]>
Commit | Line | Data |
---|---|---|
b8698a0f | 1 | /* Vectorizer Specific Loop Manipulations |
8d9254fc | 2 | Copyright (C) 2003-2020 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
c7131fb2 | 25 | #include "backend.h" |
40e23961 | 26 | #include "tree.h" |
c7131fb2 | 27 | #include "gimple.h" |
957060b5 AM |
28 | #include "cfghooks.h" |
29 | #include "tree-pass.h" | |
c7131fb2 | 30 | #include "ssa.h" |
c7131fb2 | 31 | #include "fold-const.h" |
60393bbc | 32 | #include "cfganal.h" |
45b0be94 | 33 | #include "gimplify.h" |
5be5c238 | 34 | #include "gimple-iterator.h" |
18f429e2 | 35 | #include "gimplify-me.h" |
442b4905 | 36 | #include "tree-cfg.h" |
e28030cf | 37 | #include "tree-ssa-loop-manip.h" |
442b4905 | 38 | #include "tree-into-ssa.h" |
7a300452 | 39 | #include "tree-ssa.h" |
ebfd146a | 40 | #include "cfgloop.h" |
ebfd146a IR |
41 | #include "tree-scalar-evolution.h" |
42 | #include "tree-vectorizer.h" | |
2a93954e | 43 | #include "tree-ssa-loop-ivopts.h" |
0f26839a | 44 | #include "gimple-fold.h" |
7cfb4d93 RS |
45 | #include "tree-ssa-loop-niter.h" |
46 | #include "internal-fn.h" | |
47 | #include "stor-layout.h" | |
48 | #include "optabs-query.h" | |
49 | #include "vec-perm-indices.h" | |
0a0ef238 RS |
50 | #include "insn-config.h" |
51 | #include "rtl.h" | |
52 | #include "recog.h" | |
ebfd146a IR |
53 | |
54 | /************************************************************************* | |
55 | Simple Loop Peeling Utilities | |
56 | ||
57 | Utilities to support loop peeling for vectorization purposes. | |
58 | *************************************************************************/ | |
59 | ||
60 | ||
61 | /* Renames the use *OP_P. */ | |
62 | ||
63 | static void | |
64 | rename_use_op (use_operand_p op_p) | |
65 | { | |
66 | tree new_name; | |
67 | ||
68 | if (TREE_CODE (USE_FROM_PTR (op_p)) != SSA_NAME) | |
69 | return; | |
70 | ||
71 | new_name = get_current_def (USE_FROM_PTR (op_p)); | |
72 | ||
73 | /* Something defined outside of the loop. */ | |
74 | if (!new_name) | |
75 | return; | |
76 | ||
77 | /* An ordinary ssa name defined in the loop. */ | |
78 | ||
79 | SET_USE (op_p, new_name); | |
80 | } | |
81 | ||
82 | ||
cb330ba5 | 83 | /* Renames the variables in basic block BB. Allow renaming of PHI arguments |
a6c51a12 YR |
84 | on edges incoming from outer-block header if RENAME_FROM_OUTER_LOOP is |
85 | true. */ | |
ebfd146a | 86 | |
2cfc56b9 | 87 | static void |
a6c51a12 | 88 | rename_variables_in_bb (basic_block bb, bool rename_from_outer_loop) |
ebfd146a | 89 | { |
355fe088 | 90 | gimple *stmt; |
ebfd146a IR |
91 | use_operand_p use_p; |
92 | ssa_op_iter iter; | |
93 | edge e; | |
94 | edge_iterator ei; | |
99b1c316 MS |
95 | class loop *loop = bb->loop_father; |
96 | class loop *outer_loop = NULL; | |
a6c51a12 YR |
97 | |
98 | if (rename_from_outer_loop) | |
99 | { | |
100 | gcc_assert (loop); | |
101 | outer_loop = loop_outer (loop); | |
102 | } | |
ebfd146a | 103 | |
538dd0b7 DM |
104 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi); |
105 | gsi_next (&gsi)) | |
ebfd146a IR |
106 | { |
107 | stmt = gsi_stmt (gsi); | |
108 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_ALL_USES) | |
109 | rename_use_op (use_p); | |
110 | } | |
111 | ||
2cfc56b9 | 112 | FOR_EACH_EDGE (e, ei, bb->preds) |
ebfd146a | 113 | { |
cb330ba5 JJ |
114 | if (!flow_bb_inside_loop_p (loop, e->src)) |
115 | { | |
116 | if (!rename_from_outer_loop) | |
117 | continue; | |
118 | if (e->src != outer_loop->header) | |
119 | { | |
120 | if (outer_loop->inner->next) | |
121 | { | |
122 | /* If outer_loop has 2 inner loops, allow there to | |
123 | be an extra basic block which decides which of the | |
124 | two loops to use using LOOP_VECTORIZED. */ | |
125 | if (!single_pred_p (e->src) | |
126 | || single_pred (e->src) != outer_loop->header) | |
127 | continue; | |
128 | } | |
cb330ba5 JJ |
129 | } |
130 | } | |
538dd0b7 DM |
131 | for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi); |
132 | gsi_next (&gsi)) | |
133 | rename_use_op (PHI_ARG_DEF_PTR_FROM_EDGE (gsi.phi (), e)); | |
ebfd146a IR |
134 | } |
135 | } | |
136 | ||
137 | ||
a79683d5 | 138 | struct adjust_info |
684f25f4 AO |
139 | { |
140 | tree from, to; | |
141 | basic_block bb; | |
a79683d5 | 142 | }; |
684f25f4 | 143 | |
684f25f4 AO |
144 | /* A stack of values to be adjusted in debug stmts. We have to |
145 | process them LIFO, so that the closest substitution applies. If we | |
146 | processed them FIFO, without the stack, we might substitute uses | |
147 | with a PHI DEF that would soon become non-dominant, and when we got | |
148 | to the suitable one, it wouldn't have anything to substitute any | |
149 | more. */ | |
ff4c81cc | 150 | static vec<adjust_info, va_heap> adjust_vec; |
684f25f4 AO |
151 | |
152 | /* Adjust any debug stmts that referenced AI->from values to use the | |
153 | loop-closed AI->to, if the references are dominated by AI->bb and | |
154 | not by the definition of AI->from. */ | |
155 | ||
156 | static void | |
157 | adjust_debug_stmts_now (adjust_info *ai) | |
158 | { | |
159 | basic_block bbphi = ai->bb; | |
160 | tree orig_def = ai->from; | |
161 | tree new_def = ai->to; | |
162 | imm_use_iterator imm_iter; | |
355fe088 | 163 | gimple *stmt; |
684f25f4 AO |
164 | basic_block bbdef = gimple_bb (SSA_NAME_DEF_STMT (orig_def)); |
165 | ||
166 | gcc_assert (dom_info_available_p (CDI_DOMINATORS)); | |
167 | ||
168 | /* Adjust any debug stmts that held onto non-loop-closed | |
169 | references. */ | |
170 | FOR_EACH_IMM_USE_STMT (stmt, imm_iter, orig_def) | |
171 | { | |
172 | use_operand_p use_p; | |
173 | basic_block bbuse; | |
174 | ||
175 | if (!is_gimple_debug (stmt)) | |
176 | continue; | |
177 | ||
178 | gcc_assert (gimple_debug_bind_p (stmt)); | |
179 | ||
180 | bbuse = gimple_bb (stmt); | |
181 | ||
182 | if ((bbuse == bbphi | |
183 | || dominated_by_p (CDI_DOMINATORS, bbuse, bbphi)) | |
184 | && !(bbuse == bbdef | |
185 | || dominated_by_p (CDI_DOMINATORS, bbuse, bbdef))) | |
186 | { | |
187 | if (new_def) | |
188 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) | |
189 | SET_USE (use_p, new_def); | |
190 | else | |
191 | { | |
192 | gimple_debug_bind_reset_value (stmt); | |
193 | update_stmt (stmt); | |
194 | } | |
195 | } | |
196 | } | |
197 | } | |
198 | ||
199 | /* Adjust debug stmts as scheduled before. */ | |
200 | ||
201 | static void | |
202 | adjust_vec_debug_stmts (void) | |
203 | { | |
36f52e8f | 204 | if (!MAY_HAVE_DEBUG_BIND_STMTS) |
684f25f4 AO |
205 | return; |
206 | ||
9771b263 | 207 | gcc_assert (adjust_vec.exists ()); |
684f25f4 | 208 | |
9771b263 | 209 | while (!adjust_vec.is_empty ()) |
684f25f4 | 210 | { |
9771b263 DN |
211 | adjust_debug_stmts_now (&adjust_vec.last ()); |
212 | adjust_vec.pop (); | |
684f25f4 | 213 | } |
684f25f4 AO |
214 | } |
215 | ||
216 | /* Adjust any debug stmts that referenced FROM values to use the | |
217 | loop-closed TO, if the references are dominated by BB and not by | |
218 | the definition of FROM. If adjust_vec is non-NULL, adjustments | |
219 | will be postponed until adjust_vec_debug_stmts is called. */ | |
220 | ||
221 | static void | |
222 | adjust_debug_stmts (tree from, tree to, basic_block bb) | |
223 | { | |
224 | adjust_info ai; | |
225 | ||
36f52e8f | 226 | if (MAY_HAVE_DEBUG_BIND_STMTS |
a471762f | 227 | && TREE_CODE (from) == SSA_NAME |
a52ca739 | 228 | && ! SSA_NAME_IS_DEFAULT_DEF (from) |
a471762f | 229 | && ! virtual_operand_p (from)) |
684f25f4 AO |
230 | { |
231 | ai.from = from; | |
232 | ai.to = to; | |
233 | ai.bb = bb; | |
234 | ||
9771b263 DN |
235 | if (adjust_vec.exists ()) |
236 | adjust_vec.safe_push (ai); | |
684f25f4 AO |
237 | else |
238 | adjust_debug_stmts_now (&ai); | |
239 | } | |
240 | } | |
241 | ||
242 | /* Change E's phi arg in UPDATE_PHI to NEW_DEF, and record information | |
243 | to adjust any debug stmts that referenced the old phi arg, | |
244 | presumably non-loop-closed references left over from other | |
245 | transformations. */ | |
246 | ||
247 | static void | |
355fe088 | 248 | adjust_phi_and_debug_stmts (gimple *update_phi, edge e, tree new_def) |
684f25f4 AO |
249 | { |
250 | tree orig_def = PHI_ARG_DEF_FROM_EDGE (update_phi, e); | |
251 | ||
252 | SET_PHI_ARG_DEF (update_phi, e->dest_idx, new_def); | |
253 | ||
36f52e8f | 254 | if (MAY_HAVE_DEBUG_BIND_STMTS) |
684f25f4 AO |
255 | adjust_debug_stmts (orig_def, PHI_RESULT (update_phi), |
256 | gimple_bb (update_phi)); | |
257 | } | |
258 | ||
7cfb4d93 RS |
259 | /* Define one loop mask MASK from loop LOOP. INIT_MASK is the value that |
260 | the mask should have during the first iteration and NEXT_MASK is the | |
261 | value that it should have on subsequent iterations. */ | |
0f26839a | 262 | |
7cfb4d93 | 263 | static void |
99b1c316 | 264 | vect_set_loop_mask (class loop *loop, tree mask, tree init_mask, |
7cfb4d93 RS |
265 | tree next_mask) |
266 | { | |
267 | gphi *phi = create_phi_node (mask, loop->header); | |
268 | add_phi_arg (phi, init_mask, loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
269 | add_phi_arg (phi, next_mask, loop_latch_edge (loop), UNKNOWN_LOCATION); | |
270 | } | |
0f26839a | 271 | |
7cfb4d93 | 272 | /* Add SEQ to the end of LOOP's preheader block. */ |
ebfd146a | 273 | |
7cfb4d93 | 274 | static void |
99b1c316 | 275 | add_preheader_seq (class loop *loop, gimple_seq seq) |
7cfb4d93 RS |
276 | { |
277 | if (seq) | |
278 | { | |
279 | edge pe = loop_preheader_edge (loop); | |
280 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
281 | gcc_assert (!new_bb); | |
282 | } | |
283 | } | |
ebfd146a | 284 | |
7cfb4d93 RS |
285 | /* Add SEQ to the beginning of LOOP's header block. */ |
286 | ||
287 | static void | |
99b1c316 | 288 | add_header_seq (class loop *loop, gimple_seq seq) |
7cfb4d93 RS |
289 | { |
290 | if (seq) | |
291 | { | |
292 | gimple_stmt_iterator gsi = gsi_after_labels (loop->header); | |
293 | gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT); | |
294 | } | |
295 | } | |
296 | ||
297 | /* Return true if the target can interleave elements of two vectors. | |
298 | OFFSET is 0 if the first half of the vectors should be interleaved | |
299 | or 1 if the second half should. When returning true, store the | |
300 | associated permutation in INDICES. */ | |
301 | ||
302 | static bool | |
303 | interleave_supported_p (vec_perm_indices *indices, tree vectype, | |
304 | unsigned int offset) | |
305 | { | |
306 | poly_uint64 nelts = TYPE_VECTOR_SUBPARTS (vectype); | |
307 | poly_uint64 base = exact_div (nelts, 2) * offset; | |
308 | vec_perm_builder sel (nelts, 2, 3); | |
309 | for (unsigned int i = 0; i < 3; ++i) | |
310 | { | |
311 | sel.quick_push (base + i); | |
312 | sel.quick_push (base + i + nelts); | |
313 | } | |
314 | indices->new_vector (sel, 2, nelts); | |
315 | return can_vec_perm_const_p (TYPE_MODE (vectype), *indices); | |
316 | } | |
317 | ||
318 | /* Try to use permutes to define the masks in DEST_RGM using the masks | |
319 | in SRC_RGM, given that the former has twice as many masks as the | |
320 | latter. Return true on success, adding any new statements to SEQ. */ | |
321 | ||
322 | static bool | |
0a0ef238 | 323 | vect_maybe_permute_loop_masks (gimple_seq *seq, rgroup_masks *dest_rgm, |
7cfb4d93 RS |
324 | rgroup_masks *src_rgm) |
325 | { | |
326 | tree src_masktype = src_rgm->mask_type; | |
327 | tree dest_masktype = dest_rgm->mask_type; | |
328 | machine_mode src_mode = TYPE_MODE (src_masktype); | |
0a0ef238 | 329 | insn_code icode1, icode2; |
7cfb4d93 | 330 | if (dest_rgm->max_nscalars_per_iter <= src_rgm->max_nscalars_per_iter |
0a0ef238 RS |
331 | && (icode1 = optab_handler (vec_unpacku_hi_optab, |
332 | src_mode)) != CODE_FOR_nothing | |
333 | && (icode2 = optab_handler (vec_unpacku_lo_optab, | |
334 | src_mode)) != CODE_FOR_nothing) | |
7cfb4d93 RS |
335 | { |
336 | /* Unpacking the source masks gives at least as many mask bits as | |
337 | we need. We can then VIEW_CONVERT any excess bits away. */ | |
0a0ef238 RS |
338 | machine_mode dest_mode = insn_data[icode1].operand[0].mode; |
339 | gcc_assert (dest_mode == insn_data[icode2].operand[0].mode); | |
340 | tree unpack_masktype = vect_halve_mask_nunits (src_masktype, dest_mode); | |
7cfb4d93 RS |
341 | for (unsigned int i = 0; i < dest_rgm->masks.length (); ++i) |
342 | { | |
343 | tree src = src_rgm->masks[i / 2]; | |
344 | tree dest = dest_rgm->masks[i]; | |
9bfb28ed RS |
345 | tree_code code = ((i & 1) == (BYTES_BIG_ENDIAN ? 0 : 1) |
346 | ? VEC_UNPACK_HI_EXPR | |
7cfb4d93 RS |
347 | : VEC_UNPACK_LO_EXPR); |
348 | gassign *stmt; | |
349 | if (dest_masktype == unpack_masktype) | |
350 | stmt = gimple_build_assign (dest, code, src); | |
351 | else | |
352 | { | |
353 | tree temp = make_ssa_name (unpack_masktype); | |
354 | stmt = gimple_build_assign (temp, code, src); | |
355 | gimple_seq_add_stmt (seq, stmt); | |
356 | stmt = gimple_build_assign (dest, VIEW_CONVERT_EXPR, | |
357 | build1 (VIEW_CONVERT_EXPR, | |
358 | dest_masktype, temp)); | |
359 | } | |
360 | gimple_seq_add_stmt (seq, stmt); | |
361 | } | |
362 | return true; | |
363 | } | |
364 | vec_perm_indices indices[2]; | |
365 | if (dest_masktype == src_masktype | |
366 | && interleave_supported_p (&indices[0], src_masktype, 0) | |
367 | && interleave_supported_p (&indices[1], src_masktype, 1)) | |
368 | { | |
369 | /* The destination requires twice as many mask bits as the source, so | |
370 | we can use interleaving permutes to double up the number of bits. */ | |
371 | tree masks[2]; | |
372 | for (unsigned int i = 0; i < 2; ++i) | |
373 | masks[i] = vect_gen_perm_mask_checked (src_masktype, indices[i]); | |
374 | for (unsigned int i = 0; i < dest_rgm->masks.length (); ++i) | |
375 | { | |
376 | tree src = src_rgm->masks[i / 2]; | |
377 | tree dest = dest_rgm->masks[i]; | |
378 | gimple *stmt = gimple_build_assign (dest, VEC_PERM_EXPR, | |
379 | src, src, masks[i & 1]); | |
380 | gimple_seq_add_stmt (seq, stmt); | |
381 | } | |
382 | return true; | |
383 | } | |
384 | return false; | |
385 | } | |
386 | ||
387 | /* Helper for vect_set_loop_condition_masked. Generate definitions for | |
388 | all the masks in RGM and return a mask that is nonzero when the loop | |
389 | needs to iterate. Add any new preheader statements to PREHEADER_SEQ. | |
390 | Use LOOP_COND_GSI to insert code before the exit gcond. | |
391 | ||
392 | RGM belongs to loop LOOP. The loop originally iterated NITERS | |
fcae0292 | 393 | times and has been vectorized according to LOOP_VINFO. |
7cfb4d93 | 394 | |
535e7c11 RS |
395 | If NITERS_SKIP is nonnull, the first iteration of the vectorized loop |
396 | starts with NITERS_SKIP dummy iterations of the scalar loop before | |
397 | the real work starts. The mask elements for these dummy iterations | |
398 | must be 0, to ensure that the extra iterations do not have an effect. | |
399 | ||
7cfb4d93 RS |
400 | It is known that: |
401 | ||
402 | NITERS * RGM->max_nscalars_per_iter | |
403 | ||
404 | does not overflow. However, MIGHT_WRAP_P says whether an induction | |
405 | variable that starts at 0 and has step: | |
406 | ||
407 | VF * RGM->max_nscalars_per_iter | |
408 | ||
409 | might overflow before hitting a value above: | |
410 | ||
535e7c11 | 411 | (NITERS + NITERS_SKIP) * RGM->max_nscalars_per_iter |
7cfb4d93 RS |
412 | |
413 | This means that we cannot guarantee that such an induction variable | |
414 | would ever hit a value that produces a set of all-false masks for RGM. */ | |
415 | ||
416 | static tree | |
99b1c316 | 417 | vect_set_loop_masks_directly (class loop *loop, loop_vec_info loop_vinfo, |
7cfb4d93 RS |
418 | gimple_seq *preheader_seq, |
419 | gimple_stmt_iterator loop_cond_gsi, | |
fcae0292 | 420 | rgroup_masks *rgm, tree niters, tree niters_skip, |
535e7c11 | 421 | bool might_wrap_p) |
7cfb4d93 RS |
422 | { |
423 | tree compare_type = LOOP_VINFO_MASK_COMPARE_TYPE (loop_vinfo); | |
9b884225 | 424 | tree iv_type = LOOP_VINFO_MASK_IV_TYPE (loop_vinfo); |
7cfb4d93 RS |
425 | tree mask_type = rgm->mask_type; |
426 | unsigned int nscalars_per_iter = rgm->max_nscalars_per_iter; | |
427 | poly_uint64 nscalars_per_mask = TYPE_VECTOR_SUBPARTS (mask_type); | |
fcae0292 | 428 | poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); |
7cfb4d93 RS |
429 | |
430 | /* Calculate the maximum number of scalar values that the rgroup | |
535e7c11 RS |
431 | handles in total, the number that it handles for each iteration |
432 | of the vector loop, and the number that it should skip during the | |
433 | first iteration of the vector loop. */ | |
7cfb4d93 | 434 | tree nscalars_total = niters; |
fcae0292 | 435 | tree nscalars_step = build_int_cst (iv_type, vf); |
535e7c11 | 436 | tree nscalars_skip = niters_skip; |
7cfb4d93 RS |
437 | if (nscalars_per_iter != 1) |
438 | { | |
439 | /* We checked before choosing to use a fully-masked loop that these | |
440 | multiplications don't overflow. */ | |
fcae0292 RS |
441 | tree compare_factor = build_int_cst (compare_type, nscalars_per_iter); |
442 | tree iv_factor = build_int_cst (iv_type, nscalars_per_iter); | |
7cfb4d93 | 443 | nscalars_total = gimple_build (preheader_seq, MULT_EXPR, compare_type, |
fcae0292 RS |
444 | nscalars_total, compare_factor); |
445 | nscalars_step = gimple_build (preheader_seq, MULT_EXPR, iv_type, | |
446 | nscalars_step, iv_factor); | |
535e7c11 RS |
447 | if (nscalars_skip) |
448 | nscalars_skip = gimple_build (preheader_seq, MULT_EXPR, compare_type, | |
fcae0292 | 449 | nscalars_skip, compare_factor); |
7cfb4d93 RS |
450 | } |
451 | ||
452 | /* Create an induction variable that counts the number of scalars | |
453 | processed. */ | |
454 | tree index_before_incr, index_after_incr; | |
455 | gimple_stmt_iterator incr_gsi; | |
456 | bool insert_after; | |
7cfb4d93 | 457 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); |
fcae0292 RS |
458 | create_iv (build_int_cst (iv_type, 0), nscalars_step, NULL_TREE, loop, |
459 | &incr_gsi, insert_after, &index_before_incr, &index_after_incr); | |
9b884225 | 460 | |
fcae0292 | 461 | tree zero_index = build_int_cst (compare_type, 0); |
535e7c11 | 462 | tree test_index, test_limit, first_limit; |
7cfb4d93 RS |
463 | gimple_stmt_iterator *test_gsi; |
464 | if (might_wrap_p) | |
465 | { | |
466 | /* In principle the loop should stop iterating once the incremented | |
535e7c11 RS |
467 | IV reaches a value greater than or equal to: |
468 | ||
469 | NSCALARS_TOTAL +[infinite-prec] NSCALARS_SKIP | |
470 | ||
471 | However, there's no guarantee that this addition doesn't overflow | |
472 | the comparison type, or that the IV hits a value above it before | |
473 | wrapping around. We therefore adjust the limit down by one | |
474 | IV step: | |
7cfb4d93 | 475 | |
535e7c11 RS |
476 | (NSCALARS_TOTAL +[infinite-prec] NSCALARS_SKIP) |
477 | -[infinite-prec] NSCALARS_STEP | |
7cfb4d93 RS |
478 | |
479 | and compare the IV against this limit _before_ incrementing it. | |
480 | Since the comparison type is unsigned, we actually want the | |
481 | subtraction to saturate at zero: | |
482 | ||
535e7c11 RS |
483 | (NSCALARS_TOTAL +[infinite-prec] NSCALARS_SKIP) |
484 | -[sat] NSCALARS_STEP | |
485 | ||
486 | And since NSCALARS_SKIP < NSCALARS_STEP, we can reassociate this as: | |
487 | ||
488 | NSCALARS_TOTAL -[sat] (NSCALARS_STEP - NSCALARS_SKIP) | |
489 | ||
490 | where the rightmost subtraction can be done directly in | |
491 | COMPARE_TYPE. */ | |
7cfb4d93 | 492 | test_index = index_before_incr; |
fcae0292 RS |
493 | tree adjust = gimple_convert (preheader_seq, compare_type, |
494 | nscalars_step); | |
535e7c11 RS |
495 | if (nscalars_skip) |
496 | adjust = gimple_build (preheader_seq, MINUS_EXPR, compare_type, | |
497 | adjust, nscalars_skip); | |
7cfb4d93 | 498 | test_limit = gimple_build (preheader_seq, MAX_EXPR, compare_type, |
535e7c11 | 499 | nscalars_total, adjust); |
7cfb4d93 | 500 | test_limit = gimple_build (preheader_seq, MINUS_EXPR, compare_type, |
535e7c11 | 501 | test_limit, adjust); |
7cfb4d93 | 502 | test_gsi = &incr_gsi; |
535e7c11 RS |
503 | |
504 | /* Get a safe limit for the first iteration. */ | |
505 | if (nscalars_skip) | |
506 | { | |
507 | /* The first vector iteration can handle at most NSCALARS_STEP | |
508 | scalars. NSCALARS_STEP <= CONST_LIMIT, and adding | |
509 | NSCALARS_SKIP to that cannot overflow. */ | |
510 | tree const_limit = build_int_cst (compare_type, | |
511 | LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
512 | * nscalars_per_iter); | |
513 | first_limit = gimple_build (preheader_seq, MIN_EXPR, compare_type, | |
514 | nscalars_total, const_limit); | |
515 | first_limit = gimple_build (preheader_seq, PLUS_EXPR, compare_type, | |
516 | first_limit, nscalars_skip); | |
517 | } | |
518 | else | |
519 | /* For the first iteration it doesn't matter whether the IV hits | |
520 | a value above NSCALARS_TOTAL. That only matters for the latch | |
521 | condition. */ | |
522 | first_limit = nscalars_total; | |
7cfb4d93 RS |
523 | } |
524 | else | |
525 | { | |
526 | /* Test the incremented IV, which will always hit a value above | |
527 | the bound before wrapping. */ | |
528 | test_index = index_after_incr; | |
529 | test_limit = nscalars_total; | |
535e7c11 RS |
530 | if (nscalars_skip) |
531 | test_limit = gimple_build (preheader_seq, PLUS_EXPR, compare_type, | |
532 | test_limit, nscalars_skip); | |
7cfb4d93 | 533 | test_gsi = &loop_cond_gsi; |
535e7c11 RS |
534 | |
535 | first_limit = test_limit; | |
7cfb4d93 RS |
536 | } |
537 | ||
fcae0292 RS |
538 | /* Convert the IV value to the comparison type (either a no-op or |
539 | a demotion). */ | |
9b884225 KV |
540 | gimple_seq test_seq = NULL; |
541 | test_index = gimple_convert (&test_seq, compare_type, test_index); | |
542 | gsi_insert_seq_before (test_gsi, test_seq, GSI_SAME_STMT); | |
543 | ||
fcae0292 RS |
544 | /* Provide a definition of each mask in the group. */ |
545 | tree next_mask = NULL_TREE; | |
546 | tree mask; | |
547 | unsigned int i; | |
7cfb4d93 RS |
548 | FOR_EACH_VEC_ELT_REVERSE (rgm->masks, i, mask) |
549 | { | |
550 | /* Previous masks will cover BIAS scalars. This mask covers the | |
551 | next batch. */ | |
552 | poly_uint64 bias = nscalars_per_mask * i; | |
553 | tree bias_tree = build_int_cst (compare_type, bias); | |
554 | gimple *tmp_stmt; | |
555 | ||
556 | /* See whether the first iteration of the vector loop is known | |
557 | to have a full mask. */ | |
558 | poly_uint64 const_limit; | |
559 | bool first_iteration_full | |
535e7c11 | 560 | = (poly_int_tree_p (first_limit, &const_limit) |
7cfb4d93 RS |
561 | && known_ge (const_limit, (i + 1) * nscalars_per_mask)); |
562 | ||
563 | /* Rather than have a new IV that starts at BIAS and goes up to | |
564 | TEST_LIMIT, prefer to use the same 0-based IV for each mask | |
565 | and adjust the bound down by BIAS. */ | |
566 | tree this_test_limit = test_limit; | |
567 | if (i != 0) | |
568 | { | |
569 | this_test_limit = gimple_build (preheader_seq, MAX_EXPR, | |
570 | compare_type, this_test_limit, | |
571 | bias_tree); | |
572 | this_test_limit = gimple_build (preheader_seq, MINUS_EXPR, | |
573 | compare_type, this_test_limit, | |
574 | bias_tree); | |
575 | } | |
576 | ||
535e7c11 RS |
577 | /* Create the initial mask. First include all scalars that |
578 | are within the loop limit. */ | |
7cfb4d93 RS |
579 | tree init_mask = NULL_TREE; |
580 | if (!first_iteration_full) | |
581 | { | |
582 | tree start, end; | |
535e7c11 | 583 | if (first_limit == test_limit) |
7cfb4d93 RS |
584 | { |
585 | /* Use a natural test between zero (the initial IV value) | |
586 | and the loop limit. The "else" block would be valid too, | |
587 | but this choice can avoid the need to load BIAS_TREE into | |
588 | a register. */ | |
589 | start = zero_index; | |
590 | end = this_test_limit; | |
591 | } | |
592 | else | |
593 | { | |
535e7c11 RS |
594 | /* FIRST_LIMIT is the maximum number of scalars handled by the |
595 | first iteration of the vector loop. Test the portion | |
596 | associated with this mask. */ | |
7cfb4d93 | 597 | start = bias_tree; |
535e7c11 | 598 | end = first_limit; |
7cfb4d93 RS |
599 | } |
600 | ||
601 | init_mask = make_temp_ssa_name (mask_type, NULL, "max_mask"); | |
602 | tmp_stmt = vect_gen_while (init_mask, start, end); | |
603 | gimple_seq_add_stmt (preheader_seq, tmp_stmt); | |
604 | } | |
605 | ||
535e7c11 RS |
606 | /* Now AND out the bits that are within the number of skipped |
607 | scalars. */ | |
608 | poly_uint64 const_skip; | |
609 | if (nscalars_skip | |
610 | && !(poly_int_tree_p (nscalars_skip, &const_skip) | |
611 | && known_le (const_skip, bias))) | |
612 | { | |
613 | tree unskipped_mask = vect_gen_while_not (preheader_seq, mask_type, | |
614 | bias_tree, nscalars_skip); | |
615 | if (init_mask) | |
616 | init_mask = gimple_build (preheader_seq, BIT_AND_EXPR, mask_type, | |
617 | init_mask, unskipped_mask); | |
618 | else | |
619 | init_mask = unskipped_mask; | |
620 | } | |
621 | ||
7cfb4d93 RS |
622 | if (!init_mask) |
623 | /* First iteration is full. */ | |
624 | init_mask = build_minus_one_cst (mask_type); | |
625 | ||
626 | /* Get the mask value for the next iteration of the loop. */ | |
627 | next_mask = make_temp_ssa_name (mask_type, NULL, "next_mask"); | |
628 | gcall *call = vect_gen_while (next_mask, test_index, this_test_limit); | |
629 | gsi_insert_before (test_gsi, call, GSI_SAME_STMT); | |
630 | ||
631 | vect_set_loop_mask (loop, mask, init_mask, next_mask); | |
632 | } | |
633 | return next_mask; | |
634 | } | |
635 | ||
636 | /* Make LOOP iterate NITERS times using masking and WHILE_ULT calls. | |
637 | LOOP_VINFO describes the vectorization of LOOP. NITERS is the | |
d1d20a49 RS |
638 | number of iterations of the original scalar loop that should be |
639 | handled by the vector loop. NITERS_MAYBE_ZERO and FINAL_IV are | |
640 | as for vect_set_loop_condition. | |
7cfb4d93 RS |
641 | |
642 | Insert the branch-back condition before LOOP_COND_GSI and return the | |
643 | final gcond. */ | |
644 | ||
645 | static gcond * | |
99b1c316 | 646 | vect_set_loop_condition_masked (class loop *loop, loop_vec_info loop_vinfo, |
7cfb4d93 RS |
647 | tree niters, tree final_iv, |
648 | bool niters_maybe_zero, | |
649 | gimple_stmt_iterator loop_cond_gsi) | |
650 | { | |
651 | gimple_seq preheader_seq = NULL; | |
652 | gimple_seq header_seq = NULL; | |
653 | ||
654 | tree compare_type = LOOP_VINFO_MASK_COMPARE_TYPE (loop_vinfo); | |
655 | unsigned int compare_precision = TYPE_PRECISION (compare_type); | |
7cfb4d93 RS |
656 | tree orig_niters = niters; |
657 | ||
658 | /* Type of the initial value of NITERS. */ | |
659 | tree ni_actual_type = TREE_TYPE (niters); | |
660 | unsigned int ni_actual_precision = TYPE_PRECISION (ni_actual_type); | |
9b884225 | 661 | tree niters_skip = LOOP_VINFO_MASK_SKIP_NITERS (loop_vinfo); |
7cfb4d93 RS |
662 | |
663 | /* Convert NITERS to the same size as the compare. */ | |
664 | if (compare_precision > ni_actual_precision | |
665 | && niters_maybe_zero) | |
666 | { | |
667 | /* We know that there is always at least one iteration, so if the | |
668 | count is zero then it must have wrapped. Cope with this by | |
669 | subtracting 1 before the conversion and adding 1 to the result. */ | |
670 | gcc_assert (TYPE_UNSIGNED (ni_actual_type)); | |
671 | niters = gimple_build (&preheader_seq, PLUS_EXPR, ni_actual_type, | |
672 | niters, build_minus_one_cst (ni_actual_type)); | |
673 | niters = gimple_convert (&preheader_seq, compare_type, niters); | |
674 | niters = gimple_build (&preheader_seq, PLUS_EXPR, compare_type, | |
675 | niters, build_one_cst (compare_type)); | |
676 | } | |
677 | else | |
678 | niters = gimple_convert (&preheader_seq, compare_type, niters); | |
679 | ||
9b884225 | 680 | widest_int iv_limit = vect_iv_limit_for_full_masking (loop_vinfo); |
7cfb4d93 RS |
681 | |
682 | /* Iterate over all the rgroups and fill in their masks. We could use | |
683 | the first mask from any rgroup for the loop condition; here we | |
684 | arbitrarily pick the last. */ | |
685 | tree test_mask = NULL_TREE; | |
686 | rgroup_masks *rgm; | |
687 | unsigned int i; | |
688 | vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); | |
689 | FOR_EACH_VEC_ELT (*masks, i, rgm) | |
690 | if (!rgm->masks.is_empty ()) | |
691 | { | |
692 | /* First try using permutes. This adds a single vector | |
693 | instruction to the loop for each mask, but needs no extra | |
694 | loop invariants or IVs. */ | |
695 | unsigned int nmasks = i + 1; | |
696 | if ((nmasks & 1) == 0) | |
697 | { | |
698 | rgroup_masks *half_rgm = &(*masks)[nmasks / 2 - 1]; | |
699 | if (!half_rgm->masks.is_empty () | |
0a0ef238 | 700 | && vect_maybe_permute_loop_masks (&header_seq, rgm, half_rgm)) |
7cfb4d93 RS |
701 | continue; |
702 | } | |
703 | ||
704 | /* See whether zero-based IV would ever generate all-false masks | |
705 | before wrapping around. */ | |
706 | bool might_wrap_p | |
9b884225 | 707 | = (iv_limit == -1 |
7cfb4d93 RS |
708 | || (wi::min_precision (iv_limit * rgm->max_nscalars_per_iter, |
709 | UNSIGNED) | |
710 | > compare_precision)); | |
711 | ||
712 | /* Set up all masks for this group. */ | |
713 | test_mask = vect_set_loop_masks_directly (loop, loop_vinfo, | |
714 | &preheader_seq, | |
fcae0292 | 715 | loop_cond_gsi, rgm, |
535e7c11 RS |
716 | niters, niters_skip, |
717 | might_wrap_p); | |
7cfb4d93 RS |
718 | } |
719 | ||
720 | /* Emit all accumulated statements. */ | |
721 | add_preheader_seq (loop, preheader_seq); | |
722 | add_header_seq (loop, header_seq); | |
723 | ||
724 | /* Get a boolean result that tells us whether to iterate. */ | |
725 | edge exit_edge = single_exit (loop); | |
726 | tree_code code = (exit_edge->flags & EDGE_TRUE_VALUE) ? EQ_EXPR : NE_EXPR; | |
727 | tree zero_mask = build_zero_cst (TREE_TYPE (test_mask)); | |
728 | gcond *cond_stmt = gimple_build_cond (code, test_mask, zero_mask, | |
729 | NULL_TREE, NULL_TREE); | |
730 | gsi_insert_before (&loop_cond_gsi, cond_stmt, GSI_SAME_STMT); | |
731 | ||
732 | /* The loop iterates (NITERS - 1) / VF + 1 times. | |
733 | Subtract one from this to get the latch count. */ | |
734 | tree step = build_int_cst (compare_type, | |
735 | LOOP_VINFO_VECT_FACTOR (loop_vinfo)); | |
736 | tree niters_minus_one = fold_build2 (PLUS_EXPR, compare_type, niters, | |
737 | build_minus_one_cst (compare_type)); | |
738 | loop->nb_iterations = fold_build2 (TRUNC_DIV_EXPR, compare_type, | |
739 | niters_minus_one, step); | |
740 | ||
741 | if (final_iv) | |
742 | { | |
743 | gassign *assign = gimple_build_assign (final_iv, orig_niters); | |
744 | gsi_insert_on_edge_immediate (single_exit (loop), assign); | |
745 | } | |
746 | ||
747 | return cond_stmt; | |
748 | } | |
749 | ||
750 | /* Like vect_set_loop_condition, but handle the case in which there | |
751 | are no loop masks. */ | |
752 | ||
753 | static gcond * | |
99b1c316 | 754 | vect_set_loop_condition_unmasked (class loop *loop, tree niters, |
7cfb4d93 RS |
755 | tree step, tree final_iv, |
756 | bool niters_maybe_zero, | |
757 | gimple_stmt_iterator loop_cond_gsi) | |
ebfd146a IR |
758 | { |
759 | tree indx_before_incr, indx_after_incr; | |
538dd0b7 DM |
760 | gcond *cond_stmt; |
761 | gcond *orig_cond; | |
0f26839a | 762 | edge pe = loop_preheader_edge (loop); |
ebfd146a | 763 | edge exit_edge = single_exit (loop); |
ebfd146a IR |
764 | gimple_stmt_iterator incr_gsi; |
765 | bool insert_after; | |
ebfd146a | 766 | enum tree_code code; |
0f26839a | 767 | tree niters_type = TREE_TYPE (niters); |
ebfd146a IR |
768 | |
769 | orig_cond = get_loop_exit_condition (loop); | |
770 | gcc_assert (orig_cond); | |
771 | loop_cond_gsi = gsi_for_stmt (orig_cond); | |
772 | ||
0f26839a RS |
773 | tree init, limit; |
774 | if (!niters_maybe_zero && integer_onep (step)) | |
775 | { | |
776 | /* In this case we can use a simple 0-based IV: | |
777 | ||
778 | A: | |
779 | x = 0; | |
780 | do | |
781 | { | |
782 | ... | |
783 | x += 1; | |
784 | } | |
785 | while (x < NITERS); */ | |
786 | code = (exit_edge->flags & EDGE_TRUE_VALUE) ? GE_EXPR : LT_EXPR; | |
787 | init = build_zero_cst (niters_type); | |
788 | limit = niters; | |
789 | } | |
790 | else | |
791 | { | |
792 | /* The following works for all values of NITERS except 0: | |
793 | ||
794 | B: | |
795 | x = 0; | |
796 | do | |
797 | { | |
798 | ... | |
799 | x += STEP; | |
800 | } | |
801 | while (x <= NITERS - STEP); | |
802 | ||
803 | so that the loop continues to iterate if x + STEP - 1 < NITERS | |
804 | but stops if x + STEP - 1 >= NITERS. | |
805 | ||
806 | However, if NITERS is zero, x never hits a value above NITERS - STEP | |
807 | before wrapping around. There are two obvious ways of dealing with | |
808 | this: | |
809 | ||
810 | - start at STEP - 1 and compare x before incrementing it | |
811 | - start at -1 and compare x after incrementing it | |
812 | ||
813 | The latter is simpler and is what we use. The loop in this case | |
814 | looks like: | |
815 | ||
816 | C: | |
817 | x = -1; | |
818 | do | |
819 | { | |
820 | ... | |
821 | x += STEP; | |
822 | } | |
823 | while (x < NITERS - STEP); | |
824 | ||
825 | In both cases the loop limit is NITERS - STEP. */ | |
826 | gimple_seq seq = NULL; | |
827 | limit = force_gimple_operand (niters, &seq, true, NULL_TREE); | |
828 | limit = gimple_build (&seq, MINUS_EXPR, TREE_TYPE (limit), limit, step); | |
829 | if (seq) | |
830 | { | |
831 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
832 | gcc_assert (!new_bb); | |
833 | } | |
834 | if (niters_maybe_zero) | |
835 | { | |
836 | /* Case C. */ | |
837 | code = (exit_edge->flags & EDGE_TRUE_VALUE) ? GE_EXPR : LT_EXPR; | |
838 | init = build_all_ones_cst (niters_type); | |
839 | } | |
840 | else | |
841 | { | |
842 | /* Case B. */ | |
843 | code = (exit_edge->flags & EDGE_TRUE_VALUE) ? GT_EXPR : LE_EXPR; | |
844 | init = build_zero_cst (niters_type); | |
845 | } | |
846 | } | |
847 | ||
ebfd146a IR |
848 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); |
849 | create_iv (init, step, NULL_TREE, loop, | |
850 | &incr_gsi, insert_after, &indx_before_incr, &indx_after_incr); | |
ebfd146a IR |
851 | indx_after_incr = force_gimple_operand_gsi (&loop_cond_gsi, indx_after_incr, |
852 | true, NULL_TREE, true, | |
853 | GSI_SAME_STMT); | |
0f26839a | 854 | limit = force_gimple_operand_gsi (&loop_cond_gsi, limit, true, NULL_TREE, |
ebfd146a IR |
855 | true, GSI_SAME_STMT); |
856 | ||
0f26839a | 857 | cond_stmt = gimple_build_cond (code, indx_after_incr, limit, NULL_TREE, |
ebfd146a IR |
858 | NULL_TREE); |
859 | ||
860 | gsi_insert_before (&loop_cond_gsi, cond_stmt, GSI_SAME_STMT); | |
861 | ||
71118889 | 862 | /* Record the number of latch iterations. */ |
0f26839a RS |
863 | if (limit == niters) |
864 | /* Case A: the loop iterates NITERS times. Subtract one to get the | |
865 | latch count. */ | |
866 | loop->nb_iterations = fold_build2 (MINUS_EXPR, niters_type, niters, | |
867 | build_int_cst (niters_type, 1)); | |
868 | else | |
869 | /* Case B or C: the loop iterates (NITERS - STEP) / STEP + 1 times. | |
870 | Subtract one from this to get the latch count. */ | |
871 | loop->nb_iterations = fold_build2 (TRUNC_DIV_EXPR, niters_type, | |
872 | limit, step); | |
873 | ||
874 | if (final_iv) | |
875 | { | |
876 | gassign *assign = gimple_build_assign (final_iv, MINUS_EXPR, | |
877 | indx_after_incr, init); | |
878 | gsi_insert_on_edge_immediate (single_exit (loop), assign); | |
879 | } | |
7cfb4d93 RS |
880 | |
881 | return cond_stmt; | |
882 | } | |
883 | ||
884 | /* If we're using fully-masked loops, make LOOP iterate: | |
885 | ||
886 | N == (NITERS - 1) / STEP + 1 | |
887 | ||
888 | times. When NITERS is zero, this is equivalent to making the loop | |
889 | execute (1 << M) / STEP times, where M is the precision of NITERS. | |
890 | NITERS_MAYBE_ZERO is true if this last case might occur. | |
891 | ||
892 | If we're not using fully-masked loops, make LOOP iterate: | |
893 | ||
894 | N == (NITERS - STEP) / STEP + 1 | |
895 | ||
896 | times, where NITERS is known to be outside the range [1, STEP - 1]. | |
897 | This is equivalent to making the loop execute NITERS / STEP times | |
898 | when NITERS is nonzero and (1 << M) / STEP times otherwise. | |
899 | NITERS_MAYBE_ZERO again indicates whether this last case might occur. | |
900 | ||
901 | If FINAL_IV is nonnull, it is an SSA name that should be set to | |
902 | N * STEP on exit from the loop. | |
903 | ||
904 | Assumption: the exit-condition of LOOP is the last stmt in the loop. */ | |
905 | ||
906 | void | |
99b1c316 | 907 | vect_set_loop_condition (class loop *loop, loop_vec_info loop_vinfo, |
7cfb4d93 RS |
908 | tree niters, tree step, tree final_iv, |
909 | bool niters_maybe_zero) | |
910 | { | |
911 | gcond *cond_stmt; | |
912 | gcond *orig_cond = get_loop_exit_condition (loop); | |
913 | gimple_stmt_iterator loop_cond_gsi = gsi_for_stmt (orig_cond); | |
914 | ||
915 | if (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) | |
916 | cond_stmt = vect_set_loop_condition_masked (loop, loop_vinfo, niters, | |
917 | final_iv, niters_maybe_zero, | |
918 | loop_cond_gsi); | |
919 | else | |
920 | cond_stmt = vect_set_loop_condition_unmasked (loop, niters, step, | |
921 | final_iv, niters_maybe_zero, | |
922 | loop_cond_gsi); | |
923 | ||
924 | /* Remove old loop exit test. */ | |
b5b56c2a RS |
925 | stmt_vec_info orig_cond_info; |
926 | if (loop_vinfo | |
927 | && (orig_cond_info = loop_vinfo->lookup_stmt (orig_cond))) | |
928 | loop_vinfo->remove_stmt (orig_cond_info); | |
929 | else | |
930 | gsi_remove (&loop_cond_gsi, true); | |
7cfb4d93 RS |
931 | |
932 | if (dump_enabled_p ()) | |
3c2a8ed0 DM |
933 | dump_printf_loc (MSG_NOTE, vect_location, "New loop exit condition: %G", |
934 | cond_stmt); | |
ebfd146a IR |
935 | } |
936 | ||
5ce9450f JJ |
937 | /* Helper routine of slpeel_tree_duplicate_loop_to_edge_cfg. |
938 | For all PHI arguments in FROM->dest and TO->dest from those | |
939 | edges ensure that TO->dest PHI arguments have current_def | |
940 | to that in from. */ | |
941 | ||
942 | static void | |
943 | slpeel_duplicate_current_defs_from_edges (edge from, edge to) | |
944 | { | |
945 | gimple_stmt_iterator gsi_from, gsi_to; | |
946 | ||
947 | for (gsi_from = gsi_start_phis (from->dest), | |
948 | gsi_to = gsi_start_phis (to->dest); | |
14ba8d6d | 949 | !gsi_end_p (gsi_from) && !gsi_end_p (gsi_to);) |
5ce9450f | 950 | { |
355fe088 TS |
951 | gimple *from_phi = gsi_stmt (gsi_from); |
952 | gimple *to_phi = gsi_stmt (gsi_to); | |
5ce9450f | 953 | tree from_arg = PHI_ARG_DEF_FROM_EDGE (from_phi, from); |
1a5da5b6 RB |
954 | tree to_arg = PHI_ARG_DEF_FROM_EDGE (to_phi, to); |
955 | if (virtual_operand_p (from_arg)) | |
956 | { | |
14ba8d6d RB |
957 | gsi_next (&gsi_from); |
958 | continue; | |
959 | } | |
1a5da5b6 RB |
960 | if (virtual_operand_p (to_arg)) |
961 | { | |
14ba8d6d RB |
962 | gsi_next (&gsi_to); |
963 | continue; | |
964 | } | |
1a5da5b6 RB |
965 | if (TREE_CODE (from_arg) != SSA_NAME) |
966 | gcc_assert (operand_equal_p (from_arg, to_arg, 0)); | |
d8564d45 RB |
967 | else if (TREE_CODE (to_arg) == SSA_NAME |
968 | && from_arg != to_arg) | |
1a5da5b6 RB |
969 | { |
970 | if (get_current_def (to_arg) == NULL_TREE) | |
733441e2 RB |
971 | { |
972 | gcc_assert (types_compatible_p (TREE_TYPE (to_arg), | |
973 | TREE_TYPE (get_current_def | |
974 | (from_arg)))); | |
975 | set_current_def (to_arg, get_current_def (from_arg)); | |
976 | } | |
1a5da5b6 | 977 | } |
14ba8d6d RB |
978 | gsi_next (&gsi_from); |
979 | gsi_next (&gsi_to); | |
5ce9450f | 980 | } |
1a5da5b6 RB |
981 | |
982 | gphi *from_phi = get_virtual_phi (from->dest); | |
983 | gphi *to_phi = get_virtual_phi (to->dest); | |
984 | if (from_phi) | |
985 | set_current_def (PHI_ARG_DEF_FROM_EDGE (to_phi, to), | |
986 | get_current_def (PHI_ARG_DEF_FROM_EDGE (from_phi, from))); | |
5ce9450f JJ |
987 | } |
988 | ||
ebfd146a | 989 | |
b8698a0f | 990 | /* Given LOOP this function generates a new copy of it and puts it |
5ce9450f JJ |
991 | on E which is either the entry or exit of LOOP. If SCALAR_LOOP is |
992 | non-NULL, assume LOOP and SCALAR_LOOP are equivalent and copy the | |
993 | basic blocks from SCALAR_LOOP instead of LOOP, but to either the | |
994 | entry or exit of LOOP. */ | |
ebfd146a | 995 | |
99b1c316 MS |
996 | class loop * |
997 | slpeel_tree_duplicate_loop_to_edge_cfg (class loop *loop, | |
998 | class loop *scalar_loop, edge e) | |
ebfd146a | 999 | { |
99b1c316 | 1000 | class loop *new_loop; |
3884da6f | 1001 | basic_block *new_bbs, *bbs, *pbbs; |
ebfd146a IR |
1002 | bool at_exit; |
1003 | bool was_imm_dom; | |
b8698a0f | 1004 | basic_block exit_dest; |
ebfd146a | 1005 | edge exit, new_exit; |
a6c51a12 | 1006 | bool duplicate_outer_loop = false; |
ebfd146a | 1007 | |
2cfc56b9 RB |
1008 | exit = single_exit (loop); |
1009 | at_exit = (e == exit); | |
ebfd146a IR |
1010 | if (!at_exit && e != loop_preheader_edge (loop)) |
1011 | return NULL; | |
1012 | ||
5ce9450f JJ |
1013 | if (scalar_loop == NULL) |
1014 | scalar_loop = loop; | |
1015 | ||
1016 | bbs = XNEWVEC (basic_block, scalar_loop->num_nodes + 1); | |
3884da6f BC |
1017 | pbbs = bbs + 1; |
1018 | get_loop_body_with_size (scalar_loop, pbbs, scalar_loop->num_nodes); | |
a6c51a12 YR |
1019 | /* Allow duplication of outer loops. */ |
1020 | if (scalar_loop->inner) | |
1021 | duplicate_outer_loop = true; | |
ebfd146a | 1022 | /* Check whether duplication is possible. */ |
3884da6f | 1023 | if (!can_copy_bbs_p (pbbs, scalar_loop->num_nodes)) |
ebfd146a IR |
1024 | { |
1025 | free (bbs); | |
1026 | return NULL; | |
1027 | } | |
1028 | ||
1029 | /* Generate new loop structure. */ | |
5ce9450f JJ |
1030 | new_loop = duplicate_loop (scalar_loop, loop_outer (scalar_loop)); |
1031 | duplicate_subloops (scalar_loop, new_loop); | |
ebfd146a | 1032 | |
2cfc56b9 | 1033 | exit_dest = exit->dest; |
b8698a0f L |
1034 | was_imm_dom = (get_immediate_dominator (CDI_DOMINATORS, |
1035 | exit_dest) == loop->header ? | |
ebfd146a IR |
1036 | true : false); |
1037 | ||
2cfc56b9 RB |
1038 | /* Also copy the pre-header, this avoids jumping through hoops to |
1039 | duplicate the loop entry PHI arguments. Create an empty | |
1040 | pre-header unconditionally for this. */ | |
5ce9450f | 1041 | basic_block preheader = split_edge (loop_preheader_edge (scalar_loop)); |
2cfc56b9 | 1042 | edge entry_e = single_pred_edge (preheader); |
3884da6f | 1043 | bbs[0] = preheader; |
5ce9450f | 1044 | new_bbs = XNEWVEC (basic_block, scalar_loop->num_nodes + 1); |
ebfd146a | 1045 | |
5ce9450f JJ |
1046 | exit = single_exit (scalar_loop); |
1047 | copy_bbs (bbs, scalar_loop->num_nodes + 1, new_bbs, | |
ebfd146a | 1048 | &exit, 1, &new_exit, NULL, |
3884da6f | 1049 | at_exit ? loop->latch : e->src, true); |
5ce9450f | 1050 | exit = single_exit (loop); |
3884da6f | 1051 | basic_block new_preheader = new_bbs[0]; |
ebfd146a | 1052 | |
5ce9450f JJ |
1053 | add_phi_args_after_copy (new_bbs, scalar_loop->num_nodes + 1, NULL); |
1054 | ||
1055 | if (scalar_loop != loop) | |
1056 | { | |
1057 | /* If we copied from SCALAR_LOOP rather than LOOP, SSA_NAMEs from | |
1058 | SCALAR_LOOP will have current_def set to SSA_NAMEs in the new_loop, | |
1059 | but LOOP will not. slpeel_update_phi_nodes_for_guard{1,2} expects | |
1060 | the LOOP SSA_NAMEs (on the exit edge and edge from latch to | |
1061 | header) to have current_def set, so copy them over. */ | |
1062 | slpeel_duplicate_current_defs_from_edges (single_exit (scalar_loop), | |
1063 | exit); | |
1064 | slpeel_duplicate_current_defs_from_edges (EDGE_SUCC (scalar_loop->latch, | |
1065 | 0), | |
1066 | EDGE_SUCC (loop->latch, 0)); | |
1067 | } | |
b8698a0f | 1068 | |
ebfd146a IR |
1069 | if (at_exit) /* Add the loop copy at exit. */ |
1070 | { | |
5ce9450f JJ |
1071 | if (scalar_loop != loop) |
1072 | { | |
538dd0b7 | 1073 | gphi_iterator gsi; |
5ce9450f JJ |
1074 | new_exit = redirect_edge_and_branch (new_exit, exit_dest); |
1075 | ||
1076 | for (gsi = gsi_start_phis (exit_dest); !gsi_end_p (gsi); | |
1077 | gsi_next (&gsi)) | |
1078 | { | |
538dd0b7 | 1079 | gphi *phi = gsi.phi (); |
5ce9450f JJ |
1080 | tree orig_arg = PHI_ARG_DEF_FROM_EDGE (phi, e); |
1081 | location_t orig_locus | |
1082 | = gimple_phi_arg_location_from_edge (phi, e); | |
1083 | ||
1084 | add_phi_arg (phi, orig_arg, new_exit, orig_locus); | |
1085 | } | |
1086 | } | |
2cfc56b9 RB |
1087 | redirect_edge_and_branch_force (e, new_preheader); |
1088 | flush_pending_stmts (e); | |
1089 | set_immediate_dominator (CDI_DOMINATORS, new_preheader, e->src); | |
a6c51a12 | 1090 | if (was_imm_dom || duplicate_outer_loop) |
5ce9450f | 1091 | set_immediate_dominator (CDI_DOMINATORS, exit_dest, new_exit->src); |
2cfc56b9 RB |
1092 | |
1093 | /* And remove the non-necessary forwarder again. Keep the other | |
1094 | one so we have a proper pre-header for the loop at the exit edge. */ | |
5ce9450f JJ |
1095 | redirect_edge_pred (single_succ_edge (preheader), |
1096 | single_pred (preheader)); | |
2cfc56b9 | 1097 | delete_basic_block (preheader); |
5ce9450f JJ |
1098 | set_immediate_dominator (CDI_DOMINATORS, scalar_loop->header, |
1099 | loop_preheader_edge (scalar_loop)->src); | |
ebfd146a IR |
1100 | } |
1101 | else /* Add the copy at entry. */ | |
1102 | { | |
5ce9450f JJ |
1103 | if (scalar_loop != loop) |
1104 | { | |
1105 | /* Remove the non-necessary forwarder of scalar_loop again. */ | |
1106 | redirect_edge_pred (single_succ_edge (preheader), | |
1107 | single_pred (preheader)); | |
1108 | delete_basic_block (preheader); | |
1109 | set_immediate_dominator (CDI_DOMINATORS, scalar_loop->header, | |
1110 | loop_preheader_edge (scalar_loop)->src); | |
1111 | preheader = split_edge (loop_preheader_edge (loop)); | |
1112 | entry_e = single_pred_edge (preheader); | |
1113 | } | |
1114 | ||
2cfc56b9 RB |
1115 | redirect_edge_and_branch_force (entry_e, new_preheader); |
1116 | flush_pending_stmts (entry_e); | |
1117 | set_immediate_dominator (CDI_DOMINATORS, new_preheader, entry_e->src); | |
1118 | ||
1119 | redirect_edge_and_branch_force (new_exit, preheader); | |
1120 | flush_pending_stmts (new_exit); | |
1121 | set_immediate_dominator (CDI_DOMINATORS, preheader, new_exit->src); | |
1122 | ||
1123 | /* And remove the non-necessary forwarder again. Keep the other | |
1124 | one so we have a proper pre-header for the loop at the exit edge. */ | |
5ce9450f JJ |
1125 | redirect_edge_pred (single_succ_edge (new_preheader), |
1126 | single_pred (new_preheader)); | |
2cfc56b9 RB |
1127 | delete_basic_block (new_preheader); |
1128 | set_immediate_dominator (CDI_DOMINATORS, new_loop->header, | |
1129 | loop_preheader_edge (new_loop)->src); | |
ebfd146a IR |
1130 | } |
1131 | ||
166b8799 BC |
1132 | /* Skip new preheader since it's deleted if copy loop is added at entry. */ |
1133 | for (unsigned i = (at_exit ? 0 : 1); i < scalar_loop->num_nodes + 1; i++) | |
a6c51a12 | 1134 | rename_variables_in_bb (new_bbs[i], duplicate_outer_loop); |
2cfc56b9 | 1135 | |
5ce9450f JJ |
1136 | if (scalar_loop != loop) |
1137 | { | |
1138 | /* Update new_loop->header PHIs, so that on the preheader | |
1139 | edge they are the ones from loop rather than scalar_loop. */ | |
538dd0b7 | 1140 | gphi_iterator gsi_orig, gsi_new; |
5ce9450f JJ |
1141 | edge orig_e = loop_preheader_edge (loop); |
1142 | edge new_e = loop_preheader_edge (new_loop); | |
1143 | ||
1144 | for (gsi_orig = gsi_start_phis (loop->header), | |
1145 | gsi_new = gsi_start_phis (new_loop->header); | |
1146 | !gsi_end_p (gsi_orig) && !gsi_end_p (gsi_new); | |
1147 | gsi_next (&gsi_orig), gsi_next (&gsi_new)) | |
1148 | { | |
538dd0b7 DM |
1149 | gphi *orig_phi = gsi_orig.phi (); |
1150 | gphi *new_phi = gsi_new.phi (); | |
5ce9450f JJ |
1151 | tree orig_arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, orig_e); |
1152 | location_t orig_locus | |
1153 | = gimple_phi_arg_location_from_edge (orig_phi, orig_e); | |
1154 | ||
1155 | add_phi_arg (new_phi, orig_arg, new_e, orig_locus); | |
1156 | } | |
1157 | } | |
1158 | ||
ebfd146a IR |
1159 | free (new_bbs); |
1160 | free (bbs); | |
1161 | ||
b2b29377 | 1162 | checking_verify_dominators (CDI_DOMINATORS); |
2cfc56b9 | 1163 | |
ebfd146a IR |
1164 | return new_loop; |
1165 | } | |
1166 | ||
1167 | ||
a5e3d614 BC |
1168 | /* Given the condition expression COND, put it as the last statement of |
1169 | GUARD_BB; set both edges' probability; set dominator of GUARD_TO to | |
1170 | DOM_BB; return the skip edge. GUARD_TO is the target basic block to | |
3e907b90 BC |
1171 | skip the loop. PROBABILITY is the skip edge's probability. Mark the |
1172 | new edge as irreducible if IRREDUCIBLE_P is true. */ | |
ebfd146a IR |
1173 | |
1174 | static edge | |
86290011 | 1175 | slpeel_add_loop_guard (basic_block guard_bb, tree cond, |
a5e3d614 | 1176 | basic_block guard_to, basic_block dom_bb, |
357067f2 | 1177 | profile_probability probability, bool irreducible_p) |
ebfd146a IR |
1178 | { |
1179 | gimple_stmt_iterator gsi; | |
1180 | edge new_e, enter_e; | |
538dd0b7 | 1181 | gcond *cond_stmt; |
ebfd146a IR |
1182 | gimple_seq gimplify_stmt_list = NULL; |
1183 | ||
1184 | enter_e = EDGE_SUCC (guard_bb, 0); | |
1185 | enter_e->flags &= ~EDGE_FALLTHRU; | |
1186 | enter_e->flags |= EDGE_FALSE_VALUE; | |
1187 | gsi = gsi_last_bb (guard_bb); | |
1188 | ||
f7a06a98 RG |
1189 | cond = force_gimple_operand_1 (cond, &gimplify_stmt_list, is_gimple_condexpr, |
1190 | NULL_TREE); | |
86290011 | 1191 | if (gimplify_stmt_list) |
a5e3d614 | 1192 | gsi_insert_seq_after (&gsi, gimplify_stmt_list, GSI_NEW_STMT); |
ebfd146a | 1193 | |
a5e3d614 | 1194 | cond_stmt = gimple_build_cond_from_tree (cond, NULL_TREE, NULL_TREE); |
ebfd146a IR |
1195 | gsi = gsi_last_bb (guard_bb); |
1196 | gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT); | |
1197 | ||
1198 | /* Add new edge to connect guard block to the merge/loop-exit block. */ | |
a5e3d614 | 1199 | new_e = make_edge (guard_bb, guard_to, EDGE_TRUE_VALUE); |
e78410bf | 1200 | |
e78410bf | 1201 | new_e->probability = probability; |
3e907b90 BC |
1202 | if (irreducible_p) |
1203 | new_e->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1204 | ||
357067f2 | 1205 | enter_e->probability = probability.invert (); |
a5e3d614 | 1206 | set_immediate_dominator (CDI_DOMINATORS, guard_to, dom_bb); |
5281a167 RB |
1207 | |
1208 | /* Split enter_e to preserve LOOPS_HAVE_PREHEADERS. */ | |
1209 | if (enter_e->dest->loop_father->header == enter_e->dest) | |
1210 | split_edge (enter_e); | |
1211 | ||
ebfd146a IR |
1212 | return new_e; |
1213 | } | |
1214 | ||
1215 | ||
1216 | /* This function verifies that the following restrictions apply to LOOP: | |
a6c51a12 YR |
1217 | (1) it consists of exactly 2 basic blocks - header, and an empty latch |
1218 | for innermost loop and 5 basic blocks for outer-loop. | |
1219 | (2) it is single entry, single exit | |
1220 | (3) its exit condition is the last stmt in the header | |
1221 | (4) E is the entry/exit edge of LOOP. | |
ebfd146a IR |
1222 | */ |
1223 | ||
1224 | bool | |
99b1c316 | 1225 | slpeel_can_duplicate_loop_p (const class loop *loop, const_edge e) |
ebfd146a IR |
1226 | { |
1227 | edge exit_e = single_exit (loop); | |
1228 | edge entry_e = loop_preheader_edge (loop); | |
538dd0b7 | 1229 | gcond *orig_cond = get_loop_exit_condition (loop); |
ebfd146a | 1230 | gimple_stmt_iterator loop_exit_gsi = gsi_last_bb (exit_e->src); |
a6c51a12 | 1231 | unsigned int num_bb = loop->inner? 5 : 2; |
ebfd146a | 1232 | |
62fdbf29 BC |
1233 | /* All loops have an outer scope; the only case loop->outer is NULL is for |
1234 | the function itself. */ | |
1235 | if (!loop_outer (loop) | |
a6c51a12 | 1236 | || loop->num_nodes != num_bb |
ebfd146a IR |
1237 | || !empty_block_p (loop->latch) |
1238 | || !single_exit (loop) | |
1239 | /* Verify that new loop exit condition can be trivially modified. */ | |
1240 | || (!orig_cond || orig_cond != gsi_stmt (loop_exit_gsi)) | |
1241 | || (e != exit_e && e != entry_e)) | |
1242 | return false; | |
1243 | ||
1244 | return true; | |
1245 | } | |
1246 | ||
a5e3d614 BC |
1247 | /* If the loop has a virtual PHI, but exit bb doesn't, create a virtual PHI |
1248 | in the exit bb and rename all the uses after the loop. This simplifies | |
1249 | the *guard[12] routines, which assume loop closed SSA form for all PHIs | |
1250 | (but normally loop closed SSA form doesn't require virtual PHIs to be | |
1251 | in the same form). Doing this early simplifies the checking what | |
e3969868 | 1252 | uses should be renamed. |
ebfd146a | 1253 | |
e3969868 RS |
1254 | If we create a new phi after the loop, return the definition that |
1255 | applies on entry to the loop, otherwise return null. */ | |
1256 | ||
1257 | static tree | |
99b1c316 | 1258 | create_lcssa_for_virtual_phi (class loop *loop) |
ebfd146a | 1259 | { |
538dd0b7 | 1260 | gphi_iterator gsi; |
ebfd146a | 1261 | edge exit_e = single_exit (loop); |
b8698a0f | 1262 | |
e20f6b4b | 1263 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) |
ea057359 | 1264 | if (virtual_operand_p (gimple_phi_result (gsi_stmt (gsi)))) |
e20f6b4b | 1265 | { |
538dd0b7 | 1266 | gphi *phi = gsi.phi (); |
e20f6b4b JJ |
1267 | for (gsi = gsi_start_phis (exit_e->dest); |
1268 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
ea057359 | 1269 | if (virtual_operand_p (gimple_phi_result (gsi_stmt (gsi)))) |
e20f6b4b JJ |
1270 | break; |
1271 | if (gsi_end_p (gsi)) | |
1272 | { | |
b731b390 | 1273 | tree new_vop = copy_ssa_name (PHI_RESULT (phi)); |
538dd0b7 | 1274 | gphi *new_phi = create_phi_node (new_vop, exit_e->dest); |
e20f6b4b JJ |
1275 | tree vop = PHI_ARG_DEF_FROM_EDGE (phi, EDGE_SUCC (loop->latch, 0)); |
1276 | imm_use_iterator imm_iter; | |
355fe088 | 1277 | gimple *stmt; |
e20f6b4b JJ |
1278 | use_operand_p use_p; |
1279 | ||
15b6695a RB |
1280 | SSA_NAME_OCCURS_IN_ABNORMAL_PHI (new_vop) |
1281 | = SSA_NAME_OCCURS_IN_ABNORMAL_PHI (vop); | |
9e227d60 | 1282 | add_phi_arg (new_phi, vop, exit_e, UNKNOWN_LOCATION); |
e20f6b4b JJ |
1283 | gimple_phi_set_result (new_phi, new_vop); |
1284 | FOR_EACH_IMM_USE_STMT (stmt, imm_iter, vop) | |
b5fd0440 YR |
1285 | if (stmt != new_phi |
1286 | && !flow_bb_inside_loop_p (loop, gimple_bb (stmt))) | |
e20f6b4b JJ |
1287 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) |
1288 | SET_USE (use_p, new_vop); | |
e3969868 RS |
1289 | |
1290 | return PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); | |
e20f6b4b JJ |
1291 | } |
1292 | break; | |
1293 | } | |
e3969868 | 1294 | return NULL_TREE; |
ebfd146a IR |
1295 | } |
1296 | ||
1297 | /* Function vect_get_loop_location. | |
1298 | ||
1299 | Extract the location of the loop in the source code. | |
1300 | If the loop is not well formed for vectorization, an estimated | |
1301 | location is calculated. | |
1302 | Return the loop location if succeed and NULL if not. */ | |
1303 | ||
4f5b9c80 | 1304 | dump_user_location_t |
99b1c316 | 1305 | find_loop_location (class loop *loop) |
ebfd146a | 1306 | { |
355fe088 | 1307 | gimple *stmt = NULL; |
ebfd146a IR |
1308 | basic_block bb; |
1309 | gimple_stmt_iterator si; | |
1310 | ||
1311 | if (!loop) | |
4f5b9c80 | 1312 | return dump_user_location_t (); |
ebfd146a IR |
1313 | |
1314 | stmt = get_loop_exit_condition (loop); | |
1315 | ||
502498d5 JJ |
1316 | if (stmt |
1317 | && LOCATION_LOCUS (gimple_location (stmt)) > BUILTINS_LOCATION) | |
4f5b9c80 | 1318 | return stmt; |
ebfd146a IR |
1319 | |
1320 | /* If we got here the loop is probably not "well formed", | |
1321 | try to estimate the loop location */ | |
1322 | ||
1323 | if (!loop->header) | |
4f5b9c80 | 1324 | return dump_user_location_t (); |
ebfd146a IR |
1325 | |
1326 | bb = loop->header; | |
1327 | ||
1328 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
1329 | { | |
1330 | stmt = gsi_stmt (si); | |
502498d5 | 1331 | if (LOCATION_LOCUS (gimple_location (stmt)) > BUILTINS_LOCATION) |
4f5b9c80 | 1332 | return stmt; |
ebfd146a IR |
1333 | } |
1334 | ||
4f5b9c80 | 1335 | return dump_user_location_t (); |
ebfd146a IR |
1336 | } |
1337 | ||
82570274 RS |
1338 | /* Return true if the phi described by STMT_INFO defines an IV of the |
1339 | loop to be vectorized. */ | |
a5e3d614 BC |
1340 | |
1341 | static bool | |
82570274 | 1342 | iv_phi_p (stmt_vec_info stmt_info) |
a5e3d614 | 1343 | { |
82570274 | 1344 | gphi *phi = as_a <gphi *> (stmt_info->stmt); |
a5e3d614 BC |
1345 | if (virtual_operand_p (PHI_RESULT (phi))) |
1346 | return false; | |
1347 | ||
a5e3d614 BC |
1348 | if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def |
1349 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_double_reduction_def) | |
1350 | return false; | |
1351 | ||
1352 | return true; | |
1353 | } | |
ebfd146a | 1354 | |
ebfd146a IR |
1355 | /* Function vect_can_advance_ivs_p |
1356 | ||
b8698a0f L |
1357 | In case the number of iterations that LOOP iterates is unknown at compile |
1358 | time, an epilog loop will be generated, and the loop induction variables | |
1359 | (IVs) will be "advanced" to the value they are supposed to take just before | |
ebfd146a IR |
1360 | the epilog loop. Here we check that the access function of the loop IVs |
1361 | and the expression that represents the loop bound are simple enough. | |
1362 | These restrictions will be relaxed in the future. */ | |
1363 | ||
b8698a0f | 1364 | bool |
ebfd146a IR |
1365 | vect_can_advance_ivs_p (loop_vec_info loop_vinfo) |
1366 | { | |
99b1c316 | 1367 | class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
ebfd146a | 1368 | basic_block bb = loop->header; |
538dd0b7 | 1369 | gphi_iterator gsi; |
ebfd146a IR |
1370 | |
1371 | /* Analyze phi functions of the loop header. */ | |
1372 | ||
73fbfcad | 1373 | if (dump_enabled_p ()) |
e645e942 | 1374 | dump_printf_loc (MSG_NOTE, vect_location, "vect_can_advance_ivs_p:\n"); |
ebfd146a IR |
1375 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
1376 | { | |
ebfd146a IR |
1377 | tree evolution_part; |
1378 | ||
a5e3d614 | 1379 | gphi *phi = gsi.phi (); |
91987857 | 1380 | stmt_vec_info phi_info = loop_vinfo->lookup_stmt (phi); |
73fbfcad | 1381 | if (dump_enabled_p ()) |
3c2a8ed0 DM |
1382 | dump_printf_loc (MSG_NOTE, vect_location, "Analyze phi: %G", |
1383 | phi_info->stmt); | |
ebfd146a IR |
1384 | |
1385 | /* Skip virtual phi's. The data dependences that are associated with | |
a5e3d614 | 1386 | virtual defs/uses (i.e., memory accesses) are analyzed elsewhere. |
ebfd146a | 1387 | |
a5e3d614 | 1388 | Skip reduction phis. */ |
82570274 | 1389 | if (!iv_phi_p (phi_info)) |
ebfd146a | 1390 | { |
73fbfcad | 1391 | if (dump_enabled_p ()) |
a5e3d614 BC |
1392 | dump_printf_loc (MSG_NOTE, vect_location, |
1393 | "reduc or virtual phi. skip.\n"); | |
ebfd146a IR |
1394 | continue; |
1395 | } | |
1396 | ||
ebfd146a IR |
1397 | /* Analyze the evolution function. */ |
1398 | ||
91987857 | 1399 | evolution_part = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (phi_info); |
ebfd146a IR |
1400 | if (evolution_part == NULL_TREE) |
1401 | { | |
73fbfcad | 1402 | if (dump_enabled_p ()) |
afb119be | 1403 | dump_printf (MSG_MISSED_OPTIMIZATION, |
e645e942 | 1404 | "No access function or evolution.\n"); |
ebfd146a IR |
1405 | return false; |
1406 | } | |
b8698a0f | 1407 | |
2a93954e AH |
1408 | /* FORNOW: We do not transform initial conditions of IVs |
1409 | which evolution functions are not invariants in the loop. */ | |
1410 | ||
1411 | if (!expr_invariant_in_loop_p (loop, evolution_part)) | |
1412 | { | |
1413 | if (dump_enabled_p ()) | |
1414 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1415 | "evolution not invariant in loop.\n"); | |
1416 | return false; | |
1417 | } | |
1418 | ||
b8698a0f | 1419 | /* FORNOW: We do not transform initial conditions of IVs |
ebfd146a IR |
1420 | which evolution functions are a polynomial of degree >= 2. */ |
1421 | ||
1422 | if (tree_is_chrec (evolution_part)) | |
2a93954e AH |
1423 | { |
1424 | if (dump_enabled_p ()) | |
1425 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1426 | "evolution is chrec.\n"); | |
1427 | return false; | |
1428 | } | |
ebfd146a IR |
1429 | } |
1430 | ||
1431 | return true; | |
1432 | } | |
1433 | ||
1434 | ||
1435 | /* Function vect_update_ivs_after_vectorizer. | |
1436 | ||
1437 | "Advance" the induction variables of LOOP to the value they should take | |
1438 | after the execution of LOOP. This is currently necessary because the | |
1439 | vectorizer does not handle induction variables that are used after the | |
1440 | loop. Such a situation occurs when the last iterations of LOOP are | |
1441 | peeled, because: | |
1442 | 1. We introduced new uses after LOOP for IVs that were not originally used | |
1443 | after LOOP: the IVs of LOOP are now used by an epilog loop. | |
1444 | 2. LOOP is going to be vectorized; this means that it will iterate N/VF | |
1445 | times, whereas the loop IVs should be bumped N times. | |
1446 | ||
1447 | Input: | |
1448 | - LOOP - a loop that is going to be vectorized. The last few iterations | |
1449 | of LOOP were peeled. | |
1450 | - NITERS - the number of iterations that LOOP executes (before it is | |
1451 | vectorized). i.e, the number of times the ivs should be bumped. | |
1452 | - UPDATE_E - a successor edge of LOOP->exit that is on the (only) path | |
1453 | coming out from LOOP on which there are uses of the LOOP ivs | |
1454 | (this is the path from LOOP->exit to epilog_loop->preheader). | |
1455 | ||
1456 | The new definitions of the ivs are placed in LOOP->exit. | |
1457 | The phi args associated with the edge UPDATE_E in the bb | |
1458 | UPDATE_E->dest are updated accordingly. | |
1459 | ||
1460 | Assumption 1: Like the rest of the vectorizer, this function assumes | |
1461 | a single loop exit that has a single predecessor. | |
1462 | ||
1463 | Assumption 2: The phi nodes in the LOOP header and in update_bb are | |
1464 | organized in the same order. | |
1465 | ||
1466 | Assumption 3: The access function of the ivs is simple enough (see | |
1467 | vect_can_advance_ivs_p). This assumption will be relaxed in the future. | |
1468 | ||
1469 | Assumption 4: Exactly one of the successors of LOOP exit-bb is on a path | |
b8698a0f | 1470 | coming out of LOOP on which the ivs of LOOP are used (this is the path |
ebfd146a IR |
1471 | that leads to the epilog loop; other paths skip the epilog loop). This |
1472 | path starts with the edge UPDATE_E, and its destination (denoted update_bb) | |
1473 | needs to have its phis updated. | |
1474 | */ | |
1475 | ||
1476 | static void | |
a5e3d614 BC |
1477 | vect_update_ivs_after_vectorizer (loop_vec_info loop_vinfo, |
1478 | tree niters, edge update_e) | |
ebfd146a | 1479 | { |
538dd0b7 | 1480 | gphi_iterator gsi, gsi1; |
99b1c316 | 1481 | class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
ebfd146a | 1482 | basic_block update_bb = update_e->dest; |
a5e3d614 | 1483 | basic_block exit_bb = single_exit (loop)->dest; |
ebfd146a IR |
1484 | |
1485 | /* Make sure there exists a single-predecessor exit bb: */ | |
1486 | gcc_assert (single_pred_p (exit_bb)); | |
a5e3d614 | 1487 | gcc_assert (single_succ_edge (exit_bb) == update_e); |
ebfd146a IR |
1488 | |
1489 | for (gsi = gsi_start_phis (loop->header), gsi1 = gsi_start_phis (update_bb); | |
1490 | !gsi_end_p (gsi) && !gsi_end_p (gsi1); | |
1491 | gsi_next (&gsi), gsi_next (&gsi1)) | |
1492 | { | |
ebfd146a | 1493 | tree init_expr; |
550918ca RG |
1494 | tree step_expr, off; |
1495 | tree type; | |
ebfd146a IR |
1496 | tree var, ni, ni_name; |
1497 | gimple_stmt_iterator last_gsi; | |
1498 | ||
a5e3d614 BC |
1499 | gphi *phi = gsi.phi (); |
1500 | gphi *phi1 = gsi1.phi (); | |
91987857 | 1501 | stmt_vec_info phi_info = loop_vinfo->lookup_stmt (phi); |
73fbfcad | 1502 | if (dump_enabled_p ()) |
3c2a8ed0 DM |
1503 | dump_printf_loc (MSG_NOTE, vect_location, |
1504 | "vect_update_ivs_after_vectorizer: phi: %G", phi); | |
ebfd146a | 1505 | |
a5e3d614 | 1506 | /* Skip reduction and virtual phis. */ |
82570274 | 1507 | if (!iv_phi_p (phi_info)) |
ebfd146a | 1508 | { |
73fbfcad | 1509 | if (dump_enabled_p ()) |
a5e3d614 BC |
1510 | dump_printf_loc (MSG_NOTE, vect_location, |
1511 | "reduc or virtual phi. skip.\n"); | |
ebfd146a IR |
1512 | continue; |
1513 | } | |
1514 | ||
0ac168a1 | 1515 | type = TREE_TYPE (gimple_phi_result (phi)); |
91987857 | 1516 | step_expr = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (phi_info); |
0ac168a1 | 1517 | step_expr = unshare_expr (step_expr); |
b8698a0f | 1518 | |
ebfd146a IR |
1519 | /* FORNOW: We do not support IVs whose evolution function is a polynomial |
1520 | of degree >= 2 or exponential. */ | |
0ac168a1 | 1521 | gcc_assert (!tree_is_chrec (step_expr)); |
ebfd146a | 1522 | |
0ac168a1 | 1523 | init_expr = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); |
ebfd146a | 1524 | |
550918ca RG |
1525 | off = fold_build2 (MULT_EXPR, TREE_TYPE (step_expr), |
1526 | fold_convert (TREE_TYPE (step_expr), niters), | |
1527 | step_expr); | |
0ac168a1 | 1528 | if (POINTER_TYPE_P (type)) |
5d49b6a7 | 1529 | ni = fold_build_pointer_plus (init_expr, off); |
ebfd146a | 1530 | else |
0ac168a1 RG |
1531 | ni = fold_build2 (PLUS_EXPR, type, |
1532 | init_expr, fold_convert (type, off)); | |
ebfd146a | 1533 | |
0ac168a1 | 1534 | var = create_tmp_var (type, "tmp"); |
ebfd146a IR |
1535 | |
1536 | last_gsi = gsi_last_bb (exit_bb); | |
a5e3d614 BC |
1537 | gimple_seq new_stmts = NULL; |
1538 | ni_name = force_gimple_operand (ni, &new_stmts, false, var); | |
1539 | /* Exit_bb shouldn't be empty. */ | |
1540 | if (!gsi_end_p (last_gsi)) | |
1541 | gsi_insert_seq_after (&last_gsi, new_stmts, GSI_SAME_STMT); | |
1542 | else | |
1543 | gsi_insert_seq_before (&last_gsi, new_stmts, GSI_SAME_STMT); | |
b8698a0f | 1544 | |
ebfd146a | 1545 | /* Fix phi expressions in the successor bb. */ |
684f25f4 | 1546 | adjust_phi_and_debug_stmts (phi1, update_e, ni_name); |
ebfd146a IR |
1547 | } |
1548 | } | |
1549 | ||
535e7c11 RS |
1550 | /* Return a gimple value containing the misalignment (measured in vector |
1551 | elements) for the loop described by LOOP_VINFO, i.e. how many elements | |
1552 | it is away from a perfectly aligned address. Add any new statements | |
1553 | to SEQ. */ | |
1554 | ||
1555 | static tree | |
1556 | get_misalign_in_elems (gimple **seq, loop_vec_info loop_vinfo) | |
1557 | { | |
1e5e6ff5 | 1558 | dr_vec_info *dr_info = LOOP_VINFO_UNALIGNED_DR (loop_vinfo); |
89fa689a | 1559 | stmt_vec_info stmt_info = dr_info->stmt; |
535e7c11 RS |
1560 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
1561 | ||
ca31798e AV |
1562 | poly_uint64 target_align = DR_TARGET_ALIGNMENT (dr_info); |
1563 | unsigned HOST_WIDE_INT target_align_c; | |
1564 | tree target_align_minus_1; | |
535e7c11 | 1565 | |
89fa689a RS |
1566 | bool negative = tree_int_cst_compare (DR_STEP (dr_info->dr), |
1567 | size_zero_node) < 0; | |
535e7c11 RS |
1568 | tree offset = (negative |
1569 | ? size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1) | |
1570 | : size_zero_node); | |
308bc496 RB |
1571 | tree start_addr = vect_create_addr_base_for_vector_ref (loop_vinfo, |
1572 | stmt_info, seq, | |
535e7c11 RS |
1573 | offset); |
1574 | tree type = unsigned_type_for (TREE_TYPE (start_addr)); | |
ca31798e AV |
1575 | if (target_align.is_constant (&target_align_c)) |
1576 | target_align_minus_1 = build_int_cst (type, target_align_c - 1); | |
1577 | else | |
1578 | { | |
1579 | tree vla = build_int_cst (type, target_align); | |
1580 | tree vla_align = fold_build2 (BIT_AND_EXPR, type, vla, | |
1581 | fold_build2 (MINUS_EXPR, type, | |
1582 | build_int_cst (type, 0), vla)); | |
1583 | target_align_minus_1 = fold_build2 (MINUS_EXPR, type, vla_align, | |
1584 | build_int_cst (type, 1)); | |
1585 | } | |
1586 | ||
535e7c11 RS |
1587 | HOST_WIDE_INT elem_size |
1588 | = int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
1589 | tree elem_size_log = build_int_cst (type, exact_log2 (elem_size)); | |
1590 | ||
1591 | /* Create: misalign_in_bytes = addr & (target_align - 1). */ | |
1592 | tree int_start_addr = fold_convert (type, start_addr); | |
1593 | tree misalign_in_bytes = fold_build2 (BIT_AND_EXPR, type, int_start_addr, | |
1594 | target_align_minus_1); | |
1595 | ||
1596 | /* Create: misalign_in_elems = misalign_in_bytes / element_size. */ | |
1597 | tree misalign_in_elems = fold_build2 (RSHIFT_EXPR, type, misalign_in_bytes, | |
1598 | elem_size_log); | |
1599 | ||
1600 | return misalign_in_elems; | |
1601 | } | |
1602 | ||
a5e3d614 | 1603 | /* Function vect_gen_prolog_loop_niters |
d9157f15 | 1604 | |
a5e3d614 BC |
1605 | Generate the number of iterations which should be peeled as prolog for the |
1606 | loop represented by LOOP_VINFO. It is calculated as the misalignment of | |
1607 | DR - the data reference recorded in LOOP_VINFO_UNALIGNED_DR (LOOP_VINFO). | |
1608 | As a result, after the execution of this loop, the data reference DR will | |
1609 | refer to an aligned location. The following computation is generated: | |
ebfd146a IR |
1610 | |
1611 | If the misalignment of DR is known at compile time: | |
1612 | addr_mis = int mis = DR_MISALIGNMENT (dr); | |
1613 | Else, compute address misalignment in bytes: | |
535e7c11 | 1614 | addr_mis = addr & (target_align - 1) |
ebfd146a | 1615 | |
a5e3d614 | 1616 | prolog_niters = ((VF - addr_mis/elem_size)&(VF-1))/step |
ebfd146a IR |
1617 | |
1618 | (elem_size = element type size; an element is the scalar element whose type | |
1619 | is the inner type of the vectype) | |
1620 | ||
a5e3d614 | 1621 | The computations will be emitted at the end of BB. We also compute and |
cbb22e61 | 1622 | store upper bound (included) of the result in BOUND. |
a5e3d614 | 1623 | |
ebfd146a IR |
1624 | When the step of the data-ref in the loop is not 1 (as in interleaved data |
1625 | and SLP), the number of iterations of the prolog must be divided by the step | |
1626 | (which is equal to the size of interleaved group). | |
1627 | ||
1628 | The above formulas assume that VF == number of elements in the vector. This | |
1629 | may not hold when there are multiple-types in the loop. | |
1630 | In this case, for some data-references in the loop the VF does not represent | |
1631 | the number of elements that fit in the vector. Therefore, instead of VF we | |
1632 | use TYPE_VECTOR_SUBPARTS. */ | |
1633 | ||
b8698a0f | 1634 | static tree |
a5e3d614 BC |
1635 | vect_gen_prolog_loop_niters (loop_vec_info loop_vinfo, |
1636 | basic_block bb, int *bound) | |
ebfd146a | 1637 | { |
1e5e6ff5 | 1638 | dr_vec_info *dr_info = LOOP_VINFO_UNALIGNED_DR (loop_vinfo); |
ebfd146a | 1639 | tree var; |
a5e3d614 BC |
1640 | tree niters_type = TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo)); |
1641 | gimple_seq stmts = NULL, new_stmts = NULL; | |
ebfd146a | 1642 | tree iters, iters_name; |
89fa689a | 1643 | stmt_vec_info stmt_info = dr_info->stmt; |
ebfd146a | 1644 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ca31798e | 1645 | poly_uint64 target_align = DR_TARGET_ALIGNMENT (dr_info); |
ebfd146a | 1646 | |
15e693cc | 1647 | if (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0) |
ebfd146a | 1648 | { |
15e693cc | 1649 | int npeel = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo); |
ebfd146a | 1650 | |
73fbfcad | 1651 | if (dump_enabled_p ()) |
ccb3ad87 | 1652 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1653 | "known peeling = %d.\n", npeel); |
ebfd146a | 1654 | |
720f5239 | 1655 | iters = build_int_cst (niters_type, npeel); |
cbb22e61 | 1656 | *bound = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo); |
ebfd146a IR |
1657 | } |
1658 | else | |
1659 | { | |
535e7c11 RS |
1660 | tree misalign_in_elems = get_misalign_in_elems (&stmts, loop_vinfo); |
1661 | tree type = TREE_TYPE (misalign_in_elems); | |
f702e7d4 RS |
1662 | HOST_WIDE_INT elem_size |
1663 | = int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
ca31798e AV |
1664 | /* We only do prolog peeling if the target alignment is known at compile |
1665 | time. */ | |
1666 | poly_uint64 align_in_elems = | |
1667 | exact_div (target_align, elem_size); | |
1668 | tree align_in_elems_minus_1 = | |
1669 | build_int_cst (type, align_in_elems - 1); | |
f702e7d4 | 1670 | tree align_in_elems_tree = build_int_cst (type, align_in_elems); |
f702e7d4 RS |
1671 | |
1672 | /* Create: (niters_type) ((align_in_elems - misalign_in_elems) | |
1673 | & (align_in_elems - 1)). */ | |
89fa689a RS |
1674 | bool negative = tree_int_cst_compare (DR_STEP (dr_info->dr), |
1675 | size_zero_node) < 0; | |
d8ba5b19 | 1676 | if (negative) |
f702e7d4 RS |
1677 | iters = fold_build2 (MINUS_EXPR, type, misalign_in_elems, |
1678 | align_in_elems_tree); | |
d8ba5b19 | 1679 | else |
f702e7d4 RS |
1680 | iters = fold_build2 (MINUS_EXPR, type, align_in_elems_tree, |
1681 | misalign_in_elems); | |
1682 | iters = fold_build2 (BIT_AND_EXPR, type, iters, align_in_elems_minus_1); | |
ebfd146a | 1683 | iters = fold_convert (niters_type, iters); |
ca31798e AV |
1684 | unsigned HOST_WIDE_INT align_in_elems_c; |
1685 | if (align_in_elems.is_constant (&align_in_elems_c)) | |
1686 | *bound = align_in_elems_c - 1; | |
1687 | else | |
1688 | *bound = -1; | |
ebfd146a IR |
1689 | } |
1690 | ||
73fbfcad | 1691 | if (dump_enabled_p ()) |
3c2a8ed0 DM |
1692 | dump_printf_loc (MSG_NOTE, vect_location, |
1693 | "niters for prolog loop: %T\n", iters); | |
ebfd146a IR |
1694 | |
1695 | var = create_tmp_var (niters_type, "prolog_loop_niters"); | |
a5e3d614 | 1696 | iters_name = force_gimple_operand (iters, &new_stmts, false, var); |
ebfd146a | 1697 | |
a5e3d614 BC |
1698 | if (new_stmts) |
1699 | gimple_seq_add_seq (&stmts, new_stmts); | |
ebfd146a IR |
1700 | if (stmts) |
1701 | { | |
a5e3d614 BC |
1702 | gcc_assert (single_succ_p (bb)); |
1703 | gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
1704 | if (gsi_end_p (gsi)) | |
1705 | gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT); | |
1706 | else | |
1707 | gsi_insert_seq_after (&gsi, stmts, GSI_SAME_STMT); | |
ebfd146a | 1708 | } |
b8698a0f | 1709 | return iters_name; |
ebfd146a IR |
1710 | } |
1711 | ||
1712 | ||
1713 | /* Function vect_update_init_of_dr | |
1714 | ||
535e7c11 RS |
1715 | If CODE is PLUS, the vector loop starts NITERS iterations after the |
1716 | scalar one, otherwise CODE is MINUS and the vector loop starts NITERS | |
1717 | iterations before the scalar one (using masking to skip inactive | |
1718 | elements). This function updates the information recorded in DR to | |
1719 | account for the difference. Specifically, it updates the OFFSET | |
67723321 | 1720 | field of DR_INFO. */ |
ebfd146a IR |
1721 | |
1722 | static void | |
67723321 | 1723 | vect_update_init_of_dr (dr_vec_info *dr_info, tree niters, tree_code code) |
ebfd146a | 1724 | { |
67723321 AV |
1725 | struct data_reference *dr = dr_info->dr; |
1726 | tree offset = dr_info->offset; | |
1727 | if (!offset) | |
1728 | offset = build_zero_cst (sizetype); | |
b8698a0f | 1729 | |
ebfd146a IR |
1730 | niters = fold_build2 (MULT_EXPR, sizetype, |
1731 | fold_convert (sizetype, niters), | |
1732 | fold_convert (sizetype, DR_STEP (dr))); | |
535e7c11 | 1733 | offset = fold_build2 (code, sizetype, |
587aa063 | 1734 | fold_convert (sizetype, offset), niters); |
67723321 | 1735 | dr_info->offset = offset; |
ebfd146a IR |
1736 | } |
1737 | ||
1738 | ||
1739 | /* Function vect_update_inits_of_drs | |
1740 | ||
535e7c11 RS |
1741 | Apply vect_update_inits_of_dr to all accesses in LOOP_VINFO. |
1742 | CODE and NITERS are as for vect_update_inits_of_dr. */ | |
ebfd146a | 1743 | |
97c14603 | 1744 | void |
535e7c11 RS |
1745 | vect_update_inits_of_drs (loop_vec_info loop_vinfo, tree niters, |
1746 | tree_code code) | |
ebfd146a IR |
1747 | { |
1748 | unsigned int i; | |
9771b263 | 1749 | vec<data_reference_p> datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); |
ebfd146a | 1750 | struct data_reference *dr; |
a5e3d614 | 1751 | |
adac3a68 | 1752 | DUMP_VECT_SCOPE ("vect_update_inits_of_dr"); |
a5e3d614 | 1753 | |
97c14603 AV |
1754 | /* Adjust niters to sizetype. We used to insert the stmts on loop preheader |
1755 | here, but since we might use these niters to update the epilogues niters | |
1756 | and data references we can't insert them here as this definition might not | |
1757 | always dominate its uses. */ | |
a5e3d614 | 1758 | if (!types_compatible_p (sizetype, TREE_TYPE (niters))) |
97c14603 | 1759 | niters = fold_convert (sizetype, niters); |
ebfd146a | 1760 | |
9771b263 | 1761 | FOR_EACH_VEC_ELT (datarefs, i, dr) |
5fa23466 | 1762 | { |
f5ae2856 RS |
1763 | dr_vec_info *dr_info = loop_vinfo->lookup_dr (dr); |
1764 | if (!STMT_VINFO_GATHER_SCATTER_P (dr_info->stmt)) | |
67723321 | 1765 | vect_update_init_of_dr (dr_info, niters, code); |
5fa23466 | 1766 | } |
ebfd146a IR |
1767 | } |
1768 | ||
535e7c11 RS |
1769 | /* For the information recorded in LOOP_VINFO prepare the loop for peeling |
1770 | by masking. This involves calculating the number of iterations to | |
1771 | be peeled and then aligning all memory references appropriately. */ | |
1772 | ||
1773 | void | |
1774 | vect_prepare_for_masked_peels (loop_vec_info loop_vinfo) | |
1775 | { | |
1776 | tree misalign_in_elems; | |
1777 | tree type = LOOP_VINFO_MASK_COMPARE_TYPE (loop_vinfo); | |
1778 | ||
1779 | gcc_assert (vect_use_loop_mask_for_alignment_p (loop_vinfo)); | |
1780 | ||
1781 | /* From the information recorded in LOOP_VINFO get the number of iterations | |
1782 | that need to be skipped via masking. */ | |
1783 | if (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0) | |
1784 | { | |
1785 | poly_int64 misalign = (LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
1786 | - LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo)); | |
1787 | misalign_in_elems = build_int_cst (type, misalign); | |
1788 | } | |
1789 | else | |
1790 | { | |
1791 | gimple_seq seq1 = NULL, seq2 = NULL; | |
1792 | misalign_in_elems = get_misalign_in_elems (&seq1, loop_vinfo); | |
1793 | misalign_in_elems = fold_convert (type, misalign_in_elems); | |
1794 | misalign_in_elems = force_gimple_operand (misalign_in_elems, | |
1795 | &seq2, true, NULL_TREE); | |
1796 | gimple_seq_add_seq (&seq1, seq2); | |
1797 | if (seq1) | |
1798 | { | |
1799 | edge pe = loop_preheader_edge (LOOP_VINFO_LOOP (loop_vinfo)); | |
1800 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq1); | |
1801 | gcc_assert (!new_bb); | |
1802 | } | |
1803 | } | |
1804 | ||
1805 | if (dump_enabled_p ()) | |
3c2a8ed0 DM |
1806 | dump_printf_loc (MSG_NOTE, vect_location, |
1807 | "misalignment for fully-masked loop: %T\n", | |
1808 | misalign_in_elems); | |
535e7c11 RS |
1809 | |
1810 | LOOP_VINFO_MASK_SKIP_NITERS (loop_vinfo) = misalign_in_elems; | |
1811 | ||
1812 | vect_update_inits_of_drs (loop_vinfo, misalign_in_elems, MINUS_EXPR); | |
1813 | } | |
ebfd146a | 1814 | |
a5e3d614 | 1815 | /* This function builds ni_name = number of iterations. Statements |
7078979b BC |
1816 | are emitted on the loop preheader edge. If NEW_VAR_P is not NULL, set |
1817 | it to TRUE if new ssa_var is generated. */ | |
a5e3d614 BC |
1818 | |
1819 | tree | |
7078979b | 1820 | vect_build_loop_niters (loop_vec_info loop_vinfo, bool *new_var_p) |
a5e3d614 BC |
1821 | { |
1822 | tree ni = unshare_expr (LOOP_VINFO_NITERS (loop_vinfo)); | |
1823 | if (TREE_CODE (ni) == INTEGER_CST) | |
1824 | return ni; | |
1825 | else | |
1826 | { | |
1827 | tree ni_name, var; | |
1828 | gimple_seq stmts = NULL; | |
1829 | edge pe = loop_preheader_edge (LOOP_VINFO_LOOP (loop_vinfo)); | |
1830 | ||
1831 | var = create_tmp_var (TREE_TYPE (ni), "niters"); | |
1832 | ni_name = force_gimple_operand (ni, &stmts, false, var); | |
1833 | if (stmts) | |
7078979b BC |
1834 | { |
1835 | gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1836 | if (new_var_p != NULL) | |
1837 | *new_var_p = true; | |
1838 | } | |
a5e3d614 BC |
1839 | |
1840 | return ni_name; | |
1841 | } | |
1842 | } | |
1843 | ||
cbb22e61 BC |
1844 | /* Calculate the number of iterations above which vectorized loop will be |
1845 | preferred than scalar loop. NITERS_PROLOG is the number of iterations | |
1846 | of prolog loop. If it's integer const, the integer number is also passed | |
d1d20a49 RS |
1847 | in INT_NITERS_PROLOG. BOUND_PROLOG is the upper bound (inclusive) of the |
1848 | number of iterations of the prolog loop. BOUND_EPILOG is the corresponding | |
1849 | value for the epilog loop. If CHECK_PROFITABILITY is true, TH is the | |
1850 | threshold below which the scalar (rather than vectorized) loop will be | |
1851 | executed. This function stores the upper bound (inclusive) of the result | |
1852 | in BOUND_SCALAR. */ | |
a5e3d614 BC |
1853 | |
1854 | static tree | |
1855 | vect_gen_scalar_loop_niters (tree niters_prolog, int int_niters_prolog, | |
d1d20a49 | 1856 | int bound_prolog, poly_int64 bound_epilog, int th, |
d9f21f6a RS |
1857 | poly_uint64 *bound_scalar, |
1858 | bool check_profitability) | |
a5e3d614 BC |
1859 | { |
1860 | tree type = TREE_TYPE (niters_prolog); | |
1861 | tree niters = fold_build2 (PLUS_EXPR, type, niters_prolog, | |
d1d20a49 | 1862 | build_int_cst (type, bound_epilog)); |
a5e3d614 | 1863 | |
d1d20a49 | 1864 | *bound_scalar = bound_prolog + bound_epilog; |
a5e3d614 BC |
1865 | if (check_profitability) |
1866 | { | |
cbb22e61 BC |
1867 | /* TH indicates the minimum niters of vectorized loop, while we |
1868 | compute the maximum niters of scalar loop. */ | |
1869 | th--; | |
a5e3d614 BC |
1870 | /* Peeling for constant times. */ |
1871 | if (int_niters_prolog >= 0) | |
1872 | { | |
d1d20a49 | 1873 | *bound_scalar = upper_bound (int_niters_prolog + bound_epilog, th); |
cbb22e61 | 1874 | return build_int_cst (type, *bound_scalar); |
a5e3d614 | 1875 | } |
d1d20a49 RS |
1876 | /* Peeling an unknown number of times. Note that both BOUND_PROLOG |
1877 | and BOUND_EPILOG are inclusive upper bounds. */ | |
1878 | if (known_ge (th, bound_prolog + bound_epilog)) | |
a5e3d614 | 1879 | { |
cbb22e61 | 1880 | *bound_scalar = th; |
a5e3d614 BC |
1881 | return build_int_cst (type, th); |
1882 | } | |
d9f21f6a | 1883 | /* Need to do runtime comparison. */ |
d1d20a49 | 1884 | else if (maybe_gt (th, bound_epilog)) |
d9f21f6a RS |
1885 | { |
1886 | *bound_scalar = upper_bound (*bound_scalar, th); | |
1887 | return fold_build2 (MAX_EXPR, type, | |
1888 | build_int_cst (type, th), niters); | |
1889 | } | |
a5e3d614 BC |
1890 | } |
1891 | return niters; | |
1892 | } | |
1893 | ||
0f26839a RS |
1894 | /* NITERS is the number of times that the original scalar loop executes |
1895 | after peeling. Work out the maximum number of iterations N that can | |
1896 | be handled by the vectorized form of the loop and then either: | |
ebfd146a | 1897 | |
0f26839a | 1898 | a) set *STEP_VECTOR_PTR to the vectorization factor and generate: |
d9157f15 | 1899 | |
0f26839a RS |
1900 | niters_vector = N |
1901 | ||
1902 | b) set *STEP_VECTOR_PTR to one and generate: | |
1903 | ||
1904 | niters_vector = N / vf | |
1905 | ||
1906 | In both cases, store niters_vector in *NITERS_VECTOR_PTR and add | |
1907 | any new statements on the loop preheader edge. NITERS_NO_OVERFLOW | |
1908 | is true if NITERS doesn't overflow (i.e. if NITERS is always nonzero). */ | |
ebfd146a IR |
1909 | |
1910 | void | |
a5e3d614 | 1911 | vect_gen_vector_loop_niters (loop_vec_info loop_vinfo, tree niters, |
0f26839a RS |
1912 | tree *niters_vector_ptr, tree *step_vector_ptr, |
1913 | bool niters_no_overflow) | |
ebfd146a | 1914 | { |
a5e3d614 | 1915 | tree ni_minus_gap, var; |
0f26839a | 1916 | tree niters_vector, step_vector, type = TREE_TYPE (niters); |
d9f21f6a | 1917 | poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); |
a5e3d614 | 1918 | edge pe = loop_preheader_edge (LOOP_VINFO_LOOP (loop_vinfo)); |
0f26839a | 1919 | tree log_vf = NULL_TREE; |
a5e3d614 BC |
1920 | |
1921 | /* If epilogue loop is required because of data accesses with gaps, we | |
1922 | subtract one iteration from the total number of iterations here for | |
1923 | correct calculation of RATIO. */ | |
1924 | if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)) | |
1925 | { | |
7078979b BC |
1926 | ni_minus_gap = fold_build2 (MINUS_EXPR, type, niters, |
1927 | build_one_cst (type)); | |
a5e3d614 BC |
1928 | if (!is_gimple_val (ni_minus_gap)) |
1929 | { | |
7078979b | 1930 | var = create_tmp_var (type, "ni_gap"); |
a5e3d614 BC |
1931 | gimple *stmts = NULL; |
1932 | ni_minus_gap = force_gimple_operand (ni_minus_gap, &stmts, | |
1933 | true, var); | |
1934 | gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1935 | } | |
1936 | } | |
1937 | else | |
1938 | ni_minus_gap = niters; | |
1939 | ||
d9f21f6a | 1940 | unsigned HOST_WIDE_INT const_vf; |
7cfb4d93 RS |
1941 | if (vf.is_constant (&const_vf) |
1942 | && !LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) | |
0f26839a RS |
1943 | { |
1944 | /* Create: niters >> log2(vf) */ | |
1945 | /* If it's known that niters == number of latch executions + 1 doesn't | |
1946 | overflow, we can generate niters >> log2(vf); otherwise we generate | |
1947 | (niters - vf) >> log2(vf) + 1 by using the fact that we know ratio | |
1948 | will be at least one. */ | |
d9f21f6a | 1949 | log_vf = build_int_cst (type, exact_log2 (const_vf)); |
0f26839a RS |
1950 | if (niters_no_overflow) |
1951 | niters_vector = fold_build2 (RSHIFT_EXPR, type, ni_minus_gap, log_vf); | |
1952 | else | |
1953 | niters_vector | |
1954 | = fold_build2 (PLUS_EXPR, type, | |
1955 | fold_build2 (RSHIFT_EXPR, type, | |
1956 | fold_build2 (MINUS_EXPR, type, | |
1957 | ni_minus_gap, | |
1958 | build_int_cst (type, vf)), | |
1959 | log_vf), | |
1960 | build_int_cst (type, 1)); | |
1961 | step_vector = build_one_cst (type); | |
1962 | } | |
a5e3d614 | 1963 | else |
0f26839a RS |
1964 | { |
1965 | niters_vector = ni_minus_gap; | |
1966 | step_vector = build_int_cst (type, vf); | |
1967 | } | |
a5e3d614 BC |
1968 | |
1969 | if (!is_gimple_val (niters_vector)) | |
1970 | { | |
7078979b BC |
1971 | var = create_tmp_var (type, "bnd"); |
1972 | gimple_seq stmts = NULL; | |
a5e3d614 BC |
1973 | niters_vector = force_gimple_operand (niters_vector, &stmts, true, var); |
1974 | gsi_insert_seq_on_edge_immediate (pe, stmts); | |
7078979b BC |
1975 | /* Peeling algorithm guarantees that vector loop bound is at least ONE, |
1976 | we set range information to make niters analyzer's life easier. */ | |
0f26839a | 1977 | if (stmts != NULL && log_vf) |
8e6cdc90 RS |
1978 | set_range_info (niters_vector, VR_RANGE, |
1979 | wi::to_wide (build_int_cst (type, 1)), | |
1980 | wi::to_wide (fold_build2 (RSHIFT_EXPR, type, | |
1981 | TYPE_MAX_VALUE (type), | |
1982 | log_vf))); | |
a5e3d614 BC |
1983 | } |
1984 | *niters_vector_ptr = niters_vector; | |
0f26839a | 1985 | *step_vector_ptr = step_vector; |
a5e3d614 BC |
1986 | |
1987 | return; | |
1988 | } | |
1989 | ||
1990 | /* Given NITERS_VECTOR which is the number of iterations for vectorized | |
1991 | loop specified by LOOP_VINFO after vectorization, compute the number | |
1992 | of iterations before vectorization (niters_vector * vf) and store it | |
1993 | to NITERS_VECTOR_MULT_VF_PTR. */ | |
1994 | ||
1995 | static void | |
1996 | vect_gen_vector_loop_niters_mult_vf (loop_vec_info loop_vinfo, | |
1997 | tree niters_vector, | |
1998 | tree *niters_vector_mult_vf_ptr) | |
1999 | { | |
d9f21f6a RS |
2000 | /* We should be using a step_vector of VF if VF is variable. */ |
2001 | int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo).to_constant (); | |
99b1c316 | 2002 | class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
a5e3d614 BC |
2003 | tree type = TREE_TYPE (niters_vector); |
2004 | tree log_vf = build_int_cst (type, exact_log2 (vf)); | |
2005 | basic_block exit_bb = single_exit (loop)->dest; | |
ebfd146a | 2006 | |
a5e3d614 BC |
2007 | gcc_assert (niters_vector_mult_vf_ptr != NULL); |
2008 | tree niters_vector_mult_vf = fold_build2 (LSHIFT_EXPR, type, | |
2009 | niters_vector, log_vf); | |
2010 | if (!is_gimple_val (niters_vector_mult_vf)) | |
2011 | { | |
2012 | tree var = create_tmp_var (type, "niters_vector_mult_vf"); | |
2013 | gimple_seq stmts = NULL; | |
2014 | niters_vector_mult_vf = force_gimple_operand (niters_vector_mult_vf, | |
2015 | &stmts, true, var); | |
2016 | gimple_stmt_iterator gsi = gsi_start_bb (exit_bb); | |
2017 | gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT); | |
2018 | } | |
2019 | *niters_vector_mult_vf_ptr = niters_vector_mult_vf; | |
2020 | } | |
2021 | ||
b81f2daf RB |
2022 | /* LCSSA_PHI is a lcssa phi of EPILOG loop which is copied from LOOP, |
2023 | this function searches for the corresponding lcssa phi node in exit | |
2024 | bb of LOOP. If it is found, return the phi result; otherwise return | |
2025 | NULL. */ | |
2026 | ||
2027 | static tree | |
2028 | find_guard_arg (class loop *loop, class loop *epilog ATTRIBUTE_UNUSED, | |
2029 | gphi *lcssa_phi) | |
2030 | { | |
2031 | gphi_iterator gsi; | |
2032 | edge e = single_exit (loop); | |
2033 | ||
2034 | gcc_assert (single_pred_p (e->dest)); | |
2035 | for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2036 | { | |
2037 | gphi *phi = gsi.phi (); | |
2038 | if (operand_equal_p (PHI_ARG_DEF (phi, 0), | |
2039 | PHI_ARG_DEF (lcssa_phi, 0), 0)) | |
2040 | return PHI_RESULT (phi); | |
2041 | } | |
2042 | return NULL_TREE; | |
2043 | } | |
2044 | ||
a5e3d614 BC |
2045 | /* Function slpeel_tree_duplicate_loop_to_edge_cfg duplciates FIRST/SECOND |
2046 | from SECOND/FIRST and puts it at the original loop's preheader/exit | |
2047 | edge, the two loops are arranged as below: | |
2048 | ||
2049 | preheader_a: | |
2050 | first_loop: | |
2051 | header_a: | |
2052 | i_1 = PHI<i_0, i_2>; | |
2053 | ... | |
2054 | i_2 = i_1 + 1; | |
2055 | if (cond_a) | |
2056 | goto latch_a; | |
2057 | else | |
2058 | goto between_bb; | |
2059 | latch_a: | |
2060 | goto header_a; | |
2061 | ||
2062 | between_bb: | |
2063 | ;; i_x = PHI<i_2>; ;; LCSSA phi node to be created for FIRST, | |
2064 | ||
2065 | second_loop: | |
2066 | header_b: | |
2067 | i_3 = PHI<i_0, i_4>; ;; Use of i_0 to be replaced with i_x, | |
2068 | or with i_2 if no LCSSA phi is created | |
2069 | under condition of CREATE_LCSSA_FOR_IV_PHIS. | |
2070 | ... | |
2071 | i_4 = i_3 + 1; | |
2072 | if (cond_b) | |
2073 | goto latch_b; | |
2074 | else | |
2075 | goto exit_bb; | |
2076 | latch_b: | |
2077 | goto header_b; | |
2078 | ||
2079 | exit_bb: | |
2080 | ||
2081 | This function creates loop closed SSA for the first loop; update the | |
2082 | second loop's PHI nodes by replacing argument on incoming edge with the | |
2083 | result of newly created lcssa PHI nodes. IF CREATE_LCSSA_FOR_IV_PHIS | |
2084 | is false, Loop closed ssa phis will only be created for non-iv phis for | |
2085 | the first loop. | |
2086 | ||
2087 | This function assumes exit bb of the first loop is preheader bb of the | |
2088 | second loop, i.e, between_bb in the example code. With PHIs updated, | |
2089 | the second loop will execute rest iterations of the first. */ | |
ebfd146a | 2090 | |
a5e3d614 BC |
2091 | static void |
2092 | slpeel_update_phi_nodes_for_loops (loop_vec_info loop_vinfo, | |
99b1c316 | 2093 | class loop *first, class loop *second, |
a5e3d614 BC |
2094 | bool create_lcssa_for_iv_phis) |
2095 | { | |
2096 | gphi_iterator gsi_update, gsi_orig; | |
99b1c316 | 2097 | class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
a5e3d614 BC |
2098 | |
2099 | edge first_latch_e = EDGE_SUCC (first->latch, 0); | |
2100 | edge second_preheader_e = loop_preheader_edge (second); | |
2101 | basic_block between_bb = single_exit (first)->dest; | |
2102 | ||
2103 | gcc_assert (between_bb == second_preheader_e->src); | |
2104 | gcc_assert (single_pred_p (between_bb) && single_succ_p (between_bb)); | |
2105 | /* Either the first loop or the second is the loop to be vectorized. */ | |
2106 | gcc_assert (loop == first || loop == second); | |
2107 | ||
2108 | for (gsi_orig = gsi_start_phis (first->header), | |
2109 | gsi_update = gsi_start_phis (second->header); | |
2110 | !gsi_end_p (gsi_orig) && !gsi_end_p (gsi_update); | |
2111 | gsi_next (&gsi_orig), gsi_next (&gsi_update)) | |
2112 | { | |
2113 | gphi *orig_phi = gsi_orig.phi (); | |
2114 | gphi *update_phi = gsi_update.phi (); | |
2115 | ||
2116 | tree arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, first_latch_e); | |
2117 | /* Generate lcssa PHI node for the first loop. */ | |
2118 | gphi *vect_phi = (loop == first) ? orig_phi : update_phi; | |
82570274 RS |
2119 | stmt_vec_info vect_phi_info = loop_vinfo->lookup_stmt (vect_phi); |
2120 | if (create_lcssa_for_iv_phis || !iv_phi_p (vect_phi_info)) | |
a5e3d614 BC |
2121 | { |
2122 | tree new_res = copy_ssa_name (PHI_RESULT (orig_phi)); | |
2123 | gphi *lcssa_phi = create_phi_node (new_res, between_bb); | |
2124 | add_phi_arg (lcssa_phi, arg, single_exit (first), UNKNOWN_LOCATION); | |
2125 | arg = new_res; | |
2126 | } | |
2127 | ||
2128 | /* Update PHI node in the second loop by replacing arg on the loop's | |
2129 | incoming edge. */ | |
2130 | adjust_phi_and_debug_stmts (update_phi, second_preheader_e, arg); | |
2131 | } | |
b81f2daf RB |
2132 | |
2133 | /* For epilogue peeling we have to make sure to copy all LC PHIs | |
2134 | for correct vectorization of live stmts. */ | |
2135 | if (loop == first) | |
2136 | { | |
2137 | basic_block orig_exit = single_exit (second)->dest; | |
2138 | for (gsi_orig = gsi_start_phis (orig_exit); | |
2139 | !gsi_end_p (gsi_orig); gsi_next (&gsi_orig)) | |
2140 | { | |
2141 | gphi *orig_phi = gsi_orig.phi (); | |
2142 | tree orig_arg = PHI_ARG_DEF (orig_phi, 0); | |
2143 | if (TREE_CODE (orig_arg) != SSA_NAME || virtual_operand_p (orig_arg)) | |
2144 | continue; | |
2145 | ||
2146 | /* Already created in the above loop. */ | |
2147 | if (find_guard_arg (first, second, orig_phi)) | |
2148 | continue; | |
2149 | ||
2150 | tree new_res = copy_ssa_name (orig_arg); | |
2151 | gphi *lcphi = create_phi_node (new_res, between_bb); | |
2152 | add_phi_arg (lcphi, orig_arg, single_exit (first), UNKNOWN_LOCATION); | |
2153 | } | |
2154 | } | |
a5e3d614 BC |
2155 | } |
2156 | ||
2157 | /* Function slpeel_add_loop_guard adds guard skipping from the beginning | |
2158 | of SKIP_LOOP to the beginning of UPDATE_LOOP. GUARD_EDGE and MERGE_EDGE | |
2159 | are two pred edges of the merge point before UPDATE_LOOP. The two loops | |
2160 | appear like below: | |
2161 | ||
2162 | guard_bb: | |
2163 | if (cond) | |
2164 | goto merge_bb; | |
2165 | else | |
2166 | goto skip_loop; | |
2167 | ||
2168 | skip_loop: | |
2169 | header_a: | |
2170 | i_1 = PHI<i_0, i_2>; | |
2171 | ... | |
2172 | i_2 = i_1 + 1; | |
2173 | if (cond_a) | |
2174 | goto latch_a; | |
2175 | else | |
2176 | goto exit_a; | |
2177 | latch_a: | |
2178 | goto header_a; | |
2179 | ||
2180 | exit_a: | |
2181 | i_5 = PHI<i_2>; | |
2182 | ||
2183 | merge_bb: | |
2184 | ;; PHI (i_x = PHI<i_0, i_5>) to be created at merge point. | |
2185 | ||
2186 | update_loop: | |
2187 | header_b: | |
2188 | i_3 = PHI<i_5, i_4>; ;; Use of i_5 to be replaced with i_x. | |
2189 | ... | |
2190 | i_4 = i_3 + 1; | |
2191 | if (cond_b) | |
2192 | goto latch_b; | |
2193 | else | |
2194 | goto exit_bb; | |
2195 | latch_b: | |
2196 | goto header_b; | |
2197 | ||
2198 | exit_bb: | |
2199 | ||
2200 | This function creates PHI nodes at merge_bb and replaces the use of i_5 | |
2201 | in the update_loop's PHI node with the result of new PHI result. */ | |
2202 | ||
2203 | static void | |
99b1c316 MS |
2204 | slpeel_update_phi_nodes_for_guard1 (class loop *skip_loop, |
2205 | class loop *update_loop, | |
a5e3d614 BC |
2206 | edge guard_edge, edge merge_edge) |
2207 | { | |
620e594b | 2208 | location_t merge_loc, guard_loc; |
a5e3d614 BC |
2209 | edge orig_e = loop_preheader_edge (skip_loop); |
2210 | edge update_e = loop_preheader_edge (update_loop); | |
2211 | gphi_iterator gsi_orig, gsi_update; | |
2212 | ||
2213 | for ((gsi_orig = gsi_start_phis (skip_loop->header), | |
2214 | gsi_update = gsi_start_phis (update_loop->header)); | |
2215 | !gsi_end_p (gsi_orig) && !gsi_end_p (gsi_update); | |
2216 | gsi_next (&gsi_orig), gsi_next (&gsi_update)) | |
2217 | { | |
2218 | gphi *orig_phi = gsi_orig.phi (); | |
2219 | gphi *update_phi = gsi_update.phi (); | |
2220 | ||
2221 | /* Generate new phi node at merge bb of the guard. */ | |
2222 | tree new_res = copy_ssa_name (PHI_RESULT (orig_phi)); | |
2223 | gphi *new_phi = create_phi_node (new_res, guard_edge->dest); | |
2224 | ||
2225 | /* Merge bb has two incoming edges: GUARD_EDGE and MERGE_EDGE. Set the | |
2226 | args in NEW_PHI for these edges. */ | |
2227 | tree merge_arg = PHI_ARG_DEF_FROM_EDGE (update_phi, update_e); | |
2228 | tree guard_arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, orig_e); | |
2229 | merge_loc = gimple_phi_arg_location_from_edge (update_phi, update_e); | |
2230 | guard_loc = gimple_phi_arg_location_from_edge (orig_phi, orig_e); | |
2231 | add_phi_arg (new_phi, merge_arg, merge_edge, merge_loc); | |
2232 | add_phi_arg (new_phi, guard_arg, guard_edge, guard_loc); | |
2233 | ||
2234 | /* Update phi in UPDATE_PHI. */ | |
2235 | adjust_phi_and_debug_stmts (update_phi, update_e, new_res); | |
2236 | } | |
2237 | } | |
2238 | ||
a5e3d614 BC |
2239 | /* LOOP and EPILOG are two consecutive loops in CFG and EPILOG is copied |
2240 | from LOOP. Function slpeel_add_loop_guard adds guard skipping from a | |
2241 | point between the two loops to the end of EPILOG. Edges GUARD_EDGE | |
2242 | and MERGE_EDGE are the two pred edges of merge_bb at the end of EPILOG. | |
2243 | The CFG looks like: | |
2244 | ||
2245 | loop: | |
2246 | header_a: | |
2247 | i_1 = PHI<i_0, i_2>; | |
2248 | ... | |
2249 | i_2 = i_1 + 1; | |
2250 | if (cond_a) | |
2251 | goto latch_a; | |
2252 | else | |
2253 | goto exit_a; | |
2254 | latch_a: | |
2255 | goto header_a; | |
2256 | ||
2257 | exit_a: | |
2258 | ||
2259 | guard_bb: | |
2260 | if (cond) | |
2261 | goto merge_bb; | |
2262 | else | |
2263 | goto epilog_loop; | |
2264 | ||
2265 | ;; fall_through_bb | |
2266 | ||
2267 | epilog_loop: | |
2268 | header_b: | |
2269 | i_3 = PHI<i_2, i_4>; | |
2270 | ... | |
2271 | i_4 = i_3 + 1; | |
2272 | if (cond_b) | |
2273 | goto latch_b; | |
2274 | else | |
2275 | goto merge_bb; | |
2276 | latch_b: | |
2277 | goto header_b; | |
2278 | ||
2279 | merge_bb: | |
2280 | ; PHI node (i_y = PHI<i_2, i_4>) to be created at merge point. | |
2281 | ||
2282 | exit_bb: | |
2283 | i_x = PHI<i_4>; ;Use of i_4 to be replaced with i_y in merge_bb. | |
2284 | ||
2285 | For each name used out side EPILOG (i.e - for each name that has a lcssa | |
2286 | phi in exit_bb) we create a new PHI in merge_bb. The new PHI has two | |
2287 | args corresponding to GUARD_EDGE and MERGE_EDGE. Arg for MERGE_EDGE is | |
2288 | the arg of the original PHI in exit_bb, arg for GUARD_EDGE is defined | |
2289 | by LOOP and is found in the exit bb of LOOP. Arg of the original PHI | |
2290 | in exit_bb will also be updated. */ | |
2291 | ||
2292 | static void | |
99b1c316 | 2293 | slpeel_update_phi_nodes_for_guard2 (class loop *loop, class loop *epilog, |
a5e3d614 BC |
2294 | edge guard_edge, edge merge_edge) |
2295 | { | |
2296 | gphi_iterator gsi; | |
2297 | basic_block merge_bb = guard_edge->dest; | |
2298 | ||
2299 | gcc_assert (single_succ_p (merge_bb)); | |
2300 | edge e = single_succ_edge (merge_bb); | |
2301 | basic_block exit_bb = e->dest; | |
2302 | gcc_assert (single_pred_p (exit_bb)); | |
2303 | gcc_assert (single_pred (exit_bb) == single_exit (epilog)->dest); | |
2304 | ||
2305 | for (gsi = gsi_start_phis (exit_bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2306 | { | |
2307 | gphi *update_phi = gsi.phi (); | |
2308 | tree old_arg = PHI_ARG_DEF (update_phi, 0); | |
a5e3d614 | 2309 | |
2e7a4f57 AV |
2310 | tree merge_arg = NULL_TREE; |
2311 | ||
2312 | /* If the old argument is a SSA_NAME use its current_def. */ | |
2313 | if (TREE_CODE (old_arg) == SSA_NAME) | |
2314 | merge_arg = get_current_def (old_arg); | |
2315 | /* If it's a constant or doesn't have a current_def, just use the old | |
2316 | argument. */ | |
a5e3d614 BC |
2317 | if (!merge_arg) |
2318 | merge_arg = old_arg; | |
2319 | ||
2320 | tree guard_arg = find_guard_arg (loop, epilog, update_phi); | |
2321 | /* If the var is live after loop but not a reduction, we simply | |
2322 | use the old arg. */ | |
2323 | if (!guard_arg) | |
2324 | guard_arg = old_arg; | |
2325 | ||
2326 | /* Create new phi node in MERGE_BB: */ | |
2327 | tree new_res = copy_ssa_name (PHI_RESULT (update_phi)); | |
2328 | gphi *merge_phi = create_phi_node (new_res, merge_bb); | |
2329 | ||
2330 | /* MERGE_BB has two incoming edges: GUARD_EDGE and MERGE_EDGE, Set | |
2331 | the two PHI args in merge_phi for these edges. */ | |
2332 | add_phi_arg (merge_phi, merge_arg, merge_edge, UNKNOWN_LOCATION); | |
2333 | add_phi_arg (merge_phi, guard_arg, guard_edge, UNKNOWN_LOCATION); | |
2334 | ||
2335 | /* Update the original phi in exit_bb. */ | |
2336 | adjust_phi_and_debug_stmts (update_phi, e, new_res); | |
2337 | } | |
2338 | } | |
2339 | ||
2340 | /* EPILOG loop is duplicated from the original loop for vectorizing, | |
2341 | the arg of its loop closed ssa PHI needs to be updated. */ | |
2342 | ||
2343 | static void | |
99b1c316 | 2344 | slpeel_update_phi_nodes_for_lcssa (class loop *epilog) |
a5e3d614 BC |
2345 | { |
2346 | gphi_iterator gsi; | |
2347 | basic_block exit_bb = single_exit (epilog)->dest; | |
2348 | ||
2349 | gcc_assert (single_pred_p (exit_bb)); | |
2350 | edge e = EDGE_PRED (exit_bb, 0); | |
2351 | for (gsi = gsi_start_phis (exit_bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2352 | rename_use_op (PHI_ARG_DEF_PTR_FROM_EDGE (gsi.phi (), e)); | |
2353 | } | |
2354 | ||
2355 | /* Function vect_do_peeling. | |
2356 | ||
2357 | Input: | |
2358 | - LOOP_VINFO: Represent a loop to be vectorized, which looks like: | |
2359 | ||
2360 | preheader: | |
2361 | LOOP: | |
2362 | header_bb: | |
2363 | loop_body | |
2364 | if (exit_loop_cond) goto exit_bb | |
2365 | else goto header_bb | |
2366 | exit_bb: | |
2367 | ||
2368 | - NITERS: The number of iterations of the loop. | |
2369 | - NITERSM1: The number of iterations of the loop's latch. | |
2370 | - NITERS_NO_OVERFLOW: No overflow in computing NITERS. | |
2371 | - TH, CHECK_PROFITABILITY: Threshold of niters to vectorize loop if | |
2372 | CHECK_PROFITABILITY is true. | |
2373 | Output: | |
0f26839a | 2374 | - *NITERS_VECTOR and *STEP_VECTOR describe how the main loop should |
7cfb4d93 | 2375 | iterate after vectorization; see vect_set_loop_condition for details. |
0f26839a RS |
2376 | - *NITERS_VECTOR_MULT_VF_VAR is either null or an SSA name that |
2377 | should be set to the number of scalar iterations handled by the | |
2378 | vector loop. The SSA name is only used on exit from the loop. | |
a5e3d614 BC |
2379 | |
2380 | This function peels prolog and epilog from the loop, adds guards skipping | |
2381 | PROLOG and EPILOG for various conditions. As a result, the changed CFG | |
2382 | would look like: | |
2383 | ||
2384 | guard_bb_1: | |
2385 | if (prefer_scalar_loop) goto merge_bb_1 | |
2386 | else goto guard_bb_2 | |
2387 | ||
2388 | guard_bb_2: | |
2389 | if (skip_prolog) goto merge_bb_2 | |
2390 | else goto prolog_preheader | |
2391 | ||
2392 | prolog_preheader: | |
2393 | PROLOG: | |
2394 | prolog_header_bb: | |
2395 | prolog_body | |
2396 | if (exit_prolog_cond) goto prolog_exit_bb | |
2397 | else goto prolog_header_bb | |
2398 | prolog_exit_bb: | |
2399 | ||
2400 | merge_bb_2: | |
2401 | ||
2402 | vector_preheader: | |
2403 | VECTOR LOOP: | |
2404 | vector_header_bb: | |
2405 | vector_body | |
2406 | if (exit_vector_cond) goto vector_exit_bb | |
2407 | else goto vector_header_bb | |
2408 | vector_exit_bb: | |
2409 | ||
2410 | guard_bb_3: | |
2411 | if (skip_epilog) goto merge_bb_3 | |
2412 | else goto epilog_preheader | |
2413 | ||
2414 | merge_bb_1: | |
2415 | ||
2416 | epilog_preheader: | |
2417 | EPILOG: | |
2418 | epilog_header_bb: | |
2419 | epilog_body | |
2420 | if (exit_epilog_cond) goto merge_bb_3 | |
2421 | else goto epilog_header_bb | |
2422 | ||
2423 | merge_bb_3: | |
2424 | ||
2425 | Note this function peels prolog and epilog only if it's necessary, | |
2426 | as well as guards. | |
97c14603 AV |
2427 | This function returns the epilogue loop if a decision was made to vectorize |
2428 | it, otherwise NULL. | |
2429 | ||
2430 | The analysis resulting in this epilogue loop's loop_vec_info was performed | |
2431 | in the same vect_analyze_loop call as the main loop's. At that time | |
2432 | vect_analyze_loop constructs a list of accepted loop_vec_info's for lower | |
2433 | vectorization factors than the main loop. This list is stored in the main | |
2434 | loop's loop_vec_info in the 'epilogue_vinfos' member. Everytime we decide to | |
2435 | vectorize the epilogue loop for a lower vectorization factor, the | |
2436 | loop_vec_info sitting at the top of the epilogue_vinfos list is removed, | |
2437 | updated and linked to the epilogue loop. This is later used to vectorize | |
2438 | the epilogue. The reason the loop_vec_info needs updating is that it was | |
2439 | constructed based on the original main loop, and the epilogue loop is a | |
2440 | copy of this loop, so all links pointing to statements in the original loop | |
2441 | need updating. Furthermore, these loop_vec_infos share the | |
2442 | data_reference's records, which will also need to be updated. | |
a5e3d614 BC |
2443 | |
2444 | TODO: Guard for prefer_scalar_loop should be emitted along with | |
2445 | versioning conditions if loop versioning is needed. */ | |
2446 | ||
598eaaa2 | 2447 | |
99b1c316 | 2448 | class loop * |
a5e3d614 | 2449 | vect_do_peeling (loop_vec_info loop_vinfo, tree niters, tree nitersm1, |
0f26839a RS |
2450 | tree *niters_vector, tree *step_vector, |
2451 | tree *niters_vector_mult_vf_var, int th, | |
97c14603 | 2452 | bool check_profitability, bool niters_no_overflow, |
67723321 | 2453 | tree *advance) |
a5e3d614 BC |
2454 | { |
2455 | edge e, guard_e; | |
2456 | tree type = TREE_TYPE (niters), guard_cond; | |
2457 | basic_block guard_bb, guard_to; | |
357067f2 | 2458 | profile_probability prob_prolog, prob_vector, prob_epilog; |
d9f21f6a | 2459 | int estimated_vf; |
535e7c11 | 2460 | int prolog_peeling = 0; |
97c14603 | 2461 | bool vect_epilogues = loop_vinfo->epilogue_vinfos.length () > 0; |
ca31798e AV |
2462 | /* We currently do not support prolog peeling if the target alignment is not |
2463 | known at compile time. 'vect_gen_prolog_loop_niters' depends on the | |
2464 | target alignment being constant. */ | |
2465 | dr_vec_info *dr_info = LOOP_VINFO_UNALIGNED_DR (loop_vinfo); | |
2466 | if (dr_info && !DR_TARGET_ALIGNMENT (dr_info).is_constant ()) | |
2467 | return NULL; | |
2468 | ||
535e7c11 RS |
2469 | if (!vect_use_loop_mask_for_alignment_p (loop_vinfo)) |
2470 | prolog_peeling = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo); | |
d1d20a49 RS |
2471 | |
2472 | poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
2473 | poly_uint64 bound_epilog = 0; | |
2474 | if (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) | |
2475 | && LOOP_VINFO_PEELING_FOR_NITER (loop_vinfo)) | |
2476 | bound_epilog += vf - 1; | |
2477 | if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)) | |
2478 | bound_epilog += 1; | |
2479 | bool epilog_peeling = maybe_ne (bound_epilog, 0U); | |
2480 | poly_uint64 bound_scalar = bound_epilog; | |
a5e3d614 BC |
2481 | |
2482 | if (!prolog_peeling && !epilog_peeling) | |
598eaaa2 | 2483 | return NULL; |
a5e3d614 | 2484 | |
357067f2 | 2485 | prob_vector = profile_probability::guessed_always ().apply_scale (9, 10); |
d9f21f6a RS |
2486 | estimated_vf = vect_vf_for_cost (loop_vinfo); |
2487 | if (estimated_vf == 2) | |
2488 | estimated_vf = 3; | |
357067f2 | 2489 | prob_prolog = prob_epilog = profile_probability::guessed_always () |
d9f21f6a | 2490 | .apply_scale (estimated_vf - 1, estimated_vf); |
a5e3d614 | 2491 | |
99b1c316 MS |
2492 | class loop *prolog, *epilog = NULL, *loop = LOOP_VINFO_LOOP (loop_vinfo); |
2493 | class loop *first_loop = loop; | |
3e907b90 | 2494 | bool irred_flag = loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP; |
e3969868 RS |
2495 | |
2496 | /* We might have a queued need to update virtual SSA form. As we | |
2497 | delete the update SSA machinery below after doing a regular | |
2498 | incremental SSA update during loop copying make sure we don't | |
2499 | lose that fact. | |
2500 | ??? Needing to update virtual SSA form by renaming is unfortunate | |
2501 | but not all of the vectorizer code inserting new loads / stores | |
2502 | properly assigns virtual operands to those statements. */ | |
a5e3d614 BC |
2503 | update_ssa (TODO_update_ssa_only_virtuals); |
2504 | ||
e3969868 RS |
2505 | create_lcssa_for_virtual_phi (loop); |
2506 | ||
2507 | /* If we're vectorizing an epilogue loop, the update_ssa above will | |
2508 | have ensured that the virtual operand is in SSA form throughout the | |
2509 | vectorized main loop. Normally it is possible to trace the updated | |
2510 | vector-stmt vdefs back to scalar-stmt vdefs and vector-stmt vuses | |
2511 | back to scalar-stmt vuses, meaning that the effect of the SSA update | |
2512 | remains local to the main loop. However, there are rare cases in | |
2513 | which the vectorized loop has vdefs even when the original scalar | |
2514 | loop didn't. For example, vectorizing a load with IFN_LOAD_LANES | |
2515 | introduces clobbers of the temporary vector array, which in turn | |
2516 | needs new vdefs. If the scalar loop doesn't write to memory, these | |
2517 | new vdefs will be the only ones in the vector loop. | |
2518 | ||
2519 | In that case, update_ssa will have added a new virtual phi to the | |
2520 | main loop, which previously didn't need one. Ensure that we (locally) | |
2521 | maintain LCSSA form for the virtual operand, just as we would have | |
2522 | done if the virtual phi had existed from the outset. This makes it | |
2523 | easier to duplicate the scalar epilogue loop below. */ | |
2524 | tree vop_to_rename = NULL_TREE; | |
2525 | if (loop_vec_info orig_loop_vinfo = LOOP_VINFO_ORIG_LOOP_INFO (loop_vinfo)) | |
2526 | { | |
2527 | class loop *orig_loop = LOOP_VINFO_LOOP (orig_loop_vinfo); | |
2528 | vop_to_rename = create_lcssa_for_virtual_phi (orig_loop); | |
2529 | } | |
2530 | ||
36f52e8f | 2531 | if (MAY_HAVE_DEBUG_BIND_STMTS) |
a5e3d614 BC |
2532 | { |
2533 | gcc_assert (!adjust_vec.exists ()); | |
2534 | adjust_vec.create (32); | |
2535 | } | |
ebfd146a IR |
2536 | initialize_original_copy_tables (); |
2537 | ||
d1d20a49 RS |
2538 | /* Record the anchor bb at which the guard should be placed if the scalar |
2539 | loop might be preferred. */ | |
2540 | basic_block anchor = loop_preheader_edge (loop)->src; | |
2541 | ||
2542 | /* Generate the number of iterations for the prolog loop. We do this here | |
2543 | so that we can also get the upper bound on the number of iterations. */ | |
2544 | tree niters_prolog; | |
2545 | int bound_prolog = 0; | |
2546 | if (prolog_peeling) | |
2547 | niters_prolog = vect_gen_prolog_loop_niters (loop_vinfo, anchor, | |
97c14603 | 2548 | &bound_prolog); |
d1d20a49 RS |
2549 | else |
2550 | niters_prolog = build_int_cst (type, 0); | |
2551 | ||
97c14603 AV |
2552 | loop_vec_info epilogue_vinfo = NULL; |
2553 | if (vect_epilogues) | |
2554 | { | |
2555 | epilogue_vinfo = loop_vinfo->epilogue_vinfos[0]; | |
2556 | loop_vinfo->epilogue_vinfos.ordered_remove (0); | |
2557 | } | |
2558 | ||
2559 | tree niters_vector_mult_vf = NULL_TREE; | |
2560 | /* Saving NITERs before the loop, as this may be changed by prologue. */ | |
2561 | tree before_loop_niters = LOOP_VINFO_NITERS (loop_vinfo); | |
2562 | edge update_e = NULL, skip_e = NULL; | |
2563 | unsigned int lowest_vf = constant_lower_bound (vf); | |
2564 | /* If we know the number of scalar iterations for the main loop we should | |
2565 | check whether after the main loop there are enough iterations left over | |
2566 | for the epilogue. */ | |
2567 | if (vect_epilogues | |
2568 | && LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) | |
2569 | && prolog_peeling >= 0 | |
2570 | && known_eq (vf, lowest_vf)) | |
2571 | { | |
2572 | unsigned HOST_WIDE_INT eiters | |
2573 | = (LOOP_VINFO_INT_NITERS (loop_vinfo) | |
2574 | - LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)); | |
2575 | ||
2576 | eiters -= prolog_peeling; | |
2577 | eiters | |
2578 | = eiters % lowest_vf + LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo); | |
2579 | ||
2580 | unsigned int ratio; | |
87b47251 AV |
2581 | unsigned int epilogue_gaps |
2582 | = LOOP_VINFO_PEELING_FOR_GAPS (epilogue_vinfo); | |
1c84a2d2 RS |
2583 | while (!(constant_multiple_p |
2584 | (GET_MODE_SIZE (loop_vinfo->vector_mode), | |
2585 | GET_MODE_SIZE (epilogue_vinfo->vector_mode), &ratio) | |
87b47251 | 2586 | && eiters >= lowest_vf / ratio + epilogue_gaps)) |
97c14603 AV |
2587 | { |
2588 | delete epilogue_vinfo; | |
2589 | epilogue_vinfo = NULL; | |
2590 | if (loop_vinfo->epilogue_vinfos.length () == 0) | |
2591 | { | |
2592 | vect_epilogues = false; | |
2593 | break; | |
2594 | } | |
2595 | epilogue_vinfo = loop_vinfo->epilogue_vinfos[0]; | |
2596 | loop_vinfo->epilogue_vinfos.ordered_remove (0); | |
87b47251 | 2597 | epilogue_gaps = LOOP_VINFO_PEELING_FOR_GAPS (epilogue_vinfo); |
97c14603 AV |
2598 | } |
2599 | } | |
a5e3d614 BC |
2600 | /* Prolog loop may be skipped. */ |
2601 | bool skip_prolog = (prolog_peeling != 0); | |
97c14603 AV |
2602 | /* Skip this loop to epilog when there are not enough iterations to enter this |
2603 | vectorized loop. If true we should perform runtime checks on the NITERS | |
2604 | to check whether we should skip the current vectorized loop. If we know | |
2605 | the number of scalar iterations we may choose to add a runtime check if | |
2606 | this number "maybe" smaller than the number of iterations required | |
2607 | when we know the number of scalar iterations may potentially | |
2608 | be smaller than the number of iterations required to enter this loop, for | |
2609 | this we use the upper bounds on the prolog and epilog peeling. When we | |
2610 | don't know the number of iterations and don't require versioning it is | |
2611 | because we have asserted that there are enough scalar iterations to enter | |
2612 | the main loop, so this skip is not necessary. When we are versioning then | |
2613 | we only add such a skip if we have chosen to vectorize the epilogue. */ | |
d1d20a49 RS |
2614 | bool skip_vector = (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) |
2615 | ? maybe_lt (LOOP_VINFO_INT_NITERS (loop_vinfo), | |
2616 | bound_prolog + bound_epilog) | |
97c14603 AV |
2617 | : (!LOOP_REQUIRES_VERSIONING (loop_vinfo) |
2618 | || vect_epilogues)); | |
a5e3d614 BC |
2619 | /* Epilog loop must be executed if the number of iterations for epilog |
2620 | loop is known at compile time, otherwise we need to add a check at | |
2621 | the end of vector loop and skip to the end of epilog loop. */ | |
2622 | bool skip_epilog = (prolog_peeling < 0 | |
d9f21f6a RS |
2623 | || !LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) |
2624 | || !vf.is_constant ()); | |
a5e3d614 BC |
2625 | /* PEELING_FOR_GAPS is special because epilog loop must be executed. */ |
2626 | if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)) | |
2627 | skip_epilog = false; | |
2628 | ||
a5e3d614 | 2629 | if (skip_vector) |
25c99850 BC |
2630 | { |
2631 | split_edge (loop_preheader_edge (loop)); | |
2632 | ||
2633 | /* Due to the order in which we peel prolog and epilog, we first | |
2634 | propagate probability to the whole loop. The purpose is to | |
2635 | avoid adjusting probabilities of both prolog and vector loops | |
2636 | separately. Note in this case, the probability of epilog loop | |
2637 | needs to be scaled back later. */ | |
2638 | basic_block bb_before_loop = loop_preheader_edge (loop)->src; | |
357067f2 | 2639 | if (prob_vector.initialized_p ()) |
d9f21f6a RS |
2640 | { |
2641 | scale_bbs_frequencies (&bb_before_loop, 1, prob_vector); | |
2642 | scale_loop_profile (loop, prob_vector, 0); | |
2643 | } | |
25c99850 | 2644 | } |
a5e3d614 | 2645 | |
4f5b9c80 | 2646 | dump_user_location_t loop_loc = find_loop_location (loop); |
99b1c316 | 2647 | class loop *scalar_loop = LOOP_VINFO_SCALAR_LOOP (loop_vinfo); |
97c14603 AV |
2648 | if (vect_epilogues) |
2649 | /* Make sure to set the epilogue's epilogue scalar loop, such that we can | |
2650 | use the original scalar loop as remaining epilogue if necessary. */ | |
2651 | LOOP_VINFO_SCALAR_LOOP (epilogue_vinfo) | |
2652 | = LOOP_VINFO_SCALAR_LOOP (loop_vinfo); | |
2653 | ||
a5e3d614 | 2654 | if (prolog_peeling) |
5d2eb24b | 2655 | { |
a5e3d614 BC |
2656 | e = loop_preheader_edge (loop); |
2657 | if (!slpeel_can_duplicate_loop_p (loop, e)) | |
5d2eb24b | 2658 | { |
a5e3d614 BC |
2659 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, loop_loc, |
2660 | "loop can't be duplicated to preheader edge.\n"); | |
2661 | gcc_unreachable (); | |
2662 | } | |
2663 | /* Peel prolog and put it on preheader edge of loop. */ | |
2664 | prolog = slpeel_tree_duplicate_loop_to_edge_cfg (loop, scalar_loop, e); | |
2665 | if (!prolog) | |
2666 | { | |
2667 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, loop_loc, | |
2668 | "slpeel_tree_duplicate_loop_to_edge_cfg failed.\n"); | |
2669 | gcc_unreachable (); | |
2670 | } | |
03001a35 | 2671 | prolog->force_vectorize = false; |
a5e3d614 BC |
2672 | slpeel_update_phi_nodes_for_loops (loop_vinfo, prolog, loop, true); |
2673 | first_loop = prolog; | |
2674 | reset_original_copy_tables (); | |
2675 | ||
d1d20a49 | 2676 | /* Update the number of iterations for prolog loop. */ |
0f26839a | 2677 | tree step_prolog = build_one_cst (TREE_TYPE (niters_prolog)); |
7cfb4d93 RS |
2678 | vect_set_loop_condition (prolog, NULL, niters_prolog, |
2679 | step_prolog, NULL_TREE, false); | |
a5e3d614 BC |
2680 | |
2681 | /* Skip the prolog loop. */ | |
2682 | if (skip_prolog) | |
2683 | { | |
2684 | guard_cond = fold_build2 (EQ_EXPR, boolean_type_node, | |
2685 | niters_prolog, build_int_cst (type, 0)); | |
2686 | guard_bb = loop_preheader_edge (prolog)->src; | |
25c99850 | 2687 | basic_block bb_after_prolog = loop_preheader_edge (loop)->src; |
a5e3d614 BC |
2688 | guard_to = split_edge (loop_preheader_edge (loop)); |
2689 | guard_e = slpeel_add_loop_guard (guard_bb, guard_cond, | |
2690 | guard_to, guard_bb, | |
357067f2 | 2691 | prob_prolog.invert (), |
3e907b90 | 2692 | irred_flag); |
a5e3d614 BC |
2693 | e = EDGE_PRED (guard_to, 0); |
2694 | e = (e != guard_e ? e : EDGE_PRED (guard_to, 1)); | |
2695 | slpeel_update_phi_nodes_for_guard1 (prolog, loop, guard_e, e); | |
25c99850 | 2696 | |
af2bbc51 JH |
2697 | scale_bbs_frequencies (&bb_after_prolog, 1, prob_prolog); |
2698 | scale_loop_profile (prolog, prob_prolog, bound_prolog); | |
a5e3d614 | 2699 | } |
97c14603 | 2700 | |
a5e3d614 | 2701 | /* Update init address of DRs. */ |
535e7c11 | 2702 | vect_update_inits_of_drs (loop_vinfo, niters_prolog, PLUS_EXPR); |
a5e3d614 BC |
2703 | /* Update niters for vector loop. */ |
2704 | LOOP_VINFO_NITERS (loop_vinfo) | |
2705 | = fold_build2 (MINUS_EXPR, type, niters, niters_prolog); | |
2706 | LOOP_VINFO_NITERSM1 (loop_vinfo) | |
2707 | = fold_build2 (MINUS_EXPR, type, | |
2708 | LOOP_VINFO_NITERSM1 (loop_vinfo), niters_prolog); | |
7078979b BC |
2709 | bool new_var_p = false; |
2710 | niters = vect_build_loop_niters (loop_vinfo, &new_var_p); | |
2711 | /* It's guaranteed that vector loop bound before vectorization is at | |
2712 | least VF, so set range information for newly generated var. */ | |
2713 | if (new_var_p) | |
2714 | set_range_info (niters, VR_RANGE, | |
8e6cdc90 RS |
2715 | wi::to_wide (build_int_cst (type, vf)), |
2716 | wi::to_wide (TYPE_MAX_VALUE (type))); | |
a5e3d614 | 2717 | |
cbb22e61 BC |
2718 | /* Prolog iterates at most bound_prolog times, latch iterates at |
2719 | most bound_prolog - 1 times. */ | |
2720 | record_niter_bound (prolog, bound_prolog - 1, false, true); | |
a5e3d614 BC |
2721 | delete_update_ssa (); |
2722 | adjust_vec_debug_stmts (); | |
2723 | scev_reset (); | |
5d2eb24b IR |
2724 | } |
2725 | ||
a5e3d614 BC |
2726 | if (epilog_peeling) |
2727 | { | |
2728 | e = single_exit (loop); | |
2729 | if (!slpeel_can_duplicate_loop_p (loop, e)) | |
2730 | { | |
2731 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, loop_loc, | |
2732 | "loop can't be duplicated to exit edge.\n"); | |
2733 | gcc_unreachable (); | |
2734 | } | |
97c14603 AV |
2735 | /* Peel epilog and put it on exit edge of loop. If we are vectorizing |
2736 | said epilog then we should use a copy of the main loop as a starting | |
2737 | point. This loop may have already had some preliminary transformations | |
2738 | to allow for more optimal vectorization, for example if-conversion. | |
2739 | If we are not vectorizing the epilog then we should use the scalar loop | |
2740 | as the transformations mentioned above make less or no sense when not | |
2741 | vectorizing. */ | |
2742 | epilog = vect_epilogues ? get_loop_copy (loop) : scalar_loop; | |
e3969868 RS |
2743 | if (vop_to_rename) |
2744 | { | |
2745 | /* Vectorizing the main loop can sometimes introduce a vdef to | |
2746 | a loop that previously didn't have one; see the comment above | |
2747 | the definition of VOP_TO_RENAME for details. The definition | |
2748 | D that holds on E will then be different from the definition | |
2749 | VOP_TO_RENAME that holds during SCALAR_LOOP, so we need to | |
2750 | rename VOP_TO_RENAME to D when copying the loop. | |
2751 | ||
2752 | The virtual operand is in LCSSA form for the main loop, | |
2753 | and no stmt between the main loop and E needs a vdef, | |
2754 | so we know that D is provided by a phi rather than by a | |
2755 | vdef on a normal gimple stmt. */ | |
2756 | basic_block vdef_bb = e->src; | |
2757 | gphi *vphi; | |
2758 | while (!(vphi = get_virtual_phi (vdef_bb))) | |
2759 | vdef_bb = get_immediate_dominator (CDI_DOMINATORS, vdef_bb); | |
2760 | gcc_assert (vop_to_rename != gimple_phi_result (vphi)); | |
2761 | set_current_def (vop_to_rename, gimple_phi_result (vphi)); | |
2762 | } | |
97c14603 | 2763 | epilog = slpeel_tree_duplicate_loop_to_edge_cfg (loop, epilog, e); |
a5e3d614 BC |
2764 | if (!epilog) |
2765 | { | |
2766 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, loop_loc, | |
2767 | "slpeel_tree_duplicate_loop_to_edge_cfg failed.\n"); | |
2768 | gcc_unreachable (); | |
2769 | } | |
03001a35 | 2770 | epilog->force_vectorize = false; |
a5e3d614 BC |
2771 | slpeel_update_phi_nodes_for_loops (loop_vinfo, loop, epilog, false); |
2772 | ||
2773 | /* Scalar version loop may be preferred. In this case, add guard | |
2774 | and skip to epilog. Note this only happens when the number of | |
2775 | iterations of loop is unknown at compile time, otherwise this | |
2776 | won't be vectorized. */ | |
2777 | if (skip_vector) | |
2778 | { | |
cbb22e61 | 2779 | /* Additional epilogue iteration is peeled if gap exists. */ |
a5e3d614 | 2780 | tree t = vect_gen_scalar_loop_niters (niters_prolog, prolog_peeling, |
d1d20a49 | 2781 | bound_prolog, bound_epilog, |
cbb22e61 | 2782 | th, &bound_scalar, |
a5e3d614 | 2783 | check_profitability); |
cbb22e61 BC |
2784 | /* Build guard against NITERSM1 since NITERS may overflow. */ |
2785 | guard_cond = fold_build2 (LT_EXPR, boolean_type_node, nitersm1, t); | |
a5e3d614 BC |
2786 | guard_bb = anchor; |
2787 | guard_to = split_edge (loop_preheader_edge (epilog)); | |
2788 | guard_e = slpeel_add_loop_guard (guard_bb, guard_cond, | |
2789 | guard_to, guard_bb, | |
357067f2 | 2790 | prob_vector.invert (), |
3e907b90 | 2791 | irred_flag); |
97c14603 | 2792 | skip_e = guard_e; |
a5e3d614 BC |
2793 | e = EDGE_PRED (guard_to, 0); |
2794 | e = (e != guard_e ? e : EDGE_PRED (guard_to, 1)); | |
2795 | slpeel_update_phi_nodes_for_guard1 (first_loop, epilog, guard_e, e); | |
25c99850 BC |
2796 | |
2797 | /* Simply propagate profile info from guard_bb to guard_to which is | |
2798 | a merge point of control flow. */ | |
25c99850 | 2799 | guard_to->count = guard_bb->count; |
00fa28d1 | 2800 | |
af2bbc51 JH |
2801 | /* Scale probability of epilog loop back. |
2802 | FIXME: We should avoid scaling down and back up. Profile may | |
2803 | get lost if we scale down to 0. */ | |
10ea2672 | 2804 | basic_block *bbs = get_loop_body (epilog); |
00fa28d1 JH |
2805 | for (unsigned int i = 0; i < epilog->num_nodes; i++) |
2806 | bbs[i]->count = bbs[i]->count.apply_scale | |
2807 | (bbs[i]->count, | |
2808 | bbs[i]->count.apply_probability | |
2809 | (prob_vector)); | |
af2bbc51 | 2810 | free (bbs); |
a5e3d614 | 2811 | } |
ebfd146a | 2812 | |
25c99850 | 2813 | basic_block bb_before_epilog = loop_preheader_edge (epilog)->src; |
a5e3d614 BC |
2814 | /* If loop is peeled for non-zero constant times, now niters refers to |
2815 | orig_niters - prolog_peeling, it won't overflow even the orig_niters | |
2816 | overflows. */ | |
2817 | niters_no_overflow |= (prolog_peeling > 0); | |
2818 | vect_gen_vector_loop_niters (loop_vinfo, niters, | |
0f26839a RS |
2819 | niters_vector, step_vector, |
2820 | niters_no_overflow); | |
2821 | if (!integer_onep (*step_vector)) | |
2822 | { | |
2823 | /* On exit from the loop we will have an easy way of calcalating | |
2824 | NITERS_VECTOR / STEP * STEP. Install a dummy definition | |
2825 | until then. */ | |
2826 | niters_vector_mult_vf = make_ssa_name (TREE_TYPE (*niters_vector)); | |
2827 | SSA_NAME_DEF_STMT (niters_vector_mult_vf) = gimple_build_nop (); | |
2828 | *niters_vector_mult_vf_var = niters_vector_mult_vf; | |
2829 | } | |
2830 | else | |
2831 | vect_gen_vector_loop_niters_mult_vf (loop_vinfo, *niters_vector, | |
2832 | &niters_vector_mult_vf); | |
a5e3d614 BC |
2833 | /* Update IVs of original loop as if they were advanced by |
2834 | niters_vector_mult_vf steps. */ | |
2835 | gcc_checking_assert (vect_can_advance_ivs_p (loop_vinfo)); | |
97c14603 | 2836 | update_e = skip_vector ? e : loop_preheader_edge (epilog); |
a5e3d614 BC |
2837 | vect_update_ivs_after_vectorizer (loop_vinfo, niters_vector_mult_vf, |
2838 | update_e); | |
2839 | ||
2840 | if (skip_epilog) | |
2841 | { | |
2842 | guard_cond = fold_build2 (EQ_EXPR, boolean_type_node, | |
2843 | niters, niters_vector_mult_vf); | |
2844 | guard_bb = single_exit (loop)->dest; | |
2845 | guard_to = split_edge (single_exit (epilog)); | |
2846 | guard_e = slpeel_add_loop_guard (guard_bb, guard_cond, guard_to, | |
2847 | skip_vector ? anchor : guard_bb, | |
357067f2 | 2848 | prob_epilog.invert (), |
3e907b90 | 2849 | irred_flag); |
a5e3d614 BC |
2850 | slpeel_update_phi_nodes_for_guard2 (loop, epilog, guard_e, |
2851 | single_exit (epilog)); | |
25c99850 BC |
2852 | /* Only need to handle basic block before epilog loop if it's not |
2853 | the guard_bb, which is the case when skip_vector is true. */ | |
2854 | if (guard_bb != bb_before_epilog) | |
2855 | { | |
357067f2 | 2856 | prob_epilog = prob_vector * prob_epilog + prob_vector.invert (); |
25c99850 | 2857 | |
af2bbc51 | 2858 | scale_bbs_frequencies (&bb_before_epilog, 1, prob_epilog); |
25c99850 | 2859 | } |
d9f21f6a | 2860 | scale_loop_profile (epilog, prob_epilog, 0); |
a5e3d614 BC |
2861 | } |
2862 | else | |
2863 | slpeel_update_phi_nodes_for_lcssa (epilog); | |
ebfd146a | 2864 | |
d1d20a49 RS |
2865 | unsigned HOST_WIDE_INT bound; |
2866 | if (bound_scalar.is_constant (&bound)) | |
d9f21f6a | 2867 | { |
d1d20a49 RS |
2868 | gcc_assert (bound != 0); |
2869 | /* -1 to convert loop iterations to latch iterations. */ | |
2870 | record_niter_bound (epilog, bound - 1, false, true); | |
d9f21f6a | 2871 | } |
a5e3d614 BC |
2872 | |
2873 | delete_update_ssa (); | |
2874 | adjust_vec_debug_stmts (); | |
2875 | scev_reset (); | |
2876 | } | |
97c14603 AV |
2877 | |
2878 | if (vect_epilogues) | |
2879 | { | |
2880 | epilog->aux = epilogue_vinfo; | |
2881 | LOOP_VINFO_LOOP (epilogue_vinfo) = epilog; | |
2882 | ||
2883 | loop_constraint_clear (epilog, LOOP_C_INFINITE); | |
2884 | ||
2885 | /* We now must calculate the number of NITERS performed by the previous | |
2886 | loop and EPILOGUE_NITERS to be performed by the epilogue. */ | |
2887 | tree niters = fold_build2 (PLUS_EXPR, TREE_TYPE (niters_vector_mult_vf), | |
2888 | niters_prolog, niters_vector_mult_vf); | |
2889 | ||
2890 | /* If skip_vector we may skip the previous loop, we insert a phi-node to | |
2891 | determine whether we are coming from the previous vectorized loop | |
2892 | using the update_e edge or the skip_vector basic block using the | |
2893 | skip_e edge. */ | |
2894 | if (skip_vector) | |
2895 | { | |
2896 | gcc_assert (update_e != NULL && skip_e != NULL); | |
2897 | gphi *new_phi = create_phi_node (make_ssa_name (TREE_TYPE (niters)), | |
2898 | update_e->dest); | |
2899 | tree new_ssa = make_ssa_name (TREE_TYPE (niters)); | |
2900 | gimple *stmt = gimple_build_assign (new_ssa, niters); | |
2901 | gimple_stmt_iterator gsi; | |
2902 | if (TREE_CODE (niters_vector_mult_vf) == SSA_NAME | |
2903 | && SSA_NAME_DEF_STMT (niters_vector_mult_vf)->bb != NULL) | |
2904 | { | |
2905 | gsi = gsi_for_stmt (SSA_NAME_DEF_STMT (niters_vector_mult_vf)); | |
2906 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); | |
2907 | } | |
2908 | else | |
2909 | { | |
2910 | gsi = gsi_last_bb (update_e->src); | |
2911 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); | |
2912 | } | |
2913 | ||
2914 | niters = new_ssa; | |
2915 | add_phi_arg (new_phi, niters, update_e, UNKNOWN_LOCATION); | |
2916 | add_phi_arg (new_phi, build_zero_cst (TREE_TYPE (niters)), skip_e, | |
2917 | UNKNOWN_LOCATION); | |
2918 | niters = PHI_RESULT (new_phi); | |
2919 | } | |
2920 | ||
2921 | /* Subtract the number of iterations performed by the vectorized loop | |
2922 | from the number of total iterations. */ | |
2923 | tree epilogue_niters = fold_build2 (MINUS_EXPR, TREE_TYPE (niters), | |
2924 | before_loop_niters, | |
2925 | niters); | |
2926 | ||
2927 | LOOP_VINFO_NITERS (epilogue_vinfo) = epilogue_niters; | |
2928 | LOOP_VINFO_NITERSM1 (epilogue_vinfo) | |
2929 | = fold_build2 (MINUS_EXPR, TREE_TYPE (epilogue_niters), | |
2930 | epilogue_niters, | |
2931 | build_one_cst (TREE_TYPE (epilogue_niters))); | |
2932 | ||
2933 | /* Set ADVANCE to the number of iterations performed by the previous | |
2934 | loop and its prologue. */ | |
2935 | *advance = niters; | |
2936 | ||
2937 | /* Redo the peeling for niter analysis as the NITERs and alignment | |
2938 | may have been updated to take the main loop into account. */ | |
2939 | determine_peel_for_niter (epilogue_vinfo); | |
2940 | } | |
2941 | ||
a5e3d614 | 2942 | adjust_vec.release (); |
ebfd146a | 2943 | free_original_copy_tables (); |
598eaaa2 | 2944 | |
97c14603 | 2945 | return vect_epilogues ? epilog : NULL; |
ebfd146a IR |
2946 | } |
2947 | ||
01d32b2b BC |
2948 | /* Function vect_create_cond_for_niters_checks. |
2949 | ||
2950 | Create a conditional expression that represents the run-time checks for | |
ce811fc4 | 2951 | loop's niter. The loop is guaranteed to terminate if the run-time |
01d32b2b BC |
2952 | checks hold. |
2953 | ||
2954 | Input: | |
2955 | COND_EXPR - input conditional expression. New conditions will be chained | |
2956 | with logical AND operation. If it is NULL, then the function | |
2957 | is used to return the number of alias checks. | |
2958 | LOOP_VINFO - field LOOP_VINFO_MAY_ALIAS_STMTS contains the list of ddrs | |
2959 | to be checked. | |
2960 | ||
2961 | Output: | |
2962 | COND_EXPR - conditional expression. | |
2963 | ||
2964 | The returned COND_EXPR is the conditional expression to be used in the | |
2965 | if statement that controls which version of the loop gets executed at | |
2966 | runtime. */ | |
2967 | ||
2968 | static void | |
2969 | vect_create_cond_for_niters_checks (loop_vec_info loop_vinfo, tree *cond_expr) | |
2970 | { | |
2971 | tree part_cond_expr = LOOP_VINFO_NITERS_ASSUMPTIONS (loop_vinfo); | |
2972 | ||
2973 | if (*cond_expr) | |
2974 | *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
2975 | *cond_expr, part_cond_expr); | |
2976 | else | |
2977 | *cond_expr = part_cond_expr; | |
2978 | } | |
ebfd146a | 2979 | |
9adee305 RS |
2980 | /* Set *COND_EXPR to a tree that is true when both the original *COND_EXPR |
2981 | and PART_COND_EXPR are true. Treat a null *COND_EXPR as "true". */ | |
2982 | ||
2983 | static void | |
2984 | chain_cond_expr (tree *cond_expr, tree part_cond_expr) | |
2985 | { | |
2986 | if (*cond_expr) | |
2987 | *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
2988 | *cond_expr, part_cond_expr); | |
2989 | else | |
2990 | *cond_expr = part_cond_expr; | |
2991 | } | |
2992 | ||
ebfd146a IR |
2993 | /* Function vect_create_cond_for_align_checks. |
2994 | ||
2995 | Create a conditional expression that represents the alignment checks for | |
2996 | all of data references (array element references) whose alignment must be | |
2997 | checked at runtime. | |
2998 | ||
2999 | Input: | |
3000 | COND_EXPR - input conditional expression. New conditions will be chained | |
3001 | with logical AND operation. | |
3002 | LOOP_VINFO - two fields of the loop information are used. | |
3003 | LOOP_VINFO_PTR_MASK is the mask used to check the alignment. | |
3004 | LOOP_VINFO_MAY_MISALIGN_STMTS contains the refs to be checked. | |
3005 | ||
3006 | Output: | |
3007 | COND_EXPR_STMT_LIST - statements needed to construct the conditional | |
3008 | expression. | |
3009 | The returned value is the conditional expression to be used in the if | |
3010 | statement that controls which version of the loop gets executed at runtime. | |
3011 | ||
3012 | The algorithm makes two assumptions: | |
3013 | 1) The number of bytes "n" in a vector is a power of 2. | |
3014 | 2) An address "a" is aligned if a%n is zero and that this | |
3015 | test can be done as a&(n-1) == 0. For example, for 16 | |
3016 | byte vectors the test is a&0xf == 0. */ | |
3017 | ||
3018 | static void | |
3019 | vect_create_cond_for_align_checks (loop_vec_info loop_vinfo, | |
3020 | tree *cond_expr, | |
3021 | gimple_seq *cond_expr_stmt_list) | |
3022 | { | |
7bcbf2d8 | 3023 | vec<stmt_vec_info> may_misalign_stmts |
ebfd146a | 3024 | = LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo); |
7bcbf2d8 | 3025 | stmt_vec_info stmt_info; |
ebfd146a IR |
3026 | int mask = LOOP_VINFO_PTR_MASK (loop_vinfo); |
3027 | tree mask_cst; | |
3028 | unsigned int i; | |
ebfd146a IR |
3029 | tree int_ptrsize_type; |
3030 | char tmp_name[20]; | |
3031 | tree or_tmp_name = NULL_TREE; | |
83d5977e | 3032 | tree and_tmp_name; |
355fe088 | 3033 | gimple *and_stmt; |
ebfd146a IR |
3034 | tree ptrsize_zero; |
3035 | tree part_cond_expr; | |
3036 | ||
3037 | /* Check that mask is one less than a power of 2, i.e., mask is | |
3038 | all zeros followed by all ones. */ | |
3039 | gcc_assert ((mask != 0) && ((mask & (mask+1)) == 0)); | |
3040 | ||
96f9265a | 3041 | int_ptrsize_type = signed_type_for (ptr_type_node); |
ebfd146a IR |
3042 | |
3043 | /* Create expression (mask & (dr_1 || ... || dr_n)) where dr_i is the address | |
3044 | of the first vector of the i'th data reference. */ | |
3045 | ||
7bcbf2d8 | 3046 | FOR_EACH_VEC_ELT (may_misalign_stmts, i, stmt_info) |
ebfd146a IR |
3047 | { |
3048 | gimple_seq new_stmt_list = NULL; | |
3049 | tree addr_base; | |
83d5977e RG |
3050 | tree addr_tmp_name; |
3051 | tree new_or_tmp_name; | |
355fe088 | 3052 | gimple *addr_stmt, *or_stmt; |
7bcbf2d8 | 3053 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
d8ba5b19 | 3054 | bool negative = tree_int_cst_compare |
7bcbf2d8 | 3055 | (DR_STEP (STMT_VINFO_DATA_REF (stmt_info)), size_zero_node) < 0; |
d8ba5b19 | 3056 | tree offset = negative |
aad83b7c | 3057 | ? size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1) : size_zero_node; |
ebfd146a IR |
3058 | |
3059 | /* create: addr_tmp = (int)(address_of_first_vector) */ | |
3060 | addr_base = | |
308bc496 RB |
3061 | vect_create_addr_base_for_vector_ref (loop_vinfo, |
3062 | stmt_info, &new_stmt_list, | |
3f5e8a76 | 3063 | offset); |
ebfd146a IR |
3064 | if (new_stmt_list != NULL) |
3065 | gimple_seq_add_seq (cond_expr_stmt_list, new_stmt_list); | |
3066 | ||
83d5977e RG |
3067 | sprintf (tmp_name, "addr2int%d", i); |
3068 | addr_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, tmp_name); | |
0d0e4a03 | 3069 | addr_stmt = gimple_build_assign (addr_tmp_name, NOP_EXPR, addr_base); |
ebfd146a IR |
3070 | gimple_seq_add_stmt (cond_expr_stmt_list, addr_stmt); |
3071 | ||
3072 | /* The addresses are OR together. */ | |
3073 | ||
3074 | if (or_tmp_name != NULL_TREE) | |
3075 | { | |
3076 | /* create: or_tmp = or_tmp | addr_tmp */ | |
83d5977e RG |
3077 | sprintf (tmp_name, "orptrs%d", i); |
3078 | new_or_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, tmp_name); | |
0d0e4a03 JJ |
3079 | or_stmt = gimple_build_assign (new_or_tmp_name, BIT_IOR_EXPR, |
3080 | or_tmp_name, addr_tmp_name); | |
ebfd146a IR |
3081 | gimple_seq_add_stmt (cond_expr_stmt_list, or_stmt); |
3082 | or_tmp_name = new_or_tmp_name; | |
3083 | } | |
3084 | else | |
3085 | or_tmp_name = addr_tmp_name; | |
3086 | ||
3087 | } /* end for i */ | |
3088 | ||
3089 | mask_cst = build_int_cst (int_ptrsize_type, mask); | |
3090 | ||
3091 | /* create: and_tmp = or_tmp & mask */ | |
83d5977e | 3092 | and_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, "andmask"); |
ebfd146a | 3093 | |
0d0e4a03 JJ |
3094 | and_stmt = gimple_build_assign (and_tmp_name, BIT_AND_EXPR, |
3095 | or_tmp_name, mask_cst); | |
ebfd146a IR |
3096 | gimple_seq_add_stmt (cond_expr_stmt_list, and_stmt); |
3097 | ||
3098 | /* Make and_tmp the left operand of the conditional test against zero. | |
3099 | if and_tmp has a nonzero bit then some address is unaligned. */ | |
3100 | ptrsize_zero = build_int_cst (int_ptrsize_type, 0); | |
3101 | part_cond_expr = fold_build2 (EQ_EXPR, boolean_type_node, | |
3102 | and_tmp_name, ptrsize_zero); | |
9adee305 RS |
3103 | chain_cond_expr (cond_expr, part_cond_expr); |
3104 | } | |
3105 | ||
3106 | /* If LOOP_VINFO_CHECK_UNEQUAL_ADDRS contains <A1, B1>, ..., <An, Bn>, | |
3107 | create a tree representation of: (&A1 != &B1) && ... && (&An != &Bn). | |
3108 | Set *COND_EXPR to a tree that is true when both the original *COND_EXPR | |
3109 | and this new condition are true. Treat a null *COND_EXPR as "true". */ | |
3110 | ||
3111 | static void | |
3112 | vect_create_cond_for_unequal_addrs (loop_vec_info loop_vinfo, tree *cond_expr) | |
3113 | { | |
3114 | vec<vec_object_pair> pairs = LOOP_VINFO_CHECK_UNEQUAL_ADDRS (loop_vinfo); | |
3115 | unsigned int i; | |
3116 | vec_object_pair *pair; | |
3117 | FOR_EACH_VEC_ELT (pairs, i, pair) | |
3118 | { | |
3119 | tree addr1 = build_fold_addr_expr (pair->first); | |
3120 | tree addr2 = build_fold_addr_expr (pair->second); | |
3121 | tree part_cond_expr = fold_build2 (NE_EXPR, boolean_type_node, | |
3122 | addr1, addr2); | |
3123 | chain_cond_expr (cond_expr, part_cond_expr); | |
3124 | } | |
ebfd146a IR |
3125 | } |
3126 | ||
a57776a1 RS |
3127 | /* Create an expression that is true when all lower-bound conditions for |
3128 | the vectorized loop are met. Chain this condition with *COND_EXPR. */ | |
3129 | ||
3130 | static void | |
3131 | vect_create_cond_for_lower_bounds (loop_vec_info loop_vinfo, tree *cond_expr) | |
3132 | { | |
3133 | vec<vec_lower_bound> lower_bounds = LOOP_VINFO_LOWER_BOUNDS (loop_vinfo); | |
3134 | for (unsigned int i = 0; i < lower_bounds.length (); ++i) | |
3135 | { | |
3136 | tree expr = lower_bounds[i].expr; | |
3137 | tree type = unsigned_type_for (TREE_TYPE (expr)); | |
3138 | expr = fold_convert (type, expr); | |
3139 | poly_uint64 bound = lower_bounds[i].min_value; | |
3140 | if (!lower_bounds[i].unsigned_p) | |
3141 | { | |
3142 | expr = fold_build2 (PLUS_EXPR, type, expr, | |
3143 | build_int_cstu (type, bound - 1)); | |
3144 | bound += bound - 1; | |
3145 | } | |
3146 | tree part_cond_expr = fold_build2 (GE_EXPR, boolean_type_node, expr, | |
3147 | build_int_cstu (type, bound)); | |
3148 | chain_cond_expr (cond_expr, part_cond_expr); | |
3149 | } | |
3150 | } | |
3151 | ||
ebfd146a IR |
3152 | /* Function vect_create_cond_for_alias_checks. |
3153 | ||
3154 | Create a conditional expression that represents the run-time checks for | |
3155 | overlapping of address ranges represented by a list of data references | |
3156 | relations passed as input. | |
3157 | ||
3158 | Input: | |
3159 | COND_EXPR - input conditional expression. New conditions will be chained | |
a05a89fa CH |
3160 | with logical AND operation. If it is NULL, then the function |
3161 | is used to return the number of alias checks. | |
ebfd146a IR |
3162 | LOOP_VINFO - field LOOP_VINFO_MAY_ALIAS_STMTS contains the list of ddrs |
3163 | to be checked. | |
3164 | ||
3165 | Output: | |
3166 | COND_EXPR - conditional expression. | |
ebfd146a | 3167 | |
a05a89fa | 3168 | The returned COND_EXPR is the conditional expression to be used in the if |
ebfd146a IR |
3169 | statement that controls which version of the loop gets executed at runtime. |
3170 | */ | |
3171 | ||
a05a89fa | 3172 | void |
4bdd44c4 | 3173 | vect_create_cond_for_alias_checks (loop_vec_info loop_vinfo, tree * cond_expr) |
ebfd146a | 3174 | { |
93bdc3ed | 3175 | vec<dr_with_seg_len_pair_t> comp_alias_ddrs = |
a05a89fa | 3176 | LOOP_VINFO_COMP_ALIAS_DDRS (loop_vinfo); |
ebfd146a | 3177 | |
a05a89fa | 3178 | if (comp_alias_ddrs.is_empty ()) |
ebfd146a IR |
3179 | return; |
3180 | ||
9cbd2d97 BC |
3181 | create_runtime_alias_checks (LOOP_VINFO_LOOP (loop_vinfo), |
3182 | &comp_alias_ddrs, cond_expr); | |
73fbfcad | 3183 | if (dump_enabled_p ()) |
ccb3ad87 | 3184 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 3185 | "created %u versioning for alias checks.\n", |
a05a89fa | 3186 | comp_alias_ddrs.length ()); |
ebfd146a IR |
3187 | } |
3188 | ||
3189 | ||
3190 | /* Function vect_loop_versioning. | |
b8698a0f | 3191 | |
ebfd146a IR |
3192 | If the loop has data references that may or may not be aligned or/and |
3193 | has data reference relations whose independence was not proven then | |
3194 | two versions of the loop need to be generated, one which is vectorized | |
3195 | and one which isn't. A test is then generated to control which of the | |
3196 | loops is executed. The test checks for the alignment of all of the | |
3197 | data references that may or may not be aligned. An additional | |
3198 | sequence of runtime tests is generated for each pairs of DDRs whose | |
b8698a0f L |
3199 | independence was not proven. The vectorized version of loop is |
3200 | executed only if both alias and alignment tests are passed. | |
3201 | ||
ebfd146a | 3202 | The test generated to check which version of loop is executed |
b8698a0f | 3203 | is modified to also check for profitability as indicated by the |
d9157f15 | 3204 | cost model threshold TH. |
86290011 RG |
3205 | |
3206 | The versioning precondition(s) are placed in *COND_EXPR and | |
d68d56b5 | 3207 | *COND_EXPR_STMT_LIST. */ |
ebfd146a | 3208 | |
99b1c316 | 3209 | class loop * |
9c4f0d31 RB |
3210 | vect_loop_versioning (loop_vec_info loop_vinfo, |
3211 | gimple *loop_vectorized_call) | |
ebfd146a | 3212 | { |
99b1c316 MS |
3213 | class loop *loop = LOOP_VINFO_LOOP (loop_vinfo), *nloop; |
3214 | class loop *scalar_loop = LOOP_VINFO_SCALAR_LOOP (loop_vinfo); | |
ebfd146a | 3215 | basic_block condition_bb; |
538dd0b7 DM |
3216 | gphi_iterator gsi; |
3217 | gimple_stmt_iterator cond_exp_gsi; | |
ebfd146a IR |
3218 | basic_block merge_bb; |
3219 | basic_block new_exit_bb; | |
3220 | edge new_exit_e, e; | |
538dd0b7 | 3221 | gphi *orig_phi, *new_phi; |
368117e8 | 3222 | tree cond_expr = NULL_TREE; |
d68d56b5 | 3223 | gimple_seq cond_expr_stmt_list = NULL; |
ebfd146a | 3224 | tree arg; |
af2bbc51 | 3225 | profile_probability prob = profile_probability::likely (); |
ebfd146a | 3226 | gimple_seq gimplify_stmt_list = NULL; |
fdd29374 | 3227 | tree scalar_loop_iters = LOOP_VINFO_NITERSM1 (loop_vinfo); |
9cc1fb4b XDL |
3228 | bool version_align = LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo); |
3229 | bool version_alias = LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo); | |
01d32b2b | 3230 | bool version_niter = LOOP_REQUIRES_VERSIONING_FOR_NITERS (loop_vinfo); |
a421fe9e AV |
3231 | poly_uint64 versioning_threshold |
3232 | = LOOP_VINFO_VERSIONING_THRESHOLD (loop_vinfo); | |
4e65deef JJ |
3233 | tree version_simd_if_cond |
3234 | = LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (loop_vinfo); | |
a421fe9e | 3235 | unsigned th = LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo); |
ebfd146a | 3236 | |
6eed64b9 | 3237 | if (vect_apply_runtime_profitability_check_p (loop_vinfo) |
a421fe9e | 3238 | && !ordered_p (th, versioning_threshold)) |
80be3333 | 3239 | cond_expr = fold_build2 (GE_EXPR, boolean_type_node, scalar_loop_iters, |
01d32b2b | 3240 | build_int_cst (TREE_TYPE (scalar_loop_iters), |
fdd29374 | 3241 | th - 1)); |
a696bc4f RS |
3242 | if (maybe_ne (versioning_threshold, 0U)) |
3243 | { | |
3244 | tree expr = fold_build2 (GE_EXPR, boolean_type_node, scalar_loop_iters, | |
3245 | build_int_cst (TREE_TYPE (scalar_loop_iters), | |
3246 | versioning_threshold - 1)); | |
3247 | if (cond_expr) | |
3248 | cond_expr = fold_build2 (BIT_AND_EXPR, boolean_type_node, | |
3249 | expr, cond_expr); | |
3250 | else | |
3251 | cond_expr = expr; | |
3252 | } | |
01d32b2b BC |
3253 | |
3254 | if (version_niter) | |
3255 | vect_create_cond_for_niters_checks (loop_vinfo, &cond_expr); | |
3256 | ||
3257 | if (cond_expr) | |
2778a719 RB |
3258 | cond_expr = force_gimple_operand_1 (unshare_expr (cond_expr), |
3259 | &cond_expr_stmt_list, | |
01d32b2b | 3260 | is_gimple_condexpr, NULL_TREE); |
ebfd146a | 3261 | |
9cc1fb4b | 3262 | if (version_align) |
d68d56b5 RG |
3263 | vect_create_cond_for_align_checks (loop_vinfo, &cond_expr, |
3264 | &cond_expr_stmt_list); | |
ebfd146a | 3265 | |
9cc1fb4b | 3266 | if (version_alias) |
9adee305 RS |
3267 | { |
3268 | vect_create_cond_for_unequal_addrs (loop_vinfo, &cond_expr); | |
a57776a1 | 3269 | vect_create_cond_for_lower_bounds (loop_vinfo, &cond_expr); |
9adee305 RS |
3270 | vect_create_cond_for_alias_checks (loop_vinfo, &cond_expr); |
3271 | } | |
86290011 | 3272 | |
4e65deef JJ |
3273 | if (version_simd_if_cond) |
3274 | { | |
3275 | gcc_assert (dom_info_available_p (CDI_DOMINATORS)); | |
3276 | if (flag_checking) | |
3277 | if (basic_block bb | |
3278 | = gimple_bb (SSA_NAME_DEF_STMT (version_simd_if_cond))) | |
3279 | gcc_assert (bb != loop->header | |
3280 | && dominated_by_p (CDI_DOMINATORS, loop->header, bb) | |
3281 | && (scalar_loop == NULL | |
3282 | || (bb != scalar_loop->header | |
3283 | && dominated_by_p (CDI_DOMINATORS, | |
3284 | scalar_loop->header, bb)))); | |
3285 | tree zero = build_zero_cst (TREE_TYPE (version_simd_if_cond)); | |
3286 | tree c = fold_build2 (NE_EXPR, boolean_type_node, | |
3287 | version_simd_if_cond, zero); | |
3288 | if (cond_expr) | |
3289 | cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
3290 | c, cond_expr); | |
3291 | else | |
3292 | cond_expr = c; | |
3293 | if (dump_enabled_p ()) | |
3294 | dump_printf_loc (MSG_NOTE, vect_location, | |
3295 | "created versioning for simd if condition check.\n"); | |
3296 | } | |
3297 | ||
b210f45f RB |
3298 | cond_expr = force_gimple_operand_1 (unshare_expr (cond_expr), |
3299 | &gimplify_stmt_list, | |
d68d56b5 RG |
3300 | is_gimple_condexpr, NULL_TREE); |
3301 | gimple_seq_add_seq (&cond_expr_stmt_list, gimplify_stmt_list); | |
ebfd146a | 3302 | |
2778a719 RB |
3303 | /* Compute the outermost loop cond_expr and cond_expr_stmt_list are |
3304 | invariant in. */ | |
99b1c316 | 3305 | class loop *outermost = outermost_invariant_loop_for_expr (loop, cond_expr); |
2778a719 RB |
3306 | for (gimple_stmt_iterator gsi = gsi_start (cond_expr_stmt_list); |
3307 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
5ce9450f | 3308 | { |
2778a719 RB |
3309 | gimple *stmt = gsi_stmt (gsi); |
3310 | update_stmt (stmt); | |
3311 | ssa_op_iter iter; | |
3312 | use_operand_p use_p; | |
3313 | basic_block def_bb; | |
3314 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE) | |
3315 | if ((def_bb = gimple_bb (SSA_NAME_DEF_STMT (USE_FROM_PTR (use_p)))) | |
3316 | && flow_bb_inside_loop_p (outermost, def_bb)) | |
3317 | outermost = superloop_at_depth (loop, bb_loop_depth (def_bb) + 1); | |
3318 | } | |
5ce9450f | 3319 | |
2778a719 RB |
3320 | /* Search for the outermost loop we can version. Avoid versioning of |
3321 | non-perfect nests but allow if-conversion versioned loops inside. */ | |
99b1c316 | 3322 | class loop *loop_to_version = loop; |
2778a719 RB |
3323 | if (flow_loop_nested_p (outermost, loop)) |
3324 | { | |
3325 | if (dump_enabled_p ()) | |
3326 | dump_printf_loc (MSG_NOTE, vect_location, | |
3327 | "trying to apply versioning to outer loop %d\n", | |
3328 | outermost->num); | |
3329 | if (outermost->num == 0) | |
3330 | outermost = superloop_at_depth (loop, 1); | |
3331 | /* And avoid applying versioning on non-perfect nests. */ | |
3332 | while (loop_to_version != outermost | |
3333 | && single_exit (loop_outer (loop_to_version)) | |
3334 | && (!loop_outer (loop_to_version)->inner->next | |
3335 | || vect_loop_vectorized_call (loop_to_version)) | |
3336 | && (!loop_outer (loop_to_version)->inner->next | |
3337 | || !loop_outer (loop_to_version)->inner->next->next)) | |
3338 | loop_to_version = loop_outer (loop_to_version); | |
3339 | } | |
3340 | ||
3341 | /* Apply versioning. If there is already a scalar version created by | |
b614fca2 RB |
3342 | if-conversion re-use that. Note we cannot re-use the copy of |
3343 | an if-converted outer-loop when vectorizing the inner loop only. */ | |
2778a719 | 3344 | gcond *cond; |
b614fca2 | 3345 | if ((!loop_to_version->inner || loop == loop_to_version) |
9c4f0d31 | 3346 | && loop_vectorized_call) |
2778a719 RB |
3347 | { |
3348 | gcc_assert (scalar_loop); | |
9c4f0d31 RB |
3349 | condition_bb = gimple_bb (loop_vectorized_call); |
3350 | cond = as_a <gcond *> (last_stmt (condition_bb)); | |
2778a719 RB |
3351 | gimple_cond_set_condition_from_tree (cond, cond_expr); |
3352 | update_stmt (cond); | |
3353 | ||
3354 | if (cond_expr_stmt_list) | |
3355 | { | |
9c4f0d31 | 3356 | cond_exp_gsi = gsi_for_stmt (loop_vectorized_call); |
2778a719 RB |
3357 | gsi_insert_seq_before (&cond_exp_gsi, cond_expr_stmt_list, |
3358 | GSI_SAME_STMT); | |
3359 | } | |
3360 | ||
5cee3239 RB |
3361 | /* if-conversion uses profile_probability::always () for both paths, |
3362 | reset the paths probabilities appropriately. */ | |
3363 | edge te, fe; | |
3364 | extract_true_false_edges_from_block (condition_bb, &te, &fe); | |
3365 | te->probability = prob; | |
3366 | fe->probability = prob.invert (); | |
3367 | /* We can scale loops counts immediately but have to postpone | |
3368 | scaling the scalar loop because we re-use it during peeling. */ | |
3369 | scale_loop_frequencies (loop_to_version, te->probability); | |
3370 | LOOP_VINFO_SCALAR_LOOP_SCALING (loop_vinfo) = fe->probability; | |
3371 | ||
2778a719 RB |
3372 | nloop = scalar_loop; |
3373 | if (dump_enabled_p ()) | |
3374 | dump_printf_loc (MSG_NOTE, vect_location, | |
3375 | "reusing %sloop version created by if conversion\n", | |
3376 | loop_to_version != loop ? "outer " : ""); | |
5ce9450f JJ |
3377 | } |
3378 | else | |
2778a719 RB |
3379 | { |
3380 | if (loop_to_version != loop | |
3381 | && dump_enabled_p ()) | |
3382 | dump_printf_loc (MSG_NOTE, vect_location, | |
3383 | "applying loop versioning to outer loop %d\n", | |
3384 | loop_to_version->num); | |
3385 | ||
3386 | initialize_original_copy_tables (); | |
3387 | nloop = loop_version (loop_to_version, cond_expr, &condition_bb, | |
3388 | prob, prob.invert (), prob, prob.invert (), true); | |
3389 | gcc_assert (nloop); | |
3390 | nloop = get_loop_copy (loop); | |
3391 | ||
3392 | /* Kill off IFN_LOOP_VECTORIZED_CALL in the copy, nobody will | |
3393 | reap those otherwise; they also refer to the original | |
3394 | loops. */ | |
99b1c316 | 3395 | class loop *l = loop; |
2778a719 RB |
3396 | while (gimple *call = vect_loop_vectorized_call (l)) |
3397 | { | |
3398 | call = SSA_NAME_DEF_STMT (get_current_def (gimple_call_lhs (call))); | |
3399 | fold_loop_internal_call (call, boolean_false_node); | |
3400 | l = loop_outer (l); | |
3401 | } | |
3402 | free_original_copy_tables (); | |
3403 | ||
3404 | if (cond_expr_stmt_list) | |
3405 | { | |
3406 | cond_exp_gsi = gsi_last_bb (condition_bb); | |
3407 | gsi_insert_seq_before (&cond_exp_gsi, cond_expr_stmt_list, | |
3408 | GSI_SAME_STMT); | |
3409 | } | |
3410 | ||
3411 | /* Loop versioning violates an assumption we try to maintain during | |
3412 | vectorization - that the loop exit block has a single predecessor. | |
3413 | After versioning, the exit block of both loop versions is the same | |
3414 | basic block (i.e. it has two predecessors). Just in order to simplify | |
3415 | following transformations in the vectorizer, we fix this situation | |
3416 | here by adding a new (empty) block on the exit-edge of the loop, | |
3417 | with the proper loop-exit phis to maintain loop-closed-form. | |
3418 | If loop versioning wasn't done from loop, but scalar_loop instead, | |
3419 | merge_bb will have already just a single successor. */ | |
3420 | ||
3421 | merge_bb = single_exit (loop_to_version)->dest; | |
3422 | if (EDGE_COUNT (merge_bb->preds) >= 2) | |
3423 | { | |
3424 | gcc_assert (EDGE_COUNT (merge_bb->preds) >= 2); | |
3425 | new_exit_bb = split_edge (single_exit (loop_to_version)); | |
3426 | new_exit_e = single_exit (loop_to_version); | |
3427 | e = EDGE_SUCC (new_exit_bb, 0); | |
3428 | ||
3429 | for (gsi = gsi_start_phis (merge_bb); !gsi_end_p (gsi); | |
3430 | gsi_next (&gsi)) | |
3431 | { | |
3432 | tree new_res; | |
3433 | orig_phi = gsi.phi (); | |
3434 | new_res = copy_ssa_name (PHI_RESULT (orig_phi)); | |
3435 | new_phi = create_phi_node (new_res, new_exit_bb); | |
3436 | arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, e); | |
3437 | add_phi_arg (new_phi, arg, new_exit_e, | |
3438 | gimple_phi_arg_location_from_edge (orig_phi, e)); | |
3439 | adjust_phi_and_debug_stmts (orig_phi, e, PHI_RESULT (new_phi)); | |
3440 | } | |
3441 | } | |
3442 | ||
3443 | update_ssa (TODO_update_ssa); | |
3444 | } | |
01d32b2b BC |
3445 | |
3446 | if (version_niter) | |
3447 | { | |
3448 | /* The versioned loop could be infinite, we need to clear existing | |
3449 | niter information which is copied from the original loop. */ | |
3450 | gcc_assert (loop_constraint_set_p (loop, LOOP_C_FINITE)); | |
3451 | vect_free_loop_info_assumptions (nloop); | |
3452 | /* And set constraint LOOP_C_INFINITE for niter analyzer. */ | |
3453 | loop_constraint_set (loop, LOOP_C_INFINITE); | |
3454 | } | |
9cc1fb4b | 3455 | |
4f5b9c80 | 3456 | if (LOCATION_LOCUS (vect_location.get_location_t ()) != UNKNOWN_LOCATION |
9cc1fb4b XDL |
3457 | && dump_enabled_p ()) |
3458 | { | |
3459 | if (version_alias) | |
7db960c5 DM |
3460 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS | MSG_PRIORITY_USER_FACING, |
3461 | vect_location, | |
9cc1fb4b XDL |
3462 | "loop versioned for vectorization because of " |
3463 | "possible aliasing\n"); | |
3464 | if (version_align) | |
7db960c5 DM |
3465 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS | MSG_PRIORITY_USER_FACING, |
3466 | vect_location, | |
9cc1fb4b XDL |
3467 | "loop versioned for vectorization to enhance " |
3468 | "alignment\n"); | |
3469 | ||
3470 | } | |
03001a35 RB |
3471 | |
3472 | return nloop; | |
ebfd146a | 3473 | } |