]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* SLP - Basic Block Vectorization |
f2c74cc4 | 2 | Copyright (C) 2007, 2008, 2009, 2010, 2011, 2012 |
c75c517d | 3 | Free Software Foundation, Inc. |
b8698a0f | 4 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
5 | and Ira Rosen <irar@il.ibm.com> |
6 | ||
7 | This file is part of GCC. | |
8 | ||
9 | GCC is free software; you can redistribute it and/or modify it under | |
10 | the terms of the GNU General Public License as published by the Free | |
11 | Software Foundation; either version 3, or (at your option) any later | |
12 | version. | |
13 | ||
14 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
15 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
16 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
17 | for more details. | |
18 | ||
19 | You should have received a copy of the GNU General Public License | |
20 | along with GCC; see the file COPYING3. If not see | |
21 | <http://www.gnu.org/licenses/>. */ | |
22 | ||
23 | #include "config.h" | |
24 | #include "system.h" | |
25 | #include "coretypes.h" | |
26 | #include "tm.h" | |
27 | #include "ggc.h" | |
28 | #include "tree.h" | |
29 | #include "target.h" | |
30 | #include "basic-block.h" | |
cf835838 JM |
31 | #include "tree-pretty-print.h" |
32 | #include "gimple-pretty-print.h" | |
ebfd146a IR |
33 | #include "tree-flow.h" |
34 | #include "tree-dump.h" | |
35 | #include "cfgloop.h" | |
36 | #include "cfglayout.h" | |
37 | #include "expr.h" | |
38 | #include "recog.h" | |
39 | #include "optabs.h" | |
40 | #include "tree-vectorizer.h" | |
2635892a | 41 | #include "langhooks.h" |
ebfd146a | 42 | |
a70d6342 IR |
43 | /* Extract the location of the basic block in the source code. |
44 | Return the basic block location if succeed and NULL if not. */ | |
45 | ||
46 | LOC | |
47 | find_bb_location (basic_block bb) | |
48 | { | |
49 | gimple stmt = NULL; | |
50 | gimple_stmt_iterator si; | |
51 | ||
52 | if (!bb) | |
53 | return UNKNOWN_LOC; | |
54 | ||
55 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
56 | { | |
57 | stmt = gsi_stmt (si); | |
58 | if (gimple_location (stmt) != UNKNOWN_LOC) | |
59 | return gimple_location (stmt); | |
60 | } | |
61 | ||
62 | return UNKNOWN_LOC; | |
63 | } | |
64 | ||
65 | ||
ebfd146a IR |
66 | /* Recursively free the memory allocated for the SLP tree rooted at NODE. */ |
67 | ||
68 | static void | |
69 | vect_free_slp_tree (slp_tree node) | |
70 | { | |
d092494c IR |
71 | int i; |
72 | slp_void_p child; | |
73 | ||
ebfd146a IR |
74 | if (!node) |
75 | return; | |
76 | ||
d092494c | 77 | FOR_EACH_VEC_ELT (slp_void_p, SLP_TREE_CHILDREN (node), i, child) |
d3cfd39e | 78 | vect_free_slp_tree ((slp_tree) child); |
b8698a0f | 79 | |
d3cfd39e | 80 | VEC_free (slp_void_p, heap, SLP_TREE_CHILDREN (node)); |
ebfd146a | 81 | VEC_free (gimple, heap, SLP_TREE_SCALAR_STMTS (node)); |
b8698a0f | 82 | |
ebfd146a IR |
83 | if (SLP_TREE_VEC_STMTS (node)) |
84 | VEC_free (gimple, heap, SLP_TREE_VEC_STMTS (node)); | |
85 | ||
86 | free (node); | |
87 | } | |
88 | ||
89 | ||
90 | /* Free the memory allocated for the SLP instance. */ | |
91 | ||
92 | void | |
93 | vect_free_slp_instance (slp_instance instance) | |
94 | { | |
95 | vect_free_slp_tree (SLP_INSTANCE_TREE (instance)); | |
96 | VEC_free (int, heap, SLP_INSTANCE_LOAD_PERMUTATION (instance)); | |
97 | VEC_free (slp_tree, heap, SLP_INSTANCE_LOADS (instance)); | |
98 | } | |
99 | ||
100 | ||
d092494c IR |
101 | /* Create an SLP node for SCALAR_STMTS. */ |
102 | ||
103 | static slp_tree | |
104 | vect_create_new_slp_node (VEC (gimple, heap) *scalar_stmts) | |
105 | { | |
d3cfd39e | 106 | slp_tree node; |
d092494c IR |
107 | gimple stmt = VEC_index (gimple, scalar_stmts, 0); |
108 | unsigned int nops; | |
109 | ||
110 | if (is_gimple_call (stmt)) | |
111 | nops = gimple_call_num_args (stmt); | |
112 | else if (is_gimple_assign (stmt)) | |
f7e531cf IR |
113 | { |
114 | nops = gimple_num_ops (stmt) - 1; | |
115 | if (gimple_assign_rhs_code (stmt) == COND_EXPR) | |
116 | nops++; | |
117 | } | |
d092494c IR |
118 | else |
119 | return NULL; | |
120 | ||
d3cfd39e | 121 | node = XNEW (struct _slp_tree); |
d092494c IR |
122 | SLP_TREE_SCALAR_STMTS (node) = scalar_stmts; |
123 | SLP_TREE_VEC_STMTS (node) = NULL; | |
124 | SLP_TREE_CHILDREN (node) = VEC_alloc (slp_void_p, heap, nops); | |
125 | SLP_TREE_OUTSIDE_OF_LOOP_COST (node) = 0; | |
126 | SLP_TREE_INSIDE_OF_LOOP_COST (node) = 0; | |
127 | ||
128 | return node; | |
129 | } | |
130 | ||
131 | ||
132 | /* Allocate operands info for NOPS operands, and GROUP_SIZE def-stmts for each | |
133 | operand. */ | |
134 | static VEC (slp_oprnd_info, heap) * | |
135 | vect_create_oprnd_info (int nops, int group_size) | |
136 | { | |
137 | int i; | |
138 | slp_oprnd_info oprnd_info; | |
139 | VEC (slp_oprnd_info, heap) *oprnds_info; | |
140 | ||
141 | oprnds_info = VEC_alloc (slp_oprnd_info, heap, nops); | |
142 | for (i = 0; i < nops; i++) | |
143 | { | |
144 | oprnd_info = XNEW (struct _slp_oprnd_info); | |
145 | oprnd_info->def_stmts = VEC_alloc (gimple, heap, group_size); | |
146 | oprnd_info->first_dt = vect_uninitialized_def; | |
147 | oprnd_info->first_def_type = NULL_TREE; | |
148 | oprnd_info->first_const_oprnd = NULL_TREE; | |
149 | oprnd_info->first_pattern = false; | |
150 | VEC_quick_push (slp_oprnd_info, oprnds_info, oprnd_info); | |
151 | } | |
152 | ||
153 | return oprnds_info; | |
154 | } | |
155 | ||
156 | ||
d3cfd39e JJ |
157 | /* Free operands info. */ |
158 | ||
d092494c | 159 | static void |
d3cfd39e | 160 | vect_free_oprnd_info (VEC (slp_oprnd_info, heap) **oprnds_info) |
d092494c IR |
161 | { |
162 | int i; | |
163 | slp_oprnd_info oprnd_info; | |
164 | ||
d3cfd39e JJ |
165 | FOR_EACH_VEC_ELT (slp_oprnd_info, *oprnds_info, i, oprnd_info) |
166 | { | |
d092494c | 167 | VEC_free (gimple, heap, oprnd_info->def_stmts); |
d3cfd39e JJ |
168 | XDELETE (oprnd_info); |
169 | } | |
d092494c IR |
170 | |
171 | VEC_free (slp_oprnd_info, heap, *oprnds_info); | |
172 | } | |
173 | ||
174 | ||
175 | /* Get the defs for the rhs of STMT (collect them in OPRNDS_INFO), check that | |
176 | they are of a valid type and that they match the defs of the first stmt of | |
177 | the SLP group (stored in OPRNDS_INFO). */ | |
ebfd146a IR |
178 | |
179 | static bool | |
a70d6342 | 180 | vect_get_and_check_slp_defs (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo, |
b8698a0f | 181 | slp_tree slp_node, gimple stmt, |
d092494c IR |
182 | int ncopies_for_cost, bool first, |
183 | VEC (slp_oprnd_info, heap) **oprnds_info) | |
ebfd146a IR |
184 | { |
185 | tree oprnd; | |
186 | unsigned int i, number_of_oprnds; | |
d092494c | 187 | tree def, def_op0 = NULL_TREE; |
ebfd146a | 188 | gimple def_stmt; |
d092494c IR |
189 | enum vect_def_type dt = vect_uninitialized_def; |
190 | enum vect_def_type dt_op0 = vect_uninitialized_def; | |
191 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
192 | tree lhs = gimple_get_lhs (stmt); | |
a70d6342 | 193 | struct loop *loop = NULL; |
6aa904c4 IR |
194 | enum tree_code rhs_code; |
195 | bool different_types = false; | |
d092494c IR |
196 | bool pattern = false; |
197 | slp_oprnd_info oprnd_info, oprnd0_info, oprnd1_info; | |
f7e531cf IR |
198 | int op_idx = 1; |
199 | tree compare_rhs = NULL_TREE; | |
b8698a0f | 200 | |
a70d6342 IR |
201 | if (loop_vinfo) |
202 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a | 203 | |
d092494c | 204 | if (is_gimple_call (stmt)) |
190c2236 JJ |
205 | { |
206 | number_of_oprnds = gimple_call_num_args (stmt); | |
207 | op_idx = 3; | |
208 | } | |
f7e531cf IR |
209 | else if (is_gimple_assign (stmt)) |
210 | { | |
211 | number_of_oprnds = gimple_num_ops (stmt) - 1; | |
212 | if (gimple_assign_rhs_code (stmt) == COND_EXPR) | |
213 | number_of_oprnds++; | |
214 | } | |
d092494c | 215 | else |
f7e531cf | 216 | return false; |
ebfd146a IR |
217 | |
218 | for (i = 0; i < number_of_oprnds; i++) | |
219 | { | |
f7e531cf IR |
220 | if (compare_rhs) |
221 | { | |
222 | oprnd = compare_rhs; | |
223 | compare_rhs = NULL_TREE; | |
224 | } | |
225 | else | |
226 | oprnd = gimple_op (stmt, op_idx++); | |
227 | ||
d092494c | 228 | oprnd_info = VEC_index (slp_oprnd_info, *oprnds_info, i); |
ebfd146a | 229 | |
f7e531cf IR |
230 | if (COMPARISON_CLASS_P (oprnd)) |
231 | { | |
232 | compare_rhs = TREE_OPERAND (oprnd, 1); | |
233 | oprnd = TREE_OPERAND (oprnd, 0); | |
234 | } | |
235 | ||
24ee1384 IR |
236 | if (!vect_is_simple_use (oprnd, NULL, loop_vinfo, bb_vinfo, &def_stmt, |
237 | &def, &dt) | |
d092494c | 238 | || (!def_stmt && dt != vect_constant_def)) |
ebfd146a | 239 | { |
b8698a0f | 240 | if (vect_print_dump_info (REPORT_SLP)) |
ebfd146a IR |
241 | { |
242 | fprintf (vect_dump, "Build SLP failed: can't find def for "); | |
243 | print_generic_expr (vect_dump, oprnd, TDF_SLIM); | |
244 | } | |
245 | ||
246 | return false; | |
247 | } | |
248 | ||
a70d6342 | 249 | /* Check if DEF_STMT is a part of a pattern in LOOP and get the def stmt |
ff802fa1 | 250 | from the pattern. Check that all the stmts of the node are in the |
ebfd146a | 251 | pattern. */ |
f5709183 IR |
252 | if (def_stmt && gimple_bb (def_stmt) |
253 | && ((loop && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
254 | || (!loop && gimple_bb (def_stmt) == BB_VINFO_BB (bb_vinfo) | |
255 | && gimple_code (def_stmt) != GIMPLE_PHI)) | |
ebfd146a | 256 | && vinfo_for_stmt (def_stmt) |
83197f37 | 257 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (def_stmt)) |
f5709183 IR |
258 | && !STMT_VINFO_RELEVANT (vinfo_for_stmt (def_stmt)) |
259 | && !STMT_VINFO_LIVE_P (vinfo_for_stmt (def_stmt))) | |
ebfd146a | 260 | { |
d092494c IR |
261 | pattern = true; |
262 | if (!first && !oprnd_info->first_pattern) | |
263 | { | |
264 | if (vect_print_dump_info (REPORT_DETAILS)) | |
265 | { | |
266 | fprintf (vect_dump, "Build SLP failed: some of the stmts" | |
267 | " are in a pattern, and others are not "); | |
268 | print_generic_expr (vect_dump, oprnd, TDF_SLIM); | |
269 | } | |
ebfd146a | 270 | |
d092494c | 271 | return false; |
ebfd146a IR |
272 | } |
273 | ||
274 | def_stmt = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (def_stmt)); | |
d092494c | 275 | dt = STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt)); |
ebfd146a | 276 | |
f7e531cf | 277 | if (dt == vect_unknown_def_type) |
ebfd146a IR |
278 | { |
279 | if (vect_print_dump_info (REPORT_DETAILS)) | |
280 | fprintf (vect_dump, "Unsupported pattern."); | |
281 | return false; | |
282 | } | |
283 | ||
284 | switch (gimple_code (def_stmt)) | |
285 | { | |
286 | case GIMPLE_PHI: | |
d092494c | 287 | def = gimple_phi_result (def_stmt); |
ebfd146a IR |
288 | break; |
289 | ||
290 | case GIMPLE_ASSIGN: | |
d092494c | 291 | def = gimple_assign_lhs (def_stmt); |
ebfd146a IR |
292 | break; |
293 | ||
294 | default: | |
295 | if (vect_print_dump_info (REPORT_DETAILS)) | |
296 | fprintf (vect_dump, "unsupported defining stmt: "); | |
297 | return false; | |
298 | } | |
299 | } | |
300 | ||
d092494c | 301 | if (first) |
ebfd146a | 302 | { |
d092494c IR |
303 | oprnd_info->first_dt = dt; |
304 | oprnd_info->first_pattern = pattern; | |
305 | if (def) | |
306 | { | |
307 | oprnd_info->first_def_type = TREE_TYPE (def); | |
308 | oprnd_info->first_const_oprnd = NULL_TREE; | |
309 | } | |
ebfd146a | 310 | else |
d092494c IR |
311 | { |
312 | oprnd_info->first_def_type = NULL_TREE; | |
313 | oprnd_info->first_const_oprnd = oprnd; | |
314 | } | |
ebfd146a | 315 | |
d092494c IR |
316 | if (i == 0) |
317 | { | |
318 | def_op0 = def; | |
319 | dt_op0 = dt; | |
320 | /* Analyze costs (for the first stmt of the group only). */ | |
321 | if (REFERENCE_CLASS_P (lhs)) | |
322 | /* Store. */ | |
323 | vect_model_store_cost (stmt_info, ncopies_for_cost, false, | |
324 | dt, slp_node); | |
325 | else | |
7daac345 JJ |
326 | { |
327 | enum vect_def_type dts[2]; | |
328 | dts[0] = dt; | |
329 | dts[1] = vect_uninitialized_def; | |
330 | /* Not memory operation (we don't call this function for | |
331 | loads). */ | |
332 | vect_model_simple_cost (stmt_info, ncopies_for_cost, dts, | |
333 | slp_node); | |
334 | } | |
d092494c | 335 | } |
ebfd146a | 336 | } |
ebfd146a IR |
337 | else |
338 | { | |
d092494c IR |
339 | /* Not first stmt of the group, check that the def-stmt/s match |
340 | the def-stmt/s of the first stmt. Allow different definition | |
341 | types for reduction chains: the first stmt must be a | |
342 | vect_reduction_def (a phi node), and the rest | |
343 | vect_internal_def. */ | |
344 | if (((oprnd_info->first_dt != dt | |
345 | && !(oprnd_info->first_dt == vect_reduction_def | |
346 | && dt == vect_internal_def)) | |
347 | || (oprnd_info->first_def_type != NULL_TREE | |
348 | && def | |
349 | && !types_compatible_p (oprnd_info->first_def_type, | |
350 | TREE_TYPE (def)))) | |
351 | || (!def | |
352 | && !types_compatible_p (TREE_TYPE (oprnd_info->first_const_oprnd), | |
353 | TREE_TYPE (oprnd))) | |
354 | || different_types) | |
ebfd146a | 355 | { |
d092494c | 356 | if (number_of_oprnds != 2) |
ebfd146a | 357 | { |
d092494c IR |
358 | if (vect_print_dump_info (REPORT_SLP)) |
359 | fprintf (vect_dump, "Build SLP failed: different types "); | |
360 | ||
361 | return false; | |
362 | } | |
363 | ||
364 | /* Try to swap operands in case of binary operation. */ | |
365 | if (i == 0) | |
366 | different_types = true; | |
367 | else | |
b8698a0f | 368 | { |
d092494c IR |
369 | oprnd0_info = VEC_index (slp_oprnd_info, *oprnds_info, 0); |
370 | if (is_gimple_assign (stmt) | |
371 | && (rhs_code = gimple_assign_rhs_code (stmt)) | |
372 | && TREE_CODE_CLASS (rhs_code) == tcc_binary | |
373 | && commutative_tree_code (rhs_code) | |
374 | && oprnd0_info->first_dt == dt | |
375 | && oprnd_info->first_dt == dt_op0 | |
376 | && def_op0 && def | |
377 | && !(oprnd0_info->first_def_type | |
378 | && !types_compatible_p (oprnd0_info->first_def_type, | |
379 | TREE_TYPE (def))) | |
380 | && !(oprnd_info->first_def_type | |
381 | && !types_compatible_p (oprnd_info->first_def_type, | |
382 | TREE_TYPE (def_op0)))) | |
383 | { | |
384 | if (vect_print_dump_info (REPORT_SLP)) | |
385 | { | |
386 | fprintf (vect_dump, "Swapping operands of "); | |
387 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
6aa904c4 | 388 | } |
b8698a0f | 389 | |
d092494c IR |
390 | swap_tree_operands (stmt, gimple_assign_rhs1_ptr (stmt), |
391 | gimple_assign_rhs2_ptr (stmt)); | |
392 | } | |
393 | else | |
394 | { | |
395 | if (vect_print_dump_info (REPORT_SLP)) | |
396 | fprintf (vect_dump, "Build SLP failed: different types "); | |
397 | ||
398 | return false; | |
6aa904c4 | 399 | } |
ebfd146a IR |
400 | } |
401 | } | |
402 | } | |
403 | ||
404 | /* Check the types of the definitions. */ | |
d092494c | 405 | switch (dt) |
ebfd146a IR |
406 | { |
407 | case vect_constant_def: | |
8644a673 | 408 | case vect_external_def: |
d092494c | 409 | case vect_reduction_def: |
ebfd146a | 410 | break; |
b8698a0f | 411 | |
8644a673 | 412 | case vect_internal_def: |
d092494c IR |
413 | if (different_types) |
414 | { | |
415 | oprnd0_info = VEC_index (slp_oprnd_info, *oprnds_info, 0); | |
416 | oprnd1_info = VEC_index (slp_oprnd_info, *oprnds_info, 0); | |
417 | if (i == 0) | |
418 | VEC_quick_push (gimple, oprnd1_info->def_stmts, def_stmt); | |
419 | else | |
420 | VEC_quick_push (gimple, oprnd0_info->def_stmts, def_stmt); | |
421 | } | |
ebfd146a | 422 | else |
d092494c IR |
423 | VEC_quick_push (gimple, oprnd_info->def_stmts, def_stmt); |
424 | ||
ebfd146a IR |
425 | break; |
426 | ||
427 | default: | |
428 | /* FORNOW: Not supported. */ | |
b8698a0f | 429 | if (vect_print_dump_info (REPORT_SLP)) |
ebfd146a IR |
430 | { |
431 | fprintf (vect_dump, "Build SLP failed: illegal type of def "); | |
d092494c | 432 | print_generic_expr (vect_dump, def, TDF_SLIM); |
ebfd146a IR |
433 | } |
434 | ||
435 | return false; | |
436 | } | |
437 | } | |
438 | ||
439 | return true; | |
440 | } | |
441 | ||
442 | ||
443 | /* Recursively build an SLP tree starting from NODE. | |
b8698a0f | 444 | Fail (and return FALSE) if def-stmts are not isomorphic, require data |
ff802fa1 | 445 | permutation or are of unsupported types of operation. Otherwise, return |
ebfd146a IR |
446 | TRUE. */ |
447 | ||
448 | static bool | |
b8698a0f | 449 | vect_build_slp_tree (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo, |
a70d6342 IR |
450 | slp_tree *node, unsigned int group_size, |
451 | int *inside_cost, int *outside_cost, | |
452 | int ncopies_for_cost, unsigned int *max_nunits, | |
ebfd146a | 453 | VEC (int, heap) **load_permutation, |
a70d6342 | 454 | VEC (slp_tree, heap) **loads, |
6aa904c4 | 455 | unsigned int vectorization_factor, bool *loads_permuted) |
ebfd146a | 456 | { |
ebfd146a IR |
457 | unsigned int i; |
458 | VEC (gimple, heap) *stmts = SLP_TREE_SCALAR_STMTS (*node); | |
459 | gimple stmt = VEC_index (gimple, stmts, 0); | |
2200fc49 | 460 | enum tree_code first_stmt_code = ERROR_MARK, rhs_code = ERROR_MARK; |
f7e531cf | 461 | enum tree_code first_cond_code = ERROR_MARK; |
ebfd146a IR |
462 | tree lhs; |
463 | bool stop_recursion = false, need_same_oprnds = false; | |
464 | tree vectype, scalar_type, first_op1 = NULL_TREE; | |
a70d6342 | 465 | unsigned int ncopies; |
ebfd146a IR |
466 | optab optab; |
467 | int icode; | |
468 | enum machine_mode optab_op2_mode; | |
469 | enum machine_mode vec_mode; | |
ebfd146a | 470 | struct data_reference *first_dr; |
ebfd146a IR |
471 | HOST_WIDE_INT dummy; |
472 | bool permutation = false; | |
473 | unsigned int load_place; | |
b5aeb3bb | 474 | gimple first_load, prev_first_load = NULL; |
d092494c IR |
475 | VEC (slp_oprnd_info, heap) *oprnds_info; |
476 | unsigned int nops; | |
477 | slp_oprnd_info oprnd_info; | |
f7e531cf | 478 | tree cond; |
d092494c IR |
479 | |
480 | if (is_gimple_call (stmt)) | |
481 | nops = gimple_call_num_args (stmt); | |
f7e531cf IR |
482 | else if (is_gimple_assign (stmt)) |
483 | { | |
484 | nops = gimple_num_ops (stmt) - 1; | |
485 | if (gimple_assign_rhs_code (stmt) == COND_EXPR) | |
486 | nops++; | |
487 | } | |
d092494c | 488 | else |
f7e531cf | 489 | return false; |
d092494c IR |
490 | |
491 | oprnds_info = vect_create_oprnd_info (nops, group_size); | |
ebfd146a IR |
492 | |
493 | /* For every stmt in NODE find its def stmt/s. */ | |
ac47786e | 494 | FOR_EACH_VEC_ELT (gimple, stmts, i, stmt) |
ebfd146a | 495 | { |
b8698a0f | 496 | if (vect_print_dump_info (REPORT_SLP)) |
ebfd146a IR |
497 | { |
498 | fprintf (vect_dump, "Build SLP for "); | |
499 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
500 | } | |
501 | ||
4b5caab7 IR |
502 | /* Fail to vectorize statements marked as unvectorizable. */ |
503 | if (!STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (stmt))) | |
504 | { | |
505 | if (vect_print_dump_info (REPORT_SLP)) | |
506 | { | |
507 | fprintf (vect_dump, | |
508 | "Build SLP failed: unvectorizable statement "); | |
509 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
510 | } | |
511 | ||
d3cfd39e | 512 | vect_free_oprnd_info (&oprnds_info); |
4b5caab7 IR |
513 | return false; |
514 | } | |
515 | ||
ebfd146a IR |
516 | lhs = gimple_get_lhs (stmt); |
517 | if (lhs == NULL_TREE) | |
518 | { | |
b8698a0f | 519 | if (vect_print_dump_info (REPORT_SLP)) |
ebfd146a IR |
520 | { |
521 | fprintf (vect_dump, | |
d092494c | 522 | "Build SLP failed: not GIMPLE_ASSIGN nor GIMPLE_CALL "); |
ebfd146a IR |
523 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); |
524 | } | |
b8698a0f | 525 | |
d3cfd39e | 526 | vect_free_oprnd_info (&oprnds_info); |
ebfd146a IR |
527 | return false; |
528 | } | |
529 | ||
f7e531cf IR |
530 | if (is_gimple_assign (stmt) |
531 | && gimple_assign_rhs_code (stmt) == COND_EXPR | |
532 | && (cond = gimple_assign_rhs1 (stmt)) | |
533 | && !COMPARISON_CLASS_P (cond)) | |
534 | { | |
535 | if (vect_print_dump_info (REPORT_SLP)) | |
536 | { | |
537 | fprintf (vect_dump, | |
538 | "Build SLP failed: condition is not comparison "); | |
539 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
540 | } | |
541 | ||
d3cfd39e | 542 | vect_free_oprnd_info (&oprnds_info); |
f7e531cf IR |
543 | return false; |
544 | } | |
545 | ||
b8698a0f | 546 | scalar_type = vect_get_smallest_scalar_type (stmt, &dummy, &dummy); |
ebfd146a IR |
547 | vectype = get_vectype_for_scalar_type (scalar_type); |
548 | if (!vectype) | |
549 | { | |
550 | if (vect_print_dump_info (REPORT_SLP)) | |
551 | { | |
552 | fprintf (vect_dump, "Build SLP failed: unsupported data-type "); | |
553 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
554 | } | |
d092494c | 555 | |
d3cfd39e | 556 | vect_free_oprnd_info (&oprnds_info); |
ebfd146a IR |
557 | return false; |
558 | } | |
b8698a0f | 559 | |
4ef69dfc IR |
560 | /* In case of multiple types we need to detect the smallest type. */ |
561 | if (*max_nunits < TYPE_VECTOR_SUBPARTS (vectype)) | |
a70d6342 | 562 | { |
4ef69dfc IR |
563 | *max_nunits = TYPE_VECTOR_SUBPARTS (vectype); |
564 | if (bb_vinfo) | |
565 | vectorization_factor = *max_nunits; | |
a70d6342 | 566 | } |
b8698a0f | 567 | |
4ef69dfc | 568 | ncopies = vectorization_factor / TYPE_VECTOR_SUBPARTS (vectype); |
b8698a0f | 569 | |
ebfd146a | 570 | if (is_gimple_call (stmt)) |
190c2236 JJ |
571 | { |
572 | rhs_code = CALL_EXPR; | |
573 | if (gimple_call_internal_p (stmt) | |
574 | || gimple_call_tail_p (stmt) | |
575 | || gimple_call_noreturn_p (stmt) | |
576 | || !gimple_call_nothrow_p (stmt) | |
577 | || gimple_call_chain (stmt)) | |
578 | { | |
579 | if (vect_print_dump_info (REPORT_SLP)) | |
580 | { | |
581 | fprintf (vect_dump, | |
582 | "Build SLP failed: unsupported call type "); | |
583 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
584 | } | |
585 | ||
d3cfd39e | 586 | vect_free_oprnd_info (&oprnds_info); |
190c2236 JJ |
587 | return false; |
588 | } | |
589 | } | |
ebfd146a IR |
590 | else |
591 | rhs_code = gimple_assign_rhs_code (stmt); | |
592 | ||
593 | /* Check the operation. */ | |
594 | if (i == 0) | |
595 | { | |
596 | first_stmt_code = rhs_code; | |
597 | ||
b8698a0f | 598 | /* Shift arguments should be equal in all the packed stmts for a |
ebfd146a IR |
599 | vector shift with scalar shift operand. */ |
600 | if (rhs_code == LSHIFT_EXPR || rhs_code == RSHIFT_EXPR | |
601 | || rhs_code == LROTATE_EXPR | |
602 | || rhs_code == RROTATE_EXPR) | |
603 | { | |
604 | vec_mode = TYPE_MODE (vectype); | |
605 | ||
606 | /* First see if we have a vector/vector shift. */ | |
607 | optab = optab_for_tree_code (rhs_code, vectype, | |
608 | optab_vector); | |
609 | ||
610 | if (!optab | |
947131ba | 611 | || optab_handler (optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
612 | { |
613 | /* No vector/vector shift, try for a vector/scalar shift. */ | |
614 | optab = optab_for_tree_code (rhs_code, vectype, | |
615 | optab_scalar); | |
616 | ||
617 | if (!optab) | |
618 | { | |
619 | if (vect_print_dump_info (REPORT_SLP)) | |
620 | fprintf (vect_dump, "Build SLP failed: no optab."); | |
d3cfd39e | 621 | vect_free_oprnd_info (&oprnds_info); |
ebfd146a IR |
622 | return false; |
623 | } | |
947131ba | 624 | icode = (int) optab_handler (optab, vec_mode); |
ebfd146a IR |
625 | if (icode == CODE_FOR_nothing) |
626 | { | |
627 | if (vect_print_dump_info (REPORT_SLP)) | |
628 | fprintf (vect_dump, "Build SLP failed: " | |
629 | "op not supported by target."); | |
d3cfd39e | 630 | vect_free_oprnd_info (&oprnds_info); |
ebfd146a IR |
631 | return false; |
632 | } | |
633 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
634 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
635 | { | |
636 | need_same_oprnds = true; | |
637 | first_op1 = gimple_assign_rhs2 (stmt); | |
638 | } | |
639 | } | |
640 | } | |
36ba4aae IR |
641 | else if (rhs_code == WIDEN_LSHIFT_EXPR) |
642 | { | |
643 | need_same_oprnds = true; | |
644 | first_op1 = gimple_assign_rhs2 (stmt); | |
645 | } | |
ebfd146a IR |
646 | } |
647 | else | |
648 | { | |
649 | if (first_stmt_code != rhs_code | |
650 | && (first_stmt_code != IMAGPART_EXPR | |
651 | || rhs_code != REALPART_EXPR) | |
652 | && (first_stmt_code != REALPART_EXPR | |
69f11a13 IR |
653 | || rhs_code != IMAGPART_EXPR) |
654 | && !(STMT_VINFO_STRIDED_ACCESS (vinfo_for_stmt (stmt)) | |
655 | && (first_stmt_code == ARRAY_REF | |
656 | || first_stmt_code == INDIRECT_REF | |
657 | || first_stmt_code == COMPONENT_REF | |
658 | || first_stmt_code == MEM_REF))) | |
ebfd146a | 659 | { |
b8698a0f | 660 | if (vect_print_dump_info (REPORT_SLP)) |
ebfd146a | 661 | { |
b8698a0f | 662 | fprintf (vect_dump, |
ebfd146a IR |
663 | "Build SLP failed: different operation in stmt "); |
664 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
665 | } | |
b8698a0f | 666 | |
d3cfd39e | 667 | vect_free_oprnd_info (&oprnds_info); |
ebfd146a IR |
668 | return false; |
669 | } | |
b8698a0f L |
670 | |
671 | if (need_same_oprnds | |
ebfd146a IR |
672 | && !operand_equal_p (first_op1, gimple_assign_rhs2 (stmt), 0)) |
673 | { | |
b8698a0f | 674 | if (vect_print_dump_info (REPORT_SLP)) |
ebfd146a | 675 | { |
b8698a0f | 676 | fprintf (vect_dump, |
ebfd146a IR |
677 | "Build SLP failed: different shift arguments in "); |
678 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
679 | } | |
b8698a0f | 680 | |
d3cfd39e | 681 | vect_free_oprnd_info (&oprnds_info); |
ebfd146a IR |
682 | return false; |
683 | } | |
190c2236 JJ |
684 | |
685 | if (rhs_code == CALL_EXPR) | |
686 | { | |
687 | gimple first_stmt = VEC_index (gimple, stmts, 0); | |
688 | if (gimple_call_num_args (stmt) != nops | |
689 | || !operand_equal_p (gimple_call_fn (first_stmt), | |
690 | gimple_call_fn (stmt), 0) | |
691 | || gimple_call_fntype (first_stmt) | |
692 | != gimple_call_fntype (stmt)) | |
693 | { | |
694 | if (vect_print_dump_info (REPORT_SLP)) | |
695 | { | |
696 | fprintf (vect_dump, | |
697 | "Build SLP failed: different calls in "); | |
698 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
699 | } | |
700 | ||
d3cfd39e | 701 | vect_free_oprnd_info (&oprnds_info); |
190c2236 JJ |
702 | return false; |
703 | } | |
704 | } | |
ebfd146a IR |
705 | } |
706 | ||
707 | /* Strided store or load. */ | |
708 | if (STMT_VINFO_STRIDED_ACCESS (vinfo_for_stmt (stmt))) | |
709 | { | |
710 | if (REFERENCE_CLASS_P (lhs)) | |
711 | { | |
712 | /* Store. */ | |
b8698a0f | 713 | if (!vect_get_and_check_slp_defs (loop_vinfo, bb_vinfo, *node, |
d092494c IR |
714 | stmt, ncopies_for_cost, |
715 | (i == 0), &oprnds_info)) | |
716 | { | |
d3cfd39e | 717 | vect_free_oprnd_info (&oprnds_info); |
d092494c IR |
718 | return false; |
719 | } | |
ebfd146a | 720 | } |
b5aeb3bb IR |
721 | else |
722 | { | |
723 | /* Load. */ | |
724 | /* FORNOW: Check that there is no gap between the loads. */ | |
e14c1050 IR |
725 | if ((GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) == stmt |
726 | && GROUP_GAP (vinfo_for_stmt (stmt)) != 0) | |
727 | || (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) != stmt | |
728 | && GROUP_GAP (vinfo_for_stmt (stmt)) != 1)) | |
b5aeb3bb IR |
729 | { |
730 | if (vect_print_dump_info (REPORT_SLP)) | |
731 | { | |
732 | fprintf (vect_dump, "Build SLP failed: strided " | |
733 | "loads have gaps "); | |
734 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
735 | } | |
b8698a0f | 736 | |
d3cfd39e | 737 | vect_free_oprnd_info (&oprnds_info); |
b5aeb3bb IR |
738 | return false; |
739 | } | |
2f0fa28e | 740 | |
b5aeb3bb IR |
741 | /* Check that the size of interleaved loads group is not |
742 | greater than the SLP group size. */ | |
6aa904c4 IR |
743 | if (loop_vinfo |
744 | && GROUP_SIZE (vinfo_for_stmt (stmt)) > ncopies * group_size) | |
b5aeb3bb IR |
745 | { |
746 | if (vect_print_dump_info (REPORT_SLP)) | |
747 | { | |
748 | fprintf (vect_dump, "Build SLP failed: the number of " | |
749 | "interleaved loads is greater than" | |
750 | " the SLP group size "); | |
751 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
752 | } | |
b8698a0f | 753 | |
d3cfd39e | 754 | vect_free_oprnd_info (&oprnds_info); |
b5aeb3bb IR |
755 | return false; |
756 | } | |
757 | ||
e14c1050 | 758 | first_load = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)); |
b5aeb3bb IR |
759 | if (prev_first_load) |
760 | { | |
761 | /* Check that there are no loads from different interleaving | |
ff802fa1 | 762 | chains in the same node. The only exception is complex |
b5aeb3bb IR |
763 | numbers. */ |
764 | if (prev_first_load != first_load | |
765 | && rhs_code != REALPART_EXPR | |
766 | && rhs_code != IMAGPART_EXPR) | |
767 | { | |
768 | if (vect_print_dump_info (REPORT_SLP)) | |
769 | { | |
770 | fprintf (vect_dump, "Build SLP failed: different " | |
771 | "interleaving chains in one node "); | |
772 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
773 | } | |
774 | ||
d3cfd39e | 775 | vect_free_oprnd_info (&oprnds_info); |
b5aeb3bb IR |
776 | return false; |
777 | } | |
778 | } | |
779 | else | |
780 | prev_first_load = first_load; | |
b8698a0f | 781 | |
ebfd146a IR |
782 | if (first_load == stmt) |
783 | { | |
784 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt)); | |
720f5239 | 785 | if (vect_supportable_dr_alignment (first_dr, false) |
ebfd146a IR |
786 | == dr_unaligned_unsupported) |
787 | { | |
788 | if (vect_print_dump_info (REPORT_SLP)) | |
789 | { | |
790 | fprintf (vect_dump, "Build SLP failed: unsupported " | |
791 | "unaligned load "); | |
792 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
793 | } | |
b8698a0f | 794 | |
d3cfd39e | 795 | vect_free_oprnd_info (&oprnds_info); |
ebfd146a IR |
796 | return false; |
797 | } | |
b8698a0f | 798 | |
ebfd146a IR |
799 | /* Analyze costs (for the first stmt in the group). */ |
800 | vect_model_load_cost (vinfo_for_stmt (stmt), | |
272c6793 | 801 | ncopies_for_cost, false, *node); |
ebfd146a | 802 | } |
b8698a0f | 803 | |
ff802fa1 | 804 | /* Store the place of this load in the interleaving chain. In |
ebfd146a IR |
805 | case that permutation is needed we later decide if a specific |
806 | permutation is supported. */ | |
807 | load_place = vect_get_place_in_interleaving_chain (stmt, | |
808 | first_load); | |
809 | if (load_place != i) | |
810 | permutation = true; | |
b8698a0f | 811 | |
ebfd146a | 812 | VEC_safe_push (int, heap, *load_permutation, load_place); |
b8698a0f | 813 | |
ebfd146a IR |
814 | /* We stop the tree when we reach a group of loads. */ |
815 | stop_recursion = true; | |
816 | continue; | |
817 | } | |
818 | } /* Strided access. */ | |
819 | else | |
820 | { | |
821 | if (TREE_CODE_CLASS (rhs_code) == tcc_reference) | |
822 | { | |
d092494c | 823 | /* Not strided load. */ |
b8698a0f | 824 | if (vect_print_dump_info (REPORT_SLP)) |
ebfd146a IR |
825 | { |
826 | fprintf (vect_dump, "Build SLP failed: not strided load "); | |
827 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
828 | } | |
829 | ||
830 | /* FORNOW: Not strided loads are not supported. */ | |
d3cfd39e | 831 | vect_free_oprnd_info (&oprnds_info); |
ebfd146a IR |
832 | return false; |
833 | } | |
834 | ||
835 | /* Not memory operation. */ | |
836 | if (TREE_CODE_CLASS (rhs_code) != tcc_binary | |
f7e531cf | 837 | && TREE_CODE_CLASS (rhs_code) != tcc_unary |
190c2236 JJ |
838 | && rhs_code != COND_EXPR |
839 | && rhs_code != CALL_EXPR) | |
ebfd146a | 840 | { |
b8698a0f | 841 | if (vect_print_dump_info (REPORT_SLP)) |
ebfd146a IR |
842 | { |
843 | fprintf (vect_dump, "Build SLP failed: operation"); | |
844 | fprintf (vect_dump, " unsupported "); | |
845 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
846 | } | |
847 | ||
d3cfd39e | 848 | vect_free_oprnd_info (&oprnds_info); |
ebfd146a IR |
849 | return false; |
850 | } | |
851 | ||
f7e531cf IR |
852 | if (rhs_code == COND_EXPR) |
853 | { | |
854 | tree cond_expr = gimple_assign_rhs1 (stmt); | |
855 | ||
856 | if (i == 0) | |
857 | first_cond_code = TREE_CODE (cond_expr); | |
858 | else if (first_cond_code != TREE_CODE (cond_expr)) | |
859 | { | |
860 | if (vect_print_dump_info (REPORT_SLP)) | |
861 | { | |
862 | fprintf (vect_dump, "Build SLP failed: different" | |
863 | " operation"); | |
864 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
865 | } | |
866 | ||
d3cfd39e | 867 | vect_free_oprnd_info (&oprnds_info); |
f7e531cf IR |
868 | return false; |
869 | } | |
870 | } | |
871 | ||
b8698a0f | 872 | /* Find the def-stmts. */ |
a70d6342 | 873 | if (!vect_get_and_check_slp_defs (loop_vinfo, bb_vinfo, *node, stmt, |
d092494c IR |
874 | ncopies_for_cost, (i == 0), |
875 | &oprnds_info)) | |
876 | { | |
d3cfd39e | 877 | vect_free_oprnd_info (&oprnds_info); |
d092494c IR |
878 | return false; |
879 | } | |
ebfd146a IR |
880 | } |
881 | } | |
882 | ||
883 | /* Add the costs of the node to the overall instance costs. */ | |
b8698a0f | 884 | *inside_cost += SLP_TREE_INSIDE_OF_LOOP_COST (*node); |
ebfd146a IR |
885 | *outside_cost += SLP_TREE_OUTSIDE_OF_LOOP_COST (*node); |
886 | ||
887 | /* Strided loads were reached - stop the recursion. */ | |
888 | if (stop_recursion) | |
889 | { | |
6aa904c4 | 890 | VEC_safe_push (slp_tree, heap, *loads, *node); |
ebfd146a IR |
891 | if (permutation) |
892 | { | |
6aa904c4 IR |
893 | |
894 | *loads_permuted = true; | |
35e1a5e7 | 895 | *inside_cost |
720f5239 | 896 | += targetm.vectorize.builtin_vectorization_cost (vec_perm, NULL, 0) |
35e1a5e7 | 897 | * group_size; |
ebfd146a | 898 | } |
2200fc49 | 899 | else |
6aa904c4 IR |
900 | { |
901 | /* We don't check here complex numbers chains, so we set | |
902 | LOADS_PERMUTED for further check in | |
903 | vect_supported_load_permutation_p. */ | |
2200fc49 | 904 | if (rhs_code == REALPART_EXPR || rhs_code == IMAGPART_EXPR) |
6aa904c4 | 905 | *loads_permuted = true; |
2200fc49 | 906 | } |
ebfd146a | 907 | |
d3cfd39e | 908 | vect_free_oprnd_info (&oprnds_info); |
ebfd146a IR |
909 | return true; |
910 | } | |
911 | ||
b8698a0f | 912 | /* Create SLP_TREE nodes for the definition node/s. */ |
d092494c | 913 | FOR_EACH_VEC_ELT (slp_oprnd_info, oprnds_info, i, oprnd_info) |
ebfd146a | 914 | { |
d092494c | 915 | slp_tree child; |
b8698a0f | 916 | |
d092494c IR |
917 | if (oprnd_info->first_dt != vect_internal_def) |
918 | continue; | |
ebfd146a | 919 | |
d092494c IR |
920 | child = vect_create_new_slp_node (oprnd_info->def_stmts); |
921 | if (!child | |
922 | || !vect_build_slp_tree (loop_vinfo, bb_vinfo, &child, group_size, | |
ebfd146a | 923 | inside_cost, outside_cost, ncopies_for_cost, |
a70d6342 | 924 | max_nunits, load_permutation, loads, |
6aa904c4 | 925 | vectorization_factor, loads_permuted)) |
d092494c | 926 | { |
d3cfd39e JJ |
927 | if (child) |
928 | oprnd_info->def_stmts = NULL; | |
929 | vect_free_slp_tree (child); | |
930 | vect_free_oprnd_info (&oprnds_info); | |
d092494c IR |
931 | return false; |
932 | } | |
b8698a0f | 933 | |
d3cfd39e | 934 | oprnd_info->def_stmts = NULL; |
d092494c | 935 | VEC_quick_push (slp_void_p, SLP_TREE_CHILDREN (*node), child); |
ebfd146a IR |
936 | } |
937 | ||
d3cfd39e | 938 | vect_free_oprnd_info (&oprnds_info); |
ebfd146a IR |
939 | return true; |
940 | } | |
941 | ||
942 | ||
943 | static void | |
944 | vect_print_slp_tree (slp_tree node) | |
945 | { | |
946 | int i; | |
947 | gimple stmt; | |
d092494c | 948 | slp_void_p child; |
ebfd146a IR |
949 | |
950 | if (!node) | |
951 | return; | |
952 | ||
953 | fprintf (vect_dump, "node "); | |
ac47786e | 954 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt) |
ebfd146a IR |
955 | { |
956 | fprintf (vect_dump, "\n\tstmt %d ", i); | |
b8698a0f | 957 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); |
ebfd146a IR |
958 | } |
959 | fprintf (vect_dump, "\n"); | |
960 | ||
d092494c IR |
961 | FOR_EACH_VEC_ELT (slp_void_p, SLP_TREE_CHILDREN (node), i, child) |
962 | vect_print_slp_tree ((slp_tree) child); | |
ebfd146a IR |
963 | } |
964 | ||
965 | ||
b8698a0f L |
966 | /* Mark the tree rooted at NODE with MARK (PURE_SLP or HYBRID). |
967 | If MARK is HYBRID, it refers to a specific stmt in NODE (the stmt at index | |
ff802fa1 | 968 | J). Otherwise, MARK is PURE_SLP and J is -1, which indicates that all the |
ebfd146a IR |
969 | stmts in NODE are to be marked. */ |
970 | ||
971 | static void | |
972 | vect_mark_slp_stmts (slp_tree node, enum slp_vect_type mark, int j) | |
973 | { | |
974 | int i; | |
975 | gimple stmt; | |
d092494c | 976 | slp_void_p child; |
ebfd146a IR |
977 | |
978 | if (!node) | |
979 | return; | |
980 | ||
ac47786e | 981 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt) |
ebfd146a IR |
982 | if (j < 0 || i == j) |
983 | STMT_SLP_TYPE (vinfo_for_stmt (stmt)) = mark; | |
984 | ||
d092494c IR |
985 | FOR_EACH_VEC_ELT (slp_void_p, SLP_TREE_CHILDREN (node), i, child) |
986 | vect_mark_slp_stmts ((slp_tree) child, mark, j); | |
ebfd146a IR |
987 | } |
988 | ||
989 | ||
a70d6342 IR |
990 | /* Mark the statements of the tree rooted at NODE as relevant (vect_used). */ |
991 | ||
992 | static void | |
993 | vect_mark_slp_stmts_relevant (slp_tree node) | |
994 | { | |
995 | int i; | |
996 | gimple stmt; | |
997 | stmt_vec_info stmt_info; | |
d092494c | 998 | slp_void_p child; |
a70d6342 IR |
999 | |
1000 | if (!node) | |
1001 | return; | |
1002 | ||
ac47786e | 1003 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt) |
a70d6342 IR |
1004 | { |
1005 | stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 1006 | gcc_assert (!STMT_VINFO_RELEVANT (stmt_info) |
a70d6342 IR |
1007 | || STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_scope); |
1008 | STMT_VINFO_RELEVANT (stmt_info) = vect_used_in_scope; | |
1009 | } | |
1010 | ||
d092494c IR |
1011 | FOR_EACH_VEC_ELT (slp_void_p, SLP_TREE_CHILDREN (node), i, child) |
1012 | vect_mark_slp_stmts_relevant ((slp_tree) child); | |
a70d6342 IR |
1013 | } |
1014 | ||
1015 | ||
b8698a0f | 1016 | /* Check if the permutation required by the SLP INSTANCE is supported. |
ebfd146a IR |
1017 | Reorganize the SLP nodes stored in SLP_INSTANCE_LOADS if needed. */ |
1018 | ||
1019 | static bool | |
1020 | vect_supported_slp_permutation_p (slp_instance instance) | |
1021 | { | |
1022 | slp_tree node = VEC_index (slp_tree, SLP_INSTANCE_LOADS (instance), 0); | |
1023 | gimple stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (node), 0); | |
e14c1050 | 1024 | gimple first_load = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)); |
ebfd146a IR |
1025 | VEC (slp_tree, heap) *sorted_loads = NULL; |
1026 | int index; | |
1027 | slp_tree *tmp_loads = NULL; | |
b8698a0f | 1028 | int group_size = SLP_INSTANCE_GROUP_SIZE (instance), i, j; |
ebfd146a | 1029 | slp_tree load; |
b8698a0f L |
1030 | |
1031 | /* FORNOW: The only supported loads permutation is loads from the same | |
ebfd146a | 1032 | location in all the loads in the node, when the data-refs in |
b8698a0f | 1033 | nodes of LOADS constitute an interleaving chain. |
ebfd146a IR |
1034 | Sort the nodes according to the order of accesses in the chain. */ |
1035 | tmp_loads = (slp_tree *) xmalloc (sizeof (slp_tree) * group_size); | |
b8698a0f L |
1036 | for (i = 0, j = 0; |
1037 | VEC_iterate (int, SLP_INSTANCE_LOAD_PERMUTATION (instance), i, index) | |
1038 | && VEC_iterate (slp_tree, SLP_INSTANCE_LOADS (instance), j, load); | |
ebfd146a IR |
1039 | i += group_size, j++) |
1040 | { | |
1041 | gimple scalar_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (load), 0); | |
1042 | /* Check that the loads are all in the same interleaving chain. */ | |
e14c1050 | 1043 | if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (scalar_stmt)) != first_load) |
ebfd146a IR |
1044 | { |
1045 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1046 | { | |
1047 | fprintf (vect_dump, "Build SLP failed: unsupported data " | |
1048 | "permutation "); | |
1049 | print_gimple_stmt (vect_dump, scalar_stmt, 0, TDF_SLIM); | |
1050 | } | |
b8698a0f | 1051 | |
ebfd146a | 1052 | free (tmp_loads); |
b8698a0f | 1053 | return false; |
ebfd146a IR |
1054 | } |
1055 | ||
1056 | tmp_loads[index] = load; | |
1057 | } | |
b8698a0f | 1058 | |
ebfd146a IR |
1059 | sorted_loads = VEC_alloc (slp_tree, heap, group_size); |
1060 | for (i = 0; i < group_size; i++) | |
1061 | VEC_safe_push (slp_tree, heap, sorted_loads, tmp_loads[i]); | |
1062 | ||
1063 | VEC_free (slp_tree, heap, SLP_INSTANCE_LOADS (instance)); | |
1064 | SLP_INSTANCE_LOADS (instance) = sorted_loads; | |
1065 | free (tmp_loads); | |
1066 | ||
1067 | if (!vect_transform_slp_perm_load (stmt, NULL, NULL, | |
1068 | SLP_INSTANCE_UNROLLING_FACTOR (instance), | |
1069 | instance, true)) | |
1070 | return false; | |
1071 | ||
1072 | return true; | |
1073 | } | |
1074 | ||
1075 | ||
b5aeb3bb IR |
1076 | /* Rearrange the statements of NODE according to PERMUTATION. */ |
1077 | ||
1078 | static void | |
1079 | vect_slp_rearrange_stmts (slp_tree node, unsigned int group_size, | |
1080 | VEC (int, heap) *permutation) | |
1081 | { | |
1082 | gimple stmt; | |
1083 | VEC (gimple, heap) *tmp_stmts; | |
1084 | unsigned int index, i; | |
d092494c | 1085 | slp_void_p child; |
b5aeb3bb IR |
1086 | |
1087 | if (!node) | |
1088 | return; | |
1089 | ||
d092494c IR |
1090 | FOR_EACH_VEC_ELT (slp_void_p, SLP_TREE_CHILDREN (node), i, child) |
1091 | vect_slp_rearrange_stmts ((slp_tree) child, group_size, permutation); | |
b5aeb3bb IR |
1092 | |
1093 | gcc_assert (group_size == VEC_length (gimple, SLP_TREE_SCALAR_STMTS (node))); | |
1094 | tmp_stmts = VEC_alloc (gimple, heap, group_size); | |
1095 | ||
1096 | for (i = 0; i < group_size; i++) | |
1097 | VEC_safe_push (gimple, heap, tmp_stmts, NULL); | |
1098 | ||
ac47786e | 1099 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt) |
b5aeb3bb IR |
1100 | { |
1101 | index = VEC_index (int, permutation, i); | |
1102 | VEC_replace (gimple, tmp_stmts, index, stmt); | |
1103 | } | |
1104 | ||
1105 | VEC_free (gimple, heap, SLP_TREE_SCALAR_STMTS (node)); | |
1106 | SLP_TREE_SCALAR_STMTS (node) = tmp_stmts; | |
1107 | } | |
1108 | ||
1109 | ||
ebfd146a IR |
1110 | /* Check if the required load permutation is supported. |
1111 | LOAD_PERMUTATION contains a list of indices of the loads. | |
1112 | In SLP this permutation is relative to the order of strided stores that are | |
1113 | the base of the SLP instance. */ | |
1114 | ||
1115 | static bool | |
1116 | vect_supported_load_permutation_p (slp_instance slp_instn, int group_size, | |
1117 | VEC (int, heap) *load_permutation) | |
1118 | { | |
b5aeb3bb IR |
1119 | int i = 0, j, prev = -1, next, k, number_of_groups; |
1120 | bool supported, bad_permutation = false; | |
7417f6c0 | 1121 | sbitmap load_index; |
2200fc49 | 1122 | slp_tree node, other_complex_node; |
6aa904c4 | 1123 | gimple stmt, first = NULL, other_node_first, load, next_load, first_load; |
2200fc49 | 1124 | unsigned complex_numbers = 0; |
6aa904c4 IR |
1125 | struct data_reference *dr; |
1126 | bb_vec_info bb_vinfo; | |
ebfd146a | 1127 | |
a70d6342 | 1128 | /* FORNOW: permutations are only supported in SLP. */ |
ebfd146a IR |
1129 | if (!slp_instn) |
1130 | return false; | |
1131 | ||
1132 | if (vect_print_dump_info (REPORT_SLP)) | |
1133 | { | |
1134 | fprintf (vect_dump, "Load permutation "); | |
ac47786e | 1135 | FOR_EACH_VEC_ELT (int, load_permutation, i, next) |
ebfd146a IR |
1136 | fprintf (vect_dump, "%d ", next); |
1137 | } | |
1138 | ||
b5aeb3bb IR |
1139 | /* In case of reduction every load permutation is allowed, since the order |
1140 | of the reduction statements is not important (as opposed to the case of | |
ff802fa1 | 1141 | strided stores). The only condition we need to check is that all the |
b5aeb3bb IR |
1142 | load nodes are of the same size and have the same permutation (and then |
1143 | rearrange all the nodes of the SLP instance according to this | |
1144 | permutation). */ | |
1145 | ||
1146 | /* Check that all the load nodes are of the same size. */ | |
ac47786e | 1147 | FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (slp_instn), i, node) |
2200fc49 IR |
1148 | { |
1149 | if (VEC_length (gimple, SLP_TREE_SCALAR_STMTS (node)) | |
1150 | != (unsigned) group_size) | |
1151 | return false; | |
1152 | ||
1153 | stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (node), 0); | |
1154 | if (is_gimple_assign (stmt) | |
1155 | && (gimple_assign_rhs_code (stmt) == REALPART_EXPR | |
1156 | || gimple_assign_rhs_code (stmt) == IMAGPART_EXPR)) | |
1157 | complex_numbers++; | |
1158 | } | |
1159 | ||
1160 | /* Complex operands can be swapped as following: | |
1161 | real_c = real_b + real_a; | |
1162 | imag_c = imag_a + imag_b; | |
1163 | i.e., we have {real_b, imag_a} and {real_a, imag_b} instead of | |
ff802fa1 | 1164 | {real_a, imag_a} and {real_b, imag_b}. We check here that if interleaving |
2200fc49 IR |
1165 | chains are mixed, they match the above pattern. */ |
1166 | if (complex_numbers) | |
1167 | { | |
ac47786e | 1168 | FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (slp_instn), i, node) |
2200fc49 | 1169 | { |
ac47786e | 1170 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), j, stmt) |
2200fc49 IR |
1171 | { |
1172 | if (j == 0) | |
1173 | first = stmt; | |
1174 | else | |
1175 | { | |
e14c1050 | 1176 | if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) != first) |
2200fc49 IR |
1177 | { |
1178 | if (complex_numbers != 2) | |
1179 | return false; | |
1180 | ||
1181 | if (i == 0) | |
1182 | k = 1; | |
1183 | else | |
1184 | k = 0; | |
1185 | ||
1186 | other_complex_node = VEC_index (slp_tree, | |
1187 | SLP_INSTANCE_LOADS (slp_instn), k); | |
1188 | other_node_first = VEC_index (gimple, | |
1189 | SLP_TREE_SCALAR_STMTS (other_complex_node), 0); | |
1190 | ||
e14c1050 | 1191 | if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) |
2200fc49 IR |
1192 | != other_node_first) |
1193 | return false; | |
1194 | } | |
1195 | } | |
1196 | } | |
1197 | } | |
1198 | } | |
1199 | ||
1200 | /* We checked that this case ok, so there is no need to proceed with | |
1201 | permutation tests. */ | |
1202 | if (complex_numbers == 2) | |
1203 | { | |
1204 | VEC_free (slp_tree, heap, SLP_INSTANCE_LOADS (slp_instn)); | |
1205 | VEC_free (int, heap, SLP_INSTANCE_LOAD_PERMUTATION (slp_instn)); | |
1206 | return true; | |
1207 | } | |
1208 | ||
b5aeb3bb IR |
1209 | node = SLP_INSTANCE_TREE (slp_instn); |
1210 | stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (node), 0); | |
1211 | /* LOAD_PERMUTATION is a list of indices of all the loads of the SLP | |
1212 | instance, not all the loads belong to the same node or interleaving | |
ff802fa1 | 1213 | group. Hence, we need to divide them into groups according to |
b5aeb3bb IR |
1214 | GROUP_SIZE. */ |
1215 | number_of_groups = VEC_length (int, load_permutation) / group_size; | |
1216 | ||
b010117a IR |
1217 | /* Reduction (there are no data-refs in the root). |
1218 | In reduction chain the order of the loads is important. */ | |
1219 | if (!STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt)) | |
1220 | && !GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) | |
b5aeb3bb IR |
1221 | { |
1222 | int first_group_load_index; | |
1223 | ||
1224 | /* Compare all the permutation sequences to the first one. */ | |
1225 | for (i = 1; i < number_of_groups; i++) | |
1226 | { | |
1227 | k = 0; | |
1228 | for (j = i * group_size; j < i * group_size + group_size; j++) | |
1229 | { | |
1230 | next = VEC_index (int, load_permutation, j); | |
1231 | first_group_load_index = VEC_index (int, load_permutation, k); | |
1232 | ||
1233 | if (next != first_group_load_index) | |
1234 | { | |
1235 | bad_permutation = true; | |
1236 | break; | |
1237 | } | |
1238 | ||
1239 | k++; | |
1240 | } | |
1241 | ||
1242 | if (bad_permutation) | |
1243 | break; | |
1244 | } | |
1245 | ||
1246 | if (!bad_permutation) | |
1247 | { | |
c9c1e775 IR |
1248 | /* Check that the loads in the first sequence are different and there |
1249 | are no gaps between them. */ | |
1250 | load_index = sbitmap_alloc (group_size); | |
1251 | sbitmap_zero (load_index); | |
1252 | for (k = 0; k < group_size; k++) | |
1253 | { | |
1254 | first_group_load_index = VEC_index (int, load_permutation, k); | |
1255 | if (TEST_BIT (load_index, first_group_load_index)) | |
1256 | { | |
1257 | bad_permutation = true; | |
1258 | break; | |
1259 | } | |
1260 | ||
1261 | SET_BIT (load_index, first_group_load_index); | |
1262 | } | |
1263 | ||
1264 | if (!bad_permutation) | |
1265 | for (k = 0; k < group_size; k++) | |
1266 | if (!TEST_BIT (load_index, k)) | |
1267 | { | |
1268 | bad_permutation = true; | |
1269 | break; | |
1270 | } | |
1271 | ||
1272 | sbitmap_free (load_index); | |
1273 | } | |
1274 | ||
1275 | if (!bad_permutation) | |
1276 | { | |
1277 | /* This permutation is valid for reduction. Since the order of the | |
b5aeb3bb IR |
1278 | statements in the nodes is not important unless they are memory |
1279 | accesses, we can rearrange the statements in all the nodes | |
1280 | according to the order of the loads. */ | |
1281 | vect_slp_rearrange_stmts (SLP_INSTANCE_TREE (slp_instn), group_size, | |
1282 | load_permutation); | |
1283 | VEC_free (int, heap, SLP_INSTANCE_LOAD_PERMUTATION (slp_instn)); | |
1284 | return true; | |
1285 | } | |
1286 | } | |
1287 | ||
6aa904c4 IR |
1288 | /* In basic block vectorization we allow any subchain of an interleaving |
1289 | chain. | |
1290 | FORNOW: not supported in loop SLP because of realignment compications. */ | |
1291 | bb_vinfo = STMT_VINFO_BB_VINFO (vinfo_for_stmt (stmt)); | |
1292 | bad_permutation = false; | |
1293 | /* Check that for every node in the instance teh loads form a subchain. */ | |
1294 | if (bb_vinfo) | |
1295 | { | |
1296 | FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (slp_instn), i, node) | |
1297 | { | |
1298 | next_load = NULL; | |
1299 | first_load = NULL; | |
1300 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), j, load) | |
1301 | { | |
1302 | if (!first_load) | |
1303 | first_load = GROUP_FIRST_ELEMENT (vinfo_for_stmt (load)); | |
1304 | else if (first_load | |
1305 | != GROUP_FIRST_ELEMENT (vinfo_for_stmt (load))) | |
1306 | { | |
1307 | bad_permutation = true; | |
1308 | break; | |
1309 | } | |
1310 | ||
1311 | if (j != 0 && next_load != load) | |
1312 | { | |
1313 | bad_permutation = true; | |
1314 | break; | |
1315 | } | |
1316 | ||
1317 | next_load = GROUP_NEXT_ELEMENT (vinfo_for_stmt (load)); | |
1318 | } | |
1319 | ||
1320 | if (bad_permutation) | |
1321 | break; | |
1322 | } | |
1323 | ||
1324 | /* Check that the alignment of the first load in every subchain, i.e., | |
1325 | the first statement in every load node, is supported. */ | |
1326 | if (!bad_permutation) | |
1327 | { | |
1328 | FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (slp_instn), i, node) | |
1329 | { | |
1330 | first_load = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (node), 0); | |
1331 | if (first_load | |
1332 | != GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_load))) | |
1333 | { | |
1334 | dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_load)); | |
1335 | if (vect_supportable_dr_alignment (dr, false) | |
1336 | == dr_unaligned_unsupported) | |
1337 | { | |
1338 | if (vect_print_dump_info (REPORT_SLP)) | |
1339 | { | |
1340 | fprintf (vect_dump, "unsupported unaligned load "); | |
1341 | print_gimple_stmt (vect_dump, first_load, 0, | |
1342 | TDF_SLIM); | |
1343 | } | |
1344 | bad_permutation = true; | |
1345 | break; | |
1346 | } | |
1347 | } | |
1348 | } | |
1349 | ||
1350 | if (!bad_permutation) | |
1351 | { | |
1352 | VEC_free (int, heap, SLP_INSTANCE_LOAD_PERMUTATION (slp_instn)); | |
1353 | return true; | |
1354 | } | |
1355 | } | |
1356 | } | |
1357 | ||
b8698a0f L |
1358 | /* FORNOW: the only supported permutation is 0..01..1.. of length equal to |
1359 | GROUP_SIZE and where each sequence of same drs is of GROUP_SIZE length as | |
b5aeb3bb | 1360 | well (unless it's reduction). */ |
ebfd146a IR |
1361 | if (VEC_length (int, load_permutation) |
1362 | != (unsigned int) (group_size * group_size)) | |
1363 | return false; | |
1364 | ||
1365 | supported = true; | |
7417f6c0 IR |
1366 | load_index = sbitmap_alloc (group_size); |
1367 | sbitmap_zero (load_index); | |
ebfd146a IR |
1368 | for (j = 0; j < group_size; j++) |
1369 | { | |
1370 | for (i = j * group_size, k = 0; | |
1371 | VEC_iterate (int, load_permutation, i, next) && k < group_size; | |
1372 | i++, k++) | |
1373 | { | |
1374 | if (i != j * group_size && next != prev) | |
1375 | { | |
1376 | supported = false; | |
1377 | break; | |
1378 | } | |
1379 | ||
1380 | prev = next; | |
b8698a0f | 1381 | } |
7417f6c0 IR |
1382 | |
1383 | if (TEST_BIT (load_index, prev)) | |
1384 | { | |
1385 | supported = false; | |
1386 | break; | |
1387 | } | |
1388 | ||
1389 | SET_BIT (load_index, prev); | |
ebfd146a | 1390 | } |
59eefaa6 IR |
1391 | |
1392 | for (j = 0; j < group_size; j++) | |
1393 | if (!TEST_BIT (load_index, j)) | |
1394 | return false; | |
1395 | ||
7417f6c0 | 1396 | sbitmap_free (load_index); |
ebfd146a IR |
1397 | |
1398 | if (supported && i == group_size * group_size | |
1399 | && vect_supported_slp_permutation_p (slp_instn)) | |
1400 | return true; | |
1401 | ||
b8698a0f | 1402 | return false; |
ebfd146a IR |
1403 | } |
1404 | ||
1405 | ||
b8698a0f | 1406 | /* Find the first load in the loop that belongs to INSTANCE. |
ebfd146a | 1407 | When loads are in several SLP nodes, there can be a case in which the first |
b8698a0f | 1408 | load does not appear in the first SLP node to be transformed, causing |
ff802fa1 | 1409 | incorrect order of statements. Since we generate all the loads together, |
ebfd146a IR |
1410 | they must be inserted before the first load of the SLP instance and not |
1411 | before the first load of the first node of the instance. */ | |
ff802fa1 | 1412 | |
b8698a0f L |
1413 | static gimple |
1414 | vect_find_first_load_in_slp_instance (slp_instance instance) | |
ebfd146a IR |
1415 | { |
1416 | int i, j; | |
1417 | slp_tree load_node; | |
1418 | gimple first_load = NULL, load; | |
1419 | ||
ac47786e NF |
1420 | FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (instance), i, load_node) |
1421 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (load_node), j, load) | |
ebfd146a | 1422 | first_load = get_earlier_stmt (load, first_load); |
b8698a0f | 1423 | |
ebfd146a IR |
1424 | return first_load; |
1425 | } | |
1426 | ||
1427 | ||
e4a707c4 | 1428 | /* Find the last store in SLP INSTANCE. */ |
ff802fa1 | 1429 | |
e4a707c4 IR |
1430 | static gimple |
1431 | vect_find_last_store_in_slp_instance (slp_instance instance) | |
1432 | { | |
1433 | int i; | |
1434 | slp_tree node; | |
1435 | gimple last_store = NULL, store; | |
1436 | ||
1437 | node = SLP_INSTANCE_TREE (instance); | |
1438 | for (i = 0; | |
1439 | VEC_iterate (gimple, SLP_TREE_SCALAR_STMTS (node), i, store); | |
1440 | i++) | |
1441 | last_store = get_later_stmt (store, last_store); | |
1442 | ||
1443 | return last_store; | |
1444 | } | |
1445 | ||
1446 | ||
ff802fa1 | 1447 | /* Analyze an SLP instance starting from a group of strided stores. Call |
b8698a0f | 1448 | vect_build_slp_tree to build a tree of packed stmts if possible. |
ebfd146a IR |
1449 | Return FALSE if it's impossible to SLP any stmt in the loop. */ |
1450 | ||
1451 | static bool | |
a70d6342 IR |
1452 | vect_analyze_slp_instance (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo, |
1453 | gimple stmt) | |
ebfd146a IR |
1454 | { |
1455 | slp_instance new_instance; | |
d092494c | 1456 | slp_tree node; |
e14c1050 | 1457 | unsigned int group_size = GROUP_SIZE (vinfo_for_stmt (stmt)); |
ebfd146a | 1458 | unsigned int unrolling_factor = 1, nunits; |
b5aeb3bb | 1459 | tree vectype, scalar_type = NULL_TREE; |
ebfd146a | 1460 | gimple next; |
0f900dfa | 1461 | unsigned int vectorization_factor = 0; |
b5aeb3bb | 1462 | int inside_cost = 0, outside_cost = 0, ncopies_for_cost, i; |
ebfd146a IR |
1463 | unsigned int max_nunits = 0; |
1464 | VEC (int, heap) *load_permutation; | |
1465 | VEC (slp_tree, heap) *loads; | |
b5aeb3bb | 1466 | struct data_reference *dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt)); |
6aa904c4 | 1467 | bool loads_permuted = false; |
d092494c | 1468 | VEC (gimple, heap) *scalar_stmts; |
b5aeb3bb | 1469 | |
b010117a | 1470 | if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) |
b5aeb3bb | 1471 | { |
b010117a IR |
1472 | if (dr) |
1473 | { | |
1474 | scalar_type = TREE_TYPE (DR_REF (dr)); | |
1475 | vectype = get_vectype_for_scalar_type (scalar_type); | |
1476 | } | |
1477 | else | |
1478 | { | |
1479 | gcc_assert (loop_vinfo); | |
1480 | vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt)); | |
1481 | } | |
1482 | ||
e14c1050 | 1483 | group_size = GROUP_SIZE (vinfo_for_stmt (stmt)); |
b5aeb3bb IR |
1484 | } |
1485 | else | |
1486 | { | |
1487 | gcc_assert (loop_vinfo); | |
1488 | vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt)); | |
1489 | group_size = VEC_length (gimple, LOOP_VINFO_REDUCTIONS (loop_vinfo)); | |
1490 | } | |
b8698a0f | 1491 | |
ebfd146a IR |
1492 | if (!vectype) |
1493 | { | |
1494 | if (vect_print_dump_info (REPORT_SLP)) | |
1495 | { | |
1496 | fprintf (vect_dump, "Build SLP failed: unsupported data-type "); | |
1497 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
1498 | } | |
b5aeb3bb | 1499 | |
ebfd146a IR |
1500 | return false; |
1501 | } | |
1502 | ||
1503 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
a70d6342 IR |
1504 | if (loop_vinfo) |
1505 | vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
1506 | else | |
a70d6342 IR |
1507 | vectorization_factor = nunits; |
1508 | ||
a70d6342 IR |
1509 | /* Calculate the unrolling factor. */ |
1510 | unrolling_factor = least_common_multiple (nunits, group_size) / group_size; | |
1511 | if (unrolling_factor != 1 && !loop_vinfo) | |
1512 | { | |
1513 | if (vect_print_dump_info (REPORT_SLP)) | |
e9dbe7bb IR |
1514 | fprintf (vect_dump, "Build SLP failed: unrolling required in basic" |
1515 | " block SLP"); | |
b8698a0f | 1516 | |
a70d6342 IR |
1517 | return false; |
1518 | } | |
1519 | ||
b8698a0f | 1520 | /* Create a node (a root of the SLP tree) for the packed strided stores. */ |
d092494c | 1521 | scalar_stmts = VEC_alloc (gimple, heap, group_size); |
ebfd146a | 1522 | next = stmt; |
b010117a | 1523 | if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) |
ebfd146a | 1524 | { |
b5aeb3bb IR |
1525 | /* Collect the stores and store them in SLP_TREE_SCALAR_STMTS. */ |
1526 | while (next) | |
1527 | { | |
f7e531cf IR |
1528 | if (STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (next)) |
1529 | && STMT_VINFO_RELATED_STMT (vinfo_for_stmt (next))) | |
1530 | VEC_safe_push (gimple, heap, scalar_stmts, | |
1531 | STMT_VINFO_RELATED_STMT (vinfo_for_stmt (next))); | |
1532 | else | |
1533 | VEC_safe_push (gimple, heap, scalar_stmts, next); | |
e14c1050 | 1534 | next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next)); |
b5aeb3bb IR |
1535 | } |
1536 | } | |
1537 | else | |
1538 | { | |
1539 | /* Collect reduction statements. */ | |
d092494c IR |
1540 | VEC (gimple, heap) *reductions = LOOP_VINFO_REDUCTIONS (loop_vinfo); |
1541 | for (i = 0; VEC_iterate (gimple, reductions, i, next); i++) | |
f7e531cf | 1542 | VEC_safe_push (gimple, heap, scalar_stmts, next); |
ebfd146a IR |
1543 | } |
1544 | ||
d092494c | 1545 | node = vect_create_new_slp_node (scalar_stmts); |
ebfd146a | 1546 | |
ebfd146a IR |
1547 | /* Calculate the number of vector stmts to create based on the unrolling |
1548 | factor (number of vectors is 1 if NUNITS >= GROUP_SIZE, and is | |
1549 | GROUP_SIZE / NUNITS otherwise. */ | |
1550 | ncopies_for_cost = unrolling_factor * group_size / nunits; | |
b8698a0f L |
1551 | |
1552 | load_permutation = VEC_alloc (int, heap, group_size * group_size); | |
1553 | loads = VEC_alloc (slp_tree, heap, group_size); | |
ebfd146a IR |
1554 | |
1555 | /* Build the tree for the SLP instance. */ | |
b8698a0f L |
1556 | if (vect_build_slp_tree (loop_vinfo, bb_vinfo, &node, group_size, |
1557 | &inside_cost, &outside_cost, ncopies_for_cost, | |
1558 | &max_nunits, &load_permutation, &loads, | |
6aa904c4 | 1559 | vectorization_factor, &loads_permuted)) |
ebfd146a | 1560 | { |
4ef69dfc | 1561 | /* Calculate the unrolling factor based on the smallest type. */ |
ebfd146a IR |
1562 | if (max_nunits > nunits) |
1563 | unrolling_factor = least_common_multiple (max_nunits, group_size) | |
1564 | / group_size; | |
b8698a0f | 1565 | |
4ef69dfc IR |
1566 | if (unrolling_factor != 1 && !loop_vinfo) |
1567 | { | |
1568 | if (vect_print_dump_info (REPORT_SLP)) | |
1569 | fprintf (vect_dump, "Build SLP failed: unrolling required in basic" | |
1570 | " block SLP"); | |
1571 | return false; | |
1572 | } | |
1573 | ||
1574 | /* Create a new SLP instance. */ | |
1575 | new_instance = XNEW (struct _slp_instance); | |
1576 | SLP_INSTANCE_TREE (new_instance) = node; | |
1577 | SLP_INSTANCE_GROUP_SIZE (new_instance) = group_size; | |
ebfd146a IR |
1578 | SLP_INSTANCE_UNROLLING_FACTOR (new_instance) = unrolling_factor; |
1579 | SLP_INSTANCE_OUTSIDE_OF_LOOP_COST (new_instance) = outside_cost; | |
1580 | SLP_INSTANCE_INSIDE_OF_LOOP_COST (new_instance) = inside_cost; | |
1581 | SLP_INSTANCE_LOADS (new_instance) = loads; | |
1582 | SLP_INSTANCE_FIRST_LOAD_STMT (new_instance) = NULL; | |
1583 | SLP_INSTANCE_LOAD_PERMUTATION (new_instance) = load_permutation; | |
6aa904c4 IR |
1584 | |
1585 | if (loads_permuted) | |
ebfd146a IR |
1586 | { |
1587 | if (!vect_supported_load_permutation_p (new_instance, group_size, | |
b8698a0f | 1588 | load_permutation)) |
ebfd146a IR |
1589 | { |
1590 | if (vect_print_dump_info (REPORT_SLP)) | |
1591 | { | |
1592 | fprintf (vect_dump, "Build SLP failed: unsupported load " | |
1593 | "permutation "); | |
1594 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
1595 | } | |
1596 | ||
1597 | vect_free_slp_instance (new_instance); | |
1598 | return false; | |
1599 | } | |
1600 | ||
1601 | SLP_INSTANCE_FIRST_LOAD_STMT (new_instance) | |
1602 | = vect_find_first_load_in_slp_instance (new_instance); | |
1603 | } | |
1604 | else | |
1605 | VEC_free (int, heap, SLP_INSTANCE_LOAD_PERMUTATION (new_instance)); | |
1606 | ||
a70d6342 | 1607 | if (loop_vinfo) |
b8698a0f L |
1608 | VEC_safe_push (slp_instance, heap, |
1609 | LOOP_VINFO_SLP_INSTANCES (loop_vinfo), | |
a70d6342 IR |
1610 | new_instance); |
1611 | else | |
1612 | VEC_safe_push (slp_instance, heap, BB_VINFO_SLP_INSTANCES (bb_vinfo), | |
1613 | new_instance); | |
b8698a0f | 1614 | |
ebfd146a IR |
1615 | if (vect_print_dump_info (REPORT_SLP)) |
1616 | vect_print_slp_tree (node); | |
1617 | ||
1618 | return true; | |
1619 | } | |
1620 | ||
1621 | /* Failed to SLP. */ | |
1622 | /* Free the allocated memory. */ | |
1623 | vect_free_slp_tree (node); | |
1624 | VEC_free (int, heap, load_permutation); | |
1625 | VEC_free (slp_tree, heap, loads); | |
b8698a0f | 1626 | |
a70d6342 | 1627 | return false; |
ebfd146a IR |
1628 | } |
1629 | ||
1630 | ||
ff802fa1 | 1631 | /* Check if there are stmts in the loop can be vectorized using SLP. Build SLP |
ebfd146a IR |
1632 | trees of packed scalar stmts if SLP is possible. */ |
1633 | ||
1634 | bool | |
a70d6342 | 1635 | vect_analyze_slp (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo) |
ebfd146a IR |
1636 | { |
1637 | unsigned int i; | |
b010117a IR |
1638 | VEC (gimple, heap) *strided_stores, *reductions = NULL, *reduc_chains = NULL; |
1639 | gimple first_element; | |
a70d6342 | 1640 | bool ok = false; |
ebfd146a IR |
1641 | |
1642 | if (vect_print_dump_info (REPORT_SLP)) | |
1643 | fprintf (vect_dump, "=== vect_analyze_slp ==="); | |
1644 | ||
a70d6342 | 1645 | if (loop_vinfo) |
b5aeb3bb IR |
1646 | { |
1647 | strided_stores = LOOP_VINFO_STRIDED_STORES (loop_vinfo); | |
b010117a | 1648 | reduc_chains = LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo); |
b5aeb3bb IR |
1649 | reductions = LOOP_VINFO_REDUCTIONS (loop_vinfo); |
1650 | } | |
a70d6342 IR |
1651 | else |
1652 | strided_stores = BB_VINFO_STRIDED_STORES (bb_vinfo); | |
b8698a0f | 1653 | |
b5aeb3bb | 1654 | /* Find SLP sequences starting from groups of strided stores. */ |
b010117a IR |
1655 | FOR_EACH_VEC_ELT (gimple, strided_stores, i, first_element) |
1656 | if (vect_analyze_slp_instance (loop_vinfo, bb_vinfo, first_element)) | |
a70d6342 | 1657 | ok = true; |
ebfd146a | 1658 | |
b8698a0f | 1659 | if (bb_vinfo && !ok) |
a70d6342 IR |
1660 | { |
1661 | if (vect_print_dump_info (REPORT_SLP)) | |
1662 | fprintf (vect_dump, "Failed to SLP the basic block."); | |
1663 | ||
1664 | return false; | |
1665 | } | |
ebfd146a | 1666 | |
b010117a IR |
1667 | if (loop_vinfo |
1668 | && VEC_length (gimple, LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo)) > 0) | |
1669 | { | |
1670 | /* Find SLP sequences starting from reduction chains. */ | |
1671 | FOR_EACH_VEC_ELT (gimple, reduc_chains, i, first_element) | |
1672 | if (vect_analyze_slp_instance (loop_vinfo, bb_vinfo, first_element)) | |
1673 | ok = true; | |
1674 | else | |
1675 | return false; | |
1676 | ||
1677 | /* Don't try to vectorize SLP reductions if reduction chain was | |
1678 | detected. */ | |
1679 | return ok; | |
1680 | } | |
1681 | ||
b5aeb3bb | 1682 | /* Find SLP sequences starting from groups of reductions. */ |
74500b3e | 1683 | if (loop_vinfo && VEC_length (gimple, LOOP_VINFO_REDUCTIONS (loop_vinfo)) > 1 |
b5aeb3bb IR |
1684 | && vect_analyze_slp_instance (loop_vinfo, bb_vinfo, |
1685 | VEC_index (gimple, reductions, 0))) | |
1686 | ok = true; | |
1687 | ||
ebfd146a IR |
1688 | return true; |
1689 | } | |
1690 | ||
1691 | ||
1692 | /* For each possible SLP instance decide whether to SLP it and calculate overall | |
437f4a00 IR |
1693 | unrolling factor needed to SLP the loop. Return TRUE if decided to SLP at |
1694 | least one instance. */ | |
ebfd146a | 1695 | |
437f4a00 | 1696 | bool |
ebfd146a IR |
1697 | vect_make_slp_decision (loop_vec_info loop_vinfo) |
1698 | { | |
1699 | unsigned int i, unrolling_factor = 1; | |
1700 | VEC (slp_instance, heap) *slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo); | |
1701 | slp_instance instance; | |
1702 | int decided_to_slp = 0; | |
1703 | ||
1704 | if (vect_print_dump_info (REPORT_SLP)) | |
1705 | fprintf (vect_dump, "=== vect_make_slp_decision ==="); | |
1706 | ||
ac47786e | 1707 | FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance) |
ebfd146a IR |
1708 | { |
1709 | /* FORNOW: SLP if you can. */ | |
1710 | if (unrolling_factor < SLP_INSTANCE_UNROLLING_FACTOR (instance)) | |
1711 | unrolling_factor = SLP_INSTANCE_UNROLLING_FACTOR (instance); | |
1712 | ||
ff802fa1 | 1713 | /* Mark all the stmts that belong to INSTANCE as PURE_SLP stmts. Later we |
b8698a0f | 1714 | call vect_detect_hybrid_slp () to find stmts that need hybrid SLP and |
ff802fa1 | 1715 | loop-based vectorization. Such stmts will be marked as HYBRID. */ |
ebfd146a IR |
1716 | vect_mark_slp_stmts (SLP_INSTANCE_TREE (instance), pure_slp, -1); |
1717 | decided_to_slp++; | |
1718 | } | |
1719 | ||
1720 | LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo) = unrolling_factor; | |
1721 | ||
b8698a0f L |
1722 | if (decided_to_slp && vect_print_dump_info (REPORT_SLP)) |
1723 | fprintf (vect_dump, "Decided to SLP %d instances. Unrolling factor %d", | |
ebfd146a | 1724 | decided_to_slp, unrolling_factor); |
437f4a00 IR |
1725 | |
1726 | return (decided_to_slp > 0); | |
ebfd146a IR |
1727 | } |
1728 | ||
1729 | ||
1730 | /* Find stmts that must be both vectorized and SLPed (since they feed stmts that | |
ff802fa1 | 1731 | can't be SLPed) in the tree rooted at NODE. Mark such stmts as HYBRID. */ |
ebfd146a IR |
1732 | |
1733 | static void | |
1734 | vect_detect_hybrid_slp_stmts (slp_tree node) | |
1735 | { | |
1736 | int i; | |
f2c74cc4 IR |
1737 | VEC (gimple, heap) *stmts = SLP_TREE_SCALAR_STMTS (node); |
1738 | gimple stmt = VEC_index (gimple, stmts, 0); | |
ebfd146a IR |
1739 | imm_use_iterator imm_iter; |
1740 | gimple use_stmt; | |
f2c74cc4 | 1741 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
d092494c | 1742 | slp_void_p child; |
f2c74cc4 IR |
1743 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
1744 | struct loop *loop = NULL; | |
1745 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1746 | basic_block bb = NULL; | |
ebfd146a IR |
1747 | |
1748 | if (!node) | |
1749 | return; | |
1750 | ||
f2c74cc4 IR |
1751 | if (loop_vinfo) |
1752 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1753 | else | |
1754 | bb = BB_VINFO_BB (bb_vinfo); | |
1755 | ||
ac47786e | 1756 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt) |
ebfd146a IR |
1757 | if (PURE_SLP_STMT (vinfo_for_stmt (stmt)) |
1758 | && TREE_CODE (gimple_op (stmt, 0)) == SSA_NAME) | |
1759 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, gimple_op (stmt, 0)) | |
f2c74cc4 IR |
1760 | if (gimple_bb (use_stmt) |
1761 | && ((loop && flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) | |
1762 | || bb == gimple_bb (use_stmt)) | |
1763 | && (stmt_vinfo = vinfo_for_stmt (use_stmt)) | |
99f51320 IR |
1764 | && !STMT_SLP_TYPE (stmt_vinfo) |
1765 | && (STMT_VINFO_RELEVANT (stmt_vinfo) | |
b5aeb3bb | 1766 | || VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_vinfo))) |
f2c74cc4 IR |
1767 | && !(gimple_code (use_stmt) == GIMPLE_PHI |
1768 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) | |
1769 | == vect_reduction_def)) | |
ebfd146a IR |
1770 | vect_mark_slp_stmts (node, hybrid, i); |
1771 | ||
d092494c IR |
1772 | FOR_EACH_VEC_ELT (slp_void_p, SLP_TREE_CHILDREN (node), i, child) |
1773 | vect_detect_hybrid_slp_stmts ((slp_tree) child); | |
ebfd146a IR |
1774 | } |
1775 | ||
1776 | ||
1777 | /* Find stmts that must be both vectorized and SLPed. */ | |
1778 | ||
1779 | void | |
1780 | vect_detect_hybrid_slp (loop_vec_info loop_vinfo) | |
1781 | { | |
1782 | unsigned int i; | |
1783 | VEC (slp_instance, heap) *slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo); | |
1784 | slp_instance instance; | |
1785 | ||
1786 | if (vect_print_dump_info (REPORT_SLP)) | |
1787 | fprintf (vect_dump, "=== vect_detect_hybrid_slp ==="); | |
1788 | ||
ac47786e | 1789 | FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance) |
ebfd146a IR |
1790 | vect_detect_hybrid_slp_stmts (SLP_INSTANCE_TREE (instance)); |
1791 | } | |
1792 | ||
a70d6342 IR |
1793 | |
1794 | /* Create and initialize a new bb_vec_info struct for BB, as well as | |
1795 | stmt_vec_info structs for all the stmts in it. */ | |
b8698a0f | 1796 | |
a70d6342 IR |
1797 | static bb_vec_info |
1798 | new_bb_vec_info (basic_block bb) | |
1799 | { | |
1800 | bb_vec_info res = NULL; | |
1801 | gimple_stmt_iterator gsi; | |
1802 | ||
1803 | res = (bb_vec_info) xcalloc (1, sizeof (struct _bb_vec_info)); | |
1804 | BB_VINFO_BB (res) = bb; | |
1805 | ||
1806 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1807 | { | |
1808 | gimple stmt = gsi_stmt (gsi); | |
1809 | gimple_set_uid (stmt, 0); | |
1810 | set_vinfo_for_stmt (stmt, new_stmt_vec_info (stmt, NULL, res)); | |
1811 | } | |
1812 | ||
1813 | BB_VINFO_STRIDED_STORES (res) = VEC_alloc (gimple, heap, 10); | |
1814 | BB_VINFO_SLP_INSTANCES (res) = VEC_alloc (slp_instance, heap, 2); | |
1815 | ||
1816 | bb->aux = res; | |
1817 | return res; | |
1818 | } | |
1819 | ||
1820 | ||
1821 | /* Free BB_VINFO struct, as well as all the stmt_vec_info structs of all the | |
1822 | stmts in the basic block. */ | |
1823 | ||
1824 | static void | |
1825 | destroy_bb_vec_info (bb_vec_info bb_vinfo) | |
1826 | { | |
1827 | basic_block bb; | |
1828 | gimple_stmt_iterator si; | |
1829 | ||
1830 | if (!bb_vinfo) | |
1831 | return; | |
1832 | ||
1833 | bb = BB_VINFO_BB (bb_vinfo); | |
1834 | ||
1835 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
1836 | { | |
1837 | gimple stmt = gsi_stmt (si); | |
1838 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1839 | ||
1840 | if (stmt_info) | |
1841 | /* Free stmt_vec_info. */ | |
1842 | free_stmt_vec_info (stmt); | |
1843 | } | |
1844 | ||
01be8516 SP |
1845 | free_data_refs (BB_VINFO_DATAREFS (bb_vinfo)); |
1846 | free_dependence_relations (BB_VINFO_DDRS (bb_vinfo)); | |
a70d6342 IR |
1847 | VEC_free (gimple, heap, BB_VINFO_STRIDED_STORES (bb_vinfo)); |
1848 | VEC_free (slp_instance, heap, BB_VINFO_SLP_INSTANCES (bb_vinfo)); | |
1849 | free (bb_vinfo); | |
1850 | bb->aux = NULL; | |
1851 | } | |
1852 | ||
1853 | ||
1854 | /* Analyze statements contained in SLP tree node after recursively analyzing | |
1855 | the subtree. Return TRUE if the operations are supported. */ | |
1856 | ||
1857 | static bool | |
1858 | vect_slp_analyze_node_operations (bb_vec_info bb_vinfo, slp_tree node) | |
1859 | { | |
1860 | bool dummy; | |
1861 | int i; | |
1862 | gimple stmt; | |
d092494c | 1863 | slp_void_p child; |
a70d6342 IR |
1864 | |
1865 | if (!node) | |
1866 | return true; | |
1867 | ||
d092494c IR |
1868 | FOR_EACH_VEC_ELT (slp_void_p, SLP_TREE_CHILDREN (node), i, child) |
1869 | if (!vect_slp_analyze_node_operations (bb_vinfo, (slp_tree) child)) | |
1870 | return false; | |
a70d6342 | 1871 | |
ac47786e | 1872 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt) |
a70d6342 IR |
1873 | { |
1874 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1875 | gcc_assert (stmt_info); | |
1876 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
1877 | ||
1878 | if (!vect_analyze_stmt (stmt, &dummy, node)) | |
1879 | return false; | |
1880 | } | |
1881 | ||
1882 | return true; | |
1883 | } | |
1884 | ||
1885 | ||
ff802fa1 | 1886 | /* Analyze statements in SLP instances of the basic block. Return TRUE if the |
a70d6342 IR |
1887 | operations are supported. */ |
1888 | ||
1889 | static bool | |
1890 | vect_slp_analyze_operations (bb_vec_info bb_vinfo) | |
1891 | { | |
1892 | VEC (slp_instance, heap) *slp_instances = BB_VINFO_SLP_INSTANCES (bb_vinfo); | |
1893 | slp_instance instance; | |
1894 | int i; | |
1895 | ||
1896 | for (i = 0; VEC_iterate (slp_instance, slp_instances, i, instance); ) | |
1897 | { | |
b8698a0f | 1898 | if (!vect_slp_analyze_node_operations (bb_vinfo, |
a70d6342 IR |
1899 | SLP_INSTANCE_TREE (instance))) |
1900 | { | |
1901 | vect_free_slp_instance (instance); | |
1902 | VEC_ordered_remove (slp_instance, slp_instances, i); | |
1903 | } | |
1904 | else | |
1905 | i++; | |
b8698a0f L |
1906 | } |
1907 | ||
a70d6342 IR |
1908 | if (!VEC_length (slp_instance, slp_instances)) |
1909 | return false; | |
1910 | ||
1911 | return true; | |
1912 | } | |
1913 | ||
69f11a13 IR |
1914 | /* Check if vectorization of the basic block is profitable. */ |
1915 | ||
1916 | static bool | |
1917 | vect_bb_vectorization_profitable_p (bb_vec_info bb_vinfo) | |
1918 | { | |
1919 | VEC (slp_instance, heap) *slp_instances = BB_VINFO_SLP_INSTANCES (bb_vinfo); | |
1920 | slp_instance instance; | |
1921 | int i; | |
1922 | unsigned int vec_outside_cost = 0, vec_inside_cost = 0, scalar_cost = 0; | |
1923 | unsigned int stmt_cost; | |
1924 | gimple stmt; | |
1925 | gimple_stmt_iterator si; | |
1926 | basic_block bb = BB_VINFO_BB (bb_vinfo); | |
1927 | stmt_vec_info stmt_info = NULL; | |
1928 | tree dummy_type = NULL; | |
1929 | int dummy = 0; | |
1930 | ||
1931 | /* Calculate vector costs. */ | |
ac47786e | 1932 | FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance) |
69f11a13 IR |
1933 | { |
1934 | vec_outside_cost += SLP_INSTANCE_OUTSIDE_OF_LOOP_COST (instance); | |
1935 | vec_inside_cost += SLP_INSTANCE_INSIDE_OF_LOOP_COST (instance); | |
1936 | } | |
1937 | ||
1938 | /* Calculate scalar cost. */ | |
1939 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
1940 | { | |
1941 | stmt = gsi_stmt (si); | |
1942 | stmt_info = vinfo_for_stmt (stmt); | |
1943 | ||
1944 | if (!stmt_info || !STMT_VINFO_VECTORIZABLE (stmt_info) | |
1945 | || !PURE_SLP_STMT (stmt_info)) | |
1946 | continue; | |
1947 | ||
1948 | if (STMT_VINFO_DATA_REF (stmt_info)) | |
1949 | { | |
1950 | if (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info))) | |
1951 | stmt_cost = targetm.vectorize.builtin_vectorization_cost | |
1952 | (scalar_load, dummy_type, dummy); | |
1953 | else | |
1954 | stmt_cost = targetm.vectorize.builtin_vectorization_cost | |
1955 | (scalar_store, dummy_type, dummy); | |
1956 | } | |
1957 | else | |
1958 | stmt_cost = targetm.vectorize.builtin_vectorization_cost | |
1959 | (scalar_stmt, dummy_type, dummy); | |
1960 | ||
1961 | scalar_cost += stmt_cost; | |
1962 | } | |
1963 | ||
1964 | if (vect_print_dump_info (REPORT_COST)) | |
1965 | { | |
1966 | fprintf (vect_dump, "Cost model analysis: \n"); | |
1967 | fprintf (vect_dump, " Vector inside of basic block cost: %d\n", | |
1968 | vec_inside_cost); | |
1969 | fprintf (vect_dump, " Vector outside of basic block cost: %d\n", | |
1970 | vec_outside_cost); | |
1971 | fprintf (vect_dump, " Scalar cost of basic block: %d", scalar_cost); | |
1972 | } | |
1973 | ||
1974 | /* Vectorization is profitable if its cost is less than the cost of scalar | |
1975 | version. */ | |
1976 | if (vec_outside_cost + vec_inside_cost >= scalar_cost) | |
1977 | return false; | |
1978 | ||
1979 | return true; | |
1980 | } | |
1981 | ||
1982 | /* Check if the basic block can be vectorized. */ | |
a70d6342 | 1983 | |
8e19f5a1 IR |
1984 | static bb_vec_info |
1985 | vect_slp_analyze_bb_1 (basic_block bb) | |
a70d6342 IR |
1986 | { |
1987 | bb_vec_info bb_vinfo; | |
1988 | VEC (ddr_p, heap) *ddrs; | |
1989 | VEC (slp_instance, heap) *slp_instances; | |
1990 | slp_instance instance; | |
8e19f5a1 | 1991 | int i; |
777e1f09 RG |
1992 | int min_vf = 2; |
1993 | int max_vf = MAX_VECTORIZATION_FACTOR; | |
e4a707c4 | 1994 | |
a70d6342 IR |
1995 | bb_vinfo = new_bb_vec_info (bb); |
1996 | if (!bb_vinfo) | |
1997 | return NULL; | |
1998 | ||
777e1f09 | 1999 | if (!vect_analyze_data_refs (NULL, bb_vinfo, &min_vf)) |
a70d6342 IR |
2000 | { |
2001 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
2002 | fprintf (vect_dump, "not vectorized: unhandled data-ref in basic " | |
2003 | "block.\n"); | |
b8698a0f | 2004 | |
a70d6342 IR |
2005 | destroy_bb_vec_info (bb_vinfo); |
2006 | return NULL; | |
2007 | } | |
2008 | ||
2009 | ddrs = BB_VINFO_DDRS (bb_vinfo); | |
b8698a0f | 2010 | if (!VEC_length (ddr_p, ddrs)) |
a70d6342 IR |
2011 | { |
2012 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
2013 | fprintf (vect_dump, "not vectorized: not enough data-refs in basic " | |
2014 | "block.\n"); | |
2015 | ||
2016 | destroy_bb_vec_info (bb_vinfo); | |
2017 | return NULL; | |
2018 | } | |
2019 | ||
f5709183 IR |
2020 | vect_pattern_recog (NULL, bb_vinfo); |
2021 | ||
2022 | if (!vect_analyze_data_ref_dependences (NULL, bb_vinfo, &max_vf) | |
5bfdb7d8 | 2023 | || min_vf > max_vf) |
777e1f09 RG |
2024 | { |
2025 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
2026 | fprintf (vect_dump, "not vectorized: unhandled data dependence " | |
2027 | "in basic block.\n"); | |
2028 | ||
2029 | destroy_bb_vec_info (bb_vinfo); | |
2030 | return NULL; | |
2031 | } | |
2032 | ||
a70d6342 IR |
2033 | if (!vect_analyze_data_refs_alignment (NULL, bb_vinfo)) |
2034 | { | |
2035 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
2036 | fprintf (vect_dump, "not vectorized: bad data alignment in basic " | |
2037 | "block.\n"); | |
b8698a0f | 2038 | |
a70d6342 IR |
2039 | destroy_bb_vec_info (bb_vinfo); |
2040 | return NULL; | |
2041 | } | |
b8698a0f | 2042 | |
a70d6342 IR |
2043 | if (!vect_analyze_data_ref_accesses (NULL, bb_vinfo)) |
2044 | { | |
2045 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
2046 | fprintf (vect_dump, "not vectorized: unhandled data access in basic " | |
2047 | "block.\n"); | |
b8698a0f | 2048 | |
a70d6342 IR |
2049 | destroy_bb_vec_info (bb_vinfo); |
2050 | return NULL; | |
2051 | } | |
2052 | ||
2053 | if (!vect_verify_datarefs_alignment (NULL, bb_vinfo)) | |
2054 | { | |
2055 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
2056 | fprintf (vect_dump, "not vectorized: unsupported alignment in basic " | |
2057 | "block.\n"); | |
2058 | ||
2059 | destroy_bb_vec_info (bb_vinfo); | |
2060 | return NULL; | |
2061 | } | |
2062 | ||
2063 | /* Check the SLP opportunities in the basic block, analyze and build SLP | |
2064 | trees. */ | |
2065 | if (!vect_analyze_slp (NULL, bb_vinfo)) | |
2066 | { | |
2067 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
2068 | fprintf (vect_dump, "not vectorized: failed to find SLP opportunities " | |
2069 | "in basic block.\n"); | |
2070 | ||
2071 | destroy_bb_vec_info (bb_vinfo); | |
2072 | return NULL; | |
2073 | } | |
b8698a0f | 2074 | |
a70d6342 IR |
2075 | slp_instances = BB_VINFO_SLP_INSTANCES (bb_vinfo); |
2076 | ||
2077 | /* Mark all the statements that we want to vectorize as pure SLP and | |
2078 | relevant. */ | |
ac47786e | 2079 | FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance) |
a70d6342 IR |
2080 | { |
2081 | vect_mark_slp_stmts (SLP_INSTANCE_TREE (instance), pure_slp, -1); | |
2082 | vect_mark_slp_stmts_relevant (SLP_INSTANCE_TREE (instance)); | |
b8698a0f | 2083 | } |
a70d6342 IR |
2084 | |
2085 | if (!vect_slp_analyze_operations (bb_vinfo)) | |
2086 | { | |
2087 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
2088 | fprintf (vect_dump, "not vectorized: bad operation in basic block.\n"); | |
2089 | ||
2090 | destroy_bb_vec_info (bb_vinfo); | |
2091 | return NULL; | |
2092 | } | |
2093 | ||
69f11a13 IR |
2094 | /* Cost model: check if the vectorization is worthwhile. */ |
2095 | if (flag_vect_cost_model | |
2096 | && !vect_bb_vectorization_profitable_p (bb_vinfo)) | |
2097 | { | |
2098 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
2099 | fprintf (vect_dump, "not vectorized: vectorization is not " | |
2100 | "profitable.\n"); | |
2101 | ||
2102 | destroy_bb_vec_info (bb_vinfo); | |
2103 | return NULL; | |
2104 | } | |
2105 | ||
a70d6342 | 2106 | if (vect_print_dump_info (REPORT_DETAILS)) |
e9dbe7bb | 2107 | fprintf (vect_dump, "Basic block will be vectorized using SLP\n"); |
a70d6342 IR |
2108 | |
2109 | return bb_vinfo; | |
2110 | } | |
2111 | ||
2112 | ||
8e19f5a1 IR |
2113 | bb_vec_info |
2114 | vect_slp_analyze_bb (basic_block bb) | |
2115 | { | |
2116 | bb_vec_info bb_vinfo; | |
2117 | int insns = 0; | |
2118 | gimple_stmt_iterator gsi; | |
2119 | unsigned int vector_sizes; | |
2120 | ||
2121 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2122 | fprintf (vect_dump, "===vect_slp_analyze_bb===\n"); | |
2123 | ||
2124 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2125 | { | |
2126 | gimple stmt = gsi_stmt (gsi); | |
2127 | if (!is_gimple_debug (stmt) | |
2128 | && !gimple_nop_p (stmt) | |
2129 | && gimple_code (stmt) != GIMPLE_LABEL) | |
2130 | insns++; | |
2131 | } | |
2132 | ||
2133 | if (insns > PARAM_VALUE (PARAM_SLP_MAX_INSNS_IN_BB)) | |
2134 | { | |
2135 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
2136 | fprintf (vect_dump, "not vectorized: too many instructions in basic " | |
2137 | "block.\n"); | |
2138 | ||
2139 | return NULL; | |
2140 | } | |
2141 | ||
2142 | /* Autodetect first vector size we try. */ | |
2143 | current_vector_size = 0; | |
2144 | vector_sizes = targetm.vectorize.autovectorize_vector_sizes (); | |
2145 | ||
2146 | while (1) | |
2147 | { | |
2148 | bb_vinfo = vect_slp_analyze_bb_1 (bb); | |
2149 | if (bb_vinfo) | |
2150 | return bb_vinfo; | |
2151 | ||
2152 | destroy_bb_vec_info (bb_vinfo); | |
2153 | ||
2154 | vector_sizes &= ~current_vector_size; | |
2155 | if (vector_sizes == 0 | |
2156 | || current_vector_size == 0) | |
2157 | return NULL; | |
2158 | ||
2159 | /* Try the next biggest vector size. */ | |
2160 | current_vector_size = 1 << floor_log2 (vector_sizes); | |
2161 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2162 | fprintf (vect_dump, "***** Re-trying analysis with " | |
2163 | "vector size %d\n", current_vector_size); | |
2164 | } | |
2165 | } | |
2166 | ||
2167 | ||
b8698a0f | 2168 | /* SLP costs are calculated according to SLP instance unrolling factor (i.e., |
ff802fa1 IR |
2169 | the number of created vector stmts depends on the unrolling factor). |
2170 | However, the actual number of vector stmts for every SLP node depends on | |
2171 | VF which is set later in vect_analyze_operations (). Hence, SLP costs | |
2172 | should be updated. In this function we assume that the inside costs | |
2173 | calculated in vect_model_xxx_cost are linear in ncopies. */ | |
ebfd146a IR |
2174 | |
2175 | void | |
2176 | vect_update_slp_costs_according_to_vf (loop_vec_info loop_vinfo) | |
2177 | { | |
2178 | unsigned int i, vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
2179 | VEC (slp_instance, heap) *slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo); | |
2180 | slp_instance instance; | |
2181 | ||
2182 | if (vect_print_dump_info (REPORT_SLP)) | |
2183 | fprintf (vect_dump, "=== vect_update_slp_costs_according_to_vf ==="); | |
2184 | ||
ac47786e | 2185 | FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance) |
ebfd146a | 2186 | /* We assume that costs are linear in ncopies. */ |
b8698a0f L |
2187 | SLP_INSTANCE_INSIDE_OF_LOOP_COST (instance) *= vf |
2188 | / SLP_INSTANCE_UNROLLING_FACTOR (instance); | |
ebfd146a IR |
2189 | } |
2190 | ||
a70d6342 | 2191 | |
b8698a0f L |
2192 | /* For constant and loop invariant defs of SLP_NODE this function returns |
2193 | (vector) defs (VEC_OPRNDS) that will be used in the vectorized stmts. | |
d59dc888 IR |
2194 | OP_NUM determines if we gather defs for operand 0 or operand 1 of the RHS of |
2195 | scalar stmts. NUMBER_OF_VECTORS is the number of vector defs to create. | |
b5aeb3bb IR |
2196 | REDUC_INDEX is the index of the reduction operand in the statements, unless |
2197 | it is -1. */ | |
ebfd146a IR |
2198 | |
2199 | static void | |
9dc3f7de IR |
2200 | vect_get_constant_vectors (tree op, slp_tree slp_node, |
2201 | VEC (tree, heap) **vec_oprnds, | |
b5aeb3bb IR |
2202 | unsigned int op_num, unsigned int number_of_vectors, |
2203 | int reduc_index) | |
ebfd146a IR |
2204 | { |
2205 | VEC (gimple, heap) *stmts = SLP_TREE_SCALAR_STMTS (slp_node); | |
2206 | gimple stmt = VEC_index (gimple, stmts, 0); | |
2207 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
d2a12ae7 | 2208 | unsigned nunits; |
ebfd146a | 2209 | tree vec_cst; |
d2a12ae7 RG |
2210 | tree *elts; |
2211 | unsigned j, number_of_places_left_in_vector; | |
ebfd146a | 2212 | tree vector_type; |
9dc3f7de | 2213 | tree vop; |
ebfd146a IR |
2214 | int group_size = VEC_length (gimple, stmts); |
2215 | unsigned int vec_num, i; | |
d2a12ae7 | 2216 | unsigned number_of_copies = 1; |
ebfd146a IR |
2217 | VEC (tree, heap) *voprnds = VEC_alloc (tree, heap, number_of_vectors); |
2218 | bool constant_p, is_store; | |
b5aeb3bb | 2219 | tree neutral_op = NULL; |
bac430c9 | 2220 | enum tree_code code = gimple_expr_code (stmt); |
0e93a64e IR |
2221 | gimple def_stmt; |
2222 | struct loop *loop; | |
b5aeb3bb | 2223 | |
29ed4920 IR |
2224 | if (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
2225 | && reduc_index != -1) | |
b5aeb3bb | 2226 | { |
b5aeb3bb | 2227 | op_num = reduc_index - 1; |
9dc3f7de | 2228 | op = gimple_op (stmt, reduc_index); |
b5aeb3bb | 2229 | /* For additional copies (see the explanation of NUMBER_OF_COPIES below) |
ff802fa1 | 2230 | we need either neutral operands or the original operands. See |
b5aeb3bb IR |
2231 | get_initial_def_for_reduction() for details. */ |
2232 | switch (code) | |
2233 | { | |
2234 | case WIDEN_SUM_EXPR: | |
2235 | case DOT_PROD_EXPR: | |
2236 | case PLUS_EXPR: | |
2237 | case MINUS_EXPR: | |
2238 | case BIT_IOR_EXPR: | |
2239 | case BIT_XOR_EXPR: | |
2240 | if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (op))) | |
2241 | neutral_op = build_real (TREE_TYPE (op), dconst0); | |
2242 | else | |
2243 | neutral_op = build_int_cst (TREE_TYPE (op), 0); | |
2244 | ||
2245 | break; | |
2246 | ||
2247 | case MULT_EXPR: | |
b5aeb3bb IR |
2248 | if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (op))) |
2249 | neutral_op = build_real (TREE_TYPE (op), dconst1); | |
2250 | else | |
2251 | neutral_op = build_int_cst (TREE_TYPE (op), 1); | |
2252 | ||
2253 | break; | |
2254 | ||
c1e822d5 IR |
2255 | case BIT_AND_EXPR: |
2256 | neutral_op = build_int_cst (TREE_TYPE (op), -1); | |
2257 | break; | |
2258 | ||
0e93a64e IR |
2259 | case MAX_EXPR: |
2260 | case MIN_EXPR: | |
2261 | def_stmt = SSA_NAME_DEF_STMT (op); | |
2262 | loop = (gimple_bb (stmt))->loop_father; | |
2263 | neutral_op = PHI_ARG_DEF_FROM_EDGE (def_stmt, | |
2264 | loop_preheader_edge (loop)); | |
2265 | break; | |
2266 | ||
b5aeb3bb | 2267 | default: |
0e93a64e | 2268 | neutral_op = NULL; |
b5aeb3bb IR |
2269 | } |
2270 | } | |
ebfd146a IR |
2271 | |
2272 | if (STMT_VINFO_DATA_REF (stmt_vinfo)) | |
2273 | { | |
2274 | is_store = true; | |
2275 | op = gimple_assign_rhs1 (stmt); | |
2276 | } | |
2277 | else | |
9dc3f7de IR |
2278 | is_store = false; |
2279 | ||
2280 | gcc_assert (op); | |
ebfd146a IR |
2281 | |
2282 | if (CONSTANT_CLASS_P (op)) | |
d59dc888 | 2283 | constant_p = true; |
ebfd146a | 2284 | else |
d59dc888 IR |
2285 | constant_p = false; |
2286 | ||
9dc3f7de | 2287 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
cd481d83 | 2288 | gcc_assert (vector_type); |
ebfd146a IR |
2289 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); |
2290 | ||
2291 | /* NUMBER_OF_COPIES is the number of times we need to use the same values in | |
b8698a0f | 2292 | created vectors. It is greater than 1 if unrolling is performed. |
ebfd146a IR |
2293 | |
2294 | For example, we have two scalar operands, s1 and s2 (e.g., group of | |
2295 | strided accesses of size two), while NUNITS is four (i.e., four scalars | |
f7e531cf IR |
2296 | of this type can be packed in a vector). The output vector will contain |
2297 | two copies of each scalar operand: {s1, s2, s1, s2}. (NUMBER_OF_COPIES | |
ebfd146a IR |
2298 | will be 2). |
2299 | ||
b8698a0f | 2300 | If GROUP_SIZE > NUNITS, the scalars will be split into several vectors |
ebfd146a IR |
2301 | containing the operands. |
2302 | ||
2303 | For example, NUNITS is four as before, and the group size is 8 | |
f7e531cf | 2304 | (s1, s2, ..., s8). We will create two vectors {s1, s2, s3, s4} and |
ebfd146a | 2305 | {s5, s6, s7, s8}. */ |
b8698a0f | 2306 | |
ebfd146a IR |
2307 | number_of_copies = least_common_multiple (nunits, group_size) / group_size; |
2308 | ||
2309 | number_of_places_left_in_vector = nunits; | |
d2a12ae7 | 2310 | elts = XALLOCAVEC (tree, nunits); |
ebfd146a IR |
2311 | for (j = 0; j < number_of_copies; j++) |
2312 | { | |
2313 | for (i = group_size - 1; VEC_iterate (gimple, stmts, i, stmt); i--) | |
2314 | { | |
2315 | if (is_store) | |
2316 | op = gimple_assign_rhs1 (stmt); | |
bac430c9 | 2317 | else |
f7e531cf | 2318 | { |
bac430c9 | 2319 | switch (code) |
f7e531cf | 2320 | { |
bac430c9 IR |
2321 | case COND_EXPR: |
2322 | if (op_num == 0 || op_num == 1) | |
2323 | { | |
2324 | tree cond = gimple_assign_rhs1 (stmt); | |
2325 | op = TREE_OPERAND (cond, op_num); | |
2326 | } | |
2327 | else | |
2328 | { | |
2329 | if (op_num == 2) | |
2330 | op = gimple_assign_rhs2 (stmt); | |
2331 | else | |
2332 | op = gimple_assign_rhs3 (stmt); | |
2333 | } | |
2334 | break; | |
2335 | ||
2336 | case CALL_EXPR: | |
2337 | op = gimple_call_arg (stmt, op_num); | |
2338 | break; | |
2339 | ||
2340 | default: | |
2341 | op = gimple_op (stmt, op_num + 1); | |
f7e531cf IR |
2342 | } |
2343 | } | |
b8698a0f | 2344 | |
b5aeb3bb IR |
2345 | if (reduc_index != -1) |
2346 | { | |
0e93a64e IR |
2347 | loop = (gimple_bb (stmt))->loop_father; |
2348 | def_stmt = SSA_NAME_DEF_STMT (op); | |
b5aeb3bb IR |
2349 | |
2350 | gcc_assert (loop); | |
b010117a IR |
2351 | |
2352 | /* Get the def before the loop. In reduction chain we have only | |
2353 | one initial value. */ | |
2354 | if ((j != (number_of_copies - 1) | |
2355 | || (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) | |
2356 | && i != 0)) | |
2357 | && neutral_op) | |
b5aeb3bb | 2358 | op = neutral_op; |
b010117a IR |
2359 | else |
2360 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, | |
2361 | loop_preheader_edge (loop)); | |
b5aeb3bb IR |
2362 | } |
2363 | ||
ebfd146a | 2364 | /* Create 'vect_ = {op0,op1,...,opn}'. */ |
ebfd146a | 2365 | number_of_places_left_in_vector--; |
50eeef09 RG |
2366 | if (constant_p |
2367 | && !types_compatible_p (TREE_TYPE (vector_type), TREE_TYPE (op))) | |
2368 | { | |
2369 | op = fold_unary (VIEW_CONVERT_EXPR, TREE_TYPE (vector_type), op); | |
2370 | gcc_assert (op && CONSTANT_CLASS_P (op)); | |
2371 | } | |
d2a12ae7 | 2372 | elts[number_of_places_left_in_vector] = op; |
ebfd146a IR |
2373 | |
2374 | if (number_of_places_left_in_vector == 0) | |
2375 | { | |
2376 | number_of_places_left_in_vector = nunits; | |
2377 | ||
2378 | if (constant_p) | |
d2a12ae7 | 2379 | vec_cst = build_vector (vector_type, elts); |
ebfd146a | 2380 | else |
d2a12ae7 RG |
2381 | { |
2382 | VEC(constructor_elt,gc) *v; | |
2383 | unsigned k; | |
2384 | v = VEC_alloc (constructor_elt, gc, nunits); | |
2385 | for (k = 0; k < nunits; ++k) | |
2386 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, elts[k]); | |
2387 | vec_cst = build_constructor (vector_type, v); | |
2388 | } | |
ebfd146a | 2389 | VEC_quick_push (tree, voprnds, |
d2a12ae7 RG |
2390 | vect_init_vector (stmt, vec_cst, |
2391 | vector_type, NULL)); | |
ebfd146a IR |
2392 | } |
2393 | } | |
2394 | } | |
2395 | ||
b8698a0f | 2396 | /* Since the vectors are created in the reverse order, we should invert |
ebfd146a IR |
2397 | them. */ |
2398 | vec_num = VEC_length (tree, voprnds); | |
d2a12ae7 | 2399 | for (j = vec_num; j != 0; j--) |
ebfd146a | 2400 | { |
d2a12ae7 | 2401 | vop = VEC_index (tree, voprnds, j - 1); |
ebfd146a IR |
2402 | VEC_quick_push (tree, *vec_oprnds, vop); |
2403 | } | |
2404 | ||
2405 | VEC_free (tree, heap, voprnds); | |
2406 | ||
2407 | /* In case that VF is greater than the unrolling factor needed for the SLP | |
b8698a0f L |
2408 | group of stmts, NUMBER_OF_VECTORS to be created is greater than |
2409 | NUMBER_OF_SCALARS/NUNITS or NUNITS/NUMBER_OF_SCALARS, and hence we have | |
ebfd146a IR |
2410 | to replicate the vectors. */ |
2411 | while (number_of_vectors > VEC_length (tree, *vec_oprnds)) | |
2412 | { | |
b5aeb3bb IR |
2413 | tree neutral_vec = NULL; |
2414 | ||
2415 | if (neutral_op) | |
2416 | { | |
2417 | if (!neutral_vec) | |
b9acc9f1 | 2418 | neutral_vec = build_vector_from_val (vector_type, neutral_op); |
b5aeb3bb IR |
2419 | |
2420 | VEC_quick_push (tree, *vec_oprnds, neutral_vec); | |
2421 | } | |
2422 | else | |
2423 | { | |
2424 | for (i = 0; VEC_iterate (tree, *vec_oprnds, i, vop) && i < vec_num; i++) | |
2425 | VEC_quick_push (tree, *vec_oprnds, vop); | |
2426 | } | |
ebfd146a IR |
2427 | } |
2428 | } | |
2429 | ||
2430 | ||
2431 | /* Get vectorized definitions from SLP_NODE that contains corresponding | |
2432 | vectorized def-stmts. */ | |
2433 | ||
2434 | static void | |
2435 | vect_get_slp_vect_defs (slp_tree slp_node, VEC (tree,heap) **vec_oprnds) | |
2436 | { | |
2437 | tree vec_oprnd; | |
2438 | gimple vec_def_stmt; | |
2439 | unsigned int i; | |
2440 | ||
2441 | gcc_assert (SLP_TREE_VEC_STMTS (slp_node)); | |
2442 | ||
ac47786e | 2443 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_VEC_STMTS (slp_node), i, vec_def_stmt) |
ebfd146a IR |
2444 | { |
2445 | gcc_assert (vec_def_stmt); | |
2446 | vec_oprnd = gimple_get_lhs (vec_def_stmt); | |
2447 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
2448 | } | |
2449 | } | |
2450 | ||
2451 | ||
b8698a0f L |
2452 | /* Get vectorized definitions for SLP_NODE. |
2453 | If the scalar definitions are loop invariants or constants, collect them and | |
ebfd146a IR |
2454 | call vect_get_constant_vectors() to create vector stmts. |
2455 | Otherwise, the def-stmts must be already vectorized and the vectorized stmts | |
d092494c IR |
2456 | must be stored in the corresponding child of SLP_NODE, and we call |
2457 | vect_get_slp_vect_defs () to retrieve them. */ | |
b8698a0f | 2458 | |
ebfd146a | 2459 | void |
d092494c IR |
2460 | vect_get_slp_defs (VEC (tree, heap) *ops, slp_tree slp_node, |
2461 | VEC (slp_void_p, heap) **vec_oprnds, int reduc_index) | |
ebfd146a | 2462 | { |
d092494c IR |
2463 | gimple first_stmt, first_def; |
2464 | int number_of_vects = 0, i; | |
2465 | unsigned int child_index = 0; | |
b8698a0f | 2466 | HOST_WIDE_INT lhs_size_unit, rhs_size_unit; |
d092494c IR |
2467 | slp_tree child = NULL; |
2468 | VEC (tree, heap) *vec_defs; | |
2469 | tree oprnd, def_lhs; | |
2470 | bool vectorized_defs; | |
ebfd146a IR |
2471 | |
2472 | first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0); | |
d092494c | 2473 | FOR_EACH_VEC_ELT (tree, ops, i, oprnd) |
ebfd146a | 2474 | { |
d092494c IR |
2475 | /* For each operand we check if it has vectorized definitions in a child |
2476 | node or we need to create them (for invariants and constants). We | |
2477 | check if the LHS of the first stmt of the next child matches OPRND. | |
2478 | If it does, we found the correct child. Otherwise, we call | |
2479 | vect_get_constant_vectors (), and not advance CHILD_INDEX in order | |
2480 | to check this child node for the next operand. */ | |
2481 | vectorized_defs = false; | |
2482 | if (VEC_length (slp_void_p, SLP_TREE_CHILDREN (slp_node)) > child_index) | |
ebfd146a | 2483 | { |
d092494c IR |
2484 | child = (slp_tree) VEC_index (slp_void_p, |
2485 | SLP_TREE_CHILDREN (slp_node), | |
2486 | child_index); | |
2487 | first_def = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (child), 0); | |
2488 | ||
2489 | /* In the end of a pattern sequence we have a use of the original stmt, | |
2490 | so we need to compare OPRND with the original def. */ | |
2491 | if (is_pattern_stmt_p (vinfo_for_stmt (first_def)) | |
2492 | && !STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (first_stmt)) | |
2493 | && !is_pattern_stmt_p (vinfo_for_stmt (first_stmt))) | |
2494 | first_def = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (first_def)); | |
2495 | ||
2496 | if (is_gimple_call (first_def)) | |
2497 | def_lhs = gimple_call_lhs (first_def); | |
2498 | else | |
2499 | def_lhs = gimple_assign_lhs (first_def); | |
ebfd146a | 2500 | |
d092494c IR |
2501 | if (operand_equal_p (oprnd, def_lhs, 0)) |
2502 | { | |
2503 | /* The number of vector defs is determined by the number of | |
2504 | vector statements in the node from which we get those | |
2505 | statements. */ | |
2506 | number_of_vects = SLP_TREE_NUMBER_OF_VEC_STMTS (child); | |
2507 | vectorized_defs = true; | |
2508 | child_index++; | |
2509 | } | |
2510 | } | |
ebfd146a | 2511 | |
d092494c IR |
2512 | if (!vectorized_defs) |
2513 | { | |
2514 | if (i == 0) | |
2515 | { | |
2516 | number_of_vects = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
2517 | /* Number of vector stmts was calculated according to LHS in | |
2518 | vect_schedule_slp_instance (), fix it by replacing LHS with | |
2519 | RHS, if necessary. See vect_get_smallest_scalar_type () for | |
2520 | details. */ | |
2521 | vect_get_smallest_scalar_type (first_stmt, &lhs_size_unit, | |
2522 | &rhs_size_unit); | |
2523 | if (rhs_size_unit != lhs_size_unit) | |
2524 | { | |
2525 | number_of_vects *= rhs_size_unit; | |
2526 | number_of_vects /= lhs_size_unit; | |
2527 | } | |
2528 | } | |
2529 | } | |
b5aeb3bb | 2530 | |
d092494c IR |
2531 | /* Allocate memory for vectorized defs. */ |
2532 | vec_defs = VEC_alloc (tree, heap, number_of_vects); | |
ebfd146a | 2533 | |
d092494c IR |
2534 | /* For reduction defs we call vect_get_constant_vectors (), since we are |
2535 | looking for initial loop invariant values. */ | |
2536 | if (vectorized_defs && reduc_index == -1) | |
2537 | /* The defs are already vectorized. */ | |
2538 | vect_get_slp_vect_defs (child, &vec_defs); | |
2539 | else | |
2540 | /* Build vectors from scalar defs. */ | |
2541 | vect_get_constant_vectors (oprnd, slp_node, &vec_defs, i, | |
2542 | number_of_vects, reduc_index); | |
ebfd146a | 2543 | |
d092494c | 2544 | VEC_quick_push (slp_void_p, *vec_oprnds, (slp_void_p) vec_defs); |
ebfd146a | 2545 | |
d092494c IR |
2546 | /* For reductions, we only need initial values. */ |
2547 | if (reduc_index != -1) | |
2548 | return; | |
2549 | } | |
ebfd146a IR |
2550 | } |
2551 | ||
a70d6342 | 2552 | |
b8698a0f | 2553 | /* Create NCOPIES permutation statements using the mask MASK_BYTES (by |
ebfd146a IR |
2554 | building a vector of type MASK_TYPE from it) and two input vectors placed in |
2555 | DR_CHAIN at FIRST_VEC_INDX and SECOND_VEC_INDX for the first copy and | |
2556 | shifting by STRIDE elements of DR_CHAIN for every copy. | |
2557 | (STRIDE is the number of vectorized stmts for NODE divided by the number of | |
b8698a0f | 2558 | copies). |
ebfd146a IR |
2559 | VECT_STMTS_COUNTER specifies the index in the vectorized stmts of NODE, where |
2560 | the created stmts must be inserted. */ | |
2561 | ||
2562 | static inline void | |
b8698a0f | 2563 | vect_create_mask_and_perm (gimple stmt, gimple next_scalar_stmt, |
faf63e39 | 2564 | tree mask, int first_vec_indx, int second_vec_indx, |
b8698a0f | 2565 | gimple_stmt_iterator *gsi, slp_tree node, |
2635892a | 2566 | tree vectype, VEC(tree,heap) *dr_chain, |
ebfd146a IR |
2567 | int ncopies, int vect_stmts_counter) |
2568 | { | |
faf63e39 | 2569 | tree perm_dest; |
ebfd146a IR |
2570 | gimple perm_stmt = NULL; |
2571 | stmt_vec_info next_stmt_info; | |
0f900dfa | 2572 | int i, stride; |
ebfd146a | 2573 | tree first_vec, second_vec, data_ref; |
ebfd146a | 2574 | |
ebfd146a | 2575 | stride = SLP_TREE_NUMBER_OF_VEC_STMTS (node) / ncopies; |
ebfd146a | 2576 | |
b8698a0f | 2577 | /* Initialize the vect stmts of NODE to properly insert the generated |
ebfd146a | 2578 | stmts later. */ |
b8698a0f | 2579 | for (i = VEC_length (gimple, SLP_TREE_VEC_STMTS (node)); |
ebfd146a IR |
2580 | i < (int) SLP_TREE_NUMBER_OF_VEC_STMTS (node); i++) |
2581 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (node), NULL); | |
2582 | ||
2583 | perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype); | |
2584 | for (i = 0; i < ncopies; i++) | |
2585 | { | |
2586 | first_vec = VEC_index (tree, dr_chain, first_vec_indx); | |
2587 | second_vec = VEC_index (tree, dr_chain, second_vec_indx); | |
2588 | ||
ebfd146a | 2589 | /* Generate the permute statement. */ |
2635892a RH |
2590 | perm_stmt = gimple_build_assign_with_ops3 (VEC_PERM_EXPR, perm_dest, |
2591 | first_vec, second_vec, mask); | |
ebfd146a | 2592 | data_ref = make_ssa_name (perm_dest, perm_stmt); |
2635892a | 2593 | gimple_set_lhs (perm_stmt, data_ref); |
ebfd146a | 2594 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
ebfd146a | 2595 | |
b8698a0f L |
2596 | /* Store the vector statement in NODE. */ |
2597 | VEC_replace (gimple, SLP_TREE_VEC_STMTS (node), | |
ebfd146a IR |
2598 | stride * i + vect_stmts_counter, perm_stmt); |
2599 | ||
2600 | first_vec_indx += stride; | |
2601 | second_vec_indx += stride; | |
2602 | } | |
2603 | ||
2604 | /* Mark the scalar stmt as vectorized. */ | |
2605 | next_stmt_info = vinfo_for_stmt (next_scalar_stmt); | |
2606 | STMT_VINFO_VEC_STMT (next_stmt_info) = perm_stmt; | |
2607 | } | |
2608 | ||
2609 | ||
b8698a0f | 2610 | /* Given FIRST_MASK_ELEMENT - the mask element in element representation, |
ebfd146a | 2611 | return in CURRENT_MASK_ELEMENT its equivalent in target specific |
ff802fa1 | 2612 | representation. Check that the mask is valid and return FALSE if not. |
ebfd146a IR |
2613 | Return TRUE in NEED_NEXT_VECTOR if the permutation requires to move to |
2614 | the next vector, i.e., the current first vector is not needed. */ | |
b8698a0f | 2615 | |
ebfd146a | 2616 | static bool |
b8698a0f | 2617 | vect_get_mask_element (gimple stmt, int first_mask_element, int m, |
ebfd146a | 2618 | int mask_nunits, bool only_one_vec, int index, |
22e4dee7 | 2619 | unsigned char *mask, int *current_mask_element, |
694a4f61 IR |
2620 | bool *need_next_vector, int *number_of_mask_fixes, |
2621 | bool *mask_fixed, bool *needs_first_vector) | |
ebfd146a IR |
2622 | { |
2623 | int i; | |
ebfd146a IR |
2624 | |
2625 | /* Convert to target specific representation. */ | |
2626 | *current_mask_element = first_mask_element + m; | |
2627 | /* Adjust the value in case it's a mask for second and third vectors. */ | |
694a4f61 | 2628 | *current_mask_element -= mask_nunits * (*number_of_mask_fixes - 1); |
ebfd146a IR |
2629 | |
2630 | if (*current_mask_element < mask_nunits) | |
694a4f61 | 2631 | *needs_first_vector = true; |
ebfd146a IR |
2632 | |
2633 | /* We have only one input vector to permute but the mask accesses values in | |
2634 | the next vector as well. */ | |
2635 | if (only_one_vec && *current_mask_element >= mask_nunits) | |
2636 | { | |
2637 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2638 | { | |
2639 | fprintf (vect_dump, "permutation requires at least two vectors "); | |
2640 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
2641 | } | |
2642 | ||
2643 | return false; | |
2644 | } | |
2645 | ||
2646 | /* The mask requires the next vector. */ | |
2647 | if (*current_mask_element >= mask_nunits * 2) | |
2648 | { | |
694a4f61 | 2649 | if (*needs_first_vector || *mask_fixed) |
ebfd146a IR |
2650 | { |
2651 | /* We either need the first vector too or have already moved to the | |
b8698a0f | 2652 | next vector. In both cases, this permutation needs three |
ebfd146a IR |
2653 | vectors. */ |
2654 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2655 | { | |
2656 | fprintf (vect_dump, "permutation requires at " | |
2657 | "least three vectors "); | |
2658 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
2659 | } | |
2660 | ||
2661 | return false; | |
2662 | } | |
2663 | ||
2664 | /* We move to the next vector, dropping the first one and working with | |
2665 | the second and the third - we need to adjust the values of the mask | |
2666 | accordingly. */ | |
694a4f61 | 2667 | *current_mask_element -= mask_nunits * *number_of_mask_fixes; |
ebfd146a IR |
2668 | |
2669 | for (i = 0; i < index; i++) | |
694a4f61 | 2670 | mask[i] -= mask_nunits * *number_of_mask_fixes; |
ebfd146a | 2671 | |
694a4f61 IR |
2672 | (*number_of_mask_fixes)++; |
2673 | *mask_fixed = true; | |
ebfd146a IR |
2674 | } |
2675 | ||
694a4f61 | 2676 | *need_next_vector = *mask_fixed; |
ebfd146a IR |
2677 | |
2678 | /* This was the last element of this mask. Start a new one. */ | |
2679 | if (index == mask_nunits - 1) | |
2680 | { | |
694a4f61 IR |
2681 | *number_of_mask_fixes = 1; |
2682 | *mask_fixed = false; | |
2683 | *needs_first_vector = false; | |
ebfd146a IR |
2684 | } |
2685 | ||
2686 | return true; | |
2687 | } | |
2688 | ||
2689 | ||
2690 | /* Generate vector permute statements from a list of loads in DR_CHAIN. | |
2691 | If ANALYZE_ONLY is TRUE, only check that it is possible to create valid | |
2692 | permute statements for SLP_NODE_INSTANCE. */ | |
2693 | bool | |
2694 | vect_transform_slp_perm_load (gimple stmt, VEC (tree, heap) *dr_chain, | |
2695 | gimple_stmt_iterator *gsi, int vf, | |
2696 | slp_instance slp_node_instance, bool analyze_only) | |
2697 | { | |
2698 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2699 | tree mask_element_type = NULL_TREE, mask_type; | |
2635892a | 2700 | int i, j, k, nunits, vec_index = 0, scalar_index; |
ebfd146a | 2701 | slp_tree node; |
2635892a | 2702 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a IR |
2703 | gimple next_scalar_stmt; |
2704 | int group_size = SLP_INSTANCE_GROUP_SIZE (slp_node_instance); | |
2705 | int first_mask_element; | |
22e4dee7 RH |
2706 | int index, unroll_factor, current_mask_element, ncopies; |
2707 | unsigned char *mask; | |
ebfd146a IR |
2708 | bool only_one_vec = false, need_next_vector = false; |
2709 | int first_vec_index, second_vec_index, orig_vec_stmts_num, vect_stmts_counter; | |
694a4f61 IR |
2710 | int number_of_mask_fixes = 1; |
2711 | bool mask_fixed = false; | |
2712 | bool needs_first_vector = false; | |
22e4dee7 | 2713 | enum machine_mode mode; |
ebfd146a | 2714 | |
22e4dee7 RH |
2715 | mode = TYPE_MODE (vectype); |
2716 | ||
2717 | if (!can_vec_perm_p (mode, false, NULL)) | |
ebfd146a IR |
2718 | { |
2719 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2720 | { | |
2635892a | 2721 | fprintf (vect_dump, "no vect permute for "); |
ebfd146a IR |
2722 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); |
2723 | } | |
2635892a | 2724 | return false; |
ebfd146a IR |
2725 | } |
2726 | ||
2635892a RH |
2727 | /* The generic VEC_PERM_EXPR code always uses an integral type of the |
2728 | same size as the vector element being permuted. */ | |
96f9265a RG |
2729 | mask_element_type = lang_hooks.types.type_for_mode |
2730 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
ebfd146a | 2731 | mask_type = get_vectype_for_scalar_type (mask_element_type); |
ebfd146a | 2732 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 | 2733 | mask = XALLOCAVEC (unsigned char, nunits); |
ebfd146a IR |
2734 | unroll_factor = SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance); |
2735 | ||
2736 | /* The number of vector stmts to generate based only on SLP_NODE_INSTANCE | |
2737 | unrolling factor. */ | |
b8698a0f | 2738 | orig_vec_stmts_num = group_size * |
ebfd146a IR |
2739 | SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance) / nunits; |
2740 | if (orig_vec_stmts_num == 1) | |
2741 | only_one_vec = true; | |
2742 | ||
b8698a0f | 2743 | /* Number of copies is determined by the final vectorization factor |
ebfd146a | 2744 | relatively to SLP_NODE_INSTANCE unrolling factor. */ |
b8698a0f | 2745 | ncopies = vf / SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance); |
ebfd146a | 2746 | |
b8698a0f L |
2747 | /* Generate permutation masks for every NODE. Number of masks for each NODE |
2748 | is equal to GROUP_SIZE. | |
2749 | E.g., we have a group of three nodes with three loads from the same | |
2750 | location in each node, and the vector size is 4. I.e., we have a | |
2751 | a0b0c0a1b1c1... sequence and we need to create the following vectors: | |
ebfd146a IR |
2752 | for a's: a0a0a0a1 a1a1a2a2 a2a3a3a3 |
2753 | for b's: b0b0b0b1 b1b1b2b2 b2b3b3b3 | |
2754 | ... | |
2755 | ||
2635892a | 2756 | The masks for a's should be: {0,0,0,3} {3,3,6,6} {6,9,9,9}. |
b8698a0f | 2757 | The last mask is illegal since we assume two operands for permute |
ff802fa1 IR |
2758 | operation, and the mask element values can't be outside that range. |
2759 | Hence, the last mask must be converted into {2,5,5,5}. | |
b8698a0f | 2760 | For the first two permutations we need the first and the second input |
ebfd146a | 2761 | vectors: {a0,b0,c0,a1} and {b1,c1,a2,b2}, and for the last permutation |
b8698a0f | 2762 | we need the second and the third vectors: {b1,c1,a2,b2} and |
ebfd146a IR |
2763 | {c2,a3,b3,c3}. */ |
2764 | ||
ac47786e | 2765 | FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (slp_node_instance), i, node) |
ebfd146a IR |
2766 | { |
2767 | scalar_index = 0; | |
2768 | index = 0; | |
2769 | vect_stmts_counter = 0; | |
2770 | vec_index = 0; | |
2771 | first_vec_index = vec_index++; | |
2772 | if (only_one_vec) | |
2773 | second_vec_index = first_vec_index; | |
2774 | else | |
2775 | second_vec_index = vec_index++; | |
2776 | ||
2777 | for (j = 0; j < unroll_factor; j++) | |
2778 | { | |
2779 | for (k = 0; k < group_size; k++) | |
2780 | { | |
2635892a RH |
2781 | first_mask_element = i + j * group_size; |
2782 | if (!vect_get_mask_element (stmt, first_mask_element, 0, | |
2783 | nunits, only_one_vec, index, | |
2784 | mask, ¤t_mask_element, | |
2785 | &need_next_vector, | |
2786 | &number_of_mask_fixes, &mask_fixed, | |
2787 | &needs_first_vector)) | |
2788 | return false; | |
2789 | mask[index++] = current_mask_element; | |
ebfd146a | 2790 | |
2635892a | 2791 | if (index == nunits) |
ebfd146a | 2792 | { |
d2a12ae7 RG |
2793 | tree mask_vec, *mask_elts; |
2794 | int l; | |
faf63e39 | 2795 | |
22e4dee7 RH |
2796 | if (!can_vec_perm_p (mode, false, mask)) |
2797 | { | |
2798 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2799 | { | |
2800 | fprintf (vect_dump, "unsupported vect permute { "); | |
2801 | for (i = 0; i < nunits; ++i) | |
2802 | fprintf (vect_dump, "%d ", mask[i]); | |
2803 | fprintf (vect_dump, "}\n"); | |
2804 | } | |
2805 | return false; | |
2806 | } | |
2807 | ||
d2a12ae7 RG |
2808 | mask_elts = XALLOCAVEC (tree, nunits); |
2809 | for (l = 0; l < nunits; ++l) | |
2810 | mask_elts[l] = build_int_cst (mask_element_type, mask[l]); | |
2811 | mask_vec = build_vector (mask_type, mask_elts); | |
faf63e39 RH |
2812 | index = 0; |
2813 | ||
ebfd146a IR |
2814 | if (!analyze_only) |
2815 | { | |
2816 | if (need_next_vector) | |
2817 | { | |
2818 | first_vec_index = second_vec_index; | |
2819 | second_vec_index = vec_index; | |
2820 | } | |
2821 | ||
2822 | next_scalar_stmt = VEC_index (gimple, | |
2823 | SLP_TREE_SCALAR_STMTS (node), scalar_index++); | |
2824 | ||
2825 | vect_create_mask_and_perm (stmt, next_scalar_stmt, | |
faf63e39 | 2826 | mask_vec, first_vec_index, second_vec_index, |
2635892a | 2827 | gsi, node, vectype, dr_chain, |
faf63e39 | 2828 | ncopies, vect_stmts_counter++); |
ebfd146a | 2829 | } |
b8698a0f L |
2830 | } |
2831 | } | |
2832 | } | |
2833 | } | |
ebfd146a | 2834 | |
ebfd146a IR |
2835 | return true; |
2836 | } | |
2837 | ||
2838 | ||
2839 | ||
2840 | /* Vectorize SLP instance tree in postorder. */ | |
2841 | ||
2842 | static bool | |
2843 | vect_schedule_slp_instance (slp_tree node, slp_instance instance, | |
a70d6342 | 2844 | unsigned int vectorization_factor) |
ebfd146a IR |
2845 | { |
2846 | gimple stmt; | |
2847 | bool strided_store, is_store; | |
2848 | gimple_stmt_iterator si; | |
2849 | stmt_vec_info stmt_info; | |
2850 | unsigned int vec_stmts_size, nunits, group_size; | |
2851 | tree vectype; | |
2852 | int i; | |
2853 | slp_tree loads_node; | |
d092494c | 2854 | slp_void_p child; |
ebfd146a IR |
2855 | |
2856 | if (!node) | |
2857 | return false; | |
2858 | ||
d092494c IR |
2859 | FOR_EACH_VEC_ELT (slp_void_p, SLP_TREE_CHILDREN (node), i, child) |
2860 | vect_schedule_slp_instance ((slp_tree) child, instance, | |
2861 | vectorization_factor); | |
b8698a0f | 2862 | |
ebfd146a IR |
2863 | stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (node), 0); |
2864 | stmt_info = vinfo_for_stmt (stmt); | |
2865 | ||
2866 | /* VECTYPE is the type of the destination. */ | |
b690cc0f | 2867 | vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a IR |
2868 | nunits = (unsigned int) TYPE_VECTOR_SUBPARTS (vectype); |
2869 | group_size = SLP_INSTANCE_GROUP_SIZE (instance); | |
2870 | ||
2871 | /* For each SLP instance calculate number of vector stmts to be created | |
ff802fa1 | 2872 | for the scalar stmts in each node of the SLP tree. Number of vector |
ebfd146a IR |
2873 | elements in one vector iteration is the number of scalar elements in |
2874 | one scalar iteration (GROUP_SIZE) multiplied by VF divided by vector | |
2875 | size. */ | |
2876 | vec_stmts_size = (vectorization_factor * group_size) / nunits; | |
2877 | ||
2878 | /* In case of load permutation we have to allocate vectorized statements for | |
2879 | all the nodes that participate in that permutation. */ | |
2880 | if (SLP_INSTANCE_LOAD_PERMUTATION (instance)) | |
2881 | { | |
ac47786e | 2882 | FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (instance), i, loads_node) |
ebfd146a IR |
2883 | { |
2884 | if (!SLP_TREE_VEC_STMTS (loads_node)) | |
2885 | { | |
2886 | SLP_TREE_VEC_STMTS (loads_node) = VEC_alloc (gimple, heap, | |
2887 | vec_stmts_size); | |
2888 | SLP_TREE_NUMBER_OF_VEC_STMTS (loads_node) = vec_stmts_size; | |
2889 | } | |
2890 | } | |
2891 | } | |
2892 | ||
2893 | if (!SLP_TREE_VEC_STMTS (node)) | |
2894 | { | |
2895 | SLP_TREE_VEC_STMTS (node) = VEC_alloc (gimple, heap, vec_stmts_size); | |
2896 | SLP_TREE_NUMBER_OF_VEC_STMTS (node) = vec_stmts_size; | |
2897 | } | |
2898 | ||
2899 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2900 | { | |
2901 | fprintf (vect_dump, "------>vectorizing SLP node starting from: "); | |
2902 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
b8698a0f | 2903 | } |
ebfd146a IR |
2904 | |
2905 | /* Loads should be inserted before the first load. */ | |
2906 | if (SLP_INSTANCE_FIRST_LOAD_STMT (instance) | |
2907 | && STMT_VINFO_STRIDED_ACCESS (stmt_info) | |
6aa904c4 IR |
2908 | && !REFERENCE_CLASS_P (gimple_get_lhs (stmt)) |
2909 | && SLP_INSTANCE_LOAD_PERMUTATION (instance)) | |
ebfd146a | 2910 | si = gsi_for_stmt (SLP_INSTANCE_FIRST_LOAD_STMT (instance)); |
9d5e7640 | 2911 | else if (is_pattern_stmt_p (stmt_info)) |
6aa904c4 | 2912 | si = gsi_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
ebfd146a IR |
2913 | else |
2914 | si = gsi_for_stmt (stmt); | |
b8698a0f | 2915 | |
e4a707c4 IR |
2916 | /* Stores should be inserted just before the last store. */ |
2917 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) | |
2918 | && REFERENCE_CLASS_P (gimple_get_lhs (stmt))) | |
2919 | { | |
2920 | gimple last_store = vect_find_last_store_in_slp_instance (instance); | |
a024e70e IR |
2921 | if (is_pattern_stmt_p (vinfo_for_stmt (last_store))) |
2922 | last_store = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (last_store)); | |
e4a707c4 IR |
2923 | si = gsi_for_stmt (last_store); |
2924 | } | |
2925 | ||
b010117a IR |
2926 | /* Mark the first element of the reduction chain as reduction to properly |
2927 | transform the node. In the analysis phase only the last element of the | |
2928 | chain is marked as reduction. */ | |
2929 | if (GROUP_FIRST_ELEMENT (stmt_info) && !STMT_VINFO_STRIDED_ACCESS (stmt_info) | |
2930 | && GROUP_FIRST_ELEMENT (stmt_info) == stmt) | |
2931 | { | |
2932 | STMT_VINFO_DEF_TYPE (stmt_info) = vect_reduction_def; | |
2933 | STMT_VINFO_TYPE (stmt_info) = reduc_vec_info_type; | |
2934 | } | |
2935 | ||
ebfd146a | 2936 | is_store = vect_transform_stmt (stmt, &si, &strided_store, node, instance); |
b5aeb3bb | 2937 | return is_store; |
ebfd146a IR |
2938 | } |
2939 | ||
dd34c087 JJ |
2940 | /* Replace scalar calls from SLP node NODE with setting of their lhs to zero. |
2941 | For loop vectorization this is done in vectorizable_call, but for SLP | |
2942 | it needs to be deferred until end of vect_schedule_slp, because multiple | |
2943 | SLP instances may refer to the same scalar stmt. */ | |
2944 | ||
2945 | static void | |
2946 | vect_remove_slp_scalar_calls (slp_tree node) | |
2947 | { | |
2948 | gimple stmt, new_stmt; | |
2949 | gimple_stmt_iterator gsi; | |
2950 | int i; | |
2951 | slp_void_p child; | |
2952 | tree lhs; | |
2953 | stmt_vec_info stmt_info; | |
2954 | ||
2955 | if (!node) | |
2956 | return; | |
2957 | ||
2958 | FOR_EACH_VEC_ELT (slp_void_p, SLP_TREE_CHILDREN (node), i, child) | |
2959 | vect_remove_slp_scalar_calls ((slp_tree) child); | |
2960 | ||
2961 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt) | |
2962 | { | |
2963 | if (!is_gimple_call (stmt) || gimple_bb (stmt) == NULL) | |
2964 | continue; | |
2965 | stmt_info = vinfo_for_stmt (stmt); | |
2966 | if (stmt_info == NULL | |
2967 | || is_pattern_stmt_p (stmt_info) | |
2968 | || !PURE_SLP_STMT (stmt_info)) | |
2969 | continue; | |
2970 | lhs = gimple_call_lhs (stmt); | |
2971 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2972 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2973 | set_vinfo_for_stmt (stmt, NULL); | |
2974 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2975 | gsi = gsi_for_stmt (stmt); | |
2976 | gsi_replace (&gsi, new_stmt, false); | |
2977 | SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt; | |
2978 | } | |
2979 | } | |
ebfd146a | 2980 | |
ff802fa1 IR |
2981 | /* Generate vector code for all SLP instances in the loop/basic block. */ |
2982 | ||
ebfd146a | 2983 | bool |
a70d6342 | 2984 | vect_schedule_slp (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo) |
ebfd146a | 2985 | { |
a70d6342 | 2986 | VEC (slp_instance, heap) *slp_instances; |
ebfd146a | 2987 | slp_instance instance; |
a70d6342 | 2988 | unsigned int i, vf; |
ebfd146a IR |
2989 | bool is_store = false; |
2990 | ||
a70d6342 IR |
2991 | if (loop_vinfo) |
2992 | { | |
2993 | slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo); | |
2994 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
b8698a0f | 2995 | } |
a70d6342 IR |
2996 | else |
2997 | { | |
2998 | slp_instances = BB_VINFO_SLP_INSTANCES (bb_vinfo); | |
2999 | vf = 1; | |
b8698a0f | 3000 | } |
a70d6342 | 3001 | |
ac47786e | 3002 | FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance) |
ebfd146a IR |
3003 | { |
3004 | /* Schedule the tree of INSTANCE. */ | |
3005 | is_store = vect_schedule_slp_instance (SLP_INSTANCE_TREE (instance), | |
a70d6342 | 3006 | instance, vf); |
8644a673 IR |
3007 | if (vect_print_dump_info (REPORT_VECTORIZED_LOCATIONS) |
3008 | || vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
ebfd146a IR |
3009 | fprintf (vect_dump, "vectorizing stmts using SLP."); |
3010 | } | |
3011 | ||
ac47786e | 3012 | FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance) |
b5aeb3bb IR |
3013 | { |
3014 | slp_tree root = SLP_INSTANCE_TREE (instance); | |
3015 | gimple store; | |
3016 | unsigned int j; | |
3017 | gimple_stmt_iterator gsi; | |
3018 | ||
dd34c087 JJ |
3019 | vect_remove_slp_scalar_calls (root); |
3020 | ||
b5aeb3bb IR |
3021 | for (j = 0; VEC_iterate (gimple, SLP_TREE_SCALAR_STMTS (root), j, store) |
3022 | && j < SLP_INSTANCE_GROUP_SIZE (instance); j++) | |
3023 | { | |
3024 | if (!STMT_VINFO_DATA_REF (vinfo_for_stmt (store))) | |
3025 | break; | |
3026 | ||
a024e70e IR |
3027 | if (is_pattern_stmt_p (vinfo_for_stmt (store))) |
3028 | store = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (store)); | |
b5aeb3bb IR |
3029 | /* Free the attached stmt_vec_info and remove the stmt. */ |
3030 | gsi = gsi_for_stmt (store); | |
3031 | gsi_remove (&gsi, true); | |
3032 | free_stmt_vec_info (store); | |
3033 | } | |
3034 | } | |
3035 | ||
ebfd146a IR |
3036 | return is_store; |
3037 | } | |
a70d6342 IR |
3038 | |
3039 | ||
3040 | /* Vectorize the basic block. */ | |
3041 | ||
3042 | void | |
3043 | vect_slp_transform_bb (basic_block bb) | |
3044 | { | |
3045 | bb_vec_info bb_vinfo = vec_info_for_bb (bb); | |
3046 | gimple_stmt_iterator si; | |
3047 | ||
3048 | gcc_assert (bb_vinfo); | |
3049 | ||
3050 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3051 | fprintf (vect_dump, "SLPing BB\n"); | |
3052 | ||
3053 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
3054 | { | |
3055 | gimple stmt = gsi_stmt (si); | |
3056 | stmt_vec_info stmt_info; | |
3057 | ||
3058 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3059 | { | |
3060 | fprintf (vect_dump, "------>SLPing statement: "); | |
3061 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
3062 | } | |
3063 | ||
3064 | stmt_info = vinfo_for_stmt (stmt); | |
3065 | gcc_assert (stmt_info); | |
3066 | ||
3067 | /* Schedule all the SLP instances when the first SLP stmt is reached. */ | |
3068 | if (STMT_SLP_TYPE (stmt_info)) | |
3069 | { | |
3070 | vect_schedule_slp (NULL, bb_vinfo); | |
3071 | break; | |
3072 | } | |
3073 | } | |
3074 | ||
3075 | mark_sym_for_renaming (gimple_vop (cfun)); | |
3076 | /* The memory tags and pointers in vectorized statements need to | |
3077 | have their SSA forms updated. FIXME, why can't this be delayed | |
3078 | until all the loops have been transformed? */ | |
3079 | update_ssa (TODO_update_ssa); | |
3080 | ||
3081 | if (vect_print_dump_info (REPORT_DETAILS)) | |
e9dbe7bb | 3082 | fprintf (vect_dump, "BASIC BLOCK VECTORIZED\n"); |
a70d6342 | 3083 | |
12aaf609 IR |
3084 | destroy_bb_vec_info (bb_vinfo); |
3085 | } | |
a70d6342 | 3086 |