]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* SLP - Basic Block Vectorization |
2635892a | 2 | Copyright (C) 2007, 2008, 2009, 2010, 2011 |
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 | { | |
71 | if (!node) | |
72 | return; | |
73 | ||
74 | if (SLP_TREE_LEFT (node)) | |
75 | vect_free_slp_tree (SLP_TREE_LEFT (node)); | |
b8698a0f | 76 | |
ebfd146a IR |
77 | if (SLP_TREE_RIGHT (node)) |
78 | vect_free_slp_tree (SLP_TREE_RIGHT (node)); | |
b8698a0f | 79 | |
ebfd146a | 80 | VEC_free (gimple, heap, SLP_TREE_SCALAR_STMTS (node)); |
b8698a0f | 81 | |
ebfd146a IR |
82 | if (SLP_TREE_VEC_STMTS (node)) |
83 | VEC_free (gimple, heap, SLP_TREE_VEC_STMTS (node)); | |
84 | ||
85 | free (node); | |
86 | } | |
87 | ||
88 | ||
89 | /* Free the memory allocated for the SLP instance. */ | |
90 | ||
91 | void | |
92 | vect_free_slp_instance (slp_instance instance) | |
93 | { | |
94 | vect_free_slp_tree (SLP_INSTANCE_TREE (instance)); | |
95 | VEC_free (int, heap, SLP_INSTANCE_LOAD_PERMUTATION (instance)); | |
96 | VEC_free (slp_tree, heap, SLP_INSTANCE_LOADS (instance)); | |
97 | } | |
98 | ||
99 | ||
100 | /* Get the defs for the rhs of STMT (collect them in DEF_STMTS0/1), check that | |
101 | they are of a legal type and that they match the defs of the first stmt of | |
102 | the SLP group (stored in FIRST_STMT_...). */ | |
103 | ||
104 | static bool | |
a70d6342 | 105 | vect_get_and_check_slp_defs (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo, |
b8698a0f | 106 | slp_tree slp_node, gimple stmt, |
a70d6342 | 107 | VEC (gimple, heap) **def_stmts0, |
ebfd146a IR |
108 | VEC (gimple, heap) **def_stmts1, |
109 | enum vect_def_type *first_stmt_dt0, | |
110 | enum vect_def_type *first_stmt_dt1, | |
b8698a0f | 111 | tree *first_stmt_def0_type, |
ebfd146a IR |
112 | tree *first_stmt_def1_type, |
113 | tree *first_stmt_const_oprnd, | |
114 | int ncopies_for_cost, | |
115 | bool *pattern0, bool *pattern1) | |
116 | { | |
117 | tree oprnd; | |
118 | unsigned int i, number_of_oprnds; | |
6aa904c4 | 119 | tree def[2]; |
ebfd146a IR |
120 | gimple def_stmt; |
121 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
b8698a0f | 122 | stmt_vec_info stmt_info = |
ebfd146a IR |
123 | vinfo_for_stmt (VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0)); |
124 | enum gimple_rhs_class rhs_class; | |
a70d6342 | 125 | struct loop *loop = NULL; |
6aa904c4 IR |
126 | enum tree_code rhs_code; |
127 | bool different_types = false; | |
b8698a0f | 128 | |
a70d6342 IR |
129 | if (loop_vinfo) |
130 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
131 | |
132 | rhs_class = get_gimple_rhs_class (gimple_assign_rhs_code (stmt)); | |
133 | number_of_oprnds = gimple_num_ops (stmt) - 1; /* RHS only */ | |
134 | ||
135 | for (i = 0; i < number_of_oprnds; i++) | |
136 | { | |
137 | oprnd = gimple_op (stmt, i + 1); | |
138 | ||
6aa904c4 | 139 | if (!vect_is_simple_use (oprnd, loop_vinfo, bb_vinfo, &def_stmt, &def[i], |
a70d6342 | 140 | &dt[i]) |
ebfd146a IR |
141 | || (!def_stmt && dt[i] != vect_constant_def)) |
142 | { | |
b8698a0f | 143 | if (vect_print_dump_info (REPORT_SLP)) |
ebfd146a IR |
144 | { |
145 | fprintf (vect_dump, "Build SLP failed: can't find def for "); | |
146 | print_generic_expr (vect_dump, oprnd, TDF_SLIM); | |
147 | } | |
148 | ||
149 | return false; | |
150 | } | |
151 | ||
a70d6342 | 152 | /* Check if DEF_STMT is a part of a pattern in LOOP and get the def stmt |
ff802fa1 | 153 | from the pattern. Check that all the stmts of the node are in the |
ebfd146a | 154 | pattern. */ |
a70d6342 | 155 | if (loop && def_stmt && gimple_bb (def_stmt) |
ebfd146a IR |
156 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt)) |
157 | && vinfo_for_stmt (def_stmt) | |
83197f37 IR |
158 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (def_stmt)) |
159 | && !STMT_VINFO_RELEVANT (vinfo_for_stmt (def_stmt)) | |
160 | && !STMT_VINFO_LIVE_P (vinfo_for_stmt (def_stmt))) | |
ebfd146a IR |
161 | { |
162 | if (!*first_stmt_dt0) | |
163 | *pattern0 = true; | |
164 | else | |
165 | { | |
166 | if (i == 1 && !*first_stmt_dt1) | |
167 | *pattern1 = true; | |
168 | else if ((i == 0 && !*pattern0) || (i == 1 && !*pattern1)) | |
169 | { | |
170 | if (vect_print_dump_info (REPORT_DETAILS)) | |
171 | { | |
172 | fprintf (vect_dump, "Build SLP failed: some of the stmts" | |
173 | " are in a pattern, and others are not "); | |
174 | print_generic_expr (vect_dump, oprnd, TDF_SLIM); | |
175 | } | |
176 | ||
177 | return false; | |
178 | } | |
179 | } | |
180 | ||
181 | def_stmt = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (def_stmt)); | |
182 | dt[i] = STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt)); | |
183 | ||
184 | if (*dt == vect_unknown_def_type) | |
185 | { | |
186 | if (vect_print_dump_info (REPORT_DETAILS)) | |
187 | fprintf (vect_dump, "Unsupported pattern."); | |
188 | return false; | |
189 | } | |
190 | ||
191 | switch (gimple_code (def_stmt)) | |
192 | { | |
193 | case GIMPLE_PHI: | |
6aa904c4 | 194 | def[i] = gimple_phi_result (def_stmt); |
ebfd146a IR |
195 | break; |
196 | ||
197 | case GIMPLE_ASSIGN: | |
6aa904c4 | 198 | def[i] = gimple_assign_lhs (def_stmt); |
ebfd146a IR |
199 | break; |
200 | ||
201 | default: | |
202 | if (vect_print_dump_info (REPORT_DETAILS)) | |
203 | fprintf (vect_dump, "unsupported defining stmt: "); | |
204 | return false; | |
205 | } | |
206 | } | |
207 | ||
208 | if (!*first_stmt_dt0) | |
209 | { | |
210 | /* op0 of the first stmt of the group - store its info. */ | |
211 | *first_stmt_dt0 = dt[i]; | |
6aa904c4 IR |
212 | if (def[i]) |
213 | *first_stmt_def0_type = TREE_TYPE (def[i]); | |
ebfd146a IR |
214 | else |
215 | *first_stmt_const_oprnd = oprnd; | |
216 | ||
217 | /* Analyze costs (for the first stmt of the group only). */ | |
218 | if (rhs_class != GIMPLE_SINGLE_RHS) | |
219 | /* Not memory operation (we don't call this functions for loads). */ | |
220 | vect_model_simple_cost (stmt_info, ncopies_for_cost, dt, slp_node); | |
221 | else | |
222 | /* Store. */ | |
272c6793 RS |
223 | vect_model_store_cost (stmt_info, ncopies_for_cost, false, |
224 | dt[0], slp_node); | |
ebfd146a | 225 | } |
b8698a0f | 226 | |
ebfd146a IR |
227 | else |
228 | { | |
229 | if (!*first_stmt_dt1 && i == 1) | |
230 | { | |
231 | /* op1 of the first stmt of the group - store its info. */ | |
232 | *first_stmt_dt1 = dt[i]; | |
6aa904c4 IR |
233 | if (def[i]) |
234 | *first_stmt_def1_type = TREE_TYPE (def[i]); | |
ebfd146a IR |
235 | else |
236 | { | |
b8698a0f | 237 | /* We assume that the stmt contains only one constant |
ebfd146a IR |
238 | operand. We fail otherwise, to be on the safe side. */ |
239 | if (*first_stmt_const_oprnd) | |
240 | { | |
b8698a0f | 241 | if (vect_print_dump_info (REPORT_SLP)) |
ebfd146a | 242 | fprintf (vect_dump, "Build SLP failed: two constant " |
b8698a0f | 243 | "oprnds in stmt"); |
ebfd146a IR |
244 | return false; |
245 | } | |
246 | *first_stmt_const_oprnd = oprnd; | |
247 | } | |
248 | } | |
249 | else | |
250 | { | |
b8698a0f | 251 | /* Not first stmt of the group, check that the def-stmt/s match |
b010117a IR |
252 | the def-stmt/s of the first stmt. Allow different definition |
253 | types for reduction chains: the first stmt must be a | |
254 | vect_reduction_def (a phi node), and the rest | |
255 | vect_internal_def. */ | |
b8698a0f | 256 | if ((i == 0 |
b010117a IR |
257 | && ((*first_stmt_dt0 != dt[i] |
258 | && !(*first_stmt_dt0 == vect_reduction_def | |
259 | && dt[i] == vect_internal_def)) | |
6aa904c4 | 260 | || (*first_stmt_def0_type && def[0] |
8533c9d8 | 261 | && !types_compatible_p (*first_stmt_def0_type, |
6aa904c4 | 262 | TREE_TYPE (def[0]))))) |
b8698a0f | 263 | || (i == 1 |
b010117a IR |
264 | && ((*first_stmt_dt1 != dt[i] |
265 | && !(*first_stmt_dt1 == vect_reduction_def | |
266 | && dt[i] == vect_internal_def)) | |
6aa904c4 | 267 | || (*first_stmt_def1_type && def[1] |
8533c9d8 | 268 | && !types_compatible_p (*first_stmt_def1_type, |
6aa904c4 IR |
269 | TREE_TYPE (def[1]))))) |
270 | || (!def[i] | |
8533c9d8 | 271 | && !types_compatible_p (TREE_TYPE (*first_stmt_const_oprnd), |
6aa904c4 IR |
272 | TREE_TYPE (oprnd))) |
273 | || different_types) | |
b8698a0f | 274 | { |
6aa904c4 IR |
275 | if (i != number_of_oprnds - 1) |
276 | different_types = true; | |
277 | else | |
278 | { | |
279 | if (is_gimple_assign (stmt) | |
280 | && (rhs_code = gimple_assign_rhs_code (stmt)) | |
281 | && TREE_CODE_CLASS (rhs_code) == tcc_binary | |
282 | && commutative_tree_code (rhs_code) | |
283 | && *first_stmt_dt0 == dt[1] | |
284 | && *first_stmt_dt1 == dt[0] | |
285 | && def[0] && def[1] | |
286 | && !(*first_stmt_def0_type | |
287 | && !types_compatible_p (*first_stmt_def0_type, | |
288 | TREE_TYPE (def[1]))) | |
289 | && !(*first_stmt_def1_type | |
290 | && !types_compatible_p (*first_stmt_def1_type, | |
291 | TREE_TYPE (def[0])))) | |
292 | { | |
293 | if (vect_print_dump_info (REPORT_SLP)) | |
294 | { | |
295 | fprintf (vect_dump, "Swapping operands of "); | |
296 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
297 | } | |
298 | ||
299 | swap_tree_operands (stmt, gimple_assign_rhs1_ptr (stmt), | |
300 | gimple_assign_rhs2_ptr (stmt)); | |
301 | } | |
302 | else | |
303 | { | |
304 | if (vect_print_dump_info (REPORT_SLP)) | |
305 | fprintf (vect_dump, "Build SLP failed: different types "); | |
b8698a0f | 306 | |
6aa904c4 IR |
307 | return false; |
308 | } | |
309 | } | |
ebfd146a IR |
310 | } |
311 | } | |
312 | } | |
313 | ||
314 | /* Check the types of the definitions. */ | |
315 | switch (dt[i]) | |
316 | { | |
317 | case vect_constant_def: | |
8644a673 | 318 | case vect_external_def: |
ebfd146a | 319 | break; |
b8698a0f | 320 | |
8644a673 | 321 | case vect_internal_def: |
b5aeb3bb | 322 | case vect_reduction_def: |
6aa904c4 | 323 | if ((i == 0 && !different_types) || (i == 1 && different_types)) |
ebfd146a IR |
324 | VEC_safe_push (gimple, heap, *def_stmts0, def_stmt); |
325 | else | |
6aa904c4 | 326 | VEC_safe_push (gimple, heap, *def_stmts1, def_stmt); |
ebfd146a IR |
327 | break; |
328 | ||
329 | default: | |
330 | /* FORNOW: Not supported. */ | |
b8698a0f | 331 | if (vect_print_dump_info (REPORT_SLP)) |
ebfd146a IR |
332 | { |
333 | fprintf (vect_dump, "Build SLP failed: illegal type of def "); | |
6aa904c4 | 334 | print_generic_expr (vect_dump, def[i], TDF_SLIM); |
ebfd146a IR |
335 | } |
336 | ||
337 | return false; | |
338 | } | |
339 | } | |
340 | ||
341 | return true; | |
342 | } | |
343 | ||
344 | ||
345 | /* Recursively build an SLP tree starting from NODE. | |
b8698a0f | 346 | Fail (and return FALSE) if def-stmts are not isomorphic, require data |
ff802fa1 | 347 | permutation or are of unsupported types of operation. Otherwise, return |
ebfd146a IR |
348 | TRUE. */ |
349 | ||
350 | static bool | |
b8698a0f | 351 | vect_build_slp_tree (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo, |
a70d6342 IR |
352 | slp_tree *node, unsigned int group_size, |
353 | int *inside_cost, int *outside_cost, | |
354 | int ncopies_for_cost, unsigned int *max_nunits, | |
ebfd146a | 355 | VEC (int, heap) **load_permutation, |
a70d6342 | 356 | VEC (slp_tree, heap) **loads, |
6aa904c4 | 357 | unsigned int vectorization_factor, bool *loads_permuted) |
ebfd146a IR |
358 | { |
359 | VEC (gimple, heap) *def_stmts0 = VEC_alloc (gimple, heap, group_size); | |
360 | VEC (gimple, heap) *def_stmts1 = VEC_alloc (gimple, heap, group_size); | |
361 | unsigned int i; | |
362 | VEC (gimple, heap) *stmts = SLP_TREE_SCALAR_STMTS (*node); | |
363 | gimple stmt = VEC_index (gimple, stmts, 0); | |
81f40b79 ILT |
364 | enum vect_def_type first_stmt_dt0 = vect_uninitialized_def; |
365 | enum vect_def_type first_stmt_dt1 = vect_uninitialized_def; | |
2200fc49 | 366 | enum tree_code first_stmt_code = ERROR_MARK, rhs_code = ERROR_MARK; |
ebfd146a IR |
367 | tree first_stmt_def1_type = NULL_TREE, first_stmt_def0_type = NULL_TREE; |
368 | tree lhs; | |
369 | bool stop_recursion = false, need_same_oprnds = false; | |
370 | tree vectype, scalar_type, first_op1 = NULL_TREE; | |
a70d6342 | 371 | unsigned int ncopies; |
ebfd146a IR |
372 | optab optab; |
373 | int icode; | |
374 | enum machine_mode optab_op2_mode; | |
375 | enum machine_mode vec_mode; | |
376 | tree first_stmt_const_oprnd = NULL_TREE; | |
377 | struct data_reference *first_dr; | |
378 | bool pattern0 = false, pattern1 = false; | |
379 | HOST_WIDE_INT dummy; | |
380 | bool permutation = false; | |
381 | unsigned int load_place; | |
b5aeb3bb | 382 | gimple first_load, prev_first_load = NULL; |
ebfd146a IR |
383 | |
384 | /* For every stmt in NODE find its def stmt/s. */ | |
ac47786e | 385 | FOR_EACH_VEC_ELT (gimple, stmts, i, stmt) |
ebfd146a | 386 | { |
b8698a0f | 387 | if (vect_print_dump_info (REPORT_SLP)) |
ebfd146a IR |
388 | { |
389 | fprintf (vect_dump, "Build SLP for "); | |
390 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
391 | } | |
392 | ||
4b5caab7 IR |
393 | /* Fail to vectorize statements marked as unvectorizable. */ |
394 | if (!STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (stmt))) | |
395 | { | |
396 | if (vect_print_dump_info (REPORT_SLP)) | |
397 | { | |
398 | fprintf (vect_dump, | |
399 | "Build SLP failed: unvectorizable statement "); | |
400 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
401 | } | |
402 | ||
403 | return false; | |
404 | } | |
405 | ||
ebfd146a IR |
406 | lhs = gimple_get_lhs (stmt); |
407 | if (lhs == NULL_TREE) | |
408 | { | |
b8698a0f | 409 | if (vect_print_dump_info (REPORT_SLP)) |
ebfd146a IR |
410 | { |
411 | fprintf (vect_dump, | |
412 | "Build SLP failed: not GIMPLE_ASSIGN nor GIMPLE_CALL"); | |
413 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
414 | } | |
b8698a0f | 415 | |
ebfd146a IR |
416 | return false; |
417 | } | |
418 | ||
b8698a0f | 419 | scalar_type = vect_get_smallest_scalar_type (stmt, &dummy, &dummy); |
ebfd146a IR |
420 | vectype = get_vectype_for_scalar_type (scalar_type); |
421 | if (!vectype) | |
422 | { | |
423 | if (vect_print_dump_info (REPORT_SLP)) | |
424 | { | |
425 | fprintf (vect_dump, "Build SLP failed: unsupported data-type "); | |
426 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
427 | } | |
428 | return false; | |
429 | } | |
b8698a0f | 430 | |
4ef69dfc IR |
431 | /* In case of multiple types we need to detect the smallest type. */ |
432 | if (*max_nunits < TYPE_VECTOR_SUBPARTS (vectype)) | |
a70d6342 | 433 | { |
4ef69dfc IR |
434 | *max_nunits = TYPE_VECTOR_SUBPARTS (vectype); |
435 | if (bb_vinfo) | |
436 | vectorization_factor = *max_nunits; | |
a70d6342 | 437 | } |
b8698a0f | 438 | |
4ef69dfc | 439 | ncopies = vectorization_factor / TYPE_VECTOR_SUBPARTS (vectype); |
b8698a0f | 440 | |
ebfd146a IR |
441 | if (is_gimple_call (stmt)) |
442 | rhs_code = CALL_EXPR; | |
443 | else | |
444 | rhs_code = gimple_assign_rhs_code (stmt); | |
445 | ||
446 | /* Check the operation. */ | |
447 | if (i == 0) | |
448 | { | |
449 | first_stmt_code = rhs_code; | |
450 | ||
b8698a0f | 451 | /* Shift arguments should be equal in all the packed stmts for a |
ebfd146a IR |
452 | vector shift with scalar shift operand. */ |
453 | if (rhs_code == LSHIFT_EXPR || rhs_code == RSHIFT_EXPR | |
454 | || rhs_code == LROTATE_EXPR | |
455 | || rhs_code == RROTATE_EXPR) | |
456 | { | |
457 | vec_mode = TYPE_MODE (vectype); | |
458 | ||
459 | /* First see if we have a vector/vector shift. */ | |
460 | optab = optab_for_tree_code (rhs_code, vectype, | |
461 | optab_vector); | |
462 | ||
463 | if (!optab | |
947131ba | 464 | || optab_handler (optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
465 | { |
466 | /* No vector/vector shift, try for a vector/scalar shift. */ | |
467 | optab = optab_for_tree_code (rhs_code, vectype, | |
468 | optab_scalar); | |
469 | ||
470 | if (!optab) | |
471 | { | |
472 | if (vect_print_dump_info (REPORT_SLP)) | |
473 | fprintf (vect_dump, "Build SLP failed: no optab."); | |
474 | return false; | |
475 | } | |
947131ba | 476 | icode = (int) optab_handler (optab, vec_mode); |
ebfd146a IR |
477 | if (icode == CODE_FOR_nothing) |
478 | { | |
479 | if (vect_print_dump_info (REPORT_SLP)) | |
480 | fprintf (vect_dump, "Build SLP failed: " | |
481 | "op not supported by target."); | |
482 | return false; | |
483 | } | |
484 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
485 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
486 | { | |
487 | need_same_oprnds = true; | |
488 | first_op1 = gimple_assign_rhs2 (stmt); | |
489 | } | |
490 | } | |
491 | } | |
36ba4aae IR |
492 | else if (rhs_code == WIDEN_LSHIFT_EXPR) |
493 | { | |
494 | need_same_oprnds = true; | |
495 | first_op1 = gimple_assign_rhs2 (stmt); | |
496 | } | |
ebfd146a IR |
497 | } |
498 | else | |
499 | { | |
500 | if (first_stmt_code != rhs_code | |
501 | && (first_stmt_code != IMAGPART_EXPR | |
502 | || rhs_code != REALPART_EXPR) | |
503 | && (first_stmt_code != REALPART_EXPR | |
69f11a13 IR |
504 | || rhs_code != IMAGPART_EXPR) |
505 | && !(STMT_VINFO_STRIDED_ACCESS (vinfo_for_stmt (stmt)) | |
506 | && (first_stmt_code == ARRAY_REF | |
507 | || first_stmt_code == INDIRECT_REF | |
508 | || first_stmt_code == COMPONENT_REF | |
509 | || first_stmt_code == MEM_REF))) | |
ebfd146a | 510 | { |
b8698a0f | 511 | if (vect_print_dump_info (REPORT_SLP)) |
ebfd146a | 512 | { |
b8698a0f | 513 | fprintf (vect_dump, |
ebfd146a IR |
514 | "Build SLP failed: different operation in stmt "); |
515 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
516 | } | |
b8698a0f | 517 | |
ebfd146a IR |
518 | return false; |
519 | } | |
b8698a0f L |
520 | |
521 | if (need_same_oprnds | |
ebfd146a IR |
522 | && !operand_equal_p (first_op1, gimple_assign_rhs2 (stmt), 0)) |
523 | { | |
b8698a0f | 524 | if (vect_print_dump_info (REPORT_SLP)) |
ebfd146a | 525 | { |
b8698a0f | 526 | fprintf (vect_dump, |
ebfd146a IR |
527 | "Build SLP failed: different shift arguments in "); |
528 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
529 | } | |
b8698a0f | 530 | |
ebfd146a IR |
531 | return false; |
532 | } | |
533 | } | |
534 | ||
535 | /* Strided store or load. */ | |
536 | if (STMT_VINFO_STRIDED_ACCESS (vinfo_for_stmt (stmt))) | |
537 | { | |
538 | if (REFERENCE_CLASS_P (lhs)) | |
539 | { | |
540 | /* Store. */ | |
b8698a0f L |
541 | if (!vect_get_and_check_slp_defs (loop_vinfo, bb_vinfo, *node, |
542 | stmt, &def_stmts0, &def_stmts1, | |
543 | &first_stmt_dt0, | |
544 | &first_stmt_dt1, | |
545 | &first_stmt_def0_type, | |
ebfd146a IR |
546 | &first_stmt_def1_type, |
547 | &first_stmt_const_oprnd, | |
548 | ncopies_for_cost, | |
549 | &pattern0, &pattern1)) | |
550 | return false; | |
551 | } | |
b5aeb3bb IR |
552 | else |
553 | { | |
554 | /* Load. */ | |
555 | /* FORNOW: Check that there is no gap between the loads. */ | |
e14c1050 IR |
556 | if ((GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) == stmt |
557 | && GROUP_GAP (vinfo_for_stmt (stmt)) != 0) | |
558 | || (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) != stmt | |
559 | && GROUP_GAP (vinfo_for_stmt (stmt)) != 1)) | |
b5aeb3bb IR |
560 | { |
561 | if (vect_print_dump_info (REPORT_SLP)) | |
562 | { | |
563 | fprintf (vect_dump, "Build SLP failed: strided " | |
564 | "loads have gaps "); | |
565 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
566 | } | |
b8698a0f | 567 | |
b5aeb3bb IR |
568 | return false; |
569 | } | |
2f0fa28e | 570 | |
b5aeb3bb IR |
571 | /* Check that the size of interleaved loads group is not |
572 | greater than the SLP group size. */ | |
6aa904c4 IR |
573 | if (loop_vinfo |
574 | && GROUP_SIZE (vinfo_for_stmt (stmt)) > ncopies * group_size) | |
b5aeb3bb IR |
575 | { |
576 | if (vect_print_dump_info (REPORT_SLP)) | |
577 | { | |
578 | fprintf (vect_dump, "Build SLP failed: the number of " | |
579 | "interleaved loads is greater than" | |
580 | " the SLP group size "); | |
581 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
582 | } | |
b8698a0f | 583 | |
b5aeb3bb IR |
584 | return false; |
585 | } | |
586 | ||
e14c1050 | 587 | first_load = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)); |
b5aeb3bb IR |
588 | if (prev_first_load) |
589 | { | |
590 | /* Check that there are no loads from different interleaving | |
ff802fa1 | 591 | chains in the same node. The only exception is complex |
b5aeb3bb IR |
592 | numbers. */ |
593 | if (prev_first_load != first_load | |
594 | && rhs_code != REALPART_EXPR | |
595 | && rhs_code != IMAGPART_EXPR) | |
596 | { | |
597 | if (vect_print_dump_info (REPORT_SLP)) | |
598 | { | |
599 | fprintf (vect_dump, "Build SLP failed: different " | |
600 | "interleaving chains in one node "); | |
601 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
602 | } | |
603 | ||
604 | return false; | |
605 | } | |
606 | } | |
607 | else | |
608 | prev_first_load = first_load; | |
b8698a0f | 609 | |
ebfd146a IR |
610 | if (first_load == stmt) |
611 | { | |
612 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt)); | |
720f5239 | 613 | if (vect_supportable_dr_alignment (first_dr, false) |
ebfd146a IR |
614 | == dr_unaligned_unsupported) |
615 | { | |
616 | if (vect_print_dump_info (REPORT_SLP)) | |
617 | { | |
618 | fprintf (vect_dump, "Build SLP failed: unsupported " | |
619 | "unaligned load "); | |
620 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
621 | } | |
b8698a0f | 622 | |
ebfd146a IR |
623 | return false; |
624 | } | |
b8698a0f | 625 | |
ebfd146a IR |
626 | /* Analyze costs (for the first stmt in the group). */ |
627 | vect_model_load_cost (vinfo_for_stmt (stmt), | |
272c6793 | 628 | ncopies_for_cost, false, *node); |
ebfd146a | 629 | } |
b8698a0f | 630 | |
ff802fa1 | 631 | /* Store the place of this load in the interleaving chain. In |
ebfd146a IR |
632 | case that permutation is needed we later decide if a specific |
633 | permutation is supported. */ | |
634 | load_place = vect_get_place_in_interleaving_chain (stmt, | |
635 | first_load); | |
636 | if (load_place != i) | |
637 | permutation = true; | |
b8698a0f | 638 | |
ebfd146a | 639 | VEC_safe_push (int, heap, *load_permutation, load_place); |
b8698a0f | 640 | |
ebfd146a IR |
641 | /* We stop the tree when we reach a group of loads. */ |
642 | stop_recursion = true; | |
643 | continue; | |
644 | } | |
645 | } /* Strided access. */ | |
646 | else | |
647 | { | |
648 | if (TREE_CODE_CLASS (rhs_code) == tcc_reference) | |
649 | { | |
650 | /* Not strided load. */ | |
b8698a0f | 651 | if (vect_print_dump_info (REPORT_SLP)) |
ebfd146a IR |
652 | { |
653 | fprintf (vect_dump, "Build SLP failed: not strided load "); | |
654 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
655 | } | |
656 | ||
657 | /* FORNOW: Not strided loads are not supported. */ | |
658 | return false; | |
659 | } | |
660 | ||
661 | /* Not memory operation. */ | |
662 | if (TREE_CODE_CLASS (rhs_code) != tcc_binary | |
663 | && TREE_CODE_CLASS (rhs_code) != tcc_unary) | |
664 | { | |
b8698a0f | 665 | if (vect_print_dump_info (REPORT_SLP)) |
ebfd146a IR |
666 | { |
667 | fprintf (vect_dump, "Build SLP failed: operation"); | |
668 | fprintf (vect_dump, " unsupported "); | |
669 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
670 | } | |
671 | ||
672 | return false; | |
673 | } | |
674 | ||
b8698a0f | 675 | /* Find the def-stmts. */ |
a70d6342 | 676 | if (!vect_get_and_check_slp_defs (loop_vinfo, bb_vinfo, *node, stmt, |
ebfd146a | 677 | &def_stmts0, &def_stmts1, |
b8698a0f L |
678 | &first_stmt_dt0, &first_stmt_dt1, |
679 | &first_stmt_def0_type, | |
ebfd146a IR |
680 | &first_stmt_def1_type, |
681 | &first_stmt_const_oprnd, | |
682 | ncopies_for_cost, | |
683 | &pattern0, &pattern1)) | |
684 | return false; | |
685 | } | |
686 | } | |
687 | ||
688 | /* Add the costs of the node to the overall instance costs. */ | |
b8698a0f | 689 | *inside_cost += SLP_TREE_INSIDE_OF_LOOP_COST (*node); |
ebfd146a IR |
690 | *outside_cost += SLP_TREE_OUTSIDE_OF_LOOP_COST (*node); |
691 | ||
692 | /* Strided loads were reached - stop the recursion. */ | |
693 | if (stop_recursion) | |
694 | { | |
6aa904c4 | 695 | VEC_safe_push (slp_tree, heap, *loads, *node); |
ebfd146a IR |
696 | if (permutation) |
697 | { | |
6aa904c4 IR |
698 | |
699 | *loads_permuted = true; | |
35e1a5e7 | 700 | *inside_cost |
720f5239 | 701 | += targetm.vectorize.builtin_vectorization_cost (vec_perm, NULL, 0) |
35e1a5e7 | 702 | * group_size; |
ebfd146a | 703 | } |
2200fc49 | 704 | else |
6aa904c4 IR |
705 | { |
706 | /* We don't check here complex numbers chains, so we set | |
707 | LOADS_PERMUTED for further check in | |
708 | vect_supported_load_permutation_p. */ | |
2200fc49 | 709 | if (rhs_code == REALPART_EXPR || rhs_code == IMAGPART_EXPR) |
6aa904c4 | 710 | *loads_permuted = true; |
2200fc49 | 711 | } |
ebfd146a IR |
712 | |
713 | return true; | |
714 | } | |
715 | ||
b8698a0f | 716 | /* Create SLP_TREE nodes for the definition node/s. */ |
8644a673 | 717 | if (first_stmt_dt0 == vect_internal_def) |
ebfd146a IR |
718 | { |
719 | slp_tree left_node = XNEW (struct _slp_tree); | |
720 | SLP_TREE_SCALAR_STMTS (left_node) = def_stmts0; | |
721 | SLP_TREE_VEC_STMTS (left_node) = NULL; | |
722 | SLP_TREE_LEFT (left_node) = NULL; | |
723 | SLP_TREE_RIGHT (left_node) = NULL; | |
724 | SLP_TREE_OUTSIDE_OF_LOOP_COST (left_node) = 0; | |
725 | SLP_TREE_INSIDE_OF_LOOP_COST (left_node) = 0; | |
b8698a0f L |
726 | if (!vect_build_slp_tree (loop_vinfo, bb_vinfo, &left_node, group_size, |
727 | inside_cost, outside_cost, ncopies_for_cost, | |
a70d6342 | 728 | max_nunits, load_permutation, loads, |
6aa904c4 | 729 | vectorization_factor, loads_permuted)) |
ebfd146a | 730 | return false; |
b8698a0f | 731 | |
ebfd146a IR |
732 | SLP_TREE_LEFT (*node) = left_node; |
733 | } | |
734 | ||
8644a673 | 735 | if (first_stmt_dt1 == vect_internal_def) |
ebfd146a IR |
736 | { |
737 | slp_tree right_node = XNEW (struct _slp_tree); | |
738 | SLP_TREE_SCALAR_STMTS (right_node) = def_stmts1; | |
739 | SLP_TREE_VEC_STMTS (right_node) = NULL; | |
740 | SLP_TREE_LEFT (right_node) = NULL; | |
741 | SLP_TREE_RIGHT (right_node) = NULL; | |
742 | SLP_TREE_OUTSIDE_OF_LOOP_COST (right_node) = 0; | |
743 | SLP_TREE_INSIDE_OF_LOOP_COST (right_node) = 0; | |
a70d6342 | 744 | if (!vect_build_slp_tree (loop_vinfo, bb_vinfo, &right_node, group_size, |
ebfd146a | 745 | inside_cost, outside_cost, ncopies_for_cost, |
a70d6342 | 746 | max_nunits, load_permutation, loads, |
6aa904c4 | 747 | vectorization_factor, loads_permuted)) |
ebfd146a | 748 | return false; |
b8698a0f | 749 | |
ebfd146a IR |
750 | SLP_TREE_RIGHT (*node) = right_node; |
751 | } | |
752 | ||
753 | return true; | |
754 | } | |
755 | ||
756 | ||
757 | static void | |
758 | vect_print_slp_tree (slp_tree node) | |
759 | { | |
760 | int i; | |
761 | gimple stmt; | |
762 | ||
763 | if (!node) | |
764 | return; | |
765 | ||
766 | fprintf (vect_dump, "node "); | |
ac47786e | 767 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt) |
ebfd146a IR |
768 | { |
769 | fprintf (vect_dump, "\n\tstmt %d ", i); | |
b8698a0f | 770 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); |
ebfd146a IR |
771 | } |
772 | fprintf (vect_dump, "\n"); | |
773 | ||
774 | vect_print_slp_tree (SLP_TREE_LEFT (node)); | |
775 | vect_print_slp_tree (SLP_TREE_RIGHT (node)); | |
776 | } | |
777 | ||
778 | ||
b8698a0f L |
779 | /* Mark the tree rooted at NODE with MARK (PURE_SLP or HYBRID). |
780 | If MARK is HYBRID, it refers to a specific stmt in NODE (the stmt at index | |
ff802fa1 | 781 | J). Otherwise, MARK is PURE_SLP and J is -1, which indicates that all the |
ebfd146a IR |
782 | stmts in NODE are to be marked. */ |
783 | ||
784 | static void | |
785 | vect_mark_slp_stmts (slp_tree node, enum slp_vect_type mark, int j) | |
786 | { | |
787 | int i; | |
788 | gimple stmt; | |
789 | ||
790 | if (!node) | |
791 | return; | |
792 | ||
ac47786e | 793 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt) |
ebfd146a IR |
794 | if (j < 0 || i == j) |
795 | STMT_SLP_TYPE (vinfo_for_stmt (stmt)) = mark; | |
796 | ||
797 | vect_mark_slp_stmts (SLP_TREE_LEFT (node), mark, j); | |
798 | vect_mark_slp_stmts (SLP_TREE_RIGHT (node), mark, j); | |
799 | } | |
800 | ||
801 | ||
a70d6342 IR |
802 | /* Mark the statements of the tree rooted at NODE as relevant (vect_used). */ |
803 | ||
804 | static void | |
805 | vect_mark_slp_stmts_relevant (slp_tree node) | |
806 | { | |
807 | int i; | |
808 | gimple stmt; | |
809 | stmt_vec_info stmt_info; | |
810 | ||
811 | if (!node) | |
812 | return; | |
813 | ||
ac47786e | 814 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt) |
a70d6342 IR |
815 | { |
816 | stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 817 | gcc_assert (!STMT_VINFO_RELEVANT (stmt_info) |
a70d6342 IR |
818 | || STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_scope); |
819 | STMT_VINFO_RELEVANT (stmt_info) = vect_used_in_scope; | |
820 | } | |
821 | ||
822 | vect_mark_slp_stmts_relevant (SLP_TREE_LEFT (node)); | |
823 | vect_mark_slp_stmts_relevant (SLP_TREE_RIGHT (node)); | |
824 | } | |
825 | ||
826 | ||
b8698a0f | 827 | /* Check if the permutation required by the SLP INSTANCE is supported. |
ebfd146a IR |
828 | Reorganize the SLP nodes stored in SLP_INSTANCE_LOADS if needed. */ |
829 | ||
830 | static bool | |
831 | vect_supported_slp_permutation_p (slp_instance instance) | |
832 | { | |
833 | slp_tree node = VEC_index (slp_tree, SLP_INSTANCE_LOADS (instance), 0); | |
834 | gimple stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (node), 0); | |
e14c1050 | 835 | gimple first_load = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)); |
ebfd146a IR |
836 | VEC (slp_tree, heap) *sorted_loads = NULL; |
837 | int index; | |
838 | slp_tree *tmp_loads = NULL; | |
b8698a0f | 839 | int group_size = SLP_INSTANCE_GROUP_SIZE (instance), i, j; |
ebfd146a | 840 | slp_tree load; |
b8698a0f L |
841 | |
842 | /* FORNOW: The only supported loads permutation is loads from the same | |
ebfd146a | 843 | location in all the loads in the node, when the data-refs in |
b8698a0f | 844 | nodes of LOADS constitute an interleaving chain. |
ebfd146a IR |
845 | Sort the nodes according to the order of accesses in the chain. */ |
846 | tmp_loads = (slp_tree *) xmalloc (sizeof (slp_tree) * group_size); | |
b8698a0f L |
847 | for (i = 0, j = 0; |
848 | VEC_iterate (int, SLP_INSTANCE_LOAD_PERMUTATION (instance), i, index) | |
849 | && VEC_iterate (slp_tree, SLP_INSTANCE_LOADS (instance), j, load); | |
ebfd146a IR |
850 | i += group_size, j++) |
851 | { | |
852 | gimple scalar_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (load), 0); | |
853 | /* Check that the loads are all in the same interleaving chain. */ | |
e14c1050 | 854 | if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (scalar_stmt)) != first_load) |
ebfd146a IR |
855 | { |
856 | if (vect_print_dump_info (REPORT_DETAILS)) | |
857 | { | |
858 | fprintf (vect_dump, "Build SLP failed: unsupported data " | |
859 | "permutation "); | |
860 | print_gimple_stmt (vect_dump, scalar_stmt, 0, TDF_SLIM); | |
861 | } | |
b8698a0f | 862 | |
ebfd146a | 863 | free (tmp_loads); |
b8698a0f | 864 | return false; |
ebfd146a IR |
865 | } |
866 | ||
867 | tmp_loads[index] = load; | |
868 | } | |
b8698a0f | 869 | |
ebfd146a IR |
870 | sorted_loads = VEC_alloc (slp_tree, heap, group_size); |
871 | for (i = 0; i < group_size; i++) | |
872 | VEC_safe_push (slp_tree, heap, sorted_loads, tmp_loads[i]); | |
873 | ||
874 | VEC_free (slp_tree, heap, SLP_INSTANCE_LOADS (instance)); | |
875 | SLP_INSTANCE_LOADS (instance) = sorted_loads; | |
876 | free (tmp_loads); | |
877 | ||
878 | if (!vect_transform_slp_perm_load (stmt, NULL, NULL, | |
879 | SLP_INSTANCE_UNROLLING_FACTOR (instance), | |
880 | instance, true)) | |
881 | return false; | |
882 | ||
883 | return true; | |
884 | } | |
885 | ||
886 | ||
b5aeb3bb IR |
887 | /* Rearrange the statements of NODE according to PERMUTATION. */ |
888 | ||
889 | static void | |
890 | vect_slp_rearrange_stmts (slp_tree node, unsigned int group_size, | |
891 | VEC (int, heap) *permutation) | |
892 | { | |
893 | gimple stmt; | |
894 | VEC (gimple, heap) *tmp_stmts; | |
895 | unsigned int index, i; | |
896 | ||
897 | if (!node) | |
898 | return; | |
899 | ||
900 | vect_slp_rearrange_stmts (SLP_TREE_LEFT (node), group_size, permutation); | |
901 | vect_slp_rearrange_stmts (SLP_TREE_RIGHT (node), group_size, permutation); | |
902 | ||
903 | gcc_assert (group_size == VEC_length (gimple, SLP_TREE_SCALAR_STMTS (node))); | |
904 | tmp_stmts = VEC_alloc (gimple, heap, group_size); | |
905 | ||
906 | for (i = 0; i < group_size; i++) | |
907 | VEC_safe_push (gimple, heap, tmp_stmts, NULL); | |
908 | ||
ac47786e | 909 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt) |
b5aeb3bb IR |
910 | { |
911 | index = VEC_index (int, permutation, i); | |
912 | VEC_replace (gimple, tmp_stmts, index, stmt); | |
913 | } | |
914 | ||
915 | VEC_free (gimple, heap, SLP_TREE_SCALAR_STMTS (node)); | |
916 | SLP_TREE_SCALAR_STMTS (node) = tmp_stmts; | |
917 | } | |
918 | ||
919 | ||
ebfd146a IR |
920 | /* Check if the required load permutation is supported. |
921 | LOAD_PERMUTATION contains a list of indices of the loads. | |
922 | In SLP this permutation is relative to the order of strided stores that are | |
923 | the base of the SLP instance. */ | |
924 | ||
925 | static bool | |
926 | vect_supported_load_permutation_p (slp_instance slp_instn, int group_size, | |
927 | VEC (int, heap) *load_permutation) | |
928 | { | |
b5aeb3bb IR |
929 | int i = 0, j, prev = -1, next, k, number_of_groups; |
930 | bool supported, bad_permutation = false; | |
7417f6c0 | 931 | sbitmap load_index; |
2200fc49 | 932 | slp_tree node, other_complex_node; |
6aa904c4 | 933 | gimple stmt, first = NULL, other_node_first, load, next_load, first_load; |
2200fc49 | 934 | unsigned complex_numbers = 0; |
6aa904c4 IR |
935 | struct data_reference *dr; |
936 | bb_vec_info bb_vinfo; | |
ebfd146a | 937 | |
a70d6342 | 938 | /* FORNOW: permutations are only supported in SLP. */ |
ebfd146a IR |
939 | if (!slp_instn) |
940 | return false; | |
941 | ||
942 | if (vect_print_dump_info (REPORT_SLP)) | |
943 | { | |
944 | fprintf (vect_dump, "Load permutation "); | |
ac47786e | 945 | FOR_EACH_VEC_ELT (int, load_permutation, i, next) |
ebfd146a IR |
946 | fprintf (vect_dump, "%d ", next); |
947 | } | |
948 | ||
b5aeb3bb IR |
949 | /* In case of reduction every load permutation is allowed, since the order |
950 | of the reduction statements is not important (as opposed to the case of | |
ff802fa1 | 951 | strided stores). The only condition we need to check is that all the |
b5aeb3bb IR |
952 | load nodes are of the same size and have the same permutation (and then |
953 | rearrange all the nodes of the SLP instance according to this | |
954 | permutation). */ | |
955 | ||
956 | /* Check that all the load nodes are of the same size. */ | |
ac47786e | 957 | FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (slp_instn), i, node) |
2200fc49 IR |
958 | { |
959 | if (VEC_length (gimple, SLP_TREE_SCALAR_STMTS (node)) | |
960 | != (unsigned) group_size) | |
961 | return false; | |
962 | ||
963 | stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (node), 0); | |
964 | if (is_gimple_assign (stmt) | |
965 | && (gimple_assign_rhs_code (stmt) == REALPART_EXPR | |
966 | || gimple_assign_rhs_code (stmt) == IMAGPART_EXPR)) | |
967 | complex_numbers++; | |
968 | } | |
969 | ||
970 | /* Complex operands can be swapped as following: | |
971 | real_c = real_b + real_a; | |
972 | imag_c = imag_a + imag_b; | |
973 | i.e., we have {real_b, imag_a} and {real_a, imag_b} instead of | |
ff802fa1 | 974 | {real_a, imag_a} and {real_b, imag_b}. We check here that if interleaving |
2200fc49 IR |
975 | chains are mixed, they match the above pattern. */ |
976 | if (complex_numbers) | |
977 | { | |
ac47786e | 978 | FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (slp_instn), i, node) |
2200fc49 | 979 | { |
ac47786e | 980 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), j, stmt) |
2200fc49 IR |
981 | { |
982 | if (j == 0) | |
983 | first = stmt; | |
984 | else | |
985 | { | |
e14c1050 | 986 | if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) != first) |
2200fc49 IR |
987 | { |
988 | if (complex_numbers != 2) | |
989 | return false; | |
990 | ||
991 | if (i == 0) | |
992 | k = 1; | |
993 | else | |
994 | k = 0; | |
995 | ||
996 | other_complex_node = VEC_index (slp_tree, | |
997 | SLP_INSTANCE_LOADS (slp_instn), k); | |
998 | other_node_first = VEC_index (gimple, | |
999 | SLP_TREE_SCALAR_STMTS (other_complex_node), 0); | |
1000 | ||
e14c1050 | 1001 | if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) |
2200fc49 IR |
1002 | != other_node_first) |
1003 | return false; | |
1004 | } | |
1005 | } | |
1006 | } | |
1007 | } | |
1008 | } | |
1009 | ||
1010 | /* We checked that this case ok, so there is no need to proceed with | |
1011 | permutation tests. */ | |
1012 | if (complex_numbers == 2) | |
1013 | { | |
1014 | VEC_free (slp_tree, heap, SLP_INSTANCE_LOADS (slp_instn)); | |
1015 | VEC_free (int, heap, SLP_INSTANCE_LOAD_PERMUTATION (slp_instn)); | |
1016 | return true; | |
1017 | } | |
1018 | ||
b5aeb3bb IR |
1019 | node = SLP_INSTANCE_TREE (slp_instn); |
1020 | stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (node), 0); | |
1021 | /* LOAD_PERMUTATION is a list of indices of all the loads of the SLP | |
1022 | instance, not all the loads belong to the same node or interleaving | |
ff802fa1 | 1023 | group. Hence, we need to divide them into groups according to |
b5aeb3bb IR |
1024 | GROUP_SIZE. */ |
1025 | number_of_groups = VEC_length (int, load_permutation) / group_size; | |
1026 | ||
b010117a IR |
1027 | /* Reduction (there are no data-refs in the root). |
1028 | In reduction chain the order of the loads is important. */ | |
1029 | if (!STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt)) | |
1030 | && !GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) | |
b5aeb3bb IR |
1031 | { |
1032 | int first_group_load_index; | |
1033 | ||
1034 | /* Compare all the permutation sequences to the first one. */ | |
1035 | for (i = 1; i < number_of_groups; i++) | |
1036 | { | |
1037 | k = 0; | |
1038 | for (j = i * group_size; j < i * group_size + group_size; j++) | |
1039 | { | |
1040 | next = VEC_index (int, load_permutation, j); | |
1041 | first_group_load_index = VEC_index (int, load_permutation, k); | |
1042 | ||
1043 | if (next != first_group_load_index) | |
1044 | { | |
1045 | bad_permutation = true; | |
1046 | break; | |
1047 | } | |
1048 | ||
1049 | k++; | |
1050 | } | |
1051 | ||
1052 | if (bad_permutation) | |
1053 | break; | |
1054 | } | |
1055 | ||
1056 | if (!bad_permutation) | |
1057 | { | |
c9c1e775 IR |
1058 | /* Check that the loads in the first sequence are different and there |
1059 | are no gaps between them. */ | |
1060 | load_index = sbitmap_alloc (group_size); | |
1061 | sbitmap_zero (load_index); | |
1062 | for (k = 0; k < group_size; k++) | |
1063 | { | |
1064 | first_group_load_index = VEC_index (int, load_permutation, k); | |
1065 | if (TEST_BIT (load_index, first_group_load_index)) | |
1066 | { | |
1067 | bad_permutation = true; | |
1068 | break; | |
1069 | } | |
1070 | ||
1071 | SET_BIT (load_index, first_group_load_index); | |
1072 | } | |
1073 | ||
1074 | if (!bad_permutation) | |
1075 | for (k = 0; k < group_size; k++) | |
1076 | if (!TEST_BIT (load_index, k)) | |
1077 | { | |
1078 | bad_permutation = true; | |
1079 | break; | |
1080 | } | |
1081 | ||
1082 | sbitmap_free (load_index); | |
1083 | } | |
1084 | ||
1085 | if (!bad_permutation) | |
1086 | { | |
1087 | /* This permutation is valid for reduction. Since the order of the | |
b5aeb3bb IR |
1088 | statements in the nodes is not important unless they are memory |
1089 | accesses, we can rearrange the statements in all the nodes | |
1090 | according to the order of the loads. */ | |
1091 | vect_slp_rearrange_stmts (SLP_INSTANCE_TREE (slp_instn), group_size, | |
1092 | load_permutation); | |
1093 | VEC_free (int, heap, SLP_INSTANCE_LOAD_PERMUTATION (slp_instn)); | |
1094 | return true; | |
1095 | } | |
1096 | } | |
1097 | ||
6aa904c4 IR |
1098 | /* In basic block vectorization we allow any subchain of an interleaving |
1099 | chain. | |
1100 | FORNOW: not supported in loop SLP because of realignment compications. */ | |
1101 | bb_vinfo = STMT_VINFO_BB_VINFO (vinfo_for_stmt (stmt)); | |
1102 | bad_permutation = false; | |
1103 | /* Check that for every node in the instance teh loads form a subchain. */ | |
1104 | if (bb_vinfo) | |
1105 | { | |
1106 | FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (slp_instn), i, node) | |
1107 | { | |
1108 | next_load = NULL; | |
1109 | first_load = NULL; | |
1110 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), j, load) | |
1111 | { | |
1112 | if (!first_load) | |
1113 | first_load = GROUP_FIRST_ELEMENT (vinfo_for_stmt (load)); | |
1114 | else if (first_load | |
1115 | != GROUP_FIRST_ELEMENT (vinfo_for_stmt (load))) | |
1116 | { | |
1117 | bad_permutation = true; | |
1118 | break; | |
1119 | } | |
1120 | ||
1121 | if (j != 0 && next_load != load) | |
1122 | { | |
1123 | bad_permutation = true; | |
1124 | break; | |
1125 | } | |
1126 | ||
1127 | next_load = GROUP_NEXT_ELEMENT (vinfo_for_stmt (load)); | |
1128 | } | |
1129 | ||
1130 | if (bad_permutation) | |
1131 | break; | |
1132 | } | |
1133 | ||
1134 | /* Check that the alignment of the first load in every subchain, i.e., | |
1135 | the first statement in every load node, is supported. */ | |
1136 | if (!bad_permutation) | |
1137 | { | |
1138 | FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (slp_instn), i, node) | |
1139 | { | |
1140 | first_load = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (node), 0); | |
1141 | if (first_load | |
1142 | != GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_load))) | |
1143 | { | |
1144 | dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_load)); | |
1145 | if (vect_supportable_dr_alignment (dr, false) | |
1146 | == dr_unaligned_unsupported) | |
1147 | { | |
1148 | if (vect_print_dump_info (REPORT_SLP)) | |
1149 | { | |
1150 | fprintf (vect_dump, "unsupported unaligned load "); | |
1151 | print_gimple_stmt (vect_dump, first_load, 0, | |
1152 | TDF_SLIM); | |
1153 | } | |
1154 | bad_permutation = true; | |
1155 | break; | |
1156 | } | |
1157 | } | |
1158 | } | |
1159 | ||
1160 | if (!bad_permutation) | |
1161 | { | |
1162 | VEC_free (int, heap, SLP_INSTANCE_LOAD_PERMUTATION (slp_instn)); | |
1163 | return true; | |
1164 | } | |
1165 | } | |
1166 | } | |
1167 | ||
b8698a0f L |
1168 | /* FORNOW: the only supported permutation is 0..01..1.. of length equal to |
1169 | GROUP_SIZE and where each sequence of same drs is of GROUP_SIZE length as | |
b5aeb3bb | 1170 | well (unless it's reduction). */ |
ebfd146a IR |
1171 | if (VEC_length (int, load_permutation) |
1172 | != (unsigned int) (group_size * group_size)) | |
1173 | return false; | |
1174 | ||
1175 | supported = true; | |
7417f6c0 IR |
1176 | load_index = sbitmap_alloc (group_size); |
1177 | sbitmap_zero (load_index); | |
ebfd146a IR |
1178 | for (j = 0; j < group_size; j++) |
1179 | { | |
1180 | for (i = j * group_size, k = 0; | |
1181 | VEC_iterate (int, load_permutation, i, next) && k < group_size; | |
1182 | i++, k++) | |
1183 | { | |
1184 | if (i != j * group_size && next != prev) | |
1185 | { | |
1186 | supported = false; | |
1187 | break; | |
1188 | } | |
1189 | ||
1190 | prev = next; | |
b8698a0f | 1191 | } |
7417f6c0 IR |
1192 | |
1193 | if (TEST_BIT (load_index, prev)) | |
1194 | { | |
1195 | supported = false; | |
1196 | break; | |
1197 | } | |
1198 | ||
1199 | SET_BIT (load_index, prev); | |
ebfd146a | 1200 | } |
59eefaa6 IR |
1201 | |
1202 | for (j = 0; j < group_size; j++) | |
1203 | if (!TEST_BIT (load_index, j)) | |
1204 | return false; | |
1205 | ||
7417f6c0 | 1206 | sbitmap_free (load_index); |
ebfd146a IR |
1207 | |
1208 | if (supported && i == group_size * group_size | |
1209 | && vect_supported_slp_permutation_p (slp_instn)) | |
1210 | return true; | |
1211 | ||
b8698a0f | 1212 | return false; |
ebfd146a IR |
1213 | } |
1214 | ||
1215 | ||
b8698a0f | 1216 | /* Find the first load in the loop that belongs to INSTANCE. |
ebfd146a | 1217 | When loads are in several SLP nodes, there can be a case in which the first |
b8698a0f | 1218 | load does not appear in the first SLP node to be transformed, causing |
ff802fa1 | 1219 | incorrect order of statements. Since we generate all the loads together, |
ebfd146a IR |
1220 | they must be inserted before the first load of the SLP instance and not |
1221 | before the first load of the first node of the instance. */ | |
ff802fa1 | 1222 | |
b8698a0f L |
1223 | static gimple |
1224 | vect_find_first_load_in_slp_instance (slp_instance instance) | |
ebfd146a IR |
1225 | { |
1226 | int i, j; | |
1227 | slp_tree load_node; | |
1228 | gimple first_load = NULL, load; | |
1229 | ||
ac47786e NF |
1230 | FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (instance), i, load_node) |
1231 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (load_node), j, load) | |
ebfd146a | 1232 | first_load = get_earlier_stmt (load, first_load); |
b8698a0f | 1233 | |
ebfd146a IR |
1234 | return first_load; |
1235 | } | |
1236 | ||
1237 | ||
e4a707c4 | 1238 | /* Find the last store in SLP INSTANCE. */ |
ff802fa1 | 1239 | |
e4a707c4 IR |
1240 | static gimple |
1241 | vect_find_last_store_in_slp_instance (slp_instance instance) | |
1242 | { | |
1243 | int i; | |
1244 | slp_tree node; | |
1245 | gimple last_store = NULL, store; | |
1246 | ||
1247 | node = SLP_INSTANCE_TREE (instance); | |
1248 | for (i = 0; | |
1249 | VEC_iterate (gimple, SLP_TREE_SCALAR_STMTS (node), i, store); | |
1250 | i++) | |
1251 | last_store = get_later_stmt (store, last_store); | |
1252 | ||
1253 | return last_store; | |
1254 | } | |
1255 | ||
1256 | ||
ff802fa1 | 1257 | /* Analyze an SLP instance starting from a group of strided stores. Call |
b8698a0f | 1258 | vect_build_slp_tree to build a tree of packed stmts if possible. |
ebfd146a IR |
1259 | Return FALSE if it's impossible to SLP any stmt in the loop. */ |
1260 | ||
1261 | static bool | |
a70d6342 IR |
1262 | vect_analyze_slp_instance (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo, |
1263 | gimple stmt) | |
ebfd146a IR |
1264 | { |
1265 | slp_instance new_instance; | |
1266 | slp_tree node = XNEW (struct _slp_tree); | |
e14c1050 | 1267 | unsigned int group_size = GROUP_SIZE (vinfo_for_stmt (stmt)); |
ebfd146a | 1268 | unsigned int unrolling_factor = 1, nunits; |
b5aeb3bb | 1269 | tree vectype, scalar_type = NULL_TREE; |
ebfd146a | 1270 | gimple next; |
0f900dfa | 1271 | unsigned int vectorization_factor = 0; |
b5aeb3bb | 1272 | int inside_cost = 0, outside_cost = 0, ncopies_for_cost, i; |
ebfd146a IR |
1273 | unsigned int max_nunits = 0; |
1274 | VEC (int, heap) *load_permutation; | |
1275 | VEC (slp_tree, heap) *loads; | |
b5aeb3bb | 1276 | struct data_reference *dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt)); |
6aa904c4 | 1277 | bool loads_permuted = false; |
b5aeb3bb | 1278 | |
b010117a | 1279 | if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) |
b5aeb3bb | 1280 | { |
b010117a IR |
1281 | if (dr) |
1282 | { | |
1283 | scalar_type = TREE_TYPE (DR_REF (dr)); | |
1284 | vectype = get_vectype_for_scalar_type (scalar_type); | |
1285 | } | |
1286 | else | |
1287 | { | |
1288 | gcc_assert (loop_vinfo); | |
1289 | vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt)); | |
1290 | } | |
1291 | ||
e14c1050 | 1292 | group_size = GROUP_SIZE (vinfo_for_stmt (stmt)); |
b5aeb3bb IR |
1293 | } |
1294 | else | |
1295 | { | |
1296 | gcc_assert (loop_vinfo); | |
1297 | vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt)); | |
1298 | group_size = VEC_length (gimple, LOOP_VINFO_REDUCTIONS (loop_vinfo)); | |
1299 | } | |
b8698a0f | 1300 | |
ebfd146a IR |
1301 | if (!vectype) |
1302 | { | |
1303 | if (vect_print_dump_info (REPORT_SLP)) | |
1304 | { | |
1305 | fprintf (vect_dump, "Build SLP failed: unsupported data-type "); | |
1306 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
1307 | } | |
b5aeb3bb | 1308 | |
ebfd146a IR |
1309 | return false; |
1310 | } | |
1311 | ||
1312 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
a70d6342 IR |
1313 | if (loop_vinfo) |
1314 | vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
1315 | else | |
a70d6342 IR |
1316 | vectorization_factor = nunits; |
1317 | ||
a70d6342 IR |
1318 | /* Calculate the unrolling factor. */ |
1319 | unrolling_factor = least_common_multiple (nunits, group_size) / group_size; | |
1320 | if (unrolling_factor != 1 && !loop_vinfo) | |
1321 | { | |
1322 | if (vect_print_dump_info (REPORT_SLP)) | |
e9dbe7bb IR |
1323 | fprintf (vect_dump, "Build SLP failed: unrolling required in basic" |
1324 | " block SLP"); | |
b8698a0f | 1325 | |
a70d6342 IR |
1326 | return false; |
1327 | } | |
1328 | ||
b8698a0f | 1329 | /* Create a node (a root of the SLP tree) for the packed strided stores. */ |
ebfd146a IR |
1330 | SLP_TREE_SCALAR_STMTS (node) = VEC_alloc (gimple, heap, group_size); |
1331 | next = stmt; | |
b010117a | 1332 | if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) |
ebfd146a | 1333 | { |
b5aeb3bb IR |
1334 | /* Collect the stores and store them in SLP_TREE_SCALAR_STMTS. */ |
1335 | while (next) | |
1336 | { | |
1337 | VEC_safe_push (gimple, heap, SLP_TREE_SCALAR_STMTS (node), next); | |
e14c1050 | 1338 | next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next)); |
b5aeb3bb IR |
1339 | } |
1340 | } | |
1341 | else | |
1342 | { | |
1343 | /* Collect reduction statements. */ | |
1344 | for (i = 0; VEC_iterate (gimple, LOOP_VINFO_REDUCTIONS (loop_vinfo), i, | |
1345 | next); | |
1346 | i++) | |
b010117a | 1347 | VEC_safe_push (gimple, heap, SLP_TREE_SCALAR_STMTS (node), next); |
ebfd146a IR |
1348 | } |
1349 | ||
1350 | SLP_TREE_VEC_STMTS (node) = NULL; | |
1351 | SLP_TREE_NUMBER_OF_VEC_STMTS (node) = 0; | |
1352 | SLP_TREE_LEFT (node) = NULL; | |
1353 | SLP_TREE_RIGHT (node) = NULL; | |
1354 | SLP_TREE_OUTSIDE_OF_LOOP_COST (node) = 0; | |
1355 | SLP_TREE_INSIDE_OF_LOOP_COST (node) = 0; | |
1356 | ||
ebfd146a IR |
1357 | /* Calculate the number of vector stmts to create based on the unrolling |
1358 | factor (number of vectors is 1 if NUNITS >= GROUP_SIZE, and is | |
1359 | GROUP_SIZE / NUNITS otherwise. */ | |
1360 | ncopies_for_cost = unrolling_factor * group_size / nunits; | |
b8698a0f L |
1361 | |
1362 | load_permutation = VEC_alloc (int, heap, group_size * group_size); | |
1363 | loads = VEC_alloc (slp_tree, heap, group_size); | |
ebfd146a IR |
1364 | |
1365 | /* Build the tree for the SLP instance. */ | |
b8698a0f L |
1366 | if (vect_build_slp_tree (loop_vinfo, bb_vinfo, &node, group_size, |
1367 | &inside_cost, &outside_cost, ncopies_for_cost, | |
1368 | &max_nunits, &load_permutation, &loads, | |
6aa904c4 | 1369 | vectorization_factor, &loads_permuted)) |
ebfd146a | 1370 | { |
4ef69dfc | 1371 | /* Calculate the unrolling factor based on the smallest type. */ |
ebfd146a IR |
1372 | if (max_nunits > nunits) |
1373 | unrolling_factor = least_common_multiple (max_nunits, group_size) | |
1374 | / group_size; | |
b8698a0f | 1375 | |
4ef69dfc IR |
1376 | if (unrolling_factor != 1 && !loop_vinfo) |
1377 | { | |
1378 | if (vect_print_dump_info (REPORT_SLP)) | |
1379 | fprintf (vect_dump, "Build SLP failed: unrolling required in basic" | |
1380 | " block SLP"); | |
1381 | return false; | |
1382 | } | |
1383 | ||
1384 | /* Create a new SLP instance. */ | |
1385 | new_instance = XNEW (struct _slp_instance); | |
1386 | SLP_INSTANCE_TREE (new_instance) = node; | |
1387 | SLP_INSTANCE_GROUP_SIZE (new_instance) = group_size; | |
ebfd146a IR |
1388 | SLP_INSTANCE_UNROLLING_FACTOR (new_instance) = unrolling_factor; |
1389 | SLP_INSTANCE_OUTSIDE_OF_LOOP_COST (new_instance) = outside_cost; | |
1390 | SLP_INSTANCE_INSIDE_OF_LOOP_COST (new_instance) = inside_cost; | |
1391 | SLP_INSTANCE_LOADS (new_instance) = loads; | |
1392 | SLP_INSTANCE_FIRST_LOAD_STMT (new_instance) = NULL; | |
1393 | SLP_INSTANCE_LOAD_PERMUTATION (new_instance) = load_permutation; | |
6aa904c4 IR |
1394 | |
1395 | if (loads_permuted) | |
ebfd146a IR |
1396 | { |
1397 | if (!vect_supported_load_permutation_p (new_instance, group_size, | |
b8698a0f | 1398 | load_permutation)) |
ebfd146a IR |
1399 | { |
1400 | if (vect_print_dump_info (REPORT_SLP)) | |
1401 | { | |
1402 | fprintf (vect_dump, "Build SLP failed: unsupported load " | |
1403 | "permutation "); | |
1404 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
1405 | } | |
1406 | ||
1407 | vect_free_slp_instance (new_instance); | |
1408 | return false; | |
1409 | } | |
1410 | ||
1411 | SLP_INSTANCE_FIRST_LOAD_STMT (new_instance) | |
1412 | = vect_find_first_load_in_slp_instance (new_instance); | |
1413 | } | |
1414 | else | |
1415 | VEC_free (int, heap, SLP_INSTANCE_LOAD_PERMUTATION (new_instance)); | |
1416 | ||
a70d6342 | 1417 | if (loop_vinfo) |
b8698a0f L |
1418 | VEC_safe_push (slp_instance, heap, |
1419 | LOOP_VINFO_SLP_INSTANCES (loop_vinfo), | |
a70d6342 IR |
1420 | new_instance); |
1421 | else | |
1422 | VEC_safe_push (slp_instance, heap, BB_VINFO_SLP_INSTANCES (bb_vinfo), | |
1423 | new_instance); | |
b8698a0f | 1424 | |
ebfd146a IR |
1425 | if (vect_print_dump_info (REPORT_SLP)) |
1426 | vect_print_slp_tree (node); | |
1427 | ||
1428 | return true; | |
1429 | } | |
1430 | ||
1431 | /* Failed to SLP. */ | |
1432 | /* Free the allocated memory. */ | |
1433 | vect_free_slp_tree (node); | |
1434 | VEC_free (int, heap, load_permutation); | |
1435 | VEC_free (slp_tree, heap, loads); | |
b8698a0f | 1436 | |
a70d6342 | 1437 | return false; |
ebfd146a IR |
1438 | } |
1439 | ||
1440 | ||
ff802fa1 | 1441 | /* Check if there are stmts in the loop can be vectorized using SLP. Build SLP |
ebfd146a IR |
1442 | trees of packed scalar stmts if SLP is possible. */ |
1443 | ||
1444 | bool | |
a70d6342 | 1445 | vect_analyze_slp (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo) |
ebfd146a IR |
1446 | { |
1447 | unsigned int i; | |
b010117a IR |
1448 | VEC (gimple, heap) *strided_stores, *reductions = NULL, *reduc_chains = NULL; |
1449 | gimple first_element; | |
a70d6342 | 1450 | bool ok = false; |
ebfd146a IR |
1451 | |
1452 | if (vect_print_dump_info (REPORT_SLP)) | |
1453 | fprintf (vect_dump, "=== vect_analyze_slp ==="); | |
1454 | ||
a70d6342 | 1455 | if (loop_vinfo) |
b5aeb3bb IR |
1456 | { |
1457 | strided_stores = LOOP_VINFO_STRIDED_STORES (loop_vinfo); | |
b010117a | 1458 | reduc_chains = LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo); |
b5aeb3bb IR |
1459 | reductions = LOOP_VINFO_REDUCTIONS (loop_vinfo); |
1460 | } | |
a70d6342 IR |
1461 | else |
1462 | strided_stores = BB_VINFO_STRIDED_STORES (bb_vinfo); | |
b8698a0f | 1463 | |
b5aeb3bb | 1464 | /* Find SLP sequences starting from groups of strided stores. */ |
b010117a IR |
1465 | FOR_EACH_VEC_ELT (gimple, strided_stores, i, first_element) |
1466 | if (vect_analyze_slp_instance (loop_vinfo, bb_vinfo, first_element)) | |
a70d6342 | 1467 | ok = true; |
ebfd146a | 1468 | |
b8698a0f | 1469 | if (bb_vinfo && !ok) |
a70d6342 IR |
1470 | { |
1471 | if (vect_print_dump_info (REPORT_SLP)) | |
1472 | fprintf (vect_dump, "Failed to SLP the basic block."); | |
1473 | ||
1474 | return false; | |
1475 | } | |
ebfd146a | 1476 | |
b010117a IR |
1477 | if (loop_vinfo |
1478 | && VEC_length (gimple, LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo)) > 0) | |
1479 | { | |
1480 | /* Find SLP sequences starting from reduction chains. */ | |
1481 | FOR_EACH_VEC_ELT (gimple, reduc_chains, i, first_element) | |
1482 | if (vect_analyze_slp_instance (loop_vinfo, bb_vinfo, first_element)) | |
1483 | ok = true; | |
1484 | else | |
1485 | return false; | |
1486 | ||
1487 | /* Don't try to vectorize SLP reductions if reduction chain was | |
1488 | detected. */ | |
1489 | return ok; | |
1490 | } | |
1491 | ||
b5aeb3bb | 1492 | /* Find SLP sequences starting from groups of reductions. */ |
74500b3e | 1493 | if (loop_vinfo && VEC_length (gimple, LOOP_VINFO_REDUCTIONS (loop_vinfo)) > 1 |
b5aeb3bb IR |
1494 | && vect_analyze_slp_instance (loop_vinfo, bb_vinfo, |
1495 | VEC_index (gimple, reductions, 0))) | |
1496 | ok = true; | |
1497 | ||
ebfd146a IR |
1498 | return true; |
1499 | } | |
1500 | ||
1501 | ||
1502 | /* For each possible SLP instance decide whether to SLP it and calculate overall | |
437f4a00 IR |
1503 | unrolling factor needed to SLP the loop. Return TRUE if decided to SLP at |
1504 | least one instance. */ | |
ebfd146a | 1505 | |
437f4a00 | 1506 | bool |
ebfd146a IR |
1507 | vect_make_slp_decision (loop_vec_info loop_vinfo) |
1508 | { | |
1509 | unsigned int i, unrolling_factor = 1; | |
1510 | VEC (slp_instance, heap) *slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo); | |
1511 | slp_instance instance; | |
1512 | int decided_to_slp = 0; | |
1513 | ||
1514 | if (vect_print_dump_info (REPORT_SLP)) | |
1515 | fprintf (vect_dump, "=== vect_make_slp_decision ==="); | |
1516 | ||
ac47786e | 1517 | FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance) |
ebfd146a IR |
1518 | { |
1519 | /* FORNOW: SLP if you can. */ | |
1520 | if (unrolling_factor < SLP_INSTANCE_UNROLLING_FACTOR (instance)) | |
1521 | unrolling_factor = SLP_INSTANCE_UNROLLING_FACTOR (instance); | |
1522 | ||
ff802fa1 | 1523 | /* Mark all the stmts that belong to INSTANCE as PURE_SLP stmts. Later we |
b8698a0f | 1524 | call vect_detect_hybrid_slp () to find stmts that need hybrid SLP and |
ff802fa1 | 1525 | loop-based vectorization. Such stmts will be marked as HYBRID. */ |
ebfd146a IR |
1526 | vect_mark_slp_stmts (SLP_INSTANCE_TREE (instance), pure_slp, -1); |
1527 | decided_to_slp++; | |
1528 | } | |
1529 | ||
1530 | LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo) = unrolling_factor; | |
1531 | ||
b8698a0f L |
1532 | if (decided_to_slp && vect_print_dump_info (REPORT_SLP)) |
1533 | fprintf (vect_dump, "Decided to SLP %d instances. Unrolling factor %d", | |
ebfd146a | 1534 | decided_to_slp, unrolling_factor); |
437f4a00 IR |
1535 | |
1536 | return (decided_to_slp > 0); | |
ebfd146a IR |
1537 | } |
1538 | ||
1539 | ||
1540 | /* Find stmts that must be both vectorized and SLPed (since they feed stmts that | |
ff802fa1 | 1541 | can't be SLPed) in the tree rooted at NODE. Mark such stmts as HYBRID. */ |
ebfd146a IR |
1542 | |
1543 | static void | |
1544 | vect_detect_hybrid_slp_stmts (slp_tree node) | |
1545 | { | |
1546 | int i; | |
1547 | gimple stmt; | |
1548 | imm_use_iterator imm_iter; | |
1549 | gimple use_stmt; | |
99f51320 | 1550 | stmt_vec_info stmt_vinfo; |
ebfd146a IR |
1551 | |
1552 | if (!node) | |
1553 | return; | |
1554 | ||
ac47786e | 1555 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt) |
ebfd146a IR |
1556 | if (PURE_SLP_STMT (vinfo_for_stmt (stmt)) |
1557 | && TREE_CODE (gimple_op (stmt, 0)) == SSA_NAME) | |
1558 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, gimple_op (stmt, 0)) | |
99f51320 IR |
1559 | if ((stmt_vinfo = vinfo_for_stmt (use_stmt)) |
1560 | && !STMT_SLP_TYPE (stmt_vinfo) | |
1561 | && (STMT_VINFO_RELEVANT (stmt_vinfo) | |
b5aeb3bb IR |
1562 | || VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_vinfo))) |
1563 | && !(gimple_code (use_stmt) == GIMPLE_PHI | |
1564 | && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (use_stmt)) | |
1565 | == vect_reduction_def)) | |
ebfd146a IR |
1566 | vect_mark_slp_stmts (node, hybrid, i); |
1567 | ||
1568 | vect_detect_hybrid_slp_stmts (SLP_TREE_LEFT (node)); | |
1569 | vect_detect_hybrid_slp_stmts (SLP_TREE_RIGHT (node)); | |
1570 | } | |
1571 | ||
1572 | ||
1573 | /* Find stmts that must be both vectorized and SLPed. */ | |
1574 | ||
1575 | void | |
1576 | vect_detect_hybrid_slp (loop_vec_info loop_vinfo) | |
1577 | { | |
1578 | unsigned int i; | |
1579 | VEC (slp_instance, heap) *slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo); | |
1580 | slp_instance instance; | |
1581 | ||
1582 | if (vect_print_dump_info (REPORT_SLP)) | |
1583 | fprintf (vect_dump, "=== vect_detect_hybrid_slp ==="); | |
1584 | ||
ac47786e | 1585 | FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance) |
ebfd146a IR |
1586 | vect_detect_hybrid_slp_stmts (SLP_INSTANCE_TREE (instance)); |
1587 | } | |
1588 | ||
a70d6342 IR |
1589 | |
1590 | /* Create and initialize a new bb_vec_info struct for BB, as well as | |
1591 | stmt_vec_info structs for all the stmts in it. */ | |
b8698a0f | 1592 | |
a70d6342 IR |
1593 | static bb_vec_info |
1594 | new_bb_vec_info (basic_block bb) | |
1595 | { | |
1596 | bb_vec_info res = NULL; | |
1597 | gimple_stmt_iterator gsi; | |
1598 | ||
1599 | res = (bb_vec_info) xcalloc (1, sizeof (struct _bb_vec_info)); | |
1600 | BB_VINFO_BB (res) = bb; | |
1601 | ||
1602 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1603 | { | |
1604 | gimple stmt = gsi_stmt (gsi); | |
1605 | gimple_set_uid (stmt, 0); | |
1606 | set_vinfo_for_stmt (stmt, new_stmt_vec_info (stmt, NULL, res)); | |
1607 | } | |
1608 | ||
1609 | BB_VINFO_STRIDED_STORES (res) = VEC_alloc (gimple, heap, 10); | |
1610 | BB_VINFO_SLP_INSTANCES (res) = VEC_alloc (slp_instance, heap, 2); | |
1611 | ||
1612 | bb->aux = res; | |
1613 | return res; | |
1614 | } | |
1615 | ||
1616 | ||
1617 | /* Free BB_VINFO struct, as well as all the stmt_vec_info structs of all the | |
1618 | stmts in the basic block. */ | |
1619 | ||
1620 | static void | |
1621 | destroy_bb_vec_info (bb_vec_info bb_vinfo) | |
1622 | { | |
1623 | basic_block bb; | |
1624 | gimple_stmt_iterator si; | |
1625 | ||
1626 | if (!bb_vinfo) | |
1627 | return; | |
1628 | ||
1629 | bb = BB_VINFO_BB (bb_vinfo); | |
1630 | ||
1631 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
1632 | { | |
1633 | gimple stmt = gsi_stmt (si); | |
1634 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1635 | ||
1636 | if (stmt_info) | |
1637 | /* Free stmt_vec_info. */ | |
1638 | free_stmt_vec_info (stmt); | |
1639 | } | |
1640 | ||
01be8516 SP |
1641 | free_data_refs (BB_VINFO_DATAREFS (bb_vinfo)); |
1642 | free_dependence_relations (BB_VINFO_DDRS (bb_vinfo)); | |
a70d6342 IR |
1643 | VEC_free (gimple, heap, BB_VINFO_STRIDED_STORES (bb_vinfo)); |
1644 | VEC_free (slp_instance, heap, BB_VINFO_SLP_INSTANCES (bb_vinfo)); | |
1645 | free (bb_vinfo); | |
1646 | bb->aux = NULL; | |
1647 | } | |
1648 | ||
1649 | ||
1650 | /* Analyze statements contained in SLP tree node after recursively analyzing | |
1651 | the subtree. Return TRUE if the operations are supported. */ | |
1652 | ||
1653 | static bool | |
1654 | vect_slp_analyze_node_operations (bb_vec_info bb_vinfo, slp_tree node) | |
1655 | { | |
1656 | bool dummy; | |
1657 | int i; | |
1658 | gimple stmt; | |
1659 | ||
1660 | if (!node) | |
1661 | return true; | |
1662 | ||
1663 | if (!vect_slp_analyze_node_operations (bb_vinfo, SLP_TREE_LEFT (node)) | |
1664 | || !vect_slp_analyze_node_operations (bb_vinfo, SLP_TREE_RIGHT (node))) | |
1665 | return false; | |
1666 | ||
ac47786e | 1667 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt) |
a70d6342 IR |
1668 | { |
1669 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1670 | gcc_assert (stmt_info); | |
1671 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
1672 | ||
1673 | if (!vect_analyze_stmt (stmt, &dummy, node)) | |
1674 | return false; | |
1675 | } | |
1676 | ||
1677 | return true; | |
1678 | } | |
1679 | ||
1680 | ||
ff802fa1 | 1681 | /* Analyze statements in SLP instances of the basic block. Return TRUE if the |
a70d6342 IR |
1682 | operations are supported. */ |
1683 | ||
1684 | static bool | |
1685 | vect_slp_analyze_operations (bb_vec_info bb_vinfo) | |
1686 | { | |
1687 | VEC (slp_instance, heap) *slp_instances = BB_VINFO_SLP_INSTANCES (bb_vinfo); | |
1688 | slp_instance instance; | |
1689 | int i; | |
1690 | ||
1691 | for (i = 0; VEC_iterate (slp_instance, slp_instances, i, instance); ) | |
1692 | { | |
b8698a0f | 1693 | if (!vect_slp_analyze_node_operations (bb_vinfo, |
a70d6342 IR |
1694 | SLP_INSTANCE_TREE (instance))) |
1695 | { | |
1696 | vect_free_slp_instance (instance); | |
1697 | VEC_ordered_remove (slp_instance, slp_instances, i); | |
1698 | } | |
1699 | else | |
1700 | i++; | |
b8698a0f L |
1701 | } |
1702 | ||
a70d6342 IR |
1703 | if (!VEC_length (slp_instance, slp_instances)) |
1704 | return false; | |
1705 | ||
1706 | return true; | |
1707 | } | |
1708 | ||
69f11a13 IR |
1709 | /* Check if vectorization of the basic block is profitable. */ |
1710 | ||
1711 | static bool | |
1712 | vect_bb_vectorization_profitable_p (bb_vec_info bb_vinfo) | |
1713 | { | |
1714 | VEC (slp_instance, heap) *slp_instances = BB_VINFO_SLP_INSTANCES (bb_vinfo); | |
1715 | slp_instance instance; | |
1716 | int i; | |
1717 | unsigned int vec_outside_cost = 0, vec_inside_cost = 0, scalar_cost = 0; | |
1718 | unsigned int stmt_cost; | |
1719 | gimple stmt; | |
1720 | gimple_stmt_iterator si; | |
1721 | basic_block bb = BB_VINFO_BB (bb_vinfo); | |
1722 | stmt_vec_info stmt_info = NULL; | |
1723 | tree dummy_type = NULL; | |
1724 | int dummy = 0; | |
1725 | ||
1726 | /* Calculate vector costs. */ | |
ac47786e | 1727 | FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance) |
69f11a13 IR |
1728 | { |
1729 | vec_outside_cost += SLP_INSTANCE_OUTSIDE_OF_LOOP_COST (instance); | |
1730 | vec_inside_cost += SLP_INSTANCE_INSIDE_OF_LOOP_COST (instance); | |
1731 | } | |
1732 | ||
1733 | /* Calculate scalar cost. */ | |
1734 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
1735 | { | |
1736 | stmt = gsi_stmt (si); | |
1737 | stmt_info = vinfo_for_stmt (stmt); | |
1738 | ||
1739 | if (!stmt_info || !STMT_VINFO_VECTORIZABLE (stmt_info) | |
1740 | || !PURE_SLP_STMT (stmt_info)) | |
1741 | continue; | |
1742 | ||
1743 | if (STMT_VINFO_DATA_REF (stmt_info)) | |
1744 | { | |
1745 | if (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info))) | |
1746 | stmt_cost = targetm.vectorize.builtin_vectorization_cost | |
1747 | (scalar_load, dummy_type, dummy); | |
1748 | else | |
1749 | stmt_cost = targetm.vectorize.builtin_vectorization_cost | |
1750 | (scalar_store, dummy_type, dummy); | |
1751 | } | |
1752 | else | |
1753 | stmt_cost = targetm.vectorize.builtin_vectorization_cost | |
1754 | (scalar_stmt, dummy_type, dummy); | |
1755 | ||
1756 | scalar_cost += stmt_cost; | |
1757 | } | |
1758 | ||
1759 | if (vect_print_dump_info (REPORT_COST)) | |
1760 | { | |
1761 | fprintf (vect_dump, "Cost model analysis: \n"); | |
1762 | fprintf (vect_dump, " Vector inside of basic block cost: %d\n", | |
1763 | vec_inside_cost); | |
1764 | fprintf (vect_dump, " Vector outside of basic block cost: %d\n", | |
1765 | vec_outside_cost); | |
1766 | fprintf (vect_dump, " Scalar cost of basic block: %d", scalar_cost); | |
1767 | } | |
1768 | ||
1769 | /* Vectorization is profitable if its cost is less than the cost of scalar | |
1770 | version. */ | |
1771 | if (vec_outside_cost + vec_inside_cost >= scalar_cost) | |
1772 | return false; | |
1773 | ||
1774 | return true; | |
1775 | } | |
1776 | ||
1777 | /* Check if the basic block can be vectorized. */ | |
a70d6342 | 1778 | |
8e19f5a1 IR |
1779 | static bb_vec_info |
1780 | vect_slp_analyze_bb_1 (basic_block bb) | |
a70d6342 IR |
1781 | { |
1782 | bb_vec_info bb_vinfo; | |
1783 | VEC (ddr_p, heap) *ddrs; | |
1784 | VEC (slp_instance, heap) *slp_instances; | |
1785 | slp_instance instance; | |
8e19f5a1 | 1786 | int i; |
777e1f09 RG |
1787 | int min_vf = 2; |
1788 | int max_vf = MAX_VECTORIZATION_FACTOR; | |
e4a707c4 | 1789 | |
a70d6342 IR |
1790 | bb_vinfo = new_bb_vec_info (bb); |
1791 | if (!bb_vinfo) | |
1792 | return NULL; | |
1793 | ||
777e1f09 | 1794 | if (!vect_analyze_data_refs (NULL, bb_vinfo, &min_vf)) |
a70d6342 IR |
1795 | { |
1796 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
1797 | fprintf (vect_dump, "not vectorized: unhandled data-ref in basic " | |
1798 | "block.\n"); | |
b8698a0f | 1799 | |
a70d6342 IR |
1800 | destroy_bb_vec_info (bb_vinfo); |
1801 | return NULL; | |
1802 | } | |
1803 | ||
1804 | ddrs = BB_VINFO_DDRS (bb_vinfo); | |
b8698a0f | 1805 | if (!VEC_length (ddr_p, ddrs)) |
a70d6342 IR |
1806 | { |
1807 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
1808 | fprintf (vect_dump, "not vectorized: not enough data-refs in basic " | |
1809 | "block.\n"); | |
1810 | ||
1811 | destroy_bb_vec_info (bb_vinfo); | |
1812 | return NULL; | |
1813 | } | |
1814 | ||
5bfdb7d8 IR |
1815 | if (!vect_analyze_data_ref_dependences (NULL, bb_vinfo, &max_vf) |
1816 | || min_vf > max_vf) | |
777e1f09 RG |
1817 | { |
1818 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
1819 | fprintf (vect_dump, "not vectorized: unhandled data dependence " | |
1820 | "in basic block.\n"); | |
1821 | ||
1822 | destroy_bb_vec_info (bb_vinfo); | |
1823 | return NULL; | |
1824 | } | |
1825 | ||
a70d6342 IR |
1826 | if (!vect_analyze_data_refs_alignment (NULL, bb_vinfo)) |
1827 | { | |
1828 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
1829 | fprintf (vect_dump, "not vectorized: bad data alignment in basic " | |
1830 | "block.\n"); | |
b8698a0f | 1831 | |
a70d6342 IR |
1832 | destroy_bb_vec_info (bb_vinfo); |
1833 | return NULL; | |
1834 | } | |
b8698a0f | 1835 | |
a70d6342 IR |
1836 | if (!vect_analyze_data_ref_accesses (NULL, bb_vinfo)) |
1837 | { | |
1838 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
1839 | fprintf (vect_dump, "not vectorized: unhandled data access in basic " | |
1840 | "block.\n"); | |
b8698a0f | 1841 | |
a70d6342 IR |
1842 | destroy_bb_vec_info (bb_vinfo); |
1843 | return NULL; | |
1844 | } | |
1845 | ||
1846 | if (!vect_verify_datarefs_alignment (NULL, bb_vinfo)) | |
1847 | { | |
1848 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
1849 | fprintf (vect_dump, "not vectorized: unsupported alignment in basic " | |
1850 | "block.\n"); | |
1851 | ||
1852 | destroy_bb_vec_info (bb_vinfo); | |
1853 | return NULL; | |
1854 | } | |
1855 | ||
1856 | /* Check the SLP opportunities in the basic block, analyze and build SLP | |
1857 | trees. */ | |
1858 | if (!vect_analyze_slp (NULL, bb_vinfo)) | |
1859 | { | |
1860 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
1861 | fprintf (vect_dump, "not vectorized: failed to find SLP opportunities " | |
1862 | "in basic block.\n"); | |
1863 | ||
1864 | destroy_bb_vec_info (bb_vinfo); | |
1865 | return NULL; | |
1866 | } | |
b8698a0f | 1867 | |
a70d6342 IR |
1868 | slp_instances = BB_VINFO_SLP_INSTANCES (bb_vinfo); |
1869 | ||
1870 | /* Mark all the statements that we want to vectorize as pure SLP and | |
1871 | relevant. */ | |
ac47786e | 1872 | FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance) |
a70d6342 IR |
1873 | { |
1874 | vect_mark_slp_stmts (SLP_INSTANCE_TREE (instance), pure_slp, -1); | |
1875 | vect_mark_slp_stmts_relevant (SLP_INSTANCE_TREE (instance)); | |
b8698a0f | 1876 | } |
a70d6342 IR |
1877 | |
1878 | if (!vect_slp_analyze_operations (bb_vinfo)) | |
1879 | { | |
1880 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
1881 | fprintf (vect_dump, "not vectorized: bad operation in basic block.\n"); | |
1882 | ||
1883 | destroy_bb_vec_info (bb_vinfo); | |
1884 | return NULL; | |
1885 | } | |
1886 | ||
69f11a13 IR |
1887 | /* Cost model: check if the vectorization is worthwhile. */ |
1888 | if (flag_vect_cost_model | |
1889 | && !vect_bb_vectorization_profitable_p (bb_vinfo)) | |
1890 | { | |
1891 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
1892 | fprintf (vect_dump, "not vectorized: vectorization is not " | |
1893 | "profitable.\n"); | |
1894 | ||
1895 | destroy_bb_vec_info (bb_vinfo); | |
1896 | return NULL; | |
1897 | } | |
1898 | ||
a70d6342 | 1899 | if (vect_print_dump_info (REPORT_DETAILS)) |
e9dbe7bb | 1900 | fprintf (vect_dump, "Basic block will be vectorized using SLP\n"); |
a70d6342 IR |
1901 | |
1902 | return bb_vinfo; | |
1903 | } | |
1904 | ||
1905 | ||
8e19f5a1 IR |
1906 | bb_vec_info |
1907 | vect_slp_analyze_bb (basic_block bb) | |
1908 | { | |
1909 | bb_vec_info bb_vinfo; | |
1910 | int insns = 0; | |
1911 | gimple_stmt_iterator gsi; | |
1912 | unsigned int vector_sizes; | |
1913 | ||
1914 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1915 | fprintf (vect_dump, "===vect_slp_analyze_bb===\n"); | |
1916 | ||
1917 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1918 | { | |
1919 | gimple stmt = gsi_stmt (gsi); | |
1920 | if (!is_gimple_debug (stmt) | |
1921 | && !gimple_nop_p (stmt) | |
1922 | && gimple_code (stmt) != GIMPLE_LABEL) | |
1923 | insns++; | |
1924 | } | |
1925 | ||
1926 | if (insns > PARAM_VALUE (PARAM_SLP_MAX_INSNS_IN_BB)) | |
1927 | { | |
1928 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
1929 | fprintf (vect_dump, "not vectorized: too many instructions in basic " | |
1930 | "block.\n"); | |
1931 | ||
1932 | return NULL; | |
1933 | } | |
1934 | ||
1935 | /* Autodetect first vector size we try. */ | |
1936 | current_vector_size = 0; | |
1937 | vector_sizes = targetm.vectorize.autovectorize_vector_sizes (); | |
1938 | ||
1939 | while (1) | |
1940 | { | |
1941 | bb_vinfo = vect_slp_analyze_bb_1 (bb); | |
1942 | if (bb_vinfo) | |
1943 | return bb_vinfo; | |
1944 | ||
1945 | destroy_bb_vec_info (bb_vinfo); | |
1946 | ||
1947 | vector_sizes &= ~current_vector_size; | |
1948 | if (vector_sizes == 0 | |
1949 | || current_vector_size == 0) | |
1950 | return NULL; | |
1951 | ||
1952 | /* Try the next biggest vector size. */ | |
1953 | current_vector_size = 1 << floor_log2 (vector_sizes); | |
1954 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1955 | fprintf (vect_dump, "***** Re-trying analysis with " | |
1956 | "vector size %d\n", current_vector_size); | |
1957 | } | |
1958 | } | |
1959 | ||
1960 | ||
b8698a0f | 1961 | /* SLP costs are calculated according to SLP instance unrolling factor (i.e., |
ff802fa1 IR |
1962 | the number of created vector stmts depends on the unrolling factor). |
1963 | However, the actual number of vector stmts for every SLP node depends on | |
1964 | VF which is set later in vect_analyze_operations (). Hence, SLP costs | |
1965 | should be updated. In this function we assume that the inside costs | |
1966 | calculated in vect_model_xxx_cost are linear in ncopies. */ | |
ebfd146a IR |
1967 | |
1968 | void | |
1969 | vect_update_slp_costs_according_to_vf (loop_vec_info loop_vinfo) | |
1970 | { | |
1971 | unsigned int i, vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
1972 | VEC (slp_instance, heap) *slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo); | |
1973 | slp_instance instance; | |
1974 | ||
1975 | if (vect_print_dump_info (REPORT_SLP)) | |
1976 | fprintf (vect_dump, "=== vect_update_slp_costs_according_to_vf ==="); | |
1977 | ||
ac47786e | 1978 | FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance) |
ebfd146a | 1979 | /* We assume that costs are linear in ncopies. */ |
b8698a0f L |
1980 | SLP_INSTANCE_INSIDE_OF_LOOP_COST (instance) *= vf |
1981 | / SLP_INSTANCE_UNROLLING_FACTOR (instance); | |
ebfd146a IR |
1982 | } |
1983 | ||
a70d6342 | 1984 | |
b8698a0f L |
1985 | /* For constant and loop invariant defs of SLP_NODE this function returns |
1986 | (vector) defs (VEC_OPRNDS) that will be used in the vectorized stmts. | |
d59dc888 IR |
1987 | OP_NUM determines if we gather defs for operand 0 or operand 1 of the RHS of |
1988 | scalar stmts. NUMBER_OF_VECTORS is the number of vector defs to create. | |
b5aeb3bb IR |
1989 | REDUC_INDEX is the index of the reduction operand in the statements, unless |
1990 | it is -1. */ | |
ebfd146a IR |
1991 | |
1992 | static void | |
9dc3f7de IR |
1993 | vect_get_constant_vectors (tree op, slp_tree slp_node, |
1994 | VEC (tree, heap) **vec_oprnds, | |
b5aeb3bb IR |
1995 | unsigned int op_num, unsigned int number_of_vectors, |
1996 | int reduc_index) | |
ebfd146a IR |
1997 | { |
1998 | VEC (gimple, heap) *stmts = SLP_TREE_SCALAR_STMTS (slp_node); | |
1999 | gimple stmt = VEC_index (gimple, stmts, 0); | |
2000 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
ebfd146a IR |
2001 | int nunits; |
2002 | tree vec_cst; | |
2003 | tree t = NULL_TREE; | |
2004 | int j, number_of_places_left_in_vector; | |
2005 | tree vector_type; | |
9dc3f7de | 2006 | tree vop; |
ebfd146a IR |
2007 | int group_size = VEC_length (gimple, stmts); |
2008 | unsigned int vec_num, i; | |
2009 | int number_of_copies = 1; | |
2010 | VEC (tree, heap) *voprnds = VEC_alloc (tree, heap, number_of_vectors); | |
2011 | bool constant_p, is_store; | |
b5aeb3bb | 2012 | tree neutral_op = NULL; |
d59dc888 | 2013 | enum tree_code code = gimple_assign_rhs_code (stmt); |
0e93a64e IR |
2014 | gimple def_stmt; |
2015 | struct loop *loop; | |
b5aeb3bb | 2016 | |
29ed4920 IR |
2017 | if (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
2018 | && reduc_index != -1) | |
b5aeb3bb | 2019 | { |
b5aeb3bb | 2020 | op_num = reduc_index - 1; |
9dc3f7de | 2021 | op = gimple_op (stmt, reduc_index); |
b5aeb3bb | 2022 | /* For additional copies (see the explanation of NUMBER_OF_COPIES below) |
ff802fa1 | 2023 | we need either neutral operands or the original operands. See |
b5aeb3bb IR |
2024 | get_initial_def_for_reduction() for details. */ |
2025 | switch (code) | |
2026 | { | |
2027 | case WIDEN_SUM_EXPR: | |
2028 | case DOT_PROD_EXPR: | |
2029 | case PLUS_EXPR: | |
2030 | case MINUS_EXPR: | |
2031 | case BIT_IOR_EXPR: | |
2032 | case BIT_XOR_EXPR: | |
2033 | if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (op))) | |
2034 | neutral_op = build_real (TREE_TYPE (op), dconst0); | |
2035 | else | |
2036 | neutral_op = build_int_cst (TREE_TYPE (op), 0); | |
2037 | ||
2038 | break; | |
2039 | ||
2040 | case MULT_EXPR: | |
b5aeb3bb IR |
2041 | if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (op))) |
2042 | neutral_op = build_real (TREE_TYPE (op), dconst1); | |
2043 | else | |
2044 | neutral_op = build_int_cst (TREE_TYPE (op), 1); | |
2045 | ||
2046 | break; | |
2047 | ||
c1e822d5 IR |
2048 | case BIT_AND_EXPR: |
2049 | neutral_op = build_int_cst (TREE_TYPE (op), -1); | |
2050 | break; | |
2051 | ||
0e93a64e IR |
2052 | case MAX_EXPR: |
2053 | case MIN_EXPR: | |
2054 | def_stmt = SSA_NAME_DEF_STMT (op); | |
2055 | loop = (gimple_bb (stmt))->loop_father; | |
2056 | neutral_op = PHI_ARG_DEF_FROM_EDGE (def_stmt, | |
2057 | loop_preheader_edge (loop)); | |
2058 | break; | |
2059 | ||
b5aeb3bb | 2060 | default: |
0e93a64e | 2061 | neutral_op = NULL; |
b5aeb3bb IR |
2062 | } |
2063 | } | |
ebfd146a IR |
2064 | |
2065 | if (STMT_VINFO_DATA_REF (stmt_vinfo)) | |
2066 | { | |
2067 | is_store = true; | |
2068 | op = gimple_assign_rhs1 (stmt); | |
2069 | } | |
2070 | else | |
9dc3f7de IR |
2071 | is_store = false; |
2072 | ||
2073 | gcc_assert (op); | |
ebfd146a IR |
2074 | |
2075 | if (CONSTANT_CLASS_P (op)) | |
d59dc888 | 2076 | constant_p = true; |
ebfd146a | 2077 | else |
d59dc888 IR |
2078 | constant_p = false; |
2079 | ||
9dc3f7de | 2080 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
cd481d83 | 2081 | gcc_assert (vector_type); |
ebfd146a IR |
2082 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); |
2083 | ||
2084 | /* NUMBER_OF_COPIES is the number of times we need to use the same values in | |
b8698a0f | 2085 | created vectors. It is greater than 1 if unrolling is performed. |
ebfd146a IR |
2086 | |
2087 | For example, we have two scalar operands, s1 and s2 (e.g., group of | |
2088 | strided accesses of size two), while NUNITS is four (i.e., four scalars | |
2089 | of this type can be packed in a vector). The output vector will contain | |
2090 | two copies of each scalar operand: {s1, s2, s1, s2}. (NUMBER_OF_COPIES | |
2091 | will be 2). | |
2092 | ||
b8698a0f | 2093 | If GROUP_SIZE > NUNITS, the scalars will be split into several vectors |
ebfd146a IR |
2094 | containing the operands. |
2095 | ||
2096 | For example, NUNITS is four as before, and the group size is 8 | |
2097 | (s1, s2, ..., s8). We will create two vectors {s1, s2, s3, s4} and | |
2098 | {s5, s6, s7, s8}. */ | |
b8698a0f | 2099 | |
ebfd146a IR |
2100 | number_of_copies = least_common_multiple (nunits, group_size) / group_size; |
2101 | ||
2102 | number_of_places_left_in_vector = nunits; | |
2103 | for (j = 0; j < number_of_copies; j++) | |
2104 | { | |
2105 | for (i = group_size - 1; VEC_iterate (gimple, stmts, i, stmt); i--) | |
2106 | { | |
2107 | if (is_store) | |
2108 | op = gimple_assign_rhs1 (stmt); | |
2109 | else | |
2110 | op = gimple_op (stmt, op_num + 1); | |
b8698a0f | 2111 | |
b5aeb3bb IR |
2112 | if (reduc_index != -1) |
2113 | { | |
0e93a64e IR |
2114 | loop = (gimple_bb (stmt))->loop_father; |
2115 | def_stmt = SSA_NAME_DEF_STMT (op); | |
b5aeb3bb IR |
2116 | |
2117 | gcc_assert (loop); | |
b010117a IR |
2118 | |
2119 | /* Get the def before the loop. In reduction chain we have only | |
2120 | one initial value. */ | |
2121 | if ((j != (number_of_copies - 1) | |
2122 | || (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) | |
2123 | && i != 0)) | |
2124 | && neutral_op) | |
b5aeb3bb | 2125 | op = neutral_op; |
b010117a IR |
2126 | else |
2127 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, | |
2128 | loop_preheader_edge (loop)); | |
b5aeb3bb IR |
2129 | } |
2130 | ||
ebfd146a IR |
2131 | /* Create 'vect_ = {op0,op1,...,opn}'. */ |
2132 | t = tree_cons (NULL_TREE, op, t); | |
2133 | ||
2134 | number_of_places_left_in_vector--; | |
2135 | ||
2136 | if (number_of_places_left_in_vector == 0) | |
2137 | { | |
2138 | number_of_places_left_in_vector = nunits; | |
2139 | ||
2140 | if (constant_p) | |
2141 | vec_cst = build_vector (vector_type, t); | |
2142 | else | |
2143 | vec_cst = build_constructor_from_list (vector_type, t); | |
2144 | VEC_quick_push (tree, voprnds, | |
2145 | vect_init_vector (stmt, vec_cst, vector_type, NULL)); | |
2146 | t = NULL_TREE; | |
2147 | } | |
2148 | } | |
2149 | } | |
2150 | ||
b8698a0f | 2151 | /* Since the vectors are created in the reverse order, we should invert |
ebfd146a IR |
2152 | them. */ |
2153 | vec_num = VEC_length (tree, voprnds); | |
2154 | for (j = vec_num - 1; j >= 0; j--) | |
2155 | { | |
2156 | vop = VEC_index (tree, voprnds, j); | |
2157 | VEC_quick_push (tree, *vec_oprnds, vop); | |
2158 | } | |
2159 | ||
2160 | VEC_free (tree, heap, voprnds); | |
2161 | ||
2162 | /* In case that VF is greater than the unrolling factor needed for the SLP | |
b8698a0f L |
2163 | group of stmts, NUMBER_OF_VECTORS to be created is greater than |
2164 | NUMBER_OF_SCALARS/NUNITS or NUNITS/NUMBER_OF_SCALARS, and hence we have | |
ebfd146a IR |
2165 | to replicate the vectors. */ |
2166 | while (number_of_vectors > VEC_length (tree, *vec_oprnds)) | |
2167 | { | |
b5aeb3bb IR |
2168 | tree neutral_vec = NULL; |
2169 | ||
2170 | if (neutral_op) | |
2171 | { | |
2172 | if (!neutral_vec) | |
b9acc9f1 | 2173 | neutral_vec = build_vector_from_val (vector_type, neutral_op); |
b5aeb3bb IR |
2174 | |
2175 | VEC_quick_push (tree, *vec_oprnds, neutral_vec); | |
2176 | } | |
2177 | else | |
2178 | { | |
2179 | for (i = 0; VEC_iterate (tree, *vec_oprnds, i, vop) && i < vec_num; i++) | |
2180 | VEC_quick_push (tree, *vec_oprnds, vop); | |
2181 | } | |
ebfd146a IR |
2182 | } |
2183 | } | |
2184 | ||
2185 | ||
2186 | /* Get vectorized definitions from SLP_NODE that contains corresponding | |
2187 | vectorized def-stmts. */ | |
2188 | ||
2189 | static void | |
2190 | vect_get_slp_vect_defs (slp_tree slp_node, VEC (tree,heap) **vec_oprnds) | |
2191 | { | |
2192 | tree vec_oprnd; | |
2193 | gimple vec_def_stmt; | |
2194 | unsigned int i; | |
2195 | ||
2196 | gcc_assert (SLP_TREE_VEC_STMTS (slp_node)); | |
2197 | ||
ac47786e | 2198 | FOR_EACH_VEC_ELT (gimple, SLP_TREE_VEC_STMTS (slp_node), i, vec_def_stmt) |
ebfd146a IR |
2199 | { |
2200 | gcc_assert (vec_def_stmt); | |
2201 | vec_oprnd = gimple_get_lhs (vec_def_stmt); | |
2202 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
2203 | } | |
2204 | } | |
2205 | ||
2206 | ||
b8698a0f L |
2207 | /* Get vectorized definitions for SLP_NODE. |
2208 | If the scalar definitions are loop invariants or constants, collect them and | |
ebfd146a IR |
2209 | call vect_get_constant_vectors() to create vector stmts. |
2210 | Otherwise, the def-stmts must be already vectorized and the vectorized stmts | |
2211 | must be stored in the LEFT/RIGHT node of SLP_NODE, and we call | |
b8698a0f | 2212 | vect_get_slp_vect_defs() to retrieve them. |
ebfd146a | 2213 | If VEC_OPRNDS1 is NULL, don't get vector defs for the second operand (from |
b8698a0f L |
2214 | the right node. This is used when the second operand must remain scalar. */ |
2215 | ||
ebfd146a | 2216 | void |
9dc3f7de IR |
2217 | vect_get_slp_defs (tree op0, tree op1, slp_tree slp_node, |
2218 | VEC (tree,heap) **vec_oprnds0, | |
b5aeb3bb | 2219 | VEC (tree,heap) **vec_oprnds1, int reduc_index) |
ebfd146a IR |
2220 | { |
2221 | gimple first_stmt; | |
2222 | enum tree_code code; | |
2223 | int number_of_vects; | |
b8698a0f | 2224 | HOST_WIDE_INT lhs_size_unit, rhs_size_unit; |
ebfd146a IR |
2225 | |
2226 | first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0); | |
2227 | /* The number of vector defs is determined by the number of vector statements | |
2228 | in the node from which we get those statements. */ | |
b8698a0f | 2229 | if (SLP_TREE_LEFT (slp_node)) |
ebfd146a IR |
2230 | number_of_vects = SLP_TREE_NUMBER_OF_VEC_STMTS (SLP_TREE_LEFT (slp_node)); |
2231 | else | |
2232 | { | |
2233 | number_of_vects = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
2234 | /* Number of vector stmts was calculated according to LHS in | |
2235 | vect_schedule_slp_instance(), fix it by replacing LHS with RHS, if | |
ff802fa1 | 2236 | necessary. See vect_get_smallest_scalar_type () for details. */ |
ebfd146a IR |
2237 | vect_get_smallest_scalar_type (first_stmt, &lhs_size_unit, |
2238 | &rhs_size_unit); | |
2239 | if (rhs_size_unit != lhs_size_unit) | |
2240 | { | |
2241 | number_of_vects *= rhs_size_unit; | |
2242 | number_of_vects /= lhs_size_unit; | |
2243 | } | |
2244 | } | |
2245 | ||
2246 | /* Allocate memory for vectorized defs. */ | |
2247 | *vec_oprnds0 = VEC_alloc (tree, heap, number_of_vects); | |
2248 | ||
2249 | /* SLP_NODE corresponds either to a group of stores or to a group of | |
ff802fa1 | 2250 | unary/binary operations. We don't call this function for loads. |
b5aeb3bb IR |
2251 | For reduction defs we call vect_get_constant_vectors(), since we are |
2252 | looking for initial loop invariant values. */ | |
2253 | if (SLP_TREE_LEFT (slp_node) && reduc_index == -1) | |
ebfd146a IR |
2254 | /* The defs are already vectorized. */ |
2255 | vect_get_slp_vect_defs (SLP_TREE_LEFT (slp_node), vec_oprnds0); | |
2256 | else | |
2257 | /* Build vectors from scalar defs. */ | |
9dc3f7de | 2258 | vect_get_constant_vectors (op0, slp_node, vec_oprnds0, 0, number_of_vects, |
b5aeb3bb | 2259 | reduc_index); |
ebfd146a IR |
2260 | |
2261 | if (STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt))) | |
2262 | /* Since we don't call this function with loads, this is a group of | |
2263 | stores. */ | |
2264 | return; | |
2265 | ||
b5aeb3bb IR |
2266 | /* For reductions, we only need initial values. */ |
2267 | if (reduc_index != -1) | |
2268 | return; | |
2269 | ||
ebfd146a | 2270 | code = gimple_assign_rhs_code (first_stmt); |
29ed4920 | 2271 | if (get_gimple_rhs_class (code) != GIMPLE_BINARY_RHS || !vec_oprnds1 || !op1) |
ebfd146a IR |
2272 | return; |
2273 | ||
2274 | /* The number of vector defs is determined by the number of vector statements | |
2275 | in the node from which we get those statements. */ | |
2276 | if (SLP_TREE_RIGHT (slp_node)) | |
2277 | number_of_vects = SLP_TREE_NUMBER_OF_VEC_STMTS (SLP_TREE_RIGHT (slp_node)); | |
2278 | else | |
2279 | number_of_vects = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
2280 | ||
2281 | *vec_oprnds1 = VEC_alloc (tree, heap, number_of_vects); | |
2282 | ||
2283 | if (SLP_TREE_RIGHT (slp_node)) | |
2284 | /* The defs are already vectorized. */ | |
2285 | vect_get_slp_vect_defs (SLP_TREE_RIGHT (slp_node), vec_oprnds1); | |
2286 | else | |
2287 | /* Build vectors from scalar defs. */ | |
9dc3f7de IR |
2288 | vect_get_constant_vectors (op1, slp_node, vec_oprnds1, 1, number_of_vects, |
2289 | -1); | |
ebfd146a IR |
2290 | } |
2291 | ||
a70d6342 | 2292 | |
b8698a0f | 2293 | /* Create NCOPIES permutation statements using the mask MASK_BYTES (by |
ebfd146a IR |
2294 | building a vector of type MASK_TYPE from it) and two input vectors placed in |
2295 | DR_CHAIN at FIRST_VEC_INDX and SECOND_VEC_INDX for the first copy and | |
2296 | shifting by STRIDE elements of DR_CHAIN for every copy. | |
2297 | (STRIDE is the number of vectorized stmts for NODE divided by the number of | |
b8698a0f | 2298 | copies). |
ebfd146a IR |
2299 | VECT_STMTS_COUNTER specifies the index in the vectorized stmts of NODE, where |
2300 | the created stmts must be inserted. */ | |
2301 | ||
2302 | static inline void | |
b8698a0f | 2303 | vect_create_mask_and_perm (gimple stmt, gimple next_scalar_stmt, |
faf63e39 | 2304 | tree mask, int first_vec_indx, int second_vec_indx, |
b8698a0f | 2305 | gimple_stmt_iterator *gsi, slp_tree node, |
2635892a | 2306 | tree vectype, VEC(tree,heap) *dr_chain, |
ebfd146a IR |
2307 | int ncopies, int vect_stmts_counter) |
2308 | { | |
faf63e39 | 2309 | tree perm_dest; |
ebfd146a IR |
2310 | gimple perm_stmt = NULL; |
2311 | stmt_vec_info next_stmt_info; | |
0f900dfa | 2312 | int i, stride; |
ebfd146a | 2313 | tree first_vec, second_vec, data_ref; |
ebfd146a | 2314 | |
ebfd146a | 2315 | stride = SLP_TREE_NUMBER_OF_VEC_STMTS (node) / ncopies; |
ebfd146a | 2316 | |
b8698a0f | 2317 | /* Initialize the vect stmts of NODE to properly insert the generated |
ebfd146a | 2318 | stmts later. */ |
b8698a0f | 2319 | for (i = VEC_length (gimple, SLP_TREE_VEC_STMTS (node)); |
ebfd146a IR |
2320 | i < (int) SLP_TREE_NUMBER_OF_VEC_STMTS (node); i++) |
2321 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (node), NULL); | |
2322 | ||
2323 | perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype); | |
2324 | for (i = 0; i < ncopies; i++) | |
2325 | { | |
2326 | first_vec = VEC_index (tree, dr_chain, first_vec_indx); | |
2327 | second_vec = VEC_index (tree, dr_chain, second_vec_indx); | |
2328 | ||
ebfd146a | 2329 | /* Generate the permute statement. */ |
2635892a RH |
2330 | perm_stmt = gimple_build_assign_with_ops3 (VEC_PERM_EXPR, perm_dest, |
2331 | first_vec, second_vec, mask); | |
ebfd146a | 2332 | data_ref = make_ssa_name (perm_dest, perm_stmt); |
2635892a | 2333 | gimple_set_lhs (perm_stmt, data_ref); |
ebfd146a | 2334 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
ebfd146a | 2335 | |
b8698a0f L |
2336 | /* Store the vector statement in NODE. */ |
2337 | VEC_replace (gimple, SLP_TREE_VEC_STMTS (node), | |
ebfd146a IR |
2338 | stride * i + vect_stmts_counter, perm_stmt); |
2339 | ||
2340 | first_vec_indx += stride; | |
2341 | second_vec_indx += stride; | |
2342 | } | |
2343 | ||
2344 | /* Mark the scalar stmt as vectorized. */ | |
2345 | next_stmt_info = vinfo_for_stmt (next_scalar_stmt); | |
2346 | STMT_VINFO_VEC_STMT (next_stmt_info) = perm_stmt; | |
2347 | } | |
2348 | ||
2349 | ||
b8698a0f | 2350 | /* Given FIRST_MASK_ELEMENT - the mask element in element representation, |
ebfd146a | 2351 | return in CURRENT_MASK_ELEMENT its equivalent in target specific |
ff802fa1 | 2352 | representation. Check that the mask is valid and return FALSE if not. |
ebfd146a IR |
2353 | Return TRUE in NEED_NEXT_VECTOR if the permutation requires to move to |
2354 | the next vector, i.e., the current first vector is not needed. */ | |
b8698a0f | 2355 | |
ebfd146a | 2356 | static bool |
b8698a0f | 2357 | vect_get_mask_element (gimple stmt, int first_mask_element, int m, |
ebfd146a | 2358 | int mask_nunits, bool only_one_vec, int index, |
b8698a0f | 2359 | int *mask, int *current_mask_element, |
694a4f61 IR |
2360 | bool *need_next_vector, int *number_of_mask_fixes, |
2361 | bool *mask_fixed, bool *needs_first_vector) | |
ebfd146a IR |
2362 | { |
2363 | int i; | |
ebfd146a IR |
2364 | |
2365 | /* Convert to target specific representation. */ | |
2366 | *current_mask_element = first_mask_element + m; | |
2367 | /* Adjust the value in case it's a mask for second and third vectors. */ | |
694a4f61 | 2368 | *current_mask_element -= mask_nunits * (*number_of_mask_fixes - 1); |
ebfd146a IR |
2369 | |
2370 | if (*current_mask_element < mask_nunits) | |
694a4f61 | 2371 | *needs_first_vector = true; |
ebfd146a IR |
2372 | |
2373 | /* We have only one input vector to permute but the mask accesses values in | |
2374 | the next vector as well. */ | |
2375 | if (only_one_vec && *current_mask_element >= mask_nunits) | |
2376 | { | |
2377 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2378 | { | |
2379 | fprintf (vect_dump, "permutation requires at least two vectors "); | |
2380 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
2381 | } | |
2382 | ||
2383 | return false; | |
2384 | } | |
2385 | ||
2386 | /* The mask requires the next vector. */ | |
2387 | if (*current_mask_element >= mask_nunits * 2) | |
2388 | { | |
694a4f61 | 2389 | if (*needs_first_vector || *mask_fixed) |
ebfd146a IR |
2390 | { |
2391 | /* We either need the first vector too or have already moved to the | |
b8698a0f | 2392 | next vector. In both cases, this permutation needs three |
ebfd146a IR |
2393 | vectors. */ |
2394 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2395 | { | |
2396 | fprintf (vect_dump, "permutation requires at " | |
2397 | "least three vectors "); | |
2398 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
2399 | } | |
2400 | ||
2401 | return false; | |
2402 | } | |
2403 | ||
2404 | /* We move to the next vector, dropping the first one and working with | |
2405 | the second and the third - we need to adjust the values of the mask | |
2406 | accordingly. */ | |
694a4f61 | 2407 | *current_mask_element -= mask_nunits * *number_of_mask_fixes; |
ebfd146a IR |
2408 | |
2409 | for (i = 0; i < index; i++) | |
694a4f61 | 2410 | mask[i] -= mask_nunits * *number_of_mask_fixes; |
ebfd146a | 2411 | |
694a4f61 IR |
2412 | (*number_of_mask_fixes)++; |
2413 | *mask_fixed = true; | |
ebfd146a IR |
2414 | } |
2415 | ||
694a4f61 | 2416 | *need_next_vector = *mask_fixed; |
ebfd146a IR |
2417 | |
2418 | /* This was the last element of this mask. Start a new one. */ | |
2419 | if (index == mask_nunits - 1) | |
2420 | { | |
694a4f61 IR |
2421 | *number_of_mask_fixes = 1; |
2422 | *mask_fixed = false; | |
2423 | *needs_first_vector = false; | |
ebfd146a IR |
2424 | } |
2425 | ||
2426 | return true; | |
2427 | } | |
2428 | ||
2429 | ||
2430 | /* Generate vector permute statements from a list of loads in DR_CHAIN. | |
2431 | If ANALYZE_ONLY is TRUE, only check that it is possible to create valid | |
2432 | permute statements for SLP_NODE_INSTANCE. */ | |
2433 | bool | |
2434 | vect_transform_slp_perm_load (gimple stmt, VEC (tree, heap) *dr_chain, | |
2435 | gimple_stmt_iterator *gsi, int vf, | |
2436 | slp_instance slp_node_instance, bool analyze_only) | |
2437 | { | |
2438 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2439 | tree mask_element_type = NULL_TREE, mask_type; | |
2635892a | 2440 | int i, j, k, nunits, vec_index = 0, scalar_index; |
ebfd146a | 2441 | slp_tree node; |
2635892a | 2442 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a IR |
2443 | gimple next_scalar_stmt; |
2444 | int group_size = SLP_INSTANCE_GROUP_SIZE (slp_node_instance); | |
2445 | int first_mask_element; | |
2446 | int index, unroll_factor, *mask, current_mask_element, ncopies; | |
2447 | bool only_one_vec = false, need_next_vector = false; | |
2448 | int first_vec_index, second_vec_index, orig_vec_stmts_num, vect_stmts_counter; | |
694a4f61 IR |
2449 | int number_of_mask_fixes = 1; |
2450 | bool mask_fixed = false; | |
2451 | bool needs_first_vector = false; | |
ebfd146a | 2452 | |
2635892a | 2453 | if (!can_vec_perm_expr_p (vectype, NULL_TREE)) |
ebfd146a IR |
2454 | { |
2455 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2456 | { | |
2635892a | 2457 | fprintf (vect_dump, "no vect permute for "); |
ebfd146a IR |
2458 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); |
2459 | } | |
2635892a | 2460 | return false; |
ebfd146a IR |
2461 | } |
2462 | ||
2635892a RH |
2463 | /* The generic VEC_PERM_EXPR code always uses an integral type of the |
2464 | same size as the vector element being permuted. */ | |
2465 | mask_element_type | |
2466 | = lang_hooks.types.type_for_size | |
2467 | (TREE_INT_CST_LOW (TYPE_SIZE (TREE_TYPE (vectype))), 1); | |
ebfd146a | 2468 | mask_type = get_vectype_for_scalar_type (mask_element_type); |
ebfd146a | 2469 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2635892a | 2470 | mask = (int *) xmalloc (sizeof (int) * nunits); |
ebfd146a IR |
2471 | unroll_factor = SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance); |
2472 | ||
2473 | /* The number of vector stmts to generate based only on SLP_NODE_INSTANCE | |
2474 | unrolling factor. */ | |
b8698a0f | 2475 | orig_vec_stmts_num = group_size * |
ebfd146a IR |
2476 | SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance) / nunits; |
2477 | if (orig_vec_stmts_num == 1) | |
2478 | only_one_vec = true; | |
2479 | ||
b8698a0f | 2480 | /* Number of copies is determined by the final vectorization factor |
ebfd146a | 2481 | relatively to SLP_NODE_INSTANCE unrolling factor. */ |
b8698a0f | 2482 | ncopies = vf / SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance); |
ebfd146a | 2483 | |
b8698a0f L |
2484 | /* Generate permutation masks for every NODE. Number of masks for each NODE |
2485 | is equal to GROUP_SIZE. | |
2486 | E.g., we have a group of three nodes with three loads from the same | |
2487 | location in each node, and the vector size is 4. I.e., we have a | |
2488 | a0b0c0a1b1c1... sequence and we need to create the following vectors: | |
ebfd146a IR |
2489 | for a's: a0a0a0a1 a1a1a2a2 a2a3a3a3 |
2490 | for b's: b0b0b0b1 b1b1b2b2 b2b3b3b3 | |
2491 | ... | |
2492 | ||
2635892a | 2493 | The masks for a's should be: {0,0,0,3} {3,3,6,6} {6,9,9,9}. |
b8698a0f | 2494 | The last mask is illegal since we assume two operands for permute |
ff802fa1 IR |
2495 | operation, and the mask element values can't be outside that range. |
2496 | Hence, the last mask must be converted into {2,5,5,5}. | |
b8698a0f | 2497 | For the first two permutations we need the first and the second input |
ebfd146a | 2498 | vectors: {a0,b0,c0,a1} and {b1,c1,a2,b2}, and for the last permutation |
b8698a0f | 2499 | we need the second and the third vectors: {b1,c1,a2,b2} and |
ebfd146a IR |
2500 | {c2,a3,b3,c3}. */ |
2501 | ||
ac47786e | 2502 | FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (slp_node_instance), i, node) |
ebfd146a IR |
2503 | { |
2504 | scalar_index = 0; | |
2505 | index = 0; | |
2506 | vect_stmts_counter = 0; | |
2507 | vec_index = 0; | |
2508 | first_vec_index = vec_index++; | |
2509 | if (only_one_vec) | |
2510 | second_vec_index = first_vec_index; | |
2511 | else | |
2512 | second_vec_index = vec_index++; | |
2513 | ||
2514 | for (j = 0; j < unroll_factor; j++) | |
2515 | { | |
2516 | for (k = 0; k < group_size; k++) | |
2517 | { | |
2635892a RH |
2518 | first_mask_element = i + j * group_size; |
2519 | if (!vect_get_mask_element (stmt, first_mask_element, 0, | |
2520 | nunits, only_one_vec, index, | |
2521 | mask, ¤t_mask_element, | |
2522 | &need_next_vector, | |
2523 | &number_of_mask_fixes, &mask_fixed, | |
2524 | &needs_first_vector)) | |
2525 | return false; | |
2526 | mask[index++] = current_mask_element; | |
ebfd146a | 2527 | |
2635892a | 2528 | if (index == nunits) |
ebfd146a | 2529 | { |
faf63e39 RH |
2530 | tree mask_vec = NULL; |
2531 | ||
2532 | while (--index >= 0) | |
2533 | { | |
2534 | tree t = build_int_cst (mask_element_type, mask[index]); | |
2535 | mask_vec = tree_cons (NULL, t, mask_vec); | |
2536 | } | |
2537 | mask_vec = build_vector (mask_type, mask_vec); | |
2538 | index = 0; | |
2539 | ||
2635892a | 2540 | if (!can_vec_perm_expr_p (vectype, mask_vec)) |
faf63e39 RH |
2541 | { |
2542 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2543 | { | |
2544 | fprintf (vect_dump, "unsupported vect permute "); | |
2545 | print_generic_expr (vect_dump, mask_vec, 0); | |
2546 | } | |
2547 | free (mask); | |
2548 | return false; | |
2549 | } | |
2550 | ||
ebfd146a IR |
2551 | if (!analyze_only) |
2552 | { | |
2553 | if (need_next_vector) | |
2554 | { | |
2555 | first_vec_index = second_vec_index; | |
2556 | second_vec_index = vec_index; | |
2557 | } | |
2558 | ||
2559 | next_scalar_stmt = VEC_index (gimple, | |
2560 | SLP_TREE_SCALAR_STMTS (node), scalar_index++); | |
2561 | ||
2562 | vect_create_mask_and_perm (stmt, next_scalar_stmt, | |
faf63e39 | 2563 | mask_vec, first_vec_index, second_vec_index, |
2635892a | 2564 | gsi, node, vectype, dr_chain, |
faf63e39 | 2565 | ncopies, vect_stmts_counter++); |
ebfd146a | 2566 | } |
b8698a0f L |
2567 | } |
2568 | } | |
2569 | } | |
2570 | } | |
ebfd146a IR |
2571 | |
2572 | free (mask); | |
2573 | return true; | |
2574 | } | |
2575 | ||
2576 | ||
2577 | ||
2578 | /* Vectorize SLP instance tree in postorder. */ | |
2579 | ||
2580 | static bool | |
2581 | vect_schedule_slp_instance (slp_tree node, slp_instance instance, | |
a70d6342 | 2582 | unsigned int vectorization_factor) |
ebfd146a IR |
2583 | { |
2584 | gimple stmt; | |
2585 | bool strided_store, is_store; | |
2586 | gimple_stmt_iterator si; | |
2587 | stmt_vec_info stmt_info; | |
2588 | unsigned int vec_stmts_size, nunits, group_size; | |
2589 | tree vectype; | |
2590 | int i; | |
2591 | slp_tree loads_node; | |
2592 | ||
2593 | if (!node) | |
2594 | return false; | |
2595 | ||
2596 | vect_schedule_slp_instance (SLP_TREE_LEFT (node), instance, | |
2597 | vectorization_factor); | |
2598 | vect_schedule_slp_instance (SLP_TREE_RIGHT (node), instance, | |
2599 | vectorization_factor); | |
b8698a0f | 2600 | |
ebfd146a IR |
2601 | stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (node), 0); |
2602 | stmt_info = vinfo_for_stmt (stmt); | |
2603 | ||
2604 | /* VECTYPE is the type of the destination. */ | |
b690cc0f | 2605 | vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a IR |
2606 | nunits = (unsigned int) TYPE_VECTOR_SUBPARTS (vectype); |
2607 | group_size = SLP_INSTANCE_GROUP_SIZE (instance); | |
2608 | ||
2609 | /* For each SLP instance calculate number of vector stmts to be created | |
ff802fa1 | 2610 | for the scalar stmts in each node of the SLP tree. Number of vector |
ebfd146a IR |
2611 | elements in one vector iteration is the number of scalar elements in |
2612 | one scalar iteration (GROUP_SIZE) multiplied by VF divided by vector | |
2613 | size. */ | |
2614 | vec_stmts_size = (vectorization_factor * group_size) / nunits; | |
2615 | ||
2616 | /* In case of load permutation we have to allocate vectorized statements for | |
2617 | all the nodes that participate in that permutation. */ | |
2618 | if (SLP_INSTANCE_LOAD_PERMUTATION (instance)) | |
2619 | { | |
ac47786e | 2620 | FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (instance), i, loads_node) |
ebfd146a IR |
2621 | { |
2622 | if (!SLP_TREE_VEC_STMTS (loads_node)) | |
2623 | { | |
2624 | SLP_TREE_VEC_STMTS (loads_node) = VEC_alloc (gimple, heap, | |
2625 | vec_stmts_size); | |
2626 | SLP_TREE_NUMBER_OF_VEC_STMTS (loads_node) = vec_stmts_size; | |
2627 | } | |
2628 | } | |
2629 | } | |
2630 | ||
2631 | if (!SLP_TREE_VEC_STMTS (node)) | |
2632 | { | |
2633 | SLP_TREE_VEC_STMTS (node) = VEC_alloc (gimple, heap, vec_stmts_size); | |
2634 | SLP_TREE_NUMBER_OF_VEC_STMTS (node) = vec_stmts_size; | |
2635 | } | |
2636 | ||
2637 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2638 | { | |
2639 | fprintf (vect_dump, "------>vectorizing SLP node starting from: "); | |
2640 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
b8698a0f | 2641 | } |
ebfd146a IR |
2642 | |
2643 | /* Loads should be inserted before the first load. */ | |
2644 | if (SLP_INSTANCE_FIRST_LOAD_STMT (instance) | |
2645 | && STMT_VINFO_STRIDED_ACCESS (stmt_info) | |
6aa904c4 IR |
2646 | && !REFERENCE_CLASS_P (gimple_get_lhs (stmt)) |
2647 | && SLP_INSTANCE_LOAD_PERMUTATION (instance)) | |
ebfd146a | 2648 | si = gsi_for_stmt (SLP_INSTANCE_FIRST_LOAD_STMT (instance)); |
9d5e7640 | 2649 | else if (is_pattern_stmt_p (stmt_info)) |
6aa904c4 | 2650 | si = gsi_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
ebfd146a IR |
2651 | else |
2652 | si = gsi_for_stmt (stmt); | |
b8698a0f | 2653 | |
e4a707c4 IR |
2654 | /* Stores should be inserted just before the last store. */ |
2655 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) | |
2656 | && REFERENCE_CLASS_P (gimple_get_lhs (stmt))) | |
2657 | { | |
2658 | gimple last_store = vect_find_last_store_in_slp_instance (instance); | |
2659 | si = gsi_for_stmt (last_store); | |
2660 | } | |
2661 | ||
b010117a IR |
2662 | /* Mark the first element of the reduction chain as reduction to properly |
2663 | transform the node. In the analysis phase only the last element of the | |
2664 | chain is marked as reduction. */ | |
2665 | if (GROUP_FIRST_ELEMENT (stmt_info) && !STMT_VINFO_STRIDED_ACCESS (stmt_info) | |
2666 | && GROUP_FIRST_ELEMENT (stmt_info) == stmt) | |
2667 | { | |
2668 | STMT_VINFO_DEF_TYPE (stmt_info) = vect_reduction_def; | |
2669 | STMT_VINFO_TYPE (stmt_info) = reduc_vec_info_type; | |
2670 | } | |
2671 | ||
ebfd146a | 2672 | is_store = vect_transform_stmt (stmt, &si, &strided_store, node, instance); |
b5aeb3bb | 2673 | return is_store; |
ebfd146a IR |
2674 | } |
2675 | ||
2676 | ||
ff802fa1 IR |
2677 | /* Generate vector code for all SLP instances in the loop/basic block. */ |
2678 | ||
ebfd146a | 2679 | bool |
a70d6342 | 2680 | vect_schedule_slp (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo) |
ebfd146a | 2681 | { |
a70d6342 | 2682 | VEC (slp_instance, heap) *slp_instances; |
ebfd146a | 2683 | slp_instance instance; |
a70d6342 | 2684 | unsigned int i, vf; |
ebfd146a IR |
2685 | bool is_store = false; |
2686 | ||
a70d6342 IR |
2687 | if (loop_vinfo) |
2688 | { | |
2689 | slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo); | |
2690 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
b8698a0f | 2691 | } |
a70d6342 IR |
2692 | else |
2693 | { | |
2694 | slp_instances = BB_VINFO_SLP_INSTANCES (bb_vinfo); | |
2695 | vf = 1; | |
b8698a0f | 2696 | } |
a70d6342 | 2697 | |
ac47786e | 2698 | FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance) |
ebfd146a IR |
2699 | { |
2700 | /* Schedule the tree of INSTANCE. */ | |
2701 | is_store = vect_schedule_slp_instance (SLP_INSTANCE_TREE (instance), | |
a70d6342 | 2702 | instance, vf); |
8644a673 IR |
2703 | if (vect_print_dump_info (REPORT_VECTORIZED_LOCATIONS) |
2704 | || vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) | |
ebfd146a IR |
2705 | fprintf (vect_dump, "vectorizing stmts using SLP."); |
2706 | } | |
2707 | ||
ac47786e | 2708 | FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance) |
b5aeb3bb IR |
2709 | { |
2710 | slp_tree root = SLP_INSTANCE_TREE (instance); | |
2711 | gimple store; | |
2712 | unsigned int j; | |
2713 | gimple_stmt_iterator gsi; | |
2714 | ||
2715 | for (j = 0; VEC_iterate (gimple, SLP_TREE_SCALAR_STMTS (root), j, store) | |
2716 | && j < SLP_INSTANCE_GROUP_SIZE (instance); j++) | |
2717 | { | |
2718 | if (!STMT_VINFO_DATA_REF (vinfo_for_stmt (store))) | |
2719 | break; | |
2720 | ||
2721 | /* Free the attached stmt_vec_info and remove the stmt. */ | |
2722 | gsi = gsi_for_stmt (store); | |
2723 | gsi_remove (&gsi, true); | |
2724 | free_stmt_vec_info (store); | |
2725 | } | |
2726 | } | |
2727 | ||
ebfd146a IR |
2728 | return is_store; |
2729 | } | |
a70d6342 IR |
2730 | |
2731 | ||
2732 | /* Vectorize the basic block. */ | |
2733 | ||
2734 | void | |
2735 | vect_slp_transform_bb (basic_block bb) | |
2736 | { | |
2737 | bb_vec_info bb_vinfo = vec_info_for_bb (bb); | |
2738 | gimple_stmt_iterator si; | |
2739 | ||
2740 | gcc_assert (bb_vinfo); | |
2741 | ||
2742 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2743 | fprintf (vect_dump, "SLPing BB\n"); | |
2744 | ||
2745 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
2746 | { | |
2747 | gimple stmt = gsi_stmt (si); | |
2748 | stmt_vec_info stmt_info; | |
2749 | ||
2750 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2751 | { | |
2752 | fprintf (vect_dump, "------>SLPing statement: "); | |
2753 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
2754 | } | |
2755 | ||
2756 | stmt_info = vinfo_for_stmt (stmt); | |
2757 | gcc_assert (stmt_info); | |
2758 | ||
2759 | /* Schedule all the SLP instances when the first SLP stmt is reached. */ | |
2760 | if (STMT_SLP_TYPE (stmt_info)) | |
2761 | { | |
2762 | vect_schedule_slp (NULL, bb_vinfo); | |
2763 | break; | |
2764 | } | |
2765 | } | |
2766 | ||
2767 | mark_sym_for_renaming (gimple_vop (cfun)); | |
2768 | /* The memory tags and pointers in vectorized statements need to | |
2769 | have their SSA forms updated. FIXME, why can't this be delayed | |
2770 | until all the loops have been transformed? */ | |
2771 | update_ssa (TODO_update_ssa); | |
2772 | ||
2773 | if (vect_print_dump_info (REPORT_DETAILS)) | |
e9dbe7bb | 2774 | fprintf (vect_dump, "BASIC BLOCK VECTORIZED\n"); |
a70d6342 | 2775 | |
12aaf609 IR |
2776 | destroy_bb_vec_info (bb_vinfo); |
2777 | } | |
a70d6342 | 2778 |