB |
7869 | |
7870 | running_off = newoff; | |
e09b4c37 | 7871 | group_el = 0; |
7b5fc413 | 7872 | } |
7b5fc413 | 7873 | } |
e09b4c37 | 7874 | if (nloads > 1) |
7d75abc8 | 7875 | { |
ea60dd34 | 7876 | tree vec_inv = build_constructor (lvectype, v); |
86a91c0a | 7877 | new_temp = vect_init_vector (stmt_info, vec_inv, lvectype, gsi); |
e1bd7296 | 7878 | new_stmt_info = vinfo->lookup_def (new_temp); |
ea60dd34 RB |
7879 | if (lvectype != vectype) |
7880 | { | |
e1bd7296 RS |
7881 | gassign *new_stmt |
7882 | = gimple_build_assign (make_ssa_name (vectype), | |
7883 | VIEW_CONVERT_EXPR, | |
7884 | build1 (VIEW_CONVERT_EXPR, | |
7885 | vectype, new_temp)); | |
7886 | new_stmt_info | |
86a91c0a | 7887 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
ea60dd34 | 7888 | } |
7d75abc8 MM |
7889 | } |
7890 | ||
7b5fc413 | 7891 | if (slp) |
b266b968 | 7892 | { |
b266b968 | 7893 | if (slp_perm) |
e1bd7296 | 7894 | dr_chain.quick_push (gimple_assign_lhs (new_stmt_info->stmt)); |
66c16fd9 | 7895 | else |
e1bd7296 | 7896 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
b266b968 | 7897 | } |
7d75abc8 | 7898 | else |
225ce44b RB |
7899 | { |
7900 | if (j == 0) | |
e1bd7296 | 7901 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
225ce44b | 7902 | else |
e1bd7296 RS |
7903 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
7904 | prev_stmt_info = new_stmt_info; | |
225ce44b | 7905 | } |
7d75abc8 | 7906 | } |
b266b968 | 7907 | if (slp_perm) |
29afecdf RB |
7908 | { |
7909 | unsigned n_perms; | |
7910 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
7911 | slp_node_instance, false, &n_perms); | |
7912 | } | |
7d75abc8 MM |
7913 | return true; |
7914 | } | |
aec7ae7d | 7915 | |
b5ec4de7 RS |
7916 | if (memory_access_type == VMAT_GATHER_SCATTER |
7917 | || (!slp && memory_access_type == VMAT_CONTIGUOUS)) | |
ab2fc782 RS |
7918 | grouped_load = false; |
7919 | ||
0d0293ac | 7920 | if (grouped_load) |
ebfd146a | 7921 | { |
bffb8014 RS |
7922 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
7923 | group_size = DR_GROUP_SIZE (first_stmt_info); | |
4f0a0218 | 7924 | /* For SLP vectorization we directly vectorize a subchain |
52eab378 RB |
7925 | without permutation. */ |
7926 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
bffb8014 | 7927 | first_stmt_info = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
4f0a0218 RB |
7928 | /* For BB vectorization always use the first stmt to base |
7929 | the data ref pointer on. */ | |
7930 | if (bb_vinfo) | |
b9787581 | 7931 | first_stmt_info_for_drptr = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
6aa904c4 | 7932 | |
ebfd146a | 7933 | /* Check if the chain of loads is already vectorized. */ |
bffb8014 | 7934 | if (STMT_VINFO_VEC_STMT (first_stmt_info) |
01d8bf07 RB |
7935 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. |
7936 | ??? But we can only do so if there is exactly one | |
7937 | as we have no way to get at the rest. Leave the CSE | |
7938 | opportunity alone. | |
7939 | ??? With the group load eventually participating | |
7940 | in multiple different permutations (having multiple | |
7941 | slp nodes which refer to the same group) the CSE | |
7942 | is even wrong code. See PR56270. */ | |
7943 | && !slp) | |
ebfd146a IR |
7944 | { |
7945 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
7946 | return true; | |
7947 | } | |
89fa689a | 7948 | first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info); |
9b999e8c | 7949 | group_gap_adj = 0; |
ebfd146a IR |
7950 | |
7951 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
7952 | if (slp) | |
7953 | { | |
0d0293ac | 7954 | grouped_load = false; |
ab7e60ce RS |
7955 | /* If an SLP permutation is from N elements to N elements, |
7956 | and if one vector holds a whole number of N, we can load | |
7957 | the inputs to the permutation in the same way as an | |
7958 | unpermuted sequence. In other cases we need to load the | |
7959 | whole group, not only the number of vector stmts the | |
7960 | permutation result fits in. */ | |
7961 | if (slp_perm | |
7962 | && (group_size != SLP_INSTANCE_GROUP_SIZE (slp_node_instance) | |
7963 | || !multiple_p (nunits, group_size))) | |
b267968e | 7964 | { |
ab7e60ce RS |
7965 | /* We don't yet generate such SLP_TREE_LOAD_PERMUTATIONs for |
7966 | variable VF; see vect_transform_slp_perm_load. */ | |
d9f21f6a | 7967 | unsigned int const_vf = vf.to_constant (); |
4d694b27 RS |
7968 | unsigned int const_nunits = nunits.to_constant (); |
7969 | vec_num = CEIL (group_size * const_vf, const_nunits); | |
b267968e RB |
7970 | group_gap_adj = vf * group_size - nunits * vec_num; |
7971 | } | |
91ff1504 | 7972 | else |
b267968e RB |
7973 | { |
7974 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
796bd467 RB |
7975 | group_gap_adj |
7976 | = group_size - SLP_INSTANCE_GROUP_SIZE (slp_node_instance); | |
b267968e | 7977 | } |
a70d6342 | 7978 | } |
ebfd146a | 7979 | else |
9b999e8c | 7980 | vec_num = group_size; |
44fc7854 | 7981 | |
bffb8014 | 7982 | ref_type = get_group_alias_ptr_type (first_stmt_info); |
ebfd146a IR |
7983 | } |
7984 | else | |
7985 | { | |
bffb8014 | 7986 | first_stmt_info = stmt_info; |
89fa689a | 7987 | first_dr_info = dr_info; |
ebfd146a | 7988 | group_size = vec_num = 1; |
9b999e8c | 7989 | group_gap_adj = 0; |
89fa689a | 7990 | ref_type = reference_alias_ptr_type (DR_REF (first_dr_info->dr)); |
ebfd146a IR |
7991 | } |
7992 | ||
89fa689a RS |
7993 | alignment_support_scheme |
7994 | = vect_supportable_dr_alignment (first_dr_info, false); | |
ebfd146a | 7995 | gcc_assert (alignment_support_scheme); |
70088b95 RS |
7996 | vec_loop_masks *loop_masks |
7997 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) | |
7998 | ? &LOOP_VINFO_MASKS (loop_vinfo) | |
7999 | : NULL); | |
7cfb4d93 RS |
8000 | /* Targets with store-lane instructions must not require explicit |
8001 | realignment. vect_supportable_dr_alignment always returns either | |
8002 | dr_aligned or dr_unaligned_supported for masked operations. */ | |
8003 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES | |
8004 | && !mask | |
70088b95 | 8005 | && !loop_masks) |
272c6793 RS |
8006 | || alignment_support_scheme == dr_aligned |
8007 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
8008 | |
8009 | /* In case the vectorization factor (VF) is bigger than the number | |
8010 | of elements that we can fit in a vectype (nunits), we have to generate | |
8011 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 8012 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 8013 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 8014 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 8015 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
8016 | stmts that use the defs of the current stmt. The example below |
8017 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
8018 | need to create 4 vectorized stmts): | |
ebfd146a IR |
8019 | |
8020 | before vectorization: | |
8021 | RELATED_STMT VEC_STMT | |
8022 | S1: x = memref - - | |
8023 | S2: z = x + 1 - - | |
8024 | ||
8025 | step 1: vectorize stmt S1: | |
8026 | We first create the vector stmt VS1_0, and, as usual, record a | |
8027 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
8028 | Next, we create the vector stmt VS1_1, and record a pointer to | |
8029 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 8030 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
8031 | stmts and pointers: |
8032 | RELATED_STMT VEC_STMT | |
8033 | VS1_0: vx0 = memref0 VS1_1 - | |
8034 | VS1_1: vx1 = memref1 VS1_2 - | |
8035 | VS1_2: vx2 = memref2 VS1_3 - | |
8036 | VS1_3: vx3 = memref3 - - | |
8037 | S1: x = load - VS1_0 | |
8038 | S2: z = x + 1 - - | |
8039 | ||
b8698a0f L |
8040 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
8041 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
8042 | stmt S2. */ |
8043 | ||
0d0293ac | 8044 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
8045 | |
8046 | S1: x2 = &base + 2 | |
8047 | S2: x0 = &base | |
8048 | S3: x1 = &base + 1 | |
8049 | S4: x3 = &base + 3 | |
8050 | ||
b8698a0f | 8051 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
8052 | starting from the access of the first stmt of the chain: |
8053 | ||
8054 | VS1: vx0 = &base | |
8055 | VS2: vx1 = &base + vec_size*1 | |
8056 | VS3: vx3 = &base + vec_size*2 | |
8057 | VS4: vx4 = &base + vec_size*3 | |
8058 | ||
8059 | Then permutation statements are generated: | |
8060 | ||
e2c83630 RH |
8061 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
8062 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
8063 | ... |
8064 | ||
8065 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
8066 | (the order of the data-refs in the output of vect_permute_load_chain | |
8067 | corresponds to the order of scalar stmts in the interleaving chain - see | |
8068 | the documentation of vect_permute_load_chain()). | |
8069 | The generation of permutation stmts and recording them in | |
0d0293ac | 8070 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 8071 | |
b8698a0f | 8072 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
8073 | permutation stmts above are created for every copy. The result vector |
8074 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
8075 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
8076 | |
8077 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
8078 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
8079 | we generate the following code: | |
8080 | p = initial_addr; | |
8081 | indx = 0; | |
8082 | loop { | |
8083 | p = p + indx * vectype_size; | |
8084 | vec_dest = *(p); | |
8085 | indx = indx + 1; | |
8086 | } | |
8087 | ||
8088 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 8089 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
8090 | then generate the following code, in which the data in each iteration is |
8091 | obtained by two vector loads, one from the previous iteration, and one | |
8092 | from the current iteration: | |
8093 | p1 = initial_addr; | |
8094 | msq_init = *(floor(p1)) | |
8095 | p2 = initial_addr + VS - 1; | |
8096 | realignment_token = call target_builtin; | |
8097 | indx = 0; | |
8098 | loop { | |
8099 | p2 = p2 + indx * vectype_size | |
8100 | lsq = *(floor(p2)) | |
8101 | vec_dest = realign_load (msq, lsq, realignment_token) | |
8102 | indx = indx + 1; | |
8103 | msq = lsq; | |
8104 | } */ | |
8105 | ||
8106 | /* If the misalignment remains the same throughout the execution of the | |
8107 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 8108 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
8109 | This can only occur when vectorizing memory accesses in the inner-loop |
8110 | nested within an outer-loop that is being vectorized. */ | |
8111 | ||
d1e4b493 | 8112 | if (nested_in_vect_loop |
89fa689a | 8113 | && !multiple_p (DR_STEP_ALIGNMENT (dr_info->dr), |
cf098191 | 8114 | GET_MODE_SIZE (TYPE_MODE (vectype)))) |
ebfd146a IR |
8115 | { |
8116 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
8117 | compute_in_loop = true; | |
8118 | } | |
8119 | ||
8120 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
8121 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 8122 | && !compute_in_loop) |
ebfd146a | 8123 | { |
bffb8014 | 8124 | msq = vect_setup_realignment (first_stmt_info, gsi, &realignment_token, |
ebfd146a IR |
8125 | alignment_support_scheme, NULL_TREE, |
8126 | &at_loop); | |
8127 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
8128 | { | |
538dd0b7 | 8129 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
8130 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
8131 | size_one_node); | |
ebfd146a IR |
8132 | } |
8133 | } | |
8134 | else | |
8135 | at_loop = loop; | |
8136 | ||
62da9e14 | 8137 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
a1e53f3f L |
8138 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
8139 | ||
ab2fc782 RS |
8140 | tree bump; |
8141 | tree vec_offset = NULL_TREE; | |
8142 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
8143 | { | |
8144 | aggr_type = NULL_TREE; | |
8145 | bump = NULL_TREE; | |
8146 | } | |
8147 | else if (memory_access_type == VMAT_GATHER_SCATTER) | |
8148 | { | |
8149 | aggr_type = elem_type; | |
86a91c0a | 8150 | vect_get_strided_load_store_ops (stmt_info, loop_vinfo, &gs_info, |
ab2fc782 RS |
8151 | &bump, &vec_offset); |
8152 | } | |
272c6793 | 8153 | else |
ab2fc782 RS |
8154 | { |
8155 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
8156 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
8157 | else | |
8158 | aggr_type = vectype; | |
89fa689a RS |
8159 | bump = vect_get_data_ptr_increment (dr_info, aggr_type, |
8160 | memory_access_type); | |
ab2fc782 | 8161 | } |
272c6793 | 8162 | |
c3a8f964 | 8163 | tree vec_mask = NULL_TREE; |
ebfd146a | 8164 | prev_stmt_info = NULL; |
4d694b27 | 8165 | poly_uint64 group_elt = 0; |
ebfd146a | 8166 | for (j = 0; j < ncopies; j++) |
b8698a0f | 8167 | { |
e1bd7296 | 8168 | stmt_vec_info new_stmt_info = NULL; |
272c6793 | 8169 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 8170 | if (j == 0) |
74bf76ed JJ |
8171 | { |
8172 | bool simd_lane_access_p | |
8173 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
8174 | if (simd_lane_access_p | |
89fa689a RS |
8175 | && TREE_CODE (DR_BASE_ADDRESS (first_dr_info->dr)) == ADDR_EXPR |
8176 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr_info->dr), 0)) | |
8177 | && integer_zerop (DR_OFFSET (first_dr_info->dr)) | |
8178 | && integer_zerop (DR_INIT (first_dr_info->dr)) | |
74bf76ed | 8179 | && alias_sets_conflict_p (get_alias_set (aggr_type), |
44fc7854 | 8180 | get_alias_set (TREE_TYPE (ref_type))) |
74bf76ed JJ |
8181 | && (alignment_support_scheme == dr_aligned |
8182 | || alignment_support_scheme == dr_unaligned_supported)) | |
8183 | { | |
89fa689a | 8184 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr_info->dr)); |
44fc7854 | 8185 | dataref_offset = build_int_cst (ref_type, 0); |
74bf76ed | 8186 | } |
b9787581 | 8187 | else if (first_stmt_info_for_drptr |
bffb8014 | 8188 | && first_stmt_info != first_stmt_info_for_drptr) |
4f0a0218 RB |
8189 | { |
8190 | dataref_ptr | |
b9787581 RS |
8191 | = vect_create_data_ref_ptr (first_stmt_info_for_drptr, |
8192 | aggr_type, at_loop, offset, &dummy, | |
8193 | gsi, &ptr_incr, simd_lane_access_p, | |
2d4bca81 | 8194 | byte_offset, bump); |
4f0a0218 RB |
8195 | /* Adjust the pointer by the difference to first_stmt. */ |
8196 | data_reference_p ptrdr | |
b9787581 | 8197 | = STMT_VINFO_DATA_REF (first_stmt_info_for_drptr); |
89fa689a RS |
8198 | tree diff |
8199 | = fold_convert (sizetype, | |
8200 | size_binop (MINUS_EXPR, | |
8201 | DR_INIT (first_dr_info->dr), | |
8202 | DR_INIT (ptrdr))); | |
4f0a0218 | 8203 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
86a91c0a | 8204 | stmt_info, diff); |
4f0a0218 | 8205 | } |
bfaa08b7 | 8206 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
2d4bca81 RS |
8207 | vect_get_gather_scatter_ops (loop, stmt_info, &gs_info, |
8208 | &dataref_ptr, &vec_offset); | |
74bf76ed JJ |
8209 | else |
8210 | dataref_ptr | |
bffb8014 | 8211 | = vect_create_data_ref_ptr (first_stmt_info, aggr_type, at_loop, |
74bf76ed | 8212 | offset, &dummy, gsi, &ptr_incr, |
2d4bca81 | 8213 | simd_lane_access_p, |
ab2fc782 | 8214 | byte_offset, bump); |
c3a8f964 | 8215 | if (mask) |
86a91c0a | 8216 | vec_mask = vect_get_vec_def_for_operand (mask, stmt_info, |
c3a8f964 | 8217 | mask_vectype); |
74bf76ed | 8218 | } |
ebfd146a | 8219 | else |
c3a8f964 RS |
8220 | { |
8221 | if (dataref_offset) | |
8222 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
ab2fc782 | 8223 | bump); |
bfaa08b7 | 8224 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
e4057a39 | 8225 | vec_offset = vect_get_vec_def_for_stmt_copy (vinfo, vec_offset); |
c3a8f964 | 8226 | else |
ab2fc782 | 8227 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
86a91c0a | 8228 | stmt_info, bump); |
c3a8f964 | 8229 | if (mask) |
e4057a39 | 8230 | vec_mask = vect_get_vec_def_for_stmt_copy (vinfo, 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); | |
86a91c0a | 8271 | new_stmt_info = vect_finish_stmt_generation (stmt_info, call, gsi); |
ebfd146a | 8272 | |
272c6793 RS |
8273 | /* Extract each vector into an SSA_NAME. */ |
8274 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 8275 | { |
86a91c0a | 8276 | new_temp = read_vector_array (stmt_info, gsi, scalar_dest, |
272c6793 | 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. */ | |
86a91c0a | 8282 | vect_record_grouped_load_vectors (stmt_info, dr_chain); |
3ba4ff41 RS |
8283 | |
8284 | /* Record that VEC_ARRAY is now dead. */ | |
86a91c0a | 8285 | vect_clobber_variable (stmt_info, 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, | |
86a91c0a | 8303 | stmt_info, 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 | ||
89fa689a | 8332 | align = DR_TARGET_ALIGNMENT (dr_info); |
272c6793 RS |
8333 | if (alignment_support_scheme == dr_aligned) |
8334 | { | |
89fa689a | 8335 | gcc_assert (aligned_access_p (first_dr_info)); |
644ffefd | 8336 | misalign = 0; |
272c6793 | 8337 | } |
89fa689a | 8338 | else if (DR_MISALIGNMENT (first_dr_info) == -1) |
272c6793 | 8339 | { |
89fa689a RS |
8340 | align = dr_alignment |
8341 | (vect_dr_behavior (first_dr_info)); | |
52639a61 | 8342 | misalign = 0; |
272c6793 RS |
8343 | } |
8344 | else | |
89fa689a | 8345 | misalign = DR_MISALIGNMENT (first_dr_info); |
aed93b23 RB |
8346 | if (dataref_offset == NULL_TREE |
8347 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
8348 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
8349 | align, misalign); | |
c3a8f964 | 8350 | |
7cfb4d93 | 8351 | if (final_mask) |
c3a8f964 RS |
8352 | { |
8353 | align = least_bit_hwi (misalign | align); | |
8354 | tree ptr = build_int_cst (ref_type, align); | |
8355 | gcall *call | |
8356 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, | |
8357 | dataref_ptr, ptr, | |
7cfb4d93 | 8358 | final_mask); |
c3a8f964 RS |
8359 | gimple_call_set_nothrow (call, true); |
8360 | new_stmt = call; | |
8361 | data_ref = NULL_TREE; | |
8362 | } | |
8363 | else | |
8364 | { | |
8365 | data_ref | |
8366 | = fold_build2 (MEM_REF, vectype, dataref_ptr, | |
8367 | dataref_offset | |
8368 | ? dataref_offset | |
8369 | : build_int_cst (ref_type, 0)); | |
8370 | if (alignment_support_scheme == dr_aligned) | |
8371 | ; | |
89fa689a | 8372 | else if (DR_MISALIGNMENT (first_dr_info) == -1) |
c3a8f964 RS |
8373 | TREE_TYPE (data_ref) |
8374 | = build_aligned_type (TREE_TYPE (data_ref), | |
8375 | align * BITS_PER_UNIT); | |
8376 | else | |
8377 | TREE_TYPE (data_ref) | |
8378 | = build_aligned_type (TREE_TYPE (data_ref), | |
8379 | TYPE_ALIGN (elem_type)); | |
8380 | } | |
272c6793 | 8381 | break; |
be1ac4ec | 8382 | } |
272c6793 | 8383 | case dr_explicit_realign: |
267d3070 | 8384 | { |
272c6793 | 8385 | tree ptr, bump; |
272c6793 | 8386 | |
d88981fc | 8387 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
8388 | |
8389 | if (compute_in_loop) | |
bffb8014 | 8390 | msq = vect_setup_realignment (first_stmt_info, gsi, |
272c6793 RS |
8391 | &realignment_token, |
8392 | dr_explicit_realign, | |
8393 | dataref_ptr, NULL); | |
8394 | ||
aed93b23 RB |
8395 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
8396 | ptr = copy_ssa_name (dataref_ptr); | |
8397 | else | |
8398 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
89fa689a | 8399 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr_info); |
0d0e4a03 JJ |
8400 | new_stmt = gimple_build_assign |
8401 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
8402 | build_int_cst |
8403 | (TREE_TYPE (dataref_ptr), | |
f702e7d4 | 8404 | -(HOST_WIDE_INT) align)); |
86a91c0a | 8405 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
272c6793 RS |
8406 | data_ref |
8407 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 8408 | build_int_cst (ref_type, 0)); |
89fa689a | 8409 | vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr)); |
272c6793 RS |
8410 | vec_dest = vect_create_destination_var (scalar_dest, |
8411 | vectype); | |
8412 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
8413 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
8414 | gimple_assign_set_lhs (new_stmt, new_temp); | |
86a91c0a RS |
8415 | gimple_set_vdef (new_stmt, gimple_vdef (stmt_info->stmt)); |
8416 | gimple_set_vuse (new_stmt, gimple_vuse (stmt_info->stmt)); | |
8417 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); | |
272c6793 RS |
8418 | msq = new_temp; |
8419 | ||
d88981fc | 8420 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 8421 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 8422 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
86a91c0a RS |
8423 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, |
8424 | stmt_info, bump); | |
0d0e4a03 JJ |
8425 | new_stmt = gimple_build_assign |
8426 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 | 8427 | build_int_cst |
f702e7d4 | 8428 | (TREE_TYPE (ptr), -(HOST_WIDE_INT) align)); |
aed93b23 | 8429 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 | 8430 | gimple_assign_set_lhs (new_stmt, ptr); |
86a91c0a | 8431 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
272c6793 RS |
8432 | data_ref |
8433 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 8434 | build_int_cst (ref_type, 0)); |
272c6793 | 8435 | break; |
267d3070 | 8436 | } |
272c6793 | 8437 | case dr_explicit_realign_optimized: |
f702e7d4 RS |
8438 | { |
8439 | if (TREE_CODE (dataref_ptr) == SSA_NAME) | |
8440 | new_temp = copy_ssa_name (dataref_ptr); | |
8441 | else | |
8442 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
89fa689a | 8443 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr_info); |
f702e7d4 RS |
8444 | new_stmt = gimple_build_assign |
8445 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
8446 | build_int_cst (TREE_TYPE (dataref_ptr), | |
8447 | -(HOST_WIDE_INT) align)); | |
86a91c0a | 8448 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
f702e7d4 RS |
8449 | data_ref |
8450 | = build2 (MEM_REF, vectype, new_temp, | |
8451 | build_int_cst (ref_type, 0)); | |
8452 | break; | |
8453 | } | |
272c6793 RS |
8454 | default: |
8455 | gcc_unreachable (); | |
8456 | } | |
ebfd146a | 8457 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
c3a8f964 RS |
8458 | /* DATA_REF is null if we've already built the statement. */ |
8459 | if (data_ref) | |
19986382 | 8460 | { |
89fa689a | 8461 | vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr)); |
19986382 RB |
8462 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
8463 | } | |
ebfd146a | 8464 | new_temp = make_ssa_name (vec_dest, new_stmt); |
c3a8f964 | 8465 | gimple_set_lhs (new_stmt, new_temp); |
e1bd7296 | 8466 | new_stmt_info |
86a91c0a | 8467 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
ebfd146a | 8468 | |
272c6793 RS |
8469 | /* 3. Handle explicit realignment if necessary/supported. |
8470 | Create in loop: | |
8471 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
8472 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
8473 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 8474 | { |
272c6793 RS |
8475 | lsq = gimple_assign_lhs (new_stmt); |
8476 | if (!realignment_token) | |
8477 | realignment_token = dataref_ptr; | |
8478 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
8479 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
8480 | msq, lsq, realignment_token); | |
272c6793 RS |
8481 | new_temp = make_ssa_name (vec_dest, new_stmt); |
8482 | gimple_assign_set_lhs (new_stmt, new_temp); | |
e1bd7296 | 8483 | new_stmt_info |
86a91c0a | 8484 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
272c6793 RS |
8485 | |
8486 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
8487 | { | |
8488 | gcc_assert (phi); | |
8489 | if (i == vec_num - 1 && j == ncopies - 1) | |
8490 | add_phi_arg (phi, lsq, | |
8491 | loop_latch_edge (containing_loop), | |
9e227d60 | 8492 | UNKNOWN_LOCATION); |
272c6793 RS |
8493 | msq = lsq; |
8494 | } | |
ebfd146a | 8495 | } |
ebfd146a | 8496 | |
62da9e14 | 8497 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
272c6793 | 8498 | { |
aec7ae7d JJ |
8499 | tree perm_mask = perm_mask_for_reverse (vectype); |
8500 | new_temp = permute_vec_elements (new_temp, new_temp, | |
86a91c0a | 8501 | perm_mask, stmt_info, gsi); |
e1bd7296 | 8502 | new_stmt_info = vinfo->lookup_def (new_temp); |
ebfd146a | 8503 | } |
267d3070 | 8504 | |
272c6793 | 8505 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
8506 | vect_transform_grouped_load (). */ |
8507 | if (grouped_load || slp_perm) | |
9771b263 | 8508 | dr_chain.quick_push (new_temp); |
267d3070 | 8509 | |
272c6793 RS |
8510 | /* Store vector loads in the corresponding SLP_NODE. */ |
8511 | if (slp && !slp_perm) | |
e1bd7296 | 8512 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
b267968e RB |
8513 | |
8514 | /* With SLP permutation we load the gaps as well, without | |
8515 | we need to skip the gaps after we manage to fully load | |
2c53b149 | 8516 | all elements. group_gap_adj is DR_GROUP_SIZE here. */ |
b267968e | 8517 | group_elt += nunits; |
d9f21f6a RS |
8518 | if (maybe_ne (group_gap_adj, 0U) |
8519 | && !slp_perm | |
8520 | && known_eq (group_elt, group_size - group_gap_adj)) | |
b267968e | 8521 | { |
d9f21f6a RS |
8522 | poly_wide_int bump_val |
8523 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
8524 | * group_gap_adj); | |
8e6cdc90 | 8525 | tree bump = wide_int_to_tree (sizetype, bump_val); |
b267968e | 8526 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
86a91c0a | 8527 | stmt_info, bump); |
b267968e RB |
8528 | group_elt = 0; |
8529 | } | |
272c6793 | 8530 | } |
9b999e8c RB |
8531 | /* Bump the vector pointer to account for a gap or for excess |
8532 | elements loaded for a permuted SLP load. */ | |
d9f21f6a | 8533 | if (maybe_ne (group_gap_adj, 0U) && slp_perm) |
a64b9c26 | 8534 | { |
d9f21f6a RS |
8535 | poly_wide_int bump_val |
8536 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
8537 | * group_gap_adj); | |
8e6cdc90 | 8538 | tree bump = wide_int_to_tree (sizetype, bump_val); |
a64b9c26 | 8539 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
86a91c0a | 8540 | stmt_info, bump); |
a64b9c26 | 8541 | } |
ebfd146a IR |
8542 | } |
8543 | ||
8544 | if (slp && !slp_perm) | |
8545 | continue; | |
8546 | ||
8547 | if (slp_perm) | |
8548 | { | |
29afecdf | 8549 | unsigned n_perms; |
01d8bf07 | 8550 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
29afecdf RB |
8551 | slp_node_instance, false, |
8552 | &n_perms)) | |
ebfd146a | 8553 | { |
9771b263 | 8554 | dr_chain.release (); |
ebfd146a IR |
8555 | return false; |
8556 | } | |
8557 | } | |
8558 | else | |
8559 | { | |
0d0293ac | 8560 | if (grouped_load) |
ebfd146a | 8561 | { |
2de001ee | 8562 | if (memory_access_type != VMAT_LOAD_STORE_LANES) |
86a91c0a RS |
8563 | vect_transform_grouped_load (stmt_info, dr_chain, |
8564 | group_size, gsi); | |
ebfd146a | 8565 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
8566 | } |
8567 | else | |
8568 | { | |
8569 | if (j == 0) | |
e1bd7296 | 8570 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
ebfd146a | 8571 | else |
e1bd7296 RS |
8572 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
8573 | prev_stmt_info = new_stmt_info; | |
ebfd146a IR |
8574 | } |
8575 | } | |
9771b263 | 8576 | dr_chain.release (); |
ebfd146a IR |
8577 | } |
8578 | ||
ebfd146a IR |
8579 | return true; |
8580 | } | |
8581 | ||
8582 | /* Function vect_is_simple_cond. | |
b8698a0f | 8583 | |
ebfd146a IR |
8584 | Input: |
8585 | LOOP - the loop that is being vectorized. | |
8586 | COND - Condition that is checked for simple use. | |
8587 | ||
e9e1d143 RG |
8588 | Output: |
8589 | *COMP_VECTYPE - the vector type for the comparison. | |
4fc5ebf1 | 8590 | *DTS - The def types for the arguments of the comparison |
e9e1d143 | 8591 | |
ebfd146a IR |
8592 | Returns whether a COND can be vectorized. Checks whether |
8593 | condition operands are supportable using vec_is_simple_use. */ | |
8594 | ||
87aab9b2 | 8595 | static bool |
4fc5ebf1 | 8596 | vect_is_simple_cond (tree cond, vec_info *vinfo, |
8da4c8d8 RB |
8597 | tree *comp_vectype, enum vect_def_type *dts, |
8598 | tree vectype) | |
ebfd146a IR |
8599 | { |
8600 | tree lhs, rhs; | |
e9e1d143 | 8601 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 8602 | |
a414c77f IE |
8603 | /* Mask case. */ |
8604 | if (TREE_CODE (cond) == SSA_NAME | |
2568d8a1 | 8605 | && VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (cond))) |
a414c77f | 8606 | { |
894dd753 | 8607 | if (!vect_is_simple_use (cond, vinfo, &dts[0], comp_vectype) |
a414c77f IE |
8608 | || !*comp_vectype |
8609 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
8610 | return false; | |
8611 | return true; | |
8612 | } | |
8613 | ||
ebfd146a IR |
8614 | if (!COMPARISON_CLASS_P (cond)) |
8615 | return false; | |
8616 | ||
8617 | lhs = TREE_OPERAND (cond, 0); | |
8618 | rhs = TREE_OPERAND (cond, 1); | |
8619 | ||
8620 | if (TREE_CODE (lhs) == SSA_NAME) | |
8621 | { | |
894dd753 | 8622 | if (!vect_is_simple_use (lhs, vinfo, &dts[0], &vectype1)) |
ebfd146a IR |
8623 | return false; |
8624 | } | |
4fc5ebf1 JG |
8625 | else if (TREE_CODE (lhs) == INTEGER_CST || TREE_CODE (lhs) == REAL_CST |
8626 | || TREE_CODE (lhs) == FIXED_CST) | |
8627 | dts[0] = vect_constant_def; | |
8628 | else | |
ebfd146a IR |
8629 | return false; |
8630 | ||
8631 | if (TREE_CODE (rhs) == SSA_NAME) | |
8632 | { | |
894dd753 | 8633 | if (!vect_is_simple_use (rhs, vinfo, &dts[1], &vectype2)) |
ebfd146a IR |
8634 | return false; |
8635 | } | |
4fc5ebf1 JG |
8636 | else if (TREE_CODE (rhs) == INTEGER_CST || TREE_CODE (rhs) == REAL_CST |
8637 | || TREE_CODE (rhs) == FIXED_CST) | |
8638 | dts[1] = vect_constant_def; | |
8639 | else | |
ebfd146a IR |
8640 | return false; |
8641 | ||
28b33016 | 8642 | if (vectype1 && vectype2 |
928686b1 RS |
8643 | && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1), |
8644 | TYPE_VECTOR_SUBPARTS (vectype2))) | |
28b33016 IE |
8645 | return false; |
8646 | ||
e9e1d143 | 8647 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
8da4c8d8 | 8648 | /* Invariant comparison. */ |
4515e413 | 8649 | if (! *comp_vectype && vectype) |
8da4c8d8 RB |
8650 | { |
8651 | tree scalar_type = TREE_TYPE (lhs); | |
8652 | /* If we can widen the comparison to match vectype do so. */ | |
8653 | if (INTEGRAL_TYPE_P (scalar_type) | |
8654 | && tree_int_cst_lt (TYPE_SIZE (scalar_type), | |
8655 | TYPE_SIZE (TREE_TYPE (vectype)))) | |
8656 | scalar_type = build_nonstandard_integer_type | |
8657 | (tree_to_uhwi (TYPE_SIZE (TREE_TYPE (vectype))), | |
8658 | TYPE_UNSIGNED (scalar_type)); | |
8659 | *comp_vectype = get_vectype_for_scalar_type (scalar_type); | |
8660 | } | |
8661 | ||
ebfd146a IR |
8662 | return true; |
8663 | } | |
8664 | ||
8665 | /* vectorizable_condition. | |
8666 | ||
32e8e429 RS |
8667 | Check if STMT_INFO is conditional modify expression that can be vectorized. |
8668 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized | |
b8698a0f | 8669 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it |
4bbe8262 IR |
8670 | at GSI. |
8671 | ||
32e8e429 RS |
8672 | When STMT_INFO is vectorized as a nested cycle, REDUC_DEF is the vector |
8673 | variable to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, | |
8674 | and in else clause if it is 2). | |
ebfd146a | 8675 | |
32e8e429 | 8676 | Return true if STMT_INFO is vectorizable in this way. */ |
ebfd146a | 8677 | |
4bbe8262 | 8678 | bool |
32e8e429 | 8679 | vectorizable_condition (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
1eede195 RS |
8680 | stmt_vec_info *vec_stmt, tree reduc_def, |
8681 | int reduc_index, slp_tree slp_node, | |
8682 | stmt_vector_for_cost *cost_vec) | |
ebfd146a | 8683 | { |
e4057a39 | 8684 | vec_info *vinfo = stmt_info->vinfo; |
ebfd146a IR |
8685 | tree scalar_dest = NULL_TREE; |
8686 | tree vec_dest = NULL_TREE; | |
01216d27 JJ |
8687 | tree cond_expr, cond_expr0 = NULL_TREE, cond_expr1 = NULL_TREE; |
8688 | tree then_clause, else_clause; | |
df11cc78 | 8689 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
8690 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
8691 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
5958f9e2 | 8692 | tree vec_compare; |
ebfd146a IR |
8693 | tree new_temp; |
8694 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4fc5ebf1 JG |
8695 | enum vect_def_type dts[4] |
8696 | = {vect_unknown_def_type, vect_unknown_def_type, | |
8697 | vect_unknown_def_type, vect_unknown_def_type}; | |
8698 | int ndts = 4; | |
f7e531cf | 8699 | int ncopies; |
01216d27 | 8700 | enum tree_code code, cond_code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
a855b1b1 | 8701 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
8702 | int i, j; |
8703 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
8704 | vec<tree> vec_oprnds0 = vNULL; |
8705 | vec<tree> vec_oprnds1 = vNULL; | |
8706 | vec<tree> vec_oprnds2 = vNULL; | |
8707 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 8708 | tree vec_cmp_type; |
a414c77f | 8709 | bool masked = false; |
b8698a0f | 8710 | |
f7e531cf IR |
8711 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
8712 | return false; | |
8713 | ||
bb6c2b68 RS |
8714 | vect_reduction_type reduction_type |
8715 | = STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info); | |
8716 | if (reduction_type == TREE_CODE_REDUCTION) | |
af29617a AH |
8717 | { |
8718 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
8719 | return false; | |
ebfd146a | 8720 | |
af29617a AH |
8721 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
8722 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
8723 | && reduc_def)) | |
8724 | return false; | |
ebfd146a | 8725 | |
af29617a AH |
8726 | /* FORNOW: not yet supported. */ |
8727 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
8728 | { | |
8729 | if (dump_enabled_p ()) | |
8730 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
8731 | "value used after loop.\n"); | |
8732 | return false; | |
8733 | } | |
ebfd146a IR |
8734 | } |
8735 | ||
8736 | /* Is vectorizable conditional operation? */ | |
32e8e429 RS |
8737 | gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt); |
8738 | if (!stmt) | |
ebfd146a IR |
8739 | return false; |
8740 | ||
8741 | code = gimple_assign_rhs_code (stmt); | |
8742 | ||
8743 | if (code != COND_EXPR) | |
8744 | return false; | |
8745 | ||
465c8c19 | 8746 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2947d3b2 | 8747 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
465c8c19 | 8748 | |
fce57248 | 8749 | if (slp_node) |
465c8c19 JJ |
8750 | ncopies = 1; |
8751 | else | |
e8f142e2 | 8752 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
8753 | |
8754 | gcc_assert (ncopies >= 1); | |
8755 | if (reduc_index && ncopies > 1) | |
8756 | return false; /* FORNOW */ | |
8757 | ||
4e71066d RG |
8758 | cond_expr = gimple_assign_rhs1 (stmt); |
8759 | then_clause = gimple_assign_rhs2 (stmt); | |
8760 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 8761 | |
4fc5ebf1 | 8762 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, |
4515e413 | 8763 | &comp_vectype, &dts[0], slp_node ? NULL : vectype) |
e9e1d143 | 8764 | || !comp_vectype) |
ebfd146a IR |
8765 | return false; |
8766 | ||
894dd753 | 8767 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &dts[2], &vectype1)) |
2947d3b2 | 8768 | return false; |
894dd753 | 8769 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &dts[3], &vectype2)) |
ebfd146a | 8770 | return false; |
2947d3b2 IE |
8771 | |
8772 | if (vectype1 && !useless_type_conversion_p (vectype, vectype1)) | |
8773 | return false; | |
8774 | ||
8775 | if (vectype2 && !useless_type_conversion_p (vectype, vectype2)) | |
ebfd146a IR |
8776 | return false; |
8777 | ||
28b33016 IE |
8778 | masked = !COMPARISON_CLASS_P (cond_expr); |
8779 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
8780 | ||
74946978 MP |
8781 | if (vec_cmp_type == NULL_TREE) |
8782 | return false; | |
784fb9b3 | 8783 | |
01216d27 JJ |
8784 | cond_code = TREE_CODE (cond_expr); |
8785 | if (!masked) | |
8786 | { | |
8787 | cond_expr0 = TREE_OPERAND (cond_expr, 0); | |
8788 | cond_expr1 = TREE_OPERAND (cond_expr, 1); | |
8789 | } | |
8790 | ||
8791 | if (!masked && VECTOR_BOOLEAN_TYPE_P (comp_vectype)) | |
8792 | { | |
8793 | /* Boolean values may have another representation in vectors | |
8794 | and therefore we prefer bit operations over comparison for | |
8795 | them (which also works for scalar masks). We store opcodes | |
8796 | to use in bitop1 and bitop2. Statement is vectorized as | |
8797 | BITOP2 (rhs1 BITOP1 rhs2) or rhs1 BITOP2 (BITOP1 rhs2) | |
8798 | depending on bitop1 and bitop2 arity. */ | |
8799 | switch (cond_code) | |
8800 | { | |
8801 | case GT_EXPR: | |
8802 | bitop1 = BIT_NOT_EXPR; | |
8803 | bitop2 = BIT_AND_EXPR; | |
8804 | break; | |
8805 | case GE_EXPR: | |
8806 | bitop1 = BIT_NOT_EXPR; | |
8807 | bitop2 = BIT_IOR_EXPR; | |
8808 | break; | |
8809 | case LT_EXPR: | |
8810 | bitop1 = BIT_NOT_EXPR; | |
8811 | bitop2 = BIT_AND_EXPR; | |
8812 | std::swap (cond_expr0, cond_expr1); | |
8813 | break; | |
8814 | case LE_EXPR: | |
8815 | bitop1 = BIT_NOT_EXPR; | |
8816 | bitop2 = BIT_IOR_EXPR; | |
8817 | std::swap (cond_expr0, cond_expr1); | |
8818 | break; | |
8819 | case NE_EXPR: | |
8820 | bitop1 = BIT_XOR_EXPR; | |
8821 | break; | |
8822 | case EQ_EXPR: | |
8823 | bitop1 = BIT_XOR_EXPR; | |
8824 | bitop2 = BIT_NOT_EXPR; | |
8825 | break; | |
8826 | default: | |
8827 | return false; | |
8828 | } | |
8829 | cond_code = SSA_NAME; | |
8830 | } | |
8831 | ||
b8698a0f | 8832 | if (!vec_stmt) |
ebfd146a | 8833 | { |
01216d27 JJ |
8834 | if (bitop1 != NOP_EXPR) |
8835 | { | |
8836 | machine_mode mode = TYPE_MODE (comp_vectype); | |
8837 | optab optab; | |
8838 | ||
8839 | optab = optab_for_tree_code (bitop1, comp_vectype, optab_default); | |
8840 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8841 | return false; | |
8842 | ||
8843 | if (bitop2 != NOP_EXPR) | |
8844 | { | |
8845 | optab = optab_for_tree_code (bitop2, comp_vectype, | |
8846 | optab_default); | |
8847 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8848 | return false; | |
8849 | } | |
8850 | } | |
4fc5ebf1 JG |
8851 | if (expand_vec_cond_expr_p (vectype, comp_vectype, |
8852 | cond_code)) | |
8853 | { | |
68435eb2 RB |
8854 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; |
8855 | vect_model_simple_cost (stmt_info, ncopies, dts, ndts, slp_node, | |
8856 | cost_vec); | |
4fc5ebf1 JG |
8857 | return true; |
8858 | } | |
8859 | return false; | |
ebfd146a IR |
8860 | } |
8861 | ||
f7e531cf IR |
8862 | /* Transform. */ |
8863 | ||
8864 | if (!slp_node) | |
8865 | { | |
9771b263 DN |
8866 | vec_oprnds0.create (1); |
8867 | vec_oprnds1.create (1); | |
8868 | vec_oprnds2.create (1); | |
8869 | vec_oprnds3.create (1); | |
f7e531cf | 8870 | } |
ebfd146a IR |
8871 | |
8872 | /* Handle def. */ | |
8873 | scalar_dest = gimple_assign_lhs (stmt); | |
bb6c2b68 RS |
8874 | if (reduction_type != EXTRACT_LAST_REDUCTION) |
8875 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
ebfd146a IR |
8876 | |
8877 | /* Handle cond expr. */ | |
a855b1b1 MM |
8878 | for (j = 0; j < ncopies; j++) |
8879 | { | |
e1bd7296 | 8880 | stmt_vec_info new_stmt_info = NULL; |
a855b1b1 MM |
8881 | if (j == 0) |
8882 | { | |
f7e531cf IR |
8883 | if (slp_node) |
8884 | { | |
00f96dc9 TS |
8885 | auto_vec<tree, 4> ops; |
8886 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 8887 | |
a414c77f | 8888 | if (masked) |
01216d27 | 8889 | ops.safe_push (cond_expr); |
a414c77f IE |
8890 | else |
8891 | { | |
01216d27 JJ |
8892 | ops.safe_push (cond_expr0); |
8893 | ops.safe_push (cond_expr1); | |
a414c77f | 8894 | } |
9771b263 DN |
8895 | ops.safe_push (then_clause); |
8896 | ops.safe_push (else_clause); | |
306b0c92 | 8897 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
37b5ec8f JJ |
8898 | vec_oprnds3 = vec_defs.pop (); |
8899 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
8900 | if (!masked) |
8901 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 8902 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf IR |
8903 | } |
8904 | else | |
8905 | { | |
a414c77f IE |
8906 | if (masked) |
8907 | { | |
8908 | vec_cond_lhs | |
86a91c0a | 8909 | = vect_get_vec_def_for_operand (cond_expr, stmt_info, |
a414c77f | 8910 | comp_vectype); |
894dd753 | 8911 | vect_is_simple_use (cond_expr, stmt_info->vinfo, &dts[0]); |
a414c77f IE |
8912 | } |
8913 | else | |
8914 | { | |
01216d27 JJ |
8915 | vec_cond_lhs |
8916 | = vect_get_vec_def_for_operand (cond_expr0, | |
86a91c0a | 8917 | stmt_info, comp_vectype); |
894dd753 | 8918 | vect_is_simple_use (cond_expr0, loop_vinfo, &dts[0]); |
01216d27 JJ |
8919 | |
8920 | vec_cond_rhs | |
8921 | = vect_get_vec_def_for_operand (cond_expr1, | |
86a91c0a | 8922 | stmt_info, comp_vectype); |
894dd753 | 8923 | vect_is_simple_use (cond_expr1, loop_vinfo, &dts[1]); |
a414c77f | 8924 | } |
f7e531cf IR |
8925 | if (reduc_index == 1) |
8926 | vec_then_clause = reduc_def; | |
8927 | else | |
8928 | { | |
8929 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
86a91c0a | 8930 | stmt_info); |
894dd753 | 8931 | vect_is_simple_use (then_clause, loop_vinfo, &dts[2]); |
f7e531cf IR |
8932 | } |
8933 | if (reduc_index == 2) | |
8934 | vec_else_clause = reduc_def; | |
8935 | else | |
8936 | { | |
8937 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
86a91c0a | 8938 | stmt_info); |
894dd753 | 8939 | vect_is_simple_use (else_clause, loop_vinfo, &dts[3]); |
f7e531cf | 8940 | } |
a855b1b1 MM |
8941 | } |
8942 | } | |
8943 | else | |
8944 | { | |
a414c77f | 8945 | vec_cond_lhs |
e4057a39 | 8946 | = vect_get_vec_def_for_stmt_copy (vinfo, vec_oprnds0.pop ()); |
a414c77f IE |
8947 | if (!masked) |
8948 | vec_cond_rhs | |
e4057a39 | 8949 | = vect_get_vec_def_for_stmt_copy (vinfo, vec_oprnds1.pop ()); |
a414c77f | 8950 | |
e4057a39 | 8951 | vec_then_clause = vect_get_vec_def_for_stmt_copy (vinfo, |
9771b263 | 8952 | vec_oprnds2.pop ()); |
e4057a39 | 8953 | vec_else_clause = vect_get_vec_def_for_stmt_copy (vinfo, |
9771b263 | 8954 | vec_oprnds3.pop ()); |
f7e531cf IR |
8955 | } |
8956 | ||
8957 | if (!slp_node) | |
8958 | { | |
9771b263 | 8959 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
8960 | if (!masked) |
8961 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
8962 | vec_oprnds2.quick_push (vec_then_clause); |
8963 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
8964 | } |
8965 | ||
9dc3f7de | 8966 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 8967 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 8968 | { |
9771b263 DN |
8969 | vec_then_clause = vec_oprnds2[i]; |
8970 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 8971 | |
a414c77f IE |
8972 | if (masked) |
8973 | vec_compare = vec_cond_lhs; | |
8974 | else | |
8975 | { | |
8976 | vec_cond_rhs = vec_oprnds1[i]; | |
01216d27 JJ |
8977 | if (bitop1 == NOP_EXPR) |
8978 | vec_compare = build2 (cond_code, vec_cmp_type, | |
8979 | vec_cond_lhs, vec_cond_rhs); | |
8980 | else | |
8981 | { | |
8982 | new_temp = make_ssa_name (vec_cmp_type); | |
e1bd7296 | 8983 | gassign *new_stmt; |
01216d27 JJ |
8984 | if (bitop1 == BIT_NOT_EXPR) |
8985 | new_stmt = gimple_build_assign (new_temp, bitop1, | |
8986 | vec_cond_rhs); | |
8987 | else | |
8988 | new_stmt | |
8989 | = gimple_build_assign (new_temp, bitop1, vec_cond_lhs, | |
8990 | vec_cond_rhs); | |
86a91c0a | 8991 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
01216d27 JJ |
8992 | if (bitop2 == NOP_EXPR) |
8993 | vec_compare = new_temp; | |
8994 | else if (bitop2 == BIT_NOT_EXPR) | |
8995 | { | |
8996 | /* Instead of doing ~x ? y : z do x ? z : y. */ | |
8997 | vec_compare = new_temp; | |
8998 | std::swap (vec_then_clause, vec_else_clause); | |
8999 | } | |
9000 | else | |
9001 | { | |
9002 | vec_compare = make_ssa_name (vec_cmp_type); | |
9003 | new_stmt | |
9004 | = gimple_build_assign (vec_compare, bitop2, | |
9005 | vec_cond_lhs, new_temp); | |
86a91c0a | 9006 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
01216d27 JJ |
9007 | } |
9008 | } | |
a414c77f | 9009 | } |
bb6c2b68 RS |
9010 | if (reduction_type == EXTRACT_LAST_REDUCTION) |
9011 | { | |
9012 | if (!is_gimple_val (vec_compare)) | |
9013 | { | |
9014 | tree vec_compare_name = make_ssa_name (vec_cmp_type); | |
e1bd7296 RS |
9015 | gassign *new_stmt = gimple_build_assign (vec_compare_name, |
9016 | vec_compare); | |
86a91c0a | 9017 | vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
bb6c2b68 RS |
9018 | vec_compare = vec_compare_name; |
9019 | } | |
9020 | gcc_assert (reduc_index == 2); | |
e1bd7296 | 9021 | gcall *new_stmt = gimple_build_call_internal |
bb6c2b68 RS |
9022 | (IFN_FOLD_EXTRACT_LAST, 3, else_clause, vec_compare, |
9023 | vec_then_clause); | |
9024 | gimple_call_set_lhs (new_stmt, scalar_dest); | |
9025 | SSA_NAME_DEF_STMT (scalar_dest) = new_stmt; | |
86a91c0a RS |
9026 | if (stmt_info->stmt == gsi_stmt (*gsi)) |
9027 | new_stmt_info = vect_finish_replace_stmt (stmt_info, new_stmt); | |
bb6c2b68 RS |
9028 | else |
9029 | { | |
9030 | /* In this case we're moving the definition to later in the | |
9031 | block. That doesn't matter because the only uses of the | |
9032 | lhs are in phi statements. */ | |
86a91c0a RS |
9033 | gimple_stmt_iterator old_gsi |
9034 | = gsi_for_stmt (stmt_info->stmt); | |
bb6c2b68 | 9035 | gsi_remove (&old_gsi, true); |
e1bd7296 | 9036 | new_stmt_info |
86a91c0a | 9037 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
bb6c2b68 RS |
9038 | } |
9039 | } | |
9040 | else | |
9041 | { | |
9042 | new_temp = make_ssa_name (vec_dest); | |
e1bd7296 RS |
9043 | gassign *new_stmt |
9044 | = gimple_build_assign (new_temp, VEC_COND_EXPR, vec_compare, | |
9045 | vec_then_clause, vec_else_clause); | |
9046 | new_stmt_info | |
86a91c0a | 9047 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
bb6c2b68 | 9048 | } |
f7e531cf | 9049 | if (slp_node) |
e1bd7296 | 9050 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
f7e531cf IR |
9051 | } |
9052 | ||
9053 | if (slp_node) | |
9054 | continue; | |
9055 | ||
e1bd7296 RS |
9056 | if (j == 0) |
9057 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; | |
9058 | else | |
9059 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; | |
f7e531cf | 9060 | |
e1bd7296 | 9061 | prev_stmt_info = new_stmt_info; |
a855b1b1 | 9062 | } |
b8698a0f | 9063 | |
9771b263 DN |
9064 | vec_oprnds0.release (); |
9065 | vec_oprnds1.release (); | |
9066 | vec_oprnds2.release (); | |
9067 | vec_oprnds3.release (); | |
f7e531cf | 9068 | |
ebfd146a IR |
9069 | return true; |
9070 | } | |
9071 | ||
42fd8198 IE |
9072 | /* vectorizable_comparison. |
9073 | ||
32e8e429 RS |
9074 | Check if STMT_INFO is comparison expression that can be vectorized. |
9075 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized | |
42fd8198 IE |
9076 | comparison, put it in VEC_STMT, and insert it at GSI. |
9077 | ||
32e8e429 | 9078 | Return true if STMT_INFO is vectorizable in this way. */ |
42fd8198 | 9079 | |
fce57248 | 9080 | static bool |
32e8e429 | 9081 | vectorizable_comparison (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
1eede195 | 9082 | stmt_vec_info *vec_stmt, tree reduc_def, |
68435eb2 | 9083 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
42fd8198 | 9084 | { |
e4057a39 | 9085 | vec_info *vinfo = stmt_info->vinfo; |
42fd8198 | 9086 | tree lhs, rhs1, rhs2; |
42fd8198 IE |
9087 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
9088 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
9089 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
9090 | tree new_temp; | |
9091 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
9092 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 9093 | int ndts = 2; |
928686b1 | 9094 | poly_uint64 nunits; |
42fd8198 | 9095 | int ncopies; |
49e76ff1 | 9096 | enum tree_code code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
42fd8198 IE |
9097 | stmt_vec_info prev_stmt_info = NULL; |
9098 | int i, j; | |
9099 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
9100 | vec<tree> vec_oprnds0 = vNULL; | |
9101 | vec<tree> vec_oprnds1 = vNULL; | |
42fd8198 IE |
9102 | tree mask_type; |
9103 | tree mask; | |
9104 | ||
c245362b IE |
9105 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
9106 | return false; | |
9107 | ||
30480bcd | 9108 | if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)) |
42fd8198 IE |
9109 | return false; |
9110 | ||
9111 | mask_type = vectype; | |
9112 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
9113 | ||
fce57248 | 9114 | if (slp_node) |
42fd8198 IE |
9115 | ncopies = 1; |
9116 | else | |
e8f142e2 | 9117 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
42fd8198 IE |
9118 | |
9119 | gcc_assert (ncopies >= 1); | |
42fd8198 IE |
9120 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
9121 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
9122 | && reduc_def)) | |
9123 | return false; | |
9124 | ||
9125 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
9126 | { | |
9127 | if (dump_enabled_p ()) | |
9128 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
9129 | "value used after loop.\n"); | |
9130 | return false; | |
9131 | } | |
9132 | ||
32e8e429 RS |
9133 | gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt); |
9134 | if (!stmt) | |
42fd8198 IE |
9135 | return false; |
9136 | ||
9137 | code = gimple_assign_rhs_code (stmt); | |
9138 | ||
9139 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
9140 | return false; | |
9141 | ||
9142 | rhs1 = gimple_assign_rhs1 (stmt); | |
9143 | rhs2 = gimple_assign_rhs2 (stmt); | |
9144 | ||
894dd753 | 9145 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &dts[0], &vectype1)) |
42fd8198 IE |
9146 | return false; |
9147 | ||
894dd753 | 9148 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &dts[1], &vectype2)) |
42fd8198 IE |
9149 | return false; |
9150 | ||
9151 | if (vectype1 && vectype2 | |
928686b1 RS |
9152 | && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1), |
9153 | TYPE_VECTOR_SUBPARTS (vectype2))) | |
42fd8198 IE |
9154 | return false; |
9155 | ||
9156 | vectype = vectype1 ? vectype1 : vectype2; | |
9157 | ||
9158 | /* Invariant comparison. */ | |
9159 | if (!vectype) | |
9160 | { | |
69a9a66f | 9161 | vectype = get_vectype_for_scalar_type (TREE_TYPE (rhs1)); |
928686b1 | 9162 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (vectype), nunits)) |
42fd8198 IE |
9163 | return false; |
9164 | } | |
928686b1 | 9165 | else if (maybe_ne (nunits, TYPE_VECTOR_SUBPARTS (vectype))) |
42fd8198 IE |
9166 | return false; |
9167 | ||
49e76ff1 IE |
9168 | /* Can't compare mask and non-mask types. */ |
9169 | if (vectype1 && vectype2 | |
9170 | && (VECTOR_BOOLEAN_TYPE_P (vectype1) ^ VECTOR_BOOLEAN_TYPE_P (vectype2))) | |
9171 | return false; | |
9172 | ||
9173 | /* Boolean values may have another representation in vectors | |
9174 | and therefore we prefer bit operations over comparison for | |
9175 | them (which also works for scalar masks). We store opcodes | |
9176 | to use in bitop1 and bitop2. Statement is vectorized as | |
9177 | BITOP2 (rhs1 BITOP1 rhs2) or | |
9178 | rhs1 BITOP2 (BITOP1 rhs2) | |
9179 | depending on bitop1 and bitop2 arity. */ | |
9180 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) | |
9181 | { | |
9182 | if (code == GT_EXPR) | |
9183 | { | |
9184 | bitop1 = BIT_NOT_EXPR; | |
9185 | bitop2 = BIT_AND_EXPR; | |
9186 | } | |
9187 | else if (code == GE_EXPR) | |
9188 | { | |
9189 | bitop1 = BIT_NOT_EXPR; | |
9190 | bitop2 = BIT_IOR_EXPR; | |
9191 | } | |
9192 | else if (code == LT_EXPR) | |
9193 | { | |
9194 | bitop1 = BIT_NOT_EXPR; | |
9195 | bitop2 = BIT_AND_EXPR; | |
9196 | std::swap (rhs1, rhs2); | |
264d951a | 9197 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
9198 | } |
9199 | else if (code == LE_EXPR) | |
9200 | { | |
9201 | bitop1 = BIT_NOT_EXPR; | |
9202 | bitop2 = BIT_IOR_EXPR; | |
9203 | std::swap (rhs1, rhs2); | |
264d951a | 9204 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
9205 | } |
9206 | else | |
9207 | { | |
9208 | bitop1 = BIT_XOR_EXPR; | |
9209 | if (code == EQ_EXPR) | |
9210 | bitop2 = BIT_NOT_EXPR; | |
9211 | } | |
9212 | } | |
9213 | ||
42fd8198 IE |
9214 | if (!vec_stmt) |
9215 | { | |
49e76ff1 | 9216 | if (bitop1 == NOP_EXPR) |
68435eb2 RB |
9217 | { |
9218 | if (!expand_vec_cmp_expr_p (vectype, mask_type, code)) | |
9219 | return false; | |
9220 | } | |
49e76ff1 IE |
9221 | else |
9222 | { | |
9223 | machine_mode mode = TYPE_MODE (vectype); | |
9224 | optab optab; | |
9225 | ||
9226 | optab = optab_for_tree_code (bitop1, vectype, optab_default); | |
9227 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
9228 | return false; | |
9229 | ||
9230 | if (bitop2 != NOP_EXPR) | |
9231 | { | |
9232 | optab = optab_for_tree_code (bitop2, vectype, optab_default); | |
9233 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
9234 | return false; | |
9235 | } | |
49e76ff1 | 9236 | } |
68435eb2 RB |
9237 | |
9238 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
9239 | vect_model_simple_cost (stmt_info, ncopies * (1 + (bitop2 != NOP_EXPR)), | |
9240 | dts, ndts, slp_node, cost_vec); | |
9241 | return true; | |
42fd8198 IE |
9242 | } |
9243 | ||
9244 | /* Transform. */ | |
9245 | if (!slp_node) | |
9246 | { | |
9247 | vec_oprnds0.create (1); | |
9248 | vec_oprnds1.create (1); | |
9249 | } | |
9250 | ||
9251 | /* Handle def. */ | |
9252 | lhs = gimple_assign_lhs (stmt); | |
9253 | mask = vect_create_destination_var (lhs, mask_type); | |
9254 | ||
9255 | /* Handle cmp expr. */ | |
9256 | for (j = 0; j < ncopies; j++) | |
9257 | { | |
e1bd7296 | 9258 | stmt_vec_info new_stmt_info = NULL; |
42fd8198 IE |
9259 | if (j == 0) |
9260 | { | |
9261 | if (slp_node) | |
9262 | { | |
9263 | auto_vec<tree, 2> ops; | |
9264 | auto_vec<vec<tree>, 2> vec_defs; | |
9265 | ||
9266 | ops.safe_push (rhs1); | |
9267 | ops.safe_push (rhs2); | |
306b0c92 | 9268 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
42fd8198 IE |
9269 | vec_oprnds1 = vec_defs.pop (); |
9270 | vec_oprnds0 = vec_defs.pop (); | |
9271 | } | |
9272 | else | |
9273 | { | |
86a91c0a RS |
9274 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt_info, |
9275 | vectype); | |
9276 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt_info, | |
9277 | vectype); | |
42fd8198 IE |
9278 | } |
9279 | } | |
9280 | else | |
9281 | { | |
e4057a39 | 9282 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (vinfo, |
42fd8198 | 9283 | vec_oprnds0.pop ()); |
e4057a39 | 9284 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (vinfo, |
42fd8198 IE |
9285 | vec_oprnds1.pop ()); |
9286 | } | |
9287 | ||
9288 | if (!slp_node) | |
9289 | { | |
9290 | vec_oprnds0.quick_push (vec_rhs1); | |
9291 | vec_oprnds1.quick_push (vec_rhs2); | |
9292 | } | |
9293 | ||
9294 | /* Arguments are ready. Create the new vector stmt. */ | |
9295 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
9296 | { | |
9297 | vec_rhs2 = vec_oprnds1[i]; | |
9298 | ||
9299 | new_temp = make_ssa_name (mask); | |
49e76ff1 IE |
9300 | if (bitop1 == NOP_EXPR) |
9301 | { | |
e1bd7296 RS |
9302 | gassign *new_stmt = gimple_build_assign (new_temp, code, |
9303 | vec_rhs1, vec_rhs2); | |
9304 | new_stmt_info | |
86a91c0a | 9305 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
49e76ff1 IE |
9306 | } |
9307 | else | |
9308 | { | |
e1bd7296 | 9309 | gassign *new_stmt; |
49e76ff1 IE |
9310 | if (bitop1 == BIT_NOT_EXPR) |
9311 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs2); | |
9312 | else | |
9313 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs1, | |
9314 | vec_rhs2); | |
e1bd7296 | 9315 | new_stmt_info |
86a91c0a | 9316 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
49e76ff1 IE |
9317 | if (bitop2 != NOP_EXPR) |
9318 | { | |
9319 | tree res = make_ssa_name (mask); | |
9320 | if (bitop2 == BIT_NOT_EXPR) | |
9321 | new_stmt = gimple_build_assign (res, bitop2, new_temp); | |
9322 | else | |
9323 | new_stmt = gimple_build_assign (res, bitop2, vec_rhs1, | |
9324 | new_temp); | |
e1bd7296 | 9325 | new_stmt_info |
86a91c0a | 9326 | = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); |
49e76ff1 IE |
9327 | } |
9328 | } | |
42fd8198 | 9329 | if (slp_node) |
e1bd7296 | 9330 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); |
42fd8198 IE |
9331 | } |
9332 | ||
9333 | if (slp_node) | |
9334 | continue; | |
9335 | ||
9336 | if (j == 0) | |
e1bd7296 | 9337 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; |
42fd8198 | 9338 | else |
e1bd7296 | 9339 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; |
42fd8198 | 9340 | |
e1bd7296 | 9341 | prev_stmt_info = new_stmt_info; |
42fd8198 IE |
9342 | } |
9343 | ||
9344 | vec_oprnds0.release (); | |
9345 | vec_oprnds1.release (); | |
9346 | ||
9347 | return true; | |
9348 | } | |
ebfd146a | 9349 | |
68a0f2ff RS |
9350 | /* If SLP_NODE is nonnull, return true if vectorizable_live_operation |
9351 | can handle all live statements in the node. Otherwise return true | |
82570274 | 9352 | if STMT_INFO is not live or if vectorizable_live_operation can handle it. |
68a0f2ff RS |
9353 | GSI and VEC_STMT are as for vectorizable_live_operation. */ |
9354 | ||
9355 | static bool | |
82570274 | 9356 | can_vectorize_live_stmts (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
1eede195 | 9357 | slp_tree slp_node, stmt_vec_info *vec_stmt, |
68435eb2 | 9358 | stmt_vector_for_cost *cost_vec) |
68a0f2ff RS |
9359 | { |
9360 | if (slp_node) | |
9361 | { | |
b9787581 | 9362 | stmt_vec_info slp_stmt_info; |
68a0f2ff | 9363 | unsigned int i; |
b9787581 | 9364 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, slp_stmt_info) |
68a0f2ff | 9365 | { |
68a0f2ff | 9366 | if (STMT_VINFO_LIVE_P (slp_stmt_info) |
b9787581 | 9367 | && !vectorizable_live_operation (slp_stmt_info, gsi, slp_node, i, |
68435eb2 | 9368 | vec_stmt, cost_vec)) |
68a0f2ff RS |
9369 | return false; |
9370 | } | |
9371 | } | |
82570274 RS |
9372 | else if (STMT_VINFO_LIVE_P (stmt_info) |
9373 | && !vectorizable_live_operation (stmt_info, gsi, slp_node, -1, | |
9374 | vec_stmt, cost_vec)) | |
68a0f2ff RS |
9375 | return false; |
9376 | ||
9377 | return true; | |
9378 | } | |
9379 | ||
8644a673 | 9380 | /* Make sure the statement is vectorizable. */ |
ebfd146a | 9381 | |
f4ebbd24 | 9382 | opt_result |
32e8e429 RS |
9383 | vect_analyze_stmt (stmt_vec_info stmt_info, bool *need_to_vectorize, |
9384 | slp_tree node, slp_instance node_instance, | |
9385 | stmt_vector_for_cost *cost_vec) | |
ebfd146a | 9386 | { |
6585ff8f | 9387 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 9388 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 9389 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 9390 | bool ok; |
363477c0 | 9391 | gimple_seq pattern_def_seq; |
ebfd146a | 9392 | |
73fbfcad | 9393 | if (dump_enabled_p ()) |
3c2a8ed0 DM |
9394 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: %G", |
9395 | stmt_info->stmt); | |
ebfd146a | 9396 | |
86a91c0a | 9397 | if (gimple_has_volatile_ops (stmt_info->stmt)) |
f4ebbd24 DM |
9398 | return opt_result::failure_at (stmt_info->stmt, |
9399 | "not vectorized:" | |
9400 | " stmt has volatile operands: %G\n", | |
9401 | stmt_info->stmt); | |
b8698a0f | 9402 | |
d54a098e RS |
9403 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
9404 | && node == NULL | |
9405 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) | |
9406 | { | |
9407 | gimple_stmt_iterator si; | |
9408 | ||
9409 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) | |
9410 | { | |
6585ff8f RS |
9411 | stmt_vec_info pattern_def_stmt_info |
9412 | = vinfo->lookup_stmt (gsi_stmt (si)); | |
9413 | if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info) | |
9414 | || STMT_VINFO_LIVE_P (pattern_def_stmt_info)) | |
d54a098e RS |
9415 | { |
9416 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
9417 | if (dump_enabled_p ()) | |
3c2a8ed0 DM |
9418 | dump_printf_loc (MSG_NOTE, vect_location, |
9419 | "==> examining pattern def statement: %G", | |
9420 | pattern_def_stmt_info->stmt); | |
d54a098e | 9421 | |
f4ebbd24 DM |
9422 | opt_result res |
9423 | = vect_analyze_stmt (pattern_def_stmt_info, | |
9424 | need_to_vectorize, node, node_instance, | |
9425 | cost_vec); | |
9426 | if (!res) | |
9427 | return res; | |
d54a098e RS |
9428 | } |
9429 | } | |
9430 | } | |
9431 | ||
b8698a0f | 9432 | /* Skip stmts that do not need to be vectorized. In loops this is expected |
8644a673 IR |
9433 | to include: |
9434 | - the COND_EXPR which is the loop exit condition | |
9435 | - any LABEL_EXPRs in the loop | |
b8698a0f | 9436 | - computations that are used only for array indexing or loop control. |
8644a673 | 9437 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 9438 | instance, therefore, all the statements are relevant. |
ebfd146a | 9439 | |
d092494c | 9440 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 9441 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
9442 | statements. In basic blocks we are called from some SLP instance |
9443 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
9444 | already will be part of SLP instance. */ | |
83197f37 | 9445 | |
10681ce8 | 9446 | stmt_vec_info pattern_stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); |
b8698a0f | 9447 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 9448 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 9449 | { |
9d5e7640 | 9450 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
10681ce8 RS |
9451 | && pattern_stmt_info |
9452 | && (STMT_VINFO_RELEVANT_P (pattern_stmt_info) | |
9453 | || STMT_VINFO_LIVE_P (pattern_stmt_info))) | |
9d5e7640 | 9454 | { |
83197f37 | 9455 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
10681ce8 | 9456 | stmt_info = pattern_stmt_info; |
73fbfcad | 9457 | if (dump_enabled_p ()) |
3c2a8ed0 DM |
9458 | dump_printf_loc (MSG_NOTE, vect_location, |
9459 | "==> examining pattern statement: %G", | |
9460 | stmt_info->stmt); | |
9d5e7640 IR |
9461 | } |
9462 | else | |
9463 | { | |
73fbfcad | 9464 | if (dump_enabled_p ()) |
e645e942 | 9465 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 9466 | |
f4ebbd24 | 9467 | return opt_result::success (); |
9d5e7640 | 9468 | } |
8644a673 | 9469 | } |
83197f37 | 9470 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 9471 | && node == NULL |
10681ce8 RS |
9472 | && pattern_stmt_info |
9473 | && (STMT_VINFO_RELEVANT_P (pattern_stmt_info) | |
9474 | || STMT_VINFO_LIVE_P (pattern_stmt_info))) | |
83197f37 IR |
9475 | { |
9476 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 9477 | if (dump_enabled_p ()) |
3c2a8ed0 DM |
9478 | dump_printf_loc (MSG_NOTE, vect_location, |
9479 | "==> examining pattern statement: %G", | |
9480 | pattern_stmt_info->stmt); | |
83197f37 | 9481 | |
f4ebbd24 DM |
9482 | opt_result res |
9483 | = vect_analyze_stmt (pattern_stmt_info, need_to_vectorize, node, | |
9484 | node_instance, cost_vec); | |
9485 | if (!res) | |
9486 | return res; | |
83197f37 | 9487 | } |
ebfd146a | 9488 | |
8644a673 IR |
9489 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
9490 | { | |
9491 | case vect_internal_def: | |
9492 | break; | |
ebfd146a | 9493 | |
8644a673 | 9494 | case vect_reduction_def: |
7c5222ff | 9495 | case vect_nested_cycle: |
14a61437 RB |
9496 | gcc_assert (!bb_vinfo |
9497 | && (relevance == vect_used_in_outer | |
9498 | || relevance == vect_used_in_outer_by_reduction | |
9499 | || relevance == vect_used_by_reduction | |
b28ead45 AH |
9500 | || relevance == vect_unused_in_scope |
9501 | || relevance == vect_used_only_live)); | |
8644a673 IR |
9502 | break; |
9503 | ||
9504 | case vect_induction_def: | |
e7baeb39 RB |
9505 | gcc_assert (!bb_vinfo); |
9506 | break; | |
9507 | ||
8644a673 IR |
9508 | case vect_constant_def: |
9509 | case vect_external_def: | |
9510 | case vect_unknown_def_type: | |
9511 | default: | |
9512 | gcc_unreachable (); | |
9513 | } | |
ebfd146a | 9514 | |
8644a673 | 9515 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 9516 | { |
86a91c0a RS |
9517 | tree type = gimple_expr_type (stmt_info->stmt); |
9518 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (type))); | |
9519 | gcall *call = dyn_cast <gcall *> (stmt_info->stmt); | |
0136f8f0 | 9520 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
beb456c3 | 9521 | || (call && gimple_call_lhs (call) == NULL_TREE)); |
8644a673 | 9522 | *need_to_vectorize = true; |
ebfd146a IR |
9523 | } |
9524 | ||
b1af7da6 RB |
9525 | if (PURE_SLP_STMT (stmt_info) && !node) |
9526 | { | |
9527 | dump_printf_loc (MSG_NOTE, vect_location, | |
9528 | "handled only by SLP analysis\n"); | |
f4ebbd24 | 9529 | return opt_result::success (); |
b1af7da6 RB |
9530 | } |
9531 | ||
9532 | ok = true; | |
9533 | if (!bb_vinfo | |
9534 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
9535 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
bc37759a RB |
9536 | /* Prefer vectorizable_call over vectorizable_simd_clone_call so |
9537 | -mveclibabi= takes preference over ibrary functions with | |
9538 | the simd attribute. */ | |
9539 | ok = (vectorizable_call (stmt_info, NULL, NULL, node, cost_vec) | |
9540 | || vectorizable_simd_clone_call (stmt_info, NULL, NULL, node, | |
9541 | cost_vec) | |
86a91c0a RS |
9542 | || vectorizable_conversion (stmt_info, NULL, NULL, node, cost_vec) |
9543 | || vectorizable_shift (stmt_info, NULL, NULL, node, cost_vec) | |
9544 | || vectorizable_operation (stmt_info, NULL, NULL, node, cost_vec) | |
9545 | || vectorizable_assignment (stmt_info, NULL, NULL, node, cost_vec) | |
9546 | || vectorizable_load (stmt_info, NULL, NULL, node, node_instance, | |
9547 | cost_vec) | |
86a91c0a RS |
9548 | || vectorizable_store (stmt_info, NULL, NULL, node, cost_vec) |
9549 | || vectorizable_reduction (stmt_info, NULL, NULL, node, | |
9550 | node_instance, cost_vec) | |
9551 | || vectorizable_induction (stmt_info, NULL, NULL, node, cost_vec) | |
9552 | || vectorizable_condition (stmt_info, NULL, NULL, NULL, 0, node, | |
68435eb2 | 9553 | cost_vec) |
86a91c0a RS |
9554 | || vectorizable_comparison (stmt_info, NULL, NULL, NULL, node, |
9555 | cost_vec)); | |
b1af7da6 RB |
9556 | else |
9557 | { | |
9558 | if (bb_vinfo) | |
bc37759a RB |
9559 | ok = (vectorizable_call (stmt_info, NULL, NULL, node, cost_vec) |
9560 | || vectorizable_simd_clone_call (stmt_info, NULL, NULL, node, | |
9561 | cost_vec) | |
86a91c0a RS |
9562 | || vectorizable_conversion (stmt_info, NULL, NULL, node, |
9563 | cost_vec) | |
9564 | || vectorizable_shift (stmt_info, NULL, NULL, node, cost_vec) | |
9565 | || vectorizable_operation (stmt_info, NULL, NULL, node, cost_vec) | |
9566 | || vectorizable_assignment (stmt_info, NULL, NULL, node, | |
9567 | cost_vec) | |
9568 | || vectorizable_load (stmt_info, NULL, NULL, node, node_instance, | |
68435eb2 | 9569 | cost_vec) |
86a91c0a RS |
9570 | || vectorizable_store (stmt_info, NULL, NULL, node, cost_vec) |
9571 | || vectorizable_condition (stmt_info, NULL, NULL, NULL, 0, node, | |
68435eb2 | 9572 | cost_vec) |
86a91c0a | 9573 | || vectorizable_comparison (stmt_info, NULL, NULL, NULL, node, |
68435eb2 | 9574 | cost_vec)); |
b1af7da6 | 9575 | } |
8644a673 IR |
9576 | |
9577 | if (!ok) | |
f4ebbd24 DM |
9578 | return opt_result::failure_at (stmt_info->stmt, |
9579 | "not vectorized:" | |
9580 | " relevant stmt not supported: %G", | |
9581 | stmt_info->stmt); | |
ebfd146a | 9582 | |
8644a673 IR |
9583 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
9584 | need extra handling, except for vectorizable reductions. */ | |
68435eb2 RB |
9585 | if (!bb_vinfo |
9586 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type | |
86a91c0a | 9587 | && !can_vectorize_live_stmts (stmt_info, NULL, node, NULL, cost_vec)) |
f4ebbd24 DM |
9588 | return opt_result::failure_at (stmt_info->stmt, |
9589 | "not vectorized:" | |
9590 | " live stmt not supported: %G", | |
9591 | stmt_info->stmt); | |
b8698a0f | 9592 | |
f4ebbd24 | 9593 | return opt_result::success (); |
ebfd146a IR |
9594 | } |
9595 | ||
9596 | ||
9597 | /* Function vect_transform_stmt. | |
9598 | ||
32e8e429 | 9599 | Create a vectorized stmt to replace STMT_INFO, and insert it at BSI. */ |
ebfd146a IR |
9600 | |
9601 | bool | |
32e8e429 | 9602 | vect_transform_stmt (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
b0b45e58 | 9603 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a | 9604 | { |
6585ff8f | 9605 | vec_info *vinfo = stmt_info->vinfo; |
ebfd146a | 9606 | bool is_store = false; |
1eede195 | 9607 | stmt_vec_info vec_stmt = NULL; |
ebfd146a | 9608 | bool done; |
ebfd146a | 9609 | |
fce57248 | 9610 | gcc_assert (slp_node || !PURE_SLP_STMT (stmt_info)); |
1eede195 | 9611 | stmt_vec_info old_vec_stmt_info = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 9612 | |
e57d9a82 RB |
9613 | bool nested_p = (STMT_VINFO_LOOP_VINFO (stmt_info) |
9614 | && nested_in_vect_loop_p | |
9615 | (LOOP_VINFO_LOOP (STMT_VINFO_LOOP_VINFO (stmt_info)), | |
86a91c0a | 9616 | stmt_info)); |
e57d9a82 | 9617 | |
32e8e429 | 9618 | gimple *stmt = stmt_info->stmt; |
ebfd146a IR |
9619 | switch (STMT_VINFO_TYPE (stmt_info)) |
9620 | { | |
9621 | case type_demotion_vec_info_type: | |
ebfd146a | 9622 | case type_promotion_vec_info_type: |
ebfd146a | 9623 | case type_conversion_vec_info_type: |
86a91c0a RS |
9624 | done = vectorizable_conversion (stmt_info, gsi, &vec_stmt, slp_node, |
9625 | NULL); | |
ebfd146a IR |
9626 | gcc_assert (done); |
9627 | break; | |
9628 | ||
9629 | case induc_vec_info_type: | |
86a91c0a RS |
9630 | done = vectorizable_induction (stmt_info, gsi, &vec_stmt, slp_node, |
9631 | NULL); | |
ebfd146a IR |
9632 | gcc_assert (done); |
9633 | break; | |
9634 | ||
9dc3f7de | 9635 | case shift_vec_info_type: |
86a91c0a | 9636 | done = vectorizable_shift (stmt_info, gsi, &vec_stmt, slp_node, NULL); |
9dc3f7de IR |
9637 | gcc_assert (done); |
9638 | break; | |
9639 | ||
ebfd146a | 9640 | case op_vec_info_type: |
86a91c0a RS |
9641 | done = vectorizable_operation (stmt_info, gsi, &vec_stmt, slp_node, |
9642 | NULL); | |
ebfd146a IR |
9643 | gcc_assert (done); |
9644 | break; | |
9645 | ||
9646 | case assignment_vec_info_type: | |
86a91c0a RS |
9647 | done = vectorizable_assignment (stmt_info, gsi, &vec_stmt, slp_node, |
9648 | NULL); | |
ebfd146a IR |
9649 | gcc_assert (done); |
9650 | break; | |
9651 | ||
9652 | case load_vec_info_type: | |
86a91c0a | 9653 | done = vectorizable_load (stmt_info, gsi, &vec_stmt, slp_node, |
68435eb2 | 9654 | slp_node_instance, NULL); |
ebfd146a IR |
9655 | gcc_assert (done); |
9656 | break; | |
9657 | ||
9658 | case store_vec_info_type: | |
86a91c0a | 9659 | done = vectorizable_store (stmt_info, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a | 9660 | gcc_assert (done); |
0d0293ac | 9661 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
9662 | { |
9663 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 9664 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
9665 | one are skipped, and there vec_stmt_info shouldn't be freed |
9666 | meanwhile. */ | |
bffb8014 | 9667 | stmt_vec_info group_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
2c53b149 | 9668 | if (DR_GROUP_STORE_COUNT (group_info) == DR_GROUP_SIZE (group_info)) |
ebfd146a | 9669 | is_store = true; |
f307441a | 9670 | } |
ebfd146a IR |
9671 | else |
9672 | is_store = true; | |
9673 | break; | |
9674 | ||
9675 | case condition_vec_info_type: | |
86a91c0a RS |
9676 | done = vectorizable_condition (stmt_info, gsi, &vec_stmt, NULL, 0, |
9677 | slp_node, NULL); | |
ebfd146a IR |
9678 | gcc_assert (done); |
9679 | break; | |
9680 | ||
42fd8198 | 9681 | case comparison_vec_info_type: |
86a91c0a RS |
9682 | done = vectorizable_comparison (stmt_info, gsi, &vec_stmt, NULL, |
9683 | slp_node, NULL); | |
42fd8198 IE |
9684 | gcc_assert (done); |
9685 | break; | |
9686 | ||
ebfd146a | 9687 | case call_vec_info_type: |
86a91c0a | 9688 | done = vectorizable_call (stmt_info, gsi, &vec_stmt, slp_node, NULL); |
039d9ea1 | 9689 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
9690 | break; |
9691 | ||
0136f8f0 | 9692 | case call_simd_clone_vec_info_type: |
86a91c0a RS |
9693 | done = vectorizable_simd_clone_call (stmt_info, gsi, &vec_stmt, |
9694 | slp_node, NULL); | |
0136f8f0 AH |
9695 | stmt = gsi_stmt (*gsi); |
9696 | break; | |
9697 | ||
ebfd146a | 9698 | case reduc_vec_info_type: |
86a91c0a | 9699 | done = vectorizable_reduction (stmt_info, gsi, &vec_stmt, slp_node, |
68435eb2 | 9700 | slp_node_instance, NULL); |
ebfd146a IR |
9701 | gcc_assert (done); |
9702 | break; | |
9703 | ||
9704 | default: | |
9705 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
9706 | { | |
73fbfcad | 9707 | if (dump_enabled_p ()) |
78c60e3d | 9708 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9709 | "stmt not supported.\n"); |
ebfd146a IR |
9710 | gcc_unreachable (); |
9711 | } | |
9712 | } | |
9713 | ||
225ce44b RB |
9714 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
9715 | This would break hybrid SLP vectorization. */ | |
9716 | if (slp_node) | |
d90f8440 | 9717 | gcc_assert (!vec_stmt |
1eede195 | 9718 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt_info); |
225ce44b | 9719 | |
ebfd146a IR |
9720 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
9721 | is being vectorized, but outside the immediately enclosing loop. */ | |
9722 | if (vec_stmt | |
e57d9a82 | 9723 | && nested_p |
ebfd146a IR |
9724 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
9725 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 9726 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 9727 | vect_used_in_outer_by_reduction)) |
ebfd146a | 9728 | { |
a70d6342 IR |
9729 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
9730 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
9731 | imm_use_iterator imm_iter; |
9732 | use_operand_p use_p; | |
9733 | tree scalar_dest; | |
ebfd146a | 9734 | |
73fbfcad | 9735 | if (dump_enabled_p ()) |
78c60e3d | 9736 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 9737 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
9738 | |
9739 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
9740 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
9741 | STMT). */ | |
9742 | if (gimple_code (stmt) == GIMPLE_PHI) | |
9743 | scalar_dest = PHI_RESULT (stmt); | |
9744 | else | |
4beb6642 | 9745 | scalar_dest = gimple_get_lhs (stmt); |
ebfd146a IR |
9746 | |
9747 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
6585ff8f RS |
9748 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) |
9749 | { | |
9750 | stmt_vec_info exit_phi_info | |
9751 | = vinfo->lookup_stmt (USE_STMT (use_p)); | |
9752 | STMT_VINFO_VEC_STMT (exit_phi_info) = vec_stmt; | |
9753 | } | |
ebfd146a IR |
9754 | } |
9755 | ||
9756 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
9757 | being vectorized. */ | |
68a0f2ff | 9758 | if (STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) |
ebfd146a | 9759 | { |
86a91c0a RS |
9760 | done = can_vectorize_live_stmts (stmt_info, gsi, slp_node, &vec_stmt, |
9761 | NULL); | |
ebfd146a IR |
9762 | gcc_assert (done); |
9763 | } | |
9764 | ||
9765 | if (vec_stmt) | |
83197f37 | 9766 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 9767 | |
b8698a0f | 9768 | return is_store; |
ebfd146a IR |
9769 | } |
9770 | ||
9771 | ||
b8698a0f | 9772 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
9773 | stmt_vec_info. */ |
9774 | ||
9775 | void | |
32e8e429 | 9776 | vect_remove_stores (stmt_vec_info first_stmt_info) |
ebfd146a | 9777 | { |
b5b56c2a | 9778 | vec_info *vinfo = first_stmt_info->vinfo; |
32e8e429 | 9779 | stmt_vec_info next_stmt_info = first_stmt_info; |
ebfd146a | 9780 | |
a1824cfd | 9781 | while (next_stmt_info) |
ebfd146a | 9782 | { |
a1824cfd | 9783 | stmt_vec_info tmp = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
211cd1e2 | 9784 | next_stmt_info = vect_orig_stmt (next_stmt_info); |
ebfd146a | 9785 | /* Free the attached stmt_vec_info and remove the stmt. */ |
b5b56c2a | 9786 | vinfo->remove_stmt (next_stmt_info); |
a1824cfd | 9787 | next_stmt_info = tmp; |
ebfd146a IR |
9788 | } |
9789 | } | |
9790 | ||
bb67d9c7 | 9791 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 9792 | |
bb67d9c7 | 9793 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
9794 | by the target. */ |
9795 | ||
c803b2a9 | 9796 | tree |
86e36728 | 9797 | get_vectype_for_scalar_type_and_size (tree scalar_type, poly_uint64 size) |
ebfd146a | 9798 | { |
c7d97b28 | 9799 | tree orig_scalar_type = scalar_type; |
3bd8f481 | 9800 | scalar_mode inner_mode; |
ef4bddc2 | 9801 | machine_mode simd_mode; |
86e36728 | 9802 | poly_uint64 nunits; |
ebfd146a IR |
9803 | tree vectype; |
9804 | ||
3bd8f481 RS |
9805 | if (!is_int_mode (TYPE_MODE (scalar_type), &inner_mode) |
9806 | && !is_float_mode (TYPE_MODE (scalar_type), &inner_mode)) | |
ebfd146a IR |
9807 | return NULL_TREE; |
9808 | ||
3bd8f481 | 9809 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
48f2e373 | 9810 | |
7b7b1813 RG |
9811 | /* For vector types of elements whose mode precision doesn't |
9812 | match their types precision we use a element type of mode | |
9813 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
9814 | they support the proper result truncation/extension. |
9815 | We also make sure to build vector types with INTEGER_TYPE | |
9816 | component type only. */ | |
6d7971b8 | 9817 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
9818 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
9819 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
9820 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
9821 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 9822 | |
ccbf5bb4 RG |
9823 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
9824 | When the component mode passes the above test simply use a type | |
9825 | corresponding to that mode. The theory is that any use that | |
9826 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 9827 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 9828 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
9829 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
9830 | ||
9831 | /* We can't build a vector type of elements with alignment bigger than | |
9832 | their size. */ | |
dfc2e2ac | 9833 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
9834 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
9835 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 9836 | |
dfc2e2ac RB |
9837 | /* If we felt back to using the mode fail if there was |
9838 | no scalar type for it. */ | |
9839 | if (scalar_type == NULL_TREE) | |
9840 | return NULL_TREE; | |
9841 | ||
bb67d9c7 RG |
9842 | /* If no size was supplied use the mode the target prefers. Otherwise |
9843 | lookup a vector mode of the specified size. */ | |
86e36728 | 9844 | if (known_eq (size, 0U)) |
bb67d9c7 | 9845 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); |
86e36728 RS |
9846 | else if (!multiple_p (size, nbytes, &nunits) |
9847 | || !mode_for_vector (inner_mode, nunits).exists (&simd_mode)) | |
9da15d40 | 9848 | return NULL_TREE; |
4c8fd8ac | 9849 | /* NOTE: nunits == 1 is allowed to support single element vector types. */ |
86e36728 | 9850 | if (!multiple_p (GET_MODE_SIZE (simd_mode), nbytes, &nunits)) |
cc4b5170 | 9851 | return NULL_TREE; |
ebfd146a IR |
9852 | |
9853 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
9854 | |
9855 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
9856 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 9857 | return NULL_TREE; |
ebfd146a | 9858 | |
c7d97b28 RB |
9859 | /* Re-attach the address-space qualifier if we canonicalized the scalar |
9860 | type. */ | |
9861 | if (TYPE_ADDR_SPACE (orig_scalar_type) != TYPE_ADDR_SPACE (vectype)) | |
9862 | return build_qualified_type | |
9863 | (vectype, KEEP_QUAL_ADDR_SPACE (TYPE_QUALS (orig_scalar_type))); | |
9864 | ||
ebfd146a IR |
9865 | return vectype; |
9866 | } | |
9867 | ||
86e36728 | 9868 | poly_uint64 current_vector_size; |
bb67d9c7 RG |
9869 | |
9870 | /* Function get_vectype_for_scalar_type. | |
9871 | ||
9872 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
9873 | by the target. */ | |
9874 | ||
9875 | tree | |
9876 | get_vectype_for_scalar_type (tree scalar_type) | |
9877 | { | |
9878 | tree vectype; | |
9879 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
9880 | current_vector_size); | |
9881 | if (vectype | |
86e36728 | 9882 | && known_eq (current_vector_size, 0U)) |
bb67d9c7 RG |
9883 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); |
9884 | return vectype; | |
9885 | } | |
9886 | ||
42fd8198 IE |
9887 | /* Function get_mask_type_for_scalar_type. |
9888 | ||
9889 | Returns the mask type corresponding to a result of comparison | |
9890 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
9891 | ||
9892 | tree | |
9893 | get_mask_type_for_scalar_type (tree scalar_type) | |
9894 | { | |
9895 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
9896 | ||
9897 | if (!vectype) | |
9898 | return NULL; | |
9899 | ||
9900 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
9901 | current_vector_size); | |
9902 | } | |
9903 | ||
b690cc0f RG |
9904 | /* Function get_same_sized_vectype |
9905 | ||
9906 | Returns a vector type corresponding to SCALAR_TYPE of size | |
9907 | VECTOR_TYPE if supported by the target. */ | |
9908 | ||
9909 | tree | |
bb67d9c7 | 9910 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 9911 | { |
2568d8a1 | 9912 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type)) |
9f47c7e5 IE |
9913 | return build_same_sized_truth_vector_type (vector_type); |
9914 | ||
bb67d9c7 RG |
9915 | return get_vectype_for_scalar_type_and_size |
9916 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
9917 | } |
9918 | ||
ebfd146a IR |
9919 | /* Function vect_is_simple_use. |
9920 | ||
9921 | Input: | |
81c40241 RB |
9922 | VINFO - the vect info of the loop or basic block that is being vectorized. |
9923 | OPERAND - operand in the loop or bb. | |
9924 | Output: | |
fef96d8e RS |
9925 | DEF_STMT_INFO_OUT (optional) - information about the defining stmt in |
9926 | case OPERAND is an SSA_NAME that is defined in the vectorizable region | |
9927 | DEF_STMT_OUT (optional) - the defining stmt in case OPERAND is an SSA_NAME; | |
9928 | the definition could be anywhere in the function | |
81c40241 | 9929 | DT - the type of definition |
ebfd146a IR |
9930 | |
9931 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 9932 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 9933 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 9934 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
9935 | is the case in reduction/induction computations). |
9936 | For basic blocks, supportable operands are constants and bb invariants. | |
9937 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
9938 | |
9939 | bool | |
894dd753 | 9940 | vect_is_simple_use (tree operand, vec_info *vinfo, enum vect_def_type *dt, |
fef96d8e | 9941 | stmt_vec_info *def_stmt_info_out, gimple **def_stmt_out) |
b8698a0f | 9942 | { |
fef96d8e RS |
9943 | if (def_stmt_info_out) |
9944 | *def_stmt_info_out = NULL; | |
894dd753 RS |
9945 | if (def_stmt_out) |
9946 | *def_stmt_out = NULL; | |
3fc356dc | 9947 | *dt = vect_unknown_def_type; |
b8698a0f | 9948 | |
73fbfcad | 9949 | if (dump_enabled_p ()) |
ebfd146a | 9950 | { |
78c60e3d SS |
9951 | dump_printf_loc (MSG_NOTE, vect_location, |
9952 | "vect_is_simple_use: operand "); | |
30f502ed RB |
9953 | if (TREE_CODE (operand) == SSA_NAME |
9954 | && !SSA_NAME_IS_DEFAULT_DEF (operand)) | |
9955 | dump_gimple_expr (MSG_NOTE, TDF_SLIM, SSA_NAME_DEF_STMT (operand), 0); | |
9956 | else | |
9957 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
ebfd146a | 9958 | } |
b8698a0f | 9959 | |
b758f602 | 9960 | if (CONSTANT_CLASS_P (operand)) |
30f502ed RB |
9961 | *dt = vect_constant_def; |
9962 | else if (is_gimple_min_invariant (operand)) | |
9963 | *dt = vect_external_def; | |
9964 | else if (TREE_CODE (operand) != SSA_NAME) | |
9965 | *dt = vect_unknown_def_type; | |
9966 | else if (SSA_NAME_IS_DEFAULT_DEF (operand)) | |
8644a673 | 9967 | *dt = vect_external_def; |
ebfd146a IR |
9968 | else |
9969 | { | |
30f502ed | 9970 | gimple *def_stmt = SSA_NAME_DEF_STMT (operand); |
c98d0595 RS |
9971 | stmt_vec_info stmt_vinfo = vinfo->lookup_def (operand); |
9972 | if (!stmt_vinfo) | |
30f502ed RB |
9973 | *dt = vect_external_def; |
9974 | else | |
0f8c840c | 9975 | { |
6e6b18e5 RS |
9976 | stmt_vinfo = vect_stmt_to_vectorize (stmt_vinfo); |
9977 | def_stmt = stmt_vinfo->stmt; | |
30f502ed RB |
9978 | switch (gimple_code (def_stmt)) |
9979 | { | |
9980 | case GIMPLE_PHI: | |
9981 | case GIMPLE_ASSIGN: | |
9982 | case GIMPLE_CALL: | |
9983 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
9984 | break; | |
9985 | default: | |
9986 | *dt = vect_unknown_def_type; | |
9987 | break; | |
9988 | } | |
fef96d8e RS |
9989 | if (def_stmt_info_out) |
9990 | *def_stmt_info_out = stmt_vinfo; | |
0f8c840c | 9991 | } |
30f502ed RB |
9992 | if (def_stmt_out) |
9993 | *def_stmt_out = def_stmt; | |
ebfd146a IR |
9994 | } |
9995 | ||
2e8ab70c RB |
9996 | if (dump_enabled_p ()) |
9997 | { | |
30f502ed | 9998 | dump_printf (MSG_NOTE, ", type of def: "); |
2e8ab70c RB |
9999 | switch (*dt) |
10000 | { | |
10001 | case vect_uninitialized_def: | |
10002 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
10003 | break; | |
10004 | case vect_constant_def: | |
10005 | dump_printf (MSG_NOTE, "constant\n"); | |
10006 | break; | |
10007 | case vect_external_def: | |
10008 | dump_printf (MSG_NOTE, "external\n"); | |
10009 | break; | |
10010 | case vect_internal_def: | |
10011 | dump_printf (MSG_NOTE, "internal\n"); | |
10012 | break; | |
10013 | case vect_induction_def: | |
10014 | dump_printf (MSG_NOTE, "induction\n"); | |
10015 | break; | |
10016 | case vect_reduction_def: | |
10017 | dump_printf (MSG_NOTE, "reduction\n"); | |
10018 | break; | |
10019 | case vect_double_reduction_def: | |
10020 | dump_printf (MSG_NOTE, "double reduction\n"); | |
10021 | break; | |
10022 | case vect_nested_cycle: | |
10023 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
10024 | break; | |
10025 | case vect_unknown_def_type: | |
10026 | dump_printf (MSG_NOTE, "unknown\n"); | |
10027 | break; | |
10028 | } | |
10029 | } | |
10030 | ||
81c40241 | 10031 | if (*dt == vect_unknown_def_type) |
ebfd146a | 10032 | { |
73fbfcad | 10033 | if (dump_enabled_p ()) |
78c60e3d | 10034 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 10035 | "Unsupported pattern.\n"); |
ebfd146a IR |
10036 | return false; |
10037 | } | |
10038 | ||
ebfd146a IR |
10039 | return true; |
10040 | } | |
10041 | ||
81c40241 | 10042 | /* Function vect_is_simple_use. |
b690cc0f | 10043 | |
81c40241 | 10044 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
10045 | type of OPERAND and stores it to *VECTYPE. If the definition of |
10046 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
10047 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
10048 | is responsible to compute the best suited vector type for the | |
10049 | scalar operand. */ | |
10050 | ||
10051 | bool | |
894dd753 | 10052 | vect_is_simple_use (tree operand, vec_info *vinfo, enum vect_def_type *dt, |
fef96d8e RS |
10053 | tree *vectype, stmt_vec_info *def_stmt_info_out, |
10054 | gimple **def_stmt_out) | |
b690cc0f | 10055 | { |
fef96d8e | 10056 | stmt_vec_info def_stmt_info; |
894dd753 | 10057 | gimple *def_stmt; |
fef96d8e | 10058 | if (!vect_is_simple_use (operand, vinfo, dt, &def_stmt_info, &def_stmt)) |
b690cc0f RG |
10059 | return false; |
10060 | ||
894dd753 RS |
10061 | if (def_stmt_out) |
10062 | *def_stmt_out = def_stmt; | |
fef96d8e RS |
10063 | if (def_stmt_info_out) |
10064 | *def_stmt_info_out = def_stmt_info; | |
894dd753 | 10065 | |
b690cc0f RG |
10066 | /* Now get a vector type if the def is internal, otherwise supply |
10067 | NULL_TREE and leave it up to the caller to figure out a proper | |
10068 | type for the use stmt. */ | |
10069 | if (*dt == vect_internal_def | |
10070 | || *dt == vect_induction_def | |
10071 | || *dt == vect_reduction_def | |
10072 | || *dt == vect_double_reduction_def | |
10073 | || *dt == vect_nested_cycle) | |
10074 | { | |
fef96d8e | 10075 | *vectype = STMT_VINFO_VECTYPE (def_stmt_info); |
b690cc0f | 10076 | gcc_assert (*vectype != NULL_TREE); |
30f502ed | 10077 | if (dump_enabled_p ()) |
3c2a8ed0 DM |
10078 | dump_printf_loc (MSG_NOTE, vect_location, |
10079 | "vect_is_simple_use: vectype %T\n", *vectype); | |
b690cc0f RG |
10080 | } |
10081 | else if (*dt == vect_uninitialized_def | |
10082 | || *dt == vect_constant_def | |
10083 | || *dt == vect_external_def) | |
10084 | *vectype = NULL_TREE; | |
10085 | else | |
10086 | gcc_unreachable (); | |
10087 | ||
10088 | return true; | |
10089 | } | |
10090 | ||
ebfd146a IR |
10091 | |
10092 | /* Function supportable_widening_operation | |
10093 | ||
b8698a0f L |
10094 | Check whether an operation represented by the code CODE is a |
10095 | widening operation that is supported by the target platform in | |
b690cc0f RG |
10096 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
10097 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 10098 | |
1bda738b JJ |
10099 | Widening operations we currently support are NOP (CONVERT), FLOAT, |
10100 | FIX_TRUNC and WIDEN_MULT. This function checks if these operations | |
10101 | are supported by the target platform either directly (via vector | |
10102 | tree-codes), or via target builtins. | |
ebfd146a IR |
10103 | |
10104 | Output: | |
b8698a0f L |
10105 | - CODE1 and CODE2 are codes of vector operations to be used when |
10106 | vectorizing the operation, if available. | |
ebfd146a IR |
10107 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
10108 | case of multi-step conversion (like char->short->int - in that case | |
10109 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
10110 | - INTERM_TYPES contains the intermediate type required to perform the |
10111 | widening operation (short in the above example). */ | |
ebfd146a IR |
10112 | |
10113 | bool | |
32e8e429 | 10114 | supportable_widening_operation (enum tree_code code, stmt_vec_info stmt_info, |
b690cc0f | 10115 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
10116 | enum tree_code *code1, enum tree_code *code2, |
10117 | int *multi_step_cvt, | |
9771b263 | 10118 | vec<tree> *interm_types) |
ebfd146a | 10119 | { |
ebfd146a | 10120 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); |
4ef69dfc | 10121 | struct loop *vect_loop = NULL; |
ef4bddc2 | 10122 | machine_mode vec_mode; |
81f40b79 | 10123 | enum insn_code icode1, icode2; |
ebfd146a | 10124 | optab optab1, optab2; |
b690cc0f RG |
10125 | tree vectype = vectype_in; |
10126 | tree wide_vectype = vectype_out; | |
ebfd146a | 10127 | enum tree_code c1, c2; |
4a00c761 JJ |
10128 | int i; |
10129 | tree prev_type, intermediate_type; | |
ef4bddc2 | 10130 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 10131 | optab optab3, optab4; |
ebfd146a | 10132 | |
4a00c761 | 10133 | *multi_step_cvt = 0; |
4ef69dfc IR |
10134 | if (loop_info) |
10135 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
10136 | ||
ebfd146a IR |
10137 | switch (code) |
10138 | { | |
10139 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
10140 | /* The result of a vectorized widening operation usually requires |
10141 | two vectors (because the widened results do not fit into one vector). | |
10142 | The generated vector results would normally be expected to be | |
10143 | generated in the same order as in the original scalar computation, | |
10144 | i.e. if 8 results are generated in each vector iteration, they are | |
10145 | to be organized as follows: | |
10146 | vect1: [res1,res2,res3,res4], | |
10147 | vect2: [res5,res6,res7,res8]. | |
10148 | ||
10149 | However, in the special case that the result of the widening | |
10150 | operation is used in a reduction computation only, the order doesn't | |
10151 | matter (because when vectorizing a reduction we change the order of | |
10152 | the computation). Some targets can take advantage of this and | |
10153 | generate more efficient code. For example, targets like Altivec, | |
10154 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
10155 | generate the following vectors: | |
10156 | vect1: [res1,res3,res5,res7], | |
10157 | vect2: [res2,res4,res6,res8]. | |
10158 | ||
10159 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
10160 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
10161 | iterations in parallel). We therefore don't allow to change the | |
10162 | order of the computation in the inner-loop during outer-loop | |
10163 | vectorization. */ | |
10164 | /* TODO: Another case in which order doesn't *really* matter is when we | |
10165 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
10166 | Normally, pack_trunc performs an even/odd permute, whereas the | |
10167 | repack from an even/odd expansion would be an interleave, which | |
10168 | would be significantly simpler for e.g. AVX2. */ | |
10169 | /* In any case, in order to avoid duplicating the code below, recurse | |
10170 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
10171 | are properly set up for the caller. If we fail, we'll continue with | |
10172 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
10173 | if (vect_loop | |
10174 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
86a91c0a | 10175 | && !nested_in_vect_loop_p (vect_loop, stmt_info) |
6ae6116f | 10176 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, |
86a91c0a RS |
10177 | stmt_info, vectype_out, |
10178 | vectype_in, code1, code2, | |
10179 | multi_step_cvt, interm_types)) | |
ebc047a2 CH |
10180 | { |
10181 | /* Elements in a vector with vect_used_by_reduction property cannot | |
10182 | be reordered if the use chain with this property does not have the | |
10183 | same operation. One such an example is s += a * b, where elements | |
10184 | in a and b cannot be reordered. Here we check if the vector defined | |
10185 | by STMT is only directly used in the reduction statement. */ | |
86a91c0a | 10186 | tree lhs = gimple_assign_lhs (stmt_info->stmt); |
0d0a4e20 RS |
10187 | stmt_vec_info use_stmt_info = loop_info->lookup_single_use (lhs); |
10188 | if (use_stmt_info | |
10189 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
10190 | return true; | |
ebc047a2 | 10191 | } |
4a00c761 JJ |
10192 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
10193 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
10194 | break; |
10195 | ||
81c40241 RB |
10196 | case DOT_PROD_EXPR: |
10197 | c1 = DOT_PROD_EXPR; | |
10198 | c2 = DOT_PROD_EXPR; | |
10199 | break; | |
10200 | ||
10201 | case SAD_EXPR: | |
10202 | c1 = SAD_EXPR; | |
10203 | c2 = SAD_EXPR; | |
10204 | break; | |
10205 | ||
6ae6116f RH |
10206 | case VEC_WIDEN_MULT_EVEN_EXPR: |
10207 | /* Support the recursion induced just above. */ | |
10208 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
10209 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
10210 | break; | |
10211 | ||
36ba4aae | 10212 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
10213 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
10214 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
10215 | break; |
10216 | ||
ebfd146a | 10217 | CASE_CONVERT: |
4a00c761 JJ |
10218 | c1 = VEC_UNPACK_LO_EXPR; |
10219 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
10220 | break; |
10221 | ||
10222 | case FLOAT_EXPR: | |
4a00c761 JJ |
10223 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
10224 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
10225 | break; |
10226 | ||
10227 | case FIX_TRUNC_EXPR: | |
1bda738b JJ |
10228 | c1 = VEC_UNPACK_FIX_TRUNC_LO_EXPR; |
10229 | c2 = VEC_UNPACK_FIX_TRUNC_HI_EXPR; | |
10230 | break; | |
ebfd146a IR |
10231 | |
10232 | default: | |
10233 | gcc_unreachable (); | |
10234 | } | |
10235 | ||
6ae6116f | 10236 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 10237 | std::swap (c1, c2); |
4a00c761 | 10238 | |
ebfd146a IR |
10239 | if (code == FIX_TRUNC_EXPR) |
10240 | { | |
10241 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
10242 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
10243 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
10244 | } |
10245 | else | |
10246 | { | |
10247 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
10248 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
10249 | } | |
10250 | ||
10251 | if (!optab1 || !optab2) | |
10252 | return false; | |
10253 | ||
10254 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
10255 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
10256 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
10257 | return false; |
10258 | ||
4a00c761 JJ |
10259 | *code1 = c1; |
10260 | *code2 = c2; | |
10261 | ||
10262 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
10263 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
10264 | /* For scalar masks we may have different boolean |
10265 | vector types having the same QImode. Thus we | |
10266 | add additional check for elements number. */ | |
10267 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
928686b1 RS |
10268 | || known_eq (TYPE_VECTOR_SUBPARTS (vectype), |
10269 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)); | |
4a00c761 | 10270 | |
b8698a0f | 10271 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 10272 | types. */ |
ebfd146a | 10273 | |
4a00c761 JJ |
10274 | prev_type = vectype; |
10275 | prev_mode = vec_mode; | |
b8698a0f | 10276 | |
4a00c761 JJ |
10277 | if (!CONVERT_EXPR_CODE_P (code)) |
10278 | return false; | |
b8698a0f | 10279 | |
4a00c761 JJ |
10280 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
10281 | intermediate steps in promotion sequence. We try | |
10282 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
10283 | not. */ | |
9771b263 | 10284 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
10285 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
10286 | { | |
10287 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
10288 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
10289 | { | |
7cfb4d93 | 10290 | intermediate_type = vect_halve_mask_nunits (prev_type); |
3ae0661a IE |
10291 | if (intermediate_mode != TYPE_MODE (intermediate_type)) |
10292 | return false; | |
10293 | } | |
10294 | else | |
10295 | intermediate_type | |
10296 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
10297 | TYPE_UNSIGNED (prev_type)); | |
10298 | ||
4a00c761 JJ |
10299 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); |
10300 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
10301 | ||
10302 | if (!optab3 || !optab4 | |
10303 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
10304 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
10305 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
10306 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
10307 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
10308 | == CODE_FOR_nothing) | |
10309 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
10310 | == CODE_FOR_nothing)) | |
10311 | break; | |
ebfd146a | 10312 | |
9771b263 | 10313 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
10314 | (*multi_step_cvt)++; |
10315 | ||
10316 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
10317 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff | 10318 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
928686b1 RS |
10319 | || known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type), |
10320 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)); | |
4a00c761 JJ |
10321 | |
10322 | prev_type = intermediate_type; | |
10323 | prev_mode = intermediate_mode; | |
ebfd146a IR |
10324 | } |
10325 | ||
9771b263 | 10326 | interm_types->release (); |
4a00c761 | 10327 | return false; |
ebfd146a IR |
10328 | } |
10329 | ||
10330 | ||
10331 | /* Function supportable_narrowing_operation | |
10332 | ||
b8698a0f L |
10333 | Check whether an operation represented by the code CODE is a |
10334 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
10335 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
10336 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 10337 | |
1bda738b JJ |
10338 | Narrowing operations we currently support are NOP (CONVERT), FIX_TRUNC |
10339 | and FLOAT. This function checks if these operations are supported by | |
ebfd146a IR |
10340 | the target platform directly via vector tree-codes. |
10341 | ||
10342 | Output: | |
b8698a0f L |
10343 | - CODE1 is the code of a vector operation to be used when |
10344 | vectorizing the operation, if available. | |
ebfd146a IR |
10345 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
10346 | case of multi-step conversion (like int->short->char - in that case | |
10347 | MULTI_STEP_CVT will be 1). | |
10348 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 10349 | narrowing operation (short in the above example). */ |
ebfd146a IR |
10350 | |
10351 | bool | |
10352 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 10353 | tree vectype_out, tree vectype_in, |
ebfd146a | 10354 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 10355 | vec<tree> *interm_types) |
ebfd146a | 10356 | { |
ef4bddc2 | 10357 | machine_mode vec_mode; |
ebfd146a IR |
10358 | enum insn_code icode1; |
10359 | optab optab1, interm_optab; | |
b690cc0f RG |
10360 | tree vectype = vectype_in; |
10361 | tree narrow_vectype = vectype_out; | |
ebfd146a | 10362 | enum tree_code c1; |
3ae0661a | 10363 | tree intermediate_type, prev_type; |
ef4bddc2 | 10364 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 10365 | int i; |
4a00c761 | 10366 | bool uns; |
ebfd146a | 10367 | |
4a00c761 | 10368 | *multi_step_cvt = 0; |
ebfd146a IR |
10369 | switch (code) |
10370 | { | |
10371 | CASE_CONVERT: | |
10372 | c1 = VEC_PACK_TRUNC_EXPR; | |
10373 | break; | |
10374 | ||
10375 | case FIX_TRUNC_EXPR: | |
10376 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
10377 | break; | |
10378 | ||
10379 | case FLOAT_EXPR: | |
1bda738b JJ |
10380 | c1 = VEC_PACK_FLOAT_EXPR; |
10381 | break; | |
ebfd146a IR |
10382 | |
10383 | default: | |
10384 | gcc_unreachable (); | |
10385 | } | |
10386 | ||
10387 | if (code == FIX_TRUNC_EXPR) | |
10388 | /* The signedness is determined from output operand. */ | |
b690cc0f | 10389 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
10390 | else |
10391 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
10392 | ||
10393 | if (!optab1) | |
10394 | return false; | |
10395 | ||
10396 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 10397 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
10398 | return false; |
10399 | ||
4a00c761 JJ |
10400 | *code1 = c1; |
10401 | ||
10402 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
10403 | /* For scalar masks we may have different boolean |
10404 | vector types having the same QImode. Thus we | |
10405 | add additional check for elements number. */ | |
10406 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
928686b1 RS |
10407 | || known_eq (TYPE_VECTOR_SUBPARTS (vectype) * 2, |
10408 | TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 | 10409 | |
1bda738b JJ |
10410 | if (code == FLOAT_EXPR) |
10411 | return false; | |
10412 | ||
ebfd146a IR |
10413 | /* Check if it's a multi-step conversion that can be done using intermediate |
10414 | types. */ | |
4a00c761 | 10415 | prev_mode = vec_mode; |
3ae0661a | 10416 | prev_type = vectype; |
4a00c761 JJ |
10417 | if (code == FIX_TRUNC_EXPR) |
10418 | uns = TYPE_UNSIGNED (vectype_out); | |
10419 | else | |
10420 | uns = TYPE_UNSIGNED (vectype); | |
10421 | ||
10422 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
10423 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
10424 | costly than signed. */ | |
10425 | if (code == FIX_TRUNC_EXPR && uns) | |
10426 | { | |
10427 | enum insn_code icode2; | |
10428 | ||
10429 | intermediate_type | |
10430 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
10431 | interm_optab | |
10432 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 10433 | if (interm_optab != unknown_optab |
4a00c761 JJ |
10434 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
10435 | && insn_data[icode1].operand[0].mode | |
10436 | == insn_data[icode2].operand[0].mode) | |
10437 | { | |
10438 | uns = false; | |
10439 | optab1 = interm_optab; | |
10440 | icode1 = icode2; | |
10441 | } | |
10442 | } | |
ebfd146a | 10443 | |
4a00c761 JJ |
10444 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
10445 | intermediate steps in promotion sequence. We try | |
10446 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 10447 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
10448 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
10449 | { | |
10450 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
10451 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
10452 | { | |
7cfb4d93 | 10453 | intermediate_type = vect_double_mask_nunits (prev_type); |
3ae0661a | 10454 | if (intermediate_mode != TYPE_MODE (intermediate_type)) |
7cfb4d93 | 10455 | return false; |
3ae0661a IE |
10456 | } |
10457 | else | |
10458 | intermediate_type | |
10459 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
4a00c761 JJ |
10460 | interm_optab |
10461 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
10462 | optab_default); | |
10463 | if (!interm_optab | |
10464 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
10465 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
10466 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
10467 | == CODE_FOR_nothing)) | |
10468 | break; | |
10469 | ||
9771b263 | 10470 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
10471 | (*multi_step_cvt)++; |
10472 | ||
10473 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff | 10474 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
928686b1 RS |
10475 | || known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type) * 2, |
10476 | TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 JJ |
10477 | |
10478 | prev_mode = intermediate_mode; | |
3ae0661a | 10479 | prev_type = intermediate_type; |
4a00c761 | 10480 | optab1 = interm_optab; |
ebfd146a IR |
10481 | } |
10482 | ||
9771b263 | 10483 | interm_types->release (); |
4a00c761 | 10484 | return false; |
ebfd146a | 10485 | } |
7cfb4d93 RS |
10486 | |
10487 | /* Generate and return a statement that sets vector mask MASK such that | |
10488 | MASK[I] is true iff J + START_INDEX < END_INDEX for all J <= I. */ | |
10489 | ||
10490 | gcall * | |
10491 | vect_gen_while (tree mask, tree start_index, tree end_index) | |
10492 | { | |
10493 | tree cmp_type = TREE_TYPE (start_index); | |
10494 | tree mask_type = TREE_TYPE (mask); | |
10495 | gcc_checking_assert (direct_internal_fn_supported_p (IFN_WHILE_ULT, | |
10496 | cmp_type, mask_type, | |
10497 | OPTIMIZE_FOR_SPEED)); | |
10498 | gcall *call = gimple_build_call_internal (IFN_WHILE_ULT, 3, | |
10499 | start_index, end_index, | |
10500 | build_zero_cst (mask_type)); | |
10501 | gimple_call_set_lhs (call, mask); | |
10502 | return call; | |
10503 | } | |
535e7c11 RS |
10504 | |
10505 | /* Generate a vector mask of type MASK_TYPE for which index I is false iff | |
10506 | J + START_INDEX < END_INDEX for all J <= I. Add the statements to SEQ. */ | |
10507 | ||
10508 | tree | |
10509 | vect_gen_while_not (gimple_seq *seq, tree mask_type, tree start_index, | |
10510 | tree end_index) | |
10511 | { | |
10512 | tree tmp = make_ssa_name (mask_type); | |
10513 | gcall *call = vect_gen_while (tmp, start_index, end_index); | |
10514 | gimple_seq_add_stmt (seq, call); | |
10515 | return gimple_build (seq, BIT_NOT_EXPR, mask_type, tmp); | |
10516 | } | |
1f3cb663 RS |
10517 | |
10518 | /* Try to compute the vector types required to vectorize STMT_INFO, | |
10519 | returning true on success and false if vectorization isn't possible. | |
10520 | ||
10521 | On success: | |
10522 | ||
10523 | - Set *STMT_VECTYPE_OUT to: | |
10524 | - NULL_TREE if the statement doesn't need to be vectorized; | |
10525 | - boolean_type_node if the statement is a boolean operation whose | |
10526 | vector type can only be determined once all the other vector types | |
10527 | are known; and | |
10528 | - the equivalent of STMT_VINFO_VECTYPE otherwise. | |
10529 | ||
10530 | - Set *NUNITS_VECTYPE_OUT to the vector type that contains the maximum | |
10531 | number of units needed to vectorize STMT_INFO, or NULL_TREE if the | |
10532 | statement does not help to determine the overall number of units. */ | |
10533 | ||
f4ebbd24 | 10534 | opt_result |
1f3cb663 RS |
10535 | vect_get_vector_types_for_stmt (stmt_vec_info stmt_info, |
10536 | tree *stmt_vectype_out, | |
10537 | tree *nunits_vectype_out) | |
10538 | { | |
10539 | gimple *stmt = stmt_info->stmt; | |
10540 | ||
10541 | *stmt_vectype_out = NULL_TREE; | |
10542 | *nunits_vectype_out = NULL_TREE; | |
10543 | ||
10544 | if (gimple_get_lhs (stmt) == NULL_TREE | |
10545 | /* MASK_STORE has no lhs, but is ok. */ | |
10546 | && !gimple_call_internal_p (stmt, IFN_MASK_STORE)) | |
10547 | { | |
10548 | if (is_a <gcall *> (stmt)) | |
10549 | { | |
10550 | /* Ignore calls with no lhs. These must be calls to | |
10551 | #pragma omp simd functions, and what vectorization factor | |
10552 | it really needs can't be determined until | |
10553 | vectorizable_simd_clone_call. */ | |
10554 | if (dump_enabled_p ()) | |
10555 | dump_printf_loc (MSG_NOTE, vect_location, | |
10556 | "defer to SIMD clone analysis.\n"); | |
f4ebbd24 | 10557 | return opt_result::success (); |
1f3cb663 RS |
10558 | } |
10559 | ||
f4ebbd24 DM |
10560 | return opt_result::failure_at (stmt, |
10561 | "not vectorized: irregular stmt.%G", stmt); | |
1f3cb663 RS |
10562 | } |
10563 | ||
10564 | if (VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))) | |
f4ebbd24 DM |
10565 | return opt_result::failure_at (stmt, |
10566 | "not vectorized: vector stmt in loop:%G", | |
10567 | stmt); | |
1f3cb663 RS |
10568 | |
10569 | tree vectype; | |
10570 | tree scalar_type = NULL_TREE; | |
10571 | if (STMT_VINFO_VECTYPE (stmt_info)) | |
10572 | *stmt_vectype_out = vectype = STMT_VINFO_VECTYPE (stmt_info); | |
10573 | else | |
10574 | { | |
10575 | gcc_assert (!STMT_VINFO_DATA_REF (stmt_info)); | |
10576 | if (gimple_call_internal_p (stmt, IFN_MASK_STORE)) | |
10577 | scalar_type = TREE_TYPE (gimple_call_arg (stmt, 3)); | |
10578 | else | |
10579 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); | |
10580 | ||
10581 | /* Pure bool ops don't participate in number-of-units computation. | |
10582 | For comparisons use the types being compared. */ | |
10583 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type) | |
10584 | && is_gimple_assign (stmt) | |
10585 | && gimple_assign_rhs_code (stmt) != COND_EXPR) | |
10586 | { | |
10587 | *stmt_vectype_out = boolean_type_node; | |
10588 | ||
10589 | tree rhs1 = gimple_assign_rhs1 (stmt); | |
10590 | if (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_comparison | |
10591 | && !VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (rhs1))) | |
10592 | scalar_type = TREE_TYPE (rhs1); | |
10593 | else | |
10594 | { | |
10595 | if (dump_enabled_p ()) | |
10596 | dump_printf_loc (MSG_NOTE, vect_location, | |
10597 | "pure bool operation.\n"); | |
f4ebbd24 | 10598 | return opt_result::success (); |
1f3cb663 RS |
10599 | } |
10600 | } | |
10601 | ||
10602 | if (dump_enabled_p ()) | |
3c2a8ed0 DM |
10603 | dump_printf_loc (MSG_NOTE, vect_location, |
10604 | "get vectype for scalar type: %T\n", scalar_type); | |
1f3cb663 RS |
10605 | vectype = get_vectype_for_scalar_type (scalar_type); |
10606 | if (!vectype) | |
f4ebbd24 DM |
10607 | return opt_result::failure_at (stmt, |
10608 | "not vectorized:" | |
10609 | " unsupported data-type %T\n", | |
10610 | scalar_type); | |
1f3cb663 RS |
10611 | |
10612 | if (!*stmt_vectype_out) | |
10613 | *stmt_vectype_out = vectype; | |
10614 | ||
10615 | if (dump_enabled_p ()) | |
3c2a8ed0 | 10616 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: %T\n", vectype); |
1f3cb663 RS |
10617 | } |
10618 | ||
10619 | /* Don't try to compute scalar types if the stmt produces a boolean | |
10620 | vector; use the existing vector type instead. */ | |
10621 | tree nunits_vectype; | |
10622 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) | |
10623 | nunits_vectype = vectype; | |
10624 | else | |
10625 | { | |
10626 | /* The number of units is set according to the smallest scalar | |
10627 | type (or the largest vector size, but we only support one | |
10628 | vector size per vectorization). */ | |
10629 | if (*stmt_vectype_out != boolean_type_node) | |
10630 | { | |
10631 | HOST_WIDE_INT dummy; | |
86a91c0a RS |
10632 | scalar_type = vect_get_smallest_scalar_type (stmt_info, |
10633 | &dummy, &dummy); | |
1f3cb663 RS |
10634 | } |
10635 | if (dump_enabled_p ()) | |
3c2a8ed0 DM |
10636 | dump_printf_loc (MSG_NOTE, vect_location, |
10637 | "get vectype for scalar type: %T\n", scalar_type); | |
1f3cb663 RS |
10638 | nunits_vectype = get_vectype_for_scalar_type (scalar_type); |
10639 | } | |
10640 | if (!nunits_vectype) | |
f4ebbd24 DM |
10641 | return opt_result::failure_at (stmt, |
10642 | "not vectorized: unsupported data-type %T\n", | |
10643 | scalar_type); | |
1f3cb663 RS |
10644 | |
10645 | if (maybe_ne (GET_MODE_SIZE (TYPE_MODE (vectype)), | |
10646 | GET_MODE_SIZE (TYPE_MODE (nunits_vectype)))) | |
f4ebbd24 DM |
10647 | return opt_result::failure_at (stmt, |
10648 | "not vectorized: different sized vector " | |
10649 | "types in statement, %T and %T\n", | |
10650 | vectype, nunits_vectype); | |
1f3cb663 RS |
10651 | |
10652 | if (dump_enabled_p ()) | |
10653 | { | |
3c2a8ed0 DM |
10654 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: %T\n", |
10655 | nunits_vectype); | |
1f3cb663 RS |
10656 | |
10657 | dump_printf_loc (MSG_NOTE, vect_location, "nunits = "); | |
10658 | dump_dec (MSG_NOTE, TYPE_VECTOR_SUBPARTS (nunits_vectype)); | |
10659 | dump_printf (MSG_NOTE, "\n"); | |
10660 | } | |
10661 | ||
10662 | *nunits_vectype_out = nunits_vectype; | |
f4ebbd24 | 10663 | return opt_result::success (); |
1f3cb663 RS |
10664 | } |
10665 | ||
10666 | /* Try to determine the correct vector type for STMT_INFO, which is a | |
10667 | statement that produces a scalar boolean result. Return the vector | |
10668 | type on success, otherwise return NULL_TREE. */ | |
10669 | ||
f4ebbd24 | 10670 | opt_tree |
1f3cb663 RS |
10671 | vect_get_mask_type_for_stmt (stmt_vec_info stmt_info) |
10672 | { | |
10673 | gimple *stmt = stmt_info->stmt; | |
10674 | tree mask_type = NULL; | |
10675 | tree vectype, scalar_type; | |
10676 | ||
10677 | if (is_gimple_assign (stmt) | |
10678 | && TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_comparison | |
10679 | && !VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (gimple_assign_rhs1 (stmt)))) | |
10680 | { | |
10681 | scalar_type = TREE_TYPE (gimple_assign_rhs1 (stmt)); | |
10682 | mask_type = get_mask_type_for_scalar_type (scalar_type); | |
10683 | ||
10684 | if (!mask_type) | |
f4ebbd24 DM |
10685 | return opt_tree::failure_at (stmt, |
10686 | "not vectorized: unsupported mask\n"); | |
1f3cb663 RS |
10687 | } |
10688 | else | |
10689 | { | |
10690 | tree rhs; | |
10691 | ssa_op_iter iter; | |
1f3cb663 RS |
10692 | enum vect_def_type dt; |
10693 | ||
10694 | FOR_EACH_SSA_TREE_OPERAND (rhs, stmt, iter, SSA_OP_USE) | |
10695 | { | |
894dd753 | 10696 | if (!vect_is_simple_use (rhs, stmt_info->vinfo, &dt, &vectype)) |
f4ebbd24 DM |
10697 | return opt_tree::failure_at (stmt, |
10698 | "not vectorized:can't compute mask" | |
10699 | " type for statement, %G", stmt); | |
1f3cb663 RS |
10700 | |
10701 | /* No vectype probably means external definition. | |
10702 | Allow it in case there is another operand which | |
10703 | allows to determine mask type. */ | |
10704 | if (!vectype) | |
10705 | continue; | |
10706 | ||
10707 | if (!mask_type) | |
10708 | mask_type = vectype; | |
10709 | else if (maybe_ne (TYPE_VECTOR_SUBPARTS (mask_type), | |
10710 | TYPE_VECTOR_SUBPARTS (vectype))) | |
f4ebbd24 DM |
10711 | return opt_tree::failure_at (stmt, |
10712 | "not vectorized: different sized mask" | |
10713 | " types in statement, %T and %T\n", | |
10714 | mask_type, vectype); | |
1f3cb663 RS |
10715 | else if (VECTOR_BOOLEAN_TYPE_P (mask_type) |
10716 | != VECTOR_BOOLEAN_TYPE_P (vectype)) | |
f4ebbd24 DM |
10717 | return opt_tree::failure_at (stmt, |
10718 | "not vectorized: mixed mask and " | |
10719 | "nonmask vector types in statement, " | |
10720 | "%T and %T\n", | |
10721 | mask_type, vectype); | |
1f3cb663 RS |
10722 | } |
10723 | ||
10724 | /* We may compare boolean value loaded as vector of integers. | |
10725 | Fix mask_type in such case. */ | |
10726 | if (mask_type | |
10727 | && !VECTOR_BOOLEAN_TYPE_P (mask_type) | |
10728 | && gimple_code (stmt) == GIMPLE_ASSIGN | |
10729 | && TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_comparison) | |
10730 | mask_type = build_same_sized_truth_vector_type (mask_type); | |
10731 | } | |
10732 | ||
10733 | /* No mask_type should mean loop invariant predicate. | |
10734 | This is probably a subject for optimization in if-conversion. */ | |
f4ebbd24 DM |
10735 | if (!mask_type) |
10736 | return opt_tree::failure_at (stmt, | |
10737 | "not vectorized: can't compute mask type " | |
10738 | "for statement: %G", stmt); | |
10739 | ||
10740 | return opt_tree::success (mask_type); | |
1f3cb663 | 10741 | } |